text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We investigate a fundamental question regarding a benchmark class of shapes in one of the simplest, yet most widely utilized abstract models of algorithmic tile self-assembly. More specifically, we study the directed tile complexity of a $k \times N$ thin rectangle in Winfree’s ubiquitous abstract Tile Assembly Model, assuming that cooperative binding cannot be enforced (temperature-1 self-assembly) and that tiles are allowed to be placed at most one step into the third dimension (just-barely 3D). While the directed tile complexities of a square and a scaled-up version of any algorithmically specified shape at temperature 1 in just-barely 3D are both asymptotically the same as they are (respectively) at temperature 2 in 2D, the (loose) bounds on the directed tile complexity of a thin rectangle at temperature 2 in 2D are not currently known to hold at temperature 1 in just-barely 3D. Motivated by this discrepancy, we establish new lower and upper bounds on the directed tile complexity of a thin rectangle at temperature 1 in just-barely 3D. On our way to proving our lower bound, we develop a new, more powerful type of specialized Window Movie Lemma that lets us derive our lower bound via a counting argument, where we upper bound the number of “sufficiently similar” ways to assign glues to a set (rather than a sequence) of fixed locations. Consequently, our lower bound, $\Omega\left( N^{\frac{1}{k}} \right)$, is an asymptotic improvement over the previous state of the art lower bound and is more aesthetically pleasing since it eliminates the non-constant term $k$ that used to divide $N^{\frac{1}{k}}$. The proof of our upper bound is based on the construction of a novel, just-barely 3D temperature-1 counter, organized according to “digit regions”, which affords it roughly fifty percent more digits for the same target rectangle compared to the previous state of the art counter. This increase in digit density results in an upper bound of $O\left( N^{\frac{1}{\left \lfloor \frac{k}{2} \right \rfloor}} + \log N \right)$, that is an asymptotic improvement over the previous state of the art upper bound and roughly the square of our lower bound.'
author:
- 'David Furcy[^1]'
- 'Scott M. Summers[^2]'
- 'Logan Withers[^3]'
bibliography:
- 'tam.bib'
title: '**Improved lower and upper bounds on the tile complexity of uniquely self-assembling a thin rectangle non-cooperatively in 3D**'
---
[^1]: Computer Science Department, University of Wisconsin Oshkosh, Oshkosh, WI 54901, USA,[[email protected]]([email protected]).
[^2]: Computer Science Department, University of Wisconsin Oshkosh, Oshkosh, WI 54901, USA,[[email protected]]([email protected]).
[^3]: Computer Science Department, University of Wisconsin Oshkosh, Oshkosh, WI 54901, USA,[[email protected]]([email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $\phi$ be a spherical Hecke–Maa[ß]{} cusp form on the non-compact space ${\mathrm{PGL}}_3({\mathbb{Z}}){\backslash}{\mathrm{PGL}}_3({\mathbb{R}})$. We establish various pointwise upper bounds for $\phi$ in terms of its Laplace eigenvalue $\lambda_\phi$. These imply, for $\phi$ arithmetically normalized and tempered at the archimedean place, the bound $$\|\phi\|_\infty\ll_{\varepsilon}\lambda_{\phi}^{39/40+{\varepsilon}}$$ for the global sup-norm (without restriction to a compact subset). On the way, we derive a new uniform upper bound for the ${\mathrm{GL}}_3$ Jacquet–Whittaker function.'
address:
- 'Mathematisches Institut, Bunsenstr. 3-5, D-37073 Göttingen, Germany'
- 'Alfréd Rényi Institute of Mathematics, Hungarian Academy of Sciences, POB 127, Budapest H-1364, Hungary'
- 'Central European University, Nador u. 9, Budapest H-1051, Hungary'
- 'Alfréd Rényi Institute of Mathematics, Hungarian Academy of Sciences, POB 127, Budapest H-1364, Hungary'
author:
- Valentin Blomer
- Gergely Harcos
- Péter Maga
title: 'On the global sup-norm of ${\mathrm{GL}}(3)$ cusp forms'
---
[^1]
Introduction
============
Eigenfunctions of the Laplace–Beltrami operator $\Delta$ on a Riemannian manifold $X$ are the spectral building blocks of the Hilbert space $L^2(X)$. It is therefore a classical question to study their analytic properties asymptotically as the eigenvalue tends to infinity. In addition to its intrinsic interest in global analysis, a relation to quantum physics is provided by the fact that $e^{-it\Delta}$ is the time evolution operator of the Schrödinger equation describing a freely moving particle on $X$. Therefore, an $L^2$-eigenfunction of $\Delta$ is understood in quantum physics as a bound state, and its absolute square is interpreted as the probability density of the corresponding stationary wave. Somewhat unexpectedly, number theory also enters the scene, namely when the manifold $X$ possesses some additional arithmetic structure such as a commutative family of arithmetically defined Hecke operators (which are normal and commute with $\Delta$ as well). In this case, the most interesting Laplace eigenfunctions are those that are in addition eigenfunctions of the Hecke algebra, and as such are amenable to number theoretic tools. We will present an example of this kind in a moment.
A fundamental quantity associated to an $L^2$-normalized eigenfunction $\phi$ with eigenvalue $\lambda_{\phi}$ is its sup-norm. A good upper bound for $\| \phi \|_{\infty}$ can be seen as a basic measure of equidistribution of its mass on $X$. If $X$ is compact (or $X$ is non-compact but $\phi$ is restricted to a fixed compact subset $\Omega\subseteq X$), then we have the general bound proved by Hörmander [@Ho] $$\label{laplace1}
\|\phi\|_{\infty}\ll \lambda^{(d-1)/4},\qquad d=\dim X.$$ This bound is sharp, e.g. it is attained for the $d$-sphere $S^d$ for any $d{\geqslant}1$ and for special eigenfunctions $\phi$. If $X$ is a compact locally symmetric space of dimension $d$ and rank $r$ (or we restrict to a compact subset $\Omega$ of such a space) and we require that $\phi$ is not only a Laplace eigenfunction but an eigenfunction of all differential operators invariant under the group of isometries of the universal cover of $X$, then we have the stronger bound of Sarnak [@Sa] $$\label{laplace2}
\|\phi\|_{\infty}\ll \lambda^{(d-r)/4}.$$ Even though neither nor are conjectured to be sharp for negatively curved manifolds, they provide a robust framework to work with.
The situation changes completely for the global sup-norm on non-compact manifolds, in which case and no longer need to be true. A typical example is a locally symmetric space $X = \Gamma \backslash G/K$, where $G$ is a non-compact semi-simple Lie group, $K{\leqslant}G$ is a maximal compact subgroup, and $\Gamma{\leqslant}G$ is a non-uniform lattice. It turns out that in this case the sup-norm of an eigenfunction $\phi$ is often determined by its behavior in the cuspidal regions of $X$, even though it may eventually decay very quickly. The simplest – and the only thoroughly explored – example is $G ={\mathrm{SL}}_2({\mathbb{R}})$, $K ={\mathrm{SO}}_2({\mathbb{R}})$, $\Gamma={\mathrm{SL}}_2({\mathbb{Z}})$, so that $X = \Gamma\backslash G/K = \Gamma \backslash \mathcal{H}_2$ (where $\mathcal{H}_2 = \{x + iy:\ y > 0\}$) is the familiar modular surface. It is an *arithmetic manifold* in the above sense, as it admits the standard family of Hecke operators. A joint $L^2$-eigenfunction of the Laplace and the Hecke operators is often called a *Hecke–Maa[ß]{} cusp form*. It decays exponentially as $y \rightarrow \infty$, and it is conjectured to satisfy $\|\phi|_{\Omega}\|_\infty\ll_{{\varepsilon},\Omega} \lambda_{\phi}^{{\varepsilon}}$ as $\lambda_{\phi} \rightarrow \infty$ for every compact subset $\Omega \subseteq X$ and every ${\varepsilon}> 0$, but in the cuspidal region it has considerable size. Precisely, we have $\phi(i\lambda_\phi^{1/2}/(2\pi))=\lambda_\phi^{1/12+o(1)}$, cf. [@Sa], reflecting the very similar behavior of the normalized ${\mathrm{GL}}_2$ Whittaker function $$\label{gl2whittaker}
W_{\nu}(x):=\frac{\sqrt{x}K_{\nu}(2\pi x)}{|\Gamma(1/2+\nu)|},\qquad\nu\in i{\mathbb{R}}.$$ Indeed, $W_\nu(x)$ decays exponentially as $x\rightarrow \infty$, but it has a large bump: $W_{it}(t/(2\pi)) \asymp t^{1/6}$ (see [@Ba] for a uniform asymptotic expansion).\
A systematic study of this behavior for Hecke–Maa[ß]{} cusp forms on the locally symmetric space $$X=X_n={\mathrm{GL}}_n({\mathbb{Z}}){\mathrm{Z}}({\mathbb{R}}){\backslash}{\mathrm{GL}}_n({\mathbb{R}})/{\mathrm{O}}_n({\mathbb{R}})$$ and its congruence covers was initiated by Brumley and Templier [@BT]. Here, ${\mathrm{Z}}({\mathbb{R}})$ is the center of ${\mathrm{GL}}_n({\mathbb{R}})$ and ${\mathrm{O}}_n({\mathbb{R}})$ is the orthogonal subgroup. In particular, it was proved in [@BT] that fails for $X=X_n$ when $n {\geqslant}5$, namely the *global* sup-norm on $X$ is significantly larger than the *local* sup-norm on a fixed compact subset $\Omega \subseteq X$. Precisely, the lower bound by Brumley–Templier [@BT] for the global sup-norm and the upper bound by Blomer–Maga [@BM] for the local sup-norm can be contrasted as $$\|\phi\|_{\infty}\gg_{\varepsilon}\lambda_\phi^{n(n-1)(n-2)/24-{\varepsilon}}>\lambda_\phi^{n(n-1)/8-\delta_n}\gg_{\Omega} \|\phi|_{\Omega}\|_\infty,$$ where $\delta_n>0$ is a constant depending only on $n$. In addition, in the case $n=3$, Brumley and Templier derived the upper bound [@BT Prop. 1.6] $$\label{BrumleyTemplier}
\| \phi \|_{\infty} \ll \lambda_{\phi}^{5/2}\qquad\text{on}\qquad X=X_3,$$ by using the rapid decay of $\phi$ high in the cusp and making the dependence of on the injectivity radius in the remaining piece of the manifold explicit. We note in passing that in the case $n=3$, it is known that any savings $\delta_3<1/124$ is admissible[^2] for the upper bound of the local sup-norm [@HRR], while in the case $n=2$, the global sup-norm problem has been studied extensively (see [@IS; @Sah1; @Sah2; @BHMM] and the references therein).
Despite these important advances, the investigation of the global sup-norm of eigenfunctions on non-compact symmetric spaces of rank exceeding one is still its infancy, and in this article we take a closer look at the rank two example $n=3$ with the aim of proving considerably stronger bounds than by a different technique. With this in mind, let $\phi$ be a Hecke–Maa[ß]{} cusp form on ${\mathrm{GL}}_3$ over ${\mathbb{Q}}$, which is spherical at every place and has trivial central character, regarded as a complex-valued function on the quotient space $X=X_3$. Alternatively, we may think of $\phi$ as a complex-valued function on ${\mathrm{GL}}_3({\mathbb{R}})$ satisfying $$\phi(\gamma h g k) = \phi(g)\qquad\text{for all}\qquad\gamma\in{\mathrm{GL}}_3({\mathbb{Z}}),\ h\in{\mathrm{Z}}({\mathbb{R}}),\ k\in{\mathrm{O}}_3({\mathbb{R}}),$$ which is further an eigenfunction of the invariant differential operators and the Hecke operators [@G Sections 6.1–6.4]. We recall from [@G Sections 1.2–1.3] that the symmetric space ${\mathrm{Z}}({\mathbb{R}}){\backslash}{\mathrm{GL}}_3({\mathbb{R}}) / {\mathrm{O}}_3({\mathbb{R}})$ can be represented by matrices of the generalized upper half-plane $${\mathcal{H}}_3:=\left\{z=\begin{pmatrix}1 & x_2 & x_3 \\ & 1 & x_1 \\ & & 1\end{pmatrix}
\begin{pmatrix}y_1y_2 & & \\ & y_1 & \\ & & 1\end{pmatrix}:\ y_1,y_2>0,\ x_1,x_2,x_3\in{\mathbb{R}}\right\},$$ and the quotient space $X={\mathrm{GL}}_3({\mathbb{Z}}){\backslash}{\mathcal{H}}_3$ has a fundamental domain lying in the Siegel set defined by $$\label{xybounds}
|x_1|,|x_2|,|x_3|{\leqslant}1/2\qquad\text{and}\qquad y_1,y_2{\geqslant}\sqrt{3}/2.$$
In this paper, we provide upper bounds for $|\phi(z)|$ in terms of the height parameters $y_1$, $y_2$ (assuming they are at least $\sqrt{3}/2$) and the Laplace eigenvalue $\lambda_{\phi}$, and examine what they yield for the global sup-norm $$\|\phi\|_\infty=\sup_{z\in{\mathcal{H}}_3}|\phi(z)|.$$ We shall work with the following natural normalizations:
- $\phi$ is *arithmetically normalized* if it has leading Fourier coefficient $\lambda_\phi(1,1)=1$ with respect to the standard Jacquet–Whittaker function (cf. [@G Thm. 6.4.11] and below);\
- $\phi$ is *$L^2$-normalized* if it has $L^2$-norm $1$ with respect to the measure on $X$ induced by the standard left-invariant probability measure on ${\mathcal{H}}_3$ (cf. [@G Prop. 1.5.3]).
By [@B Lemma 1] and its proof, these two normalizations differ by a positive constant times $L(1,\phi,\operatorname{Ad})^{1/2}$ when $\phi$ is tempered at the archimedean place, and this would also be true for non-tempered forms if we slightly renormalized the standard Jacquet–Whittaker function as in the display below [@B (2.13)] with the effect of correspondingly changing the notion of “arithmetically normalized”. With these conventions, our main results are as follows (see also the remarks after the theorems):
\[theorem\_2\] Let $\phi$ be an arithmetically normalized Hecke–Maa[ß]{} cusp form on $X$. Assume that $\phi$ is tempered at the archimedean place. Then for any $z\in{\mathcal{H}}_3$ with $y_1,y_2{\geqslant}\sqrt{3}/2$, and for any ${\varepsilon}>0$, we have $$\label{thm2bound}
\phi(z)\ll_{\varepsilon}\min(y_1,y_2)\left(\frac{\lambda_{\phi}^{1+{\varepsilon}}}{y_1y_2} + \frac{\lambda_{\phi}^{3/2+{\varepsilon}}}{(y_1y_2)^2}\right).$$
\[theorem\_3\] Let $\phi$ be an $L^2$-normalized Hecke–Maa[ß]{} cusp form on $X$. Then for any $z\in{\mathcal{H}}_3$ with $y_1,y_2{\geqslant}\sqrt{3}/2$, we have $$\label{thm3bound}\phi(z)\ll\lambda_{\phi}^{3/4} + \lambda_{\phi}^{5/8}y_1y_2.$$
\[theorem\_1\] Let $\phi$ be an arithmetically normalized Hecke–Maa[ß]{} cusp form on $X$. Assume that $\phi$ is tempered at the archimedean place. Then for any ${\varepsilon}>0$, we have $$\label{thm1bound}\|\phi\|_\infty\ll_{\varepsilon}\lambda_{\phi}^{39/40+{\varepsilon}}.$$
\[remarks\] Assume that $\phi$ is tempered at the archimedean place. If $\phi$ is arithmetically normalized, then holds with the extra factor $L(1,\phi,\operatorname{Ad})^{1/2}\ll_{\varepsilon}\lambda_{\phi}^{\varepsilon}$ on the right hand side by the work of Brumley [@Br Cor. 2] or Li [@L Thm. 2]. Similarly, if $\phi$ is $L^2$-normalized, then and hold with the extra factor $L(1,\phi,\operatorname{Ad})^{-1/2}\ll_{\varepsilon}\lambda_{\phi}^{1/2+{\varepsilon}}$ on the right hand side by the work of Brumley [@Br2 Thm. 3] (see also [@La Appendix]). If $\phi$ is self-dual, i.e. $\phi$ is the symmetric square of a classical (even or odd) Hecke–Maa[ß]{} cusp form on ${\mathrm{SL}}_2({\mathbb{Z}}){\backslash}{\mathcal{H}}_2$ (cf. [@R Thm. A]), we even know that $L(1,\phi,\operatorname{Ad})^{-1/2} \ll_{\varepsilon}\lambda_{\phi}^{\varepsilon}$ by a result of Ramakrishnan and Wang [@RW Cor. C], so that in this case no adjustment to and is necessary. Finally, we note that for $\phi$ self-dual, the exponent $39/40$ in cannot be lowered below $3/8$, as follows from the work of Brumley and Templier [@BT Cor. 1.10].
We prove our results by employing two very different methods. On the one hand, we estimate the Fourier–Whittaker expansion of $\phi$ termwise, which eventually leads to Theorem \[theorem\_2\]. Unlike in the rank one case $n=2$, the fact that the group of upper-triangular unipotent $3\times 3$ matrices is not commutative leads to an additional sum over an infinite subset of ${\mathrm{SL}}_2({\mathbb{Z}})$ which requires careful treatment. As a result of independent interest and as a contribution to the analytic theory of higher rank Whittaker functions, we provide in Lemma \[lemma:whittaker\_estimate\] a new uniform upper bound for the ${\mathrm{GL}}_3$ Jacquet–Whittaker function.
On the other hand, as in earlier approaches on compact spaces, we use a pre-trace formula, but we make the analysis of the geometric side uniform in the height of the considered point $z \in \mathcal{H}_3$; this is familiar for the group ${\mathrm{GL}}_2$, but seems to have not yet been worked out in higher rank. This leads to Theorem \[theorem\_3\]. A key ingredient of the proof is Lemma \[lemma:lower\_bound\_on\_cartan\_projection\], which states a simple but efficient bound for the norm of the Cartan projection of upper-triangular matrices. As we do not use amplification, the proof is independent of Hecke operators and consequently the result also holds for all $L^2$-normalized Maa[ß]{} forms, not just the Hecke eigenforms.
Our final Theorem \[theorem\_1\] is a combination of Theorems \[theorem\_2\] and \[theorem\_3\]. There is nothing particularly special about the exponent $39/40$, except that the corresponding bound is considerably stronger than and already fairly close to . Marginal improvements are possible, for instance by inserting an amplifier into the pre-trace formula.
We thank Gergő Nemes for valuable discussions concerning the special functions that appear in this paper. We also thank the referee whose careful reading and detailed comments helped us to improve the exposition.
Archimedean parameters and the Weyl group
=========================================
Associated to every Hecke–Maa[ß]{} cusp form $\phi$ on $X$ (in fact, the Hecke property is not needed for this discussion), there is an $S_3$-orbit of (archimedean) Langlands parameters $$\begin{aligned}
\label{mutriple}
(\mu_1,\mu_2,\mu_3)\in{\mathbb{C}}^3&\qquad\text{satisfying}\qquad\mu_1+\mu_2+\mu_3=0,\\
\intertext{and a corresponding $S_3$-orbit of (archimedean) spectral parameters}
\notag
(\nu_0,\nu_1,\nu_2)\in{\mathbb{C}}^3&\qquad\text{satisfying}\qquad \nu_0=\nu_1+\nu_2.\end{aligned}$$ We follow the conventions of [@Bu] except that we shift $\nu_1$ and $\nu_2$ there by $-1/3$, so that (cf. [@Bu (8.3)]) $$(\mu_1,\mu_2,\mu_3)=(\nu_1+2\nu_2,\nu_1-\nu_2,-2\nu_1-\nu_2),$$ $$\label{laplaceeigenvalue}
\lambda_\phi=1-\frac{1}{2}\left(\mu_1^2+\mu_2^2+\mu_3^2\right)=1-3\nu_1^2-3\nu_1\nu_2-3\nu_2^2,$$ and $(1/3+\nu_2,1/3+\nu_1)\in{\mathbb{C}}^2$ are the spectral parameters in [@G Section 6][^3]. The Weyl group $S_3$ acts by permutations on the Langlands parameters, hence an $S_3$-orbit of spectral parameters takes the shape $$\label{orbit}
\bigl\{(\nu_0,\nu_1,\nu_2),\ (\nu_2,-\nu_1,\nu_0),\ (-\nu_1,\nu_2,-\nu_0),\ (-\nu_0,-\nu_2,-\nu_1),\ (-\nu_2,-\nu_0,\nu_1),\ (\nu_1,\nu_0,-\nu_2)\bigr\}.$$
If $\phi$ is tempered at the archimedean place (as assumed in Theorems \[theorem\_2\] and \[theorem\_1\]), the Langlands parameters and the spectral parameters are purely imaginary. Moreover, each $S_3$-orbit of spectral parameters contains a unique triple with nonnegative imaginary parts: $$\label{eq:nu-sum}
(\nu_0,\nu_1,\nu_2)\in(i{\mathbb{R}}_{{\geqslant}0})^3\qquad\text{satisfying}\qquad \nu_0=\nu_1+\nu_2.$$ From now on we regard this triple as the spectral parameter of $\phi$. For convenience, we also introduce $$\label{T0}
T_0:=\max(2,|\nu_0|,|\nu_1|,|\nu_2|)=\max(2,|\nu_0|)\asymp\lambda_\phi^{1/2}.$$
The Fourier–Whittaker expansion
===============================
The restriction of $\phi$ to ${\mathcal{H}}_3$ admits a Fourier–Whittaker expansion (cf. [@G (6.2.1)]) $$\label{eq:f-w-expansion}
\phi(z)=\sum_{m_1=1}^{\infty} \sum_{m_2\neq 0} \frac{\lambda_{\phi}(m_1,m_2)}{|m_1m_2|} \sum_{\gamma\in{\mathrm{U}}_2({\mathbb{Z}}){\backslash}{\mathrm{SL}}_2({\mathbb{Z}})} {\mathcal{W}}_{\nu_1,\nu_2}^{\operatorname{sign}(m_2)} \left( \begin{pmatrix}
|m_1m_2| & & \\ & m_1 & \\ & & 1
\end{pmatrix} \begin{pmatrix}
\gamma & \\ & 1
\end{pmatrix} z \right),$$ where ${\mathrm{U}}_2$ is the subgroup of upper-triangular unipotent matrices $\bigl(\begin{smallmatrix}1 & * \\ & 1 \end{smallmatrix}\bigr)$ and ${\mathcal{W}}_{\nu_1,\nu_2}^{\pm}:{\mathrm{GL}}_3({\mathbb{R}})\to{\mathbb{C}}$ is the standard Jacquet–Whittaker function (cf. [@G (6.1.2)] and [@B (2.10)–(2.11)]). The function ${\mathcal{W}}_{\nu_1,\nu_2}^{\pm}$ is invariant under ${\mathrm{Z}}({\mathbb{R}})$, right-invariant under ${\mathrm{O}}_3({\mathbb{R}})$, and its restriction to ${\mathcal{H}}_3$ has the following integral representation by a result of Vinogradov–Takhtadzhyan [@VT] (cf. [@B (2.12)]): $$\begin{gathered}
\label{eq:whittaker}
\left(\prod_{j=0}^2\frac{\Gamma\left(\frac{1}{2} + \frac{3}{2}\nu_j\right)}{\pi^{\frac{1}{2}+\frac{3}{2}\nu_j}}\right)
{\mathcal{W}}_{\nu_1,\nu_2}^{\pm}\left(\begin{pmatrix}
1 & x_2 & x_3 \\ & 1 & x_1 \\ & & 1
\end{pmatrix} \begin{pmatrix}
y_1y_2 & & \\ & y_1 & \\ & & 1
\end{pmatrix}\right) = \\[6pt]
4 e(x_1 \pm x_2) y_1^{1+\frac{\nu_1-\nu_2}{2}}y_2^{1+\frac{\nu_2-\nu_1}{2}}
\int_0^{\infty} K_{\frac{3}{2}\nu_0}(2\pi y_1\sqrt{1+u}) K_{\frac{3}{2}\nu_0}(2\pi y_2\sqrt{1+u^{-1}}) u^{\frac{3}{4}(\nu_1-\nu_2)} \frac{du}{u},\end{gathered}$$ where $K_\nu$ is the usual $K$-Bessel function. This formula is a special case of [@S Thm. 2.1] with the constant $2^{n-1}$ corrected to $2^{2n-3}$ there, as recorded by [@S2 (4.3) & p. 132]. Indeed, follows readily from the first display of [@S p. 318], by updating the constant $4$ to $8$, then substituting $u^{1/2}$ for $u$, and finally noting that $(\nu_1,\nu_2)$ and $(y_1,y_2)$ there equal $(1/3+\nu_1,1/3+\nu_2)$ and $(y_2,y_1)$ in our notation[^4]. An independent verification of can be obtained by comparing carefully [@BB2 (5.8)] with [@Bu (10.1)] (or more directly with [@Bu (10.2)], after inserting a missing factor of $\pi^{-2}$ on the right hand side there).
We remark in passing that by [@G Prop. 5.5.2, Eq. (5.5.5), Thm. 5.9.8], the left hand side of is the same over the whole spectral $S_3$-orbit , hence in the tempered case (i.e. when $\nu_j\in i{\mathbb{R}}$) the corresponding Fourier coefficients $\lambda_\phi(m_1,m_2)$ only differ from each other by six constants on the unit circle. At any rate, since the unipotent part (i.e. the entries $x_1,x_2,x_3$) acts on the right hand side of via a rotation, it is convenient to introduce also $$\label{eq:whittaker_diagonal}
\begin{split}
\tilde{W}_{\nu_1,\nu_2}(y_1,y_2):= &\ 4\pi^{\frac{3}{2}} \prod_{j=0}^2\left| \Gamma\left(\frac{1}{2} + \frac{3}{2}\nu_j\right)\right|^{-1} y_1y_2 \left(\frac{y_1}{y_2}\right)^{\frac{\nu_1-\nu_2}{2}} \\ & \times \int_0^{\infty} K_{\frac{3}{2}\nu_0}(2\pi y_1\sqrt{1+u}) K_{\frac{3}{2}\nu_0}(2\pi y_2\sqrt{1+u^{-1}}) u^{\frac{3}{4}(\nu_1-\nu_2)} \frac{du}{u}.
\end{split}$$
An upper bound for the Jacquet–Whittaker function
=================================================
In this section we estimate, under the assumption , the function ${\mathcal{W}}_{\nu_1,\nu_2}^{\pm}$ given by , or equivalently the function $\tilde{W}_{\nu_1,\nu_2}$ given by . We start with a bound involving the normalized ${\mathrm{GL}}_2$ Whittaker function introduced in :
\[lemma:GL(2)\_whittaker\_integral\] Assume that $\nu\in i{\mathbb{R}}$, and put $T:=\max(1/2,|\nu|)$. Then, for any $t,A>0$, we have $$\int_t^{\infty} \frac{W_{\nu}^2(x)}{\sqrt{x^2-t^2}}\ dx \ll_A (\log(3T)) \left(1+\frac{t}{T}\right)^{-A}.$$
Applying the crude bound (cf. [@HM Prop. 9]) $$W_{\nu}(x)\ll e^{-\pi x},\qquad x>T,$$ we obtain $$\int_{\max(t,T)}^\infty \frac{W_{\nu}^2(x)}{\sqrt{x^2-t^2}}\ dx \ll
\frac{1}{\sqrt{t}} \int_t^\infty \frac{e^{-2\pi x}}{\sqrt{x-t}}\ dx \ll e^{-2\pi t},$$ hence it suffices to prove that $$\int_t^{T} \frac{W_{\nu}^2(x)}{\sqrt{x^2-t^2}}\ dx \ll \log(3T),\qquad 0<t<T.$$ Observing also that (cf. [@GR 6.576.4]) $$\int_{2t}^{\infty} \frac{W_{\nu}^2(x)}{\sqrt{x^2-t^2}}\ dx \ll \int_0^{\infty} \frac{W_{\nu}^2(x)}{x}\ dx = \frac{1}{8},$$ we are left with proving $$\int_t^{2t} \frac{W_{\nu}^2(x)}{\sqrt{x^2-t^2}}\ dx \ll \log(3T),\qquad 0<t<T.$$
To reformulate, we need to show $$\int_t^{2t} \frac{W_{\nu}^2(x)}{\sqrt{x-t}}\ dx \ll t^{1/2}\log(3T),\qquad 0<t<T,$$ and we shall accomplish this by using the bound (cf. [@BHo p. 679] and [@HM Prop. 9]) $$W_{\nu}(x)\ll \min\bigl(T^{1/6},T^{1/4}|2\pi x-T|^{-1/4}\bigr),\qquad 0<x<2T.$$ For $0<t{\leqslant}T/20$ this bound readily yields that $$\int_t^{2t} \frac{W_{\nu}^2(x)}{\sqrt{x-t}}\ dx \ll \int_t^{2t} \frac{1}{\sqrt{x-t}}\ dx \ll t^{1/2},$$ hence we can assume that $T/20<t<T$. Then, our task reduces to showing that $$\int_I \frac{T^{1/3}}{\sqrt{x-t}}\ dx + \sum_M \int_{I(M)} \frac{T^{1/2}M^{-1/2}}{\sqrt{x-t}}\ dx \ll T^{1/2}\log(3T),$$ where $$I:=\{x\in[t,2t]:\ |2\pi x-T|{\leqslant}T^{1/3}\},\qquad I(M):=\{x\in[t,2t]:\ M{\leqslant}|2\pi x-T|<2M\},$$ and $M$ runs through the numbers $M\in[T^{1/3},20T]$ of the form $M=2^kT^{1/3}$ with $k\in{\mathbb{N}}$. Note that $I\subset[t,2t]$ is either empty or an interval of length less than $T^{1/3}$, while $I(M)\subset[t,2t]$ is a disjoint union of at most two intervals of lengths less than $M$. Therefore, $$\int_I \frac{T^{1/3}}{\sqrt{x-t}}\ dx \ll T^{1/3}T^{1/6}=T^{1/2},\qquad\int_{I(M)} \frac{T^{1/2}M^{-1/2}}{\sqrt{x-t}}\ dx \ll T^{1/2}M^{-1/2}M^{1/2}=T^{1/2},$$ and we are done upon noting that the number of dyadic parameters $M$ is $\ll\log(3T)$.
\[lemma:whittaker\_estimate\] For any $y_1,y_2,A>0$ we have, under the assumption and with the notation , $$\tilde{W}_{\nu_1,\nu_2}(y_1,y_2)\ll_A
(\log T_0)\sqrt{y_1y_2}\left(1+\frac{y_1}{T_0}\right)^{-A}\left(1+\frac{y_2}{T_0}\right)^{-A}.$$
From Stirling’s approximation and , it follows that $$\prod_{j=0}^2 \left| \Gamma\left(\frac{1}{2} + \frac{3}{2}\nu_j\right)\right|^{-1} \asymp \left| \Gamma\left(\frac{1}{2}+\nu\right)\right|^{-2}
\qquad\text{with}\qquad \nu:=\frac{3}{2}\nu_0.$$ Then we see from and the Cauchy–Schwarz inequality that $$\begin{aligned}
\tilde{W}_{\nu_1,\nu_2}(y_1,y_2)
&\ll\sqrt{y_1y_2}\int_0^{\infty}\frac{\left|W_\nu(y_1\sqrt{1+u})W_\nu(y_2\sqrt{1+u^{-1}})\right|}
{(1+u)^{1/4}(1+u^{-1})^{1/4}}\ \frac{du}{u}\\
&\ll\sqrt{y_1y_2}\left(\int_0^{\infty}\frac{W_\nu^2(y_1\sqrt{1+u})}{\sqrt{1+u^{-1}}}\ \frac{du}{u}\right)^{1/2}
\left(\int_0^{\infty}\frac{W_\nu^2(y_2\sqrt{1+u^{-1}})}{\sqrt{1+u}}\ \frac{du}{u}\right)^{1/2}.\end{aligned}$$ Applying the change of variables $x:=y_1\sqrt{1+u}$ (resp. $x:=y_2\sqrt{1+u^{-1}}$) in the last two integrals, we get $$\tilde{W}_{\nu_1,\nu_2}(y_1,y_2)\ll\sqrt{y_1y_2}
\left(\int_{y_1}^{\infty} \frac{W_\nu^2(x)}{\sqrt{x^2-y_1^2}}\ dx\right)^{1/2}
\left(\int_{y_2}^{\infty} \frac{W_\nu^2(x)}{\sqrt{x^2-y_2^2}}\ dx\right)^{1/2}.$$ Finally, estimating these two integrals by Lemma \[lemma:GL(2)\_whittaker\_integral\], we are done.
Proof of Theorem \[theorem\_2\]
===============================
Let $\phi$ be an arithmetically normalized Hecke–Maa[ß]{} cusp form on $X$ as in Theorem \[theorem\_2\]. Then the Fourier coefficients $$\lambda_{\phi}(m_1,m_2)=\lambda_{\phi}(m_1,|m_2|)$$ in are actual eigenvalues of various Hecke operators (cf. [@G Thm. 6.4.11]). Moreover, summing over the representatives $\gamma=\bigl(\begin{smallmatrix}a & b \\ c & d \end{smallmatrix}\bigr)$ of ${\mathrm{U}}_2({\mathbb{Z}}){\backslash}{\mathrm{SL}}_2({\mathbb{Z}})$ in is the same as summing over the coprime pairs $(c,d)\in{\mathbb{Z}}^2$. Computing, as in [@G (6.5.4)], the ${\mathrm{GL}}_2$ Iwasawa decomposition in the upper-left $2\times 2$ block of $$\begin{pmatrix}
\gamma & \\ & 1
\end{pmatrix}
z {\mathrm{O}}_3({\mathbb{R}}) = \begin{pmatrix}
a & b & \\ c & d & \\ & & 1
\end{pmatrix}
\begin{pmatrix}
y_1y_2 & x_2y_1 & x_3 \\ & y_1 & x_1 \\ & & 1 \end{pmatrix} {\mathrm{O}}_3({\mathbb{R}}) = \begin{pmatrix}
\frac{y_1y_2}{|cz_2+d|} & * & * \\ & y_1|cz_2+d| & * \\ & & 1
\end{pmatrix} {\mathrm{O}}_3({\mathbb{R}}),$$ where $z_2=x_2+iy_2\in{\mathcal{H}}_2$, we see from – that $$\phi(z)\ll \sum_{m_1=1}^{\infty}\sum_{m_2=1}^{\infty} \frac{|\lambda_{\phi}(m_1,m_2)|}{m_1m_2}
\sum_{\substack{(c,d)\in{\mathbb{Z}}^2\\\gcd(c,d)=1}}\left|\tilde{W}_{\nu_1,\nu_2}(m_1y_1|cz_2+d|,m_2y_2|cz_2+d|^{-2})\right|.$$ Taking any ${\varepsilon}>0$, and applying the Cauchy–Schwarz inequality, we obtain $$\label{eq:after_CS}
\phi(z)\ll \left(\sum_{m_1=1}^{\infty}\sum_{m_2=1}^{\infty} \frac{|\lambda_{\phi}(m_1,m_2)|^2}{m_1^{2+2{\varepsilon}}m_2^{1+{\varepsilon}}}\right)^{1/2}
\left(\sum_{m_1=1}^{\infty}\sum_{m_2=1}^{\infty} \frac{m_1^{2{\varepsilon}}}{m_2^{1-{\varepsilon}}} F(m_1y_1,m_2y_2)^2 \right)^{1/2},$$ where $$\label{eq:F-def}
F(s_1,s_2):=\sum_{\substack{(c,d)\in{\mathbb{Z}}^2\\\gcd(c,d)=1}}\left|\tilde{W}_{\nu_1,\nu_2}(s_1|cz_2+d|,s_2|cz_2+d|^{-2})\right|.$$
It remains to estimate the two factors on the right hand side of . The first factor (the arithmetic part) will be estimated in Subsection \[sub31\]. The second factor (the analytic part) will be estimated in Subsection \[sub33\].
Estimating the arithmetic part {#sub31}
------------------------------
By [@G Sections 6.3–6.4], the dual form $$\label{dualform}
\tilde\phi(z):=\phi\left((z^{-1})^t\right),\qquad z\in{\mathcal{H}}_3,$$ is an arithmetically normalized Hecke–Maa[ß]{} cusp form on $X={\mathrm{GL}}_3({\mathbb{Z}}){\backslash}{\mathcal{H}}_3$ with spectral parameters $(\nu_0,\nu_2,\nu_1)$ and Hecke eigenvalues $$\lambda_{\tilde\phi}(m_1,m_2)=\lambda_\phi(m_2,m_1)={\overline{\lambda_\phi(m_1,m_2)}}.$$ Introducing the Rankin–Selberg $L$-function (cf. [@G Section 7.4]) $$L\bigl(s,\phi\times\tilde{\phi}\bigr) = \zeta(3s) \sum_{m_1=1}^{\infty} \sum_{m_2=1}^{\infty}
\frac{|\lambda_{\phi}(m_1,m_2)|^2}{m_1^{2s}m_2^s},\qquad\Re s>1,$$ the first double sum on the right hand side of can be estimated by the result of Brumley [@Br Cor. 2] or Li [@L Thm. 2]: $$\label{eq:arithmetic_final}
\sum_{m_1=1}^{\infty}\sum_{m_2=1}^{\infty} \frac{|\lambda_{\phi}(m_1,m_2)|^2}{m_1^{2+2{\varepsilon}}m_2^{1+{\varepsilon}}} = \frac{L\bigl(1+{\varepsilon},\phi\times\tilde{\phi}\bigr)}{\zeta(3+3{\varepsilon})} \ll_{{\varepsilon}} T_0^{{\varepsilon}}.$$
Estimating the analytic part {#sub33}
----------------------------
We decompose into dyadic subsums: for any $k,l\in{\mathbb{N}}$, we define $$D(k,l):=\left\{(c,d)\in{\mathbb{Z}}^2:\ \gcd(c,d)=1,\ 2^k<1+\frac{s_1|cz_2+d|}{T_0}{\leqslant}2^{k+1},\ 2^l<1+\frac{s_2|cz_2+d|^{-2}}{T_0}{\leqslant}2^{l+1}\right\}$$ and $$F_{k,l}(s_1,s_2):=\sum_{(c,d)\in D(k,l)} \left|\tilde{W}_{\nu_1,\nu_2}(s_1|cz_2+d|,s_2|cz_2+d|^{-2})\right|.$$ Clearly, $$F(s_1,s_2)=\sum_{k=0}^{\infty}\sum_{l=0}^{\infty} F_{k,l}(s_1,s_2),$$ which implies, via the Cauchy–Schwarz inequality, that $$\label{eq:F_after_CS_1}
F(s_1,s_2)^2\ll_{{\varepsilon}} \sum_{k=0}^{\infty}\sum_{l=0}^{\infty} 2^{{\varepsilon}(k+l)}F_{k,l}(s_1,s_2)^2$$ holds for any ${\varepsilon}>0$.
We estimate $F_{k,l}(s_1,s_2)$ on the right hand side of via Lemma \[lemma:whittaker\_estimate\] as $$F_{k,l}(s_1,s_2)\ll_{{\varepsilon},A} 2^{-A(k+l)} T_0^{{\varepsilon}} \sqrt{s_1s_2} \sum_{(c,d)\in D(k,l)} |cz_2+d|^{-1/2}.$$ From the two conditions on the range of $s_1,s_2$ in the definition of $D(k,l)$, it is immediate that $F_{k,l}(s_1,s_2)$ vanishes unless $s_1^2s_2{\leqslant}2^{2k+l+3}T_0^3$, which means that $$\label{fkl1}
F_{k,l}(s_1,s_2)\ll_{{\varepsilon},A} 2^{-A(k+l)} T_0^{{\varepsilon}} \sqrt{s_1s_2} \cdot 1_{s_1^2s_2{\leqslant}2^{2k+l+3}T_0^3} \sum_{(c,d)\in D(k,l)} |cz_2+d|^{-1/2}.$$ For the $(c,d)$-summation, we only use the estimate $|cz_2+d|{\leqslant}2^{k+1} T_0 s_1^{-1}$, which amounts to saying that we have to sum over the lattice ${\mathbb{Z}}z_2+{\mathbb{Z}}\subset {\mathbb{C}}$ in a given disk around the origin. The lattice has first successive minimum at least $\sqrt{3}/2$ and covolume $\Im z_2=y_2$ also at least $\sqrt{3}/2$. Applying again a dyadic decomposition and considering the pairs $(c,d)$ satisfying $2^m{\leqslant}|cz_2+d|<2^{m+1}$, we obtain by [@BHM Lemma 1] $$\sum_{(c,d)\in D(k,l)} |cz_2+d|^{-1/2} \ll \sum_{\substack{m\in{\mathbb{Z}}\\1/2{\leqslant}2^m{\leqslant}2^{k+1}T_0 s_1^{-1}}} 2^{-m/2}(2^m+2^{2m}y_2^{-1})
\ll 2^{k/2}T_0^{1/2}s_1^{-1/2}+2^{3k/2}T_0^{3/2}s_1^{-3/2}y_2^{-1}.$$ Plugging this into , we conclude $$F_{k,l}(s_1,s_2)\ll_{{\varepsilon},A} T_0^{{\varepsilon}} 1_{s_1^2s_2{\leqslant}2^{2k+l+3}T_0^3} \left( 2^{\left(1/2-A\right)k - Al} T_0^{1/2} s_2^{1/2} + 2^{\left(3/2-A\right)k - Al} T_0^{3/2}s_1^{-1}s_2^{1/2}y_2^{-1}\right).$$
Using the previous bound in , then substituting $s_1=m_1 y_1$ and $s_2=m_2 y_2$, and finally summing with respect to $m_1$ and $m_2$ as in , we obtain $$\label{eq:endgame_main}
\begin{split}
&\sum_{m_1=1}^{\infty}\sum_{m_2=1}^{\infty} \frac{m_1^{2{\varepsilon}}}{m_2^{1-{\varepsilon}}} F(m_1y_1,m_2y_2)^2 \ll_{{\varepsilon},A} T_0^{2{\varepsilon}} \sum_{k=0}^{\infty} \sum_{l=0}^{\infty} \sum_{m_1=1}^{\infty} \sum_{m_2=1}^{\infty} 1_{m_1^2m_2{\leqslant}2^{2k+l+3}T_0^3y_1^{-2}y_2^{-1}} \\ & \times \frac{m_1^{2{\varepsilon}}}{m_2^{1-{\varepsilon}}}\left(2^{({\varepsilon}+1-2A)k + ({\varepsilon}-2A)l}T_0 m_2y_2 + 2^{({\varepsilon}+3-2A)k + ({\varepsilon}-2A)l} T_0^3 m_1^{-2}m_2y_1^{-2}y_2^{-1}\right).
\end{split}$$ The first term in the bottom line of contributes (after the summation in the first line there) $$\ll_{{\varepsilon},A} T_0^{1+2{\varepsilon}}y_2\sum_{k=0}^{\infty}\sum_{l=0}^{\infty}2^{({\varepsilon}+1-2A)k + ({\varepsilon}-2A)l} \sum_{m_1=1}^{\infty} m_1^{2{\varepsilon}} \sum_{1{\leqslant}m_2{\leqslant}2^{2k+l+3}T_0^3 m_1^{-2}y_1^{-2}y_2^{-1}} m_2^{{\varepsilon}}.$$ Here, the $m_2$-sum is at most $(2^{2k+l+3}T_0^3 m_1^{-2}y_1^{-2}y_2^{-1})^{1+{\varepsilon}}$. Then clearly the resulting $m_1$-sum is convergent, which altogether means that this first term gives rise to $$\label{eq:endgame_term1}
\ll_{{\varepsilon},A} T_0^{4+5{\varepsilon}}y_1^{-2} \sum_{k=0}^{\infty} \sum_{l=0}^{\infty} 2^{(3{\varepsilon}+3-2A)k + (2{\varepsilon}+1-2A)l}.$$ Similarly, the second term in the bottom line of contributes (after the summation in the first line there) $$\ll_{{\varepsilon},A} T_0^{3+2{\varepsilon}} y_1^{-2}y_2^{-1} \sum_{k=0}^{\infty}\sum_{l=0}^{\infty} 2^{({\varepsilon}+3-2A)k + ({\varepsilon}-2A)l} \sum_{m_1=1}^{\infty} m_1^{2{\varepsilon}-2} \sum_{1{\leqslant}m_2{\leqslant}2^{2k+l+3}T_0^3m_1^{-2}y_1^{-2}y_2^{-1}}m_2^{{\varepsilon}}.$$ Here, the $m_2$-sum is again at most $(2^{2k+l+3}T_0^3 m_1^{-2}y_1^{-2}y_2^{-1})^{1+{\varepsilon}}$, and the resulting $m_1$-sum is convergent as above. In the end, the second term gives rise to $$\label{eq:endgame_term2}
\ll_{{\varepsilon},A} T_0^{6+5{\varepsilon}} y_1^{-4}y_2^{-2} \sum_{k=0}^{\infty} \sum_{l=0}^{\infty} 2^{(3{\varepsilon}+5-2A)k + (2{\varepsilon}+1-2A)l}.$$
We see that in both and , the double sums over $k$ and $l$ are convergent upon choosing $A:=3$, say. Returning to , this means that $$\label{eq:analytic_final}
\sum_{m_1=1}^{\infty}\sum_{m_2=1}^{\infty} \frac{m_1^{2{\varepsilon}}}{m_2^{1-{\varepsilon}}} F(m_1y_1,m_2y_2)^2 \ll_{{\varepsilon}} \frac{T_0^{4+5{\varepsilon}}}{y_1^2} + \frac{T_0^{6+5{\varepsilon}}}{y_1^4y_2^2}.$$
Conclusion
----------
The inequalities , and altogether give $$\phi(z)\ll_{{\varepsilon}} \frac{T_0^{2+3{\varepsilon}}}{y_1} + \frac{T_0^{3+3{\varepsilon}}}{y_1^2y_2} \ll_{\varepsilon}\frac{\lambda_{\phi}^{1+2{\varepsilon}}}{y_1} + \frac{\lambda_{\phi}^{3/2+2{\varepsilon}}}{y_1^2y_2},$$ and we write this in the more symmetric form $$\label{phibound}
\phi(z)\ll_{\varepsilon}y_2\left(\frac{\lambda_{\phi}^{1+{\varepsilon}}}{y_1y_2} + \frac{\lambda_{\phi}^{3/2+{\varepsilon}}}{(y_1y_2)^2}\right).$$
Finally, we utilize the dual form introduced in to show that remains valid when we interchange $y_1$ and $y_2$ on the right hand side. Indeed, we can express $\phi(z)$ as $$\phi(z)=\tilde\phi\left((z^{-1})^t\right)=\tilde\phi\left(hw(z^{-1})^t w\right),
\qquad h:=\begin{pmatrix}y_1y_2&&\\&y_1y_2&\\&&y_1y_2\end{pmatrix},
\qquad w:=\begin{pmatrix}&&1\\&1&\\1&&\end{pmatrix},$$ where we have the Iwasawa decomposition (see [@G (6.3.2)] for a similar computation) $$hw(z^{-1})^t w=\begin{pmatrix}1 & -x_1 & x_1x_2-x_3 \\ & 1 & -x_2 \\ & & 1\end{pmatrix}
\begin{pmatrix}y_1y_2 & & \\ & y_2 & \\ & & 1\end{pmatrix}.$$ Applying and noting that $\lambda_{\tilde\phi}=\lambda_\phi$ by , we infer $$\phi(z)=\tilde\phi\left(hw(z^{-1})^t w\right)\ll_{\varepsilon}y_1\left(\frac{\lambda_{\phi}^{1+{\varepsilon}}}{y_1y_2} + \frac{\lambda_{\phi}^{3/2+{\varepsilon}}}{(y_1y_2)^2}\right).$$ Together with , this proves . The proof of Theorem \[theorem\_2\] is complete.
Proof of Theorem \[theorem\_3\]
===============================
Let $\phi$ be an $L^2$-normalized Maa[ß]{} cusp form on $X$ as in Theorem \[theorem\_3\]. Note that $\phi$ is allowed to be non-tempered at the archimedean place, hence its Langlands parameters $\mu:=(\mu_1,\mu_2,\mu_3)$ introduced in are not necessarily purely imaginary, but they satisfy by unitarity and the standard Jacquet–Shalika bounds [@JS] $$\max\bigl(|\Re\mu_1|,|\Re\mu_2|,|\Re\mu_3|\bigr){\leqslant}\frac{1}{2}\qquad\text{and}\qquad
\{\mu_1,\mu_2,\mu_3\}=\{-{\overline{\mu_1}},-{\overline{\mu_2}},-{\overline{\mu_3}}\}.$$ Note also that we can assume without loss of generality, because we are already assuming $y_1,y_2{\geqslant}\sqrt{3}/2$, and replacing $z$ by $\gamma z$ for an arbitrary *unipotent* integral matrix $\gamma\in{\mathrm{U}}_3({\mathbb{Z}})$ leaves both sides of unchanged.
We shall establish with the help of Selberg’s pre-trace formula. As a preparation, we denote by $$\mathfrak{a}:=\{\operatorname{diag}(\alpha_1,\alpha_2,\alpha_3)\in{\mathrm{M}}_3({\mathbb{R}}):\ \alpha_1+\alpha_2+\alpha_3=0\}$$ the Lie algebra of the diagonal torus of ${\mathrm{PGL}}_3({\mathbb{R}})$, and by $C:{\mathrm{PGL}}_3({\mathbb{R}})\to\mathfrak{a}/S_3$ the Cartan projection induced from the Cartan decomposition for ${\mathrm{GL}}_3({\mathbb{R}})$: $$g=h k_1\exp (C(g)) k_2\qquad\text{with}\qquad h\in{\mathrm{Z}}({\mathbb{R}}),\ k_1,k_2\in{\mathrm{O}}_3({\mathbb{R}}).$$ We identify the complexified dual $\mathfrak{a}^*_{{\mathbb{C}}}$ with the set of triples (cf. ) $$(\kappa_1,\kappa_2,\kappa_3)\in{\mathbb{C}}^3\qquad\text{satisfying}\qquad\kappa_1+\kappa_2+\kappa_3=0,$$ namely such a triple acts on $\mathfrak{a}_{{\mathbb{C}}}:=\mathfrak{a}\otimes_{\mathbb{R}}{\mathbb{C}}$ by the ${\mathbb{C}}$-linear map $$\operatorname{diag}(\alpha_1,\alpha_2,\alpha_3)\mapsto \kappa_1\alpha_1+\kappa_2\alpha_2+\kappa_3\alpha_3.$$
The pre-trace formula approach
------------------------------
Following [@BM Section 2] and [@BP Sections 2 & 6], we can construct a smooth, bi-${\mathrm{O}}_3({\mathbb{R}})$-invariant function $f_{\mu}:{\mathrm{PGL}}_3({\mathbb{R}})\to{\mathbb{C}}$ supported in a fixed compact subset $K$ (which is independent of $\mu\in\mathfrak{a}^*_{{\mathbb{C}}}$) with the following properties. On the one hand, the function obeys (cf. [@BM (2.3)–(2.4)]) $$\label{eq:inverse_spherical_transform_BP_bound}
f_{\mu}(g)\ll \lambda_{\phi}^{3/2}\left(1+\lambda_{\phi}^{1/2}\|C(g)\|\right)^{-1/2},\qquad g\in{\mathrm{PGL}}_3({\mathbb{R}}),$$ where $\|\cdot\|$ stands for a fixed $S_3$-invariant norm on $\mathfrak{a}\cong{\mathbb{R}}^2$. On the other hand, its spherical transform $\tilde{f}_{\mu}:\mathfrak{a}^*_{{\mathbb{C}}}/S_3 \to{\mathbb{C}}$, defined as in [@He (17) of Section II.3] or [@BP (2.3)], satisfies $$\tilde{f}_{\mu}(\mu){\geqslant}1,\qquad \tilde{f}_{\mu}(\kappa){\geqslant}0$$ for all Langlands parameters $\kappa\in\mathfrak{a}^*_{{\mathbb{C}}}$ occurring in $L^2(X)$, including possibly non-tempered parameters. Then, using positivity in Selberg’s pre-trace formula (see [@BP (6.1)]), or more directly by a Mercer-type pre-trace inequality (cf. [@BHMM (3.15)]), we obtain $$\label{eq:pre-trace_inequality}
|\phi(z)|^2 {\leqslant}\sum_{\gamma\in{\mathrm{PGL}}_3({\mathbb{Z}})} f_\mu(z^{-1}\gamma z).$$
We can now derive quickly a somewhat weaker version of . By , we clearly have $$\label{eq:inverse_spherical_transform_trivial_bound}
f_{\mu}(g)\ll \lambda_{\phi}^{3/2},\qquad g\in{\mathrm{PGL}}_3({\mathbb{R}}),$$ whence by , $$\label{eq:pre-trace_inequality_with_number_of_matrices}
|\phi(z)|^2 \ll \lambda_{\phi}^{3/2} \sum_{\substack{\gamma\in{\mathrm{PGL}}_3({\mathbb{Z}})\\\|C(z^{-1}\gamma z)\|\ll 1}} 1.$$ Here, the uniform bound $\|C(z^{-1}\gamma z)\|\ll 1$ depends on the fixed compact subset $K$ in which $f_\mu$ is supported. Labeling the entries of $\gamma$ as $$\label{eq:labeling_gamma}
\gamma = \begin{pmatrix} a & b & c\\ d & e & f\\ g & h & i\end{pmatrix},$$ a straightforward calculation gives $$\label{eq:z_-1_gamma_z}
z^{-1}\gamma z= \begin{pmatrix} a - d x_2 + gx_1x_2 - gx_3 & \frac{b + (a - e + h x_1- dx_2 + gx_1x_2)x_2 - (h + gx_2)x_3}{y_2} & * \\ (d - g x_1) y_2 & e + d x_2 - (h + gx_2)x_1 & \frac{ f + e x_1 + dx_3 - (i + hx_1 + gx_3)x_1}{y_1}\\gy_1y_2 & (h + gx_2) y_1 & i + hx_1 + gx_3\end{pmatrix}$$ with $$\label{eq:explicit_*}
*=\frac{c + bx_1 - f x_2 + (-e+i+ hx_1)x_1x_2 + ax_3 - (i + hx_1 + dx_2 - gx_1x_2)x_3 - gx_3^2}{y_1y_2}.$$ Now $\|C(z^{-1}\gamma z)\|\ll 1$ combined with implies, in this order, $$\label{eq:entrybounds}
g\ll y_1^{-1}y_2^{-1},\qquad h\ll y_1^{-1},\qquad d\ll y_2^{-1},\qquad a,e,i\ll 1,\qquad f\ll y_1,\qquad b\ll y_2,\qquad c\ll y_1y_2.$$ Therefore, in , the number of relevant matrices $\gamma$ is $\ll y_1^2y_2^2$, and we obtain readily $$\label{eq:conclusive_pre-trace_bound}
\phi(z) \ll \lambda_{\phi}^{3/4} y_1 y_2.$$
An interlude on the Cartan projection
-------------------------------------
To improve upon the bound , we shall make full use of instead of applying its weaker version . With this aim in mind, we prove the following estimate on the Cartan projection, which works in general for ${\mathrm{PGL}}_n({\mathbb{R}})$ or ${\mathrm{SL}}_n({\mathbb{R}})$.
\[lemma:lower\_bound\_on\_cartan\_projection\] Let $g\in{\mathrm{SL}}_n({\mathbb{R}})$ be an upper-triangular matrix with positive diagonal entries, and let ${1_n}\in{\mathrm{SL}}_n({\mathbb{R}})$ denote the $n\times n$ identity matrix. If $g$ lies in a fixed compact subset of ${\mathrm{SL}}_n({\mathbb{R}})$, then $$\|C(g)\|\asymp\|g^tg-{1_n}\|\asymp\|g-{1_n}\|.$$
The first relation is standard and follows from the Cartan decomposition $g=k_1\exp(C(g))k_2$ with $k_1,k_2\in{\mathrm{SO}}_n({\mathbb{R}})$. So let us focus on the second relation. As $g$ lies in a fixed compact subset, we have that $\|g^tg-{1_n}\|\asymp\|g^t-g^{-1}\|$, so it suffices to prove that $\|g^t-g^{-1}\|\asymp\|g-{1_n}\|$. Here, $g^t$ is lower-triangular with some positive diagonal $a$, and $g^{-1}$ is upper-triangular with diagonal $a^{-1}$. Therefore, $$\begin{aligned}
\|g^t-g^{-1}\|&\asymp\|g^t-a\|+\|a-a^{-1}\|+\|a^{-1}-g^{-1}\|\\
&\asymp\|g^t-a\|+\|a-{1_n}\|+\|{1_n}-a^{-1}\|+\|a^{-1}-g^{-1}\|\\
&\asymp\|g^t-{1_n}\|+\|{1_n}-g^{-1}\|\\
&\asymp\|g-{1_n}\|,\end{aligned}$$ where in the last step we used again that $g$ lies in a fixed compact subset. We are done.
We are now ready to improve . In the light of , it is convenient to rewrite as $$\label{eq:pre-trace_inequality_cases}
|\phi(z)|^2 {\leqslant}\sum_{j=1}^4\sum_{\gamma\in M_j} f_\mu(z^{-1}\gamma z),$$ where, in the notation of , $$\begin{split}
M_1&:=\{\gamma\in{\mathrm{PGL}}_3({\mathbb{Z}}):\ g=0,\ h=0,\ d=0\},\\
M_2&:=\{\gamma\in{\mathrm{PGL}}_3({\mathbb{Z}}):\ g=0,\ h\neq 0,\ d=0\},\\
M_3&:=\{\gamma\in{\mathrm{PGL}}_3({\mathbb{Z}}):\ g=0,\ h=0,\ d\neq 0\},\\
M_4&:=\{\gamma\in{\mathrm{PGL}}_3({\mathbb{Z}}):\ g\neq 0\ \ \text{or}\ \ hd\neq 0\}.
\end{split}$$
The contribution of $M_1$
-------------------------
In this case, $g=h=d=0$ forces $a,e,i\in\{\pm 1\}$. Then – simplify to $$\label{eq:z_-1_gamma_z special 1}
kz^{-1} \gamma z = \begin{pmatrix}
1 & \pm\frac{b + (\pm 1\pm 1)x_2}{y_2} & \pm\frac{c + b x_1 - f x_2 +(\pm 1\pm 1) x_1x_2 + (\pm 1\pm 1)x_3}{y_1y_2} \\
& 1 & \pm\frac{f + (\pm 1\pm 1)x_1}{y_1} \\ & & 1
\end{pmatrix},$$ where $k$ abbreviates $\operatorname{diag}(a,e,c)\in{\mathrm{O}}_3({\mathbb{R}})$, and all combinations of the $\pm$ signs are allowed.
For a dyadic parameter $\lambda_{\phi}^{-1/2}\ll K\ll 1$, let us examine the contribution to of the matrices $\gamma\in M_1$ that satisfy $\|C(z^{-1}\gamma z)\|\asymp K$. Applying , we obtain $$\label{eq:M1bound}
\sum_{\substack{\gamma\in M_1 \\ K {\leqslant}\|C(z^{-1}\gamma z)\|< 2K}} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{5/4} K^{-1/2} \sum_{\substack{\gamma\in M_1 \\ K {\leqslant}\|C(z^{-1}\gamma z)\|< 2K}} 1.$$ The matrix count on the right hand side is given by the number of choices for the integral triple $(f,b,c)\in{\mathbb{Z}}^3$ in . Applying Lemma \[lemma:lower\_bound\_on\_cartan\_projection\] in the form $\|kz^{-1} \gamma z-1_3\|\ll K$, we infer $$\sum_{\substack{\gamma\in M_1 \\ K {\leqslant}\|C(z^{-1}\gamma z)\|< 2K}} 1\ll\#(f,b,c)\ll (1+y_1K)(1+y_2K)(1+y_1y_2K).$$ Keeping in mind that $y_1,y_2\gg 1$, we can now estimate the right hand side of as $$\lambda_{\phi}^{5/4}(K^{-1/2}+y_1y_2K^{1/2}+y_1y_2^2 K^{3/2}+y_1^2y_2K^{3/2}+y_1^2y_2^2K^{5/2}).$$ Summing up over dyadic parameters satisfying $\lambda_{\phi}^{-1/2}\ll K\ll 1$, and using $y_1,y_2\gg 1$ again, we get $$\sum_{\substack{\gamma\in M_1 \\ \lambda_{\phi}^{-1/2}{\leqslant}\|C(z^{-1}\gamma z)\|\ll 1}} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{3/2} + \lambda_{\phi}^{5/4}y_1^2y_2^2.$$
We estimate the remaining contribution of the matrices $\gamma\in M_1$ with $\|C(z^{-1}\gamma z)\|<\lambda_{\phi}^{-1/2}$ similarly, this time applying . In the end, we obtain $$\label{eq:M1contribution}
\sum_{\gamma\in M_1} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{3/2} + \lambda_{\phi}^{5/4}y_1^2y_2^2.$$
The contribution of $M_2$ and $M_3$
-----------------------------------
We work out the relevant bound only for $M_2$, as the argument for $M_3$ is very similar. As above, $g=d=0$ forces $a=\pm 1$. In this case, since $h\neq 0$, we see from that $y_1\ll 1$ and there are only $O(1)$ choices for the bottom right $2\times 2$ block $\bigl(\begin{smallmatrix} e & f \\ h & i\end{smallmatrix}\bigr)\in{\mathrm{GL}}_2({\mathbb{Z}})$ of $\gamma$. For each such block, we multiply $z^{-1}\gamma z$ by an orthogonal matrix $k\in\bigl(\begin{smallmatrix}a & \\ & {\mathrm{O}}_2({\mathbb{R}}) \end{smallmatrix}\bigr)$ on the left to arrive at an upper-triangular matrix with positive diagonal entries[^5]. This multiplies the top row by $a=\pm 1$, i.e. – become $$kz^{-1}\gamma z = \begin{pmatrix} 1 & \pm\frac{b + (\pm 1 - e + h x_1)x_2 - hx_3}{y_2} & \pm\frac{c + bx_1 - f x_2 + (-e+i+ hx_1)x_1x_2 \pm x_3 - (i + hx_1)x_3}{y_1y_2} \\ & * & * \\ & & *
\end{pmatrix}.$$ For a dyadic parameter $\lambda_{\phi}^{-1/2}\ll K\ll 1$, we estimate similarly as in , $$\sum_{\substack{\gamma\in M_2 \\ K {\leqslant}\|C(z^{-1}\gamma z)\|< 2K}} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{5/4} K^{-1/2} \sum_{\substack{\gamma\in M_2 \\ K {\leqslant}\|C(z^{-1}\gamma z)\|< 2K}} 1.$$ Applying Lemma \[lemma:lower\_bound\_on\_cartan\_projection\] again, we can bound the matrix count on the right hand side as $$\sum_{\substack{\gamma\in M_2 \\ K {\leqslant}\|C(z^{-1}\gamma z)\|< 2K}} 1\ll\#(b,c)\ll (1+y_2K)(1+y_1y_2K)\ll 1+y_2^2 K^2,$$ and a similar calculation as before leads to the bound $$\label{eq:M2contribution}
\sum_{\gamma\in M_2} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{3/2} + \lambda_{\phi}^{5/4}y_2^2.$$ The analogous argument applied to $M_3$ yields $$\label{eq:M3contribution}
\sum_{\gamma\in M_3} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{3/2} + \lambda_{\phi}^{5/4}y_1^2.$$
The contribution of $M_4$
-------------------------
We see from the first three entries of that $M_4$ can only contribute when $y_1\ll 1$ and $y_2\ll 1$, and then from and we obtain readily that $$\label{eq:M4contribution}
\sum_{\gamma\in M_4} f_\mu(z^{-1}\gamma z) \ll \lambda_{\phi}^{3/2}.$$
Conclusion
----------
By –, we arrive at , and the proof of Theorem \[theorem\_3\] is complete.
Proof of Theorem \[theorem\_1\]
===============================
Let $\phi$ be an arithmetically normalized Hecke–Maa[ß]{} cusp form on $X$ as in Theorem \[theorem\_1\]. Then , and also with an extra factor of $\lambda_\phi^{\varepsilon}$, hold in a suitable fundamental domain for $X={\mathrm{GL}}_3({\mathbb{Z}}){\backslash}{\mathcal{H}}_3$ in the light of , Theorems \[theorem\_2\]–\[theorem\_3\], and the remarks made after Theorem \[theorem\_1\]. We balance between these two bounds as follows.
If $y_1y_2>\lambda_{\phi}^{7/20}$, then we apply to obtain $$\phi(z) \ll_{\varepsilon}\min(y_1,y_2)\left(\frac{\lambda_{\phi}^{1+{\varepsilon}}}{y_1y_2} + \frac{\lambda_{\phi}^{3/2+{\varepsilon}}}{(y_1y_2)^2}\right) {\leqslant}\frac{\lambda_{\phi}^{1+{\varepsilon}}}{(y_1y_2)^{1/2}} + \frac{\lambda_{\phi}^{3/2+{\varepsilon}}}{(y_1y_2)^{3/2}} \ll \lambda_{\phi}^{39/40+{\varepsilon}}.$$ Else, if $y_1y_2{\leqslant}\lambda_{\phi}^{7/20}$, we apply with an extra factor of $\lambda_\phi^{\varepsilon}$ to obtain $$\phi(z)\ll_{\varepsilon}\lambda_{\phi}^{3/4+{\varepsilon}} + \lambda_{\phi}^{5/8+{\varepsilon}}y_1y_2 \ll \lambda_{\phi}^{39/40+{\varepsilon}}.$$ The proof of Theorem \[theorem\_1\] is complete.
[BHMM]{}
C. B. Balogh, *Asymptotic expansions of the modified Bessel function of the third kind of imaginary order*, SIAM J. Appl. Math. **15** (1967), 1315–1323.
V. Blomer, *Applications of the Kuznetsov formula on $GL(3)$*, Invent. Math. **194** (2013), 673–729.
V. Blomer, J. Buttcane, *On the subconvexity problem for L-functions on ${\mathrm{GL}}(3)$*, submitted, [arXiv:1504.02667]{}
V. Blomer, G. Harcos, D. Milićević, *Bounds for eigenforms on arithmetic hyperbolic 3-manifolds*, Duke Math. J. **165** (2016), 625–659.
V. Blomer, G. Harcos, P. Maga, D. Milićević, *The sup-norm problem for ${\mathrm{GL}}(2)$ over number fields*, submitted, [arXiv:1605.09360]{}
V. Blomer, R. Holowinsky, *Bounding sup-norms of cusp forms of large level*, Invent. Math. **179** (2010), 645–681.
V. Blomer, P. Maga, *Subconvexity for sup-norms of cusp forms on ${\mathrm{PGL}}(n)$*, Selecta Math. **22** (2016), 1269–1287.
V. Blomer, A. Pohl, *The sup-norm problem on the Siegel modular space of rank two*, Amer. J. Math. **136** (2016), 999–1027.
F. Brumley, *Second order average estimates on local data of cusp forms*, Arch. Math. **87** (2006), 19–32.
F. Brumley, *Effective multiplicity one on ${\mathrm{GL}}_N$ and narrow zero-free regions for Rankin-Selberg L-functions*, Amer. J. Math. **128** (2006), 1455–1474.
F. Brumley, N. Templier, *Large values of cusp forms on ${\mathrm{GL}}_n$*, [arXiv:1411.4317]{}
D. Bump, *Automorphic Forms on ${\mathrm{GL}}(3,{\mathbb{R}})$*, Lecture Notes in Mathematics, Vol. 1083, Springer-Verlag, Berlin, 1984.
D. Goldfeld, *Automorphic forms and L-functions for the group $GL(n,{\mathbb{R}})$*, Cambridge Studies in Advanced Mathematics, Vol. 99, Cambridge University Press, Cambridge, 2006.
I. S. Gradshteyn, I. M. Ryzhik, *Table of integrals, series, and products*, 8th edition, Elsevier/Academic Press, Amsterdam, 2015.
G. Harcos, P. Michel, *The subconvexity problem for Rankin–Selberg $L$-functions and equidistribution of Heegner points. II*, Invent. Math. **163** (2006), 581–655.
S. Helgason, *Geometric analysis on symmetric spaces*, 2nd edition, Mathematical Surveys and Monographs, Vol. 39, American Mathematical Society, Providence, RI, 2008.
R. Holowinsky, G. Ricotta, E. Royer, *On the sup-norm of $SL_3$ Hecke-Maass cusp form*, [arXiv:1404.3622]{}
L. Hörmander, *The spectral function of an elliptic operator*, Acta Math. **121** (1968), 193–218.
H. Iwaniec, P. Sarnak, *$L^\infty$ norms of eigenfunctions of arithmetic surfaces*, Ann. of Math. **141** (1995), 301–320.
H. Jacquet, J. A. Shalika, *On Euler products and the classification of automorphic representations. I*, Amer. J. Math. **103** (1981), 499–558.
E. Lapid, *On the Harish-Chandra Schwartz space of $G(F)\backslash G(\mathbb{A})$*, with an appendix by F. Brumley, in: Automorphic representations and $L$-functions, 335–377, Tata Inst. Fundam. Res. Stud. Math., Vol. 22, Mumbai, 2013.
X. Li, *Upper bounds on L-functions at the edge of the critical strip*, Int. Math. Res. Not. **2010**, no. 4, 727–755.
D. Ramakrishnan, *An exercise concerning the selfdual cusp forms on $GL(3)$*, Indian J. Pure Appl. Math. **45** (2014), 777–785.
D. Ramakrishnan, S. Wang, *On the exceptional zeros of Rankin–Selberg $L$-functions*, Compositio Math. **135** (2003), 211–244.
A. Saha, *Large values of newforms on ${\mathrm{GL}}(2)$ with highly ramified central character*, Int. Math. Res. Not. **2016**, no. 13, 4103–4131.
A. Saha, *Hybrid sup-norm bounds for Maass newforms of powerful level*, Algebra and Number Theory, to appear, [arXiv:1509.07489]{}
P. Sarnak, *Letter to Cathleen Morawetz*, August 2004, available at [http://publications.ias.edu/sarnak]{}
E. Stade, *On explicit integral formulas for $GL(n,{\mathbb{R}})$-Whittaker functions*, Duke Math. J. **60** (1990), 313–362.
E. Stade, *Mellin transforms of $GL(n,{\mathbb{R}})$ Whittaker functions*, Amer. J. Math. **123** (2001), 121–161.
I. Vinogradov, L. Takhtadzhyan, *Theory of Eisenstein Series for the group ${\mathrm{SL}}(3,{\mathbb{R}})$ and its application to a binary problem*, J. Soviet Math. **18** (1982), 293–324.
[^1]: First author partially supported by the DFG-SNF lead agency program grant BL 915/2-1. Second and third author supported by NKFIH (National Research, Development and Innovation Office) grants NK 104183, ERCHU15 118946, K 119528. Second author also supported by ERC grant AdG-321104, and third author also supported by the Postdoctoral Fellowship of the Hungarian Academy of Sciences.
[^2]: By saying this we do not mean to imply that $1/124$ is a special threshold beyond which the bound fails.
[^3]: Unfortunately, the relevant literature has many inconsistencies that are hard to track. For example, in Goldfeld’s book [@G], Definition 6.5.2 is inconsistent with Definition 9.4.3, and hence Theorem 6.5.15 contradicts Theorem 10.8.6 and the subsequent discussion. In fact $\nu_1$ and $\nu_2$ should be flipped in [@G Thm. 6.5.15], in harmony with [@Bu (8.3)]. Compare also the functions $I_\nu$ on [@Bu p. 19] and [@G p. 154].
[^4]: Accordingly, in [@G (6.1.3)], the constant $4$ should be $8$, while $\nu_1$ and $\nu_2$ should be flipped.
[^5]: In the case of $M_3$, we would multiply $z^{-1}\gamma z$ by an orthogonal matrix $k\in\bigl(\begin{smallmatrix}{\mathrm{O}}_2({\mathbb{R}})&\\&i\end{smallmatrix}\bigr)$ on the right, to achieve the same.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A $k$-noncrossing RNA pseudoknot structure is a graph over $\{1,\dots,n\}$ without $1$-arcs, i.e. arcs of the form $(i,i+1)$ and in which there exists no $k$-set of mutually intersecting arcs. In particular, RNA secondary structures are $2$-noncrossing RNA structures. In this paper we prove a central and a local limit theorem for the distribution of the numbers of $3$-noncrossing RNA structures over $n$ nucleotides with exactly $h$ bonds. We will build on the results of [@Reidys:07rna1] and [@Reidys:07rna2], where the generating function of $k$-noncrossing RNA pseudoknot structures and the asymptotics for its coefficients have been derived. The results of this paper explain the findings on the numbers of arcs of RNA secondary structures obtained by molecular folding algorithms and predict the distributions for $k$-noncrossing RNA folding algorithms which are currently being developed.'
address: |
Center for Combinatorics, LPMC-TJKLC\
Nankai University\
Tianjin 300071\
P.R. China\
Phone: \*86-22-2350-6800\
Fax: \*86-22-2350-9272
author:
- 'Emma Y. Jin and Christian M. Reidys$^{\,\star}$'
bibliography:
- 'central.bib'
date: July 2007
title: Central and Local Limit Theorems for RNA Structures
---
[^1]
Introduction
============
An RNA molecule consists of the primary sequence of the four nucleotides [**A**]{}, [**G**]{}, [**U**]{} and [**C**]{} together with the Watson-Crick ([**A-U**]{}, [**G-C**]{}) and ([**U-G**]{}) base pairing rules. The latter specify the pairs of nucleotides that can potentially form bonds. Single stranded RNA molecules form helical structures whose bonds satisfy the above base pairing rules and which, in many cases, determine their function. For instance RNA ribosomes are capable of catalytic activity, cleaving other RNA molecules. Not all possible bonds are realized, though. Due to bio-physical constraints and the chemistry of Watson-Crick base pairs there exist rather severe constraints on the bonds of an RNA molecule. In light of this three decades ago Waterman [*et.al.*]{} pioneered the concept of RNA secondary structures [@Waterman:78a; @Waterman:94a], being subject to the most strict combinatorial constraints. Any structure can be represented by drawing the primary sequence horizontally, ignoring all chemical bonds of its backbone, see Fig. \[F:1\]. Then one draws all bonds, satisfying the Watson-Crick base pairing rules as arcs in the upper half-plane, effectively identifying structure with the set of all arcs. In this representation, RNA secondary structures have no $1$-arcs, i.e. arcs of the form $(i,i+1)$ and no two arcs $(i_1,j_1)$, $(i_2,j_2)$, where $i_1<j_1$ and $i_2<j_2$ with the property $i_1<i_2<j_1<j_2$. In other words there exist no two arcs that cross in the diagram representation of the structure.
It is well-known that there exist additional types of nucleotide interactions [@Science:05a]. These bonds are called pseudoknots [@Westhof:92a] and occur in functional RNA (RNAseP [@Loria:96a]), ribosomal RNA [@Konings:95a] and are conserved in the catalytic core of group I introns. Pseudoknots appear in plant viral RNAs pseudo-knots and in [*in vitro*]{} RNA evolution [@Tuerk:92] experiments have produced families of RNA structures with pseudoknot motifs, when binding HIV-1 reverse transcriptase. Important mechanisms like ribosomal frame shifting [@Chamorro:91a] also involve pseudoknot interactions. $k$-noncrossing RNA structures introduced in [@Reidys:07rna1] capture these pseudoknot bonds and generalize the concept of the RNA secondary structures in a natural way. In the diagram representation $k$-noncrossing RNA structure has no $1$-arcs and contains at most $k-1$ mutually crossing arcs.
The starting point of this paper was the experimental finding that $3$-noncrossing RNA structures for random sequences of length $100$ over the nucleotides [**A**]{}, [**G**]{}, [**U**]{} and [**C**]{} exhibited sharply concentrated numbers of arcs (centered at $39$). It was furthermore intriguing that the numbers of arcs were significantly higher than those in RNA secondary structures. While it is evident that $3$-noncrossing RNA structures have more arcs than secondary structures, the jump from $27$ to $39$ (for $n=100$ ) with a maximum number of $50$ arcs was not anticipated. Since all these quantities were via the generating functions for $k$-noncrossing RNA structures in [@Reidys:07rna1] explicitly known we could easily confirm that the numbers of $3$-noncrossing RNA structures with exactly $h$ arcs, ${\sf S}'_3 (n,h)$ satisfy indeed almost “perfectly” a Gaussian distribution with a mean of $39$, see Fig. \[F:3\]. We also found that a central limit theorem holds for RNA secondary structures with $h$ arcs, see Figure \[F:4\]. These observation motivated us to understand how and why these limit distributions arise, which is what the present paper is about. Our main results can be summarized as follows:
[**Theorem.**]{} [*Let ${\sf S}_{3}'(n,h)$ denote the number of $3$-noncrossing RNA structures with exactly $h$ arcs. Then the random variable $X_{n}$ having distribution $\mathbb{P}(X_{n}=h)={\sf S}_{3}'(n,h)/{\sf S}_{3}(n)$ satisfies a central and local limit theorem with mean $0.39089\,n$ and variance $0.041565\,n$.* ]{} Our particular strategy is rooted in our recent work on asymptotic enumeration of $k$-noncrossing RNA structures [@Reidys:07rna2] and a paper of Bender [@Bender:73] who showed how such central limit theorems arise in case of singularities that are poles. In order to put our results into context let us provide some background on central and local limit theorems. Suppose we are given a set $A_{n}$ (of size $a_n$). For instance let $A_{n}$ be the set of subsets of $\{1,\dots,n\}$. Suppose further we are given $A_{n,k}$ (of size $a_{n,k}$), $k \in \mathbb{N}$ representing a disjoint set partition of $A_{n}$. For instance let $A_{n,k}$ be the number of subsets with exactly $k$ elements. Consider the random variable $\xi_{n}$ having the probability distribution $\mathbb{P}(\xi_{n}=k)=a_{n,k}/a_{n}$, then the corresponding probability generating function is given by $$\sum_{k \ge 0}\mathbb{P}(\xi_{n}=k)w^k=\sum_{k \ge
0}\frac{a_{n,k}}{a_{n}}w^k=\frac{\sum_{k \ge 0}a_{n,k}w^k}{\sum_{k
\ge 0}a_{n,k}1^k} \ .$$ Let $\varphi_{n}(w)=\sum_{k \ge 0}a_{n,k}w^k$, then $\frac{\varphi_{n}(w)} {\varphi_{n}(1)}$ is the probability generating function of $\xi_{n}$ and $$f(z,w)=\sum_{n \ge 0}\varphi_{n}(w)z^n
=\sum_{n \ge 0}\sum_{k\ge 0}a_{n,k}w^kz^n$$ is called the bivariate generating function. For instance, in our example we have $\mathbb{P}(\xi_{n})=\frac{{n \choose
k}}{2^n}$ and the resulting bivariate generating function is $$\label{E:bi}
\sum_{n \ge 0}\sum_{k \le n}{n \choose k}w^kz^n=\frac{1}{1-z(1+w)} \ .$$ The key idea consists in considering $f(z,w)$ as being parameterized by $w$ and to study the change of its singularity in an $\epsilon$-disc centered at $w=1$. Indeed the moment generating function is given by $$E(e^{s\xi_{n}})=\sum_{k
\ge 0}\frac{a_{n,k}}{a_{n}}e^{sk}=\frac{\varphi_{n}(e^s)}{\varphi_{n}(1)}
=\frac{[z^n]f(z,e^s)}{[z^n]f(z,1)}$$ and $\frac{[z^n]f(z,e^{it})}{[z^n]f(z,1)}=E(e^{it\xi_{n}})$ is the characteristic function of $\xi_{n}$. This shows that the coefficients of $f(z,w)$ control the distribution, which can, for large $n$, be obtained via singularity analysis. The resulting analysis can be amazingly simple. Let us showcase this in the case of the binomial distribution. Here we have the bivariate generating function $\sum_{n \ge 0}\sum_{k \le n}{n
\choose k}w^kz^n=\frac{1}{1-z(1+w)}$, eq. (\[E:bi\]). The simple pole $r(s)$ of $f(z,e^s)$ is $\frac{1}{1+e^s}$. Observe that $\frac{\varphi_{n}(e^s)}{\varphi_{n}(1)}\sim (\frac{r(0)}{r(s)})^n$ holds for $s$ uniformly in a neighborhood of 0, and Taylor expansion shows $$\frac{\varphi_{n}(e^{it})}{\varphi_{n}(1)}\sim \exp(i\cdot
\frac{n}{2}\cdot t-\frac{1}{2}\cdot \frac{n}{4}\cdot t^2+O(t^3))$$ uniformly for $t$ for any arbitrary finite interval. It remains to apply the Lévy-Cramér theorem (Theorem \[T:K\]) to the normalized characteristic function of the random variable $\frac{\eta_{n}-\frac{n}{2}}{\sqrt{\frac{n}{4}}}$, which yields the asymptotic normality of $\eta_{n}$. Thus ${n \choose k}$ is asymptotically normal distributed with mean $\frac{n}{2}$ and variance $\frac{n}{4}$. As it turns out we will have to work a bit harder to prove our main result. The complication is due to the fact that the generating function for $3$-noncrossing RNA structures is much more complex (and fascinating) than the bivariate function of eq. (\[E:bi\]) which has a simple pole as dominant singularity. For instance, the singularity of the generating function for $3$-noncrossing RNA structures is [*not*]{} a pole but of algebraic-logarithmic type [@Flajolet:05; @Gao:92; @Handbook:95].
Our two main results, Theorem \[T:gauss\] in Section \[S:CL\] and Theorem \[T:local\] in Section \[S:LL\] shed light on the distribution of $3$-noncrossing RNA structures from a global and local perspective. A central limit theorem represents the global perspective on the limiting distribution of some random variable $X_{n}$: $$\lim_{n\to\infty}\mathbb{P}\left(\frac{X_{n}-\mu_{n}}{\sigma_{n}}<x\right)
=
\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{x}e^{-\frac{t^2}{2}}dt \ .$$ Bender observed in [@Bender:73] that a central limit theorem combined with certain smoothness conditions on the coefficients $a_{n,k}$ implies a local limit theorem which considers the difference between $\mathbb{P}(x \le \frac{X_{n}-\mu_{n}} {\sigma_{n}}<x+1)$ and $\frac{1}{\sqrt{2\pi}}\int_{x}^{x+1}e^{-\frac{t^2}{2}}dt$ as $n$ tends to infinity. To be precise, $X_{n}$ satisfies a local limit theorem on some set $S\subset
\mathbb{R}$ if and only if $$\lim_{n \to \infty}\sup_{x\in S}\left|\sigma_{n}
\mathbb{P}\left(\frac{X_{n}-\mu_{n}}{\sigma_{n}}=x\right)-\frac{1}{\sqrt{2
\pi}}e^{-\frac{x^2}{2}}\right|=0$$ holds and we say $X_n$ satisfies a local limit theorem for some $S=\{x\in\mathbb{R}\mid x=o(\sqrt{n})\}$. Why is the smoothness of the $a_{n,k}$ so important? Suppose $a_{n,k}={n \choose k}+(-1)^k2{n \choose k}$, then it follows in analogy to our above argument that a central limit theorem with mean $\frac{1}{4}n$ and variance $\frac{1}{4}n$ holds. However, $\eta_{n}$ does not satisfy a local limit theorem, since $$\left|\sigma_{n}\mathbb{P}(\frac{\eta_{n}-\mu_{n}}{\sigma_{n}}=x)-
\frac{1}{\sqrt{2\pi}}e^{-\frac{x^2}{2}}\right|=\frac{1}{2}\sqrt{n}\
\mathbb{P}(\frac{\eta_{n}-\frac{1}{4}n}{\frac{1}{2}\sqrt{n}}=x)-
\frac{1}{\sqrt{2\pi}}e^{-\frac{x^2}{2}}$$ and for $S=\{\frac{\sqrt{n}}{2}|n=1,2\ldots\}$, we have $\left|\frac{1}{2}
\sqrt{n}\mathbb{P}(\eta_{n}=
\frac{1}{2}n)-\frac{1}{\sqrt{2\pi}}e^{-\frac{n}{4}}\right|\nrightarrow
0$, the key point being that $a_{n,k}$ flips between $-\binom{n}{k}$ and $3{n \choose k}$.
All results of this paper hold for $2$-noncrossing RNA structures, i.e. RNA secondary structures. This is a consequence of an analogous analysis of their respective bivariate generating function. In this case, however, no singular expansion is necessary as the generating function itself can be used. They also give rise to put the asymptotic results on RNA secondary structures of [@Schuster:98a] on a new level. We can pass from computing exponential growth rates to computing distributions for RNA secondary structures with specific properties. To be precise we have for RNA secondary structures
[**Theorem.**]{} [*Let ${\sf S}_{2}'(n,h)$ denote the number of RNA secondary structures with exactly $h$ arcs. Then the random variable $Y_{n}$ having distribution $\mathbb{P}(Y_{n}=h)={\sf
S}_{2}'(n,h)/{\sf S}_{2}(n)$ satisfies a central and local limit theorem with mean $0.27639\,n$ and variance $0.04472\, n$.* ]{}
In particular the theorem predicts a sharp concentration of the number of RNA secondary structures with $55.278\%$ unpaired bases which agrees with the statistics of RNA secondary structures obtained by folding algorithms [@Zuker:79b; @Schuster:98a; @Waterman:86; @Bauer:96; @Tacker:94a; @McCaskill:90a]. Let us finally remark that much more holds: due to the determinant formula for $k$-noncrossing matchings and the functional identity of Lemma \[L:func\], Section \[S:func\] our results can be generalized to $k$-noncrossing RNA structures, where $k$ is arbitrary. Why this is of interest can be seen in Fig. \[F:4\]. For higher $k$ the mean of the central limit theorems for $k$-noncrossing RNA structures will shift towards the maximum combinatorially possible number of arcs. We speculate that each increase in $k$ will basically cut the distance to the maximum arc number in half. This is work in progress.
The paper is structured as follows: In Section \[S:rna\] we provide some background on $k$-noncrossing RNA structures and all generating functions involved. In Section \[S:func\] we give a functional equation for the bivariate generating function of ${\sf
S}_{3}'(n,h)$ via $3$-noncrossing matchings proved in [@Reidys:07rna2]. We have included its proof in the appendix in order to keep the paper self-contained. This functional identity plays a key role in proving the central limit and local limit theorem in Section \[S:CL\] and Section \[S:LL\], respectively. The central limit theorem is proved by analyzing the singular expansion of analytic function of power series $\sum_{n \ge
0}\sum_{h\le \frac{n}{2}}{\sf S}_{3}'(n,h)w^hz^n$ and using transfer theorems [@Flajolet:05; @Gao:92; @Handbook:95] and to prove the local limit theorem, we use a theorem of Hwang [@Hwang:98] and build on our proof of the central limit theorem.
RNA structures {#S:rna}
==============
Let us begin by illustrating the concept of RNA structures. Suppose we are given the primary sequence $${\bf A}{\bf A}{\bf C}{\bf C}{\bf A}{\bf U}{\bf G}{\bf U}{\bf G}{\bf G}
{\bf U}{\bf A}{\bf C}{\bf U}{\bf U}{\bf G}{\bf A}{\bf U}{\bf G}{\bf G}
{\bf C}{\bf G}{\bf A}{\bf C} \ .$$ Structures are combinatorial graphs over the labels of the nucleotides of the primary sequence. These graphs can be represented in several ways. In Figure \[F:5\] we represent a $3$-noncrossing RNA structure with loop-loop interactions in two ways: first we display the structure as a planar graph and secondly as a diagram, where the bonds are drawn as arcs in the positive half-plane.
In the following we will consider structures as diagram representations of digraphs. A digraph $D_n$ is a pair of sets $V_{D_n},E_{D_n}$, where $V_{D_n}= \{1,\dots,n\}$ and $E_{D_n}\subset \{(i,j)\mid 1\le i< j\le n\}$. $V_{D_n}$ and $E_{D_n}$ are called vertex and arc set, respectively. A $k$-noncrossing digraph is a digraph in which all vertices have degree $\le 1$ and which does not contain a $k$-set of arcs that are mutually intersecting, i.e. $$\begin{aligned}
\not\exists\,
(i_{r_1},j_{r_1}),(i_{r_2},j_{r_2}),\dots,(i_{r_k},j_{r_k});\quad & &
i_{r_1}<i_{r_2}<\dots<i_{r_k}<j_{r_1}<j_{r_2}<\dots<j_{r_k} \ .\end{aligned}$$ We will represent digraphs as a diagrams (Figure \[F:5\]) by representing the vertices as integers on a line and connecting any two adjacent vertices by an arc in the upper-half plane. The direction of the arcs is implicit in the linear ordering of the vertices and accordingly omitted.
\[D:rna\] An RNA structure (of pseudo-knot type $k-2$), ${S}_{k,n}$, is a digraph in which all vertices have degree $\le 1$, that does not contain a $k$-set of mutually intersecting arcs and $1$-arcs, i.e. arcs of the form $(i,i+1)$, respectively. We denote the number of RNA structures by ${\sf S}_k(n)$ and the number of RNA structures with exactly $\ell$ isolated vertices and with $h$ arcs by ${\sf
S}_k(n,\ell)$ and ${\sf S}_k'(n,h)$, respectively. Note that ${\sf
S}_k'(n,h)={\sf S}_k(n,n-2h)$.
Let $f_{k}(n,\ell)$ denote the number of $k$-noncrossing digraphs with $\ell$ isolated points. We have shown in [@Reidys:07rna1] that $$\begin{aligned}
\label{E:ww0}
f_{k}(n,\ell)& ={n \choose \ell} f_{k}(n-\ell,0) \\
\label{E:ww1}
\det[I_{i-j}(2x)-I_{i+j}(2x)]|_{i,j=1}^{k-1} &=
\sum_{n\ge 1} f_{k}(n,0)\cdot\frac{x^{n}}{n!} \\
\label{E:ww2}
e^{x}\det[I_{i-j}(2x)-I_{i+j}(2x)]|_{i,j=1}^{k-1}
&=(\sum_{\ell \ge 0}\frac{x^{\ell}}{\ell!})(\sum_{n \ge
1}f_{k}(n,0)\frac{x^{n}}{n!})=\sum_{n\ge 1}
\left\{\sum_{\ell=0}^nf_{k}(n,\ell)\right\}\cdot\frac{x^{n}}{n!} \ .\end{aligned}$$ In particular we obtain for $k=2$ and $k=3$ $$\label{E:2-3}
f_2(n,\ell) = \binom{n}{\ell}\,C_{(n-\ell)/2}\quad
\text{\rm and}\quad f_{3}(n,\ell)=
{n \choose \ell}\left[C_{\frac{n-\ell}{2}+2}C_{\frac{n-\ell}{2}}-
C_{\frac{n-\ell}{2}+1}^{2}\right] \ ,$$ where $C_m$ denotes the $m$-th Catalan number. The derivation of the generating function of $k$-noncrossing RNA structures, given in Theorem \[T:cool1\] below uses advanced methods and novel constructions of enumerative combinatorics due to Chen [*et.al.*]{} [@Chen:07a; @Gessel:92a] and Stanley’s mapping between matchings and oscillating tableaux i.e. families of Young diagrams in which any two consecutive shapes differ by exactly one square. The enumeration is obtained using the reflection principle due to Gessel and Zeilberger [@Gessel:92a] and Lindström [@Lindstroem:73a] combined with an inclusion-exclusion argument in order to eliminate the arcs of length $1$. In [@Reidys:07rna1] generalizations to restricted (i.e. where arcs of the form $(i,i+2)$ are excluded) and circular RNA structures are given. The following theorem provides all data on numbers of $k$-noncrossing RNA structures with $h$ arcs and the numbers of all $k$-noncrossing RNA structures.
[@Reidys:07rna1]\[T:cool1\] Let $k\in\mathbb{N}$, $k\ge 2$, let $C_m$ denote the $m$-th Catalan number and $f_k(n,\ell)$ be the number of $k$-noncrossing digraphs over $n$ vertices with exactly $\ell$ isolated vertices. Then the number of RNA structures with $\ell$ isolated vertices, ${\sf S}_k(n,\ell)$, is given by $$\label{E:da}
{\sf S}_k(n,\ell) = \sum_{b=0}^{(n-\ell)/2}
(-1)^b\binom{n-b}{b}f_k(n-2b,\ell) \ ,$$ where $f_k(n-2b,\ell)$ is given by the generating function in eq. [(\[E:ww1\])]{}. Furthermore the number of $k$-noncrossing RNA structures, ${\sf S}_k(n)$ is $$\label{E:sum}
{\sf S}_k(n)
=\sum_{b=0}^{\lfloor n/2\rfloor}(-1)^{b}{n-b \choose b}
\left\{\sum_{\ell=0}^{n-2b}f_{k}(n-2b,\ell)\right\}$$ where $\{\sum_{\ell=0}^{n-2b}f_{k}(n-2b,\ell)\}$ is given by the generating function in eq. [(\[E:ww2\])]{}.
In principle, Theorem \[T:cool1\] contains all information about the numbers of $k$-noncrossing RNA structures. However, due to the inclusion-exclusion structure of its coefficients it is however difficult to interpret and to express their behavior for large $n$. Subsequent asymptotic analysis [@Reidys:07rna2] produced the following simple formula
[@Reidys:07rna2]\[T:asy3\] The number of $3$-noncrossing RNA structures is asymptotically given by $$\begin{aligned}
\label{E:konk3}
{\sf S}_3(n) & \sim & \frac{10.4724\cdot 4!}{n(n-1)\dots(n-4)}\,
\left(\frac{5+\sqrt{21}}{2}\right)^n \ .\\\end{aligned}$$
A functional equation {#S:func}
=====================
We have shown in the introduction that the bivariate generating function is the key to prove the central and local limit theorems. The following lemma, whose proof is given in the appendix, rewrites this bivariate generating function as a composition of two “simple” functions. This is crucial for the singularity analysis insofar as we can use a phenomenon known as persistence of the singularity of the “outer” function (the [*supercritical case*]{}) [@Flajolet:05]. It basically means that the type of the singularity is determined by the generating function of $k$-noncrossing matchings.
\[L:func\][@Reidys:07rna2] Let $x$ be an indeterminant over $\mathbb{R}$ and $w\in\mathbb{R}$ a parameter. Let $\rho_k(w)$ denote the radius of convergence of the power series $\sum_{n\ge 0} [\sum_{h\le n/2} {\sf S}'_k(n,h)
w^{2h}] x^n$. Then for $ \vert x \vert < \rho_k(w)$ $$\label{E:rr}
\sum_{n\ge 0} \sum_{h\le n/2} {\sf S}_k'(n,h) w^{2h} x^n =
\frac{1}{w^2x^2-x+1} \sum_{n\ge 0} f_k(2n,0)
\left(\frac{wx}{w^2x^2-x+1}\right)^{2n}$$ holds. In particular we have for $w=1$ $$\label{E:oha}
\sum_{n\ge 0} {\sf S}_k(n) z^{n} =\frac{1}{z^2-z+1}\, \sum_{n\ge
0} f_k(2n,0) \left(\frac{z}{z^2-z+1}\right)^{2n}$$ for $z\in \mathbb{C}$ with $\vert z\vert <\rho_{k}(1)$.
To keep the paper selfcontained we give the proof of Lemma \[L:func\] in the Appendix. While (\[E:rr\]) can only be proved on the level of formal power-series for real variables, complex analysis i.e. the interpretation of these generating functions as analytic functions allows to extend the equality to arbitrary complex variables.
\[L:bivariate\] Suppose $\epsilon>0$, $k\in\mathbb{N}$, $k\ge 2$ and $w=e^{\frac{s}{2}}$, where $\vert s\vert <\epsilon$ and $\varphi_{n,k}(s)=\sum_{h\le n/2}{\sf S}_k'(n,h) e^{hs}$. Let $\rho_k(s)\in \mathbb{R^+}$ denote the radius of convergence of $\sum_{n\ge 0} \varphi_{n,k}(s)z^n$ parameterized by $s$. Then we have $$\label{E:kl}
\forall\, s,z\in\mathbb{C};\
\vert s\vert <\epsilon, \vert z\vert< \rho_k(s);\quad
\sum_{n\ge 0}\varphi_{n,k}(s)z^n= \frac{1}{e^{s}z^2-z+1} \sum_{n\ge
0} f_k(2n,0) \left(\frac{e^{\frac{s}{2}}z}{e^{s}z^2-z+1}\right)^{2n}
\ .$$ Furthermore $\sum_{n\ge 0}\varphi_{n,3}(s)z^n$ has an analytic continuation, $\Xi_3(z,s)$. For $\epsilon$ sufficiently small and $\vert s\vert <\epsilon$, $\Xi_3(z,s)$ has exactly $6$ singularities, $4$ of which have distinct moduli.
We first prove eq. (\[E:kl\]). For this purpose we observe that $$\label{E:G}
\forall\, \vert s\vert<\epsilon,\; \vert z\vert < \rho_k(s) \qquad
G(z,s)=\frac{1}{e^{s}z^2-z+1} \sum_{n\ge 0} f_k(2n,0)
\left(\frac{e^{\frac{s}{2}}z}{e^{s}z^2-z+1}\right)^{2n}$$ considered as a power series in $e^{\frac{1}{2}s}$ is analytic in a neighborhood of $s=0$, since $G(z,0)$ is analytic for $\vert z \vert
<\rho_{k}(0)$. In addition, we can interpret $\sum_{n\ge
0}\varphi_{n,k}(s)z^n$ as a power series in $e^{\frac{1}{2}s}$: $$\label{E:ana-s}
\sum_{n\ge 0}\left(\sum_{h\le n/2}{\sf S}_k'(n,h)e^{hs}\right)z^n
=\sum_{h\ge 0}\left(\sum_{n\ge 2h}{\sf S}_k'(n,h)z^n\right)
\left(e^{s}\right)^h =\sum_{h\ge 0}\psi_{h}(z)
\left(e^{\frac{1}{2}s}\right)^{2h} \ .$$ Therefore $G(z,s)$ and the power series $\sum_{n\ge 0}\varphi_{n,k}(s)z^n$ are analytic in the indeterminant $e^{\frac{1}{2}s}$ in an $\epsilon$-disc centered at $0$. Lemma \[L:func\] implies that for $s\in ]-\epsilon,\epsilon[$ the analytic functions $G(z,s)$ and $\sum_{n\ge 0}\varphi_{n,k}(s)z^n$ are equal. Since any two functions that are analytic at $0$ and that coincide on the interval $]-\epsilon,\epsilon[$ are identical, we obtain $$\label{E:gleich}
\forall\, \vert s\vert<\epsilon,\; \vert z\vert < \rho_k(s)\qquad
G(z,s)=\sum_{n\ge 0}\varphi_{n,k}(s)z^n \ .$$ [*Claim $1$.*]{} Suppose $\vert s\vert <\epsilon$. Then $\sum_{n\ge 0}
\varphi_{n,3}(s)z^n$ has an analytic continuation, $\Xi_3(z,s)$, which has exactly $6$ singularities $4$ of which have distinct moduli.\
In order to prove Claim $1$ we observe that the power series $\sum_{n\ge 0} f_3(2n,0) y^{n}$ has the analytic continuation $\Psi(y)$ (obtained by MAPLE sumtools) given by $$\label{E:psi}
\Psi(y)= \frac{-(1-16y)^{\frac{3}{2}}
P_{\frac{3}{2}}^{-1}(-\frac{16y+1}{16y-1})} {16\, {y}^{\frac{5}{2}}}
\ ,$$ where $P_{\nu}^{m}(x)$ denotes the Legendre Polynomial of the first kind with the parameters $\nu=\frac{3}{2}$ and $m=-1$. According to eq. (\[E:gleich\]) we have $$\label{E:G2}
\sum_{n\ge 0}\varphi_{n,k}(s)z^n=\frac{1}{e^{s}z^2-z+1}
\sum_{n\ge 0} f_k(2n,0) \left(\frac{e^{\frac{s}{2}}z}{e^{s}z^2-z+1}\right)^{2n}$$ which implies that $\sum_{n\ge 0}\varphi_{n,3}(s)z^n$ has the analytic continuation $$\forall\, \vert s\vert<\epsilon,\;\qquad \Xi_3(z,s)=
\frac{1}{e^sz^2-z+1}\,
\Psi\left(\frac{e^{\frac{1}{2}s}z}{e^sz^2-z+1}\right)^2\ .$$ In particular for $s=0$, $\Xi_3(z,0)$ is the analytic continuation of the power series $\sum_{n\ge 0} {\sf S}_3(n) z^n$. We proceed by showing that $\Xi_3(z,s)$ has exactly $6$ singularities and 4 of them have different moduli in $\mathbb{C}$ parameterized by $s$. Two singularities are given by the roots of $e^sz^2-z+1$ are $\zeta_1(s)=\frac{1-\sqrt{1-4e^s}}{2e^s}$ and $\zeta_2(s)=\frac{1+\sqrt{1-4e^s}}{2e^s}$. Observe that $\vert
\zeta_1(0)\vert=\vert\zeta_2(0) \vert=1$ and polynomial $e^sz^2-z+1$ depends continuously on $e^{\frac{s}{2}}$, therefore $\zeta_1(s)$ and $\zeta_2(s)$ could potentially have equal modulus for $\vert
s\vert<\epsilon$. The remaining $4$ singularities are induced by the the unique dominant singularity $\alpha_1=\frac{1}{16}$ of analytic function $\Psi(y)$. The function $\Psi(y)$ has three singularities, two of them $\alpha_1=\frac{1}{16}$ and $\alpha_2=+\infty$ are branch points and the other $\alpha_3=0$ is a removable singularity. The function $g(z)=\left(\frac{e^{\frac{s}{2}}z}{e^sz^2-z+1}\right)^2$ with $g(0)=0$ has a radius of convergence of $1$ as $s$ tends to $0$. Therefore the singularity type only depends on $\Psi(y)$ (this is the [*supercritical case*]{} in [@Flajolet:05]). The singularity $\alpha_1=\frac{1}{16}$ gives rise to the equations $$0=e^sz^2-(1+4e^{\frac{1}{2}s})z +1 \quad \text{\rm and} \quad
0=e^sz^2+(4e^{\frac{1}{2}s}-1)z+1$$ and setting $\mu_+(s)=1+ 4e^{\frac{1}{2}s}$, $\mu_-(s)=1- 4e^{\frac{1}{2}s}$ and $\theta(s)=\sqrt{12e^s+8e^{\frac{1}{2}s}+1}$ its roots are given by $$\zeta_3(s)=\frac{\mu_+(s)-\theta(s)}{2e^s}, \
\zeta_4(s)=\frac{\mu_+(s)+\theta(s)}{2e^s},\
\zeta_5(s)=\frac{\mu_-(s)+\theta(s)}{2e^s}\ \text{\rm and}\
\zeta_6(s)=\frac{\mu_-(s)-\theta(s)}{2e^s},$$ respectively. Observe that for $\vert s \vert <\epsilon$, $e^{\frac{s}{2}}$ is in a neighborhood of $1$ over $\mathbb{C}$, hence $\theta(s)\ne 0$. That leads to $4$ distinct roots $\zeta_{3}(s),\zeta_{4}(s),\zeta_{5}(s),\zeta_{6}(s)$ over $\vert
s\vert <\epsilon$, all of them have distinct moduli for $s$ being a sufficiently small neighborhood of 0. Indeed, for $s=0$ we have $4$ distinct real valued roots $$\zeta_3(0)=\frac{5-\sqrt{21}}{2}, \
\zeta_4(0)=\frac{5+\sqrt{21}}{2},\ \zeta_5(0)=\frac{-3+\sqrt{5}}{2},
\ \text{\rm and}\ \zeta_6(0)=\frac{-3-\sqrt{5}}{2}$$ and the polynomials $e^sz^2-(1+4e^{\frac{1}{2}s})z
+1$, $e^sz^2+(4e^{\frac{1}{2}s}-1)z+1$ and $e^sz^2-z+1$ depend continuously on the parameter $e^{\frac{1}{2}s}$, whence Claim $1$ and the lemma follows.
The central limit theorem {#S:CL}
=========================
In this section we prove a central limit theorem for the numbers of $3$-noncrossing RNA structures with $h$ arcs. We will analyze for fixed but arbitrary $n$ the distribution of ${\sf S}_3'(n,h)$. Let us first prepare some methods and results used in the proof of Theorem \[T:gauss\]. $[z^n]\,f(z)$ denotes the coefficient of $z^n$ in the power series expansion of $f(z)$ around $0$. The scaling property of Taylor coefficients $$\label{E:scaling}
\forall \,\gamma\in\mathbb{C}\setminus 0;\quad
[z^n]f(z)=\gamma^n [z^n]f(\frac{z}{\gamma}) \ ,$$ shows that w.l.o.g. any singularity analysis can be reduced to the case where $1$ is the dominant singularity. We will be interested in the behavior of an analytic function “locally”, i.e. around a certain singularity $\rho$. For this purpose we use the notation $$\label{E:genau}
f(z)=O\left(g(z)\right) \ \text{\rm as $z\rightarrow \rho$}\quad
\Longleftrightarrow \quad f(z)/g(z) \ \text{\rm is bounded as
$z\rightarrow \rho$}$$ and if we write $f(z)=O(g(z))$ it is implicitly assumed that $z$ tends to a (unique) singularity. Given two numbers $\phi,R$, where $R>\vert \rho \vert>0$ and $0<\phi<\frac{\pi}{2}$ and $\rho \in\mathbb{C}$ the open domain $\Delta_{\rho}(\phi,R)$ is defined as $$\Delta_{\rho}(\phi,R)=\{ z\mid \vert z\vert < R, z\neq \rho,\, \vert
{\rm Arg}(z-\rho)\vert
>\phi\}$$ A domain is a $\Delta_{\rho}$-domain if it is of the form $\Delta_{\rho}(\phi,R)$ for some $R$ and $\phi$. A function is $\Delta_{\rho}$-analytic if it is analytic in some $\Delta_{\rho}$-domain. We use $U(a,r)=\{z\in \mathbb{C} \mid
\vert z-a\vert<r\}$ to denote the open neighborhood of $a$ in $\mathbb{C}$. Via the following theorem we can extract the coefficients of analytic functions provided these functions satisfy certain “local” properties.
\[T:transfer1\][@Flajolet:05] Let $r\in\mathbb{Z}_{\ge 0}$ and $f(z,e^s)$ be a $\Delta_{\rho(s)}$-analytic function parameterized by $s$, which satisfies in the intersection of a neighborhood of $\rho(s)$ with its $\Delta_{\rho(s)}$-domain $$\label{T:transfer2}
f(z,e^s) = b_0(s)+b_1(s)(z-\rho(s))+A(s) \
(\rho(s)-z)^{r}\ln^{}\left(\frac{1}{\rho(s)-z}\right) + R(z,s)$$ where $A(s),b_0(s),b_1(s)$ are analytic in $\vert s\vert<
\epsilon$ and $\vert R(z,s)\vert \le c \, \vert \rho(s)-z \vert$ for some absolute constant $c\in \mathbb{C}$. That is we have $f(z,e^s)=O((\rho(s)-z)^{r}\ln(\frac{1}{\rho(s)-z}))$ with uniform error bound as $s$ in a neighborhood of $0$. Then we have $$[z^n]f(z,e^s)= A(s)\, (-1)^r\frac{r!}{n(n-1)\dots(n-r)}
\left(1-O(\frac{1}{n})\right)\quad \text{\it for some
$A(s)\in\mathbb{C}$}\ ,$$ where the error term is again uniform for $s$ from a neighborhood of origin, i.e. $R(s)\le c\, \vert s\vert$, where $c>0$.
[**Remark.**]{} The equivalence between eq. (\[T:transfer2\]) and $f(z,e^s)=
O((\rho(s)-z)^{r}\ln(\frac{1}{\rho(s)-z}))$ for $r\in\mathbb{Z}_{\ge 0}$ can be seen as follows: by definition of $f(z,e^s)=
O((\rho(s)-z)^{r}\ln(\frac{1}{\rho(s)-z}))$ there exist $A(z,s)$ and $B(z,s)$, such that $f(z,e^s)=B(z,s)+A(z,s)(\rho(s)-z)^{r}\ln(\frac{1}{\rho(s)-z})$, where $A(z,s)$ and $B(z,s)$ are analytic in a neighborhood of $\rho(s)$. Taylor expansion of $A(z,s)$ and $B(z,s)$ at $z=\rho(s)$ produces $$\begin{aligned}
f(z,s)&=B(z,s)+A(z,s)(\rho(s)-z)^{r}\ln\left(\frac{1}{\rho(s)-z}\right)\\
&=b_{0}(s)+b_{1}(s)(z-\rho(s))+\cdots+\left(a_0(s)+a_{1}(z-\rho(s))+
\cdots\right)
(\rho(s)-z)^{r}\ln\left(\frac{1}{\rho(s)-z}\right)\\
&=b_{0}(s)+b_{1}(s)(z-\rho(s))+a_{0}(s)(\rho(s)-z)^{r}
\ln\left(\frac{1}{\rho(s)-z}\right)+R(z,s)\end{aligned}$$ where $R(z,s)=O((\rho(s)-z)^{r+1}\ln\left(\frac{1}{\rho(s)-z}\right))$. For $r \in \mathbb{Z}_{\ge 0}$, $\frac{|R(z,s)|}{|\rho(s)-z|}=O(|\rho(s)-z|^{r}
\ln\left(\frac{1}{|\rho(s)-z|}\right))$ is bounded by an absolute constant as $z$ tends to $\rho(s)$. That implies the error bound is uniform.
The next Theorem is a classic result on limit distributions which allows us to prove our main result via characteristic functions i.e. explicitly by showing $\lim_{n \rightarrow
\infty}\varphi_{n}(t)=\varphi(t)$ for any $t \in (-\infty,\infty)$.
\[T:K\][**(Lévy-Cramér)**]{} Let $\{\xi_{n}\}$ be a sequence of random variables and let $\{\varphi_{n}(x)\}$ and $\{F_{n}(x)\}$ be the corresponding sequences of characteristic and distribution functions. If there exists a function $\varphi(t)$, such that $\lim_{n \rightarrow
\infty}\varphi_{n}(t)=\varphi(t)$ uniformly over an arbitrary finite interval enclosing the origin, then there exists a random variable $\xi$ with distribution function $F(x)$ such that $$F_{n}(t)\Longrightarrow F(x)$$ uniformly over any finite or infinite interval of continuity of $F(x)$.
We now consider the random variable $X_n$ having the distribution $\mathbb{P}(X_n=h)={\sf S}_3'(n,h)/{\sf S}_3(n)$, where $h=0,1,\ldots \lfloor \frac{n}{2}\rfloor$. The key point in the proof of Theorem \[T:gauss\] is to compute the coefficients of the bivariate generating function whose variable, $s$ is considered as a parameter. Intuitively the particular distribution is a result of how the singularity shifts as a function of this parameter. As a result the proof is somewhat “non-probabilistic” and has two distinct parts: [(a)]{} the analytic combinatorics of the bivariate generating function and [(b)]{} the computation of the characteristic function with subsequent application of the [Lévy-Cramér]{} Theorem.
\[T:gauss\] The random variable $\frac{X_{n}-\mu n}{\sqrt{\sigma^2 n}}$ has asymptotically normal distribution with parameter $(0,1)$, i.e. $$\label{E:converge}
\lim_{n\to\infty}\mathbb{P}\left(\frac{X_n-\mu n}{\sqrt{\sigma^2 n}}< x
\right) = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{x}
e^{-\frac{1}{2}t^2} dt$$ and $\mu,\sigma^2$ are given by $$\label{E:concrete0}
\mu = -\frac{-\frac{3}{2}+\frac{13}{42}\sqrt{21}}{\frac{5}{2}-
\frac{1}{2}\sqrt{21}}=0.39089 \quad \text{\rm and}
\quad
\sigma^2 = \mu^2-\frac{1-\frac{94}{441}\sqrt{21}}
{\frac{5-\sqrt{21}}{2}}=0.041565 \ .$$
We set $w=e^{\frac{1}{2}s}$ and $\varphi_{n,3}(s)=\sum_{h\le
n/2}{\sf S}_3' (n,h)e^{hs}$. Since $$\label{E:phi}
\sum_{n\ge 0}\varphi_{n,3}(s)z^n = \sum_{n\ge 0}\left(\sum_{h\le
n/2}{\sf S}_3'(n,h)e^{hs}\right)z^n \ ,$$ we can consider the double generating function $\sum_{n\ge 0}\sum_{h\le n/2}
{\sf S}_3'(n,h)w^{2h}z^n$ as a power series in the complex indeterminant $z$, parameterized by $s$.\
[*Claim $1$.*]{} $$\Psi(z)=O\left((1-16z)^4\ln\left(\frac{1}{1-16z}\right)\right)\quad
\ \text{\rm holds uniformly for }\forall\,
z\in\Delta_{\frac{1}{16}}(\phi,R)\cap U(\frac{1}{16},\epsilon);$$ $\Psi(z)$ is $\Delta_{\frac{1}{16}}(\phi,R)$-analytic and has the singular expansion $(1-16z)^4\ln\left(\frac{1}{1-16z}\right)$ in the intersection of $U(\frac{1}{16},\epsilon)$ with the $\Delta_{\frac{1}{16}}-$domain, where $\Delta_{r}(\phi,R)=\{z\big|
|z|<R, z\ne r,\vert{\rm Arg}(z-r)\vert>\phi \}$ for some $R>r$. First $\Delta_{\frac{1}{16}}(\phi,R)$-analyticity of the function $(1-16z)^4\ln\left(\frac{1}{1-16z}\right)$ is obvious. We proceed by proving that $(1-16z)^4\ln\left(\frac{1}{1-16z}\right)$ is the singular expansion of $\Psi(z)$. The above mentioned scaling property of Taylor coefficients allows us to consider the power series $\sum_{n \ge 0}f_{3}(2n,0)(\frac{z} {16})^n$ over the $\Delta$-domain $\Delta_{1}(\phi,R)$ for some $R>1$. Using the notation of falling factorials $(n-1)_4=(n-1)(n-2)(n-3)(n-4)$ we observe $$f_3(2n,0)=C_{n+2}C_{n}-C_{n+1}^2= \frac{1}{(n-1)_4}
\frac{12(n-1)_4(2n+1)}{(n+3)(n+1)^2(n+2)^2}\,
\binom{2n}{n}^2 \ .$$ With this expression for $f_3(2n,0)$ we arrive at the formal identity $$\begin{aligned}
\sum_{n\ge 5}16^{-n}f_3(2n,0)z^n & = &
O(\sum_{n\ge 5}
\left[16^{-n}\,\frac{1}{(n-1)_4}
\frac{12(n-1)_4(2n+1)}{(n+3)(n+1)^2(n+2)^2}\,
\binom{2n}{n}^2-\frac{4!}{(n-1)_4}\frac{1}{\pi}\frac{1}{n}\right]z^n \\
& & + \sum_{n\ge 5}\frac{4!}{(n-1)_4}\frac{1}{\pi}\frac{1}{n}z^n) \ ,\end{aligned}$$ where $f(z)=O(g(z))$ denotes that the limit $f(z)/g(z)$ is bounded for $z\rightarrow 1$, eq. (\[E:genau\]). It is clear that the error bound below $$\begin{aligned}
& &\sum_{n\ge 5}\left[16^{-n}\,\frac{1}{(n-1)_4}
\frac{12(n-1)_4(2n+1)}{(n+3)(n+1)^2(n+2)^2}\,
\binom{2n}{n}^2-\frac{4!}{(n-1)_4}\frac{1}{\pi}\frac{1}{n}\right]z^n \\
&\sim & \sum_{n\ge 5} \left[16^{-n}\,\frac{1}{(n-1)_4}
\frac{12(n-1)_4(2n+1)}{(n+3)(n+1)^2(n+2)^2}\,
\binom{2n}{n}^2-\frac{4!}{(n-1)_4}\frac{1}{\pi}\frac{1}{n}\right]
<\kappa\end{aligned}$$ holds uniformly for $z$ in $\Delta_{1}(\phi,R)\cap U(1,\epsilon)$ and some absolute $\kappa< 0.0784$. Therefore we can conclude $$\sum_{n\ge 5}16^{-n}f_3(2n,0)z^n=
O(\sum_{n\ge 5}\frac{4!}{(n-1)_4}\frac{1}{\pi}\frac{1}{n}z^n) \ .$$ We proceed by interpreting the power series on the rhs, observing $$\forall\, n\ge 5\, ; \qquad
[z^n]\left((1-z)^4\,\ln\frac{1}{1-z}\right)=
\frac{4!}{(n-1)\dots (n-4)}\frac{1}{n} \, ,$$ whence $\left((1-z)^4\,\ln\frac{1}{1-z}\right)$ is the unique analytic continuation of $\sum_{n\ge 5}\frac{4!}{(n-1)_4}
\frac{1}{\pi}\frac{1}{n}z^n$. Using the scaling property of Taylor coefficients $[z^n]f(z)=\gamma^n [z^n]f(\frac{z}{\gamma})$ $$\label{E:isses}
\Psi(z) =O\left((1-16z)^4\ln\left(\frac{1}{1-16z}\right)\right)
\mbox{ holds uniformly for }\forall\,
z\in\Delta_{\frac{1}{16}}(\phi,R)\cap U(\frac{1}{16},\epsilon)$$ Therefore we have proved that $(1-16z)^{4}\ln(\frac{1}{1-16z})$ is the singular expansion of $\Psi(z)$ at $z=\frac{1}{16}$, whence Claim $1$. Our next step consists in verifying that when passing from $\Psi(z)$ to the bivariate generating function $\Psi(z,s)= \Psi((\frac{wz}{w^2z^2-z+1})^2)$, then there exists a singular expansion of the form ${O}\left((1-\frac{z}{\rho_3(s)})^4\ln(\frac{1}{1-\frac{z}
{\rho_3(s)}})\right)$, parameterized in $s$.\
[*Claim $2$*]{}. Let $0<\epsilon<1$, then for any $\vert
s\vert<\epsilon$ and $z\in \Delta_{\rho_3(s)}(\phi,R)$, we have $\Psi(z,s)
={O}\left((1-\frac{z}{\rho_3(s)})^4\ln(\frac{1}{1-\frac{z}
{\rho_3(s)}})\right)$, and the error bound is uniform for $s$ in a neighborhood of $0$.\
To prove the claim we first observe that Claim $1$ implies $$\Psi(z)=\kappa(1-16z)^4\ln\left(\frac{1}{1-16z}\right)+R(z)$$ for some absolute constant $\kappa$ and $R(z)$ is the uniform error bound for $z\in \Delta_{\frac{1}{16}}(\phi,R)\cap
U(\frac{1}{16},\epsilon)$. I.e. For $z\in
\Delta_{\frac{1}{16}}(\phi,R)\cap U(\frac{1}{16},\epsilon)$, there exists some absolute constant $c$, such that $|R(z)|\le c\cdot
|1-16z|$ holds. According to Lemma \[L:bivariate\] we have $$\begin{aligned}
\Xi_3(z,s) &=\frac{1}{e^sz^2-z+1}\ O
\,\left(\left(1-16(\frac{e^{\frac{1}{2}s}z}
{e^sz^2-z+1})^2\right)^4\ln\frac{1}{\left(1-16(
\frac{e^{\frac{1}{2}s}z}{e^sz^2-z+1})^2\right)}\right) \\
&=\frac{\kappa}{e^{s}z^2-z+1}
\left(1-16(\frac{e^{\frac{1}{2}s}z}{e^{s}z^2-z+1})^2\right)^4
\ln\left(\frac{1}{1-16(\frac{e^{\frac{1}{2}s}z}{e^{s}z^2-z+1})^2}\right)+
R(z,e^s) \ .\end{aligned}$$ We expand $\left(1-16(\frac{e^{\frac{1}{2}s}z}{e^{s}z^2-z+1})^2\right)^4
\ln\left(\frac{1}{1-16(\frac{e^{\frac{1}{2}s}z}{e^{s}z^2-z+1})^2}\right)$ around $z=\rho_{3}(s)$, where $\rho_3(s)$ is the solution of $\frac{ze^{\frac{1}{2}s}}{e^sz^2-z+1} =\frac{1}{4}$ of minimal modulus. Lemma \[L:bivariate\] implies that $$\rho_3(s)=\frac{4e^{\frac{1}{2}s}+1-\sqrt{12e^s+8e^{\frac{1}{2}s}+1}}{2e^s}$$ is the unique dominant singularity. As a function in $s$ we have $\rho_3'(0)=-\frac{3}{2}+\frac{13}{42}\sqrt{21}\neq 0$. The term $\sqrt{12e^s+8e^{\frac{1}{2}s}+1}$ in $\rho_{3}(s)$ produces two branching points parameterized by $s$. i.e. $w=e^{\frac{1}{2}s}=-\frac{1}{6}$ and $w=e^{\frac{1}{2}s}=-\frac{1}{2}$, or equivalently $s=2\ln\frac{1}{2}+2\pi
i$ and $s=2\ln\frac{1}{6}+2\pi i$, respectively. The interval between $2\ln\frac{1}{6}+2\pi i$ and $2\ln\frac{1}{2}+2\pi i$ divides the complex plane of $s$ into two analytic branches. For any $0 <\epsilon < \min\{\vert 2\ln\frac{1}{2}+2\pi i\vert, \vert
2\ln\frac{1}{6}+2\pi i\vert\}=6.4343$, the region $\vert s \vert <
\epsilon$ is disjoint to the interval $[(2\ln\frac{1}{6},2\pi),(2\ln\frac{1}{2},2\pi)]$. Therefore $\rho_{3}(s)$ is analytic for $\vert s\vert<\epsilon$. We next consider $q(z,s)=1-16(\frac{e^{\frac{1}{2}s}z} {e^sz^2-z+1})^2$ as a function of $z$ and compute the Taylor expansion at $\rho_3(s)$. $$q(z,s)=\alpha(\rho_3(s)-z)+{O}(z-\rho_3(s))^2$$ and setting $\alpha=\frac{\sqrt{21}}{5-\sqrt{21}}$ $$\begin{aligned}
\frac{1}{e^sz^2-z+1}\, \left[q(z,s)^4\ln\frac{1} {q(z,s)}\right]
&= \frac{(\alpha(\rho_3(s)-z)+{O}(z-\rho_3(s))^2)^4\ln\frac{1}
{\alpha(\rho_3(s)-z)+{O}(z-\rho_3(s))^2}}{e^s(z-\rho_3(s))^2+(2\rho_3(s)e^s-1)
(z-\rho_3(s))-3\rho_3(s)^2e^s+\rho_3(s)+1}\\
&= \frac{\left([\alpha+O(z-\rho_3(s))](\rho_3(s)-z)^4
\ln\frac{1}{[\alpha+O(z-\rho_3(s))](\rho_3(s)-z)}
\right)}{O(z-\rho_3(s))-3\rho_3(s)^2+\rho_3(s)+1} \\
&=O(\rho_{3}(s)-z)^4\ln\left(\frac{1}{\rho_{3}(s)-z}\right) \ .\end{aligned}$$ According to Theorem \[T:transfer1\] for $r=4$, we obtain the error term in the expansion of $\frac{1}{e^sz^2-z+1}\,
\left[q(z,s)^4\ln\frac{1} {q(z,s)}\right]$ is uniform for $s$ in a neighborhood of $0$. We observe that the resulting error bound for $\Xi_{3}(z,s)$ is the sum $R(z,e^s)+R_{1}(z,e^s)$, where $$\vert R(z,e^s)\vert \le
c\cdot\big|1-16\left(\frac{e^{\frac{1}{2}s}z}{e^{s}z^2-z+1}\right)\big|
=O(\rho_{3}(s)-z) \ .$$ Therefore the error bound for the expansion of bivariate $\Xi_{3}(z,s)$ is uniform and Claim $2$ is proved. We proceed by using the scaling property of Taylor coefficients $[z^n]f(z)= \gamma^n [z^n]f(\frac{z}{\gamma})$ and apply Theorem \[T:transfer1\]. Via Theorem \[T:transfer1\] we obtain the key information about the coefficients of $\Xi_3(z,s)$ which allows us to substitute $\varphi_{n,3}(\frac{it}{\sigma_n})$ in eq. (\[E:well\]) below: $$\label{E:4}
[z^n]\,\Xi_3(z,s)= K(s) \,\frac{4!}{n(n-1)\dots(n-4)}\,
\left(\rho_3(s)^{-1}\right)^n \left(1-O(\frac{1}{n})\right)\quad
\text{\rm for some $K(s)\in\mathbb{C}$}\ ,$$ where the error term is again uniform for $s$ from a neighborhood of origin.\
Suppose we are given the random variable (r.v.) $\xi_n$ with mean $\mu_n$ and variance $\sigma_n^2$. We consider the rescaled r.v. $\eta_{n}=(\xi_{n}
- \mu_{n})\sigma^{-1}_{n}$ and the characteristic function of $\eta_{n}$: $$f_{\eta_n}(t)=\mathbb{E}[e^{it\eta_{n}}]=
\mathbb{E}[e^{it\frac{\xi_{n}}{\sigma_{n}}}]
e^{-i\frac{\mu_{n}}{\sigma_{n}}t} \ .$$ In particular, for $\xi_n=X_n$ we obtain, substituting the term $\mathbb{E}[e^{it\eta_{n}}]$ $$f_{X_n}(t)=\left(
\sum_{h=0}^{n}\frac{{\sf S}_3'(n,h)}{{\sf S}_3(n)}
e^{it\frac{h}{\sigma_{n}}}\right)\,
e^{-i\frac{\mu_{n}}{\sigma_{n}}t}
\ .$$ Since $\varphi_{n,3}(s)=\sum_{h\le n/2}{\sf S}_3'(n,h)e^{hs}$, we can interpret ${\sf S}_3(n)=\sum_{h\le n/2}{\sf S}_3'(n,h)$ as $\varphi_{n,3}(0)$ and $\varphi_{n,3}(\frac{it}{\sigma_n})=
\sum_{h\le n/2}{\sf S}_3'(n,h)e^{h\frac{it}{\sigma_n}}$, respectively. Therefore we have $$\label{E:well}
f_{X_n}(t)=\frac{1}{\varphi_{n}(0)}\,\varphi_{n}(\frac{it}{\sigma_{n}})\,
e^{-i\frac{\mu_{n}}{\sigma_{n}}t} \ .$$ For $\vert s\vert < \epsilon$, eq. (\[E:4\]) yields $
\varphi(s)=[z^n]\,\Xi_3(z,s)\sim K(s)\,\frac{4!}{n(n-1)\dots(n-4)}\,
\left(\rho_3(s)^{-1}\right)^n $ with uniform error term and we accordingly obtain $$\label{E:uu}
f_{X_n}(t)\sim \frac{K(\frac{it}{\sigma_n})}{K(0)}\,
\left[\frac{\rho_3(\frac{it}{\sigma_n})}{\rho_3(0)}\right]^{-n}
e^{-i\frac{\mu_{n}}{\sigma_{n}}t} \ .$$ where the error term is uniform for $t$ from any bounded interval. Taking the logarithm we obtain $$\ln f_{X_n}(t)\sim\ln\frac{K(\frac{it}{\sigma_n})}{K(0)}- n \,
\ln\frac{\rho_3(\frac{it}{\sigma_n})}{\rho_3(0)} -i \frac{\mu_n}{\sigma_n}t
\ .$$ Expanding $g(s)=\ln\frac{\rho_3(s)}{\rho_3(0)}$ in its Taylor series at $s=0$, (note that $g(0)=0$ holds) yields $$\label{E:kkk}
\ln\frac{\rho_3(\frac{it}{\sigma_{n}})}{\rho_3(0)}=
\frac{\rho_3'(0)}{\rho_3(0)}\frac{it}{\sigma_{n}}-
\left[\frac{\rho_3''(0)}{\rho_3(0)}-
\left(\frac{\rho_3'(0)}{\rho_3(0)}\right)^2\right]
\frac{t^2}{2\sigma^2_{n}}
+{O}(\left(\frac{it}{\sigma_n}\right)^{3})$$ and therefore $$\label{E:nun}
\ln f_{n}(t) \sim\ln \frac{K(\frac{it}{\sigma_n})}{K(0)}- n \,
\left\{\frac{\rho_3'(0)}{\rho_3(0)}\frac{it}{\sigma_{n}}-\frac{1}{2}
\left[\frac{\rho_3''(0)}{\rho_3(0)}-
\left(\frac{\rho_3'(0)}{\rho_3(0)}\right)^2\right]
\frac{t^2}{\sigma^2_{n}}
+{O}(\left(\frac{it}{\sigma_n}\right)^{3})\right\}-
\frac{i\mu_{n}t}{\sigma_{n}} \ .$$ Claim $2$ implies $\Xi_{3}(z,s)=
{O}\left((\rho_3(s)-z)^4\ln\frac{1}{\rho_3(s)-z}\right)$ is analytic in $s$ where $s$ is contained in a disc of radius $\epsilon$ around $0$. Hence $\Xi_{3}(z,s)$ is in particular continuous in $s$ for $\vert s\vert <\epsilon$ and we can conclude from eq. (\[E:4\]) for fixed $t\in ]-\infty, \infty[$ $$\lim_{n\to \infty}\left(\ln K(\frac{it}{\sigma_n})-\ln K(0)\right)=0 \ .$$ In view of eq. (\[E:nun\]) we introduce $$\mu=-\frac{\rho_3'(0)}{\rho_3(0)}, \quad \quad \sigma=\left\{
\left(\frac{\rho_3'(0)}{\rho_3(0)}\right)^2-\frac{\rho_3''(0)}
{\rho_3(0)}\right\}$$ and eq. (\[E:nun\]) becomes $$\ln f_{X_n}(t)\sim -\frac{\; t^2}{2}
+{O}(\left(\frac{it}{\sigma_n}\right)^{3})$$ with uniform error term for $t$ from any bounded interval. This is equivalent to $\lim_{n\rightarrow\infty}f_{X_n}(t)=
\exp(-\frac{t^2}{2})$ with uniform error term. The Lévy-Cramér Theorem (Theorem \[T:K\]) implies now eq. (\[E:converge\]) and it remains to compute the values for $\mu$ and $\sigma$ which are given by $$\begin{aligned}
\label{E:mu}
\mu & = & -\frac{\rho_{3}'(0)}{\rho_{3}(0)}=
-\frac{-\frac{3}{2}+\frac{13}{42}
\sqrt{21}}{\frac{5}{2}-\frac{1}{2}\sqrt{21}}=0.39089 \\
\label{E:sigma}
\sigma^2 & = & \mu^2-\frac{\rho_{3}''(0)}{\rho_{3}(0)}=
\mu^2-\frac{1-\frac{94}{441}\sqrt{21}}{\frac{5-\sqrt{21}}{2}}=0.041565\end{aligned}$$ whence eq. (\[E:concrete0\]) and the proof of Theorem \[T:gauss\] is complete.
The local limit theorem {#S:LL}
=======================
In this section we complement the central limit theorem presented in the previous section $$\lim_{n\to\infty}\mathbb{P}\left(\frac{X_{n}-\mu_{n}}{\sigma_{n}}<x\right)
=
\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{x}e^{-\frac{t^2}{2}}dt$$ by considering a ”local” perspective on the limiting distribution of $X_{n}$. For the local limit theorem we analyze the difference between $\mathbb{P}(x \le \frac{X_{n}-\mu_{n}} {\sigma_{n}}<x+1)$ and $\frac{1}{\sqrt{2\pi}}\int_{x}^{x+1}e^{-\frac{t^2}{2}}dt$ as $n$ tends to infinity. $X_{n}$ satisfies a local limit theorem on some set $S\subset \mathbb{R}$ if and only if $$\label{E:51}
\lim_{n \to \infty}\sup_{x\in S}\left|\sigma_{n}
\mathbb{P}\left(\frac{X_{n}-\mu_{n}}{\sigma_{n}}=x\right)-\frac{1}{\sqrt{2
\pi}}e^{-\frac{x^2}{2}}\right|=0 \ .$$ One key condition formulated in eq. (\[E:wi\]) of Theorem \[T:local\] below for proving a local limit theorem is given by $$\frac{\varphi_{n}(s)}{\varphi_{n}(0)}\sim \exp(M(s)\beta_{n}+N(s)) \ ,$$ where $M(s)$ is differentiable and $N(s)$ is continuous in some $\epsilon$-disc centered at $0$. In view of eq. (\[E:well\]) and eq. (\[E:uu\]) in the proof of the central limit theorem, this condition alone implies the central limit theorem. In other words, the local limit theorem implies the central limit theorem. We have shown in the introduction that a central limit theorem does not imply a local limit theorem. Bender observed in [@Bender:73] that the central limit theorem combined with certain smoothness conditions does imply the local limit theorem. Accordingly, in order to prove the local limit theorem for $3$-noncrossing RNA structures with $h$ arcs our strategy will consist in verifying such smoothness conditions [@Hwang:98].
\[T:local\] Let $\varphi_{n}(s)=\sum_{k}a_{n,k}w^k$ and $w=e^s$. Suppose $$\label{E:wi}
\frac{\varphi_{n}(s)}{\varphi_{n}(0)}\sim \exp(M(s)\beta_{n}+N(s))$$ holds uniformly for $\vert s\vert \le \tau$, $s \in \mathbb{C}$ and $\tau>0$, where the following conditions are satisfied\
${\sf (i)}$ $M(s)$ is differentiable and $N(s)$ is continuous in $\vert s\vert<\epsilon$ and furthermore $M(s)$ and $N(s)$ are independent of $n$.\
${\sf (ii)}$ $\beta_{n}$ is independent of $t$, $\beta_{n} \rightarrow \infty$ and $M''(0)>0$;\
${\sf (iii)}$ there exist constant $\delta$ and $c=c(\delta,r)>0$, where $0<\delta \le \tau$ such that $$\left|\frac{\varphi_{n}(r+it)}{\varphi_{n}(r)}\right|=O(\exp(-c\beta_{n}))$$ holds uniformly for $-\tau \le r\le \tau$ and $\delta \le \vert t
\vert \le \pi$ as $n$ tends to infinity.\
Then random variable $X_{n}$ having distribution $\mathbb{P}(X_{n}=k)=a_{n,k}/a_{n}$ with mean $M'(0)\beta_{n}$ and variance $M''(0)\beta_{n}$ satisfies a local limit theorem on the real set $S=\{x\mid
x=o(\sqrt{\beta_{n}})\}$ i.e. $$\lim_{n \to \infty}\sup_{x\in S}\left|\sigma_{n}
\mathbb{P}\left(\frac{X_{n}-\mu_{n}}{\sigma_{n}}=x\right)-\frac{1}{\sqrt{2
\pi}}e^{-\frac{x^2}{2}}\right|=0 \ .$$
With the help of Theorem \[T:local\] we can now prove the local limit theorem for $3$-noncrossing RNA structures with $h$ arcs.
\[T:local2\] Let ${\sf S}'_3(n,h)$ be the number of $3$-noncrossing RNA structures with exactly $h$ arcs. Let $X_n$ be the r.v. having the distribution $$\forall\; h=0,1,\ldots \lfloor \frac{n}{2}\rfloor,\qquad
\mathbb{P}(X_n=h)=\frac{{\sf S}_3'(n,h)}{{\sf S}_3(n)}$$ Then we have for set $S=\{x\mid x=o(\sqrt{n})\}$ $$\label{E:52}
\lim_{n \to \infty}\sup_{x\in S}\left|\sqrt{\sigma^2 n}\
\mathbb{P}\left(\frac{X_{n}-n\,\mu}{\sqrt{\sigma^2 n}}
=x\right)-\frac{1}{\sqrt{2
\pi}}e^{-\frac{x^2}{2}}\right|=0 \ ,$$ where $\mu=0.39089$ and $\sigma^2 = 0.041565$.
We will show that $\{\frac{{\sf{S}}'_{3}(n,h)}{{\sf{S}}_{3}(n)}\}$ satisfies the conditions for Theorem \[T:local\]. For $\vert
s\vert \le \epsilon$, where $\epsilon$ is sufficiently small but fixed. The crucial equation implying the conditions of Theorem \[T:local\] is eq. (\[E:4\]) of the proof of Theorem \[T:gauss\]: $$\varphi_{n,3}(s)= K(s)\frac{4!}{n(n-1)\ldots
(n-4)}(\rho_{3}(s)^{-1})^n\left(1-O(\frac{1}{n})\right)\qquad
K(s)\in\mathbb{C} \ ,$$ holds uniformly for $\vert s\vert<\epsilon$. Therefore we have $$\frac{\varphi_{n,3}(s)}{\varphi_{n,3}(0)}=
\frac{K(s)}{K(0)}\left(\frac{\rho_{3}(0)}{\rho_{3}(s)}\right)^n
\left(1-O(\frac{1}{n})\right)\sim
\exp\left(n\ln\left(\frac{\rho_{3}(0)}{\rho_{3}(s)}\right)+
\ln\left(\frac{K(s)}{K(0)}\right)\right) \ .$$ uniformly for $\vert s\vert< \epsilon$. We set $$\beta_{n} = n, \quad
M(s) = \ln\left(\frac{\rho_{3}(0)}{\rho_{3}(s)}\right) \
\text{\rm and} \
N(s) = \ln\left(\frac{K(s)}{K(0)}\right) \ .$$ By construction $t$ is independent of $n$ and clearly $n
\rightarrow \infty$ and $M(s)$ is differentiable and $N(s)$ is continuous for all $s$ such that $\vert s\vert<\epsilon$. In addition $M''(0)$ is analytic for $\vert s\vert<\epsilon$ and we have $M''(0)=\mu^2-\frac{1-\frac{94}{441}\sqrt{21}}{\frac{5-\sqrt{21}}{2}}=
0.041565>0$. Let $\delta=\epsilon$ and $-\epsilon\le r\le \epsilon$, we obtain $$\frac{\varphi_{n,3}(r+it)}{\varphi_{n,3}(r)}\sim
\exp\left(n\ln\left(\frac{\rho_{3}(r)}
{\rho_{3}(r+it)}\right)+\ln\left(\frac{K(r+it)}{K(r)}\right)\right)
$$ uniformly for $-\epsilon \le r\le \epsilon$ and $\epsilon\le \vert t\vert\le \pi$. Since $\frac{K(s)}{K(0)}$ yields a constant factor and $K(s)$ is continuous for $|s|<\epsilon$, it suffices to analyze $\ln\left(\frac{\rho_{3}(r)}{\rho_{3}(r+it)}\right)$. We observe $\rho_{3}(s)=\frac{1+4e^{\frac{s}{2}}-\sqrt{12e^s-8e^{\frac{s}{2}}+1}}{2e^s}
\ne 0$ for any complex $s$ where $\vert s\vert< \epsilon$. The singularities of $\ln(\frac{\rho_{3}(0)}{\rho_{3}(s)})$ correspond to the zeros of $12e^s-8e^{\frac{s}{2}}+1=(2e^{\frac{s}{2}}+1)
(6e^{\frac{s}{2}}+1)$, that is $e^{\frac{s}{2}}=-\frac{1}{2}$ or $-\frac{1}{6}$. Observe that for $\vert s \vert < \epsilon$, $\vert
e^{\frac{s}{2}}\vert$ is close to 1. Therefore $\ln\left(\frac{\rho_{3}(r)}{\rho_{3}(r+it)}\right)$ is analytic for any $\epsilon\le|t|\le \pi$ and $r \in ]-\epsilon,\epsilon[$ and we can conclude $$\left|\frac{\varphi_{n,3}(r+it)}{\varphi_{n,3}(r)}\right|=O\left|\exp(n\cdot
\ln\left(\frac{\rho_{3}(it)}{\rho_{3}(0)}\right))\right|=
O\left(\exp\left(\mbox{Re}\left(n\cdot
\ln\left(\frac{\rho_{3}(r+it)}{\rho_{3}(r)}\right)\right)\right)\right)$$ uniformly for $-\epsilon \le r\le \epsilon$ and $\epsilon\le \vert t\vert\le \pi$. Taylor expansion of $\ln(\frac{\rho_{3}(r+it)}{\rho_{3}(r)})$ at $0$ shows (see eq. (\[E:kkk\])), that the dominant real part of $\ln(\frac{\rho_{3}(r+it)}{\rho_{3}(r)})$ is given by $$\left[\left(\frac{\rho_{3}'(r)}{\rho_{3}(r)}\right)^2-\frac{\rho_{3}''(r)}
{\rho_{3}(r)}\right]\frac{t^2}{2!}<0 \mbox{ for }
r\in]-\epsilon,\epsilon[\ .$$ Setting $c_{1}=
\left[\frac{\rho_{3}''(r)}{\rho_{3}(r)}-\left(\frac{\rho_{3}'(r)}
{\rho_{3}(r)}\right)^2\right]\frac{\pi^2}{2!}>0$ and $c_{2}=\left[\frac{\rho_{3}''(r)}{\rho_{3}(r)}-\left(\frac{\rho_{3}'(r)}
{\rho_{3}(r)}\right)^2\right]\frac{\delta^2}{2!}>0$ we can conclude $$\left|\frac{\varphi_{n,3}(r+it)}{\varphi_{n,3}(r)}\right|=O(\exp(-c\cdot
n))$$ for some $0<c_{2}<c<c_{1}$, uniformly for $-\epsilon \le r\le \epsilon$ and $\epsilon\le \vert t\vert\le \pi$ and Theorem \[T:local\] applies, whence Theorem \[T:local2\].
Appendix
========
[**Proof of Lemma \[L:func\].**]{} First we observe that for $x,w\in [-1,1]$ the term $w^2x^2-x+1$ is strictly positive. We set $$F_k(x,w)=\sum_{n\ge 0} \sum_{h\le n/2}{\sf S}_k'(n,h)w^{2h}x^n$$ and compute $$\begin{aligned}
F_k(x,w) & = & \sum_{n\ge 0}\sum_{h\le
n/2}\sum_{j=0}^h(-1)^j\binom{n-j}{j}
\binom{n-2j}{2(h-j)} f_k(2(h-j),0)w^{2h}x^n \\
& = & \sum_{n\ge 0}\sum_{j\le n/2}\sum_{h=j}^{n/2}(-1)^j\binom{n-j}{j}
\binom{n-2j}{2(h-j)}f_k(2(h-j),0)w^{2h}x^n \\
& = & \sum_{j\ge 0}\sum_{n\ge
2j}\sum_{h=j}^{n/2}(-1)^j\binom{n-j}{j}
\binom{n-2j}{2(h-j)}f_k(2(h-j),0)w^{2h}x^n \\
& = & \sum_{j\ge 0}(-1)^j\frac{(wx)^{2j}}{j!}\sum_{n\ge 2j}(n-j)!
\sum_{h=j}^{n/2}\binom{n-2j}{2(h-j)}f_k(2(h-j),0)\frac{w^{2(h-j)}}{(n-2j)!}
x^{n-2j} \ .\end{aligned}$$ We shift summation indices $n'=n-2j$ and $h'=h-j$ and derive for the rhs the following expression $$\begin{aligned}
& = & \sum_{j\ge 0}(-1)^j\frac{(wx)^{2j}}{j!}\sum_{n'\ge 0}(n'+j)!
\sum_{h=j}^{n/2}\binom{n'}{2(h-j)}f_k(2(h-j),0)\frac{w^{2(h-j)}}{n'!}
x^{n-2j} \\
& = & \sum_{j\ge 0}(-1)^j\frac{(wx)^{2j}}{j!}\sum_{n'\ge 0}(n'+j)!
\left\{
\sum_{h'=0}^{n/2-j=n'/2}\binom{n'}{2h'}f_k(2h',0)w^{2h'}\right\}
\frac{x^{n'}}{n'!}\end{aligned}$$ The idea is now to interpret the term $\sum_{h'=0}^{n'/2}\binom{n'}
{2h'}f_k(2h',0)w^{2h'}\frac{x^n}{n!}$ as a product of the two power series $e^x$ and $\sum_{n\ge
0}f_k(2n,0)\frac{(wx)^{2n}}{(2n)!}$: $$\begin{aligned}
\sum_{\ell\ge 0}\frac{x^\ell}{\ell!}\sum_{n\ge
0}f_k(2n,0)\frac{(wx)^{2n}} {(2n)!} & = & \sum_{n'\ge 0}
\sum_{2n+\ell=n'} \left\{\frac{1}{\ell!}\frac{1}{(2n)!}
f_k(2n,0)w^{2n}\right\}x^{n'}\\
& = & \sum_{n'\ge 0}\left\{\sum_{n=0}^{n'/2}
\binom{n'}{2n}f_k(2n,0)w^{2n}\right\}\frac{x^{n'}}{n'!} \ .\end{aligned}$$ We set $\eta_{n'}=\left\{\sum_{h'=0}^{n'/2}\binom{n'}{2h'}f_k(2h',0)
w^{2h'}\right\} $. By assumption we have $\vert x\vert<\rho_k(w)$ and we next derive, using the Laplace transformation and interchanging integration and summation $$\label{E:on1}
\sum_{n'\ge 0}(n'+j)!\eta_n \frac{x^{n'}}{n'!} = \int_{0}^{\infty}
\sum_{n'\ge 0}\eta_{n'} \frac{(xt)^{n'}}{n'!} t^je^{-t} dt \ .$$ Since $\vert x\vert<\rho_k(w)$ the above transformation is valid and using $$\sum_{n'\ge 0}\left\{\sum_{n=0}^{n'/2}
\binom{n'}{2n}f_k(2n,0)w^{2n}\right\}\frac{x^{n'}}{n'!}=
\sum_{\ell\ge 0}\frac{x^\ell}{\ell!}\sum_{n\ge
0}f_k(2n,0)\frac{(wx)^{2n}} {(2n)!}$$ we accordingly obtain $$\begin{aligned}
\sum_{n'\ge 0}\eta_{n'} \frac{(xt)^{n'}}{n'!} t^je^{-t} dt & = &
\int_{0}^{\infty}e^{tx}\sum_{n\ge 0}f_k(2n,0)\frac{(wxt)^{2n}}{(2n)!}
t^je^{-t}dt \ .\end{aligned}$$ The next step is to substitute the term $\sum_{n'\ge
0}(n'+j)!\eta_n \frac{x^{n'}}{n'!}$ in eq. (\[E:on1\]), whence consequently $$\begin{aligned}
F_k(x,w) & = & \sum_{j\ge 0}(-1)^j\frac{(wx)^{2j}}{j!}
\int_{0}^{\infty}e^{tx}\sum_{n\ge 0}f_k(2n,0)\frac{(wxt)^{2n}}
{(2n)!}t^je^{-t}dt\\
& = & \int_{0}^{\infty}\sum_{j\ge 0}(-1)^j\frac{(wx)^{2j}}{j!}
e^{tz}\sum_{n\ge 0}f_k(2n,0)\frac{(wxt)^{2n}} {(2n)!}t^je^{-t}dt \
.\end{aligned}$$ The summation over the index $j$ is just an exponential function and we derive $$\begin{aligned}
& = &\int_{0}^{\infty} e^{-(w^2x^2-x+1)t} \sum_{n\ge
0}f_k(2n,0)\frac{(wxt)^{2n}}
{(2n)!}dt \\
& = &\int_{0}^{\infty} e^{-(w^2x^2-x+1)t} \sum_{n\ge
0}f_k(2n,0)\frac{1}{(2n)!} \left(\frac{wx}{w^2x^2-x+1}\right)^{2n}
((w^2x^2-x+1)t)^{2n}dt\end{aligned}$$ We proceed by transforming the integral introducing $u=(w^2x^2-x+1)t$, i.e. $dt=(w^2x^2-x+1)^{-1}du$ and accordingly arrive at $$\begin{aligned}
F_k(x,w) & = & \sum_{n\ge 0}f_k(2n,0)\frac{1}{(2n)!}
\left(\frac{wx}{w^2x^2-x+1}\right)^{2n} \int_{0}^{\infty}
e^{-(w^2x^2-x+1)t} ((w^2x^2-x+1)t)^{2n}
dt \\
& = & \sum_{n\ge 0}f_k(2n,0)\frac{1}{(2n)!}
\left(\frac{wx}{w^2x^2-x+1}\right)^{2n}
\frac{1}{w^2x^2-x+1} (2n)! \\
& = & \frac{1}{w^2x^2-x+1} \sum_{n\ge 0}f_k(2n,0)
\left(\frac{wx}{w^2x^2-x+1}\right)^{2n} \ ,\end{aligned}$$ In particular for $w=1$ $$\begin{aligned}
\label{E:L1}
\sum_{n \ge 0}{\sf S}_{k}(n)x^n=\frac{1}{x^2-x+1}\sum_{n \ge
0}f_{k}(2n,0)\left(\frac{x}{x^2-x+1}\right)^{2n}\end{aligned}$$ holds for any $x\in\mathbb{R}$, satisfying $\vert
x\vert<\rho_{k}(1)$, and where $\rho_{k}(1)$ is the radius of convergence of the power series $\sum_{n \ge 0}{\sf S}_{k}(n)z^n$ over $\mathbb{C}$, that is eq. (\[E:L1\]) holds for $x\in
]-\rho_{k}(1),\rho_{k}(1)[$ . From complex analysis we know that any two functions that are analytic at $0$ and coincide on an open interval which includes $0$ are identical. Therefore eq. (\[E:L1\]) holds for $z\in\mathbb{C}$, $\vert
z\vert<\rho_k(1)$, and the proof of the lemma is complete. $\square$
[**Acknowledgments.**]{} We are grateful to Prof. Jason Gao for helpful discussions. This work was supported by the 973 Project, the PCSIRT Project of the Ministry of Education, the Ministry of Science and Technology, and the National Science Foundation of China.
[^1]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the lattice of finite-index extensions of a given finitely generated subgroup $H$ of a free group $F$. This lattice is finite and we give a combinatorial characterization of its greatest element, which is the commensurator of $H$. This characterization leads to a fast algorithm to compute the commensurator, which is based on a standard algorithm from automata theory. We also give a sub-exponential and super-polynomial upper bound for the number of finite-index extensions of $H$, and we give a language-theoretic characterization of the lattice of finite-index subgroups of $H$. Finally, we give a polynomial time algorithm to compute the malnormal closure of $H$.'
author:
- |
Pedro Silva,\
[^1]\
\
Pascal Weil,\
and [^2]
title: 'On finite-index extensions of subgroups of free groups[^3]'
---
**Keywords**: free groups, subgroups, finite-index extensions
**MSC**: 20E05
This paper is part of the study of the lattice of finitely generated subgroups of a free group of finite rank $F$. Like most of the recent work on this topic, our paper makes crucial use of the graphical representation of the subgroups of $F$ introduced in the seminal papers of Serre (1977 [@Serre]) and Stallings (1983 [@Stallings]). This representation not only makes it easier to form an intuition and to prove properties of subgroups of $F$, but it also provides a convenient framework to efficiently solve algorithmic problems and compute invariants concerning these subgroups.
The particular object of study in this paper is the lattice of extensions of a given finitely generated subgroup $H$ of $F$, and more specifically the sublattice of finite-index extensions of $H$. In this paper, all groups are subgroups of a fixed free group, and the notion of extension must be understood in this context.
It is elementary to verify that $H$ has only finitely many finite-index extensions, and it is known that if $K$ and $L$ are finite-index extensions of $H$, then the subgroup they generate, namely $\langle K,L\rangle$, has finite index over $H$ as well (Greenberg’s theorem, see [@Stallings]). Therefore $H$ has a maximum finite-index extension $H_{\fim}$, which is effectively constructible, and the finite-index extensions of $H$ form a full convex sublattice of the lattice of subgroups of $F$.
This paper contains a detailed discussion of the lattice of finite-index extensions of $H$. Our main contributions are the following.
We show that the maximum finite-index extension $H_\fim$ of $H$ is the commensurator of $H$, and we give a combinatorial (graph-theoretic) characterization of $H_\fim$. This characterization leads to efficient algorithms to compute all finite-index extensions of $H$, and to compute $H_\fim$ – the latter in time $\O(n\log n)$. We also give a rather tight upper bound on the number of finite-index extensions of $H$: there are at most $\O(\sqrt n\ n^{\frac12\log_2n})$ such extensions, where $n$ is the number of vertices in the graphical representation of $H$. Note that this upper bound is sub-exponential but super-polynomial.
The consideration of the subgroups of the form $H_\fim$, which have no proper finite-index extensions, leads us to the dual study of the lattice of finite-index subgroups of a given subgroup, and we give a combinatorial (language-theoretic) characterization of each such lattice.
Finally, we use our better understanding of the lattice of extensions of a subgroup of $F$, to give a polynomial time algorithm to compute the malnormal closure of a given subgroup.
As we already indicated, we use in a fundamental way the graphical representation of finitely generated subgroups of $F$, including a detailed study of the different steps of the computation of this representation (given a set of generators for the subgroup $H$), whose study was at the heart of an earlier paper by the authors [@SW08]. It is particularly interesting to see that language-theoretic results and arguments play an important role in this paper: that is, we sometimes consider the graphical representation of a subgroup not just as an edge-labeled graph, but as a finite state automaton. Such considerations are present in almost all the results of this paper, but they become crucial at a rather unexpected juncture: the design of an efficient algorithm to compute the maximal finite-index extension $H_\fim$ of $H$. Indeed, the very low complexity we achieve is due to the possibility of using a standard automata-theoretic algorithm, namely the computation of the minimal automaton of a regular language.
Section \[sec: background\] summarizes a number of well-known facts about free groups and the representation of their finitely generated subgroups, which will be used freely in the sequel (see [@Stallings; @Weilsurvey; @KM; @MVW; @SW08] for more details). Section \[sec: finite index\] is the heart of the paper: it starts with a technical study of the different steps of the algorithm to compute the graphical representation of a given subgroup, and a description of those steps which preserve finite-index (Section \[preserve fi\]). These technical results are then used to characterize the maximal finite-index extension $H_\fim$ (Section \[lattice fi\]), to relate the computation of $H_\fim$ and the minimization of certain finite-state automata (Section \[sec: computing\]), to evaluate the maximal number of finite-index extensions of a given subgroup (Section \[sec: counting\]), and to describe an invariant of the lattice of finite-index subgroups of a given subgroup (Section \[sec: FIS\]).
Finally, we apply the same ideas in Section \[sec: malnormal\], to study the malnormal closure of a subgroup, and to show that it can be computed in polynomial time.
Subgroups of free groups and Stallings graphs {#sec:
background}
=============================================
Let $F$ be a finitely generated free group and let $A = \{a_1, \ldots,
a_r\}$ be a fixed basis of $F$. Let $\bar A = \{\bar a_1, \ldots,
\bar a_r\}$ be a disjoint copy of $A$ and let $\tilde A = A \cup \bar
A$: as usual, we extend the map $a\mapsto \bar a$ from the set $A$ to all words by letting $\bar{\bar a} = a$ if $a\in A$ and $\overline{ua}
= \bar a\bar u$ if $a\in \tilde A$ and $u\in \tilde A^*$. As usual again, the elements of $F$ are identified with the *reduced words* over the alphabet $\tilde A$, that is, the words that do not contain a sequence of the form $a\bar a$ ($a\in \tilde A$). If $u\in
\tilde A^*$ is an arbitrary word, we denote by $\red(u)$ the corresponding reduced word, that is, the word obtained from $u$ by repeatedly deleting all sequences of the form $a\bar a$ ($a\in \tilde
A$).
A reduced word $u \in F$ is *cyclically reduced* if $u$ cannot be written as $u = av\bar a$ with $a\in \tilde A$ and $v \in F$. Every reduced word $u$ can be factored in a unique way in the form $u
= xy\bar x$, with $y$ cyclically reduced.
If $H$ is a subgroup of $F$, an *extension* of $H$ is any subgroup $G$ containing $H$ and we write $H \le G$. If $H$ is finitely generated, we also write $H \lefg G$. If $H$ has finite index in $G$, we say that $G$ is a *finite-index extension* of $H$ and we write $H \lefi G$. Finally, we write $H \leff G$ if $H$ is a free factor of $G$.
The graphical representation of a subgroup
------------------------------------------
It is well known (since Serre’s and Stalling’s fundamental work [@Serre; @Stallings]) that every finitely generated subgroup $H
\lefg F$ admits a unique graphical representation of the form $\calA(H) = (\Gamma(H),1)$, where $\Gamma(H)$ is a finite directed graph with $A$-labeled edges and 1 is a designated vertex of $\Gamma(H)$, subject to the combinatorial conditions below. Here, a graph is a pair $(V,E)$ where $V$ is the set of *vertices* and $E \subseteq V \times A \times V$ is the set of *edges*; the *in-degree* (resp. *out-degree*) of a vertex $v\in V$ is the number of edges in $E$ of the form $(v',a,v)$ (resp. $(v,a,v')$); and the *degree* of $v$ is the sum of its in- and out-degree. Every pair $\calA(H)$ satisfies the following:
- the (underlying undirected) graph is connected;
- for each $a\in A$, every vertex is the source (resp. the target) of at most one $a$-labeled edge;
- and every vertex, except possibly 1, has degree at least 2.
Moreover, every pair $(\Gamma,1)$ with these properties is said to be *admissible*, and it is the representation of a finitely generated subgroup of $F$. In addition, given a finite set of generators of $H$, the representation of $H$ is effectively computable. We refer the reader to [@Stallings; @KM; @MVW; @Weilsurvey; @SW08] for some of the literature on this construction and its many applications, and to Section \[sec: fi extensions\] below on the construction of $\calA(H)$.
We sometimes like to view the $A$-labeled graph $\Gamma(H)$ as a transition system over the alphabet $\tilde A$: if $p,q$ are vertices of $\Gamma(H)$, $a\in A$ and $(p,a,q)$ is an edge of $\Gamma(H)$, we say that $a$ labels a path from $p$ to $q$ and $\bar a$ labels a path from $q$ to $p$, written $p \mapright a q$ and $q \mapright{\bar a}
p$. If $u \in \tilde A^*$ is a word (reduced or not) and $u = va$ ($a\in \tilde A$), we say that $p \mapright u q$ (*$u$ labels a path from $p$ to $q$*) if $p \mapright v p' \mapright a q$ for some vertex $p'$. In particular, a reduced word is in $H$ if and only if it labels a loop at vertex 1. Moreover, if we have a path $p
\mapright u q$, then we also have a path $p \mapright{\red(u)} q$.
If $H \lefg G \lefg F$, then there is a *homomorphism* from $\calA(H)$ into $\calA(G)$, that is, a map $\phi$ from the vertex set of $\Gamma(H)$ to the vertex set of $\Gamma(G)$ such that
- $\phi(1) = 1$ and
- if $p \mapright a q$ in $\Gamma(H)$ ($p,q$ vertices, $a\in \tilde
A$), then $\phi(p) \mapright a \phi(q)$ in $\Gamma(G)$.
It is not difficult to verify that this morphism, if it exists, is unique, and we denote it by $\phi_{H}^{G}$. It is well known (see [@Serre; @Stallings; @KM; @MVW]) that if $\phi_H^G$ is one-to-one, then $H \leff G$.
Finally, we say that the homomorphism $\phi_H^G$ is a *cover*, if it satisfies
- if $p,q$ are vertices of $\Gamma(H)$, $a\in \tilde A$ and $\phi_H^G(p) \mapright a \phi_H^G(q)$ in $\Gamma(G)$, then $p
\mapright a q'$ in $\Gamma(H)$ for some vertex $q'$ such that $\phi_H^G(q') = \phi_H^G(q)$. In that case, all sets of the form $(\phi_H^G)\inv(q)$ ($q$ a vertex of $\Gamma(G)$) have the same cardinality.
Covers have the following property, which we will use freely in the sequel.
If $\phi\colon \calA(H) \to \calA(G)$ is a cover, $p$ is a vertex of $\calA(H)$ and $u \in F$ labels a loop of $\calA(G)$ at $\phi(p)$, then $\red(u^m)$ labels a loop of $\calA(H)$ at $p$ for some integer $m > 1$.
Covers, cyclically reduced subgroups and finite-index extensions
----------------------------------------------------------------
Let us say that $H$ is *cyclically reduced* (with respect to the basis $A$) if every vertex of $\Gamma(H)$ has degree at least equal to 2. If $H$ is not cyclically reduced, then the designated vertex 1 of $\Gamma(H)$ has degree 1 and $\Gamma(H)$ consists of two parts: $\tail(\Gamma(H))$, which contains the designated vertex 1 and all degree 2 vertices that can be connected to vertex 1 through other degree 2 vertices; and the rest of $\Gamma(H)$, which is called the *core* of $\Gamma(H)$, written $\cc(\Gamma(H))$. We let $\tt_H(1)$ be the shortest word which labels a path from 1 to a vertex in $\cc(\Gamma(H))$ and let $\tau_H(1)$ be the vertex of $\cc(\Gamma(H))$ thus reached (if $H$ is cyclically reduced, then $\tt_H(1)$ is the empty word and $\tau_H(1) = 1$). We write $\tt(1)$ and $\tau(1)$ if the subgroup $H$ is clear from the context.
The tail and the core of $\Gamma(H)$ have intrinsic characterizations. The characterization of the core is well-known (see [@Stallings Exercise 7.3(a)]) and that of the tail is an elementary consequence.
\[loop and tail\] Let $H \lefg F$ and let $p$ be a vertex of $\Gamma(H)$. Then $p$ is a vertex of $\cc(\Gamma(H))$ if and only if some cyclically reduced word $u$ labels a path from $p$ to $p$.
By definition, if $H$ is not cyclically reduced, then $\tail(\Gamma(H))$ consists of a single path from vertex 1 to vertex $\tt(1)$ (excluding the latter vertex): it is therefore elementary to verify that no non-empty cyclically reduced word labels a loop at a vertex in $\tail(\Gamma(H))$.
Let now $p$ be a vertex in $\cc(\Gamma(H))$: then $p$ has degree at least 2, and if it has degree exactly 2, then neither of the two edges adjacent to it leads to a vertex in $\tail(\Gamma(H))$. Therefore, one can find distinct letters $a,a'\in \tilde A$ such that $p
\mapright a q$ and $p \mapright{a'} q'$, with $q$ and $q'$ in $\cc(\Gamma(H))$ as well. Iterating this reasoning, one can show that there exist arbitrarily long paths within $\cc(\Gamma(H))$, starting from $p$ and labeled by reduced words of the form $au$ and $a'u'$. Since $\Gamma(H)$ is finite, vertices are repeated along these paths, and we consider the earliest such repetition after the initial $p$. If $p$ itself is the first repeated vertex along the path labeled $au$, we have a loop $p \mapright{au} p$ such that $au$ is cyclically reduced, and we are done. The situation is similar if $p$ is the first repeated vertex along the path labeled $a'u'$. Otherwise, let $r$ and $r'$ be the first repeated vertices along the two paths. Then $r \ne p$, $r' \ne p$, and $\cc(\Gamma(H))$ has paths of the form $p
\mapright{au} r$, $p \mapright{a'u'} r'$, $r \mapright v r$ and $r'
\mapright{v'} r'$ such that $auv\bar u\bar a, a'u'v'\bar u'\bar a'$ are reduced. Then the word $auv\bar u\bar aa'u'v'\bar u'\bar a'$ is cyclically reduced, and it labels a loop at $p$ in $\Gamma(H)$. This concludes the proof.
\[tail only\] Let $H \lefg F$. Then $\tt(1)$ is the maximum common prefix of the non-trivial elements of $H$.
Since every non-trivial element of $H$ is the label of a loop at 1 in $\Gamma(H)$, it is clear that $\tt(1)$ is a common prefix to all these words.
By Remark \[loop and tail\], there exists a cyclically reduced word $u$ labeling a loop at $\tau(1)$. Then both $\tt(1)u\overline{\tt(1)}$ and $\tt(1)\bar u\overline{\tt(1)}$ are reduced words in $H$, and their maximum common prefix is $\tt(1)$. This concludes the proof.
We can now state the following extension of the classical characterization of finite-index extensions of cyclically reduced subgroups in terms of covers.
\[prop: fi cover\] Let $H\lefg G \lefg F$. Then $\tt_G(1)$ is a prefix of $\tt(H)$. Moreover, $H \lefi G$ if and only if $\tt_H(1) = \tt_G(1)$ and the restriction of $\phi^G_H$ is a cover from $(\cc(\Gamma(H)),
\tau_H(1))$ onto $(\cc(\Gamma(G)), \tau_G(1))$. If that is the case, the index of $H$ in $G$ is the common cardinality of the subsets ${\phi^G_H}\inv(q)$ ($q$ a vertex of $\cc(\Gamma(G))$).
Let $\phi = \phi^G_H$. If $u$ is cyclically reduced and labels a loop at a vertex $p$ of $\cc(\Gamma(H))$, then $u$ labels a loop at $\phi(p)$ in $\Gamma(G)$, and that vertex is in the core of $\Gamma(G)$ by Fact \[loop and tail\]. It follows that $\tt_H(1)$ labels a path from the origin in $\Gamma(G)$ to a vertex in $\cc(\Gamma(G))$. In particular, $\tt_G(1)$ is a prefix of $\tt_H(1)$.
If $\tt_G(1)$ is a proper prefix of $\tt_H(1)$, we have $\tt_H(1) =
\tt_G(1)at$ for some $a\in \tilde A$ and $t\in F$. Since $\tau_G(1)$ is in $\cc(\Gamma(G))$, there exist a cyclically reduced word of the form $bu$, with first letter $b\ne a$, which labels a loop at $\tau_G(1)$ in $\cc(\Gamma(G))$. Then the words $\tt_G(1)(bu)^n\overline{\tt_G(1)}$ are all reduced, and the cosets $H\tt_G(1)(bu)^n\overline{\tt_G(1)}$ are all in $G$. Moreover, these cosets are pairwise disjoint since $H$ contains no reduced word of the form $\tt_G(1)(bu)^d\overline{\tt_G(1)}$, $d\ne 0$. Thus, if $H\lefi
G$, then $\tt_H(1) = \tt_G(1)$. It follows immediately that $\phi$ maps core vertices to core vertices and tail vertices to tail vertices.
Let us now assume that $\Gamma(H)$ and $\Gamma(G)$ have the same tails, and let us denote by $\tt(1)$ the word $\tt_H(1) = \tt_G(1)$. Then $H' = \overline{\tt(1)}H\tt(1)$ and $G' =
\overline{\tt(1)}G\tt(1)$ are cyclically reduced, and $H\lefi G$ if and only if $H' \lefi G'$. Thus we may now assume that $G$ and $H$ are cyclically reduced. If $\phi$ is not a cover, there exists a vertex $p$ of $\Gamma(H)$ such that $\Gamma(G)$ has a loop at $\phi(p)$ labeled by a cyclically reduced word $bu$ ($b\in \tilde A$) and $\Gamma(H)$ has no $b$-labeled edge out of $p$. Let $v$ label a path from 1 to $p$ in $\Gamma(H)$ (and hence in $\Gamma(G)$). By the same reasoning as above, the cosets $Hv(bu)^n\bar v$ are pairwise distinct, and contained in $G$. Thus, if $H \lefi G$, then $\phi$ is a cover.
The converse is verified as follows: if $\phi$ is a cover, let $u_1,\ldots,u_d$ be reduced words labeling paths in $\Gamma(H)$ from $1$ to the elements $1 = p_1,\ldots,p_d$ of $\phi\inv(1)$. If $g\in
G$, then $g$ labels a loop at 1 in $\Gamma(G)$, and since $\phi$ is a cover, $g$ labels a path in $\Gamma(H)$ from 1 to $p_i$ for some $i$. Therefore $g\in Hu_i$: thus $G$ is the union of finitely many $H$-cosets.
The extensions, and the finite-index subgroups of a cyclically reduced subgroup are cyclically reduced as well.
Finite-index extensions of a subgroup $H$ {#sec: finite
index}
=========================================
It follows from the characterization of finite-index extensions by covers, that if $H \lefi G$, then $\phi_H^G$ is onto[^4]. Therefore $H$ has only a finite number of finite-index extensions, and that number can be bounded above by the number of binary relations on $\Gamma(H)$: if that graph has $n$ vertices, then $H$ has at most $2^{n^2}$ finite-index extensions. We give a better upper bound in Section \[sec: counting\].
Moreover, the collection of finite-index extensions of $H$ is effectively computable. In addition, it is elementary to use these graphical representations to show that the join of two finite-index extensions of $H$ is again a finite-index extension (see Stalling’s proof of Greenberg’s theorem [@Stallings]). It follows that if $H
\lefg F$, then $H$ admits an effectively computable maximum finite-index extension $H_\fim$. The results of Section \[sec: computing\] below yield an efficient algorithm to compute the set of finite-index extensions of a given subgroup $H$, and its maximum finite-index extension $H_\fim$.
\[remark index\] We just observed that every finitely generated subgroup of $F$ has a finite number of finite-index extensions. However, it usually has infinitely many finite-index subgroups. More precisely, every non-trivial subgroup $H \lefg F$ admits a finite-index subgroup of index $r$ for each $r \ge 1$. Indeed, let $\calA(H) =
(\Gamma(H),1)$ and let $Q$ be the vertex set of $\Gamma(H)$. Define $\Gamma_r$ be the $A$-labeled graph with vertex set $Q
\times \{1,\ldots,r\}$ and with the following edge set: for each edge $(p,a,q)$ of $\Gamma(H)$, there is an edge $((p,i),a,(q,i+1))$ for each $1\le i < r$ and an edge $((p,r),a,(q,1))$. Then $\Gamma_r$ is an admissible graph, the map $\pi\colon (p,i) \mapsto p$ defines a cover from $\Gamma_r$ to $\Gamma(H)$, and if $H_r$ is the subgroup represented by $(\Gamma_r, (1,1))$, then $H_r$ has index $r$ in $H$.
i-steps and finite-index extensions {#sec: fi extensions}
-----------------------------------
Let $H \le G$ be finitely generated subgroups of $F$ and let $g_1,
\ldots, g_n$ be reduced words such that $G = \langle H, g_1, \ldots,
g_n\rangle$. Let $G_0 = H$ and let $G_i = \langle G_{i-1},
g_i\rangle$ ($1 \le i \le n$). We may of course assume that $g_i
\not\in G_{i-1}$, so $G_{i-1} \ne G_i$.
Then $\calA(G_i)$ is obtained from $\calA(G_{i-1})$ by, first, adding sufficiently many new vertices and edges to create a new path from vertex 1 to itself, labeled by $g_i$; and second, by *reducing*[^5] the resulting graph, that is, repeatedly identifying vertices $p$ and $p'$ such that $q \mapright a p$ and $q \mapright a p'$ for some vertex $q$ and some letter $a\in \tilde A$, see for instance [@Stallings; @Weilsurvey; @KM; @SW08]. Depending on the length of prefixes of $g_i$ and $\bar g_i$ that can be read from vertex $1$ in $\Gamma(G_{i-1})$, this procedure amounts to one of the two following moves:
- a *reduced expansion*, or *re-step* (we write $\calA(G_{i-1}) \restep^{(p,w,q)} \calA(G_i)$), that is, we add a new path labeled by a factor $w$ of $g_i$, from some vertex $p$ to some vertex $q$ of $\Gamma(G_{i-1})$ in such a way that the resulting graph is admissible (needs no reduction);
- or an *i-step* (we write $\calA(G_{i-1}) \istep^{p=q}
\calA(G_i)$), that is, we identify a pair of vertices $(p,q)$ of $\Gamma(G_{i-1})$, and we reduce the resulting graph.
Let us comment on these steps, with reference to Stallings’s algorithm [@Stallings]. If $H = \langle
g_1,\ldots,g_n\rangle$, Stallings produces $\calA(H)$ by reducing (folding) a bouquet of $n$ circles, labeled $g_1,\ldots, g_n$ respectively. For our purpose, we decompose this operation in $n$ steps, adding one generator at a time and producing successively the $\calA(\langle h_1,\ldots,h_i\rangle)$ ($1\le i\le n$). Each of these steps is either an re-step or an i-step.
We refer the readers to [@SW08 Section 2] for a detailed analysis of these moves and we record the following observation.
\[first lemma i-steps\] Let $G,H \lefg F$. If $H \lefi G$, then only i-steps are involved in the transformation from $\Gamma(H)$ to $\Gamma(G)$.
Let $H = G_0 \le G_1 \le \ldots \le G_n = G$ be as in the above discussion. Note that $H \lefi G$ if and only if $G_{i-1} \lefi G_i$ for each $1 \le i \le n$. If $\calA(G_{i-1}) \restep \calA(G_i)$, then the homomorphism $\phi_{G_{i-1}}^{G_i}$ is one-to-one, so $G_{i-1} \leff G_i$ and in particular, $G_i$ is not a finite-index extension of $G_{i-1}$.
Which i-steps yield finite-index extensions? {#preserve fi}
--------------------------------------------
If $p$ is a vertex of $\cc(\Gamma(H))$, we let $\widetilde L_p(H)$ be the language accepted by $\cc(\Gamma(H))$, seen as a finite state automaton with initial state $p$ and all states final: that is, the set of (possibly non-reduced) words in $\tilde A^*$ that label a path in $\cc(\Gamma(H))$ starting at vertex $p$. Let then $L_p(H)$ be the set of reduced words in $\widetilde L_p(H)$ — which is also the set of all $\red(u)$ ($u \in \widetilde L_p(H)$), and also the set of prefixes of words in the subgroup represented by the pair $(\cc(\Gamma(H)),p)$. Let us first record the following elementary remark.
\[fact: Lp Lq\] If $p \mapright u q$ is a path in $\cc(\Gamma(H))$, then $L_q(H) =
\{\red(\bar ux) \mid x \in L_p(H)\}$.
We now refine the result of Lemma \[first lemma i-steps\].
\[cns finite index\] Let $H\lefg F$ and let $p,q$ be distinct vertices in $\Gamma(H)$. Let $G$ be the subgroup of $F$ such that $\calA(H) \istep^{p=q}
\calA(G)$. Then $H \lefi G$ if and only if $p,q$ are in $\cc(\Gamma(H))$ and $L_p(H) = L_q(H)$, if and only if $p,q$ are in $\cc(\Gamma(H))$ and $\widetilde L_p(H) = \widetilde L_q(H)$.
Let us first assume that $H \lefi G$ and let $\phi = \phi^G_H$. By Proposition \[prop: fi cover\], $\phi$ is a bijection from $\tail(H)$ onto $\tail(G)$ and, since $\phi(p) = \phi(q)$, the vertices $p$ and $q$ must both be in $\cc(\Gamma(H))$.
**If $p,q \in \cc(\Gamma(H))$ and $\widetilde L_p(H) \ne
\widetilde L_q(H)$**, we consider (without loss of generality) a word $u \in \widetilde L_p(H) \setminus \widetilde L_q(H)$, with minimum length, say $u = va$ with $a\in \tilde A$. By definition, there exist paths $p \mapright v p' \mapright a p''$ and $q \mapright v q'$, but no path $q' \mapright a q''$ in $\cc(\Gamma(H))$. Observe that $\calA(H) \istep^{p' = q'} \calA(G)$. If there is a path $q'
\mapright a q''$ in $\Gamma(H)$, then we also have $\calA(H)
\istep^{p'' = q''} \calA(G)$ and since $p'' \in \cc(\Gamma(H))$ and $q'' \in \tail(\Gamma(H))$, we conclude to a contradiction by Proposition \[prop: fi cover\]. We now assume that there is no path $q' \mapright a q''$ in $\Gamma(H)$.
We claim that there exists $w\in F$ such that $aw$ is cyclically reduced and $p' \mapright{aw} q'$ in $\cc(\Gamma(H))$. Let indeed $p'' \mapright z q'$ be a path of minimal length in $\cc(\Gamma(H))$ (there exists one by connectedness). Since $p'' \in \cc(\Gamma(H))$, there exists a path $p'' \mapright b r$ for some $b \in \tilde A$, $b\ne \bar a$, and as in the proof of Remark \[loop and tail\], there exists a reduced word of the form $bt$ labeling a loop at $p''$. Let $w = \red(btz)$: then we have a path $p'' \mapright w q'$. By minimality of the length of $z$, $\bar t\,\bar b$ is not a prefix of $z$, so $w$ starts with letter $b$, and hence $aw$ is reduced. In fact, $aw$ is cyclically reduced since there is no path $q' \mapright
a q''$.
Let $1 \mapright t q'$ be a path in $\Gamma(H)$. Then $\red(taw\bar
t) \in G$, and hence there exists $m > 1$ such that $\red(t(aw)^m\bar t) \in H$. Again, since there is no path $q'
\mapright a q''$, the word $ta$ is reduced. By replacing $m$ by a sufficiently large multiple, we find that $ta$ is a prefix of $\red(t(aw)^m\bar t)$, and hence that $ta$ labels a path from 1 in $\Gamma(H)$: this contradicts the absence of a path $q' \mapright a q''$.
Thus we have proved that, if $H \lefi G$, then $p,q \in
\cc(\Gamma(H))$ and $\widetilde L_p(H) = \widetilde L_q(H)$. The latter condition immediately implies that $L_p(H) = L_q(H)$. **We now assume that $p,q \in \cc(\Gamma(H))$ and $L_p(H) =
L_q(H)$, and we show that $H \lefi G$**. We first establish a technical fact.
\[special fact\] Let $r_i \mapright{z_i} s_{i+1}$ ($z_i \in F$, $0 \le i \le k$) be paths in $\cc(\Gamma(H))$, such that $r_i, s_i \in \{p,q\}$ for each $1\le i \le k$. Then there exists a path $r_0
\mapright{\red(z_0\cdots z_k)} t$ in $\cc(\Gamma(H))$.
The proof is by induction on $k$, and is trivial for $k = 0$. If $k >
0$, then there is a path $r_1 \mapright{\red(z_1\cdots z_k)} t$ in $\cc(\Gamma(H))$. Since $L_p(H) = L_q(H)$, there is also a path $s_1
\mapright{\red(z_1\cdots z_k)} t'$ for some $t' \in \cc(\Gamma(H))$, and therefore a path $r_0 \mapright{\red(z_0\cdots z_k)} t'$ as required.
We want to show that $G$ has finitely many $H$-cosets. Let $u \in G$: then $u$ labels a loop at 1 in $\Gamma(G)$. Let $\calB$ be the automaton obtained from $\Gamma(H)$ by identifying vertices $p$ and $q$, but without performing any reduction. Then $\Gamma(G)$ is the result of the reduction of $\calB$. In particular (say, in view of [@SW08 Fact 1.4]), $u = \red(v)$ for some word $v \in \tilde
A^*$ labeling a loop at 1 in $\calB$. By definition of $\calB$, the word $v$ factors as $v = v_0\cdots v_k$, in such a way that $\Gamma(H)$ has paths of the form $1 \mapright{v_0} s_1$, $r_i \mapright{v_i} s_{i+1}$ ($1
\le i < k$) and $r_k \mapright{v_k} 1$, and the vertices $r_1, s_1,
\cdots, r_k, s_k$ are all equal to $p$ or $q$. As observed in Section \[sec: background\], $\Gamma(H)$ also has paths $$1 \mapright{\red(v_0)} s_1,\ r_i \mapright{\red(v_i)} s_{i+1}\ (1
\le i < k)\textrm{ and }r_k \mapright{\red(v_k)} 1.$$ In particular, we have $\red(v_0) = \tt(1)w_0$ and $\red(v_k) =
w_k\bar\tt(1)$ for some $w_0, w_k \in F$, and there are paths $1
\mapright{\tt(1)} \tau(1) \mapright{w_0} s_1$ and $r_k \mapright{w_k}
\tau(1) \mapright{\bar\tt(1)} 1$. Note that the paths $\tau(1)
\mapright{w_0} s_1$, $r_i \mapright{\red(v_i)} s_{i+1}$ ($1 \le i <
k$) and $r_k \mapright{w_k} \tau(1)$ are set entirely within $\cc(\Gamma(H))$, since no reduced word-labeled path between vertices in $\cc(\Gamma(H))$ can visit a vertex in $\tail(\Gamma(H))$.
By Lemma \[special fact\], there exists a path $\tau(1)
\mapright{\red(w_0v_1\cdots v_{k-1}w_k)} t$ for some vertex $t$ in $\cc(\Gamma(H))$. Let $h$ be a shortest-length word such that $t
\mapright h \tau(1)$ in $\cc(\Gamma(H))$. Then $z =
\tt(1)\red(w_0v_1\cdots v_{k-1}w_k)\ h\ \bar\tt(1)$ labels a loop at vertex 1 in $\Gamma(H)$, so $\red(z) \in H$. By construction, we have $u = \red(\tt(1)w_0v_1\cdots v_{k-1}w_k\bar\tt(1))$, so $u \in
H\red(\tt(1)\ \bar h\ \bar\tt(1))$. Since $h$ was chosen to be a geodesic in $\cc(\Gamma(H))$, it can take only finitely many values, and this completes the proof that $G$ has finitely many $H$-cosets.
We note the following consequence of this proof.
\[fact tilde or not\] Let $H \lefg F$ and let $p,q \in \cc(\Gamma(H))$. Then $L_p(H) =
L_q(H)$ if and only if $\widetilde L_p(H) = \widetilde L_q(H)$.
The lattice of finite-index extensions of $H$ {#lattice fi}
---------------------------------------------
We further refine Proposition \[cns finite index\] as follows: we consider an extension $H\lefi G$ and a pair $(r,s)$ of vertices of $\Gamma(H)$, whose identification yields a finite-index extension of $H$. Then we show that identifying the vertices of $\Gamma(G)$ corresponding to $r$ and $s$, also yields a finite-index extension of $G$.
\[preserve Lr\] Let $H \lefi G \lefg F$ and let $\phi = \phi^G_H$. Let $p$ be a vertex of $\cc(\Gamma(H))$. Then $\widetilde L_p(H) =
\widetilde L_{\phi(p)}(G)$ and $L_p(H) = L_{\phi(p)}(G)$.
If $p \mapright u r$ in $\cc(\Gamma(H))$, then the $\phi$-image of this path is a path $\phi(p) \mapright u \phi(r)$, which is entirely contained in $\cc(\Gamma(G))$ by Proposition \[prop: fi cover\]. In particular, $\widetilde L_p(H) \subseteq \widetilde L_{\phi(p)}(G)$.
Conversely, suppose that $\phi(p) \mapright u r'$ is a path in $\cc(\Gamma(G))$. Since $\phi$ is a cover from $\cc(\Gamma(H))$ onto $\cc(\Gamma(H))$, $u$ labels some path $p \mapright u r$ in $\cc(\Gamma(H))$, and hence $u \in \widetilde L_p(H)$. Thus $\widetilde L_p(H) = \widetilde L_{\phi(p)}(G)$.
Together with Proposition \[cns finite index\], Lemma \[preserve Lr\] immediately implies the following statements.
\[charact fi extensions\] Let $H \lefg F$.
1. Let $H \lefi G$ and let $\phi = \phi^G_H$. If $p,q$ are vertices of $\Gamma(H)$, $\calA(H) \istep^{p = q} \calA(K)$ and $\calA(G) \istep^{\phi(p) = \phi(q)} \calA(K')$, then $H
\lefi K$ if and only if $G \lefi K'$.
2. $H \lefi G$ if and only if $\calA(G)$ is obtained from $\calA(H)$ by identifying some pairs of vertices $(p,q)$ in $\cc(\Gamma(H))$ such that $L_p(H) = L_q(H)$, and then reducing the resulting graph.
The identification of all pairs of vertices $(p,q)$ such that $L_p(H)
= L_q(H)$ yields the minimum quotient of $\calA(H)$ and hence the maximum finite-index extension $H_\fim$ of $H$. In addition, we find that $H_\fim$ is exactly the *commensurator* of $H$ (the set $\Comm_F(H)$ of all elements $g\in F$ such that $H \cap H^g$ has finite index in both $H$ and $H^g$), a fact that can also be deduced from [@KM Lemma 8.7].
\[maximal fi extension\] Let $H\lefg F$ and let $H_\fim$ be its maximum finite-index extension.
1. $\calA(H_\fim)$ is obtained from $\calA(H)$ by identifying all pairs of vertices $p,q$ of $\cc(\Gamma(H))$ such that $L_p(H) = L_q(H)$. No reduction is necessary.
2. $H_\fim = \Comm_F(H)$.
In view of Corollary \[charact fi extensions\], $\calA(H_\fim)$ is obtained from $\calA(H)$ by identifying all pairs of vertices $p,q$ of $\cc(\Gamma(H))$ such that $L_p(H) = L_q(H)$, and then by reducing the resulting graph $\calB$. If $p \mapright a r$ and $q \mapright a s$ ($a\in \tilde A$) are paths in $\cc(\Gamma(H))$ and if $L_p(H) =
L_q(H)$, then $L_r(H) = L_s(H)$ by Remark \[fact: Lp Lq\]. Thus $\calB$ is already reduced, which concludes the proof of the first statement.
The fact that $\Comm(H)$ is a subgroup and a finite-index extension of $H$ is proved, for instance, in [@KM Prop. 8.9]. Conversely, suppose that $H \lefi G$ and $g \in G$. Since conjugation by $g$ is an automorphism of $G$, we have $H^g \lefi G$. Now the intersection of finite-index subgroups, again has finite index, so $H \cap H^g
\lefi G$ and hence $H \cap H^g \lefi H$ and $H \cap H^g \lefi H^g$. Thus $g \in \Comm_F(H)$, which concludes the proof.
Computing finite-index extensions {#sec: computing}
---------------------------------
Recall the notion of minimization of a deterministic finite-state automaton (see [@Kozen] for instance). Let $\calB = (Q,i,E,T)$ be such an automaton, over the alphabet $B$, with $Q$ the finite set of states, $i \in Q$ the initial state, $E \subseteq Q \times B \times Q$ the set of transitions and $T \subseteq Q$ the set of accepting states, and let $L$ be the language accepted by $\calB$, that is, the set of words in $B^*$ that label a path from $p$ to a state in $T$. Then the minimal automaton of $L$ is obtained by identifying the pairs of states $(p,q)$ such that the automata $(Q,p,E,T)$ and $(Q,q,E,T)$ accept the same language.
In our situation, the alphabet is $\tilde A$ and $\widetilde L_p(H)$ is the language accepted by the automaton $\calB_p$, whose states and transitions are given by $\cc(\Gamma(H))$, with initial state $p$ and all states final. Therefore Corollary \[charact fi extensions\] and Theorem \[maximal fi extension\] show that the identification of two vertices $p,q \in \cc(\Gamma(H))$ yields a finite-index extension if and only if $p$ and $q$ are identified when minimizing $\calB_{\tau(1)}$. Moreover, $\cc(\Gamma(H_\fim))$ is given by the states and transitions of the minimal automaton of $\widetilde
L_{\tau(1)}(H)$.
The classical Hopcroft algorithm (see [@Kozen]) minimizes an $n$-state automaton in time $\O(n\log n)$, so we have the following result.
\[prop compute Hfi\] Let $H \lefg F$, and let $n$ be the number of vertices of $\Gamma(H)$.
- $\cc(\Gamma(H_\fim))$ is obtained by minimizing the automaton given by the vertices and edges of $\cc(\Gamma(H))$, with all states final (the initial state does not matter in that situation).
- One can compute $\Gamma(H_\fim)$ in time $\O(n\log n)$.
- One can decide in time $\O(n\log n)$ whether identifying a given set of pairs of vertices of $\Gamma(H)$ will produce a finite-index extension of $H$.
It may be that for the particular automata at hand (over a symmetrized alphabet, with all states final), the complexity of Hopcroft’s algorithm might be better than $\O(n\log n)$, even linear. It has also been observed that in many instances, Myhill’s automata minimization algorithm exhibits a better performance than Hopcroft’s, in spite of a $\O(n^2)$ worst-case complexity. Brzozowski’s algorithm [@Brz] also performs remarkably well in practice [@ChZi].
Counting finite-index extensions {#sec: counting}
--------------------------------
Recall that, if $\Gamma$ is an $A$-labeled graph, the *product* $\Gamma \times_A \Gamma$ (also called the *fiber product*, or the *pull-back*, of two copies of $\Gamma$) is the $A$-labeled graph whose vertex set is the set of pairs $(p,q)$ of vertices of $\Gamma$ and whose edges are the triples $((p,q),a,(p',q'))$ such that $(p,a,q)$ and $(p',a,q')$ are edges of $\Gamma$. This graph is not admissible, nor even connected in general (the vertices of the form $(p,p)$ form a connected component that is isomorphic to $\Gamma$). Note that there is a $u$-labeled path in $\Gamma \times_A \Gamma$ from $(p,q)$ to $(p',q')$, if and only if $\Gamma$ has paths $p \mapright u
p'$ and $q \mapright u q'$.
If $p,q$ are vertices of $\cc(\Gamma(H))$, we let $p \sim q$ if and only if $L_p(H) = L_q(H)$.
\[LpLq covers\] Let $H \lefg F$.
- The relation $\sim$ is a union of connected components of $\cc(\Gamma(H)) \times_A \cc(\Gamma(H))$.
- Let $p,q$ be vertices of $\cc(\Gamma(H))$. Then $p \sim
q$ if and only if the first and the second component projections, from the connected component of $(p,q)$ in $\cc(\Gamma(H)) \times_A \cc(\Gamma(H))$ to $\cc(\Gamma(H))$ are both covers.
The first statement follows directly from Remark \[fact: Lp Lq\], which shows that if $p \sim q$ and there is a path $(p,q) \mapright u
(p',q')$, then $p' \sim q'$.
Let us now assume that $p \sim q$ and let us show that the first component projection is a cover from the connected component of $(p,q)$ onto $\cc(\Gamma(H))$. Let $(r,s)$ be a vertex in that connected component: then there exists $u\in F$ such that $p \mapright
u r$ and $q \mapright u s$. Let $r\mapright a r'$ ($a\in \tilde A$) be an edge in $\Gamma(H)$. Then $ua \in L_p(H)$, so $ua \in L_q(H)$, and hence (since $\Gamma(H) \times_A \Gamma(H)$ is deterministic), there exists an $a$-labeled path $s \mapright a s'$. Therefore there exists an $a$-labeled path $(r,s) \mapright a (r',s')$. Thus the first component projection is a cover. The proof concerning the second component projection is identical.
Conversely, suppose that the first and the second component projections, from the connected component of $(p,q)$ in $\cc(\Gamma(H)) \times_A \cc(\Gamma(H))$ to $\cc(\Gamma(H))$ are covers, and let $u \in L_p(H)$. Then $\cc(\Gamma(H))$ has a path $p
\mapright u r$. It is an elementary property of covers that this path can be lifted to a path in $\cc(\Gamma(H)) \times_A \cc(\Gamma(H))$, of the form $(p,q) \mapright u (r,s)$. The second component projection of that path yields a path $q \mapright u s$ in $\cc(\Gamma(H))$, and hence $u \in L_q(H)$.
Let $f(n)$ be the maximal number of finite-index extensions of a subgroup $H\lefg F$ such that $\Gamma(H)$ has at most $n$ vertices. By Proposition \[LpLq covers\], every pair $(p,q)$ such that $p\sim
q$ is in the connected component of a pair of the form $(1,r)$ for some $r > 1$. Moreover, this connected component has elements of the form $(i,j)$ for all $1\le i \le n$, so the graph resulting from the identification of $1$ and $r$ (or from $p$ and $q$) has at most $n/2$ vertices. Thus $f(1) = 1$ and $f(n) \le n\,f(\lfloor n/2\rfloor)$ for all $n \ge 2$. It follows that $f(n) \le n^{\frac12(1+\log_2 n)}$.
Let $H \lefg F$. If $\cc(\Gamma(H))$ has $n$ vertices, then $H$ has at most $n^{\frac12(1+\log_2 n)}$ finite-index extensions.
\[ex hypercube\] By means of lower bound, we consider the following example. Let $e_1,\ldots,e_k$ be the canonical basis of the vector space ${\mathbb{Z}}_2^k$, let $\phi$ be the morphism from the free group $F$ over $A = \{a_1,\ldots,a_k\}$ into the additive group ${\mathbb{Z}}_2^k$, mapping $a_i$ to $e_i$, and let $H = \ker\phi$. Then $H$ is normal and finite-index, so all its extensions have finite index and they are in bijection with the set of quotients of ${\mathbb{Z}}_2^k$, hence with the set of subgroups of ${\mathbb{Z}}^k_2$, or equivalently with the set of subspaces of ${\mathbb{Z}}_2^k$.
Let $\ell_{d,k}$ be the number of linearly independent $d$-tuples in ${\mathbb{Z}}_2^k$ ($d\ge 1$). Then $\ell_{1,k} = 2^k-1$. If $d\ge 2$, a $d$-tuple $(x_1,\ldots,x_d)$ is linearly independent if and only if $(x_1,\ldots,x_{d-1})$ is linearly independent and $x_d$ does not belong to the subspace generated by $x_1,\ldots,x_{d-1}$, so that $\ell_{d,k} = \ell_{d-1,k}(2^k - 2^{d-1})$. Now the set of cardinality $d$ linearly independent subsets has $m_{d,k} =
\ell_{d,k}/d!$ elements, and the number of dimension $d$ subspaces of ${\mathbb{Z}}_2^k$ is $$s_{d,k} = \frac{m_{d,k}}{m_{d,d}} =
\frac{\ell_{d,k}}{\ell_{d,d}} = \frac{(2^k-1)(2^k-2)(2^k-4) \cdots
(2^k-2^{d-1})}{(2^d-1)(2^d-2)(2^d-4) \cdots (2^d-2^{d-1})}.$$ Finally, the number of subspaces of ${\mathbb{Z}}_2^k$ is equal to $\sum_{d=0}^k s_{d,k}$, with $s_{0,k} = 1$.
We observe that for each $0 \le i < d < k$, $\frac{2^k-2^i}{2^d-2^i} > 2^{k-d}$, so that $s_{d,k} >
2^{(k-d)d}$. By considering $d = \lfloor\frac k2\rfloor$, we find that $\sum s_{d,k} > 2^{k^2/4}$.
Finally, we note that $\Gamma(H)$ is the Cayley graph of ${\mathbb{Z}}_2^k$ with respect to the basis $e_1,\ldots,e_d$ (a graph known as the dimension $k$ hypercube), so that $\Gamma(H)$ has $n = 2^k$ vertices. As a result, $H$ has more than $n^{\frac14\log_2n}$ finite-index extensions.
The lattice of finite-index subgroups of $G$ {#sec:
FIS}
--------------------------------------------
Let us call *-maximal* a subgroup $G \lefg F$ which has no proper finite-index extension, that is (in view of Theorem \[maximal fi extension\]), such that $G = \Comm_F(G)$. If $G$ is -maximal, let $\FIS(G)$ be the set of all finite-index subgroups of $G$, that is, the set of subgroups $H\lefg F$ such that $H_\fim = G$. Note that distinct -maximal subgroups yield disjoint lattices of finite-index subgroups.
Let $G \lefg F$ be non-trivial and -maximal. Then $\FIS(G)$ forms a convex sublattice of the lattice of subgroups of $F$, with greatest element $G$. This sublattice is always infinite (see Remark \[remark index\]) and without a least element.
Lemma \[preserve Lr\] provides us with an invariant for every sublattice of the form $\FIS(G)$ (with $G$ -maximal).
\[prop invariant\] Let $H, K \lefg F$. Then $H_\fim =
K_\fim$ if and only if $\tt_H(1) = \tt_K(1)$ and $\widetilde
L_{\tau(1)}(H) = \widetilde L_{\tau(1)}(K)$, if and only if $\tt_H(1) = \tt_K(1)$ and $L_{\tau(1)}(H) = L_{\tau(1)}(K)$.
If $H_\fim = K_\fim$, then $H,K \lefi H_\fim$, and Proposition \[prop: fi cover\] and Lemma \[preserve Lr\] show directly that $\tt_H(1) = \tt_{H_\fim}(1) = \tt_K(1)$, $\widetilde
L_{\tau(1)}(H) = \widetilde L_{\tau(1)}(H_\fim) = \widetilde
L_{\tau(1)}(K)$ and $L_{\tau(1)}(H) = L_{\tau(1)}(H_\fim) =
L_{\tau(1)}(K)$.
We now prove the converse. More precisely, we show that if $H$ and $K$ are -maximal, $\tt_H(1) = \tt_K(1)$ and $L_{\tau(1)}(H) =
L_{\tau(1)}(K)$, then $H = K$. First we note that $\calA(H^{\tt_H(1)}) = (\cc(\Gamma(H)),\tau(1))$ and in particular, $H^{\tt_H(1)}$ is cyclically reduced and -maximal. Thus, it suffices to prove the expected result (namely, that $H = K$) under the hypothesis that $H$ and $K$ are cyclically reduced.
By Remark \[fact: Lp Lq\], the set of all $L_p(H)$ ($p \in \Gamma(H)$) coincides with the set $\{\{\red(\bar ux) \mid x \in L_1(H)\} \mid u
\in L_1(H)\}$. In addition, since $H$ is -maximal, if $u,v\in
L_1(H)$ with $1 \mapright u p$ and $1 \mapright u q$ and $\{\red(\bar
ux) \mid x \in L_1(H)\} = \{\red(\bar vx) \mid x \in L_1(H)\}$, then $p = q$. It also follows from the same fact that, again if $H$ is -maximal, there is an edge $(p,a,q)$ in $\Gamma(H)$ if and only if $L_q(H) = \{\red(\bar ax) \mid x\in L_p(H)\}$. Thus, the cyclically reduced -maximal subgroup $H$ is entirely determined by the set $L_1(H)$. This concludes the proof.
The pairs $(t, L)$ that are equal to $(\tt(1),\widetilde
L_{\tau(1)}(G))$ for some subgroup $G \lefg F$ are characterized as follows. Recall that an *i-automaton* [@SilvaFIM] is a deterministic automaton $(Q,i,E,T)$ over the alphabet $\tilde A$ such that, $(p,a,q) \in E$ if and only if $(q,\bar a,p) \in E$ for all vertices $p,q$ and $a\in A$. The automata $\calB_p$ discussed in Section \[sec: computing\] are i-automata.
Let $t \in F$ and let $L \subseteq \tilde A^*$ be a rational language. The following conditions are equivalent.
- There exists a subgroup $H \lefg F$ such that $t = \tt(1)$ and $L = \widetilde L_{\tau(1)}(H)$.
- There exists a -maximal subgroup $H \lefg F$ such that $t = \tt(1)$ and $L = \widetilde L_{\tau(1)}(H)$.
- $L$ is accepted by an i-automaton with all states accepting and such that, for each state $p$, there exist transitions $(p,a,q)$ and $(p,b,r)$ for at least two distinct letters $a,b \in \tilde A$. In addition, if $t$ is not the empty word, then $t = t'a$ for a letter $a\in \tilde A$ such that $\bar a \not\in L$.
- The minimal automaton of $L$ is an i-automaton with all states accepting and such that, for each state $p$, there exist transitions $(p,a,q)$ and $(p,b,r)$ for at least two distinct letters $a,b \in \tilde A$. In addition, if $t$ is not the empty word, then $t = t'a$ for a letter $a\in
\tilde A$ such that $\bar a \not\in L$.
- $t$ and $L$ satisfy the following conditions:
- $L$ is closed under taking prefixes;
- if $u,v \in L$, then $u\bar uv \in L$;
- if $uv\bar vw \in L$, then $uw\in L$;
- if $ua \in L$ with $a\in \tilde A$, then $uab \in
L$ for some $b\in \tilde A$ such that $b\ne \bar a$.
- if $t$ is not the empty word, then $t = t'a$ for a letter $a\in \tilde A$ such that $\bar a \not\in L$.
Conditions (1) and (2) are equivalent by Proposition \[prop invariant\].
Proposition \[prop compute Hfi\] shows that (1) implies (3’), which in turn implies (3). Let us now assume that (3) holds and let $\Gamma$ be the $A$-labeled graph induced by the states and transitions of the minimal automaton of $L$. The extra condition given shows that every vertex of $\Gamma$ is visited by a loop labeled by a cyclically reduced word. It follows that, if $G$ is the subgroup whose graphical representation is $\calA(G) = (\Gamma,q_0)$, with $q_0$ the initial state, then $G$ is cyclically reduced, $L =
\widetilde L_{q_0}(G)$. The condition on the word $t$ shows that Condition (1) holds with $H = G^t$.
Condition (3) easily implies Condition (4). Let us now assume that Condition (4) holds. By [@SilvaFIM Theorem 4.1], Properties (4.2) and (4.3) show that $L$ is accepted by an i-automaton. Property (4.1) shows that all states of that automaton are final, and Property (4.4) shows that, for each state $p$, there exist transitions $(p,a,q)$ and $(p,b,r)$ for at least two distinct letters $a,b \in
\tilde A$. Thus Condition (4) implies Condition (3), which concludes the proof.
Malnormal closure {#sec: malnormal}
=================
A subgroup $H$ of $F$ is *malnormal* if $H^g \cap H = 1$ for each $g\not\in H$. Malnormality was proved decidable in [@Baumslag], and a simple decision algorithm was given in [@KM], based on the following characterization [@KM; @Jitsukawa].
\[charact malnormal\] Let $H \lefg F$. Then $H$ is malnormal if and only if every connected component of $\Gamma(H) \times_A \Gamma(H)$, except for the diagonal complement, is a tree, if and only if, for every $p\ne q\in \cc(\Gamma(H))$, $L_p(H) \cap L_q(H)$ is finite.
This yields directly an $\O(n^2\log n)$ decision algorithm, where $n$ is the number of vertices of $\cc(\Gamma(H))$. It also yields the following corollary.
Let $H\lefg F$. If $H$ is malnormal, then $H$ is -maximal, that is, $H$ has no proper finite-index extension.
By Corollary \[charact fi extensions\], if $H$ is not -maximal, then there exist vertices $p\ne q$ in $\cc(\Gamma(H))$ such that $L_p(H) = L_q(H)$, and hence such that $L_p(H) \cap L_q(H) = L_p(H)$ is infinite. In particular, $H$ is not malnormal.
It is shown in [@MVW Prop. 4.5] that for every finitely generated subgroup $H \lefg F$, there exists a least malnormal extension $H_\mal$, called the *malnormal closure* of $H$, that $H_\mal$ is finitely generated and effectively computable, and that the rank of $H_\mal$ is less than or equal to the rank of $H$ [@MVW Corol. 4.14]. In fact, [@MVW] shows that $\Gamma(H_\mal)$ is obtained from $\Gamma(H)$ by a series of i-steps. The algorithm computing $H_\mal$ then consists in computing all the quotients of $\Gamma(H)$ and verifying, for each of them, whether it represents a malnormal subgroup. We now give a much better, polynomial-time algorithm.
\[good algo malnormal\] Let $H \lefg F_n$. The malnormal closure of $H$ is computed by repeatedly applying the following construction: in $\Gamma(H)$, identify all pairs $(p,q)$ of distinct vertices in $\cc(\Gamma(H))$, such that $L_p(H) \cap L_q(H)$ is infinite and reduce the resulting graph.
Let $H_0 = H$ and let $H_{i+1}$ be the subgroup of $F$ such that $\Gamma(H_{i+1})$ is obtained from $\Gamma(H_i)$ by, first identifying all pairs $p,q$ of vertices of $\cc(\Gamma(H_i))$ such that $L_p(H_i)
\cap L_q(H_i)$ is infinite, and then reducing the resulting graph. Since $\Gamma(H_{i+1})$ has less vertices than $\Gamma(H_i)$, this defines a finite sequence of subgroups $$H = H_0 < H_1 < \ldots < H_k,$$ where $k \ge 0$ and $H_k$ is malnormal. We want to show that $H_k$ is the least malnormal extension of $H$.
Let $K$ be any malnormal subgroup of $F$ such that $H \le K$. By Takahasi’s theorem (see for instance [@MVW]), there exists a subgroup $G$ such that $H \le G \le K$ such that $G$ is obtained from $H$ by a sequence of i-steps, and $K$ is obtained from $G$ by a sequence of re-steps (this fact can also be deduced from [@SW08 Prop. 2.6]).
Let $p,q \in \cc(\Gamma(H))$ such that $L_p(H) \cap L_q(H)$ is infinite. It is elementary to verify that $L_p(H) \subseteq
L_{\phi^G_H(p)}(G) \subseteq L_{\phi^K_G(\phi^G_H(p))}(K)$. In particular, $\phi^K_G(\phi^G_H(p)) \cap \phi^K_G(\phi^G_H(q))$ is infinite. Since $K$ is malnormal, it follows from Proposition \[charact malnormal\] that $\phi^K_G(\phi^G_H(p)) =
\phi^K_G(\phi^G_H(q))$. But $\phi^K_G$ is one-to-one by definition, so we have $\phi^G_H(p) = \phi^G_H(q)$. It follows that $G$ is obtained from $H_1$ by a sequence of i-steps, and in particular, $H_1
\le G \le K$. Iterating this reasoning, we find that $H_k \le K$, which concludes the proof.
Let $H \le F$. If $\Gamma(H)$ has $n$ vertices, then one can compute $\Gamma(H_\mal)$ in time $\O(n^3\log n)$.
According to the algorithm given in Proposition \[good algo malnormal\], we first need to compute the connected components of $\cc(\Gamma(H)) \times_A \cc(\Gamma(H))$ – done in time $\O(n^2\log
n)$ –, identifying which are trees and which are not – done in time $\O(n^2)$ –, identifying the vertices of $\cc(\Gamma(H))$ involved in a non-diagonal connected component and reducing the resulting graph to obtain $\Gamma(H_1)$ – which is done in time $\O(n\log n)$. Thus $\Gamma(H_1)$ is computed from $\Gamma(H_0)$ in time $\O(n^2\log n)$.
This part of the algorithm is iterated $k$ times, to compute $\Gamma(H_k) = \Gamma(H_\mal)$, and we have $k < n$ since the number of vertices of the $\Gamma(H_i)$ forms a properly decreasing sequence. This concludes the proof.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The authors thank A. Martino and E. Ventura for their help in the computation in Example \[ex hypercube\]. They also are indebted to the anonymous referee whose careful reading helped improve this paper.
[99]{} G. Baumslag, A. Miasnikov and V. Remeslennikov. Malnormality is decidable in free groups. *Internat. J. Algebra Comput.* **9** (1999) 687–692.
J. Brzozowski. Canonical regular expressions and minimal state graphs for definite events. In *MRI Symposia Series* **12** (Polytechnic Press, Polytechnic Institute of Brooklyn, 1962), pp. 529–561.
J.-M. Champarnaud and D. Ziadi. Canonical derivatives, partial derivatives and finite automaton constructions. *Theoret. Comput. Sci.* **289** (2002) 137–163.
T. Jitsukawa. Malnormal subgroups of free groups. In *Computational and statistical group theory* (Contemp. Math. 298, Amer. Math. Soc., 2002), pp. 83–95.
I. Kapovich and A.G. Myasnikov. Stallings foldings and subgroups of free groups. [*J. Algebra*]{} [**248**]{} (2002) 608–668.
D. Kozen. *Automata and Computability* (Springer, 1997).
A. Miasnikov, E. Ventura, P. Weil. Algebraic extensions in free groups. In *Algebra and Geometry in Geneva and Barcelona* (G.N. Arzhantseva, L. Bartholdi, J. Burillo, E. Ventura eds.) (Trends in Mathematics, Birkhaüser, 2007), pp. 225–253.
J.-P. Serre. [*Arbres, amalgames, $SL_{2}$*]{} (Astérisque 46, Soc. Math. France, 1977). English translation: *Trees* (Springer Monographs in Mathematics, Springer, 2003).
P.V. Silva. On free inverse monoid languages. *Theoretical Informatics and Applications* **30** (1996) 349–378.
P.V. Silva, P. Weil. On an algorithm to decide whether a free group is a free factor of another. *Theor. Inform. Appl.* **42** (2008) 395–414.
J. Stallings. The topology of graphs. [*Invent. Math.*]{} [**71**]{} (1983) 551–565.
P. Weil. Computing closures of finitely generated subgroups of the free group. In *Algorithmic Problems in Groups and Semigroups* (J.-C. Birget, S. Margolis, J. Meakin, M. Sapir eds.) (Birkhaüser, 2000), pp. 289–307.
[^1]: Faculdade de Ciências, Universidade do Porto, R. Campo Alegre 687, 4169-007 Porto, Portugal
[^2]: LaBRI, 351 cours de la Libération, 33400 Talence, France.
[^3]: The first author acknowledges support from Project ASA (PTDC/MAT/65481/2006) and C.M.U.P., financed by F.C.T. (Portugal) through the programmes POCTI and POSI, with national and European Community structural funds. Both authors acknowledge support from ESF project <span style="font-variant:small-caps;">AutoMathA</span>.
[^4]: The converse is not true, see [@MVW] for a detailed study of the extensions $H \lefg G$ such that $\phi_H^G$ is onto.
[^5]: This reduction operation is the iteration of Stallings’s *folding* operation [@Stallings]; our terminology emphasizes the fact that this is a generalization of the reduction of a word (the iterated process of deleting factors of the form $a\bar
a$, $a\in \tilde A$).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the results of near–infrared H band (1.65 $\mu$m) imaging of 11 BL Lac objects with redshifts ranging from z = 0.05 to 0.9. We are able to clearly detect the host galaxy in seven low redshift (z$\leq$0.24) BL Lacs, while the four unresolved BL Lacs have either high or unknown redshift. The galaxies hosting the low redshift BL Lacs are large (average bulge scale length R(e) = 8.8$\pm$9.9 kpc) and luminous (average M(H) = –25.8$\pm$0.5), slightly brighter than the typical galaxy luminosity L\* (M\*(H) = –25.0$\pm$0.2), and of similar luminosity to or slightly fainter than brightest cluster galaxies (M(H) = –26.3$\pm$0.3). The average optical/near–infrared colour and colour gradient of the BL Lac hosts (R–H = 2.2$\pm$0.5; $\Delta$(R–H)/$\Delta$(log r) = –0.09$\pm$0.04) are consistent with the hosts being normal ellipticals, indicating that the nuclear activity has only a marginal effect on the star formation history and other properties of the hosts. The BL Lac hosts appear slightly less luminous than those of higher redshift flat spectrum radio quasars. The nucleus–to–galaxy luminosity ratio of the BL Lacs is similar to that of low redshift radio galaxies and consistent with what found in previous optical studies of BL Lacs. However, it is smaller that that found for flat spectrum radio quasars, suggesting there is a difference in the intrinsic brightness of the nuclear source or in the Doppler beaming factor between the two types of blazars.'
author:
- 'Jari K. Kotilainen , Renato Falomo and Riccardo Scarpa'
date: 'Accepted May 18, 1998; received April 10, 1998; in original form February 21, 1998'
title: 'The host galaxies of BL Lac objects in the near–infrared[^1]'
---
-2cm 10000 10000 \#1 \#1
to size
\#1
to size
ß \#1 PS. 2
psfig.sty
Introduction
============
BL Lac objects are active galactic nuclei (AGN) characterized by strong and rapidly variable continuum emission and polarization across the electromagnetic spectrum, strong compact flat spectrum radio emission and superluminal motion (see Kollgaard 1992 for general references). They share many properties with flat spectrum radio quasars (FSRQ) and are often grouped together as blazars. The clearest difference between them is that the latter have strong broad emission lines, while these are very weak or absent in BL Lacs.
Blazar properties are usually explained by the beaming model (Blandford & Rees 1978), where the observed emission is dominated by a synchrotron emitting relativistically boosted jet oriented close to our line–of–sight. This model is supported by the fact that almost all blazars are strong and rapidly variable $\gamma$–ray sources (von Montigny 1995). The beaming model implies the existence of a more numerous parent population of intrinsically identical objects but with their jet oriented at larger angles to our line–of–sight. In the current unified models for radio–loud AGN (Urry & Padovani 1995), BL Lac objects are unified with low luminosity core–dominated (F–R I) radio galaxies (RG) seen nearly along the jet axis, while the high luminosity lobe–dominated (F–R II) RGs represent the parents of FSRQs (Padovani & Urry 1990; Urry, Padovani & Stickel 1991). However, for potential problems in this simple unification, see Urry & Padovani (1995). For a direct test of the unification model, we need to compare orientation–independent properties of BL Lac objects with those of the parents, extended radio emission, host galaxies and environments.
Considerable amount of optical imaging exists for relatively nearby (z$\leq$0.5) BL Lac hosts (Abraham, McHardy & Crawford 1991; Stickel, Fried & Kühr 1993; Falomo 1996; Wurtz 1996; Falomo 1997b; Jannuzi, Yanny & Impey 1997). The host galaxies of nearby BL Lacs have turned out to be predominantly giant ellipticals with similar magnitude to F–R I RGs, although there appear to be some cases of disk dominated host galaxies (McHardy 1991; Abraham 1991; Stocke, Wurtz & Perlman 1995, but see contradicting views by Romanishin 1992; Stickel 1993; Falomo 1997a). The extended radio power and optical environments of BL Lacs are also consistent with those of F–R I RGs, but suggest a contribution to the parent population from F–R II RGs (Kollgaard 1992; Pesce, Falomo & Treves 1995; Wurtz 1997).
Very little near–infrared (NIR) imaging exists on BL Lac objects. However, NIR wavelengths may offer some advantages. Optical emission from BL Lacs is often dominated by the nuclear source, while the luminosity of the massive old stellar population peaks in the NIR. This leads to a better contrast of the nebulosity with respect to the nuclear source at these wavelengths. One also needs to apply much lower K–correction in the NIR than in the optical. In this paper we present NIR H–band (1.65 $\mu$m) images of 11 BL Lac objects and compare the NIR host properties with those of RGs and FSRQs. The BL Lacs were observed during our project to study the host galaxies of a complete sample of FSRQs (Kotilainen, Falomo & Scarpa 1998; hereafter KFS98) and thus they do not satisfy any criteria of completeness. However, all the low redshift BL Lacs in this sample have previously been imaged in the optical by us. The same procedure of analysis was performed on the NIR and optical datasets, thus allowing us to investigate the R–H colour of the host galaxies in a homogeneous manner. Properties of the observed objects are given in Table 1. In section 2, we briefly describe the observations, data reduction and the method of the analysis and refer the reader to a more thorough discussion given in KFS98. Our results are presented in section 3 and conclusions in section 4. Throughout this paper, H$_{0}$ = 50 and q$_{0}$ = 0 are used.
[llrllrrr]{}\
\
& Other name & z & V & M(B) & Date & Exp. time & FWHM\
& & & & & & (min) & (arcsec)\
\
PKS 0048–097 & OB-080 & $\geq$0.5(?) & 16.3 & – & 21/8/95 & 40 & 1.0\
PKS 0118–272 & OC-230.4 & $\geq$0.557 & 15.9 & $\leq$-26.8 & 18/8/95 & 37 & 1.0\
PKS 0521–365 & & 0.055 & 14.6 & -22.3 & 13/1/96 & 21 & 1.2\
PKS 0537–441 & & 0.896 & 15.0 & (-27.0) & 13/1/96 & 36 & 1.1\
PKS 0548–322 & & 0.069 & 15.5 & -22.0 & 13/1/96 & 28 & 1.0\
PKS 1514–241 & AP Lib & 0.049 & 14.9 & -21.7 & 19/8/95 & 15 & 0.9\
PKS 1538$+$149 & 4C 14.60 & 0.605 & 17.8 & -25.2 & 21/8/95 & 40 & 1.7\
PKS 2005–489 & & 0.071 & 14.4 & -24.8 & 19/8/95 & 30 & 1.0\
MS 2143.4$+$070 & & 0.237 & 18.0 & -22.8 & 18/8/95 & 40 & 0.9\
PKS 2155–305 & & 0.116 & 13.5 & -25.9 & 19/8/95 & 22 & 1.0\
PKS 2254$+$074 & OY 091 & 0.190 & 16.4 & -23.3 & 18/8/95 & 10 & 1.2\
\
Observations, data reduction and modeling of the luminosity profiles
====================================================================
The observations were carried out at the 2.2m telescope of European Southern Observatory (ESO), La Silla, Chile, in August 1995 and January 1996, using the 256x256 px IRAC2 NIR camera (Moorwood 1992) with pixel scale 0.27$''$ px$^{-1}$, giving a field of view of 69 arcsec$^2$. Log of the observations is given in Table 1. The observations, data reduction and modeling of the luminosity profiles were performed following the procedure described in KFS98. Briefly, for nearby targets object frames were interspersed with sky frames. Distant targets, on the other hand, were always kept in the field by shifting them across the array. Individual exposures were coadded to achieve the final integration time (see Table 1).
Data reduction consisted of correction for bad pixels by interpolating across neighboring pixels, sky subtraction using a median averaged sky frame (for nearby sources) or a median averaged frame of all observations (for distant sources), flat-fielding, and combination of images of the same target. Standard stars from the list of Landolt (1992) were used for photometric calibration, for which we estimate an accuracy of $\sim$0.1 mag. K–correction was applied to the host galaxy magnitudes following the method of Neugebauer (1985). The size of this correction is insignificant at low redshift in the H filter (m(H) = 0.01 at z = 0.2). No K–correction was applied to the nuclear component, assumed to have a power-law spectrum ($f_\nu \propto \nu^{-\alpha}$) with $\alpha$ $\sim$–1.
Azimuthally averaged radial luminosity profile was extracted for each object and stars in the frames down to surface brightness of $\mu$(H) = 22–23 mag arcsec$^{-2}$. Only for few objects, bright stars were present in the field for straightforward PSF determination. For most sources, the core of the PSF was derived from faint field stars, while the wing was extrapolated using a suitable Moffat (1969) function obtained from fitting bright stars in other frames of similar seeing during the same night. We stress that this extrapolation does not affect the fit for the low redshift, often host–dominated BL Lacs in our sample. The luminosity profiles were fitted into point source and galaxy components by an iterative least-squares fit to the observed profile. For low redshift objects, we attempted both elliptical (de Vaucouleurs law) and exponential disc models to represent the galaxy. However, consistently with results from optical images, in no case did the disc model give a better fit than the elliptical one for the observed sources. For higher redshift objects, data quality does not allow one to discriminate between the morphologies and for them the elliptical model was assumed. We estimate the uncertainty of the derived host galaxy magnitudes to be $\sim$$\pm$0.3 mag.
Results and discussion
======================
In Fig. 1 we show the H band contour plots of all the BL Lacs, after smoothing the images with a Gaussian filter of $\sigma$ = 1 px. We are able to clearly detect the host galaxy in all the BL Lacs at low redshift (z$\leq$0.24). In the unresolved cases, the redshift is either high (PKS 1538+14; z = 0.605, PKS 0537–441; z = 0.896) or unknown but probably high (z$\geq$0.5; PKS 0048–097 and z$\geq$0.557; PKS 0118–272). Therefore, these sources are still consistent with being hosted by giant ellipticals. On the other hand, if their redshifts are lower, they may reside in sub–luminous hosts, which could account for the failure to detect absorption lines in the spectra of the latter two BL Lacs. In Fig. 2, we show the radial luminosity profiles of each BL Lac object, together with the best–fit models overlaid. In the Appendix, we compare our NIR photometry with previous studies, and discuss in detail individual BL Lacs, including comparison with previous optical determinations of the host galaxies. The results derived from the profile fitting are summarized in Table 2, where column (1) and (2) give the name and redshift of the object; (3) the bulge scalelength in arcsec and kpc; (4) and (5) the apparent nuclear and host galaxy magnitude; (6) the nucleus/galaxy luminosity ratio; (7) and (8) the absolute nuclear and host galaxy magnitude; and (9) whether the image of object is resolved (R) or unresolved (U).
[lllllllll]{}\
\
& z & r(e)/R(e) & m(nuc) & m(g) & L(nuc)/L(gal) & M(nuc) & M(gal) & Note\
& & arcsec/kpc & & & & & &\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9)\
\
PKS 0048–097 & $\geq$0.5(?) & & 13.7 & & $\geq$100 & $\leq$-29.2 & & U\
PKS 0118–272 & $\geq$0.557 & & 13.1 & & $\geq$100 & $\leq$-30.1 & & U\
PKS 0521–365 & 0.055 & 2.55/3.8 & 12.0 & 12.4 & 1.70 & -25.8 & -25.2 & R\
PKS 0537–441 & (0.896) & & 13.0 & & $\geq$100 & (-32.0) & & U\
PKS 0548–322 & 0.069 & 4.55/8.3 & 15.4 & 12.4 & 0.087 & -23.0 & -25.7 & R\
PKS 1514–241 & 0.048 & 2.50/3.3 & 13.3 & 12.2 & 0.43 & -24.1 & -25.1 & R\
PKS 1538$+$149 & 0.605 & & 14.7 & & 3.0 & -28.7 & & U\
PKS 2005–489 & 0.071 & 1.65/3.1 & 11.7 & 11.9 & 1.38 & -26.6 & -26.3 & R\
MS 2143.4$+$070 & 0.237 & 1.10/5.5 & 16.4 & 15.1 & 0.35 & -24.7 & -25.9 & R\
PKS 2155–305 & 0.116 & 1.75/5.0 & 11.0 & 12.4 & 4.00 & -28.5 & -26.9 & R\
PKS 2254$+$074 & 0.190 & 7.60:/32.5: & 14.0: & 14.9: & 2.27: & -26.5: & -25.6: & R\
\
\
\
The near–infrared properties of the host galaxies
-------------------------------------------------
The host galaxies of all the low redshift resolved BL Lacs in this sample have previously been studied by us in the optical (Falomo 1996; Falomo & Kotilainen, in preparation). The observations presented here are, however, the first study of them in the NIR, and allow us to address for the first time the issue of the optical–NIR colour of the BL Lac host galaxies. The integrated colours of the host galaxies are given in Table 3, where column (1) gives the name of the BL Lac; (2) and (3) the absolute magnitude of the host galaxy in the H band (this work) and R band (literature value), respectively; (4) the reference for column (3); (5) the R–H colour of the host galaxy computed from columns (2) and (3); (6) the same as column (5), but computed from our colour profiles (see below); and (7) the R–H colour gradient of the host. For comparison, in Table 4 we give optical and optical–NIR colours of elliptical galaxies from literature search. The average and median R–H host colours of our BL Lac sample, R–H = 2.2$\pm$0.5 and R–H = 2.3 are consistent with those found for local giant ellipticals (R–H = 2.5; see Table 4), considering the large variations both within the sample and in the previous optical determinations of the host parameters. The average R–H colour also agrees well with recent evolutionary population synthesis models, which predict for low redshift, solar metallicity ellipticals R–H$\sim$2.2 (Poggianti 1997; age 15 Gyr) and R–H$\sim$2.4 (Fioc & Rocca-Volmerange 1997; age 13 Gyr).
[lllllll]{}\
\
& M(H) & M(R) & Ref. & R-H & R-H & $\Delta$(R-H)/$\Delta$(log r)\
(1) & (2) & (3) & (4) & (5) & (6) & (7)\
\
PKS 0521–365 & -25.4 & -23.2 & F94 & 2.2 & 2.3 & -0.08\
& & -23.0 & W96 & 2.4 & &\
PKS 0548–322 & -25.7 & -24.2 & F95 & 1.5 & 2.1 & -0.06\
& & -23.2 & W96 & 2.5 & &\
PKS 1514–241 & -25.1 & -22.8 & B87 & 2.3 & 2.1 & -0.12\
& & -22.8 & A91 & 2.3 & &\
& & -23.5 & S93 & 1.6 & &\
& & -22.9 & this work & 2.2 & &\
PKS 2005–489 & -26.3 & -24.2 & S93 & 2.1 & 2.7 & -0.15\
& & -23.7 & F96 & 2.6 & &\
MS 2143.4$+$070 & -25.9 & -23.4 & W96 & 2.5 & 2.5 & -0.04\
PKS 2155–305 & -26.8 & -24.4 & F96 & 2.4$^a)$ & &\
PKS 2254$+$074 & -25.6: & -24.1 & S93 & 1.5: & 1.8: & 0.10:\
& & -23.9 & F96 & 1.7: & &\
& & -23.9 & W96 & 1.7: & &\
\
\
\
\
\
\
\
[lllllll]{}\
\
& V–R & V–I & V–K & J–H & H–K & J–K\
\
Gregg (1989) & 0.58$\pm$0.11 & 1.23$\pm$0.22 & & & &\
Carollo (1997) & & 1.30$\pm$0.05 & & & &\
Schombert (1993) & & & 3.29$\pm$0.09 & & & 0.87$\pm$0.04\
Bressan, Chiosi & Tantalo (1996) & & & 3.28$\pm$0.11 & & &\
Recillas-Cruz (1990) & & & & 0.72$\pm$0.08 & 0.22$\pm$0.06 &\
Silva & Elston (1994) & & & & & & 0.90$\pm$0.02\
\
\
We have next computed radial R–H colour profiles for six of the BL Lacs (Fig. 3). From these profiles we have computed the R–H colour gradients of the host galaxies as reported in Table 3, column (7). With the exception of PKS 2254+074 (see Appendix), all colour profiles show a modest colour gradient (average $\Delta$(R–H)/$\Delta$(log r) = –0.09$\pm$0.04) that makes the galaxies bluer in the outer regions. The sign and amplitude of the color gradient is similar to that exhibited by normal non-active ellipticals ($\Delta$(V–K)/$\Delta$(log r) = –0.16$\pm$0.18; Peletier, Valentijn & Jameson 1990; 12 ellipticals); $\Delta$(V–K)/$\Delta$(log r) = –0.26$\pm$0.15; Schombert 1993; 16); the R–H gradient tends to be smaller than that of V–K, probably due to the smaller wavelength baseline. Finally, we show in Fig. 4 the colour–magnitude diagram of the BL Lac host galaxies, compared with those of elliptical and S0 galaxies in the Virgo and Coma clusters (Bower, Lucey & Ellis 1992a,b). With the exceptions of PKS 0548–322 and PKS 2254+074 (see Appendix for details), the BL Lac hosts follow reasonably well the established relation for elliptical galaxies.
There is increasing evidence that the photometric properties of elliptical galaxies (at least of those in nearby clusters) can be explained in terms of a single burst of star formation at high redshift, followed by passive stellar evolution (Stanford, Eisenhardt & Dickinson 1998 and references therein). The similarity of the colours and colour gradients of BL Lac hosts to those of normal elliptical galaxies then indicates that the nuclear activity has only marginal effect, if any, on the overall properties of the host galaxy. For example, galaxy interactions and mergers, often invoked to explain the BL Lac phenomenon and AGN in general (Heckman 1990), do not seem to induce strong recent star formation in the BL Lac hosts. This is perhaps due to a very low gas density in the host and/or in the interacting galaxy. Alternatively, it may be that the timescales of the star formation episodes and nuclear activity are different.
Table 5 presents a comparison of the H–band absolute magnitudes of the BL Lac hosts with the average values of relevant samples of blazars and RGs from previous optical and NIR studies in the literature. Column (1) gives the sample; (2) the filter; (3) the number of objects in the sample; (4) the average redshift of the sample; and (5) and (6) the average H band nuclear and k-corrected host galaxy absolute magnitude of the sample. All magnitudes were transformed into our adopted cosmology and into the H band. Based on the similarity of the BL Lac host galaxies to giant ellipticals (see discussion above), the transformation of published magnitudes to the H band was obtained assuming the following colours of giant ellipticals, computed from the literature values given in Table 4: V–H = 3.1, R–H = 2.5, I–H = 1.8, and H–K = 0.2. Finally, transformation from Gunn r filter (r–H = 2.8) was obtained from R–H = 2.5 (above) and r–R = 0.3, appropriate for low redshift ellipticals (Fukugita, Shimasaku & Ichikawa 1995).
[llrlll]{}\
\
& filter & N & $<z>$ & $<M_H(nuc)>$ & $<M_H(host)>^b)$\
\
(1) & (2) & (3) & (4) & (5) & (6)\
\
BL this work & H & 7 & 0.112$\pm$0.068 & -25.7$\pm$1.7 & -25.8$\pm$0.5\
& & & & &\
L\* Mobasher (1993) & K & 136 & 0.077$\pm$0.030 & & -25.0$\pm$0.2\
& & & & &\
BCM Thuan & Puschell (1989) & H & 84 & 0.074$\pm$0.026 & & -26.3$\pm$0.3\
BCM Aragon-Salamanca (1998) & K & 25 & 0.449$\pm$0.266 & & -27.0$\pm$0.3\
& & & & &\
BL Falomo (1996) & R & 11 & 0.143$\pm$0.082 & -26.3$\pm$2.1 & -26.0$\pm$0.6\
BL Wurtz (1996) & r & 35 & 0.266$\pm$0.162 & -24.5$\pm$2.3 & -26.3$\pm$0.7\
BL Falomo (1997b), Jannuzi (1997) & I & 6 & 0.351$\pm$0.172 & -26.0$\pm$0.8 & -26.0$\pm$0.8\
& & & & &\
RG Fasano (1996) & R & 25 & 0.061$\pm$0.027 & & -26.6$\pm$0.6\
RG low–z Zirbel (1996) & V & 55 & 0.107$\pm$0.047 & &-25.8$\pm$0.6\
RG F–R II high–z Zirbel (1996) & V & 25 & 0.389$\pm$0.053 & & -25.9$\pm$0.4\
RG F–R II Taylor (1996) & K & 12 & 0.214$\pm$0.049 & -25.1$\pm$0.7 & -26.1$\pm$0.8\
& & & & &\
FSRQ/R+M KFS98 & H & 9 & 0.671$\pm$0.157 & -29.7$\pm$0.8 & -26.7$\pm$1.2\
FSRQ/R KFS98 & H & 4 & 0.673$\pm$0.141 & -30.2$\pm$0.7 & -27.8$\pm$0.3\
\
\
\
\
\
\
In Fig. 5 (upper panel) we show the H–z (Hubble) diagram for the BL Lac hosts (this work), together with data for FSRQs (KFS98) and F–R II RGs (Taylor 1996), compared to the established relation for RGs (solid line; Lilly & Longair 1984; Lilly, Longair & Miller 1985; Eales 1997), and the evolutionary model for elliptical galaxies derived from passive stellar evolution models (Bressan, Chiosi & Fagotto 1994, dashed line), normalized to the average redshift and magnitude of the low redshift RGs from Taylor 1996). The resolved BL Lac hosts lie quite well on the H–z relation. In Fig. 5 (lower panel) we show the H–z diagram for the mean values of various blazar and RG samples from the literature. These values consistently follow the H–z relationship (within the errors), with the average value for the BL Lacs derived in this work well matching the relationship.
In Fig. 6 (upper panel), we show the H–band host galaxy absolute magnitude vs. redshift for the BL Lacs (this work), FSRQs (KFS98) and RGs (Taylor 1996). In Fig. 6 (lower panel), the same diagram is shown for the mean values of various blazar and RG samples. The average H–band absolute magnitude of the seven clearly resolved low redshift (z$\leq$0.2) BL Lac hosts is M(H) = –25.8$\pm$0.5, and the average and median bulge scalelength R(e) = 8.8$\pm$9.9 kpc and 5.0 kpc. The host galaxies are therefore large (all have R(e)$\geq$3 kpc, the upper boundary of normal local ellipticals; Capaccioli, Caon & D’Onofrio 1992) and slightly more luminous than an L$^*$ galaxy (M(H) = –25.0$\pm$0.2; Mobasher 1993). Indeed, we find no BL Lac host fainter than L$^*$. As found in the optical by Wurtz (1996), the BL Lac hosts have slightly, but not significantly lower luminosities in the NIR than nearby brightest cluster member galaxies (BCM; z = 0.07$\pm$0.03; M(H) = –26.3$\pm$0.3; Thuan & Puschell 1989). However, they appear significantly fainter than BCMs at higher redshift (z = 0.45$\pm$0.27; M(H) = -27.0$\pm$0.3; Aragon-Salamanca, Baugh & Kauffmann 1998).
Although the number of objects in this study is small, it is interesting to compare the NIR absolute magnitudes of BL Lac hosts to those of available samples of BL Lacs, RGs and FSRQs from the literature. These samples span a moderately large range in redshift from z$\sim$0.03 up to z$\sim$0.7. The average host galaxy magnitudes for the various samples are given in Table 5, and shown in Fig. 6. The BL Lacs in our study have host galaxies very similar in luminosity to those found by previous optical studies of BL Lacs if normal colors are assumed. Interestingly, there is good agreement between the BL Lac hosts in this study and the F-R II RGs of Taylor (1996), the only RG sample studied in the NIR and thus almost free from colour term uncertainty. The lack of direct comparison sample of F-R I RGs, however, prevents us from discussing the issue of the parent population of BL Lac objects further.
Since BL Lacs share many properties (variability and polarization) with FSRQs, it is interesting to compare the host properties of these two types of blazars. Recently, KFS98 found for clearly resolved FSRQ hosts at z$\sim$0.7 M(H) = –27.8$\pm$0.3, while the value adding their marginally resolved hosts is M(H) = –26.7$\pm$1.2 (Table 5). Because not all FSRQs were resolved, it is possible that the average host luminosity of FSRQs is even fainter. Moreover, given the large difference of redshift between the BL Lac and FSRQ samples, it is also possible that the $\sim$0.5 mag difference in host luminosity is due to cosmological evolution in the stellar population of an elliptical galaxy that makes it fade by $\sim$0.8 mag between z$\sim$1 and z$\sim$0 (models by Bressan 1994). Indeed, there is increasing observational evidence for significant evolution in the colours and luminosity of early–type galaxies (Aragon-Salamanca 1998 and references therein). The present data suggests that BL Lac hosts are less luminous than those of FSRQs but, given the limits described above, additional homogeneous observations of BL Lacs and FSRQs are needed to reach a firm conclusion.
The nuclear component
---------------------
The average absolute magnitude of the fitted nuclear component for the clearly resolved BL Lacs is M(H) = –25.7$\pm$1.7. On the other hand, the FSRQs have much brighter nuclear component (M(H) = –29.7$\pm$0.8; KFS98). This difference in the strength of the nuclear component is also more evident considering the nucleus/galaxy luminosity ratio L(nuc)/L(gal) in Table 2. None of the low redshift BL Lacs (or the RGs from Taylor 1996) have L(nuc)/L(gal)$\geq$10, whereas about half of the FSRQs are above this limit (KFS98). From Figs. 5 and 6 it is clear that the host galaxies of the various samples considered here are not dramatically different in intrinsic luminosity, especially if cosmological evolution is taken into account. Therefore, this difference suggests that FSRQs exhibit a nuclear component which is systematically brighter than that of the low redshift BL Lac objects, possibly due to either intrinsically higher luminosity or a larger Doppler beaming factor. While this result disagrees with the current unified models of AGN, further observations of larger and unbiased samples of BL Lacs and FSRQs are needed to elucidate this point.
Conclusions
===========
In this paper we have presented the results of a near–infrared imaging study of a sample of 11 BL Lac objects, for most of which the host galaxy is clearly resolved. Consistently with what is found in optical studies, we find that the host galaxies of low redshift BL Lacs are large (average bulge scale length R(e) = 8.8$\pm$9.9 kpc) and luminous (average M(H) = –25.8$\pm$0.5); they are more luminous than L$^*$ galaxies (by $\sim$1 mag) but of similar luminosity to or slightly fainter than the brightest cluster galaxies. Our NIR study was able for the first time to address the issue of the optical–NIR colour of BL Lac host galaxies. The average R–H colour and colour gradient of the BL Lac hosts are consistent with those of non-active early-type galaxies, suggesting that the nuclear activity does not have much effect on the star formation history of the host galaxies. The nucleus–to–galaxy ratio of BL Lacs is similar to that found in low redshift RGs and consistent with what found in previous optical studies of BL Lacs. However, it is smaller that that found for the higher redshift FSRQs (KFS98), suggesting there is a difference in the intrinsic brightness of the nuclear source or in the Doppler beaming factor between the two types of blazars. We finally encourage a systematic NIR multiwavelength study of a large, well defined sample of BL Lac objects and their immediate environments with the new generation large NIR arrays.
Acknowledgments {#acknowledgments .unnumbered}
===============
JKK acknowledges a research grant from the Academy of Finland during the initial part of this work. We thank the anonymous referee for prompt comments. This research has made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
Abraham,R.G., McHardy,I.M., Crawford,C.S., 1991, MNRAS 252, 482 Allen,D.A., Ward,M.J., Hyland,A.R., 1982, MNRAS 199, 969 Andrews,P.J., Glass,I.S., Hawarden,T.G., 1974, MNRAS 168, 7P Aragon-Salamanca,A., Baugh,C.M., Kauffmann,G., 1998, MNRAS, in press Baxter,D.A., Disney,M.J., Phillipps,S., 1987, MNRAS 228, 313 Bersanelli,M., Bouchet,P., Falomo,R., Tanzi,E.G., 1992, AJ 104, 28 Blandford,R.D., Rees,M.J., 1978, Pittsburgh conference on BL Lac objects, p. 328 (ed. A.M. Wolfe) Bower,R.G., Lucey,J.R., Ellis,R.S., 1992a, MNRAS 254, 589 Bower,R.G., Lucey,J.R., Ellis,R.S., 1992b, MNRAS 254, 607 Bressan,A., Chiosi,C., Fagotto,A., 1994, ApJS 94, 63 Bressan,A., Chiosi,C., Tantalo,R., 1996, A&A 311, 425 Brindle,C., Hough,J.H., Bailey,J.A., Axon,D.J., Hyland,A.R., 1986, MNRAS 221, 739 Capaccioli,M., Caon,N., D’Onofrio,M., 1992, MNRAS 259 323 Carollo,C.M., Franx,M., Illingworth,G.D., Forbes,D.A., 1997, ApJ 481, 710 Disney,M.J., Peterson,B.A., Rodgers,A.W., 1974, ApJ 194, L79 Eales,S., Rawlings,S., Law-Green,D., Cotter,G., Lacy,M., 1997, MNRAS 291, 593y Falomo,R., 1994, Msngr 77, 49 Falomo,R., 1996, MNRAS 283, 241 Falomo,R., Maraschi,L., Tanzi,E.G., Treves,A., 1987, ApJ 318, L39 Falomo,R., Giraud,E., Maraschi,L. et al, 1991, ApJ 380, L67 Falomo,R., Melnick,J., Tanzi,E.G., 1992, A&A 255, L17 Falomo,R., Pesce,J.E., Treves,A., 1993, ApJ 411, L63. Falomo,R., Pesce,J.E., Treves,A., 1995, ApJ 438, L9. Falomo,R., Kotilainen,J., Pursimo,T. , 1997a, A&A 321, 374 Falomo,R., Urry,C.M., Pesce,J.E. , 1997b, ApJ 476, 113 Fasano,G., Falomo,R., Scarpa,R., 1996, MNRAS 282, 40 Fioc,M., Rocca-Volmerange,B., 1997, A&A 326, 950 Fosbury,R.A.E., Disney,M.J., 1976, ApJ 207, L75 Fukugita,M., Shimasaku,K., Ichikawa,T., 1995, PASP 107, 945 Gear,W.K., 1993, MNRAS 264, 919 Glass,I.S., 1979, MNRAS 186, 29P Gregg,M.D., 1989, ApJS 69, 217 Heckman,T.M., 1990, in Paired and interacting galaxies (eds. J.W.Sulentic,W.C.Keel,C.M.Telesco), p. 359; NASA Conf. Publ. 3098 Jannuzi,B.T., Yanny,B., Impey,C., 1997, ApJ 491, 146 Kollgaard,R.I., Wardle,J.F.C., Roberts,D.H., Gabuzda,D.C., 1992, AJ 104, 1687 Kotilainen,J.K., Falomo,R., Scarpa,R., 1998, A&A 332, 503 (KFS98) Landolt,A.U., 1992, AJ 104, 340 Lilly,S.J., Longair,M.S., 1984, MNRAS 211, 833 Lilly,S.J., Longair,M.S., Miller,L., 1985, MNRAS 214, 109 McHardy,I.M., Abraham,R.G., Crawford,C.S. , 1991, MNRAS 249, 742 Mead,A.R.G., Ballard,K.R., Brand,P.W.J.L. , 1990, A&AS 83, 183 Mobasher,B., Sharples,R.M., Ellis,R.S, 1993, MNRAS 263, 560 Moffat,A.F.J., 1969, A&A 3, 455 Moorwood,A., Finger,G., Biereichel,P. , 1992, Msngr 69, 61 Neugebauer,G., Matthews,K., Soifer,B.T., Elias,J.H., 1985, ApJ 298, 275 Padovani,P., Urry,C.M., 1990, ApJ 356, 75 Peletier,R.F., Valentijn,E.A., Jameson,R.F., 1990, A&A 233, 62 Pesce,J.E., Falomo,R., Treves,A., 1995, AJ 110, 1554 Peterson,B.A., Jauncey,D.L., Wright,A.E., Condon,J.J., 1976, ApJ 237, L11 Poggianti,B., 1997, A&AS 122, 399 Recillas-Cruz,E., Carrasco,L., Serrano,A., Cruz-Gonzalez,I., 1990, A&A 229, 64 Romanishin,W., 1992, ApJ 401, L65 Schombert,J.M., Hanlan,P.C., Barsony,M., Rakos,K.D., 1993, AJ 106, 923 Silva,D.R., Elston,R., 1994, ApJ 428, 511 Stanford,S.A., Eisenhardt,P.R., Dickinson,M., 1998, ApJ 492, 461 Stickel,M, Fried,J.W., Kühr,H, 1988, A&A 206, L30 Stickel,M., Fried,J.W., Kühr,H., 1993, A&AS 98, 393 Stocke,J.T., Wurtz,R.E., Perlman,E.S., 1995, ApJ 454, 55 Tanzi,E.G., Barr,P., Bouchet,P. , 1986, ApJ 311, L13 Taylor,G.L., Dunlop,J.S., Hughes,D.H., Robson,E.I., 1996, MNRAS 283, 930 Thuan,T.X., Puschell,J.J., 1989, ApJ 346, 34 Urry,C.M., Padovani,P., 1995, PASP 107, 803 Urry,C.M., Padovani,P., Stickel,M., 1991, ApJ 382, 501 von Montigny,C., Bertsch,D.L., Chiang,J. , 1995, ApJ 440, 525 Wurtz,R., Stocke,J.T., Yee,H.K.C., 1996, ApJS 103, 109 Wurtz,R., Stocke,J.T., Ellingson,E., Yee,H.K.C., 1997, ApJ 480, 547 Zirbel,E., 1996, ApJ 473, 713\
Appendix: Notes on individual objects and comparison with previous NIR photometry {#appendix-notes-on-individual-objects-and-comparison-with-previous-nir-photometry .unnumbered}
=================================================================================
[**PKS 0048–097.**]{} Our H–band photometry agrees well with previous studies (Table 6). High resolution optical imaging by Falomo (1996) revealed a faint (m(R)$\sim$22.5) companion object $\sim$2.5$''$ to the E of the BL Lac nucleus. The host galaxy of PKS 0048–097 remained unresolved by Falomo (1996), but assuming an elliptical host of M(R) = –23.5, he derived a lower limit to the redshift of z$\geq$0.5. The host galaxy remains unresolved in the NIR. Some hint of the eastern companion object is visible in NIR.
[llll]{}\
\
\
& H mag$^a)$ & H mag range & References\
\
& (this work) & (literature) &\
\
PKS 0048–097 & 13.64 & 12.78 - 14.86 & M90, A82\
PKS 0118–272 & 13.29 & 12.69 - 13.10 & M90, B92\
PKS 0521–365 & 11.75 & 10.80 - 12.64 & A74, B86\
PKS 0537–441 & 13.04 & 11.69 - 13.34 & T86, B92\
PKS 0548–322 & 13.39 & 12.81 - 13.13 & G79, A82\
PKS 1514–241 & 12.47 & 11.54 - 12.58 & B92, B86\
PKS 1538$+$14 & 14.57 & 14.23 - 14.67 & B86, G93\
PKS 2005–48 & 11.34 & 11.40 & B92\
MS 2143.4$+$07 & 15.25 & 14.60 & G93\
PKS 2155–305 & 10.93 & 10.70 - 11.06 & B86, B92\
PKS 2254$+$074 & 14.38 & 12.99 - 14.09 & A82, M90\
\
\
\
\
\
\
\
[**PKS 0118–272.**]{} Our H–band magnitude is slightly fainter than that found in the literature (Table 6). The host galaxy remains unresolved in the optical (Falomo 1996), however, this is consistent with the presence of a luminous host (M(R) = –23.5) at the proposed redshift of z$\geq$0.559. The host galaxy remains unresolved in the NIR.
[**PKS 0521–365**]{} is at redshift z = 0.055. Our H-band photometry agrees well with literature values (Table 6). The host galaxy has been imaged in the optical by Falomo (1994), who found best fit for a giant elliptical (M(R) = –23.2 and R(e) = 9 kpc) with a faint stellar disk, and by Wurtz (1996), who fit the host with an elliptical galaxy having M(R) = –23.0 and R(e) = 5.4 kpc. These values are in good agreement with those found in this study (M(H) = –25.4, R(e) = 5.8 kpc).
[**PKS 0537–441**]{} is a high redshift BL Lac (z = 0.896; Peterson 1976). Our H–band magnitude agrees well with previous studies (Table 6). The existence and nature of extended emission around this BL Lac, and its relevance to gravitational lensing, has been debated in the literature ( Stickel, Fried & Kühr 1988; Falomo, Melnick & Tanzi 1992). No nebulosity surrounding this high redshift BL Lac is detected in the NIR, in agreement with Falomo (1992), who conclude that there is no evidence for a foreground lensing galaxy.
[**PKS 0548–32**]{} is the dominant member of a rich cluster of galaxies at redshift z = 0.069 (Fosbury & Disney 1976). Our H–band photometry is slightly fainter than found in previous studies (Table 6). Falomo (1995) fit the host galaxy with a giant elliptical (M(R) = –24.2 and R(e) = 51 kpc) and a faint stellar disk. Wurtz (1996) found for the host galaxy an elliptical fit, with M(R) = –23.2 and R(e) = 14 kpc. The values derived by us in the NIR are M(H) = –25.7, and R(e) = 8.3 kpc. Note that this BL Lac is surrounded by an extended halo, which was detected in the optical by Falomo (1995), but not in the NIR. Therefore, the real colour of the host is redder than that derived from difference of total magnitudes in the two bands (see table 3).
[**PKS 1514–241 = AP Lib**]{} is at redshift z = 0.0486 (Disney, Peterson & Rodgers 1974). Our H–band magnitude agrees reasonably well with literature photometry (Table 6). The elliptical host galaxy has been studied extensively in the optical. Baxter (1987) obtained M(V) = –22.8, Abraham (1991) M(R) = –22.8 and R(e) = 7.5 kpc, and Stickel (1993) derived M(R) = –23.5 and R(e) = 11.5 kpc. Finally, we have analysed our unpublished optical images of this BL Lac, obtained at the ESO 2.2m telescope, and find M(R) = –22.9 and R(e) = 4.3 kpc. The host galaxy parameters derived by us in the NIR (M(H) = –25.1, R(e) = 3.3 kpc) are in reasonable agreement with those found in the previous studies.
[**PKS 1538+149**]{} is at redshift z = 0.605 (Stickel 1993). Our H-band photometry agrees well with those found in the literature (Table 6). Stickel (1993) could not resolve the host galaxy, while Wurtz (1996) found a marginal fit for the host, with M(R) = –24.2 and R(e) = 12 kpc. Falomo (1997b) derived for the host from HST imaging M(I) = –25.2. The host galaxy remains unresolved in the NIR.
[**PKS 2005–489**]{} is at redshift z = 0.071 (Falomo 1987). Our H–band magnitude agrees well with previous studies (Table 6). Stickel (1993) derived for the host galaxy M(R) = –24.2 and R(e) = 5.2 kpc, while Falomo (1996) found the host to be an elliptical with M(R) = –23.7 and R(e) = 11 kpc. The absolute magnitude derived by us in the NIR is: M(H) = –26.3.
[**MS 2143.4+07**]{} is at redshift z = 0.237. Our H–band photometry is slightly fainter than found in previous studies (Table 6). Wurtz (1996) derived for the elliptical host galaxy M(R) = –23.4 and R(e) = 12 kpc, while Jannuzi (1997) derived from HST imaging M(I) = –24.0 and R(e) = 9.0 kpc. These values are in excellent agreement with those derived by us in the NIR (M(H) = –25.9, R(e) = 5.4 kpc).
[**PKS 2155-305**]{} is at redshift z = 0.116 (Falomo, Pesce & Treves 1993) and it is one of the brightest and most studied BL Lacs and is often considered the prototype of X–ray selected BL Lacs. Our H–band magnitude is in good agreement with literature values (Table 6). For the host galaxy, Falomo (1991) derived M(R) = –24.4 and R(e) = 13 kpc, while Wurtz (1996) could not resolve the host, with M(R)$\geq$–23.1. The NIR properties of the host (M(H) = –26.8, R(e) = 5.7 kpc) are in good agreement with Falomo (1991). The companion galaxy at 4.2$''$ E of PKS 2155-305, previously detected in the optical (Falomo 1991), is clearly seen in the NIR image (see Fig. 1).
[**PKS 2254+074**]{} is at redshift z = 0.190 (Stickel 1993). We obtained only a short exposure of this source under poor sky conditions. We also note that the BL Lac was situated in a bad area of the array during the observations. Consequently, both the NIR photometry (slightly fainter than found in literature; Table 6), and the host properties are less accurate than for the other resolved BL Lacs. Optical determinations of the host galaxy have yielded M(R) = –24.1 and R(e) = 14.5 kpc (Stickel 1993), M(R) = –23.9 and R(e) = 15 kpc (Falomo 1996), M(R) = –23.9 and R(e) = 17 kpc (Wurtz 1996) and M(I) = –24.8 and R(e) = 15 kpc (Falomo 1997b). The absolute magnitude derived by us in the NIR (M(H) = –25.5) is $\sim$1 mag fainter than implied by the optical studies, assuming normal elliptical colours.
[^1]: Based on observations collected at the European Southern Observatory, La Silla, Chile.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'These notes show that on the moduli space of semi-stable vector bundles of fixed rank and determinant (of any degree) on a smooth curve of genus at least two, the base locus of the generalized theta divisor is large provided the rank is sufficiently large. It also shows that the base locus is large for positive multiples of the theta divisor. This work extends results already known for the case where the determinant is of degree zero.'
address:
- 'Sebastian Casalaina-Martin, Harvard University, Department of Mathematics, One Oxford Street, Cambridge MA 02138, USA'
- 'Tawanda Gwena, Mathematics Department, Tufts University, Medford MA 02155, USA'
- 'Montserrat Teixidor i Bigas, Mathematics Department, Tufts University, Medford MA 02155, USA'
author:
- 'Sebastian Casalaina-Martin'
- Tawanda Gwena
- Montserrat Teixidor i Bigas
title: Some examples of vector bundles in the base locus of the generalized theta divisor
---
[^1]
Introduction {#introduction .unnumbered}
============
Let $C$ be a smooth projective curve over $\mathbb C$ of genus $g(C)=g\ge 2$. For $r,d\in \mathbb Z$ and $r\ge 1$, there is the moduli space of $\mathscr U(r,d)$ of (equivalence classes of) semi-stable vector bundles on $C$ of rank $r$ and degree $d$, and for a line bundle $L\in \operatorname{Pic}^d(C)$, there is the subspace $S\mathscr U(r,L)\subset \mathscr U(r,d)$ of vector bundles with determinant $L$. Note that because of the isomorphisms given by tensoring with a line bundle of degree one $$\begin{matrix} \mathscr U(r,d)& \rightarrow &\mathscr U(r,d+r)\\
E&\rightarrow& E\otimes L\\
\end{matrix}$$ there are at most $r$ not necessarily isomorphic moduli spaces. Therefore, when considering a property of such moduli spaces, it suffices to consider the slope modulo integers. Similarly, in the case of fixed determinant, the moduli space is independent of the specific $L$ chosen and depends only on its degree.
Recall that there is an ample line bundle $\mathscr L$ on $S\mathscr
U(r,L)$, called the determinant line bundle, such that $\operatorname{Pic}(S\mathscr U(r,L))=\mathbb Z \langle\mathscr L
\rangle$ [@DN]. We will be interested in the base locus of the linear system $|\mathscr L^{\otimes n}|$ on $S\mathscr U(r,L)$, for positive integers $n$. When $d=0$, (or equivalently when $r$ divides $d$), a great deal is known about this base locus. In [@R], Raynaud provided a finite number of vector bundles in the base locus for very high rank with respect to the genus. Later, Popa [@P] showed that there are positive dimensional families of such examples. Some improvements on those results were obtained in [@A], [@gt]. Finally, [@H] gave, at least theoretically, a characterization of all the points in the base locus (see also [@H2]).
In sharp contrast with the degree zero case, not much has been done for other degrees, the main reason being that a characterization of fixed points did not exist so far. With the strange duality conjecture now proved (see [@mo]), the situation has changed. In fact, it was already remarked in [@P] that Raynaud’s examples would allow one to construct a few fixed points in these base loci, given the (at the time still conjectural) Strange Duality Theorem.
Our aim is to generalize these results in two directions: We prove that, for any slope, one can find a positive dimensional subset in the base locus of the theta divisor in suitable rank. We also show that the base loci of $|\mathscr L^{\otimes n}|$ with $n>1$ are often quite large.
\[teomain\] Fix a slope $\mu\in \mathbb Q$, and a level $n\in \mathbb N$. For $N\in \mathbb N$ and $L\in \operatorname{Pic}^{\mu N}(C)$, the base locus of the linear system $|\mathscr L^{\otimes n}|$ on $S\mathscr U (N,L)$ is large provided $N>>0$ and $g>>0$.
For more precise conditions on $N$ and $g$, and for estimates on the dimension of the base loci, see Theorem \[teo1\] and Theorem \[teo2\]. We point out that Theorem 8.1 in [@pr] states that on $S\mathscr U(N,\mathscr O_C)$, $\mathscr L^{\otimes n}$ is globally generated for $n\ge \lfloor\frac{N^2}{4}\rfloor$, and it has been conjectured to be for $n\ge N-1$. The bound obtained in these notes is far from this.
The method of proof is similar to that in Gwena-Teixidor [@gt] and relies on results from that paper, as well as on the Strange Duality isomorphism proven in Marian-Oprea [@mo]. The idea is to construct explicit examples of vector bundles in the base locus by taking wedge powers of certain vector bundles obtained as the kernel of an evaluation map. The main technical point in the paper is Proposition \[promainexa\] which uses a result of Russo-Teixidor [@rt] to give general conditions for this process to yield a point in the base locus.
Preliminaries
=============
Given $L\in \operatorname{Pic}^d(C)$ and fixing $r_F,d_F$ such that $\mu(\xi\otimes F)=g-1$ for $\xi\in S\mathscr U(r,L)$ and $F \in \mathscr U(r_F,d_F)$, then for each $F\in \mathscr U(r_F,d_F)$, there is a set $$\Theta_F=\{\xi\in S\mathscr U (r,L):h^0(\xi\otimes F)>0\}.$$ Note that the definition of $ \Theta_F$ depends on the choice of suitable $r,L$. For simplicity, we do not include a reference to them in the notation for $ \Theta_F$. We hope that this will no cause confusion.
For some $n\in \mathbb N$, $r_F=n\frac{r}{gcd(r,d)}$, and it is well known that if $\Theta_F$ is a divisor, then $\Theta_F\in |\mathscr L^{\otimes n}|$. The strange duality theorem (Theorem 4 [@mo] Marian-Oprea) implies that the linear system $|\mathscr L^{\otimes n}|$ is spanned by such divisors, and thus the base locus of the linear system is equal to the set $$\{\xi\in S\mathscr U(r,L):h^0(F\otimes \xi)>0 \ \forall \ F \in \mathscr U(r_F,d_F)\}.$$
In the following, for $\alpha\in \mathbb Q$, denote by $\lfloor \alpha \rfloor\in \mathbb Z$ the integral part of $\alpha$, and set $\lceil \alpha \rceil=\alpha-\lfloor \alpha \rfloor$. A basic observation is the following.
Let $L\in \operatorname{Pic}^d(C)$, and suppose that on $S\mathscr U(r,L)$ the base locus of $|\mathscr L^{\otimes n}|$ has dimension $b$. If $r'\ge r$, $L'\in \operatorname{Pic}^{d'}(C)$, and $\frac{d}{r}-\frac{d'}{r'}\in \mathbb Z$, then the base locus of $|\mathscr L|$ on $S\mathscr U(r',L')$ has dimension at least $$b+(r'-r)^2(g-1)+1-\delta_{r,r'},$$ where $\delta_{r,r'}$ is the Kronecker delta. The same is true of the base locus of $|\mathscr L^{\otimes n}|$.
As pointed out in the introduction, all results of this type if true for one slope, are true for all slopes which differ by an integer. In the case that $\frac{d}{r}=\frac{d'}{r'}$, the dimension estimate for the base locus of $|\mathscr L|$ follows by considering direct sums of vector bundles of slope $\frac{d}{r}$, and of the appropriate ranks (see [@gt] Corollary 5.3). The statement for the base locus of $|\mathscr L^{\otimes n}|$ follows in the same way. One must check that the numerical conditions in the lemma insure that the level of the theta divisor is unchanged.
Our goal will be to construct some explicit examples, and then use this lemma to generalize to other cases. In order to do this, we recall the set up and some of the results from Gwena-Teixidor [@gt]. Given a globally generated vector bundle $E$ on $C$, we define a vector bundle $M_E$ as the kernel of the evaluation map $$0\rightarrow M_E \rightarrow H^0(E)\otimes \mathscr O_C \rightarrow E \rightarrow 0.$$
Theorem 1.1 and 1.2 of [@gt] states that if $d_E-2r_Eg> -g$ and $0<i<d_E-r_Eg$, then the map $$S\mathscr U(r_E,L)\rightarrow S\mathscr U\left(\binom{d_E-r_Eg}{i},L^{\otimes
\binom{d_E-r_Eg-1}{i-1}}\right)$$ given by $E\mapsto \wedge^iM_E^\vee$ is generically finite. In these notes we will give some conditions that will insure that for each $E$, $\wedge^i M_E^\vee$ is in the base locus of $|\mathscr L^{\otimes n}|$ on $ S\mathscr U\left(\binom{d_E-r_Eg}{i},L^{\otimes
\binom{d_E-r_Eg-1}{i-1}}\right)$ for certain $n$.
Examples
========
General example
---------------
The proposition below is the main technical point in these notes. All of the examples that follow will be special cases of this one.
\[promainexa\] If $E\in \mathscr U(r_E,d_E)$, and $r_F,i\in
\mathbb N$ satisfy the following condtions:
1. $\frac{r_Fid_E}{d_E-r_Eg}\in \mathbb N,$
2. $\frac{g-1}{g}\cdot\frac{d_E-r_Eg}{r_Er_F}\ge i\ge g$,
then setting $d_F=r_F(g-1)-\frac{r_Fid_E}{d_E-r_Eg}$, $$\Theta_{\wedge^iM_E^\vee} :=\{F\in \mathscr U(r_F,d_F):h^0(F\otimes \wedge^iM_E^\vee)>0\}=\mathscr U(r_F,d_F).$$
Suppose that the conditions of the proposition are satisfied. To insure that $\wedge^iM_E^\vee$ is well defined, we must check that $d_E-r_Eg>i>0$. Since $\frac{g-1}{g}\cdot\frac{1}{r_Er_F}<1$, this condition follows from condition (2) above. Condition (1) and the definition of $d_F$ imply that for $F\in \mathscr U(r_F,d_F)$, $\mu(\wedge^iM_E^\vee
\otimes F)=g-1$.
By virtue of Proposition 5.1 of Gwena-Teixidor [@gt], given a generic effective divisor $D_i$ of degree $i$, there is an immersion ${\mathcal O}(D_i)\rightarrow \wedge^iM_E^{\vee}$. Hence, in order to prove the proposition, it is enough to show that a general vector bundle $F\in \mathscr U(r_F,d_F)$ can be written as an extension of vector bundles $$0\to \mathscr O_C(-D_i) \to F \to \overline{F} \to 0$$ where $D_i$ is a general effective divisor of degree $i$ on $C$. Since $i\ge g$, it follows that $\mathscr O_C(-D_i)$ is a general line bundle of degree $-i$. It is a result of Russo-Teixidor [@rt] that a general $F\in\mathscr U(r_F,d_F)$ can be written as such an extension so long as $$\mu(\overline{F})-\mu(\mathscr O_C(-D_i))\ge g-1.$$ A little arithmetic shows that condition (2) ensures that this is the case.
We can rephrase this in terms of the slope:
\[cormu\] If $E\in \mathscr U(r_E,d_E)$ with $\mu(E)=\left(\frac{a}{b}\right) g$ for some $a,b\in
\mathbb N$, and $r_F,i\in
\mathbb N$ satisfy the following condtions:
1. $r_Fi\frac{a}{a-b}\in \mathbb N$.
2. $(g-1)\frac{a}{b}\ge r_Fi\frac{a}{a-b}\ge r_F g\frac{a}{a-b}$,
then setting $d_F=r_F(g-1)-r_Fi\frac{a}{a-b}$, $$\Theta_{\wedge^iM_E^\vee} :=\{F\in \mathscr U(r_F,d_F):h^0(F\otimes \wedge^iM_E^\vee)>0\}=\mathscr U(r_F,d_F).$$
Clear.
For later reference, with the same assumptions as in the corollary, $\mu(\wedge^iM_E^\vee)=\frac{ia}{a-b} \equiv \frac{ib}{a-b} \pmod{\mathbb Z}$. Set $r=\operatorname{rank}(\wedge^iM_E^\vee)$, and $d=\deg(\wedge^iM_E^\vee)$. If follows that $$\frac{r}{\gcd(r,d)}=\frac{a-b}{\gcd(a-b,ia)}.$$ Set $$n= r_F\cdot \frac{\gcd(a-b,ia)}{a-b}.$$ For any $L\in
\operatorname{Pic}^d(C)$ and, as pointed out before, for an appropriate line bundle $\eta$ on $C$, it follows that $\wedge^i
M_{E\otimes \eta}^\vee\in S\mathscr U(r,L)$ is a base point for the linear system $|\mathscr L^{\otimes n}|$ on $S\mathscr U(r,L)$. In other words, with the assumptions above $$\wedge^i M_{E\otimes \eta}^\vee\in \Theta_F
\subseteq S\mathscr U(r,L)\ \forall \ F\in \mathscr U(r_F,d_F).$$
Some specific examples
----------------------
The following corollary shows that there are many ways to satisfy the conditions of the statements in the previous section.
Fix $r_E,A,b,i'\in \mathbb N$ satisfying the following conditions
1. $b|r_Eg$
2. $\frac{g-1}{b}\ge i'\ge \frac{g}{A}$
Set $d_E=\left(\frac{Ar_F+b}{b}\right)r_E g$, $i=i'A$ and $d_F=r_F(g-1)-i'(Ar_F+b)$. For $E\in \mathscr U(r_E,d_E)$, $$\mu(\wedge^iM_E^\vee)=i'A+\frac{i'b}{r_F}\equiv \frac{i'b}{r_F}\pmod{\mathbb Z},$$ and $$\Theta_{\wedge^iM_E^\vee} :=\{F\in \mathscr U(r_F,d_F):h^0(F\otimes \wedge^iM_E^\vee)>0\}=\mathscr U(r_F,d_F).$$
Set $a=Ar_F+b$ in the previous corollary.
This proves the following proposition
\[mainexa\] Fix $r_E,r_F,A,b,i'\in \mathbb N$ satisfying the following conditions
1. $b|r_Eg$
2. $\frac{g-1}{b}\ge i'\ge \frac{g}{A}$
3. $r_F\ge \frac{b}{A}$
Set $n=\gcd(r_F,i'b)$, $r=\binom{\frac{Ag}{b}r_Fr_E}{i}$, $d=r(g-1)-\binom{\frac{Ag}{b}r_Fr_E}{i}\left(i'A+\frac{i'b}{r_F}\right)$, and fix a line bundle $L\in \operatorname{Pic}^d(C)$. On $S\mathscr U(r,L)$ the dimension of the base locus of $|\mathscr L^{\otimes n}|$ is at least $$(r_E^2-1)(g-1)$$
Assumption (3) in the proposition insures that the map $$S\mathscr U(r_E,N)\to S\mathscr U(d_E-r_Eg,N^\vee)$$ taking $E\mapsto M_E$ is generically finite (Gwena-Teixidor i Bigas [@gt]).
Base loci theorems
------------------
\[teo1\] Let $\mu\in \mathbb Q$, and suppose there exists numbers $i',r_F\in \mathbb N$ with $1\le i'\le g-1$ such that $$\mu\equiv \frac{i'}{r_F} \pmod{\mathbb Z}.$$
Suppose that $N,r_E\in \mathbb N$, $r_E\ge 2$, and $$N\ge N_0:=\binom{r_Fr_Eg^2}{i'g}.$$
For $L\in \operatorname{Pic}^{\mu N}(C)$, on $S\mathscr U (N,L)$, the base locus of the linear system $|\mathscr L|$ has dimension at least $$((r_E^2-1)+(N-N_0)^2)(g-1)+1-\delta_{N,N_0}.$$ The same can be said for the linear system $|\mathscr L^{\otimes
\gcd(r_F,i')}|$.
Take $A=g$ and $b=1$ in Proposition \[mainexa\].
For a fixed curve, the examples in the theorem have slopes in a range determined by the genus. On the other hand, for any given slope $\mu\in \mathbb Q$, taking a curve of high enough genus, the theorem implies that the base locus is nonempty provided the rank is sufficiently large. In particular, Theorem \[teomain\] follows from Theorem \[teo1\].
Observe also that in the notation of Corollary \[cormu\], condition (2) implies that $$\frac{g-1}{r_F}\ge\frac{ib}{a-b} \equiv \mu(\wedge^i
M_{E}^\vee) \pmod{\mathbb Z}.$$ This shows that the special cases given in the theorem above are indicative of the types of examples one expects to get using Proposition \[promainexa\].
Fixing slope zero, we can give examples with lower rank and higher level.
\[teo2\] Suppose that $N,r_E\in \mathbb N$, $r_E\ge 2$, and $$N\ge N_0:=\binom{2g(g-1)r_E}{2(g-1)}.$$ Then on $S\mathscr U (N,\mathscr O_C)$, the base locus of the linear system $|\mathscr L^{\otimes g-1}|$ has dimension at least $$((r_E^2-1)+(N-N_0)^2)(g-1)+1-\delta_{N,N_0}.$$
Take $A=2$, $b=1$, and $i'=r_F=g-1$ in Proposition \[mainexa\].
[99]{} D.Arcara, *A lower bound for the dimension of the base locus of the generalized theta divisor.* C. R. Math. Acad. Sci. Paris 340 (2005), no. 2, 131–134.
J.M.Drezet, M.S. Narasimhan, *Groupe de Picard des variétés de modules de fibres semi-stables sur les courbes algebriques* Invent. Math. [**97**]{} (1989), no. 1, 53–94.
T. Gwena and M. Teixidor i Bigas, *Maps between moduli spaces of vector bundles and the base locus of the theta divisor* arXiv:0706.3953v1
G.Hein, *Raynaud’s vector bundles and base points of the generalized theta divisor* arxiv mathAG0607338
G.Hein, *Raynaud vector bundles*. arXiv:0706.3970v1
A. Marian and D. Oprea, *The level-rank duality for non-abelian theta functions*. Invent. Math. 168 (2007), no. 2, 225–247.
M. Popa, *On the base locus of the generalized theta divisor*. C. R. Acad. Sci. Paris Sér. I Math. 329 (1999), no. 6, 507–512
M. Popa and M. Roth, *Stable maps and Quot schemes*. Invent. Math. 152 (2003), no. 3, 625–663.
M.Raynaud, *Sections des fibrés vectoriels sur une courbe.* Bull. Soc. Math. France 110 (1982), no. 1, 103–125.
B. Russo and M. Teixidor i Bigas, *On a conjecture of Lange*. J. Algebraic Geom. 8 (1999), no. 3, 483–496.
[^1]: The first author was partially supported by NSF MSPRF grant DMS-0503228
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Optical communication systems, operating in C–band, are reaching their theoretically achievable capacity limits. An attractive and economically viable solution to satisfy the future data rate demands is to employ the transmission across the full low–loss spectrum encompassing O, E, S, C and L band of the single mode fibers (SMF). Utilizing all five bands offers a bandwidth of up to $\sim$53.5 THz (365 nm) with loss below 0.4 dB/km. A key component in realizing multi–band optical communication systems is the optical amplifier. Apart from having an ultra–wide gain profile, the ability of providing arbitrary gain profiles, in a controlled way, will become an essential feature. The latter will allow for signal power spectrum shaping which has a broad range of applications such as the maximization of the achievable information rate $\times$ distance product, the elimination of static and lossy gain flattening filters (GFF) enabling a power efficient system design, and the gain equalization of optical frequency combs. In this paper, we experimentally demonstrate a multi–band (S+C+L) programmable gain optical amplifier using only Raman effects and machine learning. The amplifier achieves $>$1000 programmable gain profiles within the range from 3.5 to 30 dB, in an ultra–fast way and a very low maximum error of $1.6 \cdot 10^{-2}$ dB/THz over an ultra–wide bandwidth of 17.6–THz (140.7–nm).'
author:
- 'Uiara Celine de Moura, Md Asif Iqbal, Morteza Kamalian, Lukasz Krzczanowicz, Francesco Da Ros, Ann Margareth Rosa Brusin, Andrea Carena, Wladek Forysiak, Sergei Turitsyn, and Darko Zibar [^1] [^2] [^3]'
bibliography:
- 'IEEEabrv.bib'
- 'bibliography.bib'
title: 'Multi–band programmable gain Raman amplifier'
---
optical communications, multi–band systems, optical amplifiers, machine learning, neural networks.
Introduction {#sec:intro}
============
the past two decades, a great evolution of optical communication systems, in terms of spectral efficiency$\times$distance product, has been enabled by the advances in digital coherent detection. So far, most of the efforts, on reaching the capacity of the nonlinear fiber–optic channel, have been focusing on the C–band only [@agrell2016roadmap]. However, squeezing the information inside this transmission window will soon reach its theoretical limit [@Dar14]. To cope with the constant demand for higher throughput, novel solutions must be explored.
Optical communication systems operating across multi–band transmission, are an attractive solution for providing the future capacity scaling. They can provide up to 10$\times$ higher capacity, compared to the C–band [@Ferrari20], on the already deployed SMF fiber infrastructure. To make multi–band systems commercially deployable in the near future, large research efforts in terms of components, system and network design are needed [@ciena; @infinera; @Doerr2016; @Timurdogan19; @Messner20; @Tummidi20; @Sarwar20; @Wang20].
One of the main challenges in realizing multi–band systems is the development of optical amplifiers that are able to provide sufficiently high gains over such a wide bandwidth. Additionally, a novel feature that may become essential is the ability to provide arbitrary gain profiles in a controlled and ultra–fast way. This is because different signal channels in a multi–band system are unevenly impacted by the interaction between the Kerr nonlinearity, amplified spontaneous emission (ASE) noise and stimulated Raman scattering (SRS) [@Ferrari20]. Consequently, for the maximization of the achievable information rate (AIR) $\times$ distance product, non–flat signal channel power profiles are needed. Depending on the system configuration, signal channel power profiles will be a result of a complex optimization and may assume arbitrary shapes. Moreover, to address the future requirements on high capacity optical networks, ultra–fast gain profile re–configurability is needed [@Napoli18].
A current and by far the most dominant approach for performing programmable signal channel power profile shaping is by leveraging the use of wavelength selective switches (WSSs) whose primary function is to router the signals throughout the optical network. However, this approach is highly power inefficient since it adjusts the channel powers by means of attenuation.
A novel approach for realizing signal channel power shaping is by employing optical amplifiers with programmable (arbitrary) gain profiles. Such amplifiers could be a potential game changer as they would be able to simultaneously amplify the optical data signal and perform gain shaping. This has many impactful applications ranging from compensation of wavelength–dependent loss in devices such as modulators and frequency combs to gain–shaping in fixed-gain profile amplifiers. Especially, if integrated–combs are targeted for multi-channel sources, an efficient approach for gain shaping would be desirable. This is because for integrated–combs there is a large variation in power of their frequency components. Finally, optical amplifiers providing arbitrary gain profiles can be used in hybrid approaches to complement the gain, and overcome the limitations of other optical amplifier technologies [@Fukuchi01; @Gordienko16; @Ionescu19; @Galdino19; @Arnould20; @Ye20].
There are several approaches and technologies for realizing optical amplifiers covering multiple bands. To date, works on multi–band optical amplifiers have focused on: rare–earth–doped fiber amplifiers (xDFAs) covering 17.56 THz over O+E–band [@Wang20] and 10.7 THz over S+C–band [@Sakamoto06], semiconductor optical amplifiers (SOAs) for 12.7 THz on S+C+L–band [@Renaudier18], optic parametric amplifiers (OPAs) with 10 THz of bandwidth on S+C+L–band [@Kobayashi20], Raman amplifiers (RA) in combination with EDFAs, SOAs and OPAs achieving bandwidths ranging from 10.7 to 14 THz on C+L and S+C+L–band [@Fukuchi01; @Ye20; @Ionescu19; @Galdino19; @Arnould20; @Gordienko16], and pure RAs with bandwidths of up to 19.1–THz S+C+L–band [@Rottwitt99; @Zhou06; @Chen18; @Emori01; @Iqbal20]. So far, the majority of works in [@Fukuchi01; @Sakamoto06; @Wang20; @Renaudier18; @Kobayashi20; @Gordienko16; @Ionescu19; @Galdino19; @Arnould20; @Rottwitt99; @Zhou06; @Chen18; @Emori01; @Iqbal20] have focused on realizing flat gain profiles in C+L and S+C+L–band. Recently, an amplifier that relies on a hybrid SOA/Raman configuration has been demonstrated to achieve arbitrary loss/gain profile generation in S+C+L–band in 12.3 THz of bandwidth [@Ye20].
Among all different solutions, RAs are most suitable for realizing arbitrary gain profiles, in a controlled way. This is because the RAs allow for a flexible gain profile design by adjusting the pump powers and wavelengths, and provide gain availability across a broad range of wavelengths, when operated in multi-pump configurations.
The challenge with Raman amplifier design is on the selection of pump powers and wavelengths that would result in a targeted gain profile. Several solutions to this optimization problem have been reported in the literature but have mainly focused on realizing flat gain profiles [@Ferreira11; @Zhou01; @Chen18; @Perlin02; @Iqbal20; @Mowla08; @Jiang10; @Emori01; @Ania07]. Recently, a machine learning framework for the ultra–fast configuration of the pump powers and wavelengths has been theoretically proposed and as a proof–of–principle experimentally demonstrated in C–band only [@Zibar20; @deMoura20]. The proposed approach can be used for the design of Raman amplifiers, where an arbitrary gain profile is achievable in a controlled way. However, moving from C–band to multi–band and realizing wider gain profiles is significantly more challenging. This is partly due to the increased number of pumps that need to be controlled and also the increased nonlinearity given the higher overall powers in the optical fiber.
In this paper, we use the proposed machine learning framework for the experimental realization of multi–band RAs that can provide arbitrary gains, in a controlled way, in C+L and S+C+L–band. Up to 8 pumps are employed to provide more than 5000 arbitrary gain profiles over up to 17.6-THz of bandwidth. We achieve a highly–accurate programmable set of gain profiles with a very low average maximum error, (defined between the target and realized gain profiles), per bandwidth, $E_{MAX}/BW$, of $1.6 \cdot 10^{-2}$ dB/THz.
Machine learning for broadband gain optimisation is a topic of growing interest, which is reflected in the recent work [@Ye20] reporting a root mean squared error per bandwidth, $RMSE/BW$, of 0.033 dB/THz in a 12.3 THZ bandwidth SOA/distributed Raman link scenario. In this work, we achieve a close to an order of magnitude lower $RMSE/BW$ of 0.0045 dB/THz for transmission with discrete S+C+L–band Raman amplifier over a larger bandwidth of 17.6 THz. The presented work is first time demonstration of a programmable multi–band discrete Raman amplifier setting a benchmark, to be beaten, for the maximum error per bandwidth.
The structure of the paper is as following: Section \[sec:exp\_setups\] describes the experimental setup for realizing Raman amplifiers operating in C+L and S+C+L–band. We also give a brief overview of the ML framework used to obtain programmable arbitrary gain profiles. Section \[sec:results\] presents, discusses and evaluates the experimental results. In Section \[sec:conc\] conclusions and future work are presented and outlined.
Experimental setup {#sec:exp_setups}
==================
{width="100.00000%"}
The experimental setup for realizing the multi–band RA is shown in Fig. \[fig:setup\](a). By selecting path 1 or 2, the operation in either C+L (1) or S+C+L–band (2) can be enabled. To achieve gains in the C+L and S+C+L–band, 5 and 8 pump lasers are employed, respectively. Fig. \[fig:setup\](c) illustrates the spectral pump allocation and their individual contribution to the overall Raman gain. We only consider counter propagating pumps whose wavelengths are fixed and shown in Table \[tab:pumps\].
$P_1$ $P_2$ $P_3$ $P_4$
------------------- ------- ------- ------- -------
Wavelength \[nm\] 1508 1485 1465 1445
Frequency \[THz\] 198.8 201.9 204.6 207.5
$P_5$ $P_6$ $P_7$ $P_8$
Wavelength \[nm\] 1425 1405 1385 1365
Frequency \[THz\] 210.4 213.4 216.5 221.1
: **Pump lasers wavelengths and frequencies**
\[tab:pumps\]
The gain profile control is performed by only adjusting the pump powers. Pump lasers $P_1...P_7$ are semiconductor laser diodes. Their output power is controlled by adjusting the driving currents. The corresponding power going into the RA is in the range from $\sim$16 dBm to $\sim$27 dBm. Pump laser $P_8$ is a Raman–based fiber laser and is controlled by adjusting its voltage. It provides power to the RA ranging from $\sim$20 dBm to $\sim$27 dBm.
C+L–band Raman amplifier
------------------------
The C+L–band RA can either be operated as a discrete, (7.5 km of inverse dispersion fiber (IDF)) or distributed (75 km span of standard SMF) amplifier. An input optical signal covering the C+L–band, for testing the performance of the RA, is generated by using two ASE sources for C and L bands channelized through a WSS to generate 90 lines placed at 100 GHz ITU-T grid covering a 9.4 THz (77 nm) bandwidth. The corresponding optical spectrum is shown in Fig. \[fig:setup\](b) (inside bracket 1) and is measured with a resolution of $\Delta \lambda = 0.1$ nm. The gaps between the C and L signal bands are due to the different ASE sources for these two bands. An isolator is placed at the input to the IDF to prevent pump powers entering the C+L–band signal source and to minimize the double Rayleigh backscattering induced multipath interference [@Iqbal19JLT]. Finally, an optical spectrum analyser (OSA) is used to capture the optical spectrum.
S+C+L–band Raman amplifier
--------------------------
The S+C+L–band RA is implemented as a two–stage sequential discrete RA. The first stage is responsible for providing the gain in the S–band and it consists of 7.5 km of IDF and three pump lasers, $P_6...P_8$ used to control the gain profiles. The second stage is the same as the one used for the C+L–band RA. Note that distributing the pumps into two sequential stages reduces the strong depletion of shorter wavelength pumps [@Krummrich01]. The multi–band input optical signal (17.6 THz/140.7 nm) is generated by combining the optical signal from the C+L–band with a supercontinuum S–band source [@El-Taher:s] and a single frequency laser operating at 185 THz. The resulting signal has a total of 148 frequency lines at 100 GHz ITU-T grid. The corresponding optical spectrum is shown in Fig. \[fig:setup\](b) (inside bracket 2). Due to the amplifier configuration, two pumps from the first stage ($P_{1-2}$) fall within the S-band signal. This means that some channels from the S–band need to be removed to avoid overlapping with the Rayleigh backscattered components of the pumps, leaving the gaps as shown in Fig. \[fig:setup\](b) [@Iqbal:20].
Pump power control
------------------
The objective is to determine pump power settings that result in user defined target gain profiles such as: tilted gain, flat gain or an arbitrary gain. These settings are achieved off–line using the machine learning framework presented and then later applied on–line for the pump laser currents and voltage control [@Zibar20]. As the framework in [@Zibar20] is based on supervised learning, a data–set is required. This is achieved by varying the currents and the voltage of the pump lasers and measuring the corresponding gain profiles. The gain profiles are measured on a 100-GHz grid, as the difference in power between the output optical spectrum when the pump lasers are turned on and off, also known as the on–off gain. As the currents and the voltage, $I_1,...I_7,V_8$, are drawn from a uniform distribution, we refer to the corresponding gain profiles as arbitrary. In Fig. \[fig:datasets\], the measured on–off gain profiles, $G$, obtained for the C+L and the S+C+L–band are shown. We measure $M=5600$ and $M=4025$ gain profiles, each with $K=90$ and $K=148$ data points per gain profile, for C+L and S+C+L–band, respectively. We denote the respective data–sets as: $\mathcal{D}^{M\times (K+5)}_{C+L}=\{(G_1^i,...,G_K^i,I^i_1,...I^i_5),|i=1,...,M\}$ and $\mathcal{D}^{M\times (K+8)}_{S+C+L}=\{(G_1^i,...,G_K^i,I^i_1,...I^i_7,V^i_8),|i=1,...,M\}$.
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
To find the machine learning model with the lowest prediction error, we allocate 3400 and 3000 data points, for C+L and S+C+L–band, correspondingly. We employ 10–fold cross–validation, which means that we use 90% for training (includes hyperparameter optimization) and 10% for testing as described [@Zibar20; @bishop2006]. For a more detailed explanation on the training of the employed machine learning model, see the Appendix Section. The remaining data points are later used for the final validation of the machine learning model for the pump laser current prediction of arbitrary gains.
The procedure of obtaining pump current configuration is then as follows: 1) a single-layer neural network, $NN_{inv}$, is employed to learn the mapping between the target gain profiles and currents and voltage – inverse system learning, 2) once the neural network has learned the inverse mapping, given a set of target gain profiles, the corresponding pumps currents and voltages are predicted, 3) the predicted currents and voltages are then applied to the second multi–layer neural network, $NN_{fwd}$, that has learned the forward mapping between pump currents/voltage and gain profiles. The $NN_{fwd}$ thereby predicts the gain profile given the pump currents and the voltage. If the error between the predicted and targeted gain profile is not satisfactory pump currents and voltages are adjusted accordingly, i.e. fine–optimization. The fine–optimization uses iterative gradient descent by backpropagating the error through $NN_{fwd}$ to adjust the currents and voltage as described in [@Zibar20], 4) the obtained currents and voltages are and applied to the pump lasers in the experimental set–up, and new sets of measurements are performed, and 5) finally, to investigate the accuracy of the predicted pump currents and voltage, we calculate the maximum absolute error between the target and the newly measured gain profiles (i.e. $E_{MAX}$) and normalize it with the bandwidth ($BW$). The optimized topologies of the employed neural networks $NN_{fwd}$ and $NN_{dir}$, as well as their performance evaluation, are found in the Appendix Section.
Results and discussion {#sec:results}
======================
Arbitrary gain profiles
-----------------------
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
Fig. \[fig:results\_arb\](a)–(c) show the probability, (PDF), and the cumulative, (CDF), density functions of the $E_{MAX}/BW$ for the C+L–band (distributed and discrete) and S+C+L–band (discrete) Raman amplifiers. The error is defined between the targeted arbitrary gain profiles, taken directly from the data–set (not used for training the machine learning framework), and the predicted gain profiles obtained from the measurement using the pump currents and voltage allocation provided by the machine learning framework. We use 2100, 2600 and 1025 target arbitrary gain profiles for the distributed C+L–band, discrete C+L–band and discrete S+C+L–band validation, respectively. We compare the accuracy of allocating pump currents and voltage, by using only the inverse mapping multi–layer neural network, $(NN_{inv})$, and both the inverse and forward mapping multi–layer neural networks, $(NN_{inv}+NN_{fwd})$, which allows for fine–optimization of pump currents and the voltage.
{width="32.00000%"} {width="32.00000%"} {width="33.00000%"}
The PDFs shown in Fig. \[fig:results\_arb\](b)–(c), illustrate that for the discrete RA, highly–accurate pump current predictions, resulting in a low mean and standard deviation, can be obtained using only $NN_{inv}$. Thus, the currents and the voltage prediction is obtained in an ultra–fast way as $NN_{inv}$ only involves matrix computations. We notice that the mean and standard deviations are decreased by a factor of $\sim$2 when going from C+L to S+C+L–band. This is mainly because these two schemes have the same performance in terms of $E_{MAX}$ and S+C+L–band has almost two times wider bandwidth. However, qualitatively the results for C+L and S+C+L–band are comparable.
If $NN_{inv}+NN_{fwd}$ is used a slight increase in the mean and the standard deviation is observed. This is because the $NN_{inv}$ has already found pump current configuration that minimizes the mean square error. Applying the fine–optimization introduces some small random deviations around this minimum and worsens the performance.
For both discrete RA schemes, the CDF shows that most of the cases already present an $E_{MAX}/BW$ lower than $6\cdot10^{-2}$ dB/THz, before the fine–optimization, i.e. 97% of the cases for the C+L–band and $\sim$100% for the S+C+L–band.
Compared to the discrete RA, the resulting PDF for the distributed RA (Fig. \[fig:results\_arb\](a)) has a higher mean and standard deviation when considering only $NN_{inv}$. On the other hand, a significant reduction can be obtained after applying fine–optimization $NN_{inv}+NN_{fwd}$, as also illustrated by the CDF. Indeed, the fine–optimization significantly increases the number of cases with $E_{MAX}/BW$ lower than $6\cdot10^{-2}$ dB/THz, i.e. from 18.7% to 95.4%.
Finally, in Fig. \[fig:results\_arb\_rmse\](a)-(c), the resulting PDF and CDF of the RMSE per bandwidth is plotted for the distributed and discrete amplifiers. The Figure shows that very low mean and standard deviation values are achievable.
Flat and tilted gain profiles
-----------------------------
Next, we investigate the ability of the machine learning framework to predict accurate pump current and voltage allocations for the design of flat and tilted gain profiles using the discrete and distributed RAs, in C+L and S+C+L–band. Flat gains ranging from 6 to 12 dB (C+L–band distributed RA), 7 to 15 dB (C+L–band discrete RA), and 14 to 20 dB (S+C+L–band discrete RA) are evaluated in steps of 1 dB. For the tilted profiles, slopes of approximately 0.24 dB/THz (C+L–band RAs) and 0.20 dB/THz (S+C+L–band RA) are considered. These values were chosen to provide an overall tilt of around 1 dB on each band.
Fig. \[fig:results\_flat\_tilted\], shows the predicted and target flat ((a)-(c)) and tilted ((d)-(f)) gain profiles, as a function of frequency, for the distributed and the discrete RA operating in C+L and S+C+L–band. Just a subset of gains (2 dB step) are shown for better visualization. The corresponding $E_{MAX}/BW$ for all gains under consideration is shown in Fig. \[fig:results\_flat\_tilted\](g)-(i). We only show results obtained after using $NN_{inv}+NN_{fwd}$ as the fine–optimization significantly reduced the error for all the amplifier schemes and their evaluated gains.
A general trend observed in Fig. \[fig:results\_flat\_tilted\](a)–(f), is that the predicted gain oscillates around the target gain profile. The magnitude of the oscillations has a tendency to increase for increasing gains. Moreover, for the S+C+L–band RA, the oscillation amplitude increases with the frequency, achieving up to 2 dB of maximum error compared to the target.
To understand what is happening, it is worth mentioning that it was observed some power instabilities on the supercontinuum S–band source and the Raman-based fiber laser $P_8$. Additionally, recall that the broadband and nonuniform Raman gain spectrum for a single pump, with a peak located near 12.5 THz below the pump frequency for the IDF, is partially overlapped in the multiple-pump configurations considered in this work as illustrated in Fig. \[fig:setup\](c). On the S–band, besides pumps $P_{6-8}$, there are also contributions of pumps $P_{1-5}$ because the S–band lies within the Raman gain spectrum bandwidth of all these pumps. This makes the design more complex on this region. Thus, although it is expected that the machine learning framework is able to deal with these broadband effects when adjusting the pumps (once the two stages on the S+C+L–band discrete RA are jointly trained), it is also expected to achieve a higher error on the S–band.
It is observed in Fig. \[fig:results\_flat\_tilted\](h)-(i) that the $E_{MAX}/BW$ for the discrete RA in C+L and S+C+L–band is similar for the flat and the tilted gain profiles. The $E_{MAX}/BW$ is kept below $1.1\cdot10^{-1}$ and $0.9\cdot10^{-1}$ dB/THz for the design of flat and tilted gain profiles, respectively. On the other hand, the $E_{MAX}/BW$ for the distributed RA shown in Fig. \[fig:results\_flat\_tilted\](g) is higher for the design of the flat gains, but it is still kept below $1.3\cdot10^{-1}$ dB/THz. The reason may be related to the pump distributions, i.e. the number of pumps and wavelength being more suitable to provide a tilted gain profile. This can be observed on the experimental data–set gain profiles shown in Fig. \[fig:datasets\]. The same analysis does not apply for the S+C+L–band, since there it no clear flat/tilted profile trend on its data–set gain curves. Therefore, we also need to take into account that there will be a limitation on the theoretically achievable gain tilt and flatness given experimental set–up that has fixed wavelengths of pump lasers. Fig. \[fig:results\_flat\_tilted\](j)-(l) shows $RMSE/BW$ and it observed that the trends are very similar to as for $E_{MAX}/BW$.
To put the presented work in the perspective, in Fig. \[fig:record\], $E_{MAX}/BW$, is plotted for various experimental demonstrations of multi–band amplifiers. It is observed that the presented work results in a low–error and broad bandwidth by means of machine learning.
![$E_{MAX}/BW$ as a function of amplifier bandwidth.[]{data-label="fig:record"}](figures/record_17_6THz_with_refs.pdf){width="47.00000%"}
Conclusion {#sec:conc}
==========
A multi–band programmable gain Raman amplifier operating in C+L and S+C+L–band is experimentally demonstrated. The key enabling technique is the machine learning framework that allows for ultra–fast and highly–accurate prediction of the pump currents and voltage for providing the targeted gain profiles. The ability to generate arbitrary gain profiles in a controlled and fast way, may provide novel approaches for the intelligent utilization of the ultra–wideband spectrum and become a key feature for future optical communication systems. Moreover, the programmable gain optical amplifier may advance other areas of fundamental science requiring spectral shaping, such as optical frequency combs.
Appendix {#appendix .unnumbered}
========
The machine learning framework used in this paper to achieve highly accurate Raman amplifier (RA) programmable gains is based on two artificial neural networks. The first neural network $NN_{inv}$ models the RA inverse mapping, i.e. the mapping between gain profiles and pump lasers’ currents/voltage. Whereas the forward mapping, i.e., the mapping between the pump lasers’ currents/voltage and gain profiles, is learned by a second neural network $NN_{fwd}$. Here, we describe how these two NNs are trained for the different RA schemes considered in this paper. We also validate their prediction accuracy. Training and validation are performed on disjoint experimental data–sets, whose total number of elements are shown in Table \[tab:datasets\].
RA scheme C+L dist. C+L disc. S+C+L disc.
------------ ----------- ----------- -------------
Training 3464 3000 3000
Validation 2100 2600 1025
: **Experimental data–set distribution**
\[tab:datasets\]
Neural networks training {#sec:training}
------------------------
$NN_{inv}$ is trained using random projection (RP). This training algorithm, also known as extreme learning machine (ELM) [@Huang2011], initializes the weights of the hidden layers randomly, according to a normal distribution with mean zero and a certain standard deviation $\sigma_{NN_{init}}$, corresponding to NN initialization variance. This random weight assignment is independent from the training data–set and requires a high number of hidden nodes as these weights are kept untrained. The training data–set is used to optimize only the last layer weight by regularized least squares, with a regularization parameter $\lambda$. Since it is performed in a single step, the training time is drastically reduced when compared to standard approaches that updates all the weights in a numerical interactive routine. $NN_{inv}$ models for each RA scheme are shown in Table \[tab:NNinv\], where $f_{act}$ is the nonlinear activation function for all nodes (except the ones on the last layer, which use linear functions), $numHL$ is the number of hidden layers, and $numHN$ is the number of hidden nodes. To reduce the impact of the randomly initialized weights on the RP method, 20 parallel and independent $NN_{inv}$ are trained and the pump configuration prediction is the average of the 20 $NN_{inv}$ outputs [@Zibar20]. In Table \[tab:NNinv\], $f_{act}$, $numHN$, $\sigma_{NN_{init}}$ and $\lambda$ were obtained after a hyperparameter optimization routine using k-fold cross validation [@bishop2006].
RA scheme C+L dist. C+L disc. S+C+L disc.
---------------------- --------------------- --------------------- ---------------------
Training alg. RP RP RP
$f_{act}$ logsig sine sine
$numHL$ 1 1 1
$numHN$ 760 500 500
$\sigma_{NN_{init}}$ $6.0 \cdot 10^{-3}$ $2.6 \cdot 10^{-2}$ $1.0 \cdot 10^{-2}$
$\lambda$ $1.0 \cdot 10^{9}$ $1.0 \cdot 10^{3}$ $1.0 \cdot 10^{4}$
: **Neural network models for $NN_{inv}$**
\[tab:NNinv\]
$NN_{fwd}$ is trained differently for each RA scheme. For the C+L–band RA (discrete and distributed), $NN_{fwd}$ is trained traditionally updating all weights on the NN interactively by using the Levenberg-Marquadt (LM) method. However, the high input and output dimensions of the S+C+L–band RA scheme makes the use of LM optimization challenging due to the long convergence time. Thus, RP is applied again only for this scheme. Table \[tab:NNdir\] summarizes $NN_{fwd}$ parameters for each RA scheme, where only the RP parameters $f_{act}$, $numHN$, $\sigma_{NN_{init}}$ and $\lambda$ were obtained after a hyperparameter optimization routine. Table \[tab:NNdir\] also shows that the RP faster training comes with the cost of having a larger network, with 500 hidden nodes instead of 20 when using LM.
RA scheme C+L dist. C+L disc. S+C+L disc.
---------------------- ----------- ----------- ---------------------
Training alg. LM LM RP
$f_{act}$ tanh tanh tanh
$numHL$ 2 2 1
$numHN$ 10 10 500
$\sigma_{NN_{init}}$ \* \* $1.0 \cdot 10^{-3}$
$\lambda$ 0 0 $1.0 \cdot 10^{8}$
: **Neural network models for $NN_{fwd}$**
\[tab:NNdir\]
Neural networks validation
--------------------------
$NN_{inv}$’s performance in predicting pump currents/voltage is presented in Fig. \[fig:NNinv\]. The metric used is the absolute error relative to the maximum current/voltage excursion for each pump laser. Fig. \[fig:NNinv\] shows the probability density functions (PDF) and the cumulative density functions (CDF) over all the cases on the validation data–set and all pump lasers. Notice that the errors are kept bellow 2% for 95% of the cases for all the RA schemes. The prediction performance for the $NN_{fwd}$ is evaluated in terms of root mean squared error ($RMSE^{P}$) and maximum absolute error ($E^{P}_{MAX}$) between predicted $G^P$ and target $G^T$ gain profiles, extracted from the $K$ WDM points (spectrum), given by $$RMSE^{P} = \sqrt{\frac{1}{K}\sum_{i=1}^{N}(G_i^P-G_i^T)^2},
\label{eq:mse}$$ $$E^{P}_{MAX} = max\{|G_1^P-G_1^T|, |G_2^P-G_2^T|, \cdots, |G_N^P-G_N^T| \}
\label{eq:maxE}$$ we use index $P$ for prediction to differentiate from the experimental validation errors shown in Section \[sec:results\] and recall that $K=90$ and $K=148$ for C+L and S+C+L-band RAs, respectively. Fig. \[fig:NNdir\] shows the PDF for $RMSE^{P}$ and $E^{P}_{MAX}$ over all the cases on the validation data–set.
In Fig. \[fig:NNdir\], the overall $NN_{fwd}$ performances for both C-L–band RAs are consistent with the ones obtained in [@unpublishedBrusinJLT2020], which also considers a C+L–band RA (distributed scheme only) with same NN model and training algorithms. On the other hand, the worst performance obtained here by the S+C+L–band RA scheme in terms of $E^{P}_{MAX}$ can be explained by its more complex mapping relating more pumps to the gain over a wider bandwidth. S+C+L–band RA scheme was also the only model that used RP, but the same study presented in [@unpublishedBrusinJLT2020] showed that, for the Raman amplifier case, the performance of the LM only overcomes the RP for higher number of hidden nodes, which requires even more time to train.
![Probability density function (PDF) and cumulative density function (CDF) of the $NN_{inv}$ pump current/voltage prediction error, with indication of mean, $\mu$ and standard deviation, $\sigma$.[]{data-label="fig:NNinv"}](figures/inverseNN_all.pdf){width="40.00000%"}
![Probability density function (PDF) of the $NN_{fwd}$ gain prediction error: (a) $RMSE^{P}$ and (b) $E_{MAX}^{P}$ with indication of mean, $\mu$, standard deviation, $\sigma$, and maximum (max) values.[]{data-label="fig:NNdir"}](figures/directNN_all_rmse_maxE_test_set.pdf){width="40.00000%"}
The errors $RMSE^{P}$ and $E^{P}_{MAX}$ are non-convex and unknown functions of the pump configuration that might not share the same local minimums, i.e. the pump configuration that minimizes $RMSE^{P}$ might not minimize $E^{P}_{MAX}$. However, since the fine–optimization is a gradient-based procedure, it needs to use a differentiable cost function with respect to the pump parameters, which makes the $MSE$ the only candidate for this. When the pdf curves in Fig. \[fig:NNdir\](a) and (b) present similar shapes, like for the C+L–band RAs, it might be an indication that minimums of these two errors occur for similar pump configurations and, consequently, minimizing $MSE^P$ (which is proportional to the $RMSE^P$), may also minimizes $E^{P}_{MAX}$. For the S-C-L–band RA, on the other hand, where $E^{P}_{MAX}$ and $RMSE^{P}$ pdf curves have completely different shapes, it is more likely that minimizing $MSE^P$ is not the same as minimizing $E^{P}_{MAX}$.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported by the European Union’s H2020 program (Marie Skłodowska-Curie grant 754462 and MSCA-ITN WON grant 814276), the European Research Council (ERC CoG FRECOM grant 771878), the Villum Foundations (VYI OPTIC-AI grant no. 29344), and the UK EPSRC grants EP/M009092/1 and EP/R035342/1.
[^1]: U. C. de Moura, F. Da Ros and D. Zibar are with DTU Fotonik, Department of Photonics Engineering, Technical University of Denmark, DK-2800, Kgs. Lyngby, Denmark (e-mail: [email protected]).
[^2]: Md A. Iqbal, M. Kamalian, L. Krzczanowicz, W. Forysiak and S. Turitsyn are with Aston Institute of Photonic Technologies (AIPT), Aston University, Birmingham, B4 7ET, United Kingdom.
[^3]: A. M. Rosa Brusin and A. Carena are with Dipartimento di Elettronica e Telecomunicazioni (DET), Politecnico di Torino, Corso Duca degli Abruzzi, 24 - 10129, Torino, Italy.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This chapter will cover the search for dark matter axions based on microwave cavity experiments proposed by Pierre Sikivie. The topic begins with a brief overview of halo dark matter and the axion as a candidate. The principle of resonant conversion of axions in an external magnetic field will be described as well as practical considerations in optimizing the experiment as a signal-to-noise problem. A major focus of the lecture will be the two complementary strategies for ultra-low noise detection of the microwave photons - the “photon-as-wave” approach (i.e. conventional heterojunction amplifiers and soon to be quantum-limited SQUID devices), and “photon-as-particle” (i.e. Rydberg-atom single-quantum detection). Experimental results will be presented; these experiments have already reached well into the range of sensitivity to exclude plausible axion models, for limited ranges of mass. The section will conclude with a discussion of future plans and challenges for the microwave cavity experiment.'
author:
- Gianpaolo Carosi
- Karl van Bibber
bibliography:
- 'carosi.bib'
title: Cavity Microwave Searches for Cosmological Axions
---
Dark matter and the axion {#sec:1:new}
=========================
Recent precision measurements of various cosmological parameters have revealed a universe in which only a small fraction can be observed directly. Measurements of deuterium abundances predicted from the theory of big bang nucleosynthesis (BBN) have limited the familiar baryonic matter to a mere 4% of the universe’s total energy density [@BBN]. Evidence from the cosmic microwave background, combined with supernovae searches, galaxy surveys and other measurements leads to the fascinating conclusion that the vast majority of the universe is made of gravitating “dark matter” (26%) and a negative pressure “dark energy” (70%) [@Tegmark].
Though the evidence for “dark energy” is relatively recent (primarily resting on cosmological supernovae surveys taken over the last decade or so) the existence of “dark matter” has been known since the early 1930s. It was then that Fritz Zwicky, surveying the Coma cluster, noticed that member galaxies were moving far too quickly to be gravitationally bound by the luminous matter [@Zwicky]. Either they were unbound, which meant the cluster should have ripped apart billions of years ago or there was a large amount of unseen “dark matter” keeping the system together. Since those first observations evidence for dark matter has accumulated on scales as small as dwarf galaxies (kiloparsecs)[@JKG_DM_full] to the size of the observable universe (gigaparsecs)[@WMAP_1].
Currently the best dark matter candidates appear to be undiscovered non-baryonic particles left over from the big bang[^1]. By definition they would have only the feeblest interactions with standard model particles such as baryons, leptons and photons. Studies of structure formation in the universe suggest that the majority of this dark matter is “cold”, i.e. non-relativistic at the beginning of galactic formation. Since it is collisionless, relativistic dark matter would tend to stream out of initial density perturbations effectively smoothing out the universe before galaxies had a chance to form [@Structure]. The galaxies that we observe today tend to be embedded in large halos of dark matter which extend much further than their luminous boundaries. Measurements of the Milky Way’s rotation curves (along with other observables such as microlensing surveys) constrain the density of dark matter near the solar system to be roughly $\rho_{CDM} \approx 0.45$ GeV/cm$^3$ [@Local_Density].
The two most popular dark matter candidates are the general class of Weakly Interacting Massive Particles (WIMPs), one example being the supersymmetric neutralino, and the axion, predicted as a solution to the “Strong CP” problem. Though both particles are well motivated this discussion will focus exclusively on the axion. As described in earlier chapters the axion is a light chargeless pseudo-scalar boson (negative parity, spin-zero particle) predicted from the breaking of the Peccei-Quinn symmetry. This symmetry was originally introduced in the late 1970s to explain why charge (C) and parity (P) appear to be conserved in strong interactions, even though the QCD Lagrangian has an explicitly CP violating term. Experimentally this CP violating term should have lead to an easily detectable electric dipole moment in the neutron but none has been observed to very high precision [@Dipole].
The key parameter defining most of the axion’s characteristics is the spontaneous symmetry breaking (SSB) scale of the Peccei-Quinn symmetry, $f_a$. Both the axion couplings and mass are inversely proportional to $f_a$ with the mass defined as
$$m_a \simeq 6\;\mu eV\left(\frac{10^{12}\;GeV}{f_a}\right).
\label{equ:Mass}$$
and the coupling of axions to photons ($g_{a \gamma \gamma}$) expressed as
$$g_{a\gamma \gamma} \equiv \frac{\alpha g_{\gamma}}{\pi f_a}
\label{equ:Coupling}$$
where $\alpha$ is the fine structure constant and $g_{\gamma}$ is a dimensionless model dependent coupling parameter. Generally $g_{\gamma}$ is thought to be $\sim 0.97$ for the class of axions denoted KSVZ (for Kim-Shifman-Vainshtein-Zakharov) [@KSVZ_1; @KSVZ_2] and $\sim -0.36$ for the more pessimistic grand-unification-theory inspired DFSZ (for Dine-Fischler-Srednicki-Zhitnitshii) models [@DFSZ_1; @DFSZ_2]. Since interactions are proportional to the square of the couplings these values of $g_{\gamma}$ tend to constrain the possible axion-to-photon conversion rates to only about an order of magnitude at any particular mass.
Initially $f_a$ was believed to be around the electroweak scale ($f_a \sim 250\;GeV$) resulting in an axion mass of order $100\;keV$ [@Weinberg; @Wilczek] and couplings strong enough to be seen in accelerators. Searches for axions in particle and nuclear experiments, along with limits from astrophysics, soon lowered its possible mass to $m_a \leq 3 \times 10^{-3}\;eV$[@Cavity_review] corresponding to $f_a \geq 10^9\;GeV$. Since their couplings are inversely proportional to $f_a$ these low mass axions were initially thought to be undetectable and were termed “invisible” axions.
From cosmology it was found that a general lower limit could be placed on the axion mass as well. At the time of the big bang axions would be produced in copious amounts via various mechanisms described in previous chapters. The total contributions to the energy density of the universe from axions created via the vacuum misalignment method can then be expressed as
$$\Omega_a \sim \left(\frac{5\;\mu eV}{m_a}\right)^{7/6}
\label{Overclosure}$$
which puts a lower limit on the axion mass of $m_a \geq 10^{-6}\;eV$ (any lighter and the axions would overclose the universe, $\Omega_a \geq 1$). Combined with the astrophysical and experimental limits this results in a 3 decade mass range for the axion, from $\mu eV\;-\;meV$, with the lower masses more likely if the axion is the major component of dark matter. The axions generated in the early universe around the QCD phase transition, when the axion mass turns on, would have momenta $\sim 10^{-8}\;eV/c$ while the surrounding plasma had a temperature $T \simeq 1\;GeV$ [@Cavity_review]. Furthermore, such axions are so weakly coupling that they would never be in thermal equilibrium with anything else. This means they would constitute non-relativistic “cold” dark matter from the moment they appeared and could start to form structures around density perturbations relatively quickly.
Today the axion dark matter in the galaxy would consist of a large halo of particles moving with relative velocities of order $10^{-3}c$. It is unclear whether any or all of the axions would be gravitationally thermalized but, in order for them to be bound in the galaxy, they would have to be moving less than the local escape velocity of $2\times 10^{-3}c$. It’s possible that non-thermalized axions could still be oscillating into and out of the galaxy’s gravitational well. These axions would have extraordinarily tiny velocity dispersions (of order $10^{-17}c$) [@Caustics] and the differences in velocity from various infalls (first time falling into the galaxy, first time flying out, second time falling in, etc.) would be correlated with the galaxy’s development.
Principles of microwave cavity experiments {#sec:2:new}
==========================================
Pierre Sikivie was the first to suggest that the “invisible” axion could actually be detected [@Cavity_idea]. This possibility rests on the coupling of axions to photons given by
$$L_{a \gamma \gamma} = -\left(\frac{\alpha}{\pi} \frac{g_{\gamma}}{f_a}\right)a \vec{E} \cdot \vec{B}
\label{Lagrangian}$$
where $\vec{E}$ and $\vec{B}$ are the standard electric and magnetic field of the coupling photons, $\alpha$ is the fine structure constant and $g_{\gamma}$ is the model dependent coefficient mentioned in the previous section [@Cavity_review]. Translating this to a practical experiment Sikivie suggested that axions passing through an electromagnetic cavity permeated with a magnetic field could resonantly convert into photons when the cavity resonant frequency ($\omega$) matched the axion mass ($m_a$). Since the entire mass of the axion would be converted into a photon a 5 $\mu eV$ axion at rest would convert to a 1.2 GHz photon which could be detected with sensitive microwave receivers. The predicted halo axion velocities of order $\beta = 10^{-3}$ would predict a spread in the axion energy, from $E_a = m_a c^2 + \frac{1}{2}m_a c^2\beta^2$, of order $10^{-6}$. For our example 5 $\mu eV$ axions this would translate into a 1.2 kHz upward spread in the frequency of converted photons. The power of axions converting to photons on resonance in a microwave cavity is given by
$$\begin{aligned}
P_a & = & \left(\frac{\alpha}{\pi} \frac{g_{\gamma}}{f_a}\right)^2 V B^2_0 \rho_a C_{lmn} \frac{1}{m_a} Min\left(Q_L,Q_a\right) \label{equ:Power}\\
& = & 0.5\times 10^{-26}W \left(\frac{V}{500\;liters}\right)\left(\frac{B_0}{7\;T}\right)^2 C \left(\frac{g_{\gamma}}{0.36}\right)^2 \nonumber\\
& & \times \left(\frac{\rho_a}{0.5\times 10^{-24} g/cm^3}\right) \nonumber\\
& & \times \left(\frac{m_a}{2\pi\;(GHz)}\right)Min\left(Q_L,Q_a\right) \nonumber\end{aligned}$$
where $V$ is the cavity volume, $B_0$ is the magnetic field, $Q_L$ is the cavity’s loaded quality factor (defined as center frequency over frequency bandwidth), $Q_a=10^6$ is the quality factor of the axion signal (axion energy over spread in energy or $1/\beta^2$), $\rho_a$ is the axion mass density at the detection point (earth) and $C_{lmn}$ is the form factor for one of the transverse magnetic ($TM_{lmn}$) cavity modes (see section \[subsec:3b:new\] for more on cavity modes). This form factor is essentially the normalized overlap integral of the external static magnetic field, $\vec{B_0}(\vec{x})$, and the oscillating electric field, $\vec{E}_{\omega}(\vec{x})e^{i\omega t}$, of that particular cavity mode. It can be determined using
$$C = \frac{|\int_V d^3x \vec{E}_{\omega}\cdot \vec{B}_0|^2}{\vec{B}^2_0 V \int_V d^3x \epsilon |\vec{E}_{\omega}|^2}
\label{equ:Form_factor}$$
where $\epsilon$ is the dielectric constant in the cavity.
For a cylindrical cavity with a homogeneous longitudinal magnetic field the $TM_{010}$ mode provides the largest form factor ($C_{010}=0.69$) [@Cavity_review]. Though model dependent equation \[equ:Power\] can give an idea of the incredibly small signal, measured in yoctowatts ($10^{-24}$ W), expected from axion-photon conversions in a resonant cavity. This is much smaller than the $2.5\times 10^{-21}\;W$ of power received from the last signal of the $Pioneer\;10$ spacecraft’s 7.5-W transmitter in 2002, when it was 12.1 billion kilometers from earth [@Pioneer].
Currently the axion mass is constrained between a $\mu eV$ and a $meV$ corresponding to a frequency range for converted photons between 240 MHz and 240 GHz. To maintain the resonant quality of the cavity, however, only a few kHz of bandwidth can be observed at any one time. As a result the cavity needs to be tunable over a large range of frequencies in order to cover all possible values of the axion mass. This is accomplished using metallic or dielectric tuning rods running the length of the cavity cylinder. Moving the tuning rods from the edge to the center of the cavity shifts the resonant frequency by up to 100 MHz.
Even when the cavity is exactly tuned to the axion mass detection is only possible if the microwave receiver is sensitive enough to distinguish the axion conversion signal over the background noise from the cavity and the electronics. The signal to noise ratio (SNR) can be calculated from the Dicke radiometer equation [@Dicke_radiometer]
$$SNR = \frac{P_a}{\bar{P}_N} \sqrt{Bt} = \frac{P_a}{k_B T_S}\sqrt{\frac{t}{B}}
\label{Radiometer}$$
where $P_a$ is the axion conversion power, $\bar{P}_N=k_B B T_S$ is the average thermal noise power, $B$ is the bandwidth, $T_S$ is the total system noise temperature (cavity plus electronics) and $t$ is the signal integration time [@Cavity_review]. With the bandwidth of the experiment essentially set by the axion mass and anticipated velocity dispersion ($\beta^2 \sim 10^{-6}$) the SNR can be raised by either increasing the signal power ($P_a \propto B^2_0 V$), lowering the noise temperature or integrating for a longer period of time. Increasing the size of the magnetic field or the volume of the cavity to boost the signal power can get prohibitively expensive fairly quickly. Given the large range of possible masses the integration time needs to remain relatively short (of order 100 seconds integration for every kHz) in order to scan an appreciable amount in time scales of a year or so. If one chooses a specific SNR that would be acceptable for detection then a scanning rate can be defined as
$$\begin{aligned}
\frac{df}{dt} & = & \frac{12\;GHz}{yr} \left(\frac{4}{SNR}\right)^2\left(\frac{V}{500\;liter}\right)\left(\frac{B_0}{7\;T}\right)^4 \label{equ:Rate}\\
& & \times\;C^2\left(\frac{g_{\gamma}}{0.36}\right)^4\left(\frac{\rho_a}{5\times10^{-25}}\right)^2 \nonumber\\
& & \times \left(\frac{3K}{T_S}\right)^2\left(\frac{f}{GHz}\right)^2\frac{Q_L}{Q_a}. \nonumber\end{aligned}$$
Given that all other parameters are more or less fixed, due to physics and budgetary constraints, the sensitivity of the experiment (both in coupling reach and in scanning speed) can only practically be improved by developing ultra low noise microwave receivers. In fact some of the quietest microwave receivers in the world have been developed to detect axions [@Med_res].
Technical implementation {#sec:3:new}
========================
The first generation of microwave experiments were carried out at Brookhaven National Laboratory (BNL) [@Brookhaven] and at the University of Florida [@UofF] in the mid-1980s. These were proof-of-concept experiments and got within factors of 100 - 1000 of the sensitivity required to detect plausible dark matter axions (mostly due to their small cavity size and relatively high noise temperatures) [@Cavity_review]. In the early 1990s second generation cavity experiments were developed at Lawrence Livermore National Laboratory (LLNL) in the U.S. and in Kyoto, Japan. Though both used a microwave cavity to convert the axions to photons they each employed radically different detection techniques. The U.S. experiment focused on improving coherent microwave amplifiers (photons as waves) while the Japan experiment worked to develop a Rydberg-atom single-quantum detector (photons as particles). Since the Kyoto experiment is still in the development phase we will save its description for a later section and focus on the U.S. experiment.
![Schematic diagram of ADMX experiment including both the resonant cavity (which sits in the bore of a superconducting solenoid) and receiver electronics chain.[]{data-label="fig:ADMX_schem"}](carosi-01.eps){height="8.5cm"}
A schematic of the LLNL experiment, dubbed the **A**xion **D**ark **M**atter e**X**periment (ADMX), can be seen in figure \[fig:ADMX\_schem\]. The experiment consists of a cylindrical copper-plated steel cavity containing two axial tuning rods. These can be moved transversely from the edge of the cavity wall to its center allowing one to perturb the resonant frequency. The cavity itself is located in the bore of a superconducting solenoid providing a strong constant axial magnetic field. The electromagnetic field of the cavity is coupled to low-noise receiver electronics via a small adjustable antenna[@Cavity_review]. These electronics initially amplify the signal using two ultra-low noise cryogenic amplifiers arranged in series. The signal is then boosted again via a room temperature post-amplifier and injected into a double-heterodyne receiver. The receiver consists of an image reject mixer to reduce the signal frequency from the cavity resonance (hundreds of MHz - GHz) to an intermediate frequency (IF) of 10.7 MHz. A crystal bandpass filter is then employed to reject noise power outside of a 35 kHz window centered at the IF. Finally the signal is mixed down to almost audio frequencies (35 kHz) and analyzed by fast-Fourier-transform (FFT) electronics which compute a 50 kHz bandwidth centered at 35 kHz. Data is taken every 1 kHz or so by moving the tuning rods to obtain a new resonant $TM_{010}$ mode. In the next few sections we will expand on some of these components.
![Overview of ADMX hardware including superconducting magnet and cavity insert.[]{data-label="fig:ADMX_exp"}](carosi-02.eps){height="8.5cm"}
Magnet {#subsec:3a:new}
------
The main magnet for ADMX was designed to maximize the $B^2_0 V$ contribution to the signal power (equation \[equ:Power\]). It was determined that a superconducting solenoid would yield the most cost effective solution and its extremely large inductance (535 Henry) would have the added benefit of keeping the field very stable. The 6 ton magnet coil is housed in a 3.6 meter tall cryostat (see figure \[fig:ADMX\_exp\]) with an open magnet bore allowing the experimental insert, with the cavity and its liquid helium (LHe) reservoir, to be lowered in. The magnet itself is immersed during operations in a 4.2 K LHe bath in order to keep the niobium-titanium windings superconducting. Generally the magnet was kept at a field strength of 7.6 T in the solenoid center (falling to approximately 70 % strength at the ends) but recently its been run as high as 8.2 T [@Cavity_review].
Microwave cavities {#subsec:3b:new}
------------------
The ADMX experiment uses cylindrical cavities in order to maximize the axion conversion volume in the solenoid bore. They are made of a copper-plated steel cylinder with capped ends. The electromagnetic field structure inside a cavity can be found by solving the Helmholtz equation
$$\nabla^2\Phi + k^2\Phi = 0
\label{equ:Helmholtz}$$
where the wavenumber $k$ is given by
$$k^2 = \mu \epsilon \omega^2 - \beta^2
\label{equ:wavenumber}$$
and $\beta$ is the eigenvalue for the transverse (x,y) component [@Kinion_thesis]. The cavity modes are the standing wave solutions to equation \[equ:Helmholtz\]. The boundary conditions of an empty cavity only allow transverse magnetic (TM) modes ($\vec{B}_z = 0$) and transverse electric (TE) modes ($\vec{E}_z = 0$). Since the TE modes have no axial electric field one can see from equation \[Lagrangian\] that they don’t couple at all to axions and we’ll ignore them for the moment. The $TM_{lmn}$ modes are three dimensional standing waves where $l=0,1,2...$ is the number of azimuthal nodes, $m=1,2,3...$ is the number of radial nodes and $n=0,1,2...$ is the number of axial nodes. The axions couple most strongly to the lowest order $TM_{010}$ mode.
The resonant frequency of the $TM_{010}$ mode can be shifted by the introduction of metallic or dielectric tuning rods inserted axially into the cavity. Metallic rods raise the cavity resonant frequency the closer they get to the center while dielectric rods lower it. In ADMX these rods are attached to the ends of alumina arms which pivot about axles set in the upper and lower end plates. The axles are rotated via stepper motors mounted at the top of the experiment (see figure \[fig:ADMX\_exp\]) which swing the tuning rods from the cavity edge to the center in a circular arc. The stepper motors are attached to a gear reduction which translates a single step into a 0.15 arcsecond rotation, corresponding to a shift of $\sim$ 1 kHz at 800 MHz resonant frequency [@Cavity_review].
![Resonant cavity with top flange removed. An alumina tuning rod can be seen at the bottom right and a copper tuning rod is in the upper left.[]{data-label="fig:Cavity"}](carosi-03.eps){height="6.cm"}
With the addition of metallic tuning rods TEM modes ($\vec{B}_z = \vec{E}_z = 0$) can also be supported in the cavity. Like the TE modes they do not couple to the axions but they can couple weakly to the vertically mounted receiver antenna (due to imperfections in geometry, etc).
![Mode structure of a cavity with two copper tuning rods. The left figure displays the frequencies of the resonant modes, measured via a swept rf signal, when one tuning rod is kept at the cavity edge while the other is moved toward the center. The right figure is a sketch of a mode crossing.[]{data-label="fig:Cavity_Modes"}](carosi-04.eps){height="5.cm"}
Figure \[fig:Cavity\_Modes\] demonstrates how the various resonant modes shift as a copper tuning rod is moved from near the cavity wall toward the center. The TEM and TE modes are largely unaffected by the change in tuning rod position while TM modes rise in frequency as one of the copper rods moves toward the cavity center. This leads to regions in which a TM mode crosses a TE or TEM mode (referred to as mode mixing). These mode mixings (illustrated by the right part of figure \[fig:Cavity\_Modes\]) introduce frequency gaps which can not be scanned. As a result the cavity was later filled with LHe, which changed the microwave index of refraction to 1.027, thus lowering the mode crossings by 2.7% and allowed the previously unaccessible frequencies to be scanned.
A key feature of the resonant microwave cavity is its quality factor $Q$, which is a measure of the sharpness of the cavity response to external excitations. It is a dimensionless value which can be defined a number of ways including the ratio of the stored energy ($U$) to the power loss ($P_L$) per cycle: $Q = \omega_0 U/P_L$. The quality factor (Q) of the $TM_{010}$ mode is determined by sweeping a radio (rf) signal through the weakly coupled antenna in the cavity top plate (see figure \[fig:ADMX\_schem\]). Generally, the unloaded Q of the cavity is $\sim 2\times 10^5$ [@Cavity_review] which is very near to the theoretical maximum for oxygen-free annealed copper at cryogenic temperatures. During data taking the insertion depth of the major antenna is adjusted to make sure that it matches the 50 $\Omega$ impedance of the cavity (called critically coupling). When the antenna is critically coupled half the microwave power in the cavity enters the electronics via the antenna while half is dissipated in the cavity walls. Overcoupling the cavity would lower the Q and thus limit the signal enhancement while undercoupling the cavity would limit the microwave power entering the electronics.
Amplifier and receiver {#subsec:3c:new}
----------------------
After the axion signal has been generated in the cavity and coupled to the major port antenna it is sent to the cryogenic amplifiers. The design of the first amplifier is especially important because its noise temperature (along with the cavity’s Johnson noise) dominates the rest of the system. This can be illustrated by following a signal from the cavity as it travels through two amplifiers in series. The power contribution from the thermal noise of the cavity at temperature $T_c$ over bandwidth $B$ is given by $P_{nc} = B k_B T_c$ (where $k_B$ is Boltzmann’s constant). When this noise passes through the first amplifier, which provides gain $G_1$, the output includes the boosted cavity noise as well as extra power ($P_{N,A_1}$) from the amplifier itself. The noise from the amplifier appears as an increase in the temperature of the input source. $$P_1 = G_1 B k_B T_c + P_{N,A_1} = G_1 B k_B (T_c + T_{A_1})
\label{equ:first_stage}$$ If this boosted noise power (cavity plus first amplifier) is then sent through a second amplifier, with gain $G_2$ and noise temperature $T_{A_2}$, the power output becomes $$P_2 = G_2 P_1 + P_{N,A_2} = G_2 (G_1 B k_B (T_c + T_{A_1})) + G_2 B k_B T_{A_2}
\label{equ:noise_amps}$$ The combined noise temperature from the two amplifiers ($T_A$) can be found by matching equation \[equ:noise\_amps\] to that of a single amplifier, $P_2 = G_2 G_1 B k_B (T_c + T_A)$, which gives $$T_A = T_{A_1} + \frac{T_{A_2}}{G_1}
\label{equ:amp_temp}$$ Thus one can see that, because of the gain $G_1$ of the first stage amplifier, its noise temperature dominates all other amplifiers in the series.
The current first stage amplifiers used in ADMX are cryogenic heterostructure field-effect transistors (HFETs) developed at the National Radio Astronomy Observatory (NRAO) specifically for the ADMX experiment [@Cavity_review; @NRAO]. In these amplifiers electrons from an aluminum doped gallium arsenide layer fall into the GaAs two-dimensional quantum well (the FET channel). The FET electrons travel ballistically, with little scattering, thus minimizing electronic noise [@Physics_Today]. Currently electronic noise temperatures of under 2 K have been achieved using the HFETs. In the initial ADMX data runs, now concluded, two HFET amplifiers were used in series, each with approximately 17 dB power gain, leading to a total first stage power gain of 34 dB. Each amplifier utilized 90 degree hybrids in a balanced configuration in order to minimize input reflections, thus providing a broadband match to the 50 $\Omega$ cavity impedance (see figure \[fig:Balanced\]).
![Schematic diagram of a balanced amplifier. Every time the signal crosses through the middle of a hybrid its phase is shifted by 90 degrees. Reflections back to the input destructively interfere while reflections to the upper left constructively interfere and are dumped into a 50 $\Omega$ terminator. Signals to the output are both shifted by 90 degrees and thus add constructively.[]{data-label="fig:Balanced"}](carosi-05.eps){height="6.cm"}
Though the amplifiers worked well in the high magnetic field just above the cavity it was determined during commissioning that they should be oriented such that the magnetic field was parallel to the HFET channel electron flow. This minimized the electron travel path and thus the noise temperature [@Cavity_review].
The signal from the cryogenic amplifiers is carried by coaxial cable to a low-noise room temperature post-amplifier, which added an additional 38 dB gain between 300 MHz - 1 GHz. Though the post-amplifiers noise temperature is 90 K its contribution relative to the cryogenic amplifiers (with 38 dB initial gain) is only 0.03 K (see equation \[equ:amp\_temp\]). Including various losses the total gain from the cavity to the post-amplifier output is 69 dB [@Cavity_review].
After initial stages of amplification the signal enters the double-heterodyne receiver (essentially an AM radio). Figure \[fig:Receiver\] is a schematic of the receiver electronics. The first element is an image reject mixer which uses a local oscillator to mix the signal down to 10.7 MHz. This intermediate frequency (IF) is then sent through a programmable attenuator (used during room temperature testing so that the receiver electronics are not saturated). An IF amplifier then boosts the signal by another 20 dB before passing it by a weakly coupled signal sampler. The signal then passes through a crystal bandpass filter which suppresses noise outside a 30 kHz bandwidth center at 10.7 MHz. The signal is then boosted by an additional 20 dB before being mixed down to 35 kHz. The total amplification of the signals from the cavity is $\sim$ 106 dB [@General_art_1].
![Receiver chain that mixes the signal down from the cavity $TM_{010}$ resonant frequency to 35 kHz.[]{data-label="fig:Receiver"}](carosi-06.eps){height="4.cm"}
Once the signal has been mixed down to the 35 kHz center frequency it is passed off to a commercial FFT spectrum analyzer and the power spectrum recorded. The entire receiver, including filter, is calibrated using a white-noise source at the input. During data collection the FFT spectrum analyzer takes 8 msec single-sided spectra (the negative and positive frequency components are folded on top of each other). Each spectrum consists of 400 bins with 125 Hz width spanning a frequency range of 10-60 kHz. After 80 seconds of data taking (with a fixed cavity mode) the 10,000 spectra are averaged together and saved as raw data. This is known as the medium resolution data.
In addition there is a high resolution channel to search for extremely narrow conversion lines from late infall non-thermal axions (as mentioned at the end of section \[sec:1:new\]). For this channel the 35 kHz signal is passed through a passive LC filter with a 6.5 kHz passband, amplified, and then mixed down to a 5 kHz center frequency. A single spectrum is then obtained by acquiring $2^{20}$ points in about 53 sec and a FFT is performed. This results in about $3.4\times 10^5$ points in the 6.5 kHz passband with a frequency resolution of 19 mHz.
Data analysis {#sec:4:new}
=============
The ADMX data analysis is split into medium and high resolution channels. The medium resolution channel is analyzed using two hypotheses. The first is a “single-bin” search motivated by the possibility that some of the axions have not thermalized and therefore would have negligible velocity dispersion, thus depositing all their power into a single power-spectrum bin. The second hypothesis utilizes a “six-bin” search which assumes that axions have a velocity dispersion of order $10^{-3}c$ or less (axions with velocities greater than $2\times 10^{-3}c$ would escape the halo). The six-bin search is the most conservative and is valid regardless of whether the halo axions have thermalized or not.
Since each 80 second long medium resolution spectra is only shifted by 1 kHz from the previous integration each frequency will show up in multiple spectra (given the 50 kHz window). As a result each 125 Hz bin is weighted according to where it falls in the cavity response function and co-added to give an effective integration time of $\sim$ 25 minutes per frequency bin. For the single-bin search individual 125 Hz bins are selected if they exceed an initial power-level threshold. This is set relatively low so a large number of bins are usually selected. These bins are then rescanned to achieve a similar signal-to-noise ratio and combined with the first set of data generating a spectra with higher signal to noise. The selection process is then repeated a number of times until persistent candidates are identified. These few survivors are then carefully checked to see if there are any external sources of interference that could mimic an axion signal. If all candidates turn out to be exterior radio interference the excluded axion couplings (assuming a specific dark matter density) can be computed from the near-Gaussian statistics of the single-bin data. For the six-bin search, all six adjacent frequency bins that exceed a set power-threshold are selected from the power spectra. The large number of candidates are then whittled down using the same iterations as the single-bin analysis. If no candidates survive the excluded axion couplings are computed by Monte Carlo [@Cavity_review].
From the radiometer equation (\[Radiometer\]) one can see that the search sensitivity can be increased if strong narrow spectral lines exist. The integration times for each tuning rod setting is around 60 seconds and the resulting Doppler shift from the Earth’s rotation leads to a spread of $\sim$ mHz in a narrow axion signal. Since the actual velocity dispersions of each discrete flow is unknown multiple resolution searches were performed by combining 19 mHz wide bins. These were referred to as $n$-bin searches, where $n = $ 1, 2, 4, 8, 64, 512 and 4096. Candidate peaks were kept if they were higher then a specified threshold set for that particular $n$-bin search. These thresholds were 20, 25, 30, 40, 120, 650 and 4500 $\sigma$, for increasing order of $n$. The initial search using the high resolution analysis took data between 478-525 MHz, corresponding to axion masses between 1.98 and 2.17 $\mu eV$. This search was made in three steps. First the entire frequency range was scanned in 1 kHz increments with the candidate axion peaks recorded. Next multiple time traces were taken of candidate peaks [@High_res]. Finally persistent peaks were checked by attenuating or disconnecting various diagnostic coaxial cables leading into the cavity (see figure \[fig:ADMX\_schem\]). If the signals were external interference they would decrease in power dramatically while an axion signal would remain unchanged [@Cavity_review]. Further checks could be done by disconnecting the cavity from the receiver input and replacing it with an antenna to see if the signal persisted.
If a persistent candidate peak is found which does not have an apparent source from external interference a simple check would be to turn off the magnetic field. If the signal disappears it would be a strong indication that it was due to axions and not some unknown interference. So far, though, all candidates have been identified with an external source.
Results {#sec:5:new}
=======
So far no axions have been detected in any experiment. ADMX currently provides the best limits from microwave cavity experiments in the lowest mass range (most plausible if axions are the major component to the dark matter). Both the medium resolution data and the high resolution data yield exclusion plots in either the coupling strength of the axion (assuming a halo density of $\rho_a = 0.45\;GeV/cm^3$) or in the axion halo density (assuming a specific DFSZ or KSVZ coupling strength). Results from the medium resolution channel [@Med_res] can be seen in figure \[fig:Med\_res\_limits\] and the high resolution results [@High_res] can be seen in \[fig:High\_res\_limits\]. Both of the results are listed at 90 % confidence level.
![Results from the medium resolution channel [@Med_res]. The figure to the left is the exclusion plot for power in a thermalized spectrum assuming a halo density of $\rho_a = 0.45\;GeV/cm^3$. The figure to the right is the fractional dark matter halo density excluded as axions for two different axion models.[]{data-label="fig:Med_res_limits"}](carosi-07.eps){height="5.5cm"}
![High resolution limits given different axion couplings [@High_res]. This shows that the current high resolution channel is sensitive to fractional halo densities ($\approx$ 30%) if the axions couple via the KSVZ model. If they couple via the DFSZ model the experiment is not yet sensitive to the maximum likelihood halo density ($\rho_a \sim 0.45\;GeV/cm^3$), but would be sensitive to a single line with twice that density.[]{data-label="fig:High_res_limits"}](carosi-08.eps){height="5.cm"}
Future developments {#sec:6:new}
===================
In order to carry out a definitive search for axion dark matter various improvements to the detector technology need to be carried out. Not only do the experiments need to become sensitive enough to detect even the most pessimistic axion couplings (DFSZ) at fractional halo densities but they must be able to scan relatively quickly over a few decades in mass up to possibly hundreds of GHz. The sensitivity of the detectors (which is also related to scanning speed) is currently limited by the noise in the cryogenic HFET amplifiers. Even though they have a noise temperature under 2 K the quantum limit (defined as $T_Q \sim h\nu/k$) is almost two orders of magnitude lower (25 mK at 500 MHz). To get down to, or even past, this quantum limit two very different technologies are being developed. The first is the implementation of SQUIDs (Superconducting Quantum Interference Devices) as first stage cryogenic amplifiers. The second uses Rydberg-atoms to detect single microwave photons from axion conversions in the cavity.
Though both techniques will lead to vastly more sensitive experiments they will still be limited in their mass range. Currently all cavity experiments have been limited to the $2 - 20 \mu eV$ range, mostly due to the size of resonant cavities. For a definitive search the mass range must be increased by a factor of 50 which requires new cavity designs that increase the resonant frequency while maintaining large enough detection volumes. Detectors that work at these higher frequencies also need to be developed.
SQUID amplifiers {#sec:6a:new}
----------------
The next generation of the ADMX experiment will use SQUID amplifiers to replace the first stage HFETs. SQUIDs essentially use a superconducting loop with two parallel Josephson junctions to enclose a total amount of magnetic flux $\Phi$. This includes both a fixed flux supplied by the bias coil and the signal flux supplied by an input coil. The phase difference between the currents on the two sides of the loop are affected by changing $\Phi$ resulting in an interference effect similar to the two-slit experiment in optics [@Kinion_thesis]. Essentially the SQUID will act as flux to voltage transducers as illustrated in figure \[fig:SQUID\].
![Essentials of a SQUID microwave detector. The left figure is a schematic of the SQUID device coupling to the input signal which is converted into magnetic flux. The right figure shows how biasing the flux allows for amplification.[]{data-label="fig:SQUID"}](carosi-09.eps){height="3.5cm"}
Most SQUIDs are built using the Ketchen and Jaycox design [@Ketchen], in which the SQUID loop is an open square washer made of niobium (Nb). The loop is closed by a separate Nb electrode connected to the washer opening on either side by a Josephson junction and external shunt resistors. A spiral input coil is placed on top of the washer, separated by a layer of insulation. The original designs in which input signals were coupled into both ends of the coil tended to only work below about 200 MHz due to parasitic capacitance between the coil and the washer at higher frequencies. This was solved by coupling the input signal between one end of the coil and the SQUID washer, which would act as a ground plane to the coil and create a microstrip resonator (see figure \[fig:SQUID\_res\]). This design has been tested successfully up to 3 GHz [@Kinion_thesis].
![Diagram and picture of microstrip resonator SQUIDs to be used in ADMX upgrade.[]{data-label="fig:SQUID_res"}](carosi-10.eps){height="3.5cm"}
Unlike the HFETs, whose noise temperature bottoms out at just under 2 K regardless of how cold the amplifiers get, the SQUIDs noise temperature remains proportional to the physical temperature down to within 50% of the quantum limit. The source of this thermal noise comes from the shunt resistors across the SQUID’s Josephson junctions and future designs that minimize this could push the noise temperature even closer to the quantum limit [@Physics_Today].
Currently the ADMX experiment is in the middle of an upgrade in which SQUIDs will be installed as first stage cryogenic amplifiers. This should cut the combined noise temperature of the cavity + electronics in half allowing ADMX to become sensitive to half the KSVZ coupling (with the same scanning speed as before). Due to the SQUIDs’ sensitivity to magnetic fields this upgrade includes an entire redesign in which a second superconducting magnet is being installed in order to negate the main magnet’s field around the SQUID amplifiers. Data taking is expected to begin in the first half of 2007 and run for about a year. Future implementations of ADMX foresee using these SQUID detectors with a dilution refrigerator to set an operating temperature of $\sim$ 100 mK, allowing sensitivity to DFSZ axion couplings to be achieved with 5 times the scanning rate the current HFETs take to reach KSVZ couplings.
Rydberg-atom single-quantum detectors {#sec:6b:new}
-------------------------------------
One technique to evade the quantum noise limit is to use Rydberg atoms to detect single photons from the cavity. A Rydberg atom has a single valence electron promoted to a level with a large principal quantum number $n$. These atoms have energy spectra similar in many respects to hydrogen, and dipole transitions can be chosen anywhere in the microwave spectrum by an appropriate choice of $n$. The transition energy itself can be finely tuned by using the Stark effect to exactly match a desired frequency. That, combined with the Rydberg atom’s long lifetime and large dipole transition probability, make it an excellent microwave photon detector.
An experimental setup utilizing this technique called CARRACK has been assembled in Kyoto, Japan [@Cavity_review; @Kyoto_1] and a schematic can be seen pictured in figure \[fig:Rydberg-atom\]. The axion conversion cavity is coupled to a second “detection” cavity tuned to the same resonant frequency $\nu$. A laser excites an atomic beam (in this case rubidium) into a Rydberg state ($|0> \rightarrow |n>$) which then traverses the detection cavity. The spacing between the energy levels is adjusted to $h\nu$ using the Stark effect and microwave photons from the cavity can be efficiently absorbed by the atoms (one photon per atom, $|ns> \rightarrow |np>$). The atomic beam then exits the cavity and is subjected to selective field ionization in which electrons from atoms in the higher energy state ($|np>$) get just enough energy to be stripped off and detected [@Physics_Today].
![Schematic of single photon microwave detection utilizing Rydberg atoms.[]{data-label="fig:Rydberg-atom"}](carosi-11.eps){height="6.cm"}
Currently the Kyoto experiment has measured cavity emission at 2527 MHz down to a temperature of 67 mK, a factor of two below the quantum limit at that frequency, and is working to reach the eventual design goal of 10 mk [@Kyoto_1]. This would be the point in which the cavity blackbody radiation would become the dominant noise background. One deficiency of the Rydberg atom technique is that it can’t detect structure narrower than the bandpass ($\Delta E/E$) of the cavity (generally $\sim 10^{-5}$). As a result it is insensitive to axion halo models that predict structure down to $\Delta E/E \sim 10^{-11}$, an area in which the ADMX high resolution channel, utilizing microwave amplifiers, can cover. Despite this Rydberg atom detectors could become very useful tools for halo axion detection in the near future.
Challenge of higher frequencies {#sec:6c:new}
-------------------------------
Current microwave cavity technology has only been able to probe the lowest axion mass scale. In order to cover the entire range up to the exclusion limits set by SN 1987a of $m_a \leq meV$ new cavity and detection techniques must be investigated which can operate up to the 100 GHz range. The resonant cavity frequency essentially depends on the size the cavity and the resonant mode used. The $TM_{010}$ mode has by far the largest form factor ($C\sim 0.69$) of any mode and all other higher frequency modes have much smaller or identically zero form factors. The single 50 cm diameter cavity used in the initial ADMX experiments had a central resonant frequency ($TM_{010}$) of 460 MHz and radial translation of metallic or dielectric tuning rods could only raise or lower that frequency by about $\pm 50\%$ [@Cavity_review]. Smaller cavities could get higher frequencies but the rate of axion conversions would go down as the cavity volume decreased.
In order to use the full volume of the magnet with smaller cavities it was determined that multiple cavities could be stacked next to each other and power combined. As long as the de Broglie wavelength of the axions is larger than the total array individual cavities tuned to the same frequency can be summed in phase. Typical axion de Broglie wavelengths are $\lambda_{dB} \sim 10m - 100m$ which means they drive the $\sim 1m$ cavity volume coherently. Data taken using a four cavity array in ADMX reached KSVZ sensitivity [@Kinion_thesis] over a small mass range (see figure \[fig:Higher\_frequency\]). These initial tests had difficulties getting the piezoelectric motors working trouble free in the magnetic and cryogenic environment. Since those tests the technology has advanced to the point in which it may be feasible to create larger sets of smaller cavity areas.
To reach even higher frequencies ideas have been raised to use resonators with periodic arrays of metal posts. Figure \[fig:Higher\_frequency\] shows the electric field profile of one possible array using a 19 post hexagonal pattern. Mounting alternating posts from the cavity top and the bottom and translating them relative to each other allow the resonant frequency to be adjusted by 10% or so. The possibility of using such cavities, or other new cavity geometries, is an active area of research and progress needs to be made before the full axion mass range can be explored.
![Outline of possible cavity concepts to explore higher axion masses. The left figure includes both a picture of the 4 cavity array and its corresponding exclusion plot over the limited mass range it took data. The right figure includes field maps for multiple posts inserted in a cavity.[]{data-label="fig:Higher_frequency"}](carosi-12.eps){height="6.cm"}
Summary, conclusions {#sec:7:new}
====================
Experimentally the axion is a very attractive cold dark matter candidate. Its coupling to photons ($g_{\gamma}$) for several different models all fall within about an order of magnitude in strength and its mass scale is currently confined to a three decade window. This leaves the axion in a relatively small parameter space, the first two decades or so of which is within reach of current or near future technology.
The ADMX experiment has already begun to exclude dark matter axions with KSVZ couplings over the lowest masses and upgrades to SQUID amplifiers and a dilution refrigerator could make ADMX sensitive to DFSZ axion couplings over the first decade in mass within the next three years. Development of advanced Rydberg-atom detectors, along with higher frequency cavities geometries, could give rise to the possibility of a definitive axion search within a decade. By definitive we mean a search which would either detect axions at even the most pessimistic couplings (DFSZ) at fractional halo densities over the full mass range, or rule them out entirely.
It should be noted that if the axion is detected it would not only solve the Strong-CP problem and perhaps the nature of dark matter but could offer a new window into astrophysics, cosmology and quantum physics. Details of the axion spectrum, especially if fine structure is found, could provide new information of how the Milky Way was formed. The large size of the axions de Broglie wavelength ($\lambda_a \sim 10m - 100m$) could even allow for interesting quantum experiments to be performed at macroscopic scales. All of these tantalizing possibilities, within the reach of current and near future technologies, makes the axion an extremely exciting dark matter candidate to search for.
Acknowledgments
===============
This work was supported under the auspices of the U.S. Department of Energy under Contract W-7405-Eng-48 at Lawrence Livermore National Laboratory.
[^1]: Even without the limits from Big Bang Nucleosythesis searches for baryonic dark matter in cold gas clouds [@Gas] or MAssive Compact Halo Objects (MACHOs), like brown dwarfs [@Machos; @EROS], have not detected nearly enough to account for the majority of dark matter. Attempts to modify the laws of gravity at larger scales have also had difficulties matching observations [@MOND].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Continuing our program of spectroscopic observations of ICRF sources, we present redshifts for 120 quasars and radio galaxies. Data were obtained with five telescopes: the 3.58m ESO New Technology Telescope (NTT), the two 8.2m Gemini telescopes, the 2.5m Nordic Optical Telescope (NOT), and the 6.0m Big Azimuthal Telescope (BTA) of the Special Astrophysical Observatory in Russia.
The targets were selected from the International VLBI Service for Geodesy & Astrometry (IVS) candidate International Celestial Reference Catalogue which forms part of an observational VLBI program to strengthen the celestial reference frame. We obtained spectra of the potential optical counterparts of more than 150 compact flat-spectrum radio sources, and measured redshifts of 120 emission-line objects, together with 19 BL Lac objects. These identifications add significantly to the precise radio–optical frame tie to be undertaken by [*Gaia*]{}, due to be launched in 2013, and to the existing data available for analysing source proper motions over the celestial sphere. We show that the distribution of redshifts for ICRF sources is consistent with the much larger sample drawn from FIRST and SDSS, implying that the ultra-compact VLBI sources are not distinguished from the overall radio-loud quasar population.
In addition, we obtained NOT spectra for five radio sources from the FIRST and NVSS catalogs, selected on the basis of their red colors, which yielded three quasars with $z>4$.
author:
- 'O. Titov'
- 'Laura M. Stanford'
- 'Helen M. Johnston'
- 'T. Pursimo'
- 'Richard W. Hunstead'
- 'David L. Jauncey'
- 'K. Maslennikov'
- 'A. Boldycheva'
title: 'Optical Spectra of Candidate International Celestial Reference Frame (ICRF) Flat-Spectrum Radio Sources'
---
Introduction
============
The coming astrometric space mission, [*Gaia*]{} [@Perryman2001; @Mignard2012], to be launched in 2013 by the European Space Agency, will measure high precision positions ($\sim26\mu$as for $V=15$ mag, $\sim300\mu$as for $V=20$ mag, [@debruijne]) and proper motions of $\sim$500,000 quasars brighter than $m_{v} = 20$. This new optical astrometric catalog will be linked with the current radio astrometric catalog, ICRF2, the second realization of the International Celestial Reference Frame [@Fey2009]. Optical counterparts of the extragalactic sources are being sought to confirm their identification as quasars.
This paper is the second in our series aimed at investigating the optical spectra of radio sources in the International VLBI Service for Geodesy & Astrometry (IVS) Reference Catalog; see @Schluter2007 for a description of the IVS history. Astrometric Very Long Baseline Interferometry (VLBI) measures the differences in arrival times of radio waves from ultra-compact, flat-spectrum radio sources at telescopes positioned large distances apart. This procedure determines the positions of such sources to milliarcsecond precision.
The International Celestial Reference System (ICRS) [@arias1995] was adopted by the International Astronomical Union (IAU) as a reference system with its origin at the barycentre of the Solar system [@MacCarthy2004], and axes fixed by the positions of selected extragalactic radio sources. The first realization of the International Celestial Reference Frame (ICRF1) was used to establish the orientation of the ICRS axes [@Ma1998]. The current International Celestial Reference Frame, known as ICRF2, is based on a catalog of 295 ‘defining’ soures. The formal weighted errors in the radio positions are reported by @Fey2009 to have an upper limit to the noise floor of $41~\mu$as.
The IVS astrometric program has a total catalog of $>6000$ radio sources, where $\sim 1200$ are observed on a regular basis. In the Southern Hemisphere, there is a significant deficit in candidate sources as well as a lack of optical identifications. By 2012 February, of the 3257 objects with measured redshifts only 1213 are in the southern hemisphere and only 287 have declinations south of $-40\degr$. This paucity of redshifts in the south leads to problems in the analysis of apparent proper motions of the reference radio sources [@Titov2009]. To address this issue, an extensive program was started in 2010 to find optical counterparts and determine redshifts for southern IVS sources [@Titov2011].
The quasars at high redshift ($z\geq2$) will be used for more intensive observations at the VLBI facilities in the southern hemisphere. VLBI observations of weak sources will be undertaken with the 64-meter telescope in Parkes, Australia, and two 26-meter telescopes in Hobart (Australia) and HartRAO (South Africa). Stronger radio sources (flux density $\geq$400 mJy) will be monitored with four 12-meter telescopes recently installed in Australia: the AuScope network comprising Hobart, Yarragadee and Katherine in Australia [@Titov2013] and Warkworth in New Zealand. Several quasars found during the first observing run with the NTT in 2010 August [@Titov2011] have now been tracked with the AuScope radio telescopes in 2011–2012.
In this paper we continue our spectroscopic observations of the optical counterparts of southern IVS sources, in particular those with a long VLBI observational history. Some strong radio sources from the northern hemisphere have also been observed. Optical identifications were sought initially from the image and catalog data from the SuperCOSMOS Sky Surveys on the grounds of their small digitization pixel scale and excellent astrometric accuracy [@Hambly2001]. We also took advantage of the Sloan Digital Sky Survey (SDSS; @York2000) DR8 data release in the regions where it was available. This improved the identification process, especially in regions of high stellar density.
In addition, we observed five weaker radio sources from the NVSS (NRAO VLA Sky Survey; @Condon1998) and FIRST catalogs (Faint Images of the Radio Sky at Twenty cm; @Becker1995) that we identified with objects from the SDSS whose colors were typical of high redshift quasars. Such high redshift quasars provide unique information about the early stages of the Universe. The number of known radio sources with $z\geq4$ is small, and we are exploiting this technique in an attempt to increase their number.
The observations and data reduction procedures are described in Section 2 and we report our results, along with detailed comments on individual objects, in Section 3.
Observations
============
Spectroscopic observations were carried out at five optical facilities.
[**ESO NTT**]{}: We had a 5-night observing run in Visitor Mode at the European Southern Observatory (ESO) 3.58-meter New Technology Telescope at La Silla in 2011 December (088.A-0021 (A)) using the ESO Faint Object Spectrograph and Camera (EFOSC) system with grism \#13 covering the wavelength range 3685–9315 Å. The seeing during observations was typically $0\farcs{5} - 1\farcs{5}$, with a wavelength resolution 21.2 Å. Exposure times varied from 10 minutes to 1 hour depending on the magnitude of each target and current sky conditions. Wavelength calibration was performed using the spectra of a HeNeAr comparison lamp, resulting in an rms accuracy of 0.5 Å.
[**Gemini**]{}: A large number of targets were observed in Service Mode at the Gemini North and Gemini South 8.2-meter telescopes through the Poor Weather Program (GS-2011A-Q-89, GN-2011B-Q-109, GS-2011A-Q-94) using the Gemini Multi-Object Spectrograph (GMOS) system with grating R400. This grating covers 4500 Å centered either at 5200 Å or 6500 Å. As expected, the seeing and weather conditions were variable but the program overall was very successful. The wavelength resolution was $\sim$15 Å, and an exposure time of 20 minutes was used for all targets. Wavelength calibration was performed using the spectra of a CuAr lamp, resulting in an rms accuracy of $\sim$0.3 Å.
[**NOT**]{}: Observations with the 2.5-meter Nordic Optical Telescope on La Palma were carried out using the Andalucia Faint Object Spectrograph and Camera (ALFOSC) spectrograph, either with grism \#5, or grism \#4 with the WG345 blocking filter. The nominal wavelength range for grism \#4 is 3200–9100 Å, with the second-order blocking filter cutting below 3560 Å. The red end of the detector suffers fringing, so the effective long-wavelength limit is about 8000 Å. For grism \#5 the nominal range is 5000–10250 Å. The slit width was $1\farcs{0}$ or $1\farcs{3}$ depending on the seeing. The typical integration time was between 20 and 40 minutes. The longer integrations were divided into two and the target was offset along the slit in order to improve the fringe correction. For the single integrations, internal halogen lamp images were taken before and after the science frame. Wavelength calibration was based on a HeNe lamp exposure taken before the science frame(s), resulting in an rms accuracy of $\sim$0.5 Å.
[**BTA**]{}: Two objects were observed in Visitor Mode at the 6-meter Big Azimuthal Telescope (BTA) telescope, of the Special Astrophysical Observatory in Russia in 2011 August, using the SCORPIO multi-mode focal reducer with GR300 grism covering the wavelength range 3500–9500 Å. The seeing during observations was about $2\arcsec$. Spectral resolution was typically 20 Å.\
Data reduction was performed with the [iraf]{} software suite[^1] using standard procedures for spectral analysis. We removed the bias and pixel-to-pixel gain variations from each frame and then removed cosmic rays using the IRAF task [szap]{}. Where more than one exposure was obtained, the separate exposures were combined. Spectrum extraction, sky subtraction and wavelength calibration were then carried out and the final one-dimensional spectra were flux-calibrated with a spectrophotometric standard observed with the same instrumental setup. Because the conditions were often non-photometric, especially for observations made through the Gemini Poor Weather Program, the flux calibration should be taken as approximate.
Results
=======
Spectra of 120 IVS objects are shown in Fig. \[spectra\], along with the line identifications. A blue, dashed line indicates lines that were used for redshift calculation, while a red, dot-dashed line indicates lines that were detected, generally at a low signal-to-noise ratio (S/N), but not used in determining the mean redshift.
Table \[emlines\] lists the IVS sources with their ICRF2 coordinates (which refer to the epoch J2000.0), the telescope used for each spectrum, the identified emission lines with their rest and observed wavelengths, the mean redshift and error, and brief notes on individual sources. More detailed notes on individual sources (indicated by an asterisk in the final column) are given in section 3.2.
The quoted errors $\Delta z$ in the mean redshift $\overline{z}$ are given by
$ \Delta z = \{[(\sigma_z)^2 +
(\Delta\lambda/\overline{\lambda_0})^2]/N\}^{1/2}, $
where $\sigma_z$ is the measured standard deviation among the independent estimates of $z$, $\Delta\lambda$ is the rms error in the wavelength calibration (typically 0.5 Å), and $\overline{\lambda_0}$ is the mean rest wavelength of the $N$ lines used to measure $\overline{z}$. Single-line redshifts (mostly ) are assigned a conservative error of 0.001 if the signal-to-noise ratio is high and the line is symmetric. If the signal-to-noise ratio is low or the line is broad or asymmetric an (arbitrary) error of 0.002 has been assigned; in two extreme cases where the signal-to-noise ratio is low [*and*]{} the line is broad or asymmetric (IVS B0633$-$26B and B1129$-$161) the redshift is given with a colon(:) appended and no error.
Nineteen objects, listed in Table \[BLL\] with their ICRF2 positions [@Fey2009], were found to have a good signal-to-noise ratio (typically S/N$\sim60-110$) but featureless spectra and hence are identified as probable BL Lac objects. Their spectra are shown in Fig. \[BLLspectra\].
A further 18 targets had spectra with a signal-to-noise ratio that was too low for confident spectral classification; these are listed in Table \[fa\] with their ICRF2 positions.
There were five IVS targets that returned stellar spectra. This was assumed to be the result of foreground obscuration and, in most cases, small but significant offsets between SuperCOSMOS optical and ICRF2 radio positions. In two cases the correct identification was found when the fields were reobserved in excellent seeing. Further discussion of these five objects is given in Section \[close\_objects\].
Separation of close objects\[close\_objects\]
---------------------------------------------
Occasionally, Galactic stars are found close on the sky to the radio position, leading to possible misidentification. Here we note several cases that were encountered during the observing runs.
- IVS B0900$-$664—the spectrum obtained was that of a red M star, which is offset by $0\farcs{52}$ from the ICRF position. A faint counterpart was seen in the NTT B-band acquisition image, but a much longer integration in good seeing will be needed to secure a redshift.
- IVS B0905$-$202—the nearest object, as seen by SuperCOSMOS, was a $R=14.6^{\rm m}$ stellar object located $1\farcs{55}$ from the radio position. Our NTT acquisition image, taken in $0\farcs{6}$ seeing, showed a faint object at the radio position, but on the limb of the stellar disk. Spectroscopy of this faint $R\sim 22$ object did not reveal any clear emission lines.
- IVS B1657$-$261—located in a very crowded star field in the Galactic bulge at a latitude of $b=9\fdg{7}$. Identification of the optical counterpart may not be feasible.
- IVS B1946$-$582—based on the SuperCOSMOS optical position, the radio minus optical position difference is only $0\farcs{06}$ in right ascension and $-0\farcs{04}$ in declination, but the Gemini South observation of this $R\sim 19$ object showed a typical stellar spectrum. The radio source has a flat spectrum with flux density 200–300 mJy at cm wavelengths.
- IVS B2300$-$307—the obscuration of the optical field of this source by a foreground star was discussed previously [@Titov2011]. However, a later 80s image in $0\farcs{6}$ seeing (Fig.\[2300images\]) revealed a faint $R\sim 22$ object coincident with the radio position. Our spectroscopy yielded a redshift of $1.039\pm 0.002$ based on weak \] and emission.
![Two acquisition images of the sky field around the quasar IVS B2300$-$307 made at the NTT in 2010 August (left: seeing $2''$) and 2011 December (right: seeing $0\farcs{6}$). Each image is 50$''$ on a side; north is up and east to the left. The foreground star in the centre of the left image completely obscures the quasar due to its large seeing disk. In the right image the much better seeing reveals the faint quasar ($R\sim$22, marked between the bars), separated by $3\farcs{7}$ from the star. (The diagonal step in contrast in the right image is an artefact of fast readout using two amplifiers for the acquisition image.)[]{data-label="2300images"}](fig3b.eps "fig:"){height="3.5cm"} ![Two acquisition images of the sky field around the quasar IVS B2300$-$307 made at the NTT in 2010 August (left: seeing $2''$) and 2011 December (right: seeing $0\farcs{6}$). Each image is 50$''$ on a side; north is up and east to the left. The foreground star in the centre of the left image completely obscures the quasar due to its large seeing disk. In the right image the much better seeing reveals the faint quasar ($R\sim$22, marked between the bars), separated by $3\farcs{7}$ from the star. (The diagonal step in contrast in the right image is an artefact of fast readout using two amplifiers for the acquisition image.)[]{data-label="2300images"}](fig3a.eps "fig:"){height="3.5cm"}
Notes on individual targets
---------------------------
- IVS B0233$-$478—the broad feature attributed to and \] may also include $\lambda$1890.
- IVS B0417$-$302—prominent multiplet emission; intervening doublet absorption at $z_{\rm abs}=0.8395 \pm
0.0005$.
- IVS B0447$-$507—possible double-peaked \] line and prominent multiplet bands.
- IVS B0448$-$482—poor redshift match to \] may indicate the presence of $\lambda$1890; prominent multiplet bands.
- IVS B0521$-$262—strong associated absorption in the Ly$\alpha$ emission line and a clear detection of the Lyman limit at the emission redshift.
- IVS B0529$-$031—faint optical counterpart with a very red spectrum, showing a single broad emission line identified as with associated absorption at $z_{\rm abs}=1.8744 \pm 0.0004$, corresponding to a relative blue-shift of $\sim$1500kms$^{-1}$; a further absorption line, at 8182.1Å remains unidentified. Given the low signal blueward of 6000Å, together with the faint SDSS blue magnitudes ($u= 24.41,\ g=21.93$), it is not surprising that \] and are not detected.
- IVS B0548$-$527—redshift in Table \[emlines\] is for the forbidden lines; the permitted lines of and H$\beta$ show a redshift systematically higher by $830\pm 80$kms$^{-1}$.
- IVS B0554+242—narrow self-absorption in Ly$\alpha$ and at $z_{\rm abs}=3.2319 \pm 0.0010$.
- IVS B0608$-$230—wavelengths for \] and are not consistent, suggesting the presence of additional lines or asymmetric structure; in addition, the Ly$\alpha$ wavelength is likely to be affected by absorption in the blue wing.
- IVS B0633$-$26B—a very faint galaxy; a low signal-to-noise ratio spectrum, with a single broad emission line, assumed to be .
- IVS B0810$-$180—very broad line wings in and Ly$\alpha$, extending $\sim$30,000kms$^{-1}$ redward of the line peaks.
- IVS B0828$-$064—poor consistency in redshift between and \], possibly due to associated absorption in the blue wing of . is present with a low signal-to-noise ratio just redward of the atmospheric A-band.
- IVS B0844$-$557—redshift in Table \[emlines\] is for the forbidden lines; permitted lines and H$\beta$ are displaced $\sim$1000kms$^{-1}$ to higher redshift.
- IVS B0948$-$860—single emission line, assumed to be Ly$\alpha$; the adopted redshift, $z=3.696$, is consistent with the 40% continuum depression blueward of the emission line due to the Ly$\alpha$ forest, and the possible detection of at the red edge of the spectrum.
- IVS B0952$-$185—associated absorption at the emission redshift in both Ly$\alpha$ and .
- IVS B0956$-$409—Ly$\alpha$ is strongly self-absorbed.
- IVS B1004$-$125—reported as having $z=0.24$ in Simbad (no reference given), clearly inconsistent with our Gemini spectrum; strong intervening heavy-element system at $z_{\rm abs}=1.5786 \pm 0.0002$ based on $\lambda\lambda$2344, 2382, 2586, 2600 and $\lambda\lambda$2796, 2803 absorption.
- IVS B1020+270—extended red wing in ; \] is not detected due to strong CCD fringing.
- IVS B1039$-$474—, and \] emission lines all show strongly extended blue wings; redshift is based on the peak positions of the lines.
- IVS B1127$-$443—single-line redshift is supported by stellar absorption features (G-band, b) noted in the plotted spectrum.
- IVS B1143$-$696—spectrum is very similar to that of 3C273, with strong multiplet emission in the region around H$\beta$ [@wampler1967].
- IVS B1722+562—strong associated absorption system at $z_{\rm
abs}=2.2463 \pm 0.0002$ seen in , , and Ly$\alpha$. This has led to a relatively large uncertainty in emission redshift.
- IVS B2235$-$556—strong self-absorption in the red wing of .
- IVS B2334$-$525—broad , possibly double-peaked with an extended red wing.
- IVS B2341+295—very blue object based on POSSII sky survey images; rise in the spectrum redward of 7000Å may be due to the underlying galaxy or an unrelated object on the spectrograph slit. Intervening absorber at $z=0.8644 \pm 0.0004$.
Spectra of color-selected quasars
---------------------------------
In an attempt to find more high-redshift radio quasars for our VLBI proper motion studies as a function of redshift, we selected five weak radio sources from the FIRST and NVSS catalogs for which the SDSS colors suggested the likelihood of $u$- or $g$-band dropouts and the possibility of high redshifts. Spectra were obtained at the NOT on the nights of 2012 May 24–26. The SDSS colors, wavelengths and redshifts for the five color-selected radio quasars are given in Table \[nvss\_red\]. The ‘bluest’ of the five, NVSS J125944+240707, was a $z = 1.139$ radio galaxy, but the other four proved to be broad emission-line quasars with redshifts in excess of 3.5, including three with $z>4$. This is an excellent return for our search and a striking result from the NOT, the smallest of the telescopes used in this program. The five spectra are shown in Fig.\[nvss\] and further information is given in Table \[nvss\_red\]. Redshifts for the $z>4$ quasars are based on estimated wavelengths for Ly$\alpha$ (and, in one case, Ly$\beta$) and are very uncertain because of strong absorption in the blue wing of the line.
A comparison between the FIRST and NVSS flux densities at 1.4GHz for the five sources showed that one, NVSS J145459+110928, appeared to show evidence of variability, with $S_{\rm NVSS}=9.8 \pm 0.5$ mJy and $S_{\rm
FIRST}= 15.07\pm 0.14$ mJy. This is also the only source that is unresolved in FIRST. The other four sources all show minor extension at the $1\arcsec$ level.
Redshift distributions
======================
Together with the 31 redshifts reported in [@Titov2011], we have now accumulated over 150 redshifts for IVS sources, mostly in the south. Their optical counterparts are systematically fainter than those in the [@Titov2009] compilation. Since the IVS selection process is directed at compact, milliarcsecond radio quasars, it is important to test whether their redshift distribution differs from that of the FIRST-SDSS quasar sample [@kimball2011], which is selected without regard to morphology, spectral index or angular size and extends to much lower flux densities.
The redshift distribution of the sources from this paper and [@Titov2011] is shown in Figure \[zdist\](a). For comparison, Figure \[zdist\](b) shows the distribution of 1594 point sources with known redshifts from ICRF2, using redshift data from [@Titov2009]. The SuperCOSMOS [@Hambly2001] morphological classification was used to remove galaxies from the [@Titov2009] list, and the sample was further restricted to those with radio-optical offsets $<1''$ and Galactic latitudes $>10^{\circ}$ (Schaefer et al. 2013, in prep.).
The redshift distribution for the sources from this paper and the earlier paper [@Titov2011] is completely consistent with having been drawn from the same redshift distribution as FIRST-SDSS quasars [@kimball2011], with a Kolmogorov-Smirnov test giving a probability $p=0.84$ of the two samples being drawn from the same distribution. Thus the sharp drop-off that we see in the number of quasars at $z \gtrsim 2$is a reflection of the underlying quasar distribution, and not a selection effect of the IVS candidate sources. While the K-S plot for the ICRF2 sample shows an apparent excess of low-redshift sources compared with FIRST-SDSS, the difference is not significant, with a K-S probability of $p=0.24$. We attribute this excess to selection effects arising from the redshifts in [@Titov2009] being drawn from the literature, and therefore likely to be biased towards brighter, lower-redshift quasars.
\
(a) (b)
Summary and conclusion
======================
We present redshifts and spectra for 120 emission-line objects identified with radio sources from the candidate International Celestial Reference Catalog. Most of the target objects are in the south and many had not previously been optically identified. While redshifts were usually based on two or more lines, those for 22 objects were based on a single emission line, in most cases assumed to be ; many of these single-line redshifts were supported by other spectral information and most are considered reliable. In addition, we classed 19 sources as probable BL Lac objects, based on a high signal-to-noise ratio but featureless spectra. A further 18 targets were considered to have a signal-to-noise ratio too low for confident spectral classification.
The distribution of redshifts from this paper, together with those from our earlier paper [@Titov2011], is consistent with the much larger sample drawn from FIRST and SDSS [@kimball2011]. This implies that the ultra-compact, flat-spectrum sources that make up the IVS Reference Catalog are not distinguished from the radio quasar population at large. On the other hand, the distribution of redshifts from the much larger sample drawn from ICRF2 [@Titov2009] has a small, but not significant excess of low-redshift quasars, almost certainly the result of observational selection.
Acknowledgments
===============
This paper is based on observations collected at five telescopes:
1\. ESO New Technology Telescope, under the European Organisation for Astronomical Research in the Southern Hemisphere, Chile under program 088.A-0021(A).
2\. Two Gemini Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the National Science Foundation on behalf of the Gemini partnership: the National Science Foundation (United States), the Science and Technology Facilities Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), Ministurio da Ciuncia, Tecnologia e Inovacio (Brazil) and Ministerio de Ciencia, Tecnologia e Innovacion Productiva (Argentina) under programs GS-2011A-Q-89 and GS-2011B-Q-94 (Gemini South), and GN-2011B-Q-109 (Gemini North).
3\. Six-meter Big Azimuthal Telescope (BTA) operated by the Special Astrophysical Observatory (Russia)
4\. Nordic Optical Telescope, operated on the island of La Palma jointly by Denmark, Finland, Iceland, Norway and Sweden, in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofisica de Canarias.
Two of us, Titov and Jauncey, were supported by a travel grant from the Australian Nuclear Science Technology Organisation (ANSTO) in their Access to Major Research Facilities Program (AMRFP) (reference number AMRFP 10/11-O-31) to travel to the BTA telescope in Russia.
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, University of Cambridge, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
AuScope is funded under the National Collaborative Research Infrastructure Strategy (NCRIS), an Australian Commonwealth Government Programme.
This paper is published with the permission of the CEO, Geoscience Australia.\
© Commonwealth of Australia (Geoscience Australia) 2013. This product is released under the Creative Commons Attribution 3.0 Australia Licence. http://creativecommons.org/licenses/by/3.0/au/deed.en
Afanas’ev, V. L., Dodonov, S. N., Moiseev, A. V., et al. 2009, ARep, 53, 287
Arias, E. F., Charlot P., Feissel, M., & Lestrade, J.-F. 1995, A&A, 303, 604
Becker, R. H., White, R. L., & Helfand, D. J., 1995, , 450, 559
Condon, J. J., Cotton, W. D., Greisen, E. W., et al. 1998, , 115, 1693
de Bruijne, J. H. J., 2012, Ap&SS, Vol 341, Issue 1, 31
Fey, A., Gordon, G., & Jacobs, C. (eds.), 2009, The Second realization of the International Celestial Reference Frame by Very Long Baseline Interferometry, IERS Technical Notes 35, Verlad des Bundesamts fur Kartographie und Geodasie, Frankfurt am Main
Hambly, N. C., Davenhall, A. C., Irwin, M. J., & MacGillivray, H. T., 2001, , 326, 1315
Jones, D. H., Read, M. A., Saunders, W., et al. 2009, , 399,683
Kimball, A. E., Ivezic, Z., Wiita, P. J., & Schneider, D. P., 2011, , 141, 182
Ma, C., Arias, E. F., & Eubanks, T. M., 1998, , 116, 516
MacCarthy, D. & Petit G. (eds.), 2004, IERS Covnentions (2003), IERS Technical Notes 32, Verlad des Bundesamts fur Kartographie und Geodasie, Frankfurt am Main
Mignard, F. & Klioner, S., 2012, , 547. A59
Ojha, R., Fey, A. L., Charlot, P., et al. 2005, , 130, 2529
Perryman, M. A. C., de Boer, K. S., Gilmore, G., et al. 2001, , 369, 339
Quiniento, Z. M. & Cersosimo, J. C. 1993, , 97, 435
Schl[ü]{}ter, W. & Behrend D. 2007, JG, 81, 379
Shaw, M. S., Romani, R. W., Cotter, G. H., et al. 2012, , 748, 49
Titov, O., & Malkin, Z. 2009, , 506, 1477
Titov, O., Jauncey, D. L, Johnston, H. M., et al. 2011, , 142, 165
Titov, O., Dickey, J. M., Lovell, J. E. J., & McCulloch, P. M., 2013, Proc. IAG Symp. 139
Wampler, E. J. & Oke, J. B. 1967, , 148, 695
York, D. G., Adelman, J., Anderson, J. E., Jnr, et al. 2000, , 120, 1579
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
{width="85.00000%"}
[^1]: IRAF is distributed by the National Optical Astronomical Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under contract to the National Science Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Decision making in uncertain and risky environments is a prominent area of research. Standard economic theories fail to fully explain human behaviour, while a potentially promising alternative may lie in the direction of Reinforcement Learning (RL) theory. We analyse data for 46 players extracted from a financial market online game and test whether Reinforcement Learning (Q-Learning) could capture these players behaviour using a risk measure based on financial modeling. Moreover we test an earlier hypothesis that players are “na[ï]{}ve” (short-sighted). Our results indicate that a simple Reinforcement Learning model which considers only the selling component of the task captures the decision-making process for a subset of players but this is not sufficient to draw any conclusion on the population. We also find that there is not a significant improvement of fitting of the players when using a full RL model against a myopic version, where only immediate reward is valued by the players. This indicates that players, if using a Reinforcement Learning approach, do so na[ï]{}vely.'
author:
-
-
-
title: |
Modelling Stock-market Investors as\
Reinforcement Learning Agents \[Correction\]
---
Introduction
============
{width="7.25in"}
One of the most challenging fields of research is human decision-making. Understanding the processes involved and trying to predict or replicate behaviours has been, historically, the ultimate goal of many disciplines. Economics for example, has a long tradition of trying to formalise human behaviour into descriptive or normative models. These models have been employed for several years (e.g. Expected Utility model [@J.vonNeumann1947]) but have been proven to be inadequate [@Starmer2000; @ShaneFriederickGeorgeLoewenstein2002; @Tversky1974; @Kahneman1991], giving rise to new research areas like behavioural and experimental economics. Psychology as well, is natively concerned with decision-making. Sequential decision problems have been used to evaluate people’s risk attitude, in order to predict actual risk proneness in real life scenarios [@Hoffrage2003; @Pleskac2008; @Wallsten2005]. While economics and psychology are focused on the high-level manifestations and implications of decision-making, neuroscience aims at understanding the biological machinery and the neural processes behind human (or animal) behaviour [@Britten1996; @Britten1992; @Gold2001; @Gold2002; @Gold2007; @Shadlen1996; @Shadlen1996a].
Recently these fields of research have started to collaborate, contributing to the rise of an emerging multi-disciplinary field called neuroeconomics [@Sanfey2006; @Glimcher2004; @Loewenstein2008; @Barraclough2004; @Soltani2006]. This discipline approaches the problem from several perspectives and on different levels of abstraction. RL is a theoretical framework [@Sutton1998], extensively used in neuroeconomics literature for addressing a wide array of problems involving learning in partially observable environments [@Doya2007; @Tesauro1994; @AbbeelPCoatesAMorganQNg2006; @Hafner2007; @Walker2000; @Mnih2015]. RL is based on the concept of reward/punishment for the actions taken. The agents act in an environment of which they possess only partial knowledge. To be able to achieve the best behaviour, i.e. maximise their reward, the agents have to learn through experience and update their beliefs. Learning happens as a result of the agent’s interpretation of the interactions with its surroundings and the consequences of a “reward” feedback signal. The ability of this framework to model and therefore understand behavioural data and its underlying neural implications, is of pivotal importance in decision making [@Dayan2008].
RL can accurately capture human and animal learning patterns and has been proven effective at describing the functioning of some areas of the human brain, like the basal ganglia, and the functions of neurotransmitters such as dopamine [@Schultz1997; @Doya2000; @Daw2006; @Doya2007]. One of the most remarkable similarities between biological functioning and RL models is the one about Temporal Difference (TD) error [@Sutton1998; @Watkins1989; @Werbos1990; @Hikosaka2006] and the activation of mid-brain dopamine neurons [@SchultzWRomoRLjungbergTMirenowiczJHollermanJRandDickson1995; @Suri1998; @Waelti2001; @Satoh2003; @Nakahara2004; @Morris2006]. These findings supported the notion that TD Learning is implemented in the brain with dopaminergic neurons in the striatum [@Schultz1997; @Barto1995; @Montague1996; @Samejima2005; @Kawagoe1998; @Kawagoe2004; @Hikosaka2006; @Schultz2002; @Day2007; @Costa2007; @Hyman2006; @Joel2002; @Wickens2007], making it a reasonable first choice for a modelling attempt. Humans and animals are very advanced signal detectors whose behaviour is susceptible to changes in the rewards resulting from their choices [@Stocker2006; @Kording2007]. Both neuroscience and psychology have extensively employed tasks in which the exploration-exploitation trade-off was of crucial importance [@Daw2006a; @Soltani2006; @Luksys2009; @Frey2015; @Benartzi2015]. It is crucial for the individuals to maximise their reward using the information at their disposal but to do so advantageously they need to learn which actions lead to better rewards. Decision making in uncertain environments is a challenging problem because of the competing importance of the two strategies: exploitation is, of course, the best course of action, but only when enough knowledge about the quality of the actions is available, while exploration increases the knowledge about the environment.
A complicated task that encompasses all these features is stocks selection in financial markets, where investors have to choose among hundreds of possible securities to create their portfolio. Stock trends are non-monotonic because they are not guaranteed to achieve a global maximum and the future distribution of reward is intrinsically stochastic. After purchasing a stock, investors are faced with the decisions on when to sell it (Market timing problem [@Benartzi2015]). To be able to achieve the best return from their investments, people need to be careful in considering how to maximise their profit in the long term and not only in a single episode. We speculate that RL is part of the decision making process for investors. This speculation is supported by Choi et al. [@Choi2009], who studied individual investors decisions on 401(k) savings plans. Over the years, investors could decide to increase or decrease the percentage of their salary to commit to this retirement plan. Their results suggest that investors’ decisions are influenced by personal experiences: they show that those investors who have experienced a positive reward from previous investment in their 401(k) fund, tend to increase the investment in that particular product, compared to those who experienced a lower reward outcome. This kind of behaviour follows a “na[ï]{}ve reinforcement learning” and is in contrast with the disposition effect [@Barber2013; @Odean1998](the unwillingness of investors to sell “losing” investments). Huang et al. investigated how personal experience in investments affects future decisions about the selection of stocks [@Huang2012]. They used data that spans from 1991 to 1996, from a large discount broker. Again, the pattern of repurchasing financial product which yielded positive return was found. As Huang suggests, by understanding the way past experience affects investors’ decisions, it might be possible to make predictions about the financial markets involved. RL has also been used, with promising results, to develop Stock Market Trading Systems [@Chen2007; @Lee2001; @O2006; @Moody2001] and to build Agent Based Stock Market Simulations [@Rutkauskas2009]. While these works use RL to predict future prices, they do not try to describe human behaviour. With these notions as background we decided to investigate and try to model human choices in a stochastic, non-stationary environment. We hypothesise that RL is a component of decision making and to test this we compare two RL models against a purely random one. Our modelling attempts are based on two assumptions. First, we assume that risk is a proxy of the internal representation of the actions for some players. To test this we use a measure of systematic risk widely used in finance and economics to categorise the different choices into three discrete classes. We also assume that the reward signal is based on the cash income arising from the sales an investor makes. This assumption follows a widely researched behaviour referred to as “disposition effect” in literature [@Barber2013; @Odean1998], the tendency of individual investors to sell stocks which increased in value since when they were purchased, while holding onto the stocks which lost value. This phenomenon is stronger for individual investors but it also exhibited by institutional investors, such as mutual funds and corporations [@Brown2006; @Barber2007; @Frazzini2006; @Grinblatt2001; @Heath1999; @Dean1998; @Shapira2001]. Following these indications we mapped the sell transactions to a reward signal to fit our models. Finally, we hypothesise that not all players are short-sighted, to test this we compare a full RL model (3 free parameters) against a myopic model (2 free parameters, no gamma). The difference is that the latter can be considered a na[ï]{}ve RL as it does not take into account future rewards, it only seeks to maximise immediate rewards.
![Comparison of the ranked vs scrambled discretisation. Test for the assumption that investors use risk to internally classify the different stocks in discrete degrees of risk. The errorbars represent the probability and 99$\%$ confidence intervals of the comparison of the stock discretisation based on the $\beta_F$ risk measure as opposed to 500 randomly generated scrambled discretisations. The 7 players for which the RL model outperformed the random model all have a statistically significative probability of using a representation of actions discretisation based on risk. The comparison of the entire dataset resulted in 31 players out of 46 ($\sim 67 \%$) to have a probability and CI above chance. []{data-label="fig:ranked_v_scrambled"}](ranked_v_scrambled_big_dots.png){width="3.5in"}
Method
======
Dataset
-------
The dataset has been extracted from the publicly accessible online trading simulation game VirtualTrader[^1], which is managed by IEX Media Group BV in the Netherlands. Players can subscribe for free and start playing the game with an assigned virtual cash budget of 100k GBP. The players will then pick the stocks they prefer from the FTSE100 stock index pool (107 stocks at the time of data collection) and create their own portfolio. These competitors are ranked according to the return of their investment. This is composed of “holdings” and “cash”. The former represent the shares possessed by a player while the latter is the amount of money not invested (i.e. deriving from sold stocks or never invested). The simulation follows real world data evolution, for example price fluctuations and price splits. The delay is usually in the order of 10-15 minutes and the player can access a visual representation of the stocks time series. All the transactions are stored for each player. For this study we considered transactions that span from the 1st of January 2014 to the 31st of May 2014. This time period has been chosen because at that time the player ranking was determining the winner of the monthly prize giveaway. Two possible rewards can be identified: a psychological one, consisting of the ranking position and a tangible one being the prize for the highest achiever.
The transactions have been stored in a database in order to be manipulated and used to fit models with different combination of free parameters. The rows are structured in 6 fields: Date, Type, Stock, Volume, Price and Total. The dataset initially contained about 100k transactions that were reduced to about 1.4k. This was due to preprocessing, which removed the many instances of inactive players who played only at the beginning and/or at the end of the time frame considered. In the final version of the dataset there are 46 players. The average amount of transactions per player is 30. The player who played the most during the six months performed 107 transactions. We considered the full amount of transactions each player operated in the game.
Reinforcement Learning Setup
----------------------------
We adopted a widely used off-policy RL framework called Q-learning [@Sutton1998]. The learning rule of this model is:
$$\label{q-learning_equation}
\Delta Q(s_t,a_t) = \alpha \bigg [ r_{t+1} + \gamma \max\limits_{a} Q(s_{t+1},a) - Q(s_t,a_t) \bigg]$$
where $Q(s_t,a_t)$ represents the value of action $a$ while in state $s$, at time $t$. $\alpha \in [0,2]$ is the step-size parameter and controls the rate of learning. $\gamma \in [0,1]$ is the discount factor and represents how far-sighted the model is, It encodes how much a future reward is worth at time $t$. When $\gamma = 0$ only immediate rewards are taken into account by the player.
To test this framework the task has been mapped as follows. There are two states (win, loss) calculated according to the profit of the player (details in equations \[eq:profit\_equation\] and \[eq:State\_equation\]). These two states reflect the dichotomy rooted in the Prospect Theory’s value function gain/loss spectrum [@Kahneman1979].
Since all players begin with the same initial budget our calculation of the profit uses the returns accumulated by selling stocks. This choice reduces the scope of the model, focusing on the cash component of the players assets. This will be referred to as the “Sell” model. The actions are mapped to the stocks available for trading.
In order to avoid dimensionality issues, 107 stocks for 2 states give rise to 214 potential actions, we decided to classify the stocks in 3 classes of risk using a widely used financial modelling measure, $CAPM \ Beta$. The acronym stands for Capital Asset Pricing Model, a model developed by Sharpe[@Sharpe1964] used to explain the relationship between the expected return of a security and its risk . In this report we will refer to financial volatility measure $CAPM \ Beta$ as $\beta_F$:
$$\label{beta_equation_covariance_variance}
\beta_F = \frac{Cov(r_a,r_b)}{Var(r_b)}$$
This financial modelling measure quantifies the volatility of a security in comparison to the market or a reference benchmark [@Beninga2000]. Relatively safe investments like utilities stocks (e.g. gas and power) have a low $\beta_F$, while high-tech stocks (e.g. Nasdaq or MSH Morgan Stanley High-Tech) have a high $\beta_F$.
As an example, the $\beta_F$ of the index of reference (that represents the portion of the market considered) is exactly 1. A $\beta_F \in (0,1)$ indicates that the asset has a lower volatility compared to the market or low correlation of the asset price movements compared to the market. While if $\beta_F > 1$ it signifies an investment with higher volatility compared to the benchmark. Following the previous example, high-tech securities with a $\beta_F > 1$ could yield better returns compared to their benchmark index, when the market is going up. This also poses more risk because in case the market loses value, the security would lose value at a higher rate than the index.
$\beta_F$ is considered a measure of the systematic risk and can be estimated by regression. Considering the returns of an asset $a$ and the returns of the corresponding benchmark $b$:
$$\label{beta_equation_regression}
r_a = \alpha + \beta r_b$$
$\beta_F$ has been calculated for each stock in the FTSE100 at the time of the game by considering daily returns in the year between 1st June 2013 and 31st May 2014. The measure associated with each stock is used to rank them and subdivide them in three classes, containing respectively 36, 36 and 35 stocks each.
Reward $r$ at time step $t+1$ is defined by the gain (or loss) made in a sell transaction. Buying transactions are kept into account to track players portfolios and to calculate the price difference. They were not used as actions, but we might extend our modelling scenario by integrating a “Buy” model in the future and consider purchase actions by changing the reward scheme. The reward is calculated as:
$$\label{eq:reward_calculation}
r_{t+1} = v_{t+1}\left (p_{t+1} - \frac{1}{t} \sum_{i=1}^{ t} v_i p_i \right )$$
where $v_{i}$ and $p_{i}$ are the volume and the price of the stock traded at the i-th time step. The second term of the difference is a weighted average of the stock prices at previous times.
To avoid numerical instabilities, reward has been flattened with a sigmoid function into the range $[-1,+1]$. Specifically a hyperbolic tangent has been used, with $\rho = 500$ to capture most of the variability of the rewards, only flattening the extreme values. This choice is in line with prospect theory value function which is concave for gains and convex for losses [@Kahneman1979].
$$\label{eq:hyperbolic_tangent}
tanh(r) = \frac{1-e^{-r/\rho }}{1+e^{-r/\rho }}$$
As in this study we focused on the sell subset of the players interactions, the states are based solely on profit, which in turn is based on the reward of the sell transactions. The profit and states are defined as:
$$\label{eq:profit_equation}
Profit = \sum_{t} tanh(r_t)$$
$$\label{eq:State_equation}
State=\begin{cases}
0, & \text{if $Profit<0$}.\\
1, & \text{otherwise}.
\end{cases}$$
The RL framework is composed of a learning model (eq. \[q-learning\_equation\]) and an action-selection model that is responsible for picking the best action. In our setup the former is Q-learning and the latter is Soft-Max. An action $a$ is picked at time $t$ with probability:
$$\label{eq:softmax_equation}
P(a_t) = \frac{e^{Q_{t} (a) \beta}}{\sum_{b=1}^{n} e^{Q_{t}(b) \beta}}$$
where $n$ is the number of available actions (i.e. 3 in this study) and $\beta$ is the inverse temperature parameter and represents the greediness of the action-selection model. In the limit $\beta \rightarrow 0$ the actions become equiprobable and the model reverts to random. Higher values of $\beta$ approximate a greedy model which picks the best known action (fully exploitative). The full model has 3 free parameters: $\alpha$ (step-size parameter or learning rate), $\beta$ (exploration-exploitation trade-off) and $\gamma$ (discount factor). For this study we used a bounded gradient descent search with 27 combinations of initial guess points. These are the combination of values of the free parameters from where the search starts. By having different entry points we hope to reduce the chance of the search getting stuck in a local minimum solution. The search has been performed with the following boundaries:
- $\alpha \in (0.0001, 2)$
- $\beta \in (0, 50)$
- $\gamma \in (0, 0.9999)$ (for the myopic model $\gamma = 0$)
The entry points are the combinations rising from the following values:
- $\alpha \in \left \{0, 0.5, 2\right \}$
- $\beta \in \left \{0, 25, 50\right \}$
- $\gamma \in \left \{0, 0.5, 0.9999\right \}$
The search results have been obtained on `python 2.7.9` and `scipy.optimize.minimize` with `scipy 0.17.1`.
Model Testing Routine
---------------------
Maximum Likelihood Estimate has been used as a measure of the model fitness, following Daw’s comprehensive analysis of methodology [@Daw2009]. MLE is the appropriate method to assess model performance because it evaluates which set of model parameters are more likely to generate the data using a probabilistic approach. Data likelihood is a powerful method because it keeps into account the presence of noise in the choices. It does so by using probability estimates for the potential actions. Given a model $M$ and its corresponding set of parameters $\theta_M$ the likelihood function is defined as $P(D|M,\theta_M)$, where $D$ is the dataset (the list of choices and the associated rewards). Applying Bayes’ rule:
$$\label{eq:bayes_rule_likelihood}
P(\theta_M|D,M) \propto P(D|M,\theta_M) \cdot P(\theta_M|M)$$
The left hand side of the proportionality is the posterior probability distribution over the free parameters, given the data. This quantity is proportional to the product of the likelihood of the data, given the parameters and the prior probability of the parameters. Treating the latter as flat we obtain that the most probable value for $\theta_M$ (the best set of free parameters) is the Maximum Likelihood Estimate (MLE), that is the set of parameter which maximises the likelihood function, $P(D|M,\theta_M)$ and it is commonly denoted $\hat\theta_M$. The likelihood function is maximised through the following process. At each timestep, for every action, the observation model (Soft-Max) estimates a probability. These probabilities are then multiplied together. To avoid numerical problems that could arise when multiplying probabilities, the sum of their logarithm is calculated instead. The negative of this value, also known as Negative Log-Likelihood, is then used. The aim is then to minimise this quantity, which is the equivalent of maximising the likelihood function, $P(D|M,\theta_M)$. In the future we will refer to the Negative Log-Likelihood as MLE for simplicity, keeping in mind that lower values represent better fit.
The values of MLE generated represent the goodness of fit of the model with its associated set of parameters. To compare the selected model with a random model and for statistical significance we adopted the Likelihood Ratio Test [@Crandall1997]. This statistical test uses the likelihood ratio to compare two nested models and takes into account the different number of free parameters of the two. It encapsulates this information, when testing for significance, using the difference of the two amounts as degrees of freedom for the chi-square ($\chi^2$) test. Since the test statistic is distributed $\chi^2$ it is straightforward to estimate the p-value associated with the $\chi^2$ value.
The baseline for comparison is a random model which has 0 parameters as there is no learning involved and the action-selection policy is random ($\frac{1}{3}$ chance of picking any of the three stock bins). The first comparison is between this random model and the simpler of the proposed models, which has only two free parameters ($\gamma = 0$). This setup represent the na[ï]{}ve learning procedure that could explain investors’ behaviour showed in literature[@Choi2009]. The full model, with all the three free parameters, has also been tested against the myopic model to assess whether some players are better fitted by a more complex version of the framework. Finally the model goodness of fit has been evaluated with the adopted action classification (based on risk) against 500 randomly generated stock classifications. This has been done to test the assumption that players internally classify the stock range into discrete degrees of risk.
Results
=======
Results for the test of the hypothesis that RL is a component of decision making are shown in Fig. \[fig:models\_comparisons\]. The best set of parameters was found according to MLE through gradient descent search. The best model MLE has been compared to the random model MLE using the Likelihood Ratio Test [@Crandall1997]. The random model MLE is easily estimated as:
$$\label{eq:random_MLE_estimation}
P(D|M_{rand})\ =\ log \prod^{N_t}\frac{1}{3} \ = \ \sum^{N_t}log\frac{1}{3}$$
where $N_t$ is the number of transactions for each player in the dataset. As shown in Fig. \[fig:models\_comparisons\] (a) and (b) 15$\%$ of the players in our dataset is better fitted by a myopic RL model as opposed to a random model. In Fig. \[fig:models\_comparisons\] (c) and (d) we report an improvement in the fitting for some players using a full RL model against the myopic (nested) version of the model. This improvement is not reflected in the comparison of the full RL model against the random model, as shown in Fig. \[fig:models\_comparisons\] (e) and (f). Most of the players that can be fitted with our models are well represented by a myopic model. These results follow what found by Choi et al.[@Choi2009] and Huang et al. [@Huang2012]. We made the assumption that players, when faced with the choice to trade many stocks (107 for this task), internally model these in discrete groups of risk using readily available information such as stock historical prices and returns, which in turn are used to estimate their volatility ($\beta_F$). To test whether this assumption holds true for the players in our dataset, we ran the simpler version of our model on the risk-ranked discretisation and on 500 independent and randomly scrambled discretisations. The results are shown in Fig. \[fig:ranked\_v\_scrambled\] and are generated using Bayesian Information Criterion (BIC) as a measure for comparison of fitness and Binomial Proportion Confidence Interval calculated with Clopper-Pearson method using `Matlab 8.4.0.150421 (R2014b)` function `binofit`. The BIC has been used as the Likelihood Ratio Test can only be used to compare nested models, while in this case the comparison is between models with the same number of parameters that are tested on different data arrangements. This procedure estimates the probability that the ranked discretisation is better than the 500 scrambled discretisations ($BIC_{ranked} < BIC_s$, where $BIC_s$ is the BIC for the s-th scrambled). The results shown in Fig. \[fig:ranked\_v\_scrambled\] are for 99$\%$ confidence interval. The results shown are only for those 7 players who are fitted significantly by the myopic model. As shown in Fig. \[fig:ranked\_v\_scrambled\], all the players are well above the chance threshold. This indicates that risk based on historical data could be considered a proxy for action selection for the players who are well fitted by our RL myopic model.
Conclusion
==========
We investigated a publicly available dataset consisting of trading transactions operated by players of an investment game. We based the discretisation of the actions on the assumption that risk can capture the internal modelling that players operate when facing this task. This assumption was shown to hold true and be statistically significant for a subset of the players, 31 out of 46 and specifically for the 7 players who are best fitted by a RL model. This could signify that the remaining players might use other types of discretisation techniques based on different measures (or a combination of them) or they do not use technical analysis but fundamental analysis (e.g. using financial statements and reports). In this work, we investigated a model which combines two versions of a Reinforcement Learning framework using Q-learning as an update rule and Soft-Max as action-selection policy on a discretised action space according to the risk measure $\beta_F$. It is possible that different model combinations, which use different learning rules or different measures of risk, fit the players population in our dataset better. It is also likely that, by restricting our focus on the sell model, we missed some features of what constitutes the reward signal that players receive. In the full version of the game, in fact, players might try to maximise both holdings and cash simultaneously, in order to compete in the ranking.
The myopic model is a nested version with only two free parameters, representing the learning rate ($\alpha$) and the degree of greediness ($\beta$). The full version extends the simpler model with a discount factor ($\gamma$) which regulates how much of the future rewards is taken into account when updating the values of present state-action pairs. 15$\%$ of the players are well fitted by a RL model with $\gamma = 0$ and there is no significant improvement of fitting by extending this model including gamma as a free parameter. Previous literature pointed in the direction of investors being na[ï]{}ve (short-sighted) [@Choi2009; @Huang2012] and these results, albeit for a subset of the dataset, confirm this indication. The hypothesis that RL is a component of the decision making process for some investors is not confirmed as either version of the tested model (short or far-sighted) is statistically better than chance only for a subset of the players. This subset, within this population, is not large enough to draw a statistically meaningful positive result. By means of a Binomial Proportion Confidence Interval calculated with Clopper-Pearson method we get a negative result for the entire population within a 99$\%$ confidence interval (Fig. \[fig:population\_statistic\] in the Appendix). While this exploratory study gives some perspectives on how Reinforcement Learning can be used to model learning and action-selection for investing problems, future work will focus on different models and risk classification techniques as well as on a deeper investigation of the typical parameters of the best performing players and the correlation of different strategies and performance of stock trading together with a study of different RL models.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank their colleagues of the Sheffield Neuroeconomics interdisciplinary research group for insightful discussion.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
J. [Von Neumann]{} and O. Morgenstern, *[Theory of games and economic behavior]{}*. 1947.
C. Stramer, “[Developments in Non-Expected Utility Theory: The Hunt for a Descriptive Theory of Choice under Risk]{}.” pp. 332–382, 2000.
S. Frederick, G. Loewenstein, and T. O’Donoghue, “[Time Discounting and time preference: a critical review]{}.” *Journal of Economic Literature*, vol. XL, pp. 351–401, 2002.
A. Tversky and D. Kahneman, “[Judgment under uncertainty : heuristics and biases]{}.” *Science*, vol. 185, no. 4157, pp. 1124–1131, 1974.
D. Kahneman, J. L. Knetsch, and R. H. Thaler, “[Anomalies: The Endowment Effect, Loss Aversion, and Status Quo Bias]{}.” *Journal of Economic Perspectives*, vol. 5, no. 1, pp. 193–206, 1991.
U. Hoffrage, A. Weber, R. Hertwig, and V. M. Chase, “[How to Keep Children Safe in Traffic: Find the Daredevils Early]{}.” *Journal of Experimental Psychology: Applied*, vol. 9, no. 4, pp. 249–260, 2003.
T. J. Pleskac, “[Decision making and learning while taking sequential risks]{}.” *Journal of Experimental Psychology: Learning*, vol. 34, no. 1, pp. 167–185, 2008.
T. S. Wallsten, T. J. Pleskac, and C. W. Lejuez, “[Modeling Behavior in a Clinically Diagnostic Sequential Risk-Taking Task.]{}” *Psychological Review,*, vol. 112, no. 4, pp. 862–880, 2005.
K. H. Britten, W. T. Newsome, M. N. Shadlen, S. Celebrini, and J. a. Movshon, “[A relationship between behavioral choice and the visual responses of neurons in macaque MT.]{}” *Visual neuroscience*, vol. 13, no. 1, pp. 87–100, 1996.
K. H. Britten, M. N. Shadlen, W. T. Newsome, and J. a. Movshon, “[The analysis of visual motion: a comparison of neuronal and psychophysical performance]{}.” *The Journal of neuroscience: the official journal of the Society for Neuroscience*, vol. 12, no. 12, pp. 4745–4765, 1992.
J. I. Gold and M. N. Shadlen, “[Neural computations that underlie decisions about sensory stimuli]{}.” *Trends in Cognitive Sciences*, vol. 5, no. 1, pp. 10–16, 2001.
——, “[Banburismus and the brain: Decoding the relationship between sensory stimuli, decisions, and reward]{}.” *Neuron*, vol. 36, no. 2, pp. 299–308, 2002.
——, “[The neural basis of decision making.]{}” *Annual review of neuroscience*, vol. 30, pp. 535–574, 2007.
M. N. Shadlen, K. H. Britten, W. T. Newsome, and J. a. Movshon, “[A computational analysis of the relationship between neuronal and behavioral responses to visual motion]{}.” *J Neurosci*, vol. 16, no. 4, pp. 1486–1510, 1996.
M. N. Shadlen and W. T. Newsome, “[Motion perception: seeing and deciding.]{}” *Proceedings of the National Academy of Sciences of the United States of America*, vol. 93, no. 2, pp. 628–633, 1996.
A. G. Sanfey, G. Loewenstein, S. M. McClure, and J. D. Cohen, “[Neuroeconomics: cross-currents in research on decision-making.]{}” *Trends in cognitive sciences*, vol. 10, no. 3, pp. 108–16, mar 2006.
P. W. Glimcher and A. Rustichini, “[Neuroeconomics: the consilience of brain and decision.]{}” *Science (New York, N.Y.)*, vol. 306, no. 5695, pp. 447–52, oct 2004.
G. Loewenstein, S. Rick, and J. D. Cohen, “[Neuroeconomics.]{}” *Annual review of psychology*, vol. 59, pp. 647–72, jan 2008. \[Online\].
D. J. Barraclough, M. L. Conroy, and D. Lee, “[Prefrontal cortex and decision making in a mixed-strategy game.]{}” *Nature neuroscience*, vol. 7, no. 4, pp. 404–410, 2004.
A. Soltani, D. Lee, and X. J. Wang, “[Neural mechanism for stochastic behaviour during a competitive game]{}.” *Neural Networks*, vol. 19, no. 8, pp. 1075–1090, 2006.
R. S. Sutton and A. G. Barto, *[Reinforcement Learning: An introduction]{}*, [A Bradford Book]{}, Ed.1em plus 0.5em minus 0.4emMIT Press, Cambridge, MA, 1998.
K. Doya, “[Reinforcement learning: Computational theory and biological mechanisms]{}.” *HFSP Journal*, vol. 1, no. 1, p. 30, 2007.
G. Tesauro, “[TD-Gammon, a Self-Teaching Backgammon Program, Achieves Master-Level Play]{}.” *Neural Computation*, vol. 6, no. 2, pp. 215–219, 1994.
A. [Abbeel, P, Coates, A, Morgan, Q, Ng]{}, “[An Application of Reinforcement Learning to Aerobatic Helicopter Flight.]{}” *[Advances in Neural Information Processing Systems 19: Proceedings of the 2006 Conference, 1-9]{}*. 2006.
R. Hafner and M. Riedmiller, “[Neural reinforcement learning controllers for a real robot application]{}.” *Proceedings - IEEE International Conference on Robotics and Automation*, no. April, pp. 2098–2103, 2007.
M. a. Walker, “[An Application of Reinforcement Learning to Dialogue Strategy Selection in a Spoken Dialogue System for Email]{}.” *Journal of Artificial Intelligence Research*, vol. 12, p. 387, 2000.
V. Mnih, K. Kavukcuoglu, D. Silver, A. a. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “[Human-level control through deep reinforcement learning]{}.” *Nature*, vol. 518, no. 7540, pp. 529–533, 2015.
P. Dayan and N. D. Daw, “[Decision theory, reinforcement learning, and the brain.]{}” *Cognitive, affective [&]{} behavioral neuroscience*, vol. 8, no. 4, pp. 429–53, dec 2008.
W. Schultz, P. Dayan, and P. R. Montague, “[A neural substrate of prediction and reward.]{}” *Science (New York, N.Y.)*, vol. 275, no. 5306, pp. 1593–9, mar 1997.
K. Doya, “[Complementary roles of basal ganglia and cerebellum in learning and motor control]{}.” *Current Opinion in Neurobiology*, vol. 10, pp. 732–739, 2000.
N. D. Daw and K. Doya, “[The computational neurobiology of learning and reward]{}.” *Current Opinion in Neurobiology*, vol. 16, no. 2, pp. 199–204, 2006.
C. J. C. H. Watkins,, “[Learning from delayed rewards]{}.” *University of Cambridge* , 1989.
P. Werbos, “[A menu of designs for reinforcement learning over time. In Neural Networks for Control ]{}.” *Neural Networks for Control*, MIT Press, Cambridge, Massachusetts, pp. 67–95, 1990.
O. Hikosaka, K. Nakamura, and H. Nakahara, “[Basal ganglia orient eyes to reward.]{}” *Journal of neurophysiology*, vol. 95, no. 2, pp. 567–584, 2006.
A. [Schultz, W, Romo, R, Ljungberg, T, Mirenowicz, J, Hollerman, JR, and Dickson]{}, “[Reward-related signals carried by dopamine neurons]{}.” in *Models of Information Processing in the Basal Ganglia*, M. Cambridge, Ed., 1995, pp. 233–248.
R. E. Suri and W. Schultz, “[Learning of sequential movements by neural network model with dopamine-like reinforcement signal]{}.” *Experimental Brain Research*, vol. 121, no. 3, pp. 350–354, 1998.
P. Waelti, a. Dickinson, and W. Schultz, “[Dopamine responses comply with basic assumptions of formal learning theory.]{}” *Nature*, vol. 412, no. 6842, pp. 43–48, 2001.
T. Satoh, S. Nakai, T. Sato, and M. Kimura, “[Correlated coding of motivation and outcome of decision by dopamine neurons.]{}” *The Journal of neuroscience : the official journal of the Society for Neuroscience*, vol. 23, no. 30, pp. 9913–9923, 2003.
H. Nakahara, H. Itoh, R. Kawagoe, Y. Takikawa, and O. Hikosaka, “[Dopamine Neurons Can Represent Context-Dependent Prediction Error]{}.” *Neuron*, vol. 41, no. 2, pp. 269–280, 2004.
G. Morris, A. Nevet, D. Arkadir, E. Vaadia, and H. Bergman, “[Midbrain dopamine neurons encode decisions for future action.]{}” *Nature neuroscience*, vol. 9, no. 8, pp. 1057–1063, 2006.
A. Barto, “[Adaptive critics and the basal ganglia]{}.” in *Models of Information Processing in the Basal Ganglia*, M. Cambridge, Ed., 1995, pp. 215–232.
P. R. Montague, P. Dayan, and T. J. Sejnowski, “[A framework for mesencephalic dopamine systems based on predictive Hebbian learning.]{}” *The Journal of neuroscience : the official journal of the Society for Neuroscience*, vol. 16, no. 5, pp. 1936–1947, 1996.
K. Samejima, Y. Ueda, K. Doya, and M. Kimura, “[Representation of action-specific reward values in the striatum.]{}” *Science (New York, N.Y.)*, vol. 310, no. 5752, pp. 1337–1340, 2005.
R. Kawagoe, Y. Takikawa, and O. Hikosaka, “[Expectation of reward modulates cognitive signals in the basal ganglia.]{}” *Nature neuroscience*, vol. 1, no. 5, pp. 411–416, 1998.
——, “[Reward-predicting activity of dopamine and caudate neurons–a possible mechanism of motivational control of saccadic eye movement.]{}” *Journal of neurophysiology*, vol. 91, no. 2, pp. 1013–1024, 2004.
W. Schultz, “[Getting formal with dopamine and reward]{}.” *Neuron*, vol. 36, no. 2, pp. 241–263, 2002.
J. J. Day, M. F. Roitman, R. M. Wightman, and R. M. Carelli, “[Associative learning mediates dynamic shifts in dopamine signaling in the nucleus accumbens.]{}” *Nature neuroscience*, vol. 10, no. 8, pp. 1020–1028, 2007.
S. E. Hyman, R. C. Malenka, and E. J. Nestler, “[Neural mechanisms of addiction: the role of reward-related learning and memory.]{}” *Annual review of neuroscience*, vol. 29, pp. 565–598, 2006.
D. Joel, Y. Niv, and E. Ruppin, “[Actor-critic models of the basal ganglia: new anatomical and computational perspectives]{}.” *Neural Networks*, vol. 15, no. 4-6, pp. 535–547, 2002.
J. R. Wickens, J. C. Horvitz, R. M. Costa, and S. Killcross, “[Dopaminergic mechanisms in actions and habits.]{}” *The Journal of neuroscience : the official journal of the Society for Neuroscience*, vol. 27, no. 31, pp. 8181–8183, 2007.
A. a. Stocker and E. P. Simoncelli, “[Noise characteristics and prior expectations in human visual speed perception.]{}” *Nature neuroscience*, vol. 9, no. 4, pp. 578–585, 2006.
K. K[ö]{}rding, “[Decision theory: what “should” the nervous system do?]{}” *Science (New York, N.Y.)*, vol. 318, no. 5850, pp. 606–610, 2007.
N. D. Daw, J. P. O’Doherty, P. Dayan, B. Seymour, and R. J. Dolan, “[Cortical substrates for exploratory decisions in humans.]{}” *Nature*, vol. 441, no. 7095, pp. 876–879, 2006.
G. Luksys, W. Gerstner, and C. Sandi, “[Stress, genotype and norepinephrine in the prediction of mouse behavior using reinforcement learning.]{}” *Nature neuroscience*, vol. 12, no. 9, pp. 1180–6, sep 2009.
R. Frey and R. Hertwig, “[Sell in May and Go Away ? Learning and Risk Taking in Nonmonotonic Decision Problems]{}.” *Journal of Experimental Psychology*, vol. 41, no. 1, pp. 193–208, 2015.
S. Benartzi and R. H. Thaler, “[Heuristics and Biases in Retirement Savings Behavior]{}.” *Journal of Economic Perspectives*, vol. 21, no. 3, pp. 81–104, 2007.
J. Choi and D. Laibson, “[Reinforcement learning and savings behavior]{}.” *The Journal of Finance*, vol. 64, no. 6, 2009.
X. Huang, “[Industry Investment Experience and Stock Selection]{}.” *Available at SSRN 1786271*, no. November, 2012.
T. Odean, “[Are Investors Reluctant to Realize Their Losses ?]{}” vol. LIII, no. 5, pp. 1775–1798, 1998.
Y. Chen, S. Mabu, K. Hirasawa, and J. Hu, “[Trading rules on stock markets using genetic network programming with sarsa learning]{}.” *Proceedings of the 9th annual conference on Genetic and evolutionary computation GECCO 07*, vol. 12, p. 1503, 2007.
J. Lee, “[Stock price prediction using reinforcement learning]{}.” *Industrial Electronics. Proceedings. ISIE 2001. IEEE International Symposium on*, vol. 1, pp. 690–695, 2001.
J. O, J. Lee, J. Lee, and B. Zhang, “[Adaptive stock trading with dynamic asset allocation using reinforcement learning]{}.” *Information Sciences*, vol. 176, no. 15, pp. 2121–2147, 2006.
J. Moody and M. Saffell, “[Learning to trade via direct reinforcement]{}.” *IEEE Transactions on Neural Networks*, vol. 12, no. 4, pp. 875–889, 2001.
A. V. Rutkauskas and T. Ramanauskas, “[Building an artificial stock market populated by reinforcement‐learning agents]{}.” *Journal of Business Economics and Management*, vol. 10, no. 4, pp. 329–341, 2009.
B. M. Barber and T. Odean, “[The Behavior of Individual Investors]{}.” *Handbook of the Economics of Finance*, vol. 2, PB, pp. 1533–1570, 2013
B. M. Barber, Y. T. Lee, Y. J. Liu, and T. Odean, “[Is the aggregate investor reluctant to realise losses? Evidence from Taiwan]{}.” *European Financial Management*, vol. 13, no. 3, pp. 423–447, 2007.
P. Brown, N. Chappel, R. [Da Silva Rosa]{}, and T. Walter, “[The Reach of the Disposition Effect: Large Sample Evidence Across Investor Classes]{}.” *International Review of Finance*, vol. 6, no. 1-2, p. 43, 2006.
A. Frazzini, “[The disposition effect and underreaction to news]{}.” *Journal of Finance*, vol. 61, no. 4, pp. 2017–2046, 2006.
M. Grinblatt and M. Keloharju, “[What Makes Investors Trade?]{}” *The Journal Of Finance*, vol. 56 (2), no. 2, pp. 549–578, 2001.
C. Heath and M. Lang, “[Psychological Factors and Stock Option Exercise]{}.” *Quarterly Journal of Economics* vol. 114, no. 2, pp. 601–627, 1999.
T. Odean, “[Do Investors Trade Too Much?]{}” *American Economic Review*, vol. 89 (5), pp. 1279-1298, 1998.
Z. Shapira and I. Venezia, “[Patterns of behavior of professionally managed and independent investors]{}”, *Journal of Banking and Finance*, vol. 25, no. 8, pp. 1573–1587, 2001.
Kahneman, D., and Tversky, A. “[Prospect Theory: An Analysis of Decision under Risk.]{}”, *Econometrica: Journal of the Econometric Society*, vol. 47(3), 263–291, 1979.
W. Sharpe, “[Capital Asset Prices: A Theory of Market Equilibrium Under Conditions of Risk]{},” *The Journal of Finance*, vol. XIX, no. 3, pp. 425–442, 1964.
S. Beninga, *[Financial Modeling]{}*, 2000.
N. D. Daw, “[Trial-by-trial data analysis using computational models]{},” in *Decision Making, Affect, and Learning: Attention and Performance XXIII*, 2011
J. P. Huelsenbeck and K. a. Crandall, “[Phylogeny Estimation and Hypothesis Testing Using Maximum Likelihood]{},” *Annual Review of Ecology and Systematics*, vol. 28, no. 1, pp. 437–466, 1997.
R. M. Costa, “[Plastic Corticostriatal Circuits for Action Learning What’s Dopamine Got to Do with It?]{},” *Annals of the New York Academy of Sciences*, pp. 172-191, 2007
Appendix {#App:Appendix}
========
This manuscript is a correction to the article “Modelling Stock-market Investors as Reinforcement Learning Agents” by the same authors, issued in the proceedings of the 2015 IEEE Conference on Evolving and Adaptive Intelligent Systems. The corrections include fixing a bug in the script which estimated the probabilities used in the calculation of model fitness. In the previous work we applied some constraints and used a different measure: the number of transactions considered for each player was capped at 25 and the measure of risk used to rank the stocks and classify them into discrete categories was defined by the authors as:
$$\label{eq:riskiness_equation}
R(s_{j}) = \left\lvert {\beta_F}(s_{j}) \cdot \frac{\sigma(s_{j})}{\max \sigma(s)} \right\rvert$$
where $\beta_F$ is the financial modelling measure of volatility of a security used in the present work, $\sigma(s_{j})$ is the standard deviation of the j-th stock and the $\max \sigma(s)$ is the largest standard deviation in the stock pool. This measure of riskiness was used as it was believed to take into account the graphical interpretation of the fluctuations in time series ($\sigma$) and the overall trend of the security compared to the market ($\beta_F$).
![Population model fitting statistic using Binomial Proportion Confidence Interval calculated with Clopper-Pearson method. The results are negative as only 7 out of 46 players are better than random. The errorbar represent the 99$\%$ confidence interval for the myopic model to be correctly fitting the players in the dataset. As the errorbar lies entirely below chance threshold we conclude that the models investigated do not correctly represent the behaviour of the players in the data.[]{data-label="fig:population_statistic"}](population_statistic.png){width="3.5in"}
![Updated model comparison of Fig 1 (a) and (b) in the original paper. This figure shows the comparison of the full RL model (grey bars) against the random model (+ signs) according to their MLE. This is the same comparison as in \[fig:models\_comparisons\] (c) but for the original paper configuration: players transactions capped at 25 and the combination of $\beta_F$ and $\sigma$, as in Eq. \[eq:riskiness\_equation\], used to determine stocks risk degree and their classification. In this case the portion of players captured by the full RL model is only 8$\%$ (4 players, 14, 18, 24, 26). Only 2 players are captured by the myopic model (14 and 18).[]{data-label="fig:RL_v_random_original_updated"}](RL_vs_Random_previous.png){width="3.5in"}
![Updated model comparison of Fig 1 (c) and (d) in the original paper. This figure shows the comparison of the full RL model (grey bars) against the myopic model (filled diamonds signs) according to their MLE. This is the same comparison as in \[fig:models\_comparisons\] (b) but for the original paper configuration: players transactions capped at 25 and the combination of $\beta_F$ and $\sigma$, as in Eq. \[eq:riskiness\_equation\], used to determine stocks risk degree and their classification. The 4 players who are significantly better fitted by a full model are 14, 21, 24, 26.[]{data-label="fig:RL_v_nogamma_original_updated"}](RL_vs_NoGamma_previous.png){width="3.5in"}
[^1]: http://www.virtualtrader.co.uk - Copyright IEX Media Group BV
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We illustrate a general method for calculating spectral statistics that combines the universal (Random Matrix Theory limit) and the non-universal (trace-formula-related) contributions by giving a heuristic derivation of the three-point correlation function for the zeros of the Riemann zeta function. The main idea is to construct a generalized Hermitian random matrix ensemble whose mean eigenvalue density coincides with a large but finite portion of the actual density of the spectrum or the Riemann zeros. Averaging the random matrix result over remaining oscillatory terms related, in the case of the zeta function, to small primes leads to a formula for the three-point correlation function that is in agreement with results from other heuristic methods. This provides support for these different methods. The advantage of the approach we set out here is that it incorporates the determinental structure of the Random Matrix limit.'
author:
- 'E. Bogomolny$^{1}$, J.P. Keating$^2$'
title: 'A method for calculating spectral statistics based on random-matrix universality with an application to the three-point correlations of the Riemann zeros'
---
Introduction
============
Random matrix theory (RMT) was introduced in the 1950s to describe the highly-excited energy levels of heavy nuclei [@wigner_1]-[@bohigas]. It has subsequently found numerous and important applications in many different branches of physics. A review of various research directions in RMT is given, for example, in [@nina; @RMTbook]. Of all these directions, one is particularly unexpected: the use of RMT in number theory. Though statistical methods have many applications in number theory (see e.g. [@kac]) the importance of RMT has attracted particular attention in recent years. This line of investigation started with a theorem of Montgomery [@montgomery] relating to the pair correlation of the non-trivial zeros of the Riemann zeta function, and Dyson’s remark that Montgomery’s formula can be interpreted as saying that statistical properties of the zeros are the same as those of eigenvalues of large hermitian matrices with independent random elements. Odlyzko’s extensive numerical computations [@Odlyzko] provided compelling support for this conjecture, but also drew attention to the slow approach to the limit where RMT is expected to hold. In the case of pair correlation, the approach to the limit was shown to be arithmetical in origin by Berry [@Berry] and a precise formula was derived in [@bogomolny_keating] which matches the numerical data extremely well [@SIAM]. An analogous formula for the zeros of Dirichlet $L$-functions was recently derived in [@bogomolny_keating2].
Further developments in this direction came from a conjecture that the moments of the Riemann zeta function (as well other $L$-functions) can be calculated using the moments of characteristic polynomials in RMT [@keating; @keating2]. Today there exists a large collection of conjectures which predict (in a good agreement with existing numerics) mean values of many different quantities related to number-theoretic zeta and $L$-functions using ideas from RMT. For a review of the background to this area, see [@KS; @KS2].
Unfortunately, only a limited number of rigorous results have been obtained in this field. Most conjectures are based on heuristic arguments which are very difficult (if not impossible) to justify mathematically. As different heuristics stress different points, it is of interest and importance to compare different methods of calculation. It is also often the case that these heuristic methods lead to new insights into similar problems concerning the spectral statistics of quantum chaotic systems (see, e.g., [@bogomolny_keating; @KM]).
The purpose of this paper is to calculate a formula, like that derived in [@bogomolny_keating], for the three-point correlation function for the Riemann zeros using a method proposed in [@varenna]. This function has already been obtained in [@triple] by Conrey and Snaith using the ratio conjecture [@ratio], which follows from heuristic manipulations of the approximate functional equation of the zeta function developed in [@CFKRS]. Our calculation was carried out independently of, and at the same time as [@triple] but we hesitated to publish the result mainly because both methods are general, permit to calculate, in principle, all correlation functions, and should lead to the same formulae.
We decided to present our calculations now because there has recently been renewed interest in these kinds of formulae (see, e.g. [@Zeev; @CS]) and we believe that the approach we take here sheds significant new light on their structure. We emphasize that this approach is completely different to the one involving the ratios conjecture. It is based on the idea of exploiting random matrix universality, and it incorporates universal and non-universal (arithmetic) components in a novel way from the outset. In particular, it combines the determinental structure of random matrix theory with the arithmetical terms in a way that appears more natural than in other approaches. We see this as a significant advantage, because it is often a major difficulty to identify the combinatorial identities underpinning this structure [@nonlinearity; @RudSar; @Zeev; @CS]. It is also worth remarking that the formulae that emerge from this kind of approach are useful for applications (cf. [@nearest; @DHKMS]). And finally, there has recently been considerable focus on random matrix universality (see, e.g., [@RMTbook]), and our approach is likely to be of interest in that context too.
The plan of the paper is the following. Section \[rmt\_universality\] is devoted to a short discussion of random matrix universality. It is well known that standard random matrix ensembles with different one-body potentials lead to different (non-universal) densities of eigenvalues. Nevertheless, it is widely accepted that after unfolding, local statistical properties of eigenvalues for all ’reasonable’ ensembles are the same. This universality leads to an explicit expression for the random matrix kernel, conjecturally valid for any mean density. In Section \[riemann\] an expression for the density of Riemann zeros calculated by taking into account a large but finite number of prime numbers is presented. The primes entering this formula are chosen in such a way that they can be considered as independent under an average over a large window of heights on the critical axis. Inserting the finite expression for the Riemann zeros density into the random matrix kernel gives us formal correlation functions of Riemann zeros. But the result inevitably has oscillations related with short primes. Averaging such oscillations over a large window leads to our main conjecture for correlation functions presented in Section \[conjecture\]. In Section \[two\_point\] it is demonstrated how it is possible to derive from this conjecture the two-point correlation function of Riemann zeros. Section \[three\_point\] contains an explicit calculation of the tree-point correlation function. The final expression agrees with the result of [@triple] obtained by a completely different method. Section \[summary\] is a brief summary of important formulae.
Random matrix universality {#rmt_universality}
==========================
It is well known (see e.g. [@mehta]) that the standard Gaussian Unitary Ensemble of $N\times N$ random matrices (GUE) is determined as an ensemble of Hermitian matrices ($M_{mn}=M_{mn}^{\dag}$) whose elements are random variables with the joint probability $$P(M_{mn})=C_N \mathrm{e}^{-\mathrm{Tr}\, M^2}\prod_{j=1}^N \mathrm{d}M_{nn}
\prod_{1\leq m<n\leq N}\mathrm{d} \mathrm{Re}\,M_{mn}\ \mathrm{d} \mathrm{Im}\,M_{mn}
\label{gaussian}$$ where $C_N$ is a normalization constant.
A natural generalization of this ensemble consists in choosing instead of $\mathrm{Tr}\, M^2$ in the exponent an ’arbitrary’ function $\mathrm{Tr}\, f(M)$ (called often the one-body potential) $$P(M_{mn})=C_N \mathrm{e}^{-\mathrm{Tr}\,f( M)}\prod_{j=1}^N \mathrm{d}M_{nn}
\prod_{1\leq m<n\leq N}\mathrm{d} \mathrm{Re}\,M_{mn}\ \mathrm{d} \mathrm{Im}\,M_{mn}\ .
\label{ensemble_f}$$ All such ensembles permit one to integrate over angle-type variables and get the joint probability density for the eigenvalues $\lambda_j$ of the matrices $M_{mn}$ [@mehta] $$P(\lambda_1,\lambda, \ldots, \lambda_N)=C^{\prime}_N \prod_{1\leq i<j\leq N}(\lambda_j-\lambda_i)^2 \exp \left (-\sum_{k=1}^Nf(\lambda_k) \right )\ .$$ To calculate the $n$-point correlation function one has to fix $n$ eigenvalues $x_j=\lambda_j$ with $j=1,\ldots, n$ and to integrate over the remaining $N-n$ variables. For the ensemble considered, this can be done by the method of orthogonal polynomials [@mehta]. One first introduces polynomials, $p_k(x)$ orthogonal with respect to the measure $\exp \left (-f(x) \right )$ $$\int \mathrm{e}^{-f(x) }p_k(x)p_r(x)\mathrm{d}x=\delta_{kr}\ .$$ Then the $n$-point correlation function for the ensemble takes the form of the $n\times n$ determinant [@mehta] $$R(x_1,\ldots,x_n)=\det( K_N(x,y) )_{x,y=x_1,\ldots, x_n}
\label{n_corr_functions}$$ where the kernel $K_N(x,y)$ is expressed through the orthogonal polynomials as follow $$K_N(x,y)=\sum_{j=0}^{N-1} p_j(x)p_j(y)\mathrm{e}^{-f(x)/2-f(y)/2}\ .$$ In particular, the mean level density, $\bar{\rho}(x)$, (i.e. one-point correlation function) is $$\bar{\rho}(x)=K_N(x,x)\ .
\label{density}$$ Usually one is interested in the limit $N\to\infty$ and the main question is what is then the limiting behaviour of this kernel. For the Gaussian ensemble the answer is well known [@mehta] since the orthogonal polynomials in this case are the usual Hermite polynomials $$p_j(x)=\frac{1}{\sqrt{2^jj!\sqrt{\pi}}}H_j(x)\ .$$ The mean level density in this case is given by the famous “semicircle law”, $$\bar{\rho}(x)=\left \{ \begin{array}{cc}\dfrac{1}{\pi}\sqrt{2N-x^2},& |x|<\sqrt{2N}\\0,&|x|>\sqrt{2N}\end{array}\right . \ ,$$ and there exists an explicit formula for the kernel $K_N(x,y)$ when $N$ is large. This formula takes an especially simple form in the bulk of the spectrum when $|x|,|y|\ll \sqrt{N}$ $$K_N(x,y)=\frac{\sin \pi \bar{\rho}(0)(x-y)}{\pi (x-y)}
\label{kernel_gue}$$ where $\bar{\rho}(0)=\sqrt{2N}/\pi$ is the level density at small $x$.
For a general one-body potential $f(x)$ the situation is more difficult. The mean level density can be calculated for large $N$ from the Dyson equation [@mehta] $$\mathrm{P}\int_{-\infty}^{\infty}\frac{\bar{\rho}(z)}{x-z}\mathrm{d}z=\tfrac{1}{2}f^{\prime}(x)
\label{dyson}$$ where P indicates the principal value integral, but correlation functions are more difficult to obtain though there exists a vast literature on this subject (see e.g. [@orthogonal] and references therein).
Instead of using rigorous asymptotic formulae, we shall argue as follows. It is well known that the mean level density is not a universal quantity. Different one-body potentials, $f(x)$, lead to different densities (cf. ). On the other hand, it is widely accepted that after unfolding all correlation functions in the local scale should be universal. The unfolding here means that one calculates statistical properties not of the true levels, $\lambda_n$, (which, in general, have a non-universal mean density $\bar{\rho}(x)$) but of new quantities $$e_n=\bar{N}(\lambda_n)$$ where $\bar{N}(x)$ is the mean number of levels with $\lambda_j<x$, $\bar{N}(x)=\int^x \bar{\rho}(y) \mathrm{d}y$.
By construction, the new levels, $e_n$, have unit mean density, and random matrix universality asserts that for these quantities the kernel has exactly the same form as in with $\bar{\rho}(0)=1$ $$K(x,y)=\frac{\sin \pi (\bar{N}(x)-\bar{N}(y) )}{\pi (x-y)}\ .
\label{k_general}$$ This formula is assumed to be valid when (i) points $x$ and $y$ are far from the ends of the spectrum, and (ii) the mean number of levels $\bar{N}(x)$ is a smooth function of $x$ i.e. it changes slowly in the scale of the nearest levels.
This expression is the concise manifestation of random matrix universality but we are not aware that it has been proved in full generality. Nevertheless, it agrees with all we know (or conjecture) about universal behaviour of random matrix ensembles (at least for GUE types) and we shall apply it below for the Riemann zeta function.
Riemann zeta function {#riemann}
=====================
The Riemann zeta function is defined when Re$s>1$ as the sum over all integers (see e.g. [@riemann]) $$\zeta(s)=\sum_{n=1}^{\infty}\frac{1}{n^s}$$ or as the Euler product over prime numbers $$\zeta(s)=\prod_{p}\left (1-\frac{1}{p^s}\right )^{-1}\ .
\label{euler}$$ It has an analytic continuation to the rest of the complex $s$-plane, except for a pole at $s=1$. The celebrated Riemann Hypothesis states that all non-trivial zeros of this function, $\zeta(s_j)=0$, have the form $$s_j=\tfrac{1}{2}+\mathrm{i}E_j$$ with real $E_j$ (which may be thought of as analogous to quantum energies).
Assuming the Riemann Hypothesis, it is easy to write down a formal expression for the density of these zeros (called in the mathematical literature Weil’s explicit formula [@weil]). Indeed if one writes $\zeta(1/2+\mathrm{i}E)\sim \prod_j(E-E_j)$ then the density of $E_j$ is $$d(E)=-\frac{1}{\pi }\mathrm{Im}\,\frac{\partial }{\partial E}\ln \zeta \left (\tfrac{1}{2}+\mathrm{i}(E+\mathrm{i}\varepsilon)\right )_{\varepsilon\to +0}\ .$$ Using the functional relation for the zeta function and one gets that, as usual, the density of zeros is a sum of two terms $$d(E)=\overline{d(E)} +d^{(\mathrm{osc})}(E) \ ,$$ where $ \overline{d(E)}$ is the smooth part of the density, which, as $E\rightarrow\infty$ is given by $$\overline{d(E)}\approx\frac{1}{2\pi}\ln \frac{E}{2\pi} \ ,$$ and $d^{(\mathrm{osc})}(E)$ is the oscillating part of the density $$d^{(\mathrm{osc})}(E)=-\frac{1}{2\pi}\sum_{p}\sum_{n=1}^{\infty}
\frac{\ln p}{ p^{n/2}}\left (\mathrm{e}^{\mathrm{i} n E\ln p}+\mathrm{e}^{-\mathrm{i} n E\ln p}\right )\ .
\label{osc}$$ Of course, the sum over all primes $p$ diverges at real $E$ and this expression has no (clear) mathematical meaning (similar to all “physical” trace formulae). It gains such a meaning when integrated against a sufficiently smooth test function.
Nevertheless, it is legitimate to write the ’true’ density as a finite sum over primes with $p<p^*$ and an unknown remainder related with large primes satisfying $p>p^*$ $$d(E)=\bar{\rho}(E)+\mathrm{large\;primes},\qquad \bar{\rho}(E)=\overline{d(E)} +\widetilde{d(E,p^*)}
\label{full_density}$$ where $\widetilde{d(E,p^*)}$ is the same sum as in but taken over a finite set of primes with $p<p^*$ (the value of $p^*$ will be chosen below) $$\widetilde{d(E,p*)}=-\frac{1}{2\pi}\sum_{p<p*}\sum_{n=1}^{\infty}
\frac{\ln p}{ p^{n/2}}\left (\mathrm{e}^{\mathrm{i} n E\ln p}+\mathrm{e}^{-\mathrm{i} n E\ln p}\right )=
\frac{1}{2\pi \mathrm{i}}\frac{\partial}{\partial E}\sum_{p<p^*}
\ln\frac{1-A_p\mathrm{e}^{\mathrm{i}\Phi_p(E)}}{1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(E)}}\ .
\label{d}$$ Here for further convenience we introduce the notation $$A_p=\frac{1}{\sqrt{p}},\qquad \Phi_p(E)=E\ln p \ .$$ The knowledge of $\bar{\rho}(E)$ permits easily to calculate the mean number of levels corresponding to this density $$\bar{N}(x,p^*)\equiv \int^x_0 \bar{\rho}(E)\mathrm{d}E= \int^x_0 ( \overline{d(E)}+\widetilde{d(E,p^*)} )\mathrm{d}E\ .$$ It is plain that $$\mathrm{e}^{2\pi \mathrm{i} \bar{N}(E,p^*)}=\mathrm{e}^{2\pi \mathrm{i} \overline{N(E)}}\prod_{p<p^*}
\frac{1-A_p \mathrm{e}^{\mathrm{i}\Phi_p(E)} }{1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(E)} }
\label{N}$$ where $$\overline{N(E)}=\frac{E}{2\pi}\ln\frac{E}{2\pi \mathrm{e}}+\mathrm{const}\ .$$ For the Riemann zeta function the constant is known ($7/8$) but is irrelevant for our purpose.
Main conjecture {#conjecture}
===============
The principal point in the approach to statistical properties of Riemann zeros advocated here consists in the assumption that the large primes indicated in give rise to GUE correlations with random matrix kernel where $\bar{N}(E,p*)$ is determined by small primes . Precisely, $$K(E_i,E_j)=\frac{\sin(\pi (\bar{N}(E_i,p*)-\bar{N}(E_j,p*)))}{\pi (E_i-E_j)}\ .
\label{main_kernel}$$ Of course, in such an approach the exact mechanism by which large primes conspire to give this kernel is completely ignored. But as we shall show below this assumption permits us to calculate all low order terms for correlation functions of Riemann zeros in agreement with ones calculated by different mathods.
When $n$-point correlation functions are calculated from using the kernel , the result necessarily has oscillations related with oscillations in the “mean” density of zeros produced by short primes.
Usually one is looking for statistical properties of a set of zeros close to a large value of $E$. In this case it is natural to write $E_j=E+e_j$ and then to average around $E$. It means that we propose to calculate correlation functions of the Riemann zeros from the following expression $$R_n(e_1,e_2,\ldots, e_n)=\left \langle \left \langle \det \Big ( K(E+e_i,E+e_j )\Big )_{i,j=1,\ldots,n}\right \rangle \right \rangle_{\Delta E} \ .
\label{R_n_definition}$$ Here the average indicated by $\langle \langle \ldots \rangle \rangle_{\Delta E}$ is to be carried out over a large window of heights $E$ $$\langle \langle F(E) \rangle \rangle_{\Delta E} \equiv \frac{1}{\Delta E}\int_{E-\Delta E/2}^{E+\Delta E/2} F(E^{\prime})\mathrm{d}E^{\prime} \ .$$ Let us choose the cut-off prime, $p^*$, and the window, $\Delta E$, to fulfil the inequalities $$1\ll p^*\ll\Delta E\ll E\ .
\label{inequalities}$$ This choice permits one, at least formally, to calculate all necessary mean values. In particular, one has $$\langle \langle \mathrm{e}^{\mathrm{i} n E\ln p } \rangle \rangle_{\Delta E} =0,\qquad \mathrm{for}\; p<p^*,\qquad n\in Z^*\ ,$$ and $$\langle \langle \mathrm{e}^{ \mathrm{i} E(n_1\ln p_1 -n_2\ln p_2)} \rangle \rangle_{\Delta E} =\delta_{n_1,n_2} \delta_{p_1,p_2}\qquad \mathrm{for}\; p_1,p_2<p^*,\qquad n_1,n_2\in Z^* \ .
\label{orthogonality}$$ Therefore $$\langle \langle \tilde{d}(E,p^*) \rangle \rangle_{\Delta E} =0,\qquad \langle \langle \overline{d(E)} \rangle \rangle_{\Delta E} = \overline{d(E)}\ .
\label{rho_average}$$ More generally, phases $E\ln p$ associated with different primes $p<p^*$ can be considered as independent random phases and the procedure of averaging a quasi-periodic function of these phases is reduced to the integration over them $$\left \langle \left \langle F\Big (\mathrm{e}^{\mathrm{i}E\ln p_1},\ldots, \mathrm{e}^{\mathrm{i}E\ln p_n} \Big ) \right \rangle \right\rangle_{\Delta E}
=\int_0^{2\pi}\frac{\mathrm{d}\phi_1}{2\pi}\cdots \int_0^{2\pi}\frac{\mathrm{d}\phi_n}{2\pi}
F\Big (\mathrm{e}^{\mathrm{i}\phi_1},\ldots, \mathrm{e}^{\mathrm{i}\phi_n} \Big ) \ .
\label{good_average}$$ Any averaging procedure for which this relation is fulfilled is suitable for our purposes and it can serve as the definition of ’good’ averaging.
Eq. together with and are our main formulae for correlation functions of zeros of the Riemann zeta function. In the next Sections we show how such formula can be used to calculate explicitly the two and three-point correlation functions. When performing the calculations we shall see that, after averaging, the remaining terms can be divided into two groups. The first contains various sums over primes such that they have well defined values in the formal limit $p^*\to\infty$. The second, which includes divergent contributions, can be transformed to the formally divergent product (with imaginary $s$) $$f(s,p^*)= \prod_{p<p^*}\dfrac{1-p^{-1}}{1-p^{-1-s}}\ .$$ Under the assumption that $$1 \ll \ln(p^*)\ll 1/|s|\;.
\label{ll}$$ This can be done as follows $$\begin{aligned}
& &f(s,p^*)=\lim_{t\to 0}\prod_{p<p^*}\frac{1-1/p^{1+t}}{1-1/p^{1+s}}
=\lim_{t\to 0}\frac{\zeta(1+s)}{\zeta(1+t)}
\prod_{p>p^*}\frac{1-1/p^{1+s}}{1-1/p^{1+t}}\nonumber \\
&\approx &\lim_{t\to 0}\frac{\zeta(1+s)}{\zeta(1+t)}
\exp (\int_{\ln(p^*)}^{\infty}\frac{\mathrm{d}u}{u}(\mathrm{e}^{-tu}-\mathrm{e}^{-su}))
=\lim_{t\to 0}\frac{\zeta(1+s)}{\zeta(1+t)}\exp \ln(s/t)=s\zeta(1+s)\ .\end{aligned}$$ In the last step we use that according to our assumption $s\ln (p^*)\ll 1$ (with, of course, $t\ln(p^*)\ll 1$), and $$\int_{0}^{\infty}\frac{\mathrm{d} u}{u}(\mathrm{e}^{-tu}-\mathrm{e}^{-su})=\ln s -\ln t\ .$$ (A more careful derivation can be given by using Eq. 3.14.1 of [@riemann].)
This leads to the conclusion that under $$\prod_{p<p^*}\frac{1-p^{-1}}{1-p^{-1-s}}\underset{p^*\to\infty}{\longrightarrow} s\zeta(1+s)
\label{zeta}$$ and we shall use this expression throughout the paper.
Two-point correlation function of Riemann zeros {#two_point}
===============================================
The simplest non-trivial example of a correlation function of Riemann zeros is the two-point correlation function. It was calculated in [@bogomolny_keating] by using the explicit form of the Hardy-Littlewood conjecture concerning the distribution of prime pairs. See [@bogomolny_keating2] for an extension to Dirichlet $L$-functions. A formula identical to that obtained in [@bogomolny_keating] was shown also to follow from the ratios conjecture [@ratiosCS]. Here we show that exactly the same result is obtained form Eq. based on completely different assumptions.
From one gets $$\begin{aligned}
R_2(e_1,e_2)&=& \left \langle \left \langle \det \left ( \begin{array}{cc} K(E+e_1,E+e_1 ) & K(E+e_1,E+e_2 \\ K(E+e_2,E+e_1 )& K(E+e_2,E+e_2) \end{array} \right )\right \rangle \right \rangle_{\Delta E} \nonumber\\
& =& \langle \langle \bar{\rho}(E+e_1)\bar{\rho}(E+e_2)\rangle \rangle_{\Delta E} -
\langle \langle K^2(E+e_1,E+e_2) \rangle \rangle_{\Delta E} \ .\end{aligned}$$ Here we use $K(E,E)=\bar{\rho}(E)$ and $K(E_1,E_2)=K(E_2,E_1)$. It is worth remarking that in this approach we start with a determinant - in other approaches the main difficultly lies in identifying the combinatorial identities that match with a determinental form. Noting that $$\begin{aligned}
K(E_1,E_2)&=&\frac{\sin^2(\pi (\bar{N}(E_1,p*)-\bar{N}(E_2,p*)))}{\pi^2 (E_1-E_2)^2}\\
&=&-\frac{1}{4\pi^2(E_1-E_2)^2}\left (\mathrm{e}^{2\pi \mathrm{i} (\bar{N}(E_1,p^*)-\bar{N}(E_2,p^*)}-2+ \mathrm{e}^{-2\pi \mathrm{i} (\bar{N}(E_1,p^*)-\bar{N}(E_2,p^*)} \right ) \nonumber\end{aligned}$$ and assuming the validity of , one gets $$R_{2}(e_1,e_2)=\overline{d(E)}^{\,2}+R_{2}^{\mbox{c}}(e_{1},e_2)\ .$$ $R_{2}^{\mbox{c}}(e_{1},e_2)$ is the connected part of the two-point correlation function equals to the sum of two terms, the smooth term, $R_{2}^{\mbox{diag}}(e_1, e_2)$, and the oscillatory term, $R_{2}^{\mbox{osc}}(e_1,e_2)$, $$\begin{aligned}
R_2^{\mbox{c}}(e_1,e_2)&\equiv & \langle \langle K^2(E+e_1,E+e_2) \rangle \rangle_{\Delta E} =
R_{2}^{\mbox{diag}}(e_1, e_2)+R_{2}^{\mbox{osc}}(e_1,e_2) \ ,
\label{rtwo}\end{aligned}$$ where $$R_2^{\mbox{diag}}(e_1,e_2)= \langle \langle d_1\ d_2 \rangle \rangle_{\Delta E} -\frac{1}{2\pi^2\epsilon^2}
\label{R_diag}$$ and $$R_{2}^{\mbox{osc}}(e_1,e_2)=\frac{1}{4\pi^2 \epsilon^2} \langle \langle \mathrm{e}^{2\pi \mathrm{i} (N_1-N_2)}+\mathrm{e}^{-2\pi \mathrm{i} (N_1-N_2)}\rangle \rangle_{\Delta E}\ .
\label{R_osc}$$ Here and below we use the following notations: $ d_j=\widetilde{d(E+e_j, p*)}$, $N_j=\bar{N}(E+e_j,p*)$, and $\epsilon=e_1-e_2$.
Smooth terms
------------
The calculation of $\langle \langle d_1\ d_2 \rangle \rangle_{\Delta E}$ is straightforward and corresponds to the so-called diagonal approximation. From one has $$\begin{aligned}
&4\pi^2 & \langle \langle d_1\ d_2 \rangle \rangle_{\Delta E} =\sum_{p_{1,2}<p*}\sum_{n_{1,2}=1}^{\infty}
\frac{\ln p_1 \ln p_2}{p_1^{n_1/2}p_2^{n_2/2}} \Big \langle \Big \langle \left (\mathrm{e}^{\mathrm{i} n_1 (E+e_1)\ln p_1}+\mathrm{e}^{-\mathrm{i} n_1 (E+e_1)\ln p_1}\right ) \nonumber \\
&\times & \left (\mathrm{e}^{\mathrm{i} n_2 (E+e_2)\ln p_2}+\mathrm{e}^{-\mathrm{i} n_2 (E+e_2)\ln p_2}\right )\Big \rangle \Big \rangle_{\Delta E}
\nonumber\\
&=&\sum_{p<p*}\sum_{n=1}^{\infty}\frac{\ln^2 p}{p^{n} }
\Big (\mathrm{e}^{\mathrm{i} n \epsilon \ln p}+\mathrm{e}^{-\mathrm{i} n \epsilon\ln p}\Big )
= - \frac{\partial^2}{\partial \epsilon^2}\sum_{p<p*}\sum_{n=1}^{\infty}\frac{1}{n^2 p^{n} }
\Big(\mathrm{e}^{\mathrm{i} n \epsilon \ln p}+\mathrm{e}^{-\mathrm{i} n \epsilon\ln p}\Big )\nonumber\\
&=& - \frac{\partial^2}{\partial \epsilon^2}\sum_{p<p*}\sum_{n=2}^{\infty}\frac{1}{n^2 p^{n} }
\left (\mathrm{e}^{\mathrm{i} n \epsilon\ln p}+\mathrm{e}^{-\mathrm{i} n \epsilon\ln p}\right )
- \frac{\partial^2}{\partial \epsilon^2}\sum_{p<p*}\frac{1}{ p }
\left (\mathrm{e}^{\mathrm{i} \epsilon \ln p}+\mathrm{e}^{-\mathrm{i} \epsilon\ln p}\right )
\label{d_1_d_2}\end{aligned}$$ When $p^*\to\infty$ only the last term diverges. To calculate this it is convenient to use Eq. . By taking the logarithm of the both parts of this relation and of its complex conjugate one obtains $$\sum_{p<p^*}\sum_{n=1}^{\infty}\frac{1}{n p^n}\left (\mathrm{e}^{\mathrm{i} n s \ln p}+\mathrm{e}^{-\mathrm{i}n s\ln p}\right )=
2\ln s +\ln |\zeta(1+\mathrm{i}s)|^2 + C$$ with a constant $C=-2\ln \prod_{p<p^*}(1-1/p)$. Consequently, $$\sum_{p<p*}\frac{1}{ p }
\left (\mathrm{e}^{\mathrm{i} s\ln p}+\mathrm{e}^{-\mathrm{i} s\ln p}\right )=2\ln s +\ln |\zeta(1+\mathrm{i}s)|^2 -\sum_{p<p^*} \sum_{n=2}^{\infty}\frac{1}{n p^n}\left (\mathrm{e}^{\mathrm{i} n s \ln p}+\mathrm{e}^{-\mathrm{i}n s\ln p}\right )+ C\ .$$ Substituting this relation (with $s=\epsilon$) into and taking into account the fact that the sums with $n\geq 2$ converge and $C$ is independent on $\epsilon$, we conclude that $$\langle \langle d_1\ d_2 \rangle \rangle_{\Delta E}=\frac{1}{2\pi^2 \epsilon^2}-\frac{1}{4\pi^2} \frac{\partial^2}{\partial \epsilon^2}\ln |\zeta(1+\mathrm{i} \epsilon )|^2
-\frac{1}{4\pi^2} \frac{\partial^2}{\partial \epsilon^2}\sum_{p<p*}\sum_{n=2}^{\infty}\frac{1-n}{n^2 p^{n} }
\left (\mathrm{e}^{\mathrm{i} n \epsilon\ln p}+\mathrm{e}^{-\mathrm{i} n\epsilon \ln p}\right )\ .$$ Combining this expression and Eq. , and using $\sum_{n=2}^{\infty}(1-n)x^n=-x^2/(1-x)^2$ gives $$R_{2}^{\mbox{diag}}(\epsilon)=-\frac{1}{4\pi^2}\frac{\partial^2}{\partial \epsilon^2}
\ln |\zeta(1+\mathrm{i}\epsilon)|^2-
\frac{1}{4\pi^2}\sum_p \ln^2 p\left( \frac{1}{(p^{1+\mathrm{i}\epsilon }-1)^2}+
\frac{1}{(p^{1-\mathrm{i} \epsilon}-1)^2} \right )\ .
\label{r_2diag}$$
Oscillatory terms
-----------------
The next step consists in calculating the oscillatory part of the two-point correlation function given by . Substituting Eq. into , one gets $$\left \langle \left \langle \mathrm{e}^{2\pi \mathrm{i} (N_1-N_2)} \right \rangle \right \rangle_{\Delta E} =
\mathrm{e}^{2\pi \mathrm{i} \overline{d(E)}(e_1-e_2)} \langle \langle R_p(E;e_1,e_2) \rangle \rangle_{\Delta E}
\label{mean_exp_N}$$ where $$R_p(E;e_1,e_2)= \prod_{p<p^*} \frac{[1-A_p \mathrm{e}^{-\mathrm{i}(\Phi_p(E)+e_2\ln p )}][1-A_p \mathrm{e}^{\mathrm{i}(\Phi_p(E)+e_1\ln p)}] }
{[1-A_p\mathrm{e}^{-\mathrm{i}(\Phi_p(E)+e_1\ln p)}][1-A_p \mathrm{e}^{\mathrm{i}(\Phi_p(E)+e_2\ln p)}] }\ .$$ The averaging of $R_p(E,e_1,e_2)$ over $E$ can be done by using Eq. \[good\_average\]. Therefore, the average over $E$ corresponds to the independent integration over phases $\Phi_p(E)=E\ln p_j$ $$\langle \langle R_p(E;e_1,e_2) \rangle \rangle_{\Delta E} =\prod_{p<p^*} \langle \mathrm{R}_p (\Phi_p; e_1,e_2) \rangle_{\Phi_p}$$ where the average $\langle \mathrm{R}_p (\Phi_p; e_1,e_2) \rangle_{\Phi_p}$ is simply the mean value over all $\Phi_p$ $$\langle \mathrm{R}_p (\Phi_p; e_1,e_2) \rangle_{\Phi_p}=\frac{1}{2\pi}\int_0^{2\pi}\mathrm{R}_p (\Phi_p; e_1,e_2)\mathrm{d}\Phi_p
\ ,
\label{phi_average}$$ and $$\mathrm{R}_p (\Phi_p; e_1,e_2)=
\frac{[1-A_p \mathrm{e}^{-\mathrm{i}(\Phi_p +e_2\ln p)}][1-A_p \mathrm{e}^{\mathrm{i}(\Phi_p +e_1\ln p)} ] }
{[1-A_p\mathrm{e}^{-\mathrm{i}(\Phi_p +e_1\ln p)}][1-A_p \mathrm{e}^{\mathrm{i}(\Phi_p +e_2\ln p)}] }\ .
\label{R_p}$$ The calculation of the integral can conveniently be performed by complex integration. Putting $z=\mathrm{e}^{\mathrm{i}\Phi_p}$, one gets $$\langle \mathrm{R}_p (\Phi_p;e_1,e_2) \rangle_{\Phi_p} =
\frac{1}{2\pi} \oint \frac{[1-A_p z^{-1}\mathrm{e}^{-\mathrm{i}e_2\ln p }][1-A_p z \mathrm{e}^{\mathrm{i}e_1\ln p}] }
{[1-A_p z^{-1} \mathrm{e}^{-\mathrm{i}e_1\ln p}][1-A_p z \mathrm{e}^{\mathrm{i}e_2\ln p}] }\frac{\mathrm{d}z}{z}$$ where the integral is taken over the unit circle in the complex plane.
As $A_p=p^{-1}<1$, inside the integration contour there are two poles. The first is at $z=0$ and the second at $z=A_p \mathrm{e}^{-\mathrm{i}e_1\ln p}$. Straightforward calculation gives $$\begin{aligned}
\langle \mathrm{R}_p (\Phi_p;e_1,e_2) \rangle_{\Phi_p}&=&\mathrm{e}^{\mathrm{i}\epsilon \ln p}
+\frac{(1-A_p^2)(1-\mathrm{e}^{\mathrm{i}\epsilon \ln p })}{1-A_p^2\mathrm{e}^{-\mathrm{i}\epsilon \ln p }}=
\frac{1-2A_p^2+A_p^2\mathrm{e}^{\mathrm{i} \epsilon \ln p}}{1-A_p^2 \mathrm{e}^{-\mathrm{i}\epsilon \ln p }} \label{ww}\\
&=& \frac{(1-A_p^2)^2}{|1-A_p^2\, \mathrm{e}^{\mathrm{i}\epsilon \ln p }|^2}
\left ( 1-\frac{A_p^4(1-\mathrm{e}^{\mathrm{i}\epsilon \ln p })^2}{(1-A_p^2)^2} \right ) \nonumber \ ,
\label{average_R_p}\end{aligned}$$ where as above $\epsilon=e_1-e_2$.
To calculate the product over $p<p^*$ we use Eq. ; the final answer is $$R_{2}^{\mbox{osc}}(\epsilon)=\frac{1}{4\pi^2}\mathrm{e}^{2\pi \mathrm{i} \overline{d(E)}\epsilon}|\zeta(1+\mathrm{i}\epsilon)|^2
\prod_p \left (1-\frac{(1-p^{\mathrm{i}\epsilon })^2}{(p-1)^2}\right ) + \mbox{ c.c. }\ .
\label{r_2osc}$$ Expressions and constitute two parts of the connected two-point correlation function of zeros of the Riemann zeta function. It is important to stress that the same expressions were obtained in [@bogomolny_keating] by a completely different method based on the Hardy-Littlewood conjecture concerning the distribution of prime pairs. Exactly the same formulae are also derived from the ratio conjecture [@ratiosCS]. Of course, all these methods are heuristic and cannot be considered as a true mathematical proof. Nevertheless, their mutual agreement means that if such a formula exists, it is likely to be given by the above expressions.
It is plain that when $\epsilon\to 0$, Eq. \[r\_2osc\] reproduces the two-point correlation function for the GUE ensemble of random matrices $$R_2^{\mbox{c}}(\epsilon)\underset{\epsilon\to 0}{\longrightarrow} -\frac{\sin^2(\pi \bar{d}\epsilon)}{\pi^2 \epsilon^2}\ .$$ In [@nonlinearity] using an averaged version of the Hardy-Littlewood conjecture it was shown that at small separations all correlation functions of the Riemann zeros agree with random matrix predictions.
Three-point correlation function {#three_point}
================================
The purpose of this Section is to calculate explicitly the three-point correlation function of zeros of the Riemann zeta function using the method discussed in the previous Sections.
By definition $$R_3(e_1,e_2,e_3)=\left \langle \left \langle \left |
\begin{array}{ccc}
\bar{\rho}_1&K_{12}&K_{13}\\
K_{21}&\bar{\rho}_2&K_{23}\\
K_{31}&K_{32}&\bar{\rho}_3
\end{array}\right | \right \rangle \right \rangle_{\Delta E}$$ where $\bar{\rho}_j=\rho(E+e_j)$ and $K_{ij}=K(E+e_i, E+e_j)$.
Expanding the determinant one gets $$R_3(e_1,e_2,e_3)=\bar{\rho}_1\, \bar{\rho}_2\, \bar{\rho}_3+K_{12}\, K_{23}\, K_{31}+K_{21}\, K_{13}\, K_{32}-
\bar{\rho}_1\, K_{23}K_{32}-\bar{\rho}_2\, K_{13}\, K_{31}-\bar{\rho}_3 \, K_{12}\, K_{21} \ .$$ Each $K_{ij}$ is given by and it is straightforward to check that $$K_{12}\, K_{23}\, K_{31}=\frac{1}{(2\pi \mathrm{i})^3 e_{12} e_{23 }e_{31}}
\left (
\mathrm{e}^{2\pi \mathrm{i} (N_1-N_2)}+\mathrm{e}^{2\pi \mathrm{i} (N_2-N_3)}+\mathrm{e}^{2\pi \mathrm{i} (N_3-N_1)}-
\mathrm{ c.c. }\right )$$ with $e_{ij}=e_i-e_j$.
From $\bar{\rho}_j\equiv \bar{\rho}(E+e_j)$ is the sum of two terms, $\bar{\rho}_j=\overline{d(E+e_j)} +\widetilde{d(E+e_j,p^*)}$. Using $\overline{d(E+e_j)}=\overline{d(E)}+ \mathcal{O}(e_j/E)$, ignoring the last correction terms, and, as above, denoting $d_j=\widetilde{d(E+e_j,p^*)}$ one obtains $$R_3(e_1,e_2,e_3)=\overline{d(E)}^{\ 3}+\overline{d(E)} \, R_2^{\mbox{c}}(e_{12})+
\overline{d(E)} \, R_2^{\mbox{c}}(e_{23})+
\overline{d(E)} \, R_2^{\mbox{c}}(e_{31}) +R_{3}^{\mbox{c}}(e_1,e_2,e_3)\ .$$ Here $R_2^{\mbox{c}}(e_{ij})$ is the connected two-point correlation function calculated in the previous Section, and $R_{3}^{\mbox{c}}(e_1,e_2,e_3)$ is the connected three-point function conveniently written as the sum of two terms $$R_{3}^{\mbox{c}}(e_1,e_2,e_3)= R_{3}^{\mbox{diag}}(e_1,e_2,e_3)+R_{3}^{\mbox{osc}}(e_1,e_2,e_3) .
\label{rthree}$$ Here $R_{3}^{\mbox{diag}}(e_1,e_2,e_3)$ is a smooth part $$R_{3}^{\mbox{diag}}(e_1,e_2,e_3)=\langle \langle d_1 d_2 d_3 \rangle \rangle_{\Delta E}\ ,
\label{diag}$$ and $R_{3}^{\mbox{osc}}(e_1,e_2,e_3)$ is an oscillatory part $$R_{3}^{\mbox{osc}}(e_1,e_2,e_3)=-\frac{1}{(2\pi i)^3}\left (
\frac{r_{12 }}{e_{12}^2}+
\frac{r_{23}}{e_{23}^2}+
\frac{r_{31}}{e_{31}^2}+\mbox{ c.c. }\right )$$ where $$\begin{aligned}
r_{12}&=&\left \langle \left \langle \left [ 2\pi \mathrm{i}
d_3-\frac{2 e_{12}}{e_{23}e_{31}}\right ]\mathrm{e}^{2\pi \mathrm{i} (N_1-N_2)}\right \rangle \right \rangle_{\Delta E} \ ,\\
r_{23}&=&\left \langle \left \langle \left [ 2\pi \mathrm{i}
d_1-\frac{2 e_{23}}{e_{12}e_{31}}\right ]\mathrm{e}^{2\pi \mathrm{i} (N_2-N_3)} \right \rangle \right \rangle_{\Delta E}\ ,\\
r_{31}&=&\left \langle \left \langle \left [ 2\pi \mathrm{i}
d_2 -\frac{2e_{31}}{e_{12}e_{23}}\right ]\mathrm{e}^{2\pi \mathrm{i} (N_3-N_1)}\right \rangle \right \rangle_{\Delta E}\ . \end{aligned}$$
Smooth terms
------------
The calculation of smooth (diagonal) contributions for the three-point correlation function of Riemann zeros is simplified by the fact that after averaging terms divergent when $p^*\to \infty$ disappear and only convergent sums remain. Indeed, the average over $E$ removes all products with different primes (cf. ). Therefore $$\begin{aligned}
&&R_{3}^{\mbox{diag}}(e_1,e_2,e_3)\equiv \langle \langle d_1\, d_2\, d_3\rangle \rangle_{\Delta E} =-\frac{1}{(2\pi)^3}\sum_p \ln^3 p \sum_{n_1,n_2,n_3=1}^{\infty}A_p^{n_1+n_2+n_3} \nonumber\\
&\times &
\Big \langle
\left (\mathrm{e}^{\mathrm{i} n_1(\Phi_p(E)+e_1\ln p)}+\mathrm{e}^{-\mathrm{i} n_1(\Phi_p(E)+e_1\ln p)}\right)
\left (\mathrm{e}^{\mathrm{i} n_2(\Phi_p(E)+e_2\ln p)}+\mathrm{e}^{-\mathrm{i} n_2(\Phi_p(E)+e_2\ln p)}\right)\nonumber\\
&\times &\left(\mathrm{e}^{\mathrm{i} n_3(\Phi_p(E)+e_3\ln p)}+\mathrm{e}^{-\mathrm{i} n_3(\Phi_p(E)+e_3\ln p)}\right )
\Big \rangle_{\Phi_p}\ .\end{aligned}$$ After averaging over $\Phi_p$, non-zero result gives only diagonal terms with $n_1=n_2+n_3$, $n_2=n_1+n_3$, and $n_3=n_1+n_2$. So $$\begin{aligned}
&&R_{3}^{\mbox{diag}}(e_1,e_2,e_3)= -\frac{1}{(2\pi)^3}\sum_p \ln^3 p \left [
\frac{1}{(p^{1-\mathrm{i} e_{12}}-1)(p^{1-\mathrm{i} e_{13}}-1)}+\right . \nonumber\\
&&+\left .\frac{1}{(p^{1-\mathrm{i} e_{21}}-1)(p^{1-\mathrm{i} e_{23}}-1)}
+ \frac{1}{(p^{1-\mathrm{i} e_{32}}-1)(p^{1-\mathrm{i} e_{31}}-1)}\right ]+\mathrm{c.c.}\ .
\label{rdiag}\end{aligned}$$
Oscillatory terms
-----------------
Using Eqs. (\[N\]) and (\[d\]) we obtain $$\begin{aligned}
&& \left \langle \left \langle 2\pi \mathrm{i} d_3 \mathrm{e}^{2\pi \mathrm{i} (N_1-N_2)}\right \rangle \right \rangle_{\Delta E} \nonumber\\
&&=\mathrm{e}^{2\pi \mathrm{i} \overline{d(E)} e_{12}}\frac{\partial}{\partial e_3}
\left \langle \left \langle \left [ \sum_{p<p^*}
\ln \frac{1-A_p\mathrm{e}^{\mathrm{i}\Phi_p(3)}}{1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(3)}} \right ]
\prod_{p<p^*} \frac{(1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(2)})(1-A_p \mathrm{e}^{\mathrm{i}\Phi_p(1)})}
{(1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(1)})(1-A_p\mathrm{e}^{\mathrm{i}\Phi_p(2)})} \right \rangle \right \rangle_{\Delta E}
\label{sum_product}\end{aligned}$$ where $\Phi_p(j)\equiv \Phi_p(E+e_j)=\Phi_p(E)+e_p\ln p$.
According to Eq. the average over $E$ is equivalent to the mean value over all phases. In Eq the sum over primes is multiplied by the product over the same primes. Therefore the average of each term in the sum, say with $p=q$, reduces to the following product of the averages $$\begin{aligned}
& &\left \langle \left \langle \frac{\partial}{\partial e_3}
\ln \left [ \frac{1-A_q\mathrm{e}^{\mathrm{i}\Phi_q(3)}}{1-A_q\mathrm{e}^{-\mathrm{i}\Phi_q(3)}} \right ]
\prod_{p<p^*} \frac{(1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(2)})(1-A_p \mathrm{e}^{\mathrm{i}\Phi_p(1)})}
{(1-A_p\mathrm{e}^{-\mathrm{i}\Phi_p(1)})(1-A_p\mathrm{e}^{\mathrm{i}\Phi_p(2)})} \right \rangle \right \rangle_{\Delta E}\nonumber \\
&=& T_q(e_1,e_2,e_3)\ \prod_{p\neq q} \langle R_p(\Phi_p ; e_1,e_2)\rangle_{\Phi_p} \end{aligned}$$ where $$\begin{aligned}
T_q( e_1,e_2,e_3)&=&\int_0^{2\pi} \frac{\mathrm{d}\Phi_q}{2\pi} \frac{\partial}{\partial e_3}
\ln \left [\frac{1-A_q\mathrm{e}^{\mathrm{i}(\Phi_q+e_3\ln q) }}{1-A_q\mathrm{e}^{-\mathrm{i}(\Phi_q+e_3\ln p)}}\right ]\nonumber\\
&\times &
\frac{(1-A_q\mathrm{e}^{-\mathrm{i}(\Phi_q+e_2\ln q)})(1-A_q \mathrm{e}^{\mathrm{i}(\Phi_q+e_1\ln q)})}
{(1-A_q\mathrm{e}^{-\mathrm{i}(\Phi_q+e_1\ln q)})(1-A_q\mathrm{e}^{\mathrm{i}(\Phi_q+e_2\ln q)})}\ , \end{aligned}$$ with $R_p(\Phi_p ; e_1,e_2)$ given by Eq. , and its average by Eq. .
As in the previous Section it is convenient to calculate $T_q(\Phi_q; e_1,e_2,e_3)$ by complex integration. Denoting $z=\mathrm{e}^{\mathrm{i}(\Phi_q}$, $A_q=A$, and $f_j=e_j\ln q$ one has $$T_q( e_1,e_2,e_3) = - \frac{\ln q}{2\pi}\oint \left [\frac{Az\mathrm{e}^{\mathrm{i}f_3} }{1-Az\mathrm{e}^{\mathrm{i}f_3}} + \frac{A\mathrm{e}^{-\mathrm{i}f_3}}{z-A\mathrm{e}^{-\mathrm{i}f_3}} \right ] \frac{(1-A z \mathrm{e}^{\mathrm{i} f_1})(z-A \mathrm{e}^{-\mathrm{i}f_2})}{(z-A\mathrm{e}^{-\mathrm{i}f_1})(1-Az\mathrm{e}^{\mathrm{i}f_2})}\frac{\mathrm{d}z}{z}\ .$$ Inside the unit circle the integrand has 3 poles, $z=0$, $z=A\mathrm{e}^{-\mathrm{i}f_1}$, and $z=A\mathrm{e}^{-\mathrm{i}f_3}$. Direct calculations give ($\phi_{ij}=f_i-f_j$) $$\begin{aligned}
&&T_q( e_1,e_2,e_3) = -\mathrm{i}\ln q \left [- \mathrm{e}^{\mathrm{i}\phi_{12}} +\Big (
\frac{A^2 \mathrm{e}^{\mathrm{i}\phi_{31}} }{1-A^2 \mathrm{e}^{\mathrm{i}\phi_{31}}}+
\frac{1}{\mathrm{e}^{\mathrm{i}\phi_{31}}-1} \Big )
\frac{(1-A^2)(1-\mathrm{e}^{\mathrm{i}\phi_{12}})}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{21}}} \right .\nonumber \\
&&+ \left . \frac{(1-A^2\mathrm{e}^{\mathrm{i}\phi_{13}})(1-\mathrm{e}^{\mathrm{i}\phi_{32}})}{(1-\mathrm{e}^{\mathrm{i}\phi_{31}})(1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}})}\right ]
= -\mathrm{i}\ln q \frac{(1-A^2)(1-\mathrm{e}^{\mathrm{i}\phi_{12}})}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{21}}}\left [
\frac{A^2 \mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2 \mathrm{e}^{\mathrm{i}\phi_{31}}}+\frac{A^2 \mathrm{e}^{\mathrm{i}\phi_{23}}}{1-A^2 \mathrm{e}^{\mathrm{i}\phi_{23}}}\right ] \ .\end{aligned}$$ Re-introducing the full notation it follows that ($e_{ij}=e_i-e_j$) $$T_q( e_1,e_2,e_3)=\frac{\partial}{\partial e_3} \ln \left [ \frac{1-A_q^2 \mathrm{e}^{\mathrm{i}e_{31}\ln q}}{1-A_q^2 \mathrm{e}^{\mathrm{i}e_{23}\ln q}} \right ] \frac{(1-A_q^2)(1-\mathrm{e}^{\mathrm{i}e_{12}\ln q})}{1-A_q^2\mathrm{e}^{\mathrm{i}e_{21}\ln q}} \ .$$ Using Eq. one gets $$\begin{aligned}
r_{12}&=&\mathrm{e}^{2\pi \mathrm{i} \overline{d(E)}e_{12}}\prod_{p<p^*}
\frac{(1-A_p^2)^2}{|1-A_p^2 \mathrm{e}^{\mathrm{i}e_{12}\ln p}|^2}
\left (1-\frac{A_p^4}{(1-A_p^2)^2}(1-\mathrm{e}^{\mathrm{i}e_{12}\ln p} )^2\right )\times \\
&\times &\frac{\partial}{\partial e_3}\left [\sum_{q<p^*}
\frac{(1-A_q^2)(1-\mathrm{e}^{\mathrm{i}e_{12}\ln q})}{1-2A_q^2+A_q^2 \mathrm{e}^{\mathrm{i}e_{12}\ln q}}
\ln \frac{1-A_q^2\mathrm{e}^{\mathrm{i}e_{31}\ln q}}{1-A_q^2\mathrm{e}^{\mathrm{i}e_{23}\ln q}}+ \ln \frac{e_{31}^2}{e_{32}^2} \right ] \ .
\end{aligned}$$ The summand in the square brackets can be transformed as follows $$\begin{aligned}
&&\frac{(1-A^2)(1-\mathrm{e}^{\mathrm{i}\phi_{12}})}{1-2A^2+A^2\mathrm{e}^{\mathrm{i}\phi_{12}}}
\ln \frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}=
\ln \frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}+\ln \frac{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{23}}} +\nonumber \\
&&+\frac{A^2-\mathrm{e}^{\mathrm{i}\phi_{12}}}{1-2A^2+A^2\mathrm{e}^{\mathrm{i}\phi_{12}}}\ln \frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}-\ln \frac{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{23}}}=\nonumber \\
&&=\ln \left |\frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}\right |^2 + A^2\frac{(1-\mathrm{e}^{\mathrm{i}\phi_{12}})^2}{1-2A^2+A^2\mathrm{e}^{\mathrm{i}\phi_{12}}}\ln \frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}-\nonumber \\
&&-\left [\mathrm{e}^{\mathrm{i}\phi_{12}}\ln \frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}+ \ln \frac{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{23}}}\right ]\ .
\label{terms}\end{aligned}$$ The expansions of all terms except the first one starts with $A^4\equiv A_q^4$ and, consequently, their sum over $q$ converge for large primes. In the divergent part, as above, we use , and $$\sum_{q<p^*} \ln \left |\frac{1-A_q^2\mathrm{e}^{\mathrm{i}e_{31}\ln q}}{1-A_q^2\mathrm{e}^{\mathrm{i}e_{23}\ln q}}\right |^2=
\ln \left |\prod_{q<p^*} \frac{1-A^2\mathrm{e}^{\mathrm{i}e_{31}\ln q}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}\ln q}}\right |^2
\underset{p^*\to\infty}{\longrightarrow} \ln \left |\frac{e_{23}\zeta(1+\mathrm{i} e_{23})}{e_{31}\zeta(1+\mathrm{i} e_{31})}\right |^2\ .$$ The other terms in can be transform as follows $$\begin{aligned}
&&\frac{\partial}{\partial e_3}\left [\mathrm{e}^{\mathrm{i}\phi_{12}}\ln\frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{
1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}+ \ln\frac{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{23}}}\right ]\nonumber \\
&&=\mathrm{i}\ln q\ A^4\left [ \frac{\mathrm{e}^{-\mathrm{i}\phi_{23}}(\mathrm{e}^{\mathrm{i}\phi_{32}}-\mathrm{e}^{\mathrm{i}\phi_{31}})}
{(1-A^2\mathrm{e}^{\mathrm{i}\phi_{32}})(1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}})} +\frac{\mathrm{e}^{-\mathrm{i}\phi_{31}}(\mathrm{e}^{-\mathrm{i}\phi_{31}}-\mathrm{e}^{-\mathrm{i}\phi_{32}})}{(1-A^2\mathrm{e}^{-\mathrm{i}\phi_{31}})(1-A^2\mathrm{e}^{-\mathrm{i}\phi_{32}})}\right ]\ .\end{aligned}$$ Similarly, the other terms in Eq. (\[terms\]) takes the form $$\begin{aligned}
&&\frac{\partial}{\partial e_3}\left [
A^2\frac{(1-\mathrm{e}^{\mathrm{i}\phi_{12}})^2}{1-2A^2+A^2\mathrm{e}^{\mathrm{i}\phi_{12}}}\ln\frac{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{23}}}\right ]=\nonumber \\
&&=-\mathrm{i}\ln q \ A^4\frac{(1-\mathrm{e}^{\mathrm{i}\phi_{12}})^2}{1-2A^2+A^2\mathrm{e}^{\mathrm{i}\phi_{12}}}\left (\frac{\mathrm{e}^{\mathrm{i}\phi_{31}}}{1-A^2\mathrm{e}^{\mathrm{i}\phi_{31}}} + \frac{\mathrm{e}^{-\mathrm{i}\phi_{32}}}{1-A^2\mathrm{e}^{-\mathrm{i}\phi_{32}}} \right )\ .\end{aligned}$$ Combining all terms together one finds $$\begin{aligned}
&&R_3^{\mbox{osc}}(e_1,e_2,e_3)=
-\frac{\mathrm{e}^{2\pi \mathrm{i}\overline{d(E)} e_{12}}}{(2\pi \mathrm{i})^3}
|\zeta(1+\mathrm{i}e_{12})|^2
\prod_p\left (1 -\frac{(1-p^{\mathrm{i}e_{12}})^2}{(p-1)^2}\right )
\left [\frac{\partial}{\partial e_3}
\ln\left |\frac{\zeta(1+\mathrm{i}e_{32})}{\zeta(1+\mathrm{i}e_{31})}\right |^2 \right . \nonumber\\
&&\left . -\mathrm{i} \sum_q \ln q \left ( \frac{ q^{\mathrm{i} e_{12}}-1}
{(q^{1+\mathrm{i} e_{23}}-1)(q^{1+\mathrm{i} e_{13}} -1)}\right . \right . +\frac{q^{\mathrm{i} e_{12}}-1}{(q^{1+\mathrm{i} e_{31}}-1)(q^{1+\mathrm{i} e_{32}}-1)}+\label{rosc}\\
&&+ \left . \left .\frac{(1-q^{\mathrm{i} e_{12}})^2}{q-2+q^{\mathrm{i} e_{12}}}
(\frac{1}{q^{1-\mathrm{i}e_{31}}-1}+\frac{1}{q^{1-\mathrm{i}e_{23}}-1}) \right )\right ]+\mathrm{cyclic\; permutations}
+\mathrm{c.c.}\ .
\nonumber\end{aligned}$$ Here “cyclic permutations” means that one has to add 2 other terms corresponding to cyclic permutations of indices $(1,2,3)$, i.e. terms with substitutions: $1\to 3,\; 2\to 1,\; 3\to 2$ and $1\to 2,\; 2\to 3,\; 3\to 1$.
Summary
=======
The principal ingredients of the proposed method for calculating correlation functions for the Riemann zeros are the following:
- A ’universal’ formula for the kernel of a GUE-type ensemble of random matrices with a given mean eigenvalue density, $\bar{\rho}(E)$, $$K(x,y)=\frac{\sin \pi \int_{x}^y \bar{\rho}(E)\mathrm{d}E }{\pi (x-y)}\ .$$
- The relation with the Riemann zeta function is established by fixing $\bar{\rho}(E)$ in the above expression as the finite part of the density of the zeros $$\bar{\rho}(E)=\frac{1}{2\pi}\ln \frac{E}{2\pi} -\frac{1}{2\pi}\sum_{p<p*}\sum_{n=1}^{\infty}
\frac{\ln p}{ p^{n/2}}\left (\mathrm{e}^{\mathrm{i} n E\ln p}+\mathrm{e}^{-\mathrm{i} n E\ln p}\right )$$ where the summation is performed over all prime numbers up to a certain cut-off value of $p^*$.
- Correlation functions are calculated by the averaging the GUE determinantal formula over a large window of $E$ $$R(e_1,\ldots, e_n)=\left \langle \left \langle \det( K (x,y) )_{x,y=E+e_1,\ldots, E+e_n}\right \rangle \right\rangle_{\Delta E}
\ .$$
- The average is assumed to be such that phases $E\ln p$ with different primes $p<p^*$ can be considered as independent random phases and the procedure of averaging is carried out by integration over these phases $$\left \langle \left \langle F\Big (\mathrm{e}^{\mathrm{i}E\ln p_1},\ldots, \mathrm{e}^{\mathrm{i}E\ln p_n} \Big ) \right \rangle \right\rangle_{\Delta E}
=\int_0^{2\pi}\frac{\mathrm{d}\phi_1}{2\pi}\cdots \int_0^{2\pi}\frac{\mathrm{d}\phi_n}{2\pi}
F\Big (\mathrm{e}^{\mathrm{i}\phi_1},\mathrm{e}^{\mathrm{i}\phi_2},\ldots, \mathrm{e}^{\mathrm{i}\phi_n} \Big )\ .$$
- After the averaging, the result consists of different sums over prime less than $p^*$. Those sums which converge when $p^*\to\infty$ are substituted by the sums over all primes. Sums divergent at large $p^*$ can be transformed to one particular product (or its logarithm or its derivative) whose limiting value is $$\prod_{p<p^*}\dfrac{1-p^{-1} }{1-p^{-1-s}}\underset{p^*\to\infty}{\longrightarrow} s\zeta(1+s)\ .$$
When the above rules are accepted, the calculation of correlation functions of Riemann zeros at a large height $E$ on the critical line reduces to purly algebraic manipulations (cf. Sections \[two\_point\] and \[three\_point\]). Our purpose here was to explain this new method, which, as we have emphasized, has the advantage over other heuristic approaches that it incorporates the determinanetal structure of RMT at the beginning (usually, it requires delicate combinatorial manipulations to establish this [@nonlinearity; @RudSar; @Zeev; @CS]). The method also extends straightforwardly to similar quantum chaotic problems.
For convenience we rewrite the obtained expressions given by Eqs. (\[rthree\]), (\[rosc\]), and (\[rdiag\]) together with Eqs. , , for the two-point and three-point correlation functions.
**Two-point correlation function**
$$R_2(\epsilon)=\overline{d(E)}^{\, 2}+R_2^{\mbox{c}}(\epsilon),\qquad
R_2^{\mbox{c}}(\epsilon)=R_{2}^{\mbox{diag}}(\epsilon)+R_{2}^{\mbox{osc}}(\epsilon)\ .$$
where $$\begin{aligned}
R_{2}^{\mbox{diag}}(\epsilon)&=&-\frac{1}{4\pi^2}\frac{\partial^2}{\partial \epsilon^2}
\ln|\zeta(1+\mathrm{i}\epsilon)|^2 -
\frac{1}{4\pi^2}\sum_p \ln^2 p \left (\frac{1}{(p^{1+\mathrm{i}\epsilon}-1)^2}
+\frac{1}{(p^{1-\mathrm{i}\epsilon}-1)^2}\right )\ ,
\label{twodiag}\\
R_{2}^{\mbox{osc}}(\epsilon)&=&\frac{\mathrm{e}^{2\pi i \overline{d(E)}\epsilon}}{4\pi^2}|\zeta(1+\mathrm{i}\epsilon)|^2
\prod_p\left (1-\frac{(1-p^{\mathrm{i}\epsilon})^2}{(p-1)^2}\right ) + \mbox{ c.c.}\ .
\label{twoosc}\end{aligned}$$
**Three-point correlation function**
$$\begin{aligned}
R_3(e_1,e_2,e_3)&=&\overline{d(E)}^{\, 3}+\overline{d(E)}R_2^{\mbox{c}}(e_{12})+
\overline{d(E)} R_2^{\mbox{c}}(e_{23})+
\overline{d(E)} R_2^{\mbox{c}}(e_{31})+R_3^{\mbox{c}}(e_1,e_2,e_3)\ , \nonumber\\
R_3^{\mbox{c}}(e_1,e_2,e_3)&=& R_{3}^{\mbox{diag}}(e_1,e_2,e_3)+R_{3}^{\mbox{osc}}(e_1,e_2,e_3)\ ,\end{aligned}$$
where $$\begin{aligned}
&&R_{3}^{\mbox{diag}}(e_1,e_2,e_3)= -\frac{1}{(2\pi)^3}
\sum_p \ln^3 p\left (
\frac{1}{(p^{1-\mathrm{i} e_{12}}-1)(p^{1-\mathrm{i} e_{13}}-1)}\right . +
\frac{1}{(p^{1-\mathrm{i} e_{21}}-1)(p^{1-\mathrm{i} e_{23}}-1)}\nonumber\\
&& \left . +\frac{1}{(p^{1-\mathrm{i} e_{32}}-1)(p^{1-\mathrm{i} e_{31}}-1)}\right ) +\mbox{c.c.} \end{aligned}$$ and $$\begin{aligned}
&&R_3^{\mbox{osc}}(e_1,e_2,e_3)=-\frac{\mathrm{e}^{2\pi \mathrm{i}\overline{d(E)}e_{12}}}{(2\pi \mathrm{i})^3}
|\zeta(1+\mathrm{i}e_{12})|^2
\prod_p \left (1 -\frac{(1-p^{\mathrm{i}e_{12}})^2}{(p-1)^2}\right ) \left [
\frac{\partial}{\partial e_3}\ln
\left |\frac{\zeta(1+\mathrm{i}e_{32})}{\zeta(1+\mathrm{i}e_{31})}\right |^2 \right .\nonumber\\
&&\left .-\mathrm{i} \sum_p \ln p \left ( \frac{ p^{\mathrm{i} e_{12}}-1}
{(p^{1+\mathrm{i} e_{23}}-1)(p^{1+\mathrm{i} e_{13}}-1)}\right .\right .+\frac{p^{\mathrm{i} e_{12}}-1}{(p^{1-\mathrm{i} e_{13}}-1)(p^{1-\mathrm{i} e_{22}}-1)}+
\label{threeosc}\\
&&+ \left . \left .\frac{(1-p^{\mathrm{i} e_{12}})^2}{p-2+p^{\mathrm{i} e_{12}}}
\Big (\frac{1}{p^{1-\mathrm{i}e_{31}}-1}+\frac{1}{p^{1-\mathrm{i}e_{23}}-1}\Big ) \right )\right ]+
\mathrm{ cyclic \;permutations}+\mathrm{c.c.}\ .
\nonumber\end{aligned}$$
[99]{} E. P. Wigner, *On the statistical distribution of the widths and spacings of nuclear resonance levels*, Math. Proc. Cambridge Phil. Soc.,**47**, 790-798 (1951). E. P. Wigner, *Random matrices in physics*, SIAM Review, **9**, 1-23 (1967). F. J. Dyson, *Statistical theory of the energy levels of complex systems*, J. Math. Phys., **3**, 140; ibid. II, 157; ibid. III, 166 (1962). C. E. Porter (editor), *Statistical theories of spectra: fluctuations*, Academic Press, New York (1965). M.L. Mehta, *Random matrices*, 2nd edition, Academic Press, New York. O. Bohigas, *Random matrix theories and chaotic dynamics*, in M.J. Giannoni, A. Voros, and J. Zinn-Justin, eds., *Chaos and quantum physics*, Les Houches, Session LII, 1989, 87-199, North-Holland, (1991). P.J. Forrester, N.C. Snaith, and J.J.M. Verbaarschot, *Developments in Random Matrix Theory*, J. Phys. A **36**, R1 (2003). G. Akemann, J. Baik and P. Di Francesco (editors), *The Oxford Handbook of Random Matrix Theory*, Oxford University Press (2011). M. Kac, *Statistical independence in probability, analysis and number theory*, Mathematical Association of America (1959). H. L. Montgomery, *The pair correlation of zeros of the zeta function*, Proc. Symp. Pure Math. **24**, 181-193 (1973). A.M. Odlyzko, *The $10^{20th}$ zero of the Riemann zeta function and 70 million of its neighbours*, Preprint 1989, unpublished, http://www.dtc.umn.edu/ odlyzko/ M.V. Berry, *Semiclassical formula for the number variance of the Riemann zeros*, Nonlinearity **1**, 399-407 (1988). E. Bogomolny and J. P. Keating, *Gutzwiller’s trace formula and spectral statistics: beyond the diagonal approximation*, Phys. Rev. Lett. **77**, 1472-1475 (1996). M.V. Berry and J.P. Keating, *The Riemann zeros and eigenvalue asymptotics*, SIAM Review **41**, 236, 1999. E. Bogomolny and J. P. Keating, *Two-point correlation function for Dirichlet $L$-functions*, J. Phys. A **46**, 095202 (2013). J. P. Keating and N. C. Snaith, *Random matrix theory and $\zeta (1/2+ \mathrm{i}t)$*, Com. Math. Phys., **214**, 57-89 (2000). J. P. Keating and N. C. Snaith, *Random matrix theory and $L$-functions at $s=1/2$*, Com. Math. Phys., **214**, 91-110 (2000). J.P. Keating and N.C. Snaith, *Random matrices and ${L}$-functions*, J. Phys. A **36**, 2859 (2003) J.P. Keating and N.C. Snaith, chapter 24 in [@RMTbook]. J.P. Keating and S. Müller, *Resummation and the semiclassical theory of spectral statistics*, Proc. R. Soc. Lond. A **463**, 3241-3250 (2007). E. Bogomolny, *Spectral statistics and periodic orbits*, in G. Casati, I. Gaurneri, and U. Smilansky (eds.), *New directions in quantum chaos*, Proc. Inter. School of Physics “Enrico Fermi”, Varenna, 333-368 (1999). J. B. Conrey and N. C. Snaith, *Triple correlation of the Riemann zeros*, *Journal de théorie des nombres de Bordeaux*, **20**, 61-106 (2008). J.B. Conrey, D.W. Farmer, and M.R. Zirnbauer, *Autocorrelation of ratios of ${L}$-functions*, Comm. Number Theory and Physics, **2**, 593 (2008). J.B. Conrey, D.W. Farmer, J.P. Keating, M.O. Rubinstein, and N.C. Snaith, *Integral moments of ${L}$-functions*, Proc. London Math. Soc., **91**, 33 (2005). A. Entin, E. Roditty-Gershon and Z. Rudnick, *Low-lying zeros of quadratic Dirichlet L-functions, hyper-elliptic curves and Random Matrix Theory*, preprint, arXiv:1208.5962. J. B. Conrey and N. C. Snaith, *$n$-correlation with restricted support*, arXiv:1212.5537 (2012). E.B.Bogomolny and J.Keating, *Random matrix theory and the Riemann zeros I: three- and four-point correlations*, Nonlinearity **8**, 1115-1131 (1995); ibid *Random matrix theory and the Riemann zeros II: n-point correlations*, Nonlinearity **9**, 911-935 (1995). Z. Rudnick and P. Sarnak, *Zeros of principal L-functions and random matrix theory*, Duke Math. J. **81**, 269-322 (1996). E. Bogomolny, O. Bohigas, P. Leboeuf, and A.G. Monastra, *On the spacing distribution of the Riemann zeros: corrections to the asymptotic result*, J. Phys. A [**39**]{}, 10743-10754 (2006). E. Duenez, D.K. Huynh, J.P. Keating, S.J. Miller and N.C. Snaith, *A random matrix model for elliptic curve L-functions of finite conductor*, J. Phys. A **45**, 115207 (2012). P. Deift, T. Kriecherbauer, K. T-R. McLaughlin, S. Venakides, and X. Zhou, *Strong Asymptotics of Orthogonal Polynomials with Respect to Exponential Weights*, Comm. Pure Appl. Math. **52**, 1491–1552 (1999). E. C. Titchmarsh, *The Theory of the Riemann zeta-function*, Oxford Science Publications, (1987). A. Weil, *Sur les “formules explicites” de la théorie des nombres premiers*, Comm. Sém. Math. Univ. Lund, 252–265 (1952). J.B. Conrey and N.C. Snaith, *Applications of the [$L$]{}-functions ratios conjectures*, Proc. London Math. Soc., **94**, 594 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In gravitational wave (GW) astronomy accurate measurement of the source parameters, such as mass, relies on accurate waveform templates. Currently, the templates are developed assuming that the source, such as a binary black hole (BBH), is residing in a vacuum. However, astrophysical models predict that BBHs could form in gaseous environments, such as common envelops, stellar cores, and accretion disks of active galactic nuclei. Here we revisit the impact of gas on the GW waveforms of stellar-mass BBHs with a focus on the early inspiral phase when the GW frequency is around milli-Hertz. We show that for these BBHs, gas friction could dominate the dynamical evolution and hence duplicate chirp signals. The relevant hydrodynamical timescale, $\tau_{\rm gas}$, could be much shorter than the GW radiation timescale, $\tau_{\rm gw}$, in the above astrophysical scenarios. As a result, the observable chirp mass is higher than the real one by a factor of $(1+\tau_{\rm gw}/\tau_{\rm gas})^{3/5}$ if the gas effect is ignored in the data analysis. Such an error also results in an overestimation of the source distance by a factor of $(1+\tau_{\rm gw}/\tau_{\rm gas})$. By performing matched-filtering analysis in the milli-Hertz band, we prove that the gas-dominated signals are practically indistinguishable from the chirp signals of those more massive BBHs residing in a vacuum environment. Such fake massive objects in the milli-Hertz band, if not appropriately accounted for in the future, may alter our understanding of the formation, evolution, and detection of BBHs.'
author:
- Xian Chen
- 'Ze-Yuan Xuan'
- Peng Peng
title: 'Fake massive black holes in the milli-Hertz gravitational-wave band'
---
Introduction {#sec:intro}
============
The majority of the black holes (BHs) detected by the ground-based gravitational wave (GW) observatories (i.e, LIGO and Virgo) turn out to be several times more massive than those previously detected in X-ray binaries [@ligo18a]. Such a discrepancy has important implications for the formation and evolution of stellar-mass binary BHs [BBHs, @ligo16astro]. However, mass is not a direct observable in GW astronomy. It is inferred either from the chirp signal, i.e., an increase of the GW frequency with time, or from the merger and ringdown signals. If the signal gets distorted, either at the moment of generation or during the propagation, an error would be induced in the measurement of the mass.
Redshift is such a disturbing factor. It stretches the signal during its propagation. As a result, mass ($m$) is degenerate with redshift ($z$) so that one can only measure from GW signals the redshifted mass $m(1+z)$ [@schutz86]. In two astrophysical scenarios, the mass-redshift degeneracy could lead to an significant overestimation of the masses of BBHs. In the first scenario, a BBH is at a high cosmological redshift and, nevertheless, is detected because the GW signal is magnified due to gravitational lensing [@broadhurst18; @smith18]. So far, there is no convincing evidence supporting this scenario [@hannuksela19]. This scenario also has difficulties explaining the positive correlation between the apparent masses and distances of the detected BBHs [@ligo18a]. In the second scenario, the BBH is captured by a supermassive black hole (SMBH) to a small distance, so that Doppler and gravitational redshifts become significant, which then lead to the mass-redshift degeneracy [@chen_han_2018; @chen19]. The uncertainty of this latter scenario lies mainly in the poor understanding of the event rate. In both scenarios the signal in the LIGO/Virgo band (centered around $10-10^2$ Hz) is similar to that of a not-redshifted, but more massive BBH. Distinguishing them would be difficult partly because the signal is short, which normally lasts no more than one second, too short to reveal any signature of gravitational lensing [@hannuksela19] or a nearby SMBH [@chen19].
The difficulty would be alleviated if the BBH can be detected by a space-borne GW observatory, such as the Laser Interferometer Space Antenna [LISA, @danzmann_2017]. LISA is sensitive to milli-Hertz (mHz) GWs and hence can capture a BBH at the early inspiral phase, weeks to millennia before it enters the LIGO/Virgo band [@miller02; @sesana16; @moore19]. Using the chirp signal, LISA can also measure the (redshifted) mass of the BBH. In this way, LISA may detect hundreds of massive BBHs during its mission duration of $4-5$ years [@sesana16; @kyutoku16; @lamberts18; @kremer18] and compare their masses with those from LIGO/Virgo observations.
Moreover, by tracking the BBHs with a decent signal-to-noise ratio (SNR, e.g., $>10$ ) for several months to years, LISA may reveal multiple images of a GW source if it is strongly lensed [@seto04; @sereno11]. The long signal may also reveal a shift of the GW phase caused by the wave effect of gravitational lensing [@nakamura98; @takahashi03lensing]. Moreover, if a BBH is close to a SMBH, the long waveform should also contain imprints of the orbital motion of the binary around the SMBH [@inayoshi17; @meiron17; @robson18; @chamberlain19; @tamanini19; @wong19; @torres-orjuela20] or the perturbation of the binary orbit by the tidal force of the SMBH [@meiron17; @hoang19; @randall19; @fang19]. These signatures can help us identify the BBHs affected by the redshift effects.
Redshift is not the only factor in GW astronomy that could affect the measurement of mass. Gas, for example, can exert a frictional force on a binary and hence lead to a faster orbital decay [@ostriker99; @kim07; @kim08]. The resulting GW signal is expected to deffer from the real chirp signal due to GW radiation only. The impact on the mass measurement deserves further investigation since a large fraction of BBHs may form in gaseous environments. For example, BBHs can be produced by binary-star evolution, and in this case the mergers may happen inside a common envelope [@ivanova13; @macleod17; @ginat19] or the fallback material from the previous supernovae [@tagawa18]. Moreover, some BBHs may form in the accretion disks of active galactic nuclei [AGNs, @mckernan12; @bartos17; @stone17]. When these BBHs merge, it is likely that they are surrounded by dense gas. Furthermore, the dense cores of massive stars may also produce BBHs, and hence the mergers would also be accompanied by gas [@loeb16; @fedrow17; @dorazio18]. The density of the gas can reach $10^{8}-10^{14}\,{\rm cm^{-3}}$ in the case of AGN disks, $10^{16}-10^{19}\,{\rm cm^{-3}}$ in common envelops, and even higher in stellar cores [see @antoni19 for a summary].
Several earlier works studied the impact of gas on the GW signal of merging binaries, including BBHs in the LIGO/Virgo band [@fedrow17; @ginat19; @cardoso19] and extreme-mass-ratio inspirals (another type of binary composed of a stellar-mass BH orbiting a SMBH) in the LISA band [@yunes11; @kocsis11; @barausse14]. They focused on the final evolutionary stage when the semi-major axes ($a$) of the binaries are only $10-10^2$ times the Schwarzschild radius ($r_S$) of the bigger BHs. At this stage, GW radiation predominates and gas plays a minor role. Nevertheless, these works showed that gas could induce a small phase shift to the GW signal. The phase shift does not significantly affect the mass measurement but can be used to identify the mergers happening in gas.
The BBHs in the LISA band are very different from those considered in the earlier works. They have much greater semi-major axes [$a\sim10^3-10^4\,r_S$, e.g., @chen17] since GW frequency is proportional to $a^{-3/2}$ . Such wide binaries have much weaker GW radiation because the GW power scales with $a^{-5}$ [@peters64]. Consequently, gas could play a dominant role in the evolution of these binaries and, in this way, produce a fake chirp signal. If this factor is not accounted for in the LISA data analysis, one may overestimate the masses, as we have shown in one example of our preliminary study [@chen19gas Paper I].
Here we present our full analysis of the problem and discuss the detectability of such “fake” massive BBHs in the LISA band. The paper is organized as follows. In §\[sec:theory\] we explain how mass is measured from a chirp signal and why it is affected by gas friction. Then in §\[sec:compare\] we show that in realistic astrophysical scenarios gas friction indeed can overcome GW radiation and dominate the orbital evolution of a inspiraling BBH. We compute the gas-dominated chirp signals in §\[sec:signal\] and show that they resemble the chirp signals from more massive BBHs in vacuum environments. In §\[sec:MF\], we employ the “matched-filtering” technique to quantify the similarity between the chirp signals in the cases with and without gas. Finally, in §\[sec:Dis\] we discuss the detectability of the fake massive BBHs in the LISA band.
Chirp signal and the effect of gas {#sec:theory}
==================================
We focus on the early inspiral phase because the majority of the LISA BBHs are in this evolutionary phase. Without gas, the semi-major axis $a$ of a BBH decays approximately as $$\dot{a}=\dot{a}_{\rm gw}:=\frac{64}{5}\frac{G^3m_1m_2m_{12}}{c^5a^3}\label{eq:adotgw}$$ [@peters64 assuming near-Keplerian circular orbits], where the dot symbol denotes the time derivative, $G$ is the gravitational constant, $c$ is the speed of light, $m_1$ and $m_2$ are the masses of the two BH members, and $m_{12}=m_1+m_2$ is the total mass of the binary. During the orbital decay, the GW frequency, which can be calculated with $f=\pi^{-1}(Gm_{12}/a^3)^{1/2}$, increases at a rate of $\dot{f}\propto f^{11/3}$. Such a characteristic signature is called the “chirp signal”. From it, one can derive a characteristic mass scale $${\cal
M}=\frac{c^3}{G}\,\left(\frac{5f^{-11/3}\dot{f}}{96\pi^{8/3}}\right)^{3/5}
=\frac{(m_1m_2)^{3/5}}{(m_1+m_2)^{1/5}},\label{eq:M}$$ which is known as the “chirp mass”. It uniquely determines the time evolution of $f$.
From the chirp signal one can also derive the distance $d$ of the BBH [@schutz86]. This is because from $f$ and $\dot{f}$ one can infer the energy-loss rate of the orbit, $\dot{E}\propto f^{-4}\dot{f}^2$, which also equals the GW power. In addition, the frequency $f$ and the GW amplitude $h$, together, determine a flux $S\propto h^2f^2$ which is the GW flux. From the power and the flux, one can derive the distance of the source $$d=\frac{4G}{c^2}\frac{{\cal M}}{h}\left(\frac{G}{c^3}\pi\,f\,{\cal M}\right)^{2/3}.\label{eq:d}$$
If the BBH is at a cosmological distance, the mass and distance encoded in the chirp signal will have slightly different meanings. First, both $f$ and $\dot{f}$ will be distorted by the redshift so that the observed frequency becomes $f_o=f(1+z)^{-1}$ and the chirping rate appears to be $\dot{f}_o=\dot{f}(1+z)^{-2}$. As a result, the chirp mass that one will derive from the redshifted GW signal becomes $${\cal
M}_o:=\frac{c^3}{G}\,\left(\frac{5f_o^{-11/3}\dot{f_o}}{96\pi^{8/3}}\right)^{3/5}={\cal M}(1+z).\label{eq:M_o}$$ This apparent chirp mass is bigger than the intrinsic one by a redshift factor $1+z$. Second, the GW amplitude will be determined by the transverse comoving distance $d_C$ in the following way, $$h_o=\frac{4G}{c^2}\frac{{\cal M}}{d_C}\left(\frac{G}{c^3}\pi\,f\,{\cal M}\right)^{2/3}.\label{eq:h_o}$$ If one uses the observed $f_o$, $\dot{f}_o$, and $h_o$ to infer a distance $d_o$, one will get $$d_o=\frac{4G}{c^2}\frac{{\cal M}_o}{h_o}\left(\frac{G}{c^3}\pi\,f_o\,{\cal M}_o\right)^{2/3}=d_C(1+z).\label{eq:d_o}$$ Such a distance is identical to the luminosity distance $d_L$ in a $\Lambda$CDM cosmology.
Gas will accelerate the orbital shrinkage and, in this way, affect the chirp signal. Because of the gas friction [or viscosity, e.g., @haiman09], a BBH would shrink at a faster rate of $\dot{a}=\dot{a}_{\rm gw}+\dot{a}_{\rm gas}$, where the additional term, $\dot{a}_{\rm gas}<0$, is due to gas. Correspondingly, $\dot{f}_o$ increases more rapidly, as $$\frac{\dot{f}_o}{f_o}=-\frac{3}{2}\left[\frac{\dot{a}_{\rm gw}+\dot{a}_{\rm gas}}{a(1+z)}\right].\label{eq:fdottot}$$
The apparent increase of $\dot{f}_o$ will lead to an overestimation of the mass of the BBH, as well as the distance. To see this effect, it is useful to first define an acceleration factor $$\Gamma:=\dot{a}_{\rm gas}/\dot{a}_{\rm gw}.$$ From this definition, it follows that $\dot{f}_o=(1+\Gamma)(1+z)^{-2}\dot{f}$, i.e., the chirp signal evolves faster by a factor of $1+\Gamma$. Finally, by revisiting Equations (\[eq:M\_o\]) and (\[eq:d\_o\]), we find that the apparent mass and distance become $$\begin{aligned}
{\cal M}_o&=(1+\Gamma)^{3/5}{\cal M}(1+z),\label{eq:calM_o}\\
d_o&=(1+\Gamma)\,d_C(1+z)=(1+\Gamma)\,d_L.\label{eq:bigdo}\end{aligned}$$
The factor $\Gamma$ in general is a function of $a$, because both $\dot{a}_{\rm
gw}$ and $\dot{a}_{\rm gas}$ depend on $a$. Such a dependence has two consequences. (i) If the semi-major axis changes substantially during the observational period, one would see a significant variation of ${\cal M}_o$ and $d_o$ with time. This result is inconsistent with the dynamical evolution of a BBH in a vacuum, and hence can be used to prove the presence of a environmental factor, such as gas. (ii) Otherwise, if $a$ evolves very slowly, i.e., the corresponding evolutionary timescale $a/|\dot{a}_{\rm gw}+\dot{a}_{\rm gas}|$ is much longer than the observational period $T_{\rm obs}$, the acceleration factor $\Gamma$ would be more or less a constant. In this case, the measurement of ${\cal M}_o$ and $d_o$ would be relatively consistent during the observational period, and both values would be greater than the intrinsic ones.
For LIGO/Virgo, the relevant BBHs are normally in the first case because the signal is typically less than a second but each LIGO/Virgo observing run lasts several weeks to several months. This is why the previous studies found that the gas effect could be discerned in the LIGO/Virgo waveforms (see §\[sec:intro\]). For LISA, however, the majority of the in-band BBHs belong to the latter case because a BBH could dwell in the band for as long as millions of years but the canonical mission duration of LISA is only $4-5$ years. During such a short observing time, $\Gamma$ is almost constant so that discerning the gas effect is more difficult. An overestimation of the mass and distance becomes more likely.
Hydrodynamics versus GW radiation {#sec:compare}
=================================
To evaluate the efficiency of the hydrodynamical drag, we compare the GW radiation timescale, defined as $\tau_{\rm
gw}:=|a/\dot{a}_{\rm gw}|$, and the hydrodynamical timescale, defined as $\tau_{\rm gas}:=|a/\dot{a}_{\rm gas}|$. Following these definitions, the acceleration factor $\Gamma$ equals $\tau_{\rm gw}/\tau_{\rm gas}$. Because the most sensitive band of LISA is around $3$ mHz, the corresponding BBHs have a typical semi-major axis of $$a=\left(\frac{Gm_{12}}{\pi^2 f^2}\right)^{1/3}\simeq0.0021
\left(\frac{m_{12}}{20~M_\odot}\right)^{1/3}
\left(\frac{f}{3\,{\rm mHz}}\right)^{-2/3}\,{\rm AU}.$$ According to Equation (\[eq:adotgw\]), without gas these binaries have a typical evolutionary timescale of $$\begin{aligned}
\tau_{\rm gw}&=\frac{5}{64}\frac{c^5a^4}{G^3m_1m_2m_{12}}\\
&\simeq\frac{9.1\times10^3}{q(1+q)^{-1/3}}
\left(\frac{m_1}{10\,M_\odot}\right)^{-5/3}
\left(\frac{f}{3\,{\rm mHz}}\right)^{-8/3}\,{\rm years},\end{aligned}$$ where $q$ denotes the mass ratio $m_2/m_1$ of the two BHs (we assume $m_1\ge m_2$).
As for $\tau_{\rm gas}$, we first use the hydrodynamic drag derived in @ostriker99 [@sanchez99] to estimate its value. The drag force on the secondary BH ($m_2$) is calculated with $F\sim4\pi\rho(Gm_2/v)^2$, where $\rho$ is the mass density of the background gas and $v$ is the Kepler velocity of the secondary. For circular orbits, we have $v\propto a^{-1/2}$, so that $a/\dot{a}_{\rm gas}=-v/(2\dot{v})$. Moreover, since $|\dot{v}|=F/m_2$, we derive that $$\tau_{\rm gas}=\frac{m_2v}{2F}\simeq\frac{1.1\times10^4}{q(1+q)^{2}}
\left(\frac{n}{10^{16}\,{\rm cm^{-3}}}\right)^{-1}
\left(\frac{f}{3\,{\rm mHz}}\right) \,{\rm yrs},$$ where $n$ is the number density of hydrogen atoms in the gas background. It follows that $$\begin{aligned}
\Gamma\simeq4.3\,\left(\frac{1+q}{2}\right)^{7/3}
\left(\frac{n}{10^{16}\,{\rm cm^{-3}}}\right)
\left(\frac{m_1}{10\,M_\odot}\right)^{-5/3}
\left(\frac{f}{3\,{\rm mHz}}\right)^{-11/3}.\label{eq:Gamm}\end{aligned}$$
In the above derivation of $\tau_{\rm gas}$, it is assumed that the gas background is homogeneous and the small body is moving in a straight line. However, for the BHs in binaries, which move along Keplerian orbits, it has been shown that the formula for the drag force will be modified, because the shape of the density wake is different [@sanchez01; @escala04; @kim07; @kim08]. More recently, @antoni19 showed that when $a$ is smaller than the Bondi accretion radius $R_{\rm
acc}=Gm_{12}/c_s^2$ ($c_s$ being the sound speed of the gas medium), the gas density close to the binary will be much higher than the background density due to accretion. This is normally the case for those BBHs embedded in common envelopes and AGN accretion disks. If we use $\tilde{n}\sim
n\,(R_{\rm acc}/a)^{3/2}$ to correct the gas density around the binary [@bondi52; @antoni19], we find that the timescale due to hydrodynamical drag becomes $$\begin{aligned}
{T}_{\rm gas}\simeq&8.5\times10^{4}q^{-1}(1+q)^{-3}
\left(\frac{n}{10^{11}\,{\rm cm^{-3}}}\right)^{-1}\nonumber\\
&\times
\left(\frac{m_1}{10\,M_\odot}\right)^{-1}
\left(\frac{c_s}{10^2\,{\rm km\,s^{-1}}}\right)^{3}\,{\rm years}.\label{eq:Tgas}\end{aligned}$$ Note that the new timescale does not depend on $a$ or $f$. A similar result can be found in [@bartos17 see their Fig. 2]. Moreover, in the last equation we have rescaled $n$ with $10^{11}\,{\rm cm^{-3}}$. Given this timescale, we derive that $$\begin{aligned}
{\Gamma}&\simeq1.1\,\left(\frac{1+q}{2}\right)^{10/3}
\left(\frac{n}{10^{11}\,{\rm cm^{-3}}}\right)\nonumber\\
&\times\left(\frac{m_1}{10\,M_\odot}\right)^{-2/3}
\left(\frac{c_s}{10^2\,{\rm km\,s^{-1}}}\right)^{-3}
\left(\frac{f}{3\,{\rm mHz}}\right)^{-8/3}.\label{eq:Gamm2}\end{aligned}$$ For illustrative purposes, we show in Figure \[fig:Gamm\] the dependence of the ${\Gamma}$ computed in the last equation on $m_1$ and $n$. The black dashed curve marks the location where ${\Gamma}=1$. Above it, gas dominates the dynamical evolution of a BBH, and hence the chirp signal is determined by gas dynamics, not GW radiation.
Equations (\[eq:Gamm\]) and (\[eq:Gamm2\]) suggest that for LIGO/Virgo BBHs, which typically have $m_1\sim10\,M_\odot$ and $f\sim10^2$ Hz, the gas effect is negligible unless the gas density $n$ is orders of magnitude higher than $10^{16}\,{\rm cm^{-3}}$. This is the reason that for LIGO/Virgo sources, significant gas effect is expected only in stellar cores, where $n$ could be as high as $(10^{28}-10^{31})\,{\rm cm^{-3}}$ [e.g. @fedrow17]. For LISA sources with $f\sim3$ mHz, however, gas effect is already important when $n\sim10^{16}\,{\rm cm^{-3}}$ according to Equation (\[eq:Gamm\]) or $n\sim10^{12}\,{\rm cm^{-3}}$ according to Equation (\[eq:Gamm2\]). These two characteristic densities can be found, respectively, in common envelopes and AGN accretion disks [e.g. @antoni19]. Therefore, the gaseous environments common for BBHs would affect the LISA signals more than the LIGO/Virgo ones.
Faking a chirp signal {#sec:signal}
=====================
Having understood the effect and the relative importance of gas, we now compute the chirp signal of a BBH embedded in a gaseous environment. In the following, we assume $z=0$ for simplicity ($f_o=f$). When there is no gas, we calculate the time derivative of the GW frequency ($\dot{f}_{\rm gw}$) using a 3.5 post-Newtonian (PN) approximation presented in @sathyaprakash09. When gas is present, we have $\dot{f}_o=\dot{f}_{\rm gw}+\dot{f}_{\rm gas}$, where $\dot{f}_{\rm gas}$ is calculated with $\dot{f}_{\rm gas}=-3f/(2\tau_{\rm
gas})$ according to Equation (\[eq:fdottot\]). In this way, the gas effect is included in the model through a parameter $\tau_{\rm gas}$.
Figure \[fig:long\] compares the long-term evolution of the chirp signal in the cases with and without the gas drag. The blue solid curve corresponds to a $10M_\odot-10M_\odot$ BBH (${\cal M}\simeq8.7\,M_\odot$) embedded in a gaseous medium. The two BHs coalesce at the time $t=t_c$. The hydrodynamical timescale $\tau_{\rm gas}$ is computed according to Equation (\[eq:Tgas\]) and the model parameters are chosen such that $\tau_{\rm gw}/\tau_{\rm gas}=10$ when $f=3$ mHz. This chirp signal around $f=3$ mHz, according to Equation (\[eq:calM\_o\]), should resemble a more massive binary with a chirp mass of $(1+\Gamma)^{3/5}{\cal M}\simeq37\,M_\odot$ in a vacuum. The chirp signal of the latter more massive BBH is shown as the dot-dashed curve, and we can see that, indeed, at $f=3$ mHz it matches the signal of the smaller binary embedded in gas. Eventually, the two signals diverge, since $\Gamma$ is decreasing, but the divergence appears more than one hundred years later. If we focus on the LISA observational window of five years (marked by the two vertical lines), the two signals are almost identical. Finally, during the coalescence, the blue curve recovers the chirp signal of a $10M_\odot-10M_\odot$ binary in a vacuum (red dashed curve), because GW radiation predominates during the merger.
![Evolution of the GW frequency as a function of the time before coalescence. The blue solid curve shows a $10M_\odot-10M_\odot$ circular BBH embedded in a gaseous environment. The model parameters are chosen such that ${\Gamma}=10$ when the GW frequency is $f=3$ mHz (see Eq.(\[eq:Gamm2\])). The two vertical lines mark the typical LISA observational window of five years. The red dashed curve shows the chirp signal of the same BBH but placed in a vacuum. The coalescence times of the previous two BBHs are aligned for easier comparison. The dot-dashed curve shows another BBH merging in vacuum but with higher masses. It is offset in time so that the GW frequency in the LISA observational window is also $3$ mHz. []{data-label="fig:long"}](longscale.pdf){width="50.00000%"}
To see more clearly the chirp signal in the observation window of LISA, we show in Figure \[fig:short\] the evolution of $f$ during a period of $1-2$ years, around the moment when $f$ is approximately $3$ mHz. Now the chirp signals look like straight lines because the observational period is orders of magnitude shorter than the evolutionary timescales of the BBHs. Although the variation of $f$ is small during the observation period, it is detectable by LISA because LISA’s resolution is approximately $10^{-8}(1\,{\rm yr}/T_{\rm obs})(10/ {\rm
SNR})$ Hz [@seto02]. Comparing the blue solid and the red dashed lines, we see that the presence of gas increases the slop the chirp signal. The steeper line resembles the chirp signal of a more massive BBH in a vacuum with a chirp mass of $37\,M_\odot$.
![Chirp signals in the LISA observational window. The blue solid curve is computed with the gas effect and the red dashed one without. The model parameters are the same as in Fig. \[fig:long\]. The gray dashed curves show the dependence of the chirp signal on the chirp mass when there is no gas.[]{data-label="fig:short"}](shortscale.pdf){width="50.00000%"}
We have seen that with or without gas, the chirp signals in the LISA band are almost straight lines and are relatively featureless compared to those in the LIGO/Virgo band. This is the main reason that in this band a BBH in a gaseous environment could be misidentified as a more massive binary in a vacuum.
Misidentification {#sec:MF}
=================
LISA uses a technique called “matched filtering” to search for BBHs in the data stream [@Finn92; @cutler94]. In this section, we will show that this method cannot distinguish light BBHs embedded in certain gaseous environments from those massive ones in vacuum.
Matched filtering
-----------------
Given two waveforms, $h_1$ and $h_2$, their similarity is quantified by a “fitting factor” (FF), which is defined as $${\rm FF}=\frac{\left<h_1|h_2\right>}{\sqrt{\left<h_1|h_1\right>\left<h_2|h_2\right>}}.$$ The term $\left<h_1|h_2\right>$ means an inner product of $$\left<h_1|h_2\right>=2\int_0^\infty\frac{\tilde{h}_1(f)\tilde{h}_2^*(f)+\tilde{h}_1^*(f)\tilde{h}_2}{S_n(f)}df,$$ where the tilde symbols stand for the Fourier transformation, the stars stand for the complex conjugation, and $S_n(f)$ is the spectral noise density of LISA [@klein16]. Identical waveforms have ${\rm FF}=1$.
In our problem, $h_1(t)$ is the chirp signal of a BBH embedded in a gaseous environment, and $h_2(t)$ is the waveform of a inspiraling BBH in a vacuum. By tuning the parameters of $h_2$, we want to maximize the FF. We follow @cutler94 and compute the waveforms using $$h(t)=\frac{Q(\theta, \varphi, \psi, \iota) \mu M}{d_L\, a(t)} \cos \left(\int 2 \pi f d t\right),$$ where $Q(\theta, \varphi, \psi, \iota)$ is a function depending on the sky location and orientation of the BBH. In the integrand, the frequency $f$ is a function of $a$. It is computed using the 3.5 PN approximation for $h_2$ [@sathyaprakash09] and using the gas model described in §\[sec:signal\] for $h_1$.
Because the two evolutionary timescales $\tau_{\rm gas}$ and $\tau_{\rm gw}$ are both much longer than the observational period of LISA, $f$ is almost a constant in our model. In this case the computation of the inner product $\left<h_1|h_2\right>$ can be performed in the time domain and the calculation of the FF can be simplified. First, the noise curve $S_n(f)$ can be taken out of the integration because of the small variation of $f$, so that $$\left<h_{1} | h_{2}\right>\approx \frac{2}{S_n} \int_{0}^{\infty} [\tilde{h}_{1}^{*}(f) \tilde{h}_{2}(f)+\tilde{h}_{1}(f) \tilde{h}_{2}^{*}(f)] d f.$$ Second, using Parseval’s theorem, we further derive $$\left <h_{1} | h_{2}\right >\approx \frac{4}{S}\int_{0}^{\infty} h_{1}(t) h_{2}(t)d t.$$ Finally, the FF can be written as $${\rm FF}=\frac{\int_{0}^{\infty} h_{1}(t) h_{2}(t)d t}{\sqrt{\int_{0}^{\infty} h_{1}(t) h_{1}(t)d t\int_{0}^{\infty} h_{2}(t) h_{2}(t)d t}}.$$ Note that there is no more dependence on $d_L$, $Q(\theta, \varphi, \psi,
\iota)$, or $S_n(f)$, because they all cancel out. Given $h_1$, i.e., the signal, we want to find a template $h_2$ that maximizes the FF. The parameter space in which we conduct this search is $({\cal M}_c,\,q,\,\phi)$, where $\phi$ is the initial phase.
Examples {#sec:examp}
--------
Figure \[fig:heatmap\] shows one example of our search. The signal is generated using a BBH with ${\cal M}_c=8.7\,M_\odot$, $q=0.7$, and a hydrodynamical timescale of $\tau_{\rm gas}=10^3$ years. Initially, the GW frequency is $f=3$ mHz, and the corresponding GW radiation timescale is about $10^4$ years. In this particular example the observational period is set to $T_{\rm obs}=1.25$ years, but later we will show the FF for different $T_{\rm
obs}$. We match the signal using the templates developed for vacuum BBHs, i.e., the 3.5 PN approximation described above. We use a simulated annealing algorithm to search for the highest FF in the parameter space of $({\cal
M}_c,\,q,\,\phi)$. The best FF is found at a chirp mass of ${\cal
M}_c\simeq37\,M_\odot$. It is offset from the real chirp masses by a factor of about $4.2$, which is consistent with our Equation (\[eq:calM\_o\]). This result confirms our prediction that ignoring the gas effect could result in a significant overestimation of the mass of a LISA BBH.
![Dependence of the FF on the chirp mass ${\cal M}_c$ and initial phase $\phi$ of the template $h_2$. The dependence on $q$ is relatively weak and is not shown here. The signal $h_1$ is generated using the parameters ${\cal M}_c=8.7\,M_\odot$, $q=0.7$, and $\tau_{\rm gas}=10^3$ years. The plus symbol marks the location of the maximum FF, which is offset from the real chirp mass ($8.7\,M_\odot$).[]{data-label="fig:heatmap"}](heat_map.pdf){width="50.00000%"}
We note that the best FF and the corresponding best-match ${\cal M}_c$ are both functions of $T_{\rm obs}$. Figure \[fig:FFsingle\] shows such a dependence on time. The model parameters are the same as in Figure \[fig:heatmap\]. During the first $1-2$ years, the FF remains close to $1$ and afterwards decreases with time. The FF deteriorates on a long timescale because the perturbation on the GW phase by the gas effect is accumulative. For the best-fit ${\cal M}_c$, it decreases with time but remains close to $37\,M_\odot$.
![Variation of the maximum FF and the best-fit chirp mass with the observational period. The real chirp mass of the BBH is $8.7\,M_\odot$.[]{data-label="fig:FFsingle"}](FFtime.pdf){width="50.00000%"}
Whether or not LISA is able to distinguish $h_1$ from $h_2$ depends on not only the FF but also the SNR, since noise also plays a role. The SNR is defined as $$\label{eq:SNR2}
{\rm SNR}^2:=\left<h|h\right>.$$ According to @lindblom08, two waveforms are distinguishable when $\left<\delta h|\delta h\right> >1$, where $\delta h:=\tilde{h}_1(f)-\tilde{h}_2(f)$. This criterion can be simplified in our problem because we are often in a situation where $h_1\simeq h_2$. In this case we have ${\rm SNR}^2\simeq\left<h_1|h_1\right>\simeq\left<h_2|h_2\right>$, and the criterion reduces to $${\rm
FF}<1-1/(2\,{\rm SNR}^2).\label{eq:FF}$$ If we take ${\rm SNR}\simeq10$ as the threshold for LISA to claim a detection, the corresponding criterion of distinguishing two different waveforms becomes ${\rm FF}<0.995$.
According to this criterion, the gas and vacuum waveforms used in Figure \[fig:FFsingle\] are indistinguishable during the first $1-2$ years of observation. Only in the third year could one start to tell the difference and prove that the signal is not produced by a massive BBH of ${\cal
M}_c\simeq37\,M_\odot$ residing in a vacuum environment.
For completeness, we show in Figure \[fig:FFmultiple\] the FF derived assuming different values for $\tau_{\rm gas}$. The other model parameters are the same as in Figure \[fig:heatmap\]. As $\tau_{\rm gas}$ increases, the FF gets better at later times because the gas effect becomes weaker. We note that when $\tau_{\rm gas}\ga5,000$ years, the FF is better than $0.995$ for almost five years, which is equivalent to the canonical mission duration of LISA. As a result, LISA may identify our BBHs of a chirp mass of $8.7\,M_\odot$ as more massive binaries. The measured chirp mass is approximately $18\,M_\odot$ when $\tau_{\rm gas}=5,000$ years and $14\,M_\odot$ when $\tau_{\rm gas}=10^4$ years. In both cases, the overestimation of the mass, and hence the distance (see Eq. (\[eq:bigdo\])), is substantial.
![Variation of the maxim FF with time assuming different values of $\tau_{\rm gas}$.[]{data-label="fig:FFmultiple"}](FF-T.pdf){width="50.00000%"}
Discussions {#sec:Dis}
===========
We have seen that those BBHs with $\Gamma=\tau_{\rm gw}/\tau_{\rm gas}\ga1$ could be misidentified by LISA as more massive binaries residing in vacuum environments. To what distance could LISA detect such fake massive binaries? The standard way of addressing this question is to derive the maximum distance at which the SNR drops to a threshold, say ${\rm SNR}=10$. This distance is known as the “detection horizon”. In the case without gas, the detection horizon has been derived in several works assuming different LISA configurations. For example, @kyutoku17 showed that $$\begin{aligned}
d_L\simeq&13\,\left(\frac{{\cal M}_c}{10\,M_\odot}\right)^{5/3}
\left(\frac{T_{\rm obs}}{5\,{\rm yr}}\right)^{1/2}
\left(\frac{{\rm SNR}}{10}\right)^{-1}\nonumber \\
&\times
\left(\frac{S_n(f)}{10^{-40}\,{\rm Hz^{-1}}}\right)^{-1/2}
\left(\frac{f}{3\,{\rm mHz}}\right)^{2/3}\,{\rm Mpc}\label{eq:SNR}\end{aligned}$$ for the N2A5 configuration of LISA. Here ${\cal M}_c$ and $d_L$ refer to the real chirp mass and real distance of the source.
In fact, the last equation is valid also in the case with gas. This is so because of the following reasons. (i) The stationary phase approximation [@thorne87] in which the last equation is derived remains valid, since the evolutionary timescale $f/\dot{f}$ for the frequency is much longer than the GW period $1/f$. We note that $\dot{f}$ here stands for the observed frequency. We dropped the subscript $o$ in $\dot{f}_o$ for simplicity. (ii) In this approximation, the characteristic amplitude defined as $h_c(f):=2f\tilde{h}(f)$ becomes proportional to $A\dot{f}^{-1/2}$, where $A$ is a function of ${\cal M}_c$, $q$, $f$, and $d_L$, as well as the sky location and orientation of the binary. Noticing that $\dot{f}=\dot{f}_{\rm gw}+\dot{f}_{\rm gas}>\dot{f}_{\rm gw}$, we find that gas in general reduces the characteristic amplitude at any frequency. The reduction is due to a faster drift of the signal in the frequency domain. (iii) Using $h_c$, we can rewrite the SNR defined in Equation (\[eq:SNR2\]) as $${\rm SNR}^2=\int_{f_1}^{f_2}\frac{|h_c(f)|^2}{f^2S_n(f)}df,$$ where $f_1$ and $f_2$ denote the minimum and maximum frequencies during the observational period. Because $\Delta f=f_2-f_1\ll f$, the integration becomes proportional to $|h_c|^2\Delta f/(f^2S_n)$. (iv) In our problem we have $T_{\rm
obs}\ll |f/\dot{f}|$. Therefore, we can write $\Delta f\simeq \dot{f}T_{\rm
obs}$. Again, by noticing that $\dot{f}>\dot{f}_{\rm gw}$, we find that $\Delta f$ is broader when gas is present. (v) Finally, the $\dot{f}^{-1}$ from the term $|h_c|^2$ cancels the $\dot{f}$ from the term $\Delta f$, so that the SNR does not depend on $\dot{f}$. Physically, this means the reduction of the characteristic amplitude is compensated by the larger frequency drift.
The above conclusion that gas does not affect the SNR is derived in the scenario of LISA observations. It does not apply to LIGO/Virgo because in the latter case the assumption $T_{\rm obs}\ll |f/\dot{f}|$ is invalid. In fact, gas will reduce the SNR for LIGO/Virgo sources by suppressing $|h_c|$. Nevertheless, the corresponding change of SNR is small because, as has been explained in §\[sec:compare\], the acceleration factor $\Gamma$ is small when a BBH enters the LIGO/Virgo band.
Therefore, we can use Equation (\[eq:SNR\]) to estimate the detection horizon and, based on it, discuss the detectability of the fake massive BBHs embedded in gaseous environment. For ${\cal M}_c=10\,M_\odot$, the detection horizon corresponding to a SNR of $10$ is approximately $13$ Mpc, assuming $5$ years of observation. If ${\cal M}_c=30\,M_\odot$, as the LIGO/Virgo observations tend to suggest [@ligo18a], the detection horizon elongates to about $80$ Mpc.
To estimate the number of fake massive BBHs within the detection horizon, we start from the event rate in the LIGO/Virgo band, which is estimated to be ${\cal O}(10^2)\,{\rm Gpc^{-3}\,yr^{-1}}$ [@ligo18a]. According to this rate, the number of BBHs in the last year before their coalescence is about $N(\tau=1\,{\rm yr})=100$ per ${\rm Gpc}^{3}$, where $\tau=|a/\dot{a}|$ denotes the orbital evolutionary timescale. For the other BBHs at an earlier evolutionary stage, we can follow the continuity equation [e.g., see Sec. V in @amaro-seoane19] and derive that $dN/d\ln a\propto \tau$. In our problem, $\tau=(1/\tau_{\rm gw}+\tau_{\rm gas})^{-1}$. At a frequency of $f=3$ mHz, where LISA is the most sensitive, $\tau_{\rm gw}\simeq(1500-9000)$ years when ${\cal M}_c$ varies from $30\,M_\odot$ to $10\,M_\odot$. (i) Without gas, $\tau=\tau_{\rm gw}$, and we find that the number density of BBHs at $f\sim3$ mHz is about $(1.5-9)\times10^{5}\,{\rm Gpc}^{-3}$. The number of BBHs inside the detection horizon is $8-320$. (ii) With gas, the number would be smaller because $\tau$ is shortened by gas friction. In the extreme case that all BBHs are embedded in gas, if we assume $\tau_{\rm gas}=10^3$ years, we find that $\tau\simeq(600-900)$ years when ${\cal M}_c$ varies from $30\,M_\odot$ to $10\,M_\odot$. Correspondingly, the number density of BBHs at $f\sim3$ mHz is $(6-9)\times10^{4}\,{\rm Gpc}^{-3}$. The number of BBHs inside the detection horizon decreases to $0.8-130$, but is not zero.
Since BBHs embedded in gaseous environments could be common, the effect of hydrodynamics should be considered more carefully in the waveform modeling. Otherwise, as our results suggest, LISA may provide a biased demography of BBHs. Such a bias may also affect future cosmology studies, given the possibility of using BBHs as standard sirens to measure cosmological parameters.
[*Acknowledgement.*]{}–This work is supported by the NSFC grants No. 11873022 and 11991053. XC is supported partly by the Strategic Priority Research Program “Multi-wavelength gravitational wave universe” of the Chinese Academy of Sciences (No. XDB23040100 and XDB23010200). The computation in this work was performed on the High Performance Computing Platform of the Center for Life Science, Peking University.
[64]{} natexlab\#1[\#1]{}href \#1\#2urllinklabel adsurllinklabel
, B. P., [Abbott]{}, R., [Abbott]{}, T. D., [Abernathy]{}, M. R., [Acernese]{}, F., [Ackley]{}, K., [Adams]{}, C., [Adams]{}, T., [Addesso]{}, P., [Adhikari]{}, R. X., & et al. 2016, , 818, L22
, P. 2019, , 99, 123025
, P., [Audley]{}, H., [Babak]{}, S., [Baker]{}, J., [Barausse]{}, E., [Bender]{}, P., [Berti]{}, E., [Binetruy]{}, P., [Born]{}, M., [Bortoluzzi]{}, D. et al. 2017, arXiv:1702.00786
, A., [MacLeod]{}, M., & [Ramirez-Ruiz]{}, E. 2019, , 884, 22
, E., [Cardoso]{}, V., & [Pani]{}, P. 2014, , 89, 104059
, I., [Kocsis]{}, B., [Haiman]{}, Z., & [M[á]{}rka]{}, S. 2017, , 835, 165
, H. 1952, , 112, 195
, T., [Diego]{}, J. M., & [Smoot]{}, George, I. 2018, arXiv e-prints, arXiv:1802.05273
, V. & [Maselli]{}, A. 2019, arXiv e-prints, arXiv:1909.05870
, K., [Moore]{}, C. J., [Gerosa]{}, D., & [Yunes]{}, N. 2019, , 99, 024025
, X. & [Amaro-Seoane]{}, P. 2017, , 842, L2
, X. & [Han]{}, W.-B. 2018, Communications Physics, 1, 53
, X., [Li]{}, S., & [Cao]{}, Z. 2019, , 485, L141
Chen, X. & Shen, Z. 2019, Proceedings, 17, 4
, C. & [Flanagan]{}, [É]{}. E. 1994, , 49, 2658
, D. J. & [Loeb]{}, A. 2018, , 97, 083008
, A., [Larson]{}, R. B., [Coppi]{}, P. S., & [Mardones]{}, D. 2004, , 607, 765
, Y., [Chen]{}, X., & [Huang]{}, Q.-G. 2019, , 887, 210
, J. M., [Ott]{}, C. D., [Sperhake]{}, U., [Blackman]{}, J., [Haas]{}, R., [Reisswig]{}, C., & [De Felice]{}, A. 2017, , 119, 171103
, L. S. 1992, Phys. Rev. D, 46, 5236
, Y. B., [Glanz]{}, H., [Perets]{}, H. B., [Grishin]{}, E., & [Desjacques]{}, V. 2020,
, Z., [Kocsis]{}, B., & [Menou]{}, K. 2009, , 700, 1952
Hannuksela, O. A., Haris, K., Ng, K. K. Y., Kumar, S., Mehta, A. K., Keitel, D., Li, T. G. F., & Ajith, P. 2019, The Astrophysical Journal, 874, L2
, B.-M., [Naoz]{}, S., [Kocsis]{}, B., [Farr]{}, W. M., & [McIver]{}, J. 2019, , 875, L31
, K., [Tamanini]{}, N., [Caprini]{}, C., & [Haiman]{}, Z. 2017, , 96, 063014
, N., [Justham]{}, S., [Chen]{}, X., [De Marco]{}, O., [Fryer]{}, C. L., [Gaburov]{}, E., [Ge]{}, H., [Glebbeek]{}, E., [Han]{}, Z., [Li]{}, X.-D., [Lu]{}, G., [Marsh]{}, T., [Podsiadlowski]{}, P., [Potter]{}, A., [Soker]{}, N., [Taam]{}, R., [Tauris]{}, T. M., [van den Heuvel]{}, E. P. J., & [Webbink]{}, R. F. 2013, , 21, 59
, H. & [Kim]{}, W.-T. 2007, , 665, 432
, H., [Kim]{}, W.-T., & [S[á]{}nchez-Salcedo]{}, F. J. 2008, , 679, L33
Klein, A., Barausse, E., Sesana, A., Petiteau, A., Berti, E., Babak, S., Gair, J., Aoudia, S., Hinder, I., Ohme, F., & Wardell, B. 2016, Phys. Rev. D, 93, 024003
, B., [Yunes]{}, N., & [Loeb]{}, A. 2011, , 84, 024032
, K., [Chatterjee]{}, S., [Breivik]{}, K., [Rodriguez]{}, C. L., [Larson]{}, S. L., & [Rasio]{}, F. A. 2018, , 120, 191103
, K. & [Seto]{}, N. 2016, , 462, 2177
—. 2017, , 95, 083525
, A., [Garrison-Kimmel]{}, S., [Hopkins]{}, P. F., [Quataert]{}, E., [Bullock]{}, J. S., [Faucher-Gigu[è]{}re]{}, C. A., [Wetzel]{}, A., [Kere[š]{}]{}, D., [Drango]{}, K., & [Sand erson]{}, R. E. 2018, , 480, 2704
, L., [Owen]{}, B. J., & [Brown]{}, D. A. 2008, , 78, 124020
, A. 2016, , 819, L21
MacLeod, M., Antoni, A., Murguia-Berthier, A., Macias, P., & Ramirez-Ruiz, E. 2017, The Astrophysical Journal, 838, 56
, B., [Ford]{}, K. E. S., [Lyra]{}, W., & [Perets]{}, H. B. 2012, , 425, 460
, Y., [Kocsis]{}, B., & [Loeb]{}, A. 2017, , 834, 200
, M. C. 2002, , 581, 438
, C. J., [Gerosa]{}, D., & [Klein]{}, A. 2019, , 488, L94
, T. T. 1998, , 80, 1138
, E. C. 1999, , 513, 252
, P. C. 1964, Physical Review, 136, 1224
, L. & [Xianyu]{}, Z.-Z. 2019, arXiv:1902.08604
, T., [Cornish]{}, N. J., [Tamanini]{}, N., & [Toonen]{}, S. 2018, , 98, 064012
, F. J. & [Brandenburg]{}, A. 1999, , 522, L35
—. 2001, , 322, 67
, B. S. & [Schutz]{}, B. F. 2009, Living Reviews in Relativity, 12, 2
, B. F. 1986, , 323, 310
, M., [Jetzer]{}, P., [Sesana]{}, A., & [Volonteri]{}, M. 2011, , 415, 2773
, A. 2016, Physical Review Letters, 116, 231102
, N. 2002, , 333, 469
—. 2004, , 69, 022002
, G. P., [Jauzac]{}, M., [Veitch]{}, J., [Farr]{}, W. M., [Massey]{}, R., & [Richard]{}, J. 2018, , 475, 3823
, N. C., [Metzger]{}, B. D., & [Haiman]{}, Z. 2017, , 464, 946
Tagawa, H., Saitoh, T. R., & Kocsis, B. 2018, Phys. Rev. Lett., 120, 261101
, R. & [Nakamura]{}, T. 2003, , 595, 1039
, N., [Klein]{}, A., [Bonvin]{}, C., [Barausse]{}, E., & [Caprini]{}, C. 2020, , 101, 063002
& [the Virgo Collaboration]{}. 2019, Physical Review X, 9, 031040
, K. S. [Gravitational radiation.]{}, ed. , S. W. [Hawking]{}W. [Israel]{} (Cambridge University Press), 330–458
, A., [Chen]{}, X., & [Amaro-Seoane]{}, P. 2020, arXiv e-prints, arXiv:2001.00721
, K. W. K., [Baibhav]{}, V., & [Berti]{}, E. 2019, , 488, 5665
, N., [Kocsis]{}, B., [Loeb]{}, A., & [Haiman]{}, Z. 2011, , 107, 171103
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A set of points and a positive integer $m$ are given and our goal is to cover the maximum number of these point with $m$ disks. We devise the first output sensitive algorithm for this problem. We introduce a parameter $\rho$ as the maximum number of points that one disk can cover. In this paper first we solve the problem for $m=2$ in $O({n\rho} + {\rho ^3}\log \rho ))$ time. The previous algorithm for this problem runs in $O({n^3}\log n)$ time. Our algorithm outperforms the previous algorithm because $\rho$ is much smaller than $n$ in many cases. Then we extend the algorithm for any value of $m$ and we solve the problem in $O(m{n\rho} + {(m\rho )^{2m - 1}}\log m\rho )$ time. The previous algorithm for this problem runs in $O({n^{2m - 1}}\log n)$ time. Our algorithm runs faster than the previous algorithm because $m\rho$ is smaller than $n$ in many cases. Our technique to obtain an output sensitive algorithm is to use a greedy algorithm to confine the areas that we should search to obtain the result. Our technique in this paper may be applicable in other set covering problems that deploy a greedy algorithm, to obtain faster solutions.'
address: |
SSRD Lab, Computer Engineering and Information technology department, Amirkabir University of technology, Tehran, Iran.\
School of Computer Science, Institute for Research in Fundamental Sciences(IPM), P.O. Box:19395-5746, Tehran, Iran.
author:
- 'Hossein Ghasemalizadeh[^1]'
- 'Mohammadreza Razzazi[^2]'
title: Output sensitive algorithm for covering many points
---
Introduction
============
In the classic covering problem, a set of points are given, and the goal is to place the minimum number of unit disks to cover all the input points. In the maximum coverage problem, the number of disks to be used, $m$, is given and the goal is to cover the maximum number of points with $m$ disks. We call this problem $MostPoints(P,m)$. This problem was introduced by Drezner [@1] for $m=1$ and it was solved in $O({n^2}\log n)$ time. To solve the problem, he replaced every point with a disk centered at that point and he obtained the maximum depth in the arrangement of disks. Later, Chazelle and Lee [@2] solved the problem in $O(n^2)$ time. This problem belongs to the 3-SUM hard complexity class [@3], which means that, the $O(n^2)$ running time algorithm is the best one for this problem. However, there are some algorithms which approximately solve this problem in less than $O(n^2)$ time. These algorithms approximate the disk radius and the number of covered points. Figueiredo and Fonesca [@5] gave an algorithm to cover the maximum number of points that a unit disk in ${\Re^d}$ can cover, with a $(1 + \varepsilon )$-radius disk in $O(n/{\varepsilon ^{d - 1}})$ time. In [@4] Aronov and Har-peled gave a $(1 - \varepsilon )$-approximation algorithm in the number of covered points which runs in $O(n{\varepsilon ^{ - 2}}\log n)$ time. $MostPoints(P,m)$, for $ m > 1 $, is NP-Hard unless we consider $m$ as a constant. A trivial greedy algorithm is a $(1 - \frac{1}{e})$-approximation algorithm for it [@6]. The greedy algorithm first finds a disk which covers the maximum number of points in $O(n^2)$ time. To pick the next disks, it removes the points located in the first disk, and finds the disk which covers the maximum total weight of points as the second disk. It repeats this process until $m$ disks are picked up. This yields a $(1 - \frac{1}{e})$-approximation algorithm which runs in $O(m{n^2})$ time. The first $(1 - \varepsilon )$-approximation algorithm for this problem was given in [@7] which runs in $O(n\log n + n{\varepsilon ^{ - 6m + 6}}\log (\frac{1}{\varepsilon }))$ time. We presented a polynomial time approximation schema for this problem in our previous work [@8], which runs in $O((1 + \varepsilon )mn + {\varepsilon ^{ - 1}}{n^{4\sqrt 2 {\varepsilon ^{ - 1}} + 2}})$ time.
The trivial method to obtain the optimal result of the $MostPoints(P,m)$ is to consider all subsets of size $m$ of $n^2$ possible disks and finds the subset which covers the most number of points. This method takes $O(n^{2m})$ time. In [@7] de Berg et. al gave an algorithm for this problem. They first solved the problem for $m=2$ in $O({n^3}\log n)$ time. To solve the problem for $m>2$, they fixed every subset of $m-2$ disks, and they found the best 2 disks after removing the points contained in the $m-2$ disks. This takes $O({n^{2m - 1}}\log n)$ time.\
In section 2, we present an output sensitive algorithm to obtain the optimal result of $MostPoints(P,2)$ which runs in $O({n\rho} + {\rho ^3}\log \rho )$ time where $\rho$ is the maximum number of points that one disk can cover. In Section 3 we extends the algorithm for $MostPoints(P,m)$ . In section 4 we compare the implementation results of our algorithm and the algorithm of [@7] and in section 5 we conclude the paper. Our innovation in this algorithm is to use the greedy algorithm to find the regions in which the $m$ resultant disks may reside, and searching in those regions only. In the rest of the paper when we use disks we mean unit disks.
Output sensitive algorithm for MostPoints(P,2)
==============================================
In this section we describe our algorithm for $MostPoints(P,2)$. Let $g_1 =MostPoints(P,1)$ be the disk that covers the maximum number points from the point set $P$. Let $Points(g_1)$ refers to the points of $P$ which are located inside $g_1$. Define $g_2=MostPoints(P - Points(g_1),1)$ as the disk that covers the maximum number of points after removing the points located in $g_1$. $g_1$ and $g_2$ are the result of the greedy algorithm for $MostPoints(P,2)$. An example is given in Figure 1. In this example, the optimal solution covers 18 points whereas the greedy algorithm covers 15 points. As illustrated in this example, the two disks of the optimal solution have common points with $g_1$. Intuitively, either the disks of the optimal solution have common points with $g_1$ or, the greedy algorithm obtains the optimal solution. Lemma 1 proves this claim.
![This figure shows the greedy solution and the optimal solution for $MostPoints(P,2)$, in a sample point set. In this example, the greedy algorithm returns $g_1$ and $g_2$ which cover 15 points together, whereas the disks in the optimal solution, the two disks specified with the dash lines, cover 18 points together. []{data-label="fig:fig1"}](Picture1){width=".5\columnwidth"}
By the best two disks we mean the optimal solution of $MostPoints(P,2)$. Let $D$ be a set of disks. The function $Cover(D)$ denotes the number of points covered by the disks in $D$.
**Lemma 1**: Let $g_1$ be one of the disks that covers the maximum number of points and $g_2$ be one of the disks that covers the maximum number of points after removing the points located in $g_1$. Let $o_1$ and $o_2$ be the two disks that together cover the maximum number of points among any combination of two disks: Either both $o_1$ and $o_2$ have common points with $g_1$, or $g_1$ and $g_2$ cover the maximum number of points.
Suppose that at least one of $o_1$ and $o_2$, say $o_2$, does not have any common point with $g_1$. Then we have $Cover(\{g_1, g_2\}) \geq Cover(\{g_1 , o_2\})$, because $g_2$ covers the maximum number of points not covered by $g_1$. We also have $Cover(\{g_1,o_2\}) \geq Cover(\{o_1 , o_2\})$ because $g_1$ covers at least the same weight of points as $o_1$ (since it is a disk that covers the maximum weight of points) and $g_1$ does not have any point in common with $o_2$. Thus $Cover(\{g_1 , g_2\}) \geq Cover(\{o_1, o_2\})$, which implies that the total number of the points covered by $g_1$ and $g_2$ is maximal.
Based on Lemma 1, we should look for the best two disks in the disks which have common points with $g_1$. These disks are located around $g_1$ in a circle having the same center as $g_1$ and radius $3$, where the radius of $g_1$ is 1. We call this region $Ng_1$. Figure 2 shows $Ng_1$. Lemma 2 bounds the number of points in $Ng_1$.
![The disks which have common points with $g_1$ are located in the dashed region around $g_1$. We call this region $Ng_1$[]{data-label="fig:fig2"}](Picture2){width=".5\columnwidth"}
![We can pack the dashed region with at most 21 unit disks[]{data-label="fig:fig3"}](Picture3){width=".5\columnwidth"}
\[lem:Lemma 2\] **Lemma 2**: Let $\rho$ be the maximum number of points that can be covered by a disk. The number of points in $Ng_1$ is $O(\rho)$.
We can pack $Ng_1$ with at most 21 unit disks, which is shown in Figure \[fig:fig3\]. Every one of these 21 disks covers equal or less than $ \rho$ points. So the number of points in $Ng_1$ is at most $21\rho $ points, which is $O(\rho)$ points.
\
We can also devise an output sensitive algorithm for $MostPoints(P,1)$ that solves the problem in $O(n\rho)$ time. The algorithm is as follows. Cover the plane using four shifted grids $G_1,...,G_4$. Each grid covers the whole plane, but the grids are shifted relative to each other, 2 units from each side. The cells of the grids have size 4x4. For any unit disk $d$ there is a grid $G_i$ such that $d$ is fully contained in a single cell of $G_i$. Now, for each non-empty grid cell $C$ of every grid, compute the optimal disk on the set of points lying inside $C$. $C$ contains at most $c\rho$ points, for a constant value of $c$ and the time spent for cell $C$ is $O(|C|^2)$ where $|C|$ denotes the number of points in the cell. The total number of points over all cells $C$ is $4n$, since each point is contained in exactly one cell per grid. To obtain the running time of the algorithm, we count the cells of the grids $C_1,...,C_l$. The total running time of the algorithm is:\
$\sum\limits_{i = 1}^l {|{C_i}{|^2}} \le \sum\limits_{i = 1}^l {|{C_i}| \times c\rho \le c\rho \times 4n = O(\rho n)} $\
Using Lemma 1 and Lemma 2, the best two disks can be $g_1$ and $g_2$, or the two disks around $g_1$ that cover the maximum number of points. To find the best two disks in $P$, first we find $g_1$ and $g_2$ in $O(n\rho)$ time. Then, we determine the points in $Ng_1$, and we run the algorithm of [@7] for the points located in $Ng_1$. This algorithm obtains the best two disks in $Ng_1$, in $O(\rho^3\log\rho)$ time. Finally, we compare this result, the best two disks in $Ng_1$, with $g_1$ and $g_2$ and consider the one that covers more weight as the optimal solution for $MostPoints(P,2)$. Since $\rho$ can be much smaller than $n$ in practice, our algorithm which runs in $O({n\rho} + {\rho ^3}\log \rho )$ time, is an improvement over the algorithm of [@7] that requires $O({n^3}\log n)$ time.
The algorithm for MostPoints(P,m)
=================================
The algorithm can be extended to cover the maximum number of points with $m$ disks. From now on, by the best $j$ disks, we mean a set of $j$ disks that cover the maximum total weight of points. The proposed algorithm is an iterative algorithm. At the first iteration, the algorithm finds two disks, which cover the maximum number of points as explained before. To find the best 3 disks, the algorithm uses the following fact, which will be proven in Lemma 3: the best 3 disks are among the disks that have some common points with the best 2 disks; Otherwise, the best 2 disks plus one disk that covers the maximum number of points, after removing the points located in the best two disks, are the best 3 disks. The algorithm repeats this process until the best $m$ disks are found. Now we explain the algorithm formally. Remember that the function $MostPoints(Q,j)$ returns a set of $j$ disks that cover the maximum number of points in the point set $Q$. Let $D$ be a set of unit disks. Define $Points(D) = \{ p|p \in (P \cap D)\} $ as the subset of $P$ that is located in the disks of $D$. Define $Neighbor(D)$ as the points in the circles with the same centers as the disks in $D$ and radius $3$. Let $OPT_i$ denotes the best $i$ disks. The outline of the algorithm is as follows:
*\
The exact output sensitive algorithm for the most points covering problem\
Inputs: A set P of n points in the plane, and a positive integer m.\
Outputs: A set of m disks which cover the maximum number of points from P.\
Definitions: The function MostPoints(Q,j), Points(D), and Neighbor(D) as defined in the text.*\
\
1 begin\
2 $OPT_0= \emptyset $\
3 $g_1= MostPoints(P,1) $\
4 $OPT_1 = g_1 $\
5 For $i=2$ to $m$\
6 $rp= P-Points(OPT_{i-1})$\
7 $g_i= MostPoints(rp,1)$\
8 $O_i= MostPoints(Neighbor(OPT_{i-1}), i)$\
9 If ( $Cover(OPT_{i-1} \cup g_i) > Cover(O_i)$ )\
10 $OPT_i= OPT_{i-1} \cup g_i$\
11 else\
12$OPT_i = O_i$\
\
13 return $OPT_m$\
14end\
In Lemma 3 we prove the correctness of the above algorithm. Before proving Lemma 3 we have the following definitions.\
**Definition 1**: Let $D$ and $E$ be two sets of disks. Define $ExculsiveCover(D,E)$ as the function which returns the number of points covered by $D$ after removing the points covered by $E$. By the definition of $ExculsiveCover(D,E)$ and $Cover(D)$ functions we have the following facts: $$ExculsiveCover\left( {D,E} \right){\rm{ }} \le Cover(D)$$ $$Cover(D \cup E) = Cover(D) + ExclusiveCover(E,D)$$ $$Cover(D \cup E) \le Cover(D) + Cover(E)$$
**Lemma 3**: The exact output sensitive algorithm for the most points covering problem obtains $m$ disks that cover the maximum number of points.
We prove the correctness of the algorithm by induction on the number of iterations. Assume that at the end of iteration $i$ we have correctly computed $OPT_i$. Consider $OPT_{i+1}$, which consists of $i+1$ disks. These disks may have common points with $Points(OPT_i)$ or not. We consider both cases. In case 1 we consider that all $i+1$ disks of $OPT_{i+1}$ have common points with $Points(OPT_i)$, and in case 2 we consider that some of these $i$ disks do not have any common point with $Points(OPT_i)$.\
\
Case1: All $i+1$ disks of $OPT_{i+1}$ have common points with $Points(OPT_i)$:\
\
All disks that have common points with $Points(OPT_i)$, have some points in $Neighbor(OPT_i)$. The algorithm, in line 8, obtains the best $i+1$ disks in the points of $Neighbor(OPT_{i})$. So in this case the algorithm exactly computes $OPT_{i+1}$.\
\
Case 2: Some of the disks in $OPT_{i+1}$ do not have any common point with $Points(OPT_i)$:\
\
Call one of these disks $F$. We can partition $OPT_{i+1}$ to $\{ OPT_{i + 1} - F\} $ and $F$. Based on the Definition 1, we have:\
\
$\begin{array}{l}
OPT_{i + 1} = \{ OPT_{i + 1} - F\} \cup \{ F\} \\
Cover(OPT_{i + 1}) = ExclusiveCover(OPT_{i + 1},F) \\
+ Cover(F) (1)
\end{array}$\
$\{ OP{T_{i + 1}} - F\} $ is a set of $i$ disks. As $OPT_i$ is a set of $i$ disks which cover the maximum number of points, we have:\
$Cover(OP{T_{i + 1}} - F) \le Cover(OP{T_i})$ (2)\
\
Furthermore, $g_{i+1}$ is the disk that covers the maximum number of points after removing the points in $Points(OPT_i)$. Therefore, $g_{i+1}$ covers the maximum number of points among the disks that have not common points with $OPT_i$. Thus, we have:\
$Cover(F)\; \le \;Cover({g_{i + 1}})$ (3)\
\
From (2) and (3) we have:\
\
$Cover(\{ OP{T_{i + 1}} - F\} ) + Cover(F)\; \le \;Cover(OP{T_i}) + Cover({g_{i + 1}})$ (4)\
\
From (1) and (4) we have:\
\
$Cover(OP{T_{i + 1}}) \le \;Cover(OP{T_{i - 1}}) + Cover({g_{i + 1}})$\
Thus in case 2, $OPT_i$ and $g_{i+1}$ cover the maximum number of points.\
The algorithm compares the results of case 1 with the result of case 2, and considers the best one as $OPT_{i+1}$. So, the algorithm correctly computes $OPT_{i+1}$. The base case, $i=1$, is correct and can be satisfied using the algorithm devised for $MostPoints(P,1)$.
In the following lemma, we obtain the running time of the algorithm.
**Lemma 4**: The exact output sensitive algorithm for the most points covering problem runs in $O(m{n\rho} + {(m\rho )^{2m - 1}}\log m\rho )$ time, where $\rho$ is the maximum number of points that one disk can cover in the points set $P$.
The algorithm runs in $m$ iterations. At the start of the iteration $i$, we have found the best $i-1$ disks, and we are getting to find the best $i$ disks. To find the best $i$ disks, we obtain $g_i$ in line 7, which takes $O(n\rho)$ time. We also look for the $i$ disks covering the maximum number of points in $Neighbor(OPT_{i-1})$, in line 8. $Neighbor(OPT_{i-1})$ consists of $i-1$ disks, and based on Lemma 2, the neighborhood of each disk can be packed with at most 21 disks. So the maximum number of points in $Neighbor(OPT_{i-1})$ is $21\rho(i - 1)$, which is $O(i\rho)$ points. We search for the $i$ disks that cover the maximum number of points in $O(i \rho )$ points. The algorithm of [@7] is applied for these points, which takes $O({(i \rho )^{2i - 1}}\log (i\rho ))$ time. The algorithm runs in $m$ iterations so, the running time is bounded by $O(n\rho) + \sum_{i=2}^{m} ( O(n\rho) + (i\rho)^{2i-1} \log (m\rho) )$ which is $O( mn\rho + (m\rho)^{2m-1} \log(m\rho))$.\
Implementation results
=======================
We have implemented our algorithm and the algorithm of [@7] for $MostPoints(P,2)$. In each algorithm, we observed the number of pairs of disks that each algorithms consider to obtain the optimal solution. We distributed different number of points randomly in squares with different side length and we run both algorithms on the points. Table 1 shows the result of the comparison. The results show that our algorithm computes significantly less number of pairs than the algorithm of [@7] in most cases. When $\frac{\rho }{n}$ becomes close to 1, the number of pair of disks that are considered in both algorithms become close to each other.
Extending the algorithm for other shapes and weighted point sets
================================================================
This algorithm can be applied for other types of maximum coverage problem such as covering point sets with other shapes and maximum coverage problem in sets. The algorithm uses the parameter $\rho$ which is the cardinality of the largest subset. Lemma 1 and 3 can be applied for general set systems. So we should search in the subsets that have common members with the largest subsets. To bound the running time of the algorithm there should be a constant number of subsets that have common points with the largest subsets.
Conclusion and future works
===========================
In this paper we presented an exact output sensitive algorithm for covering many points problem. In the algorithm, we confined the region in which the resultant disks may reside using the greedy algorithm, and we searched in that regions only. This led to an output sensitive algorithm for this problem which runs in $O(m{n\rho} + {(m\rho )^{2m - 1}}\log m\rho )$ time, where $\rho$ is the maximum number of points that one disk can cover. Our algorithm can be improved by improving the running time of the algorithm in [@7]. If better algorithms for covering most points with two disks are suggested, they can improve the running time of our algorithm and the algorithm of [@7]. Furthermore, considering $\rho$ as a parameter, and using our technique, it may be possible to improve the analysis of other covering algorithms.
Acknowledgement
===============
We would like to acknowledge anonymous referees for their helpful comments and the improvements they suggested for the paper.
--------------- ------------- -------- ----------------------------------- -------------------------------
Num of points square side $\rho$ Num of pairs in algorithm of [@7] Num of pairs in our algorithm
\[.5ex\] 1000 200 19 7007 201
1000 100 44 22878 1131
1000 50 125 45414 4259
500 50 105 26041 6402
500 100 34 8605 252
--------------- ------------- -------- ----------------------------------- -------------------------------
: Number of pairs of disks processed in our algorithm an algorithm of \[7\]
\[table: Table 1\]
[99]{}
Z. Drezner On a modified one-center model. , 27(7):848-851, 1981.
B. M. Chazelle and D. T. Lee On a circle placement problem. , 36(1):1-16, 1986.
A. Gajentaan and M. H. Overmars On a class of $O(n^2)$ problems in computational geometry. , 5(3):165-185, 1995.
B. Aronov and S. Har-Peled On approximating the depth and related problems. , 886-894, 2005.
C. M. H. de Figueiredo and G. D. da Fonseca Enclosing weighted points with an almost-unit ball , 109(21-22):1216-1221, 2009.
D. S. Hochbaum and A. Pathria Analysis of the greedy approach in problems of maximum k-coverage. , 45(6):615-627, 1998.
M. de Berg, S. Cabello, and S. Har-Peled Covering many or few points with unit disks. , 45(3):446-469, 2009.
H. Ghasemalizadeh and M. Razzazi An Improved Approximation Algorithm for the Most Points Covering Problem. , 50(3): 545-558, 2012.
[^1]: Email:
[^2]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the effect of radiative feedback on accretion onto intermediate mass black holes (IMBHs) using the hydrodynamical code ZEUS-MP with a radiative transfer algorithm. In this paper, the first of a series, we assume accretion from a uniformly dense gas with zero angular momentum and extremely low metallicity. Our 1D and 2D simulations explore how X-ray and UV radiation emitted near the black hole regulates the gas supply from large scales. Both 1D and 2D simulations show similar accretion rate and period between peaks in accretion, meaning that the hydro-instabilities that develop in 2D simulations do not affect the mean flow properties. We present a suite of simulations exploring accretion across a large parameter space, including different radiative efficiencies and radiation spectra, black hole masses, density and temperature, $T_\infty$, of the neighboring gas. In agreement with previous studies we find regular oscillatory behavior of the accretion rate, with duty cycle $\sim 6\%$, mean accretion rate 3% $(T_{\infty}/10^4~{\rm
K})^{2.5}$ of the Bondi rate and peak accretion $\sim 10$ times the mean for $T_{\infty}$ ranging between $3000~$K and $15000~$K. We derive parametric formulas for the period between bursts, the mean accretion rate and the peak luminosity of the bursts and thus provide a formulation of how feedback regulated accretion operates. The temperature profile of the hot ionized gas is crucial in determining the accretion rate, while the period of the bursts is proportional to the mean size of the Strömgren sphere and we find qualitatively different modes of accretion in the high vs. low density regimes. We also find that softer spectrum of radiation produces higher mean accretion rate. However, it is still unclear what is the effect of a significant time delay between the accretion rate at our inner boundary and the output luminosity. Such a delay is expected in realistic cases with non-zero angular momentum and may affect the time-dependent phenomenology presented here. This study is a first step to model the growth of seed black holes in the early universe and to make a prediction of the number and the luminosity of ultra-luminous X-ray sources in galaxies produced by IMBHs accreting from the interstellar medium.
author:
- KwangHo Park
- Massimo Ricotti
bibliography:
- 'pr10\_revIII.bib'
title: |
Accretion onto Intermediate Mass Black Holes Regulated by Radiative Feedback.I\
Parametric Study for Spherically Symmetric Accretion
---
INTRODUCTION
============
The occurrence of gas accretion onto compact gravitating sources is ubiquitous in the universe. The Bondi accretion formula [@BondiH:44; @Bondi:52], despite the simplifying assumption of spherical symmetry, provides a fundamental tool for understanding the basic physics of the accretion process. Angular momentum of accreted gas, in nearly all realistic cases, leads to the formation of an accretion disk on scales comparable to or possibly much greater than the gravitational radius of the black hole, $r_g \sim GM/c^2$, thus breaking the assumption of spherical symmetry in the Bondi solution. However, the fueling of the disk from scales larger than the circularization radius $r_c \sim j^2/GM$, where $j$ is the gas specific angular momentum, can be approximated by a quasi-radial inflow. Thus, assuming that numerical simulations resolve the sonic radius, $r_s$, the resolved gas flow is quasi-spherical if $r_c \ll
r_s$. The Bondi formula, which links the accretion rate to the properties of the environment, such as the gas density and temperature, or Eddington-limited rate are often used in cosmological simulations to model the supply of gas to the accretion disk from galactic scales [@Volonteri:05; @DiMatteo:08; @Pelupessy:07; @Greif:08; @AlvarezWA:09].
However, the Bondi formula is a crude estimation of the rate of gas supply to the accretion disk because it does not take into account the effect of accretion feedback loops on the surrounding environment. Radiation emitted by black holes originates from gravitational potential energy of inflowing gas [@Shapiro:73] and a substantial amount of work has been performed to understand the simplest case of spherical accretion onto compact X-ray sources or quasars. Several authors have used hydrodynamical simulations to explore how feedback loops operates and whether they produce time-dependent or a steady accretion flows. A variety of feedback processes have been considered: X-ray preheating, gas cooling, photo-heating and radiation pressure [@OstrikerWYM:76; @CowieOS:78; @BB:80; @KrolikL:83; @Vitello:84; @WandelYM:84; @MiloBCO:09; @NovakOC:10; @OstrikerCCNP:10]. Typically, the dominance of one process over the others depends on the black hole mass and the properties of the gas accreted by the black hole. The qualitative description of the problem is simple: gravitational potential energy is converted into other forms of energy such as UV and X-ray photons or jets, which act to reduce and reverse the gas inflow, either by heating the gas or by producing momentum driven outflows [@CiottiO:07; @CiottiOP:09; @Proga:07; @ProgaOK:08]. In general these feedback processes reduce the accretion rate and thus the luminosity of the accreting black hole [@OstrikerWYM:76; @Begelman:85; @Ricotti:08]. Consequently, the time averaged accretion rate differs from Bondi’s solution. There have been works on self-regulated accretion of supermassive black holes (SMBHs) at the center of elliptical galaxies [@Sazonov:05; @CiottiO:07; @CiottiOP:09] and radiation-driven axisymmetric outflow in active galactic nuclei [@Proga:07; @ProgaOK:08; @KurosawaPN:09; @KurosawaP:09a; @KurosawaP:09b]. However, far less has been done in order to quantify the simplest case of spherical accretion onto IMBHs as a function of the properties of the environment. Recent theoretical [@MiloBCO:09 hereafter MBCO09] and numerical [@MiloCB:09 hereafter MCB09] works explore accretion of protogalactic gas onto IMBHs in the first galaxies. MCB09 describes the accretion onto a 100 M$_\odot$ black hole from protogalactic gas of density $n_{H,\infty}=10^7$ cm$^{-3}$ and temperature $T_\infty =10^4$ K. Our study, which complements this recent numerical work, is a broader investigation of accretion onto IMBHs for a set of several simulations with a wide range of radiative efficiencies, black hole masses, densities and sound speeds of the ambient gas. Our aim is to use simulations to provide a physically motivated description of how radiation modifies the Bondi solution and provide an analytical formulation of the problem (see MBCO09).
The results of the present study will help to better understand the accretion luminosities of IMBHs at high z and in the present-day universe [@Ricotti:09]. Applications of this work include studies on the origin of ultra-luminous X-ray sources (ULXs) [@Krolik:81; @Krolik:84; @Ricotti:07; @MackOR:07; @Ricotti:08], the build up of an early X-ray background [@VenkatesanGS:01; @RicottiO:04; @MadauV:04; @RicottiOG:05] and growth of SMBHs [@VolonteriHM:03; @Volonteri:05; @JohnsonB:07; @Pelupessy:07; @AlvarezWA:09]. For example, different scenarios have been proposed for the formation of quasars at $z \sim 6$ [@Fan:03]: growth by mergers, accretion onto IMBHs, or direct formation of larger seed black holes from collapse of quasi-stars [@Carr:84; @HaehneltNR:98; @Fryer:01; @BegelmanVR:06; @VolonteriLN:08; @OmukaiSH:08; @ReganH:09; @MayerKEC:10] that may form from metal free gas at the center of rare dark matter halos [@OhH:02]. Understanding the properties which determine the efficiency of self-regulated accretion onto IMBHs is important to estimate whether primordial black holes produced by Pop III stars can accrete fast enough to become SMBHs by redshift $z \sim 6$ [@MadauR:01; @VolonteriHM:03; @YooM:04; @Volonteri:05; @JohnsonB:07; @Pelupessy:07; @AlvarezWA:09].
In this paper we focus on simulating accretion onto IMBH regulated by photo-heating feedback in 1D and 2D hydrodynamic simulations, assuming spherically symmetric initial conditions. We provide fitting formulas for the mean and peak accretion rates, and the period between accretion rate bursts as a function of the parameters we explore, including radiative efficiency, black hole mass, gas density, temperature and spectrum of radiation. In § 2 we introduce basic concepts and definitions in the problem. Numerical procedures and physical processes included in the simulations are discussed in § 3. Our simulation results and the parameter study are shown in § 4. In § 5 we lay out a physically motivated model that describes the results of the simulations. Finally, a summary and discussion are given in § 6.
BASIC DEFINITIONS
=================
Bondi Accretion and Eddington luminosity
----------------------------------------
The assumption of spherical symmetry allows to treat accretion problems analytically. The solution [@Bondi:52] provides the typical length scale at which gravity affects gas dynamics and the typical accretion rate as a function of the black hole mass $M_{bh}$, ambient gas density $\rho_\infty$ and sound speed $c_{s,\infty}$. The Bondi accretion rate for a black hole at rest is $$\begin{aligned}
\dot{M}_B &=& 4\pi\lambda_{B} r_b^2 \rho_{\infty} c_{s,\infty} \nonumber\\
& = & 4 \pi \lambda_{B} \rho_\infty \frac{G^2 M_{bh}^2 }{c_{s,\infty}^3}, \end{aligned}$$ where $r_b = GM_{bh}c_{s,\infty}^{-2}$ is the Bondi radius, and $\lambda_B$ is the dimensionless mass accretion rate, which depends on the polytropic index, $\gamma$, of the gas equation of state $P=K\rho^\gamma$: $$\begin{aligned}
\lambda_{B} &=& \frac{1}{4} \left[ \frac{2}{5-3\gamma} \right]^{\frac{5-3\gamma}{2(\gamma -1)}}.\end{aligned}$$ The value of $\lambda_{B}$ ranges from $e^{3/2}/4 \simeq 1.12$ for an isothermal gas ($\gamma = 1$ ) to $1/4$ for an adiabatic gas ($\gamma=5/3$).
However, a fraction of the gravitational potential energy of the inflowing gas is necessarily converted into radiation or mechanical energy when it approaches the black hole, significantly affecting the accretion process. Photons emitted near the black hole heat and ionize nearby gas, creating a hot bubble which exerts pressure on the inflowing gas. Radiation pressure may also be important in reducing the rate of gas inflow (see MBCO09). These processes may act as self-regulating mechanisms limiting gas supply to the disk from larger scales and, thus, controlling the luminosity of the black hole. We quantify the reduction of the accretion rate with respect to the case without radiative feedback by defining the dimensionless accretion rate $$\begin{aligned}
\lambda_{rad} \equiv \frac{\dot{M}}{\dot{M}_B}, \end{aligned}$$ where $\dot{M}_B$ is the Bondi accretion rate for isothermal gas ( $\dot{M}_B = e^{3/2} \pi G^2 M_{bh}^2 \rho_{\infty}
c_{s,\infty}^{-3}$). This definition of $\lambda_{rad}$ is consistent with the one adopted by MBCO09.
Luminosity and Radiative efficiency
-----------------------------------
The Eddington luminosity sets an upper limit on the luminosity of a black hole. In this limit the inward gravitational force on the gas equals the radiation pressure from photons interacting with electrons via Compton scattering. Although this limit can be evaded in some special cases, observations suggest that black hole and SMBH luminosity is sub-Eddington. The Eddington luminosity is thus, $$\begin{aligned}
L_{Edd} &=& \frac{4\pi GM_{bh}m_{p}c}{\sigma_{T}} \simeq 3.3\times
10^6 L_{\sun} \left( \frac{M_{bh}}{100~M_{\sun}} \right).\end{aligned}$$ The luminosity of an accreting black hole is related to the accretion rate via the radiative efficiency $\eta$: $L =\eta \dot{M}c^2$. From the Eddington luminosity, we define the Eddington gas accretion rate $\dot{M}_{Edd} \equiv L_{Edd}c^{-2} $, and the dimensionless accretion rate and luminosity as $$\begin{aligned}
\dot{m} \equiv \frac{\dot{M}}{\dot{M}_{Edd}}~~~ {\rm and}~~~
l \equiv \frac{L}{L_{Edd}}.\end{aligned}$$ Hence, in dimensionless units, the bolometric luminosity of the black hole is $l=\eta \dot{m}$, where $\dot{m}$ is the accretion rate onto the black hole. Note, that our definition of $\dot{M}_{Edd}$ is independent of the radiative efficiency $\eta$. Therefore, if we impose sub-Eddington luminosity of the black hole, the dimensionless accretion rate ranges between $0 < \dot{m} \leq \frac{1}{\eta} $. The radiative efficiency, $\eta$, depends on the geometry of the accretion disk and on $\dot{m}$. For a thin disk, $\eta \simeq 0.1$, whereas $\eta \propto \dot{m}$ for an advection dominated thick disk or for spherical accretion [@Shapiro:73; @Park:01]. In this study we consider two idealized cases for the radiative efficiency. The case of constant radiative efficiency $\eta=const$; and the case in which the radiative efficiency has a dependence on the dimensionless accretion rate and luminosity: $\eta=const$ for $l \ge 0.1$ and $\eta \propto
\dot{m}$ for $l < 0.1$. The second case we explored accounts for the lower radiative efficiency expected when the accreted gas does not settle into a thin disk. In both formulations the the radiative efficiency is one of the free parameter we allow to vary and we do not find important differences between the two cases. Observations of Sgr A\*, the best studied case of low accretion rate onto a SMBH, suggest that the radiative efficiency is indeed low but not as low as implied by the scaling $\eta \propto \dot{m}$. Recent theoretical work by [@Sharma:07] demonstrates that there is indeed a floor on the radiative efficiency.
Because the Bondi rate, $\dot{M}_B$ does not include radiation feedback effect, it provides an upper limit on the accretion rate from large scales to radii near the black hole. The Eddington rate provides the maximum accretion rate onto the black hole, limited by radiation feedback at small radii. Thus, numerical simulations are necessary to obtain realistic estimates of the accretion rates. If the accretion rate onto the black hole is lower than the gas accretion from large scales, the accreted material accumulates near the black hole, creating a disk whose mass grows with time. We cannot simulate such a scenario because it is too computationally challenging to resolve a range of scales from the Bondi radius to the accretion disk in the same simulations. Here we assume that accretion onto the black hole is not limited by physical processes taking place on radial distances much smaller than the sonic radius. For instance, even if angular momentum of accreted gas is small and the circularization radius $r_c\ll r_s$, further inflow will be slowed down with respect to the free-fall rate. The rate of inflow will be controlled by angular momentum loss (e.g. torques due to MHD turbulence) and there will be a delay between the accretion rate at the inner boundary of our simulation ($r_{min}$) and the accretion luminosity associated with it. The effect of the aforementioned time delay on the feedback loop is not considered in this paper but will be considered in future works. We also assume that the effect of self-gravity is negligible in our simulations since we have estimate that the mass within the HII region around the black hole is smaller than the black hole mass for $M_{bh}< 1000$ M$_\odot$.
If the rate of gas supply to the disk is given by the Bondi rate, accretion onto the black hole is sub-Eddington for black hole masses $$M_{bh} < {c_{s,\infty}^3 \over G n_{H,\infty} \sigma_T c \eta} \sim ~
40M_\odot~ T_{\infty,4}^{1.5} n_{H,5}^{-1} \eta_{-1}^{-1},
\label{eq:edd}$$ where we use the notations of $T_{\infty,4} \equiv T_{\infty}/(10^4~{\rm K})$, $n_{H,5}\equiv n_{H,\infty}/(10^5$ cm$^{-3})$ and $\eta_{-1}\equiv\eta/10^{-1}$. Thus, in this regime we may assume that the accretion is quasi-steady in the sense that the mean accretion rate onto the black hole equals the gas supply from large scales when the accretion rate is averaged over a sufficiently long time scale.
NUMERICAL SIMULATIONS
=====================
ZEUS-MP and Radiative Transfer Module
-------------------------------------
We perform a set of hydrodynamic simulations to understand accretion onto IMBHs regulated by radiation feedback. Numerical simulations of radiative feedback by black holes are challenging because they involve resolving a large dynamical range in length scales. In this study we use ZEUS-MP [@Hayes:06], a modified parallel version of the non-relativistic hydrodynamics code ZEUS [@StoneN:92]. For the present work we add a radiative transfer module [@RicottiGS:01] to ZEUS-MP to simulate radiative transfer of UV and X-ray ionizing photons emitted near the black hole. A detailed description of the numerical methods used to solve radiative transfer and tests of the code are presented in the Appendix.
As X-ray and UV photons ionize the surrounding medium, different reactions take place depending on the density and composition of the gas. Photo-ionization changes the ionization fraction of H and He. The detailed evolution of the Strömgren sphere depends on the cooling function $\Lambda(T,Z)$ of the gas and thus on the metallicity, $Z$, and the fraction of gas in the molecular phase. For a gas of primordial composition, the cooling rate depends on the formation rates of $H^{-}$ and $H_2$, which depend on both the redshift and the intensity of the local dissociating background in the $H_2$ Lyman-Werner bands [e.g. @ShapiroK:87; @AbelAnninos:98; @RicottiGS:02a; @RicottiGS:02b]. In addition, the cooling function may depend on redshift due to Compton cooling of the electrons by CMB photons. We adopt atomic hydrogen cooling for temperatures $T>10^4$ K, and use a simple parametric function to model complicated cooling physics of gas at $T <
10^4~$K. Thus, the temperature structure inside the ionized bubble is appropriate only for a low metallicity gas. For a subset of simulations we also include the effect of helium photo-heating and cooling. We assume that gas cooling at temperatures below $T_{\infty}$ is negligible in order to achieve thermal equilibrium in the initial conditions far from the black hole. For the parameter space in which we can neglect the effect of radiation pressure we find (see § 5.1) that the accretion rate is a function of the temperature both outside and inside the HII region. The temperature outside the HII region depends on the cooling function of gas at $T < 10^4~$K and on the heating sources. The temperature inside the HII region depends on the spectrum of radiation and cooling mechanism of gas at $T >
10^4$ K. Thus, it depends on the gas metallicity and the redshift at which Compton cooling might become important. However, for the parameter space we have explored we find that Compton cooling has a minor effect on the temperature inside and outside the Strömgren sphere.
The gas heating rate depends on flux and spectral energy distribution (SED) of the radiation emitted near the black hole. We assume a luminosity of the black hole $l=\eta \dot{m}$ (see § 2.2), where $\dot{m}$ is calculated at the inner boundary in our simulation (typically $r_{min}\sim 3\times10^{-5}$ pc). We adopt a single power law $\nu^{-\alpha}$ for the SED, where the spectral index $\alpha$ is one of the parameters we vary in our set of simulations.
We use an operator-split method to calculate the hydrodynamic step and the radiative transfer and chemistry steps. The hydrodynamic calculation is done using ZEUS-MP, then for the radiative transfer calculation we use a ray tracing module [@RicottiGS:01]. The radiative transfer module calculates chemistry, cooling and heating processes along rays outgoing from the central black hole, and thus is easily parallelized in the polar angle direction.
We perform 1D and 2D simulations in spherical coordinates. In both cases we use a logarithmically spaced grid in the radial direction typically with 256 to 512 cells to achieve high resolution near the black hole. The size ratio between consecutive grids is chosen according to the free parameters of the simulation to resolve the ionization front and resolve the region where the gas is in free fall. In the 2D simulations we use evenly spaced grids in the polar angle direction and compute radiative transfer solutions in each direction. Flow-out inner boundary conditions and flow-in outer boundary conditions are used in the radial direction ($r$), whereas in polar angle directions ($\theta$), reflective boundary conditions are used.
To determine the optimal box size of the simulations we make sure that we resolve important length scales in the problem: the inner Bondi radius, $r_{b,in}$, the outer Bondi radius, $r_{b,\infty}$, the sonic radius, $r_s$ and the ionization front, $R_s$. We select the value of the inner boundary (typically ${\sim} 3\times 10^{-5}$ pc for $M_{bh}=100~$M$_{\sun}$) to be smaller than the sonic point or the inner Bondi radius (both still far larger than the Schwarzschild radius of the black holes). We find that once the sonic radius is resolved, reducing the inner boundary box size does not create significant differences in the results. In most cases the ionization front is located outside of the outer Bondi radius and the box size is selected to be large enough to cover both length scales. We select a box size that achieves the highest possible resolution with a given number of grids, making sure that the physical quantities around boundaries remain constant during the simulations. The box is sufficiently large to minimize the effect of spurious wave reflections at the outer boundary.
In this paper, the first of a series, we adopt idealized initial conditions of uniform density and temperature, zero velocity and zero angular momentum of the gas relative to the black hole. In future work we will relax some of these assumptions by adding turbulence in the initial condition and considering the effect of black hole motion with respect to the ambient medium and considering the effect of a time-delay between the accretion rate at the inner boundary of our simulations and the accretion luminosity. We assume monatomic, non-relativistic ideal gas with $\gamma = 5/3$ which is initially neutral (electron fraction $x_e \sim 10^{-5}$). In this paper we also neglect the effect of radiation pressure. Our goal is to add to the simulations one physical process at a time to understand which feedback loop is dominant in a given subset of the parameter space. We take this approach to attempt an interpretation of the simulation results in the context of a physically motivated analytical description of the accretion cycle. We will explore the effect of radiation pressure due to HI ionization and Lyman-alpha scattering in future works. However, a simple inspection of the relevant equations suggests that radiation pressure is increasingly important for large values of the ambient gas density ($n_{H,\infty} \sim 10^7$ cm$^{-3}$, see MBCO09 and § 6) since accretion rate approaches Eddington limit.
RESULTS
=======
Qualitative description of accretion regulated by radiative feedback
--------------------------------------------------------------------
Our simulations show that UV and X-ray photons modify the thermal and dynamical structure of the gas in the vicinity of the Bondi radius. A hot bubble of gas is formed due to photo-heating by high energy photons and sharp changes of physical properties such as density, temperature, and ionization fraction occur at the ionization front. Figure \[evolution\] shows 8 snapshots from one of our 2D simulations. Top half of each snapshot shows the gas density and the bottom half shows the hydrogen ionization fraction. We show the periodic oscillation of the density and the ionization fraction from a 2D simulation in Figure \[evolution\]. The time evolution of the density, temperature and ionization fraction profiles for the 1D simulation are shown in Figure \[profile\]. We can identify 3 evolutionary phases that repeat cyclically:
1\) Once the Strömgren sphere is formed, it expands and the gas density inside that hot bubble decreases maintaining roughly pressure equilibrium across the ionization front. At the front, gas inflow is stopped by the hot gas and the average gas density inside the bubble decreases due to the following two physical processes. First, the black hole continues accreting hot gas within an accretion radius, $r_{acc}$, defined as the radius where the gravitational force of the black hole dominates the thermal energy of the hot gas. The accretion radius is similar to the Bondi radius defined by the temperature inside Strömgren sphere, but there exists a difference between them since the kinematic and thermal structure of gas is modified significantly by the photo-heating and cooling. Second, the gas between $r_{acc}$ and the ionization front moves towards the ionization front due to pressure gradients. The left panel of Figure \[vel\] shows inflowing gas within $r_{acc}$ and outflowing gas outside $r_{acc}$. A dense shell forms just outside the ionization front. Thus, the mass of the shell grows because gravity pulls distant gas into the system at the same time that gas within the the hot bubble is pushed outwards.
2\) As the average density inside the hot bubble decreases, the accretion rate diminishes. During this process the radius of the Strömgren sphere remains approximately constant since the reduced number of ionizing UV and X-ray photons is still sufficient to ionize the rarefied hot bubble. Figure \[ifront\] illustrates this. Thus, the average gas temperature, ionization fraction and the size of the HII region remain constant. As the accretion rate increases during the burst, it produces a rapid expansion of the Strömgren sphere radius. During one cycle of oscillation, there are small peaks in the Strömgren sphere radius which are associated with minor increases in the accretion rate. Rayleigh-Taylor (RT) instabilities develop quickly when the accretion rate increases. In these phases, the acceleration of the dense shell is directed toward the black hole, so the dense shell, supported by more rarefied gas, becomes RT unstable.
3\) As gas depletion continues, the pressure inside the hot bubble decreases to the point where equilibrium at the ionization front breaks down. The outward pressure exerted by the hot bubble becomes too weak to support the gravitational force exerted on the dense shell. The dense shell of gas collapses toward the black hole, increasing dramatically the accretion rate and creating a burst of ionizing photons. The ionization front propagates outward in a spherically symmetric manner, creating a large Strömgren sphere and returning to the state where the high pressure inside the Strömgren sphere suppresses gas inflow from outside.
Comparison of 1D and 2D simulations
-----------------------------------
In agreement with previous studies, our simulations show that radiation feedback induces regular oscillations of the accretion rate onto IMBH. This result is in good agreement with numerical work by MCB09 for accretion onto a $100~$M$_\sun$ black hole from a high density ($n_{H,\infty}=10^7$ cm$^{-3}$) and high temperature ($T_{\infty}=10^4$ K) gas. Periodic oscillatory behavior is found in all our simulations for different combinations of parameters, when assuming spherically symmetric initial conditions and a stationary black hole. This oscillation pattern is quite regular and no sign of damping is observed for at least $\sim 10$ cycles.
For the same parameters, our 1D and 2D simulations are nearly identical in terms of oscillatory behavior in accretion rate and Strömgren sphere size. Figure \[1d2d\] shows accretion rate in 1D and 2D simulations for $M_{bh}$=100 M$_\sun$, $T_{\infty}=10^4$ K and $n_{H,\infty}=10^5$ cm$^{-3}$. Note the similar pattern in accretion rate and period between bursts. This indicates that the 1D result adequately represents 2D cases when the accretion flow does not have significant angular momentum.
Moreover, this result demonstrates that RT instabilities which we observe in the 2D simulations do not affect the mean accretion rate or the period of oscillations. The RT instability develops during the phase when the dense shell in front of the ionization front is supported against gravitational accretion by the low density medium inside the hot bubble [@Whalen:08a; @Whalen:08b]. The top panels in Figure \[evolution\] show small instabilities when ionization fronts move outward, which largely decay over time. The pressure gradient inside the Strömgren sphere creates an outward force which helps suppress the development of the instability.
In summary, we believe that 1D simulations can be used in place of higher dimension simulations to determine the cycle and magnitude of the periodic burst of gas accretion onto IMBH. This allows us to reduce the computational time required to explore a large range of parameter space.
Parameter space exploration
----------------------------
In this section we present the results of a set of 1D simulations aimed at exploring the dependence of the accretion rate and the period of oscillations of the black hole luminosity as a function of the black hole mass, $M_{bh}$, the ambient gas density, $n_{H,\infty}$, temperature, $T_\infty$, and the radiative efficiency $\eta$. In $\S~$5 we present results in which we allow the spectrum of ionizing radiation to vary as well. The accretion can be described by three main parameters: $\tau_{cycle}$, the mean period between bursts, $\lambda_{rad,max}$, the maximum value of the dimensionless accretion rate (at the peak of the burst), and $\langle \lambda_{rad} \rangle$, the time-averaged dimensionless accretion rate. These parameters are typically calculated as the mean over $\sim$ 5 oscillation cycles and the error bars represent the standard deviation of the measurements.
After reaching the peak, the luminosity decreases nearly exponentially on a time scale $\tau_{on}$, that we identify as the duration of the burst. Both $\tau_{on}$ and the duty cycle, $f_{duty}$, of the black hole activity ([*i.e.,*]{} the fraction of time the black hole is active), can be expressed as a function of $\tau_{cycle}$, $\lambda_{rad,max}$ and $\langle \lambda_{rad} \rangle$: $$\begin{aligned}
\tau_{on} &\equiv& \frac{\langle \lambda_{rad}
\rangle}{\lambda_{rad,max}} \tau_{cycle} , \\
f_{duty} &\equiv& \frac{\tau_{on}}{\tau_{cycle}} = \frac{\langle \lambda_{rad}
\rangle}{\lambda_{rad,max}}.
\label{eq:fduty} \end{aligned}$$ The values of $\lambda_{rad,max}$ and $f_{duty}$ as a function of the black hole mass, the density and the temperature of the ambient medium are important for estimating the possibility of detection of IMBHs in the local universe because these values provide an estimate of the maximum luminosity and the number of active sources in the local universe at any time. On the other hand, the mean accretion rate, $\langle \lambda_{rad}
\rangle$ is of critical importance for estimating IMBH growth rate in the early universe.
The four panels in Figure \[para\] summarize the results of a set of simulations in which we vary the free parameters one at a time. We find that, in most of the parameter space that we have explored, the period of the oscillations and the accretion rates are described by a single or a split power law with slope $\beta$. In the following paragraphs we report the values of $\beta$ derived from weighted least squares fitting of the simulation results. The weight is $1/\sigma$ where $\sigma$ is the standard deviation of $\langle \lambda_{rad}
\rangle$ or $\lambda_{rad,max}$ over several oscillations.
[*a) Dependence on the radiative efficiency*]{}\
First, we explore how the accretion depends on the radiative efficiency $\eta$. This parameter describes the fraction, $\eta$, of the accreting rest mass energy converted into radiation while the remaining fraction, $1-\eta$, is added to the black hole mass. We have explored both constant values of the radiative efficiency and the case $\eta \propto {\dot m}$ for $l < 0.1$ (see § 2.2). The simulation results shown in this section are obtained assuming $\eta$ is constant. We find similar results for $\langle
\lambda_{rad} \rangle$, $\lambda_{rad,max}$ and $\tau_{cycle}$ when we assume $\eta \propto \dot{m}$. The radiative efficiency for a thin disk is about $10\%$. Here, we vary $\eta$ in the range: $0.2\%$ to $10\%$. The other free parameters are kept constant with values $n_{H,\infty}=10^5$ cm$^{-3}$, $M_{bh}=100$ M$_{\sun}$ and $T_{\infty}=10^4$ K. Figure \[eta1\] shows the accretion rate as a function of time for different values of the radiative efficiency: $\eta= 0.1, 0.03, 0.01$ and $0.003$. Panel (a) in Figure \[para\] shows the dependence on $\eta$ of the three parameters that characterize the accretion cycle. The maximum accretion rate increases mildly with increasing $\eta$ (log slope $\beta=0.13 \pm 0.06$). The average accretion rate is $\langle \lambda_{rad} \rangle \sim 2.9\%
\pm 0.2\% $, is nearly independent of $\eta$ ($\beta
=-0.04\pm0.01$). The period of the oscillations increases with $\eta$ as $\tau_{cycle} \propto \eta^{1/3}$. We also show the simulation results including helium photo-heating and cooling, shown as open symbols in the same panel of Figure \[para\]. We find that including helium does not change the qualitative description of the results, but does offset the mean accretion rate, that is $\sim 41 \%$ lower and the period of the accretion bursts, that is $ \sim 42\%$ shorter. This offset of the accretion rate and period with respect to the case without helium is due to the higher temperature of the gas inside the HII region surrounding the black hole.
[*b) Dependence on black hole mass*]{}\
We explore a range in black hole mass from 100 M$_{\sun}$ to 800 M$_{\sun}$, while keeping the other parameters constant ($\eta =
0.1$, $n_{H,\infty}=10^5$ cm$^{-3}$ and $T_{\infty}= 10^4$ K). The results are shown in panel (b) of Figure \[para\]. The mean accretion rate is $\langle \lambda_{rad} \rangle \sim 2.7\% \pm 0.4\%$ and the maximum accretion rate is $\lambda_{rad,max} \sim 42\% \pm
12\% $ ($\beta = -0.26 \pm 0.20$). They are both independent of $M_{bh}$ within the error of the fit. The period of the bursts is well described by a power-law relation $\tau_{cycle} \propto M_{bh}^{2/3}$.
[*c) Dependence on gas density of the ambient medium*]{}\
Panel (c) in Figure \[para\] shows the dependence of accretion rate and burst period on the ambient gas density, $n_{H,\infty}$. We explore a range of $n_{H,\infty}$ from $5\times 10^3$ cm$^{-3}$ to $10^7$ cm$^{-3}$, while keeping the other parameters constant at $\eta = 0.1$, $M_{bh}=100$ M$_{\sun}$ and $T_{\infty}= 10^4$ K. For densities $n_{H,\infty} \ge 10^5$ cm$^{-3}$, $\langle \lambda_{rad}
\rangle$ and $\lambda_{rad, max}$ are insensitive to $n_{H,\infty}$ ($\beta = -0.04\pm0.08$ and $\beta=-0.18\pm0.13$, respectively). However, for $n_{H,\infty} \le 10^5$ cm$^{-3}$, $\langle
\lambda_{rad}\rangle$ and $\lambda_{rad, max}$ are proportional to $n_{H,\infty}^{1/2}$ ($\beta = 0.44 \pm 0.02$ and $\beta = 0.37 \pm
0.09$, respectively).
The bottom of Figure \[para\](c) shows the effect of density in determining the oscillation period. For densities $n_{H,\infty} \ge
10^5$ cm$^{-3}$, $\tau_{cycle}$ is fitted well by a power law with $\tau_{cycle} \propto n_{H,\infty}^{-1/3}$ and for the densities $n_{H,\infty} \le 10^5$ cm$^{-3}$ it is fitted well by a power law $\tau_{cycle} \propto n_{H,\infty}^{-1/6}$. However, $\tau_{cycle}$ at $n_{H,\infty}=10^7~$cm$^{-3}$ is lower than predicted by the power law fit for $n_{H,\infty} \ge 10^5~$cm$^{-3}$. Although Figure \[1d2d\], \[eta1\] do not show clearly the magnitude of accretion rate during the inactive phase, it is evident in a log-log plot that accretion rate at minima is 4 orders of magnitude lower than during the peak of the burst. This is the case for all simulations but the ones with $n_{H,\infty}=10^7~$cm$^{-3}$ in which the accretion rate at minima is 2 orders of magnitude higher than in all other simulations. The simulations show that the ambient gas density is an important parameter in determining the accretion luminosity and period between bursts of the IMBH. One of the reasons is that the gas temperature inside the hot ionized bubble and the thickness and density of the dense shell in front of it depend on the density via the cooling function. The drop in the accretion rate we observe at low densities can be linked to an increase of the temperature within the sonic radius with respect to simulations with higher ambient density. This results in an increase in the pressure gradient within the ionized bubble that reduces the accretion rate significantly.
[*d) Dependence on the temperature of the ambient medium*]{}\
Panel (d) in Figure \[para\] shows the dependence of accretion rate and period of the bursts on the temperature of the ambient medium, $T_\infty$. We vary $T_{\infty}$ from $3000$ K to $15000$ K while keeping the other parameters constant at $\eta = 0.1$, $M_{bh}=100$ M$_{\sun}$ and $n_{H,\infty}= 10^5$ cm$^{-3}$. We find $\langle \lambda_{rad} \rangle$ and $\lambda_{rad,max}$ depend steeply on $T_\infty$ as $T_{\infty}^{5/2}$ ($\beta=2.44 \pm 0.06$). Except for the simulation with $T_\infty=3000$ K, the period of the accretion cycle is fitted well by a single power law $\tau_{cycle}\propto
T_{\infty}^{-1/2}$.
Analytical Formulation of Bondi Accretion with Radiative Feedback
=================================================================
In this section we use the fitting formulas for $\langle \lambda_{rad}
\rangle$, $\lambda_{rad, max}$ and $\tau_{cycle}$ obtained from the simulations, to formulate an analytic description of the accretion process. For ambient densities $n_{H,\infty} \ge 10^5$ cm$^{-3}$, we have found that the dimensionless mean accretion rate $\langle
\lambda_{rad} \rangle$ depends only on the temperature of the ambient medium. It is insensitive to $\eta$, $M_{bh}$ and $n_{H,\infty}$. Thus, for $n_{H,\infty} \ge10^5$ cm$^{-3}$ we find $$\begin{aligned}
\langle \lambda_{rad} \rangle \sim
3.3\%~T_{\infty,4}^{5/2}~n_{H,5}^{-0.04} \sim
3.3\%~T_{\infty,4}^{5/2} ,
\label{eq:alambda_a}\end{aligned}$$ while for $n_{H,\infty} \le 10^5$ cm$^{-3}$ we find $$\begin{aligned}
\langle \lambda_{rad} \rangle \sim 3.3\%~
T_{\infty,4}^{5/2}~n_{H,5}^{1/2}.
\label{eq:alambda_b}\end{aligned}$$
As mentioned above, the dependence of $\langle \lambda_{rad} \rangle$ on the density is due to the increasing temperature inside the ionized bubble at low densities. The period of the accretion cycle depends on all the parameters we have investigated in our simulation. In the range of densities $n_{H,\infty} \ge 10^5$ cm$^{-3}$ we find $$\begin{aligned}
\tau_{cycle} &=& (6\times 10^3~{\rm yr})~\eta_{-1}^{\frac{1}{3}}~
M_{bh,2}^{\frac{2}{3}}~n_{H,5}^{-\frac{1}{3}}~T_{\infty,4}^{-\frac{1}{2}}\end{aligned}$$ where we use the notation of $M_{bh,2}\equiv M_{bh}/(10^2$ M$_{\sun})$. However, at lower densities $n_{H,\infty} \le 10^5$ cm$^{-3}$, we find $$\begin{aligned}
\tau_{cycle} &=& (6\times 10^3~{\rm yr})~\eta_{-1}^{\frac{1}{3}}~
M_{bh,2}^{\frac{2}{3}}~n_{H,5}^{-\frac{1}{6}}~T_{\infty,4}^{-\frac{1}{2}} \end{aligned}$$ in which only the dependence on $n_{H,5}$ changes. The different dependence of $\tau_{cycle}$ on $n_{H,\infty}$ is associated with a change of the mean accretion rate $\langle \lambda_{rad} \rangle$ for each density regime. The deviation of $\tau_{cycle}$ from the power law fit at $n_{H,\infty} =10^7$ cm$^{-3}$ is not associated with any variation of the mean accretion rate. Our value of $\tau_{cycle}$ for $n_{H,\infty} =10^7$ cm$^{-3}$ is in good agreement with the value found by MCB09.
Dimensionless accretion rate : $\langle \lambda_{rad}
\rangle $
------------------------------------------------------
In this section we seek a physical explanation for the relationship between the mean accretion rate $\langle \lambda_{rad} \rangle$ and the temperature of the ambient medium found in the simulations. The model is valid in all the parameter space we have explored with a caveat in the low density regime ($n_{H,\infty}<3\times
10^5$ cm$^{-3}$) and at low ambient temperatures ($T_{\infty}<3000$ K).
Figure \[temp\_all\] shows the time-averaged temperature profiles for simulations in which we vary $\eta$, $M_{bh}$, $n_{H,\infty}$ and $T_{\infty}$. In the case of different $M_{bh}$ the radii are rescaled so that direct comparisons can be made with the case of $100~$M$_{\sun}$. Vertical lines indicate the accretion radius $r_{acc}$, inside of which gas is accreted and outside of which gas is pushed out to the ionization front. We find that the value of $r_{acc}$ is generally insensitive to the parameters of the simulation as is the gas temperature at $r_{acc}$.
Accretion onto the black hole of gas inside the hot ionized sphere is limited by the thermal pressure of the hot gas and by the outflow velocity of the gas that is produced by the pressure gradient inside the Strömgren sphere. Thus, the accretion radius, $r_{acc}$, is analogous to the inner Bondi radius, $r_{b,in}$, modified to take into account temperature and pressure gradient inside the hot bubble.
Let us assume that the average accretion rate onto the black hole is $$\langle {\dot M} \rangle = 4\pi \lambda_{B}
r_{acc}^2\rho_{in}c_{s,in},
\label{eq:acc}$$ where $\rho_{in}$ and $c_{s,in}$ (and the corresponding temperature $T_{in}$) are the density and the sound speed at $r_{acc}$. Based on the results illustrated in Figure \[temp\_all\], we expect the accretion rate to depend only on $\rho_{in}$, since $r_{acc}$ and $c_{s,in}$ can be taken to be constants.
When a Strömgren sphere is formed, the gas inside the hot bubble expands and its density decreases. Inside the ionization front the temperature is about $10^4-10^5$K. Thus, assuming pressure equilibrium across the ionization front we find the dependence of $\rho_{in}$ on $T_{\infty}$: $$\begin{aligned}
\rho_{in} \approx \rho_{\infty}{T_{\infty} \over T_{in}}=\rho_\infty
\left({c_{s,\infty} \over c_{s,in}}\right)^{2}.
\label{eq:pres_eq}\end{aligned}$$ We find $f=r_{acc}/r_{b,in} \sim 1.8$ and the temperature at $r_{acc}$ is $T_{in}\sim~4\times 10^4$ K independent of $\eta$, $M_{bh}$, $n_{H,\infty}$ and $T_{\infty}$ for a fixed spectral index of radiation $\alpha=1.5$. The dimensionless accretion rate inside of the Strömgren sphere normalized by the Bondi accretion rate in the ambient medium is then : $$\begin{aligned}
\langle \lambda_{rad} \rangle &\simeq& \lambda_B \frac{r_{acc}^2
\rho_{in} c_{s,in}}{r_{b,\infty}^2\rho_{\infty}c_{s,\infty}} \nonumber \\
&\simeq& \frac{1}{4} (1.8)^2\left( \frac{\rho_{in}}{\rho_{\infty}}
\right) \left( \frac{c_{s,in}}{c_{s,\infty}} \right)^{-3} \nonumber \\
&\simeq& 3\%~T_{\infty,4}^{2.5}
\label{eq:alambda}\end{aligned}$$ where we have used $\lambda_{B}=1/4$ appropriate for an adiabatic gas. Thus, $\langle \lambda_{rad} \rangle \propto T_{\infty}^{5/2}$ which is in agreement with the simulation result, given that $r_{acc}$ and $T_{in}$ remain constant when we change the simulation parameters. However, $r_{acc}$ and $T_{in}$ may not stay constant if we modify the cooling or heating function, for instance by increasing the gas metallicity or by changing the spectrum of radiation; this result suggests that the accretion rate is very sensitive to the details of the temperature structure inside the Strömgren sphere which shows a dependence on $n_{H,\infty}$. The temperature profile changes significantly for $n_{H,\infty}<3\times 10^4$ cm$^{-3}$ and this is probably the reason why our model does not fit perfectly $\langle \lambda_{rad} \rangle$ from the simulations in the lower density regime. In the next section we test whether Equation (\[eq:acc\]) is still a good description of our results when we change the thermal structure inside the HII region.
### Dependence on temperature at accretion radius
In this section we study the dependence of the accretion rate on the time-averaged temperature $T_{in}$ at $r_{acc}$. We change the temperature $T_{in}$ by varying the spectral index $\alpha$ of the radiation spectrum and by including Compton cooling of the ionized gas by CMB photons. Here we explore the spectral index of the radiation spectrum in the range $\alpha = 0.5,1.0,1.5,2.0,2.5$ with the energy of photons from $10$ keV $100$ keV. The other parameters are kept constant at $\eta = 0.1$, $M_{bh}=100$ M$_{\sun}$, $n_{H,\infty}=10^{5}$ cm$^{-3}$ and $T_{\infty}= 10^4$ K.
Figure \[pr10\_alpha\] shows the different time-averaged temperature profiles for different values of $\alpha$. Spectra with lower values of the spectral index $\alpha$ produce more energetic photons for a given bolometric luminosity, increasing the temperature inside the ionized bubble. Simulations show that the averaged accretion rate $\langle \lambda_{rad} \rangle$ increases for softer spectrum of radiation. Different slopes ($0.5 \le \alpha \le 2.5$) of the power-law spectrum lead to different $T_{in}$ ($59000~$K to $36000~$K) and $\langle \lambda_{rad} \rangle$ (0.0076 to 0.0509). Adopting a harder spectrum (with $\alpha = 0.5$) instead of the softer ($\alpha = 2.5$) increases $T_{in}$ by a factor of $1.6$ and $\langle \lambda_{rad} \rangle$ decreases by a factor $6.7$. The fit to the simulation results in Figure \[pr10\_alpha\] show that $\langle \lambda_{rad} \rangle$ depends on temperature at $r_{acc}$ as $$\begin{aligned}
\langle \lambda_{rad} \rangle \propto T_{in}^{-4} \propto c_{s,in}^{-8}.\end{aligned}$$
The dependence on $c_{s,in}$ differs from equation (\[eq:alambda\]). However this is not surprising because in these simulations the values of $r_{acc}$ and $c_{s,in}$ do not remain constant while we vary the value of the spectral index $\alpha$. This is due to a change of the temperature and pressure gradients within the HII region. The accretion radius, $r_{acc}$, can be expressed as a function of the Bondi radius inside the hot bubble, $r_{b,in}=GM_{bh}c_{s,in}^{-2}$. From the simulations we obtain the following relationship between these two radii: $$\begin{aligned}
f=\frac{r_{acc}}{r_{b,in}} \simeq 1.8
\left(\frac{T_{in}}{4\times10^4~{\rm K}} \right)^{-0.7\pm0.2}.\end{aligned}$$
Thus, if our model for the accretion rate summarized by equation (\[eq:acc\]) is valid, we should have: $$\begin{aligned}
\langle \lambda_{rad} \rangle &\simeq& \frac{1}{4} \frac{r_{acc}^2
\rho_{in} c_{s,in}}{r_{b,\infty}^2\rho_{\infty}c_{s,\infty}}
\simeq \frac{(1.8)^2}{4}\left( \frac{\rho_{in}}{\rho_{\infty}}
\right) c_{s,in}^{-5.9} c_{s,\infty}^3 \nonumber \\ &\simeq&
3\%~T_{\infty,4}^{2.5} \left(\frac{T_{in}}{4\times10^4~{\rm K}}\right)^{-4},\end{aligned}$$ in agreement with the simulation results $\langle \lambda_{rad}
\rangle \propto T_{\infty}^{2.5} T_{in}^{-4}$ where the dependence on $T_{in}$ was not explored initially. Thus, Bondi-like accretion on the scale of $r_{acc}$ is indeed a good explanation of our results. Given the steep dependence of the value of accretion rate $\langle
\lambda_{rad} \rangle$ on $T_{in}$ it is clear that it is very sensitive on the details of the thermal structure inside the HII region. This means that $\langle \lambda_{rad}\rangle$ depends on the spectrum of radiation and gas metallicity.
Accretion rate at peaks and duty cycle: $\lambda_{rad,max}$, $f_{duty}$
-----------------------------------------------------------------------
We estimate $f_{duty}$ by comparing $\lambda_{rad,max}$ and $\langle
\lambda_{rad} \rangle$ using Equation (\[eq:fduty\]). This quantity gives an estimate of what fraction of black holes are accreting gas at the rate close to the maximum. Within the fitting errors, the log slopes of $\lambda_{rad,max}$ and $\langle \lambda_{rad} \rangle$ as a function of the parameters $M_{bh}$, $T_{\infty}$ are zero. Thus, we assume that the dimensionless accretion rates are independent of these parameters.
For $n_{H,\infty} \ge 10^5$ cm$^{-3}$, $\lambda_{rad,max}$ can be expressed as $\lambda_{rad,max} \sim
0.55~\eta_{-1}^{0.13}~n_{H,5}^{-0.18}~T_{\infty, 4}^{2.0} $ and the dependence of $f_{duty}$ on these parameters can be expressed using equation (\[eq:alambda\_a\]) as $$\begin{aligned}
f_{duty} &\sim& 6\%~\eta_{-1}^{-0.13}~n_{H,5}^{0.14}~T_{\infty,4}^{0.5}\end{aligned}$$ where we include the mild dependence of $\langle \lambda_{rad}
\rangle$ on the density. For $n_{H,\infty}\le~10^5$ cm$^{-3}$, $\lambda_{rad,max} \sim
0.55~\eta_{-1}^{0.13}~n_{H,5}^{0.37}~T_{\infty,4}^{2.0}$ has a different power law dependence on the density and we get $f_{duty}$ as $$\begin{aligned}
f_{duty} &\sim& 6\%~\eta_{-1}^{-0.13}~n_{H,5}^{0.07}~T_{\infty,4}^{0.5}\end{aligned}$$ where $f_{duty}$ shows a milder dependence on the gas density. Thus, we expect about 6% of IMBHs to be accreting near the maximum rate at any given time. This value depends weakly on $\eta$, $n_{H,\infty}$ and $T_{\infty}$.
Average period between bursts : $\tau_{cycle}$
------------------------------------------------
In this section we derive an analytical expression for the period of the luminosity bursts as a function of all the parameters we tested. Although $\tau_{cycle}$ shows a seemingly complicated power law dependencies on the free parameters, we find that $\tau_{cycle}$ is proportional to the time-averaged size of the Strömgren sphere. This is shown in Figure \[rs\_period\]. The linear relation between $\tau_{cycle}$ and the average Strömgren radius $\langle
R_s \rangle$ explains the dependence of $\tau_{cycle}$ on every parameter considered in this work.
The number of ionizing photons created by accretion onto a black hole is determined by the average accretion rate and the radiative efficiency $\eta$. The average accretion rate itself can be expressed as a fraction of the Bondi accretion rate $\langle \lambda_{rad}
\rangle$. Therefore, the average number of ionizing photons emitted near the black hole can be expressed as $$\begin{aligned}
N_{ion} &\propto& \eta \langle \lambda_{rad} \rangle \dot{M}_{B}
\\ \nonumber &\propto& \eta \langle \lambda_{rad} \rangle \frac{G^2
M_{bh}^2 }{c_{s,\infty}^3} \rho_{\infty} .\end{aligned}$$ It follows that: $$\begin{aligned}
\tau_{cycle} &=& t_{out} \approx {\langle R_s \rangle \over v_{out}}
\propto \left( \frac{3 N_{ion}}{4\pi\alpha_{rec}
n_H^2}\right)^{\frac{1}{3}} \nonumber \\ &\propto& \left(
\frac{1}{n_H^2}\right)^{\frac{1}{3}} \left( \eta \langle \lambda_{rad}
\rangle \frac{G^2 M_{bh}^2 }{c_{s,\infty}^3} \rho_{\infty}
\right)^{\frac{1}{3}},\end{aligned}$$ where we find $v_{out} \sim {1 \over 3} c_{s,in}$. Ignoring constant coefficients and using equation (\[eq:alambda\_a\]) for $n_{H,\infty} \ge
10^5$ cm$^{-3}$, we find : $$\begin{aligned}
\tau_{cycle} &\propto& \eta^{\frac{1}{3}} M_{bh}^{\frac{2}{3}}
n_{H,\infty}^{-\frac{1}{3}} T_{\infty}^{-\frac{1}{2}}, \end{aligned}$$ or using equation (\[eq:alambda\_b\]) for $n_{H,\infty} \le
10^5$ cm$^{-3}$, we find : $$\begin{aligned}
\tau_{cycle} &\propto& \eta^{\frac{1}{3}} M_{bh}^{\frac{2}{3}}
n_{H,\infty}^{-\frac{1}{6}} T_{\infty}^{-\frac{1}{2}} \end{aligned}$$ which are exactly as in the empirical fitting formulas in both density regimes and also in good agreement with the analytical work by MBCO09. This explains the dependence of $\tau_{cycle}$ on any tested parameter $\eta$, $M_{bh}$, $n_{H,\infty}$ and $T_{\infty}$. In Figure \[rs\_period\] we also show simulation results assuming $\eta
\propto \dot{m}$. All simulations show the same relationship between $\tau_{cycle}$ and $\langle R_s\rangle$. However, the simulation with the highest ambient density ($n_{H,\infty}=10^7$ cm$^{-3}$) deviates from the linear relationship, but is in agreement with the numerical simulation by MCB09. It appears that in the high density regime $\tau_{cycle}$ decreases steeply with decreasing $\langle R_s\rangle$.
We can interpret $\tau_{cycle}$ as the time scale at which the gas inside HII region gets depleted. If the gas depletion inside the Strömgren sphere is dominated by the outward gas flow, then $\tau_{cycle} \propto \langle R_{s} \rangle/c_{s,in} $ in agreement with the empirical linear relation in Figure \[rs\_period\]. However, the depletion time scale may be different if the accretion by the black hole dominates gas consumption inside the Strömgren sphere. We can derive this time scale as $$t_{in}= {M_{HII} \over \dot{M}}= \left({\langle R_s \rangle \over
r_{acc}}\right)^2 {\langle R_s \rangle \over 3~c_{s,in}} \sim
\left({\langle R_s \rangle \over r_{acc}}\right)^2 {t_{out} \over 9}.$$ Roughly, we expect $\tau_{cycle}=\min{(t_{out}, t_{in})}$. So, for $\langle R_s \rangle/ r_{acc} \le 3$, the period of the cycle scales as $\langle R_s \rangle^3$. This may explain the deviation of the period for $n_{H,\infty}=10^7$ cm$^{-3}$ from the linear relation. We see in Figure \[temp\_all\] that the ratio $\langle R_s \rangle
/r_{acc} \sim 5$ for $n_{H,\infty}=10^7$ cm$^{-3}$ which is much smaller than the ratio found for other densities.
### Rayleigh-Taylor instability
In 2D simulations we find that RT instability develops across the Strömgren radius, but it decays on short time scales. This can be explained by the pressure gradient inside the Strömgren sphere which does not allow the RT grow. In the linear regime the growth time scale of the RT instability of wavelength $\lambda$ is $$\tau_{RT} \simeq \sqrt{ {{\rho_{sh}+\rho_{in}} \over
{\rho_{sh}-\rho_{in}}}
\frac{2\pi \lambda}{g}} \simeq \sqrt{\frac{2\pi \lambda}{g}}$$ where $\rho_{sh}$ is the density of the shell and $g \simeq
GM_{bh}\langle R_s \rangle ^{-2}$ is the gravitational acceleration at the shell radius. Thus, RT timescale can be expressed as $$\begin{aligned}
\tau_{RT}
\simeq \frac{ \langle R_s \rangle}{c_{s,in}} \sqrt{\frac{2\pi\lambda}{r_{b,in}}}.\end{aligned}$$ So during one cycle perturbations grow on scales: $$\lambda_{RT} < \left( {\tau_{RT} \over \tau_{cycle}} \right) ^2
\frac{r_{b,in}}{2\pi} < \frac{r_{b,in}}{2\pi}$$ where $r_{b,in}$ is the inner Bondi radius. Thus only instability on angular scales $\theta \sim \lambda_{RT}/2\pi \langle R_s
\rangle \le r_{b,in}/(2\pi)^2 \langle R_s \rangle$ grow in our simulation.
Summary and Discussion
======================
In this paper we simulate accretion onto IMBHs regulated by radiative feedback assuming spherical symmetric initial conditions. We study accretion rates and feedback loop periods while varying radiative efficiency, mass of black hole, density and temperature of the medium, and spectrum of radiation. The aim of this work is to simulate feedback-regulated accretion in a wide range of the parameter space to formulate an analytical description of processes that dominate the self-regulation mechanism. Thus, in this first paper we keep the physics as simple as possible, neglecting the effect of angular momentum of the gas, radiation pressure and assuming a gas of primordial composition (i.e. metal and dust free). We will relax some of these assumptions in future works. However, the parametric formulas for the accretion presented in this paper should provide a realistic description of quasi-spherical accretion onto IMBH for ambient gas densities $n_{H, \infty} \lesssim 10^5-10^6$ cm$^{-3}$, as radiation pressure should be minor for these densities.
We find an oscillatory behavior of the accretion rate that can be explained by the effect of UV and X-ray photo-heating. The ionizing photons produced by the black hole near the gravitational radius increase gas pressure around the black hole. This pressure prevents the surrounding gas from being accreted. An over-dense shell starts to form just outside the Strömgren sphere. Due to the decreased accretion rate, the number of emitted ionizing photons decreases and the density inside the Strömgren sphere also decreases with time. Gas accretion onto the black hole is dominant in decreasing the density inside the HII region only for ambient gas density $n_{H,
\infty} \gtrsim 10^7$ cm$^{-3}$; for lower values of the ambient gas density the gas inside the HII region is pushed outward toward the dense shell by a pressure gradient that develops behind the ionization front. Eventually, the pressure gradient inside the Strömgren sphere is not able to support the weight of the over-dense shell that starts to fall toward the black hole. The accretion rate rapidly increases and the Strömgren sphere starts to expand again.
However, the introduction of a small, non-zero, angular momentum in the flow could change the time-dependent behavior of accretion and feedback loop. The inflow rate in the accretion disk that will necessarily develop, and that is not resolved in our simulations, is typically much slower than the free-fall rate since the viscous time scale in units of the free-fall time is $ t_{visc}/t_{ff} \sim
\alpha^{-1} \mathcal{M}^2 $ where $\alpha$ is the dimensionless parameter for a thin disk [@ShakuraS:1973] and $\mathcal{M}$ is the gas Mach number. Therefore, angular momentum may produce a long delay between changes in the accretion rate at the inner boundary of our simulation and their mirror in terms of output luminosity. Hence $\alpha \simeq
0.01-0.1$ and $\mathcal{M}$ at the inner boundary of our simulations is of order of unity, time delays of 10-100 free-fall times are shorter if the disk is smaller than the inner boundary of the simulation. We have started investigating the effect of such a delay on the periodic oscillations and preliminary results show that the period of the oscillations can be modified by the time delay but the oscillatory behavior is still present (at least for delays of 10-100 free fall times calculated at the inner boundary). As long as the time delay is shorter than the oscillation period, that depends mainly on the gas density, it does not seem to affect the results. We are carefully investigating this in the low and high density regimes where the oscillation pattern and the periods are different. At low densities a time delay of a few hundred free-fall times is much smaller compared to the oscillation period, whereas at the high densities the maximum time delay that we have tested is comparable to the oscillation period. We will publish more extensive results on this effect in our next paper in this series.
We find that the average accretion rate is sensitive to the temperature of the ambient medium and to the temperature profile inside the ionized bubble, and so depends on the gas cooling function and spectral energy distribution of the radiation. The period of the accretion bursts is insensitive to the temperature structure of the HII region, but is proportional to its radius.
Our simulations show that 1D results adequately reproduce 2D results in which instabilities often develop. We find that the accretion rate is expressed as $\lambda_{rad} \simeq 3\%~T_{\infty,4}^{2.5} \left(
T_{in}/4\times 10^4~{\rm K} \right)^{-4}$. We also derive $\tau_{cycle}$ as a function of $\eta$, $M_{bh}$, $n_{H,\infty}$ and $T_{\infty}$. The dependencies of $\langle \lambda_{rad} \rangle$ and $\tau_{cycle}$ on our free parameters can be explained analytically. Assuming pressure equilibrium across the Strömgren sphere is a key ingredient to derive the dependence of $\langle
\lambda_{rad} \rangle$ on $T_\infty$, whereas the linear relation between the average size of the Strömgren sphere and $\tau_{cycle}$ is used to derive the dependence of $\tau_{cycle}$ on all the parameters we varied.
The qualitative picture of the feedback loop agrees with the description of X-ray bursters in [@CowieOS:78]. After extrapolating our analytical formulas to black holes of a few solar masses studied by [@CowieOS:78], we find that the average accretion rate is in good agreement ($L \sim 2\times
10^{35}$ erg/s). However, the details of the accretion rate as a function of time, the burst period and peak accretion rates show qualitative differences. [@CowieOS:78] simulations do not show periodic oscillation while our simulations have well-defined fast rise and exponential decay of accretion followed by quiescent phases of the accretion rate. This regular pattern of accretion bursts is possible only when spherical symmetry is maintained on relatively large scales during oscillations. An axisymmetric radiation source [@Proga:07; @ProgaOK:08; @KurosawaPN:09; @KurosawaP:09a; @KurosawaP:09b] or inhomogeneous initial condition on scale of the Bondi radius can break the symmetry.
Our simulations are also in excellent qualitative agreement with simulations by MCB09 that studied accretion onto $100$ M$_\odot$ black hole for the case $n_{H,\infty}=10^7$ cm$^{-3}$. However, we find a dimensionless accretion rate $\langle \lambda_{rad} \rangle \sim 3\%$ ($\langle \lambda_{rad} \rangle \sim 2\%$ including helium heating/cooling) that is about one order of magnitude larger than in MCB09. The cycle period, $\tau_{cycle}$, is in better agreement since $\tau_{cycle}\propto \langle R_s \rangle \propto
\langle\lambda_{rad}\rangle^{1/3}$. The discrepancy in the mean accretion is likely produced by the effect radiation pressure on HI, that becomes important for large $n_{H, \infty}$ and that we have neglected. In addition, our results indicate that the qualitative description of the feedback loop starts to change at ambient densities $>10^6-10^7$ cm$^{-3}$: the oscillation period decreases much more rapidly with increasing ambient density as gas depletion inside the ionized bubble becomes dominated by accretion onto the IMBH. The accretion luminosity during the quiescent phase of the accretion also increases and the two phases of growth and collapse of the dense shell become blended into a smoother modulation of the accretion rate. Hence, further numerical studies are required to characterize accretion onto IMBH in the high-density regime. As mentioned above, in this study we have neglected three important physical processes that may further reduce the accretion rate: 1) Compton heating, 2) radiation pressure and 3) Lyman-$\alpha$ scattering processes. The importance of these processes is thoroughly discussed in MBCO09. @Ricotti:08 and MBCO09 find that Compton heating is not an important feedback mechanism in regulating the accretion rate onto IMBHs in which the gas density inside the Strömgren sphere is roughly independent of the radius. However, MBCO09 suggest that both radiation pressure on HI and Lyman-$\alpha$ radiation pressure can contribute to reducing the accretion rate onto IMBH. At higher densities accretion becomes Eddington limited ($\langle \lambda_{rad} \rangle \dot{M}_B \lesssim \dot{M}_{Edd}$). By assuming $\langle \lambda_{rad} \rangle \sim 1\%~T_{\infty,4}^{2.5}$ which is suggested by simulations including helium and manipulating Equation (\[eq:edd\]) we obtain $M_{BH,2}~T_{\infty,4}~n_{H,5}~\eta_{-1} \gtrsim 40$ as a criteria for Eddington limited condition. This expression predicts that radiation pressure becomes important at gas density $n_{H,5} \sim 10-100$ with other parameters fixed to unity. It also implies that this critical density depends on the black hole mass, gas temperature and radiative efficiency. We expect radiation pressure to play a minor role at low densities also because the accretion luminosity is negligible during the quiescent phases of accretion (between accretion bursts). However, at densities $n_{H, \infty}\ge 10^7$ cm$^{-3}$ the accretion rate during the quiescent phases is not negligible, thus radiation pressure can be important in this regime. The main effect of Thomson radiation pressure is to prevent the accretion luminosity to exceed the Eddington limit. The continuum radiation pressure due to HI ionization can instead be important for sub-Eddington luminosities, but will be strong only at the location of the ionization front during the peak of the accretion burst. We will present more extensive results in the next paper of this series.
The results of this study provide a first step to estimate the maximum X-ray luminosity and period of oscillations of an accreting IMBH from a medium with given physical conditions. Hence, they may be useful for modeling detection probability of ULX originating from accreting IMBH in the local universe. From the average growth rate of IMBHs accreting in this manner it is also possible to estimate the maximum masses of quasars at a given redshift starting from seed primordial black holes. One of the main motivations of this study is to derive simple analytical prescriptions to incorporate growth of seed black holes from Population III stars into large scale cosmological simulation. However, before being able to do use these results in cosmological simulation we need to understand the effects of relaxing our assumption of spherical symmetric accretion: we need to simulate accretion onto moving black holes [@HoyleL:39; @Shima:85; @Ruffert:94; @Ruffert:96] and use more realistic initial conditions, including gas with non-zero angular momentum or a multi-phase turbulent ISM [@Krumholz:05; @Krumholz:06].
The simulations presented in this paper were carried out using high performance computing clusters administered by the Center for Theory and Computation of the Department of Astronomy at the University of Maryland (“yorp”), and the Office of Information Technology at the University of Maryland (“deepthought”). This research was supported by NASA grants NNX07AH10G and NNX10AH10G. The authors thank the anonymous referee for constructive comments and feedback.
BASIC TESTS OF THE CODE
=======================
We test the Bondi accretion formula using ZEUS-MP for the adiabatic indexes $\gamma =1.2, 1.4$ and $1.6$. For a given equation of state, the sonic point where the gas inflow becomes supersonic must be resolved not to overestimate the accretion rate $\lambda_{B}$. The left panel of Figure \[acc\_gamm\] shows the steady accretion rate as a function of the radius at the inner boundary normalized by the Bondi radius. Different lines show results for $\gamma=$1.2, 1.4 and 1.6.
We also test whether our radiative transfer module produces radii of the Strömgren spheres in agreement with the analytical prediction: $(4\pi/3)R_s^3 n_e n_H \alpha_{rec} = N_{ion}$, where $R_s$ is the Strömgren radius and $N_{ion}$ is the number of ionizing photons emitted per unit time. The right panel of Figure \[acc\_gamm\] shows the test of the 1D radiative transfer module without hydrodynamics. Different symbols indicates the radii for the different ionization fractions: $x_e =$ 0.99 (circle), 0.90 (square), 0.50 (triangle).
RADIATIVE TRANSFER MODULE AND TIME STEPPING
===========================================
Our hydrodynamic calculation is performed using ZEUS-MP, returning the density and gas energy at each time step to the radiative transfer module. The operator-splitting method is applied to mediate between hydrodynamics and radiative transfer with a photon-conserving method [@Whalen:06]. For each line of sight radiative transfer equations are solved in the following order:
1. At the inner boundary, the average inflow mass flux $\dot{M}$ is calculated.
2. The mass flux is converted into accretion luminosity $L$, and thus into the number of ionizing photons for a given radiative efficiency $\eta$.
3. The photon spectrum is determined using a power law spectral energy distribution with the spectral index $\alpha$. We use up to $300$ logarithmically spaced frequency bins for photons between $10$ eV up to $100$ keV.
4. The ordinary differential equation for time-dependent radiative transfer cooling/heating and chemistry of the gas are solved using a Runge-Kutta or Semi-Implicit solver for each line of sight with a maximum of 10% error. Photo-heating, cooling for a given cooling function and Compton cooling are calculated.
5. The energy density and the abundances of neutral and ionized hydrogen are updated.
Parallelization is easily implemented in polar angle direction because radiative transfer calculations along each ray are independent of one another.
RESOLUTION STUDIES
==================
We perform a resolution study to confirm that the number of grid zones does not affect the results. Number of zones from 384 to 768 are tested and they all show the similar outputs in terms of accretion rate at peaks, average accretion rate, decaying shape and the period between peaks . Figure \[strom\] shows that the details of the accretion rate history from simulations are not identical but the physical quantities which we are interested in (average accretion rate, peak accretion rate and period of the bursts) do not show significant deviation from each other. In general, a Courant number of 0.5 is used for most simulations, but we try a Courant number which is one order of magnitude smaller to investigate how the results are affected by reducing the hydro-time step by an order of magnitude. The chemical/cooling time steps are calculated independently by the radiation transfer module.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Dept. of Physics, University of Waterloo, Waterloo,\
ONT N2L 3G1, CANADA
author:
- 'R.B. MANN'
title: QUANTUM ENTROPY OF CHARGED ROTATING BLACK HOLES
---
\#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
Introduction
============
For more than two decades physicists have come to appreciate black holes as thermodynamic systems characterized by only a few macroscopic parameters such as mass ($m$), charge ($q$) and angular momentum ($\Omega$). The generic representative of such a hole in general relativity is the Kerr-Newman metric [@KN]. The thermodynamic analogy suggests that there is an entropy associated with this hole that is proportional to the area of the event horizon.
For all other thermodynamic systems, the entropy is proportional to the logarithm of the number of hidden degrees of freedom. Are there analogous degrees of freedom are for a black hole? If so, where do they come from? Statistical explanations [@4] of their origin in terms of a gas of quantum fields have been proposed [@6; @FS]. Unfortunately the resultant expressions for the entropy can be understood as one-loop corrections to the classical black hole entropy, and so do not give any explanation of the classical entropy itself.
For arbitrary static black holes, the divergences of the entropy have the same origin as the UV-divergences of the quantum effective action and can be removed by remormalization of the gravitational couplings in the tree-level gravitational action[@FS]. I report here on work carried out with S. Solodukhin for the analogous problem in the stationary case[@RS]. Remarkably, the UV-divergences for the one-loop entropy of a Kerr-Newman black hole are renormalized in the same way as for a static black hole.
The Euclidean Kerr-Newman Metric
================================
The Euclidean Kerr-Newman metric can be in the form $$ds^2_E={\hat{\rho}^2\over \hat{\Delta}}dr^2+{\hat{\Delta}\hat{\rho}^2\over
(r^2-\hat{a}^2)^2}\omega^2+\hat{\rho}^2(d\theta^2+\sin^2\theta \tilde{\omega}^2)
\label{5}$$ where the Euclidean time is $t=\imath \tau$ and the rotation and charge parameters have also been transformed $a=\imath \hat{a},~q=\imath \hat{q}$, so that the metric (\[5\]) is purely real. Here $\hat{\Delta}(r)=(r-\hat{r}_+)(r-\hat{r}_-)$, where $r_{\pm}=m\pm \sqrt{m^2+\hat{a}^2+\hat{q}^2}$, the quantities $\omega$ and $\tilde{\omega}$ take the form $$\omega={(r^2-\hat{a}^2)\over \hat{\rho}^2}
(d\tau-\hat{a}\sin^2\theta d \phi ) \qquad
\tilde{\omega}={(r^2-\hat{a}^2)\over
\hat{\rho}^2}(d\phi+{\hat{a}\over (r^2-\hat{a}^2)}d\tau)
\label{4}$$ with $\hat{\rho}^2=r^2-\hat{a}^2 \cos^2 \theta$. This space-time has a pair of orthogonal Killing vectors $$K=\partial_\tau-{\hat{a}\over r^2-\hat{a}^2}\partial_\phi~,
~~\tilde{K}=\hat{a}\sin^2\theta \partial_\tau+\partial_\phi \label{4a}$$ which are the respective analogs of the vectors $\partial_\tau$ and $\partial_\phi$ in the (Euclidean) Schwarzchild case. The horizon surface $\Sigma$ defined by $r=\hat{r}_+$ is the stationary surface of the Killing vector $K$. Near this surface the metric (\[5\]) is approximately $ds^2_E=ds^2_\Sigma+\hat{\rho}^2_+ ds^2_{C_2}$ where $\hat{\rho}^2_+=\hat{r}^2_+ -\hat{a}^2\cos^2\theta$ and $$ds^2_\Sigma
=\hat{\rho}_+^2d\theta^2+{(\hat{r}^2_+-\hat{a}^2)^2\over\hat{\rho}_+^2}
\sin^2\theta d\psi^2
\label{8}$$ is the metric on the horizon surface $\Sigma$ up to $O(x^2)$, where $(r-\hat{r}_+)={\gamma x^2 \over 4}$ and $\gamma=2\sqrt{m^2+\hat{a}^2+\hat{q}^2}$. The angle co-ordinate $\psi=\phi+{\hat{a} \over (\hat{r}^2_+-\hat{a}^2)} \tau$ and is well-defined on $\Sigma$. The metric $ds^2_{C_2}$ is that of a two-dimensional disk $C_2$ $$ds^2_{C_2}=dx^2+{\gamma^2 x^2\over 4 \hat{\rho}_+^4}d\chi^2~~.
\label{11}$$ attached to $\Sigma$ at a point ($\theta,~\psi$), where $\chi=\tau-\hat{a}\sin^2\theta\ \phi$ is an angle co-ordinate on $C_2$.
Regularity of the metric near the horizon implies the identifications $\psi \leftrightarrow \psi +2\pi$ and $\chi \leftrightarrow \chi+4\pi\gamma^{-1}\hat{\rho}^2_+$. For this latter condition to hold independently of $\theta$ on the horizon, it is also necessary to identify $(\tau,~\phi )$ with $(\tau+2\pi \beta_H,~\phi-2\pi \Omega \beta_H)$, where $\Omega={\hat{a} \over(\hat{r}^2_+- \hat{a}^2)}$ is the (complex) angular velocity and $\beta_H={(\hat{r}_+^2-\hat{a}^2)/
\sqrt{m^2+\hat{a}^2+\hat{q}^2}}$. The identified points have the same coordinate $\psi$.
Near $\Sigma$ we therefore have the following description of the Euclidean Kerr-Newman geometry: attached to every point $(\theta, \psi$) of the horizon is a two-dimensional disk $C_2$ with coordinates ($x, \chi$). The periodic identification of points on $C_2$ holds independently of any point on the horizon $\Sigma$, even though $\chi$ is not a global coordinate. As in static case, there is an abelian isometry generated by the Killing vector $K$, whose fixed set is $\Sigma$. Locally we have $K=\partial_\chi$. The periodicity is in the direction of the vector $K$ and the resulting Euclidean space $E$ is regular manifold.
Now consider closing the trajectory of $K$ with an arbitrary period $\beta\neq \beta_H$. This implies the identification $(\tau+2\pi \beta,~\phi-2\pi \Omega \beta)$, and the metric on $C_2$ becomes $$ds^2_{C_{2,\alpha}}=dx^2+\alpha^2x^2d\bar{\chi}^2
\label{12}$$ where $\chi=\beta \hat{\rho}^2_+ (\hat{r}_+^2-\hat{a}^2)^{-1}\bar{\chi}$ is a new angle coordinate with period $2\pi$. This is the metric of a two dimensional cone with angular deficit $\delta=2\pi (1-\alpha)$, $\alpha\equiv {\beta\over \beta_H}$. With this new identification the metric (\[5\]) now describes the Euclidean conical space $E_\alpha$ with singular surface $\Sigma$.
For static metrics it is known that curvature tensors behave as $\alpha$-dependent distribution functions. This can also be shown to be true for stationary metrics by regulating (\[12\]) so that $ds^2_{C_{2,\alpha , b}}=f(x,b)dx^2+\alpha^2x^2d\bar{\chi}^2$ where $f(x,b)$ is some smooth regulating function such that $\lim_{b\to 0} f(x,b) = 1$. An evaluation of the curvature tensors then yields [@RS] $$\begin{aligned}
&&R^{\mu\nu}_{\ \ \alpha\beta} = \bar{R}^{\mu\nu}_{\ \
\alpha\beta}+
2\pi (1-\alpha) \left( (n^\mu n_\alpha)(n^\nu n_\beta)- (n^\mu
n_\beta)
(n^\nu n_\alpha) \right) \delta_\Sigma \nonumber \\
&&R^{\mu}_{ \ \nu} = \bar{R}^{\mu}_{ \ \nu}+2\pi(1-\alpha)(n^\mu
n_\nu)
\delta_\Sigma \quad
R = \bar{R}+4\pi(1-\alpha) \delta_\Sigma
\label{27}\end{aligned}$$ in the $b \rightarrow 0$ limit, where $\delta_\Sigma$ is the delta-function $\int_{\cal M}^{}f\delta_\Sigma = \int_{\Sigma}^{}f$ and $(n_\mu n_\nu)=\sum_{a=1}^{2}n^a_\mu n^a_\nu$, where $n^a_\mu$ are both normal to $\Sigma$. Barred quantities denoted tensors evaluated with the unregulated metric (\[12\]). Quadratic curvature invariants may also be shown to have a structure that is formally identical to the static case [@RS].
Heat Kernel Expansion
=====================
In the Euclidean path integral approach to a statistical field system at temperature $T=(2\pi\beta )^{-1}$ one considers the fields which are periodic with respect to imaginary time $\tau$ with period $2\pi\beta$. For a (regulated) rotating black hole metric this entails closing the integral curves of $K$ (\[4\]) for arbitrary $\beta$. The partition function $Z(\beta)$ then becomes the functional integral of the matter Euclidean action on $E_\alpha$, with periodicity conditions imposed on the matter field(s).
For the matter action $I_{E}={1\over 2}\int_{E_\alpha}(\nabla \varphi )^2$ standard techniques yield $$\ln Z (\beta )=-{1\over 2}\ln det (-\Box_{E_\alpha})
{1\over (4\pi s)^2}\sum^{\infty}_{n=0} a_n s^n
\label{40}$$ where the heat-kernel coefficients $a_n$ are a sum of standard and conical coefficients. An evaluation of these indicates that the UV-divergent part of the entropy is renormalized in a manner identical to that for static black holes. Once this is taken into account, the contribution to the entropy $S=-(\beta \partial_\beta-1)\ln Z(\beta )|_{\beta=\beta_H}$ for a Kerr-Newman black hole becomes [@RS] $$S_{div}={1 \over 48\pi \epsilon^2} A_\Sigma
+{1\over 45}
\left( 1 -{3q^2\over 4r^2_+}(1+{(r^2_++a^2)\over a r_+} \tan^{-1}({a\over r_+}))
\right) \ln {L \over \epsilon}
\label{46}$$ in the Lorentzian section, where $A_\Sigma=4\pi (r^2_++a^2)$ is area of the horizon $\Sigma$. When $q=0$ the quantum-corrected part of the entropy in (\[46\]) is the same as that for a Schwarzchild black hole. At present there is no explanation for this.
Concluding Remarks
==================
It is still an open question as to what degrees of freedom are counted by the entropy of black hole. The conical-deficit methods employed here clearly indicate that the entropy of a Kerr-Newman black hole is associated with the horizon. A proper treatment of the statistical-mechanical calculation of the quantum entropy should provide us with a better understanding of the relationship between the different methods of assigning entropy to a black hole.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank the organizers of the 2nd Sakharov Conference for their invitation to me to speak at this meeting and for their kind hospitality. This work was supported by the Natural Sciences and Engineering Research Council of Canada.
References {#references .unnumbered}
==========
[99]{}
R.P. Kerr, Phys. Rev. Lett. [**11**]{} 237 (1963); E.T. Newman [*et.al.*]{}, J. Math. Phys. [**6**]{} 918 (1965). J.D. Bekenstein qr-qc/9409015; V.P.Frolov, hep-th/9510156. G.’t Hooft, Nucl. Phys. [**B256**]{}, 727 (1985); R.B.Mann, L.Tarasov and A.Zelnikov, Class.Quant.Grav. [**9**]{}, 1487 (1992). D.V.Fursaev, S.N.Solodukhin, Phys.Lett. [**B365**]{}, 51 (1996); Phys.Rev. [**D52**]{}, 2133 (1995). J.D.Brown, E.A.Martinez, J.W.York, Jr. Phys.Rev.Lett. [**66**]{}, 2281 (1991). R.B. Mann and S.N. Solodukhin, Physical Review [**D**]{} (to be published).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $G$ be a reductive algebraic group over $\Q$, and suppose that $\Gamma \subset G(\R)$ is an arithmetic subgroup defined by congruence conditions. A basic problem in arithmetic is to determine the multiplicities of discrete series representations in $L^2(\Gamma \bks G(\R))$, and in general to determine the traces of Hecke operators on these spaces. In this paper we give a conjectural formula for the traces of Hecke operators, in terms of stable distributions. It is based on a stable version of Arthur’s formula for $L^2$-Lefschetz numbers , which is due to Kottwitz [@SVAF]. We reduce this formula to the computation of elliptic p-adic orbital integrals and the theory of endoscopic transfer. As evidence for this conjecture, we demonstrate the agreement of the central terms of this formula with the unipotent contributions to the multiplicity coming from Selberg’s trace formula in Wakatsuki [@Wak; @Mult], in the case $G=\GSp_4$ and $\Gamma=\GSp_4(\Z)$.'
author:
- Steven Spallone
title: Stable Trace Formulas and Discrete Series Multiplicities
---
Introduction
============
Let $G$ be a reductive algebraic group over $\Q$, and $\Gamma$ an arithmetic subgroup of $G(\R)$ defined by congruence conditions. Then $G(\R)$ acts on $L^2(\Gamma \bks G(\R))$ via right translation; let us write $R$ for this representation. A fundamental problem in arithmetic is to understand $R$. As a first step, we may decompose $R$ as $$R= R_{\disc} \oplus R_{\cont},$$ where $R_{\disc}$ is a direct sum of irreducible representations, and $R_{\cont}$ decomposes continuously. The continuous part may be understood inductively through Levi subgroups of $G$ as in [@Langlands; @c]. We are left with the study of $R_{\disc}$. Given an irreducible representation $\pi$ of $G(\R)$, write $R_{\disc}(\pi)$ for the $\pi$-isotypic subspace of $R_{\disc}$. Then $$R_{\disc}(\pi) \cong \pi^{\oplus m_{\disc}(\pi)}$$ for some integer $m_{\disc}(\pi)$. (We may also write $m_{\disc}(\pi,\Gamma)$.) A basic problem is to compute these integers.
There is more structure than simply these dimensions, however. Arithmetic provides us with a multitude of Hecke operators $h$ on $L^2(\Gamma \bks G(\R))$ which commute with $R$. Write $R_{\disc}(\pi,h)$ for the restriction of $h$ to $R_{\disc}(\pi)$. The general problem is to find a formula for the trace of $R_{\disc}(\pi,h)$.
We focus on discrete series representations $\pi$. These are representations which behave like representations of compact or finite groups, in the sense that their associated matrix coefficients are square integrable. Like other smooth representations, they have a theory of characters developed by Harish-Chandra. They separate naturally into finite sets called $L$-packets. For an irreducible finite-dimensional representation $E$ of $G(\C)$, there is a corresponding $L$-packet $\Pi_E$ of discrete series representations, consisting of those with the same infinitesimal and central characters as $E$. Say that a discrete series representation is regular if $\pi \in \Pi_E$, with the highest weight of $E$ regular.
We follow the tradition of computing $\tr R_{\disc}(\pi,h)$ through trace formulas. This method has gone through several incarnations, beginning in a paper [@Selberg] of Selberg’s for $\GL_2$, in which he also investigated the continuous Eisenstein series. A goal was to compute dimensions of spaces of modular forms, and traces of Hecke operators on these spaces. These spaces of modular forms correspond to the spaces $R_{\disc}(\pi)$ we are discussing in this case. His trace formula is an integral, over the quotient of the upper half space $X$ by $\Gamma$, of a sum of functions $H_{\gm}$, one for each element of $\Gamma$. Let us write it roughly as
$$\dim_{\C} S(\Gamma)= \int_{\Gamma \bks X} \sum_{\gm \in \Gamma} H_{\gm}(Z) dZ,$$
for some space $S(\Gamma)$ of cusp forms with a suitable $\Gamma$-invariance condition.
Here $dZ$ is a $G(\R)$-invariant measure on $X$. When the quotient $\Gamma \bks X$ is compact, the sum and integral may be interchanged, leading to a simple expression for the dimensions in terms of orbital integrals. The interference of the Eisenstein series precludes this approach in the noncompact quotient case. Here there are several convergence difficulties, which Selberg overcomes by employing a truncation process. Unfortunately the truncation process leads to notoriously complicated expressions which are far from being in closed form. This study of $R_{\disc}(\pi)$ has been expanded to other reductive groups using what is called the Arthur-Selberg trace formula. (See Arthur [@Arthur; @Clay].)
Generally, a trace formula is an equality of distributions on $G(\R)$, or on the adelic group $G(\Aff)$. One distribution is called the geometric side; it is a sum of terms corresponding to conjugacy classes of $G$. Given a test function $f$, the formula is essentially made up of combinations $I_M(\gm,f)$ of weighted integrals of $f$ over the conjugacy classes of elements $\gm$. (Here $M$ is a Levi subgroup of $G$.) Another distribution is called the spectral side, involving the Harish-Chandra transforms $\tr \pi(f)$ for various representations $\pi$. Here, the operator $\pi(f)$ is given by weighting the representation $\pi$ by $f$. The geometric and spectral sides agree, and in applications we can learn much about the latter from the former. Some of the art is in picking test functions to extract information about both sides.
The best general result using the trace formula to study $\tr R_{\disc}(\pi,h)$ seems to be in Arthur . He produces a formula for $$\label{stableR}
\sum_{\pi \in \Pi} \tr R_{\disc}(\pi,h),$$ where $\Pi$ is a given discrete series $L$-packet for $G(\R)$. He uses test functions $f$ which he calls “stable cuspidal”. Their Fourier transforms $\pi \mapsto \tr \pi(f)$ are “stable” in the sense that they are constant on $L$-packets, and “cuspidal" in the sense that, considered as a function defined on tempered representations, they are supported on discrete series. (Tempered representations are those which appear in the Plancherel formula for $G(\R)$.) Using his invariant trace formula (Arthur [@ITF1], [@ITF2]), he obtains (\[stableR\]) as the spectral side. The geometric side is a combination of orbital integrals for $h$ and values of Arthur’s $\Phi$-function, which describes the asymptotic values of discrete series characters averaged over an $L$-packet.
In particular, he produces a formula for $$\label{stablemult}
\sum_{\pi \in \Pi} m_{\disc}(\pi),$$ for a $L$-packet $\Pi$ of (suitably regular) discrete series representations.
In the case of $G=\GL_2$, there is a discrete series representation $\pi_k$ for each integer $k \geq 1$. In this case $m_{\disc}(\pi_k)$ is the dimension of the space $S_k(\Gamma)$ of $\Gamma$-cusp forms of weight $k$ on the upper half plane. Restriction to $\SL_2(\R)$ gives two discrete series $\{ \pi_k^+, \pi_k^- \}$ in each $L$-packet. However we may still use Arthur’s formula here since $m_{\disc}(\pi_k^+, \Gamma)=m_{\disc}(\pi_k^-, \Gamma)$ for every arithmetic subgroup $\Gamma$. (Endoscopy does not play a role.)
For the group $\GSp_4(\R)$ there are two discrete series representations in each $L$-packet: one “holomorphic" and one “large" discrete series. Let $\pi$ be a holomorphic discrete series, and write $\pi'$ for the large discrete series representation in the same $L$-packet as $\pi$. The multiplicity $m_{\disc}(\pi,\Gamma)$ is also the dimension of a certain space of vector-valued Siegel cusp forms (see [@Wallach]) on the Siegel upper half space, an analogue of the usual cusp forms on the upper half plane. For $\Gamma=\Sp_4(\Z)$, the dimensions of these spaces of cusp forms were calculated in Tsushima [@Tsu; @1], [@Tsu; @2] using the Riemann-Roch-Hirzebruch formula, and later in Wakatsuki [@Wak; @Dim] using the Selberg trace formula and the theory of prehomogeneous vector spaces. Wakatsuki then evaluated Arthur’s formula in [@Wak; @Mult] to compute $$m_{\disc}(\pi,\Gamma)+m_{\disc}(\pi',\Gamma),$$ thereby deducing a formula for $m_{\disc}(\pi',\Gamma)$.
A natural approach to isolating the individual $m_{\disc}(\pi)$, or generally the individual $\tr R_{\disc}(\pi,h)$, is to apply a trace formula to a matrix coefficient, or more properly, a pseudocoefficient $f$. This means that $f$ is a test function whose Fourier transform picks out $\pi$ rather than the entire packet $\Pi $ containing $\pi$. (See Definition \[pseudoscience\] below.) Such a function will not be stable cuspidal, but merely cuspidal. Arthur (, see also [@Arthur; @Clay]) showed that $I_M(\gm,f)$ vanishes when $f$ is stable cuspidal and the unipotent part of $\gm$ is nontrivial. If we examine the geometric side of Arthur’s formula for a pseudocoefficient $f$, we must evaluate the more complicated terms $I_M(\gm,f)$ for elements $\gm$ with nontrivial unipotent part. At the time of this writing, such calculations have not been made in general; we take another approach.
Distinguishing the individual representations $\pi$ from others in its $L$-packet leads to the theory of endoscopy, and stable trace formulas. The grouping of representations $\pi$ into packets $\Pi$ on the spectral side mirrors the fusion of conjugacy classes which occurs when one extends the group $G(\R)$ to the larger group $G(\C)$. If $F$ is a local or global field, then a stable conjugacy class in $G(F)$ is, roughly, the union of classes which become conjugate in $G(\ol{F})$. (See [@Langlands; @defn] for a precise definition.)
The distribution which takes a test function to its integral over a regular semisimple stable conjugacy class is a basic example of a stable distribution. Indeed, a stable distribution is defined to be a closure of the span of such distributions (see [@Langlands; @debuts], [@Langlands; @defn]). A distribution on $G(F)$ is stabilized if it can be written as a sum of stable distributions, the sum being over smaller subgroups $H$ related to $G$. These groups $H$ are called endoscopic groups for $G$; they are tethered to $G$ not as subgroups but through their Langlands dual groups. As part of a series of techniques called endoscopy, one writes unstable distributions on $G$ as combinations of stable distributions on the groups $H$. Part of this process is the theory of transfer, associating suitable test functions $f^H$ on $H(F)$ to test functions $f$ on $G(F)$ which yield a matching of orbital integrals. Indeed this was the drive for [@Ngo]. As the name suggests, the theory of endoscopy, while laborious, leads to an intimate understanding of $G$.
There has been much work in stabilizing Arthur’s formula: See for example [@Langlands; @debuts], [@A-STF1], [@A-STF2], [@A-STF3]. In Kottwitz’s preprint [@SVAF], he defines a stable version of Arthur’s Lefschetz formula, which we review below. (See also [@Morel].) It is a combination $$\Kot(f)=\sum_H \iota(G,H) ST_g(f^H)$$ of distributions $f \mapsto f^H \mapsto ST_g(f^H)$ over endoscopic groups $H$ for $G$. Here the distributions $ST_g$, defined for each $H$, are stable. (See Section \[Various Invariants\] for the definition of the rational numbers $\iota(G,H)$.) Each $ST_g$ is a sum of terms corresponding to stable conjugacy classes of elliptic elements $\gm \in H(\Q)$. The main result of [@SVAF] is that $\Kot$ agrees with Arthur’s distribution, at least for functions $f$ which are stable cuspidal at the real place.
As part of the author’s thesis [@Sp], he evaluated the identity terms of Kottwitz’s distribution for the group $G=\SO_5$ at a function $f$ which was a pseudocoefficient for a discrete series representation at the real place. Later, Wakatsuki noted that the resulting expressions match up with the terms in his multiplicity formulas $m_{\disc}(\pi, \Gamma)$ and $m_{\disc}(\pi',\Gamma)$ corresponding to unipotent elements. Moreover, the contribution in [@Sp] from the endoscopic group accounted for the difference in these multiplicity formulas, while the stable part corresponded to the sum. After further investigation, we conjecture simply that Kottwitz’s distribution evaluated at a function $f=f_{\pi,\Gamma}$ suitably adapted to $\pi$ and $\Gamma$ is equal to $m_{\disc}(\pi,\Gamma)$, under a regularity condition on $\pi$. (See Section \[Conjecture\] for the precise statement.) Of course this is compatible with Arthur’s results in .
In this paper we give some computational evidence for this conjecture. We also reduce the computation of each $ST(f_{\pi,\Gamma}^H)$ to evaluating elliptic orbital $p$-adic integrals for the transfer $f^{\infty H}$ at the finite places. The rest breaks naturally into a problem at the real points and a global volume computation.
The main ingredient at the archimedean place is Arthur’s $\Phi$-function $\Phi_M(\gm, \Theta^E)$, which we review. This quantity gives the contribution from the real place to the trace formulas in and [@GKM]. It also plays a prominent role in Kottwitz’s formula. This function, originally defined by the asymptotic behaviour of a stable character near a singular elliptic element $\gm$, was expressed in closed form by the author in [@SpalPhi].
There are two volume-related constants that enter into any explicit computation of $ST_g$. The first is $\ol{v}(G)$, which is essentially the volume of an inner form of $G$ over $\R$. It depends on the choice of local measure $dg_{\infty}$. The second comes about from orbital integrals at the finite ideles, and depends on the choice of local measure $dg_f$. These integrals may frequently be written in terms of the volumes of open compact subgroups $K_f$ of $G(\Aff_f)$. In practice, one is left computing expressions such as $\ol{v}(G)^{-1} \vol_{dg_f}(K_f)^{-1}$, which are independent of the choice of local measures. More specifically, we define $$\chi_{K_f}(G)=\ol{v}(G)^{-1} \vol_{dg_f}(K_f)^{-1} \tau(G) d(G).$$ Here $\tau(G)$ is the Tamagawa number of $G$ and $d(G)$ is the index of the real Weyl group in the complex Weyl group. A main general result of this paper, Theorem \[chi-theorem\], interprets $\chi_{K_f}(G)$ via Euler characteristics of arithmetic subgroups. It extends a computation of Harder [@Harder], which was for semisimple simply connected groups, to the case of reductive groups, under some mild hypotheses on $G$.
We work out two examples in this paper, one for $\SL_2$ and another for $\GSp_4$. It is easy to verify our conjecture for $G=\SL_2$ and $\Gamma=\SL_2(\Z)$ using the classic dimension formula for cusp forms. In this case endoscopy does not appear. The calculations for $\GSp_4$ are more complex; we content ourselves with working out the central terms of Kottwitz’s formula.
If $\pi$ is a holomorphic discrete series representation of $\GSp_4(\R)$, write $H_{1}^{\pi}$ for the central-unipotent terms of the Selberg trace formula, as evaluated in [@Wak; @Mult] to compute $m_{\disc}(\pi,\Gamma)$. Here $\Gamma=\GSp_4(\Z)$. If $\pi$ is a large discrete series representation, write $H_{1}^{\pi}$ for the central-unipotent terms in [@Wak; @Mult] contributing to $m_{\disc}(\pi,\Gamma)$. In both cases, write $f=f_{\pi,\Gamma}=f_{\infty} f^{\infty}$, with $f_{\infty}$ a pseudocoefficient for $\pi$, and $f^{\infty}$ the (normalized) characteristic function of the integer adelic points of $G$. Write $\Kot(f, \pm 1)$ for the central terms of Kottwitz’s formula applied to $f$.
As evidence for our conjecture, we show:
\[Theorem\] For each regular discrete series representation $\pi$ of $G(\R)$ we have $$\Kot(f_{\pi,\Gamma}, \pm 1)=H_1^{\pi}.$$
We believe that the central terms of Kottwitz’s formula will generally match up with the difficult central-unipotent terms of the Arthur-Selberg formula, as in this case.
Our conjecture reduces the computation of discrete series multiplicities to the computation of stable elliptic orbital integrals of various transfers $f_p^H$, written for functions on $G(\Qp)$. Let us write this as $SO_{\gm_H}(f_p^H)$. Here $f_p$ are characteristic functions of congruence subgroups of $G(\Qp)$ related to $\Gamma$. Certainly at suitably regular elements, $SO_{\gm_H}(f_p^H)$ is an unstable combination of orbital integrals of $f_p$, however there are also contributions from elliptic singular $\gm_H$, notably $\gm_H=1$. At present, there are expressions for $f_p^H$ in the parahoric case and of course for $G(\Zp)$, but less seems to be known for smaller congruence subgroups. On the other hand, there are many formulas for dimensions of Siegel cusp forms and discrete series multiplicities for these cases (e.g. [@Wak; @Mult]). This suggests that one could predict stable singular elliptic orbital integrals $SO_{\gm_H}(f_p^H)$ for the transfer $f_p^H$ of characteristic functions of congruence subgroups (e.g. Klingen, Iwahori, Siegel), by comparing our formulas.
We now describe the layout of this paper.
In Section \[prelim\] we set up the conventions for this study. We explain how we are setting up the orbital integrals, and indicate our main computational tools. We also review the Langlands correspondence for real groups.
The theory of Arthur’s $\Phi$-function is reviewed in Section \[APF\]. In Section \[Kottwitz’s Formula\], we review Kottwitz’s stable version of Arthur’s formula from [@SVAF]. We also state our conjecture here. The heart of the volume computations in this paper is in Section \[Euler Characteristics\], where we determine $\chi_K(G)$. As a warm-up, we work out the classic case of $\SL_2$, with $\Gamma=\SL_2(\Z)$ in Section \[Special Linear\].
The case of $G=\GSp_4$ is considerably more difficult. We must work out several isomorphisms of real tori. These are described in Section \[Real Tori\]. The basic structure of $G$ and its Langlands dual $\hat{G}$ is set up in Section \[Structure of\]. In Section \[Discrete Series for G\] we work out the Langlands parameters for discrete series of $G(\R)$. There is only one elliptic endoscopic group $H$ for $G$. We describe $H$ in Section \[Endoscopic\]. In Section \[Discrete Series for H\], we describe the Langlands parameters for discrete series of $H(\R)$ and describe the transfer of discrete series in this case. In Section \[Levi Subgroups\] we describe the Levi subgroups of $G$ and $H$ and compute various constants which occur in Kottwitz’s formula for these groups. In Section \[Computing for G Levis\] we compute explicitly Arthur’s $\Phi$-function for Levi subgroups of $G$, and we do this for Levi subgroups of $H$ in Section \[Computing for H Levis\]. In Section \[Final Central\], we write out the terms of Kottwitz’s formula corresponding to central elements of $G$ and $H$, for a general arithmetic subgroup $\Gamma$. In Section \[Integral Points\], we specialize to the case of $\Gamma=\GSp_4(\Z)$, and in Section \[Comparison\] we gather our results to demonstrate Theorem \[Theorem\].
[**Acknowledgements:**]{} This paper is founded on the author’s thesis under the direction of Robert Kottwitz. The author would like to thank him for his continual help with this project. This paper is also indebted to Satoshi Wakatsuki for predicting Theorem \[Theorem\], and for much useful correspondence. We would also like to thank Ralf Schmidt for helpful conversations.
Preliminaries and Notation {#prelim}
==========================
We denote by $\Aff$ the ring of adeles over $\Q$. We denote by $\Aff_f$ the ring of finite adeles over $\Q$, so that $\Aff=\Aff_f \times \R$. Write $\OO_f$ for the integral points of $\Aff_f$.
If $G$ is a real Lie group, we write $G^+$ for the connected component of $G$ (using the classical topology rather than any Zariski topology).
Let $G$ be a connected reductive group over $\R$. A torus $T$ in $G$ is elliptic if $T/A_G$ is anisotropic (as an $\R$-torus). Say that $G$ is cuspidal if it contains a maximal torus $T$ which is elliptic. An element of $G(\R)$ is elliptic if it is contained in an elliptic maximal torus of $G$. Having fixed an elliptic maximal torus $T$, the absolute Weyl group $\Omega_G$ of $T$ in $G$ is the quotient of the normalizer of $T(\C)$ in $G(\C)$ by $T(\C)$. The real Weyl group $\Omega_{G,\R}$ of $T$ in $G$ is the quotient of the normalizer of $T(\R)$ in $G(\R)$ by $T(\R)$. We may drop the subscript “$G$” if it is clear from context. Also fix a maximal compact subgroup $K_{\R}$ of $G(\R)$.
Write $q(G)$ for half the dimension of $G(\R)/K_{\R}Z(\R)$. If we write $R$ for the roots of $G$, with a set of positive roots $R^+$, then $$q(G)=\half \left( |R^+|+ \dim(X) \right),$$ where $X$ is the span of $R$.
If $F$ is a field, write $\Gamma_F$ for the absolute Galois group of $F$. Suppose $G$ is an algebraic group over $F$. If $E$ is an extension field of $F$, we write $G_E$ for $G$ viewed as an algebraic group over $E$ (by restriction). If $\gm$ is an element of $G(F)$, we denote by $G_\gm$ the centralizer of $\gm$ in $G$. By $G^\circ$ we denote the identity component of $G$ (with the Zariski topology). Write $G_{\der}$ for the derived group of $G$. If $G$ is a reductive group, write $G_{\simp}$ for the simply connected cover of $G_{\der}$. Let $X^*(G)=\Hom(G_{\ol F},\Gm)$ and $X_*(G)=\Hom(\Gm,G_{\ol F})$. These are abelian groups. Write $X^*(G)_{\C}$ and $X_*(G)_{\C}$ for the tensor product of these groups over $\Z$ with $\C$. Similarly with the subscript “$\R$". Write $A_G(F)$ for the maximal $F$-split torus in the center of $G$.
If $G$ is an algebraic group over $\Q$, let $G(\Q)^+=G(\R)^+ \cap G(\Q)$.
Endoscopy
---------
In this section we review the theory of based root data and endoscopy in the form we will use in this paper.
The notion of a based root datum is defined in [@Springer]. Briefly, it is a quadruple $\Psi=(X, \Delta, X^{\vee}, \Delta^{\vee})$ where:
- $X$ and $X^{\vee}$ are free, finitely generated abelian groups, in duality by a pairing $$\lip, \rip: X \times X^{\vee} \to \Z.$$
- $\Delta$ (resp. $\Delta^{\vee}$) is a finite subset of $X$ (resp. of $X^{\vee}$).
- There is a bijection $\alpha \mapsto \alpha^{\vee}$ from $\Delta$ onto $\Delta^{\vee}$.
- For all $\alpha$, we have $\lip \alpha, \alpha^{\vee} \rip=2$.
- If $s_{\alpha}$ (resp. $s_{\alpha^{\vee}}$) is the reflection of $X$ (resp., of $X^{\vee}$) determined by $\alpha$ and $\alpha^{\vee}$, then $s_{\alpha}(\Delta) \subset \Delta$ (resp., $s_{\alpha^{\vee}}(\Delta^{\vee}) \subset \Delta^{\vee}$).
The dual of $\Psi=(X, \Delta, X^{\vee}, \Delta^{\vee})$ is given simply by $\Psi^{\vee}=(X^{\vee}, \Delta^{\vee}, X, \Delta)$.
Let $\Psi=(X, \Delta, X^{\vee}, \Delta^{\vee})$ and $\Psi'=(X', \Delta', X^{' \vee}, \Delta^{' \vee})$ be two root data. Then an isomorphism between $\Psi$ and $\Psi'$ is a homomorphism $f: X \to X'$ so that $f$ induces a bijection of $\Delta$ onto $\Delta'$ and so that the transpose of $f$ induces a bijection of $\Delta^{\vee}$ onto $\Delta^{' \vee}$.
Let $G$ be a connected reductive group over an algebraically closed field $F$. Fix a maximal torus $T$ and a Borel subgroup $B$ of $G$ with $T \subseteq B$. We say in this situation that $(T,B)$ is a pair (for $G$). The choice of pair determines a based root datum $$\Psi_0(G,T,B)=(X^*(T), \Delta(T,B), X_*(T), \Delta^{\vee}(T,B))$$ for $G$. Here $\Delta(T,B)$ is the set of simple $B$-positive roots of $T$, and $\Delta^{\vee}(T,B)$ is the set of simple $B$-positive roots of $T$. If another pair $T' \subseteq B'$ is chosen, the new based root datum obtained is canonically isomorphic to the original via an inner automorphism $\alpha$ of $G$. We have $\alpha(T')=T$ and $\alpha(B')=B$. Although the inner automorphism $\alpha$ need not be unique, its restriction to an isomorphism $T' \isom T$ is unique.
We may remove the dependence of the based root datum on the choice of pair as follows. Write $X^*$ (resp. $\Delta$, $X_*$, and $\Delta^{\vee}$) for the inverse limit over the set of pairs $(T,B)$ of $X^*(T)$ (resp. $\Delta(T,B)$, $X_*(T)$, $\Delta^{\vee}(T,B)$). Then we simply define the based root datum of $G$ to be $$\Psi_0(G)=(X^*,\Delta, X_*, \Delta^{\vee}).$$
Conversely, given a based root datum, the group $G$ over $F$ is uniquely determined up to isomorphism.
Let $G$ be a connected reductive group over a field $F$, and $\Psi_0(G)$ a based root datum of $G_{\ol F}$. Then $\Gamma_F$ acts naturally (via isomorphisms) on $\Psi_0(G)$. The action of $\Gamma_F$ on $G$ is said to be an $L$-action if it fixes some splitting of $G$ (see Section 1.3 of [@K84]).
Conversely, given a based root datum with a $\Gamma_F$-action, the group $G$ over $F$ is uniquely determined up to isomorphism.
A dual group for $G$ is the following data:
1. A connected complex reductive group with a based root datum $\Psi_0(\hat{G})$. We write its complex points as $\hat{G}$.
2. An $L$-action of $\Gamma_{F}$ on $\hat{G}$.
3. A $\Gamma_{F}$-isomorphism from $\Psi_0(\hat{G})$ to the dual of $\Psi_0(G)$.
To specify the isomorphism for iii) above, one typically fixes pairs $(T_0,B_0)$ of $G$ and $(\hat S_0, \hat B_0)$ of a dual group $\hat G$ and an isomorphism from $\Psi_0(\hat{G},\hat S_0, \hat B_0)$ to the dual of $\Psi_0(G,T_0,B_0)$.
In the case that $G$ is a torus $T$, the dual group $\hat{T}$ is simply given by $$\label{dual torus}
\hat{T}=X^*(T) \otimes_{\Z} \C^{\times},$$ with the $\Gamma_F$-action induced from $X^*(T)$. There are canonical $\Gamma_F$-isomorphisms $X^*(\hat T) \isom X_*(T)$ and $X_*(\hat T) \isom X^*(T)$.
The formalism for dual groups encodes canonical isomorphisms between tori. If $T$ and $T'$ are tori, and $\varphi: T \to T'$ is a homomorphism, it induces a homomorphism $\hat{T'} \to \hat{T}$ in the evident way.
Suppose that $(T,B)$ is a pair for $G$ and $(\hat S, \hat B)$ is a pair for $\hat{G}$. By iii) above, one has an in particular a fixed isomorphism from $\Psi_0(G,T,B)$ to the dual of $\Psi_0(\hat G, \hat S, \hat B)$. In particular this yields an isomorphism from $X^*(T)$ to $X_*(\hat S)$, which induces an isomorphism $$\label{determined}
\hat T \isom \hat S.$$
Next, let $G$ be a connected reductive group over a field $F$, which is either local or global.
An endoscopic group for $G$ is a triple $(H,s, \eta)$, where
- $H$ is a quasi-split connected group, with a fixed dual group $\hat{H}$ as above.
- $s \in Z(\hat{H})$.
- $\eta: \hat{H} \to \hat{G}$ is an embedding.
- The image of $\eta$ is $(\hat{G})_{\eta(s)}^{\circ}$, the connected component of the centralizer in $\hat{G}$ of $\eta(s)$.
- The $\hat{G}$-conjugacy class of $\eta$ is fixed by $\Gamma_F$.
Cohomology of $\Gamma_F$-modules then yields a boundary map $$\left[ Z(\hat{H})/ Z(\hat{G}) \right]^{\Gamma_F} \to H^1(F,Z(\hat{G})).$$
- The image of $s$ in $Z(\hat{H})/Z(\hat{G})$ is fixed by $\Gamma$, and its image under the above boundary map is trivial if $F$ is local and locally trivial if $F$ is global.
An endoscopic group is elliptic if the identity components of $Z(\hat{G})^{\Gamma_F}$ and $Z(\hat{H})^{\Gamma_F}$ agree.
The notion of an isomorphism of endoscopic groups is defined in Section 7.5 of [@K84]; we do not review it here.
Langlands Correspondence {#Langlands packets}
------------------------
Let $G$ be a connected reductive group over $\R$. In this section we review elliptic Langlands parameters for $G$ and the corresponding $L$-packets for discrete series representations of $G(\R)$. Our main references are [@Borel] and [@K90]. Write $W_\R$ for the Weil group of $\R$, and $W_{\C}$ for the canonical image of $\C^{\times}$ in $W_{\R}$. There is an exact sequence $$1 \to W_{\C} \to W_\R \to \Gamma_\R \to 1.$$ The Weil group $W_\R$ is generated by $W_{\C}$ and a fixed element $\tau$ satisfying $\tau^2=-1$ and $\tau z \tau^{-1}=\ol{z}$ for $z \in W_{\C}$. The action of $\Gamma_{\R}$ on $\hat{G}$ inflates to an action of $W_{\R}$ on $\hat{G}$, and through this action we form the $L$-group ${}^LG=\hat{G} \rtimes W_\R$.
A Langlands parameter $\varphi$ for $G$ is an equivalence class of continuous homomorphism $\varphi: W_\R \to {}^LG$ commuting with projection to $\Gamma_{\R}$, satisfying a mild hypothesis on the image (see [@Borel]). The equivalence relation is via inner automorphisms from $\hat{G}$. One associates to a Langlands parameter $\varphi$ an $L$-packet $\Pi(\varphi)$ of irreducible admissible representations of $G$.
Suppose that $G$ is cuspidal, so that there is a discrete series representation of $G(\R)$. This implies that the longest element $w_0$ of the Weyl group $\Omega$ acts as $-1$ on $X_*(T)$. If $\varphi$ is a Langlands parameter, write $C_{\varphi}$ for the centralizer of $\varphi(W_{\R})$ in $\hat{G}$ and $\hat{S}$ for the centralizer of $\varphi(W_{\C})$ in $\hat{G}$. Write $S_{\varphi}$ for the product $C_{\varphi}Z(\hat{G})$. We say $\varphi$ is elliptic if $S_\varphi/Z(\hat{G})$ is finite, and describe the $L$-packet $\Pi(\varphi)$ in this case.
Since $\varphi$ is elliptic, the centralizer $\hat{S}$ is a maximal torus in $\hat{G}$. Since $\varphi$ commutes with the projection to $\Gamma_{\R}$, it restricts to a homomorphism $$W_{\C} \to \hat{S} \times \{ 1 \}.$$ We may view this restriction as a continuous homomorphism $\varphi: \C^{\times} \to \hat{S}$, which may be written in exponential form $$\varphi(z)=z^{\mu} \ol{z}^{\nu}$$ with $\mu$ and $\nu$ regular elements of $X_*(\hat{T})_{\C}$. Write $\hat{B}$ for the unique Borel subgroup of $\hat{G}$ containing $\hat{S}$ so that $\lip \mu, \alpha \rip$ is positive for every root $\alpha$ of $\hat{S}$ that is positive for $\hat{B}$. We say that $\varphi$ determines the pair $(\hat{S}, \hat{B})$, at least up to conjugacy in $\hat{G}$.
Let $B$ be a Borel subgroup of $G_\C$ containing $T$. Then $\varphi$ and $B$ determine a quasi-character $\chi_B=\chi(\varphi,B)$, as follows. There is a canonical (up to $\hat{G}$-conjugacy) homomorphism $\eta_B: {}^LT \to {}^LG$ described in [@K90] so that for $z \in W_{\C}$, $$\eta_B(z)=z^\rho \ol{z}^{-\rho} \times z \in \hat{G} \rtimes W_\R.$$ Here $\rho=\rho_G$ is the half sum of the $B$-positive roots for $T$. Then a Langlands parameter $\varphi_B$ for $T$ may be chosen so that $\varphi=\eta_B \circ \varphi_B$. Finally $\chi_B$ is the quasi-character associated to $\varphi_B$ by the Langlands correspondence for $T$ (as described in Section 9.4 of [@Borel]).
Write $\BB$ for the set of Borels of $G_\C$ containing $T$. The $L$-packet associated to $\varphi$ is indexed by $\Omega_\R \backslash
\BB$. For $B \in \Omega_\R \backslash \BB$, a representation $\pi(\varphi,B)$ in the $L$-packet is given by the irreducible discrete series representation of $G(\R)$ whose character $\Theta_\pi$ is given on regular elements $\gm$ of $T(\R)$ by
$$(-1)^{q(G)} \sum_{\omega \in \Omega_\R} \chi_{\omega(B)}(\gm) \cdot \Delta_{\omega(B)}(\gm)^{-1}.$$
Here $\Delta_B$ is the usual discriminant $$\Delta_B(\gm)= \prod_{\alpha >0 \text{ for $B$}} (1-\alpha(\gm)^{-1}).$$
Finally, let $$\Pi(\varphi)= \{ \pi(\varphi,B) \mid B \in \Omega_\R \backslash
\BB \}.$$
It has order $d(G)=|\Omega / \Omega_\R|$. There is a unique irreducible finite-dimensional complex representation $E$ of $G(\C)$ with the same infinitesimal character and central character as the representations in this $L$-packet. It has highest weight $\mu-\rho \in X^*(T)$ with respect to $B$. The isomorphism classes of such $E$ are in one-to-one correspondence with elliptic Langlands parameters $\varphi$, and we often write $\Pi_E$ for $\Pi(\varphi)$.
We say that a discrete series representation $\pi \in \Pi_E$ is regular if the highest weight of $E$ is regular.
Measures and Orbital Integrals {#algorithm}
------------------------------
Let $G$ be a locally compact group with Haar measure $dg$. If $f$ is a continuous function on $G$, write $fdg$ for the measure on $G$ given by $$\varphi \mapsto \int_G \varphi(g)f(g)dg,$$ for $\varphi$ continuous and compactly supported in $G$. We will refer to the measures obtained in this way simply as “measures". If $G$ is a $p$-adic, real, or adelic Lie group, we require that $f$ be suitably smooth.
In this paper, we will view orbital integrals and Fourier transforms as distributions defined on measures, rather than on functions. This approach eases their dependence on choices of local measures, choices which do not matter in the end.
If $K$ is an open compact subset of $G$, then write $e_K$ for the measure given by $fdg$, where $f$ is the characteristic function of $K$ divided by $\vol_{dg}(K)$. Note that the measure $e_K$ is independent of the choice of Haar measure $dg$.
Let $G$ be a reductive group defined over a local field $F$. Fix a Haar measure $dg$ on $G(F)$. Let $fdg$ be a measure on $G(F)$, and take a semisimple element $\gamma \in G(F)$. Fix a Haar measure $dt$ of $G(F)_{\gm}^{\circ}$. Then we write $O_\gm(f dg; dt)$ for the usual orbital integral
$$O_\gm(f dg; dt)=\int_{G_\gm^{\circ}(F) \bks G(F)} f(g^{-1} \gm g) \frac{dg}{dt}.$$
Many cases of finite orbital integrals are easy to compute by the following result, a special case extracted from Section 7 of [@K86]:
\[orbitalintegrals\] Let $F$ be a $p$-adic field with ring of integers $\OO$. Let $G$ be a split connected reductive group defined over $\OO$, and $K=G(\OO)$. Suppose that $\gm \in K$ is semisimple, and that $1-\alpha(\gm)$ is either $0$ or a unit for every root $\alpha$ of $G$. Let $\gm'$ be stably conjugate to $\gm$. Then $O_{\gm'}(e_{K}; dt)$ vanishes unless $\gm'$ is conjugate to $\gm$, in which case
$$O_{\gm'}(e_{K}; dt)=\vol_{dt}(G_\gm^{\circ}(F) \cap K)^{-1}.$$
Now let $G$ be a reductive group defined over $\Q$.
Let $f^\infty dg_f$ be a measure on $G(\Aff_f)$ and take a semisimple element $\gm \in G(\Aff_f)$. Fix a Haar measure $dt_f$ of $G_{\gm}^{\circ}(\Aff_f)$. Write $O_\gm(f^\infty dg_f; dt_f)$ for the orbital integral
$$O_\gm(f^\infty dg_f; dt_f)=\int_{G_\gm^{\circ}(\Aff_f) \bks G(\Aff_f)} f^{\infty}(g^{-1} \gm g) \frac{dg_f}{dt_f}.$$
We also have the stable orbital integrals $$SO_\gm(f^\infty dg_f; dt_f)= \sum_i e(\gm_i) O_{\gm_i}(f^\infty dg_f; dt_{i,f}),$$ the sum being over $\gm_i \in G(\Aff_f)$ (up to $G(\Aff_f)$-conjugacy) whose local components are stably conjugate to $\gm$. The centralizers of $\gm$ and a given $\gm_i$ are inner forms of each other, and we use corresponding measures $dt_f$ and $dt_{i,f}$. The number $e(\gm_i)$ is defined as follows: Recall that for a reductive group $A$ over a local field, Kottwitz has defined an invariant $e(A)$ in [@Kottwitz; @sign]. It is equal to $1$ if $A$ is quasi-split. For each place $v$ of $\Q$, write $\gm_{i,v}$ for the $v$th component of $\gm_i$. Let $$e(\gm_{i,v})=e(G_{\gm_{i,v}}^{\circ}(\Q_v)).$$ Finally, let $$e(\gm_i)=\prod_v e(\gm_{i,v}).$$
Let $M$ be a Levi component of a parabolic subgroup $P$ of $G$, and $dm_f$ a Haar measure on $M(\Aff_f)$. Given a measure $f^\infty dg_f$, its “$M$-constant term” is the measure $f_M^{\infty} dm_f$, where $f_M^{\infty}$ is defined via $$f_M^\infty(m)= \delta_{P(\Aff_f)}^{-\half}(m) \int_{N(\Aff_f)} \int_{K_f} f^\infty(k^{-1}nmk)dk_fdn_f.$$ Here we fix the Haar measure $dk_f$ on $K_f$ giving it mass one, and the Haar measure $dn_f$ on $N(\Aff_f)$ is chosen so that $dg_f=dk_f dn_f dm_f$. The function $\delta_{P(\Aff_f)}$ is the modulus function on $P(\Aff_f)$.
It is independent of the choice of parabolic subgroup $P$.
Let $G$ be a split group defined over $\Z$ and let $K_f=G(\OO_f)$. Then $$(e_{K_f})_M= e_{M(\Aff_f) \cap K_f}.$$
Write $e_{K_f}=f^{\infty} dg_f$. Then it is easy to see that $f_M^{\infty}(m)=0$ unless $m \in K_f$. If $m \in K_f$, we compute that $$f_M^{\infty}(m)=\frac{\vol_{dk_f}(K_f) \vol_{dn_f}(K_f \cap N(\Aff_f))}{\vol_{dg_f}(K_f)}.$$ The result follows since $\vol_{dg_f}(K_f)=\vol_{dm_f}(M(\Aff_f) \cap K_f) \vol_{dn_f}(N(\Aff_f) \cap K_f) \vol_{dk_f}(K_f)$.
Pseudocoefficients {#PC}
------------------
We continue with a connected reductive group $G$ over $\Q$, and adopt some terminology from . Fix a maximal compact subgroup $K_{\R}$ of $G(\R)$. We put $K_{\R}'=K_{\R}A_G(\R)^+$. Given a quasicharacter (smooth homomorphism to $\C^{\times}$) $\xi$ on $A_G(\R)^+$, write $\mathcal{H}_{\ac}(G(\R),\xi)$ for the set of smooth, $K_{\R}'$-finite functions on $G(\R)$ which are compactly supported modulo $A_G(\R)^+$, and transform under $A_G(\R)^+$ according to $\xi$. Write $\Pi(G(\R),\xi)$ for the set of irreducible representations of $G(\R)$ whose central character restricted to $A_G(\R)^+$ is equal to $\xi$.
Given a function $f \in \mathcal{H}_{\ac}(G(\R),\xi^{-1})$, a representation $\pi \in \Pi(G(\R),\xi)$, and a Haar measure $dg_{\infty}$ on $G(\R)$, write $\pi(f dg_{\infty})$ for the operator on the space of $\pi$ given by the formula:
$$\pi(f dg_{\infty})= \int_{G(\R)/A_G(\R)^+} f(x) \pi(x) dg_{\infty}.$$
Here we give $A_G(\R)^+$ the measure corresponding to Lebesgue measure on $\R^n$, if $A_G$ is $n$-dimensional. The operator is of trace class.
Write $\Pi_{\tmp}(G(\R),\xi)$ (resp. $\Pi_{\disc}(G(\R),\xi))$ for the subset of tempered (resp. discrete series) representations in $\Pi(G(\R),\xi)$.
Suppose that $f \in \mathcal{H}_{\ac}(G(\R),\xi^{-1})$. Then we say that the measure $fdg_{\infty}$ is cuspidal if $\tr \pi(f dg_{\infty})$, viewed as a function on $\Pi_{\tmp}(G(\R),\xi)$, is supported on $\Pi_{\disc}(G(\R),\xi)$.
Write $\tilde{E}$ for the contragredient of the representation $E$. In , Arthur employs cuspidal measures $f_E \in \mathcal{H}_{\ac}(G(\R),\xi^{-1})$ whose defining property is that, for all $\pi \in \Pi_{\tmp}(G(\R),\xi)$, $$\label{packet ps}
\tr \pi(f_E dg_{\infty})=
\begin{cases} & (-1)^{q(G)} \text{ if } \pi \in \Pi_{\tilde{E}}, \\
& 0, \text{ otherwise}. \end{cases}$$
Such measures can be broken down further.
\[pseudoscience\] Fix a representation $\pi_0 \in \Pi_{\disc}(G(\R), \xi^{-1})$, and let $f_0 \in \mathcal{H}_{\ac}(G(\R),\xi^{-1})$. Suppose that the measure $f_0 dg_{\infty}$ satisfies, for all $\pi \in \Pi_{\tmp}(G(\R),\xi)$, $$\tr \pi(f_0 dg_{\infty})=
\begin{cases} & (-1)^{q(G)} \text{ if } \pi \cong \tilde \pi_0, \\
& 0, \text{ otherwise}. \end{cases}$$
It follows from the corollary in Section 5.2 of [@CD] that such functions exist. Pick such a function $f_0$, and put $$e_{\pi_0}=f_0 dg_{\infty}.$$
Suppose that for each $\pi \in \Pi_E$ we fix measures $e_{\pi}$ as above. Let $$f_E dg_{\infty} =\sum_{\pi} e_{\pi},$$ the sum being over $\pi \in \Pi_E$. Then clearly $f_E dg_{\infty}$ satisfies Arthur’s condition (\[packet ps\]).
[**Remark:**]{} The measure $(-1)^{q(G)}e_{\pi}$ is called a pseudocoefficient of $\tilde \pi$.
Transfer {#Transfer}
========
We sketch the important theory of transfer in the form that we will use in this paper.
Suppose that $G$ is a real connected reductive group, and that $(H,s,\eta)$ is an elliptic endoscopic group for $G$. Fix an elliptic maximal torus $T_H$ of $H$, an elliptic maximal torus $T$ of $G$, and an isomorphism $j: T_H \isom T$ between them. Also fix a Borel subgroup $B$ of $G_\C$ containing $T$ and a Borel subgroup $B_H$ of $H_\C$ containing $T_H$.
Suppose that $\xi$ is a quasicharacter on $A_G(\R)$, and that $f_{\infty} \in \mathcal{H}_{\ac}(G(\R),\xi^{-1})$, with $f_{\infty} dg_{\infty}$ cuspidal. There is a corresponding quasicharacter $\xi_H$ on $A_H(\R)$ described in Section 5.5 of [@SVAF].
There is also a measure $f_{\infty}^H dh_{\infty}$ on $H(\R)$ with $f_{\infty}^H \in \mathcal{H}_{\ac}(H(\R),\xi_H^{-1})$, having matching character values. (See [@Shelstad], [@CD], [@CD2], [@LS; @90].) More specifically, let $\varphi_H$ be a tempered Langlands parameter for $H_{\R}$, and write $\Pi_H=\Pi(\varphi_H)$ for the corresponding $L$-packet of discrete series representations of $H(\R)$. Transport $\varphi_H$ via $\eta$ to a tempered Langlands parameter $\varphi_G$ for $G$. The parameters $\varphi_G$ and $\varphi_H$ determine pairs $(\hat S, \hat B)$ and $(\hat S_H, \hat B_H)$ as in Section \[Langlands packets\].
Then $$\label{charidentity}
\tr \Pi_H(f_{\infty}^H dh_{\infty})= \sum_{\pi \in \Pi} \Delta_{\infty}(\varphi_H,\pi) \cdot \tr \pi(f_{\infty} dg_{\infty}),$$ using Shelstad’s transfer factors $\Delta_{\infty}(\varphi_H,\pi)$. Both sides of (\[charidentity\]) vanish unless $\Pi_H$ is a discrete series packet. In particular, $f_{\infty}^H dh_{\infty}$ is cuspidal, and it may be characterized by (\[charidentity\]). (The transfer $f_{\infty}^H dh_{\infty}$ is only defined up to the kernel of stable distributions.) We may use this formula to identify it as a combination of pseudocoefficients.
It is a delicate matter to specify the transfer factors. We will use a formula for $\Delta_{\infty}(\varphi_H,\pi)$ from [@K90], which is itself a reformulation of [@Shelstad]. One must carefully specify the duality between $G$ and $\hat{G}$, and between $H$ and $\hat{H}$, because this factor depends on precisely how this is done. It also depends on the isomorphism $j: T_H \isom T$, which must be compatible with correspondences of tori determined by the Langlands parameters, as specified below.
The triple $(j,B_T, B_{T_H})$ is aligned with $\varphi_H$ if the following diagram commutes:
$$\label{alignment}
\begin{CD}
\hat{T} @ >>> \hat{S} \\
@V \hat{j} V V @ AA \eta A \\
\hat{T}_H @ >>> \hat{S}_H
\end{CD}.$$
Here the isomorphism $\hat{T} \to \hat{S}$ (resp., $\hat{T}_H \to \hat{S}_H$) is determined, as in (\[determined\]), by $(B,\hat{B})$ (resp., $(B_H,\hat{B}_H)$). The map $\hat{j}$ is the dual map to $j$ using the identification (\[dual torus\]) of the dual tori.
For each $\omega \in \Omega$, there is a character $$a_{\omega}: \left( \hat{T} / Z(\hat{G}) \right)^{\Gamma_{\R}} \to \{ \pm 1 \}$$ described in [@K90].
If the triple $(j,B_T,B_{T_H})$ is aligned with $\varphi_H$, then we may take as transfer factors $$\Delta_{\infty}(\varphi_H, \pi(\varphi, \omega^{-1}(B)))= \lip a_{\omega}, \hat j^{-1}(s) \rip.$$
Next, let $G$ be a connected reductive algebraic group over $\Q$, and $(H,s,\eta)$ an endoscopic group for $G$. Given a measure $f^{\infty}dg_f$ on $G(\Aff_f)$, there is a measure $f^{\infty H} dh_f$ on $H(\Aff_f)$ so that for all $\gm_H \in H(\Aff_f)$ suitably regular, one has $$SO_{\gm_H}(f^{\infty H} dh_f )= \sum_{\gm} \Delta^{\infty}(\gm_H,\gm) O_{\gm}(f^{\infty} dg_f).$$ The sum is taken over $G(\Aff_f)$-conjugacy classes of “images" $\gm \in G(\Aff_f)$ of $\gm_H$. We have written $\Delta^{\infty}(\gm_H,\gm)$ for the the Langlands-Shelstad transfer factors. One takes matching measures on the centralizers of $\gm_H$ and the various $\gm$ in forming the quotient measures for the orbital integrals. We have left out many details; please see Langlands-Shelstad [@LS; @90] and Kottwitz-Shelstad [@KS] for definitions, and Ngô [@Ngo] for the celebrated proof.
Arthur’s $\Phi$-Function {#APF}
========================
In this section we consider a reductive group $G$ defined over $\R$. Let $T$ be a maximal torus contained in a Borel subgroup $B$ of $G_{\C}$. Let $A$ be the split part of $T$, let $T_e$ be the maximal elliptic subtorus of $T$, and $M$ the centralizer of $A$ in $G$. It is a Levi subgroup of $G$ containing $T$. Let $E$ be an irreducible finite-dimensional representation of $G(\C)$, and consider the $L$-packet $\Pi_E$ of discrete series representations $\pi$ of $G(\R)$ which have the same infinitesimal and central characters as $E$. Write $\Theta_\pi$ for the character of $\pi$, and put
$$\Theta^E= (-1)^{q(G)} \sum_{\pi \in \Pi_E} \Theta_\pi.$$
Note that $\Theta^E(\gm)$ will not extend continuously to all elements $\gm \in
T(\R)$, in particular to $\gm= 1$. Define the function $D^G_M$ on $T$ by $$D^G_M(\gm) = \det(1-\Ad(\gm);\Lie(G)/\Lie(M)).$$
Then a result of Arthur and Shelstad states that the function $$\gm \mapsto |D^G_M(\gm)|^\half \Theta^E(\gm),$$ defined on the set of regular elements $T_{\reg} (\R)$, extends continuously to $T(\R)$. We denote this extension by $\Phi_M(\gm,\Theta^E)$. The following closed expression for $\Phi_M(\gm,\Theta^E)$ when $\gm \in T_e$ is given in [@SpalPhi].
\[MPI\] If $\gm \in T_e(\R)$, then $$\label{SpalPhiformula}
\Phi_M(\gm, \Theta^E)=(-1)^{q(L)}|\Omega_L| \sum_{\omega \in \Omega^{LM}} \eps(\omega) \tr(\gm; V^M_{\omega(\lambda_B+\rho_B)-\rho_B}).$$
In particular,
1. If $T$ is elliptic then $M=G$ and $\Phi_G(\gm,\Theta^E)= \tr(\gm; E)$.
2. If $T$ is split and $z \in A_G(\R)$ then $M=A$ and $\Phi_A(z,\Theta^E)=(-1)^{q(G)} |\Omega_G| \lambda_0(z)$.
The notation needs to be explained. Here $L$ is the centralizer of $T_e$ in $G$. The roots of $T$ in $L$ (resp. $M$) are the real (resp. imaginary) roots of $T$ in $G$. Write $\Omega_L$ and $\Omega_M$ for the respective Weyl groups. Write $\Omega^{LM}$ for the set of elements which are simultaneously Kostant representatives for both $L$ and $M$, relative to $B$. We write $\eps$ for the sign character of $\Omega_G$. Finally by $V^M_{\omega(\lambda_B+\rho_B)-\rho_B}$ we denote the irreducible finite-dimensional representation of $M(\C)$ with highest weight $\omega(\lambda_B+\rho_B)-\rho_B$, where $\lambda_B$ is the $B$-dominant highest weight of $E$. Finally, $\lambda_0$ is the character by which $A_G(\R)$ acts on $E$.
For the case of central $\gm=z$, computing $\Phi_M(z,\Theta^{E})$ amounts to computing the dimensions of finite-dimensional representations of $M(\C)$ with various highest weights. For this we use the Weyl dimension formula, in the following form:
\[Weyl\] (Weyl Dimension Formula) Let $G$ be a complex reductive group, $T$ a maximal torus in $G$, contained in a Borel subgroup $B$. Write $\rho_B$ for the half-sum of the positive roots for $T$ in $G$ (with respect to $B$). Let $\lambda_B \in X^*(T)$ be a positive weight. Then there is a unique irreducible representation $V_{\lambda_B}$ of $G$ with highest weight $\lambda_B$. Its dimension is given by $$\dim_\C V_{\lambda_B}= \prod_{\alpha >0} \frac{ \lip \alpha, \lambda_B+ \rho_B \rip}{\lip \alpha, \rho_B \rip}.$$
Here $\lip, \rip$ is a $\Omega_G$-invariant inner product on $X^*(T)_\R$, which is unique up to a scalar.
Kottwitz’s Formula {#Kottwitz's Formula}
==================
Various Invariants {#Various Invariants}
------------------
In this section we introduce some invariants involved in Kottwitz’s formula.
By $\ol{G}$ we generally denote an inner form of $G_{\R}$ such that $\ol{G}/A_G$ is anisotropic over $\R$.
Let $G$ be a cuspidal reductive group over $\R$, and $dg_{\infty}$ a Haar measure on $G(\R)$. Let $$\ol{v}(G; dg_{\infty})=e(\ol{G}) \vol(\ol{G}(\R)/ A_G(\R)^+).$$
This is a stable version of the constant $v(G)$ which appears in .
Here $e(\ol{G})$ is the sign defined in [@Kottwitz; @sign]. (Note that $e(\ol{G})=(-1)^{q(G)}$ when $G$ is quasisplit.) In both cases the Haar measure on $\ol{G}(\R)$ is transported from $dg_\infty$ on $G(\R)$ in the usual way, and the measure on $A_G(\R)^+$ is the standard Lebesgue measure.
Let $G$ be a cuspidal connected reductive group over $\Q$. Then $G$ contains a maximal torus $T$ so that $T/ A_G$ is anisotropic over $\R$. Write $T_{\simp}$ for the inverse image in $G_{\simp}$ of $T$. Then $k(G)$ is the cardinality of the set $\im[ H^1(\R,T_{\simp}) \to H^1(\R,T)]$.
If $G$ is a reductive group over $\Q$, write $\tau(G)$ for the Tamagawa number of $G$, as defined in [@Ono].
By [@K88] or [@SVAF], the Tamagawa numbers $\tau(G)$ for a reductive group $G$ over $\Q$ may be computed using the formula $$\tau(G)=|\pi_0(Z(\hat{G})^{\Gamma_\Q})| \cdot |\ker^1(\Q,Z(\hat{G}))|^{-1}.$$
Here $\pi_0$ denotes the topological connected component.
Let $M$ be a Levi subgroup of $G$. Then put $$n^G_M=[N_G(M)(\Q) : M(\Q)].$$
Here $N_G(M)$ denotes the normalizer of $M$ in $G$.
Let $\gm \in M(\Q)$ be semisimple. Then put $$\ol{\iota}^M(\gm)=|(M_{\gm} / M_{\gm}^{ \circ})(\Q)|$$ and $$\iota^M(\gm)=[M_{\gm}(\Q) : M_{\gm}^{ \circ}(\Q)].$$
Let $(H,s,\eta)$ be an endoscopic triple for $G$, and write $\Out(H,s,\eta)$ for its outer automorphisms. Put $$\iota(G,H)= \tau(G) \tau(H)^{-1} |\Out(H,s,\eta)|^{-1}.$$
The Formula
-----------
In this section we give Kottwitz’s formula from [@SVAF].
Our $G$ will now be a cuspidal connected reductive group over $\Q$. Let $f^{\infty} \in C_c^{\infty}(G(\Aff_f))$ and $f_{\infty} \in \mathcal{H}_{\ac}(G(\R),\xi)$ for some $\xi$. We consider measures $fdg$ of the form $fdg=f^{\infty}dg_f \cdot f_{\infty}dg_{\infty} \in C_c^{\infty}(G(\Aff))$, for some decomposition $dg=dg_f dg_{\infty}$ of the Tamagawa measure on $G(\Aff_f)$. Also choose such decompositions for every cuspidal Levi subgroup $M$ of $G$.
First we define the stable distribution $S\Phi_M$ at the archimedean place:
Let $M$ be a cuspidal Levi subgroup of $G$. Let $\gm \in M(\Q)$ be elliptic, and pick a Haar measure $dt_{\infty}$ of $M_{\gm}^{\circ}(\R)$. Then $S \Phi_M(\gm,f_\infty dg_\infty; dt_{\infty})$ is defined to be $$(-1)^{\dim(A_M/A_G)} k(M)k(G)^{-1} \ol{v}(M_\gm^{\circ}; dt_{\infty})^{-1} \sum_\Pi \Phi_M(\gm^{-1},\Theta_\Pi) \tr \Pi(f_\infty dg_\infty),$$
the sum being taken over $L$-packets of discrete series representations.
Here is the basic building block of Kottwitz’s formula:
Let $M$ be a cuspidal Levi subgroup of $G$, and $\gm \in M(\Q)$ an elliptic element. Pick Haar measures $dt_f$ on $M_{\gm}^{\circ}(\Aff_f)$ and $dt_{\infty}$ on $M_{\gm}^{\circ}(\R)$ whose product is the Tamagawa measure $dt$ on $M_{\gm}^{\circ}(\Aff)$.
We define $$ST_g(fdg,\gm, M)= (n^G_M)^{-1} \tau(M) \ol{\iota}^M(\gm)^{-1} SO_\gm(f_M^\infty dm_f; dt_f) S \Phi_M(\gm,f_\infty dg_\infty; dt_{\infty}).$$
Here $f_M^{\infty}dm_f$ is the $M$-constant term of $f^{\infty}dg_f$. The product $SO_{\gm}(f_M^{\infty}dm_f; dt_f) \ol v(M; dt_{\infty})$ is independent of the decompositions of $dt$ and $dg$. We will therefore often write this simply as $SO_{\gm}(f_M^{\infty} dm_f) \ol v(M)$. Similarly for other such products.
Kottwitz defines $$ST_g(fdg)= \sum_M \sum_{\gm \in M} ST_g(fdg,\gm, M).$$
Here $M$ runs over $G(\Q)$-conjugacy classes of cuspidal Levi subgroups in $G$, and the second sum runs over stable $M(\Q)$-conjugacy classes of semisimple elements $\gm \in M(\Q)$ which are elliptic in $M(\R)$.
For convenience we also define, for $\gm \in G(\Q)$ semisimple, $$ST_g(f dg, \gm)= \sum_M ST_g( fdg, \gm, M),$$ The sum being taken over cuspidal Levi subgroups of $G$ with $\gm \in M(\Q)$.
Kottwitz’s stable version of Arthur’s trace formula is given by $$\Kot(f dg)=\sum_{(H,s,\eta) \in \funE_0} \iota(G,H) ST_g(f^H dh),$$ where $\funE_0$ is the set of (equivalence classes of) elliptic endoscopic groups for $G$.
We record here the simpler form of $ST_g(fdg, \gm, M)$ when $\gm=z$ is in the rational points $Z(\Q)$ of the center of $G$. We have $$ST_g(f dg, z, M)= (-1)^{\dim(A_M/A_G)} \frac{k(M)}{k(G)} (n^G_M)^{-1} \tau(M) f_M^\infty(z) \ol{v}(M; dm_{\infty})^{-1} \Phi_M(z^{-1}, \Theta_{\Pi}).$$
Conjecture {#Conjecture}
----------
Recall the stable cuspidal measure $f_E dg_{\infty}$ from Section \[PC\]. Fix any test function $f^{\infty} dg_f$ and put $f=f^{\infty}f_E dg$.
Let $$T_g(fdg)= \sum_M (n^G_M)^{-1} \sum_\gm \iota^M(\gm)^{-1} \tau(M_\gm) O_\gm(f_M^\infty dm_f) \Phi_M(\gm,f_E dg_{\infty}).$$
Here as in , $\Phi_M(\gm,-)$ is the unnormalized form of Arthur’s distribution $I_M$ defined in [@ITF1].
Now suppose that $\pi \in \Pi_{\disc}(G(\R), \xi)$, and let $K_f$ be an open compact subgroup of $G(\Aff_f)$. Write $$L^2(G(\Q) \bks G(\Aff) / K_f, \xi)$$ for the space of functions on this double coset space which transform by $A_G(\R)^+$ according to $\xi$ and are square integrable modulo center. Write $R_{\disc}(\pi, K_f)$ for the $\pi$-isotypical subspace of $L^2(G(\Q) \bks G(\Aff) / K_f, \xi)$; it is finite-dimensional. If $f^{\infty}dg_f$ is $K_f$-biinvariant, then convolution gives an operator $R_{\disc}(\pi,f^{\infty}dg_f)$ on $R_{\disc}(\pi, K_f)$. According to Arthur’s Corollary 6.2 of , if the highest weight of $E$ is regular, then
$$\sum_{\pi \in \Pi_E} \tr R_{\disc}(\pi,f^{\infty}dg_f) = T_g(fdg).$$
The main result of Kottwitz [@SVAF] is that when $f_{\infty} dg_{\infty}$ is stable cuspidal, then $$T_g(f dg)=\Kot(f dg).$$
Since we may assume $f_Edg_{\infty}= \sum_{\pi \in \Pi_E} e_{\pi}$, the following conjecture is plausible:
Fix a regular discrete series representation $\pi$ of $G(\R)$. Let $f_{\infty} dg_{\infty}=e_{\pi}$ as in Section \[PC\]. Pick a measure $f^{\infty} dg_f $ with $f^{\infty} \in C_c(G(\Aff_f))$, and $dg_f dg_{\infty}=dg$ the Tamagawa measure on $G(\Aff)$. Put $f=f^{\infty} f_{\infty}$. Then
$$\Kot(f dg)=\tr R_{\disc}(\pi,f^{\infty} dg_f).$$
In particular, if we pick a compact open subgroup $K_f$ of $G(\Aff_f)$, and put $f^\infty dg_f=e_{K_f}$, we obtain $$m_{\disc}(\pi,K_f)=\Kot(e_{\pi} e_{K_f}).$$
In this paper we give some evidence for this conjecture. Moreover, we will see that $\Kot(fdg)$ is given by a closed algebraic expression, which is straightforward to evaluate, so long as one can compute the transfers $e_{\pi}^H$ at the real place, and evaluate the elliptic orbital integrals of $f^{\infty H} dh_f$ at the finite adeles.
Euler Characteristics {#Euler Characteristics}
=====================
We have finished our discussion of Kottwitz’s formula, and now solve the arithmetic volume problem mentioned in the introduction. For simplicity we will write $K$ rather than $K_f$ for open compact subgroups of $G(\Aff_f)$ in this section.
For $K$ a compact open subgroup of $G(\Aff_f)$, we define $$\chi_{K}(G)=\ol{v}(G; dg_{\infty})^{-1} \vol_{dg_f}(K)^{-1} \tau(G) d(G),$$ if $G$ is cuspidal. If $G$ is not cuspidal, then $\chi_K(G)=0$.
Note that if $K_0$ is another compact open subgroup of $G(\Aff_f)$, with $K \subseteq K_0$ of finite index, then $\chi_K(G)=[K_0:K] \chi_{K_0}(G)$. In this section we compute the quantities $\chi_{K}(G)$ under some mild hypotheses on $G$.
Statement of Theorem
--------------------
Before getting embroiled in details, let us sketch the idea of the computation of $\chi_K(G)$. The computation is considerably easier if $K$ is sufficiently small. In this case, $\chi_K(G)$ is the classical Euler characteristic of a Shimura variety. This in turn may be written in terms of Euler characteristics of an arithmetic subgroup of $G_{\ad}(\R)$. For $G$ a semisimple and simply connected Chevalley group, such Euler characteristics were computed in [@Harder].
Our work is to reduce to this case. Given a compact open subgroup $K_0$ of $G(\Aff_f)$, we will pick a sufficiently small subgroup $K$ of $K_0$. By the above we know the analogue of $\chi_K(G)$ for $G^{\simp}$. To compute $\chi_{K_0}(G)$ we have two tasks: to transition between $G$ and $G^{\simp}$, and to transition between $K$ and $K_0$.
The resulting formula entails several standard definitions:
Write $G(\R)_+ \subseteq G(\R)$ for the inverse image of $G_{\ad}(\R)^+$. Let $G(\Q)_+=G(\Q) \cap G(\R)_+$. Write $\nu: G \twoheadrightarrow C$ for the quotient of $G$ by $G_{\der}$. Let $C(\R)^\dagger=\nu(Z(\R))$, and $C(\Q)^\dagger=C(\Q) \cap C(\R)^\dagger$. Write $\rho: G_{\simp} \to G_{\der}$ for the usual covering of $G_{\der}$ by $G_{\simp}$. For $K$ a compact open subgroup of $G(\Aff_f)$, let $K^{\der}=G_{\der}(\Aff_f) \cap K$, and let $K^{\simp}$ be the preimage of $K$ in $G_{\simp}(\Aff_f)$. Let $\Gamma_K=G(\Q)_+ \cap K$, let $\Gamma^{\der}_K=G_{\der}(\Q)_+ \cap K$, let $\Gamma^{\simp}_K=K^{\simp} \cap G_{\simp}(\Q)_+$, and write $\Gamma^{\ad}_K$ for the image of $\Gamma_K$ in $G_{\ad}(\Q)$.
In this section we avoid certain awkward tori for simplicity, preferring the following kind:
A torus $T$ over $\Q$ is $\Q \R$-equitropic if the largest $\Q$-anisotropic torus in $T$ is $\R$-anisotropic.
Here are some basic facts about $\Q \R$-equitropic tori.
If $T$ is a $\Q \R$-equitropic torus then $T(\Q)$ is discrete in $T(\Aff_f)$. If $G$ is a reductive group, and the connected component $Z^\circ$ of the center of $G$ is $\Q \R$-equitropic, then its derived quotient $C$ is also $\Q \R$-equitropic.
The first statement follows from Theorem 5.26 of [@Milne]. The second is straightforward.
In [@Serre], Serre introduces an Euler characteristic $\chi_{\alg}(\Gamma) \in \Q$ applicable to any group $\Gamma$ with a finite index subgroup $\Gamma_0$ which is torsion-free and has finite cohomological dimension. In particular, it applies to our congruence subgroups $\Gamma=\Gamma_K$. Here are some simple properties of $\chi_{\alg}$:
- For an exact sequence of the form $$1 \to A \to B \to C \to 1,$$ with $A,B$ and $C$ groups as above, we have $\chi_{\alg}(B)=\chi_{\alg}(A) \cdot \chi_{\alg}(C)$.
- If $\Gamma$ is a finite group, then $\chi_{\alg}(\Gamma)= |\Gamma|^{-1}$.
The theorem of this section relates $\chi_K(G)$ to $\chi_{\alg}(\Gamma^{\simp}_{K})$. More precisely :
\[chi-theorem\] Let $G$ be a reductive group over $\Q$. Assume that $G_{\simp}$ has no compact factors, and that the connected component $Z^{\circ}$ of the center of $G$ is $\Q \R$-equitropic. Let $K_0 \subset G(\Aff_f)$ be a compact open subgroup. Then $\chi_{K_0}(G)$ is equal to $$\frac{ |\ker (\rho(\Q))| [G_{\der}(\Aff_f):G_{\der}(\Q)_+ K^{\der}_0] \left[\Gamma^{\der}_{K_0}:G_{\der}(\Q)_+ \cap \rho(K_0^{\simp}) \right] \cdot [C(\Aff_f) : C(\Q)^\dagger \nu(K_0)]}{ [G(\R): G(\R)_+] |\nu(K_0) \cap C(\Q)^\dagger|} \chi_{\alg}(\Gamma^{\simp}_{K_0}).$$
Here $\rho(\Q)$ denotes the map $\rho(\Q): G_{\simp}(\Q) \to G(\Q)$ on $\Q$-points. The assumption on the absence of compact factors is needed for strong approximation, and is discussed in [@Milne].
When $G_{\simp}$ is a Chevalley group, and $\Gamma^{\simp}_{K_0}=G_{\simp}(\Z)$, this reduces the problem to the calculation of Harder [@Harder]:
\[H2\] Let $G$ be a simply connected, semisimple Chevalley group over $\Z$. Write $m_1, \ldots, m_r$ for the exponents of its Weyl group $\Omega$, and put $\Gamma=G(\Z)$. We have $$\chi_{\alg}(\Gamma)=\left(-\half \right)^r|\Omega_{\R}|^{-1} \prod_{i=1}^r B_{m_i+1}.$$
Here $B_n$ denotes the $n$th Bernoulli number. Recall that $\Omega_{\R}$ is the real Weyl group of $G$.
Shimura Varieties
-----------------
To prove Theorem \[chi-theorem\], we will use some basic Shimura variety theory, which may be found in [@Deligne] or [@Milne]. Much of the theory holds only for $K$ sufficiently small. For simplicity, we will say “$K$ is small” rather than “$K$ is a sufficiently small finite index subgroup of $K_0$”.
For convenience, we gather here many simplifying properties of small $K$, which we will often use without comment. For the rest of this section assume that $Z(G)^\circ$ is $\Q \R$-equitropic, and that $G_{\simp}$ has no compact factors.
\[small\] Let $K$ be small. Then
1. \[lily\] $K \cap Z(\Q)= \{1\}$.
2. $\nu(K) \cap C(\Q)= \{1\}$.
3. \[abby\] $G(\Q) \cap K G_{\der}(\Aff_f) \subseteq G_{\der}(\Q)$.
4. $G_{\der}(\Aff_f) \cap G(\Q) K= G_{\der}(\Q) K_{\der}$.
5. \[joy\] $K \cap G_{\der}(\Q) \subseteq \rho(G_{\simp}(\Q))$.
6. \[reu\] $K \cap G(\Q) \subseteq G(\Q)^+$.
The first two items follow because $Z^\circ$, and thus $C$, are $\Q \R$-equitropic. Item \[abby\] follows from Corollaire 2.0.12 in [@Deligne], and the next item is a corollary. Items \[joy\] and \[reu\] follow from Corollaire 2.0.5 and 2.0.14 in [@Deligne], respectively.
Recall that we have chosen a maximal compact subgroup $K_{\R}$ of $G(\R)$.
Let $$X=G(\R)/ K_{\R}^+ Z(\R),$$
$$\ol{X}=G(\R) / K_{\R} Z(\R),$$
and $$S_K=G(\Q) \bks X \times G(\Aff_f)/K$$ be the double coset space obtained through the action $q(x,g)k=(qx,qgk)$ of $q \in G(\Q)$ and $k \in K$.
Similarly, let $$\ol{S}_K=G(\Q) \bks \ol{X} \times G(\Aff_f)/K,$$ with the action of $G(\Q) \times K$ defined in the same way.
The component group of $S_K$ is finite and given (see 2.1.3 in [@Deligne]) by $$\label{pease}
\pi_0(S_K)= G(\Aff_f)/ G(\Q)_+ K.$$
There is some variation in the literature regarding the use of $X$ versus $\ol{X}$. Deligne and Milne implicitly use $X$ in [@Deligne] and [@Milne] (in light of Proposition 1.2.7 in [@Deligne]). Harder uses $\ol{X}$ in [@Harder]. Arthur uses $$G(\R)/K_{\R}'$$ in . (Recall that $K_{\R}'=A_G(\R)^{+}K_{\R}$.) Since for us $Z^{\circ}$ is $\Q \R$-equitropic, we have $$K_{\R}'=Z(\R)K_{\R},$$ and so this quotient is equal to $\ol{X}$.
Since we would like to combine results stated in terms of $X$ with others stated in terms of $\ol{X}$, we must understand the precise relationship between the two. This is the purpose of Proposition \[fibre computation\] below.
Let $G$ be a real group, and $Z$ its center. Write $$\ad: G(\R) \to G(\R)/Z(\R)$$ for the quotient map.
Note that $\ad(G(\R))$ has finite index in $G_{\ad}(\R)$.
\[real groups\] For this lemma, let $G$ be a Zariski-connected reductive real group, and $K_{\R}$ a maximal compact subgroup of $G(\R)$. Let $L_{\R}$ be a maximal compact subgroup of $G_{\ad}(\R)$ containing $\ad(K_{\R})$. Then the following hold:
1. \[do\] $K_{\R}$ meets all the connected components of $G(\R)$.
2. \[re\] $K_{\R} \cap G(\R)^+=K_{\R}^+$.
3. \[mi\] $\ad(K_{\R})$ is a maximal compact subgroup of $\ad(G(\R))$.
4. \[fa\] $\ad(K_{\R}^+)=L_{\R}^+$.
5. \[so\] $K_{\R}Z(\R) \cap G(\R)_+=K_{\R}^+Z(\R)$.
The first two statements follow from the Cartan decomposition (Corollary 4.5 in [@Satake]).
For the third statement, suppose that $C$ is a subgroup of $G(\R)$ with $\ad(K_{\R}) \subseteq \ad(C)$ and $\ad(C)$ compact. If $\ad(K_{\R}) \neq \ad(C)$, then there is an element $a \in CZ(\R)-K_{\R}Z(\R)$. By the Cartan decomposition, we may assume that $a=\exp(H)$, with $H$ a semisimple element of $\Lie(G)$, and $\alpha(H)$ real and nonnegative for every root $\alpha$ of $G$. Since $a \notin Z(\R)$, we have $\alpha(H)>0$ for some root $\alpha$. Thus $\ad(C)$ is not compact, a contradiction. Thus $\ad(K_{\R})=\ad(C)$, and statement \[mi\] follows.
For the fourth statement, note that $L_{\R} \cap \ad(G)=\ad(K_{\R})$, and so $L_{\R}/\ad(K_{\R})$ injects into $G_{\ad}(\R)/\ad(G(\R))$. It follows that $\ad(K_{\R}^+)$ has finite index in $L_{\R}$. Since it is connected, statement \[fa\] follows.
For the fifth statement, let $g \in K_{\R}Z(\R) \cap G(\R)_+$. Then $\ad(g) \in L_{\R} \cap G_{\ad}(\R)^+$, so by statement \[re\], we see $\ad(g) \in L_{\R}^+=\ad(K_{\R}^+)$. Thus $g \in K_{\R}^+Z(\R)$. The other inclusion is obvious.
\[fibre computation\]
1. The natural projection $p_X: X \to \ol{X}$ has fibres of order $[G(\R): G(\R)_+]$.
2. Let $X^+$ be a connected component of $X$. It is stabilized by $G(\R)_+$, and the restriction of $p_X$ to $X^+$ is a $G(\R)_+$-isomorphism onto $\ol{X}$.
3. Let $K$ be small. Then the natural projection $p_S: S_K \to \ol{S}_K$ has fibres of order $[G(\R): G(\R)_+]$.
Consider the natural map $$\label{amap}
K_{\R}Z(\R)/K_{\R}^+Z(\R) \to G(\R)/G(\R)_+.$$ It is surjective because $K_{\R}$ meets every connected component of $G(\R)$. It is injective because $K_{\R}Z(\R) \cap G(\R)_+ \subseteq K_{\R}^+Z(\R)$. It follows that (\[amap\]) is an isomorphism, and the first statement follows.
We now prove the second statement. Note that $p_X$ is both an open and closed map so that $p_X(X^+)$ is a component of $\ol{X}$. Since $K_{\R}$ meets every connected component of $G(\R)$, the set $\ol{X}$ is connected. Therefore $p_X(X^+)=\ol{X}$. By Proposition 5.7 of [@Milne], there are $[G(\R): G(\R)_+]$ connected components of $X$, each stabilized by $G(\R)_+$. Thus the fibre over a point in $\ol{X}$ is comprised of exactly one point from each component of $X$. So $p_X$ restricted to $X^+$ is an isomorphism; it is clear that it respects the $G(\R)_+$-action.
To prove the third statement, we require $K$ to be sufficiently small, in the following way. Suppose $K_*$ is an open compact subgroup of $G(\Aff_f)$ satisfying $K_* \cap G(\Q) \subseteq G(\Q)^+$. Let $g_1, \ldots, g_r$ be representatives of the finite quotient group $G(\Q) K_*\bks G(\Aff_f)$. Then we require that $$\label{amina}
K \subseteq \bigcap_{i=1}^r g_i^{-1}K_* g_i.$$
Now for $x \in X$, let $\Fib(x)$ be the fibre of $p_X$ containing $x$. If we further fix $g \in G(\Aff_f)$, let $\Fib(x,g)$ be the fibre of $p_S$ containing $(x,g)$. (Here we understand $(x,g)$ as an element of $S_K$.) We claim that for all such $x$ and $g$, the map $$\label{fibremap}
\Fib(x) \to \Fib(x,g)$$ given by $x' \mapsto (x',g)$ is a bijection. This will imply the third statement.
For surjectivity of (\[fibremap\]), pick $(x',g') \in \Fib(x,g)$. Then there are $q \in G(\Q)$ and $k \in G(\Aff_f)$ so that $qp_X(x')=p_X(x)$ and $qg'k=g$. Let $x''=qx'$. Then $x'' \in \Fib(x)$ and $(x'',g)=(x',g')$.
For injectivity of (\[fibremap\]), suppose that $(x_1,g)=(x_2,g)$ in $S_K$ with $x_1,x_2 \in \Fib(x)$. Then in particular, there is an element $q \in G(\Q)$ and $k \in K$ so that $qgk=g$ and $qx_1=x_2$. Write $g=q_0k_0g_i$ with $q_0 \in G(\Q)$ and $k_0 \in K_*$. Then we have $$q(q_0k_0g_i)k=q_0k_0g_i,$$ which we rewrite as $$q_0^{-1}q q_0= k_0 g_i k^{-1}g_i^{-1}k_0^{-1}.$$ Using this and (\[amina\]) we see that $q_0^{-1}qq_0 \in G(\Q) \cap K_* \subseteq G(\Q)^+$. Since $G(\Q)^+$ is normal in $G(\Q)$, in fact $q \in G(\Q)^+$.
Meanwhile, pick $\xi_1,\xi_2 \in G(\R)$ representing $x_1$ and $x_2$, respectively. Since $x_1,x_2 \in \Fib(x)$ we have $\xi_1^{-1}\xi_2 \in K_{\R}Z(\R)$. Write $\xi_2=\xi_1 kz$, with $k \in K_{\R}$ and $z \in Z(\R)$. Since $qx_1=x_2$, we have $\xi_2^{-1}q \xi_1 \in K_{\R}^+Z(\R)$, and therefore $z^{-1}k^{-1}\xi_1^{-1}q\xi_1 \in K^+_{\R}Z(\R)$. Using the fact that $q$ is in the normal subgroup $G(\R)_+$ of $G(\R)$, it follows that $k \in G(\R)_+ \cap K_{\R} \subseteq K_{\R}^+ Z(\R)$. Thus $x_1=x_2$, as desired.
We will need the following two known results:
(Harder, see [@Harder], [@Serre]) \[H1\] If $G$ is semisimple and $K$ is small, then $\chi_{\Top}( \Gamma_K \bks \ol{X})=\chi_{\alg}(\Gamma_K)$.
\[???\] (See , [@GKM]) If $K$ is small, then $\chi_K(G)=\chi_{\Top}(\ol{S}_K)$.
Computations
------------
The next three lemmas will allow us to convert our computation for $K_0$ to a computation for $K$.
\[frank\] If $K$ is small, then $$| C(\Q)^\dagger \bks C(\Aff_f) / \nu(K)|=[\nu(K_0): \nu(K)] |\nu(K_0) \cap C(\Q)^\dagger|^{-1} | C(\Q)^\dagger \bks C(\Aff_f) / \nu(K_0)|.$$
This follows from the exactness of the sequence $$1 \to \nu(K_0) \cap C(\Q)^\dagger \to \frac{\nu(K_0)}{\nu(K)} \to C(\Q)^\dagger \bks C(\Aff_f) / \nu(K) \to C(\Q)^\dagger \bks C(\Aff_f) / \nu(K_0) \to 1.$$
\[jeremiah\] If $K \subseteq K_0$ is small, then $$\label{jade}
[\Gamma^{\ad}_{K_0}: \Gamma^{\ad}_K]=\frac{ [\Gamma_{K_0}: \rho(\Gamma_{K_0}^{\simp})][K_0: K]}{|K_0 \cap Z(\Q)| [\nu(K_0) : \nu(K)] [K_0^{\der}: K^{\der} \rho(K_0^{\simp})] }.$$
In the proof we refer to conditions of Proposition \[small\].
Consider the map $$\Gamma^{\der}_{K_0}/ \Gamma^{\der}_K \to \Gamma^{\ad}_{K_0} / \Gamma^{\ad}_K.$$
The kernel of this map sits in the middle of the exact sequence $$1 \to \Gamma^{\der}_{K_0} \cap Z(\Q) \to (\Gamma_K Z(\Q) \cap \Gamma^{\der}_{K_0})/\Gamma^{\der}_K \to (\Gamma_K Z(\Q) \cap \Gamma^{\der}_{K_0}) \left/ \Gamma^{\der}_K (\Gamma^{\der}_{K_0} \cap Z(\Q)) \right. \to 1,$$ using condition \[lily\]. This last quotient is trivial, because actually $\Gamma_K=\Gamma_K^{\der}$ by condition \[abby\].
We have established the exactness of the sequence $$1 \to \Gamma^{\der}_{K_0} \cap Z(\Q) \to \Gamma^{\der}_{K_0} / \Gamma^{\der}_K \to \Gamma^{\ad}_{K_0} / \Gamma^{\ad}_K \to
\Gamma_{K_0}Z(\Q)/ \Gamma^{\der}_{K_0}Z(\Q) \to 1.$$ The last quotient is isomorphic to $$\Gamma_{K_0} \left / (Z(\Q) \cap K_0)\Gamma^{\der}_{K_0} \right.,$$ which itself sits inside the exact sequence $$1 \to K_0 \cap Z(\Q) \left / \Gamma^{\der}_{K_0} \cap Z(\Q) \right. \to \Gamma_{K_0} / \Gamma^{\der}_{K_0} \to
\Gamma_{K_0} \left / (Z(\Q) \cap K_0)\Gamma^{\der}_{K_0} \right. \to 1.$$
The quantity $|\Gamma^{\der}_{K_0} \cap Z(\Q)|$ cancels, and it follows that $$\label{ac/dc}
[\Gamma^{\ad}_{K_0}: \Gamma^{\ad}_K]=\frac{ [\Gamma^{\der}_{K_0}: \Gamma^{\der}_K] \cdot [\Gamma_{K_0}:\Gamma^{\der}_{K_0}]}{|K_0 \cap Z(\Q)|}.$$
By condition \[joy\] we have $$1 \to \rho(\Gamma_{K_0}^{\simp})/ \rho(\Gamma_K^{\simp}) \to \Gamma_{K_0}^{\der}/ \Gamma_K^{\der} \to \Gamma_{K_0}^{\der}/ \rho(\Gamma_{K_0}^{\simp}) \to 1.$$
Strong approximation tells us that $G_{\simp}(\Q)$ is dense in $G_{\simp}(\Aff_f)$. Therefore we have isomorphisms $$\rho(\Gamma_{K_0}^{\simp})/ \rho(\Gamma_K^{\simp}) \isom \Gamma^{\simp}_{K_0}/ \Gamma^{\simp}_K \isom K_0^{\simp}/ K^{\simp} \isom \rho(K_0^{\simp})/ \rho(K^{\simp}).$$ Combining this with the exact sequences $$1 \to K_0^{\der}/ K^{\der} \to K_0/K \to \nu(K_0)/\nu(K) \to 1$$ and $$\label{custer}
1 \to \rho(K_0^{\simp})/ \rho(K^{\simp}) \to K_0^{\der}/ K^{\der} \to K_0^{\der} \left / K^{\der} \rho(K_0^{\simp}) \right. \to 1,$$ we obtain $$[\Gamma_{K_0}^{\der}: \Gamma_K^{\der}]=\frac{[\Gamma_{K_0}^{\der}: \rho(\Gamma_{K_0}^{\simp})][K_0:K]}
{[K_0^{\der}: K^{\der} \rho(K_0^{\simp})][\nu(K_0): \nu(K)]}.$$ Plugging this into (\[ac/dc\]) gives the lemma.
\[reparation\] Suppose that $K \subseteq K_0$ is small, and $g \in G(\Aff_f)$ with $gKg^{-1} \subseteq K_0$ also small. Then $$[\Gamma^{\ad}_{K_0}: \Gamma^{\ad}_{gKg^{-1}}]=[\Gamma^{\ad}_{K_0}: \Gamma^{\ad}_K].$$
We show that the expression (\[jade\]) does not change when $K$ is replaced with $gKg^{-1}$. Clearly $\nu(K)=\nu(gKg^{-1})$. Since $$[K_0:K]= \frac{\vol_{dg_f}(K_0)}{\vol_{dg_f}(K)},$$ we have $[K_0:gKg^{-1}]=[K_0:K]$. Finally, we claim that $$[K_0^{\der}: (gKg^{-1})^{\der} \rho(K_0^{\simp})] = [K_0^{\der}: K^{\der} \rho(K_0^{\simp})] .$$ From the exact sequence (\[custer\]), it is enough to show that $[K_0^{\der}:(gKg^{-1})^{\der}]=[K_0^{\der}:K^{\der}]$ and $[\rho(K_0^{\simp}):\rho((gKg^{-1})^{\simp})]=[\rho(K_0^{\simp}):\rho(K^{\simp})]$. These hold because $(gKg^{-1})^{\der}=gK^{\der}g^{-1}$ and $\rho((gKg^{-1})^{\simp})=g\rho(K^{\simp})g^{-1}$.
\[schloendorn\] If $G$ is semisimple and $K$ is small, then $$|\pi_0(S_K)|= [K_0: K \rho(K_0^{\simp})][ \Gamma_{K_0}: G(\Q)_+ \cap \rho(K_0^{\simp})] |\pi_0(S_{K_0})|.$$
The kernel of the projection $\pi_0(S_K) \twoheadrightarrow \pi_0(S_{K_0})$ is isomorphic to $$K_0 \left / (KG(\Q)_+ \cap K_0) \right. .$$ By Section 2.1.3 in [@Deligne], we have $ \rho(G_{\simp}(\Aff_f)) \subseteq K G(\Q)_+$. Using the exact sequence $$1 \to (K_0 \cap KG(\Q)_+) \left /K \rho(K_0^{\simp}) \right. \to K_0 \left / K \rho(K_0^{\simp}) \right. \to K_0 \left/ (KG(\Q)_+ \cap K_0) \right. \to 1,$$ we are reduced to computing the order of $$(K_0 \cap KG(\Q)_+) \left / K \rho(K_0^{\simp}) \right. \isom \Gamma_{K_0} \left/ (K \rho(K_0^{\simp}) \cap G(\Q)_+) \right..$$ This group sits in the sequence $$1 \to (G(\Q)_+ \cap K \rho(K_0^{\simp})) \left / (G(\Q)_+ \cap \rho(K_0^{\simp})) \right. \to \Gamma_{K_0} \left / (G(\Q)_+ \cap \rho(K_0^{\simp})) \right. \to
\Gamma_{K_0} \left / (K \rho(K_0^{\simp}) \cap G(\Q)_+) \right. \to 1.$$
We claim the kernel is trivial. Note that $K \rho(K_0^{\simp}) \subseteq K \rho(G_{\simp}(\Q) K^{\simp})$ by strong approximation. So $$\begin{split}
G(\Q)_+ \cap K \rho(K_0^{\simp}) &\subseteq G(\Q)_+ \cap K \rho(G_{\simp}(\Q)) \\
&= G(\Q)_+ \cap (K \cap G(\Q)) \rho(G_{\simp}(\Q)).
\end{split}$$ Since $K \cap G(\Q) \subseteq \rho(G_{\simp}(\Q))$ by condition \[joy\] of Proposition \[small\], we have $G(\Q)_+ \cap K \rho(K_0^{\simp}) \subseteq G(\Q)_+ \cap \rho(K_0^{\simp})$. This proves the claim, and the lemma follows.
In the course of proving the theorem, we will pass to the adjoint group to apply Harder’s theorem (Proposition \[H1\]), but lift to $G_{\simp}$ to apply Harder’s calculation (Proposition \[H2\]). We must record the difference between Serre’s Euler characteristic at $G_{\ad}$ and $G_{\simp}$.
\[kristin\] We have $$\chi_{\alg}(\Gamma_{K_0}^{\ad})= \frac{|\ker (\rho(\Q))| | K_0 \cap Z(\Q)|}{[\Gamma_{K_0}^{\der}: \rho(\Gamma_{K_0}^{\simp})][\Gamma_{K_0}: \Gamma_{K_0}^{\der}]} \chi_{\alg}(\Gamma_{K_0}^{\simp}).$$
This follows from the properties of $\chi_{\alg}$ mentioned earlier.
Proof of Theorem \[chi-theorem\]
--------------------------------
Pick a set $g_1, \ldots, g_r$ of representatives of $\pi_0(S_{K_0})$, viewed as a quotient of $G(\Aff_f)$ as in (\[pease\]).
Let $K$ be small subgroup of finite index in $K_0$. Possibly by intersecting finitely many conjugates of $K$, we may assume
- $K$ is normal in $K_0$
- $g_iKg_i^{-1}$ is a small subgroup of $K_0$ for all $i$.
By Proposition \[???\], $\chi_K(G)=\chi_{\Top}(\ol{S}_K)$. By Proposition \[fibre computation\], this is equal to $[G(\R): G(\R)_+]^{-1}\chi_{\Top}(S_K)$. Write $\Gamma_g$ for $\Gamma^{\ad}_{gKg^{-1}}$. By 2.1.2 of [@Deligne], the components of $S_K$ are each isomorphic to $\Gamma_g \bks X^+$, where $X^+$ is a component of $X$. Here $g$ runs over $\pi_0(S_K)$.
By Proposition \[fibre computation\], the topological spaces $\Gamma_g \bks X^+$ and $\Gamma_g \bks \ol{X}$ are isomorphic. Therefore we have $$\chi_{\Top}(\Gamma_g \bks X^+) = \chi_{\Top}(\Gamma_g\bks \ol{X}).$$
Applying Proposition \[H1\] to $G_{\ad}$, this is equal to $\chi_{\alg}(\Gamma_g) $. Therefore $$\chi_K(G)= [G(\R): G(\R)_+]^{-1} \sum_{g \in \pi_0(S_K)} \chi_{\alg}(\Gamma_g).$$
Every element in $\pi_0(S_K)$ may be written as the product of an element of $\pi_0(S_{K_0})$ with an element of $K_0$. Since $K$ is normal in $K_0$, the groups $\Gamma_{gk_0}$ and $\Gamma_{g}$ are equal for $k_0 \in K_0$. It follows that $$\chi_K(G)=\frac{|\pi_0(S_{K})|}{ [G(\R): G(\R)_+]|\pi_0(S_{K_0})|} \sum_{i=1}^r \chi_{\alg}(\Gamma_{g_i}).$$
By Corollary \[reparation\] we have $$\begin{split}
\chi_{\alg}(\Gamma_{g_i}) &=[\Gamma^{\ad}_{K_0}: \Gamma^{\ad}_{g_iKg_i^{-1}}] \chi_{\alg}(\Gamma^{\ad}_{K_0}) \\
&=[\Gamma^{\ad}_{K_0}: \Gamma^{\ad}_K] \chi_{\alg}(\Gamma^{\ad}_{K_0}).
\end{split}$$
This gives $$\chi_K(G)= [G(\R): G(\R)_+]^{-1} [\Gamma^{\ad}_{K_0}: \Gamma_K^{\ad}] \left| \pi_0(S_K) \right| \chi_{\alg}(\Gamma^{\ad}_{K_0}).$$
The component group $\pi_0(S_K)$ fits into the exact sequence $$1 \to G_{\der}(\Aff_f) / (G_{\der}(\Aff_f) \cap G(\Q)_+K) \to \pi_0(S_K) \to C(\Q)^\dagger \bks C(\Aff_f) / \nu(K) \to 1$$
This gives $$\chi_K(G)=[G(\R): G(\R)_+]^{-1} |\pi_0(S_{K^{\der}})| |C(\Q)^\dagger \bks C(\Aff_f) / \nu(K)| [\Gamma^{\ad}_{K_0}: \Gamma_K^{\ad}] \chi_{\alg}(\Gamma^{\ad}_{K_0}).$$ where here $\pi_0(S_{K^{\der}})=G_{\der}(\Aff_f)/ G_{\der}(\Q)_+ K^{\der}$.
Using $\chi_{K_0}(G)=[K_0:K]^{-1} \chi_K(G)$ together with Lemma \[frank\] gives $$\chi_{K_0}(G)= \frac{|\pi_0(S_{K^{\der}})| [\nu(K_0):\nu(K)] |C(\Q)^\dagger \bks C(\Aff_f) / \nu(K_0)|
[\Gamma^{\ad}_{K_0}: \Gamma_K^{\ad}]}
{[G(\R): G(\R)_+] |\nu(K_0) \cap C(\Q)^\dagger|[K_0:K]} \chi_{\alg}(\Gamma^{\ad}_{K_0}).$$
By Lemmas \[jeremiah\] and \[kristin\],
$$\chi_{K_0}(G)=\frac{ |\ker (\rho(\Q))| |\pi_0(S_{K^{\der}})| |C(\Q)^\dagger \bks C(\Aff_f) / \nu(K_0)|}{ [G(\R): G(\R)_+] |\nu(K_0) \cap C(\Q)^\dagger|[ K_0^{\der}: K^{\der}\rho(K_0^{\simp})]} \chi_{\alg}(\Gamma^{\simp}_{K_0}).$$
The theorem then follows from Lemma \[schloendorn\].
Examples
--------
We now use Theorem \[chi-theorem\] and Proposition \[H2\] to explicitly compute some cases of $\chi_{K_0}(G)$. Recall that we write $\OO_f$ for the integer points of $\Aff_f$.
If $T$ is a torus and $K_0 \subset T(\Aff_f)$ is a compact open subgroup, then $$\chi_{K_0}(T)=|T(\Q) \bks T(\Aff_f) / K_0| \cdot |K_0 \cap T(\Q)|^{-1}.$$
$\bullet$ Let $T=\Gm$, and $K_0=T(\OO_f)$. Then $\chi_{K_0}(T)=\half$.
$\bullet$ Let $T$ be the norm-one subgroup of an imaginary quadratic extension $E$ of $\Q$. Let $K_0=T(\OO_f)$. Write $\OO(E)$ for the integer points of the adeles $\Aff_E$ over $E$. Then $T(\Q) \bks T(\Aff_f) / K_0$ injects into $E^\times \bks \Aff_{E,f}^\times/ \OO(E)^{\times}$, which is in bijection with the class group. If the class number of $E$ is trivial, it follows that $\chi_{K_0}(T)=|T(\Z)|^{-1}$.
If $G$ is semisimple and simply connected, then $$\chi_{K_0}(G)= [G(\R): G(\R)_+]^{-1} \chi_{\alg}(\Gamma_{K_0}).$$
$\bullet$ Let $G=\SL_2$ and $K_0=G(\OO_f)$. Then $$\begin{split}
\chi_{K_0}(G) &=\chi_{\alg}(\SL_2(\Z)) \\
&= - \half B_2 \\
&=-2^{-2}3^{-1}.
\end{split}$$
$\bullet$ Let $G=\Sp_4$ and $K_0=G(\OO_f)$. Then $$\begin{split}
\chi_{K_0}(G) &=\chi_{\alg}(\Sp_4(\Z))\\
&= -\frac{1}{8}B_2 B_4\\
&=-2^{-5}3^{-2}5^{-1}.
\end{split}$$
When the derived group is simply connected the calculation is not much harder.
If $G_{\der}$ is simply connected, then $$\chi_{K_0}(G)=\frac{ |C(\Q)^\dagger \bks C(\Aff_f) / \nu(K_0)|}{ [G(\R): G(\R)_+] |\nu(K_0) \cap C(\Q)^\dagger|} \chi_{\alg}(\Gamma^{\der}_{K_0}).$$
$\bullet$ Let $G=\GL_2$ and $K_0=G(\OO_f)$. Then $\chi_{K_0}(G)=\half \chi_{\alg}(\SL_2(\Z))=-2^{-3}3^{-1}$.
$\bullet$ Let $G=\GSp_4$ and $K_0=G(\OO_f)$. Then $\chi_{K_0}(G)=\half \chi_{\alg}(\Sp_4(\Z))=-2^{-6}3^{-2}5^{-1}$
If all the points of $\ker \rho$ are $\Q$-rational, then $\left[\Gamma^{\der}_{K_0}:G_{\der}(\Q)_+ \cap \rho(K_0^{\simp}) \right]=1$.
By Section 2.0.3 of [@Deligne], we have an injection $$G_{\der}(\Q) / \rho(G_{\simp}(\Q)) \hookrightarrow H^1(\im( \Gal(\ol{\Q}/\Q)), (\ker \rho)(\ol{\Q})),$$ using the cohomology group defined in that paper. We also have an injection $$\Gamma^{\der}_{K_0} / (G_{\der}(\Q)_+ \cap \rho(K_0^{\simp})) \hookrightarrow G_{\der}(\Q) / \rho(G_{\simp}(\Q)).$$
Since all the points of $\ker \rho$ are $\Q$-rational, all these groups are trivial.
$\bullet$ Let $G=\PGL_2$ and $K_0=G(\OO_f)$. The only nontrivial factors in the formula are $[G(\R): G(\R)_+]=2$, $|\ker \rho(\Q)|=2$, and $\chi_{\alg}(\SL_2(\Z))=-2^{-2}3^{-1}$. Thus $\chi_{K_0}(G)= -2^{-2}3^{-1}$.
The Case of $\SL_2$ {#Special Linear}
===================
In this section we work out the example of $\SL_2$. Let $G=\SL_2$, defined over $\Q$. Let $A$ be the subgroup of diagonal matrices in $G$, and let $T$ be the maximal elliptic torus of $G$ given by matrices $$\label{rotation}
\gm_{a,b}=\left(\begin{array}{cc}
a & -b \\
b & a \\
\end{array} \right),$$ with $a^2+b^2=1$.
The characters and cocharacters of $T$ are both isomorphic to $\Z$. We identify $\Z \isom X^*(T)$ via $n \mapsto \chi_n$, where $\chi_n(\gm_{a,b})=(a+bi)^n$. We specify $\Z \isom X_*(T)$ by identifying $n$ with the cocharacter taking $\alpha$ to $\left(\begin{array}{cc}
\alpha & 0 \\
0 & \alpha^{-1} \\
\end{array} \right)$. The roots of $T$ in $G$ are then $\{ \pm 2 \}$, and the coroots of $T$ in $G$ are $\{ \pm 1 \}$. The Weyl group $\Omega$ of these systems has order $2$ and the compact Weyl group $\Omega_\R$ is trivial. Thus each $L$-packet of discrete series has order $2$. The dual group to $G$ is $\hat{G}=\PGL_2(\C)$ in the usual way.
Pick an element $\xi \in G(\C)$ so that
$$\Ad(\xi) \left( \begin{array}{cc}
a &-b \\
b&a \\ \end{array} \right)=\left( \begin{array}{cccc}
a+ib & \\
& a-ib \\ \end{array} \right),$$
and put $B_T=\Ad(\xi^{-1})B_A$. Then $B_T$ is a Borel subgroup of $G(\C)$ containing $T$.
Consider the Langlands parameter $\varphi_G: W_\R \to \hat{G}$ given by $\varphi_G(\tau)=\left(\begin{array}{cc}
0 & 1 \\
1 & 0 \\
\end{array} \right) \times 1$, and $$\varphi_G(z)=\left(\begin{array}{cc}
z^n & 0 \\
0 & \ol{z}^n \\
\end{array} \right) \times z =z^\mu \ol{z}^\nu \times z,$$ where $\mu$ corresponds to $n \in X_*(\hat{T}) \isom X^*(T)$ and $\nu$ corresponds to $-n$. The corresponding representation $E$ of $G(\C)$ has highest weight $\lambda_B=n-1 \in X^*(T)$. It is the $(n-1)$-th symmetric power representation. Its central character is $\lambda_0(z)=z^{n-1}$, where $z=\pm 1$.
We put $\pi_G=\pi(\varphi_G,B_T)$, in the notation from Section \[Langlands packets\]. Write $\pi_G'$ for the other discrete series representation in $\Pi_E$. Thus the $L$-packet determined by $\varphi_G$ is $$\Pi_E= \{ \pi_G, \pi_G' \}.$$
We will put $f_{\infty}dg_{\infty}=e_{\pi_G}$ as in Section \[PC\].
Main Term
---------
First we consider the terms $ST_g(fdg,\pm 1)$.
We have $S \Phi_G(1,e_{\pi_G})=-n\ol{v}(G; dg_{\infty})^{-1}$, and so $$ST_g(fdg,\pm 1, G)=( \pm 1)^n n \ol{v}(G; dg_{\infty})^{-1} f^{\infty}(\pm 1).$$
We have $S \Phi_A(1,e_{\pi_G})=-\ol{v}(G; dg_{\infty})^{-1}$, and so $$ST_g(fdg,\pm 1 , A)= (\pm 1)^n \half \ol{v}(G; dg_{\infty})^{-1} f_A^\infty( \pm 1).$$
If $\gm$ is a regular semisimple element of $G(\C)$ with eigenvalues $\alpha, \alpha^{-1}$, then according to the Weyl character formula, $$\tr(\gm; E)= \frac{\alpha^n-\alpha^{-n}}{\alpha-\alpha^{-1}}.$$
Define $$t_4(n)=\tr \left( \left(\begin{array}{cc}
i & 0 \\
0 & -i \\
\end{array} \right); E \right),$$ where $i$ is a fourth root of unity. Then $t_4(n)=0$ if $n$ is even, and $t_4(n)=(-1)^{\frac{n-1}{2}}$ if $n$ is odd.
Similarly, define $$t_3(n)=\tr \left( \left(\begin{array}{cc}
\zeta & 0 \\
0 & \zeta^2 \\
\end{array} \right); E \right),$$ where $\zeta$ is a third root of unity. Then $t_3(n)=[0,1,-1;3]_n$, meaning that $$t_3(n) =
\begin{cases} & 0 \text{ if } n \equiv 0, \\
& 1 \text{ if } n \equiv 1, \\
& -1 \text{ if } n \equiv 2. \end{cases}$$ Here the congruence is modulo $3$.
There are three stable conjugacy classes of elliptic $\gm \in G(\Q)$, which we represent by $$\gm_3=\left(\begin{array}{cc}
-1 & -1 \\
1 & 0 \\
\end{array} \right),
\gm_4=\left(\begin{array}{cc}
0 & -1 \\
1 & 0 \\
\end{array} \right),
\text{ and }
\gm_6=\left(\begin{array}{cc}
0 & -1 \\
1 & 1 \\
\end{array} \right).$$
Note that $-\gm_4 \sim \gm_4$, $\gm_6^2=\gm_3$, and $-\gm_3 \sim \gm_6$.
Write $T_3$ for the elliptic torus consisting of elements $$\left(\begin{array}{cc}
a & a-b \\
b-a & b \\
\end{array} \right),$$ with $a^2-ab+b^2=1$.
We have $S \Phi_G(\gm_3,e_{\pi_G})=-\ol{v}(T_3)^{-1} t_3(n)$, and so $$ST_g(fdg,\gm_3 , G)=-\ol{v}(T_3)^{-1} SO_{\gm_3}(f^\infty dg_f) t_3(n).$$
We have $S \Phi_G(\gm_4,e_{\pi_G})=-\ol{v}(T)^{-1} t_4(n)$, and so $$ST_g(fdg,\gm_4 , G)=-\ol{v}(T)^{-1} SO_{\gm_4}(f^\infty dg_f) t_4(n).$$
Finally $S \Phi_G(\gm_6,e_{\pi_G})=-\ol{v}(T_3) t_3(n)(- 1)^{n-1}$, and so $$ST_g(fdg, \gm_6 , G)=-\ol{v}(T_3)^{-1} SO_{-\gm_3}(f^\infty dg_f) t_3(n)(- 1)^{n-1}.$$
Thus, $ST_g(fdg)$ is equal to the sum of the following terms: $$-n \ol{v}(G; dg_{\infty})^{-1} f^\infty(1) +n \ol{v}(G; dg_{\infty})^{-1} f^\infty(-1) (- 1)^{n}$$ $$-\half \ol{v}(A; da_{\infty})^{-1}f_A^\infty(1)+\half \ol{v}(A; da_{\infty})^{-1}f_A^\infty(-1)(-1)^{n}$$ $$-\ol{v}(T_3)^{-1} SO_{\gm_3}(f^\infty dg_f) t_3(n)-\ol{v}(T)^{-1} SO_{\gm_4}(f^\infty dg_f) t_4(n)+\ol{v}(T_3)^{-1} SO_{-\gm_3}(f^\infty dg_f) t_3(n)(- 1)^{n}.$$
Endoscopic Terms
----------------
Let $E$ be an imaginary quadratic extension of $\Q$. Write $H_E$ for the kernel of the norm map $\operatorname{Res}^E_{\Q} \Gm \to \Gm$.
The $H_E$ comprise the (proper) elliptic endoscopic groups for $G=\SL_2$. For each $H=H_E$ one finds $\tau(H)=2$ and $| \Out(H,s,\eta)|=1$ (see [@K84], Section 7). Therefore $\iota(G,H)=\half$.
$$e_{\pi_G}^H=e_{\chi_n}+ e_{\chi_n^{-1}}.$$
Write $f^H dh=f^{\infty H} dh_f e_{\pi_G}^H$, where $f^{\infty H}dh_f$ is the transfer of $f^{\infty} dg_f$. Choose $dh_{\infty}$ so that $dh_f dh_{\infty}$ is the Tamagawa measure on $H$. Then we obtain $$ST_g(f^H dh)=2 \ol{v}(H; dh_{\infty}) \sum_{\gm_H} f^{\infty,H}(\gm_H) \Tr_{\Q}^E(\gm_H^n),$$ the sum being taken over $\gm_H \in H(\Q)$.
[ **Remark:**]{} Consider the local transfer, where $f_p dg_p$ is a spherical (invariant under $G(\Zp)$) measure on $G(\Qp)$. Then if $H$ ramifies over $p$, a representation $\pi_p$ in one of the $L$-packets transferring from $H$ will also be ramified. This means that $\tr \pi_p(f_pdg_p)=0$. So we take $f_p^H=0$ in this case. Thus $\Kot(f dg)=ST_g(fdg)$; there is no (proper) endoscopic contribution. This is compatible with the fact that $m_{\disc}$ is constant on $L$-packets in this case.
Case of $\Gamma=\SL_2(\Z)$
--------------------------
We take $K_f=K_0$ to be the integral points of $G(\Aff_f)$. Also let $K_A=K_0 \cap A(\Aff_f)$ and $K_T=K_0 \cap T(\Aff_f)$. Each of these breaks into a product of local groups $K_{0,p}$, etc.
We put $f^\infty dg_f=e_{K_0}$. Note that $f^\infty(g)=f^\infty(-g)$ for all $g \in G(\Aff_f)$ and $f^\infty_A(a)=f^\infty_A(-a)$ for all $a \in A(\Aff_f)$. Therefore, if $n$ is even, then $ST_g(fdg)=0$. So assume henceforth that $n$ is odd. Then our expression is equal to: $$-2n \ol{v}(G; dg_{\infty})^{-1} f^\infty(1)- \ol{v}(A; da_{\infty})^{-1}f_A^\infty(1) -2\ol{v}(T_3)^{-1} SO_{\gm_3}(f^\infty dg_f) t_3(n)+\ol{v}(T)^{-1} SO_{\gm_4}(f^\infty dg_f) (-1)^{\frac{n+1}{2}}.$$
We have $$\begin{split}
-2n \ol{v}(G; dg_{\infty})^{-1} f^\infty(1) &= -2n \ol{v}(G; dg_{\infty})^{-1} \vol_{dg_f}(K_0)^{-1} \\
&= -2n \tau(G)^{-1} d(G)^{-1} \chi_{K_0}(G) \\
&=\dfrac{n}{12},
\end{split}$$ $$\begin{split}
- \ol{v}(A; da_{\infty})^{-1} f_A^\infty(1) &= - \ol{v}(A; da_{\infty})^{-1} \vol_{da_f}(K_A)^{-1} \\
&= -\tau(A)^{-1} d(A)^{-1} \chi_{K_A}(A) \\
&=-\half.
\end{split}$$
Now we consider $SO_{\gm_4}(f^\infty dg_f; dt_f)$. We have $1-\alpha(\gm_4)=2$ for the positive root $\alpha$ of $G$. Therefore by Proposition \[orbitalintegrals\], the local orbital integrals are equal to $\vol_{dt_p}(K_{T,2})^{-1}$ for $p \neq 2$. At $p=2$, one has two stable conjugacy classes $\gm_4$ and $\gm_4'$ in the conjugacy class of $\gm_4$, where $\gm_4'=\left(\begin{array}{cc}
0 & 1 \\
-1 & 0 \\
\end{array} \right)$.
It follows that
$$SO_{\gm_4}(f^\infty dg_f; dt_f)= \left( O_{\gm_4}(e_{K_2}; dt_2)+ O_{\gm_4'}(e_{K_2}; dt_2) \right) \prod_{p \neq 2} \vol_{dt_p}(T(\Qp) \cap K_p)^{-1} .$$
To compute the local integral at $p=2$, we reduce to a $\GL_2$-computation by the following lemma. Its proof is straightforward.
Let $F$ be a $p$-adic local field with ring of integers $\OO$. Put $G=\SL_2$, $\tilde{G}=\GL_2$, and $Z$ for the center of $\tilde{G}$. Pick Haar measures $dg$ on $G(F)$, $d \tilde g$ on $\tilde G(F)$, and $dz$ on $Z(F)$. Let $f \in C_c(Z(F) \bks \tilde G(F))$. Then $$\frac{\vol_{dz}(Z(\OO))}{\vol_{d \tilde g}(\tilde G(\OO))} \int_{Z(F) \bks \tilde G(F)} f(g) \frac{d \tilde g}{dz}= \vol_{dg}(G(\OO))^{-1} |\OO^{\times}/\OO^{\times 2}|^{-1} \sum_{\alpha} \int_{G(F)} f(t_\alpha g) dg.$$ Here $\alpha$ runs over the square classes in $F^\times$, and $t_\alpha=\left(\begin{array}{cc}
\alpha & 0 \\
0 & 1 \\
\end{array} \right)$.
We have $$O_{\gm_4}(e_{K_2}; dt_2)+ O_{\gm_4'}(e_{K_2}; dt_2)= 2 \vol_{dt_2}(K_{T,2})^{-1}.$$
Write $ \tilde f_2$ for the characteristic function of $\GL_2(\Z_2)Z(\Q_2)$. By the lemma, $$\int_{Z(\Q_2) \bks \GL_2(\Q_2)}\tilde f_2(g^{-1}\gm_4 g) \frac{d \tilde g}{dz}= \vol_{dt_2}(K_{T,2}) |\Z_2^{\times}/\Z_2^{\times 2}|^{-1} \sum_\alpha O_{\Ad(t_{\alpha})(\gm_4)}(e_{K_0}; dt_2).$$
Here we are normalizing $d \tilde g$ and $dz$ so that $\vol_{dz}(Z(\Z_2))=\vol_{d \tilde g}(\GL_2(\Z_2))=1$.
In fact, $\Ad(t_{\alpha})(\gm_4)$ is conjugate in $G(\Q_2)$ to $\gm_4$ if and only if $\alpha$ is a norm from $\Q_2(\sqrt{-1})$, and in the contrary case, it is conjugate to $\gm_4'$. It follows that
$$\int_{Z(\Q_2) \bks \GL_2(\Q_2)} f_2(g^{-1}\gm_4 g) \frac{d \tilde g}{dz}= \left(O_{\gm_4}(e_{K_2}; dt_2)+O_{\gm'_4}(e_{K_2}; dt_2) \right) \vol_{dt_2}(K_{T,2}).$$
By an elliptic orbital integral computation in [@KClay], the left hand side is equal to $2$.
We conclude that $$SO_{\gm_4}(f^\infty dg_f; dt_f)=2 \vol_{dt_f}(T(\Aff_f) \cap K_0)^{-1},$$ and so $$\begin{split}
-\ol{v}(T)^{-1} SO_{\gm_4}(f^\infty dg_f)t_4(n) &= -2\ol{v}(T)^{-1} \vol_{dt_f}(T(\Aff_f) \cap K_0)^{-1} t_4(n) \\
&= -2\tau(T)^{-1} \chi_{K_T}(T) t_4(n) \\
&=2^{-2} (-1)^{\frac{n+1}{2}}.
\end{split}$$
Similarly, we find that $$SO_{\gm_3}(f^\infty dg_f)=2 \vol_{dt_{3,f}}(T_3(\Aff_f) \cap K_0)^{-1},$$ and so $$-2\ol{v}(T_3)^{-1} SO_{\gm_3}(f^\infty dg_f) t_3(n)=-3^{-1}t_3(n).$$
We conclude that in this case, $$ST_g(fdg)= \frac{n}{12}-\half+ \frac{1}{4} (-1)^{\frac{n+1}{2}}-\frac{1}{3}t_3(n).$$
Note that for $n >1$ this agrees precisely with the discrete series multiplicities. For $n=1$, this expression is equal to $-1$, but of course in this case $\pi$ is not regular.
Real Tori {#Real Tori}
=========
We have finished our discussion of $\SL_2$. Starting with this section, we begin to work out the example of $\GSp_4$. Various isomorphisms of tori must be written carefully, so we begin by explicitly working out their parametrizations.
The Real Tori $\Gm$, $S$, and $T_1$
-----------------------------------
We identify the group of characters of $\Gm$ with $\Z$ in the usual way, via $\left(a \mapsto a^n \right) \leftrightarrow n$.
Let $A_0= \Gm \times \Gm$, viewed as a maximal torus in $\GL_2$ in the usual way. Via the above identification we obtain $X^*(A_0) \cong \Z^2$ and $X_*(A_0) \cong \Z^2$.
Let $S= \operatorname{Res}_\R^\C \Gm$. Recall that $\operatorname{Res}^\C_\R \Gm$ denotes the algebraic group over $\R$ whose $\mathcal{A}$-points are $(\mathcal{A} \otimes \C)^\times$ for an $\R$-algebra $\mathcal{A}$. By choosing the basis $\{1,i \}$ of $\C$ over $\R$, we have an injection $(\mathcal{A} \times \C)^\times \to \GL(\mathcal{A} \otimes \C) \cong \GL_2(\mathcal{A})$. Thus we have an embedding $\iota_S:S \to \GL_2$ as an elliptic maximal torus.
There is a ring isomorphism $\varphi: \C \otimes \C \isom \C \times \C$ so that $\varphi(z_1 \otimes z_2)=(z_1z_2,z_1 \ol{z_2})$, which restricts to an isomorphism $\varphi: S(\C) \isom \Gm(\C) \times \Gm(\C)$. This isomorphism is also actualized by conjugation within $\GL_2(\C)$. Fix $x \in \GL_2(\C)$ so that $$\Ad(x) \left(\begin{array}{cc}
a & -b \\
b & a \\
\end{array} \right)=
\left(\begin{array}{cc}
a+ib & \\
& a-ib \\
\end{array} \right);$$ then $\Ad(x): S(\C) \isom A_0(\C)$ is identical to $\varphi$, viewing these two tori under the embeddings above.
We fix the isomorphism from $\Z^2$ to $X^*(S)$ which sends $(1,0)$ (resp. $(0,1)$) to the character $\varphi$ composed with projection to the first (resp. second) component of $\Gm \times \Gm$. Similarly we fix the isomorphism from $\Z^2$ to $X_*(S)$ which sends $(1,0)$ (resp. $(0,1)$) to the cocharacter $a \mapsto \varphi^{-1}(a,1)$ (resp. $a \mapsto \varphi^{-1}(1,a)$).
Write $\hat{S}$ for the Langlands dual torus to $S$. It is isomorphic to $\C^\times \times \C^\times$ as a group, with $\Gamma_\R$-action defined by $\sigma(\alpha,\beta)=(\beta, \alpha)$. We fix the isomorphism $X^*(S) \isom X_*(\hat{S})$ given by $$(a,b) \mapsto (z \mapsto (z^a,z^b)).$$
We have an inclusion $\iota_S: \Gm \to S$ given on $\mathcal A$-points by $a \mapsto a \otimes 1$. Write $\sigma_S$ for the automorphism of $S$ given by $1 \otimes \sigma$ on $\mathcal A$-points. Note that the fixed point set of $\sigma_S$ is precisely the image of $\iota_S$.
Write $\Nm: S \to \Gm$ for the norm map given by $s \mapsto s \cdot \sigma_S(s)$. Note that the product $s \cdot \sigma_S(s)$ is in $\iota_S(\Gm)$, which we identify here with $\Gm$. One computes that the norm map induces the map $n \mapsto (n,n)$ from $X^*(\Gm)$ to $X^*(S)$ with the above identifications.
Write $T_1$ for the kernel of this norm map. Its group of characters fits into the exact sequence $$0 \to X^*(\Gm) \to X^*(S) \to X^*(T_1) \to 0.$$ We identify $X^*(T_1)$ with $\Z$ in such a way so that the restriction map $X^*(S) \to X^*(T_1)$ is given by $(a,b) \mapsto a-b$. The corresponding map $\hat{S} \to \hat{T}$ is given by $(\alpha,\beta) \mapsto \alpha \beta^{-1}$.
The Kernel and Cokernel Tori
----------------------------
We define $\Aker$ to be the kernel of the map from $\Gm^4 \to \Gm$ given by $(a,b,c,d) \mapsto \dfrac{ab}{cd}$. We define $\Acok$ to be the cokernel of the map from $\Gm$ to $\Gm^4$ given by $x \mapsto (x,x,x^{-1}, x^{-1})$. Write $\Tker$ for the kernel of the map $$S \times S \to \Gm$$ given by $$(\alpha,\beta) \mapsto \Nm(\alpha/\beta),$$ and $\Tcok$ for the cokernel of the map $$\Gm \to S \times S$$ given by $$x \mapsto (\iota_S(x),\iota_S(x^{-1})).$$
Identifying $X_*(\Gm)$ and $X^*(\Gm)$ with $\Z$ as before, we obtain exact sequences $$0 \to X_*(\Aker) \to \Z^4 \to \Z \to 0,$$ $$0 \to \Z \to \Z^4 \to X^*(\Aker) \to 0,$$ $$0 \to \Z \to \Z^4 \to X_*(\Acok) \to 0,$$ $$0 \to X^*(\Acok) \to \Z^4 \to \Z \to 0.$$
Here the maps from $\Z \to \Z^4$ are both $n \mapsto (n,n,-n,-n)$, and the maps from $\Z^4 \to \Z$ are both $(n_1,n_2,n_3,n_4) \mapsto n_1+n_2-n_3-n_4$.
Thus we obtain isomorphisms $$g_{\kc}: X^*(A_{\ker}) \isom X_*(A_{\cok})$$ and $$g_{\ck}: X^*(A_{\cok}) \isom X_*(A_{\ker}),$$ obtained from the exact sequences defining $\Aker$ and $\Acok$. In this way we view $\Acok(\C)$ and $\Aker(\C)$ as the dual tori $\hat{A}_{\ker}$ and $\hat{A}_{\cok}$, respectively.
The isomorphism $\varphi \times \varphi: S(\C) \times S(\C) \isom (\C^\times)^4$ gives isomorphisms $\Phi_{\ker}: \Tker(\C) \isom \Aker(\C)$ and $\Phi_{\cok}: \Tcok(\C) \isom \Acok(\C)$.
Consider the map from $S \times S$ to $S \times S$ given by $(a,b) \mapsto (ab,a \sigma_S(b))$. This fits together with the previous maps to form an exact sequence $$1 \to \Gm \to S \times S \to S \times S \to \Gm \to 1,$$ and yields an isomorphism $\Psi_T: \Tcok \isom \Tker$.
Consider the map from $\Gm^4$ to $\Gm^4$ given by $(a,b,c,d) \mapsto (ac,bd,ad,bc)$. This fit together with the previous maps to form an exact sequence $$1 \to \Gm \to \Gm^4 \to \Gm^4 \to \Gm \to 1$$ and yields an isomorphism $\Psi_A: \Acok \isom \Aker$. On $\C$-points we have $$\label{100students}
\Phi_{\ker} \circ \Psi_T(\C)=\Psi_A(\C) \circ \Phi_{\cok}.$$
Structure of $\GSp_4(F)$ {#Structure of}
========================
The General Symplectic Group
----------------------------
Let $F$ be a field of characteristic $0$. Put $$J= \left( \begin{array}{cccc}
&&&1 \\
&&-1& \\
&1&& \\
-1 &&& \\ \end{array} \right).$$
Take $G$ to be the algebraic group $\GSp_4= \{ g \in \GL_4 \mid gJg^t= \mu J, \text{ some } \mu=\mu(g) \}$. It is closely related to the group $G'=\Sp_4= \{g \in \GSp_4 \mid \mu(g)=1 \}$. Write $A$ for the subgroup of diagonal matrices in $G$, and $Z$ for the subgroup of scalar matrices in $G$.
We fix the isomorphism $\iota_A: \Aker \isom A$ given by $$\label{abcd}
(a,b,c,d) \mapsto \left( \begin{array}{cccc}
a &&& \\
&c&& \\
&& d& \\
&&& b \\ \end{array} \right).$$
Let $B_A$ be the Borel subgroup of upper triangular matrices in $G$.
Root Data
---------
Although $A$ and $\Aker$ are isomorphic tori, we prefer to parametrize their character and cocharacter groups differently, since the isomorphism $\iota_A$ permutes the order of the components.
So we express $X^*(A)=\Hom(A,\Gm)$ as the cokernel of the map $$\label{exactchar}
i: \Z \to \Z^4,$$ given by $i(n)=(n,-n,-n,n)$.
Write $e_1, \ldots, e_4$ for the images in $X^*(A)$ of $(1,0,0,0), \ldots, (0,0,0,1)$. Thus $e_1+e_4=e_2+e_3$. The basis $\Delta_G$ of simple roots corresponding to $B_A$ is $ \{ e_1-e_2, e_2-e_3 \}$. The corresponding positive roots are $\{ e_1-e_2,e_1-e_4,e_2-e_3,e_1-e_3 \}$. The half-sum of the positive roots is then $\rho_B=\half(4e_1-e_2-3e_3) \in X^*(A)$.
Write $\Omega$ for the Weyl group of $A$ in $G$. Write $w_0,w_1$, and $w_2$ for the elements of $\Omega$ which conjugate $\diag(a,b,c,d) \in A$ to $$\diag(d,c,b,a), \diag(a,c,b,d), \text{ and } \diag(b,a,d,c),$$ respectively.
$\Omega$ has order $8$ and is generated by $w_0,w_1$, and $w_2$.
Express $X_*(A)$ as the kernel of the map $$\label{exactco}
p: \Z^4 \to \Z,$$ given by $p(a,b,c,d)=a-b-c+d$.
Let $\vt_1=(1,0,0,-1)$ and $\vt_2=(0,1,-1,0) \in X_*(A)$. Then the coroots of $A$ in $G$ are given by $R^\vee=R^\vee(A,G)= \{ \pm \vt_1 \pm \vt_2, \pm \vt_1, \pm \vt_2 \}$. The basis $\Delta^\vee_G$ of simple coroots dual to $\Delta_G$ is $\{ \vt_1-\vt_2, \vt_2 \}$. Then $(X^*(A),\Delta_G, X_*(A), \Delta^\vee_G)$ is a based root datum for $G$.
The Dual Group $\hat{G}$
------------------------
We will take $\hat{G}$ to be $\GSp_4(\C)$, with trivial $L$-action, and the same based root data as already discussed for $G$. The isomorphism $$\label{Kimball}
X^*(A) \stackrel{(\iota_A)^*}{\to} X^*(\Aker) \stackrel{(\Psi_A)^*}{\to} X^*(\Acok) \stackrel{g_{ck}}{\to}
X_*(\Aker) \stackrel{(\iota_A)_*}{\to} X_*(A)$$ (and its inverse) furnish the required isomorphism of based root data. Let us write this out more explicitly. Note that $(\iota_A)_*$ and $(\iota_A)^*$ are given by: $$(\iota_A)_*(a,b,c,d)=(a,c,d,b)$$ and $$(\iota_A)^*(a,b,c,d)=(a,d,b,c).$$
The isomorphism in (\[Kimball\]) is induced from the linear transformation $\Sigma: \Z^4 \to \Z^4$ represented by the matrix $$\left( \begin{array}{cccc}
1&1&0&0 \\
1&0&1&0 \\
0&1&0&1 \\
0&0&1&1 \\ \end{array} \right),$$ which gives the exact sequence $$0 \to \Z \stackrel{i}{\to} \Z^4 \stackrel{\Sigma}{\to} \Z^4 \stackrel{p}{\to} \Z \to 0,$$ and thus an isomorphism
$$\label{bat}
X^*(A) \stackrel{\Sigma}{\isom} X_*(A).$$
(This agrees with the isomorphism used in Section 2.3 in [@Ralf].)
We have $\Sigma(e_1-e_2)=\vt_2$ and $\Sigma(e_2-e_3)=\vt_1-\vt_2$. Thus the based root datum above is self-dual. Note that $\Sigma(\rho)=\frac{3}{2} \vt_1+ \half \vt_2$. Write $\hat A$ for $A(\C)$; it is the dual torus to $A$ via the isomorphism in (\[bat\]).
Discrete Series for $\GSp_4(\R)$ {#Discrete Series for G}
================================
The maximal elliptic torus $T$ of $G$
-------------------------------------
Consider the map $\GL_2 \times \GL_2 \to \GL_4$ given by $$\left( \begin{array}{cc}
a &b \\
c&d \\ \end{array} \right) \times
\left( \begin{array}{cc}
e &f \\
g&h \\ \end{array} \right) \mapsto
\left( \begin{array}{cccc}
a &&&b \\
&e&f& \\
&g&h& \\
c&&&d \\ \end{array} \right).$$
The composition of this with the natural inclusion $S \times S \to \GL_2 \times \GL_2$ gives an embedding of $S \times S$ into $\GL_4$. This restricts to an embedding of $\Tker$ into $G$, whose image is an elliptic maximal torus $T$ of $G$. Thus we have $\iota_T: \Tker \isom T$.
$T(\R)$ is the subgroup of matrices of the form $$\label{gagamma}
\gamma_{r,\theta_1,\theta_2}= \left( \begin{array}{cccc}
r\cos(\theta_1) && &-r\sin(\theta_1) \\
&r\cos(\theta_2)&-r\sin(\theta_2)& \\
&r\sin(\theta_2)&r\cos(\theta_2) & \\
r\sin(\theta_1)&&&r\cos(\theta_1) \\ \end{array} \right),$$ for $r>0$ and angles $\theta_1,\theta_2$.
Pick an element $\xi \in G(\C)$ so that
$$\Ad(\xi) \left( \begin{array}{cccc}
a &&&-b \\
&c&-d& \\
&d&c& \\
b&&&a \\ \end{array} \right)=\left( \begin{array}{cccc}
a+ib &&& \\
&c+id&& \\
&& c-id& \\
&&& a-ib \\ \end{array} \right),$$
and put $B_T=\Ad(\xi^{-1})B_A$. Then $B_T$ is a Borel subgroup of $G_{\C}$ containing $T$, and $\Ad(\xi): T(\C) \isom A(\C)$ is the canonical isomorphism associated to the pairs $(T,B_T)$ and $(A,B_A)$. The definitions have been set up so that $$\iota_A \circ \Phi_{\ker}=\Ad(\xi) \circ \iota_T.$$
We identify $A(\C)$ as the dual torus $\hat{T}$ to $T$ via the isomorphisms $$\label{Tdual}
X^*(T) \stackrel{(\iota_T)^*}{\to} X^*(\Tker) \stackrel{\Phi_{\ker}^*}{\to} X^*(\Aker)\stackrel{(\Psi_A)^*}{\to}
X^*(\Acok) \stackrel{g_{ck}}{\to} X_*(\Aker) \stackrel{(\iota_A)_*}{\to} X_*(A).$$
Real Weyl Group
---------------
We use $\Ad(\xi)$ to identify $\Omega$ with the Weyl group of $T(\C)$ in $G(\C)$. Recall that $\Omega_{\R}$ denotes the Weyl group of $T(\R)$ in $G(\R)$. By [@Warner], Proposition 1.4.2.1, we have $$\Omega_\R=N_{K_{\R}}(T(\R))/(T(\R) \cap K_{\R}).$$
When discussing maximal compact subgroups of $\GSp_4(\R)$, it is convenient to use a different realization of these symplectic groups. Following [@Bessel], take for $J$ the symplectic matrix $$\left( \begin{array}{cccc}
&&1& \\
&&&1 \\
-1&&& \\
&-1&& \\ \end{array} \right).$$ Take for $K_{\R}$ the standard maximal compact subgroup of $\GSp_4(\R)$ (the intersection of $G(\R)$ with the orthogonal group), and $SK_{\R}$ the intersection of $K_{\R}$ with $\Sp_4(\R)$. One finds that $SK_{\R}$ is isomorphic to the compact unitary group $U_2(\R)$, and yields the Weyl group element $w_2$. The element $\diag(1,1,-1,-1) \in N_{G(\R)}(T(\R)) \cap K_{\R}$ gives $w_0 \in \Omega_\R$, and these two elements generate $\Omega_\R$. This subgroup has index $2$ in $\Omega$, and does not contain the element $w_1$.
Admissible Embeddings
---------------------
Consider the following admissible embedding $\eta_B: {}^LT \to {}^LG$. Write $\theta(z)=\frac{z}{|z|}$ for $z \in \C^\times$. We have ${}^LT= \hat T \rtimes W_{\R}$, with $\tau$ acting as the longest Weyl group element on $\hat T$.
Writing ${}^LT=\hat{T} \times W_\R$, we put $$\eta_B(1 \times z)=\left( \begin{array}{cccc}
\theta(z)^3 &&& \\
&\theta(z)&& \\
&&\theta(z)^{-1}& \\
&&&\theta(z)^{-3} \\ \end{array} \right) \times z$$ for $z \in \C^\times \cong W_\C$,
$$\eta_B(\hat{t} \times 1)=\hat{t} \times 1,$$
for $\hat{t} \in \hat{T}$, and $$\eta_B(1 \times \tau)=J \times \tau.$$
Elliptic Langlands Parameters {#ELP}
-----------------------------
Let $a,b$ be odd integers with $a>b>0$. Let $t$ be an even integer. Put $\mu=\half [(t,t,t,t)+(a,b,-b,-a)]$ and $\nu=\half[ (t,t,t,t)+(-a,-b,b,a) ]$, viewed in $X_*(\hat{T})_\C$. Then we may define a Langlands parameter $\varphi_G: W_\R \to {}^L G$ by $$\varphi_G(z)=z^\mu \ol{z}^{\nu} \times z=|z|^t \left( \begin{array}{cccc}
\theta(z)^a &&& \\
&\theta(z)^b&& \\
&&\theta(z)^{-b}& \\
&&&\theta(z)^{-a} \\ \end{array} \right) \times z,$$ and $\varphi_G(\tau)=J \times \tau$.
Note that the centralizer of $\varphi_G(W_\C)$ in $\hat{G}$ is simply $\hat A$, and that $\lip \mu, \alpha \rip$ is positive for every root of $A$ that is positive for $B_A(\C)$. Thus $\varphi_G$ determines the pair $(\hat A, \hat B_A)$, where $\hat B_A$ is simply $B_A(\C)$.
Define a Langlands parameter $\varphi_B: W_\R \to {}^LT$ by $$\varphi_B(z)=|z|^t \diag(\theta(z)^{a-3},\theta(z)^{b-1},\theta(z)^{1-b},\theta(z)^{3-a}) \times z,$$ and $\varphi_B(\tau)=1 \times \tau$. Then $\varphi_G=\eta_B \circ \varphi_B$.
Let $\pi_G= \pi(\varphi_G,B_T)$ and $\pi_G'= \pi(\varphi_G,w_1(B_T))$, recalling notation from Section \[Langlands packets\]. The $L$-packet determined by $\varphi_G$ is $$\Pi=\{ \pi_G, \pi_G' \}.$$ Here $\pi_G$ is called a holomorphic discrete series representation, and $\pi_G'$ is called a large discrete series representation.
The highest weight for the associated representation $E$ of $G(\C)$ is $$\lambda_B= \half(a+b-4,t-b+1,t-a+3,0) \in X^*(A).$$
From this we may read off the central character $\lambda_0(zI)=z^t$ for $zI \in A_G(\C)$.
The Elliptic Endoscopic Group $H$ {#Endoscopic}
=================================
Root Data
---------
Let $H$ be the cokernel of the map $\Gm \to \GL_2 \times \GL_2$ given by $t \mapsto tI \times t^{-1}I$. Write $A^H$ for the diagonal matrices in $H$, and $B_H$ for the pairs of upper triangular matrices in $H$. Fix $\iota_{A^H}: \Acok \isom A^H$ given by $$(a,b,c,d) \mapsto \left( \begin{array}{cc} a & \\
&b \\ \end{array} \right) \times
\left( \begin{array}{cc}
d & \\
&c \\ \end{array} \right).$$
Write $T_H$ for the image of $S \times S$ in $H$. It is an elliptic maximal torus in $H$. Fix $\iota_{T_H}: \Tcok \isom T_H$ obtained from the map $S \times S \to \GL_2 \times \GL_2$ given by $\alpha \mapsto (\iota_S(\alpha), \iota_S(\alpha))$. Put $B_{T_H}=\Ad(x \times x)^{-1}B_H$, a Borel subgroup of $H_\C$ containing $T_H$. Then $\Ad(x \times x)$ is the canonical isomorphism $T_H(\C) \isom A^H(\C)$ associated to the pairs $(T_H,B_{T_H})$ and $(A^H,B_H)$. We view $X^*(T_H)$ as the kernel of the map $p: \Z^2 \times \Z^2 \to \Z$ given by $(a,b) \times (c,d) \mapsto a+b-c-d$. We have a basis of roots $\Delta_H$ given by $$\label{Eych}
\Delta_H=\{ (1,-1) \times (0,0), (0,0) \times (1,-1) \},$$ and $\rho_H=\half(1,-1) \times \half(1,-1)$.
Furthermore, $X_*(T_H)$ is the cokernel of the map $\iota: \Z \to \Z^2 \times \Z^2$ given by $a \mapsto (a,a) \times (-a,-a)$. We have a basis of coroots $\Delta_H^\vee$ given by $$\label{Hvee}
\Delta_H^\vee= \{ (1,-1) \times (0,0), (0,0) \times (1,-1) \},$$ viewed in the quotient $X_*(T_H)$.
Dual Group $\hat{H}$
--------------------
Let $\hat{H}= \{ (g,h) \in \GL_2(\C) \times \GL_2(\C) \mid \det(g)=\det(h) \}$. We have an inclusion $\Aker(\C) \to \hat{H}$ given by $$(a,b,c,d) \mapsto \left( \begin{array}{cc}
a & \\
&b \\ \end{array} \right) \times
\left( \begin{array}{cc}
d & \\
&c \\ \end{array} \right).$$
Write $\hat{A}^H \subset \hat{H}$ for the image. We thus have an isomorphism $\iota_{\hat{A}^H}: \Aker(\C) \isom \hat{A}^H$.
Also write $\hat{B}_H$ for the subgroup of upper triangular matrices in $\hat{H}$. This Borel subgroup determines a based root datum for $\hat{H}$.
Giving $\hat{H}$ the trivial $L$-action, we view it as a dual group to $H$ via the isomorphisms $$X^*(A^H) \stackrel{(\iota_{A^H})^*}{\to} X^*(\Acok) \stackrel{g_{\ck}}{\to} X_*(\Aker) \stackrel{(\iota_{\hat{A}^H})_*}{\to} X_*(\hat{A}^H),$$
$$X^*(\hat{A}^H) \stackrel{(\iota_{\hat{A}^H})^*}{\to} X^*(\Aker) \stackrel{g_{\kc}}{\to} X_*(\Acok) \stackrel{(\iota_{A^H})_*}{\to} X_*(A^H).$$
We identify $\hat{A}^H$ as the dual torus $\hat{T}_H$ to $T_H$ via the isomorphisms $$\label{THdual}
X^*(T_H)\stackrel{(\iota_{T_H})^*}{\to} X^*(\Tcok) \stackrel{\Phi_{\cok}^*}{\to} X^*(\Acok) \stackrel{g_{ck}}{\to}
X_*(\Aker) \stackrel{(\iota_{\hat{A}^H})^*}{\to} X_*(\hat{A}^H).$$
Let $\eta: {}^LH \to {}^LG$ be given by $$\label{eta}
\left( \begin{array}{cc}
a &b \\
c&d \\ \end{array} \right) \times
\left( \begin{array}{cc}
e &f \\
g&h \\ \end{array} \right) \times w \mapsto
\left( \begin{array}{cccc}
a &&&b \\
&e&f& \\
&g&h& \\
c&&&d \\ \end{array} \right) \times w.$$
Let $$s=\left( \begin{array}{cc}
1 & \\
&1 \\ \end{array} \right) \times
\left( \begin{array}{cc}
-1 & \\
&-1 \\ \end{array} \right)
\in \hat{H}.$$
The image $\eta(\hat{H})$ is the connected centralizer in $\hat{G}$ of $\eta(s)$. Thus, $(H,s,\eta)$ is an elliptic endoscopic triple for $G$. In fact it is the only one, up to isomorphism.
Moreover note that $\eta$ restricted to $\hat{A}^H$ is given by $$\label{hulk}
\eta|_{\hat{A}^H}=\iota_A \circ (\iota_{\hat{A}^H})^{-1}.$$ (Recall that $\hat{A}=A(\C)$.)
Transfer for $H(\R)$ {#Discrete Series for H}
====================
The goal of this section is Proposition \[TransferToH\], in which we identify $e_{\pi_G}^H$ and $e_{\pi_G'}^H$. This is part of the global transfer $f^Hdh$ which is to be entered into $ST_g$ for the endoscopic group $H$. We will recognize it using the character theory of transfer reviewed in Section \[Transfer\].
Parametrization of Discrete Series
----------------------------------
First we must set up the Langlands parameters for discrete series representations of $H(\R)$, and describe how they transfer to $L$-packets in $G(\R)$. Recall that we have fixed three integers $a,b,t$, with $a,b$ odd, $t$ even, and $a>b>0$. Define the Langlands parameter $\varphi_H: W_\R \to {}^LH= \hat H \times W_{\R}$ by $$\varphi_H(z)=|z|^t\left( \begin{array}{cc}
\theta(z)^a & \\
&\theta(z)^{-a} \\ \end{array} \right) \times
|z|^t \left( \begin{array}{cc}
\theta(z)^b & \\
&\theta(z)^{-b} \\ \end{array} \right) \times z$$ for $z \in W_{\C}$, and $$\varphi_H(\tau)=\left( \begin{array}{cc}
& 1\\
-1& \\ \end{array} \right) \times
\left( \begin{array}{cc}
& -1 \\
1& \\ \end{array} \right) \times \tau.$$
Then $\varphi_H$ determines the pair $(\hat{A}_H,\hat{B}_H)$. The $L$-packet is a singleton $\{ \pi_H \}$. The corresponding representation $E_H$ of $H(\C)$ has highest weight $$\lambda_H= \half(t+a-1,t-a+1) \times \half(t+b-1,t-b+1).$$ From this we read off the central character $\lambda_0^H(z_1,z_2)=(z_1z_2)^t$. Most importantly, we have $\varphi_G=\eta \circ \varphi_H$.
There is another Langlands parameter $\varphi_H'$ given by $$\varphi_H'(z)=|z|^t\left( \begin{array}{cc}
\theta(z)^b & \\
&\theta(z)^{-b} \\ \end{array} \right) \times
|z|^t \left( \begin{array}{cc}
\theta(z)^a & \\
&\theta(z)^{-a} \\ \end{array} \right) \times z,$$ and by $\varphi_H'(\tau)=\varphi_H(\tau)$ as above.
Again the $L$-packet is a singleton $\{ \pi_H' \}$. The corresponding representation $E_H'$ has highest weight $$\lambda_H'=\half(t+b-1,t-b+1) \times \half(t+a-1,t-a+1),$$ and the same central character $\lambda_0^H$ as $E_H$.
Let $\varphi_G'=\eta \circ \varphi_H'$. Then $\varphi_G'=\Int(w_2) \circ \varphi_G$, so it is equivalent to $\varphi_G$. In particular, both $L$-packets $\{ \pi_H \}$ and $\{ \pi_H' \}$ transfer to $\Pi=\{ \pi_G, \pi_G' \}$.
Alignment
---------
Recall the definition of alignment from Section \[Transfer\].
\[Pallavi\] Define $j: T_H \isom T$ by $j=\iota_T \circ \Psi_T \circ (\iota_{T_H})^{-1}$. Then $(j,B_T,B_{T_H})$ is aligned with $\varphi_H$, and $(j,w_1B_T,B_{T_H})$ is aligned with $\varphi_H'$.
Since the parameter $\varphi_G$ gives the pair $(\hat{A},\hat{B})$, the parameter $\varphi_G'$ gives the pair $(\hat{A},w_1 \hat{B})$, and since $\varphi_H$ and $\varphi_H'$ both give $(\hat{A},\hat{B}_H)$, the horizontal maps in (\[alignment\]) are identities. The map $\hat{j}: \hat{T} \to \hat{T}_H$ may be computed by composing the isomorphism $X_*(\hat{T}) \isom X^*(T)$ in (\[Tdual\]) with the induced map $j^*: X^*(T) \isom X^*(T_H)$ and finally with the inverse of the isomorphism $X_*(\hat{T}_H) \isom X^*(T_H)$ in (\[THdual\]). Using Equations (\[100students\]) and (\[hulk\]), one finds that $$\hat{j}=\iota_{\hat{A}_H} \circ (\iota_A)^{-1}= \eta^{-1},$$ as desired.
Transfer for $H_\R$
-------------------
\[TransferToH\] Let $\pi_G=\pi(\varphi_G,B_T)$ and $\pi_G'= \pi(\varphi_G, \omega^{-1}(B_T))$ as described in Section \[ELP\]. Then (using notation from Section \[PC\]) we may take $e_{\pi_G}^H=e_{\pi_H}+ e_{\pi_H'} $, where $\pi_H$ (resp., $\pi_H'$) is the discrete series representation determined by $\varphi_H$ (resp., by $\varphi_H'$) as above. Furthermore, we may take $e_{\pi_G'}^H=-e_{\pi_G}^H$.
By Lemma \[Pallavi\], we may use $$\Delta_{\infty}(\varphi_H,\pi(\varphi_G,\omega^{-1}(B_T)))= \lip a_\omega, \hat j^{-1}(s) \rip$$ and $$\Delta_{\infty}(\varphi_H',\pi(\varphi_G,\omega^{-1}(w_1B_T)))= \lip a_{w_1 \omega}, \hat j^{-1}(s) \rip$$ for $\omega \in \Omega$. In both cases, this is given by $$\lip a_{\omega},s \rip= \begin{cases} & 1, \text{ if } \omega \in \Omega_{\R}, \\
& -1, \text{ if } \omega \notin \Omega_{\R}. \end{cases}$$ Note that $ \lip a_{w_1 \omega}, \hat j^{-1}(s) \rip=- \lip a_\omega, \hat j^{-1}(s) \rip$. Therefore the characterization (\[charidentity\]) becomes, for a general measure $f_{\infty}dg_{\infty}$ at the real place,
$$\begin{split}
\Theta_{\pi_H}(f_{\infty}^Hdh_{\infty}) &= \sum_{\pi \in \Pi(\varphi_G)} \Delta_{\infty}(\varphi_H,\pi) \Theta_{\pi}(f_{\infty}dg_{\infty}) \\
&=\Theta_{\pi_G}(f_{\infty} dg_{\infty})-\Theta_{\pi_G'}(f_{\infty}dg_{\infty})
\end{split}$$
and similarly $$\Theta_{\pi_H'}(f_{\infty}^Hdh_{\infty}) =\Theta_{\pi_G}(f_{\infty}dg_{\infty})-\Theta_{\pi_G'}(f_{\infty}dg_{\infty}).$$ In our case, we obtain $$\Theta_{\pi_H}(e_{\pi_G}^H)=\Theta_{\pi_H'}(e_{\pi_G}^H)=(-1)^{q(G)},$$ and $$\Theta_{\pi_H}(e_{\pi_G'}^H)=\Theta_{\pi_H'}(e_{\pi_G'}^H)=-(-1)^{q(G)}.$$
The proposition follows.
Levi Subgroups {#Levi Subgroups}
==============
Levi Subgroups {#levi-subgroups}
--------------
We give the standard Levi subgroups of $G$, which are those of the parabolic subgroups containing $B_A$. We have the group $A$, the group $G$ itself, and the following two Levi subgroups:
$$M_1= \left\{ \left( \begin{array}{cc}
g & \\
& \lambda g \\ \end{array} \right) \mid g \in \GL_2, \lambda \in \Gm \right\}.$$
$$M_2= \left\{ \left( \begin{array}{ccc}
a && \\
&g& \\
&& b \\ \end{array} \right) \mid g \in \GL_2, a,b \in \Gm, \det(g)=ab \right\}.$$
Note that both $M_1$ and $M_2$ are isomorphic to $\Gm \times \GL_2$.
The group $H$ also has four Levi subgroups, namely $A^H$, the group $H$ itself, the image $M_1^H$ of $\GL_2 \times A_0$ in $H$, and the image $M_2^H$ of $A_0 \times \GL_2$ in $H$. Note that both $M_1^H$ and $M_2^H$ are isomorphic to $\GL_2 \times \Gm$.
Miscellaneous constants
-----------------------
We now compute the invariants from Section \[Various Invariants\] for the Levi subgroups of $G$ and $H$.
First, we compute the various $k(M)$. When $M$ is the split torus $A$ its derived group is trivial and so $k(A)=1$. For $i=1,2$, the Levi subgroup $M_i$ is isomorphic to $\GL_2 \times \Gm$, and the torus is isomorphic to $S \times \Gm$. Since $S$ and $\Gm$ have trivial first cohomology, again $k(M_1)=1$.
We have $k(G)=2$.
Write $T$ as before for the elliptic torus of $G$.
Recall that $T_1$ is the kernel of $\Nm$ and $H^1(\R,T_1)$ has order $2$.
Recall that the torus $T$ is isomorphic to the kernel of the map $$S \times S \to \Gm$$ given by $$(\alpha,\beta) \mapsto \Nm(\alpha/\beta).$$ Projection to the first (or second) component followed by $\Nm$ gives an exact sequence $$\label{T_0}
1 \to T_1 \times T_1 \to T \to \Gm \to 1.$$
We have that $G_{\simp}=G_{\der}$ and the inclusion $T_{\simp} = G_{\der} \cap T \subset T$ may be identified with the map $T_1 \times T_1 \to T$ in the sequence above. In particular, $H^1(\R,T_{\simp})$ has order $4$.
Taking the cohomology of (\[T\_0\]) gives the exact sequence $$1 \to \R^\times/\R^{\times 2} \to H^1(\R,T_{\simp}) \to H^1(\R,T) \to 1,$$ from which we conclude that $H^1(\R,T_{\simp}) \to H^1(\R,T)$ is surjective and $H^1(\R,T)$ has order $2$.
One must also compute $k(M_H)$ for Levi subgroups $M_H$ of $H$. The intermediate Levi subgroups are again isomorphic to $\GL(2) \times \Gm$, and for $A_H$ the derived group is trivial. So $k(M_H)=1$ for each of these.
We have $k(H)=1$.
We have $T=P(S \times S)$, $H_{\simp}=\SL_2 \times \SL_2$, and $T_{\simp}=T_1 \times T_1$. The map $T_{\simp} \to T$ factors through $T_1 \times T_1 \to S \times S$. As above we conclude that $k(H)=1$.
Secondly, we compute the Tamagawa numbers. Recall that $$\tau(G)=|\pi_0(Z(\hat{G})^{\Gamma_\Q})| \cdot |\ker^1(\Q,Z(\hat{G}))|^{-1}.$$
We have $\tau(M)=1$ for all Levi subgroups of $G$, and for all proper Levi subgroups of $H$, and $\tau(H)=2$.
For each of these groups, $Z(\hat{M})$ is either the group $\C^\times$ with trivial $\Gamma_\Q$-action, or a product of such groups. By the Chebotarev Density Theorem, the homomorphism
$$\Hom(\Gamma_{\Q},\C^\times) \to \prod_v \Hom(\Gamma_{\Q_v},\C^{\times})$$
is injective. So $|\ker^1(\Q,Z(\hat{G}))|$ is trivial for our examples. Computing the component group of each $Z(\hat{M})$ is straightforward.
The quantities $n^G_M$ are easy to compute, using $N_G(M) \subseteq N_G(Z(M))$. When $M$ is a maximal torus, $n^G_M$ is of course the order of the Weyl group. For the intermediate cases, one finds that $n^G_{M_i}=n^H_{M^H_i}=2$.
For $\gm=1$, we have $\ol{\iota}^M(\gm)=1$ for each $M$, since each $M$ is connected. Note that for Levi subgroups $M$ of $G$, all proper Levi subgroups $M$ of $H$, and all semisimple elements $\gm$ in $G$ or $H$, we have $\ol{\iota}^M(\gm)=1$ since in all these cases the derived groups are simply connected.
Finally, we compute $\iota(G,H)$, which we recall is given by $$\iota(G,H)= \tau(G) \tau(H)^{-1} |\Out(H,s,\eta)|^{-1}.$$ One may compute the order of $\Out(H,s,\eta)$ through Section 7.6 of [@K84], which shows that this set is in bijection with $\bigwedge(\eta(s),\rho)$, in the notation of that paper. This last set is represented by $\{1,g \}$, where $$g= \left( \begin{array}{cccc}
&1&& \\
1&&& \\
&&&1 \\
&&1& \\ \end{array} \right).$$
The conclusion is that $\iota(G,H)=\dfrac{1}{4}$.
Computing $S\Phi_M$ for Levi Subgroups of $G$ {#Computing for G Levis}
=============================================
Recall from Proposition \[MPI\] the formula
$$\Phi_M(\gm, \Theta^E)=(-1)^{q(L)}|\Omega_L| \sum_{\omega \in \Omega^{LM}} \eps(\omega) \tr(\gm; V^M_{\omega(\lambda_B+\rho_B)-\rho_B}),$$
for $\gm \in T_e(\R)$.
In this section, the maximal torus will be conjugate to $A$, and the character group will be identified with $X^*(A)$. We specify an inner product we use on $X^*(A)_\R$ for the Weyl Dimension Formula (Proposition \[Weyl\]).
The usual dot product gives an inner product $(,)$ on $X_*(A)_\R$, viewing it as a hypersurface in $\R^4$.
Consider the isomorphism $$\pr: X^*(A)_\R \isom X_*(A)_\R$$ given by $$\pr(a,b,c,d)=(a,b,c,d)-\frac{a+d-b-c}{4}(1,-1,-1,1),$$ and let $$\lip \lambda, \mu \rip=(\pr(\lambda),\pr(\mu)).$$
For instance, $$\pr(\lambda_B)=\frac{1}{4} (a+b+t-4,a-b+t-2,-a+b+t+2,-a-b+t+4).$$ It will also be necessary to compute $\Omega^{LM}$ for each example. Recall that this is the set of $w \in \Omega$ so that $w^{-1}\alpha >0$ for positive roots $\alpha$ which are either real or imaginary.
The term $\Phi_G$
-----------------
By Proposition \[SpalPhiformula\] we have $\Phi_G(\gm,\Theta^E)=\tr(\gm; E)$. Using the Weyl dimension formula, we compute $$S\Phi_G(1,e_{\pi_G})= -\frac{1}{24} ab(a+b)(a-b) \ol{v}(G)^{-1}.$$
The term $S\Phi_{M_1}$
----------------------
Consider the torus $T_{M_1}$ given by $$\left(\begin{array}{cccc}
a & b & & \\
-b & a && \\
&& \lambda a & \lambda b\\
&& -\lambda b & \lambda a\\
\end{array} \right),$$ with $a^2+b^2 \neq 0$ and $\lambda \neq 0$. This is an elliptic torus in $M_1$.
There is one positive real root, $e_1-e_3$, and one positive imaginary root, $\alpha_{M_1}=e_1-e_2$. We have $\Omega^{LM}=\{1,w_1\}$, $q(L)=1$, and $|\Omega_L|=2$. This gives $$\Phi_{M_1}(1, \Theta^E)= (-2) \left[ \dim_\C V^{M_1}_{\lambda_B}- \dim_\C V^{M_1}_{\lambda_B'} \right],$$ where $\lambda_B'=\half(a+b-4,t-a+1,t-b+3,0)\in X^*(T)$.
Note that $\lip \alpha_{M_1},\lambda_B \rip=\half (b-1)$. The Weyl dimension formula yields $\dim_{\C} V_{\lambda_B}^{M_1}=b$ and $\dim_{\C} V_{\lambda_B'}^{M_1}=a$. Thus, $$S \Phi_{M_1}(1,e_{\pi_G})=-(b-a)\ol{v}(M_1)^{-1}.$$
The term $S\Phi_{M_2}$
----------------------
Consider the torus $T_{M_2}$ given by $$\left(\begin{array}{cccc}
s & & & \\
& a &-b& \\
&b&a & \\
&& & t\\
\end{array} \right),$$ with $st=a^2+b^2 \neq 0$. This is an elliptic torus in $M_2$.
We may conjugate this in $G(\C)$ to matrices of the form $$\gamma= \left( \begin{array}{cccc}
s &&& \\
&a+ib&& \\
&& a-ib& \\
&&& t \\ \end{array} \right)$$ in $A(\C)$. Composing the roots of $A$ with this composition, we determine the positive imaginary root $\alpha_{M_2}=e_2-e_3$. We have $\Omega^{LM}=\{1,w_2 \}$.
This gives $$\Phi_{M_2}(1, \Theta^E)= (-2) \left[ \dim_\C V^{M_2}_{\lambda_B}- \dim_\C V^{M_2}_{\lambda_B''} \right],$$ where $\lambda_B''=\half(t-b-1,a+b-2,0,t-a+3)\in X^*(T)$. Note that $$\pr(\lambda_B'')=\frac{1}{4}(t+a-b-4,t+a+b-2,t-a-b+2,t-a+b+4).$$
The Weyl dimension formula yields $\dim_{\C} V_{\lambda_B}^{M_2}=\half(a-b)$ and $\dim_{\C} V_{\lambda_B''}^{M_2}=\half(a+b)$, and so $$S\Phi_{M_2}(1, e_{\pi_G})=b \cdot \ol{v}(M_2)^{-1}.$$
The term $S\Phi_A$
------------------
By Proposition \[SpalPhiformula\], we have $\Phi_A(1, \Theta^E)= (-1)^{q(G)} |\Omega_G|=-8$, and so $$S \Phi_A(1, e_{\pi_G})=4 \ol{v}(A)^{-1}.$$
Computing $S\Phi_{M_H}$ for Levi subgroups of $H$ {#Computing for H Levis}
=================================================
Since $e_{\pi_G}^H=e_{\pi_H}+e_{\pi_H'}$, we have $$S \Phi_{M_H}(1,e_{\pi_G}^H )=(-1)^{q(G)}(-1)^{\dim(A_{M_H}/A_H)}\ol{v}(M_H)^{-1} \left[ \Phi_{M_H}(1,\Theta_{\pi_H})+ \Phi_{M_H}(1,\Theta_{\pi_H'}) \right].$$
The term $S\Phi_H(1,e_{\pi_G}^H)$
---------------------------------
In this case $H$ has the elliptic torus $T_H$.
From Proposition \[SpalPhiformula\], we obtain $\Phi_H(1,\Theta_{\pi_H})=\dim_\C E_H$. To apply the dimension formula, we compute for instance $\lip \alpha_1,\lambda_H \rip=a-1$, $\lip \alpha_2, \lambda_H \rip = b-1$, and $\lip \alpha_i,\rho_H \rip =1$.
We find that $$\Phi_H(1,\Theta^{E_H})=\Phi_H(1,\Theta^{E_H'})=ab.$$
Therefore $$S \Phi_H(1,e_{\pi_G}^H)=-2 \ol{v}(H)^{-1}ab.$$
The term $S\Phi_{A^H}(1,e_{\pi_G}^H)$
-------------------------------------
From Proposition \[SpalPhiformula\], we obtain $$\Phi_{A^H}(1,\Theta^{E_H})=\Phi_{A^H}(1,\Theta^{E_H'})=4.$$ Therefore $$S \Phi_{A^H}(1,e_{\pi_G}^H)=-8 \ol{v}(A^H)^{-1}.$$
The terms $S\Phi_{M_H}(1,e_{\pi_G}^H)$ for the intermediate Levi subgroups
--------------------------------------------------------------------------
For both $M=M_H^1$ and $M=M_{H}^2$, we have $\Omega_G=\Omega_L \Omega_M$, and so formula (\[SpalPhiformula\]) becomes simply $\Phi_{M_H}(1,\Theta^{E_H})=(-2)\dim_\C V_{\lambda_H}^{M_H}$ for both of these Levi subgroups.
We obtain $$\Phi_{M_H^1}(1,\Theta^{E_H})= \Phi_{M_H^2}(1,\Theta^{E_H'})= -2a$$ and $$\Phi_{M_H^2}(1,\Theta^{E_H})= \Phi_{M_H^1}(1,\Theta^{E_H'})= -2b.$$
Therefore $$S \Phi_{M_H^1}(1,e_{\pi_G}^H)=S \Phi_{M_H^2}(1,e_{\pi_G}^H)=-2\ol{v}(M_H^1)^{-1}(a+b).$$
Final Form: $\gm$ central {#Final Central}
=========================
Recall that $G=\GSp_4$. For the convenience of the reader, we recall the set-up.
Let $a,b$ be odd integers with $a>b>0$, and $t$ an even integer. Consider the Langlands parameter $\varphi_G: W_\R \to {}^L G$ given by $$\varphi_G(z)=|z|^t \left( \begin{array}{cccc}
\theta(z)^a &&& \\
&\theta(z)^b&& \\
&&\theta(z)^{-b}& \\
&&&\theta(z)^{-a} \\ \end{array} \right) \times z,$$ and $\varphi_G(\tau)=J \times \tau$.
Let $\pi_G$ be the discrete series representation $\pi(\varphi_G,B_T)$ of $G(\R)$ as in Section \[Langlands packets\]. Write $\pi_G'$ for the other representation in $\Pi(\varphi_G)$.
Put $f_{\infty}dg_{\infty}=e_{\pi_G}$ as in Section \[PC\] for $\pi_G$ and any measure $f^\infty dg_f$ on $G(\Aff_f)$. Let $fdg=e_{\pi_G} f^\infty dg_f$, a measure on $G(\Aff)$. By the theory of endoscopic transfer there is a matching measure $f^H dh$ on $H(\Aff)$, where $H$ is the elliptic endoscopic group $P(\GL_2 \times \GL_2)$ discussed above.
If $z \in A_G(\Q)$, then $\sum_M ST_g(fdg,z , M)$ is given by the product of $\lambda_0(z)=z^t$ with: $$-\frac{1}{24} ab(a+b)(a-b) \ol{v}(G)^{-1} f^\infty(z)+ \half (a-b)\ol{v}(M_1)^{-1}f^\infty_{M_1}(z)+ \half b \ol{v}(M_2)^{-1}f^\infty_{M_2}(z)+\half \ol{v}(A)^{-1} f^\infty_A(z).$$ If $z=(z_1,z_2) \in A_H(\Q)$, then $\sum_{M_H} ST_g(f^Hdh,z , M_H)$ is given by the product of $\lambda_0^H(z)=(z_1z_2)^t$ with: $$-4ab \ol{v}(H)^{-1}f^{H,\infty}(z)-2(a+b)\ol{v}(M_H^1)^{-1}f^\infty_{M_2}(z)- 2\ol{v}(A^H)^{-1}f^\infty_{A^H}(z).$$
The case $\Gamma=\Sp_4(\Z)$ {#Integral Points}
===========================
Let $f^\infty dg_f=e_{K_0}$, where $K_0=G(\OO_f)$. Here $dg_f$ is an arbitrary Haar measure on $G(\Aff_f)$ so that $dg=dg_f dg_{\infty}$ is the Tamagawa measure on $G(\Aff)$.
Central terms in $G$
--------------------
Note that $f^\infty_{M}(z)=0$ for all $z \in Z(\Q)$ unless $z= \pm 1$, and that $f_M^\infty(1)=f_M^\infty(-1)$ for all Levi subgroups $M$.
First we compute $ST_g(f dg,\pm 1, G)$. We have $$\begin{split}
-\frac{1}{2^33} ab(a+b)(a-b) \ol{v}(G)^{-1} f^\infty(\pm 1) &= -\frac{1}{2^33} ab(a+b)(a-b) \tau(G)^{-1} d(G)^{-1} \chi_{K_0}(G) \\
&= 2^{-10}3^{-3}5^{-1}ab(a+b)(a-b).
\end{split}$$
Next we treat the $\pm 1 \in M_i$ terms, for the intermediate Levi subgroups. We have $$\begin{split}
ST_g(f dg,\pm 1, M_1) &=\half (a-b)\ol{v}(M_1)^{-1}f^\infty_{M_1}(\pm 1) \\
&= -2^{-5}3^{-1}(a-b),
\end{split}$$ and $$\begin{split}
ST_g(f dg,\pm 1 \in M_2] &= \half b \ol{v}(M_2)^{-1}f^\infty_{M_2}(\pm 1) \\
&=-2^{-5}3^{-1}b.
\end{split}$$
Next we treat the $\pm 1 \in A$ terms. We have $f_A(1)=\vol_{da_f}(K \cap A(\Aff_f))^{-1}$, which is $1$. Moreover we take Lebesgue measure on $A(\R)$ so that $\ol{v}(A)=8$. It follows that $$ST_g(f dg,\pm 1 , A)= \half \ol{v}(A)^{-1} f^\infty_A(\pm 1)=2^{-4}.$$
Doubling these terms to account for both central elements, we compute
$$\label{stablepart}
\sum_{z, M} ST_g(f dg,z , M)=2^{-9}3^{-3}5^{-1}ab(a+b)(a-b)-2^{-4}3^{-1}(a-b)-2^{-4}3^{-1}b+2^{-3}.$$
Central Terms in H
------------------
By the Fundamental Lemma (See Hales [@Hales] and Weissauer [@Weissauer] for $\GSp_4$, of course Ngô [@Ngo] in general), we may write $(e_{K_0})^H=e_{K_H}$, where $K_H=H(\OO_f)$. Thus $(f^\infty)_M^H(z)=0$ for all $z \in H(\Q)$ unless $z=(1,\pm 1)$, and $$f_M^{H \infty}(1,1)=f_M^{H \infty}(1,-1)$$ for all Levi subgroups $M=M_H$ of $H$.
The only nontrivial factors in the formula of Theorem \[chi-theorem\] are $|\ker \rho(\Q)|=2$, $[H(\R): H(\R)_+]=4$, and $\chi_{\alg}(H^{\simp}(\Z))$. Note that $H^{\simp}=\SL_2 \times \SL_2$.
Therefore $$\begin{split}
\chi_{K_H}(H) &= 2^{-1} \chi_{\alg}(\SL_2(\Z))^2\\
&= 2^{-5}3^{-2}.
\end{split}$$
We conclude that
$$\begin{split}
ST_g(f^Hdh,(1,\pm 1), H) &=-4ab \ol{v}(H)^{-1}\vol(K_H)^{-1} \\
&= -2^{-4} 3^{-2}ab.
\end{split}$$
Next we find that $\sum_{i=1}^2 ST_g(f^H dh,(1,\pm 1), M_i^H)$ is equal to $$\begin{split}
\sum_{i=1}^2 ST_g(f^H dh,(1,\pm 1), M_i^H) &=-2(a+b)\ol{v}(M_1^H)^{-1}\vol(K_M)^{-1} \\
&=2^{-3} 3^{-1}(a+b).
\end{split}$$
Finally, we have
$$\begin{split}
ST_g(f^H dh,(1,\pm 1) , A^H) &= - 2 \ol{v}(A)^{-1}\vol (K_A)^{-1} \\
&=-2^{-2}.
\end{split}$$
Multiplying by $\iota(G,H)=4^{-1}$, then doubling to account for both central elements, we compute
$$\label{labilepart}
\iota(G,H)\sum_{z,M_H} ST_g(f^H dh, z , M_H)= -2^{-5} 3^{-2}ab +2^{-4} 3^{-1}(a+b) -2^{-3}.$$
Comparison {#Comparison}
==========
As mentioned in the introduction, Wakatsuki in [@Wak; @Mult], [@Wak; @Dim] has used the Selberg Trace Formula, and Arthur’s $L^2$-Lefschetz number formula to compute the discrete series multiplicities $m_{\disc}(\pi, \Gamma)$ for $\pi$ both holomorphic and large discrete series representations for $\Sp_4(\R)$, and for many cases of arithmetic subgroups $\Gamma$. We will compare our formula to his when $\Gamma$ is the full modular group. (Note that if $\pi$ is a discrete series representation of $\GSp_4(\R)$ with trivial central character, and $\pi_1$ is its restriction to $\Sp_4(\R)$, then $m_{\disc}(\pi,\Gamma)=m_{\disc}(\pi_1, \Gamma_1)$, where $\Gamma_1= \Sp_4(\Z)$.) Since he is using the Selberg trace formula, his formula breaks into contributions from each conjugacy class in $\Gamma$. In particular, he identifies the central-unipotent contributions $H_1^{\Hol}$ and $H_1^{\largess}$ to $m_{\disc}(\pi_G)$ and $m_{\disc}(\pi_G')$, respectively. Namely,
$$H_{1}^{\Hol}=2^{-9}3^{-3}5^{-1}ab(a-b)(a+b)-2^{-5}3^{-2}ab+2^{-4}3^{-1}b$$
and $$H_{1}^{\largess}=2^{-9}3^{-3}5^{-1}ab(a-b)(a+b)+2^{-5}3^{-2}ab-2^{-3}3^{-1}b+2^{-2}.$$
(To translate from his notation to ours, use $j=b-1$ and $k= \half(a-b)+2$.)
Using these formulas and and our formulas above, we find
$$H_{1}^{\Hol}= \sum_M ST_g(f dg, \pm 1, M)+\iota(G,H)\sum_{M_H} ST_g(f^H dh, (1, \pm 1),M_H)$$
when $fdg=e_{\pi_G} e_{K_0}$ and $$H_{1}^{\largess}= \sum_M ST_g(f dg, \pm 1, M)+\iota(G,H)\sum_{M_H} ST_g(f^H dh, (1, \pm 1),M_H).$$ when $fdg=e_{\pi_G'} e_{K_0}$.
This proves Theorem \[Theorem\].
[99]{}
J. Arthur, [*The invariant trace formula I. Local theory*]{}, J. Am. Math. Soc. [**1**]{} (1988), 323-383.
J. Arthur, [*The invariant trace formula II. Global theory*]{}, J. Am. Math. Soc. [**1**]{}(1988), 501-554.
J. Arthur, [*The $L^2$-Lefschetz numbers of Hecke operators*]{}, Invent. Math. [**97**]{} (1989), 257-290.
J. Arthur, [*Stable Trace Formula I. General Expansions*]{} J. Inst. of Math. Jussieu [**1**]{}(2) (2002), 175-277.
J. Arthur, [*Stable Trace Formula II. Global Descent*]{} Invent. Math. [**143**]{} (2001), 157-220.
J. Arthur, [*Stable Trace Formula III. Proof of the Main Theorems*]{} Ann. of Math. [**158**]{} (2003), 769-873.
J. Arthur, [*An Introduction to the Trace Formula*]{} Clay Math. Proc. [**4**]{} (2005), 1-263.
A. Borel, [*Automorphic $L$-functions*]{}, in [*Automorphic Forms, Representations and $L$-functions*]{}, Proc. Sympos. Pure Math. vol. 33, Part 2, Amer. Math. Soc. (1979), 27-61.
L. Clozel, P. Delorme, [*Le théorème de Paley-Wiener invariant pour les groupes de Lie réductifs*]{}, Invent. Math. [**77**]{} (1984), 427-453.
L. Clozel, P. Delorme, [*Le théorème de Paley-Wiener invariant pour les groupes de Lie réductifs II*]{}, Ann. Sci. École Norm. Sup. (4) [**23**]{} (1990), 193-228.
P. Deligne, [*Variétés de Shimura: interprétation modulaire, et techniques de construction de modèles canoniques*]{}, in [*Automorphic Forms, Representations and $L$-functions*]{}, Proc. Sympos. Pure Math. vol. 33, Part 2, Amer. Math. Soc. (1979), 247-289.
M. Goresky, R. E. Kottwitz, R. MacPherson, [*Discrete Series Characters and the Lefschetz Formula for Hecke Operators*]{}, Duke Math. J. [**89**]{}, No. 3 (1997), 477-554.
T. Hales, [*The Fundamental Lemma for Sp(4)*]{}, Proc. Amer. Math. Soc. [**125**]{}, No. 1 (1997), 301-308.
G. Harder, [*A Gauss-Bonnet formula for discrete arithmetically defined groups*]{}, Ann. Ec. Norm. Sup. [**4**]{} (1971), 409-455.
R. E. Kottwitz, [*Sign changes in harmonic analysis on reductive groups*]{}, Trans. Amer. Math. Soc. [**278**]{} (1983), 289-297.
R. E. Kottwitz, [*Stable Trace Formula: Cuspidal Tempered Terms*]{}, Duke Math. J. [**51**]{}, No. 3 (1984), 611-650.
R. E. Kottwitz, [*Stable Trace Formula: Elliptic Singular Terms*]{}, Math. Ann. [**275**]{}, (1986), 365-399.
R. E. Kottwitz, [*Tamagawa numbers*]{}, Ann. of Math. (2) [**127**]{} (1988), 629-646.
R. E. Kottwitz, [*Shimura Varieties and $\lambda$-adic representations*]{}, Automorphic Forms, Shimura Varieties, and $L$-functions, Vol. 1(1990), 161-209.
R. E. Kottwitz [*Harmonic analysis on reductive $p$-adic groups and Lie algebras*]{}, Clay Math. Proc. [**4**]{} (2005), 393-522.
R. E. Kottwitz, [*Stable Version of Arthur’s Formula*]{}, preprint.
R. E. Kottwitz, D. Shelstad, [*Foundations of Twisted Endoscopy*]{}, Asterisque [**255**]{} (1999).
R. Langlands, [*On the functional equations satisfied by Eisenstein Series*]{}, Lect. Notes in Math. [**544**]{}, Springer-Verlag (1976), 361-500.
R. Langlands, [*Stable conjugacy: definitions and lemmas*]{}, Canad. J. Math. [**31**]{} (1979), 700-725.
R. Langlands, [*Les debuts d’une formule des traces stable*]{}, Publ. Math. Univ. Paris VII [**13**]{} (1983).
R. Langlands, D. Shelstad, [*On the definition of transfer factors*]{}, Math. Ann. [**278**]{} (1987), 219-271.
J. S. Milne, [*Introduction to Shimura Varieties*]{}, Clay Math. Proc. [**4**]{} (2005), 265-378.
S. Morel, [*On the Cohomology of Certain Non-Compact Shimura Varieties*]{}, Ann. Math. Stud. [**173**]{} (2010).
B. C. Ngô, [*Le lemme fondamental pour les algebres de Lie*]{}, arXiv:0801.0446 (2008).
T. Ono, [*On Tamagawa Numbers*]{}, Algebraic Groups and Discontinuous Subgroups, Proc. Sympos. Pure Math. [**9**]{}, Amer. Math. Soc. (1966) 122-132.
A. Pitale, R. Schmidt, [*Bessel Models for Lowest Weight Representations of $\GSp(4,\R)$*]{}, IMRN, Article ID rnn 156, 54 pages.
B. Roberts, R. Schmidt, [*Local Newforms for $\GSp(4)$*]{}, Lecture Notes in Mathematics [**1918**]{}, Springer (2007).
I. Satake, [*Algebraic structures of symmetric domains*]{}, Vol. 4 of Kanô Memorial Lectures, Iwanami Shoten, Tokyo (1980).
A. Selberg, [*Harmonic analysis and discontinuous groups in weakly symmetric spaces with applications to Dirichlet series*]{}, J. Indian Math. Soc. [**20**]{} (1956), 47-87.
J-P. Serre, [*Cohomologie des groupes discrets*]{}, Prospects in Mathematics, Ann. of Math. Studies, 70, Princeton University Press, 1971.
D. Shelstad, [*$L$-indistinguishability for real groups*]{}, Math. Ann. [**259**]{} (1982), 385-430.
S. Spallone, [*Arthur’s Trace Formula for $\SO(5)$ and Individual Discrete Series Matrix Coefficients*]{}, Thesis, University of Chicago (2004).
S. Spallone, [*Stable Discrete Series Characters at Singular Elements*]{}, Can. Jour. Math., [**61**]{} (2009), no.6, 1375-1382.
T. A. Springer, [*Reductive Groups*]{}, in [*Automorphic Forms, Representations and $L$-functions*]{}, Proc. Sympos. Pure Math. vol. 33, Part 1, Amer. Math. Soc. (1979), 3-27.
R. Tsushima, [*An explicit dimension formula for the spaces of automorphic forms with respect to $\Sp(2,\Z)$*]{}, Proc. Japan Acad. Ser A [**59**]{} (1983), 139-142.
R. Tsushima, [*Dimension Formula for the spaces of Siegel cusp forms and a certain exponential sum*]{}, Mem. Inst. Sci. Tech. Meiji Univ. [**36**]{} (1997), 1-56.
S. Wakatsuki, [*Dimension Formulas for Spaces of Vector Valued Siegel Cusp Forms of Degree Two*]{}, preprint.
S. Wakatsuki, [*Multiplicity Formulas for Discrete Series Representations in $L^2(\Gamma \bks \Sp(2,\R))$*]{}, preprint.
N. Wallach, [*On the constant term of a square integrable automorphic form*]{}, Operator algebras and group representations, Vol. II (Neptun, 1980), 227-237, Monogr. Stud. Math. [**18**]{}, Pritman, Boston, MA, 1984.
G. Warner, [*Harmonic Analysis on Semisimple Lie Groups I*]{}, Springer-Verlag, Berlin (1972).
R. Weissauer, [*A special case of the fundamental lemma, Parts I, II, III, IV*]{}, preprints.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a construction of curved analogues of the nonstandard operators on Grassmannians parallel to the construction of the Paneitz operator in [@CCCex], but technically more demanding. In particular, the construction breaks down in the presence of torsion. In the second part, we prove that the nonstandard operators are not strongly invariant.'
address: 'Dept. of Mathematics and Statistics, Masaryk University, Kotl'' ařsk'' a 2, 611 37 Brno, Czech Republic, E-mail: [email protected]'
author:
- ALEŠ NÁVRAT
title: Nonstandard operators in Grassmannian geometry
---
\[counter\][Lemma]{} \[counter\][Corollary]{} \[counter\][Proposition]{} \[counter\][Remark]{}
Introduction
============
The curved Casimir operators for parabolic geometries were originally introduced in [@CCCBGG]. As described in that paper, their strong naturality properties together with the fact that they act by a multiplication by a scalar on irreducible bundles enable us to use them to construct higher order invariant operators. Concrete examples of such constructions in conformal geometry are discussed in [@CCCex]. The purpose of this article is to apply curved Casimir operators in Grassmannian geometry to construct a family of fourth order operators which are intrinsic to any Grassmannian structure and which coincide with nonstandard operators on locally flat structures. The existence of such operators was proved by a different method by Slov' ak and Gover in [@Gover theorem 5.1]. The approach via curved Casimirs gives an alternative proof of the existence and also yields new formulae for the operators, c.f. corollary \[factorization\]. In the second part we prove by algebraic methods that these operators have the exceptional property that they are not strongly invariant.
The original inspiration comes from conformal geometry. Namely, the equivalence of Grassmannian structures and split signature conformal structures in dimension four is well known. It is obvious then that the curved version of the nonstandard operator corresponds to the conformal square of the Laplacian $\Delta^2:{{\mathcal E}}\to{{\mathcal E}}[-4]$. Hence the construction of $\Delta^2$ via curved Casimirs described in [@CCCex] gives us a recipe how to construct curved analogues of nonstandard operators in Grassmannian geometry. The construction is rather subtle since the direct application of the construction scheme described in [@CCCBGG] yields a trivial operator. This reflects the fact that $\Delta^2$ is not strongly invariant in dimension four.
The structure of the paper is as follows. First we recall basic facts about Grassmannian geometry following from the general theory of parabolic geometries. It includes also the definitions of curved Casimir operators and nonstandard operators. The first part of the second section concerns the initial tractor bundle for the curved Casimir procedure. We derive its composition series, an explicit form of the action of one-forms on that bundle and Casimir eigenvalues on its irreducible pieces. This is used in the second part to state and prove proposition \[thm\] which is the first main result of this paper. The third section is rather independent on the previous two sections. It contains the second main result, stated in proposition \[weak\_invariance\].
Grassmannian structures of type $(2,n)$ {#GS}
---------------------------------------
We shall use the conventions and the abstract index notation of [@prolongation] and [@Gover]. An almost Grassmannian structure of type $(2,n)$ on a smooth manifold $M$ of dimension $2n$ is given by two auxiliary vector bundles ${{\mathcal E}}^A$, ${{\mathcal E}}^{A'}$ of ranks $n$ and $2$ respectively, and identifications $$\label{AG}
TM={{\mathcal E}}^a\cong{{\mathcal E}}_{A'}\otimes{{\mathcal E}}^A={{\mathcal E}}_{A'}^A, \quad
\Lambda^2{{\mathcal E}}_{A'}\cong\Lambda^n{{\mathcal E}}^A.$$ Equivalently, the structure is a classical first order G-structure with reduction of the structure group $GL(2n,{{\mathbb R}})$ to its subgroup $S(GL(2,{{\mathbb R}})\times GL(n,{{\mathbb R}}))$. It is well known that one can construct a unique Cartan connection associated to this structure which makes it into the $|1|$-graded normal parabolic geometry of type $(G,P)$, where $G=SL(2+n,{{\mathbb R}})$ and $P$ is the stabilizer of ${{\mathbb R}}^2$ in ${{\mathbb R}}^{2+n}$. The corresponding flat model $G/P$ are Grassmannians $\operatorname{Gr}_2({{\mathbb R}}^{2+n})$. In the diagram notation, the $|1|$-grading of Lie algebra ${{\mathfrak g}}={\mathfrak{sl}}(2+n,{{\mathbb R}})$ defining the geometry is given by the $A_{n+1}$-diagram with the second node crossed. Hence ${{\mathfrak g}}={{\mathfrak g}}_{-1}\oplus{{\mathfrak g}}_0\oplus{{\mathfrak g}}_1$, where $${{\mathfrak g}}_0\cong\mathfrak{s}({\mathfrak{gl}}(2,{{\mathbb R}})\oplus{\mathfrak{gl}}(n,{{\mathbb R}})),\quad {{\mathfrak g}}_1\cong({{\mathfrak g}}_{-1})^* \cong{{\mathbb R}}^2\otimes{{\mathbb R}}^{n*}.$$ In general, the full obstruction against local flatness of a normal parabolic geometry is encoded in the harmonic curvature $\kappa_H$. In the case of the almost Grassmannian geometry of type $(2,n)$, this curvature consists of two parts: the homogeneity one component $(\kappa_H)_1$ and the homogeneity two component $(\kappa_H)_2$. For detailed description see [@parabook Section 4.1.3]. The former one can be interpreted geometrically as the torsion of a linear connection on the tangent bundle. The connections with such torsion $T=(\kappa_H)_1$ form a class of distinguished connections modelled on one-forms. They are also called Weyl connections and can be viewed as analogues of Levi-Civita connections in conformal geometry, see [@parabook section 5.1]. The harmonic curvature component $(\kappa_H)_2$ is a component of the curvature of any of these connections.
A Grassmannian geometry is usually defined as an almost Grassmannian geometry which admits a torsion free connection moreover. By [@parabook Theorem 4.1.1], the full obstruction against existence of such a connection is exactly the harmonic curvature component $(\kappa_H)_1$. Thus a Grassmannian geometry can be equivalently described as an almost Grassmannian geometry with $(\kappa_H)_1=0$ and hence it is also called semi flat in the literature. Obviously, the distinguished connections in such a case are exactly the torsion-free connections compatible with the isomorphisms in . In subsequent formulas for differential operators, we will use such distinguished connections which moreover induce a flat connection on $\Lambda^2{{\mathcal E}}_{A'}\cong\Lambda^n{{\mathcal E}}^A$. They are called closed since they form a subclass modeled on closed one forms. For more details see [@parabook section 5.1.7].
Let us note that there is another geometry with a similar behavior, namely the quaternionic geometry. It is given by the same grading as the Grassmannian geometry but on the quaternionic real form of ${{\mathfrak g}}^{{\mathbb C}}$ rather than the split form. Nevertheless, considering complexifications and complex bundles, the identifications are satisfied. Thus we may include this geometry into our framework and all the constructions and results hold also for quaternionic geometry.
Curvature of distinguished connections
--------------------------------------
Theorem 4.1.1 in [@parabook] also states that for any normal $|1|$-graded parabolic geometry the equation $(\kappa_H)_1=0$ implies that the harmonic curvature component $(\kappa_H)_2$ coincides with the so-called Weyl curvature $W:=R+\partial({{\sf P}})$. Here, $R$ is the usual curvature tensor while $\partial$ is the bundle map induced by the Lie algebra differential and ${{\sf P}}\in{\Omega}^1(M,T^*M)$ is a uniquely defined piece of curvature called Rho-tensor. Hence according to the description of the harmonic curvature in Grassmannian geometry, the Weyl curvature is given by $W_{ab}{}^d{}_{c}=W^{A'B'D}_{A{\phantom{'}}B{\phantom{'}}C}{{\delta}}^{C'}_{D'}$, where $$\label{W_sym}
W^{A'B'D}_{A{\phantom{'}}B{\phantom{'}}C}\in \Gamma({{\mathcal E}}^{[A'B']}\otimes({{\mathcal E}}_{(AB C)}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}D})_0),$$ and where $()_0$ denotes the trace free part. See e.g. [@parabook section 4.1.3]. Thus the whole curvature $R_{ab}{}^d{}_{c}$ of a distinguished connection $\nabla_a$ is given by the sum of such $W_{ab}{}^d{}_{c}$ and terms which are linear and zero order in ${{\sf P}}_{ab}$. The former part is obviously invariant while the latter depend on the choice of $\nabla_a$. Namely, if the change from $\nabla_a$ to another distinguished connection $\hat{\nabla}_a$ is described by one form $\Upsilon_a$, then the linearized change of the Rho-tensor is given by $\hat{{{\sf P}}}_{ab}={{\sf P}}_{ab}+\nabla_a\Upsilon_b.$ See e.g. [@Gover] or section 5.1.8 of [@parabook]. Moreover, it follows from the algebraic Bianchi identity that the Rho-tensor corresponding to a closed distinguished connection is symmetric, see equation (9) in [@Gover]. The next consequence of theorem 4.1.1 in [@parabook] is that the tensor $d^\nabla{{\sf P}}$, called the Cotton-York tensor, can be expressed in terms of $W$. The exact formula can be deduced from the differential Bianchi identity, see [@thesis Lemma A.5]. In particular, it follows that $$\label{Pe_sym}
(d^\nabla{{\sf P}})^{A'B'C'}_{A{\phantom{'}}B{\phantom{'}}C}\in \Gamma({{\mathcal E}}^{[A'B']C'}\otimes{{\mathcal E}}_{(AB C)}).$$
Grassmannian bundles {#Grassmannian bundles}
--------------------
Following [@Gover] and [@prolongation], we adopt the convention ${{\mathcal E}}[-1]:=\Lambda^2{{\mathcal E}}^{A'}\cong\Lambda^n{{\mathcal E}}_A,$ and then for $w\in\mathbb Z_-$ we put ${{\mathcal E}}[w]:={{\mathcal E}}[-1]^{-w}$ and ${{\mathcal E}}[-w]:={{\mathcal E}}[w]^*$. In analogy with conformal geometry, the bundle ${{\mathcal E}}[w]$ will be called the bundle of densities of weight $w$, and adding $[w]$ to the notation for a bundle indicates a tensor product by ${{\mathcal E}}[w]$. Motivated by the case $n=2$ when the structure is equivalent to the spin conformal structure, the basic bundles ${{\mathcal E}}^A$ and ${{\mathcal E}}_{A'}$ are called spinor bundles, and the corresponding indices are called spinor indices. They can be raised and lowered similarly as the abstract indices in conformal or Riemannian geometry. However, we use a skew symmetric object instead of a metric which is symmetric, and thus the order of indices is important. Namely, to raise primed indices, we will use the canonical section ${{\epsilon}}^{A'B'}$ of ${{\mathcal E}}^{[A'B']}[1]$ which gives the isomorphism ${{\mathcal E}}[-1]\ni f\mapsto f{{\epsilon}}^{A'B'}\in{{\mathcal E}}^{[A'B']}$, and we use it in such a way that $v^{B'}=v_{A'}{{\epsilon}}^{A'B'}.$ The inverse ${{\epsilon}}_{A'B'}$ is used to lower primed indices as follows $v_{B'}=v^{A'}{{\epsilon}}_{A'B'}.$
The Grassmannian standard tractor bundle will be denoted by ${{\mathcal E}}^{{\alpha}}$. It is a vector bundle of rank $n+2$ induced by the standard representation of ${\mathfrak{sl}}(2+n)$ on ${{\mathbb R}}^{2+n}$. The obvious filtration ${{\mathbb R}}^{2+n}\supset{{\mathbb R}}^2$ gives rise to the filtration ${{\mathcal E}}^{{\alpha}}\supset{{\mathcal E}}^{A'}$, where ${{\mathcal E}}^{{\alpha}}/{{\mathcal E}}^{A'}\cong{{\mathcal E}}^A$. Similarly, the standard cotractor bundle ${{\mathcal E}}_{{\alpha}}$ is endowed with filtration ${{\mathcal E}}_{{\alpha}}\supset{{\mathcal E}}_A$, where ${{\mathcal E}}_{{\alpha}}/{{\mathcal E}}_A\cong{{\mathcal E}}_{A'}$. These data define a composition series for ${{\mathcal E}}_{{{\alpha}}}$ that will be denoted by ${{\mathcal E}}_{{{\alpha}}}={{\mathcal E}}_{A'}{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}_A$ henceforth. This notation is motivated by the fact that summands include while there is a projection onto direct summands. Under a choice of a torsion-free connection $\nabla_a$, the composition series splits. The exact formulas for transformations of the splittings of tractor bundles under a change from $\nabla_a$ to another torsion-free connection $\hat{\nabla}_a$ can be found by making explicit the general formulas in [@parabook Section 5.1], see e.g. [@Gover]. However, we will not need these formulas in this article since we will always deal with objects and operations which are known to be invariant. The splittings will be used only to compute explicit formulas for these objects and operations.
Following [@CCCex] and [@prolongation], sections of tractor bundles will be written either as a formula in injectors $X^A_{{{\alpha}}}\in {{\mathcal E}}^A_{{\alpha}}$, $Y^{A'}_{{{\alpha}}}\in {{\mathcal E}}^{A'}_{{\alpha}}$ or simply denoted by a “vector”. For example, a section $v_{{\alpha}}$ of the standard cotractor bundle reads as $$\label{standard_cotractor}
v_{{{\alpha}}}=v_{A'}Y^{A'}_{{{\alpha}}}+v_A X^A_{{{\alpha}}}
=\begin{pmatrix}v_{A'}\\v_A\end{pmatrix},$$ where $v_{A'}$ and $X^A_{{\alpha}}$ are invariant while $v_A$ and $Y^{A'}_{{\alpha}}$ depend on a choice of splitting.
Notation for forms and irreducible bundles {#Notation}
------------------------------------------
To simplify subsequent expressions let us adopt the index notation for forms of [@prolongation]. The usual notation ${{\mathcal E}}_{[AB\dots C]}$ for $k$th skew symmetric power of ${{\mathcal E}}_A$ will be abbreviated by the following multi-indices $$\begin{array}{l}
{{\bf A}}^k:=[A^1\cdots A^k], \quad k\geq 0,
\\
{{\bf \dot{A}}}^k:=[A^2\cdots A^k], \quad k\geq 1,
\\
{{\bf \ddot{A}}}^k:=[A^3\cdots A^k], \quad k\geq 2.
\end{array}$$ The subscript $k$ will be usually omitted, i.e. ${{\mathcal E}}_{{\bf A}}$ means ${{\mathcal E}}_{{{\bf A}}^k}$ throughout the article. Also the square bracket may be absent. Indices labelled with sequential superscripts automatically indicate a completely skew set of indices. The same notation will be used also for skew symmetric powers of tractor bundles. For example, the bundle ${{\mathcal E}}_{{{\alpha}}^2\cdots {{\alpha}}^k}={{\mathcal E}}_{[{{\alpha}}^2\cdots {{\alpha}}^k]}$ will be denoted by ${{\mathcal E}}_{{{\bf \dot{{\boldsymbol{{{\alpha}}}}}}}}$. We will combine this notation with the usual Young diagram notation for irreducible bundles induced by $SL(n)$-modules as follows. We adopt the convention that we symmetrized over sets of indices corresponding to the rows of the diagram first and then with the result skew over sets of indices corresponding to the columns of the diagram. The sets of indices in columns may be then shortened by multi-indices as above. For instance suppose we have a general valence four spinor $A_{A^1A^2B^1B^2}\in{{\mathcal E}}_{A^1A^2B^1B^2}[w].$ If we first symmetrized as follows $B_{A^1A^2B^1B^2}:=A_{(A^1|A^2|B^1)B^2},$ $C_{A^1A^2B^1B^2}:=B_{A^1(A^2|B^1|B^2)}$ and then on this result skew over the nonsymmetric indices $D_{{{\bf A}}^2{{\bf B}}^2}:=C_{[A^1A^2][B^1B^2]},$ then $$D_{{{\bf A}}^2{{\bf B}}^2}\in{\mbox{$
\begin{picture}(13,10)(-1,1)
\put(0,10){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\end{picture}
$}}{{\mathcal E}}_{{{\bf A}}^2{{\bf B}}^2}[w].$$
Nonstandard operators {#NO}
---------------------
Using the above notation, we can explicitly write down irreducible subbundles of $k$-forms. Namely, since $T^*M={{\mathcal E}}^{A'}\otimes{{\mathcal E}}_A$ by , each such subbundle is given by a product of an irreducible $SL(2)$-bundle in $\otimes^k{{\mathcal E}}^{A'}$, which is a symmetric power of ${{\mathcal E}}^{A'}$, and an irreducible $SL(n)$-bundle in $\otimes^k{{\mathcal E}}_{A}$ indicated by the corresponding diagram. For example, the bundle of two-forms splits into ${{\mathcal E}}^{(A'B')}\otimes{{\mathcal E}}_{{{\bf A}}^2}$ and ${{\mathcal E}}_{(AB)}[-1]$, while the decomposition of $\Lambda^4T^*M$ reads as $$\Lambda^4({{\mathcal E}}^{A'}_A)={\mbox{$
\begin{picture}(13,10)(-1,1)
\put(0,10){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\end{picture}
$}}{{\mathcal E}}_{{{\bf A}}^2{{\bf B}}^2}[-2]\oplus{{\mathcal E}}^{(A'B')}\otimes{\mbox{$
\begin{picture}(13,15)(-1,1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,5){\line(1,0){5}}
\put(0,0){\line(1,0){5}}
\put(0,0){\line(0,1){15}}
\put(10,10){\line(0,1){5}}
\put(5,0){\line(0,1){15}}
\end{picture}
$}}{{\mathcal E}}_{{{\bf A}}^3{{\bf B}}^1}[-1]\oplus{{\mathcal E}}^{(A'B'C'D')}_{{{\bf A}}^4}.$$ The decomposition of $\Lambda^kT^*M$ can be written a similar form for any $k$. In particular, the number of columns of Young diagrams that occur is never more than two since this would correspond to a $k$-form which would be skew symmetric in more than two primed spinor indices, and such forms vanish due to the low rank of ${{\mathcal E}}^{A'}$. For instance, the irreducible bundles of $2k$-forms are bundles which are given for each $0\leq \ell\leq k$ by the tensor product of $S^{2\ell}{{\mathcal E}}^{A'}[\ell-k]$ with the subbundle of $\otimes^{2k}{{\mathcal E}}_A$ which is given by the Young diagram with columns of height $k+\ell$ and $k-\ell$. Of course, if $k+\ell>n$ then the component vanishes, and if $k+\ell=n$ then we get a one more copy of ${{\mathcal E}}[-1]\cong\Lambda^n{{\mathcal E}}_A$ and the Young diagram contains $n$ boxes less. In particular, we see that for $2k\leq n$, the bundle of $2k$-forms decomposes into $k+1$ components, while for $2k>n$ we have $n-k+1$ components. The decomposition of odd degree forms is similar, and so the De Rham sequence splits into a triangular pattern, see [@Subcomplexes] for more details.
We mainly are interested in the long side of this triangle. In the case of a locally flat Grassmannian structure, there occur invariant operators which correspond to nonstandard homomorphisms of generalized Verma modules, cf. [@Lepowsky] and [@Boe]. Concretely, for each $2\leq k\leq n$ there is a fourth order invariant operator $$\widetilde{\square}_{ABCD}: { \mbox{$\mbox{$\begin{picture}(21,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k-2}$}}
\put(10,9){\vector(0,1){6}}
\put(10,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(13,13)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}[-k+2]\to{ \mbox{$\mbox{$\begin{picture}(11,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k}$}}
\put(4,9){\vector(0,1){6}}
\put(4,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(13,13)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf B}}}[-k].$$ These operators are called the nonstandard operators for Grassmannian structures. In terms of preferred flat connections (which are known to exist on locally flat structures), the operator $\widetilde{\square}_{ABCD}$ is obviously given by $
\nabla_{I'A}\nabla^{I'}_B\nabla_{J'C}\nabla^{J'}_D,
$ followed by the projection to the target bundle. Let us note that such projection is unique since the commutativity of $\nabla^{A'}_A$ ensures that $
\nabla_{I'A}\nabla^{I'}_B\nabla_{J'C}\nabla^{J'}_D$ is a section of ${{\mathcal E}}_{[AB][CD]}[-2],
$ and the target bundle appears only in the tensor product of the initial bundle with ${\mbox{$
\begin{picture}(13,10)(-1,1)
\put(0,10){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\end{picture}
$}}{{\mathcal E}}_{ABCD}[-2]$ which occurs with multiplicity one in ${{\mathcal E}}_{[AB][CD]}[-2]$.
All other invariant operators between exterior forms are components of exterior derivative and their nonzero compositions (of order two). Therefore, they exist also on general almost Grassmannian structures. On the other hand, it may be proved that the nonstandard operators are not strongly invariant, see [@Navrat]. Hence it is not clear whether these extend to invariant operators to a larger class of (almost) Grassmannian structures. However, it was proved by J. Slov' ak and R. Gover in [@Gover Theorem 5.1] that they all do exist on any Grassmannian structure of type $(2,n)$. In sequel, we construct them via so called curved Casimir operators, and thus we give an alternative proof of their existence and alternative formulae for them.
A formula for the curved Casimir operator, construction principle
-----------------------------------------------------------------
First recall from Theorem 3.4 of [@CCCBGG] that the curved Casimir ${{\mathcal{C}}}$ is and invariant operator which on an irreducible bundle $W\to M$ acts by a real multiple of the identity. We denote the corresponding scalar by $\beta_W$ and we call it the Casimir eigenvalue. By the theorem, this scalar can be computed in terms of weights of the representation which induces $W$. Namely, if the lowest weight of this representation is $-\lambda$, then $$\label{eigenvalue}
\beta_W=\langle\lambda,\lambda+2\rho\rangle,$$ where $\rho$ is the lowest form which by definition equals half the sum of all positive roots or equivalently, it equals the sum of all fundamental weights.
A suitable formula for ${{\mathcal{C}}}$ on an arbitrary natural Grassmannian bundle $V\to M$ can be obtained from the formula in terms of an adapted local frame for the adjoint tractor bundle from Proposition 3.3 of [@CCCBGG]. Following the proof of Proposition 2.2 in [@CCCex], which gives a formula for ${{\mathcal{C}}}$ in conformal geometry, one gets a formula which has precisely the same form as the formula in this proposition. Namely, having fixed a connection $\nabla$ from the class of the distinguished connections, the adjoint tractor bundle splits as $TM\oplus\operatorname{End}_0(TM)\oplus T^*M$, and having chosen a local orthonormal frame $\xi_\ell$ for $TM$ with dual frame $\varphi^\ell$ for $T^*M$ (with respect to the Cartan-Killing form), the formula for the curved Casimir operator reads as $$\label{Ca_short}
{{\mathcal{C}}}(s)=\beta(s)-2\sum_\ell{{\varphi}}^\ell\bullet \nabla_{\psi_\ell}s-2\sum_\ell{{\varphi}}^\ell\bullet{{\sf P}}(\psi_\ell)\bullet s,$$ where $\beta:V\to V$ is the bundle map which acts on each irreducible component $W\subset V$ by multiplication by $\beta_W$. For the sake of simplicity, we write this formula in a shortened form as ${{\mathcal{C}}}=\beta_W-2\nabla\bullet -2{{\sf P}}\bullet\bullet{\phantom{'}}.$
The construction principle we use is inspired by the construction of splitting operators in section 3.5 of [@CCCBGG]. Let $\mathcal{T}\to M$ be a tractor bundle. The filtration of the standard tractor bundle induces a natural filtration $\mathcal{T}=\mathcal{T}^0\supset\mathcal{T}^1\supset
\cdots\supset\mathcal{T}^N$, which we write as $\mathcal{T}=\mathcal{T}^0/\mathcal{T}^1{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}\mathcal{T}^1/\mathcal{T}^2
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}\cdots{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}\mathcal{T}^N.$ Each of the subquotients $\mathcal{T}^i/\mathcal{T}^{i+1}$ splits into a direct sum of irreducible tensor bundles. We know from above that on sections of each such irreducible $W\subset\mathcal{T}^i/\mathcal{T}^{i+1}$, the curved Casimir acts by multiplication with $\beta_W$. Let $\beta_j^1,\dots,\beta_j^{n_j}$ denote for all $j>i$ the different eigenvalues that occur in the decomposition of $\mathcal{T}^j/\mathcal{T}^{j+1}$. Then the natural operator on $\Gamma(VM)$ defined by $L:=\Pi_{j=i+1}^N\Pi_{\ell=1}^{n_j}({{\mathcal{C}}}-\beta^\ell_j)$ descends to an operator $\Gamma(WM)\to\Gamma(V^iM)$. Moreover, if for $j>i$ all eigenvalues $\beta^\ell_j$ are different from $\beta_W$, then $L$ defines a natural splitting operator. On the other hand, if an eigenvalue $\beta^\ell_j$ corresponding to an irreducible bundle $W'M\subset V^jM/V^{j+1}M$ coincides with $\beta_W$, then $L$ restricted to $WM$ defines an invariant operator $\Gamma(WM)\to\Gamma(W'M)$. A detailed explanation of the construction principle can be found in section 2.3 of [@CCCex].
Curved Casimir construction
============================
The first step of the construction of nonstandard operators is a choice of a suitable tractor bundle. Following the construction of Paneitz operator in [@CCCex], it is natural to take as the input bundle such a tractor bundle, which has the initial bundle and the target bundle of the operator in the top slot and the bottom slot respectively. Hence from the description of $\widetilde{\square}_{ABCD}$ in section \[NO\] we conclude that the right input tractor bundle for our construction is $$\label{tractor_bundle}
\mathcal{T}:={ \mbox{$\mbox{$\begin{picture}(11,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k}$}}
\put(4,9){\vector(0,1){6}}
\put(4,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(13,13)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}}}[-k].$$
The composition series of $\mathcal{T}$
---------------------------------------
The composition series for this tractor bundle can be computed from the series of the simpler bundles. Namely, it follows from the structure of the standard cotractor bundle that ${{\mathcal E}}_{[{{\alpha}}{{\beta}}]}={{\mathcal E}}_{[A'B']}{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}_{A'B}{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}_{[AB]}$. In terms of densities and multi indices, defined in \[Grassmannian bundles\] and \[Notation\] respectively, this can be also written as ${{\mathcal E}}_{{\boldsymbol{{{\alpha}}}}^2}={{\mathcal E}}[1]{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}^{A'}_{A}[1]{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}_{{{\bf A}}^2}.$ It is then easy to see that for an arbitrary $k$ the composition series of $k$-forms is $${{\mathcal E}}_{{\boldsymbol{{{\alpha}}}}}={{\mathcal E}}_{\ddot{{{\bf A}}}}[1]{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}^{A'}_{\dot{{{\bf A}}}}[1]{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}_{{\bf A}}.$$ From this composition series, one easily deduces that the second symmetric power satisfies $${{\mathcal E}}_{({\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}})}=
{{\mathcal E}}_{({{\bf \ddot{A}}}{{\bf \ddot{B}}})}[2]
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}^{A'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}[2]
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}(
{{\mathcal E}}^{(A'B')}_{({{\bf \dot{A}}}{{\bf \dot{B}}})}[2]
\oplus
{{\mathcal E}}_{[{{\bf \dot{A}}}{{\bf \dot{B}}}]}[1]
\oplus
{{\mathcal E}}_{{{\bf A}}{{\bf \ddot{B}}}}[1]
)
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}^{A'}_{{{\bf A}}{{\bf \dot{B}}}}[1]
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2,0.1){$+$}
\put(6.2,2.5){\oval(8,8)[l]}
\end{picture}$}}{{\mathcal E}}_{({{\bf A}}{{\bf B}})}.$$ The bundles displayed are not irreducible however. On the left-hand side, there appears $\mathcal{T}$ as a direct summand. It is an easy observation that the diagrams corresponding to the other summands have also two columns at most which means that the height of one of them is greater then $k$. Hence the desired composition series of $\mathcal{T}$ can be detected in the right-hand side of the previous equation by considering only such terms that vanish under alternations over more than $k$ tractor indices. Its exact form is given in the lemma below.
In order to write down the representatives of the the bundles in the composition series explicitly, it is convenient to use a notation analogous to the notation of [@prolongation]. Concretely, we use the injectors $X^A_{{\alpha}}$, $Y^{A'}_{{\alpha}}$ from to define new injectors $$\label{injectors}
\begin{array}{l}
{{\mathbb X}}^{{{\bf A}}}_{{\boldsymbol{{{\alpha}}}}}:=X^{A_1}_{[{{\alpha}}_1}\dots X^{A_k}_{{{\alpha}}_k]}\in{{\mathcal E}}^{{{\bf A}}}_{{\boldsymbol{{{\alpha}}}}},
\\
{{\mathbb W}}^{A'{{\bf \dot{A}}}}_{{\boldsymbol{{{\alpha}}}}}:=Y^{A'}_{[{{\alpha}}_1}X^{A_2}_{{{\alpha}}_2}\dots X^{A_k}_{{{\alpha}}_k]}\in{{\mathcal E}}^{A'{{\bf \dot{A}}}}_{{\boldsymbol{{{\alpha}}}}},
\\
{{\mathbb Y}}^{A'B'{{\bf \ddot{A}}}}_{{\boldsymbol{{{\alpha}}}}}:=Y^{A'}_{[{{\alpha}}_1}Y^{B'}_{{{\alpha}}_2}X^{A_3}_{{{\alpha}}_3}\dots X^{A_k}_{{{\alpha}}_k]}\in{{\mathcal E}}^{[A'B']{{\bf \ddot{A}}}}_{{\boldsymbol{{{\alpha}}}}}.
\end{array}$$ In terms of these injectors, a tractor $k$-form ${{\varphi}}_{{\boldsymbol{{{\alpha}}}}}$ can be simply expressed as $${{\varphi}}_{{\boldsymbol{{{\alpha}}}}}=
\sigma_{{{\bf \ddot{A}}}}{{\epsilon}}_{A'B'}{{\mathbb Y}}^{A'B'{{\bf \ddot{A}}}}_{{\boldsymbol{{{\alpha}}}}}
+
\mu^{A'}_{{{\bf \dot{A}}}}{{\epsilon}}_{A'B'}{{\mathbb W}}^{B'{{\bf \dot{A}}}}_{{\boldsymbol{{{\alpha}}}}}
+
\rho_{{{\bf A}}}{{\mathbb X}}^{{{\bf A}}}_{{\boldsymbol{{{\alpha}}}}},$$ where $\sigma_{{{\bf \ddot{A}}}}\in{{\mathcal E}}_{\ddot{{{\bf A}}}}[1]$, $\mu^{A'}_{{{\bf \dot{A}}}}\in{{\mathcal E}}^{A'}_{\dot{{{\bf A}}}}[1]$, $\rho_{{{\bf A}}}\in{{\mathcal E}}_{{\bf A}}$. In a similar way, one can explicitly write down the sections of ${{\mathcal E}}_{({\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}})}$ and also the sections of its subbundles, in particular the sections of our bundle $\mathcal{T}$.
\[composition\_series\]
$$\label{comp_series}
\mathcal{T}:={ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}}}[-k]=
\begin{pmatrix}
{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}[-k+2]
\\
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2.2,0.2){$+$}
\put(6,2.8){\oval(8,8)[t]}
\end{picture}$}}\\
{{\mathcal E}}^{A'}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-5)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-10){\line(0,1){25}}
\put(10,-5){\line(0,1){20}}
\put(5,-10){\line(0,1){12}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}[-k+2]
\\
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2.2,0.2){$+$}
\put(6,2.8){\oval(8,8)[t]}
\end{picture}$}}\\
{{\mathcal E}}^{(A'B')}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \dot{A}}}{{\bf \dot{B}}}}[-k+2]
\oplus
{ \mbox{$
\mbox{$\begin{picture}(13,19)(-1,-6)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-15){\line(1,0){5}}
\put(0,-15){\line(0,1){30}}
\put(10,-5){\line(0,1){20}}
\put(5,-15){\line(0,1){17}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf \ddot{B}}}}[-k+1]
\\
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2.2,0.2){$+$}
\put(6,2.8){\oval(8,8)[t]}
\end{picture}$}}\\
{{\mathcal E}}^{A'}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-5)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-10){\line(0,1){25}}
\put(10,-5){\line(0,1){20}}
\put(5,-10){\line(0,1){12}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf \dot{B}}}}[-k+1]
\\
{\mbox{$
\begin{picture}(12.7,8)(-.5,-1)
\put(2.2,0.2){$+$}
\put(6,2.8){\oval(8,8)[t]}
\end{picture}$}}\\
{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf B}}}[-k]
\end{pmatrix}$$
The composition series of tractor bundle $\mathcal{T}$ from has the form as displayed in figure \[composition\_series\]. In terms of injectors defined by , its section $v_{{\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}}}$ can be expressed as $$\label{comp_injectors}
\begin{array}{l}
v_{{\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}}}=
{\sigma}_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}{{\epsilon}}_{A'B'}{{\epsilon}}_{C'D'}{{\mathbb Y}}^{A'B'{{\bf \ddot{A}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb Y}}^{C'D'{{\bf \ddot{B}}}}_{{\boldsymbol{{{\beta}}}})}
+
\mu^{A'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}{{\epsilon}}_{A'B'}{{\epsilon}}_{C'D'}{{\mathbb W}}^{B'{{\bf \dot{A}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb Y}}^{C'D'{{\bf \ddot{B}}}}_{{\boldsymbol{{{\beta}}}})}
\\
+
A^{A'B'}_{{{\bf \dot{A}}}{{\bf \dot{B}}}}{{\epsilon}}_{A'C'}{{\epsilon}}_{B'D'}{{\mathbb W}}^{C'{{\bf \dot{A}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{D'{{\bf \dot{B}}}}_{{\boldsymbol{{{\beta}}}})}
+
{{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}{{\epsilon}}_{C'D'}({{\mathbb X}}^{{{\bf A}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb Y}}^{C'D'{{\bf \ddot{B}}}}_{{\boldsymbol{{{\beta}}}})}
+\frac{k}{2}{{\mathbb W}}^{C'A_1{{\bf \ddot{B}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{D'{{\bf \dot{A}}}}_{{\boldsymbol{{{\beta}}}})})
\\
+
\nu^{A'}_{{{\bf A}}{{\bf \dot{B}}}}{{\epsilon}}_{A'B'}{{\mathbb X}}^{{{\bf A}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{B' {{\bf \dot{B}}}}_{{\boldsymbol{{{\beta}}}})}
+
\rho_{{{\bf A}}{{\bf B}}}{{\mathbb X}}^{{{\bf A}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb X}}^{{{\bf B}}}_{{\boldsymbol{{{\beta}}}})},
\end{array}$$ where ${\sigma}_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}$, $\mu^{A'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}$, $A^{A'B'}_{{{\bf \dot{A}}}{{\bf \dot{B}}}}$, ${{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}$, $\nu^{A'}_{{{\bf A}}{{\bf \dot{B}}}}$, $\rho_{{{\bf A}}{{\bf B}}}$ are representatives of the corresponding bundles appearing in the composition series of $\mathcal{T}$.
By definition, sections of our bundle are such sections of ${{\mathcal E}}_{({\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}})}[-k]$ that vanish under all alternations over more than $k$ indices. It is an easy observation that the low dimension of ${{\mathcal E}}^{A'}$ implies that this condition translates from tractor indices to unprimed spinor indices. Hence the height of Young diagrams of all bundles in the composition series of our bundle is less or equal to $k$. Looking at the composition series of ${{\mathcal E}}_{({\boldsymbol{{{\alpha}}}}{\boldsymbol{{{\beta}}}})}$, this condition determines a unique irreducible bundle in each summand. Twisting everything by weight $-k$, these bundles already are the bundles appearing in . Their representatives in our tractor bundle are obvious except for the bundle with Young diagram with columns of heights $k$ and $k-2$. One copy of this bundle sits in ${{\mathcal E}}_{{{\bf \dot{A}}}{{\bf \dot{B}}}}[-k+1]$ and the second in ${{\mathcal E}}_{{{\bf A}}{{\bf \ddot{B}}}}[-k+1]$, and the condition on tractor indices yields a nontrivial relation between these two copies. Namely, we will show that the corresponding sections $B_{{{\bf \dot{A}}}{{\bf \dot{B}}}}$ and ${{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}$ are related by $$\label{B_al}
B_{{{\bf \dot{A}}}{{\bf \dot{B}}}}=\frac{k}{2}{{\alpha}}_{[A_2|{{\bf \dot{B}}}| {{\bf \ddot{A}}}]},
\quad
{{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}=(-1)^k(k-1)B_{{{\bf A}}{{\bf \ddot{B}}}}.$$ This will finish the proof since the first equation shows that a section ${{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}$ of the bundle with Young diagram with columns of heights $k$ and $k-2$ is represented in our tractor bundle by $${{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}{{\epsilon}}_{C'D'}({{\mathbb X}}^{{{\bf A}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb Y}}^{C'D'{{\bf \ddot{B}}}}_{{\boldsymbol{{{\beta}}}})}
+\frac{k}{2}{{\mathbb W}}^{C'A_1{{\bf \ddot{B}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{D'{{\bf \dot{A}}}}_{{\boldsymbol{{{\beta}}}})}).$$ Equations relating these two sections can be found by looking at the middle slots of $v_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}=0$. Namely, expanding the alternation in ${\boldsymbol{{{\beta}}}}$ with respect to ${{\beta}}_1$ and ${{\bf \dot{{\boldsymbol{{{\beta}}}}}}}$ yields $${{\mathbb W}}^{C'{{\bf \dot{A}}}}_{[{\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{D' {{\bf \dot{B}}}}_{{{\beta}}_1]{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}=
(-1)^{k-1}\frac1{k}{{\mathbb Y}}^{C'D'{{\bf \dot{A}}}}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb X}}^{{{\bf \dot{B}}}}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}
+\frac1{k}\sum\limits_{i=2}^{k}(-1)^{i-1}{{\mathbb W}}^{C'{{\bf \dot{A}}}B_i }_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb W}}^{D'B_2\cdots\hat{B}_i\cdots B_k}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}$$ and since $B_{{{\bf \dot{A}}}{{\bf \dot{B}}}}$ is skew symmetric in $B_2,\dots, B_k$, the alternation over tractor indices ${{\alpha}}_1,\dots,{{\alpha}}_k$ and ${{\beta}}_1$ of term $B_{{{\bf \dot{A}}}{{\bf \dot{B}}}}{{\epsilon}}_{C'D'}{{\mathbb W}}^{C'{{\bf \dot{A}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{D' {{\bf \dot{B}}}}_{{\boldsymbol{{{\beta}}}})}$ is equal to $$\label{B_skew}
B_{{{\bf \dot{A}}}{{\bf \dot{B}}}}{{\epsilon}}_{C'D'}\left((-1)^{k-1}\frac1{k}{{\mathbb Y}}^{C'D'{{\bf \dot{A}}}}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb X}}^{{{\bf \dot{B}}}}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}-\frac{k-1}{k}{{\mathbb W}}^{C'{{\bf \dot{A}}}B_2 }_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb W}}^{D'{{\bf \ddot{B}}}}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}\right),$$ while the alternation of term $A^{A'B'}_{{{\bf \dot{A}}}{{\bf \dot{B}}}}{{\epsilon}}_{A'C'}{{\epsilon}}_{B'D'}{{\mathbb W}}^{C'{{\bf \dot{A}}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb W}}^{D' {{\bf \dot{B}}}}_{{\boldsymbol{{{\beta}}}})}$ over the same tractor indices obviously vanishes. Similarly, for the other injectors the expansion of ${\boldsymbol{{{\beta}}}}$ with respect to ${{\beta}}_1$ and ${{\bf \dot{{\boldsymbol{{{\beta}}}}}}}$ yields $${{\mathbb X}}^{{{\bf A}}}_{[{\boldsymbol{{{\alpha}}}}}{{\mathbb Y}}^{C'D' {{\bf \ddot{B}}}}_{{{\beta}}_1]{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}=
(-1)^k\frac2{k}{{\mathbb W}}^{[C'|{{\bf A}}|}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb W}}^{D']{{\bf \ddot{B}}}}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}
-\frac1{k}\sum\limits_{i=3}^{k}(-1)^i{{\mathbb X}}^{{{\bf A}}B_i}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb Y}}^{C'D'B_3\cdots\hat{B}_i\cdots B_k}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}$$ and $${{\mathbb X}}^{{{\bf A}}}_{[{{\beta}}_1|{\boldsymbol{{{\beta}}}}|}{{\mathbb Y}}^{C'D' {{\bf \ddot{B}}}}_{{\boldsymbol{{{\alpha}}}}]}=
\frac1{k}\sum\limits_{i=1}^{k}(-1)^{i-1}{{\mathbb Y}}^{C'D' {{\bf \ddot{B}}}A_i}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb X}}^{A_1\cdots\hat{A}_i\cdots A_k}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}.$$ Concerning the first formula, the second term on the right vanishes when applied to ${{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}$ since each summand is skew symmetric in $k+1$ unprimed indices. Skewing the second formula in $A_1,\dots,A_k$ one obtains that the alternation over tractor indices ${{\alpha}}_1,\dots,{{\alpha}}_k$ and ${{\beta}}_1$ of term ${{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}{{\epsilon}}_{C'D'}{{\mathbb X}}^{{{\bf A}}}_{({\boldsymbol{{{\alpha}}}}}{{\mathbb Y}}^{C'D' {{\bf \ddot{B}}}}_{{\boldsymbol{{{\beta}}}})}$ is $$\label{alpha_skew}
{{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}{{\epsilon}}_{C'D'}\left((-1)^k\frac1{k}({{\mathbb W}}^{C'{{\bf A}}}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb W}}^{D'{{\bf \ddot{B}}}}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}+\frac12{{\mathbb Y}}^{C'D' {{\bf \ddot{B}}}A_1}_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]}{{\mathbb X}}^{{{\bf \dot{A}}}}_{{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}\right).$$ Then $v_{[{\boldsymbol{{{\alpha}}}}{{\beta}}_1]{{\bf \dot{{\boldsymbol{{{\beta}}}}}}}}=0$ implies that the sum of and is equal to zero, and this immediately gives equations .
The action of one forms on $\mathcal{T}$
----------------------------------------
The dual to standard representation of ${\mathfrak{sl}}(2+n)$ on ${{\mathbb R}}^{2+n}$ restricted to ${{\mathfrak g}}_1$ translates to bundles and gives rise to an action of ${{\mathcal E}}_a$ on ${{\mathcal E}}_{{{\alpha}}}$. It is easy to see that for an one-form ${{\varphi}}^{A'}_A\in{{\mathcal E}}^{A'}_A={{\mathcal E}}_a$ and a standard cotractor $v_{{\alpha}}=v_{A'}Y^{A'}_{{\alpha}}+v_aX^A_{{\alpha}}$ its explicit form reads as ${{\varphi}}\bullet v_{{\alpha}}=-{{\varphi}}^{A'}_A v_{A'} X^A_{\alpha}$. Hence the action of ${{\varphi}}^{A'}_A$ on the injectors is given by equations $$\begin{array}{cccc}
({{\varphi}}\bullet Y)^{A'}_{\alpha}=-{{\varphi}}^{A'}_A X^A_{\alpha},&{{\varphi}}\bullet X=0.
\end{array}$$ Using these basic relations one is able to compute easily the action of one forms on more complicated tractor bundles.
The action of an one form ${{\varphi}}^{A'}_A\in{{\mathcal E}}^{A'}_A$ on a section of $\mathcal{T}$ is given by $$\label{action}
{{\varphi}}^{A'}_A\bullet
\begin{pmatrix}
{\sigma}_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
\\
\mu^{A'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}
\\
A^{A'B'}_{{{\bf \dot{A}}}{{\bf \dot{B}}}} | {{\alpha}}_{{{\bf A}}{{\bf \ddot{B}}}}
\\
\nu^{A'}_{{{\bf A}}{{\bf \dot{B}}}}
\\
\rho_{{{\bf A}}{{\bf B}}}
\end{pmatrix}
=
\begin{pmatrix}
0\\
{{\varphi}}^{A'}_{[A_2}{\sigma}_{{{\bf \ddot{A}}}]{{\bf \ddot{B}}}}
\\
{{\varphi}}^{(A'}_{[A_2}\mu^{B')}_{|{{\bf \dot{B}}}|{{\bf \ddot{A}}}]}
+{{\varphi}}^{(A'}_{[B_2}\mu^{B')}_{|{{\bf \dot{A}}}|{{\bf \ddot{B}}}]}
\quad|\quad
{{\varphi}}_{I'[A_1}\mu^{I'}_{{{\bf \dot{A}}}]{{\bf \ddot{B}}}}
\\
2{{\varphi}}_{I'[A_1} A^{I'A'}_{{{\bf \dot{A}}}]{{\bf \dot{B}}}}
+2{{\varphi}}^{A'}_{[B_2}\alpha_{|{{\bf A}}|{{\bf \ddot{B}}}]}
-k(-1)^k{{\varphi}}^{A'}_{[A_1}\alpha_{{{\bf \dot{A}}}]{{\bf \dot{B}}}}
\\
{{\varphi}}_{I'[A_1}\nu^{I'}_{|{{\bf B}}|{{\bf \dot{A}}}]}
+{{\varphi}}_{I'[B_1}\nu^{I'}_{|{{\bf A}}|{{\bf \dot{B}}}]}
\end{pmatrix}$$
From the defining equations for the injectors ${{\mathbb X}}^{{{\bf A}}}_{{\boldsymbol{{{\alpha}}}}}$, ${{\mathbb W}}^{A'{{\bf \dot{A}}}}_{{\boldsymbol{{{\alpha}}}}}$, ${{\mathbb Y}}^{A'B'{{\bf \ddot{A}}}}_{{\boldsymbol{{{\alpha}}}}}$ and from the basic relations for the action of an one form ${{\varphi}}^{A'}_A\in{{\mathcal E}}^{A'}_A$ on injectors $X^A_{{\alpha}},Y^{A'}_{{\alpha}}$, one immediately derives $$\begin{array}{l}
({{\varphi}}\bullet {{\mathbb Y}})^{A'B'{{\bf \ddot{A}}}}_{{\boldsymbol{{{\alpha}}}}}=2{{\varphi}}^{[A'}_{I}{{\mathbb W}}^{B']I{{\bf \ddot{A}}}}
\\
({{\varphi}}\bullet {{\mathbb W}})^{A'{{\bf \dot{A}}}}_{{\boldsymbol{{{\alpha}}}}}=-{{\varphi}}^{A'}_I{{\mathbb X}}^{I{{\bf \dot{A}}}}_{{\boldsymbol{{{\alpha}}}}}
\\
\phantom{(}
{{\varphi}}\bullet{{\mathbb X}}=0.
\end{array}$$ Now a direct use of these equations to yields the result.
Casimir eigenvalues on $\mathcal{T}$
------------------------------------
Knowing the irreducible bundles occurring in the composition series of $\mathcal{T}$, we can directly compute the corresponding Casimir eigenvalues using formula . Since the minus lowest weights of basic bundles ${{\mathcal E}}^{A'}$, ${{\mathcal E}}[1]$ and $\Lambda^k{{\mathcal E}}_{A}={{\mathcal E}}_{{\bf A}}$ are given by ${\omega}_1-{\omega}_2$, ${\omega}_2$ and ${\omega}_{k+2}-{\omega}_2$ respectively, where ${\omega}_k$ for $k=1,\dots, n-1$ denotes fundamental weights of ${{\mathfrak g}}$, the minus lowest weights corresponding to the bundles in are $$\begin{pmatrix}
&\lambda_0&
\\
&\lambda_1&
\\
\lambda_2^1&|&\lambda_2^2
\\
&\lambda_3&
\\
&\lambda_4&
\end{pmatrix}
=
\begin{pmatrix}
2{\omega}_{k}-k{\omega}_2
\\
{\omega}_{k+1}+{\omega}_{k}-(k+1){\omega}_2+{\omega}_1
\\
2{\omega}_{k+1}-(k+2){\omega}_2+2{\omega}_1
\quad|\quad
{\omega}_{k+2}+{\omega}_k-(k+1){\omega}_2
\\
{\omega}_{k+2}+{\omega}_{k+1}-(k+2){\omega}_2+{\omega}_1
\\
2{\omega}_{k+2}-(k+2){\omega}_2
\end{pmatrix}.$$ The computation of Casimir eigenvalues from formula is parallel to [@CCCex] using these weights as an input. We get the following.
\[eigenvalues\] The eigenvalues of the curved Casimir operator on the irreducible bundles appearing in the composition series of $\mathcal{T}$ read as $$\begin{pmatrix}
&\beta_0&
\\
&\beta_1&
\\
\beta_2^1&|&\beta_2^2
\\
&\beta_3&
\\
&\beta_4&
\end{pmatrix}
=
\begin{pmatrix}
&0&
\\
&0&
\\
4&|&-4
\\
&0&
\\
&0&
\end{pmatrix}.$$
Construction of nonstandard operators
-------------------------------------
Looking at Casimir eigenvalues for irreducible components of the tractor bundle $\mathcal{T}$ displayed in lemma \[eigenvalues\] we see that the candidates for curved analogues of the nonstandard operators $\widetilde{\square}_{ABCD}$ are induced by ${{\mathcal{C}}}^3\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$. By the construction principle, such a composition of the curved Casimirs gives an invariant operator between the top slot and the bottom slot which are exactly the bundles where the nonstandard operators is defined. However, we will show that such direct construction yields always a trivial operator. The reason for this might be seen in the high degeneracy of this case – there appear four zeros among the Casimir eigenvalues. On the other hand, the curved Casimirs induce more operators due to the degeneracy. Namely, the operator ${{\mathcal{C}}}$ itself evidently gives an invariant operator between first two slots and an invariant operator between the last two slots. It is easy to see that these operators are nothing else but the exterior derivatives. The next operator we get is due to the coincidence of eigenvalues $\beta_1=\beta_3$. By the construction principle, an invariant operator between the respective bundles $${\mbox{\sf M}}_{AB}:{{\mathcal E}}^{A'}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-5)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-10){\line(0,1){25}}
\put(10,-5){\line(0,1){20}}
\put(5,-10){\line(0,1){12}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}[-k+2]\rightarrow
{{\mathcal E}}^{A'}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-5)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-10){\line(0,1){25}}
\put(10,-5){\line(0,1){20}}
\put(5,-10){\line(0,1){12}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf \dot{B}}}}[-k+1]$$ is induced by ${{\mathcal{C}}}\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$. Similarly, the coincidences $\beta_0=\beta_3$ and $\beta_1=\beta_4$ show that ${{\mathcal{C}}}^2\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$ gives rise to invariant operators between the corresponding bundles. Obviously, these operators are given by the compositions ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$. But, there are no such operators in the classification of invariant operators on flat structures, and so the compositions must vanish in that case. We will prove that ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ vanish identically even in the case of a general torsion-free structure with a nonzero curvature. It implies then that ${{\mathcal{C}}}^2\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$ actually induces an invariant operator from the top slot to the bottom slot, and we will show that this is the curved analogue of the nonstandard operator that we wanted to construct. Let us remark that this construction breaks down in the case that the torsion does not vanish, see Remark \[torsion\] bellow.
\[thm\] Let $M$ be a manifold endowed with Grassmannian or quaternionic structure. For each integer $k$ such that $2\leq k\leq n$ the action of operator ${{\mathcal{C}}}^2\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$ on tractor bundle $\mathcal{T}$ gives rise to a fourth order invariant operator $$\square_{ABCD}: { \mbox{$\mbox{$\begin{picture}(21,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k-2}$}}
\put(10,9){\vector(0,1){6}}
\put(10,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(13,13)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}[-k+2]\to{ \mbox{$\mbox{$\begin{picture}(11,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k}$}}
\put(4,9){\vector(0,1){6}}
\put(4,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(13,13)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf B}}}[-k]$$ which coincides with the corresponding nonstandard operator on flat structures.
First, we prove that for each $k$ such that $0\leq k\leq n-2$ the operators ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ vanish identically. In the second step, we prove that the principal part of $\square_{ABCD}$ is given by a nonzero scalar multiple of the corresponding nonstandard operator $\widetilde{\square}_{ABCD}$, and thus the operators coincide on the category of locally flat structures.
The direct way how to prove vanishing of the two operators is to express the operator ${\mbox{\sf M}}$ in terms of the Weyl connection and Rho-tensor. This can be easily done in a way since the operator is obtained by acting with ${{\mathcal{C}}}\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$, and we have formula for ${{\mathcal{C}}}$ at disposal. Namely, denoting the projections to the left slot (corresponding to $\beta^1_2$) and right slot (corresponding to $\beta^2_2$) by $({\phantom{'}})_1$ and $({\phantom{'}})_2$ respectively, it is an easy exercise to show that up to a scalar multiple ${\mbox{\sf M}}$ can be written as $$\label{M}
{\mbox{\sf M}}=\nabla \bullet (\nabla\bullet{\phantom{'}})_1-\nabla\bullet(\nabla\bullet{\phantom{'}})_2 +2{{\sf P}}\bullet\bullet{\phantom{'}}.$$ For more details see [@thesis section 3.1.4]. Obviously, the compositions ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ are given by ${\mbox{\sf M}}(\nabla\bullet{\phantom{'}})$ and $\nabla\bullet{\mbox{\sf M}}$ respectively. Having these abstract formulas in hand, we can directly use the form of the action $\bullet$ given by equation to show explicitly that the operators vanish. The computation becomes tedious soon however. Therefore, we rather prove the vanishing by analyzing terms that might occur in formulae for these operators.
The first observation is that the principal (third order) part of operators ${\mbox{\sf M}}(\nabla\bullet{\phantom{'}})$ and $\nabla\bullet{\mbox{\sf M}}$ vanishes in the torsion-free case, and thus ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ are operators of order at most one with the curvature in their leading part. This follows from the classification of invariant operators in the flat case – it is well known that there are no operators between the respective bundles, see e.g. [@BoeCollingwood]. It can be also checked directly by use of for the first two terms in . This is doable since we may freely commute the derivatives. It can be also partially deduced from the computation of the principal part of $\square_{ABCD}$ below.
Since there are no terms of order three in ${\mbox{\sf M}}\circ d (\varphi)$ and $ d \circ{\mbox{\sf M}}(\psi)$ for each $\varphi$ and $\psi$, the operators are given by projections to target bundles of a linear combination of terms of the form $R\cdot\nabla\phi$ and $\nabla R\cdot\phi$, where $\phi=\varphi$ and $\phi=\psi$ respectively. According to the description of curvature in section \[GS\], the operators may be also written as projections of a linear combination of terms of the form $W\cdot\nabla\phi, \nabla W\cdot\phi$ and ${{\sf P}}\cdot\nabla\phi, \nabla {{\sf P}}\cdot\phi$, where the tensor ${{\sf P}}_{ab}$ is symmetric, and the symmetries of tensor $W_{ab}{}^d{}_{c}$ and exterior derivative of the Rho-tensor are described in and respectively. Since ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ are natural by construction, given a point $x$ we can compute $M( d \varphi)(x)$ and $ d ({\mbox{\sf M}}\psi)(x)$ for each $\varphi, \psi$ in terms of any distinguished connection. So let us choose a distinguished connection $\nabla_a$ such that ${{\sf P}}_{(ab)}=0$ and also $\nabla_{(a}{{\sf P}}_{bc)}=0$ at $x$. Its existence follows from the transformation formulas of the Rho-tensor and its covariant derivative. One can actually demand that the total symmetrizations of all derivatives of ${{\sf P}}_{ab}$ vanish at $x$. This is a general feature which refers to the so called normal scale for a parabolic geometry, see [@parabook theorem 5.1.12]. Evidently, $M( d \varphi)(x)$ and $ d ({\mbox{\sf M}}\psi)(x)$ for such a connection are then given by projections of terms which contain either tensor $W_{ab}{}^d{}_{c}$ or tensor $(d^\nabla{{\sf P}})_{abc}=\nabla_{[a}{{\sf P}}_{b]c}$. However, by and both tensors are symmetric in three unprimed spinor indices while the Young diagrams of target bundles has only two columns - recall that the target bundles of the projections are the two bundles in the bottom of composition series . Thus all terms vanish under the projections. Since $x$, $\varphi$, $\psi$ were arbitrary, we conclude that the invariant operators ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ are the zero operators.
Now we know that $\square_{ABCD}$ is invariant. To finish the proof, we need to show that its principal part is given by a nonzero scalar multiple of $\widetilde{\square}_{ABCD}$. First, we find even a full formula in an abstract form. By our construction, $\square_{ABCD}$ is given by a formula in the bottom slot of the section of tractor bundle which is given by the action of ${{\mathcal{C}}}^2\circ({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$ on the section with $\sigma$ in the top slot and zeros elsewhere. Acting with ${{\mathcal{C}}}^2$ first, by formula for ${{\mathcal{C}}}$ we get zeros everywhere except for the middle slot and the slot below, where up to the factor 4 we get $(\nabla\bullet\nabla\bullet\sigma)_{1,2}\mp 2({{\sf P}}\bullet\bullet\sigma)_{1,2}$ and ${{\sf P}}\bullet\bullet\nabla\bullet\sigma$ respectively. Further, acting with $({{\mathcal{C}}}-4)\circ({{\mathcal{C}}}+4)$ we get zeros everywhere except for the bottom slot (in the slot above the bottom there is ${\mbox{\sf M}}\operatorname({ d }\sigma)=0$), and it is easy to deduce that, up to a scalar multiple, the formula there reads as $$\label{Do}
\begin{array}{l}
\square \sigma=
\nabla\bullet\nabla\bullet (\nabla\bullet\nabla\bullet \sigma)_1+\nabla\bullet\nabla\bullet (\nabla\bullet\nabla\bullet \sigma)_2
-2{{\sf P}}\bullet\bullet (\nabla\bullet\nabla\bullet \sigma)_1
\\
+2{{\sf P}}\bullet\bullet(\nabla\bullet\nabla\bullet \sigma)_2
-2\nabla\bullet\nabla\bullet({{\sf P}}\bullet\bullet \sigma)_1+2\nabla\bullet\nabla\bullet({{\sf P}}\bullet\bullet \sigma)_2.
\end{array}$$ A detailed derivation of this formula can be found in [@thesis section 3.1.6]. Thus we see that the principal part is given by the sum of the two possible paths from the top slot to the bottom slot. Now we make the corresponding two terms explicit by a multiple use of formula for $\bullet$. We directly get that the term corresponding to the path through slot $({\phantom{'}})_1$ is given by $
8\nabla_{J'A_1}\nabla_{I'B_1}\nabla^{(I'}_{(B_2}\nabla^{J')}_{A_2)}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}},
$ followed by the (unique) projection to the target bundle. Recall that the target bundle is the bundle in the bottom slot of , i.e. $$\label{target}
{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf B}}}[-k]\subset{{\mathcal E}}_{({{\bf A}}{{\bf B}})}[-k]$$ and thus the projection is given by alternation in $A_1,\dots,A_k$ and in $B_1,\dots,B_k$, followed by symmetrization in ${{\bf A}}$ and ${{\bf B}}$ and projection to the joint kernel of alternations in more than $k$ inputs. Note that we may freely commute the derivatives since we are only interested in the principal part. In particular, the symmetry of primed indices of covariant derivatives translates to unprimed indices and vice versa. Hence expanding the symmetrizations in the previous formula, we get an equivalent formula $$4\nabla_{J'A_1}\nabla^{J'}_{A_2}\nabla_{I'B_1}\nabla^{I'}_{B_2}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
+4\nabla_{J'A_1}\nabla^{I'}_{A_2}\nabla_{I'B_1}\nabla^{J'}_{B_2}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}.$$ Moreover, since the explicit form of the isomorphism ${{\mathcal E}}^{[A'B']}\cong{{\mathcal E}}[-1]$ reads as $v^{[A'B']}=-1/2 \cdot v_{I'}{}^{I'}{{\epsilon}}^{A'B'}$, the second summand is the previous display is equivalent to $-\nabla_{J'A_1}\nabla^{J'}_{A_2}\nabla_{I'B_1}\nabla^{I'}_{B_2}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}.$ Therefore, the principal part of $\nabla\bullet\nabla\bullet (\nabla\bullet\nabla\bullet \sigma)_1$ is given by the projection of $
3\nabla_{J'A_1}\nabla^{J'}_{A_2}\nabla_{I'B_1}\nabla^{I'}_{B_2}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}},
$ and thus it coincides with $3\cdot\widetilde{\square}_{ABCD}$. Similarly, we deduce directly by a multiple use of that the leading term corresponding to the path through slot $({\phantom{'}})_2$ is given by $$\label{path_2}
4\nabla_{I'A_1}\nabla^{I'}_{A_2}\nabla_{J'B_1}\nabla^{J'}_{B_2}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
-2k(-1)^k\nabla_{I'A_1}\nabla^{I'}_{B_1}\nabla_{J'[B_2}\nabla^{J'}_{B_3}
\sigma_{{{\bf \dddot{B}}}A_2] {{\bf \ddot{A}}}},$$ followed by the projection to target bundle . The second term can be rewritten by expanding the displayed alternation with respect to $A_2$ and ${{\bf \ddot{B}}}$ as follows $$\begin{array}{l}
\nabla_{J'[B_2}\nabla^{J'}_{B_3}
\sigma_{{{\bf \dddot{B}}}A_2] {{\bf \ddot{A}}}}=
\frac{1}{k}(-1)^{k-1}\nabla_{J'A_2}\nabla^{J'}_{[B_2}\sigma_{{{\bf \ddot{B}}}]{{\bf \ddot{A}}}}
+\frac{1}{k}(-1)^{k-2}\nabla_{J'[B_2}\nabla^{J'}_{|A_2|}\sigma_{{{\bf \ddot{B}}}]{{\bf \ddot{A}}}}
\\
+\frac{k-2}{k}\nabla_{J'[B_2}\nabla^{J'}_{B_3}\sigma_{{{\bf \dddot{B}}}]A_2{{\bf \ddot{A}}}}.
\end{array}$$ Since $\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}$ is a section of the bundle which corresponds to Young diagram with two columns of height $k-2$, the last term on the right hand side in the previous equation vanishes under the alternation over $A_2,\dots,A_k$. Therefore, we can substitute $\nabla_{J'[B_2}\nabla^{J'}_{B_3}
\sigma_{{{\bf \dddot{B}}}A_2] {{\bf \ddot{A}}}}$ in by $\frac{2}{k}(-1)^{k-1}\nabla_{J'[A_2}\nabla^{J'}_{B_2]}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}$, and thus we conclude that the principal part of $\nabla\bullet\nabla\bullet (\nabla\bullet\nabla\bullet \sigma)_2$ is given by the projection of $$4\nabla_{I'A_1}\nabla^{I'}_{A_2}\nabla_{J'B_1}\nabla^{J'}_{B_2}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
+4\nabla_{I'[A_1}\nabla^{I'}_{B_1]}\nabla_{J'[A_2}\nabla^{J'}_{B_2]}\sigma_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}$$ to the target bundle. Now we observe that this is actually the same formula as in the case of the first path up to a commutation of derivatives. Hence by the same reasons as above, the principal part of $\nabla\bullet\nabla\bullet (\nabla\bullet\nabla\bullet \sigma)_2$ coincides with $3\cdot\widetilde{\square}_{ABCD}$. The principal part of $\square_{ABCD}$ then equals $6\cdot\widetilde{\square}_{ABCD}$.
The proposition shows in particular that in the torsion-free case there exist a curved analogue for each of the nonstandard operators. Hence it gives an alternative proof of Theorem 5.1 of [@Gover]. Moreover, our construction directly yields a formula for each of these curved analogues, cf. formula . Making this formula explicit and using the symmetries of the Rho-tensor and its exterior derivative, we get the following.
\[factorization\] The Grassmannian nonstandard operator can be written as the projection of an operator $ d \circ A\circ d $ to bundle , where $A$ is a noninvariant operator $$A_{AB}:{{\mathcal E}}^{A'}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-5)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-10){\line(0,1){25}}
\put(10,-5){\line(0,1){20}}
\put(5,-10){\line(0,1){12}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}[-k+2]\rightarrow
{{\mathcal E}}^{A'}\otimes{ \mbox{$
\mbox{$\begin{picture}(13,17)(-1,-5)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-10){\line(1,0){5}}
\put(0,-10){\line(0,1){25}}
\put(10,-5){\line(0,1){20}}
\put(5,-10){\line(0,1){12}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathcal E}}_{{{\bf A}}{{\bf \dot{B}}}}[-k+1]$$ which is given by $$\begin{array}{l}
(A\mu)^{A'}_{{{\bf A}}{{\bf \dot{B}}}}=
2\nabla_{J'A_1}\nabla^{(A'}_{A_2}\mu^{J')}_{{{\bf \dot{B}}}{{\bf \ddot{A}}}}
+2\nabla_{J'A_1}\nabla^{(A'}_{B_2}\mu^{J')}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}
+2\nabla^{A'}_{B_2}\nabla_{J'A_1}\mu^{J'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}
\\
\phantom{qqq}
+\nabla^{A'}_{A_1}\nabla_{J'B_2}\mu^{J'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}
-\nabla^{A'}_{A_1}\nabla_{J'A_2}\mu^{J'}_{{{\bf \dot{B}}}{{\bf \ddot{A}}}}
+16{{\sf P}}^{(A'J')}_{(B_2A_1)}\mu_{J'{{\bf \dot{A}}}{{\bf \ddot{B}}}}
+8{{\sf P}}_{J'A_1A_2}^{{\phantom{'}}{\phantom{'}}J'}\mu^{A'}_{{{\bf \dot{B}}}{{\bf \ddot{A}}}}
\\
\phantom{qqq}
+8{{\sf P}}_{J'[A_1B_2]}^{{\phantom{'}}{\phantom{'}}J'}\mu^{A'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}.
\end{array}$$
Applying the explicit form of the action $\bullet$ from to formula , we directly get that $(\square f)_{{{\bf A}}{{\bf B}}}$ is given by the action of $$\begin{array}{l}
\nabla_{A_1A_2B_1B_2}=
8\nabla_{I'(A_1}\nabla_{|J'|B_1)}\nabla^{(I'}_{(A_2}\nabla^{J')}_{B_2)}
+2\nabla_{I'A_1}\nabla^{I'}_{A_2}\nabla_{J'B_1}\nabla^{J'}_{B_2}
\\
\phantom{qqq}
+2\nabla_{I'B_1}\nabla^{I'}_{B_2}\nabla_{J'A_1}\nabla^{J'}_{A_2}
+4\nabla_{I'[A_1}\nabla^{I'}_{B_1]}\nabla_{J'[A_2}\nabla^{J'}_{B_2]}
\\
\phantom{qqq}
-16{{\sf P}}_{I'(A_1|J'|B_1)}\nabla^{(I'}_{(A_2}\nabla^{J')}_{B_2)}
+4{{\sf P}}_{I'{\phantom{'}}A_1A_2}^{{\phantom{'}}{\phantom{'}}I'}\nabla_{J'B_1}\nabla^{J'}_{B_2}
\\
\phantom{qqq}
+4{{\sf P}}_{I'{\phantom{'}}B_1B_2}^{{\phantom{'}}{\phantom{'}}I'}\nabla_{J'A_1}\nabla^{J'}_{A_2}
+8{{\sf P}}_{I'{\phantom{'}}[A_1B_1]}^{{\phantom{'}}{\phantom{'}}I'}\nabla_{J'[A_2}\nabla^{J'}_{B_2]}
\\
\phantom{qqq}
-16\nabla_{I'(A_1}\nabla_{|J'|B_1)}{{\sf P}}^{(I'J')}_{(A_2B_2)}
+4\nabla_{I'A_1}\nabla^{I'}_{A_2}{{\sf P}}_{J'{\phantom{'}}B_1B_2}^{{\phantom{'}}{\phantom{'}}J'}
\\
\phantom{qqq}
+4\nabla_{I'B_1}\nabla^{I'}_{B_2}{{\sf P}}_{J'{\phantom{'}}A_1A_2}^{{\phantom{'}}{\phantom{'}}J'}
+8\nabla_{I'[A_1}\nabla^{I'}_{B_1]}{{\sf P}}_{J'[A_2B_2]}^{{\phantom{'}}{\phantom{'}}J'},
\end{array}$$ on a section $f_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}$. Now we simplify this formula by using symmetries of the covariant derivative of the Rho-tensor. Namely, we can replace each tensor $\nabla{{\sf P}}$ by its totally symmetric part since all other components are symmetric in three unprimed spinor indices and thus all terms containing them vanish when projected to target . Hence the we get the identity $$-2\nabla_{J'B_1}{{\sf P}}^{(I'J')}_{(A_2B_2)}
+\nabla^{I'}_{A_2}{{\sf P}}^{{\phantom{'}}{\phantom{'}}J'}_{J'B_1B_2}
+\nabla^{I'}_{B_1}{{\sf P}}^{{\phantom{'}}{\phantom{'}}J'}_{J'[A_2B_2]}=0$$ Applying this equation to the previous formula we conclude that the sum of terms of type $\nabla\nabla{{\sf P}}f$ is equal to the sum of terms of type ${{\sf P}}\nabla\nabla f$ and that the lower order terms may be written as $$\begin{array}{l}
\nabla_{I'A_1}(-16{{\sf P}}^{(I'J')}_{(A_2B_2)}\nabla_{J'B_1}
+8{{\sf P}}^{{\phantom{'}}{\phantom{'}}J'}_{J'B_1B_2}\nabla^{I'}_{A_2}+8{{\sf P}}^{{\phantom{'}}{\phantom{'}}J'}_{J'[A_2B_2]}\nabla^{I'}_{B_1})f_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
\\
+\nabla_{I'B_1}(-16{{\sf P}}^{(I'J')}_{(B_2A_2)}\nabla_{J'A_1}
+8{{\sf P}}^{{\phantom{'}}{\phantom{'}}J'}_{J'A_1A_2}\nabla^{I'}_{B_2}+8{{\sf P}}^{{\phantom{'}}{\phantom{'}}J'}_{J'[B_2A_2]}\nabla^{I'}_{A_1})f_{{{\bf \ddot{A}}}{{\bf \ddot{B}}}}.
\end{array}$$ Rewriting the leading part accordingly, the result follows by applying formulas for the exterior derivative $(df)^{A'}_{{{\bf \dot{A}}}{{\bf \ddot{B}}}}=\nabla^{A'}_{[A_2}f_{{{\bf \ddot{A}}}]{{\bf \ddot{B}}}}$ and $(d\nu)_{{{\bf A}}{{\bf B}}}=\nabla_{I'[A_1}\nu^{I'}_{{{\bf \dot{A}}}]{{\bf B}}}
+\nabla_{I'[B_1}\nu^{I'}_{{{\bf \dot{B}}}]{{\bf A}}}$.
\[torsion\] [The torsion-freeness of the structure is important for the proof of proposition \[thm\]. In the case of nonvanishing torsion, the Weyl curvature does not lie in the irreducible bundle but consists of the harmonic part and some other irreducible components which can be expressed in terms of torsion. Of course, also the derivative of the Rho-tensor does not lie in . The consequence of these facts is that the operators ${\mbox{\sf M}}\circ d $ and $ d \circ {\mbox{\sf M}}$ do not vanish. Namely, it is easy to show that they both are second order operators with the torsion in their leading part. Our construction breaks down therefore. However, this does not mean that curved analogues of the nonstandard operators do not exist in such a case. Indeed, it is proved in [@thesis Section 3.2] that the operator which acts on functions exists also in the presence of a nonzero torsion. On the other hand, it is proved there that this operator cannot be written as in corollary \[factorization\], i.e. as a composition of operators with the exterior derivatives in the beginning and at the end. ]{}
Weak invariance of nonstandard operators
========================================
The invariance of operators $\square_{ABCD}$ constructed above obviously depends on the vanishing of operators ${\mbox{\sf M}}\circ d $ and $ d \circ{\mbox{\sf M}}$ in the torsion-free case. The first step of the proof of the vanishing of these operators was an observation that their leading (third order) terms can be rewritten in terms of the first derivative and the curvature. Hence we need to commute covariant derivatives in order to prove the invariance of operators $\square_{ABCD}$. This shows in turn that if we replace the distinguished connection $\nabla$ by the coupled distinguished tractor connection, defined by the Leibnitz rule in the usual way, then the formula obtained from the construction does not define invariant operator on tractor bundles. The transformation of such operator will consist of terms containing tractor curvature in general. A natural question now is whether there exist a formula for $\square_{ABCD}$ which is universal in the sense that it defines invariant operators also on tractor bundles or there is no such formula, i.e. whether $\square_{ABCD}$ is strongly invariant or not. We will mainly use a slightly different notion of strong invariance. Namely, an operator is strongly invariant in the sense of [@Eastwood] if it factors through semi-holonomic jets. In the dual picture, it means that it is induced by a homomorphism of semi-holonomic Verma modules. Let us remark that such an operator then translates to tractor bundles, see [@translation]. We will prove in this section that operators $\square_{ABCD}$ are not strongly invariant in the algebraic sense. In the subsequent remark, we will argue that the operators are also not strongly invariant in the sense of the existence of a universal formula.
Nonstandard operators are not strongly invariant
------------------------------------------------
Let us digress to the case of locally flat structures for a moment. It is well known that then the jet bundles are associated to the Cartan bundle. Hence invariant differential operators are in a bijective correspondence with homomorphisms between jet prolongations of representations which induce the bundles in question. According to the definition of the flat nonstandard operators $\widetilde{\square}_{ABCD}$, the respective homomorphisms are $\mathfrak{p}$-homomorphisms $\Phi: \mathcal{J}^4(\mathbb{V}_{k-2})\to\mathbb{V}_{k}$, where for each $2\leq k\leq n$ we set $$\mathbb{V}_k:={ \mbox{$\mbox{$\begin{picture}(11,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k}$}}
\put(4,9){\vector(0,1){6}}
\put(4,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(13,13)(-1,-1)
\put(0,15){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathbb R}}^{n*}[-k]$$ for the module inducing bundle , and where $\mathcal{J}^4(\mathbb{V}_{k})$ denotes a module inducing the fourth jet prolongation of this bundle. By the description of the flat nonstandard operators in section \[NO\], for each convenient $k$ the map $\Phi$ is given by a composition of (up to a multiple) unique ${{\mathfrak g}}_0$-homomorphism $\phi: S^4{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}
\to\mathbb{V}_{2}\otimes\mathbb{V}_{k-2}$, which in terms of abstract indices reads $$\label{Phi}
\phi({\omega})_{ABCD{{\bf E}}{{\bf F}}}=
\frac12({\omega}_{I'[AB]J'[CD]{{\bf E}}{{\bf F}}}^{{\phantom{'}}{\phantom{'}}I' {\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}J'}
+{\omega}_{I'[CD]J'[AB]{{\bf E}}{{\bf F}}}^{{\phantom{'}}{\phantom{'}}I' {\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}J'})
-{\omega}_{I'[AB|J'|CD]{{\bf E}}{{\bf F}}}^{{\phantom{'}}{\phantom{'}}I' {\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}{\phantom{'}}J'},$$ with a unique projection $\mathbb{V}_{2}\otimes\mathbb{V}_{k-2}\to\mathbb{V}_{k}$. We will manifest in the course of the forthcoming proof that $\Phi$ is indeed a $\mathfrak{p}$-homomorphism.
In contrary to ordinary jet bundles, the semi-holonomic jet bundles are associated to the Cartan bundle also on structures with nonzero curvature. The respective representation, denoted by $\bar{\mathcal{J}}^k()$, is called the semi-holonomic jet prolongation. The dual module is the so called semi-holonomic Verma module. For more details see [@CSS1] and [@Eastwood]. The operators which are strongly invariant in the algebraic sense are exactly the operators which are induced by homomorphisms between these modules. We will prove that operator $\square_{ABCD}$ does not belong to them. That is, there does not exist any $\mathfrak{p}$-homomorphism $\widetilde{\Phi}: \bar{\mathcal{J}}^4(\mathbb{V}_{k-2})\to \mathbb{V}_k$ such that its restriction to the holonomic jets $ \mathcal{J}^4(\mathbb{V}_{k-2})\subset \bar{\mathcal{J}}^4(\mathbb{V}_{k-2})$ coincides with $\Phi$ (this condition says in other words that $\square_{ABCD}$ coincides with $\widetilde{\square}_{ABCD}$ on locally flat structures).
\[weak\_invariance\] The Grassmannian nonstandard operators are not strongly invariant.
We shall prove that there exists no $\mathfrak{p}$-homomorphism $\widetilde{\Phi}$ which extends $\Phi$. We prove it in two steps. By [@CSS1 Lemma 5.8], a ${{\mathfrak g}}_0$-homomorphism $\widetilde{\Phi}$ is $\mathfrak{p}$-homomorphism if and only if it factors through $\mathcal{J}^4(\mathbb{V}_{k-2})\to\otimes^{4}{{\mathfrak g}}_{-1}^*\otimes \mathbb{V}_{k-2}$ and it vanishes on the image of $\otimes^3{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}$ under the action of ${{\mathfrak g}}_1$. Hence we analyze first the space of ${{\mathfrak g}}_0$-homomorphisms $\widetilde{\Phi}: \otimes^4{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}\to\mathbb{V}_{k}$, and then we describe explicitly the image of the action of ${{\mathfrak g}}_1$ on $\bar{\mathcal{J}}^4(\mathbb{V}_{k-2})$ and we show that it never lies in $\operatorname{Ker}(\widetilde{\Phi})$ for any $\widetilde{\Phi}$ which lifts $\Phi$.
Step 1. An important observation is that any ${{\mathfrak g}}_0$-homomorphism $\otimes^4{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}\to\mathbb{V}_{k}$ factors through $\mathbb{V}_{2}\otimes\mathbb{V}_{k-2}$ and a unique projection $\mathbb{V}_{2}\otimes\mathbb{V}_{k-2}\to\mathbb{V}_{k}$. Therefore, we only need to analyze ${{\mathfrak g}}_0$-homomorphisms $\otimes^4{{\mathfrak g}}_{-1}^*\to\mathbb{V}_{2}$. This map is a complete contraction on the ${{\mathbb R}}^2$-part of ${{\mathfrak g}}_{-1}^*$ and thus a linear combination of $c_1({\omega})={\omega}_{I'}{}^{I'}{}_{J'}{}^{J'}$, $c_2({\omega})={\omega}_{I'J'}{}^{J'I'}$ and $c_3({\omega})={\omega}_{I'}{}^{J'I'}{}_{J'}$, which are related by $c_1+c_2+c_3=0.$ Similarly, on the ${{\mathbb R}}^{n*}$-part of ${{\mathfrak g}}_{-1}^*$ we denote the three projections $\otimes^4{{\mathbb R}}^{n*}\to{\mbox{$
\begin{picture}(13,10)(-1,1)
\put(0,10){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,0){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\end{picture}
$}}{{\mathbb R}}^{n*}$ as follows $$p_1({\omega})_{{{\bf A}}^2{{\bf B}}^2}=\tfrac12({\omega}_{A_1A_2B_1B_2}+{\omega}_{B_1B_2A_1A_2})
-{\omega}_{[A_1A_2B_1B_2]},$$ $$p_2({\omega})_{{{\bf A}}^2{{\bf B}}^2}=\tfrac12({\omega}_{A_1B_1B_2A_2}+{\omega}_{B_1A_1A_2B_2})
-{\omega}_{[A_1A_2B_1B_2]},$$ $$p_3({\omega})_{{{\bf A}}^2{{\bf B}}^2}=\tfrac12({\omega}_{A_1B_2A_2B_1}+{\omega}_{B_1A_2B_2A_1})
-{\omega}_{[A_1A_2B_1B_2]}.$$ They are obviously related by a similar equation $p_1+p_2+p_3=0.$ Hence the ${{\mathfrak g}}_0$-homomorphisms $\otimes^4{{\mathfrak g}}_{-1}^*\to\mathbb{V}_{2}$ form a vector space of dimension four. We choose $c_i\circ p_j$, $i,j=1,2$ as a basis and we set $\Phi_{ii}:=(c_i\circ p_i)\otimes{{\textup{id}}}$ and $\Phi_{ij}:=-2(c_i\circ p_j)\otimes{{\textup{id}}}$ for $i\neq j$. Then any ${{\mathfrak g}}_0$-homomorphism $\widetilde{\Phi}: \otimes^4{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}\to\mathbb{V}_{k}$ is given by a linear combination $K\Phi_{11}+L\Phi_{12}+M\Phi_{21}+N\Phi_{22}$, followed by the unique projection to $\mathbb{V}_{k}$. Moreover, the uniqueness in symmetric case implies that the restriction of each $\Phi_{ij}$ to $S^4{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}$ is a multiple of $\phi$, which is given by and which defines the holonomic map $\Phi$. Precisely, it it is easy to compute that upon the restriction to $\mathcal{J}^4(\mathbb{V}_{k-2})$ all maps $\Phi_{ij}$ coincide with $\phi$. This shows that $\widetilde{\Phi}$ covers $\Phi$ if and only if it is given by a linear combination of maps $\Phi_{ij}$ such that the coefficients satisfy $K+L+ M+N=1.$
Step 2. By [@CSS1 Lemma 5.10], the action of $Z\in {{\mathfrak g}}_1$ on $\psi\in\otimes^{3}{{\mathfrak g}}_{-1}^*\otimes \mathbb{V}_{k-2}$, regarded as a ${{\mathfrak g}}_0$-submodule of $\bar{\mathcal{J}}^4(\mathbb{V}_{k-2})$, is given by $$\label{Verma_action}
\begin{array}{l}
(Z\cdot\psi)(X_1,\cdots,X_4)=
\sum\limits_{1\leq i\leq 4} [Z,X_i]\bullet \psi(X_1,\cdots,\hat{X}_i,\cdots,X_{4})
\\
-\sum\limits_{1\leq i<j\leq 4}\psi(X_1,\cdots,[[Z,X_i],X_j]\hat{X}_j,\cdots, X_{4})
\end{array}$$ In terms of abstract indices with conventions from section \[GS\], the element $[Z,X_i]$ of ${{\mathfrak g}}_0$ is given by $[Z,X_i]=(Z^{A'}_I(X_i)^I_{B'},Z^{I'}_{A}(X_i)^{B}_{I'})$, and its action on $v^{A'}\in{{\mathbb R}}^2$, $v_A\in{{\mathbb R}}^{q*}$, $\sigma\in{{\mathbb R}}[w]$ and its adjoint action on $X_j\in{{\mathfrak g}}_{-1}$reads as follows $$\begin{array}{c}
[Z,X_i]\bullet v^{A'} = Z^{A'}_I(X_i)^I_{I'} v^{I'},
\\
[Z,X_i]\bullet v_A = Z^{I'}_{A}(X_i)^{I}_{I'} v_I,
\\
[Z,X_i]\bullet\sigma =wZ^{I'}_{I}(X_i)^{I}_{I'} \sigma,
\end{array}$$ and $$[[Z,X_i],X_j]^{A}_{A'}=-Z^{I'}_{I}(X_i)^{A}_{I'}(X_j)^{I}_{A'}-Z^{I'}_I(X_i)^I_{A'}(X_j)^A_{I'}.$$ Now we apply these equations to . It is easy to see that for $Z_{a_i}=Z^{A'_i}_{A_i}\in {{\mathfrak g}}_{-1}^*$ and $\psi_{a_1a_2a_3{{\bf E}}{{\bf F}}}\in\otimes^3{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}$ the terms $[Z,X_i]\bullet \psi$ appearing in the first sum are given by $$\begin{array}{l}
( [Z,X_i]\bullet \psi)_{a_1a_2a_3a_4{{\bf E}}{{\bf F}}}
=(k-2)Z^{A'_i}_{A_i}\psi_{a_1\cdots\hat{a}_i\cdots a_4 {{\bf E}}{{\bf F}}}
\\
+(k-2)Z^{A'_i}_{E_1}\psi_{a_1\cdots\hat{a}_i\cdots a_4 A_i{{\bf \dot{E}}}{{\bf F}}}
+(k-2)Z^{A'_i}_{F_1}\psi_{a_1\cdots\hat{a}_i\cdots a_4 {{\bf E}}A_i{{\bf \dot{F}}}},
\end{array}$$ Now it is an easy observation that the unprimed spinor indices displayed on the right-hand side of the previous formula lie in $${ \mbox{$\mbox{$\begin{picture}(21,13)(-1,-1)
\put(1,3){\mbox{\tiny ${k-2}$}}
\put(10,9){\vector(0,1){6}}
\put(10,1){\vector(0,-1){6}}
\end{picture}$}
\mbox{$\begin{picture}(19,13)(-1,-1)
\put(0,15){\line(1,0){15}}
\put(0,10){\line(1,0){15}}
\put(0,0){\line(1,0){10}}
\put(0,-5){\line(1,0){10}}
\put(0,-5){\line(0,1){20}}
\put(10,-5){\line(0,1){20}}
\put(15,10){\line(0,1){5}}
\put(5,-5){\line(0,1){7}}
\put(5,8){\line(0,1){7}}
\put(1,2){.}
\put(1,4){.}
\put(1,6){.}
\put(6,2){.}
\put(6,4){.}
\put(6,6){.}
\end{picture}$}
$}}{{\mathbb R}}^{n*}.$$ It means that each term $[Z,X_i]\bullet \psi$ vanishes under the projection to $\mathbb{V}_{k}$, and thus only the second sum in remains modulo Ker($\widetilde{\Phi}$). For $\psi={\omega}\otimes v$, where ${\omega}\in\otimes^3{{\mathfrak g}}_{-1}^*$ and $v\in\mathbb{V}_{k-2}
$, it reads as follows $$\begin{array}{l}
(Z\cdot({\omega}\otimes v))^{A'B'C'D'}_{ABCD}=(Z^{A'}_B{\omega}^{B'C'D'}_{ACD}+Z^{B'}_A{\omega}^{A'C'D'}_{BCD}
+Z^{A'}_C{\omega}^{B'C'D'}_{BAD}+Z^{C'}_A{\omega}^{B'A'D'}_{BCD}
\\
+Z^{A'}_D{\omega}^{B'C'D'}_{BCA}+Z^{D'}_A{\omega}^{B'C'A'}_{BCD}+Z^{B'}_C{\omega}^{A'C'D'}_{ABD}
+Z^{C'}_B{\omega}^{A'B'D'}_{ACD}+Z^{B'}_D{\omega}^{A'C'D'}_{ACB}
\\
+Z^{D'}_B{\omega}^{A'C'B'}_{ACD}+Z^{C'}_D{\omega}^{A'B'D'}_{ABC}+Z^{D'}_C{\omega}^{A'B'C'}_{ABD})\otimes v.
\end{array}$$ In order to show that for some ${\omega}$ it does not lie in Ker($\widetilde{\Phi}$) for any lift $\widetilde{\Phi}$, we express its image under each $\Phi_{ij}$. A straightforward computation yields $$\begin{array}{l}
c_1(Z\cdot{\omega})_{ABCD}=
-Z_{I'C}({\omega}^{I'J'}_{BADJ'}+{\omega}_{J'ABD}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}I'J'})
\\
+Z_{I'D}(-{\omega}^{I'J'}_{BCAJ'}+{\omega}^{I'J'}_{ACBJ'}+{\omega}_{J'ABC}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}J'I'})
\end{array}$$ and $$\begin{array}{l}
c_2(Z\cdot{\omega})_{ABCD}=
Z_{I'C}({\omega}_{J'BAD}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}J'I'}+{\omega}_{J'ABD}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}J'I'})
\\
+Z_{I'D}({\omega}_{J'BCA}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}J'I'}+{\omega}_{J'ACB}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}I'J'}-{\omega}_{J'ABC}^{{\phantom{'}}{\phantom{'}}{\phantom{'}}I'J'}).
\end{array}$$ Now it is easy to see that both contractions vanish provided that ${\omega}$ is symmetric. And since holonomic homomorphism $\Phi$ factors through a complete contraction, it also vanishes on the image of the action of ${{\mathfrak g}}_1$ on $S^3{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}$. This shows that $\Phi$ is indeed a $\mathfrak{p}$-homomorphism. On the other hand, the contractions $c_1$, $c_2$ for a nonsymmetric ${\omega}$ are nonzero in general. Namely, if we set $${\omega}^{A'B'C'}_{ABC}={{\epsilon}}^{A'B'}\bar{{\omega}}^{C'}_{ABC}\in\operatorname{Ker}({{\mathcal E}}^{C'}_{[AB]C}[-1]\to{{\mathcal E}}^{C'}_{[ABC]}[-1])\subset\otimes^3{{\mathfrak g}}_{-1}^*,$$ then they are equal to $$c_1(Z\cdot\bar{{\omega}})_{ABCD}=-3Z_{I'D}\bar{{\omega}}^{I'}_{ABC}$$ $$c_2(Z\cdot\bar{{\omega}})_{ABCD}=-3Z_{I'D}\bar{{\omega}}^{I'}_{BCA}.$$ Now it is easy to compute that the compositions with projections $p_1,p_2$ yield $$p_1\circ c_1(Z\cdot\bar{{\omega}})=-2 p_2\circ c_1(Z\cdot\bar{{\omega}})
=-2p_1\circ c_2(Z\cdot\bar{{\omega}})= p_2\circ c_2(Z\cdot\bar{{\omega}}).$$ This fact shows that for $\bar{\psi}:=\bar{{\omega}}\otimes v$ the element $Z\cdot\bar{\psi}$ has the same image under all maps $\Phi_{ij}$. Hence the image of $Z\cdot\bar{\psi}$ under a lift $\widetilde{\Phi}$ of $\Phi$, which is given by a linear combination of maps $\Phi_{ij}$ with coefficients $K,L,M,N$, is given by the projection of $$(K+ L+ M+N)\Phi_{11}(Z\cdot\bar{\psi})=\Phi_{11}(Z\cdot\bar{\psi})$$ to $\mathbb{V}_{k}$. In particular, it does not depend on the coefficients $K,L,M,N$ and it is not zero. Precisely, we get $$\widetilde{\Phi}(Z\cdot\bar{\psi})_{{{\bf A}}{{\bf B}}}=
\frac12(
Z_{I'B_2}\bar{\psi}^{I'}_{A_1A_2B_1{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
+Z_{I'A_2}\bar{\psi}^{I'}_{B_1B_2A_1{{\bf \ddot{A}}}{{\bf \ddot{B}}}}
)
\neq 0.$$
[A careful computation reveals that the maps $\Phi_{ij}$ coincide on whole image of $\otimes^3{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}$ in $\otimes^4{{\mathfrak g}}_{-1}^*\otimes\mathbb{V}_{k-2}$ under the action of ${{\mathfrak g}}_1$. Since the action of ${{\mathfrak g}}_1$ gives exactly the transformation of the four-fold covariant derivative which is algebraic and linear in the one-form $\Upsilon$ describing the change $\nabla\to\hat{\nabla}$, the coincidence of maps $\Phi_{ij}$ shows in turn that the projection of the algebraic linearized transformation of $\nabla^4f$ to $\mathbb{V}_{k}$ does not depend on the succession of covariant derivatives. Indeed, acting with $\square_{ABCD}$ on a tractor bundle, the leading terms with different orders of derivatives differ by terms of the form $\nabla^2\Omega$, $\nabla\Omega\nabla$ and $\Omega\nabla^2$ where $\Omega$ is the tractor curvature. And it is easy to prove that projections to the target bundle of all these terms are invariant. Hence any formula for $\square_{ABCD}$ will have the same linearized transformation of the leading part as the formula obtained from the curved Casimir construction. If we trace where we commuted derivatives back in the proof of proposition \[thm\], we conclude that this transformation (described by an one form $\Upsilon$) is given by the projection of $$\Upsilon\nabla\Omega+\Upsilon\Omega\nabla.$$ Moreover, it is easy to see that this cannot be cancelled by transformations of lower order terms and thus it is contained in the transformation of any formula for $\square_{ABCD}$. A straightforward computation shows that the action of this curvature expression on the standard tractor bundle vanishes if and only if the harmonic part of the Cartan curvature vanishes, i.e. in the flat case. Hence there exists no universal formula for $\square_{ABCD}$ which would define invariant operator between tractor bundles on manifolds with a general Grassmannian structure. ]{}
[99]{}
A. Čap and J. Slovák, Parabolic Geometries I: Background and General Theory, Math. Surv. and Monographs, [**154**]{}, Amer. Math. Soc., Providence, RI, 2009.
A. Čap, V. Souček, Curved Casimir operators and the BGG machinery, SIGMA Symmetry Integrability Geom. Methods Appl. **3** (2007) 111, 17 pp.
A. Čap, A.R. Gover and V. Souček, Conformally Invariant Operators via Curved Casimirs: Examples, Pure Appl. Math. Q., to appear, available at arXiv:0808.1978.
M. Hammerl, P. Somberg, V. Souček, J. Šilhan, Invariant prolongation of overdetermined PDE’s in projective, conformal and Grassmannian geometry, Annals of Global Analysis and Geometry, Springer, 2012, 0232-704X.
Gover A.R., Slov' ak J., [*Invariant local twistor calculus for quaternionic structures and related geometries,*]{} J. Geom. Phys. 32, No.1 (1999) 14-56
J. Lepowsky, A generalization of the Bernstein–Gelfand–Gelfand resolution, J. of Algebra 49 (1977) 496–511
B. Boe. Homomorphisms between generalized Verma modules. [*Trans. Amer. Math. Soc.*]{} 356 (1): 159-184, (2004).
Návrat A., Nonstandard invariant operators on quaternionic geometries, MSc Thesis, Masaryk University in Brno, 2004
Návrat A., Nonstandard operators in almost Grassmannian geometry, PhD Thesis, University of Vienna, 2012
Boe, Brian D.; Collingwood, David H. Multiplicity free categories of highest weight representations. I, II. Commun. Algebra 18, No.4, 947-1032, 1033-1070 (1990)
A. Čap, V. Souček, Subcomplexes in curved BGG sequences, Ann. Math. (2012) arXiv:math/0508534
Eastwood M., Slov' ak J., [*Semi-holonomic Verma modules*]{}, J. of Algebra, 197 (1997), 424-448
A. Čap, J. Slovák, V. Souček, Invariant operators on manifolds with almost Hermitian symmetric structures, I. Invariant differentiation, Acta Math. Univ. Commenianae, [**66**]{} (1997), 33–69, electronically available at www.emis.de
A. Čap, [*Translation of natural operators on manifolds with AHS-structures*]{}, Archivum Mathematicum (Brno), Tomus 32 (1996), 249-266.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In complementary images of coordinate-space and momentum-space density in a trapped 2D Bose gas, we observe the emergence of presuperfluid behavior. As phase-space density $\rho$ increases toward degenerate values, we observe a gradual divergence of the compressibility $\kappa$ from the value predicted by a bare-atom model, $\kappa_{ba}$. $\kappa/\kappa_{ba}$ grows to 1.7 before $\rho$ reaches the value for which we observe the sudden emergence of a spike at $p=0$ in momentum space. Momentum-space images are acquired by means of a 2D focusing technique. Our data represent the first observation of non-mean-field physics in the presuperfluid but degenerate 2D Bose gas.'
author:
- 'S. Tung, G. Lamporesi, D. Lobser, L. Xia, and E. A. Cornell[^1]'
title: 'Observation of the Presuperfluid Regime in a Two-Dimensional Bose Gas'
---
Because of the enhanced role of fluctuations in low-dimensional systems [@Mermin], a two-dimensional (2D) Bose gas at nonzero temperature does not have long-range phase coherence. In a homogeneous system there can be at best only a quasicondensate, no true Bose-Einstein condensation (BEC). Under the combined effect of interactions and quantum degeneracy, however, there is nonetheless a phase transition known as Berezinskii-Kosterlitz-Thouless (BKT) associated with the unbinding of vortex pairs [@BKT]. Below the critical temperature $T_{\mathrm{BKT}}$, the system is superfluid.
Experiments in 2D atomic gases [@Hadzibabic; @Kruger; @Rath; @Phillips] are usually conducted in the presence of an inhomogeneous trapping potential. In the complete absence of interactions, the confining potential can resurrect a traditional BEC [@Bagnato], but for realistic experimental parameters, interatomic interactions tend to suppress BEC by smoothing out the spatial profile [@Petrov; @Hadzibabic; @Hadzibabic2; @Kruger; @Phillips; @Rath] of the mean density to the point where the sample can be understood as a collection of locally uniform spatial regions, each of which is characterized by a particular local density and thus a particular local value of $T_{\mathrm{BKT}}$. Although these local regions may be too small to test in detail the coherence-related predictions of BKT theory, qualitative effects have been observed in experiment [@Hadzibabic; @Phillips].
Our particular interest is in the region just to the warm side of $T_{\mathrm{BKT}}$. In an earlier experiment on bosons trapped in a 2D optical lattice, we observed a proliferation of vortices as we warmed through the discrete-case equivalent of $T_{\mathrm{BKT}}$ [@Schweikhard]. But in that experiment a great many mesoscopic condensates were present, one at each lattice site, on both sides of the BKT transition, because they had condensed at a $T_{\mathrm{BEC}}$ distinct from and well above $T_{\mathrm{BKT}}$. For the continuous case, in contrast, there is no corresponding second transition temperature above $T_{\mathrm{BKT}}$. But if the cooling gas has by $T_{\mathrm{BKT}}$ already become a medium that can support vortices, whether bound or not, then heuristically we see that it must have continuously evolved from a fully fluctuating nondegenerate gas into a sort of presuperfluid with suppressed density fluctuations [@Hadzibabic3]. Theory [@Petrov; @Holzmann1; @Blakie; @Prokofev; @Holzmann2; @Hadzibabic3] validates this intuition, and experiments [@Rath] have in turn been consistent with predictions of that theory. Up until now, however, experiments have not been directly sensitive to the properties of the presuperfluid, $T \gtrsim T_{\mathrm{BKT}}$ gas. The goal of the present work is to provide a minimal-assumption, first empirical look at this exotic regime. We emphasize key features of this approach: 1. Our line of sight is along the axis of tight confinement: we do not need to do a deconvolution of our images to get the 2D density distribution. Steps are taken to minimize systematic errors in density measurements. 2. We analyze our *in situ* images to extract the local compressibility, a quantity directly sensitive to local microscopic physics. 3. We use a 2D focusing technique to record high-resolution 2D momentum-space images complementary to the coordinate-space images. We make corresponding inferences about nonlocal coherence. 4. We use a simple but robust “bare-atom" model to correct the observed density for the presence of a small population in excited states in the tight confinement direction, and to determine a bare-atom compressibility with which to compare our observations.
Experimentally, we create a stack of well-isolated quasi-2D layers by superimposing a one-dimensional, blue-detuned optical lattice with lattice spacing 3.8 $\mu m$ onto a magnetically trapped, evaporatively cooled cloud of Rb-87 atoms. Within each layer, approximately $6.2\times10^5$ atoms feel a harmonic potential characterized by frequencies $(\omega_{r},\omega_{z})$=2$\pi$(10,1400) Hz. The characteristic dimensionless 2D interaction strength is $\tilde{g}=\sqrt{8\pi}(a_s/a_{ho})=0.093$, where $a_s$ is the 3D scattering length and $a_{ho}$ is the $\hat{z}$ harmonic-oscillator length [@gtilde]. The atoms are allowed to equilibrate in their 2D geometry before probing occurs. Right before probing the resulting coordinate- or momentum-space distribution, we apply a microwave pulse together with a transient magnetic field to pump atoms in the central layer into another hyperfine state, resonant with the probe light. The strength of the microwave selection pulse is adjusted to keep the peak optical density of the imaged fraction within the linear dynamic range of our imaging.
Our momentum-space imaging makes use of a focusing technique which is an extension to 2D of a procedure developed for imaging 1D momentum distributions [@Walraven; @Druten]. It yields a much cleaner separation of momentum and coordinate-space distributions than is obtained in earlier experiments [@Phillips; @Kruger]. After selecting a single layer, we turn off the optical lattice and let the layer expand into a purely magnetic trap with frequencies $(\omega_{r},\omega_{z})$=2$\pi$(5.2,10.4) Hz. Because of the 140:1 aspect ratio of the cloud, the expansion is initially purely axial, very rapidly reducing the 3D density while not affecting the in-plane coordinate- or momentum-space distributions. After this near-instantaneous suppression of the repulsive atom-atom interactions, each atom undergoes free harmonic motion in the $x-y$ plane. After a dwell time $t=\frac{1}{4}\frac{2\pi}{\omega_r}$, just as the initial 2D momentum- and coordinate-space distributions have swapped, we take an absorption image \[Fig. 1(a)\]. Scaling the spatial coordinate by $m\omega_r$ yields the momentum-space distribution that existed just as the probe sequence began. We take azimuthal averages of the absorption images before plotting and fitting the data.
For coordinate-space images the procedure is similar, but after the lattice turn-off, we wait only 1 ms before taking the absorption image \[Fig. 1(b)\]. The 2D density remains essentially frozen while the 3D density – and related imaging artifacts [@Rath] – are much reduced.
![(a) A trap-focused, momentum-space image. (b) An in-situ, coordinate-space image. The corresponding azimuthal averages are shown in Figs. 2(g) and 2(c). []{data-label="fig1"}](fig1_v9.eps){width="45.00000%"}
To extract a signal for many-body physics from our data, we compare our data to a fully fluctuating, bare-atom model, in essence the Paris group’s mean-field, Hartree-Fock, local-density model [@Hadzibabic2]. The mean occupation of a single-atom state $k$ is given by the Bose-Einstein distribution, $N_{k}=\frac{1}{e^{(E_k-\mu)/k_BT}-1}$, where $E_k$ is the energy of the state. For our system, $k_BT \gg \hbar\omega_{r}$, but $k_BT
\sim \hbar\omega_{z}$. We treat the atomic motion semiclassically in the in-plane direction, while preserving discrete harmonic-oscillator levels in the $\hat{z}$ direction. The 2D coordinate-space density in the $j$th axial level is $$n_j(\vec{r})=\frac{1}{h^2}\iint \mathrm{d}^{2}\vec{p}
\frac{1}{e^{[\varepsilon(\vec{p})+\theta_j(\vec{r})-\mu_j(\vec{r})]/k_BT}-1},
\label{eq2}$$ where the free particle dispersion $\varepsilon(\vec{p})=\frac{p^2}{2m}$. The local chemical potential for the $j$th level is given by $$\begin{split}
\mu_j(\vec{r})& =\mu^{global}-\frac{1}{2}m\omega_r^2r^2- j \hbar\omega_z \\
&-\sum\limits_{l\neq j}2\left(\frac{4\pi\hbar^2}{m} a_s f_{jl}
n_l(\vec{r})\right), \label{eq3}
\end{split}$$ whereas the intralevel interaction energy is $$\label{eq4}
\theta_j(\vec{r})=2(4\pi\hbar^2a_s/m) f_{jj}n_{j}.$$ The relevant mean-field interaction energies depend on $f_{jl}$ which are the normalized density overlap integrals over the axial dimension between densities associated with axial quantum states $j$ and $l$. Interaction energies are comfortably less than the axial spacing $\hbar\omega_{z}$, justifying our treating the axial wave functions as frozen. We define a quantity $u_{00} \equiv
(4\pi\hbar^2a_s/m)f_{00}$, such that we can write $\theta_0(\vec{r}) = 2 u_{00} n_0 = 2 (\hbar^2\tilde{g}/m)
n_0(\vec{r})$. Evaluating the integral in Eq. , we get $$n_j(\vec{r})=-\ln(1-e^{-[\theta_j(\vec{r})-\mu_j(\vec{r})]/k_BT})/\lambda_{db}^2
\label{eq5}$$ where the de Broglie wavelength $\lambda_{db}
=\sqrt{2\pi\hbar^2/mk_BT}$. For any given value of $\mu^{global}$ and $T$, $n_j$ are determined self-consistently. For $k_BT
\lesssim \hbar\omega_z$, the model converges in just a few iterations.
The bare-atom model is a no-condensate model from which all the many-body effects associated with degenerate bosons has been intentionally omitted: the additional factor of 2 in front of the parentheses in Eq. and arises from an implicit assumption that the second-order correlation function at zero distance is 2, as it would be for fully fluctuating, nondegenerate ideal bosons, and not 1, as for a 3D BEC. Furthermore, $\varepsilon(\vec{p})=p^2/2m$ is the dispersion relation for independent atoms moving in a mean-field potential. There are no collective excitations such as phonons.
All the same, the bare-atom model should do very well where phase-space density $\rho_j \equiv n_j \lambda_{db}^2 < 1$, true for $j>0$ in our system. As for the calculated value of $n_0(\vec{r})$, this will begin to fail for $\rho_0 \gtrsim 3.5$, but a comparison observations with the naive, bare-atom $n_0$ will allow us to quantify the telltale discrepancy.
An analysis of a coordinate-space image proceeds as follows. We measure the integrated density in the $z$-direction with contributions from all populated excited axial levels. $n_{meas}(\vec{r}) = \sum_{j} n_j (\vec{r})$, but the interesting physics lies in $n_0(\vec{r})$. We compare the results of the bare-atom model to observed $n_{meas}$ and fit the parameters $T$ and $\mu^{global}$ to the low phase-space (hence well-understood) regions of the cloud. With the chemical potential and temperature obtained from the fit, we can use Eq. to evaluate the excited state distributions. Then we numerically find a self-consistent solution to get $n_{k}$ ($k > 0$), with the constraint $n_{0}=n_{meas}- \sum_{k>0} n_k$; see Figs. \[fig2\](a)–\[fig2\](d).
![(a)–(d) Coordinate-space distributions and (e)–(h) corresponding momentum distributions. Two distributions in the same row are taken under near-identical conditions. The thin black curves give the azimuthal averages of $n_{meas}$ from the raw images. The thick blue curves in the coordinate-space distributions are the ground-state distribution $n_0$ after correcting for $n_{k>0}$. The spike in momentum space that first appears in (g) has no corresponding dramatic change in coordinate space (c). The vertical dotted line in (c) represents the inverse of the momentum resolution limit indicated in (g) and is thus a lower limit on the spatial extent of the coherence of the population of low-p (high-coherence) atoms represented by the area (about 1.4 $\%$ of total) under the spike in (g).[]{data-label="fig2"}](fig4_v15.eps){width="45.00000%"}
Once the corrected ground-state distribution $n_{0}$ is extracted from $n_{meas}$, we calculate the scaled (by $n^2$) isothermal compressibility $\kappa$ at each imaging pixel $$\kappa =d n_0/d \mu_{0}=(d n_0/d r)(d \mu_0/d r)^{-1}. \label{eq6}$$ Although $\mu^{global}$ is not a quantity we can know with great accuracy in a model-independent way, $d\mu_0 / d r \approx -
d(\frac{1}{2}m\omega_r^2r^2)/dr = - m \omega_r^2 r$ is known quite precisely, as the contributions to $\mu_0$ arising from the mean field of axially excited atoms are small and correctable. $dn_0/dr$ is determined from our coordinate-space images. Equation then gives $\kappa$ at each discrete radius in an image. We plot the result vs the local phase-space density $\rho_0=n_0 \lambda_{db}^2$ in Fig. \[fig3\]. We compare $\kappa$ to the value $\kappa_{ba}$ that the bare-atom model would predict at the same density. For an observed value of $n_0$, we numerically solve the bare-atom prediction $n_0 = -\ln
[1-exp(\mu_0 - \theta_0 (n_0))/k_BT]/\lambda_{db}^2$ for $\mu_0$, and determine how $n_0$ changes for small changes in $\mu_0$, and thus extract $\kappa_{ba}$ (Fig. \[fig3\]).
As a test of the local-density approximation that is central to our analysis, we determine $\kappa$ using images from two very different classes of samples: clouds with $T$ = 171 nK and central $\rho_0$ of about 9 \[Fig. \[fig3\](a)\], and clouds with $T$ = 128 nK and central $\rho_0$ of about 30 \[Fig. \[fig3\](b)\]. The shape of $\kappa(\rho)$ is the same and, in particular, the value of $\rho_0$ for which the extracted value of $\kappa$ becomes distinguishable from $\kappa_{ba}$ is in both cases about 4.
A note on our preferred method of fixing of $\alpha$, the calibration scale factor that relates $n_{meas}(\vec{r})$ to the observed optical density profile: we considered (i) an error-prone calculation based on optical absorption cross section and (ii) a model-dependent (even unto logical circularity) fitting of $\alpha$ in the image analysis but settled finally on (iii) interleaving our data runs with auxiliary measurements of very low $T$ clouds in which the atoms are in a near-pure Thomas-Fermi inverted parabola with negligible noncondensed wings. In this limit, we assume $\mu^{global}$=$u_{00}n_0(0)$ and thus fix $\alpha$. This assumption means that our measured values of $\kappa$, scaled as in Fig. \[fig3\], are constrained to saturate to an average of 2.0 at very high $\rho_0$. In essence, we get accurate measurements of the sample density and temperature in the exotic, intermediate regime of $\rho_0$, by assuming prior good understanding of behavior in the experimentally well-characterized regimes of low degeneracy, mean-field at high $T$ and of high-coherence, pure condensates at low $T$.
![Scaled compressibility $\kappa$ vs phase-space density $\rho_0$. (a) Measured $\kappa$ extracted from images of samples at $T = 171$ nK, as in the image in Fig.2(c). Black circles are data averaged over the values calculated from images of three separate clouds. The blue curves are $\kappa_{ba}$ calculated from the bare-atom model. (b) Same but with $\kappa$ extracted from images of samples at $T = 128$ nK. The paired vertical dotted lines indicate the location of the jump in coherence discussed in the text. \[fig3\]](fig3_v14.eps){width="40.00000%"}
In Fig. \[fig2\], we present side-by-side pairs of coordinate-space and momentum-space distribution taken in a sequence such that pairs represent images of clouds with very similar temperatures and total atom number, such that the preimaging values of $\rho_0(0)$ are the same between pairs to within 10%. As $T$ decreases, there is no obvious sudden change in the coordinate-space distribution $n_0(\vec{r})$, \[Figs. \[fig2\](a)–(d)\], while in momentum space \[Figs. \[fig2\](e)–(h)\] a central spike suddenly emerges at $T$ =171 nK \[Fig. \[fig2\](g)\]. The inverse width of this central spike is a measure of the spatial extent of the coherent fraction in the highest-density region of the cloud. Our momentum-space resolution is such that the presence of a resolution-limited peak implies that at least some coherence extends over a central disk of radius 4.5 $\mu$m, or $\gtrsim 10
\lambda_{db}$. From coordinate-space images taken under the same conditions for which the coherence spike first appears in momentum space, we determine that it happens when the central value of $\rho_0$ = 8.0(0.7) (this critical value $\rho_c$ is determined from looking at many more pairs of images than are presented in Fig. \[fig2\]). We emphasize that from the coordinate-space distribution alone, the identification of a transition temperature would require model-dependent analysis of the smoothly varying distribution, while with access to both distributions at once, we readily see that a modest change in the central phase-space density of $<15\%$ causes the distribution at $p=0$ to jump by a factor of 3.
What have our observations told us about the 2D Bose gas as it cools towards the BKT transition? We can say empirically that as $\rho_0$ varies from about 7.2 to 8.7, we see a dramatic increase of coherence in a range $>$4.5 $\mu$m, jumping by a factor of 3. The transition may be even sharper, but temporal drifts limit resolution. The predicted [@Prokofev] critical value is $\rho_c=ln
(380/\tilde{g})=8.3$.
Our most interesting observation is that in warmer gases, for $\rho_0 \approx 4$, well before the sudden onset of coherence, we can already resolve that compressibility is above what a bare-atom model of degenerate bosons can account for. As $\rho_0$ reaches its coherence-jump value, 8.0, $\kappa/\kappa_{ba}$ has already increased to 1.7 (Fig. \[fig3\]). It is natural to associate the increase in $\kappa$/$\kappa_{ba}$ with the gradual changing of the interaction energy from its value in a fully fluctuating gas, $2
u_{00} n_0$, to its fully condensed value, $u_{00} n_0$, and to draw a corresponding inference about the zero-range second-order correlation function. For $\tilde{g}=0.093$ the corresponding theoretical prediction (from Ref. [@Prokofev] and Sec. 3.3 of Ref. [@Hadzibabic3]) for $\kappa$/$\kappa_{ba}$ at $\rho_c$ would be 1.59, in reasonable agreement with our observed 1.7. As a caveat, our observation of an anomalous $\kappa/\kappa_{ba}$ establishes definitively only the breakdown of the bare-atom model, which could be partially due to the violation of the other key bare-atom assumption, that in-plane excitations correspond to individual atoms with kinetic energy $\epsilon(\vec{p}) = p^2/2m$. In any case, our data provide a first definitive observation of non-mea-nfield physics in the presuperfluid 2D Bose gas.
We are very pleased to acknowledge useful conversations with Z. Hadzibabic, J. Dalibard, W. Phillips, M. Holzmann, C. Chin, A. Imambekov, L.-K. Lim and V. Gurarie. This work was supported by NSF and ONR.
[*Note added.*]{}—Recently, a relevant experimental preprint appeared [@Hung].
[19]{}
N. D. Mermin and H. Wagner, Phys. Rev. Lett. [**17,**]{} 1133 (1966).
J. M. Kosterlitz and D. J. Thouless, J. Phys. C [**6**]{}, 1181 (1973); J. M. Kosterlitz, J. Phys. C [**7**]{}, 1046 (1974); V. L. Berezinskii, Sov. Phys. JETP [**32**]{} 493 (1971); [**34**]{} 610 (1972).
Z. Hadzibabic, P. Krüger, M. Cheneau, B. Battelier, and J. Dalibard, Nature (London) [**441**]{}, 1118 (2006).
P. Krüger, Z. Hadzibabic, and J. Dalibard, Phys. Rev. Lett. [**99**]{}, 040402 (2007).
S. P. Rath *et al.*, Phys. Rev. A [**82**]{}, 013609 (2010).
P. Cladé, C. Ryu, A. Ramanathan, K. Helmerson, and W. D. Phillips, Phys. Rev. Lett. [**102**]{}, 170401 (2009).
V. Bagnato and D. Kleppner, Phys. Rev. A [**44,**]{} 7439 (1991).
D. S. Petrov, M. Holzmann, and G. V. Shlyapnikov, Phys. Rev. Lett. [**84**]{}, 2551 (2000).
Z. Hadzibabic, P. Krüger, M. Cheneau, S. P. Rath, and J. Dalibard, New J. of Phys. [**10**]{}, 045006 (2008).
V. Schweikhard, S. Tung, and E. A. Cornell, Phys. Rev. Lett. [**99**]{}, 030401 (2007).
Z.Hadzibabic and J. Daibard, arXiv:0912.1490v1.
M. Holzmann, G. Baym, J.-P. Blaizot, and F. Lalo$\ddot{e}$, Proc. Natl. Acad. Sci. U.S.A. [**104**]{}, 1476 (2007).
R. N. Bisset, M. J. Davis, T. P. Simula, and P. B. Blakie, Phys. Rev. A [**79**]{}, 033626 (2009).
N. Prokof’ev, O. Ruebenacker, and B. Svistunov, Phys. Rev. Lett. [**87**]{}, 270402 (2001); N. Prokof’ev and B. Svistunov, Phys. Rev. A [**66**]{}, 043608 (2002).
M. Holzmann and W. Krauth, Phys. Rev. Lett. [**100**]{}, 190402 (2008).
c.f. $\tilde{g}$ = 0.13 [@Kruger] and 0.146 [@Rath] in Paris group. $\tilde{g}$ = 0.02 in Gaithersburg group [@Phillips].
I. Shvarchuck *et al.*, Phys. Rev. Lett. [**89**]{}, 270404 (2002).
A. H. van Amerongen *et al.*, Phys. Rev. Lett. [**100**]{}, 090402 (2008).
C.-L. Hung, X. Zhang, N. Gemelke, and C. Chin, arXiv:1009.0016.
[^1]: NIST
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce and analyze a model for the transport of particles or energy in extended lattice systems. The dynamics of the model acts on a discrete phase space at discrete times but has nonetheless some of the characteristic properties of Hamiltonian dynamics in a confined phase space : it is deterministic, periodic, reversible and conservative. Randomness enters the model as a way to model ignorance about initial conditions and interactions between the components of the system. The orbits of the particles are non-intersecting random loops. We prove, by a weak law of large number, the validity of a diffusion equation for the macroscopic observables of interest for times that are arbitrary large, but small compared to the minimal recurrence time of the dynamics.'
address: 'Laboratoire de Probabilités et Modèles Aléatoires (CNRS UMR 7599), Université Paris 7 – Denis Diderot, UFR Mathématiques, Case 7012, Bâtiment Chevaleret, 75205 Paris Cedex 13, France'
author:
- Raphaël Lefevere
title: 'Macroscopic diffusion from a Hamilton-like dynamics'
---
Fick’s law of diffusion or Fourier’s law of heat conduction describe phenomena which are part of everyday life : think of the diffusion of sugar in a cup of coffee or the exponential decay in time of the temperature of the same hot coffee. The laws of microscopic physics possess features that makes them look contradictory at first sight with the phenomenological laws of macroscopic physics. In particular, microscopic laws are reversible and when the dynamics is confined in phase space, the Poincaré recurrence theorem ensures quasi-periodicity of the orbits. This a well-known problem which, in various guises, has generated a large number of debates among physicists and mathematicians alike, see for instance [@Bricmont; @Lebowitz] for excellent reviews. Boltzmann made a decisive contribution to the issue by insisting on the fact that in large systems, the usual laws of macroscopic physics correspond only to a [*typical*]{} (with respect to the initial conditions) behaviour and not a uniform one. One should note that in general the argument can not rest on typicality of the initial conditions alone. There exist quite a few systems for which macroscopic laws of normal diffusion are not obeyed : Fourier’s law is not observed in non-interacting gases, lattices of harmonic oscillators or the Toda lattice. Still, given its pervasiveness, one can not expect “normal" macroscopic behaviour to depend on delicate microscopic dynamical properties. On the contrary, in the absence of any information on microscopic initial conditions or interactions reflected in a large-scale structure, Fick’s law or Fourier’s law seems to be the typical behaviour in Nature. Therefore, Boltzmann’s solution to the problem should rather refer to typical initial conditions [*and*]{} typical dynamics.
We propose an instructive toy model for Fick’s law or Fourier’s law observed in systems which have a lattice spatial structure at the microscopic level . This model is deterministic, reversible, periodic, conservative and amenable to a full rigorous treatment. Randomness enters the model as a way to model ignorance about microscopic initial conditions and microscopic interactions between the components of the system. With large probability with respect to this randomness, as the number of components increase, one can show that a Boltzmann equation holds and is equivalent to a macroscopic diffusion. The key point is that the minimal recurrence time of the dynamics increases as the number of components does. This prevents any “recollision" occurring for times smaller than the minimal recurrence time. The model is reminiscent of the Kac ring model [@Kac] but possesses a rich orbit structure that makes it more similar to Hamiltonian dynamics. The orbits of the particles are non-intersecting random loops.
The dynamics and its first properties
=====================================
Le us consider a finite interval in ${{\ensuremath{\mathbb Z}} }$ : $$\Lambda_N=\{i\in{{\ensuremath{\mathbb Z}} }:|i|\leq N\}.$$ To each site of $i\in\Lambda_N$, we attach a ring ${{\ensuremath{\mathcal R}} }_i$ carrying $R$ sites $k\in\{1,\ldots,R\}$. The model consists of particles moving on $${{\ensuremath{\mathcal C}} }_N=\prod_{i\in\Lambda_N}{{\ensuremath{\mathcal R}} }_i=\{(k,i):k\in\{1,\ldots,R\},\;i\in\Lambda_N\}.$$
The second ingredient of the model is the presence of “scatterers" that are located in between pairs of sites $(k,i)$ and $(k,i+1)$. We define variables $\xi(k,i)$ taking values in $\{0,1\}$ and $\xi(k,i)=1$ if and only if there is a scatterer between sites $(k,i)$ and $(k,i+1)$.
We put particles on the sites of ${{\ensuremath{\mathcal C}} }_N$ and a site carries at most one particle. At fixed discrete times, all particles move forward on the rings. A particle located at site $(k,i)$, namely at site $k$ on the ring ${{\ensuremath{\mathcal R}} }_i$ will jump to site $k+1$ on ring ${{\ensuremath{\mathcal R}} }_{i+1}$ (resp. ${{\ensuremath{\mathcal R}} }_{i-1}$), if and only if the following conditions are simultaneously satisfied.
1. There is a scatterer between $(k,i)$ and $(k,i+ 1)$ (resp. $(k,i- 1)$), namely $\xi(k,i)=1$ (resp. $\xi(k,i-1)=1$).
2. There are no other scatterers around that pair.
In every other case, the particle located at site $(k,i)$ simply moves forward on its own ring to $(k+1,i)$. Formally, we define the map $\tau:{{\ensuremath{\mathcal C}} }_N\to{{\ensuremath{\mathcal C}} }_N$ giving the one-step evolution of a particle located at $(k,i)\in{{\ensuremath{\mathcal C}} }_N$ : $$\begin{aligned}
\tau(k,i)&=&J(k,i)(k+1,i+1)+J(k,i-1)(k+1,i-1)\nonumber\\
&+&(1-J(k,i))(1-J(k,i-1))(k+1,i)
\label{taudef}\end{aligned}$$ where $$J(k,i)=\xi(k,i)(1-\xi(k,i-1))(1-\xi(k,i+1))
\label{defJ}$$ and we set $$\xi(k,-N-1)=\xi(k,N)=0 .$$ whenever they appear in the above expressions. Here and in the following addition and substraction on the first component of a point in ${{\ensuremath{\mathcal C}} }_N$ are understood to be modulo $R$. The first factor in (\[defJ\]) accounts for the presence or absence of a scatterer, the two others are there to guarantee that $\tau$ is an injective map and has thus orbits similar to Hamiltonian dynamics.
To each $(k,i)\in{{\ensuremath{\mathcal C}} }_N$, one can attach an occupation variable $\sigma(k,i)\in\{0,1\}$. The map $\tau:{{\ensuremath{\mathcal C}} }_N\to{{\ensuremath{\mathcal C}} }_N$ allows to define the evolution of those variables by the relation : $$\sigma(k,i;t)=\sigma(\tau^{-t}(k,i);0),\; t\in{{\ensuremath{\mathbb N}} }^*.
\label{ocup}$$ This is equivalent to the recursion relation : $$\begin{aligned}
\sigma(k,i;t)&=&(1-J(k-1,i))(1-J(k-1,i-1))\sigma(k-1,i;t-1)\nonumber\\
&+&J(k-1,i-1)\sigma(k-1,i-1;t-1)+J(k-1,i)\sigma(k-1,i+1;t-1).\nonumber\\
\label{recurevol}\end{aligned}$$ We note that at any time $t\in{{\ensuremath{\mathbb N}} }^*$, the configuration $\sigma(\cdot,t)$ is obtained as a permutation of the initial occupation variables $\sigma(\cdot;0)$. Thus the dynamics is [*conservative*]{}.
$\tau$ is a well-defined bijective map of ${{\ensuremath{\mathcal C}} }_N$ into ${{\ensuremath{\mathcal C}} }_N$. $\tau$ is therefore invertible. \[taumap\]
We first note that for any configuration $\xi$ of scatterers $J(k,i)\in\{0,1\}$. Next only one of the three terms in (\[taudef\]) is non-zero. Indeed, assume that $J(k,i)=1$ then the third term is obviously zero but so is also the second one because one must have $\xi(k,i)=1$ and thus $J(k,i-1) =\xi(k-1,i)(1-\xi(k,i-2))(1-\xi(k,i))=0$. Other cases are treated similarly. Thus, for any $\xi$, $\tau(k,i)\in\{(k+1,i+1),(k+1,i-1),(k+1,i)\}\subset{{\ensuremath{\mathcal C}} }_N$ and $\tau$ is well-defined. We prove now that $\tau$ is injective. Let us assume that we have $x$ and $x'$ such that $\tau(x)=\tau(x')=(k,i)$, then both $x$ and $x'$ must belong to $\{(k-1,i),(k-1,i-1),(k-1,i+1)\}$. Suppose that $x=(k-1,i)$, then because of the definition (\[taudef\]), we must have $J(k-1,i)=J(k-1,i-1)=0$. This implies that $\tau(k-1,i-1)\neq (k,i)$ and $\tau(k-1,i+1)\neq (k,i)$ and thus that $x=x'=(k-1,i)$. Now, if $x=(k-1,i-1)$ and $\tau(x)=(k,i)$, then $J(k-1,i-1)=1$ and $J(k-1,i)=0$. Thus $\xi(k-1,i-1)=1$ and $\xi(k-1,i)=0$ and therefore $\tau(k-1,i)=(k,i-1)$ and $\tau(k-1,i+1)\neq(k,i)$. Thus, in that case also, if $x'$ is such that $\tau(x')=\tau(x)$ then $x'=x$. The last possible case (when $x=(k-1,i+1)$) is treated in a similar way. We have thus proven that $\tau:{{\ensuremath{\mathcal C}} }_N\to{{\ensuremath{\mathcal C}} }_N$ is an injective map. It is obvious that $\tau$ is onto, we have thus shown that $\tau$ is a bijection.
One could add a degree of freedom (“velocity") to the model by deciding that all particles move either in the positive (as above) or negative direction on the rings. We would then have two maps $\tau_+$ and $\tau_-$ with $\tau_+=\tau$ and obviously $\tau_-=\tau^{-1}$. The dynamics is thus [*reversible*]{} in a analogous sense to Hamilton dynamics.
To each $x\in {{\ensuremath{\mathcal C}} }_N$, we associate an orbit $$O(x)=\{z\in{{\ensuremath{\mathcal C}} }_N: \exists n\in{{\ensuremath{\mathbb N}} }\,\;\tau^n(x)=z\}$$ and a period $$T(x)=\inf\{n>0:\tau^n(x)=x\}.$$ Since the map $\tau$ is injective, the orbits of two different sites are either identical or do not intersect. Moreover, any orbit is self-avoiding.
Every point of ${{\ensuremath{\mathcal C}} }_N$ is periodic (i.e. has a finite period) and we have, $$R\leq T(x)\leq R(2N+1),\;\;\forall x\in{{\ensuremath{\mathcal C}} }_N.$$
Since the cardinality of ${{\ensuremath{\mathcal C}} }_N$ is finite, the fact that $\tau$ is injective implies that every orbit is periodic. This is actually a special case of the Poincaré recurrence theorem and we do not repeat the proof. An upper bound on the period is given by the number of sites contained in ${{\ensuremath{\mathcal C}} }_N$. A lower bound is obtained by observing that at each step one moves one step ahead on the rings. Thus, before coming back to the same site one must have performed at least one full rotation and this gives the lower bound on the period.
We sum up some obvious but crucial facts in the following lemma, for further reference.
1. $\tau^{-t}(k,i)$ depends on the configuration of scatterers only through the set of variables $\{\xi(k,i):\xi(k-n,i),\; n=1,\ldots, t,\;i=-N,\ldots,N\}$.
2. $\sigma(k,i;t)$ is a function of $\{\xi(k,i):\xi(k-n,i),\; n=1,\ldots, t,\;i=-N,\ldots,N\}$ and $\sigma(\cdot,0)$.
\[dependance\]
To summarize : for [*every*]{} configuration of scatterers, the dynamics has all the characteristic of a Hamiltonian dynamics in a confined domain : it is deterministic, reversible, (quasi-) periodic and conservative.
Randomness and diffusion equation
=================================
We let randomness enter the dynamics by taking a random configuration of scatterers and random initial occupation variables. Then $\tau:{{\ensuremath{\mathcal C}} }_N\to{{\ensuremath{\mathcal C}} }_N$ becomes a random map and (\[ocup\]) implies that the configuration $\sigma(\cdot,t)$ at any time is simply a random permutation of a collection of random variables. Likewise, orbits become non-intersecting random loops on ${{\ensuremath{\mathcal C}} }_N$. Scatterers stand for interactions between individual components, namely the rings. In the absence of any macroscopic information on those interactions except its average “density", the most natural choice is to take the scatterers $\xi(k,i)$ to be a set of independent Bernoulli variables of parameter $0<\mu<1$. Similar considerations lead to take the initial configuration of occupation variables as a set of independent Bernoulli random variables but with a ring-dependent parameter to account for an initial inhomogeneous distribution of density of particles. We define the empirical density of particles on the rings ${{\ensuremath{\mathcal R}} }_i$ which are our main “macroscopic" quantities : $$\rho^R(i,t):=\frac 1 R\sum_{k=1}^{R}\sigma(k,i;t).$$ Our goal is to show that for times $t$ smaller than the size $R$ of the rings, this empirical density follows the solution of a diffusion equation, with large probability as $R$ grows to infinity. As we indicate in the last section, it is easy to exhibit configurations of scatterers that would lead to an anomalous behaviour.
Let $\hat \rho(i,t)$ be a solution of the system of discrete-time evolution equations : $$\left\{\begin{array}{lll}
&\hat\rho(i,t)-\hat\rho(i,t-1)=\mu(1-\mu)^2\left(\hat\rho(i-1,t-1)+\hat\rho(i+1,t-1)-2\hat\rho(i,t-1)\right),\; |i|<N\\
&\hat\rho(-N,t)-\hat\rho(-N,t-1)=\mu(1-\mu)(\hat\rho(-N+1,t-1)-\hat\rho(-N,t-1))\\
&\hat\rho(N,t)-\hat\rho(N,t-1)=\mu(1-\mu)(\hat\rho(N-1,t-1)-\hat\rho(N,t-1))
\end{array}\right.
\label{diffusion}$$ This is a discrete laplace equation in discrete time. Given initial data, the solution is unique for any positive integer $t$. We state now our main result.
Let the $\{\sigma(k,i,0): (k,i)\in {{\ensuremath{\mathcal C}} }_N\}$ be a set of independent Bernoulli random variables and $\{\hat\rho_i:0<\hat \rho_i<1, \; i\in\Lambda_N\}$ such that ${{\ensuremath{\mathbb E}} }(\sigma(k,i,0))=\hat\rho_i$, $\forall k \in \{1,\ldots,R\}$. Let also $\hat\rho(\cdot,t)$ be the solution of the above system with initial condition $\hat\rho(i,0)=\hat\rho_i$, then $\forall \epsilon>0$ and $\forall 0<\alpha<1$, $$\lim_{R\to\infty}\sup_{t\in[0,R^\alpha]}{{\ensuremath{\mathbb P}} }\left[\bigcup_{i=-N}^N\{|\rho^R(i,t)-\hat\rho(i,t)|>\epsilon\}\right]=0.$$
By sub-additivity, it’s enough to show : $$\lim_{R\to\infty}\sup_{t\in[0,R^\alpha]}{{\ensuremath{\mathbb P}} }\left[\{|\rho^R(i,t)-\hat\rho(i,t)|>\epsilon\}\right]=0, \;i\in\Lambda_N.
\label{sub}$$ First, we show that $${{\ensuremath{\mathbb E}} }[\rho^R(i,t)]=\hat\rho(i,t),\;i\in\Lambda_N,\; 0<t<R^\alpha.
\label{average}$$ We consider the cases $i\neq -N,N$, it is easy to see that the cases $i=-N,N$ are similar. Using (\[recurevol\]) and noting that : $$J(k-1,i)J(k-1,i-1)=0$$ and $${{\ensuremath{\mathbb E}} }[J(k-1,i)]={{\ensuremath{\mathbb E}} }[J(k-1,i-1)]=\mu(1-\mu)^2,\;\forall\; 1\leq k\leq R,$$ we compute from (\[recurevol\]) and (\[defJ\]): $${{\ensuremath{\mathbb E}} }[\rho^R(i,t)]-{{\ensuremath{\mathbb E}} }[\rho^R(i,t-1)]=\mu(1-\mu)^2\left({{\ensuremath{\mathbb E}} }[\rho^R(i-1,t-1)]+{{\ensuremath{\mathbb E}} }[\rho^R(i+1,t-1)-2{{\ensuremath{\mathbb E}} }[\rho^R(i,t-1)]\right).$$ We have used $t<R^\alpha<R$ so that $\sigma(k,i;t)$ and $\xi(k,j)$ are independent for all $k,i,j$ by Lemma (\[dependance\]). Since ${{\ensuremath{\mathbb E}} }[\rho^R(i,0)]=\hat\rho_i$ and because the solution to (\[diffusion\]) is unique, the relation (\[average\]) is proven. We now look at the variance of $\rho^R(i,t)$. $$\begin{aligned}
{\rm Var}[\rho^R(i,t)]&=&\frac 1 {R^2}{{\ensuremath{\mathbb E}} }[\left(\sum_{k=1}^R\sigma(k,i;t)-\sum_{k=1}^R{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)]\right)^2]\nonumber\\
&=&\frac 1{R^2}\left({{\ensuremath{\mathbb E}} }[\sum_{k,k'=1}^R\sigma(k,i;t)\sigma(k',i;t)]-(\sum_{k=1}^R{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)])^2\right)
\label{variance}\end{aligned}$$ We can write for the second term : $$\begin{aligned}
{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)]&=&\sum_{x\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(\tau^{-t}(k,i);0)|\tau^{-t}(k,i)=x]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x]\nonumber\\
&=&\sum_{x\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x].\end{aligned}$$ Proceeding in the same way for the first one we obtain, $$\begin{aligned}
{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)\sigma(k',i;t)]&=&\sum_{x,x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0) \sigma(x';0)] {{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x,\tau^{-t}(k',i)=x'].\nonumber\\\end{aligned}$$ When $k\neq k'$, we get : $$\begin{aligned}
{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)\sigma(k',i;t)]&=&\sum_{x\neq x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb E}} }[\sigma(x';0)] {{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x,\tau^{-t}(k',i]=x')\nonumber\\\end{aligned}$$ because if $k\neq k'$, then $\tau^{-t}(k,i)\neq\tau^{-t}(k',i)$ for any configuration of scatterers, since $\tau$ (and $\tau^{-t}$) is a bijection. The factorization of the expectation is obtained because the initial occupation variables are independent. Going back to (\[variance\]), we take the first expectation : $$\begin{aligned}
{{\ensuremath{\mathbb E}} }[\sum_{k,k'=1}^R\sigma(k,i;t)\sigma(k',i;t)]&=&\sum_{k=1}^R\sum_{x\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma^2(x;0)]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x]\nonumber\\
&+& \sum_{k\neq k'\in{{\ensuremath{\mathcal C}} }_N}\sum_{x\neq x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb E}} }[ \sigma(x';0)] {{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x,\tau^{-t}(k',i)=x']\nonumber\\
\label{firstterm}\end{aligned}$$ Since ${{\ensuremath{\mathbb E}} }[\sigma^2(x;0)]\leq 1$, the first term above may be bounded by $R$ and we get : $$\sum_{k,k'=1}^R{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)\sigma(k',i;t)]\leq R+\sum_{k\neq k'}\sum_{x\neq x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb E}} }[ \sigma(x';0)] {{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x,\tau^{-t}(k',i)=x'].
\label{firsttermbound}$$ For the second term of (\[variance\]), we have : $$\begin{aligned}
(\sum_{k=1}^R{{\ensuremath{\mathbb E}} }[\sigma(k,i;t)])^2
&\geq&\sum_{k\neq k'}\sum_{x,x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb E}} }[ \sigma(x';0)] {{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k',i)=x']\nonumber\\
\label{secondtermbound}\end{aligned}$$ Combining (\[firsttermbound\]) and (\[secondtermbound\]), we get for the variance (\[variance\]) $$\begin{aligned}
{\rm Var}[\rho^R(i,t)]\leq\frac 1 R+&&\frac 1{R^2}|\sum_{k\neq k'}\sum_{x,x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb E}} }[ \sigma(x';0)]\cdot\nonumber\\
&&\left( {{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x,\tau^{-t}(k',i)=x']-{{\ensuremath{\mathbb P}} }[\tau^{-t}(k,i)=x]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k',i)=x']\right)|\nonumber\\\end{aligned}$$ Using rotational invariance of both the distribution of the scatterers and the distribution of initial occupation variables, we get
$$\begin{aligned}
{\rm Var}[\rho^R(i,t)]\leq\frac 1 R+&&\frac 1{R}|\sum_{ k'\neq 1}\sum_{x,x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb E}} }[\sigma(x;0)]{{\ensuremath{\mathbb E}} }[ \sigma(x';0)]\cdot\nonumber\\
&&\left( {{\ensuremath{\mathbb P}} }[\tau^{-t}(1,i)=x,\tau^{-t}(k',i)=x']-{{\ensuremath{\mathbb P}} }[\tau^{-t}(1,i)=x]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k',i)=x']\right)|.\nonumber\\\end{aligned}$$
Now, by Lemma (\[dependance\]), if $t+1<k'\leq R-t+1$ then $\tau^{-t}(0,i)$ and $\tau^{-t}(k',i)$ are independent random variables and for those $k'$, we have : $${{\ensuremath{\mathbb P}} }[\tau^{-t}(1,i)=x,\tau^{-t}(k',i)=x']-{{\ensuremath{\mathbb P}} }[\tau^{-t}(1,i)=x]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k',i)=x']=0.
\label{recollisions}$$
We are thus left with : $$\begin{aligned}
{\rm Var}[\rho^R(i,t)]\leq&&\frac 1 R+\frac 1{R}\sum_{\begin{array}{ll} {\scriptstyle R-t+1< k'\leq R}\\{\scriptstyle 1<k'\leq t+1}\end{array} }\sum_{x,x'\in{{\ensuremath{\mathcal C}} }_N} {{\ensuremath{\mathbb P}} }[\tau^{-t}(1,i)=x,\tau^{-t}(k',i)=x']\nonumber\\
&+&\frac 1 R\sum_{\begin{array}{ll} {\scriptstyle R-t+1< k'\leq R}\\{\scriptstyle 1<k'\leq t+1}\end{array} }\sum_{x,x'\in{{\ensuremath{\mathcal C}} }_N}{{\ensuremath{\mathbb P}} }[\tau^{-t}(1,i)=x]{{\ensuremath{\mathbb P}} }[\tau^{-t}(k',i)=x']\nonumber\\
\leq && \frac 1 R+\frac {4(t-1)}{R}\nonumber\\
\leq&&\frac 6 {R^{1-\alpha}}, {\rm for } \;R\; {\rm large \; enough}.\end{aligned}$$ In the first inequality, we have used independence and ${{\ensuremath{\mathbb E}} }[\sigma(x;0)]\leq 1$, in the second one the fact that probabilities sum up to $1$ and the fact that there are $2t-2$ terms in each sum over $k'$. Finally, we have used the hypothesis that $t<R^\alpha$. We conclude by using Chebyshev’s inequality and get : $${{\ensuremath{\mathbb P}} }\left[\{|\rho^R(i,t)-\hat\rho(i,t)|>\epsilon\}\right]\leq \frac{6}{\epsilon^2 R^{1-\alpha}}.
\label{final}$$ This in turn, implies (\[sub\]).
Relation to Boltzmann equation
==============================
By using a “molecular chaos" hypothesis, one can derive heuristically a Boltzmann equation for the evolution of the proportion of occupied sites on ring $\rho^R(i,t)$. This equation coincides with the diffusion equation (\[diffusion\]). We start by an exact relation for the evolution of the number of particles contained in ${{\ensuremath{\mathcal R}} }_i$, $|i|< N$. $$\begin{aligned}
R\rho^R(i,t+1)-R\rho^R(i,t)&=&X_{(i+1)\to i}(t)-X_{i\to (i+1)}(t)\nonumber\\
&+&X_{(i-1)\to i}(t)-X_{i\to( i-1)}(t)\nonumber\\
\label{xdef}\end{aligned}$$ $X_{i\to (i+1)}(t)$ counts the number of pairs of sites on rings ${{\ensuremath{\mathcal R}} }_i$ and ${{\ensuremath{\mathcal R}} }_{i+1}$ for which, at time $t$, a particle will jump from ring ${{\ensuremath{\mathcal R}} }_i$ to ${{\ensuremath{\mathcal R}} }_{i+1}$ and no particle jump from ${{\ensuremath{\mathcal R}} }_{i+1}$ to ${{\ensuremath{\mathcal R}} }_{i}$. Namely, $$X_{i\to i+1}(t)=|\{k\in\{1,\ldots,R\}: \sigma(k,i,t)=1,\;\sigma(k,i+1,t)=0, J(k,i)=1\}|.$$ We define also $$\hat X_{i\to i+1}(t)=|\{k\in\{1,\ldots,R\}: \sigma(k,i,t)=1,\;\sigma(k,i+1,t)=0\}|.$$
All other $X$’s appearing in (\[xdef\]) are defined in an analogous way. In this context, the molecular chaos hypothesis amounts to assume :
1. The proportion $\hat X_{i\to i+1}(t)/R$ is independent of the scatterer distribution between the rings and therefore : $$\begin{aligned}
X_{i\to i+1}(t)/R&\simeq&\frac 1 R\sum_{k=1}^RJ(k,i)\; \hat X_{i\to i+1}(t)/R\nonumber\\
&\simeq&{{\ensuremath{\mathbb E}} }[J(k,i)]\; \hat X_{i\to i+1}(t)/R\nonumber\\
&\simeq&\mu(1-\mu)^2\hat X_{i\to i+1}(t)/R\end{aligned}$$ as $R\to\infty$.
2. The proportion $\hat X_{i\to i+1}(t)/R$ is given by the product of the proportions of occupied sites on ${{\ensuremath{\mathcal R}} }_i$ and vacant sites on ${{\ensuremath{\mathcal R}} }_{i+1}$ : $$\hat X_{i\to i+1}(t)/R=\rho^R(i,t)(1-\rho^R(i+1,t))$$
The first assumption is similar to the one made in the Kac model. The second one is more similar to a “genuine" Boltzmann assumption on the factorization of distribution of pairs of particles. One is thus led to write : $$X_{i\to (i+1)}(t)/R=\mu(1-\mu)^2\rho^R(i,t)(1-\rho^R(i+1,t)),\; i<N.$$ A simple computation leads then to the evolution equation for $|i|<N$. $$\rho^R(i,t)-\rho^R(i,t-1)=\mu(1-\mu)^2\left(\rho^R(i-1,t-1)+\rho^R(i+1,t-1)-2\rho^R(i,t-1)\right).$$ This is identical to (\[diffusion\]) for $|i|<N$, equations for the evolution of the boundary densities are derived in a similar way.
Conclusions
===========
We observe that taking a diffusive scaling limit (when $N\to\infty$) of the discrete-time discrete-space diffusion equation (\[diffusion\]) will yield the usual diffusion equation $$\partial_t\rho(x,t)=\mu(1-\mu)^2\partial^2_x\rho(x,t),\; x\in[0,1], t\in{{\ensuremath{\mathbb R}} }_+,
\label{heat}$$ with Neuman boundary conditions. A more challenging task would be to derive (\[heat\]) for a fixed parameter $R$ directly in the diffusive scaling limit (when $N\to\infty$). In that case, the absence of “collisions" of sufficiently distant paths expressed in (\[recollisions\]) is no more automatic and one should rely on a decay of spatio-temporal correlations. In order to obtain such a decay, it might be necessary to define the model on a higher-dimensional lattice. More generally, the possibilities of modifying the parameters of the model are numerous : one could also take correlated distribution of scatterers and study the effect of the correlations on the conduction properties of the model.
We note that it is easy to find interactions (namely a fixed distribution of scatterers) between rings such that the empirical densities do not follow the diffusion equation (\[diffusion\]). In contrast to the seemingly “chaotic" orbits that are typical of systems obeying normal diffusion, those configurations of scatterers tend to have a long-range order and give rise to very ordered orbits. In the set of possible interactions they represent “integrable" dynamics. By analogy, this might teach us a useful point to keep in mind when studying systems described by real Hamiltonian dynamics. We tend to conceive and study models that are (random or not) perturbations of dynamics that are “simple" or lend themselves to computations . But those systems might be rather untypical in the space of all possible interactions or indeed in real physical systems that obey the ordinary laws of macroscopic physics.
[**Acknowledgments.**]{} I thank Jean Bricmont and Carlos Mejia-Monasterio for discussions. This work was supported by the ANR SHEPI grant.
[10]{} J. Bricmont, [*Science of Chaos or Chaos in Science?*]{} Physicalia Magazine, 17 (3-4), (1995) , 159-208 M. Kac, [*Some remarks on the use of probability in classical statistical mechanics*]{}, Acad. Roy. Belg. Bull. Cl. Sci. (5) 42 (1956), 356Ð361 J. Lebowitz, [*Boltzmann’s Entropy and Time’s Arrow*]{}, Physics Today, 46 (1993), 32-38.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have conducted an angular differential imaging survey with NIRC2 at Keck in search of close-in substellar companions to a sample of seven systems with confirmed planetary-mass companions (PMCs) on wide orbits ($>$50 AU). These wide-separation PMCs pose significant challenges to all three possible formation mechanisms: core accretion plus scattering, disk instability, and turbulent fragmentation. We explore the possibility that these companions formed closer in and were scattered out to their present-day locations by searching for other massive bodies at smaller separations. The typical sensitivity for this survey is $\Delta K$ $\sim$ 12.5 at 1$"$. We identify eight candidate companions, whose masses would reach as low as one Jupiter mass if gravitationally bound. From our multi-epoch astrometry we determine that seven of these are conclusively background objects, while the eighth near DH Tau is ambiguous and requires additional monitoring. We rule out the presence of $>$7 M$_{\rm Jup}$ bodies in these systems down to 15 – 50 AU that could be responsible for scattering. This result combined with the totality of evidence suggests that dynamical scattering is unlikely to have produced this population of PMCs. We detect orbital motion from the companions ROXs 42B b and ROXs 12 b, and from this determine 95$\%$ upper limits on the companions’ eccentricities of 0.58 and 0.83 respectively. Finally, we find that the 95$\%$ upper limit on the occurrence rate of additional planets with masses between 5 – 15 M$_{\rm Jup}$ outside of 40 AU in systems with PMCs is 54$\%$.'
author:
- 'Marta L. Bryan, Brendan P. Bowler, Heather A. Knutson, Adam L. Kraus, Sasha Hinkley, Dimitri Mawet, Eric L. Nielsen, Sarah C. Blunt'
title: 'Searching for Scatterers: High-Contrast Imaging of Young Stars Hosting Wide-Separation Planetary-Mass Companions'
---
=1
Introduction
============
Observational studies of exoplanet systems present a unique opportunity to probe the mechanisms behind planet formation. Over the past decade, surveys using a variety of techniques (radial velocity, transit, microlensing, direct imaging) have revealed a multitude of new systems with astoundingly diverse properties. Many of these systems are difficult to explain within the framework of standard planet formation theories [e.g. @Pollack1996; @Boss2006], and have forced theorists and observers alike to re-evaluate their narratives for planet formation and migration. Perhaps one of the biggest challenges for planet formation models comes from direct imaging surveys, which have uncovered of a new population of young planetary-mass companions (PMCs) ($<$ 15 M$_{\rm Jup}$) located beyond 50 AU.
In 2004 Chauvin et al. discovered a 5 M$_{\rm Jup}$ companion 2M1207 b orbiting 55 AU away from a 25 M$_{\rm Jup}$ brown dwarf. Shortly afterwards, additional discoveries of other wide-separation PMCs such as AB Pic b [@Chauvin2005], DH Tau b [@Itoh2005], and CHXR 73 b [@Luhman2006] drove observers and theorists to question how this growing population of objects formed [@Lodato2005; @Boss2006]. To date, fifteen PMCs at large orbital distances have been confirmed, most of which are extremely young, $<$10 Myr old [@Bowler2014]. Three possible formation routes have been proposed for these wide-separation planets, including direct collapse from molecular cloud fragmentation, disk instability, and core accretion plus gas capture, but all three have significant problems explaining this population of PMCs.
In the process of turbulent fragmentation, all stellar and substellar objects begin as opacity-limited fragments with masses of a few Jupiter masses and subsequently begin to accrete gas from the molecular cloud [@Low1976]. Hydrodynamical star formation simulations have shown that in order to stop accretion at brown dwarf or planetary masses, PMCs must either form at nearly the same time that the circumstellar envelope is exhausted, or else they must be dynamically ejected from the densest regions of gas before they are able to accrete much additional mass [@Bate2002; @Bate2009; @Bate2012]. This mechanism has a very difficult time producing binaries with the high mass ratios needed to match the observed wide-separation planetary systems.
In models of disk instability, gas giant planets form rapidly via fragmentation of a gravitationally unstable disk. For this model to work, the disk needs to be massive enough and cold enough to gravitationally collapse. In the majority of scenarios, the disk surface densities beyond 100 AU are too low for gravitational instability to operate. While some models show that disk fragmentation can occur outside 100 AU [@Dodson2009; @Boss2006; @Vorobyov2013], the fragments rarely survive to become full-fledged planetary embryos. This low survival probability is due to processes such as inward migration and accretion onto the host star, or ejection from the system due to dynamical interactions. While it has been suggested that disk instability could be effective for exceptionally massive disks, this is an extremely limited region of disk parameter space [@Vorobyov2013].
Finally, in the core accretion model, cores grow via successions of two-body collisions between solids until they are massive enough to start runaway gas accretion [@Pollack1996; @Alibert2005]. In situ formation of massive wide-separation planets through core accretion is unlikely since the timescale to grow massive cores at these separations is longer than the observed lifetimes of protoplanetary disks. However, recent simulations of core formation via pebble accretion have shown that gas giant cores can form at separations out to 50 AU comfortably before the gas in the disk dissipates [@Lambrechts2012]. Furthermore, it might be possible for these giant planets to form closer to the star and be subsequently scattered out beyond 100 AU by another planet in the system. One potential scenario is that if multiple planet-planet scatterings occur, these giant planets could permanently end up in stable, wide-separation orbits [@Scharf2009]. Simulations have shown that in this case, these wide-separation planets have high eccentricities of $>$0.5 [@Scharf2009; @Nagasawa2011]. While planet-planet scattering seems to be a potential solution, it requires another body in the system at least as massive as the wide separation planets.
Thus far, despite hundreds of hours of AO imaging, only one multi-planet system has been confirmed with this technique, HR 8799 [@Marois2008; @Marois2010]. Recently, two surveys have found evidence of additional planets in two more systems: LkCa 15 and HD 100546 [@Kraus2012; @Quanz2015; @Sallum2015; @Currie2015]. Searching for additional planets in these directly imaged systems is critical to understanding the formation and orbital evolution of planets at wide separations, a parameter space currently explored solely by the direct imaging technique.
In this study, we explore the possibility that the observed wide-separation PMCs formed closer in to their host stars, and were scattered out to their present day locations by another massive companion within the system. We conducted an angular differential imaging (ADI) survey with NIRC2 at Keck in search of close-in substellar companions to a sample of seven systems with confirmed PMCs on extremely wide orbits. Our observations are sensitive to companions at significantly lower masses and smaller separations than previous studies of these systems, and allow us to place much stronger constraints on the presence of inner companions. We also use these same systems to calculate the first estimate of the multiplicity of directly imaged planetary systems.
This paper is structured in the following manner. In Section 2 we describe the selected sample of systems and the methods for obtaining the ADI imaging data. In Section 3 we describe the PCA reduction of the images as well as a new method to simultaneously calculate astrometry and relative photometry of candidate companions. Finally, in Section 4 we discuss our results and their implications for the formation mechanisms of this population of wide-separation PMCs.
Observations
============
Target Selection
----------------
We selected our targets from the sample of 15 systems with confirmed companions beyond 50 AU with mass ranges that are either below or straddle the deuterium burning limit ($<$ 15 M$_{\rm Jup}$). These systems are as a whole extremely young, which translates into higher sensitivity to lower mass planets at smaller separations. From this larger sample, we selected targets that were observable from Keck and that had previously been imaged only with short integrations. This would allow our deeper follow-up imaging to achieve unprecedented levels of sensitivity in these systems. Altogether, we targeted seven systems: ROXs 42B, ROXs 12, HN Peg, HD 203030, DH Tau, LP 261–75, and 2MASS J012250–243950. Table 1 summarizes the properties of this sample. In addition, we targeted 2MASS J162627744–2527247, which does not have a previously confirmed wide-separation PMC. This star is a wide separation stellar companion to ROXs 12, located $\sim$40" away.[^1] Not only do the two stars show identical proper motion, but 2M1626–2527 and ROXs 12 also exhibit WISE excesses, indicating that these objects form a wide binary, are disk-bearing, and are young.
[lccccccccc]{} 2M0122–2439 & 01 22 50.94 & –24 39 50.6 & M3.5 & $9.20 \pm 0.03$ & $13.6$ & $36 \pm 4$ & 12 – 25 & $120 \pm 10$ & 1, 2, 3\
DH Tau & 04 29 41.56 & +26 32 58.3 & M1 & $8.18 \pm 0.03$ & $12.1$ & $145 \pm 15$ & $12^{+10}_{-4}$ & $1-2$ & 4, 6, 10, 13\
LP 261–75 & 09 51 04.60 & +35 58 09.8 & M4.5 & $9.69 \pm 0.02$ & $14.4$ & $32.9^{+3}_{-2}$ & $20^{+10}_{-5}$ & $100-200$ & 1, 6, 7, 10\
2M1626–2527 & 16 26 27.75 & –25 27 24.7 & M0 & $9.21 \pm 0.03$ & $15.8$ & $120 \pm 10$ & $\cdots$ & $8^{+4}_{-3}$ & 2, 4, 6\
ROXs 12 & 16 26 28.10 & –25 26 47.1 & M0 & $9.10 \pm 0.03$ & $13.5$ & $120 \pm 10$ & 12 – 20 & $8^{+4}_{-3}$ & 2, 4, 6, 12\
ROXs 42B & 16 31 15.02 & –24 32 43.7 & M1 & $8.67 \pm 0.02$ & $13.4$ & $120 \pm 10$ &6 – 14 & $7^{+3}_{-2}$ & 2, 4, 5, 6, 8\
HD 203030 & 21 18 58.22 & +26 13 49.9 & G8 & $6.65 \pm 0.02$ & 7.9 & $40.9 \pm 1.2$ & $23^{+8}_{-11}$ & $130-400$ & 6, 9, 10\
HN Peg & 21 44 31.33 & +14 46 19.0 & G0 & $4.56 \pm 0.04$ & 5.6 & $18.4 \pm 0.3$ & $21 \pm 9$ & $300-400$ & 6, 10, 11
[lcccccccc]{}
2M0122-2439 & 2014 Nov 9 & $K_{S}$ & 600 & 30 & 30 & 11.0 & 1.44 & 46.5 $\pm$ 0.6\
DH Tau & 2014 Dec 7 & $K_{S}$ & 600& 25 &25 & 57.6 & 1.07& 45.9 $\pm$ 1.9\
DH Tau & 2015 Nov 04 & $K_{S}$ & 600& 25 & 25 & 36.7 &1.01& 46.3 $\pm$ 1.2\
2M1626–2527 & 2014 May 13 & $K_{S}$& 600 & 28 & 28 & 11.3 & 1.44& 73.6 $\pm$ 11.6\
2M1626–2527 & 2015 Jun 23 & $K_{S}$ & 600& 25 & 25& 9.5 & 1.48& 45.8 $\pm$ 1.5\
ROXs 12 & 2011 Jun 23 & $K_{P}$ &300 & 27 & 13.5 & 5.4 & 1.56 & 47.7 $\pm$ 1.6\
ROXs 12 & 2015 Aug 27 & $K_{S}$ &600& 20 & 20& 6.4 & 1.54 & 53.0 $\pm$ 9.3\
ROXs 42B & 2011 Jun 23 & $K_{P}$ & 300& 46 & 23 & 13.9 & 1.44& 45.2 $\pm$ 4.9\
ROXs 42B & 2014 May 13 & $K_{S}$ &600 &30 &30 & 12.3 & 1.41 & 60.1 $\pm$ 10.1\
HD 203030 & 2014 Nov 9 & $K_{S}$ &600 & 60 & 30 & 12.4 &1.06 & 43.0 $\pm$ 0.3\
HD 203030 & 2015 Jun 3 & $K_{S}$ &600 & 80 & 40& 80.9 & 1.02& 40.4 $\pm$ 1.6\
HN Peg & 2014 Aug 4 & $K_{S}$ & 600& 50 & 25& 102.3 & 1.01& 47.7 $\pm$ 0.7\
HN Peg & 2015 Jun 2 & $K_{S}$ & 600& 70 & 35& 32.1 &1.04 & 39.6 $\pm$ 1.2
NIRC2 Imaging
-------------
We used the near-infrared imaging camera NIRC2 at the Keck II 10 m telescope for all of our observations. Adaptive optics imaging was carried out in natural guide star mode using the narrow camera. Due to the realignment of the Keck II AO and NIRC2 system in April 2015, for epochs taken prior to this date we used a plate scale of $0.009952 \pm 0.000002$ arcsec/pixel [@Yelda2010], and for epochs taken after this date we used a plate scale of $0.009971 \pm 0.000004$ arcsec/pixel [@Service]. The field of view of the 1024$\times$1024 array is 10.2“$\times$10.2”. For each system we obtained a total integration time of $\sim 30$ minutes in ADI mode with an average of $\sim 30$ degrees of sky rotation. For each image, the star was centered behind a coronagraph, which for nearly all images was the 600 mas diameter coronagraph. This coronagraph is partially transparent with 6.65 +/- 0.10 magnitudes of attenuation in Ks band [@Bowler2015]. Due to high noise levels in the lower left quadrant of the detector, we positioned the occulting spot (already fixed at row 430) at column 616. When possible, we observed these systems as they were transiting to maximize the rotation achieved during the observation, which makes post-processing PSF subtraction more effective. The inner working angle achieved for these observations is 300 mas, while the outer working angle for complete FOV coverage is $\sim 4"$. All observations were taken with the $Ks$ filter, which maximizes the Strehl ratio while avoiding the high sky backgrounds encountered in L band. We took second epoch data for the systems where we identified a candidate companion.
Analysis
========
PCA Image Reduction
-------------------
After removing bad pixels and cosmic rays and flat-fielding, we applied the @Yelda2010 distortion correction to raw images taken before April 2015, and applied the updated @Service distortion solution to later epochs. We then used principle component analysis (PCA) to further reduce these images. PCA is an algorithm that has recently been applied to high contrast imaging for increasing the contrast achievable next to a bright star. In short, PCA is a mathematical technique that relies on the assumption that every image in a stack can be represented as a linear combination of its principle orthogonal components, selecting structures that are present in most of the images. The stellar PSF, composed of a sum or orthogonal components, is subtracted from each image, thereby providing access to faint companions at contrasts below the speckle noise. We used a PCA routine following the method presented in @Soummer2012 which uses the KLIP algorithm.
The optimal number of principle components to use in a reduction is set by the trade-off between speckle noise and self-attenuation of the signal of interest. Too few components might not subtract enough speckle noise near the star, and too many may lead to self-subtraction of the planetary signal, reducing the achievable contrasts. In our analysis, we optimized the number of components used for each individual system empirically. We reduced the data for each system with different numbers of principle components, then compared resulting contrast curves that were calibrated for self-attenuation by injecting fake companions. For each system we adopted the number of principle components that corresponded to the most favorable contrast as the optimal number. These ranged from 5 – 20 principle components for systems in our sample.
We found a total of nine candidate companions at a wide variety of separations in the eight systems that we observed. These candidate companions are shown in Figure 1, and the contrast curves for all systems observed are shown in Figure 2. We determine our contrast curves by calculating the noise level in our images as a function of radial distance using the standard deviation on concentric annuli of width FWHM of the stellar PSF. The noise level at each radius is corrected for self-subtraction by dividing by the self-attenuation at that radius. This self-attenuation factor is calculated by injection and recovery of sources with known magnitudes at different radii. We present the 5$\sigma$ contrast curves in Figure 2, which are simply our noise levels divided by the self-attenuation factor multiplied by a factor of 5. We list 5$\sigma$ contrast values for a range of angular separations for each target in Table 3. We note that these contrasts are often limited by small PA rotation and subsequent ADI self-subtraction.
{width="100.00000%"}
[lccccccc]{} 2M0122-2439 & $1.4\times10^{-3}$ & 2.5$\times10^{-5}$ & 1.0$\times10^{-5}$ & $5.9\times10^{-6}$ & $5.6\times10^{-6}$ & $5.1\times10^{-6}$ & $4.8\times10^{-6} $\
DH Tau & $2.7\times10^{-4}$ & $3.2\times10^{-5}$ & $9.1\times10^{-6}$ & $7.4\times10^{-6}$ & $6.9\times10^{-6}$ & $5.2\times10^{-6}$ & $5.1\times10^{-6}$\
LP261-75 & $1.6\times10^-{4}$ & $2.2\times10^{-5}$ & $1.5\times10^{-5}$ & $1.6\times10^{-5}$ & $1.5\times10^{-5}$ & $1.4\times10^{-5}$ &$1.5\times10^{-5}$\
2M1626–2527 & $1.6\times10^{-3}$ & $8.4\times10^{-5}$ & $2.7\times10^{-5}$ & $1.6\times10^{-5}$ & $1.7\times10^{-5}$ & $1.5\times10^{-5}$ & $1.5\times10^{-5}$\
ROXs 12 & $2.7\times10^{-2}$ & $3.0\times10^{-4}$ & $2.9\times10^{-5}$ & $2.0\times10^{-5}$ & $1.5\times10^{-5}$ & $1.2\times10^{-5}$ & $1.3\times10^{-5}$\
ROXs 42B & $5.1\times10^{-4}$ & $3.6\times10^{-5}$ & $1.1\times10^{-5}$ & $5.8\times10^{-6}$ & $5.1\times10^{-6}$ & $4.7\times10^{-6}$ & $4.8\times10^{-6}$\
HD 203030 & $1.2\times10^{-4}$ & $1.0\times10^{-5}$ & $2.4\times10^{-6}$ & $9.7\times10^{-7}$ & $6.8\times10^{-7}$ & $6.2\times10^{-7}$ & $4.9\times10^{-7}$\
HN Peg & $3.4\times10^{-4}$ & $1.4\times10^{-5}$ & $2.4\times10^{-6}$ & $8.2\times10^{-7}$ & $4.2\times10^{-7}$ & $3.5\times10^{-7}$ & $2.9\times10^{-7}$
{width="100.00000%"}
Simultaneous Astrometry and Relative Photometry
-----------------------------------------------
Using second epoch data, we can determine whether or not these candidate companions are co-moving. Typical methods for determining the astrometry and photometry of candidate companions use the final post-processed images for these calculations. However, we note that because of self-subtraction, using the final LOCI or PCA images to calculate separations, position angles, and their uncertainties can lead to significant biases in the corresponding photometry and astrometry [e.g. @Marois2010]. To avoid this, we developed an MCMC algorithm that simultaneously calculates the astrometry and relative photometry of these candidate companions. For each iteration in the MCMC process, we injected a negative PSF into each individual science image in the vicinity of the companion of interest prior to de-rotation, where we modeled these negative PSFs as Moffat distributions. There were three parameters that we varied with each step during the MCMC routine, namely the negative PSF amplitude, separation, and position angle. We fixed all other free parameters to the values determined from fitting the Moffat distribution to the stellar PSF. The science images with injected negative PSFs were then run through the PCA reduction routine. The smaller the RMS noise at the location of the candidate companion, the better the fit of the negative PSF. The result of this MCMC analysis is a posterior distribution for the amplitude, separation, and position angle of the candidate companions, an example of which is shown in Figure 3. Figure 4 compares a reduced image with and without the best-fit negative PSF injected at the best fit separation and position angle of the candidate companion.
{width="100.00000%"}
{width="50.00000%"}
In addition to the errors from the candidate companion PDFs for the separation and position angle, after the MCMC program is finished we also account for uncertainties from the registration of the stellar position in each science image, uncertainties from the distortion correction, and uncertainties associated with the plate scale. We also include the $+0.252$ degree correction for north alignment to the NIRC2 header position angles [@Yelda2010] for epochs taken before April 2015, and for subsequent epochs include the $+0.262$ degree correction for north alignment [@Service]. This MCMC technique calculates robust uncertainties from the posterior distributions, without the systematic uncertainties that occur when the reduced LOCI or PCA images are used. The best-fit separations and position angles for each candidate companion, as well as confirmed companions in these systems, are presented in Table 4.
While the separation, position angle, and amplitude are the three parameters that were actively varied with each step in the MCMC program, we also track the total flux of the candidate companion. For each link in the chain, we place an aperture at the separation and position angle of that step and sum the number of counts in the aperture. While the size of the aperture for a given companion remains fixed, the aperture size ranged from 4 – 6 pixels depending on the candidate companion FWHM. The aperture moves with the changes in position angle and separation as the MCMC program progresses, producing a posterior distribution of counts for the candidate companion. We note that since we calculate the astrometry and relative photometry simultaneously using this MCMC program, we use the same aperture sizes for both of these steps for a given candidate companion.
We next calculate the contrast relative to the host star, $\Delta m$, for each candidate companion. To determine the flux from the star, the throughput of the occulting spot ($0.0022 \pm 0.0002$, Bowler et al 2015), and the sky noise need to be accounted for. The measured counts are a combination of the flux from the star plus the flux from the sky, both attenuated by the throughput of the mask. We denote this combined and attenuated star plus sky flux as $F_{b,\star}$. The corrected flux for the star is shown in Equation 1. $T$ is the throughput (0.0022) of the mask in $K_{S}$, $F_{sky}$ is the sky flux, and $F_{\star}$ is the flux from the star corrected for both throughput losses and sky noise.
$$F_{\star} = \frac{F_{\star,b}}{T} - F_{sky}$$
This calculation is performed for each image in the stack; for a stack of $N$ images, there are $N$ values of $F_{\star}$. The error on this flux value is:
$$\sigma_{F, \star} = \sqrt{\bigg(\frac{F_{b,\star}}{T}\bigg)^2 \times \bigg[\bigg(\frac{\sigma_{\star,b}}{F_{\star,b}}\bigg)^2 + \bigg(\frac{\sigma_{T}}{T}\bigg)^2\bigg] + \sigma_{sky}^2} .$$
Here, $\sigma_{\star,b}$ is the standard deviation of the $F_{\star,b}$ values for each image in the stack, $\sigma_T$ is the measured error on the throughput, and $\sigma_{sky}$ is the standard deviation of the sky values calculated for each image.
We obtain the companion flux and its uncertainty from the posterior distribution generated from the MCMC analysis, and subtract off sky noise. The flux ratio between the star and the companion is simply $F_{ratio} = \frac{F_{\star}}{F_{\rm comp}}$, and the error on this flux ratio can be propagated analytically: $$\sigma_{fr} = F_{ratio} \times \sqrt{\bigg(\frac{\sigma_{F,\star}}{F_{\star}}\bigg)^2 + \bigg(\frac{\sigma_{comp}}{F_{comp}}\bigg)^2} .$$
The contrast ratio $\Delta m$ in magnitudes is:
$$\Delta m = -2.5\log_{10}{F_{ratio}} .$$
Finally, the error on $\Delta m$ is given by:
$$\sigma_{\Delta m} = \frac{2.5}{\ln 10} \times \frac{\sigma_{fr}}{F_{ratio}} .$$
We present the $\Delta m$ values for each candidate and confirmed companion in Table 4. We note that because 2M0122-2439 cc1 is so faint we were unable to use our MCMC analysis to calculate the astrometry of that companion (the MCMC chains failed to converge). Instead, we used centroiding on the final image to obtain the separation and position angle of the candidate companion, and adopted the robust errors calculated for the faint HD 203030 cc1. Furthermore, since the candidate companion near 2M0122-2439 appears to be extended (its FWHM is about twice that of the stellar PSF), we conclude that is likely a background galaxy, not a bound planet, and exclude it from the rest of the analysis.
{width="100.00000%"}
{width="100.00000%"}
{width="100.00000%"}
[lcccccc]{} 2M0122–2439 b & 2014.8575 & $K_{S}$ & $1450^{+1}_{-1}$ & $216.48^{+0.02}_{-0.02}$ & $4.79 \pm 0.11$ & 52\
2M0122–2439 cc1 & 2014.8575 & $K_{S}$ & $5238^{+12}_{-16}$ & $355.75^{+0.14}_{-0.14}$ & $9.60 \pm 0.20$ & 188\
DH Tau cc1 & 2014.9342 & $K_{S}$ & $2726^{+7}_{-7}$ & $351.82^{+0.16}_{-0.16}$ & $8.66 \pm 0.25$ & 395\
DH Tau cc1 & 2015.8438 & $K_{S}$ & $2734^{+4}_{-4}$ & $351.35^{+0.10}_{-0.10}$ & $8.45 \pm 0.23$ & 396\
DH Tau b & 2014.9342 & $K_{S}$ & $2343^{+1}_{-1}$ &$140.25^{+0.02}_{-0.02}$ & $5.91 \pm 0.20$ & 340\
DH Tau b & 2015.8438 & $K_{S}$ & $2339^{+1}_{-1}$ & $139.94^{+0.02}_{-0.02}$ & $5.72 \pm 0.28$ & 340\
2M1626–2527 cc1 & 2014.3644 & $K_{S}$ & $1193^{+11}_{-14}$ & $266.88^{+0.26}_{-0.21}$ & $7.36 \pm 0.23$ & 143\
2M1626–2527 cc1 & 2015.4767 & $K_{S}$ & $1189^{+2}_{-1}$ & $267.93^{+0.03}_{-0.02}$ & $6.93 \pm 0.13$ & 143\
2M1626–2527 cc2 & 2014.3644 & $K_{S}$ & $4883^{+8}_{-7}$ & $107.99^{+0.06}_{-0.06}$ & $7.62 \pm 0.25$ & 586\
2M1626–2527 cc2 & 2015.4767 & $K_{S}$ & $4901^{+4}_{-3}$ & $107.82^{+0.03}_{-0.03}$ & $7.11 \pm 0.13$ & 589\
ROXs 12 cc1 & 2011.4767 & $K_{P}$ & $3811^{+10}_{-10}$ & $65.29^{+0.08}_{-0.07}$ & $\dots$ & 457\
ROXs 12 cc1 & 2015.6548 & $K_{S}$ & $3877^{+15}_{-14}$ & $64.12^{+0.09}_{-0.09}$ & $8.51 \pm 0.13$ & 465\
ROXs 12 b & 2011.4767 & $K_{P}$ & $1778^{+1}_{-1}$ & $8.90^{+0.08}_{-0.08}$ & $\dots$ & 213\
ROXs 12 b & 2015.6548 & $K_{S}$ & $1786^{+1}_{-1}$ & $8.18^{+0.29}_{-0.30}$ & $4.30 \pm 0.13$ & 214\
ROXs 42B cc1 & 2011.4767 & $K_{P}$ & $ 580^{+1}_{-1} $& $224.06^{+0.06}_{-0.06}$ & $\dots$ & 70\
ROXs 42B cc1 & 2014.3644 & $K_{S}$ & $525^{+1}_{-1}$ & $227.27^{+0.06}_{-0.06}$ & $6.23 \pm 0.27$ & 63\
ROXs 42B cc2 & 2011.4767 & $K_{P}$ & $3037^{+14}_{-11}$ & $138.99^{+0.14}_{-0.15}$ & $\dots$ & 365\
ROXs 42B cc2 & 2014.3644 & $K_{S}$ & $2991^{+8}_{-7}$ & $138.27^{+0.10}_{-0.12}$ & $8.37 \pm 0.25$ & 359\
ROXs 42B b & 2011.4767 & $K_{P}$ & $1173^{+1}_{-1}$ & $270.06^{+0.01}_{-0.01}$ & $\dots$ & 141\
ROXs 42B b & 2014.3644 & $K_{S}$ & $1170^{+1}_{-1}$ & $270.55^{+0.01}_{-0.01}$ & $6.16 \pm 0.35$ & 140\
HD 203030 cc1 & 2014.8575 & $K_{S}$ & $3379^{+12}_{-16}$ & $89.33^{+0.14}_{-0.14}$ & $11.17 \pm 0.15$ & 139\
HD 203030 cc1 & 2015.4219 & $K_{S}$ & $3263^{+10}_{-7}$ & $89.76^{+0.16}_{-0.22}$ & $11.46 \pm 0.20$ & 134\
HN Peg cc1 & 2014.5918 & $K_{S}$ & $2931^{+4}_{-5}$ & $19.06^{+0.13}_{-0.11}$ & $12.63 \pm 0.6$ & 54\
HN Peg cc1 & 2015.4192 & $K_{S}$ & $2933^{+2}_{-2}$ & $15.22^{+0.05}_{-0.05}$ & $12.13 \pm 0.48$ & 54
Our astrometry conclusively shows that seven of the remaining eight candidate companions are background objects, while the nature of the candidate companion near DH Tau is ambiguous. Figures 5 and 6 show the relative astrometry of each candidate companion compared to the expected background track of a stationary object. The candidate companion background track plots clearly show that the second epoch astrometry falls on or near the predicted track for a stationary background object. While some of the second epoch astrometry measurements don’t fall precisely on the expected track of a stationary background object, we note that this could be due to small errors in proper motion or distance, which would affect the predicted trajectory of a distant stationary object. We note that ROXs 42B cc1 was previously identified as likely a background object in the literature [@Kraus2014; @Currie2014] but our astrometry conclusively shows that it is a background object. For DH Tau cc1, while the second epoch astrometry falls close to the co-moving line, uncertainties on the expected trajectory of a background object make comovement ambiguous. The separation of a stationary object at the second epoch differs by $\sim 2.9\sigma$ from the separation we find for DH Tau cc1. @Zhou2014 published $HST$ UVIS optical photometry for the DH Tau system but they did not report a detection of our candidate companion. They presented detection limits in both $i$ and $z$ filters, which we can use to place limits on the colors of DH Tau cc1. We find that the bluest DH Tau cc1 could be is $i$ – $K$ = 7.8 mag. Furthermore, line of sight visual extinction is low, 0.0 - 1.5 mag [@Strom1989; @White2001]. This apparently red color further motivates additional follow-up for the potentially bound DH Tau cc1. A third epoch taken when DH Tau is next observable end of 2016 would conclusively determine whether or not DH Tau cc1 is a bound object.
We also plot the relative astrometry for the previously confirmed companions to ROXs 12, ROXs 42B, and 2M0122-2439 in Figure 7. We have included astrometry from the literature in addition to the data presented in this paper. These plots show that follow-up astrometry generally fall near the dotted line denoting co-moving objects. We do not plot the relative astrometry for DH Tau b, since there are significant systematic offsets for measurements of the companion position angle and separation amongst previous epochs spanning 1999 through 2013. Table 5 lists literature astrometry measurements that we used for the confirmed companions in each of these three systems.
[lcccc]{} 2M0122-2439 b & 2012.7808 & $1444 \pm 7$ & $216.7 \pm 0.2$ & 3\
2M0122-2439 b & 2013.0493 & $1448.6 \pm 0.6 $ & $216.4 \pm 0.08$ & 3\
2M0122-2439 b & 2013.4959 & $1448 \pm 4$ & $216.47 \pm 0.07$ & 3\
2M0122-2439 b & 2013.6258 & $1488 \pm 3$ & $216.52 \pm 0.09$ & 5\
ROXs 12 b & 2001.5014 & $1747 \pm 30$ & $10.3 \pm 0.9$ & 2\
ROXs 12 b & 2012.2575 & $1783.0 \pm 1.8$ & $8.85 \pm 0.06$ & 1\
ROXs 42B b & 2001.5014 & $1137 \pm 30$ & $268.0 \pm 1.5$ & 2\
ROXs 42B b & 2005.2904 & $1157 \pm 10$ & $268.8 \pm 0.6$ & 4\
ROXs 42B b & 2008.5479 & $1160 \pm 10$ & $269.7 \pm 1.0$ & 4\
ROXs 42B b & 2012.2575 & $1172.0 \pm 1.2$ & $270.03 \pm 0.10$ & 1\
ROXs 42B b & 2013.3233 & $1172.5 \pm 1.2$ & $270.25 \pm 0.10$ & 1
Orbital Motion
--------------
We tested ROXs 42B b, 2M0122-2439 b, and ROXs 12b for evidence of orbital motion. Assuming a face on, circular orbit, we find that between the first and last epoch of ROXs 42B b, the maximum amount of change we would expect to see in position angle is 0.6 degrees. The actual change in PA between the first and last epochs is $0.49 \pm 0.02$ degrees. We performed a linear fit to all epochs with uncertainties in PA and in separation and compared these to the best-fit constants using evidence ratios. Evidence ratios use Akaike’s information criterion (AIC) to quantitatively compare models. They are equal to the ratio of each model’s Akaike weights, which are a measure of the strength of evidence for a model. An evidence ratio of 9 comparing model 1 to model 2 would mean that model 1 is 9 times more likely than model 2 given the data. We label linear fits as preferred if the slope of the line differs from zero by 2 - 4$\sigma$, and highly preferred if this slope is $>4\sigma$ away.
In PA we find the evidence ratio comparing a linear to constant fit for ROXs 42B b to be $>10^4$, and in separation the evidence ratio is 62. The best fit slope of the linear fit in PA is $0.1703 \pm 0.0049$ deg/yr, and in separation is $-0.00132 \pm 0.00029$ arcsec/yr. We therefore conclude that the linear fits are highly preferred, suggesting that the displacements that we see in PA and separation over time are due to the orbital motion of ROXS 42B b.
For the confirmed companion orbiting 2M0122-2439, at a separation of only 52 AU we would expect this companion to have moved by 1.3 degrees in PA between the first and last epochs assuming a circular, face-on orbit. However, we only find a change in PA of $0.2 \pm 0.2$ degrees between the first and last epochs. Given that the change in PA is consistent with zero, and the evidence ratio for the separation of the companion favors a constant over a linear fit, we conclude that we do not find evidence of orbital motion for 2M0122-2439 b.
Finally, we assess whether orbital motion is evident for the confirmed companion ROXs 12 b. Between the first and last epochs, assuming a face-on circular orbit we would expect to see a change in PA of 1.6 degrees. We find a change of $2.1 \pm 0.9$ degrees. Evidence ratios comparing best linear fits to best fit constants through all four epochs including uncertainties are 4.5$\times$10$^3$ in PA and 10$^4$ in separation. The best fit slope of the linear fit in PA is $-0.164 \pm 0.048$ deg/yr and in separation is $0.00058 \pm 0.00032$ arcsec/yr. We conclude that linear fits are preferred, and that we likely see orbital motion from ROXs 12 b.
Using multiple epochs of astrometry allows us to constrain the orbits of ROXs 42B b and ROXs 12 b. To fit each orbit we use an updated implementation of the Rejection Sampling Monte Carlo method described in De Rosa et al. (2015), based on the method of Ghez et al. (2008). This technique generates an initial orbit with semi-major axis ($a$) of unity and position angle of nodes ($\Omega$) of 0, with eccentricity ($e$), inclination angle ($i$), argument of periastron ($\omega$), and epoch of periastron passage ($T_0$) drawn from the appropriate probability distribution: uniform for $e$, $\omega$, T$_0$, and uniform in cos($i$), and we use Kepler’s third law to generate the period from a fixed system mass. We then scale $a$ and rotate $\Omega$ to fit a single observational epoch, with observational errors included by adding Gaussian random noise to the observed separation and position angle for that epoch with $\sigma$ equal to the observational errors. Stellar mass and distance for each trial are both drawn from Gaussian distributions with medians at the measurements and standard deviations of the measurement uncertainties. Unlike De Rosa et al. (2015) where all potential orbits where shifted and scaled to the earliest epoch, here we randomly select an epoch for each orbit, which avoids the fit being biased toward the first epoch.
The algorithm has also been modified at the rejection sampling step: previously we proceeded one epoch at a time, rejecting ill-fitting orbits at each epoch. In this version the chi-square for the newly-scaled orbit is calculated for all the remaining epochs, and then the orbit is accepted if a uniform random variable is less than $e^{-\frac{\chi^2}{2}}$ and rejected otherwise. Mathematically this is the same operation as we used previously, but it allows for higher computational efficiency in the face of outliers, since the rejection test can be scaled to the minimum value of $\chi^2$ reached for the given astrometry, with orbits now accepted if the random variable is less than $e^{-\frac{\chi^2}{2}}$ / $e^{-\frac{\chi_{min}^2}{2}}$. This method will be described in more detail in Blunt et al. (2016, in prep).
This rejection sampling technique produces identical posterior probability distributions to those generated by MCMC, but requires much less computational time for astrometry covering short arcs of an orbit, as demonstrated in De Rosa et al. (2015) for the exoplanet 51 Eri b. In the very long-period orbits presented here we find that even after $10^{10}$ steps of Metropolis Hastings MCMC the chains have not converged, though the posteriors are broadly similar to those generated by the Rejection Sampling method.
Figures 8 and 9 show the range of Keplerian orbits consistent with the available astrometry for ROXs 12 b and ROXs 42B b respectively, while Figures 10 and 11 show the posterior distributions for the orbital parameters that were fit for ROXs 12 b and ROXs 42B b, respectively. We note that even with a small fraction of orbital coverage, fitting orbits and obtaining marginal constraints on the corresponding parameters is useful. For example, several recent studies have fit the small orbital coverage observed for Fomalhaut b, and find that they can constrain the eccentricity of this object to high values [@Kalas2013; @Beust2015]. Furthermore, detected orbital motion of the low mass brown dwarfs PZ Tel b and GQ Lup b appears to constrain their eccentricities to high values [@Ginski2014].
{width="100.00000%"}
{width="100.00000%"}
{width="100.00000%"}
{width="100.00000%"}
While the eccentricities of these PMCs are poorly constrained, we do find that low to moderate eccentricities are favored. The 95$\%$ upper limits on the eccentricities of ROXs 42B b and ROXs 12 b are 0.58 and 0.83 respectively. Previous studies have run scattering simulations to test if these wide-separation ($>$ 100 AU) PMCs can form via planet-planet scattering. These simulations showed that for giant planets that end up outside of 100 AU, their eccentricities are significantly pumped up to $>$ 0.5 [@Scharf2009; @Nagasawa2011]. The fact that the eccentricity distributions for ROXs 42B b and ROXs 12 b favor moderate to low eccentricities argues against the scattering hypothesis for these companions. Note that while a uniform prior on the eccentricity is used in these fits, the eccentricity posterior is significantly different. We can conclude that the eccentricity posterior is a reflection of the underlying companion eccentricity and not of the prior chosen. To further test this, we ran these orbit fits with two additional eccentricity priors, the $\beta$ distribution [@Kipping2013] and the thermal distribution [@Ambartsumian1937]. The thermal distribution of eccentricities, which is proportional to $2e$ d$e$, is the distribution that binary companions should follow if they are distributed solely as a function of energy. The eccentricity posteriors using these priors are overplotted with the eccentricity posterior found using a uniform prior in the top right plot in Figures 10 and 11. For both ROXs 12 b and ROXs 42B b, while the eccentricity posterior using the thermal distribution prior pushes to higher eccentricities, in general lower to moderate eccentricities are favored.
Detection Probability
---------------------
We calculate the detection probability for additional companions in these eight systems over a range of masses and separations. Our contrast curves can be converted into sensitivity maps in mass and semi-major axis using evolutionary models, the age and distance of the system and the uncertainties on these values, and an underlying distribution of planet eccentricities. Following Bowler et al. (2015), we generate a population of artificial companions on random, circular Keplerian orbits with a given mass and semi-major axis. Each synthetic planet is assigned an apparent magnitude using an interpolated grid of the Cond hot-start evolutionary models (Baraffe et al 2003), the distance and age of the host star, and the companion mass. We use the Cond evolutionary models because they extend down to planetary masses, although we note that different models can vary significantly in their predictions for the same planet mass. We do not explicitly account for this model-dependent error in our final analysis. The fraction of companions falling above a contrast curve compared to those falling below it yields the fractional sensitivity at that grid point. We further take into account the fractional field of view coverage for each target, which is uniformly complete out to 4" for our sample and drops to zero beyond that. Iterating over masses between 0.5–100 M$_{\rm Jup}$ and semi-major axes between 1–1000 AU yields sensitivity maps for each target, which are shown in Figure 12 for this sample. Depending on the distance and age of the target, our observations are generally sensitive to 1–10 M$_{\rm Jup}$ companions beyond about 30 AU.
{width="100.00000%"}
Discussion
==========
Can Dynamical Scattering Explain This Population?
-------------------------------------------------
Except for DH Tau cc1, which remains ambiguous, none of the new candidate companions detected in our sample are bound. While we can generally rule out the existence of massive scatterers above our detection limits (outside of 15-50 AU for massive planets), we cannot unambiguously rule out scattering as a formation mechanism based on our results alone, since the massive scatterers might be located closer in. However, our results combined with complementary lines of evidence suggest that formation close to the host star plus subsequent scattering is probably not the dominant formation mechanism for these wide-separation PMCs. Note that in order for the scattering scenario to operate, there must be a body that is often at least as massive as these already massive PMCs closer in to the host stars [@Veras2004]. We present a comprehensive list of this evidence below.
- In this study, we do not find any potential scatterers down to $\sim$ 15 - 50 AU in this sample of seven systems which host wide-separation PMCs. Furthermore, other studies with comparably deep imaging of wide separation PMCs also did not find any potential scatterers in HD 106906 and 1 RXS 1609-2105 [e.g. @Bailey2014; @Lafreniere2010; @Lagrange2015; @Kalas2015]. Efforts using non-redundant aperture masking techniques have probed higher masses ($>$15 M$_{\rm Jup}$) down to smaller separations ($>$ $\sim 5$ AU), and have likewise found a dearth of inner companions [@Cheetham2015; @Kraus2011]. If additional inner gas giant planets or brown dwarfs are present, they must be located within a few tens of AU of their host stars.
- Moderate to low eccentricities are favored for ROXs 12 b and ROXs 42B b, which both exhibit orbital motion. This is in contrast to the predictions of scattering simulations, which show that giant planets that get scattered out to $>$100 AU typically have high eccentricities $>$0.5 [@Nagasawa2011; @Scharf2009].
- From RV studies, it is clear that high-mass planets are rare. This is evident from the significantly negative power law in mass found by @Cumming2008 for a sample of giant planets 0.3 – 10 M$_{\rm Jup}$ out to 3 AU, where for a power law $m^{\alpha}$, $\alpha = -0.31 \pm 0.2$. Within this semi-major axis range, this power law implies that the occurrence rate of giant planets in the range 5 - 10 M$_{\rm Jup}$ is $1.3\%$. Similarly, work by Bryan et al (2016) suggests that for a sample of gas giant planets outside 5 AU, lower-mass planets are more frequent than higher-mass planets. This implies that planets massive enough to be potential scatterers for the wide-separation directly imaged planets in this study ($>$ 5 M$_{\rm Jup}$) are intrinsically rare in RV surveys. However, even if massive planets are disfavored generally, this doesn’t necessarily mean that there would only be one super-massive planet per system. For example, perhaps unusual disk properties are required to form a $>$ 5 M$_{\rm Jup}$ planet, but once this kind of disk is formed it is easy to form multiple massive planets at a range of separations. Given the low estimated occurrence rate of wide-separation PMCs (less than a few percent), this might be consistent with the low occurrence rate of massive planets found in RV surveys.
- Dynamical interactions between planets preferentially scatter out lower-mass planets [@Veras2009]. However, lower-mass planets ($<$ 5 M$_{\rm Jup}$) have not been discovered at distances greater than 100 AU despite the fact that many surveys were sensitive down to a few Jupiter masses [e.g. @Bowler2013; @Biller2013]. This implies that the companion mass function truncates at $\sim$5 M$_{\rm Jup}$, inconsistent with scattering.
- Rough estimates of the occurrence rate of these massive ($>$ 5 M$_{\rm Jup}$) wide-separation PMCs yield at most a frequency of a few percent [@Ireland2011; @Aller2013]. In contrast, scattering simulations, which began with 100 systems populated with 10 planets each with masses between 0.1 - 10 M$_{\rm Jup}$ drawn from a uniform distribution in logM and with separations $<$ 30 AU drawn from a uniform distribution in semi-major axis, find that the occurrence rate of scattered planets from 1 – 10 M$_{\rm Jup}$ outside of 300 AU is $\sim$0.2$\%$ at $\sim$10 Myr [@Scharf2009]. Furthermore, of the 0.2$\%$ occurrence rate of all planets from 1 – 10 M$_{\rm Jup}$ that get scattered beyond 300 AU, only a few percent of that population are $>$ 4 M$_{\rm Jup}$ [@Veras2009]. This implies that the occurrence rate of massive ($>$ 5 M$_{\rm Jup}$) scattered planets predicted by these simulations is of the order several hundredths of a percent, which is orders of magnitude smaller than any of the occurrence rate measurements from surveys thus far.
- Many of these widely-separation PMCs are actively accreting from a circumplanetary disk [@Zhou2014; @Bowler2011]. However, if these objects were dynamically scattered, we might expect the circumplanetary disk to be partially or completely stripped away [@Bowler2011]. This implies that many if not all of the PMCs that we find did not undergo such a violent evolution and were thus able to keep their disks.
Taken together, these lines of evidence indicate that the most likely origin for these wide-separation PMCs is in situ formation. Evidence for in situ formation, by cloud fragmentation or disk instability, includes the fact that PMCs have been found orbiting low-mass brown dwarfs with decidedly non-planetary mass ratios, implying that the tail of the initial mass function appears to continue down to at least 5 –10 M$_{\rm Jup}$. In addition, @Brandt2014 found that a single power law distribution is consistent with a sample of 5 – 70 M$_{\rm Jup}$ objects from the SEEDS survey. Given that results from many other surveys are well fit by this same power law distribution, this suggests that this population immediately below the deuterium-burning threshold are the end of a smooth mass function, sharing a common origin with more massive brown dwarfs.
Occurrence Rate
---------------
We now consider the multiplicity of directly imaged planetary systems. Although we did not find any new companions in our sample, we can place an upper limit on the occurrence rate of inner, massive planets in systems with previously known wide-separation PMCs. Since we have no detections, our occurrence rate is simply:
$$O = \frac{N_{m}}{N_{sys}}$$
where $N_{m}$ is the number of planets that we missed in our survey due to incompleteness, and $N_{sys}$ is the number of systems in our sample.
The number of planets that we missed due to survey incompleteness can be expressed as:
$$N_{m} = \Sigma_{i = 0}^{N_{sys}} \bigg[\int_{a_1}^{a_2} d\log a \int_{m_1}^{m_2} d\log m \hspace{0.1cm}f(m,a)(1 - P_i(m,a))\bigg] .$$
Here, $P_i(m,a)$ is the probability of detecting a planet of mass $m$ at semi-major axis $a$ for system $i$. We have these values for a grid of masses and semi-major axes from our detection probability calculations, described in section 3.4 and shown in Figure 8. The quantity $f(m,a)$ is the assumed distribution in mass and semi-major axis for the population of planets whose occurrence rate we wish to calculate. In our calculation we adopt the underlying distribution in @Clanton2015, which combines five different exoplanet surveys compiled using three different detection methods to derive a double power law distribution in mass and semi-major axis for giant planets. This power law takes the form:
$$f(m,a) = \frac{dN}{d\log m_p d\log a} = A \bigg (\frac{m_p}{M_{Sat}} \bigg )^{\alpha}\bigg (\frac{a}{2.5 AU} \bigg )^{\beta}$$
In this equation, $A = 0.21^{+0.20}_{-0.15}$, $\alpha = -0.86^{+0.21}_{-0.19}$, and $\beta = 1.1^{+1.9}_{-1.4}$. We note that this power law was derived specifically for M dwarf host stars. Out of our seven systems with previously confirmed PMCs, five of the host stars are M stars. We then create a 30$\times$30 grid evenly spaced in logarithmic bins with masses from 1 – 100 M$_{\rm Jup}$ and semi-major axes ranging from 1 – 1000 AU, and determine the power law distribution values at each grid point.
Since we want to determine the probability of finding an inner planet given that an outer PMC has been detected, we calculated the occurrence rate of PMCs between 5 – 15 M$_{\rm Jup}$ and from 40 AU to the location of each PMC. The inner limit on the separation was chosen because we are reasonably complete for massive planets beyond 40 AU for most of our systems. In order to take into account the large uncertainties on the power law parameters, we calculated the occurrence rate using a Monte Carlo method with $10^6$ trials, each time drawing a new $A$, $\alpha$, and $\beta$ value from a Gaussian distribution with widths equal to the parameter uncertainties. This yielded a distribution of missed planets, which we converted to a distribution in occurrence rate.
We found that the 95$\%$ confidence upper limit on the occurrence rate of planetary companions interior to our sample of previously known wide separation PMCs is 54$\%$. This result assumes the companion distribution shown in equation 8 as well as hot start evolutionary models. This first estimate of the occurrence rate upper limit will be better constrained with the discovery and analysis of more PMC systems. Note that covariances between parameters have not been taken into account in this method, which inflates our upper limit.
Conclusions
===========
We conducted a deep angular differential imaging (ADI) survey with NIRC2 at Keck in search of close-in substellar companions to a sample of seven systems with confirmed PMCs on extremely wide orbits ($>$100 AU). We explored the possibility that the wide-separation PMCs formed closer in to their host stars and were subsequently scattered out to their present day locations by a more massive body in the system. In this survey we obtained deep imaging for each target, for the first time probing significantly lower masses and smaller separations in all systems.
Within our sample we found eight candidate companions. Using second epoch data, we measured the astrometry for each candidate and determined whether or not they were co-moving by using an MCMC technique that calculates robust uncertainties from the posterior distributions, without the systematics that occur when the reduced images are used.
Seven candidate companions are unequivocally background objects, while the candidate companion near DH Tau remains ambiguous. Although our results alone do not conclusively rule out formation closer in to the host star followed by scattering as a formation mechanism for these wide-separation PMCs, the totality of evidence suggests that scattering is not a dominant formation mechanism. Instead, formation of these objects in situ appears to be more likely.
If we wish to better understand how these wide separation PMCs formed, there are several possible approaches to consider. $Gaia$ will allow us to carve out the immediate environment around these young stars, which has been extremely difficult with our current imaging capabilities (due to unfavorable contrasts close to the star), and radial velocity capabilities (due to high jitter values for young stars). Furthermore, studying the composition of these PMCs by obtaining high resolution spectra might allow us to distinguish amongst formation mechanisms [@Konopacky2013; @Barman2015]. While the core accretion model predicts that planets should have enhanced metallicities relative to their host stars, formation via disk instability or turbulent fragmentation should result in compositions matching those of the host star. Finally, large high-contrast imaging surveys of young star forming regions conducted homogeneously would give us a more precise measurement of the occurrence rates and orbital architectures of this population of planetary-mass objects.
{#section .unnumbered}
The data presented herein were obtained at the W.M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation. We acknowledge the efforts of the Keck Observatory staff. The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain.
Alibert, Y., Mordasini, C., Benz, W. et al. 2005, A$\&$A, 434, 1
Aller, K. M., Kraus, A. L., Liu, M. C. et al 2013, ApJ, 773, 1
Ambartsumian, V. A. 1937, Astron. Zh., 14, 207
Bailey, V., Meshkat, T., Reiter, M. et al 2014, ApJL, 780, 1
Baraffe, I., Chabrier, G., Barman, T. S. et al 2003, A$\&$A, 402
Barman, T. S., Konopacky, Q. M., Macintosh, B. et al 2015, ApJ, 804, 1
Bate, M. R., Bonnell, I. A., $\&$ Bromm, V. 2002, MNRAS, 332
Bate, M. R. 2009, MNRAS, 392
Bate, M. R. 2012, MNRAS, 419
Beust, H., Bonnefoy, M., Maire, A. L. et al 2015, A$\&$A, accepted
Biller, B. A., Liu, M. C., Wahhaj, Z. et al 2013, ApJ, 777, 2
Boss, A. P. 2006, ApJ, 637
Bowler, B. P., Liu, M. C., Kraus, A. L. et al 2011, ApJ, 743, 2
Bowler, B. P., Liu, M. C., Shkolnik, E. L. et al 2013, ApJ, 774, 1
Bowler, B. P., Liu, M. C., Kraus, A. L. et al 2014, ApJ, 784, 1
Bowler, B. P., Shkolnik, E. L., Liu, M. C. et al 2015, ApJ, 806, 1
Bowler, B. P., Liu, M. C., Shkolnik, E. L. et al 2015, ApJS, 216, 1
Brandt, T. D., McElwain, M. W., Turner, E. L. et al 2014, ApJ, 794, 2
Bryan, M. L., Knutson, H. A., Howard, A. W. et al 2016, ApJ, 821, 89
Carlsberg Meridian Catalogue 15, 2011
Chauvin, G., Lagrange, A. M., Lacombe, F. et al 2004, A$\&$A, 425
Chauvin, G., Lagrange, A. M., Zuckerman, B. et al 2005, A$\&$A, 438, 3
Cheetham, A. C., Kraus, A. L., Ireland, M. J. et al 2015, ApJ, 813, 2
Clanton, C. $\&$ Gaudi, S. 2015, arXiv:150804434C
Cumming, A., Butler, R. P., Marcy, G. W. et al. 2008, PASP, 120, 531
Currie, T., Cloutier, R., Brittain, S. et al 2015, ApJL, 814, 2
Currie, T., Daemgen, S., Debes, J. et al 2014, ApJ, 780, 2
Cutri, R. M., Skrutskie, M. F., van Dyk, S. et al 2013, 2MASS All Sky Catalog
Cutri, R. M., Wright, E. L., Conrow, T. et al 2013, AllWISE Data Release
De Rosa, R. J., Nielsen, E. L., Blunt, S. C. et al 2015, ApJ, 814, 3
Dodson-Robinson, S. E., Veras, D., Ford, E. B. et al 2009, ApJ, 707
Dupuy, T. J. $\&$ Liu, M. C. 2011, ApJ, 733, 2
Ghez, A. M., Salim, S., Weinberg, N. N. et al 2008, ApJ, 689, 1044
Ginski, C., Schmidt, T. O. B., Mugrauer, M. et al 2014, MNRAS, 444, 3
Ireland, M. J., Kraus, A., Martinache, F. et al 2011, ApJ, 726, 2
Itoh, Y., Hayashi, M., Tamura, M. et al 2005, ApJ, 620, 2
Kalas, P. G., Rajan, A., Wang, J. J. et al 2015, ApJ, 814, 1
Kalas, P., Graham, J. R., Fitzgerald, M. P. et al 2013, ApJ, 775, 1
Kipping, D. M. 2013, MNRAS, 434, L51
Konopacky, Q. M. et al 2013, Science, 229, 1398
Kraus, A. L., Ireland, M. J., Cieza, L. A. et al 2014, ApJ, 781, 1
Kraus, A. L., Ireland, M. J., Martinache, F. et al 2011, ApJ, 731, 1
Kraus, A. $\&$ Ireland, M. J. 2012, ApJ, 745, 1
Lafreniere, D., Jayawardhana, R., $\&$ van Kerkwijk, M. H. 2010, ApJ, 719, 1
Lagrange, A. M., Langlois, M., Gratton, R. et al 2015, arXiv: 1510.02511
Lambrechts, M. $\&$ Johansen, A. 2012, A$\&$A, 544
Lodato, G., Delgado-Donate, E., $\&$ Clarke, C. J. 2005, MNRAS, 364, 1
Low, C. $\&$ Lynden-Bell, D. 1976, MNRAS, 176
Luhman, K. L., Patten, B. M., Marengo, M. et al 2007, ApJ, 654, 1
Luhman, K. L., Wilson, J. C., Brandner, W. et al 2006, ApJ, 649, 2
Marois, C., Macintosh, B., Barman, T. et al 2008, Science, 322, 5906
Marois, C., Machintosh, B., $\&$ Veran, J. P. 2010, Proceedings of the SPIE, 7736
Metchev, S. A. $\&$ Hillenbrand, L. A. 2006, ApJ, 651, 2
Nagasawa, M. $\&$ Ida, S. 2011, ApJ, 742, 2
Pollack, J. B., Hubickyj, O., Bodenheimer, P. et al 1996, Icarus, 124, 1
Quanz, S. P., Amara, A., Meyer, M. R. et al 2015, ApJ, 807, 1
Ratzka, T., Kohler, R., $\&$ Leinert, C. 2005, A$\&$A, 437, 2
Reid, I. N. $\&$ Walkowicz, L. M. 2006, PASP, 118, 843
Sallum, S., Follette, K. B., Eisner, J. A. et al 2015, Nature, 527, 7578
Scharf, C. $\&$ Menou, K. 2009, ApJ, 693
Service, M., Lu, J. R., Campbell, R. et al 2016, submitted
Skiff, B. A. 2013, General Catalogue of Stellar Spectral Classifications
Soummer, R., Pueyo, L., $\&$ Larkin, J. 2012, ApJL, 755, 2
Strom, K. M., Strom, W. E., Edwards, S. et al 1989, AJ, 97, 1451
Veras, D., Crepp, J. R., Ford, E. B. 2009, ApJ, 696
Veras, D. $\&$ Armitage, P. J. 2004, MNRAS, 347, 2
Vorobyov, E. I. 2013, A$\&$A, 552
White, R. J. $\&$ Ghez, A. M. 2001, ApJ, 556, 265
Yelda, S., Lu, J. R., Ghez, A. M. et al 2010, ApJ, 725, 1
Zacharias, N., Monet, D. G., Levine, S. E. et al 2005, NOMAD Catalog
Zacharias, N., FInch, C. T., Girard, T. M. et al 2012, UCAC4
Zhou, Y., Herczeg, G. J., Kraus, A. L. et al 2014, ApJL, 783, L17
[^1]: There was some confusion with regards to follow-up observations of ROXs 12. The coordinates for ROXs 12 listed in Simbad and in both the discovery and confirmation papers of the PMC ROXs 12b [@Ratzka2005; @Kraus2014] are for 2M1626-2527, which does not have a confirmed PMC. The correct coordinates for ROXs 12 are listed in Table 1.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Suppose $M$ is a non-compact connected $n$-manifold without boundary, ${\mathcal D}(M)$ is the group of $C^\infty$-diffeomorphisms of $M$ endowed with the Whitney $C^\infty$-topology and ${\mathcal D}_0(M)$ is the identity connected component of ${\mathcal D}(M)$, which is an open subgroup in the group ${\mathcal D}_c(M)\subset{\mathcal D}(M)$ of compactly supported diffeomorphisms of $M$. It is shown that ${\mathcal D}_c(M)$ is homeomorphic to $N \times {\mathbb R}^\infty$ for an $l_2$-manifold $N$ whose topological type is uniquely determined by the homotopy type of ${\mathcal D}_0(M)$. For instance, ${\mathcal D}_0(M)$ is homeomorphic to $l_2 \times {\mathbb R}^\infty$ if $n = 1, 2$ or $n = 3$ and $M$ is orientable and irreducible.'
address:
- 'Instytut Matematyki, Uniwersytet Humanistyczno-Przyrodniczy Jana Kochanowskiego w Kielcach, Poland, andDepartment of Mathematics, Ivan Franko National University of Lviv, Lviv, 79000, Ukraine'
- 'Division of Mathematics, Kyoto Institute of Technology, Kyoto, 606-8585, Japan'
author:
- Taras Banakh
- Tatsuhiko Yagasaki
title: |
Diffeomorphism groups of non-compact manifolds\
endowed with the Whitney $C^\infty$-topology
---
Introduction
============
This paper is a continuation of the study of topological types of diffeomorphism groups of non-compact smooth manifolds endowed with the Whitney $C^\infty$-topology. Suppose $M$ is a $\sigma$-compact smooth $n$-manifold without boundary. Let ${\mathcal D}(M)$ denote the group of diffeomorphisms of $M$ endowed with the Whitney $C^\infty$-topology ($=$ the very-strong $C^\infty$-topology in [@Illman]) and ${\mathcal D}_0(M)$ the identity connected component of ${\mathcal D}(M)$. The group ${\mathcal D}(M)$ includes the normal subgroup ${\mathcal D}_c(M)$ consisting of diffeomorphisms with compact support.
In [@BMSY1 Theorem 4, Theorem 6.8] we have shown that ${\mathcal D}_c(M)$ is an $(l_2 \times {\mathbb R}^\infty)$-manifold and ${\mathcal D}_0(M)$ is an open subgroup of ${\mathcal D}_c(M)$. Here $l_2$ is the separable Hilbert space and ${\mathbb R}^\infty$ is the direct limit of the sequence $({\mathbb R}^n)_{n \in {\omega}}$, where ${\mathbb R}^n$ is identified with the hyperspace ${\mathbb R}^n \times \{ 0 \}$ in ${\mathbb R}^{n+1}$.
In a series of papers [@BR1; @BR2; @BR3] T. Banakh and D. Repovš studied topological properties of direct limits in the categories of uniform spaces. These results were applied in [@BMRSY] to yield a simple criterion for recognizing topological groups homeomorphic to open subspaces of $l_2 \times {\mathbb R}^\infty$ (see Theorem \[thm\_BR\] in Section 2).
In this paper we apply the above criterion to obtain the following important conclusion on the group ${\mathcal D}_0(M)$.
\[thm\_diff\] For any non-compact $\sigma$-compact smooth $n$-manifold without boundary the group ${\mathcal D}_c(M)$ is homeomorphic to an open subspace of $l_2 \times {\mathbb R}^\infty$.
In [@MS] K. Mine and K. Sakai obtained the triangulation theorem of open subsets of $l_2 \times {\mathbb R}^\infty$ (see Theorem \[thm\_MS\] in Section 2). This means that any open subset $U$ of $l_2 \times {\mathbb R}^\infty$ is homeomorphic to $N \times {\Bbb R}^\infty$ for some $l_2$-manifold $N$ whose topological type is uniquely determined by the homotopy type of $U$. Thus we obtain the following conclusion of the group ${\mathcal D}_0(M)$.
\[cor\_factor\] The group ${\mathcal D}_0(M)$ is homeomorphic to $N \times {\Bbb R}^\infty$ for some $l_2$-manifold $N$ whose topological type is uniquely determined by the homotopy type of ${\mathcal D}_0(M)$.
In some specific cases we can detect the homotopy type of ${\mathcal D}_0(M)$.
\[cor\_top\_type\] Let $M$ be a non-compact connected smooth $n$-manifold without boundary.
1. If $1\le n\le 2$, then ${\mathcal D}_0(M) \approx l_2 \times {\mathbb R}^\infty$.
2. If $n = 3$ and the manifold $M$ is orientable and irreducible (i.e., any smooth $2$-sphere in $M$ bounds a $3$-ball in $M$), then ${\mathcal D}_0(M) \approx l_2 \times {\mathbb R}^\infty$.
3. If $M$ is the interior ${\mathrm{Int}}X = X \setminus \partial X$ of a compact connected smooth $n$-manifold $X$ with boundary, then $${\mathcal D}_0(M) \approx {\mathcal D}_0(X, \partial X) \times {\mathbb R}^\infty.$$ In particular, if ${\mathcal D}_0(X, \partial X)$ is contractible, then ${\mathcal D}_0(M) \approx l_2 \times {\mathbb R}^\infty$.
For example, if $M$ is the 3-dimensional Euclidean space ${\Bbb R}^3$ or the Whitehead contractible 3-manifold [@Hempel], then ${\mathcal D}_0(M) \approx l_2 \times {\mathbb R}^\infty$.
Open subspaces of LF-spaces
===========================
In this preliminary section we recall the criterion for recognizing topological groups homeomorphic to open subsets of $l_2 \times {\mathbb R}^\infty$. First we recall some necessary definitions. Below a [*Polish space*]{} means a separable completely metrizable space; a [*Polish group*]{} is a topological group whose underlying topological space is Polish.
A subgroup $H$ of a topological group $G$ is called [*locally topologically complemented*]{} (LTC) in $G$ if $H$ is closed in $G$ and the quotient map $q : G \to G/H=\{xH:x\in G\}$ is a locally trivial bundle. This condition is equivalent to saying that $q$ has a local section at some point of $G/H$. Here, a [*local section*]{} of a map $q: X \to Y$ at a point $y \in Y$ means a continuous map $s: U \to X$ defined on a neighborhood $U$ of $y$ in $Y$ such that $q\circ s={\mathrm{id}}_U$.
Following [@BMRSY], we say that a topological group $G$ carries the [*strong topology*]{} with respect to a tower of subgroups $$G_0\subset G_1\subset G_2\subset\cdots$$ if $G=\bigcup_{n\in{\omega}}G_n$ and for any neighborhood $U_n$ of the neutral element $e$ in $G_n$, $n \in {\omega}$, the group product $${\operatornamewithlimits{\overrightarrow{\textstyle\prod}}}_{n\in{\omega}}U_n=\bigcup_{n\in{\omega}} U_0U_1\cdots U_n$$is a neighborhood of $e$ in $G$. In this case the topology of $G$ coincides with the topology of the direct limit ${\mathrm{g}\mbox{-}\kern-2pt\varinjlim}G_n$ of the tower $(G_n)_{n\in{\omega}}$ in the category of topological groups, which means that $G$ carries the strongest group topology such that the idenity maps $G_n\to G$, $n\in{\omega}$, are continuous.
The following criterion is obtained in [@BMRSY Theorem 11].
\[thm\_BR\] A non-metrizable topological group $G$ is homeomorphic to an open subset of ${\mathbb R}^\infty$ or $l_2 \times {\mathbb R}^\infty$ if $G$ carries the strong topology with respect to an LTC-tower of Polish ANR-groups $(G_n)_{n\in{\omega}}$.
Open subspaces of $l_2\times{\mathbb R}^\infty$ were studied in [@MS; @MS2] and the following Triangulation Theorem was obtained.
\[thm\_MS\] [(1)]{} Each open subspace $X$ of $l_2\times{\mathbb R}^\infty$ is homeomorphic to the product $K\times l_2\times{\mathbb R}^\infty$ for a locally finite simplicial complex $K$.
[(2)]{} Two open subspaces of $l_2\times{\mathbb R}^\infty$ are homeomorphic if and only if they are homotopically equivalent.
Note that the product $N = K \times l_2$ is an $l_2$-manifold and its topological type is determined by its homotopy type.
Diffeomorphism groups of non-compact manifolds {#Diffeo groups}
==============================================
Suppose $M$ is a non-compact $\sigma$-compact smooth $n$-manifold without boundary. We can represent $M$ as the countable union $M=\bigcup_{i\in{\omega}}M_i$ of compact $n$-submanifold $M_i\subset M$, $i\in{\omega}$, such that $M_i \subset {\mathrm{Int}}M_{i+1}$. Let $M_{-1}=\emptyset$ and consider the $n$-submanifolds $K_i = M \setminus {\mathrm{Int}}M_i$, $i\in{\omega}$, of $M$ and closed subgroups ${\mathcal D}(M; K_i) = \{ h \in {\mathcal D}(M) : h|_{K_i} = {\mathrm{id}}_{K_i}\}$ of the diffeomorphism group ${\mathcal D}(M)$ endowed with the Whitney $C^\infty$-topology, see [@Illman]. Thus we obtain the group $G = {\mathcal D}_c(M)$ and the tower $(G_i)_{i\in{\omega}}$ of closed subgroups $G_i = {\mathcal D}(M; K_i)$ of $G$.
The small box product ${\boxdot}_{i \in {\omega}} G_i$ is defined by $${\boxdot}_{i\in{\omega}} G_i = \big\{(x_i)_{i\in{\omega}}\in\square_{i\in{\omega}}G_i :
\exists \,k\in{\omega}\;\forall\, i\ge k, \ x_i= e \,\big\}.$$ This space is endowed with the box topology generated by the base consisting of boxes ${\boxdot}_{i\in{\omega}}U_i$, where $U_i$ is an open set of $G_i$. The left multiplication map $$\pi : {\boxdot}_{i \in {\omega}} G_i \to G, \ \pi(x_0,\dots,x_k,e,e,\dots) = x_0 \cdot x_1\cdots x_k$$ is continuous ([@BMSY1 Lemma 2.10]).
We shall show that the tower $(G_i)_{i \in {\omega}}$ has the properties listed in Lemma \[lemma\_G\_i\] below. Hence, Theorem \[thm\_diff\] now follows from Theorem \[thm\_BR\].
\[lemma\_G\_i\] [(1)]{} $G = \bigcup_{i\in{\omega}} G_i$ and the group $G$ is not metrizable.
- $G_i$ is a separable $l_2$-manifold for each $i \in {\omega}$.
- $G_i$ is TLC in $G_{i+1}$ for each $i \in {\omega}$.
- The left multiplication map $\pi : {\boxdot}_{i \in {\omega}} G_i \to G$ admits a local section at any points. Hence, the map $\pi$ is an open map and the group $G$ carries the strong topology with respect to the sequence $G_i$ $(i\in{\omega})$.
- $G = {\mathrm{g}\mbox{-}\kern-2pt\varinjlim}G_i$.
- Let $H$ and $H_i$ $(i \in {\omega})$ denote the identity connected components of $G$ and $G_i$ $(i \in {\omega})$ respectively. Then $H$ is an open subgroup of $G$ and the sequence $H_i$ $(i\in{\omega})$ of closed subgroups of $H$ also have the above properties $(1) \sim (5)$.
The statement (1) easily follows from the definitions of $G$ and $G_i$ themselves. The statement (4) follows directly from [@BMSY1 Proposition 5.5(2)]. (The proof of Theorem 6.8 in [@BMSY1] assures that Proposition 5.5 can be applied to this setting.) The statement (5) now follows from (4) and [@BMRSY Proposition 1].
\(2) Since $M - {\mathrm{Int}}\,K_i = M_i$ is compact, the group $G_i$ is an infinite-dimensional separable Fréchet manifold (cf. [@Hmlt], [@Les]).
\(3) The assertion follows directly from the bundle theorem, Theorem \[thm\_bdl\], explained below.
\(6) Since $H_i$ is an open subgroup of $G_i$ for each $i \in {\omega}$, it suffices to show that $H = \cup_{i \in {\omega}} H_i$. The group $H$ is path-connected, since it is the identity connected component of $G$ and the latter is locally path-connected by (2) and (4). Hence any $h \in H$ can be joined to ${\mathrm{id}}_M$ by an arc $A$ in $H$. By [@BMSY1 Proposition 3.3] the compact subset $A$ lies in $G_n$ for some $n \in {\omega}$. This means that $h \in H_n$.
For an $n$-submanifold $L $ of $M$ and a subset $K \subset L$ let ${\mathcal E}_K(L, M)$ denote the space of $C^\infty$-embeddings $f : L \to M$ with $f|_K = {\mathrm{id}}_K$ endowed with the compact-open $C^\infty$-topology. There is a natural restriction map $$r : {\mathcal D}(M, K) \to {\mathcal E}_K(L, M), \quad r(h) = h|_L.$$
The following is the classical bundle theorem in codimension 0 (cf. [@Cerf], [@Hmlt], [@Palais1], [@Seeley]).
\[thm\_bdl\] Suppose $K, L$ are $n$-submanifolds of $M$ which are closed subsets of $M$, $K \subset {\mathrm{Int}}L$ and ${\mathrm{cl}}_M(L \setminus K)$ is compact. Then, for any closed subset $C$ of $M$ with $C \cap L = \emptyset$, the restriction map $r : {\mathcal D}(M, K) \to {\mathcal E}_K(L, M)$ has a local section $$s : {\mathcal U}\to {\mathcal D}_0(M, K \cup C) \subset {\mathcal D}(M, K)$$ at the inclusion $i_L : L \subset M$ such that $s(i_L) = {\mathrm{id}}_M$.
For the proof of Corollary \[cor\_top\_type\] we need a preliminary. For any pairs of spaces $(X, A)$ and $(Y, B)$ let $[X,A; Y, B]$ denote the set of homotopy classes of maps of pairs. Any map of pairs $f : (Y, B) \to (Z, C)$ induces a function $$f_{\#} : [X, A; Y, B] \to [X, A; Z, C], \ f_{\#} : [g] \to [fg].$$
Suppose $L$ is a compact space and $K$ is a closed subset of $L$. The inclusion maps $H_i \subset H_{i+1}$ and $H_i \subset H$ $(i \in {\omega})$ induce the associated functions between pointed sets: $$\xymatrix@M+1pt{
[L, K; H_i, {\mathrm{id}}_M] \ar[rr]^{} \ar[dr]_{} & & [L, K; H_{i+1}, {\mathrm{id}}_M] \ar[dl]^{} \\
& [L, K; H, {\mathrm{id}}_M] &
}$$ Taking the direct limit, we obtain a function between pointed sets $$\iota : \varinjlim\,[L, K ; H_i, {\mathrm{id}}_M] \longrightarrow [L, K; H, {\mathrm{id}}_M].$$ Since any compact subset of $H$ is included in some $H_i$, we have the following conclusion.
For any pair of compact spaces $(L, K)$ the inclusion induced function $$\iota : \varinjlim\,[L, K ; H_i, {\mathrm{id}}_M] \longrightarrow [L, K; H, {\mathrm{id}}_M]$$ is a bijection.
For $m = 0, 1, \cdots, \infty$, a map $f : X \to Y$ between path-connected spaces is called an $m$-equivalence if for some base point $x \in X$, the induced homomorphism on the $k$-th homotopy group $$f_\# : \pi_k(X, x) \to \pi_k(Y, f(x))$$ is an isomorphism for $k = 0, 1, \cdots, m-1$ and an epimorphism for $k = m$. An $\infty$-equivalence is called a weak equivalence. If both $X$ and $Y$ have the homotopy type of CW-complexes, then every weak equivalence is a homotopy equivalence. Note that the groups $H$ and $H_i$ $(i \in {\omega})$ are path-connected and have the homotopy type of CW-complexes.
\[cor\_homotopy\_type\] For $m = 0, 1, \cdots, \infty$, if each inclusion $H_i \subset H_{i+1}$ is an $m$-equivalence, then so is the inclusion $H_1 \subset H$. For example, each $H_i$ is contractible, then so is $H$ and hence $H \approx l_2 \times {\mathbb R}^\infty$.
We keep the notations $M_i$, $K_i$, $H_i$ $(i \in {\omega})$ and $H$.
(1), (2) Since $M$ is connected, we may assume that for each $i \in {\omega}$ (a) $M_i$ is connected and (b) each connected component of $K_i = M \setminus {\mathrm{Int}}M_i$ is non-compact. By Corollary \[cor\_homotopy\_type\] it suffices to show that each $H_i$ is contractible. Note that the inclusion $H_i \subset {\mathcal D}_0(M_i, \partial M_i)$ is a homotopy equivalence.
For $n = 1 ,2$ the assertion follows from [@EE], [@Ivanov3 Section 2.7], [@Sc], [@Sm], etc. In the case $n = 3$, if $M_i$ is a 3-ball, then ${\mathcal D}_0(M_i, \partial M_i)$ is contractible by the Smale conjecture [@Hat2 Appendix (1)]. If $M_i$ is not a 3-ball, then by the assumption, $M_i$ is an orientable Haken 3-manifold with boundary [@Hempel; @Waldhausen] and ${\mathcal D}_0(M_i, \partial M_i)$ is contractible by [@Hat1], [@Ivanov1], [@Ivanov2].
\(3) Take a collar $\partial X \times [0,1]$ of $\partial X = \partial X \times \{0\}$ in $X$ and let $M_i = X \setminus (\partial X \times [0, 1/i))$ and $K_i = M - {\mathrm{Int}}M_i = \partial X \times (0, 1/i]$ ($i\in{\mathbb N}$). First we shall show that the inclusion $H_i \subset H_{i+1}$ is a homotopy equivalence. Consider the restriction map $\pi : H_{i+1} \to {\mathcal E}_{K_{i+1}}(K_i, M)$. Since ${\mathcal E}_{K_{i+1}}(K_i, M)$ is the space of embeddings of the collar $K_i$ relative to $K_{i+1}$, it is seen that ${\mathcal E}_{K_{i+1}}(K_i, M)$ is contractible. Since ${\mathcal E}_{K_{i+1}}(K_i, M)$ is path-connected, from Theorem \[thm\_bdl\] it follows that the map $\pi$ is onto and is a principal bundle with the structure group $H_{i+1}\cap G_i$. Since ${\mathcal E}_{K_{i+1}}(K_i, M)$ is contractible and paracompact, this bundle is trivial and $H_{i+1} \approx {\mathcal E}_{K_{i+1}}(K_i, M) \times (H_{i+1}\cap G_i)$. Since $H_{i+1}$ is path-connected, it follows that $H_i = H_{i+1}\cap G_i$ and the inclusion $H_i \subset H_{i+1}$ is a homotopy equivalence.
From Corollary \[cor\_homotopy\_type\] it follows that $H \simeq H_1 \simeq {\mathcal D}_0(M_1, \partial M_1) \approx
{\mathcal D}_0(X, \partial X)$. Since the last one is an $l_2$-manifold, we have $H \approx {\mathcal D}_0(X, \partial X) \times {\mathbb R}^\infty$ by Corollary \[cor\_factor\].
[MM]{}
T. Banakh, K. Mine, D. Repovš, K. Sakai, T. Yagasaki, [*On topological groups (locally) homeomorphic to LF-spaces*]{}, preprint (arXiv:1004.0305).
T. Banakh, K. Mine, K. Sakai, T. Yagasaki, [*Homeomorphism and diffeomorphism groups of non-compact manifolds with the Whitney topology*]{}, Topology Proc. (arXiv:0802.0337v2).
T. Banakh, K. Mine, K. Sakai, T. Yagasaki, [*On homeomorphism groups of non-compact surfaces, endowed with the Whitney topology*]{}, preprint (arXiv:1004.3015).
T. Banakh, D. Repovš, [*The topological structure of direct limits in the category of uniform spaces*]{}, Topology Appl. [**157**]{} (2010) 1091-1100.
T. Banakh, D. Repovš, [*Direct limit topologies in the categories of topological groups and of uniform spaces*]{}, preprint (ArXiv:0911.3985).
T. Banakh, D. Repovš, [*A topological characterization of LF-spaces*]{}, preprint (arXiv:0911.0609).
J. Cerf, [*Topologie de certains espaces de plongements*]{}, Bull. Soc. Math. France [**89**]{} (1961), 227–380.
T. Dobrowolski, H.Torunczyk, [*Separable complete ANRs admitting a group structure are Hilbert manifolds*]{}, Topology Appl. [**12:3**]{} (1981) 229–235.
Earle, C. J. and Eells, J., [*A fiber bundle discription of Teichmüller theory,*]{} [J. Diff. Geom.,]{} [**3**]{} (1969) 19–43.
R. S. Hamilton, [*The inverse function theorem of Nash and Moser*]{}, Bull. Amer. Math. Soc. (N.S.) [**7**]{}:1 (1982), 65–222.
A. Hatcher, [*Homeomorphisms of sufficiently large $P\sp{2}$-irreducible $3$-manifolds*]{}, Topology [**15**]{}, no.4 (1976), 343–347.
A. Hatcher, [*A proof of the Smale conjecture, $\mathrm{Diff}(S^3)\simeq O(4)$*]{}, Ann. of Math. [**117**]{}(2) (1983), 553–607.
A. Hatcher, Algebraic Topology, Cambridge Univ. Press, 2002.
J. Hempel, $3$-Manifolds, Ann. of Math. Studies, [**86**]{}, Princeton Univ. Press, Princeton, N.J., 1976.
S. Illman [*The very-strong $C^\infty$ topology on $C^\infty(M,N)$ and $K$-equivariant maps*]{}, Osaka J. Math. [**40**]{}, no. 2 (2003) 409–428.
N. V. Ivanov, [*Diffeomorphism groups of Waldhausen manifolds*]{}, J. Soviet Math. [**12**]{}, No. 1 (1979), 115–118. Russian original: Research in Topology, II, Notes of LOMI scientific seminars, v. [**66**]{} (1976), 172–176.
N. V. Ivanov, [*Spaces of surfaces in Waldhausen manifolds*]{}, Preprint LOMI P-5-80 (1980) (in Russian).
N. V. Ivanov, [*Mapping class groups*]{}, in: R.J. Daverman and R.B. Sher, (eds.), Handbook of Geometric Topology, Elsevier Sci. Publ. B.V., Amsterdam, 2002, pp. 523–633
J. A. Leslie, [*On a differential structure for the group of diffeomorphisms*]{}, Topology [**6**]{} (1967), 263–271.
K. Mine, K. Sakai, [*Open subsets of LF-spaces*]{}, Bull. Pol. Acad. Sci. Math. [**56**]{}:1 (2008), 25–37.
K. Mine, K. Sakai, [*Simplicial complexes and open subsets of LF-spaces*]{}, preprint.
R. S. Palais, [*Local triviality of the restriction map for embeddings*]{}, Comment. Math. Helv. [**34**]{} (1960), 305–312.
G. P. Scott, The space of homeomorphisms of 2-manifold, [*Topology,*]{} 9 (1970) 97–109.
R. T. Seeley, [*Extension of $C\sp{\infty }$ functions defined in a half space*]{}, Proc. Amer. Math. Soc. [**15**]{} (1964), 625–626.
S. Smale, Diffeomorphisms of the 2-sphere, [*Proc. Amer. Math. Soc.,*]{} 10 (1959) 621–626.
F. Waldhausen, [*On irreducible $3$-manifolds which are sufficiently large*]{}, Ann. of Math. (2) [**87**]{} (1968), 56–88.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define a symmetric derivative on an arbitrary nonempty closed subset of the real numbers and derive some of its properties. It is shown that real-valued functions defined on time scales that are neither delta nor nabla differentiable can be symmetric differentiable.'
address:
- 'Escola Superior de Tecnologia de Setúbal, Estefanilha, 2910-761 Setúbal, Portugal Center for Research and Development in Mathematics and Applications, University of Aveiro, 3810-193 Aveiro, Portugal'
- 'Center for Research and Development in Mathematics and Applications, Department of Mathematics, University of Aveiro, 3810-193 Aveiro, Portugal'
- 'Center for Research and Development in Mathematics and Applications, Department of Mathematics, University of Aveiro, 3810-193 Aveiro, Portugal'
author:
- 'Artur M. C. Brito da Cruz'
- Natália Martins
- 'Delfim F. M. Torres'
title: Symmetric Differentiation on Time Scales
---
[^1]
Introduction
============
Symmetric properties of functions are very useful in a large number of problems. Particularly in the theory of trigonometric series, applications of such properties are well known [@Zygmund]. Differentiability is one of the most important properties in the theory of functions of real variables. However, even simple functions such as $$\label{ts:1}
f\left( t\right)
=\left \vert t\right \vert,
\quad g\left( t\right)
= \begin{cases}
t\sin \frac{1}{t} \, , & t\neq 0 \\
0 \, , & t=0,
\end{cases}
\quad h\left( t\right) = \frac{1}{t^{2}} \, , t \neq 0,$$ do not have (classical) derivative at $t=0$. Authors like Riemann, Schwarz, Peano, Dini, and de la Vallée-Poussin, extended the classical derivative in different ways, depending on the purpose [@Zygmund]. One of those notions is the symmetric derivative: $$\label{eq:sd:r}
f^{s}\left( t\right)
=\lim_{h\rightarrow 0}\frac{f\left( t+h\right) -f\left(t-h\right) }{2h}.$$ While the functions in do not have ordinary derivatives at $t=0$, they have symmetric derivatives: $f^{s}\left( 0\right) = g^{s}\left( 0\right) = h^{s}\left( 0\right) =0$. For a deeper understanding of the symmetric derivative and its properties, we refer the reader to the specialized monograph [@Thomson]. Here we note that the symmetric quotient $\left(f\left( t+h\right) -f\left( t-h\right)\right)/(2h)$ has, in general, better convergence properties than the ordinary difference quotient [@Serafin], leading naturally to the so-called $h$-symmetric quantum calculus [@Kac]. A more recent theory is the general time-scale calculus. In 1988, Hilger introduced the calculus on time scales as a generalization of continuous and discrete time theories, obviating the need for separate proofs and highlighting the differences between them [@Hilger; @Bohner:1]. Here we introduce the notion of symmetric derivative on time scales, initiating the corresponding theory and putting into context some of the recent results found in the literature.
The article is organized as follows. In Section \[ts:sec:pre\] we review the necessary concepts and we fix notations. The results are then given in Section \[ts:sec:dif\], where we define the time scale symmetric derivative and derive some of its properties. Applications are found in the context of quantum calculus [@Kac]. Finally, we show in Section \[ts:sec:fc\] that the new symmetric derivative is a generalization of the diamond-$\alpha$ derivative [@Sheng], which brings new insights.
Preliminary notions and notations {#ts:sec:pre}
=================================
A nonempty closed subset of $\mathbb{R}$ is called a time scale and is denoted by $\mathbb{T}$. We assume that a time scale has the topology inherited from $\mathbb{R}$ with the standard topology. Two jump operators are considered: the forward jump operator $\sigma : \mathbb{T} \rightarrow \mathbb{T}$, defined by $\sigma \left( t\right) :=\inf \left \{ s\in \mathbb{T}:s>t\right \}$ with $\inf \emptyset =\sup \mathbb{T}$ (i.e., $\sigma \left( M\right) =M$ if $\mathbb{T}$ has a maximum $M$), and the backward jump operator $\rho :\mathbb{T} \rightarrow \mathbb{T}$ defined by $\rho \left( t\right) :=\sup \left \{ s\in \mathbb{T}:s<t\right \}$ with $\sup \emptyset =\inf \mathbb{T}$ (i.e., $\rho \left( m\right) =m$ if $\mathbb{T}$ has a minimum $m$). A point $t\in \mathbb{T}$ is said to be right-dense, right-scattered, left-dense or left-scattered if $\sigma \left( t\right) =t$, $\sigma \left( t\right) >t$, $\rho \left( t\right) =t$ or $\rho \left( t\right) <t$, respectively. A point $t\in \mathbb{T}$ is dense if it is right and left dense; isolated if it is right and left scattered. If $\sup\mathbb{T}$ is finite and left-scattered, then we define $\mathbb{T}^\kappa := \mathbb{T}\setminus \{\sup\mathbb{T}\}$, otherwise $\mathbb{T}^\kappa :=\mathbb{T}$; if $\inf\mathbb{T}$ is finite and right-scattered, then $\mathbb{T}_\kappa := \mathbb{T}\setminus\{\inf\mathbb{T}\}$, otherwise $\mathbb{T}_\kappa := \mathbb{T}$. We set $\mathbb{T}_{\kappa}^{\kappa}:=\mathbb{T}_{\kappa}\cap \mathbb{T}^{\kappa}$ and we denote $f\circ\sigma$ by $f^{\sigma}$ and $f\circ\rho$ by $f^{\rho}$.
Main Results {#ts:sec:dif}
============
In quantum calculus, the $h$-symmetric difference and the $q$-symmetric difference, $h>0$ and $0<q<1$, are defined by $$\label{eq:sd:hc}
\tilde{D}_{h}\left( t\right) =\frac{f\left( t+h\right) -f\left( t-h\right) }{2h}$$ and $$\label{eq:sd:qc}
\tilde{D}_{q}\left( t\right) =\frac{f\left( qt\right) -f\left(
q^{-1}t\right) }{\left( q-q^{-1}\right) t}, \quad t\neq 0,$$ respectively [@Kac]. Here we propose a general notion of symmetric derivative on time scales that encompasses all the three definitions , , and .
We say that a function $f:\mathbb{T} \rightarrow \mathbb{R}$ is symmetric continuous at $t\in $ $\mathbb{T}$ if, for any $\varepsilon >0$, there exists a neighborhood $U_{t} \subset \mathbb{T}$ of $t$ such that for all $s\in U_{t}$ for which $2t-s \in U_{t}$ one has $\left \vert f\left( s\right) -f\left( 2t-s\right) \right \vert \leqslant \varepsilon$.
It is easy to see that continuity implies symmetric continuity.
Let $\mathbb{T}$ be a time scale. If $f: \mathbb{T} \rightarrow \mathbb{R}$ is a continuous function, then $f$ is symmetric continuous.
Since $f$ is continuous at $t\in \mathbb{T}$, then, for any $\varepsilon >0$, there exists a neighborhood $U_{t}$ such that $\left\vert f\left( s\right) -f\left( t\right) \right\vert <\frac{\varepsilon}{2}$ and $\left\vert f\left( 2t-s\right) -f\left( t\right) \right\vert
<\frac{\varepsilon }{2}$ for all $s\in U_{t}$ for which $2t-s \in U_{t}$. Thus, $\left\vert f\left( s\right) -f\left( 2t-s\right) \right\vert
\leqslant \left\vert f\left( s\right) -f\left( t\right) \right\vert +\left\vert
f\left( t\right) -f\left( 2t-s\right) \right\vert <\varepsilon$.
The next example shows that symmetric continuity does not imply continuity.
Consider the function $f:\mathbb{R} \rightarrow \mathbb{R}$ defined by $$f\left( t\right) =
\begin{cases}
0 \, , & \text{ if } \, t\neq 0, \\
1 \, , & \text{ if } \, t=0.
\end{cases}$$ Function $f$ is symmetric continuous at $0$: for any $\varepsilon>0$, there exists a neighborhood $U_{t}$ of $t=0$ such that $\left\vert f\left( s\right) -f\left( -s\right) \right\vert =0<\varepsilon$ for all $s\in U_{t}$ for which $-s\in U_{t}$. However, $f$ is not continuous at $0$.
Let $f:\mathbb{T} \rightarrow \mathbb{R}$ and $t\in \mathbb{T}_{\kappa}^{\kappa}$. The symmetric derivative of $f$ at $t$, denoted by $f^{\diamondsuit }\left( t\right)$, is the real number (provided it exists) with the property that, for any $\varepsilon >0$, there exists a neighborhood $U \subset \mathbb{T}$ of $t$ such that $$\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) \right] -f^{\diamondsuit }\left(
t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right) \right]
\right \vert
\leqslant \varepsilon \left \vert \sigma \left( t\right)
+2t-2s-\rho \left(t\right) \right \vert$$ for all $s\in U$ for which $2t-s \in U$. A function $f$ is said to be symmetric differentiable provided $f^{\diamondsuit }\left(t\right)$ exists for all $t\in \mathbb{T}_{\kappa}^{\kappa}$.
Some useful properties of the symmetric derivative are given in Theorem \[ts:propriedade\].
\[ts:propriedade\] Let $f:\mathbb{T} \rightarrow \mathbb{R}$ and $t\in \mathbb{T}_{\kappa}^{\kappa}$. The following holds:
1. Function $f$ has at most one symmetric derivative at $t$.
2. If $f$ is symmetric differentiable at $t$, then $f$ is symmetric continuous at $t$.
3. If $f$ is continuous at $t$ and $t$ is not dense, then $f$ is symmetric differentiable at $t$ with $f^{\diamondsuit }\left( t\right)
=\frac{f^\sigma(t) -f^\rho(t)}{\sigma \left( t\right)
-\rho\left( t\right)}$.
4. If $t$ is dense, then $f$ is symmetric differentiable at $t$ if and only if the limit $\lim_{s\rightarrow t}\frac{f\left( 2t-s\right)-f\left( s\right)}{2t-2s}$ exists as a finite number. In this case $f^{\diamondsuit }\left( t\right)
=\lim_{s\rightarrow t}\frac{f\left( 2t-s\right)-f\left(s\right)}{2t-2s}
=\lim_{h\rightarrow 0}\frac{f\left( t+h\right) -f\left( t-h\right) }{2h}$.
5. If $f$ is symmetric differentiable and continuous at $t$, then $f^{\sigma }\left( t\right) =f^{\rho }\left( t\right) +f^{\diamondsuit}\left( t\right)
\left[ \sigma \left( t\right) -\rho \left( t\right) \right]$.
\(i) Suppose that $f$ has two symmetric derivatives at $t$, $f_{1}^{\diamondsuit }\left( t\right)$ and $f_{2}^{\diamondsuit }\left(t\right)$. Then, there exists a neighborhood $U_{1}$ of $t$ such that $$\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right)
+f\left(2t-s\right) -f^{\rho }\left( t\right) \right]
-f_{1}^{\diamondsuit }\left(t\right) \left[ \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right]\right \vert
\leqslant \frac{\varepsilon }{2}\left \vert \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right \vert$$ for all $s\in U_{1}$ for which $2t-s\in U_{1}$, and a neighborhood $U_{2}$ of $t$ such that $$\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) \right] -f_{2}^{\diamondsuit }\left(
t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right) \right]
\right \vert
\leqslant\frac{\varepsilon }{2}\left \vert \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right \vert$$ for all $s\in U_{2}$ for which $2t-s\in U_{2}$. Therefore, for all $s\in U_{1}\cap U_{2}$ for which $2t-s \in U_{1}\cap U_{2}$, $$\begin{split}
\bigg{|}&f_{1}^{\diamondsuit}\left( t\right) -f_{2}^{\diamondsuit }\left(
t\right) \bigg{|} =\left \vert \left[ f_{1}^{\diamondsuit }\left( t\right)
-f_{2}^{\diamondsuit }\left( t\right) \right] \frac{\sigma \left( t\right)
+2t-2s-\rho \left( t\right) }{\sigma \left( t\right) +2t-2s-\rho \left(
t\right) }\right \vert \\
&=\frac{1}{\left \vert \sigma \left( t\right) +2t-2s-\rho \left(
t\right) \right \vert } \bigg{|}\left[ f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) \right] -f_{2}^{\diamondsuit }\left(
t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right) \right]\\
&\qquad -\left[ f^{\sigma }\left( t\right) -f\left( s\right) +f\left( 2t-s\right)
-f^{\rho }\left( t\right) \right] +f_{1}^{\diamondsuit }\left( t\right)
\left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right) \right] \bigg{|}\\
&\leqslant \frac{1}{\left \vert \sigma \left( t\right) +2t-2s-\rho \left(
t\right) \right \vert } \bigg{(}\left \vert \left[ f^{\sigma }\left( t\right) -f\left(
s\right) +f\left( 2t-s\right) -f^{\rho }\left( t\right) \right]
-f_{2}^{\diamondsuit }\left( t\right) \left[ \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right] \right \vert \\
&\qquad +\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) \right] -f_{1}^{\diamondsuit }\left(
t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right) \right]
\right \vert \bigg{)} \\
&\leqslant \varepsilon.
\end{split}$$
\(ii) From hypothesis, for any $\epsilon^{*}>0$, there exists a neighborhood $U$ of $t$ such that $$\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) \right] -f^{\diamondsuit }\left(
t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right) \right]
\right \vert
\leqslant \varepsilon ^{\ast }\left \vert \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right \vert$$ for all $s\in U$ for which $2t-s \in U$. Therefore, for all $s\in U \cap \left] t-\varepsilon^{\ast },t
+\varepsilon ^{\ast }\right[$, $$\begin{split}
\vert f\left( 2t-s\right) -f\left( s\right) \vert
&\leqslant \left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right)
+f\left( 2t-s\right) -f^{\rho }\left( t\right) \right] -f^{\diamondsuit
}\left( t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right)
\right] \right \vert \\
&\qquad +\left \vert \left[ f^{\sigma }\left( t\right) -f^{\rho }\left( t\right)
\right] -f^{\diamondsuit }\left( t\right) \left[ \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right] \right \vert \\
&\leqslant \left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right)
+f\left( 2t-s\right) -f^{\rho }\left( t\right) \right] -f^{\diamondsuit
}\left( t\right) \left[ \sigma \left( t\right) +2t-2s-\rho \left( t\right)
\right] \right \vert \\
&\qquad +\left \vert \left[ f^{\sigma }\left( t\right) -f\left( t\right) +f\left(
t\right) -f^{\rho }\left( t\right) \right] -f^{\diamondsuit }\left( t\right)
\left[ \sigma \left( t\right) +2t-2t-\rho \left( t\right) \right]
\right \vert \\
&\qquad +2\left \vert f^{\diamondsuit }\left( t\right) \right \vert \left \vert
t-s\right \vert\\
&\leqslant \varepsilon ^{\ast }\left \vert \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right \vert +\varepsilon ^{\ast }\left \vert
\sigma \left( t\right) +2t-2t-\rho \left( t\right) \right \vert +2\left \vert
f^{\diamondsuit }\left( t\right) \right \vert \left \vert t-s\right \vert \\
&\leqslant \varepsilon ^{\ast }\left \vert \sigma \left( t\right) -\rho
\left( t\right) \right \vert +2\varepsilon ^{\ast }\left \vert t-s\right \vert
+\varepsilon ^{\ast }\left \vert \sigma \left( t\right) -\rho \left( t\right)
\right \vert +2\left \vert f^{\diamondsuit }\left( t\right) \right \vert
\left \vert t-s\right \vert \\
&= 2\varepsilon ^{\ast }\left \vert \sigma \left( t\right) -\rho
\left( t\right) \right \vert +2\left( \varepsilon ^{\ast }+\left \vert
f^{\diamondsuit }\left( t\right) \right \vert \right) \left \vert
t-s\right \vert \\
&\leqslant 2\varepsilon ^{\ast }\left \vert \sigma \left( t\right) -\rho \left(
t\right) \right \vert +2\left( \varepsilon ^{\ast }+\left \vert
f^{\diamondsuit }\left( t\right) \right \vert \right) \varepsilon ^{\ast } \\
&= 2\varepsilon ^{\ast }\left[ \left \vert \sigma \left( t\right)
-\rho \left( t\right) \right \vert +\varepsilon ^{\ast }+\left \vert
f^{\diamondsuit }\left( t\right) \right \vert \right],
\end{split}$$ proving that $f$ is symmetric continuous at $t$.
\(iii) Suppose that $t\in\mathbb{T}_{\kappa}^{\kappa}$ is not dense and $f$ is continuous at $t$. Then, $$\lim_{s\rightarrow t}\frac{f^{\sigma }\left( t\right) -f\left( s\right)
+f\left( 2t-s\right) -f^{\rho }\left( t\right) }{\sigma \left( t\right)
+2t-2s-\rho \left( t\right) }=\frac{f^{\sigma }\left( t\right)
-f^{\rho}\left( t\right) }{\sigma \left( t\right) -\rho \left( t\right)}.$$ Hence, for any $\varepsilon >0$, there exists a neighborhood $U$ of $t$ such that $$\left \vert \frac{f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) }{\sigma \left( t\right) +2t-2s-\rho
\left( t\right) }-\frac{f^{\sigma }\left( t\right)
-f^{\rho }\left( t\right)}{\sigma \left( t\right)
-\rho \left( t\right) }\right \vert \leqslant \varepsilon$$ for all $s\in U$ for which $2t-s \in U$. It follows that $$\begin{gathered}
\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right)
+f\left(2t-s\right) -f^{\rho }\left( t\right) \right]
-\frac{f^{\sigma }\left(t\right) -f^{\rho }\left( t\right)}{\sigma \left( t\right)
-\rho \left(t\right) }\left[ \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right]\right\vert\\
\leqslant \varepsilon \left[ \sigma \left( t\right)
+2t-2s-\rho \left(t\right) \right],\end{gathered}$$ which proves that $f^{\diamondsuit }\left( t\right) =\left(f^{\sigma }\left( t\right)
-f^{\rho}\left( t\right)\right)/\left(\sigma \left( t\right) -\rho \left( t\right)\right)$.
\(iv) Assume that $f$ is symmetric differentiable at $t$ and $t$ is dense. Let $\varepsilon >0$ be given. Then, there exists a neighborhood $U$ of $t$ such that $$\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right)
+f\left(2t-s\right) -f^{\rho }\left( t\right) \right]
-f^{\diamondsuit }\left(t\right) \left[ \sigma \left( t\right)
+2t-2s-\rho \left( t\right) \right]\right \vert
\leqslant \varepsilon \left \vert \sigma \left( t\right)
+2t-2s-\rho \left(t\right) \right \vert$$ for all $s\in U$ for which $2t-s \in U$. Since $t$ is dense, $\left \vert \left[ -f\left( s\right) +f\left( 2t-s\right) \right]
-f^{\diamondsuit }\left( t\right) \left[ 2t-2s\right] \right \vert \leqslant
\varepsilon \left \vert 2t-2s\right \vert$ for all $s\in U$ for which $2t-s \in U$. It follows that $\left \vert \frac{f\left( 2t-s\right)-f\left( s\right)}{2t-2s}
-f^{\diamondsuit }\left( t\right) \right \vert \leqslant \varepsilon$ for all $s\in U$ with $s\neq t$. Therefore, we get the desired result: $f^{\diamondsuit }\left( t\right)
=\lim_{s\rightarrow t}\frac{f\left( 2t-s\right)-f\left(s\right)}{2t-2s}$. Conversely, let us suppose that $t$ is dense and the limit $\lim_{s\rightarrow t}\frac{f\left( 2t-s\right)-f\left( s\right)}{2t-2s} =: L$ exists. Then, there exists a neighborhood $U$ of $t$ such that $\left \vert \frac{f\left( 2t-s\right)-f\left( s\right)}{2t-2s}
-L\right \vert \leqslant \varepsilon$ for all $s\in U$ for which $2t-s \in U$. Because $t$ is dense, we have $\left \vert \frac{f^{\sigma }\left( t\right) -f\left( s\right) +f\left(
2t-s\right) -f^{\rho }\left( t\right) }{\sigma \left( t\right) +2t-2s-\rho
\left( t\right) }-L\right \vert \leqslant \varepsilon$. Therefore, $$\left \vert \left[ f^{\sigma }\left( t\right) -f\left( s\right)
+f\left( 2t-s\right) -f^{\rho }\left( t\right) \right] -L\left[ \sigma
\left( t\right) +2t-2s-\rho \left( t\right) \right] \right \vert
\leqslant \varepsilon \left \vert \sigma \left( t\right) +2t-2s-\rho \left(
t\right) \right \vert,$$ which leads us to the conclusion that $f$ is symmetric differentiable and $f^{\diamondsuit }\left( t\right) = L$. Note that if we use the substitution $s=t+h$, then $f^{\diamondsuit }\left( t\right)
=\lim_{h\rightarrow 0}\frac{f\left( t+h\right) -f\left( t-h\right) }{2h}$.
\(v) If $t$ is a dense point, then $\sigma \left( t\right) =\rho \left(t\right) $ and $f^{\sigma }\left( t\right) =f^{\rho }\left( t\right)
+f^{\diamondsuit }\left( t\right) \left[ \sigma \left( t\right)
-\rho \left(t\right) \right]$. If $t$ is not dense, and since $f$ is continuous, then $f^{\diamondsuit }\left( t\right) =\frac{f^{\sigma }\left( t\right)
-f^{\rho }\left( t\right) }{\sigma \left( t\right) -\rho \left( t\right) }
\Leftrightarrow f^{\sigma }\left( t\right) =f^{\rho }\left( t\right)
+f^{\diamondsuit }\left( t\right) \left[ \sigma \left( t\right) -\rho \left(
t\right) \right]$.
Let $\mathbb{T} = \mathbb{R}$. Then our symmetric derivative coincides with the classic symmetric derivative : $f^{\diamondsuit} = f^{s}$.
Let $\mathbb{T} = h \mathbb{Z}$, $h>0$. Then the symmetric derivative is the symmetric difference operator : $f^{\diamondsuit} = \tilde{D}_{h}$.
Let $\mathbb{T=}\overline{q^{\mathbb{Z}}}$, $0<q<1$. Then the symmetric derivative coincides with the $q$-symmetric difference operator : $f^{\diamondsuit } = \tilde{D}_{q}$.
Independently of the time scale $\mathbb{T}$, the symmetric derivative of a constant is zero and the symmetric derivative of the identity function is one.
An alternative way to define the symmetric derivative of $f$ at $t\in\mathbb{T}_{\kappa}^{\kappa}$ consists in saying that the limit $f^{\diamondsuit }\left( t\right) =\lim_{s\rightarrow t}\frac{f^{\sigma
}\left( t\right) -f\left( s\right) +f\left( 2t-s\right) -f^{\rho }\left(
t\right) }{\sigma \left( t\right) +2t-2s-\rho \left( t\right) }
=\lim_{h\rightarrow 0}\frac{f^{\sigma }\left( t\right) -f\left( t+h\right)
+f\left( t-h\right) -f^{\rho }\left( t\right) }{\sigma \left( t\right)
-2h-\rho \left( t\right)}$ exists.
Let $f,g:\mathbb{T} \rightarrow \mathbb{R}$ be two symmetric differentiable functions at $t\in \mathbb{T}_{\kappa}^{\kappa}$ and $\lambda \in \mathbb{R}$. The following holds:
1. Function $f+g$ is symmetric differentiable at $t$ with $\left( f+g\right) ^{\diamondsuit }\left( t\right)
=f^{\diamondsuit }\left(t\right) +g^{\diamondsuit }\left( t\right)$.
2. Function $\lambda f$ is symmetric differentiable at $t$ with $\left( \lambda f\right)^{\diamondsuit }\left( t\right)
=\lambda f^{\diamondsuit }\left( t\right)$.
3. If $f$ and $g$ are continuous at $t$, then $fg$ is symmetric differentiable at $t$ with $\left(fg\right)^{\diamondsuit }\left( t\right)
=f^{\diamondsuit }\left(t\right) g^{\sigma }\left( t\right)
+f^{\rho }\left( t\right) g^{\diamondsuit }\left( t\right)$.
4. If $f$ is continuous at $t$ and $f^{\sigma }\left( t\right) f^{\rho}\left( t\right) \neq 0$, then $1/f$ is symmetric differentiable at $t$ with $\left(\frac{1}{f}\right)^{\diamondsuit }\left( t\right)
=-\frac{f^{\diamondsuit }\left( t\right) }{f^{\sigma }\left(
t\right) f^{\rho}\left( t\right)}$.
5. If $f$ and $g$ are continuous at $t$ and $g^{\sigma }\left( t\right) g^{\rho }\left( t\right) \neq 0$, then $f/g$ is symmetric differentiable at $t$ with $\left( \frac{f}{g}\right) ^{\diamondsuit }\left( t\right)
=\frac{f^{\diamondsuit }\left( t\right) g^{\rho }\left( t\right) -f^{\rho }\left(
t\right) g^{\diamondsuit }\left( t\right)}{g^{\sigma }\left( t\right)
g^{\rho }\left( t\right)}$.
\(i) For $t\in \mathbb{T}_{\kappa}^{\kappa}$ we have $$\begin{split}
\left( f+g\right)^{\diamondsuit }\left( t\right)
&=\lim_{s\rightarrow t}\frac{\left( f+g\right) ^{\sigma }\left( t\right)
-\left( f+g\right) \left( s\right) +\left( f+g\right) \left( 2t-s\right)
-\left( f+g\right) ^{\rho }\left( t\right) }{\sigma \left( t\right)
+2t-2s-\rho \left( t\right) } \\
&=\lim_{s\rightarrow t}\frac{f^{\sigma }\left( t\right) -f\left( s\right)
+f\left( 2t-s\right) -f^{\rho }\left( t\right) }{\sigma \left( t\right)
+2t-2s-\rho \left( t\right)}
+ \lim_{s\rightarrow t}\frac{g^{\sigma }\left( t\right)
-g\left( s\right) +g\left( 2t-s\right) -g^{\rho }\left( t\right) }{\sigma
\left( t\right) +2t-2s-\rho \left( t\right) } \\
&= f^{\diamondsuit }\left( t\right) +g^{\diamondsuit }\left( t\right).
\end{split}$$
\(ii) Let $t\in \mathbb{T}_{\kappa}^{\kappa}$ and $\lambda \in \mathbb{R}$. Then, $$\begin{split}
\left( \lambda f\right) ^{\diamondsuit }\left( t\right)
&=\lim_{s\rightarrow t}\frac{\left( \lambda f\right) ^{\sigma }\left(
t\right) -\left( \lambda f\right) \left( s\right) +\left( \lambda f\right)
\left( 2t-s\right) -\left( \lambda f\right) ^{\rho }\left( t\right) }{\sigma
\left( t\right) +2t-2s-\rho \left( t\right) } \\
&=\lambda \lim_{s\rightarrow t}\frac{f^{\sigma }\left( t\right) -f\left(
s\right) +f\left( 2t-s\right) -f^{\rho }\left( t\right) }{\sigma \left(
t\right) +2t-2s-\rho \left( t\right)}
=\lambda f^{\diamondsuit }\left( t\right) .
\end{split}$$
\(iii) Let us assume that $t\in \mathbb{T}_{\kappa}^{\kappa}$ and $f$ and $g$ are continuous at $t$. If $t$ is dense, then $$\begin{split}
\left( fg\right) ^{\diamondsuit }\left( t\right)
&=\lim_{h\rightarrow 0}\frac{\left( fg\right) \left( t+h\right) -\left(
fg\right) \left( t-h\right) }{2h} \\
&=\lim_{h\rightarrow 0}\frac{f\left( t+h\right) -f\left( t-h\right) }{2h}
g\left( t+h\right) +\lim_{h\rightarrow 0}\frac{g\left( t+h\right) -g\left(
t-h\right) }{2h}f\left( t-h\right) \\
&= f^{\diamondsuit }\left( t\right) g^{\sigma }\left( t\right)
+f^{\rho}\left( t\right) g^{\diamondsuit }\left( t\right) .
\end{split}$$ If $t$ is not dense, then $$\begin{split}
\left( fg\right) ^{\diamondsuit }\left( t\right)
&= \frac{\left( fg\right)^{\sigma }\left( t\right)
-\left( fg\right) ^{\rho }\left( t\right) }{\sigma\left( t\right)
-\rho \left( t\right) }
=\frac{f^{\sigma }\left( t\right) -f^{\rho }\left( t\right) }{\sigma
\left( t\right) -\rho \left( t\right) }g^{\sigma }\left( t\right)
+\frac{g^{\sigma }\left( t\right) -g^{\rho }\left( t\right) }{\sigma \left(
t\right) -\rho \left( t\right) }f^{\rho }\left( t\right) \\
&=f^{\diamondsuit }\left( t\right) g^{\sigma }\left( t\right)
+f^{\rho}\left( t\right) g^{\diamondsuit }\left( t\right)
\end{split}$$ proving the intended equality.
\(iv) Using the relation $\left( \frac{1}{f}\times f\right) \left( t\right) =1$ we can write that $0 =\left( \frac{1}{f}\times f\right) ^{\diamondsuit }\left( t\right)
= f^{\diamondsuit }\left( t\right) \left(
\frac{1}{f}\right)^{\sigma}\left( t\right)
+ f^{\rho }\left( t\right) \left(
\frac{1}{f}\right)^{\diamondsuit }\left( t\right)$. Therefore, $\left( \frac{1}{f}\right) ^{\diamondsuit }\left( t\right)
=-\frac{f^{\diamondsuit }\left( t\right)}{
f^{\sigma }\left( t\right) f^{\rho}\left( t\right) }$.
\(v) Let $t\in \mathbb{T}_{\kappa}^{\kappa}$. Then, $$\begin{split}
\left( \frac{f}{g}\right)^{\diamondsuit }\left( t\right)
&= \left( f\times \frac{1}{g}\right) ^{\diamondsuit }\left( t\right)
= f^{\diamondsuit }\left( t\right) \left( \frac{1}{g}\right)^{\sigma}\left(
t\right) +f^{\rho }\left( t\right) \left( \frac{1}{g}\right)^{\diamondsuit }\left( t\right) \\
&= \frac{f^{\diamondsuit }\left( t\right) }{g^{\sigma }\left( t\right) }
+f^{\rho }\left( t\right) \left(
-\frac{g^{\diamondsuit }\left( t\right)}{g^{\sigma }\left( t\right) g^{\rho }\left( t\right) }\right)
= \frac{f^{\diamondsuit }\left( t\right) g^{\rho }\left( t\right)
-f^{\rho}\left( t\right) g^{\diamondsuit }\left( t\right) }{g^{\sigma }\left(
t\right) g^{\rho }\left( t\right)}.
\end{split}$$
The symmetric derivative of $f\left( t\right) =t^{2}$ is $f^{\diamondsuit }\left( t\right) =\sigma \left( t\right) +\rho \left(t\right)$.
The symmetric derivative of $f\left( t\right) =1/t$ is $f^{\diamondsuit }\left( t\right)
=-\frac{1}{\sigma \left( t\right) \rho\left( t\right) }$.
Particular Cases {#ts:sec:fc}
================
In Section \[ts:sec:dif\] we introduced the symmetric derivative on a time scale $\mathbb{T}$ and derived some of its properties. It has been shown that the new notion unifies the symmetric derivatives of classical analysis [@Thomson] and quantum calculus [@Kac]. Here we note that our symmetric derivative is different from the delta and nabla derivatives considered in the time scale literature [@Hilger; @Bohner:1; @Bohner:2]. A simple example of a function that is neither delta nor nabla differentiable, in the sense of time scales, but that is symmetric differentiable, is the absolute value function.
\[ex:dos:nts\] Let $\mathbb{T}$ be a time scale with $0\in \mathbb{T}_{\kappa}^{\kappa}$ and $f:\mathbb{T} \rightarrow \mathbb{R}$ be defined by $f\left( t\right) =\left \vert t\right \vert$. This function is not differentiable at point $t = 0$ in the sense of time scales [@Hilger; @Bohner:1; @Bohner:2; @Sheng]. However, the symmetric derivative is always well defined: $$f^{\diamondsuit }\left( 0\right)
=\lim_{h\rightarrow 0}\frac{f^{\sigma}\left( 0\right)
-f\left( 0+h\right) +f\left( 0-h\right) -f^{\rho }\left(
0\right) }{\sigma \left( 0\right) -2h-\rho \left( 0\right) }
= \lim_{h\rightarrow 0}\frac{\sigma \left( 0\right)
+\rho \left( 0\right) }{\sigma \left( 0\right) -2h-\rho \left( 0\right)},$$ so that $f^{\diamondsuit}(0) = 0$ if $0$ is dense, and $f^{\diamondsuit}(0) = (\sigma(0)+\rho(0))/(\sigma(0)-\rho(0))$ otherwise.
In the particular case a function is simultaneously delta and nabla differentiable [@Bohner:1; @Bohner:2], Proposition \[ts:delta nabla\] below shows that a relation can be done between our symmetric derivative and the recent diamond-alpha derivative [@Sheng] (Corollary \[cor:diam\]).
\[ts:delta nabla\] If $f$ is delta and nabla differentiable, then $f$ is symmetric differentiable and, for each $t\in\mathbb{T}_{\kappa}^{\kappa}$, $f^{\diamondsuit }\left( t\right) =\gamma\left( t\right)
f^{\Delta}\left( t\right) +\left(1-\gamma\left( t\right)\right) f^{\nabla }\left( t\right)$, where $$\label{eq:gamma}
\gamma\left( t\right) =\lim_{s\rightarrow t}\frac{\sigma \left(
t\right) -s}{\sigma \left( t\right) +2t-2s-\rho \left( t\right) }.$$
Note that $$\begin{split}
f^{\diamondsuit }\left( t\right)
&=\lim_{s\rightarrow t}\frac{f^{\sigma}\left( t\right)
-f\left( s\right) +f\left( 2t-s\right) -f^{\rho }\left(
t\right) }{\sigma \left( t\right) +2t-2s-\rho \left( t\right) } \\
&= \lim_{s\rightarrow t}\bigg{(}\frac{\sigma \left( t\right) -s}{\sigma
\left( t\right) +2t-2s-\rho \left( t\right) }\frac{f^{\sigma }\left(
t\right) -f\left( s\right) }{\sigma \left( t\right) -s}
+\frac{\left( 2t-s\right) -\rho \left( t\right) }{\sigma \left( t\right)
+2t-2s-\rho \left( t\right) }\frac{f\left( 2t-s\right) -f^{\rho }\left(
t\right) }{\left( 2t-s\right) -\rho \left( t\right) }\bigg{)} \\
&= \lim_{s\rightarrow t}\left( \frac{\sigma \left( t\right) -s}{\sigma
\left( t\right) +2t-2s-\rho \left( t\right) }f^{\Delta }\left( t\right)
+\frac{\left( 2t-s\right) -\rho \left( t\right) }{\sigma \left( t\right)
+2t-2s-\rho \left( t\right) }f^{\nabla }\left( t\right) \right) .
\end{split}$$ For each $t\in\mathbb{T}$, define $\gamma\left( t\right)
:= \lim_{s\rightarrow t}\frac{\sigma \left(
t\right) -s}{\sigma \left( t\right) +2t-2s-\rho \left( t\right) }$ and $\tilde{\gamma}\left( t\right)
:= \lim_{s\rightarrow t}\frac{\left(
2t-s\right) -\rho \left( t\right) }{\sigma \left( t\right) +2t-2s-\rho
\left( t\right)}$. It is clear that $\gamma\left( t\right) +\tilde{\gamma}\left( t\right) =1$. Note that if $t\in \mathbb{T}$ is dense, then $$\gamma\left( t\right) = \lim_{s\rightarrow t}\frac{\sigma \left(
t\right) -s}{\sigma \left( t\right) +2t-2s-\rho \left( t\right) }
=\lim_{s\rightarrow t}\frac{t-s}{2t-2s}
=\frac{1}{2}$$ and, therefore, $\tilde{\gamma}\left( t\right) =1/2$. On the other hand, if $t\in \mathbb{T}$ is not dense, then $$\gamma\left( t\right) = \lim_{s\rightarrow t}\frac{\sigma \left(
t\right) -s}{\sigma \left( t\right) +2t-2s-\rho \left( t\right)}
= \frac{\sigma \left( t\right) -t}{\sigma \left( t\right)
-\rho \left(t\right)}$$ and $\tilde{\gamma}\left( t\right) =\frac{t-\rho \left( t\right) }{\sigma \left(
t\right) -\rho \left( t\right) }$. Hence, functions $\gamma, \tilde{\gamma} :\mathbb{T}\rightarrow \mathbb{R}$ are well defined and, if $f$ is delta and nabla differentiable, then $f^{\diamondsuit }\left( t\right) = \gamma\left( t\right)
f^{\Delta}\left( t\right) +\tilde{\gamma}\left( t\right) f^{\nabla }\left( t\right)
= \gamma\left( t\right) f^{\Delta }\left( t\right)
+\left( 1-\gamma\left( t\right) \right) f^{\nabla }\left( t\right)$.
Functions $\gamma, \tilde{\gamma} : \mathbb{T}\rightarrow \mathbb{R}$ are bounded and nonnegative: $0 \leqslant \gamma(t), \tilde{\gamma}(t) \leqslant 1$. This is due to the fact that $\rho \left( t\right)\leqslant t \leqslant\sigma \left( t\right)$ for every $t\in \mathbb{T}$.
\[cor:diam\] If $f$ is delta and nabla differentiable and if function $\gamma(\cdot)$ in is a constant, $\gamma(t) \equiv \alpha$, then the symmetric derivative coincides with the diamond-$\alpha$ derivative: $f^{\diamondsuit}(t) = \alpha f^{\Delta}\left( t\right)
+(1-\alpha) f^{\nabla}(t)$.
In the classical case $\mathbb{T} = \mathbb{R}$ it can be proved that “A continuous function is necessarily increasing in any interval in which its symmetric derivative exists and is positive” [@Thomson]. We note that this result is not valid for the symmetric derivative on time scales. For instance, consider the time scale $\mathbb{T}=\mathbb{N}$ and function $f(n) = n$ if $n$ is odd and $f(n) = 10 n$ if $n$ is even. The symmetric derivative of $f$ is given by $f^{\diamondsuit }\left( n\right) = \frac{f^{\sigma }\left( n\right)
-f^{\rho }\left( n\right) }{\sigma \left( n\right) -\rho \left( n\right) }
=\frac{10\left( n+1\right) -10\left( n-1\right) }{\left( n+1\right)
-\left( n-1\right)} = 10$ for $n$ odd, while for $n$ even one has $f^{\diamondsuit }\left( n\right) = \frac{f^{\sigma }\left( n\right)
-f^{\rho }\left( n\right) }{\sigma \left( n\right) -\rho \left( n\right)}
=\frac{\left( n+1\right) -\left( n-1\right) }{\left( n+1\right)
-\left(n-1\right)} = 1$. Clearly, function $f$ is non-increasing although its symmetric derivative is always positive. In this example, the symmetric derivative coincides with the diamond-$\alpha$ derivative, $\alpha=1/2$, showing that there is an inconsistency in Corollary 2.1 of [@Ozkan].
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by [*FEDER*]{} funds through [*COMPETE*]{} — Operational Programme Factors of Competitiveness (“Programa Operacional Factores de Competitividade”) and by Portuguese funds through the [*Center for Research and Development in Mathematics and Applications*]{} (University of Aveiro) and the Portuguese Foundation for Science and Technology (“FCT–Fundação para a Ciência e a Tecnologia”), within project PEst-C/MAT/UI4106/2011 with COMPETE number FCOMP-01-0124-FEDER-022690. Brito da Cruz is also supported by FCT through the Ph.D. fellowship SFRH/BD/33634/2009. The authors are grateful to the referees for their valuable comments and helpful suggestions.
[9]{}
B. Aulbach and S. Hilger, A unified approach to continuous and discrete dynamics, in [*Qualitative theory of differential equations (Szeged, 1988)*]{}, 37–56, Colloq. Math. Soc. János Bolyai, 53 North-Holland, Amsterdam, 1990.
M. Bohner and A. Peterson, [*Dynamic equations on time scales*]{}, Birkhäuser Boston, Boston, MA, 2001.
M. Bohner and A. Peterson, [*Advances in dynamic equations on time scales*]{}, Birkhäuser Boston, Boston, MA, 2003.
V. Kac and P. Cheung, [*Quantum calculus*]{}, Universitext, Springer, New York, 2002.
U. M. Özkan and B. Kaymakçalan, Basics of diamond-$\alpha$ partial dynamic calculus on time scales, Math. Comput. Modelling [**50**]{} (2009), no. 9-10, 1253–1261.
R. A. Serafin, On the symmetric difference quotient and its application to the correction of orbits, Celestial Mech. [**26**]{} (1982), no. 4, 383–393.
Q. Sheng, M. Fadag, J. Henderson and J. M. Davis, An exploration of combined dynamic derivatives on time scales and their applications, Nonlinear Anal. Real World Appl. [**7**]{} (2006), no. 3, 395–413.
B. S. Thomson, [*Symmetric properties of real functions*]{}, Monographs and Textbooks in Pure and Applied Mathematics, 183, Dekker, New York, 1994.
A. Zygmund, [*Trigonometric series. Vol. I, II*]{}, reprint of the 1979 edition, Cambridge Mathematical Library, Cambridge Univ. Press, Cambridge, 1988.
[^1]: This is a preprint of a paper whose final and definite form will be published in *Applied Mathematics Letters*. Submitted 30-Jul-2012; revised 07-Sept-2012; accepted 10-Sept-2012.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'X-ray echo spectroscopy, a counterpart of neutron spin-echo, is being introduced here to overcome limitations in spectral resolution and weak signals of the traditional inelastic x-ray scattering (IXS) probes. An image of a point-like x-ray source is defocused by a dispersing system comprised of asymmetrically cut specially arranged Bragg diffracting crystals. The defocused image is refocused into a point (echo) in a time-reversal dispersing system. If the defocused beam is inelastically scattered from a sample, the echo signal acquires a spatial distribution, which is a map of the inelastic scattering spectrum. The spectral resolution of the echo spectroscopy does not rely on the monochromaticity of the x-rays, ensuring strong signals along with a very high spectral resolution. Particular schemes of x-ray echo spectrometers for 0.1–0.02-meV ultra-high-resolution IXS applications (resolving power $> 10^8$) with broadband $\simeq$ 5–13 meV dispersing systems are introduced featuring more than $10^3$ signal enhancement. The technique is general, applicable in different photon frequency domains.'
author:
- 'Yuri Shvyd’ko'
title: 'X-ray echo spectroscopy'
---
Introduction
============
The spectroscopic signal strength decreases abruptly with improving spectral resolution. An x-ray echo spectroscopy, introduced here, offers a potential for achieving much higher, yet unattainable, spectral resolution in the hard x-ray regime without compromising the signal strength.
The origin of the proposed technique is in spin echo, a phenomenon discovered by Erwin Hahn in 1950 [@Hahn50]. Spin echo is the refocusing in the time domain of the defocused spin magnetization by time reversal. The spin echo technique is widely used in NMR. There is a photon echo analog in the optics. Neutron spin echo spectroscopy is an inelastic neutron scattering technique invented by Mezei, which uses time reversal of the neutron spin evolution to measure energy loss in an inelastic neutron scattering process [@Mezei80]. Photon polarization precession spectroscopy was proposed recently by Röhlsberger [@Roehle14] for studies of spin waves that exhibit similarities to the neutron spin echo. Fung et al. proposed a space-domain analog of the echo spectroscopy for resonant inelastic soft x-ray scattering applications [@Fung04]. Defocusing and refocusing of the spectral components is achieved by angular dispersion from curved diffraction gratings. This approach has been recently demonstrated by Lai et al. [@Lai14].
Here, we propose a hard x-ray version of the echo spectroscopy, which can be applied for non-resonant and resonant high-resolution inelastic x-ray scattering applications. Diffraction gratings are not practical in the hard x-ray regime. However, as was demonstrated in [@Shvydko-SB; @SLK06], the angular dispersion in the hard x-ray regime can be achieved by Bragg diffraction from asymmetrically cut crystals or from special arrangements of asymmetrically cut crystals [@SSM13; @Shvydko15], which are a hard x-ray analog of the optical diffraction gratings and optical prisms. In the space-domain echo-spectrometer proposed here, an image of a point-like x-ray source is defocused by a dispersing system comprised of asymmetrically cut Bragg diffracting crystals. The defocused image is refocused into a point (echo) in a time-reversal dispersing system. We show, if the defocused beam is inelastically scattered from a sample, the echo signal acquires a spatial distribution, which is a map of the energy transfer spectrum in the scattering process. The spectral resolution of the echo spectroscopy does not rely on the monochromaticity of the x-rays, thus ensuring strong signals along with a very high spectral resolution.
In the present paper, we use an analytical ray-transfer matrix approach and the dynamical theory of x-ray diffraction in crystals to calculate and analyze the performance of a generic echo spectrometer comprised of defocusing and refocusing dispersing elements, to derive conditions for refocusing and expressions for the spectral resolution of the echo spectrometer. Specific designs of the hard x-ray echo spectrometers are introduced with a spectral resolution $\Delta
{\varepsilon}=0.1-0.02$ meV at photon energies $E\simeq 9.1$ keV and $E\simeq 4.6$ keV, comprised of defocusing and refocusing systems with multi-crystal inline dispersing elements featuring both large cumulative dispersion rates ${{{\mathcal D}}_{\cup_{{_{}}}}} \gtrsim 25-60~\mu$rad/meV, transmission bandwidths $\Delta E_{{_{\cup}}}\simeq 5-13$ meV, and a dynamical range $\Delta E_{{_{\cup}}}/\Delta {\varepsilon}\simeq
100-500$. Because of much greater dispersion rates which are feasible in the crystal systems, as compared to the diffraction gratings, the spectral resolving power in the hard x-ray regime can be as large as $\gtrsim 10^8-10^9$, i.e., more than three orders of magnitude higher than in the soft x-ray regime.
Theory
======
We start here by considering optical systems featuring a combination of focusing and energy dispersing capabilities. We assume that such systems can, first, focus [*monochromatic*]{} x-rays from a source of a linear size $\Delta x_{{_{0}}}$ in a source plane (reference plane $0$ perpendicular to the optical axis $z$ in Fig. \[fig001\]) onto an intermediate image plane (reference plane $1$ in Fig. \[fig001\]) with an image linear size $\Delta x_{{_{1}}}= |A| \Delta
x_{{_{0}}}$, where $A$ is a magnification factor of the optical system. In addition, the system can disperse photons in such a way that the location of the source image for photons with an energy $E+\delta E$ is displaced in the image plane by $G \delta E$ from the location of the image for photons with energy $E$. Here, $G$ is a linear dispersion rate of the system. As a result, although monochromatic x-rays are focused, the whole spectrum of x-rays is defocused, due to linear dispersion.
We will use the ray-transfer matrix technique [@KL66; @MK80-1; @Siegman] to propagate paraxial x-rays through such optical systems and to determine linear and angular sizes of the x-ray beam along the optical axis. A paraxial ray in any reference plane is characterized by its distance $x$ from the optical axis, by its angle $\xi$ with respect to that axis, and the deviation $\delta E$ of the photon energy from a nominal value $E$. The ray vector ${\mbox{\boldmath $r$}}_{{_{0}}}=(x_{{_{0}}},\xi_{{_{0}}},\delta E)$ at an input source plane is transformed to ${\mbox{\boldmath $r$}}_{{_{1}}}=(x_{{_{1}}},\xi_{{_{1}}},\delta
E)=\hat{O}{\mbox{\boldmath $r$}}_{{_{0}}}$ at the output reference plane (image plane), where $\hat{O}=\{ABG;CDF;001\}$ is a ray-transfer matrix of an optical element placed between the planes [^1]. Only elastic processes in the optical systems are taken into account, that is reflected by zero and unity terms in the lowest row of the ray-transfer matrices.
![Graphical presentation of the echo spectroscopy principles. (a) Photons from a source with a linear size $\Delta
x_{{_{\mathrm {0}}}}$ in the reference source plane $0$ are focused onto a spot $\Delta x_{{_{\mathrm {1}}}}$ in the intermediate image plane $1$ by a focusing-dispersing system $\hat{O}_{{_{\mathrm {{D}}}}}$. Each spectral component, indicated by different color, is focused at a different location due to dispersion in $\hat{O}_{{_{\mathrm {{D}}}}}$. All spectral components of the x-rays are refocused by a consecutive time-reversal focusing-dispersing system $\hat{O}_{{_{\mathrm {{R}}}}}$ onto the same spot $\Delta x_{{_{\mathrm {2}}}}$ (echo) in the image plane $2$. (b) Inelastic x-ray scattering with an energy transfer ${\varepsilon}$ (indicated by changed color) from a sample in the reference plane $1$ results in a lateral shift $G_{{_{\mathrm {{R}}}}} {\varepsilon}$ of the echo signal equal for all spectral components. Here $G_{{_{\mathrm {{R}}}}}$ is a linear dispersion rate in the time-reversal system $\hat{O}_{{_{\mathrm {{R}}}}}$. Spectral resolution of the x-ray echo spectrometer is $\Delta {\varepsilon}=\Delta
x_{{_{\mathrm {2}}}}/G_{{_{\mathrm {{R}}}}}$. []{data-label="fig001"}](1.pdf){width="50.00000%"}
Focusing of the monochromatic spectral components requires that matrix element $B=0$. The ray-transfer matrix of any focusing-dispersing system in a general case therefore reads as $$\hat{O} =
\{A \,0\,G;\,CDF;001\}
\label{matrix}$$ with $A$ and $G$ elements defined above. The system blurs the polychromatic source image, because of linear dispersion, as mentioned earlier and graphically presented in Fig. \[fig001\](a). However, another focusing-dispersing system can be used to refocus the source onto reference plane 2. Indeed, propagation of x-rays through the defocusing system $\hat{O}_{{_{\mathrm {{D}}}}}$ and a second system, which we will refer to as a refocusing or time-reversal system $\hat{O}_{{_{\mathrm {{R}}}}}$ (see Fig. \[fig001\]) is given by a combined ray-transfer matrix $$\begin{gathered}
\hat{O}_{{_{\mathrm {{C}}}}} = \hat{O}_{{_{\mathrm {{R}}}}} \hat{O}_{{_{\mathrm {{D}}}}} = \{A_{{_{\mathrm {{C}}}}}\,0\,G_{{_{\mathrm {{C}}}}};C_{{_{\mathrm {{C}}}}}D_{{_{\mathrm {{C}}}}}F_{{_{\mathrm {{C}}}}};001\}\\ = \left(\!\! \begin{array}{ccc} A_{{_{\mathrm {{R}}}}} A_{{_{\mathrm {{D}}}}} & 0 & A_{{_{\mathrm {{R}}}}}G_{{_{\mathrm {{D}}}}}+G_{{_{\mathrm {{R}}}}} \\ C_{{_{\mathrm {{R}}}}}A_{{_{\mathrm {{D}}}}}\!+\!D_{{_{\mathrm {{R}}}}}C_{{_{\mathrm {{R}}}}} & D_{{_{\mathrm {{R}}}}} D_{{_{\mathrm {{D}}}}} & C_{{_{\mathrm {{R}}}}}G_{{_{\mathrm {{D}}}}}\!+\!D_{{_{\mathrm {{R}}}}}F_{{_{\mathrm {{D}}}}}\!+\!F_{{_{\mathrm {{R}}}}} \\ 0 & 0 & 1 \end{array}\!\! \right),
\label{comb}\end{gathered}$$ and by a ray vector ${\mbox{\boldmath $r$}}_{{_{2}}}=(x_{{_{2}}},\xi_{{_{2}}},\delta
E)=\hat{O}_{{_{\mathrm {{C}}}}}{\mbox{\boldmath $r$}}_{{_{0}}}$.
Here we arrive at a crucial point. If $$G_{{_{\mathrm {{C}}}}}=A_{{_{\mathrm {{R}}}}}G_{{_{\mathrm {{D}}}}}+G_{{_{\mathrm {{R}}}}}=0,
\label{refocus}$$ the linear dispersion at the exit of the combined system vanishes, because dispersion in the defocusing system is compensated (time reversed) by dispersion in the refocusing system. As a result, the combined system refocuses all photons independent of the photon energy to the same location, $x_{{_{2}}}$ in image plane $2$, to a spot with a linear size $$\Delta x_{{_{2}}}= |A_{{_{\mathrm {{R}}}}} A_{{_{\mathrm {{D}}}}}| \Delta x_{{_{0}}}\equiv |A_{{_{\mathrm {{R}}}}} | \Delta x_{{_{1}}},
\label{magnification}$$ as shown schematically in Fig. \[fig001\](a). Such behavior is an analog of the echo phenomena. Here, however, it takes place in space, rather than in the time domain.
Now, what happens if a sample is placed into the intermediate image plane $1$, \[Fig. \[fig001\](b)\], which can scatter photons inelastically? In an inelastic scattering process, a photon with an arbitrary energy $E+\delta E$, changes its value to $E+\delta E+
{\varepsilon}$. Here ${\varepsilon}$ is an energy transfer in the inelastic scattering process. The ray vector ${\mbox{\boldmath $r$}}_{{_{1}}}=(x_{{_{1}}},\xi_{{_{1}}},\delta E)$ before scattering transforms to ${\mbox{\boldmath $r$}}_{{_{1}}}^{\prime}=(x_{{_{1}}},\xi_{{_{1}}}^{\prime},\delta
E+{\varepsilon})$ after inelastic scattering. Propagation of ${\mbox{\boldmath $r$}}_{{_{1}}}^{\prime}$ through the time-reversal system results in a ray vector ${\mbox{\boldmath $r$}}_{{_{2}}}^{\prime}=(x_{{_{2}}}^{\prime},\xi_{{_{2}}}^{\prime},\delta
E+{\varepsilon})= \hat{O}_{{_{\mathrm {{R}}}}}{\mbox{\boldmath $r$}}_{{_{1}}}^{\prime}$. Assuming that refocusing condition holds, we come to a decisive point: all photons independent of the incident photon energy $E+\delta E$ are refocused to the same location $$x_{{_{2}}}^{\prime}=x_{{_{2}}}+G_{{_{\mathrm {{R}}}}} {\varepsilon},\hspace{0.5cm} x_{{_{2}}}=A_{{_{\mathrm {{R}}}}} A_{{_{\mathrm {{D}}}}} x_{{_{o}}},
\label{magnification2}$$ which is, however, shifted from $x_{{_{2}}}$ by $G_{{_{\mathrm {{R}}}}}
{\varepsilon}$, a value proportional to the energy transfer ${\varepsilon}$ in the inelastic scattering process. The essential point is that, the combined defocusing-refocusing system maps the inelastic scattering spectrum onto image plane $2$. The image is independent of the spectral composition $E+\delta E$ of the photons in the incident beam.
The spectral resolution $\Delta {\varepsilon}$ of the echo spectrometer is calculated from the condition, that the shift due to inelastic scattering $x_{{_{2}}}^{\prime}-x_{{_{2}}}=G_{{_{\mathrm {{R}}}}}
{\varepsilon}$ is at least as large as the linear size $\Delta
x_{{_{2}}}$ of the echo signal : $$\Delta {\varepsilon}= \frac{\Delta x_{{_{2}}}}{|G_{{_{\mathrm {{R}}}}}|} \equiv \frac{|A_{{_{\mathrm {{R}}}}}|\Delta x_{{_{1}}} }{|G_{{_{\mathrm {{R}}}}}|} \equiv \frac{|A_{{_{\mathrm {{R}}}}} A_{{_{\mathrm {{D}}}}}| \Delta x_{{_{o}}}}{|G_{{_{\mathrm {{R}}}}}|}.
\label{resolution}$$ Here it is assumed that the spatial resolution of the detector is better than $\Delta x_{{_{2}}}$.
These results constitute the underlying principle of x-ray echo spectroscopy. Noteworthy, angular dispersion always results in an inclined intensity front, i.e., in dispersion both perpendicular to and along the beam propagation direction [@SL12]. Therefore, x-rays are defocused and refocused also in the time domain, as in spin-echo. As a result, inelastic scattering spectra can be also mapped by measuring time distributions in the detector, given a short-pulse source.
Perfect refocusing takes place if the linear dispersion of the combined system $G_{{_{\mathrm {{C}}}}}=A_{{_{\mathrm {{R}}}}}G_{{_{\mathrm {{D}}}}}+G_{{_{\mathrm {{R}}}}}$ vanishes, as in Eq. . Refocusing can still take place with good accuracy if $|G_{{_{\mathrm {{C}}}}}|$ is sufficiently small: $$|G_{{_{\mathrm {{C}}}}}| \Delta E_{\cup}\ll \Delta x_{{_{2}}}.
\label{tolerances}$$ Here $\Delta E_{\cup}$ is the bandwidth of x-rays in image plane $2$. Tolerances on the echo spectrometer parameters and on the sample shape can be calculated with Eq. .
The above approach is general and applicable to any frequency domain. A particular version was proposed in the soft x-ray domain, for applications in resonant IXS spectroscopy at $L$-edges of $3d$ elements [@Fung04; @Lai14]. The dispersing elements in the soft x-ray and visible light domains are diffraction gratings. with diffraction gratings as dispersing elements [@Fung04; @Lai14].
Optical Design
==============
Diffraction gratings are not practical in the hard x-ray regime. Extension into the hard x-ray regime is therefore nontrivial. In this regard, as was demonstrated in [@Shvydko-SB; @SLK06], the angular dispersion in the hard x-ray regime can be achieved by Bragg diffraction from asymmetrically cut crystals, i.e., from crystals with the reflecting atomic planes not parallel to the entrance surface. This is a hard x-ray analog of the optical diffraction gratings or optical prisms. A large dispersion rate is a key for achieving high spectral resolution in angular-dispersive x-ray spectrometers [@SSS14; @Shvydko15], including echo spectrometers, see Eq. (6). This is achieved, first, by using strongly asymmetric Bragg reflections close to backscattering [@Shvydko-SB; @SLK06], and, second, by enhancing the single-reflection dispersion rate considerably by subsequent asymmetric Bragg reflections from crystals in special arrangements [@SSM13] exemplified below. In the following two steps, we will show how the principle scheme of a generic echo spectrometer presented above, can be realized in the hard x-ray regime by using multi-crystal arrangements as dispersing elements.
![Principle optical scheme of an x-ray echo spectrometer, comprised of the defocusing $\hat{O}_{{_{\mathrm {{D}}}}}$ and refocusing $\hat{O}_{{_{\mathrm {{R}}}}}$ dispersing systems; the x-ray source in reference plane $0$; the sample in plane $1$; and the position-sensitive detector in $2$. The spectrometer is shown in the vertical dispersion plane for elastic (v$_{{_{\mathrm {e}}}}$) and inelastic (v$_{{_{\mathrm {i}}}}$) scattering, and in the horizontal scattering plane (h) with the refocusing system at a scattering angle $\Phi$. The defocusing system $\hat{O}_{{_{\mathrm {{D}}}}}$ consists of a dispersing Bragg diffracting (multi)crystal element D$_{{_{\mathrm {{D}}}}}$, and of a focusing element $f$. The refocusing system $\hat{O}_{{_{\mathrm {{R}}}}}$ is of a spectrograph type comprising collimating element $f_{{_{1}}}$; a dispersing Bragg diffracting (multi)crystal element D$_{{_{\mathrm {{R}}}}}$; and a focusing element $f_{{_{2}}}$. []{data-label="fig002"}](2-revised-more-in-and-elastic-prl.pdf){width="50.00000%"}
In the first step, we propose a principle optical scheme of a hard x-ray echo spectrometer (Fig. \[fig002\]) comprised of the defocusing $\hat{O}_{{_{\mathrm {{D}}}}}$ and refocusing $\hat{O}_{{_{\mathrm {{R}}}}}$ dispersing systems. The x-ray source is in reference plane $0$, the sample is in plane $1$, and the position-sensitive detector is in plane $2$. The defocusing system $\hat{O}_{{_{\mathrm {{D}}}}}$ is proposed here as a combination of a Bragg (multi)crystal dispersing element D$_{{_{\mathrm {{D}}}}}$ and a focusing element $f$. As has been shown in [@Shvydko15], such a system can be represented by a ray-transfer matrix with the magnification $A_{{_{\mathrm {{D}}}}}$ and linear dispersion $G_{{_{\mathrm {{D}}}}}$ matrix elements given by $$A_{{_{\mathrm {{D}}}}}\!=\! - \frac{1}{{b_{\cup_{{_{{D}}}}}}}\!\frac{l_{{_{3}}}}{l_{{_{12}}}},\hspace{0.20cm}
G_{{_{\mathrm {{D}}}}}\!=\!{{{\mathcal D}}_{\cup_{{_{{D}}}}}} \frac{l_{{_{3}}} l_{{_{1}}}}{{b_{\cup_{{_{{D}}}}}}^2 l_{{_{12}}}},\hspace{0.20cm} l_{{_{12}}}\! =\!\frac{l_{{_{1}}}}{{b_{\cup_{{_{{D}}}}}}^2} + l_{{_{2}}}.
\label{defocusing}$$ Here, $l_{{_{1}}}$, $l_{{_{2}}}$, and $l_{{_{3}}}$, are the distances between the x-ray source, the dispersing element D$_{{_{\mathrm {{D}}}}}$, the focusing element $f$ with focal length $f=(l_{{_{12}}}^{-1}+l_{{_{3}}}^{-1})^{-1}$, and the sample in the image plane $1$, respectively (Fig. \[fig002\]). The dispersing (multi)crystal system D$_{{_{\mathrm {{D}}}}}$ is characterized by the cumulative angular dispersion rate ${{{\mathcal D}}_{\cup_{{_{{D}}}}}}$, and cumulative asymmetry factor ${b_{\cup_{{_{{D}}}}}}$, which are defined in [@Shvydko15] (see also Appendix \[raytransfer\] and Table \[tab2\]).
![Dispersing element D$_{{_{\mathrm {{D}}}}}$ (a) of the defocusing system $\hat{O}_{{_{\mathrm {{D}}}}}$ (see Fig. \[fig002\]) and its spectral transmittance function (b). D$_{{_{\mathrm {{D}}}}}$ is an example of an in-line four-crystal CDDW-type dispersing optic, comprised of collimating (C), dispersing (D$_{{_{1}}}$, D$_{{_{2}}}$), and wavelength-selecting (W) crystals in a ($\pi+$,$0-$,$0+$,$0-$) scattering configuration. With the crystal parameters provided in Table \[tab1\] of Appendix \[cddwoptic\], the dispersing element D$_{{_{\mathrm {D}}}}$ features a spectral transmission function with a $\Delta E_{{_{\cup}}}=8.8$ meV bandwidth (b), a cumulative angular dispersion rate ${{{\mathcal D}}_{\cup_{{_{{D}}}}}} = -25 ~\mu$rad/meV, and a cumulative asymmetry factor ${b_{\cup_{{_{{D}}}}}} = 1.4$. The sharp line in (b) presents the 0.1-meV spectral resolution of an x-ray echo spectrometer, designed for use with 9.1-keV photons. []{data-label="fig003"}](cddw-pi+0-0+0-comb-art.pdf){width="50.00000%"}
![Dispersing element D$_{{_{\mathrm {{R}}}}}$ (a) of the refocusing system $\hat{O}_{{_{\mathrm {{R}}}}}$ (see Fig. \[fig002\]) and its spectral transmittance function (b). Similar to D$_{{_{\mathrm {{D}}}}}$ in Fig. \[fig003\], D$_{{_{\mathrm {{R}}}}}$ is an example of the in-line four-crystal CDDW-type dispersing optic, but, in a ($\pi+$,$\pi+$,$\pi-$,$0-$) scattering configuration. With the crystal parameters provided in Table \[tab1\] of Appendix \[cddwoptic\], the dispersing element D$_{{_{\mathrm {D}}}}$ features a spectral transmission function with a $\Delta E_{{_{\cup}}}=5.5$ meV bandwidth (b), a cumulative angular dispersion rate ${{{\mathcal D}}_{\cup_{{_{{D}}}}}} =
-43 ~\mu$rad/meV, a cumulative asymmetry factor ${b_{\cup_{{_{{D}}}}}} =
0.36$, and ${{{\mathcal D}}_{\cup_{{_{{D}}}}}}/{b_{\cup_{{_{{D}}}}}} = -120 ~\mu$rad/meV. []{data-label="fig004"}](cddw-pi+pi+pi-0-comb-v2-art.pdf){width="50.00000%"}
For the spectrometer to feature a large throughput, the refocusing system $\hat{O}_{{_{\mathrm {{R}}}}}$ has to be capable of collecting x-ray photons in a large solid angle scattered from the sample. For this purpose, we propose using a hard x-ray focusing-dispersing system of a spectrograph-type considered in [@Shvydko15], and schematically shown in Fig \[fig002\]. A collimating focusing element $f_{{_{1}}}$ collects photons in a large solid angle and makes x-ray beams of each spectral component parallel. The collimated beams impinge upon the Bragg (multi)crystal dispersing element D$_{{_{\mathrm {{R}}}}}$ with the cumulative angular dispersion rate ${{{\mathcal D}}_{\cup_{{_{{R}}}}}}$, and the cumulative asymmetry factor ${b_{\cup_{{_{{R}}}}}}$. The focusing element $f_{{_{2}}}$ focuses x-rays in the vertical dispersion plane onto the detector placed in the image plane $2$. As shown in [@Shvydko15] (see also Appendix \[raytransfer\] and Table \[tab2\]) such a system is described by a ray-transfer matrix with the magnification $A_{{_{\mathrm {{R}}}}}$ and linear dispersion $G_{{_{\mathrm {{R}}}}}$ matrix elements given by $$A_{{_{\mathrm {{R}}}}}=-\frac{{b_{\cup_{{_{{R}}}}}}f_{{_{2}}}\!}{f_{{_{1}}}\!},\hspace{0.5cm}
G_{{_{\mathrm {{R}}}}}={{{\mathcal D}}_{\cup_{{_{{R}}}}}} f_{{_{2}}} .
\label{refocusing}$$ Using Eqs. , , and we obtain for the refocusing condition in the hard x-ray echo spectrometer $$\frac{l_{{_{3}}} l_{{_{1}}}}{l_{{_{1}}} + {b_{\cup_{{_{{D}}}}}}^2 l_{{_{2}}}} {{{\mathcal D}}_{\cup_{{_{{D}}}}}} = f_{{_{1}}} \frac{{{{\mathcal D}}_{\cup_{{_{{R}}}}}}}{{b_{\cup_{{_{{R}}}}}}}.
\label{refocus2}$$ The dispersing element D$_{{_{\mathrm {{D}}}}}$ can be placed from the source at a large distance $l_{{_{1}}} \gg {b_{\cup_{{_{{D}}}}}}^2 l_{{_{2}}}$. In this case, the refocusing condition reads $$l_{{_{3}}}{{{\mathcal D}}_{\cup_{{_{{D}}}}}} \simeq f_{{_{1}}} \frac{{{{\mathcal D}}_{\cup_{{_{{R}}}}}}}{{b_{\cup_{{_{{R}}}}}}}.
\label{refocus23}$$ For the spectral resolution $\Delta {\varepsilon}$ of a hard x-ray echo spectrometer we obtain from Eqs. , , and : $$\Delta {\varepsilon}= \frac{|{b_{\cup_{{_{{R}}}}}}|}{|{{{\mathcal D}}_{\cup_{{_{{R}}}}}}|} \frac{\Delta x_{{_{1}}}}{f_{{_{1}}}}.
\label{resolution2}$$ As follows from Eq. , the spectral resolution of the echo spectrometer is determined solely by the parameters of the refocusing system, i.e., by the resolution of the hard x-ray spectrograph [@Shvydko15]. The parameters of the defocusing system determine only the size of the secondary monochromatic source on the sample $\Delta x_{{_{1}}}=|A_{{_{\mathrm {{D}}}}}|\Delta
x_{{_{0}}}$.
In the second step, we consider the particular optical designs of x-ray echo spectrometers with a very high spectral resolution $\Delta
{\varepsilon}\lesssim 0.1$ meV. For practical reasons, we will assume that the secondary [*monochromatic*]{} source size is $\Delta
x_{{_{1}}}\simeq 5~\mu$m, and the focal length is $f_{{_{1}}}\simeq 0.4$ m of the collimating element in the refocusing system. Then, Eq. requires the ratio ${|{{{\mathcal D}}_{\cup_{{_{{R}}}}}}|} /{|{b_{\cup_{{_{{R}}}}}}|}\simeq 125~\mu$rad/meV for the dispersive element D$_{{_{\mathrm {{R}}}}}$. Assuming the distance $l_{{_{3}}}\simeq 2$ m from the focusing element to the sample in the defocusing system, and $l_{{_{1}}} \gg {b_{\cup_{{_{{D}}}}}}^2
l_{{_{2}}}$, we estimate from Eq. for the required cumulative dispersing rate ${|{{{\mathcal D}}_{\cup_{{_{{R}}}}}}|} \simeq
25~\mu$rad/meV in the defocusing dispersive element. These are relatively large values. Typically, in a single Bragg reflection, a maximum dispersion rate is ${{\mathcal D}}\simeq 6-8~\mu$rad/meV for photons with energy $E\simeq 10$ keV [@SLK06; @ShSS11]. As mentioned before, multi-crystal arrangements can be used to enhance the dispersion rate [@SSM13].
Such large dispersion rates ${{{\mathcal D}}_{\cup_{{_{}}}}}$, unfortunately, tend to decrease the transmission bandwidths $\Delta E_{{_{\cup}}}$ of the dispersing elements [@SSM13; @Shvydko15]. Achieving strong signals in IXS experiments, however, requires $\Delta E_{{_{\cup}}} \gg
\Delta {\varepsilon}$. Therefore, optical designs of the dispersing elements have to be found featuring both large ${{{\mathcal D}}_{\cup_{{_{}}}}}$ and $\Delta
E_{{_{\cup}}}$. Figures \[fig003\] and \[fig004\] show representative examples of multi-crystal CDDW-type inline dispersing elements [@ShSS11; @SSS13; @SSS14] of the defocusing and refocusing systems, respectively, with the required cumulative dispersion rates, asymmetry factors, and with bandwidths $\Delta E_{{_{\cup}}}\simeq
5.5-9$ meV, i.e., $\Delta E_{{_{\cup}}}/\Delta {\varepsilon}\simeq
55-90$, designed for use with 9.1-keV photons. These examples are modifications of the dispersing elements designs presented in [@Shvydko15]. The CDDW optic in the ($\pi+$,$0-$,$0+$,$0-$) scattering configuration is preferred for the defocusing dispersing element D$_{{_{\mathrm {{D}}}}}$ (Fig. \[fig003\]) as it provides the required dispersion rate ${|{{{\mathcal D}}_{\cup_{{_{{D}}}}}}|}\simeq 25~\mu$rad/meV, significant transmission bandwidth $\Delta E_{{_{\cup}}}\simeq 9$ meV, and is compact. The CDDW optic in the ($\pi+$,$\pi+$,$\pi-$,$0-$) configuration is better suited for the refocusing dispersing element D$_{{_{\mathrm {{R}}}}}$ (Fig. \[fig004\]). It provides the large ratio ${|{{{\mathcal D}}_{\cup_{{_{{R}}}}}}|} /{|{b_{\cup_{{_{{R}}}}}}|}\simeq 120~\mu$rad/meV required for the high spectral resolution \[see Eq. \] and substantial transmission bandwidth $\Delta E_{{_{\cup}}}\simeq
5.5$ meV, although smaller than in the D$_{{_{\mathrm {{D}}}}}$-case.
The total beam size on the sample is $\simeq G_{{_{\mathrm {{D}}}}}\Delta E_{{_{\cup}}}$. For the spectrometer exemplified here, it is estimated to be $\simeq 275~\mu$m. Equation together with Eqs. and can be used to estimate tolerances on the admissible variations of the focal distances, sample displacement, sample surface imperfections, etc. As discussed in Appendix \[echotollerances\] the tolerances are in a millimeter range in the case of the 0.1-meV spectrometer.
Details on the optical designs and examples of the dispersing elements designed for use with a lower photon energy of 4.57 keV and larger $\Delta E_{{_{\cup}}}\simeq 13-10$ meV, can be found in Appendix \[cddwoptic\]. All these examples showcase the applicability of echo spectroscopy in a wide spectral range, and its feasibility both with synchrotron radiation and x-ray free electron laser sources.
The point is that even higher spectral resolution $\Delta {\varepsilon}\lesssim 0.02$ meV can be achieved with x-ray echo spectrometers by increasing the dispersion rates ${{{\mathcal D}}_{\cup_{{_{}}}}}$ in the dispersing elements. This, however, will result in their narrower transmission bandwidths $\Delta E_{{_{\cup}}}$. Still, an approximately constant ratio $\Delta E_{{_{\cup}}}/\Delta {\varepsilon}$ holds. Alternatively, the spectral resolution can be improved by increasing the focal length $f_{{_{1}}}$ in the refocusing system, see Eq. .
The essential feature of the echo spectrometers is that the signal strength, which is proportional to the product of the bandwidths of the photons on the sample and on the detector, is enhanced by $(\Delta E_{{_{\cup}}}/\Delta
{\varepsilon})^2 \simeq 10^3-10^4$ compared to what is possible with the standard scanning-IXS-spectrometer approach.
Conclusions
===========
In conclusion, x-ray echo spectroscopy, a counterpart of neutron spin-echo, is introduced here to overcome limitations in spectral resolution and weak signals of the traditional inelastic hard x-ray scattering (IXS) probes. Operational principles, refocusing conditions, and spectral resolutions of echo spectrometers are substantiated by an analytical ray-transfer-matrix approach. A principle optical scheme for a hard x-ray echo spectrometer is proposed with multi-crystal arrangements as dispersing elements. Concrete schemes are discussed with 5–13-meV transmission bandwidths, a spectral resolution of 0.1-meV (extension to 0.02-meV is realistic), and designed for use with 9.1-keV and 4.6-keV photons. The signal in echo spectrometers is enhanced by at least three orders of magnitude compared to what is possible with the standard scanning-IXS-spectrometer approach.
X-ray echo spectrometers require a combination of the CDDW dispersing elements and focusing optics as major optical components. Such components have been experimentally demonstrated recently [@SSM13; @SSS14]. Implementation of x-ray echo spectrometers is, therefore, realistic.
Acknowledgments
===============
Stimulating discussions with D.-J. Huang (NSRRC) are greatly appreciated. S.P. Collins (DLS) is acknowledged for reading the manuscript and for valuable suggestions. Work at Argonne National Laboratory was supported by the U.S. Department of Energy, Office of Science, under Contract No. DE-AC02-06CH11357.
[21]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ed., **, vol. of ** (, , ).
, ****, ().
, , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, **, vol. of ** (, , ).
, , , , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, ****, ().
, , , , , , , **** ().
, , , , ****, ().
, , , , , , , , , , , ****, ().
, **, Optical Sciences (, , ).
, ****, ().
, , , , ****, ().
, , , , , , , , ****, ().
, , , , , , ****, ().
Ray-transfer matrices {#raytransfer}
=====================
Ray-transfer matrices $\{A0G,CDF,001\}$ of the defocusing $\hat{O}_{{_{\mathrm {D}}}}$ and refocusing $\hat{O}_{{_{\mathrm {R}}}}$ systems of the x-ray echo spectrometers used in the paper are given in the last two rows of Table \[tab2\]. They are equivalent to the derived in Ref. [@Shvydko15] ray-transfer matrices of x-ray focusing monochromators and spectrographs. The matrices of the multi-element systems $\hat{O}_{{_{\mathrm {D}}}}$ and $\hat{O}_{{_{\mathrm {R}}}}$ are obtained by successive multiplication of the matrices of the constituent optical elements, which are given in the upper rows of Table \[tab2\].
In the first three rows, 1–3, matrices are shown for the basic optical elements, such as propagation in free space $\hat{{P}}(l)$, thin lens or focusing mirror $\hat{{L}}(f)$, and Bragg reflection from a crystal $\hat{{C}}(b,{s}{{\mathcal D}})$. Scattering geometries in Bragg diffraction from crystals are defined in Fig. \[fig005\]. In the following rows of Table \[tab2\], ray-transfer matrices are shown for arrangements composed of several basic optical elements, such as successive multiple Bragg reflections from crystals $\hat{{C}}({b_{\cup_{{_{n}}}}},{{{\mathcal D}}_{\cup_{{_{n}}}}})$ and $\hat{{K}}({b_{\cup_{{_{n}}}}},{{{\mathcal D}}_{\cup_{{_{n}}}}},l)$, rows 4–5; and a focusing system $\hat{{F}}(l_{{_{2}}},f,l_{{_{1}}})$, row 6.
The matrices of the defocusing $\hat{O}_{{_{\mathrm {D}}}}$ and refocusing $\hat{O}_{{_{\mathrm {R}}}}$ systems presented in Table \[tab2\], rows 7 and 8, respectively, are calculated using the multi-crystal matrix $\hat{{C}}({b_{\cup_{{_{n}}}}},{{{\mathcal D}}_{\cup_{{_{n}}}}})$ from row 4, assuming zero free space between crystals in successive Bragg reflections. Generalization to a more realistic case of nonzero distances between the crystals requires the application of matrix $\hat{{K}}({b_{\cup_{{_{n}}}}},{{{\mathcal D}}_{\cup_{{_{n}}}}},l)$ from row 5.
We refer to Ref. [@Shvydko15] for details on the derivation of these matrices. Here, we provide only the final results, notations, and definitions.
![Definitions of scattering geometries in Bragg diffraction from a crystal: (a) $0+$, (b) $0-$, (c) $\pi+$, and (d) $\pi-$.[]{data-label="fig005"}](5.pdf){width="50.00000%"}
---------------- --------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------
Optical system Matrix notation Ray-transfer matrix Definitions and remarks
\[-5pt\]
\[-3.8mm\] $\hat{{P}}(l)$ $\left( \begin{array}{ccc} 1 & l & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{array} \right)$
\[-0.5mm\]
\[-3.8mm\] $\hat{{L}}(f)$ $\left( \begin{array}{ccc} 1 & 0 & 0\\ -\frac{1}{f} & 1 & 0 \\ 0 & 0 & 1 \end{array} \right)$
\[-0.5mm\]
\[-4.2mm\] $\hat{{C}}(b,{s}{{\mathcal D}})$ $\left( \begin{array}{ccc} {1}/{b} & 0 & 0 \\0 & b & {s}{{\mathcal D}}\\ 0 & 0 & 1 \end{array} \right)$
\[+0.5mm\]
\[-3.8mm\] $\hat{{C}}({b_{\cup_{{_{n}}}}},{{{\mathcal D}}_{\cup_{{_{n}}}}})$ $\left( \begin{array}{ccc} {1/b_{\cup_{{_{n}}}}} & 0 & 0 \\ 0 & {b_{\cup_{{_{n}}}}} & {{{\mathcal D}}_{\cup_{{_{n}}}}} \\ 0 & 0 & 1 \end{array} \right)$
\[-0.5mm\]
\[-3.8mm\] $\hat{{K}}({b_{\cup_{{_{n}}}}},{{{\mathcal D}}_{\cup_{{_{n}}}}},l)$ $\left( \begin{array}{ccc} {1/b_{\cup_{{_{n}}}}} & {B_{\cup_{{_{n}}}}} & {G_{\cup_{{_{n}}}}} \\ 0 & {b_{\cup_{{_{n}}}}} & {{{\mathcal D}}_{\cup_{{_{n}}}}} \\ 0 & 0 & 1 \end{array} \right)$
\[-0.5mm\]
\[-3.8mm\] $\hat{{F}}(l_{{_{2}}},f,l_{{_{1}}})$ $\left( \begin{array}{ccc} 1-\frac{l_{{_{2}}}}{f} & B_{{_{\mathrm {F}}}} & 0 \\ -\frac{1}{f} & 1-\frac{l_{{_{1}}}}{f} & 0\\ 0 & 0 & 1 \end{array} \right)$
\[-0.5mm\]
\[-3.8mm\] $\hat{O}_{{_{\mathrm {D}}}}$ $\left(\!\! \begin{array}{ccc} \frac{1}{{b_{\cup_{{_{n}}}}}\!}\!\left(\!1\!-\!\frac{l_{{_{3}}}}{f}\!\right) & 0 & X {{{\mathcal D}}_{\cup_{{_{n}}}}} \\
-\frac{1}{f{b_{\cup_{{_{n}}}}}} & {b_{\cup_{{_{n}}}}}\!\!\left(\!1\!-\!\frac{l_{12}}{f}\!\right) & \left(\!1\!-\!\frac{l_{{_{2}}}}{f}\!\right)\!{{{\mathcal D}}_{\cup_{{_{n}}}}} \\
0 & 0 & 1 \end{array}\!\! \right)
$
\[-0.5mm\]
\[-3.8mm\] $\hat{O}_{{_{\mathrm {R}}}}$ $
\left(\!\! \begin{array}{ccc} -\frac{{b_{\cup_{{_{n}}}}}f_{{_{2}}}\!}{f_{{_{1}}}\!} & 0 & f_{{_{2}}} {{{\mathcal D}}_{\cup_{{_{n}}}}} \\ \frac{(\!l_{{_{1}}}\!-\!f_{{_{1}}}\!)\!+\!(\!l_{{_{2}}}\!-\!f_{{_{2}}}\!){b_{\cup_{{_{n}}}}}^2\!}{\!{b_{\cup_{{_{n}}}}}f_{{_{1}}}f_{{_{2}}}\!} & -\frac{f_{{_{1}}}}{{b_{\cup_{{_{n}}}}}\!f_{{_{2}}}} & \left(\!1\!-\!\frac{l_{{_{2}}}}{f_{{_{2}}}}\!\right)\!{{{\mathcal D}}_{\cup_{{_{n}}}}}\! \\ 0 & 0 & 1 \end{array}\!\! \right)
$
---------------- --------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------
CDDW optic as dispersing element {#cddwoptic}
================================
![Dispersing element D$_{{_{\mathrm {{D}}}}}$ (a) of the defocusing system $\hat{O}_{{_{\mathrm {{D}}}}}$ (Fig. 2) and its spectral transmittance function (b). D$_{{_{\mathrm {{D}}}}}$ is an example of an in-line four-crystal CDDW-type dispersing optic, comprised of collimating (C), dispersing (D$_{{_{1}}}$, D$_{{_{2}}}$), and wavelength-selecting (W) crystals in a ($\pi+$,$0-$,$0+$,$0-$) scattering configuration. With the crystal parameters provided in Table \[tab14\], the dispersing element D$_{{_{\mathrm {D}}}}$ features a spectral transmission function with a $\Delta E_{{_{\cup}}}=13$ meV bandwidth (b), a cumulative angular dispersion rate ${{{\mathcal D}}_{\cup_{{_{{D}}}}}} =
-42 ~\mu$rad/meV, and a cumulative asymmetry factor ${b_{\cup_{{_{{D}}}}}} =
1.4$. The sharp line in (b) presents the 0.1-meV design spectral resolution of the x-ray echo spectrometer. $E=4.5686$ keV. []{data-label="fig00399"}](cddw-pi+0-0+0-comb-E4o57keV.pdf){width="50.00000%"}
![Dispersing element D$_{{_{\mathrm {{R}}}}}$ (a) of the refocusing system $\hat{O}_{{_{\mathrm {{R}}}}}$ (Fig. 2) and its spectral transmittance function (b). Similar to D$_{{_{\mathrm {{D}}}}}$ in Fig. \[fig00399\], D$_{{_{\mathrm {{R}}}}}$ is an example of the in-line four-crystal CDDW-type dispersing optic, but, in a ($\pi+$,$\pi+$,$\pi-$,$0-$) scattering configuration. With the crystal parameters provided in Table \[tab14\], the dispersing element D$_{{_{\mathrm {D}}}}$ features a spectral transmission function with a $\Delta
E_{{_{\cup}}}=9.9$ meV bandwidth (b), a cumulative angular dispersion rate ${{{\mathcal D}}_{\cup_{{_{{D}}}}}} = -64 ~\mu$rad/meV, a cumulative asymmetry factor ${b_{\cup_{{_{{D}}}}}} = 0.39$, and ${{{\mathcal D}}_{\cup_{{_{{D}}}}}}/{b_{\cup_{{_{{D}}}}}}
= -161 ~\mu$rad/meV. The sharp line in (b) presents the 0.1-meV design spectral resolution of the x-ray echo spectrometer. $E=4.5686$ keV. []{data-label="fig0049"}](cddw-pi+pi+pi-0-comb-4o57keV.pdf){width="50.00000%"}
In-line four-crystal CDDW-type dispersing optics [@ShSS11; @SSS13; @SSS14] are proposed in the paper for use as dispersing elements D$_{{_{\mathrm {D}}}}$, D$_{{_{\mathrm {R}}}}$ of the defocusing $\hat{O}_{{_{\mathrm {D}}}}$ and refocusing systems $\hat{O}_{{_{\mathrm {{R}}}}}$ of the echo spectrometer, respectively. The in-line four-crystal CDDW-type dispersing optic (see schematics in Figs. 3-4 and Figs. \[fig00399\]-\[fig0049\]), comprises collimating (C), dispersing (D$_{{_{1}}}$, D$_{{_{2}}}$), and wavelength-selecting (W) crystals, which can be arranged in different scattering configurations. In a general case, the scattering configuration is defined as $(\phi_{{_{1}}}s_{{_{1}}},\phi_{{_{2}}}s_{{_{2}}},\phi_{{_{3}}}s_{{_{3}}},\phi_{{_{4}}}s_{{_{4}}})$. Here, for each crystal $n=1,2,3,4$ (1=C, 2=D$_{{_{1}}}$, 3=D$_{{_{2}}}$, 4=W) the value $\phi_{{_{n}}}=0$ corresponds to the grazing reflection, see Fig. \[fig005\](a)-(b); while $\phi_{{_{n}}}=\pi$ corresponds to the grazing incidence, see Fig. \[fig005\](c)-(d). The sign $s_{{_{n}}}=+1$ corresponds to a reflection in the counterclockwise direction, see Figs. \[fig005\](a),(c); while $s_{{_{n}}}=-1$ means the clockwise direction, see Figs. \[fig005\](b),(d). Those scattering geometries have been selected for use as dispersing elements of the x-ray echo spectrometer in the paper, which feature the largest cumulative dispersion rates ${{{\mathcal D}}_{\cup_{{_{n}}}}}$.
The cumulative dispersion rate ${{{\mathcal D}}_{\cup_{{_{4}}}}}$ in a four-crystal system is given in a general case by $${{{\mathcal D}}_{\cup_{{_{4}}}}} =
b_{{_{4}}}b_{{_{3}}}b_{{_{2}}}{s}_{{_{1}}}{{\mathcal D}}_{{_{1}}}+b_{{_{4}}}b_{{_{3}}}{s}_{{_{2}}}{{\mathcal D}}_{{_{2}}}+b_{{_{4}}}{s}_{{_{3}}}{{\mathcal D}}_{{_{3}}}+{s}_{{_{4}}}{{\mathcal D}}_{{_{4}}},$$ with the asymmetry parameters $b$ and dispersion rates ${{\mathcal D}}$ defined in Table \[tab2\], rows (3) and (4). For the CDDW-type dispersing elements considered in this paper the dispersion rate of the D-crystals ($n=2,3$) is much larger than those of the C- and W-crystals ($n=1,4$), see Tables \[tab1\] and \[tab14\]. In this case, the cumulative dispersion rate can be approximated by ${{{\mathcal D}}_{\cup_{{_{4}}}}}
\simeq
b_{{_{4}}}b_{{_{3}}}{s}_{{_{2}}}{{\mathcal D}}_{{_{2}}}+b_{{_{4}}}{s}_{{_{3}}}{{\mathcal D}}_{{_{3}}}=b_{{_{4}}}(b_{{_{3}}}{s}_{{_{2}}}{{\mathcal D}}_{{_{2}}}+{s}_{{_{3}}}{{\mathcal D}}_{{_{3}}})$. The largest dispersion rates can be achieved in systems, in which the product is ${s}_{{_{2}}}{s}_{{_{3}}}{{\mathcal D}}_{{_{2}}}{{\mathcal D}}_{{_{3}}}<0$.
There are four high-dispersion-rate CDDW configurations featuring ${{\mathcal D}}_{{_{2}}}{{\mathcal D}}_{{_{3}}}>0$ and ${s}_{{_{2}}}{s}_{{_{3}}}<0$: $(\pi+,\pi-,\pi+,0+)$; $(\pi+,\pi+,\pi-,0-)$; $(\pi+,0-,0+,0-)$; and $(\pi+,0+,0-,0-)$. These configurations are especially interesting because of the incident and transmitted x-rays being parallel (in-line scheme).
There are four other high-dispersion-rate CDDW configurations featuring ${{\mathcal D}}_{{_{2}}}{{\mathcal D}}_{{_{3}}}<0$ and ${s}_{{_{2}}}{s}_{{_{3}}}>0$: $(\pi+,\pi-,0-,0-)$; $(\pi+,\pi+,0+,0-)$; $(\pi+,0-,\pi-,0-)$; and $(\pi+,0+,\pi+,0-)$. However, they are not in-line. The angle between the incident and reflected beams is $4(\theta_{\mathrm D}-\pi/2)$.
In the present paper, we have chosen the in-line high-dispersion-rate CDDW optic in the ($\pi+$,$0-$,$0+$,$0-$) configuration, as a dispersing element D$_{{_{\mathrm {{D}}}}}$ of the defocusing system $\hat{O}_{{_{\mathrm {{D}}}}}$, see Figs. 3 and \[fig00399\]. The in-line high-dispersion-rate CDDW optic in the ($\pi+$,$\pi+$,$\pi-$,$0-$) configuration, was chosen as a dispersing element D$_{{_{\mathrm {{R}}}}}$ of the refocusing system $\hat{O}_{{_{\mathrm {{R}}}}}$, see Figs. 4 and \[fig0049\].
Tables \[tab1\] and \[tab14\] present crystal parameters and cumulative parameters of the dispersing elements designed for operations with x-rays with photon energies of 9.131385 keV and 4.5686 keV, respectively.
[|l|lllllll|]{} crystal & ${\mbox{\boldmath $H$}}_{{_{\mathrm {{e}}}}}$ &$\eta_{{_{\mathrm {{e}}}}} $ &$\theta_{{_{\mathrm {{e}}}}} $ & ${\Delta E_{{_{\mathrm {{e}}}}}} $ & ${\Delta \theta_{{_{\mathrm {{e}}}}}}$ & $b_{{_{\mathrm {{e}}}}}$ & $s_{{_{\mathrm {{e}}}}}{{\mathcal D}}_{{_{\mathrm {{e}}}}} $\
element ([e]{}) & & & & & & &\
\[material\] & $(hkl)$ & deg & deg & meV & $\mu$rad & & $\frac{\mu {\mathrm {rad}}}{\mathrm {meV}}$\
\
1 C \[Ge\]& (1 1 1) & -10.0 & 12.0 & 3013 & 71 & -0.09 & -0.02\
2 D$_{{_{1}}}$ \[Si\] & (8 0 0) & 77.5 & 89 & 27 & 341 & -1.17 & -1.07\
3 D$_{{_{2}}}$ \[Si\] & (8 0 0) & 77.5 & 89 & 27 & 341 & -1.17 & +1.07\
4 W \[Ge\] & (1 1 1) & 10.0 & 12.0 & 3013 & 71 & -10.8 & -0.22\
& $\Delta E_{{_{\cup}}}$ & $\Delta \theta_{{_{\cup}}}^{\prime}$ & ${b_{\cup_{{_{ }}}}}$ & ${{{\mathcal D}}_{\cup_{{_{ }}}}}$\
& meV & $\mu$rad & & $\frac{\mu {\mathrm {rad}}}{\mathrm {meV}}$\
& 8.8 & -218 & 1.38 & -25.0\
\
1 C \[Ge\]& (1 1 1) & -10.0 & 12.0 & 3013 & 71 & -0.09 & -0.02\
2 D$_{{_{1}}}$ \[Si\] & (8 0 0) & -86 & 89 & 27 & 341 & -0.6 & -2.50\
3 D$_{{_{2}}}$ \[Si\] & (8 0 0) & -86 & 89 & 27 & 341 & -0.6 & +2.50\
4 W \[Ge\] & (1 1 1) & 10.0 & 12.0 & 3013 & 71 & -10.8 & -0.22\
& 5.5 & -237 & 0.36 & -43.5\
[|l|lllllll|]{} crystal & ${\mbox{\boldmath $H$}}_{{_{\mathrm {{e}}}}}$ &$\eta_{{_{\mathrm {{e}}}}} $ &$\theta_{{_{\mathrm {{e}}}}} $ & ${\Delta E_{{_{\mathrm {{e}}}}}} $ & ${\Delta \theta_{{_{\mathrm {{e}}}}}}$ & $b_{{_{\mathrm {{e}}}}}$ & $s_{{_{\mathrm {{e}}}}}{{\mathcal D}}_{{_{\mathrm {{e}}}}} $\
element ([e]{}) & & & & & & &\
\[material\] & $(hkl)$ & deg & deg & meV & $\mu$rad & & $\frac{\mu {\mathrm {rad}}}{\mathrm {meV}}$\
\
1 C \[Ge\]& (1 1 1) & -22.0 & 24.60 & 1542 & 154 & -0.06 & -0.09\
2 D$_{{_{1}}}$ \[Si\] & (4 0 0) & 68 & 88 & 110 & 691 & -1.19 & -1.18\
3 D$_{{_{2}}}$ \[Si\] & (4 0 0) & 68 & 88 & 110 & 691 & -1.19 & +1.18\
4 W \[Ge\] & (1 1 1) & 22.0 & 24.55 & 1542 & 154 & -16.4 & -1.53\
& 13.0 & -547 & 1.44 & -41.9\
\
1 C \[Ge\]& (1 1 1) & -22.0 & 24.60 & 1542 & 154 & -0.06 & -0.09\
2 D$_{{_{1}}}$ \[Si\] & (4 0 0) & -81.5 & 88 & 110 & 691 & -0.62 & -2.37\
3 D$_{{_{2}}}$ \[Si\] & (4 0 0) & -81.5 & 88 & 110 & 691 & -0.62 & +2.37\
4 W \[Ge\] & (1 1 1) & 22.0 & 24.55 & 1542 & 154 & -16.4 & -1.53\
& 9.9 & -638 & 0.39 & -64.0\
Focusing and collimating optics
===============================
Focusing and collimating optic elements are another key components of the x-ray echo spectrometers. There are no principal preferences of using either curved mirrors or compound-refractive lenses (CRL) for this purpose. However, in practical terms, mirrors maybe a preferable choise ensuring higher efficiency, because photoabsorption of 9.1-keV and especially of 4.5-keV photons is substantial in the CRLs [@SKSL; @LST99].
The focusing element $f$ in the defocusing dispersing system $\hat{O}_{{_{\mathrm {{D}}}}}$ (see Fig. 2) can be a standard K-B mirror system, ensuring a $\Delta x_{{_{1}}}\simeq 5~\mu$m vertical spot size in the echo-spectrometer example considered in the present paper. Tight focusing in the horizontal plane is also advantageous to minitage the negative effect on the spectral resolution of a “projected” scattering source size with increasing scattering angle.
The collimating element $f_{{_{1}}}$ in the refocusing dispersing system $\hat{O}_{{_{\mathrm {{R}}}}}$ collects photons in a large solid angle $\Omega_{{_{\mathrm {h}}}}\times\Omega_{{_{\mathrm {v}}}}$, with $\Omega_{{_{\mathrm {h}}}}\simeq 10$ mrad, $\Omega_{{_{\mathrm {h}}}}\simeq
1-10$ mrad (depending on the required momentum transfer resolution, and makes the x-ray beam parallel. Laterally graded multilayer Montel mirrors recently proved to be useful exactly in this role [@MSL13; @SSS14].
The focusing element $f_{{_{2}}}$ in the refocusing dispersing system $\hat{O}_{{_{\mathrm {{R}}}}}$ focuses x-rays in the vertical dispersion plane onto the detector. Because the vertical beamsize after the dispersing element D$_{{_{\mathrm {{R}}}}}$ is increased by $1/{b_{\cup_{{_{{D}}}}}}$ (an inverse of the cumulative asymmetry factor ${b_{\cup_{{_{{D}}}}}}$) the $f_{{_{2}}}$ element has to have a relatively large vertical geometrical aperture $\simeq 2-5$ mm (depending on the required spectral resolution). One-dimensional parabolic mirrors should be able to deal effectively with this problem.
Echo spectrometer tolerances {#echotollerances}
============================
Tolerances on the echo spectrometer parameters can be calculated from Eq. (7) in a general case. The equation can be rewritten as $$|G_{{_{\mathrm {{D}}}}}+G_{{_{\mathrm {{R}}}}}/A_{{_{\mathrm {{R}}}}}| \Delta E_{\cup}\ll \Delta x_{{_{1}}}
\label{tollerances2}$$ using Eq. (3) and the relationship $\Delta
x_{{_{2}}}= |A_{{_{\mathrm {{R}}}}} | \Delta x_{{_{1}}}$ from Eq. (4),
In a particular case of the echo spectrometer, which optical scheme is shown in Fig. 2, the tolerances on the spectrometer parameters can be calculated from equation $$\left| {{{\mathcal D}}_{\cup_{{_{{D}}}}}} \frac{l_{{_{3}}} l_{{_{1}}}}{{b_{\cup_{{_{{D}}}}}}^2 l_{{_{12}}}} - \frac{ {{{\mathcal D}}_{\cup_{{_{{R}}}}}} f_{{_{1}}}\! }{ {b_{\cup_{{_{{R}}}}}}\!} \right| \Delta E_{\cup}\ll \Delta x_{{_{1}}}
\label{tolerances2}$$ which is obtained by combining Eq. and Eqs.(8)-(9). If the dispersing element D$_{{_{\mathrm {{D}}}}}$ is placed from the source at a large distance $l_{{_{1}}} \gg {b_{\cup_{{_{{D}}}}}}^2 l_{{_{2}}}$, in this case, the tolerance equation simplifies to $$\left| {{{\mathcal D}}_{\cup_{{_{{D}}}}}} l_{{_{3}}} - \frac{ {{{\mathcal D}}_{\cup_{{_{{R}}}}}} f_{{_{1}}}\! }{ {b_{\cup_{{_{{R}}}}}}\!} \right| \Delta E_{\cup}\ll \Delta x_{{_{1}}}.
\label{tolerances3}$$ As an example, we assume that the spectrometer parameters are perfectly adjusted, except for the distance $l_{{_{3}}}$ from the focusing mirror to the secondary source (to the sample). The tolerance interval $\Delta l_{{_{3}}}$ in this case can be estimated using Eq. as $$\left|\Delta l_{{_{3}}} \right| \ll \frac{\Delta x_{{_{1}}}}{ |{{{\mathcal D}}_{\cup_{{_{{D}}}}}}| \Delta E_{\cup} }.
\label{toleranceintervall3}$$ If the distance $f_{{_{1}}}$ from the secondary source (sample) to the collimating mirror is not perfectly adjusted, the tolerance interval $\Delta f_{{_{1}}}$ can be estimated in this case as $$\left|\Delta f_{{_{1}}} \right| \ll \frac{\Delta x_{{_{1}}} {b_{\cup_{{_{{R}}}}}}}{ |{{{\mathcal D}}_{\cup_{{_{{R}}}}}}| \Delta E_{\cup} }.
\label{toleranceintervalf1}$$ With the parameters of the 0.1-meV-resolution echo spectrometer provided in the paper ($\Delta x_{{_{1}}}=5~\mu$m; $\Delta E_{\cup}
=5.5$ meV; $|{{{\mathcal D}}_{\cup_{{_{{D}}}}}}|=25~\mu$rad/meV; $|{b_{\cup_{{_{{R}}}}}}/{{{\mathcal D}}_{\cup_{{_{{R}}}}}}|=125~\mu$rad/meV), these tolerance intervals are estimated to be $\left|\Delta l_{{_{3}}} \right| \ll 36$ mm, and $\left|\Delta f_{{_{1}}} \right| \ll 7$ mm, respectively. These numbers are not extremely demanding.
Since the variations of $l_{{_{3}}}$ and $f_{{_{1}}}$ could be related to sample position displacement and surface imperfections or to the sample being installed at some angle to the incident beam, the above estimated numbers also provide constraints on imperfections in the sample shape in this particular case.
Tuning the refocusing condition up
==================================
In practice, the refocusing condition given by Eq. (10) can be exactly satisfied by tuning the distance $l_{{_{3}}}$, between the focusing element $f$ and the sample, see Fig. 2. Given that the source-to-sample distance $l=l_{{_{1}}}+l_{{_{2}}}+l_{{_{3}}}$, as well as the focal distance $f=l_{{_{12}}}l_{{_{3}}}/(l_{{_{12}}}+l_{{_{3}}})$ are fixed, the distances $l_{{_{1}}}$ and $l_{{_{2}}}$ also have to be corrected by the positioning of the dispersing system D$_{{_{\mathrm {{D}}}}}$, appropriately. The distances $l_{{_{1}}}$ and $l_{{_{2}}}$ are defined from the above mentioned constraints, by solving the equations $$\begin{aligned}
l_{{_{1}}}+l_{{_{2}}}& = l-l_{{_{3}}}\\
\frac{l_{{_{1}}}}{{b_{\cup_{{_{{D}}}}}}^2}+l_{{_{2}}}& = \frac{f l_{{_{3}}}}{l_{{_{3}}}-f}.\end{aligned}$$
Spectral window of imaging
==========================
The spectral window of the imaging of the echo spectrometer is defined by the bandwidths and their relative shifts of the defocusing and refocusing systems. The shape of the window of imaging can be measured by measuring the elastically scattered signal and scanning one bandwidth against another. The window of imaging can be shifted by shifting one of the bandwidths against another.
[^1]: The beam size $\Delta x$, the angular spread $\Delta \xi$, and the energy spread $\Delta E$ are obtained by the propagation of second-order statistical moments, using transport matrices derived from the ray-transfer matrices, and assuming zero cross-correlations (i.e., zero mixed second-order moments).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider reflecting random walks on the nonnegative integers with drift of order $1/x$ at height $x$. We establish explicit asymptotics for various probabilities associated to such walks, including the distribution of the hitting time of 0 and first return time to 0, and the probability of being at a given height $k$ at time $n$ (uniformly in a large range of $k$.) In particular, for drift of form $-\delta/2x + o(1/x)$ with $\delta>-1$, we show that the probability of a first return to 0 at time $n$ is asymptotically $n^{-c}\varphi(n)$, where $c = (3+\delta)/2$ and $\varphi$ is a slowly varying function given in terms of the $o(1/x)$ terms.'
address: |
Department of Mathematics KAP 108\
University of Southern California\
Los Angeles, CA 90089-2532 USA
author:
- 'Kenneth S. Alexander'
title: 'Excursions and local limit theorems for Bessel-like random walks'
---
[^1]
Introduction
============
We consider random walks on ${\mathbb{Z}}_+ = \{0,1,2,\dots\}$, reflecting at 0, with steps $\pm 1$ and transition probabilities of the form $$\label{Bessellike}
p(x,x+1) = p_x = {\frac{1}{2}}\left( 1 - \frac{\delta}{2x} + o\left( \frac{1}{x} \right) \right) \ \text{as } x \to \infty, \quad p(x,x-1) = q_x
= 1-p_x,$$ for $x \geq 1$. We call such processes *Bessel-like walks*, as their drift is asymptotically the same as that of a Bessel process of (possibly negative) dimension $1-\delta$. We call $\delta$ the *drift parameter*. Bessel-like walks are a special case of what is called the Lamperti problem—random walks with asymptotically zero drift. A Bessel-like walk is recurrent if $\delta > -1$, positive recurrent if $\delta>1$, and transient if $\delta < -1$; for $\delta=-1$ recurrence or transience depends on the $o(1/x)$ terms. Here we consider the recurrent case, with primary focus on $\delta>-1$, as the case $\delta=-1$ has additional complexities which weaken our results. Bessel-like walks arise for example when (reflecting) symmetric simple random walk (SSRW) is modified by a potential proportional to $\log x$.
Bessel-like walks have been extensively studied since the 1950’s. Hodges and Rosenblatt [@HR53] gave conditions for finiteness of moments of certain passage times, and Lamperti [@La62] established a functional central limit theorem (with non-normal limit marginals) for $\delta<1$; for $-1<\delta<1$ our Theorem \[location\] below is a local version of his CLT. In [@La63] Lamperti related the first and second moments of the step distribution to finiteness of integer moments of first-return-time distributions. He worked with a wider class of Markov chains with drift of order $1/x$, showing in particular that for return times of Bessel-like walks, moments of order less than $\kappa = (1+\delta)/2$ are finite while those of order greater than $\kappa$ are infinite. Lamperti’s results were generalized and extended to noninteger moments in [@AI97], [@AIM96], and to expected values of more general functions of return times in [@AI99]. “Upper and lower” local limit theorems were established in [@MP95] for certain positive recurrent processes which include our $\delta>1$. Bounds for the growth rate of processes with drift of order $1/x$ were given in [@MVW08], and the domain of attraction of the excursion length distribution was examined in [@Fa73].
Karlin and McGregor ([@KM57a], [@KM57b], [@KM59]) showed that, for general birth-death processes, many quantities of interest could be expressed in terms of a family of polynomials orthogonal with respect to a measure on $[-1,1]$. This measure can in principle be calculated (see Section 8 of [@KM57b]) but not concretely enough, apparently, for some computations we will do here. An exception is the case of $p_x = {\frac{1}{2}}(1 - \frac{\delta} {2x+\delta} )$ considered in [@De01] (for $\delta=1$) and [@DDH08]; we will call this the *rational-form case*. Birth-death processes dual to the rational form case were considered in [@Ro66]. Further results for birth-death processes via the Karlin-McGregor representation are in [@CV98], [@Fa81].
Our interest in Bessel-like walks originates in statistical physics. These walks were used in [@DDH09] in a model of wetting. Additionally, in polymer pinning models of the type studied in [@Gi07] and the references therein, there is an underlying Markov chain which interacts with a potential at times of returns to 0. The location of the $i$th monomer is given by the state of the chain at time $i$. There may be quenched disorder, in the form of random variation in the potential as a function of the time of the return. Let $\tau_0$ denote the return time to 0 for the Markov chain started at 0. For many models of interest, e.g. SSRW on ${\mathbb{Z}}^d$, the distribution of $\tau_0$ for the underlying Markov chain has a power-law tail: $$\label{powerlaw}
P(\tau_0 = n) = n^{-c}\varphi(n)$$ for some $c \geq 1$ and slowly varying $\varphi$. Considering even $n$, for $d=1$ one has $c=3/2$ and $\varphi(n)$ converging to $\sqrt{2/\pi}$; for $d=2$ one has $c=1$ and $\varphi(n)$ proportional to $(\log n)^{-2}$ [@JP72]; for $d \geq 3$ one has $c=d/2$ and $\varphi(n)$ asymptotically constant. In general the value of $c$ is central to the critical behavior of the polymer with the presence of the disorder altering the critical behavior for $c>3/2$ but not for $c<3/2$ ([@Al08],[@AZ08],[@GT05].) In the “marginal” case $c=3/2$, the slowly varying function $\varphi$ determines whether the disorder has such an effect [@GLT09]. As we will see, for Bessel-like walks, holds in the approximate sense that $$\label{powerlaw2}
P(\tau_0 = n) \sim n^{-c}\varphi(n) \quad \text{as } n \to \infty,$$ with $c=(3+\delta)/2$ and $\varphi(n)$ determined explicitly by the $o(1/x)$ terms. Here $\sim$ means the ratio converges to 1. Thus Bessel-like walks provide a single family of Markov chains in $(1+1)$-dimensional space-time in which can be realized (at least asymptotically) for arbitrary $c$ and $\varphi$.
A related model is the directed polymer in a random medium (DPRM), in which the underlying Markov chain is generally taken to be SSRW on ${\mathbb{Z}}^d$ and the polymer encounters a random potential at every site, not just the special site 0. The DPRM has been studied in both the physics literature (see the survey [@HZ95]) and the mathematics literature (see e.g. [@CH02], [@CY07], [@La10].) In place of SSRW, one could use a Markov chain on ${\mathbb{Z}}^d$ in which each coordinate is an independent Bessel-like walk. In this manner one could study the effect on the DPRM of the behavior , or more broadly, study the effect of the drift present in the Bessel-like walk. As with the pinning model, via Bessel-like walks, all drifts and all tail exponents $c$ (not just the half-integer values occurring for SSRW) can be studied using the same space of trajectories. This will be pursued in future work.
For the DPRM, an essential feature is the overlap, that is, the value $$\sum_{i=1}^N \delta_{\{X_i=X_i'\}},$$ where $\{X_i\}, \{X_i'\}$ are two independent copies of the Markov chain; see ([@CH02], [@CY07], [@La10].) To determine the typical behavior of the overlap one should know the probabilities $P(X_i=y), y \in {\mathbb{Z}}^d$, as precisely as possible, with as much uniformity in $y$ as possible..
For this paper we thus have two goals: given the transition probabilities $p_x,q_x$ of a Bessel-like walk, determine
- the value $c$ and slowly varying function $\varphi$ for which holds, and
- the probabilities $P(X_i=y), y \in {\mathbb{Z}}$, asymptotically as $i \to \infty$, as uniformly in $y$ as possible.
We will not make use of the methods of Karlin and McGregor ([@KM57a], [@KM57b], [@KM59]) due to the difficulty of calculating the measure explicitly enough, and obtaining the desired uniformity in $y$. Instead we take a more probabilistic approach, comparing the Bessel-like walk to a Bessel process with the same drift, while the walk is at high enough heights. This leads to estimates of probabilities of form $P(\tau_0 \in [a,b])$ when $a/b$ is bounded away from 1. Then to obtain we use special coupling properties of birth-death processes which force regularity on the sequence $\{P(\tau_0 = n), n \geq 1\}$. These properties, given in Lemma \[latticepath\] and Corollary \[latticepath2\], may be of some independent interest.
Main Results
============
Consider a Bessel-like random walk $\{X_n\}$ on the nonnegative integers with drift parameter $\delta \geq -1$, with transition probabilities $p_x = p(x,x+1),q_x=p(x,x-1)=1-p_x$. The walk is reflecting, i.e. $p_0=1$. We assume uniform ellipticity: there exists ${\epsilon}>0$ for which $$\label{elliptic}
p_x,q_x \in [{\epsilon},1-{\epsilon}] \quad \text{for all } x \geq 1.$$ Define $R_x$ by $$\label{pxRx}
p_x = {\frac{1}{2}}\left( 1 - \frac{\delta}{2x} + \frac{R_x}{2} \right),$$ where $R_x = o(1/x)$. Note that in the rational-form case we have $$R_x = \frac{\delta^2}{2x^2} + O\left( \frac{1}{x^3} \right).$$ The drift at $x$ is $$p_x - q_x = 2p_x-1 = -\frac{\delta}{2x} + \frac{R_x}{2}.$$ Let $\lambda_0 = 1, M_0=0$ and for $x \geq 1$, $$\lambda_x = \prod_{k=1}^x \frac{q_k}{p_k}, \quad M_x = \sum_{k=0}^{x-1} \lambda_k,
\quad L(x) = \exp\left( R_1 + \dots + R_x\right).$$ $M_x$ is the scale function. Note $M_1=1$, and $M_{X_n \wedge \tau_0}$ is a martingale. It is easily checked that the assumption $R_x = o(1/x)$ ensures $L$ is slowly varying. By linearly interpolating between integers, we can extend $L$ to a function on $[1,\infty)$ which is still slowly varying. Let $\tau_j$ be the hitting time of $j \in {\mathbb{Z}}_+$, let $P_j$ denote probability for the walk started from height $j$ and let $$\label{Hdef}
H = \max \{X_i: i \leq \tau_0\}$$ be the height of an excursion from 0. From the martingale property we have $$\label{height}
P_0(H\geq h) = P_1(\tau_h<\tau_0) = \frac{M_1}{M_h}$$ so since $M_1=1$, $$P_0(H=h) = \frac{M_1}{M_h} - \frac{M_1}{M_{h+1}} = \frac{\lambda_h}{M_h M_{h+1}}.$$ In place of $\delta$, a more convenient parameter is often $$\kappa = \frac{1+\delta}{2} \geq 0.$$ We have $$\frac{p_x}{q_x} = 1 - \frac{\delta}{x} + R_x + O\left( \frac{1}{x^2} \right),$$ and hence $$\label{lambda}
\lambda_x \sim K_0x^{2\kappa-1}L(x)^{-1} \ \text{as $x \to \infty$, for some } K_0>0,$$ so for $\kappa>0$, $$\label{Mxapprox}
M_x \sim \frac{K_0}{2\kappa} x^{2\kappa} L(x)^{-1}.$$ Our assumption of recurrence is equivalent to $M_x \to \infty$.
Define the slowly varying function $$\nu(n) = \sum_{l \leq n,\ l \text{ even}} \frac{1}{lL(\sqrt{l})}.$$
Throughout the paper, $K_0, K_1,\dots$ are constants which depend only on $\{p_x, x \geq 1\}$, except as noted; for example, $K_i(\theta,\chi)$ means that $K_i$ depends on some previously-specified $\theta$ and $\chi$. Further, to avoid the notational clutter of pervasive integer-part symbols, we tacitly assume that all indices which appear are integers, as may be arranged by slightly modifying various arbitrarily-chosen constants, or more simply by mentally inserting the integer-part symbol as needed.
\[tau0tail\] Assume and . For $\delta>-1$, $$\label{exctail}
P_0(\tau_0 \geq n) \sim \frac{2^{1-\kappa}}{K_0 \Gamma(\kappa)} n^{-\kappa} L(\sqrt{n}) \quad \text{as } n \to \infty,$$ and for $n$ even, $$\label{excpoint}
P_0(\tau_0 = n) \sim \frac{2^{2-\kappa}\kappa}{K_0 \Gamma(\kappa)} n^{-(\kappa+1)} L(\sqrt{n}).$$ For $\delta = -1$, assuming recurrence (i.e. $M_x \to \infty$ as $x \to \infty$), $$\label{exctailsv}
P_0(\tau_0 \geq n) \sim \frac{1}{K_0 \nu(n)}.$$
For the case of SSRW, in contrast to , the excursion length distribution is easily given exactly [@Fe68]: for $n$ even, $$P_0(\tau_0=n) = \frac{1}{n-1} {n \choose n/2} 2^{-n} \sim \frac{1}{2\sqrt{\pi}}n^{-3/2}.$$
By we have for fixed $\eta \in (0,1)$ that $$\label{exctail2}
P_0\big((1-\eta)n \leq \tau_0 \leq (1+\eta) n \big)
\sim \frac{ 2^{2-\kappa} }{ K_0 \Gamma(\kappa) } \eta \Upsilon(\eta) n^{-\kappa}L(\sqrt{n}),$$ where $$\label{psidef}
\Upsilon(\eta) = \frac{1}{2\eta} \left( (1-\eta)^{-\kappa} - (1+\eta)^{-\kappa} \right) \to \kappa \quad \text{as } \eta \to 0.$$ Heuristically, one expects that conditionally on the event on the left side of , $\tau_0$ should be approximately uniform over even numbers in the interval $[(1-\eta)n,(1+\eta) n]$, leading to . The precise statement we use is Lemma \[convexity\].
It follows from , and Theorem \[tau0tail\] that $\tau_0$ and $H^2$ have asymptotically the same tail, to within a constant: $$\label{sametail}
P_0(H^2 \geq n) \sim 2^\kappa \kappa \Gamma(\kappa) P_0(\tau_0 \geq n)
\sim P_0\left( 2 (\kappa \Gamma(\kappa))^{1/\kappa} \tau_0 \geq n \right) \quad \text{as } n \to \infty.$$ This says roughly that the typical height of an excursion becomes a large multiple of the square root of its length (i.e. duration), as $\kappa$ grows, meaning the downward drift becomes stronger. In this sense the random walk climbs higher to avoid the strong drift.
By reversing paths we see that $$\label{reverse}
P_k(X_n = 0) = p_k \lambda_k P_0(X_n = k).$$ Hence to obtain an approximation for $P_0(X_n=k)$, we need an approximation for $P_k(X_n=0)$, and for that we first need an approximation for $P_k(\tau_0 = m)$. In this context, keeping in mind the similarity between $\tau_0$ and $H^2$, for a given constant $\chi<1$ we say that a starting (or ending) height $k$ is *low* if $k < \sqrt{\chi m}$, *midrange* if $\sqrt{m\chi} \leq k \leq \sqrt{m/\chi}$ and *high* if $k > \sqrt{m/\chi}$.
\[hittime\] Suppose $\delta>-1$. Given $\theta>0$, for $\chi>0$ sufficiently small, there exists $m_0(\theta,\chi)$ as follows. For all $m \geq m_0$ and $1 \leq k < \sqrt{\chi m}$ (low starting heights) with $m-k$ even, $$\begin{aligned}
\label{hitapprox}
(1-\theta) \frac{2^{2-\kappa}\kappa}{K_0\Gamma(\kappa)} m^{-(1+\kappa)} L(\sqrt{m}) M_k
&\leq P_k(\tau_0 =m) \\
&\leq (1+\theta) \frac{2^{2-\kappa}\kappa}{K_0\Gamma(\kappa)} m^{-(1+\kappa)} L(\sqrt{m}) M_k. \notag
\end{aligned}$$ For all $\sqrt{m\chi} \leq k \leq \sqrt{m/\chi}$ (midrange starting heights) with $m-k$ even, $$\label{hitapprox2}
(1-\theta)\frac{2}{\Gamma(\kappa)m} \left( \frac{k^2}{2m} \right)^\kappa e^{-k^2/2m}
\leq P_k(\tau_0 =m)
\leq (1+\theta)\frac{2}{\Gamma(\kappa)m} \left( \frac{k^2}{2m} \right)^\kappa e^{-k^2/2m}.$$ For all $k > \sqrt{m/\chi}$ (high starting heights) with $m-k$ even, $$\label{hitapprox3}
P_k(\tau_0 =m) \leq \frac{1}{m} e^{-k^2/8m}.$$
In general, for high starting heights, as in we accept upper bounds, rather than sharp approximations as in and .
Note that by , when $k$ is large and differ only in the factor $e^{-k^2/2m}$, which is near 1 for low starting heights. (Here “large” does not depend on $m$.) Further, by , one can replace with $$\begin{aligned}
\label{hitapprox4}
(1-\theta) P_0(\tau_0 = m) M_k \leq P_k(\tau_0 =m) \leq (1+\theta)P_0(\tau_0 = m) M_k.
\end{aligned}$$
We will see below that the left and right sides of represent approximately the probabilities for a Bessel process, with the same drift parameter $\delta$ and starting height $k$, to hit 0 in $[m-1,m+1]$. But the Bessel approximation is not necessarily valid for low starting heights, where holds, because the analog of $M_k$ for the Bessel process may be quite different from its value for the Bessel-like RW, and because $L(\sqrt{m})/L(k)$ need not be near 1, whereas the analog of $L(\cdot)$ for the Bessel process is a constant. Even if a RW has asymptotically constant $L(\cdot)$, the constant $K_0$ may be different from the related Bessel case.
From , for midrange starting heights the distribution of $\tau_0$ is nearly the same as for the approximating Bessel process. For low starting heights, this is not true in general—the Bessel-like RW in this case will typically climb to a height of order $\sqrt{m}$ for paths with $\tau_0 = m$, and this climb is what is affected by the dissimilarity between the two processes, as reflected in the errors $R_x$.
If $\delta>1$ (i.e. $\kappa>1$), or if $\delta=1$ and $E_0(\tau_0) < \infty$, then $$\label{finitemean3}
P_0(X_n=0) \to \frac{2}{E_0(\tau_0)} \quad \text{as } n \to \infty \quad (n \text{ even}),$$ and of course when it is finite, $E_0(\tau_0)$ can be expressed explicitly in terms of the transition probabilities $p_x$ and $q_x$, by using reversibility. If $-1<\delta<1$ (i.e. $0<\kappa<1$), then by and a result of Doney [@Do97], $$\label{infinitemean3}
P_0(X_n=0) \sim \frac{2^\kappa K_0}{\Gamma(1-\kappa) } n^{-(1-\kappa)} L(\sqrt{n})^{-1} \quad (n \text{ even}),$$ and if $\delta=1$ (i.e. $\kappa=1$) with $E_0(\tau_0) = \infty$, then by and a result of Erickson [@Er70], $$\label{c2case3}
P_0(X_n=0) \sim \frac{2}{\mu_0(n)} \quad (n \text{ even}),$$ where $\mu_0(n)$ is the truncated mean: $$\mu_0(n) = \sum_{l=1}^n l P_0(\tau_0 = l) \sim \frac{2}{K_0} \sum_{l\leq n,\ l\text{ even}} \frac{L(\sqrt{l})}{l},$$ which is a slowly varying function.
The next theorem, approximating the left side of , is based on Theorem \[hittime\] and —, together with the fact that $$\label{renewal}
P_k(X_n=0) = \sum_{j=0}^n P_k(\tau_0=n-j)P_0(X_j=0).$$
\[distrib\] Given $\theta>0$, for $\chi$ sufficiently small there exists $n_0(\theta,\chi)$ such that for all $n \geq n_0$, the following hold.
\(i) For $k < \sqrt{\chi n}$ (low starting heights) with $n-k$ even, $$\label{lowheight}
(1-\theta)P_0(X_{{\tilde{n}}} = 0) \leq P_k(X_n = 0) \leq (1+\theta)P_0(X_{{\tilde{n}}} = 0),$$ where ${\tilde{n}}= n$ if $n$ is even, ${\tilde{n}}= n+1$ if $n$ is odd.
\(ii) If $E_0(\tau_0) < \infty$ (which is always true for $\delta>1$), then for $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$ (midrange starting heights) with $n-k$ even, $$\begin{aligned}
\label{finitemean2}
\frac{2-\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du &\leq P_k(X_n = 0) \\
&\leq \frac{2+\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du, \notag
\end{aligned}$$ and for $k > \sqrt{n/\chi}$ (high starting heights) with $n-k$ even, $$\begin{aligned}
\label{finitemean4}
P_k(X_n = 0) &\leq \frac{8}{E_0(\tau_0)} e^{-k^2/8n}.\end{aligned}$$
\(iii) If $-1<\delta<1$, then for $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$ (midrange starting heights) with $n-k$ even, $$\begin{aligned}
\label{infinitemean}
\frac{ (1-\theta)2^\kappa K_0 }{ \Gamma(1-\kappa) } n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n}
&\leq P_k(X_n=0) \\
&\leq \frac{ (1+\theta)2^\kappa K_0 }{ \Gamma(1-\kappa) }
n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n}, \notag
\end{aligned}$$ and there exists $K_{1}(\kappa)$ such that for $k > \sqrt{n/\chi}$ (high starting heights) with $n-k$ even, $$\begin{aligned}
\label{infinitemean3a}
P_k(X_n = 0) &\leq K_{1} e^{-k^2/8n} n^{-(1-\kappa)} L(\sqrt{n})^{-1}.\end{aligned}$$
\(iv) If $\delta=1$ and $E_0(\tau_0) = \infty$, then for $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$ (midrange starting heights) with $n-k$ even, $$\begin{aligned}
\label{borderline}
\frac{2-\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du &\leq P_k(X_n = 0) \\
&\leq \frac{2+\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du, \notag
\end{aligned}$$ and for $k > \sqrt{n/\chi}$ (high starting heights) with $n-k$ even, $$\begin{aligned}
\label{borderline3}
P_k(X_n = 0) &\leq \frac{8}{\mu_0(n)} e^{-k^2/8n}.\end{aligned}$$
From [@GJY03], the integral that appears in and is the probability that the approximating Bessel process started at $k$ hits 0 by time $n$.
We may of course replace $P_0(X_n=0)$ with the appropriate approximation from —, in .
We now combine with Theorem \[distrib\] to approximate the left side of .
\[location\] Given $\theta>0$, for $\chi>0$ sufficiently small, there exists $n_0(\theta,\chi)$ such that for all $n \geq n_0$, the following hold.
\(i) For $1 \leq k < \sqrt{\chi n}$ (low ending heights) with $n-k$ even, $$\label{lowheight5}
\frac{1-\theta}{\lambda_k p_k} P_0(X_n = 0) \leq P_0(X_n = k) \leq \frac{1+\theta}{\lambda_k p_k} P_0(X_n = 0).$$
\(ii) If $E_0(\tau_0) < \infty$ (which is always true for $\delta>1$), then for $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$ (midrange ending heights) with $n-k$ even, $$\begin{aligned}
\label{finitemean5}
(1-\theta)&\frac{4}{K_0 E_0(\tau_0)} k^{1-2\kappa} L(k) \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \\
&\leq P_0(X_n = k) \leq (1+\theta)\frac{4}{K_0 E_0(\tau_0)} k^{1-2\kappa} L(k) \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)}
u^{\kappa-1} e^{-u}\ du, \notag
\end{aligned}$$ and for $k > \sqrt{n/\chi}$ (high ending heights) with $n-k$ even, $$\label{finitemean6}
P_0(X_n = k) \leq\frac{32}{K_0 E_0(\tau_0)} k^{1-2\kappa} L(k) e^{-k^2/8n}.$$
\(iii) If $-1<\delta<1$, then for $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$ (midrange ending heights) with $n-k$ even, $$\begin{aligned}
\label{infinitemean5}
&(1-\theta)\frac{ 2^{\kappa+1}}{ \Gamma(1-\kappa) } \left( \frac{k}{\sqrt{n}}
\right)^{1-2\kappa} e^{-k^2/2n} n^{-1/2} \\
&\qquad \qquad \leq P_0(X_n = k) \leq (1+\theta) \frac{ 2^{\kappa+1}}{ \Gamma(1-\kappa) }
\left( \frac{k}{\sqrt{n}} \right)^{1-2\kappa} e^{-k^2/2n} n^{-1/2}, \notag
\end{aligned}$$ and for $k > \sqrt{n/\chi}$ (high ending heights) with $n-k$ even, for $K_{1}$ of , $$\label{infinitemean6}
P_0(X_n = k) \leq \frac{4K_{1}}{K_0} e^{-k^2/8n} n^{-1/2}.$$
\(iv) If $\delta=1$ and $E_0(\tau_0) = \infty$, then for $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$ (midrange ending heights) with $n-k$ even, $$\begin{aligned}
\label{borderline5}
(1-\theta)&\frac{4}{K_0 \mu_0(n)} \frac{L(k)}{k} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \\
&\leq P_0(X_n = k) \leq (1+\theta)\frac{4}{K_0 \mu_0(n)} \frac{L(k)}{k} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du, \notag
\end{aligned}$$ and for $k > \sqrt{n/\chi}$ (high ending heights) with $n-k$ even, $$\label{borderline6}
P_0(X_n = k) \leq\frac{44}{K_0\mu_0(n)} \frac{L(k)}{k} e^{-k^2/8n}.$$
A version of for the RW dual to the rational-form case, with $\delta=-1$, was proved in [@Ro66], with the statement that the proof works for general $\delta<1$.
For large $k$ we can use the approximation in . For example, in the case $-1<\delta<1$, there exists $k_1(\theta)$ such that for $n \geq n_0$ and $k_1 \leq k < \sqrt{\chi n}$ we have $$\begin{aligned}
\label{uselambda}
(1-\theta)& \frac{2^{2-\kappa}}{\Gamma(1-\kappa)} n^{-(1-\kappa)} k^{-\delta} \frac{L(k)}{ L(\sqrt{n}) } \\
&\leq P_0(X_n = k) \leq (1+\theta) \frac{2^{2-\kappa}}{\Gamma(1-\kappa)} n^{-(1-\kappa)} k^{-\delta} \frac{L(k)}{ L(\sqrt{n}) }. \notag
\end{aligned}$$
We can use Theorem \[location\] to approximately describe the distribution of $X_n$ only because its statement gives uniformity in $k$. This requires uniformity in $k$ in Theorems \[hittime\] and \[distrib\], which points us toward our probabilistic approach.
The factors 8 in the exponent in , and is not sharp. For $-2<\delta<0$, bounds on tail (not point) probabilities with sharper exponents are established in [@BRS71].
We are unable to extend our results to random walks with drift which is asymptotically 0 but not of order $1/x$, because we rely on known properties of the Bessel process.
Coupling {#Coupling}
========
Let us consider the random walk with steps $\pm 1$ imbedded in a Bessel process $Y_t \geq 0$ with drift $-\delta/2Y_t$: $$dY_t = -\frac{\delta}{2Y_t}\ dt + dB_t,$$ where $B_t$ is Brownian motion. (We need only consider this process until the time, if any, that it hits 0, which avoids certain technical complications.) The imbedded walk is defined in the standard way: we start both the RW and the Bessel process at the same integer height $k$. The first step of the RW is to $k \pm 1$, whichever the Bessel process hits first, at some time $S_1$. The second step is to $Y_{S_1} \pm 1$, whichever the Bessel process hits first starting from time $S_1$, and so on.
Let $g(x) = x^{1+\delta}$; then $g(Y_t)$ is a martingale, in fact a time change of Brownian motion (see [@RY91].) Write $P^{\operatorname{Be}}$ for probability for the Bessel process, $P^{\operatorname{BI}}$ for the imbedded RW and $P^{\operatorname{sym}}$ for symmetric simple random walk (not reflecting at 0.) For the imbedded RW, for $x \geq 1$, the downward transition probability is $$q_x^{\operatorname{BI}} = P_x^{\operatorname{Be}}(\tau_{x-1} < \tau_{x+1}) = \frac{g(x+1) - g(x)}{g(x+1)-g(x-1)}
= {\frac{1}{2}}\left(1 + \frac{\delta}{2x} + \frac{\delta^2(1-\delta)}{12x^3} + O\left(\frac{1}{x^4}\right) \right)$$ so the corresponding value of $R_x$ is $$R_x^{\operatorname{BI}} = -\frac{\delta^2(1-\delta)}{6x^3} + O\left(\frac{1}{x^4}\right).$$ We write $\{X_n\}$, $\{X_n^{\operatorname{BI}}\}$ and $\{X_n^{\operatorname{sym}}\}$ for the Bessel-like RW, imbedded RW, and symmetric simple RW, respectively, and $\tau_j,\tau_j^{\operatorname{BI}},\tau_j^{\operatorname{sym}}$ for the corresponding hitting times.
Here is a special construction of $\{X_n\}$ that couples it to $\{X_n^{\operatorname{sym}}\}$, when $p_x \leq q_x$ for all $x$. (A similar construction works in case $p_x \geq q_x$ for all $x$.) Let $\xi_0,\xi_1,\dots$ be i.i.d. uniform in \[0,1\]. For each $i \geq 0$ we have an alarm independent of $\xi_i$. If $X_i=x$, the alarm sounds with probability $q_x - p_x = \frac{\delta}{2x} - \frac{R_x}{2}$. If there is no alarm, $X_{i+1} = x+1$ if $\xi_i > 1/2$, and $X_{i+1}=x-1$ if $\xi_i \leq 1/2$. If the alarm sounds, then $X_{i+1}=x-1$, regardless of $\xi_i$. $\{X_n^{\operatorname{sym}}\}$ ignores the alarm and always takes its step according to $\xi_i$.
A second special construction, coupling $\{X_n\}$ to $\{X_n^{\operatorname{BI}}\}$, is as follows; a related coupling appears in [@CFR08]. If $X_i=x$, the alarm sounds independently with probability $a(x)$ given by $$a(x) = \begin{cases} \frac{p_x - p_x^{\operatorname{BI}}}{q_x^{\operatorname{BI}}} = \frac{R_x}{2} + \frac{\delta^2(1-\delta)}{12x^3}
+ O\left( \frac{|R_x|}{x} + \frac{1}{x^4} \right) \quad &\text{if } p_x \geq p_x^{\operatorname{BI}}, \\
\frac{q_x - q_x^{\operatorname{BI}}}{p_x^{\operatorname{BI}}} = -\frac{R_x}{2} - \frac{\delta^2(1-\delta)}{12x^3}
+ O\left( \frac{|R_x|}{x} + \frac{1}{x^4} \right) \quad &\text{if } p_x < p_x^{\operatorname{BI}}.
\end{cases}$$ Whenever the alarm sounds, $\{X_i\}$ takes a step up in the case $p_x \geq p_x^{\operatorname{BI}}$, and down in the case $p_x < p_x^{\operatorname{BI}}$. If there is no alarm, $\{X_n\}$ goes up if $\xi_i > q_x^{\operatorname{BI}}$ and down if $\xi_i \leq q_x^{\operatorname{BI}}$. By contrast, $\{X_n^{\operatorname{BI}}\}$ ignores the alarm and always takes its step according to $\xi_i$. Under this construction, if $p_x \geq p_x^{\operatorname{BI}}$, the probability of an up step for $\{X_i\}$ from $x$ is $$(1-a(x))p_x^{\operatorname{BI}} + a(x) \cdot 1 = p_x,$$ and if $p_x < p_x^{\operatorname{BI}}$, the probability of a down step for $\{X_i\}$ is $$(1-a(x))q_x^{\operatorname{BI}} + a(x) \cdot 1 = q_x,$$ which shows that this second construction does indeed couple $\{X_n\}$ to $\{X_n^{\operatorname{BI}}\}$. Note that in the second construction, unlike the first, the frequency of alarms is $o(1/x)$. The coupling to $\{X_n^{\operatorname{BI}}\}$ is more complicated because the transition probabilities for $\{X_n^{\operatorname{BI}}\}$ depend on location. Even when no alarm sounds, the two walks may take opposite steps if $X_i=x$, $X_i^{\operatorname{BI}}=y$ and $\xi_i$ falls between $q_x^{\operatorname{BI}}$ and $q_y^{\operatorname{BI}}$. When (i) there is no alarm, (ii) $X_i=x, X_i^{\operatorname{BI}}=y$ for some $x,y$, and (iii) $\xi_i$ falls between $q_x^{\operatorname{BI}}$ and $q_y^{\operatorname{BI}}$, we say a [*discrepancy*]{} occurs at time $i$. A [*misstep*]{} means either an alarm or a discrepancy. For $h$ sufficiently large, for $x \geq h, y \geq h$, conditioned on $X_i=x, X_i^{\operatorname{BI}}=y$ and no alarm, the probability of a discrepancy is $$\label{discrep}
|q_x^{\operatorname{BI}} - q_y^{\operatorname{BI}}| \leq \frac{\delta}{2h^2}|x-y|.$$ We let $N(k)$ denote the number of missteps which occur up to time $k$.
Note that if $\delta=0$, the imbedded RW is symmetric and there are no discrepancies.
When we couple $\{X_n\}$ and $\{X_n^{\operatorname{BI}}\}$ in the above manner, with both processes starting at $k$, we denote the corresponding measure by $P_k^*$. Where confusion seems possible, for hitting times we then use a superscript to designate the process that the hitting time refers to, e.g. $\tau_0^{\operatorname{Be}}$ and $\tau_0^{\operatorname{BI}}$ for the Bessel process and its imbedded RW, respectively.
Proof of the tail approximation
================================
Recall that for we have $\delta>-1$. Let $\theta>0$, $0<\rho<1/8$, $0 < {\epsilon}_1 < {\epsilon}_2 < \sqrt{\rho}$ and $h_i = {\epsilon}_i \sqrt{m}$. Let $0 < \eta < {\epsilon}_1/4$ and $h_{1\pm} = ({\epsilon}_1 \pm 2\eta)\sqrt{m}$. To prove we will show that provided $\rho,\theta$ are sufficiently small, one can choose the other parameters so that the following sequence of six inequalities holds, for large $m$: $$\begin{aligned}
\label{toprove}
\frac{1-3\theta}{M_{h_2}} &P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1+2\rho)m\big) \\
&\leq \frac{1-\theta}{M_{h_2}} P_{h_2}^{\operatorname{BI}}(\tau_{h_{1+}} \geq m) \notag \\
&\leq \frac{1}{M_{h_2}} P_{h_2}(\tau_{h_1} \geq m) \notag \\
&\leq P_0(\tau_0 \geq m) \notag \\
&\leq \frac{1+\theta}{M_{h_2}} P_{h_2}(\tau_{h_1} \geq (1 - 2\rho)m) \notag \\
&\leq \frac{1+2\theta}{M_{h_2}} P_{h_2}^{\operatorname{BI}}\big(\tau_{h_{1-}} \geq (1-2\rho)m\big) \notag \\
&\leq \frac{1+4\theta}{M_{h_2}} P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1-3\rho)m\big). \notag\end{aligned}$$ These may be viewed as three “sandwich” bounds on $P_0(\tau_0 \geq m)$, with the outermost sandwich readily yielding the desired result, as we will show. The innermost sandwich (the 3rd and 4th inequalities) may be interpreted as follows. For convenience we assume the $h_i$ are even integers. Recall $H$ from ; when $H \geq h_2$, we let $T$ denote the first hitting time of $h_1$ after $\tau_{h_2}$. We can decompose an excursion of height at least $h_2$ and length at least $m$ into 3 parts: 0 to $\tau_{h_2}$, $\tau_{h_2}$ to $T$, and $T$ to the end. The idea is that for a typical excursion of length at least $m$, most of the length $\tau_0$ of the full excursion will be in the middle interval $[\tau_{h_2},T]$; the first and last intervals will have length at most $\rho m$. The middle sandwich (2nd and 5th inequalities) comes from approximating the original RW by the imbedded RW from a Bessel process, during the interval $[\tau_{h_2},T]$. Then the outermost sandwich (1st and 6th inequalities) comes from approximating the imbedded RW by the actual Bessel process, and from showing that the third interval, from $T$ to excursion end, is typically relatively short.
A useful inequality is as follows: for $h>k \geq 0$ and $m \geq 1$, $$\begin{aligned}
\label{excdecomp2}
P_0&(\tau_0 \geq m, H \geq h) \geq P_0\left( \tau_h < \tau_0 \right) P_{h}(\tau_k \geq m)
= \frac{1}{M_h} P_h(\tau_k \geq m).
\end{aligned}$$ As a special case we have $$\begin{aligned}
\label{excdecomp}
P_0(\tau_0 \geq m) \geq P_0&(\tau_0 \geq m, H \geq h_2) \geq \frac{1}{M_{h_2}} P_{h_2}(\tau_{h_1} \geq m),
\end{aligned}$$ which establishes the 3rd inequality in .
By there exists $l_1 \geq 1$ such that for all $x \geq l_1$, $$x|R_x| \leq {\frac{1}{2}}, \quad \frac{2\kappa M_x}{K_0x^{2\kappa}L(x)^{-1}}
\in \left( \frac{7}{8},\frac{9}{8} \right), \quad \frac{2\kappa(M_{2x}-M_x)}{K_0(2^{2\kappa}-1)x^{2\kappa}L(x)^{-1}}
\in \left( \frac{7}{8},\frac{9}{8} \right),$$ If $\delta\neq 0$, enlarging $l_1$ if necessary, we also have $$\left| x(2p_x-1) + \frac{\delta}{2} \right| < \frac{|\delta|}{4}.$$
We turn to the 4th inequality in . We have $$\label{Hsplit}
P_0(\tau_0 \geq m) = P_0(\tau_0 \geq m, H \geq h_2) + P_0(\tau_0 \geq m, H<h_2).$$ The main contribution should come from the first probability on the right. To show this, we first need two lemmas. We begin with the following bound on strip-confinement probabilities.
\[strip\] Assume and . There exists $K_{2}({\epsilon},l_1)$ as follows. For all $h \geq 1, m \geq 2h^2$ and $0<q<h$, $$P_q(X_n \in (0,h) \text{ for all } n \leq m) \leq e^{-K_{2} m/h^2}.$$
Consider first $\delta \neq 0,\ h > l_1$. We claim that $$P_q(X_n \in (l_1,h) \text{ for all } n \leq h^2 - l_1)$$ is bounded away from 1 uniformly in $q,h$ with $l_1 \leq q < h$. In fact, from the definition of $l_1$, the drift $p_x - q_x$ has constant sign for $x \geq l_1$. Suppose the drift is positive; then $\{X_n\}$ and $\{X_n^{\operatorname{sym}}\}$ can be coupled so that $X_n \geq X_n^{\operatorname{sym}}$ for all $n$ up to the first exit time of $\{X_n\}$ from $(l_1,h)$. Therefore $$P_q(X_n \in (l_1,h) \text{ for all } n \leq h^2 - l_1) \leq P_q^{\operatorname{sym}}(\tau_h > h^2 - l_1) \leq 1 - P_0^{\operatorname{sym}}(\tau_h \leq h^2 - l_1).$$ Since $X_n^{\operatorname{sym}}$ is a non-reflecting symmetric RW, for $Z$ a standard normal r.v. we have $$P_0^{\operatorname{sym}}(\tau_h \leq h^2 - l_1) \geq P_0^{\operatorname{sym}}(\tau_h \leq h^2/2) \geq
P_0^{\operatorname{sym}}(X_{\lfloor h^2/2 \rfloor}^{\operatorname{sym}} \geq h) \to P(Z > \sqrt{2})$$ as $h \to \infty$, so $P_0^{\operatorname{sym}}(\tau_h \leq h^2 - l_1)$ is bounded away from 0 uniformly in $h>l_1$, and the claim follows. Similarly if the drift is negative, we can couple so that $X_n \leq X_n^{\operatorname{sym}}$ until the time that $\{X_n\}$ hits $l_1$, and therefore $$P_q(X_n \in (l_1,h) \text{ for all } n \leq h^2 - l_1) \leq P_q^{\operatorname{sym}}(\tau_{l_1} > h^2-l_1) \leq 1 - P_h^{\operatorname{sym}}(\tau_0 \leq h^2-l_1),$$ and the claim again follows straightforwardly. Then since $q_x \geq {\epsilon}$ for all $x \leq l_1$, we have $$\begin{aligned}
\label{belowl1}
P_q(X_n \notin (0,h) \text{ for some } n \leq h^2) &\geq {\epsilon}^{l_1}P_q(X_n \notin (l_1,h) \text{ for some } n \leq h^2 - l_1),
\end{aligned}$$ which together with the claim shows that there exists $\gamma = \gamma(l_1,{\epsilon})$ such that for all $l_1 \leq q < h$ we have $$\label{shortstrip}
P_q(X_n \notin (0,h) \text{ for some } n \leq h^2) \geq \gamma.$$ Therefore by straightforward induction, since $m \geq 2h^2$, $$\label{stripbound}
P_q(X_n \in (0,h) \text{ for all } n \leq m) \leq (1-\gamma)^{\lfloor m/h^2 \rfloor} \leq e^{-K_{2} m/h^2},$$ completing the proof for $\delta \neq 0,\ h>l_1$.
For $\delta \neq 0, h \leq l_1$, the left side of is bounded below by ${\epsilon}^{l_1}$, and follows similarly.
For $\delta=0$, it seems simplest to proceed by comparison. Instead, in place of we have $$\label{delta0}
P_q(X_n \notin (0,h) \text{ for some } n \leq h^2) \geq P_q(\tau_0 \leq q^2 + 1).$$ We can change the value of the (downward) drift parameter from $\delta=0$ to $\tilde{\delta} \in (-1,0)$ by subtracting $\tilde{\delta}/4x$ from $p_x$ for each $x \geq 1$. By an obvious coupling, this reduces the probability on the right side of . But by Proposition \[intervalmid\] below, this reduced probability is bounded away from 0 in $q \geq 1$. Thus and then hold in this case as well.
It should be pointed out that the proof of Proposition \[intervalmid\] makes use of Theorem \[tau0tail\] which in turn makes use of Lemma \[strip\]. Since the application of Proposition \[intervalmid\] in the proof of Lemma \[strip\] is only for $\tilde{\delta} \neq 0$, and since this application is only used to prove the lemma in the case $\delta=0$, this is not circular—all proofs can be done for nonzero drift parameter first, and then this can be applied to obtain the result for 0 drift parameter.
If we start the RW at 0, we can strengthen the bound in Lemma \[strip\], as follows. Let $Q_n = \max_{0 \leq k \leq n} X_k$, so $H = Q_{\tau_0}$.
\[lowheightlemma\] Assume $\delta>-1$. There exist $K_{3}({\epsilon},l_1),K_{4}({\epsilon},l_1)$ as follows. For all $h > l_1$ and $m \geq 4h^2$, $$P_0(X_n \in (0,h) \text{ for all } 1 \leq n \leq m) \leq \frac{K_{3}}{M_h} e^{-K_{4} m/h^2}.$$
Let $k_1 = \min\{k: 2^{k-2} > l_1\}$ and $k_2 = \max\{k: 2^{k-1} < h\}$. Then for some constants $K_i({\epsilon},l_1)$, [$$\begin{aligned}
\label{excdecomp00a}
P_0&(X_n \in (0,h) \text{ for all } 1 \leq n \leq m) \notag \\
&\leq P_0\left( X_n \in (0,2^{k_1-1}) \text{ for all } 1 \leq n \leq m \right)
+ \sum_{k=k_1}^{k_2} P_0\big(Q_m \in [2^{k-1},2^k), \tau_0 > m \big) \notag \\
&\leq e^{-K_{5}m} +
\sum_{k=k_1}^{k_2} \bigg[ P_0\big(Q_m \in [2^{k-1},2^k), \tau_0 > m, \tau_{2^{k-2}} \leq \frac{m}{2} \big) \notag \\
&\qquad \qquad \qquad \qquad + P_0\big(Q_m \in [2^{k-1},2^k), \tau_0 > m, \tau_{2^{k-2}} > \frac{m}{2} \big) \bigg] \notag \\
&\leq e^{-K_{5}m} + \sum_{k=k_1}^{k_2} \bigg[ P_0\left( \tau_{2^{k-2}} \leq \frac{m}{2}, \tau_0 > \tau_{2^{k-1}} \right)
P_{2^{k-2}}\left( X_n \in (0,2^k) \text{ for all } n \leq \frac{m}{2} \right) \notag \\
&\qquad \qquad \qquad \qquad + P_0\left( \tau_0 > \tau_{2^{k-1}} > \tau_{2^{k-2}} > \frac{m}{2} \right) \bigg] \notag \\
&\leq e^{-K_{5}m} + \sum_{k=k_1}^{k_2} \bigg[ P_1\left( \tau_0 > \tau_{2^{k-1}} \right)
e^{-K_{2} m/2^{2k+1}} \notag \\
&\qquad \qquad \qquad \qquad + \frac{1}{ p_{2^{k-1}} \lambda_{2^{k-1}} }
P_{2^{k-1}}\left( \tau_0 < \tau_{2^{k-1}}, \tau_0 - \tau_{2^{k-2}} > \frac{m}{2} \right) \bigg] \notag \\
&\leq e^{-K_{5}m} + \sum_{k=k_1}^{k_2} \bigg[ \frac{1}{M_{2^{k-1}}} e^{-K_{2} m/2^{2k+1}} \notag \\
&\qquad \qquad + \frac{1}{ p_{2^{k-1}} \lambda_{2^{k-1}} }
P_{2^{k-1}}\left( \tau_{2^{k-2}} < \tau_{2^{k-1}} \right) P_{2^{k-2}}\left( X_n \in (0,2^{k-1})
\text{ for all } n \leq \frac{m}{2} \right) \bigg] \notag \\
&\leq e^{-K_{5}m} + \sum_{k=k_1}^{k_2} \bigg[ \frac{1}{M_{2^{k-1}}} e^{-K_{2} m/2^{2k+1}}
+ \frac{1}{ p_{2^{k-1}} \lambda_{2^{k-1}} }
\frac{ q_{2^{k-1}}( M_{2^{k-1}} - M_{2^{k-1}-1} ) }{ M_{2^{k-1}} - M_{2^{k-2}} } e^{-K_{2} m/2^{2k-1}} \bigg] \\
&\leq e^{-K_{5}m} + \sum_{k=k_1}^{k_2} \bigg[ \frac{1}{M_{2^{k-1}}} + \frac{1}{ M_{2^{k-1}} - M_{2^{k-2}} } \bigg]
e^{-K_{2} m/2^{2k+1}}\notag \\
&\leq e^{-K_{5}m} + K_{6} \sum_{k=k_1}^{k_2} \frac{ L(2^k) }{ 2^{2k\kappa} } e^{-K_{2} m/2^{2k+1}} \notag \\
&\leq e^{-K_{5}m} + K_{7} h^{-2\kappa} L(h) e^{-K_{2} m/8h^2} \notag \\
&\leq K_{8} h^{-2\kappa} L(h) e^{-K_{9} m/h^2}, \notag
\end{aligned}$$ ]{} and the lemma follows from this and . Here in the 2nd inequality we used the ellipticity condition , in the 4th inequality we used Lemma \[strip\] and reversal of the path from time 0 to time $\tau_{2^{k-1}}$, in the 5th inequality we used , in the 6th inequality we used Lemma \[strip\], in the 8th inequality we used , and in the last three inequalities we used the fact that $L$ is slowly varying.
We return to the proof of the 4th inequality in . We have for $m$ sufficiently large that $$\begin{aligned}
\label{excdecomp0}
P_0&(\tau_0 \geq m, H \geq h_2) \\
&\leq P_0\left( \tau_{h_2} < \tau_0 \right)
P_{h_2}(\tau_{h_1} \geq (1 - 2\rho)m) \notag \\
&\qquad + P_0\left( \rho m < \tau_{h_2} < \tau_0 \right)
+ P_0\left( \tau_{h_2} < \tau_0 \right) P_{h_1}(\tau_0 > \rho m) \notag \\
&\leq \frac{1}{M_{h_2}} P_{h_2}(\tau_{h_1} \geq (1 - 2\rho)m) + P_0\left( \rho m < \tau_{h_2} < \tau_0 \right) \notag \\
&\qquad + \frac{1}{M_{h_2}} \left[ P_{h_1}(\tau_{h_2} < \tau_0) + P_{h_1}(\rho m < \tau_0 < \tau_{h_2}) \right] \notag \\
&\leq \frac{1}{M_{h_2}} P_{h_2}(\tau_{h_1} \geq (1 - 2\rho)m)
+ P_0\left( X_n \in (0,h_2) \text{ for all } 1 \leq n \leq \rho m \right) \notag \\
&\qquad + \frac{1}{M_{h_2}} \left[ \frac{M_{h_1}}{M_{h_2}}
+ P_{h_1}( X_n \in (0,h_2) \text{ for all } n \leq \rho m ) \right] \notag \\
&\leq \frac{1}{M_{h_2}} P_{h_2}(\tau_{h_1} \geq (1 - 2\rho)m) + \frac{K_{3}}{M_{h_2}} e^{-K_{4}\rho/{\epsilon}_2^2}
+ \frac{2}{M_{h_2}} \left( \frac{{\epsilon}_1}{{\epsilon}_2} \right)^{2\kappa}
+ \frac{1}{M_{h_2}} e^{-K_{2}\rho/{\epsilon}_2^2} \notag \\
&= (I) + (II) + (III) + (IV). \notag
\end{aligned}$$ The 4th inequality in uses and Lemmas \[strip\] and \[lowheightlemma\]. We want to show that $(II), (III), (IV)$ are much smaller than $(I)$. We will show that if ${\epsilon}_1 \ll {\epsilon}_2$ the probability in $(I)$ is of the same order as $$\label{longorder}
P_{h_2}(\tau_{\sqrt{m}} < \tau_{h_1}) = \frac{M_{h_2} - M_{h_1}}{M_{\sqrt{m}} - M_{h_1}} \sim {\epsilon}_2^{2\kappa}.$$ This means that $(III) \ll (I)$ provided ${\epsilon}_1 \ll {\epsilon}_2^2$.
To complement we have the following bound from Lemma \[lowheightlemma\]: $$\begin{aligned}
\label{excdecomp00}
P_0&(\tau_0 \geq m, H < h_2) \leq P_0\left( X_n \in (0,h_2) \text{ for all } 1 \leq n < m \right) \leq
\frac{K_{3}}{M_{h_2}} e^{-K_{4}/{\epsilon}_2^2}.
\end{aligned}$$ We will later prove the following lower bound for (I).
[**Claim 1.**]{} There exists $K_{10}(\delta)$ such that provided ${\epsilon}_1 < {\epsilon}_2/2$ and $m$ is sufficiently large, we have $$\label{claim1}
P_{h_2}(\tau_{h_1} \geq (1 - 2\rho)m) \geq P_{h_2}\left(\tau_{h_1} \geq m \right) \geq K_{10}{\epsilon}_2^{2\kappa}$$ and $$\label{claim2}
P_{h_2}^{\operatorname{BI}}(\tau_{h_1} \geq (1 - 2\rho)m) \geq P_{h_2}^{\operatorname{BI}}\left(\tau_{h_1} \geq m \right) \geq K_{10}{\epsilon}_2^{2\kappa}.$$
Assuming Claim 1, given $\theta>0$, provided ${\epsilon}_2$ and ${\epsilon}_1/{\epsilon}_2^2$ are sufficiently small (depending on $\delta,\rho,\theta$), the 4th inequality in follows from and .
Our next task is to use the coupling of $\{X_n\}$ to $\{X_n^{\operatorname{BI}}\}$, from Section \[Coupling\], to prove the 2nd and 5th inequaltites in . Here $h_{1\pm}$ should be viewed as substitutes for $h_1$ which allow an error of $\eta\sqrt{m}$ in the coupling construction. Fix $m/2 \leq l \leq m$. We begin with the 5th inequality. From the coupling construction we have $$\begin{aligned}
\label{deviation}
P_{h_2}\big(\tau_{h_1} \geq l \big) &\leq P_{h_2}^{\operatorname{BI}}(\tau_{h_{1-}}^{\operatorname{BI}} \geq l \big)
+ P_{h_2}^*( N(\tau_{h_{1-}}^{\operatorname{BI}}) \geq \eta \sqrt{m}, \tau_{h_{1-}}^{\operatorname{BI}} < l \wedge \tau_{h_1}).
\end{aligned}$$ We need to bound the last probability. Consider first $\delta \neq 0$. Let $A(x) = \sup_{y \geq x} a(y)$, so $A(x) = o(1/x)$, and let $d_0 = h_{1-}^2A(h_{1-})/|\delta|$. Suppose that for some time $i$ and some even integers $d_0 \leq d \leq \eta \sqrt{m}$, the gap $|X_i - X_i^{\operatorname{BI}}| \leq d$ and $X_i^{\operatorname{BI}} \geq h_{1-}$. Provided $h_{1-}$ is large, by the misstep probability for the next step is then at most $$A(h_{1-}) + \frac{|\delta| d}{h_{1-}^2} \leq \frac{2|\delta| d}{h_{1-}^2}.$$ Let $G_{d_0}, G_{d_0+2}, \dots, G_{2\eta \sqrt{m}-2}$ be independent geometric random variables, with $G_d$ having parameter $2|\delta| d/h_{1-}^2$, and $S = G_{d_0} + G_{d_0+2} + \dots + G_{2\eta\sqrt{m}-2}$. The gap $|X_i - X_i^{\operatorname{BI}}|$ can change (always by 2) only at times of missteps. Therefore if we start from the time (if any) before $\tau_{h_{1-}}^{\operatorname{BI}}$ when the gap first reaches $d_0$, the time until the next misstep (if any) before $\tau_{h_{1-}}^{\operatorname{BI}}$ is stochastically larger than $G_{d_0}$, and then the time until the misstep after that (if any) before $\tau_{h_{1-}}^{\operatorname{BI}}$ is stochastically larger than $G_{d_0+2}$, and so on. It follows that $$\label{Nbound}
P_{h_2}^*( N(\tau_{h_{1-}}^{\operatorname{BI}}) \geq \eta \sqrt{m}, \tau_{h_{1-}}^{\operatorname{BI}} < l \wedge \tau_{h_1}) \leq P(S \leq l) \leq P(S \leq m).$$ Note that for $h_{1-}$ large (depending on $\eta/{\epsilon}_1$), $$\frac{E(S)}{m} = \sum_{d_0\leq d < 2\eta\sqrt{m},\atop d-d_0\text{ even}} \frac{h_{1-}^2}{2|\delta| dm}
\geq \frac{h_{1-}^2}{4|\delta| m} \log \frac{\eta \sqrt{m}}{d_0}
\geq \frac{{\epsilon}_1^2}{32|\delta|} \log \frac{|\delta|}{h_{1-}A(h_{1-})},$$ which grows to infinity as $m \to \infty$; thus $E(S) \gg m$. In fact by standard computations using exponential moments, we obtain that for some $K_{11}(\eta,\delta,{\epsilon}_1)$ we have $$\label{Sbound}
P(S \leq m) \leq e^{-K_{11} \sqrt{m}}$$ for all sufficiently large $m$, and hence by Claim 1, $$\label{Sbound2}
P(S \leq m) \leq \frac{\theta}{2} P_{h_2}(\tau_{h_1} \geq l).$$ In the case $\delta=0$, $\{X_n^{\operatorname{BI}}\}$ is a symmetric simple RW so there are no discrepancies, only alarms, which have probability at most $A(h_1)$ when the original RW is above height $h_1$. Hence in place of we have the left side of bounded above by the probability that a Binomial($l,A(h_1)$) exceeds $\eta \sqrt{m}$, and this probability is also bounded by $e^{-K_{11} \sqrt{m}}$, and then the same argument applies. Now , , and show that provided $m$ is large, the 5th inequality in holds.
Turning to the 2nd inequality in , the analog of is still valid, so from the coupling construction, and (trivially modified to allow $h_{1+}$ in place of $h_1$), we have $$\begin{aligned}
\label{deviation2}
P_{h_2}(\tau_{h_1} \geq m) &\geq P_{h_2}^{\operatorname{BI}}(\tau_{h_{1+}}^{\operatorname{BI}} \geq m) -
P_{h_2}^*( N(\tau_{h_1}) \geq \eta \sqrt{m}, \tau_{h_1} < m \wedge \tau_{h_{1+}}^{\operatorname{BI}}) \\
&\geq P_{h_2}^{\operatorname{BI}}(\tau_{h_{1+}} \geq m) - e^{-K_{11}\sqrt{m}} \notag \\
&\geq (1-\theta)P_{h_2}^{\operatorname{BI}}(\tau_{h_{1+}} \geq m), \notag
\end{aligned}$$ proving the desired inequality.
The next step is to prove the first and last inequalities in , by relating the probabilities for $\{X_n^{\operatorname{BI}}\}$ to probabilities for the continuous-time Bessel process $Y_t$. We need to establish the following.
[**Claim 2.**]{} Given $0 < {\epsilon}_1 < {\epsilon}_2, 0<\rho<1/3$ and $\theta>0$, for sufficiently large $m$, $$\label{BesselvsBI}
P_{h_2}^{\operatorname{BI}}\big(\tau_{h_{1-}} \geq (1-2\rho)m\big) \leq (1+\theta) P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1-}} \geq (1-3\rho)m\big),$$ and $$\label{BesselvsBI2}
P_{h_2}^{\operatorname{BI}}(\tau_{h_{1+}} \geq m) \geq (1-\theta) P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1+}} \geq (1+\rho)m\big).$$
Suppose Claim 2 is proved. For the Bessel process we have the obvious inequality $$\label{h1vs0}
P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1-}} \geq (1-3\rho)m\big) \leq P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1-3\rho)m\big),$$ while $$\label{h2h10}
P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1+}} \geq (1+\rho)m\big) \geq P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1+2\rho)m\big)
- P_{h_{1+}}^{\operatorname{Be}}\big(\tau_0 \geq \rho m\big).$$ It follows from (15) in [@GJY03] that for $\delta>-1$ and ${\epsilon}>0$, $$\label{GammaRV}
P_{{\epsilon}\sqrt{t}}^{\operatorname{Be}}(\tau_0 \geq t) = \int_0^{{\epsilon}^2/2} \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du
\sim K_{12} {\epsilon}^{2\kappa} \quad \text{as } {\epsilon}\to 0,$$ where $K_{12} = (2^\kappa \kappa \Gamma(\kappa))^{-1}$. (Strictly speaking this seems to be stated in [@GJY03] only for Bessel processes with dimension in $(0,2)$, i.e. $\delta \in (-1,1)$, but the same proof works for nonpositive dimension, i.e. $\delta \geq 1$. The key is the 3 lines after (57) in Appendix B of [@GJY03].) Applying this to each probability on the right side of we see that for $\rho$ and then ${\epsilon}_1/{\epsilon}_2$ taken sufficiently small and then $m$ large, we have $$P_{h_{1+}}^{\operatorname{Be}}\big(\tau_0 \geq \rho m\big) \leq \theta P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1+2\rho)m\big),$$ and therefore by , $$\label{h1vs0part2}
P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1+}} \geq (1+\rho)m\big) \geq (1-\theta) P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1+2\rho)m\big).$$ Combining and we obtain the first inequality in , while the last inequality in is a consequence of and . This completes the proof of . Since $\rho, \theta$ can be taken arbitrarily small, together with and proves .
[*Proof of Claim 2.*]{} Let $T_0=0$ and let $T_1,T_2,\dots$ be the stopping times when the Bessel process reaches an integer different from the last integer it has visited, so that $X_n^{\operatorname{BI}} = Y_{T_n}$. Denote the hitting times of $h_{1-}$ in the two processes by $\tau_{h_{1-}}^{\operatorname{BI}}$ and $\tau_{h_{1-}}^{\operatorname{Be}}$ and let $\sigma_i = \min\{t: Y_t \in \{i-1,i+1\} \}$. Given $k$ and $x_1,x_2,\dots,x_k$, with $x_i \geq h_{1-}$, let $$A = \{\tau_{h_{1-}}^{\operatorname{BI}}=k\} \cap \{X_0^{\operatorname{BI}} = h_2, X_1^{\operatorname{BI}} = x_1,\dots,X_k^{\operatorname{BI}} = x_k \}.$$ Conditionally on $A$, the random variables $T_i - T_{i-1},\ i \leq k,$ are independent, with the distribution of $T_i - T_{i-1}$ being $$P_{x_{i-1}}^{\operatorname{Be}}\big( \sigma_{x_{i-1}} \in \cdot \mid Y_{\sigma_{x_{i-1}}} = x_i \big).$$ The mean of this distribution is $$\label{mean}
E_{h_2}^{\operatorname{Be}}(T_i - T_{i-1} \mid A) = \frac{ E_{x_{i-1}}^{\operatorname{Be}}\big( \sigma_{x_{i-1}} \delta_{\{ Y_{\sigma_{x_{i-1}}} = x_i \}} \big) }{
P_{x_{i-1}}^{\operatorname{Be}}( Y_{\sigma_{x_{i-1}}} = x_i ) }.$$ We need estimates for the quantities $$E_x^{\operatorname{Be}}(\sigma_x \delta_{\{Y_{\sigma_x} = x - 1\}}), \quad E_x^{\operatorname{Be}}(\sigma_x) \quad
\text{and} \quad P_x^{\operatorname{Be}}(Y_{\sigma_x} = x - 1).$$ Let $$s(x) = \begin{cases} x^{1+\delta} &\text{if } \delta \neq -1,\\ \log x &\text{if } \delta=-1 \end{cases}$$ be the scale function for the Bessel process and let ${\mathcal{L}}f$ given by $$({\mathcal{L}}f)(x) = {\frac{1}{2}}f''(x) - \frac{\delta}{2x}f'(x)$$ be its infinitesmal generator. For fixed $x$ and $z \in [x-1,x+1]$ the functions $f=f_x,g=g_x,h^\pm=h_x^\pm$ given by $$f(z) = P_z^{\operatorname{Be}}(Y_{\sigma_x} = x - 1), \quad g(z) = E_z^{\operatorname{Be}}(\sigma_x), \quad
\frac{h^\pm(z)}{s(x+1)-s(x-1)} = E_z^{\operatorname{Be}}(\sigma_x \delta_{\{Y_{\sigma_x} = x \pm 1\}})$$ satisfy $${\mathcal{L}}f \equiv 0, \quad f(x-1) = 1, \quad f(x+1) = 0;$$ $${\mathcal{L}}g \equiv -1, \quad g(x-1) = g(x+1) = 0;$$ $$({\mathcal{L}}h^+)(z) = s(x-1) - s(z), \quad h^+(x-1) = h^+(x+1) = 0;$$ $$({\mathcal{L}}h^-)(z) = s(z) - s(x+1), \quad h^-(x-1) = h^-(x+1) = 0.$$ These can be solved explicitly, yielding that for $\delta>-1$, $$f(z) = \frac{s(x+1) -s(z)}{s(x+1) - s(x-1)},$$ $$g(z) = \begin{cases} -\frac{1}{1-\delta} z^2 + \frac{4x}{1-\delta}\ \frac{1}{(x+1)^{1+\delta} - (x-1)^{1+\delta}} z^{1+\delta} + A_x
&\text{if } \delta \neq 1,\\
-z^2 \log z + \frac{(x+1)^2 \log (x+1) - (x-1)^2 \log (x-1)}{4x}z^2 +A_x'
&\text{if } \delta=1, \end{cases}$$ $$h^+(z) = \begin{cases} \frac{(x-1)^{1+\delta}}{1-\delta} z^2 - \frac{1}{3+\delta} z^{3+\delta} + B_x z^{1+\delta} + D_x &\text{if } \delta \neq 1,\\
(x-1)^2 z^2 \log z - \frac{1}{4} z^4 + B_x' z^2 + D_x' &\text{if } \delta=1, \end{cases}$$ $$h^-(z) = \begin{cases} -\frac{(x+1)^{1+\delta}}{1-\delta} z^2 + \frac{1}{3+\delta} z^{3+\delta} + B_x'' z^{1+\delta} +
D_x'' &\text{if } \delta \neq 1,\\
-(x+1)^2 z^2 \log z + \frac{1}{4} z^4 + B_x''' z^2 + D_x''' &\text{if } \delta=1. \end{cases}$$ Note the formulas here for $\delta=1$ are determined by the formulas for $\delta\neq 1$, by continuity in $\delta$. Here $B_x$ is given by $$(1-\delta)B_x = -\frac{4x(x-1)^{1+\delta}}{(x+1)^{1+\delta} - (x-1)^{1-\delta}} + \frac{1-\delta}{1+\delta}(x-1)^2\psi_1\left( \frac{2}{x-1} \right)$$ with $$\psi_1(u) = \frac{1+\delta}{3+\delta}\ \frac{(1+u)^{3+\delta}-1}{(1+u)^{1+\delta}-1}
= 1 + u + \frac{2+\delta}{6}u^2 + O(u^3) \quad \text{as } u \to 0,$$ $B_x'$ is given by $$B_x' = {\frac{1}{2}}(x^2+1) - \frac{(x-1)^2(x+1)^2}{4x}\log\left( 1 + \frac{2}{x-1} \right) - (x-1)^2\log(x-1),$$ and $B_x''$ and $B_x''$ are given by $$(1-\delta)B_x'' = \frac{4x(x+1)^{1+\delta}}{(x+1)^{1+\delta} - (x-1)^{1-\delta}} - \frac{1-\delta}{1+\delta}(x-1)^2 \psi_1\left( \frac{2}{x-1} \right)$$ and $$B_x''' = -{\frac{1}{2}}(x^2+1) + \frac{(x-1)^2(x+1)^2}{4x}\log\left( 1 + \frac{2}{x-1} \right) + (x+1)^2 \log(x+1) .$$ Finally, $A_x,A_x'$ and $D_x,D_x'$ and $D_x'',D_x'''$ are determined by $g(x-1)=0,h^+(x-1)=0$ and $h^-(x+1)=0$, respectively, but we do not need these values because we can use for example $g(x) = g(x) - g(x-1)$, and $A_x$ or $A_x'$ cancels in the latter expression. From these computations we readily obtain $$\label{limits}
f(x) \to {\frac{1}{2}}, \quad g(x) \to 1, \quad \frac{h^\pm(x)}{s(x+1)-s(x-1)} \to {\frac{1}{2}}\quad \text{as } x \to \infty,$$ and then also $$E_x^{\operatorname{Be}}(\sigma_x \mid Y_{\sigma_x} = x-1) \to 1, \quad E_x^{\operatorname{Be}}(\sigma_x \mid Y_{\sigma_x} = x+1) \to 1
\quad \text{as } x \to \infty.$$ Therefore, uniformly in those $A$ with all $x_i \geq h_{1-}$, as $m \to \infty$ we have $$\label{near1}
E_{h_2}^{\operatorname{Be}}(T_i - T_{i-1} \mid A) \to 1.$$ It is easily seen by comparison to “Brownian motion plus small constant” that $P_z^{\operatorname{Be}}(\sigma_x > 1)$ is bounded away from 1 uniformly in (large) $x$ and in $z \in [x-1,x+1]$. Hence by the Markov property $P_x^{\operatorname{Be}}\big(\sigma_x > t)$ decays exponentially in $t$, uniformly in large $x$. By , this means there exist $K_{13},K_{14}$ such that $$P_x^{\operatorname{Be}}\big(\sigma_x > t \mid Y_{\sigma_x} = x \pm 1 \big) \leq \max\left( \frac{1}{f(x)},\frac{1}{1-f(x)} \right) P_x^{\operatorname{Be}}\big(\sigma_x > t)
\leq K_{13} e^{-K_{14} t},$$ for all $t\geq 0$ and all (large) $x$. Therefore for $m$ sufficiently large, for all $A$ and $t$, $$\label{expbound}
P_{h_2}^{\operatorname{Be}}(T_i - T_{i-1} > t \mid A) \leq K_{13} e^{-K_{14} t}.$$ By standard methods, it follows from and that for some $K_{15}(\rho),K_{16}(\rho)$ not depending on $A$, $$\label{taugap}
P_{h_2}^{\operatorname{Be}}\left( \left| \tau_{h_{1-}}^{\operatorname{Be}} - \tau_{h_{1-}}^{\operatorname{BI}} \right| > \rho \tau_{h_{1-}}^{\operatorname{BI}}\ \big|\ A \right) =
P_{h_2}^{\operatorname{Be}}\left( \left| T_k - k \right| > \rho k\ \big|\ A \right) \leq K_{15} e^{-K_{16}k}.$$ Therefore the same bound holds unconditionally, so $$\begin{aligned}
\label{BesselvsBI3}
P_{h_2}^{\operatorname{BI}}&\big(\tau_{h_{1-}}^{\operatorname{BI}} \geq (1-2\rho)m\big) \\
&\leq P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1-}}^{\operatorname{Be}} \geq (1-3\rho)m\big)
+ P_{h_2}^{\operatorname{Be}}\left( \tau_{h_{1-}}^{\operatorname{BI}} \geq (1-2\rho)m, \left| \tau_{h_{1-}}^{\operatorname{Be}} - \tau_{h_{1-}}^{\operatorname{BI}} \right| > \rho \tau_{h_{1-}}^{\operatorname{BI}} \right)
\notag \\
&\leq P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1-}}^{\operatorname{Be}} \geq (1-3\rho)m\big) + K_{15} e^{-(1-2\rho)K_{16}m} \notag \\
&\leq (1+\theta) P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1-}}^{\operatorname{Be}} \geq (1-3\rho)m\big), \notag
\end{aligned}$$ where the last inequality follows from and , for large $m$. Thus is proved. We have similarly from (with $\tau_{h_{1-}}$ trivially replaced by $\tau_{h_{1+}}$) that $$\begin{aligned}
\label{BesselvsBI4}
P_{h_2}^{\operatorname{Be}}&\big(\tau_{h_{1+}}^{\operatorname{Be}} \geq (1+\rho)m\big) \\
&\leq P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1+}}^{\operatorname{BI}} \geq m\big)
+ P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1+}}^{\operatorname{Be}} \geq (1+\rho)m, \left| \tau_{h_{1+}}^{\operatorname{Be}} - \tau_{h_{1+}}^{\operatorname{BI}} \right| > \rho \tau_{h_{1+}}^{\operatorname{BI}} \big)
\notag \\
&\leq P_{h_2}^{\operatorname{BI}}\big(\tau_{h_{1+}}^{\operatorname{BI}} \geq m\big) + K_{15}e^{-K_{16}m} \notag \\
&\leq \frac{1}{1-\theta} P_{h_2}^{\operatorname{BI}}\big(\tau_{h_{1+}}^{\operatorname{BI}} \geq m\big), \notag
\end{aligned}$$ so , and thus Claim 2, are also proved.
[*Proof of Claim 1.*]{} From , and then, we have $$\begin{aligned}
P_{h_2}^{\operatorname{BI}}(\tau_{h_{1+}} \geq m) &\geq (1-\theta) P_{h_2}^{\operatorname{Be}}\big(\tau_{h_{1+}} \geq (1+\rho)m\big) \\
&\geq (1-2\theta) P_{h_2}^{\operatorname{Be}}\big(\tau_0 \geq (1+2\rho)m\big) \\
&\geq K_{17} {\epsilon}_2^{2\kappa},
\end{aligned}$$ and it is straightforward to replace $\tau_{h_{1+}}$ here by $\tau_{h_1}$, proving the second inequality in . The first inequality there is trivial.
The first inequality in is also trivial, so we prove the second one. Using and slight variants of and we get that for large $m$, $$\begin{aligned}
P_{h_2}(\tau_{h_1} \geq m) &\geq P_{h_2}^*\big( \tau_{h_{1+}}^{\operatorname{BI}} \geq m, N(m) \leq \eta \sqrt{m} \big) \\
&= P_{h_2}^*\big( \tau_{h_{1+}}^{\operatorname{BI}} \geq m \big) - P_{h_2}^*\big( \tau_{h_{1+}} \geq m, N(m) > \eta \sqrt{m} \big) \\
&\geq K_{10} {\epsilon}_2^{2\kappa} - e^{-K_{2} \sqrt{m}} \\
&\geq {\frac{1}{2}}K_{10} {\epsilon}_2^{2\kappa},
\end{aligned}$$ completing the proof of Claim 1.
This also completes the proof of , as noted after Claim 2.\
Proof of and
=============
For even numbers $0<m<n$, let $$f_m = P_0(\tau_0 = m), \qquad A_{m,n} = \frac{2}{n-m+2} \sum_{j=m}^n f_j.$$ $A_{m,n}$ is the average of the even-index $f_j$’s with $j \in [m,n]$. We use and the following convexity property of $\{f_m\}$.
\[convexity\] For all even numbers $0<k<m$, $$\label{convex}
f_m \leq\frac{f_{m+k} + f_{m-k}}{2}$$ and $$\label{average}
f_m \leq A_{m-k,m+k}.$$
Let ${\mathbf{x}}= \{x_0,\dots,x_m\}$ be the trajectory of an excursion of length $m$ starting at time 0, and ${\mathbf{x}}' = \{x_k',\dots,x_{m+k}'\}$ the trajectory of an excursion of length $m$ starting at time $k$. (Necessarily, then, $x_0 = x_m = x_k' = x_{m+k}' = 0$ and all other $x_j$ and $x_j'$ are positive.) Since $k$ is even, there must be an $s \in (k,m)$ with $x_s = x_s'$; let $T = T({\mathbf{x}},{\mathbf{x}}')$ denote the least such $s$ and $D_t = \{({\mathbf{x}},{\mathbf{x}}'): T({\mathbf{x}},{\mathbf{x}}') = t\}$. For ${\mathbf{x}},{\mathbf{x}}' \in D_t$, by switching the two trajectories after time $t$, we obtain an excursion ${\mathbf{y}}= {\mathbf{y}}({\mathbf{x}},{\mathbf{x}}') = \{x_0,\dots,x_t,x_{t+1}',\dots,x_{m+k}'\}$ of length $m+k$ and an excursion ${\mathbf{y}}' = {\mathbf{y}}'({\mathbf{x}},{\mathbf{x}}') = \{x_k',\dots,x_t',x_{t+1},\dots,x_m\}$ of length $m-k$. The map $({\mathbf{x}},{\mathbf{x}}') \mapsto ({\mathbf{y}},{\mathbf{y}}')$ is one to one and satisfies $$P_0({\mathbf{x}}) P({\mathbf{x}}' \mid X_k = 0) = P_0({\mathbf{y}}) P({\mathbf{y}}' \mid X_k = 0).$$ It follows that $$\begin{aligned}
f_m^2 &= \sum_{t:k<t<m}\ \sum_{({\mathbf{x}},{\mathbf{x}}') \in D_t} P_0({\mathbf{x}}) P({\mathbf{x}}' \mid X_k = 0) \\
&= \sum_{t:k<t<m}\ \sum_{({\mathbf{x}},{\mathbf{x}}') \in D_t} P_0({\mathbf{y}}({\mathbf{x}},{\mathbf{x}}')) P({\mathbf{y}}'({\mathbf{x}},{\mathbf{x}}') \mid X_k = 0) \\
&\leq \left( \sum_{{\mathbf{y}}} P_0({\mathbf{y}}) \right) \left( \sum_{{\mathbf{y}}'} P({\mathbf{y}}' \mid X_k=0) \right) \\
&= f_{m+k} f_{m-k} \\
&\leq \left( \frac{f_{m+k} + f_{m-k}}{2} \right)^2.
\end{aligned}$$ Equation is an immediate consequence of .
Let $\theta>0$. Provided $\eta$ is sufficiently small (depending on $\theta$), we have from , and Lemma \[convexity\] that for $n$ large and even and $k = 2\lfloor \eta n/2 \rfloor$, $$\begin{aligned}
\label{pointupper}
P_0(\tau_0 = n) &= f_n \\
&\leq A_{n-k,n+k} \notag \\
&= \frac{1}{k+1} P_0\left( (1-\eta)n \leq \tau_0 \leq (1+\eta)n \right) \notag \\
&\leq (1+\theta) \frac{2^{2-\kappa}\kappa}{K_0 \Gamma(\kappa)} n^{-(\kappa+1)}L(\sqrt{n}). \notag
\end{aligned}$$ In the reverse direction, suppose $f_n < (1-\theta) A_{n-k,n-2}$ for some $0<k<n/2$, with $k,n$ even. By Lemma \[convexity\] we have $$\begin{aligned}
\sum_{j=1}^{k/2} f_{n-2j} &\leq \frac{k}{4} f_n + {\frac{1}{2}}\sum_{j=1}^{k/2} f_{n-4j} \\
&\leq \frac{1-\theta}{2} \sum_{j=1}^{k/2} f_{n-2j} + \frac{1}{4} \sum_{j=1}^{k/2} f_{n-4j}
+ \frac{1}{4} \sum_{j=1}^{k/2} \frac{f_{n-4j+2} + f_{n-4j-2}}{2} \notag \\
&= \frac{1-\theta}{2} \sum_{j=1}^{k/2} f_{n-2j} + \frac{1}{4} \sum_{j=1}^{k/2} f_{n-4j}
+ \frac{1}{8} \sum_{j=1}^{k/2} f_{n-4j+2} + \frac{1}{8} \sum_{j=2}^{(k+2)/2} f_{n-4j+2} \notag \\
&= \frac{1-\theta}{2} \sum_{j=1}^{k/2} f_{n-2j} + \frac{1}{4} \sum_{m=2}^k f_{n-2m} + \frac{1}{8} f_{n-2}
+ \frac{1}{8} f_{n-2k-2}, \notag
\end{aligned}$$ and therefore $$\begin{aligned}
\frac{1+\theta}{2} \sum_{j=1}^{k/2} f_{n-2j} \leq \frac{1}{4} \sum_{j=1}^k f_{n-2j} + \frac{1}{8} f_{n-2k-2},
\end{aligned}$$ which in the case $k = 2\lfloor \eta n/2 \rfloor$ gives $$\begin{aligned}
\label{contra}
(1+\theta)&P_0\left( (1-\eta)n \leq \tau_0 \leq n-2 \right) \\
&\leq {\frac{1}{2}}P_0\left( (1 - 2\eta)n \leq \tau_0 \leq n-2 \right)
+ \frac{1}{4} P_0\left(\tau_0 = n - 4\left\lfloor \frac{\eta n}{2} \right\rfloor - 2 \right). \notag\end{aligned}$$ For small $\eta$ and large $n$, this contradicts , showing that we cannot have $f_n < (1-\theta) A_{n-k,n-2}$. Therefore for large $n$, using we have $$\begin{aligned}
P_0(\tau_0 = n) &= f_n \\
&\geq (1-\theta)A_{n-k,n-2} \notag \\
&= \frac{2(1-\theta)}{k} P_0\left( (1 - \eta)n \leq \tau_0 \leq n-2 \right) \notag \\
&\geq (1-2\theta) \frac{2^{2-\kappa}\kappa}{K_0 \Gamma(\kappa)} n^{-(\kappa+1)}L(\sqrt{n}). \notag\end{aligned}$$ This and prove .
We now prove . Let ${\tilde{P}}$ denote the distribution of the Bessel-like RW dual to $P$, that is, the walk with transition probabilities $\tilde{p}_x = q_x, \tilde{q}_x = p_x$ for $x \geq 1$. In [@DFPS97] it is proved that for $n$ even, $$\label{duality}
P(\tau_0 > n) = {\tilde{P}}(X_n = 0).$$ For $\delta=-1$, the dual walk has drift parameter $\tilde{\delta}=1$, so follows by applying and to the dual walk.
Proof of Theorem \[hittime\]
============================
We want to use so we need to approximate $$f_n^{(k)} = P_k(\tau_0 = n) \quad \text{ and} \quad P_k(X_n=0).$$ We sometimes omit the superscript $(k)$ when it is equal to 0. We start with the following relative of Lemma \[convexity\].
\[latticepath\] Let $0 \leq p \leq q \leq r \leq s \leq \infty$ and $0 \leq k < l$. Then for $l-k$ even, $$\label{intervals}
P_l(\tau_0 \in [p,q]) P_k(\tau_0 \in [r,s]) \leq P_l(\tau_0 \in [r,s]) P_k(\tau_0 \in [p,q]),$$ and for $l-k$ odd, $$\label{intervals2}
P_l(\tau_0 \in [p,q]) P_k(\tau_0 \in [r,s]) \leq P_l(\tau_0 \in [r+1,s+1]) P_k(\tau_0 \in [p-1,q-1]).$$
Suppose first that $l-k$ is even. Consider a lattice path ${\mathbf{x}}$ starting at $(0,k)$ in space-time which first hits the horizontal axis at a time in $[r,s]$, and a lattice path ${\mathbf{x}}'$ starting at $(0,l)$ which first hits the axis at a time in $[p,q]$. Since $l-k$ is even, there must be a $t \in (0,q]$ with $x_t = x_t'$. Switching the two trajectories after the first such $t$ and proceeding as in Lemma \[convexity\] we obtain .
For $l-k$ odd we repeat this argument but with the path ${\mathbf{x}}$ shifted one unit to the right, that is, started from $(1,k)$.
Here are some special cases of interest for Lemma \[latticepath\], particularly when comparing point versus interval probabilities for $\tau_0$.
\[latticepath2\] (i) For all $0 \leq k<l \leq n$ and $j>0$ with $n-l$ and $n+j-k$ even, $$\label{lattice1}
\frac{f_{n+j}^{(k)}}{f_n^{(k)}} \leq \frac{f_{n+j}^{(l)}}{f_n^{(l)}} \quad \text{if $l-k$ is even},$$ and $$\label{lattice2}
\frac{f_{n+j}^{(k)}}{f_{n-1}^{(k)}} \leq \frac{f_{n+j+1}^{(l)}}{f_n^{(l)}} \quad \text{if $l-k$ is odd}.$$ (ii) For all $0 \leq l \leq m$, $$\label{lattice3}
P_l(\tau_0 = m) \leq M_l P_0(\tau_0 = m) \quad \text{if $l$ is even},$$ and $$\label{lattice4}
P_l(\tau_0 = m) \leq M_l P_0(\tau_0 = m-1) \quad \text{if $l$ is odd}.$$ (iii) For all $l>0$ and $0 \leq p < q < m$, $$\label{lattice5}
P_l(\tau_0 = m) \geq M_l P_0(\tau_0 = m) \frac{ P_0(\tau_0 - \tau_l \in [p,q]) }{ P_0(\tau_0 \in [p,q]) } \quad \text{if $l$ is even},$$ and $$\label{lattice5a}
P_l(\tau_0 = m) \geq M_l P_0(\tau_0 = m-1) \frac{ P_0(\tau_0 - \tau_l \in [p,q]) }{ P_0(\tau_0 \in [p-1,q-1]) } \quad \text{if $l$ is odd}.$$
By Corollary \[latticepath2\], to show that $P_l(\tau_0 = m)$ can be well approximated by $M_l P_0(\tau_0 = m)$ (or $M_l P_0(\tau_0 = m-1)$, depending on parity), it is sufficient to find, given $m$, values $p<q \leq m$ for which the fraction in or is almost 1. We will see that this can be done for $m \gg l^2$.
\(i) Take $p=q=n$ and $r=s=n+j$ in Lemma \[latticepath\] to get $f_n^{(l)} f_{n+j}^{(k)} \leq f_{n+j}^{(l)} f_n^{(k)}$ in the case of even $l-k$, and similarly for odd $l-k$.
\(ii) Consider even $l$. We may assume $m$ is also even, for otherwise the left side of is 0. Applying Lemma \[latticepath\] with $k=0$, $p=q=r=m$ and $s=\infty$ we get $$\label{lattice6}
P_l(\tau_0 = m) \leq \frac{ P_l(\tau_0 \geq m) }{ P_0(\tau_0 \geq m) } P_0(\tau_0 = m),$$ while by , $$\begin{aligned}
\frac{1}{M_l} P_l(\tau_0 \geq m) \leq P_0(\tau_0 \geq m).
\end{aligned}$$ Together these prove . For odd $l$ we may assume $m$ is odd, and in place of we get $$\label{lattice7}
P_l(\tau_0 = m) \leq \frac{ P_l(\tau_0 \geq m+1) }{ P_0(\tau_0 \geq m) } P_0(\tau_0 = m-1),$$ and the rest of the proof is essentially unchanged, since $P_l(\tau_0 \geq m+1) \leq P_l(\tau_0 \geq m)$.
\(iii) Consider even $l$. We may assume $m$ is even, for otherwise the right side of is 0. Applying Lemma \[latticepath\] with $k=0$, $[r,s] = \{m\}$ we obtain $$\label{lattice8}
P_l(\tau_0 = m) \geq \frac{ P_l(\tau_0 \in [p,q]) }{ P_0(\tau_0 \in [p,q]) } P_0(\tau_0 = m),$$ while by , $$\label{1overMl}
\frac{1}{M_l} P_l(\tau_0 \in [p,q]) = P_0(\tau_l < \tau_0) P_l(\tau_0 \in [p,q]) = P_0(\tau_0 - \tau_l \in [p,q]),$$ and together these prove . For odd $l$ we may again assume $m$ is odd and take $[r,s] = \{m-1\}$, so that in place of , using we get $$\begin{aligned}
\label{lattice9}
P_l(\tau_0 = m) &\geq \frac{ P_l(\tau_0 \in [p,q]) }{ P_0(\tau_0 \in [p-1,q-1]) } P_0(\tau_0 = m-1) \\
&= M_l \frac{ P_0(\tau_0 - \tau_l \in [p,q]) }{ P_0(\tau_0 \in [p-1,q-1]) } P_0(\tau_0 = m-1). \notag\end{aligned}$$
Note that if we take $k=0$ and $j \ll n$ in , we see from that the left side of is close to 1, so the right side cannot be much less than 1 for any $l>0$.
For the Bessel process we have by that for $0<a<b$, recalling $\kappa = (1+\delta)/2$, $$\begin{aligned}
\label{Besselcase}
P_k^{\operatorname{Be}}(\tau_0 \in [a,b]) &= \int_{k^2/2b}^{k^2/2a} \frac{1}{\Gamma(\kappa)}u^{\kappa-1} e^{-u}\ du.
\end{aligned}$$ As a step toward approximating $P_k(X_n=0)$ we have the following “interval” version of Theorem \[hittime\], for midrange starting heights ($k$ of order $\sqrt{m}$); for these we apparently cannot get sharp results from Corollary \[latticepath2\](ii) and (iii).
\[intervalmid\] Let $\theta>0,\chi>0$, $0 < {\Delta_{\min}}< {\Delta_{\max}}$ and $0<a<b$. Provided $\chi$ is sufficiently small (depending on $\theta$), ${\Delta_{\max}}$ is sufficiently small (depending on $\theta,\chi$), $$\label{ab}
\frac{b-a}{b} \in [\Delta_{\min},\Delta_{\max}],$$ the starting height $k$ is midrange, that is, $$\label{mk2}
\sqrt{a\chi} \leq k \leq \sqrt{a/\chi},$$ and $a$ is sufficiently large (depending on $\theta,\chi,{\Delta_{\min}},{\Delta_{\max}}$), we have $$\begin{aligned}
\label{sandwich7}
(1-\theta) &P_k^{\operatorname{Be}}(\tau_0 \in [a,b]) \leq P_k(\tau_0 \in [a,b])
\leq (1+\theta) P_k^{\operatorname{Be}}(\tau_0 \in [a,b])
\end{aligned}$$ and $$\begin{aligned}
\label{sandwich6}
\frac{1-\theta}{\Gamma(\kappa)} \frac{b-a}{b} \left( \frac{k^2}{2a} \right)^\kappa e^{-k^2/2a}
&\leq P_k(\tau_0 \in [a,b]) \\
&\leq \frac{1+\theta}{\Gamma(\kappa)} \frac{b-a}{b} \left( \frac{k^2}{2a} \right)^\kappa e^{-k^2/2a}. \notag
\end{aligned}$$
Let $0<\rho<\Delta_{\min}/8$ and $\zeta>2\beta>0$. We always select our constants in the following manner: $\theta$ is given; we choose $\chi$ then $\Delta_{\max}$, and then $\Delta_{\min} < \Delta_{\max}$ is arbitrary, then we choose $\rho$ and then $\zeta$ and $\beta$ (which appear in below.) Finally we choose $[a,b]$ as specified. Each choice may depend only on the preceding choices, and when we say a parameter is “sufficiently large” (or small), the required size may depend on the previous choices.
The general outline is similar to the proof of . Analogously to , we will establish the following sequence of ten inequalities: $$\begin{aligned}
\label{toprove2}
(1-&6\theta)P_k^{\operatorname{Be}}(\tau_0^{\operatorname{Be}} \in [a,b]) \\
&\leq (1-5\theta)P_k^{\operatorname{Be}}\big(\tau_0^{\operatorname{Be}} \in [(1+2\rho)a,(1-3\rho)b] \big) \notag \\
&\leq (1-4\theta)P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1+\rho)a,(1-3\rho)b] \big) \notag \\
&\leq (1-3\theta) P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big) \notag\\
&\leq (1-\theta) P_k\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \big) \notag \\
&\leq P_k(\tau_0 \in [a,b]) \notag \\
&\leq (1+\theta) P_k\big( \tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big) \notag \\
&\leq (1+4\theta)P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b] \big) \notag \\
&\leq (1+5\theta)P_k^{\operatorname{Be}}(\tau_{\zeta k}^{\operatorname{Be}} \in [(1-2\rho)a,(1+2\rho)b]) \notag \\
&\leq (1+7\theta)P_k^{\operatorname{Be}}(\tau_0^{\operatorname{Be}} \in [(1-2\rho)a,(1+3\rho)b]) \notag \\
&\leq (1+8\theta)P_k^{\operatorname{Be}}(\tau_0^{\operatorname{Be}} \in [a,b]). \notag
\end{aligned}$$ As with , this should be viewed as five “sandwich” bounds on $P_k(\tau_0 \in [a,b])$, with the outermost sandwich yielding the desired result.
Provided $\Delta_{\max}/\theta\chi$ is sufficiently small and the second inequality in holds, the gamma density $$f_\kappa(u) = \frac{1}{\Gamma(\kappa)}u^{\kappa-1} e^{-u}, \quad u \geq 0,$$ satisfies $$(1-\theta) f_\kappa\left( \frac{k^2}{2a} \right) \leq f_\kappa(u) \leq (1+\theta) f_\kappa\left( \frac{k^2}{2a} \right) \quad \text{for all } u \in
\left[ \frac{k^2}{2b},\frac{k^2}{2a} \right].$$ Then by , $$\label{Bessellower}
P_k^{\operatorname{Be}}(\tau_0 \in [a,b]) \geq \frac{1-\theta}{\Gamma(\kappa)} \frac{b-a}{b} \left( \frac{k^2}{2a} \right)^\kappa e^{-k^2/2a}$$ and, using also the second inequality in , $$\label{Besselupper}
P_k^{\operatorname{Be}}(\tau_0 \in [a,b]) \leq \frac{1+\theta}{\Gamma(\kappa)} \frac{b-a}{b} \left( \frac{k^2}{2a} \right)^\kappa e^{-k^2/2a} .$$ Therefore follows from . The inequalities and , with minor modifications made to $\theta,a$ and $b$, also prove the first and last inequalities in , provided $\rho$ is suficiently small (depending on $\theta,{\Delta_{\min}},\chi$.)
Turning to the 2nd and 9th inequalities in , provided $\zeta^2/\rho \chi$ is sufficiently small (depending on $\theta$), using we have $$\begin{aligned}
\label{zetakvs0}
P_k^{\operatorname{Be}}(\tau_0 \in [(1-2\rho)a,(1+3\rho)b]) &\geq P_k^{\operatorname{Be}}\big(\tau_{\zeta k} \in [(1-2\rho)a,(1+2\rho)b] \big)
P_{\zeta k}^{\operatorname{Be}}\big( \tau_0 \leq \rho b \big) \\
&= P_k^{\operatorname{Be}}\big(\tau_{\zeta k} \in [(1-2\rho)a,(1+2\rho)b] \big) \int_{\zeta^2k^2/2\rho b}^\infty
\frac{1}{\Gamma(\kappa)}u^{\kappa-1} e^{-u}\ du \notag\\
&\geq (1-\theta) P_k^{\operatorname{Be}}\big(\tau_{\zeta k} \in [(1-2\rho)a,(1+2\rho)b] \big) . \notag
\end{aligned}$$ This proves the 9th inequality in . In the other direction, $$\begin{aligned}
\label{zetakvs0b}
P_k^{\operatorname{Be}}(\tau_0 \in [(1+2\rho)a,(1-3\rho)b]) &\leq P_k^{\operatorname{Be}}\big(\tau_{\zeta k} \in [(1+\rho)a,(1-3\rho)b] \big)
+ P_{\zeta k}^{\operatorname{Be}}\big( \tau_0 > \rho a \big).
\end{aligned}$$ From , and , $$\label{Bessellower2}
P_k^{\operatorname{Be}}(\tau_0 \in [(1+2\rho)a,(1-3\rho)b])
\geq \frac{(1-\theta){\Delta_{\min}}}{\Gamma(\kappa)} \left( \frac{k^2}{2a} \right)^\kappa e^{-1/2\chi},$$ and hence by , provided $\zeta^2/\rho$ is sufficiently small (depending on $\theta,{\Delta_{\min}},\chi$), $$\begin{aligned}
\label{Besselupper2}
P_{\zeta k}^{\operatorname{Be}}\big( \tau_0 > \rho a \big) &= \int_0^{\zeta^2 k^2/2\rho a} \frac{1}{\Gamma(\kappa)}u^{\kappa-1} e^{-u}\ du \\
&\leq \frac{1}{\kappa \Gamma(\kappa)} \left( \frac{\zeta^2 k^2}{2\rho a} \right)^\kappa \notag \\
&\leq \theta P_k^{\operatorname{Be}}(\tau_0 \in [(1+2\rho)a,(1-3\rho)b]). \notag
\end{aligned}$$ With this shows that $$\label{zetakvs0c}
P_k^{\operatorname{Be}}(\tau_0 \in [(1+2\rho)a,(1-3\rho)b]) \leq \frac{1}{1-\theta} P_k^{\operatorname{Be}}\big(\tau_{\zeta k} \in [(1+\rho)a,(1-3\rho)b] \big),$$ which proves the 2nd inequality in .
Next we consider the 3rd and 8th inequalities in , in which Bessel-process probabilities are compared to similar probabilities for the imbedded RW. First, for the 8th inequality, analogously to we have for some $K_i = K_i(\rho)$ that $$\begin{aligned}
\label{BIvsBe1}
P_k^{\operatorname{BI}}&\big(\tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b] \big) \\
&\leq P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1-2\rho)a,(1+2\rho)b] \big) \notag \\
&\qquad + P_k^{\operatorname{Be}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b],
|\tau_{\zeta k}^{\operatorname{BI}} - \tau_{\zeta k}^{\operatorname{Be}}| > \rho a \big) \notag \\
&\leq P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1-2\rho)a,(1+2\rho)b] \big) + K_{18}e^{-K_{19}a}. \notag
\end{aligned}$$ By , and , there exist $K_{20} = K_{20}(\rho,{\Delta_{\min}},\chi)$ and $K_{21} = K_{21}(\rho,{\Delta_{\min}},\chi,\theta)$ such that for $a \geq K_{21}$, $$\label{expsmall}
P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1-2\rho)a,(1+2\rho)b] \big) \geq K_{20} a^{-\kappa} \geq \frac{1}{\theta}e^{-K_{19}a},$$ which with and shows that $$\begin{aligned}
\label{BIvsBe2}
P_k^{\operatorname{BI}}\big(\tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b] \big) &\leq (1+\theta)P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1-2\rho)a,(1+2\rho)b] \big),
\end{aligned}$$ so the 8th inequality in is proved. For the 3rd inequality, similarly to and we get $$\label{BIvsBe3}
P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1+\rho)a,(1-3\rho)b] \big) \leq P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big)
+ e^{-K_{19}a},$$ which together with a slight modification of gives $$\label{BIvsBe4}
(1-\theta)P_k^{\operatorname{Be}}\big(\tau_{\zeta k}^{\operatorname{Be}} \in [(1+\rho)a,(1-3\rho)b] \big) \leq P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big),$$ yielding the desired result.
Now we consider the 4th through 7th inequalities in , comparing probabilities for the imbedded RW to similar probabilities for the original RW, and comparing the hitting times of $(\zeta-2\beta)k$ and 0; for this we use the coupling of $\{X_n\}$ and $\{X_n^{\operatorname{BI}}\}$. First, for the 4th inequality, observe that for walks starting at $k$, if $\tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b]$ and the number of missteps by time $\tau_{\zeta k}^{\operatorname{BI}}$ is less than $\beta k$, then at time $\tau_{\zeta k}^{\operatorname{BI}}$, the stopping time $\tau_{(\zeta-2\beta)k}$ for the RW $\{X_n\}$ has not yet occurred and this RW is located in $((\zeta-2\beta)k,(\zeta+2\beta)k)$. Therefore $$\begin{aligned}
\label{BIvsRW1}
P_k^{\operatorname{BI}}&\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big) \\
&\leq P_k^*\big( \tau_{(\zeta-2\beta)k} < \tau_{\zeta k}^{\operatorname{BI}} \wedge (1-2\rho)b, N(\tau_{(\zeta-2\beta)k}) \geq \beta k \big) \notag \\
&\quad + P_k^*\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b], \tau_{(\zeta-2\beta)k} > \tau_{\zeta k}^{\operatorname{BI}},
X_{\tau_{\zeta k}^{\operatorname{BI}}} \in ((\zeta-2\beta)k,(\zeta+2\beta)k) \big). \notag\end{aligned}$$ Let $$D = \{ \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b], \tau_{(\zeta-2\beta)k} > \tau_{\zeta k}^{\operatorname{BI}},
X_{\tau_{\zeta k}^{\operatorname{BI}}} \in ((\zeta-2\beta)k,(\zeta+2\beta)k) \}$$ denote the last event in . When $D$ occurs, the RW $\{X_n^{\operatorname{BI}}\}$ reaches height $\zeta k$ at some time $l$, and when it does, the RW $\{X_n\}$ is at some height $j$ close to $\zeta k$, so $\{X_n\}$ has a high probability to reach height $(\zeta - 2\beta)k$ within an additional time $\rho b$. More precisely, for $j \in ((\zeta-2\beta)k,(\zeta+2\beta)k)$ and $l \in [a,(1-2\rho)b]$, provided $\zeta^2/\rho\chi$ is sufficiently small (depending on $\theta$), using , , and our assumption $a \geq \chi k^2$ we have $$\begin{aligned}
\label{finish}
P_k^*&\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \mid D \cap \{\tau_{\zeta k}^{\operatorname{BI}} = l,X_l= j\} \big) \\
&= P_j\big( \tau_{(\zeta-2\beta)k} \leq (1-\rho)b - l \big) \notag \\
&\geq P_j(\tau_{(\zeta-2\beta)k} \leq \rho b \big) \notag \\
&\geq 1 - M_j P_0(\tau_0 \geq \rho b) \notag \\
&\geq 1-\theta. \notag
\end{aligned}$$ Since $l,j$ are arbitrary, the same bound holds if we just condition on $D$. From this and we get $$\begin{aligned}
\label{BIvsRW2}
P_k^{\operatorname{BI}}&\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big) \\
&\leq P_k^*\big( \tau_{(\zeta-2\beta)k} < \tau_{\zeta k}^{\operatorname{BI}} \wedge (1-2\rho)b, N(\tau_{(\zeta-2\beta)k}) \geq \beta k \big) \notag \\
&\quad + \frac{1}{1-\theta} P_k\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \big). \notag
\end{aligned}$$ Reasoning similarly to using , and then using and , we get that for some $K_{22}(\zeta,\beta)$ and $K_{23}(\zeta,\beta,\theta,{\Delta_{\min}},\chi)$, for $a \geq K_{23}$, $$\begin{aligned}
\label{missteps}
P_k^*&\big( \tau_{(\zeta-2\beta)k} < \tau_{\zeta k}^{\operatorname{BI}} \wedge (1-2\rho)b, N(\tau_{(\zeta-2\beta)k}) \geq \beta k \big) \\
&\leq e^{-K_{22}\sqrt{a}} \notag \\
&\leq \theta P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big). \notag
\end{aligned}$$ With this shows that $$\begin{aligned}
\label{BevsRW1}
&(1-\theta)^2 P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [a,(1-2\rho)b] \big)
\leq P_k\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \big),
\end{aligned}$$ which yields the 4th inequality in .
For the 5th inequality in , from , , and , provided $\zeta^2/\rho\chi$ is sufficiently small (depending on $\theta$), we have $$P_{(\zeta-2\beta)k}\big( \tau_0 > \rho b \big) \leq M_{(\zeta-2\beta)k}P_0(\tau_0 \geq \rho b) \leq \theta.$$ Hence $$\begin{aligned}
\label{0vshigher}
(1-\theta) &P_k\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \big) \\
&\leq P_k\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \big) P_{(\zeta-2\beta)k}\big( \tau_0 \leq \rho b \big) \notag \\
&\leq P_k\big( \tau_0 \in [a,b] \big), \notag\end{aligned}$$ which proves the 5th inequality.
Next, to prove the 7th inequality in , we can repeat — with $\{X_n\}$ and $\{X_n^{\operatorname{BI}}\}$ interchanged, and with $\zeta k,(\zeta-2\beta)k$ replaced by $(\zeta+2\beta)k,\zeta k$, respectively, to obtain first the following analog of and : $$\begin{aligned}
\label{BIvsRW3}
P_k&\big( \tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big) \\
&\leq P_k^*\big( \tau_{\zeta k}^{\operatorname{BI}} < \tau_{(\zeta+2\beta)k} \wedge b, N(\tau_{\zeta k}^{\operatorname{BI}}) \geq \beta k \big)
+ \frac{1}{1-\theta} P_k\big( (\tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b] \big) \notag \\
&\leq \theta P_k\big( \tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big) + \frac{1}{1-\theta} P_k\big( (\tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b] \big), \notag\end{aligned}$$ and from this the analog of : $$\begin{aligned}
\label{BevsRW3}
P_k\big( \tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big)
&\leq (1+3\theta)P_k^{\operatorname{BI}}\big( \tau_{\zeta k}^{\operatorname{BI}} \in [(1-\rho)a,(1+\rho)b] \big),
\end{aligned}$$ so the 7th inequality is proved. Here for the second inequality in , analogously to , we require a lower bound for $P_k\big( \tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big)$, and this follows from and the inequality $$P_k\big( \tau_{(\zeta-2\beta)k} \in [a,(1-\rho)b] \big) \geq \frac{1-6\theta}{1-\theta} P_k^{\operatorname{Be}}(\tau_0 \in [a,b])$$ which is contained in the first four inequalities of , with trivial modification to replace $\zeta-2\beta$ with $\zeta+2\beta$ and $[a,(1-\rho)b]$ with $[(1-\rho)a,b]$.
For the 6th inequality, we have $$\begin{aligned}
\label{Bevshigher2}
P_k\big( \tau_0 \in [a,b] \big) \leq P_k\big( &\tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big) \\
&+ P_k\big( \tau_{(\zeta+2\beta)k} < (1-\rho)a, \tau_0 \in [a,b] \big). \notag\end{aligned}$$ Let us show that the last probability in is much smaller than the first one. The Markov property at $\tau_{(\zeta+2\beta)k}$, together with , , and , yields that for some $K_{24}$, provided $a$ is sufficiently large, $$\begin{aligned}
\label{smallbit}
P_k\big( \tau_{(\zeta+2\beta)k} < (1-\rho)a, \tau_0 \in [a,b] \big) &\leq P_{(\zeta+2\beta)k}\big( \tau_0 \geq \rho a \big) \\
&\leq M_{(\zeta+2\beta)k}P_0(\tau_0 \geq \rho a) \notag \\
&\leq K_{24}\left( \frac{ \zeta^2k^2 }{ \rho a } \right)^\kappa. \notag
\end{aligned}$$ From , and the first half of we have that for some $K_{25}$, $$P_k\big( \tau_0 \in [a,b] \big) \geq (1-6\theta)P_k^{\operatorname{Be}}(\tau_0 \in [a,b]) \geq K_{25} {\Delta_{\min}}e^{-1/2\chi} \left( \frac{k^2}{a} \right)^\kappa.$$ From this and we obtain that provided $\zeta^2/\rho$ is sufficiently small (depending on ${\Delta_{\min}},\theta,\chi$), the ratio of the last to the first probability in is at most $\theta$, which with shows that $$(1-\theta)P_k\big( \tau_0 \in [a,b] \big) \leq P_k\big( \tau_{(\zeta+2\beta)k} \in [(1-\rho)a,b] \big),$$ proving the 6th inequality in , which completes the full proof of . Statement is then immediate, and then, as we have noted, follows.
Let us now prove Theorem \[hittime\] for low starting heights—suppose that $$1 \leq k < \sqrt{\chi m}.$$ Let $a \in [m/2,m)$. We will use Corollary \[latticepath2\](ii) and (iii), with $[p,q] = [a/2,a]$, together with . By , provided $\rho$ is sufficiently small (depending on $\theta$) and then $a$ is sufficiently large, we have $$\begin{aligned}
\label{latticelower}
P_0&\left( \tau_0 - \tau_k \in \left[ \frac{a}{2},a \right] \right)) \\
&\geq P_0\left( \tau_k \leq \rho a, \tau_0 \in \left[ \left( {\frac{1}{2}}+ \rho \right)a,a \right] \right) \notag \\
&= P_0\left( \tau_0 \in \left[ \left( {\frac{1}{2}}+ \rho \right)a,a \right] \right)
\left[ 1 - P_0\left( \tau_k > \rho a\ \bigg|\ \tau_0 \in \left[ \left( {\frac{1}{2}}+ \rho \right)a,a \right] \right) \right] \notag \\
&\geq (1-\theta) P_0\left( \tau_0 \in \left[ \frac{a}{2},a \right] \right)
\left[ 1 - P_0\left( \tau_k > \rho a\ \bigg|\ \tau_0 \in \left[ \left( {\frac{1}{2}}+ \rho \right)a,a \right] \right) \right]. \notag
\end{aligned}$$ We need an upper bound for the conditional probability on the right side of . For some $K_{26},K_{27}$ we have from , , and Lemma \[lowheightlemma\] that provided $\chi$ is sufficiently small (depending on $\theta,\rho$), $$\begin{aligned}
\label{slowclimb2}
P_0&\left( \tau_k > \rho a\ \bigg|\ \tau_0 \in \left[ \left( {\frac{1}{2}}+ \rho \right)a,a \right] \right) \\
&\leq \frac{ P_0( \tau_k \wedge \tau_0 > \rho a) }
{ P_0\left( \tau_0 \in \left[ \left( {\frac{1}{2}}+ \rho \right)a,a \right] \right) } \notag \\
&\leq K_{26} \frac{ e^{-K_{4} \rho a/k^2} }{ M_k a^{-\kappa}L(\sqrt{a}) } \notag \\
&\leq K_{27} \frac{L(k)}{L(\sqrt{a})} \left( \frac{a}{k^2} \right)^\kappa e^{-K_{4} \rho a/k^2} \notag \\
&\leq \theta. \notag
\end{aligned}$$
Now , , and show that $$\begin{aligned}
\label{latticelower2}
P_0\left( \tau_0 - \tau_k \in \left[ \frac{a}{2},a \right] \right)) &\geq (1-2\theta) P_0\left( \tau_0 \in \left[ \frac{a}{2},a \right] \right) \\
&\geq (1-3\theta) P_0\left( \tau_0 \in \left[ \frac{a}{2}-1,a-1 \right] \right) \notag
\end{aligned}$$ which with Corollary \[latticepath2\](ii), (iii) shows that $$\label{corollaryuse}
(1-3\theta) M_k P_0(\tau_0 = m) \leq P_k(\tau_0 = m) \leq M_k P_0(\tau_0 = m), \quad m \text{ even},$$ $$\label{corollaryuse2}
(1-3\theta) M_k P_0(\tau_0 = m-1) \leq P_k(\tau_0 = m) \leq M_k P_0(\tau_0 = m-1), \quad m \text{ odd}.$$ This and prove .
Next we prove Theorem \[hittime\] for midrange starting heights–suppose $$\sqrt{m\chi} \leq k \leq \sqrt{ \frac{m}{\chi} }.$$ Let $\theta>0$ and let $0 < {\Delta_{\min}}< {\Delta_{\max}}$ be as in Proposition \[intervalmid\]. For the first inequality in we use Corollary \[latticepath2\](i) and Proposition \[intervalmid\]. From Corollary \[latticepath2\](i) and Theorem \[tau0tail\], for $m$ large with $m-k$ even, and $0 \leq j < {\Delta_{\min}}m$ with $m-j$ even, provided ${\Delta_{\min}}$ is small enough (depending on $\theta$), we have $$\label{comparedowne}
f_m^{(k)} \geq \frac{ f_m^{(0)} }{ f_{m-j}^{(0)} } f_{m-j}^{(k)} \geq (1-\theta) f_{m-j}^{(k)} \quad \text{if $k$ is even},$$ $$\label{comparedowno}
f_m^{(k)} \geq \frac{ f_{m-1}^{(0)} }{ f_{m-j-2}^{(0)} } f_{m-j-1}^{(k)} \geq (1-\theta) f_{m-j-1}^{(k)} \quad \text{if $k$ is odd},$$ so that, averaging over $j$ and applying Proposition \[intervalmid\], provided ${\Delta_{\min}}$ is small enough (depending on $\chi,\theta$), $$\begin{aligned}
\label{averaging}
P_k(\tau_0 = m) &\geq (1-2\theta) \frac{2}{{\Delta_{\min}}m} P_k( (1-{\Delta_{\min}})m < \tau_0 < m) \\
&\geq (1-3\theta) \frac{2}{\Gamma(k)m} \left( \frac{k^2}{2m} \right)^\kappa e^{-k^2/2m}. \notag
\end{aligned}$$ For the second inequality in the proof is similar: in place of and we have that for $m+j$ even, $$\label{compareupe}
f_m^{(k)} \leq \frac{f_m^{(0)}}{f_{m+j}^{(0)}} f_{m+j}^{(k)} \leq (1+\theta) f_{m+j}^{(k)} \quad \text{if $k$ is even},$$ $$\label{compareupo}
f_m^{(k)} \leq \frac{f_{m-1}^{(0)}}{f_{m+j}^{(0)}} f_{m+j+1}^{(k)} \leq (1+\theta) f_{m+j+1}^{(k)} \quad \text{if $k$ is odd}.$$ and then as with , $$\begin{aligned}
\label{averaging2}
P_k(\tau_0=m) &\leq (1+\theta) \frac{2}{{\Delta_{\min}}m} P_k\big( \tau_0 \in [m,(1+{\Delta_{\min}})m] \big) \notag \\
&\leq (1+3\theta)\frac{2}{\Gamma(\kappa)m} \left( \frac{k^2}{2m} \right)^\kappa e^{-k^2/2m},\end{aligned}$$ completing the proof of .
Last, we prove Theorem \[hittime\] for high starting heights. We may assume $k \leq m$. From the first inequalities in and and from Theorem \[tau0tail\], averaging over $j \in [0,m/8]$ we obtain that for $m$ large and $0<h<k/3$ we have $$\begin{aligned}
\label{averaging3}
P_k(\tau_0=m) &\leq \left( \frac{9}{8} \right)^\kappa \frac{32}{m}
P_k\left( \tau_0 \in \left[m,\frac{9}{8}m\right] \right) \notag \\
&\leq \left( \frac{9}{8} \right)^\kappa \frac{32}{m} P_k\left( \tau_h \leq \frac{9}{8}m \right).\end{aligned}$$ To bound the last probability we couple our Bessel-like RW to a symmetric simple RW. Recall that $N(t)$ denotes the number of alarms by time $t$, and let $$N^* = \left| \left\{i \leq \frac{9}{8}m: X_i \geq h, X_i^{\operatorname{sym}} \geq h, \text{and an alarm occurs at time } i \right\} \right|.$$ Analogously to we have $$\begin{aligned}
\label{twoways}
P_k\left( \tau_h \leq \frac{9}{8}m \right) &= P_k^*\left( \tau_{3h}^{\operatorname{sym}} > \tau_h, \tau_h \leq \frac{9}{8}m \right)
+ P_k^*\left( \tau_{3h}^{\operatorname{sym}} \leq \tau_h \leq \frac{9}{8}m \right) \notag \\
&\leq P_k^*\left( N^* > h \right) + P_k^{\operatorname{sym}}\left( \tau_{3h}^{\operatorname{sym}} \leq \frac{9}{8}m \right).\end{aligned}$$ We now take $h= k/8$; we assume for convenience that $h$ is an integer. If $m_0$ (and hence $k$) is large enough, then $\sup_{x \geq k/8} |p_x - {\frac{1}{2}}| \leq 2(1+|\delta|)/k$. Then $N^*$ is stochastically smaller than a Binomial($9m/8,2(1+|\delta|)/k$) random variable. We apply Bennett’s Inequality (see Hoeffding [@Ho63]), which states that for a Binomial($n,p$) random variable $Y$ and $\lambda > np$, $$P(Y \geq \lambda) \leq e^{-\lambda\psi(np(1-p)/\lambda)},$$ where $\psi$ is the decreasing function $$\psi(x) = (1+x)\log\left( 1 + \frac{1}{x} \right) - 1.$$ For $x \leq 1/4$ we have $\psi(x) \geq 1$ and hence $\psi(x) \geq {\frac{1}{2}}(1+x)\log\left( 1 + \frac{1}{x} \right)\geq {\frac{1}{2}}\log\frac{1}{x}$. Therefore for $\lambda \geq 4np$, $$P(Y \geq \lambda) \leq e^{-(\lambda/2) \log(\lambda/np)},$$ and in particular, provided $\chi$ is sufficiently small we have $$\label{binombound}
P_k^*\left( N^* > \frac{k}{8} \right) \leq e^{-(k/16) \log( k^2/18(1+|\delta|) m)} \leq e^{-k/4} \leq e^{-k^2/4m}.$$ Also, again provided $\chi$ is small, by Hoeffding’s inequality [@Ho63], $$P_k^{\operatorname{sym}}\left( \tau_{k/4}^{\operatorname{sym}} \leq \frac{9}{8}m \right) \leq 2P_0^{\operatorname{sym}}\left(X_{9m/8}^{\operatorname{sym}} > \frac{3}{4}k \right)
\leq e^{-k^2/4m},$$ which with , and yields $$\label{highconcl}
P_k(\tau_0=m) \leq \left( \frac{9}{8} \right)^\kappa \frac{64}{m} e^{-k^2/4m} \leq \frac{1}{m} e^{-k^2/8m},$$ completing the proof of .\
Proof of Theorems \[distrib\] and \[location\]
==============================================
Theorem \[location\] is a straightforward consequence of Theorem \[distrib\], and , so we prove Theorem \[distrib\]. We use , applying Theorem \[hittime\] and — to approximate the products on the right side.
We consider first part (i), for low starting heights, i.e. $1 \leq k < \sqrt{\chi n}$. By , and and Theorem \[hittime\], given $\theta>0$, taking $\rho$ and then $\chi$ sufficiently small, for $n$ large, we have the following sandwich bound for $P_k(X_n = 0)$: $$\begin{aligned}
\label{lowstart}
(1&-2\theta) P_0(X_{\tilde{n}} = 0) \\
&\leq (1-\theta) \min_{(1-\rho)n \leq j \leq n} P_0(X_{\tilde{j}} = 0) \notag \\
&\leq P_k(\tau_0 \leq \rho n) \min_{(1-\rho)n \leq j \leq n} P_0(X_{\tilde{j}} = 0) \notag \\
&\leq P_k(X_n = 0) \notag \\
&\leq \max_{(1-\rho)n \leq j \leq n} P_0(X_{\tilde{j}} = 0) + \sum_{0 \leq j < (1-\rho)n} P_k(\tau_0 = n-j) P_0(X_j = 0) \notag \\
&\leq (1+\theta) P_0(X_{\tilde{n}} = 0) + \max_{0 \leq i < (1-\rho)n} P_k(\tau_0 = n-i) \sum_{0 \leq j < (1-\rho)n} P_0(X_j = 0) \notag \\
&\leq (1+\theta) P_0(X_{\tilde{n}} = 0) + K_{28} (\rho n)^{-(\kappa+1)}L(\sqrt{n}) M_{\sqrt{\chi n}} \sum_{0 \leq j < (1-\rho)n} P_0(X_j = 0). \notag
\end{aligned}$$ We need to show that the second term on the right side of is small compared to the first term on the right side. From — we see that for some $K_{29}$, in all three cases, the sum in that second term is bounded by $K_{29}n P_0(X_{\tilde{n}} = 0)$. Therefore, using and , if $\chi$ is sufficiently small (depending on $\theta,\rho$) then for large $n$, the second term is bounded above by $$2K_{28}K_{29} \frac{ \chi^\kappa }{ \rho^{\kappa + 1} } P_0(X_{\tilde{n}} = 0) \leq \theta P_0(X_{\tilde{n}} = 0).$$ With this gives $$\label{lowstart1}
(1-2\theta) P_0(X_{\tilde{n}} = 0) \leq P_k(X_n = 0) \leq (1+2\theta) P_0(X_{\tilde{n}} = 0),$$ as desired.
Next we consider part (ii), for $E_0(\tau_0)<\infty$ and midrange starting heights, $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$. By there exists $n_1$ such that $$\frac{2-\theta}{E_0(\tau_0)} \leq P_0(X_n = 0) \leq \frac{2+\theta}{E_0(\tau_0)} \quad \text{for all even } n \geq n_1.$$ Let $0 < {\tilde{\chi}}< \chi$. Then using and Theorem \[hittime\] (with ${\tilde{\chi}}$ in place of $\chi$), provided ${\tilde{\chi}}$ is sufficiently small, and then $n$ (and hence $k$) is sufficiently large, $$\begin{aligned}
\label{midstart1}
P_k(X_n=0) &\geq \sum_{j=n_1}^{n-{\tilde{\chi}}k^2} P_k(\tau_0 = n-j) P_0(X_j = 0) \\
&\geq \frac{2-\theta}{E_0(\tau_0)} \sum_{m:{\tilde{\chi}}k^2 \leq m \leq n-n_1,m-k\text{ even}} P_k(\tau_0=m) \notag \\
&\geq \frac{2-3\theta}{E_0(\tau_0)} \sum_{m:{\tilde{\chi}}k^2 \leq m \leq n-n_1,m-k\text{ even}}
\frac{2}{\Gamma(\kappa)m} \left( \frac{k^2}{2m} \right)^\kappa e^{-k^2/2m} \notag \\
&\geq \frac{2-4\theta}{E_0(\tau_0)} \int_{{\tilde{\chi}}k^2}^{n-n_1} \frac{1}{\Gamma(\kappa)x} \left( \frac{k^2}{2x} \right)^\kappa e^{-k^2/2x}\ dx \notag \\
&= \frac{2-4\theta}{E_0(\tau_0)} \int_{k^2/2(n-n_1)}^{1/2{\tilde{\chi}}} \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \notag \\
&\geq \frac{2-5\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du. \notag
\end{aligned}$$ In the other direction, we have similarly $$\label{midstart2}
\sum_{j=n_1}^{n-{\tilde{\chi}}k^2} P_k(\tau_0 = n-j) P_0(X_j = 0)
\leq \frac{2+5\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du.$$ Also similarly to , given $\alpha>0$ we have for sufficiently small ${\tilde{\chi}}$ that provided $n$ is large, $$\label{mostprob}
P_k({\tilde{\chi}}k^2 \leq \tau_0 \leq k^2/{\tilde{\chi}}) \geq (1-\alpha) \int_{{\tilde{\chi}}/2}^{1/2{\tilde{\chi}}} \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \geq 1-2\alpha,$$ so in particular, for small ${\tilde{\chi}}$, $$P_k(\tau_0 < {\tilde{\chi}}k^2) \leq \theta \int_{1/2\chi}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du.$$ With , , and Theorem \[hittime\] this gives $$\begin{aligned}
\label{midstart3}
P_k(X_n=0) &\leq \frac{2+5\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \\
&\qquad + P_k(n-n_1 < \tau_0 \leq n) + P_k(\tau_0 < {\tilde{\chi}}k^2) \max_{n-{\tilde{\chi}}k^2/2 < j \leq n} P_0(X_j = 0) \notag \\
&\leq \frac{2+5\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \notag \\
&\qquad + \frac{K_{30}n_1}{n} \left( \frac{k^2}{2n} \right)^\kappa e^{-k^2/2n}
+ \frac{3\theta}{E_0(\tau_0)} \int_{1/2\chi}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \notag \\
&\leq \frac{2+9\theta}{E_0(\tau_0)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du, \notag
\end{aligned}$$ which with proves Theorem \[distrib\](ii) for midrange starting heights.
Now consider part (ii) for high starting heights, $k > \sqrt{n/\chi}$. We may assume $\theta<1$. Analogously to we have using and Theorem \[hittime\] that $$\begin{aligned}
\label{highstart1}
\sum_{j=n_1+1}^{n-k} P_k(\tau_0 = n-j) P_0(X_j = 0)
&\leq \frac{3}{E_0(\tau_0)} \sum_{m:k \leq m \leq n-n_1-1,m-k\text{ even}}
\frac{1}{m} e^{-k^2/8m} \notag \\
&\leq \frac{3}{E_0(\tau_0)} \frac{k+1}{k} \int_k^{n-n_1} \frac{1}{x} e^{-k^2/8x}\ dx \notag \\
&= \frac{3}{E_0(\tau_0)} \frac{k+1}{k} \int_{k^2/8(n-n_1)}^{k/8} e^{-u}\ du \notag \\
&\leq \frac{4}{E_0(\tau_0)} e^{-k^2/8n}.
\end{aligned}$$ Further, as in , using Theorem \[hittime\], $$\begin{aligned}
\label{highstart2}
\sum_{j=0}^{n_1} P_k(\tau_0 = n-j) P_0(X_j = 0)
&\leq \left( \max_{j \leq n_1} P_k(\tau_0 = n-j) \right) \sum_{j=0}^{n_1-1} P_0(X_j = 0) \notag \\
&\leq \frac{n_1+1}{n} e^{-k^2/8n}.
\end{aligned}$$ Now , and prove .
We turn now to part (iii), for $-1<\delta<1$ and midrange starting heights $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$. We use the fact that $$\label{Psik}
\int_0^1 \frac{1}{(1-u)^{1+\kappa} u^{1-\kappa}} e^{-a/(1-u)}\ du = \Gamma(\kappa) a^{-\kappa} e^{-a} \quad
\text{for all } a, \kappa > 0,$$ as can easily be seen via the change of variable $v = (1-u)^{-1}$. By there exists $n_2=n_2(\theta)$ such that $$\label{returnzero}
\frac{(1-\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} \leq P_0(X_n = 0)
\leq \frac{(1+\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1}$$ for all even $n \geq n_2$. Analogously to , provided $\tilde{\chi}/\chi$ is sufficiently small, using , and we then obtain that for large $n$, $$\begin{aligned}
\label{midstart4}
P_k&(X_n=0) \\
&\geq \sum_{j=n_2}^{n-{\tilde{\chi}}k^2} P_k(\tau_0 = n-j) P_0(X_j = 0) \notag \\
&\geq \frac{(2-\theta)2^\kappa K_0}{\Gamma(\kappa)\Gamma(1-\kappa)} \sum_{n_2 \leq j \leq n-{\tilde{\chi}}k^2 \atop j\text{ even}} \frac{1}{n-j}
\left( \frac{k^2}{2(n-j)} \right)^\kappa e^{-k^2/2(n-j)} j^{-(1-\kappa)}L(\sqrt{j})^{-1} \notag \\
&\geq \frac{(1-\theta)2^\kappa K_0}{\Gamma(\kappa)\Gamma(1-\kappa)} L(\sqrt{n})^{-1}
\int_{n_2}^{n-{\tilde{\chi}}k^2} \frac{1}{n-x} \left( \frac{k^2}{2(n-x)} \right)^\kappa e^{-k^2/2(n-x)} \frac{1}{x^{1-\kappa}}\ dx \notag \\
&= \frac{(1-\theta)2^\kappa K_0}{\Gamma(\kappa)\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} \left( \frac{k^2}{2n} \right)^\kappa
\int_{n_2/n}^{1-{\tilde{\chi}}k^2/n} \frac{1}{(1-u)^{1+\kappa} u^{1-\kappa}} e^{-k^2/2n(1-u)}\ du \notag \\
&\geq \frac{(1-2\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n}. \notag
\end{aligned}$$ In the other direction, analogously to , from a calculation similar to we get $$\label{midstart5}
\sum_{j=n_2}^{n-{\tilde{\chi}}k^2} P_k(\tau_0 = n-j) P_0(X_j = 0) \leq
\frac{(1+2\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n}.$$ With , and Theorem \[hittime\] this gives the analog of : provided ${\tilde{\chi}}$ is taken sufficiently small and then $n$ sufficiently large, $$\begin{aligned}
\label{midstart6}
P_k&(X_n=0) \\
&\leq \frac{(1+2\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n} \notag \\
&\qquad + n_2 \max_{0 \leq j < n_2} P_k(\tau_0 = n-j) + P_k(\tau_0 < {\tilde{\chi}}k^2) \max_{n-{\tilde{\chi}}k^2 < j \leq n} P_0(X_j=0) \notag \\
&\leq \frac{(1+2\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n} \notag \\
&\qquad + K_{31}(\chi) n^{-1}
+ \theta \frac{(1+2\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} \notag \\
&\leq \frac{(1+4\theta)2^\kappa K_0}{\Gamma(1-\kappa)} n^{-(1-\kappa)} L(\sqrt{n})^{-1} e^{-k^2/2n}. \notag\end{aligned}$$ Here the second inequality uses the fact that by , we can make $P_k(\tau_0 < {\tilde{\chi}}k^2)$ as small as desired by taking ${\tilde{\chi}}$ small. Together and prove Theorem \[distrib\](iii) for midrange starting heights.
We turn next to part (iii) for high starting heights, $k > \sqrt{n/\chi}$. There exists $K_{32}$ such that for $0<\alpha \leq K_{32}^2$, $$\label{decaysum}
\sum_{j=1}^\infty e^{-\alpha j} \frac{1}{ j^{1-\kappa}L(\sqrt{j}) } \leq \frac{2}{\kappa} \alpha^{-\kappa} L\left( \frac{1}{\sqrt{\alpha}} \right)^{-1}.$$ Then analogously to and , when $k \leq K_{32}n$, using , and Theorem \[hittime\] we have for large $n$, $$\begin{aligned}
\label{highstart3}
\sum_{j=n_2}^{n-k} &P_k(\tau_0 = n-j) P_0(X_j = 0) \notag \\
&\leq \frac{2^{1+\kappa} K_0}{\Gamma(1-\kappa)} \sum_{j=n_2}^{n-k}
\frac{1}{n-j} e^{-k^2/8(n-j)} \frac{1}{ j^{1-\kappa}L(\sqrt{j}) } \notag \\
&\leq \frac{2^{2+\kappa} K_0}{\Gamma(1-\kappa)} \bigg[ \frac{2}{n} e^{-k^2/8n} \sum_{n_2\leq j\leq n/2} e^{-k^2j/8n^2}
\frac{1}{ j^{1-\kappa}L(\sqrt{j}) } \notag \\
&\qquad \qquad \qquad + \sum_{n/2<j\leq n-k}
\frac{1}{n-j} e^{-k^2/8(n-j)} \frac{1}{ j^{1-\kappa}L(\sqrt{j}) }\bigg] \notag \\
&\leq \frac{2^{2+\kappa} K_0}{\Gamma(1-\kappa)} \bigg[ \frac{4\cdot 4^\kappa}{\kappa n} e^{-k^2/8n} \left( \frac{n}{k} \right)^{2\kappa}
L\left( \frac{n}{k} \right)^{-1} + \frac{2}{n} e^{-k^2/4n} \sum_{j=1}^n
\frac{1}{ j^{1-\kappa}L(\sqrt{j}) }\bigg] \notag \\
&\leq \frac{2^{2+\kappa} K_0}{\Gamma(1-\kappa)} \bigg[ \frac{4\cdot 4^\kappa}{\kappa n} e^{-k^2/8n} \left( \frac{n}{k} \right)^{2\kappa}
L\left( \frac{n}{k} \right)^{-1} + \frac{4}{\kappa n} e^{-k^2/4n} n^\kappa L(\sqrt{n})^{-1} \bigg].\end{aligned}$$ Here in the third inequality we used the fact that $(n-j)^{-1} e^{-k^2/8(n-j)}$ is a decreasing function of $j$, and the fact that $L$ is slowly varying. Provided $\chi$ is sufficiently small, the second term inside the brackets on the right side of is smaller than the first term; using this, and Theorem \[hittime\] we obtain that for some $K_{33}(\kappa)$, provided $\chi$ is small enough, $$\begin{aligned}
\label{highstart4}
P_k(X_n = 0) &\leq \frac{2^{4+3\kappa} K_0K_{32}}{\kappa\Gamma(1-\kappa)n} e^{-k^2/8n} \left( \frac{n}{k} \right)^{2\kappa}
L\left( \frac{n}{k} \right)^{-1} + \sum_{0 \leq j < n_2} P_k(\tau_0 = n-j) \notag \\
&\leq \frac{2^{4+3\kappa} K_0K_{32}}{\kappa\Gamma(1-\kappa)n} e^{-k^2/8n} \left( \frac{n}{k} \right)^{2\kappa}
L\left( \frac{n}{k} \right)^{-1} + \frac{2n_2}{n} e^{-k^2/8n} \notag \\
&\leq K_{33} e^{-k^2/8n} n^{-(1-\kappa)} L(\sqrt{n})^{-1}.\end{aligned}$$ Here $n_2=n_2(1)$. This proves when $k \leq K_{32}n$. If $K_{32}n<k\leq n$, in place of and we have using that $$\begin{aligned}
\label{highstart7}
P_k(X_n = 0) &= \sum_{j=0}^{n-k} P_k(\tau_0 = n-j) P_0(X_j = 0) \notag \\
&\leq \sum_{j=0}^{n-k} \frac{1}{n-j} e^{-k^2/8n} \notag \\
&\leq K_{34} e^{-k^2/8n},\end{aligned}$$ from which follows.
Next we consider part (iv), in which $\delta=1$, $E_0(\tau_0) = \infty$, in the case of midrange starting heights $\sqrt{n\chi} \leq k \leq \sqrt{n/\chi}$. In this case $\mu_0$ is slowly varying, and by there exists $n_3=n_3(\theta)$ such that $$\label{returnzero2}
\frac{2-\theta}{\mu_0(n)} \leq P_0(X_n = 0) \leq \frac{2+\theta}{\mu_0(n)} \quad \text{for all even } n \geq n_3.$$ Then analogously to , using Theorem \[hittime\], and , for large $n$, $$\begin{aligned}
\label{midstart7}
P_k(X_n = 0) &\geq \sum_{j=n_3}^{n-{\tilde{\chi}}k^2} P_k(\tau_0 = n-j) P_0(X_j = 0) \\
&\geq \frac{4-3\theta}{\Gamma(\kappa)} \sum_{{\tilde{\chi}}k^2 \leq m \leq n-n_3 \atop n-m\text{ even}} \frac{1}{m}
\left( \frac{k^2}{2m} \right)^\kappa e^{-k^2/2m} \frac{1}{\mu_0(n-m)} \notag \\
&\geq \frac{2-2\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du, \notag\end{aligned}$$ and similarly $$\label{midstart8}
\sum_{j=n_3}^{n-{\tilde{\chi}}k^2} P_k(\tau_0 = n-j) P_0(X_j = 0)
\leq \frac{2+2\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du.$$ Using , , and Theorem \[hittime\], and taking $\tilde{\chi}$ sufficiently small, we obtain the analog of : $$\begin{aligned}
\label{midstart9}
P_k&(X_n = 0) \\
&\leq \frac{2+2\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \notag \\
&\qquad + n_3 \max_{0 \leq j < n_3} P_k(\tau_0 = n-j) + P_k(\tau_0 < {\tilde{\chi}}k^2) \max_{n-{\tilde{\chi}}k^2 < j \leq n} P_0(X_j=0) \notag \\
&\leq \frac{2+2\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \notag \\
&\qquad + K_{35}(\chi) n^{-1}
+ \frac{\theta}{\mu_0(n)} \int_{1/2\chi}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du \notag \\
&\leq \frac{2+4\theta}{\mu_0(n)} \int_{k^2/2n}^\infty \frac{1}{\Gamma(\kappa)} u^{\kappa-1} e^{-u}\ du. \notag\end{aligned}$$ Together and prove Theorem \[distrib\](iv) for midrange starting heights.
Last we consider part (iv) for high starting heights, $k > \sqrt{n/\chi}$. We may assume $\theta<1$. When $k \leq K_{32}n$ and $k$ is sufficiently large, we have analogously to , using and Theorem \[hittime\], $$\begin{aligned}
\label{highstart5}
\sum_{j=n_3}^{n-k} &P_k(\tau_0 = n-j) P_0(X_j = 0) \notag \\
&\leq \sum_{j=n_3}^{n-k} \frac{1}{n-j} e^{-k^2/8(n-j)}\frac{3}{\mu_0(j)} \notag \\
&\leq \frac{6}{n} e^{-k^2/8n} \sum_{n_3\leq j\leq n/2} e^{-k^2j/8n^2}
\frac{1}{ \mu_0(j) } + \frac{6}{n} e^{-k^2/4n} \sum_{n/2<j\leq n-k} \frac{1}{ \mu_0(j) } \notag \\
&\leq \frac{96}{n} e^{-k^2/8n} \frac{n^2}{k^2} \mu_0\left( \frac{n^2}{k^2} \right)^{-1} + e^{-k^2/4n} \frac{6}{\mu_0(n)} \notag \\
&\leq e^{-k^2/8n} \frac{7}{\mu_0(n)}.\end{aligned}$$ In the last inequality we have bounded $(n^2/k^2)\mu_0(n^2/k^2)^{-1}$ by $n/96\mu_0(n)$, valid for $\chi$ sufficiently small because $n^2/k^2 \leq \chi^2 n$ and $\mu_0$ is slowly varying. Then using and , $$\begin{aligned}
\label{highstart6}
P_k(X_n = 0) &\leq e^{-k^2/8n} \frac{7}{\mu_0(n)} + \sum_{1 \leq j < n_3} P_k(\tau_0 = n-j) \notag \\
&\leq e^{-k^2/8n} \frac{7}{\mu_0(n)} + \frac{n_3}{n} e^{-k^2/8n} \notag \\
&\leq e^{-k^2/8n} \frac{8}{\mu_0(n)}.
\end{aligned}$$ If instead $K_{32}n<k\leq n$, then is valid. In fact, a look at shows that, by reducing $\chi$ if necessary, we can replace 8 on the right side of with any constant greater than 4. Therefore in place of we have for large $n$ that $$\begin{aligned}
\label{highstart8}
P_k(X_n = 0) \leq K_{34} e^{-k^2/6n} \leq \frac{1}{\mu_0(n)} e^{-k^2/8n}.\end{aligned}$$ Thus holds in both cases.
Acknowledgements
================
The author thanks P. Baxendale, F. Dunlop, T. Huillet, J. Pitman and A. Wade for helpful discussions, and thanks a referee for multiple references and suggestions.
[99]{}
Alexander, K. S. (2008). The effect of disorder on polymer depinning transitions. *Commun. Math. Phys.* **279** 117–146. MR2377630
Alexander, K.S. and Zygouras, N. (2009). Quenched and annealed critical points in polymer pinning models. *Commun. Math. Phys.* **291** 659–689. MR2534789
Aspandiiarov, S. and Iasnogorodski, R. (1997). Tails of passage-times and an application to stochastic processes with boundary reflection in wedges. *Stoch. Proc. Appl.* **66**, 115–145.
Aspandiiarov, S. and Iasnogorodski, R. (1999). General criteria of integrability of functions of passage-times for nonnegative stochstic processes and their applications. *Theor. Probab. Appl.* **43**, 343–369.
Aspandiiarov, S., Iasnogorodski, R. and Menshikov, M. (1996). Passage-time moments for nonnegative stochastic processes and an application to reflected random walks in a quadrant. *Ann. Probab.* **24**, 932–960.
Brezis, H., Rosenkrantz, W. and Singer, B. (1971). An extension of Khintchine’s estimate for large deviations to a class of Markov chains converging to a singular diffusion. *Comm. Pure Appl. Math.* **24**, 705–726.
Carmona, P. and Hu, Y. (2002). On the partition function of a directed polymer in a random environment. *Probab. Theory. Rel. Fields*, **124**, 431-457.
Coolen-Schrijner, P. and van Doorn, E. A. (1998). Analysis of random walks using orthogonal polynomials. *J. Comp. Appl. Math.* **99**, 387–399.
Comets, F. and Yoshida, N. (2007). Directed polymers in random environments are diffusive at weak disorder. *Ann. Probab.* **34**, 1746-1770.
Csáki, E., Földes, A. and Révész, P. (2009). Transient nearest neighbor random walk and Bessel process. *J. Theoret. Probab.* **22**, 992–1009.
de Coninck, J., Dunlop, F. and Huillet, T. (2008). Random walk weakly attracted to a wall. *J. Stat. Phys.* **133**, 271–280.
de Coninck, J., Dunlop, F. and Huillet, T. (2009). Random walk versus random line. *Phys. A* **388**, 4034–4040.
Dette, H. (2001). First return probabilities of birth and death chains and associated orthogonal polynomials. *Proc. Amer. Math. Soc.* **129**, 1805–1815.
Dette, H., Fill, J. A., Pitman, J. and Studden, W. J. (1997). Wall and Seigmund duality relations for birth and death chains with reflecting barrier. *J. Theor. Probab.* **10**, 349–374.
Doney, R.A. (1997). One-sided local large deviation and renewal theorems in the case of infinite mean. *Probab. Theory Rel. Fields* **107**, 451–465.
Erickson, K. B. (1970). Strong renewal theorems with infinite mean. *Trans. Amer. Math. Soc.* **151**, 263–291.
Fal’, A. M. (1981). Certain limit theorems for an elementary Markov random walk. *Ukrainian Math. J.* **33**, 433–435. \[Translated from *Ukrain. Mat. Zh.* **33** (1981), 564–566.\]
Fal’, A. M. (1973). On the simplest Markov random walk. *Soviet Math. Dokl.* **14**, 1071–1074. \[Translated from *Dokl. Akad. Nauk SSSR* **211** (1973), 540–542.\]
Feller, W. (1968). *An Introduction to Probability Theory and its Applications*, 3rd ed. Wiley, New York.
Giacomin, G. (2007). *Random Polymer Models*. Imperial College Press, London.
Giacomin, G., Lacoin, H. and Toninelli, F. (2009). Disorder relevance at marginality and critical point shift. arXiv:0906.1942v1 \[math-ph\]
Giacomin, G. and Toninelli, F. (2005). Smoothing effect of quenched disorder on polymer depinning transitions. *Commun. Math. Phys.* **266** 1–16. MR2231963
Göing-Jaeschke, A. and Yor, M. (2003). A survey and some generalizations of Bessel processes. *Bernoulli* **9**, 313–349.
Halpin-Healy, T. and Zhang, Y.-C. (1995). Kinetic roughening phenomena, stochastic growth, directed polymers and all that. *Phys. Reports* **254**, 215–414.
Hodges, J. L., Jr. and Rosenblatt, M. (1953). Recurrence-time moments in random walks. *Pacific J. Math.* **3**, 127–136.
Hoeffding, W. (1963). Probability inequalities for sums of bounded random variables. *J. Amer. Statist. Assoc.* **58**, 13–30.
Jain, N. C. and Pruitt, W.E. (1972). The range of random walk. In *Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. III: Probability Theory*, 31-50. Univ. California Press, Berkeley, Calif. MR0410936
Karlin, S. and McGregor, J. L. (1957a). The differential equations of birth-and-death processes, and the Stieltjes moment problem. *Trans. Amer. Math. Soc.* **85**, 489–546.
Karlin, S. and McGregor, J. L. (1957b). The classification of birth and death processes. *Trans. Amer. Math. Soc.* **86**, 366–400.
Karlin, S. and McGregor, J. (1959). Random walks. *Ill. J. Math.* **3**, 66–81.
Lacoin, H. (2010). New bounds for the free energy of directed polymers in dimension $1+1$ and $1+2$. *Commun. Math. Phys.* **294** 471–503.
Lamperti, J. (1962). A new class of probability limit theorems. *J. Math. Mech.* **11**, 749–772.
Lamperti, J. (1963). Criteria for stochastic processes II: Passage-time moments. *J. Math. Anal. Appl.* **7**, 127–145.
Menshikov, M. V. and Popov, S. Yu. (1995). Exact power estimates for countable Markov chains. *Markov. Proc. Rel. Fields* **1**, 57–78.
Menshikov, M. V., Vachkovskaia, M. and Wade, A. R. (2008). Asymptotic behaviour of randomly reflecting billiards in unbounded tubular domains. *J. Stat. Phys.* **132**, 1097–1133.
Revuz, D. and Yor, M. (1991). *Continuous Martingales and Brownian Motion*. Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Sciences\], **293**. Springer-Verlag, Berlin, New York.
Rosenkrantz, W. A. (1966). A local limit theorem for a certain class of random walks. *Ann. Math. Statist.* **37**, 855–859.
[^1]: This research was supported by NSF grant DMS-0804934.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have developed a multi-phase SPH method to simulate arbitrary interfaces containing surface active agents (surfactants) that locally change the properties of the interface, such the surface tension coefficient [@Adami2010]. Our method incorporates the effects of surface diffusion, transport of surfactant from/to the bulk phase to/from the interface and diffusion in the bulk phase. Neglecting transport mechanisms, we use this method to study the impact of insoluble surfactants on drop deformation and breakup in simple shear flow and present the results in a fluid dynamics video.'
author:
- |
S. Adami, X.Y. Hu and N.A. Adams\
\
Institute of Aerodynamics,\
Technical University Munich, 85748 Garching, Germany
bibliography:
- 'bibdata.bib'
title: Tipstreaming of a drop in simple shear flow in the presence of surfactant
---
Introduction
============
Exposing drops to extensional flows such as e.g. a simple Couette flow, the viscous forces along the interface tend to deform the drop and elongate it to an ellipsoid-type shape. The balancing force to stop the deformation due to the shearing is the surface tension. When two pure fluids of different types are in contact, the resulting surface tension force is only proportional to the local curvature and is normal to the interface. Depending on the strength of this force and the viscosity ratio between the two fluids $\lambda$, drops are deformed to a steady ellipsoid shape or break up. The correlation between the breakup behaviour and the flow parameter expressed by the Capillar number $Ca\left(\lambda\right)$ is known as the Grace curve [@Grace1982].
Adding surface active agents (surfactants) to a multiphase system can strongly alter the flow phenomena. Surfactants mainly affect the surface tension coefficients between two fluids when replacing fluid molecules at the interface with surfactant molecules. Hence, surface tension gradients along the interface can occur, resulting in Marangoni forces [@Scriven1960].
Here, we only focus on the case of insoluble surfactants, i.e. surfactants are initially added to the interface and cannot dissolve to the adjacent fluid phases. Bazhlekov et al. [@Bazhlekov2006] studied the effect of insoluble surfactants on drop deformation and breakup in simple shear flow with a boundary-integral method and clearly describe the different breakup modes. However, due to the nature of their method, an interface capturing scheme is required and breakup detection requires special procedures. By the use of a Lagrangian particle method we can avoid these steps and can handle strong interface deformations naturally.
Governing equations
===================
Following the well-known weakly compressible SPH method [@Monaghan2005], we use an equation of state (EOS) to relate the pressure to the density and solve the isothermal Navier-Stokes equations on a moving Lagrangian frame. Formulating the surface force as the gradient of the surface stress tensor [@Brackbill1992] $$\textbf{F}^{\left(s\right)} = \nabla \cdot \left[ \alpha \left( \mathbf{I} -
\mathbf{n}\otimes \mathbf{n}\right)\delta_{\Sigma} \right]
= - \left( \alpha \kappa \mathbf{n} + \nabla_{s} \alpha\right) \delta_{\Sigma} ~,$$ surface tension can be split into a normal Capillary force $\alpha \kappa \mathbf{n} \delta_{\Sigma}$ with the curvature $\kappa$, the normal vector of the interface $\mathbf{n}$ and the surface delta function $\delta_{\Sigma}$ and the tangential Marangoni force $\nabla_s \alpha \delta_{\Sigma}$ ($\nabla_{s}$ is the surface gradient operator $\nabla_{s} = \left(\mathbf{I} - \mathbf{n}\otimes \mathbf{n}\right)\nabla$).
Assuming insolubility in the phases, the evolution of surfactant on the interface is governed by a diffusion equation $$\frac{d \Gamma}{dt} = \nabla_s \cdot \mathbf{D_s} \nabla_s \Gamma~,
\label{eq:surfactant}$$ where $\Gamma$ and $\mathbf{D_s}$ are the interfacial surfactant concentration and the diffusion coefficient matrix (in case of isotropic diffusion $\mathbf{D_s} = D_s\cdot \mathbf{I}$), respectively.
To close our model, we relate the interfacial surfactant concentration $\Gamma$ to the surface-tension coefficient $\alpha$ by a constitutive equation. Widely used in literature, the Frumkin isotherm or the Langmuir model [@Tryggvason2001] are known to agree reasonably well with experimental data. By the purpose of visualization, however, a simple linear relation between $\alpha$ and $\Gamma$ is sufficient and is employed here.
Numerical results
=================
The results of our simulations are presented in [Video 1](file://anc/movie-tipstreaming-highres.mp4).
As a reference, in the first part of the movie we show the effect of the Capillary number on the droplet behaviour in the shear flow. At sub-critical conditions, the drop deforms to a steady ellipsoid. When the shear forces are strong enough, the surface forces cannot counterbalance the shearing of the drop and breakup occurs.
When surfactants are present at the interface, another very complex phenomenon can be observed, namely the so-called tipstreaming. At the tips of the deformed drop the interface is compressed and consequently, the surfactant concentration increases locally. Thus, surface tension gradients develop and Marangoni-forces change the behaviour of the drop dramatically. As the surface tension is very low at the tips of the drop, very sharp tips with a high curvature develop. Under critical conditions, this effect results in a singularity where the tipstreaming starts. In our simulations we can show that the thickness of the stream is of the order of the size of a particle.
At the end, we show the possibilities to influence the tipstreaming by changing the strength of the surface diffusion. When diffusion is strong, concentration gradients are smoothed and the Marangoni-forces are weaker. Hence, the curvature at the tips can be reduced and tipstreaming is suppressed. Another way of controlling the tipstreaming is by changing the properties of the surfactant, which is not shown here.
Acknowledgment {#acknowledgment .unnumbered}
==============
We want to acknowledge the support of J. Biddiscombe [@Biddiscombe2008] with *pv-meshless*, which was used to visualize the simulation data. Furthermore, we are grateful to R. Fraedrich, who provided the volume rendering of our results [@Fraedrich2010].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the relationship between (non-)Markovian evolutions, established correlations, and the entropy production rate. We consider a system qubit in contact with a thermal bath and in addition the system is strongly coupled to an ancillary qubit. We examine the steady state properties finding that the coupling leads to effective temperatures emerging in the composite system, and show that this is related to the creation of correlations between the qubits. By establishing the conditions under which the system reaches thermal equilibrium with the bath despite undergoing a non-Markovian evolution, we examine the entropy production rate, showing that its transient negativity is a sufficient sign of non-Markovianity.'
author:
- 'Maria Popovic,$^1$, Bassano Vacchini$^{1,2}$, and Steve Campbell$^{2}$'
bibliography:
- 'entropy\_production.bib'
title: 'Entropy production and correlations in a controlled non-Markovian setting'
---
Introduction
============
The inevitable interaction of a system with its surroundings necessitates we find suitable means for modelling the dynamics of open quantum systems [@Breuer2002]. Under certain situations, in particular when the coupling between the system and its environment is sufficiently weak such that the environment is unaffected by this interaction, the system evolves in a Markovian (memoryless) manner. For these settings, a well-known and widely used approach is to model the time evolution using a dynamical semi-group. Conversely, allowing for memory effects the scenario drastically changes [@CampbellPRA2012], and different approaches to the very definition of a non-Markovian dynamics have been recently introduced [@Vacchini2011a; @Vacchini2012a; @AddisPRA], as discussed in the recent reviews [@Rivas2014a; @BreuerRMP].
An equally important issue is developing a clear thermodynamic framework for quantum systems [@GooldJPA; @Alicki2018a]. Indeed, thermodynamic quantities, such as work and heat, must be carefully re-examined when the working materials are inherently quantum. In the thermodynamic characterization of a given process, the (irreversible) entropy production and the associated entropy production rates are crucial [@Spohn; @Alicki1979; @EspositoNJP; @DeffnerPRL2011; @DeffnerPRE2015]. The entropy production can be naturally defined as the difference between the change in entropy of the reduced system state and the mean exchanged heat with a bath at fixed temperature, $T$, divided by $T$. For the case of a quantum dynamical semigroup with a stationary state in Gibbs form the entropy production is guaranteed to be positive or zero, and can naturally be associated with a statement of the second law. An equivalent expression for the entropy production in a semigroup dynamics can be introduced also in the presence of an invariant state [@Spohn]. However, in general, this definition lacks a clear thermodynamic interpretation since the invariant state is not necessarily a thermal equilibrium state. To date several significant advances have been made in defining and understanding the thermodynamic entropy production for quantum systems [@PaternostroPRL2017; @PaternostroArXiv; @GooldArXiv; @BarbieriArXiv; @BrunelliArXiv; @SerraPRL2016; @DeffnerPRE2017], however only recently has the explicit consideration of non-Markovian maps, where negative entropy production rates can appear, been explored [@MarcantoniSciRep; @PatiPRA; @ManiscalcoSciRep].
It is in this direction that the present work progresses. We consider a two-level quantum system (qubit) immersed in a Markovian bath. The system is in turn strongly coupled to an ancillary two-level system such that the joint dynamics of the two qubits is Markovian, while the reduced dynamics of the system alone is manifestly non-Markovian. Thus, our setting significantly differs from other recent works, for example Refs. [@MarcantoniSciRep; @PatiPRA], as we have a direct access to the state of both the system and ancilla. Indeed, in our model we can consider the ancilla as a special subset of environmental degrees of freedom with which the system interacts and which gives rise to a non-Markovian evolution. Thus, at variance with other studies, one of the main goals of the present work is to assess the role that the establishment of correlations plays in the thermodynamic characterization of the evolution. By first characterizing the steady state properties, we show that the strong coupling can lead to a non-equilibrium steady state exhibiting correlations between the two qubits. These correlations can be related to the emergence of effective temperatures, different from that of the bath, for the system and the ancilla, making a thermodynamic description of the process more complex. Despite this, as the overall system+ancilla evolves under a Markovian map, a meaningful, albeit not necessarily thermodynamically meaningful, entropy production can be studied. By identifying the conditions under which the reduced system qubit reaches thermal equilibrium, we then study its associated entropy production and entropy production rate. While the long-time entropy production is consistent with the Markovian case, the coupling induced non-Markovian dynamics can lead to transiently negative rates. While such negative entropy production rates are due to the non-Markovian dynamics, the two notions are not necessarily commensurate, which we show by examining the trace distance measure of non-Markovianity.
The model
=========
We consider a bipartite system consisting of two coupled spin-$1/2$ particles (qubits), labelled $S$ and $A$. The free evolution of the qubits are governed by their respective Hamiltonians $\mathcal{H}_{S(A)}\!=\!\omega_{S(A)} \sigma_z$. Throughout we will denote ${\left\vert0\right\rangle}({\left\vert1\right\rangle})$ as their ground (excited) state and assume units such that $\hbar=k_B=1$. We will assume that only qubit $S$ feels the effects of an external thermal environment (bath) and thus the dynamics of the total system can be described by the Markovian master equation (omitting the explicit time dependence) [@CampbellPRA2010] $$\label{master}
\dot{\varrho}_{SA} = \mathcal{L}(\varrho_{SA}) = -i[\mathcal{H}_I +\mathcal{H}_S + \mathcal{H}_A , \varrho_{SA}] + \mathcal{D}(\varrho_{SA}),$$ with $\mathcal{H}_I= \left(J_x \sigma_x \otimes \sigma_x + J_y \sigma_y \otimes \sigma_y +J_z \sigma_z \otimes \sigma_z\right)$ defining the interaction between the two qubits and $$\begin{aligned}
\mathcal{D}(\varrho_{SA}) & = \gamma \left( (\sigma_- \otimes \openone) \varrho_{SA} (\sigma_+ \otimes \openone) - \frac{1}{2} \Big\{ \varrho_{SA} , (\sigma_+ \otimes \openone) (\sigma_- \otimes \openone) \Big\} \right) \\ & + \Gamma \left( (\sigma_+ \otimes \openone) \varrho_{SA} (\sigma_- \otimes \openone) - \frac{1}{2} \Big\{ \varrho_{SA} , (\sigma_- \otimes \openone) (\sigma_+ \otimes \openone) \Big\} \right),
\end{aligned}$$ where $\sigma_{+}\!=\!\sigma_-^\dagger\!=\!{\left\vert1\right\rangle}{\left\langle0\right\vert}$ are the spin raising and lowering operators, while $\gamma$ and $\Gamma$ fix the dissipation rates. The inverse temperature of the bath is given by $\beta=\tfrac{1}{2\omega_S}\ln\left(\tfrac{\gamma}{\Gamma}\right)$. The expression $\mathcal{H}_I$ of the interaction captures a broad range of important physical processes, including energy and excitation preserving models. The dynamics of the reduced state of the system $$\varrho_S = \text{Tr}_A \left[ \varrho_{SA} \right],$$ provides a physically legitimate evolution, crucially however, it is no longer Markovian due to the strong interaction with $A$. We remark similar settings have been explored in the literature [@CampbellPRA2010; @DeffnerPRL2015], and moreover, its general validity was recently examined [@ManiscalcoPRA].
Our model therefore provides a versatile setting to explore the interrelation between the establishment of strong correlations, both quantum and classical, non-Markovian dynamics, and the behavior of the entropy production. Indeed, our model not only allows to seamlessly move from a Markovian to a non-Markovian picture, but also allows for the study of non-equilibrium steady states with respect to the bath and clearly establish a relation with the emergence of such states and the correlations shared between $S$ and $A$.
Steady State Properties
=======================
As the steady state, $\varrho_{SA}^{\infty}$, plays a crucial role in evaluating the entropy production, in this section we focus on its characteristics. By solving the LHS of Eq. set equal to zero $\varrho_{SA}^{\infty}$ can be obtained fully analytically, however given its cumbersome form we do not report it here. Regardless of the explicit expressions, some properties are immediately evident, in particular the steady state is in $X$-form and independent of $J_z$, while all other parameters enter non-trivially. The presence of off-diagonal terms implies some correlations are established between the two qubits. We can examine the quantum correlations present using the entanglement of formation (EoF) $$\mathcal{E}\,{=}\,h\left(\frac{1}{2}\left[1+\sqrt{1-\mathcal{C}^2}\right]\right),$$ where $h(x)\,{=}\,-x\text{log}_2 x - (1-x)\text{log}_2 (1-x)$ is the binary entropy function and $\mathcal{C}$ is the concurrence of the state. The latter is an equally valid entanglement measure and can be found in terms of the eigenvalues $\lambda_1\,{\geq}\, \lambda_{2,3,4}$ of the spin-flipped density matrix $\rho_{AB}(\sigma_y\otimes\sigma_y)\rho^*_{AB} (\sigma_y\otimes\sigma_y)$ as $\mathcal{C}=\text{max}\left[0,\sqrt{\lambda_1}-\sum_{i=2}^{4}\sqrt{\lambda_i}\right]$. A more general measure of correlations in a state is given by the quantum mutual information (MI) $$\label{mutinfoeq}
{\cal I}(\rho_{12})={\cal S}(\rho_1)+{\cal S}(\rho_2)-{\cal S}(\rho_{12}),$$
[**(a)**]{}\
![[**(a)**]{} Steady state EoF and [**(b)**]{} MI (in base 2) versus the frequency, $\omega_A$ of the ancillary system $A$ and coupling strength $J_x$. [**(c)**]{} Effective inverse temperature, $\beta_\text{eff}$ of the system $S$. In all panels we have fixed $\gamma=10$, $\Gamma = 1$, $J_y=1$, $J_z=0$, and $\omega_S=1$.[]{data-label="fig1"}](fig1a.png "fig:"){width="0.85\columnwidth"}\
[**(b)**]{}\
![[**(a)**]{} Steady state EoF and [**(b)**]{} MI (in base 2) versus the frequency, $\omega_A$ of the ancillary system $A$ and coupling strength $J_x$. [**(c)**]{} Effective inverse temperature, $\beta_\text{eff}$ of the system $S$. In all panels we have fixed $\gamma=10$, $\Gamma = 1$, $J_y=1$, $J_z=0$, and $\omega_S=1$.[]{data-label="fig1"}](fig1b.png "fig:"){width="0.85\columnwidth"}\
[**(c)**]{}\
![[**(a)**]{} Steady state EoF and [**(b)**]{} MI (in base 2) versus the frequency, $\omega_A$ of the ancillary system $A$ and coupling strength $J_x$. [**(c)**]{} Effective inverse temperature, $\beta_\text{eff}$ of the system $S$. In all panels we have fixed $\gamma=10$, $\Gamma = 1$, $J_y=1$, $J_z=0$, and $\omega_S=1$.[]{data-label="fig1"}](fig1c.pdf "fig:"){width="0.8\columnwidth"}\
where ${\cal S}(\rho)\,{=}\,{-}\text{Tr}[\rho\log\rho]$ is the von Neumann entropy of a generic state $\rho$. This measure accounts for all correlations both classical and quantum. Fixing $J_y\!=\!1$ and $\omega_S\!=\!1$, in Fig. \[fig1\] [**(a)**]{} and [**(b)**]{} we find that both quantifiers depend non-trivially on the frequency detuning between the two spins and the coupling strength $J_x$. Immediately we see that there are significant parameter ranges where the steady state exhibits no entanglement, while the MI exhibits a markedly different dependence on the parameters, in particular, only being identically zero when $J_x\!=\!J_y$, as highlighted by the vertical white line. For other interaction strengths it is clear that the presence of correlations indicates that the strong coupling between $S$ and $A$ results in a non-equilibrium steady state with respect to the bath. As expected, the actual steady state compares to a thermal ansatz of $S\!+\!A$ given by a canonical Gibbs state, determined by the total Hamiltonian $\mathcal{H}_S + \mathcal{H}_A +\mathcal{H}_I$ and the inverse temperature of the bath $\beta$ only for $J_x\!=\!J_y$ and $\gamma\!\approx \!\Gamma$, which corresponds to an excitation preserving interaction in the limiting case of infinite temperature for the bath. As we show in the following, for other parameter choices we find that additional characteristic temperatures can emerge from $\varrho_{SA}^{\infty}$.
Due to the $X$-shape of the steady state, the reduced state of $S$, i.e. $\varrho_S^{\infty}\!=\!\text{Tr}_A[\varrho_{SA}^{\infty}]$ is diagonal for all parameter values, and therefore we can define an effective temperature, $\beta_\text{eff}$, for the system. In Fig. \[fig1\] [**(c)**]{} we examine how the strength of the coupling between $S$ and $A$ and their relative detuning affects $\beta_\text{eff}$. When correlations are present in the steady state, as captured by the MI, the effective temperature that the system reaches is higher than that of of the bath and the discrepancy is enhanced for smaller values of $\omega_A$, corresponding to the fact that a transfer of thermal excitations to the system is favoured. Interestingly, for $J_x\!=\!J_y$ we find that $S$ thermalizes with the environment regardless of the relative detuning between the two qubits. For this value of coupling the steady state is diagonal and takes the form $$\label{diagonalss}
\varrho_{SA}^{\infty} = \frac{\mathrm{e}^{-\beta \mathcal{H}_S}}{\mathcal{Z}_S} \otimes \frac{\mathrm{e}^{-\tilde{\beta} \mathcal{H}_A}}{\mathcal{Z}_A},$$ where $\tilde{\beta}=(\beta \omega_S)/\omega_A$. The factorized form of this state puts into clear evidence the vanishing MI shown in Fig. \[fig1\] [**(b)**]{}. It is worth stressing at this point that although $S$ is in thermal equilibrium with the bath for $J_x\!=\!J_y$, by virtue of the interaction, a second characteristic temperature emerges for $A$. The appearance of the two distinct parameters $\beta$ and $\tilde{\beta}$ can be understood as follows. For $J_x\!=\!J_y$ we have an excitation preserving interaction and the temperature associated to each system is naturally introduced by imposing a detailed balance condition for the energy exchange. For the system $S$ we have $\Gamma=\gamma \exp{(-\beta \omega_S)}$ to be compared with $\Gamma_{\mathrm{eff}}=\gamma_{\mathrm{eff}} \exp{(-\beta \omega_A)}$ for qubit $A$. Note that for this special choice of coupling the stationary state is left invariant by the Hamiltonian and the dissipative contribution separately. While the overall state can exhibit correlations, which for $J_x\!\not =\!J_y$ can also amount to entanglement, the local state of $S$ and $A$ are in Gibbs form. Indeed the interplay between global entanglement and local thermal states appears to be a typical quantum feature [@Kaufman2016a], and we remark there is evidence of a thermodynamic role played by (quantum) correlations [@HuberNJP; @MossyNJP; @PaternostroArXiv; @GooldArXiv]. Therefore, although $S$ always reaches a canonical Gibbs state at [*some*]{} temperature regardless of the particular parameter choice, the interaction can introduce other characteristic temperatures to the overall system. Thus, while an expression for the entropy production can be found that is meaningful, the emergence of the additional temperatures makes discussing any thermodynamic aspect more difficult. In this regard, the choice of $J_x\!=\!J_y$ and $\omega_S\!=\!\omega_A$ is special as the model retains the unique temperature defined by the bath, $\beta$. In what follows we will restrict to this setting to allow for a more meaningful and consistent thermodynamic interpretation.
Entropy Production
==================
Preliminaries {#entropy}
-------------
Let us consider a system in contact with a bath at inverse temperature $\beta$, with which it can exchange heat. The irreversible contribution to the entropy production for a given transformation is then defined as [[@Sagawa1]]{} $$\begin{aligned}
\langle \Sigma \rangle & = & \Delta S- \beta \Delta Q, \label{eq:irr}\end{aligned}$$ where $\Delta S$ is the change in entropy of the system and $\Delta Q$ denotes the mean exchanged heat, so that $\langle \Sigma \rangle$ indeed provides the contribution in entropy change which cannot be traced back to a reversible heat flow. Assuming as initial and final times of the transformation zero and $t$, respectively, and defining $Q= {\ensuremath{\operatorname{Tr}}} \rho H$, where $H$ is the system Hamiltonian, this expression can be equivalently rewritten as $$\begin{aligned}
\langle \Sigma \rangle & = & S ( \rho (0)|| \rho_{\beta} ) -S ( \rho (t)||
\rho_{\beta} ) , \label{eq:irr1}\end{aligned}$$ where $\rho_{\beta}$ denotes a Gibbs state for the system at inverse temperature $\beta$, and we have introduce the Umegaki’s quantum relative entropy [[@Nielsen2000]]{} $$\begin{aligned}
S ( \rho ||w) & = & {\ensuremath{\operatorname{Tr}}} \rho \log \rho - {\ensuremath{\operatorname{Tr}}} \rho \log w.
\nonumber\end{aligned}$$ If the dynamics of the system is given by a collection of time dependent completely positive trace preserving maps $\{\Phi (t,0)\}_{t}$, admitting $\rho_{\beta}$ as an invariant state, the irreversible entropy production as defined by Eq. (\[eq:irr1\]) is a positive quantity, in accordance to the second law. One can further consider the quantity $$\begin{aligned}
\sigma (t) & = & - \frac{{\mathrm{d}}}{{\ensuremath{\operatorname{dt}}}} S ( \rho (t)|| \rho_{\beta} ) ,
\label{eq:rate1}\end{aligned}$$ which can be naturally interpreted as the (instantaneous) entropy production rate. Consider the case in which the collection of completely positive trace preserving maps is $P$-divisible, in the sense that the following composition law holds $$\begin{aligned}
\Phi (t,0) & = & \Phi (t,s) \Phi (s,0), \hspace{2em} t \geqslant s \geqslant 0,
\label{eq:pdiv}\end{aligned}$$ with $\Phi (t,s)$ a positive map $\forall$ $t \geqslant s \geqslant 0$, and where by definition $$\begin{aligned}
\Phi (t,0) \rho (0) & = & \rho (t) . \nonumber\end{aligned}$$ Due to the fact that the relative entropy is a contraction under the action of a completely positive trace preserving map [@Lindblad1975a], and as recently shown also for a positive trace preserving map [[@Reeb2017a]]{}, in this case also the entropy production is a positive quantity.
As already observed in Ref. [[@Spohn]]{} considering the special case of quantum dynamical semigroups, the very existence of an invariant state of the dynamics, say $\bar{\rho}$, not necessarily in Gibbs form, is sufficient to introduce via $$\begin{aligned}
\langle \bar{\Sigma} \rangle & = & S ( \rho (0)|| \bar{\rho} ) -S ( \rho
(t)|| \bar{\rho} ) , \label{eq:irr2}\end{aligned}$$ a quantifier of entropy production which is always positive, and whose associated entropy production rate $\bar{\sigma} (t)$ is also positive provided the dynamics is $P$-divisible according to Eq. (\[eq:pdiv\]). For the special case of a quantum dynamical semigroup with generator, $\mathcal{G}$, the entropy production rate is given by the explicit expression $$\begin{aligned}
\bar{\sigma} (t) & = & {\ensuremath{\operatorname{Tr}}} \{ \mathcal{G} [ \rho (t)]( \log
\bar{\rho} - \log \rho (t))\} , \label{eq:rate2}\end{aligned}$$ whose positivity, following from the divisibility of the dynamics, is also known as Spohn’s inequality [@Spohn; @Alicki1979] $$\begin{aligned}
{\ensuremath{\operatorname{Tr}}} \{ \mathcal{G} [ \rho (t)]( \log \bar{\rho} - \log \rho (t))\}
& \geqslant & 0. \nonumber\end{aligned}$$ Both definitions for the entropy production rate provide convex functions of the system state, thus ensuring stability, and they are positive in the presence of a $P$-divisible dynamics. However, only $\sigma (t)$ defined in Eq. (\[eq:rate1\]) via its relation to Eq. (\[eq:irr1\]), and therefore heat transfer, can be directly connected to a thermodynamic interpretation.
Turning our attention to the dynamics, recently different notions of non-Markovianity have been introduced, related to a notion of divisibility of the quantum dynamical map or motivated by information backflow between system and environment, see [@Breuer2012a; @Rivas2014a; @BreuerRMP; @Devega2017a] for recent reviews. The definition of non-Markovianity related to divisibility was originally conceived in terms of $CP$-divisibility, that is asking that $\Phi (t,s)$ in Eq. (\[eq:pdiv\]) is a completely positive map [[@Rivas2010a]]{}. On the other hand, in order to connect the presence of memory effects with an information exchange between system and environment, it has been suggested to consider the behavior in time of the trace distance between time evolved distinct initial system states given by $$\label{tracedist}
D ( \rho_{1} ( t ) , \rho_{2} ( t ) ) = \frac{1}{2} \| \rho_{1} ( t ) - \rho_{2} ( t ) \| ,$$ where the trace norm of an operator $\| A \| = {\ensuremath{\operatorname{Tr}}} \sqrt{A^{\dag} A}$ has been introduced, reducing to the sum of the modulus of its eigenvalues for a self-adjoint operator. Revivals in time of the trace distance, Eq. (\[tracedist\]), for at least a pair of initial states is then assumed as a definition of non-Markovian dynamics [@Breuerprl]. Despite the difference in the formulation, it has been realized that there is a close connection between the notions of divisibility and information back-flow. In particular it has been shown [[@Chruscinski2011a; @Wissmann2015a; @Breuer2018a]]{} that the trace distance criterion is equivalent to $P$-divisibility, provided the map is invertible as a linear transformation. Thus, it is natural to study the behavior of the entropy production rate, and in particular its sign, and compare it with the assessment of non-Markovianity, without necessarily using it as a new definition of non-Markovian dynamics [[@Chen2017a]]{}.
Dynamical behavior
------------------
We next study the behavior of the entropy production rate in our model, which provides a simple controlled way to go from a Markovian to a non-Markovian dynamics. As stressed previously, we will restrict to the case of $J_x\!=\!J_y$ and $\omega_S=\omega_A$ where the steady state takes the form shown in Eq. with $\tilde{\beta}=\beta$. For this model an analytic treatment is feasible as shown in the Appendix, and in particular one can provide the equations describing the reduced dynamics in the form of a time-convolutionless master equation.
[**(a)**]{}\
![Trace distance for the joint $S\!+\!A$ system (upper, black) and reduced system $S$ (lower, gray) setting $\varrho_S(0)$ to be [**(a)**]{} $\{ {\left\vert0\right\rangle}{\left\langle0\right\vert},{\left\vert1\right\rangle}{\left\langle1\right\vert} \}$. [**(b)**]{} $\{ {\left\vert+\right\rangle}{\left\langle+\right\vert},{\left\vert-\right\rangle}{\left\langle-\right\vert} \}$. [*Insets:*]{} MI (upper, purple) and EoF (lower, gray). In both panels $\varrho_A(0) = {\left\vert+\right\rangle}{\left\langle+\right\vert}$ and we have fixed $\gamma=10 \Gamma = 1$, $J_x=J_y=1$, $J_z=0$, and $\omega_S=\omega_A=1$.[]{data-label="fig2"}](fig2a.pdf "fig:"){width="0.85\columnwidth"}\
[**(b)**]{}\
![Trace distance for the joint $S\!+\!A$ system (upper, black) and reduced system $S$ (lower, gray) setting $\varrho_S(0)$ to be [**(a)**]{} $\{ {\left\vert0\right\rangle}{\left\langle0\right\vert},{\left\vert1\right\rangle}{\left\langle1\right\vert} \}$. [**(b)**]{} $\{ {\left\vert+\right\rangle}{\left\langle+\right\vert},{\left\vert-\right\rangle}{\left\langle-\right\vert} \}$. [*Insets:*]{} MI (upper, purple) and EoF (lower, gray). In both panels $\varrho_A(0) = {\left\vert+\right\rangle}{\left\langle+\right\vert}$ and we have fixed $\gamma=10 \Gamma = 1$, $J_x=J_y=1$, $J_z=0$, and $\omega_S=\omega_A=1$.[]{data-label="fig2"}](fig2b.pdf "fig:"){width="0.85\columnwidth"}\
While the combined system and ancilla state obeys a semigroup dynamics described by Eq. (\[master\]), the reduced dynamics becomes non-Markovian once the interaction with the ancillary system is switched on. We see this clearly in Fig. \[fig2\] where we show the trace distance for $S\!+\!A$ (black) and the reduced system $S$ (gray) for two different pairs of orthogonal initial states. While the total state results in a monotonically decreasing behavior, since the dynamics is strictly Markovian, the interaction can lead to points of inflection/plateaus, cf. panel [**(a)**]{}. In contrast, the trace distance for $S$ clearly shows the non-Markovian nature of the dynamics. In additional we see from the insets that despite the steady state exhibiting zero EoF and MI, nevertheless dynamically significant amounts of correlations can be established. As we will see, these correlations contribute non-trivially to the entropy production.
For $S+A$ the entropy production rate $\bar{\sigma}_{{\ensuremath{\operatorname{SA}}}} (t)$ is known to be positive and given in particular by the expression $$\bar{\sigma}_{{\ensuremath{\operatorname{SA}}}} (t) = {\ensuremath{\operatorname{Tr}}} \{ \mathcal{L} [ \varrho_{SA} (t)]( \log \varrho^{\infty}_{SA} - \log \varrho_{SA} (t))\} .
\label{SAentprod}$$ Similarly, the entropy production rate for the reduced system $S$, is captured via $$\begin{aligned}
\bar{\sigma}_{S} (t)=& - \frac{{\mathrm{d}}}{{\ensuremath{\operatorname{dt}}}} S ( \varrho_{S} (t)|| {\ensuremath{\operatorname{Tr}}}_{A} \varrho^{\infty}_{SA} ) \\
=& - \frac{{\mathrm{d}}}{{\ensuremath{\operatorname{dt}}}} S ( \varrho_{S} (t)|| \rho_{\beta} ) = \sigma_{S} (t).
\label{Sentprod}
\end{aligned}$$ Notice that Eq. retains a clear thermodynamic meaning since in this case, due to the form of coupling, the system thermalizes with the bath. In fact the long-time entropy production, $\langle \Sigma \rangle$, for $S$ is a positive quantity and, furthermore, achieves the same value regardless of whether the underlying dynamics is Markovian ($\mathcal{H}_I\!=\!0$) or non-Markovian ($\mathcal{H}_I\!\neq\!0$). However, in line with the results of Ref. [@MarcantoniSciRep], we see the entropy production can become transiently negative in the non-Markovian case, as shown in Fig. \[fig3\].
![Behavior of entropy production, $\langle \Sigma \rangle$, of system $S$ for non-Markovian dynamics (solid, red) compared with the corresponding Markovian case when the $S$-$A$ coupling is switched off (dashed, magenta). Here, $\varrho_S(0) = {\left\vert1\right\rangle}{\left\langle1\right\vert}$, $\varrho_A(0) = {\left\vert+\right\rangle}{\left\langle+\right\vert}$ and we have fixed $\gamma=10 \Gamma = 1$, $J_x=J_y=1$, $J_z=0$, and $\omega_S=\omega_A=1$. In the long time limit the two entropy productions coincide.[]{data-label="fig3"}](fig3.pdf "fig:"){width="0.85\columnwidth"}\
In the present setting, thanks to the general identity $$\begin{aligned}
\label{sss}
S ( \rho_{SA} | | \tau_{S} \otimes w_{A} ) & = & \mathcal{I} ( \rho_{SA} )
+S ( \rho_{S} | | \tau_{S} ) +S ( \rho_{A} | | w_{A} ), \end{aligned}$$ relying on the structure of the equilibrium state Eq. , we find a simple relation between the various contributions to the entropy production rate for $S+A$ and the establishment of correlations $$\label{eq:sigmasum}
\bar{\sigma}_{{\ensuremath{\operatorname{SA}}}} (t) = \bar{\sigma}_{{\ensuremath{\operatorname{S}}}} (t) +\bar{\sigma}_{{\ensuremath{\operatorname{A}}}} (t) - \frac{{\mathrm{d}}}{{\ensuremath{\operatorname{dt}}}} {\cal I}(\rho_{SA}(t)).$$ Such a relation immediately puts into evidence the non-trivial role that the dynamical build-up of correlations between the constituents of the total system plays in the proper thermodynamical characterization of the process, both in open and closed system settings [@EspositoNJP; @CampisiJPA].
For zero interaction between the qubits, the entropy production rate, Eq. , is a strictly monotonically decreasing function as shown by the dashed curves in Fig. \[fig4\]. We examine the effect that the interaction term has on $\bar{\sigma}_{{\ensuremath{\operatorname{SA}}}}$ in Fig. \[fig4\] [**(a)**]{} for various initial states of $S$ when $A$ is initialized in ${\left\vert+\right\rangle}$. While in line with Markovian dynamics $\bar{\sigma}_{SA}\geq 0$ for all parameter choices, the general behavior deviates significantly from the Markovian case and we see strong oscillations occurring. In Fig. \[fig4\] [**(b)**]{} we examine the entropy production rate for the system, $S$, for the same parameters. We now see that the $\bar{\sigma}_{S}$ can dynamically become negative and the periods during which this happens are closely related to when oscillations occur in $\bar{\sigma}_{SA}$. Such a behavior is consistent with other studies assessing entropy production rates in non-Markovian settings [@MarcantoniSciRep; @PatiPRA], where the stationary state is always in Gibbs form.
[**(a)**]{}\
![[**(a)**]{} Entropy production rate for $S+A$ Eq. . [**(b)**]{} Entropy production rate for the reduced system $S$ Eq. . In both panels $\varrho_A(0)\!=\!{\left\vert+\right\rangle}{\left\langle+\right\vert}$ and we have fixed $\gamma\!=\!10 \Gamma\!=\!1$, $J_x\!=\!J_y\!=\!1$, $J_z\!=\!0$, and $\omega_S\!=\!\omega_A\!=\!1$. In both panels the dashed magenta curve corresponds to the respective entropy production rate when $\mathcal{H}_I=0$ and $\varrho_S(0)={\left\vert0\right\rangle}{\left\langle0\right\vert}$. The curve labels correspond to the initial states of $S$.[]{data-label="fig4"}](fig4a.pdf "fig:"){width="0.85\columnwidth"}\
[**(b)**]{}\
![[**(a)**]{} Entropy production rate for $S+A$ Eq. . [**(b)**]{} Entropy production rate for the reduced system $S$ Eq. . In both panels $\varrho_A(0)\!=\!{\left\vert+\right\rangle}{\left\langle+\right\vert}$ and we have fixed $\gamma\!=\!10 \Gamma\!=\!1$, $J_x\!=\!J_y\!=\!1$, $J_z\!=\!0$, and $\omega_S\!=\!\omega_A\!=\!1$. In both panels the dashed magenta curve corresponds to the respective entropy production rate when $\mathcal{H}_I=0$ and $\varrho_S(0)={\left\vert0\right\rangle}{\left\langle0\right\vert}$. The curve labels correspond to the initial states of $S$.[]{data-label="fig4"}](fig4b.pdf "fig:"){width="0.85\columnwidth"}\
Connecting this to non-Markovianity as described by an information back-flow between system and environment, as discussed in Sec. \[entropy\], this exchange in information can be traced back to a change in time of the distinguishability between distinct system states. In the present setting it is natural to consider the relative entropy as a quantifier of the distinguishability between system states. We can therefore study $S(\varrho^1_S(t)~\|~\varrho^2_S(t))$ for different choices of $\varrho^{1,2}_S(0)$. For the special case of $\varrho^{2}_S(0)=\varrho_S^{\infty}$ one describes the entropy production rate as a change of distinguishability, and therefore an indicator of non-Markovianity, thus, under this viewpoint, negative entropy production rates and non-Markovianity are directly related. However, considering the behavior of the trace distance shown in Fig. \[fig2\], the revivals, which indicate periods of non-Markovianity, are not in one-to-one correspondence with the negative entropy production rates for $S$ in Fig. \[fig4\] [**(b)**]{}. Therefore, periods of non-Markovian dynamics under one figure of merit do not directly imply a negative entropy production rate.
We can therefore conclude that a violation of $P$-divisibility alone can lead to dynamically negative entropy production rates. However this must be caveated, for $J_x\!=\!J_y$ there are initial states of systems $A$ and $S$ for which $\sigma_S\geq0$ at all times. For example, setting $\varrho_A(0)\!=\!\varrho_S(0)\!=\!\tfrac{1}{2}\openone$ leads to a positive entropy production rate for $S$ at all times despite its dynamics being non-Markovian. Thus, we have that satisfying the $P$-divisibility property is a sufficient but not necessary condition for a positive entropy production rate, since there are non $P$-divisible dynamical maps that still lead to $\sigma_S\geq0$ $\forall t$ for particular initial conditions.
Conclusions
===========
We have examined the steady state correlation properties and dynamical entropy production in a versatile setting, consisting of a single qubit, $S$, embedded in a Markovian bath interacting with a clean ancillary qubit, $A$. While the overall dynamics remains Markovian, this is no longer true for the reduced dynamics of the system due to its interaction with $A$. We have shown that the interaction can lead to strongly correlated steady states for the joint system. Examining the resulting non-equilibrium steady state we found that when such correlations are present, $S$ exhibits an effective temperature which is higher than the bath. We highlight the special choice of an excitation preserving interaction, which ensured no correlations were present in the steady state. In this setting $S$ reached thermal equilibrium with the bath and therefore allowed for a meaningful assessment of the thermodynamic features of the dynamics. By computing the entropy production rate we showed that the non-Markovianity induced by the interaction with $A$ could lead to negative entropy production rates, while the overall entropy production was still strictly positive.
Our study reveals the highly non-trivial role that the establishment of correlations plays in the thermodynamic characterization of quantum systems. At the level of the system, we have shown that the absence of quantum correlations, i.e. entanglement, is not sufficient to ensure meaningful thermodynamic quantities. Rather, [*any*]{} correlations can greatly complicate both the dynamical and steady state properties. In line with previous studies we have shown that witnessing a negative entropy production rate for the system due to a non-Markovian dynamic does not imply a violation of the second law. For instance, Ref. [@MarcantoniSciRep] insisted that one should keep track of the entropy changes from both the system and the environment, our study goes further by showing that one must also take into account the correlations established between the two, i.e. Eq. . The versatile nature of our model further reveals that, while clearly interrelated, the establishment of correlations, non-Markovian dynamics, and (negative) entropy productions rates is quite complex. In particular, as our study has revealed, non-Markovianity appears to be sufficient, but not necessary, to realize negative entropy production rates.
We are grateful to Matteo Brunelli and Sebastian Deffner for enlightening discussions and comments. BV acknowledges support from the EU Collaborative project QuProCS (grant agreement 641277) and FFABR.
Appendix {#appendix .unnumbered}
========
In the following appendix we provide the analytic solution of the model Eq. in the particular case when $J_{x}=J_{y}=\frac{J}{8}$, $J_{z}=0$, and $\omega_{S}=\omega_{A}=\omega$. The set of differential equations for the elements of density matrix $\varrho_{SA}$ are given by $$\begin{aligned}
\dot{\varrho}_{SA}^{0000}\left(t\right)&=&-\gamma\varrho_{SA}^{0000}\left(t\right)+\Gamma\varrho_{SA}^{1010}\left(t\right)\\
\dot{\varrho}_{SA}^{0001}\left(t\right)&=&i\,\frac{J}{4}\,\varrho_{SA}^{0010}\left(t\right)-\left(2i\,\omega+\gamma\right)\varrho_{SA}^{0001}\left(t\right)+\Gamma\varrho_{SA}^{1011}\left(t\right)\\
\dot{\varrho}_{SA}^{0010}\left(t\right)&=&-\left(\frac{1}{2}(\gamma+\Gamma)+2i\omega\right)\varrho_{SA}^{0010}\left(t\right)+i\,\frac{J}{4}\varrho_{SA}^{0001}\left(t\right)\\
\dot{\varrho}_{SA}^{0011}\left(t\right)&=&-\frac{1}{2}\left(\gamma+\Gamma+8i\omega\right)\varrho_{SA}^{0011}\left(t\right)\\
\dot{\varrho}_{SA}^{0101}\left(t\right)&=&-\left(\gamma+\Gamma\right)\varrho_{SA}^{0101}\left(t\right)+\Gamma\left(1-\varrho_{SA}^{0000}\left(t\right)-\varrho_{SA}^{1010}\left(t\right)\right)\\
&&-i\,\frac{J}{4}\left(\varrho_{SA}^{0110*}\left(t\right)-\varrho_{SA}^{0110}\left(t\right)\right)\\
\dot{\varrho}_{SA}^{0110}\left(t\right)&=&i\,\frac{J}{4}\left(\varrho_{SA}^{0101}\left(t\right)-\varrho_{SA}^{1010}\left(t\right)\right)-\frac{1}{2}\left(\gamma+\Gamma\right)\varrho_{SA}^{0110}\left(t\right)\\
\dot{\varrho}_{SA}^{0111}\left(t\right)&=&-\left(\frac{1}{2}\left(\gamma+\Gamma\right)+2i\omega\right)\varrho_{SA}^{0111}\left(t\right)-i\,\frac{J}{4}\varrho_{SA}^{1011}\left(t\right)\\\end{aligned}$$ $$\begin{aligned}
\dot{\varrho}_{SA}^{1010}\left(t\right)&=&\gamma\varrho_{SA}^{0000}\left(t\right)-\Gamma\varrho_{SA}^{1010}\left(t\right)+i\,\frac{J}{4}\left(\varrho_{SA}^{0110*}\left(t\right)-\varrho_{SA}^{0110}\left(t\right)\right)\\
\dot{\varrho}_{SA}^{1011}\left(t\right)&=&-i\,\frac{J}{4}\,\varrho_{SA}^{0111}\left(t\right)-\left(2i\,\omega+\Gamma\right)\varrho_{SA}^{1011}\left(t\right)+\gamma\varrho_{SA}^{0001}\left(t\right)\end{aligned}$$
For the reduced system’s dynamics the density matrix, $\varrho_{S}$, obeys the differential equations $$\begin{aligned}
\dot{\varrho}_{S}^{00}\left(t\right)&=&\Gamma-\left(\gamma+\Gamma\right)\varrho_{S}^{00}\left(t\right)-i\,\frac{J}{4}\left(\varrho_{SA}^{0110*}\left(t\right)-\varrho_{SA}^{0110}\left(t\right)\right)\\
\dot{\varrho}_{S}^{01}\left(t\right)&=&-\frac{1}{2}\left(\gamma+\Gamma+4i\omega\right)\varrho_{S}^{01}\left(t\right)+i\,\frac{J}{4}\left(\varrho_{SA}^{0001}\left(t\right)-\varrho_{SA}^{1011}\left(t\right)\right).\end{aligned}$$ For both the total $S+A$ and reduced system, the remaining density matrix elements can be readily obtained by exploiting normalization and Hermicity.
From the above equations for the reduced system we can determine the explicit form of the generator, $\mathcal{K}_{S}$, of the non-Markovian dynamics. Fixing the initial state of the ancilla to be $\varrho_{A}\left(0\right)=\frac{1}{2}\openone$ in the basis $\left\{\frac{\openone}{\sqrt{2}},\,\sigma_{-},\sigma_{+},\frac{\sigma_{z}}{\sqrt{2}}\right\}$ the generator has the form $$\begin{aligned}
\mathcal{K_{S}}\left(\varrho_{S}\right) &=-i\left[\mathcal{H}_{S},\varrho_{S}\right]+\frac{1}{D}\left[ \gamma_{1}^{S}\left(t\right) \left( \sigma_{z} \varrho_{S} \sigma_{z}-\varrho_{S} \right) \right. \\
&+ \gamma_{2}^{S}\left(t\right) \left( \sigma_{-}\varrho_{S}\sigma_{+}-\frac{1}{2}\left\{ \sigma_{+}\sigma_{-},\varrho_{S}\right\} \right)\\
&+ \gamma_{3}^{S}\left(t\right)\left(\sigma_{+}\varrho_{S}\sigma_{-}-\frac{1}{2}\left\{ \sigma_{-}\sigma_{+},\varrho_{S}\right\} \right)\\
&+ e^{-i(2\omega t + \frac{\pi}{4})}\gamma_{4}^{S}\left(t\right)\left(\sigma_{z}\varrho_{S}\sigma_{+}-\frac{1}{2}\left\{ \sigma_{+}\sigma_{z},\varrho_{S}\right\} \right)\\
&+ e^{-i(2\omega t + \frac{\pi}{4})}\gamma_{4}^{S}\left(t\right) \left(\sigma_{+}\varrho_{S}\sigma_{z}-\frac{1}{2}\left\{ \sigma_{z}\sigma_{+},\varrho_{S}\right\} \right)\\
&+ e^{i(2\omega t + \frac{\pi}{4})}\gamma_{4}^{S}\left(t\right)\left(\sigma_{z}\varrho_{S}\sigma_{-}-\frac{1}{2}\left\{ \sigma_{-}\sigma_{z},\varrho_{S}\right\} \right)\\
&+\left. e^{i(2\omega t + \frac{\pi}{4})}\gamma_{4}^{S}\left(t\right)\left(\sigma_{-}\varrho_{S}\sigma_{z}-\frac{1}{2}\left\{ \sigma_{z}\sigma_{-},\varrho_{S}\right\} \right)\right]
\end{aligned}$$ where $$\begin{aligned}
\gamma_{1}^{S}\left(t\right)&=&4\dot{\lambda_{1}}\left(\lambda_{3}-\lambda_{4}\right)-2\lambda_{1}\left( \dot{\lambda_{3}} -\dot{\lambda_{4}}\right)\\
\gamma_{2}^{S}\left(t\right)&=&-2\lambda_{1}\left(\lambda_{4}\left(\dot{\lambda_{2}}+\dot{\lambda_{3}}\right)+\left( \lambda_{2}-1 \right)\left(\dot{\lambda_{3}}-\dot{\lambda_{4}}\right)-\lambda_{3}\left( \dot{\lambda_{2}}+\dot{\lambda_{4}}\right)\right)\\
\gamma_{3}^{S}\left(t\right)&=&4\lambda_{1}\left(\dot{\lambda_{3}}-\dot{\lambda_{4}}\right)-\gamma_{2}^{S}\left(t\right)\\
\gamma_{4}^{S}\left(t\right)&=&\sqrt{2} \left(\lambda_{5}-\lambda_{2}\right)\left(\dot{\lambda_{3}}-\dot{\lambda_{4}}\right)- \sqrt{2} \lambda_{4}\left(\dot{\lambda_{2}}+\dot{\lambda_{3}}-\dot{\lambda_{5}}\right)\\
&+& \sqrt{2} \lambda_{3}\left(\dot{\lambda_{2}}+\dot{\lambda_{4}}-\dot{\lambda_{5}}\right)\end{aligned}$$ and $D=4\lambda_{1}\left(\lambda_{4}-\lambda_{3}\right)$, with $\left\{ \lambda_{i}\right\}$ being dimensionless functions defined below and $\left\{ \dot{\lambda_{i}}\right\}$ being their time derivatives. It is clear from the generator $\mathcal{K}_{S}$, which is written here in Lindblad form, that the non-Markovian nature of the reduced dynamics arises from the time dependency of the rates $\left\{ \gamma_{j}\left(t\right)\right\}$. This time dependency is exponential, as can be seen from the explicit form of the $\left\{ \lambda_{i}\right\}$ functions that appear in $\left\{ \gamma_{j}\left(t\right)\right\}$ $$\begin{aligned}
\lambda_{1}&=&\frac{e^{-\frac{1}{2}t\eta}\left(\sqrt{\Omega_{-}}\sinh\left(\frac{t\sqrt{\Omega_{+}}}{2\sqrt{2}}\right)-\sqrt{\Omega_{+}}\sinh\left(\frac{t\sqrt{\Omega_{-}}}{2\sqrt{2}}\right)\right)}{\sqrt{2 \Delta}}\\ \nonumber
&+& \frac{e^{-\frac{1}{2}t\eta}\left(\Sigma_{+}\cosh\left(\frac{t\sqrt{\Omega_{-}}}{2\sqrt{2}}\right)-\Sigma_{-}\cosh\left(\frac{t\sqrt{\Omega_{+}}}{2\sqrt{2}}\right)\right)}{2\sqrt{\Delta}}\\ \nonumber\end{aligned}$$ $$\begin{aligned}
\lambda_{2}&=&\left(\sqrt{\Delta}\eta(\Gamma-\gamma)+\gamma \eta e^{-\frac{1}{2}t\left(\eta+\sqrt{\Delta}\right)}\left(\Sigma_{+}-\Sigma_{-}e^{\sqrt{\Delta}t}\right)\right)\Delta^{-3/2}\\ \nonumber\end{aligned}$$ $$\begin{aligned}
\lambda_{3}&=&\frac{J^{2}\left(\sqrt{\Delta}(\gamma-\Gamma)+2\gamma e^{-\frac{1}{2}t\eta}\left(\eta\sinh\left(\frac{\sqrt{\Delta}t}{2}\right)-\sqrt{\Delta}\right)\right)}{\Delta^{3/2}\eta}\\ \nonumber\end{aligned}$$ $$\begin{aligned}
\lambda_{4}&=&\frac{J^{2}}{\eta}\left(\frac{1}{2\Delta}(\gamma-\Gamma)\left(e^{\frac{\sqrt{\Delta}t}{2}}-1\right)^{2}e^{-\frac{1}{2}t\left(\eta+\sqrt{\Delta}\right)}-\frac{\,\lambda_{2}}{\eta}\right)\\\end{aligned}$$ $$\begin{aligned}
\lambda_{5}&=&\frac{e^{-\frac{1}{2}t\eta}\left(2J^{2}\Gamma+\Delta(\Gamma-\gamma)e^{\frac{1}{2}t\eta}\right)}{\Delta\eta}\\
&-&\frac{e^{-\frac{1}{2}t\left(\eta+\sqrt{\Delta}\right)}\Gamma\left(\sqrt{\Delta}\left(1-e^{\sqrt{\Delta}t}\right)+\eta\left(1+e^{\sqrt{\Delta}t}\right)\right)}{\Delta}\end{aligned}$$ where $\eta =\gamma+\Gamma$, $\Delta =\eta^{2}-J^{2}$, $\Omega_{\pm} =\eta^{2}-\frac{J^{2}}{2} \pm\eta\sqrt{\Delta}$, and $\Sigma_{\pm} =\left(\eta \pm\sqrt{\Delta}\right)$. Note that if condition $\eta>J$ is satisfied, functions $\left\{ \lambda_{i}\right\}$ are real functions $\forall i$. Consequently, functions $\left\{ \gamma_{j}\left(t\right)\right\}$ are also real functions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The number of earthquakes as a function of magnitude decays as a power law. This trend is usually justified using spring-block models, where slips with the appropriate global statistics have been numerically observed. However, prominent spatial and temporal clustering features of earthquakes are not reproduced by this kind of modeling. We show that when a spring-block model is complemented with a mechanism allowing for structural relaxation, realistic earthquake patterns are obtained. The proposed model does not need to include a phenomenological velocity weakening friction law, as traditional spring-block models do, since this behavior is effectively induced by the relaxational mechanism as well. In this way, the model provides also a simple microscopic basis for the widely used phenomenological rate-and-state equations of rock friction.'
author:
- 'E. A. Jagla'
- 'A. B. Kolton'
title: The mechanisms of spatial and temporal earthquake clustering
---
[^1]
Introduction
============
The distribution of earthquakes in nature follows non-trivial patterns, some of which are captured by well known empirical laws. The Gutenberg-Richter (GR) law [@scholz; @gr] states that the number of earthquakes as a function of magnitude $N(M)$ scales as $N(M)\sim 10^{-bM}$. The exponent $b$ is very nearly 1. The Omori law refers to temporal correlations between earthquakes, in particular to aftershocks, namely the temporal clustering of earthquakes following a large one, usually called the main shock. The Omori law of aftershocks[@omori] states that the number of aftershocks per unit of time decays as $\sim (t+c)^{-p}$ with the time $t$ from the main shock. The exponent $p$ is typically very close to 1, and $c$ is a time constant of the order between minutes and hours. Aftershocks occur mainly in the spatial region where the rupture of the main shock took place.
The GR law has been shown to be compatible with a state of (at least partial) critical organization of the system[@bak; @bak2; @langer] that is understandable in terms of spring-block models, when an appropriate velocity weakening friction law (i.e., a friction force decreasing with the relative velocity of the sliding elements) is assumed to hold. This kind of modeling was pioneered by Burridge and Knopoff[@bk] (BK), and was extended along different directions afterwards, particularly in the works on self-organized-criticality of the eighties [@bak; @bak2]. The BK model reproduces the global statistical behavior implied by the GR law, but it fails to account for the existence of spatial and temporal correlations observed in actual seismicity. On searching for the origin of the aftershock phenomenon, Dieterich[@dieterich94] followed by others [@marcellini; @ziv; @moreno; @shaw] have shown that an analysis based on rate-and-state equations[@dr1; @dr2] is able to justify the appearance of aftershocks following the Omori decay. On this perspective, it is puzzling that the use of a BK model with a rate-and-state friction law does not produce realistic aftershocks[@kawamura]. Although aftershocks usually are responsible for less than about 5 $\%$ of the total released seismic moment, the finding that the GR law is also obeyed within individual aftershock sequences strongly suggests that consistent and compatible explanations for GR and Omori laws should exist. We may thus say that at present, there is not a single, unified picture of the physics behind some of the most robust features of seismicity, namely GR and Omori laws. In addition, the use of rate-and-state equations, although widely supported by experimental results, remains essentially a crude phenomenological approach.
We show here that when a spring-block model without any a priori velocity dependent frictional force is complemented with an appropriate relaxational term as discussed below, it produces: 1) earthquake patterns and in particular aftershock sequences quantitatively comparable with real ones, 2) a velocity weakening friction law, and in general, agreement with the predictions of the rate-and-state equations, and 3) a power law decay of number of earthquakes with magnitude compatible with the GR law, with an exponent $b$ that compares well with actual values.
Model in the absence of relaxation
==================================
Our modeling is based on the original BK model[@bk], with the important difference that the friction law between the blocks and the substrate is not a priori assumed to have any particular form such as the velocity-weakening form commonly used. A velocity dependent friction law emerges naturally at large scales from the characteristic collective dynamics of elastic manifolds in random media[@brazovskii]. In this context, an elastic interface (which corresponds to the blocks joined by springs in the BK model - we already describe the two dimensional case, more appropriate to real faults) is driven through a disordered potential energy landscape (the ‘substrate’) that models the random nature of asperities. The velocity-dependent frictional force between the blocks and the substrate of the BK model is therefore replaced by a disordered potential energy landscape that is chosen randomly and uncorrelated for each point block of the discretized interface. In concrete, our model is described by the overdamped equation $$\lambda \frac{\partial u_{i,j}}{\partial t}=k_0\nabla^2 u_{i,j}+f_{i,j}+k_1(X_0(t)-u_{i,j})
\label{lambda}$$ where $u_{i,j}$ is a continuous variable representing the displacement of the block labeled by the indices $(i,j)$ in a two dimensional grid, $X_0(t)$ is the driving variable (usually $X_0(t)=Vt$, we will refer also to $X_0$ and to $k_1(X_0(t)-u_{i,j})$ as the strain and local stress in the system, respectively), $\nabla^2$ is the discrete Laplacian operator, and $f_{i,j}$ is the pinning force at each block, which is assumed to be short-range correlated along the direction of the block displacements. For numerical convenience this spatially random force is chosen in the following way: a random position $u^0_{i,j}$ is selected for each $i,j$, then the force is $f_{i,j}=k(u_{i,j}-u^0_{i,j})$. When, upon the dynamical evolution, $f_{i,j}$ reaches some threshold value $f^{th}_{i,j}$ (that is chosen randomly distributed between $1+\kappa$ and $1-\kappa$), the corresponding $u^0_{i,j}$ is given the value $u_{i,j}+ \delta$, where $\delta$ is randomly chosen between $-1$ and 1. Also, a new threshold is assigned to site $i,j$. Similar results are obtained by using standard, though computationally more demanding, short-ranged correlated smooth pinning forces such as the ones used in Refs. [@lacombe; @alberto].
We use periodic boundary conditions, and from now on, we set the values $k_0=0.1$, $k=1$, $\kappa=0.8$. Taking also the numerical lattice constant as unity, this renders our time, distance, and forces, dimensionless. We have tried other parameter sets, finding no qualitatively new results. The value of $\lambda$ in Eq. (\[lambda\]) fixes the time scale necessary for the surface to adapt to the conditions dictated by $u^0$ and $X_0$. We work in the case $\lambda\rightarrow 0$, i.e., we give the surface time to relax to what we call a meta-stable configuration, defined by equating the right hand side of Eq. (\[lambda\]) to zero, for fixed values of $X_0$ and $u^0_{i,j}$. Note in particular that this also means that the duration of individual earthquakes is zero in our implementation.
Abrupt rearrangements occur in the system whenever at some particular position $i,j$ the force from the pinning potential of the substrate is not able to sustain any more the surface pinned to it (see an example of this situation in Fig. \[f0\]A). In this situation the local rearrangement of the surface can trigger instability events in neighbor sites, and the process continues until the surface finds a new globally stable configuration. The full sequence of all rearrangements triggered by an initial instability is what we call an event, or an earthquake, being the position of the triggering instability its epicenter. We measure the seismic moment $m_0$ of events as $m_0 = \sum_{i,j} \Delta_{i,j}$ where $\Delta_{i,j}$ is the displacement caused by the event at position $i,j$. In order to compare with real earthquakes, the magnitude $M$ of an event is defined from the seismic moment as $M = 2/3\log_{10}m_0$.
Results
-------
![One dimensional sketch of main processes that occur in our model. (A) In the absence of relaxation, the interface (vertical line, with coordinates $u_{i,j}$) is driven to the right by an external force (not shown) through a set of randomly placed pinning centers, represented by the gray rectangles. The mid point of the rectangles have coordinates $u^0_{i,j}$, and the horizontal length is the range in which pinning is effective. Different lengths indicate different threshold values $f^{th}_{i,j}$. Upon driving, the system passes from the configuration indicated by the continuous line to the dashed line, and an event is triggered at the site indicated by the arrow, where the maximum local pinning force is overpassed. The system goes through a cascade process (not fully indicated) onto a new meta-stable configuration (dotted line) in which some pinning centers have been refreshed (outlined rectangles). In (B) structural relaxation is acting. A quite relaxed (and therefore more coherently pinned) initial configuration (continuous line) is driven until a main shock occurs, at the configuration corresponding to the dashed line. After the system has reached a meta-stable configuration (dotted line and outlined rectangles) relaxation continuous to act modifying the position of the pinning centers. The arrows at the center of the rectangles indicate the local value of $(u-u^0)$. The arrows just outside the rectangles indicate the values of $du^0/dt$ according to Eq. (\[eq2\]), that produce a drift in the position of the pinning centers. This drift may cause a further instability as can be seen in (C). The process in (C) is an aftershock to the main shock in (B). []{data-label="f0"}](pnasf1.eps){width=".5\textwidth"}
![Results without relaxation. ([**A**]{}) Magnitude histogram for systems of 512x512 sites, with $k_1=0.01$ (full symbols) and $k_1=0.001$ (open symbols). Continuous line has a slope $b=0.4$. ([**B**]{}) Magnitude-time plot for a system of size 256x256, $k_1=0.01$. []{data-label="S1"}](pnasf2.eps){width=".4\textwidth"}
Our model in the absence of relaxation has been widely studied, and is known to have a well defined size distribution of events[@lacombe; @alberto; @zapperi; @pierre] (note that the usual definition of the decay exponent $\tau$ is given as a function of our $b$ as $\tau=2b/3+1$). In Fig. \[S1\]A we show this distribution. We observe a power law with exponent $b\simeq 0.4$, which is consistent with the expected results for two dimensional elastic interfaces both from scaling arguments[@zapperi] and from recent analytical and numerical calculations[@pierre]. This value is however well different from the $b\simeq 1$ observed for actual earthquakes. An exponential cut-off for large event size exists due to confinement. This cut-off is controlled by the rigidity $k_1$ of the driving spring and occurs when the spatial extent of the events in the direction of the displacements is of order $k_1^{-\zeta/2}$, with $\zeta$ the interface roughness exponent at low velocities. The crossover to the exponential behavior thus moves to larger magnitudes as $k_1$ is decreased [@lacombe; @alberto; @zapperi; @pierre].
In Fig. \[S1\]B we show a time-magnitude plot of all events occurring in a particular time interval. It is apparent that no obvious temporal correlations occur in this case, and more quantitative observations confirm this fact. Spatial correlations are not observed neither. This is qualitatively similar to what has been obtained for the original BK model[@bk].
Model in the presence of relaxation
===================================
![Results for the model with relaxation as compared to earthquakes in California[@California]. ([**A**]{})Magnitude-time plot for a 512x512 system in the presence of relaxation ($k_1=0.01$, $R/V=500$). ([**B**]{}) Actual earthquakes in the California area. []{data-label="512r500"}](pnasf3.eps){width=".4\textwidth"}
So far, in the present form, the model does not give any clue on the reason for earthquake clustering, or the origin of a velocity weakening friction law. However, the inclusion of a simple additional ingredient changes this scenario drastically. This ingredient turns out to be what we have called [*structural relaxation*]{} [@jagla]. The primary physical justification of its inclusion is the following. It is known that in solid friction the friction coefficient at rest increases with the time the surfaces have been in contact [@marone; @persson]. This is telling us about the existence of a temporal dependent mechanism that makes the sliding surfaces get more attached or pinned to each other when they remain in contact for a longer period of time. In the present model, a rather simple way to include such an effect is to consider, in addition to the random pinning force that the substrate performs on the sliding surface, the reaction that the surface performs onto the substrate. If we give the substrate the possibility to react to this force, the system will gain pinning energy by making the substrate more correlated, so to pin better the correlated interface structure. This is in general a slow process, and the longer the surface remains in contact with the substrate, the stronger the join. This process of attachment is however stopped and restarted when a slip event occurs, since the values of the disordered potential refreshen, becoming uncorrelated again. We will show in the following that this simple mechanism is enough to explain, in particular, the appearance of a robust sequence of aftershocks, and the occurrence of a velocity weakening friction law. Our structural relaxation mechanism is what in other contexts is called the ageing of the material.
![Spatial distribution of aftershocks in the simulations. The slip surface of a large event (shadowing proportional to the local slip) and the before- (full) and after-events (open symbols) of magnitude larger than 0 occurring in a symmetric time interval $\delta t=0.0022$ around the main event are shown. The increase in seismicity after the main shock is clearly observable, as well as the localization of aftershocks mainly at and near the slip surface of the main shock. The figure depicts a portion of size 350x350 of a system of 512x512. []{data-label="locate"}](pnasf4.eps){width=".4\textwidth"}
The modification to the model is as follows. We allow the values of $u^0_{i,j}$ to relax in time according to $$\frac{\partial u^0_{i,j}}{\partial t}=-R \nabla^2 f_{i,j}= Rk \nabla^2 (u^0_{i,j}-u_{i,j}).
\label{eq2}$$ This conserved dynamics for the shift of the disorder potential at different block points is a generic way to introduce the back effect of the surface on the substrate (the actualization of the $u^0$’s when a slip event occurs is made as before). The coefficient $R$ is a measure of the intensity or rate of relaxation, and can thus be related to experimental relaxation times. Equation (\[eq2\]) generates a tendency for the local forces $f_{i,j}$ to become uniform across the system, generating a stronger contact between surface and substrate. This relaxational effect competes with the driving, which forces the movement of the surface onto the substrate at a fixed average velocity $V$. The relevant parameter that measures the competition between the two effects is the ratio $R/V$.
The mechanism by which earthquake clustering occurs can be summarized as follows (see Fig. \[f0\]). If a particular region of the sample has not experienced a large event in a rather long period of time, the structural relaxation has made this region stronger (Fig. \[f0\](B)). When an event occurs (driven by the overall displacement between surface and substrate) the contacts refreshen and large variations in the local forces remain. Relaxation continues to act, trying to uniformize the local forces. In this process, particular points that were originally stable immediately after the main shock, may destabilize and generate a new event (Fig. \[f0\](C)). Note that in this description it is obvious that aftershocks will occur at, or near, the rupture region of the main shock. It is also worth noting here that aftershocks are triggered by the inner dynamics of the system, and that most aftershocks occur also if we stop the driving of the system after the main shock. The seemingly contradictory fact that aftershocks (which must be triggered by an initial instability) are originated in a relaxation mechanism is understood when one realizes that due to the disorder, relaxation according to Eq. (\[eq2\]) may produce local increases in the forces $f_{i,j}$, and this can trigger aftershocks if a local threshold is overcome. Note in this respect the opposite direction between $(u-u^0)$ and $du^0/dt$ at the aftershock epicenter, in Fig. \[f0\]C.
Results and comparison with an actual earthquake sequence
---------------------------------------------------------
![The Omori law. Histogram of the number of events after main shocks in the simulation (open symbols), averaged over 120 main shocks, and the histogram of aftershocks for events in the California area (full symbols), averaged over 7 events of magnitude $M > 6.0$ in the time period considered. $\Delta t$ is the time since main shock. Curves have been vertically rescaled, setting the value 1 for large $\Delta t$. Continuous lines are fittings to Omori law with $p=1$. The time shifts are $c_1=3\times 10^{-4}$, $c_2=0.05$ days. []{data-label="S2"}](pnasf5.eps){width=".4\textwidth"}
\[t\]
![Cumulative number of events and cumulative seismic moment for the sequences presented in Fig. \[512r500\]. []{data-label="S3"}](pnasf6.eps){width=".4\textwidth"}
In Fig. \[512r500\]A we show a magnitude-time plot of events in a simulation of a system of 512x512 sites, in the presence of relaxation ($R/V=500$). For comparison, the same plot for the earthquakes in the California area [@California] is presented as Fig. \[512r500\]B. The visual similarity is striking. In both graphs, the very large activity immediately after large events, i.e., the existence of aftershocks, is apparent. It is worth noting here that a minimum value around $R/V \sim 100$ of relaxation is necessary in order to observe aftershocks in the model, so the structural relaxation is the crucial ingredient behind these particular effects.
The spatial location of aftershocks are strongly correlated with the slip surface of the main event. In Fig. \[locate\] we show the region that has slip in a large event in the simulations, together with the epicenters of all events occurring in a symmetric time interval around the main shock. Events before and after the main shock are shown in a separate way. We see that there is a rather uniform spatial distribution of before-events, whereas once the main shock has occurred, aftershocks occur at and near the region in which the main slip occurred.
In Fig. \[S2\] we plot the histogram with the number of events after a main shock as a function of time. In the simulations, we average over 120 large events in a single simulation of size 512x512. The continuous lines correspond to Omori laws of the form $N(t)=A/(t-t_0+c)^{p} +N_0$, where $t_0$ is the time of the main shock and $N_0$ is the value of background seismicity. For reference, we also plot the number of aftershocks of the seven earthquake with $M>6.0$ in the California area in the considered time period. Both cases are well fitted by an Omori law with $p\sim 1$.
Figure \[S3\] shows plots of cumulated number of events and seismic moment corresponding to the sequences presented in Fig.\[512r500\], and Fig. \[S4\] is a detail after the events indicated by vertical arrows in Fig. \[S3\]. The cumulative number of events is fitted in both cases with a cumulative Omori law, with $p=1$ with very good agreement. The evolution of the accumulated seismic moment is also qualitatively similar in both cases, with the main shock accounting for most of the released seismic moment of the whole sequence.
Finally, in Fig. \[S5\] we present an analysis of the time intervals $\Delta t$ between successive events of magnitude larger than some defined threshold $M_0$. The main characteristics observed for the real sequence, that are reproduced by our model are the following. The curves are roughly independent of the threshold value $M_0$ chosen, and the global behavior represents almost an exponential decay with $\Delta t$, but with a reproducible excess of events at low $\Delta t$. This excess is accounted for by the aftershocks.
Averaged frictional properties
==============================
![Detail to Fig. \[S3\]. Cumulative number of events and cumulative seismic moment (taken as zero just before the main shock), following the events indicated by vertical arrows in Fig. \[S3\]. In both cases, dotted lines are fitting to the (cumulative) Omori law with $p=1$. []{data-label="S4"}](pnasf7.eps){width=".4\textwidth"}
![Time intervals distribution. Distribution of the times (normalized by the mean time $\Delta t_m$) between events of magnitude larger than $M_0$ for our data ([**A**]{}) and for earthquakes in California ([**B**]{}). Independence of $M_0$, and a rather exponential distribution with an excess due to aftershocks for small $\Delta t$ is observed in both cases. []{data-label="S5"}](pnasf8.eps){width=".4\textwidth"}
The second set of results that we present corresponds to the stress-strain relation of the model for different driving protocols. First of all we recall that the model without relaxation shows a stress $\sigma$ that is independent of the strain rate, since the internal time scale of the model is very rapid compared to the driving. In Eq. (\[lambda\]) this means that $\lambda/V\to 0$. The inclusion of relaxation introduces a new time scale (set by the parameter $R$ in Eq. (\[eq2\])) and now the average stress in the system depends on the ratio $R/V$. When $R/V$ is small, the effect of relaxation is negligible, and the stress will be similar to that in the absence of relaxation. However, if $R/V$ is high enough, relaxation will act by effectively correlating the pinning potential in a larger spatial region. The size of this region increases with $R/V$. A spatially more correlated pinning potential produces, in turn, a larger average stress in the system. We conclude that the larger is $R/V$ the larger is the average stress. In other words, the model will display velocity weakening. In Fig. \[ss\]A we show a plot of the stress in the system as a function of strain rate where this weakening is clearly observed. For large strain rates the stress converges to the value corresponding to no relaxation, whereas the behavior for very small strain shows a saturation at a larger value. The transition between these two values is logarithmic and spans about a factor of 100 of strain rate. Note that the values reported above as necessary to observe aftershocks ($R/V \gtrsim 100$) correspond to the limit of small velocity in this plot. A closer examination at the instantaneous stress-strain relation reveals that the lower the strain rate, the more pronounced the fluctuations in the instantaneous stress.
Additional information on the frictional behavior is obtained by studying the system response to abrupt changes of the strain rate. We show in Fig. \[ss\]B in particular, the stress on a system in which driving is stopped during some time interval (the hold time) and then is re-initiated. First of all, a logarithmic decrease of stress during the hold time is observed. This occurs because the system continues to relax during the hold time and some instability events continue to occur for some time. This is related to our previous statement that aftershocks also occur if driving is stopped after a main shock. Despite the stress reduction during the hold time, a stress peak occurs after re-initiation of sliding. The height of this peak increases logarithmically with the hold time. This peak is a consequence of the more stable configuration that the system reached due to relaxation during the hold time. The phenomenon is similar to the one observed in glass forming materials, where it has been explained using the same ideas [@jagla]. These results are in remarkable agreement with those obtained in laboratory measurements [@marone; @marone2].
Gutenberg-Richter behavior
==========================
![Global frictional properties of the model. ([**A**]{}) Mean stress in a system of 256x256 as a function of relative velocity. A detail of the temporal dependence of stress is given for two points, emphasizing the larger fluctuations that appear when relative velocity is lower. ([**B**]{}) Time evolution of stress in a system in which velocity is changed from $V/R=0$ in the hold periods (indicated by arrows), to $V/R=1$ in the rest of time (results shown correspond to an average over ten realizations). Inset: The value of the stress peak as a function of the hold time. []{data-label="ss"}](pnasf9.eps){width=".4\textwidth"}
![The decaying of number of events with event magnitude. The $R=0$ case is included again for reference. The thin continuous line shows the case $b=1$, followed rather accurately by actual seismic events. The results in our model in the presence of relaxation show a behavior compatible with the actual seismicity, with a cut-off at large event size that increases upon decreasing the spring constant $k_1$. However, finite size effects seem to be appreciably for the system sizes used. The larger decaying rate observed for $M < 1$ is an spurious effect associated with events comparable in size with our numerical mesh. []{data-label="S6"}](pnasf10.eps){width=".4\textwidth"}
The inclusion of relaxation produces also a change in the decaying exponent $b$ of the GR law. In Fig. \[S6\] we see that a power law decaying is maintained in the presence of relaxation, with a $b$ value substantially larger than that corresponding to no relaxation. Once a minimum value of relaxation has been over passed ($R/V \sim 20$), the $b$ decaying exponent is quite insensitive to the precise value of relaxation. There seems to be an excess of events of large magnitude, before the cut off is reached. The cut off and the peak corresponding to large events are mainly dependent on the value $k_1$ of the spring driving the system. The smaller this value, the larger is the cut-off. It is not clear however if this tendency can be extrapolated to very small values of $k_1$. Unfortunately, to simulate decreasing values of this spring constant requires an increase in system size, and we reach rapidly very time consuming runs. The obtained decaying exponent in the presence of relaxation is compatible with the value $b\sim 1$ observed in actual seismicity.
The fact that the $b$ exponent takes a value close to 1 in the presence of relaxation, quite independent of the precise value of the relaxation parameter and other details of the model seems to indicate that relaxation takes the system out of its original universality class with $b\simeq 0.4$, to a new one with $b\simeq 1.0$. Coincidence of this value with actual ones is another indication that we are capturing essential features of the seismic process with the inclusion of the relaxation mechanism.
Conclusions
===========
Summarizing, in the present paper we have presented a modeling that combines a spring-block type system in the spirit of the BK model but without a priori velocity weakening friction, with a rather generic implementation of ageing effects within the sliding materials. The motivation for this approach was to introduce, in a spring-block model, a mechanism that generates (and not merely assumes) non trivial frictional effects, which can produce realistic temporal and spatial clustering of earthquakes.
Our model allows to obtain a time sequence of events that globally follow the GR law with a $b\simeq 1$ exponent, and at the same time highly non-trivial spatial and temporal correlations compatible, in particular, with the Omori law. In addition we have shown that frictional properties of the model compare very well with laboratory results. We think this model gives a unified and comprehensive physical picture of all these phenomena.
Acknowledgments
===============
This research was financially supported by Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Argentina. Partial support from grants PIP/5596 (CONICET) and PICT 32859/2005 (ANPCyT, Argentina) is also acknowledged.
C. H. Scholz, [*The Mechanics of Earthquakes and Faulting*]{}, (Cambridge University Press, Cambridge, England, 2002).
Gutenberg B, Richter C F (1956) Magnitude and energy of earthquakes. [*Ann Geophys*]{} (C.N.R.S) 9:1.
Bak P, Tang C, Wiesenfeld K (1987) Self-organized criticality: An explanation of the 1/f noise. [*Phys Rev Lett*]{} 59:381-384.
Bak P, Tang C (1989) Earthquakes as a Self-Organized Critical Phenomenon. [*J Geophys Res*]{} 94:15635-15637.
Carlson J M, Langer J S. , Shaw B-E (1994) Dynamics of earthquake faults. [*Rev Mod Phys*]{} 66:657-670.
Burridge R, Knopoff L (1967) Model and theoretical seismicity. [*Bull Seismol Soc Am*]{} 57:341-362.
Omori F (1894) On the aftershocks of earthquakes. [J Coll Sci Imp Univ Tokio]{} 7:111-200.
Dieterich J-H (1994) A constitutive law for rate of earthquake production and its application to earthquake clustering. [J Geophys Res]{} 99:2601-2618.
Marcellini A (1997) Physical model of aftershock temporal behavior. [*Tectonophysics*]{} 277:137-146.
Ziv A, Rubin A M (2003) Earthquakes and Acoustic Emission. [*J Geophys Res B: Solid Earth*]{} 108(B1):2051.
Moreno Y, Correig A M, Gomez J B, Pacheco A F (2001) A model for complex aftershock sequences. [*J Geophys Res B: Solid Earth*]{} 106:6609-6619.
Helmstetter A., Shaw B. E. (2006) Relation between stress heterogeneity and aftershock rate in the rate-and-state model. [*J Geophys Res*]{} 111:B07304.
Dieterich J H (1979) Modeling of rock friction .1. Experimental Results and Constitutive Equations. [*J Geophys Res*]{} 84:2161-2168.
Ruina A (1983) Slip Instability and State Variable Friction Laws. [J Geophys Res]{} 88:10359-10370.
Ohmura A, Kawamura H (2007) Rate- and state-dependent friction law and statistical properties of earthquakes. \*\*[*Europhys Lett*]{} 77:690011-5.
Fisher D S (1998) Collective transport in random media: from superconductors to earthquakes. [*Physics Reports*]{} 301:113-150. Lacombe F, Zapperi S, Hermann H J (2001) Force fluctuation in a driven elastic chain.[*Phys Rev B*]{} 63:1041041-7.
Rosso A, Le Doussal P, Wiese K J (2007) Numerical calculation of the functional renormalization group fixed-point functions at the depinning transition. [*Phys Rev B*]{} 75:220201(R).
Zapperi S, Cizeau P, Durin G, Stanley H E (1998) Dynamics of a ferromagnetic domain wall: Avalanches, depinning transition, and the Barkhausen effect. [*Phys Rev B*]{} 58:6353-6366.
Le Doussal P, Middleton A, Wiese K J (2008) Statistics of static avalanches in a random pinning landscape. [*arXiv*]{}:0803.1142v1.
Jagla E A (2007) Strain localization driven by structural relaxation in sheared amorphous solids. [*Phys Rev E*]{} 76:0461191-7.
Marone C (1998) The effect of loading rate on static friction and the rate of fault healing during the earthquake cycle. [*Nature*]{} 391:69-72.
B. N. J. Persson, [*Sliding Friction, Physical Principles and Applications*]{}, (Springer, Berlin, 2000).
By this, we mean all earthquakes in the region between 115 and 119 W, and between 32 and 35 N, of $M > 2$ between January 1st (1980), and May 20th (2008), as reported at http://quake.geo.berkeley.edu/anss.
Marone C (1998) Laboratory-Derived Friction Laws and their Application to Seismic Faulting. [*Annu Rev Earth Planet Sci*]{} 26:643-696.
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We summarize the existing experimental data on electromagnons in multiferroic $R$Mn$_2$O$_5$ compounds, where $R$ denotes a rare earth ion, Y or Bi, and discuss a realistic microscopic model of these materials based on assumption that the microscopic mechanism of magnetically-induced ferroelectricity and electromagnon absorption relies entirely on the isotropic Heisenberg exchange and magnetostrictive coupling of spins to a polar lattice mode and does not involve relativistic effects. This model explains many magnetic and optical properties of $R$Mn$_2$O$_5$ manganites, such as the spin re-orientation transition, magnetically-induced polarisation, appearance of the electromagnon peak in the non-collinear spin state and the polarisation of light for which this peak is observed. We compare experimental and theoretical results on electromagnons in $R$Mn$_2$O$_5$ and $R$MnO$_3$ compounds.'
address:
- '$^{1}$ Materials Research Science and Engineering Center, University of Maryland, College Park, Maryland 20742'
- '$^{2}$ Zernike Institute for Advanced Materials, University of Groningen, 9747 AG Groningen, The Netherlands'
- '$^{3}$ Rutgers Center for Emergent materials and Department of Physics and Astronomy, Rutgers University, Piscataway, New Jersey 08854'
author:
- 'A B Sushkov$^{1}$, M Mostovoy$^{2}$, R Valdés Aguilar$^{1}$, S-W Cheong$^{3}$ and H D Drew$^{1}$'
---
Introduction {#sec:introduction}
============
Multiferroic materials that exhibit simultaneous magnetic and ferroelectric order have attracted much attention recently because of the fundamental interest of systems with coupled order parameters and because of their potential for cross electric and magnetic functionality [@Smol-Chupis; @Fiebig2005; @Prellier2005; @Khomskii2006; @Eerenstein2006; @Cheong2007; @Ramesh2007]. Many recently discovered multiferroics, e.g. TbMnO$_3$ [@Kimura-113], TbMn$_2$O$_5$ [@Hur-nature], and Ni$_3$V$_2$O$_8$ [@Lawes-vanadate], are improper ferroelectrics in which the electric polarisation is induced by spin ordering. These materials show striking cross-coupling effects such as magnetic field induced polarisation switching [@Kimura-113; @Hur-nature] and giant magnetocapacitance [@Goto2004] and a strong coupling between spin and lattice excitations that leads to electric dipole excitation of magnons, or electromagnons [@Smol-Chupis; @Pimenov-Nature; @Sushkov125; @Katsura-em], which is the subject of this paper. The fundamental interest in multiferrocity also derives from the strong interplay between magnetic frustration, ferroelectric order and unusual symmetry breaking in phase transformations that characterize these materials [@Cheong2007; @Kimura07].
The known microscopic mechanisms of magnetically-induced ferroelectricity include lattice distortion (exchange striction) and redistribution of electron density in response to spin ordering. Such processes occur locally in all magnetic materials. However, only when a spin ordering breaks inversion symmetry do these local electric dipoles add into a macroscopic electric polarisation. Spin orders that break inversion symmetry are rare and the best systems to look for them are frustrated magnets, where competing interactions and the geometry of spin lattice favor unconventional magnetic states. In most of the recently discovered multiferroic materials, such as TbMnO$_3$, Ni$_3$V$_2$O$_8$, MnWO$_4$ and CuO, competing interactions force spins to form a cycloidal spiral. This non-collinear spin order breaks inversion symmetry and activates antisymmetric Dzyaloshinskii-Moriya interaction proportional to the cross-product of spins, $\bi{S}_{1} \times \bi{S}_{2}$ [@Moriya]. The concomitant lattice and electronic distortion induces electric polarisation [@Katsura2005; @SergienkoDagotto; @Jia].
When inversion symmetry is broken by a collinear magnetic ordering the strongest spin interaction that can shift ions and polarise electronic clouds is the symmetric Heisenberg exchange, proportional to the scalar product of spins, $\bi{S}_{1} \cdot \bi{S}_{2}$. This mechanism was proposed to explain multiferroic properties of $R$Mn$_2$O$_5$, where $R$ denotes a rare earth ion, Y or Bi, and orthorombic manganites showing the E-type antiferromagnetic ordering [@Kadomtseva; @ChaponPRL2006; @SergienkoPRL2006].
These two microscopic mechanisms of magnetically-induced ferroelectricity give rise to two different forms of phenomenological magnetoelectric coupling: electric polarisation induced by a low-pitch spiral is described by the third-order coupling term $P \left(L_{1} \partial L_{2} - L_{2} \partial L_{1}\right)$, where $P$ is electric polarisation and $L_{1,2}$ are magnetic order parameters describing the sinusoidal and cosinusoidal components of the spiral [@Baryachtar; @Mostovoy-spiral], while the coupling working in collinear spin states has the form $P\left(L_{1}^2-L_{2}^2\right)$, where $L_{1}$ and $L_{2}$ are components of a two-dimensional irreducible representation describing magnetic states with opposite electric polarisations [@Levanyuk; @Kadomtseva; @SergienkoPRL2006].
![(Colour online) Feynman diagrams describing photo-excitation of magnons by the electric field of light (here photon and phonon are represented, respectively, by dashed and solid lines, while the wavy line corresponds to magnon): (a) Photoexcitation of two magnons and one phonon via the fourth-order magnetoelectric coupling (the Lorenzana-Sawatzky mechanism), (b) Excitation of the two-magnon continuum via the third-order magnetoelectric coupling (‘charged magnons’) and (c) Photoexcitation of a single magnon (electromagnon) due to the third-order magnetoelectric coupling, where photon with the frequency $\omega$ and zero wave vector scatters off the static spin modulation with the wave vector ${\bf Q}$ producing a magnon with the same wave vector and the frequency ${\omega}$.[]{data-label="fig:photoex"}](figure1.eps){width="8cm"}
The magnetoelectric interactions that induce electric polarisation in magnets can also couple oscillations of magnetisation to polar lattice vibrations. The oscillations of polarisation at magnon frequency and vice versa give rise to dynamic magnetoelectric effects, such as electromagnon excitations. In usual magnets, an oscillating electric field of photons can excite a three-particle continuum consisting of two magnons and one phonon [@Lorenzana1995]. This process results from the fourth-order spin-lattice coupling (see figure \[fig:photoex\]a). The third-order couplings in multiferroics, discussed above, make possible photo-excitation of two-magnon continuum without a phonon (‘charged magnons’ [@Damascelli]) shown in figure \[fig:photoex\]b). Replacing one of the magnons by the static modulation of spin density appearing in the ordered spin state, we obtain a process that converts photon into a single magnon, which is the electromagnon (see figure \[fig:photoex\]c). This process is usually mediated by a polar phonon linearly coupled to both magnons and light, which for low-frequency phonons can lead to a resonant enhancement of the photo-excitation of electromagnons.
As there are two possible contributions to the polarisation — from ionic displacements and from electronic density redistribution — one can think of two corresponding electric contributions to electromagnon — from phonons and from electronic excitations. This means a transfer of the electric dipole spectral weight from phonons ($\hbar\omega \sim 10 - 100$ meV) and/or electronic excitations ($\hbar\omega \sim 2$ eV) down to magnons ($\hbar\omega \sim 1 - 5$ meV). Such a transfer, in turn, leads to a step-like anomaly in the temperature dependence of the dielectric constant. Current research on electromagnons, including this paper, is focused mainly on magnon-phonon coupling while magnon-electron aspect is much less explored as it is generally expected to be weak because of the large energy difference between the electronic excitations and the magnetic excitations.
The possibility of electromagnon excitations in multiferroics has long been anticipated theoretically [@Smol-Chupis], but only recently were they observed in experiment. Pimenov [@Pimenov-Nature] reported observation of electric active modes at magnon frequencies in GdMnO$_3$ and TbMnO$_3$ which exist only in some magnetically ordered phases and can be suppressed by magnetic field. Further, they found in GdMnO$_3$ the spectral weight transfer from the lowest frequency phonon down to the electromagnon mode [@Pimenov-PRB]. Sushkov [@Sushkov125] reported observation of electromagnons in YMn$_2$O$_5$ and TbMn$_2$O$_5$. Electromagnons in both these $R$Mn$_2$O$_5$ compounds have very similar spectrum and exist only in magnetic ferroelectric phases that proves their magnetic origin. Valdés Aguilar [@Valdes-EuY] showed that electromagnon absorption in Eu$_{0.75}$Y$_{0.25}$MnO$_3$ occurs over a broad band with two peaks, both of which exist only in the magnetic ferroelectric phase below 30 K. Pimenov [@Pimenov-EuY] explored a composition set Eu$_{1-x}$Y$_{x}$MnO$_3$ for $0 \leq x \leq 0.5$ and confirmed main features of electromagnons in $R$MnO$_3$. Kida [@Kida2008] found out that the electromagnon polarisation in DyMnO$_3$ stays with the lattice when the spin plane is rotated by the external magnetic field. Electromagnons, observed so far only in non-collinear spin phases of $R$MnO$_3$ and $R$Mn$_2$O$_5$ compounds, have common features for both families: they are active only in one polarisation $\bi{e} \| \bi{a}$ axis for $R$MnO$_3$ and $\bi{e} \| \bi{b}$ axis for $R$Mn$_2$O$_5$, where $\bi{e}$ is electric field of light; well defined peaks exist only in the low-temperature magnetic ferroelectric phase. A theory of electromagnons for the circular magnetic spiral was developed by Katsura [@Katsura-em] and its continuum version was given by de Sousa [@deSoussaPRB2008; @deSousa] who applied it to describe low-energy excitations in BiFeO$_3$. Fang and Hu [@FangCM] discussed electromagnon absorption in $R$Mn$_2$O$_5$ compounds assuming Heisenberg interactions between spins.
The outstanding fundamental questions for electromagnon are the microscopic origin of these novel excitations (Heisenberg or Dzyaloshinskii-Moriya type exchange) in the different classes of compounds ($R$MnO$_3$ and $R$Mn$_2$O$_5$), the explanation of the observed selection rules, and whether electromagnons may occur in a wider range of materials. The practical issues are enhancing the magneto-capacitance effect and its temperature range and possibly applying these new excitations to metamaterials and/or achieving negative index of refraction in the magnon range of frequencies.
The remainder of this paper is organized as follows. In section \[sec:exp125\], we discuss optical absorption spectra of multiferroic YMn$_2$O$_5$ and TbMn$_2$O$_5$ compounds taken at different temperatures. We present experimental evidence allowing us to identify the low-frequency peaks appearing in the non-collinear phase of these materials as electromagnons. For comparison, in section \[sec:exp113\] we present data on electromagnons peaks in the spiral multiferroic material Eu$_{0.75}$Y$_{0.25}$MnO$_3$. In section \[sec:model\], we discuss a simple microscopic model of YMn$_2$O$_5$ manganites, based on isotropic Heisenberg exchange interactions between spins, with which we describe magnetic, multiferroic and optical properties of these compounds. We also briefly discuss phenomenological description of magnetic orders and magnetoelectric coupling in $R$Mn$_2$O$_5$ compounds based on their symmetry. In section \[sec:discussion\], we discuss the microscopic origin of the electromagnon peaks in both $R$Mn$_2$O$_5$ and $R$MnO$_3$ and how the electromagnon relates to the spontaneous polarisation. Finally, we summarize our experimental and theoretical results.
Far infrared spectroscopy of $R$Mn$_2$O$_5$ compounds {#sec:exp125}
=====================================================
![(Colour online) Optical conductivity of YMn$_2$O$_5$ and TbMn$_2$O$_5$ for the electric field of light $\bi{e} \| \bi{b}$ in three phases. Strong peaks at 113 and 97 cm$^{-1}$ are the lowest phonons, other peaks are electromagnons.[]{data-label="s1"}](figure2.eps){width="8cm"}
The family of $R$Mn$_2$O$_5$ compounds has long been studied [@Bertaut1967]. Small variations of the exchange integrals with temperature lead to the complex phase diagram for this multiferroic family [@Harris2008]. Earlier spectroscopic works revealed far infrared absorption modes activated at low temperatures in EuMn$_2$O$_5$ [@Sanina1988], YMn$_2$O$_5$ [@Mukhin2000] and GdMn$_2$O$_5$ [@Golovenchits2003]. Recently, we have shown that in YMn$_2$O$_5$ and TbMn$_2$O$_5$ strong electric dipole active modes emerge at magnon frequencies in the lowest in temperature ferroelectric phase [@Sushkov125]. As in the work by Pimenov [@Pimenov-Nature] on $R$MnO$_3$ compounds, we assign this modes as electromagnons.
In figure \[s1\], we show the optical conductivity spectra of YMn$_2$O$_5$ and TbMn$_2$O$_5$ at three different temperatures corresponding to three magnetic/ferroelectric phases. These spectra were obtained by fitting measured transmission spectra [@Sushkov125] with the model dielectric function [@reffit], assuming all modes to be of the electric dipole nature. The spectra measured at 7 K (blue curves) show the characteristic strong and sharp (electromagnon) absorption peaks at lowest frequencies. In this phase, the spontaneous electric polarisation is relatively small and the angles between neighbouring spins along the $b$ axis are large. The spectra measured at $25$ K (green curves) have one broad absorption peak near 20 cm$^{-1}$. In this phase spins are almost collinear and electric polarisation is large. Red curves, taken just above the Néel temperature, show a single broad absorption band below the phonon frequencies.
Identifying the low-frequency excitations as electromagnons requires addressing several questions. First, to avoid confusion with possible transitions between $f$-levels of rare earth ions, we have studied YMn$_2$O$_5$. The second issue is electric versus magnetic dipole (antiferromagnetic resonance) activity. We measured transmission spectra for various mutual orientations of the electric and magnetic field of light (respectively, $\bi{e}$ and $\bi{h}$) with respect to the crystal axes and we found the absorption only for $\bi{e} \| \bi{b} \| \bi{P}$, where $\bi{P}$ is the spontaneous polarisation vector, independent of the orientation of $\bi{h}$, which implies that the excitations are electric dipole active. Can these resonances be new phonons, activated in the low temperature phase? We have performed shell model calculations that put the lowest phonon near 100 cm$^{-1}$ for any reasonable parameters. Thus, we believe that we can reliably identify these low-frequency peaks as electromagnons.
Another check of the electromagnon origin of this modes is comparison of their contribution to the step-like anomaly in the temperature dependence of $\varepsilon_{1}$, calculated using the Kramers-Kronig relation, with the measured one.
![(Colour online) Dielectric constant of TbMn$_2$O$_5$ from fits of infrared spectra (lower curve) in comparison with kHz measurements. The whole step-like anomaly is due to electromagnons (figure \[s1\]a, blue curve).[]{data-label="figeps"}](figure3.eps){width="8cm"}
Figure \[figeps\] shows such a comparison. We have chosen TbMn$_2$O$_5$ for this purpose because of the larger sample size and higher frequency of the electromagnon peaks. It is clear from figure \[figeps\] that the whole step-like anomaly in $\varepsilon_1(T)$ comes entirely from the sharp electromagnon peak.
The frequency and temperature behaviour, presented in figures \[s1\] and \[figeps\] is typical for a set of $R$Mn$_2$O$_5$ compounds: $R$ = Er, Ho, Y, Dy, Tb, Gd and Eu [@Kobayashi2004; @Sirenko-Ho; @Sushkov125; @Golovenchits2003; @Sanina1988; @RolandoPhD]. The appearance of electromagnon peaks as well as the $\varepsilon$ step-like anomaly seem to be correlated with the transition into the non-collinear spin state. Notably, BiMn$_2$O$_5$ whose spin ordering is nearly collinear at all temperatures, shows neither the $\varepsilon$ step-like anomaly nor electromagnon absorption [@RolandoPhD].
An inelastic neutron scattering study of YMn$_2$O$_5$ is in progress. Preliminary data by S.-H. Lee [@LeeAPS] show several scattering peaks in energy scans at the wave vector of the static spin structure. A strong neutron feature is observed at 1 meV in good agreement with the sharp low frequency feature in the low temperature infrared spectrum (figure \[s1\]a).
Katsura [@Katsura-em] predict that the electromagnon originating from antisymmetric exchange has a ‘transversal’ polarisation with respect to the spontaneous polarisation: $\bi{e} \perp \bi{P}$. The observed polarisation selection rule for electromagnons in $R$Mn$_2$O$_5$ compounds is ‘longitudinal’: $\bi{e} \| \bi{P}$. We will show that such ‘longitudinal’ electromagnon can be obtained in a model based on the isotropic Heisenberg exchange.
Far infrared spectroscopy of $R$MnO$_3$ {#sec:exp113}
=======================================
It is interesting to compare electromagnons in the two families of multiferroic manganites. We begin with Eu$_{0.75}$Y$_{0.25}$MnO$_3$ — a compound that mimics lattice parameters of TbMnO$_3$ but does not have $f$–$f$ transitions in the far infrared [@Valdes-EuY].
![(Colour online) Optical conductivity of Eu$_{0.75}$Y$_{0.25}$MnO$_3$ for $\bi{e} \| \bi{a}$ in the ferroelectric (blue), the spin density wave (dashed), and the paramagnetic (red) phases. Electromagnon band consists of a broad background and two peaks. []{data-label="s1EuY"}](figure4.eps){width="8cm"}
To extract the parameters of the oscillators, we fit the transmission spectra with a Lorentzian model of the dielectric constant $\varepsilon(\omega)$ for electric dipole transitions: $$\varepsilon(\omega) = \varepsilon_\infty+\sum_j
\frac{S_j}{\omega_{j}^2-\omega^2-\imath\omega\gamma_j}
\label{Lorentz}$$ where $\varepsilon_\infty $ is the high frequency dielectric constant, $j$ enumerates the oscillators, $S_j$ is the spectral weight, $\omega_{j}$ is the resonance frequency, and $\gamma_j$ is the damping rate.
Figure \[s1EuY\] shows optical conductivity spectra for three phases of Eu$_{0.75}$Y$_{0.25}$MnO$_3$. Figure \[swEuY\] shows temperature dependence of the fit parameter $S_j$ (spectral weight of the peaks in figure \[s1EuY\]).
![(Colour online) Spectral weight of the absorption peaks below 140 cm$^{-1}$ in figure \[s1EuY\]. The numbers from 1 to 5 enumerate the curves. Frequencies in the legend are the lowest temperature resonance frequencies of each mode. Inset: Total spectral weight of the eight phonons above 140 cm$^{-1}$.[]{data-label="swEuY"}](figure5.eps){width="8cm"}
A broad absorption (electromagnon) band exists well above the Néel temperature 47 K (red curve in figure \[s1EuY\] and curve 3 in figure \[swEuY\]). Upon cooling, the spectral weight $S$ of this broad band is growing down to $T_{FE}$=30 K and stays at this level at lower temperatures. Such a behaviour of the background can be seen from the spectra — the absorption at 40 cm$^{-1}$, the frequency least affected by two absorption peaks, stays constant at all $T < T_{FE}$. This absorption produces smooth growth of the $\varepsilon_a(T)$ at $T_{FE} < T$. $T_N$=47 K is the inflection point of the curve 3 which is an evidence for the magnetic origin of this low frequency broad electric absorption. Two electromagnon peaks emerge sharply at the $T_{FE}$ (blue curve in figure \[s1EuY\] and curves 1 and 4 in figure \[swEuY\]). They produce all further growth of the $\varepsilon_a(T)$ at $T < T_{FE}$. The electromagnon spectrum of TbMnO$_3$ is very similar to that one of Eu$_{0.75}$Y$_{0.25}$MnO$_3$, except the large electromagnon peak is at 60 instead of 80 cm$^{-1}$ .
It is also interesting to follow the temperature dependence of the redistribution of the electric dipole spectral weight $S$. Since electromagnons result from a small admixture of phonons to magnons, the total spectral weight should be conserved. Comparing the curves 2 and 5 in figure \[swEuY\], one can estimate how much spectral weight the 120 cm$^{-1}$ phonon is loosing and how much electromagnons acquire. As electromagnons gain more spectral weight than the phonon is loosing, we checked the rest of the phonon peaks for the same polarisation of light. The inset in figure \[swEuY\] shows that, indeed, phonons lose just enough of the spectral weight to conserve total spectral weight.
In their recent inelastic neutron scattering work, Senff [@SenffPRL] reported a set of modes at the incommensurate zone center of TbMnO$_3$. The frequency of the lowest mode is equal to the frequency of the lowest infrared peak (24 cm$^{-1}$). We agree with Senff in assignment of this mode as electromagnon. However, despite the satisfied polarisation selection rule $\bi{e} \perp \bi{P}$, this is not the electromagnon predicted by Katsura [@Katsura-em], as we discuss below.
The model of $R$Mn$_2$O$_5$ {#sec:model}
===========================
The presence of two different types of magnetic ions and geometric frustration of spin interactions give rise to rather complex magnetic structures $R$Mn$_2$O$_5$ compounds. Nonetheless, a number of salient properties of these materials, such as the magnetically-induced electric polarisation, photo-excitation of magnons as well as the spin re-orientation transition, can be understood within a simplified microscopic model, which we discuss in this section. Our starting point is the assumption that multiferroic and optical properties of these materials are governed by the isotropic Heisenberg exchange, although we do include magnetic anisotropy to explain the spin re-orientation transition that has a strong effect on the low-frequency absorption spectrum. We discuss ordered spin states of the model, the mechanism of magnetoelectric coupling and calculate the optical absorption spectrum at magnon frequencies for different magnetic states.
We start by considering a single magnetic $ab$-layer including Mn$^{3+}$ and Mn$^{4+}$ ions. The model describes interactions between the spins and their coupling to a polar phonon mode: $$\begin{array}{ll}
H = & \frac{1}{2}\sum_{i,j}J_{ij}\left( {\bf P}\right) \left( \bi{S}_{i}
\cdot \bi{S}_{j}\right) -\frac{1}{2}\sum_{i\alpha } K_{i\alpha } \left( \bi{S}_{i} \cdot \hat{k}_{i\alpha }\right)^{2}- \\
& -\sum_{i}\mu _{i}\left( \bi{S}_{i} \cdot \bi{H}\right) +V\left(\frac{\bi{P}^{2}}{2\chi _{1}^{(0)}}-\bi{PE} - \frac{\chi_{2}
E^{2}}{2} \right).
\end{array}
\label{eq:phi}$$ Here the first term is the isotropic Heisenberg spin exchange, while the second term is the single-ion anisotropy. The antisymmetric Dzyaloshinskii-Moriya exchange as well as other types of anisotropic exchange interactions are not included, as in the scenario discussed below they play no role. We assume for simplicity that in the ordered state all spins lie in the $ab$ plane [@ChaponPRL2006] and neglect the small out-of-plane components found in recent neutron diffraction experiments on single crystals [@Noda2006; @Vecchini2008]. Thus, the easy and intermediate magnetic axes on each Mn site ($\alpha = 1,2$) lie in the $ab$ plane, while the hard axis $\hat{k}_{i3}\parallel \hat{c}$. The third term in (\[eq:phi\]) is the interaction of spins with an applied magnetic field and the last term describes the dielectric response of the system, where $\chi _{1}^{(0)}$ is the ‘bare’ dielectric susceptibility related to the polar lattice mode (not including the magnetic contribution calculated below) and $\chi _{2}$ is the remaining part of the dielectric susceptibility of non-magnetic origin. Finally, $V$ is the system volume.
The coupling between the spins and the polar phonon mode results from the dependence of the exchange coupling on the electric polarisation, which in $R$Mn$_2$O$_5$ is parallel to the $b$ axis: $$%J_{ij}\left( P\right) =J_{ij}\left( 0\right) +J_{ij}^{\prime }\left( 0\right) P \label{eq:J(P)}
J_{ij}(P_{b}) = J_{ij}(0) + J_{ij}^{\prime }(0)P_{b} + \frac{1}{2}
J_{ij}^{\prime\prime }(0)P_{b}^2 + \ldots \label{eq:J(P)}$$ The last two terms give rise to the cubic and the quartic magneto-electric couplings.
Magnetic ordering and spin re-orientation transition
----------------------------------------------------
Here, we adopt the model of Chapon [@ChaponPRL2006] with 5 exchange constants between pairs of nearest-neighbour Mn ions: $J_{1}$ and $J_{2}$ couple Mn$^{4+}$ ions along the $c$ direction, $J_{3}$ and $J_{5}$ couple the spins of neighbouring Mn$^{3+}$ and Mn$^{4+}$ ions and $J_{4}$ is the coupling between two neighbouring Mn$^{3+}$ ions (see figure \[fig:ordering\]).
![(Colour online) Minimal energy spin configurations for $J_{4}=J_{5}=40$ K and the anisotropy parameter $K_{a}(\mbox{Mn}^{3+})=0.6$ K for all Mn$^{3+}$ ions (red) and $K_{a}(\mbox{Mn}^{4+}) =0.1$ K for all Mn$^{4+}$ ions (blue). The value of the interchain coupling $J_{3}$ is $-2$ K for the structure in panel (a) and $-4$ K for the one panel (b).[]{data-label="fig:ordering"}](figure6.eps){width="8cm"}
Figure \[fig:ordering\] shows the minimal energy spin configurations obtained by the numerical minimization of the spin energy (\[eq:phi\]) on the subspace of the commensurate magnetic states with the wave vector $\bi{Q} = \left(1/2,0,0\right)$ for two different sets of exchange constants. The exchange constants $J_{4}$ and $J_{5}$ were chosen to be positive and large compared to other exchange constants, which gives rise to antiferromagnetic zig-zag chains along the $a$ axis with nearly collinear spins (marked by dashed lines) observed in neutron experiments [@ChaponPRL2006; @Vecchini2008].
The angle between spins in neighbouring chains sensitively depends on the ratio between the interchain coupling $J_{3}$ and the magnetic anisotropy parameters $\bi{K}_{i}$. We assume that the easy magnetic axis is parallel to the $a$ axis. The interplay between magnetic anisotropy and interchain interaction determines the angle between spins in neighbouring antiferromagnetic chains.
We first note that if spins in each antiferromagnetic chain would be perfectly collinear than the interchain interactions would cancel as a result of geometric frustration. Conversely, the interchain coupling $J_{3}$ results in spin rotations, which destroy the collinearity of spins in each chain. Consider, for example, the spin of the Mn$^{4+}$ ion, marked in figure \[fig:ordering\](a) by an arrow. Because of a nonzero angle between spins in the neighbouring $a$-chains, the interaction of the marked spin with the spin of the Mn$^{3+}$ ion from the neighbouring chain will give rise to a rotation of the marked spin. These small spin rotations lift the frustration and lead to some energy gain due to interchain interactions. It easy to show that this energy gain is maximal when spins in neighbouring antiferromagnetic chains are orthogonal to each other. Thus, while the magnetic anisotropy favours an almost collinear spin configuration, interchain interactions favour the $90^{\circ}$ angle between spins of neighbouring chains.
This competition gives rise to a very strong sensitivity of the angle between spins in neighbouring chains to the interchain coupling $J_{3}$. For weak interchain coupling $J_{3} = -2$ K, this angle is relatively small and minimal-energy spin configuration shown in figure \[fig:ordering\](a) is similar to the one observed in the high-temperature ‘collinear’ phase of YMn$_2$O$_5$ by Chapon [@ChaponPRL2006]. A small change in $J_{3}$ from $-2$ K to $-4$ K transforms the configuration shown in figure \[fig:ordering\](a) into the one shown in figure \[fig:ordering\](b), which may explain the spin re-orientation observed in $R$Mn$_2$O$_5$ with $R = $Tb, Ho, Dy and Y, provided that the interchain coupling is temperature-dependent. Although the rotations that make spins in each chain non-collinear are barely visible, they are sufficient to produce the large changes in the spin configuration, since the magnetic anisotropy is relatively weak. In $R$Mn$_2$O$_5$, the spin re-orientation transition is accompanied by the loss of commensurability of the spin structure in the $a$ and $c$ directions (which is also a consequence of magnetic frustration). This latter aspect of the transition is, however, less important for the photo-excitation of magnons, discussed below, than the re-orientation of spins.
Magnetically-induced polarisation
---------------------------------
Minimizing (\[eq:phi\]) with respect to $P_{y}$, we obtain expression for the magnetically-induced electric polarisation, $$P_{b} \approx -\frac{\chi _{1}^{(0)}}{2V}\sum_{i,j}J_{ij}^{\prime }\left(
0\right) \left( \bi{S}_{i}\bi{S}_{j}\right), \label{eq:P0}$$ which only involves scalar products of spins.
![A cartoon of the magnetic ordering in the high-temperature collinear phase.[]{data-label="fig:cartoon"}](figure7.eps){width="8cm"}
![Electric polarisation induced by magnetostriction along the $b$-chains. The two order parameters $\eta_{1}$ and $\eta_{2}$ describe degenerate ferroelectric states with opposite directions of electric polarisation.[]{data-label="fig:dipole"}](figure8.eps){width="12cm"}
Figures \[fig:cartoon\] and \[fig:dipole\] show why in the high-temperature ferroelectric phase the polarisation vector is oriented along the $b$ axis. Figure \[fig:cartoon\] gives a simplified view of the Mn layer in the $ab$ plane, in which the spins inside squares depict the spins of Mn$^{4+}$ ions located inside oxygen octahedra, while the spins inside triangles are the spins of Mn$^{3+}$ ions in oxygen pyramids. The nearly collinear magnetic ordering in the high-temperature ferroelectric phase consists of antiferromagnetic chains along the $a$ direction.
The mechanism responsible for electric polarisation in this magnetic state involves, however, the $\uparrow\uparrow\downarrow$ and $\downarrow\downarrow\uparrow$ spin chains along the $b$ axis, such as shown in figure \[fig:dipole\]. These chains contain the polar Mn$^{4+}$ – Mn$^{3+}$ bonds, connecting parallel spins, and the Mn$^{3+}$ – Mn$^{4+}$ bonds with opposite polarity, connecting antiparallel spins. Importantly, the charge and spin modulations in the chains have the same period, in which case the conventional exchange striction destroys the cancellation of electric dipoles of the polar bonds and induces electric polarisation along the chains, as illustrated in figure \[fig:dipole\].
This mechanism also works in the low-temperature (incommensurate) ferroelectric phase, which has the same periodicity in the $b$-direction. The amplitude of the exchange striction is, however, the largest for collinear spins \[see (\[eq:P0\])\], which explains the drop of the polarisation at the transition to the low-temperature phase. For example, if the value of the magnetoelectric coupling $\propto J_{3}^{\prime} - J_{4}^{\prime} - J_{5}^{\prime}$, is chosen such that the electric polarisation induced by the ‘high-temperature’ configuration shown in figure \[fig:ordering\](a) is $1000$ $\mu $C m$^{-2}$, then for the ‘low-temperature’ configuration shown in figure \[fig:ordering\](b) it equals $500$ $\mu $C m$^{-2}$.
Phenomenological approach
-------------------------
In this subsection we discuss the phenomenological description of spin states and the magnetoelectric coupling mechanism discussed above, which will clarify the similarities between $R$Mn$_2$O$_5$ and other multiferroic materials. The positions of Mn$^{3+}$ and Mn$^{4+}$ in the paramagnetic unit cell are shown in table \[tab:coordinates\] and the eight vector order parameters can be found in table \[tab:orderparameters\] (we use the notations of Bertaut [@Bertaut1967]).
------------------------------------------------------- -- --------------------------------------
Mn$^{3+}$ Mn$^{4+}$
\[0.5ex\] $\bi{r}_{1} = \left(x,y,1/2\right)$ $\bi{r}_{5} = \left(1/2,0,z\right)$
\[0.5ex\] $\bi{r}_{2} = \left(-x,-y,1/2\right)$ $\bi{r}_{6} = \left(1/2,0,-z\right)$
\[0.5ex\] $\bi{r}_{3} = \left(1/2-x,1/2+y,1/2\right)$ $\bi{r}_{7} = \left(0,1/2,z\right)$
\[0.5ex\] $\bi{r}_{4} = \left(1/2+x,1/2-y,1/2\right)$ $\bi{r}_{8} = \left(0,1/2,-z\right)$
\[1ex\]
------------------------------------------------------- -- --------------------------------------
: The coordinates of Mn ions, where $x \approx 0.41$, $y \approx 0.35$ and $z \approx 0.26$ (for BiMn$_2$O$_5$).
\[tab:coordinates\]
------------------------------------------------------------------ -- -----------------------------------------------------------------
Mn$^{3+}$ Mn$^{4+}$
\[0.5ex\] $\bi{F} = \bi{S}_{1}+\bi{S}_{2}+\bi{S}_{3}+\bi{S}_{4}$ $\bi{F}^{\prime} = \bi{S}_{5}+\bi{S}_{6}+\bi{S}_{7}+\bi{S}_{8}$
\[0.5ex\] $\bi{C} = \bi{S}_{1}+\bi{S}_{2}-\bi{S}_{3}-\bi{S}_{4}$ $\bi{C}^{\prime} = \bi{S}_{5}+\bi{S}_{6}-\bi{S}_{7}-\bi{S}_{8}$
\[0.5ex\] $\bi{G} = \bi{S}_{1}-\bi{S}_{2}+\bi{S}_{3}-\bi{S}_{4}$ $\bi{G}^{\prime} = \bi{S}_{5}-\bi{S}_{6}+\bi{S}_{7}-\bi{S}_{8}$
\[0.5ex\] $\bi{A} = \bi{S}_{1}-\bi{S}_{2}-\bi{S}_{3}+\bi{S}_{4}$ $\bi{A}^{\prime} = \bi{S}_{5}-\bi{S}_{6}-\bi{S}_{7}+\bi{S}_{8}$
\[1ex\]
------------------------------------------------------------------ -- -----------------------------------------------------------------
: Magnetic order parameters.
\[tab:orderparameters\]
---------------------- ------------------------------- ------------------------------ ---------------------------------- ------------------------------ ----------------------------- ---------------------------------- ----------------------------------
2$_{x}$ 2$_{y}$ 2$_{z}$ m$_{x}$ m$_{y}$ m$_{z}$ I
\[0.1ex\]
$\Gamma_{1}$ $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left(
\begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc}
\!\!\!\!0\!\! & \!\!1\!\!\\ \!\!1\!\! & \!\!0\!\!\!\!\\ \!\!\!\!0\!\! & \!\!-1\!\!\!\!\\ \!\!1\!\! & \!\!0\!\!\!\!\\ \!\!\!\!0\!\! & \!\!1\!\!\\ 1 & 0\\ \!\!\!\!0\!\! & \!\!-1\!\!\!\!\\
\!\!\!\!-1\!\! & \!\!0\!\! \!\!0\!\! & \!\!-1\!\!\!\! \!\!\!\!-1\!\! & \!\!0\!\!\!\! \!\!0\!\! & \!\!-1\!\!\!\! \!\!\!\!-1\!\! & \!\!0\!\! 0 & 1 \!\!\!\!-1\!\! & \!\!0\!\!\!\!
\end{array} \end{array} \end{array} \end{array} \end{array} \end{array} \end{array}
\right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$
\[5ex\] $\Gamma_{2}$ $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left( $\!\!\left(
\begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc} \begin{array}{cc}
\!\!0\!\! & \!\!-1\!\!\!\!\\ \!\!1\!\! & \!\!0\!\!\!\!\\ 0 & 1\\ \!\!\!\!-1\!\! & \!\!0\!\!\\ \!\!\!\!0\!\! & \!\!1\!\!\\ \!\!\!\!-1\!\! & \!\!0\!\!\!\!\\ \!\!\!\!0\!\! & \!\!-1\!\!\!\!\\
\!\! 1\!\! & \!\!0\!\!\!\! \!\!0\!\! & \!\!-1\!\!\!\! 1 & 0 \!\!\!\!0\!\! & \!\!1\!\! \!\!\!\!-1\!\! & \!\!0\!\! \!\!\!\!0\!\! & \!\!-1\!\!\!\! \!\!\!\!-1\!\! & \!\!0\!\!\!\!
\end{array} \end{array} \end{array} \end{array} \end{array} \end{array} \end{array}
\right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$ \right)\!\!$
---------------------- ------------------------------- ------------------------------ ---------------------------------- ------------------------------ ----------------------------- ---------------------------------- ----------------------------------
: Irreducible representations of the space group $Pbam$ for $\bi{Q} = \left(1/2,0,1/2\right)$.
\[tab:representations\]
---------------------- ------------------ ------------------- ------------------ ------------------------- -------------------------- ------------------------- -------------------------- -------------------------- -------------------------- ---------------------------------- ----------------------------------- ----------------------------------
$\Gamma_{1}$ $\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left(
\begin{array}{c} \begin{array}{c} \begin{array}{c} \begin{array}{c} \begin{array}{c} \begin{array}{c} \begin{array}{c}
\!\!F_{x}\!\!\\ \!\! C_{y} \!\!\\ \!\!\!\!G_{x}\!\!\!\!\\ \!\!\!\!-A_{y}\!\!\!\!\\ \!\!C_{z}^{\prime}\!\!\\ \!\!\!\!G_{x}^{\prime}\!\!\!\!\\ \!\!\!\!-A_{y}^{\prime}\!\!\!\!\\
\!\!C_{x}\!\! \!\! F_{y} \!\! \!\!\!\!-A_{x}\!\!\!\! \!\!\!\!G_{y}\!\!\!\! \!\!F_{z}^{\prime}\!\! \!\!\!\!-A_{x}^{\prime}\!\!\!\! \!\!\!\!G_{y}^{\prime}\!\!\!\!
\end{array} \end{array} \end{array} \end{array} \end{array} \end{array} \end{array}
\right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$
\[5ex\] $\Gamma_{2}$ $\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left( $\!\!\!\!\left(
\begin{array}{c} \begin{array}{c} \begin{array}{c} \begin{array}{c} \begin{array}{c}
\!\!F_{z}\!\!\\ \!\!\!\!G_{z}\!\!\!\!\\ \!\!C_{x}^{\prime}\!\!\\ \!\!F_{y}^{\prime}\!\!\\ \!\!\!\!G_{z}^{\prime}\!\!\!\!\\
\!\!C_{z}\!\! \!\!\!\!-A_{z}\!\!\!\! \!\!F_{x}^{\prime}\!\! \!\!C_{y}^{\prime}\!\! \!\!\!\!-A_{z}^{\prime}\!\!\!\!
\end{array} \end{array} \end{array} \end{array} \end{array}
\right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$ \right)\!\!\!\!$
---------------------- ------------------ ------------------- ------------------ ------------------------- -------------------------- ------------------------- -------------------------- -------------------------- -------------------------- ---------------------------------- ----------------------------------- ----------------------------------
: Basis vectors of the space group $Pbam$ for $\bi{Q} = \left(1/2,0,1/2\right)$.
\[tab:vectors\]
For discussion of phenomenological description of the magnetoelectric coupling, we have chosen the relatively simple case of BiMn$_2$O$_5$, which shows the commensurate spin ordering with $\bi{Q} = \left(1/2,0,1/2\right)$. In this case the components of the order parameters belong to one of the two two-dimensional representations, $\Gamma_{1}$ or $\Gamma_{2}$, of the $Pbam$ group [@Bertaut1967; @Munoz2002]. BiMn$_2$O$_5$ only shows the ‘collinear’ state with the $a$-components of the Mn$^{3+}$ and Mn$^{4+}$ spins described, respectively, by the order parameters $F_{x} = -3.1 \mu_{\rm B}$ and $G_{x}^{\prime} = 2.4 \mu_{\rm B}$ with small $b$ components $C_{y} = - 0.8 \mu_{\rm B}$ and $A_{y}^{\prime} = 0.6 \mu_{\rm B}$ [@Munoz2002] corresponding to a small rotation between spins in neighbouring antiferromagnetic chains.
Since $F_{x}$ is the part two-dimensional representation $
\left(
\begin{array}{c}
F_{x} \\ C_{x}
\end{array}
\right)
\in \Gamma_{1},
$ the state described by the order parameter $C_{x}$ is another ground state of the system. These two states are related by inversion, which transforms $F_{x}$ into $-C_{x}$ and vice versa. It is easy to check that the magnetoelectric coupling of the form $-\lambda_{x} P_{y} \left(F_{x}^2 - C_{x}^{2}\right)$ is invariant upon all symmetry transformations of the paramagnetic phase, so that the order parameters $F_{x}$ and $C_{x}$ describe two ferroelectric states with opposite directions of electric polarisation. It also easy to check that the couplings $-\lambda_{y} P_{y} \left(F_{y}^2 - C_{y}^{2}\right)$ and $-\lambda_{z} P_{z} \left(F_{z}^2 - C_{z}^{2}\right)$ are also allowed by symmetry, which is a strong indication that the mechanism inducing electric polarisation in magnetically ordered state is invariant upon the global spin rotation and the coupling can be written in the form $\lambda P_{y} \left( \bi{F}^2 - \bi{C}^{2} \right)$ [@Kadomtseva].
Due to the exchange coupling between the Mn$^{3+}$ and Mn$^{4+}$ ions, the order parameters $G_{x}^{\prime}$ and $F_{x}$ are strongly coupled. Since $
\left(
\begin{array}{c}
G_{x}^{\prime} \\ -A_{x}^{\prime}
\end{array}
\right)
$ also belongs to $\Gamma_{1}$ representation, the coupling between the two spin subsystems is phenomenologically described by $-g \left(F_{x} G_{x}^{\prime} - C_{x} A_{x}^{\prime}\right)$ [@Kadomtseva]. Thus, more generally, the magnetoelectric coupling should be written in the form $$\Phi_{\rm me} = - \lambda P_{y} \left(\eta_{1}^{2} - \eta_{2}^{2} \right),$$ where $
\left(
\begin{array}{c}
\eta_{1} \\ \eta_{2}
\end{array}
\right)
$ belongs to $\Gamma_{1}$ representation. This form of the third-order magnetoelectric coupling was discussed previously in the context of the orthorombic manganites with the $E$-type magnetic ordering [@SergienkoPRL2006] and is typical for improper ferroelectrics [@Levanyuk].
Static and dynamic dielectric susceptibility
--------------------------------------------
The contribution of the coupled spin-lattice degrees of freedom to static dielectric susceptibility is given by $$\chi _{1}^{-1}\approx \left( \chi _{1}^{\left( 0\right) }\right)
^{-1}-\frac{1}{V}\sum_{i,j}I_{i}A_{ij}^{-1}I_{j}+\frac{1}{2V}\sum_{i,j}J_{ij}^{\prime
\prime }\left( 0\right) \left( \bi{S}_{i}\bi{S}_{j}\right) ,
\label{eq:chistat}$$ where $$A_{ij}= J_{ij}\left( \bi{S}_{i}\bi{S}_{j}\right) +\delta
_{ij}\left[\sum_{\alpha =1,2}K_{i\alpha }\left( 2\left(
\bi{S}_{i}\hat{k}
_{i\alpha }\right) ^{2}-S_{i}^{2}\right) - \sum_{k}J_{ik}\left( \bi{S}_{i}
\bi{S}_{k}\right) \right]
\label{eq:A}$$
and $$I_{i}=\sum_{j}J_{ij}^{\prime }\left[ \bi{S}_{i}\times \bi{S}_{j}
\right] _{c}, \label{eq:I}$$ The second term in (\[eq:chistat\]) is the spin contribution to the dielectric constant due to virtual excitations of magnons by electric field (this will become more apparent in the discussion of the dynamic susceptibility).
The last term in (\[eq:chistat\]) describes the shift of the phonon frequency due to a change of the spring constants in magnetically ordered states. This effect is known in condensed matter spectroscopy as spin-phonon coupling [@Baltensperger; @Sushkov-PRL2005]. Phenomenologically, this effect is described by the fourth-order magnetoelectric coupling of the type $P^2L^2$, where $L$ is a magnetic order parameter. In most cases, magnon and phonon branches coupled through this term experience ‘repulsion’ and phonon hardens. However, in the magnetically frustrated compounds, this contribution to the dielectric constant can have either sign and can result in either hardening or softening of phonons in the magnetic phase (see, e. g., [@Wakamura1976] on CdCr$_2$S$_4$ spinel).
Equations of motion describing the coupled spin-lattice dynamics have the form, $$\left\{
\begin{array}{ccc}
\ddot{P_{b}} & = & -\frac{\chi _{0}\omega _{0}^{2}}{V}\frac{\partial H }{\partial P_{b}}, \\
& & \\
{\dot \bi{S}}_{i} & = & \left[ \frac{\partial H }{\partial
\bi{S}_{i}}\times \bi{S}_{i}\right] ,
\end{array}
\right.
\label{eq:motion}$$ where $\chi_{0} = \chi_{1}^{(0)} + \chi_{2}$ and $\omega _{0}$ is the bare frequency of the polar phonon. Omitting the fourth order coupling term and solving linearized equations of motion, we obtain the dynamic dielectric susceptibility: $$\chi ^{-1}\left( \omega \right) \approx {\chi_{0}}^{-1}\left(1-\frac{\omega ^{2}}{\omega _{0}^{2}}\right) - \frac{1}{V}\sum_{i,j}I_{i}\left[ \left( BA-\omega ^{2}\right) ^{-1}B\right] _{ij}I_{j},
\label{chidyn}$$ where $$B_{ij} = J_{ij}+\frac{\delta_{ij}}{S_{i}^{2}} \sum_{\alpha =1,2}\left(K_{i\alpha }\left( \bi{S}_{i}{\hat k}_{i\alpha }\right)^{2}-K_{i3}S_{i}^{2}\right) -
\frac{\delta_{ij}}{S_{i}^{2}} \sum_{k}J_{ik}\left( \bi{S}_{i}\bi{S}_{k}\right) .
\label{eq:B}$$
The second term in (\[chidyn\]) describes the transfer of a part of electric dipole spectral weight from phonon to magnon frequencies, which turns magnons coupled to phonons into electromagnons. If such an electromagnon has lower frequency than the phonon, the static dielectric constant $\varepsilon(0)=1+4\pi\chi(0)$ increases as a result of the coupling showing a step-like anomaly. Furthermore, frequencies of the mixed spin-lattice excitations (poles of the dielectric susceptibility (\[chidyn\])) are shifted down with respect to the ‘bare’ magnon frequencies, found from $$\det \left( BA-\omega^{2}\right) =0.
\label{eq:magfreq}$$ Note that the electromagnon term in (\[chidyn\]) disappears for collinear spin states, as $I_{i}$, defined by (\[eq:I\]), is zero in this case. This can be understood as follows. Classically, magnons correspond to spin oscillations that are orthogonal to ordered spin vectors. The change of the scalar product of a pair of collinear spins is then proportional to the second power of the amplitude of the oscillations. Since the magnetoelectric coupling in our model originates solely from Heisenberg exchange and only involves scalar products of spins, the linear coupling of electric field to magnons is absent for collinear spins and the lowest-order process is the photo-excitation of a pair of magnons.
![(Colour online) The model calculation results: Frequency dependence of the real ($\varepsilon_1$) and imaginary ($\varepsilon_2$) parts of the dielectric function (red and blue lines, respectively) for the ‘high-temperature’ collinear state (panel a) and the ‘low-temperature’ non-collinear state (panel b) shown, respectively, in panels (a) and (b) of figure \[fig:ordering\]. Red points are selected magnon frequencies of spins decoupled from the lattice found as the roots of (\[eq:magfreq\]). A magnon at 11.8 cm$^{-1}$ couples to a polar phonon at 100 cm$^{-1}$ and becomes an electromagnon observable as the peak of the $\varepsilon_2$.[]{data-label="fig:modelspectra"}](figure9.eps){width="8cm"}
In figures \[fig:modelspectra\](a) and (b), we plot the real and imaginary parts of the dielectric function (red and blue lines, respectively) for the ‘high-temperature’ and ‘low-temperature’ states shown in, respectively, figures \[fig:ordering\](a) and (b) (the imaginary part was obtained by the shift $\omega \rightarrow $ $\omega +i\frac{\gamma }{2}$ with $\gamma =1$ K). As was explained above, the coupling of magnons to the electric component of light and significant electric dipole absorption at magnon frequencies is only present in the non-collinear ‘low-temperature’ phase, in agreement with experiment. This result may seem somewhat counterintuitive: while the spontaneous electric polarisation induced by the Heisenberg exchange striction is maximum for the collinear state, the excitation of magnons by the electric component of light (electromagnons) requires non-collinear spins and is only observed below the spin re-orientation transition.
The red points in figure \[fig:modelspectra\] mark the bare frequencies of the softest magnons with zero wave vector, found from (\[eq:magfreq\]). Since the magnetic unit cell in this calculation contains 16 mangetic ions, the total number of such magnons is also 16. However, only one of them is strongly coupled to the electric component of light and significantly contributes to the dielectric constant. This electromagnon corresponds to relative rotation of spins of the neighbouring antiferromagnetic chains, which gives rise to oscillations of the induced electric polarisation in the $b$ direction. The frequency of this uncoupled magnon for the non-collinear spin configuration shown in figure \[fig:ordering\](b) is 11.8 cm$^{-1}$. As the position of the absorption peak in figure \[fig:modelspectra\](b) is clearly lower that this magnon frequency, our parameters correspond to the strong magnetoelectric coupling case. The strong coupling is apparently necessary, if the large difference between dielectric constants of the ‘high’- and ‘low’-temperature phases $\Delta \varepsilon ^{\prime }\left( 0\right) $ ($=3.25$ in our calculation) is associated solely with the absorption peak emerging in the non-collinear state.
The results of our model calculations presented in figure \[fig:modelspectra\](b) show that the symmetry properties of one magnon and the lattice allow an electromagnon in this system. Next, we discuss the strength of coupling. In our model, the first derivatives of the exchange integrals $J_{ij}$ are the coupling constants. Both magnetically induced polarisation and electromagnon were calculated using the same set of $J^{\prime}$ values. Characteristic value of $J^{\prime}$ used in our calculation was $dJ_3/dy$ = 4 meV/$\AA$ which is much less than, for example, 40 meV/$\AA$ for ZnCr$_2$O$_4$ [@LeePRL].
Discussion {#sec:discussion}
==========
Our model calculations show that the observed values of the polarisation and electromagnon strength are obtained with realistic symmetric exchange constants. Also, the electromagnon coupling has been shown to be of Heisenberg type by the observed selection rules. These are strong arguments in favour of the Heisenberg exchange origin of the polarisation and electromagnons in $R$Mn$_2$O$_5$ family. However, in comparing the experimental and theoretical results, we should keep in mind that in the compounds of interest both symmetric and antisymmetric exchange mechanisms are operative, and they both may produce polarisation and electromagnons. Therefore, it is important to compare both the spontaneous polarisation and the electromagnon selection rules in both $R$MnO$_3$ and $R$Mn$_2$O$_5$ with these two mechanisms.
In particular, the origin of ferroelectricity in $R$Mn$_2$O$_5$ remains controversial. An alternative to the magnetostriction scenario discussed here is electric polarisation induced by the $bc$ spiral, as observed in recent experiments ([@Noda2006; @Vecchini2008; @Kim2008]) via the inverse Dzyaloshinskii-Moriya mechanism of relativistic origin [@Katsura2005; @SergienkoDagotto; @Mostovoy-spiral]. One natural question is whether the electromagnon study can clarify this controversy.
Electromagnon excitations for the cycloidal spiral state were studied by Katsura [@Katsura-em]. According to their theory, the electric field of light that excites electromagnon has to be orthogonal to the direction of the electric polarisation induced by the spiral, which can be easily understood as follows. The spin spiral induces electric polarisation that lies in the spiral plane and is orthogonal to the propagation wave vector of the spiral [@Mostovoy-spiral]. Thus the $bc$-spiral in $R$Mn$_2$O$_5$ with the wave vector along the $c$ axis would induce polarisation along the $b$ axis, in agreement with experiment. An electric field applied in the direction perpendicular to the spiral plane will result in a small rotation of this plane, as a result of which the spontaneous polarisation vector acquires a component parallel to the applied field. An oscillating electric field orthogonal to the spiral plane will then excite oscillations of this plane, which is precisely the electromagnon of Katsura [@Katsura-em]. Since the polarisation lies in the spiral plane, the polarisation of light should be orthogonal to the spontaneous polarisation. Thus, to excite the oscillations of the $bc$ spiral would require $\bi{e} \| \bi{a}$, whereas in experiment electric field is parallel to the $b$ axis and the direction of electric polarisation. Indeed, this can be an argument against the spiral scenario of multiferroicity of $R$Mn$_2$O$_5$.
However, the situation is complicated by the fact that even in $R$MnO$_3$ compounds, where the spiral origin of the magnetically-induced electric polarisation is well established, the selection rule $\bi{e} \perp \bi{P}$ is not obeyed. This was demonstrated by optical measurements in magnetic field. Magnetic field applied to DyMnO$_3$ [@Kida2008] and TbMnO$_3$ [@RolandoPhD] gives rise to the magnetic flop transition at which the $bc$-spiral is replaced by the $ab$-spiral. The spiral flop, however, has no effect on the polarisation of electromagnons, which are always observed for $\bi{e} \| \bi{a}$. Also in Eu$_{0.75}$Y$_{0.25}$MnO$_3$, where the spin spiral lies in the $ab$ plane already for zero magnetic field and the induced electric polarisation is parallel to $a$, the selection rule for electromagnons remains the same: ${\bi e} \| \bi{a}$ [@Valdes-EuY]. Therefore, the selection rules in both $R$MnO$_3$ and $R$Mn$_2$O$_5$ are tied to the lattice, not the spin plane. The selection rule for $R$MnO$_3$ appears to originate from the GdFeO$_3$ distortions of the perovskite structure of orthorombic manganites, which generates new types of magnetoelectric interactions and couples light to magnetic excitations that are different from the oscillations of the spiral plane [@Mostovoy113]. Furthermore, these distortions may also couple to the zone edge magnons and account for the broad peak observed in $R$MnO$_3$ at high frequencies. However, an alternative mechanism for the broad peak is the photoexcitation of bi-magnons and so the precise origin of this high frequency peak in $R$MnO$_3$ remains an open question.
In any case, it appears that electromagnons in $R$MnO$_3$ are also induced by symmetric exchange even though the spontaneous polarisation is produced by the DM-type antisymmetric exchange mechanism. This is understood as a consequence of the vanishing of any static polarisation induced by symmetric exchange due to the symmetry of the lattice [@SergienkoDagotto]. Another important question then is why the antisymmetric exchange does not produce electromagnon excitations? We suspect that the relativistic interactions are just too weak to produce observable signals. The observation of these weak DM induced electromagnons would help clarify the overall picture of of spin-lattice interactions in multiferroics.
Consequently, while our experimental and theoretical results strongly suggest that the magnetoelectric coupling in $R$Mn$_2$O$_5$ is governed by Heisenberg exchange, we cannot safely rule out the spiral scenario for the spontaneous polarisation in these materials on the basis of optical data. However, the results of recent neutron scattering experiments [@Kim2008] suggest rather that the spiral components appear as a consequence of the existence of spontaneous polarisation.
Summary
=======
We have presented experimental evidence and theoretical analysis that demonstrate that electromagnon excitations are present in both $R$Mn$_2$O$_5$ and $R$MnO$_3$ multiferroics. These electric dipole mixed magnon-phonon modes are observed in the far infrared and match excitations observed in inelastic neutron spectra in both classes of materials. The polarisation selection rules observed in the infrared experiments show that the electromagnon excitations are generated only by symmetric exchange in both classes of materials. The observed selection rules are tied to the lattice, not the spin plane, in contrast to the predictions of the antisymmetric exchange model.
To theoretically account for electromagnons, we have considered third and fourth order coupling between the lattice and spins. The fourth order terms produce spin-phonon interactions that lead to shifts of the magnon and phonon frequencies near magnetic phase transitions observed in many magnetic materials. The third order coupling terms can produce mixed magnon-phonon excitations –– the electromagnons –– only for non-collinear spin orders. The third order terms are also responsible for the spontaneous polarisation in multiferroics [@Cheong2007]. We also developed a simple microscopic model of $R$Mn$_2$O$_5$ that explains the transition between collinear and non-collinear spin states, the magnetically-induced electric polarisation in the collinear state and the appearance of the electromagnon peak in the non-collinear state as well the polarisation of light for which this peak is observed. The mechanism of magnetoelectric coupling in this model relies entirely on isotropic Heisenberg exchange and magnetostrictive coupling of spins to a polar phonon mode.
This work was supported in part by the National Science Foundation MRSEC under Grant No. DMR-0520471. MM gratefully acknowledges the support by DFG (Mercator fellowship) and the hospitality of Cologne University. We acknowledge S. Park, C.L. Zhang and Y.J. Choi for the growth and characterization of the single crystals.
References {#references .unnumbered}
==========
[99]{}
Smolenskii G A and Chupis I E 1982 [*Sov. Phys. Usp.*]{} [**25**]{} 475
Fiebig M 2005 [*J. Phys. D*]{} [**38**]{} R123
Prellier W, Singh M P and Murugavel P 2005 R803
Khomskii D I 2006 1
Eerenstein W, Mathur N D and Scott J F 2006 [*Nature*]{} [**442**]{} 759
Cheong S-W and Mostovoy M 2007 [*Nature Mater*]{} [**6**]{} 13
Ramesh R and Spaldin N A 2007 [*Nature Mater*]{} [**6**]{} 21
Kimura T, Goto T, Shintani H, Ishizaka K, Arima T and Tokura Y 2003 [*Nature*]{} [**426**]{} 55
Hur N, Park S, Sharma P A, Ahn J S, Guha S and Cheong S-W 2004 [*Nature*]{} [**429**]{} 392
Lawes G, Harris A B, Kimura T, Rogado N, Cava R J, Aharony A, Entin-Wohlman O, Yildirim T, Kenzelmann M, Broholm C and Ramirez A P 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 087205
Goto T, Kimura T, Lawes G, Ramirez A P, Tokura Y 2004 [*Phys. Rev. Lett.*]{} [**92**]{} 257201
Pimenov A, Mukhin A A, Ivanov V Yu, Travkin V D, Balbashov A M and Loidl A 2006 [*Nature Phys.*]{} [**2**]{} 97
Sushkov A B, Valdés Aguilar R, Park S, Cheong S-W and Drew H D 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 027202
Katsura H, Balatsky A V, and Nagaosa N 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 027203
Kimura T 2007 [*Annu. Rev. Mater. Res.*]{} [**37**]{} 387
Moriya T 1960 [*Phys. Rev.*]{} [**120**]{} 91
Katsura H, Nagaosa N and Balatsky A V 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 057205
Sergienko I A and Dagotto E 2006 [*Phys. Rev.*]{} B [**73**]{} 094434
Jia C, Onoda S, Nagaosa N and Han J H 2007 [*Phys. Rev.*]{} B [**76**]{} 144424
Kadomtseva A M, Krotov S S, Popov Yu F, Vorob’ev G P and Lukina M M 2005 [*JETP*]{} [**100**]{} 305
Chapon L C, Radaelli P G, Blake G R, Park S and Cheong S-W 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 097601
Sergienko I A, Sen C and Dagotto E 2006 [*Phys. Rev. Lett.*]{} [**97**]{} 227204
Bary’achtar V G, L’vov V A and Jablonskii D A 1983 [*JETP Letters*]{} [**37**]{} 673
Mostovoy M 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 067601
Levanyuk A P and Sannikov D G 1974 [*Sov. Phys. Usp.*]{} [**17**]{} 199
Lorenzana J and Sawatzky G A 1995 [*Phys. Rev. Lett.*]{} [**74**]{} 1867
Damascelli A, van der Marel D, M. Grüninger M, Presura C, Palstra T T M, Jegoudez J and Revcolevschi A 1998 [*Phys. Rev. Lett.*]{} [**81**]{} 918
Pimenov A, Rudolf T, Mayr F, Loidl A, Mukhin A A and Balbashov A M 2007 [*Phys. Rev.*]{} B [**74**]{} 100403(R)
Valdés Aguilar R, Sushkov A B, Zhang C L, Choi Y J, Cheong S-W and Drew H D 2007 [*Phys. Rev.*]{} B [**76**]{} 060404(R)
Pimenov A, Loidl A, Mukhin A A, Travkin V D, Ivanov V Yu and Balbashov A M 2008 [*Phys. Rev.*]{} B [**77**]{} 014438
Kida N, Ikebe Y, Takahashi Y, He J P, Kaneko Y, Yamasaki Y, Shimano R, Arima T, Nagaosa N and Tokura Y 2007 arXiv:0711.2733
de Sousa R and Moore E J 2008 [*Phys. Lett.*]{} B [**77**]{} 012406
de Sousa R and Moore E J 2008 [*Appl. Phys. Lett.*]{} [**92**]{} 022514
Fang C and Hu J 2008 [*Europhys. Lett.*]{} [**82**]{} 57005
Bertaut E F, Buisson G, Quezel-Ambrunaz S and Quezel G 1967 [*Solid State Commun.*]{} [**5**]{} 25
Harris A B, Aharony A and Entin-Wohlman O 2008 [*Phys. Rev. Lett.*]{} [**100**]{} 217202
Sanina V A, Sapozhnikova L M, Golovenchits E I and Morozov N V 1988 [*Sov. Phys. Solid State*]{} [**30**]{} 1736
Mukhin A A, Travkin V D, Ivanov V Yu, Lebedev S P, Prokhorov A S and Kon K 2001 [*Bulletin of the Lebedev Physics Institute*]{} [**4**]{} 34
Golovenchits E I and Sanina V A 2003 [*JETP Lett.*]{} [**78**]{} 88
Kuzmenko A B 2004 Reffit: Software to fit Optical Spectra (http://optics.unige.ch/alexey/reffit.html)
Kobayashi S, Osawa T, Kimura H, Noda Y, Kagomiya I and Kohn K 2004 [*J. Phys. Soc. Jpn.*]{} [**73**]{} 1031
Sirenko A A, O Malley S M, Ahn K H, Park S, Carr G L and Cheong S-W 2007 arXiv:cond-mat/0703255v1
Valdés Aguilar R 2008 PhD Thesis, University of Maryland
Lee S-H, Kim J H, Chung J-H, Qiu Y, Kenzelmann M, Sato T J, Park S Y, Cheong S-W 2007 [*Bull. Am. Phys. Soc.*]{} [**52**]{} 330
Senff D, Link P, Hradil K, Hiess A, Regnault L P, Sidis Y, Aliouane N, Argyriou D N and Braden M 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 137206
Noda Y, Kimura H, Kamada Y, Osawa T, Fukuda Y, Ishikawa Y, Kobayashi S, Wakabayashi Y, Sawa H, Ikeda N and Kohn K 2006 [*Physica*]{} B 385
Vecchini C, Chapon L C, Brown P J, Chatterji T, Park S, Cheong S-W and Radaelli P G 2008 [*Phys. Rev.*]{} B [**77**]{} 134434
Muñoz A, Alonso J A, Casais M T, Martínez-Lope M J, Martínez J L and Fernández-Díaz M T 2002 [*Phys. Rev.*]{} B [**65**]{} 144423
Baltensperger W and Helman J S 1968 [*Helv. Phys. Acta*]{} [**41**]{} 668
Sushkov A B, Tchernyshyov O and Ratcliff II W, Cheong S-W and Drew H D 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 137202
Wakamura K and Arai T 1988 [*J. Apl. Phys.*]{} [**63**]{} 5824
Lee S-H, Broholm C, Kim T H, Ratcliff II W. and Cheong S-W 2000 [*Phys. Rev. Lett.*]{} [**84**]{} 3718
Kim T H 2008 arXiv:0803.1123v2
Mostovoy M to be published
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Jean Daniel Mukam
- |
Jean Daniel Mukam ([email protected])\
African Institute for Mathematical Sciences (AIMS)\
Senegal\
\
[ Supervised by : Dr. Antoine Tambue]{}\
[ AIMS-South Africa and University of Cape Town ]{}\
[ [email protected]]{}
date: |
[ 13 June 2015]{}\
[*Submitted in Partial Fulfillment of a Masters II at AIMS Senegal*]{}\
title: |
Stochastic Calculus with Jumps Processes : Theory and Numerical Techniques\
(Master Thesis)
---
Abstract {#abstract .unnumbered}
========
In this work we consider a stochastic differential equation (SDEs) with jump. We prove the existence and the uniqueness of solution of this equation in the strong sense under global Lipschitz condition. Generally, exact solutions of SDEs are unknowns. The challenge is to approach them numerically. There exist several numerical techniques. In this thesis, we present the compensated stochastic theta method (CSTM) which is already developed in the literature. We prove that under global Lipschitz condition, the CSTM converges strongly with standard order 0.5. We also investigated the stability behaviour of both CSTM and stochastic theta method (STM). Inspired by the tamed Euler scheme developed in [@Martin1], we propose a new scheme for SDEs with jumps called compensated tamed Euler scheme. We prove that under non-global Lipschitz condition the compensated tamed Euler scheme converges strongly with standard order $0.5$. Inspired by [@Xia2], we propose the semi-tamed Euler for SDEs with jumps under non-global Lipschitz condition and prove its strong convergence of order $0.5$. This latter result is helpful to prove the strong convergence of the tamed Euler scheme. We analyse the stability behaviours of both tamed and semi-tamed Euler scheme We present also some numerical experiments to illustrate our theoretical results.\
\
**Key words** : Stochastic differential equation, strong convergence, mean-square stability, Euler scheme, global Lipschitz condition, polynomial growth condition, one-sided Lipschitz condition.
Declaration {#declaration .unnumbered}
-----------
I, the undersigned, hereby declare that the work contained in this essay is my original work, and that any work done by others or by myself previously has been acknowledged and referenced accordingly.
Jean Daniel Mukam, 25 May 2015
INTRODUCTION {#introduction .unnumbered}
============
In many branches of sciences like finance, economics, biology, engineering, ecology one often encountered some problems influenced by uncertainties. For example, in finance, the unpredictable nature of events such as markets crashes and booms may have significant and sudden impact on the stock price fluctuations. Therefore, in order to have more realistic prediction of these phenomena, it is natural to model them with equations which involves the deterministic part and the random part including jump. The SDEs with jumps is the generalization of both deterministic part and random part with jumps. SDEs with jumps have probability theory and stochastic process as prerequisites. We refer to [@Oksa], [@Oks], [@Phi] for general notions in probability theory and stochastic process.
In this thesis, under global Lipschitz condition, we prove the existence and uniqueness of solution of SDEs with jumps. We focus on the strong convergence of the compensated stochastic theta methods (CSTM) of these equations under global Lipschitz condition. In particular, we prove that CSTM have strong convergence of order $0.5$. We investigate the stability of both CSTM and stochastic theta method (STM). For the linear case, we prove that under the assumption $\dfrac{1}{2}\leq \theta\leq 1$, CSTM holds the A-stability property. For the general nonlinear problem, we study the stability for $\theta=1$. In this case, when the drift coefficient have a negative one-sided Lipschitz coefficient, the diffusion coefficient and the jump coefficient satisfy the global Lipschitz condition, we prove that STM reproduce stability under certain step-size and the CSTM is stable for any step-size.
Most phenomena are modelised by SDEs with jumps where the drift coefficient is one-sided Lipschitz and satisfies the polynomial growth condition. For such equations, it is proved in [@Martin3] that Euler explicit method fails to converge strongly to the exact solution while Euler implicit method converges strongly, but requires much computational efforts. Recently, a new explicit and efficient method was developed in [@Martin1] called tamed Euler scheme. In [@Martin1], the authors proved that the tamed Euler converges strongly with order 0.5 to the exact solution of SDEs under non-global Lipschitz condition. In this thesis, we extend the tamed Euler scheme by introducing a compensated tamed Euler scheme for SDEs with jumps. We prove that this scheme converges strongly with standard order $0.5$. We also extend the semi-tamed Euler developed in [@Xia2] and we prove that this scheme converge strongly with order $0.5$ for SDEs with jumps. As a consequence of this latter result, we prove the strong convergence of the tamed Euler scheme for SDEs with jumps. The stability analysis of both tamed Euler and semi-tamed Euler are done in this thesis.
This thesis is organized as follows. In chapter 1, we recall some basic notions in probability theory and stochastic process. Chapter 2 is devoted to the proof of the existence and uniqueness of SDEs with jumps under global Lipschitz condition. In chapter 3, we focus on the strong convergence of the CSTM and the stability analysis of both CSTM and STM. In chapter 4, under non-global Lipschitz condition we investigate the strong convergence of the compensated tamed Euler scheme. In chapter 5, under non-global Lipschitz condition, we investigate the strong convergence and stability of both semi-tamed Euler scheme and tamed Euler scheme. Our theoretical results are illustrated by numerical examples at the end of chapter 3, chapter 4 and chapter 5. \[chp: chap1 Basic Notions in probability theory and stochastic process\]
Basic notions in probability theory and stochastic process
==========================================================
Basic notions in probability theory
-------------------------------------
In this chapter, we present some basic concepts and results in probability theory and stochastic process useful to understand the notion of stochastic differential equations.
More details for this chapter can be found in [@Oksa], [@Oks] and [@Phi].
### **Basic notions in probability theory**
**\[$\sigma$- algebra\]**
Let $\Omega$ be a non-empty set.
1. A $\sigma$-algebra (or $\sigma$-field) $\mathcal{F}$ on $\Omega$ is a family of subsets of $\Omega$ satisfying
1. $\Omega \in \mathcal{F}$.
2. $\forall A\in \mathcal{F}$, $A^c\in\mathcal{F}$.
3. If $(A_i)_{i\in I}$ is a countable collection of set in $\mathcal{F}$, then $\cup_{i\in I} A_i\in \mathcal{F}$.
2. Let $\mathcal{F}_1$ and $\mathcal{F}_2$ be two $\sigma$-algebra on $\Omega$. $\mathcal{F}_1$ is said to be a sub-$\sigma$-algebra of $\mathcal{F}_2$ if $\mathcal{F}_1\subset\mathcal{F}_2$.
<!-- -->
1. Given any family $\mathcal{B}$ of subset of $\Omega$, we denote by $$\begin{aligned}
\sigma(\mathcal{B}):=\cap\{\mathcal{C}: \hspace{0.2cm}\mathcal{C}, \hspace{0.2cm}\sigma- \text{algebra of}\hspace{0.2cm} \Omega, \hspace{0.2cm}\mathcal{B}\subset \mathcal{C}\}\end{aligned}$$ the smallest $\sigma$-field of $\Omega$ containing $\mathcal{B}$, $\sigma(\mathcal{B})$ is called the $\sigma$-field generated by $\mathcal{B}$.
When $\mathcal{B}$ is a collection of all open sets of a topological space $\Omega$, $\sigma(\mathcal{B})$ is called the Borel $\sigma$-algebra on $\Omega$ and the elements of $\sigma(\mathcal{B})$ are called Borel sets.
2. If $ X: \Omega \longrightarrow\mathbb{R}^n$ is a function, then the $\sigma$-algebra generated by $X$ is the smallest $\sigma$-algebra on $\Omega$ containing all the sets of the form $$\begin{aligned}
\{X^{-1}(U) : U\subset \mathbb{R}^n, \hspace{0.5cm}\text{open}\}.\end{aligned}$$
**\[Probability measure\]**.
Let $\mathcal{F}$ be a $\sigma$-field on $\Omega$. A probability measure is an application $\mathbb{P} : \mathcal{F}\longrightarrow [0,1]$ satisfying
1. $\mathbb{P}(\Omega)=1-\mathbb{P}(\emptyset)=1$.
2. If $(A_i)_{i\in I}$ is a countable collection of elements of $\mathcal{F}$ pairwise disjoints, then
$\mathbb{P}(\cup_{i\in I}A_i)=\sum\limits_{i\in I}\mathbb{P}(A_i)$.
**\[Probability space\]**.
Let $\Omega$ be a non-empty set, $\mathcal{F}$ a $\sigma$-field on $\Omega$ and $\mathbb{P}$ a probability measure on $\mathcal{F}$.
The triple $(\Omega, \mathcal{F}, \mathbb{P})$ is called a probability space.
**\[Negligeable set\]**
1. Given a probability space $(\Omega, \mathcal{F}, \mathbb{P})$, $A\subset\Omega$ is said to be $\mathbb{P}$-null or negligeable if $\mathbb{P}(A)=0$
2. A property is said to be true almost surely (a.s) if the set on which this property is not true is negligeable.
**\[Measurability and random variable\]**
1. Let $(\Omega, \mathcal{F}, \mathbb{P})$ and $(\Omega', \mathcal{F}', \mathbb{P}')$ be two probability spaces. A function $X : \Omega \longrightarrow \Omega'$ is said to be $\mathcal{F}$-measurable if and only if $$\begin{aligned}
X^{-1}(U) := \{ \omega\in \Omega : X(\omega) \in U\}\subset \mathcal{F}, \hspace{0.5cm} \forall\hspace{0.2cm} U\in \mathcal{F}'\end{aligned}$$
2. A random variable $X$ is a function $X : \Omega \longrightarrow \Omega'$ $\mathcal{F}$-measurable.
3. If $\Omega'=\mathbb{R}$, then $X$ is called a real random variable.
4. If $\Omega'=\mathbb{R}^n$, $n>1$ then $X$ is called a vector random variable.
In the following, unless otherwise state, $(\Omega, \mathcal{F}, \mathbb{P})$ denote a probability space and $X$ a random variable, $X :\Omega \longrightarrow \mathbb{R}^n$.
.
Every random variable induces a probability measure on $\mathbb{R}^n$ denoted $\mu_X$ and define by
$\mu_X(B) :=\mathbb{P}(X^{-1}(B))$, $\forall\, B$ open set of $\mathbb{R}^n$. $\mu_X$ is called the distribution function of $X$.
**\[Expected value\]**
1. If $X$ is a random variable such that $\int_{\Omega}||X(\omega)||d\mathbb{P}(\omega)<\infty$ almost surely, the quantity $$\begin{aligned}
\mathbb{E}(X) :=\int_{\Omega}X(\omega)d\mathbb{P}(\omega) =\int_{\mathbb{R}^n}d\mu_X(x)
\end{aligned}$$ is called the expected value of $X$, where $||.||$ denote the euclidean norm on $\mathbb{R}^n$.
2. In general, if $f : \mathbb{R}^n\longrightarrow \mathbb{R}^m$ is measurable and $\int_{\Omega}||f(X(\omega))||d\mathbb{P}(\omega)<\infty $ almost surely, then the qauntity $\mathbb{E}(f(X))$ define by $$\begin{aligned}
\mathbb{E}(f(X)) :=\int_{\Omega}f(X(\omega))d\mathbb{P}(\omega)=\int_{\mathbb{R}^n}f(x)d\mu_X(x)\end{aligned}$$ is called expected value of $f(X)$.
**\[Independent random variables\]**
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space.
1. Two elements $A$ and $B$ of $\mathcal{F}$ are independent if $$\begin{aligned}
\mathbb{P}(A\cap B)=\mathbb{P}(A)\cap\mathbb{P}(B).\end{aligned}$$
2. Two random variables $X_1$ and $X_2$ of $(\Omega, \mathcal{F}, \mathbb{P})$ are independent if for every choice of different borel sets $B_1$ and $B_2$ the following holds : $$\begin{aligned}
\mathbb{P}(X_1\in B_1, X_2\in B_2)=\mathbb{P}(X_1\in B_1)\times\mathbb{P}(X_2\in B_2).\end{aligned}$$
The following proposition is from [@Oksa].
Two random variables $X_1$ and $X_2$ are independent if and only if for any measurable positive functions $f_1$ and $f_2$, the following equality holds $$\begin{aligned}
\mathbb{E}(f_1(X_1)f_2(X_2))=\mathbb{E}(f_1(X_1)) \mathbb{E}(f_2(X_2)).\end{aligned}$$
**\[Conditional probability\]**
For any event $A$ such that $P(A)>0$, the conditional probability on $A$ is the probability measure define by : $$\begin{aligned}
\mathbb{P}(B/A) :=\dfrac{\mathbb{P}(A\cap B)}{\mathbb{P}(A)},\hspace{0.3cm} \forall B\in \mathcal{F}.\end{aligned}$$
### Conditional expectation
Let $X$ be a random variable such that $\int_{\Omega}|X(\omega)|d\mathbb{P}(\omega)<\infty$ almost surely. Let $\mathcal{G}$ a sub $\sigma$-algebra of $\mathcal{F}$. The conditional expectation of $X$ relative to the $\sigma$-algebra $\mathcal{G}$ is a random variable denoted by $\mathbb{E}(X/\mathcal{G})$ satisfying
1. $ \mathbb{E}(X/\mathcal{G})$ is $\mathcal{G}$-measurable.
2. $\int_G\mathbb{E}(X/\mathcal{G})d\mathbb{P}=\int_GXd\mathbb{P}, \hspace{0.5cm}\forall \hspace{0.2cm} G\in \mathcal{G}$.
In the litterature, $\mathbb{E}(X/\mathcal{G})$ is called the projection of $X$ upon $\mathcal{G}$.
The proof of the following theorem can be seen in [@Phi].
1. $ \mathbb{E}(\mathbb{E}(X/\mathcal{G}))=\mathbb{E}(X)$.
2. If $X$ is $\mathcal{G}$-measurable, then $\mathbb{E}(X/\mathcal{G})=X$.
3. $\mathbb{E}((X+Y)/\mathcal{G})=\mathbb{E}(X/\mathcal{G})+\mathbb{E}(Y/\mathcal{G})$.
4. If $\mathcal{G}\subset \mathcal{G}'$ then $\mathbb{E}(X/\mathcal{G}')=\mathbb{E}.(\mathbb{E}(X/\mathcal{G})/\mathcal{G}')$.
5. If $\sigma(X)$ and $\mathcal{G}$ are independent, then $\mathbb{E}(X/\mathcal{G})=\mathbb{E}(X)$.
6. If $X\leq Y$ a.s, then $\mathbb{E}(X/\mathcal{G})\leq \mathbb{E}(Y/\mathcal{G})$.
7. If $X$ is $\mathcal{G}$ measurable, then $\mathbb{E}(XY/\mathcal{G})=X\mathbb{E}(Y/\mathcal{G})$.
### Convergence of random variables
.
Let $p\in[1,\infty)$, we denote by $\mathbb{L}^p(\Omega, \mathbb{R}^n)$ the equivalence class of measurable functions $X :\Omega : \longrightarrow \mathbb{R}^n $, $\mathcal{F}_t$-measurable such that $$\begin{aligned}
||X||^p_{\mathbb{L}^p(\Omega,\mathbb{R}^n)} := \mathbb{E}(||X||^p) =\int_{\Omega}||X(\omega)||^pd\mathbb{P}(\omega)<+\infty.\end{aligned}$$ Let $(X_n)\subset \mathbb{L}^p(\Omega, \mathbb{R}^n)$ be a sequence of random variables and $X\in \mathbb{L}^p(\omega, \mathbb{R}^n)$ a random variable. Let $$\begin{aligned}
N:=\{\omega : \lim_{n\longrightarrow \infty}X_n(\omega)=X(\omega)\}\end{aligned}$$
1. $(X_n)$ converges to $X$ almost surely if $N^c$ is negligeable.
2. $(X_n)$ converges in probability to $X$ if $$\begin{aligned}
\forall\hspace{0.2cm}\epsilon >0\hspace{0.3cm}\lim_{n\longrightarrow\infty}\mathbb{P}(||X_n-X||>\epsilon)=0.\end{aligned}$$
3. $(X_n)$ converges in $\mathbb{L}^p$ to $X$ if $$\begin{aligned}
\lim_{n\longrightarrow +\infty}\mathbb{E}(||X_n-X||^p)=0.
\end{aligned}$$
: **\[Frobenius norm\]**
The Frobenius norm of a $m\times n$ matrix $A=(a_{ij})_{1\leq i\leq n; 1\leq j\leq m}$ is defined by $$\begin{aligned}
||A||:=\sqrt{\sum_{j=1}^n\sum_{i=1}^n|a_{ij}|^2}.\end{aligned}$$
Frobenius norm and euclidean norm are the same for vectors.
\[ch1Minkowski\]**\[Minkowski inequality : Integral form\]**.
Let $1\leq p<+\infty$ and let $(X, \mathcal{A}, dx)$ and $(Y, \mathcal{B}, dy)$ be $\sigma$-finite measures spaces. Let $F$ be a measurable function on the product space $X\times Y$. Then $$\begin{aligned}
\left(\int_X\left|\int_YF(x,y)dy\right|^pdx\right)^{1/p}\leq \int_Y\left(\int_X|F(x,y)|^pdx\right)^{1/p}dy.\end{aligned}$$ The above inequality can be writen as $$\begin{aligned}
\left\|\int_YF(.,y)dy\right\|_{L^p(X,\mathcal{A}, dx)}\leq \int_Y||F(.,y)||_{L^p(X, \mathcal{A}, dx)}dy.\end{aligned}$$
**\[Gronwall inequality\] : Continous form**
Let $a(t)$ and $b(t)$ be two continuous and positives functions defines on $\mathbb{R}_+$ such that $$\begin{aligned}
a(t)\leq b(t)+c\int_0^ta(s)ds, \hspace{0.5cm} \forall\, t\in \mathbb{R}_+,\end{aligned}$$ then $$\begin{aligned}
a(t)\leq b(t)+c\int_0^tb(s)e^{c(t-s)}ds, \hspace{0.5cm}\forall\, t\in \mathbb{R}_+.\end{aligned}$$ **\[Gronwall inequality\] : Discrete form**.
Let $\theta$ and $K$ be two constants and $(v_n)$ be a sequence satisfying : $$\begin{aligned}
v_{n+1}\leq (1+\theta)v_n+K,\end{aligned}$$ then $$\begin{aligned}
v_n\leq e^{n\theta}v_0+K\dfrac{e^{n\theta}-1}{e^{\theta}-1}.\end{aligned}$$
[@Gron].
**\[Borel Cantelli\]**
Let $(A_n)_{n\in \mathbb{N}}$ be a family of subset of $\Omega$.
1. If $\sum\limits_{n\in\mathbb{N}}\mathbb{P}(A_n)<\infty$, then $\mathbb{P}\left[\limsup\limits_{n\longrightarrow\infty}A_n\right]=0$.
2. If the events $(A_n)$ are independent and $\sum\limits_{n\in\mathbb{N}}\mathbb{P}(A_n)=0$, then $\mathbb{P}\left[\limsup\limits_{n\longrightarrow \infty}A_n\right]=1$.
[@Phi].
Stochastic processes
--------------------
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space. A family $(\mathcal{F}_t)_{t\geq 0}$ of sub $\sigma$-algebra of $\mathcal{F}$ is called filtration if $\mathcal{F}_s\subset \mathcal{F}_t$, $\forall\hspace{0.2cm} 0\leq s\leq t$.
If $(\mathcal{F}_t)$ is such that $\mathcal{F}_t=\cap_{t>s}\mathcal{F}_s$, then $(\mathcal{F}_t)_{t\geq 0}$ is said to be right continuous.
A stochastic process is a family of vector random variables $(X_t)_{t\geq 0}$. That is for all $t> 0$, the application
$\begin{array}{cccc}
X_t & : \Omega &\longrightarrow &\mathbb{R}^n\\
& w & \longmapsto& X_t(\omega)
\end{array}$ is measurable.
If $(X_t)_{t\geq 0}$ is a stochastic process, then for all $t\geq 0$, the application $t\longmapsto X_t$ is called sample path.
Let $(\mathcal{F}_t)$ be a filtration on $(\Omega, \mathcal{F}, \mathbb{P})$. A stochastic process $(X_t)$ is said to be $\mathcal{F}_t$-adapted if $\forall\hspace{0.2cm} t\geq 0$ $X_t$ is $\mathcal{F}_t$-measurable.
**\[Martingale\]** Let $(\mathcal{F}_t)_{t\geq 0}$ be a filtration on $(\Omega, \mathcal{F}, \mathbb{P})$. A stochastic process $(M_t)_{t\geq 0}$ is called $\mathcal{F}_t$- martingale if the following properties holds
1. $(M_t)$ is $\mathcal{F}_t$-adapted.
2. $\mathbb{E}||M_t||<\infty $, $\forall\hspace{0.1cm} t\geq 0$.
3. $\mathbb{E}(M_t/\mathcal{F}_s)=M_s$, $\forall\hspace{0.1cm} 0\leq s\leq t$.
<!-- -->
1. If the condition $(iii)$ of the previous definition is replaced by $\mathbb{E}(M_t/\mathcal{F}_s)\geq M_s$, $\forall\hspace{0.1cm} 0\leq s\leq t$, then $(M_t)$ is called submartingale.
2. If the condition $(iii)$ of the previous definition is replaced by $\mathbb{E}(M_t/\mathcal{F}_s)\leq M_s$, $\forall\hspace{0.1cm} 0\leq s\leq t$, then $(M_t)$ is called supermartingale.
3. A positive submartingale is a submartingale $(X_t)_{t\geq 0}$ satisfying $X_t\geq 0$ for all $t\geq 0$.
**\[Predictable process\]**
Let $(\mathcal{F}_t)_{t\geq 0}$ be a filtration on $(\Omega, \mathcal{F}, \mathbb{P})$. A stochastic process $(X_t)_{t\geq 0}$ is called $\mathcal{F}_t$- predictable process if for all $t> 0$, $X_t$ is measurable with respect to the $\sigma$-algebra generated by $\{X_s, \hspace{0.2cm}s<t\}$.
Let $M=(M_t)$ be a submartingale. Then for $1<p<\infty$, we have
1. **Markov’s inequality** $$\begin{aligned}
\mathbb{P}\left(\sup_{0\leq s \leq t}||M_t||\geq \alpha\right)\leq\dfrac{\mathbb{E}(||M_t||)}{\alpha},\hspace{0.6cm} \forall\, \alpha>0.
\end{aligned}$$
2. **Doop’s maximal inequality** $$\begin{aligned}
\mathbb{E}\left[\left(\sup_{0\leq s \leq t}||M_t||\right)^p\right]^{1/p}\leq \dfrac{p}{p-1}\mathbb{E}\left[||M_t||^p\right]^{1/p}.
\end{aligned}$$
[@Phi].
**\[Wiener process or Brownian motion\]**
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space and $(\mathcal{F}_t)_{t\geq 0}$ a filtration on this space. A $\mathcal{F}_t$-adapted stochastic process $(W_t)_{t\geq 0}$ is called Wiener process or Brownian motion if :
1. $W_0=0$.
2. $ t \longmapsto W_t$ is almost surely continous.
3. $(W_t)_{t\geq 0}$ has independent increments $($i.e $W_t-W_s$ is independent of $W_r,\hspace{0.2cm} r\leq s)$.
4. $W_t-W_s\rightsquigarrow \mathcal{N}(0, t-s)$, for $0\leq s \leq t$. Usually, this property is called stationarity.
If $(W_t)$ is an $\mathcal{F}_t$- Brownian motion, then the following process are $\mathcal{F}_t$- martingales
1. $W_t$.
2. $W_t^2-t$.
3. $\exp\left(\gamma W_t-\gamma^2\dfrac{t}{2}\right),\hspace{0.3cm}\forall\; \gamma\in \mathbb{R}$.
Let $0\leq s\leq t$, then
1. $$\begin{aligned}
\mathbb{E}(W_t/\mathcal{F}_s)&=&\mathbb{E}(W_t-W_s+W_s/\mathcal{F}_s)\\
&=&W_s+\mathbb{E}(W_t-W_s/\mathcal{F}_s) \hspace{0.1cm}\text{since} \hspace{0.2cm} W_s \hspace{0.1cm} is \hspace{0.1cm} \mathcal{F}_s-\text{measurable}\\
& =& W_s+ \mathbb{E}(W_t-W_s)\hspace{0.1cm} (\text{since the increments are independents }) \\
&=&W_s \hspace{0.3cm}(\text{since} \hspace{0.1cm} W_t-W_s\rightsquigarrow\mathcal{N}(0,t-s)).
\end{aligned}$$
2. $$\begin{aligned}
\mathbb{E}(W_t^2-t/\mathcal{F}_s)&=&\mathbb{E}(W_t^2+W_s^2-2W_sW_t+2W_sW_t-W_s^2/\mathcal{F}_s)-t\\
&=&\mathbb{E}((W_t-W_s)^2/\mathcal{F}_s)+W_s\mathbb{E}((2W_t-W_s)/\mathcal{F}_s)-t \\
& & \hspace{0.1cm}(\text{since} \hspace{0.2cm} W_s \hspace{0.1cm} is \hspace{0.1cm} \mathcal{F}_s-\text{measurable })\\
&=&\mathbb{E}((W_t-W_s)^2)+ W_s\mathbb{E}(W_t-W_s)+W_s\mathbb{E}(W_t/\mathcal{F}_s)-t\\
& &(\text{since the increments are independents})\\
&=&t-s+0+W_s^2-t\hspace{0.2cm} \text{since} \hspace{0.2cm} W_t-W_s\rightsquigarrow\mathcal{N}(0, t-s)\\
&=&W_s^2-s.
\end{aligned}$$
Using the same argument as above, we have : $$\begin{aligned}
\mathbb{E}(e^{\gamma W_t}/\mathcal{F}_s)&=& e^{\gamma W_s}\mathbb{E}(e^{\gamma(W_t-W_s)}/\mathcal{F}_s)\\
&=& e^{\gamma W_s}\mathbb{E}(e^{\gamma W_{t-s}})\\
&=&e^{\gamma W_s}\int_{-\infty}^{+\infty}\dfrac{e^{-x^2/2(t-s)}}{\sqrt{2\pi(t-s)}}dx\\
&=&e^{\gamma W_s}e^{\gamma^2(t-s)/2}=e^{\gamma W_s+\gamma^2(t-s)/2}.
\end{aligned}$$ Therefore, $$\begin{aligned}
\mathbb{E}\left(exp\left(\gamma W_t-\gamma^2\dfrac{t}{2}\right)/\mathcal{F}_s\right)&=&\mathbb{E}(e^{\gamma W_t}/\mathcal{F}_s)e^{-\gamma^2t/2}\\
&=&e^{\gamma W_s+\gamma^2(t-s)/2}e^{-\gamma^2t/2}\\
&=& \exp(\gamma W_s-{\gamma^2 s}/2).
\end{aligned}$$
The following proposition is from [@Oks].
Almost all sample paths of a Brownian motion are nowhere differentiable.
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space and $(\mathcal{F}_t)$ a filtration on this space. Let $(S_k)_{k\geq 1}$ be an $\mathcal{F}_t$-adapted stochastic process on $(\Omega, \mathcal{F}, \mathbb{P})$ with $0\leq S_1(\omega)\leq S_2(\omega)\leq ...$ for all $k\geq 1$ and $\omega \in \Omega$. The $\mathcal{F}_t$- adapted process $N=(N_t)_{t\geq 0}$ defined by : $$\begin{aligned}
N_t:=\sum_{k\geq 1}\mathbf{1}_{\{S_k\leq k\}}\end{aligned}$$ is called counting process with jump times $S_k$.
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space and $(\mathcal{F}_t)$ a filtration on this space. A counting process $(N)_t$, $\mathcal{F}_t$- adapted is called poisson process of intensity $\lambda >0$ if :
1. $N_0=0$.
2. $\forall\; 0\leq t_0<t_1<...<t_n$, the random variables $\{N_{t_j}-N_{t_{j-1}} \hspace{0.3cm} 1\leq j\leq n\}$ are independent.
3. For $0\leq s\leq t$, $N_t-N_s\approx N_{t-s}$, where $\approx$ stand for the equality in probability law.
4. For all $t>0$, $N_t$ follows a poisson law with parameter $\lambda t$ (and we denote $N_t\leftrightsquigarrow \mathcal{P}(\lambda t)$). That is $$\begin{aligned}
\mathbb{P}(N_t=k)=e^{-\lambda t}\dfrac{(\lambda t)^k}{k!},\hspace{0.5cm} k\in \mathbb{N}.
\end{aligned}$$
**\[Compound poisson process\]**
Let $(Z_n)$ be a sequence of discrete independent identically distributed random variables with probability law $\nu_Z$. Let $N=(N_t)$ be a poisson process with parameter $\lambda$. Let’s assume that $(N_t)$ and $(Z_n)$ are independent. A compound poisson process with intensity $\lambda>0$ with a jump law $\nu_Z$ is a $\mathcal{F}_t$- adapted stochastic process $(Y_t)$ defined by : $$\begin{aligned}
Y_t: =\sum_{k=1}^{N_t}Z_k.
\end{aligned}$$
**\[Compensated poisson process\]**
A compensated poisson process associated to a poisson process $N$ with intensity $\lambda$ is a stochastic process $\overline{N}$ defined by : $$\begin{aligned}
\overline{N}(t) := N(t)-\lambda t.\end{aligned}$$
\[ch1quadratic\] Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space and $(\mathcal{F}_t)$ a filtration on this space.
If $(N)_t$ is a $\mathcal{F}_t$- adapted poisson process with intensity $\lambda$, then
1. $\overline{N}$ is a $\mathcal{F}_t$- adapted martingale.
2. $\mathbb{E}(\overline{N}(t+s)-\overline{N}(t))=0$.
3. $\mathbb{E}[\overline{N}(t+s)-\overline{N}(t)]^2=\lambda s, \hspace{0.5cm} \forall\; t,s \geq 0$.
4. $\overline{N}_t^2-\lambda t$ is a martingale.
<!-- -->
1. Let $\leq s\leq t$, then $$\begin{aligned}
\mathbb{E}(\overline{N}_t/\mathcal{F}_s)&=& \mathbb{E}(\overline{N}_t-\overline{N}_s+\overline{N}_s/ \mathcal{F}_s)\\
&=&\mathbb{E}(\overline{N}_t-\overline{N}_s/\mathcal{F}_s)+\overline{N}_s\\
&=&\mathbb{E}(N_t-N_s-\lambda t+\lambda s/\mathcal{F}_s)+N_s-\lambda s\\
&=&\mathbb{E}(N_t-N_s)-\lambda t+\lambda s +N_s-\lambda s \hspace{0.1cm}\\
& & \text{since the increments of the poisson process are independents}\\
&=& \lambda (t-s)-\lambda t+ N_s\hspace{0.2cm}\text( since\hspace{0.1cm} N_t-N_s\rightsquigarrow\mathcal{P}(\lambda(t-s)))\\
&=& N_s-\lambda s\\
&=& \overline{N}(s).\end{aligned}$$
2. $$\begin{aligned}
\mathbb{E}(\overline{N}(t+s)-\overline{N}(t))&=& \mathbb{E}(N(t+s)-N(t)-\lambda s)\\
&=& \lambda (t+s-t)-\lambda s= 0.\end{aligned}$$
3. $$\begin{aligned}
[\overline{N}(t+s)-\overline{N}(t)]^2&=&[N(t+s)-N(t)-\lambda s]^2\\
&=&[N(t+s)-N(t)]^2+\lambda^2s^2-2\lambda s(E(t+s)-N(t)).\end{aligned}$$
Since $N(t)\rightsquigarrow \mathcal{P}(\lambda t)$, using the relation $\mathbb{E}(N_t)=var(N_t)= \lambda t$, it follows that : $$\begin{aligned}
\mathbb{E}[\overline{N}(t+s)-\overline{N}(s)]^2&=&\lambda(t+s-t)+\lambda^2(t+s-t)^2+\lambda^2s^2-2\lambda s(\lambda s)=\lambda s. \end{aligned}$$
4. $$\begin{aligned}
\mathbb{E}[\overline{N}^2_t-\lambda t/\mathcal{F}_s]&=& \mathbb{E}[\overline{N}^2_t/\mathcal{F}_s]-\lambda t\\
&=&\mathbb{E}[\overline{N}^2_t+\overline{N}^2_s-2\overline{N}_t\overline{N}_s+2\overline{N}_t\overline{N}_s-
\overline{N}_s^2/\mathcal{F}_s]-\lambda t\\
&=&\mathbb{E}[(\overline{N}_t-\overline{N}_s)^2/\mathcal{F}_s]+\mathbb{E}[\overline{N}_s(\overline{N}_t-\overline{N}_s)/\mathcal{F}_s]+\mathbb{E}[\overline{N}_t\overline{N}_s/\mathcal{F}_s].\end{aligned}$$
Using the fact that $\overline{N}_t$ have independent increments and using the first part of the theorem, it follows that $$\begin{aligned}
\mathbb{E}[\overline{N}^2_t-\lambda t/\mathcal{F}_s]&=&\mathbb{E}[(\overline{N}_t-\overline{N}_s)^2]+\overline{N}_s\mathbb{E}(\overline{N}_t-\overline{N}_s)+\mathbb{E}[\overline{N}_t-\overline{N}_s+\overline{N}_s/\mathcal{F}_s]-\lambda t\\
&=&\lambda (t-s)+0+\overline{N}_s\mathbb{E}[\overline{N}_t-\overline{N}_s+\overline{N}_s/\mathcal{F}_s]-\lambda t\\
&=&\lambda t-\lambda s+0+0+\overline{N}^2_s-\lambda t\\
&=&\overline{N}^2_s-\lambda s\end{aligned}$$ This complete the proof.
Stochastic integral
-------------------
Let $\mathbb{M}^p([0,T], \mathbb{R})$ be the subspace of $\mathbb{L}^p([0,T], \mathbb{R})$ such that for any process
$(X_t)\in \mathbb{M}^p([0,T], \mathbb{R})$ we have $$\begin{aligned}
\mathbb{E}\left(\int_0^T|X(t)|^pdt\right)<\infty.\end{aligned}$$
Consider a Brownian motion $W$ and a stochastic process $(X_t)$ both adapted to a given filtration $(\mathcal{F}_t)$. We will define the following expression called stochastic integral $$\begin{aligned}
I_t(X)=\int_0^tX(s)dW(s).\end{aligned}$$ We will also give some of its properties.
Let’s start with the stochastic integral of simple process.
**\[Elementary process or simple process\]**
A process $(X_t)_{t\in\mathbb{R}}\in\mathbb{L}^p([0,T], \mathbb{R})$ is called simple or elementary process if there exist a partition $0=t_0<t_1<...<t_n=T$ such that $$\begin{aligned}
X_s(\omega)=\sum_{j=0}^{n}1_{]t_j, t_{j+1}]}\theta_j(\omega),\end{aligned}$$ where $\theta_j$ is a bounded $\mathcal{F}_{t_j}$-measurable random variable.
**\[ Itô’s integral\]**
The Itô’s Integral of the simple process $(X_t)_{t\in\mathbb{R}}\in\mathbb{L}^2([0,T], \mathbb{R})$ is defined by $$\begin{aligned}
I_t(X)=\int_0^tX(s)dW(s) :=\sum_{j=0}^{n-1}\theta_j(W_{t_{j+1}}-W_{t_j}).\end{aligned}$$
If $f$ is an elementary function in $\mathbb{L}^2([a,b],\mathbb{R})$ and $W_t$ a Brownian motion, then :
1. $\mathbb{E}\left(\int_a^bf(t)dW_t\right)=0$.
2. $\mathbb{E}\left(\int_a^bf(t)dW_t\right)^2=\int_a^b\mathbb{E}(f^2(t))dt$.
<!-- -->
1. By definition we have $$\begin{aligned}
\int_a^bf(t)dW_t=\sum_{j=0}^{n-1}f_j(W_{t_j+1}-W_{t_j}).\end{aligned}$$ By taking expectation in both sides, we obtain $$\begin{aligned}
\mathbb{E}\left[\int_a^bf(t)dW_t\right]=\sum_{j=0}^{n-1}\mathbb{E}(f_j)\mathbb{E}(W_{t_{j+1}}-W_{t_j})=0,\end{aligned}$$ since $W_{t_{j+1}}-W_{t_j}$ is a normal distribution with mean $0$ and standard deviation $\sqrt{t_{j+1}-t_j}$.
2. $$\begin{aligned}
\left(\int_a^bf(t)dW_t\right)^2&=&\left[\sum_{j=0}^{n-1}f_j(B_{t_{j+1}}-W_{t_j})\right]^2\\
&=&\sum_{j=0}^{n-1}(f_j)^2(W_{t_{j+1}}-W_{t_j})^2+\sum_{l=0}^{n-1}\sum_{k=0, k\neq l}^{n-1}f_lf_k(W_{t_{l+1}}-W_{t_l})(W_{t_{k+1}}-W_{t_k}).\end{aligned}$$
Taking expectation in both sides and using independence of the increments of Brownian motion, we get $$\begin{aligned}
\mathbb{E}\left(\int_a^bf(t)dW_t\right)^2&=&\sum_{j=0}^{n-1}\mathbb{E}(f_j)^2E\left(W_{t_{j+1}}-W_{t_j}\right)^2\\
&=&\sum_{j=0}^{n-1}\mathbb{E}(f_j)^2(t_{j+1}-t_j)\\
&=&\int_a^b\mathbb{E}(f^2(t))dt.\end{aligned}$$
The following proposition can be seen in [@Oks].
For any process $X=(X_t)_{t\geq 0}\in\mathbb{M}^2([0,T], \mathbb{R})$, such that $\mathbb{E}|X_t|^2<\infty $ for all $t\geq 0$, there exist a sequence $(f^{(n)}_t)_{t\geq 0}$ of simple process such that $\mathbb{E}|f^{(n)}_t|^2<\infty$ and $$\begin{aligned}
\lim_{n\longrightarrow \infty}\mathbb{E}\left[\int_0^t|X_s-f_s^{(n)}|^2ds\right]=0.\end{aligned}$$
For any process $X=(X_t)_{t\geq 0}\in\mathbb{M}^2([0,T], \mathbb{R})$, we define a stochastic integral of $X$ with respect to a Brownian motion $W$ by : $$\begin{aligned}
\int_0^tX_sdW(s)=\lim_{n\longrightarrow \infty}\int_0^tf^{(n)}_sdW(s),
\end{aligned}$$ where $(f^{(n)}_t)$ is the sequence of simple process converging almost surely to $X$ according to the previous proposition. Moreover, using Itô isometry for elementaries functions one can prove that the limit on this definition does not depend on the actual choice of $(f^{(n)}).$
**\[Properties of Itô integral\]**.
For any process $X=(X_t)_{t\geq 0}\in\mathbb{M}^2([0,T], \mathbb{R})$ such that $\mathbb{E}|X_t|^2<\infty$, for any functions
$f,g\in \mathbb{M}^2([0,T], \mathbb{R})$ and $0\leq S<U<T$, the following holds :
1. $\int_S^TfdW(t)=\int_S^UfdW(t)+\int_U^TfdW(t)$ almost surely.
2. $\int_S^T(cf+g)dW(t)=c\int_S^TfdW(t)+\int_S^TgdW(t)$, for any constant $c$.
3. $\int_S^TfdW(t)$ is $\mathcal{F}_T$-measurable.
4. $\mathbb{E}\left(\int_0^tX_sdW(s)\right)=0$.
5. $\mathbb{E}\left(\int_0^tX_sdW(s)\right)^2=\int_0^t\mathbb{E}(X_s^2)ds$.
[@Oks]
[@Oks] For any elementary function $f^{(n)}$ $\mathcal{F}_t$-adapted, the integral $$\begin{aligned}
I_n(t, \omega)=\int_0^tf^{(n)}dW(r)
\end{aligned}$$ is a martingale with respect to $\mathcal{F}_t$.
For $t\leq s$, we have : $$\begin{aligned}
\mathbb{E}[I_n(s,\omega)/\mathcal{F}_t]&=&\mathbb{E}\left[\left(\int_0^sf^{(n)}dW(r)\right)/\mathcal{F}_t\right]\\
&=&\mathbb{E}\left[\left(\int_0^tf^{(n)}dW(r)\right)/\mathcal{F}_t\right]+
\mathbb{E}\left[\left(\int_t^sf^{(n)}dW(r)\right)/\mathcal{F}_t\right]\\
&=&\int_0^tf^{(n)}dW(r)+\mathbb{E}\left[\sum_{t\leq t^{(n)}_j\leq t^{(n)}_{j+1}\leq s}f^{(n)}_j\Delta W_j/\mathcal{F}_t\right]\\
&=&\int_0^tf^{(n)}dW(r)+\sum_{t\leq t^{(n)}_j\leq t^{(n)}_{j+1}\leq s}\mathbb{E}[f^{(n)}_j\Delta W_j/\mathcal{F}_t]\\
&=&\int_0^tf^{(n)}dW(r)+\sum_{t\leq t^{(n)}_j\leq t^{(n)}_{j+1}\leq s}\mathbb{E}[\mathbb{E}[f^{(n)}_j\Delta W_j/\mathcal{F}_{t_j}]/\mathcal{F}_t]\\
&=&\int_0^tf^{(n)}dW(r)+\sum_{t\leq t^{(n)}_j\leq t^{(n)}_{j+1}\leq s}\mathbb{E}[f^{(n)}_j\mathbb{E}[\Delta W_j/\mathcal{F}_{t_j}]/\mathcal{F}_t]\\
&=&\int_0^tf^{(n)}dW(r),\hspace{0.2cm} \text{since} \hspace{0.2cm} E[\Delta W_j/\mathcal{F}_{t_j}]=\mathbb{E}[\Delta W_j]=0\\
&=&I_n(t, \omega).
\end{aligned}$$
**\[Generalisation\]**
Let $f(t, \omega)\in \mathbb{M}^2([0, T], \mathbb{R})$ for all $t$. Then the integral $$\begin{aligned}
M_t(\omega)=\int_0^tf(s, \omega)dW(s)
\end{aligned}$$ is a martingale with respect to $\mathcal{F}_t$ and $$\begin{aligned}
\mathbb{P}\left[\sup_{0\leq t\leq T}|M_t|\geq \lambda\right]\leq\dfrac{1}{\lambda^2}\mathbb{E}\left[\int_0^Tf^2(s, \omega)ds\right], \hspace{0.3cm} \forall\; \lambda >0
\end{aligned}$$
[@Oks].
### One dimensional Itô Formula
**\[1-dimensional Itô process\]**
Let $W_t$ be a $1$-dimensional Brownian motion on $(\Omega, \mathcal{F}, \mathbb{P})$. An Itô process (or Stochastic integral) is any stochastic process $X_t$ of the form $$\begin{aligned}
X_t=X_0+\int_0^tu(s,\omega)ds+\int_0^tv(s, \omega)dW(s),
\label{ch1Ito1}
\end{aligned}$$ where $u\in \mathbb{L}^1([0,T], \mathbb{R})$ and $v\in\mathbb{L}^2([0, T], \mathbb{R})$.
**\[ first $1$- dimensional Itô formula\]**
Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a complete probability space, $(W_t)_{t\in \mathbb{R}_{+}}$ a one-dimensional Brownian motion and $f : \mathbb{R}\longrightarrow \mathbb{R}$ such that $f$ is once derivable. If $(X_t)$ is any process of the form ,
then $f(X_t)$ is an Itô processes and $$\begin{aligned}
f(X_t)=f(X_0)+\int_0^t f'(X_s)u_sds+\dfrac{1}{2}\int_0^tf''(X_s)v_s^2ds+\int_0^tf'(X_s)v_sdW_s.
\end{aligned}$$
[@Oksa].
**\[ second $1$- dimensional Itô formula\]**
If in the previous proposition we consider $f :[0, \infty)\times\mathbb{R}\longrightarrow\mathbb{R}$ such that $f$ is once differentiable with respect to the first variable $t$ and twice differentiable with respect to the second variable $x$,
then $f(t, X_t)$ is an Itô process and $$\begin{aligned}
f(t, X_t)=f(0, X_0)+\int_0^t\dfrac{\partial f}{\partial t}(s,X_s)ds+\int_0^t\dfrac{\partial f}{\partial x}(s,X_s)u_sds+\int_0^t\dfrac{\partial f}{\partial x}(s,X_s)v_sdW_s+\dfrac{1}{2}\int_0^t\dfrac{\partial^2 f}{\partial x^2}(s,X_s)v_s^2ds,
\end{aligned}$$ or in its differential form : $$\begin{aligned}
df(t, X_t)=\dfrac{\partial f}{\partial t}(t, X_t)dt+\dfrac{\partial f}{\partial x}(t, X_t)dX_t+\dfrac{1}{2}\dfrac{\partial^2 f}{\partial x^2}(t, X_t)(dX_t)^2.
\end{aligned}$$ $(dX_t)^2=dX_tdX_t$ is computed according to the rules $$\begin{aligned}
dtdt=dW_tdt=dtdW_t=0,\hspace{1cm} dW_tdW_t=dt.
\end{aligned}$$
[@Oksa].
### Multi-dimensional Itô integral
**\[m-dimensional Brownian motion\]**[@Oks].
Let $W_1, \cdots W_m$ be $m$ Brownian motions. The random variable $W=(W_1, W_2, ..., W_m)$ is called $m$-dimensional Brownian motion. Let $\mathbb{L}^{n\times m}([0, T], \mathbb{R}^{n\times m})$ denotes the set of $n\times m$ matrices $v=[v_{ij}(t, \omega)]$, $1\leq i\leq n$, $1\leq j\leq m$. Where $v_{ij}(t, \omega)\in \mathbb{L}^2([0, T], \mathbb{R})$. $\int_0^tv_sdW_s$ denotes the Itô integral of $v$ with respect to the m-dimensional Brownian motion $W$. It can be written into its matrix form $$\begin{aligned}
\int_0^TvdW(s)=\int_0^T\left(\begin{array}{ccc}
v_{11}&\cdots& v_{1m}\\
.& &.\\
.& &.\\
.& &.\\
v_{n1}&\cdots &v_{nm}
\end{array}
\right)\left(\begin{array}{c}
dW_1(s)\\
.\\
.\\
.\\
dW_m(s)
\end{array}
\right),
\end{aligned}$$ which is a $n\times 1$ matrix (column vector) whose $i^{th}$ components are given by $$\begin{aligned}
\sum_{j=1}^m\int_0^Tv_{ij}(s, \omega)dW_j(s).
\end{aligned}$$
**\[$n$-dimensional Itô process\]**[@Oks].
Let $W$ be an $m$- Brownain motion and $v=[v_{i,j}, 1\leq i\leq n \hspace{0.2cm} 1\leq j\leq m]$ an element of $\mathbb{L}^{n\times m}([0, t], \mathbb{R}^{n\times m})$. Let $u=(u_i)_{i=1}^n$ such that $u_i\in\mathbb{L}^2([0, T])$ pour tout $1\leq i\leq n$.
The $n$-dimensional Itô process is any stochastic process of the form $$\begin{aligned}
dX(t)=udt+vdW(t),
\end{aligned}$$ which is a system of $n$ Itô process, where the $i^{th}$ process is given by : $$\begin{aligned}
dX_i(t)=u_idt+\sum_{j=1}^mv_{ij}dW_j(t).
\end{aligned}$$
**\[General Itô formula\]**
Let $ dX(t)=udt+vdW(t)$ be an $n$-dimensional Itô process. Let $g(t, x)=(g_1(t,x),..., g_p(t,x))$ be a function once differentiable with respect to $t$ and twice differentiable with respect to $x$.
Then the process $Y(t)=g(t, X(t))$ is also a $p$-dimensional Itô process, whose component $Y_k$ are given by : $$\begin{aligned}
Y_k(t)=\dfrac{\partial g_k}{\partial t}(t, X_t)dt+\sum_{i=1}^n\dfrac{\partial g_k}{\partial x_i}(t, X_t)dX_i+\dfrac{1}{2}+\sum_{i=1}^n\sum_{j=1}^n\dfrac{\partial^2 g_k}{\partial x_i\partial x_j}(t, X_t)dX_idX_j,
\end{aligned}$$ where $dW_idW_j=\delta_{ij}dt$ and $dW_idt=dtdW_i=0$.
[@Oks].
Stochastic process with jumps and Stochastic integral with jumps
----------------------------------------------------------------
1. A function $f : [0,T]\longrightarrow \mathbb{R}^n$ is said to be right continuous with left limit at $t\in[0,T]$ if $$\begin{aligned}
f(t^+) : =\lim_{s\longrightarrow t^+}f(s) \hspace{0.5cm}\text{and}\hspace{0.2cm}f(t^-) : =\lim_{s\longrightarrow t^-}f(s)\hspace{0.2cm}\text{exist}\hspace{0.2cm} \text{and} \hspace{0.5cm} f(t^+)=f(t).
\end{aligned}$$
2. A function $f : [0,T]\longrightarrow \mathbb{R}^n$ is said to be left continuous with right limit if $$\begin{aligned}
f(t^+) : =\lim_{s\longrightarrow t^+}f(s) \hspace{0.5cm}\text{and}\hspace{0.2cm}f(t^-) : =\lim_{s\longrightarrow t^-}f(s)\hspace{0.2cm}\text{exist}\hspace{0.2cm} \text{and} \hspace{0.5cm} f(t^-)=f(t).\end{aligned}$$
In the litterature, the french short forms “cádlág” and “cáglád” denote respectively functions which are right continous with left limit and left continous with right limit.
- If $f$ is right continous with left limit at $t$, then $\Delta f(t)=f(t)-f(t^-)$ is called the jump of $f$ at $t$.
- If $f$ is left continous with right limit at $t$, then $\Delta f(t)=f(t^+)-f(t)$ is called the jump of $f$ at $t$.
A stochastic process $X=(X_t)_{t\geq 0}$ is called jump process if the sample path $s\longmapsto X_s$ is left continuous (cáglág) or right continuous (cádlág) $\forall s\geq 0$.
**\[Lévy process\]**
A stochastic process $X=\{X_t,\hspace{0.3cm} t\geq 0\}$ is a Lévy process if the following conditions are fulfilled
1. The increments on disjoint time intervals are independent. That is for $0\leq t_0<t_1<...<t_n$ $\{X_{t_j}-X_{t_{j-1}}\hspace{0.3cm} 1\leq j\leq n\}$ are independent.
2. The increments of sample paths are stationary : $X_t-X_s\approx X_{t-s}$ for $0\leq t\leq s$.
3. The sample paths are right continuous with left limit.
The Brownian motion and the poisson process starting at $0$ are Lévy process.
[@Oksa] Let $\textbf{D}_{ucp}$ denote the space of càdlàg adapted process equipped with the topology of the uniform convergence in probability (ucp) on compact sets. ucp : $H_n\longrightarrow H$ if $\forall\; t\geq 0$ $\sup\limits_{0\leq s\leq t}|H_n(s)-H(s)|\longrightarrow 0$ in probability ($A_n\longrightarrow A$ in probability if $\forall\; \epsilon>0, \exists\; n_{\epsilon}\in \mathbb{N}$ such that $n>n_{\epsilon}\Longrightarrow \mathbb{P}(|A_n-A|>\epsilon)<\epsilon$).
In the sequel $\textbf{L}_{ucp}$ denote the space of adapted càdlàg processes (left continous with right limit) equiped with the ucp topology.
[@Oksa] Let $H$ be an elementary function. i.e there exist a partition
$0=t_0<t_1...<t_n=T$ such that $$\begin{aligned}
H=\sum_{j=0}^nH_j1_{[t_j, t_{j+1})},
\end{aligned}$$ where $H_j$ are $\mathcal{F}_{t_j}$-measurable. Let $(X_t)$ be a Lévy process. The stochastic integral $\int_0^tH(s)dX(s)$ is defined by $$\begin{aligned}
J_XH(t) : =\int_0^tH(s)dX(s): =\sum_{j=0}^n H_j(X(t_{j+1})-X(t_j)) \hspace{0.3cm} t\geq 0.\end{aligned}$$
[@Oksa] Let $X$ be a semimartingale, then the mapping $J_X$ can be extended to the continuous linear map $$\begin{aligned}
J_X : \textbf{L}_{ucp}\longrightarrow \textbf{D}_{ucp}.\end{aligned}$$
The above proposition allows us to define a stochastic integral of the form $$\begin{aligned}
\int_0^tH(s)dX(s),\end{aligned}$$ where $H\in \textbf{L}_{ucp}$.
[@Oksa] For $H\in \textbf{L}_{ucp}$ we define $\int_0^tH(s)dX(s)$ : $$\begin{aligned}
\int_0^tH(s)dX(s) : =\lim_{n\longrightarrow +\infty}\int_0^tH^{n}(s)dX(s),\end{aligned}$$ where $(H^{n})$ is a sequence of simple process converging to $H$.
Let $f\in\textbf{L}_{ucp}$ and $(\overline{N_t})$ be a compensated poisson process. The following holds
1. $\mathbb{E}\left(\int_0^tf(s)d\overline{N}(s)\right)=0$.
2. $\mathbb{E}\left(\int_0^tf(s)d\overline{N}(s)\right)^2=\lambda\int_0^t\mathbb{E}(f(s))^2ds$.
[@Oksa]
### Itô formula for jump process
**\[Itô jump-diffusion process\]**
The Itô jump-diffusion process is any process of the form $$\begin{aligned}
X_t=X_0+\int_0^ta(X_s)ds+\int_0^tb(X_s)dW_s+\int_0^tc(X_s)dN_s.
\label{ch1ito1}
\end{aligned}$$ The coefficient $a\in\mathbb{L}^1([0,T], \mathbb{R}^n)$ is called the drift coefficient, $b\in \mathbb{L}^2([0,T], \mathbb{R}^{n\times m})$ is called the diffusion coefficient and $c(X_s)\in\mathbb{L}^2([0,T], \mathbb{R}^n)$ is called the jump coefficient. $W$ is a $m$-dimentional Brownian motion and $N$ a one dimentional poisson process.
[@Oksa pp 6]**\[Itô formula for jump process\]**
If $(X_t)$ is a jump-diffusion process of the form and $f :[0,\infty)\longrightarrow \mathbb{R}^n$ any function twice derivable, then $f(X_t)$ is a jump-diffusion process and satisfies the following equation $$\begin{aligned}
f(X_t)=f(X_0)+\int_0^t\dfrac{\partial f}{\partial x}(X_s)a_sds+\dfrac{1}{2}\int_0^t\dfrac{\partial^2f}{\partial x^2}(X_s)b_s^2ds +\int_0^t\dfrac{\partial f}{\partial x}b_sdW_s+\int_0^t(f(X_{s^-}+c( X_s))-f(X_{s^-}))dN_s.
\end{aligned}$$
\[ch1Itoproduct\] **\[Itô’s lemma for product\]**
If $X_t$ and $Y_t$ are two Itô’s jump-diffusion process, Then $X_tY_t$ is an Itô jump-diffusion process and $$\begin{aligned}
d(X_tY_t)=Y_tdX_t+X_tdY_t+dX_tdY_t.\end{aligned}$$ $dX_1dX_2$ is called the Itô’s corrective term and it is computed according to the relations $$\begin{aligned}
dt.dt=dN_tdt=dtdN_t=dW_tdN_t=dN_tdW_t=0,\hspace{0.5cm}
dN_tdN_t=dN_t.\end{aligned}$$
[@Oksa]
After being familiar with some basic notions in probability theory and stochastic process, we are now ready to provide in the following chapter the proof of the existence and uniqueness of solution of SDEs with jumps under global Lipschitz conditions.
Existence and uniqueness of solution of the jump-diffusion Itô’s stochastic differential equations
==================================================================================================
In this chapter, we give the general formulation of the compensated stochastic differential equation (CSDE) with jumps which will be helpful to prove the existence and the uniqueness solutions of the stochastic differential equation with jump.
General formulation
-------------------
Along this work, $||.||$ denote the Frobenius matrix norm, $(\Omega, \mathcal{F}_t, \mathbb{P})$ denote a complete probability space. For all $x, y\in\mathbb{R}^n$, $\langle x,y\rangle=x_1y_1+\cdots, x_ny_n$ is the inner product. For all $a, b\in\mathbb{R}$,
$ a\vee b :=\max(a,b)$.
Throughout this work, we consider a jump diffusion Itô’s stochastic differential (SDEs) of the form $$\begin{aligned}
dX(t)=f(X(t^-))dt+g(X(t^-))dW(t)+h(X(t^-))dN(t),\hspace{0.5cm} X(0^-)=X_0,
\label{ch2jdi1}\end{aligned}$$ where $X_0$ is the initial condition, $X(t^-)=\lim\limits_{s\longrightarrow t^-}X(s)$, $W(t)$ is an $m$-dimensional Brownian motion and $N(t)$ is a $1$-dimensional poisson process with intensity $\lambda>0$. We assume $W_t$ and $N_t$ to be both $\mathcal{F}_t$-measurable. $f : \mathbb{R}^n\longrightarrow \mathbb{R}^n$, $g : \mathbb{R}^n \longrightarrow\mathbb{R}^{n\times m}$ and $h : \mathbb{R}^n\longrightarrow \mathbb{R}^n$.
Our aim in this chapter is to prove the existence and the uniqueness of solution of equation in the strong sense.
**\[Strong solution\]**
A stochastic process $X=\{X_t\}_{t\in[0,T]}$ is called strong solution of Itô jump-diffusion differential equation if :
1. $X_t$ is $\mathcal{F}_t-$measurable $\forall\; t\in[0,T]$.
2. $\mathbb{P}\left[\int_0^t|f(X_s,s)|ds<\infty\right]=\mathbb{P}\left[\int_0^t|g(X_s,s)|^2ds<\infty\right]
=\mathbb{P}\left[\int_0^t|h(X_s,s)|^2ds<\infty\right]=1$.
3. $X$ satisfies equation almost surely.
\[ch2assump\] Troughout this chapter, we make the following hypothesis :
There exist positive constants $L$ and $K$ such that for all $x, y\in\mathbb{R}^n$,
1. $\mathbb{E}||X(0)||^2<+\infty$ and $X(0)$ is independent of the Weiner process $W(t)$ and of the poisson process $N(t)$.
2. $f$, $g$ and $h$ satisfy the global Lipschitz condtion : $$\begin{aligned}
||f(x)-f(y)||^2\vee ||g(x)-g(y)||^2\vee ||h(x)-h(y)||^2\leq L||x-y||^2\end{aligned}$$
3. $f$, $g$ and $h$ satisfy the linear growth condition : $$\begin{aligned}
||f(x)||^2\vee ||g(x)||^2\vee ||h(x)||^2\leq K(1+||x||^2).\end{aligned}$$
The globaly Lipschitz condition implies the linear growth condition. So it is enough to make the assumptions only on the global Lipschitz condition.
A soluion $\{X_t\}$ of is said to be pathwise unique if any other solution $\overline{X}$ is a stochastilly indistinguishable from it, that is $\mathbb{P}\{X(t)=\overline{X}(t)\}=1$, $\forall\; t\in [0,T]$ almost surely.
In order to prove the existence and uniqueness solution of , it is useful to write in its compensated form.
### Compensated stochastic differential equation
(CSDE)
From the relation $\overline{N}(t)= N(t)-\lambda t$, we have $d\overline{N}(t)=dN_t-\lambda dt$. Substituting this latter relation in leads to : $$\begin{aligned}
dX(t)=f_{\lambda}(X(t^-))dt+g(X(t^-))dW(t)+h(X(t-))d\overline{N}(t),
\label{ch2jdi2}\end{aligned}$$ where $$\begin{aligned}
f_{\lambda}(x) :=f(x)+\lambda h(x).
\label{ch2jdi4}\end{aligned}$$ can be rewriten into its integral form $$\begin{aligned}
X(t)=X_0+\int_0^tf_{\lambda}(X(s^-))ds+\int_0^tg(X(s^-))dW(s)+\int_0^th(X(s^-))d\overline{N}(s),
\label{ch2jdi3}\end{aligned}$$
\[ch2lemma1\] If assumptions \[ch2assump\] are satisfied, then the function $f_{\lambda}$ satisfies the global Lipschitz and the linear growth conditions with constants $L_{\lambda}=(1+\lambda)^2L$ and $K_{\lambda}=(1+\lambda)^2K$ respectively.
1. Using the global lipschitz condition satisfied by $f$ and $h$, it follows from that: $$\begin{aligned}
||f_{\lambda}(x)-f_{\lambda}(y)||^2&=&||f(x)-f(y)+\lambda (h(x)-h(y))||^2\\
&\leq& \left(\sqrt{L}||x-y||+\lambda\sqrt{L}||x-y||\right)^2=(1+\lambda)^2L||x-y||^2.\end{aligned}$$
2. Along the same lines as above, we obtain the linear growth condition satisfies by $f_{\lambda}$.
Well-posedness problem
----------------------
Based on Lemma \[ch2lemma1\] and using the fact that equations and are equivalent, the existence and uniqueness of solution of equation is equivalent to the existence and uniqueness of solution of equation .
\[ch2th1\] If Assumptions \[ch2assump\] are fulfilled, then there exist a unique strong solution of equation .
In order to prove Theorem \[ch2th1\], we need the following lemma.
\[ch2lemma2\] let $X^0(t)=X_0(t), \hspace{0.5cm} \forall t\in [0,T]$ and $$\begin{aligned}
X^{n+1}(t)=X_0+\int_0^tf(X^n(s^-))ds+\int_0^tg(X^n(s^-))dW(s)+\int_0^th(X^n(s^-))d\overline{N}(s)ds,\end{aligned}$$ then $$\begin{aligned}
\mathbb{E}||X^{n+1}(t)-X^n(t)||^2 \leq \dfrac{(Mt)^{n+1}}{(n+1)!},
\label{lemI1}\end{aligned}$$
where $M$ is a positive constant depending on $ \lambda, K, L, X_0$.
: By induction
1. For $n=0$ $$\begin{aligned}
||X^1(t)-X^0(t)||^2&=&\left\|\int_0^tf_{\lambda}(X_0(s^-))ds+\int_0^tg(X_0(s^-))dW(s)+\int_0^th(X_0(s^-))d\overline{N}(s)\right\|^2\nonumber\\
&\leq& 3\left\|\int_0^tf_{\lambda}(X_0(s^-))ds\right\|^2+3\left\|\int_0^tg(X_0(s^-))dW(s)\right\|^2\nonumber\\
&+&3\left\|\int_0^th(X_0(s^-))d\overline{N}(s)\right\|^2.
\label{E1}\end{aligned}$$ Using Cauchy-Schwartz inequality and the linear growth condition, it follows that: $$\begin{aligned}
\mathbb{E}(I_1(t)):=\mathbb{E}\left(3\left\|\int_0^tf_{\lambda}(X_0(s^-))ds\right\|^2\right)&\leq & 3T\int_0^t\mathbb{E}||f_{\lambda}(X_0(s^-))||^2ds\nonumber\\
&\leq & 3TK_{\lambda}\int_0^t(1+E||X_0||^2)ds.
\label{EI1}\end{aligned}$$ From the martingale property of $\overline{N}(t)$ and the linear growth condition, it follows that : $$\begin{aligned}
\mathbb{E}(I_2(t)) := \mathbb{E}\left(3\left\|\int_0^th(X_0(s^-))d\tilde{N}(s)\right\|^2\right)&=&3\lambda\int_0^t\mathbb{E}||h(X_0(s^-))||^2ds\nonumber\\
&\leq & 3\lambda K\int_0^t(1+\mathbb{E}|X_0|^2)ds.
\label{EI2}\end{aligned}$$ Using the martingale property of $W(t)$ and the linear growth condition, it follows that : $$\begin{aligned}
\mathbb{E}(I_3(t)) := \mathbb{E}\left(3\left\|\int_0^tg(X_0(s^-))dW(s)\right\|^2\right)\leq 3K\int_0^t(1+\mathbb{E}||X_0||^2)ds.
\label{EI3}\end{aligned}$$ Taking expectation in both sides of and using estimations , and leads to : $$\begin{aligned}
\mathbb{E}||X^1(t)-X^0(t)||^2\leq (3TK_{\lambda}+3K+3\lambda K)\int_0^t(1+\mathbb{E}||X_0||^2)ds\leq Mt,
\label{E2}\end{aligned}$$ where $M=(3TK_{\lambda}+3K+3\lambda K) \vee (3TL_{\lambda} +3L+3\lambda L)$.
2. Let’s assume that inequality holds up to a certain rank $n\geq 0$. We have to show that it remains true for $n+1$. That is, we have to prove that $\mathbb{E}||X^{n+2}(t)-X^{n+1}(t)||^2\leq \dfrac{(Mt)^{n+2}}{(n+2)!}$. $$\begin{aligned}
X^{n+2}(t)&=&X_0+\int_0^tf_{\lambda}(X^{n+1}(s^-))ds+\int_0^tg(X^{n+1}(s^-))dW(s)+\int_0^th(X^{n+1}(s^-))d\overline{N}(s),\\
X^{n+1}(t)&=&X_0+\int_0^tf_{\lambda}(X^{n}(s^-))ds+\int_0^tg(X^{n}(s^-))dW(s)+\int_0^th(X^{n}(s^-))d\overline{N}(s)\end{aligned}$$ $$\begin{aligned}
||X^{n+2}(t)-X^{n+1}(t)||^2&=&\left\|\int_0^t\left[f_{\lambda}(X^{n+1}(s^-)-f_{\lambda}(X^n(s^-))\right]ds\right.\\
&+&\int_0^t[g(X^{n+1}(s^-))-g(X^n(s^-))]dW(s)\\
&+&\left.\int_0^t[h(X^{n+1}(s^-))-h(X^n(s^-))]d\overline{N}(s)\right\|^2.\end{aligned}$$ Using the inequality $(a+b+c)^2\leq 3a^2+3b^2+3c^2$ for all $a, b,c\in\mathbb{R}$, it follows that : $$\begin{aligned}
||X^{n+2}(t)-X^{n+1}(t)||^2&\leq & 3\left\|\int_0^t[f_{\lambda}(X^{n+1}(s^-))-f_{\lambda}(X^n(s^-))ds\right\|^2\\ &+&3\left\|\int_0^t[g(X^{n+1}(s^-))-g(X^n(s^-))dW\right\|^2\\
&+&3\left\|\int_0^t[h(X^{n+1}(s^-))-h(X^n(s^-))]d\overline{N}\right\|^2.\end{aligned}$$ Using the martingale properties of $W(s)$ and $\overline{N}(s)$ and the global Lipschitz condition satisfied by $f_{\lambda}$, $g$ and $h$, it follows that : $$\begin{aligned}
\mathbb{E}||X^{n+2}(t)-X^{n+1}(t)||^2&\leq&(3TL_{\lambda}+3L+3\lambda L)\int_0^t\mathbb{E}||X^{n+1}(s^-)-X^n(s^-)||^2ds.\end{aligned}$$ Using the hypothesis of induction, it follows that : $$\begin{aligned}
\mathbb{E}||X^{n+2}(t)-X^{n+1}(t)||^2&\leq &M\int_0^t\dfrac{(Ms)^{n+1}}{(n+1)!}ds=\dfrac{(Mt)^{n+2}}{(n+2)!}.\end{aligned}$$ This complete the proof of the lemma.
**\[Theorem \[ch2th1\]\]**
**** : Let $X_1$ and $X_2$ be two solutions of . Then : $$\begin{aligned}
X_1(t)=X_0+\int_0^tf_{\lambda}(X_1(s^-))ds+\int_0^tg(X_1(s^-))dW(s)+\int_0^th(X_1(s^-))d\overline{N}(s),\end{aligned}$$ $$\begin{aligned}
X_2(t)=X_0+\int_0^tf_{\lambda}(X_2(s^-))ds+\int_0^tg(X_2(s^-))dW(s)+\int_0^th(X_2(s^-))d\overline{N}(s).\end{aligned}$$ Therefore, $$\begin{aligned}
||X_1(t)-X_2(t)||^2&=&\left\|\int_0^t\left[f_{\lambda}(X_1(s^-))-f_{\lambda}(X_2(s^-))\right]ds +\int_0^t[g(X_1(s^-))-g(X_2(s^-))]dW(s)\nonumber\right.\\
&+&\left.\int_0^t[h(X_1(s^-))-h(X_2(s^-))]d\overline{N}(s)\right\|^2\nonumber\\
&\leq& 3\left\|\int_0^t[f_{\lambda}(X_1(s^-))-f_{\lambda}(X_2(s^-))]ds\right\|^2+3\left\|\int_0^t[g(X_1(s^-))-g(X_2(s^-))]dW(s)\right\|^2\nonumber\\
&+&3\left\|\int_0^t[h(X_1(s^-))-h(X_2(s^-))]d\overline{N}(s)\right\|^2.
\label{tI1}\end{aligned}$$ Using Cauchy-Schwartz inequality and the globaly Lipschitz condition, it follows that : $$\begin{aligned}
\mathbb{E}(I_1(t)) &:=&\mathbb{E}\left(3\left\|\int_0^t[f_{\lambda}(X_1(s^-))-f_{\lambda}(X_2(s^-))]ds\right\|^2\right)\nonumber\\
&\leq& 3t\int_0^t\mathbb{E}\left\|f_{\lambda}(X_1(s^-))-f_{\lambda}(X_2(s^-))\right\|^2\nonumber\\
&\leq& 3tL_{\lambda}\int_0^t\mathbb{E}||X_1(s^-)-X_2(s^-)||^2ds.
\label{I1}\end{aligned}$$ Using the martingale property of $W(s)$ and the global Lipschitz condition, it follows that : $$\begin{aligned}
\mathbb{E}(I_2(t))&:=& \mathbb{E}\left(3\left\|\int_0^t[g(X_1(s^-))-g(X_2(s^-))]dW(s)\right\|^2\right)\nonumber\\
&\leq& 3\int_0^t\mathbb{E}\left\|g(X_1(s^-))-g(X_2(s^-))\right\|^2ds\nonumber\\
&\leq& 3L\int_0^t\mathbb{E}\left\|X_1(s^-)-X_2(s^-)\right\|^2ds.
\label{I2}\end{aligned}$$ Along the same lines as above, we obtain : $$\begin{aligned}
\mathbb{E}(I_3(t))&:=& \mathbb{E}\left(3\left\|\int_0^t[h(X_1(s^-))-h(X_2(s^-))]d\overline{N}(s)\right\|^2\right)\nonumber\\
&\leq&
3L\int_0^tE\left\|X_1(s^-)-X_2(s^-)\right\|^2ds.
\label{I3}\end{aligned}$$ Taking expectation in both sides of and using estimations , and leads to : $$\begin{aligned}
\mathbb{E}||X_1(t)-X_2(t)||^2\leq (3tL_{\lambda}+3L+3\lambda L)\int_0^t\mathbb{E}||X_1(s^-)-X_2(s^-)||^2ds.
\label{tI2}\end{aligned}$$ Applying Gronwall lemma (contonuous form) to inequality leads to : $$\begin{aligned}
\mathbb{E}||X_1(t)-X_2(t)||^2=0, \hspace{0.5cm} \forall t\;\in[0,T].\end{aligned}$$ It follows from Markov’s inequality that : $$\begin{aligned}
\forall\; a>0,\hspace{0.5cm} \mathbb{P}(||X_1-X_2||^2>a)=0.\end{aligned}$$ Therefore, $$\begin{aligned}
\mathbb{P}\left(\{X_1(t)=X_2(t), \hspace{0.2cm} t\in[0,T]\}\right)=1 \hspace{0.2cm} a.s.\end{aligned}$$
****:
From the sequence $X^n(t)$ defined in Lemma \[ch2lemma2\], it follows that : $$\begin{aligned}
||X^{n+1}(t)-X^{n}(t)||^2&=&\left\|\int_0^t[f_{\lambda}(X^n(s^-))-f_{\lambda}(X^{n-1}(s^-))]ds+\int_0^t[g(X^n(s^-))-g(X^{n-1}(s^-))]dW(s)\right.\nonumber\\
&+&\left.\int_0^t[h(X^n(s^-))-h(X^{n-1}(s^-))]d\overline{N}(s)\right\|^2.
\label{ch2I3}\end{aligned}$$ Taking expectation and the supremum in the both sides of inequality and using inequality $(a+b+c)^2\leq 3a^2+3b^2+3c^2$ for all $a, b, c \in\mathbb{R}$ leads to : $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t\leq T}||X^{n+1}(t)-X^n(t)||^2\right)&\leq& 3T\sup_{0\leq t\leq T}\left(\mathbb{E}\int_0^t||f_{\lambda}(X^n(s^-)-f_{\lambda}(s^-)||^2ds\right)\nonumber\\
&+&3\mathbb{E}\left(\sup_{0\leq t\leq T}M_1(t)\right)+3\mathbb{E}\left(\sup_{0\leq t\leq T}M_2(t)\right),
\label{ch2I4}\end{aligned}$$ where $$\begin{aligned}
M_1(t)&=&\left\|\int_0^t[g(X^n(t))-g(X^{n-1}(t))]dW(s)\right\|^2\\
M_2(t)&=&\left\|\int_0^t[h(X^n(t))-h(X^{n-1}(t))]d\overline{N}(s)\right\|^2.\end{aligned}$$ Using the global Lipschitz condition, it follows that $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t\leq T}\int_0^t||f_{\lambda }(X^n(s^-))-f_{\lambda}(X^{n-1}(s^-))||^2ds\right)\leq L_{\lambda}\int_0^T\mathbb{E}||X^n(s^-)-X^{n-1}(s^-)||^2ds.
\label{ch2M0}\end{aligned}$$ Using respectively Doop’s maximal inequality, martingale property of $W(s)$ and global Lipschitz condition satisfied by $g$, it follows that : $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t\leq T}M_1(t)\right)\leq 4M_1(t)&=&4\mathbb{E}\left\|\int_0^T[g(X^n(s^-))-X^{n-1}(s^-))]dW(s)\right\|^2\nonumber\\
&=&4\int_0^T\mathbb{E}||g^(X^n(s^-))-g(X^{n-1}(s^-))||^2ds\nonumber\\
&\leq& 4L\int_0^T\mathbb{E}||X^n(s^-)-X^{n-1}(s^-)||^2ds.
\label{ch2M1}\end{aligned}$$ Along the same lines as above, we obtain $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t\leq T}M_2(t)\right)\leq 4\lambda L\int_0^T\mathbb{E}||X^n(s^-)-X^{n-1}(s^-)||^2ds.
\label{ch2M2}\end{aligned}$$ Inserting , and in leads to : $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t\leq T}||X^{n+1}(t)-X^n(t)||^2\right)\leq (3L_{\lambda}+12L+12\lambda L)\int_0^T\mathbb{E}||X^n(s^-)-X^{n-1}(s^-)||^2ds.\end{aligned}$$ Using Lemma \[ch2lemma2\], it follows that : $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t\leq T}||X^{n+1}(t)-X^n(t)||^2\right)\leq C\int_0^T\dfrac{(Ms^-)^n}{n!}ds=C\dfrac{(MT^-)^{n+1}}{(n+1)!},\end{aligned}$$ where $C=3L_{\lambda}+12L+12\lambda L$.
It follows from Doop’s and Markov’s inequalities that : $$\begin{aligned}
\mathbb{P}\left(\sup_{0\leq t\leq T}||X^{n+1}(t)-X^n(t)||>\dfrac{1}{2^{n+1}}\right)\leq \dfrac{\mathbb{E}||X^{n+1}(t)-X^n(t)||^2}{\left(\dfrac{1}{(2^{n+1})}\right)^2}\leq C\dfrac{(2^2MT)^{n+1}}{(n+1)!}.\end{aligned}$$ Moreover, $$\begin{aligned}
\sum_{n=0}^{\infty}\dfrac{(2^2MT)^{n+1}}{(n+1)!}=e^{2^2MT}-1<\infty.\end{aligned}$$
Using Borel Cantelli’s Lemma, it follows that : $$\begin{aligned}
\mathbb{P}\left(\sup_{0\leq t\leq T}||X^{n+1}(t)-X^n(t)||>\dfrac{1}{2^{n+1}}\right)=0, \hspace{0.5cm} \text{almost surely}.\end{aligned}$$ Therefore, for almost every $\omega\hspace{0.1cm}\in\hspace{0.1cm} \Omega, \hspace{0.1cm}\exists\hspace{0.1cm} n_0(\omega)$ such that $$\begin{aligned}
\sup_{0\leq t\leq T}||X^{n+1}(t)-X^n(t)||\leq\dfrac{1}{2^{n+1}},\hspace{0.2cm} \forall\: n\geq n_0(\omega).\end{aligned}$$ Therefore $X^n$ converge uniformly on $[0,T]$. Let $X$ be its limit.
Futhermore, since $X^n_t$ is continuous and $\mathcal{F}_t$-measurable for all $n\in\mathbb{N}$, it follows that $X(t)$ is continuous and $\mathcal{F}_t$- measurable. It remains to prove that $X$ is a solution of equation \[ch2jdi1\].
One can see that $(X^n)$ converges in $\mathbb{L}^2([0,T]\times \Omega, \mathbb{R}^n)$. Indeed, $$\begin{aligned}
||X^m_t-X^n_t||^2_{\mathbb{L}^2}\leq \sum_{k=n}^{m-1}||X^{k+1}_t-X^k_t||_{\mathbb{L}^2}
\leq \sum_{k=n}^{\infty}\dfrac{(MT)^{k+1}}{(k+1)!}\longrightarrow 0, \hspace{0.5cm}\text{as}\hspace{0.5cm} n\longrightarrow \infty.
\end{aligned}$$ Therefore, $(X^n)$ is a Cauchy sequence in a Banach $\mathbb{L}^2([0,T]\times \Omega, \mathbb{R}^n)$, so its converges to $X$.
Using Fatou’s lemma, it follows that : $$\begin{aligned}
\mathbb{E}\left[\int_0^T||X_t-X^n_t||^2dt\right]\leq\liminf_{m\longrightarrow +\infty}\mathbb{E}\left[\int_0^T||X^m_t-X^n_t||^2dt\right]\longrightarrow 0,\hspace{0.3cm}\text{when} \hspace{0.3cm}m\longrightarrow +\infty.
\end{aligned}$$ Using the global Lipschitz condition and the Ito isometry, it follows that $$\begin{aligned}
\mathbb{E}\left\|\int_0^t[g(X_s)-g(X^n_s)]dW_s\right\|^2\leq L\mathbb{E}\int_0^t||X_s-X^n_s||^2ds\longrightarrow 0\hspace{0.3cm}\text{when}\hspace{0.3cm}n\longrightarrow+\infty.
\end{aligned}$$ So we have the following convergence in $\mathbb{L}^2([0,T]\times\mathbb{R}^n)$ $$\begin{aligned}
\int_0^tg(X^n_s)dW_s\longrightarrow \int_0^tg(X_s)dW_s.
\end{aligned}$$ Along the same lines as above, the following holds convergence holds in $\mathbb{L}^2([0,T]\times\mathbb{R}^n)$ $$\begin{aligned}
\int_0^tg(X^n_s)d\overline{N}_s\longrightarrow \int_0^tg(X_s)d\overline{N}_s.
\end{aligned}$$ Using Holder inequality and the global Lipschitz condition, it follows that : $$\begin{aligned}
\mathbb{E}\left\|\int_0^t[f_{\lambda}(X_s)-f_{\lambda}(X^n_s)]ds\right\|^2\leq L_{\lambda}\mathbb{E}\int_0^t||X_s-X^n_s||^2ds\longrightarrow 0\hspace{0.3cm}\text{when}\hspace{0.3cm}n\longrightarrow+\infty.\end{aligned}$$ So the following convergence holds in $\mathbb{L}^2([0,T]\times\mathbb{R}^n)$ $$\begin{aligned}
\int_0^tf_{\lambda}(X^n_s)d\overline{N}_s\longrightarrow \int_0^tf_{\lambda}(X_s)d\overline{N}_s.
\end{aligned}$$ Therefore, taking the limit in the sense of $\mathbb{L}^2([0,T]\times\mathbb{R}^n)$ in the both sides of the following equality : $$\begin{aligned}
X^{n+1}(t)=X_0+\int_0^tf(X^n(s^-))ds+\int_0^tg(X^n(s^-))dW(s)+\int_0^th(X^n(s^-))d\overline{N}(s)ds\end{aligned}$$ leads to : $$\begin{aligned}
X(t)=X_0+\int_0^tf(X(s^-))ds+\int_0^tg(X(s^-))dW(s)+\int_0^th(X(s^-))d\overline{N}(s)ds.\end{aligned}$$ So $X(t)$ is a strong solution of . This complete the proof of Theorem \[ch2th1\].
Generally, analitycal solutions of SDEs are unknows. Knowing that the exact solution exist, one tool to approach it, is the numerical resolution. In the following chapters, we provide some numerical schemes for SDEs with jumps.
Strong convergence and stabilty of the compensated stochastic theta methods
============================================================================
Our goal in this chapter is to prove the strong convergence of the compensated stochastic theta method (CSTM) and to analyse the stability behavior of both stochastic theta method(STM) and CSTM under global Lipschitz conditions. The strong convergence and stability of STM for SDEs with jumps has been investigated in [@Desmond1], while the strong convergence and stability of the CSTM for SDEs with jumps has been investigated in [@Xia1]. Most results presented in this chapter are from [@Desmond1] and [@Xia1]. In the following section, we recall the theta method which will be used to introduce the STM and the CSTM.
Theta Method
------------
Let’s consider the following deterministic differential equation $\left\{\begin{array}{ll}
u'(t)=f(t, u(t))\\
u(t_0)=u_0,
\end{array}
\right.$
which can be writen into its following integral form : $$\begin{aligned}
u(t)=u_0+\int_{t_0}^tf(s,u(s))ds.
\label{ch3Euler1}\end{aligned}$$
### Euler explicit method
This method use the following approximation : $$\begin{aligned}
\int_a^bf(s)ds\simeq (b-a)f(a).
\end{aligned}$$ So for a constant step $\Delta t$, the Euler explicit approximation of is given by : $$\begin{aligned}
u_{k+1}=u_k+\Delta tf(t_k,u_k),\end{aligned}$$ where $u_k :=u(t_k)$.
### Euler implicit method
This method use the following approximation $$\begin{aligned}
\int_a^bf(s)\simeq (b-a)g(b).\end{aligned}$$ Therefore, for a constant step $\Delta t$, the Euler implicit approximation of is : $$\begin{aligned}
u_{k+1}=u_k+\Delta tf(t_k,u_{k+1}).\end{aligned}$$
### Theta Euler method
In order to have a better approximation of the integral, we can take a convex combinaison of Euler explict and Euler implicit method. So we have the following approximation $$\begin{aligned}
\int_a^bf(s)ds\simeq (b-a)[(1-\theta)f(a)+\theta f(b)],\end{aligned}$$ where $\theta$ is a constant satisfying $0\leq \theta\leq 1$.
Hence, for a constant step $\Delta t$, the Euler theta approximation of is : $$\begin{aligned}
u_{k+1}=u_k+\Delta t[(1-\theta)f(t_k, u_k)+\theta f(t_k,u_{k+1}) ].\end{aligned}$$ For $\theta=1$, the Euler theta method is called Euler backward method, which is also the Euler implicit method.
### Stochastic theta method and compensated stochastic theta method
(STM and CSTM)
In order to have an approximate solution of equation , we use the theta Euler method for the deterministic integral and the Euler explicit method for the two random parts. So we have the following approximate solution of called Stochastic theta method (STM) : $$\begin{aligned}
Y_{n+1}=Y_n+(1-\theta)f(Y_n)\Delta t+\theta f(Y_{n+1})\Delta t+g(Y_n)\Delta W_n+h(Y_n)\Delta N_n,
\label{ch2approxi1}\end{aligned}$$ where $Y_n :=X(t_n)$, $\Delta W_n :=W(t_{n+1})-W(t_n)$ and $ \Delta N_n := N(t_{n+1})-N(t_n)$
Applying the same rules as for the STM to equation leads to : $$\begin{aligned}
Y_{n+1}=Y_n+(1-\theta)f_{\lambda}(Y_n)\Delta t+\theta f_{\lambda}(Y_{n+1})\Delta t +g(Y_n)\Delta W_n+h(Y_n)\Delta\overline{N}_n,
\label{ch3comp1}\end{aligned}$$ where $$\begin{aligned}
f_{\lambda}(x)=f(x)+\lambda h(x).\end{aligned}$$
The numerical approximation is called compensated stochastic theta method (CSTM).
Strong convergence of the CSTM on a finite time interval \[0,T\]
----------------------------------------------------------------
In this section, we prove the strong convergence of order $0.5$ of the CSTM. Troughout, $T$ is a fixed constant. For $t\in [t_n,t_{n+1})$ we define the continuous time approximation of as follows : $$\begin{aligned}
\overline{Y}(t):=Y_n +(1-\theta)(t-t_n)f_{\lambda}(Y_n)+\theta(t-t_n)f_{\lambda}(Y_{n+1})+g(Y_n)\Delta W_n(t)+h(Y_n)\Delta\overline{N}_n(t),
\label{ch3contapproxi1}\end{aligned}$$ where $\Delta W_n(t):=W(t)-W(t_n)$, $\Delta\overline{N}_n(t):=\overline{N}(t)-\overline{N}(t_n)$
The continuous approximation can be writen into its following integral form : $$\begin{aligned}
\overline{Y}(t)&=&Y_0+\int_0^t\left[(1-\theta)f_{\lambda}(Y(s))+\theta f_{\lambda}(Y(s+\Delta t))\right]ds+\int_0^tg(Y(s))dW(s)\nonumber\\
&+&\int_0^th(Y(s))d\overline{N}(s),
\label{ch3contapproxi2}\end{aligned}$$ where $$\begin{aligned}
Y(s):=\sum_{n=0}^{\infty}\mathbf{1}_{\{t_n\leq s<t_{n+1}\}}Y_n.\end{aligned}$$ It follows from that $\overline{Y}(t_n)=Y_n$. In others words, $\overline{Y}(t)$ and $Y_n$ coincide at the grid points. The main result of this section is formulated in the following theorem.
\[ch3th1\] Under Assumptions \[ch2assump\], the continuous time approximation solution $\overline{Y}(t)$ given by converges to the true solution $X(t)$ of in the mean square sense. More precisely, $$\begin{aligned}
\mathbb{E}\left(\sup_{0\leq t \leq T}||\overline{Y}(t)-X(t)||^2\right)\leq C(1+\mathbb{E}||X_0||^2)\Delta t,\end{aligned}$$ where $C$ is a positive constant independent of the stepsize $\Delta t$.
In order to prove Theorem \[ch3th1\], we need the following two lemmas.
\[ch3lemma1\] Under Assumptions \[ch2assump\], there exist a fixed constant $\Delta t_0$ such that for any stepsize $\Delta t$ satisfying $0<\Delta t<\Delta t_0<\dfrac{1}{K_{\lambda}+1}$, the following bound of the numerical solution holds $$\begin{aligned}
\sup_{0\leq n\Delta t\leq T}\mathbb{E}||Y_n||^2\leq C_1(1+\mathbb{E}||X_0||^2),\end{aligned}$$ where $C_1$ is a positive constant independent of $\Delta t$.
From , it follows that : $$\begin{aligned}
||Y_{n+1}-\theta \Delta t f_{\lambda}(Y_{n+1})||^2=||Y_n+(1-\theta)f_{\lambda}(Y_n)\Delta t+g(Y_n)\Delta W_n+h(Y_n)\Delta \overline{N}_n||^2.
\label{ch3eq1}\end{aligned}$$ Taking expectation in both sides of leads to : $$\begin{aligned}
\mathbb{E}||Y_{n+1}-\Delta t\theta f_{\lambda}(Y_{n+1})||^2&=&\mathbb{E}||Y_n||^2+(1-\theta)^2(\Delta t)^2\mathbb{E}||f_{\lambda}(Y_n)||^2+\mathbb{E}||g(Y_n)\Delta W_n||^2+\mathbb{E}||h(Y_n)\Delta\overline{N}_n||^2\nonumber\\
&+&2\mathbb{E}\langle Y_n,(1-\theta)f_{\lambda}(Y_n)\Delta t\rangle.
\label{ch3eq2}\end{aligned}$$ Since $W$ is a Brwonian motion, $\Delta W_n=W_{t_{n+1}}-W_{t_n}\leftrightsquigarrow \mathcal{N}(0,t_{n+1}-t_n)$. So $\mathbb{E}(\Delta W_n)=0$.
Using the properties $\mathbb{E}(\Delta W_n)=0$ and $\mathbb{E}(\Delta \overline{N}_n)=0$, we have $$\begin{aligned}
\mathbb{E}\langle Y_n,g(Y_n)\Delta W_n\rangle=\mathbb{E}\langle f_{\lambda}(Y_n)\Delta, g(Y_n)\Delta W_n\rangle=\mathbb{E} \langle f_{\lambda}(Y_n)\Delta t, h(Y_n)\Delta \overline{N}_n\rangle=0.\end{aligned}$$ The martingale properties of $\Delta W_n$ and $\Delta \overline{N}_n$ leads to : $$\begin{aligned}
\mathbb{E}||g(Y_n)\Delta W_n||^2=\mathbb{E}||g(Y_n)||^2\Delta t\hspace{0.5cm} \text{and}\hspace{0.5cm}
\mathbb{E}||h(Y_n)\Delta \overline{N}_n||^2=\lambda \Delta t\mathbb{E}||h(Y_n)||^2.\end{aligned}$$ Hence equality becomes : $$\begin{aligned}
\mathbb{E}||Y_{n+1}-\Delta t\theta f_{\lambda}(Y_{n+1})||^2&=&\mathbb{E}||Y_n||^2+(1-\theta)^2(\Delta t)^2\mathbb{E}||f_{\lambda}(Y_n)||^2+\mathbb{E}||g(Y_n)||^2\Delta t\nonumber\\
&+&2(1-\theta)\Delta t\mathbb{E}\langle Y_n, f_{\lambda}(Y_n)\rangle+\lambda \Delta t\mathbb{E}||h(Y_n)||^2.
\label{ch3eq3}\end{aligned}$$ Using Cauchy-Schwartz inequality and the linear growth condition, it follows that : $$\begin{aligned}
\mathbb{E}\langle Y_n, f_{\lambda}(Y_n)\rangle=\mathbb{E}||Y_nf_{\lambda}(Y_n)||&\leq&\sqrt{\mathbb{E}||Y_n||^2\mathbb{E}||f_{\lambda}(Y_n)||^2}\\
&\leq& \dfrac{1}{2}\mathbb{E}||Y_n||^2+\dfrac{1}{2}\mathbb{E}||f_{\lambda}(Y_n)||^2\\
&\leq& \dfrac{K_{\lambda}}{2}+\dfrac{1}{2}(1+K_{\lambda})\mathbb{E}||Y_n||^2.\end{aligned}$$ By the same arguments as above, it follows that : $$\begin{aligned}
\mathbb{E}\langle Y_{n+1}, f_{\lambda}(Y_{n+1})\rangle\leq \dfrac{K_{\lambda}}{2}+\dfrac{1}{2}(1+K_{\lambda})\mathbb{E}||Y_{n+1}||^2.\end{aligned}$$ Since $\theta\in [0,1]$ and $\Delta t\in]0,1[$, it follows from that : $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2&\leq& 2\Delta t\mathbb{E}|\langle Y_{n+1},f_{\lambda}(Y_{n+1})\rangle|+\mathbb{E}||Y_n||^2+\Delta t \mathbb{E}||f_{\lambda}(Y_n)||^2+\Delta t\mathbb{E}||g(Y_n)||^2\\
&+& 2\Delta t\mathbb{E}Y_n, f_{\lambda}(Y_n)+\lambda\Delta t\mathbb{E}||h(Y_n)||^2\\
&\leq&2\Delta t\left[\dfrac{1}{2}(1+K_{\lambda})\mathbb{E}||Y_{n+1}||^2+\dfrac{1}{2}K_{\lambda}\right]+\mathbb{E}||Y_n||^2+\Delta tK_{\lambda}(1+\mathbb{E}||Y_n||^2)+\Delta tK(1+\mathbb{E}||Y_n||^2)\\
&+&2\Delta t\left[\dfrac{1}{2}(1+K_{\lambda})\mathbb{E}||Y_n||^2+\dfrac{1}{2}K_{\lambda}\right]+\lambda \Delta tK(1+\mathbb{E}||Y_n||^2).\end{aligned}$$ Therefore, from the above inequality the following holds : $$\begin{aligned}
(1-\Delta t-\Delta tK_{\lambda})\mathbb{E}||Y_{n+1}||^2&\leq &(1+\Delta tK_{\lambda}+\Delta tK+\Delta t+\Delta tK_{\lambda}+\lambda \Delta tK)\mathbb{E}||Y_n||^2\\
&+&\Delta tK_{\lambda}+\Delta tK_{\lambda}+\Delta tK+\Delta tK_{\lambda}+\lambda \Delta tK\\
&\leq & (1+2\Delta tK_{\lambda}+\Delta tK+\Delta t+\Delta t\lambda K)\mathbb{E}||Y_n||^2+3\Delta t K_{\lambda}+\Delta tK+\lambda\Delta t K.\end{aligned}$$ Then it follows from the previous inequality that : $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2\leq\left(1+\dfrac{3K_{\lambda}\Delta t+K\Delta t+\lambda K\Delta t+2\Delta t}{1-\Delta t-K_{\lambda}\Delta t}\right)\mathbb{E}||Y_n||^2+\dfrac{3K_{\lambda}\Delta t+K\Delta t+\lambda K\Delta t+2\Delta t}{1-\Delta t-K_{\lambda}\Delta t}.\end{aligned}$$ Since $\Delta t<\Delta t_0<\dfrac{1}{K_{\lambda}+1}$, we have $1-\Delta t-K_{\lambda}\Delta t>1-\Delta t_0-K_{\lambda}\Delta t_0>0$ and then $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2\leq\left(1+\dfrac{3K_{\lambda}+K+\lambda K t+2 }{1-\Delta t_0-K_{\lambda}\Delta t_0}\Delta t\right)\mathbb{E}||Y_n||^2+\dfrac{3K_{\lambda}+K+\lambda K t+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}\Delta t_0.\end{aligned}$$ In the short form, we have $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2\leq (1+A)\mathbb{E}||Y_{n}||^2+B,
\label{ch3eq4}\end{aligned}$$ where $$\begin{aligned}
A=\dfrac{3K_{\lambda}+K+\lambda K+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}\Delta t\hspace{0.5cm}\text{and}\hspace{0.5cm}
B=\dfrac{3K_{\lambda}+K+\lambda K+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}\Delta t_0.\end{aligned}$$ Applying Gronwall lemma (discrete form) to leads to : $$\begin{aligned}
\mathbb{E}||Y_n||^2<e^{nA}\mathbb{E}||X_0||^2+B\dfrac{e^{nA}-1}{e^A-1},
\label{ch3eq5}\end{aligned}$$ $$\begin{aligned}
nA=\dfrac{3K_{\lambda}+K+\lambda K+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}n\Delta t \leq \dfrac{3K_{\lambda}+K+\lambda K+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}T,\hspace{0.2cm} \text{since}\hspace{0.2cm} n\Delta t\leq T.\end{aligned}$$ Therefore, it follows from that : $$\begin{aligned}
\mathbb{E}||Y_n||^2\leq e^C\mathbb{E}||X_0||^2+B\dfrac{e^C-1}{e^D-1},
\label{ch3eq6}\end{aligned}$$ where $$\begin{aligned}
C=\dfrac{3K_{\lambda}+K+\lambda K+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}T\hspace{0.2cm}\text{and}\hspace{0.5cm}
D=\dfrac{3K_{\lambda}+K+\lambda K+2}{1-\Delta t_0-K_{\lambda}\Delta t_0}\Delta t_0.\end{aligned}$$ It is straightforward to see that $B$, $C$ and $D$ are independents of $\Delta t$.
can be rewritten into the following appropriate form : $$\begin{aligned}
\mathbb{E}||Y_n||^2\leq C_1(1+\mathbb{E}||X_0||^2)\hspace{0.3cm} \hspace{0.2cm} C_1=\max\left(e^C, B\dfrac{e^C-1}{e^D-1}\right).\end{aligned}$$ This complete the proof of the lemma.
\[ch3lemma2\] If the conditions of Lemma \[ch3lemma1\] are satisfied, then there exist a positive constant $C_2$ independent of $\Delta t$ such that for $s\in [t_n, t_{n+1})$ $$\begin{aligned}
\mathbb{E}||\overline{Y}(s)-Y(s)||^2\vee \mathbb{E}||\overline{Y}(s)-Y(s+\Delta t)||^2\leq C_2(1+\mathbb{E}||X_0||^2)\Delta t.\end{aligned}$$
1. The continous interpolation of the numerical solution is given by $$\begin{aligned}
\overline{Y}(s)=Y_n+(1-\theta)(s-t_n)f_{\lambda}(Y_n)+\theta(s-t_n)f_{\lambda}(Y_{n+1})+g(Y_n)\Delta W_n(s)+h(Y_n)\Delta \overline{N}(s),\end{aligned}$$ where $$\begin{aligned}
Y(s)=\sum_{n=0}^{\infty}1_{\{t_n\leq s<t_{n+1}\}}Y_n.\end{aligned}$$ For $s\in[t_n, t_{n+1})$, we have $Y(s)=Y_n$. Then, we have the following equality : $$\begin{aligned}
\overline{Y}(s)-Y(s)&=&(1-\theta)(s-t_n)f_{\lambda}(Y_n)+\theta(s-t_n)f_{\lambda}(Y_{n+1})+g(Y_n)\Delta W_n(s)\nonumber\\
&+&h(Y_n)\Delta\overline{N}(s).
\label{ch3eq7}\end{aligned}$$ By squaring both sides of and taking expectation, using the martingale properties of $\Delta W_n$ and $\Delta\overline{N}_n$ leads to : $$\begin{aligned}
\mathbb{E}||\overline{Y}(s)-Y(s)||^2&\leq &3(1-\theta)^2(s-t_n)^2\mathbb{E}||f_{\lambda}(Y_n)||^2+3\theta^2(s-t_n)^2\mathbb{E}|f_{\lambda}(Y_{n+1})|^2+3\mathbb{E}||g(Y_n)\Delta W_n(s)||^2\nonumber\\
&+&3\mathbb{E}||h(Y_n)\Delta\overline{N}_n(s)||^2\nonumber\\
&\leq & 3(1-\theta)^2\Delta t^2\mathbb{E}||f_{\lambda}(Y_n)||^2+3\theta ^2\Delta t^2\mathbb{E}||f_{\lambda}(Y_{n+1})||^2+3\Delta t\mathbb{E}||g(Y_n)||^2\nonumber\\
&+&3\lambda\Delta t\mathbb{E}||h(Y_n)||^2.
\label{ch3eq8}\end{aligned}$$ By using the linear growth condition and the fact that $\theta\in[0,1]$, it follows from that : $$\begin{aligned}
\mathbb{E}||\overline{Y}(s)-Y(s)||^2&\leq &3\Delta tK_{\lambda}(1+\mathbb{E}||Y_n||^2)+3\Delta K_{\lambda}(1+\mathbb{E}||Y_{n+1}||^2)\nonumber\\
&+&3\Delta tK(1+\mathbb{E}||Y_n||^2)+3\Delta t\lambda K(1+\mathbb{E}||Y_n||^2).
\label{ch3eq9}\end{aligned}$$ Now by application of Lemma \[ch3lemma2\] to , it follows that there exist a constant $C_1>0$ independent of $\Delta t$ such that : $$\begin{aligned}
\mathbb{E}||\overline{Y}(s)-Y(s)||^2\leq C_1(1+\mathbb{E}||X_0||^2)\Delta t.\end{aligned}$$
2. For $s\in[t_n, t_{n+1})$, $s+\Delta t \in[t_{n+1}, t_{n+2})$ and then $Y(s+\Delta t)=Y_{n+1}$. So it follows from that : $$\begin{aligned}
Y(s+\Delta t)&=&Y_{n+1}\\
&=&Y_n+(1-\theta)(t_{n+1}-t_n)f_{\lambda}(Y_n)+\theta(t_{n+1}-t_n)f_{\lambda}(Y_{n+1})+g(Y_n)\Delta W_n+h(Y_n)\Delta\overline{N}_n.\end{aligned}$$ So we have $$\begin{aligned}
\overline{Y}(s)-Y(s+\Delta t)&=&(1-\theta)(s-t_{n+1})f_{\lambda}(Y_n)+\theta(s-t_{n+1})f_{\lambda}(Y_{n+1})+g(Y_n)\left(W(s)-W(t_{n+1})\right)\nonumber\\
&+&h(Y_n)\left(\overline{N}(s)-\overline{N}(t_{n+1})\right).
\label{ch3eq10}\end{aligned}$$ By squaring both sides of , taking expectation and using martingale properties of $\Delta W_n$ and $\Delta\overline{N}_n$, it follows that $$\begin{aligned}
\mathbb{E}||\overline{N}(s)-Y(s+\Delta t)||^2\leq 3\Delta t\mathbb{E}||f_{\lambda}(Y_n)||^2+3\Delta t\mathbb{E}||f_{\lambda}(Y_{n+1})||^2+3\Delta t\mathbb{E}||g(Y_n)||^2+3\lambda \Delta t \mathbb{E}||h(Y_n)||^2.\end{aligned}$$ Applying respectively the linear growth condition and Lemma \[ch3lemma2\] to the previous inequality, it follows that there exist a positive constant $C_2$ independent of $\Delta t$ such that : $$\begin{aligned}
\mathbb{E}||\overline{Y}(s)-Y(s+\Delta t)||^2\leq C_2(1+\mathbb{E}||X_0||^2)\Delta t.\end{aligned}$$
This complete the proof of Lemma \[ch3lemma2\].
Now, we are ready to give the proof of Theorem \[ch3th1\].
**\[Theorem \[ch3th1\]\]**
From equations and , it follows that : $$\begin{aligned}
||\overline{Y}(s)-X(s)||^2&=&\left\|\int_0^s[(1-\theta)(f_{\lambda}(Y(r))-f_{\lambda}(X(r^-))+\theta(f_{\lambda}(Y+\Delta t))-f_{\lambda}(X(r^-))]dr\right.\\
&+&\left.\int_0^s(g(Y(r))-g(X(r^-)))dW(r)+\int_0^s(h(Y(r))-h(X(r^-)))d\overline{N}(r)\right\|^2.\end{aligned}$$ Taking expectation in both sides of the above equality and using the inequality
$(a+b+c)^2\leq 3a^2+3b^2+3c^2$ for all $a, b, c\in\mathbb{R}$ leads to : $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}||\overline{Y}(s)-X(s)||^2\right]\leq 3\mathbb{E}\left(\sup_{0\leq s\leq t}M_1(t)\right)+3\mathbb{E}\left(\sup_{0\leq s\leq t}M_2(t)\right)+3\mathbb{E}\left(\sup_{0\leq s\leq t}M_3(t)\right),
\label{ch3eq11}\end{aligned}$$ where $$\begin{aligned}
M_1(t)=\left\|\int_0^s[(1-\theta)(f_{\lambda}(Y(r))-f_{\lambda}(X(r^-))]dr\right\|^2,\hspace{0.5cm}
M_2(s)=\left\|\int_0^s[g(Y(r))-g(X(r^-))]dW(r)\right\|^2\end{aligned}$$ $$\begin{aligned}
\text{and}\hspace{0.5cm} M_3(s)=\left\|\int_0^s[h(Y(r))-h(X(r^-))]d\overline{N}(r)\right\|^2.\end{aligned}$$ Using Holder inequality, it follows that : $$\begin{aligned}
M_1(s)\leq s\int_0^s\left\|(1-\theta)(f_{\lambda}(Y(r))-f_{\lambda}(X(r^-))+\theta(f_{\lambda}(Y(r+\Delta t))-f_{\lambda}(X(r^-))\right\|^2dr.
\label{ch3eq12}\end{aligned}$$ Using the convexity of the application $x\longmapsto ||x||^2$, it follows from that $$\begin{aligned}
M_1(s)\leq s\int_0^s(1-\theta)||f_{\lambda}(Y(r))-f_{\lambda}(X(r^-))||^2dr+s\int_0^s\theta||f_{\lambda}(Y(r+\Delta t))-f_{\lambda}(X(r^-))||^2dr.\end{aligned}$$ Taking the supremum in both sides of the above inequality and using the global Lipschitz condition satisfied by $f_{\lambda}$, and then taking expectation it follows that $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}M_1(s)\right]\leq t(1-\theta)L_{\lambda}\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr+t\theta L_{\lambda}\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr.
\label{ch3eq13}\end{aligned}$$ Using Doop’s inequality, it follows that : $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}M_2(s)\right]\leq 4\sup_{0\leq s\leq t}\mathbb{E}[M_2(s)]=4\sup_{0\leq s\leq t}\int_0^s\mathbb{E}||g(Y(r))-g(X(r^-))||^2dr.\end{aligned}$$ Using the global Lipschitz condition satisfied by $g$, it follows that : $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}M_2(s)\right]\leq 4L\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr.
\label{ch3eq14}\end{aligned}$$ Along the same lines as above, we have $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}M_3(s)\right]=4\lambda L\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr.
\label{ch3eq15}\end{aligned}$$ Inserting , and in leads to : $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}||\overline{Y}(s)-X(s)||^2\right]&\leq& 3T(1-\theta)L_{\lambda}\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr\nonumber\\
&+&3T\theta L_{\lambda}\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr\nonumber\\
&+&12L\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr\nonumber\\
&+&12\lambda L\int_0^t\mathbb{E}||Y(r)-X(r^-)||^2dr.
\label{ch3eq16}\end{aligned}$$ Using the fact that $$\begin{aligned}
||Y(r)-X(r^-)||^2=||(Y(r)-\overline{Y}(r)-(X(r^-)-\overline{Y}(r))||^2\leq 2||Y(r)-\overline{Y}(r)||^2+2||X(r^-)-\overline{Y}(r)||^2,\end{aligned}$$ it follows from that : $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}||\overline{Y}(s)-X(s)||^2\right]&\leq & 6T(1-\theta)L_{\lambda}\int_0^t\left[\mathbb{E}||Y(r)-\overline{Y}(r)||^2+\mathbb{E}||\overline{Y}(r)-X(r^-)||^2\right]dr\\
&+&6T\theta L_{\lambda}\int_0^t\left[\mathbb{E}||Y(r)-\overline{Y}(r)||^2+\mathbb{E}||\overline{Y}(r)-X(r^-)||^2\right]dr\\
&+&24L(1+\lambda)\int_0^t\left[\mathbb{E}||Y(r)-\overline{Y}(r)||^2+\mathbb{E}||\overline{Y}(r)-X(r^-)||^2\right]dr.\end{aligned}$$ Using lemma \[ch3lemma2\] in the above inequality, it follows that $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}||\overline{Y}(s)-X(s)||^2\right]&\leq & [6TL_{\lambda}+24L(1+\lambda)]\int_0^t\mathbb{E}\left[\sup_{0\leq r\leq s}||\overline{Y}(r)-X(r)||^2\right]ds\nonumber\\
&+&[6T^2L_{\lambda}+24TL(1+\lambda)]C_2(1+\mathbb{E}||X_0||^2)\Delta t.
\label{ch3eq17}\end{aligned}$$ Applying Gronwall lemma (continous form) to leads to the existence of a positive constant $C$ independent of $\Delta t$ such that $$\begin{aligned}
\mathbb{E}\left[\sup_{0\leq s\leq t}||\overline{Y}(s)-X(s)||^2\right]\leq C(1+\mathbb{E}||X_0||^2)\Delta t.\end{aligned}$$ This complete the proof of Theorem \[ch3th1\].
The strong convergence of the STM has been studied in [@Desmond1]. Since STM and CSTM convergence strongly to the exact slution, it is interesting to study their stability behaviours.
Linear mean-square stability of the CSTM
-----------------------------------------
In this section, we focus on the linear mean-square stability. Let’s consider the following linear test equation with real coefficients $$\begin{aligned}
dX(t)=aX(t^-)dt+bX(t^-)dW(t)+cX(t^-)dN(t), \hspace{0.5cm} X(0)=X_0.
\label{ch3lin1}\end{aligned}$$
The exact solution $X$ of SDEs is said to be exponentially mean-square stable if there exist constants $\alpha>0$ and $L>0$ such that : $$\begin{aligned}
\mathbb{E}||X(t)||^2\leq Le^{-\alpha t}\mathbb{E}||X(0)||^2.\end{aligned}$$
1. The numerical solution $X_n$ of SDEs is said to be exponentially mean-square stable if there exist constants $\alpha>0$ and $L>0$ such that : $$\begin{aligned}
\mathbb{E}||X_n||^2\leq Le^{-\alpha t}\mathbb{E}||X(0)||^2.\end{aligned}$$
2. The numerical solution $X_n$ of SDEs is said to be mean-square stable if there exist constants $0<L<1$ such that : for all $n\in[0,T]$ $$\begin{aligned}
\mathbb{E}||X_{n+1}||^2<L\mathbb{E}||X_n||^2.\end{aligned}$$
3. A numerical method is said to be A-stable if it is stable for any stepsize.
It is proved in [@Desmond1] that the exact solution of have the following stability property : $$\begin{aligned}
\lim_{t\longrightarrow \infty}\mathbb{E}||X(t)||^2=0\Longleftrightarrow l := 2a+b^2+\lambda c(2+c)<0,
\label{ch3lin2}\end{aligned}$$ where $\lambda$ is the intensity of the poisson precess $(N_t)_{t\geq 0}$.
Under condition , the numerical solution of produced by compensated stochastic theta method is mean-square stable for any stepsize $\Delta t>0$ if and only if
$\dfrac{1}{2}\leq\theta \leq 1$. For $0\leq \theta<\dfrac{1}{2}$ this numerical solution is mean-square stable for any stepsize $\Delta t>0$ satisfying : $$\begin{aligned}
\Delta t<\dfrac{-l}{(1-2\theta)(a+\lambda c)^2}.\end{aligned}$$
Applying the compensated theta method to gives $$\begin{aligned}
Y_{n+1}=Y_n+(1-\theta)\Delta t(a+\lambda c)Y_n+\theta \Delta t(a+\lambda c)Y_{n+1}+bY_n\Delta W_n+cY_n\Delta\overline{N}_n.\end{aligned}$$ So we have $$\begin{aligned}
(1-\theta \Delta ta-\theta \Delta t\lambda c)Y_{n+1}=Y_n+(1-\theta)\Delta t(a+\lambda c)Y_n+bY_n\Delta W_n+c\Delta\overline{N}_n.\end{aligned}$$ It follows that : $$\begin{aligned}
(1-\theta \Delta ta-\theta \Delta t\lambda c)^2\mathbb{E}||Y_{n+1}||^2=[1+(1-\theta)\Delta t(a+\lambda c)]^2\mathbb{E}||Y_n||^2+b^2\Delta t\mathbb{E}||Y_n||^2+c^2\lambda\Delta t\mathbb{E}||Y_n||^2.\end{aligned}$$ Therefore, $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2=\dfrac{1+[2(1-\theta)(a+\lambda c)+b^2+c^2\lambda]\Delta t+(1-\theta)^2(a+\lambda c)^2\Delta t^2}{(1-\theta \Delta t a-\theta \Delta t\lambda c)^2}\mathbb{E}||Y_n||^2.\end{aligned}$$ It follows that $\mathbb{E}||Y_n||^2$ is a geometric sequence which converge if and only if $$\begin{aligned}
\dfrac{1+[2(1-\theta)(a+\lambda c)+b^2+c^2\lambda]\Delta t+(1-\theta)^2(a+\lambda c)^2\Delta t^2}{(1-\theta \Delta t a-\theta \Delta t\lambda c)^2}<1.\end{aligned}$$ That is if and only if $$\begin{aligned}
(1-2\theta)(a+\lambda c)^2\Delta t<-l.
\label{ch3lin3}\end{aligned}$$ It follows that :
- If $\dfrac{1}{2}\leq \theta\leq 1$, then the condition is satisfied for any stepsize. And then the numerical solution is mean-square stable for any stepsize.
- If $0\leq\theta<\dfrac{1}{2}$, then it follows from that if $0<\Delta t<\dfrac{-l}{(1-2\theta)(a+\lambda c)^2}$, the numerical method is stable.
Changing $c$ to $-2-c$ does not affect the mean-square stability conditon . Hence the exact solution of have the same stability property under this transformation. It is interesting to look for what happens to the numerical solution under this transformation.
A numerical method applied to is said to be jump symmetric if whenever stable (unstable) for $\{a,b,c,\lambda , \Delta t\}$ it is also stable (unstable) for $\{a,b,-2-c,\lambda , \Delta t\}$.
The compensated stochastic theta method applied to is jump symmetric if and only if $\theta =\dfrac{1}{2}$.
1. For $\theta=\dfrac{1}{2}$, clearly the stability condition of the numerical solution is equivalent to the stability condition of the exact solution. Since the stability condition is invariant under the transformation $c\longmapsto -2-c$, it follows that the jump symmetric property holds.
2. If $\theta \neq \dfrac{1}{2}$, the right hand side of remains the same under the transformation $c\longmapsto -2-c$, but the left hand side changes. Therefore the jump symmetry property does not holds.
If the exact solution of the problem is mean-square stable, then for $\dfrac{1}{2}<\theta\leq 1$, it follows from the stability property of the CSTM is preserved under the transformation
$c\longmapsto -2-c$.
Nonlinear mean-square stability
-------------------------------
This section is devoted to the nonlinear mean-square analysis.
Troughout, this section, we make the following assumptions.
\[ch3assump1\] We assume that there exist constants $\mu, \sigma, \gamma$ such that for all $x,y\in\mathbb{R}^n$ $$\begin{aligned}
\langle x-y, f(x)-f(y)\rangle\leq \mu||x-y||^2\label{ch3nonlin1}\\
||g(x)-g(y)||^2\leq \sigma||x-y||^2\nonumber\\
||h(x)-h(y)||^2\leq \gamma||x-y||^2.\nonumber\end{aligned}$$ Usually, condition is called “one-sided Lipschitz condition”.
### Nonlinear mean-square stability of the exact solution
[@Desmond2 Theorem 4, pp 13] \[ch3th2\]
Under assumptions \[ch3assump1\], any two solutions $X(t)$ and $Y(t)$ of the SDEs with jumps with $\mathbb{E}||X_0||^2<\infty$ and $\mathbb{E}||Y_0||^2<\infty$ satisfy the following property $$\begin{aligned}
\mathbb{E}||X(t)-Y(t)||^2\leq \mathbb{E}||X_0-Y_0||^2e^{\alpha t},\end{aligned}$$ where $\alpha : =2\mu+\sigma +\lambda\sqrt{\gamma}(\sqrt{\gamma}+2)$.
Hence, the condition $\alpha <0$ is sufficient for the exponential mean-square stability property.
The two solutions $X(t)$ and $Y(t)$ of satisfy respectively $$\begin{aligned}
dX(t)=f(X(t^-))dt+g(X(t^-))dW(t)+h(X(t^-))dN(t)\end{aligned}$$ and $$\begin{aligned}
dY(t)=f(Y(t^-))dt+g(Y(t^-))dW(t)+h(Y(t^-))dN(t).\end{aligned}$$ Applying Itô’s lemma for product (Lemma \[ch1Itoproduct\]) to the stochastic process $Z(t)=||X(t)-Y(t)||^2$ leads to $$\begin{aligned}
dZ(t)&=&2\langle X(t^-)-Y(t^-),d(X(t^-))-d(Y(t^-))\rangle+||d(X(t^-))-d(Y(t^-))||^2\\
&=&\left[2\langle X(t^-)-Y(t^-), f(X(t^-))-f(Y(t^-))\rangle+2\lambda \langle X(t^-)-Y(t^-), h(X(t^-))-h(Y(t^-))\rangle\right.\\
&+&\left.||g(X(t^-))-g(Y(t^-))||^2+\lambda||h(X(t^-))-h(Y(t^-))||^2\right]dt+dM_t,\end{aligned}$$ where $M_t$ is a martingale and where we used the following rule of calculation $$\begin{aligned}
dtdt=dtdW(t)=0,\hspace{0.5cm} dN_tdN(t)=dN_t,\hspace{0.5cm} dW_tdW_t=dt\hspace{0.5cm} and \hspace{0.2cm}dtdN(t)=dW_tdN_t=0.\end{aligned}$$ Using Assumptions \[ch3assump1\], ones get : $$\begin{aligned}
d||X(t)-Y(t)||^2&\leq& [2\mu||X(t^-)-Y(t^-)||^2+2\lambda\sqrt{\gamma}||X(t^-)-Y(t^-)||^2+\sigma||X(t^-)-Y(t^-)||^2\\
&+&\lambda\gamma||X(t^-)-Y(t^-)||^2]dt +dM(t),\end{aligned}$$
So we have $$\begin{aligned}
d||X(t^-)-Y(t^-)||^2\leq[2\mu+\sigma+\lambda\sqrt{\gamma}(\sqrt{\gamma}+2)]||X(t^-)-Y(t^-)||^2dt+dM(t),\end{aligned}$$ which can be writen into its following integral form : $$\begin{aligned}
||X(t^-)-Y(t^-)||^2\leq [2\mu+\sigma+\lambda\sqrt{\gamma}(\sqrt{\gamma}+2)]\int_0^t||X(s^-)-Y(s^-)||^2ds+\int_0^tdM(s).
\label{ch3nonlin2}\end{aligned}$$ Taking expectation in both sides of and using the fact that $\mathbb{E}\left(\int_0^tdM(s)\right)=0$ leads to : $$\begin{aligned}
\mathbb{E}||X(t^-)-Y(t^-)||^2\leq [2\mu+\sigma+\lambda\sqrt{\gamma}(\sqrt{\gamma}+2)]\int_0^t\mathbb{E}||X(s^-)-Y(s^-)||^2ds.
\label{ch3nonlin3}\end{aligned}$$ Applying Gronwall lemma ( continuous form ) to leads to : $$\begin{aligned}
\mathbb{E}||X(t^-)-Y(t^-)||^2\leq \mathbb{E}||X_0-Y_0||^2e^{[2\mu+\sigma+\lambda\sqrt{\gamma}(\sqrt{\gamma}+2)]t}.\end{aligned}$$ This complete the proof of Theorem \[ch3th2\].
For the linear test equation , the one-sided Lipschitz and the global Lipschitz condition become $$\begin{aligned}
\langle x-y, f(x)-f(y)\rangle=a|x-y|^2\\
|g(x)-g(y)|^2=b^2|x-y|^2\\
|h(x)-h(y)|^2=c^2|x-y|^2.\end{aligned}$$ Along the same lines as for the nonlinear case, we obtain : $$\begin{aligned}
\mathbb{E}|X(t^-)|^2=\mathbb{E}|X_0|^2e^{(2a+b^2+\lambda c(2+c))t}.\end{aligned}$$ Therefore, we have the following equivalence for the linear mean-square stability $$\begin{aligned}
\lim_{t\longrightarrow +\infty}\mathbb{E}|X(t)|^2=0\Longleftrightarrow l:= 2a+b^2+\lambda c(2+c)<0.\end{aligned}$$
Based on Theorem \[ch3th2\], it is interesting to analyse whether or not the numerical solution of reproduce the mean-square stability of the exact solution.
### Nonlinear mean-square stability of the numerical solutions
**\[Stability of the stochastic theta method\]**\[ch3th3\]
Under Assumptions \[ch3assump1\] and the further hypothesis $\alpha<0$, for $$\begin{aligned}
\Delta t< \dfrac{-\alpha}{\lambda^2\gamma},\end{aligned}$$ the Euler backward method (STM with $\theta=1$) applied to equation is mean-square stable in the sense that $$\begin{aligned}
\mathbb{E}||X_n-Y_n||^2\leq \mathbb{E}||X_0-Y_0||^2e^{\beta_1(\Delta t)n\Delta t},\end{aligned}$$ where $$\begin{aligned}
\beta_1(\Delta t):=\dfrac{1}{\Delta t}ln\left(\dfrac{1+(\sigma+\lambda\gamma+2\lambda\sqrt{\gamma})\Delta t+\lambda^2\gamma\Delta t^2}{1-\mu\Delta t}\right).\end{aligned}$$
Let’s introduce the following notations : $$\begin{aligned}
\Delta Z_n=X_n-Y_n, \hspace{0.2cm} \Delta f_n=f(X_n)-f(Y_n),\hspace{0.2cm}\Delta g_n=g(X_n)-g(Y_n), \hspace{0.2cm} \Delta h_n=h(X_n)-h(Y_n)\end{aligned}$$ If $\theta =1$, the numerical approximation applied to $X$ and $Y$ gives : $$\begin{aligned}
Y_{n+1}=Y_n+f(Y_{n+1})\Delta t+g(Y_n)\Delta W_n +h(Y_n)\Delta N_n\\
X_{n+1}=X_n+f(X_{n+1})\Delta t+g(X_n)\Delta W_n +h(X_n)\Delta N_n.\end{aligned}$$ So we have : $$\begin{aligned}
||\Delta Z_{n+1}-\Delta f_{n+1}\Delta t||^2=||\Delta Z_n+\Delta g_n\Delta W_n+\Delta h_n\Delta N_n||^2.
\label{ch3nonlin5}\end{aligned}$$ Using the independence of $\Delta W_n$ and $\Delta N_n$ and the fact that $$\begin{aligned}
\mathbb{E}|\Delta N_n|^2&=&var(\Delta N_n)+\left(\mathbb{E}(\Delta N_n)\right)^2=\lambda\Delta t+\lambda^2\Delta t^2\\
\mathbb{E}||\Delta W_n||^2&=&\Delta t, \hspace{0.5cm} \mathbb{E}||\Delta W_n||=0, \hspace{0.5cm} \mathbb{E}|\Delta N_n|=\lambda\Delta t,\end{aligned}$$ we obtain from the following estimation : $$\begin{aligned}
\mathbb{E}||\Delta Z_{n+1}||^2-2\Delta t\mathbb{E}\langle \Delta Z_{n+1}, \Delta f_{n+1}\rangle &\leq& \mathbb{E}||\Delta Z_n||^2+\Delta t\mathbb{E}||\Delta g_n||^2+\lambda \Delta t(1+\lambda \Delta t)\mathbb{E}||\Delta h_n||^2\\
&+&2\lambda\Delta t\mathbb{E}\langle\Delta Z_n, \Delta h_n\rangle.\end{aligned}$$ Using the one-sided Lipschitz condition and the global Lipschitz condition, it follows that : $$\begin{aligned}
\mathbb{E}||\Delta Z_{n+1}||^2 &\leq& 2\Delta t\mu \mathbb{E}||\Delta Z_{n+1}||^2+\mathbb{E}||\Delta Z_n||^2+\sigma\Delta t\mathbb{E}||\Delta Z_n||^2+\lambda\Delta t(1+\lambda\Delta t)\gamma \mathbb{E}||\Delta Z_n||^2\\
(1-2\mu\Delta t)\mathbb{E}||\Delta Z_{n+1}||^2&\leq&[1+(\sigma + \lambda \gamma +2\lambda\sqrt{\gamma})\Delta t+\lambda^2\gamma\Delta t^2]\mathbb{E}||\Delta Z_n||^2.\end{aligned}$$ The latter inequality leads to : $$\begin{aligned}
\mathbb{E}||\Delta Z_{n+1}||^2 \leq \left[\dfrac{1+(\sigma + \lambda \gamma + 2\lambda\sqrt{\gamma})\Delta t+\lambda^2\gamma\Delta t^2}{1-2\mu \Delta t}\right]\mathbb{E}||\Delta Z_n||^2.\end{aligned}$$ Therefore, we have : $$\begin{aligned}
\mathbb{E}||\Delta Z_n||^2 \leq \left[\dfrac{1+(\sigma + \lambda \gamma + 2\lambda\sqrt{\gamma})\Delta t+\lambda^2\gamma\Delta t^2}{1-2\mu \Delta t}\right]^n\mathbb{E}||\Delta Z_0||^2.
\label{ch3nonlin6}\end{aligned}$$ In order to have stability, we impose the following condition : $$\begin{aligned}
\dfrac{1+(\sigma + \lambda \gamma + 2\lambda\sqrt{\gamma})\Delta t+\lambda^2\gamma\Delta t^2}{1-2\mu \Delta t}<1.
\label{ch3nonlin7}\end{aligned}$$ The hypothesis $\alpha<0$ implies that $\mu<0$. So $1-2\mu\Delta t>0$, for all positive stepsize. It follows that is equivalent to $$\begin{aligned}
\Delta t< \dfrac{-\alpha}{\lambda^2\gamma}.\end{aligned}$$ Applying the equality $a^n=e^{n\ln a}$, for all $a>0$ and all $n\in\mathbb{N}$ to complete the proof of Theorem \[ch3th3\].
**\[A-stability of the compensated Euler backward method\]**\[ch3th3\]
Under Assumptions \[ch3assump1\] and the further hypothesis $\alpha<0$, for any stepsize, the compensated backward Euler method ( CSTM with $\theta=1$) for equation is mean square stable in the sense that : $$\begin{aligned}
\mathbb{E}||X_n-Y_n||^2\leq \mathbb{E}||X_0-Y_0||^2e^{\beta_2(\Delta t)n\Delta t},\end{aligned}$$ where $$\begin{aligned}
\beta_2(\Delta t) :=\dfrac{1}{\Delta t}ln\left(\dfrac{1+(\sigma +\lambda\gamma)\Delta t}{1-2(\mu+\lambda\sqrt{\gamma})\Delta t}\right).\end{aligned}$$
We use the same notations as for the proof of Theorem \[ch3th3\] except for $\Delta f^{\lambda}_n$ for which we have $\Delta f^{\lambda}_n=f_{\lambda}(X_n)-f_{\lambda}(Y_n)$.
Along the same line as for the proof of Theorem \[ch3th3\], we obtain : $$\begin{aligned}
||\Delta Z_{n+1}-\Delta t\Delta f^{\lambda}_{n+1}||^2=||\Delta Z_n+\Delta g_n\Delta W_n+\Delta h_n\Delta\overline{N}_n||^2.
\label{ch3nonlin8}\end{aligned}$$ Futhermore, the relations $\mathbb{E}|\Delta\overline{N}_n|=0$ and $ \mathbb{E}|\Delta\overline{N}_n|^2=\lambda\Delta t$ leads to : $$\begin{aligned}
\langle x-y, f_{\lambda}(x)-f_{\lambda}(y)\rangle&=&\langle x-y, f(x)-f(y)\rangle+\lambda\langle x-y, h(x)-h(y)\rangle\\
&\leq& (\mu+\lambda\sqrt{\gamma})||x-y||^2.\end{aligned}$$ Using the independence of $\Delta W_n$ and $\Delta\overline{N}_n$, it follows from that $$\begin{aligned}
\mathbb{E}||\Delta Z_{n+1}||^2\leq 2\Delta t\mathbb{E}\langle \Delta Z_{n+1}, \Delta f_{n+1}\rangle+\mathbb{E}||\Delta Z_n||^2+\Delta t \mathbb{E}||\Delta g_n||^2+\lambda \Delta t\mathbb{E}||\Delta h_n||^2.
\label{ch3nonlin9}\end{aligned}$$ Using the one-sided Lipschitz and the global Lipschitz condition, it follows that : $$\begin{aligned}
(1-2(\mu+\lambda\sqrt{\gamma})\Delta t)\mathbb{E}||\Delta Z_{n+1}||^2\leq(1+\sigma\Delta t+\lambda\gamma\Delta t)\mathbb{E}||\Delta Z_n||^2.\end{aligned}$$ Therefore, $$\begin{aligned}
\mathbb{E}||\Delta Z_n||^2\leq \left[\dfrac{1+\sigma\Delta t+\lambda\gamma\Delta t}{1-2(\mu+\lambda\sqrt{\gamma})\Delta t}\right]^n\mathbb{E}||Z_0||^2.
\label{ch3nonlin10}\end{aligned}$$ In order to have stability, we need the following condition to be fulfilled $$\begin{aligned}
\dfrac{1+\sigma\Delta t+\lambda\gamma\Delta t}{1-2(\mu+\lambda\sqrt{\gamma})\Delta t}<1.
\label{ch3nonlin11}\end{aligned}$$ From the hypothesis $\alpha<0$, we have $2(\mu+\lambda\sqrt{\gamma})<0$ and then $1-2(\mu+\lambda\sqrt{\gamma})\Delta t>0$ for any stepsize. Hence condition is equivalent to $\alpha\Delta t<0$, which is satisfied for any stepsize.
Applying the relation $a^n=e^{n\ln a}$ to complete the proof of the theorem.
Numerical Experiments
---------------------
The purpose of this section is to illustrate our theorical results of strong convergence and stability. We will focus in the linear case. We consider the linear jump-diffusion Itô’s stochastic integral (SDEs) $$\begin{aligned}
\label{ch3num1}
\left\{\begin{array}{ll}
dX(t)=aX(t^-)dt+bX(t^-)dW(t)+cX(t^-)dN(t), \hspace{0.5cm} t\geq 0,\hspace{0.5cm} c>-1,\\
X(0)=1.
\end{array}
\right.\end{aligned}$$
### Strong convergence illustration
In order to illustrate the strong convergence result, we need the exact solution of problem .
The problem has the following process as a unique solution $$\begin{aligned}
X(t)=X_0\exp\left[\left(a-\dfrac{b^2}{2}\right)t+bW(t)\right](1+c)^{N(t)},\end{aligned}$$ which can be written in the following equivalent form $$\begin{aligned}
X(t)=X_0\exp\left[\left(a-\dfrac{b^2}{2}\right)t+bW(t)+\ln(1+c)N(t)\right].\end{aligned}$$
1. Obviously, the functions $f(x)=ax$, $g(x)=bx$ and $h(x)=cx$ satisfy the global Lipschitz condition and the linear growth condition. Therefore from Theorem \[ch2th1\], it follows that the problem admit a unique solution.
2. Let’s consider the following Itô’s jump-diffusion process $$\begin{aligned}
Z(t)=\left(a-\dfrac{b^2}{2}\right)t+bW(t)+N(t)\ln(1+c).\end{aligned}$$ The function $f: [0,\infty)\longrightarrow \mathbb{R}, \hspace{0.3cm} x\longmapsto x_0\exp(x)$ is infinitely differentiable.
Then applying Itô formula for jump process to the process $Z(t)$ leads to : $$\begin{aligned}
f(Z_t)&=&f(Z_0)+\int_0^t\left(a-\dfrac{b^2}{2}\right)f'(Z_{s^-})ds+\dfrac{1}{2}\int_0^tb^2f''(Z_{s^-})ds+\int_0^tbf'(Z_{s^-})dW(s)\nonumber\\
&+&\int_0^t(f(Z_s)-f(Z_{s^-}))dN(s),
\label{ch3num2}\end{aligned}$$ where $$\begin{aligned}
f(Z_s)-f(Z_{s^-})&=&X_0\exp[Z_{s^-}+\ln(1+c)]-X_0\exp(Z_{s^-})\nonumber\\
&=&(1+c)X_0\exp(Z_{s^-})-X_0\exp(Z_{s^-})\nonumber\\
&=&cX_0\exp(Z_{s^-})=cf(Z_{s^-})
\label{ch3num3}\end{aligned}$$ and $$\begin{aligned}
X(s^-)=f(Z_{s^-})=f'(Z_{s^-})=f''(Z_{s^-}).
\label{ch3num4}\end{aligned}$$ Substituting and in and rewriting the result into its differential form leads to $$\begin{aligned}
dX(t)=aX(t^-)dt+bX(t^-)dW(t)+cX(t^-)dN(t).\end{aligned}$$ So $X(t)$ satisfies the desired equation.
For the numerical simulation, we take $a=b=1$, $c=0.5$ and $\lambda=1$. We have the following graphs for the strong error. We use $5000$ sample paths. The algorithms for simulation are based on [@Desmond3]. We take $d t=2^{14}$ and $\Delta t=2^{p-1}$ for $p=1,2,3,4,5$. The error is computing at the end point $T=1$.
### Mean-square stability illustration
In order to illustrate our theoretical result of A-stability, we first consider two examples
**Example I** $a=b=2$, $c=-0.9$ and $\lambda=9$.
**Example II** $a=-7$, $b=c=1$ and $\lambda=4$.
In both examples, the stability condition is satisfied. So exact solutions of both examples are mean-square stable. For $\theta$ slightly less than $0.5$ (for instance $\theta=0.495$) both solutions may be unstable for a large stepsize ($\Delta t=60, 25$), but for $\dfrac{1}{2}\leq \theta\leq 1$, numerical solutions of both examples are stable. From the top to the bottom, we present numerical examples of example I and example II respectively.
The following curves provide the stability comparison between CSTM and STM. We focus on the example I. Here, $a>0$ and $c<0$. So the jump part can stabilise the problem. In this case, from the theoretical result the STM is stable for $\Delta t<0.0124829$. For $\Delta t=0.005$, both CSTM and STM stabilities behavior look the same. But for $\Delta t=0.5$ CSTM is stable while STM produce an oscillation. For $\Delta t=0.1$, numerical solution of STM grows rapidly to the scale $10^7$ and is unstable while the numerical solution of CSTM is stable. So CSTM works better than STM.
In this chapter, we provided the proof of the strong convergence of order $0.5$ of the CSTM under global Lipschitz condition. We also studied the stability behaviour of both STM and CSTM. We proved that the CSTM works better than the STM. Some situations in real life are modelised by SDEs with jumps, where the drift coefficient does not satisfy the global Lipschitz condtion. It is proved in [@Martin3] that the Euler explicit method for such equations diverge strongly. The tamed Euler scheme for SDEs without jump is the currently investigated by many authors. The compensated tamed Euler scheme for SDEs with jumps is not yet well developped in the litterature. In the following chapter, we establish the strong convergence of the compensated tamed Euler scheme for SDEs with jumps under non-global Lipschitz condition. This scheme is slightly different to what is already done in the litterature.
Strong convergence of the compensated tamed Euler scheme for stochastic differential equation with jump under non-global Lipschitz condition
============================================================================================================================================
Under non-global Lipschitz condition, Euler Explicit method fails to converge strongly to the exact solution, while Euler implicit method converges but requires more computational efforts. The strong convergence of the tamed Euler scheme has been investigated in [@Martin1]. This scheme is explicit and requires less computational efforts than the Euler implicit method. In this chapter, we extend the strong convergence of the tamed Euler scheme by introducing its compensated form for stochastic differential equations with jumps. More precisely, we prove that under non-global Lipschitz condition, the compensated tamed Euler scheme converges strongly with order $0.5$ to the exact solution of the SDEs with jumps. This scheme is different to the one proposed in [@Kon]. As opposed to what is done in [@Kon], here we obtain the strong convergence and the rate of convergence simultaneously under more relaxed conditions. The contents of this chapter can also be found in [@atjdm1].
Compensated tamed Euler scheme {#ch4intro}
------------------------------
In this chapter, we still consider the jump-diffusion Itô’s stochastic differential equations (SDEs) of the form $$\begin{aligned}
dX(t)= f(X(t^{-}))dt +g(X(t^{-}))dW(t)+h(X(t^{-}))dN(t), \hspace{0.5cm}
X(0)=X_0,
\label{ch4exactsol}\end{aligned}$$ where $W_t$ is an $m$-dimensional Brownian motion, $f :\mathbb{R}^d\longrightarrow\mathbb{R}^d$ satisfies the one-sided Lipschitz condition and the polynomial growth condition. The functions $g : \mathbb{R}^d \longrightarrow\mathbb{R}^{d\times m}$ and $h :\mathbb{R}^d \longrightarrow\mathbb{R}^d$ satisfy the global Lipschitz condition, $N_t$ is a one dimensional poisson process with parameter $\lambda$.
We recall that the compensated poisson process $\overline{N}(t) := N(t)-\lambda t$ is a martingale satisfying the following properties : $$\begin{aligned}
\mathbb{E}\left(\overline{N}(t+s)-\overline{N}(t)\right)=0,\,\qquad \qquad \mathbb{E}\vert \overline{N}(t+s)-\overline{N}(t)\vert^2=\lambda s,\qquad s, t \geqslant 0.\end{aligned}$$ We can rewrite the jump-diffusion SDEs in the following equivalent form $$\begin{aligned}
dX(t)= f_\lambda(X(t^{-}))dt +g(X(t^{-}))dW(t)+h(X(t^{-}))d\overline{N}(t),\end{aligned}$$ where $$f_\lambda(x)=f(x)+\lambda h(x).
\label{ch4flambda}$$ To easy notation, we will use $X(t)$ instead of $X(t^{-})$.
If $T$ is the final time, the tamed Euler scheme is defined by : $$X_{n+1}^{M}=X_{n}^{M}+\dfrac{\Delta t f(X_{n}^{M})}{1+ \Delta t\Vert f(X_{n}^{N}) \Vert }+g(X_{n}^{M}) \Delta W_n +h(X_{n}^{M})\Delta N_n
\label{ch4tam}$$ and the compensated tamed Euler scheme is given by : $$\begin{aligned}
Y_{n+1}^{M}=Y_{n}^{M}+\dfrac{\Delta t f_\lambda(Y_{n}^{M})}{1+ \Delta t \Vert f_{\lambda}(Y_{n}^{M}) \Vert }+g(Y_{n}^{M}) \Delta W_n +h(Y_{n}^{M})\Delta\overline{N}_n,
\label{ch4tamc}\end{aligned}$$ where $M\in\mathbb{N}$ is the number of steps and $\Delta t=\dfrac{T}{M}$ is the stepsize.
Inspired by [@Martin1], we prove the strong convergence of the numerical approximation to the exact solution of .
Moments bounded of the numerical solution
-----------------------------------------
\[ch4nota1\] Throughout this chapter $(\Omega, \mathcal{F}, \mathbb{P})$ denote a complete probability space with a filtration $(\mathcal{F}_t)_{t\geq 0}$, $||X||_{L^p(\Omega, \mathbb{R}^d)}$ is equal to $(\mathbb{E}||X^p||)^{1/p}$, for all $p\in[1,+\infty)$ and for all $(\mathcal{F}_t)-$adapted process $X$. For all $x, y\in\mathbb{R}^d$, we denote by $\langle x, y\rangle=x.y= x_1y_1+x_2y_2+\cdots+x_dy_d$, $||x||=\left(\langle x, x\rangle\right)^{1/2}$ and $||A||=\sup_{x\in\mathbb{R}^d, ||x||\leq 1}||Ax||$ for all $A\in\mathbb{R}^{m\times d}$. We use also the following convention : $\sum_{i=u}^na_i=0$ for $u>n$.
We define the continuous time interpolation of the discrete numerical approximation of by the family of processes $\left(\overline{Y}^M\right)_M $, $
\overline{Y}^M : [0,T]\times\Omega \longrightarrow \mathbb{R}^d
$ such that : $$\begin{aligned}
\overline{Y}^M_t =Y^M_n+\dfrac{(t-n\Delta t)f_{\lambda}(Y^M_n)}{1+\Delta t||f_{\lambda}(Y^M_n)||}+g(Y^M_n)(W_t-W_{n\Delta t})+ h(Y^M_n)(\overline{N}_t-\overline{N}_{n\Delta t}),
\label{ch4continoussolu}\end{aligned}$$ for all $M\in\mathbb{N}$, all $n\in\{0,\cdots, M-1\}$, and all $t\in[n\Delta t, (n+1)\Delta t[$.
\[ch4assumption1\] Throughout this chapter, We make the following assumptions :
$(A.1)$ $f,g,h\in C^1$.
$(A.2)$ For all $p>0$, there exist a finite $M_p>0$ such that $\mathbb{E}||X_0||\leq M_p$.
$(A.3)$ $g$ and $h$ satisfy the global Lipschitz condition: $$\begin{aligned}
||g(x)-g(y)||\vee ||h(x)-h(y)||\leq C||x-y||, \hspace{0.5cm} \forall\;x,y\in \mathbb{R}^d.\end{aligned}$$ $(A.4)$ $f$ satisfies the one-sided Lipschitz condition : $$\begin{aligned}
\langle x-y, f(x)-f(y)\rangle\leq C||x-y||^2,\hspace{0.5cm} \forall\; x,y\in \mathbb{R}^d.\end{aligned}$$ $(A.5)$ $f$ satisfies the superlinear growth condition : $$\begin{aligned}
||f(x)-f(y)||\leq C(K+ ||x||^c+||y||^c)||x-y||, \hspace{0.5cm} \forall\; x,y\in \mathbb{R}^d,\end{aligned}$$ where $K$, $C$ and $c$ are strictly positive constants.
Under conditions $(A.1)$, $(A.2)$ and $(A.3)$ of Assumptions \[ch4assumption1\], it is proved in [@Desmond2 Lemma 1] that has a unique solution with all moments bounbed.
We note that if Assumptions \[ch4assumption1\] are satisfied, the function $f_{\lambda}$ defined in satisfies the one-sided Lipschitz condition and the superlinear growth condition with constants $C_{\lambda} :=C(1+\lambda)$ and $K_{\lambda} : =K+\lambda$.
Indeed, for all $x,y\in\mathbb{R}^d$, $$\begin{aligned}
\langle x-y, f_{\lambda}(x)-f_{\lambda}(y)\rangle&=&\langle x-y,f(x)\rangle+\lambda\langle x-y, h(x)-h(y)\rangle \\
&\leq& C(1+\lambda)||x-y||,\\
||f_{\lambda}(x)-f_{\lambda}(y)||&\leq& ||f(x)-f(y)||+\lambda||h(x)-h(y)||\\
&\leq& C(K+\lambda +||x||^c+||y||^c)||x-y|| \\
&=&C(K_{\lambda}+||x||^c+||y||^c)||x-y||.\end{aligned}$$ Since the value of the constant does not matter too much, we will use $C$ instead of $C_{\lambda}$ and $K$ instead of $K_{\lambda}$. Throughout this work, the generic constants $C_p$ may change the value from one line to another one. We will sometimes use $Y_n^M$ instead of $Y_n^M(\omega)$ to simplify notations.
The main result of this section is given in the following theorem.
\[ch4theorem1\]
Let $Y_n^M : \Omega\longrightarrow \mathbb{R}^d$ be defined by for all $M\in\mathbb{N}$ and all $ n\in\{0,\cdots, M\}$. Then the following inequality holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\mathbb{E}\left[||Y_n^M||^p\right]<+\infty,
\end{aligned}$$ for all $p\in[1,\infty)$.
In order to prove Theorem \[ch4theorem1\] we introduce the following notations facilitating computations.
\[ch4notation1\]
$$\begin{aligned}
\alpha^M_k := \mathrm{1}_{\{||Y^M_k||\geq 1\}}\left\langle\dfrac{Y^M_k}{||Y^M_k||}, \dfrac{g(Y^M_k)}{||Y^M_k||}\Delta W^M_k\right\rangle,\\\\
\beta^M_k := \mathrm{1}_{\{||Y^M_k||\geq 1\}}\left\langle\dfrac{Y^M_k}{||Y^M_k||}, \dfrac{h(Y^M_k)}{||Y^M_k||}\Delta\overline{N}^M_k\right\rangle,
\end{aligned}$$
$$\begin{aligned}
\beta :=\left(1+K+2C+KTC+TC+||f_{\lambda}(0)||+||g(0)||+||h(0)||\right)^4,\\\\
D^M_n := (\beta+||\varepsilon||)\exp\left(\dfrac{3\beta}{2}+\sup_{u\in\{0,\cdots,n\}}\sum_{k=u}^{n-1}\left[\dfrac{3\beta}{2}||\Delta W^M_k||^2+\dfrac{3\beta}{2}||\Delta\overline{N}^M_k||+\alpha^M_k+\beta^M_k\right]\right),\\
\Omega^M_n :=\{\omega\in \Omega : \sup_{k\in\{0,1,\cdots, n-1\}}D^M_k(\omega)\leq M^{1/2c}, \sup_{k\in\{0,1,\cdots,n-1\}}||\Delta W^M_k(\omega)||\leq 1,\\ \sup_{k\in\{0,1,\cdots,n-1\}}||\Delta \overline{N}^M_k(\omega)||\leq 1\}.
\end{aligned}$$
In order to prove Theorem \[ch4theorem1\], we need the following lemmas.
\[ch4lemma1\] For all positive real numbers $a$ and $b$, the following inequality holds $$\begin{aligned}
1+a+b^2\leq e^{a+\sqrt{2}b}.
\end{aligned}$$
For $a\geq 0$ fixed, let’s define the function $f(b)=e^{a+\sqrt{2}b}-1-a-b^2$. It can be easily checked that $f'(b)=\sqrt{2}e^{a+\sqrt{2}b}-2b$ and $f''(b)=2(e^{a+\sqrt{2}b}-1)$. Since $a$ and $b$ are positive, it follows that $f''(b)\geq 0$ for all $b\geq 0$. So $f'$ is a non-decreasing function. Therefore, $f'(b)\geq f'(0)=\sqrt{2}e^a>0$ for all $b\geq 0$. This implies that $f$ is a non-decreasing function. Hence $f(b)\geq f(0)=e^a-1-a$ for all $b\geq 0$. Since $1+a\leq e^a$ for all positive number $a$, it follows that $f(b)\geq 0$ for all positive number $b$. i.e $1+a+b^2\leq e^{a+\sqrt{2}b}$, $\forall\;b\geq0$. Therefore for all $a\geq 0$ fixed, $1+a+b^2\leq e^{a+\sqrt{2}b}$, $\forall\;b\geq0$. The proof of lemma is complete.
Following closely , we have the following main lemma.
\[ch4lemma2\] The following inequality holds for all $M\in \mathbb{N}$ and all $n\in\{0,1,\cdots, M\}$ $$\begin{aligned}
\mathbf{1}_{\Omega^M_n}||Y^M_n||\leq D^M_n,
\label{ch4Denobound}
\end{aligned}$$ where $D_n^M$ and $\Omega^M_n$ are given in Notation \[ch4notation1\].
Using the inequality $\dfrac{\Delta t}{1+\Delta t||f_{\lambda}(x)||}\leq T$, the global Lipschitz condition of $g$ and $h$ and the polynomial growth condition of $f_{\lambda}$ we have the following estimation on $\Omega^M_{n+1}\cap\{\omega\in \Omega : ||Y^M_n(\omega)||\leq 1\}$, for all $n\in\{0,1,\cdots, M-1\}$ $$\begin{aligned}
||Y^M_{n+1}||&\leq ||Y^M_n||+\dfrac{\Delta t||f_{\lambda}(Y^M_n)||}{1+\Delta t||f_{\lambda}(Y^M_n)||}+||g(Y^M_n)||||\Delta W^M_n||+||h(Y^M_n)||||\Delta\overline{N} ^M_n|| \nonumber \\
&\leq ||Y^M_n||+T||f_{\lambda}(Y^M_n)-f_{\lambda}(0)||+T||f_{\lambda}(0)||+ ||g(Y^M_n)-g(0)||+||g(0)||\nonumber\\
&+||h(Y^M_n)-h(0)||+||h(0)||\nonumber\\
&\leq ||Y^M_n||+TC(K+||Y^M_n||^c+||0||^c)||Y^M_n-0||+T||f_{\lambda}(0)||+C||Y^M_n||+C||Y^M_n||\nonumber\\
&+||g(0)||+||h(0)||.\nonumber
\end{aligned}$$ Since $||Y^M_n||\leq 1$, it follows that : $$\begin{aligned}
||Y^M_{n+1}||
&\leq 1+KTC +TC+2C+T||f_{\lambda}(0)||+||g(0)||+||h(0)||\leq \beta.
\label{ch4normY}
\end{aligned}$$ Futhermore, from the numerical approximation , we have $$\begin{aligned}
\label{ch4partnorm2}
||Y^M_{n+1}||^2&=&||Y^M_n||^2+\dfrac{\Delta t^2||f_{\lambda}(Y^M_n)||^2}{(1+\Delta t||f_{\lambda}(Y^M_n)||)^2}+||g(Y^M_n)\Delta W^M_n||^2+||h(Y^M_n)\Delta\overline{N}^M_n||^2\nonumber\\
&+&\dfrac{2\Delta t\langle Y^M_n, f_{\lambda}(Y^M_n)\rangle}{1+\Delta t||f_{\lambda}(Y^M_n)||}+2\langle Y^M_n,g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&+&\dfrac{2\langle \Delta tf_{\lambda}(Y^M_n),g(Y^M_n)\Delta W^M_n\rangle}{1+\Delta t||f_{\lambda}(Y^M_n)||}+\dfrac{2\langle\Delta tf_{\lambda}(Y^M_n),h(Y^M_n)\Delta\overline{N}^M_n\rangle}{1+\Delta t||f_{\lambda}(Y^M_n)||}\nonumber\\
&+&2\langle g(Y^M_n)\Delta W^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle.
\end{aligned}$$ Using Cauchy-Schwartz inequality and the estimation $
\dfrac{1}{1+\Delta t||f_{\lambda}(Y^M_n)||}\leq 1$, we obtain the following inequality from : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq& ||Y^M_n||^2+\Delta t^2||f_{\lambda}(Y^M_n)||^2+||g(Y^M_n)||^2||\Delta W^M_n||^2+||h(Y^M_n)||^2|\Delta\overline{N}^M_n|^2\nonumber\\
&+&2\Delta t\left|\langle Y^M_n, f_{\lambda}(Y^M_n)\rangle\right|+2\langle Y^M_n,g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n, h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&+&2\Delta t\left|\langle f_{\lambda}(Y^M_n),g(Y^M_n)\Delta W^M_n\rangle\right|+2\Delta t\left|\langle f_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle\right|\nonumber\\
&+&2\langle g(Y^M_n)\Delta W^M_n, h(Y^M_n)\Delta\overline{N}^M_n\rangle.
\label{ch4ine14}
\end{aligned}$$ Using the estimation $2ab\leq a^2+b^2$, inequality becomes : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq & ||Y^M_n||^2+\Delta t^2||f_{\lambda}(Y^M_n)||^2+||g(Y^M_n)||^2||\Delta W^M_n||^2+||h(Y^M_n)||^2|\Delta\overline{N}^M_n|^2\nonumber\\
&+& 2\Delta t\left|\langle Y^M_n,f_{\lambda}(Y^M_n)\rangle\right|+2\langle Y^M_n,g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&+&\Delta t^2||f_{\lambda}(Y^M_n)||^2+||g(Y^M_n)||^2||\Delta W^M_n||^2+\Delta t^2||f_{\lambda}(Y^M_n)||^2\nonumber\\
&+&||h(Y^M_n)||^2|\Delta\overline{N}^M_n|^2 +||g(Y^M_n)||^2||\Delta W^M_n||^2+||h(Y^M_n)||^2|\Delta\overline{N}^M_n|^2.
\label{ch4ine15}
\end{aligned}$$ Putting similars terms of inequality together, we obtain : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq & ||Y^M_n||^2+3\Delta t^2||f_{\lambda}(Y^M_n)||^2+3||g(Y^M_n)||^2||\Delta W^M_n||^2+3||h(Y^M_n)||^2|\Delta\overline{N}^M_n|^2\nonumber\\
&+&2\Delta t\left|\langle Y^M_n,f_{\lambda}(Y^M_n)\rangle\right|+ 2\langle Y^M_n,g(Y^M_n)\Delta W^M_n\rangle\nonumber\\
&+&2\langle Y^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle
\label{ch4ine16}
\end{aligned}$$ on $\Omega$, for all $M\in\mathbb{N}$ and all $n\in\{0,1,\cdots, M-1\}$.
In addition, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, the global Lipschitz condition satisfied by $g$ and $h$ leads to : $$\begin{aligned}
||g(x)||^2&\leq& (||g(x)-g(0)||+||g(0)||)^2\nonumber\\
&\leq & (C||x||+||g(0)||)^2\nonumber\\
&\leq & (C+||g(0)||)^2||x||^2\nonumber\\
&\leq &\beta||x||^2.
\label{ch4normdeg2}
\end{aligned}$$ Along the same lines as above, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, we have : $$\begin{aligned}
||h(x)||^2\leq \beta||x||^2.
\label{ch4normdeh2}
\end{aligned}$$ Also, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, the one-sided Lipschitz condition satisfied by $f_{\lambda}$ leads to : $$\begin{aligned}
\langle x,f_{\lambda}(x)\rangle&=&\langle x,f_{\lambda}(x)-f_{\lambda}(0)+f_{\lambda}(0)\rangle=\langle x,f_{\lambda}(x)-f_{\lambda}(0)\rangle+\langle x,f_{\lambda}(0)\rangle\nonumber\\
&\leq & C||x||^2+||x||||f_{\lambda}(0)||\nonumber\\
&\leq &(C+||f_{\lambda}(0)||)||x||^2\nonumber\\
&\leq &\sqrt{\beta}||x||^2.
\label{ch4crochetf}
\end{aligned}$$ Futhermore, for all $x\in\mathbb{R}^d$ such that $1\leq ||x||\leq M^{1/2c}$ and for all $M\in\mathbb{N}$, using the polynomial growth condition of $f_{\lambda}$, the following inequality holds $$\begin{aligned}
||f_{\lambda}(x)||^2&\leq&\left(||f_{\lambda}(x)-f_{\lambda}(0)||+||f_{\lambda}(0)||\right)^2\nonumber\\
&\leq & \left(C(K+||x||^c)||x||+||f_{\lambda}(0)||\right)^2\nonumber\\
&\leq & \left(C(K+1)||x||^{c+1}+||f_{\lambda}(0)||\right)^2\nonumber\\
&\leq & (KC+C+||f_{\lambda}(0)||)^2||x||^{2(c+1)}\nonumber\\
&\leq & M\sqrt{\beta}||x||^2.
\label{ch4normf2}
\end{aligned}$$ Now combining inequalities , , , and , we obtain : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq& ||Y^M_n||^2+\dfrac{3T^2\sqrt{\beta}}{M}||Y^M_{n}||^2+3\beta||Y^M_n||^2||\Delta W^M_n||^2+3\beta||Y^M_n||^2|\Delta \overline{N}^M_n|^2\nonumber\\
&+&\dfrac{2T\sqrt{\beta}}{M}||Y^M_n||^2+2\langle Y^M_n, g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n, h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&\leq &||Y^M_n||^2+\dfrac{(3T^2+2T)\sqrt{\beta}}{M}||Y^M_n||^2+3\beta||Y^M_n||^2||\Delta W^M_n||^2+3||Y^M_n||^2|\Delta\overline{N}^M_n|^2\nonumber\\
&+&2\langle Y^M_n, g(Y^M_n)\Delta W^M_n\rangle
+2\langle Y^M_n, h(Y^M_n)\Delta\overline{N}^M_n\rangle.
\end{aligned}$$ Using the inequality $3T^2+2T\leq 3\sqrt{\beta}$, it follows that : $$\begin{aligned}
||Y^M_{n+1}||^2&\leq& ||Y^M_n||^2+\dfrac{3\beta}{M}||Y^M_n||^2+3\beta||Y^M_n||^2||\Delta W^M_n||^2+3\beta||Y^M_n||^2|\Delta\overline{N}^M_n|^2\nonumber\\
&+&2\langle Y^M_n,g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&=&||Y^M_n||^2 \left(1+\dfrac{3\beta}{M}+3\beta||\Delta W^M_n||^2+3\beta||\Delta\overline{N}^M_n||^2+2\left<\dfrac{Y^M_n}{||Y^M_n||}, \dfrac{g(Y^M_n)}{||Y^M_n||}\Delta W^M_n\right> \right.\nonumber\\
&+&\left. 2\left\langle\dfrac{Y^M_n}{||Y^M_n||}, \dfrac{h(Y^M_n)}{||Y^M_n||}\Delta\overline{N}^M_n\right\rangle\right)\nonumber\\
&=&||Y^M_n||^2\left(1+\dfrac{3\beta}{M}+3\beta||\Delta W^M_n||^2+3\beta|\Delta\overline{N}^M_n|^2+2\alpha^M_n+2\beta^M_n\right).
\label{ch4expY1}
\end{aligned}$$ Using Lemma \[ch4lemma1\] for $a=\dfrac{3\beta}{M}+3\beta||\Delta W^M_n||^2+2\alpha^M_n+2\beta^M_n$ and $b=\sqrt{3\beta}|\Delta\overline{N}^M_n| $ it follows from that : $$\begin{aligned}
||Y^M_{n+1}||^2\leq ||Y^M_n||^2\exp\left(\dfrac{3\beta}{M}+3\beta||\Delta W^M_n||^2+3\beta|\Delta\overline{N}^M_n|+2\alpha^M_n+2\beta^M_n\right)
\label{ch4expY2}
\end{aligned}$$ on $\{w\in\Omega : 1\leq ||Y^M_n(\omega)||\leq M^{1/2c}\}$, for all $M\in\mathbb{N}$ and all $n\in\{0,1,\cdots,M-1\}$.
In order to complete our proof, we need the following map $$\begin{aligned}
\tau^M_l : \Omega \longrightarrow\{-1,0,1,\cdots,l\},\hspace{0.5cm}l\in\{0,1,\cdots, M\},
\end{aligned}$$ such that : $$\begin{aligned}
\tau^M_l(\omega) :=\max\left(\{-1\}\cup\{n\in\{0,1,\cdots,l-1\} : ||Y^M_n(\omega)||\leq 1\}\right),
\end{aligned}$$ for all $\omega\in\Omega$, $M\in\mathbb{N}$ and all $l\in\{0,1,\cdots,M\}$.
For $M\in\mathbb{N}$ fixed we prove by induction on $n\in\{0,1,\cdots,M\}$ that $$\begin{aligned}
\mathbf{1}_{\Omega^M_n}||Y^M_n||\leq D^M_n.
\end{aligned}$$ \[ch4borneD1\]
- For $n=0$, $D_0^M =(\beta+||X_0||)\exp(\beta)$ and $||Y^M_0||=||X_0||$. Since $\beta\geq 1 $ we have $\exp(\beta)\geq 1$. So the following inequality holds $$\begin{aligned}
\mathbf{1}_{\Omega^M_0}||Y^M_0||\leq D^M_0.
\end{aligned}$$
- Let $l\in\{0,1,\cdots,M-1\}$ be arbitrary and let’s assume that
$\mathbf{1}_{\Omega^M_n}||Y^M_n||\leq D^M_n$ for all $n\in\{0,1,\cdots,l\}$. We want to prove that inequality holds for $n=l+1$.
Let $\omega\in\Omega^M_{l+1}$ we have to prove that $||Y^M_{l+1}(\omega)||\leq D^M_{l+1}(\omega)$.
Since $(\Omega^M_n)$ is a decreasing sequence and $\omega\in\Omega^M_{l+1}$, we have $\omega\in\Omega^M_k$ and it follows from the hypothesis of induction that : $||Y^M_k(\omega)||\leq D^M_k(\omega)$, for all $k\in\{0,\cdots, l\}$.
Also, since $\omega\in\Omega^M_{k+1}$, by definition of $\Omega^M_{k+1}$ it follows that $D^M_k(\omega)\leq M^{1/2c}$, for all $k\in\{0, \cdots, l\}$. So for all $k\in\{0,1,\cdots,l\}$, $$\begin{aligned}
||Y^M_k(\omega)||\leq D^M_k(\omega)\leq M^{1/2c}.
\end{aligned}$$ For all $k\in\{\tau^M_{l+1}(\omega)+1,\tau^M_{l+1}(\omega)+2,\cdots,l\}$ we have $$\begin{aligned}
1\leq ||Y^M_k(\omega)||\leq M^{1/2c}.
\label{ch4con}
\end{aligned}$$ Since holds, it follows from , that $$\begin{aligned}
||Y^M_{k+1}(\omega)||&\leq& ||Y^M_k(\omega)||\exp\left(\dfrac{3\beta}{2M}+\dfrac{3\beta}{2}||\Delta W^M_k(\omega)||^2\right.\\
&+&\left.\dfrac{3\beta}{2}|\Delta\overline{N}^N_k(\omega)|+\alpha^M_k(\omega)+\beta^M_k(\omega)\right)
\end{aligned}$$ for all $k\in\{\tau^M_{l+1}(\omega)+1,\tau^M_{l+1}+2,\cdots,l\}$.
For $k=l$ from the previous inequality, we have : $$\begin{aligned}
||Y^M_{l+1}(\omega)||&\leq &||Y^M_l(\omega)||\exp\left(\dfrac{3\beta}{2M}+\dfrac{3\beta}{2}||\Delta W^M_l(\omega)||^2+\dfrac{3\beta}{2}|\Delta\overline{N}^N_l(\omega)|\right.\nonumber\\
&+&\left.\alpha^M_l(\omega)+\beta^M_l(\omega)\right).
\label{ch4iter}
\end{aligned}$$ Iterating $l-\tau^M_{l+1}(\omega)$ times leads to $$\begin{aligned}
||Y^M_{l+1}(\omega)|| &\leq &||Y^M_{\tau^M_{l+1}(\omega)+1}(\omega)||\exp\left(\sum_{m=\tau^M_{l+1}(\omega)+1}^l\left[\dfrac{3\beta}{2M}+\dfrac{3\beta}{2}||\Delta W^M_m(\omega)||^2\right.\right.\\
&+&\left.\left.\dfrac{3\beta}{2}|\Delta\overline{N}^M_m(\omega)|+\alpha^M_m(\omega)+\beta^M_m(\omega)\right]\right).
\end{aligned}$$ By definition of $\tau^M_l(\omega)$, we have $||Y^M_{\tau^M_{l+1}(\omega)}(\omega)||\leq 1$.
Then it follows from that $||Y^M_{\tau^M_{l+1}(\omega)+1}(\omega)||\leq \beta$. So the above estimation of $||Y^M_{l+1}(\omega)||$ becomes : $$\begin{aligned}
||Y^M_{l+1}(\omega)||&\leq& \beta\exp\left(\sum_{m=\tau^M_{l+1}(\omega)+1}^l\left[\dfrac{3\beta}{2M}\right.\right.\\
&+&\dfrac{3\beta}{2}||\Delta W^M_m(\omega)||^2+\dfrac{3\beta}{2}|\Delta\overline{N}^M_m(\omega)|+\left.\left.\alpha^M_m(\omega)+\beta^M_m(\omega)\right]\right)\\
&\leq& (\beta+||X_0||)\exp\left(\dfrac{3\beta}{2}+\sup_{u\in\{0,1,\cdots,l+1\}}\sum_{m=u}^l\left[\dfrac{3\beta}{2}||\Delta W^M_m(\omega)||^2\right.\right.\\
&+&\left.\left.\dfrac{3\beta}{2}|\Delta\overline{N}^M_m(\omega)|+\alpha^M_m(\omega)+\beta^M_m(\omega)\right]\right)=D^M_{l+1}(\omega).
\end{aligned}$$ Therefore $||Y^M_{l+1}(\omega)||\leq D^M_{l+1}(\omega)$. This complete the proof of Lemma \[ch4lemma2\].
The following is from [@Martin1 Lemma 3.2 pp 15].
\[ch4lemma3\] Let $n\in\mathbb{N}$ and $Z : \Omega \longrightarrow \mathbb{R}^m$ be an $m-$dimensional standard normal random variable. Then for all $a\in\left[0,\dfrac{1}{4}\right]$ the following inequality holds $$\begin{aligned}
\mathbb{E}\left[\exp(a||Z||^2)\right]=(1-2a)^{-m/2}\leq e^{2am}.
\end{aligned}$$
Using the relation $||Z||^2=|Z_1|^2+|Z_2|^2+\cdots+|Z_n|^2$ and the fact that $(Z_i)$ are independent and identically distributed, we have : $$\begin{aligned}
\mathbb{E}\left[\exp(a||Z||^2)\right]=\mathbb{E}\left[\exp\left(\sum_{i=1}^ma|Z_i|^2\right)\right]=\mathbb{E}\left[\prod_{i=1}^m\exp\left(a|Z_i|^2\right)\right]=\left[\mathbb{E}\left(\exp(a|Z_1|^2)\right)\right]^m.
\label{ch4normal1}
\end{aligned}$$ From the definition of the expected value of the standard normal random variable, we have : $$\begin{aligned}
\mathbb{E}[\exp(a|Z_1|^2)]=\int_{-\infty}^{+\infty}e^{ax^2}\dfrac{1}{\sqrt{2\pi}}e^{-x^2/2}dx=\dfrac{1}{\sqrt{1-2a}}.
\end{aligned}$$ Using the inequality $\dfrac{1}{1-x}\leq e^{2x} \hspace{0.3cm }\forall x\in\left[0,\dfrac{1}{2}\right]$, it follows that $$\begin{aligned}
\mathbb{E}[\exp(a|Z_1|^2)]=\dfrac{1}{\sqrt{1-2a}}\leq e^{2a},\hspace{0.5cm}\forall a\in\left[0,\dfrac{1}{4}\right].
\label{ch4normal2}
\end{aligned}$$ Combining and leads to : $$\begin{aligned}
\mathbb{E}\left[\exp(a||Z||^2)\right]=(1-2a)^{-m/2}\leq e^{2am},\hspace{0.5cm}\forall a\in\left[0,\dfrac{1}{4}\right].
\end{aligned}$$
The following lemma and its proof are based on with only different value of the coefficient $\beta$.
\[ch4lemma4\] The following inequality holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}, M\geq 4\beta pT}\mathbb{E}\left[\exp\left(\beta p\sum_{k=0}^{M-1}||\Delta W^M_k||^2\right)\right]<\infty.
\end{aligned}$$
Let $Z=\mathcal{N}(0,1)$ be an $m$-dimensional standard normal random variable. Since for
$k=0,\cdots, M-1$, $\Delta W^M_k$ are independent, stationary and follows the normal distribution with mean $0$ and variance $\dfrac{T}{M}$, $||\Delta W^M_n||^2=\dfrac{T}{M}||Z||^2$, it follows that : $$\begin{aligned}
\mathbb{E}\left(\exp\left[\beta p\sum_{k=0}^{M-1}||\Delta W^M_k||^2\right]\right)&=&\prod_{k=0}^{M-1}\mathbb{E}\left[\exp(\beta p||\Delta W^M_k||^2)\right]\\
&=&\left(\mathbb{E}\left[\exp\left(\beta p\dfrac{T}{M}||Z||^2\right)\right]\right)^M \\
&\leq &\left[\exp\left(2\beta pm\dfrac{T}{M}\right)\right]^M\text{(using Lemma \ref{ch4lemma3})}\\
&\leq &\exp(2\beta pTm)<\infty,
\end{aligned}$$ for all $p\in[1,+\infty)$ and all $M\in\mathbb{N}\cap[4\beta pT, \infty)$.
\[ch4lemma5\] Let $Y$ be a standard normal random of dimension $m$ variable and $c\in\mathbb{R}^m$, then $$\begin{aligned}
\mathbb{E}[\exp(cY)]=\exp\left(\dfrac{c^2}{2}\right).
\end{aligned}$$
$\mathbb{E}[\exp(cY)]$ is the moment generating function of $Y$ at $c$. Since the mean $\mu=0$ and the standard deviation $\sigma=1$, it follows directly that $$\begin{aligned}
\mathbb{E}[\exp(cY)]=\exp\left(\mu+\dfrac{1}{2}\sigma^2c^2\right)=\exp\left(\dfrac{c^2}{2}\right).
\end{aligned}$$
The following lemma is from [@Martin2 Lemma 5.7, pp 15].
\[ch4lemma6\] The following inequality holds $$\begin{aligned}
\mathbb{E}\left[\left|pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right|^2\right]\leq \exp\left[\dfrac{p^2T(C+||g(0)||)^2}{M}\right],
\end{aligned}$$ for all $x\in\mathbb{R}^d, k\in\{0,1,\cdots,M-1\}, p\in[1,\infty)$ and all $z\in\{-1,1\}$.
Let the notation $a^{\top}$ stand for the transposed of a vector $a$ and $Y$ the $m$ column vector define by $Y=\sqrt{\dfrac{T}{M}}(1,\cdots, 1)$. Then we have : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]&=&\mathbb{E}\left[\exp\left(pz\dfrac{g(x)^{\top}x}{||x||^2}\Delta W^M_k\right)\right]\\
&=&\mathbb{E}\left[\exp\left(pz\dfrac{g(x)^{\top}x}{||x||^2}\sqrt{\dfrac{T}{M}}\mathcal{N}(0,1)\right)\right]\\
&=&\mathbb{E}\left[\exp\left(pzY\dfrac{g(x)^{\top}x}{||x||^2}\right)\right]
\end{aligned}$$ Using Lemma \[ch4lemma6\], it follows that $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]&=&\exp\left[\dfrac{1}{2}\left|pz\dfrac{g(x)^{\top}x}{||x||^2}Y\right|^2\right]\\
&\leq &\exp\left[p^2\dfrac{||g(x)||^2}{||x||^2}||Y||^2\right]\\
&\leq &\exp\left[\dfrac{p^2T}{M}\dfrac{||g(x)||^2}{||x||^2}\right].
\end{aligned}$$ Using the global Lipschitz condition and the fact that $||x||\geq 1$, we have : $$\begin{aligned}
\dfrac{||g(x)||^2}{||x||^2}\leq\dfrac{(||g(x)-g(0)||+||g(0)||)^2}{||x||^2}\leq\dfrac{(C+||g(0)||)^2||x||^2}{||x||^2}\leq (C+||g(0)||)^2.
\end{aligned}$$ Therefore, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, we have $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]\leq \exp\left[\dfrac{p^2T(C+||g(0)||)^2}{M}\right],
\end{aligned}$$ for all $M\in\mathbb{N}$, $k\in\{0,\cdots,M-1\}$, all $p\in[1,\infty)$ and $z\in\{-1,1\}$.
Following closely we have the following lemma.
\[ch4lemma7\] Let $\alpha^M_n :\Omega\longrightarrow\mathbb{R}$ for $M\in\mathbb{N}$ and $n\in\{0,1,\cdots,M\}$ defined in Notation \[ch4notation1\], then the following inequality holds : $$\begin{aligned}
\sup_{z\in\{-1,1\}}\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,1,\cdots,M\}}\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^p(\Omega, \mathbb{R})}<\infty,
\end{aligned}$$ for all $p\in[2,+\infty)$.
The time discrete stochastic process $z\sum_{k=0}^{n-1}\alpha^M_k$, $n\in\{0,1,\cdots, M\}$ is an $(\mathcal{F}_{nT/M})_{n\in\{0,\cdots,M\}}-$ martingale for every $z\in\{-1,1\}$ and $M\in \mathbb{N}$. So $\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)$ is a positive $(\mathcal{F}_{nT/M})_{n\in\{0,\cdots,M\}}-$ submartingale for every $z\in\{-1,1\}$ and $M\in\mathbb{N}$ since $\exp$ is a convex function.
Applying Doop’s maximal inequality leads to : $$\begin{aligned}
\left\|\sup_{n\in\{0,\cdots,M\}}\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^p{(\Omega, \mathbb{R})}}&=&\left(\mathbb{E}\left|\sup_{n\in\{0,\cdots,M\}}\exp\left(pz\sum_{k=0}^{n-1}\alpha^M_k\right)\right|\right)^{1/p}\nonumber\\
&\leq &\left(\dfrac{p}{p-1}\right)\left(\mathbb{E}\left |\exp\left(pz\sum_{k=0}^{M-1}\alpha^M_k\right)\right|\right)^{1/p}\nonumber\\
&= &\dfrac{p}{p-1}\left\|\exp\left(z\sum_{k=0}^{M-1}\alpha^M_k\right)\right\|_{L^p{(\Omega,\mathbb{R})}}.
\label{ch4alpha1}
\end{aligned}$$ Using Lemma \[ch4lemma6\], it follows from the previous inequality that : $$\begin{aligned}
\mathbb{E}\left[\exp(pz\alpha^M_k)/\mathcal{F}_{kT/M}\right]\leq\exp\left(\dfrac{p^2T(C+||g(0)||)^2}{M}\right).
\label{ch4alpha2}
\end{aligned}$$ Using inequality , it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-1}\alpha^M_k\right)\right]&=&\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-2}
\alpha^M_k\right)\mathbb{E}[\exp(p\alpha^M_{M-1}/\mathcal{F}_{(M-1)T/M}\right]\\
&\leq &\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-2}\alpha^M_k\right)\right]\exp\left(\dfrac{p^2T(C+||g(0)||)^2}{M}\right).
\end{aligned}$$ Iterating the previous inequality $M$ times gives : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-1}\alpha^M_k\right)\right] \leq \exp(p^2T(C+||g(0)||)^2).
\label{ch4alpha3}
\end{aligned}$$ Now combining inequalities and leads to $$\begin{aligned}
\sup_{z\in\{-1,1\}}\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,\cdots,M\}}\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^p(\Omega,\mathbb{R})} \leq 2\exp(p^2T(C+||g(0)||)^2) <\infty,
\end{aligned}$$ for all $p\in[2,\infty)$.
\[ch4lemma8\] For all $c\in\mathbb{R}$, we have : $$\begin{aligned}
\mathbb{E}[\exp(c\Delta \overline{N}^M_n)]=\exp\left[\dfrac{(e^c+c-1)\lambda T}{M}\right],
\end{aligned}$$ for all $M\in\mathbb{N}$ and all $n\in\{0,\cdots,M\}$.
It is known that if $Y$ is a random variable following the poisson law with parameter $\lambda$, then its moment generating function is given by : $$\begin{aligned}
\mathbb{E}[\exp(cY)]=\exp(\lambda(e^c-1)).
\end{aligned}$$ \[gene\] Since $\Delta N_n$ follows a poisson law with parameter $\lambda\Delta t$, it follows that $$\begin{aligned}
\mathbb{E}[\exp(c\Delta\overline{N}^M_n)]&=&\mathbb{E}[\exp(c\Delta N^M_n+c\lambda\Delta t)]\\
&=&\mathbb{E}\left[\exp\left(\dfrac{\lambda T}{M}\right)\exp(c\Delta N^M_n)\right]\\
&=&\exp\left(\dfrac{c\lambda T}{M} \right)\exp\left[ \dfrac{\lambda T}{M}(e^c-1)\right]\\
&=&\exp\left[\dfrac{(e^c+1-1)\lambda T}{M}\right].
\end{aligned}$$
\[ch4lemma9\] The following inequality holds $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right]\leq\exp\left[\dfrac{\lambda \left(e^{p(C+||h(0)||)}+p(C+||h(0)||\right)}{M}\right],
\end{aligned}$$ for all $M\in\mathbb{N}$, $z\in\{-1,1\}$, all $p\in[1, +\infty)$ and all $n\in\{0,\cdots,M\}$.
For $x\in\mathbb{R}^d$ such that $||x||\neq 0$, we have : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right] &\leq& \mathbb{E}\left[\exp\left(pz\dfrac{||x||||h(x)||}{||x||^2}\Delta\overline{N}^M_n\right)\right]\\
&=& \mathbb{E}\left[\exp\left(pz\dfrac{||h(x)||}{||x||}\Delta \overline{N}^M_n\right)\right].
\end{aligned}$$ For all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, using the global Lipschitz condition satisfied by $h$, we have : $$\begin{aligned}
\dfrac{||h(x)||}{||x||}\leq \dfrac{||h(x)-h(0)||+||h(0)||}{||x||}\leq C+||h(0)||. \label{ch4normeh}
\end{aligned}$$ So from inequality and using Lemma \[ch4lemma8\] it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right] &\leq &\mathbb{E}[\exp(pz(C+||h(0)||)\Delta\overline{N}^M_n)]\\
&\leq &\exp\left[\dfrac{\left(e^{p(C+||h(0)||)}+p(C+||h(0)||-1\right)\lambda T}{M}\right]\\
&\leq &\exp\left[\dfrac{\left(e^{p(C+||h(0)||)}+p(C+||h(0)||\right)\lambda T}{M}\right].
\end{aligned}$$
\[ch4lemma10\] Let $\beta^M_n :\Omega \longrightarrow \mathbb{R}$ define as in Notation \[ch4notation1\] for all $M\in\mathbb{N}$ and all $n\in\{0,\cdots,M\}$, then we have the following inequality $$\begin{aligned}
\sup_{z\in\{-1,1\}}\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(z\sum_{K=0}^{n-1}\beta^M_k\right)\right\|_{L^p(\Omega, \mathbb{R})}<+\infty.
\end{aligned}$$
For the same reason as for $\alpha^M_k$, $\beta^M_k$ is an $(\mathcal{F}_{nT/M})$- martingale. So $\exp\left(pz\sum_{k=0}^{n-1}\beta^M_k\right)$ is a positive $(\mathcal{F}_{nT/M})$- submartingale for all $M\in\mathbb{N}$ and all $n\in\{0,\cdots, M\}$. Using Doop’s maximal inequality we have : $$\begin{aligned}
\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(z\sum_{k=0}^{n-1}\beta^M_k\right)\right\|_{L^p(\Omega, \mathbb{R})} &\leq &\left(\dfrac{p}{p-1}\right)\left\|\exp\left(z\sum_{k=0}^{M-1}\beta^M_k\right)\right\|_{L^p(\Omega,\mathbb{R})},
\label{ch4beta1}
\end{aligned}$$ $$\begin{aligned}
\left\|\exp\left(z\sum_{k=0}^{M-1}\beta^M_k\right)\right\|^p_{L^p(\Omega,\mathbb{R})}&=&\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-1}\beta^M_k\right)\right]=\mathbb{E}
\left[\exp\left(pz\left(\sum_{k=0}^{M-2}\beta^M_k\right)+pz\beta_{M-1}^M\right)\right]\\
&=&\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-2}\beta^M_k\right)
\mathbb{E}\left[\exp\left(pz\beta^M_{M-1}\right)/\mathcal{F}_{(M-1)T/M}\right]
\right].
\end{aligned}$$ Using Lemma \[ch4lemma9\] it follows that $$\begin{aligned}
\left\|\exp\left(z\sum_{k=0}^{M-1}\beta^M_k\right)\right\|^p_{L^p(\Omega,\mathbb{R})} \leq \mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-2}\beta^M_k\right)\right]\exp\left[\dfrac{\left(e^{p(C+||h(0)||)}+p(C+||h(0)||)\right)\lambda T}{M}\right].
\end{aligned}$$ Iterating this last inequality $M$ times leads to : $$\begin{aligned}
\left(\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-1}\beta^M_k\right)\right]\right)^p\leq \exp\left[\lambda T\left(e^{p(C+||h(0)||)}+Tp(C+||h(0)||\right)\right],
\label{ch4beta2}
\end{aligned}$$ for all $M\in\mathbb{N}$, all $p\in (1,\infty)$ and all $z\in\{-1,1\}$.
Combining inequalities and complete the proof of Lemma \[ch4lemma10\]
\[ch4lemma11\] The following inequality holds $$\begin{aligned}
\sup_{M\in\mathbb{N}}\mathbb{E}\left[\exp\left(p\beta\sum_{k=0}^{M-1}||\Delta\overline{N}^M_k||\right)\right]<+\infty,\end{aligned}$$ for all $p\in[1, +\infty)$.
Using independence, stationarity of $\Delta\overline{N}_k^M$ and Lemma \[ch4lemma8\], it follows that : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\mathbb{E}\left[\exp\left(p\beta\sum_{k=0}^{M-1}||\Delta\overline{N}^M_k||\right)\right]&=&\prod_{k=0}^{M-1}\mathbb{E}[\exp(p\beta||\Delta\overline{N}^M_k||)]\\
&=&\left(\mathbb{E}[\exp(p\beta||\Delta\overline{N}^M_k||)]\right)^M\\
&=&\left(\exp\left[\dfrac{(e^{p\beta}+p\beta-1)\lambda T}{M}\right]\right)^M\\
&=&\exp[e^{p\beta}+p\beta-1]<+\infty,
\end{aligned}$$ for all $p\in[1, +\infty)$.
Inspired by [@Martin1 Lemma 3.5, pp 15], we have the following estimation.
\[ch4lemma12\] \[Uniformly bounded moments of the dominating stochastic processes\].
Let $M\in\mathbb{N}$ and $D_n^M : \Omega \longrightarrow [0,\infty)$ for $n\in\{0,1,\cdots, M\}$ be define as above, then we have : $$\begin{aligned}
\sup_{M\in\mathbb{N}, M\geq8\lambda pT}\left\|\sup_{n\in\{0,1,\cdots, M\}}D_n^M\right\|_{L^p(\Omega, \mathbb{R})}<\infty,
\end{aligned}$$ for all $p\in[1,\infty)$.
Let’s recall that : $$\begin{aligned}
D_n^M =(\beta+||\varepsilon||)\exp\left(\dfrac{3\beta}{2}+\sup_{u\in\{0,\cdots,n\}}\sum_{k=u}^{n-1}\dfrac{3\beta}{2}||\Delta W^M_k||^2+\dfrac{3\beta}{2}|\Delta\overline{N}^M_k|+\alpha^M_k+\beta^M_k\right).
\end{aligned}$$ Using Holder inequality, it follows that : $$\begin{aligned}
\sup_{M\in\mathbb{N}, M\geq 8\lambda pT}\left\|\sup_{n\in\{0,\cdots, M\}}D_n^M\right\|_{L^p(\Omega, \mathbb{R})}&\leq &e^{3\beta/2}\left(\beta+||\varepsilon||_{L^{4p}(\Omega, \mathbb{R})}\right)\\
&\times& \sup_{M\in\mathbb{N}, M\geq 8\lambda pT}\left\|\exp\left(\dfrac{3\beta}{2}\sum_{k=0}^{M-1}||\Delta W^M_k||^2\right)\right\|_{L^{2p}(\Omega, \mathbb{R})}\\
&\times &\sup_{M\in\mathbb{N}}\left\|\exp\left(\dfrac{3\beta}{2}\sum_{k=0}^{M-1}|\Delta\overline{N}_k^M|\right)\right\|_{L^{8p}(\Omega, \mathbb{R})}\\
&\times &\left(\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(\sup_{u\in\{0,\cdots,n\}}\sum_{k=u}^{n-1}\alpha_k^M\right)\right\|_{L^{16p}(\Omega, \mathbb{R})}\right)\\
&\times &\left(\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(\sup_{u\in\{0,\cdots,n\}}\sum_{k=u}^{n-1}\beta_k^M\right)\right\|_{L^{16p}(\Omega, \mathbb{R})}\right)\\
&=& A_1\times A_2\times A_3\times A_4\times A_5.
\end{aligned}$$ By assumption $A_1$ is bounded. Lemma \[ch4lemma4\] and \[ch4lemma11\] show that $A_2$ and $A_3$ are bounded. Using again Holder inequality and Lemma \[ch4lemma7\] it follows that : $$\begin{aligned}
A_4=\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(\sup_{u\in\{0,\cdots,n\}}\sum_{k=u}^{n-1}\alpha_k^M\right)\right\|_{L^{16p}(\Omega, \mathbb{R})}
\end{aligned}$$ $$\begin{aligned}
\leq\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^{32p}(\Omega,\mathbb{R})}
\times \left\|\sup_{u\in\{0,\cdots, M\}}\exp\left(-\sum_{k=0}^{u-1}\alpha^M_k\right)\right\|_{L^{32p}(\Omega,\mathbb{R})} <+\infty,
\end{aligned}$$ for all $M\in\mathbb{N}$ and all $p\in[1,\infty)$.
Along the same lines as above, we prove that $A_5$ is bounded.
Since each of the terms $A_1, A_2, A_3, A_4$ and $A_5$ is bounded, this complete the proof of Lemma \[ch4lemma12\].
The following lemma is an extension of [@Martin1 Lemma 3.6, pp 16]. Here, we include the jump part.
\[ch4lemma13\] Let $M\in\mathbb{N}$ and $\Omega_M^M\in\mathcal{F}$. The following holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\left(M^p\mathbb{P}[(\Omega_M^M)^c]\right)<+\infty,
\end{aligned}$$ for all $p\in[1,\infty)$.
Using the subadditivity of the probability measure and the Markov’s inequality, it follows that $$\begin{aligned}
\mathbb{P}[(\Omega_M^M)^c] &\leq & \mathbb{P}\left[\sup_{n\in\{0,\cdots, M-1\}}D_n^M>M^{1/2c}\right]+M\mathbb{P}\left[\|W_{T/M}\|>1\right]+M\mathbb{P}\left[|\overline{N}_{T/M}|>1\right]\nonumber\\
&\leq & \mathbb{P}\left[\sup_{n\in\{0,\cdots, M-1\}}|D_n^M|>M^{q/2c}\right]+M\mathbb{P}\left[\|W_{T}\|>\sqrt{M}\right]+M\mathbb{P}\left[|\overline{N}_{T}|>M\right]\nonumber\\
&\leq & \mathbb{P}\left[\sup_{n\in\{0,\cdots, M-1\}}|D_n^M|>M^{q/2c}\right]+M\mathbb{P}\left[\|W_{T}\|^q>M^{q/2}\right]+M\mathbb{P}\left[|\overline{N}_{T}|^q>M^q\right]\nonumber\\
&\leq &\mathbb{E}\left[\sup_{n\in\{0,\cdots, M-1\}}|D_n^M|^q\right]M^{-q/2c}+\mathbb{E}[\|W_T\|^q]M^{1-q/2}+ \mathbb{E}[|\overline{N}_{T}|^q] M^{1-q} \nonumber,\\
\end{aligned}$$ for all $q>1$.
Multiplying both sides of the above inequality by $M^p$ leads to $$\begin{aligned}
M^p\mathbb{P}[(\Omega_M^M)^c]\leq \mathbb{E}\left[\sup_{n\in\{0,\cdots, M-1\}}|D_n^M|^q\right]M^{p-q/2c}+\mathbb{E}[\|W_T\|^q]M^{p+1-q/2}+\mathbb{E}[|\overline{N}_{T}|^q] M^{p+1-q}
\end{aligned}$$ for all $q>1$.
For $q>\max\{2pc, 2p+2\}$, we have $M^{p+1-q/2}<1$, $M^{p-q/2c}<1$ and $ M^{p+1-q}<1$. It follows for this choice of $q$ that $$\begin{aligned}
M^p\mathbb{P}[(\Omega_M^M)^c]\leq \mathbb{E}\left[\sup_{n\in\{0,\cdots, M-1\}}|D_n^M|^p\right]+\mathbb{E}[\|W_T\|^q]+\mathbb{E}[|\overline{N}_{T}|^q].
\end{aligned}$$ Using Lemma \[ch4lemma12\] and the fact that $W_T$ and $\overline{N}_{T}$ are independents of $M$, it follows that $$\begin{aligned}
\sup_{M\in\mathbb{N}}\left(M^p\mathbb{P}[(\Omega_M^M)^c]\right)<+\infty.
\end{aligned}$$
The following lemma can be found in [@Protter Theorem 48 pp 193] or in [@Gundy Theorem 1.1, pp 1].
\[ch4lemma18b\]\[Burkholder-Davis-Gundy inequality\]
Let $M$ be a martingale with càdlàg paths and let $p\geq 1$ be fixed. Let $M_t^*=\sup\limits_{s\leq t}||M_s||$. Then there exist constants $c_p$ and $C_p$ such that for any $M$ $$\begin{aligned}
c_p \left[\mathbb{E}\left([M,M]_t\right)^{p/2}\right]^{1/p}\leq \left[\mathbb{E}(M_t^*)^p\right]^{1/p}\leq C_p\left[\mathbb{E}\left([ M, M]_t\right)^{p/2}\right]^{1/p},
\end{aligned}$$ for all $0\leq t\leq \infty$, where $[M,M]$ stand for the quadratic variation of the process $M$. The constants $c_p$ and $C_p$ are universal : They does not depend on the choice of $M$.
The following lemma can be found in [@Martin1 Lemma 3.7, pp 16].
\[ch4lemma15\]
Let $k\in\mathbb{N}$ and let $Z : [0,T]\times \Omega \longrightarrow \mathbb{R}^{k\times m}$ be a predictable stochastic process satisfying $\mathbb{P}\left[\int_0^T||Z_s||^2ds<+\infty\right]=1$. Then we have the following inequality $$\begin{aligned}
\left\|\sup_{s\in[0,t]}\left\|\int_0^sZ_udW_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\int_0^t\sum_{i=1}^m||Z_s\vec{e}_i||^2_{L^p(\Omega, \mathbb{R}^k)}ds\right)^{1/2}
\end{aligned}$$ for all $t\in[0,T]$ and all $p\in[1,\infty)$. Where $(\vec{e}_1,\cdots,\vec{e}_m)$ is the canonical basis of $\mathbb{R}^m$.
Since $W$ is a continuous martingale satisfying $d[ W,W]_s= ds$, it follows from the property of the quadratic variation (see [@Fima 8.21, pp 219] ) that $$\begin{aligned}
\left[ \int_0Z_sdW_s, \int_0Z_sdW_s\right]_t=\int_0^t||Z_s||^2d[W,W]_s=\int_0^t||Z_s||^2 ds.
\label{ch4Bur2}
\end{aligned}$$ Applying Lemma \[ch4lemma18b\] for $M_t=\sup\limits_{0\leq s\leq T}\int_0^tZ_sdW_s$ and using leads to :
$$\begin{aligned}
\left[\mathbb{E}\left[\sup_{0\leq t\leq T}\left\|\int_0^tZ_udW_u\right\|^p\right]\right]^{1/p}\leq C_p\left[\mathbb{E}\left(\int_0^T||Z_s||^2ds\right)^{p/2}\right]^{1/p},
\end{aligned}$$
where $C_p$ is a positive constant depending on $p$ :
Using the definition of $||X||_{L^p(\Omega, \mathbb{R})}$ for any random variable $X$, it follows that $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_udW_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}&\leq& C_p\left\|\int_0^T||Z_s||^2ds\right\|^{1/2}_{L^{p/2}(\Omega, \mathbb{R})}\\
&\leq&C_p\left\|\int_0^T\sum_{i=1}^m||Z_s.\vec{e}_i||^2ds\right\|^{1/2}_{L^{p/2}(\Omega, \mathbb{R})}.
\end{aligned}$$ Using Minkowski inequality in its integral form (see Proposition \[ch1Minkowski\]) yields : $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_udW_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\int_0^T\sum_{i=1}^m\left\|||Z_s.\vec{e}_i||^2\right\|_{L^{p/2}(\Omega, \mathbb{R}^k)}ds\right)^{1/2}.
\end{aligned}$$ Using Holder inequality, it follows that : $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_udW_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\int_0^T\sum_{i=1}^m||Z_s.\vec{e}_i||^2_{L^p(\Omega, \mathbb{R}^k)}ds\right)^{1/2}.
\end{aligned}$$ This complete the proof of the lemma.
The following lemma and its proof can be found in [@Martin1 Lemma 3.8, pp 16].
\[ch4lemma16\]
Let $k\in\mathbb{N}$ and let $ Z^M_l : \Omega \longrightarrow \mathbb{R}^{k\times m}$, $l\in\{0,1,\cdots, M-1\}$, $M\in\mathbb{N}$ be a familly of mappings such that $Z^M_l$ is $\mathcal{F}_{lT/M}/\mathcal{B}(\mathbb{R}^{k\times m})$-measurable for all $l\in\{0,1,\cdots,M-1\}$ and $M\in\mathbb{N}$. Then the following inequality holds : $$\begin{aligned}
\left\|\sup_{j\in\{0,1,\cdots,n\}}\left\|\sum_{l=0}^{j-1}Z_l^M\Delta W^M_l\right\|\right\|_{L^p(\Omega, \mathbb{R})} \leq C_p\left(\sum_{l=0}^{n-1}\sum_{i=1}^m||Z^M_l.\vec{e}_i||^2_{L^p(\Omega, \mathbb{R}^k)}\dfrac{T}{M}\right)^{1/2}.
\end{aligned}$$
Let $\overline{Z}^M : [0,T]\times\Omega \longrightarrow \mathbb{R}^{k\times m}$ such that $\overline{Z}_s :=Z^M_l$ for all $s\in\left[\dfrac{lT}{M},\dfrac{(l+1)T}{M}\right)$,
$l\in\{0,1,\cdots,M-1\}$ and all $M\in\mathbb{N}$.
Using Lemma \[ch4lemma15\], one obtain : $$\begin{aligned}
\left\|\sup_{j\in\{0,1,\cdots,n\}}\left\|\sum_{l=0}^{j-1}Z_l^M\Delta W^M_l\right\|\right\|_{L^p(\Omega, \mathbb{R})} &=&\left\|\sup_{j\in\{0,1,\cdots, n\}}\left\|\int_0^{jT/M}\overline{Z}^M_udW_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\\
&\leq & \left\|\sup_{s\in\left[0,\dfrac{nT}{M}\right]}\left\|\int_0^s\overline{Z}^M_udW_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\\
&\leq &C_p\left(\int_0^{nT/M}\sum_{i=1}^m||Z_s.\vec{e}_i||^2_{L^p(\Omega,\mathbb{R}^k)}ds\right)^{1/2}\\
&=&C_p\left(\sum_{l=0}^{n-1}\sum_{i=1}^m||Z^M_l.\vec{e}_i||^2_{L^p(\Omega, \mathbb{R}^k)}\dfrac{T}{M}\right)^{1/2}.
\end{aligned}$$
\[ch4lemma18\] Let $k\in\mathbb{N}$ and $Z :[0, T]\times\Omega \longrightarrow \mathbb{R}^k$ be a predictable stochastic process satisfying $\mathbb{E}\left(\int_0^T||Z_s||^2ds\right)<+\infty$. Then the following inequality holds : $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_ud\overline{N}_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\int_0^T||Z_s||^2_{L^p(\Omega, \mathbb{R}^k)}ds\right)^{1/2},
\end{aligned}$$ for all $t\in[0, T]$ and all $p\in[1,+\infty)$.
Since $\overline{N}$ is a martingale with càdlàg paths satisfying $d[ \overline{N},\overline{N}]_s=\lambda s$ (see Proposition \[ch1quadratic\]), it follows from the property of the quadratic variation (see [@Fima 8.21, pp 219]) that $$\begin{aligned}
\left[\int_0Z_sd\overline{N}_s, \int_0Z_sd\overline{N}_s\right]_t=\int_0^t||Z_s||^2\lambda ds.
\label{ch4Bur1}
\end{aligned}$$ Applying Lemma \[ch4lemma18b\] for $M_t=\sup\limits_{0\leq s\leq T}\int_0^tZ_sd\overline{N}_s$ and using leads to :
$$\begin{aligned}
\left[\mathbb{E}\left[\sup_{0\leq t\leq T}\left\|\int_0^tZ_ud\overline{N}_u\right\|^p\right]\right]^{1/p}\leq C_p\left[\mathbb{E}\left(\int_0^T||Z_s||^2ds\right)^{p/2}\right]^{1/p},
\end{aligned}$$
where $C_p$ is a positive constant depending on $p$ and $\lambda$.
Using the definition of $||X||_{L^p(\Omega, \mathbb{R})}$ for any random variable $X$, it follows that $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_ud\overline{N}_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left\|\int_0^T||Z_s||^2ds\right\|^{1/2}_{L^{p/2}(\Omega, \mathbb{R})}.
\end{aligned}$$ Using Minkowski inequality in its integral form (see Proposition \[ch1Minkowski\]) yields : $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_ud\overline{N}_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\int_0^T\left\|||Z_s||^2\right\|_{L^{p/2}(\Omega, \mathbb{R}^k)}ds\right)^{1/2}.
\end{aligned}$$ Using Holder inequality leads to : $$\begin{aligned}
\left\|\sup_{s\in[0,T]}\left\|\int_0^sZ_ud\overline{N}_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\int_0^T||Z_s||^2_{L^p(\Omega, \mathbb{R}^k)}ds\right)^{1/2}.
\end{aligned}$$ This complete the proof of the lemma.
\[ch4lemma19\] Let $k\in\mathbb{N}$, $M\in\mathbb{N}$ and $Z^M_l : \Omega \longrightarrow \mathbb{R}^k, l\in\{0,1,\cdots, M-1\}$ be a family of mappings such that $Z^M_l$ is $\mathcal{F}_{lT/M}/\mathcal{B}(\mathbb{R}^k)$-measurable for all $l\in\{0,1,\cdots, M-1\}$, then $\forall\; n\in\{0,1\cdots, M\}$ the following inequality holds : $$\begin{aligned}
\left\|\sup_{j\in\{0,1,\cdots, n\}}\left\|\sum_{l=0}^{j-1}Z^M_l\Delta\overline{N}^M_l\right\|\right\|_{L^p(\Omega, \mathbb{R})}\leq C_p\left(\sum_{j=0}^{n-1}||Z^M_j||^2_{L^p(\Omega, \mathbb{R}^k)}\dfrac{T}{M}\right)^{1/2},
\end{aligned}$$ where $C_p$ is a positive constant independent of $M$.
Let’s define $\overline{Z}^M : [0, T]\times\Omega\longrightarrow \mathbb{R}^k$ such that $\overline{Z}^M_s := Z^M_l$ for all $s\in\left[\dfrac{lT}{M}, \dfrac{(l+1)T}{M}\right)$, $l\in\{0, 1,\cdots, M-1\}$.
Using the definition of stochastic integral and Lemma \[ch4lemma18\], it follows that : $$\begin{aligned}
\left\|\sup_{j\in\{0,1,\cdots, n\}}\left\|\sum_{l=0}^{j-1}Z^M_l\Delta\overline{N}^M_l\right\|\right\|_{L^p(\Omega, \mathbb{R})}
&= &\left\|\sup_{j\in\{0,1,\cdots, n\}}\left\|\int_0^{jT/M}\overline{Z}^M_ud\overline{N}^M_u\right\|\right\|_{L^p(\Omega, \mathbb{R})}\\
&\leq &\left\|\sup_{s\in[0, nT/M]}\left\|\int_0^s\overline{Z}^M_ud\overline{N}_u\right\|\right\|_{L^p(\Omega, \mathbb{R}^k)}\\
&\leq & C_p\left(\int_0^{nT/M}||\overline{Z}^M_u||^2_{L^p(\Omega, \mathbb{R}^k)}ds\right)^{1/2}\\
&=&C_p\left(\sum_{j=0}^{n-1}||Z_j^M||^2_{L^p(\Omega,\mathbb{R}^k)}\dfrac{T}{M}\right)^{1/2}.
\end{aligned}$$ This complete the proof of lemma.
Now we are ready to prove Theorem \[ch4theorem1\].
**\[ Theorem \[ch4theorem1\]\]**
Let’s first represent the numerical approximation $Y^M_n$ in the following appropriate form : $$\begin{aligned}
Y^M_n&=&Y_{n-1}^M+\dfrac{\Delta tf_{\lambda}(Y^M_{n-1})}{1+\Delta t||f_{\lambda}(Y^M_{n-1})||}+g(Y_{n-1})\Delta W^M_{n-1}+h(Y^M_{n-1})\Delta\overline{N}^M_{n-1}\\
&=&X_0+\sum_{k=0}^{n-1}\dfrac{\Delta t f_{\lambda}(Y_k^M)}{1+\Delta t||f_{\lambda}(Y^M_k)||}+\sum_{k=0}^{n-1}g(Y^M_k)\Delta W^M_k+\sum_{k=0}^{n-1}h(Y^M_k)\Delta\overline{N}^M_k\\
&=& X_0+ \sum_{k=0}^{n-1}g(0)\Delta W^M_k+\sum_{k=0}^{n-1}h(0)\Delta\overline{N}^M_k+\sum_{k=0}^{n-1}\dfrac{\Delta tf_{\lambda}(Y^M_{n-1})}{1+\Delta t||f_{\lambda}(Y^M_{n-1})||}\\
&+&\sum_{k=0}^{n-1}(g(Y^M_k)-g(0))\Delta W^M_k+\sum_{k=0}^{n-1}(h(Y^M_k)-h(0))\Delta\overline{N}^M_k,
\end{aligned}$$ for all $M\in\mathbb{N} $ and all $n\in\{0,\cdots,M\}$.
Using the inequality $$\begin{aligned}
\left\|\dfrac{\Delta tf_{\lambda}(Y^M_k)}{1+\Delta t||f_{\lambda}(Y^M_k)||}\right\|_{L^P(\Omega, \mathbb{R}^d)} <1\end{aligned}$$ it follows that : $$\begin{aligned}
||Y^M_n||_{L^p(\Omega, \mathbb{R}^d)} &\leq &||X_0||_{L^p(\Omega,\mathbb{R}^d)}+\left\|\sum_{k=0}^{n-1}g(0)\Delta W^M_k\right\|_{L^p(\Omega, \mathbb{R}^d)}+\left\|\sum_{k=0}^{n-1}h(0)\Delta\overline{N}^M_k\right\|_{L^p(\Omega,\mathbb{R}^d)}+M\\
&+&\left\|\sum_{k=0}^{n-1}(g(Y^M_k)-g(0))\Delta W^M_k\right\|_{L^p(\Omega,\mathbb{R}^d)}+\left\|\sum_{k=0}^{n-1}(h(Y^M_k)-h(0))\Delta\overline{N}^M_k\right\|_{L^p(\Omega,\mathbb{R}^d)}.\end{aligned}$$ Using Lemma \[ch4lemma16\] and Lemma \[ch4lemma19\], it follows that : $$\begin{aligned}
||Y^M_n||_{L^p(\Omega,\mathbb{R}^d)} &\leq &||X_0||_{L^p(\Omega,\mathbb{R})}+C_p\left(\sum_{k=0}^{n-1}\sum_{i=1}^{m}||g_i(0)||^2\dfrac{T}{M}\right)^{1/2}+C_p\left(\sum_{k=0}^{n-1}||h(0)||^2\dfrac{T}{M}\right)^{1/2}\nonumber\\
&+&M+ C_p\left(\sum_{k=0}^{n-1}\sum_{i=1}^m||(g_i(Y_k^M)-g_i(0))\Delta W^M_k||^2_{L^p(\Omega, \mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}\nonumber\\ &+&C_p\left(\sum_{k=0}^{n-1}\lambda||(h(Y_k^M)-h(0))\Delta W^M_k||^2_{L^p(\Omega, \mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}\nonumber\\
&\leq&||X_0||_{L^p(\Omega, \mathbb{R}^d)}+C_p\left(\dfrac{nT}{M}\sum_{i=1}^m||g_i(0)||^2\right)^{1/2}+C_p\left(\dfrac{nT}{M}||h(0)||^2\right)^{1/2}\nonumber\\
&+&M+C_p\left(\sum_{k=0}^{n-1}\sum_{i=1}^m||g_i(Y^M_k)-g_i(0)||^2_{L^p(\Omega,\mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}\nonumber\\
&+&C_p\left(\sum_{k=0}^{n-1}||h(Y^M_k)-h(0)||^2_{L^p(\Omega,\mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}.
\label{ch4MB1}\end{aligned}$$ From $||g_i(0)||^2\leq ||g(0)||^2$ and the global Lipschitz condition satisfied by $g$ and $h$, we obtain
$||g_i(Y^M_k)-g_i(0)||_{L^p(\Omega, \mathbb{R}^d)}\leq C||Y^M_k||_{L^p(\Omega,\mathbb{R}^d)}$ and $||h(Y^M_k)-h(0)||_{L^p(\Omega, \mathbb{R}^d)}\leq C||Y^M_k||_{L^p(\Omega,\mathbb{R}^d)}$. So using , we obtain $$\begin{aligned}
||Y^M_n||_{L^p(\Omega, \mathbb{R}^d)} &\leq & ||X_0||_{L^p(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\\
&+&C_p\left(\dfrac{Tm}{M}\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega,\mathbb{R}^d)}\right)^{1/2}
+C_p\left(\dfrac{T}{M}\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega,\mathbb{R}^d)}\right)^{1/2}.\end{aligned}$$ Using the inequality $(a+b+c)^2\leq 3a^2+3b^2+3c^2$, it follows that : $$\begin{aligned}
||Y^M_n||^2_{L^p(\Omega,\mathbb{R}^d)}&\leq& 3\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)^2\nonumber\\
&+&\dfrac{3T(C_p\sqrt{m}+C_p)^2}{M}\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega, \mathbb{R}^d)},
\label{ch4MB2}\end{aligned}$$ for all $p\in[1,\infty)$.
Using the fact that $
\dfrac{3T(p\sqrt{m}+C_p)^2}{M}<3(p\sqrt{m}+C_p)^2
$ we obtain the following estimation $$\begin{aligned}
||Y^M_n||^2_{L^p(\Omega,\mathbb{R}^d)}&\leq& 3\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)^2\nonumber\\
&+&3T(C_p\sqrt{m}+C_p)^2\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega, \mathbb{R}^d)},
\label{ch4MB}\end{aligned}$$
Applying Gronwall lemma to leads to $$\begin{aligned}
||Y^M_n||^2_{L^p(\Omega, \mathbb{R}^d)}\leq 3e^{3(C_p\sqrt{m}+C_p)^2}\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)^2.
\label{ch4MB3}\end{aligned}$$ Taking the square root and the supremum in the both sides of leads to :
$
\sup\limits_{n\in\{0,\cdots, M\}}||Y^M_n||_{L^p(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
\leq \sqrt{3}e^{3(C_p\sqrt{m}+C_p)^2}\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)
\label{ch4MB4}\end{aligned}$$ Unfortunately, is not enough to conclude the proof of the lemma due to the term $M$ in the right hand side. Using the fact that $(\Omega_n^M)_n$ is a decreasing sequence and by exploiting Holder inequality, we obtain : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left\|\mathbf{1}_{(\Omega_n^M)^c}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}& \leq &\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega,\mathbb{R}^d)}\left\|Y^M_n\right\|_{L^{2p}(\Omega,\mathbb{R}^d)}\nonumber\\
&\leq &\left(\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left(M\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega,\mathbb{R}^d)}\right)\right)\nonumber\\
&\times &\left(\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left(M^{-1}||Y^M_n||_{L^{2p}(\Omega,\mathbb{R}^d)}\right)\right).
\label{ch4MB5}\end{aligned}$$ Using inequality we have
$
\left(\sup\limits_{M\in\mathbb{N}}\sup\limits_{n\in\{0,\cdots,M\}}\left(M^{-1}||Y^M_n||_{L^{2p}(\Omega,\mathbb{R}^d)}\right)\right)$ $$\begin{aligned}
\leq \sqrt{3}e^{3(C_p\sqrt{m}+C_p)^2}\left(\dfrac{||X_0||_{L^{2p}(\Omega, \mathbb{R}^d)}}{M}+\dfrac{C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||}{M}+1\right)\nonumber\\
\leq \sqrt{3}e^{3(C_p\sqrt{m}+C_p)^2}\left(||X_0||_{L^{2p}(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+1\right)<+\infty,
\label{ch4MB6}\end{aligned}$$ for all $p\geq 1$. From the relation $$\begin{aligned}
\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega, \mathbb{R}^d)}=
\mathbb{E}\left[\mathbf{1}_{(\Omega^M_M)^c}\right]^{1/2p}=
\mathbb{P}\left[(\Omega^M_M)^c\right]^{1/2p},\end{aligned}$$ it follows using Lemma \[ch4lemma13\] that : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left(M\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega,\mathbb{R}}\right)=\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left(M^{2p}\mathbb{P}\left[(\Omega^M_M)^c\right]\right)^{1/2p}<+\infty,
\label{ch4MB7}\end{aligned}$$ for all $p\geq 1$.
So plugging and in leads to : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)^c}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty.
\label{ch4MB8}\end{aligned}$$ Futhermore, we have $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)} &\leq &\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}\nonumber\\
&+&\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)^c}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}.
\label{ch4MB9}\end{aligned}$$ From , the second term of inequality is bounded, while using Lemma \[ch4lemma2\] and Lemma \[ch4lemma12\] we have : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}\leq \sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|D_n^M\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty.
\label{ch4MB10}\end{aligned}$$ Finally plugging and in leads to : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty.\end{aligned}$$
Strong convergence of the compensated tamed Euler scheme
--------------------------------------------------------
The main result of this chapter is given in the following theorem.
\[ch4theorem2\] Under Assumptions \[ch4assumption1\], for all $p\in[1,+\infty)$ there exist a positive constant $C_p$ such that : $$\begin{aligned}
\left(\mathbb{E}\left[\sup_{t\in[0,T]}\left\|X_t-\overline{Y}^M_t\right\|^p\right]\right)^{1/p}\leq C_p\Delta t^{1/2},
\label{ch4inetheo}
\end{aligned}$$ for all $M\in \mathbb{N}$.
Where $X : [0,T]\times \Omega\longrightarrow \mathbb{R}^d$ is the exact solution of equation and $\overline{Y}^M_t$ is the time continous approximation defined by .
In order to prove Theorem \[ch4theorem2\], we need the following two lemmas.
Following closely , we have the following lemma.
\[ch4lemma21\] Let $Y_n^M$ be defined by for all $M\in\mathbb{N}$ and all $n\in\{0,1,\cdots, M\}$, then we have $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,1,\cdots, M\}}\left(\mathbb{E}\left[||f_{\lambda}(Y_n^M)||^p\right]\vee \mathbb{E}\left[\left\|g(Y_n^M)\right\|^p\right]\vee \mathbb{E}\left[\left\|h(Y_n^M)\right\|^p\right]\right)<+\infty,\end{aligned}$$ for all $p\in[1,\infty)$.
From the polynomial growth condition of $f_{\lambda}$, for all $x\in\mathbb{R}^d$ we have $$\begin{aligned}
||f_{\lambda}(x)||\leq C(K+||x||^c)||x||+||f_{\lambda}(0)||=CK||x||+C||x||^{c+1}+||f_{\lambda}(0)||.\end{aligned}$$
- If $||x||\leq 1$, then $CK||x||\leq CK$, hence $$\begin{aligned}
||f_{\lambda}(x)||&\leq& CK+C||x||^{c+1}+||f_{\lambda}(0)||\nonumber\\
&\leq & KC+KC||x||^{c+1}+C+C||x||^{c+1}+||f_{\lambda}(0)||+||f_{\lambda}(0)||||x||^{c+1}\nonumber\\
&=&(KC+C+||f_{\lambda}(0)||)(1+||x||^{c+1}).
\label{ch4eq1}\end{aligned}$$
- If $||x||\geq 1$, then $C||x||\leq C||x||^{c+1}$, hence $$\begin{aligned}
||f_{\lambda}(x)||&\leq & KC||x||^{c+1}+C||x||^{c+1}+||f_{\lambda}(0)||\nonumber\\
&\leq & KC+KC||x||^{c+1}+C+C||x||^{c+1}+||f_{\lambda}(0)||+||f_{\lambda}(0)||||x||^{c+1}\nonumber\\
&=&(KC+C+||f_{\lambda}(0)||)(1+||x||^{c+1}).
\label{ch4eq2}\end{aligned}$$
So it follows from and that $$\begin{aligned}
||f_{\lambda}(x)|| \leq (KC+C+||f_{\lambda}(0)||)(1+||x||^{c+1}), \hspace{0.5cm} \text{for all} \hspace{0.3cm} x\in\mathbb{R}^d.
\label{ch4eq3}\end{aligned}$$ Using inequality and Theorem \[ch4theorem1\], it follows that : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left\|f_{\lambda}(Y_n^M)\right\|_{L^p(\Omega, \mathbb{R}^d)}&\leq& (KC+C+||f_{\lambda}(0)||)\\
&\times&\left(1+\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|Y^M_n\right\|^{c+1}_{L^{p(c+1)}(\Omega, \mathbb{R}^d)}\right)\\
&<&+\infty,\end{aligned}$$ for all $p\in[1,\infty)$.
In other hand, using the global Lipschitz condition satisfied by $g$ and $h$, it follows that : $$\begin{aligned}
||g(x)|| \leq C||x||+ ||g(0)|| \hspace{0.2cm} \text{and}\hspace{0.2cm} ||h(x)||\leq C||x||+||h(0)||.
\label{ch4stron2}\end{aligned}$$
Using once again Theorem \[ch4theorem1\], it follows from that : $$\begin{aligned}
\sup_{M\in\mathbb{N}, n\in\{0, \cdots, M\}}\left\|g(Y^M_n)\right\|_{L^p(\Omega, \mathbb{R}^d)}\leq||g(0)||+C\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|Y^M_n\right\|_{L^p(\Omega, \mathbb{R}^d)}<+\infty,\end{aligned}$$ for all $p\in[1,\infty)$.
Using the same argument as for $g$ the following holds $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|h(Y^M_n)\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty, \end{aligned}$$ for all $p\in[1, +\infty)$.
This complete the proof of Lemma \[ch4lemma21\].
For $s\in[0,T]$ let $\lfloor s\rfloor$ be the greatest grid point less than $s$. We have the following lemma.
\[ch4lemma22\] For any stepsize $\Delta t$, the following inequalities holds $$\begin{aligned}
\sup_{t\in[0,T]}\left\|\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\leq C_p\Delta t^{1/2},
\end{aligned}$$ $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{t\in[0,T]}\left\|\overline{Y}^M_t\right\|_{L^p(\Omega, \mathbb{R}^d)}<\infty,
\end{aligned}$$ $$\begin{aligned}
\sup_{t\in[0,T]}\left\|f_{\lambda}(\overline{Y}^M_t)-f_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})\right\|_{L^p(\Omega, \mathbb{R}^d)}\leq C_p\Delta t^{1/2}.
\end{aligned}$$
- Using Lemma \[ch4lemma18\], Lemma \[ch4lemma15\] and the time continous approximation , it follows that :
$
\sup\limits_{t\in[0,T]}\left\|\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq &\dfrac{T}{M}\left(\sup_{t\in[0,T]}\left\|\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})||}\right\|_{L^2(\Omega, \mathbb{R}^d)}\right)+\sup_{t\in[0, T]}\left\|\int^t_{\lfloor t\rfloor} g(\overline{Y}^M_{\lfloor t\rfloor})dW_s\right\|_{L^p(\Omega, \mathbb{R}^d)}\nonumber\\
&+& \sup_{t\in[0,T]}\left\|\int^t_{\lfloor t\rfloor}h(\overline{Y}^M_{\lfloor t\rfloor})d\overline{N}_s\right\|_{L^p(\Omega, \mathbb{R}^d)}\nonumber\\
&\leq &\dfrac{T}{\sqrt{M}}\left(\sup_{n\in\{0,\cdots, M\}}||f_{\lambda}(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)+\sup_{t\in[0,T]}\left(\dfrac{T}{M}\sum_{i=1}^m\int^t_{\lfloor t\rfloor}||g_i(\overline{Y}^M_s)||^2_{L^p(\Omega, \mathbb{R}^k)}ds\right)^{1/2}\nonumber\\
&+& \sup_{t\in[0,T]}\left(\dfrac{TC_p}{M}\int^t_{\lfloor t\rfloor}||h(\overline{Y}^M_s)||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right)^{1/2}\nonumber\\
&\leq &\dfrac{T}{\sqrt{M}}\left(\sup_{n\in\{0,\cdots, M\}}||f_{\lambda}(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)+\dfrac{\sqrt{Tm}}{\sqrt{M}}\left(\sup_{i\in\{1,\cdots, m\}}\sup_{n\in\{0,\cdots, M\}}||g_i(Y^M_n)||_{L^p(\Omega, \mathbb{R}^k)}\right)\nonumber\\
&+&\dfrac{C_p\sqrt{T}}{\sqrt{M}}\left(\sup_{n\in\{0, \cdots,M\}}||h(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right),
\label{ch4Thcontinous}
\end{aligned}$$ for all $M\in\mathbb{N}$.
Using inequality and Lemma \[ch4lemma21\], it follows that : $$\begin{aligned}
\left[\sup_{t\in[0,T]}\left\|\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\right]<C_p\Delta t^{1/2},
\label{Ch4bon1}\end{aligned}$$ for all $p\in[1,\infty)$ and all stepsize $\Delta t$.
- Using the inequalities , $||a||\leq ||a-b||+||b||$ for all $a,b\in\mathbb{R}^d$ and Theorem \[ch4theorem1\] it follows that $$\begin{aligned}
\sup_{t\in[0,T]}||\overline{Y}^M_t||_{L^p(\Omega, \mathbb{R}^d)}&\leq&\left[\sup_{t\in[0,T]}\left\|\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\right]+\sup_{t\in[0,T]}\left\|\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\\
&\leq&\dfrac{C_p}{M^{1/2}}+\sup_{t\in[0,T]}\left\|\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\\
&<&C_pT^{1/2}+\sup_{t\in[0,T]}\left\|\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\\
&<&\infty,\end{aligned}$$ for all $p\in[1,+\infty)$ and all $M\in\mathbb{N}$.
- Further, using the polynomial growth condition : $$\begin{aligned}
||f_{\lambda}(x)-f_{\lambda}(y)||\leq C(K+||x||^c+||y||^c)||x-y||,\end{aligned}$$ for all $x, y\in\mathbb{R}^d$, it follows using Holder inequality that : $$\begin{aligned}
\sup_{t\in[0,T]}||f_{\lambda}(\overline{Y}^M_t)-f_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})||_{L^p(\Omega, \mathbb{R}^d)} &\leq &C\left(K+2\sup_{t\in[0,T]}||\overline{Y}^M_t||^c_{L^{2pc}(\Omega, \mathbb{R}^d)}\right)\nonumber\\
&\times & \left(\sup_{t\in[0,T]}||\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}||_{L^{2p}(\Omega, \mathbb{R}^d)}\right)
\label{ch4Thfcontinou}\end{aligned}$$ Using and the first part of Lemma \[ch4lemma22\], the following inequality holds $$\begin{aligned}
\left[\sup_{t\in[0,T]}||f_{\lambda}(\overline{Y}^M_t)-f_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})||_{L^p(\Omega, \mathbb{R}^d)}\right]<C_p\Delta t^{1/2},
\label{ch4Thffinal}\end{aligned}$$ for all $p\in[1,\infty)$ and for all stepsize $\Delta t$.
Now we are ready to give the proof of Theorem \[ch4theorem2\].
**\[ Theorem \[ch4theorem2\]\]**
Let’s recall that for $s\in[0,T]$, $\lfloor s\rfloor$ denote the greatest grid point less than $s$. The time continuous solution can be writen into its integral form as bellow : $$\begin{aligned}
\overline{Y}^M_s=X_0+\int_0^s\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}du+ \int_0^s g(\overline{Y}^M_{\lfloor u\rfloor})dW_u+\int_0^s h(\overline{Y}^M_{\lfloor u\rfloor})d\overline{N}_u,
\label{ch4continoussol2}
\end{aligned}$$ for all $s\in[0, T]$ almost surely and all $M\in\mathbb{N}$.
Let’s estimate first the quantity $||X_s-\overline{Y}^M_s||^2$ $$\begin{aligned}
X_s-\overline{Y}_s&=&\int_0^s\left(f_{\lambda}(X_u)-\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right)du+\int_0^s\left(g(X_u)-g(\overline{Y}^M_{\lfloor u\rfloor})\right)dW_u\\
&+&\int_0^s\left(h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right)d\overline{N}_u.
\end{aligned}$$ Using the relation $d\overline{N}_u=dN_u-\lambda du$, it follows that $$\begin{aligned}
X_s-\overline{Y}_s&=&\int_0^s\left[\left(f_{\lambda}(X_u)-\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right)-\lambda \left(h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right)\right]du\\
&+&\int_0^s\left(g(X_u)-g(\overline{Y}^M_{\lfloor u\rfloor})\right)dW_u+\int_0^s\left(h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right)dN_u.
\end{aligned}$$ The function $ k :\mathbb{R}^m\longrightarrow \mathbb{R}$, $x \longmapsto ||x||^2$ is twice differentiable. Applying Itô’s formula for jumps process to the process $X_s-\overline{Y}^M_s$ leads to : $$\begin{aligned}
\left\|X_s-\overline{Y}^M_s\right\|^2&=&2\int_0^s\left<X_u-\overline{Y}^M_u, f_{\lambda}(X_u)-\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right>du\\
&-&2\lambda\int_0^s\left<X_u-\overline{Y}^M_u, h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right>du +\sum_{i=1}^m\int_0^s||g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})||^2du\\
&+&2\sum_{i=1}^m\int_0^s\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>dW^i_u\\
&+& \int_0^s\left[||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2-||X_u-\overline{Y}^M_u||^2\right]dN_u.
\end{aligned}$$ Using again the relation $d\overline{N}_u=dN_u-\lambda du$ leads to : $$\begin{aligned}
\left\|X_s-\overline{Y}^M_s\right\|^2&=&2\int_0^s\left<X_u-\overline{Y}^M_u, f_{\lambda}(X_u)-\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right>du\nonumber\\
&-&2\lambda\int_0^s\left<X_u-\overline{Y}^M_u, h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right>du+\sum_{i=1}^m\int_0^s||g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&+&2\sum_{i=1}^m\int_0^s\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>dW^i_u\nonumber\\
&+& \int_0^s\left[||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2-||X_u-\overline{Y}^M_u||^2\right]d\overline{N}_u\nonumber\\
&+&\lambda\int_0^s\left[||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u \rfloor})||^2-||X_u-\overline{Y}^M_u||^2\right]du\nonumber\\
&=&A_1+A_2+A_3+A_4+A_5+A_6.
\label{ch4Th1}
\end{aligned}$$ In the next step, we give some useful estimations of $A_1, A_2, A_3$ and $A_6$. $$\begin{aligned}
A_1& : =&2\int_0^s\left<X_u-\overline{Y}^M_u,f_{\lambda}-\dfrac{f_{\lambda}(\overline{Y}_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right>du\\
&=&2\int_0^s\left\langle X_s-\overline{Y}^M_u,f_{\lambda}(X_u)-f_{\lambda}(\overline{Y}^M_u)\right\rangle du\\
&+&2\int_0^s\left<X_s-\overline{Y}^M_u,f_{\lambda}(\overline{Y}^M_u)-\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right>du.\\
&=& A_{11}+A_{12}
\end{aligned}$$ Using the one-sided Lipschitz condition satisfied by $f_{\lambda}$ leads to : $$\begin{aligned}
A_{11} &: =&2\int_0^s\left\langle X_s-\overline{Y}^M_u,f_{\lambda}(X_u)-f_{\lambda}(\overline{Y}^M_u)\right\rangle du\nonumber\\
&\leq& 2C\int_0^u||X_u-\overline{Y}^M_u||^2du.
\label{ch4ThA11}
\end{aligned}$$ Moreover, using the inequality $\langle a, b\rangle\leq |a||b|\leq \dfrac{a^2}{2}+\dfrac{b^2}{2}$ leads to : $$\begin{aligned}
A_{12}&=& 2\int_0^s\left<X_u-\overline{Y}^M_u,f_{\lambda}(\overline{Y}^M_u)-\dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right>du\nonumber\\
&=&2\int_0^s\left<X_u-\overline{Y}^M_u, f_{\lambda}(\overline{Y}^M_u)-f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})\right>ds\nonumber\\
&+&2\Delta t\int_0^s\left<X_u-\overline{Y}^M_u, \dfrac{f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}{1+\Delta t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||}\right>du\nonumber\\
&\leq &\int_0^s||X_u-\overline{Y}^M_u||^2du+\int_0^s||f_{\lambda}(\overline{Y}^M_u)-f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&+&\int_0^s||X_u-\overline{Y}^M_u||^2du+\dfrac{T^2}{M^2}\int_0^s||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^4du\nonumber\\
&\leq &2\int_0^s||X_u-\overline{Y}^M_u||^2du+\int_0^s||f_{\lambda}(\overline{Y}^M_u)-f_{\lambda}(\overline{Y}_{\lfloor u\rfloor})||^2du\nonumber\\
&+&\dfrac{T^2}{M^2}\int_0^s||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^4du.
\label{ch4ThA12}
\end{aligned}$$ Combining and give the following estimation of $A_1$ : $$\begin{aligned}
A_1 &\leq & (2C+2)\int_0^s||X_u-\overline{Y}^M_u||^2du+\int_0^s||f_{\lambda}(\overline{Y}^M_u)-f_{\lambda}(\overline{Y}_{\lfloor u\rfloor})||^2du\nonumber\\
&+&\dfrac{T^2}{M^2}\int_0^s||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^4du.
\label{ch4ThA1}
\end{aligned}$$ Using again the inequality $2\langle a, b\rangle\leq 2|a||b|\leq a^2+b^2$ and the global Lipschitz condition satisfied by $h$ leads to : $$\begin{aligned}
A_2 & : =&-2\lambda\int_0^s\left<X_u-\overline{Y}^M_u, h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right>du\nonumber\\
&=&-2\lambda\int_0^s\left<X_u-\overline{Y}^M_u, h(X_u)-h(\overline{Y}^M_u)\right>du-2\lambda\int_0^s\left<X_u-\overline{Y}^M_u, h(\overline{Y}^M_u)-h(\overline{Y}^M_{\lfloor u\rfloor})\right>du\nonumber\\
&\leq &(2\lambda+\lambda C^2)\int_0^s||X_u-\overline{Y}^M_u||^2du+\lambda C^2\int_0^s||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2du.
\label{ch4ThA2}
\end{aligned}$$ Using the inequalities $||g_i(x)-g_i(y)||\leq ||g(x)-g(y)||$ and $(a+b)^2\leq 2a^2+2b^2$ and the global Lipschitz condition we have $$\begin{aligned}
A_3 &: =&\sum_{i=1}^m\int_0^s||g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&\leq &m\int_0^s||g(X_u)-g(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&\leq &m\int_0^s||g(X_u)-g(\overline{Y}^M_u)+g(\overline{Y}^M_u)-g(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&\leq &2m\int_0^s||g(X_u)-g(\overline{Y}^M_u)||^2du+2m\int_0^s||g(\overline{Y}^M_u)-g(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&\leq& 2mC^2\int_0^s||X_u-\overline{Y}^M_u||^2du+2mC^2\int_0^s||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2du.
\label{ch4ThA3}
\end{aligned}$$ Using the same idea as above we obtain the following estimation of $A_6$ : $$\begin{aligned}
A_6 & : =&\lambda\int_0^s\left[X_u-\overline{Y}^M_u+h(\overline{Y}^M_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2-||X_u-\overline{Y}^M_u||^2\right]du\nonumber\\
&\leq &3\lambda\int_0^s||X_u-\overline{Y}^M_u||^2du+2\lambda\int_0^s||h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&\leq &3\lambda\int_0^s||X_u-\overline{Y}^M_u||^2du+4\lambda\int_0^s||h(X_u)-h(\overline{Y}^M_u)||^2du\nonumber\\
&+ &4\lambda\int_0^s||h(\overline{Y}^M_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2du\nonumber\\
&\leq &(3\lambda+4\lambda C^2)\int_0^s||X_u-\overline{Y}^M_u||^2du+4\lambda C^2\int_0^s||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2du.
\label{ch4ThA6}
\end{aligned}$$ Inserting , , and in we obtain : $$\begin{aligned}
\left\|X_s-\overline{Y}^M_s\right\|^2&\leq &(2C+2+2mC^2+5\lambda+5\lambda C^2)\int_0^s||X_u-\overline{Y}^M_u||^2du\nonumber\\
&+&(2mC^2+5\lambda C^2)\int_0^s||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2du\nonumber\\
&+&\int_0^s||f_{\lambda}(\overline{Y}^M_u)-f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^2du+\dfrac{T^2}{M^2}\int_0^s||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^4du\nonumber\\
&+&2\sum_{i=1}^m\int_0^s\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>dW^{i}_u\nonumber\\
&+&\int_0^s\left[||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2-||X_u-\overline{Y}^M_u||^2\right]d\overline{N}_u.
\end{aligned}$$ Taking the supremum in both sides of the previous inequality leads to $$\begin{aligned}
\sup_{s\in[0,t]}\left\|X_s-\overline{Y}^M_s\right\|^2&\leq &(2C+2+2mC^2+5\lambda+5\lambda C^2)\int_0^t||X_u-\overline{Y}^M_u||^2du\nonumber\\
&+&(2mC^2+5\lambda C^2)\int_0^t||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2du\nonumber\\
&+&\int_0^t||f_{\lambda}(\overline{Y}^M_u)-f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^2du+\dfrac{T^2}{M^2}\int_0^t||f_{\lambda}(\overline{Y}^M_{\lfloor u\rfloor})||^4du\nonumber\\
&+&2\sup_{s\in[0,t]}\left|\sum_{i=1}^m\int_0^s\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>dW^{i}_u\right|\nonumber\\
&+&\sup_{s\in[0,t]}\left|\int_0^s\left[||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2\right]d\overline{N}_u\right|\nonumber\\
&+&\sup_{s\in[0,t]}\left|\int_0^s||X_u-\overline{Y}^M_u||^2d\overline{N}_u\right|
\label{ch4Th2}\end{aligned}$$ Using Lemma \[ch4lemma15\] we have the following estimation for all $p\geq 2$ $$\begin{aligned}
B_1& :=&\left\|2\sup_{s\in[0,t]}\left|\sum_{i=1}^m\int_0^s\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>dW^i_u\right|\right\|_{L^{p/2}(\Omega, \mathbb{R})}\\
&\leq &C_p\left(\int_0^t\sum_{i=1}^m\left\|\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>\right\|^2_{L^{p/2}(\Omega, \mathbb{R})}ds\right)^{1/2}.
\end{aligned}$$ Moreover, using Holder inequality, the inequalities $ab\leq \dfrac{a^2}{2}+\dfrac{b^2}{2}$ and $(a+b)^2\leq 2a^2+b^2$ we have the following estimations for all $p\geq 2$ $$\begin{aligned}
B_1 &\leq& C_p\left(\int_0^t\sum_{i=1}^m\left\|\left<X_u-\overline{Y}^M_u, g_i(X_u)-g_i(\overline{Y}^M_{\lfloor u\rfloor})\right>\right\|^2_{L^{p/2}(\Omega, \mathbb{R})}ds\right)^{1/2}\nonumber\\
&\leq &C_p\left(\int_0^t\sum_{i=1}^m||X_u-\overline{Y}_u^M||^2_{L^p(\Omega, \mathbb{R})}||g_i(X_u)-g_u(\overline{Y}^M_{\lfloor u\rfloor})||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right)^{1/2}\nonumber\\
&\leq &\dfrac{C_p}{\sqrt{2}}\left(\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||_{L^p(\Omega, \mathbb{R}^d)}\right)\left(2pC^2m\int_0^t||X_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right)^{1/2}\nonumber\\
&\leq &\dfrac{1}{4}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}+p^2C_p^2m\int_0^t||X_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}ds\nonumber\\
&\leq &\dfrac{1}{4}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}+2p^2C_p^2m\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega,\mathbb{R}^d)}ds\nonumber\\
&+&2p^2C_p^2m\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}ds,
\label{ch4ThB1}
\end{aligned}$$ Using Lemma \[ch4lemma18\] and the inequality $(a+b)^4\leq 16a^4+16b^4$, it follows that $$\begin{aligned}
B_2 &: =&\left\|\sup_{s\in[0,t]}\left|\int_0^s||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^2d\overline{N}_u\right|\right\|_{L^{p/2}(\Omega, \mathbb{R}^d)}\nonumber\\
&\leq &C_p\left(\int_0^t||X_u-\overline{Y}^M_u+h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}\nonumber\\
&\leq &C_p\left(\int_0^t16||X_u-\overline{Y}^M_u||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}+16||h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2},
\end{aligned}$$ for all $p\geq 2$.
Using the inequality $\sqrt{a+b}\leq\sqrt{a}+\sqrt{b}$, it follows that : $$\begin{aligned}
B_2&\leq &2C_p\left(\int_0^t||X_u-\overline{Y}^M_u||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2} +2C_p\left(\int_0^t||h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}\nonumber\\
& =& B_{21}+B_{22}.
\label{ch4ThB}
\end{aligned}$$ Using Holder inequality, it follows that $$\begin{aligned}
B_{21} &: = &2 C_p\left(\int_0^t||X_u-\overline{Y}^M_u||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}\nonumber\\
&\leq &2C_p\left(\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}du\right)^{1/2}\nonumber\\
&\leq &\dfrac{1}{4}\sup_{u\in[0,t]}||X_u-\overline{Y}^M_u||_{L^p(\Omega,\mathbb{R}^d)}8C_p\left(\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega,\mathbb{R}^d)}du\right)^{1/2}.
\end{aligned}$$ Using the inequality $2ab\leq a^2+b^2$ leads to : $$\begin{aligned}
B_{21}&\leq &\dfrac{1}{16}\sup_{u\in[0,t]}||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}
+16C^2_p\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}du.
\label{ch4ThB21}
\end{aligned}$$ Using the inequalities $(a+b)^4\leq 4a^4+4b^4$ and $\sqrt{a+b}\leq\sqrt{a}+\sqrt{b}$, we obtain $$\begin{aligned}
B_{22}&: = &2C_p\left(\int_0^t||h(X_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}\nonumber\\
&\leq &2C_p\left(\int_0^t\left[4||h(X_u)-h(\overline{Y}^M_u)||^4_{L^{p/2}(\Omega,\mathbb{R}^d)}+4||h(\overline{Y}^M_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^4_{L^{p/2}(\Omega,\mathbb{R}^d)}\right]du\right)^{1/2}\\
&\leq &4C_p\left(\int_0^t||h(X_u)-h(\overline{Y}^M_u)||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}
+4C_p\left(\int_0^t||h(\overline{Y}^M_u)-h(\overline{Y}^M_{\lfloor u\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}.
\end{aligned}$$ Using the global Lipschitz condition, leads to : $$\begin{aligned}
B_{22}&\leq &4C_p\left(\int_0^tC||X_u-\overline{Y}^M_u||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}+4C_p\left(\int_0^tC||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}.
\end{aligned}$$ Using the same estimations as for $B_{21}$, it follows that : $$\begin{aligned}
B_{22} &\leq &\dfrac{1}{16}\sup_{u\in[0,t]}||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)} +64C_p\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}du\nonumber\\
&+&\dfrac{1}{4}\sup_{u\in[0,t]}||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}+64C_p\int_0^t||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}du.\nonumber
\end{aligned}$$ Taking the supremum under the integrand in the last term of the above inequality and using the fact that we don’t care about the value of the constant leads to : $$\begin{aligned}
B_{22}&\leq &\dfrac{1}{16}\sup_{u\in[0,t]}||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)} +64C_p\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}du\nonumber\\
&+&C_p\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}.
\label{ch4ThB22}
\end{aligned}$$ Inserting and into gives : $$\begin{aligned}
B_2&\leq &\dfrac{1}{8}\sup_{u\in[0,t]}||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}+C_p\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}du\nonumber\\
&+&C_p\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}.
\label{ch4ThB2}
\end{aligned}$$ Using again Lemma \[ch4lemma18\] leads to : $$\begin{aligned}
B_3 &: =& \left\|\sup_{u\in[0,t]}\left(\int_0^s||X_u-\overline{Y} ^M_u||^2d\overline{N}_u\right)^{1/2}\right\|_{L^{p/2}(\Omega, \mathbb{R}^d)}\\
&\leq &C_p\left(\int_0^t||X_u-\overline{Y}^M_u||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}du\right)^{1/2}.
\end{aligned}$$ Using the same argument as for $B_{21}$, we obtain : $$\begin{aligned}
B_3 &\leq &\dfrac{1}{8}\sup_{u\in[0,t]}||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}+C_p\int_0^t||X_u-\overline{Y}^M_u||^2_{L^p(\Omega, \mathbb{R}^d)}du.
\label{ch4ThB3}
\end{aligned}$$ Taking the $L^{p}$ norm in both side of , inserting inequalities , , and using Holder inequality in its integral form (see Proposition \[ch1Minkowski\]) leads to : $$\begin{aligned}
\left\|\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^p(\Omega, \mathbb{R}^d)}&=&\left\|\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2\right\|_{L^{p/2}(\Omega, \mathbb{R}^d)}\end{aligned}$$ $$\begin{aligned}
&\leq & C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds+C_p\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}ds\\
&+&\int_0^t||f_{\lambda}(X_s)-f_{\lambda}(\overline{Y}^M_{\lfloor s\rfloor})||^2_{L^{p}(\Omega, \mathbb{R}^d)}ds+C_p\sup_{u\in[0,t]}||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}\\
&+&\dfrac{T^2}{M^2}\int_0^t||f_{\lambda}(\overline{Y}^M_{\lfloor s\rfloor})||^4_{L^{2p}(\Omega, \mathbb{R}^d)}ds +2C_p\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds\\
&+&\dfrac{1}{2}\left\|\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^{p}(\Omega, \mathbb{R}^d)}.
\end{aligned}$$ for all $t\in[0,T]$ and all $p\in[2,+\infty)$.
The previous inequality can be rewrite in the following appropriate form :
$
\dfrac{1}{2}\left\|\sup\limits_{s\in[0,t]}\left\|X_s-\overline{Y}^M_s\right\|\right\|^2_{L^{p}(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq & C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds+C_p\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}ds\\
&+&\int_0^t||f_{\lambda}(X_s)-f_{\lambda}(\overline{Y}^M_{\lfloor s\rfloor})||^2_{L^{p}(\Omega, \mathbb{R}^d)}ds+C_p\sup_{u\in[0,t]}||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}\\
&+&\dfrac{T^2}{M^2}\int_0^t||f_{\lambda}(\overline{Y}^M_{\lfloor s\rfloor})||^4_{L^{2p}(\Omega, \mathbb{R}^d)}ds +2C^2m\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds.
\end{aligned}$$
Applying Gronwall lemma to the previous inequality leads to :
$
\dfrac{1}{2}\left\|\sup\limits_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^p(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq &C_pe^{C_p}\left(\int_0^T||f_{\lambda}(\overline{Y}^M_s)-f_{\lambda}(\overline{Y}^M_{\lfloor s\rfloor})||^2_{L^p(\Omega, \mathbb{R}^d)}ds+C_p\sup_{t\in[0,t]}||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}\right.\\
& &\left. +\dfrac{T^2}{M^2}\int_0^T||f_{\lambda}(\overline{Y}^M_{\lfloor s\rfloor})||^4_{L^{2p}(\Omega, \mathbb{R}^d)}ds +C_p\int_0^T||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right).
\end{aligned}$$ From the inequality $\sqrt{a+b+c}\leq \sqrt{a}+\sqrt{b}+\sqrt{c}$, it follows that
$
\left\|\sup\limits_{t\in[0, T]}||X_t-\overline{Y}^M_t||\right\|_{L^p(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq &C_pe^{C_p}\left(\sup_{t\in[0,T]}||f_{\lambda}(\overline{Y}^M_t)-f_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})||_{L^p(\Omega, \mathbb{R}^d)}+C_p\sup_{t\in[0,t]}||\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}||_{L^p(\Omega, \mathbb{R}^d)} \right.\nonumber\\
&+& \left. \dfrac{T}{M}\left[\sup_{n\in\{0,\cdots, M\}}||f_{\lambda}(Y^M_n)||^2_{L^p(\Omega, \mathbb{R}^d)}\right]+C_p\sup_{t\in[0,T]}||\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}||_{L^p(\Omega, \mathbb{R}^d)}\right),
\label{ch4Gronwall}
\end{aligned}$$ for all $p\in[2,\infty)$.
Using Lemma \[ch4lemma21\] and Lemma \[ch4lemma22\] it follows from \[ch4Gronwall\] that $$\begin{aligned}
\left(\mathbb{E}\left[\sup_{t\in[0,T]}\left\|X_t-\overline{Y}^M_t\right\|^p\right]\right)^{1/p}\leq C_p(\Delta t)^{1/2},
\label{ch4final}
\end{aligned}$$ for all $p\in[2,\infty)$. Using Holder inequality, one can prove that holds for all $p\in[1,2]$. The proof of the theorem is complete.
Numerical Experiments
---------------------
In order to illustrate our theoretical result, we consider the following stochastic differential equation
$\left\{\begin{array}{ll}
dX_t=-X_t^4dt+X_tdW_t+X_tdN_t\\
X_0=1,
\end{array}
\right.$
$\lambda=1$. It is straighforward to verify that Assumptions \[ch4assumption1\] are satisfied. We use Monte carlo method to evaluate the error. The exact solution is consider as the numerical one with small stepsize $dt=2^{-14}$. We have the following curve for $5000$ paths.
In this chapter, we proposed a compensated tamed Euler scheme to solve numerically SDEs with jumps under non-global Lipschitz condition. We proved the strong convergence of order $0.5$ of the compensated Euler scheme. This scheme is explicit and then requires less computational efforts than the implicit scheme. In some situations, the drift part can be equipped with the Lipschitz continuous part and the non-Lipschitz continuous part. In the following chapter, we combine the tamed Euler scheme and the Euler scheme and obtain another scheme called semi-tamed Euler scheme in order to solve numerically the kind of equation mentioned above.
Strong convergence and stability of the semi- tamed Euler and the tamed Euler scheme for stochastic differential equations with jumps, under non-global Lipschitz continous coefficients
========================================================================================================================================================================================
Explicit numerical method called compensated tamed Euler scheme is developped in the previous chapter. More precisely, it is proved that such numerical approximation have strong convergence of order $0.5$ for stochastic differential equations with jumps under non-global Lipschitz condition. In this chapter, following the idea of [@Xia2], we propose a semi-tamed Euler scheme to solve stochastic differential equations with jumps, where the drift coefficient is equipped with the Lipschitz continuous part and the non-Lipschitz continuous part. We prove that for SDEs with jumps, the semi-tamed Euler scheme converges strongly with order $0.5$. We use this result to deduce a strong convegrence of order $0.5$ of the tamed Euler scheme for SDEs with jumps, where the drift coefficient satisfies the non-global Lipschitz condition. We also investigate the stability analysis of both semi-tamed Euler scheme and tamed Euler scheme. The contents of this chapter can also be found in [@atjdm1] and [@atjdm2].
Semi-tamed Euler scheme {#ch5intro}
-----------------------
In this chapter, we consider again a jump-diffusion Itô’s stochastic differential equations (SDEs) of the form : $$\begin{aligned}
dX(t)= f((X(t^{-}))dt +g((X(t^{-}))dW(t)+h(X(t^{-}))dN(t), \hspace{0.5cm}
X(0)=X_0,
\label{ch5exactsol1}\end{aligned}$$ where $W_t$ is a $m$-dimensional Brownian motion, $f :\mathbb{R}^d\longrightarrow\mathbb{R}^d$ such that $f(x)=u(x)+v(x)$ satisfies the global one-sided Lipschitz condition. $u, v : \mathbb{R}^d\longrightarrow\mathbb{R}^d$, $u$ is the global Lipschitz continous part while $v$ is the non-global Lipschitz continous part. The functions $ g : \mathbb{R}^d \longrightarrow\mathbb{R}^{d\times m}$ and $h :\mathbb{R}^d \longrightarrow\mathbb{R}^d$ satisfy the global Lipschitz condition, $N_t$ is a one dimensional poisson process with parameter $\lambda$. Using the relation $f=u+v$, equation can be rewritten into its equivalent form : $$\begin{aligned}
dX(t)=u(X(t^-))dt+v(X(t^-))dt+g(X(t^-))dW(t)+h(X(t^-))dN(t), \hspace{0.2cm} X(0)=X_0.\end{aligned}$$
We can rewrite the jump-diffusion SDEs in the following equivalent form $$\begin{aligned}
dX(t)= f_\lambda(X(t^{-}))dt +g(X(t^{-}))dW(t)+h(X(t^{-}))d\overline{N}(t),\end{aligned}$$ where $$f_\lambda(x)=f(x)+\lambda h(x)= u(x)+\lambda h(x)+v(x).
\label{ch5flambda}$$ If $T$ is the final time we consider the tamed Euler scheme $$X_{n+1}^{M}=X_{n}^{M}+\dfrac{\Delta t f(X_{n}^{M})}{1+ \Delta t\Vert f(X_{n}^{N}) \Vert }+g(X_{n}^{M}) \Delta W_n +h(X_{n}^{M})\Delta N_n
\label{ch5tam}$$ and the semi-tamed Euler scheme $$\begin{aligned}
Y_{n+1}^{M}=Y_{n}^{M}+u(Y^M_n)\Delta t+\dfrac{\Delta t v(Y_{n}^{M})}{1+ \Delta t \Vert v(Y_{n}^{M}) \Vert }+g(Y_{n}^{M}) \Delta W_n +h(Y_{n}^{M})\Delta N_n,
\label{ch5semi}\end{aligned}$$ where $\Delta t=\dfrac{T}{M}$ is the time step-size, $M\in\mathbb{N}$ is the number of steps.
Inspired by [@Xia2] and [@Martin1] we prove the strong convergence of the numerical approximation and deduce the strong convergence of to the exact solution of .
Moments bounded of the numerical solution
------------------------------------------
Throughout this chapter, we use Notations \[ch4nota1\].
Note that the numerical approximation can be writen into its following equivalent form $$\begin{aligned}
Y_{n+1}^{M}=Y_{n}^{M}+u(Y^M_n)\Delta t+\lambda h(Y^M_n)\Delta t+\dfrac{\Delta t v(Y_{n}^{M})}{1+ \Delta t \Vert v(Y_{n}^{M}) \Vert }+g(Y_{n}^{M}) \Delta W_n +h(Y_{n}^{M})\Delta \overline{N}_n.
\label{ch5semitam}\end{aligned}$$ We define the continous time interpolation of the discrete numerical approximation of $\eqref{ch5semitam}$ by the familly of processes $\left(\overline{Y}^M\right)_M $, $
\overline{Y}^M : [0,T]\times\Omega \longrightarrow \mathbb{R}^d
$ such that : $$\begin{aligned}
\overline{Y}^M_t &=&Y^M_n+u(Y^M_n)(t-n\Delta t)+\lambda h(Y^M_n)(t-n\Delta t)+\dfrac{(t-n\Delta t)v(Y^M_n)}{1+\Delta t||v(Y^M_n)||}\nonumber\\
&+&g(Y^M_n)(W_t-W_{n\Delta t})+ h(Y^M_n)(\overline{N}_t-\overline{N}_{n\Delta t}),
\label{ch5continoussolu}\end{aligned}$$ for all $M\in\mathbb{N}$, all $n\in\{0,\cdots, M-1\}$, and all $t\in[n\Delta t, (n+1)\Delta t)$.
\[ch5assumption1\] We assume that :
$(A.1)$ $f,g,h\in C^1$.
$(A.2)$ For all $p>0$, there exist a finite $M_p>0$ such that $\mathbb{E}||X_0||^p\leq M_p$.
$(A.3)$ $g$, $h$ and $u$ satisfy global Lipschitz condition: $$\begin{aligned}
||g(x)-g(y)||\vee ||h(x)-h(y)||\vee ||u(x)-u(y)|| \leq C||x-y||, \hspace{0.5cm} \forall\; x,y\in \mathbb{R}^d.\end{aligned}$$ $(A.4)$ $v$ satisfies the one-sided Lipschitz condition : $$\begin{aligned}
\langle x-y, f(x)-f(y)\rangle\leq C||x-y||^2,\hspace{0.5cm} \forall\;x,y\in \mathbb{R}^d,\end{aligned}$$ $(A.5)$ $v$ satisfies the superlinear growth condition : $$\begin{aligned}
||v(x)-v(y)||\leq C(K+ ||x||^c+||y||^c)||x-y||, \hspace{0.5cm} \forall\;x,y\in \mathbb{R}^d,\end{aligned}$$ where $K$, $C$ and $c$ are constants strictetly positives.
Under conditions $(A.1)$, $(A.2)$ and $(A.3)$ of Assumptions \[ch5assumption1\] it is proved in [@Desmond2 Lemma 1] that has a unique solution with all moments bounded.
Let’s define $u_{\lambda}(x)=u(x)+\lambda h(x)$. From Assumptions \[ch5assumption1\], it is straightforward to prove that $u_{\lambda}$ satisfies the global Lipschitz condition with constant $C_{\lambda}=(1+\lambda)C$ and $v$ satisfies the one-sided Lipschitz condition. We denote by $C_p$ a generic constant. Throughout this work, this constant may change the value from one line to another one. We will sometimes use $Y_n^M$ instead of $Y_n^M(\omega)$ to simplify notations.
The main result of this section is formulated in the following theorem. which is based on [@Martin1 Lemma 3.9 pp 16]. Here, we include the jump part.
\[ch5theorem1\]
Let $Y_n^M : \Omega\longrightarrow \mathbb{R}^d$ be defined by for all $M\in\mathbb{N}$ and all $ n\in\{0,\cdots, M\}$. The following inequality holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\mathbb{E}\left[||Y_n^M||^p\right]<+\infty,
\end{aligned}$$ for all $p\in[1,\infty)$.
In order to prove Theorem \[ch5theorem1\] we introduce the following notations facilitating computations.
\[ch5notation1\]
$$\begin{aligned}
\alpha^M_k := \mathrm{1}_{\{||Y^M_k||\geq 1\}}\left\langle\dfrac{Y^M_k+u_{\lambda}(Y^M_n)\Delta t}{||Y^M_k||}, \dfrac{g(Y^M_k)}{||Y^M_k||}\Delta W^M_k\right\rangle,\\\\
\beta^M_k := \mathrm{1}_{\{||Y^M_k||\geq 1\}}\left\langle\dfrac{Y^M_k+u_{\lambda}(Y^M_n)\Delta t}{||Y^M_k||}, \dfrac{h(Y^M_k)}{||Y^M_k||}\Delta\overline{N}^M_k\right\rangle,
\end{aligned}$$
$$\begin{aligned}
\beta=\left(1+K+3C_{\lambda}+KTC_{\lambda}+TC_{\lambda}+||u_{\lambda}(0)||+||g(0)||+||h(0)||\right)^4,\\\\
D^M_n := (\beta+||\varepsilon||)\exp\left(4\beta+\sup_{u\in\{0,\cdots,n\}}\sum_{k=u}^{n-1}\left[2\beta||\Delta W^M_k||^2+2\beta||\Delta\overline{N}^M_k||+\alpha^M_k+\beta^M_k\right]\right),\\
\Omega^M_n :=\{\omega\in \Omega : \sup_{k\in\{0,1,\cdots, n-1\}}D^M_k(\omega)\leq M^{1/2c}, \sup_{k\in\{0,1,\cdots,n-1\}}||\Delta W^M_k(\omega)||\leq 1,\\ \sup_{k\in\{0,1,\cdots,n-1\}}||\Delta \overline{N}^M_k(\omega)||\leq 1\},
\end{aligned}$$
for all $M\in\mathbb{N}$ and $k\in\{0,\cdots M\}$.
Following closely [@Xia2 Lemma 2.1] we have the following main lemma.
\[ch5lemma2\] The following inequality holds for all $M\in \mathbb{N}$ and all $n\in\{0,1,\cdots, M\}$ $$\begin{aligned}
\mathbf{1}_{\Omega^M_n}||Y^M_n||\leq D^M_n.
\label{ch5Denobound}
\end{aligned}$$
Using the inequality $\dfrac{\Delta t}{1+\Delta t||u_{\lambda}(x)||}\leq T$, the global Lipschitz condition of $g$ and $h$ and the polynomial growth condition of $v$ we have the following estimation on $\Omega^M_{n+1}\cap\{\omega\in \Omega : ||Y^M_n(\omega)||\leq 1\}$, for all $n\in\{0,1,\cdots, M-1\}$ $$\begin{aligned}
||Y^M_{n+1}||&\leq ||Y^M_n||+||u_{\lambda}(Y^M_n)||\Delta t+\dfrac{\Delta t||v(Y^M_n)||}{1+\Delta t||v(Y^M_n)||}+||g(Y^M_n)||||\Delta W^M_n||+||h(Y^M_n)||||\Delta\overline{N} ^M_n|| \nonumber \\
&\leq ||Y^M_n||+T||u_{\lambda}(Y^M_n)-u_{\lambda}(0)||+T||u_{\lambda}(0)||+T||v(Y^M_n)-v(0)||+T||v(0)||\nonumber\\
&+ ||g(Y^M_n)-g(0)||+||g(0)||+||h(Y^M_n)-h(0)||+||h(0)||\nonumber\\
&\leq 1+C||Y^M_n||+T||u_{\lambda}(0)||+TC(K+||Y^M_n||^c+||0||^c)||Y^M_n-0||+T||v(0)||\nonumber\\
&+C||Y^M_n||+C||Y^M_n||+||g(0)||+||h(0)||\nonumber\\
&\leq 1+KTC +TC+3C++T||u_{\lambda}(0)||+T||v(0)||+||g(0)||+||h(0)||\leq \beta.
\label{ch5normY}
\end{aligned}$$ Futhermore, from the numerical approximation , we have $$\begin{aligned}
\label{ch5partnorm2}
||Y^M_{n+1}||^2&=&||Y^M_n||^2+||u_{\lambda}(Y^M_n)||^2\Delta t^2+\dfrac{\Delta t^2||v(Y^M_n)||^2}{(1+\Delta t||f_{\lambda}(Y^M_n)||)^2}+||g(Y^M_n)\Delta W^M_n||^2\nonumber\\
&+& ||h(Y^M_n)\Delta\overline{N}^M_n||^2+ 2\Delta t\langle Y^M_n, u_{\lambda}(Y^M_n)\rangle+\dfrac{2\Delta t\langle Y^M_n, v(Y^M_n)\rangle}{1+\Delta t||f_{\lambda}(Y^M_n)||}\nonumber\\
&+&2\langle Y^M_n,g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle+\dfrac{2\Delta t^2\langle u_{\lambda}(Y^M_n), v(Y^M_n)\rangle}{1+\Delta t||v(Y^M_n)||}\nonumber\\
&+&2\Delta t\langle u_{\lambda}(Y^M_n), g(Y^M_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&+&\dfrac{2\langle \Delta tv(Y^M_n),g(Y^M_n)\Delta W^M_n\rangle}{1+\Delta t||v(Y^M_n)||}+\dfrac{2\langle \Delta tv(Y^M_n),h(Y^M_n)\Delta\overline{N}^M_n\rangle}{1+\Delta t||v(Y^M_n)||}\nonumber\\
&+&2\langle g(Y^M_n)\Delta W^M_n,h(Y^M_n)\Delta\overline{N}^M_n\rangle.
\end{aligned}$$ Using the estimations $a\leq |a|$ and $
\dfrac{1}{1+\Delta t||v(Y^M_n)||}\leq 1$, we obtain the following inequality from equation : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq& ||Y^M_n||^2+||u_{\lambda}(Y^M_n)||\Delta t^2+\Delta t^2||v(Y^M_n)||^2+||g(Y^M_n)||^2||\Delta W^M_n||^2\nonumber\\
&+&||h(Y^M_n)||^2||\Delta\overline{N}^M_n||^2+2\Delta t\langle Y^M_n, u_{\lambda}(Y^M_n)\rangle+2\Delta t|\langle Y^M_n, v(Y^M_n)\rangle|\nonumber\\
&+&2\langle Y^M_n, g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n, h(Y^M_n)\Delta\overline{N}^M_n\rangle+2\Delta t^2|\langle u_{\lambda}(Y^M_n), v(Y^M_n)\rangle|\nonumber\\
&+&2\langle \Delta tu_{\lambda}(Y^M_n), g(Y^M_n)\Delta W_n^M\rangle+2\langle \Delta tu_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&+&2\Delta t|\langle v(Y^M_n), g(Y^M_n)\Delta W^M_n\rangle|+2\Delta t|\langle v(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle|\nonumber\\
&+&2\langle g(Y^M_n)\Delta W^M_n, h(Y^M_n)\Delta\overline{N}^M_n\rangle.
\label{ch5ine14}
\end{aligned}$$ Using the estimation $2ab\leq a^2+b^2$, the inequality becomes : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq & ||Y^M_n||^2+||u_{\lambda}(Y^M_n)||^2\Delta t^2+||v(Y^M_n)||^2\Delta t^2+||g(Y^M_n)||^2||\Delta W^M_n||^2\nonumber\\
&+&||h(Y^M_n)||^2||\Delta\overline{N}^M_n||^2+2\Delta t\langle Y^M_n, u_{\lambda}(Y^M_n)\rangle+2\Delta t|\langle Y^M_n, v(Y^M_n)\rangle|\nonumber\\
&+&2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), g(Y^M_n)\Delta W^M_n\rangle+2\rangle Y^M_n+\Delta tu_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&+&||u_{\lambda}(Y^M_n)||^2\Delta t^2+||u_{\lambda}(Y^M_n)||^2\Delta t^2+||v(Y^M_n)||^2\Delta t^2+||h(Y^M_n)||^2||\Delta\overline{N}^M_n||^2\nonumber\\
&+&||v(Y^M_n)||^2\Delta t^2+||g(Y^M_n)||^2||\Delta W^M_n||^2+||v(Y^M_n)||^2\Delta t^2+||h(Y^M_n)||^2||\Delta\overline{N}^M_n||^2\nonumber\\
&+&||g(Y^M_n)||^2||\Delta W^M_n||^2+||h(Y^M_n)||^2||\Delta\overline{N}^M_n||^2
\label{ch5ine15}
\end{aligned}$$ Putting similars terms of inequality together, we obtain : $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq & ||Y^M_n||^2+3||u_{\lambda}(Y^M_n)||^2\Delta t^2+4||v(Y^M_n)||^2\Delta t^2+3||g(Y^M_n)||^2||\Delta W^M_n||^2\nonumber\\
&+&4||h(Y^M_n)||^2||\Delta\overline{N}^M_n||^2+2\Delta t\langle Y^M_n, u_{\lambda}(Y^M_n)\rangle+2\Delta t|\langle Y^M_n, v(Y^M_n)\rangle|\nonumber\\
&+&2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), g(Y^M_n)\Delta W^M_n\rangle\nonumber\\
&+&2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle,
\label{ch5ine16}
\end{aligned}$$ on $\Omega$, for all $M\in\mathbb{N}$ and all $n\in\{0,1,\cdots, M-1\}$.
In addition, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, the global Lipschitz condition satisfied by $g$, $h$ and $u_{\lambda}$ leads to : $$\begin{aligned}
||g(x)||^2&\leq& (||g(x)-g(0)||+||g(0)||)^2\nonumber\\
&\leq & (C||x||+||g(0)||)^2\nonumber\\
&\leq & (C+||g(0)||)^2||x||^2\nonumber\\
&\leq &\beta||x||^2
\label{ch5normdeg2}
\end{aligned}$$ Along the same lines as above, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, we have : $$\begin{aligned}
||h(x)||^2\leq \beta||x||^2 \hspace{0.3cm} \text{and} \hspace{0.3cm} ||u_{\lambda}(x)||\leq \beta ||x||.
\label{ch5normdeh2}
\end{aligned}$$ Also, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, the one-sided Lipschitz condition satisfied by $v$ leads to : $$\begin{aligned}
\langle x,v(x)\rangle&=&\langle x,v(x)-v(0)+v(0)\rangle=\langle x-0,v(x)-v(0)\rangle +\langle x,v(0)\rangle\nonumber\\
&\leq & C||x||^2+||x||||v(0)||\nonumber\\
&\leq &(C+||v(0)||)||x||^2\nonumber\\
&\leq &\sqrt{\beta}||x||^2.
\label{ch5crochetv}
\end{aligned}$$ Along the same lines as above, $\forall x\in\mathbb{R}^d$ such that $||x||\geq 1$, we have $$\begin{aligned}
\langle x, u(x)\rangle\leq \sqrt{\beta}||x||^2.
\label{ch5crochetu}
\end{aligned}$$ Futhermore, using the polynomial growth condition satisfied by $v$, the following inequality holds for all $x\in\mathbb{R}^d$ with $1\leq ||x||\leq M^{1/2c}$ and for all $M\in\mathbb{N}$ $$\begin{aligned}
||v(x)||^2&\leq&(||v(x)-v(0)||+||v(0)||)^2\nonumber\\
&\leq & (C(K+||x||^c)||x||+||v(0)||)^2\nonumber\\
&\leq & (C(K+1)||x||^{c+1}+||v(0)||)^2\nonumber\\
&\leq & (KC+C+||v(0)||)^2||x||^{2(c+1)}\nonumber\\
&\leq & M\sqrt{\beta}||x||^2.
\label{ch5normv2}
\end{aligned}$$ Using the global-Lipschitz condition of $u_{\lambda}$ leads to $$\begin{aligned}
||u_{\lambda}(x)||^2\leq \sqrt{\beta}||x||^2.
\label{ch5normu2}
\end{aligned}$$ Now combining inequalities , , , ,, and we obtain $$\begin{aligned}
||Y^M_{n+1}||^2 &\leq& ||Y^M_n||^2+\dfrac{3T^2\sqrt{\beta}}{M}||Y^M_n||^2+\dfrac{4T^2\sqrt{\beta}}{M}||Y^M_n||^2+3\beta||Y^M_n||^2||\Delta W^M_n||^2\nonumber\\
&+&+4\beta||Y^M_n||^2||\Delta\overline{N}^M_n||^2+\dfrac{2T\sqrt{\beta}}{M}||Y^M_n||^2+\dfrac{2T\sqrt{\beta}}{M}||Y^M_n||^2\nonumber\\
&+&2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle\\
&\leq &||Y^M_n||^2+\left(\dfrac{8T^2\sqrt{\beta}}{M}+\dfrac{4T\sqrt{\beta}}{M}\right)||Y^M_n||^2+4\beta||Y^M_n||^2||\Delta W^M_n||^2+4\beta||Y^M_n||^2||\Delta\overline{N}^M_n||^2\\
&+&2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n), h(Y^M_n)\Delta\overline{N}^M_n\rangle.
\end{aligned}$$ Using the inequality $8T^2+4T\leq 8\sqrt{\beta}$, it follows that : $$\begin{aligned}
||Y^M_{n+1}||^2&\leq& ||Y^M_n||^2+\dfrac{8\beta}{M}||Y^M_n||^2+4\beta||Y^M_n||^2||\Delta W^M_n||^2+4\beta||Y^M_n||^2||\Delta\overline{N}^M_n||^2\nonumber\\
&+&2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n),g(Y^M_n)\Delta W^M_n\rangle+2\langle Y^M_n+\Delta tu_{\lambda}(Y^M_n),h(Y^M_n)\Delta\overline{N}^M_n\rangle\nonumber\\
&=&||Y^M_n||^2 \left(1+\dfrac{8\beta}{M}+4\beta||\Delta W^M_n||^2+4\beta||\Delta\overline{N}^M_n||^2 \right.\nonumber\\
&+&\left. \left\langle\dfrac{Y^M_n+\Delta tu_{\lambda}(Y^M_n)}{||Y^M_n||}, \dfrac{g(Y^M_n)}{||Y^M_n||}\Delta W^M_n\right\rangle+2\left\langle\dfrac{Y^M_n+\Delta tu_{\lambda}(Y^M_n)}{||Y^M_n||}, \dfrac{h(Y^M_n)}{||Y^M_n||}\Delta\overline{N}^M_n\right\rangle\right)\nonumber\\
&=&||Y^M_n||^2\left(1+\dfrac{8\beta}{M}+4\beta||\Delta W^M_n||^2+4\beta||\Delta\overline{N}^M_n||^2+2\alpha^M_n+2\beta^M_n\right).
\label{ch5expY1}
\end{aligned}$$ Using Lemma \[ch4lemma1\] for $a=\dfrac{8\beta}{M}+4\beta||\Delta W^M_n||^2+2\alpha^M_n+2\beta^M_n$ and $b=2\sqrt{\beta}||\Delta\overline{N}^M_n|| $ it follows from that : $$\begin{aligned}
||Y^M_{n+1}||^2\leq ||Y^M_n||^2\exp\left(\dfrac{8\beta}{M}+4\beta||\Delta W^M_n||^2+4\beta||\Delta\overline{N}^M_n||+2\alpha^M_n+2\beta^M_n\right),
\label{ch5expY2}
\end{aligned}$$ on $\{w\in\Omega : 1\leq ||Y^M_n(\omega)||\leq M^{1/2c}\}$, for all $M\in\mathbb{N}$ and $n\in\{0,\cdots,M-1\}.$ Now combining and and using mathematical induction as used in the proof of Lemma \[ch4lemma2\] complete the proof of Lemma \[ch5lemma2\].
The following lemma and its proof are similars to [@Martin1 Lemma 3.3 pp 15] with only different value of the coefficient $\beta$.
\[ch5lemma4\] The following inequality holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}, M\geq 4\beta pT}\mathbb{E}\left[\exp\left(\beta p\sum_{k=0}^{M-1}||\Delta W^M_k||^2\right)\right]<\infty.
\end{aligned}$$
The following lemma is based on [@Martin2 Lemma 5.7].
\[ch5lemma6\] The following inequality holds $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]\leq \exp\left[\dfrac{p^2T(1+TC+T||u(0)||)^2(C+||g(0)||)^2}{M}\right].
\end{aligned}$$
Let $a^{\top}$ stand for the transposed of a vector $a$, let $Y$ be the $m$ column vector defined by : $Y=\sqrt{\dfrac{T}{M}}(1,\cdots,1)$ and let $\mathcal{N}(0,1)$ be a $1$-dimensional standard normal random variable. Then we have $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]= \mathbb{E}\left[\exp\left(pz\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{g(x)}{||x||}Y\mathcal{N}(0,1)\right>\right)\right].
\end{aligned}$$ Using Lemma \[ch4lemma5\] it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]&\leq&\exp\left[p^2z^2\left\|\dfrac{g(x)^{\top}(x+u(x)T/M)Y||}{||x||^2}\right\|^2\right]\\
&=&\exp\left[\dfrac{p^2T}{M}\dfrac{||g(x)^{\top}(x+u(x)T/M)||^2}{||x||^4}\right]\nonumber\\
&\leq & \exp\left[\dfrac{p^2T}{M}\dfrac{||g(x)||^2||x+u(x)T/M||^2}{||x||^4}\right].
\end{aligned}$$ From the global Lipschitz condition, for all $x\in\mathbb{R}^d$ such that $||x||\geq1$ we have $$\begin{aligned}
||g(x)||^2\leq (||g(x)-u(0)||+||g(0)||)^2\leq (C||x||^2+||g(0)||)^2\leq (C+||g(0)||)^2||x||^2
\end{aligned}$$ $$\begin{aligned}
||x+u(x)T/M||\leq ||x||+T/M||u(x)||&\leq&||x||+T/M||u(x)-u(0)||+T/M||u(0)||\\
&\leq &||x||+TC||x||+T||u(0)||\\
&\leq &(1+TC+T||u(0)||)||x||.
\end{aligned}$$ Therefore, it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz_{\{||x||\geq 1\}}\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{g(x)}{||x||}\Delta W^M_k\right>\right)\right]\leq \exp\left[\dfrac{p^2T(1+TC+T||u(0)||)^2(C+||g(0)||)^2}{M}\right].
\end{aligned}$$
for all $M\in\mathbb{N}$, $k\in\{0,\cdots,M-1\}$, $p\in[1,\infty)$ and $z\in\{-1,1\}$.
Following closely [@Xia2 Lemma 2.3 ] we have the following lemma.
\[ch5lemma7\] Let $\alpha^M_n :\Omega\longrightarrow\mathbb{R}$ for $M\in\mathbb{N}$ and $n\in\{0,1,\cdots,M\}$ define as in notation . Then the following inequality holds : $$\begin{aligned}
\sup_{z\in\{-1,1\}}\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,1,\cdots,M\}}\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^p(\Omega, \mathbb{R})}<\infty,
\end{aligned}$$ for all $p\in[2,+\infty)$
The time discrete stochastic process $z\sum_{k=0}^{n-1}\alpha^M_k$, $n\in\{0,1,\cdots, M\}$ is an $(\mathcal{F}_{nT/M})_{n\in\{0,\cdots,M\}}-$ martingale for every $z\in\{-1,1\}$ and $M\in \mathbb{N}$. So $\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)$ is a positive $(\mathcal{F}_{nT/M})_{n\in\{0,\cdots,M\}}-$ submartingale for every $z\in\{-1,1\}$ and $M\in\mathbb{N}$ since $\exp$ is a convex function.
Applying Doop’s maximal inequality leads to : $$\begin{aligned}
\left\|\sup_{n\in\{0,\cdots,M\}}\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^p{(\Omega, \mathbb{R})}}&=&\left(\mathbb{E}\left|\sup_{n\in\{0,\cdots,M\}}\exp\left(pz\sum_{k=0}^{n-1}\alpha^M_k\right)\right|\right)^{1/p}\nonumber\\
&\leq &\left(\dfrac{p}{p-1}\right)\left(\mathbb{E}\left |\exp\left(pz\sum_{k=0}^{M-1}\alpha^M_k\right)\right|\right)^{1/p}\nonumber\\
&= &\dfrac{p}{p-1}\left\|\exp\left(z\sum_{k=0}^{M-1}\alpha^M_k\right)\right\|_{L^p{(\Omega,\mathbb{R})}}.
\label{ch5alpha1}
\end{aligned}$$
Using Lemma \[ch5lemma6\], it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp(pz\alpha^M_k)/\mathcal{F}_{kT/M}\right]\leq\exp\left(\dfrac{p^2T(C+||g(0)||)^2(1+TC+T||u(0)||)^2}{M}\right).
\label{ch5alpha4}
\end{aligned}$$ Using inequality , it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-1}\alpha^M_k\right)\right]&=&\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-2}
\alpha^M_k\right)\mathbb{E}[\exp(p\alpha^M_{M-1}/\mathcal{F}_{(M-1)T/M}\right]\\
&\leq &\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-2}\alpha^M_k\right)\right]\exp\left(\dfrac{p^2T(C+||g(0)||)^2(1+TC+T||u(0)||)^2}{M}\right).
\end{aligned}$$ Iterating the previous inequality $M$ times gives : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\sum_{k=0}^{M-1}\alpha^M_k\right)\right] \leq \exp(p^2T(C+||g(0)||)^2(1+TC+T||u(0)||)^2).
\label{ch5alpha5}
\end{aligned}$$ Now combining inequalities and leads to $$\begin{aligned}
\sup_{z\in\{-1,1\}}\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,\cdots,M\}}\exp\left(z\sum_{k=0}^{n-1}\alpha^M_k\right)\right\|_{L^p(\Omega,\mathbb{R})} &\leq& 2\exp(p^2T(C+||g(0)||)^2(1+TC+T||u(0)||)^2) \\
&< &+\infty,
\end{aligned}$$ for all $p\in[2,\infty)$.
\[ch5lemma9\] The following inequality holds
$
\mathbb{E}\left[\exp\left(pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right]
$ $$\begin{aligned}
\leq\left[\exp\left(\dfrac{\left[e^{p(C+||h(0)||)(1+TC+T||u(0)||)}+p(C+||h(0)||)(1+TC+T||u(0)||)\right]\lambda T}{M}\right)\right],
\end{aligned}$$ for all $M\in\mathbb{N}$, all $p\in[1, +\infty)$ and all $n\in\{0,\cdots,M\}$, $z\in\{-1,1\}$.
For $x\in\mathbb{R}^d$ such that $||x||\neq 0$, we have : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right] &\leq& \mathbb{E}\left[\exp\left(pz\dfrac{||x+u(x)T/M||||h(x)||}{||x||^2}\Delta\overline{N}^M_n\right)\right].
\end{aligned}$$ Using the global Lipschitz condition on $h$, for all $x\in\mathbb{R}^d$ such that $||x||\geq 1$, we have : $$\begin{aligned}
||h(x)||\leq ||h(x)-h(0)||+||h(0)||\leq (C+||h(0)||)||x||. \label{ch5normeh}\\
||x+u(x)T/M||\leq (1+TC+T||u(0)||)||x||\label{ch5normeh1}
\end{aligned}$$ So using inequalities and , it follows that : $$\begin{aligned}
\mathbb{E}\left[\exp\left(pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right]
\leq\mathbb{E}\left( \exp[pz(C+||h(0)||)(1+TC+T||u(0)||)\Delta\overline{N}^M_n \right).
\end{aligned}$$ Using Lemma \[ch4lemma8\], it follows that :
$
\left[\exp\left(pz\mathbf{1}_{\{||x||\geq 1\}}\left<\dfrac{x+u(x)T/M}{||x||},\dfrac{h(x)}{||x||}\Delta\overline{N}^M_n\right>\right)\right]
$ $$\begin{aligned}
&\leq &\mathbb{E}[\exp(pz(C+||h(0)||)(1+TC+T||u(0)||)\Delta\overline{N}^M_n)]\\
&\leq &\left[\exp\left(\dfrac{\left[e^{p(C+||h(0)||)(1+TC+T||u(0)||)}+p[(C+||h(0)||)(1+TC+T||u(0)||)-1\right]\lambda T}{M}\right)\right]\\
&\leq &\left[\exp\left(\dfrac{\left[e^{p(C+||h(0)||)(1+TC+T||u(0)||)}+p[(C+||h(0)||)(1+TC+T||u(0)||)\right]\lambda T}{M}\right)\right].
\end{aligned}$$
The following lemma is similar to [@Xia2 Lemma 2.3].
\[ch5lemma10\] Let $\beta^M_n :\Omega \longrightarrow \mathbb{R}$ defined in Notation \[ch5notation1\] for all $M\in\mathbb{N}$ and all $n\in\{0,\cdots,M\}$ then we have the following inequality $$\begin{aligned}
\sup_{z\in\{-1,1\}}\sup_{M\in\mathbb{N}}\left\|\sup_{n\in\{0,\cdots, M\}}\exp\left(z\sum_{K=0}^{n-1}\beta^M_k\right)\right\|_{L^p(\Omega, \mathbb{R})}<+\infty.
\end{aligned}$$
Following the proof of [@Martin1 Lemma 3.4 ], the result is straightforward using lemmas \[ch5lemma9\] and \[ch5lemma6\].
\[ch5lemma11\] The following inequality holds $$\begin{aligned}
\sup_{M\in\mathbb{N}}\mathbb{E}\left[\exp\left(p\beta\sum_{k=0}^{M-1}||\Delta\overline{N}^M_k||\right)\right]<+\infty,\end{aligned}$$ for all $p\in[1, +\infty)$.
The proof is similar to the proof of Lemma \[ch4lemma11\] with only different value of $\beta$.
The following lemma is based on Lemma \[ch5lemma10\].
\[ch5lemma12\] \[Uniformly bounded moments of the dominating stochastic processes\].
Let $M\in\mathbb{N}$ and $D_n^M : \Omega \longrightarrow [0,\infty)$ for $n\in\{0,1,\cdots, M\}$ be define as in notation . Then we have : $$\begin{aligned}
\sup_{M\in\mathbb{N}, M\geq8\lambda pT}\left\|\sup_{n\in\{0,1,\cdots, M\}}D_n^M\right\|_{L^p(\Omega, \mathbb{R})}<\infty,
\end{aligned}$$ for all $p\in[1,\infty)$.
The following lemma is an extension of [@Martin1 Lemma 3.6, pp 16 ]. Here, we include the jump part.
\[ch5lemma13\] Let $M\in\mathbb{N}$ and $\Omega_M^M\in\mathcal{F}$. The following inequality holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\left(M^p\mathbb{P}[(\Omega_M^M)^c]\right)<+\infty,
\end{aligned}$$ for all $p\in[1,\infty)$.
**\[Theorem \[ch5theorem1\]\]**.
Let’s first represent the numerical approximation $Y^M_n$ in the following appropriate form : $$\begin{aligned}
Y^M_n&=&Y_{n-1}^M +u_{\lambda}(Y^M_{n-1})T/M+\dfrac{\Delta tv(Y^M_{n-1})}{1+\Delta t||v(Y^M_{n-1})||}+g(Y_{n-1})\Delta W^M_{n-1}+h(Y^M_{n-1})\Delta\overline{N}^M_{n-1}\\
&=&X_0+\sum_{k=0}^{n-1}u(Y^M_{k})T/M+\sum_{k=0}^{n-1}\dfrac{\Delta t v(Y_k^M)}{1+\Delta t||v(Y^M_k)||}+\sum_{k=0}^{n-1}g(Y^M_k)\Delta W^M_k+\sum_{k=0}^{n-1}h(Y^M_k)\Delta\overline{N}^M_k\\
&=& X_0+u(0)nT/M+ \sum_{k=0}^{n-1}g(0)\Delta W^M_k+\sum_{k=0}^{n-1}h(0)\Delta\overline{N}^M_k+\sum_{=0}^{n-1}T/M(u(^M_k)-u(0))\\
&+&\sum_{k=0}^{n-1}\dfrac{\Delta tv(Y^M_{k})}{1+\Delta t||v(Y^M_{k})||}+\sum_{k=0}^{n-1}(g(Y^M_k)-g(0))\Delta W^M_k+\sum_{k=0}^{n-1}(h(Y^M_k)-h(0))\Delta\overline{N}^M_k,
\end{aligned}$$ for all $M\in\mathbb{N} $ and all $n\in\{0,\cdots,M\}$.
Using the inequality $$\begin{aligned}
\left\|\dfrac{\Delta tv(Y^M_k)}{1+\Delta t||v(Y^M_k)||}\right\|_{L^P(\Omega, \mathbb{R}^d)} <1\end{aligned}$$ it follows that : $$\begin{aligned}
||Y^M_n||_{L^p(\Omega, \mathbb{R}^d)} &\leq &||X_0||_{L^p(\Omega,\mathbb{R}^d)}+ ||u(0)||nT/M+\left\|\sum_{k=0}^{n-1}g(0)\Delta W^M_k\right\|_{L^p(\Omega, \mathbb{R}^d)}+M\\
&+&\left\|\sum_{k=0}^{n-1}h(0)\Delta\overline{N}^M_k\right\|_{L^p(\Omega,\mathbb{R}^d)}+\left\|\sum_{k=0}^{n-1}(g(Y^M_k)-g(0))\Delta W^M_k\right\|_{L^p(\Omega,\mathbb{R}^d)}\\
&+&\left\|\sum_{k=0}^{n-1}(h(Y^M_k)-h(0))\Delta\overline{N}^M_k\right\|_{L^p(\Omega,\mathbb{R}^d)}.\end{aligned}$$ Using Lemma \[ch4lemma16\] and Lemma \[ch4lemma19\], it follows that $$\begin{aligned}
||Y^M_n||_{L^p(\Omega,\mathbb{R}^d)} &\leq &||X_0||_{L^p(\Omega,\mathbb{R})}+||u(0)||nT/M+C_p\left(\sum_{k=0}^{n-1}\sum_{i=1}^{m}||g_i(0)||^2\dfrac{T}{M}\right)^{1/2}+C_p\left(\sum_{k=0}^{n-1}||h(0)||^2\dfrac{T}{M}\right)^{1/2}\nonumber\\
&+&M+ C_p\left(\sum_{k=0}^{n-1}\sum_{i=1}^m||(g_i(Y_k^M)-g_i(0))\Delta W^M_k||^2_{L^p(\Omega, \mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}\nonumber\\ &+&C_p\left(\sum_{k=0}^{n-1}\lambda||(h(Y_k^M)-h(0))\Delta W^M_k||^2_{L^p(\Omega, \mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}\nonumber\\
&\leq&||X_0||_{L^p(\Omega, \mathbb{R}^d)}+T||u(0)||+C_p\left(\dfrac{nT}{M}\sum_{i=1}^m||g_i(0)||^2\right)^{1/2}+C_p\left(\dfrac{nT}{M}||h(0)||^2\right)^{1/2}\nonumber\\
&+&M+C_p\left(\sum_{k=0}^{n-1}\sum_{i=1}^m||g_i(Y^M_k)-g_i(0)||^2_{L^p(\Omega,\mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}\nonumber\\
&+&C_p\left(\sum_{k=0}^{n-1}||h(Y^M_k)-h(0)||^2_{L^p(\Omega,\mathbb{R}^d)}\dfrac{T}{M}\right)^{1/2}.
\label{ch5MB1}\end{aligned}$$ From $||g_i(0)||^2\leq ||g(0)||^2$ and the global Lipschitz condition satisfied by $g$ and $h$, we obtain $||g_i(Y^M_k)-g_i(0)||\leq C||Y^M_k||_{L^p(\Omega,\mathbb{R}^d)}$ and $||h(Y^M_k)-h(0)||\leq C||Y^M_k||_{L^p(\Omega,\mathbb{R}^d)}$. So using , we have : $$\begin{aligned}
||Y^N_n||_{L^p(\Omega, \mathbb{R}^d)} &\leq & ||X_0||_{L^p(\Omega,\mathbb{R}^d)}+T||u(0)||+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\\
&+&C_p\left(\dfrac{Tm}{M}\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega,\mathbb{R}^d)}\right)^{1/2}
+C_p\left(\dfrac{T}{M}\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega,\mathbb{R}^d)}\right)^{1/2}.\end{aligned}$$ Using the inequality $(a+b+c)^2\leq 3a^2+3b^2+3c^2$, it follows that : $$\begin{aligned}
||Y^M_n||^2_{L^p(\Omega,\mathbb{R}^d)}&\leq& 3\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+T||u(0)||+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)^2\nonumber\\
&+&\dfrac{3T(C_p\sqrt{m}+C_p)^2}{M}\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega, \mathbb{R}^d)},\end{aligned}$$ Using the inequality the inequality $
\dfrac{3T(C_p\sqrt{m}+C_p)^2}{M}<3(C_p\sqrt{m}+C_p)^2
$, it follows that :
$$\begin{aligned}
||Y^M_n||^2_{L^p(\Omega,\mathbb{R}^d)}&\leq& 3\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+T||u(0)||+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)^2\nonumber\\
&+&3T(C_p\sqrt{m}+C_p)^2\sum_{k=0}^{n-1}||Y^M_k||^2_{L^p(\Omega, \mathbb{R}^d)},
\label{ch5MB2}\end{aligned}$$
for all $p\in[2,\infty)$.
Applying Gronwall inequality to leads to : $$\begin{aligned}
||Y^M_n||^2_{L^p(\Omega,\mathbb{R}^d)}\leq 3e^{3(C_p\sqrt{m}+C_p)^2}\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+T||u(0)||+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right)^2
\label{ch5MB3}\end{aligned}$$ Taking the square root and the supremum in both sides of the previous inequality leads to :
$\sup\limits_{n\in\{0,\cdots, M\}}||Y^M_n||_{L^p(\Omega, \mathbb{R}^d)}$ $$\begin{aligned}
\leq\sqrt{3}e^{3(C_p\sqrt{m}+C_p)^2}\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+T||u(0)||+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+M\right).
\label{ch5MB4}\end{aligned}$$ Unfortunately, is not enough to conclude the proof of the theorem due to the term $M$ in the right hand side. Using the fact that $(\Omega_n^M)_n$ is a decreasing sequence and exploiting Holder inequality, we obtain : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left\|\mathbf{1}_{(\Omega_n^M)^c}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}& \leq &\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega,\mathbb{R}^d)}\left\|Y^M_n\right\|_{L^{2p}(\Omega,\mathbb{R}^d)}\nonumber\\
&\leq &\left(\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left(M\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega,\mathbb{R}^d)}\right)\right)\nonumber\\
&\times &\left(\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots,M\}}\left(M^{-1}||Y^M_n||_{L^{2p}(\Omega,\mathbb{R}^d)}\right)\right).
\label{ch5MB5}\end{aligned}$$ From inequality we have
$
\left(\sup\limits_{M\in\mathbb{N}}\sup\limits_{n\in\{0,\cdots,M\}}\left(M^{-1}||Y^M_n||_{L^p(\Omega,\mathbb{R}^d)}\right)\right)$ $$\begin{aligned}
\leq \sqrt{2}e^{(C_p\sqrt{m}+C_p)^2}\left(\dfrac{||X_0||_{L^p(\Omega, \mathbb{R}^d)}}{M}+\dfrac{C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||}{M}+1\right)\nonumber\\
\leq \sqrt{2}e^{(p\sqrt{m}+C_p)^2}\left(||X_0||_{L^p(\Omega,\mathbb{R}^d)}+C_p\sqrt{Tm}||g(0)||+C_p\sqrt{T}||h(0)||+1\right)<+\infty,
\label{ch5MB6}\end{aligned}$$ From the relation $$\begin{aligned}
\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega, \mathbb{R}^d)}=
\mathbb{E}\left[\mathbf{1}_{(\Omega^M_M)^c}\right]^{1/2p}=
\mathbb{P}\left[(\Omega^M_M)^c\right]^{1/2p},\end{aligned}$$ it follows using Lemma \[ch5lemma13\] that : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left(M\left\|\mathbf{1}_{(\Omega^M_M)^c}\right\|_{L^{2p}(\Omega,\mathbb{R}}\right)=\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left(M^{2p}\mathbb{P}\left[(\Omega^M_M)^c\right]\right)^{1/2p}<+\infty.
\label{ch5MB7}\end{aligned}$$ So plugging and in leads to : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)^c}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty.
\label{ch5MB8}\end{aligned}$$ Futhermore, we have $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)} &\leq &\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}\nonumber\\
&+&\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)^c}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}.
\label{ch5MB9}\end{aligned}$$ From the second term of inequality is bounded, while using Lemma \[ch5lemma2\] and Lemma \[ch5lemma12\] we have : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|\mathbf{1}_{(\Omega^M_n)}Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}\leq \sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|D_n^M\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty.
\label{ch5MB10}\end{aligned}$$ Finally plugging and in leads to : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,\cdots, M\}}\left\|Y^M_n\right\|_{L^p(\Omega,\mathbb{R}^d)}<+\infty.\end{aligned}$$ This complete the proof of Theorem \[ch5theorem1\].
Strong convergence of the semi-tamed Euler scheme
-------------------------------------------------
\[ch5theorem2\] Under Assumptions \[ch5assumption1\], for all $p\in[1,+\infty)$ there exist a positive constant $C_p$ such that : $$\begin{aligned}
\left(\mathbb{E}\left[\sup_{t\in[0,T]}\left\|X_t-\overline{Y}^M_t\right\|^p\right]\right)^{1/p}\leq C_p\Delta t^{1/2},
\label{ch5inetheo}
\end{aligned}$$ for all $M\in \mathbb{N}$.
Where $X : [0,T]\times \Omega\longrightarrow \mathbb{R}^d$ is the exact solution of equation and $\overline{Y}^M_t$ is the time continous solution defined in .
In order to prove Theorem \[ch5theorem2\], we need the following two lemmas.
\[ch5lemma14\]\[Based on [@Martin1 Lemma 3.10, pp 16]\].
Let $Y_n^M$ be defined by for all $M\in\mathbb{N}$ and all $n\in\{0,1,\cdots, M\}$. Then the following inequalities holds : $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{n\in\{0,1,\cdots, M\}}\left(\mathbb{E}\left[||u_{\lambda}(Y_n^M)||^p\right]\vee \mathbb{E}[||v(Y^M_n)||^p]\vee \mathbb{E}\left[||g(Y_n^M)||^p\right]\vee \mathbb{E}\left[||h(Y_n^M)||^p\right]\right)<+\infty,\end{aligned}$$ for all $p\in[1,\infty)$.
The proof is similar to the proof [@Martin1 Lemma 3.10, pp 16].
For $s\in[0,T]$ let $\lfloor s\rfloor$ be the greatest grid point less than $s$. We have the following lemma.
\[ch5lemma16\] The following inequalities holds for any stepsize $\Delta t$. $$\begin{aligned}
\sup_{t\in[0,T]}\left\|\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\leq C_p\Delta t^{1/2},
\end{aligned}$$ $$\begin{aligned}
\sup_{M\in\mathbb{N}}\sup_{t\in[0,T]}\left\|\overline{Y}^M_t\right\|_{L^p(\Omega, \mathbb{R}^d)}<\infty,
\end{aligned}$$ $$\begin{aligned}
\sup_{t\in[0,T]}\left\|v(\overline{Y}^M_t)-v(\overline{Y}^M_{\lfloor t\rfloor})\right\|_{L^p(\Omega, \mathbb{R}^d)}\leq C_p\Delta t^{1/2}.
\end{aligned}$$
- Using the time continous approximation , Lemma \[ch4lemma19\] and Lemma \[ch4lemma16\], it follows that :
$
\sup\limits_{t\in[0,T]}||\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}||_{L^p(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq &\dfrac{T}{M}\sup_{t\in[0,T]}\left\|u_{\lambda}(\overline{Y}^M_{\lfloor t\rfloor})\right\|_{L^p(\Omega, \mathbb{R}^d)}+\dfrac{T}{M}\left(\sup_{t\in[0,T]}\left\|\dfrac{v(\overline{Y}^M_{\lfloor t\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor t\rfloor})||}\right\|_{L^p(\Omega, \mathbb{R}^d)}\right)\nonumber\\
&+& \sup_{t\in[0, T]}\left\|\int^t_{\lfloor t\rfloor} g(\overline{Y}^M_{\lfloor t\rfloor})dW_s\right\|_{L^p(\Omega, \mathbb{R}^d)}+\sup_{t\in[0,T]}\left\|\int^t_{\lfloor t\rfloor}h(\overline{Y}^M_{\lfloor t\rfloor})d\overline{N}_s\right\|_{L^p(\Omega, \mathbb{R}^d)}\nonumber\\
&\leq &\dfrac{T}{\sqrt{M}}\left(\sup_{n\in\{0,\cdots M\}}\|u_{\lambda}(Y^M_n)\|_{L^p(\Omega,\mathbb{R}^d)}\right)+\dfrac{T}{\sqrt{M}}\left(\sup_{n\in\{0,\cdots, M\}}||v(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)\nonumber\\
&+& C_p\sup_{t\in[0,T]}\left(\dfrac{T}{M}\sum_{i=1}^m\int^t_{\lfloor t\rfloor}||g_i(\overline{Y}^M_s)||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right)^{1/2}+C_p\sup_{t\in[0,T]}\left(\dfrac{TC_p}{M}\int^t_{\lfloor t\rfloor}||h(\overline{Y}^m_s)||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right)^{1/2}\nonumber\\
&\leq &\dfrac{T}{\sqrt{M}}\left(\sup_{n\in\{0,\cdots, M\}}||u_{\lambda}(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)+\dfrac{T}{\sqrt{M}}\left(\sup_{n\in\{0,\cdots, M\}}||v(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)\nonumber\\
&+& \dfrac{C_p\sqrt{Tm}}{\sqrt{M}}\left(\sup_{i\in\{1,\cdots, m\}}\sup_{n\in\{0,\cdots, M\}}||g_i(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)\nonumber\\
&+&\dfrac{C_p\sqrt{T}}{\sqrt{M}}\left(\sup_{n\in\{0, \cdots,M\}}||h(Y^M_n)||_{L^p(\Omega, \mathbb{R}^d)}\right)
\label{Thcontinous}
\end{aligned}$$ for all $p\in[1,+\infty)$ and all $M\in\mathbb{N}$. Using inequality and Lemma \[ch5lemma14\] it follows that : $$\begin{aligned}
\left[\sup_{t\in[0,T]}||\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}||_{L^p(\Omega, \mathbb{R}^d)}\right]<C_p\Delta t^{1/2},
\label{Ch5bon1}\end{aligned}$$ for all $p\in[1, +\infty)$ and for all stepsize $\Delta t$.
- Using inequality , inequality $||a||\leq ||a-b||+||b||$ for all $a,b\in\mathbb{R}^d$ and Theorem \[ch5theorem1\] it follows that $$\begin{aligned}
\sup_{t\in[0,T]}||\overline{Y}^M_t||_{L^p(\Omega, \mathbb{R}^d)}&\leq&\left[\sup_{t\in[0,T]}\left\|\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\right]+\sup_{t\in[0,T]}\left\|\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\\
&\leq&C_p\Delta t^{1/2}+\sup_{t\in[0,T]}\left\|\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}\\
&<&C_pT^{1/2}+\sup_{t\in[0,T]}\left\|\overline{Y}^M_{\lfloor t\rfloor}\right\|_{L^p(\Omega, \mathbb{R}^d)}<\infty,\end{aligned}$$ for all $p\in[1,+\infty)$ and all $M\in\mathbb{N}$.
- Further, using the polynomial growth condition : $$\begin{aligned}
||v(x)-v(y)||\leq C(K+||x||^c+||y||^c)||x-y||,\end{aligned}$$ for all $x, y\in\mathbb{R}^d$, it follows using Holder inequality that : $$\begin{aligned}
\sup_{t\in[0,T]}||v(\overline{Y}^M_t)-v(\overline{Y}^M_{\lfloor t\rfloor})||_{L^p(\Omega, \mathbb{R}^d)} &\leq &C\left(K+2\sup_{t\in[0,T]}||\overline{Y}^M_t||^c_{L^{2pc}(\Omega, \mathbb{R}^d)}\right)\nonumber\\
&\times & \left(\sup_{t\in[0,T]}||\overline{Y}^M_t-\overline{Y}^M_{\lfloor t\rfloor}||_{L^{2p}(\Omega, \mathbb{R}^d)}\right)
\label{ch4Thfcontinous}\end{aligned}$$ Using and the first part of Lemma \[ch5lemma16\], the following inequality holds for all $p\in[1,+\infty)$ $$\begin{aligned}
\left[\sup_{t\in[0,T]}||v(\overline{Y}^M_t)-v(\overline{Y}^M_{\lfloor t\rfloor})||_{L^p(\Omega, \mathbb{R}^d)}\right]<C_p\Delta t^{1/2},
\label{ch4Thffinal}\end{aligned}$$ for all $p\in[1,\infty)$ and all $M\in\mathbb{N}$.
Now we are ready to prove Theorem \[ch5theorem2\].
**\[ Theorem \[ch5theorem2\]\]**
Let’s recall that for $z\in[0,T]$, $\lfloor z\rfloor$ is the greatest grid point less than $z$. The time continuous solution can be written into its integral form as bellow : $$\begin{aligned}
\overline{Y}^M_s=\varepsilon+\int_0^su(\overline{Y}^M_{\lfloor z\rfloor})dz+\int_0^s\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}dz+ \int g(\overline{Y}^M_{\lfloor z\rfloor})dW_z+\int h(\overline{Y}^M_{\lfloor z\rfloor})d\overline{N}_z,
\label{ch5continoussol2}
\end{aligned}$$ for all $z\in[0, T]$ almost sure (a.s) and all $M\in\mathbb{N}$.
Let’s estimate first the quantity $||X_s-\overline{Y}^M_s||^2$ $$\begin{aligned}
X_s-\overline{Y}_s&=&\int_0^s\left(u_{\lambda}(X_z)-u_{\lambda}(\overline{Y}^M_{\lfloor z \rfloor})\right)dz+\int_0^s\left(v(X_z)-\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right)dz\\
&+&\int_0^s\left(g(X_z)-g(\overline{Y}^M_{\lfloor z\rfloor})\right)dW_z+\int_0^s\left(h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})\right)d\overline{N}_z.
\end{aligned}$$ Using the relation $d\overline{N}_z=dN_z-\lambda dz$, it follows that $$\begin{aligned}
X_s-\overline{Y}_s&=&\int_0^s\left[\left(v(X_z)-\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right)+\left(u(X_z)-u(\overline{Y}^M_{\lfloor z\rfloor})\right)+\lambda\left(h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})\right)\right]dz\\
&+&\int_0^s\left(g(X_z)-g(\overline{Y}^M_{\lfloor z\rfloor})\right)dW_z+\int_0^s\left(h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})\right)dN_z.
\end{aligned}$$ The function $ k :\mathbb{R}^n\longrightarrow \mathbb{R}$, $x \longmapsto ||x||^2$ is twice differentiable. Applying Itô’s formula for jump process to the process $X_s-\overline{Y}^M_s$ with the function $k$ leads to : $$\begin{aligned}
||X_s-\overline{Y}^M_s||^2&=&2\int_0^s\left<X_z-\overline{Y}^M_z, v(X_z)-\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right>dz+2\lambda\int_0^s\left\langle X_z-\overline{Y}^M_z, h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz\\
&+&2\int_0^s\left\langle X_z-\overline{Y}^M_z, u(X_z)-u(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz +\sum_{i=1}^m\int_0^s||g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\\
&+&2\sum_{i=1}^m\int_0^s\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>dW^i_z\\
&+& \int_0^s\left[||X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2-||X_z-\overline{Y}^M_z||^2\right]dN_z.
\end{aligned}$$ Using again the relation $d\overline{N}_z=dN_z-\lambda dz$ leads to $$\begin{aligned}
||X_s-\overline{Y}^M_s||^2&=&2\int_0^s\left<X_z-\overline{Y}^M_z, v(X_z)-\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right>dz+2\lambda\int_0^s\left\langle X_z-\overline{Y}^M_z, h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz\nonumber\\
&+&2\int_0^s\left\langle X_z-\overline{Y}^M_z, u(X_z)-u(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz+\sum_{i=1}^m\int_0^s||g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})||^2dz \nonumber\\
&+&2\sum_{i=1}^m\int_0^s\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>dW^i_z\nonumber\\
&+& \int_0^s\left[||X_z-\overline{Y}^M_u+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2-||X_z-\overline{Y}^M_z||^2\right]d\overline{N}_z\nonumber\\
&+&\lambda\int_0^s\left[||X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z \rfloor})||^2-||X_z-\overline{Y}^M_z||^2\right]dz\nonumber\\
&=&A_1+A'_1+A_2+A_3+A_4+A_5+A_6.
\label{ch5Th1}
\end{aligned}$$ In the next step, we give some useful estimations of $A_1, A'_1, A_2, A_3$ and $A_6$. $$\begin{aligned}
A_1&=&2\int_0^s\left\langle X_z-\overline{Y}^M_z,v(X_z)-\dfrac{v(\overline{Y}_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right\rangle dz\\
&=&2\int_0^s<X_s-\overline{Y}^M_z,v(X_z)-v(\overline{Y}^M_z)>dz\\
&+&2\int_0^s\left\langle X_s-\overline{Y}^M_z,v(\overline{Y}^M_z)-\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right\rangle dz\\
&=& A_{11}+A_{12}.
\end{aligned}$$ Using the one-sided Lipschitz condition satiasfied by $v$ leads to $$\begin{aligned}
A_{11} &=&2\int_0^s\langle X_s-\overline{Y}^M_z,v(X_z)-v(\overline{Y}^M_z)\rangle dz\nonumber\\
&\leq& 2C\int_0^s||X_z-\overline{Y}^M_z||^2dz.
\label{ch5ThA11}
\end{aligned}$$ Moreover, using the inequality $\langle a, b\rangle\leq |a||b|\leq \dfrac{a^2}{2}+\dfrac{b^2}{2}$ leads to : $$\begin{aligned}
A_{12}&=& 2\int_0^s\left\langle X_z-\overline{Y}^M_z,v(\overline{Y}^M_z)-\dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right\rangle dz\nonumber\\
&=&2\int_0^s\left\langle X_z-\overline{Y}^M_z, v(\overline{Y}^M_z)-v(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz\nonumber\\
&+&2\Delta t\int_0^s\left\langle X_z-\overline{Y}^M_z, \dfrac{v(\overline{Y}^M_{\lfloor z\rfloor})||v(\overline{Y}^M_{\lfloor z\rfloor})||}{1+\Delta t||v(\overline{Y}^M_{\lfloor z\rfloor})||}\right\rangle dz\nonumber\\
&\leq &\int_0^s||X_z-\overline{Y}^M_z||^2dz+\int_0^s||v(\overline{Y}^M_z)-v(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&+&\int_0^s||X_z-\overline{Y}^M_z||^2dz+\dfrac{T^2}{M^2}\int_0^s||v(\overline{Y}^M_{\lfloor z\rfloor})||^4dz\nonumber\\
&\leq &2\int_0^s||X_z-\overline{Y}^M_z||^2dz+\int_0^s||v(\overline{Y}^M_z)-v(\overline{Y}_{\lfloor z\rfloor})||^2dz\nonumber\\
&+&\dfrac{T^2}{M^2}\int_0^s||v(\overline{Y}^M_{\lfloor z\rfloor})||^4dz
\label{ch5ThA12}
\end{aligned}$$ Combining and give the following estimation for $A_1$ : $$\begin{aligned}
A_1 &\leq & (2C+2)\int_0^s||X_z-\overline{Y}^M_z||^2dz+\int_0^s||v(\overline{Y}^M_z)-v(\overline{Y}_{\lfloor z\rfloor})||^2dz\nonumber\\
&+&\dfrac{T^2}{M^2}\int_0^s||v(\overline{Y}^M_{\lfloor z\rfloor})||^4dz.
\label{ch5ThA1}
\end{aligned}$$ Using again the inequality $\langle a, b\rangle\leq |a||b|\leq \dfrac{a^2}{2}+\dfrac{b^2}{2}$ and the global-Lipschitz condition satisfied by $u$ leads to : $$\begin{aligned}
A_2 &=&2\int_0^s\left\langle X_z-\overline{Y}^M_z, u(X_z)-u(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz\nonumber\\
&=&2\int_0^s\left\langle X_z-\overline{Y}^M_z, u(X_z)-u(\overline{Y}^M_z)\right\rangle dz+\int_0^s\left\langle X_z-\overline{Y}^M_z, u(\overline{Y}^M_z)-u(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz\nonumber\\
&\leq &2C\int_0^s||X_z-\overline{Y}^M_z||^2dz+2C\int_0^s||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2dz.
\label{ch5ThA2}
\end{aligned}$$ Using the same arguments as for $A_2$ leads to the following estimation of $A'_1$ $$\begin{aligned}
A'_1&=&2\lambda\int_0^s\left\langle X_z-\overline{Y}^M_z, h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})\right\rangle dz \nonumber\\
&\leq& 2\lambda C\int_0^s||X_z-\overline{Y}^M_z||^2dz+2\lambda C\int_0^s||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2dz.
\label{ch5ThA1n}
\end{aligned}$$ Using the inequalities $||g_i(x)-g_i(y)||\leq ||g(x)-g(y)||$ and $(a+b)^2\leq 2a^2+2b^2$ and the global Lipschitz condition satisfyed by $g$, we have $$\begin{aligned}
A_3 &=&\sum_{i=1}^m\int_0^s||g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&\leq &m\int_0^s||g(X_z)-g(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&= &m\int_0^s||g(X_z)-g(\overline{Y}^M_z)+g(\overline{Y}^M_z)-g(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&\leq &2m\int_0^s||g(X_z)-g(\overline{Y}^M_z)||^2dz+2m\int_0^s||g(\overline{Y}^M_z)-g(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&\leq& 2mC^2\int_0^s||X_z-\overline{Y}^M_z||^2dz+2mC^2\int_0^s||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2dz
\label{ch5ThA3}
\end{aligned}$$ Using the same reasons as above we obtain the following estimation for $A_6$ : $$\begin{aligned}
A_6 &=&\lambda\int_0^s\left[X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2-||X_z-\overline{Y}^M_z||^2\right]dz\nonumber\\
&\leq &3\lambda\int_0^s||X_z-\overline{Y}^M_z||^2dz+2\lambda\int_0^s||h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&\leq &3\lambda\int_0^s||X_z-\overline{Y}^M_z||^2dz+4\lambda\int_0^s||h(X_z)-h(\overline{Y}^M_z)||^2dz\nonumber\\
&+ &4\lambda\int_0^s||h(\overline{Y}^M_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2dz\nonumber\\
&\leq &(3\lambda+4\lambda C^2)\int_0^s||X_z-\overline{Y}^M_z||^2dz+4\lambda C^2\int_0^s||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2dz.
\label{ch5ThA6}
\end{aligned}$$ Inserting , , , and in we obtain : $$\begin{aligned}
||X_s-\overline{Y}^M_s||^2&\leq &(4C+2+2mC^2+3\lambda+4\lambda C^2+2\lambda C)\int_0^s||X_z-\overline{Y}^M_z||^2dz\nonumber\\
&+&(2C+2mC^2+4\lambda C^2+2\lambda C)\int_0^s||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2dz\nonumber\\
&+&\int_0^s||v(\overline{Y}^M_z)-v(\overline{Y}^M_{\lfloor z\rfloor})||^2dz+\dfrac{T^2}{M^2}\int_0^s||v(\overline{Y}^M_{\lfloor z\rfloor})||^4dz\nonumber\\
&+&2\sum_{i=1}^m\int_0^s\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>dW^{i}_z\nonumber\\
&+&\int_0^s\left[||X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2-||X_z-\overline{Y}^M_z||^2\right]d\overline{N}_z.
\end{aligned}$$
Taking the supremum in both sides of the previous inequality leads to $$\begin{aligned}
\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2&\leq &(4C+2+2mC^2+3\lambda+4\lambda C^2+2\lambda C)\int_0^t||X_z-\overline{Y}^M_z||^2dz\nonumber\\
&+&(2C+2mC^2+4\lambda C^2+2\lambda C)\int_0^t||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2dz\nonumber\\
&+&\int_0^t||v(\overline{Y}^M_z)-v(\overline{Y}^M_{\lfloor z\rfloor})||^2dz+\dfrac{T^2}{M^2}\int_0^t||v(\overline{Y}^M_{\lfloor z\rfloor})||^4dz\nonumber\\
&+&2\sup_{s\in[0,t]}\left|\sum_{i=1}^m\int_0^s\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>dW^{i}_z\right|\nonumber\\
&+&\sup_{s\in[0,t]}\left|\int_0^s\left[||X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2\right]d\overline{N}_z\right|\nonumber\\
&+&\sup_{s\in[0,t]}\left|\int_0^s||X_z-\overline{Y}^M_z||^2d\overline{N}_z\right|.
\label{ch5Th2}\end{aligned}$$ Using Lemma \[ch4lemma15\] we have the following estimations $$\begin{aligned}
B_1& :=&\left\|2\sup_{s\in[0,t]}\left|\sum_{i=1}^m\int_0^s\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>dW^i_z\right|\right\|_{L^{p/2}(\Omega, \mathbb{R})}\\
&\leq &C_p\left(\int_0^t\sum_{i=1}^m\left\|\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>\right\|^2_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2},
\end{aligned}$$ for all $p\in[2,\infty)$.
Moreover, using Holder inequality, the inequality $ab\leq \dfrac{a^2}{2}+\dfrac{b^2}{2}$ and the global Lipschitz condition satisfied by $g$, we have the following estimation for $B_1$. $$\begin{aligned}
B_1 &\leq& C_p\left(\int_0^t\sum_{i=1}^m\left\|\left<X_z-\overline{Y}^M_z, g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})\right>\right\|^2_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &C_p\left(\int_0^t\sum_{i=1}^m||X_z-\overline{Y}_z^M||^2_{L^p(\Omega, \mathbb{R}^d)}||g_i(X_z)-g_i(\overline{Y}^M_{\lfloor z\rfloor})||^2_{L^p(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &C_p\left(\int_0^t\dfrac{1}{2}||X_z-\overline{Y}_z^M||^2_{L^p(\Omega, \mathbb{R}^d)}2m||g(X_z)-g(\overline{Y}^M_{\lfloor z\rfloor})||^2_{L^p(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &\dfrac{C_p}{\sqrt{2}}\left(\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||_{L^p(\Omega, \mathbb{R}^d)}\right)\left(2C^2m\int_0^t||X_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &\dfrac{1}{4}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}+C_p^2m\int_0^t||X_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}dz\nonumber\\
&\leq &\dfrac{1}{4}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}+2C_p^2m\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega,\mathbb{R}^d)}dz\nonumber\\
&+&2C_p^2m\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}dz.
\label{ch5ThB1}
\end{aligned}$$ Using Lemma \[ch4lemma18\] and the inequality $(a+b)^4\leq 4a^4+4b^4$, it follows that : $$\begin{aligned}
B_2 &=&\left\|\sup_{s\in[0,t]}\left|\int_0^s||X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^2d\overline{N}_z\right|\right\|_{L^{p/2}(\Omega, \mathbb{R}^d)}\nonumber\\
&\leq &C_p\left(\int_0^t||X_z-\overline{Y}^M_z+h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &C_p\left(\int_0^t\left[4||X_z-\overline{Y}^M_z||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}+4||h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}\right]dz\right)^{1/2},
\end{aligned}$$ for all $p\in[2,\infty)$.
Using the inequality $\sqrt{a+b}\leq\sqrt{a}+\sqrt{b}$, it follows that : $$\begin{aligned}
B_2&\leq &2C_p\left(\int_0^t||X_z-\overline{Y}^M_z||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2} +2C_p\left(\int_0^t||h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
& =& B_{21}+B_{22}.
\label{ch5ThB}
\end{aligned}$$ Using Holder inequality, it follows that : $$\begin{aligned}
B_{21} &: = &2 C_p\left(\int_0^t||X_z-\overline{Y}^M_z||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &2C_p\left(\int_0^t||X_z-\overline{Y}^M_z||^2_{L^p(\Omega, \mathbb{R}^d)}||X_z-\overline{Y}^M_z||^2_{L^p(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&= &2C_p\left(\int_0^t\dfrac{1}{16}||X_z-\overline{Y}^M_z||^2_{L^p(\Omega, \mathbb{R}^d)}16||X_z-\overline{Y}^M_z||^2_{L^p(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &\dfrac{1}{4}\sup_{s\in[0,t]}||X_z-\overline{Y}^M_z||_{L^p(\Omega,\mathbb{R}^d)}8C_p\left(\int_0^t||X_z-\overline{Y}^M_z||^2_{L^p(\Omega,\mathbb{R}^d)}dz\right)^{1/2}.
\end{aligned}$$ Using the inequality $2ab\leq a^2+b^2$ leads to : $$\begin{aligned}
B_{21}&\leq &\dfrac{1}{16}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}
+16C_p\int_0^t||X_s-\overline{Y}^M_z||^2_{L^p(\Omega, \mathbb{R}^d)}dz.
\label{ch5ThB21}
\end{aligned}$$ Using the inequalities $(a+b)^4\leq 4a^4+4b^4$ and $\sqrt{a+b}\leq\sqrt{a}+\sqrt{b}$, we have the following bound for $B_{22}$ $$\begin{aligned}
B_{22}&: = &2C_p\left(\int_0^t||h(X_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\nonumber\\
&\leq &2C_p\left(\int_0^t4||h(X_z)-h(\overline{Y}^M_z)||^4_{L^{p/2}(\Omega,\mathbb{R}^d)}+4||h(\overline{Y}^M_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^4_{L^{p/2}(\Omega,\mathbb{R}^d)}dz\right)^{1/2}\\
&\leq &4C_p\left(\int_0^t||h(X_z)-h(\overline{Y}^M_z)||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}\\
&+&4C_p\left(\int_0^t||h(\overline{Y}^M_z)-h(\overline{Y}^M_{\lfloor z\rfloor})||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}.
\end{aligned}$$ Using the global Lipschitz condition satisfied by $h$ leads to : $$\begin{aligned}
B_{22}&\leq &4C_p\left(\int_0^tC^4||X_z-\overline{Y}^M_z||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}+4C_p\left(\int_0^tC^4||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}.
\end{aligned}$$ Using the same idea as for $B_{21}$, it follows that : $$\begin{aligned}
B_{22} &\leq &\dfrac{1}{16}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)} +64C_p\int_0^t||X_z-\overline{Y}^M_z||^2_{L^p(\Omega, \mathbb{R}^d)}dz\nonumber\\
&+&\dfrac{1}{16}\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}+64C_p\int_0^t||\overline{Y}^M_z-\overline{Y}^M_{\lfloor z\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}dz.\nonumber
\end{aligned}$$ Taking the supremum under the integrand of the last term in the above inequality and using the fact that we don’t care about the value of the constant leads to : $$\begin{aligned}
B_{22}&\leq &\dfrac{1}{16}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)} +64C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds\nonumber\\
&+&C_p\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}.
\label{ch5ThB22}
\end{aligned}$$ Inserting and into gives : $$\begin{aligned}
B_2&\leq &\dfrac{1}{8}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}+C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds\nonumber\\
&+&C_p\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}.
\label{ch5ThB2}
\end{aligned}$$ Using again Lemma \[ch4lemma18\] leads to : $$\begin{aligned}
B_3 &: =& \left\|\sup_{s\in[0,t]}\int_0^s||X_z-\overline{Y} ^M_z||^2d\overline{N}_z\right\|_{L^{p/2}(\Omega, \mathbb{R}^d)}\\
&\leq &C_p\left(\int_0^t||X_z-\overline{Y}^M_z||^4_{L^{p/2}(\Omega, \mathbb{R}^d)}dz\right)^{1/2}.
\end{aligned}$$ Using the same argument as for $B_{21}$, we obtain : $$\begin{aligned}
B_3 &\leq &\dfrac{1}{8}\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}+C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds
\label{ch5ThB3}
\end{aligned}$$ Taking the $L^{p}$ norm in both side of and inserting inequalities , and leads to : $$\begin{aligned}
\left\|\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^{p}(\Omega, \mathbb{R}^d)}&=&\left\|\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||^2\right\|_{L^{p/2}(\Omega, \mathbb{R}^d)}\end{aligned}$$ $$\begin{aligned}
&\leq & C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds+C_p\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}ds\\
&+&\int_0^t||v(X_s)-v(\overline{Y}^M_{\lfloor s\rfloor})||^2_{L^{p}(\Omega, \mathbb{R}^d)}ds+C_p\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}\\
&+&\dfrac{T^2}{M^2}\int_0^t||v(\overline{Y}^M_{\lfloor s\rfloor})||^4_{L^{2p}(\Omega, \mathbb{R}^d)}ds +2C_p\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds\\
&+&\dfrac{1}{2}\left\|\sup_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^{p/2}(\Omega, \mathbb{R}^d)},
\end{aligned}$$ for all $t\in[0,T]$ and all $p\in[2,+\infty)$. Where $C_p$ is the generic constant.
The previous inequality can be writen in the following appropriate form
$
\dfrac{1}{2}\left\|\sup\limits_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^{p}(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq & C_p\int_0^t||X_s-\overline{Y}^M_s||^2_{L^p(\Omega, \mathbb{R}^d)}ds+C_p\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega,\mathbb{R}^d)}ds\\
&+&\int_0^t||v(X_s)-v(\overline{Y}^M_{\lfloor s\rfloor})||^2_{L^{p}(\Omega, \mathbb{R}^d)}ds+C_p\sup_{s\in[0,t]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}\\
&+&\dfrac{T^2}{M^2}\int_0^t||v(\overline{Y}^M_{\lfloor s\rfloor})||^4_{L^{2p}(\Omega, \mathbb{R}^d)}ds +2C^2m\int_0^t||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds.
\end{aligned}$$
Applying Gronwall lemma to the previous inequality leads to :
$
\dfrac{1}{2}\left\|\sup\limits_{s\in[0,t]}||X_s-\overline{Y}^M_s||\right\|^2_{L^p(\Omega, \mathbb{R}^d)}
$
$$\begin{aligned}
&\leq &C_pe^{C_p}\left(\int_0^T||v(\overline{Y}^M_s)-v(\overline{Y}^M_{\lfloor s\rfloor})||^2_{L^p(\Omega, \mathbb{R}^d)}ds+C_p\sup_{u\in[0,t]}||\overline{Y}^M_u-\overline{Y}^M_{\lfloor u\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}\right.\\
& &\left. +\dfrac{T^2}{M^2}\int_0^T||v(\overline{Y}^M_{\lfloor s\rfloor})||^4_{L^{2p}(\Omega, \mathbb{R}^d)}ds +C_p\int_0^T||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||^2_{L^p(\Omega, \mathbb{R}^d)}ds\right).
\end{aligned}$$
Using the inequality $\sqrt{a+b+c}\leq \sqrt{a}+\sqrt{b}+\sqrt{c}$, it follows that
$
\left\|\sup\limits_{t\in[0, T]}||X_t-\overline{Y}^M_t||\right\|_{L^p(\Omega, \mathbb{R}^d)}
$ $$\begin{aligned}
&\leq &C_pe^{C_p}\left(\sup_{s\in[0,T]}||v(\overline{Y}^M_s)-v(\overline{Y}^M_{\lfloor s\rfloor})||_{L^p(\Omega, \mathbb{R}^d)}+C_p\sup_{s\in[0,T]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||_{L^p(\Omega, \mathbb{R}^d)} \right.\nonumber\\
&+& \left. \dfrac{T^2}{M}\left[\sup_{n\in\{0,\cdots, M\}}||v(Y^M_n)||^2_{L^{2p}(\Omega, \mathbb{R}^d)}\right]+C_p\sup_{s\in[0,T]}||\overline{Y}^M_s-\overline{Y}^M_{\lfloor s\rfloor}||_{L^p(\Omega, \mathbb{R}^d)}\right),
\label{ch5Gronwall}
\end{aligned}$$ for all $p\geq 2$. Using Lemma \[ch5lemma14\] and Lemma \[ch5lemma16\], it follows that : $$\begin{aligned}
\mathbb{E}\left[\sup_{t\in[0,T]}\left\|X_t-\overline{Y}^M_t\right\|^p\right]^{1/p}\leq C_p(\Delta t)^{1/2},
\end{aligned}$$ for all $M\in\mathbb{N}$ and all $p\in[2,\infty)$. The application of Holder inequality shows that the latter inequality is satisfied for all $p\in[1,\infty)$, and this complete the proof of Theorem \[ch5theorem2\].
Strong convergence of the tamed Euler Scheme
--------------------------------------------
\[ch5theorem2b\] Under Assumptions \[ch5assumption1\], for all $p\in[1, +\infty)$ there exist a constant $C_p>0$ such that $$\begin{aligned}
\left(\mathbb{E}\left[\sup_{t\in[0,T]}\left\|X_t-\overline{X}^M_t\right\|^p\right]\right)^{1/p}\leq C_p\Delta t^{1/2},
\end{aligned}$$ for all $M\in\mathbb{N}$. Where $X : [0, T]\times\Omega \longrightarrow \mathbb{R}^d$ is the exact solution of and $\overline{X}^M_t$ the continous interpolation of the numerical solution defined by : $$\begin{aligned}
\overline{X}^M_t &=&X^M_n+\dfrac{(t-n\Delta t)f(X^M_n)}{1+||f(X^M_n)||}+g(X^M_n)(W_t-W_{n\Delta t})+h(X^M_n)(\overline{N}_t-\overline{N}_{n\Delta t}),
\label{ch5continoussol1}
\end{aligned}$$ for all $M\in\mathbb{N}$, all $n\in\{0,\cdots, M\}$ and all $t\in[n\Delta t, (n+1)\Delta t)$.
Using the relation $\Delta\overline{N}^M_n=\Delta N^M_n-\lambda\Delta t$, the continous interpolation of can be express in the following form $$\begin{aligned}
\overline{X}^M_t =X^M_n+\lambda(t-n\Delta t)h(X^M_n)+\dfrac{(t-n\Delta t)f(X^M_n)}{1+||f(X^M_n)||}+g(X^M_n)(W_t-W_{n\Delta t})+h(X^M_n)(N_t-N_{n\Delta t}),
\end{aligned}$$ for all $t\in[n\Delta t, (n+1)\Delta t[$.
From the numerical solution and using the relation $\Delta N^M_n=\Delta\overline{N}^M_n+\lambda\Delta t$, it follows that: $$\begin{aligned}
X^M_{n+1}&=&X^M_n+\dfrac{\Delta tf(X^M_n)}{1+\Delta t||f(X^M_n)||}+g(X^M_n)\Delta W^M_n+h(X^M_n)\Delta N^M_n\nonumber\\
&=&X^M_n+\lambda h(X^M_n)T/M+\dfrac{\Delta tf(X^M_n)}{1+\Delta t||f(X^M_n)||}+g(X^M_n)\Delta W^M_n\nonumber\\
&+&h(X^M_n)\Delta \overline{N}^M_n.
\label{ch5contam}
\end{aligned}$$ The functions $\lambda h$ and $f$ satisfy the same conditions as $u_{\lambda}$ and $v$ respectively. So from it follows that the numerical solution satisfied the same hypothesis as the numerical solution . Hence, it follows from Theorem \[ch5theorem2\] that there exist a constant $C_p>0$ such that $$\begin{aligned}
\left(\mathbb{E}\left[\sup_{t\in[0,T]}\left\|X_t-\overline{X}^M_t\right\|^p\right]\right)^{1/p}\leq C_p\Delta t^{1/2},
\end{aligned}$$ for all $p\in[1,\infty)$.
Linear mean-square stability
----------------------------
The goal of this section is to find a stepsize for which the tamed euler scheme and the semi-tamed Euler scheme are stable. The first approach to the stability analysis of a numerical method is to study the stability behavior of the method for a scalar linear equation. So we will focus in the linear case. Let’s consider a linear test equation with real and scalar coefficients. $$\begin{aligned}
dX(t)= aX(t^{-})dt +bX(t^{-})dW(t)+cX(t^{-})dN(t), \hspace{0.5cm}
X(0)=X_0.
\label{ch5linearequation}\end{aligned}$$ It is proved in [@Desmond2] that the exact solution of is mean-square stable if and only if
$l :=2a+b^2+\lambda c(2+c)<0$. That is : $$\begin{aligned}
\label{ch5stabilitycondition}
\lim_{t\longrightarrow \infty}|X(t)|^2=0 \Longleftrightarrow l := 2a+b^2+\lambda c(2+c) <0.\end{aligned}$$
In this section, to easy notations, the tamed Euler appoximation $X^M$ will be replaced by $X$ and the semi-tamed Euler approximation $Y^M$ will be replaced by $Y$.
We have the following result for the numerical method .
Under Assumption \[ch5stabilitycondition\], the semi-tamed Euler is mean-square stable if and only if $$\begin{aligned}
\Delta t<\dfrac{-l}{(a+\lambda c)^2}.
\end{aligned}$$
Aplying the semi-tamed Euler scheme to leads to $$\begin{aligned}
Y_{n+1}=Y_n+aY_n\Delta t+\lambda cY_n\Delta t+bY_n\Delta W_n+cY_n\overline{N}_n.
\label{ch5compen1}
\end{aligned}$$ Squaring both sides of leads to $$\begin{aligned}
Y_{n+1}^2&=&Y_n^2+(a+\lambda c)^2\Delta t^2Y_n^2+b^2Y_n^2\Delta W_n^2+c^2Y_n^2\Delta\overline{N}_n^2+2(a+\lambda c)Y_n^2\Delta t+2bY_n^2\Delta W_n\nonumber\\
&+&2cY_n^2\Delta\overline{N}_n+2b(a+\lambda c)Y_n^2\Delta t\Delta W_n+2c(a+\lambda c)Y_n^2\Delta t\Delta\overline{N}_n+2bcY_n^2\Delta W_n\Delta\overline{N}_n.
\label{ch5compen2}
\end{aligned}$$ Taking expectation in both sides of and using the relations $\mathbb{E}(\Delta W_n^2)=\Delta t$, $\mathbb{E}(\Delta\overline{N}_n^2)=\lambda \Delta t$ and $\mathbb{E}(\Delta W_n)=\mathbb{E}(\Delta\overline{N}_n)=0$ leads to $$\begin{aligned}
\mathbb{E}|Y_{n+1}|^2=(1+(a+\lambda c)^2\Delta t^2+(b^2+\lambda c^2+2a+2\lambda c)\Delta t)\mathbb{E}|Y_n|^2.
\end{aligned}$$ So the numerical method is stable if and only if $$\begin{aligned}
1+(a+\lambda c)^2\Delta t^2+(b^2+\lambda c^2+2a+2\lambda c)\Delta t<1.
\end{aligned}$$ That is if and only if $$\begin{aligned}
\Delta t<\dfrac{-l}{(a+\lambda c)^2}.
\end{aligned}$$
\[thmncts\] Under Assumption \[ch5stabilitycondition\], the tamed Euler scheme is mean-square stable if one of the following conditions is satisfied :
- $a(1+\lambda c\Delta t)\leq 0$, $2a-l>0$ and $\Delta t<\dfrac{2a-l}{a^2+\lambda^2c^2}$.
- $a(1+\lambda c\Delta t)>0$ and $\Delta t<\dfrac{-l}{(a+\lambda c)^2}$.
Applying the tamed Euler scheme to equation leads to : $$\begin{aligned}
X_{n+1}=X_n+\dfrac{aX_n\Delta t}{1+\Delta t|aX_n|}+bX_n\Delta W_n+cX_n\Delta N_n.
\label{ch5eq1}\end{aligned}$$ By squaring both sides of leads to : $$\begin{aligned}
X_{n+1}^2&=&X_n^2+\dfrac{a^2X^2_n\Delta t^2}{(1+\Delta t|aX_n|)^2}+b^2X_n^2\Delta W_n^2+c^2X_n^2\Delta N_n^2+\dfrac{2aX_n^2\Delta t}{1+\Delta t|aX_n|}+2bX_n^2\Delta W_n\\
&+&2cX_n^2\Delta N_n+\dfrac{2abX_n^2}{1+\Delta t|aX_n|}\Delta W_n+\dfrac{2acX_n^2\Delta t}{1+\Delta t|aX_n|}\Delta N_n+2bcX_n^2\Delta W_n\Delta N_n.\end{aligned}$$ Using the inequality $ \dfrac{a^2\Delta t^2}{1+\Delta t|aX_n|}\leq a^2\Delta t^2$, the previous equality becomes $$\begin{aligned}
X_{n+1}^2&\leq &X_n^2+a^2X2_n\Delta t^2+b^2X_n^2\Delta W_n^2+c^2X_n^2\Delta N_n^2+\dfrac{2aX_n^2\Delta t}{1+\Delta t|aX_n|}+2bX_n^2\Delta W_n\\
&+&2cX_n^2\Delta N_n+\dfrac{2abX_n^2}{1+\Delta t|aX_n|}\Delta W_n+\dfrac{2acX_n^2\Delta t}{1+\Delta t|aX_n|}\Delta N_n+2bcX_n^2\Delta W_n\Delta N_n.\end{aligned}$$ Taking expectation in both sides of the previous equality and using independence and the fact that $\mathbb{E}(\Delta W_n)=0$, $\mathbb{E}(\Delta W_n^2)=\Delta t$, $\mathbb{E}(\Delta N_n)=\lambda\Delta t$, $\mathbb{E}(\Delta N_n^2)=\lambda \Delta t+\lambda^2\Delta t^2$ leads to : $$\begin{aligned}
\mathbb{E}|X_{n+1}|^2&\leq& \left[1+a^2\Delta t^2+b^2\Delta t+\lambda^2c^2\Delta t^2+(2+ c)\lambda c\Delta t\right]\mathbb{E}|X_n|^2\nonumber\\
&+&\mathbb{E}\left(\dfrac{2aX^2_n\Delta t(1+\lambda c\Delta t)}{1+\Delta t|aX_n|}\right).
\label{ch5eq2}\end{aligned}$$
- If $a(1+\lambda c\Delta t)\leq 0$, it follows from that $$\begin{aligned}
\mathbb{E}|X_{n+1}|^2\leq \{1+(a^2+\lambda^2c^2)\Delta t^2+[b^2+\lambda c(2+c)]\Delta t\}\mathbb{E}|X_n|^2.\end{aligned}$$ Therefore, the numerical solution is stable if $$\begin{aligned}
1+(a^2+\lambda^2c^2)\Delta t^2+[b^2+\lambda c(2+c)]\Delta t<1.\end{aligned}$$ That is $\Delta t<\dfrac{2a-l}{a^2+\lambda^2c^2}$.
- If $a(1+\lambda c\Delta t)> 0$, using the fact that $\dfrac{2aX_n^2\Delta t(1+\lambda c\Delta t)}{1+\Delta t|aX_n|}< 2aY_n^2\Delta t(1+\lambda c\Delta t)$, inequality becomes $$\begin{aligned}
\mathbb{E}|X_{n+1}|^2\leq\left[1+a^2\Delta t^2+b^2\Delta t+\lambda^2c^2\Delta t^2+2\lambda ac\Delta t^2+(2+ c)\lambda c\Delta t+2a\Delta t\right]\mathbb{E}|X_n|^2.
\label{ch5eq3}\end{aligned}$$ Therefore, it follows from that the numerical solution is stable if
$1+a^2\Delta t^2+b^2\Delta t+\lambda^2c^2\Delta t^2+2\lambda ac\Delta t^2+(2+ c)\lambda c\Delta t+2a\Delta t<1$. That is $\Delta t<\dfrac{-l}{(a+\lambda c)^2}$.
In Theorem \[thmncts\], we can easily check that if $l<0$, we have: $$\begin{aligned}
\left\lbrace \begin{array}{l}
a(1+\lambda c\Delta t)\leq0,\\
2a-l>0 \\
\Delta t<\dfrac{2a-l}{a^2+\lambda^2c^2}\\
\end{array} \right.
\Leftrightarrow
\left\lbrace \begin{array}{l}
a \in (l/2,0], c\geq 0,\\
\Delta t <\dfrac{2a-l}{a^2+\lambda^2 c^2}\\
\end{array} \right.
\bigcup
\left\lbrace \begin{array}{l}
a \in (l/2,0), c<0,\\
\Delta t <\dfrac{2a-l}{a^2+\lambda^2 c^2} \\
\Delta t\leq \dfrac{-1}{ \lambda c} \\
\end{array} \right.\\
\bigcup
\left\lbrace \begin{array}{l}
a>0, c<0 \\
\Delta t <\dfrac{2a-l}{a^2+\lambda^2 c^2} \\
\Delta t \geq \dfrac{-1}{ \lambda c}
\end{array} \right. \end{aligned}$$ $$\begin{aligned}
\left\lbrace \begin{array}{l}
a(1+\lambda c\Delta t)>0,\\
\Delta t<\dfrac{-l}{(a+\lambda c)^2}\\
\end{array} \right.
\Leftrightarrow
\left\lbrace \begin{array}{l}
a >0, c>0,\\
\Delta t < \dfrac{-l}{(a+\lambda c)^2}\\
\end{array} \right.
\bigcup
\left\lbrace \begin{array}{l}
a >0, c<0,\\
\Delta t <\dfrac{-l}{(a+\lambda c)^2} \wedge \dfrac{-1}{ \lambda c} \\
\end{array} \right.\\
\bigcup
\left\lbrace \begin{array}{l}
a<0, c<0 \\
\Delta t <\dfrac{-l}{(a+\lambda c)^2} \\
\Delta t > \dfrac{-1}{ \lambda c}
\end{array} \right. \end{aligned}$$
Nonlinear mean-square stability
-------------------------------
In this section, we focus on the mean-square stability of the approximation . It is proved in [@Desmond2] that under the following conditions, $$\begin{aligned}
\langle x-y, f(x)-f(y)\rangle&\leq& \mu||x-y||^2,\\
||g(x)-g(y)||^2&\leq& \sigma||x-y||^2,\\
||h(x)-h(y)||^2 &\leq &\gamma ||x-y||^2,\end{aligned}$$ where $\mu$, $\sigma$ and $\gamma$ are constants, the eaxct solution of SDE is mean-square stable if $$\begin{aligned}
\alpha :=2\mu+\sigma+\lambda\sqrt{\gamma}(\sqrt{\gamma}+2)<0.
\end{aligned}$$
In this section, to easy notations, the tamed Euler appoximation $X^M$ will be replaced by $X$ and the semi-tamed Euler approximation $Y^M$ will be replaced by $Y$.
Following the literature of [@Xia2], in order to examine the mean-square stability of the numerical solution given by , we assume that $f(0)=g(0)=h(0)=0$. Also we make the following assumptions.
\[ch5assumption2\] There exist a positive constants $\rho$, $\beta$, $\theta$, $K$, $C$ and $a>1$ such that $$\begin{aligned}
\langle x-y, u(x)-u(y)\rangle\leq -\rho||x-y||^2,\hspace{1cm} ||u(x)-u(y)||\leq K||x-y||,\\
\langle x-y, v(x)-v(y)\rangle\leq-\beta||x-y||^{a+1}, \hspace{1.5cm} ||v(x)||\leq \overline{\beta}||x||^a,\\
||g(x)-g(y)||\leq\theta||x-y||,\hspace{2cm} ||h(x)-h(y)||\leq C||x-y|.
\end{aligned}$$
We define $\alpha_1= -2\rho+\theta^2+\lambda C(2+C)$.
Under Assumptions \[ch5assumption2\] and the further hypothesis $2\beta-\overline{\beta}>0$, for any stepsize $\Delta t<\dfrac{-\alpha_1}{(K+\lambda C)^2}\wedge\dfrac{2\beta}{[2(K+\lambda C)+\overline{\beta}]\overline{\beta}}\wedge\dfrac{2\beta-\overline{\beta}}{2(K+\lambda C)\overline{\beta}}$, the numerical solution is exponentiallly mean-square stable.
The numerical solution is given by $$\begin{aligned}
Y_{n+1}=Y_n+\Delta tu_{\lambda}(Y_n)+\dfrac{\Delta tv(Y_n)}{1+\Delta t||v(Y_n)||}+g(Y_n)\Delta W_n+h(Y_n)\Delta\overline{N}_n,
\end{aligned}$$ where $u_{\lambda}=u+\lambda h$.
Taking the inner product in both sides of the previous equation leads to $$\begin{aligned}
||Y_{n+1}||^2&=&||Y_n||^2+\Delta t^2||u_{\lambda}(Y_n)||^2+\dfrac{\Delta t^2||v(Y_n)||^2}{\left(1+\Delta t||v(Y_n)||\right)^2}+||g(Y_n)||^2||\Delta W_n||^2+||h(Y_n)||^2|\Delta\overline{N}_n|^2\nonumber\\
&+&2\Delta\langle Y_n, u_{\lambda}(Y_n)\rangle+2\Delta t\left\langle Y_n, \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}\right\rangle+2\langle Y_n, g(Y_n)\Delta W_n\rangle+2\langle Y_n, h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&+2\Delta t^2\left\langle u_{\lambda}(Y_n), \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}\right\rangle+2\Delta t\langle u_{\lambda}(Y_n), g(Y_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y_n), h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, g(Y_n)\Delta W_n\right\rangle+2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, h(Y_n)\Delta\overline{N}_n\right\rangle\nonumber\\
&+&2\langle g(Y_n)\Delta W_n, h(Y_n)\Delta\overline{N}_n\rangle.
\label{ch5meansemi1}
\end{aligned}$$ Using Assumptions \[ch5assumption2\], it follows that : $$\begin{aligned}
2\Delta t\left\langle Y_n, \dfrac{v(Y_n)}{1+\Delta ||v(Y_n)||}\right\rangle\leq\dfrac{-2\beta\Delta t||Y_n||^{a+1}}{1+\Delta t||v(Y_n)||}
\label{ch5meansemi2}
\end{aligned}$$ $$\begin{aligned}
2\Delta t^2\left<u_{\lambda}(Y_n), \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}\right>&\leq &\dfrac{2\Delta t^2||u_{\lambda}(Y_n)||||v(Y_n)||}{1+\Delta t||v(Y_n||}\nonumber\\
&\leq&\dfrac{2\Delta t^2(K+\lambda C)\overline{\beta}||Y_n|^{a+1}}{1+\Delta t||v(Y_n)||}.
\label{ch5meansemi3}
\end{aligned}$$ Let’s define $\Omega_n=\{\omega\in\Omega : ||Y_n||>1\}$.
- On $\Omega_n$ we have $$\begin{aligned}
\dfrac{\Delta t^2||v(Y_n)||^2}{\left(1+\Delta t||v(Y_n)||\right)^2}\leq\dfrac{\Delta t||v(Y_n)||}{1+\Delta t||v(Y_n)||}\leq\dfrac{\overline{\beta}\Delta t||Y_n||^{a+1}}{1+\Delta t||v(Y_n)||}.
\label{ch5meansemi4}
\end{aligned}$$ Therefore using , and , equality becomes :
yields $$\begin{aligned}
\|Y_{n+1}\|^2& \leq &\|Y_n\|^2+\Delta t^2 \|u_{\lambda}(Y_n)\|^2+ \|g(Y_n)\|^2\|\Delta W_n\|^2+\|h(Y_n)\|^2|\Delta\overline{N}_n|^2\nonumber\\
&+&2\Delta t\langle Y_n, u_{\lambda}(Y_n)\rangle+2\langle Y_n, g(Y_n)\Delta W_n
+2\langle Y_n, h(Y_n)\Delta\overline{N}_n\rangle \nonumber\\
&+&2\Delta t\langle u_{\lambda}(Y_n), g(Y_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y_n), h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t\|v(Y_n)\|}, g(Y_n)\Delta W_n\right\rangle+2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t\|v(Y_n)\|}, h(Y_n)\Delta\overline{N}_n\right\rangle\nonumber\\
&+&2\langle g(Y_n)\Delta W_n, h(Y_n)\Delta\overline{N}_n\rangle+\dfrac{\left[-2\beta\Delta t+2(K+\lambda c)\overline{\beta}\Delta t^2+\overline{\beta}\Delta t\right]\|Y_n\|^{a+1}}{1+\Delta t\|v(Y_n)\|}.
\label{ch5meansemi4a}
\end{aligned}$$ The hypothesis $\Delta t<\dfrac{2\beta-\overline{\beta}}{2(K+\lambda C)\overline{\beta}}$ implies that $-2\beta\Delta t+2(K+\lambda C)\overline{\beta}\Delta t^2+\overline{\beta}\Delta t<0$. Therefore, becomes
$$\begin{aligned}
||Y_{n+1}||^2&\leq &||Y_n||^2+\Delta t^2||u_{\lambda}(Y_n)||^2+2\Delta t<Y_n, u_{\lambda}(Y_n)>+||g(Y_n)||^2||\Delta W_n||^2\nonumber\\
&+&||h(Y_n)||^2||\Delta\overline{N}_n||^2+2\langle Y_n,g(Y_n)\Delta W_n\rangle+2\langle Y_n, h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\langle u_{\lambda}(Y_n), g(Y_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y_n), h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\left\langle \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, g(Y_n)\Delta W_n\right\rangle+2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, h(Y_n)\Delta\overline{N}_n\right\rangle\nonumber\\
&+&2\langle g(Y_n)\Delta W_n, h(Y_n)\Delta\overline{N}_n\rangle.
\label{ch5meansemi4b}
\end{aligned}$$
- On $\Omega_n^c$ we have $$\begin{aligned}
\dfrac{\Delta t^2||v(Y_n)||^2}{\left(1+\Delta t||v(Y_n)||\right)^2}\leq\dfrac{\Delta t^2||v(Y_n)||^2}{1+\Delta t||v(Y_n)||}\leq\dfrac{\overline{\beta}^2\Delta t^2||Y_n||^{2a}}{1+\Delta t||v(Y_n)||}\leq\dfrac{\overline{\beta}^2\Delta t^2||Y_n||^{a+1}}{1+\Delta t||v(Y_n)||}.
\label{ch5meansemi5}
\end{aligned}$$ Therefore, using , and , equality becomes
$$\begin{aligned}
||Y_{n+1}||^2&\leq &||Y_n||^2+\Delta t^2||u_{\lambda}(Y_n)||^2+2\Delta t<Y_n, u_{\lambda}(Y_n)>+||g(Y_n)||^2||\Delta W_n||^2\nonumber\\
&+&||h(Y_n)||^2||\Delta\overline{N}_n||^2+2\langle Y_n,g(Y_n)\Delta W_n\rangle+2\langle Y_n, h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\langle u_{\lambda}(Y_n), g(Y_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y_n), h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\left\langle \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, g(Y_n)\Delta W_n\right\rangle+2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, h(Y_n)\Delta\overline{N}_n\right\rangle\nonumber\\
&+&2\langle g(Y_n)\Delta W_n, h(Y_n)\Delta\overline{N}_n\rangle+\dfrac{\left[-2\beta\Delta t+2(K+\lambda c)\overline{\beta}\Delta t^2+\overline{\beta}^2\Delta t^2\right]||Y_n||^{a+1}}{1+\Delta t t||v(Y_n)||}.
\label{ch5meansemi5a}
\end{aligned}$$
The hypothesis $\Delta t<\dfrac{2\beta}{[2(K+\lambda C)+\overline{\beta}]\overline{\beta}}$ implies that $-2\beta\Delta t+2(K+\lambda C)\overline{\beta}\Delta t^2+\overline{\beta}^2\Delta t^2<0$. Therefore, becomes
$$\begin{aligned}
||Y_{n+1}||^2&\leq &||Y_n||^2+\Delta t^2||u_{\lambda}(Y_n)||^2+2\Delta t<Y_n, u_{\lambda}(Y_n)>+||g(Y_n)||^2||\Delta W_n||^2\nonumber\\
&+&||h(Y_n)||^2||\Delta\overline{N}_n||^2+2\langle Y_n,g(Y_n)\Delta W_n\rangle+2\langle Y_n, h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\langle u_{\lambda}(Y_n), g(Y_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y_n), h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\left\langle \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, g(Y_n)\Delta W_n\right\rangle+2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, h(Y_n)\Delta\overline{N}_n\right\rangle\nonumber\\
&+&2\langle g(Y_n)\Delta W_n, h(Y_n)\Delta\overline{N}_n\rangle.
\label{ch5meansemi5b}
\end{aligned}$$
Finally, from the discussion above on $\Omega_n$ and $\Omega_n^c$, it follows that on $\Omega$, the following inequality holds for all $\Delta t<\dfrac{2\beta}{[2(K+\lambda C)+\overline{\beta}]\overline{\beta}}\wedge\dfrac{2\beta-\overline{\beta}}{2(K+\lambda C)\overline{\beta}}$ $$\begin{aligned}
||Y_{n+1}||^2&\leq &||Y_n||^2+\Delta t^2||u_{\lambda}(Y_n)||^2+2\Delta t<Y_n, u_{\lambda}(Y_n)>+||g(Y_n)||^2||\Delta W_n||^2\nonumber\\
&+&||h(Y_n)||^2||\Delta\overline{N}_n||^2+2\langle Y_n,g(Y_n)\Delta W_n\rangle+2\langle Y_n, h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\langle u_{\lambda}(Y_n), g(Y_n)\Delta W_n\rangle+2\Delta t\langle u_{\lambda}(Y_n), h(Y_n)\Delta\overline{N}_n\rangle\nonumber\\
&+&2\Delta t\left\langle \dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, g(Y_n)\Delta W_n\right\rangle+2\Delta t\left\langle\dfrac{v(Y_n)}{1+\Delta t||v(Y_n)||}, h(Y_n)\Delta\overline{N}_n\right\rangle\nonumber\\
&+&2\langle g(Y_n)\Delta W_n, h(Y_n)\Delta\overline{N}_n\rangle.
\label{ch5meansemi6}
\end{aligned}$$ Taking expectation in both sides of and using the martingale properties of $\Delta W_n$ and $\Delta\overline{N}_n$ leads to : $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2&\leq&\mathbb{E}||Y_n||^2+\Delta t^2\mathbb{E}||u_{\lambda}(Y_n)||^2+2\Delta t\mathbb{E}\langle Y_n, u_{\lambda}(Y_n)\rangle+\Delta t\mathbb{E}||g(Y_n)||^2\nonumber\\
&+&\lambda\Delta t\mathbb{E}||h(Y_n)||^2.
\label{ch5meansemi7}
\end{aligned}$$ From Assumptions \[ch5assumption2\], we have $$\begin{aligned}
||u_{\lambda}(Y_n)||^2\leq (K+\lambda C)^2||Y_n||^2 \hspace{0.5cm} \text{and} \hspace{0.5cm}\langle Y_n, u_{\lambda}(Y_n)\rangle\leq (-\rho+\lambda C)||Y_n||^2.
\end{aligned}$$ So inequality gives $$\begin{aligned}
\mathbb{E}||Y_{n+1}||^2&\leq&\mathbb{E}||Y_n||^2+(K+\lambda C)^2\Delta t^2\mathbb{E}||Y_n||^2+2(-\rho+\lambda C)\Delta t\mathbb{E}||Y_n||^2+\theta^2\Delta t\mathbb{E}||Y_n||^2\\
&+&\lambda C^2\Delta t\mathbb{E}||Y_n||^2\\
&=&\left[1-2\rho\Delta t+(K+\lambda C)^2\Delta t^2+2\lambda C\Delta t+\theta^2\Delta t+\lambda C^2\Delta t\right]\mathbb{E}||Y_n||^2.
\end{aligned}$$ Iterating the previous inequality leads to $$\begin{aligned}
\mathbb{E}||Y_n||^2\leq\left[1-2\rho\Delta t+(K+\lambda C)^2\Delta t^2+2\lambda C\Delta t+\theta^2\Delta t+\lambda C^2\Delta t\right]^n\mathbb{E}||Y_0||^2.
\end{aligned}$$ In oder to have stability, we impose : $$\begin{aligned}
1-2\rho\Delta t+(K+\lambda C)^2\Delta t^2+2\lambda C\Delta t+\theta^2\Delta t+\lambda C^2\Delta t<1.
\end{aligned}$$ That is $$\begin{aligned}
\Delta t<\dfrac{-[-2\rho+\theta^2+\lambda C(2+C)]}{(K+\lambda C)^2} =\dfrac{-\alpha_1}{(K+\lambda C)^2}.
\end{aligned}$$
In the following , we analyse the mean-square stability of the tamed Euler. We make the following assumptions which are essentially consequences of Assumptions \[ch5assumption2\].
\[ch5assumption3\] we assume that there exists positive constants $\beta$, $\overline{\beta}$, $\theta$, $\mu$, $K$, $C$, $\rho$, and $a>1$ such that : $$\begin{aligned}
\langle x-y, f(x)-f(y)\rangle\leq &-\rho ||x-y||^2-\beta||x-y||^{a+1},\nonumber\\
||f(x)|| \leq \overline{\beta}||x||^a+K||x||,\nonumber\\
||g(x)-g(y)|| \leq &\theta||x-y||,\nonumber\\
||h(x)-h(y)|| \leq &C||x-y||,\nonumber\\
\langle x-y, h(x)-h(y)\rangle \leq &-\mu ||x-y||^2.
\label{assumparticular}
\end{aligned}$$
Apart from , Assumption \[ch5assumption3\] is a consequence of Assumption \[ch5assumption2\].
Under Assumptions \[ch5assumption3\] and the further hypothesis $\beta-C\overline{\beta}>0$, $\overline{\beta}(1+2C)-2\beta<0$, $K+\theta^2+\lambda C^2-2\mu\lambda+2\lambda CK<0$, the numerical solution is mean-square stable for any stepsize $$\begin{aligned}
\Delta t<\dfrac{-[K+\theta^2+\lambda C^2-2\mu\lambda+2\lambda CK]}{2K^2+\lambda^2C^2}\wedge\dfrac{\beta-C\overline{\beta}}{\overline{\beta}^2}.
\end{aligned}$$
From equation , we have $$\begin{aligned}
||X_{n+1}||^2&=&||X_n||^2+\dfrac{\Delta t^2||f(X_n)||^2}{\left(1+\Delta t||f(X_n)||\right)^2}+||g(X_n)\Delta W_n||^2+||h(X_n)\Delta N_n||^2\nonumber\\
&+&\left\langle X_n,\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}\right\rangle+2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, g(X_n)\Delta W_n\right\rangle\nonumber\\
&+&2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, h(X_n)\Delta N_n\right\rangle+2\langle g(X_n)\Delta W_n, h(X_n)\Delta N_n\rangle.
\label{ch5meantamed1}
\end{aligned}$$ Using assumptions \[ch5assumption3\], it follows that : $$\begin{aligned}
2\left\langle X_n, \dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}\right\rangle &\leq &\dfrac{-2\Delta t\rho||X_n||^2}{1+\Delta t||f(X_n)||}-\dfrac{2\beta \Delta t||X_n||^{a+1}}{1+\Delta t||f(X_n)||}
\leq -\dfrac{2\beta \Delta t||X_n||^{a+1}}{1+\Delta t||f(X_n)||}.
\label{ch5meantamed2}
\end{aligned}$$
$$\begin{aligned}
||g(X_n)\Delta W_n||^2 \leq \theta^2 ||X_n||^2||\Delta W_n||^2\hspace{0.5cm}\text{and}\hspace{0.5cm}
||h(X_n)\Delta N_n||^2 \leq C^2||X_n||^2|\Delta N_n|^2.\label{ch5meantamed2a}
\end{aligned}$$
$$\begin{aligned}
2\langle X_n, h(X_n)\Delta N_n\rangle =2\langle X_n, h(X_n)\rangle\Delta N_n \leq -2\mu||X_n||^2|\Delta N_n|.
\label{ch5meantamed3}
\end{aligned}$$
$$\begin{aligned}
2\left\langle\dfrac{\Delta tf(X_n)}{1+h||f(X_n)||}, h(X_n)\Delta N_n\right\rangle &\leq &\dfrac{2\Delta t||f(X_n)||||h(X_n)|||\Delta N_n|}{1+\Delta t||f(X_n)||}\nonumber\\
&\leq &\dfrac{2\Delta tC\overline{\beta}||X_n||^{a+1}}{1+\Delta t||f(X_n)||}|\Delta N_n|+2CK||X_n||^2|\Delta N_n|.
\label{ch5meantamed4}
\end{aligned}$$
So from Assumptions \[ch5assumption3\], we have $$\begin{aligned}
\label{ch5assmeantamed} \left\{
\begin{array}{lllll}
\left\langle X_n, \dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}\right\rangle\leq -\dfrac{2\beta \Delta t||X_n||^{a+1}}{1+\Delta t||f(X_n)||}\\
||g(X_n)\Delta W_n||^2 \leq \theta^2 ||X_n||^2||\Delta W_n||^2\\
||h(X_n)\Delta N_n||^2 \leq C^2||X_n||^2|\Delta N_n|^2\\
2\langle Y_n, h(X_n)\Delta N_n\rangle \leq -2\mu||X_n||^2|\Delta N_n|\\
2\left\langle\dfrac{\Delta tf(X_n)}{1+h||f(X_n)||}, h(X_n)\Delta N_n\right\rangle \leq\dfrac{2\Delta tC\overline{\beta}||X_n||^{a+1}}{1+\Delta t||f(X_n)||}|\Delta N_n|+2CK||X_n||^2|\Delta N_n|
\end{array}
\right.
\end{aligned}$$
Let’s define $\Omega_n :=\{w\in\Omega : ||X_n(\omega)||>1\}$.
- On $\Omega_n$ we have : $$\begin{aligned}
\dfrac{\Delta t^2||f(X_n)||^2}{\left(1+\Delta t||f(X_n)||\right)^2}&\leq& \dfrac{\Delta t||f(X_n)||}{1+\Delta t||f(X_n)||}
\leq \dfrac{\Delta t\overline{\beta}||X_n||^a}{1+\Delta t||f(X_n)||}+K\Delta t||X_n||\nonumber\\
&\leq& \dfrac{\Delta t\overline{\beta}||X_n||^{a+1}}{1+\Delta t||f(X_n)||}+K\Delta t||X_n||^2.
\label{ch5meantamed5}
\end{aligned}$$ Therefore using and , equality becomes $$\begin{aligned}
||X_{n+1}||^2&\leq&||X_n||^2+K\Delta t||X_n||^2+\theta^2||X_n||^2||\Delta W_n||^2+C^2||X_n||^2|\Delta N_n|^2\nonumber\\
&+&2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, g(X_n)\Delta W_n\right\rangle-2\mu||X_n||^2|\Delta N_n|+2CK|\Delta N_n|\nonumber\\
&+&2\langle g(X_n)\Delta W_n, h(X_n)\Delta N_n\rangle+\dfrac{\left[-2\beta\Delta t+\overline{\beta}\Delta t+2\overline{\beta}C\Delta t\right]||X_n||^{a+1}}{1+\Delta t||f(X_n)||}.
\label{ch5meantamed6}
\end{aligned}$$ Using the hypothesis $\overline{\beta}(1+2C)-2\beta<0$, becomes $$\begin{aligned}
||X_{n+1}||^2&\leq&||X_n||^2+K\Delta t||X_n||^2+\theta^2||X_n||^2||\Delta W_n||^2+C^2||X_n||^2|\Delta N_n|^2\nonumber\\
&+&2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, g(X_n)\Delta W_n\right\rangle-2\mu||X_n||^2|\Delta N_n|+2CK|\Delta N_n|\nonumber\\
&+&2\langle g(X_n)\Delta W_n, h(X_n)\Delta N_n\rangle.
\label{ch5meantamed7}
\end{aligned}$$
- On $\Omega_n^c$, we have : $$\begin{aligned}
\dfrac{\Delta t^2||f(X_n)||^2}{\left(1+\Delta t||f(X_n)||\right)^2}&\leq& \dfrac{\Delta t^2||f(X_n)||^2}{1+\Delta t||f(X_n)||}
\leq \dfrac{2\Delta t^2\overline{\beta}^2||X_n||^{2a}}{1+\Delta t||f(X_n)||}+2K^2\Delta t^2||X_n||^2\\
&\leq& \dfrac{2\Delta t^2\overline{\beta}^2||X_n||^{a+1}}{1+\Delta t||f(X_n)||}+2K^2\Delta t^2||X_n||^2.
\label{ch5meantamed8}
\end{aligned}$$ Therefore, using , and , becomes $$\begin{aligned}
||X_{n+1}||^2&\leq&||X_n||^2+2K^2\Delta t^2||X_n||^2+\theta^2||X_n||^2||\Delta W_n||^2+C^2||X_n||^2|\Delta N_n|^2\nonumber\\
&+&2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, g(X_n)\Delta W_n\right\rangle-2\mu||Y_n||^2|\Delta N_n|+2CK|\Delta N_n|\nonumber\\
&+&2\langle g(X_n)\Delta W_n, h(X_n)\Delta N_n\rangle+\dfrac{\left[2C\overline{\beta}\Delta t-2\beta\Delta t+2\overline{\beta}^2\Delta t^2\right]||X_n||^{a+1}}{1+\Delta t||f(X_n)||}.
\label{ch5meantamed9}
\end{aligned}$$ The hypothesis $\Delta t<\dfrac{\beta-C\overline{\beta}}{\overline{\beta}^2}$ implies that $2C\overline{\beta}\Delta t-2\beta\Delta t+2\overline{\beta}^2\Delta t^2<0$. Therefore, becomes
$$\begin{aligned}
||X_{n+1}||^2&\leq&||X_n||^2+2K^2\Delta t^2||X_n||^2+\theta^2||X_n||^2||\Delta W_n||^2+C^2||X_n||^2|\Delta N_n|^2\nonumber\\
&+&2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, g(X_n)\Delta W_n\right\rangle-2\mu||X_n||^2|\Delta N_n|+2CK|\Delta N_n|\nonumber\\
&+&2\langle g(X_n)\Delta W_n, h(X_n)\Delta N_n\rangle.
\label{ch5meantamed10}
\end{aligned}$$
From the above discussion on $\Omega_n$ and $\Omega_n^c$, the following inequality holds on $\Omega$ for all $\Delta t<\dfrac{\beta-C\overline{\beta}}{\overline{\beta}^2}$ and $\overline{\beta}(1+2C)-\beta<0$ $$\begin{aligned}
||X_{n+1}||^2&\leq&||X_n||^2+K\Delta t||X_n||^2+2K^2\Delta t^2||X_n||^2+\theta^2||X_n||^2||\Delta W_n||^2+C^2||X_n||^2|\Delta N_n|^2\nonumber\\
&+&2\left\langle X_n+\dfrac{\Delta tf(X_n)}{1+\Delta t||f(X_n)||}, g(X_n)\Delta W_n\right\rangle-2\mu||X_n||^2|\Delta N_n|+2CK|\Delta N_n|\nonumber\\
&+&2\langle g(X_n)\Delta W_n, h(X_n)\Delta N_n\rangle.
\label{ch5meantamed11}
\end{aligned}$$
Taking Expectation in both sides of , using the relation $\mathbb{E}||\Delta W_n||=0$, $\mathbb{E}||\Delta W_n||^2=\Delta t$, $\mathbb{E}|\Delta N_n|=\lambda\Delta t$ and $\mathbb{E}|\Delta N_n|^2=\lambda^2\Delta t^2+\lambda\Delta t$ leads to : $$\begin{aligned}
\mathbb{E}||X_{n+1}||^2&\leq& \mathbb{E}||X_n||^2+2K^2\Delta t^2\mathbb{E}||X_n||^2+\theta^2\Delta t\mathbb{E}||X_n||^2+\lambda^2C^2\Delta t^2\mathbb{E}||X_n||^2+\lambda C^2\Delta t\mathbb{E}||X_n||^2\\
&-&2\mu\lambda\Delta t\mathbb{E}||X_n||^2+\lambda CK\Delta t\mathbb{E}||X_n||^2\\
&=&\left[1+(2K^2+\lambda^2C^2)\Delta t^2+(K+\theta^2+\lambda C^2-2\mu\lambda+2\lambda CK)\Delta t\right]\mathbb{E}||X_n||^2.
\end{aligned}$$ Iterating the last inequality leads to $$\begin{aligned}
\mathbb{E}||X_n||^2\leq \left[1+(2K^2+\lambda^2C^2)\Delta t^2+(K+\theta^2+\lambda C^2-2\mu\lambda+2\lambda CK)\Delta t\right]^n\mathbb{E}||X_0||^2.
\end{aligned}$$ In order to have stability, we impose $$\begin{aligned}
1+(2K^2+\lambda^2C^2)\Delta t^2+(K+\theta^2+\lambda C^2-2\mu\lambda+2\lambda CK)\Delta t<1.
\end{aligned}$$ That is $$\begin{aligned}
\Delta t<\dfrac{-[K+\theta^2+\lambda C^2-2\mu\lambda+2\lambda CK]}{2K^2+\lambda^2C^2}.
\end{aligned}$$
Numerical Experiments
---------------------
In this section, we present some numerical experiments that illustrate our theorical strong convergence and stability results. For the strong convergence illustration of and , let’s consider the stochastic differential equation $$\begin{aligned}
dX_t=(-4X_t-X^3_t)dt+X_tdW_t+X_tdN_t,
\label{ch5numeric1}
\end{aligned}$$ with initial $X_0=1$. $N$ is the scalar poisson process with parameter $\lambda=1$. Here $u(x)=-4x$, $v(x)=-x^3$ $g(x)=h(x)=x$. It is easy to check that $u, v, g$ and $h$ satisfy the Assumptions \[ch5assumption1\].
For the illustration of the linear mean-square stability , we consider a linear test equation
$\left\{\begin{array}{ll}
dX(t)=aX(t^-)+bX(t^-)dW(t)+cX(t^-)dN(t)\\
X(0)=1
\end{array}
\right.$
We consider the particular case $a=-1$, $b=2$, $c=-0.9$ and $\lambda=9$. In this case $l=-0.91$, $\dfrac{-l}{(a+\lambda c)^2}<0.084$ and $\dfrac{2a-l}{a^2+\lambda c^2}<0.074$. $a(1+\lambda c\Delta t)<0$ for $\Delta t<0.124$. We test the stability behaviour of semi-tamed and of tamed Euler for different step-size, $\Delta t=0.02, 0.05$ and $0.08$. We use $7\times 10^3$ sample paths. For all step-size $\Delta t<0.083$ the semi-tamed Euler is stable. But for the step-size $\Delta t=0.08>0.074$, the tamed Euler scheme is unstable while the semi-tamed Euler scheme is stable. So the semi-tamed Euler scheme works better than the tamed Euler scheme.
Conclusion {#conclusion .unnumbered}
==========
In this thesis, we provided an overview in probability theory which allowed us to define some basic concepts in stochastic process. Under global Lipschitz condition, we proved the existence and the uniqueness of solution of stochastic differential equation (SDEs) with jumps. In general, it is difficult to find the exact solution of most SDEs. one tool to approach the exact solution is the numerical resolution. We provided in this dissertation some numerical techniques to solve SDEs with jumps. More precisely, we investigated the strong convergence of the compensated stochastic theta method (CSTM) under global Lipschitz condition. We investigated the stability of both CSTM and stochastic theta method (STM) and we proved that for the linear test equation, when $\dfrac{1}{2}\leq\theta\leq 1$ the CSTM is A-stable while the STM is not. So CSTM works better than STM.
Under non-global Lipschitz condition Euler explicit method fails to converge strongly while Euler implicit method converges strongly but requires much computational efforts. We extended the tamed Euler scheme by introducing the compensated tamed Euler scheme for SDEs with jumps, which converges strongly with standard order $0.5$ to the exact solution. We also extended the semi-tamed Euler scheme proposed in [@Xia2] for SDEs with jumps under non-global Lipschitz condition and proved his strong convergence. This latter enable us define the tamed Euler scheme and to prove his strong convergence which was not yet done in the literature. In this thesis, we also analysis the stability behaviours of both tamed and semi-tamed Euler schemes in the linear and the nonlinear case. We proved that these two numerical schemes reproduce the exponentially mean-square property of the exact solution.
All the numerical scheme presented in this work are of rate of convergence $0.5$. The tamed Misltein scheme was introduced in [@Xia3], where the authors proved the strong convergence of order $1$ of this scheme for SDEs without jumps. The case with jumps is not yet well developped in the litterature. The weak convergence under non-global Lipschitz condition is not yet investigated. In the future, We would like to focus on the opened topics mentioned above.
Appendix {#appendix .unnumbered}
========
The goal of this section is to present some Scilab codes for simulations.
A.1 Code for simulation of the mean square error {#a.1-code-for-simulation-of-the-mean-square-error .unnumbered}
------------------------------------------------
lambda = 1; Xzero = 1;
T = 1; N = 2^(14); dt = T/N; a=1, b=1, c=0.5, theta=1, A=a+lambda*c
M = 5000;
Xerr = zeros(M,5);
for s = 1:M,
dW = sqrt(dt)*grand(1,N,'nor',0,1)
W = cumsum(dW);
dN=grand(1,N,'poi',dt*lambda)-dt*lambda*ones(1,N);
W(1)=0
dP=dN+dt*lambda*ones(1,N);
P=cumsum(dP,'c');
P(1)=0;
X=linspace(0,1,N);
Xtrue=exp((a-1/2*b^2)*X+b*W+log(1+c)*P);
for p = 1:5
R = 2^(p-1); Dt = R*dt; L = N/R;
Xtemp = Xzero;
for j = 1:L
Winc = sum(dW(R*(j-1)+1:R*j));
Ninc=sum(dN(R*(j-1)+1:R*j));
Xtemp=(Xtemp+(1-theta)*A*Xtemp*Dt+b*Xtemp*Winc+c*Xtemp*Ninc)/(1-A*theta*Dt);
end
Xerr(s,p) = abs(Xtemp - Xtrue(N))^2;
end
end
Dtvals = dt*(2.^([0:4]));
T=mean(Xerr,'r')^(1/2);
disp(T);
plot2d('ll',Dtvals,T,[5])
plot2d('ll',Dtvals,Dtvals^(1/2),[3])
legends([ 'mean square error for CSTM for theta=1','Reference line' ],[5,3,2],2)
xtitle("Mean square stability for CSTM");
xlabel("Time")
ylabel("E|Y_L-X(T)|^0.5")
A.2 Code for the simulation of the mean square stability {#a.2-code-for-the-simulation-of-the-mean-square-stability .unnumbered}
--------------------------------------------------------
T=2500, M=5000, Xzero=1,
a=-7,b=1, c=1, lambda=4, Dt=25,
A=a+lambda*c // drift coefficient for compensated equation
N=T/Dt;
theta=[0.4995, 0.50,0.51];// different value for theta
Xms=zeros(3,N);//initialisation of the mean
for i=1:3
Xtemp=Xzero*ones(M,1);// initialization of the solution
for j=1:N
Winc=grand(M,1,'nor',0,sqrt(Dt)); // generation of random
//variables following the normal distribution
Ninc=grand(M,1,'poi',Dt*lambda)-Dt*lambda*ones(M,1);// generation of
// compensated poisson process
B=1-theta(i)*A*Dt
Xtemp=(Xtemp+(1-theta(i))*A*Xtemp*Dt+b*Xtemp.*Winc+c*Xtemp.*Ninc)/B;
Xms(i,j)=mean(Xtemp^2);
end
end
X=linspace(0,T,N);
plot(X,Xms(1,:),'b',X,Xms(2,:),'r',X,Xms(3,:),'g')
legends(['Theta=0.499' 'theta=0.50' 'Theta=0.51' ], [2,5,3],2)
xtitle("Mean-square stability for CSTM");
xlabel("tn")
ylabel("E|Yn|^2")
Acknowledgements {#acknowledgements .unnumbered}
================
This dissertation would not have been possible without the guidance and the help of several individuals who in one way or another contributed and extended their valuable assistance in the preparation and completion of this study.
I would like to take this opportunity to express my gratitude to my supervisor Dr. Antoine Tambue to give me the chance to work with him; for his availability, his time to guide my researches, for sincerity, kindness, patience, understanding and encouragement that I will never forget. I am very happy for the way you introduced me in this nice topic which gave us two preprints papers.
I would also like to thank Dr. Mousthapha Sene, my tutor for the time he took to read this work and give a valuable suggestions.
Special thanks to Neil Turok for his wonderful initiative. Thanks to all staff of AIMS-Senegal, tutors and my classmates. I would like to express my gratitude to all lecturers who taught me at AIMS-Senegal, I would like to mention here Pr. Dr. Peter Stollmann. I would like to extend my thanks to the Ph.D. students at the chair of Mathematics of AIMS-senegal for their moral support during my stay in Senegal.
Many thanks to all lecturers in my country (Cameroon) who have being training me since my undergraduate.
Big thanks to my family, to my friends and to everyone who supported me during my studies.
[00]{}
Avner Friedman, Stochastic Differential Equations and Apllications, volume 1 and 2 [A]{} Document Preparation System,
Bernt Oksendal and Agnés Sulem Applied Stochastic Control of Jump Diffusions, Third Edition,
Bernt Oksendal Stochastic Differential Equations: An Introduction with Applications Fift Edition, Corrected printing,
Jean Jacod and Philip Protter, Essentiel en Théorie des Probabilités,
Desmond J. Highan, Peter E. Kloeden Convergence and Stability of Implicit Methods for Jump-Diffusion Systems
D.J. Higham, P.E. Kloeden, Numerical methods for nonlinear stochastic differential equations with jumps,
Xiaojie Wang, Siqing Gan, Compensated stochastic theta methods for stochastic differential equations with jumps,
Martin Hutzenthaler, Arnult Jentzen and Peter E. Kloeden, Strong convergence of an explicit numerical method for SDES with nonglobally Lipschitz continous coefficients,
Martin Hutzenthaler, Arnult Jentzen, Convergence of the stochastic Euler scheme for locally Lipschitz coefficients,
Martin Hutzenthaler, Arnult Jentzen and Peter E. Kloeden, Divergence of the multilevel Monte Carlo Euler Euler Method for nonlinear Stochastic Differential Equations, .
Xiaofeng Zong, Fuke Wu, Chengming Huang, Convergence and stability of the semi-tamed Euler scheme for stochastic differential equations with non-Lipschitz continuous coefficients,
Xiaofeng, Siqing Gan, The tamed Milstein method for commutative stochastic differential equations with non-globally Lipschitz continuous coefficients,
Giueseppe Da Prato, Jerzy Zabczyk, Stochastic equations in infinite dimensions,
Fima C. Klebaner, Introduction To Stochastic Calculus With Applications, SECOND EDITION,
Philip E. Protter, Stochastic Integration and Differential Equations, Second Edition,
Sever Silvestru Dragomir, Some Gronwall Type Inequalities and Application, .
Desmond J. Highan, An Algorithmic Introduction to Numerical Simulation of Stochastic Differential Equations,
Carlo Marinelli, Michael Rockner, On the maximal inequalities of Burkholder, Davis and Gundy (2013), ,
Konstantinos Dareiotis, Chaman Kumar and Sotirios Sabanis, On Tamed Euler Approximations of SDEs with Random Coefficients and Jumps, .
A. Tambue and J. D. Mukam. Strong convergence of the tamed and the semi-tamed Euler schemes for stochatic differential equations with jumps under non-global Lipschitz condition .
A. Tambue and J. D. Mukam, Stability of the semi-tamed and tamed schemes for stochastic differential equations with jumps under non-global Lipschitz condition. .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'While detecting low mass exoplanets at tens of au is beyond current instrumentation, debris discs provide a unique opportunity to study the outer regions of planetary systems. Here we report new ALMA observations of the 80-200 Myr old Solar analogue HD 107146 that reveal the radial structure of its exo-Kuiper belt at wavelengths of 1.1 and 0.86 mm. We find that the planetesimal disc is broad, extending from 40 to 140 au, and it is characterised by a circular gap extending from 60 to 100 au in which the continuum emission drops by about 50%. We also report the non-detection of the CO J=3-2 emission line, confirming that there is not enough gas to affect the dust distribution. To date, HD 107146 is the only gas-poor system showing multiple rings in the distribution of millimeter sized particles. These rings suggest a similar distribution of the planetesimals producing small dust grains that could be explained invoking the presence of one or more perturbing planets. Because the disk appears axisymmetric, such planets should be on circular orbits. By comparing N-body simulations with the observed visibilities we find that to explain the radial extent and depth of the gap, it would require the presence of multiple low mass planets or a single planet that migrated through the disc. Interior to HD 107146’s exo-Kuiper belt we find extended emission with a peak at $\sim20$ au and consistent with the inner warm belt that was previously predicted based on 22$\mu$m excess as in many other systems. This warm belt is the first to be imaged, although unexpectedly suggesting that it is asymmetric. This could be due to a large belt eccentricity or due to clumpy structure produced by resonant trapping with an additional inner planet.'
author:
- |
S. Marino,$^{1}$[^1], J. Carpenter,$^{2}$ M. C. Wyatt$^{1}$, M. Booth$^{3}$, S. Casassus$^{4,5}$, V. Faramaz$^{6}$,\
$^{1}$Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK\
$^{2}$Joint ALMA Observatory (JAO), Alonso de Cordova 3107 Vitacura - Santiago de Chile, Chile\
$^{3}$Astrophysikalisches Institut und Universitätssternwarte, Friedrich-Schiller-Universität Jena, Schillergäßchen 2-3, 07745 Jena, Germany\
$^{4}$Departamento de Astronomía, Universidad de Chile, Casilla 36-D, Santiago, Chile\
$^{5}$Millennium Nucleus “Protoplanetary Disks”, Chile\
$^{6}$Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove drive, Pasadena CA 91109, USA.\
$^{7}$Department of Astronomy, Van Vleck Observatory, Wesleyan University, 96 Foss Hill Drive, Middletown, CT 06459, USA\
$^{8}$Department of Physics and Astronomy, Rice University, 6100 Main Street, MS-108, Houston, Texas 77005, USA\
$^{9}$Department of Physics, University of Warwick, Gibbet Hill Road, Coventry, CV4 7AL, UK\
$^{10}$Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA\
$^{11}$Department of Physics and Astronomy, California State University Northridge, 18111 Nordhoff Street, Northridge, CA 91130, USA\
$^{12}$National Radio Astronomy Observatory, 520 Edgemont Road, Charlottesville, VA 22903-2475, USA\
bibliography:
- 'SM\_pformation.bib'
date: 'Accepted XXX. Received YYY; in original form ZZZ'
title: A gap in the planetesimal disc around HD 107146 and asymmetric warm dust emission revealed by ALMA
---
\[firstpage\]
circumstellar matter - planetary systems - planets and satellites: dynamical evolution and stability - techniques: interferometric - methods: numerical - stars: individual: HD 107146.
Introduction {#sec:intro}
============
While exoplanet campaigns have discovered thousands of close in planets in the last decade, at separations greater than 10 au it has only been possible to detect a few gas giants, mainly through direct imaging [^2] [@Marois2008; @Lagrange2009betapicb; @Rameau2013]. Protoplanetary disc observations, on the other hand, have shown that enough mass in both dust and gas to form massive planets resides at large stellocentric distances [see review by @Andrews2015]. In addition, the detection of cold dusty debris discs at tens of au shows that planetesimals can and do form at tens and hundreds of au in extrasolar systems [e.g. @Su2006; @Hillenbrand2008; @Wyatt2008; @Carpenter2009; @Eiroa2013; @Absil2013; @Matthews2014pp6; @Thureau2014; @Montesinos2016; @Hughes2018], although the exact planetesimal belt formation mechanism is a matter of debate [e.g. @Matra2018mmlaw].
It is natural then to wonder *how far out can planets form?* In situ formation of the imaged distant gas giants is challenging as the growth timescale of their cores can easily take longer than the protoplanetary gas-rich phase [@Pollack1996; @Rafikov2004; @Levison2010]. Gravitational instability was thought to be the only potential pathway towards in-situ formation at tens of au [@Boss1997; @Boley2009], but the revisited growth timescale of embryos through pebble accretion could be fast enough to form ice giants or the core of gas giants during the disc lifetime [@Johansen2010; @Ormel2010; @Lambrechts2012; @Morbidelli2012pa; @Bitsch2015; @Johansen2017]. Alternatively, the observed giant planets at tens of au may have formed closer in and evolved to their current orbits by migrating outward [@Crida2009], as could be the case for HR 8799 with four gas giants in mean motion resonances [@Marois2008; @Marois2010] surrounded by an outer debris disc [@Su2009; @Matthews2014hr8799; @Booth2016], or may have been scattered from closer in onto a highly eccentric orbit [@Ford2008; @Chatterjee2008; @Juric2008], as has been suggested for Fomalhaut b [@Kalas2008; @Kalas2013; @Faramaz2015]. On the other hand, after the dispersal of gas and dust, planetesimals could continue growing to form icy planets at tens of au over 100 Myr timescales; however, numerical studies show that once a Pluto size object is formed at 30-150 au within a disc of planetesimals, these are inevitably stirred, stopping growth and the formation of higher mass planets through oligarchic growth [@Kenyon2002; @Kenyon2008; @Kenyon2010]. Thus, it is not yet clear how far from their stars planets can form. Moreover, the discovery of vast amounts of gas (possibly primordial) in systems with low, debris-like levels of dust [so-called “hybrid discs”, e.g. @Moor2017] has opened the possibility for long lived gaseous discs that could facilitate the formation of both ice and gas giant planets at tens of au.
Broad debris discs provide a unique tool to investigate planet formation at tens of au. Planets formed at large radii or evolved onto a wide orbit should leave an imprint in the parent planetesimal belt, and thus in the dust distribution around the system. Gaps have been tentatively identified in a few young debris discs using scattered light observations suggesting the presence of planets at large orbital radii clearing their orbits from debris, e.g. HD 92945 [@Golimowski2011] and HD 131835 [@Feldt2017]. However, alternative scenarios without planets that could also reproduce the observed structure have not been ruled out yet in these systems. For example, multiple ring structures can arise from gas-dust interactions if gas and dust densities are similar [@Lyra2013Natur; @Richert2017], which might explain HD 131835’s rings since large amounts of CO gas (likely primordial origin) have been found in this system [@Moor2017]. Moreover, the double ring structure around HD 92945 and HD 131835 has only been identified in scattered light images, tracing small dust grains whose distribution can be highly affected by radiation forces [@Burns1979], therefore not necessarily tracing the distribution of planetesimals [e.g. @Wyatt2006].
Only HD 107146, an $\sim80-200$ Myr old G2V star [@Williams2004 and references therein] at a distance of $27.5\pm0.3$ pc [@Gaiamission; @Gaiadr1], has a double debris ring structure tentatively identified at longer wavelengths thanks to the Atacama Large Millimeter/submillimeter Array [ALMA, @Ricci2015]. At these wavelengths observations trace mm-sized dust for which radiation forces are negligible, therefore indicating that the double ring structure is imprinted in the planetesimal distribution as well. Moreover, these observations ruled out the presence of gas at densities high enough to be responsible for such structure. The debris disc surrounding HD 107146 was first discovered by its infrared (IR) excess using IRAS data [@Silverstone2000], but it was not until recently that the disc was resolved by the *Hubble Space Telescope* (HST) in scattered light, revealing a nearly face on disc with a surface brightness peak at 120 au and extending out to $\sim160$ au [@Ardila2004; @Ertel2011; @Schneider2014]. Despite HST’s high resolution, limitations in subtracting the stellar emission or a smooth distribution of small dust likely kept the double ring structure hidden. Using ALMA’s unprecedented sensitivity and resolution, [@Ricci2015] showed that this broad disc extended from about 30 to 150 au, but that it had a decrease in the dust density at intermediate radii, which could correspond to a gap produced by a planet of a few Earth masses clearing its orbit at 80 au through scattering. Finally, analysis of *Spitzer* spectroscopic and photometric data revealed the presence of an extra unresolved warm dust component in the system, at a temperature of $\sim$120 K and thus inferred to be located between 5-15 au from the star [@Morales2011; @Kennedy2014].
[max width=1.0]{}
------------------ ------------------------ ------------------ ------------- ------------------------------------------ ----------------------------- ----------------- --------------------- ------------------
Observation Dates t$_\mathrm{sci}$ Image rms beam size (PA) Min and max baselines \[m\] Flux calibrator Bandpass calibrator Phase calibrator
\[min\] \[$\mu$Jy\] (5th and 95th percentiles)
Band 6 - 12m 24, 27-30 Apr 2017 236.4 6.8 $0\farcs67\times0\farcs66$ ($-2.9\degr$) 41 and 312 J1229+0203 J1229+0203 J1215+1654
Band 7 - 12m 11 Dec 2016 48.8 30.0 $0\farcs46\times0\farcs37$ ($21.2\degr$) 48 and 410 J1229+0203 J1229+0203 J1215+1654
Band 7 - ACA 20 Oct and 2 Nov 2016 135.1 245 $4\farcs3\times3\farcs4$ ($-76.4\degr$) 9 and 44 Titan J1256+0547 J1224+2122
22 Mar and 13 Apr 2017
Band 7 - 12m+ACA - - 31.1 $0\farcs47\times0\farcs38$ ($20.8\degr$) - - - -
------------------ ------------------------ ------------------ ------------- ------------------------------------------ ----------------------------- ----------------- --------------------- ------------------
Despite the tentative evidence of planets producing these gaps around their orbits, neither the HD 107146 ALMA observations, nor the scattered light observations of HD 92945 and HD 131835, ruled out alternative scenarios in which planets are not orbiting within these gaps, but similar structure is created in the dust distribution through different mechanisms. These different scenarios have important implications for the inferred dynamical history of the system and planet formation. While a planet formed in situ could explain the data reasonably well, questions arise regarding how a planet of a few Earth masses could have formed at such large separations, where coagulation and planetesimal growth timescales are significantly longer. Alternative scenarios such as the one suggested by [@Pearce2015doublering] to explain HD 107146’s gap, could avoid these issues. In that scenario a broad gap is produced by secular interactions between a planetesimal disc and a similar mass planet on an eccentric orbit, which formed closer in and was scattered out by an additional massive planet. That scenario also predicts that the planet’s orbit should become nearly circular and the planet would be located at the inner edge of the disc at the current epoch. The model also predicts the presence of asymmetries in the disc such as spiral features that would be detectable in deeper ALMA observations.
A second alternative scenario was proposed by [@Golimowski2011] to explain the double-ring structure around HD 92945 seen in scattered light. As shown by [@Wyatt2003], planet migration can trap planetesimals in mean motion resonances; resulting in overdensities that are stationary in the reference frame co-rotating with the planet. Small dust released from these trapped planetesimals can exit the resonances due to radiation pressure, forming a double-ring structure that could be observable in scattered light [@Wyatt2006]. On the other hand, in this planet migration scenario the distribution of mm-sized dust should match the planetesimal distribution, with prominent clumps that could be seen in millimetre observations.
Finally, secular resonances produced by a single eccentric planet in a massive gaseous disc [@Zheng2017] or by two planets formed interior to the disc could also explain some of the wide gaps [@Yelvertoninprep], but possibly also leaving asymmetric features. Hence debris disc observations at multiple wavelengths can disentangle these different scenarios, and provide insights into the dynamical history of the outer regions of planetary systems, testing the existence and origin of planets at tens of au, which otherwise would remain invisible. Since radiation forces acting on the smallest dust grains have a significant effect on their distribution, the planetary perturbations discussed above can be best studied using ALMA observations which trace the distribution of large ($\sim$0.1-10mm) grains for which radiation forces are negligible, and thus follow the distribution of their parent planetesimals.
In this paper we present new ALMA observations of HD 107146 in both band 6 and 7 (1.1 and 0.86 mm). These observations resolve the broad debris disc around this system at higher sensitivity and resolution than the data presented by [@Ricci2015], which showed tentative evidence of a gap as commented above. This paper is outlined as follows. In §\[sec:obs\] we present the new ALMA observations of the dust continuum and line emission of HD 107146. Then in §\[sec:model\] we model the data using both parametric models and the output of N-body simulations to quantify the disc structure and assess whether a single planet could explain the observations. In §\[sec:dis\] we discuss our results, the origin of the gap and implications for planet formation; the total mass of HD 107146’s outer disc; the detection of an inner component that could be warm dust; and our gas non-detection. Finally the main conclusions of this paper are summarised in §\[sec:conclusions\].
Observations {#sec:obs}
============
HD 107146 was observed both in band 7 (0.86mm, project 2016.1.00104.S, PI: S. Marino) and band 6 (1.1mm, project 2016.1.00195.S, PI: J. Carpenter). Band 7 observations were carried out between October and December 2016 (see Table \[tab:obs\]) both using the 12m array and the Atacama Compact Array (ACA) to recover small and large scale structures. The total number of antennas for the 12m array was 42, with baselines ranging from 48 to 410 m (5th and 95th percentiles), and between 9 to 11 ACA 7m antennas with baselines ranging from 9 to 44 m. The correlator was set up with two spectral windows centered at 343.13 and 357.04 GHz with 2 GHz bandwidths and 15.625 MHz spectral resolution, and the other two centered at 345.03 and 355.14 GHz with 1.875 GHz bandwidths and 0.976 MHz spectral resolution. The four windows are used together to study the dust continuum emission, while the latter two are also used specifically to search for line emission from CO and HCN molecules in the disc (see §\[sec:gas\]). The weather varied between the multiple ACA band 7 observations, with average PWV values of 0.38, 0.72, 0.89 and 1.2 mm. The average PWV during the single 12m observation was 0.56 mm.
Band 6 observations were carried out in April 2017 (see Table \[tab:obs\]) using the 12m array only. We requested observations using two different 12m antenna configurations to recover well the large scale structure and, at the same time, achieve high spatial resolution, but due to time constraints the compact configuration observations were never carried out. We find however that the observations with the extended configuration have baselines short enough to recover the large scale structure (see §\[sec:continuum\] below). The total number of 12m antennas varied between 41 and 42, with baselines ranging from 41 to 312 m (5th and 95th percentiles). The correlator was set up with four spectral windows centered at 253.60, 255.60, 269.61 and 271.61 GHz with 2 GHz bandwidths and 15.625 MHz spectral resolution. The four are used together to study the dust continuum emission only. The weather also varied between the multiple band 6 observations, with PWV values of 0.89, 0.33, 0.30, 0.82 and 1.56 mm. Despite these variations we decided to use all the data sets to obtain the highest possible S/N. Calibrations were applied using the pipeline provided by ALMA and CASA 4.7. The total time on source for band 6 was 236 min, and 184 min for band 7 (49 and 135 min for the 12m array and ACA, respectively). Below we present the image analysis of continuum and line observations.
Continuum {#sec:continuum}
---------
Continuum maps at band 6 and 7 are created using the <span style="font-variant:small-caps;">clean</span> task in CASA 4.7 [@casa] and are presented in Figure \[fig:almab6b7\]. We adopt natural weighting for band 6 and 7 (12m+ACA combined) for a higher signal to noise. These maps have a rms of 6.3 and 27 $\mu$Jy beam$^{-1}$ at the center, which increases towards the edges as the maps are corrected by the primary beam, reaching values of 7.0 and 34 $\mu$Jy beam$^{-1}$ at $5\arcsec$ (140 au) for band 6 and 7, respectively. The band 6 and band 7 synthesised beams have dimensions of $0\farcs80\times0\farcs79$ and $0\farcs58\times0\farcs47$[^3], respectively, reaching an approximate resolution of 22 au for band 6 and 15 au for band 7. The higher resolution and sensitivity of these new data reveal a nearly axisymmetric broad disc with a large decrease or gap in the surface brightness centered at a radius of $\sim80$ au as suggested by [@Ricci2015]. The disc extends from nearly $40$ to 180 au, being one of the widest discs resolved at millimetre wavelengths [@Matra2018mmlaw]. We compute the total flux by integrating the disc emission inside an elliptical mask with a 180 au semi-major axis and the same aspect ratio as the disc (see §\[sec:parmodel\]). We find a total flux of $16.1\pm1.6$ and $34.4\pm3.5$ mJy at 1.1 and 0.86 mm (including 10% absolute flux uncertainties), leading to a spatially unresolved millimetre spectral index of $2.64\pm0.48$, consistent with results from [@Ricci2015atca] which combined ALMA and ATCA observations. The total flux in band 6 is consistent with that measured by [@Ricci2015] at similar wavelengths and using a more compact ALMA configuration, thus proving that our band 6 observations do not suffer from flux loss or miss large scale structure.
In order to study in more detail the disc radial structure, the top panel of Figure \[fig:radial\_profile\] shows the radial intensity profile, computed by azimuthally averaging the disc emission over ellipses as in [@Marino2016; @Marino2017etacorvi; @Marino201761vir]. Both in band 6 and 7, we find that the disc surface brightness peaks near the disc inner edge at $\sim45$ au, from which it decreases reaching a minimum at 80 au that is deeper in the band 7 profile likely due to the higher resolution. Beyond this minimum, the surface brightness increases until 120 au where it peaks and then decreases steeply with radius. No significant positive emission is recovered beyond 180 au. Within the gap, we find that the radial profile is not symmetric with respect to the minimum, with the outer section (80-100 au) having a steeper slope than the inner part (60-80 au), a feature that is present both in band 6 and 7 data. This is also visible in Figure \[fig:almab6b7\] and is an important feature that could shed light on the origin of this gap. We also compute a spectral index map using multi-frequency clean (<span style="font-variant:small-caps;">nterms</span>=2) and natural weights. The bottom panel of Figure \[fig:radial\_profile\] shows the azimuthally averaged radial profile of the spectral index (${\alpha_\mathrm{mm}}$). From 40 to 150 au the disc has spectral index of roughly $2.7\pm0.1$ (assuming ${\alpha_\mathrm{mm}}$ is constant over radii), consistent with the overall spectral index estimated above. Note that uncertainties on ${\alpha_\mathrm{mm}}$ do not include the 10% absolute flux uncertainties as we are only interested in relative differences as a function of radius.
In addition to the outer disc, we detect emission from within 30 au that peaks near the stellar position in the azimuthally averaged profile. However, it has a much higher level than the photospheric emission if we extrapolate this from available photometry at wavelengths shorter than 10$\mu$m using a Rayleigh-Jeans spectral index ($F_{\nu \star}=32\pm1$ and $18\pm1$ $\mu$Jy at 0.86 and 1.1 mm, respectively). Moreover, between 20 and 30 au we also find that ${\alpha_\mathrm{mm}}$ has a peak of $\sim3.7$, although still within $2\sigma$ from the average spectral index. In §\[sec:parmodel\] we recover this inner component in more detail after subtracting the disc emission using a parametric model and we find that it is inconsistent with point source emission, it is significantly offset from the stellar position, and unlikely to be a background sub-millimetre galaxy.
The fact that the disc emission is consistent with being axisymmetric disfavours the scenario in which a planet is scattered out from the inner regions opening a gap through secular interactions with the disc [@Pearce2015doublering see their figure 6]. In their model, spiral density features are present in the planetesimal disc for hundreds of Myr, which should be imprinted in the mm-sized dust distribution as well, thus being detectable by our observations. In §\[sec:parmodel\] we fit parametric models to the data to study with more detail the level of axisymmetry of HD 107146’s disc. Moreover, in §\[sec:nbody\] we compare our observations with N-body simulations of a planet on a circular orbit clearing a gap in a planetesimal disc.
CO J=3-2 and HCN J=4-3 {#sec:gas}
----------------------
Despite the increasing number of CO gas detections in nearby debris discs [e.g. @Dent2014; @Moor2015gas; @Marino2016; @Marino2017etacorvi; @Lieman-Sifry2016; @Matra2017fomalhaut; @Moor2017], no CO v=0 J=3-2 emission is detected around HD 107146 in the dirty continuum-subtracted data cube. This non detection is, however, not surprising since HD 107146 is significantly older and fainter than the other systems in which primordial gas has been found (i.e. the hybrid discs). Furthermore, the sensitivity of these observations is expected to be insufficient to detect CO gas if being released through collisions of volatile-rich planetesimals [@Kral2017CO]. We search more carefully for CO emission by applying a matched filter technique [@Matra2015; @Marino2016; @Marino2017etacorvi; @Matra2017fomalhaut] in which we integrate the emission over an elliptic mask with the same orientation as the disc on the sky (see §\[sec:model\]), but in each spatial pixel we integrate only over those frequencies (i.e. radial velocities) where line emission is expected taking into account the Doppler shift due to Keplerian rotation. For this, we assume a stellar mass of 1 $M_{\sun}$, inclination of $19\degr$, PA of $153\degr$, and deprojected minimum and maximum radii of 40 and 150 au, respectively. Using this method we obtain an integrated line flux $3\sigma$ upper limit of 74 mJy km s$^{-1}$. Note that we consider the two possible directions of the rotation, obtaining similar limits. We also search for emission that could be present at a specific radius by integrating the emission azimuthally, however no significant emission is found. In §\[dis:gas\], we use this total flux upper limit to estimate an upper limit on the CO gas mass that could be present in the disc.
Similarly, we search for HCN emission, finding no significant emission. Based on this non detection we place a $3\sigma$ upper limit of 91 mJy km s$^{-1}$, which we also use in §\[dis:gas\] to constrain its abundance in planetesimals in this system. HCN is of particular interest for exocometary studies as, besides being abundant in Solar System comets [@Mumma2011], it has recently been suggested that it could play a key role for prebiotic chemistry in habitable planets [@Patel2015; @Sutherland2017].
Modelling {#sec:model}
=========
In this section we model the data using parametric models to constrain the density distribution of solids in the system (§\[sec:parmodel\]), and N-body simulations of a planet embedded in a planetesimal disc tailored to HD 107146 to constrain the mass and orbit of a putative planet carving the observed gap (§\[sec:nbody\]). In both approaches we model the central star as a G2V type star with a mass of 1 $M_{\sun}$, an effective temperature of 5750 K and a radius of 1 $R_{\sun}$.
Parametric model {#sec:parmodel}
----------------
We first use a set of parametric models to study the underlying density distribution of mm-sized dust in the system, which we fit directly to the observed visibilities as in [@Marino2016; @Marino2017etacorvi; @Marino201761vir]. Inspired by the radial profile of the dust emission (Figure \[fig:radial\_profile\]), we first choose as a disc model an axisymmetric disc with a surface density that is parametrized as a triple power law that divides the disc into an inner edge, an intermediate section (where the bulk of the dust mass is) and an outer edge. On top of this, the triple power law surface density distribution has a gap, which we parametrise with a Gaussian profile, to reproduce the depression seen in the ALMA data. This parametrization introduces a total of 9 parameters that define the surface density as follows,
$$\begin{aligned}
&\Sigma(r)=&\Sigma_0 f_\mathrm{gap}(r) \begin{cases}
\left(\frac{r}{r_{\min}}\right)^{\gamma_1} & \text{$r<r_{\min}$}, \\
\left(\frac{r}{r_{\min}}\right)^{\gamma_2} & \text{$r_{\min}<r<r_{\max}$}, \\
\left(\frac{r}{r_{\max}}\right)^{\gamma_3}\left(\frac{r_{\max}}{r_{\min}}\right)^{\gamma_2} & \text{$r>r_{\max}$}, \\
\end{cases} \label{eq:self-stir}\\
&f_\mathrm{gap}(r)=& 1-\delta_\mathrm{g}\exp\left[-\frac{(r-r_\mathrm{g})^2}{2\sigma_\mathrm{g}^2}\right],\end{aligned}$$
where ${r_{\min}}$ and ${r_{\max}}$ are the inner and outer radii of the disc, $\gamma_{1,2,3}$ determine how the surface density varies interior to the disc inner radius, within the disc and beyond the disc outer radius. The gap is parametrized with a fractional depth ${\delta_\mathrm{g}}$, a center ${r_\mathrm{g}}$ and a full width half maximum (FWHM) ${w_\mathrm{g}}=2\sqrt{2\ln(2)}{\sigma_\mathrm{g}}$. We leave as a free parameter the total dust mass ${M_\mathrm{d}}$, which is the surface integral of $\Sigma(r)$. Although the disc is close to face on, we still model the dust distribution in three dimensions adding the scale height $h$ as an extra parameter (i.e. vertical standard deviation of $hr$), and imposing a prior of $h>0.03$.
We solve for the dust equilibrium temperature and compute images at 0.86 and 1.13 mm using <span style="font-variant:small-caps;">RADMC-3D</span>[^4]. We assume a weighted mean dust opacity corresponding to dust grains made of a mix of astrosilicates [@Draine2003], amorphous carbon and water ice [@LiGreenberg1998], with mass fractions of 70%, 15% and 15%, respectively, and assuming a size distribution with an exponent of -3.5 and minimum and maximum sizes of 1$\mu$m and 1cm. This translates to a dust opacity of 1.5 cm$^2$ g$^{-1}$ at 1.1 mm. We note that these choices in dust composition and size distribution have no significant effect on our modeling apart from the derived total dust mass. We then use these images to compute model visibilities at the same *uv* points as the 12m band 6, 12m band 7 and ACA band 7 observations by taking the Fast Fourier Transform after multiplying the images by the corresponding primary beam. Additionally, we leave as free parameters the disc inclination ($i$), position angle (PA), RA and Dec offsets for the three observation sets, and a disc spectral index (${\alpha_\mathrm{mm}}$) that sets the flux at 0.86 mm given the dust mass and opacity at 1.1 mm, i.e. the size distribution is assumed to be the same throughout the disc. In total, our model has 19 parameters, 10 for the density distribution and 9 for the disc centre, orientation, and spectral index.
To find the best fit parameters we sample the parameter space using the <span style="font-variant:small-caps;">PYTHON</span> module <span style="font-variant:small-caps;">emcee</span>, which implements Goodman & Weare’s Affine Invariant MCMC Ensemble sampler [@GoodmanWeare2010; @emcee]. The posterior probability distribution is defined as the product of the likelihood function (proportional to $\exp[-\chi^{2}/2]$) and prior distributions which we assume uniform, although we impose a lower limit for $h$ of 0.03 due to model resolution constraints. In computing the $\chi^2$ over the three visibility sets we applied three constant re-weighting factors for band 6, 12m band 7 and for ACA band 7 that ensures that the final reduced $\chi^2$ of each of the three sets is approximately 1 without affecting the relative weights within each of these data sets that are provided by ALMA. The re-scaling is necessary as the absolute uncertainty of ALMA visibilities can be offset by a factor of a few, even after re-weighting the visibilities with the task <span style="font-variant:small-caps;">statwt</span> in CASA 4.7. These factors could be alternatively left as free parameters by adding an extra term to the likelihood function, however we find no differences in our results compared to leaving them fixed during multiple tries. Therefore we opt for leaving them fixed.
[lrl]{} Parameter & best fit value & description\
\
${M_\mathrm{d}}$ \[$M_{\earth}$\]& $0.250\pm0.004$ & total dust mass\
${r_{\min}}$ \[au\] & $46.6^{+1.4}_{-1.5}$ & disc inner radius\
${r_{\max}}$ \[au\] & $135.6^{+1.1}_{-1.2}$ & disc outer radius\
$\gamma_1$ & $2.6^{+0.3}_{-0.2}$ & inner edge’s slope\
$\gamma_2$ & $0.26^{+0.08}_{-0.10}$ & disc slope\
$\gamma_3$ & $-10.5^{+0.9}_{-1.0}$ & outer edge’s slope\
${r_\mathrm{g}}$ \[au\] & $75.5^{+1.1}_{-1.2}$ & radius of the gap\
${w_\mathrm{g}}$ \[au\] & $38.6^{+4.5}_{-3.6}$& FWHM of the gap\
${\delta_\mathrm{g}}$ & $0.52^{+0.03}_{-0.02}$ & fractional depth of the gap\
$h$ & $0.12^{+0.04}_{-0.05}$ & scale height\
PA \[$\degr$\] & $153\pm3$& disc PA\
$i$ \[$\degr$\]& $19.3\pm1.0$ & disc inclination from face-on\
${\alpha_\mathrm{mm}}$ & $2.57\pm0.11$& millimetre spectral index\
\
${r_\mathrm{g}}$ \[au\] & $75.4^{+0.8}_{-0.7}$ & radius of the gap\
${w_\mathrm{g}}$ \[au\] & $42.2^{+1.7}_{-2.2}$ & width of the gap\
${\delta_\mathrm{g}}$ & $0.43\pm0.02$ & depth of the gap\
\
${e_\mathrm{d}}$ & <0.03 & disc global eccentricity\
\
${r_{\min}}$ \[au\] & $41.9^{+1.2}_{-1.4}$ & disc inner radius\
$\gamma_1$ & $11.6^{+3.0}_{-2.7}$ & inner edge’s slope\
$\gamma_2$ & $0.03^{+0.19}_{-0.26}$ & disc slope\
${r_\mathrm{g}}$ \[au\] & $72.1^{+2.2}_{-2.9}$ & radius of the gap\
${w_\mathrm{g}}$ \[au\] & $51^{+12}_{-8}$& FWHM of the gap\
${\delta_\mathrm{g}}$ & $0.58^{+0.08}_{-0.06}$ & fractional depth of the gap\
${M_\mathrm{c}}$ \[$M_{\earth}$\] & $3.0^{+0.9}_{-0.6}\times10^{-3}$ & dust mass inner component\
${r_\mathrm{c}}$ \[au\] & $19.3^{+2.8}_{-2.8}$ & radius of inner component\
${\Delta r_\mathrm{c}}$ \[au\] & $35.8^{+9.1}_{-6.7}$ & radial width of inner component\
${\omega_\mathrm{c}}$ \[$\degr$\] & $85^{+9}_{-9}$ & PA of inner component south of disc PA,\
& & and in the disc plane\
${\sigma_{\phi}}$ \[$\degr$\] & $94^{+15}_{-12}$ & azimuthal width of inner component\
In Table \[tab:mcmc\] we present the best fit parameters of our 3-power law model with a Gaussian gap. We find a disc inner radius of 47 au that is significantly larger than the inner edge of $\sim25-30$ au derived by [@Ricci2015]. This is because in their model they considered a sharp inner edge, while in ours we allow for the presence of dust within this inner radius, but decreasing towards smaller radii. We find that $r_{\min}$ matches well the radius at which the surface density peaks as seen in Figure \[fig:radial\_profile\]. Similarly, our estimate of the outer radius (136 au) is significantly smaller than their outer edge (150 au). We find that the disc inner edge has a power law index (slope hereafter) of $\sim2.6$, while the outer edge is much steeper with a slope of $-11$. The intermediate component has a very flat slope of $\sim0.3$, $2.7\sigma$ flatter than the previous estimate (0.59), although the difference could be simply due to different parametrizations, i.e. considering a three vs a single power law parametrization or leaving the gap’s depth as a free parameter vs fixed to 1. Note that the slope derived from the intermediate component does not imply that the mass surface density of planetesimals is flat. In fact, from collisional evolution models we expect the surface density of mm-sized dust and optical depth to have a lower slope than the total mass surface density in regions of the disc where the largest planetesimals are not yet in collisional equilibrium [i.e. have a lifetime longer than the age of the system, @Schuppler2016; @Marino201761vir; @Geiler2017]. Moreover, we also expect that in the inner regions where the largest planetesimals are in collisional equilibrium, the surface density of material should have a slope close to $7/3$ [@Wyatt2007hotdust; @Kennedy2010], as we find interior to $r_{\min}$. This suggests that the regions interior to 47 au might be relatively depleted of solids simply due to collisional evolution rather than clearing by planets or inefficient planetesimal formation. We compare the derived inner radius and surface density of millimetre grains ($\sim3\times10^{-6}$ [M$_\oplus$]{} au$^{-2}$ at 50 au) with collisional evolution models by [@Marino201761vir], which include how the size distribution evolves at different radii, to estimate the maximum planetesimal size and initial total surface density of solids. We find that the best match has a maximum planetesimal size of $\sim10$ km and an initial disc surface density of 0.015 [M$_\oplus$]{} au$^{-2}$ at 50 au, i.e. 5 times the surface density of the Minimum Mass Solar Nebula [@Weidenschilling1977mmsn; @Hayashi1981] extrapolated to large radii, or a total solid disc mass of 300 [M$_\oplus$]{}. This implies a very massive initial disc and efficient planetesimal formation at large radii. These conclusions assume, however, that the observed structure in the surface brightness profile arise from collisional evolution neglecting alternative origins. Finally, regarding the disc orientation, we find values that are consistent with previous estimates, but with tighter constraints (see Table \[tab:mcmc\]).
Our results show that the gap is centered at $75.5^{+1.1}_{-1.2}$ au, consistent with the previous estimate. However, we find a FWHM of $\sim$40 au that is much larger than the previous estimate of 9 au, likely due to [@Ricci2015] assuming a gap depth of 100%. Instead, we fit the depth of the gap finding a best fit value of 0.5. We also fit an alternative model in which the gap is a step function with a constant depth, finding best fit values that are similar to the model with the Gaussian gap (see Table \[tab:mcmc\]). The difference in widths between the previous study and this work is interesting as [@Ricci2015] derived the mass of a putative planet clearing a gap in the disc based on the gap’s width and assuming that it should be roughly equal to the planet’s chaotic zone. A 3-4 times wider gap would imply a much higher planet mass as the width of the chaotic zone scales as ${M_\mathrm{p}}^{2/7}$ [@Wisdom1980; @Duncan1989]. Such planet mass estimates assume that the system is in steady state, that the gap is devoid of material, and that the gap’s width is simply equal to the chaotic zone’s width. However, given the age of the system ($\sim$80-200 Myr), and that any planet may be younger, it is reasonable to consider that the distribution of particles in the planet’s vicinity could still be evolving. Therefore, instead of using the gap’s width to estimate a planet mass, in §\[sec:nbody\] we estimate this by comparing with N-body simulations tailored to HD 107146.
Using the best fit PA and $i$ we deproject the observed visibilities and bin them to compare them with our axisymmetric model in Figure \[fig:vis\] (continuous line). The 12m and ACA band 7 observations are consistent with each other within errors, with some systematic differences due to the different primary beams. We find that our best fit model fits well both the real and imaginary components, with the imaginary part being consistent with zero as expected for an axisymmetric disc. However, we find a significant deviation between the real components of the 12m band 7 data and model at around 50 k$\lambda$ (corresponding to an angular scale of $4\arcsec$), with the model visibility predicting a slightly lower value than the data. This could be due to large scale variations in the spectral index as the same model fits well those baselines in the band 6 data. This deviation is the same when comparing the model with a step gap, which has an indistinguishable visibility profile for baselines shorter than 200 k$\lambda$. Beyond 150 k$\lambda$ we also find deviations between the data and model, expected since as shown in Figure \[fig:radial\_profile\] the radial profile seems to have structure that is more complex than a simple Gaussian gap.
A more intuitive way to study the goodness of fit of our model is to look at the dirty map of the residuals. The left column of Figure \[fig:res\] shows the dirty maps of the residuals after subtracting our best fit model from the data (in visibility space). These are computed using natural weights, plus an outer taper of $0\farcs8$ for band 7 to obtain a synthesised beam of similar size compared to the band 6 beam. Both band 6 and 7 residual images show significant residuals (with peaks of 9 and 5 times the rms) within the disc inner edge and with a peak that is significantly offset from the stellar position by $\sim0\farcs5$, corresponding to a projected distance of $\sim15$ au. Moreover, the emission is marginally resolved, extending radially by more than a beam (i.e. larger than 10 au) and with an integrated flux significantly higher than the predicted photospheric emission at these wavelengths assuming a Rayleigh-Jeans spectral index (18 and 32 $\mu$Jy at 1.1 and 0.86 mm). We estimate the flux of these components by integrating the dirty maps over a circular region of $2\arcsec$ diameter, finding an integrated flux of $85\pm13$ $\mu$Jy and $290\pm70$ $\mu$Jy at 1.1 and 0.86 mm, respectively (without including 10% absolute flux uncertainties). Because this inner emission is resolved, it cannot arise from a compact source such as a planet or circumplanetary material.
We identify two plausible origins for this inner component. It could be the same warm dust that was inferred based on Spitzer data [@Morales2011], as the clump is at a radial distance that is consistent with the inferred black body radius (5-15 au) from its spectral energy distribution (SED), although our detection would imply that the warm component has an asymmetric distribution (see discussion in §\[dis:warmdust\]). On the other hand, the residual could be also due to a background sub-millimeter galaxy that are often detected incidentally as part of ALMA deep observations [see @Simpson2015; @Carniani2015; @Marino201761vir; @Su2017]. Sub-millimeter galaxies have typical sizes of the order of 1 and spectral indices ranging between $\sim3-5$ at mm-wavelengths, thus consistent with the observed clump. We estimate a probability of $\sim50\%$ of finding a sub-millimeter galaxy as bright as 0.3 mJy at 0.86 mm or 0.1 mJy at 1.1 mm, respectively, within the entire band 6 and 7 primary beams [@Simpson2015; @Carniani2015]. Hence, the detection of such a bright background object is not a rare event. However, we find that the probability of finding such a bright sub-millimeter galaxy at 0.86 mm and 1.1 mm and co-located with warm dust (i.e within $1\arcsec$ from the star) is only 0.6% and 1%, respectively; therefore favouring the warm dust scenario. In fact, we do not detect any other compact emission above $5\sigma$ within the band 7 and 6 primary beams, consistent with the number counts of sub-millimeter galaxies. We also discard that this emission could originate from the galaxy detected using HST in 2004, 2005 and 2011 [@Ardila2004; @Ertel2011; @Schneider2014], as its position would only have changed by $1.4\arcsec$, and therefore still lie at $\sim5\arcsec$ ($\sim140$ au) from the star in 2017, and is not detected in our observations.
### Disc global eccentricity
As shown by [@Pearce2014; @Pearce2015doublering], a planet on an eccentric orbit can force an eccentricity in a disc of planetesimals through secular interactions. Here we aim to assess if HD 107146’s debris disc could have a global eccentricity and pericentre, using the same parametrization as in [@Marino2017etacorvi], i.e. taking into account the expected apocentre glow [@Pan2016]. We find that the disc is consistent with being axisymmetric, with a 2$\sigma$ upper limit of 0.03 for the forced eccentricity. We find though that the marginalised distribution of ${e_\mathrm{d}}$ peaks at 0.02 with a pericentre that is opposite to the residual inner clump. This peak is likely produced as the residuals are lower when the disc is eccentric and with an apocentre oriented towards the clump’s position angle due to apocentre glow. We therefore conclude that the fit is biased by the inner clump and that the disc is probably not truly eccentric.
### Inner component {#sec:modelin}
In order to constrain the geometry or distribution of the inner emission found in the residuals, we add an extra inner component by introducing five additional parameters to our reference parametric model (3-power law surface density with a Gaussian gap). We parametrise its surface density as a 2D Gaussian in polar coordinates, with a total dust mass ${M_\mathrm{c}}$ and centered at a radius ${r_\mathrm{c}}$ and azimuthal angle ${\omega_\mathrm{c}}$ (measured in the plane of the disc from the disc PA and increasing in an anti-clockwise direction). The width of this Gaussian is parametrized with a radial FWHM ${\Delta r_\mathrm{c}}$ and an azimuthal standard deviation ${\sigma_{\phi}}$. Best fit values are presented in Table \[tab:mcmc\]. We find that this inner component is extended both radially and azimuthally, but concentrated around $19\pm3$ au and orthogonal to the disc PA (${\omega_\mathrm{c}}\sim90\degr$), as we found in the residuals of our axisymmetric model. In the right panel of Figure \[fig:res\] we present the model image of the best fit model and its residuals, which are below $3\sigma$ in both band 6 and 7 within the disc inner edge. We also find that the total dust mass of this inner component is $3.0^{+0.9}_{-0.6}\times10^{-3}$ [M$_\oplus$]{} ($\sim$1% of the outer disc mass).
When adding this extra inner component we find that the slope of the inner edge is steeper and the inner radius smaller compared with our previous model (symmetric disc model hereafter), as it was probably compensating for the emission within 30 au with a less steep inner edge. We also find a slightly smaller gap radius of 72 au, a larger and deeper gap (51 au wide and 0.58 deep), and a flatter surface density slope of $0.0\pm0.2$. These differences in the gap’s structure and disc slope are overall consistent within $3\sigma$ with our previous estimates, but significantly improve the fit at large baselines as Figure \[fig:vis\] shows (dashed blue and orange lines). These improvements in the fit at large baselines are not due to the addition of the clump, but due to the different best fit surface density profile of the outer disc. In fact, the visibilities of the inner component are negligible beyond 200 k$\lambda$. The rest of the parameters are consistent within $1\sigma$ with the values presented for the symmetric disc model. To have a better estimate of the flux of this inner component, we subtract the new best fit of the outer component (its inner edge is steeper), finding an integrated flux over a circular region of $3\arcsec$ diameter of $0.82\pm0.14$ and $0.31\pm0.04$ mJy at 0.86 and 1.1 mm, respectively (including 10% absolute flux uncertainties)[^5]. Note that this flux is a factor two higher than that estimated from the residuals of the symmetric disc model, as without the inner component the model tries to compensate for the emission interior to 40 au. From these fluxes we estimate a spectral index of $3.3\pm0.6$, thus still consistent with the typical observed spectral indices of debris discs. Based on this new flux estimate at 0.86 mm, we find an even lower probability of 0.1% and 0.3% of finding a sub-mm galaxy as bright as this inner emission at 0.86 and 1.1 mm, respectively, and within $1\arcsec$ from the star [@Simpson2015; @Carniani2015]. These results also confirm that the peak of this inner component is significantly offset from the star and it is incompatible with an axisymmetric inner component. If this emission is produced by warm dust, then it could bring valuable insights into the origin of warm dust emission in general, as it is inconsistent with an axisymmetric asteroid belt (see §\[dis:warmdust\]).
Because the new estimate of the inner edge slope is too steep to be consistent with being set by collisional evolution, the maximum planetesimal size is only constrained to be $\gtrsim10$ km. Despite this, the relative brightness between the inner and outer components can still be explained simply by collisional evolution. This has also been found for other systems with warm dust components, e.g. q$^1$ Eridani [@Schuppler2016], suggesting the presence of planets clearing the material in between [@Shannon2016].
N-body simulations {#sec:nbody}
------------------
In this section we compare the observations with dynamical N-body simulations of a planet embedded in a planetesimal disc. To simulate the gravitational interactions between the planet and particles we use the N-body software package <span style="font-variant:small-caps;">REBOUND</span> [@rebound], using the hybrid integrator <span style="font-variant:small-caps;">Mercurius</span>[^6] that switches from a fixed to a variable time-step when a particle is within a given distance from the planet (here chosen to be 8 Hill radii). The fixed time step is chosen to be 4% of the planet’s period, which is lower than 17% of the orbital period of all particles in the simulation. Although the total mass of the disc could be tens of [M$_\oplus$]{}, and thus comparable with the mass of the simulated planets, we assume particles have zero or a negligible mass. In §\[sec:massparticles\] we discuss the effect of considering a massive planetesimal disc on planet disc interactions.
Particles are initially randomly distributed in the system with a uniform distribution in semi-major axis ($a$) between 20 and 170 au (i.e. with a surface density proportional to $r^{-1}$), with an eccentricity and inclination uniformly distributed between 0-0.02 and 0-0.01 radians, respectively. We use a total number of $10^{4}$ particles, sufficient to sample the 150 au span in semi-major axes and recover smooth images of the disc density distribution. The planet is placed on a circular orbit at 80 au and assumed to have a bulk density of 1.64 g cm$^{-3}$ (Neptune’s density). We integrate the evolution of the system up to 200 Myr, roughly the upper limit for the age of the system. We run a set of simulations with planet masses varying from 10 to 100 $M_{\earth}$ with a 5 $M_{\earth}$ spacing.
In order to translate the outcome of these simulations to density distributions used to produce synthetic images, we populate the orbits of each particle with 200 points randomly distributed uniformly in mean anomaly as in [@Pearce2014], but in the frame co-rotating with the planet in order to see if there are resonant structures (e.g. loops or Trojan regions). We find however no significant azimuthal structures due to first or second order resonances in the derived surface density of particles. We weight the mass of each particle based on its initial semi-major axis to impose an initial surface density proportional to $r^{\gamma}$ between a minimum and maximum semi-major axis ($a_{\min}, a_{\max}$). Finally, because we only run simulations with a fixed planet semi-major axis of 80 au to save computational time, we scale all the distances in the output of the simulation and leave ${a_\mathrm{p}}$ as a free parameter (only varying roughly between 75-80 au). This linear scaling is motivated by the fact that some of the features that we are interested in should scale with semi-major axis (e.g. Hill radius, mean-motion resonances, chaotic zone), although some other important quantities, such as the scattering diffusion timescale [@Tremaine1993], have a dependency on $a$ that is different from linear. In Appendix \[appendix\] we test the validity of the scaling approximation for the narrow range of planet semi-major axis that we explore.
In Figure \[fig:nbody\] we present, as an example, the evolution of $a$, $e$ and the surface density of particles for planet masses of 10, 30 and 90 [M$_\oplus$]{}. The width of the chaotic zone is overlayed in grey (note that this is a region in semi-major axis rather than radius), to compare with the gap in the surface density cleared by the planet. Although by 100 Myr the chaotic zone is almost empty of material (except for particles in the co-rotation zone) the surface density is not zero inside the gap as some particles have apocentres or pericentres within this region while being scattered by the planet. Some particles also remain on stable tadpole or Trojan orbits until the end of the simulation, creating an overdensity within the gap at 80 au. Interior and exterior to the planet, a small fraction of particles are in mean motion resonance with the planet and have their eccentricities increased. In agreement with previous work, we find that the gap’s width approximates to the chaotic zone, and its width does not vary after 10 Myr for the planet masses explored.
To compare with observations, we use snapshots of the simulations at 50, 100 and 200 Myr since the age of the system is uncertain and could vary roughly within this range. These ages assume that the putative planet formed (or grew to its current mass) early during the evolution of this system, rather than recently, or that this is the time since the planet formed. For each assumed epoch, we explore the parameter space using the same MCMC technique as in §\[sec:parmodel\], varying ${a_{\min}}$ and ${a_{\max}}$, the surface density exponent $\gamma$, the semi-major axis of the planet, and its mass by interpolating the resulting surface density between two neighbouring simulations. We also leave as free parameters the disc orientation, pointing offsets and spectral index, and instead of varying the disc scale height we assume a flat disc. We find a best fit mass of $30\pm5$ [M$_\oplus$]{} and semi-major axis of $77\pm1$ au, independently of the assumed age since after 10 Myr there is no significant evolution in the orbits of particles near the planet (see Figure \[fig:nbody\]). We find, however, that our best fit cannot reproduce well the width and depth of the gap. This is illustrated in the top and middle panels in Figure \[fig:radial\_profile\_nbody\], where the model has a narrower gap compared with the band 6 radial profile, and thus overall larger residuals. Although larger planet masses could produce wider gaps, these would also be significantly deeper and thus inconsistent with the observations. Moreover, the planet gap could be even deeper if, for example, we assume a different starting condition with a depleted surface density within the planet’s feeding zone as it accreted a large fraction of that material while growing. To test this, we repeat the fitting procedure, but removing those particles that start the simulation within two Hill radii from the planet’s orbit (planet’s feeding zone). We find a slightly lower best fit planet mass of $26\pm3$ [M$_\oplus$]{}, but overall the fit is worse with a difference of $\sim300$ in the total $\chi^2$, strongly preferring the model with particles starting near the planet. Therefore, we conclude that a single planet on a circular orbit that was born within the outer disc is unable to explain the ALMA observations.
Discussion {#sec:dis}
==========
The gap’s origin
----------------
As discussed in §\[sec:intro\], we identify multiple scenarios where a single planet might open a gap in a planetesimal disc. Below we discuss these and how well they could fit the data.
### Single planet on a circular orbit
As we showed in §\[sec:nbody\], a single planet on a circular orbit and formed inside the gap could clear its orbit of debris through scattering, opening a gap with a width similar to the chaotic zone. Such a planet, however, needs to be very massive to produce a $\sim40$ au wide gap ($\gtrsim3$ [M$_\mathrm{Jup}$]{}), which, on the other hand, results in a gap that is significantly deeper than our observations suggest. Based on these constraints we find that a 30 [M$_\oplus$]{} planet is the best trade-of between the gap’s width and depth.
If, however, the planet could migrate (inwards or outwards) then these two observables could become compatible. A low mass planet that migrated through the disc could carve a sufficiently wide gap to explain the $\sim40$ au gap’s width. Migration could also explain the relative depth of $\sim50\%$ that is observed as after migrating, the planet would also leave behind debris that was scattered onto excited orbits, but that do no longer cross the planet’s orbit [e.g. @Kirsh2009], therefore the gap would still contain a significant fraction of the original material. Planet migration could be induced by planetesimal scattering, which is discussed in §\[sec:massparticles\].
### Multiple planets on circular orbits
If we assume planet migration is negligible (e.g. disc mass around the planet’s orbit is much lower than its mass), then multiple planets would need to be present to carve such a wide and shallow gap. Let us assume that the gap was carved by multiple equal mass planets orbiting between 60 and 90 au. As shown by [@Shannon2016] and in §\[sec:nbody\], the width of the gap and age of the system place tight constraints on the minimum mass of a planet to clear the region surrounding its orbit, and on the maximum number of planets that could be orbiting within the gap based on a stability criterion. Given HD 107146’s age limits of $\sim50$ and 200 Myr, only planets with masses greater than 10 [M$_\oplus$]{} would have enough time to clear their orbits via scattering. On the other hand, only planet masses $\lesssim30$ [M$_\oplus$]{} can create a gap that is not too deep compared with our observations. Given this range of planet masses, we estimate that a maximum of three 10 [M$_\oplus$]{} planets could orbit between 60 and 90 au spaced by 8 mutual Hill radii at the limits of long term stability [@Chambers1996; @Smith2009stability]. If we require planets to be spaced by more than 8 mutual Hill radii, this multiplicity reduces considerably to two or only one planet if planet masses are slightly higher. We tested this by running new simulations with the same parameters as in §\[sec:nbody\], but instead with a pair of 30 or 10 [M$_\oplus$]{} planets and semi-major axes ranging between 60 to 90 au. After a few iterations we found a good fit using two 10 [M$_\oplus$]{} planets with semi-major axes of 60 and 83 au (i.e. spaced by 12 Mutual Hill radii). Figure \[fig:radial\_profile\_nbody\_2plt\] shows the radial profile for this model reproducing the width and depth of the gap. The relative depth of the gap is similar to the one observed as there are still particles on stable orbits at around 70 au and trapped in the co-orbital regions of the two planets — note that there are residuals above $3\sigma$ at the inner and outer edge of the disc because this model has sharp boundaries in semi-major axis. If we assumed that no particles were present near the planet at the start of the simulation as commented in §\[sec:nbody\], then even lower mass planets creating narrower gaps would be needed in order to achieve an overall gap depth of 50%.
### Planet(s) on eccentric orbits
Although a single or multiple low mass planets might explain the observed gap, the formation of an ice giant planet at $\sim80$ au encounters significant difficulties compared to within a few tens of au (see §\[sec:pformation\]). The scenario proposed by [@Pearce2015doublering], where a planet opens a broad gap through secular interactions, circumvents some of these problems as the planet is formed closer in at 10-20 au, and scattered out by a more massive planet onto an eccentric orbit with a larger semi-major axis. This scenario was able to fit the mean radial profile derived by [@Ricci2015] and predicted the presence of asymmetries in the form of spiral features and a small offset between the inner and outer regions of the disc due to secular interactions with the planet. However, no significant offset between the inner and outer regions is present in our observations. By fitting an ellipse to the inner and outer bright arcs in the band 6 image (roughly at 50 and 110 au), we constrain the offset to be $0\farcs05\pm0\farcs02$, (i.e. consistent with zero and lower than 1.6 au). This translates to a maximum eccentricity of 0.03 if we assume that the outer bright arc is circular while the inner arc is eccentric and thus offset from the star. Moreover, our observations suggest that the disc inner edge is much steeper than predicted by [@Pearce2015doublering], thus inconsistent with their model. A last recently proposed scenario involves two interior planets with low but non-zero eccentricity, which open a gap in the outer regions due to secular resonances [see @Yelvertoninprep]. Although this scenario can produce an observable gap, in its simplest form the model produces a gap that is narrower than seen for HD 107146.
### Dust-gas interactions
There are other scenarios that might not require the presence of planets to produce a gap or multiple ring structures in the dust distribution. Photoelectric instability [@Klahr2005; @Besla2007; @Lyra2013Natur; @Richert2017] is one of those scenarios and has received particular attention lately as the presence of vast amounts of gas around a few debris discs suggests that this mechanism might be common. However, this mechanism is only important when dust-to-gas ratios are comparable and it is not yet clear if relevant for the dust distribution of large millimetre-sized grains that are not well coupled to the gas. Assuming some residual primordial gas could still be present in the disc, we convert our CO gas mass upper limit of $5\times10^{-6}$ [M$_\oplus$]{} (see §\[dis:gas\] below) to a total gas mass upper limit of $\sim0.05$ [M$_\oplus$]{} (assuming a CO/H$_2$ ratio of $10^{-4}$) or a surface density of $\sim10^{-5}$ g cm$^{-2}$. This upper limit is comparable to the total dust mass in millimetre grains, thus photoelectric instability could occur. However, using this gas surface density upper limit we estimate a Stokes number of $10^{5}$ for millimetre grains, thus the stopping time is much longer than the collisional lifetime of millimetre dust and of the order of the age of the system. Therefore, we conclude that photoelectric instability does not play an important role in the formation of the structure observed by ALMA around HD 107146.
Planet formation at tens of au {#sec:pformation}
------------------------------
Based on these new ALMA observations, if the gap was cleared by a single or multiple planets, these must have a low mass ($\lesssim$30 [M$_\oplus$]{}) and formed between 50 and 100 au. Otherwise, if the planets were formed closer in and scattered out onto an eccentric orbit, the disc would appear asymmetric [@Pearce2015doublering]. Such low planet masses and orbits at tens of au resemble the ice giant planets in the Solar System, but with a semi-major axis 2-3 times larger than Neptune’s. *Could such planets have formed in situ as these observations suggest?* HD 107146’s broad and massive debris disc indicates that planetesimals efficiently formed at a large range of radii from 40 to 140 au. Moreover, the mass of these putative planets ($\lesssim30$ [M$_\oplus$]{}) is consistent with the solid mass available in their feeding zones ($\sim4$ Hill radii wide), based on the dust surface density derived in §\[sec:parmodel\] and extrapolating it to the total mass surface density (assuming a size distribution $dN\propto D^{-3.5}dD$ and planetesimals up to sizes of 10 km). Nevertheless, in situ formation encounters the two following problems. First, although a planet at $\sim80$ au might grow through pebble accretion fast enough to form an ice giant before gas dispersal [@Johansen2010; @Ormel2010; @Lambrechts2012; @Morbidelli2012pa; @Bitsch2015], it requires the previous formation of a massive planetesimal (or protoplanet) of $10^{-2}-10^{-1}$ [M$_\oplus$]{} (so-called transition mass). However, newly born planetesimals through streaming instability [@Youdin2005] have characteristic masses of rather $10^{-6}-10^{-4}$ [M$_\oplus$]{} [@Johansen2015; @Simon2016]. These planetesimals can grow through the accretion of pebbles and smaller planetesimals, but this growth (so-called Bondi accretion) is very slow for low mass bodies at large stellocentric distances [@Johansen2015]. A possible solution to this problem is that the protoplanetary disc around HD 107146 was unusually massive and long-lived, which would increase the chances of forming such a protoplanet. Alternatively, the low mass protoplanet could have formed closer in and been scattered out and circularised during the protoplanetary disc phase. The second problem that in-situ formation faces is related to planet migration. A protoplanet growing to form an ice giant is expected to migrate inwards through type-I migration [@Tanaka2002]. While this might imply that the single or multiple putative planets might have formed at larger radii, our observations suggest that they would have attained most of their mass within the observed gap, i.e. within 100 au. If pebble accretion is fast enough, the planet could grow fast and migrate only slightly before disc dispersal.
*Why would planets form only between 60-90 au within this 100 au wide disc of planetesimals?* While no planets might have formed beyond 90 au as planetesimal accretion was too slow for a planetesimal to reach the transition mass and efficiently accrete pebbles, this is not the case for planetesimals formed at smaller radii between the inner edge of the disc and the orbit of the innermost putative planet. *Why did only planetesimals form between 40-60 au?* Planetesimal growth might have been hindered there if the orbits of large solids were stirred by inner planets, making their accretion rate slower. Alternatively, planetesimal growth between 60-90 au could have been more efficient due to a local enhancement in the available solid mass. As stated before, it is also possible that the low mass protoplanet did not form in situ, but it was scattered out from further in.
Massive planetesimal disc {#sec:massparticles}
-------------------------
Here we discuss the effect that a massive planetesimal disc could have on the conclusions stated above where we assumed a disc of negligible mass ($\ll10$ [M$_\oplus$]{}). As shown in §\[sec:model\], this debris disc is likely very massive and thus affects the dynamics of this system, e.g. because of the gravitational force on the planet and disc self-gravity. A massive disc can induce planetesimal driven migration where the planet migrates through a planetesimal disc due to the angular momentum exchange in close encounters [e.g. @Fernandez1984; @Ida2000; @Gomes2004]. This type of migration has been well studied in the context of the outer Solar System, as it could have driven an initially compact orbital configuration to a more extended and current configuration [@Hahn1999; @Hahn2005], or towards an orbital instability [e.g. @Tsiganis2005]. In all these models the outer planets scatter material in to Jupiter which ejects most of it, leading to an outward migration of the outer planets and a small inward migration of Jupiter. However, [@Ida2000] showed that even in the absence of interior planets self-sustained migration could have led Neptune’s orbit to expand due to the asymmetric planetesimal distribution around it. A more recent work by [@Kirsh2009], however, showed that a single planet embedded in a planetesimal disc migrates preferentially inwards due to the timescale difference between the inner and outer feeding zones [see also @Ormel2012]. Therefore, we expect that the putative planet at $\sim$80 au around HD 107146 is likely migrating or has migrated inwards since it formed.
Simple scaling relations from [@Ida2000] predict that the migration rate should be of the order of $$\left| \frac{da}{dt} \right| \approx \frac{4\pi\Sigma a^2}{M_{\star}}\frac{a}{T} \label{eq:migration}$$ which agrees roughly with numerical simulations [e.g. @Kirsh2009]. Given the expected surface density of planetesimals at 80 au around HD 107146 ($\Sigma\gtrsim10^{-2}$ [M$_\oplus$]{} au$^{-2}$ for a maximum planetesimal size of 10 km), Equation \[eq:migration\] predicts that the migration rate is such that the planet would have crossed the whole disc reaching its inner edge in only a few Myr. [@Kirsh2009] found, however, that when the planet mass exceeds that of the planetesimals within a few Hill radii, the migration rate decreases strongly with planet mass. Assuming that the gap is indeed caused by a single or multiple planets between 60-90 au, given the 30 [M$_\oplus$]{} upper limit for the planet mass we conclude that the surface density of planetesimals must be much lower than 0.01 [M$_\oplus$]{} au$^{-2}$ to hinder planetesimal driven migration, i.e. a total disc mass $\lesssim430$ [M$_\oplus$]{}. If not the surface density profile would probably be significantly different without a well defined gap. This dynamical upper limit together with the lower limit derived from collisional models constrain the total disc mass to be between $\sim300-400$ [M$_\oplus$]{}, which is close to the maximum solid mass available in a protoplanetary disc under standard assumptions (e.g. disc-to-stellar mass ratio of 0.1 and gas-to-dust mass ratio of 100). This particularly high disc mass at the limits of feasibility is not unique, but a confirmation of the so-called “disc mass problem” as many other young and bright discs need similar or even higher masses according to collisional evolution models [see discussion in @Krivov2018].
Although we expect that the gap produced by a migrating planet might look significantly different compared with the no migration scenario (e.g. wider and radially and perhaps azimuthally asymmetric), none of the above studies provided a prediction for the resulting surface density of particles during planetesimal driven migration that we could compare with our observations. In experimental runs with a 10 or 30 [M$_\oplus$]{} planet and a similar mass planetesimal disc (using massive test particles) we find that the planet migrates inward $\sim$5-40 au in 100 Myr depending on the planet and disc mass, and that the gap has an asymmetric radial profile that is significantly distinct from the no-migration scenario, which could explain why we find a slight asymmetry within the gap (see §\[sec:continuum\]). Future comparison with numerical simulations of planetesimal-driven migration could provide evidence for inward or outward migration, and thus set tighter constraints on the disc and planet mass. Dynamical estimates of the disc mass could also shed light on the disc mass problem, confirming the high mass derived from collisional models or rather indicating that these need to be revisited.
Warm inner dust component {#dis:warmdust}
-------------------------
While IR excess detections can be generally explained by a single temperature black body, there is a large number of systems that show evidence for a broad range of temperatures that are hard to explain simply as due to a single dusty narrow belt, even with temperatures varying as a function of grain size [e.g. @Backman2009; @Morales2009; @Chen2009; @Ballering2014; @Kennedy2014]. Although this might be explained by material distributed over a broad range of radii (like in protoplanetary discs), an alternative and very attractive explanation for these systems is the presence of a two-temperature disc, with an inner asteroid belt and an outer exo-Kuiper belt, analogous to the Solar System [e.g. @Kennedy2014; @Schuppler2016; @Geiler2017]. HD 107146 is a good example of this type of system [@Ertel2011; @Morales2011], with a significant excess at 22 $\mu$m that cannot be reproduced by models of a single outer belt, but rather is indicative of dust located within $\sim30$ au produced in an asteroid belt. These type of belts are typically hard to resolve due to the small separation to the star (hindering scattered light observations) where current instruments are not able to resolve and due to their low emissivity that peaks in mid-IR.
*Is the detected inner emission related to the $\sim20$ $\mu$m excess?* In order to test if the inner emission seen by ALMA is compatible with being warm dust, we use the parametric model developed in §\[sec:parmodel\] to see if it can reproduce the available photometry of this system, including the mid-IR excess. We introduce, though, two changes to the model assumptions: first, we modify the size distribution index from -3.5 to -3.36 to be consistent with the derived spectral index in this work and previous studies [@Ricci2015atca]; and second, we extend the size distribution from 1 to 5 cm. The latter is necessary to reproduce the photometry at 7 mm as the contribution from cm-sized grains is significant at these wavelengths. In Figure \[fig:sed\] we compare the model and observed SED, including the new ALMA photometric points of the inner component. Despite the simplicity of our model, it reproduces successfully both the photometry at 22 $\mu$m and at millimetre wavelengths, confirming that our detection is consistent with being warm dust emission.
These new ALMA observations provide unique information on the nature of the warm dust emission as it is marginally resolved (see Figure \[fig:continuum\_clump\]). The surface brightness peaks at $\sim20$ au, thus roughly in agreement with the predicted location based on SED modelling. The emission, however, is far from originating in an axisymmetric asteroid belt, but is rather asymmetric with a maximum brightness towards the south west side of the disc. *What could cause such an asymmetric dust distribution?* We identify three scenarios proposed in the literature that could cause long- or short-term brightness asymmetries in a disc or belt.
First, the inner emission could correspond to an eccentric disc, which at millimetre wavelengths would be seen brightest at apocentre. This is known as apocentre glow, which is caused by the increase in dust densities at apocentre for a coherent disc [@Wyatt2005secular; @Pan2016]. Disc eccentricities can be caused by perturbing planets [e.g. @Wyatt1999; @Nesvold2013; @Pearce2014], as it has been suggested to explain Fomalhaut’s eccentric debris disc [@Quillen2006; @Chiang2009; @Kalas2013; @Acke2012; @MacGregor2017] and HD 202628 [@Krist2012; @Thilliez2016; @Faramazinprep]. At long wavelengths, the contrast between apocentre and pericentre brightness is expected to be approximately [@Pan2016] $$\left(\frac{1-e/2}{1+e/2}\right)\left(\frac{1+e}{1-e}\right),$$ thus to reach a contrast higher than 2 (as observed for HD 107146’s inner disc), the disc eccentricity would need to be higher than 0.5. Assuming the disc eccentricity corresponds to the forced eccentricity (i.e. free eccentricities are much smaller), then the perturbing planet should have a very high eccentricity $\gtrsim0.5$. One potential problem with this scenario is that the outer disc does not show any hint of being influenced by an eccentric planet. This problem could be circumvented if the true eccentricity of the inner belt is lower than derived based on the image residuals (e.g. $\lesssim0.2$) and the inner planet has a semi-major axis of only a few au, as the forced eccentricity on the outer disc would be much smaller. Alternatively, if the disc or putative outer planet(s) are much more massive than the inner eccentric planet, the disc might remain circular as it is observed.
A second potential scenario relates to a recent collision between planetary embryos, which would release large amounts of dust at the collision point producing an asymmetric dust distribution that could last for $\sim1000$ orbits or $\sim1$ Myr at 20 au [@Kral2013; @Jackson2014]. In such a scenario the *pinch point* would appear brightest since the orbits of the generated debris converge where the impact occurred and more debris is created from collisions. Thus, the pinch point would appear radially narrow. However, both our band 6 and 7 datasets show that the emission is radially broad at its brightest point, spanning $\gtrsim20$ au. This could be circumvented if the collision occurred within a broad axisymmetric disc. Higher resolution observations are necessary to discard this scenario or confirm these two inner components.
Finally, a third possible scenario is that the asymmetric structure is caused by planetesimals trapped in mean motion resonances (typically 3:2 and 2:1) with an interior planet that migrated through the planetesimal disc [@Wyatt2002; @Wyatt2003; @Wyatt2006; @Reche2008]. The exact dust spatial distribution depends on the planet mass, migration rate and eccentricity. The single clump inferred from our observations suggests that planetesimals would be trapped predominantly on the 2:1 resonance rather than 3:2 as the latter only creates a two clump symmetric structure. Simulations by [@Reche2008] showed that in order to trap planetesimals in the 2:1 resonance, a Saturn mass planet (or higher) with a very low eccentricity was needed, therefore, placing a lower limit on the planet mass if the asymmetry is due to resonant trapping.
Future observations could readily distinguish between scenarios 1-2 and 3, as in the first two scenarios the orientation of the asymmetry should stay constant (precession timescales are orders of magnitude longer than the orbital period), while in the third scenario it should rotate at the same rate as the putative inner planet orbits the star. Moreover, higher resolution and more sensitive observations could reveal if the disc is eccentric, smooth with a radially narrow clump, or smooth with a radially broad clump (scenarios 1 to 3, respectively). Observations at shorter wavelength would also be useful. In the eccentric disc scenario we expect the disc to be eccentric and broader due to radiation pressure on small dust grains, while resonant structure would be completely absent and the disc should look axisymmetric as small grains are not trapped also due to radiation pressure. Finally, future ALMA observations should be able to definitely rule out the possibility of the inner emission arising from a sub-mm galaxy. Given HD 107146’s proper motion (-174 and -148 mas yr$^{-1}$ in RA and Dec. direction, respectively), we would expect that in 2 years any background object should shift by $0\farcs23$ towards the north east with respect to HD 107146, thus enough to be measured with ALMA observations of similar sensitivity and resolution to the ones presented here.
Gas non detections {#dis:gas}
------------------
In §\[sec:gas\] we search for CO and HCN secondary origin gas around HD 107146. Although we did not find any, here we use the flux upper limits, including also the 40 mJy km s$^{-1}$ upper limit for CO J=2-1 from [@Ricci2015], to derive an exocometary gas mass upper limit in this disc. It has been demonstrated that in the low density environments around debris discs, gas species are not necessarily in local thermal equilibrium (LTE), which typically leads to an underestimation of CO gas masses [@Matra2015]. Here we use the code developed by [@Matra2018] to estimate the population of the CO rotational level in non-LTE based on the radiation environment, densities of collisional partners, and also taking into account UV fluorescence. We consider as sources of radiation the star, the CMB and the mean intensity due to dust thermal emission within the disc (calculated using our model and RADMC-3D). We choose 80 au as the representative radius, which is approximately the middle radius of the disc, and left as a free parameter the density of collisional partners (here assumed to be electrons). Based on our model, the dust temperature should roughly vary between 50 to 30 K between the disc inner and outer edge. Figure \[fig:COmass\] shows the CO gas mass upper limit as a function of electron density for different kinetic temperatures. We find that for electron densities lower than $\sim10^2$ cm$^{-3}$, the CO J=2-1 upper limit is more constraining than J=3-2 and vice versa. Overall, the CO gas mass must be lower than $5\times10^{-6}$ [M$_\oplus$]{}. We then use equation 2 in [@Matra2017fomalhaut] and the estimated photodissociation timescale of 120 yr to estimate an upper limit on the CO+CO$_{2}$ mass fraction of planetesimals in the disc, but we find no meaningful constraint because this CO gas mass upper limit is much higher than the predicted $5\times10^{-7}$ [M$_\oplus$]{} if gas were released in collisions of comet-like bodies [e.g. @Marino2016; @Kral2017CO; @Matra2017betapic; @Matra2017fomalhaut].
In the absence of a tool to calculate the population of rotational levels for HCN, we estimate a mass upper limit assuming LTE. For temperatures ranging between 30-50 K, we find an upper limit of $3\times10^{-9}$ [M$_\oplus$]{}, which translates to an upper limit on the mass fraction of HCN in planetesimals of 3%, an order of magnitude higher than the observed abundance in Solar System comets [@Mumma2011]. Note that this HCN upper limit could be much higher due to non-LTE effects, thus this 3% limit must be taken with caution.
Conclusions {#sec:conclusions}
===========
In this work, we have analysed new ALMA observations of HD 107146’s debris disc at 1.1 and 0.86 mm, to study a possible planet-induced gap suggested by [@Ricci2015] with a higher resolution and sensitivity. These new observations show that HD 107146, a 80-200 Myr old G2V star, is surrounded by a broad disc of planetesimals from 40 to 140 au, that is divided by a gap $\sim$40 au wide (FWHM), centered at 80 au and 50% deep, i.e. the gap is not devoid of material. We constrained the disc morphology, mass and spectral index by fitting parametric models to the observed visibilities using an MCMC procedure. We find that the disc is consistent with being axisymmetric, and we constrain the disc eccentricity to be lower than 0.03.
The observed morphology of HD 107146’s debris disc suggests the presence of a planet on a wide circular orbit opening a gap in a planetesimal disc through scattering. We run a set of N-body simulations of a planet embedded in a planetesimal disc that we compare with our observations. We find that the observed morphology is best fit with a planet mass of 30 [M$_\oplus$]{}, but significant residuals appear after subtracting the best fit model. We conclude that the observed gap cannot be reproduced by the dynamical clearing of such a planet as the gap it creates is significantly deeper and narrower than observed. We discuss that this could be circumvented by allowing the planet to migrate (e.g. due to planetesimal driven migration) or by allowing multiple planets to be present. We discuss how a planet could have formed in situ if the primordial disc was massive and long-lived, and possibly grew to its final mass very quickly and by the end of the disc lifetime (e.g. through pebble accretion), avoiding significant inward migration and runaway gas accretion. Moreover, because the putative planet(s) could undergo very fast planetesimal driven migration, we set an upper limit on the surface density and total mass of the disc.
These ALMA observations also revealed unexpected emission near the star that is best seen when subtracting our best fit parametric model of the outer disc. This inner component has a total flux of 0.8 and 0.3 mJy at 0.86 mm and 1.1 mm, respectively, a peak intensity that is significantly offset from the star by $0\farcs5$ (15 au), and we resolve the emission both radially and azimuthally. Its radial location indicates that it could be the same warm dust that had been inferred to be between 10-15 au to explain HD 107146’s excess at $22\mu$m. Indeed, we fit this emission with an extra inner asymmetric component finding a good match with these ALMA observations and also with HD 107146’s excess at $22\mu$m. We constrain its peak density at 19 au, and a radial width of at least 20 au. We hypothesise that this asymmetric emission could be due to a disc that is eccentric due to interactions with an eccentric inner planet, asymmetric due to a recent giant collision, or clumpy due to resonance trapping with a migrating inner planet. On the other hand, we find that this inner emission is unlikely to be a background sub-mm galaxy, as the probability of finding one as bright as 0.8 mJy at 0.86 mm within the disc inner edge (i.e. co-located with the warm dust) is 0.1%.
Finally, although it had been demonstrated by [@Ricci2015] that no primordial gas is present around HD 107146, we search for CO and HCN gas that could be released from volatile-rich solids throughout the collisional cascade in the outer disc. However, we find no gas, but we place upper limits on the total gas mass and HCN abundance inside planetesimals, being consistent with comet-like composition.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Bertram Bitsch for useful discussion on planet formation and pebble accretion. JC and VG acknowledge support from the National Aeronautics and Space Administration under grant No. 15XRP15\_20140 issued through the Exoplanets Research Program. MB acknowledges support from the Deutsche Forschungsgemeinschaft (DFG) through project Kr 2164/15-1. VF’s postdoctoral fellowship is supported by the Exoplanet Science Initiative at the Jet Propulsion Laboratory, California Inst. of Technology, under a contract with the National Aeronautics and Space Administration. GMK is supported by the Royal Society as a Royal Society University Research Fellow. LM acknowledges support from the Smithsonian Institution as a Submillimeter Array (SMA) Fellow. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2016.1.00104.S and 2016.1.00195.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan) and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. Simulations in this paper made use of the REBOUND code which can be downloaded freely at http://github.com/hannorein/rebound.
Linear scaling of semi-major axis {#appendix}
=================================
In §\[sec:nbody\] we assumed that we can approximate the surface density of particles interacting with a planet at a given semi-major axis (between $\sim75-80$ au), by linearly scaling the distances in the output of a N-body simulation of a planet on a fixed semi-major axis of 80 au. In order to check the validity of this, we run three additional simulations with a planet semi-major axis of 75 au and masses of 10, 20 and 30 [M$_\oplus$]{}. Figure \[fig:nbody\_check\] compares the resulting surface density of these new simulations (dashed line) with simulations of a planet at 80 au, but scaled to 75 au. These three new simulations show that the linear scaling is a reasonable approximation, matching very well the structure seen in the simulations with a planet at 75 au. Only minor differences are present, which are of the order of the expected noise due to the random initial semi-major axis of particles in our simulations.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: http://exoplanet.eu/
[^3]: Note that the rms and beam sizes are different to the one reported in Table \[tab:obs\] as the imaging is done with natural weights to increase the S/N
[^4]: http://www.ita.uni-heidelberg.de/ dullemond/software/radmc-3d/
[^5]: We also measured the integrated flux of the inner component using the task <span style="font-variant:small-caps;">uvmodelfit</span> in CASA 4.7, obtaining values of $0.35\pm0.03$ and $0.81\pm0.14$ mJy at 1.1 and 0.86 mm, respectively.
[^6]: We also used Hermes in several trial runs; however, we decided to use Mercurius instead because with Hermes we obtained results that differ significantly from other integrators such as ias15, whfast and the hybrid integrator within Mercury. We found that some particles outside the chaotic zone were driven to unstable orbits and did not conserve their Tisserand parameter. The possibility of a potential bug in Hermes was confirmed by private communication with Hanno Rein.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Estimation of a deterministic quantity observed in non-Gaussian additive noise is explored via order statistics approach. More specifically, we study the estimation problem when measurement noises either have positive supports or follow a mixture of normal and uniform distribution. This is a problem of great interest specially in cellular positioning systems where the wireless signal is prone to multiple sources of noises which generally have a positive support. Multiple noise distributions are investigated and, if possible, minimum variance unbiased (MVU) estimators are derived. In case of uniform, exponential and Rayleigh noise distributions, unbiased estimators without any knowledge of the hyper parameters of the noise distributions are also given. For each noise distribution, the proposed order statistic-based estimator’s performance, in terms of mean squared error, is compared to the best linear unbiased estimator (BLUE), as a function of sample size, in a simulation study.'
author:
- |
Kamiar Radnosrati\
Department of Electrical Engineering\
Linköping University\
Linköping, Sweden\
`[email protected]` Gustaf Hendeby\
Department of Electrical Engineering\
Linköping University\
Linköping, Sweden\
`[email protected]` Fredrik Gustafsson\
Department of Electrical Engineering\
Linköping University\
Linköping, Sweden\
`[email protected]`
bibliography:
- 'Exploring-positive-noise-in-estimation-theory.bib'
title: Exploring positive noise in estimation theory
---
Introduction {#sec:introduction}
============
We consider the problem of estimating the mean $x$ observed in noise as $y_k=x+e_k$, for $k=1,2,\dots, N$, also known as “estimation of location" [@article:PI_kassam_85], where the noise $e_k$ has positive support. We will refer to such distributions as [*positive noise*]{}. Examples of distributions we will study include uniform, exponential, Rayleigh, Pareto.
A bias compensated linear estimator as the sample mean has a variance that decays as $1/N$, while it is well-known from the statistical literature, see for example [@book:ET_kay_93; @book:TPE_lehmann], that the minimum has a variance that decays as $1/N^2$. The minimum is the simplest example of order statistics. Certain care has to be taken for the cases where the parameters in the distributions are unknown, in which case bias compensation becomes tricky. This paper derives all combinations of known/unknown parameters for order statistics/BLUE (best linear unbiased estimator) for some selected and common distribution that allow for analytical solutions.
Problems involving positive noise can be motivated from applications where the arrival times of radio or sound waves are used. Such waves travel with the speed of the medium, and non line of sight conditions give rise to delayed arrival times. Physics does simply not allow for negative noise, only positive one. This case occur in a variety of applications such as target tracking using radar or lidar, and localisation using radio waves such as is done in for instance global satellite navigation systems [@article:ITVT_kok_15; @article:ITVT_chen_09; @article:ISPM_gustafsson_05; @article:IME_eling_12]. For example, the error histograms of time-of-arrival measurements collected from three separate cellular antennas are given in Figure \[fig:kista\]. For detailed description of hardware and the measurement campaign see [@conf:PIMRC_medbo_09]
To deal with the estimation’s performance degradation in non-Gaussian error conditions, conventional estimation techniques which are developed based on Gaussian assumptions need to be adjusted properly. As discussed in [@article:ITSP_yin_13], “identify and discard”, “mathematical programming”, and “robust estimation” are the three broad categories of estimation methods which are robust against non-Gaussian errors. Robustness of the estimator has been a concern for many years in both research [@article:JASA_stigler_73] and different engineering topics [@article:PI_kassam_85; @book:SDNGN_kassam; @article:SIAM_stewart_99; @book:NSP_arce] for a long time now. A more recent survey on this topic containing more references can be found in [@article:ISPM_zoubir_12].
The maximum likelihood estimator (MLE), developed under Gaussian assumptions, can be modified to become robust in presence of non-Gaussian noises. The authors in [@conf:ICML_eskin_00] first detect and then reject the outliers by learning the probability density function (PDF) of the measurements and develop a mixture model of outliers and clean data. A similar idea to k-nearest-neighbor approach is used in [@article:DMKD_chawla_10] to classify outliers as the data points that are less similar to a number of neighboring measurements. Surveys of advances in clustering the data into outliers and clean data can be found in [@article:AIR_hodge_04; @article:ITSP_yin_13; @conf:ICASSP_fritsche_09]. While these approaches might result in high estimation accuracy, they typically require large datasets [@article:ISPM_zoubir_12].
M-estimators [@book:RS_huber], in the contrary to identification-based methods, do not require pre-processing and can be used in non-Gaussian noise conditions. In principle, M-estimators can be seen as generalization of MLE and rely on solving a minimization problem of some loss function. For a detailed discussion on different loss functions, see [@book:RS_huber]. Since minimization problems are typically solved numerically based on the derivative of the loss function [@book:RSTM_maronna], they might converge to local minima.
In this work, we strive to find minimum variance unbiased (MVU) estimators for the location of estimation problems for non-Gaussian noise distributions where multiple distributions with positive support are considered. In case where MVU is not found, we introduce unbiased order-statistic-based estimators and compare their variances against the BLUE. The MVU estimators without any knowledge of the hyper parameters of the noise distributions are also derived, if possible. Finally, we derive an estimator for the case in which the noise follows a mixture of normal and uniform distribution. The rest of this paper is structured as follows. In Section \[sec:marginal\_distribution\_of\_order\_statistics\] the marginal distribution of order statistics is introduced. In Section \[sec:location\_estimation\_problem\] the location estimation problem is formulated. The problem is then investigated for different noise distributions and estimators for each distribution are derived in Sections \[sec:uniform\_distribution\]–\[sec:other\_distributions\]. The proposed estimators are evaluated in a simulation study in Section \[sec:performance\_evaluation\] followed by the concluding remarks given in Section \[sec:conclusions\].
Marginal Distribution of Order Statistics {#sec:marginal_distribution_of_order_statistics}
=========================================
The marginal distribution of order statistics, in this work is computed by differentiating the corresponding cumulative distribution function (CDF). In this section, we first introduce the minimum, also know as first or extreme, order statistic and then give the generalization to any statistics of order $k$. Let $F$ denote the common CDF of $N$ independent and identically distributed sample of random variables $y_1,\ldots,y_N$. We let $y^{(k)}$ denote the $k$:th order statistic of the sample, defined as the $k$:th smallest value of the set $\{y_i\}_{i=1}^N$. We define $f_{(k,N)}(y)$ as the marginal PDF of the $k$:th order statistics corresponding to a sample of size $N$. The PDF $f_{(k,N)}(y)$ is then calculated by differentiating $F_{(k,N)}(y)$ with respect to $y$.
Marginal distribution of minimum order statistic {#subsec:marginal_distribution}
------------------------------------------------
To further illustrate the problem, consider fist an example in which we have drawn $N=5$ independent random variables $\{y_i\}_{i=1}^5$ each from a common distribution with PDF $f(y)$. Assume that we are interested in the PDF of the first order statistic, $f_{(1,5)}(y)$. The CDF $F_{(1,5)}(y)$ is defined as $P(y^{(1)} < y)$. We note that the minimum order statistic $Y^{(1)}$ would be less than $y$ if at least $1$ of the random variables $y_1, y_2, y_3, y_4, y_5$ are less than $y$. In other words, we need to count the number of ways that can happen such that at least one random variable is less than $y$. This leads to a binomial probability calculation. The ’success’ is considered to be the event $\{y_i < y\}$, $i = 1$ and we let $\zeta$ denote the number of successes in five trials, then $$\begin{aligned}
F_{(1,5)}(y) &= P(y_{(1)}<y) = P(\zeta=1)+\ldots + P(\zeta=5),\\
f_{(1,5)}(y) &= \frac{\,d}{\,dy}F_{(1,5)}(y).\end{aligned}$$ To generalize the example, let $y_{(1)} < y_{(2)} < \ldots < y_{(N)}$ be the order statistics of $N$ independent observations from a continuous distribution with cumulative distribution function $F(y)$ and probability density function $f(y)=F'(y)$. The marginal PDF $f_{(1,N)}(y)$ of the minimum order statistic can be obtained by considering the event $\{Y_i \leq y\}, i = 1$ as a “success,” and letting $\zeta$ = the number of such successes in $N$ mutually independent trials. $\zeta$ is a binomial random variable with $N$ trials and probability of success $P(y_i\leq y)$. Hence, the CDF of the minimum order statistic is given by,
$$\begin{aligned}
F_{(1,N)}(y)=\sum_{n=1}^{N}P(\zeta=n).
\label{eq:cdf_order_minimum}
\end{aligned}$$
Noting that the probability mass function of this binomial distribution is given by, $$\begin{aligned}
P(\zeta=n) = \begin{pmatrix}N\\n\end{pmatrix}[F(y)]^n[1-F(y)]^{N-n}.
\label{eq:pmf_order_minimum}
\end{aligned}$$ Substituting into and taking the last term out of the sum, we get $$\begin{aligned}
F_{(1,N)}(y)=\sum_{n=1}^{N-1}\begin{pmatrix}N\\n\end{pmatrix}[F(y)]^n[1-F(y)]^{N-n}+[F(y)]^N.
\label{eq:cdf_order_minimum_2}
\end{aligned}$$
Differentiating with respect to $y$ gives a telescoping sum of the form, $$\begin{aligned}
f_{(1,N)}(y)&=\sum_{n=1}^{N-1}\frac{N!}{(n-1)!(N-n)!}[F(y)]^{n-1}f(y)[1-F(y)]^{N-n}\nonumber\\ &+ \sum_{n=1}^{N-1}\frac{N!}{n!(N-n-1)!}[F(y)]^n[1-F(y)]^{N-n-1}(-f(y)) \nonumber\\&+ N[F(y)]^{N-1}f(y),\end{aligned}$$ in which, except the first term, all other terms cancel each other out. Hence, the marginal probability density function of the minimum order statistic of a set of $N$ independent and identically random variables with common CDF $F(y)$ and PDF $f(y)$ is given by, $$\begin{aligned}
f_{(1,N)}(y) = Nf(y)\left(1-F(y)\right)^{N-1}.
\label{eq:density_order_minimum}\end{aligned}$$
Marginal distribution of general order statistic {#subsec:marginal_distribution_of_general_order_statistic}
------------------------------------------------
The marginal PDF $f_{(k,N)}(y)$ of the general order statistic $k$ can be obtained by generalizing the results of the previous section, and considering the event $\{y_i \leq y\}, i = 1, 2, \ldots, k$ as a “success,” and letting $\zeta$ = the number of such successes in $N$ mutually independent trials, $$\begin{aligned}
F_{(k,N)}(y)=\sum_{n=k}^{N-1}\begin{pmatrix}N\\n\end{pmatrix}[F(y)]^n[1-F(y)]^{N-n}+[F(y)]^N.
\label{eq:cdf_generic2}
\end{aligned}$$ Differentiating with respect to $y$ gives a telescoping sum of the form, $$\begin{aligned}
f_{(k,N)}(y)&=\sum_{n=k}^{N-1}\frac{N!}{(n-1)!(N-n)!}[F(y)]^{n-1}f(y)[1-F(y)]^{N-n}\nonumber\\ &+ \sum_{n=k}^{N-1}\frac{N!}{n!(N-n-1)!}[F(y)]^n[1-F(y)]^{N-n-1}(-f(y)) \nonumber\\&+ N[F(y)]^{N-1}f(y),\end{aligned}$$ in which, except the first term, all other terms cancel each other. Hence, the marginal probability density function of the $k$:th order statistic of a set of $N$ independent and identically random variables with common CDF $F(y)$ and PDF $f(y)$ is given by, $$\begin{aligned}
f_{(k,N)}(y) = Nf(y)\begin{pmatrix}N-1\\k-1\end{pmatrix}F(y)^{k-1}\left(1-F(y)\right)^{N-k}.
\label{eq:density_order}\end{aligned}$$
Location Estimation Problem {#sec:location_estimation_problem}
===========================
Consider the location estimation problem in which we have measurements $y_k$, $k=1,\ldots,N$ of the unknown parameter $x$. Assuming that the measurements are corrupted with additive noise $e_k\sim p_e(\theta)$, where $\theta$ denotes the parameter(s) of the noise distribution, the measurement model is given by $$\begin{aligned}
y_k=x+e_k, \quad k=1,\ldots,N.
\label{eq:generic_model}\end{aligned}$$ The BLUE for the estimation problem is given by $$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{p_e(\theta)}(y_{1:N},\theta) &= \frac{1}{N}\sum_{k=1}^N y_k- \delta(\theta),
\label{eq:mean_estimator}\end{aligned}$$ where $y_{1:N} = \{y_k\}_{k=1}^{N}$ and $\delta(\theta)=\mathbb{E}(e_k)$ is the bias compensation term.
In the following sections, closed-form expressions for the mean squared error (MSE) of the BLUE estimator for multiple noise distributions with positive support are provided. Given hyperparameter $\theta$, the MVU estimator for each noise distribution $p_e$ is denoted by $\hat{x}_{\mathrm{MVU}}^{p_e}(y_{1:N},\theta)$. MVU estimators with unknown hyperparameter are denoted by $\hat{x}_{\mathrm{MVU}}^{p_e}(y_{1:N})$. If the MVU cannot be found, an unbiased order-statistics-based estimator is derived and denoted by $\hat{x}^{p_e}(y_{1:N},\theta)$ and $\hat{x}^{p_e}(y_{1:N})$ for known and unknown hyperparameter cases, respectively. For example, $\hat{x}^{\mathcal{U}}_{\mathrm{MVU}}(y_{1:N},\beta)$ denotes the MVU estimator when $e_k\sim\mathcal{U}[0,\beta]$ and $\beta$ is known. $\hat{x}^{\mathcal{U}}_{\mathrm{MVU}}(y_{1:N})$, on the other hand, corresponds to the MVU estimator of uniform noise with unknown hyper parameters of the distribution. Table \[tbl:notation\] summarizes the notation used throughout this work.
--------------------------------------------------------- ------------------------------------------------------------
$\{y_k\}_{k=1}^N$ noisy measurements of the unknown parameter $x$
\[2mm\] $\left(y^{(m)}\right)_{m=1}^N$ ordered measurement sequence
\[2mm\] $\theta$ parameters of the noise distribution
\[2mm\] $\delta(\theta)$ bias compensation term
\[2mm\] $\hat{x}^{p_e}_{\mathrm{BLUE}}(y_{1:N},\theta)$ BLUE when $e_k\sim p_e$ for known $\theta$
\[2mm\] $\hat{x}^{p_e}_{\mathrm{MVU}}(y_{1:N},\theta)$ MVU estimator when $e_k\sim p_e$ for known $\theta$
\[2mm\] $\hat{x}^{p_e}_{\mathrm{MVU}}(y_{1:N})$ MVU estimator when $e_k\sim p_e$ for unknown $\theta$
\[2mm\] $\hat{x}^{p_e}(y_{1:N},\theta)$ unbiased estimator when $e_k\sim p_e$ for known $\theta$
\[2mm\] $\hat{x}^{p_e}(y_{1:N})$ unbiased estimator when $e_k\sim p_e$ for unknown $\theta$
--------------------------------------------------------- ------------------------------------------------------------
: Notation.
\[tbl:notation\]
For each noise distribution, we also consider the minimum order statistic estimator, denoted by $\hat{x}^{p_e}_{\mathrm{min}}(y_{1:N})$. Let $\left(y^{(m)}_{1:N}\right)_{m=1}^{N}$ denote the ordered sequence obtained from sorting $y_{1:N}$ in an ascending order, $\hat{x}^{p_e}_{\mathrm{min}}(y_{1:N})$ is defined as $$\begin{aligned}
\hat{x}^{p_e}_{\mathrm{min}}(y_{1:N}) = y_{1:N}^{(1)} \triangleq \min_{k} y_k.\end{aligned}$$ Noting that for any generic estimator $\hat{x}$ the MSE is given by $$\begin{aligned}
\label{eq:mse_generic}
{\mathrm{MSE}}(\hat{x}) = {\mathrm{Var}}(\hat{x}) + b^2(\hat{x}).\end{aligned}$$ The MSE for BLUE and MVU or any other bias compensated estimator coincides with the estimator’s variance. In case of $\hat{x}^{p_e}_{\mathrm{min}}$, the existing bias enters the MSE.
In order to find the MVU estimator, the first step is to find the PDF $f(y_{1:N};\theta)$ with $\theta$ denoting the parameters of the distribution. If the PDF satisfies regularity conditions, the CRLB can be determined. Any unbiased estimator that satisfies CRLB is thus the MVU estimator. However, the considered PDFs do not satisfy the regularity conditions, $$\begin{aligned}
{\mathbb{E}}\left[\frac{\partial\ln f(y_k;\theta)}{\partial \theta}\right]\neq0.\end{aligned}$$ Hence, the CRLB approach is not applicable. Instead, we rely on the RBLS theorem [@article:IJS_lehmann_1; @article:IJS_lehmann_2; @book:ET_kay_93], to find the MVU estimator. The RBLS theorem [@article:IJS_lehmann_1] states that for any unbiased estimator $\tilde{x}$ and sufficient statistics $T(y_{1:N})$, $\hat{x}={\mathbb{E}}(\tilde{x}\mid T(y_{1:N}))$ is unbiased and ${\mathrm{Var}}(\hat{x})\leq{\mathrm{Var}}(\tilde{x})$. Additionally, if $T(y_{1:N})$ is complete, then $\hat{x}$ is MVU.
As shown in [@book:ET_kay_93], if the dimension of the sufficient statistics is equal to the dimension of the parameter, then the MVU estimator is given by $\hat{x}=g(T(y_{1:N}))$ for any function $g(\cdot)$ that satisfies $$\begin{aligned}
{\mathbb{E}}(g(T)) = x.\end{aligned}$$ Hence, the problem of MVU estimator turns into the problem of finding a complete sufficient statistic. The Neyman-Fisher theorem [@article:fisher_22; @article:AMS_halmos_49] gives the sufficient statistic $T(y_{1:N})$, if the PDF can be factorized as follows $$\begin{aligned}
f(y_{1:N};\Psi) =g(T(y_{1:N}),\Psi)h(y_{1:N}),\end{aligned}$$ where $\psi$ is the union of the noise hyper parameter(s) $\theta$ and $x$. The estimators in this work are derived in the order statistics framework.
Uniform Distribution {#sec:uniform_distribution}
====================
As the first scenario, consider the case in which the additive noise $e_k$ has a uniform distribution with a positive support, $p_e(\theta)=\mathcal{U}[0,\beta]$, $\beta>0$ and $\theta=\beta$. The BLUE is given by
$$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{\mathcal{U}}(y_{1:N},\beta) &= \frac{1}{N}\sum_{k=1}^N y_k- \frac{\beta}{2}.
\label{eq:sample_mean_example}
\end{aligned}$$
The MSE of BLUE for this case is given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{BLUE}}^{\mathcal{U}}(y_{1:N},\beta)\right) &= \frac{1}{N^2}\sum_{k=1}^{N}{\mathrm{Var}}\left(y_k-\frac{\beta}{2}\right)= \frac{\beta^2}{12N}.
\label{eq:mse_blue_uniform}
\end{aligned}$$
In order to find the MSE of the minimum order statistics estimator, $\hat{x}_{\mathrm{min}}^{\mathcal{U}}(y_{1:N}) $, we need to find the first two moments of the estimator. Let $\tilde{y}_k= \frac{1}{\beta}y_k$. Since $y_k\sim\mathcal{U}[x,x+b]$, then for any constant $\beta>0$, $\tilde{y_k}\sim\mathcal{U}[\frac{x}{\beta},\frac{x}{\beta}+1]$. Hence, $f(\tilde{y}_k)=1$ and $F(\tilde{y}_k)=\frac{1}{\beta}(y_k-x)$. From we get,
$$\begin{aligned}
f&_{(k,N)}^{\mathcal{U}[0,\beta]}(\tilde{y}) = N\begin{pmatrix}N-1\\k-1\end{pmatrix}\left(\frac{\tilde{y}-x}{\beta}\right)^{k-1}\left(\frac{\beta-(\tilde{y}-x)}{\beta}\right)^{N-k}\nonumber\\
&= \frac{(N)!}{(k-1)!(N-k)!}\left(\frac{\tilde{y}-x}{\beta}\right)^{k-1}\left(\frac{\beta-(\tilde{y}-x)}{\beta}\right)^{N-k}.
\end{aligned}$$
since $N\in\mathbb{N}^+$, $k\in\mathbb{N}^+$, and $k\in[1,N]$ we can the change the factorials to gamma functions, $$\begin{aligned}
f_{(k,N)}^{\mathcal{U}[0,\beta]}(\tilde{y}) = \frac{\Gamma(N+1)}{\Gamma(k)\Gamma(N-k+1)} \left(\frac{\tilde{y}-x}{\beta}\right)^{k-1}\left(\frac{\beta-(\tilde{y}-x)}{\beta}\right)^{N-k}.
\label{eq:order_uniform_general}
\end{aligned}$$
The marginal distribution is a generalized beta distribution, also known as four parameters beta distribution [@article:Mcdonal_JE_95]. The support of this distribution is from $0$ to $\beta>0$ and $f_{(k,N)}^{\mathcal{U}[0,\beta]}(\cdot)=\frac{1}{\beta}f_{(k,N)}^{\mathcal{U}[0,1]}(\cdot)$. The bias and variance of the general $k$:th order statistic estimator $\hat{x}_{(k)}^{\mathcal{U}}(y_{1:N})$ in case of uniform noise with support on $[0,\beta]$ are given by
\[eq:moments\_uniform\_order\] $$\begin{aligned}
{\mathbb{E}}(\hat{x}_{(k)}^{\mathcal{U}}(y_{1:N})) &= \frac{\beta k}{N+1},\\
{\mathrm{Var}}(\hat{x}_{(k)}^{\mathcal{U}}(y_{1:N})) &= \frac{k(N-k+1)\beta^2}{(N+1)^2(N+2)}.
\end{aligned}$$
The first two moments of the minimum order statistic estimator are obtained by letting $k=1$ in
\[eq:moments\_uniform\_min\] $$\begin{aligned}
b\left(\hat{x}_{\mathrm{min}}^{\mathcal{U}}(y_{1:N})\right) &= \frac{\beta}{N+1}\\
{\mathrm{Var}}\left(\hat{x}_{\mathrm{min}}^{\mathcal{U}}(y_{1:N})\right) &= \frac{N\beta^2}{(N+1)^2(N+2)}.
\end{aligned}$$
The MSE of $\hat{x}_{\mathrm{min}}^{\mathcal{U}}(y_{1:N})$ is then given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{min}}^{\mathcal{U}}(y_{1:N})\right) = \frac{2\beta^2}{(N+1)(N+2)}.
\label{eq:mse_min_uniform}\end{aligned}$$
MVU estimator {#subsec:mvu_estimator}
-------------
In order to find the MVU estimator, we note that the PDF can be written in a compact form using the step function $\sigma(\cdot)$ as
$$\begin{aligned}
f(y_k;x,\beta) = \frac{1}{\beta}\left[\sigma(y_k-x) - \sigma(y_k-x-\beta)\right].
\label{eq:uniform_pdf_single}
\end{aligned}$$
which gives $$\begin{aligned}
f(y_{1:N};&x,\beta) = \frac{1}{\beta^N}\prod_{k=1}^{N}\left[\sigma(y_k-x) - \sigma(y_k-x-\beta)\right]\nonumber\\
&=\frac{1}{\beta^N}\left[\sigma( y^{(1)}_{1:N} -x) - \sigma(y^{(N)}_{1:N}-x -\beta)\right], \label{eq:unknown}
\end{aligned}$$
where $y^{(N)}_{1:N}\triangleq\max_k y_k,\quad k=1,\ldots,N$. The expressions for the MVU estimator is derived for two different scenarios. We first assume that the hyper parameter $\beta$ of the noise distribution is known and then further discuss the unknown hyper parameter case. In the general case, let $\Psi = [x,\beta]^\top$ denote the unknown parameter vector, the Neyman-Fisher factorization gives $h(y_{1:N})=1$ and $$\begin{aligned}
T(y_{1:N}) = \begin{bmatrix}
y^{(1)}_{1:N} \\ y^{(N)}_{1:N}
\end{bmatrix} = \begin{bmatrix}
T_1(y_{1:N}) \\ T_2(y_{1:N})
\end{bmatrix}.
\label{eq:uniform_ss}\end{aligned}$$
### Known hyper parameter $\beta$ {#subsubsec:known_hyper_parameter_uniform}
When the maximum support of the uniform noise $\beta$ is known, the dimensionality of the sufficient statistic is larger than that of the unknown parameter $x$. As discussed in [@book:ET_kay_93], the RBLS theorem can be extended to address this case if the form of a function $g(T_1(y_{1:N}),T_2(y_{1:N}))$ can be found that combines $T_1$ and $T_2$ into a single unbiased estimator of $x$.
Let $Z = T_1(y_{1:N})+T_2(y_{1:N})=u+v$. Since $T_1$ and $T_2$ are dependent,
$$\begin{aligned}
\label{eq:dist_sum_gen}
f_Z(z) = \int_{-\infty}^{\infty} f_{y^{(1)},y^{(N)}}(u,z-u)\,d_u,
\end{aligned}$$
where $f_{y^{(1)},y^{(N)}}(u,z-u)$ is the joint density of minimum and maximum order statistics. As shown in [@book:OS_david_04], for $-\infty<u<v<\infty$, the joint density of two order statistics $y^{(i)}$ and $y^{(j)}$ is given by $$\begin{aligned}
f_{y^{(i)},y^{(j)}}(u,v) = &\frac{N!}{(i-1)!(j-1-i)!(N-j)!}\nonumber\\&\times f_Y(u)f_Y(v)\left[F_Y(u)\right]^{i-1}\nonumber\\
&\times\left[F_Y(v)-F_Y(u)\right]^{j-1-i}\left[1-F_Y(v)\right]^{N-j},
\end{aligned}$$ that for the extreme orders, $i=1$ and $j=N$ can be simplified such that for $u<v$ $$\begin{aligned}
\label{eq:order_joint_gen}
f_{y^{(1)},y^{(N)}}(u,v) = N(N-1)\left[F_Y(v)-F_Y(u)\right]^{N-2} f_Y(u)f_Y(v).
\end{aligned}$$ and zero otherwise. Substituting into , we get $$\begin{aligned}
f_Z(z) = \frac{1}{2}N\beta^{-N}(2x+2\beta-z)^{N-1},
\end{aligned}$$ for $2x+\beta<z<2(x+\beta)$ and $$\begin{aligned}
f_Z(z) = -\frac{1}{2}N\beta^{-N}\frac{(z-2x)^{N}}{2x -z},
\end{aligned}$$ for $2x<z\leq2x+\beta$ and zero otherwise. It can be shown that $$\begin{aligned}
{\mathbb{E}}(f_Z(z)) = 2x + \beta.
\end{aligned}$$
Hence, noting that $\beta$ is known, the function $g(T_1(y_{1:N}),T_2(y_{1:N}))$ that gives an unbiased estimator should be of the form of
$$\begin{aligned}
\hat{x}_{\mathrm{MVU}}^{\mathcal{U}}(y_{1:N},\beta) &= g(T_1(y_{1:N}),T_2(y_{1:N}))\nonumber\\
&= \frac{1}{2}(y^{(1)}_{1:N}+ y^{(N)}_{1:N}) - \frac{\beta}{2}.
\end{aligned}$$
The MSE of the MVU estimator is given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{MVU}}^{\mathcal{U}}(y_{1:N},\beta) \right) = \frac{\beta^2}{2N(N+3)+4}.\label{eq:mse_mvu_unif_known}
\end{aligned}$$
Comparing to , the order statistics based MVU estimator outperforms the BLUE one order of magnitude.
### Unknown hyper parameter $\beta$ {#subsec:unknown_hyper_parameter_uniform}
In this case, the MVU estimators for the parameter vector $\Psi=[x,\beta]^\top$ can be derived from sufficient statistics , $$\begin{aligned}
\hat{\Psi}=g(T(y_{1:N})),\quad \mathrm{s.t.}\quad {\mathbb{E}}\left(g\left(T(y_{1:N})\right)\right) = \Psi.\end{aligned}$$ In this case, we have $$\begin{aligned}
{\mathbb{E}}(T(y_{1:N})) = \begin{bmatrix}
x+\frac{\beta}{N+1}\\\\x+\frac{N\beta}{N+1}
\end{bmatrix}
\label{eq:uniform_unknown_E_gen}\end{aligned}$$ To find the transformation that makes unbiased, we define
$$\begin{aligned}
g(T(y_{1:N}))=\begin{bmatrix}\frac{1}{N-1}\left(NT_1(y_{1:N})-T_2(y_{1:N})\right)\\\\ \frac{N+1}{N-1}\left(T_2(y_{1:N})-T_1(y_{1:N})\right)\end{bmatrix}
\end{aligned}$$
that gives $$\begin{aligned}
{\mathbb{E}}\left(g(T(y_{1:N}))\right) = \begin{bmatrix}x\\\beta\end{bmatrix}.
\end{aligned}$$
Finally, the MVU estimator of $x$ when the hyper parameter $\beta$ is unknown is given by
$$\begin{aligned}
\hat{x}_{\mathrm{MVU}}^{\mathcal{U}}(y_{1:N}) = \frac{N}{N-1}y^{(1)}_{1:N} - \frac{1}{N-1} y^{(N)}_{1:N}.
\end{aligned}$$
and its MSE is $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{MVU}}^{\mathcal{U}}(y_{1:N})\right) = \frac{N\beta^2}{(N+2)(N^2-1)}.
\end{aligned}$$
This is naturally slightly larger than for finite $N$.
Distributions in the exponential family {#sec:distributioons_in_the_exponential_family}
=======================================
The exponential family of probability distributions, in their most general form, is defined by $$\begin{aligned}
f(y;\theta) = h(y)g(\theta)\exp\left\{A(\theta)\cdot T(y)\right\},\end{aligned}$$ where $\theta$ is the parameter of the distribution, and $h(y)$, $g(\theta)$, $A(\theta)$, and $T(y)$ are all known functions. In this section, we only consider some example distributions of this family and show that the minimum order statistic estimator gets the same form of distribution as the noise distribution but with modified parameters. For the selected distributions, if possible, MVU estimators for both cases of known and unknown hyperparameter are derived. Otherwise, unbiased estimators with less variance than BLUE are proposed.
Exponential distribution {#subsec:exponential_distribution}
------------------------
Exponential distributions are members of the gamma family with shape parameter 1; strongly skewed with no left sided tail ($y_k\in[x,\infty]$). Let $\beta>0$ denote the scale parameter, the PDF of an exponential distribution is then given by
\[eq:exponential\] $$\begin{aligned}
f^{\mathrm{Exp}}(y_k;x,\beta)=\left\{\begin{matrix}
\frac{1}{\beta}\exp(-\frac{(y_k-x)}{\beta})&y_k\geq x, \\
0&y_k< x.
\end{matrix}\right.
\label{eq:exponential_pdf}
\end{aligned}$$ and the CDF, for $y\geq x$, is given by $$\begin{aligned}
F^{\mathrm{Exp}}(y_k;x,\beta)=1-\exp(-\frac{(y_k-x)}{\beta}).
\label{eq:exponential_cdf}
\end{aligned}$$
For the BLUE estimator , from the properties of exponential distribution, we have $$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{\mathrm{Exp}}(y_{1:N},\beta) &= \frac{1}{N}\sum_{k=1}^{N}y_k - \beta,
&\quad
{\mathrm{MSE}}(\hat{x}_{\mathrm{BLUE}}^{\mathrm{Exp}}) &= \frac{\beta^2}{N}.\end{aligned}$$ Substituting into , the marginal density of the $k$:th order statistic is given by $$\begin{aligned}
f^{\mathrm{Exp}}_{(k,N)}(y;x,\beta) = \frac{N}{\beta}\begin{pmatrix}N-1\\k-1\end{pmatrix}\left(1-\exp(-\frac{(y-x)}{\beta})\right)^{k-1}\exp\left(-\frac{(N-k+1)(y-x)}{\beta}\right).
\label{eq:exponential_order}\end{aligned}$$ The first order statistic density is then given by letting $k=1$ in that results in another exponential distribution, $$\begin{aligned}
f^{\mathrm{Exp}}_{(1,N)}(y;x,\beta) = f^{\mathrm{Exp}}(y;x,\bar{\beta}),\end{aligned}$$ where $\bar{\beta}=\frac{\beta}{N}$. Hence, the MSE of the minimum order statistics estimator is given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{min}}^{\mathrm{Exp}}(y_{1:N})\right) = \frac{2\beta^2}{N^2}.\end{aligned}$$ In order to find the MVU estimator, we re-write the PDF as $$\begin{aligned}
f(y_{1:N};x,\beta) &= \frac{1}{\beta^N}\exp\left[-\frac{1}{\beta}\sum_{k=1}^{N}y_k\right]\exp\left[-\frac{N}{\beta}x\right]\times\sigma(y^{(1)}_{1:N} - x).
\label{eq:exponential_pdf_2}\end{aligned}$$
### Known hyper parameter $\beta$ {#known-hyper-parameter-beta}
In case of the known hyper parameter $\beta$, the Neyman-Fisher factorization of PDF gives
$$\begin{aligned}
T(y_{1:N}) &= y^{(1)}_{1:N}\\
h(y_{1:N}) &= \frac{1}{\beta^N}\exp\left[-\frac{1}{\beta}\sum_{k=1}^{N}y_k\right].
\label{eq:exponential_known_ss}
\end{aligned}$$
The MVU estimator can then be obtained from a transformation of the minimum order statistic that makes it an unbiased estimator. Finally, in case of exponential noise with known hyper parameter of the distribution, the MVU estimator and its MSE are given by
$$\begin{aligned}
\hat{x}_{MVU}^{\mathrm{Exp}}(y_{1:N},\beta) &= y^{(1)}_{1:N} - \frac{\beta}{N}
\\
{\mathrm{MSE}}\left(\hat{x}_{MVU}^{\mathrm{Exp}}(y_{1:N},\beta)\right) &= \frac{\beta^2}{N^2}.
\end{aligned}$$
### Unknown hyper parameter $\beta$ {#unknown-hyper-parameter-beta}
If the hyper parameter $\beta$ is unknown, the factorization gives
$$\begin{aligned}
T(y_{1:N}) = \begin{bmatrix}
y^{(1)}_{1:N}\\ \sum_{k=1}^{N} y_k
\end{bmatrix} = \begin{bmatrix}
T_1(y_{1:N}) \\ T_2(y_{1:N})
\end{bmatrix}.
\label{eq:uniform_unknown_ss}
\end{aligned}$$
Noting that sum of exponential random variables results in a Gamma distribution, we have $T_2(y_{1:N})\sim\Gamma(N,\beta)$. Hence, $$\begin{aligned}
{\mathbb{E}}(T(y_{1:N})) = \begin{bmatrix}
x+\frac{\beta}{N}\\\\N(x+\beta)
\end{bmatrix}.
\label{eq:exponential_unknown_E_gen}
\end{aligned}$$
Following the same line of reasoning as in Section \[subsec:unknown\_hyper\_parameter\_uniform\], the unbiased estimator is given by the transformation
$$\begin{aligned}
g(T(y_{1:N}))=\begin{bmatrix}\frac{1}{N-1}\left(NT_1(y_{1:N})-\frac{1}{N}T_2(y_{1:N})\right)\\\\ \frac{1}{N-1}\left(T_2(y_{1:N})-NT_1(y_{1:N})\right)\end{bmatrix},
\end{aligned}$$
that gives $$\begin{aligned}
{\mathbb{E}}\left(g(T(y_{1:N}))\right) = \begin{bmatrix}x\\\beta \end{bmatrix}.
\end{aligned}$$
Finally, the MVU estimator when the hyper parameter $\beta$ is unknown, is given by
$$\begin{aligned}
\hat{x}_{\mathrm{MVU}}^{\mathrm{Exp}}(y_{1:N}) &= \frac{N}{N-1}y^{(1)}_{1:N} - \frac{1}{N(N-1)} \sum_{k=1}^{N} y_k \nonumber\\&= \frac{N}{N-1}y^{(1)}_{1:N}- \frac{1}{N-1} \bar{y},
\end{aligned}$$
where $\bar{y}$ is the sample mean. Assuming that $N$ is large $\min_k y_k$ and $\bar{y}$ are independent and the MSE of the estimator, asymptotically, is given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{MVU}}^{\mathrm{Exp}}(y_{1:N})\right) = \frac{\beta^2(N+1)}{N(N-1)^2}.
\end{aligned}$$
Rayleigh distribution {#subsec:rayleigh_distribution}
---------------------
One generalization of the exponential distribution is obtained by parameterizing in terms of both a scale parameter $\beta$ and a shape parameter $\alpha$. Rayleigh distribution is a special case obtained by setting $\alpha=2$
\[eq:rayleigh\] $$\begin{aligned}
f^{\mathrm{Rayleigh}}(y_k;x,\beta)=\left\{\begin{matrix}
\frac{y_k-x}{\beta^2}\exp(-\frac{(y_k-x)^2}{2\beta^2})&y_k> x, \\
0&y_k\leq x.
\end{matrix}\right.
\label{eq:rayleigh_pdf}
\end{aligned}$$ and the CDF, for $y_k>x$ is given by $$\begin{aligned}
F^{\mathrm{Rayleigh}}(y_k;x,\beta)=1-\exp(-\frac{(y_k-x)^2}{2\beta^2}).
\label{eq:rayleigh_cdf}
\end{aligned}$$
Hence, the BLUE estimator , becomes
$$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{\mathrm{Rayleigh}}(y_{1:N},\beta) &= \frac{1}{N}\sum_{k=1}^{N}y_k - \sqrt{\frac{\pi}{2}}\beta,
\\
{\mathrm{MSE}}\left(\hat{x}^{\mathrm{Rayleigh}}_{\mathrm{BLUE}}(y_{1:N},\beta)\right) &= \frac{(4-\pi)\beta^2}{2N}.
\end{aligned}$$
The marginal density of the $k$:th order statistic is given by $$\begin{aligned}
f^{\mathrm{Rayleigh}}_{(k,N)}(y;x,\beta) =\left\{\begin{matrix}
\frac{Ny}{\beta^2}\begin{pmatrix}N-1\\k-1\end{pmatrix}\left(1-\exp(-\frac{(y-x)^2}{2\beta^2})\right)^{k-1}\exp\left(-\frac{(N-k+1)(y-x)^2}{2\beta^2}\right)&y> x, \\
0&y\leq x.
\end{matrix}\right.
\label{eq:rayleigh_order}\end{aligned}$$ Hence, the minimum order statistics density also is Rayleigh distributed $$\begin{aligned}
f^{\mathrm{Rayleigh}}_{(1,N)}(y;x,\beta) = f^{\mathrm{Rayleigh}}(y;x,\bar{\beta}),\end{aligned}$$ where $\bar{\beta}=\frac{\beta}{\sqrt{N}}$. The MSE of the minimum order statistics is given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}^{\mathrm{Rayleigh}}_{\mathrm{min}}(y_{1:N})\right) = \frac{2\beta^2}{N}.\end{aligned}$$ The joint PDF of $N$ independent observations $y_{1:N}$ is given by
\[eq:rayleigh\_pdf\_2\] $$\begin{aligned}
f(y_{1:N};x,\beta) &= \frac{\prod_{k=1}^{N}(y_k-x)}{\beta^{2N}}\exp\left[\sum_{k=1}^{N}-\frac{(y_k-x)^2}{2\beta^2}\right]\sigma(y^{(1)}_{1:N}- x).
\label{eq:rayleigh_pdf_2_1}
\end{aligned}$$ Noting that $$\begin{aligned}
\sum_{k=1}^{N}(y_k-x)^2 = \sum_{k+1}^{N}(y_k)^2 - 2x\sum_{k=1}^{N}y_k+Nx^2,
\label{eq:rayleigh_pdf_2_2}
\end{aligned}$$ the PDF becomes $$\begin{aligned}
f(y_{1:N};x,\beta) &= \beta^{-2N}\prod_{k=1}^{N}(y_k-x)\exp\left[\frac{-1}{2\beta^2}\sum_{k=1}^{N}y_k^2\right]\nonumber\\&\times\exp\left[-\frac{Nx^2}{2\beta^2}\right]\exp\left[\frac{x}{\beta^2}\sum_{k=1}^{N}y_k\right]\sigma(y^{(1)}_{1:N} - x).
\label{eq:rayleigh_pdf_2_3}
\end{aligned}$$
### Known hyper parameter $\beta$ {#known-hyper-parameter-beta-1}
Since cannot be factorized in the form of $f(y_{1:N};x,\beta) = g(T(y_{1:N}),x)h(y_{1:N})$, the RBLS theorem cannot be used. Hence, even if an MVU estimator exists for this problem, we may not be able to find it. Thus, in case of Rayleigh-distributed measurement noise, we propose unbiased estimators based on order statistics.
If the hyper parameter of the distribution is known, the unbiased order statistic based estimator $\hat{x}^{\mathrm{Rayleigh}}(y_{1:N},\beta)$ is then given by,
$$\begin{aligned}
\hat{x}^{\mathrm{Rayleigh}}(y_{1:N},\beta) &= y^{(1)}_{1:N} - \frac{\sqrt{\pi}\beta}{\sqrt{2N}},
\\
{\mathrm{MSE}}\left(\hat{x}^{\mathrm{Rayleigh}}(y_{1:N},\beta)\right) &=\frac{(4-\pi)\beta^2}{2N}.
\end{aligned}$$
which has the same variance as the BLUE estimator.
### Unknown hyper parameter $\beta$ {#unknown-hyper-parameter-beta-1}
In case of unknown hyper parameters, as for the known case, no factorization that enables us to use the RBLS theorem can be found. In this case, we propose the following unbiased estimator $$\begin{aligned}
\hat{x}^{\mathrm{Rayleigh}}(y_{1:N}) &= \frac{\sqrt{N}}{\sqrt{N}-1}y^{(1)}_{1:N} - \frac{1}{N(\sqrt{N}-1)}\sum_{k=1}^{N}y_k\nonumber\\&= \frac{1}{\sqrt N-1}(\sqrt{N}y^{(1)}_{1:N} - \bar{y}).\end{aligned}$$ Asymptotically, for large $N$, the sample mean and minimum order statistic are independent and the estimator MSE is given by $$\begin{aligned}
{\mathrm{MSE}}\left(\hat{x}^{\mathrm{Rayleigh}}(y_{1:N})\right)=\frac{(1+N)(4-\pi)\beta^2}{2N(\sqrt{N}-1)^2}.
\label{eq:minimum_orde_unknown_rayleigh}\end{aligned}$$
Weibull distribution {#subsec:weibull_distribution}
--------------------
Weibull distribution is a generalization of the Rayleigh, distribution that is parameterized by two parameters–scale parameter $\beta$ and shape parameter $\alpha>0$. In fact Weibull distribution is obtained by relaxing the assumption $\alpha=2$ in the Rayleigh distribution and its density function is given by
\[eq:weibull\] $$\begin{aligned}
f^{\mathrm{Weibull}}(y_k;x,\beta,\alpha)=\left\{\begin{matrix}
\frac{\alpha}{\beta}\left(\frac{y_k-x}{\beta}\right)^{\alpha-1}\exp(-(\frac{y_k-x}{\beta})^\alpha)&y_k> x, \\
0&y_k\leq x.
\end{matrix}\right.
\label{eq:weibull_pdf}
\end{aligned}$$ and the CDF, for $y_k\geq x$ is given by $$\begin{aligned}
F^{\mathrm{Weibull}}(y_k;x,\beta,\alpha)=1-\exp(-(\frac{y_k-x}{\beta})^\alpha).
\label{eq:weibull_cdf}
\end{aligned}$$ The BLUE estimator, in case of Weibull-distributed measurement noises is given by $$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{\mathrm{Weibull}}(y_{1:N},\beta,\alpha) &= \frac{1}{N}\sum_{k=1}^{N}y_k - \beta\Gamma(1+\frac{1}{\alpha}),
&\quad
{\mathrm{MSE}}(\hat{x}_{\mathrm{BLUE}}^{\mathrm{Weibull}}(y_{1:N},\beta,\alpha)) &= \frac{\beta^2}{N}\left[\Gamma(\frac{\alpha+2}{\alpha})-\left(\Gamma(\frac{\alpha+1}{\alpha})\right)^2\right].\end{aligned}$$
The marginal density of the $k$:th order statistic is given by $$\begin{aligned}
f^{\mathrm{Weibull}}_{(k,N)}(y;x,\beta,\alpha) = \frac{N\alpha}{\beta}\begin{pmatrix}N-1\\k-1\end{pmatrix}(\frac{y-x}{\beta})^{\alpha-1}\left(1-\exp^{-(\frac{y-x}{\beta})^\alpha}\right)^{k-1}\exp\left(-(N-k+1)(\frac{y-x}{\beta})^\alpha\right).
\label{eq:weibull_order}\end{aligned}$$ Hence, the first order statistic density in case of $e_k\sim\mathrm{Weibull}(\beta,\alpha)$, is another Weibull distribution, $$\begin{aligned}
f^{\mathrm{Weibull}}_{(1,N)}(y;x,\beta,\alpha) = f^{\mathrm{Weibull}}(y;x,\bar{\beta},\alpha),\end{aligned}$$ where $\bar{\beta}=\sqrt[-\alpha]{N}\beta$. This gives the MSE of the minimum order statistic estimator as $$\begin{aligned}
{\mathrm{MSE}}(\hat{x}_{\mathrm{min}}^{\mathrm{Weibull}}(y_{1:N})) = \beta^2N^{\frac{-2}{\alpha}}\Gamma(\frac{\alpha+2}{\alpha})\end{aligned}$$ Given $N$ independent observations, the joint density is given by $$\begin{aligned}
f^{\mathrm{Weibull}}(y_{1:N};x,\beta,\alpha)=(\frac{\alpha}{\beta})^N\prod_{k=1}^{N}\left(\frac{y_k-x}{\beta}\right)^{\alpha-1}\exp(-\sum_{k=1}^{N}(\frac{y_k-x}{\beta})^\alpha)\sigma(\min y_k - x)
\label{eq:weibull_pdf_joint}\end{aligned}$$ Since cannot be factorized using Neyman-Fisher factorization, RBLS is not applicable. Additionally, in this case, it is not possible to find an unbiased estimator when the hyper parameters $\alpha$ and $\beta$ are unknown. In case of known hyper parameters, the unbiased minimum order statistic estimator, however, can be computed. The unbiased estimator based on minimum order statistic is given by, $$\begin{aligned}
\hat{x}^{\mathrm{Weibull}}(y_{1:N},\beta,\alpha) &= y^{(1)}_{1:N}- \beta N^{-\frac{1}{\alpha}}\Gamma(1+\frac{1}{\alpha}),
&\nonumber\\
{\mathrm{MSE}}(\hat{x}^{\mathrm{Weibull}}(y_{1:N},\beta,\alpha)) &=\beta^2N^{\frac{-2}{\alpha}}\left[\Gamma(\frac{\alpha+2}{\alpha})-\left(\Gamma(\frac{\alpha+1}{\alpha})\right)^2\right].
\end{aligned}$$ An order-statistics-based unbiased estimator with unknown hyper parameters of the distribution could not be obtained.
Other Distributions {#sec:other_distributions}
===================
In this section, we further study the location estimation problem for two other noise distributions. In the rest, the Pareto distribution with positive support is first studied followed by the mixture of uniform and normal distribution.
Pareto distribution
-------------------
Let the scale parameter $\beta$ (necessarily positive) denote the minimum possible value of $y_k$, and $\alpha>0$ denote the shape parameter. The Pareto Type I distribution is characterized by $\beta$ and $\alpha$
\[eq:pareto\] $$\begin{aligned}
f^{\mathrm{Pareto}}(y_k;x,\beta,\alpha)=\left\{\begin{matrix}
\alpha\beta^\alpha (y_k-x)^{-(\alpha+1)}&y_k\geq x+\beta, \\
0&y_k< x+\beta.
\end{matrix}\right.
\label{eq:pareto_pdf}
\end{aligned}$$ and the CDF is given by $$\begin{aligned}
F^{\mathrm{Pareto}}(y_k;x,\beta,\alpha)=1-\left(\frac{\beta}{y-x}\right)^\alpha.
\label{eq:pareto_cdf}
\end{aligned}$$
For the BLUE we get,
$$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{\mathrm{Pareto}}(y_{1:N},\beta,\alpha) &= \frac{1}{N}\sum_{k=1}^{N}y_k - \frac{\alpha\beta}{\alpha-1},&\quad \alpha>1
\\
{\mathrm{MSE}}(\hat{x}_{\mathrm{BLUE}}^{\mathrm{Pareto}}(y_{1:N},\beta,\alpha)) &=\frac{\alpha\beta^2}{N(\alpha-1)^2(\alpha-2)},&\quad \alpha>2
\end{aligned}$$
The RBLS theorem cannot be used in case Pareto-distributed noises. We provide an unbiased estimator using minimum order statistics and its variance. The marginal density of the $k^{\mathrm{th}}$ order statistic, for $y\geq x+\beta$ is given by $$\begin{aligned}
f^{\mathrm{Pareto}}_{(k,N)}(y;x,\beta,\alpha) = N\alpha\beta^\alpha (y-x)^{-(\alpha+1)}\begin{pmatrix}N-1\\k-1\end{pmatrix}\left(1-(\frac{\beta}{y-x})^\alpha\right)^{k-1}\left(\frac{\beta}{y-x}\right){\alpha(N-k)}.
\label{eq:pareto_order}\end{aligned}$$ The minimum order statistic has the same form of distribution $$\begin{aligned}
f^{\mathrm{Pareto}}_{(1,N)}(y;\beta,\alpha) = f^{\mathrm{Pareto}}(y;\beta,\bar{\alpha}),\end{aligned}$$ where $\bar{\alpha}=N\alpha$. The MSE of the minimum order statistic estimator is $$\begin{aligned}
{\mathrm{MSE}}(\hat{x}_{\mathrm{min}}^{\mathrm{Pareto}}(y_{1:N})) = \frac{N\alpha\beta^2}{N\alpha-2}\end{aligned}$$ The unbiased estimator is thus given by,
$$\begin{aligned}
\hat{x}^{\mathrm{Pareto}}(y_{1:N},\beta,\alpha) &= y^{(1)}_{1:N} - \frac{N\alpha\beta}{N\alpha-1},&\quad N\alpha>1
\\
{\mathrm{MSE}}(\hat{x}^{\mathrm{Pareto}}(y_{1:N},\beta,\alpha)) &=\frac{N\alpha\beta^2}{(N\alpha-1)^2(N\alpha-2)},&\quad N\alpha>2
\end{aligned}$$
No unbiased estimator for unknown hyper parameter case could be found for Pareto distribution.
Mixture of Normal and Uniform Noise Distribution
================================================
Suppose the error is distributed as $$e_k\sim \alpha\mathcal{N}(0,\sigma^2) + (1-\alpha) \mathcal{U}[0,\beta],$$ where $\alpha$ is the mixing probability of the mixture distribution. Define $f^{\mathrm{\mathcal{U},\mathcal{N}}}(y_k)$ as the probability density function of the considered mixture distribution given by $$\begin{aligned}
\label{eq:mixture_pdf}
f&^{\mathrm{\mathcal{U},\mathcal{N}}}(y_k;x,\alpha,\sigma^2,\beta) =\nonumber\\
&\begin{cases}
\begin{aligned}
& \frac{a}{\sqrt{2\pi\sigma^2}}\exp\left[-\frac{(y_k-x)^2}{2\sigma^2}\right] +\frac{1-a}{\beta} \end{aligned} & \parbox[t]{4cm}{\raggedright $0\leq y_k-x\leq \beta$}\\
\frac{a}{\sqrt{2\pi\sigma^2}}\exp\left[-\frac{(y_k-x)^2}{2\sigma^2}\right] & \parbox[t]{4cm}{\raggedright Otherwise.}\\
\end{cases} \end{aligned}$$ The BLUE, in case of the mixture of normal and uniform measurement noises is given by
$$\begin{aligned}
\hat{x}_{\mathrm{BLUE}}^{\mathrm{\mathrm{\mathcal{U},\mathcal{N}}}}(y_{1:N},\alpha,\beta,\sigma^2) &= \frac{1}{N}\sum_{k=1}^{N}y_k - \frac{\beta(1-\alpha)}{2},
\\
{\mathrm{MSE}}\left(\hat{x}_{\mathrm{BLUE}}^{\mathrm{\mathrm{\mathcal{U},\mathcal{N}}}}(y_{1:N},\alpha,\beta,\sigma^2)\right) &=\nonumber\\ &\frac{\beta^2 \left(1+(2-3\alpha)\alpha\right)+12\alpha\sigma^2}{12N}.
\end{aligned}$$
Noting that at $y_k-x=0$ contributions of the uniform distribution and the mean (mode) of the normal distribution are added together, is maximized at this point. The order statistics PDF for $0\leq y-x\leq \beta$ is given by $$\begin{aligned}
f_{(k,N)}^{\mathrm{\mathrm{\mathcal{U},\mathcal{N}}}}&(y;\alpha,\beta,\sigma^2,x,k) =\nonumber\\ &N\begin{pmatrix}N-1\\k-1\end{pmatrix}\left(\frac{\alpha\exp(-\frac{(y-x)^2}{2\sigma^2})}{\sqrt{2\pi\sigma^2}}+\frac{1-\alpha}{\beta}\right)\nonumber\\
&\times\left(\frac{(1-\alpha)(y-x)}{\beta}+\frac{\alpha}{2}(1+\mathrm{Erf}\left[\frac{y-x}{\sqrt{2\sigma^2}}\right])\right)^{k-1}\nonumber\\
&\times\left(1+\frac{(\alpha-1)(y-x)}{\beta}-\frac{\alpha}{2}(1+\mathrm{Erf}\left[\frac{y-x}{\sqrt{2}\sigma}\right])\right)^{N-k},\end{aligned}$$ where $\mathrm{Erf}(\cdot) = \frac{2}{\sqrt{\pi}}\int_{0}^{\cdot}e^{-t^2}dt$ is the error function. In order to find the best order statistic estimator, we maximize the likelihood function $\ell(k\mid y=x,a,\beta,\sigma^2)$
$$\begin{aligned}
\ell(k\mid y=x,\alpha,\beta,\sigma^2) &= N\begin{pmatrix}N-1\\k-1\end{pmatrix}2(2-\alpha)^{-k}(1-\frac{\alpha}{2})^N\alpha^{k-1}\nonumber\\&\times\left(\frac{1-\alpha}{\beta}+\frac{\alpha}{\sqrt{2\pi}\sigma}\right).
\label{eq:best_order_likelihood_general}
\end{aligned}$$
Noting that $\left(\frac{1-\alpha}{\beta}+\frac{\alpha}{\sqrt{2\pi}\sigma}\right)$ is always positive and independent of $k$, we extract it from the likelihood function. Simplifying by means of manipulating the terms, we get
$$\begin{aligned}
2(2-\alpha)^{-k}&=2^{1-k}(1-\frac{\alpha}{2})^{-k},\\
\alpha^{k-1}&=(2\frac{\alpha}{2})^{k-1}=2^{k-1}(\frac{\alpha}{2})^{k-1}.
\end{aligned}$$
the likelihood function to be maximized can be re-written as $$\begin{aligned}
\ell(k\mid y=x,\alpha,,\sigma^2) \propto \begin{pmatrix}N-1\\k-1\end{pmatrix}(\frac{\alpha}{2})^{k-1}(1-\frac{\alpha}{2})^{N-k}.
\label{eq:best_order_likelihood_modified}
\end{aligned}$$
In order to find the maximum likelihood estimate $\hat{k} = \operatorname*{arg\,max}_k \ell(k\mid y-x=0)$, we note that is a binomial distribution with probability of success $\frac{\alpha}{2}$. Hence, the maximum of the function is given at the mode of the distribution, $$\begin{aligned}
\hat{k}=\lfloor \frac{N\alpha}{2}\rfloor+1\quad \mathrm{or}\quad \lceil\frac{N\alpha}{2}\rceil.\end{aligned}$$ This gives the best order statistic estimator for the case when noise is a mixture of normal and uniform distribution as $$\begin{aligned}
\hat{x}^{\mathrm{\mathrm{\mathcal{U},\mathcal{N}}}}(y_{1:N},\alpha)= y^{(\hat{k})}_{1:N}.\end{aligned}$$
distribution bias MSE
------------------------------------------ --------------------------------------------------------- ----------------------------------------------------------
$\mathcal{U}[0,\beta]$ $\frac{\beta}{N+1}$ $\frac{2\beta^2}{(N+1)(N+2)}$
\[2mm\] $\mathrm{Exp}(\beta)$ $\frac{\beta}{N}$ $\frac{2\beta^2}{N^2}$
\[2mm\] $\mathrm{Rayleigh}(\beta)$ $\frac{\sqrt{\pi}\beta}{\sqrt{2N}}$ $\frac{2\beta^2}{N}$
\[2mm\] $\mathrm{Weibull}(\beta,\alpha)$ $\beta N^{-\frac{1}{\alpha}}\Gamma(1+\frac{1}{\alpha})$ $\beta^2N^{-\frac{2}{\alpha}}\Gamma(1+\frac{2}{\alpha})$
\[2mm\] $\mathrm{Pareto}(\beta,\alpha)$ $\frac{N\alpha\beta}{N\alpha -1}$ $\frac{N\alpha\beta^2}{N\alpha -2}$
: Bias and MSE of minimum order statistics estimators $\hat{x}_{\mathrm{min}}^{p_e}$.
\[tbl:estimators\_biased\]
\[tbl:estimators\]
Performance Evaluation {#sec:performance_evaluation}
======================
The estimators (both unbiased and the ones without bias compensation) derived in sections \[sec:distributioons\_in\_the\_exponential\_family\]–\[sec:other\_distributions\] for different noise distributions together with their MSE are summarized in Tables \[tbl:estimators\] and \[tbl:estimators\_biased\]. The biased minimum order statistics based estimators and their MSE are also The estimators derived for each noise distribution are compared against each other as a function of the sample size $N\in[2,\ldots,2000]$. Additionally, in order to verify the analytical derivations of the estimator variances, they are compared against the numerical variances obtained from $M=5000$ Monte Carlo runs.
Simulation Setup
----------------
For each sample size, $N$ noisy measurements of the unknown parameter $x$ are generated. The hyper parameters of the noise distributions are randomly selected in each repetition. In order to have a fair comparison, the hyper parameters are randomly drawn such that the error densities are mostly in the same range for all scenarios. The noise realizations are generated from the six considered distributions with the following hyper parameters
- Uniform noise: $\beta\sim\mathcal{U}[6,50]$
- Exponential noise: $\beta\sim\mathcal{U}[5,14]$
- Rayleigh noise: $\beta\sim\mathcal{U}[5,12]$
- Weibull noise:$\beta=1$, $\alpha\sim\mathcal{U}[5,10]$
- Pareto noise: $\beta=6$, $\alpha\sim\mathcal{U}[2.1,2.5]$
- Mixture noise: $\sigma\sim\mathcal{U}[1,9]$, $\beta\sim\mathcal{U}[1,50]$
The empirical CDF of the error values used in the simulations are presented in Figure \[fig:cdf\_error\]. The support of the noise values, as can be read from the figure, is $\bm{e}_m\in[0,60]$ unit.
Let $\hat{x}_N^{(m)}$ denote the estimated value of the unknown parameter $x$ in the $m$:th repetition obtained from a sample of size $N$. For each noise distribution, the estimators’ performances are evaluated in terms of the obtained MSEs. The theoretical MSE of each estimator, as defined in Table \[tbl:estimators\] and Table \[tbl:estimators\_biased\], is compared against the numerical MSE obtained in simulations.
We let ${\mathbb{E}}[\hat{x}_N] = \frac{1}{M}\sum_{m=1}^{M}\hat{x}_N^{(m)}$ and define
$$\begin{aligned}
\hat{b}_N &= {\mathbb{E}}[\hat{x}_N] - x\\
\hat{\sigma}^2_N &= \frac{1}{M}\sum_{m=1}^{M}(\hat{x}_N^{(m)}-{\mathbb{E}}[\hat{x}_N])^2.
\end{aligned}$$
The numerical MSE for each sample size $N$ is then computed by $$\begin{aligned}
\hat{\mathrm{MSE}}(\hat{x}_N) = \hat{\sigma}^2_N + \hat{b}_N^2.
\end{aligned}$$
![Analytical (marked with solid lines) and numerical (marked with crosses) MSE for uniform noise distribution as a function of the sample size $N$.[]{data-label="fig:uniform_blue_mvu"}](cdf_error.pdf){width="1.14\linewidth"}
![Analytical (marked with solid lines) and numerical (marked with crosses) MSE for uniform noise distribution as a function of the sample size $N$.[]{data-label="fig:uniform_blue_mvu"}](estimationMse_unif.pdf){width="\linewidth"}
Simulation Results
------------------
Figure \[fig:uniform\_blue\_mvu\] presents the performance of the four estimators when the noise is uniformly distributed. The solid lines correspond to the theoretical MSEs and the crosses are the numerical MSEs obtained from $M=5000$ repetitions. Both MVU estimators, with and without any knowledge of the hyper parameters of the underlying noise, result in noticeably less MSE compared to the BLUE estimator. The minimum order statistics estimator also outperforms BLUE when measurements are corrupted with additive, uniformly distributed, noise. It can be further observed that if the hyper parameter $\beta$ is unknown, the MSE of the proposed estimator is negligibly larger than the case with known $\beta$.
\
For the exponential noise distribution, as shown in Figure \[fig:exponential\_blue\_mvu\_all\], there is still a non-negligible difference between BLUE and the other three estimators in terms of estimators’ MSE. However, the two MVU estimators, specially for large values of $N$, behave similarly. In order to verify their performance for smaller sample sizes, Figure \[fig:exponential\_blue\_mvu\_zoom\] illustrates the variances of all estimators for $N\leq 20$. At the beginning, $N\in[2,4]$ the estimator with unknown hyperparameter has the largest MSE. However, for larger sample sizes, the two MVU estimators are almost equal and both have less MSE than the BLUE estimator. As in case of uniformly distributed measurement noise, the minimum order statistics estimator outperforms BLUE specially for large sample sizes.
In case of Rayleigh noise distribution, as given in Table \[tbl:estimators\], the minimum order statistics estimator has the largest MSE while the BLUE and the proposed unbiased estimator with known hyper parameter, result in similar estimation variance. This can be verified also in the simulation results presented in Figure \[fig:rayleigh\_blue\_order\_all\]. For large sample sizes, $N>20$, these two estimators and the proposed estimator with unknown hyperparameter have similar values. However, for the smaller sample sizes, as illustrated in Figure \[fig:rayleigh\_blue\_order\_zoom\], the BLUE (and order statistic with known hyper parameter) estimator has smaller variance compared to the case with unknown hyper parameter. The minimum order statistics estimator results in larger MSE compared to the other three estimators in case of Rayleigh noise distribution.
As Table \[tbl:estimators\] suggests, for Pareto and Weibull noise distributions, we only derived BLUE and an unbiased order statistics based estimators when the two hyperparameters of the distributions are known. For both noise distributions, the MSE of the two unbiased estimators as well as the MSE of the minimum order statistics estimator are compared and the results are presented in Figure. \[fig:pareto\_weibull\_blue\_order\]. In both cases, the proposed estimators outperform the BLUE in terms of variance. The minimum order statistics estimator results in a lower MSE than the BLUE for Weibull noise distributions. However, in case of Pareto noise, the BLUE has a better performance compared to the minimum order statistics estimator.
In case of mixture noise distribution, we consider three different scenarios based on the mixing probabilities; two extreme cases with dominant contribution from uniform noise, $\alpha=0.01$, and dominant contribution from normal noise, $\alpha=0.99$, and the case with $\alpha=0.5$. Fig. \[fig:mixture\_pdf\] illustrates the histogram of the noise realizations of the considered mixture noise distributions $e_k\sim \alpha\mathcal{N}(0,8^2) + (1-\alpha)\mathcal{U}(0,60)$ and the fitted densities. The empirical CDFs of the errors for the three cases are presented in Figure \[fig:mixture\_cdf\].
In order to estimate the unknown parameter $x$, in each Monte Carlo run, we sort the measurements and then find the $(\lfloor\frac{N\alpha}{2}\rfloor+1)$:th component. Figure \[fig:mixture\_order\] presents the estimation MSE for the three different scenarios with different mixing probabilities. As the results indicate, when the main contribution of the noise is from uniform distribution, $\alpha=0.01$, BLUE outperforms the proposed estimator. In this case, a periodic behavior for the MSE can be observed. The jumps in the MSEe occur exactly at pints where $\lfloor\frac{N\alpha}{2}\rfloor+1$ switches from the $k$:th measurement to the $k+1$:th measurement. For instance, for $N\in[1,199]$, $\lfloor\frac{N\alpha}{2}\rfloor=0$, hence $\hat{x}=y^{(1)}$. However, at $N=200$, $\lfloor\frac{N\alpha}{2}\rfloor=1$, resulting in $\hat{x}=y^{(2)}$.
The proposed estimator and the BLUE result in similar estimation MSE for $\alpha=0.99$, as shown in Figure \[fig:estimationMse\_mixture99\], in which the normal component is the dominant source of error. However, the most interesting results are obtained when both distributions have equal contributions in the measurement noise, [*i.e*]{} $\alpha=0.5$. In this case, as Figure \[fig:estimationMse\_mixture5\] suggests, the proposed estimator outperforms the BLUE.
Conclusions {#sec:conclusions}
===========
In this work, the location estimation problem was studied in which an unknown parameter was estimated from observations under additive noise. Multiple noise distributions were considered and, in some cases, MVU estimators were proposed. In other cases an unbiased estimator based on minimum order statistic was derived. Furthermore, if applicable, MVU and minimum order statistic estimators without any knowledge of the hyper parameters of the underlying noise distributions were provided. The results of all the estimators were compared with BLUE in terms of variance for various measurement sample sizes. The results indicate better performance of the proposed estimators compared to BLUE. Additionally, the location estimation problem under mixture of normal and uniform noise distribution was studied and the numerical MSE of the proposed estimator were evaluated. The simulation results indicate that for the extreme cases where either of the two components, Gaussian or uniform, are dominant, the proposed estimator cannot beat the BLUE. However, when the mixing probability is not in the extreme region, [*e.g*]{} larger than $1$ percent, the proposed estimator has a noticeably less MSE compared to the BLUE.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report the discovery of the second accreting millisecond X-ray pulsar (AMXP) in the globular cluster NGC 6440. Pulsations with a frequency of 205.89 Hz were detected with the Rossi X-Ray Timing Explorer on August 30th, October 1st and October 28th, 2009, during the decays of $\lesssim4$ day outbursts of a newly X-ray transient source in NGC 6440. By studying the Doppler shift of the pulsation frequency, we find that the system is an ultra-compact binary with an orbital period of 57.3 minutes and a projected semi-major axis of 6.22 light-milliseconds. Based on the mass function, we estimate a lower limit to the mass of the companion to be 0.0067 M$_{\odot}$ (assuming a 1.4 M$_{\odot}$ neutron star). This new pulsar shows the shortest outburst recurrence time among AMXPs ($\sim1$ month). If this behavior does not cease, this AMXP has the potential to be one of the best sources in which to study how the binary system and the neutron star spin evolve. Furthermore, the characteristics of this new source indicate that there might exist a population of AMXPs undergoing weak outbursts which are undetected by current all-sky X-ray monitors. NGC 6440 is the only globular cluster to host two known AMXPs, while no AMXPs have been detected in any other globular cluster.'
author:
- |
D. Altamirano A. Patruno, C. O. Heinke, C. Markwardt, T. E. Strohmayer,\
M. Linares, R. Wijnands, M. van der Klis, J. H. Swank
title: 'Discovery of a 205.89 Hz accreting–millisecond X-ray pulsar in the globular cluster NGC 6440'
---
[*Keywords*]{}:Galaxy: globular clusters: individual (NGC 6440): X-rays: binaries, stars: neutron
Introduction {#sec:intro}
============
Accreting millisecond X-ray pulsars (AMXPs) are rapidly spinning neutron stars accreting matter from a low-mass stellar companion. These systems are thought to be the progenitors of the millisecond radio pulsars [@Alpar82; @Backer82]. Therefore, their study allows us to better understand how neutron stars in low-mass X-ray binaries (LMXB) evolve into radio millisecond pulsars. Furthermore, AMXPs are perfect laboratories to study the interaction between the neutron star magnetosphere and the accretion disk [see, e.g., @Poutanen06 and references therein].
A total of ten AMXPs among the $\sim100$ neutron star LMXBs were known by August 2009. Seven of these AMXPs showed outbursts lasting a few days to months, and recurrence times of at least 2 years. In all seven cases, the pulsations were detected persistently during the whole outburst. The remaining three AMXPs showed pulsations only intermittently [@Galloway07a; @Gavriil07; @Casella08; @Altamirano08b; @Patruno09], bridging the gap between the small number of AMXPs and the large group of non-pulsating LMXBs [@Altamirano08b]. The reason why only a small amount of neutron star LMXBs show pulsations is still under debate. One of the proposed scenarios for the non-pulsating systems is that the neutron-star magnetic field could be temporarily buried by the accreted matter [@Cumming01]. If the time-averaged mass accretion rate of the accretors is relatively high, the accreted matter can bury the field throughout the life of the X-ray binary so the neutron star does not pulsate. However, if the average accretion rate is low enough that the magnetic field can diffuse through the accreted matter faster than it is buried, then these systems will probably exhibit pulsations. Another of the proposed scenarios that explains why the majority of neutron stars in LMXBs do not pulsate is that of @Lamb09a; if neutron stars undergo long periods of accretion (at high rates), then their magnetic poles are naturally forced to align to their spin axes as they spin up. When the accretion rate decreases, the magnetic poles can move away from the rotation axis (due to magnetic dipole and other braking torques which cause neutron stars to spin down) and oscillations powered by accretion should become visible. Both @Cumming01 and @Lamb09a present clear predictions: we should find pulsations in many (if not all) of the systems which show low time-averaged mass accretion rates. Very recently, a Chandra X-ray observation serendipitously discovered a new X-ray transient in the globular cluster NGC 6440 [@Heinke09a; @Heinke09b; @Heinke09c identified as NGC 6440 X-2]. RXTE and Swift follow-up observations not only showed that this source has a very low time–average accretion rate [$\lesssim2 \cdot 10^{-12}$ M$_{\odot}$ year$^{-1}$, see @Heinke09d] but, as predicted by the models discussed above, also millisecond X-ray pulsations [@Altamirano09]. One of the most interesting properties of this new AMXP is its very short recurrence time [$\sim1$ month, see also @Heinke09d]. If its outbursts continue, it will be possible to study the evolution of spin and orbital parameters on timescales never before probed. In this Letter we report the discovery of X-ray pulsations in NGC 6440 X-2. In a companion work [@Heinke09d], we report on a detailed description of the evolution of the source’s outbursts based on our multiwavelength campaign on this source. Our results demonstrate that there may exist a population of AMXPs undergoing low-luminosity outbursts which are undetected by current all-sky X-ray monitors.
LMXBs in the globular cluster NGC 6440
======================================
NGC 6440 is a globular cluster at $8.5\pm0.4$ kpc [@Ortolani94]. Using Chandra images, @Pooley02 studied the population of the low-luminosity X-ray sources in NGC 6440. These authors found 24 X-ray sources above a limiting luminosity of $\sim2 \cdot 10^{31}$ ergs s$^{-1}$ (0.5–2.5 keV) inside the half-mass radius of the cluster (all of which lie within $\sim$2 core radii of the cluster center); there is strong evidence that 8 of these systems are LMXBs in quiescence [@Heinke03]. @Pooley02 also found excess emission in and around the core of NGC 6440, suggesting unresolved point sources.
Bright X-ray outbursts from a LMXB were observed in 1971, 1998, 2001 and 2005 [@Markert75; @Zand99; @Verbunt00; @Zand01; @Markwardt05]. From X-ray and optical observations, @Zand01 concluded that the 1998 and 2001 outbursts were from the same object, which they designated SAX J1748.9–2021. @Altamirano08b detected intermittent X-ray pulsations at $\sim442$ Hz in the 2001 and 2005 outbursts [see also @Gavriil07]. Whether the 1971 outburst was from SAX J1748.9–2021 is uncertain.
Between MJD 54875 and 55075 at least 5 short-duration ($\sim$days) outbursts have been detected in NGC 6440. As discussed in @Heinke09d, Chandra, RXTE and Swift data indicates that all outbursts were from the same source, denoted NGC 6440 X-2.
Observations, data analysis and results {#sec:dataanalysis}
=======================================
We used data from the RXTE Proportional Counter Array [PCA, for instrument information see @Jahoda06]. Between August 1998 and November 4th, 2009, there were 49 pointed observations of the globular cluster NGC 6440 each covering 1 to 5 consecutive 90-min satellite orbits. Usually, an orbit contains between 1 and 5 ksec of useful data separated by 1–4 ksec data gaps due to Earth occultations and South Atlantic Anomaly passages. The first 27 observations sample 3 outbursts of the Intermittent AMXP SAX J1748.9–2021 [@Altamirano08b]. The remaining 22 were triggered based on results of the PCA Bulge Scan program or ToO Swift observations, and sample the 5 outbursts of NGC 6440 X-2 [see also @Heinke09d].
We performed a timing analysis using the high-time resolution data collected in the Event (E\_125us\_64M\_0\_1s), Good Xenon and Single-Bit modes. Fourier power spectra were constructed for each observation, using data segments of 128, 256 and 512 seconds and with a Nyquist frequency of 4096 Hz. No background or dead-time corrections were made prior to the calculation of the power spectra, but all reported rms amplitudes are background corrected; deadtime corrections are negligible.
Parameter Value
--------------------------------------------------------- ---------------
Orbital period, P$_{orb}$(seconds) 3438(33)
Projected semi major axis, $a_x sin i$ (lt-ms) 6.22(7)
Time of ascending node, $T_{asc}$ (MJD/TDB) 55073.0344(6)
Eccentricity, e (95% c.l.) $<0.07$
Spin frequency $\nu_0$ (Hz) 205.89215(2)
Pulsar mass function, $f_x$ ($\times 10^{-7} M_{\sun}$) $1.6(1)$
Minimum companion mass$^1$, $M_c$ ($M_{\sun}$) $0.0067$
: Timing parameters for NGC 6440
\
All errors are at $\Delta\chi^2=1$.\
$^1$: The minimum companion mass is estimated assuming a neutron star mass of 1.4 M$_{\odot}$. \[table:data\]
Light curves and power spectra
------------------------------
In Figure \[fig:lc\] we plot the intensity (2.0–10.0 keV cts/sec/PCA) of the globular cluster NGC 6440 as measured with PCA Bulge Scans [@Swank01]. The data suggest that between MJD 54875 and 55075 the globular cluster has brightened for short periods ($\sim$days) on $\sim5$ occasions [see also @Heinke09d]. However, we note that some of these points might be spurious given that the intensity measured is systematically uncertain due to the contributions from diffuse galactic emission and other nearby sources in the PCA field of view. In none of the pointed nor the Bulge Scan observations did we detect thermonuclear X-ray bursts.
As can be seen in Figure \[fig:lc\], on August 30th, 2009 (MJD 55072) the PCA bulge scan detected the highest intensity in the direction of NGC 6440 over a period of more than 250 days. A ToO RXTE observation (ObsID: 94044-04-02-01) found NGC 6440 at a luminosity of $L_{2-10 {\rm keV}}\sim1.7 \cdot 10^{36}$ ergs s$^{-1}$ [using a distance of 8.5 kpc, see @Ortolani94]. The energy spectrum is well fitted with an absorbed power law with index of $1.83\pm0.3$ ($\chi^2/dof=0.84$ for 47 degrees of freedom; the interstellar absorption $N_h$ was fixed to a value of $5.9 \cdot
10^{21}$ cm$^{-2}$, consistent with the cluster value and the absorption measured by Chandra, see @Heinke09d). The power spectrum of this observation shows two features with characteristic frequencies [@Belloni02] of $7.4\pm0.8$ and $1.13\pm0.06$ Hz, quality factor Q of $1.3\pm0.6$ and $0.71\pm0.07$, and rms amplitudes of $28\pm3$% and $42\pm2$%, respectively. We searched for kHz QPOs, but found none. The average (background-subtracted) count rate during this observation was $\sim22$ cts/sec and only one proportional counter unit (PCU) was active (2–60 keV). Additional RXTE observations were performed on the 1st and 2nd of September; fluxes were consistent with background emission. These measurements are consistent with the fluxes measured by Swift between the 1st and the 4th of September [@Heinke09d]. If we take into account that a PCA bulge scan observation did not detect significant flux in the direction of NGC 6440 on August 28th, we can estimate an outburst duration of no more than $\sim3$ days (above $\sim10^{35}$ erg s$^{-1}$).
On July 30th, October 1st and October 28th, 2009, RXTE sampled three other outbursts of NGC 6440 [@Heinke09d] for $\sim1.5-2.5$ ksec. We did not find any significant QPOs nor broad band noise in their power spectra, probably due to poor statistics (note that the total count rates measured are dominated by background counts). These observations ranged from $L_{2-10 {\rm
keV}}=1\times10^{35}$ to $5.2\times10^{35}$ ergs s$^{-1}$, thus providing poorer statistics than on August 30th. We found no evidence for dips or eclipses in any of the RXTE pointed observations of NGC 6440 X-2.
Pulsations
----------
Adopting a source position $\alpha=17^h 48^m 52^s.75$, $\delta = -20^o
21^{'} 24^{''}.0$ [from Chandra images, see @Heinke09d], we converted the photon arrival times to the Solar System barycenter with the FTOOL faxbary, which uses the JPL DE-405 ephemeris along with the spacecraft ephemeris and fine clock corrections to provide an absolute timing accuracy of 3.4 $\rm\,\mu s$ [@Jahoda06].
We found a strong signal at a frequency of $\sim205.89$ Hz in the August 30th RXTE observation. In Figure \[fig:pds\] we show the Leahy normalized power spectrum. We calculated a dynamical power spectrum with a 512 sec sliding window and found that the pulse frequency is modulated in a way that is typical from Doppler shifts due to the orbital motion of a pulsar in a binary system (see Figure \[fig:dyna\]); the data sample a little less than a full orbital cycle. To improve the signal to noise of the pulsations, we first modeled the pulse drifts with a sinusoid (assuming zero eccentricity). We found that the frequency drift is consistent with a system with an orbital period of 57 minutes and a projected semi-major axis of 6.2 lt-ms. With this provisional solution, we folded the 3 ks of lightcurve into 10 pulse profiles of $\sim 300$ s each and fitted them with a constant plus two sinusoids representing the pulse frequency and its first overtone. We then phase-connected the pulse phases by fitting a constant pulse frequency plus a circular Keplerian orbital model. The procedure is described in detail in @Patruno09. In Table \[table:data\] we report the best fit. In the inner panel of Figure \[fig:pds\] we show the folded pulse profile in the 2–14.8 keV range (i.e. channels 0–35) after correcting for the effects of the orbital modulation. The fractional rms amplitudes of the fundamental and first overtone are $7.5\pm0.6$ and $2.3\pm0.6$%, respectively (2.0–14.8 keV).
We applied the above technique to three energy bands: 2–5.7 keV, 6.1–14.8 keV and 15.2–60 keV. (We note that we could only choose these bands, as the data were split in two Single-bit modes – SB\_125us\_0\_13\_1s & SB\_125us\_14\_35\_1s – and one Event mode – E\_62us\_32M\_36\_1s.) The pulsed fractional rms amplitudes of the fundamental were $(7.5\pm1.0)$% and $(8.2\pm1.0)$% in the 2-5.7 keV and 6.1-14.8 energy band. The first overtone was significantly detected in the latter band, with an amplitude of $3.3\pm0.8$% rms. Due to the low quality of the data at high energies, only upper limits of 21% rms (at 95% confidence level) can be put on the fundamental in the 15.2-60 keV band. (NGC 6440 X-2 was more than 100 times brighter than the other 24 X-ray sources in NGC 6440 [@Pooley02], so their contributions can be ignored. However, Galactic Ridge emission does provide additional background at this location, not included in standard RXTE background estimates, and therefore the rms amplitudes we quote above might be slight underestimates).
We also searched for pulsations in the observations on July 30th, October 1st and October 28th, 2009, by correcting for the orbital motion of the system and folding the data around the best spin period determination. No pulsations were found on July 30th, with 95% c.l. upper limits of 8% rms on the pulse fractional amplitude. On October 1st and 28th we found $3.4\sigma$ and $8.9\sigma$ detections, respectively, of pulsations with characteristics consistent with those we found on August 30th.
RXTE observed NGC 6440 during the three outbursts of the AMXP SAX J1748.9–2021 [see, e.g., @Altamirano08b; @Patruno09]. Although improbable, it is possible that NGC 6440 X-2 was also active during these periods. To further investigate this, we searched all previous RXTE observations of NGC 6440 for the presence of pulsations at $\sim205-206$ Hz. We barycentered all data in the same manner as above and divided the data in 3 groups, corresponding to the 1998, 2001 and 2005 outbursts. For each group we performed Fourier transforms of all the data and calculated a single averaged power spectra. We found no significant pulsations in the $\sim205-206$ Hz range. We also searched for evidence of intermittent pulsations such as those found in other AMXPs [@Galloway07a; @Gavriil07; @Altamirano08b; @Casella08], but found none. In the above searches, we did not correct for the orbital motion of the binary; although these corrections would improve our sensitivity, such analysis is beyond the scope of this Letter. Finally, we note that it was not possible to search for pulsations in the Swift observations. These data-sets were obtained in photon-counting mode with a time resolution of 2.5 seconds.
Discussion {#sec:discussion}
==========
We have discovered the second accreting millisecond X-ray pulsar in the globular cluster NGC 6440. Pulsations were detected in three RXTE pointed observations performed during the decay phase of three outbursts, each of which lasted less than $\sim3$ days [@Heinke09d]. The frequency drifts we observe are consistent with an orbital period of 57.3 minutes and a semi-major axis ($a_x$ sin$i$) of 6.22 light-msec, which shows that this new AMXP is in an ultra-compact binary system. The characteristics of this new pulsar are similar to those of other AMXPs, particularly SWIFT J1756.9–2508 [@Krimm07; @Patruno10], which showed pulsations at a frequency of 182 Hz, an orbital period of 54.7 minutes, a projected semi-major axis of 5.94 lt-ms and a minimum companion mass of 0.0067 M$_{\odot}$. As pointed out by @Krimm07, such a low mass for a companion in this type of system is probably inconsistent with brown dwarf models while white dwarf models suggest that the companion is probably a He-dominated donor.
The reason why only twelve[^1] NS LMXBs exhibit coherent pulsations out of about 100 systems is still unclear. There is a strong tendency for the detected AMXPs to have rather low time-averaged accretion rates [over many outbursts, see, e.g., @Chakrabarty05; @Galloway06a; @Wijnands08]. The time-averaged accretion rate of NGC 6440 X-2 is at most similar to that of the AMXP SAX J1808.4–3658, but could easily be much lower [see discussion by @Heinke09d]. This is consistent with the general tendency and furthermore, with the theoretical predictions [@Cumming01; @Lamb09a] that a significant fraction (if not all) of the NS LMXB systems with low time-averaged accretion rates harbor accreting millisecond X-ray pulsars [@Wijnands08]. The outburst durations ($\sim3$ days) of the newly discovered AMXP are short compared with those of the other AMXPs, although not uniquely so. For example, short-duration outbursts of the AMXP XTE J1751–305 have been detected in 2005, 2007 and 2009 [@Grebenev05; @Swank05; @Markwardt07b; @Markwardt07c; @Linares07a; @Markwardt09a]. Our discovery of NGC 6440 X-2 indicates that there might exist a population of AMXPs undergoing weak outbursts. Unfortunately, short low-luminosity outbursts cannot be detected with the current all-sky X-ray monitors (e.g., ASM aboard RXTE). More sensitive instruments like the “Monitor of All-sky X-ray Image” [MAXI, which should detect X-ray – 2–30 keV – sources of about 20 mCrab from 90 minutes observations, and sources of about 4.5 mCrab after a day observation, see @Matsuoka09] will probably allow us to discover many more potential low-luminosity AMXPs. As discussed in @Heinke09d, NGC 6440 X-2 has been in outburst at least five times in the last 250 days. If true, this is the AMXP with the shortest recurrence time. If NGC 6440 X-2 continues undergoing this type of short outbursts and RXTE is able to observe them, then NGC 6440 X-2 has the potential to be the best millisecond X-ray pulsar in which to study the evolution of the binary as well as that of the neutron star spin (by means of coherent timing analysis).
The globular cluster NGC 6440 is known for having at least 24 faint X-ray sources to a limiting luminosity of $\sim2 \cdot 10^{31}$ ergs s$^{-1}$ [0.5–2.5 keV, see @Pooley02]. NGC 6440 is also known for hosting at least 6 millisecond radio pulsars, of which 3 are in binary systems [@Freire08]. None of the known radio pulsars match the position (nor the characteristics) of NGC 6440 X-2. Although millisecond radio pulsars in binary systems have been found in other globular clusters [see, e.g., @Ransom05; @Freire08 and references therein], it is an intriguing question why NGC 6440 is the only globular cluster known today to host (two) AMXPs (SAX J1748.9–2021 – @Gavriil07 and @Altamirano08b – and NGC 6440 X-2 – this paper), while no AMXPs have been detected in any other globular cluster. This might be only a selection effect, as current all-sky monitors cannot detect short low-luminosity outbursts. Monitoring X-ray observations of different globular clusters would be needed to further investigate this.
[42]{} natexlab\#1[\#1]{}url \#1[[\#1]{}]{}urlprefix
M.A., [Cheng]{} A.F., et al., 1982, , 300, 728
D., [Casella]{} P., et al., Feb. 2008, , 674, L45
D., [Strohmayer]{} T.E., et al., Aug. 2009, The Astronomer’s Telegram, 2182, 1
D.C., [Kulkarni]{} S.R., et al., 1982, , 300, 615
T., [Psaltis]{} D., [van der Klis]{} M., 2002, , 572, 392
P., [Altamirano]{} D., et al., Feb. 2008, , 674, L41
D., Jul. 2005, In: [F. A. Rasio & I. H. Stairs]{} (ed.) Binary Radio Pulsars, vol. 328 of Astronomical Society of the Pacific Conference Series, 279–+
A., [Zweibel]{} E., [Bildsten]{} L., 2001, , 557, 958
P.C.C., [Ransom]{} S.M., et al., Mar. 2008, , 675, 670
D.K., Jun. 2006, In: [Braga]{} J., [D’Amico]{} F., [Rothschild]{} R.E. (eds.) The Transient Milky Way: A Perspective for MIRAX, vol. 840 of American Institute of Physics Conference Series, 50–54
D.K., [Morgan]{} E.H., et al., 2007, , 654, L73
F.P., [Strohmayer]{} T.E., et al., 2007, , 669, L29
S.A., [Molkov]{} S.V., [Sunyaev]{} R.A., Mar. 2005, The Astronomer’s Telegram, 446, 1
C.O., [Budac]{} S.A., Jul. 2009, The Astronomer’s Telegram, 2139, 1
C.O., [Grindlay]{} J.E., et al., Nov. 2003, , 598, 501
C.O., [Altamirano]{} D., et al., Nov. 2009, ArXiv e-prints:0911.0444
C.O., [Altamirano]{} D., [Markwardt]{} C., Aug. 2009, The Astronomer’s Telegram, 2180, 1
C.O., [Budac]{} S.A., et al., Aug. 2009, The Astronomer’s Telegram, 2143, 1
J.J.M., [Verbunt]{} F., et al., 1999, , 345, 100
J.J.M., [van Kerkwijk]{} M.H., et al., 2001, , 563, L41
K., [Markwardt]{} C.B., et al., 2006, , 163, 401
H.A., [Markwardt]{} C.B., et al., Oct. 2007, , 668, L147
F.K., [Boutloukos]{} S., et al., Nov. 2009, , 706, 417
D.A., [Darbro]{} W., et al., 1983, , 266, 160
M., [Wijnands]{} R., [van der Klis]{} M., Apr. 2007, The Astronomer’s Telegram, 1055, 1
T.H., [Backman]{} D.E., et al., 1975, , 257, 32
C.B., [Swank]{} J.H., 2005, The Astronomer’s Telegram, 495, 1
C.B., [Swank]{} J.H., Apr. 2007, The Astronomer’s Telegram, 1045, 1
C.B., [Pereira]{} D., [Swank]{} J.H., Apr. 2007, The Astronomer’s Telegram, 1051, 1
C.B., [Altamirano]{} D., et al., Sep. 2009, The Astronomer’s Telegram, 2197, 1
C.B., [Altamirano]{} D., et al., Oct. 2009, The Astronomer’s Telegram, 2237, 1
M., [Kawasaki]{} K., et al., Jun. 2009, ArXiv e-prints
S., [Barbuy]{} B., [Bica]{} E., 1994, , 108, 653
A., [Altamirano]{} D., et al., Jan. 2009, , 690, 1856
A., [Altamirano]{} D., [Messenger]{} C., Feb. 2010, , 207–+
D., [Lewin]{} W.H.G., et al., 2002, , 573, 184
J., 2006, Advances in Space Research, 38, 2697
S.M., [Hessels]{} J.W.T., et al., Feb. 2005, Science, 307, 892
J., [Markwardt]{} K., 2001, in ASP Conf. Ser. 251, New Century of X-ray Astronomy, eds. H. Inoue & H. Kunieda (San Francisco: ASP), 94
J.H., [Markwardt]{} C.B., [Smith]{} E.A., Mar. 2005, The Astronomer’s Telegram, 449, 1
F., [van Kerkwijk]{} M.H., et al., 2000, , 359, 960
R., May 2008, In: [R. M. Bandyopadhyay, S. Wachter, D. Gelino, & C. R. Gelino]{} (ed.) A Population Explosion: The Nature & Evolution of X-ray Binaries in Diverse Environments, vol. 1010 of American Institute of Physics Conference Series, 382–386
[^1]: See also the recent discovery of a 245 Hz AMXP by @Markwardt09.
| {
"pile_set_name": "ArXiv"
} |
[OG 4.1.14\
]{} [VERY HIGH ENERGY OBSERVATIONS\
OF PSR B1951+32\
]{} R.Srinivasan$^1$, P.J.Boyle$^2$, J.H.Buckley$^3$, A.M.Burdett$^4$, J.Bussons Gordo$^2$, D.A.Carter-Lewis$^5$,
M.F.Cawley$^6$, M.Catanese$^5$, E. Colombo$^8$, D.J.Fegan$^2$, J.P.Finley$^1$, J.A.Gaidos$^1$, A.M.Hillas$^4$,
R.C.Lamb$^6$, F.Krennrich$^5$, R.W.Lessard$^1$, C.Masterson$^2$, J.E.McEnery$^2$, G.Mohanty$^5$, P. Moriarty$^7$, J.Quinn$^2$, A.J.Rodgers$^4$, H.J.Rose$^4$, F.W.Samuelson$^5$, G.H.Sembroski$^1$, T.C.Weekes$^3$, and J.Zweerink$^5$\
[*$^1$Department of Physics, Purdue University, West Lafayette, IN 47907\
$^2$ Physics Department, University College, Dublin 4, Ireland\
$^3$ Whipple Observatory, Harvard-Smithsonian CfA, P.O. Box 97, Amado, AZ 85645-0097\
$^4$ Department of Physics and Astronomy, University of Leeds, Leeds, LS2 9JT, Yorkshire, UK\
$^5$ Department of Physics and Astronomy, Iowa State University, Ames, IA 50011-3160\
$^6$ Space Radiation Lab, California Institute of Technology, Pasadena, CA 91125\
$^7$ Regional Technical College, Galway, Ireland\
$^8$ Present address: CONAE, Paseo Colon 751, Argentina\
*]{}
[ABSTRACT\
]{}
PSR B1951+32 is a [[$\gamma$]{}]{}-ray pulsar detected by the [*[Energetic Gamma Ray Experiment Telescope]{}*]{} (EGRET) and identified with the 39.5 ms radio pulsar in the supernova remnant CTB 80. The EGRET data shows no evidence for a spectral turnover. Here we report on the first observations of PSR B1951+32 beyond 30 GeV. The observations were carried out with the 10m [[$\gamma$]{}]{}-ray telescope at the Whipple Observatory on Mt. Hopkins, Arizona. In 8.1 hours of observation we find no evidence for steady or periodic emission from PSR B1951+32 above $\sim$ 260 GeV. FLux upper limits are derived and compared with model extrapolations from lower energies and the predictions of emission models.
INTRODUCTION
============
The pursuit of Very High Energy (VHE) astrophysics has resulted in the discovery of five sources, of which three are associated with young spin-powered pulsars. VHE emission has been detected from the direction of the Crab Nebula (Vacanti et.al., 1991), the Vela pulsar (Takanori 1996) and PSR B1706-44 (Kifune et al., 1995) but no evidence has been found for periodic emission at these energies in these experiments.
PSR B1951+32 has been detected as a pulsating X-ray source (Safi-Harb et al., 1995) and as a high energy [[$\gamma$]{}]{}-ray pulsar at E $\ge$ 100 MeV at the radio period (Ramanamurthy et al., 1995). It can be inferred from the five pulsars seen in the MeV to GeV $\gamma$-ray region that longer period or older ( $\sim$ 10$^5$ years ) pulsars have a greater fraction of spin down energy emitted as high energy [[$\gamma$]{}]{}-rays. The best fit outer gap model of Zhang and Cheng (1997) suggests that PSR B1951+32 should emit detectable levels of TeV [[$\gamma$]{}]{}-rays (Figure \[fig2\]). The multiwavelength spectrum of PSR B1951+32 (Figure 1b) indicates a maximum power per decade at energies consistent with a few GeV and still rising at 10 GeV. These factors make PSR B1951+32 a good candidate for observations with the ACT above 100 GeV.
OBSERVATIONS
============
The observations of PSR B1951+32 reported here were acquired with the 10m reflector located at the Whipple Observatory on Mt. Hopkins in Arizona. A total of 14 $\it{Tracking}$ runs and 4 $\it{On/Off}$ pairs taken between 13th May, 1996 and 17th July, 1996 constitute the database for all subsequent discussion. The total $\it{On}$ source observing time is 8.1 hrs. The radio position (J2000) of PSR B1951+32 ($\alpha$ = 19$^h$ 52$^m$ 58.25$^s$, $\delta$= 32$^{\circ}$ 52$^{\prime}$ 40.9$^{\prime\prime}$) was assumed for the subsequent timing analysis.
---------- ------- --------------------- ---------- ------------- ---------------------
PSR P $\dot{P}$ Distance log$_{10}$B Log$_{10}$$\dot{E}$
msec 10$^{-15}$ss$^{-1}$ kpc Gauss ergs/s
B1951+32 39.53 5.8494 2.5 11.69 36.57
---------- ------- --------------------- ---------- ------------- ---------------------
: Pulsar Parameters[]{data-label="tab1"}
ANALYSIS AND RESULTS
====================
Standard Analysis
-----------------
The event selection criteria are collectively called $\it{Supercuts95}$ and a detailed description can be found elsewhere (Catanese et al., 1995). $\it{Supercuts95}$ raises the effective energy threshold of the detector with its software trigger and $\it{size}$ cuts. PSR B1951+32 appears to have a steep spectrum at EGRET energies and since the pulsar spectrum is expected to cut off, it behooves us to reduce the threshold of our analysis to search for a lower energy signal. The dominant background at lower energies is due to muons whose images appear in the camera as arcs and can be discriminated by a cut on their large $\it{length/size}$ values. Hence the selection criteria used $\it{Supercuts95}$ on images with sizes larger than 400 p.e. and $\it{Smallcuts}$ (Table \[tab2\]) for images with sizes less than 400 p.e.
Parameter Supercuts95 Smallcuts
------------- ------------------------------------ --------------------------------------------
length 0${{^\circ}}$16 - 0${{^\circ}}$30 unchanged
width 0${{^\circ}}$073 - 0${{^\circ}}$15 unchanged
distance 0${{^\circ}}51$ - 1${{^\circ}}$1 unchanged
alpha $<$ 15$^{{^\circ}}$ unchanged
max1 $>$ 100 p.e. 45 p.e. - 100 p.e.
max2 $>$ 80 p.e. 45 p.e. - 80 p.e.
size $\geq$ 400 p.e. 0 - 400 p.e.
length/size not used $<$ 7.5 $\times$ 10$^{-4}$ $^{\circ}$/p.e.
: Parameter ranges for selecting [[$\gamma$]{}]{}-ray images[]{data-label="tab2"}
------------------------- --------------------------- --------------------------- -------- ---------------
Selection Source Events Background Events Excess Significance
$\alpha$ $<$ 15$^{\circ}$ $\alpha$ $<$ 15$^{\circ}$
Supercuts95 292 254 38 1.16$\sigma$
Smallcuts 618 672 -54 -1.10$\sigma$
Supercuts95 + Smallcuts 910 926 -16 -0.24$\sigma$
------------------------- --------------------------- --------------------------- -------- ---------------
: Selected Events for Steady Emission analysis[]{data-label="tab3"}
No steady emission is detected from PSR B1951+32 and 3$\sigma$ flux upper limits are displayed in Table \[tab5\]. The effective area for $\it{Supercuts95}$, that was used to calculate the upper limit, was taken as A$_{\it{eff}}$ $\sim$ 3.5 $\times$ 10$^8$ cm$^{2}$: the same area was used for the dataset that resulted from a combination of $\it{Supercuts95}$ and $\it{Smallcuts}$ although here there is more systematic uncertainty. The energy threshold was obtained from simulations and extrapolating the Crab Nebula [[$\gamma$]{}]{}-ray rate for each set of cuts used assuming a spectrum $\sim$ E$^{-2.4}$.
Periodic Analysis
-----------------
The arrival times of the [Čerenkov]{} events were registered by a GPS clock with an absolute resolution of 250 $\mu$sec. An oscillator calibrated by GPS second marks was used to interpolate to a resolution of 0.1 $\mu$sec. After an oscillator drift correction was applied, all arrival times were transformed to the solar system barycenter and folded to produce the phases, $\phi$$_{j}$, of the events modulo the pulse period. The ephemeris frequency parameters used were $\nu$= 25.2963719901267 s$^{-1}$ and $\dot{\nu}$=-3.73940$\times$10$^{-12}$s$^{-2}$, at the epoch t$_0$=JD 2450177.5. This frequency was extrapolated 72 days to obtain a timing solution relevant to the epoch of observation. The datasets, however, were taken within the validity interval of the above ephemeris.
To check the Whipple Observatory timing systems an $\underline{optical}$ observation of the Crab pulsar was undertaken on December 2nd (UT) 1996 using the 10m reflector. The phase analysis of the event arrival times yielded a clear detection of optical Crab pulsar signal which is in phase with the radio pulse and demonstrates the validity of the timing, data acquisition and software in the presence of a pulsed signal. No evidence of pulsed emission from PSR B1951+32 at the radio period exists. To calculate a pulsed flux upper limit we assumed the same pulse profile as seen at EGRET energies, i.e. with the phase range for the main pulse and secondary pulse as 0.12 - 0.22 and 0.48 - 0.74 respectively (Ramanamurthy et al., 1995).
------------------------- -------------------------- ------------------------ ------------
Steady Emission Periodic Emission Threshold
cm$^{-2}$s$^{-1}$) (cm$^{-2}$s$^{-1}$) (GeV)
Supercuts95 0.97 $\times$ 10$^{-11}$ 3.7$\times$ 10$^{-12}$ $\geq$ 370
Supercuts95 + Smallcuts 1.95$\times$ 10$^{-11}$ 6.7$\times$ 10$^{-12}$ $\geq$ 260
------------------------- -------------------------- ------------------------ ------------
: Integral Flux Upper limits[]{data-label="tab5"}
[r]{}[7.5cm]{}
DISCUSSION
==========
PSR B1951+32 is surrounded by a compact nebula which gives a plerionic nature to the supernova remnant, CTB80. X-ray plerions are good candidates for VHE emission since the electrons responsible for nebular synchrotron X-rays should also create VHE [[$\gamma$]{}]{}-rays via the inverse Compton (IC) process. It is expected that for plerions, such as that associated with PSR B1951+32 where the density of nebular synchrotron photons is too low for SSC to take place, detectable VHE emission should be produced by the IC scattering of the 2.7K cosmic microwave background by the same electrons radiating synchrotron X-ray photons. Interpreting the unpulsed X-ray emission form CTB80 as the synchrotron emission from a plerion, the estimated IC flux $>$ 1 TeV is 6.6 $\times$ 10$^{-13}$ TeV/cm$^2$/s/TeV (De Jager et. al., 1995). This represents the lower limit on the IC flux since there can be other sources of soft photons in addition to the microwave background.
To model the pulsed high energy spectrum, a function of the form $$dN_{\gamma}/dE = K E^{-\Gamma}e^{(-E/E_{0})}
\label{eq}$$ was used where E is the photon energy, $\Gamma$ is the photon spectral index and E$_0$ is the cut off energy. The pulsed upper limit reported here is two orders of magnitude lower than the extrapolated EGRET power law. Equation (2) was used to extrapolate the EGRET spectrum to VHE energies constrained by the TeV upper limit reported here and indicates a cut off energy of E$_{0}$ $\leq$ 75 GeV for pulsed emission (Figure 1a).
[r]{}[7.6cm]{}
The strength of the cut off provides a good discriminant between the various pulsar emission models. The status of current observations and the derived cutoff discussed above indicates that the cutoff is beyond 10 GeV. In polar cap models this would indicate a sharp cutoff since the pair production optical depth increases exponentially with photon energy (Harding 1997). However, it is not possible to constrain the shape of the cut off with the non detection of pulsed TeV flux reported here. The most relavant comparison of the Whipple upper limit with emission models is the outer gap model of Zhang and Cheng (see Figure \[fig2\]). This model includes the effect of geometry in the treatment of pulsed emission via a parameter $\alpha$ = r/r$_{L}$, the radial distance to the synchrotron emitting region near the outer gap, r, as a function of the light cylinder radius r$_{L}$. Our pulsed upper limits are consistent with the outer gap model if $\alpha$ $>$ 0.6 implying an emission region far out in the magnetosphere.
The result reported here is the first observation of PSR B1951+32 beyond 30 GeV. PSR B1951+32 exhibits very similar spectral behavior and morphological features, such as an associated synchrotron nebula, to PSR B1706-44 (Finley et al., 1997). If these factors are any indication of similar emission mechanisms in pulsars then the lack of unpulsed emission from PSR B1951+32 is puzzling considering that PSR B1706-44 was detected as a VHE source of unpulsed emission $>$ 1 TeV (Kifune et al., 1995). Lack of pulsed emission indicates that the processes producing pulsed high energy photons over two decades of energy in the EGRET energy range somehow become ineffective over a decade of energy to result in a lack of VHE [[$\gamma$]{}]{}-rays. The low magnetic field of PSR B1951+32 relative to the average pulsar field implies that attenuation of [[$\gamma$]{}]{}-rays by magnetic absorption is not a likely explanation for the non-detection.
ACKNOWLEDGEMENTS
================
We acknowledge the technical assistance of K. Harris. This research is supported by grants from the U.S. Department of Energy, NASA, PPARC in the UK and by Forbairt in Ireland. The authors wish to thank A. Lyne for providing the radio ephemeris of PSR B1951+32 and D. J. Thompson for providing the multiwavelength spectrum for PSR B1951+32.
REFERENCES
==========
[ -5mm]{}
Catanese, M. et al. 1995, Towards a Major Atmospheric Cherenkov Detector -IV, Padova, 335
De Jager, O.C. et al. 1995, in Proc. 24th ICRC (Rome), 2, 528
Finley, J.P. et al. 1997, [ApJ]{}, in preparation
Kifune, T. et al. 1995, [ApJ]{}, 438, L91
Ramanamurthy, P.V., et al. 1995, [ApJ]{}, 447, L109
Safi-Harb, S., Ogelman, H., & Finley, J.P. 1995, [ApJ]{}, 439, 722
Harding, A. 1997, private communication
Takanori, Y. 1996, Ph.D. thesis, University of Tokyo
Vacanti, G., 1991, [ApJ]{}, 377, 467
Zhang, L. and Cheng, K.S. 1997, [ApJ]{}, submitted
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
[We study diffusion on a multilayer network where the contact dynamics between the nodes is governed by a random process and where the waiting time distribution differs for edges from different layers. We study the impact on a random walk of the competition that naturally emerges between the edges of the different layers. In opposition to previous studies which have imposed a priori inter-layer competition, the competition is here induced by the heterogeneity of the activity on the different layers. We first study the precedence relation between different edges and by extension between different layers, and show that it determines biased paths for the walker. We also discuss the emergence of cyclic, rock-paper-scissors random walks, when the precedence between layers is non-transitive. Finally, we numerically show the slowing-down effect due to the competition on a heterogeneous multilayer as the walker is likely to be trapped for a longer time either on a single layer, or on an oriented cycle .\
*Keywords:* random walks; multilayer networks; dynamical systems on networks; models of networks; simulations of networks; competition between layers.]{}
author:
- |
Martin Gueuning\
naXys, Université de Namur, Belgium\
ICTEAM, Université catholique de Louvain, Belgium\
`[email protected]` Sibo Cheng\
Électricité de France R$\&$D\
LIMSI, CNRS, Univ. Paris-Sud, Université Paris-Saclay, France\
ICTEAM, Université catholique de Louvain, Belgium\
Renaud Lambiotte\
Mathematical Institute, University of Oxford, UK \
Jean-Charles Delvenne\
ICTEAM, Université catholique de Louvain, Belgium\
title: 'Rock-Paper-Scissors Random Walks on Temporal Multilayer Networks'
---
Introduction
============
The study of random walks has a long tradition in network science [@Masuda2017]. Random walks are at the heart of many algorithms to uncover central nodes or communities of densely connected nodes, and they often serve as a first model to understand how the topology of a network affects diffusive processes. Random Walks have also been studied mathematically and numerically when the underlying topology is a network enriched with additional features. An important family of models consists of temporal networks [@Holme2012; @Guide2016], where edges are dynamical entities and diffusion can only propagate when they are active. In this case, random walks are affected by the interplay between the network topology and the statistical properties of the edge dynamics [@Starnini2012; @Hoffmann2013; @Delvenne2015]. Another important family is made of multiplex networks [@Domenico2016; @Kivela2014; @Aleta2018], where different types of connections exist between the nodes, as in social networks [@szell2010multirelational; @Magnani2013] or transportation networks [@Cardillo2013; @Gallotti2015] for instance. Multiplex networks have a layered organisation and are usually represented as tensors or by means of a so-called supra-adjacency matrix. Random walks have also been studied in this context, to uncover how the presence of multiple layers affects diffusion [@Domenico2013; @Domenico2014] or to define generalized versions of Pagerank [@de2015ranking].
The main objective of this work is to explore phenomena emerging in the case of networks that are temporal and multiplex [@Starnini2017]. To do so, we extend a generic model of continuous-time random walk on networks [@Hoffmann2013], where the waiting time distribution for an edge, that is the time before an edge gets active, depends solely on its layer. We study the properties of the resulting stochastic process and show that the presence of temporal heterogeneities across layers results in a competition between them and biases in the random walk dynamics. Here competition means that edges in one layer may have a higher probability to be selected by a walker than edges in another layer, due to the statistical properties of their temporal ordering. In other words, competition between layers emerges due to the temporality of the graph, and not as a model parameter as in previous works [@Ding2017; @Gomez-Gardenes2015; @Kleineberg2016; @Jang2015]. In addition, we show and explain some apparently counter-intuitive situations, such as the emergence of a cyclic, rock-paper-scissors precedence between the layers. Note that the notion of non-transitivity is well-known in statistics, and that it has mostly focused on systems having a finite set of possible states, such as in non-transitive dice [@Gardner1970]. Our work can be seen as an extension to continuous variables and a study of its impact on diffusion over multilayer networks. As a second step, we explore numerically the impact of the above mechanism on the dynamics of a walker and, specifically, we study the coverage of a walker on temporal multilayer network.
Random walks on multiplex temporal networks
===========================================
We consider a diffusion process defined as follows [@Masuda2017]. The system is made of a given network of nodes and edges, and a random walker jumping between neighbouring nodes according to a given protocol as we now discuss. In a standard framework of temporal networks, which is a natural model for contacts in social networks, we assume that when the random walker arrives on a node, it triggers on each incident outgoing edge an associated random waiting time for so-called ‘activation’ of the edge, and the first edge to reach activation is then followed by the random walker. The process is restarted on the new node reached by the random walker. Here activation is therefore seen as an event of infinitesimal duration, allowing the passage of the random walker. The time at which each edge reaches activation is independently drawn from its own waiting time distribution. The walk is called active since waiting times are drawn once the walker arrives on a node, as opposite to the passive walk where contacts are taking place on edges regardless of the presence of a random walker. Note that, when the network is directed and has no short cycles, a passive walk may be approximated by an active one provided that waiting time distributions are adapted accordingly [@Gueuning2017; @Masuda2017]. Because only the first edge to reach activation is taken by the walker, there is a underlying competition between different edges. Even though a general master equation can be derived [@Hoffmann2013], specific results have often been studied with homogeneous waiting time distribution [@Karsai2017; @lala; @Gueuning2015]. Here, we consider a situation where edges in the same layer of the multiplex network have the same waiting time distribution but distributions may differ for edges in different layers. The system thus exhibits intra-layer homogeneity and inter-layer heterogeneity.
Emergence of biased paths
-------------------------
Let us consider the trajectory of a random walker. Arriving on node $u$, the walker will leave through the first edge that reaches activation among the $k$ edges connected to $u$. Denote $T_{1},\ldots,T_{k}$ the random variables associated to the waiting times associated to each edge, with distributions $f_1(t),\ldots,f_k(t)$. The transition time $T$ of the walker, defined as the time before its next jump is given by: $$\begin{aligned}
T&=&\min{\left(T_{1},T_{2},....,T_{k}\right)} \label{eq:T}.\end{aligned}$$
In our case, each edge belongs to a layer and it is thus natural to determine which layer will be selected by the walker. In particular, we will be interested in a notion of precedence between the random variables of the layers. The random variable $A$ is said to precede the random variable $B$, written $A \prec B$, if $P(A<B)>0.5$, that is if the edge associated to $A$ is more likely to reach activation before the edge associated to $B$.\
The existence of precedence relations translates into biased paths for the random walker traveling on the network since at each step of the walk some edges may be statistically more likely to be selected by the walker. Therefore, understanding the precedence relation between the random variables associated to the dynamics of the network is of paramount interest in order to understand the resulting diffusion. In the following, we will exemplify the somewhat counter-intuitive properties of precedence, then we will numerically investigate its impact on the coverage of a random walker on a multilayer network with inter-layer heterogeneity.
Basic properties of precedence
-------------------------------
### Rock-Paper-Scissors
One may see precedence as a relation of dominance between edges in competition to attract the random walker. As such one may expect transitivity. However it is not the case, and we encounter circular, rock-paper-scissors situations as follows. We focus on the triangle network of Figure \[fig:sim\] and the three following distributions with expectation equal to $1$:\
$$\begin{aligned}
\label{distr}
X &=&\left\{
\begin{array}{ll}
U\left[0.65,0.75\right] &\quad \textrm{with probability} \quad 1-\frac{1}{\varphi}\\
U\left[0.3\varphi+0.65,0.3\varphi+0.75\right] &\quad \textrm{with probability} \quad \frac{1}{\varphi};
\end{array}
\right.\notag\\
&&\notag\\
Y&\sim&\quad ~U\left[0.9,1.1\right] ;\\
&&\notag\\
Z&\sim&\left\{
\begin{array}{ll}
U~[0.75,0.85] &\quad\quad\quad \textrm{with probability} \quad \frac{1}{\varphi} \\
U\left[\frac{\varphi}{5}+0.95,\frac{\varphi}{5}+1.05\right] &\quad\quad\quad \textrm{with probability} \quad 1-\frac{1}{\varphi} ,
\end{array}
\right. \notag\end{aligned}$$ where $U[a,b]$ stands for the uniform distribution on the interval $[a,b]$, and $\varphi=\frac{1+\sqrt{5}}{2}\approx 1.618$.
It is straightforward to show that $Y\prec X$, $Z\prec Y$ and $X\prec Z$. As a consequence, a walker jumping on the network will have a tendency to jump clockwise on the triangle, as illustrated numerically in Figure \[fig:sim\]. The emergence of a circular flow leads to correlations between the successive edges on the random walk trajectory, even if edges are chosen independently at random at each step. Competition becomes more complex when more than two edges interact together. For instance, in the above example, even if the pairwise precedence is uniform, that is $$\begin{aligned}
P(Y<X)=P(Z<Y)=P(X<Z)=\frac{1}{\varphi},
\end{aligned}$$ when considering the competition between 3 different edges, one finds as $P(X<\min(Y,Z))>\frac{1}{3}$, and thus $X$ will tend to be favoured by the walker against the others two in the presence of the three types of edges simultaneously.
![Numerical simulations of a random walk on a multilayer triangle. Each (colored) pairwise relation belongs to a different layer and distributions activity of the layers are given by the distribution of (\[distr\]). The walker position is incremented $+1$ at each clockwise jump, and $-1$ at each anti-clockwise jump. On the right hand side, the four dash-dot lines represent four realisations of random walk starting at $0$ while the blue line represent the average over 1000 independent walks. The precedence relation is not transitive as a random walker jumping on the triangle will have a tendency to jump clockwise.[]{data-label="fig:sim"}](trian "fig:") ![Numerical simulations of a random walk on a multilayer triangle. Each (colored) pairwise relation belongs to a different layer and distributions activity of the layers are given by the distribution of (\[distr\]). The walker position is incremented $+1$ at each clockwise jump, and $-1$ at each anti-clockwise jump. On the right hand side, the four dash-dot lines represent four realisations of random walk starting at $0$ while the blue line represent the average over 1000 independent walks. The precedence relation is not transitive as a random walker jumping on the triangle will have a tendency to jump clockwise.[]{data-label="fig:sim"}](1d3 "fig:")
### There is no most preceding edge
It is clear that comparing either means or variances of two random variables is not sufficient to determine which one precedes the other, as the random variables $X$, $Y$ and $Z$ in the previous example share equal mean but have different variance. Even more, it is impossible to find a random variable that precedes any random variable with equal mean. Without loss of generality, we prove this statement for random variables with a mean of $1$. First, we observe that for any random variable $X$ following distribution $f(t)$, it is possible to find a random variable $Y_n$ of distribution $g_n(t)$ such that $Y_n$ precedes $X$, by setting $$\begin{aligned}
g_n(t)= & &\left\{
\begin{array}{ll}
\displaystyle\frac{n}{2} (n-1) & \mbox{if } t\in\left[\frac{1}{n}-\frac{1}{n^2}~,~\frac{1}{n}+\frac{1}{n^2}\right] \\
\displaystyle\frac{1}{2 \sqrt{n}}&\mbox{if } t\in\left[n-1+\frac{1}{n}-\frac{1}{\sqrt{n}}~,~ n-1+\frac{1}{n}+\frac{1}{\sqrt{n}}\right], \\
0 & \mbox{otherwise}
\end{array}
\right.\end{aligned}$$ where $n\in\N_0$ is chosen large enough. One possible choice of $n$ is such that $\frac{1}{n}+\frac{1}{n^2}<\beta$, where $\beta$ is the $10\%$ quantile of $f$. From the sequence of random variables $\left(Y_n\right)_{n\in \N}$ with distribution respectively given by $\left(g_n\right)_{n\in \N}$, it is possible to extract a subsequence of random variables $\left(Z_n\right)_{n\in \N}$ whose respective distributions have non-overlapping supports. Such a sequence is increasingly precedent, that is, for any $n \in \N$ we have $Z_{n+1} \prec Z_n$.
As the series $\left(g_n\right)_{n\in \N}$ does not converge to a probability distribution, there exists no most preceding random variable. In terms of diffusion on multiplex networks, this results implies that there is no optimal distribution ensuring that a given layer always captures a majority of the random walk flow, independently of the distributions in the other layers. After the choice of the other layers have been made, a layer may always find a distribution that will allow it to be the most precedent.
### Layer Precedence and Node Out-degree
The notion of precedence between random variables naturally extends to precedence between layers when waiting time distributions inside a layer are homogeneous but vary across layers. In this case, the layer $L_1$ precedes the layer $L_2$ at node $u$ if the walker sitting on $u$ is more likely to perform her next jump through an edge of $L_1$, that is if $L_1$ is more likely to capture the flow passing through $u$. Moreover, in the presence of several layers, $L_1$ precedes $L_2$ and $L_3$ jointly if $L_1$ precedes the artificial layer $L_2 L_3$ obtained by the union of the layers $L_2$ and $L_3$.
Precedence of layers at a node depends on its out-degree $k_i$ on each layer $L_i$, as the layer of the edge selected by a walker is determined by comparing the smallest time $m_i$ to reach activation on each layer $L_i$, where $m_i$ is the minimum of $k_i$ random variables with identical distributions associated to the layer $L_i$. In the simplest case of two duplicate layers $L_1$ and $L_2$ with waiting time on edges of the type $X$ and $Y$ respectively, that is $L_1$ and $L_2$ share the same nodes and have the same edges, the out-going edges consist of $k$ edges of each type, and $L_1$ precedes $L_2$ if $$\begin{aligned}
\displaystyle\min_{j=1,\ldots,k}X_j \prec \displaystyle\min_{j=1 \ldots,k}Y_j ,\end{aligned}$$ where $X_j$ and $Y_j$ are duplicates of random variables following the same distribution than $X$ and $Y$ respectively.
The layer precedence may therefore vary between nodes depending on their out-degree, in particular for high and low out-degree nodes. However, it is worth to note that there always exists a threshold out-degree $\nu^*$ above which one layer will always precede the other one. Indeed, the distribution $f$ is said to have a larger minimal weight than the distribution $g$ if there exists $\epsilon > 0$ such that $P(X<\epsilon) > P(Y<\epsilon)$ and $\forall ~0< \sigma < \epsilon$, $P(X<\sigma) \geq P(Y<\sigma)$, where $X$ and $Y$ are two random variables following the distribution $f$ and $g$ respectively. Then, there always exists a threshold out-degree $\nu^*$ above which the minimum of at least $\nu^*$ realizations of the distribution with the larger minimal weight will precede the other one.\
In order to investigate this effect in a real-world setting, we construct a multiplex network as follows. We use a dataset of private messages sent on an online social network at the University of California Irvine [@Pietro2009]. This typical social network is then duplicated to create a hypothetical two-layer social network, where each layer can be thought of as corresponding to a different medium of communication. Each layer is thus identical, and is composed of $1899$ nodes and $20296$ edges. The difference between the layers is induced by the choice of two different distributions of $X$ and $Y$ defined in Eq \[distr\], where $Y\prec X$. As $X$ has a larger minimal weight than $Y$, there exists a switch in the precedence between the variables of type $X$ and $Y$. In this case, it is straightforward to show that this switch occurs when at least two edges of each type are competing, that is $$\begin{aligned}
\displaystyle\min_{i=1,\ldots,\nu}X_i \prec \displaystyle\min_{i=1 \ldots,\nu}Y_i~~ \forall \nu\leq2.\end{aligned}$$ Figure \[fig:switch\_precedence\] shows the probability of taking an edge of type $X$ instead of $Y$ with respect to the out-degree of a node in the static aggregated network (by construction, twice the out-degree in each layer), where each edge has two activation times associated to the random variable $X$ and $Y$ respectively. The numerical results confirm a switch of the precedence when the out-degree of the nodes increases.\
![ We display the average probability ($\pm$ the standard deviation) of taking an edge of $L_1$ versus one of $L_2$, with respect to the out-degree of each node on the replicate layers. We observe a switch in the layers precedence: as the nodes of out-degree equal to one tend to favor the layer $L_2$, nodes of out-degree $\geq2$ favor the layer $L_1$. The distribution of the waiting times of the edges of $L_1$ and $L_2$ are the distributions of the random variables $X$ and $Y$ respectively, defined in Eq \[distr\]. Dashed line corresponds to a probability equal to $0.5$. []{data-label="fig:switch_precedence"}](errorbar_precedence)
Impact of Inter-layer Heterogeneity on the Coverage of the Walker {#Mixing}
=================================================================
Finally, we investigate numerically the impact of competition in the case of diffusion in a multiplex social network with more than two layers. To do so, we use publicly available data introduced in [@Magnani2013] where the layers consist in five kinds of social relationships between employees. In the numerical simulations, we consider the three types of random variables defined in Eq.\[distr\] and focus on the layers associated to Facebook, Leisure and Co-authorship relations. The structure of the graph thus corresponds to real-world interactions, while the inter-events time are chosen for the sake of illustration. Indeed, the presence of three layers and the choice of the specific distributions allow the emergence of properties of precedence we have previously shown, such as a rock-paper-scissors situation, and thus enables the investigation of its impact on the random walk.
We compute numerically the exploration rate of a walker, or coverage, defined as the percentage of nodes visited by the walker over time [@Domenico2014]. We always assume homogeneous waiting time distribution inside a given layer, and consider homogeneous as well as heterogeneous inter-layer distributions.
A typical simulation is given in Figure \[fig:walk\_real\_crop\], where we observe that the exploration rate of a random walk tends to grow faster under inter-layer homogeneity compared to inter-layer heterogeneity, irrespective of the choice of the unique distribution. The slow-down induced by the inter-layer heterogeneity is mainly due to the fact that the flow is captured inside a single layer. Indeed, when graph density of the layers is large enough, one layer precedes the others two jointly, the one with edges of type $X$ in our example. Since the random walker has a tendency to stay in this layer, the walk will mainly take place on this layer, that is less connected than the aggregated network. However, when the graph density of the layers is weak, a rock-paper-scissors situation may arise, as the switch in precedence between layers might not occur for lower out-degree nodes. This emergence promotes the switch of the walker between different layers through lower out-degree nodes, resulting in a more efficient exploratory walk across the network, as illustrated in Figure \[fig:deg1\] where each node has one outgoing edge in each layer, ensuring low density and homogeneous out-degree distribution. In this case, the exploration rate of the walk is similar to the exploration rate of an homogeneous inter-layer network (or monolayer network), which is the most efficient in terms of coverage rate.
Discussion
==========
The main purpose of this work is to investigate the competition between different layers of a multiplex network in situations when the network is temporal. In our framework, edges activations are modeled as independent renewal processes, each layer being characterised by a different distribution, and we highlight implications of the concept of precedence on diffusion. In particular, we show that precedence may lead to biases between the different layers of the network. Despite the simplicity of the process, it may lead to counter-intuitive properties, such as non-transitivity, out-degree dependence, or rock-paper-scissors situation. Our numerical results also show that precedence may have important quantitative effects on the speed of diffusion on a multilayer network, as the precedence of one layer over others may hinder the number of edges available to the walker, and hence slow down its exploration of the graph. We also show that high out-degree nodes are more prone to favor one single layer, while low out-degree nodes exhibit a different effect, and may lead to a cyclic exploration between the layers. We study the impact of the precedence on an active random walk, however it is worth noticing that its impact on the passive random walk may lead to opposite bias towards layers. Indeed, in the presence of short cycles for the passive walker, an additional competition will emerge induced by the short cycles [@Petit2018]. For non-exponential distributions, this effect results in a backtracking bias towards or against the last traveled edges [@Gueuning2017], typically leading to the emergence of short cycle patterns in human-related network [@Saramaki2015].
These results remain mostly mathematical as we used toy-model distributions instead of ones modelled on real-life data, and an important next step would be to test the resulting ideas in empirical data of multiplex networks where different layers are associated to different time scales, for instance between physical, mobile phone and social media interactions [@sekara2016fundamental].
Acknowledgment {#acknowledgment .unnumbered}
==============
This work has been partially supported by the Concerted Research Action (ARC) supported by the Federation Wallonia-Brussels Contract ARC 14/19-060 ; and Flagship European Research Area Network (FLAG-ERA) Joint Transnational Call “FuturICT 2.0”.
[99]{} <span style="font-variant:small-caps;">Masuda, N., Porter, M. A. [&]{} Lambiotte, R.</span> (2017) Random walks and diffusion on networks. *Physics Reports*, **716-717**, 1–58.
<span style="font-variant:small-caps;">Holme, P. [&]{} Saramäki, J.</span> (2012) Temporal networks. *Physics Reports*, **519**(3), 97–125.
<span style="font-variant:small-caps;">Masuda, N. [&]{} Lambiotte, R.</span> (2016) *A guide to temporal networks*. World Scientific Publishing Europe Ltd.
<span style="font-variant:small-caps;">Delvenne, J.-C., Lambiotte, R. [&]{} Rocha, L. E. C.</span> (2015) Diffusion on networked systems is a question of time or structure. *Nature Communications*, **6**, 7366.
<span style="font-variant:small-caps;">Hoffmann, T., Porter, M. A. [&]{} Lambiotte, R.</span> (2013) Random Walks on Stochastic Temporal Networks. in *Understanding Complex Systems*, pp. 295–313. Springer Berlin Heidelberg.
<span style="font-variant:small-caps;">Starnini, M., Baronchelli, A., Barrat, A. [&]{} Pastor-Satorras, R.</span> (2012) Random walks on temporal networks. *Physical Review E*, **85**(5).
<span style="font-variant:small-caps;">Aleta, A. [&]{} Moreno, Y.</span> (2019) Multilayer Networks in a Nutshell. *Annual Review of Condensed Matter Physics*, **10**(1).
<span style="font-variant:small-caps;">De Domenico, M., Granell, C., Porter, M. A. [&]{} Arenas, A.</span> (2016) The physics of spreading processes in multilayer networks. *Nature Physics*, **12**(10), 901–906.
<span style="font-variant:small-caps;">Kivela, M., Arenas, A., Barthelemy, M., Gleeson, J. P., Moreno, Y. [&]{} Porter, M. A.</span> (2014) Multilayer networks. *Journal of Complex Networks*, **2**(3), 203–271.
<span style="font-variant:small-caps;">Magnani, M., Micenkova, B. [&]{} Rossi, L.</span> (2013) Combinatorial Analysis of Multiple Networks. *arXiv preprint*.
<span style="font-variant:small-caps;">Szell, M., Lambiotte, R. [&]{} Thurner, S.</span> (2010) Multirelational organization of large-scale social networks in an online world. *Proceedings of the National Academy of Sciences*, **107**(31), 13636–13641.
<span style="font-variant:small-caps;">Cardillo, A., G[ó]{}mez-Garde[ñ]{}es, J., Zanin, M., Romance, M., Papo, D., del Pozo, F. [&]{} Boccaletti, S.</span> (2013) Emergence of network features from multiplexity. *Scientific Reports*, **3**(1).
<span style="font-variant:small-caps;">Gallotti, R. [&]{} Barthelemy, M.</span> (2015) The multilayer temporal network of public transport in Great Britain. *Scientific Data*, **2**, 140056.
<span style="font-variant:small-caps;">De Domenico, M., Sol[é]{}-Ribalta, A., G[ó]{}mez, S. [&]{} Arenas, A.</span> (2013) Random Walks on Multiplex Networks. *CoRR*, **abs/1306.0519**.
<span style="font-variant:small-caps;">Domenico, M. D., Sole-Ribalta, A., Gomez, S. [&]{} Arenas, A.</span> (2014) Navigability of interconnected networks under random failures. *Proceedings of the National Academy of Sciences*, **111**(23), 8351–8356.
<span style="font-variant:small-caps;">De Domenico, M., Sol[é]{}-Ribalta, A., Omodei, E., G[ó]{}mez, S. [&]{} Arenas, A.</span> (2015) Ranking in interconnected multilayer networks reveals versatile nodes. *Nature communications*, **6**, 6868.
<span style="font-variant:small-caps;">Starnini, M., Baronchelli, A. [&]{} Pastor-Satorras, R.</span> (2017) Effects of temporal correlations in social multiplex networks. *Scientific Reports*, **7**(1).
<span style="font-variant:small-caps;">Ding, C. [&]{} Li, K.</span> (2017) Topologically biased random walk for diffusions on multiplex networks. *Journal of Computational Science*.
<span style="font-variant:small-caps;">G[ó]{}mez-Garde[ñ]{}es, J., de Domenico, M., Guti[é]{}rrez, G., Arenas, A. [&]{} G[ó]{}mez, S.</span> (2015) Layerlayer competition in multiplex complex networks. *Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences*, **373**(2056), 20150117.
<span style="font-variant:small-caps;">Jang, S., Lee, J. S., Hwang, S. [&]{} Kahng, B.</span> (2015) Ashkin-Teller model and diverse opinion phase transitions on multiplex networks. *Physical Review E*, **92**(2).
<span style="font-variant:small-caps;">Kleineberg, K.-K. [&]{} Bogu[ñ]{}[á]{}, M.</span> (2016) Competition between global and local online social networks. *Scientific Reports*, **6**(1).
<span style="font-variant:small-caps;">Gardner, M.</span> (1970) The paradox of the nontransitive dice and the elusive principle of indifference. *Scientific American*, **223**, 110–114.
<span style="font-variant:small-caps;">Gueuning, M., Delvenne, J.-C. [&]{} Lambiotte, R.</span> (2015) Imperfect spreading on temporal networks. *The European Physical Journal B*, **88**(11).
<span style="font-variant:small-caps;">Karsai, M., Jo, H.-H. [&]{} Kaski, K.</span> (2017) *Bursty Human Dynamics (SpringerBriefs in Complexity)*. Springer.
<span style="font-variant:small-caps;">Lambiotte, R., Tabourier, L. [&]{} Delvenne, J.-C.</span> (2013) Burstiness and spreading on temporal networks. *The European Physical Journal B*, **98**(7), 052307.
<span style="font-variant:small-caps;">Pietro, P., Tore, O. [&]{} M., C. K.</span> (2009) Patterns and dynamics of users’ behavior and interaction: Network analysis of an online community. *Journal of the American Society for Information Science and Technology*, **60**(5), 911–932.
<span style="font-variant:small-caps;">Petit, J., Gueuning, M., Carletti, T., Lauwens, B. [&]{} Lambiotte, R.</span> (2018) Random walk on temporal networks with lasting edges. *Phys. Rev. E*, **98**, 052307.
<span style="font-variant:small-caps;">Gueuning, M., Lambiotte, R. [&]{} Delvenne, J.-C.</span> (2017) Backtracking and Mixing Rate of Diffusion on Uncorrelated Temporal Networks. *Entropy*, **19**(10), 542.
<span style="font-variant:small-caps;">Saramäki, J. [&]{} Holme, P.</span> (2015) Exploring temporal networks with greedy walks. *The European Physical Journal B*, **88**(12).
<span style="font-variant:small-caps;">Sekara, V., Stopczynski, A. [&]{} Lehmann, S.</span> (2016) Fundamental structures of dynamic social networks. *Proceedings of the national academy of sciences*, **113**(36), 9977–9982.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on the universality of height fluctuations at the crossing point of two interacting $1+1$-dimensional Kardar-Parisi-Zhang interfaces with curved and flat initial conditions. We introduce a control parameter $p$ as the probability for the initially flat geometry to be chosen and compute the phase diagram as a function of $p$. We find that the distribution of the fluctuations converges to the Gaussian orthogonal ensemble Tracy-Widom (TW) distribution for $p<0.5$, and to the Gaussian unitary ensemble TW distribution for $p>0.5$. For $p=0.5$ where the two geometries are equally weighted, the behavior is governed by an emergent Gaussian statistics in the universality class of Brownian motion. We propose a phenomenological theory to explain our findings and discuss possible applications in nonequilibrium transport and traffic flow.'
author:
- Abbas Ali Saberi
- 'Hor Dashti-N.'
- Joachim Krug
bibliography:
- 'refs.bib'
title: ' Competing Universalities in Kardar$-$Parisi$-$Zhang Growth Models '
---
Scale invariant fluctuations play a central role in the emergence of universal properties in complex random systems interconnecting various areas of physics, mathematics and statistical mechanics. Whereas the concept of universality classes is well established in the theory of equilibrium phase transitions [@Binder1981], our understanding of systems driven out of equilibrium is much less complete [@Henkel2008]. The Kardar$-$Parisi$-$Zhang (KPZ) equation [@KPZ1986] governing the evolution of the surface height $h$(**x**,t), \[Eq1\]\_t h(**x**,t)=\^2 h+(h)\^2+(**x**,t), is a prototypical model for describing nonequilibrium growing interfaces with a wide range of theoretical and experimental applications [@Krug1991; @Stanley1995; @HHZ1995; @Krug1997]. The first term in (\[Eq1\]) represents relaxation of the interface caused by a surface tension $\nu$, the second describes the nonlinear growth locally normal to the surface, and the last term is uncorrelated Gaussian white noise in space and time with zero average $
\langle\eta(\textbf{x},t)\rangle=0 $ and $\langle
\eta(\textbf{x},t)\eta(\textbf{x}',t')\rangle=2D\delta^d(\textbf{x}-\textbf{x}')\delta(t-t')$, representing the stochastic nature of the growth process. One recovers the Edwards$-$Wilkinson equation for $\lambda=0$.
The universality class of randomly growing interfaces is usually characterized by the scaling exponents defined by Family$-$Vicsek scaling [@Family1985] i.e., $w^2(t,l)\sim t^{2\beta}
f(l/t^{\beta/\alpha})$, in terms of the second moment $w^2(t,l)$ of the height fluctuations at a measurement scale $l$ at time $t$, where $f(x)\rightarrow$ const as $x\rightarrow\infty$ and $f(x)\sim
x^{2\alpha}$ as $x\rightarrow 0$. Thus $w^2$ grows with time like $t^{2\beta}$ until it saturates to $l^{2\alpha}$ when $t\sim l^{\alpha/\beta}$. The universality class is characterized by the exponents $\alpha$ and $\beta$ (the roughness and the growth exponents, respectively), whose exact values for the KPZ equation are known only in 1+1 dimensions (1+1 D) as $\alpha=1/2$ and $\beta=1/3$.
In a series of pioneering works, it has been shown that the universality in various growth models belonging to the KPZ class holds beyond the second moment [@Krug1992; @Krug2010RMT; @Takeuchi2011]. Unexpectedly, the height fluctuations of the 1+1 D single-step model (SSM) [@Meakin1986SSM] grown from a point seed were found to be governed [@Johansson2000] by the Tracy-Widom (TW) distribution of the Gaussian unitary random matrix ensemble (GUE) [@Mehta2004]. Thereafter, it was reported [@prahofer2000statistical; @prahofer2000universal] that the radial 1+1 D polynuclear growth (PNG) model also follows the TW GUE distribution, and in addition, the Gaussian orthogonal ensemble (GOE) determines the universality of the 1+1 D KPZ growth models on a flat substrate [@prahofer2000universal]. Recently, exact solutions of the 1+1 D KPZ equation have confirmed the TW GUE distribution for the height fluctuations on the curved (wedge-like) [@sasamoto2010one; @amir2011probability] and the TW GOE distribution on the flat geometries [@calabrese2011exact]. The key question of interest in this Letter is how these two GOE and GUE universalities compete when two different 1+1 D KPZ growth models adopting the flat and curved geometries meet each other at a single common point (Fig. \[fig:fw\]).
![\[fig:fw\] (color online) Schematic of the crossing flat$-$wedge geometry with a single common site in the middle. ](fig1.pdf){width="0.8\columnwidth"}
The SSM is a solid on solid growth model in the KPZ class in which at each time step on a 1 D (flat or wedge-like) lattice of size $L$, one site $-L/2 \leq j <
L/2 $ is randomly chosen, and if it is a local minimum the height $h(j)$ is increased by $2$. The initial conditions at $t=0$ are $h^f_0(j) = [1-(-1)^j]/2$ and $ h^w_0(j)=|j|$ for the flat and wedge geometries, respectively. This definition guarantees that at each step, the height difference between two neighboring sites is $\pm 1$. The SSM is the growth model representation [@Rost1981] of the totally asymmetric simple exclusion process (TASEP) in 1 D, a paradigmatic model for driven transport of a single conserved quantity [@Krug2010RMT].
![\[fig:fw\_w2\] (color online) Main: Second moment of the height fluctuations at the crossing point of the flat-wedge geometry as a function of time for several $p$ from bottom to top. The dashed line shows the scaling prediction $w^2\sim t^{2\beta}$ for the 1+1 D KPZ equation with growth exponent $\beta=1/3$. All curves are shifted by a constant for ease of comparison. Inset: The crossover from 1+1 D KPZ scaling at earlier times to the Brownian motion (BM) statistics at long time limit for $p=0.5$. In order to clearly observe the crossover to the BM regime, the simulations for $p=0.5$ were carried out up to time $t=10^6$. ](fig2.pdf){width="0.8\columnwidth"}
Here we consider growth on two crossing flat-wedge substrates subject to the same growth rules but with an exception at the origin $\textbf{x}=\textbf{0}$, where the two geometries meet. The origin is the only site with four nearest neighbors, the heights of which have to exceed the height at $\textbf{0}$ by one for growth to take place. This Letter studies the statistics of the fluctuations of the height $h(\textbf{0},t)$ at the crossing point at time $t$. Here time is defined in terms of the number of deposition trials per lattice site, either successful or not. The initial conditions are set as mentioned above for each geometry so that $h^f_0(\textbf{0})=h^w_0(\textbf{0})=0$. Periodic boundary conditions are applied along both geometries. At each time step, one of the two flat or wedge crossing geometries is chosen with probability $p$—the only parameter in our study— and then a site $j$ is randomly chosen for the growth process. The flat geometry is chosen with probability $p$ and the wedge geometry with probability $1-p$. In the TASEP representation this corresponds to two single-lane exclusion processes which meet at an intersection. The growth rule at the origin implies that the particles on the two lanes are forced to cross the intersection simultaneously. TASEP-like traffic flow models with intersections have been studied before, but with different crossing rules and without considering the current fluctuations at the intersection [@Nagatani1993; @Ishibashi1996; @Foulaadvand2004; @Foulaadvand2007; @Belbasi2008; @Embley2009; @Hilhorst2012; @Raguin2013].
Let us first examine the Family$-$Vicsek scaling for the second moment of the height fluctuations at the origin i.e., $w^2(t)=\langle
h^2(\textbf{0},t)\rangle-\langle h(\textbf{0},t)\rangle^2$, for different values of $p$. As Fig. \[fig:fw\_w2\] demonstrates, all curves for $p\ne0.5$ follow the scaling law $w^2\sim t^{2\beta}$ with the growth exponent $\beta=1/3$ predicted for the $1+1$ KPZ equation. A remarkable observation is that for $p=0.5$ when both geometries are picked with equal probability, the variance of the height at earlier times behaves as in the KPZ class, but later it crosses over to the universality of the Brownian motion (BM) i.e., $w^2\sim t$, with Gaussian statistics (see below).
![\[fig:v\_inf(p)\] (color online) $v_\infty$ (main panel) and $\Gamma_n$ (inset) for the flat$-$wedge geometry as a function of $p$. ](fig3.pdf){width="0.8\columnwidth"}
Until now our analysis has revealed two interesting facts: First, the point with $p=0.5$ acts as a distinguished fixed point with a characteristic Gaussian statistics in the universality of Brownian motion, and, second, for $p\ne 0.5$ the statistics of the height fluctuations at the crossing point—despite the existence of four nearest neighbors—is compatible with that of the 1+1 D KPZ equation whose long time statistics converges to the TW GUE/GOE distribution depending on the narrow-wedge/flat initial condition. One might naively expect that for $p>0.5$ for which the flat geometry is chosen with higher probability, the height fluctuations would converge to the GOE statistics and for $p<0.5$ where the wedge geometry is more likely to be picked, they should be compatible with the GUE distribution. As we will show in the following, our results unveil exactly the opposite behavior. The local height of an 1+1 D KPZ interface is asymptotically given by the following relation [@Takeuchi2011], $$\label{eq:h}
h = v_\infty t + s_\lambda (\Gamma t)^{1/3} \chi,$$ where $s_\lambda=\operatorname{sgn}(\lambda)$ is the sign of the nonlinear parameter $\lambda$ in the KPZ Eq. (\[Eq1\]), $ v_\infty $ and $ \Gamma $ are non-universal parameters and $\chi$ is a stochastic variable with a universal TW distribution depending on the flat/wedge growth geometry. We estimate the parameter $v_\infty$ by extrapolating ${\langle h \rangle}/t $ versus $t^{-2/3}$, as an intercept in a linear regression in the $ t \to \infty $ limit, i.e., ${\langle h \rangle}/t = v_\infty +
s_\lambda \Gamma^{1/3} {\langle \chi \rangle} t^{-2/3}$ [@Krug1990]. We carried out extensive simulations to generate height profiles of SSM on the flat$-$wedge geometry of linear size $L=2^{13}$ up to time $t=2\times 10^{4}$ for several values $p=0.2$, $0.3$, $0.4$, $0.45$, $0.48$, $0.49$, $0.5$, $0.51$, $0.52$, $0.55$, $0.6$, $0.7$, $0.8$. For each dataset, an ensemble of $7\times10^5$ independent realizations have been generated.
As shown in Fig. \[fig:v\_inf(p)\], we numerically find a simple relation for $v_\infty$ as a function of the parameter $p$, $$\label{eq:min}
v_\infty(p)=\min(p, 1-p).$$ Contrary to the naive expectation, this implies that the substrate with the *smaller* growth probability dominates the coupled process. To see why this is so, recall that the asymptotic growth rate of a single 1+1 D SSM interface with periodic boundary conditions is given by $v_\infty = \frac{\gamma}{2} (1-u^2)$, where $\gamma$ is the rate of deposition attempts and $u \in [-1,1]$ is the surface slope [@Krug1992; @Krug2010RMT]. Because the growth rate is maximal at $u=0$, an SSM interface can lower its growth rate by developing a nonzero slope, but it cannot increase its growth rate beyond $\gamma/2$ [@Wolf1990; @Krug1997]. In the present setting $\gamma = 2p$ for the flat geometry and $\gamma = 2(1-p)$ for the wedge geometry, respectively. To accomodate a common growh rate at the origin, for $p < 0.5$ the flat interface grows at maximal speed $v_\infty = p$ whereas the wedge interface maintains a nonzero tilt $u = \sqrt{\frac{1-2p}{1-p}}$. For $p > 0.5$ the roles of the two substrates are interchanged and the initially flat interface becomes wedge-shaped (Fig. \[fig:ht\]).
![\[fig:ht\] (color online) Snapshots for the time evolution of the height profiles on the flat$-$wedge geometry for $t=0$ (left column), $t=200$ (second column), and $t=2000$ (right column) for $p=0.3$ (first row), $p=0.5$ (second row) and $p=0.7$ (third row) corresponding to the GOE, Gaussian (BM) and GUE universality classes, respectively. ](fig6.pdf){width="1\columnwidth"}
We next show that the dominance of the slower geometry extends also to the height fluctuations at the origin. In order to estimate the parameter $\Gamma$ in Eq. (\[eq:h\]) we define $g_n \equiv {\langle h^n \rangle}_c/s^n_\lambda t^{n/3} = \Gamma^{n/3} {\langle \chi^n \rangle}_c$, where $ {\langle \chi^n \rangle}_c $ denotes the $n$th cumulant of the random variable $\chi$. We write $\Gamma_n = [g_n/{\langle \chi^n \rangle}_c]^{3/n}$ for the value of $\Gamma$ estimated from the $n$th cumulant. All estimates have to give rise to the same value assuming that the cumulants of $\chi$ are those of the corresponding TW GOE or GUE distributions. To find the possible TW distributions, we use two dimensionless $\Gamma$-independent measures, i.e., the skewness $S=g_3/g_2^{3/2}$ and the kurtosis $K=g_4/g_2^2$, and compare them with those of the TW distributions. Figure \[fig:fw\_SK\] represents the most remarkable finding of our study: For $p<0.5$ the statistics of the height fluctuations of the crossing point in the wedge$-$flat geometry is determined by the TW GOE distribution, and, for $p>0.5$ it is governed by the TW GUE distribution. Therefore we adopt the corresponding cumulants of the TW distributions into the above relations to extract $\Gamma_n$. We find that all $\Gamma_n$ follow the same simple relation with $p$ as we found for $v_\infty(p)$, i.e., $\Gamma(p)=\min(p, 1-p)$—see the inset of Fig. \[fig:v\_inf(p)\]. The relation $\Gamma = v_\infty$ is a known property of the SSM [@Krug1992].
![\[fig:fw\_SK\] (color online) Skewness (main panel) and kurtosis (inset) for the flat$-$wedge geometry as a function of $p$. ](fig4.pdf){width="0.8\columnwidth"}
Now we can directly check for universality by comparing the height fluctuation distribution with the analytic TW predictions. For this, we define a new variable $q=(h-v_\infty
t)/s_\lambda (\Gamma t)^{1/3},$ and plot the rescaled distribution functions $P(q)$ for several values of $p$. Figure \[fig:fw\_P\_q\] shows an excellent agreement with the corresponding TW distributions for $p\ne 0.5$. The figure also shows the distribution function of height fluctuations for $p=0.5$ which is in perfect agreement with the Gaussian distribution.
![\[fig:fw\_P\_q\] (color online) Rescaled distribution functions of the height fluctuations for the crossing point of the flat$-$wedge geometry for several values of $p$ (symbols), compared with the TW GOE distribution for $p<0.5$, TW GUE distribution for $p>0.5$, and Gaussian distribution for $p=0.5$ (solid lines). ](fig5.pdf){width="0.9\columnwidth"}
The fact that the fluctuations at the crossing point are determined by the slowly growing interface can be most easily understood in the TASEP representation. The growth rule at the origin implies that a particle on the fast lane has to wait for a particle on the slow lane to appear before it can cross the intersection. Therefore the statistics of the crossing events is determined by the slower lane, and follows TW-GOE (TW-GUE) statistics for $p < 0.5$ ($p > 0.5$), respectively. Whereas the dynamics on the slow lane is asymptotically unaffected by the intersection, the particles on the fast lane effectively experience a blockage, which leads to the buildup of a density discontinuity across the origin. In the interface representation this implies the formation of a wedge (Fig. \[fig:ht\]).
The physics of inhomogeneous growth processes [@Wolf1990] and exclusion processes with a blockage [@Janowsky1994; @Basu2014; @Schmidt2015] is also key to understanding the emergent Gaussian statistics that we observe at $p=0.5$. Consider first a single, initially flat SSM interface where deposition attempts occur at unit rate at all sites except a single defect site with deposition rate $r$. This corresponds to a TASEP with a single slow ($r < 1$) or fast ($r > 1$) bond. Recent work has established that the defect induces a macroscopic inhomogeneity for any $r < 1$, whereas it is asymptotically irrelevant when $r > 1$ [@Basu2014; @Schmidt2015]. We have numerically studied the height fluctuations at the defect site, finding TW-GOE statistics for $ r > 1$ but Gaussian BM statistics for $r < 1$. The latter behavior can be rationalized within the directed polymer (DP) representation of the process, where the defect site extends to a defect line in space-time which pins the polymer when $r< 1$ [@HHZ1995; @Krug1997; @Basu2014; @Tang1993]. In the pinned phase the energy of the polymer, which translates into the height of the SSM surface, is the sum of uncorrelated contributions accumulated along the one-dimensional defect line, which satisfies a central limit theorem and therefore displays Gaussian statistics.
The crossing geometry at $p=0.5$ is similar to the SSM with a defect site, in the sense that deposition occurs at the same rate at all sites except for the origin, where it is enhanced by a factor of $r = 2$. By analogy with the 1+1 D SSM, one might anticipate the existence of a critical value $r_c$, such that the fluctuations display Gaussian BM statistics for $r < r_c$ and KPZ TW statistics for $r > r_c$. However, our simulations of a crossing flat-flat geometry with a variable deposition probability $r$ at the crossing point indicate that the critical point, which is at $r_c=1$ for the single lane problem, is shifted to large $r_c\rightarrow\infty$, introducing the BM statistics as the dominant process in the long-time limit for any $r$. This may reflect the dynamic nature of the defect: Even when $r$ is very large, a TASEP particle attempting to cross the intersection still has to wait for a particle on the second lane to arrive, which happens at unit rate irrespective of $r$. In marked contrast to the 1+1 D SSM, however, we observe BM statistics in the absence of a macroscopically tilted, wedge-like surface profile. To clarify the origin of this behavior, a DP representation of the crossing growth geometry would be needed.
To conclude, we have considered 1+1 D KPZ growth models on a weighted flat$-$curved geometry and analyzed the statistics of the height fluctuations at the crossing point. We found a rich and unexpectedly non-trivial phase diagram comprising, in addition to the known TW GUE/GOE phases, an emergent Gaussian BM phase at $p=\frac{1}{2}$. It is important to note that the dominance of the more slowly growing geometry in the SSM is linked to the fact that the coefficient $\lambda$ of the KPZ nonlinearity is negative in this case [@Krug1992; @Wolf1990]. When $\lambda > 0$, the argument based on the slope-dependence of the asymptotic growth rate $v_\infty$ predicts that the faster geometry determines the behavior, which implies that the phase diagram is reflected around the point $p=\frac{1}{2}$. We have indeed verified that simulations of the restricted-solid-on-solid (RSOS) model, which also has $\lambda < 0$, lead to the same phase diagram.
At the critical point $p=\frac{1}{2}$, the TASEP representation of the model relates to previous work on exclusion processes with intersections [@Foulaadvand2007; @Embley2009; @Raguin2013], with the seemingly innocuous modification that particles are forced to cross the intersection in a correlated manner. Our results suggest that this makes the transport across the intersections much more efficient, in that macroscopic density discontinuities do not appear, while a signature of the intersection is retained in the form of anomalously large, BM-type current fluctuations. Importantly, the correlated hopping of particles moving along perpendicular directions is a fundamental feature of any particle representation of higher-dimensional growth processes, which is enforced by the integrability condition on the height field [@Krug1991; @Odor2009]. As such, by introducing a single site with a two-dimensional growth environment into an otherwise one-dimensional setting, the model may provide an inroad for progress towards an understanding of the elusive 2+1 D KPZ problem [@HH2012].
*Acknowledgment.* We thank Andreas Schadschneider for useful discussions. A.A.S. would like to acknowledge support from the Alexander von Humboldt Foundation and partial financial support from the research council of the University of Tehran. J.K. was supported by the German Excellence Initiative through the UoC Forum *Classical and Quantum Dynamics of Interacting Particle Systems.*
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this review I consider modern theoretical models of coupled star–disk magnetospheres. I discuss a number of models, both stationary and time-dependent, and examine what physical conditions govern the selection of a preferred model.'
author:
- 'Dmitri A. [^1]'
title: Magnetic Interaction Between Stars And Accretion Disks
---
Introduction {#sec-intro}
============
In this paper I review recent theoretical progress in our understanding of magnetic interaction between Young Stellar Objects (YSOs), in particular, Classical T-Tauri Stars (CTTSs), and their accretion disks. That such interaction takes place, we have no doubt, as there is now ample observational evidence of strong (of order $10^3$ G) magnetic fields in these systems (e.g., ; ).
Most of the theoretical work on magnetically linked star–disk systems, both analytical and numerical, has focussed on examining the structure and role of a large-scale axisymmetric magnetic field with (at least initially) dipole-like topology (see Fig. \[fig-geometry\]). This field presumably arises due to the star’s internal magnetic dipole moment. Studying such a large-scale star–disk magnetosphere will also be the focus of this paper. I will thus ignore the effects of any small-scale intermittent loops that may be generated by the turbulent dynamo action in the disk.
![The general geometry of a magnetically-linked star–disk system.[]{data-label="fig-geometry"}](geometry.eps){width="8cm"}
While I am restricting myself to only large [*spatial*]{} scales, I will consider a variety of [*temporal*]{} scales. The shortest relevant time-scale is the rotation period, typically a few days for CTTSs; the longest time-scale is the accretion disk life-time, which can be $\sim 10^6$ years or more (; ). Among the models developed to date, there exist a dichotomy with respect to the system’s behavior on the rotation time-scale. More specifically, in some models a direct magnetic connection between the disk and the star is maintained in a stationary configuration, whereas in other models it is not. In this paper I will review both of these classes of models.
Before I proceed, I would like to list some of the most important questions related to the subject of this review:\
1) What physical parameters determine whether a direct star–disk coupling via a large-scale dipole-like magnetic field can be maintained on the rotation-period time-scale?\
2) If a quasi-stationary magnetically-coupled configuration does exist, what is its structure and how does it evolve on the longer (e.g., accretion) time-scale?\
3) If the magnetic link is disrupted, then what is the non-steady process? Are there periodic or quasi-periodic openings and closings of the field (due to magnetic reconnection) or there is a transition to a wind-supported permanent stationary open-field configuration without the link?\
4) Is it possible that both scenarios are possible under different physical circumstances?\
5) In either scenario, what are the effects of turbulent viscosity and magnetic diffusivity? And what is the role of winds and jets?\
6) What are the implications for the time-variability of the accretion flow and for the angular momentum and energy exchange? What are the observational consequences that would allow one to discriminate between the models?
Although I will not be able to answer all of these questions in this review, I will use them as the main guiding themes in my discussion.
Non-stationary models {#sec-non-stationary}
=====================
The study of magnetic interaction between YSOs and their disks has been pioneered by . He has successfully applied the model of , developed originally for accreting neutron stars, to explain several important observational features of T-Tauri stars, such as their relatively long rotation periods, UV-excesses, and inverse P-Cygni profiles. In this steady-state model, the stellar magnetic field penetrates the disk over a finite range of radii both inside and outside of the corotation radius $r_{\rm co}\equiv
(GM_*/\Omega_*^2)^{1/3}$, where $M_*$ is the mass of the star and $\Omega_*$ is its angular velocity. The spin-up magnetic torque due to the field lines connecting to the disk inside $r_{\rm co}$ is balanced by the spin-down torque by the lines connecting to the disk outside $r_{\rm co}$. K[ö]{}nigl has proposed that a magnetic field of $10^3$ G at the stellar surface (a value consistent with observations, see ) can disrupt the disk at a few stellar radii (but well inside the corotation radius) and channel the accretion flow to higher stellar latitudes. He also has estimated the typical time needed to bring the star into the spin equilibrium with the disk to be $\sim 10^5$ years, much shorter than the typical accretion time for CTTSs.
Although this model has been very successful in explaining many spectral and variability features, it has not considered the dynamics of the magnetic field itself. The presence of a strong magnetic field has just been inferred, but no equations governing the magnetosphere structure have been solved. It turns out that there exists a very robust mechanism that leads to the breaking of the magnetic link on the rotation-period time-scale. This presents a serious obstacle for all steady-state models and thus gives us the motivation to consider nonstationary models.
The basic idea can be explained as follows. Both the star and the disk are fairly good conductors and so the magnetic field can generally be considered frozen into them. In addition, they rotate with different angular velocities (except at $r_{\rm co}$). Therefore, the field lines are twisted by the differential rotation and toroidal magnetic flux is generated out of the poloidal flux. As the toroidal field builds up, the corresponding field pressure tends to push the field lines outward and inflate them. At first, the poloidal field structure changes very little, but, after the relative star-disk twist angle exceeds one radian or so, the field lines start to expand faster and faster (at an angle of $\sim 60^\circ$ with respect to the rotation axis) and tend to open up, thereby destroying the magnetic link between the star and the disk.
This opening process is essentially identical to a similar process that have been studied extensively in solar-corona context (e.g., , [-@Aly-1995]; , [-@Low-2001]; ; ). Indeed, the opening of coronal magnetic arcades, brought about by displacements of the field-line footpoints on the photosphere, is one of the leading mechanisms for Coronal Mass Ejections [@Low-2001]. A significant amount of work on this process has also been done in the accretion-disk context. Thus, in the force-free approximation, it has been shown, using both simple analytical and semi-analytical arguments (; ; ; ; ; ), and via numerical solutions of the force-free Grad–Shafranov equation [@UKL-1] that such an opening occurs at a finite twist angle (see for a review). In addition, full numerical 2D MHD simulations (without the force-free assumption) have demonstrated the opening process at work as a part of the overall cycle (, [-@Hayashi-2000]; , [-@Goodson-1999a]; ; ). Thus, at present there is no doubt that, if stellar dipole magnetic field penetrates a conducting disk over a wide range of radii, the twisting of the field lines will open them, thereby breaking the star–disk connection everywhere, with perhaps the exception of the inner disk region where the magnetic field is strong enough to make the matter corotate with the star. This process naturally results in a non-steady behavior, which has lead to the development of a number of time-dependent models. In addition, however, there exist several alternatives, leading to a small number of distinct stationary models. I shall discuss the non-steady models first.
After the field lines expand and effectively open up, a natural question to ask is: what happens next? Currently, the situation is not entirely clear and there is no unique answer to this question. There are two drastically-different possibilities that are most often discussed.
In the first scenario [@Lovelace-1995], developed in the neutron star context, once the field lines open, they stay open indefinitely. A steady state is then achieved, although it is very different from the original one, as the magnetic link between the disk and the star has been severed on most of the field lines (see Fig. \[fig-LRBK\]). One can identify three topologically-distinct regions in the magnetosphere: the stellar wind region (region I), where the field lines extend from the star to infinity, the disk wind region (region II), where the field lines return from infinity to the disk, and the closed field region (region III) — the remnant of the linked magnetosphere, where the field enforces corotation of the disk with the star. Thus, the configuration here is stationary, but the magnetic link extends only over a small part of the disk.
![The magnetic configuration of Lovelace et al. (1995).[]{data-label="fig-LRBK"}](LRBK.eps){width="8cm"}
In the second scenario, the situation is really time-dependent, with quasi-periodic cycles of field inflation and opening due to twisting followed by the field closing through reconnection and subsequent contraction back to the initial state. This picture has been suggested by and has subsequently been studied in extensive numerical simulations by a number of authors (, [-@Goodson-1999a]; ; ; ). Some recent observational results also seem to favor this point of view (e.g., ). Let us consider this most interesting scenario in more detail.
First, notice that in the scenario the poloidal magnetic field reverses across the separatrix between regions I and II. This makes the separatrix an obvious prospective site for reconnection. Indeed, the presence of a rather large anomalous or numerical resistivity has routinely lead to reconnection in the numerical simulations by @Hayashi-1996 ([-@Hayashi-1996], [-@Hayashi-2000]) and by @Goodson-1997 ([-@Goodson-1997], [-@Goodson-1999a]) and , .[^2]
It is also important to realize that most of the toroidal flux, generated in the twisting process, has now been evacuated radially to infinity (since the toroidal flux on an inflated flux tube is concentrated near the tube’s apex). Therefore, magnetic field in both open-field regions is essentially poloidal (it is exactly poloidal in the force-free framework, but some toroidal field may be present in the MHD-wind regime where matter inertia is important). This means that if magnetic reconnection does occur somewhere along the separatrix, the inner newly-reconnected field lines (connected to both the star and the disk) find themselves out of force-free balance: they have very strong poloidal-field tension that tries to pull them back towards the star but almost no toroidal-field pressure. As a result, in the absence of a powerful outgoing wind (see the discussion below), these inner reconnected field lines contract on the alfv[é]{}nic time-scale. If both reconnection (exhibited as a flare) and the subsequent contraction and relaxation occur quickly enough, then the resulting closed field lines have very little residual twist, similar to the original dipole-like state. This sets the stage for a new cycle. The continuing differential rotation gradually twists the lines up again and the whole sequence of events repeats, with the natural period of the order of the rotation period. As for the other, outer, newly-reconnected field lines, they, together with the apex regions of the field lines that are still expanding somewhere far away, form a toroidal plasmoid (in a sense, a flying spheromak). The closed magnetic flux surfaces comprising such a plasmoid have the shape of tori nested around a circular line (an O-point in poloidal projection). Each plasmoid is not magnetically connected to either the disk or the star and is out of equilibrium; it then just flies away. If the motion of these plasmoids is collimated towards the axis (i.e., if they are flying mostly vertically), then they can feed the jet, providing an explanation for the observed knotty jet structure [@Goodson-1999a]. The plasmoids will be ejected out with the time intervals equal to the opening/closing period. For CTTSs, however, this period is expected to be too short compared with the observed interval between jet knots [@Goodson-1999a].
Whereas the effective field-opening time is about a fraction of the rotation period, the time between opening and reconnection is not certain. It depends on the intricate details of the reconnection process and, in particular, is intimately related to the so-called reconnection-trigger, or sudden-onset, problem, very well known in studies of flares in solar physics [@Priest-1984]. Here is what it means in the context of our problem. As the field lines start to open, one by one, and the current sheet is formed along the separatrix, how long does one wait before reconnection starts? In other words, how much flux is opened before it is reconnected back? For example, one can imagine that reconnection is triggered as soon as the first few field lines have opened; then one will see the ejection of small islands, separated by the time it takes the critical amount of flux to open (much shorter than the rotation period). Or, in the opposite extreme, it may be that a large portion or all of the flux opens and only long after that reconnection somehow starts; then one will see finite-size plasmoids ejected with the time interval equal to the sum of the opening time (days) and the uncertain time delay before reconnection onset \[for example, in the simulations by the total cycle period was about a month\].
Also not clear is how much flux is reconnected in each event before the reconnection process shuts off. This question is important because it determines the size of the ejected plasmoids. Indeed, it may be that reconnection proceeds until a large fraction of the flux is reconnected, in which case there will be large-amplitude oscillations in all of the system’s parameters (, [-@Hayashi-2000]; ). Alternatively, reconnection may stop very quickly after it has begun, and then one will see very small plasmoids ejected but the larger-scale open-field structure will stay intact, as seen in numerical simulations by . In fact, the model of can be viewed as an extreme manifestation of this latter scenario. Indeed, in this model it has been assumed, without much discussion or argumentation, that there is no reconnection at all. One can actually bring forth some arguments in favor of this point of view.
In general, there is a competition between field-line closing via reconnection and field-line opening by the wind. Reconnection will be stopped if the wind flowing along the open field lines is so strong that any newly-reconnected field lines are swept open by it (B.C. Low, private communication). More specifically, if there were no flows along the unreconnected field lines (i.e., no wind), then a newly reconnected closed field line on the inner side of the reconnection region would contract rapidly, with the field-line apex moving out of the reconnection region towards the star with the poloidal alfv[é]{}n velocity $V_{A,\rm pol}$. However, if there is a background outflow such as a wind, then one has to add the velocity of this outflow. If the latter is larger than $V_{A,\rm pol}$, then the resulting apex motion will be directed outward, i.e., the field line will open again. Thus, I suggest the following physical criterion for determining when the re-closing of the open field lines via reconnection will occur. I propose that if the prospective reconnection site is located outside of the Alfv[é]{}n radius (along the separatrix field line), so that the wind there is super-alfv[é]{}nic with respect to the reconnecting poloidal field, then everything will be swept outward by the wind and hence reconnection will not take place and the magnetic link will not be re-established. One then will get a helmet-streamer configuration like that of . In the opposite case, reconnection will occur and will lead to the closing of (a portion of) the field lines, leading to a cyclic behavior.
To summarize, the system’s behavior depends on both the reconnection physics (one needs to know where and when reconnection will be triggered) and on the wind physics: one needs to have a model for the wind to determine everything self-consistently.
Steady-state models {#sec-stationary}
===================
I shall now switch gears and discuss the few existing steady-state models in which the magnetic link between the star and the disk remains unbroken. First, to maintain the link, one must find a mechanism that could stop the twisting process. One obvious possibility is the toroidal resistive slippage of the field lines with respect to the plasma in the disk. Let us examine this possibility in more detail and see under what conditions it can work.
The situation depends critically on the disk’s effective magnetic diffusivity (which we shall sometimes call the resistivity). Unfortunately, the value of this diffusivity is not very well known (e.g., ). Nevertheless, one can set a reasonable upper limit by assuming that it is caused by the same turbulence that facilitates angular momentum transport across the disk. Thus, one can set the magnetic diffusivity $\eta$ to be equal to the Shakura–Sunyaev kinematic viscosity: $\eta\sim \nu_{\rm turb}
=\alpha c_s h$, where $c_s$ is the speed of sound, $h$ is the disk half-thickness, and $\alpha \simeq 0.01-0.1$. This range of the $\alpha$-values is consistent with the results of numerical MHD simulations of the Magneto-Rotational Instability (e.g., ; ). It is also consistent with the level of MHD turbulence that is necessary for the ejection of disk winds, as follows from the work of combined with the results of .[^3] The effective magnetic diffusivity of this kind leads to the toroidal slippage of field lines with respect to the disk with the relative drift velocity $\Delta v_{\phi}=(\eta/h)|B_\phi/B_z|_d\sim\alpha c_s|B_\phi/B_z|_d$, where the subscript $d$ designates the disk’s surface. For a thin disk, $c_s/v_K \sim h/r \ll 1$; thus, the slippage velocity is usually much smaller than the differential rotation velocity $r\Delta\Omega(r)\equiv
r[\Omega_K(r)-\Omega_*]$. There are, however, two special circumstances when this is not so. They are very important as they point us toward the ways to get a steady-state configuration. I would like to stress, however, that both of these circumstances are somewhat unusual and hence the resulting steady states are not very natural. The first of the two schemes is realized when the field lines under consideration are very close to the corotation radius (so that $\Delta\Omega\ll\Omega_K$); it provides the conceptual basis for the model developed by (see also ; ; ). The second scheme requires a very large ratio $|B_\phi/B_z|_d$ and provides the basis for the model developed by and by .
Let us discuss the first scheme first. Usually, its main idea can be readily dismissed because for a steady state to exist globally, it must exist for all the field lines, and in the majority of models most of the magnetic flux crosses the disk a finite distance away from $r_{\rm co}$. However, the model developed by solves this problem by assuming that almost all of the magnetic flux is “trapped” and concentrated in the so-called X-region, a very close vicinity of the corotation radius (which the authors of the model call the X-point). This model is one of the most promising, well-developed, and sophisticated models and has gained a lot of popularity and observational support (e.g., ) over the last few years. The disk’s magnetosphere consists of three parts (see Fig. \[fig-shu\]). Field lines emanating from the inner part of the X-region connect to the star and form the magnetic funnel that directs the accretion flow. A second portion of the field lines also connects to the star but carries no mass flow; it forms what the authors call the dead zone. Finally, the remaining field lines, those emanating from the outermost part of the X-region, are open and carry the wind that plays a key role in removing the excess angular momentum from the disk. In addition, there are some open stellar field lines that extend from the star to infinity. Thus, the general topology of the poloidal magnetic field is similar to the helmet streamer configuration of . An important difference however is that in the model, the poloidal flux is spread smoothly over the entire disk surface and the corotation radius plays no special role, whereas in the Shu model the flux is concentrated close to the X-point, with almost no field at $r>r_{\rm co}$. Because of that, the differential rotation in the latter model is weak ($\Delta\Omega \ll \Omega_*$), and even a small disk resistivity is sufficient to eliminate the twisting and hence to ensure a steady state. In addition, inside of the corotation radius, the disk density drops rapidly as the plasma is uplifted to form the funnel flow; as a result, the corotation with the star is enforced by the strong magnetic field there.
![Schematic diagram of the X-wind configuration in the Shu et al. (1994) model.[]{data-label="fig-shu"}](shu.eps){width="8cm"}
The inner radius of the disk in the Shu model essentially coincides with the X-point \[see, however, , who put it at $r_{\rm in}=0.74 r_{\rm co}$\]. As the accreting matter enters the X-region, some part of it is loaded onto the open field lines and forms the outgoing X-wind. The rest of the plasma gradually diffuses through the magnetic field in the X-region and is loaded onto the funnel-region field lines and then falls onto the star. At the same time, most of the angular momentum of this falling matter is taken away by the magnetic field and is transported back to the inner portion of the disk, while only a small fraction ends up on the star. This provides an effective control mechanism for the star’s spin and suggests a plausible explanation of the relatively-long rotation periods of CTTSs [@Shu-1994a].
This very interesting model is not free of its own problems and inconsistencies, however. Thus, for example, it is highly unlikely that the poloidal field on the inner side of the X-region will not diffuse towards the star. Indeed, these field lines are very strongly bent so that there is a highly concentrated current, essentially an equatorial current sheet, between the X-point and the inner edge of the disk \[i.e., the kink-point of \]. Any small amount of resistivity will then cause the field to slip inward through this current layer. As for the plasma flow, it will not be able to counter this diffusion because it is in the same direction. Thus, the resistive-MHD Ohm’s law immediately tells us that this configuration cannot be in a steady state.
The Shu model has also been criticized by . He pointed out that their solution requires very fine tuning, so that the inner disk radius (determined by the balance between the stellar magnetic field and the accretion flow) is equal to $r_{\rm co}$ (determined by $\Omega_*$). Hartmann considers this situation unacceptable, citing an example of DR Tau, where accretion rate has been observed to change on the time scale far too short for $\Omega_*$ to adjust. He seems to favor the viewpoint and concludes that “the entire magnetosphere might be a complicated, time dependent structure”. He also emphasizes the importance of magnetospheric reconnection, noting that it “should lead to substantial heating and flare activity” [@Hartmann-1997].
Now let us consider the second possibility for a steady state. In this scenario, the balance between the twisting due to the differential rotation and the turbulent resistive slippage is made possible by a very large ratio of the toroidal to vertical magnetic field at the disk surface, of the order of $\alpha^{-1}
r/h \gg 1$. Such high values are usually considered to be unlikely. Indeed, the angle between the field lines and the disk is determined by the entire solution in the magnetosphere and cannot be arbitrary. The density of matter above the disk is typically so low that magnetic forces completely dominate the dynamics there. In this force-free regime, the toroidal field at the disk surface, $B_{\phi,d}$, increases in proportion to the twist at first, but then reaches a maximum and starts to decline during the rapid-expansion phase; it goes to zero as the field approaches the open state. The maximum value of $B_{\phi,d}$, achievable in a force-free magnetosphere, depends sensitively on the way the poloidal magnetic flux is distributed across the disk, that is on the function $\Psi_d(r)$. Usually, as it turns out, this maximum value is of the same order as the vertical magnetic field and hence the minimum angle between the disk and the projection of the magnetic field vector on the $\theta-\phi$ plane is of order one. In this case, the differential rotation cannot produce the required very large values of the disk toroidal field. The primary physical reason for this is that most of the toroidal magnetic flux, which is being continuously generated by twisting, becomes concentrated near the field-line apex (i.e., the farthest from the star point on a field line). As the field expands, it becomes energetically favorable for the toroidal flux to escape to infinity by opening the poloidal field lines. Coming back to the question of the effects of the disk resistivity, we see that, with the disk toroidal field limited by the opening and flux-escape process, the toroidal resistive slippage, even in a turbulent disk, cannot be fast enough to significantly affect the twisting process.
On the other hand, for a certain class of functions $\Psi_d(r)$, the maximum value $|B_\phi/B_z|_{\rm d, max}$ of the ratio of the toroidal to vertical field components at the disk surface, allowed by the force-free solution in the magnetosphere, can be large. In particular, if $\Psi_d(r)\sim r^{-n}$, then $|B_\phi/B_z|_{\rm d,max}
\sim O(1/n)$ in the limit $n\rightarrow 0$ (; ; ; ).
One can then picture the following evolutionary scenario. Let us start with a non-steady cyclic configuration such as that described by , . During the first part of the cycle, as the field expands and approaches the open state, the field lines at the disk surface are inclined away from the star \[i.e., $(B_r/B_z)_d>0$\] and hence diffuse a little bit outward.[^4] Then, during the second part of the cycle, as the reconnected field contracts back to the nearly potential state, the field lines may be inclined towards the star at the disk surface \[i.e., $(B_r/B_z)_d<0$\]; they will then diffuse inward. (Note also that, for the field lines inside $r_{\rm co}$, such an inclination is conducive to loading of matter onto the field lines; thus, accretion can take place during this phase of the cycle.) Then, one can ask what happens on a much longer time scale, when the radial diffusion of magnetic field has to be included. Here, two possibilities immediately come to mind.
It may be, as suggested by , that the little diffusive displacements will, over time, redistribute the disk’s magnetic flux so that the net displacement over one cycle will become zero. Then, an averaged steady state will be established, i.e., the cycles of field opening, reconnection, and closing will produce no net secular evolution in the magnetic flux distribution. Since the amount of the outward radial displacement depends on the exact moment of reconnection, it follows that the physics of reconnection again plays a crucial role in determining the long-term magnetic flux distribution.
On the other hand, as I have discussed above, if $\Psi_d(r)\sim r^{-n}$, then $|B_\phi/B_z|_{\rm d,max} \sim O(1/n)$ in the limit $n\rightarrow 0$. Thus, it is in principle possible for the system to achieve exact, not time-averaged, steady state if the disk’s flux redistributes in such a way that the corresponding value of $n\equiv -d\ln \Psi_d/d\ln r$ becomes very small. Since the value of $|B_\phi/B_z|_d$, necessary for the balance between the differential rotation and toroidal resistive slippage, is inversely proportional to the disk’s effective resistivity, we see that in this case the disk flux distribution $\Psi_d(r)$ is essentially determined by the resistivity. In the case of Shakura–Sunyaev turbulent resistivity, one can obtain the following upper limit: $ n < C |B_\phi/B_z|_{\rm d, max}^{-1} \sim \eta/rh\Delta\Omega =
O(\alpha h/r) \ll 1$, where $C$ is a finite number. Note also that for a steady state to be maintained, one must worry not only about the toroidal direction, but also about the radial direction. This requirement gives not just an upper limit, but in fact determines implicitly the entire function $n(\eta)$, or, more generally, the dependence $\Psi_d(r)[\eta(r)]$ (; ). Such a stationary field configuration, possible in principle, is very different from the dipole field; in particular, it leads to a dramatic decrease in the torque between the star and the disk[@Agapitou-2000].
Summary {#sec-summary}
=======
In conclusion, I would like to give the following approximate list of the major theoretical approaches to the problem of magnetically-coupled star–disk magnetospheres:
1\) Very rich non-stationary scenario (; ; , [-@Hayashi-2000]; , [-@Goodson-1999a]; ; , [-@UKL-2]; ) with cycles of field inflation, opening, reconnection, contraction, and accretion. Both accretion and outflows occur intermittently, with variability on the differential rotation period (or somewhat longer) time-scale. The amplitude of these oscillations (e.g., how much poloidal flux is opened and then reconnected in each cycle) depends strongly on the physics of reconnection and is not very well constrained. For example, the steady-state model of can be considered a limiting case where no reconnection takes place at all, and thus the oscillation amplitude is zero.
2\) The steady-state X-wind model of . The model of can be considered a bridge model between the Goodson and Shu models.
3\) Finally, a steady-state closed magnetosphere with the poloidal vertical magnetic field that threads the disk scaling as $B_z(r)\sim r^{-[2+O(\eta)]}$ — models of and of . These models take into account the field’s radial diffusion in the disk over a long (compared with $\Omega_*^{-1}$) time scale.
At present, it is apparently too early to select one of these models as the preferred one based on purely theoretical considerations. More rigorous theoretical work, in conjunction with more sophisticated and thorough numerical simulations and comparison with observations, is needed to sort things out.
I am deeply indebted to Arieh K[ö]{}nigl and Christof Litwin for many, many insightful and productive discussions, and to Ana G[ó]{}mez de Castro for her very thoughtful comments and useful suggestions. I am also very grateful to the organizers of the International Workshop on Magnetic Fields and Star Formation (Madrid, April 21–25, 2003) for the invitation to write this review. This research was supported by the National Science Foundation under Grant No. PHY99-07949.
Agapitou, V. and Papaloizou, J. C. B.: 2000 [*MNRAS*]{}, [**Vol. no. 317**]{}, p. 273.
Aly, J. J.: 1984, [*ApJ*]{}, [**Vol. no. 283**]{}, p. 349.
Aly, J. J.: 1995, [*ApJ*]{}, [**Vol. no. 439**]{}, p. L63.
Aly, J. J., and Kuijpers, J.: 1990, [*A&A*]{}, [**Vol. no. 227**]{}, p. 473.
Bardou, A., and Heyvaerts, J.: 1996, [*A&A*]{}, [**Vol. no. 307**]{}, p. 1009.
Bouvier, J., Alencar, S. H. P., and Dougados, C.: 2003, preprint (astro-ph/0306553).
Brandenburg, A., Nordlund, A., Stein, R. F., and Torkelsson, U.: 1996, [*ApJ*]{}, [**Vol. no. 458**]{}, p. L45.
Fendt, C.: 2002, preprint (astro-ph/0211601).
Ferreira, J.: 1997, [*A&A*]{}, [**Vol. no. 319**]{}, p. 340.
Ferro-Fontan, C., and Gomez de Castro, A. I.: 2003, [*MNRAS*]{}, [**Vol. no. 342**]{}, p. 427.
Ghosh, P., and Lamb, F. K.: 1978, [*ApJ*]{}, [**Vol. no. 223**]{}, p. L83.
Goodson, A. P., Winglee, R. M., and B[ö]{}hm, K.-H.: 1997, [*ApJ*]{}, [**Vol. no. 489**]{}, p. 199.
Goodson, A. P., B[ö]{}hm, K.-H., and Winglee, R. M.: 1999, [*ApJ*]{}, [**Vol. no. 524**]{}, p. 142.
Goodson, A. P., and Winglee, R. M.: 1999, [*ApJ*]{}, [**Vol. no. 524**]{}, p. 159.
Guenther, E.W., Lehmann, H., Emerson, J.P., and Staude, J.: 1999, [*A&A*]{}, [**Vol. no. 341**]{}, p. 768.
Hartmann, L.: in [*Herbig-Haro Flows and the the Birth of Low Mass Stars*]{}, eds. B. Reipurth and C. Bertout, Kluwer, Dordrecht, 1997, p. 391.
Hayashi, M. R., Shibata, K., and Matsumoto, R.: 1996, [*ApJ*]{}, [**Vol. no. 468**]{}, p. L37.
Hayashi, M., Shibata, K., and Matsumoto, R.: 2000, [*Adv. Space Res.*]{}, [**Vol. no. 26**]{}, p. 567.
Johns-Krull, C.M., Valenti, J.A., and Koresko, C.: 1999, [*ApJ*]{}, [**Vol. no. 516**]{}, p. 900.
Johns-Krull, C.M., Gafford, A.D.: 2002, [*ApJ*]{}, [**Vol. no. 573**]{}, p. 685.
Kenyon, S. J., and Hartmann, L.: 1995, [*ApJ Suppl.*]{}, [**Vol. no. 101**]{}, p. 117.
K[ö]{}nigl, A.: 1991, [*ApJ*]{}, [**Vol. no. 370**]{}, p. L39.
Lovelace, R. V. E., Romanova, M. M., and Bisnovatyi-Kogan, G. S.: 1995, [*MNRAS*]{}, [**Vol. no. 275**]{}, p. 244 (LRBK).
Low, B. C.: 1977, [*ApJ*]{}, [**Vol. no. 212**]{}, p. 234.
Low, B. C.: 2001, [*Journal Geophys. Res.*]{}, [**Vol. no. 106**]{}, p. 25,141.
Lynden-Bell, D., and Boily, C.: 1994, [*MNRAS*]{}, [**Vol. no. 267**]{}, p. 146.
Lynden-Bell, D.: 2003, [*MNRAS*]{}, [**Vol. no. 341**]{}, p. 1360.
Matt, S., Goodson, A.P., Winglee, R.M., and B[öhm]{}, K.-H.: 2002, [*ApJ*]{}, [**Vol. no. 574**]{}, p. 232.
Miki[ć]{}, Z., and Linker, J. A.: 1994, [*ApJ*]{}, [**Vol. no. 430**]{}, p. 898.
Najita, J. R., and Shu, F. H.: 1994, [*ApJ*]{}, [**Vol. no. 429**]{}, p. 808.
Ostriker, E. C., and Shu, F. H.: 1995, [*ApJ*]{}, [**Vol. no. 447**]{}, p. 813.
Priest, E. R.: 1984, [*Solar Magnetohydrodynamics*]{}, Reidel, Dordrecht, 1984.
Shu, F., Najita, J., Ostriker, E., Wilkin, F., Ruden., S., and Lizano, S.: 1994, [*ApJ*]{}, [**Vol. no. 429**]{}, p. 781.
Shu, F. H., Najita, J., Ruden., S. P., and Lizano, S.: 1994, [*ApJ*]{}, [**Vol. no. 429**]{}, p. 797.
Stone, J. M., Hawley, J. F., Gammie, C. F., and Balbus, S. A.: 1996, [*ApJ*]{}, [**Vol. no. 463**]{}, p. 656.
Uzdensky, D. A., K[ö]{}nigl, A., and Litwin, C.: 2002, [*ApJ*]{}, [**Vol. no. 565**]{}, p. 1191.
Uzdensky, D. A., K[ö]{}nigl, A., and Litwin, C.: 2002, [*ApJ*]{}, [**Vol. no. 565**]{}, p. 1205.
Uzdensky, D. A.: 2002, [*ApJ*]{}, [**Vol. no. 572**]{}, p. 432.
Uzdensky, D. A.: 2002, [*ApJ*]{}, [**Vol. no. 574**]{}, p. 1011.
van Ballegooijen, A. A.: 1994, [*Space Sci. Rev.*]{}, [**Vol. no. 68**]{}, p. 299.
[^1]: KITP preprint NSF-KITP-03-83
[^2]: We also have to mention that were sceptical about the possibility of reconnection, but this is because they had used the self-similar model describing a uniformly-rotating disk. In that model finite-time field-opening occurs without current-sheet formation along the separatrix. In a more realistic case of non-uniform rotation, has argued that there will be finite-time partial field opening accompanied by asymptotic thinning of the separatrix current-concentration region, which can be regarded as current-sheet formation. As the current layer becomes thinner and thinner, a reconnection process may be triggered by anomalous resistivity or the Hall effect.
[^3]: The conditions for launching MHD winds from accretion disks are found to be close to those necessary for the operation of the Magneto-Rotational Instability [@Ferro-Fontan-2003].
[^4]: with the initial velocity of order the rms turbulent velocity in the disk, $v_{\rm turb}\sim \alpha c_s$; hence the characteristic radial footpoint displacement over a rotation period scales as $\Delta r\sim v_{fp}\Delta\Omega^{-1} \sim
\alpha c_s/\Delta\Omega \sim \alpha h \Omega_K(r)/
\Delta\Omega \ll r$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an efficient analytical method to predict the maximum transit timing variations of a circumbinary exoplanet, given some basic parameters of the host binary. We derive an analytical model giving limits on the potential location of transits for coplanar planets orbiting eclipsing binaries, then test it against numerical N-body simulations of a distribution of binaries and planets. We also show the application of the analytic model to Kepler-16b, -34b and -35b. The resulting method is fast, efficient and is accurate to approximately 1% in predicting limits on possible times of transits over a three-year observing campaign. The model can easily be used to, for example, place constraints on transit timing while performing circumbinary planet searches on large datasets. It is adaptable to use in situations where some or many of the planet and binary parameters are unknown.'
author:
- '\'
bibliography:
- 'papers\_17042013.bib'
date: 'Accepted 2013 July 1. Received 2013 June 28; in original form 2013 May 15'
title: Placing Limits on the Transit Timing Variations of Circumbinary Exoplanets
---
\[firstpage\]
(stars:) planetary systems; (stars:) binaries: eclipsing
Introduction {#sectintro}
============
To date seven transiting circumbinary exoplanets have been discovered, all from the NASA Kepler mission data [@Doyle:2011ev; @Welsh:2012kl; @Schwamb:2012ts; @Orosz:2012ip; @Orosz:2012ku]. Such circumbinary planets provide interesting tests of planet formation theories, having formed in a complex environment. Recently several studies have been performed on their formation [e.g. @Gong:2012il; @Pelupessy:2012gl; @Meschiari:2012ts; @Meschiari:2012er], orbital stability [e.g. @Jaime:2012br; @Doolin:2011ib; @Pichardo:2008et; @Pichardo:2005fb] and variations in insolation from a habitability perspective [e.g. @Kane:2012fl; @OMalleyJames:2012jf]. They are beginning to be subjected to analytical models, such as that provided by @Leung:2013fm (hereafter LL). Such planets are valuable objects for our understanding of planetary formation and evolution, with further discoveries needed to provide observational constraints on these interesting and complex systems. Their signal may be present in datasets from other transit surveys such as WASP [@Pollacco:2006gb] or NGTS [@Wheatley:2013uw]. Detecting these planets via the transit method presents observational challenges, as they exhibit transit timing variations (TTVs) on the order of days in magnitude, in addition to changes in the shape and duration of transits.
The purpose of this paper is to present constraints on the observational characteristics of a transiting circumbinary exoplanet through our knowledge of the host binary system, using a fast method which requires no complex modelling. In this way we aim to aid detection through reducing the problems generated by the large scale TTVs mentioned above. Specifically we address TTVs in coplanar circumbinary systems, placing general limits on the magnitude of such variations, through constraining the location of possible transits. A similar analysis was carried out for the system KIC002856960 [@Armstrong:2012ie; @Lee:2013ee], which shows similar large scale TTVs, and multiple transits per orbit, albeit in a triple star scenario. These constraints are of use to surveys for such planets, where we can place limits on and aid the design of new automated searches, such as the QATS algorithm [@Carter:2013bg]. While it is possible with numerical simulations to predict exact times of transit for circumbinary systems, our analytical model allows (under some approximations) constraints to be placed on systems where some or many orbital parameters are not yet known, including the majority of eclipsing binaries in the Kepler Eclipsing Binary Catalogue [@Prsa:2011dx; @Slawson:2011fg].
TTVs on the transits of circumbinary planets have two main sources. The first is a geometrical timing variation (we refer to this as Effect I) resulting from the changing positions of the host binary stars. This leads to a range in time in which transits can occur, similar to more ‘usual’ TTVs, and is derived in Section \[sectGTVderiv\]. The second is a precessional variation (referred to as Effect II), a long term oscillation in time around a constant periodicity of the potential location of transits, caused by precession of the planet’s orbit (which is itself caused by torques arising from the non point mass nature of the binary). It is treated in Section \[sectPTVderiv\]. There are other possible sources of TTVs, such as another planet in the system. The effect of such a planet, or any other known source of TTVs, is negligible compared to the above in circumbinary systems (c.f. Kepler-47b,c where planet-planet interactions are negligible [@Orosz:2012ku]).
We make use of several unusual terms in this paper, and define them here for clarity. First, a ‘crossing’, or ‘crossing region’. This is the region of a circumbinary exoplanet’s orbit where the planet crosses the binary star orbit, from the observer’s perspective. It may only transit the stars within this crossing region, but will generally spend most of its time in the region out of transit. Second we use extensively the ‘azimuthal’ period of a circumbinary planet, mentioned in LL. There are several periods which may be relevant to a circumbinary planet, and we make use of two here - the azimuthal period and the Keplerian period. The azimuthal period is the period which on average the planet takes between successive alignments with the observer, i.e. to traverse $2\pi$ radians relative to a fixed reference vector and plane. The Keplerian period is an osculating period taken at a particular epoch, derivable from Kepler’s third law via the binary mass and planet semi-major axis. These two periods are not equivalent, and are discussed further in Section \[sectdiscuss\].
The structure of the paper is as follows. Section \[sectkepapprox\] describes a Keplerian approximation which can be used to estimate the possible location of transits for a general planet and binary. Section \[sectnummod\] describes the implementation of a numerical model used to test this approximation, with application to a demonstration simulated system. Section \[sectresults\] shows the results of testing the analytical model against a distribution of binaries and planets modelled numerically, and applies the models to Kepler-16b, -34b and -35b. Section \[sectdiscuss\] discusses the accuracy and usefulness of these results, as well as discussing observational issues in the search for circumbinary transiting exoplanets.
Models
======
Analytic Approximation {#sectkepapprox}
----------------------
We present here a derivation which allows the potential location of transits of a circumbinary planet to be estimated without the need for detailed modelling or any free parameters. It proceeds using Keplerian orbital equations for both the stars and planets of a circumbinary system, and hence is an approximation only, as it does not consider three-body effects that perturb the orbits of the binary and planet (although precession of the planet’s argument of periapse is included). We consider the TTVs of transits of only one star at a time, through this paper star 1. To consider transits of star 2, swap the indices 1 and 2 in Equation \[betaeqn\].
### Geometrical Timing Variations - Effect I {#sectGTVderiv}
These variations arise from the movement of the binary stars within their orbit. As such we use the limits of this orbit, coupled with the time the planet takes to cross said orbit. We make use of an equation for the duration of a transit in a single star/planet system (Equation \[durationeqn\], from @exoseagerdur, their Equation 14). A crossing (defined in Section \[sectintro\]) of a circumbinary planet is analogous to the transit of a single star by a planet passing in front of it; conceptually, we just replace the single star with a ‘metastar’ of diameter equal to the maximum extent of the binary’s orbit, giving
$$\label{durationeqn}
T_{\textrm{GTV}} = \frac{P_\textrm{p}}{\pi}\arcsin\left(\frac{R_{\textrm{metastar}}}{a_\textrm{p}}\right) \frac{\sqrt{1-e_\textrm{p}^2}}{1+e_\textrm{p}\sin(\omega_\textrm{p})},$$
where $ T_{\textrm{GTV}}$ represents the duration of the crossing, subscript p represents the planet, $P$ the azimuthal period, $a$ the semi-major axis, $e$ the eccentricity and $\omega$ the argument of periapse. We have made the approximation that the impact parameter $b_\textrm{p}\ll R_{\textrm{metastar}}$, the inclination of the planet $i_\textrm{p} = \pi/2$ and $R_\textrm{p}\ll R_{\textrm{metastar}}$. To find $R_{\textrm{metastar}}$ we must derive the extent of the binary’s orbit, projected onto the sky.
Consider the eclipsing binary orbit to be in the x-z plane, with the z axis being along the line of site of the observer. By doing this we take the binary orbit to have inclination $\pi/2$, a reasonable approximation for detached eclipsing binaries and for this purpose. Take the motion of star 1 in the x plane, projected onto the sky. From @exoseagerskyproj (their Equation 53, with $\Omega = 0$), this is given by
$$\label{skyprojx}
X = \beta(f)a_\textrm{b},$$
where $$\label{betaeqn}
\beta(f_\textrm{b}) = \frac{M_2}{M_1+M_2}\frac{(1-e_\textrm{b}^2)}{1+e_\textrm{b}\cos(f_\textrm{b})}\cos(\omega_\textrm{b}+f_\textrm{b}),$$ and subscript b represents the binary, $f$ the true anomaly and $M_{1,2}$ the mass of stars 1 and 2 respectively. Taking the zero points of the differential with respect to $f_\textrm{b}$ of Equation \[skyprojx\] gives us the minimum and maximum values of $X$ - the extent of the star’s motion projected onto the sky. The values of the true anomaly of the binary at these points are given by
$$\label{minmaxf}
f_0,f_1 = \arcsin[-e_\textrm{b}\sin(\omega_\textrm{b})] - \omega_\textrm{b}.$$
Equation \[minmaxf\] has two solutions within the range $0,2\pi$. Inserting both into Equation \[skyprojx\] gives the maximum and minimum values for $X$. We term these $X_1$ and $X_0$. Which of $X_0$ and $X_1$ is the minimum and which the maximum depends on $\omega_\textrm{b}$, but is unimportant here.
The radius of the ‘metastar’ is given by
$$\label{rmetstareqn}
R_{\textrm{metastar}} = \frac{\vert X_1\vert + \vert X_0\vert}{2},$$
and a scaled radius by $$\label{rscaledeqn}
R_{\textrm{m,scaled}} = \frac{R_{\textrm{metastar}}}{a_\textrm{b}} = \frac{\vert \beta(f_1)\vert + \vert \beta(f_0)\vert}{2}.$$ Substituting Equation \[rmetstareqn\] into Equation \[durationeqn\] leads to
$$\label{phaserangeeqn}
T_{\textrm{GTV}} = \frac{P_\textrm{p}}{\pi}\arcsin\left[R_{\textrm{m,scaled}}\left(\frac{P_\textrm{b}}{P_\textrm{p}}\right)^{\frac{2}{3}}\right] \frac{\sqrt{1-e_\textrm{p}^2}}{1+e_\textrm{p}\sin(\omega_\textrm{p})},$$
where the ratio of semi-major axes has been substituted to the equivalent ratio of periods using Kepler’s third law, allowing the use of the azimuthal period outlined in Section \[sectintro\]. In the presented form $ T_{\textrm{GTV}}$ represents the duration of a crossing, and as such a range of time within which transits can occur. The argument of periapse, $\omega_\textrm{p}$, is a function of time due to precession of the planetary orbit; assuming a constant precession rate it can be estimated analytically using Equation 5 of @Doolin:2011ib, hereafter DB, which is derived from that of @Farago:2010ev.
Lacking knowledge of the present system alignment, it is possible to take a ‘safe’ approximation by using the value of $\omega_\textrm{p}$ which gives the maximum $ T_{\textrm{GTV}}$, i.e. $\omega_\textrm{p} = 3\pi/2$. This corresponds to when the planet transits near its apoapse, and hence is travelling relatively slowly so that the range of transit times is extended. Using this constant value of $ T_{GTV}$ is often more practical. For systems with low planetary eccentricity the variation caused by varying $\omega_\textrm{p}$ is small (on the order of a few percent in $ T_{GTV}$).
### Precessional Timing Variation - Effect II {#sectPTVderiv}
This variation is caused by the precession of the planet’s orbit. For an eccentric planetary orbit, this precession will result in shifts in the time of potential transits away from the ‘expected’ time for a constant periodic signal. The magnitude of these shifts at a given time depends on the instantaneous value of $\omega_\textrm{p}$.
We assume a constant precession rate for the planetary orbit, such that
$$\label{eqnomrate}
\frac{d \omega_\textrm{p}}{dt} = \frac{2\pi}{P_{\omega}},$$
where $P_{\omega}$ represents the period of precession of the planet’s periapse, and can be estimated analytically through the equation of DB.
For a planet precessing in the prograde direction, this change in $\omega_\textrm{p}$ represents time ‘gained’, a portion of its orbit which it does not have to cover before aligning with the observer once more. The differential amount of time saved (i.e. period shifted) in this way is given by
$$\label{eqntimeloss}
\frac{dP_\textrm{p}}{d \omega_\textrm{p}} = \frac{dt}{df_\textrm{p}},$$
where $dP$ represents an apparent change in the period of the planet, and $f_\textrm{p}$ is the true anomaly of the planet, with $dt/df_\textrm{p}$ evaluated at $f_\textrm{p}=\pi/2 - \omega_\textrm{p}$, the value of $f_\textrm{p}$ at transit conjunction.
There are two contributions here, a constant term from the precession and a varying oscillation induced by the effect of the eccentricity of the planet’s orbit. The constant term can be found simply, by realising that the planet ‘loses’ one full orbit of time in one precessional period. For a constant precession rate, this gives a constant rate of time loss of $P_\textrm{p}/P_{\omega}$, which must be subtracted from Equation \[eqntimeloss\] to find the oscillation term. When using the azimuthal period of the planet (as defined in Section \[sectintro\]), or searching observationally for transits this constant term is automatically accounted for, which is why it must be removed here.
Continuing the derivation, we take the standard Keplerian orbital equation for $df_\textrm{p}/dt$ [@exoseagerdfbydt their Equation 32] evaluated at $f_\textrm{p} = \pi/2 - \omega_\textrm{p}$, $$\label{eqndfbydt}
\frac{df_\textrm{p}}{dt} = \frac{2\pi}{P_\textrm{p}} \frac{[1+e_\textrm{p}\sin(\omega_\textrm{p})]^2}{(1-e_\textrm{p}^2)^{\frac{3}{2}}},$$ where we have approximated $P_\textrm{p} \simeq P_\textrm{p}(1+1/P_{\omega})$. Combining Equations \[eqnomrate\] and \[eqntimeloss\] gives us the oscillation term
$$T_{\textrm{PTV}} = \int_{t_0}^{t}dP_\textrm{p} = \int_{t_0}^{t}\left(\frac{dt}{df_\textrm{p}} - \frac{P_\textrm{p}}{P_{\omega}}\right) d \omega_\textrm{p},$$
which, after inserting Equation \[eqndfbydt\], becomes $$\label{eqnPTV}
T_{\textrm{PTV}} = -\frac{P_\textrm{b}}{P_{\omega}} \int_{t_0}^{t} \left(\frac{(1-e_\textrm{p}^2)^{\frac{3}{2}}}{(1+e_\textrm{p}\sin[\omega_\textrm{p}(t)])^2} - 1\right) dt,$$ where the negative sign accounts that this is time gained or equivalently an apparent shortening of the planetary period, and applies for prograde precession. The quantity $ T_{\textrm{PTV}}$ represents an oscillation of the location of possible transits with time. We give an example of its effect through application to a demonstration simulated system in Section \[sectnummod\].
### Combined TTV Limits - Practical Use {#sectcombinedlimits}
Equations \[phaserangeeqn\] and \[eqnPTV\] can be combined to provide limits on the TTVs of transiting coplanar circumbinary planets. At a given epoch, $ T_{\textrm{PTV}}$ represents the offset around some zero point that the range of possible transit times would be centred around, whereas $ T_{\textrm{GTV}}$ represents the extent of the range around this offset. We present constraints here for practical use, in the situation where one or more transits have been detected, and limits need placing on the times of as yet undetected transits. A period must be estimated, either from the separation of two transits (or fractions of this) or by using a succession of trial periods. In the case where only one or two transits are known, as we do not know where in the possible transit range the transit falls we must use double the range to cover all possible times, giving the following limits: $$\label{eqntmin}
t_{\textrm{min}} (i) = t_0 + iP_\textrm{p} + T_{\textrm{PTV}}(t_0 + iP_\textrm{p}) - T_{\textrm{GTV}}(t_0 + iP_\textrm{p})$$ and $$\label{eqntmax}
t_{\textrm{max}} (i) = t_0 + iP_\textrm{p} + T_{\textrm{PTV}}(t_0 + iP_\textrm{p}) + T_{\textrm{GTV}}(t_0 + iP_\textrm{p}),$$ where $t_0$ represents the time of first transit, and $i$ an index for the orbit under consideration (each orbit may contain more than one transit, though in practice this is unusual). The quantities $t_{\textrm{min}}$ and $t_{\textrm{max}} $ represent the minimum and maximum times between which possible undetected transits must fall within on each orbit, for the case of one or two known transits.
Over short ($\ll P_{\omega}$) timescales $T_{\textrm{GTV}}$ is the dominant contribution (in some systems, such as those with low eccentricity planets, it is always so), and $T_{\textrm{PTV}}$ may be neglected. Using the maximum possible value of $T_{\textrm{GTV}}$ (by setting $\omega_\textrm{p}=3\pi/2$ in Equation \[phaserangeeqn\]) provides a ‘safe’ (in that the result will always be an overestimate) way of neglecting the time and $\omega_\textrm{p}$ dependence of $T_{\textrm{GTV}}$. Similarly, if little is known about a proposed circumbinary system, parameters in the above equations can be easily approximated with only small and quantifiable errors introduced.
The effects of $T_{\textrm{GTV}}$ and $T_{\textrm{PTV}}$ are shown for a demonstration circumbinary planet in Section \[sectdemo\].
Numerical Model {#sectnummod}
---------------
### Approach
We use a numerical model to test the above analytical framework. The N-body equations of motion were integrated using a fourth-order Runge-Kutta algorithm. Since this integrator does not inherently conserve energy, the total system energy was calculated over time to ensure that it was conserved such that the energy loss fraction remained below approximately $10^{-7}$.
To calculate the azimuthal period numerically we averaged the time intervals between the planet passing each of the two boundaries of the projected star orbit. The azimuthal period is the mean of these two averages. An alternative method is to average the interval between system centre of mass crossing times, which will converge to the same value but more slowly because it is only based on one crossing point, not two. Over time, the average interval between consecutive transits will converge to the azimuthal period.
### Demonstration {#sectdemo}
We here apply the numerical and analytical models to a simulated system (chosen from the simulations of Section \[sectresults\] as a system with a typical error) to demonstrate the effects of the derived timing variations. This system has a binary star with period 14.1 days, eccentricity 0.13, stellar masses of 1.22 and 1.07 $M_\odot$ and argument of periapse 282.3 deg. The planet has azimuthal period 191.5 days and eccentricity 0.16, leading to a precessional period for the planet of 84.2 years from the numerical model. Figure \[t1t2demo\] shows the potential locations of planetary transits derived from the numerical model, using times of potential transit phase-folded at the above azimuthal period. Potential transits must occur on each solid line. The variations seen are discussed in Sections \[T1var\] and \[T2var\].
### Geometrical Timing Variations - Effect I {#T1var}
The Effect I geometrical timing variations introduced in Section \[sectintro\] and derived in Section \[sectGTVderiv\] arise from the significant motion of the host binary stars. The planet can take several days to traverse the full extent of the binary orbit, and it is during this time that transits will occur. The TTVs, given by Equation \[phaserangeeqn\], can therefore be very large. By considering circular orbits and Solar-mass stars, Equation \[phaserangeeqn\] can be approximated by $T_{\textrm{GTV}} \approx (P_\textrm{p} P_\textrm{b}^2)^{1/3}/(2\pi)$, with periods in days, which demonstrates the size of the TTVs and their period dependence.
This geometrical contribution to the TTVs corresponds to the length of the lines in Figure \[t1t2demo\]. The magnitude of the Effect I term itself oscillates with the precession period of the planet, due to the changing speed of the planet at crossing, as different regions of its eccentric orbit line up with the observer.
### Precessional Timing Variations - Effect II {#T2var}
The other variation, an oscillation in phase or equivalently oscillation in apparent period, is due to the precession of the planet causing transits to correspond to different phases, as seen in Figure \[t1t2demo\](right). The oscillation is in particular caused by the changing instantaneous effect of the precession on a planet in an eccentric orbit. This is different to the contribution of precession in the Effect I geometrical case, which varies $T_{\textrm{GTV}}$ due to the changing planetary velocity. The Effect II precessional variation becomes significant over timescales approaching the planetary precessional period, typically decades. The amplitude of this variation is strongly dependent upon the planetary precessional period and eccentricity.
Results {#sectresults}
=======
Setup
-----
The accuracy of the model of Section \[sectkepapprox\] was tested using the numerical model (Section \[sectnummod\]) applied to a simulated distribution of 1000 single-planet circumbinary systems, 799 of which were stable over 1200 years (longer than the maximum planetary precession period found, and significantly longer than the majority). A more thorough stability analysis was not deemed necessary for the purposes of testing the equations in this paper. The binary star periods and eccentricities were taken from @Halbwachs:2003, which presented an unbiased distribution taken from radial velocity surveys, expanding upon the work of @Duquennoy:1991wk. The primary star masses were taken from the Kepler catalog of all stars monitored, using an empirical calibration from Torres [@Torres:2010eoa] to calculate the mass based on the metallicity, effective temperature and log $g$. The secondary star mass was determined using the mass ratio distributions found in @Halbwachs:2003, for binaries with periods less than and greater than 50 days. The radii of the stars were unimportant for this test.
For the planets, since no circumbinary planet distribution is known as yet, the period and eccentricity distributions were taken from data for planets orbiting single stars. Only radial velocity data were used to avoid the bias towards small periods seen in transit surveys. The planet was taken as a massless test particle, as its mass has a minimal effect on the dynamics. The planet radius was also unimportant for this simulation, as it has no effect on the dynamics. For each circumbinary system the minimum planet period was four times that of the binary, as a rough stability constraint [@Holman:1999aa], although some systems still proved to be unstable (particularly those with high eccentricities). The maximum planet period was set at 500 days, long enough that TTVs in such systems are unlikely to be of interest in the near future. All systems were exactly coplanar. Each of these systems was integrated numerically over its expected precession period (calculated from the equation of DB) with a time step of 30 minutes. The system’s azimuthal period was then calculated from the time it took the planet to orbit the system centre of mass on average.
To test the analytical model we used Equations \[phaserangeeqn\] and \[eqnPTV\] to predict the limits on possible transit time of the simulated planets. The precession period was split into three-year baselines (chosen as the length of a representative observing campaign). At each of the three-year baseline for each system, the predicted and numerical limits were initially aligned (as would be the case when detecting the first transit of a candidate planet) and then the system and predicted limits were allowed to evolve. At each crossing, the deviations between the upper analytical and numerical limits and lower analytical and numerical limits were averaged, and the same averaged for all crossings within each of the three-year baselines.
Test Results
------------
Results are represented as a percentage of the numerically integrated crossing time found at each planetary crossing. As such, an error of 100% represents analytically predicted transit limits which are misaligned by one crossing time on average. Figure \[figtest\] shows the histogram of percentage errors found for the 799 stable systems. The peak shows an error of 0.4%. The median error is 0.84%. For clarity, 43 systems are not shown in Figure \[figtest\]. These represent badly predicted single systems, with percentage errors higher than 20% (four of them have errors over 100%). These larger error systems are discussed in Section \[sectdiscuss\].
Application to Kepler-16b, -34b and -35b
----------------------------------------
The numerical model was applied to the known systems Kepler-16b, -34b and -35b, and times of possible transit were extracted. We find azimuthal periods of 227.06, 283.13 and 127.30 days for -16b, -34b and -35b, respectively. These are slightly offset from those found by LL. These are compared to Keplerian periods from the respective discovery papers of 228.78, 288.82 and 131.46 days [@Doyle:2011ev; @Welsh:2012kl]. We note that care must be taken regarding the different reference frames parameters for these planets can be published under, and also regarding the instantaneous and highly variable nature of many of the usual planetary parameters. Figure \[phasevar\] shows the potential locations of planetary transits derived from the numerical model, using times of potential transit phase-folded at the above azimuthal periods. Potential transits must occur within the thick band for each planet. The thickness of each band represents the Effect I, geometrical timing variation, and the oscillation in phase of the band represents the Effect II, precessional variation. The amplitude of this Effect II variation is strongly dependent upon the planetary precessional period and eccentricity. The period of the Effect II oscillations is equivalent to the planet’s precessional period, [.17ex]{}48, [.17ex]{}63 and [.17ex]{}21 years for Kepler-16b, -34b and -35b, respectively.\
Whilst the previous large-scale test used massless test particles, in this application the planet masses were included in the N-body code. To demonstrate that the planet mass has only a small effect on the transit times, we also simulated the Kepler planets with zero mass. The transit times of the mass and massless simulations were compared over a 200 year period in blocks of three years. On average the difference, as a percentage of the TTV range, was only 3.6%, 0.8% and 0.7% for Kepler-16b, -34b and -35b, respectively. Future work may include updated equations that incorporate the planet mass, but this will likely only be beneficial for very massive circumbinary planets.
A typical three-year region is shown for each planet in Figures \[16dbtest\], \[34dbtest\] and \[35dbtest\], with the analytical model prediction for each crossing. We note the slight secondary oscillation in Figure \[16dbtest\]. This is an additional dynamical effect likely due to a non-Keplerian effect of the host binary, and is stronger for Kepler-16b than for -34b or -35b. We do not attempt to predict this effect in this work.
Discussion {#sectdiscuss}
==========
Overview
--------
We have derived and validated a fast and simple to implement framework for placing limits on the possible locations of transits for a transiting coplanar circumbinary exoplanet. These variations can be split into two parts - Effect I, geometrical, caused by the changing positions of the binary stars as they orbit, and Effect II, precessional, caused by the long term precession of the planet’s orbit. With the likelihood of future searches for circumbinary exoplanets high, and the possibility of discovering such planets in the extant data from previous surveys, being able to place limits on the location in time of potential signals is particularly useful.
Accuracy {#sectdiscussacc}
--------
Figure \[figtest\] shows the accuracy of Equations \[phaserangeeqn\] and \[eqnPTV\] in predicting the possible times of transit of coplanar circumbinary planets - a median percentage error of 0.84% of the planet crossing time across the test set of 799 stable systems, over three years of observations. This can be used as an error when using Equations \[eqntmin\] and \[eqntmax\] to predict possible times of transit, where the percentage error should be applied to both $t_\textrm{max}$ and $t_\textrm{min}$. We note that our stated errors depend on the time baseline covered - they will be reduced for baselines lower than three years, and increased for those higher. The stated errors should, however, be indicative for a general observing campaign. Limitations on the accuracy arise primarily from non-Keplerian effects (beyond simple constant precession of the planet’s orbit, which we account for). This is demonstrated by the 43 systems with errors greater than 20%, including four with errors greater than 100%. These, and the scattered systems found at over 5% in Figure \[figtest\], are systems which appear to be stable but which show strong dynamical effects we have not accounted for, such as shorter period additional oscillations of $\omega_\textrm{p}$ or other effects we do not investigate here. The underlying dynamics behind these are beyond the scope of this paper. Encouragingly, it seems that such effects are strong only in a small minority of cases - the analytical model missed the possible transit range entirely in only 0.5% of the tested stable systems.
Applications
------------
We anticipate that Equations \[phaserangeeqn\] and \[eqnPTV\] (and in practice Equations \[eqntmin\] and \[eqntmax\]) will prove useful particularly for current and future searches for circumbinary planets. They provide a link between our theoretical knowledge of a circumbinary planetary system and the observational transit signatures which may arise from it, without requiring complex modelling or N-body integrations. This can be used to place limits on the potential transit times of candidate planets around a binary star, for the purpose of constraining searches for the transits of unknown planets. As a specific example, Equations \[eqntmin\] and \[eqntmax\] can be used to set the parameters $\triangle_{\textrm{min}}$ and $\triangle_{\textrm{max}}$ in the QATS search algorithm [@Carter:2013bg]. Importantly, this analytic framework can be used on systems where detailed knowledge of the component stars and orbital parameters is lacking, something impossible for N-body models. Full use of Equations \[eqntmin\] and \[eqntmax\] requires knowledge of the binary system, specifically the individual stellar masses, binary orbital eccentricity, argument of periapse and binary period, as well as the argument of periapse and eccentricity of the planet (while the planetary period is involved, we envisage that for general searches for unknown planets a series of trial periods would be used). Lacking some or all of these details, it is possible to make useful conclusions through using simplifying assumptions - taking $M_2\ll M_1$ for example removes the need for knowledge of the stellar masses while only overestimating the Effect I timing variation limit by at most a factor of two (i.e. placing loose but still useful limits on transit timing).
We repeat that it is possible in general to neglect the time and $\omega_\textrm{p}$ dependent part of $T_{\textrm{GTV}}$ (by setting $\omega_\textrm{p} = 3\pi/2$, as the value which gives the maximum value of $T_{\textrm{GTV}}$). It is also possible for low eccentricity planets to neglect $T_{\textrm{PTV}}$. This makes Equations \[eqntmin\] and \[eqntmax\] simply a constant limit on the TTV of a planet. These additional terms are however included in this work so that they can be utilised if necessary, particularly for highly eccentric planets or those whose precessional periods approach the baseline of observations used. Note that Equations \[eqntmin\] and \[eqntmax\] represent double the range of transit times predicted by Equations \[phaserangeeqn\] and \[eqnPTV\], as it would not be known where in this range a first detected transit fell.
These equations are also useful in reverse, for making first estimates of planet parameters using the observed transit variations of a newly discovered planet candidate. In this situation, the planet azimuthal period must be estimated using the mean transit interval. With this, the maximum observed transit timing variation around this period can be obtained. Neglecting $T_{\textrm{PTV}}$, this represents a lower limit on Equation \[phaserangeeqn\]. In the situation where the binary period, eccentricity and argument of periapse are known through the binary light curve, this gives a constraint on a combination of the planet eccentricity, argument of periapse and the binary mass ratio. The geometrical contribution $T_{\textrm{GTV}}$ is only weakly dependent on the planet eccentricity for moderate eccentricities, so by setting $e_\textrm{p}=0$ an approximate lower limit can be found on the binary mass ratio (for $e_\textrm{p}=0.2$ this approximation has an error of at most [.17ex]{}20%, depending on the precise value of $\omega_\textrm{p}$). Conversely, if the lower limit on $T_{\textrm{GTV}}$ found from the observed transits is especially high for the known binary parameters, this is an indication of high planetary eccentricity. Such constraints can be of use when attempting to find best fitting orbital solutions for these systems.
In the situation where for example only a few transits are detected, and the orbital solution is degenerate or poorly constrained (such that N-body integration is unfeasible), these expressions can be used for placing limits on the time period for which an object should be surveyed from the ground to detect future transits. This makes such follow up work much more efficient, and becomes relevant when continuous space based observations are not available.
Observational Considerations
----------------------------
We summarise here some issues which have become apparent affecting observational searches for circumbinary exoplanets. While this paper aims to reduce the difficulty caused by TTVs, these other limitations to detection of circumbinary planets remain and should be noted:
**Azimuthal Period** This is the time which on average the planet takes to traverse $2\pi$ radians in a fixed reference frame - i.e. the time interval between successive conjunctions. It is offset from, for example, the Keplerian period which can be derived from the planet’s semi-major axis and the binary mass. In LL it is shown that the azimuthal period is shorter than the Keplerian orbital period for circumbinary planets. The effect of this can be seen in many of the published transiting circumbinary planets so far. If we take the observed times of transit of these planets and estimate a period from the mean transit interval (which is equivalent to the azimuthal period), the estimated period is generally found to be a few days under the published Keplerian period. This is not an error, but a mark of the difference between the azimuthal period and Keplerian period that LL mention. The effect is clear for Kepler-16b: The maximum TTV at the published Keplerian period (228.78 days) is [.17ex]{}13 days, but at the azimuthal period we find (225.72 days), it is [.17ex]{}4.5 days, significantly lower. This azimuthal period is the important quantity when considering circumbinary planets from an observational perspective.
**Non-Coplanarity** If a circumbinary planet is not close to coplanarity with its host binary (such that it is within a few degrees of the binary orbital plane), then due to the motion of the binary stars it will often ‘miss’ them while crossing, exhibiting transits only on some orbits and again making detection much less likely. This constraint is relaxed for binary stars where the mass of one star is much greater than that of its companion (such that the more massive star’s orbit is smaller than its radius) or for contact binaries. Furthermore, for systems that are not exactly coplanar, the precession of the planetary orbit will take it in and out of a transiting configuration. This is the case for Kepler-16, where the transits across the larger star A are predicted to cease in early 2018 and return in approximately 2042.
**Eccentricity** As part of the source of TTVs of circumbinary exoplanets is due to precession of the planet’s orbit, highly eccentric planets will show more variations. While this does not reduce their detection chances as much as the above points, it increases the difficulty caused by these variations, further ‘blurring’ the planet’s transit signal. The ‘blurring’ effect of eccentricity is then scaled by the period of the precession of the planet’s orbit. Planets that precess faster will experience more transit timing variations over a given timescale.
Conclusion {#sectconc}
==========
1. There are two key contributions to the timing variations affecting transits of circumbinary planets. These are geometrical, Effect I, from the motion of the binary stars, and precessional, Effect II, from the precession of the planet’s orbit. Other contributions, from for example other planets in the system, are generally on the order of minutes or less in amplitude and negligible compared to these.
2. We have derived and validated an analytic framework to quickly estimate each of these terms, for a planet coplanar with its host binary.
3. This can be used to place limits on the location of possible transits. In particular, the equations can be approximated using minimal knowledge of the system (in contrast to a more detailed numerical integrator), making them useful for searching datasets for transits of such planets or in reverse making first estimates of parameters using the observed transit variations. Specifically, full use of the equations require the individual stellar masses, binary eccentricity, argument of periapse and binary period, as well as the period, argument of periapse and eccentricity of the planet. It is simple to approximate the parameters or use trial values where necessary, as described at various points above.
4. We have also summarised some observational issues which have become clear affecting the prospects of detection of circumbinary planets.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors are grateful for the useful comments of an anonymous referee, which helped improve the paper. D. V. Martin is funded by the Swiss National Science Foundation. A. H.M.J. Triaud is a Swiss National Science Foundation fellow under grant number PBGEP2-145594.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We analyze the various interstellar components of the region Sh2-132. The main stellar source is the double binary system that includes the Wolf-Rayet star WR153ab. We use radio continuum images at 408 and 1420 MHz, and 21cm line data taken from the Canadian Galactic Plane Survey, molecular observations of the $^{12}$CO(1-0) line at 115 GHz from the Five College Radio Astronomy Observatory, and available mid and far IR observations obtained with the MSX and IRAS satellites, respectively.
Sh2-132 is composed of two shells showing radio continuum counterparts at both frequencies. The emission is thermal in nature. The estimated rms electron density and ionized mass of the nebula are $n_e \simeq 20\,\rm cm^{-3}$ and $M_{\rm HII}\simeq 1500 M_{\odot}$. The distribution of the CO emission shows molecular gas bordering the ionized nebula and interacting with it. The velocities of the molecular gas is in the range –38 to –53 , similar to the velocity of the ionized gas.
The emission at 8.3 $\mu$m reveals a ring like feature of about 15that encircles the bright optical regions. This emission is due to the PAHs and marks the location of photodissociation regions.
The gas distribution in the environs of Sh2-132 can be explained in a scenario where the massive stars in the region photodissociated, ionized, and swept-up the dense molecular material from the parental cloud through their strong stellar winds and intense UV photon flux.
author:
- |
J. Vasquez$^{1,2}$[^1], C.E. Cappa$^{1,2}$, S. Pineault$^{1,3}$ and N.U. Duronea$^{1,2}$\
$^{1}$Instituto Argentino de Radioastronomia, CCT-La Plata, CONICET, C.C.5., 1894, Villa Elisa, Argentina\
$^{2}$Facultad de Ciencias Astronómicas y Geofísicas, Universidad Nacional de La Plata, La Plata, Argentina\
$^{3}$Départment de physique, de génie physique et d’optique and Centre de recherche en astrophysique du Québec (CRAQ),\
Université Laval, Québec, Canada GIVOA6
date: 'Accepted 1988 December 15. Received 1988 December 14; in original form 1988 October 11'
title: 'Ionized gas, molecules, and dust in Sh2-132'
---
\[firstpage\]
ISM: bubbles – ISM: molecules – ISM: regions – ISM: individual: Sh2-132 – stars: Wolf-Rayet – stars: individual: WR153ab
Introduction {#intro}
============
The interaction between massive stars and the interstellar medium leads to the formation of regions, interstellar bubbles, and new stars. More specifically, the interaction between the molecular medium and the UV stellar radiation is at the origin of photodissociation regions (PDRs). Molecules are dissociated by UV photons in the range 6 eV to 13.6 eV. The gas component ranges from ionized atoms at the PDR surface to molecules deeper into the molecular region as the FUV flux is absorbed. The FUV flux illuminating these regions range between 1$\le G_0 \le$ 10$^6$, where $G_0$ is the intensity of the radiation field in the solar neighborhood (Habing 1968), while the density lies between 10 $\le n_H \le$ 10$^5$ (Tielens & Hollembach 1985, Sternberg & Dalgarno 1989, Sternberg & Dalgarno 1995, Bakes & Tielens 1998). PDRs are formed at the edge of regions, planetary nebulae, reflexion nebulae, and stellar forming regions. Forbidden spectral lines in the far infrared as well as molecular lines are observed towards PDRs (e.g. Hogerheijde et al. 1995, Kaufman et al. 1999).
An important component of PDRs is interstellar dust. Its presence plays a major role absorbing stellar radiation and reemitting it in the mid and far infrared. Polycyclic aromatic hydrocarbons (PAHs), which emit at 3.3, 6.2, 7.7, 8.6, and 11.3 $\mu$m, are tracers of PDRs (e.g. Watson et al. 2008).
In this paper, we study the interstellar environment of the region Sh2-132 (Sharpless 1959), located in the Perseus spiral arm, paying particular attention to the presence of PDRs at the interfase between the ionized and molecular gas.
Sh2-132 was detected in H$\alpha$, \[O III\], and \[S II\] lines by Heckathorn et al. (1982). The nebula is centered approximately at $(l,b)\simeq$ (10250, –042) and presents two bright regions, which are surrounded by diffuse emission regions. They can be identified in Fig. 1, which displays the DSS-R image of the nebula.
Churchwell & Walmsley (1973) performed radio continuum observations at 2695 MHz with an angular resolution of 182, and, using measurements at several frequencies, derived a thermal spectral index. Harten et al. (1978) investigated the characteristics and morphology of the nebula based on radio continuum observations at 610 MHz using the WSRT with a synthesized beam of 60. Most of the emission at this frequency originates from two optically bright regions named Shell A and Shell B, located at $(l,b)\simeq$ (10257, and $(l,b)\simeq$ (10247, –042). Both shells coincide with the bright optical regions within the nebula and are indicated in Fig. 1.
According to Harten et al. (1978), the optically bright section of Sh2-132 (at $l>$10243) is an ordinary region excited by the massive O-type star BD+552722 and the WR star WR153ab (= HD211853), which are located close to the center of the nebula and linked to Shell B. Chu et al. (1983) reached to similar conclusions, based on the presence of the stars near the center of the nebula and on the absence of velocity changes in the nebular spectra. The main parameters of these stars, as well as of other OB stars listed by Harten et al. (1978) that appear projected onto the region, are summarized in Table 1, which lists their galactic coordinates, spectral classification, visual magnitude $V$, colour indices ($B-V$) and ($B-V$)$_0$, visual absorption $A_V$, absolute magnitude $M_V$, and derived spectrophotometric distance $d$. Uncertainties in the derived spectrophotometric distances of some of the stars were estimated using $M_V$ values from Vacca et al. (1996) and Landolt-B$\ddot{o}$rnstein (1982). The large distace derive for LSIII+5537 casts doubts on their relation to the nebula. Note that, based on photometric data, Panov & Seggewiss (1990) concluded that HD211853 is a system with two pairs of stars, both hosting a WR component and an O-type star.
![DSS-R image of the brightest section of Sh2-132. The grey scale is arbitrary. The large cross marks the position of the WR star, while smaller crosses indicate the location of the OB stars.[]{data-label="opt1-sh2-132"}](vasquez-1.eps){width="84mm"}
Chu & Treffers (1981), based on Fabry Perot observations of the H$\alpha$ line, found that the ionized gas has LSR velocities in the range –53 to –45 , in agreement with previous results from Georgelin & Georgelin (1976, –50.9 ), Reynolds (1988, –48$\pm$1 ), and Fich et al. (1990, –47.1 ). Quireza et al. (2006) detected ionized gas at –50.46 and –49.48 from He and H radio recombination lines, respectively. Harten et al. (1978) estimated an emission measure $EM$ = 3000 pc cm$^{-6}$ for the entire nebula.
------------ -------------- -------------------- --------------- -------------- --------------- -------------- ---------------------- --------------------------
Star $(l,b)$ ST $V$ $(B-V)$ $(B-V)_0$ $A_v$ $M $d$
_V$
mag mag mag mag mag kpc
HD211853 102468,–039 WN6/WC+O6I$^{(1)}$ 9.08$^{(2)}$ 0.27$^{(2)}$ $^{(2)}$ 2.28$^{(2)}$ –6.4$^{(2)}$ 2.75$^{(2)}$,5.0$^{(4)}$
(WR153ab)
LSIII+5537 102468,–0402 O7V$^{(3)}$ 11.01 0.32$^{(3)}$ –0.29$^{(6)}$ 1.86 –5.2$^{(5)}$ 7.5
BD+552722 102486,–0402 O8.5V$^{(3)}$ 9.91$^{(3)}$ 0.45$^{(3)}$ –0.28$^{(6)}$ 2.26 –4.7${(5)}$ 3.0
LSIII+5540 102564,–0426 O9V$^{(3)}$ 12.08$^{(3)}$ 0.64$^{(3)}$ –0.28$^{(6)}$ 2.85 –4.5${(5)}$ 5.6
LSIII+5542 102564,–0438 B$^{(7)}$ 12.38 0.52 –0.26$^{(6)}$ 2.42 –2.45, –5.1$^{(*5)}$ 5-0-10.3
LSIII+5539 102576,–0384 B$^{(7)}$ 11.74 0.45 –0.26$^{(6)}$ 2.20 –2.45, –5.1$^{(*5)}$ 2.5-8.5
LSIII+5545 102594,–0456 O8$^{(3)}$ 10.45$^{(3)}$ 0.70$^{(3)}$ -0.27$^{(5)}$ 3.01 –4.9, –5.8$^{(**5)}$ 2.9-4.9
LSIII+5549 103024,–0486 B$^{(7)}$ 11.50 0.57 –0.26$^{(6)}$ 2.5 7 –2.45, –5.1$^{(*5)}$ 1.9-6.4
------------ -------------- -------------------- --------------- -------------- --------------- -------------- ---------------------- --------------------------
\
New H$\alpha$ and \[O III\] images taken by Miller & Chu (1993) revealed an arc of 4 in size to the south of the WR star (at $[l,b]$ = \[10246, –043\]). They concluded that the filament probably indicates the position of an ionization front and considered it to be a ring nebula related to the star. Based on optical data, Esteban & Rosado (1995) found velocities in the range –41 to –59 for the arc and derived an electron density of 290$\pm$100 cm$^{-3}$ from line ratios. The \[OIII\]/H$\alpha$ ratio is consistent with values derived for ring nebulae, suggesting that WR153ab is the main source of ionization (Esteban & Rosado 1995). These authors classify the arc as $R_s$-type based on the morphology of this feature and on the absence of a clear evidence of expanding motions in the line profiles. This last point suggests that the massive progenitor of the WR star contributed in the shaping of the nebula (see Chu et al. 1981).
Distance estimates for the WR star based on different methods range from 2.75 to 5.0 kpc. Massey (1981) derived a spectrophotometric distance of 5.0 kpc, while van der Hucht (2001) estimated 2.75 kpc based on its association to CepOB1, with a distance uncertainty of 40%. Foster & Routledge (2003), using a method based on column densities, estimated 3.2$\pm$0.5 kpc. A distance of 3.68 kpc was adopted by Churchwell & Wamsley (1973), based on the WN6+BOI spectral type of the WR star, known at that time. Georgelin & Georgelin (1976) derived a distance of 3.6 kpc, while Fich, Blitz & Stark (1989) obtained 4.2$\pm$1.5 kpc.
Adopting a mean LSR velocity of –48 for the ionized gas, and taking into account the presence of non-circular motions in this section of the Galaxy (see Fig. 2a by Brand & Blitz 1993), a kinematical distance $d_k$= 3.5$\pm$1.0 kpc can be predicted, in close agreement with distances derived optically for the WR star. We adopt this distance in the present study.
With the aim of investigating in some detail the nature of Sh2-132, we analyzed the distribution of the different gas components in its environs. To accomplish this goal we performed a multi-frequency study using radio continuum data at several frequencies, 21cm line data, $^{12}$CO(1-0) line data, and IR observations at different wavelengths. These data revealed that photodissociation regions are present at the interface between the ionized and molecular gas.
Database
========
Neutral hydrogen 21cm line data from the Canadian Galactic Plane Survey (CGPS, Taylor et al. 2003) was used to analyze the gas distribution, while $^{12}$CO(1-0) data at 115 GHz from the Five College Radio Astronomy Observatory, allowed us to study the molecular distribution. Intensities are shown as main beam brightness temperatures $T_{mb}$. The data were smoothed to $1\farcm 4\times1\farcm 4$ to facilitate the identification of the features. The main observational parameters of these data bases are detailed in Table \[cubos\].
Radio continuum observations at 408 and 1420 MHz were also extracted from the CGPS. The image at 408 MHz was obtained with a synthesized beam and an rms noise of 34$\times$28 and $\sim$1.1 K, respectively; the corresponding values for 1420 MHz are 10$\times$081, and 0.063 K.
Mid-infrared (MIR) images were extracted from the MSX Galactic Plane Survey (Price et al. 2001) at 8.3$\mu$m (A-band), 12.1$\mu$m (C-band), 14.7$\mu$m (D-band), and 21.3$\mu$m (E-band), with an angular resolution of 184. MIR and far-infrared (FIR) images at 12, 25, 60, and 100 $\mu$m from the IRAS satelite (HIRES data) are also included. These images have angular resolutions in the range 05 to 20.
$^{12}$CO(1-0)
--------------------- ---------------- ----------------
Original FWHM 118$\times$098 $\sim$1
Final FWHM 14$\times$14
Vel. range (LSR) () –165,+57 –165, +57
Vel. resolution () 1.32 0.824
RMS noise $T_b$ (K) 1.7 0.15
: Observational parameters of the and $^{12}$CO data[]{data-label="cubos"}
Results
=======
Radio continuum emission {#continuo}
------------------------
Figures \[figlarge\] and \[figwherett\] show the CGPS radio continuum emission at 1420 MHz. The grayscale in Fig. 2 has been chosen to emphasize the large-scale region of faint emissivity to the north of WR152 (Cappa et al. 2010, hence the “burnt-out” appearance of the region around WR 153ab). The emission distribution at 1420 MHz resembles an upside-down “U”. This semicircular structure consists of two areas of strong emission coincident with Shells A and B detected by Harten et al. (1978), centered at $(l,b)$ $\simeq$ (10234424, –004698) and at $(l,b)$ $\simeq$ (10226358,–0043102), respectively. WR153ab, indicated by the larger cross in Fig. 1, appears projected onto the central region of Shell B. Thus, the bright radio emission at $l\geq$ 1025 is in excellent agreement with the optical emission (see Fig. \[opt1-sh2-132\]) and with the image at 610 MHz by Harten et al. (1978). Both Shells A and B can be identified at 1420 MHz. Note in particular the small arc of emission at $(l,b)$ = (10277, –073) (best identified in Fig. 3), which is the radio counterpart of the 4-arc detected at optical wavelengths by Miller & Chu (1993).
![Full resolution CGPS image at 1420 MHz emphasizing faint emission to the north of WR 152 (indicated by [**a cross symbol**]{} in the lower right part of the figure). The arrow points to a continuum feature corresponding to the large HI structure.[]{data-label="figlarge"}](vasquez-2.eps){width="84mm"}
![Full resolution 1420 MHz CGPS image showing the eastern and western regions of Sh2-132 used for TT-plot analyses. The radio sources in direction to Sh2-132 are indicated by 10, 13 and 14. The white contour line marks the 9 K contour of the convolved 1420 MHz image used for the analysis. See text for details.[]{data-label="figwherett"}](vasquez-3.eps){width="84mm"}
The faint structure centered at $(l,b)$ = (1022, –083) is related to WR152 and has been analyzed in a previous paper (Cappa et al. 2010). Of particular interest is the curved filament indicated by an arrow in Fig. 2. This filament can be followed from $b \approx$ –075 to nearly $b \approx$ –005. The figure also shows a faint barely visible northeastern section. This structure is likely the radio continuum counterpart of the $\sim$60 diameter structure discussed in Cappa et al. (2010).
Three bright point sources, best seen on Figure \[figwherett\], are projected onto the nebula near Shell A. Their coordinates, flux densities and identifications are included in Table 3. The spectral index estimates indicate that sources 10 and 14 are non-thermal in nature and probably extragalactic, while source 13 is thermal. The value of the spectral index indicates that the radio source is optically thick at least at 408 MHz.
The CGPS image at 408 MHz does not show as many details as the one at 1420 MHz, as it is of much lower resolution. However it can be used to study the spectral index distribution of the radio continuum emission around WR 153ab. As a first step, we removed the point sources which are present over the region of interest and convolved the 1420 and 408 MHz images to the same spatial resolution of 34. We then constructed a TT-plot, in which the brightness temperature $T_b$ at one frequency is plotted point-by-point against the brightness temperature at the other frequency. The brightness temperature spectral index $\beta$, where $T_b \propto \nu^{\beta}$, is directly related to the slope of the straight line fitted by regression, $\beta = -\,\log(\rm slope)/\log(1420/408)$. The usual flux density spectral index $\alpha$ ($S_{\nu} \propto \nu^{\alpha}$) is simply $\alpha = \beta + 2$. The TT-plot method is immune from errors in the zero-level of the individual images (which the simpler technique of dividing one map by the other, even after background subtraction, is not).
Three different TT-plots were constructed, one for the entire Sh2-132 region and one for the eastern and western halves, which correspond to the two shells identified by Harten et al. (1978). Figure \[figwherett\], which is the full resolution image at 1420 MHz, shows these different regions, the dashed lines delimiting the eastern and western parts. The three bright small diameter sources coincide with sources 10, 13 and 14 in Table 2 of Harten et al. (1978) and had been subtracted from the original image before convolution to a circular beam of 34.
![TT-plots corresponding to (a) the whole structure, (b) the eastern half and (c) the western half. For all plots, a background of 9 K and 80 K has been subtracted from the convolved images at 1420 and 408 MHz, respectively.[]{data-label="figtt"}](vasquez-4.eps){width="70mm"}
The results of the TT-plot analysis are presented in Figure\[figtt\]. In each case, the analysis was limited to regions brighter than 9 K at 1420 MHz and 80 K at 408 MHz. Although a significant amount of scatter is present on the diagram, especially for the eastern sector, there is no evidence for the presence of significant variations of the spectral index and a further study of spatial variations is not warranted. The best-fit spectral index is $\alpha$ = –0.03$\pm$0.03, corresponding to thermal emission.
In order to obtain some radio continuum parameters, we first derive the total number $N_u$ of UV ionizing photons per second required to produce the observed free-free emission. This can be written as (e.g. Chaisson 1976): $$N_u = 0.76 \times 10^{47}\,\,T_4^{-0.45}\,\nu_{\rm GHz}
^{0.1}\,S_{\nu}\,\,d_{\rm kpc}^{2},$$ where $S_{\nu}$ is the flux density in Jy, $T_4$ is the electron temperature in units of $10^4$ K and $\nu_{\rm GHz}$ the frequency in GHz. We use the CGPS image at 1420 MHz. The derived flux density is significantly dependent on how the source dimension is defined. The surrounding background is somewhat variable, nevertheless it is reasonable to assume that the source emission corresponds to brightness temperatures above 6.5 to 7 K. Using an average background of 7 K, we obtain a flux density at 1420 MHz of $23 \pm 5$ Jy, the error being determined from a second estimate using a background of 6.5 K. From the above equation for $N_u$ and taking $T_4 = 0.8$ (Quireza et al. 2006), we obtain $N_u = 2.2 \times 10^{49} \,\rm s^{-1}$. Considering two O6I-type stars and two WN6 as components of the multiple system, and the spectral type from BD+552722, LSIII+5545, and LS+5540 (see Table 1, related to the region, we obtain the UV ionizing photons per second emited by the stars, $Q_0$=(16.6$\pm$3.0)$\times$10$^{49}$ s$^{-1}$, for solar metalicity from Smith et al. (2002).
The comparison between the number of UV photons emitted by the stars $Q_0$ and the number of UV photons used to ionize the gas $N_u$ indicates that the massive stars in the nebula are responsible for the ionization. A large fraction of the UV photons are also used to heat the associated interstellar dust. This result differs from the one by Harten et al. (1978) since their estimate of $Q_0$ was too low.
----- ----------- ------------ ----------- ---------- -------------- ------------- ----------- --------------- ------------ ---------------------
\# $l$ S$_{1420}$ S$_{408}$ $\alpha$ S$_{178}$ S$_{365}$ S$_{610}$ S$_{1420}$ S$_{4850}$ Identification
$b$ mJy mJy mJy mJy mJy mJy
(1) (2) (2) (2) (3) (4) (5) (6) (5)
10 102595090 195$\pm$10 550 –0.8 819$\pm$ 87 330 $<$40 7C 2218+558$^{(7)}$
–0402335
13 10333024 365$\pm$15 280 +0.2 160 222.6$\pm$7.5 $<$40 NVSS J222034+561438
–0412814
14 10344004 450$\pm$20 1200 –0.8 3100$\pm$500 1738$\pm$62 810 60 TXS 2219+559
–0472704
----- ----------- ------------ ----------- ---------- -------------- ------------- ----------- --------------- ------------ ---------------------
References: (1) Number from the list of Harten et al. (1978); (2) this paper, derived using CGPS data; (3) Gower et al. (1967); (4) Douglas et al. (1996); (5) Harten et al. (1978); (6) Condon et al. (1998); (7) Waldram et al. (1996).
We can now use the models of Mezger & Henderson (1967) to infer the properties of the extended ionized gas, namely its mass $M_{\rm HII}$ and the rms electron number density $n_e$. Letting $\theta_{Ga}$ be the observed half-power width of the extended emission (in minutes of arc, after correction for the beam), we obtain: $$n_e/{\rm cm^{-3}} = 6.35 \times 10^2 \,u_1\,a^{-1/2}\,T_4^{0.175}\,\nu_{\rm GHz}
^{0.05}\,S_{\nu}^{0.5}\,\theta_{\rm Ga}^{-1.5}\,d_{\rm kpc}
^{-0.5}$$ $$M_{\rm HII}/M_{\odot}=0.386\,u_2\,a^{-1/2}\,T_4^{0.175}\,\nu_{\rm GHz}
^{0.05}\,S_{\nu}^{0.5}\,\theta_{\rm Ga}^{1.5}\,d_{\rm kpc}
^{2.5}$$ where $a(\nu,T_e)$ is the Gaunt factor (about unity for our purposes), $d_{\rm kpc}$ is the distance in kpc and the parameters $u_1$ and $u_2$, both of order unity, depend on the assumed model for the source. These parameters are given by Mezger & Henderson (1967) for three different idealized models. From Figure\[figwherett\], it is apparent that the nebulosity can be roughly broken up into two different regions separated by a nearly vertical line running at a longitude of $\sim$10287. These two regions correspond to shells A and B from Harten et al. (1978). The flux densities of the two sub-regions are $11 \pm 2$ Jy and $12 \pm 3$ Jy, respectively. For simplicity, we model each half-region as a sphere of constant density and of angular size 12. From Mezger & Henderson (1967), we find $u_1 = 0.78$, $u_2 =1.29$. Assuming $T_4 =0.8$, from which we determine $a^{-1/2} = 1.04$, we finally have an rms $n_e = 20\,\rm cm^{-3}$ and $M_{\rm HII} =
1500 M_{\odot}$ for each region, where, given the error in estimating fluxes, we have simply taken $S_{\nu} = 10 \,\rm Jy$ for each half. The derived rms electron density is compatible with estimates by Harten et al. (1978) for Shells A and B.
![[*Upper pannel*]{} Sh2-132 at 8.3 $\mu$m (MSX band A). The grayscale goes from 8.0$\times$10$^{-7}$ to 3.5$\times$10$^{-6}$ W m$^{-2}$ ster$^{-1}$. Extended filaments and point-like sources are indicated (see text). [*Bottom pannel:*]{} Overlay of the DSS-R image and the image at 60$\mu$m. IR contours are 50 to 190 MJy ster$^{-1}$ in steps of 20 MJy ster$^{-1}$, and from 300 to 900 MJy ster$^{-1}$ in steps of 200 MJy ster$^{-1}$.[]{data-label="fig:imagenesIR"}](vasquez-5.eps){width="84mm"}
IRAS and MSX emission
---------------------
The IRAS (60$\mu$m) and MSX band A (8.3$\mu$m) emissions are shown in Fig.\[fig:imagenesIR\]. Although we analyzed the four IRAS bands, we include the 60 $\mu$m-image only since the extended emission is clearly seen at this wavelength. Emission at 8.3$\mu$m (upper panel) in diffuse nebulae has a contribution from both PAH and dust continuum. As pointed out by Povich et al. (2007), in very energetic regions like M17, dust continuum can be very strong. However, for the case of Sh2-132, the assumption that MSX band A is dominated by PAH emission seems to be reasonable. Consequently, emisison in this band indicate the presence of photodissociation regions, while the emission in the far IR originates in large dust grains.
The image at 8.3 $\mu$m (upper panel) reveals the existence of a ring-like feature centered at $(l,b)$ = (10253, –0040) of about $\sim$15 in size. This feature can be also identified at 60 $\mu$m (lower panel). The comparison of the optical emission with the emission at 8.3 $\mu$m shows that the mid IR emission encircles most of the nebula, following the outer rim of the ionized gas. This fact is more noticeable for Shell B and is particularly striking near $(l,b)$ = (10247,–0043), where the 4 ring nebula is present. The bottom panel shows that the emission distribution at 60 $\mu$m resembles that in the optical.
In order to estimate some dust parameters, we have divided this ring-like structure in three main filaments, named f1, f2, and f3, which are indicated in the upper panel of Fig. 5. Note that f3 borders the region with high extinction seen in the optical image near $(l,b)$ = (1030,–0047).
A number of interesting point-like sources located onto the filaments can also be identified in the MSX and IRAS images. They are indicated in the figure as ps1, ps2, and ps3. Infrared source ps1 at $(l,b)$ = (1023945,–003624) is the MSX source G102.6627-00.6067, coincident with IRAS22160+5551; ps2 at $(l,b)$ = (1024818,–00436) corresponds to G102.8051-00.7184 and IRAS22172+5549; while ps3 at $(l,b)$ = (1030344,–004127) is G103.0624-00.6911 and IRAS22187+5559.
After background substraction, the IR flux densities of the filaments in the four IRAS bands and in the MSX band A were calculated. These values are listed in Table \[table:flujos\]. The quoted errors reflect the uncertainties in the determination of the background. Emission in bands C, D, and E is very weak in comparison with the strong background emission leading to large uncertainties in flux density estimates. Values for these bands are not included. The three point sources, whose flux densities in the MSX and IRAS bands are included in the table, can be classified as YSO candidates following criteria by Junkes et al. (1992) and Lumsden et al. (2002) for IRAS and MSX point sources, respectively. In the case of the IRAS criteria, Junkes et al. (1992) consider (i) $S_{100}>$20 Jy, (ii) 1.2$<S_{100}/S_{60}<$6, (iii) $S_{60}/S_{25}\geq$1, and $Q_{60}$+$Q_{100}\geq$4, where $S_i$ is the flux density in the bands centered at 25, 60 and 100 $\mu$m, and $Q_i$ is the quality of the fluxes at these bands. Lumsden et al.’s (2002) criteria are (i) $S_{21}/S_8>2$ and (ii) $S_{14}/S_{12}>1$, where $S_i$ is the flux density centered at 8.3, 12.1, 14.7, and 21.3 $\mu$m.
It has been demonstrated (Scoville & Kwan 1976) that the infrared spectrum of radiatively excited dust grains can be fitted by a modified Planck function $B(T_d)\nu^m$, were $\nu^{m}$ is the grain emissivity and $m$ increases from 1 to 2 depending on the wavelength and composition of dust grains (Schwartz 1982). Adopting $m$ = 1.5, according with the molecular environment (Whittet, 1992), and using the measured flux densities at 60 and 100 $\mu$m, we derived the dust color temperature T$_d$ summarized in Table \[table:flujos\]. IR luminosities for the point sources are also included. Infrared filaments are excited by the intense stellar radiation field of WR153ab and the other massive stars in the region. The slightly higher dust color temperature derived for f1 can be explained bearing in mind that this filament is closer to the exciting stars than f2 and f3. In Fig. \[fig:sed\], the MIR and FIR spectral energy distribution corresponding to the filaments are shown. Note that the measured fluxes at shorter wavelengths exceed the values of our model, possibly as the result of the emission of nebular lines which may contribute to the infrared flux densities.
![IR spectral energy distribution for the IR filaments. The lines in the diagram corresponding to $f1$, $f2$, and $f3$ show the emission of a blackbody with a [**$T$ = 32, 29 and 32 K, respectively,**]{} and a [**$\nu^{1.5}$**]{}-emissivity law. The boxes denote flux densities in the four IRAS bands, and the circle, that corresponding to the MSX band A.[]{data-label="fig:sed"}](vasquez-6.eps){width="84mm"}
![Overlay of the averaged $^{12}$CO emission distribution for the molecular clouds between –46.0 to –38.6 (upper panel) and –53.4 to –46.0 (lower panel) and the MSX band A emission (grayscale). Contours correspond to 3, 7, and 15$\sigma$. The grayscale goes from 0.1$\times$10$^{-7}$ to 1$\times$10$^{-7}$ Watts m$^{-2}$ ster$^{-1}$. The crosses indicate the positions of the massive stars.[]{data-label="co-msx"}](vasquez-7.eps){width="84mm"}
![[*Upper panel:*]{} Averaged $^{12}$CO velocity distribution for $\Delta b$ = (–55$\arcmin$30$\arcsec$,–39$\arcmin$54$\arcsec$). The grayscale goes from 0.15 to 3 K. Contours are 0.45, 0.75 and 1.05 K. [*Bottom panel:*]{} Averaged $^{12}$CO velocity distribution between $\Delta b$ = (–39$\arcmin$54$\arcsec$,–29$\arcmin$06$\arcsec$). Contours and grayscale are the same as in the upper panel.[]{data-label="co-v-l"}](vasquez-8.eps){width="84mm"}
-------- -------------- ------------- ------------------------ ------------- ------------- ------------ -------------- -------------- ---------- -------
Source $T_d$ (K) $L_{IRAS} (L_{\odot})$
8.3 $\mu$m 12.1 $\mu$m 14.7 $\mu$m 21.3 $\mu$m 12 $\mu$m 25 $\mu$m 60 $\mu$m 100 $\mu$m
f1 46$\pm$1 — — — 72$\pm$4 192$\pm$24 1370$\pm$130 2765$\pm$145 32$\pm$7 43800
f2 16.0$\pm$3.5 — — — 35$\pm$6 77$\pm$11 484$\pm$51 1201$\pm$160 29$\pm$6 16900
f3 1.8$\pm$0.3 — — — 4.9$\pm$0.5 32$\pm$11 133$\pm$40 343$\pm$37 32$\pm$6 5000
ps1 1.45 1.85 0.97 2.14 3.03 4.28 74.1 274 28$\pm$5 2700
ps2 2.22 3.93 3.78 7.02 7.8 16.1 106 318 30$\pm$5 3900
ps3 1.0 1.6 1.40 4.26 7.02 17.2 243 453 36$\pm$8 6800
-------- -------------- ------------- ------------------------ ------------- ------------- ------------ -------------- -------------- ---------- -------
Molecular gas distribution {#CO}
--------------------------
The analysis of the $^{12}$CO(1-0) emission distribution leads to the detection of two structures with velocities in the ranges from –53.4 to 46.0 , and from –46.0 to –38.6 . Fig. \[co-msx\] shows the averaged CO emission (contours) for the two molecular structures superposed onto the emission at 8.3 $\mu$m.
The CO feature present between –46.0 and –38.6 (upper panel) exhibits a ring-like appearance between $l\sim$ 10235 and 102 50, while for $l>$ 10245, the emission is concentrated along $b\simeq$ –035. The whole structure will be referred as cloud A. Two smaller molecular clouds are detected at $(l,b)\sim$ (1032,–030) (indicated in the figure as cloud C) and at $(l,b)\sim$ (1034,–042) (cloud B). The emission at 8.3 $\mu$m correlates very well with the molecular emission. In particular, the brighter emission regions in the mid infrared coincide with clumps in cloud A \[for example at $(l,b)\sim$ (10243,–035), (10244,–043), (10247,–044), and (1035,–042)\]. Cloud B coincides with the strong radio source NVSS J222034+561438, which is an YSO candidate. The comparison of the ionized and molecular gas distributions shows that the molecular emission encircles the brightest optical emission regions analyzed by Sánchez-Monge et al. (2008).
The lower panel displays the molecular emission between –53.4 to –46.0 . The CO emission is clumpy, being concentrated in clouds D, E and F. Cloud C is also detected in this velocity range. Cloud E overlaps a section of cloud A and follows the ring-like feature identified at 8.3 $\mu$m. Cloud F coincides with part of the IR filament at $(l,b)\sim$ (10244,–042). The section of Cloud D running from $(l,b)$ = (10256,–0051) to $(l,b)$ = (10302,–0044) coincides with f3. Cloud D also displays a clump at $(l,b)\sim$ (10305,–044), which coincides with a region of high optical extincion and is probably connected with cloud B.
Table \[tabla\_co\] gives the physical parameters of each molecular cloud. The masses of the different clouds were estimated by integrating the CO line intensity as $W(CO) = \int T_{mb}dv$ = $T_{mb}\Delta v$, where $T_{mb}$ is the average main beam brightness temperature of the molecular cloud over the velocity interval in which the cloud is observed. To estimate the $H_2$ column density $N(H_2)$, the relation $X = N(H_2)/W(CO)$ = (2.3$\pm$1.2)$\times$10$^{20}$ mol cm$^{-2}$ K$^{-1}$ km$^{-1}$ s (Grenier & Lebrun 1990) was used. The molecular mass was derived from $M$() = 4.2$\times$10$^{20}N(H_2)d^2 \Omega$, where $d$ is the distance in pc and $\Omega$ is the solid angle in steradians. Errors in $T_{mb}$ and in $N(H_2)$ arise in background uncertainties, while errors in masses come from background and distance uncertainties. The adopted distance is 3.5$\pm$1.0 kpc.
To determine the volume density, we assumed a spherical geometry for all the clouds, except for Cloud A, for which an elongated geometry was adopted.
------------------------------------- --------------- --------------- --------------- --------------- --------------- ---------------
A B C D E F
Average $T_{mb}$ (K) 1.30$\pm$0.30 0.90$\pm$0.15 0.60$\pm$0.15 1.20$\pm$0.30 1.60$\pm$0.15 0.70$\pm$0.15
Velocity range () –38.6,–46.0 –38.6,–46.0 –38.6,–53.4 –46.0,–53.4 –46.0,–53.4 –46.0,–53.4
$N(H_2)$ column density (10$^{20}$) 6.50$\pm$1.60 3.30$\pm$0.60 4.20$\pm$1.05 5.90$\pm$2.40 9.0$\pm$0.90 1.95$\pm$0.20
$H_2$ mass (10$^3$ ) 4.25$\pm$1.90 0.80$\pm$0.35 0.04$\pm$0.02 1.30$\pm$0.60 1.55$\pm$0.70 0.95$\pm$0.45
Volume density () 37$\pm$22 2150$\pm$1290 110$\pm$65 95$\pm$55 115$\pm$70 320$\pm$190
$N(HI)$ column density (10$^{21}$) 3.7$\pm$0.9 3.5$\pm$0.5 3.7$\pm$0.9 3.3$\pm$0.8 3.4$\pm$0.3 3.3$\pm$0.6
Visual absorption $A_v$ (mag) 2.5$\pm$0.8 2.2$\pm$0.5 2.8$\pm$0.7 2.4$\pm$0.7 2.8$\pm$0.6 2.7$\pm$0.5
------------------------------------- --------------- --------------- --------------- --------------- --------------- ---------------
Figure \[co-v-l\] shows two velocity-position maps of the CO emission. We have plotted Galactic longitude in the y-axis since the molecular morphology is mostly elongated in this direction. The map in the upper panel is the result of integrating within $\Delta b$ = , while the map in the bottom panel was integrated within $\Delta b$ = (–03954,–02906). The upper panel shows two clouds at $l \sim$ 10245 and $l \sim$ 10300, having velocities in the range $\Delta v \sim$ (–40,–45) and (–45,–50) , respectively (two small cloudlets are also detected at $v\simeq$ –30 at $l \sim$102 50 and $l \sim$103 03). Clouds F and D are clearly identified. The bottom panel shows cloud A extending from –10235 to 10313. The velocity range goes from –40 at $l \sim$ 10310 to –50 at $l \sim$ 10235. The image shows some small cloudlets at $l \sim$ 10323 at $v\simeq$ –38 , $l<$ 10240 with velocities in the range $\sim$ –8 to –5 , and at $l >$ 10325 at $v\simeq$ 0 . Molecular gas with velocities higher than –10 is unconnected to Sh2-132.
The agreement in velocity between the ionized and molecular gas and the fact that the molecular emission borders the brightest ionized regions indicate that molecular gas with velocities between –38 and –54 is associated with Sh2-132 and the massive stars in the region.
We adopt –45 as the systemic velocity of the molecular gas. Taking into account the non-circular motions in this section of the Galaxy shown by Brand & Blitz (1993), a kinematical distance $d_k$ = 3.0$\pm$1.0 kpc can be estimated, compatible with the distance adopted in Sect.\[intro\].
The distribution of the emission {#HI}
----------------------------------
The upper panel of Fig.\[hi-corte\] displays the emission distribution in the velocity range where the CO emission is present, i.e. between –31.1 and –58.4 . The white circle has the dimensions of the region. Neither a region lacking gas coincident in position with Sh2-132 as a whole nor an shell surrounding it are evident in the image, although the small cavity near $(l,b) \simeq$ (10258,–040) is probably linked to shell A.
![[*Upper panel:*]{} Average emission distribution between –31.1 and –58.4 . The grayscale goes from 50 to 93 K, while contours correspond to 60, 70, 80, and 90 K. The white circle has the dimensions of the region Sh2-132. [*Bottom panel:* ]{} Diagram showing $T_b$ vs. $b$ from $b$= –90$\arcmin$ to 0$\arcmin$ for $l$ = 10250 from the image above. The dashed line shows the best fit to the data.[]{data-label="hi-corte"}](vasquez-9.eps){width="84mm"}
However, a gradient in the emission distribution is evident for $b <$ –010. The lower panel displays the emission corresponding to $l$ = 10250 from $b$ = 000 to –130from the image in the upper panel. $T_b$ and $N(HI)$ are shown on left and right y-axes, respectively. The image shows that the brightness temperature is $T_b\simeq$ 85 K at $b \simeq$ –015, while 65% this value is detected at $b \simeq$ –130. The [**fitted curve**]{}, showed as a dashed line, confirms the presence of a $T_b$-gradient.
The fitting curve was used to estimate the contribution of the gas to the visual absorption $A_v$. We calculated the visual absorption $A_v$ due to the molecular and neutral hydrogen related to the region by using the relation $(N(HI) + 2N(H_2))/E(B-V) = 5.8\times10^{21}$ atoms cm$^{-2}$ mag$^{-1}$ (Bohlin et al. 1978), where $N(HI)$ and $N(H_2)$ are the and H$_2$ column densities, respectively, while $E(B-V)$ is the color excess. We adopted $E(B-V) = A_v/3.05$ and read off $N(HI)(b)$ from the diagram. The estimated column densities, as well as the derived $A_v$-values for each molecular cloud are included in Table \[tabla\_co\].
Photodissociated regions
========================
The presence of PAH emission at the interfase between the ionized and molecular materia, e.i. at the border of many regions, suggests that molecular gas is being photodissociated by the UV photons emitted by the massive stars. In the following subsections we analyze the presence of PDRs in two selected areas within the region, one around WR153ab and BD+552722, linked to shell B, and the other around LS+5539, linked to shell A.
Region near WR153ab and BD+552722
---------------------------------
The upper panel of Figure 10 displays an overlay of the DSS-R image and the CO emission distribution for a small region of Sh2-132 near WR153ab and BD+552722. Thin contours correspond to Cloud A with velocities in the range –38.6 to –46.0 , while thick ones delineate Clouds E and F having velocities between –46.0 and –53.4 . The crosses mark the position of the stars.
The morphological agreement between the optical emission and the inner borders of the molecular clouds is excellent, particularly for the 4 optical filament at $(l,b) \sim$ (1024630, –043), where the contour corresponding to 2.25 K closely follows the optical emission. Note also that difusse emission is also encircled by molecular emission. The existence of emission at 8.3 $\mu$m bordering the ionized region (see Fig. 7) strongly indicates the presence of a PDR.
To confirm the existence of a PDR in this region we compared the distribution of the ionized and molecular gas, and that of PAHs. After convolving the DSSR and MSX band A images to the CO beam (46), we averaged the optical, molecular and band A emissions in concentric rings from the position of the WR star within an angle of 300 from $(l,b) \sim$ (10252,–038) clockwise up to $(l,b) \sim$ (10249,–033). The result of this average is displayed in the bottom panel of the figure, where intensities in the different bands are normalized to the maximum value of each band. The values in the abscisae indicates distance from the WR star. The diagram reveals a clear stratification, with the optical emission peaking near the stellar position, and the emission at 8.3 $\mu$m peaking 3 far from the star. The molecular emission increases up to 6 from the star, where the maximum is present. The observed distribution is typical of molecular clouds exposed to UV radiation, which photodissociates them (van der Werf et al. 1996).
![[*Upper panel:*]{} Overlay of the DSS-R image and the CO emission in the region near WR153 and LS+552722, indicated as crosses. Thin lines delineate Cloud A and thick lines, clouds E and F. [*Bottom panel:*]{} Normalized intensities of the optical, MSX band A, and CO emissions in concentric rings as a function of the distance to the WR star. []{data-label="BD"}](vasquez-10.eps){width="84mm"}
Region near LS+5539
--------------------
The region around LS+5539 is particularly interesting. The upper panel of Fig. 11 shows that the star is close to a bright optical filament of $\sim$2 in size. The fact that the optical filament partially follows the border of cloud A indicates that the UV photons of the star are photodissociating and ionizing the dense gas. The high extinction region at $(l,b) \sim$ (10257,–040) coincides with a section of cloud D.
With the aim of investigating the presence of a PDR near the optical filament, we plotted the normalized intensity of optical, MSX band A, and CO emissions as a function of the distance to LS+5539. The plotted emission corresponds to the region indicated by the black arrow in the upper panel. As for the region around WR153ab and BD+552722, the optical emission peaks closer to the excitation source than the emission in MSX band A and the molecular emission, revealing the presence of a PDR.
Taking into account that LS+5539 is a B0V star with $\log Q(H^0)$ = 46.23 (Vacca et al. 1996), and assuming that it is immersed in an ionized interstellar medium with a mean electron density $n_e$ = 20 cm$^{-3}$, we can estimate the Strömgren radius $r_s$ of the region created by the star from $\frac{3Q(H^0)}{4\pi \alpha_B n_e ^2} = r_s^3$ (Mezger & Henderson 1967), where $\alpha_B$ = 2$\times$10$^{-13}$ s$^{-1}$. The result is $r_s$ = 5.3 pc. Bearing in mind that the angular distance from the star to the optical filament is 2.2, or 2.25$\pm$0.7 pc at 3.5$\pm$1.0 kpc, we can conclude that the B-type star is responsible for ionizing the gas in the region and contributing to the photodissociation the the molecular gas.
![[*Upper panel:*]{} Overlay of the DSS-R image and the CO emission in the region near LS+5539, indicated by a cross. Thin lines delineate Clouds A and D. [*Bottom panel:*]{} Normalized intensities of the optical, MSX band A, and CO emissions in concentric rings as a function of the distance to the B star.[]{data-label="LS"}](vasquez-11.eps){width="84mm"}
Scenario and conclusions
========================
In this paper we have analyzed the distribution of the ionized, neutral atomic and molecular gas and that of the interstellar dust in Sh2-132.
This study revealed the presence of a number of molecular clouds encircling the ionized nebula, with a mass of 9000 . The velocity of the molecular gas, which is in the range from –38.6 to –53.4 coincides with the velocities of the $H\alpha$ line, and $He$ and $H$ radio recombination lines, strongly indicating that the ionized gas is associated with the nebula and interacting with the ionized gas.
The ring-like appearance of the gas and dust distribution in the environs of the WR star suggests the action of stellar winds, that sweep up and compress the gas. The mechanical energy $E_\mathrm{w}$ released by the multiple system WR153ab into the ISM can be roughly estimated taking into account the two WR components of the binary system during the duration of the WR phase of a massive star, the previous main sequence phase of the current WR stars, and the two O-star components. We adopt a mass loss rate $\log$ ($\dot{M}$) = –6.64 and a terminal velocity $v_{\infty}$ = 1950 , corresponding to a O8.5V star (Smith et al. 2002) as parameters of the stellar wind of the O-type components. For the WR components we adopt a conservative mass loss rate $\dot{M} = $5$\times$10$^{-6}$M$_\odot$yr$^{-1}$ and a terminal velocity $v_{\infty}$ = 2000 . Adopting lifetimes of 3$\times$10$^6$ yr and 0.5$\times$10$^6$ yr for the O and WR phases of the stars, respectively (Conti & Vacca 1990, Meynet & Maeder 2005), $E_\mathrm{w}$ results to be $\sim$ 7.2$\times$10$^{50}$ erg.
Considering the molecular clouds E, F and a section of cloud A, and the dust and ionized components, with an expansion velocity of 7.4 , the kinetic energy of shell B is $\sim$2.80$\times$10$^{48}$ erg. This result indicates that the multiple system can shape shell B.
The emission at 8.3 $\mu$m appears concentrated in a ring-like structure bordering the brightest sections of the nebula, indicating that molecular material is being photodissociated at the interface between the ionized and molecular gas. This configuration resembles the IR bubbles analyzed by Watson et al. (2008). To confirm the presence of photodissociation regions we have analyzed in some detail two regions within the complex. One in Shell B around WR153ab and BD+552722, and the other in Shell A around LS+5539. The relative positions of the ionized gas, PAHs, and molecular gas in these regions show the stratification typical in PDRs. Thus, the distribution of $H_2$ plays a main role in shaping the region.
The emission distribution at 60 $\mu$m shows the same ring-like structure, along with emission inside the ring. The presence of emission in the far IR inside the region, which originates from large dust grains, is compatible with the existence of PDRs since PAHs are destroyed inside regions, while large grains are not.
The number of UV photons emitted by the massive stars in the region is seven times higher than the UV photons used to ionize the gas, indicating that the massive stars are responsible for the ionization of the region. Clearly, a large fraction of the available UV photons are used in the heating of the interstellar gas and in the photodissociation of the dense material.
Different studies of regions such as those towards RCW79 (Zavagno et al. 2006), RCW82 (Pomarés et al. 2009), and Sh2-212 (Deharveng et al 2008) show how the interaction between the massive stars with the surrounding molecular material induces star forming regions through the “collect and collapse” process (Elmegreen and Lada 1977) in the borders of these regions. For example, the paper on RCW79 (Zavagno et al. 2006) reveals the formation of YSOs inside pre-existing molecular cores located at the border of the ionized region. By using the analytical model by Whitworth et al. (1994) to determine the time at which the fragmentation occurs and the size of the structure at that time, they conclude that the collect and collapse process may not be the only one at work. Sh2-212 (Deharveng et al. 2008) and Sh2-104 (Deharveng et al. 2003) represent good examples of stellar formation in which the molecular clouds surround the regions and the PAH emission is located at the interfase between the ionized and the molecular gases. In particular, Sh2-104 is encircled by a molecular ring showing four dense molecular clouds containing many dense cores. This configuration is a strong evidence in favour of the triggering star formation process descripted by the “collect and collapse” model. In all of the preceding examples, we have an region generated by an O-type star or a cluster of massive stars.
The region around Shell B presents a similar interstellar scenario as the preceding examples, e.i an optical shell bounded by a strong arc-like feature of PAH emission which are surrounding by molecular material (Clouds A, E, and F). In such scenario, star formation can be induced by the action of the massive stars through the ionizing flux and stellar winds. Indeed, the presence of YSOs towards the molecular clouds linked to Sh2-132 indicates that star formation is very active in this region of the Galaxy (Vasquez et al. in preparation).
acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge the anonymous referee of her/his comments. This project was partially financed by the Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) of Argentina under projects PIP 112-200801-02488 and PIP 112-200801-01299, Universidad Nacional de La Plata (UNLP) under project 11/G093, Universidad de Buenos Aires under project UBACyT X482, and Agencia Nacional de Promoción Científica y Tecnológica (ANPCYT) under projects PICT 00812 and 2007-00902. The Digitized Sky Survey (DSS) was produced at the Space Telescope Science Institute under US Government grant NAGW-2166. This work was partly (S.P.) supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Fonds FQRNT of Québec. The DRAO Synthesis Telescope is operated as a national facility by the National Research Council of Canada. The CGPS is a Canadian project with international partners and is supported by grants from NSERC. Data from the CGPS is publicly available through the facilities of the Canadian Astronomy Data Centre (http://cadc.hia.nrc.ca) operated by the Herzberg Institute of Astrophysics, NRC.
[99]{} Bakes, E.L.O., Tielens, A.G.G.M. 1998, ApJ, 499, 258\
Bohlin, R.C., Savage, B.D., Drake, J.F. 1978, ApJ, 224, 132\
Brand J., Blitz L. 1993, A&A, 275, 67\
Cappa C.E., Vasquez J., Pineault, S., Cichowolski S. 20010, MNRAS, in press\
Chaisson, E.J. 1976 in [*Frontiers of Astrophysics*]{}, edited by E.H. Avrett, Cambridge: Harvard Univ. Press, 259\
Chu, Y.-H., 1981, Apj, 249, 195\
Chu, Y.-H., Treffers R.R. 1981, ApJ, 250, 615\
Chu, Y.-H., Treffers R.R., Kwitter, K.B. 1983, ApJS, 53, 937\
Churchwell E., Walmsley C.M., 1973, A&A, 23, 117\
Condon, J.J., Cotton, W.D., Greisen, E.W., Yin, Q.F., Perley, R.A., Taylor, G.B., Broderick, J.J. 1998, AJ, 115, 1693\
Conti P.S., Vacca W.D. 1990, AJ, 100, 431\
Deharveng L., Lefloch B., Zavagno A., Caplan J., Whitworth A. P., Nadeau D., Martín S., 2003, A&A, 408, 25\
Deharveng L., Zavagno A., Caplan J., 2005, A&A, 433, 565\
Deharveng L., Lefloch B., Kurtz S., Nadeau D., Pomarés M., Caplan J., Zavagno A., 2008, A&A, 482, 585\
Douglas, J.N., Bash F.N., Bozyan F.A., Torrence G.W., Wolfe C. 1996, AJ, 111, 1945\
Elmegreen B.G., Lada C.J., 1977, ApJ, 214, 725\
Esteban C. Rosado M. 1995, A&A, 304, 491\
Fich M., Blitz M., Stark A.A. 1989, ApJ, 342, 272\
Foster, T., Routledge, D. 2003, ApJ, 598, 1005\
Georgelin Y. M., Georgelin Y. P. 1976, A&A, 49, 57\
Gower J.F.R., Scott P.F., Wills D. 1967, MNRAS, 71, 49\
Grenier I.A., Lebrun F. 1990, ApJ, 360, 129\
Habing H.J. 1968, [*Bull. Astron. Inst. Astrophys.*]{}, 204, 253\
Harten R. H., Felli M., Tofani G. 1978, A&A, 70, 205\
Heckathorn J. N., Bruhweiler F. C., Gull T. R. 1982, ApJ, 252, 230\
Hogerheijde M.R., Jansen D.J., van Dishoeck E.F. 1995, AA, 294, 792\
Junkes N., Fuerst E., Reich W. 1992, A&A, 261, 289\
Kaufman M.J., Wolfire M.g. Hollembach D.J., Luhman M.L. 1999, ApJ, 527, 795\
Landalt-R$\ddot{o}$rmstein, 1982, [*Numerical Data and Functional Relationships in Science and Technology*]{}, Group VI: Astronomy, Astrophysics and Space Research, Volumen 2: Astronomy and Astrophysics, edited by Hellwege K.H.,\
Lumsden S.L., Hoare M.G., Oudmaijer R.D., Richards D. 2002, MNRAS, 336, 621\
Massey, P. 1981, PASP, 93, 549\
Meynet G., Maeder A. 2003, A&A, 404, 975\
Mezger P.G, Henderson A.P. 1967, ApJ, 147, 471\
Miller G.,J., Chu Y.-H. 1993, ApJS, 85, 137\
Nugis T., Lamers H. J. G. L. M., 2000, A&A, 360, 227\
Pomarés M., Zavagno A., Deharveng L., Cunningham M., Jones P., Kurtz S., Russeil D., Caplan J., Comerón F., 2009, A&A, 494, 987\
Panov K.P., Seggewiss W., 1990, A&A, 227, 117\
Povich M.S., Stone J.M., Churchwell E., Zweibel E.G., Wolfire M.G., Babler B.L., Indebetouw R., Meade M.R., Whitney B.A., 2007, ApJ, 660, 346\
Price S.D. Egan M.P., Carey S.J., Mizuno D.R., Kuchar T.A., 2001, AJ, 121, 2819\
Prinja R. K., Barlow M.J., Howarth I.D., 1990, ApJ, 361, 607\
Quireza C., Rood R.T., Balser D.S., Bania T. M. 2006, ApJS, 165, 338\
Quireza C., Rood R.T., Bania T. M., Balser, D.S., Maciel, W.J. 2006, ApJ, 653, 1226\
Reich W., Fuerst E., Reich P., Reif K., 1990, A&A, 85, 633\
Reynolds R. J., 1988, ApJ, 333, 341\
Sánchez-Monge A., Palau A., Estalella R., Beltrán M.T., Girart J.M., 2008, A$\&$A, 485, 497\
Sharpless S., 1959, ApJS, 4, 257\
Scoville N.Z., Kwan J. 1976, ApJ, 206, 718\
Smith L.J., Norris R.P.F., Crowther P.A., 2002, MNRAS, 337, 1309\
Smith L.F., Shara M.M., Moffat A.F.J., 1998, MNRAS, 281, 163\
Sota A., Maíz-Apellániz J., Walborn N.R., Shida R.Y. 2008, RMxAC, 33, 56\
Sternberg, A. and Dalgarno, A., ApJS, 1995, 99, 565\
Sternberg, A. and Dalgarno, A., ApJ, 1989, 338, 197\
Schwartz P.R., ApJ, 1982, 252, 589\
Taylor, A.R., Gibson S.J., Peracaula M., Martin P.G., Landecker T.L., Brunt C.M., Dewdney P.E., Dougherty S.M., Gray A.D., Higgs L.A., Kerton C.R., Knee L.B.G., Kothes R., Purton C.R., Uyaniker B., Wallace B.J., Willis A.G., & Durand D., 2003, AJ, 125, 3145\
Tielens A.G.G.M., Hollenbach, D. 1985, ApJ, 291, 722\
Vacca W.D., Garmany C.D., Shull J.M., 1996, ApJ, 460, 914\
van der Hucht K.A., 2001, New Astronomy. Rev., 45, 135\
Waldram E.M., Yates J.A., Riley J.M., Warner P.J., 1996, MNRAS, 282, 779\
Wegner W., 1994, MNRAS, 270, 229\
Watson C., Povich M.S., Churchwell, E.B., Babler B.L., Chunev G. et al. 2008, ApJ, 681, 1341\
Whittet D.C.B., from the book “ Dust in the galactic environment”, Cambridge University Press\
Whitworth A.P., Bhattal A.S., Chapman S.J., Disney M.J., Turner J.A., 1994, A&A, 290, 421\
van der Werf P.P., Stutzki J., Sternberg A., Krabbe A. 1996, AA, 313, 633\
Zavagno A., Deharveng L., Comerón F., Brand J., Massi F., Caplan J., Russeil D., 2006, A&A, 446, 171\
[^1]: E-mail: [email protected]; [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'New telescopes like the Square Kilometre Array (SKA) will push into a new sensitivity regime and expose systematics, such as direction-dependent effects, that could previously be ignored. Current methods for handling such systematics rely on alternating best estimates of instrumental calibration and models of the underlying sky, which can lead to inadequate uncertainty estimates and biased results because any correlations between parameters are ignored. These deconvolution algorithms produce a single image that is assumed to be a true representation of the sky, when in fact it is just one realisation of an infinite ensemble of images compatible with the noise in the data. In contrast, here we report a Bayesian formalism that simultaneously infers both systematics and science. Our technique, Bayesian Inference for Radio Observations (BIRO), determines all parameters directly from the raw data, bypassing image-making entirely, by sampling from the joint posterior probability distribution. This enables it to derive both correlations and accurate uncertainties, making use of the flexible software MeqTrees to model the sky and telescope simultaneously. We demonstrate BIRO with two simulated sets of Westerbork Synthesis Radio Telescope datasets. In the first, we perform joint estimates of 103 scientific (flux densities of sources) and instrumental (pointing errors, beam width and noise) parameters. In the second example, we perform source separation with BIRO. Using the Bayesian evidence, we can accurately select between a single point source, two point sources and an extended Gaussian source, allowing for ‘super-resolution’ on scales much smaller than the synthesised beam.'
author:
- |
Michelle Lochner$^{1,2,3,4}$, Iniyan Natarajan$^5$, Jonathan T. L. Zwart$^{5,6}$, Oleg Smirnov$^{7,8}$, Bruce A. Bassett$^{1,2,3}$, Nadeem Oozeer$^{1,8,9}$ and Martin Kunz$^{1,10}$\
$^1$African Institute for Mathematical Sciences, 6 Melrose Road, Muizenberg, 7945, South Africa\
$^2$Department of Mathematics and Applied Mathematics, University of Cape Town, Rondebosch, Cape Town, 7700, South Africa\
$^3$South African Astronomical Observatory, Observatory Road, Observatory, Cape Town, 7935, South Africa\
$^4$Department of Physics & Astronomy, University College London, London, WC1E 6BT, U.K.\
$^5$Astrophysics, Cosmology and Gravity Centre (ACGC), Department of Astronomy,\
University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa\
$^6$Department of Physics & Astronomy, University of the Western Cape, Private Bag X17, Bellville 7535, South Africa\
$^7$Department of Physics and Electronics, Rhodes University, PO Box 94, Grahamstown, 6140, South Africa\
$^8$SKA South Africa, 3rd Floor, The Park, Park Road, Pinelands, 7405, South Africa\
$^9$Centre for Space Research, North-West University, Potchefstroom 2520, South Africa\
$^{10}$Département de Physique Théorique and Center for Astroparticle Physics, Université de Genève,\
Quai E. Ansermet 24, CH-1211 Genève 4, Switzerland
title: Bayesian Inference for Radio Observations
---
\[firstpage\]
methods: data analysis – methods: statistical – techniques: inteferometric.
Introduction
============
The high sensitivity of the SKA (up to 50 times more sensitive than current instruments [@ska]) combined with a relatively cheap antenna design means a far more careful and detailed treatment of systematics will be required to fully exploit this telescope [@noordam2]. The current approach to this calibration problem iteratively applies deconvolution methods such as CLEAN [@clean], alternating with sky and instrumental modelling to determine the best-fitting, calibrated image [@selfcal; @kamezi1; @kamezi2; @bhatnagar]. This provides only a point estimate of the model parameters which will in general differ from the true parameters due to random noise [@ensslin].
A more rigorous approach is to infer the science and instrumental parameters simultaneously, deriving accurate uncertainties and correlations between them. Work in this direction includes improvements on the self-calibration algorithm [@selfcal; @ensslin; @ensslin2; @dorn] and some extensions to the RESOLVE algorithm [@junklewitz2; @junklewitz]. There has also been considerable effort in this direction in producing a maximum posterior image for the data and dealing with certain calibration parameters [@sutton; @sutter; @sutter2]. These works each solve specific aspects of the calibration and deconvolution problem, but so far do not explore the full posterior distribution, giving an inaccurate estimation of the uncertainties and correlations, and still rely on producing a single image (i.e. a point estimate).
We propose instead a new technique, called Bayesian Inference for Radio Observations (BIRO), which is able to: include any source of instrumental uncertainty, such as ionospheric effects, pointing errors and primary beam uncertainties, jointly determine the science and instrumental parameters and provide reliable estimates of the uncertainties and correlations on these parameters, in a holistic and mathematically rigorous manner.
A simultaneous analysis requires the full posterior probability distribution of the parameters, which can naturally be sampled in the Bayesian formalism by using (for example) MCMC [@metropolis; @hastings] or nested sampling [@skilling]. Our new technique, BIRO, fits models including both instrumental and science parameters directly to the raw visibility data. We use the MeqTrees [@noordam] software, which implements the Radio Interferometry Measurement Equation (RIME) [@hamaker], for the modelling of the sky and instrumental effects. This technique thus obviates the need for intermediate imaging and map-making. The rigorous statistical use of all available information allows this technique to open new discovery windows, solving previously intractable problems, and is applicable to all interferometers and problems in radio interferometry.
This paper is arranged as follows: in section \[sec:bayes\] we provide an introduction to Bayesian statistics and illustrate the use of the RIME for modelling in the BIRO algorithm in section \[sec:rime\]. We then apply BIRO to two key simulated datasets to demonstrate its power: In section \[sec:joint\], we jointly fit all scientific (source flux densities) and instrumental parameters (pointing errors, primary beam parameters and receiver noise) to a dataset suffering from direction-dependent instrumental effects. In section \[sec:source\_sep\], we focus on the problem of reliably distinguishing between an extended source, point source and a pair of close point sources, for sources on sub-synthesised beam scales. We conclude in section \[sec:conclusions\].
Bayesian statistics {#sec:bayes}
===================
The problem of obtaining the most information possible from an incomplete dataset, such as obtained by an interferometer, is perfectly suited to the application of Bayesian statistics. These allow the fitting of arbitrarily complex models to data, providing reliable uncertainty estimates for the parameters. Bayes’ theorem allows the use of a familiar quantity, the likelihood, to answer the question one is really interested in: what is the probability of an hypothesis, given the data in hand? This probability is known as the posterior and indicates by how much our degree of belief in the hypothesis has been updated by the new data. Simple application of Bayes’ theorem also allows a robust and intuitive way to compare models, which we will require for the second example problem in this paper. What follows here is a brief overview of Bayesian theory, see @trotta for a more in-depth review.
From Bayes’ theorem, the probability distribution, $\mathcal{P}\left(\mathbf{\Theta}|\mathrm{\mathbf{D}},H\right)$, of the values of parameters $\mathbf{\Theta}$, the quantity that is actually sought, given the data $\mathrm{\mathbf{D}}$ that are in-hand and a model $H$ (hypothesis plus any assumptions), is: $$\mathcal{P}\left(\mathbf{\Theta}|\mathrm{\mathbf{D}},H\right)
= \frac{
\mathcal{L}\left(\mathrm{\mathbf{D}}|\mathbf{\Theta},H\right)
\mathit{\Pi}\left(\mathbf{\Theta}| H\right)}
{\mathcal{Z}\left(\mathrm{\mathbf{D}}| H\right)}.$$ This is known as the posterior probability distribution. The likelihood $\mathcal{L}\left(\mathrm{\mathbf{D}}|\mathbf{\Theta},H\right)$, which encodes any constraints imposed by observations, is the probability distribution of the data given parameter values and a model.
The prior $\mathit{\Pi}\left(\mathbf{\Theta}| H\right)$ includes any prior knowledge of or prejudices about the parameter values. $\mathcal{Z}\left(\mathrm{\mathbf{D}}| H\right)$ is the integral of $\mathcal{L}\left(\mathrm{\mathbf{D}}|\mathbf{\Theta},H\right)
\mathit{\Pi}\left(\mathbf{\Theta}| H\right)$ over all $\mathbf{\Theta}$, not simply normalizing the posterior $\mathcal{P}\left(\mathbf{\Theta}|\mathrm{\mathbf{D}},H\right)$, but also allowing selection of different models by comparing their values quantitatively. This so-called evidence, $\mathcal{Z}\left(\mathrm{\mathbf{D}}| H\right)$, automatically includes an Occam’s razor effect, penalising models with a large number of parameters that are not preferred by the data. By computing the evidence for a range of models we can select the best model by maximising the evidence.
For this work, the likelihood function is $$\label{eq:likelihood}
\mathcal{L}\left(\mathrm{\mathbf{D}}|\mathbf{\Theta},H\right)=\frac{1}{(2\pi\sigma^2)^{N/2}} \, \text{exp}\bigg[ -\bigg(\displaystyle \sum_{i=1}^N \Big({\mathsf{{V_i}}_{{}}}(\mathbf{\Theta}) - {\mathsf{{\widetilde{V}}}_{{i}}} \Big)^2\bigg)/2\sigma^2 \bigg],$$ where ${\mathsf{{V_i}}_{{}}}(\mathbf{\Theta})$ are the model visibilities produced by MeqTrees (see section \[sec:rime\]), with the parameters $\mathbf{\Theta}$ as input, ${\mathsf{{\widetilde{V}}}_{{i}}}$ are the data visibilities, $N$ is the number of data points. Here we assume the uncertainties on the visibilities are Gaussian and have the same value, $\sigma$, for all datapoints. The best-fitting model corresponds to maximum posterior.
The inferred posterior distributions are full probability distributions rather than a summary mean/median value and a (perhaps covariant) uncertainty, since this represents the total inference about the problem at hand. These distributions may be highly non-Gaussian, making such summary parameters inaccurate.
The application of Bayesian statistics allows one to marginalise out the effects of nuisance parameters, which are parameters such as the beam shape and pointing errors that are not of primary interest, but are unknown and can affect the estimates of the parameters of interest (i.e. science parameters) because of correlations and degeneracies. The marginalised posterior can be written as a function of the parameters of interest, $\mathbf{\Phi}$, the nuisance parameters, $\mathbf{\Psi}$, and the data, $\mathbf{D}$: $$P(\mathbf{\Phi}|\mathbf{D},H) = \int P(\mathbf{\Phi},\mathbf{\Psi}|\mathbf{D},H) \text{d}\mathbf{\Psi},$$ where the integral is performed over the parameter space of $\mathbf{\Psi}$.
The posterior is, thanks to advances in modern computing, fairly easily determined using numerical techniques. In this paper, we use the Markov Chain Monte Carlo (MCMC) Metropolis-Hastings [@metropolis; @hastings] algorithm for the joint scientific and instrumental parameter inference example. We chose MCMC due to its simplicity and the ease with which it handles large numbers of parameters (we have 103 parameters for the first example problem). For our second example, that of model selection related to source separation and extended structure, we require efficient calculation of the Bayesian evidence, something provided naturally by the nested sampling algorithm. We utilise the public code MultiNest [@multinest1; @multinest2] to determine both the parameters and the evidence for model comparison [@jeffreys; @trotta], but for a smaller set of parameters as nested sampling grows rapidly in complexity with increasing number of parameters.
Using the RIME for modelling {#sec:rime}
============================
Previous Bayesian visibility analyses [@lancaster; @feroz; @zwart; @ami; @sutter] focused on the sky model and were not generalised to include arbitrary instrumental effects (or were attempting to solve for a much more general sky model resulting in many more parameters, thus needing to fix instrumental parameters). The Radio Interferometry Measurement Equation (RIME) [@hamaker; @smirnov; @smirnov2] provides a powerful framework to easily describe exactly what happens to a signal as it travels from source to telescope, where it is converted into voltages. The RIME is a natural way to model the instrumental and scientific effects that we are inferring through our Bayesian technique. For example, the RIME for a single point source is given by $$\label{eq:rime}
{\mathsf{{V}}_{{pq}}} = \boldsymbol{J}_p {\mathsf{{B}}_{{}}} \boldsymbol{J}_q^H.$$ where ${\mathsf{{B}}_{{}}}$ is the brightness matrix, which describes the sky flux distribution, $\boldsymbol{J}_p$ is the Jones matrix [@jones] for antenna $p$, containing all instrumental and atmospheric effects that interfere with the signal, $\boldsymbol{J}_q$ is the Jones matrix for antenna $q$, $H$ indicates the Hermitian of a matrix and ${\mathsf{{V}}_{{pq}}}$ are the visibilities, the outputs of the telescope correlator for baseline $pq$.
The effects that interfere with the signal on its route to the output of the telescope can each be described by a Jones matrix, with each effect adding a pair of Jones matrices in the ‘onion’ form of the RIME: $${\mathsf{{V}}_{{pq}}} = \boldsymbol{J}_{pn} (\ldots(\boldsymbol{J}_{p2}(\boldsymbol{J}_{p1} {\mathsf{{B}}_{{}}} \boldsymbol{J}_{q1}^H)\boldsymbol{J}_{q2}^H)\ldots)\boldsymbol{J}_{qm}^H.$$
We can go a few steps further and consider the full-sky RIME by integrating over the direction cosines, $l$ and $m$: $$\label{eq:rime_full}
{\mathsf{{V}}_{{pq}}} = \boldsymbol{G}_p \left( \int \int_{lm} \boldsymbol{E}_p \boldsymbol{K}_p {\mathsf{{B}}_{{}}} \boldsymbol{K}_q^H \boldsymbol{E}_q^H \text{d}l \text{d}m \right)\boldsymbol{G}_q^H.$$ Here, $\boldsymbol{K}_p$ and $\boldsymbol{K}_q$ are the Jones matrices describing the geometric delay between antennas $p$ and $q$, $\boldsymbol{G}_p$ represents the *direction-independent* gains for antenna $p$, which we set to unity for all antennas, and $\boldsymbol{E}_p$ is the Jones matrix containing all the *direction-dependent* effects for antenna $p$. We focus in this paper on the more difficult to handle direction-dependent effects, but direction-independent can also be handled with our technique. As with all other Jones matrices, $\boldsymbol{E}_p$ can be written as a product of Jones matrices, each describing a different effect. In section \[sec:joint\], we consider both primary beam effects and pointing errors as examples of direction-dependent effects each with their own Jones matrix.
The RIME is implemented in the general, flexible software MeqTrees [@noordam; @smirnov3; @smirnov4], which allows us to apply it to any sky model and for any telescope. MeqTrees has been useful for predicting the capabilities of future experiments and for understanding the intricacies of current telescopes. Here, we go a step further and use MeqTrees as the modelling step in our Bayesian analysis. In order to test BIRO and compare it with the standard deconvolution approach, we use datasets simulated with MeqTrees over which we have complete control and thus would know if we were correctly recovering the true input parameters.
MeqTrees takes from the user a sky model (such as the number and distribution of sources, their fluxes, shapes etc.) as well as instrumental details (such as the telescope configuration, primary beam pattern, pointing errors, noise, atmospheric effects, ionospheric effects etc.) and uses the measurement equation to produce realistic simulated visibilities that such a telescope would observe.
In order to test the validity of our technique, we only work with simulations in this paper. We use MeqTrees to simulate the data and also to model the sky, to test if we recover the input parameters. MeqTrees can be used to model any telescope configuration and any sky and instrumental effects that can be described with the RIME. While we only concentrate on primary beam and pointing error effects in section \[sec:joint\], in principle, a wide variety of source types and instrumental corruptions can be added in MeqTrees.
Fig. \[fig:flowchart\] shows a schematic overview of the BIRO approach. At each step in the chain of MCMC or MultiNest, MeqTrees is called with new values for the parameters. MeqTrees then returns a visibility set that can be compared directly with the simulated data, to determine how well the parameters fit. This iterative process allows the determination of the full posterior for the parameters. We do not as yet have a public release of the BIRO code, but plan to in the future where we will integrate MontBlanc, a GPU implementation of the RIME [@perkins] with BIRO. MontBlanc is already publicly available meaning it can be combined with any sampler to allow the user to implement BIRO for themselves.
![The BIRO algorithm. Fixed or initialized inputs are shown in yellow, while the sampling loop is represented by the pink boxes. Data products are in blue. The main iteration loop occurs within either the MCMC or MultiNest algorithm (depending on the problem), where new parameters are used in each iteration to compute the likelihood. The initial parameters are drawn from the prior, which generally restricts the parameter ranges. In the final step, the ensemble of sky realisations can be generated with MeqTrees using the parameter samples in the posterior, if required.[]{data-label="fig:flowchart"}](fig1){width="1\linewidth"}
Example 1: Joint inference of scientific and instrumental parameters {#sec:joint}
====================================================================
In this example, we use BIRO to jointly estimate the scientific parameters and nuisance instrumental parameters. Below we describe the model and simulated dataset used, and details of the MCMC analysis, and show that the instrumental parameters studied are tightly correlated with the scientific parameters, a fact that cannot be ignored when determining these parameters.
Simulated data and parameters of the model
------------------------------------------
### Telescope configuration
We use MeqTrees to simulate observations with the Westerbork Synthesis Radio Telescope (WSRT) [@wsrt], a 14-element East-West array with 25m diameter dishes. All our WSRT simulations use an integration time of 30 seconds and a total observation time of 12 hours at a frequency of 1.4 GHz. We use a narrow bandwidth of 125kHz, a single channel (for simplicity) and include noise with a standard deviation of 0.1 Jy/visibility. At this frequency, WSRT has a field of view of 0.5-0.6 degrees and a synthesised beam width of around 13 arcsec FWHM (full width at half maximum)[^1].
### Scientific parameters
![The simulated, noise-free sky model with 17 sources with flux densities varying between 0.03 and 3.13 Jy.[]{data-label="fig:field"}](fig2){width="1\linewidth"}
![The dirty dataset for the model of Fig. \[fig:field\], as the telescope would see it (the colours are histogram-equalised to improve contrast). The image is produced directly from the visbilities and shows the typical ring structure around bright sources that is seen in interferometric data, due to the missing angular-scale information in the dataset. The rms noise in flux density is about 0.28 mJy.[]{data-label="fig:dirty"}](fig3){width="1\linewidth"}
The simulated field consists of 17 unpolarised, point sources with known positions. The science goal was to determine the flux densities of these sources. We based the simulation on an existing field observed by WSRT, consisting of sources with a range of fluxes (from $0.03-3.13$ Jy). This is a very simple sky model, consisting only of point sources, whereas in the second example of the paper, we address modelling of extended sources. We do not explore the possibility of extended sources of arbitrary shapes, as this is out of the scope of this paper, but this should be possible using shapelets, such as employed in the existing PyBDSM software[^2]. The brightness matrix in Eq.(\[eq:rime\_full\]) for an unpolarised point source is written as: $${\mathsf{{B}}_{{}}}^{\scriptscriptstyle \text{POINT}} =
\begin{pmatrix}
I & 0\\
0 & I
\end{pmatrix},$$ where $I$ is the intensity.
Fig. \[fig:field\] shows an image of the true input model without any instrumental effects, while Fig. \[fig:dirty\] shows the dirty image of the sky.
### Instrumental parameters
[**Beam width**]{}\
Knowing the primary beam pattern is critical for any astronomical survey. Current practice is to determine the primary beam pattern using a technique such as holography [@scott], then fix a beam model, without propagating any uncertainty information into the estimates of the science parameters. Since the primary beam directly attenuates the flux distribution of the sky, even a small error in the beam model can lead to large biases. We thus include beam parameters in our analysis. WSRT commonly adopts a simple model for the primary beam, namely: $\text{cos}^3(c\nu\theta)$, where $\nu$ is the observing frequency (in GHz), $\theta$ is the distance from the pointing centre in degrees and $c$ is the beam factor (in 1/GHz). The beam factor (or beam width) is known to vary slightly with frequency. As proof of concept, we assume it is unknown, and include it as a further instrumental parameter. One could provide a more complex model for the primary beam and easily fit those parameters with this technique as well, comparing the models with the Bayesian evidence. The model for the beam enters the RIME of Eq.(\[eq:rime\_full\]) as a direction-dependent Jones matrix: $$\label{eq:jones_beam}
\boldsymbol{E}^{\scriptscriptstyle \text{BEAM}}(l,m) = \text{cos}^3(c\nu\sqrt{l^2+m^2}) \,\boldsymbol{I},$$ where $\boldsymbol{I}$ is the identity matrix.\
[**Pointing errors**]{}\
Pointing errors can substantially corrupt radio observations and are known to be a limiting factor in deep observations with WSRT [@smirnov3] and other telescopes. The greatest effect is on sources on the flank of the primary beam, where the gradient of the beam pattern is steep, and a small pointing error produces a larger error in apparent flux (compared to the centre of the beam). Since the errors can be different from antenna to antenna, this produces errors on the observed visibility amplitudes, which translates into artefacts in the image. Essentially each source is ‘defocussed’ in a complicated way. Thus, we can immediately suspect there will be a correlation between the pointing errors and source flux densities. Two prior approaches to inferring pointing errors directly from the data have hinged on maximum-likelihood estimates. These are the pointing selfcal algorithm [@pointing-selfcal] and direct fitting with MeqTrees (Smirnov 2011[^3]). Neither approach estimates the correlation between pointing errors and source parameters, which the Bayesian approach naturally provides. We inject time-varying polynomial pointing errors for each of the 14 WSRT antennas. We use a second order polynomial for each pointing error and fit for the coefficients. A polynomial pointing error in each orthogonal direction for each antenna results in a total of 84 pointing-error parameters. The pointing errors are written as a Jones matrix in Eq.(\[eq:rime\_full\]): $$\label{eq:jones_pe}
\boldsymbol{E}^{\scriptscriptstyle \text{PE}}_{p}(l,m)= \boldsymbol{E}^{\scriptscriptstyle \text{BEAM}}(l+\delta l_p,m+\delta m_p)\,$$ where $\delta l_p$ and $\delta m_p$ are the pointing errors in the right ascension and declination direction respectively, for antenna $p$. The pointing errors are taken to be time-varying polynomials, written as: $$\delta l_p = c_2t^2+c_1t+c_0,$$ and similarly for $\delta m_p$, where $t$ is time (rescaled over the observation) and $c_k$ are the coefficients we determine with MCMC.\
[**Noise**]{}\
The noise on the visibilities is expected to be Gaussian, stationary and uncorrelated. Noise level can be estimated with some precision from the known system temperature, here however we show than it can also be inferred accurately directly from the data. We thus included one final parameter for the standard deviation of the noise on the visibilities.
### Resulting measurement equation
The RIME for this example problem is thus: $$\resizebox{0.99\linewidth}{!}{${\mathsf{{V}}_{{pq}}} = \displaystyle \sum_s \left( \boldsymbol{E}^{\scriptscriptstyle \text{BEAM}}(l_s+\delta l_p,m_s+\delta m_p) \boldsymbol{K}^{(s)}_p {\mathsf{{B}}_{{s}}}^{\scriptscriptstyle \text{POINT}} \boldsymbol{K}_q^{(s),H} (\boldsymbol{E}^{\scriptscriptstyle \text{BEAM}})^H(l_s+\delta l_p,m_s+\delta m_p)\right),$}
\label{eq:rime1}$$
where $s$ runs from 1 to 17 over all the sources. This brings the total to 103 parameters: 17 scientific (the flux densities of the sources) and 86 instrumental (84 pointing error parameters, the beam width and the noise). The full model can be visualised in the Bayesian factor graph of Fig. \[fig:factor\] and a more detailed description of factor graphs is given in section \[sec:factor\] of the appendix.
![Bayesian factor graph (see section \[sec:factor\] of the appendix) of the model for the first simulated dataset. All parameters we estimate with MCMC are the constants, without any circles around them, coloured blue. The ${\mathsf{{V}}_{{pq}}}$ are the observed visibilities, which are drawn from a normal distribution of mean ${\mathsf{{\widetilde{V}}}_{{pq}}}$ (the unobserved, true visbilities) and standard deviation $\sigma$, which is one of the parameters we estimate with MCMC. These ‘true’ visibilities are governed by the RIME, which is here simplified graphically to two components, the brightness matrix, ${\mathsf{{B}}_{{}}}$, and the Jones’ matrices of the antennas, $\boldsymbol{J}_p,\,\boldsymbol{J}_q$. The flux densities of the 17 sources are represented by $f_i$, which form components of ${\mathsf{{B}}_{{}}}$. The coefficients of the polynomial time-varying pointing errors, $l_jc_k$ and $m_jc_k$ (where $j$ represents the antenna number and $k$ is the number of polynomial coefficient) enter the Jones matrices, along with the beam width, $bw$.[]{data-label="fig:factor"}](fig4){width="1\linewidth"}
Using MCMC for joint parameter inference
----------------------------------------
The initial step of our analysis was to choose an appropriate sky model in MeqTrees (specifying the brightness matrix in Eq.(\[eq:rime\_full\])) and select the telescope configuration corresponding to the dataset including all known sources of interference and instrumental errors (the Jones matrices in Eq.(\[eq:rime\_full\])). We vary all the parameters within the model – the flux densities, pointing errors, beam width and noise – using MCMC. Fig. \[fig:flowchart\] illustrates how the sampling algorithm repeatedly calls MeqTrees with new parameter values and evaluates the likelihood. MCMC uses the likelihood (Eq.(\[eq:likelihood\])) to determine the best-fitting parameter values and to explore the surrounding parameter space, thus determining the uncertainties and correlations for all parameters.
Technical details and priors
----------------------------
Due to the large volume of the parameter space, we use a standard, gradient-based optimisation algorithm to get close to the best-fitting parameter values and provide a good starting point for the MCMC. We run several chains in parallel, each of around $500, 000$ steps, repeatedly computing and diagonalising the covariance matrix to improve convergence, and we test convergence using the Gelman-Rubin statistic [@gelman]. The estimated parameters and their uncertainties are determined by finding the mean and standard deviation (using percentiles) from the marginalised one-dimensional posterior for each parameter. For this particular setup, MeqTrees takes about 0.4s for one likelihood calculation, parallelised using 4 cores of 2.2 GHz each. As 10 chains were run, 40 cores in total were used resulting in approximately 55 CPU hours for convergence per dataset.
We apply a uniform prior to the pointing error parameters, restricting them to the broad range of $\pm200$ arcseconds. We also restrict the beam width to be positive, and vary the noise on the visibilities in logarithmic space (with an infinitely broad prior in log-space). We do not restrict the ranges of the flux densities.
Comparison with CLEAN plus source extraction {#sec:clean}
--------------------------------------------
To compare our technique with the standard approach, we apply CLEAN followed by a source-extraction algorithm to determine the flux densities of the sources (we call this combination CLEAN+SE), without any instrumental calibration. We do not use any calibration algorithms such as self-cal, because it would have no benefit: our dataset only has direction-dependent instrumental effects, whereas self-cal can only correct for direction-independent effects. Current approaches to direction-dependent calibration are of no help here because:
1. Direction-dependent solutions (such as peeling, or differential gains) can in principle solve for the variable gains induced by pointing error, given a prior source model. However, this destroys information on the source, since deviations between the true sky and the prior model are completely absorbed by such gain solutions.
2. Pointing selfcal should in principle improve the CLEAN maps and thus produce better source model estimates. However, implementations of this remain unavailable to the public.
3. MeqTrees should in principle be able to do a maximum-likelihood solution for the source parameters and pointing errors simultaneously. However, only solutions for the latter has been demonstrated to work in practice and as we have argued, a maximum-likelihood solution produces a point estimate for the parameters which may be biased due to correlations.
Instead, we apply a naïve CLEAN algorithm, followed by source extraction, to compare with BIRO as a worst case scenario in the case of time-varying pointing errors. Note that we do provide prior information on the positions of the sources to CLEAN, in the form of CLEAN boxes.
We use the CLEAN implementation (specifically the Cotton-Schwab algorithm) in the software package CASA[^4] to image the simulated datasets. The images were made with robust weighting with a robustness parameter of $-1.0$. We did 1000 iterations of CLEAN with a loop gain of 0.1. Interactive cleaning was performed on the visibility data twice, once with masks defined around known source positions and then with masks defined around only those sources that were found during the cleaning procedure. The source extraction was performed interactively using PyBDSM to ensure that the artefacts were not wrongly identified as sources.
Results
-------
To illustrate fitting a model to the raw data, we plot a subset of the visibilities in Fig. \[fig:vis\] with the best-fit visibilities as obtained by BIRO. Fig. \[fig:compare\] (with numerical details in Table \[tab:clean\]) shows the comparison between the flux densities obtained by CLEAN+SE and those by BIRO. The flux densities of CLEAN+SE are on average biased due to undealt-with correlations with the pointing errors and underestimated uncertainties. Additionally, because of the time-varying pointing errors corrupting the data, CLEAN+SE only manages to find 5 of the 17 sources. With polynomial pointing errors included in the simulations, bright artefacts dominated the final image resulting in the weaker sources being swamped. In contrast, because these correlations are taken into account, the Bayesian approach is able to recover the true flux densities for all sources and to determine error bars that include the effects of all nuisance parameters. Without the instrumental errors, BIRO achieves similar flux estimates to CLEAN+SE.
![Example of fitting a model to the raw data. Plotted are the real component of the visibilities for a single baseline (between antenna 0 and 1) and for the single channel of the data, in black. The best fitting model line is overplotted in dark blue, with a band of uncertainty of 0.1 Jy (the original noise added to the simulation) in pale blue.[]{data-label="fig:vis"}](fig5){width="1\linewidth"}
Fig. \[fig:covmat\] shows a subset of the covariance matrix between parameters and Fig. \[fig:contour\] shows an example 1$\sigma$ and 2$\sigma$ contour plot between pairs of parameters. The key result of Fig. \[fig:covmat\] is that it highlights the significant and complex correlations between the pointing errors and flux densities, i.e. the instrumental and science parameters, which therefore need to be estimated jointly allowing for the correlations.
The (anti-)correlations between pointing errors and flux densities are easy to understand qualitatively. Consider a source on the flank of the main lobe of the primary beam, e.g. on the half-power point. If a given antenna mispoints *towards* the source, the source will be subject to a higher primary beam gain, in other words, it will be perceived as brighter by all baselines involving that antenna. Mispointing *away* from the source has the opposite effect. The nature of the correlation will also strongly depend on the position of the source with respect to the pointing centre. For example, a source near the centre of the main lobe (i.e. on a ‘flat’ part of the primary beam pattern) will correlate very weakly with pointing error, while a source on the inner flank of the first sidelobe will correlate with mispointing *away* rather than *towards*. Since different baselines contribute to different Fourier mode measurements, pointing error will also have a complicated interaction with perceived source structure. Similar arguments apply to beamwidth.
Deriving the exact quantitative nature of this correlation analytically is highly impractical, which is why a technique like BIRO proves so powerful. This covariance matrix could be used to assist in calibration, study calibration parameters or as input to future MCMC analyses on similar datasets.
{width="0.9\linewidth"}
Example 2: Model comparison {#sec:source_sep}
===========================
In this example problem, we show that BIRO is able, using model selection [@jeffreys; @trotta], to choose the correct model in each of three different cases, distinguishing between an extended source, an unresolved point source and two close (sub-synthesised-beam) sources. The sources recovered are all smaller than the synthesised beam. This is known as super-resolution and has recently been shown to be possible with compressive sensing [@wiaux; @li; @carrillo1; @carrillo2; @honma] (and to some extent [@marti]). Here we use the Bayesian evidence to determine the correct model of these sub-synthesised-beam sources, with statistical significance. Although in this example problem we exclude instrumental effects, they can, in general, be included as in example 1.
Simulated datasets and models
-----------------------------
The datasets for this example use the same frequency, bandwidth, integration time and noise characteristics as the dataset simulated in section \[sec:joint\]. We simulate three datasets with three different sky models with all the sources away from the phase centre: a point source, a sub-synthesised-beam extended source modelled as a Gaussian and two point sources separated by the distance the size of that Gaussian. No instrumental effects were included in the model-selection simulations and the beam width and noise were assumed to be known. Fig. \[fig:source\_sep\] shows the input model for all three cases in the left column.
The point sources are parametrized by the Stokes I flux density and the position as the distance from the phase centre, along two mutually perpendicular axes, $l$ and $m$. The extended Gaussian source has three more parameters in the form of the projections of the major axis on the $l$ and $m$ axes and the ratio of the minor to major axis, defined as: $$\begin{gathered}
\label{eq:gaussshape}
l_{{\scriptscriptstyle \perp}}=e_{\text{maj}}\,\text{sin}(\alpha)\\
m_{{\scriptscriptstyle \perp}}=e_{\text{maj}}\,\text{cos}(\alpha)\\
r= e_{\text{min}}/e_{\text{maj}},
\end{gathered}$$ where $e_{\text{maj}}$ and $e_{\text{min}}$ are the major and minor axes of the Gaussian source and $\alpha$ is the position angle (the angle of rotation of the extended source). See Fig. \[fig:gaussian\] for a visual description. The brightness matrix of Eq.(\[eq:rime\_full\]) for an extended Gaussian is simply the product of a Gaussian and the brightness matrix for a point source. The RIME is simple in this example, since there are no instrumental effects apart from the usual phase shift between antennas: $${\mathsf{{V}}_{{pq}}} = \displaystyle \sum_s \left( \int \int_{lm} \boldsymbol{K}^{(s)}_p f(l,m) {\mathsf{{B}}_{{s}}}^{\scriptscriptstyle \text{POINT}} \boldsymbol{K}_q^{(s),H} \text{d}l \text{d}m \right),$$ where $f(l,m)$ is a Gaussian in $l$ and $m$ for the extended source case and $f$ is a delta function for the one and two-source models. Also in the one and two-source models, $l$ and $m$ reduce to single points $l_s$ and $m_s$, as in Eq.(\[eq:rime1\]).
![Estimated vs true flux densities of the sources with error bars as estimated by BIRO (blue circles) and by a CLEAN+Source Extraction algorithm (red triangles). Note that CLEAN+SE only detects 5 out of 17 sources. The BIRO error bars are the standard deviation of marginalised one-dimensional posterior for each flux parameter. While the BIRO results are unbiased, CLEAN+SE has two problems: it underestimates the error bars and yields biased estimates of the flux densities of up to $44\sigma$. The reader is reminded that this dataset contains no direction-independent effects that may normally cause biases in a CLEAN analysis; these biases are instead due entirely to the complexities in the dataset introduced by the time-varying pointing errors.[]{data-label="fig:compare"}](fig7){width="0.95\linewidth"}
![Credible interval contour plots between a subset of parameters. The 1$\sigma$ and 2$\sigma$ probability densities are shown in dark and light colours respectively. The true (input) parameters are marked with a black star. The pairs of parameters are: *Upper left:* The two highest order coefficients of the pointing error in the right ascension direction for antenna 9. *Upper right:* Flux densities of two of the sources. *Lower left:* The flux density of the 17th source vs the beam width. *Lower right:* The constant term from the polynomial pointing error in the right ascension direction for antenna 10 vs the flux density of the 15th source.[]{data-label="fig:contour"}](fig8){width="1\linewidth"}
![The parameterisation of a Gaussian extended source in MeqTrees. Here, $e_{\text{maj}}$ and $e_{\text{min}}$ are the major and minor axes of the Gaussian and $\alpha$ is the position angle. MeqTrees uses $l_{{\scriptscriptstyle \perp}}$, $m_{{\scriptscriptstyle \perp}}$ and $r= e_{\text{min}}/e_{\text{maj}}$ in its parameterisation of a Gaussian.[]{data-label="fig:gaussian"}](fig9){width="0.6\linewidth"}
Using MultiNest for model selection
-----------------------------------
We use MultiNest for calculating the Bayesian evidence (see section \[sec:bayes\]) and MeqTrees for predicting the model visibilities from the sampled source parameters from which the likelihood is computed iteratively. The likelihood is computed according to Eq.(\[eq:likelihood\]). The posterior probability distributions are obtained as a by-product along with the uncertainties in the best-fit parameter values and the Bayesian evidence.
For the single-point-source model, we vary three parameters: the flux density and relative source position, $l$ and $m$. We similarly vary the flux densities and positions of the two sources in the two-source model. The Gaussian extended source model has six parameters: the flux density, position coordinates, and the shape parameters ($l_{{\scriptscriptstyle \perp}}$, $m_{{\scriptscriptstyle \perp}}$, $r$).
We generate a unique, simulated dataset for each of the three cases and then fit each of the three models to them, to see if the correct model is selected in each case. MultiNest fits for the parameters, their uncertainties and correlations (just as MCMC does in example 1), but also returns the evidence, $\mathcal{Z}\left(\mathrm{\mathbf{D}}| H\right)$ (the probability of the data, given the hypothesis). By taking the ratio of evidences, one can determine whether one model is favoured over another, and by how much. The Jeffrey’s scale [@jeffreys; @trotta] provides an intuitive way of deciding whether the evidence is strong enough to select a model, based on odds derived directly from the evidence.
Technical details and priors
----------------------------
We use uniform priors for all the source parameters. The flux density is restricted to the range 0 to 2 Jy. The position parameters are allowed to be both positive and negative in the range -$25\arcsec$ to $25\arcsec$ since the position is measured relative to the phase centre. For the shape parameters of the extended source, ($l_{{\scriptscriptstyle \perp}}$ and $m_{{\scriptscriptstyle \perp}}$), we allow the prior ranges to be big enough to encompass the point-spread-function (PSF) of the interferometer and no more, since we are dealing with sub-synthesised-beam sources. This translates to a range of $0\arcsec$ to $20\arcsec$ for $l_{{\scriptscriptstyle \perp}}$ and -$20\arcsec$ to $20\arcsec$ for $m_{{\scriptscriptstyle \perp}}$. Finally, we restrict the minor-to-major axis ratio ($r$) to be positive, but less than unity to be physically meaningful. We found that using 1000 live points achieved good results from MultiNest.
Results
-------
The relative logarithmic evidences are computed for each model giving the relative confidence with which one model is preferred over another (see Table \[tab:evmatrix\]). We find that the correct hypothesis is selected in all cases, at odds of $10^{593}$:1, $10^{993}$:1 and 62:1, for the two-point-source, extended-source and single-point-source models respectively. Using model selection, BIRO is able to select the correct model in all three cases (the model with the highest evidence), showing it can perform source separation even on sub-synthesised-beam scales.
We computed a ‘best-fitting’ image by running MeqTrees with the maximum posterior model and parameters in each of the three cases, to compare with the CLEANed image (see Fig. \[fig:source\_sep\]). We use the same CLEAN parameters as in section \[sec:clean\]. The CLEANed images (at least in this case, without an enforced smaller beam size) are unable to reach the sub-synthesised beam scales achievable by BIRO.
In Fig. \[fig:crossover\], we determine the point at which model selection fails to distinguish an extended source from a point source for different source sizes and signal-to-noise ratios (SNRs). Any evidence lower than ‘strong’ is not usually considered high enough to say either way which model is correct. Perhaps obviously, at high SNR extremely small sources can be detected (around 1.0 arcseconds) and sources become more difficult to distinguish as the SNR is reduced.
Video \[vid:gaussian\] in the online-only content shows visually how MultiNest converges to the correct model, exploring the posterior as it goes, for the extended source model. Each frame is an image generated using the parameters from every $40^{\text{th}}$ step of the chain.
![*Left column:* The true sky for the extended Gaussian, single point source and two point sources models (from top to bottom). *Middle column:* The CLEANed image for the three models. *Right column:* The maximum posterior BIRO image for the three models. The purple contour in each image indicates the size of the synthesised beam, as returned by CLEAN (note that the sources are all much smaller than the synthesised beam). BIRO recovers the correct input model each time while CLEAN is unable to distinguish between the models at the same SNR (in this case the SNR was 1000).[]{data-label="fig:source_sep"}](fig10){width="1\linewidth"}
![Relative natural log-evidence (i.e. the natural logarithm of the ratio of the Bayesian evidence for the true model to that of a single point source) as a function of Gaussian source size, for the extended source input model, showing the evidence-crossover points for different source sizes and signal-to-noise ratios (peak flux to background noise). The horizontal axis gives the size of the circular Gaussian source in the input model (the reader is reminded that the FWHM of the synthesised beam is around 13 arcseconds). The vertical axis gives the odds in favour of the Gaussian source model when model comparison is performed for the Gaussian model against a point-source model. The more positive the relative log-evidence is, the more strongly is the Gaussian model favoured. Each curve on the graph is for a different noise level with the approximate (map) SNRs shown in the legend.[]{data-label="fig:crossover"}](fig11){width="0.99\linewidth"}
Discussion and conclusions {#sec:conclusions}
==========================
We have introduced the technique Bayesian Inference for Radio Observations (BIRO), a Bayesian approach to the deconvolution problem of radio interferometry. Instead of making an image and then performing source extraction, BIRO uses MCMC or nested sampling to fit models directly to the visibility data and obtain the posterior for the parameters of interest, as well as nuisance parameters.
In the first example problem, we focused on the relationship between scientific and instrumental parameters. It was found that all parameter estimates from BIRO were consistent within their error bars with the true values. As well as determining the uncertainties of the parameters, BIRO also returns the covariance matrix between them, as a by-product of the full posterior. Our work shows these correlations are complicated and non-negligible. BIRO effortlessly incorporates the effects of the correlations in the estimates of the marginalised uncertainties on the individual parameters, as well as providing a way to study these correlations in the form of the covariance matrix. We compared our results to a standard CLEAN algorithm, without calibration (since our simulated data contains *only* direction-dependent effects and publicly available calibration algorithms only deal with direction-independent effects). Because of the time-varying pointing errors we introduce to the dataset, CLEAN is only able to find 5 out of the 17 sources and returns biased flux densities for them, while BIRO returns unbiased flux densities for all sources. BIRO is also able to correctly determine the coefficients of the time-varying pointing errors, the primary beam width and the noise on the visibilities.
In the second example problem, we addressed the issue of how to determine the best sky model for the data. We worked with three models: a single point source, a Gaussian extended source and two point sources. We simulated data for each of the three models and then, for each dataset, ran MultiNest to fit each of the three models and determine the Bayesian evidence. The evidence then determines the selection of the correct model. All of the sources detected were several times smaller than the synthesised beam, hence we successfully achieved super-resolution as well as source-separation.
This paper constitutes a proof of concept but more work is required before the technique can be easily applied to interferometric dataset:
1. Firstly, while using a WSRT simulation has relevance to the SKA due to the similar instrumental setup, the SKA will have many more antennas (on the order of a thousand) which will of course result in many more instrumental parameters (and indirectly more science parameters as the source count increases with sensitivity). Fortunately, while the number of instrumental parameters scales as the number of antennas, $N$, the number of datapoints scales as the number of baselines, i.e. ${\cal O}$($N^2$), meaning it is plausible that one could simultaneously determine the sky and instrumental parameters for large $N$. While the precedent for sampling an extremely large parameter space exists [@jasche], new and sophisticated sampling techniques [@hmc1; @hmc2; @affine1; @affine2] (which are also easily parallelised) will be required to improve convergence in the thousand-parameter regime, especially as the non-linear nature of the modelling makes sampling inefficient (as addressed in [@jasche]).
2. Secondly, the Bayesian approach is far more computationally intensive than standard deconvolution, taking hours (55 CPU hours in the case of example 1) to converge to the correct posterior distribution. The complexity of the likelihood computation scales as the number of antennas squared (i.e. the number of baselines), making an SKA-like computation difficult with the current setup. However, the RIME is intrinsically highly parallelisable allowing an efficient implementation of MeqTrees on GPUs. Preliminary work on a GPU implementation indicates a speed-up of the likelihood computation of about 250 times [@perkins]. This means this technique can be applied to data from existing telescopes such as ALMA [@alma] and LOFAR [@lofar], using current computer clusters.
3. Thirdly, we need to address the problem of not knowing the sky model beforehand, which is a common difficulty when dealing with calibration but is particularly important here, as a Bayesian analysis relies on a good model. There are a number of ways to tackle this issue which we hope to address in future publications. A simple, but computationally-intensive, solution would be to run several different models (with increasing numbers of sources) and select between them using the Bayesian evidence. Another possible approach is to use a deconvolution algorithm, like CLEAN or RESOLVE, to get an initial set of sources and then iterate between deconvolution and the best fit of BIRO to get a subsequently better model. A more rigorous solution would be to use an algorithm like birth-death [@stephens] or reversible jump [@green] MCMC, which is able to determine both the number of parameters required and the posterior for them simultaneously. A further possibility is to combine the more general approach proposed in [@sutter2] and [@junklewitz2], that divides up the field into many ‘pixels’ that are then allowed to vary, with the calibration capabilities of BIRO to produce estimates of the sky model. This is even more computationally challenging however but would provide a more general and robust solution.\
BIRO is not only useful for dealing with systematics, which will become more important as telescopes become more sensitive, but it is also a powerful technique for lending statistical strength to topical scientific questions. Potential applications include: structures of black hole systems, jet emission in active galaxies, time variability of objects and radio weak lensing. BIRO allows a holistic way to include instrumental effects while at the same time returning the science we are interested in. By leveraging the power of Bayesian statistics, BIRO uses all information available to get the most out of interferometric datasets.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank Paul Sutter for his useful referee comments. M.L. and J.Z. are grateful to the South Africa National Research Foundation Square Kilometre Array Project for financial support. M.L. acknowledges support from the University of Cape Town and resources from the African Institute for Mathematical Sciences. O.S. is supported by the South African Research Chairs Initiative of the Department of Science and Technology and National Research Foundation. B.B. acknowledges funding from the South African National Research Foundation. I.N. acknowledges the MeerKAT HPC for Radio Astronomy Programme. Part of the computations were performed using facilities provided by the University of Cape Town’s ICTS High Performance Computing team: <http://hpc.uct.ac.za>. This project was initiated at the SuperJEDI Mauritius conference.
Author Contributions {#author-contributions .unnumbered}
====================
M.L. performed the research for the joint estimation example and wrote the majority of the paper. I.N. performed the research for the source separation problem and wrote the corresponding sections of the paper and did some CLEAN analysis and wrote the corresponding section. J.Z. assisted with the source separation research and wrote part of the paper. O.S. provided the simulated datasets and led the technical aspects of radio observation modeling. B.B. developed the original idea. N.O. performed some CLEAN analysis. M.K. contributed to the conceptual development of ideas and provided support with the statistical methods. All authors commented on the research and edited the paper.
\[References\]
Bayesian Factor Graphs {#sec:factor}
======================
Here we introduce Bayesian factor graphs, useful tools for visualising Bayesian models, which we use to describe the model of section \[sec:joint\]. We make use of the directed factor graph notation, developed in [@dietz], to visualise how the parameters in our models depend on one another. Table \[table:factor\] defines the graphical primitives of a factor graph. Figure \[fig:factor\_ex\] demonstrates the use of the factor graph notation in a simple example.
![A simple example factor graph. In this model, the data are represented by a vector $x_i$, which we suspect is normally distributed. This is modelled by a normal distribution (represented by the factor labeled $\mathcal{N}$) which is governed by the parameters $\mu$ and $\sigma$. These constants would be the parameters we would want to estimate with an MCMC or MultiNest analysis.[]{data-label="fig:factor_ex"}](figA1){width="0.3\linewidth"}
{width="0.7\linewidth"} \[table:factor\]
\[lastpage\]
[^1]: WSRT Guide to Observations, [www.astron.nl/radio-observatory/astronomers/wsrt-guide-observations/5-technical-information/5-technical-informatio](www.astron.nl/radio-observatory/astronomers/wsrt-guide-observations/5-technical-information/5-technical-informatio)
[^2]: \[note1\]Python Blob Detection and Source Measurement software, [www.lofar.org/wiki/doku.php?id=public:user\_software:pybdsm](www.lofar.org/wiki/doku.php?id=public:user_software:pybdsm)
[^3]: <https://indico.skatelescope.org/getFile.py/access?contribId=20&sessionId=9&resId=0&materialId=0&confId=171>
[^4]: Common Astronomy Software Applications, <http://casa.nrao.edu/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a novel classification scheme for stars evolving in the transition phase between the Asymptotic Giant Branch (AGB) to the Planetary Nebula (PN) stage based on the results obtained with ISO-SWS. With the better sensitivity and higher spatial resolution of Spitzer this analysis can now be extended to a larger number of sources located in the Galactic Bulge, in the Magellanic Clouds and in other Galaxies of the Local Group, offering an excellent opportunity to study the validity of the proposed scheme in environments which are free of the distance scale bias that hinder the observations made with ISO on galactic sources located at uncertain distances. The new observations will be used to test the current evolutionary models which predict the dredge-up of processed material to the surface of low- and intermediate-mass stars as a function of the progenitor mass and of the metallicity.'
author:
- 'P. García-Lario$^1$, J.V. Perea Calderón$^2$'
title: 'From ISO to Spitzer: a new view of the AGB-PN transition phase'
---
A novel classification scheme based on ISO results
==================================================
Based on the analysis of $\sim$350 ISO SWS spectra of sources in the AGB-PN transition phase retrieved from the ISO Data Archive we have been able to propose a novel classification scheme which takes into account the evolution of the overall shape of the infrared spectrum in combination with the gas-phase molecular bands and solid state features detected in the SWS spectral range. We identify three main chemical evolutionary branches which are interpreted as the result of the evolution of low-mass ($<$1.5–2.0 M$_{\odot}$), intermediate-mass (2.0–3.0 M$_{\odot}$) and high-mass ($>$ 3.0–4.0 M$_{\odot}$) AGB stars, respectively.
The sequences reflect the increase of optical thickness in the circumstellar shell of AGB stars followed by its cool down as a consequence of the shell expansion after the end of the strong mass loss phase. It is also a sequence in which we can follow the process of condensation and growth of the dust grains in the envelope until the star becomes a PN. In addition, there is a clear evolution of carbonaceous material from aliphatic to aromatic structures and of the silicates from amorphous to crystalline, which is still not very well understood.
The proposed scheme is consistent with the theoretical model predictions based on the dredge-up of processed material to the surface of AGB stars which result in the transformation of the initially O-rich AGB star into a C-rich AGB star, after a few thermal pulses. The number and efficiency of the thermal pulses in low-mass stars is expected to be small. As a consequence of this, low-mass stars would stay as O-rich during the whole AGB-PN evolution. Only part of them will become low-mass O-rich type II or type III PNe. In the most extreme cases very low-mass stars will not develop an observable PN. Intermediate-mass stars will soon become C-rich and will further evolve as such until the PN stage. High-mass stars will develop very thick envelopes and thus strong silicate absorption bands and will activate the [*hot bottom burning*]{} mechanism which prevents the formation of carbon and favours the production of nitrogen, instead, evolving as O-rich stars until they become high-mass, N-rich type I PNe.
What can Spitzer do that ISO could not do?
==========================================
More observations...
--------------------
The number of sources in the AGB-PN evolutionary phase observed with ISO is small. Some of the new results derived from ISO data will need to be confirmed by extending the observations to a larger number of sources. Spitzer expected lifetime of $\sim$5 yr could provide a huge increase of available data to further study the many questions still left open by ISO.
Of fainter sources...
---------------------
Below the 1–5 Jy level, the sensitivity of ISO-SWS was not enough to obtain spectra with enough quality to derive reliable conclusions. With the much higher sensitivity of Spitzer it will be possible to observe under much better conditions these sources and extend the analysis to other sources several orders of magnitude fainter.
Located at homogeneous distances...
-----------------------------------
The study of the stars in the ISO sample above described is hampered by the poor knowledge of their distances. This problem can be overcome with Spitzer if the analysis concentrates on well-defined samples located at distances which are known with a reasonable accuracy. This includes the Galacic Bulge, the Small and Large Magellanic Clouds and other galaxies of the Local Group, which were forbidden for ISO (because of its poor sensitivity). As a byproduct, we will be able to analyze also the influence of metallicity on the proposed evolutionary scheme, expected to play an important role according to the existing models.
With a higher spatial resolution...
-----------------------------------
It has been suggested that PAHs may be predominantly present in the scattering lobes, while the crystalline silicates are expected to be present in the disks of bipolar sources showing a mixed chemistry. With the high spatial resolution of Spitzer it will be possible to resolve the different emitting regions in sources extended over just a few arcsec showing this dual chemistry.
Extending the search to other galaxies
--------------------------------------
The use of multi-filter photometry with Spitzer (IRS + MIPS) can be a powerful tool to search for new sources in the AGB-PN evolutionary phase. Color-color diagrams based on large-scale maps of galaxies of the Local Group in the adequate filters could provide us with large datasets of new candidate sources which may have escaped from detection in classical optical surveys.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe SMS, our submission to the exact treedepth track of PACE 2020. SMS computes the treedepth of a graph by branching on the **S**mall **M**inimal **S**eparators of the graph.'
author:
- Tuukka Korhonen
bibliography:
- 'paper.bib'
title: SMS in PACE 2020
---
Overview
========
SMS is an exact algorithm implementation for computing treedepth, available in [@tuukka_korhonen_2020_3872898] and in <https://github.com/Laakeri/pace2020-treedepth-exact>. SMS was developed for the fifth PACE challenge (PACE 2020). The main algorithm implemented in SMS is a recursive procedure that branches on minimal separators [@DBLP:journals/dam/DeogunKKM99]. Two variants of the branching algorithm are implemented, one with a heuristic algorithm for enumerating minimal separators and one with an exact algorithm [@DBLP:conf/sea2/Tamaki19]. Several lower bound techniques are implemented within the branching algorithm. Before applying the branching algorithm, preprocessing techniques are applied and a heuristic upper bound for treedepth is computed.
This arXiv version contains an appendix containing proofs for novel techniques used.
Notation
========
Let $G$ be a graph with vertex set $V(G)$ and edge set $E(G)$. The graph $G[X]$ is the induced subgraph of $G$ with vertex set $X$. The set $N(v)$ is the neighborhood of a vertex $v$ and $N(X)$ is the neighborhood of a vertex set $X$. The treedepth of $G$ is denoted by ${\texttt{\textbf{td}}}(G)$. A minimal $a,b$-separator of $G$ is a subset-minimal vertex set $S$ such that the vertices $a$ and $b$ are in different connected components of $G[V(G) \setminus S]$. The set of minimal separators of $G$ for all pairs $a,b \in V(G)$ is denoted by ${\Delta}(G)$ and the set of minimal separators with size at most $k$ by ${\Delta}_k(G)$. The set of vertex sets of connected components of $G$ is denoted by ${\mathcal{C}}(G)$.
The Algorithm
=============
Branching
---------
SMS is based on the following characterization of treedepth.
\[pro:td\_rec\] Let $G$ be a graph. If $G$ is a clique then ${\texttt{\textbf{td}}}(G) = |V(G)|$. Otherwise $${\texttt{\textbf{td}}}(G) = \min_{S \in {\Delta}(G)} \left( |S| + \max_{C \in {\mathcal{C}}(G[V(G) \setminus S])} {\texttt{\textbf{td}}}(G[C]) \right).$$
Proposition \[pro:td\_rec\] is implemented as a recursive algorithm that takes a vertex set $X$ as input and computes ${\texttt{\textbf{td}}}(G[X])$ by first enumerating the minimal separators of $G[X]$ and then branching from each minimal separator $S$ to smaller induced subgraphs $G[C]$ for each component $C \in {\mathcal{C}}(G[X \setminus S])$. We make use of upper bounds by implementing Proposition \[pro:td\_rec\] as a decision procedure which, given a vertex set $X$ and a number $k$, decides if ${\texttt{\textbf{td}}}(G[X]) \le k$. Clearly, in this case we may consider only the minimal separators in ${\Delta}_{k-1}(G[X])$. Moreover, we handle the minimal separators with sizes $k-1$ and $k-2$ as special cases and thus consider only the minimal separators in ${\Delta}_{k-3}(G[X])$ in the main recursion. A minimal separator $S$ with $|S| = k-1$ such that ${\texttt{\textbf{td}}}(G[X \setminus S]) = 1$ must be a vertex cover of $G[X]$ and therefore is a neighborhood of a vertex. A minimal separator $S$ with $|S| = k-2$ such that ${\texttt{\textbf{td}}}(G[X \setminus S]) \le 2$ has also a somewhat special structure, and we handle them with a modification of Berry’s algorithm [@DBLP:journals/ijfcs/BerryBC00] for enumerating minimal separators.
Enumerating Small Minimal Separators
------------------------------------
SMS spends most of its runtime in a subroutine which given a number $k$ and a graph $G$ enumerates ${\Delta}_k(G)$. To make use of the fact that heuristic enumeration of small minimal separators is more efficient than exact enumeration, two variants of the main branching algorithm are ran: first a variant using a heuristic minimal separator enumeration algorithm and then a variant using an exact minimal separator enumeration algorithm.
The heuristic enumeration algorithm is a simple modification of Berry’s algorithm [@DBLP:journals/ijfcs/BerryBC00]. The modification prunes all minimal separators with more than $k$ vertices immediately during the execution, outputting a set ${\Delta}'_k \subseteq {\Delta}_k(G)$ in $O(|{\Delta}'_k| n^3)$ time. As observed in [@DBLP:conf/sea2/Tamaki19], there are cases in which ${\Delta}'_k \neq {\Delta}_k(G)$. However, in practice the algorithm seems to often find all small minimal separators on the values of $k$ that are relevant.
As an exact small minimal separator enumeration algorithm we implement the algorithm of Tamaki [@DBLP:conf/sea2/Tamaki19], including also the optimizations discussed in the paper. To the best of our knowledge there are no better bounds than $n^{k+O(1)}$ for the runtime of this algorithm. In practice it appears to usually have only a factor of 2-10 runtime overhead compared to the heuristic algorithm.
In cases when $G[C]$ is a child of $G$ in the recursion, obtained by branching on a minimal separator $N(C) \in {\Delta}(G)$, and $|C| > |V(G)|/2$ we make use of the small minimal separators of $G$ to enumerate the small minimal separators of $G[C]$. In particular, for all minimal separators $S \in {\Delta}_k(G[C])$, there exists a minimal separator $S' \in {\Delta}_{k+|N(C)|}(G)$ such that $S = C \cap S'$. Note that in this case $|N(C)|$ is exactly the difference in the values of $k$ in recursive calls on $G[C]$ and $G$, and therefore ${\Delta}_{k+|N(C)|}(G)$ is already enumerated.
Lower Bounds
------------
To avoid unnecessary re-computation, the known upper and lower bounds for ${\texttt{\textbf{td}}}(G[X])$ are stored for each handled induced subgraph $G[X]$. To this end, an open addressing hashtable with linear probing is implemented. Also, we implement an ad-hoc data structure so that given a vertex set $X$, a vertex set $X' \subset X$ with the highest known lower bound for ${\texttt{\textbf{td}}}(G[X'])$ can be found. This data structure uses the idea of computing subset-preserving hashes by using the intersection $X \cap V'$, where $V'$ is a subset of vertices with size $O(\log n)$, where $n$ is the number of elements in the data structure. Other implemented algorithms for computing lower bounds on ${\texttt{\textbf{td}}}(G[X])$ are the MMD+ algorithm [@DBLP:journals/iandc/BodlaenderK11] which finds large clique minors, a depth-first search algorithm which finds long paths and cycles, and a graph isomorphism hashtable which finds already processed induced subgraphs $G[X']$ that are isomorphic to $G[X]$ and applies the lower bounds of $G[X']$ to $G[X]$.
Preprocessing Techniques
------------------------
The preprocessing techniques implemented in SMS are *tree elimination* and the kernelization procedures described in [@DBLP:conf/iwpec/KobayashiT16]. Tree elimination finds a subgraph $G[T]$ such that $G[T]$ is a tree and $|N(V(G) \setminus T)| = 1$, i.e., the subgraph is attached to the rest of the graph only on a single vertex. Then it uses an exact algorithm to compute a list of length ${\texttt{\textbf{td}}}(G[T])$ that characterizes the behavior of $G[T]$ with respect to treedepth of $G$ [@DBLP:journals/ipl/Schaffer89], and replaces $G[T]$ with a construction of $O({\texttt{\textbf{td}}}(G[T])^2)$ vertices whose behavior is the same. The simplicial vertex kernelization rule from [@DBLP:conf/iwpec/KobayashiT16] is implemented as it is described there, but the shared neighborhood rule is generalized. In particular, if there are two non-adjacent vertices $u,v \in V(G)$, and the minimum $u,v$-vertex cut is at least $k$, where $k$ is an upper bound for treedepth, then an edge can be added between $u$ and $v$.
Upper Bounds
------------
To compute upper bounds on treedepth we implement a novel heuristic algorithm. The algorithm first finds a triangulation (chordal completion) $H$ of $G$ using the LB-Triang algorithm [@DBLP:journals/jal/BerryBHSV06] with a heuristic aiming to minimize the number of fill-edges in each step. Then it uses the branching algorithm, with some additional heuristics making it non-exact, to compute a treedepth decomposition of $H$. Any treedepth decomposition of $H$ is also a treedepth decomposition of $G$. The properties of chordal graphs interplay nicely with the branching algorithm: chordal graphs have a linear number of minimal separators and the treewidth of a chordal graph can be computed in linear time [@DBLP:conf/wg/GalinierHP95]. Moreover, there exists a triangulation $H$ of $G$ with ${\texttt{\textbf{td}}}(H) = {\texttt{\textbf{td}}}(G)$, because treedepth can be formulated as a completion problem to a graph class that is a subset of chordal graphs [@DBLP:journals/dam/DeogunKKM99].
Details and Proofs
==================
For a vertex set $X \subseteq V(G)$ let $G \setminus X = G[V(G) \setminus X]$. For a vertex $v \in V(G)$ let $N[v] = N(v) \cup \{v\}$.
A minimal separator $S \in {\Delta}(G)$ has at least two full components, i.e., components $C \in {\mathcal{C}}(G \setminus S)$ with $S = N(C)$.
Minimal Separators with $k-1$ Vertices
--------------------------------------
In the branching algorithm, minimal separators with $k-1$ vertices are handled as special cases.
\[pro:sc1\] Let $S$ be a minimal separator of a graph $G$ such that ${\texttt{\textbf{td}}}(G \setminus S) \le 1$. It holds that $S = N(v)$ for a vertex $v \in V(G)$.
The set $V(G) \setminus S$ is not empty, so ${\texttt{\textbf{td}}}(G \setminus S) = 1$. A graph with treedepth $1$ is an independent set, so the full components of $S$ are single vertices.
By Proposition \[pro:sc1\], it can be decided in polynomial time if there is a minimal separator of size $k-1$ that can be used to obtain a treedepth decomposition of depth $k$.
Minimal Separators with $k-2$ Vertices
--------------------------------------
Handling minimal separators with size $k-2$ is more complicated. A minimal separator $S$ with $|S| = k-2$ can be used to obtain a treedepth decomposition of depth $k$ if and only if each connected component of $G \setminus S$ is a star. To find such minimal separators, we find for each vertex $a \in V(G)$ the minimal separators $S$ such that $a$ is in a full component $C_a$ of $S$ and $C_a$ is a star with $C_a \subseteq N[a]$. The following proposition can be adapted from [@DBLP:journals/ijfcs/BerryBC00].
\[pro:berry\_gen\] Let $S$ be a minimal separator of a graph $G$ and $C_a$ a full component of $S$ containing $a$. The separator $S$ is either a separator close to $a$, obtained as $S = N(C)$ for a component $C \in {\mathcal{C}}(G \setminus N[a])$, or can generated from a separator $S'$ with a full component $C'_a \subset C_a$ with $a \in C'_a$ by $S = N(C)$ where $C \in {\mathcal{C}}(G \setminus (S' \cup N(v)))$ for a vertex $v \in S'$.
By Proposition \[pro:berry\_gen\], the minimal separators such that $C_a$ is a star can be generated from other minimal separators such that $C_a$ is a star. Thus we can modify Berry’s algorithm to prune the minimal separators for which $C_a$ is not a star. We also use an optimization that if $G \setminus (N[a] \cup N(N[a]))$ has a component that is not a star, then there is no minimal separator $S$ such that $a$ is in a full component $C_a \subseteq N[a]$ of $S$ and all components of $G \setminus S$ are stars.
Inducing Small Minimal Separators
---------------------------------
We use small minimal separators of $G$ to generate small minimal separators of $G[X]$.
Let $G$ be a graph, $X \subseteq V(G)$ and $S \in {\Delta}_k(G[X])$. There exists a minimal separator $S' \in {\Delta}_{k+|N(X)|}(G)$ such that $S = S' \cap X$.
Let $S$ be a minimal $a,b$-separator in $G[X]$. Let $Y = V(G) \setminus N(X)$. Now $S$ is a minimal $a,b$-separator in $G[Y]$. Let us add each vertex $v \in (V(G) \setminus Y)$ to $G[Y]$ one by one. If $v$ can be reached from both $a$ and $b$ in $G[Y \cup \{v\} \setminus S]$, then $S \cup \{v\}$ is a minimal $a,b$-separator of $G[Y \cup \{v\}]$, having the same full components in $G[Y \cup \{v\}]$ as $S$ has in $G[Y]$. Otherwise $S$ is a minimal separator of $G[Y \cup \{v\}]$. With this process we obtain a minimal separator $S' \in {\Delta}(G)$ with $S \subseteq S' \subseteq S \cup N(X)$.
Tree Elimination
----------------
We adapt the algorithm for computing the treedepth of a tree in linear time [@DBLP:journals/ipl/Schaffer89] to locally kernelize tree subgraphs of the input graph. Treedepth can be formulated as a problem of finding a vertex ranking $c : V(G) \rightarrow \{1, \ldots, {\texttt{\textbf{td}}}(G)\}$ such that if $c(v) = c(u)$ for distinct vertices $u,v \in V(G)$, then all paths between $u$ and $v$ contain a vertex $w$ with $c(w) > c(v)$. Now, for a vertex $v$ we can define the set of ranks $c^+(v)$ that can be “seen” from $v$ as the ranks $c(u)$ (including $u = v$) such that there is a path from $v$ to $u$ such that all vertices $w$ of the path have $c(w) \le c(u)$. Let us say that a set of ranks $R$ is smaller than a set of ranks $R'$ if the maximum element of $R \setminus R'$ is smaller than the maximum element of $R' \setminus R$.
Let $T$ be a tree and $v \in V(T)$. A vertex ranking $c$ of $T$ such that $c^+(v)$ is the smallest possible can be computed in linear time.
Our local kernelization algorithm finds a subset-maximal subgraph $G[T]$ such that $G[T]$ is a tree and is attached to the rest of the graph only on a single vertex $v \in V(T)$, i.e., $N(V(G) \setminus T) = \{v\}$ and then computes a vertex ranking of $G[T]$ such that $c^+(v)$ is the smallest possible. Then the subgraph $G[T]$ is replaced with a subgraph determined by $c^+(v)$. This is done by first replacing $v$ with a path $v = w_1, w_2, \ldots, w_{|c^+(v)|}$ of length $|c^+(v)|$, attached to the rest of the graph only on $v$. Then, a clique with $c^+(v)(i)-1$ vertices is attached to each vertex $w_i$ of the path, where $c^+(v)(i)$ is the $i$-th smallest rank in $c^+(v)$. Now, $G[T]$ has been replaced with a construction consisting of $|c^+(v)|$ vertex-disjoint cliques whose sizes are the ranks in $c^+(v)$. The size of this construction is $\sum_{r \in c^+(v)} r = O({\texttt{\textbf{td}}}(G[T])^2)$ vertices.
Let $G$ be a graph and $G'$ be a graph obtained from $G$ by applying the above described procedure. It holds that ${\texttt{\textbf{td}}}(G) = {\texttt{\textbf{td}}}(G')$.
Assume that the procedure was applied on a single tree $G[T]$ with $N(V(G) \setminus T) = \{v\}$. Let $c_o^+(v)$ be the smallest possible set of ranks seen from $v$ in $G[T]$.
First, consider a vertex ranking $c$ of $G$ and let us construct a vertex ranking $c'$ of $G'$. Let $c_t^+(v)$ be the set of ranks on $c$ seen from $v$ in $G[T]$. If the maximum value of $c_t^+(v)$ is the same as the maximum value of $c_o^+(v)$, then let $c'(w_{|c_o^+(v)|})$ be this value. The clique attached to $w_{|c_o^+(v)|}$ can now be arbitrarily ranked with smaller values. Now we can remove the maximum value from both $c_t^+(v)$ and $c_o^+(v)$ and remove the vertex $w_{|c_o^+(v)|}$ and the attached clique and continue by induction. If the maximum value of $c_t^+(v)$ is different than the maximum value of $c_o^+(v)$, then it must be higher than the maximum value of $c_o^+(v)$ because $c_o^+(v)$ is the smallest. In that case, we can rank $v$ with the maximum value, and now clearly the rest of the structure of cliques can be ranked with smaller values.
Then, consider a vertex ranking $c'$ of $G'$ and let us construct a vertex ranking $c$ of $G$. Let $c_t^{'+}(v)$ be the set of ranks on $c'$ seen from $v$ in $G'[T']$, where $T'$ is the set of vertices which replaced $T$. Our strategy is to show that (1) $c_t^{'+}(v)$ is not smaller than $c_o^+(v)$ and (2) given any set of ranks $R$ not smaller than $c_o^+(v)$, we can construct a ranking $c$ of $G[T]$ such that $c^+(v) \subseteq R$. In particular, we can construct a ranking for the induced subgraph $G[T]$ which does not add any ranks to the set of seen ranks from $v$, and thus this ranking can be plugged into the rest of the ranking for $G$. To show (1), if the maximum value of $c_t^{'+}(v)$ is higher the maximum value of $c_o^+(v)$ we are done. Otherwise, the maximum value must be taken by the clique attached to $w_{|c_o^+(v)|}$. Let us then remove the maximum value from both $c_t^{'+}(v)$ and $c_o^+(v)$ and remove $w_{|c_o^+(v)|}$ and the attached clique and proceed by induction. To show (2), let $R$ be a set of ranks not smaller than $c_o^+(v)$ and let us construct a ranking $c$ of $G[T]$ such that $c^+(v) \subseteq R$. If the maximum value of $R$ is higher than the maximum value of $c_o^+(v)$ then let $c(v)$ be the maximum value, now $c^+(v) = \{c(v)\}$ and the rest of the tree can be ranked with the optimal ranking. Otherwise find the subtree whose root in the optimal ranking of $G[T]$ has the maximum value of $c_o^+(v)$, rank the subtree with the optimal ranking, remove the subtree, and remove the maximum value from both $R$ and $c_o^+(v)$ and proceed by induction.
The local kernelization and the reconstruction can be implemented in $O((n + m) \log n)$ time.
Generalized Shared Neighborhood Rule
------------------------------------
In [@DBLP:conf/iwpec/KobayashiT16] it was shown that if there are non-adjacent vertices $u,v \in V(G)$ with $|N(v) \cap N(u)| \ge k$, where $k$ is an upper bound for treedepth, then an edge can be added between $u$ and $v$. We generalize this rule.
Let $u$ and $v$ be non-adjacent vertices in a graph $G$. If the minimum vertex cut between $u$ and $v$ is at least $k$ and ${\texttt{\textbf{td}}}(G) \le k$, then ${\texttt{\textbf{td}}}(G) = {\texttt{\textbf{td}}}(G')$, where $G'$ is a graph obtained from $G$ by adding an edge between $u$ and $v$.
Consider a treedepth decomposition of $G$ with depth at most $k$ such that $u$ is not an ancestor of $v$ and $v$ is not an ancestor of $u$. Now, the ancestors of $v$ form a vertex cut between $u$ and $v$ with size at most $k-1$. Therefore, in any treedepth decomposition of $G$ with depth at most $k$ either $v$ is an ancestor of $u$ or $u$ is an ancestor of $v$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Invariant Subspace Problem (“ISP”) for Hilbert space operators is known to be equivalent to a question that, on its surface, seems surprisingly concrete: [*For composition operators induced on the Hardy space ${H^2}$ by hyperbolic automorphisms of the unit disc, is every nontrivial minimal invariant subspace one dimensional (i.e., spanned by an eigenvector)?*]{} In the hope of reviving interest in the contribution this remarkable result might offer to the studies of both composition operators and the ISP, I revisit some known results, weaken their hypotheses and simplify their proofs. [*Sample results:*]{} If ${\varphi}$ is a hyperbolic disc automorphism with fixed points at $\alpha$ and $\beta$ (both necessarily on the unit circle), and ${C_{\varphi}}$ the composition operator it induces on ${H^2}$, then for every $f\in\sqrt{(z-\alpha)(z-\beta)}\,{H^2}$, the doubly ${C_{\varphi}}$-cyclic subspace generated by $f$ contains many independent eigenvectors; more precisely, the point spectrum of ${C_{\varphi}}$’s restriction to that subspace intersects the unit circle in a set of positive measure. Moreover, this restriction of ${C_{\varphi}}$ is hypercyclic (some forward orbit is dense). Under the stronger restriction $f\in\sqrt{(z-\alpha)(z-\beta)}\,H^p$ for some $p>2$, the point spectrum of the restricted operator contains an open annulus centered at the origin.'
address: 'Department of Mathematics and Statistics, Portland State University, Portland, OR 97207 USA'
author:
- 'Joel H. Shapiro'
title: |
Eigenfunctions for hyperbolic\
composition operators—redux
---
Introduction
============
More than twenty years ago Nordgren, Rosenthal, and Wintrobe [@NRW] made a surprising connection between composition operators on the Hardy space ${H^2}$ and the Invariant Subspace Problem—henceforth, the “ISP”. The ISP asks if every operator on a separable Hilbert space has a nontrivial invariant subspace (following tradition: “operator” means “bounded linear operator,” “subspace” means “closed linear manifold,” and for a subspace, “nontrivial” means “neither the whole space nor the zero-subspace”). Nordgren, Rosenthal, and Wintrobe proved the following [@NRW Corollary 6.3, page 343]:
> [*Suppose ${\varphi}$ is a hyperbolic automorphism of the open unit disc ${\mathbb{U}}$. Let ${C_{\varphi}}$ denote the composition operator induced by ${\varphi}$ on the Hardy space ${H^2}$. Then the ISP has a positive solution if and only if every nontrivial minimal ${C_{\varphi}}$-invariant subspace of ${H^2}$ has dimension one.*]{}
It is easy to see that, for each nontrivial minimal invariant subspace $V$ of a Hilbert space operator $T$, every non-zero vector $x\in V$ is cyclic, i.e., ${{\rm span}\,\{T^nx: n = 0, 1, 2, \ldots\}}$ is dense in $V$. If, in addition, $T$ is invertible, then so is its restriction to $V$ (otherwise the range of this restriction would be a nontrivial invariant subspace strictly contained in $V$, contradicting minimality). Thus for $T$ invertible, $V$ a nontrivial minimal invariant subspace of $T$, and $0\neq x \in V$, $$V = {{\overline{\rm span}\,}{\{T^n x: n = 0, 1, 2, \ldots\}}} = {{\overline{\rm span}\,}{\{T^n x: n\in{\mathbb{Z}}\}}},$$ where now “$\overline{\rm span}$ ” means “closure of the linear span.”
The result of Nordgren, Rosenthal, and Wintrobe therefore suggests that for ${\varphi}$ a hyperbolic disc automorphism we might profitably study how the properties of a function $f$ in ${H^2}\backslash\{0\}$ influence the operator-theoretic properties of ${{C_{\varphi}}|_{D_f}}$, the restriction of ${C_{\varphi}}$ to the “doubly cyclic” subspace subspace $$\label{df_defn}
D_f := {{\overline{\rm span}\,}{\{{C_{\varphi}}^n f: n\in{\mathbb{Z}}\}}} = {{\overline{\rm span}\,}{\{f\circ{\varphi}_n: n \in {\mathbb{Z}}\}}},$$ with particular emphasis on the question of when the point spectrum of the restricted operator is nonempty. (Here, for $n$ is a positive integer, ${\varphi}_n$ denotes the $n$-th compositional iterate of ${\varphi}$, while ${\varphi}_{-n}$ is the $n$-th iterate of ${\varphi}{^{-1}}$; ${\varphi}_0$ is the identity map.)
Along these lines, Valentin Matache [@Mat 1993] obtained a number of interesting results on minimal invariant subspaces for hyperbolic-automorphically induced composition operators. He observed, for example, that if a minimal invariant subspace for such an operator were to have dimension larger than $1$, then, at either of the fixed points of ${\varphi}$, none of the non-zero elements of that subspace could be both continuous and non-vanishing (since ${\varphi}$ is a hyperbolic automorphism of the unit disc, its fixed points must necessarily lie on the unit circle; see §\[hypautos\] below). Matache also obtained interesting results on the possibility of minimality for invariant subspaces generated by inner functions.
Several years later Vitaly Chkliar [@Chk 1996] proved this result for hyperbolic-automorphic composition operators ${C_{\varphi}}$:
> [*If $f\in{H^2}\backslash\{0\}$ is bounded in a neighborhood of one fixed point of ${\varphi}$, and at the other fixed point vanishes to some order ${\varepsilon}>0$, then the point spectrum of ${{C_{\varphi}}|_{D_f}}$ contains an open annulus centered at the origin.*]{}
Later Matache [@Mat2] obtained similar conclusions under less restrictive hypotheses.
In the work below, after providing some background (in §\[background\]), I revisit in §\[main\_results\] and §\[complements\] the work of Chkliar and Matache, providing simpler proofs of stronger results. Here is a sample: [*for ${\varphi}$ a hyperbolic automorphism of ${\mathbb{U}}$ with fixed points $\alpha$ and $\beta$ (necessarily on ${\partial {\mathbb{U}}}$):* ]{}
- [*If $f\in\sqrt{(z-\alpha)(z-\beta)}\,{H^2}\backslash\{0\}$, then $\sigma_p({{C_{\varphi}}|_{D_f}})$ intersects the unit circle in a set of positive measure.*]{}
- [*If $f\in\sqrt{(z-\alpha)(z-\beta)}\,H^p\backslash\{0\}$ for some $p>2$, then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains an open annulus centered at the origin.*]{}
Note that the function $\sqrt{(z-\alpha)(z-\beta)}$ is an outer function, so the set of functions $f$ being singled out in both parts (a) and (b) is dense in ${H^2}$.
Finally, observe that, in the hypotheses of both (a) and (b), the exponent “$\frac{1}{2}$” is best possible in the sense that for any smaller exponent the function $f\equiv 1$, for which $D_f$ is the one dimensional subspace of constant functions, would satisfy both hypotheses. This comment applies throughout the sequel.
Acknowledgement {#acknowledgement .unnumbered}
---------------
I wish to thank Professor Paul Bourdon of Washington and Lee University for many suggestions that corrected and improved a preliminary version of this paper.
Background material {#background}
===================
Disc automorphisms {#hypautos}
------------------
An [*automorphism*]{} of a domain in the complex plane is a univalent holomorphic mapping of that domain onto itself. Every automorphism of the open unit disc ${\mathbb{U}}$ is a linear fractional map [@Rud Theorem 12.6, page 255].
Linear fractional maps can be regarded as homeomorphisms of the Riemann Sphere; as such, each one that is not the identity map has one or two fixed points. The maps with just one fixed point are the [*parabolic*]{} ones; each such map is conjugate, via an appropriate linear fractional map, to one that fixes only the point at infinity, i.e., to a translation. A linear fractional map that fixes two distinct points is conjugate, again via a linear fractional map, to one that fixes both the origin and the point at infinity, i.e., to a dilation $w{\rightarrow}\mu w$ of the complex plane, where $\mu\neq 1$ is a complex number called the [*multiplier*]{} of the original map (actually $1/\mu$ can just as well occur as the multiplier—depending on which fixed point of the original map is taken to infinity by the conjugating transformation). The original map is called [*elliptic*]{} if $|\mu|=1$, hyperbolic if $\mu$ is positive, and [*loxodromic*]{} in all other cases (see, for example, [@S Chapter 0] for more details).
Suppose ${\varphi}$ is a hyperbolic automorphism of ${\mathbb{U}}$. Then the same is true of its inverse. The fixed points of ${\varphi}$ must necessarily lie on ${\partial {\mathbb{U}}}$, the unit circle. Indeed, if the attractive fixed point of ${\varphi}$ lies outside the closed unit disc, then the compositional iterates of ${\varphi}$ pull ${\mathbb{U}}$ toward that fixed point, and hence outside of ${\mathbb{U}}$, which contradicts the fact that ${\varphi}({\mathbb{U}})={\mathbb{U}}$. If, on the other hand, the attractive fixed point lies in ${\mathbb{U}}$, then its reflection in the unit circle is the repulsive fixed point, which is the attractive one for ${\varphi}^{-1}$. Thus ${\varphi}^{-1}$ can’t map ${\mathbb{U}}$ into itself, another contradiction. Conclusion: both fixed points lie on ${\partial {\mathbb{U}}}$.
Let’s call a hyperbolic automorphism ${\varphi}$ of ${\mathbb{U}}$ [*canonical*]{} if it fixes the points $\pm 1$, with $+1$ being the attractive fixed point. We’ll find it convenient to move between the open unit disc ${\mathbb{U}}$ and the open right half-plane ${\rm \Pi^+}$ by means of the [*Cayley transform*]{} $\kappa:\Pi^+ {\rightarrow}{\mathbb{U}}$ and its inverse $\kappa{^{-1}}:{\mathbb{U}}{\rightarrow}\Pi^+$, where $$\kappa(w) = \frac{w-1}{w+1}
\quad {\rm and} \quad
\kappa{^{-1}}(z) = \frac{1+z}{1-z}
\qquad (z\in{\mathbb{U}}, w\in\Pi^+).$$ In particular, if ${\varphi}$ is a canonical hyperbolic automorphism of ${\mathbb{U}}$, then $\Phi:=\kappa{^{-1}}\circ{\varphi}\circ\kappa$ is an automorphism of ${\rm \Pi^+}$ that fixes $0$ and $\infty$, with $\infty$ being the attractive fixed point. Thus $\Phi(w) = \mu w$ for some $\mu >1$, and ${\varphi}= \kappa\circ\Phi\circ\kappa{^{-1}}$, which yields, after a little calculation, $$\label{canonical_auto}
{\varphi}(z) = \frac{r+z}{1+rz} \qquad {\rm where}
\qquad {\varphi}(0)= r = \frac{\mu-1}{\mu+1} \in (0,1).$$ If ${\varphi}$ is a hyperbolic automorphism of ${\mathbb{U}}$ that is not canonical, then it can be conjugated, via an appropriate automorphism of ${\mathbb{U}}$, to one that is. This is perhaps best seen by transferring attention to the right half-plane ${\rm \Pi^+}$, and observing that if $\alpha <\beta$ are two real numbers, then the linear fractional map $\Psi$ of ${\rm \Pi^+}$ defined by $$\Psi(w) ~=~ i \, \frac{w-i\beta}{w-i\alpha}$$ preserves the imaginary axis, and takes the point 1 into ${\rm \Pi^+}$. Thus it is an automorphism of ${\rm \Pi^+}$ that takes the boundary points $i\beta$ to zero and $i\alpha$ to infinity. Consequently if $\Phi$ is any hyperbolic automorphism of ${\rm \Pi^+}$ with fixed points $i\alpha$ (attractive) and $i\beta$ (repulsive), then $\Psi\circ\Phi\circ\Psi^{-1}$ is also hyperbolic automorphism with attractive fixed point $\infty$ and repulsive fixed point $0$. If, instead, $\alpha > \beta$ then $-\Psi$ does the job.
Since any hyperbolic automorphism ${\varphi}$ of ${\mathbb{U}}$ is conjugate, via an automorphism, to a canonical one, ${C_{\varphi}}$ is similar, via the composition operator induced by the conjugating map, to a composition operator induced by a canonical hyperbolic automorphism. For this reason the work that follows will focus on the canonical case.
Spectra of hyperbolic-automorphic composition operators {#spectra}
-------------------------------------------------------
Suppose ${\varphi}$ is a hyperbolic automorphism of ${\mathbb{U}}$ with multiplier $\mu>1$. Then it is easy to find lots of eigenfunctions/eigenvalues for ${C_{\varphi}}$ on ${H^2}$. We may without loss of generality assume that ${\varphi}$ is canonical, and then move, via the Cayley map, to the right half-plane where ${\varphi}$ morphs into the dilation $\Phi(w) = \mu w$. Let’s start by viewing the composition operator $C_\Phi$ as just a linear map on ${\rm Hol\,({\rm \Pi^+})}$, the space of all holomorphic functions on ${\rm \Pi^+}$. For any complex number $a$ define $E_a(w) = w^a$, where $w^a=\exp(a\log w)$, and “log” denotes the principal branch of the logarithm. Then $E_a\in{\rm Hol\,({\rm \Pi^+})})$ and $C_\Phi(E_a) = \mu^a E_a$, i.e., $E_a$ is an eigenvector of $C_\Phi$ (acting on ${\rm Hol\,({\rm \Pi^+})}$) and the corresponding eigenvalue is $\mu^a$ (again taking the principal value of the “$a$-th power”). Upon returning via the Cayley map to the unit disc, we see that, when viewed as a linear transformation of ${\rm Hol\,({\mathbb{U}})}$, the operator ${C_{\varphi}}$ has, for each $a\in{\mathbb{C}}$, the eigenvector/eigenvalue combination $(f_a, \mu^a)$, where the function $$\label{eigenfunction_defn}
f_a(z) = \left(\frac{1+z}{1-z}\right)^a \quad\qquad (z\in{\mathbb{U}})$$ belongs to ${H^2}$ if and only if $|{{\rm Re\,}}(a)| < 1/2$. Thus the corresponding ${H^2}$-eigenvalues $\mu^a$ cover the entire open annulus $$\label{spectral_annulus}
A:= \{\lambda\in{\mathbb{C}}: \frac{1}{\sqrt{\mu}}<|\lambda|<\sqrt{\mu}\}.$$ In particular $\sigma({C_{\varphi}})$, the ${H^2}$-spectrum of ${C_{\varphi}}$, contains this annulus, and since the map $a{\rightarrow}\mu^a$ takes the strip $|{{\rm Re\,}}(a)| < 1/2$ infinitely-to-one onto $A$, each point of $A$ is an eigenvalue of ${C_{\varphi}}$ having infinite multiplicity.
As for the rest of the spectrum, an elementary norm calculation shows that $\sigma({C_{\varphi}})$ is just the closure of $A_\mu$. To see this, note first that the change-of-variable formula from calculus shows that for each $f\in{H^2}$ and each automorphism ${\varphi}$ of ${\mathbb{U}}$ (not necessarily hyperbolic): $$\label{ch_var}
{\|{C_{\varphi}}f\|}^2 = {\int_{{\partial {\mathbb{U}}}}}|f|^2 P_a \, dm$$ where $m$ is normalized arc-length measure on the unit circle ${\partial {\mathbb{U}}}$, and $P_a$ is the Poisson kernel for $a={\varphi}(0)$; more generally, for any $a\in{\mathbb{U}}$: $$\label{poisson_kernel_formula}
P_a(\zeta) ~=~ \frac{1-|a|^2}{|\zeta - a|^2} \qquad (\zeta\in{\partial {\mathbb{U}}}) \,$$ (see also Nordgren’s neat argument [@Nor Lemma 1, page 442], which shows via Fourier analysis that holds for any inner function).
Now suppose ${\varphi}$ is the canonical hyperbolic automorphism of ${\mathbb{U}}$ with multiplier $\mu>1$. Then ${\varphi}$ is given by , so by $$P_r(\zeta) ~=~ \frac{1-r^2}{|\zeta-r|^2} ~\le~ \frac{1+r}{1-r} ~=~ \mu$$ which, along with shows that $$\label{norm_ineq}
{\|{C_{\varphi}}\|} ~\le~ \sqrt{\mu} \, .$$ Since also $$P_r(\zeta) ~\ge~ \frac{1-r}{1+r} ~=~ \mu^{-1}$$ we have, for each $f\in{H^2}$ $${\|{C_{\varphi}}f\|} ~\ge~ \frac{1}{\sqrt{\mu}} \, {\|f\|},$$ which shows that holds with ${C_{\varphi}}$ replaced by ${C_{\varphi}}^{-1}$. Thus the spectra of both ${C_{\varphi}}$ and its inverse lie in the closed disc of radius $\sqrt{\mu}$ centered at the origin, so by the spectral mapping theorem, $\sigma({C_{\varphi}})$ is contained in the closure of the annulus . Since we have already seen that this closed annulus contains the spectrum of ${C_{\varphi}}$ we’ve established the following result, first proved by Nordgren [@Nor Theorem 6, page 448] using precisely the argument given above:
\[spectrum\_thm\] If ${\varphi}$ is a hyperbolic automorphism of ${\mathbb{U}}$ with multiplier $\mu ~(>1)$, then $\sigma({C_{\varphi}})$ is the closed annulus $
\{\lambda\in{\mathbb{C}}:1/\sqrt{\mu}\le |\lambda|\le \sqrt{\mu}\}.
$ The interior of this annulus consists entirely of eigenvalues of ${C_{\varphi}}$, each having infinite multiplicity.
In fact the interior of $\sigma({C_{\varphi}})$ is precisely the point spectrum of ${C_{\varphi}}$; see [@Mat3] for the details.
Poisson kernel estimates {#poisson_estimate}
------------------------
Formula , giving the Poisson kernel for the point $a=\rho e^{i\theta_0} \in{\mathbb{U}}$, can be rewritten $$P_a({e^{i\theta}}) ~=~ \frac{1-\rho^2}{1-2\rho\cos(\theta-\theta_0) + \rho^2}
\qquad (0\le\rho<1, \theta\in{\mathbb{R}}) \, .$$ We will need the following well-known estimate, which provides a convenient replacement (cf. for example [@A page 313]).
\[poisson\_est\_lem\] For $0\le\rho<1$ and $|\theta|\le\pi$: $$\label{poisson_int_est}
P_\rho({e^{i\theta}}) ~\le~ 4 \, \frac{(1-\rho)}{(1-\rho)^2+(\theta/\pi)^2}$$
$$P_\rho({e^{i\theta}})
:= \frac{1-\rho^2}{1-2\rho\cos\theta+\rho^2}
= \frac{1-\rho^2}{(1-\rho)^2+\rho(2\sin\frac{\theta}{2})^2}
\le \frac{2(1-\rho)}{(1-\rho)^2+4\rho(\theta/\pi)^2}$$ so, at least when $\rho\ge \frac{1}{4}$, inequality holds with constant “2” in place of “4”. For the other values of $\rho$ one can get inequality by checking that, over the interval $[0,\pi]$, the minimum of the right-hand side exceeds the maximum of the left-hand side.
\[constant\_remark\][*The only property of the constant “4” on the right-hand side of that matters for our purposes is its independence of $\rho$ and $\theta$.* ]{}
For the sequel (especially Theorem \[main\_thm\_2\] below) we will require the following upper estimate of certain infinite sums of Poisson kernels.
\[sum\_estimate\] For ${\varphi}$ the canonical hyperbolic automorphism of ${\mathbb{U}}$ with multiplier $\mu$: $$\label{poisson_sum_estimate}
{\sum_{n=0}^\infty}P_{{\varphi}_n(0)}({e^{i\theta}})
~\le~ \frac{16\mu}{\mu-1} \, \frac{\pi}{|\theta|}
\qquad (|\theta|\le \pi) \,.$$
In the spirit of Remark \[constant\_remark\] above, the precise form of the positive constant that multiplies $\pi/|\theta|$ on the right-hand side of is unimportant (as long as it does not depend on $\theta$).
The automorphism ${\varphi}$ is given by equations . For each integer $n\ge 0$ the $n$-th iterate ${\varphi}_n$ of ${\varphi}$ is just the canonical hyperbolic automorphism with multiplier $\mu^n$, so upon substituting $\mu^n$ for $\mu$ in we obtain $$\label{nth_iterate}
{\varphi}_n(z) = \frac{r_n+z}{1+r_nz} \qquad{\rm where} \qquad
{\varphi}_n(0) = r_n=\frac{\mu^n-1}{\mu^n+1} \in (0,1).$$ Thus $1-r_n = 2/(\mu^n+1)$, and so $$\label{mu_power_est}
\mu^{-n} ~<~ 1-r_n ~<~ 2 \, \mu^{-n} \quad (n=0, 1, 2, ...),$$ (in particular, $r_n$ approaches the attractive fixed point $+1$ with exponential speed as $n{\rightarrow}\infty$; this is true of the ${\varphi}$-orbit of any point of the unit disc).
Fix $\theta\in[-\pi,\pi]$. We know from and that for each integer $n\ge 0$, $${P_{r_n}}({e^{i\theta}}) ~\le~ \frac{4(1-r_n)}{(1-r_n)^2 + (\theta/\pi)^2}
~\le~ \frac{8\mu^{-n}}{\mu^{-2n}+(\theta/\pi)^2}~,$$ whereupon, for each non-negative integer $N$: $$\begin{aligned}
\frac{1}{8} {\sum_{n=0}^\infty}{P_{r_n}}({e^{i\theta}})
&\le& {\sum_{n=0}^\infty}\frac{\mu^{-n}}{\mu^{-2n} + (\theta/\pi)^2} \\
& &\\
&\le& \sum_{n=0}^{N-1}\frac{\mu^{-n}}{\mu^{-2n}}
~+~ \left(\frac{\pi}{\theta}\right)^2
\sum_{n=N}^\infty\mu^{-n} \\
& &\\
&=& \sum_{n=0}^{N-1} \mu^n
~+~ \left(\frac{\pi}{\theta}\right)^2
\sum_{n=N}^\infty \mu^{-n} \\
& &\\
&=& \frac{\mu^N - 1}{\mu - 1}
~+~
\left(\frac{\pi}{\theta}\right)^2 \mu^{-N} (1-\mu^{-1})^{-1}
\end{aligned}$$ where the geometric sum in the next-to-last line converges because $\mu>1$.
We need a choice of $N$ that gives a favorable value for the quantity in the last line of the display above. Let $\nu = \log_\mu(\pi/|\theta|)$, so that $\mu^\nu = \pi/|\theta|$. Since $|\theta|\le\pi$ we are assured that $\nu\ge 0$. Let $N$ be the least integer $\ge \nu$, i.e., the unique integer in the interval $[\nu,\nu+1)$. The above estimate yields for any integer $N\ge 0$, upon setting $C:=8\mu/(\mu-1)$ (which is $>0$ since $\mu>1$), $$\label{first_sum_est}
{\sum_{n=0}^\infty}{P_{r_n}}({e^{i\theta}})
\le C\left[\mu ^{N-1} + \left(\frac{\pi}{\theta}\right)^2 \mu^{-N} \right]
\le C \frac{\pi}{|\theta|}
\left[\frac{|\theta|}{\pi}\mu^\nu
+ \left(\frac{|\theta|}{\pi}\mu^\nu\right){^{-1}}\right].$$ By our choice of $\nu$, both summands in the square-bracketed term at the end of have the value 1 and this implies .
Main results {#main_results}
============
Here I extend work of Chkliar [@Chk] and Matache [@Mat2] that provides, for a hyperbolic-automorphically induced composition operator ${C_{\varphi}}$, sufficient conditions on $f\in{H^2}$ for the doubly-cyclic subspace $D_f$, as defined by (\[df\_defn\]), to contain a rich supply of linearly independent eigenfunctions. I’ll focus mostly on canonical hyperbolic automorphisms, leaving the general case for the next section. Thus, until further notice, ${\varphi}$ will denote a canonical hyperbolic automorphism of ${\mathbb{U}}$ with multiplier $\mu>1$, attractive fixed point at $+1$ and repulsive one at $-1$, i.e., ${\varphi}$ will be given by equations .
Following both Chkliar and Matache, I will use an ${H^2}$-valued Laurent series to produce the desired eigenvectors. The idea is this: for $f\in{H^2}$, and $\lambda$ a non-zero complex number, if the series $$\label{eigen_series}
\sum_{n\in{\mathbb{Z}}} \lambda^{-n} (f\circ{\varphi}_n)$$ converges strongly enough (for example, in ${H^2}$) then the sum $F_\lambda$, whenever it is not the zero-function, will be a $\lambda$-eigenfunction of ${C_{\varphi}}$ that lies in $D_f$. Clearly the convergence of the series will depend crucially on the behavior of ${\varphi}$ at its fixed points, as the next result indicates. For convenience let’s agree to denote by $A(R_1,R_2)$ the open annulus, centered at the origin, of inner radius $R_1$ and outer radius $R_2$ (where, of course, $0<R_1<R_2<\infty$).
\[main\_thm\_1\] [(cf. [@Chk])]{} Suppose $0<{\varepsilon},\delta\le 1/2$, and that $$f\in (z-1)^{\frac{1}{2}+{\varepsilon}} (z+1)^{\frac{1}{2}+\delta}{H^2}\backslash\{0\}.$$ Then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, except possibly for a discrete subset, $A(\mu^{-{\varepsilon}},\mu^\delta)$.
Our hypothesis on the behavior of $f$ at the point $+1$ (the attractive fixed point of ${\varphi}$) is that $f=(z-1)^{\frac{1}{2}+{\varepsilon}}g$ for some $g\in{H^2}$, i.e., that $$\label{boundary_hypothesis}
\infty ~>~ {\int_{{\partial {\mathbb{U}}}}}|g|^2dm ~=~ {\int_{{\partial {\mathbb{U}}}}}\frac{|f(\zeta)|^2}{|\zeta-1|^{2{\varepsilon}+1}}\,dm(\zeta)
~\ge~ {\frac{1}{2\pi}}{\int_{-\pi}^\pi}\frac{|f({e^{i\theta}})|^2}{~~|\theta|^{2{\varepsilon}+1}} \, d\theta \, .$$ Upon setting $a={\varphi}_n(0) := r_n$ in we obtain $$\label{ch_var_htwo}
{\|f\circ{\varphi}_n\|}^2 ~=~ \int |f|^2 P_{r_n}\,dm \, ,
\qquad (n\in{\mathbb{Z}})$$ which combines with estimates and to show that if $n$ is a non-negative integer (thus insuring that $r_n > 0$): $$\begin{aligned}
{\|f\circ{\varphi}_n\|}^2
&\le & 2\pi {\int_{-\pi}^\pi}|f({e^{i\theta}})|^2 \frac{1-r_n}{(1-r_n)^2+\theta^2} \,d \theta \\
& &\\
&\le& 4\pi{\int_{-\pi}^\pi}|f({e^{i\theta}})|^2
\frac{\mu^{-n}}{\mu^{-2n}+\theta^2} \, d\theta \\
& &\\
&=& 4\pi \mu^{-2n{\varepsilon}} {\int_{-\pi}^\pi}\frac{|f({e^{i\theta}})|^2}{|\theta|^{1+2{\varepsilon}}} \,
\left\{\frac{(\mu^n|\theta|)^{1+2{\varepsilon}}}{1+(\mu^n|\theta|)^2}
\right\} \, d\theta\\
& &\\
&\le& 4\pi \mu^{-2n{\varepsilon}}
\sup_{x\in{\mathbb{R}}}\left\{ \frac{|x|^{1+2{\varepsilon}}}{1+x^2}\right\}
{\int_{-\pi}^\pi}\frac{|f({e^{i\theta}})|^2}{|\theta|^{1+2{\varepsilon}}} \, d\theta ~. \end{aligned}$$ By the integral in the last line is finite, and since $0<{\varepsilon}\le 1/2$, the supremum in that line is also finite. Thus $${\|f\circ{\varphi}_n\|} ~=~ {\mbox{O}}(\mu^{-n{\varepsilon}})
\quad {\rm as}\quad n{\rightarrow}\infty,$$ which guarantees that the subseries of with positively indexed terms converges in ${H^2}$ for all $\lambda\in{\mathbb{C}}$ with $|\lambda| > \mu^{-{\varepsilon}}$.
As for the negatively indexed subseries of , note from that ${\varphi}{^{-1}}(z) = -{\varphi}(-z)$, so ${\varphi}_{-n}(z) = -{\varphi}_n(-z)$ for each integer $n$. Let $g(z) = f(-z)$, so our hypothesis on $f$ implies that $g\in (z-1)^{{\frac{1}{2}}+\delta}\,{H^2}\backslash\{0\}$. Let $\psi_n(z) = {\varphi}_n(-z)$ (the subscript on $\psi$ does not now indicate iteration). Then for each positive integer $n$ we have $\psi_n(0)={\varphi}_n(0) = r_n$, hence: $${\|f\circ{\varphi}_{-n}\|}^2
~=~ {\|g\circ\psi_n\|}^2 ~=~ {\int_{{\partial {\mathbb{U}}}}}|g|^2 {P_{r_n}}\, dm$$ so by the result just obtained, with $g$ in place of $f$ and ${\varepsilon}$ replaced by $\delta$, $${\|f\circ{\varphi}_{-n}\|} ~=~ {\mbox{O}}(\mu^{-n\delta})
\quad {\rm as} \quad n{\rightarrow}\infty.$$ Thus the negatively indexed subseries of converges in ${H^2}$ for all complex numbers $\lambda$ with $|\lambda|<\mu^\delta$.
Conclusion: For each $\lambda$ in the open annulus $A(\mu^{-{\varepsilon}}, \mu^\delta)$ the ${H^2}$-valued Laurent series converges in the norm topology of ${H^2}$ to a function $F_\lambda \in {H^2}$. Now $F_\lambda$, for such a $\lambda$, will be a ${C_{\varphi}}$-eigenfunction unless it is the zero-function, and—just as for scalar Laurent series—this inconvenience can occur for at most a discrete subset of points $\lambda$ in the annulus of convergence (the relevant uniqueness theorem for ${H^2}$-valued holomorphic functions follows easily from the scalar case upon applying bounded linear functionals).
[ *Chkliar [@Chk] has a similar result, where there are uniform conditions on the function $f$ at the fixed points of ${\varphi}$ (see also Remark \[Chkliar\_remark\] below); as he suggests, it would be of interest to know whether or not the “possible discrete subset” that clutters the conclusions of results like Theorem \[main\_thm\_1\] can actually be nonempty.* ]{}
[*The limiting case $\delta=0$ of Theorem \[main\_thm\_1\] still holds (see Theorem \[main\_thm\_3\] below); it is a slight improvement on Chkliar’s result (see also the discussion following Theorem \[main\_thm\_3\]).* ]{}
[*Note that the restriction ${\varepsilon}, \delta \le 1/2$ in the hypothesis of Theorem \[main\_thm\_1\] cannot be weakened since, as mentioned at the end of §\[spectra\], the point spectrum of ${C_{\varphi}}$ is the open annulus $A(\mu^{-{\frac{1}{2}}}, \mu^{\frac{1}{2}})$.* ]{}
Here is a companion to Theorem \[main\_thm\_1\], which shows that even in the limiting case $\delta={\varepsilon}=0$ (in some sense the “weakest” hypothesis on $f$) the operator ${{C_{\varphi}}|_{D_f}}$ still has a significant supply of eigenvalues.
\[main\_thm\_2\] If $f\in\sqrt{(z+1)(z-1)} \, {H^2}$ then $\sigma_p({C_{\varphi}}|_{D_f})$ intersects ${\partial {\mathbb{U}}}$ in a set of positive measure.
We will work in the Hilbert space $L^2({H^2},dm)$ consisting of ${H^2}$-valued ($m$-equivalence classes of) measurable functions $F$ on ${\partial {\mathbb{U}}}$ with $$|||F|||^2 ~:=~ \int_{{\partial {\mathbb{U}}}} \|F(\omega)\|^2 dm(\omega) <\infty.$$ I will show in a moment that the hypothesis on $f$ implies $$\label{series_conv}
\sum_{n\in{\mathbb{Z}}} {\|f\circ{\varphi}_n\|}^2 < \infty \,.$$ Granting this, it is easy to check that the ${H^2}$-valued Fourier series $$\label{vector_fs}
\sum_{n\in{\mathbb{Z}}} (f\circ{\varphi}_n) \, \omega^{-n} \qquad (\omega\in{\partial {\mathbb{U}}})$$ converges unconditionally in $L^2({H^2},dm)$, so at least formally, we expect that for a.e. $\omega\in{\partial {\mathbb{U}}}$ we’ll have ${C_{\varphi}}(F(\omega)) = \omega F(\omega)$. This is true, but a little care is needed to prove it. The “unconditional convergence” mentioned above means this: If, for each finite subset $E$ of ${\mathbb{Z}}$, $$S_E(\omega) := \sum_{n \in E} (f\circ{\varphi}_n) \omega^{-n}
\qquad (\omega\in{\partial {\mathbb{U}}}) \, ,$$ then the net $
(S_E: E~{\rm a~finite~subset~of}~ {\mathbb{Z}})
$ converges in $L^2({H^2}, dm)$ to $F$. In particular, if for each non-negative integer $n$ we define $F_n = S_{[-n,n]}$, then $F_n{\rightarrow}F$ in $L^2({H^2},dm)$, hence some subsequence $(F_{n_k}(\omega))_{k=1}^\infty$ converges in ${H^2}$ to $F(\omega)$ for a.e. $\omega\in{\partial {\mathbb{U}}}$. Now for any $n$ and any $\omega\in{\partial {\mathbb{U}}}$: $${C_{\varphi}}F_n(\omega) = \omega F_n(\omega)
- \omega^{n+1}f\circ{\varphi}_{-n}
+\omega^{-n} f\circ{\varphi}_{n+1}$$ which implies, since guarantees that ${\|f\circ{\varphi}_n\|}{\rightarrow}0$ as $n{\rightarrow}\infty$, that $${C_{\varphi}}F_n(\omega) - \omega F_n(\omega)
{\rightarrow}0~~{\rm in}~~{H^2}\qquad (n{\rightarrow}\infty).$$ This, along with the a.e. convergence of the subsequence $(F_{n_k})$ to $F$, shows that ${C_{\varphi}}F(\omega) = \omega F(\omega)$ for a.e. $\omega\in{\partial {\mathbb{U}}}$. Now the ${H^2}$-valued Fourier coefficients $f\circ{\varphi}_n$ are not all zero (in fact, none of them are zero) so at least for a subset of points $\omega\in{\partial {\mathbb{U}}}$ having positive measure we have $F(\omega)\neq 0$. The corresponding ${H^2}$-functions $F(\omega)$ are therefore eigenfunctions of ${C_{\varphi}}$ that belong to $D_f$, thus $\sigma_p({{C_{\varphi}}|_{D_f}})\cap{\partial {\mathbb{U}}}$ has positive measure.
It remains to prove . As usual, we treat the positively and negatively indexed terms separately. Since $f\in\sqrt{z-1}\, {H^2}$ we have $$\frac{1}{2\pi}\int_{-\pi}^\pi \frac{|f({e^{i\theta}})|^2}{|\theta|} d\theta
~\le~ \int_{{\partial {\mathbb{U}}}} \frac{|f(\zeta)|^2}{|\zeta-1|} dm(\zeta)
~<~\infty$$ so successive application of and yields $${\sum_{n=0}^\infty}{\|f\circ{\varphi}_n\|}^2
= {\int_{{\partial {\mathbb{U}}}}}|f|^2 \left({\sum_{n=0}^\infty}{P_{r_n}}\right)dm
~\le~ {{\rm const.\,}}{\int_{-\pi}^\pi}\frac{|f({e^{i\theta}})|^2}{|\theta|}\, d\theta
~<~ \infty \, .$$
For the negatively indexed terms in , note that our hypothesis on $f$ guarantees that $$\label{second_hyp}
\frac{1}{2\pi}\int_{-\pi}^\pi \frac{|f(e^{i(\theta-\pi)})|^2}{|\theta|} d\theta
~\le~ \int_{{\partial {\mathbb{U}}}} \frac{|f(\zeta)|^2}{|\zeta+1|} dm(\zeta)
~<~ \infty \, .$$ Recall from the proof of Theorem \[main\_thm\_1\] that ${\varphi}_{-n}(z) = -{\varphi}_n(-z)\,$ for $z\in{\mathbb{U}}$ and $n>0$, and so $${\|f\circ{\varphi}_{-n}\|}^2 = \int |f|^2P_{-r_n} dm
= \frac{1}{2\pi} \int_{-\pi}^\pi |f({e^{i\theta}})|^2P_{r_n}(\theta-\pi)\,d\theta.$$ Thus $$\begin{aligned}
\sum_{n=1}^\infty{\|f\circ{\varphi}_{-n}\|}^2
&=&
\frac{1}{2\pi} \int_{-\pi}^\pi |f({e^{i\theta}})|^2
\left(\sum_{n=1}^\infty P_{r_n}(\theta-\pi)\right) d\theta \\
&=&
\frac{1}{2\pi} \int_{-\pi}^\pi |f(e^{i(\theta-\pi)})|^2
\left(\sum_{n=1}^\infty P_{r_n}(\theta)\right) d\theta \\
& &\\
& \le& {{\rm const.\,}}\, \int_{-\pi}^\pi
\frac{|f(e^{i(\theta-\pi)})|^2}{|\theta|} d\theta \\
& &\\
&<& \infty \end{aligned}$$ where the last two lines follow, respectively, from inequalities and . This completes the proof of , and with it, the proof of the Theorem.
It would be of interest to know just how large a set $\sigma_p({{C_{\varphi}}|_{D_f}})$ has to be in Theorem \[main\_thm\_2\]. Might it always be the whole unit circle? Might it be even larger? What I do know is that if the hypothesis of the Theorem is strengthened by replacing the hypothesis “$f\in \sqrt{(z+1)(z-1)}\,{H^2}$” with the stronger “$f\in \sqrt{(z+1)(z-1)}\,H^p$ for some $p>2$”, then the conclusion improves dramatically, as shown below by the result below, whose proof reprises the latter part of the proof of Theorem \[main\_thm\_1\].
\[main\_thm\_3\] [(cf. [@Mat2 Theorem 5.5])]{} If $f\in \sqrt{(z+1)(z-1)}\, H^p\backslash\{0\}$ for some $p>2$, then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, except possibly for a discrete subset, the open annulus $A(\mu^{-{\varepsilon}},\mu^{\varepsilon})$ where ${\varepsilon}= {\frac{1}{2}}-\frac{1}{p}$.
I will show that the hypothesis implies that $f\in[(z-1)(z+1)]^{{\frac{1}{2}}+\delta}{H^2}$ for each positive $\delta<{\varepsilon}$. This will guarantee, by the proof of Theorem \[main\_thm\_1\], that the series converges in the open annulus $A(\mu^{-\delta},\mu^\delta)$ for each such $\delta$, and hence it converges in $A(\mu^{-{\varepsilon}},\mu^{\varepsilon})$, which will, just as in the proof of Theorem \[main\_thm\_1\] finish the matter. The argument below, suggested by Paul Bourdon, greatly simplifies my original one. Our hypotheses of $f$ imply that for some $g\in H^p$, $$f=[(z-1)(z+1)]^{{\frac{1}{2}}+ \delta} h \quad
{\rm where} \quad h = [(z-1)(z+1)]^{-\left({\frac{1}{2}}+ \delta\right)} g.$$ To show: $h\in{H^2}$. The hypothesis on $\delta$ can be rewritten: $2p\delta/(p-2)<1$, so the function $[(z-1)(z+1)]^{-\delta}$ belongs to $H^{\frac{2p}{p-2}}$, hence an application of Hölder’s inequality shows that $h$ is in ${H^2}$ with norm bounded by the product of the $H^p$-norm of $g$ and the $H^{\frac{2p}{p-2}}$-norm of $[(z-1)(z+1)]^{-\delta}$.
In both [@Chk] and [@Mat2 Theorem 5.3] there are results where the hypotheses on $f$ involve uniform boundedness for $f$ at one or both of the fixed points of ${\varphi}$. In [@Mat2 Theorem 5.4] Matache shows that these uniform conditions can be replaced by boundedness of a certain family of Poisson integrals, and from this he derives the following result.
> [*[@Mat2 Theorem 5.5] If $f\in (z-1)^{\frac{2}{p}}\,H^p$ for some $p>2$, and $f$ is bounded in a neighborhood of $-1$, then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains an open annulus centered at the origin.* ]{}
I’ll close this section by presenting some results of this type, where uniform boundedness at one of the fixed points is replaced by boundedness of the Hardy-Littlewood maximal function. This is the function, defined for $g$ non-negative and integrable on ${\partial {\mathbb{U}}}$, and $\zeta\in{\partial {\mathbb{U}}}$, by: $$M[g](\zeta) := \sup\left\{\frac{1}{m(I)}\int_I g \, dm: I
~{\rm an~arc~of}~ {\partial {\mathbb{U}}}~{\rm centered~at}~\zeta\right\}.$$ The [*radial maximal function*]{} $R[g]$ of $g$ at $\zeta\in{\partial {\mathbb{U}}}$ is the supremum of the values of the Poisson integral of $g$ on the radius $[0,\zeta)$. It is easy to check that $M[g]$ is dominated pointwise on ${\partial {\mathbb{U}}}$ by a constant multiple of $R[g]$. What is perhaps surprising, but still elementary, is the fact that there is a similar inequality in the other direction: the radial maximal function of the non-negative integrable function $g$ is dominated pointwise on ${\partial {\mathbb{U}}}$ by a constant multiple of its Hardy-Littlewood maximal function (see [@Rud Theorem 11.20, page 242]). This and yield
\[max\_fcn\_lemma\] For $f\in{H^2}$, $$M[|f|^2](-1) < \infty
~~\implies~~
\sup\{{\|f\circ{\varphi}_n\|}: n<0\} < \infty.$$
To see that the hypotheses of Lemma \[max\_fcn\_lemma\] can be satisfied by functions in ${H^2}$ that are unbounded as $z{\rightarrow}-1$, one need only observe that $$M[|f|^2](-1) ~\le~ {{\rm const.\,}}\,\int \frac{|f(\zeta)|^2}{|1+\zeta|}\,dm(\zeta) \, ,$$ hence, along with , the Lemma implies:
\[zplusone\_cor\] If $f\in\sqrt{z+1} \, {H^2}$ then $\sup\{{\|f\circ{\varphi}_n\|}: n<0\} < \infty$.
Thus if $f\in\sqrt{z+1}\,{H^2}$, or more generally if $M[|f|^2](-1)<\infty$, the negatively indexed subseries of will converge in ${H^2}$ for all $\lambda\in{\mathbb{U}}$. We have seen in the proof of Theorem \[main\_thm\_1\] that if $f\in (z-1)^{{\frac{1}{2}}+ {\varepsilon}}\,{H^2}$ for some ${\varepsilon}\in(0,1/2]$ then the positively indexed subseries of converges for $|\lambda| > \mu^{-{\varepsilon}}$. Putting it all together we obtain the promised “$\delta=0$” case of Theorem \[main\_thm\_1\]:
\[main\_thm\_4\] Suppose $f\in (z+1)^{{\frac{1}{2}}}(z-1)^{{\frac{1}{2}}+{\varepsilon}}\,{H^2}\backslash\{0\}$ for some $0<{\varepsilon}<1/2$. Then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, with the possible exception of a discrete subset, the open annulus $A(\mu^{-{\varepsilon}}, 1)$.
\[Chkliar\_remark\][*By the discussion preceding this theorem, the hypothesis on $f$ could be replaced by the weaker: “$f\in (z-1)^{{\frac{1}{2}}+{\varepsilon}}\,{H^2}\backslash\{0\}$ and $M[|f|^2](-1)<\infty$, ” (cf. [@Chk]). If, in either version, the hypotheses on the attractive and repulsive fixed points are reversed, then the conclusion will assert that $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, except for perhaps a discrete subset, the annulus $A(1, \mu^{\varepsilon})$ (see §\[non-canonical\], especially the discussion preceding Corollary \[inverse\_cor\]).* ]{}
[*Note how the previously mentioned Theorem 5.5 of [@Mat2] follows from the work above. Indeed, if $f\in (z-1)^{2/p}\,H^p$ for some $p>2$ then by Hölder’s inequality $f\in (z-1)^{{\frac{1}{2}}+{\varepsilon}}\,{H^2}$, for each ${\varepsilon}<1/p$. Thus, as in the proof of Theorem \[main\_thm\_1\], the positively indexed subseries of converges for $|\lambda|>\mu^{-1/p}$, and by Lemma \[max\_fcn\_lemma\] the boundedness of $f$ in a neighborhood of $-1$ insures that the negatively indexed subseries of converges in the open unit disc. Thus as in the proof of Theorem \[main\_thm\_1\], $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, with the possible exception of a discrete subset, the open annulus $A(\mu^{-1/p},1)$.* ]{}
Complements and comments {#complements}
========================
In this section I collect some further results and say a few more words about the theorem of Nordgren, Rosenthal, and Wintrobe.
Non-canonical hyperbolic automorphisms {#non-canonical}
--------------------------------------
The results of §\[main\_results\], which refer only to canonical hyperbolic automorphisms ${\varphi}$, can be easily “denormalized”. Here is a sample:
\[non\_canonical\_thm\] Suppose ${\varphi}$ is a hyperbolic automorphism of ${\mathbb{U}}$ with attractive fixed point $\alpha$, repulsive one $\beta$, and multiplier $\mu>1$. Then
- [(cf. Theorem \[main\_thm\_1\])]{} Suppose, for $0<{\varepsilon},\delta<1/2$ we have $$f\in (z-\alpha)^{\frac{1}{2}+{\varepsilon}} (z-\beta)^{\frac{1}{2}+\delta}{H^2}\backslash\{0\}.$$ Then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, except possibly for a discrete subset, the open annulus $A(\mu^{-{\varepsilon}},\mu^\delta)$.
- [(cf. Theorem \[main\_thm\_2\])]{} If $f\in\sqrt{(z-\alpha)(z-\beta)} \, {H^2}$ then $\sigma_p({C_{\varphi}}|_{D_f})$ intersects ${\partial {\mathbb{U}}}$ in a set of positive measure.
- [(cf. Theorem \[main\_thm\_3\])]{} If $f\in \sqrt{(z-\alpha)(z-\beta)}\, H^p\backslash\{0\}$ for some $p>2$, then $\sigma_p({{C_{\varphi}}|_{D_f}})$ contains, except possibly for a discrete subset, the open annulus $A(\mu^{-{\varepsilon}},\mu^{\varepsilon})$ where ${\varepsilon}= {\frac{1}{2}}-\frac{1}{p}$.
I’ll just outline the idea, which contains no surprises. Suppose $\alpha$ and $\beta$ (both on ${\partial {\mathbb{U}}}$) are the fixed points of ${\varphi}$, and—for the moment—that ${\tilde{\alpha}}$ and ${\tilde{\beta}}$ are any two distinct points of ${\partial {\mathbb{U}}}$. Then, as we noted toward the end of §\[hypautos\], there is an automorphism $\psi$ of ${\mathbb{U}}$ that takes ${\tilde{\alpha}}$ to $\alpha$ and ${\tilde{\beta}}$ to $\beta$. Thus ${\tilde{{\varphi}}}:= \psi{^{-1}}\circ{\varphi}\psi$ is a hyperbolic automorphism of ${\mathbb{U}}$ that is easily seen to have attractive fixed point ${\tilde{\alpha}}$ and repulsive one ${\tilde{\beta}}$. Furthermore:
- $C_{{\tilde{{\varphi}}}} = {C_\psi}{C_{\varphi}}{C_\psi}{^{-1}}$, so $C_{{\tilde{{\varphi}}}}$ is similar to ${C_{\varphi}}$.
- For $f\in{H^2}$: ${C_\psi}D_f = D_{f\circ\psi}$.
- $F\in{H^2}$ is a $\lambda$-eigenvector for ${C_{\varphi}}$ if and only if ${C_\psi}=F\circ\psi$ is one for $C_{{\tilde{{\varphi}}}}$.
- For $f\in{H^2}$, $M[|f|^2](\beta)<\infty
\iff M[|f\circ\psi|^2](\beta)<\infty. $
- For any $\gamma>0$, $f\in (z-\alpha)^\gamma\, {H^2}\iff {C_\psi}f \in (z-{\tilde{\alpha}})^\gamma \, {H^2}$.
Only the last of these needs any comment. If $f\in (z-\alpha)^\gamma\, {H^2}$ then $$\begin{aligned}
{C_\psi}f &\in& (\psi(z) - \alpha)^\gamma {C_\psi}({H^2}) \\
& &\\
&=& \left(\frac{\psi(z)- \psi({\tilde{\alpha}})}{z-{\tilde{\alpha}}}\right)^\gamma
(z-{\tilde{\alpha}})^\gamma {H^2}\\
& &\\
&=& (z-{\tilde{\alpha}})^\gamma \, {H^2}\end{aligned}$$ where the last line follows from the fact that the quotient in the previous one is, in a neighborhood of the closed unit disc, analytic and non-vanishing (because $\psi$ is univalent there), hence both bounded and bounded away from zero on the closed unit disc. Thus ${C_\psi}((z-\alpha)^\gamma\,{H^2}) \subset (z-{\tilde{\alpha}})\, {H^2}$, and the opposite inclusion follows from this by replacing $\psi$ by $\psi{^{-1}}$ and applying ${C_\psi}$ to both sides of the result.
Theorem \[non\_canonical\_thm\] now follows, upon setting $({\tilde{\alpha}},{\tilde{\beta}}) = (+1,-1)$, from Theorems \[main\_thm\_1\], \[main\_thm\_2\], and \[main\_thm\_3\].
What happens if we interchange attractive and repulsive fixed points of ${\varphi}$ in the hypotheses of Theorem \[non\_canonical\_thm\](a)? Then the hypotheses apply to ${\varphi}{^{-1}}$, hence so does the conclusion. Since $C_{{\varphi}{^{-1}}} = {C_{\varphi}}{^{-1}}$, Theorem \[non\_canonical\_thm\](a) and the spectral mapping theorem yield, for example, the following complement to Theorem \[main\_thm\_4\]:
\[inverse\_cor\] If ${\varphi}$ is a hyperbolic automorphism of ${\mathbb{U}}$ with attractive fixed point $\alpha$, repulsive one $\beta$, and multiplier $\mu>1$. Suppose $f \in (z-\alpha)^\frac{1}{2}(z-\beta)^{\frac{1}{2}+{\varepsilon}}$ for some ${\varepsilon}\in(0,\frac{1}{2})$, then $\sigma_p({C_{\varphi}}|_{D_f})$ contains, except possibly for a discrete subset, the open annulus $A(1,\mu^{{\varepsilon}})$
The reader can easily supply similar “reversed” versions of the other results on the point spectrum of ${C_{\varphi}}|_{D_f}$.
The Nordgren-Rosenthal-Wintrobe Theorem {#RNW_thm}
---------------------------------------
Recall that this result equates a positive solution to the Invariant Subspace Problem for Hilbert space with a positive answer to the question: “For ${\varphi}$ a hyperbolic automorphism of ${\mathbb{U}}$, does does every nontrivial minimal ${C_{\varphi}}$-invariant subspace of ${H^2}$ contain an eigenfunction?” The theorem comes about in this way: About forty years ago Caradus [@Car] proved the following elementary, but still remarkable, result:
> [ *If an operator $T$ maps a separable, infinite dimensional Hilbert space onto itself and has infinite dimensional null space, then [*every*]{} operator on a separable Hilbert space is similar to a scalar multiple of the restriction of $T$ to one of its invariant subspaces.* ]{}
Consequently the invariant subspace lattice of $T$ contains that of every operator on a separable Hilbert space.
Now all composition operators (except the ones induced by constant functions) are one-to-one, so none of these obeys the Caradus theorem’s hypotheses. However Nordgren, Rosenthal, and Wintrobe were able to show that if ${\varphi}$ is a hyperbolic automorphism, then for every eigenvalue $\lambda$ of ${C_{\varphi}}$ the operator ${C_{\varphi}}-\lambda I$, which has infinite dimensional kernel (recall Theorem \[spectrum\_thm\]), maps ${H^2}$ onto itself. Their restatement of the Invariant Subspace Problem follows from this via the Caradus theorem and the fact that ${C_{\varphi}}$ and ${C_{\varphi}}-\lambda I$ have the same invariant subspaces.
Cyclicity
---------
Minimal invariant subspaces for invertible operators are both cyclic and doubly invariant—this was the original motivation for studying the subspaces $D_f$. Thus it makes sense, for a given doubly invariant subspace, and especially for a doubly cyclic one $D_f$, to ask whether or not it is cyclic. Here is a result in that direction in which the cyclicity is the strongest possible: [*hypercyclicity*]{}—some orbit (with no help from the linear span) is dense. I state it for canonical hyperbolic automorphisms; the generalization to non-canonical ones follows from the discussion of §\[non-canonical\] and the similarity invariance of the property of hypercyclicity.
\[cyclic\_prop\] Suppose ${\varphi}$ is a canonical hyperbolic automorphism of ${\mathbb{U}}$ and $f\in\sqrt{(z+1)(z-1)} \, {H^2}$. Then ${{C_{\varphi}}|_{D_f}}$ is hypercyclic.
A sufficient condition for an invertible operator on a Banach space $X$ to be hypercyclic is that for some dense subset of the space, the positive powers of both the operator and its inverse tend to zero pointwise in the norm of $X$ (see [@S Chapter 7, page 109], for example; much weaker conditions suffice). In our case the dense subspace is just the linear span of $S:=\{f\circ{\varphi}_n: n\in{\mathbb{Z}}\}$. As we saw in the proof of Theorem \[main\_thm\_2\], our hypothesis on $f$ insures that $\sum_{n\in{\mathbb{Z}}}{\|f\circ{\varphi}_n\|}^2 <\infty$ so both $({C_{\varphi}}^n)_0^\infty$ and $({C_{\varphi}}^{-n})_0^\infty$ converge pointwise to zero on $S$, and therefore pointwise on its linear span.
\[hc\_remark\][*One can obtain the conclusion of Proposition \[cyclic\_prop\] under different hypotheses. For example if $f$ is continuous with value zero at both of the fixed points of ${\varphi}$, then the same is true of the restriction of $|f|^2$ to ${\partial {\mathbb{U}}}$. Thus the Poisson integral of $|f|^2$ has radial limit zero at each fixed point of ${\varphi}$ (see [@Rud Theorem 11.3, page 244], for example), so by , just as in the proof of Proposition \[cyclic\_prop\], ${{C_{\varphi}}|_{D_f}}$ satisfies the sufficient condition for hypercyclicity. In fact, all that is really needed for this argument is that the measure $$E{\rightarrow}\int_E |f|^2 \, dm \qquad (E~{\rm measurable}\subset{\partial {\mathbb{U}}})$$ have symmetric derivative zero at both fixed points of ${\varphi}$ (see the reference above to [@Rud]).* ]{}
[99]{}
Patrick Ahern, [*The mean modulus and the derivative of an inner function*]{}, Indiana Univ. Math. J. 28 (1979), 311–347.
S.R. Caradus, [*Universal operators and invariant subspaces,*]{} Proc. Amer. Math. Soc. 23 (1969), 526–527.
Vitali Chkliar, [*Eigenfunctions of the hyperbolic composition operator,*]{} Integr. Equ. Oper. Theory 29 (1997), 264–367.
Valentin Matache, [*On the minimal invariant subspaces of the hyperbolic composition operator,*]{} Proc. Amer. Math. Soc. 119 (1993), 837–841.
Valentin Matache, [*The eigenfunctions of a certain composition operator,*]{} Contemp. Math. 213 (1998), 121–136.
Valentin Matache, [*Notes on hypercyclic operators,*]{} Acta Sci. Math. (Szeged) 58 (1993), 397–406.
Eric Nordgren, [*Composition operators,*]{} Canadian J. Math. 20 (1968), 442–449.
Eric Nordgren, Peter Rosenthal, and F.S. Wintrobe, [*Invertible composition operators on $H^p$,*]{} J. Functional Anal. 73 (1987), 324–344.
Walter Rudin, [*Real and Complex Analysis,*]{} third ed., McGraw-Hill, New York, 1987.
Joel H. Shapiro, [*Composition Operators and Classical Function Theory,*]{} Springer-Verlag 1993.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'g2bibliography\_abbrv.bib'
---
Acknowledgements
================
This work was supported in part by the US DOE, Fermilab and Argonne National Laboratory under contract No. KA2201020.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Hardness magnification reduces major complexity separations (such as $\mathsf{\mathsf{EXP}} \nsubseteq \mathsf{NC}^1$) to proving lower bounds for some natural problem $Q$ against weak circuit models. Several recent works [@OS18_mag_first; @MMW_STOC_paper; @CT19_STOC; @OPS19_CCC; @CMMW_CCC_paper; @DBLP:conf/icalp/Oliveira19; @Magnification_FOCS19] have established results of this form. In the most intriguing cases, the required lower bound is known for problems that appear to be significantly easier than $Q$, while $Q$ itself is susceptible to lower bounds but these are not yet sufficient for magnification.
In this work, we provide more examples of this phenomenon, and investigate the prospects of proving new lower bounds using this approach. In particular, we consider the following essential questions associated with the hardness magnification program:
– *Does hardness magnification avoid the natural proofs barrier of Razborov and Rudich *[@DBLP:journals/jcss/RazborovR97]*?*
– *Can we adapt known lower bound techniques to establish the desired lower bound for $Q$?*
We establish that some instantiations of hardness magnification overcome the natural proofs barrier in the following sense: slightly superlinear-size circuit lower bounds for certain versions of the minimum circuit size problem [${\sf MCSP}$]{}imply the non-existence of natural proofs. As a corollary of our result, we show that certain magnification theorems not only imply strong worst-case circuit lower bounds but also rule out the existence of efficient learning algorithms.
Hardness magnification might sidestep natural proofs, but we identify a source of difficulty when trying to adapt existing lower bound techniques to prove strong lower bounds via magnification. This is captured by a *locality barrier*: existing magnification theorems *unconditionally* show that the problems $Q$ considered above admit highly efficient circuits extended with small fan-in oracle gates, while lower bound techniques against weak circuit models quite often easily extend to circuits containing such oracles. This explains why direct adaptations of certain lower bounds are unlikely to yield strong complexity separations via hardness magnification.
author:
- 'Lijie Chen[^1]\'
- 'Shuichi Hirahara[^2]\'
- 'Igor C. Oliveira[^3]\'
- 'Ján Pich[^4]\'
- 'Ninad Rajgopal[^5]\'
- |
Rahul Santhanam[^6]\
\
bibliography:
- 'refs.bib'
title: |
Beyond Natural Proofs:\
Hardness Magnification and Locality
---
Preliminaries {#s:preliminaries}
=============
Magnification Frontiers {#a:improved_magnification_MCSP}
=======================
Hardness Magnification and Natural Proofs {#s:non-natural}
=========================================
The Locality Barrier {#s:difficulties}
====================
Acknowledgements {#acknowledgements .unnumbered}
================
Part of this work was completed while some of the authors were visiting the Simons Institute for the Theory of Computing. We are grateful to the Simons Institute for their support. This work was supported in part by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2014)/ERC Grant Agreement no. 615075. Ján Pich was supported in part by Grant 19-05497S of GA ČR. Lijie Chen is supported by NSF CCF-1741615 and a Google Faculty Research Award. Igor C. Oliveira was supported in part by a Royal Society University Research Fellowship.[^7]
Review of Hardness Magnification in Circuit Complexity {#a:review_magnification}
======================================================
[^1]: `[email protected]`
[^2]: `[email protected]`
[^3]: `[email protected]`
[^4]: `[email protected]`
[^5]: `[email protected]`
[^6]: `[email protected]`
[^7]: Most of this work was completed while Igor C. Oliveira was affiliated with the University of Oxford.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Thanks to incredible advances in instrumentation, surveys like the Sloan Digital Sky Survey have been able to find and catalog billions of objects, ranging from local M dwarfs to distant quasars. Machine learning algorithms have greatly aided in the effort to classify these objects; however, there are regimes where these algorithms fail, where interesting oddities may be found. We present here an X-ray bright quasar misidentified as a red supergiant/X-ray binary, and a subsequent search of the SDSS quasar catalog for X-ray bright stars misidentified as quasars.'
---
Introduction/Overview
=====================
Red Supergiant X-ray Binaries
-----------------------------
Over the past decade, many exotic close binary systems with supergiant components have been discovered. Systems like NGC 300 X-1 — a Wolf-Rayet/black hole X-ray binary ([@crowther10 Crowther 2010]) — and SN2010da — a sgB\[e\]/neutron star X-ray binary ([@villar16 Villar 2016]) — are two such examples of a coupling between a massive star in a short-lived evolutionary phase and a compact stellar remnant. Interestingly, no X-ray binaries with confirmed red supergiant (RSG) counterparts have been discovered (RSGs have been proposed as the candidate donor star for a few Ultraluminous X-ray Sources, see [@heida16 Heida 2016]). This may be partially explained by the rarity of RSGs; however, though rare, RSGs are both longer-lived and more common (due to the smaller — 10 - 25 $M_\odot$ — initial masses of their zero age main sequence progenitors) than most other evolved massive stars.
RSG X-ray binaries, if they exist, offer a view into an interesting edge case of accretion; their extended envelopes and strong winds ($M \sim 10^{-4}\:M_\odot$ yr$^{-1}$, [@vanloon05 van Loon 2005]) could allow for accretion from both the wind and Roche-Lobe Overflow in an environment continually enriched with dust produced by the RSG. RSG X-ray binaries are also the immediate progenitors of Thorne-Żytkow Objects — stars with embedded neutron star cores ([@thorne75 Thorne & Żytkow 1975]) — assuming the neutron star plunges into the RSG as it expands ([@taam78 Taam 1978]).
J0045+41
--------
To search for RSG X-ray binaries, we used the photometry of the Local Group Galaxy Survey (LGGS, [@massey03; @massey06; @massey07 Massey & Olsen 2003, Massey 2006, 2007]), which covers M31, M33, the Magellanic Clouds and 7 dwarf galaxies in the Local Group. Following [@massey98] to find RSGs among the nearly-identical foreground dwarfs, we cross-referenced the positions of the LGGS RSGs with the [*Chandra*]{} Source Catalog (CSC, [@evans10 Evans 2010]), and found one RSG coincident with an X-ray source.
LGGS J004527.30+413254.3 (J0045+41 hereafter) is a bright ($V \approx 19.9$) object of previously-unknown nature in the disk of M31. [@vilardell06] classify J0045+41 as an eclipsing binary with a period of $\sim 76$ days. J0045+41 was also observed with the Palomar Transient Factory (PTF); the $g$-band lightcurve shows evidence for a $\sim650$ day period. On the other hand [@kim07] identify J0045+41 as a globular cluster, and it has been included in catalogs of M31 globular clusters as recently as 2014 ([@wang14 Wang 2014]). The LGGS photometry is consistent with the color and brightness of a RSG. Indeed, following [@levesque06], we found that, as an RSG, J0045+41 would have an effective temperature of $\sim$3500 K and bolometric magnitude of -6.67, consistent with a 12-15 M$_{\odot}$ RSG. However, a complete SED fit to photometry from the Panchromatic Hubble Andromeda Treasury (PHAT, [@dalcanton12 Dalcanton 2012]) using the Bayesian Extinction and Stellar Tool (BEAST, [@gordon16 Gordon 2016]) yields an unphysical result of 300 M$_{\odot}$, 10$^{\rm 5}$ K star, extincted by $A_V\sim$4 magnitudes. Furthermore, the object appears extended in the PHAT images (though its radial profile appears similar to that of other nearby stars).
J0045+41 is separated by $\sim1.18^{\prime\prime}$ from an X-ray source. The source, CXO J004527.3\
+413255, is bright ($F_X = 1.98\times10^{-13}$ erg s$^{-1}$ cm$^{-2}$) and hard; fitting a spectrum obtained by Williams et al. (in prep) yields a power law with $\Gamma\sim1.5$. The best-fit neutral hydrogen column density is $1.7\times10^{21}$ cm$^{-2}$, which corresponds to $A_V\sim1$.
### Observations and Data Reduction
The apparent periodicity of J0045+41 and its apparent association with a hard and unabsorbed X-ray source prompted us to obtain follow-up spectroscopic observations to determine the true nature of this object. We obtained a longslit spectrum of J0045+41 using the Gemini Multi-Object Spectrograph (GMOS) on Gemini North. Four 875 second exposures were taken 2016 July 5 using the `B600` grating centered on 5000 Å, and four 600 second exposures were taken 2016 July 9 using the `R400` grating centered on 7000 Å, with a blocking filter to remove 2$^{nd}$-order diffraction. Due to the gaps between GMOS’s three CCDs, two of each set of exposures were offset by +50 Å. The data were reduced using the standard `gemini` `IRAF` package. The final reduced spectrum has continuous signal from $\sim$4000 to $\sim$9100 Å at a resolution of $R \sim 1688\:({\rm blue})/1918\:({\rm red})$.
### Spectrum and Redshift Determination
![GMOS spectrum of J0045+41 with all identified lines labeled. The Na II feature is intrinsic to the Milky Way/M31, and the $\oplus$ line is telluric.[]{data-label="fig:spectrum"}](J0045_41_long_lineID.eps){width="90.00000%"}
The spectrum (Figure \[fig:spectrum\]) shows that J0045+41 is a quasar at $z\approx0.21$ (measured with H$\alpha$, H$\beta$, \[OIII\]$\lambda5007$ and Ca II H and K). While a false positive, this quasar is quite interesting on its own. The (low-significance) detections of periodicity on short timescales by multiple sources are difficult to explain. Furthermore, mistaking a blue quasar for a red star would imply a high reddening along the line of sight, which is consistent with a sightline through M31; however, the low H column density implied by the X-ray counterpart’s spectrum and our inability to achieve a satisfactory fit to the optical spectrum by reddening the (redshifted) quasar template spectrum from [@vandenberk01] indicate that J0045+41 belongs to a small and intriguing class of intrinsically red quasars, observed through a low-extinction region of the ISM in M31 ([@richards03 Richards 2003]).
Stars in the SDSS Quasar Catalog
================================
If a quasar can be misidentified as a RSG, are there red stars — especially X-ray bright stars — in already-existing quasar catalogs? To answer this question, we turned to the quasar catalog of the Sloan Digital Sky Survey (SDSS, [@york00 York 2000]).
Sample Selection
----------------
We selected all SDSS objects automatically tagged as quasars that were within 0.2 magnitudes of J0045+41 in $g-r$ vs. $r-i$ vs. $i-z$ color-space — J0045+41 is too faint in $u$ to utilize $u-g$ — and ignored any warning flags to avoid throwing out interesting objects that were not easily identified by the SDSS algorithm. 1098 objects in this sample had associated spectroscopic observations. Interestingly, many of the spectroscopically-determined redshifts were unbelievably small or even negative, implying that these objects are in a regime of color-space where classification algorithms may fail. Indeed, on visual inspection of these spectra, many of them are stellar.
Stars
-----
We used `emcee`, a Python implementation of Markov Chain Monte Carlo (MCMC) by [@formanmackey13], to fit Gaussian profiles to the Ca II triplet ($\lambda = 8498,\: 8542,\: {\rm and}\: 8662$ Å), which we use to identify stars. The posterior distributions of the parameters allow us to determine if the triplet is well fit, and estimate errors for each parameter. Because the relative centroids and strengths of the lines are fixed, a good fit guarantees the lines are actually detected, while simultaneously measuring — with accurate errors — the radial velocity and equivalent width ($W_\lambda$) of the triplet. After a follow-up inspection by eye of spectra that were noisy or missing data, we find 344 confirmed cool stars, representing $\sim$31% of the total sample. Figure \[fig:ew\_hist\] shows the distribution of $W_\lambda$ for $W_{\lambda}/\sigma_{W_\lambda} > 1$. We follow [@jennings16] to estimate luminosity from $W_\lambda$, and find that most stars are dwarfs ($W_{\lambda} \lesssim 6.5$ Å) but $\sim$40 stars have larger equivalent widths indicating they are likely giants or supergiants (the relationship is dependent on effective temperature and metallicity, so these labels are approximate).
![Distribution of measured equivalent widths for confirmed stars.[]{data-label="fig:ew_hist"}](EW_hist_hand_bw.eps){width="90.00000%"}
Discussion and Future Work
==========================
This result demonstrates that, when looking for rare objects like RSG X-ray binaries, it is important to look in unlikely places; e.g., a red and X-ray bright star may be confused for a quasar if the classification algorithm mistakes the continuum between the TiO bands in a RSG spectrum for an emission line. The fact that some of the color-space containing M-dwarfs — by far the most common type of star — is a regime where classification algorithms fail underlines the importance of improving on these algorithms until they perform as well as the human eye. Indeed, many of these stars were previously identified (see [@west11 West 2011]), but are still listed as quasars on the SDSS online data portal.
Future work will focus on improving our star-finding algorithm to use alternate spectral features when the Ca II triplet is missing or obscured by noise, and on finding which areas of color-space contain significant numbers of these misidentified stars, with the goal of finding RSG X-ray binaries as well as improving our knowledge of where classification algorithms fail.
2010, [[*MNRAS*]{}]{}, 403, L41
Dalcanton, J. J., Williams, B. F., Lang, D., et al. 2012, [[*ApJS*]{}]{}, 200, 18
Davidsen, A., Malina, R., & Bowyer, S. 1977, [[*ApJ*]{}]{}, 211, 866
Evans, I. N., Primini, F. A., Glotfelty, K. J., et al. 2010, [[*ApJS*]{}]{}, 189, 37-82
Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, [[*PASP*]{}]{}, 125, 306
Gordon, K. D., Fouesneau, M., Arab, H., et al. 2016, [[*ApJ*]{}]{}, 826, 104
2016, [[*MNRAS*]{}]{}, 459, 771
Jennings, J., & Levesque, E. M. 2016, [[*ApJ*]{}]{}, 821, 131
Kim, S. C., Lee, M. G., Geisler, D., et al. 2007, [[*AJ*]{}]{}, 134, 706
Levesque, E. M., Massey, P., Olsen, K. A. G., et al. 2006, [[*ApJ*]{}]{}, 645, 1102
Massey, P. 1998, [[*ApJ*]{}]{}, 501, 153
Massey, P., & Olsen, K. A. G. 2003, [[*AJ*]{}]{}, 126, 2867
Massey, P., Olsen, K. A. G., Hodge, P. W., et al. 2006, [[*AJ*]{}]{}, 131, 2478
Massey, P., Olsen, K. A. G., Hodge, P. W., et al. 2007, [[*AJ*]{}]{}, 133, 2393
Richards, G. T., Hall, P. B., Vanden Berk, D. E., et al. 2003, [[*AJ*]{}]{}, 126, 1131
Taam, R. E., Bodenheimer, P., & Ostriker, J. P. 1978, [[*ApJ*]{}]{}, 222, 269
Thorne, K. S., & Żytkow, A. N. 1975, [[*ApJL*]{}]{}, 199, L19
Vanden Berk, D. E., Richards, G. T., Bauer, A., et al. 2001, [[*AJ*]{}]{}, 122, 549
van Loon, J. T., Cioni, M.-R. L., Zijlstra, A. A., & Loup, C. 2005, [[*A&A*]{}]{}, 438, 273
Vilardell, F., Ribas, I., & Jordi, C. 2006, [[*A&A*]{}]{}, 459, 321
2016, [[*ApJ*]{}]{}, 830, 11
Wang, S., Ma, J., Wu, Z., & Zhou, X. 2014, [[*AJ*]{}]{}, 148, 4
West, A. A., Morgan, D. P., Bochanski, J. J., et al. 2011, [[*AJ*]{}]{}, 141, 97
York, D. G., Adelman, J., Anderson, J. E., Jr., et al. 2000, [[*AJ*]{}]{}, 120, 1579
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The cosparse analysis model has been introduced recently as an interesting alternative to the standard sparse synthesis approach. A prominent question brought up by this new construction is the analysis pursuit problem – the need to find a signal belonging to this model, given a set of corrupted measurements of it. Several pursuit methods have already been proposed based on $\ell_1$ relaxation and a greedy approach. In this work we pursue this question further, and propose a new family of pursuit algorithms for the cosparse analysis model, mimicking the greedy-like methods – compressive sampling matching pursuit (CoSaMP), subspace pursuit (SP), iterative hard thresholding (IHT) and hard thresholding pursuit (HTP). Assuming the availability of a near optimal projection scheme that finds the nearest cosparse subspace to any vector, we provide performance guarantees for these algorithms. Our theoretical study relies on a restricted isometry property adapted to the context of the cosparse analysis model. We explore empirically the performance of these algorithms by adopting a plain thresholding projection, demonstrating their good performance.'
address:
- 'The Department of Computer Science, Technion – Israel Institute of Technology, Haifa 32000, Israel'
- 'INRIA Rennes - Bretagne Atlantique, Campus de Beaulieu, F-35042 Rennes Cedex, France'
- |
School of Engineering and Electronics, The University of Edinburgh,\
The King’s Buildings, Mayfield Road, Edinburgh EH9 3JL, UK
author:
- 'R. Giryes'
- 'S. Nam'
- 'M. Elad'
- 'R.Gribonval'
- 'M. E. Davies'
bibliography:
- 'AnalysisGreedyLike.bib'
title: 'Greedy-Like Algorithms for the Cosparse Analysis Model'
---
Sparse representations ,Compressed sensing ,Synthesis ,Analysis ,CoSaMP ,Subspace-pursuit ,Iterative hard threshodling ,Hard thresholding pursuit.
94A20 ,94A12 ,62H12
Introduction {#sec:intro}
============
Many natural signals and images have been observed to be inherently low dimensional despite their possibly very high ambient signal dimension. It is by now well understood that this phenomenon lies at the heart of the success of numerous methods of signal and image processing. Sparsity-based models for signals offer an elegant and clear way to enforce such inherent low-dimensionality, explaining their high popularity in recent years. These models consider the signal $\x\in {\mathbb{R}^{\sdim}} $ as belonging to a finite union of subspaces of dimension $k\ll \sdim$ [@Lu08Theory]. In this paper we shall focus on one such approach – the cosparse analysis model – and develop pursuit methods for it.
Before we dive into the details of the model assumed and the pursuit problem, let us first define the following generic inverse problem that will accompany us throughout the paper: For some unknown signal $\x \in {\mathbb{R}^{\sdim}}$, an incomplete set of linear observations $\y \in {\mathbb{R}^{\mdim}}$ (incomplete implies $\mdim<\sdim$) is available via $$\label{eq:inverseProblem}
\y = \M \x + \e,$$ where $\e \in {\mathbb{R}^{\mdim}}$ is an additive bounded noise that satisfies ${\left\Vert\e\right\Vert}_2^2 \leq \epsilon^2$. The task is to recover or approximate $\x$. In the noiseless setting where $\e = 0$, this amounts to solving $\y = \M \x$. Of course, a simple fact in linear algebra tells us that this problem admits infinitely many solutions (since $\mdim<\sdim$). Therefore, when all we have is the observation $\y$ and the measurement/observation matrix $\M\in {\mathbb{R}^{\mdim\times \sdim}}$, we are in a hopeless situation to recover $\x$.
The Synthesis Approach
----------------------
This is where ‘sparse signal models’ come into play. In the sparse synthesis model, the signal $\x$ is assumed to have a very sparse representation in a given fixed dictionary $\Dict \in {\mathbb{R}^{\sdim\times\ddim}}$. In other words, there exists ${{\boldsymbol{\alpha}}}$ with few nonzero entries, as counted by the “$\ell_{0}$-norm” ${\left\Vert{{\boldsymbol{\alpha}}}\right\Vert}_{0}$, such that $$\label{eq:synthesisModel}
\x = \Dict {{\boldsymbol{\alpha}}}, \quad \text{and} \quad \spar := {\left\Vert{{\boldsymbol{\alpha}}}\right\Vert}_0 \ll \sdim.$$ Having this knowledge we solve (\[eq:inverseProblem\]) using $$\label{eq:synthesisL0}
\hat\x_{\ell_0} = \Dict \hat{{{\boldsymbol{\alpha}}}}_{\ell_0}, \quad \text{and} \quad \hat{{{\boldsymbol{\alpha}}}}_{\ell_0} = {\operatornamewithlimits{argmin}}_{{{\boldsymbol{\alpha}}}} {\left\Vert{{\boldsymbol{\alpha}}}\right\Vert}_0 \subjectto {\left\Vert\y - \M \Dict {{\boldsymbol{\alpha}}}\right\Vert}_2 \le \epsilon.
$$ More details about the properties of this problem can be found in [@Donoho02Optimally; @Gribonval03spars]. Since solving is an NP-complete problem [@Davis97Adaptive], approximation techniques are required for recovering ${{\bfx}}$. One strategy is by using relaxation, replacing the $\ell_0$ with $\ell_1$ norm, resulting with the $\ell_1$-synthesis problem $$\begin{aligned}
\label{eq:l1_synthesis}
\hat\x_{\ell_1} = \Dict \hat{{{\boldsymbol{\alpha}}}}_{\ell_1}, \quad \text{and} \quad \hat{{{\boldsymbol{\alpha}}}}_{\ell_1} = {\operatornamewithlimits{argmin}}{\left\Vert{{\boldsymbol{\alpha}}}\right\Vert}_1 & s.t. & {\left\Vert{{\bfy}} - {{\bfM}}{{\bfD}}{{\boldsymbol{\alpha}}}\right\Vert}_2 \le \epsilon.\end{aligned}$$ For a unitary matrix ${{\bfD}}$ and a vector ${{\bfx}}$ with $k$-sparse representation ${{\boldsymbol{\alpha}}}$, if $\delta_{2k} < \delta_{\ell_1}$ then $$\begin{aligned}
\label{eq:l1_rec_error}
{\left\Vert\hat{{{\bfx}}}_{\ell_1} - {{\bfx}}\right\Vert}_2 \le C_{\ell_1}{\left\Vert\e\right\Vert}_2,\end{aligned}$$ where $\hat{{{\bfx}}}_{\ell_1} = {{\bfD}}\hat{{{\boldsymbol{\alpha}}}}_{\ell_1}$, $\delta_{2k}$ is the constant of the restricted isometry property (RIP) of ${{\bfMD}}$ for 2k sparse signals, $C_{\ell_1}$ is a constant greater than $\sqrt{2}$ and $\delta_{\ell_1}$ ($\simeq 0.4931$) is a reference constant [@Candes06Near; @foucart10Sparse; @Mo11New]. Note that this result implies a perfect recovery in the absence of noise. The above statement was extended also for incoherent redundant dictionaries [@Rauhut08Compressed].
Another option for approximating is using a greedy strategy, like in the thresholding technique or orthogonal matching pursuit (OMP) [@Pati93OMP; @MallatZhang93]. A different related approach is the greedy-like family of algorithms. Among those we have compressive sampling matching pursuit (CoSaMP) [@Needell09CoSaMP], subspace pursuit (SP) [@Dai09Subspace], iterative hard thresholding (IHT) [@Blumensath09Iterative] and hard thresholding pursuit (HTP) [@Foucart11Hard]. CoSaMP and SP were the first greedy methods shown to have guarantees in the form of assuming $\delta_{4k}< \delta_{\text{\tiny CoSaMP}}$ and $\delta_{3k} \le \delta_{\text{\tiny SP}}$ [@Needell09CoSaMP; @Dai09Subspace; @foucart10Sparse; @Giryes12RIP]. Following their work, iterative hard thresholding (IHT) and hard thresholding pursuit (HTP) were shown to have similar guarantees under similar conditions [@Blumensath09Iterative; @Foucart11Hard; @Garg09Gradient; @foucart10Sparse]. Recently, a RIP based guarantee was developed also for OMP [@Zhang11Sparse].
The Cosparse Analysis Model
---------------------------
Recently, a new signal model called *cosparse analysis model* was proposed in [@Nam12Cosparse; @Nam11CosparseConf]. The model can be summarized as follows: For a fixed analysis operator $\OM\in{\mathbb{R}^{\pdim\times\sdim}}$ referred to as the analysis dictionary, a signal $\x\in{\mathbb{R}^{\sdim}}$ belongs to the cosparse analysis model with cosparsity $\cosp$ if $$\label{eq:analysisModel}
\cosp := \pdim - {\left\Vert\OM\x\right\Vert}_0.$$ The quantity $\cosp$ is the number of rows in $\OM$ that are orthogonal to the signal. The signal $\x$ is said to be $\cosp$-cosparse, or simply cosparse. We denote the indices of the zeros of the analysis representation as the *cosupport* $\Lambda$ and the sub-matrix that contains the rows from $\OM$ that belong to $\Lambda$ by $\OM_\Lambda$. As the definition of cosparsity suggests, the emphasis of the cosparse analysis model is on the zeros of the analysis representation vector $\OM\x$. This contrasts the emphasis on ‘few non-zeros’ in the synthesis model . It is clear that in the case where every $\cosp$ rows in $\OM$ are independent, $\x$ resides in a subspace of dimension $\sdim - \cosp$ that consists of vectors orthogonal to the rows of $\OM_\Lambda$. In the general case where dependencies occur between the rows of $\OM$, the dimension is $d$ minus the rank of $\OM_\Lambda$. This is similar to the behavior in the synthesis case where a $k$-sparse signal lives in a $k$-dimensional space. Thus, for this model to be effective, we assume a large value of $\ell$.
In the analysis model, recovering $\x$ from the corrupted measurements is done by solving the following minimization problem [@elad07Analysis]: $$\label{eq:analysisL0}
\x_{A-\ell_0} = {\operatornamewithlimits{argmin}}_{\x} {\left\Vert\OM\x\right\Vert}_0 \subjectto {\left\Vert\y - \M \x\right\Vert}_2 \le \epsilon.$$ Solving this problem is NP-complete [@Nam12Cosparse], just as in the synthesis case, and thus approximation methods are required. As before, we can use an $\ell_1$ relaxation to , replacing the $\ell_0$ with $\ell_1$ in , resulting with the $\ell_1$-analysis problem [@Nam12Cosparse; @elad07Analysis; @Candes11Compressed; @Vaiter12Robust]. Another option is the greedy approach. A greedy algorithm called Greedy Analysis Pursuit (GAP) has been developed in [@Nam12Cosparse; @Nam11CosparseConf; @Nam11GAPN] that somehow mimics Orthogonal Matching Pursuit [@Pati93OMP; @MallatZhang93] with a form of iterative reweighted least Squares (IRLS) [@IRLS10Deubechies]. Other alternatives for OMP, backward greedy (BG) and orthogonal BG (OBG), were presented in [@Rubinstein12Cosparse] for the case that ${{\bfM}}$ is the identity. For the same case, the parallel to the thresholding technique was analyzed in [@Peleg12Performance].
This Work
---------
Another avenue exists for the development of analysis pursuit algorithms – constructing methods that will imitate the family of greedy-like algorithms. Indeed, we have recently presented preliminary and simplified versions of analysis IHT (AIHT), analysis HTP (AHTP), analysis CoSaMP (ACoSaMP) and Analysis SP (ASP) in [@Giryes11Iterative; @Giryes12CoSaMP] as analysis versions of the synthesis counterpart methods. This paper re-introduces these algorithms in a more general form, ties them to their synthesis origins, and analyze their expected performance. The main contribution of the paper is our result on the stability of these analysis pursuit algorithms. We show that after a finite number of iterations and for a given constant $c_0$, the reconstruction result $\hat\x$ of AIHT, AHTP, ACoSaMP and ASP all satisfy $$\begin{aligned}
{\left\Vert\x - \hat\x\right\Vert}_2 \le c_0{\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ under a RIP-like condition on $\M$ and the assumption that we are given a good near optimal projection scheme. A bound is also given for the case where $\x$ is only nearly $\cosp$-cosparse. Similar results for the $\ell_1$ analysis appear in [@Candes11Compressed; @Vaiter12Robust]. More details about the relation between these papers and our results will be given in Section \[sec:guarantees\]. In addition to our theoretical results we demonstrate the performance of the four pursuit methods under a thresholding based simple projection scheme. Both our theoretical and empirical results show that linear dependencies in $\OM$ that result with a larger cosparsity in the signal $\x$, lead to a better reconstruction performance. [*This suggests that, as opposed to the synthesis case, strong linear dependencies within $\OM$ are desired.*]{}
This paper is organized as follows:
- In Section \[sec:notation\] we present the notation used along the paper.
- In Section \[sec:omega\_RIP\] we define a RIP-like property, the $\OM$-RIP, for the analysis model, proving that it has similar characteristics like the regular RIP. In Section \[sec:near\_opt\_proj\] the notion of near optimal projection is proposed and some nontrivial operators for which a tractable optimal projection exists are exhibited. Both the $\OM$-RIP and the near optimal projection are used throughout this paper as a main force for deriving our theoretical results.
- In Section \[sec:analysis\_alg\] the four pursuit algorithms for the cosparse analysis framework are defined, adapted to the general format of the pursuit problem we have defined above.
- In Section \[sec:guarantees\] we derive the success guarantees for all the above algorithms in a unified way. Note that the provided results can be easily adapted to other union-of-subspaces models given near optimal projection schemes for them, in the same fashion done for IHT with an optimal projection scheme in [@Blumensath09Sampling]. The relation between the obtained results and existing work appears in this section as well.
- Empirical performance of these algorithms is demonstrated in Section \[sec:exp\] in the context of the cosparse signal recovery problem. We use a simple thresholding as the near optimal projection scheme in the greedy-like techniques.
- Section \[sec:conc\] discuss the presented results and concludes our work.
Notations and Preliminaries {#sec:notation}
===========================
We use the following notation in our work:
- $\sigma_{{{\bfM}}}$ is the largest singular value of ${{\bfM}}$, i.e., $\sigma_{{{\bfM}}}^2 = {\left\Vert{{\bfM}}^*{{\bfM}}\right\Vert}_2$.
- ${\left\Vert\cdot\right\Vert}_2$ is the euclidian norm for vectors and the spectral norm for matrices. ${\left\Vert\cdot\right\Vert}_1$ is the $\ell_1$ norm that sums the absolute values of a vector and ${\left\Vert\cdot\right\Vert}_0$, though not really a norm, is the $\ell_0$-norm which counts the number of non-zero elements in a vector.
- Given a cosupport set $\Lambda$, ${{\bf\Omega}}_{\Lambda}$ is a sub-matrix of ${{\bf\Omega}}$ with the [*rows*]{} that belong to $\Lambda$.
- For given vectors ${{\bfv}},{{\bfz}}\in {\mathbb{R}^{d}}$ and an analysis dictionary $\OM$, $\operatorname{cosupp}(\OM{{\bfv}})$ returns the cosupport of $\OM{{\bfv}}$ and $\operatorname{cosupp}(\OM{{\bfz}}, \cosp)$ returns the index set of $\cosp$ smallest (in absolute value) elements in $\OM{{\bfz}}$. If more than $\ell$ elements are zero all of them are returned. In the case where the $\cosp$-th smallest entry is equal to the $\cosp+1$ smallest entry, one of them is chosen arbitrarily.
- In a similar way, in the synthesis case ${{\bfD}}_T$ is a sub-matrix of ${{\bfD}}$ with [*columns*]{}[^1] corresponding to the set of indices $T$, $\operatorname{supp}(\cdot)$ returns the support of a vector, $\operatorname{supp}(\cdot, k)$ returns the set of $k$-largest elements and $\lceil \cdot \rceil_k$ preserves the $k$-largest elements in a vector. In the case where the $k$-th largest entry is equal to the $k+1$ largest entry, one of them is chosen arbitrarily.
- $\Q_\Lambda = {{\bfI}} - \OM_\Lambda^\dag\OM_\Lambda$ is the orthogonal projection onto the orthogonal complement of $\operatorname{range}({{\bf\Omega}}_\Lambda^*)$.
- $\P_\Lambda = {{\bfI}} - \Q_\Lambda = \OM_\Lambda^\dag\OM_\Lambda$ is the orthogonal projection onto $\operatorname{range}({{\bf\Omega}}_\Lambda^*)$.
- $\hat\x_{\text{\tiny AIHT}}$/$\hat\x_{\text{\tiny AHTP}}$/$\hat\x_{\text{\tiny ACoSaMP}}$/$\hat\x_{\text{\tiny ASP}}$ are the reconstruction results of AIHT/ AHTP/ ACoSaMP/ ASP respectively. Sometimes when it is clear from the context to which algorithms we refer, we abuse notations and use $\hat\x$ to denote the reconstruction result.
- A cosupport $\Lambda$ has a corank $r$ if $\operatorname{rank}(\OM_\Lambda)=r$. A vector ${{\bfv}}$ has a corank $r$ if its cosupport has a corank $r$.
- $[p]$ denotes the set of integers $[1 \dots p]$.
- $\L_{{{\bf\Omega}},\cosp} = \{\Lambda \subseteq [p], {\left\vert\Lambda\right\vert}\ge\cosp\}$ is the set of $\cosp$-cosparse cosupports and $\L_{{{\bf\Omega}},r}^{\text{corank}} = \{\Lambda \subseteq [p], \operatorname{rank}({{\bf\Omega_{\Lambda}}})\ge r\}$ is the set of all cosupports with corresponding corank $r$.
- $\aspace_\Lambda =\operatorname{span}^\perp(\OM_\Lambda) = \{ \Q_\Lambda{{\bfz}}, {{\bfz}}\in {\mathbb{R}^{\sdim}}\}$ is the subspace spanned by a cosparsity set $\Lambda$.
- $\A_{{{\bf\Omega}},\cosp} = \bigcup_{\Lambda \in \L_{{{\bf\Omega}},\cosp}}\aspace_\Lambda$ is the union of subspaces of $\cosp$-cosparse vectors and $\A_{{{\bf\Omega}},r}^{\text{corank}} = \bigcup_{\Lambda \in \L_{{{\bf\Omega}},r}^{\text{corank}}}\aspace_\Lambda$ is the union of subspaces of all vectors with corank $r$. In the case that every $\cosp$ rows of ${{\bf\Omega}}$ are independent it is clear that $\A_{{{\bf\Omega}},\cosp} = \A_{{{\bf\Omega}},r}^{\text{corank}}$. When it will be clear from the context, we will remove ${{\bf\Omega}}$ from the subscript.
- $\x \in {\mathbb{R}^{\sdim}}$ denotes the original unknown $\cosp$-cosparse vector and $\Lambda_{\x}$ its cosupport.
- ${{\bfv}},{{\bfu}} \in \A_\cosp$ are used to denote general $\cosp$-cosparse vectors and ${{\bfz}} \in {\mathbb{R}^{\sdim}}$ is used to denote a general vector.
$\OM$-RIP Definition and its Properties {#sec:omega_RIP}
=======================================
We now turn to define the $\OM$-RIP, which parallels the regular RIP as used in [@Candes06Near]. This property is a very important property for the analysis of the algorithms which holds for a large family of matrices ${{\bfM}}$ as we will see hereafter.
\[def:omega\_RIP\] A matrix ${{\bfM}}$ has the ${{\bf\Omega}}$-RIP property with a constant $\delta_\cosp$, if $\delta_\cosp$ is the smallest constant that satisfies $$\begin{aligned}
\label{eq:omega_RIP}
&& (1-\delta_{\cosp}){\left\Vert{{\bfv}}\right\Vert}_2^2 \le {\left\Vert{{\bfM}}{{\bfv}}\right\Vert}_2^2 \le (1+\delta_\cosp){\left\Vert{{\bfv}}\right\Vert}_2^2,\end{aligned}$$ whenever $\OM{{\bfv}}$ has at least $\cosp$ zeroes.
Note that though $\delta_\cosp$ is also a function of $\OM$ we abuse notation and use the same symbol for the $\OM$-RIP as the regular RIP. It will be clear from the context to which of them we refer and what $\OM$ is in use with the $\OM$-RIP. A similar property that looks at the corank of the vectors can be defined
\[def:omega\_RIP\_corank\] A matrix ${{\bfM}}$ has the corank-${{\bf\Omega}}$-RIP property with a constant $\delta_{r}^{\text{corank}}$, if $\delta_{r}^{\text{corank}}$ is the smallest constant that satisfies $$\begin{aligned}
\label{eq:omega_RIP_rank}
&& (1-\delta_{r}^{\text{corank}}){\left\Vert{{{\bfu}}}\right\Vert}_2^2 \le {\left\Vert{{\bfM}}{{{\bfu}}}\right\Vert}_2^2 \le (1+\delta_{r}^{\text{corank}}){\left\Vert{{{\bfu}}}\right\Vert}_2^2\end{aligned}$$ whenever the corank of ${{\bfu}}$ with respect to $\OM$ is greater or equal to $r$.
The $\OM$-RIP, like the regular RIP, inherits several key properties. We present only those related to $\delta_\cosp$, while very similar characteristics can be derived also for the corank-$\OM$-RIP. The first property we pose is an immediate corollary of the $\delta_\cosp$ definition.
\[cor:MQ\_RIP\_norm\] If ${{\bfM}}$ satisfies the ${{\bf\Omega}}$-RIP with a constant $\delta_{\cosp}$ then $$\begin{aligned}
\label{eq:MQ_RIP_norm}
&& {\left\Vert{{\bfM}}\Q_{\Lambda}\right\Vert}_2^2 \le 1+\delta_{\cosp}
$$ for any $\Lambda \in \L_{\cosp}$.
[*Proof:*]{} Any ${{\bfv}} \in \A_{\cosp}$ can be represented as ${{\bfv}} = \Q_{\Lambda}{{\bfz}}$ with $\Lambda \in \L_{\cosp}$ and ${{\bfz}} \in {\mathbb{R}^{\sdim}}$. Thus, the $\OM$-RIP in can be reformulated as $$\begin{aligned}
\label{eq:omega_RIP_Qz}
&& (1-\delta_{\cosp}){\left\Vert\Q_\Lambda{{\bfz}}\right\Vert}_2^2 \le {\left\Vert{{\bfM}}\Q_\Lambda{{\bfz}}\right\Vert}_2^2 \le (1+\delta_\cosp){\left\Vert\Q_\Lambda{{\bfz}}\right\Vert}_2^2\end{aligned}$$ for any ${{\bfz}} \in {\mathbb{R}^{\sdim}}$ and $\Lambda \in \L_{\cosp}$. Since $\Q_\Lambda$ is a projection ${\left\Vert\Q_\Lambda{{\bfz}}\right\Vert}_2^2 \le {\left\Vert{{\bfz}}\right\Vert}_2^2$. Combining this with the right inequality in gives $$\begin{aligned}
\label{eq:omega_RIP_Qz2}
{\left\Vert{{\bfM}}\Q_\Lambda{{\bfz}}\right\Vert}_2^2 \le (1+\delta_\cosp){\left\Vert{{\bfz}}\right\Vert}_2^2\end{aligned}$$ for any ${{\bfz}} \in {\mathbb{R}^{\sdim}}$ and $\Lambda \in \L_{\cosp}$. The first inequality in follows from by the definition of the spectral norm. $\Box$
\[lem:l\_inequality\] For $\tilde{\cosp} \le \cosp$ it holds that $\delta_{\cosp} \le \delta_{\tilde{\cosp}}$.
[*Proof:*]{} Since $\A_{\cosp} \subseteq \A_{\tilde{\cosp}}$ the claim is immediate. $\Box$
\[lem:omega\_RIP\_norm\] ${{\bfM}}$ satisfies the ${{\bf\Omega}}$-RIP if and only if $$\begin{aligned}
\label{eq:omega_RIP_norm}
&& {\left\Vert\Q_{\Lambda}({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\Lambda}\right\Vert}_2 \le \delta_\cosp
$$ for any $\Lambda \in \L_{\cosp}$.
[*Proof:*]{} The proof is similar to the one of the regular RIP as appears in [@foucart10Sparse]. As a first step we observe that Definition \[def:omega\_RIP\] is equivalent to requiring $$\begin{aligned}
{\left\vert{\left\Vert{{\bfM}}{{\bfv}}\right\Vert}_2^2 - {\left\Vert{{\bfv}}\right\Vert}_2^2\right\vert} \le \delta_{\cosp}{\left\Vert{{\bfv}}\right\Vert}_2^2\end{aligned}$$ for any ${{\bfv}} \in \A_{\cosp}$. The latter is equivalent to $$\begin{aligned}
{\left\vert{\left\Vert{{\bfM}}\Q_{\Lambda}{{{\bfz}}}\right\Vert}_2^2 - {\left\Vert\Q_{\Lambda}{{{\bfz}}}\right\Vert}_2^2\right\vert} \le \delta_{\cosp}{\left\Vert\Q_{\Lambda}{{{\bfz}}}\right\Vert}_2^2\end{aligned}$$ for any set $\Lambda \in \L_{\cosp}$ and any ${{{\bfz}}} \in {\mathbb R}^d$, since $\Q_{\Lambda}{{{\bfz}}} \in \A_{\cosp}$. Next we notice that $$\begin{aligned}
&& {\left\Vert{{\bfM}}\Q_{\Lambda}{{{\bfz}}}\right\Vert}_2^2 - {\left\Vert\Q_{\Lambda}{{{\bfz}}}\right\Vert}_2^2
= {{{\bfz}}}^*\Q_{\Lambda}{{\bfM}}^*{{\bfM}}\Q_{\Lambda}{{{\bfz}}} - {{{\bfz}}}^*\Q_{\Lambda}{{{\bfz}}}
= \langle\Q_{\Lambda}({{\bfM}}^*{{\bfM}} - {{\bfI}})\Q_{\Lambda}{{{\bfz}}}, {{{\bfz}}} \rangle.\end{aligned}$$ Since $\Q_{\Lambda}({{\bfM}}^*{{\bfM}} - {{\bfI}})\Q_{\Lambda}$ is Hermitian we have that $$\begin{aligned}
&&\max_{{{{\bfz}}}}\frac{{\left\vert\langle\Q_{\Lambda}({{\bfM}}^*{{\bfM}} - {{\bfI}})\Q_{\Lambda}{{{\bfz}}}, {{{\bfz}}} \rangle\right\vert}}{{\left\Vert{{{\bfz}}}\right\Vert}_2}
= {\left\Vert\Q_{\Lambda}({{\bfM}}^*{{\bfM}} - {{\bfI}})\Q_{\Lambda}\right\Vert}_2.\end{aligned}$$ Thus we have that Definition \[def:omega\_RIP\] is equivalent to for any set $\Lambda \in \L_{\cosp}$. $\Box$
\[cor:omega\_RIP\_norm\_diff\] If ${{\bfM}}$ satisfies the ${{\bf\Omega}}$-RIP then $$\begin{aligned}
\label{eq:omega_RIP_norm_diff}
&& {\left\Vert\Q_{\Lambda_1}({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\Lambda_2}\right\Vert}_2 \le \delta_\cosp,
$$ for any $\Lambda_1$ and $\Lambda_2$ such that $\Lambda_1 \cap \Lambda_2 \in \L_{\cosp}$.
[*Proof:*]{} Since $\Lambda_1 \cap \Lambda_2 \subseteq \Lambda_1$ and $\Lambda_1 \cap \Lambda_2 \subseteq \Lambda_2$ $$\begin{aligned}
\nonumber
{\left\Vert\Q_{\Lambda_1}({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\Lambda_2}\right\Vert}_2 \le {\left\Vert\Q_{\Lambda_2 \cap \Lambda_1}({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\Lambda_2 \cap \Lambda_1}\right\Vert}_2.\end{aligned}$$ Using Lemma \[lem:omega\_RIP\_norm\] completes the proof. $\Box$
As we will see later, we require the $\OM$-RIP to be small. Thus, we are interested to know for what matrices this hold true. In the synthesis case, where ${{\bf\Omega}}$ is unitary and the $\OM$-RIP is identical to the RIP, it was shown for certain family of random matrices, such as matrices with Bernoulli or Subgaussian ensembles, that for any value of $\epsilon_k$ if $m \ge C_{\epsilon_k} k \log(\frac{m}{k\epsilon_k})$ then $\delta_k \le \epsilon_k$ [@Candes06Near; @Rauhut08Compressed; @Mendelson08Uniform], where $\delta_k$ is the RIP constant and $C_{\epsilon_k}$ is a constant depending on $\epsilon_k$ and ${{\bfM}}$. A similar result for the same family of random matrices holds for the analysis case. The result is a special case of the result presented in [@Blumensath09Sampling].
\[thm:analysis\_RIP\_cond\] Let ${{\bfM}}\in {\mathbb{R}^{m \times d}}$ be a random matrix such that for any ${{\bfz}}\in {\mathbb{R}^{d}}$ and $0<\tilde{\epsilon} \le \frac{1}{3}$ it satisfies $$\begin{aligned}
P\left({\left\vert{\left\Vert{{\bfM}}{{\bfz}}\right\Vert}_2^2 - {\left\Vert{{\bfz}}\right\Vert}_2^2\right\vert} \ge \tilde{\epsilon}{\left\Vert{{\bfz}}\right\Vert}_2^2\right) \le e^{-\frac{C_{{{\bfM}}}m\tilde{\epsilon}}{2}},\end{aligned}$$ where $C_{{{\bfM}}} >0$ is a constant. For any value of $\epsilon_\cosp >0$, if $$\begin{aligned}
\label{eq:omega_RIP_m_size_cond}
m \ge \frac{32}{C_M\epsilon_r^2}\left( \log({\left\vert\L_{r}^{\text{corank}}\right\vert}) + (d-r)\log({9}/{\epsilon_r})+t\right),\end{aligned}$$ then $\delta_{r}^{\text{corank}} \le \epsilon_r$ with probability exceeding $1-e^{-t}$.
The above theorem is important since it shows that the $\OM$-RIP holds with a small constant for a large family of matrices – the same family that satisfy the RIP property. In a recent work it was even shown that by randomizing the signs of the columns in the matrices that satisfy the RIP we get new matrices that also satisfy the RIP [@Krahmer11New]. Thus, requiring the $\OM$-RIP constant to be small, as will be done hereafter, is legitimate.
For completeness we present a proof for theorem \[thm:analysis\_RIP\_cond\] in \[sec:analysis\_RIP\_proof\] based on [@Rauhut08Compressed; @Mendelson08Uniform; @Baraniuk08Simple]. We include in it also the proof of Theorem \[thm:analysis\_RIP\_cond\_dependencies\] to follow. In the case that ${{\bf\Omega}}$ is in general position ${\left\vert\L_{r}^{\text{corank}}\right\vert} = {p \choose r} \le (\frac{ep}{p-r})^{p-r}$ (inequality is by Stirling’s formula) and thus $m \ge (p-r)\log(\frac{ep}{p-r})$. Since we want $m$ to be smaller than $d$ we need $p -\cosp$ to be smaller than $d$. This limits the size of $p$ for $\OM$ since $r$ cannot be greater than $d$. Thus, we present a variation of the theorem which states the results in terms of $\delta_{\cosp}$ and $\cosp$ instead of $\delta_{r}^{\text{corank}}$ and $r$. The following theorem is also important because of the fact that our theoretical results are in terms of $\delta_{\cosp}$ and not $\delta_{r}^{\text{corank}}$. It shows that $\delta_{\cosp}$ is small in the same family of matrices that guarantees $\delta_{r}^{\text{corank}}$ to be small.
\[thm:analysis\_RIP\_cond\_dependencies\] Under the same setup of Theorem \[thm:analysis\_RIP\_cond\], for any $\epsilon_\cosp > 0$ if $$\begin{aligned}
\label{eq:omega_RIP_m_size_cond_dependencies}
m \ge \frac{32}{C_M\epsilon_\cosp^2}\left( (p-\cosp)\log\left(\frac{9p}{(p-\cosp)\epsilon_\cosp}\right)+t \right),\end{aligned}$$ then $\delta_\cosp \le \epsilon_\cosp$ with probability exceeding $1-e^{-t}$.
Remark that when $\OM$ is in general position $\cosp$ cannot be greater than $d$ and thus $p$ cannot be greater than $2d$ [@Nam12Cosparse]. For this reason, if we want to have large values for $p$ we should allow linear dependencies between the rows of $\OM$. In this case the cosparsity of the signal can be greater than $d$. This explains why linear dependencies are a favorable thing in analysis dictionaries [@Rubinstein12Cosparse]. In Section \[sec:exp\] we shall see that also empirically we get a better recovery when $\OM$ contains linear dependencies.
Near Optimal Projection {#sec:near_opt_proj}
=======================
As we will see hereafter, in the proposed algorithms we will face the following problem: Given a general vector ${{\bfz}} \in {\mathbb{R}^{\sdim}}$, we would like to find an $\cosp$-cosparse vector that is closest to it in the $\ell_2$-norm sense. In other words, we would like to project the vector to the closest $\cosp$-cosparse subspace. Given the cosupport $\Lambda$ of this space the solution is simply $\Q_\Lambda{{\bfz}}$. Thus, the problem of finding the closest $\cosp$-cosparse vector turns to be the problem of finding the cosupport of the closest $\cosp$-cosparse subspace. We denote the procedure of finding this cosupport by $$\begin{aligned}
\label{eq:optimal_cosparse_projection}
\CF^*_{\cosp}({{\bfz}}) = {\operatornamewithlimits{argmin}}_{\Lambda \in \L_{\cosp}} {\left\Vert{{\bfz}} - \Q_{\Lambda}{{\bfz}}\right\Vert}_2^2.\end{aligned}$$ In the representation domain in the synthesis case, the support of the closest $k$-sparse subspace is found simply by hard thresholding, i.e., taking the support of the $k$-largest elements. However, in the analysis case calculating is NP-complete with no efficient method for doing it for a general $\OM$ [@Gribonval12Projection]. Thus an approximation procedure $\hat{\CF}_\cosp$ is needed. For this purpose we introduce the definition of a near-optimal projection [@Giryes11Iterative].
\[def:C\_optimal\_proj\] A procedure $\hat{\CF}_\cosp$ implies a near-optimal projection $\Q_{\hat{\CF}_\cosp(\cdot)}$ with a constant $C_\cosp$ if for any ${{\bfz}} \in {\mathbb R}^d$ $$\begin{aligned}
\label{eq:C_optimal_proj}
&& {\left\Vert{{\bfz}}-\Q_{\hat\CF_\cosp({{\bfz}})}{{\bfz}}\right\Vert}_2^2 \le C_\cosp{\left\Vert{{\bfz}} - \Q_{\CF^*_\cosp({{\bfz}})}{{\bfz}}\right\Vert}_2^2.\end{aligned}$$
A clear implication of this definition is that if $\hat\CF_\cosp$ implies a near-optimal projection with a constant $C_\cosp$ then for any vector ${{\bfz}}\in {\mathbb R}^d$ and an $\cosp$-cosparse vector ${{\bfv}} \in {\mathbb R}^d$ $$\begin{aligned}
\label{eq:C_optimal_ineq}
&& {\left\Vert{{\bfz}}-\Q_{\hat\CF_\cosp({{\bfz}})}{{{\bfz}}}\right\Vert}_2^2 \le C_\cosp {\left\Vert{{\bfz}}-{{\bfv}}\right\Vert}_2^2.\end{aligned}$$
Similarly to the $\OM$-RIP, the above discussion can be directed also for finding the closest vector with corank $r$ defining $\CF_{r}^{\text{corank}*}$ and near optimal projection for this case in a very similar way to and Definition \[def:C\_optimal\_proj\] respectively.
Having a near-optimal cosupport selection scheme for a general operator is still an open problem and we leave it for a future work. It is possible that this is also NP-complete. We start by describing a simple thresholding rule that can be used with any operator. Even though it does not have any known (near) optimality guarantee besides the case of unitary operators, the numerical section will show it performs well in practice. Then we present two tractable algorithms for finding the optimal cosupport for two non-trivial analysis operators, the one dimensional finite difference operator $\OM_{\text{1D-DIF}}$ [@Han04Optimal] and the fused Lasso operator $\OM_{\text{FUS}}$ [@Tibshirani05Sparsity]. Later in the paper, we propose theoretical guarantees for algorithms that use operators that has an optimal or a near-optimal cosupport selection scheme. We leave the theoretical study of the thresholding technique for a future work but demonstrate its performance empirically in Section \[sec:exp\] where this rule is used showing that also when near-optimality is not at hand reconstruction is feasible.
Cosupport Selection by Thresholding {#sec:thresholding}
-----------------------------------
One intuitive option for cosupport selection is the simple thresholding $$\begin{aligned}
\label{eq:thresh_cosupp_selection}
\hat\CF_\cosp({{\bfz}}) = \operatorname{cosupp}(\OM{{\bfz}},\cosp),\end{aligned}$$ which selects as a cosupport the indices of the $\cosp$-smallest elements after applying $\OM$ on $\z$. As mentioned above, this selection method is optimal for unitary analysis operators where it coincides with the hard thresholding used in synthesis. However, in the general case this selection method is not guaranteed to give the optimal cosupport. Its near optimality constant $C_\cosp$ is not close to one and is equal to the fraction of the largest and smallest eigenvalues (which are not zero) of the submatrices composed of $\cosp$ rows from $\OM$ [@Giryes11Iterative].
One example for an operator for which the thresholding is sub-optimal is the 1D-finite difference operator $\OM_{\text{1D-DIF}}$. This operator is defined as: $$\label{1D Diff}
\OM_{\text{1D-DIF}} = \begin{pmatrix}
-1 & 1 &\cdots &~ &~ & ~ \\
\vdots &-1 & 1 & ~&~ & ~ \\
~ &~ &~ &\ddots &~ & ~ \\
~ &~ &~ &~ &-1&1
\end{pmatrix}$$ In this case, given a signal ${{\bfz}}$, applying $\OM_{\text{1D-DIF}}$ on it, result with a vector of coefficients that represents the differences in the signal. The thresholding selection method will select the indices of the $\cosp$ smallest elements in $\OM_{}{{\bfz}}$ as the cosupport $\Lambda_{{\bfz}}$. For example, for the signal ${{\bfz}}\in {\mathbb{R}^{201}}$ in Fig \[fig:z\_sig\] that contains $100$ times one, $100$ times minus one and $1.5$ as the last element, the thresholding will select the cosupport to be the first $199$ coefficients in $\OM_{\text{1D-DIF}}{{\bfz}}$ that appears in Fig \[fig:omega\_z\_sig\] and thus the projected vector will be the one in Fig \[fig:z\_thresh\_proj\]. Its error in the $\ell_2$-norm sense is $\sqrt{200}$. However, selecting the cosupport to be the first $99$ elements and last $100$ elements result with the projected vector in Fig. \[fig:z\_opt\_proj\], which has a smaller projection error ($2.5$). Thus, it is clear that the thresholding is sub-optimal for $\OM_{\text{1D-DIF}}$. In a similar way it is also sub-optimal for the 2D-finite difference operator $\OM_{\text{2D-DIF}}$ that returns the vertical and horizontal differences of a two dimensional signal. Though not optimal, the use of thresholding with this operator is illustrated in Section \[sec:exp\] demonstrating that also when a good projection is not at hand, good reconstruction is still possible.
[ ]{}
Optimal Analysis Projection Operators
-------------------------------------
As mentioned above, in general it would appear that determining the optimal projection is computationally difficult with the only general solution being to fully enumerate the projections onto all possible cosupports. Here we highlight two cases where it is relatively easy (polynomial complexity) to calculate the optimal cosparse projection.
### Case 1: 1D finite difference
For the 1D finite difference operator the analysis operator is not redundant ($\pdim = \sdim-1$) but neither is it invertible. As we have seen, a simple thresholding does not provide us with the optimal cosparse projection. Thus, in order to determine the best $\cosp$-cosparse approximation for a given vector ${{\bfz}}$ we take another route and note that we are looking for the closest (in the $\ell_2$-norm sense to ${{\bfz}}$) piecewise constant vector with $\pdim-\cosp$ change-points. This problem has been solved previously in the signal processing literature using dynamic programming (DP), see for example: [@Han04Optimal]. Thus for this operator it is possible to calculate the best cosparse representation in ${\cal
O}(\sdim^2)$ operations. The existence of a DP solution follows from the ordered localized nature of the finite difference operator. To the best of our knowledge, there is no known extension to 2D finite difference.
### Case 2: Fused Lasso Operator
A redundant operator related to the 1D finite difference operator is the so-called fused Lasso operator, usually used with the analysis $\ell_1$-minimization [@Tibshirani05Sparsity]. This usually takes the form:
$$\label{fusion operator}
\OM_{\text{FUS}} = \begin{pmatrix} \OM_{\text{1D-DIF}} \\ \epsilon {{\bfI}}
\end{pmatrix}.$$
Like $\OM_{\text{1D-DIF}}$ this operator works locally and therefore we can expect to derive a DP solution to the approximation problem. This is presented below.
Note that in terms of the cosparsity model the $\epsilon$ parameter plays no role. This is in contrast to the traditional convex optimization solutions where the value of $\epsilon$ is pivotal [@Vaiter12Robust]. It is possible to mimic the $\epsilon$ dependence within the cosparsity framework by considering a generalized fused Lasso operator of the form: $$\label{fusion operator_epsilon}
\OM_{\epsilon {\text{FUS}}} = \begin{pmatrix} \OM_{\text{1D-DIF}}\\ \OM_{\text{1D-DIF}}\\
\vdots \\
\OM_{\text{1D-DIF}} \\ {{\bfI}}
\end{pmatrix}.$$ where the number of repetitions of the $\OM_{\text{1D-DIF}}$ operator (and possibly the ${{\bfI}}$ operator) can be selected to mimic a weight on the number of nonzero coefficients of each type. For simplicity we only consider the case indicated by
### A recursive solution to the optimal projector for $\OM_{\text{FUS}}$
Rather than working directly with the operator $\OM_{\text{FUS}}$ we make use of the following observation. An $\cosp$-cosparse vector ${{\bfv}}$ (or $k$-sparse vector) for $\OM_{\text{FUS}}$ is a piecewise constant vector with $k_1$ change points and $k_2$ non-zero entries such that $k_1 + k_2 = k = p-\cosp$, where $p=2d-1$. To understand better the relation between $k_1$ and $k_2$, notice that $k_1=0$ implies equality of all entries, so $k_2 = 0$ or $d$, hence $\cosp = p$ or $d-1$. Conversely, considering $d\le\cosp<p$ or $0\le\cosp<d-1$ implies $k_1 \neq 0$. It also implies that there is at least one nonzero value, hence $k_2 \neq 0$.
Thus, an $\cosp$-cosparse vector ${{\bfv}}$ for $\OM_{\text{FUS}}$ can be parameterized in terms of a set of change points, $\{n_i\}_{i=0:k_1+1}$, and a set of constants, $\{\mu_i\}_{i=1:k_1+1}$, such that: $${{\bfv}}_j = \mu_i, n_{i-1} < j \leq n_i$$ with the convention that $n_0 = 0$ and $n_{k_1+1} = d$, unless stated otherwise. We will also make use of the indicator vector, $\s$, defined as: $$\s_i =
\begin{cases}
0& \text{if $\mu_i = 0$},\\
1& \text{otherwise}
\end{cases} \text{for $1\le i \le k_1+1$}.$$ Using this alternative parametrization we can write the minimum distance between a vector ${{\bfz}}$ and the set of $k$-sparse fused Lasso coefficients as: $$\begin{split}
F_k({{\bfz}}) &= \min_{1\leq k_1 \leq k} \min_{\substack{
\{n_i\}_{i=1:k_1}\\
\{\mu_i\}_{i=1:k_1+1}\\
n_{k_1}<d}}
\sum_{i=1}^{k_1+1} \sum_{j = n_{i-1}+1}^{n_i}({{\bfz}}_j-\mu_i)^2,\\
&\mbox{~subject to~} \sum_{i = 1}^{k_1+1} \s_i (n_i-n_{i-1}) = k-k_1
\end{split}$$ Although this looks a formidable optimization task we now show that it can be computed recursively through a standard DP strategy, modifying the arguments in [@Han04Optimal].
Let us define the optimal cost, $I_k(L,\omega,k_1)$, for the vector $[{{\bfz}}_1, \ldots ,
{{\bfz}}_L]^T$ with $k_1$ change points and $\s_{k_1+1} = \omega$, as: $$\begin{split}
I_k(L,\omega,k_1) &= \min_{\substack{
\{n_i\}_{i=1:k_1}\\
\{\s_i\}_{i=1:k_1+1}\\
n_{k_1} <L,~ n_{k_1+1} =L\\
\s_{k_1+1}=\omega}}
\sum_{i=1}^{k_1+1} \sum_{j = n_{i-1}+1}^{n_i}({{\bfz}}_j-\mu_i)^2,\\
&\mbox{~subject to~} \sum_{i = 1}^{k_1+1} \s_i (n_i-n_{i-1}) = k-k_1\\
&\mbox{~and~} \mu_i = \frac{\s_i}{n_i-n_{i-1}}
\sum_{l=n_{i-1}+1}^{n_i} {{\bfz}}_l
\end{split}$$ where we have set $\mu_i$ to the optimal sample means. Notice that calculating $I_k(L,\omega,k_1)$ is easy for $k_1\le k \le 1$. Thus, we calculate it recursively considering two separate scenarios:
Case 1: $\omega = 0$
: where the last block of coefficients are zero. This gives: $$\begin{split}
I_k(L,0,k_1) = & \min_{n_{k_1<L}}\left( \sum_{j = n_{k_1}+1}^L ({{\bfz}}_j)^2 +
\min_{\substack{\{n_i\}_{i=1:k_1-1}\\
\{\s_i\}_{i=1:k_1-1}\\
n_{k_1-1}<n_{k_1}\\
\s_{k_1} = 1}} \sum_{i=1}^{k_1} \sum_{j =
n_{i-1}+1}^{n_i}({{\bfz}}_j-\mu_i)^2 \right),\\
~&\mbox{~subject to~} \sum_{i = 1}^{k_1} s_i (n_i-n_{i-1}) = (k-1)-(k_1-1)\\
~&\mbox{~and~} \mu_i = \frac{\s_i}{n_i-n_{i-1}} \sum_{l=n_{i-1}+1}^{n_i} {{\bfz}}_l,
\end{split}$$ (noting that if $\s_{k_1+1} = 0$ then $\s_{k_1} = 1$ since otherwise $n_{k_1}$ would not have been a change point). This simplifies to the recursive formula: $$\label{eq:recusive1}
I_k(L,0,k_1) = \min_{n_{k_1}< L}\left( \sum_{j = n_{k_1}+1}^L ({{\bfz}}_j)^2 +
I_{k-1}(n_{k_1},1,k_1-1)\right)$$
Case 2: $\omega = 1$
: when the final block of coefficients are non-zero we have: $$\begin{split}
I_k(L,1,k_1) = & \min_{
\substack{n_{k_1}<L \\n_{k_1+1}=L\\ \s_{k_1}} }
\left( \sum_{j = n_{k_1}+1}^L ({{\bfz}}_j-\mu_{k_1+1})^2 +
\min_{\substack{\{n_i\}_{i=1:k_1-1}\\
\{\s_i\}_{i=1:k_1-1}\\
n_{k_1-1}<n_{k_1}}}
\sum_{i=1}^{k_1} \sum_{j = n_{i-1}+1}^{n_i}({{\bfz}}_j-\mu_i)^2 \right),\\
~&\mbox{~subject to~} \sum_{i = 1}^{k_1} \s_i (n_i-n_{i-1}) = (k-L+n_{k_1}-1)-(k_1-1)\\
~&\mbox{~and~} \mu_i = \frac{\s_i}{n_i-n_{i-1}} \sum_{l=n_{i-1}+1}^{n_i} {{\bfz}}_l.
\end{split}$$ This simplifies to the recursive relationship: $$\label{eq:recursive2}
\begin{split}
I_k(L,1,k_1) &= \min_{\substack{n_{k_1}<L\\ \s_{k_1}} } \left(\sum_{j = n_{k_1}+1}^L ({{\bfz}}_j-\mu_{k_1+1})^2 +
I_{k-L+n_{k_1}-1}(n_{k_1},\s_{k_1},k_1-1) \right)\\
&\text{subject to~} \mu_{k_1+1} = \sum_{l=n_{k_1}+1}^{L} {{\bfz}}_l/\big(L-n_{k_1}\big)
\end{split}$$
Equations and are sufficient to enable the calculation of the optimal projection in polynomial time,starting with $k_1\le k\le 1$ and recursively evaluating the costs for $k\ge k_1 \ge 1$. Finally, we have $F_k({{\bfz}}) = \min_{k_1\le k, \omega \in \{0,1\}} I_k(d,\omega,k_1)$. The implementation details are left as an exercise for the reader.
New Analysis algorithms {#sec:analysis_alg}
=======================
Quick Review of the Greedy-Like Methods
---------------------------------------
Before we turn to present the analysis versions of the greedy-like techniques we recall their synthesis versions. These use a prior knowledge about the cardinality $k$ and actually aim at approximating a variant of $$\label{eq:synthesisL0_k}
{\operatornamewithlimits{argmin}}_{{{\boldsymbol{\alpha}}}} {\left\Vert\y - \M \Dict {{\boldsymbol{\alpha}}}\right\Vert}_2^2 \subjectto {\left\Vert{{\boldsymbol{\alpha}}}\right\Vert}_0 \le k.
$$ For simplicity we shall present the greedy-like pursuits for the case ${{\bfD}} = {{\bfI}}$. In the general case ${{\bfM}}$ should be replaced with ${{\bfM}}{{\bfD}}$, ${{\bfx}}$ with ${{\boldsymbol{\alpha}}}$ and the reconstruction result should be $\hat\x = {{\bfD}}\hat{{\boldsymbol{\alpha}}}$. In addition, in the algorithms’ description we do not specify the stopping criterion. Any standard stopping criterion, like residual’s size or relative iteration change, can be used. More details can be found in [@Needell09CoSaMP; @Dai09Subspace].
[*IHT and HTP:*]{} IHT [@Blumensath09Iterative] and HTP [@Foucart11Hard] are presented in Algorithm \[alg:IHT\_HTP\]. Each IHT iteration is composed of two basic steps. The first is a gradient step, with a step size $\mu_t$, in the direction of minimizing ${\left\Vert\y - \M \x\right\Vert}_2^2$. The step size can be either constant in all iterations ($\mu^t = \mu$) or changing [@Kyrillidis11Recipes]. The result vector $\x_g$ is not guaranteed to be sparse and thus the second step of IHT projects $\x_g$ to the closest $k$-sparse subspace by keeping its largest $k$ elements. The HTP takes a different strategy in the projection step. Instead of using a simple projection to the closest $k$-sparse subspace, HTP selects the vector in this subspace that minimizes ${\left\Vert\y - \M \x\right\Vert}_2^2$ [@Foucart11Hard; @Blumensath12Accelerated].
$k, {{\bfM}}, {{\bfy}}$ where ${{\bfy}} = {{\bfM}}{{\bfx}}
+ {{\bfe}}$, $k$ is the cardinality of ${{\bfx}}$ and ${{\bfe}}$ is an additive noise.
$\hat{{{\bfx}}}_{\text{\tiny IHT}}$ or $\hat{{{\bfx}}}_{\text{\tiny HTP}}$: $k$-sparse approximation of ${{\bfx}}$.
Initialize representation $\hat{\x}^0 = {{\bf0}}$ and set $t = 0$.
$t = t + 1$.
Perform a gradient step: ${{\bfx}}_g = \hat{{{\bfx}}}^{t-1} + \mu^t \M^*(\y - \M\hat{\x}^{t-1})$
Find a new support: $T^t = \operatorname{supp}(\x_g,k)$
Calculate a new representation: $\hat{{{\bfx}}}_{\text{\tiny IHT}}^t = ({{\bfx}}_g)_{T^t}$ for IHT, and $\hat{{{\bfx}}}_{\text{\tiny HTP}}^t = {{\bfM}}^{\dag}_{T^t}{{\bfy}}$ for HTP.
Form the final solution $\hat{{{\bfx}}}_{\text{\tiny IHT}} = \hat{{{\bfx}}}_{\text{\tiny IHT}}^t$ for IHT and $\hat{{{\bfx}}}_{\text{\tiny HTP}} = \hat{{{\bfx}}}_{\text{\tiny HTP}}^t$ for HTP.
[*CoSaMP and SP:*]{} CoSaMP [@Needell09CoSaMP] and SP [@Dai09Subspace] are presented in Algorithm \[alg:CoSaMP\_SP\]. The difference between these two techniques is similar to the difference between IHT and HTP. Unlike IHT and HTP, the estimate for the support of $\x$ in each CoSaMP and SP iteration is computed by observing the residual ${{\bfy}}_{\text{resid}}^t = {{\bfy}} - \M\x^t$. In each iteration, CoSaMP and SP extract new support indices from the residual by taking the indices of the largest elements in ${{\bfM}}^*{{\bfy}}_{\text{resid}}^t$. They add the new indices to the estimated support set from the previous iteration creating a new estimated support $\tilde{T}^t$ with cardinality larger than $k$. Having the updated support, in a similar way to the projection in HTP, an objective aware projection is performed resulting with an estimate ${{\bfw}}$ for $\x$ that is supported on $\tilde{T}^t$. Since we know that $\x$ is $k$-sparse we want to project ${{\bfw}}$ to a $k$-sparse subspace. CoSaMP does it by simple hard thresholding like in IHT. SP does it by an objective aware projection similar to HTP.
$k, {{\bfM}}, {{\bfy}}$ where ${{\bfy}} = {{\bfM}}{{\bfx}}
+ {{\bfe}}$, $k$ is the cardinality of ${{\bfx}}$ and ${{\bfe}}$ is an additive noise. $a = 1$ (SP), $a = 2$ (CoSaMP).
$\hat{{{\bfx}}}_{\text{\tiny CoSaMP}}$ or $\hat{{{\bfx}}}_{\text{\tiny SP}}$: $k$-sparse approximation of ${{\bfx}}$.
Initialize the support $T^0 =\emptyset$, the residual ${{\bfy}}_{\text{resid}}^0 = {{\bfy}}$ and set $t = 0$.
$t = t + 1$.
Find new support elements: $T_\Delta =
\operatorname{supp}({{\bfM}}^*{{\bfy}}^{t - 1}_{\text{resid}},a k)$.
Update the support: $\tilde{T}^t = T^{t -1} \cup
T_\Delta$.
Compute a temporary representation: ${{\bfw}} =
{{\bfM}}^{\dag}_{\tilde{T}^t}{{\bfy}}$.
Prune small entries: $T^t =
\operatorname{supp}({{\bfw}},k)$.
Calculate a new representation: $\hat{{{\bfx}}}_{\text{\tiny CoSaMP}}^t = {{\bfw}}_{T^t}$ for CoSaMP, and $\hat{{{\bfx}}}_{\text{\tiny SP}}^t = {{\bfM}}^{\dag}_{T^t}{{\bfy}}$ for SP.
Update the residual: ${{\bfy}}_{\text{resid}}^t = {{\bfy}} - {{\bfM}}\hat{{{\bfx}}}_{\text{\tiny CoSaMP}}^t$ for CoSaMP, and ${{\bfy}}_{\text{resid}}^t = {{\bfy}} -{{\bfM}}\hat{{{\bfx}}}_{\text{\tiny SP}}^t$ for SP.
Form the final solution $\hat{{{\bfx}}}_{\text{\tiny CoSaMP}} = \hat{{{\bfx}}}_{\text{\tiny CoSaMP}}^t$ for CoSaMP and $\hat{{{\bfx}}}_{\text{\tiny SP}} = \hat{{{\bfx}}}_{\text{\tiny SP}}^t$ for SP.
Analysis greedy-like methods
----------------------------
[||C[3.6cm]{}|C[3.5cm]{}|| C[4cm]{}| C[3.5cm]{}||]{} [**Synthesis operation name**]{} & [**Synthesis operation**]{} & [**Analysis operation name** ]{} & [**Analysis operation** ]{}\
Support selection & Largest $k$ elements: $T=\operatorname{supp}(\cdot, k)$ & Cosupport selection & Using a near optimal projection: $\Lambda = \hat\CF_\cosp(\cdot)$\
Orthogonal Projection of ${{\bfz}}$ to a $k$-sparse subspace with support $T$ & ${{\bfz}}_T$ & Orthogonal projection of ${{\bfz}}$ to an $\cosp$-cosparse subspace with cosupport $\Lambda$ & $\Q_{\Lambda}{{\bfz}}$\
Objective aware projection to a $k$-sparse subspace with support $T$ & $\M_T^\dag \y$ = ${\operatornamewithlimits{argmin}}_{{{\bfv}}}{\left\Vert\y - \M{{\bfv}}\right\Vert}_2^2$ s.t. ${{\bfv}}_{T^C} = 0$ & Objective aware projection to an $\cosp$-cosparse subspace with cosupport $\Lambda$ & ${\operatornamewithlimits{argmin}}_{{{\bfv}}}{\left\Vert\y - \M{{\bfv}}\right\Vert}_2^2$ s.t. $\OM_{\Lambda}{{\bfv}} = 0$\
Support of ${{\bfv}}_1 + {{\bfv}}_2$ where $\operatorname{supp}({{\bfv}}_1)= T_1$ and $\operatorname{supp}({{\bfv}}_2)= T_2$ & $\operatorname{supp}({{\bfv}}_1+{{\bfv}}_2)\subseteq T_1 \cup T_2$ & Cosupport of ${{\bfv}}_1 + {{\bfv}}_2$ where $\operatorname{cosupp}({{\bfv}}_1)= \Lambda_1$ and $\operatorname{cosupp}({{\bfv}}_2)= \Lambda_2$ & $\operatorname{cosupp}({{\bfv}}_1+{{\bfv}}_2) \supseteq \Lambda_1 \cap \Lambda_2$\
Maximal size of $T_1 \cup T_2$ where ${\left\vertT_1\right\vert} \le k_1$ and ${\left\vertT_2\right\vert} \le k_2$ & ${\left\vertT_1 \cup T_2\right\vert} \le k_1+k_2$ & Minimal size of $\Lambda_1 \cap \Lambda_2$ where ${\left\vert\Lambda_1\right\vert} \ge \cosp_1$ and ${\left\vert\Lambda_2\right\vert} \ge \cosp_2$ & ${\left\vert\Lambda_1 \cap \Lambda_2\right\vert} \ge \cosp_1 + \cosp_2 - p$\
Given the synthesis greedy-like pursuits, we would like to define their analysis counterparts. For this task we need to ’translate’ each synthesis operation into an analysis one. This gives us a general recipe for converting algorithms between the two schemes. The parallel lines between the schemes are presented in Table \[tbl:Synthesis\_Analysis\_parallels\]. Those become more intuitive and clear when we keep in mind that while the synthesis approach focuses on the non-zeros, the analysis concentrates on the zeros.
For clarity we dwell a bit more on the equivalences. For the cosupport selection, as mentioned in Section \[sec:near\_opt\_proj\], computing the optimal cosupport is a combinatorial problem and thus the approximation $\hat\CF_\cosp$ is used. Having a selected cosupport $\Lambda$, the projection to its corresponding cosparse subspace becomes trivial, given by $\Q_\Lambda$.
Given two vectors ${{\bfv}}_1 \in \A_{\cosp_1}$ and ${{\bfv}}_2 \in \A_{\cosp_2}$ such that $\Lambda_1 = \operatorname{cosupp}(\OM{{\bfv}}_1)$ and $\Lambda_2 = \operatorname{cosupp}(\OM{{\bfv}}_2)$, we know that ${\left\vert\Lambda_1\right\vert} \ge \cosp_1$ and ${\left\vert\Lambda_2\right\vert} \ge \cosp_2$. Denoting $T_1 = \operatorname{supp}(\OM{{\bfv}}_1)$ and $T_2 = \operatorname{supp}(\OM{{\bfv}}_2)$ it is clear that $\operatorname{supp}(\OM({{\bfv}}_1 + {{\bfv}}_1)) \subseteq T_1 \cup T_2$. Noticing that $\operatorname{supp}(\cdot) =\operatorname{cosupp}(\cdot)^C$ it is clear that ${\left\vertT_1\right\vert}\le p-\cosp_1$, ${\left\vertT_2\right\vert}\le p-\cosp_2$ and $\operatorname{cosupp}(\OM({{\bfv}}_1 + {{\bfv}}_2)) \supseteq (T_1 \cup T_2)^C =T_1^C \cap T_2^C = \Lambda_1 \cap \Lambda_2$. From the last equality we can also deduce that ${\left\vert\Lambda_1 \cap \Lambda_2\right\vert} = p -{\left\vertT_1 \cup T_2\right\vert} \ge p - (p-\cosp_1) - (p-\cosp_2) = \cosp_1 + \cosp_2 - p$.
With the above observations we can develop the analysis versions of the greedy-like algorithms. As in the synthesis case, we do not specify a stopping criterion. Any stopping criterion used for the synthesis versions can be used also for the analysis ones.
$\cosp, {{\bfM}}, {{\bf\Omega}}, {{\bfy}}$ where ${{\bfy}} = {{\bfM}}{{\bfx}}
+ {{\bfe}}$, $\cosp$ is the cosparsity of ${{\bfx}}$ under ${{\bf\Omega}}$ and ${{\bfe}}$ is the additive noise.
$\hat{{{\bfx}}}_{\text{\tiny AIHT}}$ or $\hat{{{\bfx}}}_{\text{\tiny AHTP}}$: $\cosp$-cosparse approximation of ${{\bfx}}$.
Initialize estimate $\hat{\x}^0 = {{\bf0}}$ and set $t = 0$.
$t = t + 1$.
Perform a gradient step: ${{\bfx}}_g = \hat{{{\bfx}}}^{t-1} + \mu^t \M^*(\y - \M\hat{\x}^{t-1})$
Find a new cosupport: $\hat\Lambda^t = \hat{\CF}_\cosp(\x_g)$
Calculate a new estimate: $\hat{{{\bfx}}}_{\text{\tiny AIHT}}^t = \Q_{\hat\Lambda^t}{{\bfx}}_g$ for AIHT, and $\hat{{{\bfx}}}_{\text{\tiny AHTP}}^t = {\operatornamewithlimits{argmin}}_{\tilde{{{\bfx}}}}{\left\Vert\y - \M\tilde{{{\bfx}}}\right\Vert}_2^2$ s.t. $\OM_{\hat\Lambda^t}\tilde{{{\bfx}}} = 0$ for AHTP.
Form the final solution $\hat{{{\bfx}}}_{\text{\tiny AIHT}} = \hat{{{\bfx}}}_{\text{\tiny AIHT}}^t$ for AIHT and $\hat{{{\bfx}}}_{\text{\tiny AHTP}} = \hat{{{\bfx}}}_{\text{\tiny AHTP}}^t$ for AHTP.
[*AIHT and AHTP:*]{} Analysis IHT (AIHT) and analysis HTP (AHTP) are presented in Algorithm \[alg:Analysis\_IHT\_HTP\]. As in the synthesis case, the choice of the gradient stepsize $\mu^t$ is crucial: If $\mu^t$’s are chosen too small, the algorithm gets stuck at a wrong solution and if too large, the algorithm diverges. We consider two options for $\mu^t$.
In the first we choose $\mu^t = \mu$ for some constant $\mu$ for all iterations. A theoretical discussion on how to choose $\mu$ properly is given in Section \[sec:AIHT\_AHTP\_guarantees\].
The second option is to select a different $\mu$ in each iteration. One way for doing it is to choose an ‘optimal’ stepsize $\mu^t$ by solving the following problem $$\label{eq:muOptimalM}
\mu^t := {\operatornamewithlimits{argmin}}_\mu {\left\Vert\y - \M\hat\x^t \right\Vert}_2^2.$$ Since $\hat\Lambda^t = \hat\CF_\cosp(\hat\x^{t-1} + \mu^t\M^*(\y-\M\hat\x^{t-1}))$ and $\hat\x^t = \Q_{\hat\Lambda^t}(\x_g)$, the above requires a line search over different values of $\mu$ and along the search $\hat\Lambda^t$ might change several times. A simpler way is an adaptive step size selection as proposed in [@Kyrillidis11Recipes] for IHT. In a heuristical way we limit the search to the cosupport $\tilde{\Lambda} = \hat{\CF}_\cosp(\M^*(\y-\M\hat\x^{t-1})) \cap \hat{\Lambda}^{t-1}$. This is the intersection of the cosupport of $\hat\x^{t-1}$ with the $\cosp$-cosparse cosupport of the estimated closest $\cosp$-cosparse subspace to $\M^*(\y-\M\hat\x^{t-1})$. Since $\hat\x^{t-1} = \Q_{\tilde\Lambda} \hat\x^{t-1}$, finding $\mu$ turns to be $$\label{eq:muOptimalM_approx}
\mu^t := {\operatornamewithlimits{argmin}}_\mu {\left\Vert\y - \M(\hat\x^{t-1} + \mu\Q_{\tilde{\Lambda}}\M^*(\y-\M\hat\x^{t-1}))\right\Vert}_2^2.$$ This procedure of selecting $\mu^t$ does not require a line search and it has a simple closed form solution.
To summarize, there are three main options for the step size selection:
- Constant step-size selection – uses a constant step size $\mu^t = \mu$ in all iterations.
- Optimal changing step-size selection – uses different values for $\mu^t$ in each iterations by minimizing ${\left\Vert{{\bfy}} -\M\hat\x^t\right\Vert}_2$.
- Adaptive changing step-size selection – uses .
$\cosp, {{\bfM}}, {{\bf\Omega}}, {{\bfy}}, a$ where ${{\bfy}} = {{\bfM}}{{\bfx}}
+ {{\bfe}}$, $\cosp$ is the cosparsity of ${{\bfx}}$ under ${{\bf\Omega}}$ and ${{\bfe}}$ is the additive noise.
$\hat{{{\bfx}}}_{\text{\tiny ACoSaMP}}$ or $\hat{{{\bfx}}}_{\text{\tiny ASP}}$: $\cosp$-cosparse approximation of ${{\bfx}}$.
Initialize the cosupport $\Lambda^0 =\{i,1\le i \le p\}$, the residual ${{\bfy}}_{\text{resid}}^0 = {{\bfy}}$ and set $t = 0$.
$t = t + 1$.
Find new cosupport elements: $\Lambda_\Delta=\hat{\CF}_{a\cosp}({{\bfM}}^*{{\bfy}}^{t - 1}_{\text{resid}})$.
Update the cosupport: $\tilde{\Lambda}^t = \hat\Lambda^{t -1} \cap
\Lambda_\Delta$.
Compute a temporary estimate: ${{\bfw}} = {\operatornamewithlimits{argmin}}_{\tilde{{{\bfx}}}} {\left\Vert{{\bfy}} - {{\bfM}}\tilde{{{\bfx}}}\right\Vert}_2^2$ s.t. ${{\bf\Omega}}_{\tilde{\Lambda}^t}\tilde{{{\bfx}}} = 0$.
Enlarge the cosupport: $\hat\Lambda^t =
\hat{\CF}_{\cosp}({{\bfw}})$.
Calculate a new estimate: $\hat{{{\bfx}}}_{\text{\tiny ACoSaMP}}^t = \Q_{\hat\Lambda^t}{{\bfw}}$ for ACoSaMP, and $\hat{{{\bfx}}}_{\text{\tiny ASP}}^t = {\operatornamewithlimits{argmin}}_{\tilde{{{\bfx}}}} {\left\Vert{{\bfy}} - {{\bfM}}\tilde{{{\bfx}}}\right\Vert}_2^2$ s.t. ${{\bf\Omega}}_{\hat\Lambda^t}\tilde{{{\bfx}}} = 0$ for ASP.
Update the residual: ${{\bfy}}_{\text{resid}}^t = {{\bfy}} - {{\bfM}}\hat{{{\bfx}}}_{\text{\tiny ACoSaMP}}^t$ for ACoSaMP, and ${{\bfy}}_{\text{resid}}^t = {{\bfy}} -{{\bfM}}\hat{{{\bfx}}}_{\text{\tiny ASP}}^t$ for ASP.
Form the final solution $\hat{{{\bfx}}}_{\text{\tiny ACoSaMP}} = \hat{{{\bfx}}}_{\text{\tiny ACoSaMP}}^t$ for ACoSaMP and $\hat{{{\bfx}}}_{\text{\tiny ASP}} = \hat{{{\bfx}}}_{\text{\tiny ASP}}^t$ for ASP.
[*ACoSaMP and ASP:*]{} analysis CoSaMP (ACoSaMP) and analysis SP (ASP) are presented in Algorithm \[alg:Analysis\_CoSaMP\_SP\]. The stages are parallel to those of the synthesis CoSaMP and SP. We dwell a bit more on the meaning of the parameter $a$ in the algorithms. This parameter determines the size of the new cosupport $\Lambda_\Delta$ in each iteration. $a=1$ means that the size is $\cosp$ and according to Table \[tbl:Synthesis\_Analysis\_parallels\] it is equivalent to $a=1$ in the synthesis as done in SP in which we select new $k$ indices for the support in each iteration. In synthesis CoSaMP we use $a=2$ and select $2k$ new elements. $2k$ is the maximal support size of two added $k$-sparse vectors. The corresponding minimal size in the analysis case is $2\cosp-p$ according to Table \[tbl:Synthesis\_Analysis\_parallels\]. For this setting we need to choose $a = \frac{2\cosp-p}{\cosp}$.
The Unitary Case
----------------
For ${{\bf\Omega}} = {{\bfI}}$ the synthesis and the analysis greedy-like algorithms become equivalent. This is easy to see since in this case we have $p=d$, $k = d-\cosp$, $\Lambda = T^C$, $\Q_\Lambda\x = \x_T$ and $T_1 \cup T_2 = \Lambda_1 \cap \Lambda_2$ for $\Lambda_1 = T_1^C$ and $\Lambda_2 = T_2^C$. In addition, $\hat{\CF}_{\cosp} = \CF_{\cosp}^*$ finds the closest $\cosp$-cosparse subspace by simply taking the smallest $\cosp$ elements. Using similar arguments, also in the case where $\OM$ is a unitary matrix the analysis methods coincide with the synthesis ones. In order to get exactly the same algorithms $\M$ is replaced with $\M\OM^*$ in the synthesis techniques and the output is multiplied by $\OM^*$.
Based on this observation, we can deduce that the guarantees of the synthesis greedy-like methods apply also for the analysis ones in a trivial way. Thus, it is tempting to assume that the last should have similar guarantees based on the $\OM$-RIP. In the next section we develop such claims.
Relaxed Versions for High Dimensional Problems {#sec:relaxed_alg}
----------------------------------------------
Before moving to the next section we mention a variation of the analysis greedy-like techniques. In AHTP, ACoSaMP and ASP we need to solve the constrained minimization problem $\min_{\tilde{{{\bfx}}}}{\left\Vert{{\bfy}}-{{\bfM}}\tilde{{{\bfx}}}\right\Vert}_2^2$ s.t. ${\left\Vert\OM_{\Lambda}\tilde{{{\bfx}}}\right\Vert}_2^2 =0$. For high dimensional signals this problem is hard to solve and we suggest to replace it with minimizing ${\left\Vert{{\bfy}}-{{\bfM}}\tilde{{{\bfx}}}\right\Vert}_2^2 + \lambda{\left\Vert\OM_{\Lambda}\tilde{{{\bfx}}}\right\Vert}_2^2$, where $\lambda$ is a relaxation constant. This results in a relaxed version of the algorithms. We refer hereafter to these versions as relaxed AHTP (RAHTP) relaxed ASP (RASP) and relaxed ACoSaMP (RACoSaMP).
Algorithms Guarantees {#sec:guarantees}
=====================
In this section we provide theoretical guarantees for the reconstruction performance of the analysis greedy-like methods. For AIHT and AHTP we study both the constant step-size and the optimal step-size selections. For ACoSaMP and ASP the analysis is made for $a = \frac{2\cosp-p}{\cosp}$, but we believe that it can be extended also to other values of $a$, such as $a=1$. The performance guarantees we provide are summarized in the following two theorems. The first theorem, for AIHT and AHTP, is a simplified version of Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\] and the second theorem, for ASP and ACoSaMP, is a combination of Corollaries \[cor:ACoSaMP\_bound\] and \[cor:ASP\_bound\], all of which appear hereafter along with their proofs. Before presenting the theorems we recall the problem we aim at solving:
Consider a measurement vector $\y \in {\mathbb{R}^{\mdim}}$ such that $\y=\M\x + \e$ where $\x\in {\mathbb{R}^{\sdim}}$ is $\cosp$-cosparse, $\M\in {\mathbb{R}^{\mdim\times \sdim}}$ is a degradation operator and $\e\in {\mathbb{R}^{\mdim}}$ is a bounded additive noise. The largest singular value of $\M$ is $\sigma_\M$ and its $\OM$-RIP constant is $\delta_\cosp$. The analysis operator $\OM \in {\mathbb{R}^{\pdim \times \sdim}}$ is given and fixed. A procedure $\hat{\CF}_\cosp$ for finding a cosupport that implies a near optimal projection with a constant $C_\cosp$ is assumed to be at hand. Our task is to recover $\x$ from $\y$. The recovery result is denoted by $\hat\x$.
\[thm:AIHT\_AHTP\_general\_bound\] Consider the problem $\cal P$ and apply either AIHT or AHTP with a certain constant step-size or an optimal changing step-size, obtaining $\hat \x^t$ after $t$ iterations. If $$\begin{aligned}
\label{eq:AIHT_AHTP_general_bound_cond}
\frac{(C_\cosp - 1) \sigma_{\M}^2}{C_\cosp} < 1\end{aligned}$$ and $$\deltaB< \delta_1(C_\cosp,\sigma^2_\M),$$ where $\delta_1(C_\cosp,\sigma^2_\M)$ is a constant guaranteed to be greater than zero whenever is satisfied and $C_\cosp$ is the near-optimal projection constant for cosparsity $\cosp$ (Definition \[def:C\_optimal\_proj\]), then after a finite number of iterations $t^*$ $$\begin{aligned}
{\left\Vert{\x} - {\hat\x}^{t^*}\right\Vert}_2 \le c_1{\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ implying that these algorithms lead to a stable recovery. The constant $c_{1}$ is a function of $\deltaB$, $C_\cosp$ and $\sigma_\M^2$, and the constant step-size used is dependent on $\delta_1(C_\cosp,\sigma^2_\M)$.
\[thm:ACoSaMP\_ASP\_general\_bound\] Consider the problem $\cal P$ and apply either ACoSaMP or ASP with $a = \frac{2\cosp - \pdim}{\cosp}$, obtaining $\hat \x^t$ after $t$ iterations. If $$\begin{aligned}
\label{eq:ACoSaMP_ASP_general_bound_cond}
\frac{(C_{\hat\CF}^2-1)\sigma_{{{\bfM}}}^2 }{C_{\hat\CF}^2} < 1,\end{aligned}$$ and $$\begin{aligned}
\deltaD< \delta_2(C_{\hat\CF},\sigma^2_\M),\end{aligned}$$ where $C_{\hat\CF} = \max(C_\cosp, C_{2\cosp-\pdim})$ and $\delta_2(C_{\hat\CF},\sigma^2_\M)$ is a constant guaranteed to be greater than zero whenever is satisfied, then after a finite number of iterations $t^*$ $$\begin{aligned}
{\left\Vert\x - \hat{{{\bfx}}}^{t^*} \right\Vert}_2 \le c_2{\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ implying that these algorithms lead to a stable recovery. The constant $c_{2}$ is a function of $\deltaD$, $C_\cosp$, $C_{2\cosp-\pdim}$ and $\sigma_\M^2$.
Before we proceed to the proofs, let us comment on the constants in the above theorems. Their values can be calculated using Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\], and Corollaries \[cor:ACoSaMP\_bound\] and \[cor:ASP\_bound\]. In the case where $\OM$ is a unitary matrix, and are trivially satisfied since $C_{\cosp}=C_{2\cosp - \pdim} = 1$. In this case the $\OM$-RIP conditions become $\deltaB < \delta_1(1,\sigma^2_\M) = 1/3$ for AIHT and AHTP, and $\deltaD < \delta_2(1,\sigma^2_\M) = 0.0156$ for ACoSaMP and ASP. In terms of synthesis RIP for $\M\OM^{*}$, the condition $\deltaB < 1/3$ parallels $\delta_{2k}(\M\OM^*) < 1/3$ and similarly $\deltaD < 0.0156$ parallels $\delta_{4k}(\M\OM^*) < 0.0156$. Note that the condition we pose for AIHT and AHTP in this case is the same as the one presented for synthesis IHT with a constant step size [@Garg09Gradient]. Better reference constants were achieved in the synthesis case for all four algorithms and thus we believe that there is still room for improvement of the reference constants in the analysis context. In the non-unitary case, the value of $\sigma_\M$ plays a vital role, though we believe that this is just an artifact of our proof technique. For a random Gaussian matrix whose entries are i.i.d with a zero-mean and a variance $\frac{1}{\mdim}$, $\sigma_\M$ behaves like $\frac{d}{m}\left(1+\sqrt{\frac{\sdim}{\mdim}}\right)$. This is true also for other types of distributions for which the fourth moment is known to be bounded [@Rudelson10Non]. For example, for $\sdim/\mdim = 1.5$ we have found empirically that $\sigma_\M^2 \simeq 5$. In this case we need $C_\cosp \le \frac{5}{4}$ for to hold and $C_{\hat\CF} \le 1.118$ for to hold, and both are quite demanding on the quality of the near-optimal projection. For $C_\cosp = C_{\hat\CF} = 1.05$ we have the conditions $\deltaB \le 0.289 $ for AIHT and AHTP, and $\deltaD \le 0.0049$ for ACoSaMP and ASP; and for $C_\cosp = C_{\hat\CF} = 1.1$ we have $\deltaB \le 0.24 $ for AIHT and AHTP, and $\deltaD \le 0.00032$ for ACoSaMP and ASP.
As in the synthesis case, the $\OM$-RIP requirements for the theoretical bounds of AIHT and AHTP are better than those for ACoSaMP and ASP. In addition, in the migration from the synthesis to the analysis we lost more precision in the bounds for ACoSaMP and ASP than in those of AIHT and AHTP. In particular, even in the case where $\OM$ is the identity we do not coincide with any of the synthesis parallel RIP reference constants. We should also remember that the synthesis bound for SP is in terms of $\delta_{3k}$ and not $\delta_{4k}$ [@Dai09Subspace]. Thus, we expect that it will be possible to give a condition for ASP in terms of $\deltaC$ with better reference constants. However, our main interest in this work is to show the existence of such bounds, and in Section \[sec:thm\_conds\] we dwell more on their meaning.
We should note that here and elsewhere we can replace the conditions on $\deltaB$ and $\deltaD$ in the theorems to conditions on $\delta_{2r-p}^{\text{corank}}$ and $\delta_{4r-3p}^{\text{corank}}$ and the proofs will be almost the same[^2]. In this case we will be analyzing a version of the algorithms which is driven by the corank instead of the cosparsity. This would mean we need the near-optimal projection to be in terms of the corank. In the case where ${{\bf\Omega}}$ is in a general position, there is no difference between the cosparsity $\cosp$ and the corank $r$. However, when we have linear dependencies in $\OM$ the two measures differ and an $\cosp$-cosparse vector is not necessarily a vector with a corank $r$.
As we will see hereafter, our recovery conditions require $\deltaB$ and $\deltaD$ to be as small as possible and for this we need $2\cosp -\pdim$ and $4\cosp-3\pdim$ to be as large as possible. Thus, we need $\cosp$ to be as close as possible to $\pdim$ and for highly redundant $\OM$ this cannot be achieved without having linear dependencies in $\OM$. Apart from the theoretical advantage of linear dependencies in $\OM$, we also show empirically that an analysis dictionary with linear dependencies has better recovery rate than analysis dictionary in a general position of the same dimension. Thus, we deduce that linear dependencies in $\OM$ lead to better bounds and restoration performance.
Though linear dependencies allow $\cosp$ to be larger than $d$ and be in the order of $\pdim$, the value of the corank is always bounded by $d$ and cannot be expected to be large enough for highly redundant analysis dictionaries. In addition, we will see hereafter that the number of measurements $m$ required by the $\OM$-RIP is strongly dependent on $\cosp$ and less effected by the value of $r$. From the computational point of view we note also that using corank requires its computation in each iteration which increases the overall complexity of the algorithms. Thus, it is more reasonable to have conditions on $\deltaB$ and $\deltaD$ than on $\delta_{2r-p}^{\text{corank}}$ and $\delta_{4r-3p}^{\text{corank}}$, and our study will be focused on the cosparsity based algorithms.
AIHT and AHTP Guarantees {#sec:AIHT_AHTP_guarantees}
------------------------
A uniform guarantee for AIHT in the case that an optimal projection is given, is presented in [@Blumensath09Sampling]. The work in [@Blumensath09Sampling] dealt with a general union of subspaces, $\A$, and assumed that $\M$ is bi-Lipschitz on the considered union of subspaces. In our case $\A = \A_\cosp$ and the bi-Lipschitz constants of $\M$ are the largest $B_{L}$ and smallest $B_{U}$ where $0<B_{L} \le B_{U}$ such that for all $\cosp$-cosparse vectors ${{\bfv}}_1,{{\bfv}}_2$: $$\begin{aligned}
\label{eq:ORIP}
B_{L} {\left\Vert{{\bfv}}_1 + {{\bfv}}_2\right\Vert}_2^2 \le {\left\Vert\M({{\bfv}}_1 + {{\bfv}}_2)\right\Vert}_2^2 \le B_{U} {\left\Vert{{\bfv}}_1+{{\bfv}}_2\right\Vert}_2^2.\end{aligned}$$ Under this assumption, one can apply Theorem 2 from [@Blumensath09Sampling] to the idealized AIHT that has access to an optimal projection and uses a constant step size $\mu^t = \mu$. Relying on Table \[tbl:Synthesis\_Analysis\_parallels\] we present this theorem and replace $B_{L}$ and $B_{U}$ with $1-\deltaB$ and $1+\deltaB$ respectively.
\[thm:AIHT\_optimal\_case\_theorem\] Consider the problem $\cal P$ with $C_\cosp = 1$ and apply AIHT with a constant step size $\mu$. If $1+\deltaB \le \frac{1}{\mu} < 1.5(1-\deltaB)$ then after a finite number of iterations $t^*$ $$\begin{aligned}
{\left\Vert{\x}-{\hat\x}^{t^*}\right\Vert}_{2} \le c_3{\left\Vert\mathbf{e}\right\Vert}_{2},\end{aligned}$$ implying that AIHT leads to a stable recovery. The constant $c_3$ is a function of $\deltaB$ and $\mu$.
In this work we extend the above in several ways: First, we refer to the case where optimal projection is not known, and show that the same flavor guarantees apply for a near-optimal projection[^3]. The price we seemingly have to pay is that $\sigma_\M$ enters the game. Second, we derive similar results for the AHTP method. Finally, we also consider the optimal step size and show that the same performance guarantees hold true in that case.
\[thm:AIHT\_AHTP\_theorem\_noisy\_iter\] Consider the problem $\cal P$ and apply either AIHT or AHTP with a constant step size $\mu$ or an optimal changing step size. For a positive constant $\eta > 0$, let $$b_1 := \frac{\eta}{1+\eta} \quad\text{and}\quad b_2 := \frac{(C_\ell - 1)\sigma_\M^2 b_1^2}{C_\ell(1-\deltaB)}.$$ Suppose $\frac{b_2}{b_1^2} = \frac{(C_\ell - 1)\sigma_\M^2}{C_\ell(1-\deltaB)} < 1$, $1+\deltaB \le \frac{1}{\mu} < \left(1+\sqrt{1-\frac{b_2}{b_1^2}}\right)b_1(1-\deltaB)$ and $\frac{1}{\mu} \le \sigma_\M^2$. Then for $$\begin{aligned}
\label{eq:AIHT_AHTP_iter_num}
t \ge t^* \triangleq \frac{\log\left(\frac{\eta{\left\Vert\e\right\Vert}_2^2}{{\left\Vert\y\right\Vert}_2^2}\right)}{\log\left((1+\frac{1}{\eta})^2(\frac{1}{\mu(1-\deltaB)} - 1)C_\cosp + (C_\cosp -1)(\mu\sigma_\M^2 -1) + \frac{C_\cosp}{\eta^2}\right)},\end{aligned}$$ $$\begin{aligned}
\label{eq:AIHT_AHTP_iter_bound}
{\left\Vert\x-\hat\x^t\right\Vert}_2^2 \le \frac{(1+\eta)^2}{1-\deltaB}{\left\Vert\e\right\Vert}_2^2,\end{aligned}$$ implying that AIHT and AHTP lead to a stable recovery. Note that for an optimal changing step-size we set $\mu = \frac{1}{1+\deltaB}$ in $t^*$ and the theorem conditions turn to be $\frac{b_2}{b_1^2} < 1$ and $1+\deltaB <(1+\sqrt{1-\frac{b_2}{b_1^2}})b_1(1-\deltaB)$.
This theorem is the parallel to Theorems 2.1 in [@Garg09Gradient] for IHT. A few remarks are in order for the nature of the theorem, especially in regards to the constant $\eta$. One can view that $\eta$ gives a trade-off between satisfying the theorem conditions and the amplification of the noise. In particular, one may consider that the above theorem proves the convergence result for the noiseless case by taking $\eta$ to infinity; one can imagine solving the problem $\cal P$ where $\e \to 0$, and applying the theorem with appropriately chosen $\eta$ which approaches infinity. It is indeed possible to show that the iterate solutions of AIHT and AHTP converges to $\x$ when there is no noise. However, we will not give a separate proof since the basic idea of the arguments is the same for both cases.
As to the minimal number of iterations $t^*$ given in , one may ask whether it can be negative. In order to answer this question it should be noted that according to the conditions of the Theorem the term inside the log in the denominator is always greater than zero. Thus, $t^*$ will be negative only if ${\left\Vert{{\bfy}}\right\Vert}_2^2 < \eta{\left\Vert{{\bfe}}\right\Vert}_2^2$. Indeed, in this case $0$ iterations suffice for having the bound in .
The last remark is on the step-size selection. The advantage of the optimal changing step-size over the constant step-size is that we get the guarantee of the optimal constant step-size $\mu = \frac{1}{1+\deltaB}$ without computing it. This is important since in practice we cannot evaluate the value of $\deltaB$. However, the disadvantage of using the optimal changing step-size is its additional complexity for the algorithm. Thus, one option is to approximate the optimal selection rule by replacing it with an adaptive one, for which we do not have a theoretical guarantee. Another option is to set $\mu = 6/5$ which meets the theorem conditions for small enough $\deltaB$, in the case where an optimal projection is at hand.
We will prove the theorem by proving two key lemmas first. The proof technique is based on ideas from [@Garg09Gradient] and [@Blumensath09Sampling]. Recall that the two iterative algorithms try to reduce the objective ${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2$ over iterations $t$. Thus, the progress of the algorithms can be indirectly measured by how much the objective ${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2$ is reduced at each iteration $t$. The two lemmas that we present capture this idea. The first lemma is similar to Lemma 3 in [@Blumensath09Sampling] and relates ${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2$ to ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2$ and similar quantities at iteration $t-1$. We remark that the constraint $\frac{1}{\mu} \le \sigma_\M^2$ in Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\] may not be necessary and is added only for having a simpler derivation of the results in this theorem. Furthermore, this is a very mild condition compared to $\frac{1}{\mu} < \left(1+\sqrt{1-\frac{b_2}{b_1^2}}\right)b_1(1-\deltaB)$ and can only limit the range of values that can be used with the constant step size versions of the algorithms.
\[lem:AIHT\_lemma\] Consider the problem $\cal P$ and apply either AIHT or AHTP with a constant step size $\mu$ satisfying $\frac{1}{\mu} \ge 1+\delta_{2\cosp - \pdim}$ or an optimal step size. Then, at the $t$-th iteration, the following holds: $$\begin{aligned}
\label{eq:AIHT_lemma}
&&{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2
\le C_\ell\left({\left\Vert\y-\M\x\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2\right)\\
&& \nonumber ~~~~ + C_\ell \left(\frac{1}{\mu(1-\deltaB)}-1\right){\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2^2
+ (C_\ell-1)\mu\sigma_\M^2 {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2.\end{aligned}$$ For the optimal step size the bound is achieved with the value $\mu = \frac{1}{1+\deltaB}$.
The proof of the above lemma appears in \[sec:AIHT\_lemma\_proof\]. The second lemma is built on the result of Lemma \[lem:AIHT\_lemma\]. It shows that once the objective ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2$ at iteration $t-1$ is small enough, then we are guaranteed to have small ${\left\Vert\y-\M\hat\x^{t}\right\Vert}_2^2$ as well. Given the presence of noise, this is quite natural; one cannot expect it to approach $0$ but may expect it not to become worse. Moreover, the lemma also shows that if ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2$ is not small, then the objective in iteration $t$ is necessarily reduced by a constant factor.
\[lem:AIHT\_AHTP\_lemma\_noisy\] Suppose that the same conditions of Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\] hold true. If ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2$, then ${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2$. Furthermore, if ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 > \eta^2{\left\Vert\e\right\Vert}_2^2$, then $$\begin{aligned}
\label{eq:AIHT_AHTP_lemma_noisy}
{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 \le c_4 {\left\Vert\y-\M\x^{t-1}\right\Vert}_2^2\end{aligned}$$ where $$c_4 := \left(1+\frac{1}{\eta}\right)^2\left(\frac{1}{\mu(1-\deltaB)} - 1\right)C_\cosp + (C_\ell -1)(\mu\sigma_\M^2 -1) + \frac{C_\cosp}{\eta^2} < 1.$$
Having the two lemmas above, the proof of the theorem is straightforward.
[*Proof:*]{}\[Proof of Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\]\] When we initialize $\hat\x^0 = \mathbf{0}$, we have ${\left\Vert\y-\M\hat\x^0\right\Vert}_2^2 = {\left\Vert\y\right\Vert}_2^2$. Assuming that ${\left\Vert\y\right\Vert}_2 > \eta{\left\Vert\e\right\Vert}_2$ and applying Lemma \[lem:AIHT\_AHTP\_lemma\_noisy\] repeatedly, we obtain $${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 \le \max(c_4^t{\left\Vert\y\right\Vert}_2^2, \eta^2{\left\Vert\e\right\Vert}_2^2).$$ Since $c_4^t{\left\Vert\y\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2$ for $t \ge t^*$, we have simply $$\begin{aligned}
\label{eq:AIHT_AHTP_theorem_noisy_iter_step1}
{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2\end{aligned}$$ for $t \ge t^*$. If ${\left\Vert\y-\M\hat\x^0\right\Vert}_2 = {\left\Vert\y\right\Vert}_2 \le \eta{\left\Vert\e\right\Vert}_2$ then according to Lemma \[lem:AIHT\_AHTP\_lemma\_noisy\], $\eqref{eq:AIHT_AHTP_theorem_noisy_iter_step1}$ holds for every $t > 0$. Finally, we observe $$\begin{aligned}
\label{eq:AIHT_AHTP_theorem_noisy_iter_step2}
{\left\Vert\x-\hat\x^t\right\Vert}_2^2 \le \frac{1}{1-\deltaB}{\left\Vert\M(\x-\hat\x^t)\right\Vert}_2^2\end{aligned}$$ and, by the triangle inequality, $$\begin{aligned}
\label{eq:AIHT_AHTP_theorem_noisy_iter_step3}
{\left\Vert\M(\x-\hat\x^t)\right\Vert}_2 \le {\left\Vert\y-\M\hat\x^t\right\Vert}_2 + {\left\Vert\e\right\Vert}_2.\end{aligned}$$ By plugging into and then the resulted inequality into , the result of the Theorem follows. $\Box$
As we have seen, the above AIHT and AHTP results hold for the cases of using a constant or an optimal changing step size. The advantage of using an optimal one is that we do not need to find $\mu$ that satisfies the conditions of the theorem – the knowledge that such a $\mu$ exists is enough. However, its disadvantage is the additional computational complexity it introduces. In Section \[sec:analysis\_alg\] we have introduced a third option of using an approximated adaptive step size. In the next section we shall demonstrate this option in simulations, showing that it leads to the same reconstruction result as the optimal selection method. Note, however, that our theoretical guarantees do not cover this case.
ACoSaMP Guarantees
------------------
Having the results for AIHT and AHTP we turn to ACoSaMP and ASP. We start with a theorem for ACoSaMP. Its proof is based on the proof for CoSaMP in [@foucart10Sparse].
\[thm:ACoSaMP\_iter\_bound\] Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. Let $C_{\hat\CF} = \max({C_\cosp},C_{2\cosp-\pdim})$ and suppose that there exists $\gamma >0$ such that $$\begin{aligned}
\label{eq:C_l_tilda_C_2lp_cond}
(1 + C_{\hat\CF})\left(1 -\bigg(\frac{C_{\hat\CF}}{(1+\gamma)^2} -(C_{\hat\CF}-1)\sigma_{{{\bfM}}}^2 \bigg)\right) < 1.\end{aligned}$$ Then, there exists $\delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma) > 0$ such that, whenever $\deltaD \le \delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$, the $t$-th iteration of the algorithm satisfies $$\begin{aligned}
\label{eq:ACoSaMP_iter_bound}
&& \hspace{-0.3in} {\left\Vert\x-\hat\x^t\right\Vert}_2 \le \rho_1\rho_2{\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2 + \left(\eta_1 + \rho_1\eta_2 \right){\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ where $$\eta_1 \triangleq \frac{\sqrt{\frac{2+C_\cosp}{1+C_\cosp}+2\sqrt{C_\cosp}+C_\cosp}\sqrt{1+\deltaC}}{1-\deltaD},$$ $$\eta_2^2 \triangleq \bigg(\frac{1+\deltaC}{\gamma(1+\alpha)}
+\frac{(1+\deltaB)C_{2\cosp-\pdim}}{\gamma(1+\alpha)(1+\gamma)}
+ \frac{(C_{2\cosp-\pdim}-1)(1+\gamma)\sigma_{{{\bfM}}}^2}{(1+\alpha)(1+\gamma)\gamma}\bigg),$$ $$\rho_1^2 \triangleq \frac{1+2\deltaD\sqrt{C_\cosp} + C_\cosp}{1-\deltaD^2},$$ $$\small {\rho_2^2 \triangleq 1 -\bigg(\sqrt{\deltaD} -
\sqrt{\frac{C_{2\cosp-\pdim}}{(1+\gamma)^2}\left(1-\sqrt{\deltaB}\right)^2
-(C_{2\cosp-\pdim}-1)(1+\deltaB)\sigma_{{{\bfM}}}^2} \bigg)^2}$$ and $$\small {\alpha = \frac{\sqrt{\delta_{4\cosp-3p}}}{\sqrt{\frac{C_{2\cosp-p}}{(1+\gamma)^2}\left(1-\sqrt{\delta_{2\cosp-p}}\right)^2 -(C_{2\cosp-p}-1)(1+\delta_{2\cosp-p})\sigma_{{{\bfM}}}^2} - \sqrt{\delta_{4\cosp-3p}}}}.$$ Moreover, $\rho_1^2\rho_2^2 < 1$, i.e., the iterates converges.
The constant $\gamma$ plays a similar role to the constant $\eta$ of Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\]. It gives a tradeoff between satisfying the theorem conditions and the noise amplification. However, as opposed to $\eta$, the conditions for the noiseless case are achieved when $\gamma$ tends to zero. An immediate corollary of the above theorem is the following.
\[cor:ACoSaMP\_bound\] Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. If holds and $\deltaD< \delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$, where $C_{\hat\CF}$ and $\gamma$ are as in Theorem \[thm:ACoSaMP\_iter\_bound\] and $\delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$ is a constant guaranteed to be greater than zero whenever is satisfied, then for any $$t \ge t^* = {\left\lceil\frac{\log({\left\Vert{{\bfx}}\right\Vert}_2/{\left\Vert{{\bfe}}\right\Vert}_2)}{\log(1/\rho_1\rho_2)}\right\rceil},$$ $$\begin{aligned}
\label{eq:ACoSaMP_bound}
&& \hspace{-0.5in} {\left\Vert{{\bfx}} - \hat\x^{t^*}_{\text{\tiny ACoSaMP}}\right\Vert}_2 \le
\left(1 + \frac{1-(\rho_1\rho_2)^{t^*}}{1-\rho_1\rho_2}\left(\eta_1 + \rho_1\eta_2 \right)\right){\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ implying that ACoSaMP leads to a stable recovery. The constants $\eta_1$, $\eta_2$, $\rho_1$ and $\rho_2$ are the same as in Theorem \[thm:ACoSaMP\_iter\_bound\].
[*Proof:*]{} By using and recursion we have that $$\begin{aligned}
\label{eq:ACoSaMP_bound_step1}
&& \hspace{-0.3in} {\left\Vert{{\bfx}} - \hat\x^{t^*}_{\text{\tiny ACoSaMP}}\right\Vert}_2 \le (\rho_1\rho_2)^{t^*}{\left\Vert{{\bfx}} - \hat\x^{0}_{\text{\tiny ACoSaMP}}\right\Vert}_2
\\ \nonumber && \hspace{-0.3in}
+(1+\rho_1\rho_2+(\rho_1\rho_2)^2+\dots (\rho_1\rho_2)^{t^*-1})\left(\eta_1 + \rho_1\eta_2 \right){\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$ Since $\hat\x^{0}_{\text{\tiny ACoSaMP}}=0$, after $t^*$ iterations, one has $$\begin{aligned}
\label{eq:ACoSaMP_bound_step2}
(\rho_1\rho_2)^{t^*}{\left\Vert{{\bfx}} - \hat\x^{0}_{\text{\tiny ACoSaMP}}\right\Vert}_2 = (\rho_1\rho_2)^{t^*}{\left\Vert\x\right\Vert}_2 \le {\left\Vert\e\right\Vert}_2.\end{aligned}$$ By using the equation of geometric series with and plugging into it, we get . $\Box$
We turn now to prove the theorem. Instead of presenting the proof directly, we divide the proof into several lemmas. The first lemma gives a bound for ${\left\Vert\x - {{\bfw}}\right\Vert}_2$ as a function of ${\left\Vert{{\bfe}}\right\Vert}_2$ and ${\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2$.
\[lem:ACoSaMP\_xp\_bound\] Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. For each iteration we have $$\begin{aligned}
\label{eq:ACoSaMP_xp_bound}
{\left\Vert\x - {{\bfw}}\right\Vert}_2 &\le& \frac{1}{\sqrt{1-\deltaD^2}}{\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}} )\right\Vert}_2+ \frac{\sqrt{1+\deltaC}}{1-\deltaD}{\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$
The second lemma bounds ${\left\Vert\x - \hat\x^{t}_{\text{\tiny ACoSaMP}}\right\Vert}_2$ in terms of ${\left\Vert\P_{\tilde{\Lambda}^t}(\x- \hat\x^{t}_{\text{\tiny ACoSaMP}} )\right\Vert}_2$ and ${\left\Vert{{\bfe}}\right\Vert}_2$ using the first lemma.
\[lem:ACoSaMP\_xt\_bound1\] Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. For each iteration we have $$\begin{aligned}
&& \hspace{-0.5in} {\left\Vert\x - \hat\x^t\right\Vert}_2 \le \rho_1{\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2 + \eta_1{\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ where $\eta_1$ and $\rho_1$ are the same constants as in Theorem \[thm:ACoSaMP\_iter\_bound\].
The last lemma bounds ${\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2$ with ${\left\Vert\x - \hat\x^{t-1}_{\text{\tiny ACoSaMP}}\right\Vert}_2$ and ${\left\Vert{{\bfe}}\right\Vert}_2$.
\[lem:ACoSaMP\_Pxp\_bound\] Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. if $$\begin{aligned}
\label{eq:C_2lp_cond}
C_{2\cosp-p} < \frac{\sigma_{{{\bfM}}}^2(1+\gamma)^2}{\sigma_{{{\bfM}}}^2(1+\gamma)^2-1},\end{aligned}$$ then there exists $\tilde{\delta}_{\text{ \tiny ACoSaMP}}(C_{2\cosp - \pdim}, \sigma_{{{\bfM}}}^2, \gamma) >0$ such that for any $\deltaB < \tilde{\delta}_{\text{ \tiny ACoSaMP}}(C_{2\cosp - \pdim}, \sigma_{{{\bfM}}}^2, \gamma)$ $$\begin{aligned}
\label{eq:ACoSaMP_Pxp_bound}
&& {\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2 \le
\eta_2{\left\Vert{{\bfe}}\right\Vert}_2
+ \rho_2{\left\Vert\x - \hat\x^{t-1}\right\Vert}_2.\end{aligned}$$ The constants $\eta_2$ and $\rho_2$ are as defined in Theorem \[thm:ACoSaMP\_iter\_bound\].
The proofs of Lemmas \[lem:ACoSaMP\_xp\_bound\], \[lem:ACoSaMP\_xt\_bound1\] and \[lem:ACoSaMP\_Pxp\_bound\] appear in \[sec:ACoSaMP\_xp\_bound\_proof\], \[sec:ACoSaMP\_xt\_bound1\_proof\] and \[sec:ACoSaMP\_Pxp\_bound\_proof\] respectively. With the aid of the above three lemmas we turn to prove Theorem \[thm:ACoSaMP\_iter\_bound\].
[*Proof:*]{}\[Proof of Theorem \[thm:ACoSaMP\_iter\_bound\]\] Remark that since $1+C_{\hat{\CF}} > 1$ we have that implies $\frac{C_{\hat\CF}}{(1+\gamma)^2} -(C_{\hat\CF}-1)\sigma_{{{\bfM}}}^2 \ge 0$. Because of that the condition in in Lemma \[lem:ACoSaMP\_Pxp\_bound\] holds. Substituting the inequality of Lemma \[lem:ACoSaMP\_Pxp\_bound\] into the inequality of Lemma \[lem:ACoSaMP\_xt\_bound1\] gives . The iterates convergence if $\rho_1^2\rho_2^2 = \frac{1+2\deltaD\sqrt{C_\cosp} + C_\cosp}{1 -\deltaD^2}\rho_2^2 < 1$. By noticing that $\rho_2^2<1$ it is enough to require $\frac{1 + C_\cosp}{1 -\deltaD^2}\rho_2^2 + \frac{2\deltaD\sqrt{C_\cosp}}{1 -\deltaD^2} < 1$. The last is equivalent to $$\begin{aligned}
\label{eq:ACoSaMP_rho1rho2_cond_step1}
&& \hspace{-0.4in} (1 + C_\cosp)
\left(1 -\left(\sqrt{\deltaD} -
\sqrt{\frac{C_{2\cosp - \pdim}}{(1+\gamma)^2}\left(1-\sqrt{\deltaB}\right)^2
-(C_{2\cosp - \pdim}-1)(1+\deltaB)\sigma_{{{\bfM}}}^2} \right)^2\right) \\ \nonumber &&
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~+2\deltaD\sqrt{C_\cosp} -1 +\deltaD^2 < 0.\end{aligned}$$ It is easy to verify that $\zeta(C,\delta) \triangleq \frac{C}{(1+\gamma)^2}\left(1-\sqrt{\delta}\right)^2
-(C-1)(1+\delta)\sigma_{{{\bfM}}}^2$ is a decreasing function of both $\delta$ and $C$ for $0 \le \delta \le 1$ and $C>1$. Since $1 \le C_{2\cosp - \pdim}\le C_{\hat\CF}$, $\deltaB \le \deltaD$ and $\delta \ge 0$ we have that $\zeta(C_{\hat\CF},\deltaD) \le \zeta(C_{2\cosp - \pdim},\deltaD) \le \zeta(C_{2\cosp - \pdim},\deltaB) \le \zeta(1,0) =\frac{1}{(1+\gamma)^2}\le 1$. Thus we have that $-1 \le-(\sqrt{\deltaD} -\zeta(C_{2\cosp-\pdim},\deltaB))^2 \le -\deltaD +2\sqrt{\deltaD} - \zeta(C_{\hat\CF},\deltaD)$. Combining this with the fact that $C_\cosp \le C_{\hat\CF}$ provides the following guarantee for $\rho_1^2\rho_2^2 < 1$, $$\begin{aligned}
\label{eq:ACoSaMP_rho1rho2_cond_step2}
&& \hspace{-0.3in} (1+ C_{\hat\CF})\bigg(1 -\deltaD +2\sqrt{\deltaD}
\\ \nonumber && \hspace{-0.3in}
-\frac{C_{\hat\CF}}{(1+\gamma)^2}\left(1-2\sqrt{\deltaD}+\deltaD\right)
+(C_{\hat\CF}-1)(1+\deltaD)\sigma_{{{\bfM}}}^2 \bigg)+2\deltaD\sqrt{C_{\hat\CF}} -1 + \deltaD^2 <0.\end{aligned}$$ Let us now assume that $\deltaD \le \frac{1}{2}$. This necessarily means that $\delta_{\text{\tiny ACoSaMP}} \le \frac{1}{2}$ in the end. This assumption implies $\deltaD^2\le \frac{1}{2}\deltaD$. Using this and gathering coefficients, we now consider the condition $$\begin{aligned}
\label{eq:ACoSaMP_rho1rho2_cond_step3}
&& \hspace{-0.3in} (1+ C_{\hat\CF})\left(1-\frac{C_{\hat\CF}}{(1+\gamma)^2}+(C_{\hat\CF}-1)\sigma_{{{\bfM}}}^2\right)-1 + 2(1+ C_{\hat\CF})\left(1+\frac{C_{\hat\CF}}{(1+\gamma)^2}\right)\sqrt{\deltaD}
\\ \nonumber && \hspace{-0.3in}
+\left((1+ C_{\hat\CF})\left(-1-\frac{C_{\hat\CF}}{(1+\gamma)^2}+(C_{\hat\CF}-1)\sigma_{{{\bfM}}}^2\right)+ 2\sqrt{C_{\hat\CF}} +\frac{1}{2}\right)\deltaD <0.\end{aligned}$$ The expression on the LHS is a quadratic function of $\sqrt{\deltaD}$. Note that since holds the constant term in the quadratic function is negative. This guarantees the existence of a range of values $[0,\delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)]$ for $\deltaD$ for which holds, where $\delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$ is the square of the positive solution of the quadratic function. In case of two positive solutions we should take the smallest among them – in this case the coefficient of $\deltaD$ in will be positive.
Looking back at the proof of the theorem, we observe that the value of the constant $\delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$ can potentially be improved: at the beginning of the proof, we have used $\rho_2^2\le 1$. By the end, we obtained $\rho_2^2\le \rho_1^{-2} \le 0.25$ since $\rho_1 > 2$. If we were to use this bound at the beginning, we would have obtained better constant $\delta_{\text{\tiny ACoSaMP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$. $\Box$
ASP Guarantees
--------------
Having the result of ACoSaMP we turn to derive a similar result for ASP. The technique for deriving a result for ASP based on the result of ACoSaMP is similar to the one we used to derive a result for AHTP from the result of AIHT.
\[thm:ASP\_iter\_bound\] Consider the problem $\cal P$ and apply ASP with $a = \frac{2\cosp - \pdim}{\cosp}$. If holds and $\deltaD \le
\delta_{\text{\tiny ASP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$, where $C_{\hat\CF}$ and $\gamma$ are as in Theorem \[thm:ACoSaMP\_iter\_bound\], and $\delta_{\text{\tiny ASP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$ is a constant guaranteed to be greater than zero whenever is satisfied, then the $t$-th iteration of the algorithm satisfies $$\begin{aligned}
\label{eq:ASP_iter_bound}
&& \hspace{-0.5in} {\left\Vert{{\bfx}} - \hat\x^t_{\text{\tiny ASP}}\right\Vert}_2 \le \frac{1+\deltaB}{1-\deltaB
}\rho_1\rho_2{\left\Vert{{\bfx}} - \hat{{{\bfx}}}^{t-1}_{\text{\tiny ASP} }\right\Vert}_2 +\left(\frac{1+\deltaB}{1-\deltaB}\left(\eta_1 + \rho_1\eta_2 \right)+ \frac{2}{1-\deltaB}\right){\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$ and the iterates converges, i.e., $\rho_1^2\rho_2^2 < 1$. The constants $\eta_1$, $\eta_2$, $\rho_1$ and $\rho_2$ are the same as in Theorem \[thm:ACoSaMP\_iter\_bound\].
[*Proof:*]{} We first note that according to the selection rule of $\hat{{{\bfx}}}_{\text{\tiny ASP}}$ we have that $$\begin{aligned}
{\left\Vert{{\bfy}} - {{\bfM}}\hat{{{\bfx}}}^t_{\text{\tiny ASP}}\right\Vert}_2 \le {\left\Vert{{\bfy}} - {{\bfM}}{{\bfQ}}_{\hat\Lambda^t}{{\bfw}}\right\Vert}_2.\end{aligned}$$ Using the triangle inequality and the fact that ${{\bfy}} = {{\bfM}}{{\bfx}}+\e$ for both the LHS and the RHS we have $$\begin{aligned}
{\left\Vert{{\bfM}}({{\bfx}} - \hat{{{\bfx}}}^t_{\text{\tiny ASP}})\right\Vert}_2 -{\left\Vert{{\bfe}}\right\Vert}_2 \le {\left\Vert {{\bfM}}({{\bfx}} - {{\bfQ}}_{\hat\Lambda^t}{{\bfw}})\right\Vert}_2 + {\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$ Using the ${{\bf\Omega}}$-RIP property of ${{\bfM}}$ with the fact that ${{\bfx}}$, $\hat{{{\bfx}}}_{\text{\tiny ASP}}$ and ${{\bfQ}}_{\hat\Lambda^t}{{\bfw}}$ are $\cosp$-cosparse we have $$\begin{aligned}
{\left\Vert{{\bfx}} - \hat{{{\bfx}}}^t_{\text{\tiny ASP}}\right\Vert}_2 \le
\frac{1+\deltaB}{1-\deltaB}{\left\Vert{{\bfx}} - {{\bfQ}}_{\hat\Lambda^t}{{\bfw}}\right\Vert}_2
+ \frac{2}{1-\deltaB}{\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$ Noticing that ${{\bfQ}}_{\hat\Lambda^t}{{\bfw}}$ is the solution we get in one iteration of ACoSaMP with initialization of $\hat{{{\bfx}}}^{t-1}_{\text{\tiny ASP} }$, we can combine the above with the result of Theorem \[thm:ACoSaMP\_iter\_bound\] getting . For $\frac{1+\deltaB}{1-\deltaB}\rho_1\rho_2<1$ to hold we need that $$\begin{aligned}
\label{eq:ASP_rho1rho2_cond1}
&& \hspace{-0.3in} \frac{1+2\deltaD\sqrt{C_\cosp} + C_\cosp}{(1-\deltaD)^2}
\left(1-\left( \left(\frac{\sqrt{\tilde{C}_{2\cosp-p}}}{1+\gamma} +1 \right)\sqrt{\deltaD} -\frac{\sqrt{\tilde{C}_{2\cosp-p}}}{1+\gamma}
\right)^2\right) < 1.\end{aligned}$$ Remark that the above differs from what we have for ACoSaMP only in the denominator of the first element in the LHS. In ACoSaMP $1-\deltaD^2$ appears instead of $(1-\deltaD)^2$. Thus, Using a similar process to the one in the proof of ACoSaMP we can show that holds if the following holds $$\begin{aligned}
\label{eq:ASP_rho1rho2_cond_step3}
&& \hspace{-0.3in} (1+ C_{\hat\CF})\left(1-\frac{C_{\hat\CF}}{(1+\gamma)^2}+(C_{\hat\CF}-1)\sigma_{{{\bfM}}}^2\right)-1 + 2(1+ C_{\hat\CF})\left(1+\frac{C_{\hat\CF}}{(1+\gamma)^2}\right)\sqrt{\deltaD}
\\ \nonumber && \hspace{-0.3in}
+\left((1+ C_{\hat\CF})\left(-1-\frac{C_{\hat\CF}}{(1+\gamma)^2}+(C_{\hat\CF}-1)\sigma_{{{\bfM}}}^2\right)+ 2\sqrt{C_{\hat\CF}} +2\right)\deltaD <0.\end{aligned}$$ Notice that the only difference of the above compared to is that we have $+2$ instead of $+0.5$ in the coefficient of $\deltaD$ and this is due to the difference we mentioned before in the denominator in . The LHS of is a quadratic function of $\sqrt{\deltaD}$. As before, we notice that if holds then the constant term of the above is positive and thus $\delta_{\text{\tiny ASP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma) \ge 0$ exists and is the square of the positive solution of the quadratic function. $\Box$
Having Theorem \[thm:ASP\_iter\_bound\] we can immediately have the following corollary which is similar to the one we have for ACoSaMP. The proof resembles the one of Corollary \[cor:ACoSaMP\_bound\] and omitted.
\[cor:ASP\_bound\] Consider the problem $\cal P$ and apply ASP with $a = \frac{2\cosp - \pdim}{\cosp}$. If holds and $\deltaD \le
\delta_{\text{\tiny ASP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$, where $C_{\hat\CF}$ and $\gamma$ are as in Theorem \[thm:ACoSaMP\_iter\_bound\], and $\delta_{\text{\tiny ASP}}(C_{\hat\CF},\sigma_{{{\bfM}}}^2,\gamma)$ is a constant guaranteed to be greater than zero whenever is satisfied, then for any $$t\ge t^* = {\left\lceil\frac{\log({\left\Vert{{\bfx}}\right\Vert}_2/{\left\Vert{{\bfe}}\right\Vert}_2)}{\log(1/\frac{1+\deltaB}{1-\deltaB}\rho_1\rho_2)}\right\rceil},$$ $$\begin{aligned}
\label{eq:ASP_bound}
&& \hspace{-0.5in} {\left\Vert{{\bfx}}^{t}_{\text{\tiny ASP}} -{{\bfx}}\right\Vert}_2 \le
\Bigg(1 + \frac{1-\left(\frac{1+\deltaB}{1-\deltaB}\rho_1\rho_2\right)^{t}}{1-\frac{1+\deltaB}{1-\deltaB}\rho_1\rho_2}\cdot
\left(\frac{1+\deltaB}{1-\deltaB}\left(\eta_1 + \rho_1\eta_2 \right)+ \frac{2}{1-\deltaB}\right)\Bigg){\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$ implying that ASP leads to a stable recovery. The constants $\eta_1$, $\eta_2$, $\rho_1$ and $\rho_2$ are the same as in Theorem \[thm:ACoSaMP\_iter\_bound\].
Non-Exact Cosparse Case
-----------------------
In the above guarantees we have assumed that the signal $\x$ is $\cosp$-cosparse. In many cases, it is not exactly $\cosp$-cosparse but only nearly so. Denote by $\x^\cosp = \Q_{\CF^*_\cosp(\x)}\x$ the best $\cosp$-cosparse approximation of $\x$, we have the following theorem that provides us with a guarantee also for this case. Similar result exists also in the synthesis case for the synthesis-$\ell_1$ minimization problem [@Candes08TheRIP].
\[thm:non\_exact\_noisy\_result\] Consider a variation of problem $\cal P$ where $\x$ is a general vector, and apply either AIHT or AHTP both with either constant or changing step size; or ACoSaMP or ASP with $a = \frac{2\cosp - \pdim}{\cosp}$, and all are used with a zero initialization. Under the same conditions of Theorems \[thm:AIHT\_AHTP\_general\_bound\] and \[thm:ACoSaMP\_ASP\_general\_bound\] we have for any $t\ge t^*$ $$\begin{aligned}
\label{eq:non_exact_bound}
{\left\Vert\x - {\hat\x}\right\Vert}_2 \le {\left\Vert\x - \x^\cosp\right\Vert}_2 + c{\left\Vert\M(\x - \x^\cosp)\right\Vert}_2 + c{\left\Vert\e\right\Vert}_2,\end{aligned}$$ where $t^*$ and $c$ are the constants from Theorems \[thm:AIHT\_AHTP\_general\_bound\] and \[thm:ACoSaMP\_ASP\_general\_bound\].
[*Proof:*]{} First we notice that we can rewrite $\y = \M\x^\cosp + \M(\x - \x^\cosp) + {{\bfe}}$. Denoting ${{\bfe}}^\cosp = \M(\x - \x^\cosp) + {{\bfe}}$ we can use Theorems \[thm:AIHT\_AHTP\_general\_bound\] and \[thm:ACoSaMP\_ASP\_general\_bound\] to recover $\x^\cosp$ and have $$\begin{aligned}
{\left\Vert\x^\cosp - \hat\x\right\Vert}_2 \le c{\left\Vert\e^\cosp\right\Vert}_2.\end{aligned}$$ Using the triangle inequality for ${\left\Vert\x - \hat\x\right\Vert}_2$ with the above gives $$\begin{aligned}
{\left\Vert\x - {\hat\x}\right\Vert}_2 \le {\left\Vert\x - \x^\cosp\right\Vert}_2 +{\left\Vert\x^\cosp - {\hat\x}\right\Vert}_2 \le {\left\Vert\x - \x^\cosp\right\Vert}_2+ c{\left\Vert\e^\cosp\right\Vert}_2.\end{aligned}$$ Using again the triangle inequality for ${\left\Vert\e^\cosp\right\Vert}_2 \le {\left\Vert\e\right\Vert}_2 + {\left\Vert\M(\x - \x^\cosp)\right\Vert}_2$ provides us with the desired result. $\Box$
Theorem Conditions {#sec:thm_conds}
------------------
Having the results of the theorems we ask ourselves whether their conditions are feasible. As we have seen in Section \[sec:omega\_RIP\], the requirement on the $\OM$-RIP for many non-trivial matrices. In addition, as we have seen in the introduction of this section we need $C_\cosp$ and $C_{2\cosp -\pdim}$ to be one or close to one for satisfying the conditions of the theorems. Using the thresholding in for cosupport selection with a unitary $\OM$ satisfies the conditions in a trivial way since $C_\cosp = {C}_{2\cosp-p} = 1$. This case coincides with the synthesis model for which we already have theoretical guarantees. As shown in Section \[sec:near\_opt\_proj\], optimal projection schemes exist for $\OM_{\text{1D-DIF}}$ and $\OM_{\text{FUS}}$ which do not belong to the the synthesis framework. For a general ${{\bf\Omega}}$, a general projection scheme is not known and if the thresholding method is used the constants in do not equal one and are not even expected to be close to one [@Giryes11Iterative]. It is interesting to ask whether there exists an efficient general projection scheme that guarantees small constants for any given operator ${{\bf\Omega}}$, or for specifically structured $\OM$. We leave these questions as subject for future work. Instead, we show empirically in the next section that a weaker projection scheme that does not fulfill all the requirements of the theorems leads to a good reconstruction result. This suggests that even in the absence of good near optimal projections we may still use the algorithms practically.
Comparison to Other Works {#sec:comp_other}
-------------------------
Among the existing theoretical works that studied the performance of analysis algorithms [@Nam12Cosparse; @Vaiter12Robust; @Peleg12Performance], the result that resembles ours is the result for $\ell_1$-analysis in [@Candes11Compressed]. This work analyzed the $\ell_1$-analysis minimization problem with a synthesis perspective. The analysis dictionary $\OM$ was replaced with the conjugate of a synthesis dictionary ${{\bfD}}$ which is assumed to be a tight frame, resulting with the following minimization problem. $$\begin{aligned}
\label{eq:l1_D_analysis}
\hat{{{\bfx}}}_{A-\ell_1} = {\operatornamewithlimits{argmin}}_{{{\bfz}}} {\left\Vert{{\bfD}}^*{{\bfz}}\right\Vert}_1 & s.t. & {\left\Vert{{\bfy}} - {{\bfM}}{{\bfz}}\right\Vert}_2 \le \epsilon.\end{aligned}$$ It was shown that if ${{\bfM}}$ has the ${{\bfD}}$-RIP [@Candes11Compressed; @Blumensath09Sampling] with $\delta_{7k} < 0.6$, an extension of the synthesis RIP, then $$\begin{aligned}
\label{eq:l1_candes_rec_error}
{\left\Vert\hat{{{\bfx}}}_{A-\ell_1} - {{\bfx}}\right\Vert}_2 \le \tilde{C}_{\ell_1}\epsilon + \frac{{\left\Vert{{\bfD}}^*{{\bfx}}-[{{\bfD}}^*{{\bfx}}]_k\right\Vert}_1}{\sqrt{k}}.\end{aligned}$$ We say that a matrix ${{\bfM}}$ has a ${{\bfD}}$-RIP with a constant $\delta_k$ if for any signal ${{\bfz}}$ that has a $k$-sparse representation under ${{\bfD}}$ $$\begin{aligned}
(1-\delta_k){\left\Vert{{{\bfz}}}\right\Vert}_2^2 \le {\left\Vert{{\bfM}}{{{\bfz}}}\right\Vert}_2^2 \le (1+\delta_k){\left\Vert{{{\bfz}}}\right\Vert}_2^2.\end{aligned}$$ The authors in [@Candes11Compressed] presented this result as a synthesis result that allows linear dependencies in ${{\bfD}}$ at the cost of limiting the family of signals to be those for which ${\left\Vert{{\bfD}}^*{{\bfx}}-[{{\bfD}}^*{{\bfx}}]_k\right\Vert}_1$ is small. However, having the analysis perspective, we can realize that they provided a recovery guarantee for $\ell_1$-analysis under the new analysis model for the case that $\OM$ is a tight frame. An easy way to see it is to observe that for an $\cosp$-cosparse signal ${{\bfx}}$, setting $k=\pdim-\cosp$, we have that ${\left\Vert\OM{{\bfx}}-[\OM^*{{\bfx}}]_{\pdim-\cosp}\right\Vert}_1 =0$ and thus in the case $\epsilon = 0$ we get that $\eqref{eq:l1_candes_rec_error}$ guarantees the recovery of ${{\bfx}}$ by using with ${{\bfD}}^* = \OM$. Thus, though the result in [@Candes11Compressed] was presented as a reconstruction guarantee for the synthesis model, it is actually a guarantee for the analysis model.
Our main difference from [@Candes11Compressed] is that the proof technique relies on the analysis model and not on the synthesis one and that the results presented here are for general operators and not only for tight frames. For instance, the operators $\OM_{\text{1D-DIF}}$ and $\OM_{\text{FUS}}$ for which the guarantees hold are not tight frames where $\OM_{\text{1D-DIF}}$ is not even a frame. However, the drawback of our approached compared to the work in [@Candes11Compressed] is that it is still not known how to perform an optimal or a near optimal projection for a tight frame.
In the non-exact sparse case our results differ from the one in in the sense that it looks at the projection error and not at the values of $\OM\x$. It would be interesting to see if there is a connection between the two and whether one implies the other.
A recent work has studied the $\ell_1$-analysis minimization with the 2D-DIF operator, also known as anisotropic two dimensional total-variation (2D-TV) [@Needell12Stable]. It would be interesting to see whether similar results can be achieved for the greedy-like techniques proposed here with 2D-DIF.
Experiments {#sec:exp}
===========
In this section we repeat some of the experiments performed in [@Nam12Cosparse] for the noiseless case (${{\bfe}} =0$) and some of the experiments performed in [@Nam11GAPN] for the noisy case[^4].
Targeted Cosparsity {#sec:targ_cosp}
-------------------
Just as in the synthesis counterpart of the proposed algorithms, where a target sparsity level $\spar$ must be selected before running the algorithms, we have to choose the targeted cosparsity level which will dictate the projection steps. In the synthesis case it is known that it may be beneficial to over-estimate the sparsity $\spar$. Similarly in the analysis framework the question arises: In terms of recovery performance, does it help to under-estimate the cosparsity $\cosp$? A tentative positive answer comes from the following heuristic: Let $\tilde\Lambda$ be a subset of the cosupport $\Lambda$ of the signal $\x$ with $\tilde\cosp := |\tilde\Lambda| < \cosp = |\Lambda|$. According to Proposition 3 in [@Nam12Cosparse] $$\begin{aligned}
\label{eq:kappa_uniqueness}
\kappa_{\OM}(\tilde{\cosp}) \le \frac{m}{2}\end{aligned}$$ is a sufficient condition to identify $\tilde\Lambda$ in order to recover $\x$ from the relations $\y = \M\x$ and $\OM_{\tilde\Lambda}\x = 0$. $\kappa_{\OM}(\tilde{\cosp}) = \max_{\tilde{\Lambda} \in \L_{\tilde\cosp}} \dim (\aspace_{\tilde{\Lambda}})$ is a function of $\tilde\cosp$. Therefore, we can replace $\cosp$ with the smallest $\tilde\cosp$ that satisfies as the effective cosparsity in the algorithms. Since it is easier to identify a smaller cosupport set it is better to run the algorithm with the smallest possible value of $\tilde\cosp$, in the absence of noise. In the presence of noise, larger values of $\cosp$ allows a better denoising. Note, that in some cases the smallest possible value of $\tilde\cosp$ will be larger than the actual cosparsity of $\x$. In this case we cannot replace $\cosp$ with $\tilde{\cosp}$.
We take two examples for selecting $\tilde{\cosp}$. The first is for $\OM$ which is in general position and the second is for $\OM_{2D-DIF}$, the finite difference analysis operator that computes horizontal and vertical discrete derivatives of an image which is strongly connected to the total variation (TV) norm minimization as noted before. For $\OM$ that is in general position $\kappa_{\OM}(\tilde{\cosp}) = \max(\sdim -\cosp,0)$ [@Nam12Cosparse]. In this case we choose $$\begin{aligned}
\tilde{\cosp} = \min\left(\sdim -\frac{\mdim}{2}, \cosp\right).\end{aligned}$$ For $\OM_{DIF}$ we have $\kappa_{\OM_{DIF}}(\tilde{\cosp}) \ge d - \frac{\cosp}{2} - \sqrt{\frac{\cosp}{2}}-1$ [@Nam12Cosparse] and $$\begin{aligned}
\tilde{\cosp} = \lceil\min((-1/\sqrt{2}+\sqrt{2d-m-1.5})^2, \cosp)\rceil.\end{aligned}$$
Replacing $\cosp$ with $\tilde\cosp$ is more relevant to AIHT and AHTP than ACoSaMP and ASP since in the last we intersect cosupport sets and thus the estimated cosupport set need to be large enough to avoid empty intersections. Thus, for $\OM$ in general position we use the true cosparsity level for ACoSaMP and ASP. For $\OM_{DIF}$, where linear dependencies occur, the corank does not equal the cosparsity and we use $\tilde{\cosp}$ instead of $\cosp$ since it will be favorable to run the algorithm targeting a cosparsity level in the middle. In this case $\cosp$ tends to be very large and it is more likely to have non-empty intersections .
Phase Diagrams for Synthetic Signals in the Noiseless Case
----------------------------------------------------------
We begin with with synthetic signals in the noiseless case. We test the performance of AIHT with a constant step-size, AIHT with an adaptive changing step-size, AHTP with a constant step-size, AHTP with an adaptive changing step-size, ACoSaMP with $a=\frac{2\cosp-p}{\cosp}$, ACoSaMP with $a=1$, ASP with $a=\frac{2\cosp-p}{\cosp}$ and ASP with $a=1$. We compare the results to those of A-$\ell_1$-minimization [@elad07Analysis] and GAP [@Nam12Cosparse]. We use a random matrix ${{\bfM}}$ and a random tight frame with $d=120$ and $p=144$, where each entry in the matrices is drawn independently from the Gaussian distribution.
We draw a phase transition diagram [@Donoho09countingfaces] for each of the algorithms. We test $20$ different possible values of $m$ and $20$ different values of $\cosp$ and for each pair repeat the experiment $50$ times. In each experiment we check whether we have a perfect reconstruction. White cells in the diagram denotes a perfect reconstruction in all the experiments of the pair and black cells denotes total failure in the reconstruction. The values of $m$ and $\cosp$ are selected according to the following formula: $$\begin{aligned}
m = \delta d && \cosp = d - \rho m,\end{aligned}$$ where $\delta$, the sampling rate, is the x-axis of the phase diagram and $\rho$, the ratio between the cosparsity of the signal and the number of measurements, is the y-axis.
Figure \[fig:phaseDiagramAll1\_2\] presents the reconstruction results of the algorithms. It should be observed that AIHT and AHTP have better performance using the adaptive step-size than using the constant step-size. The optimal step-size has similar reconstruction result like the adaptive one and thus not presented. For ACoSaMP and ASP we observe that it is better to use $a=1$ instead of $a= \frac{2\cosp-p}{\cosp}$. Compared to each other we see that ACoSaMP and ASP achieve better recovery than AHTP and AIHT. Between the last two, AHTP is better. Though AIHT has inferior behavior, we should mention that with regards to running time AIHT is the most efficient. Afterwards we have AHTP and then ACoSaMP and ASP. Compared to $\ell_1$ and GAP we observe that ACoSaMP and ASP have competitive results.
[ccccc]{} & & & &\
& ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](AIHT_adaptive2 "fig:"){width="1.3in"} & ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](AHTP_adaptive2 "fig:"){width="1.3in"} & ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](ACoSaMP2 "fig:"){width="1.3in"} & ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](ASP2 "fig:"){width="1.3in"}\
& ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](AIHT_adaptive2_TV "fig:"){width="1.3in"} & ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](AHTP_adaptive2_TV "fig:"){width="1.3in"} & ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](ACoSaMP2_TV "fig:"){width="1.3in"} & ![Recovery rate for a random tight frame with $p=240$ and $d=120$ (up) and a finite difference operator (bottom). From left to right: AIHT and AHTP with an adaptive changing step-size, and ACoSaMP and ASP with $a=1$.[]{data-label="fig:phaseDiagramAll2"}](ASP2_TV "fig:"){width="1.3in"}
With the above observations, we turn to test operators with higher redundancy and see the effect of linear dependencies in them. We test two operators. The first is a random tight frame as before but with redundancy factor of $2$. The second is the two dimensional finite difference operator $\OM_{\text{2D-DIF}}$. In Fig. \[fig:phaseDiagramAll2\] we present the phase diagrams for both operators using AIHT with an adaptive changing step-size, AHTP with an adaptive changing step-size, ACoSaMP with $a=1$, and ASP with $a=1$. As observed before, also in this case the ACoSaMP and ASP outperform AIHT and AHTP in both cases and AHTP outperform AIHT. We mention again that the better performance comes at the cost of higher complexity. In addition, as we expected, having redundancies in $\OM$ results with a better recovery.
Reconstruction of High Dimensional Images in the Noisy Case
-----------------------------------------------------------
We turn now to test the methods for high dimensional signals. We use RASP and RACoSaMP (relaxed versions of ASP and ACoSaMP defined in Section \[sec:relaxed\_alg\]) for the reconstruction of the *Shepp-Logan phantom* from few number of measurements. The sampling operator is a two dimensional Fourier transform that measures only a certain number of radial lines from the Fourier transform. The cosparse operator is $\OM_{\text{2D-DIF}}$ and the cosparsity used is the actual cosparsity of the signal under this operator ($\cosp = 128014$). The phantom image is presented in Fig. \[fig:shepploganphantom\]. Using the RACoSaMP and RASP we get a perfect reconstruction using only $15$ radial lines, i.e., only $m=3782$ measurements out of $d=65536$ which is less then $6$ percent of the data in the original image. The algorithms requires less than $20$ iterations for having this perfect recovery. For AIHT and RAHTP we achieve a reconstruction which is only close to the original image using $35$ radial lines. The reconstruction result of AIHT is presented in Fig \[fig:shepploganphantom\_AIHT\]. The advantage of the AIHT, though it has an inferior performance, over the other methods is its running time. While the others need several minutes for each reconstruction, for the AIHT it takes only few seconds to achieve a visually reasonable result.
Exploring the noisy case, we perform a reconstruction using RASP of a noisy measurement of the phantom with $22$ radial lines and signal to noise ratio (SNR) of $20$. Figure \[fig:shepploganphantom\_noisy\] presents the noisy image, the result of applying inverse Fourier transform on the measurements, and Fig. \[fig:shepploganphantom\_noisy\_RASP\] presents its reconstruction result. Note that for the minimization process we solve conjugate gradients, in each iteration and take only the real part of the result and crop the values of the resulted image to be in the range of $[0,1]$. We get a peak SNR (PSNR) of $36dB$. We get similar results using RACoSaMP but using more radial lines (25).
Discussion and Conclusion {#sec:conc}
=========================
In this work we presented new pursuits for the cosparse analysis model. A theoretical study of these algorithms was performed giving guarantees for stable recovery under the assumptions of the $\OM$-RIP and the existence of an optimal or a near optimal projection. We showed that optimal projections exists for some non-trivial operators, i.e., operators that do not take us back to the synthesis case. In addition, we showed experimentally that using simpler kind of projections is possible in order to get good reconstruction results. We demonstrated both in the theoretical and the empirical results that linear dependencies within the analysis dictionary are favorable and enhance the recovery performance.
We are aware that there are still some open questions in this work and we leave them for future research. This should deal with following:
- Our work assumed the existence of a procedure that finds a cosupport that implies a near optimal projection with a constant $C_\cosp$. Two examples for optimal cosupport slection schemes were given. However, the existence of an optimal or a near optimal scheme for a general operator is still an open question. The question is: for which types of $\OM$ and values of $C_\cosp$ we can find an efficient procedure that implies a near optimal projection.
- As we have seen in the simulations, the thresholding procedure, though not near optimal with the theorems required constants, provides good reconstruction results. A theoretical study of the analysis greedy-like techniques with this cosupport selection scheme is required.
- A family of analysis dictionaries that deserves a special attention is the family of tight frame operators. In synthesis, there is a parallel between the guarantees of $\ell_1$-synthesis and the greedy like algorithms. The fact that a guarantee with a tight frame $\OM$ exists for $\ell_1$-analysis encourage us to believe that similar guarantees exist also for the analysis greedy-like techniques.
- In this paper, the noise $\e$ was considered to be adversarial. Random white Gaussian case was considered for the synthesis case in [@Giryes12RIP] resulting with near-oracle performance guarantees. It would be interesting to verify whether this is also the case for the analysis framework.
Proofs of Theorem \[thm:analysis\_RIP\_cond\] and Theorem \[thm:analysis\_RIP\_cond\_dependencies\] {#sec:analysis_RIP_proof}
===================================================================================================
[*Theorem \[thm:analysis\_RIP\_cond\] (Theorem 3.3 in [@Blumensath09Sampling]):*]{} Let ${{\bfM}}\in {\mathbb{R}^{m \times d}}$ be a random matrix that satisfies that for any ${{\bfz}}\in {\mathbb{R}^{d}}$ and $0<\tilde{\epsilon} \le \frac{1}{3}$ $$\begin{aligned}
P\left({\left\vert{\left\Vert{{\bfM}}{{\bfz}}\right\Vert}_2^2 - {\left\Vert{{\bfz}}\right\Vert}_2^2\right\vert} \ge \tilde{\epsilon}{\left\Vert{{\bfz}}\right\Vert}_2^2\right) \le e^{-\frac{C_{{{\bfM}}}m\tilde{\epsilon}}{2}},\end{aligned}$$ where $C_{{{\bfM}}} >0$ is a constant. For any value of $\epsilon_r >0$, if $$\begin{aligned}
m \ge \frac{32}{C_M\epsilon_r^2}\left( \log({\left\vert\L_{r}^{\text{corank}}\right\vert}) + (d-r)\log({9}/{\epsilon_r})+t\right),\end{aligned}$$ then $\delta_{r}^{\text{corank}} \le \epsilon_r$ with probability exceeding $1-e^{-t}$.
[*Theorem \[thm:analysis\_RIP\_cond\_dependencies\]:*]{} Under the same setup of Theorem \[thm:analysis\_RIP\_cond\], for any $\epsilon_\cosp > 0$ if $$\begin{aligned}
m \ge \frac{32}{C_M\epsilon_\cosp^2}\left( (p-\cosp)\log\left(\frac{9p}{(p-\cosp)\epsilon_\cosp}\right)+t \right),\end{aligned}$$ then $\delta_\cosp \le \epsilon_\cosp$ with probability exceeding $1-e^{-t}$.
[*Proof:*]{} Let $\tilde{\epsilon} = \epsilon_{r}/4$, $B^{\sdim-r} = \{{{\bfz}}\in {\mathbb{R}^{d-r}},{\left\Vert{{\bfz}}\right\Vert}_2\le 1\}$ and $\Psi$ an $\tilde{\epsilon}$-net for $B^{\sdim-r}$ with size ${\left\vert\Psi\right\vert} \le \left(1+ \frac{2}{\tilde{\epsilon}} \right)^{\sdim-r}$ [@Mendelson08Uniform]. For any subspace $\aspace_{\Lambda}^B = \aspace_{\Lambda} \cap B^{d-r}$ such that $\Lambda \in \L_{r}^{\text{corank}}$ we can build an orthogonal matrix ${{\bfU}}_{{{\bf\Omega}}} \in {\mathbb{R}^{\sdim\times (\sdim - r)}}$ such that $\aspace_{\Lambda}^B = \{{{\bfU}}_{\Lambda}{{\bfz}}, {{\bfz}}\in {\mathbb{R}^{\sdim-r}}, {\left\Vert{{\bfz}}\right\Vert}_2\le 1 \} = {{\bfU}}_{\Lambda}B^{\sdim-r}$. It is easy to see that $\Psi_{\Lambda}={{\bfU}}_\Lambda \Psi^{\sdim-r}$ is an $\tilde{\epsilon}$-net for $W_{\Lambda}^B$ and that $\Psi_{\A_{r}^{\text{corank}}} =\cup_{\Lambda \in \L_{r}^{\text{corank}}} \Psi_\Lambda$ is an $\tilde{\epsilon}$-net for $\A_{r}^{\text{corank}}\cap B^{\sdim}$, where ${\left\vert\Psi_{\A_{r}^{\text{corank}}}\right\vert} \le {\left\vert\L_{r}^{\text{corank}}\right\vert}(1+\frac{2}{\tilde{\epsilon}})^{\sdim-r}$.
We could stop here and use directly Theorem 2.1 from [@Mendelson08Uniform] to get the desired result for Theorem \[thm:analysis\_RIP\_cond\]. However, we present the remaining of the proof using a proof technique from [@Baraniuk08Simple; @Rauhut08Compressed]. Using union bound and the properties of ${{\bfM}}$ we have that with probability exceeding $1- {\left\vert\L_{r}^{\text{corank}}\right\vert}(1+\frac{2}{\tilde{\epsilon}})^{\sdim-r}e^{-\frac{C_\M m\tilde{\epsilon}^2}{2}}$ every ${{\bfv}} \in \Psi_{\A_{r}^{\text{corank}}}$ satisfies $$\begin{aligned}
\label{eq:epsilon_net_JL}
(1- \tilde{\epsilon}){\left\Vert{{\bfv}}\right\Vert}_2^2 \le {\left\Vert{{\bfM}}{{\bfv}}\right\Vert}_2^2 \le (1+\tilde{\epsilon}){\left\Vert{{\bfv}}\right\Vert}_2^2.\end{aligned}$$ According to the definition of $\delta_{r}^{\text{corank}}$ it holds that $\sqrt{1+\delta_{r}^{\text{corank}}} = \sup_{{{\bfv}}\in \A_{r}^{\text{corank}}\cap B^d}{\left\Vert{{\bfM}}{{\bfv}}\right\Vert}_2$. Since $\A_{r}^{\text{corank}} \cap B^\sdim$ is a compact set there exists ${{\bfv}}_0 \in \A_{r}^{\text{corank}} \cap B^\sdim$ that achieves the supremum. Denoting by $\tilde{{{\bfv}}}$ its closest vector in $\Psi_{\A_{r}^{\text{corank}}}$ and using the definition of $\Psi_{\A_{r}^{\text{corank}}}$ we have ${\left\Vert{{\bfv}}_0 - \tilde{{{\bfv}}}\right\Vert}_2 \le \tilde{\epsilon}$. This yields $$\begin{aligned}
\label{eq:Omega_RIP_epsilon_stage}
\sqrt{1+\delta_{r}^{\text{corank}}} = {\left\Vert{{\bfM}}{{\bfv}}_0\right\Vert}_2 &\le& {\left\Vert{{\bfM}}\tilde{{{\bfv}}}\right\Vert}_2 + {\left\Vert{{\bfM}}({{\bfv}}_0-\tilde{{{\bfv}}})\right\Vert}_2
\\ \nonumber &\le& \sqrt{1+\tilde{\epsilon}} + {\left\Vert{{\bfM}}\frac{{{\bfv}}_0-\tilde{{{\bfv}}}}{{\left\Vert{{\bfv}}_0-\tilde{{{\bfv}}}\right\Vert}}_2\right\Vert}_2{\left\Vert{{\bfv}}_0-\tilde{{{\bfv}}}\right\Vert}_2
\le \sqrt{1+\tilde{\epsilon}} + \sqrt{1+\delta_{r}^{\text{corank}}}\tilde{\epsilon}.\end{aligned}$$ The first inequality is due to the triangle inequality; the second one follows from and arithmetics; and the last inequality follows from the definition of $\delta_{r}^{\text{corank}}$, the properties of $\tilde{\epsilon}$-net and the fact that ${\left\Vert\frac{{{\bfv}}_0-\tilde{{{\bfv}}}}{{\left\Vert{{\bfv}}_0-\tilde{{{\bfv}}}\right\Vert}_2}\right\Vert}_2 =1$. Reordering gives $$\begin{aligned}
\label{eq:delta_l_epsilon_l_ineq}
1+\delta_{r}^{\text{corank}} \le \frac{1+\tilde{\epsilon}}{(1-\tilde{\epsilon})^2} \le 1 + 4\tilde{\epsilon} = 1+\epsilon_{r}.\end{aligned}$$ where the inequality holds because $\epsilon_{r}\le0.5$ and $\tilde{\epsilon} = \frac{\epsilon_{r}}{4} \le \frac{1}{8}$. Since we want to hold with probability greater than $1-e^{-t}$ it remains to require ${\left\vert\L_{r}^{\text{corank}}\right\vert}(1+\frac{8}{\epsilon_r})^{\sdim-r}e^{-\frac{C_M m\epsilon_r^2}{32}} \le e^{-t}$. Using the fact that $(1+\frac{8}{\epsilon_r})\ge \frac{9}{\epsilon_r}$ and some arithmetics we get and this completes the proof of the theorem.
We turn now to the proof of Theorem \[thm:analysis\_RIP\_cond\_dependencies\]. Its proof is almost identical to the previous proof but with the difference that instead of $r$, $\L_{r}^{\text{corank}}$ and $\delta_{r}^{\text{corank}}$ we look at $\cosp$, $\L_{\cosp}$ and $\delta_{\cosp}$. In this case we do not know what is the dimension of the subspace that each cosupport implies. However, we can have a lower bound on it using $\pdim - \cosp$. Therefore, we use $B^{\pdim- \cosp}$ instead of $B^{\sdim-r}$. This change provides us with a condition similar to but with $\pdim-\cosp$ in the second coefficient instead of $\sdim-r$. By using some arithmetics, noticing that the size of $\L_{\cosp}$ is $\pdim \choose \cosp$ and using Stirling’s formula for upper bounding it we get and this completes the proof.
Proof of Lemma \[lem:AIHT\_lemma\] {#sec:AIHT_lemma_proof}
==================================
[*Lemma \[lem:AIHT\_lemma\]:*]{} Consider the problem $\cal P$ and apply either AIHT or AHTP with a constant step size $\mu$ satisfying $\frac{1}{\mu} \ge 1+\delta_{2\cosp - \pdim}$ or an optimal step size. Then, at the $t$-th iteration, the following holds: $$\begin{aligned}
\label{eq:AIHT_lemma}
&&{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2
\le C_\ell\left({\left\Vert\y-\M\x\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2\right)\\
&& \nonumber ~~~~ + C_\ell \left(\frac{1}{\mu(1-\deltaB)}-1\right){\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2^2
+ (C_\ell-1)\mu\sigma_\M^2 {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2.\end{aligned}$$ For the optimal step size the bound is achieved with the value $\mu = \frac{1}{1+\deltaB}$.
[*Proof:*]{} We consider the AIHT algorithm first. We take similar steps to those taken in the proof of Lemma 3 in [@Blumensath09Sampling]. Since $\frac{1}{\mu} \ge 1+\deltaB$, we have, from the $\OM$-RIP property of $\M$, $${\left\Vert\M(\hat\x^t-\hat\x^{t-1})\right\Vert}_2^2 \le \frac{1}{\mu}{\left\Vert\hat\x^t-\hat\x^{t-1}\right\Vert}_2^2.$$ Thus, $$\begin{aligned}
&&{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2
= -2{\langle \M(\hat\x^t-\hat\x^{t-1}), \y-\M\hat\x^{t-1} \rangle} + {\left\Vert\M(\hat\x^t-\hat\x^{t-1})\right\Vert}_2^2 \\
&&~~~~~~~~~~\le -2{\langle \M(\hat\x^t-\hat\x^{t-1}), \y-\M\hat\x^{t-1} \rangle} + \frac{1}{\mu}{\left\Vert\hat\x^t-\hat\x^{t-1}\right\Vert}_2^2 \\
&&~~~~~~~~~~= -2{\langle \hat\x^t-\hat\x^{t-1}, \M^*(\y-\M\hat\x^{t-1}) \rangle} + \frac{1}{\mu}{\left\Vert\hat\x^t-\hat\x^{t-1}\right\Vert}_2^2 \\
&&~~~~~~~~~~= -\mu{\left\Vert\M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2 + \frac{1}{\mu}{\left\Vert\hat\x^t-\hat\x^{t-1}-\mu \M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Note that by definition, $\hat\x^t = \Q_{\hat \CF_\cosp}\left(\hat\x^{t-1}+\mu \M^*(\y-\M\hat\x^{t-1})\right)$. Hence, by the $C_\cosp$-near optimality of the projection, we get $$\label{eq:aihtlemlast}
{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2
\le -\mu{\left\Vert\M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2 + \frac{C_\ell}{\mu}{\left\Vert\x -\hat\x^{t-1}-\mu \M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2.$$ Now note that $$\begin{aligned}
&& {\left\Vert\x -\hat\x^{t-1}-\mu \M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2 \\
&&~~~~~= {\left\Vert\x -\hat\x^{t-1}\right\Vert}_2^2 - 2\mu{\langle \M(\x-\hat\x^{t-1}), \y-\M\hat\x^{t-1} \rangle} + \mu^2 {\left\Vert\M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2 \\
&&~~~~~ \le \frac{1}{1-\deltaB} {\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2^2 - 2\mu{\langle \M(\x-\hat\x^{t-1}), \y-\M\hat\x^{t-1} \rangle} + \mu^2 {\left\Vert\M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2 \\
&&~~~~~ = \frac{1}{1-\deltaB} {\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2^2 + \mu\left({\left\Vert\y-\M\x\right\Vert}_2^2 - {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 - {\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2^2\right) \\
&&~~~~~~~~~ + \mu^2 {\left\Vert\M^*(\y-\M\hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Putting this into , we obtain the desired result for the AIHT algorithm.
We can check that the same holds true for the AHTP algorithm as follows: suppose that $\hat\x^{t-1}_{\text{\tiny AHTP}}$ is the $(t-1)$-st estimate from the AHTP algorithm. If we now initialize the AIHT algorithm with this estimate and obtain the next estimate $\hat\x^t_{\tilde{\text{\tiny AIHT}}}$, then the inequality of the lemma holds true with $\hat\x^t_{\tilde{\text{\tiny AIHT}}}$ and $\hat\x^{t-1}_{\text{\tiny AHTP}}$ in place of $\hat\x^t$ and $\hat\x^{t-1}$ respectively. On the other hand, from the algorithm description, we know that the $t$-th estimate $\hat\x^t_{\text{\tiny AHTP}}$ of the AHTP satisfies $${\left\Vert\y-\M\hat\x^t_{\text{\tiny AHTP}}\right\Vert}_2^2 \le {\left\Vert\y-\M\hat\x^t_{\tilde{\text{\tiny AIHT}}}\right\Vert}_2^2.$$ This means that the result holds for the AHTP algorithm as well.
Using a similar argument for the optimal changing step size we note that it selects the cosupport that minimizes ${\left\Vert\M\x - \M{\hat\x}^{t}\right\Vert}_2^2$. Thus, for AIHT and AHTP we have that ${\left\Vert\M\x - \M{\hat\x_{\text{\tiny Opt}}}^{t}\right\Vert}_2^2 \le {\left\Vert\M\x - \M{\hat\x}_{\mu}^{t}\right\Vert}_2^2$ for any value of $\mu$, where $\hat\x_{\text{\tiny Opt}}^{t}$ and $\hat\x_{\mu}^{t}$ are the recovery results of AIHT or AHTP with an optimal changing step-size and a constant step-size $\mu$ respectively. This yields that any theoretical result for a constant step-size selection with a constant $\mu$ holds true also to the optimal changing-step size selection. In particular this is true also for $\mu = \frac{1}{1+\deltaB}$. This choice is justified in the proof of Lemma \[lem:AIHT\_AHTP\_lemma\_noisy\]. $\Box$
Proof of Lemma \[lem:AIHT\_AHTP\_lemma\_noisy\]
===============================================
[*Lemma \[lem:AIHT\_AHTP\_lemma\_noisy\]:*]{} Suppose that the same conditions of Theorem \[thm:AIHT\_AHTP\_theorem\_noisy\_iter\] hold true. If ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2$, then ${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2$. Furthermore, if ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 > \eta^2{\left\Vert\e\right\Vert}_2^2$, then $${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 \le c_4 {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2$$ where $$c_4 := \left(1+\frac{1}{\eta}\right)^2\left(\frac{1}{\mu(1-\deltaB)} - 1\right)C_\cosp + (C_\ell -1)(\mu\sigma_\M^2 -1) + \frac{C_\cosp}{\eta^2} < 1.$$
[*Proof:*]{} First, suppose that ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 > \eta^2{\left\Vert\e\right\Vert}_2^2$. From Lemma \[lem:AIHT\_lemma\], we have $$\begin{aligned}
\label{eq:aihtsecond}
{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2
&\le& C_\cosp{\left\Vert\y-\M\x\right\Vert}_2^2 + (C_\cosp - 1) (\mu\sigma_\M^2 - 1) {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2
\\ \nonumber && + C_\cosp \left(\frac{1}{\mu(1-\deltaB)} - 1\right) {\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Remark that all the coefficients in the above are positive because $1 + \deltaB \le \frac{1}{\mu} \le \sigma_\M^2$ and $C_\cosp \ge 1$. Since $\y-\M\x = \e$, we note $${\left\Vert\y-\M\x\right\Vert}_2^2 < \frac{1}{\eta^2}{\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2$$ and, by the triangle inequality, $${\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2 \le {\left\Vert\y-\M\x\right\Vert}_2 + {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2
< \left( 1 + \frac{1}{\eta} \right) {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2.$$ Therefore, from , $${\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 < c_4 {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2.$$ This is the second part of the lemma.
Now, suppose that ${\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2^2 \le \eta^2{\left\Vert\e\right\Vert}_2^2$. This time we have $${\left\Vert\M(\x-\hat\x^{t-1})\right\Vert}_2 \le {\left\Vert\y-\M\x\right\Vert}_2 + {\left\Vert\y-\M\hat\x^{t-1}\right\Vert}_2
\le (1+\eta) {\left\Vert\e\right\Vert}_2.$$ Applying this to , we obtain $$\begin{aligned}
{\left\Vert\y-\M\hat\x^t\right\Vert}_2^2 &\le
C_\cosp {\left\Vert\e\right\Vert}_2^2 + (C_\ell - 1) (\mu\sigma_\M^2 - 1) \eta^2 {\left\Vert\e\right\Vert}_2^2
+ C_\cosp \left(\frac{1}{\mu(1-\deltaB)} - 1\right) (1+\eta)^2 {\left\Vert\e\right\Vert}_2^2 \\
&= \left(C_\cosp + (C_\cosp - 1) (\mu\sigma_M^2 - 1) \eta^2 + C_\cosp \left(\frac{1}{\mu(1-\deltaB)} - 1\right) (1+\eta)^2 \right) {\left\Vert\e\right\Vert}_2^2 =c_4 \eta^2 {\left\Vert\e\right\Vert}_2^2.\end{aligned}$$ Thus, the proof is complete as soon as we show $c_4 < 1$, or $c_4-1 < 0$.
To see $c_4 -1 < 0$, we first note that it is equivalent to–all the subscripts are dropped from here on for simplicity of notation– $$\frac{1}{\mu^2} - \frac{2(1-\delta)}{1+\frac{1}{\eta}} \frac{1}{\mu}
+ \frac{(C-1)\sigma^2 (1-\delta)}{C\left(1+\frac{1}{\eta}\right)^2} < 0,$$ or $$\frac{1}{\mu^2} - 2 (1-\delta) b_1 \frac{1}{\mu} + (1-\delta)^2 b_2 < 0.$$ Solving this quadratic equation in $\frac{1}{\mu}$, we want $$(1-\delta) \left(b_1 - \sqrt{b_1^2 - b_2}\right) < \frac{1}{\mu} <
(1-\delta) \left(b_1 + \sqrt{b_1^2 - b_2}\right).$$ Such $\mu$ exists only when $\frac{b_2}{b_1^2} < 1$. Furthermore, we have already assumed $1 + \delta \le \frac{1}{\mu}$ and we know $(1-\delta) \left(b_1 - \sqrt{b_1^2 - b_2}\right) < 1+\delta$, and hence the condition we require is $$1 + \delta \le \frac{1}{\mu} <
(1-\delta) \left(b_1 + \sqrt{b_1^2 - b_2}\right),$$ which is what we desired to prove.
As we have seen in Lemma \[lem:AIHT\_lemma\], for changing optimal step-size selection, holds for any value of $\mu$ that satisfies the above conditions. Thus, in the bound of changing optimal step-size we put a value of $\mu$ that minimizes $c_4$. This minimization result with $\frac{1}{\mu}=\sqrt{b_2}(1-\deltaB)$. However, since we need $\frac{1}{\mu} \ge 1+\deltaB$ and have that $\sqrt{b_2}(1-\deltaB) < b_1(1-\deltaB) <1+\deltaB$ we set $\frac{1}{\mu} = 1+\deltaB$ in $c_4$ for the bound in optimal changing step-size case. $\Box$
Proof of Lemma \[lem:ACoSaMP\_xp\_bound\] {#sec:ACoSaMP_xp_bound_proof}
=========================================
[*Lemma \[lem:ACoSaMP\_xp\_bound\]:*]{} Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. For each iteration we have $$\begin{aligned}
{\left\Vert\x - {{\bfw}}\right\Vert}_2 &\le& \frac{1}{\sqrt{1-\deltaD^2}}{\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}} )\right\Vert}_2+ \frac{\sqrt{1+\deltaC}}{1-\deltaD}{\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$
[*Proof:*]{} Since ${{\bfw}}$ is the minimizer of ${\left\Vert{{\bfy}} - {{\bfM}}{{\bfv}}\right\Vert}^2_2$ with the constraint ${{\bf\Omega}}_{\tilde{\Lambda}^t}{{\bfv}} =0$, then $$\begin{aligned}
\langle {{\bfM}} {{\bfw}} - {{\bfy}}, {{\bfM}}{{\bfu}} \rangle =0,\end{aligned}$$ for any vector ${{\bfu}}$ such that ${{\bf\Omega}}_{\tilde{\Lambda}^t}{{\bfu}} =0 $. Substituting ${{\bfy}} = {{\bfM}}{{\bfx}} + {{\bfe}}$ and moving terms from the LHS to the RHS gives $$\begin{aligned}
\label{eq:xp_x_property}
\langle {{\bfw}} - {{\bfx}}, {{\bfM}}^*{{\bfM}}{{\bfu}} \rangle = \langle {{\bfe}}, {{\bfM}}{{\bfu}} \rangle,\end{aligned}$$ where ${{\bfu}}$ is a vector satisfying ${{\bf\Omega}}_{\tilde{\Lambda}^t}{{\bfu}} =0$. Turning to look at ${\left\Vert\Q_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2^2$ and using with ${{\bfu}} = \Q_{\tilde{\Lambda}^t}(\x - {{\bfw}})$, we have $$\begin{aligned}
\label{eq:Q_xp_x_norm}
&& \hspace{-0.3in} {\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2^2 = \langle {{\bfx}} - {{\bfw}}, \Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}}) \rangle \\
\nonumber && \hspace{-0.3in} = \langle {{\bfx}} - {{\bfw}}, ({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}}) \rangle - \langle {{\bfe}}, {{\bfM}}\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}}) \rangle
\\ \nonumber && \hspace{-0.3in} \le {\left\Vert {{\bfx}} - {{\bfw}}\right\Vert}_2 {\left\Vert\Q_{\Lambda \cap \tilde{\Lambda}^t} ({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\tilde{\Lambda}^t}\right\Vert}_2 {\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2 + {\left\Vert {{\bfe}}\right\Vert}_2{\left\Vert {{\bfM}}\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2
\\ \nonumber && \hspace{-0.3in} \le \deltaD{\left\Vert {{\bfx}} - {{\bfw}}\right\Vert}_2 {\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2 + {\left\Vert {{\bfe}}\right\Vert}_2\sqrt{1+\deltaC}{\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2.\end{aligned}$$ where the first inequality follows from the Cauchy-Schwartz inequality, the projection property that $\Q_{\tilde{\Lambda}^t} = \Q_{\tilde{\Lambda}^t}\Q_{\tilde{\Lambda}^t}$ and the fact that ${{\bfx}} - {{\bfw}} = \Q_{\Lambda \cap \tilde{\Lambda}^t }({{\bfx}} - {{\bfw}})$. The last inequality is due to the ${{\bf\Omega}}$-RIP properties, Corollary \[cor:omega\_RIP\_norm\_diff\] and that according to Table \[tbl:Synthesis\_Analysis\_parallels\] $|{\tilde{\Lambda}^t}| \ge 3\cosp - 2\pdim$ and $|{\Lambda\cap\tilde{\Lambda}^t}| \ge 4\cosp - 3\pdim$. After simplification of by ${\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2$ we have $$\begin{aligned}
\nonumber {\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2 \le \deltaD{\left\Vert{{\bfx}} - {{\bfw}}\right\Vert}_2 + \sqrt{1+\deltaC}{\left\Vert{{\bfe}}\right\Vert}_2.\end{aligned}$$ Utilizing the last inequality with the fact that ${\left\Vert{{\bfx}} - {{\bfw}}\right\Vert}_2^2 = {\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2^2+ {\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2^2$ gives $$\begin{aligned}
&& {\left\Vert{{\bfx}} - {{\bfw}}\right\Vert}_2^2 \le {\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - {{\bfw}})\right\Vert}_2^2 + \left(\deltaD{\left\Vert{{\bfx}} - {{\bfw}}\right\Vert}_2 + \sqrt{1+\deltaC}{\left\Vert{{\bfe}}\right\Vert}_2 \right)^2.\end{aligned}$$ By moving all terms to the LHS we get a quadratic function of ${\left\Vert\x-{{\bfw}}\right\Vert}_2$. Thus, ${\left\Vert\x-{{\bfw}}\right\Vert}_2$ is bounded from above by the larger root of that function; this with a few simple algebraic steps gives the inequality in . $\Box$
Proof of Lemma \[lem:ACoSaMP\_xt\_bound1\] {#sec:ACoSaMP_xt_bound1_proof}
==========================================
[*Lemma \[lem:ACoSaMP\_xt\_bound1\]:*]{} Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. For each iteration we have $$\begin{aligned}
&& \hspace{-0.5in} {\left\Vert\x - \hat\x^t\right\Vert}_2 \le \rho_1{\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2 + \eta_1{\left\Vert{{\bfe}}\right\Vert}_2,\end{aligned}$$ where $\eta_1$ and $\rho_1$ are the same constants as in Theorem \[thm:ACoSaMP\_iter\_bound\].
[*Proof:*]{} We start with the following observation $$\begin{aligned}
\label{eq:xt_x_diff_norm}
&& \hspace{-0.3in} {\left\Vert{{\bfx}} -\hat\x^t\right\Vert}_2^2 = {\left\Vert{{\bfx}} -{{\bfw}} + {{\bfw}}-\hat\x^t\right\Vert}_2^2
= {\left\Vert{{\bfx}}-{{\bfw}}\right\Vert}_2^2 + {\left\Vert\hat\x^t - {{\bfw}}\right\Vert}_2^2 +2({{\bfx}} - {{\bfw}})^*({{\bfw}} - \hat\x^t),\end{aligned}$$ and turn to bound the second and last terms in the RHS. For the second term, using the fact that $\hat\x^t = \Q_{\hat\CF_\cosp({{\bfw}})}{{\bfw}}$ with gives $$\begin{aligned}
\label{eq:xt_x_diff_norm_coeff1}
{\left\Vert\hat\x^t - {{\bfw}}\right\Vert}_2^2 \le C_\cosp{\left\Vert\x -{{\bfw}}\right\Vert}_2^2.\end{aligned}$$ For bounding the last term, we look at its absolute value and use with ${{\bfu}} = {{\bfw}} - \hat\x^t = \Q_{\tilde{\Lambda}^t}({{\bfw}} - \hat\x^t)$. This leads to $$\begin{aligned}
&& \hspace{-0.3in}{\left\vert({{\bfx}} - {{\bfw}})^*({{\bfw}} - \hat\x^t)\right\vert} = {\left\vert({{\bfx}} - {{\bfw}})^*({{\bfI}} - {{\bfM}}^*{{\bfM}})({{\bfw}} - \hat\x^t) - {{\bfe}}^*{{\bfM}}({{\bfw}} - \hat\x^t)\right\vert}.\end{aligned}$$ By using the triangle and Cauchy-Schwartz inequalities with the fact that ${{\bfx}}-{{\bfw}} = \Q_{\Lambda \cap \tilde{\Lambda}^t}({{\bfx}}-{{\bfw}})$ and ${{\bfw}}-\hat\x^t = \Q_{\tilde{\Lambda}^t}({{\bfw}}-\hat\x^t)$ we have $$\begin{aligned}
\label{eq:xt_x_diff_norm_coeff3}
&& \hspace{-0.55in} {\left\vert({{\bfx}} - {{\bfw}})^*({{\bfw}} - \hat\x^t)\right\vert}
\le {\left\Vert{{\bfx}} - {{\bfw}}\right\Vert}_2{\left\Vert\Q_{\Lambda \cap \tilde{\Lambda}^t}({{\bfI}} - {{\bfM}}^*{{\bfM}})\Q_{\tilde{\Lambda}^t}\right\Vert}_2{\left\Vert{{\bfw}} - \hat\x^t\right\Vert}_2
+ {\left\Vert{{\bfe}}\right\Vert}_2{\left\Vert{{\bfM}}({{\bfw}} - \hat\x^t)\right\Vert}_2
\\ \nonumber && \hspace{0.83in} \le \deltaD{\left\Vert{{\bfx}}-{{\bfw}}\right\Vert}_2{\left\Vert{{\bfw}} - \hat\x^t\right\Vert}_2
+\sqrt{1+\deltaC}{\left\Vert{{\bfe}}\right\Vert}_2{\left\Vert{{\bfw}} - \hat\x^t\right\Vert}_2,\end{aligned}$$ where the last inequality is due to the ${{\bf\Omega}}$-RIP definition and Corollary \[cor:omega\_RIP\_norm\_diff\].
By substituting and into we have $$\begin{aligned}
\label{eq:xt_x_diff_norm_all}
&& \hspace{-0.3in}{\left\Vert\x -\hat\x^t\right\Vert}_2^2 \le (1+C_\cosp){\left\Vert{{\bfx}} -{{\bfw}}\right\Vert}_2^2
+ 2\deltaD\sqrt{C_\cosp}{\left\Vert{{\bfx}}-{{\bfw}}\right\Vert}_2^2 +2\sqrt{1+\deltaC}\sqrt{C_\cosp}{\left\Vert{{\bfe}}\right\Vert}_2{\left\Vert\x - {{\bfw}}\right\Vert}_2
\\ \nonumber && \le \Big((1 + 2\deltaD\sqrt{C_\cosp} +C_\cosp){\left\Vert{{\bfx}} -{{\bfw}}\right\Vert}_2
+2\sqrt{(1+\deltaC)C_\cosp}{\left\Vert{{\bfe}}\right\Vert}_2\Big) {\left\Vert{{\bfx}} - {{\bfw}}\right\Vert}_2 \\ \nonumber
&& \le \frac{1 + 2\deltaD\sqrt{C_\cosp} +C_\cosp}{1-\deltaD^2}{\left\Vert\P_{\tilde{\Lambda}^t}(\x -{{\bfw}} )\right\Vert}_2^2 \\ \nonumber &&+ \frac{2\sqrt{1+\deltaC}(1 +(1+\deltaD)\sqrt{C_\cosp}+C_\cosp)}{(1-\deltaD)\sqrt{1-\deltaD^2}}{\left\Vert\P_{\tilde{\Lambda}^t}(\x-{{\bfw}} )\right\Vert}_2{\left\Vert{{\bfe}}\right\Vert}_2 + \frac{(1+\deltaC)(1 +2\sqrt{C_\cosp}+C_\cosp)}{(1-\deltaD)^2}{\left\Vert{{\bfe}}\right\Vert}_2^2
\\ \nonumber && \le \Bigg(\frac{\sqrt{1+2\deltaD\sqrt{C_\cosp} + C_\cosp}}{\sqrt{1-\deltaD^2}}
{\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}} )\right\Vert}_2 +\frac{\sqrt{\frac{2+C_\cosp}{1+C_\cosp}+2\sqrt{C_\cosp}+C_\cosp}\sqrt{1+\deltaC}}{1-\deltaD}{\left\Vert{{\bfe}}\right\Vert}_2 \Bigg)^2,\end{aligned}$$ where for the second inequality we use the fact that $\deltaD \le 1$ combined with the inequality of Lemma \[lem:ACoSaMP\_xp\_bound\], and for the last inequality we use the fact that $(1 +(1+\deltaD)\sqrt{C_\cosp}+C_\cosp)^2 \le (1 + 2\deltaD\sqrt{C_\cosp} +C_\cosp)(\frac{2+C_\cosp}{1+C_\cosp}+2\sqrt{C_\cosp}+C_\cosp)$ together with a few algebraic steps. Taking square-root on both sides of provides the desired result. $\Box$
Proof of Lemma \[lem:ACoSaMP\_Pxp\_bound\] {#sec:ACoSaMP_Pxp_bound_proof}
==========================================
[*Lemma \[lem:ACoSaMP\_Pxp\_bound\]:*]{} Consider the problem $\cal P$ and apply ACoSaMP with $a = \frac{2\cosp - \pdim}{\cosp}$. if $$\begin{aligned}
C_{2\cosp-p} < \frac{\sigma_{{{\bfM}}}^2(1+\gamma)^2}{\sigma_{{{\bfM}}}^2(1+\gamma)^2-1},\end{aligned}$$ then there exists $\tilde{\delta}_{\text{ \tiny ACoSaMP}}(C_{2\cosp - \pdim}, \sigma_{{{\bfM}}}^2, \gamma) >0$ such that for any $\deltaB < \tilde{\delta}_{\text{ \tiny ACoSaMP}}(C_{2\cosp - \pdim}, \sigma_{{{\bfM}}}^2, \gamma)$ $$\begin{aligned}
&& {\left\Vert\P_{\tilde{\Lambda}^t}(\x - {{\bfw}})\right\Vert}_2 \le
\eta_2{\left\Vert{{\bfe}}\right\Vert}_2
+ \rho_2{\left\Vert\x - \hat\x^{t-1}\right\Vert}_2.\end{aligned}$$ The constants $\eta_2$ and $\rho_2$ are as defined in Theorem \[thm:ACoSaMP\_iter\_bound\].
In the proof of the lemma we use the following Proposition.
[*Proposition E.1:*]{} For any two given vectors ${{\bfx}}_1$, ${{\bfx}}_2$ and any constant $c>0$ it holds that $$\begin{aligned}
\label{eq:norm2_ineq}
{\left\Vert{{\bfx}}_1+{{\bfx}}_2\right\Vert}_2^2 \le (1+c){\left\Vert{{\bfx}}_1\right\Vert}_2^2 + \left(1+\frac{1}{c}\right){\left\Vert{{\bfx}}_2\right\Vert}^2.\end{aligned}$$
The proof of the proposition is immediate using the inequality of arithmetic and geometric means. We turn to the proof of the lemma.
[*Proof:*]{} Looking at the step of finding new cosupport elements one can observe that $\Q_{\Lambda_{\Delta}}$ is a near optimal projection for ${{\bfM}}^*{{\bfy}}^{t-1}_{\text{resid}} = {{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})$ with a constant $C_{2\cosp-p}$. The fact that ${\left\vert{\hat\Lambda^{t-1} \cap \Lambda}\right\vert} \ge 2\cosp -p$ combined with gives $$\begin{aligned}
{\left\Vert({{\bfI}} - \Q_{{\Lambda}_{\Delta}}){{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2 \le C_{2\cosp -\pdim}{\left\Vert({{\bfI}} - \Q_{\hat\Lambda^{t-1} \cap \Lambda}){{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Using simple projection properties and the fact that $\tilde{\Lambda}^t \subseteq \Lambda_{\Delta}$ with ${{\bfz}} = {{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})$ we have $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq}
&& \hspace{-0.3in} {\left\Vert\Q_{\tilde{\Lambda}^t}{{\bfz}}\right\Vert}_2^2 \ge {\left\Vert\Q_{{\Lambda}_\Delta}{{\bfz}}\right\Vert}_2^2
={\left\Vert{{\bfz}}\right\Vert}_2^2 - {\left\Vert({{\bfI}}-\Q_{{\Lambda}_\Delta}){{\bfz}}\right\Vert}_2^2
\ge {\left\Vert{{\bfz}}\right\Vert}_2^2 - C_{2\cosp-p}{\left\Vert({{\bfI}}-\Q_{\hat\Lambda^{t-1} \cap \Lambda}){{\bfz}}\right\Vert}_2^2
\\ \nonumber && ={\left\Vert{{\bfz}}\right\Vert}_2^2 - C_{2\cosp-p}\left({\left\Vert{{\bfz}}\right\Vert}_2^2 - {\left\Vert\Q_{\hat\Lambda^{t-1} \cap \Lambda}{{\bfz}}\right\Vert}_2^2\right)
= C_{2\cosp-p}{\left\Vert\Q_{\hat\Lambda^{t-1} \cap \Lambda}{{\bfz}}\right\Vert}_2^2- (C_{2\cosp-p}-1){\left\Vert{{\bfz}}\right\Vert}_2^2.\end{aligned}$$
We turn to bound the LHS of from above. Noticing that $\y = \M\x+\e$ and using with a constant $\gamma_1>0$ gives $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_lhs_step1}
{\left\Vert\Q_{\tilde{\Lambda}^t}{{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2 \le
\left(1+\frac{1}{\gamma_1}\right){\left\Vert\Q_{\tilde{\Lambda}^t}{{\bfM}}^*{{\bfe}}\right\Vert}_2^2 + (1+\gamma_1){\left\Vert\Q_{\tilde{\Lambda}^t}{{\bfM}}^*{{\bfM}}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Using again, now with a constant $\alpha>0$, we have $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_lhs_step2}
&&\hspace{-0.32in}{\left\Vert\Q_{\tilde{\Lambda}^t}{{\bfM}}^*{{\bfM}}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
\le (1+\alpha){\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
+\left(1+\frac{1}{\alpha}\right){\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfI}}-{{\bfM}}^*{{\bfM}})({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
\\ \nonumber && \le (1+\alpha){\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2 - (1+\alpha){\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
+\left(1+\frac{1}{\alpha}\right){\left\Vert\Q_{\tilde{\Lambda}^t}({{\bfI}}-{{\bfM}}^*{{\bfM}})({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Putting into and using and Corollary \[cor:MQ\_RIP\_norm\] gives $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_lhs}
&& \hspace{-0.5in} {\left\Vert\Q_{\tilde{\Lambda}^t}{{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2
\le \frac{(1+\gamma_1)(1+\delta_{3\cosp-2p})}{\gamma_1}{\left\Vert{{\bfe}}\right\Vert}_2^2
- (1+\alpha)(1+\gamma_1){\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
\\ \nonumber && ~~~~~~~~~~~~~~~~~~~~~~~ +\left(1+\alpha +\delta_{4\cosp-3p} + \frac{\delta_{4\cosp-3p}}{\alpha}\right)(1+\gamma_1){\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2.\end{aligned}$$
We continue with bounding the RHS of from below. For the first element of the RHS we use an altered version of with a constant $\gamma_2>0$ and have $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_rhs1_step1}
{\left\Vert\Q_{{\hat\Lambda^{t-1} \cap \Lambda}}{{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2
\ge \frac{1}{1+\gamma_2}{\left\Vert\Q_{\hat\Lambda^{t-1} \cap \Lambda}{{\bfM}}^*{{\bfM}}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
-\frac{1}{\gamma_2}{\left\Vert\Q_{\hat\Lambda^{t-1} \cap \Lambda}{{\bfM}}^*{{\bfe}}\right\Vert}_2^2.\end{aligned}$$ Using the altered form again, for the first element in the RHS of , with a constant $\beta >0$ gives $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_rhs1_step2}
{\left\Vert\Q_{\hat\Lambda^{t-1} \cap \Lambda}{{\bfM}}^*{{\bfM}}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
\ge \frac{1}{1+\beta}{\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2
-\frac{1}{\beta}{\left\Vert\Q_{\hat\Lambda^{t-1} \cap \Lambda}({{\bfM}}^*{{\bfM}} - {{\bfI}})({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2.\end{aligned}$$ Putting in and using the RIP properties and provide $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_rhs1}
{\left\Vert\Q_{{\hat\Lambda^{t-1} \cap \Lambda}}{{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2
\ge \left(\frac{1}{1+\beta}-\frac{\delta_{2\cosp-p}}{\beta}\right)\frac{1}{1+\gamma_2}{\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2
-\frac{(1+\delta_{2\cosp-p})}{\gamma_2}{\left\Vert{{\bfe}}\right\Vert}_2^2.\end{aligned}$$ Using , with a constant $\gamma_3 >0$, , and some basic algebraic steps we have for the second element in the RHS of $$\begin{aligned}
\label{eq:ACoSaMP_QMy_Mx_ineq_rhs2}
\hspace{-0.3in} {\left\Vert{{\bfM}}^*({{\bfy}} - {{\bfM}}\hat\x^{t-1})\right\Vert}_2^2
&\le& (1+\gamma_3){\left\Vert{{\bfM}}^*{{\bfM}}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2 + \left(1+\frac{1}{\gamma_3}\right){\left\Vert{{\bfM}}^*{{\bfe}}\right\Vert}_2^2
\\ \nonumber &\le& (1+\gamma_3)(1+\delta_{2\cosp-p})\sigma_{{{\bfM}}}^2{\left\Vert({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2 + \left(1+\frac{1}{\gamma_3}\right)\sigma_{{{\bfM}}}^2{\left\Vert{{\bfe}}\right\Vert}_2^2.\end{aligned}$$
By combining , and with we have $$\begin{aligned}
&& \hspace{-0.15in}(1+\alpha)(1+\gamma_1){\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
\le \frac{(1+\gamma_1)(1+\delta_{3\cosp-2p})}{\gamma_1}{\left\Vert{{\bfe}}\right\Vert}_2^2
+C_{2\cosp-p}\frac{(1+\delta_{2\cosp-p})}{\gamma_2}{\left\Vert{{\bfe}}\right\Vert}_2^2
\\ \nonumber && + (C_{2\cosp-p}-1)\left(1+\frac{1}{\gamma_3}\right)\sigma_{{{\bfM}}}^2{\left\Vert{{\bfe}}\right\Vert}_2^2
+ \left(1+\alpha +\delta_{4\cosp-3p} + \frac{\delta_{4\cosp-3p}}{\alpha}\right)(1+\gamma_1){\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2
\\ \nonumber &&
+ (C_{2\cosp-p}-1)(1+\gamma_3)(1+\delta_{2\cosp-p})\sigma_{{{\bfM}}}^2{\left\Vert({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2
- C_{2\cosp-p}\left(\frac{1}{1+\beta}-\frac{\delta_{2\cosp-p}}{\beta}\right)\frac{1}{1+\gamma_2}{\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2.\end{aligned}$$ Dividing both sides by $(1+\alpha)(1+\gamma_1)$ and gathering coefficients give $$\begin{aligned}
&& \hspace{-0.3in}{\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2 \le
\bigg(\frac{1+\delta_{3\cosp-2p}}{\gamma_1(1+\alpha)}
+\frac{(1+\delta_{2\cosp-p})C_{2\cosp-p}}{\gamma_2(1+\alpha)(1+\gamma_1)}
+ \frac{(C_{2\cosp-p}-1)(1+\gamma_3)\sigma_{{{\bfM}}}^2}{(1+\alpha)(1+\gamma_1)\gamma_3}\bigg){\left\Vert{{\bfe}}\right\Vert}_2^2
\\ \nonumber && \hspace{-0.3in} ~~~~~~~~~~~~~~~~~~~~~~~~~~
+ \bigg(1 +\frac{\delta_{4\cosp-3p}}{\alpha} + \frac{(C_{2\cosp-p}-1)(1+\gamma_3)(1+\delta_{2\cosp-p})\sigma_{{{\bfM}}}^2}{(1+\alpha)(1+\gamma_1)}
\\ \nonumber && ~~~~~~~~~~~~~~~~~~~~~~~~~~ - \frac{C_{2\cosp-p}}{(1+\alpha)(1+\gamma_1)(1+\gamma_2)}\left(\frac{1}{1+\beta}-\frac{\delta_{2\cosp-p}}{\beta}\right)\bigg){\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2.\end{aligned}$$ The smaller the coefficient of ${\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2$, the better convergence guarantee we obtain. Thus, we choose $\beta = \frac{\sqrt{\delta_{2\cosp-p}}}{1- \sqrt{\delta_{2\cosp-p}}}$ and $ \small {\alpha = \frac{\sqrt{\delta_{4\cosp-3p}}}{\sqrt{\frac{C_{2\cosp-p}}{(1+\gamma_1)(1+\gamma_2)}\left(1-\sqrt{\delta_{2\cosp-p}}\right)^2 -\frac{(C_{2\cosp-p}-1)(1+\gamma_3)(1+\delta_{2\cosp-p})\sigma_{{{\bfM}}}^2}{1+\gamma_1}} - \sqrt{\delta_{4\cosp-3p}}}}$ so that the coefficient is minimized. The values of $\gamma_1, \gamma_2, \gamma_3$ provide a tradeoff between the convergence rate and the size of the noise coefficient. For smaller values we get better convergence rate but higher amplification of the noise. We make no optimization on their values and choose them to be $\gamma_1 = \gamma_2 = \gamma_3 = \gamma$ for an appropriate $\gamma>0$. Thus, the above yields $$\begin{aligned}
\label{eq:ACoSaMP_Pxp_bound_square}
&& \hspace{-0.3in}{\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2 \le
\Bigg(\frac{1+\delta_{3\cosp-2p}}{\gamma(1+\alpha)}
+\frac{(1+\delta_{2\cosp-p})C_{2\cosp-p}}{\gamma(1+\alpha)(1+\gamma)}
+ \frac{(C_{2\cosp-p}-1)(1+\gamma)\sigma_{{{\bfM}}}^2}{(1+\alpha)(1+\gamma)\gamma}\bigg){\left\Vert{{\bfe}}\right\Vert}_2^2
\\ \nonumber &&
+ \left(1 -\left(\sqrt{\delta_{4\cosp-3p}} -
\sqrt{\frac{C_{2\cosp-p}}{(1+\gamma)^2}\left(1-\sqrt{\delta_{2\cosp-p}}\right)^2 -(C_{2\cosp-p}-1)(1+\delta_{2\cosp-p})\sigma_{{{\bfM}}}^2}
\right)^2\right){\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2.\end{aligned}$$ Since $\P_{\tilde{\Lambda}^t}{{\bfw}} = \P_{\tilde{\Lambda}^t}\hat\x^{t-1} =0$ the above inequality holds also for ${\left\Vert\P_{\tilde{\Lambda}^t}({{\bfx}} - \hat\x^{t-1})\right\Vert}_2^2$. Inequality follows since the right-hand side of is smaller than the square of the right-hand side of .
Before ending the proof, we notice that $\rho_2$, the coefficient of ${\left\Vert{{\bfx}} - \hat\x^{t-1}\right\Vert}_2^2$ is defined only when $$\begin{aligned}
\label{eq:rho_2_delta_ineq}
(C_{2\cosp-p}-1)(1+\deltaB)\sigma_{{{\bfM}}}^2 \le \frac{C_{2\cosp-p}}{(1+\gamma)^2}\left(1-\sqrt{\deltaB}\right)^2.\end{aligned}$$ First we notice that since $1+\delta_{2\cosp-p} \ge \left(1-\sqrt{\delta_{2\cosp-p}}\right)^2$ a necessary condition for to hold is $(C_{2\cosp-p}-1)\sigma_{{{\bfM}}}^2 < \frac{C_{2\cosp-p}}{(1+\gamma)^2}$ which is equivalent to . By moving the terms in the RHS to the LHS we get a quadratic function of $\sqrt{\deltaB}$. The condition in guarantees that its constant term is smaller than zero and thus there exists a positive $\deltaB$ for which the function is smaller than zero. Therefore, for any $\deltaB < \tilde{\delta}_{\text{ \tiny ACoSaMP}}(C_{2\cosp - \pdim}, \sigma_{{{\bfM}}}^2, \gamma)$ holds, where $\tilde{\delta}_{\text{ \tiny ACoSaMP}}(C_{2\cosp - \pdim}, \sigma_{{{\bfM}}}^2, \gamma) >0$ is the square of the positive solution of the quadratic function.
$\Box$
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank Jalal Fadili for fruitful discussion, and the unknown reviewers for the important remarks that helped to improved the shape of the paper. Without both of them, the examples of the optimal projections would not have appeared in the paper. This research was supported by New York Metropolitan Research Fund. R. Giryes is grateful to the Azrieli Foundation for the award of an Azrieli Fellowship. This work was supported in part by the EU FP7, SMALL project under FET-Open grant number 225913, and EPSRC grants EP/J015180/1 and EP/F039697/1. R. Gribonval acknowledges the support of the European Research Council, PLEASE project, under grant ERC-StG- 2011-277906. MED acknowledges support of his position from the Scottish Funding Council and their support of the Joint Research Institute with the Heriot-Watt University as a component part of the Edinburgh Research Partnership.
[^1]: By the abuse of notation we use the same notation for the selection sub-matrices of rows and columns. The selection will be clear from the context since in analysis the focus is always on the rows and in synthesis on the columns.
[^2]: At a first glance one would think that the conditions should be in terms of $\delta_{2r-d}^{\text{corank}}$ and $\delta_{4r-3d}^{\text{corank}}$. However, given two cosparse vectors with coranks $r_1$ and $r_2$ the best estimation we can have for the corank of their sum is $r_1+r_2-\pdim$.
[^3]: Remark that we even improve the condition of the idealized case in [@Blumensath09Sampling] to be $\deltaB \le \frac{1}{3}$ instead of $\deltaB \le \frac{1}{5}$.
[^4]: A matlab package with code for the experiments performed in this paper is in preparation for an open source distribution.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe the construction of a nested family of Hilbert spaces $\{ \RKHS_\tau : \tau>0 \}$ of functions on a torus, which also have a $C^*$ algebra structure under the pointwise product of functions. This space is built using the eigenfunctions of the Laplace Beltrami operator (Laplacian) on the torus, and is dense in both the space of smooth functions, and the space of functions square integrable with respect to the normalized Lebesgue volume measure on the torus. This makes the $\RKHS_\tau$ good candidates for an approximation space, or for a hypothesis space in conditional expectation problems. In the $1$-dimensional case of the circle, these spaces are related to a fractional diffusion operator. We propose that these spaces $\RKHS_\tau$ should also be useful in a range of applications, including Galerkin methods, signal processing, ergodic theory, exterior calculus and analysis on manifolds. We construct these spaces by directly describing an orthonormal basis. The elements of these basis have the additional advantage of being easily approximable from data-driven or finite rank approximations of the kernel integral operators.'
address: 'Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA'
author:
- Suddhasattwa Das
- Dimitrios Giannakis
bibliography:
- 'References.bib'
title: Reproducing kernel Hilbert algebras on compact Abelian Lie groups
---
Banach algebra ,Reproducing kernel Hilbert space,Kernel integral operator ,Diffusion operator 47A60 ,46F12 ,65R10 ,47D07 ,46J10 ,46C05 ,46J15
Introduction. {#sec:intro}
=============
In many areas of applied and theoretical mathematics, such as harmonic analysis, analysis on manifolds, signal processing and Galerkin methods, one works with operators or performs approximations of functions using some choice of a functional space. The choice of the functional space strongly affects the outcome and applicability of the methods being developed. The purpose of this paper is to provide a constructive definition of a functional space which satisfies many aspects of a good functional space, especially in the context of dynamical systems. We have listed some properties desirable from a functional space in Table \[tab:spaces\], along with some common choices of spaces.
Property $L^p(\mu)$, $p \neq 2, \infty$ $L^2(\mu)$ $L^\infty(\mu)$ $C^r(M)$ RKHS RKHA
--------------------------------- -------------------------------- ------------ ----------------- ---------- ------ ------
Banach space (normed, complete)
Inner-product structure
Pointwise evaluation
Algebra structure
: Summary of the common choices of functional spaces. On the left are some properties commonly desired from functional spaces, the table shows which space satisfies which property. The goal of this paper is to construct a space(s) called RKHA or *Reproducing kernel Hilbert algebra*. See Section \[sec:intro\] for more discussions on RKHS or RKHA. []{data-label="tab:spaces"}
Two basic choices for a functional space are $L^p(\mu)$ or $C^r(M)$, where $L^p(\mu)$ is the collection of $L^p$-equivalence classes with respect to a measure $\mu$; and $C^r(M)$ denotes the collection of $C^r$-smooth functions on a manifold $M$. The $L^p(\mu)$ has the advantage of not requiring any manifold structure on the underlying space. When $p=2$ it is also a Hilbert space and thus allows for the crucial operations of orthogonal projections and inner-products. Results in harmonic analysis and the operator theoretic formulation of dynamical systems are often stated in the context of $L^p(\mu)$ spaces. In certain situations [@Giannakis15; @DasGiannakis_delay_Koop], one may enforce some regularity on the functions by defining a more general notion of Sobolev spaces.
The disadvantage of working in $L^2(\mu)$ is that the elements of this space are equivalence classes of functions, and the basic task of evaluation at points cannot be performed. This poses a challenge when trying to state rigorous results on prediction of observables. This shortcoming is overcome by $C^r(M)$ functions, which have the additional property of being regular. However, these spaces are not Hilbert spaces and there is no concept of an inner-product or a unique way to define projections. The lack of an inner product also makes it difficult to determine the coefficients of a vector with respect to a given Schauder basis.
A bridge between these two spaces are reproducing kernel Hilbert spaces, or RKHSs. These have been used extensively in learning and extrapolation, and have recently been used in the spectral analysis of dynamical systems [@DGJ_compactV_2018; @DasGiannakis_RKHS_2018] and data-assimilation in partially observed dynamical systems [@AlxndrGian2019]. In typical situations, these are dense as subspaces of both $L^2(\mu)$ and $C^r(M)$, and have a Hilbert structure of their own. The are however not an algebra, i.e., these spaces are not preserved by taking products of functions. The extra structure on an algebra provides many more tools to study linear operators on these spaces. Some classical examples are the spectral exterior calculus of a manifold [@DimitrisBerry_SEC_2018], and the use of the GNS representation theorem for operator algebras to study operators in quantum mechanics [@Bratteli2006]. The latter technique has also been extended to data-analysis in a stochastic dynamical system [@Giannakis_quantum_2019]. The present work describes the construction of Hilbert spaces of smooth functions that also carry the algebra property. We next state more precisely the desirable properties of our functional space, using a notion called a *Hilbert algebra*.
#### Hilbert algebras
For our purposes, a Hilbert algebra is a Hilbert space $H$ with an additional binary multiplication operation $\cdot : H\times H \to H$ which makes it a $C^*$ algebra [e.g. @FewsterRejzner2019]. Additionally we require that $$\label{eqn:jonv0}
\|ab\| \leq \|a\| \|b\|, \quad \langle ab, c \rangle_H = \langle b, a^*c \rangle_H = \langle a, c b^* \rangle_H, \quad \forall a,b,c\in H.$$ Equation along with the distributive property of algebras imply that for each $a\in H$, the operations of left and right multiplication by $a$ is a norm preserving, \*-homomorphism between $H$ and $\mathcal{L}_b(H;H)$, the collection of bounded linear maps from $H$ to $H$. The algebra will be called associative / commutative, if the multiplication operation is associative / commutative. $H$ will be called *unital* if it has a special *unit* element, denoted as $1_H$, such that for every $a\in H$, $a 1_H = 1_H a = a$. For our purposes, we will call a family of functions $\left\{ H_\tau : \tau>0 \right\}$ a *nested family* of commutative, unital Hilbert \*algebras if for each $0<\tau<\tau'$ $$H_{\tau'} \subset H_\tau, \quad \|h\|_{\tau'} > \|h\|_{\tau}, \quad \forall h\in H_{\tau'} .$$ where $\norm{\cdot}_\tau$ denotes the $H_{\tau}$-norm. In the next section, we will describe a construction of these spaces. Since these will additionally also be RKHS, we will call them *reproducing kernel Hilber algebra*-s or RKHAs. Our notion of a Hilbert algebra is somewhat different from some traditional definitions [e.g. @Nakano1950; @Dixmier2011], mainly by our requirement of a $C^*$ algebra property.
Hilbert algebras on the circle and torus
========================================
We will now consider the cases when $M=\mathbb{T}^d$, the $d$-dimensional torus which is the $d$-fold Cartesian product of $\mathbb{T}^1$, the unit circle. These are compact Abelian Lie groups and their Haar probability measure $\mu$ is also the volume measure associated to the canonical flat Riemannian metric, suitable scaled to be a probability measure. We will first describe the construction of a RKHA on the circle and then on $\mathbb{T}^d$. The construction is closely related to a notion central to analysis on manifolds : the *Laplacian*.
#### Laplacian operator
Given a Riemannian metric on a manifold $M$, one can define a Laplace-Beltrami operator [e.g. @Rosenberg_laplacian_1997] $\Delta : C^\infty(M) \to C^\infty(M)$, which is an unbounded, self-adjoint operator. The operator $\Delta$ is also referred to as the Laplacian or *diffusion* operator. It has many equivalent definitions, one of them being the divergence with respect to the volume measure of the gradient of a smooth function. The Laplacian is the key component of the heat equation $du/dt = \Delta u$, the basic model for the study of heat conduction and other diffusive processes through material medium. It is usual to view the Laplacian as an unbounded operator $ \Delta : \dom(\Delta) \to L^2(\mu)$, where $\dom(\Delta)$ is a dense subspace of $L^2(\mu)$, and $L^2(\mu)$ is the collection of functions square integrable with respect to the Riemannian volume measure $\mu$ of $M$. On $\mathbb{T}^1$ and with the metric chosen as above, $\Delta$ is simply the second-order differential operator $$(\Delta f)(\theta_0) = \left. \frac{\partial^2}{\partial \theta^2} \right|_{\theta= \theta_0} f(\theta), \quad \forall f\in C^2\left( \mathbb{T}^1 \right), \, \forall \theta_0\in S^1.$$ Although $\Delta$ is an unbounded operator, it always has a purely discrete spectrum with its eigenfunctions forming an orthonormal basis of $L^2(\mu)$. For $M = \mathbb{T}^1$, these eigenpairs take the form $$\label{eqn:def:Lap_T1}
\Delta \varphi_{j} = \lambda_{j} \varphi_{j}, \quad \lambda_{j} = j^2, \quad \varphi_{j}( x) = e^{ \iota j x} \quad \forall x \in \mathbb{T}^1,$$ where the index $j$ runs over $\integer$. A related notion is that of a *fractional diffusion* process [e.g. @Wyss1986], which models phenomenon such as continuous-time random walks. It is the evolution group $e^{-\tau \Delta^{1/2}}$ generated by the operator $\Delta^{1/2}$. The fractional diffusion generator $\Delta^{1/2}$ is defined uniquely by its action on the orthonormal basis $\{ \varphi_j : j\in\integer \}$, namely : $$\Delta^{1/2} \varphi_{j} = \lambda_{j}^{1/2} \varphi_{j} = |j| \varphi_{j} .$$ The action of the fractional diffusion generator $\Delta^{1/2}$ can be described using a integral operator as $$(\Delta^{1/2} f)(x) = \int_{\mathbb{T}^1} k_\tau(x, y) f(y) d\mu(y)$$ where $k_\tau : \mathbb{T}^1\times \mathbb{T}^1 \to \real$ is defined as $$\label{eqn:def:heatker_T1}
k_\tau(x, y) := \sum_{j\in\integer} \lambda_{\tau,j} \varphi_{j}(x) \varphi^*_{j}(y),$$ where $*$ denotes complex conjugation, and $$\label{eqn:def:lambdatau_T1}
\lambda_{\tau,j} := e^{-\tau \lambda_{j}^{1/2} } = e^{-\tau |j| } .$$ We will prove that the series for $k_\tau$ converges uniformly and in $C^0$ norm for every $r>0$. A more important consideration for us will be a certain Hilbert space of functions associated to $k_\tau$, called *RKHS*.
#### RKHS
A *kernel* function such as $k_\tau$ above is a function $k:M\times M\to\real$, which measures the similarity between pairs of points on $M$. Kernel functions can be of various designs, and are meant to capture the geometric structures of the underlying manifold; see for example [@BelkinNiyogi03; @Kernel1; @CoifmanLafon06]. A kernel $k$ is be said to be *positive semidefinite* if for every $x_0, x_1, \ldots, x_n \in M $ and $ c_0, c_1, \ldots, c_n \in \cmplx$, $ \sum_{i,j=0}^{n-1} c_i^* k(x_i,x_j) c_j \geq 0$. It will called *strictly positive definite* if equality holds in this equation iff each of the $c_i$-s are zero. We will make the following general assumptions on kernels.
\[A:ker\] $k:M\times M\to \real$ is a $C^r$ symmetric, positive semidefinite kernel, for some $r\geq 0$.
A *reproducing kernel Hilbert space* or RKHS is a Hilbert space of pointwise evaluated functions that one can associate with any symmetric, positive semi-definite kernel $k :M\times M\to \real$. If the kernel $k$ is $C^r$, the corresponding Hilbert space $\RKHS$ will also consist of $C^r$ functions. To define $\RKHS$, first note that for every $x\in M$, $k_x := k(\cdot, x)$ is a continuous function. Next, consider the pre-Hilbert space $\RKHS'$ consisting of finite linear combinations of $\{ k_x : x\in M \}$, and with the inner-product structure $$\left\langle \sum_{i=1}^{m} a_i k(\cdot, x_i), \sum_{j=1}^{n} b_j k(\cdot, y_j) \right\rangle := \sum_{i=1}^{m} \sum_{j=1}^{n} a_i^* b_j k( x_i, y_j) .$$ The RKHS $\RKHS$ can now be defined as the Hilbert space completion of $\RKHS'$, with the inner-product structure as above. By the Moore-Aronszajn theorem [@Aronszajn1950], the elements in this Hilbert space completion are themselves well defined $C^r$ functions.
\[thm:T1\] For every $\tau>0$, the series for $k_\tau$ in converges uniformly and absolutely. Moreover, $k_\tau$ satisfies Assumption \[A:ker\]. Hence one also has a sequence of associated RKHSs $\RKHS_\tau$. Then $\{ \RKHS_\tau : \tau>0 \}$ is a nested family of commutative, unital Hilbert \*algebras.
#### Remark
The kernel $k_\tau$ in has the simple expression $k_\tau(\theta, \theta') = \frac{ \sinh(\tau) }{ \cosh(\tau) - \cos(\theta-\theta') }$. To see why, note that $$\begin{split}
k_\tau(\theta, \theta') &= \sum_j e^{-|j|\tau} e^{\iota j (\theta-\theta')} = \sum_{j=0}^{\infty} e^{-j\tau} e^{\iota j (\theta-\theta')} + \sum_{j=0}^{\infty} e^{-j\tau} e^{-\iota j (\theta-\theta')} - 1 \\
&= \left[ 1 - e^{-\tau + \iota(\theta-\theta')} \right]^{-1} + \left[ 1 - e^{-\tau - \iota(\theta-\theta')} \right]^{-1} - 1 = \frac{ \sinh(\tau) }{ \cosh(\tau) - \cos(\theta-\theta') }.
\end{split}$$ We will prove Theorem \[thm:T1\] as a special case of a more general result on $\mathbb{T}^d$ for general $d\in\num$.
#### The $d$-dimensional torus
As before, the spaces of functions that we will construct will be RKHSs associated to a kernel function, which in turn will be related to the Laplacian. The Laplacian has a purely discrete spectrum as usual on $M = \mathbb{T}^d$, with its eigenfunctions forming an orthonormal basis of $L^2(\mu)$. These eigenpairs take the form $$\label{eqn:def:Lap_Td}
\Delta \varphi_{\vec{j}} = \lambda_{\vec{j}} \varphi_{\vec{j}}, \quad \lambda_{\vec{j}} = \exp(-\|\vec{j}\|_2^2), \quad \varphi_{\vec{j}}( \vec{\theta}) = \exp\left( \iota \vec{\theta} \cdot \vec{j} \right) \quad \forall \vec{\theta}\in \mathbb{T}^d,$$ where the eigenpairs have been indexed by $\vec{j} \in \integer^d$ (instead of a scalar $j\in\integer$). Here $\norm{\cdot}_2$ denotes the Euclidean 2-norm. The Laplacian eigenfunctions are therefore, simply the Fourier functions on the torus. Similar to , define $$\label{eqn:def:lambdatau_Td}
\lambda_{\tau,\vec{j}} := \exp\left( -\tau \norm{\vec{j}}_1 \right).$$ where $\norm{\cdot}_1$ denotes the Euclidean $1$-norm. Similarly to , we define a kernel $k_\tau$ as $$\label{eqn:def:heatker_Td}
k_\tau(x, y) := \sum_{ \vec{j} \in \integer^d} \lambda_{\tau, \vec{j}} \varphi_{\vec{j}}(x) \varphi^*_{\vec{j}}(y),$$ We now have similarly to Theorem \[thm:T1\]:
\[thm:Td\] For every $\tau>0$, the series for $k_\tau$ in converges uniformly and absolutely. Moreover, $k_\tau$ satisfies Assumption \[A:ker\]. Hence one also has a sequence of associated RKHSs $\RKHS_\tau$. Then $\{ \RKHS_\tau : \tau>0 \}$ is a nested family of commutative, unital Hilbert \*algebras.
Note that when $d=1$, is equivalent to and thus Theorem \[thm:T1\] is a special case of Theorem \[thm:Td\]. We will prove Theorem \[thm:Td\] in Section \[sec:proof:Td\]. We will show that they are a special case of a more general method on arbitrary manifolds which is based on kernel integral operators.
Hilbert algebras from kernel integral operators {#sec:construction}
===============================================
We now give a more general construction of a parameterized family of RKHSs on a manifold. It is based on kernel integral operators, and we show in Theorem \[thm:RKHS\_tau\] that if the kernel satisfies certain assumptions, one gets a nested family of Hilbert spaces. The following will be a standing assumption for the rest of the paper.
\[A:mu\] $M$ is a $C^r$ manifold, and $\mu$ is a Borel probability measure with a compact support $X$.
The measure $\mu$ could be a smooth volume measure, a purely atomic sampling measure or an invariant measure corresponding to a fractal, chaotic set of a dynamics. Given any such $\mu$, one can consider the Hilbert space $L^2(\mu)$ of equivalence classes of functions which are square integrable with respect to (wrt) $\mu$. The norm on this space will be denoted as $\norm{\cdot}_{\mu}$ and it arises from the following inner product: $$\langle f , g \rangle_{\mu} = \int_X f^* g d\mu.$$ A kernel function $k:M\times M\to \real$ will be called *square integrable* (w.r.t. $\mu$) if $\int_{M\times M} |k(x,y)| d\mu(x) \mu(y) < \infty$, i.e., $k\in L^2(\mu \times \mu )$. Note that if $k$ satisfies Assumption \[A:ker\], then under the compactness assumption in Assumption \[A:mu\], $k$ is guaranteed to be in $L^2(\mu \times \mu )$. Associated to a $k\in L^2(\mu \times \mu)$ is a compact integral operator $K:L^2(\mu)\to L^2(\mu) $ defined as $$Kf(x) := \int_X k(x,y)f(y) \, d\mu(y).$$By the symmetry of $k$, $K$ is self-adjoint and has a spectrum confined to the real-line. Since $\mu$ has compact support, $K$ is also a compact and thus has a bounded, purely atomic spectrum. Suppose that Assumption \[A:ker\] holds. Then strict positive definiteness implies that the spectrum is positive, the eigenvalues admit the ordering $ \lambda_0 > \lambda_1 \geq \lambda_2 \geq \cdots $ and the corresponding eigenfunctions, $ \phi_j $ form a real orthonormal basis of $L^2(\mu) $. On certain occasions, it is advantageous to have the additional property of being a *Markov* kernel, which means that the top eigenvalue $\lambda_0$ of $K$ equals $1$ and the corresponding eigenvector $\phi_0$ equals $1_M$, the constant function on $M$, equal to $1$. If Assumptions \[A:ker\]\[A:mu\] are satisfied and $k$ is also strictly positive definite, then $k$ can be transformed into a symmetric, strictly-positive definite Markov kernel using the method of binormalization [@CoifmanHirn2013]. Thus, in addition if the Markov property holds, then $\lambda_0 = 1$, $ \phi_0 =1_M$, and the spectrum of $K$ is discrete and confined to the interval $(0,1]$.
#### Universal applicability of integral operators
Due to the unspecified nature of $\mu$, one can use kernel integral operators and their related techniques in a wide variety of situations. In general, the eigenbasis of kernel integral operators have a strong connection with the underlying geometry [@CoifmanLafon06b]. When $X = M$ and $\mu$ is a a Riemannian volume measure, then $K$ suitably designed have been shown to yield the eigenfunctions of the associated Laplace Beltrami operator [@CoifmanLafon06]. Even when $X$ is non-smooth or fractal, using the method of delay-coordinates, the eigenfunctions of $K$ has been shown to yield salient dynamic features [@DasGiannakis_delay_Koop].
#### Integral operators and RKHS
The operator $K$ plays an important role in the construction of $\RKHS$, the RKHS associated to the kernel $k$. First of all, the range of $K$ is a dense subspace of $\RKHS$, and thus $K$ can be viewed as the operator $$K : L^2(\mu) \to \RKHS .$$ Moreover for each $j\in \num_0$, define the functions $\varphi_j$ and $\psi_j$ as $$\label{eqn:def:psi}
\varphi_j := \lambda_j^{-1} K \phi_j, \quad \psi_j := \lambda_j^{-1/2} K \phi_j = \lambda_j^{1/2} \varphi_j.$$ The functions $\varphi_j$ are $C^r$ representations of the $L^2(\mu)$ functions $\phi_j$, and the set $\{ \psi_j : j\in \num_0 \}$ is an orthonormal basis for $\RKHS$. If $K$ is viewed as an operator with codomain $\RKHS$ rather than $L^2(\mu)$, one can compose it with its adjoint $K^* : \RKHS\to L^2(\mu)$ to effectively get an integral operator mapping $L^2(\mu)$ into $L^2(\mu)$, namely, $$G = K^* K : L^2(\mu) \to L^2(\mu).$$ It is a standard result that the map $K^*$ is simply the inclusion map of the continuous functions in $\RKHS$ into their respective $L^2(\mu)$ equivalence classes. As a result, the operator $G$ is the same as our former definition of $K$ as an operator on $L^2(\mu)$. Some important relations between $K$ and $K^*$ are summarized as follows: $$\label{eqn:jkbc9}
K^*K \phi_j = \lambda_j \phi_j, \quad K^* \psi_{j} = \lambda_j^{1/2} \phi_j.$$
#### A general cascade framework
We will now provide a framework to build a cascade of RKHSs $\{ \RKHS_\tau : \tau>0 \}$, where $\tau$ will be a parameter. It is based on a technique developed in [@DGJ_compactV_2018], and is reminiscent of the construction of rigged Hilbert spaces [@MadridEtAl2002]. The RKHSs $\RKHS_\tau$ will be associated to a parameterized family of reproducing kernels $k_\tau$ and its associated integral operator $K_\tau$. Instead of constructing $k_\tau$ directly, we will explicitly construct the discrete spectrum $$\lambda_{\tau, 0} > \lambda_{\tau,1} \geq \lambda_{\tau,2} \geq \ldots >0$$ of $K_\tau$ and then construct $k_\tau$ using the $\lambda_{\tau,j}$ and $\varphi_j$ from , in a manner analogous to and . The exact definitions of the $\lambda_{\tau,j}$ will depend on the case in consideration. But assuming that the $\lambda_{\tau,j}$ are defined, we can define, at least formally, a kernel $k_\tau$ as follows: $$\label{eqn:def:ktau}
k_{\tau} :M\times M\to \real, \quad k_{\tau} (x,y) := \sum_{j=-\infty}^{\infty} \psi_{\tau,j}(x) \psi^*_{\tau,j}(y) .$$ where $$\label{eqn:def:psi_tau}
\psi_{\tau,j} := \lambda_{\tau,j}^{1/2} \varphi_j, \quad \forall j\in \num_0.$$ We will show that under certain conditions on $M,\mu,k$ and the $\lambda_{\tau,j}$s, the infinite sum in is well-defined. If $k_\tau$ is a well defined function, then one has the smoothing operator : $$K_\tau : L^2(\mu) \to \RKHS_\tau, \; K_\tau f := \int_X k_\tau(\cdot, y) f(y) d\mu(y)$$ and the compact symmetric operator $$G_\tau : L^2(\mu) \to L^2(\mu), \quad G_\tau := K_\tau^* K_\tau .$$ We show in Theorem \[thm:RKHS\_tau\] that the $\psi_{\tau,j}$ play the same role for $\RKHS_\tau$ as the $\psi_{j}$ does for $\RKHS$. For that, we need the following assumptions.
\[A:phi\] The functions $\varphi_j$ are bounded in $\norm{}_{\sup}$ norm, uniformly over $j$. Moreover for each $\tau>0$, $\sum_{j=0}^{\infty} \lambda_{\tau, j} < \infty$
\[A:lambdatau\] For each $j\in\num_0$, $\lambda_{\tau,j}$ depends continuously on $\tau$ and converges monotonically to $0$ as $\tau\to\infty$.
Given a reproducing kernel $k:M\times M\to \real$ and its corresponding RKHS $\RKHS$, $\RKHS(X)$ will denote the subspace of $\RKHS$ spanned by the collection $\{ k(x,\cdot) : x\in X \}$. Then we have:
\[thm:RKHS\_tau\] Let $M$ be a manifold and Assumptions \[A:ker\], \[A:mu\] and \[A:phi\] hold. Then:
(i) The series for $k_\tau$ in converges absolutely and uniformly over $X\times X$.
(ii) The kernel $k_\tau$ also satisfies Assumptions \[A:ker\]. Thus there is an RKHS $\RKHS_\tau$ associated to the kernel $k_\tau$.
(iii) The $\psi_{\tau,j}$ from form an orthonormal basis for $\RKHS_\tau(X)$. Moreover $$\label{eqn:kbvc9}
\psi_{\tau,j} = \lambda_{\tau,j}^{-1/2} K_\tau \phi_j; \quad \varphi_{\tau,j} = \lambda_{\tau,j}^{-1} K_\tau \phi_j .$$
(iv) In particular, $\RKHS_\tau(X)$ is a dense subset of $L^2(\mu)$ and $C^0(X)$, for each $\tau>0$.
(v) Further, if Assumption \[A:lambdatau\] holds, then the $\RKHS_\tau$ form a nested sequence of Hilbert spaces, i.e., for each $0<\tau<\tau'$, $$\RKHS_{\tau'} \subset \RKHS_\tau, \quad \|h\|_{\tau'} > \|h\|_{\tau}, \quad \forall h\in \RKHS_{\tau'} .$$
We first prove Theorem \[thm:RKHS\_tau\] in Section \[sec:proof:RKHS\_tau\]. We shall also establish some lemmas which follow from Theorem \[thm:RKHS\_tau\]. In Section \[sec:proof:Td\] we show that Assumption \[A:phi\] is satisfied on the torus and thereby prove Theorem \[thm:Td\].
Proof of Theorem \[thm:RKHS\_tau\] {#sec:proof:RKHS_tau}
==================================
We begin with two results which reveal parts of the claims of Theorem \[thm:RKHS\_tau\]. The hypothesis of these results are themselves consequences of the assumptions of Theorem \[thm:RKHS\_tau\]. These results are continuations of the authors’ work in [@DGJ_compactV_2018]. These results difer from notable results in kernel integral operator theory [e.g. @SriperumbudurEtAl2011; @Steinwart_inflnc_2001] which begin with the assumption that the kernel $k$ is semidefinite. The following proposition gives sufficient conditions under which a kernel is semidefinite.
\[prop:D\] Let Assumption \[A:mu\] hold and $k:X\times X\to \real$ be a continuous kernel. Then if the integral operator $G:L^2(\mu)\to L^2(\mu)$ associated to $k$ and $\mu$ is positive semidefinite, then so is the kernel $k$.
It is equivalent to show the contrapositive, i.e., if $k$ is not positive semidefinite, then neither is $G$. So let there be $a_i\in\cmplx$ and $x_i\in X$ for $i=1,\ldots,n$ such that $\sum_{i=1}^{n} \sum_{j=1}^{n} a_i^* a_j k(x_i, x_j) = -\delta$ for some $\delta>0$. For each $\epsilon>0$, let $U_i(\epsilon)$ be a neighborhood of $x_i$ such that $$\sup \left\{ \norm{ k(\cdot, x_i) - k(\cdot, y) }_{\sup(X)} \;:\; y\in U_i(\epsilon) \right\} < \epsilon .$$ Since the $x_i\in\support(\mu)$, we have $\mu\left( U_i(\epsilon) \right) >0$ for each $\epsilon>0$. Define $$\phi_{\epsilon} := \sum_{i=1}^{n} b_i 1_{U_i(\epsilon)}, \quad b_i := a_i / \mu(U_i(\epsilon)) .$$ Then $\phi_{\epsilon}$ lies in $L^2(\mu)$. To show that $G$ is not positive semidefinite, it is enough to show that for some $\epsilon>0$, the inner-product $\left\langle \phi_{\epsilon}, G \phi_{\epsilon} \right\rangle_{\mu}$ is negative. Note that $$\left( G \phi_{\epsilon} \right)(x) = \int_X k(x,y) \phi_{\epsilon}(y) d\mu(y) = \sum_{j=1}^{n} b_j \int_{U_j(\epsilon)} k(x,y) d\mu(y) .$$ Therefore, $$\begin{split}
\left\langle \phi_{\epsilon}, G \phi_{\epsilon} \right\rangle_{\mu} &= \left\langle \sum_{i=1}^{n} b_i 1_{U_i(\epsilon)}, G \phi_{\epsilon} \right\rangle_{\mu} = \sum_{i=1}^{n} \sum_{j=1}^{n} b_i^* b_j \int_{U_i(\epsilon)} \int_{U_j(\epsilon)} k(x,y) d\mu(x) d\mu(y) \\
&= \sum_{i=1}^{n} \sum_{j=1}^{n} b_i^* b_j \int_{U_i(\epsilon)} \int_{U_j(\epsilon)} \left[ k(x_i,x_j) - k(x,y) \right] d\mu(x) d\mu(y) \\
& \quad \quad \quad \quad \quad \quad + \sum_{i=1}^{n} \sum_{j=1}^{n} b_i^* b_j \int_{U_i(\epsilon)} \int_{U_j(\epsilon)} k(x_i,x_j) d\mu(x) d\mu(y) \\
&= \sum_{i=1}^{n} \sum_{j=1}^{n} b_i^* b_j \int_{U_i(\epsilon)} \int_{U_j(\epsilon)} \left[ k(x_i,x_j) - k(x,y) \right] d\mu(x) d\mu(y) + \sum_{i=1}^{n} \sum_{j=1}^{n} a_i^* a_j k(x_i, x_j) \\
&= \sum_{i=1}^{n} \sum_{j=1}^{n} b_i^* b_j \int_{U_i(\epsilon)} \int_{U_j(\epsilon)} \left[ k(x_i,x_j) - k(x,y) \right] d\mu(x) d\mu(y) - \delta.
\end{split}$$ Therefore, $$\left| \left\langle \phi_{\epsilon}, G \phi_{\epsilon} \right\rangle_{\mu} + \delta \right| \leq \sum_{i=1}^{n} \sum_{j=1}^{n} |b_i^* b_j| \int_{U_i(\epsilon)} \int_{U_j(\epsilon)} \left| k(x_i,x_j) - k(x,y) \right| d\mu(x) d\mu(y) \leq \epsilon \sum_{i=1}^{n} \sum_{j=1}^{n} |a_i^* a_j|.$$ Since this inequality holds for every $\epsilon>0$ and since $\sum_{i=1}^{n} \sum_{j=1}^{n} |a_i^* a_j|$ is bounded and constant, for $\epsilon$ small enough, the RHS is less than $0.5 \delta$. In that case, $\left\langle \phi_{\epsilon}, G \phi_{\epsilon} \right\rangle_{\mu}$ must the less than $-0.5\delta$ and is thus negative. This completes the proof of the claim.
Next, we give sufficient conditions under which the operator $G$ is semidefinite.
\[prop:E\] Let Assumption \[A:mu\] hold and and suppose that $\{ \varphi_j : j\in\num_0 \}$ be a sequence of continuous functions which also form an orthonormal basis for $L^2(\mu)$. Let there be bounded, positive constants $a_0,a_1,\ldots$ such that the series $$k(x,y) = \sum_{j=0}^{\infty} a_j \varphi_j(x) \varphi^*_j(y)$$ converges uniformly and absolutely. Then $k:X\times X\to \real$ is a uniformly continuous, symmetric, positive semidefinite kernel.
The uniform continuity and symmetry of $k$ follows from the assumption on the series. To prove semidefiniteness, let $G:L^2(\mu)\to L^2(\mu)$ be the integral operator associated to $k$ and $\mu$. Then $G$ is a compact symmetric operator, and the $\{ \varphi_j : j\in\num_0 \}$ form a complete set of eigenvectors for $G$, and $a_j$ are the corresponding eigenvalues. To check this, note that $$\begin{split}
(G \varphi_i)(x) &= \int_X k(x,y) \varphi_i(y) d\mu(y) = \int_X \sum_{j=0}^{\infty} a_j \varphi_j(x) \varphi^*_j(y) \varphi_j(y) d\mu(y) \\
&= \sum_{j=0}^{\infty} a_j \varphi_j(x) \int_X \varphi^*_j(y) \varphi_j(y) d\mu(y) = \sum_{j=0}^{\infty} a_j \varphi_j(x) \delta_{i,j} = a_i \varphi_i(x) .
\end{split}$$ In the above equations, the integral can be brought inside the infinite sum because of the uniform absolute convergence of the series for $k$. Since all the $a_j$ are positive by assumption, $G$ is positive definite. Thus Proposition \[prop:D\] applies and $k$ is positive semidefinite.
#### Proof of Theorem \[thm:RKHS\_tau\] {#proof-of-theoremthmrkhs_tau}
By assumption, there is a constant $C>0$ such that for every $j$, $\norm{\varphi_j}_{\sup} \leq C$. Therefore, $$\sup_{x,y\in X} \abs{ \sum_{|j|\geq N} \psi_{\tau,j}(x) \psi^*_{\tau,j}(y) } \leq \sup_{x,y\in X} \sum_{|j|\geq N} \abs{ \psi_{\tau,j}(x) \psi^*_{\tau,j}(y) } \stackrel{ \text{by \eqref{eqn:def:psi_tau}} }{=} \sup_{x,y\in X} \sum_{|j|\geq N} \lambda_{\tau,j} \abs{ \varphi_{j}(x)} \abs{ \varphi_{j}(y) } \leq C^2 \sum_{|j|\geq N} \lambda_{\tau,j} .$$ Thus by the summability of the $\lambda_{\tau,j}$s, $$\label{eqn:mpxdv7}
\lim_{N\to\infty} \sup_{x,y\in X} \left| \sum_{j=N}^{\infty} \psi_{\tau,j}(x) \psi^*_{\tau,j}(y) \right| = C^2 \lim_{N\to\infty} \sum_{|j|\geq N} \lambda_{\tau,j} = 0,$$ and the series for $k_\tau$ converges absolutely and uniformly, proving Claim (i). Claim (ii) follows immediately from Proposition \[prop:E\].
#### Proof of Claim (iii)
To prove the orthonormality of the $\psi_{\tau,j}$, we begin with the observation $$\label{eqn:khb8xs}
\int_X \psi^*_{\tau,i}(y) \phi_j(y) d\mu(y) \stackrel{ \text{by \eqref{eqn:def:psi_tau}} } { = } \lambda_{\tau,i}^{1/2} \int_X \varphi^*_{i}(y) \phi_j(y) d\mu(y) = \lambda_{\tau,i}^{1/2} \left\langle \varphi_i, \phi_j \right\rangle_{\mu} = \lambda_{\tau,i}^{1/2} \left\langle \phi_i, \phi_j \right\rangle_{\mu} = \lambda_{\tau,i}^{1/2} \delta_{i,j} .$$ Now note that $$K_\tau \phi_k = \int_X k_\tau(\cdot,y) \phi_j(y) d\mu(y)= \int_X \left[ \sum_{j=0}^{\infty} \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \phi_k(y) \right] d\mu(y) .$$ Therefore for every $N>k$, $$\begin{gathered}
\left\| K_\tau \phi_k - \lambda_{\tau,k} \varphi_k \right\|_{\sup} = \left\| \int_X \left[ \sum_{j=0}^{\infty} \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \phi_j(y) \right] d\mu(y) - \lambda_{\tau,k} \varphi_k \right\|_{\sup} \\
\leq \left\| \sum_{j=0}^{N-1} \int_X \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \phi_k(y)d\mu(y) - \lambda_{\tau,k} \varphi_k \right\|_{\sup} + \left\| \int_X \left[ \sum_{j=N}^{\infty} \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \phi_k(y) \right] d\mu(y) \right\|_{\sup} \\
= \left\| \lambda_{\tau,k}^{1/2} \psi_{\tau,k} - \lambda_{\tau,k} \varphi_k \right\|_{\sup} + \left\| \int_X \left[ \sum_{j=N}^{\infty} \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \phi_k(y) \right] d\mu(y) \right\|_{\sup} \, \text{by \eqref{eqn:khb8xs}}, \\
= \left\| \int_X \left[ \sum_{j=N}^{\infty} \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \phi^*_k(y) \right] d\mu(y) \right\|_{\sup} \, \text{by \eqref{eqn:def:psi_tau}}, \, \\
\leq \int_X \left\| \sum_{j=N}^{\infty} \psi_{\tau,j}(\cdot) \psi_{\tau,j}(y) \right\|_{\sup} \phi_k(y) d\mu(y) \leq \left\| \sum_{j=N}^{\infty} \psi_{\tau,j}(\cdot) \psi^*_{\tau,j}(y) \right\|_{\sup}.
\end{gathered}$$ Thus taking $\lim_{N\to\infty}$ on both sides gives $$\left\| K_\tau \phi_k - \lambda_{\tau,k} \varphi_k \right\|_{\sup} = \lim_{N\to\infty} \left\| K_\tau \phi_j - \lambda_{\tau,k} \varphi_k \right\|_{\sup} \leq \lim_{N\to\infty} \left\| \sum_{j=N}^{\infty} \psi_{\tau,j}(\cdot) \psi_{\tau,j}(y) \right\|_{\sup} = 0, \, \text{by \eqref{eqn:mpxdv7}}.$$ Thus $ K_\tau \phi_k = \lambda_{\tau,k} \varphi_k = \lambda_{\tau,k}^{1/2} \psi_{\tau,k}$, proving . Next note that since $\RKHS_\tau$ is defined to be the RKHS associated to $k_\tau$, the range of the integral operator $K_\tau$ will be dense in $\RKHS_\tau(X)$. By , each of the $\psi_{\tau,j}$s are in the range of $K_\tau$ and hence lie in $\RKHS_\tau(X)$. We will next prove that $\left\{ \psi_{\tau,j} : j\in\num_0 \right\}$ is an orthonormal set in $\RKHS_\tau$: $$\begin{split}
\left\langle \psi_{\tau,i} , \psi_{\tau,j} \right\rangle_{\RKHS_\tau} &\stackrel{ \text{by \eqref{eqn:kbvc9}} }{ = } \left\langle \lambda_{\tau,i}^{-1/2} K_\tau \phi_i , \lambda_{\tau,j}^{-1/2} K_\tau \phi_j \right\rangle_{\RKHS_\tau} = \left( \lambda_{\tau,i} \lambda_{\tau,j} \right)^{-1/2} \left\langle K_\tau^* K_\tau \phi_i , \phi_j \right\rangle_{\mu} \\
&= \left( \lambda_{\tau,i} \lambda_{\tau,j} \right)^{-1/2} \lambda_{\tau,i} \left\langle \phi_i , \phi_j \right\rangle_{\mu} = \delta_{i,j} .
\end{split}$$ Here we have used the fact that $ K_\tau^* K_\tau \phi_i = \lambda_{\tau,i} \phi_i $, analogously to . It remains to be shown that $\left\{ \psi_{\tau,j} : j\in\num_0 \right\}$ is a basis. Since they are orthonormal, this is true iff their linear span is dense. Thus it is equivalent to show that if there is a $\psi\in\RKHS_\tau(X)$ which is orthogonal to each of the $\psi_{\tau,j}$, then $\psi$ must be zero. But then for each $j\in\num_0$, $$0 = \left\langle \psi, \psi_{\tau,j} \right\rangle_{\RKHS_\tau} \stackrel{ \text{by \eqref{eqn:kbvc9}} }{ = } \lambda_{\tau,j}^{-1/2} \left\langle \psi, K_\tau \phi_j \right\rangle_{\RKHS_\tau} = \lambda_{\tau,j}^{-1/2} \left\langle K_\tau^* \psi, \phi_j \right\rangle_{\mu} .$$ Since the $\{ \phi_j : j\in\num_0 \}$ form an orthonormal basis for $L^2(\mu)$, this proves that $K_\tau^* \psi$ must be $0$. But $$\ker K_\tau^* = \ran (K_\tau)^\bot = \{0\}$$ since $\ran K_\tau$ is dense. Thus $\psi = 0$, proving the claim.
#### Proof of Claim (iv)
This follows from the fact that the $\varphi_j$ and the $\psi_{\tau,j}$ are orthonormal bases for $L^2(\mu)$ and $\RKHS_\tau(X)$ respectively. By , the $\psi_{\tau,j}$ are scalar multiples of the $\varphi_j$. Thus $\RKHS_\tau(X)$ must be dense in $L^2(\mu)$.
#### Proof of Claim (v)
Note that for each $j\in\num_0$, $\varphi_j$ lies in $\RKHS_\tau$ for every $\tau>0$. Thus for each $\tau'>0$, $\psi_{\tau,j}$ also lies in $\RKHS_{\tau'}$ and $$\norm{ \psi_{\tau,j} }_{\tau'} = \norm{ \lambda_{\tau,j}^{1/2} \varphi_{j} }_{\tau'} = \left( \lambda_{\tau,j} / \lambda_{\tau',j} \right)^{1/2} \norm{ \lambda_{\tau',j}^{1/2} \varphi_{j} }_{\tau'} = \left( \lambda_{\tau,j} / \lambda_{\tau',j} \right)^{1/2} \norm{ \psi_{\tau',j} }_{\tau'} = \left( \lambda_{\tau,j} / \lambda_{\tau',j} \right)^{1/2}$$ Thus depending on whether $\tau'$ is greater than or less than $\tau$, $\norm{ \psi_{\tau,j} }_{\tau'}$ is greater than or less than $\norm{ \psi_{\tau,j} }_{\tau}$, which equals $1$. The inclusions and inequality of Claim (v) follow immediately from this observation. This completes the proof of Theorem \[thm:RKHS\_tau\].
We end this section with two important applications of Theorem \[thm:RKHS\_tau\], in the form of lemmas. We shall need them later in the paper.
\[lem:mn6x\] Let the assumptions of Theorem \[thm:RKHS\_tau\] hold, $\left( a_j \right)_j$ be a sequence of numbers, and $\tau>0$. Then the functions $f_N = \sum_{|j|\leq N} a_j \varphi_{j}$ converges as $N\to \infty$ to a function in $\RKHS_\tau$ iff the sequence $\left( \lambda_{\tau,j}^{-1/2} |a_j| \right)_{j}$ is $\ell^2$, i.e., $\sum_j \lambda_{\tau,j}^{-1} |a_j|^2 < \infty$.
The proof of this lemma follows from basic Hilbert space theory and will be omitted.
\[lem:nuc0x\] Let $M$ be a manifold and Assumptions \[A:ker\], \[A:mu\], \[A:phi\] and \[A:lambdatau\] hold. Then for every $T>0$ and every $f\in \RKHS_{T}$, the map $\tau \mapsto \norm{f}_{\tau}$ defined for $\tau\in (0, T]$, is continuous.
#### Proof of Lemma \[lem:nuc0x\]
Note that the map is well defined because the $\RKHS_\tau$ form a nested family of spaces, so $f\in \RKHS_\tau$ for every $\tau\in (0, T]$. Let $f=\sum_j a_j \psi_{T,j}$. Then we also have $$f=\sum_j a_j \psi_{T,j} \stackrel{ \text{by \eqref{eqn:def:psi_tau}} }{=} \sum_j a_j \lambda_{T,j}^{1/2} \varphi_j = \sum_j a_j \left[ \lambda_{T,j} / \lambda_{\tau,j} \right] \lambda_{\tau,j}^{1/2} \varphi_j \stackrel{ \text{by \eqref{eqn:def:psi_tau}} }{=} \sum_j a_j \left[ \lambda_{T,j} / \lambda_{\tau,j} \right]^{1/2} \psi_{\tau,j}.$$ Now define $r(j;\tau,T) := \lambda_{T,j} \lambda_{\tau,j}^{-1}$. Then the above equation can be rewritten as $$\norm{f}_{\tau}^2 = \sum_j |a_j|^2 \lambda_{T,j} \lambda_{\tau,j}^{-1} = \sum_j |a_j|^2 r(j;\tau,T), \quad .$$ By Assumption \[A:lambdatau\], for every $j$, $r(j;\tau,T)$ varies continuously with $\tau$, and $$\label{eqn:bpljc9}
\lim_{\tau'\to \tau} \lambda_{\tau',j} = \lambda_{\tau,j}; \quad \lim_{\tau'\to \tau} r(j;\tau',T) = r(j;\tau,T), \quad r(j;\tau,T) < 1, \quad \forall 0<\tau<T.$$ Now note that $$\norm{f}_{\tau}^2 - \norm{f}_{\tau'}^2 = \sum_{|j|<N} |a_j|^2 \left[ r(j;\tau,T) - r(j;\tau',T) \right] + \sum_{|j|\geq N} |a_j|^2 \left[ r(j;\tau,T) - r(j;\tau',T) \right]$$ We can derive an upper-bound for the left-hand side (LHS) using as $$\abs{ \norm{f}_{\tau}^2 - \norm{f}_{\tau'}^2 } \leq \sum_{|j|<N} |a_j|^2 \left| r(j;\tau,T) - r(j;\tau',T) \right| + \sum_{|j|\geq N} |a_j|^2$$ Therefore, taking the limit $\tau'\to \tau$ gives $$\limsup_{\tau'\to \tau} \abs{ \norm{f}_{\tau}^2 - \norm{f}_{\tau'}^2 } \leq \limsup_{\tau'\to \tau} \sum_{|j|<N} |a_j|^2 \left| r(j;\tau,T) - r(j;\tau',T) \right| + \sum_{|j|\geq N} |a_j|^2 \stackrel{ \text{by \eqref{eqn:bpljc9}} }{ = } \sum_{|j|\geq N} |a_j|^2.$$ Taking the limit $N\to\infty$ on both sides gives $$\limsup_{\tau'\to \tau} \abs{ \norm{f}_{\tau}^2 - \norm{f}_{\tau'}^2 } = \lim_{N\to\infty} \limsup_{\tau'\to \tau} \abs{ \norm{f}_{\tau}^2 - \norm{f}_{\tau'}^2 } = \lim_{N\to\infty} \sum_{|j|\geq N} |a_j|^2 = 0,$$ since $\{a_j\}_j$ is by definition, an $\ell^2$ sequence. This completes the proof of the lemma.
Proof of Theorem \[thm:Td\] {#sec:proof:Td}
===========================
For simplicity of notation a bit we shall use $j$ to denote a vector in $\integer^d$, instead of the vectorial $\vec{j}$ notation from . Continuing the simplification, the $L^1$ norm $\norm{\vec{j}}_1$ will be denoted as $|j|$.
#### Outline of the proof
We will first show that the Fourier / Laplacian eigenfunctions in arise from a kernel integral operator. We will next show that Assumptions \[A:ker\], \[A:mu\], \[A:phi\] and \[A:lambdatau\] hold. Then Theorem \[thm:RKHS\_tau\] will be applicable and the kernels $k_\tau$ and RKHS $\RKHS_\tau$ become well defined. Finally, we will prove the algebraic structure on $\RKHS_\tau$.
#### The Laplacian on $\mathbb{T}^d$
For the torus, $M=\mathbb{T}^d$ and $\mu$ will be the normalized Riemannian volume measure, as described earlier. Thus Assumption \[A:mu\] is satisfied with $X=M$. It is well known [e.g. @Giannakis_QntmOsc_2019; @Steinwart_inflnc_2001] that the action of the Laplacian on $\mathbb{T}^d$ can be defined in terms of a kernel integral operator $K_{\text{Lap}}$ with a $C^0$ kernel $k_{\text{Lap}}$. Thus the Fourier eigenfunctions $\varphi_j$ from are also the eigenfunctions of a kernel integral operator whose kernel satisfies Assumption \[A:ker\]. The definition of the $\lambda_{\tau,j}$ , and the fact that Fourier functions have constant magnitude equal to one, implies that both Assumption \[A:phi\] and \[A:lambdatau\] hold. Thus Theorem \[thm:RKHS\_tau\] is applicable. Next note that the series for $k_\tau$ in which is in terms of the Laplacian eigenfunctions, is the same as the series in which is in terms of the kernel integral eigenfunctions. Thus by Theorem \[thm:RKHS\_tau\], $\{ \RKHS_\tau : \tau>0\}$ is a well defined, nested family of Hilbert spaces. It remains to be shown that for for each $\tau>0$, $\RKHS_\tau$ is a unital, commutative Hilbert \*algebra. For that we need the following key property of pointwise products between various RKHS functions.
\[prop:lp9cc\] Let the assumptions and notations of Theorem \[thm:Td\] hold. Then for every $\tau>0$, $c>0$, $f\in \RKHS_{(c+1)\tau}$ and $g\in\RKHS_\tau$, the pointwise product $fg$ lies in $\RKHS_\tau$. Moreover, there is a constant $C_\tau>0$ depending only on $\tau$ such that $\norm{ fg }_{\tau} \leq C_\tau \norm{f}_{(c+1)\tau} \norm{g}_{\tau} .
$
Theorem \[thm:Td\] now follows from Proposition \[prop:lp9cc\]. The continuity of the norm $\norm{\cdot}_{\tau}$ with $\tau$ from Lemma \[lem:nuc0x\] implies that for every $f\in\RKHS_{(c+1)\tau}$ and $g\in\RKHS_{\tau}$, $$\label{eqn:lkjbc8}
\norm{ fg }_{\tau} \leq C_\tau \norm{f}_{\tau} \norm{g}_{\tau} .$$ Since $\RKHS_{(c+1)\tau}$ is dense in $\RKHS_\tau$, the above inequality applies to $f\in\RKHS_\tau$ as well. Thus $\RKHS_\tau$ is an algebra. Since $\varphi_0 = 1_X$, it serves as a unit element and $\RKHS_\tau$ is unital. It is commutative since pointwise multiplication is commutative. The $*$-algebra property is satisfied by any functional algebra. The normed algebra property (ii) of a Hilbert algebra follows from . This proves that for each $\tau>0$, $\RKHS_\tau$ is indeed a Hilbert algebra, and completes the proof of Theorem \[thm:Td\].
The rest of this section will be devoted to the Proposition.
#### Proof of Proposition \[prop:lp9cc\]
One of the key properties of the Laplacian eigenbasis on the torus that we will rely on is that $$\label{eqn:oinv0}
\phi_i \phi_j = \phi_{i+j}, \quad \phi_i \phi^*_j = \phi_{i-j}, \quad \forall i,j\in\integer^d.$$ Let $f = \sum_j f_j \psi_{(c+1)\tau,j} \in \RKHS_{(c+1)\tau}$ and $g = \sum_j g_j \psi_{\tau,j} \in \RKHS_\tau$. Since $f,g$ are continuous functions, their product $fg$ is also continuous and hence is in $L^2(\mu)$. Therefore for every $j\in\integer^d$, the following inner products exist: $$\begin{split}
a_k &:= \left\langle \phi_k , fg \right\rangle_{\mu} = \left\langle f^* \phi_k , g \right\rangle_{\mu} = \left\langle f^* \phi_k , \sum_j g_j \psi_{\tau,j} \right\rangle_{\mu} = \sum_j g_j \left\langle f^* \phi_k ,\psi_{\tau,j} \right\rangle_{\mu}= \sum_j g_j \left\langle \psi^*_{\tau,j} \phi_k ,f \right\rangle_{\mu} \\
&= \sum_j g_j \left\langle \psi^*_{\tau,j} \phi_k , \sum_i f_i \psi_{(c+1)\tau,i} \right\rangle_{\mu} = \sum_j g_j \sum_i f_i \left\langle \psi^*_{\tau,j} \phi_k , \psi_{(c+1)\tau,i} \right\rangle_{\mu} = \sum_{i,j} f_i g_j \left\langle \phi_k , \psi_{(c+1)\tau,i} \psi_{\tau,j} \right\rangle_{\mu}.
\end{split}$$ In the equalities above, we were able to bring the $L^2(\mu)$-inner products inside the infinite sums because since the series expansions for $f$ and $g$ converge in their respective RKHS norms, the series also converge in $C^0(X)$ norm and thus in $L^2(\mu)$ norm too. Continuing the expansion, we get $$\begin{split}
a_k &= \sum_{i,j} f_i g_j \left\langle \phi_k , \psi_{(c+1)\tau,i} \psi_{\tau,j} \right\rangle_{\mu} \stackrel{ \text{by \eqref{eqn:def:psi_tau}} }{ = } \sum_{i,j} f_i g_j \left[ \lambda_{(c+1)\tau,i} \lambda_{\tau,j} \right]^{1/2} \left\langle \phi_k , \varphi_{i} \varphi_{j} \right\rangle_{\mu} \\
& \stackrel{\text{by \eqref{eqn:oinv0}}}{=} \sum_{i,j} f_i g_j \left[ \lambda_{(c+1)\tau,i} \lambda_{\tau,j} \right]^{1/2} \left\langle \phi_k , \varphi_{i+j} \right\rangle_{\mu} = \sum_{j} f_{k-j} g_j \left[ \lambda_{(c+1)\tau,k-j} \lambda_{\tau,j} \right]^{1/2} .
\end{split}$$ In conclusion, $$\label{eqn:kbx8zq}
a_k := \left\langle \phi_k , fg \right\rangle_{\mu} = \sum_{j} f_{k-j} g_j \left[ \lambda_{(c+1)\tau,k-j} \lambda_{\tau,j} \right]^{1/2} .$$ We make a small side note here that $$\label{eqn:dfnc80x}
\sum_k \sum_{j} \abs{f_{k-j} g_j}^2 \xrightarrow{i=k-j} = \sum_i \sum_{j} \abs{f_{i} g_j}^2 = \sum_i |f_{i}|^2 \sum_{j}|g_j|^2 = \norm{f}_{(c+1)\tau}^2 \norm{g}^2_{\tau},$$ To prove that $fg\in\RKHS_\tau$, by Lemma \[lem:mn6x\], it is equivalent to show that $\left( \lambda_{\tau,k}^{-1/2} a_k \right)_k$ is an $\ell^2$ sequence. But $$\begin{split}
\norm{ \left( \lambda_{\tau,k}^{-1/2} a_k \right)_k }^2_{\ell^2} &= \sum_k \abs{ \sum_{j} f_{k-j} g_j \mathcal{E}_{k,j}^{1/2} }, \quad \mathcal{E}_{k,j}:= \lambda_{(c+1)\tau,k-j} \lambda_{\tau,j} \lambda_{\tau,k}^{-1} \\
& \leq \sum_k \left[\sum_{j} \abs{f_{k-j} g_j}^2 \right] \left[ \sum_{j} \mathcal{E}_{k,j} \right], \quad \mbox{by Cauchy Schwarz},\\
&\leq \norm{f}_{(c+1)\tau}^2 \norm{g}^2_{\tau} \left[ \sup_k \sum_{j} \mathcal{E}_{k,j} \right], \quad \mbox{by \eqref{eqn:dfnc80x}} .
\end{split}$$ Since the LHS equals the squared norm $\norm{fg}_{\tau}^2$, the Proposition will be proven if it can be shown that there is a constant $C_\tau>0$ such that $$\label{eqn:jbc7}
\sup_k \sum_{j} \mathcal{E}_{k,j} = \sup_k \sum_j \exp\left( -\tau \left( (c+1)|k-j| + |j| - |k| \right) \right) \leq C_\tau.$$ So it only remains to prove . Some more simplifying assumptions will be made. We will first prove the case when $d=1$ and later extend the proof to higher values of $d$. Secondly, we may assume WLOG that $f_0 = g_0 = 0$, as both the functions $1_X g = g$ and $f 1_X = f$ lie in $\RKHS_\tau$. Now define the subspaces $\RKHS_\tau^{\pm}$ to be the spans of $\left\{ \psi_{\tau,\pm j} : j>0 \right\}$. Thus $\RKHS_\tau$ has the orthogonal splitting $$\RKHS_\tau = \spn(1_X) \oplus \RKHS_\tau^{+} \oplus \RKHS_\tau^{-} .$$ We can therefore assume WLOG that $f$ belongs to either of $\RKHS_{(c+1)\tau}^{\pm}$, and $g$ belongs to either of $\RKHS_{\tau}^{\pm}$. We will split our analysis into four cases, depending on these cases.
#### Case 1
$f\in \RKHS_{(c+1)\tau}^{-}$ and $g\in \RKHS_{\tau}^{-}$ : The proof for this case is exactly analogous to Case 2 below.
#### Case 2
$f\in \RKHS_{(c+1)\tau}^{+}$ and $g\in \RKHS_{\tau}^{+}$ : In this case $f_i=0$ unless $i>0$, and $g_j=0$ unless $j>0$. Thus when summing over $j$ in we can be restricted to the values of $j$ for which $k-j>0$ and $j>0$. The same restriction also applied to the LHS in . When $k>0$, the sum becomes $$\begin{split} \sum_j \mathcal{E}_{k,j} &= \sum_{j=1}^{k-1} e^{ -\tau \left( (c+1)|k-j| + |j| - |k| \right) } = \sum_{j=1}^{k-1} e^{-\tau \left( (c+1)(k-j) + j - k \right)} \\
&= \sum_{m=1}^{k-1} e^{-\tau cm} = \frac{e^{-\tau} (1-e^{-n\tau}) }{ 1-e^{-\tau} } < \left( e^{\tau} - 1 \right)^{-1} < \infty.
\end{split}$$ When $k<0$, the range for $j$ is empty and the sum is zero. This completes the proof of this case.
#### Case 3
$f\in \RKHS_{(c+1)\tau}^{-}$ and $g\in \RKHS_{\tau}^{+}$ : The proof for this case is exactly analogous to Case 4 below.
#### Case 4
$f\in \RKHS_{(c+1)\tau}^{+}$ and $g\in \RKHS_{\tau}^{-}$ : In this case $f_i=0$ unless $i>0$, and $g_j=0$ unless $j<0$. Thus when summing over $j$ in we can be restricted to the values of $j$ for which $k-j>0$ and $j<0$. The same restriction also applies to the LHS in . When $k>0$, the sum becomes $$\sum_j \mathcal{E}_{k,j} = \sum_{j<0} e^{ -\tau \left( (c+1)|k-j| + |j| - |k| \right) } = \sum_{j>0} e^{-\tau \left( (c+1)(k+j) + j - k \right)} = e^{-\tau ck} \left( e^{(c+2) \tau} - 1 \right)^{-1} < \left( e^{(c+2) \tau} - 1 \right)^{-1} < \infty.$$ When $k<0$, the sum becomes $$\begin{split}
\sum_j \mathcal{E}_{k,j} &= \sum_{j=-\infty}^{k-1} \exp\left( -\tau \left( (c+1)|k-j| + |j| - |k| \right) \right), \quad \mbox{take } m = k-j \\
& = \sum_{m>0} \exp\left( -\tau \left( (c+1)m + (m-k) + k \right) \right) = \sum_{m>0} e^{-\tau( c+2)m} = \left( e^{(c+2) \tau} - 1 \right)^{-1} < \infty.
\end{split}$$
#### The case when $d>1$
We will now switch back to the vectorial notation $\vec{j}$ and denote by $j_1, \ldots, j_d$ its $d$ components. The sum on the LHS of can then be written as $$\begin{split}
\sum_{\vec{j} \in \integer^d} \mathcal{E}_{ \vec{k}, \vec{j}} &= \sum_{\vec{j} \in \integer^d} \exp\left( -\tau \sum_{i=1}^{d} \left( (c+1) \abs{ k_i - j_i } + \abs{j_i} - \abs{k_i} \right) \right) = \sum_{\vec{j} \in \integer^d} \prod_{i=1}^{d} \exp\left( -\tau \left( (c+1)\abs{ k_i - j_i } + \abs{j_i} - \abs{k_i} \right) \right) \\
&= \prod_{i=1}^{d} \left[ \sum_{j \in \integer} \exp\left( -\tau \left( (c+1)\abs{ k_i - j } + |j| - \abs{k_i} \right) \right) \right] .
\end{split}$$ Similar to the case when $d=1$, we can consider the $2^d \times 2^d$ cases of $f$ lying in either of the $2^d$ spaces $\oplus^d \RKHS_{(c+1)\tau}^{\pm}$ and $g$ lying in either of the $2^d$ spaces $\oplus^d \RKHS_\tau^{\pm}$. In each of these cases, for every $\vec{k} \in \integer^d$, the sum $\sum_{\vec{j} \in \integer^d} \mathcal{E}_{ \vec{k}, \vec{j}}$ is the product of $d$ sums $$\sum_{j \in \integer} \exp\left( -\tau \left( (c+1)\abs{ k_i - j } + |j| - \abs{k_i} \right) \right)$$ as shown above. Each of these sums are bounded by the constant $C_\tau$ obtained for the $d=1$ case. Thus we have $$\sup_{ \vec{k} \in \integer^d } \sum_{\vec{j} \in \integer^d} \mathcal{E}_{ \vec{k}, \vec{j}} \leq C_\tau^d .$$ This completes the proof of Proposition \[prop:lp9cc\].
#### Acknowledgments
This research was supported by NSF grant DMS 1854383, ONR MURI grant N00014-19-1-242, and ONR YIP grant N00014-16-1-2649.
| {
"pile_set_name": "ArXiv"
} |
amssym.def amssym -1.3cm 0.0cm
ł Ł ø c
\[section\] \[thm\][Proposition]{} \[thm\][Corollary]{} \[thm\][Conjecture]{} \[thm\][Lemma]{} \[thm\][Remark]{} \[thm\][Definition]{} \[thm\][Hypothesis]{} addtoreset[equation]{}[section]{}
[**Regularity of rational vertex operator algebras**]{}
Chongying Dong[^1], Haisheng Li and Geoffrey Mason[^2]\
Department of Mathematics, University of California, Santa Cruz, CA 95064
Introduction
============
Rational vertex operator algebras, which play a fundamental role in rational conformal field theory (see \[BPZ\] and \[MS\]), single out an important class of vertex operator algebras. Most vertex operator algebras which have been studied so far are rational vertex operator algebras. Familiar examples include the moonshine module $V^{\natural}$ (\[B\], \[FLM\], \[D2\]), the vertex operator algebras $V_L$ associated with positive definite even lattices $L$ (\[B\], \[FLM\], \[D1\]), the vertex operator algebras $L(l,0)$ associated with integrable representations of affine Lie algebras \[FZ\] and the vertex operator algebras $L(c_{p,q},0)$ associated with irreducible highest weight representations for the discrete series of the Virasoro algebra (\[DMZ\] and \[W\]).
A rational vertex operator algebra as studied in this paper is a vertex operator algebra such that any [*admissible*]{} module is a direct sum of simple ordinary modules (see Section 2). It is natural to ask if such complete reducibility holds for an arbitrary weak module (defined in Section 2). A rational vertex operator algebra with this property is called a [*regular*]{} vertex operator algebra. One motivation for studying such vertex operator algebras arises in trying to understand the appearance of negative fusion rules (which are computed by the Verlinde formula) for vertex operator algebras $L(l,0)$ for certain rational $l$ (cf. \[KS\] and \[MW\]).
In this paper we give several sufficient conditions under which a rational vertex operator algebra is regular. We prove that the rational vertex operator algebras $V^{\natural},$ $L(l,0)$ for positive integers $l,$ $L(c_{p,q},0)$ and $V_L$ for positive definite even lattices $L$ are regular. Our result for $L(l,0)$ implies that any restricted integrable module of level $l$ for the corresponding affine Lie algebra is a direct sum of irreducible highest weight integrable modules. This result is expected to be useful in comparing the construction of tensor product of modules for $L(l,0)$ in \[F\] based on Kazhdan-Lusztig’s approach \[KL\] with the construction of tensor product of modules \[HL\] in this special case. We should remark that $V_L$ in general is a vertex algebra in the sense of \[DL\] if $L$ is not positive definite. In this case we establish the complete reducibility of any weak module.
Since the definition of vertex operator algebra is by now well-known, we do not define vertex operator algebra in this paper. We refer the reader to \[FLM\] and \[FHL\] for their elementary properties. The reader can find the details of the constructions of $V^{\natural}$ and $V_L$ in \[FLM\], and $L(l,0)$ and $L(c_{p,q},0)$ in \[DMZ\], \[DL\], \[FLM\], \[FZ\], \[L1\] and \[W\].
The paper is organized as follows: In Section 2, after defining the notion of weak module for a vertex operator algebra and the definition of rational vertex operator algebra, we discuss the rational vertex operator algebras $V^{\natural},$ $V_L,$ $L(l,0)$ and $L(c_{p,q},0).$ Section 3 is devoted to regular vertex operator algebras. We begin this section with the definition of regular vertex operator algebra. We show that the tensor product of regular vertex operator algebras is also regular and that a rational vertex operator algebra is regular under either of the assumptions (i) it contains a regular vertex operator subalgebra, or (ii) any weak module contains a simple ordinary module. These results are then used to prove that $V^{\natural},$ $V_L$ ($L$ is positive definite), $L(l,0)$ and $L(c_{p,q},0)$ are regular. We also discuss the complete reducibility of weak $V_L$-modules for an arbitrary even lattice $L.$ Based on these results, we conjecture that [*any*]{} rational vertex operator algebra is regular.
We thank Yi-Zhi Huang for pointing out a mistake in a prior version of this paper.
Rational vertex operator algebras
=================================
Let $(V,Y,{\bf 1},\omega)$ be a vertex operator algebra (cf. \[B\], \[FHL\] and \[FLM\]). A [*weak module*]{} $M$ for $V$ is a vector space equipped with a linear map $$\begin{array}{l}
V\to (\mbox{End}\,M)[[z^{-1},z]]\label{map}\\
v\mapsto\displaystyle{ Y_M(v,z)=\sum_{n\in\Z}v_nz^{-n-1}\ \ \ (v_n\in
\mbox{End}\,M)}
\end{array}$$ (where for any vector space $W,$ we define $W[[z^{-1},z]]$ to be the vector space of $W$-valued formal series in $z$) satisfying the following conditions for $u,v\in V$, $w\in M$: $$\begin{aligned}
& &Y_M(v,z)=\sum_{n\in \Z}v_nz^{-n-1}\ \ \ \ \mbox{for}\ \
v\in V;\label{1/2}\\
& &v_nw=0\ \ \
\mbox{for}\ \ \ n\in \Z \ \ \mbox{sufficiently\ large};\label{vlw0}\\
& &Y_M({\bf 1},z)=1;\label{vacuum}\end{aligned}$$ $$\label{jacobi}
\begin{array}{c}
\displaystyle{z^{-1}_0\delta\left(\frac{z_1-z_2}{z_0}\right)
Y_M(u,z_1)Y_M(v,z_2)-z^{-1}_0\delta\left(\frac{z_2-z_1}{-z_0}\right)
Y_M(v,z_2)Y_M(u,z_1)}\\
\displaystyle{=z_2^{-1}\delta\left(\frac{z_1-z_0}{z_2}\right)
Y_M(Y(u,z_0)v,z_2)}.
\end{array}$$ $$\label{vir}
[L(m),L(n)]=(m-n)L(m+n)+\frac{1}{12}(m^3-m)\delta_{m+n,0}(\mbox{rank}\,V)$$ for $m, n\in {\Z},$ where $$\begin{aligned}
& &L(n)=\omega_{n+1}\ \ \ \mbox{for}\ \ \ n\in{\Z}, \ \ \
\mbox{i.e.},\ \ \ Y_M(\omega,z)=\sum_{n\in{\Z}}L(n)z^{-n-2};\nonumber\\
& &\frac{d}{dz}Y_M(v,z)=Y_M(L(-1)v,z).\label{6.72}\end{aligned}$$ This completes the definition. We denote this module by $(M,Y_M)$ (or briefly by $M$).
\[r2.2\] An (ordinary) $V$-module is a weak $V$-module which carries a $\C$-grading $$M=\coprod_{\lambda \in{\C}}M_{\lambda}$$ such that $\dim M_{\l}$ is finite and $M_{\l+n}=0$ for fixed $\l$ and $n\in {\Z}$ small enough. Moreover one requires that $M_{\l}$ is the $\l$-eigenspace for $L(0):$ $$L(0)w=\l w=(\mbox{wt}\,w)w, \ \ \ w\in M_{\l}.$$
This definition is weaker than that of \[FLM\], for example, where the grading on $M$ is taken to be rational. The extra flexibility attained by allowing $\C$-gradings is important $-$ see for example \[DLM1\] and \[Z\].
We observe some redundancy in the definition of weak module:
\[r2.1\] Relations (\[vir\]) and (\[6.72\]) in the definition of weak module are consequences of (\[1/2\])-(\[jacobi\]).
To establish (\[6.72\]) note that $L(-1)u=L(-1)u_{-1}{\bf 1}=u_{-2}{\bf 1}$ for $u\in V.$ Then $$\begin{aligned}
& &\ \ \ Y_{M}(L(-1)u,z_{2})\nonumber\\
& &=Y_{M}(u_{-2}{\bf 1},z_{2})\nonumber\\
& &={\rm Res}_{z_{0}}z_{0}^{-2}Y_{M}(Y(u,z_{0}){\bf 1},z_{2})\nonumber\\
& &={\rm Res}_{z_{0}}{\rm Res}_{z_{1}}z_{0}^{-2}
\left(z^{-1}_0\delta\left(\frac{z_1-z_2}{z_0}\right)
Y_M(u,z_1)Y_M({\bf 1},z_2)\right.\nonumber\\
& &\ \ \ \ \ \ \left.-z^{-1}_0\delta\left(\frac{z_2-z_1}{-z_0}\right)
Y_M({\bf 1},z_2)Y_M(u,z_1)\right)\nonumber\\
& &={\rm Res}_{z_{0}}{\rm
Res}_{z_{1}}z_{0}^{-2}z_2^{-1}\delta\left(\frac{z_1-z_0}{z_2}\right)
Y_{M}(u,z_{1})\nonumber\\
& &={\rm Res}_{z_{0}}{\rm
Res}_{z_{1}}z_{0}^{-2}z_1^{-1}\delta\left(\frac{z_2+z_0}{z_1}\right)
Y_{M}(u,z_{2}+z_{0})\nonumber\\
& &={\rm Res}_{z_{0}}z_{0}^{-2}Y_{M}(u,z_{2}+z_{0})\nonumber\\
& &={\rm Res}_{z_{0}}z_{0}^{-2}e^{z_0\frac{d}{d z_2}}Y_{M}(u,z_{2})\nonumber\\
& &={d\over dz_{2}}Y_{M}(u,z_{2}).\end{aligned}$$ This establishes (\[6.72\]), and together with (\[jacobi\]) and $Y(\omega,z_0)\omega=\frac{1}{2}({\rm rank}V)z_0^{-4}
+2\omega z_0^{-2}+L(-1)\omega z_0^{-1}+$ regular terms we can easily deduce (\[vir\]). $\Box$
Thus we may just use (\[1/2\])-(\[jacobi\]) as the axioms for a weak $V$-module.
\[d2.2\] An [*admissible*]{} $V$-module is a weak $V$-module $M$ which carries a ${\Z}_{+}$-grading $$M=\coprod_{n\in {\Z}_{+}}M(n)$$ ($\Z_+$ is the set all nonnegative integers) satisfying the following condition: if $r, m\in {\Z} ,n\in {\Z}_{+}$ and $a\in V_{r}$ then $$\begin{aligned}
a_{m}M(n)\subseteq M(r+n-m-1).\label{2.7}\end{aligned}$$
We call an admissible $V$-module $M$ [*simple*]{} in case $0$ and $M$ are the only $\Z_+$-graded submodules.
$V$ is called [*rational*]{} if every admissible $V$-module is a direct sum of simple admissible $V$-modules. That is, we have complete reducibility of admissible $V$-modules.
\[r2.4\] (i) Note that any ordinary $V$-module is admissible.
\(ii) It is proved in \[DLM1\] that if $V$ is rational then conversely, every simple admissible $V$-module is an ordinary module. Moreover $V$ has only a finite number of inequivalent simple modules.
\(iii) Zhu’s definition of rational vertex operator algebra $V$ is as follows \[Z\]: (a) all admissible $V$-module are completely reducible, (b) each simple admissible $V$-module is an ordinary $V$-module, (c) $V$ only has finitely many inequivalent simple modules. Thanks to (ii), Zhu’s definition of rational thus coincides with our own.
We next introduce a certain category $\cal{O}$ of admissible $V$-modules in analogy with the well-known category $\cal O$ of Bernstein-Gelfand-Gelfand. First some notation: for any weak $V$-module $M$ we set for $h\in \C:$ $$M_{h}=\{m\in M|
(L(0)-h)^km=0\ {\rm for\ some} \ k\in\Z_{+}\}.$$ So $M_h$ is a [*generalized*]{} eigenspace for $L(0),$ and in particular $M_h$ is the $h$-eigenspace for $L(0)$ if $L(0)$ is a semisimple operator.
Now define $\cal{O}$ to be the category of weak $V$-modules $M$ satisfying the following two conditions:
\(1) $L(0)$ is locally finite in the sense that if $m\in M$ then there is a finite-dimensional $L(0)$-stable subspace of $M$ which contains $m.$
\(2) There are $h_{1}, \cdots, h_{k}\in\C$ such that $$M=\oplus
_{i=1}^{k}\oplus _{n\in {\Z}_{+}}M_{n+h_{i}}.$$ These are the [*objects*]{} of $\cal O.$ Morphisms may be taken to be $V$-module homomorphisms, though we will not make use of them in the sequel.
\[r2.5\] (i) Any weak $V$-module which belongs to $\cal O$ is necessarily admissible: a $\Z_+$-grading obtains by defining $M(n)=\oplus_{i=1}^kM_{h_i+n}.$ Condition (\[2.7\]) follows in the usual way.
\(ii) Suppose that $M$ is a weak $V$-module and that $W$ is a weak $V$-submodule of $M.$ Then $M$ lies in $\cal O$ if, and only if, both $W$ and $M/W$ lie in $\cal O.$
\(iii) If $V$ is rational, any weak $V$-module in $\cal{O}$ is a direct sum of simple $V$-modules (use Remark \[r2.4\] (ii)).
Next we briefly discuss some familiar examples of rational vertex operator algebras. The reader is referred to the references for notation and the details of the constructions.
\(1) Let $L$ be an even lattice and $V_{L}$ the corresponding vertex algebra (see \[B\], \[DL\] and \[FLM\]). It is proved in \[D1\] that if $L$ is positive definite then $V_{L}$ is rational and its simple modules are parametrized by $L'/L$ where $L'$ is the dual lattice of $L.$
\(2) Let ${\frak g}$ be a finite-dimensional simple Lie algebra with a Cartan subalgebra ${\frak h}$ and $\hat \frak g=\C[t,t^{-1}]\otimes \frak g\oplus\C c$ the corresponding affine Lie algebra. Fix a positive integer $l.$ Then any $\lambda\in \frak h^*$ can be viewed as a linear form on $\C c\oplus \frak h\subset \hat \frak g$ by sending $c$ to $l.$ Let us denote the corresponding irreducible highest weight module for $\hat\frak g$ by $L(l,\lambda).$ Then $L(\ell,0)$ is a rational vertex operator algebra (\[DL\], \[FZ\], \[L1\]).
\(3) Let $c$ and $h$ be two complex numbers and let $L(c,h)$ be the lowest weight irreducible module for the Virasoro algebra with central charge $c$ and lowest weight $h$. Then $L(c,0)$ has a natural vertex operator algebra structure (cf. \[FZ\]). Moreover, $L(\ell,0)$ is rational if, and only if, $c=c_{p,q}=1-\frac{6(p-q)^{2}}{pq}$ for $p,q\in \{2,3,4,\cdots\}$ and $p, q$ are relatively prime (see \[DMZ\] and \[W\]).
\(4) Let $V^{\natural}$ be the moonshine module vertex operator algebra constructed by Frenkel, Meurman and Lepowsky \[FLM\] (see also \[B\]). It is established in \[D2\] that $V^{\natural}$ is [*holomorphic*]{} in the sense that $V^{\natural}$ is rational and the only simple module is $V^{\natural}$ itself.
\(5) Let $V^{1},\cdots, V^{k}$ be vertex operator algebras. Then $V=\otimes_{i=1}^{k}V^{i}$ is a vertex operator algebra of rank $\sum_{i=1}^{k}{\rm rank} V^{i}$ and any simple $V$-module $M$ is isomorphic to a tensor product module $M^{1}\otimes \cdots \otimes M^{k}$ for some simple $V^{i}$-module $M^i$ \[FHL\]. Furthermore, $V^{1}\otimes V^{2}\otimes \cdots\otimes V^{k}$ is rational if, and only if, each $V^i$ is rational \[DMZ\].
\(6) Let $V^{1}, \cdots, V^{k}$ be vertex operator algebras of the same rank. Then $\oplus_{i=1}^{k}V^{k}$ is a vertex operator algebra \[FHL\]. It is clear that $\oplus_{i=1}^{k}V^{i}$ is rational if each $V^{i}$ is rational. The vacuum space of the resulting vertex operator algebra is not one-dimensional, and we will not consider this particular example further in this paper.
Regular vertex operator algebras
================================
In Section 2 we made use of the complete reducibility of admissible modules in order to define rational vertex operator algebras. The study of complete reducibility of an arbitrary weak module for a rational vertex operator algebra leads us to the following notion of regular vertex operator algebra:
\[d2.4\] A vertex operator algebra $V$ is said to be [*regular*]{} if any weak $V$-module $M$ is a direct sum of simple ordinary $V$-modules.
\[r3.2\] A regular vertex operator algebra $V$ is necessarily rational. Indeed if $M$ is a weak $V$-module then, being a direct sum of ordinary simple $V$-module, it is admissible (Remark \[r2.4\] (i)) and, for the same reason, a direct sum of simple admissibles.
The main result of the present paper is to show that the rational vertex operator algebras in examples (1)-(4) of Section 2 are each regular. First we have some general results which will be useful later.
\[p2.5\] Let $V^{1}, \cdots, V^{k}$ be regular vertex operator algebras. Then $V=V^{1}\otimes V^{2}\otimes \cdots \otimes V^{k}$ is regular.
[**Proof.**]{} Let $M$ be a weak $V$-module. For each $1\le i\le k$, we may regard $V^{i}$ as a vertex operator subalgebra with a different Virasoro element. Then $M$ is a weak $V^{i}$-module. Since $V^i$ is regular, $M$ is a direct sum of simple ordinary $V^{i}$-modules. Note that there are only finitely many simple $V^{i}$-modules up to equivalence. Thus $M$ is a $V^{i}$-module in category $\cal{O}$ of weak $V^i$-modules. Denote the generators of the Virasoro algebra of $V^i$ by $L_i(n).$ Then $L(0)=L_{1}(0)+\cdots +L_{k}(0)$ and $L_{i}(0)$’s commute with each other. This implies that $M$ is in category $\cal{O}$ of weak $V$-modules. So $M$ is completely reducible by Remarks \[r3.2\] and \[r2.5\] (iii). $\;\;\;\;\Box$
\[p2.6\] Let $V$ be a rational vertex operator algebra such that there is a regular vertex operator subalgebra $U$ with the same Virasoro element $\omega$. Then $V$ is regular.
[**Proof.**]{} Let $M$ be a weak $V$-module. Then $M$ is a weak $U$-module, so that $M$ is a direct sum of simple ordinary $U$-modules. Thus $L(0)$ acts semisimply on $M$. Let $W^{1},\cdots, W^{k}$ be all simple $U$-modules up to equivalence. Then we can write $M=\sum_{i=1}^{k}\oplus_{n\in {\Z}_{+}}M_{n+h_{i}}$, where $h_{i}$ is the lowest weight of $W^{i}$. Thus $M$ is in the category $\cal{O}$ for $V.$ The complete reducibility of $M$ follows immediately as $V$ is rational.$\;\;\;\;\Box$
\[p2.8\] Let $V$ be a rational vertex operator algebra such that any nonzero weak $V$-module contains a simple ordinary $V$-submodule. Then $V$ is regular.
[**Proof.**]{} Let $M$ be any weak $V$-module and let $W$ be the sum of all simple ordinary submodules. We have to prove that $W=M$. If $M\ne W$, $M/W$ is a nonzero weak $V$-module so that by assumption, there is a simple ordinary $V$-submodule $M^{1}/W$ of $M/W$. Then both $W$ and $M^{1}/W$ are in the category $\cal{O}$. Thus $M^{1}$ is in the category $\cal{O}$ and $M^{1}$ is a direct sum of simple ordinary $V$-modules. This contradicts the choice of $W$. $\;\;\;\;\Box$
Now we are ready to show that the vertex operator algebras $L(l,0),$ $L(c_{p,q},0),$ $V^{\natural}$ and $V_L$ are regular.
Recall from example (2) that $\frak g$ is a finite-dimensional simple Lie algebra with Cartan subalgebra $\frak h;$ $L(l,0)$ is a vertex operator algebra. We shall denote the corresponding root system by $\Delta.$
\[l2.9\] There is a basis $\{a^{1},\cdots, a^{m}\}$ for ${\frak
g}$ such that for $1\leq i,j\leq m$ we have $$\begin{aligned}
\label{eb}
[Y(a^{i},z_{1}),Y(a^{i},z_{2})]=0\;\;\mbox{ and }Y(a^{i},z)^{3\ell+1}=0\end{aligned}$$ as operators on $L(\ell,0).$
[**Proof.**]{} Let $\alpha\in\Delta$ and $e\in {\frak g}_{\alpha}$. If ${\frak g}$ is of type $A, D,$ or $ E$, it is proved in \[MP1\] and \[DL\] that $Y(e,z)^{\ell+1}=0$. In general, it is proved in \[L1\] and \[MP2\] that $Y(e,z)^{3\ell+1}=0$. It is well known (cf. \[H\], \[K1\]) that there are elements $e_{\alpha}\in {\frak g}_{\alpha}, f_{\alpha}\in {\frak g}_{-\alpha},
h_{\alpha}\in \frak h$ which linearly span a subalgebra naturally isomorphic to $sl_{2}$. Set $\sigma_{\alpha}=e^{(e_{\alpha})_{0}}$ where $(e_{\alpha})_0$ is the component operator of $Y(e_{\alpha},z)$ (cf. (\[1/2\])) corresponding to $z^{-1}.$ Then $\sigma_{\alpha}$ is an automorphism of the vertex operator algebra $L(\ell,0)$ (see Chap. 11 of \[FLM\]). A straightforward calculation gives $\sigma_{\alpha}(f_{\alpha})=f_{\alpha}+h_{\alpha}-2e_{\alpha}$. Since $e_{\alpha}, f_{\alpha}, h_{\alpha}$ form a basis of ${\frak g}$ for $\alpha\in \Delta,$ $e_{\alpha}, f_{\alpha}, \sigma_{\alpha}(f_{\alpha})$ also form a basis of ${\frak g}$. It is clear that this basis satisfies condition (\[eb\]). $\;\;\;\;\Box$
Let $\ell$ be a positive integer. Then the vertex operator algebra $L(\ell,0)$ is regular.
[**Proof.**]{} By Proposition \[p2.8\], it is enough to prove that any nonzero weak $L(\ell,0)$-module $M$ contains a simple $L(\ell,0)$-module. This will be established in three steps.
[**Claim 1:**]{} [*There exists a nonzero $u\in M$ such that $(t{\C}[t]\otimes {\frak g})u=0$.*]{} Set $\frak g(n)=t^n\otimes \frak g$ for $n\ne 0.$ For any nonzero $u\in M$, by the definition of a weak module, ${\frak g}(n)u=0$ for sufficiently large $n.$ So $(t{\C}[t]\otimes {\frak g})u$ is finite-dimensional. For any $u\in M$, we define $d(u)=\dim (t{\C}[t]\otimes {\frak g})u$. If there is a $0\ne u\in M$ such that $d(u)=0$, then $(t{\C}[t]\otimes {\frak g})u=0$. Suppose that $d(u)>0$ for any $0\ne u\in M$. Take $0\ne u\in M$ such that $d(u)$ is minimal.
Let $a^{i}$ ($1\le i\le m$) be a basis of ${\frak g}$ satisfying condition (\[eb\]) and let $k$ be the positive integer such that ${\frak g}(k)u\ne 0$ and ${\frak g}(n)u=0$ whenever $n>k$. By definition of $k$, $a^{i}(k)u\ne 0$ for some $1\le i\le m$. Since $Y(a^{i},z)^{3\ell+1}=0$, by Proposition 13.16 in \[DL\], $Y_{M}(a^{i},z)^{3\ell+1}=0$. Extracting the coefficient of $z^{-(k+1)(3\ell+1)}$ from $Y_{M}(a^{i},z)^{3\ell+1}u=0$ we obtain $(a^{i}_{k})^{3\ell+1}u=0$.
Let $r$ be a nonnegative integer such that $(a^{i}_{k})^{r}u\ne 0$ and $(a^{i}_{k})^{r+1}u=0$. Set $v=(a^{i}_{k})^{r}u$. We will obtain a contradiction by showing that $d(v)<d(u)$. First we prove that if $a_{n}u=0$ for some $a\in {\frak g}, 1\le n\in {\Z}$, then $a_{n}v=0$. In the following we will show by induction on $m$ that $a_n(a_k^i)^mu=0$ for any $a\in \frak g$ and $m\in \Z$ nonnegative. If $m=0$ this is immediate by the choice of $u.$ Now assume that the result holds for $m.$ Since $[a,a^{i}]_{k+n}u=0$ (from the definition of $k$) and $a_{n}u=0,$ by the induction assumption that $a_n(a_k^i)^mu=0$ we have: $$\begin{aligned}
[a,a^{i}]_{k+n}(a^{i}_{k})^{m}u=0,\;a_{n}(a^{i}_{k})^{m}u=0.\end{aligned}$$ Thus $$\begin{aligned}
&
%% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR
&a_{n}(a^{i}_{k})^{m+1}u=[a_{n},a^{i}_{k}](a^{i}_{k})^{m}u+a^{i}_{k}a_{n}(a^{i}_{k})^{m}u\nonumber\\
& &\ \ \ \
=[a,a^{i}]_{k+n}(a^{i}_{k})^{m}u+a^{i}_{k}a_{n}(a^{i}_{k})^{m}u\nonumber\\
& &\ \ \ \ =0,\end{aligned}$$ as required. In particular, we see that $a_{n}v=a_{n}(a^{i}_{k})^{r}u=0$. Therefore, $d(v)\le d(u)$. Since $a^{i}_{k}v=0$ and $a^{i}_{k}u\ne 0$, we have $d(v)< d(u)$.
[**Claim 2:**]{} [*There is a nonzero $u\in M$ such that ${\frak g}(n)u=0$ for $n>0$ and ${\frak g}_+u=0$ where $\frak g_+=\sum_{\alpha\in \Delta_+}\frak
g_{\alpha}$ for a fixed positive root system $\Delta_+.$*]{}
Set $$\begin{aligned}
\Omega(M)=\{u\in M| (t{\C}[t]\otimes {\frak g})u=0\}.\end{aligned}$$ Then $\Omega(M)$ is a nonzero ${\frak g}$-submodule of $M$ by Claim 1. Let $0\ne e_{\theta}\in {\frak g}_{\theta}$ where $\theta$ is the longest positive root in $\Delta$. Then $Y_{M}(e_{\theta},z)^{\ell+1}=0$ (see \[DL\] and \[FZ\]). Extracting the coefficient of $z^{-\ell-1}$ from $Y_{M}(e_{\theta},z)^{\ell+1}\Omega(M)=0$, we obtain $e_{\theta}^{\ell+1}\Omega(M)=0$. By Proposition 5.1.2 of \[L1\] $\Omega(M)$ is a direct sum of finite-dimensional irreducible ${\frak g}$-modules. Then any highest weight vector for ${\frak g}$ in $\Omega(M)$ meets our need.
[**Claim 3:**]{} [*Any lowest weight vector for $\hat{{\frak g}}$ in $M$ generates a simple $L(\ell,0)$-module.*]{} Let $u$ be a lowest weight vector for $\hat{{\frak g}}$ in $M$. Extracting the constant term from $Y_{M}(e_{\theta},z)^{\ell+1}u=0$, we obtain $(e_{\theta})_{-1}^{\ell+1}u=0$. Then $u$ generates an integrable highest weight $\hat{{\frak g}}$-module. It follows from \[K1\] that $u$ generates an irreducible $\hat{{\frak g}}$-module of level $\ell$. Since any submodule of $M$ for the affine Lie algebra is a submodule of $M$ for $L(\ell,0)$, such $u$ generates a simple $L(\ell,0)$-module.$\;\;\;\;\Box$
\[r2.3’\] This theorem has been proved in \[DLM2\] under the assumption that $t\C[t]\otimes \frak h$ acts locally nilpotently on any weak module. See Proposition 5.6 in \[DLM2\].
Recall that a $\hat{{\frak g}}$-module $M$ is called [*restricted* ]{} (cf. \[K1\]) if for any $u\in M$, there is an integer $k$ such that $(t^{n}\otimes {\frak g})u=0$ for $n>k;$ $M$ is called an [*integrable*]{} module if the Chevalley generators $e_i,f_i$ of $\hat{{\frak g}}$ act locally finitely on $M$ \[K2\] (note that in the definition of integrable module, we do not assume that the action of $\frak h$ is semisimple). At affine Lie algebra level, Proposition \[p2.8\] and Theorem \[t2.10\] essentially assert that any restricted integrable $\hat{{\frak g}}$-module is a direct sum of irreducible highest weight integrable $\hat{{\frak g}}$-modules.
Next we turn our attention to the vertex operator algebras $L(c_{p,q},0).$
First we recall some results from \[DL\]. Let $V$ be a vertex operator algebra and $M$ be a weak $V$-module. Then for any $u,v\in V$ we have $$\begin{aligned}
\label{ef}
Y(u_{-1}v,z)=Y(u,z)^{-}Y(v,z)+Y(v,z)Y(u,z)^{+},\end{aligned}$$ where $$\begin{aligned}
Y(u,z)^{+}=\sum_{n\ge 0}u_{n}z^{-n-1}, \; Y(u,z)^{-}=\sum_{n<0}u_{n}z^{-n-1}.\end{aligned}$$ Note that $Y(u,z)=Y(u,z)^{+}+Y(u,z)^{-}$ and that $Y(u,z)^{+}$ (reps. $Y(u,z)^{-}$) involves only nonpositive (resp. nonnegative) powers of $z$.
For convenience we will write $c=c_{p,q}.$ For any nonnegative integer $n$, we set $\omega^{(n)}={1\over n!}L(-1)^{n}\omega$. Then $L(-n-2)=\omega^{(n)}_{-1}.$ We need the following lemmas.
\[l2.12\] Let $M$ be a weak $L(c,0)$-module and $u\in M.$ Let $k$ be a positive integer such that $L(k)u\ne 0$ and that $L(n)u=0$ whenever $n>k$. Then for any nonnegative integers $n_{1},..., n_{r}$ the lowest power of $z$ in $Y_{M}(\omega^{(n_{1})}_{-1}\cdots \omega^{(n_{r})}_{-1}{\bf 1},z)u$ (in the sense that the coefficients of $z^m$ is zero whenever $m$ is smaller than the lowest power) is $-r(k+2)-n_{1}-\cdots -n_{r}$ with coefficient $\displaystyle{\prod_{i=1}^{r}{-k-2\choose n_{i}}L(k)^{r}u}.$
[**Proof.**]{} We prove this lemma by induction on $r$. If $r=1$, we have: $$\begin{aligned}
& &Y_M(\omega^{(n_{1})}_{-1}{\bf 1},z)u=Y_M(\omega^{(n_{1})},z)u\nonumber\\
& &\ \ \ \ \ =\frac{1}{n_1!}\left({d\over
dz}\right)^{n_{1}}Y_M(\omega,z)u\nonumber\\
& &\ \ \ \ \ =\sum_{n\in {\Z}}{-n-2\choose n_{1}}z^{-n-2-n_{1}}L(n)u
\nonumber\\
& &\ \ \ \ \ =\sum_{n\le k}{-n-2\choose n_{1}}z^{-n-2-n_{1}}L(n)u.\end{aligned}$$ Then the lowest power of $z$ is $-(k+2)-n_{1}$ with a coefficient ${-k-2\choose n_{1}}L(k)u$. That is, the lemma holds for $r=1$.
Suppose that this lemma holds for some positive integer $r$. By formula (\[ef\]) we have: $$\begin{aligned}
& &Y_{M}(\omega^{(n_{1})}_{-1}\cdots
\omega^{(n_{r})}_{-1}\omega^{(n_{r+1})}_{-1}{\bf
1},z)u=Y_{M}(\omega^{(n_{1})},z)^{-}Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)
u\nonumber\\
& &\ \ \ \ \ \ +Y_{M}(\omega^{(n_{2})}_{-1}\cdots \omega^{(n_{r+1})}_{-1}{\bf
1},z)
Y_{M}(\omega^{(n_{1})},z)^{+}u.\label{a2.15}\end{aligned}$$ Since $Y_{M}(\omega^{(n_{1})},z)^{-}$ involves only nonnegative powers of $z$, it follows from the inductive assumption, the lowest power of $z$ in the first term of the right hand side of (\[a2.15\]) is $-r(k+2)-n_{2}-\cdots -n_{r+1}$. It is easy to observe that for any $v$ in the algebra, $$Y_M(L(-1)v,z)^+=\frac{d}{dz}Y_M(v,z)^+.$$ Thus $$\begin{aligned}
& &\ \ \ \ Y_{M}(\omega^{(n_{2})}_{-1}\cdots \omega^{(n_{r+1})}_{-1}{\bf 1},z)
Y_{M}(\omega^{(n_{1})},z)^{+}u\\
& &=\frac{1}{n_1!}Y_{M}(\omega^{(n_{2})}_{-1}\cdots \omega^{(n_{r+1})}_{-1}
{\bf 1},z)\left(\frac{d}{dz}\right)^{n_1}Y_{M}(\omega,z)^{+}u\\
& &=\sum_{n=-1}^k{-n-2\choose n_1}Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)L(n)uz^{-n-2-n_1}\\
& &=\sum_{n=0}^k{-n-2\choose n_1}Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)L(n)uz^{-n-2-n_1}\\
& &\ \ \ \ \ +{-1\choose n_1}L(-1)Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)uz^{-1-n_1}\\
& &\ \ \ \ \ +{-1\choose
n_1}\left(\frac{d}{dz}Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)\right)uz^{-1-n_1}.\end{aligned}$$ Note that $L(m)L(n)u=(m-n)L(m+n)u+L(n)L(m)u=0$ for $0\leq n\leq k$ and $m>k.$ Applying the inductive hypothesis to $L(n)u$ we see that the lowest power of $z$ in $$\sum_{n=0}^k{-n-2\choose n_1}Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)L(n)uz^{-n-2-n_1}$$ is $-(r+1)(k+2)-n_1-\cdots -n_{r+1}$ with coefficient $\displaystyle{\prod_{i=1}^{r+1}{-k-2\choose n_i}L(k)^{r+1}u.}$ Also by the induction assumption the lowest power of $z$ in $L(-1)Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)uz^{-1-n_1}$ is $-r(k+2)-n_1-\cdots -n_{r+1}-1$ and the lowest power of $z$ in $\left(\frac{d}{dz}Y_{M}(\omega^{(n_{2})}_{-1}\cdots
\omega^{(n_{r+1})}_{-1}{\bf 1},z)\right)uz^{-1-n_1}$ is $-r(k+2)-n_1-\cdots -n_{r+1}-2.$ Thus the lowest power of $z$ in the second term of right hand side of (\[a2.15\]) is $-(r+1)(k+2)-n_{1}-\cdots -n_{r+1}$ with coefficient $\displaystyle{\prod_{i=1}^{r+1}\left(\begin{array}{c}-k-2\\n_{i}\end{array}\right)
L(k)^{r+1}u},$ as desired. $\;\;\;\;\Box$
Let $V$ be a vertex operator algebra and let $A(V)$ (which is a certain quotient space of $V$ modulo a subspace $O(V)$) be the corresponding associative algebra defined in \[Z\]. We refer the reader to \[Z\] for details. Recall from \[DLM1\] or \[L2\] that for any weak $V$-module $M,$ $\Omega(M)$ consists of vectors $u\in M$ such that $a_{m}u=0$ for any homogeneous element $a\in V$ and for any $m>{\rm wt}\,a-1$. In other words, $u\in \Omega(M)$ if and only if $z^{m}Y_{M}(a,z)u\in M[[z]]$ for any homogeneous element $a\in V$ and for any $m>{\rm wt}a-1$. The following result can be found in \[DLM1\], \[L2\] and \[Z\].
\[al\] (1) $\omega+O(V)$ is in the center of $A(V).$
\(2) $A(V)$ is semisimple if $V$ is rational.
\(3) $\Omega(M)$ is an $A(V)$-module under the action $a+O(V)\mapsto
a_{\wt\,a-1}$ for homogeneous $a\in V.$
Now we take $V=L(c,0).$ Set $\bar{\Omega}(M)=\{ u\in M|L(n)u=0\;\mbox{ for any }n>0\}$. Then it is clear that $\Omega(M)\subseteq \bar{\Omega}(M)$.
\[l2.13\] Let $M$ be a weak $L(c,0)$-module. Then $\Omega(M)=\bar{\Omega}(M)$.
[**Proof.**]{} It suffices to prove that $a_{m}u=0$ for any $u\in
\bar{\Omega}(M)$ and for any homogeneous element $a\in L(c,0)$ whenever $m>{\rm wt}a-1$. We shall prove this by induction on the weight of $a$. If ${\rm wt}a=0$, $a={\bf 1}$. Since ${\bf 1}_{m}=0$ for $m\ge 0$, there is nothing to prove. Suppose that $a_{m}u=0$ for any homogeneous element $a\in L(c,0)$ of weight less than $n$ and for any $m>{\rm wt}a-1$. Let $b\in L(c,0)$ be a homogeneous element of weight $n$ and let $m\in {\Z}$ such that $m>{\rm wt}b-1$.
Let $a\in L(c,0)$ be any homogeneous element of weight less than $n$, let $k$ be any positive integer and let $m>{\wt}\,(L(-k)a)-1$ $(={\rm wt}\,a+k-1)$. Then from the Jacobi identity (\[jacobi\]) we have: $$\begin{aligned}
& &(L(-k)a)_{m}u={\rm Res}_{z_{0}}{\rm
Res}_{z_{2}}z_{0}^{1-k}z_{2}^{m}Y_{M}(Y(\omega,z_{0})a,z_{2})u\nonumber\\
& &\ \ ={\rm Res}_{z_{1}}{\rm Res}_{z_{0}}{\rm
Res}_{z_{2}}z_{0}^{1-k}z_{2}^{m}\cdot \nonumber\\
& &\ \ \ \cdot\left(
%% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR
z_{0}^{-1}\delta\left(\frac{z_{1}-z_{2}}{z_{0}}\right)Y_{M}(\omega,z_{1})Y_{M}(a,z_{2})u
%% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR
-z_{0}^{-1}\delta\left(\frac{z_{2}-z_{1}}{-z_{0}}\right)Y_{M}(a,z_{2})Y_{M}(\omega,z_{1})u\right)
\nonumber\\
& &\ \ ={\rm Res}_{z_{1}}{\rm
Res}_{z_{2}}(z_{1}-z_{2})^{1-k}z_{2}^{m}Y_{M}(\omega,z_{1})Y_{M}(a,z_{2})u
\nonumber\\
& &\ \ \ -{\rm Res}_{z_{1}}{\rm
Res}_{z_{2}}(-z_{2}+z_{1})^{1-k}z_{2}^{m}Y_{M}(a,z_{2})Y_{M}(\omega,z_{1})u.\end{aligned}$$ Since $m>{\rm wt}\,(L(-k)a)-1={\rm wt}a+k-1>{\rm wt}a-1$, we have: $${\rm Res}_{z_{1}}{\rm
Res}_{z_{2}}(z_{1}-z_{2})^{1-k}z_{2}^{m}Y_{M}(\omega,z_{1})Y_{M}(a,z_{2})u=0.$$ For the second term, we have: $$\begin{aligned}
& &\ \ \ -{\rm Res}_{z_{1}}{\rm
Res}_{z_{2}}(-z_{2}+z_{1})^{1-k}z_{2}^{m}Y_{M}(a,z_{2})Y_{M}(\omega,z_{1})u
\nonumber\\
& &=-{\rm Res}_{z_{2}}(-1)^{1-k}z_{2}^{m+1-k}Y_{M}(a,z_{2})L(-1)u
-{\rm Res}_{z_{2}}(-1)^{-k}(1-k)z_{2}^{m-k}Y_{M}(a,z_{2})L(0)u\nonumber\\
& &=-{\rm Res}_{z_{2}}(-1)^{1-k}z_{2}^{m+1-k}L(-1)Y_{M}(a,z_{2})u
+{\rm Res}_{z_{2}}(-1)^{1-k}z_{2}^{m+1-k}{d\over
dz_{2}}Y_{M}(a,z_{2})u\nonumber\\
& &\ \ \ -{\rm
Res}_{z_{2}}(-1)^{-k}(1-k)z_{2}^{m-k}Y_{M}(a,z_{2})L(0)u\nonumber\\
& &=-{\rm Res}_{z_{2}}(-1)^{1-k}z_{2}^{m+1-k}L(-1)Y_{M}(a,z_{2})u
-{\rm Res}_{z_{2}}(-1)^{1-k}(m+1-k)z_{2}^{m-k}Y_{M}(a,z_{2})u\nonumber\\
& &\ \ \ -{\rm
Res}_{z_{2}}(-1)^{-k}(1-k)z_{2}^{m-k}Y_{M}(a,z_{2})L(0)u\nonumber\\
&
&=(-1)^kL(-1)a_{m+1-k}u+(-1)(m+1-k)a_{m-k}u+(-1)^{k-1}a_{m-k}L(0)u.\label{2.17}\end{aligned}$$ Since $L(0)u\in \bar{\Omega}(M)$ and $m-k>{\rm wt}a-1$ all the three terms in (\[2.17\]) are zero by the inductive hypothesis. Thus $(L(-k)a)_mu=0.$ Note that $b$ is a linear combination of all $L(-k)a$, where ${\rm wt}a<n$ and $k$ is a positive integer. This shows $b_{m}\bar{\Omega}(M)=0$ for $m>{\rm wt}\,b-1,$ as desired. $\;\;\;\;\Box$
Now we are in a position to prove The vertex operator algebra $L(c,0)$ associated with the lowest weight irreducible module for the Virasoro algebra with central charge $c=c_{p,q}$ is regular.
[**Proof:**]{} By Proposition \[p2.8\], it is enough to prove that any nonzero weak $L(c,0)$-module $M$ contains a simple $L(c,0)$-module.
[**Claim 1:**]{} [*The space $\Omega(M)$ is not zero.*]{} For any $0\ne u\in M$, we define $l(u)$ to be the integer $k$ such that $L(k)u\ne 0$ and $L(n)u=0$ whenever $n>k$. Since $L(n)u\ne 0$ for some $n$ (because $c\ne 0$), $l(u)$ is well-defined. Suppose that $\Omega(M)=0$. Then by Lemma \[l2.13\] $l(u)\ge 1$ for any $0\ne u\in M$. Let $0\ne u\in M$ such that $l(u)=k$ is minimal. It is well known \[FF\] that there are two singular vectors in the Verma module $M(c,0)$ for the Virasoro algebra. One singular vector is $L(-1){\bf 1}$ and the other is: $$\begin{aligned}
v=L(-2)^{pq}{\bf 1}+\sum a_{n_{1},\cdots, n_{r}}\omega^{(n_{1})}_{-1}\cdots
\omega^{(n_{r})}_{-1}{\bf 1},\end{aligned}$$ where the sum is over some $(n_{1},\cdots, n_{r})\in {\Z}_{+}^{r}$ such that $2pq=2r+n_{1}+\cdots +n_{r}$ and $n_{1}+\cdots +n_{r}\ne
0$. By Lemma \[l2.12\], the lowest power of $z$ in $$Y_{M}(L(-2)^{pq}{\bf 1},z)u=Y_{M}((\omega_{-1})^{m}{\bf 1},z)u$$ is $-pq(k+2)$ with $L(k)^{pq}u$ as its coefficient and the lowest power of $z$ in $$Y_{M}(\omega^{(n_{1})}_{-1}\cdots \omega^{(n_{r})}_{-1}{\bf
1},z)u$$ is greater than $-pq(k+2)$ for any nonnegative integers $n_{1},\cdots, n_{r}$ such that $2pq=2r+n_{1}+\cdots +n_{r}$ and $n_{1}+\cdots +n_{r}\ne 0$. Thus the coefficient of $z^{-pq(k+2)}$ in $Y_{M}(v,z)u$ is $L(k)^{pq}u$. Since $v=0$ in $L(c,0)$ we have $Y_{M}(v,z)=0$. In particular the coefficient $L(k)^{pq}u$ of $z^{-pq(k+2)}$ in $Y_M(v,z)$ is zero. Let $s$ be the nonnegative integer such that $L(k)^{s}u\ne 0$ and $L(k)^{s+1}u=0$ and set $u'=L(k)^{s}u$. Then it is clear that $l(u')<l(u)$. This is a contradiction.
[**Claim 2:**]{} [*Any weak $L(c,0)$-module $M$ contains a simple ordinary $L(c,0)$-module.*]{} Since $L(c,0)$ is rational, Lemma \[al\] tells us that $A(L(c,0))$ is semisimple and that the central element $\omega+O(L(c,0))$ acts semisimply on $\Omega(M)$ as $L(0).$ Since $\Omega(M)$ is nonzero by Claim 1 we can take $0\ne u\in \Omega(M)$ such that $L(0)u=hu$ where $h\in \C.$ Again since $L(c,0)$ is rational \[W\], $u$ generates a simple (ordinary) $L(c,0)$-module. The proof is complete. $\;\;\;\;\Box$
\[cm\] The moonshine module vertex operator algebra $V^{\natural}$ is regular.
[**Proof.**]{} From \[DMZ\], $V^{\natural}$ contains $L({1\over
2},0)^{\otimes 48}$ as a vertex operator subalgebra. Then the result follows from Theorem \[t2.11\], Propositions \[p2.5\] and \[p2.6\]. $\;\;\;\;\Box$
Finally we discuss the complete reducibility of weak $V_L$-modules for an even lattice $L.$ We refer the reader to \[FLM\] and \[D1\] for the construction of $V_L$ and related notations.
Let $M$ be any weak $V_L$-module. Define the vacuum space $$\Omega_M=\{u\in M| \alpha(i)u=0\ {\rm for}\ \a\in L,\ i>0\}.$$
\[p2.14\] Let $L$ be an even lattice. Then for any weak $V_{L}$-module $M,$ $\Omega_M\ne 0.$
[**Proof.**]{} For $u\in M$ then $A_u=\span\{\a(n)u|\alpha\in L, n>0\}$ is finite-dimensional as $\a(n)u=0$ if $n$ is sufficiently large and as the rank of $L$ is finite. Set $d(u)=\dim A_u.$ Note that $d(u)=0$ if and only if $u\in \Omega_M.$ So it is enough to show that $d(u)=0$ for some nonzero $u\in M.$ Assume this is false, and take $0\ne u\in M$ such that $d(u)$ is minimal.
Let $k$ be the smallest positive integer such that $\a(k)u\ne 0$ and $\b(n)u=0$ whenever $n>k$ for some $\a\in L$ and all $\b\in L.$
Let $a\in \hat L$ such that $\bar a=\a.$ Then from the formula (3.4) of \[D1\] we have $$\frac{d}{dz}Y(\iota(a),z)=Y(L(-1)\iota(a),z)=Y(\alpha(-1)\iota(a),z)
=\alpha(z)^-Y(\iota(a),z)+Y(\iota(a),z)\alpha(z)^+$$ where $$\alpha(z)^-=\sum_{n<0}\alpha(n)z^{-n-1},\ \
\alpha(z)^+=\sum_{n\geq 0}\alpha(n)z^{-n-1}.$$ Clearly the submodule generated by $u$ is not zero. Note that the vertex algebra $V_L$ is simple (see \[D1\]). By Proposition 11.9 of \[DL\], $Y(\iota(a),z)u\ne 0.$ Let $r$ be an integer such that $\iota(a)_{r+m}u=0$ and $\iota(a)_{r}u\ne 0$ for any positive integer $m.$ Thus the lowest power of $z$ in $$\frac{d}{dz}Y(\iota(a),z)u=-\sum_{m\leq r}(m+1)\iota(a)_muz^{-m-2}$$ is at most $-r-2.$ It is obvious that the lowest power of $z$ in $
\alpha(z)^-Y(\iota(a),z)u$ is at most $-r-1.$
Use the following commutator formula which is a result from the Jacobi identity $$\label{gr}
[\beta(m),\iota(a)_n]=\<\a,\b\>\iota(a)_{m+n}$$ to obtain $$\iota(a)_m\alpha(n)u=-\<\a,\a\>\iota(a)_{m+n}u+\alpha(n)\iota(a)_mu=0$$ if $m>r$ and $n\geq 0.$ This gives $$Y(\iota(a),z)\alpha(z)^+u=
\sum_{m\leq r}\sum_{n=0}^k\iota(a)_m\a(n)z^{-m-n-2}.$$ Thus the coefficient $\iota(a)_{r}\a(k)u$ of $z^{-r-k-2}$ in the formula above is zero as $k$ is positive. This shows by (\[gr\]) again that $\alpha(n)\iota(a)_ru=0$ for any positive integer greater than or equal to $k.$
Note from (\[gr\]) that if $\b(m)u=0$ for positive $m$ then $\b(m)\iota(a)_ru=0.$ Thus $d(\iota(a)_ru)< d(u).$ This is a contradiction. $\;\;\;\;\Box$
Let $L$ be an even lattice. Then any weak $V_L$-module is completely reducible and any simple weak $V_L$-module is isomorphic to $V_{L+\beta}$ for some $\b$ in the dual lattice of $L.$ In particular, $V_L$ is regular if $L$ is positive definite.
[**Proof.**]{} By Lemma \[p2.14\], $\Omega_M\ne 0$ for a weak $V_L$-module $M$. It is proved in \[D1\] that if $M$ is also simple then it is necessarily isomorphic to $V_{L+\beta}$ for some $\b$ in the dual lattice of $L.$ So it remains to show the complete reducibility of any weak $V_L$-module $M.$
Let $W$ be the sum of all simple submodules of $M.$ Assume that $M'=M/W$ is not zero. Then $\Omega_{M'}\ne 0.$ It is essentially proved in \[D1\] that $M'$ contains a simple module $W^1/W$ (here $W^1$ is a weak $V_L$-submodule of $M$ which contains $W$) generated by $w^1+W$ where $w^1$ is a common eigenvector for the operators $\a(0)$ for $\a\in L.$ It follows from the proof of Theorem 3.1 of \[D1\] that the submodule of $M$ generated by $w^1$ is simple. Thus $W^1$ is a sum of certain simple submodules of $V_L.$ This is a contradiction. $\;\;\;\;\Box$
At this point we have proved that almost all known rational vertex operator algebras are regular. We conclude this paper by presenting the following conjecture:
Any rational vertex operator algebra is regular.
[FLM1]{}
A. Belavin, A. M. Polyakov, A. B. Zamolodchikov, Infinite conformal symmetries in two-dimensional quantum field theory, [*Nucl. Phys. B*]{} [**241**]{} (1984), 333-380.
R. E. Borcherds, Vertex algebras, Kac-Moody algebras, and the Monster, [*Proc. Natl. Acad. Sci. USA*]{} [**83**]{} (1986), 3068-3071.
C. Dong, Vertex algebras associated with even lattices, [*J. of Algebra*]{} [**161**]{} (1993), 245-265.
C. Dong, Representations of the moonshine module vertex operator algebra, [*Contemporary Math.*]{}, [**175**]{} (1994), 27-36.
C. Dong and J. Lepowsky, Generalized Vertex Algebras and Relative Vertex Operators, Progress in Math., [**Vol. 112**]{}, Birkhaüser, Boston, 1993.
C. Dong, H.-S. Li and G. Mason, Twisted representations of vertex operator algebras, preprint.
C. Dong, H.-S. Li and G. Mason, Simple currents and extensions of vertex operator algebras, preprint, q-alg/9504008.
C. Dong, G. Mason and Y. Zhu, Discrete series of the Virasoro algebra and the moonshine module, [*Proc. Symp. Pure. Math., American Math. Soc.*]{} [**56**]{} II (1994), 295-316.
B. L. Feigin and D. B. Fuchs,Verma modules over the Virasoro algebra, Lecture Notes in Math., Vol. 1060, Springer-Verlag, Berlin and New York, 1984.
M. Finkelberg, Fusion categories, Ph.D. thesis, Harvard University, 1993.
I. Frenkel, Y.-Z. Huang and J. Lepowsky, On axiomatic approaches to vertex operator algebras and modules, preprint, 1989; Memoirs Amer. Math. Soc. [**104**]{}, 1993.
I. Frenkel, J. Lepowsky and A. Meurman, [*Vertex Operator Algebras and the Monster*]{}, Pure and Appl. Math., [**Vol. 134**]{}, Academic Press, Boston, 1988.
I. Frenkel and Y.-C. Zhu, Vertex operator algebras associated to representations of affine and Virasoro algebras, [*Duke Math. J.*]{} [**66**]{} (1992), 123-168.
Y.-Z. Huang and J. Lepowsky, Toward a theory of tensor products for representations of a vertex operator algebra, in: [*Proc. 20th Intl. Conference on Differential Geometric Methods in Theoretical Physics, New York, 1991,*]{} ed. S. Catto and A. Rocha, World Scientific, Singapore, 1992, Vol. 1, 344-354.
J. Humphreys, An introduction to Lie algebras and representation theory, Graduate Textbook in Mathematics, Springer, New York, 1978.
V. G. Kac, [*Infinite-dimensional Lie Algebras*]{}, 3rd ed., Cambridge Univ. Press, Cambridge, 1990.
V. G. Kac, Constructing groups associated to infinite-dimensional Lie algebras, in [*Infinite-dimensional Groups with Applications,*]{} MSRI Publ. 4, Springer-Verlag, 1985, 167-216.
D. Kazhdan and G. Lusztig, Affine Lie algebras and quantum groups, [*Duke Math. J.*]{} [**2**]{} (1991), 21-29.
I. G. Koh and P. Sorba, Fusion rules and (sub)modular invariant partition functions in non-unitary theories, [*Phys. Lett.*]{} [**215**]{} (1988), 723-739.
H.-S. Li, Local systems of vertex operators, vertex superalgebras and modules, [*Journal of Pure and Applied Algebra*]{}, to appear.
H.-S. Li, Representation theory and tensor product theory for a vertex operator algebra, Ph.D. thesis, Rutgers University, 1994.
P. Mathieu and A. A. Walton, Fractional-level Kac-Moody algebras and nonunitary coset conformal theories, [*Progress of Theoretical Physics Supplement*]{} [**102**]{} (1990), 229-254.
A. Meurman and M. Primc, Annihilating ideals of standard modules of $\tilde{sl}(2,{\bf C})$ and combinatorial identities, [*Advances in Math.*]{} [**64**]{} (1987), 177-240.
A. Meurman and M. Primc, Annihilating fields of standard modules of $\tilde{sl}(2,{\bf C})$ and combinatorial identities, preprint.
G. Moore and N. Seiberg, Classical and quantum conformal field theory, [*Comm. Math. Phys.*]{} [**123**]{} (1989), 177-254.
W. Wang, Rationality of Virasoro vertex operator algebras, [*Duke Math. J. IMRN*]{}, [**Vol. 71**]{}, No. 1 (1993), 197-211. Y.-C. Zhu, Vertex operator algebras, elliptic functions and modular forms, Ph.D. thesis, Yale University, 1990.
[^1]: Supported by NSF grant DMS-9303374 and a research grant from the Committee on Research, UC Santa Cruz.
[^2]: Supported by NSF grant DMS-9401272 and a research grant from the Committee on Research, UC Santa Cruz.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'H. N. Smitha'
- 'R. Holzreuter'
- 'M. van Noort'
- 'S. K. Solanki'
title: |
The influence of NLTE effects in Fe [i]{} lines on an inverted atmosphere\
I. 6301Å and 6302Å lines formed in 1D NLTE
---
[Ultraviolet (UV) overionisation of iron atoms in the solar atmosphere leads to deviations in their level populations based on Saha-Boltzmann statistics. This causes their line profiles to form in non-local thermodynamic equilibrium (NLTE) conditions. When inverting such profiles to determine atmospheric parameters, the NLTE effects are often neglected and other quantities are tweaked to compensate for deviations from the local thermodynamic equilibrium (LTE). ]{} [We investigate how the routinely employed LTE inversion of iron lines formed in NLTE underestimates or overestimates atmospheric quantities, such as temperature ($T$), line-of-sight velocity ($v_{\rm LOS}$), magnetic field strength ($B$), and inclination ($\gamma$) while the earlier papers have focused mainly on $T$. Our findings has wide-ranging consequences since many results derived in solar physics are based on inversions of Fe [i]{} lines carried out in LTE.]{} [We synthesized the Stokes profiles of Fe [i]{} 6301.5Å and 6302.5Å lines in both LTE and NLTE using a snapshot of a 3D magnetohydrodynamic simulation. The profiles were then inverted in LTE. We considered the atmosphere inferred from the inversion of LTE profiles as the fiducial model and compared it to the atmosphere resulting from the inversion of NLTE profiles. The observed differences have been attributed to NLTE effects.]{} [Neglecting the NLTE effects introduces errors in the inverted atmosphere. While the errors in $T$ can go up to $13\%$, in $v_{\rm LOS}$ and $B,$ the errors can go as high as $50\%$ or above. We find these errors to be present at all three inversion nodes. Importantly, they survive degradation from the spatial averaging of the profiles.]{} [We provide an overview of how neglecting NLTE effects influences the values of $T, v_{\rm LOS}$, $B,$ and $\gamma$ that are determined by inverting the Fe [i]{} 6300Å line pair, as observed, for example, by Hinode/SOT/SP. Errors are found at the sites of granules, intergranular lanes, magnetic elements, and basically in every region susceptible to NLTE effects. For an accurate determination of the atmospheric quantities and their stratification, it is, therefore, important to take the NLTE effects into account.]{}
{width="70.00000%"}
Introduction {#sec:intro}
============
The photospheric spectral lines of iron at 6301.5Å and 6302.5Å are among the most widely observed lines for both space and ground-based telescopes. Spectral analysis and the inversion of their Stokes profiles is often carried out based on the assumption of local thermodynamic equilibrium (LTE). The overionization of iron atoms by ultraviolet (UV) radiation, however, results in an underpopulation of the Fe [i]{} atomic levels, which leads to deviations from their LTE values. Moreover, with a core formation height between 250km – 300km above the photosphere , their line source function is less than the Planck function, which introduces additional non-LTE (NLTE) effects in the lines themselves. While modeling their spectral profiles, these effects should be taken into account.
One of the first investigations of the influence of NLTE on iron lines was by [@1972ApJ...176..809A], which was later followed by others, such as . Their studies were, however, limited to one-dimensional (1D) radiative transfer. Recently, in a series of papers, investigated the effects of three-dimensional (3D) NLTE radiative transfer on iron lines formed in thin flux tubes and flux sheets in a snapshot of a 3D radiation hydrodynamic (HD) simulation and in a snapshot of a 3D magnetohydrodynamic (MHD) simulation. In each case, the profiles of the iron lines computed based on the assumption of LTE and 1D NLTE were compared with the 3D NLTE profiles. They reported that in some regions of the atmospheric cube, the NLTE line profiles show line weakening as compared to LTE due to UV irradiation, while a few other regions demonstrated line strengthening due to the departure of the source function from the Planck function, which is in agreement with the findings of @2001ApJ...550..970S. The NLTE effects were shown to be particularly significant in spatially resolved profiles.
Accounting for the NLTE effects in iron lines is important for the determination of solar and stellar atmospheric parameters, including elemental abundance . In earlier papers involving the inversion of iron line profiles for the determination of atmospheric models and iron abundances, the NLTE effects have generally been neglected . To our knowledge, there has, so far, been only one exception. Using the NICOLE NLTE inversion code, inverted observations from the 6302.5Å line taken by the *Hinode* satellite to infer a 3D solar atmospheric model. The NLTE effects were taken into account by using the departure coefficients computed from a 3D HD model. They found that with the NLTE correction, the line cores are better fitted and the inverted atmosphere is less noisy.
The current paper aims to investigate the errors introduced in the inferred atmospheres when the NLTE effects in the Fe [i]{} 6301.5Å and 6302.5Å lines are neglected during inversions. For this, we first computed the Stokes profiles from an MHD cube in both LTE and in NLTE. Since we could not at present carry out a fully self-consistent NLTE inversion of Fe [i]{} lines, we decided to test how well LTE inversions (referred to as just an “inversion”) of NLTE lines work by comparing them with inversions of LTE lines. That is, we inverted both NLTE and LTE line profiles using an LTE inversion code. To get a rough idea of the errors introduced by neglecting NLTE effects during the inversion, the model atmosphere obtained from the inversion of profiles computed in LTE from the MHD snapshot was used as a reference and the atmosphere recovered from the inversion of NLTE profiles was compared with this reference. Any differences were then classified as errors that the inversion makes by neglecting the NLTE effects.
Earlier studies by focused mainly on the influence of NLTE effects on temperature measurements by analysing the intensity profiles. In the present paper, we also discuss the influence of NLTE effects on the line-of-sight (LOS) velocity and magnetic field diagnostics by comparing the Stokes profiles ($I,Q,U,V$) computed in LTE and NLTE. Here we focus only on the 1D NLTE effects (simply referred to as NLTE) and neglect the effects of horizontal radiative transfer. The influence of horizontal radiative transfer will be investigated in a follow-up paper.
Finally, we tested whether the spatial averaging of the Stokes profiles influences the accuracy of the atmosphere derived in LTE. For this, we rebinned the profiles to match the spatial scales resolved by *Hinode* and inverted the profiles as before. We find that because the NLTE effects indeed appear to survive such spatial degradation, it is important to account for them to infer atmospheric quantities accurately.
{width="80.00000%"}
Model atmosphere and Stokes profiles synthesis {#sec:atmos}
==============================================
A Snapshot of a realistic 3D MHD model atmosphere generated by the MURaM code was used for the synthesis of the Stokes profiles. The cube extends 6 Mm $\times$ 6 Mm $\times$ 2 Mm in the x, y, and z directions, respectively, with a resolution of 5.82 km $\times$ 5.82 km $\times$ 7.85 km. The cube represents the quiet Sun with mixed magnetic polarities.
Maps of the temperature ($T$), LOS velocity ($v_{\rm LOS}$) and magnetic field strength ($B$) at $log(\tau)=-2.0, -0.9,$ and $0.0$ from the 3D MHD cube are shown in Figure \[fig:maptau\]. In the rest of the paper, by $log(\tau)$ we mean $log(\tau_{5000})$. We choose these three $\tau$ surfaces because the nodes for inversions are placed there, as will be discussed in Section \[sec:inversion\]. As we move higher up in the atmosphere, from $log(\tau)=0$ to $-2.0$, the temperature maps clearly change from normal to reverse granulation, with some granules still hotter than others at $log(\tau)=-2.0$, but clearly cooler than the intergranular lanes. As the height of the $log(\tau)=-0.9$ surface is approximately midway through this transition, the distribution of the temperature across the granules appears more homogeneous. Regions with a high magnetic field strength stand out as hot regions between the granules, with temperatures much higher than the average in intergranular lanes. A narrowing of the intergranular lanes and increase in the convective velocities (upflows and downflows) with depth can be observed.
The Stokes profiles of the 6301.5Å and 6302.5Å iron lines were synthesized in 1D using the RH code [@2001ApJ...557..389U] in both LTE and NLTE, at $\mu=1$. By 1D computations, we really mean 1.5D, that is 1D radiative transfer carried out along each column of the MHD cube, while neglecting all 3D radiative transfer effects in the profile synthesis. All calculations were done in the same way as given in . Similar to the computations in that paper, we neglect the influence of NLTE effects on the concentration of H- ions, which can affect the continuum . The full Stokes profiles were computed using the field-free approximation [@1969SoPh...10..268R]. The logarithmic iron abundance was taken to be 7.50 on a scale on which hydrogen has an abundance of 12.
We used the same iron atomic model that was used in . The model contains 23 atomic levels coupled by 33 bound-bound transitions and 22 bound-free transitions. The bound-free transitions represent the coupling of atomic levels in each term in Fe [i]{} to its parent term in Fe [ii]{} through photoionization. For the photoionization cross-sections, hydrogen-like approximations have been used. The missing iron line opacity in the UV was approximated by enhancing the opacities in the relevant wavelength range, which is known as opacity fudging, as proposed by . have tested the effect of such an opacity fudging on the synthesized Fe [i]{} lines. They find that the 6301.5Å line is less sensitive to the opacity fudging compared to the 5250Å line. Removing the opacity fudging will likely enhance the NLTE effects in these lines possibly to unrealistically large values.
The continuum intensity map at 6300Å is shown in Figure \[fig:cont1\]. To facilitate the comparison of the different cases, we choose five representative spatial points within the white box, inside a granule (pos. 1 and 2), in an intergranular lane (pos. 3), intergranular lane next to a magnetic element (pos. 4) and in the middle of a strong magnetic element (pos. 5), the last two of which, following the convention in , may be referred to as a flux sheet and a flux tube, respectively. Although we discuss the full maps of the different atmospheric quantities obtained from inversions, at these five spatial locations we also discuss the Stokes profiles in detail.
Inversion of Stokes profiles {#sec:inversion}
============================
The Stokes profiles computed assuming LTE and NLTE conditions were inverted using the 1D LTE inversion code Stokes-Profiles-INversion-O-Routines, , which makes use of the STOPRO routines [@1987PhDT.......251S]. The code assumes a strongly simplified model of the atmospheric stratification, consisting of a spline representation of the temperature, magnetic field strength, inclination ($\gamma$), azimuth, and LOS velocity as a function of optical depth. The splines are controlled by the values of each atmospheric quantity at three points, referred to as nodes, placed at strategically chosen optical depth values throughout the formation region of the spectral lines to be inverted. To avoid any arbitrariness as much as possible, for the inversions presented in this paper, the nodes were placed at $log(\tau)=-2.0, -0.9$ and $0.0$, close to the optimum found by . No microturbulent velocity was used in fitting the Stokes profiles since none was used to synthesize them in the first place. The SPINOR code only returns the values of the atmospheric quantities at the positions of the three nodes and the comparison between the atmospheres from the inversion of LTE and NLTE profiles are, therefore, shown only on these three iso-$(\tau)$ surfaces.
In Figure \[fig:icomp\], we show the input intensity maps and the best fits at six wavelength points; specifically, at the centers of the two lines, and $\pm 0.05$Å on either side of the line centers. The plotted intensities are normalized to $I_c$, the continuum intensity, spatially averaged over the surface of the whole cube. The accuracy of the fit is demonstrated using the scatter density plots. The intensity maps look almost identical in both cases at all six points. The scatter plots are narrow, with the majority of the data points lying along the diagonal. The same is true for the Stokes $Q, U$ and $V$ profiles (figures not shown). The low scatter of the values around the diagonal shows that the inversion code is able to fit the LTE and NLTE profiles equally well and, therefore, any difference in the two (LTE and NLTE) input profiles is translated by the inversion into differences in the atmospheric parameters.
{width="49.00000%"} {width="49.00000%"}
{width="85.00000%"}
{width="101.00000%"}
{width="38.00000%"} {width="59.50000%"}
Comparison of the inverted atmospheres {#sec:atm_comp}
======================================
An ideal way to quantify the quality of the model atmospheres returned by the inversions would be to compare them with the MHD cube. For Milne-Eddington inversions, one could follow a strategy similar to that of and compare the two. However, when we have a stratified atmosphere, comparison with the MHD cube is complicated and there is no unique way to do so. The method used for such a comparison would introduce additional uncertainties. To avoid this, we first inverted the LTE profiles in LTE. This case is consistent so that the results should be reasonably correct (although grossly simplified compared with the original MHD snapshot). We used this atmosphere as the fixed reference and called it the “reference model”. We then inverted the NLTE profiles in LTE and referred to the atmosphere obtained as the “test model”. This latter case is not self-consistent and differences between the reference model and test model are likely due to the NLTE effects. A similar approach was used in but the input was re-synthesized from the fitted atmosphere, which is not done in the present paper
The reference and test models are compared in two ways, one simply using the difference $$\Delta x = x^{\rm LTE}-x^{\rm NLTE},
\label{eqn:abs_diff}$$ and the other using the relative difference defined as $$\delta x = \frac{x^{\rm LTE}-x^{\rm NLTE}}{x^{\rm LTE}},
\label{eqn:rel_diff}$$ where $x$ represents an atmospheric quantity, such as temperature, velocity, magnetic field strength, or inclination.
The two spectral lines at 6301.5Å and 6302.5Å are formed approximately 100 km apart in the middle to upper photosphere , with the 6301.5Å line forming somewhat higher than the 6302.5Å line, but with a smaller Landé g-factor. Due to its formation height, the 6301.5Å line is more affected by the NLTE conditions than the 6302.5Å line. Since the two lines are inverted together, the NLTE sensitivity of both lines contributes towards shaping the final inverted atmosphere, making the interpretation more complex, especially with a stratified atmosphere.
Different parts of a spectral line, from continuum to the line core, are formed at different heights in the atmosphere. So the contributions to the inferred atmosphere at the three nodes come from different parts of the spectral line. This information is best captured by the so-called response functions, which tells us how the values of the Stokes parameters at each wavelength point responds to a change of a physical quantity in the atmosphere at a particular depth. In Figure \[fig:rf\_int\], we show the LTE response functions of intensity to temperature at the five chosen spatial points. The value of the response functions, normalized to unity and represented by a color value, is plotted as a function of the spectral range and the optical depth $log(\tau)$. Over-plotted on this are the LTE intensity profiles at the five locations. Each spectral point has a non-zero response over a range of optical depths. For example, the response function at the line core is strongest at lower optical depths (larger heights) in the atmosphere and decreases with depth. As a first approximation, we assume that the contribution to the top node $log(\tau)=-2.0$ comes mainly from the line core. Such an assumption will help us explain the differences in temperature maps by correlating them to the differences in the line core intensities of the LTE and NLTE profiles. In the right panels of the same figure, we also compare the LTE and NLTE intensity profiles at the five points. We compute the relative differences in intensity (or residual intensity) and equivalent widths similar to Equations 1 and 2 of . The expressions are repeated here for convenience, $$\begin{aligned}
\delta I = \frac{I^{\rm LTE}-I^{\rm NLTE}}{I^{\rm NLTE}};
\delta E = \frac{EW^{\rm LTE}-EW^{\rm NLTE}}{EW^{\rm NLTE}},
\label{eqn:int_ew}\end{aligned}$$ where $I^{\rm LTE, NLTE}$ is the minimum intensity of an LTE/NLTE profile at a specific spatial location and $EW^{\rm LTE, NLTE}$ is the equivalent width. A $\delta I < 0$ or a $\delta E >0$ corresponds to NLTE weakening of the line. An overview of the variation in $\delta E$ for the 6301.5Å line across different atmospheric structures from an MHD cube, like granules, intergranular lanes, magnetic elements (flux tubes and flux sheets) etc., can be found in Figure 3 of . In Figure \[fig:rf\_int\], we indicate the values of $\delta I$ and $\delta E$ only at the representative five spatial points computed from the intensity profiles, for both the 6301.5Å and the 6302.5Å lines.
Temperature {#sec:temp}
-----------
### Top and central nodes {#sec:temp_nodes12}
The temperature maps from the reference and test models at $log(\tau)=-2.0, -0.9,$ and $0.0$ are compared in Figure \[fig:temp\] (*first and second columns*). We first discuss the maps in the top two nodes ($log(\tau)=-2.0$ and $-0.9$) and consider the bottom node ($log(\tau)=0.0$) separately in Section \[sec:temp\_node0\]. In Figure \[fig:temp\], the temperature maps from the test model look quite similar to those of the reference model. However, the difference map in the *third column* indicates that the $\Delta T$ is mostly negative (i.e., the temperature in the test model is higher than in the reference model) in the granules and can reach values as high as 300K in some patches. The higher temperature in the test model is also reflected in the histogram of $\Delta T$ (*fourth column*), which displays an increased skewness towards negative values. The mean and standard deviation of the histogram distribution (first and second numbers on the left side, respectively) are less than 100 K. The density plots of the relative difference $\delta T$, computed using Equation \[eqn:rel\_diff\], are shown in the last column. In the top node, we find $\delta T$ reaching up to -13% in regions where $T^{\rm LTE}$ $<$ 5000K, which look like reverse granulation. In regions of higher temperature ($T^{\rm LTE} > 5000 K$), $\delta T$ can be positive or negative depending on the temperature gradients in those regions, and it remains below 5%. Overall, the density of points is within $\pm 5\%$ in the top node. In the middle node, $\delta T$ is again prominently negative in the granules reaching up to 10% while in the intergranular lanes it is less than 5%.
{width="69.00000%"}
[ccccccc]{} & $log(\tau)$ & Pos 1 & Pos 2 & Pos 3 & Pos 4 & Pos 5\
\
&-2.0 & 16.8 (4721) & -158.5 (4490) & -126.2 (4404) & 121.8 (4919)& -130.7 (5299)\
(K) &-0.9 & -108.5 (5036) & -217.0 (5087) & -78.3 (5005)& 8.6 (5097) & -45.9 (5963)\
& 0.0 & 35.9 (6718) & 80.2 (6968) & 36.9 (6096)& 0.5 (6295)& 29.9 (6761)\
\
&-2.0 & 0.08 (-0.27) & -0.01 (-1.1) & -0.53 (-0.1) & 0.55 (0.18) & -0.55 (1.6)\
(km/s) &-0.9 & 0.07 (-0.64) & 0.03 (-1.1)& -0.35 (1.4) & -0.59 (-0.7) & 0.02 (2.8)\
& 0.0 & -0.15 (-1.75)& 0.10 (-1.9)& -0.09 (4.1) & 0.66 (3.7)& 0.15 (1.5)\
\
&-2.0 & 0.0 (10) & 112.5 (122) & -57.14 (67) & -170.1 (577)& -327.8 (1190)\
(G) &-0.9 & 25.7 (131) & -14.1 (10) & 26.4 (180) & 454.0 (571)& 1.5 (1856)\
& 0.0 &-84.7 (255)& -23.7 (224) & -62.4 (234)& -203.8 (217)& -123.2 (1720)\
\
&-2.0 & -33.6 (70.2) & -30.2 (95.4) & -38.9 (63.1) & 22.3 (179.5)& -1.5 (162.9)\
(deg) &-0.9 & -2.7 (103.1) & -6.7 (99.8) & -5.0 (126.5) & 10.0 (128.9)& -4.0 (175.6)\
& 0.0 & 3.0 (107.7)& 4.5 (94.0) & 5.1 (71.2)& 1.6 (1.6)& -1.6 (158.2)\
\[tab:table1\]
Now we consider the temperature differences at the 5 spatial points in detail. The differences in temperature between the reference and test models at the five points in the three nodes are given in Table \[tab:table1\]. The temperatures in the reference model are indicated in parentheses. From Figure \[fig:rf\_int\], the $\delta E$ is positive in the middle of the granules (pos. 1 and 2), in intergranular lanes (pos. 3) and in the middle of a magnetic concentration (pos. 5). It is negative only in the intergranular lane next to a magnetic element (pos. 4), as discussed also in . The weakening of the NLTE line relative to LTE ($\delta E$ > 0) is due to the opacity deficit resulting from the UV overionization of iron atoms. When this weak NLTE line is fit by the inversion code using an LTE line, the temperature inferred is higher than that inferred from the inversion of an LTE line profile. However, in some cases, the NLTE line can also be stronger than its LTE version. This happens due to the departure of source function from the Planck function. When this effect dominates over the line weakening caused from the UV overionization [for a detailed discussion, see @2001ApJ...550..970S], we observe line strengthening due to NLTE effects. For example, in the top node, $\Delta T$ > 0 in pos. 1 but negative in pos. 2 (see *third column* in Figure \[fig:temp\] and also in Table \[tab:table1\]). In other words, although both pos. 1 and pos. 2 are in the granules, temperature from the inversion of NLTE lines is higher than from the inversion of LTE lines only in pos. 2 and exactly the opposite is observed at pos. 1. To explain this further, we plot the temperature profiles at these locations from the MHD cube as a function of $log(\tau)$ in Figure \[fig:maptau\_grad\]. The temperature stratification is quite similar at both pos.1 and pos.2 up to $log(\tau)=-1.0$. In the higher layers, the temperature gradient is smaller at pos.1 compared to pos.2. The smaller gradient leads to the strengthening of the line formed in NLTE compared to the LTE line. The residual intensity, $\delta I$, for both lines at pos.1, indicated in Figure 4, are positive and large, 30% for 6301.5Å line and 11% for 6302.5Å line. This indicates line strengthening. When the inversion code fits both these lines in LTE, the temperature from fitting the NLTE line will be lower than that from fitting the LTE line, resulting in a positive $\Delta T$ at pos.1. Such a line strengthening at several locations in the model atmosphere was observed also by .
In the intergranular lanes, predicted that applying LTE inversions would underestimate the temperature by 100 K - 200 K [also discussed by @2001ApJ...550..970S]. This conjecture is in agreement with the difference image at the top node (*third column*, Figure \[fig:temp\]). For example, at pos. 3, the NLTE intensity profiles are shallower than in LTE, $\delta I<0$ for both the lines (Figure \[fig:rf\_int\]), and $\Delta T$ is negative (Table \[tab:table1\]). In intergranular lanes next to magnetic elements, for example pos. 4, the temperature gradients are small, NLTE effects are not as strong, and $\delta E$ < 0, $\delta I$ >0. discuss that in the border of granules, the LTE temperature can be 200 K-400 K higher than in NLTE. This effect is seen in the top node, where along the boundaries of the granules, the difference image shows the temperature in reference model to be higher than in the test model (see Figure \[fig:temp\], *third column*). Since in both and [@2001ApJ...550..970S], these predictions were mainly based on the differences in the LTE and NLTE line core depths, they show a good correlation with our results in the top most node.
From Figure \[fig:temp\], we see that in the middle node, the temperature in the test model is higher in all the granules compared to the reference model. The histogram of $\Delta T$ has a mean value around 90K with a small secondary peak on the negative side which is due to the negative $\Delta T$ from the granules. From Figure \[fig:rf\_int\], the response at $log(\tau)=-0.9$ is strongest near the line wings. Differences between the NLTE and the LTE profiles in the line wings result in differences in the equivalent widths which manifest themselves as differences in temperature. In the intergranular lanes, the $\Delta T$ is much smaller as compared to that in the granules. In the intergranular lanes next to magnetic elements (e.g., pos. 4) or in the boundaries of the granules the difference is close to zero.
![*Top panel:* input LTE and NLTE intensity profiles at a pixel in a granule. *Bottom panel:* difference of the two intensity profiles.[]{data-label="fig:cont_diff"}](f7.pdf){width="49.00000%"}
### Bottom node {#sec:temp_node0}
At $log(\tau)=0.0$, the temperature in the reference model is slightly higher than in the test model (Figure \[fig:temp\], *bottom* row). In principle, the inversion code should return the same temperature in both cases since the continuum intensity is computed strictly in LTE, as in . However, in between and away from the two spectral lines, the difference between the NLTE and LTE intensity profiles is non-zero and positive. An example from a pixel in the granule is shown in Figure \[fig:cont\_diff\]. This indicates that within the spectral window used for synthesis, the profiles do not yet reach the real continuum. The inversion code is thus constraining the temperature at $log(\tau)=0$ by fitting the intensity in the far wings of the spectral lines and not from the continuum intensity. From Figure \[fig:temp\], this results in a difference of about $50-100$K in the granules with a relative difference $\delta T$ of less than 2% .
{width="38.00000%"} {width="59.50000%"}
Line-of-sight velocity {#sec:vel}
----------------------
The LOS velocity maps at the three nodes from the reference and test models are compared in Figure \[fig:vel\]. Again the two maps look similar to each other at all three nodes. reported only a small difference in the statistical distribution of the Doppler shifts between the LTE and NLTE profiles. From our analysis, the histograms of the differences between the velocities in the two models, in Figure \[fig:vel\], peak close to zero at all three nodes (*fourth column*). The mean of the differences is between 30 ms$^{-1}$-70 ms$^{-1}$ and the standard deviation is less than 0.5 kms$^{-1}$. From the difference maps (*third column*) regions of non-zero $\Delta v_{\rm LOS}$ are concentrated in the intergranular lanes, edges of granules and in the middle of magnetic elements. In the higher nodes differences of the order of 0.5 kms$^{-1}$ are seen in the middle of a few granules. The relative differences, ($\delta v_{\rm LOS}$), in the last column, are concentrated around zero in the bottom node and spread out to larger values, even up to and beyond 100%, in the higher nodes. $\delta v_{\rm LOS}$ is large for smaller velocities. To avoid small denominator values in the computation of $\delta v_{\rm LOS}$, we have neglected pixels with $v_{\rm LOS} < 10$ ms$^{-1}$. In the top node, a large number of points is clustered around $\pm 1$ km/s, arising from intergranular lanes and granular boundaries. In these regions, inversion of NLTE profiles results in errors of up to 50% as compared to the LTE case. When the velocities are higher, they are well constrained in the inversions and the relative differences are smaller. This is similar to the trend in the middle node.However, the absolute error is more or less independent from the actual value. This explains why the relative error in the LTE inversions of the NLTE data increases with decreasing velocity.
If we consider the individual intensity profiles in Figure \[fig:rf\_int\], at pos. 3, 4 and 5, we see a difference in the Doppler shifts, most evident at pos. 3, which is in an intergranular lane. From Table \[tab:table1\], the $\Delta v_{\rm LOS}$ is large in pos. 3, 4, and 5 compared to the granular points pos. 1 and 2. At pos. 4 $\Delta v_{\rm LOS}$ also changes sign from one node to the other. The reason for this is clear from the middle panel of Figure \[fig:maptau\_grad\] where the variation in $v_{\rm LOS}$ with $log(\tau)$ is plotted for all five points. At pos. 4, this variation is extreme with changes in direction of the flow along the vertical coordinate. Pos. 4 lies in the intergranular lane next to a magnetic element, , and tilting of such regions into the granule can lead to drastic variations in the flow velocity. Gradients in velocity are known to modify the NLTE line source function [@1967ApJ...147.1063K; @1971ApJ...165..543K; @1971ApJ...169..157C; @1974AuJPh..27..157V], and such scenarios are common in simulated atmospheres . Lower UV opacity due to overionization of iron atoms in combination with strong velocity gradients results in the LTE and NLTE profiles sampling different heights in the atmosphere leading to significant differences in velocity between the two cases.
{width="38.00000%"} {width="59.50000%"}
{width="38.00000%"} {width="59.50000%"}
{width="31.00000%"} {width="31.00000%"} {width="31.00000%"}
Magnetic field
--------------
### Strength {#sec:bfield}
investigated the NLTE effects in flux sheets and discussed their influence on the Stokes Zeeman profiles computed using 1D and 3D radiative transfer with a simple model and no inversions. The authors found that the field strengths recovered from the 1D-NLTE profiles of the Fe [i]{} 5250Å line are higher than those from LTE. In the present paper we use the two lines at 6300Å, which are formed slightly higher in the atmosphere than the 5250Å line, and thus are more sensitive to NLTE effects. Among the two lines at 6300Å, the 6301.5Å line is formed higher but has a smaller Landé g-factor (g=1.67), whereas the 6302.5Å line forms $\approx 100$km lower, but is more sensitive to the magnetic field (g=2.5). From the intensity profiles in Figure \[fig:rf\_int\] at pos. 5, which corresponds to a point in a magnetic element, the 6302.5Å line is completely split, with the LTE profiles being deeper and stronger ($\delta E$>0) than the NLTE profiles. we recover higher field strengths from the NLTE profiles as compared to the LTE ones, in two out of three nodes in the magnetic element (see Table \[tab:table1\]).
The errors in the magnetic field strength determination due to the neglect of NLTE effects, across different structures in an MHD cube is shown in Figure \[fig:bfield\]. The magnetic field strength maps at all three nodes are, overall, similar in both reference and test models. The histograms of the differences (*fourth column*) are highly non-Gaussian with narrow cores, broad wings with mean values less than 50 G. The standard deviation of the distribution is of the order of 100 G, with the main contribution coming from the wings of the distribution. From the relative difference, $\delta B$ plotted in the last column, the scatter is large for field strengths below 500G which correspond to the weak fields in and around the granules. Here, due to the fields being weak, the polarization signals are quite small, making it difficult to constrain the magnetic field. From the difference maps in *third column*, $\Delta B$ is non-zero in the intergranular lanes, edges of granules and in regions with a strong magnetic field. Also from Table \[tab:table1\], at pos. 3, 4 and 5, which are chosen from these three regions, not only is $\Delta B$ large, $\delta B$ is large as well (seen from the LTE values in parentheses). The MHD cube used here has large magnetic flux concentrations between granules (flux sheets lying in lanes) and at the intersections of multiple lanes (more tube-like structures). In these regions, neglecting the NLTE effects results in retrieval of stronger fields. Thus the points in the density plots are shifted slightly towards the negative $\delta B$, especially for fields greater than 500 G, particularly in the top node. Figure \[fig:maptau\_grad\] shows that, strong magnetic field gradients are present at the five chosen spatial points, which contributes to the strengthening of NLTE effects. The errors are smallest for the central node, which is also the one that is best constrained in the inversions. The qualitative picture is, however, reasonably consistent between nodes: fields are overestimated in lanes and in magnetic elements, but are underestimated at the edges of granules.
### Inclination
Neglecting the NLTE effects introduces errors in the determination of the magnetic field inclination as well. In Figure \[fig:gamma\] we estimate the magnitude of this error by comparing the inclination in the test model with the reference model. Large patches of non-zero $\Delta \gamma$ are seen in the top and the central nodes in the middle of granules (*third column*). In the last column, unlike in Figures \[fig:temp\], \[fig:vel\], and \[fig:bfield\], we show a scatter plot of the inclination in the test model vs inclination in the reference model, instead of the relative difference. In the top and the central nodes, the scatter increases as the field becomes more horizontal. These come from regions of low (< 200G) magnetic field strength (see Figure \[fig:bfield\]). Pixels with larger $\Delta \gamma$ are seen along the intergranular lanes as well where magnetic fields are strong and vertical but the main departures are seen in the granules. Of the three nodes the standard deviation of the histogram of $\Delta \gamma$ is the smallest at the central node as it is better constrained in the inversions. The mean of this histogram is less than $1\degree$ in all three nodes.
When we examine the Stokes $Q,U$ and $V$ profiles at the five selected spatial positions in Figure \[fig:quv\], we find big differences between the LTE and NLTE profiles. The magnitude of the error introduced by neglecting the NLTE effects at these five positions are given in Table \[tab:table1\]. The difference $\Delta \gamma$ is large at positions 1, 2, 3 and 4, especially in the top node. At pos. 1 and 2, the magnetic field from the reference model is weak and nearly horizontal. The $Q, U$ signals are small and it is difficult to constrain the magnetic field inclination. This also contributes to $\Delta \gamma$ being large. At pos. 4 and 5., the field (Table \[tab:table1\]) is strong and vertical in the reference model. Here the $V$ signals are strong and $\Delta \gamma$ is also smaller compared to the other positions. The gradients in magnetic field strength, inclination, LOS velocity which result in multi-lobed asymmetric $V$ profile, enhance the NLTE effects.
{width="25.80000%"} {width="22.00000%"} {width="22.00000%"} {width="22.00000%"}
{width="80.00000%"}
[cccccc]{} & $log(\tau)$ & < 100K & 100K–200K & 200K–400K & >400K\
\
&-2.0 & 76.4 & 18.7 & 4.8 & 0.1\
&-0.9 & 56.4 & 32.2 & 11.2 & 0.1\
& 0.0 & 97.8 & 2.2 & – & –\
\
&-2.0 & 76.8 & 18.0 & 5.0 & 0.04\
(rebinned)&-0.9 & 53.6 & 34.0 & 12.0 & 0.4\
& 0.0 & 96.7 & 3.1 & 0.1 & 0.04\
& & & & &\
& $log(\tau)$ & < 100ms$^{-1}$ & 100ms$^{-1}$–500ms$^{-1}$ & 500ms$^{-1}$–1kms$^{-1}$ & >1kms$^{-1}$\
\
&-2.0 & 42.0 & 53.1 & 3.8 & 1.0\
&-0.9 & 49.5 & 46.1 & 3.4 & 1.0\
& 0.0 & 41.0 & 49.6 & 6.5 & 2.8\
\
&-2.0 & 43.0 & 50.7 & 4.7 & 1.6\
(rebinned) &-0.9 & 46.1 & 46.0 & 6.1 & 1.8\
& 0.0 & 37.4 & 48.4 & 8.6 & 5.5\
& & & & &\
& $log(\tau)$ & < 100G & 100G–500G & 500G–1kG & >1kG\
\
&-2.0 & 76.2 & 23.2 & 0.4 & 0.05\
&-0.9 & 87.5 & 12.1 & 0.3 & 0.07\
& 0.0 & 74.1 & 23.6 & 1.8 & 0.5\
\
&-2.0 & 72.6 & 26.5 & 0.9 & 0.04\
(rebinned) &-0.9 & 81.5 & 17.8 & 0.6 & 0.11\
& 0.0 & 68.0 & 28.2 & 3.3 & 0.6\
& $log(\tau)$ & $< 10\degree$ & $10\degree-45\degree$ & $45\degree-90\degree$ & $>90\degree$\
\
&-2.0 & 50.6 & 37.9 & 10.3 & 1.2\
&-0.9 & 69.2 & 26.8 & 3.9 & 0.1\
& 0.0 & 60.3 & 33.0 & 5.9 & 0.8\
\
&-2.0 & 38.7 & 39.8 & 17.7 & 3.8\
(rebinned) &-0.9 & 59.5 & 34.7 & 5.3 & 0.5\
& 0.0 & 52.2 & 38.6 & 8.1 & 1.1\
\[tab:table2\]
Inversion of rebinned profiles {#sec:rebin}
==============================
The high resolution 3D MHD cube used for the analyses in the present paper was generated on a very fine grid with a spacing of 5.8km/pixel. This is much smaller than the spatial scales that can be resolved with current instrumentation. The *Hinode* satellite, which is extensively used for solar photospheric observations of the 6300Å lines, has a resolution of $0.16^{\prime\prime}$/pixel, which corresponds to about 116 km/pixel, 20 times larger than the resolution of our MHD cube. Many of the locations where the differences between the reference and test models were discussed in the previous sections lie within small-scale fine structures, resolved in the MHD cube, but not in any observations. It is therefore important to know if the errors in the test model discussed in Section \[sec:atm\_comp\] persist when the resolution is reduced, for example, to roughly match that of *Hinode*.
To test this, ideally, the Stokes profiles should be convolved with the point spread function (PSF) of the instrument, the profiles rebinned and noise added. Since our aim is to test if spatial averaging weakens the NLTE effects as discussed in @2001ApJ...550..970S, we only rebin the profiles to match the spatial scales observed by *Hinode* and do not convolve them with a PSF or add noise. The Stokes profiles computed in LTE and NLTE, respectively, are then inverted in LTE using the exact same setting as before. The temperature, LOS velocity, magnetic field strength, and inclination maps from the inversion of LTE and NLTE profiles are compared in a way similar to that used for the full resolution case, discussed in the previous sections. In Figure \[fig:rebin\], we show only the difference maps for the four atmospheric parameters. Comparing these maps with the original resolution maps in Figures \[fig:temp\], \[fig:vel\], \[fig:bfield\], and \[fig:gamma\] shows a striking similarity. For the temperature, as in the original resolution case, differences are seen mostly in the granules with $T_{\rm LTE} < T_{\rm NLTE}$ in many of them, this of course also depends on the node. A few exceptions are seen at $log(\tau)=-2.0$ in locations with weaker temperature gradients, as discussed in detail in Section \[sec:temp\_nodes12\].
For the LOS velocity and magnetic field strength, the locations with significant differences are more pronounced and appear to be clustered in the intergranular lanes and regions with strong magnetic fields. These are the regions with strong vertical gradients and the Stokes profiles, even after rebinning, contain these imprints, which are reflected in the corresponding inverted atmospheres. Differences in velocity are seen also in the middle of granules, like in Figure \[fig:vel\]. The magnitude of the errors is as large as in the original case.
In the maps showing errors in the determined magnetic field inclination, we see a large number of pixels with increased errors or differences, especially in the top node. The errors are smallest in the middle node just as in the full resolution case.
Statistics of the differences
=============================
To further quantify the NLTE effects, we divide the absolute difference $|\Delta T|, |\Delta v_{\rm LOS}|$, $|\Delta B|,$ and $|\Delta \gamma|$ into four different ranges. In Table \[tab:table2\], we indicate the percentage of pixels in each range and in Figure \[fig:diff\_range\_maps\], we map the four ranges to the regions where they originate. In temperature maps, the majority of the pixels have $|\Delta T|$ < 200K in the top two nodes. At the same time, pixels with $|\Delta T|$ between 200K-400K are non-negligible and are concentrated in the middle of granules. Nearly 5% and 11% of the pixels fall in this range in the top and middle node, respectively. NLTE effects from both line core and the wings contribute to the temperature determination in the middle node (see response functions in Figure \[fig:rf\_int\]) resulting in a higher percentage of pixels with large $|\Delta T|$. In the bottom node, $|\Delta T|$ < 100K with a few pixels here and there having larger differences. Ideally $|\Delta T|$ should be zero but, due to a poor separation of the continuum from the line wings, it is non-zero, as already discussed in Section \[sec:temp\_node0\].
Regarding the determined LOS velocity, the majority of the pixels have differences $ < 0.5$kms$^{-1}$. Less than $10\%$ have $|\Delta v_{\rm LOS}|$ between 0.5 kms$^{-1}$ and 1.0 kms$^{-1}$. A small fraction $\approx 1\%$ of the pixels show differences larger than 1kms$^{-1}$. As discussed in Section \[sec:vel\] and as can be seen from Figure \[fig:diff\_range\_maps\], pixels with large differences are found along the intergranular lanes. Nearly all pixels, at all three nodes, have differences less than 500G in the determined magnetic field strength. Only in the bottom node, $\approx 2\%$ of the pixels have $|\Delta B|>$1kG. These pixels are found in small groups, sparsely distributed across the whole region (see Figure \[fig:diff\_range\_maps\]).
Although the error in the magnetic field inclination is less than $10\degree$ in most of the pixels, a significant fraction of pixels ($30\%-40\%$) have errors between $10\degree-45\degree$, in all three nodes. In the top node, nearly $10\%$ of the pixels have errors greater than $45\degree$. These problematic pixels are found in the centers of the granules where the magnetic field strength is low. When the spatial resolution is reduced, the percentage of pixels with error $>45\degree$ increases to $~22\%$.
An important result from Table \[tab:table2\] is that the percentage of pixels with quantitative errors in $T, v_{\rm LOS},B$ and $\gamma$ remains the same, or even increases for $\gamma$, after spatially averaging the Stokes profiles. Although earlier papers such as @2001ApJ...550..970S predicted the spatial averaging to weaken the NLTE effects, our investigations show that they do survive. From Figure \[fig:diff\_range\_maps\], the errors in test model are not just found in a few isolated pixels but are seen as patches. Spatial averaging of the Stokes profiles does not cancel out these errors, at least when the horizontal transfer effects are not taken into account.
Summary and conclusions
=======================
The influence of NLTE conditions on the formation of photospheric iron lines has been well-known for several decades. However, due to the complexities involved in NLTE radiative transfer theory, these lines are often assumed to be formed in LTE, in particular when doing inversions of Stokes profiles. When inverting their Stokes profiles, not only are the effects of NLTE neglected, but also those of 3D radiative transfer for the same reasons. In continuation to the work of , we investigate how the assumption of LTE in the inversions of NLTE iron profiles introduces errors in the inferred atmospheres. In this first step, we assume that the Fe [i]{} 6301.5Å and 6302.5Å lines are formed in 1D NLTE. We compare the atmosphere determined from the inversion of line profiles computed in NLTE with the atmosphere from the inversion of lines formed in LTE. The latter is used as a reference model. The atmospheric model chosen for the inversion is a simple 3-node atmosphere which can be inadequate to reproduce extreme variations of physical quantities in the MHD cube. Such artifacts of the inversion can also contribute to the scatter in relative differences and lead to uncertainties. However, the strategy of using an inverted atmosphere as a reference model instead of the actual MHD cube is expected to mitigate this issue.
While most of the previous studies have focused on discussing how the NLTE conditions can alter the measured temperature, we examine the LOS velocity and magnetic field as well. In the inferred temperature, we find that neglecting NLTE can result in departures of up to 13% in both top and central inversion nodes. In the LOS velocity and magnetic field strength, we find relative difference as high as 50% or more in multiple nodes. A large fraction of the pixels have errors up to $45\degree$ in the determined magnetic field inclination. It is not possible to generalize whether neglecting the NLTE effects results in an overestimation or an underestimation of a particular parameter, since this depends on the vertical gradients in the region of interest and the amount of UV radiation from the layers below.
We chose five representative spatial positions to compare the LTE and NLTE Stokes profiles. We tried to correlate the errors in the inferred atmosphere at these locations with the differences observed in the Stokes profiles. In the intergranular lane and in the magnetic elements, the LOS velocity is off by$~500$m/s compared to the reference model. This can be a significant error when the LOS velocity itself is of the order of a few 100ms$^{-1}$ such as in the quiet Sun bright points [@Riethm_ller_2010; @Romano2012]. Errors in magnetic field strength and inclination are evident in both weak and strong magnetic field regions. Given that the 6300Å are often used in the construction of 1D model atmospheres based on LTE inversions , neglecting NLTE effects will introduce errors in these models.
A large scatter in the relative difference ($\delta B$) is observed at all three nodes in the test model in Figure \[fig:bfield\]. The small-scale weak magnetic field close to the photosphere is difficult to constrain with the simple inversion model as the polarization signals are small. While this contributes to the observed scatter in the test model, the role of NLTE effects cannot be ruled out. This is accompanied by errors of magnitude $10\degree-45\degree$ (Figure \[fig:gamma\]) in the determined magnetic field inclination. LTE inversions of observations at Fe [i]{} 6300Å lines are often used for measuring the strength, inclination of internetwork magnetic field and their variation with height . Understanding the errors due to the neglect of NLTE effects is important for the interpretation of such measurements.
While most the papers discussing NLTE effects in iron lines mainly focus on the quiet Sun, a detailed study in active regions is missing. The only known work is by @1997ASPC..118..207S, where the authors conclude that neglecting the NLTE effects can lead to an under-estimation of sunspot temperature. Observations at the Fe [i]{} 6300Å line pair are extensively used in understanding the physical properties of active regions such as sunspots and plages under the assumption that the lines are formed in LTE. How important are NLTE effects in active regions? How do they affect the temperature, velocity and magnetic field? These questions can be answered by carrying out an investigation similar to the one presented in the current paper using a sunspot simulation from @2012ApJ...750...62R. This will be the subject of an interesting follow-up work.
With the availability of a few NLTE inversion codes such as NICOLE, SNAPI , and STiC [@2018arXiv181008441D], it should now become possible to carry out the complex NLTE inversion of the photospheric iron lines at least using 1D radiative transfer. Finally, according to , the effects of horizontal radiative transfer can intensify or weaken the effects of NLTE. How the combination of horizontal radiative transfer with NLTE affects the inverted atmosphere will be investigated in a future publication.
We thank L. P. Chitta and I. Milić for comments on the manuscript. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 695075). This work has also been partially supported by the BK21 plus program through the National Research Foundation (NRF) funded by the Ministry of Education of Korea. This research has made use of NASA’s Astrophysics Data System.
[51]{} natexlab\#1[\#1]{}
, M., [Nordlund]{}, [Å]{}., [Trampedach]{}, R., & [Stein]{}, R. F. 2000, , 359, 743
, R. G. & [Lites]{}, B. W. 1972, , 176, 809
, L. R. & [Borrero]{}, J. M. 2002, , 391, 331
, L. R., [Ruiz Cobo]{}, B., & [Collados]{}, M. 1997, The Astrophysical Journal, 478, L45
, J. M. & [Bellot Rubio]{}, L. R. 2002, , 385, 1056
, J. M. & [Kobel]{}, P. 2012, , 547, A89
, J. M., [Lites]{}, B. W., [Lagg]{}, A., [Rezaei]{}, R., & [Rempel]{}, M. 2014, A&A, 572, A54
, J. H. M. J., [Rutten]{}, R. J., & [Shchukina]{}, N. G. 1992, , 265, 237
, D., [Lagg]{}, A., [Solanki]{}, S. K., & [van Noort]{}, M. 2015, , 576, A27
, C. J. & [Rees]{}, D. E. 1971, , 169, 157
, A. & [Ermolli]{}, I. 2017, , 841, 115
, S., [van Noort]{}, M., & [Rempel]{}, M. 2016, , 593, A93
, J., [Leenaarts]{}, J., [Danilovic]{}, S., & [Uitenbroek]{}, H. 2018, ArXiv e-prints \[\]
, M., [Ricort]{}, G., & [Aime]{}, C. 2013, , 554, A116
, C., [Solanki]{}, S. K., [Fligge]{}, M., & [Bruls]{}, J. H. M. J. 2000, , 358, 1109
, C., [Uitenbroek]{}, H., [Faurobert]{}, M., & [Aime]{}, C. 2010, , 514, A91
, R. & [Solanki]{}, S. K. 2012, , 547, A46
, R. & [Solanki]{}, S. K. 2013, A&A, 558, A20
, R. & [Solanki]{}, S. K. 2015, , 582, A101
, J., [Lagg]{}, A., [Hirzberger]{}, J., [Solanki]{}, S. K., & [Tiwari]{}, S. K. 2017, , 599, A35
, J., [Rezaei]{}, R., [Gonz[á]{}lez]{}, N. B., [Schlichenmaier]{}, R., & [Vomlel]{}, J. 2018, , 611, L4
, J. L. 1967, , 147, 1063
, J. L. 1971, , 165, 543
, A., [Solanki]{}, S. K., [van Noort]{}, M., & [Danilovic]{}, S. 2014, , 568, A60
, K., [Bergemann]{}, M., & [Asplund]{}, M. 2012, , 427, 50
, B. W. 2002, , 573, 431
, I. & [van Noort]{}, M. 2018, , 617, A24
, D., [Bellot Rubio]{}, L. R., [del Toro Iniesta]{}, J. C., [et al.]{} 2007, , 670, L61
, D. E. 1969, , 10, 268
, M. 2012, , 750, 62
, T. L., [Solanki]{}, S. K., & [Lagg]{}, A. 2008, , 678, L157
, T. L., [Solanki]{}, S. K., [van Noort]{}, M., & [Tiwari]{}, S. K. 2013, , 554, A53
Riethmüller, T. L., Solanki, S. K., Pillet, V. M., [et al.]{} 2010, The Astrophysical Journal, 723, L169
Romano, P., Berrilli, F., Criscuoli, S., [et al.]{} 2012, Solar Physics, 280, 407
, R. J. 1988, in Astrophysics and Space Science Library, Vol. 138, IAU Colloq. 94, ed. R. [Viotti]{}, A. [Vittone]{}, & M. [Friedjung]{}, 185–210
, R. J. & [Kostik]{}, R. I. 1982, , 115, 104
, A. I., [Schmutz]{}, W., [Schoell]{}, M., [Haberreiter]{}, M., & [Rozanov]{}, E. 2010, , 517, A48
, N. & [Trujillo Bueno]{}, J. 2001, , 550, 970
, N. G. & [Trujillo Bueno]{}, J. 1997, in Astronomical Society of the Pacific Conference Series, Vol. 118, 1st Advances in Solar Physics Euroconference. Advances in Physics of Sunspots, ed. B. [Schmieder]{}, J. C. [del Toro Iniesta]{}, & M. [Vazquez]{}, 207
, N. G., [Trujillo Bueno]{}, J., & [Asplund]{}, M. 2005, , 618, 939
, H. 2011, , 529, A37
, H., [de la Cruz Rodr[í]{}guez]{}, J., [Asensio Ramos]{}, A., [Trujillo Bueno]{}, J., & [Ruiz Cobo]{}, B. 2015, , 577, A7
, S. K. 1987, PhD thesis, ETH, Zürich
, S. K. & [Steenbock]{}, W. 1988, , 189, 243
, S. K., [van Noort]{}, M., [Lagg]{}, A., & [Solanki]{}, S. K. 2013, , 557, A25
, S. K., [van Noort]{}, M., [Solanki]{}, S. K., & [Lagg]{}, A. 2015, , 583, A119
, H. 2001, ApJ, 557, 389
, M. 2012, , 548, A5
, M., [Lagg]{}, A., [Tiwari]{}, S. K., & [Solanki]{}, S. K. 2013, , 557, A24
, I. M. & [Cannon]{}, C. J. 1974, Australian Journal of Physics, 27, 157
, A., [Shelyag]{}, S., [Sch[ü]{}ssler]{}, M., [et al.]{} 2005, , 429, 335
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This study assessed the effects of migration on the dynamics of a species population. It was considered that the species in its natural state and without the presence of migration exhibited Allee effect. This work also considered migration as a density-dependent function, which, from a maximum rate, decreases to a minimum of zero when the population reaches its carrying capacity.'
author:
- |
G.A. Ossandón & Ricardo Castro Santis\
\
\
title: 'Effects of density-dependent migration on a population subjected to Allee effect'
---
Introducción {#section}
============
A recurring problem in situations of endangered species preservation is that mere protection through hunting prohibition is not sufficient to ensure the recovery of the population. This fact occurs when population density decreases to a level in which reproduction and maintenance becomes very difficult, thus decreasing the reproductive rate to negative levels, which will finally end in the extinction of the species. This phenomenon is called *Efecto Allee* [@ALLE][@E0][@E1].In such a situation, it is necessary to intervene in order to promote the recovery of the population levels that will allow the species to survive in a natural way. One of these interventions is the introduction of new individuals in the habitat of the species. This strategy is often effective; however, it has high costs. This way, it is necessary to perform it in a way able to optimise the outcomes and resources[@BC; @CKI]. Within this context, a natural way is to generate a migratory flow dependent on the population level of the species, so that migration is greater than low population levels and ceases completely when the population reaches the carrying capacity of the environment. There are many studies on models with Alle effect, including deterministic and stochastic models, in ordinary differential equations and in partial derivatives. It is important to point out that the use of differential equations in partial derivatives is one of the possible approaches, very useful when the Allee effect is produced by the low rate of encounters between individuals. In Petroski&Lian[@petrovskii2003exactly] they propose a model with unidimensional spatial diffusion with Allee effect and dense dependent migration, whose migration function is a linear function. This considered migration is the product of environmental factors and biological mechanisms.\
The first thing to observe, and perhaps most importantly, is that it is not possible to maintain a migration above the carrying capacity of the environment, as the habitat does not support a larger population , will cause the surplus to be quickly eliminated. From this point of view it is natural that a migratory flow ceases when the population reaches the carrying capacity.\
These hypotheses can be translated mathematically in the following way: Let $f(x)$ be the immigration rate. We will consider $f$ as a decreasing function in the variable $x$ such that $f(K) = 0$, where $K$ is the capacity of environment. Therefore
$$\displaystyle\max_{x\in[0,K]}\{f(x)\}=f(0),$$
If we call $f(0)=\alpha$, the migration rate is a decreasing function joining the points $(0, \alpha)$ with the point $(K, 0)$\
In other words, if $K$ indicates the carrying capacity of the environment inhabited by the species, $x(t)$ will refer to the population of the species in time $t>0$, considering a positive and continuous function of migration rate $f_{\alpha}(x)$, strictly decreasing in the interval $[0,K]$, so that $f_{\alpha}(0)=\alpha$, and $f_{\alpha}(K)=0$. A natural choice in this context is to assume that the $f_{\alpha}$ function is linear, as shown in Figure 1.
{width="8cm"}\
Figure 1: Inmigration rate.
Therefore, the explicit equation of $f_{\alpha}$ is: $$\label{eq:inmigracion}
f_\alpha(x)=\alpha\left(1-\frac{x}{K}\right)\,\, , \,\, \forall \,\, x\in [0,K]$$
The parameter $\alpha$ represents the maximum value of the migration rate, which occurs in the total absence of population in the environment, which can be interpreted as the colonization rate of the habitat. The carrying capacity $K$ will be considered as a fixed parameter exclusively dependent on environmental conditions and that is not affected by the population of the species.
In the present study, we considered a logistic type of natural growth of the species subjected to Allee effect [@BC][@CKI][@EK] at a population level $m$, with intrinsic rate growth $r$, and carrying capacity $K$, i.e., the growth rate of the species in the absence of migration will be:
$$\label{eq:Allee}
g(x)=r\left(1-\frac{x}{K}\right)(x-m)x\quad\mbox{with}\quad 0<m<K\quad \mbox{and}\quad r>0$$
The Model
=========
Taken into consideration the hypotheses discused in the introduction, the model for population dynamics will have the following equation: $$\label{eq:model}
\left\{\begin{array}{rcl}
\dfrac{d}{dt}x(t)&=& g(x)+f_\alpha(x)\\
x(0)&=&x_0
\end{array}\right.$$ where the initial population $x_{0} \in [0,K]$, and the functions $g(x)$ and $f_{\alpha}(x)$ are those defined in equations and , respectively.\
Equation can be written as follows: $$\begin{array}{rcl}
x'(t)&=&g(x)+f_\alpha(x)\\
{}\\
{}&=&r\left(1-\dfrac{x}{K}\right)(x-m)x+\alpha\left(1-\dfrac{x}{K}\right)\\
{}\\
{}&=&\left(1-\dfrac{x}{K}\right)\left(r(x-m)x+\alpha\right)\\
{}\\
{}&=&\left(1-\dfrac{x}{K}\right)\left(rx^2-mrx+\alpha\right)
\end{array}$$ which indicates that $x=K$ remains an equilibrium point of the system. The other equilibrium points will depend on the quadratic factor $(rx^{2}-mrx+\alpha)$.\
A discriminant analysis of equation $rx^{2}-mrx+\alpha$ leads to: $$\label{eq:discriminante}
\Delta=m^2r^2-4r\alpha$$
This equation divides the analysis into three situations relating to the possible values that can achieve maximum migration rate $\alpha$ with respect to the intrinsic growth rate $r$ and the Allee level $m$. This fact is illustrated in the graph of the phase diagram, as shown in the figures 2, 3 nd 4:\
{width="3cm"} {width="3cm"} {width="3cm"}\
Figure 2Figure 3Figure 4\
$\Delta>0$$\Delta=0$$\Delta<0$
**NOTE:**In this model, $x=0$ is not the equilibrium point of the system. This way, it is possible that the initial population rate of the species is zero. In this situation, it can be affirmed that there is a colonization of the habitat by the migratory flow.
Case $\alpha>\dfrac{m^2r}{4}$
-----------------------------
If $\alpha>\frac{m^{2}r}{4}$, the discriminant factor defined in equation will be negative and the only equilibrium point of the system will be $x=K$, as illustrated in Figure 2. In this condition $x'(t)>0$ for all $x_0 \in [0,K]$, therefore, the population will grow from the initial value $x_{0}$ to the carrying capacity $K$.\
In this situation, for any initial value, including zero initial population, the species will asymptotically tend to the carrying capacity of the environment.\
Figure 5 shows the temporal evolution of the population of the species, which starts with a $x_{0}$ value.\
{width="6cm"}\
Figure 5: Trajectory case $\Delta<0$
Case $\alpha=\dfrac{m^2r}{4}$
-----------------------------
There is a new equilibrium point as a solution to equation $rx^{2}-mrx+\alpha=0$, which, with the condition $\alpha=\frac{m^2r}{4}$, becomes $\,r\left(x-\frac{m}{2}\right)^2=0$, thus obtaining the equilibrium point $x=\frac{m}{2}$, which corresponds to half the value of the natural Allee level of the species. The equilibrium point $x=\frac{m}{2}$ is unstable of attractor-repulsor design.
As shown in Figure 3, for all $x_0\in[0,K]$; the trajectories will be non-decreasing for any initial value $0\le x_0<K$.\
If $0\le x_0<\frac{m}{2}$, the trajectories will grow asymptotically at $\frac{m}{2}$, whereas, if $\frac{m}{2}<x_0<K$, the trajectories will grow asymptotically with respect to carrying capacity $K$.\
In ecological terms, the species will finally achieve its recovery. Even though the initial population is less than $\frac{m}{2}$, a positive disturbance may eventually raise that level. This way, final recovery will start to achieve carrying capacity, as shown in Figure 6.
{width="6cm"}\
Figure 6: : Trajectory case $\Delta=0$
Case $\alpha<\dfrac{m^2r}{4}$
-----------------------------
This case exhibits the greatest dynamic richness. There are two equilibrium points relating to value $\frac{m}{2}$. The discriminant $\Delta=m^2r^2-4r\alpha>0$ and the equation solutions $rx^2-mrx+\alpha=0$ will be
$$\label{eq:equi-point}
x_1=\frac{m}{2}-\frac{\sqrt{\Delta}}{2r}\qquad\mbox{y}\qquad x_2=\frac{m}{2}+\frac{\sqrt{\Delta}}{2r}$$
as $\displaystyle\sqrt{m^2r^2-4\alpha}<mr$, the following inequalities are obtained
$$0<x_1<\dfrac{m}{2}<x_2<m$$
Figure 4 illustrates the situation described, from which it can be easily observed that $x'>0$if$0\le x<x_1$or$x_2<x<K$ and, conversely, $x'<0$if$x_1<x<x_2$. Therefore, the trajectories will have the following behaviour depending on the initial population $x_0$ value.
- If $0\le x_0<x_1$, the population will asymptotically grow to level $x_1$.
- If $x_1<x_0<x_2$, the population will asymptotically decrease to level $x_1$.
- If $x_2<x_0<K$, the population will asymptotically grow to carrying capacity $K$.
Figure 7 illustrates the situation described.
{width="6cm"}\
Figure 7: : Trajectory case $\Delta>0$
Some conclusions
================
- The Allee level, which, in the absence of migration is related to a population abundance of $m$, decreases to the value $x_2 \in\left]\frac{m}{2},m\right[$.
- The presence of migration prevents the population from becoming extinct, with an attractor equilibrium point of $x_1\in\left]0,\frac{m}{2}\right[$.
- This equilibrium is exclusively maintained by the migratory flow and can be interpreted as precarious subsistence equilibrium, because, since the migrant population is far below the natural Allee effect level, it fails to survive and is continuously replaced by new migrants, which will never be able to recover the population with respect to the carrying capacity of the environment.
- From the perspective of the problem of induced immigration as a technique for recovery of species, it can be observed that, if the initial population is very small $\frac{m}{2}$, the only possibility for recovering will be a strong migratory flow, i.e., a rate given by $\alpha>\frac{m^2r}{4}$.
- If the initial population is greater than$\frac{m}{2}$but less than $m$ , a migration rate will be necessary to ensure that$x_0>x_2$and the following result will be obtained from equation : $\alpha> rx_{0}(m-x_{0})$
[99]{} B. Dennis. Allee effects: population growth, critical density, and chance of extinction. Nat. Resour. Model., 3 (1989), 381-538.
E. Liz, A. Ruiz-Herrera, Delayed population models with Allee effects and exploitation, Math. Biosci. Eng. 12 (2015), in press.
P. Amarasekari. Allee effect in metapopuilation dinamycs. A.M. Nat., 152 (1998) , 298-302
Brauer, F. and Castillo Chávez, C.\[2001\], Mathematical Models in Population Biology and epidemology. Texts in Applied mathematics. Springer-Verlag.
Clark, C.W. \[1990\], Mathematical Bioeconomic: The optimal management of renovavle resources, second edition. John Wiley and Sons
Edelstein-Kesner, L. \[1988\], Mathematical Models in Biology, Birkhauser mathematics series, Mc-Graw Hill, Inc.
Petrovskii, Sergei and Li, Bai-Lian. An exactly solvable model of population dynamics with density-dependent migrations and the Allee effect. Mathematical Biosciences. (2003) Vol 1. 79 - 91
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Institute for Nuclear Studies, Department of Physics, George Washington University, Washington DC 20052, USA; E-mail:
title: Assessing Theory Uncertainties in EFT Power Countings from Residual Cutoff Dependence
---
Motivation: Serious Theorists Have Error Bars
=============================================
That our understanding of natural phenomena is based on concrete, falsifiable predictions is deeply ingrained in the scientific method. It is insufficient to compare numbers; one also must judge their reliability. And since we do not trust experiments without error bars, why should it be acceptable for a theorist to not assess uncertainties in a calculation independently of the data to be explained? Simply stating that this is “difficult” is certainly no sufficient excuse, especially after the recent surge of articles on theory errors; see e.g. [@editorial; @JPhysG]. The prospect of a reproducible, objective, quantitative estimate of theoretical uncertainties lies thus at the heart not only of any Effective Field Theory (EFT). But EFTs claim to possess well-defined schemes to find just that. It is therefore befitting to explore how the validity of such prescriptions can be gauged.
More than a decade ago, Lepage used a highly influential lecture to quantify convergence to data [@Lepage:1997cs]. In contradistinction, the test presented here hopes to quantify the internal consistency of an EFT and takes *minimal resort* to experimental information. Ideally, theorists would perform “double-blind” calculations in which theory errors are assessed under the pretence that no or only very limited data is available. Such “post-dictions” are of course predictions when information is indeed experimentally unknown or hard to access, or when data consistency must be checked.
EFT uses scale-separation to expand interactions and observables in a dimension-less quantity $$Q=\frac{\text{typical low momenta } k,{p_\text{typ}}}{\text{breakdown scale } {\overline{\Lambda}_\text{EFT}}}<1\;\;.$$ The numerator contains the relative momentum $k$ between scattering particles and other intrinsic low scales, here summarily denoted by ${p_\text{typ}}$. At the breakdown scale ${\overline{\Lambda}_\text{EFT}}$, new dynamical degrees of freedom enter which are not explicitly accounted but whose effects at these short distances are simplified into Low-Energy Coefficients (LECs). In [$\chi$EFT]{}, ${p_\text{typ}}$ includes the pion mass and the inverse scattering lengths of the $\mathrm{NN}$ system. ${\overline{\Lambda}_{\chi}}\approx[700\dots1000]\;{\ensuremath{\mathrm{MeV}}}$ is consistent with the masses of the $\omega$ and $\rho$ as the next-lightest exchange mesons, and with the chiral symmetry breaking scale. The power counting (PC) is then determined by Naïve Dimensional Analysis [@NDA; @NDA2]. When all interactions are perturbative, as in the mesonic and one-baryon sectors, this amounts to not much more than counting powers of $k$ and ${p_\text{typ}}$.
The situation is more complicated when some interactions must be treated non-perturbatively at leading order because of shallow real or virtual bound states with scales ${p_\text{typ}}\ll{\overline{\Lambda}_\text{EFT}}$ in the EFT’s range of validity. In $\mathrm{NN}$ scattering, all terms in the LO Lippmann-Schwinger equation, including the potential, must be of the same order when all nucleons are close to their non-relativistic mass-shell. If not, one term could be treated as perturbation of the others and there would be no shallow bound-state. Picking the nucleon-pole in the energy-integration, $q_0\sim\frac{q^2}{2M}$, leads to the consistency condition that both amplitude $T_\mathrm{NN}$ (ellipse) and potential $V_\mathrm{NN}$ (rectangle) must be of order $Q^{-1}$: $$\label{eq:consistency}
\begin{split}
\includegraphics[clip=]{consistentPC.pdf}
\end{split}$$ In particular, the one-pion exchange appears to scale as $(\vec{\sigma}_1\cdot\vec{q})(\vec{\sigma}_2\cdot\vec{q})/(\vec{q}^2+{\ensuremath{m_\pi}}^2)\sim
Q^0$ when one counts only explicit low-momentum scales, but must be of order $Q^{-1}$ if its iteration is to be mandated. In a straightforward extension, amplitude and interaction between $n$ nucleons scale as $$\label{eq:consistency-n}
T_{n\mathrm{N}}\sim V_{n\mathrm{N}}\sim Q^{1-n}$$ if it is nonperturbative at LO. Since both the interactions and the LECs themselves carry inverse powers of $Q$, finding their importance by counting momenta is insufficient. This result only assumes the existence of unnaturally small scales, irrespective of the form of the interaction. It does not reveal *which* terms constitute the LO potential; only how those terms must be power-counted.
This behaviour has been long-recognised in pionless EFT ([EFT($\slashed{\pi}$)]{}) and its variants (like Halo-EFT and EFT of point-like interactions), where the scaling of operators and the $\beta$ functions of couplings between up to 3 nucleons are well-established [@Bedaque:2002mn; @Platter:2009gz]. For example, analytic results in well-controlled limits show one momentum-independent $3N$ operator at LO. Likewise, non-relativistic QED and QCD count the Coulomb potential as $Q^{-1}$ to allow its resummation.
The situation in [$\chi$EFT]{}for two and more nucleons is less obvious. Weinberg suggested to still count LECs as $Q^0$, and to apply the perturbative counting of momenta not to amplitudes but to the few-nucleon potential, which is then iterated to produce shallow bound states. How this translates into a PC of observables is under dispute. Further disagreement persists about the interpretation of approximate solutions (large off-shell momenta, semi-classical limit, etc.), and about unrelated numerical problems (deeply bound states etc.). In addition, a cutoff $\Lambda$ becomes numerically necessary. It is conceptually quite different from the breakdown scale ${\overline{\Lambda}_\text{EFT}}$, albeit the two symbols are similar. It cannot be much smaller than the breakdown scale in order not to “cut out” physical, low-resolution momenta in loops. But even how far $\Lambda$ should be varied is under dispute: Is any value $\Lambda\gtrsim{\overline{\Lambda}_\text{EFT}}$ legitimate, including $\Lambda\to\infty$; or should the range be constrained to $\Lambda\approx{\overline{\Lambda}_\text{EFT}}$?
[|l||p[0.18]{}|p[0.187]{}|p[0.215]{}|p[0.205]{}|]{}
------------------------------------------------------------------------
**order**&**Weinberg** (modified) [@Weinberg]&**Birse** 2005 [@Birse:2005um; @Birse:2009my]&**Pavon et al.** 2006 [@PavonValderrama:2005wv; @PavonValderrama:2005uj; @Valderrama:2009ei; @Valderrama:2011mv]&**Long/Yang** 2012 [@Long:2011xw; @Long:2012ve]
\
------------------------------------------------------------------------
[$Q^{-1}$]{}& LO of [${}^{1}\mathrm{S}_{0}$]{}, [${}^{3}\mathrm{S}_{1}$]{}, OPE& LO of [${}^{1}\mathrm{S}_{0}$]{}, [${}^{3}\mathrm{S}_{1}$]{}, OPE, [${}^{3}\mathrm{D}_{1}$]{}, [${}^{3}\mathrm{SD}_{1}$]{} & LO of [${}^{1}\mathrm{S}_{0}$]{}, [${}^{3}\mathrm{S}_{1}$]{}, OPE, [${}^{3}\mathrm{P}_{0,2}$]{}, [${}^{3}\mathrm{D}_{2}$]{}&LO of [${}^{1}\mathrm{S}_{0}$]{}, [${}^{3}\mathrm{S}_{1}$]{}, OPE, [${}^{3}\mathrm{P}_{0,2}$]{}
\
------------------------------------------------------------------------
$Q^{-{\frac{1}{2}}}$&none& LO of [${}^{3}\mathrm{P}_{0,1,2}$]{}, [${}^{3}\mathrm{PF}_{2}$]{}, [${}^{3}\mathrm{F}_{2}$]{}, [${}^{3}\mathrm{D}_{2}$]{} & LO of [${}^{3}\mathrm{SD}_{1}$]{}, [${}^{3}\mathrm{D}_{1}$]{}, [${}^{3}\mathrm{PF}_{2}$]{}, [${}^{3}\mathrm{F}_{2}$]{}&none
\
------------------------------------------------------------------------
$Q^{0}$&none&NLO of [${}^{1}\mathrm{S}_{0}$]{} &NLO of [${}^{1}\mathrm{S}_{0}$]{}& NLO of [${}^{1}\mathrm{S}_{0}$]{}
\
------------------------------------------------------------------------
$Q^{{\frac{1}{2}}}$&none & NLO of [${}^{3}\mathrm{S}_{1}$]{}, [${}^{3}\mathrm{D}_{1}$]{}, [${}^{3}\mathrm{SD}_{1}$]{} &none&none
\
------------------------------------------------------------------------
&LO of [${}^{3}\mathrm{SD}_{1}$]{},[${}^{1}\mathrm{P}_{1}$]{}, [${}^{3}\mathrm{P}_{0,1,2}$]{}, TPE; NLO of [${}^{1}\mathrm{S}_{0}$]{}, [${}^{3}\mathrm{S}_{1}$]{}&&&LO of [${}^{3}\mathrm{SD}_{1}$]{},[${}^{1}\mathrm{P}_{1}$]{}, [${}^{3}\mathrm{P}_{1}$]{}, [${}^{3}\mathrm{PF}_{2}$]{}, TPE; NLO of [${}^{3}\mathrm{S}_{1}$]{}, [${}^{3}\mathrm{P}_{0}$]{}, [${}^{3}\mathrm{P}_{2}$]{}; [N${}^{2}$LO]{} of [${}^{1}\mathrm{S}_{0}$]{}
\
------------------------------------------------------------------------
$Q^\frac{3}{2}$&none&NLO of [${}^{3}\mathrm{D}_{2}$]{}, [${}^{3}\mathrm{P}_{0,1,2}$]{}, [${}^{3}\mathrm{PF}_{2}$]{}, [${}^{3}\mathrm{F}_{2}$]{} &none&none
\
------------------------------------------------------------------------
&& LO of TPE, [${}^{1}\mathrm{P}_{1}$]{}; NLO of OPE; [N${}^{2}$LO]{} of [${}^{1}\mathrm{S}_{0}$]{}& LO of TPE, [${}^{1,3}\mathrm{P}_{1}$]{}; NLO of [${}^{3}\mathrm{S}_{1}$]{}, [${}^{3}\mathrm{D}_{1,2}$]{}, [${}^{3}\mathrm{SD}_{1}$]{}, [${}^{3}\mathrm{P}_{0,2}$]{}, [${}^{3}\mathrm{PF}_{2}$]{}; [N${}^{2}$LO]{} of [${}^{1}\mathrm{S}_{0}$]{}&
\
------------------------------------------------------------------------
[\# at $Q^{-1}$]{}&$2$&$4$&$5$&$4$
\
------------------------------------------------------------------------
[\# at $Q^{0}$]{}&$+0$&$+7$&$+5$&$+1$
\
------------------------------------------------------------------------
[\# at $Q^{1}$]{}&$+7$&$+3$&0&$+8$
\
------------------------------------------------------------------------
total at $Q^{1}$&$9$&$14$&$10$&$13$
\
It is thus no surprise that four active PC proposals emerged in [$\chi$EFT]{}, all with the same degrees of freedoms: nucleons and pions only [@Weinberg; @Birse:2005um; @Birse:2009my; @PavonValderrama:2005wv; @PavonValderrama:2005uj; @Valderrama:2009ei; @Valderrama:2011mv; @Long:2011xw; @Long:2012ve]. Table \[tab:pc\] lists their predictions for the order at which a LEC enters in the lower $\mathrm{NN}$ partial waves. Each finds a different number of LECs at any given order – and each claims consistency. Not all can be right, though. Articles, panels and sessions at Chiral Dynamics and other conferences as well as dedicated workshops led to little consensus (see e.g. Daniel Phillips’ even-handed account at the last Chiral Dynamics [@Phillips:2013fia]); some additional notable contributions include Refs. .
This is not just stamp-collecting or a philosophical question which potentially exposes the soft underbelly of [$\chi$EFT]{}and the credibility of its error assessments, but which is “otherwise” of little practical consequence. A central EFT promise is that it encodes the unresolved short-distance information at a given accuracy into not just some, but the *smallest-possible* number of independent LECs. For example, the PC proposals of $\mathrm{NN}$ [$\chi$EFT]{}differ most for attractive triplet waves: the [${}^{3}\mathrm{P}_{2}$]{}-[${}^{3}\mathrm{F}_{2}$]{} system at order $Q^0$ has no LEC parameter [@Weinberg] – or $3$ of similar size [@Birse:2005um; @Birse:2009my] – or $3$, but with different weights [@PavonValderrama:2005wv; @PavonValderrama:2005uj; @Valderrama:2009ei; @Valderrama:2011mv] – or $1$ [@Long:2011xw; @Long:2012ve]. To bring it to a boil: If all proposals are renormalised and fit $\mathrm{NN}$ data with the same $\chi^2$, the one with the least number of parameters wins.
For the sake of this note, I am agnostic about this dispute. Rather, I propose to test if a predicted convergence pattern is reflected in the answers, i.e. if a proposed power counting is consistent. On a historical note, the origin of these remarks goes back to publications in 2003 and 2005 [@Bedaque:2002yg; @improve3body], and to lectures at the 2008 US National Nuclear Physics Summer School [@NNPSS]. When the issue was revisited at two more recent workshops [@Saclay; @Benasque], its conclusions were generally perceived as not immediately straightforward or widely known. Input on some aspects was also provided for two recent publications [@Furnstahl:2014xsa; @Epelbaum:2014efa]. It seems therefore fitting to present the test in the form of an expanded Technical Note. This article summarises an upcoming publication [@hgrie].
The Test: Turning Cutoff Dependence into an Advantage
=====================================================
Assume we calculated an observable ${\mathcal{O}}$ up to and including order $Q^n$ in an EFT: $$\label{eq:observable}
{\mathcal{O}}(k,{p_\text{typ}};\Lambda;{\overline{\Lambda}_\text{EFT}})=
\sum\limits_i^n\left(\frac{k,{p_\text{typ}}}{{\overline{\Lambda}_\text{EFT}}}\right)^i
{\mathcal{O}}_i(k,{p_\text{typ}};{\overline{\Lambda}_\text{EFT}})\; +\;{\mathcal{C}}_n(\Lambda;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})\left(\frac{k,{p_\text{typ}}}{{\overline{\Lambda}_\text{EFT}}}\right)^{n+1}$$ \[Non-integer $n$ and non-integer steps from one order to the next will be discussed in Sect. \[sec:principle\].\] The notation indicates that numerators may depend on both $k$ and ${p_\text{typ}}$. In a properly renormalised result, effects attributed to a regulator $\Lambda$ appear only at orders which are higher than the last order $n$ which is known in full. The residual ${\mathcal{C}}_n(\Lambda;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})$ may still depend on ${\overline{\Lambda}_\text{EFT}}$, $k$ and ${p_\text{typ}}$, but it should be of natural size for all $k,{p_\text{typ}}<{\overline{\Lambda}_\text{EFT}}$, so that its contribution is parametrically suppressed by $\left(\frac{k,{p_\text{typ}}}{{\overline{\Lambda}_\text{EFT}}}\right)^{n+1}$ relative to the known terms of the series. If not, cutoff variations produce corrections which are comparable in size to the regulator-independent terms ${\mathcal{O}}_i(k,{p_\text{typ}};{\overline{\Lambda}_\text{EFT}})$ and contradict the EFT assumption that higher-order corrections are parametrically small.
The relative difference of ${\mathcal{O}}(k,{p_\text{typ}};\Lambda)$ at any two cutoffs is then: $$\label{eq:master}
\frac{{\mathcal{O}}_n(k,{p_\text{typ}};\Lambda_1)-{\mathcal{O}}_n(k,{p_\text{typ}};\Lambda_2)}
{{\mathcal{O}}_n(k,{p_\text{typ}};\Lambda_1)}=\left(\frac{k,{p_\text{typ}}}{{\overline{\Lambda}_\text{EFT}}}\right)^{n+1}
\times\;\frac{{\mathcal{C}}_n(\Lambda_1;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})-
{\mathcal{C}}_n(\Lambda_2;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})}
{{\mathcal{C}}_n(\Lambda_1;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})}\;\;.$$ One can now vary $k$ or ${p_\text{typ}}$ to read off both the order $n$ to which the calculation is complete and the breakdown scale ${\overline{\Lambda}_\text{EFT}}$ – if the cutoff behaviour cannot be eliminated in its entirety, i.e. ${\mathcal{C}}_n(\Lambda_1)\ne{\mathcal{C}}_n(\Lambda_2)$ for at least some cutoff pairs, and if the residuals ${\mathcal{C}}_n$ vary more slowly with $k$ and ${p_\text{typ}}$ than with $\Lambda$. The results of such a fit may certainly be inconclusive; see extended remarks in Sect. \[sec:notesofnote\]. But if higher orders are *not* parametrically suppressed and the slope comes out *smaller* than the PC prediction $n+1$, then the EFT is necessarily not properly renormalised. As will be discussed in Sect. \[sec:principle\], a slope $\ge n+1$ does not suffice to demonstrate consistency or establish failure.
Equation is formulated in terms of renormalised quantities only and therefore holds for any regulator, but it is most useful for cutoffs: Answers in nonperturbative EFTs are usually found only numerically and for a *cutoff regulator*, i.e. for a regulator which explicitly suppresses high momenta $q\gtrsim\Lambda$ in loops. It is this case which we use to our advantage from now on.
Cutoffs are of course only sensible if all loop momenta are sampled which lie in the domain of validity of the EFT, i.e. if $\Lambda\gtrsim{\overline{\Lambda}_\text{EFT}}$. Only then can the coefficients ${\mathcal{C}}_n$ be expected to be of natural size relative to ${\overline{\Lambda}_\text{EFT}}$ (with the caveats mentioned around eq. ). Except for this, no particular assumption is necessary as to the size of $\Lambda$ relative to ${\overline{\Lambda}_\text{EFT}}$ in eq. . In dimensional regularisation and other analytic schemes, on the other hand, renormalisation can be performed exactly and no cutoff $\Lambda$ or residual regulator scale appears in observables at all. Equation is then an exact zero, with no information about $n$ and ${\overline{\Lambda}_\text{EFT}}$. But doubts about proper renormalisation of a calculation which is analytic at each step do not arise, so the test is moot anyway.
An Application: Confirming the Hierarchy of $3N$ Interactions in [EFT($\slashed{\pi}$)]{} {#sec:application}
=========================================================================================
Before continuing the discussion of the parameters of the test in Sect. \[sec:notesofnote\], consider the first application (to my knowledge) of this test: the [${}^{2}\mathrm{S}_{\frac{1}{2}}$]{} $Nd$ wave in [EFT($\slashed{\pi}$)]{}. It is well-known that its $3N$ interaction without derivatives does not follow simplistic PC rules (“just count momenta”) which predict $H_0$ at [N${}^{2}$LO]{} or ${\mathcal{O}}(Q^0)$ [@Bedaque:2002mn; @Platter:2009gz]. Instead, it is needed at LO to stabilise the system (Thomas-collapse, Efimov effect); its scaling, $H_0\sim Q^{-2}$, follows from eq. for $n=3$. If the first momentum-dependent $3N$ interaction $k^2H_2$ follows the simplistic argument and scales as $Q^2$, then new LECs need to be determined from $3N$ data only at [N${}^{4}$LO]{}. Therefore, one could find $2N$ interaction strengths from few-N data with only one new $3N$ datum up to an accuracy of better than $1$% at low momenta. This is crucial for example for hadronic flavour-conserving parity violation since it considerably extends the number of targets and observables [@Griesshammer:2010nd].
Based on the asymptotic off-shell amplitudes, Refs. [@Bedaque:2002yg; @improve3body] proposed that $H_2$ is only suppressed by $Q^2$ relative to LO, i.e. that calculations at [N${}^{2}$LO]{} or on the $10$%-level do already need one additional $3N$ datum as input. In Ref. [@effrange], this was confirmed and extended to a general scheme to find the order at which any given $3N$ interaction starts contributing. The argument analyses perturbations to the asymptotic form of the LO integral equation. It is not immediately transparent, as witnessed by a subsequent claim that a $k$-dependent $3N$ interaction enters not earlier than [N${}^{3}$LO]{} [@Platter:2006ev]. Upon closer inspection, it was later refuted [@Ji:2012nj].
Refs. [@Bedaque:2002yg; @improve3body] also supplied numerical evidence from solutions of the Faddeev equations in momentum space with a step-function cutoff: a double-logarithmic plot of eq. for the inverse $K$ matrix, ${\mathcal{O}}=k\cot\delta$ at $\Lambda_1=900\;{\ensuremath{\mathrm{MeV}}}$ and $\Lambda_2=200\;{\ensuremath{\mathrm{MeV}}}$, both well above the breakdown scale ${\overline{\Lambda}_{\slashed{\pi}}}\approx{\ensuremath{m_\pi}}$ of [EFT($\slashed{\pi}$)]{}. A slight variant is reproduced here as Fig. \[fig:nonLepage\]. The cutoff dependence decreases order-by-order as expected when the theory is perturbatively renormalised in the EFT sense. There is no decrease from NLO to [N${}^{2}$LO]{} when $H_2\equiv0$. That by itself could be accidental – after all, would one not expect better convergence with one more parameter to tune?
More informative is a look at the slopes. Lines at different orders are near-parallel for small $k$ because there are additional natural low-energy scales ${p_\text{typ}}$, namely the binding momenta of the deuteron ($\gamma_t\approx45\;{\ensuremath{\mathrm{MeV}}}$) and of the virtual singlet-S state ($\gamma_s\approx8\;{\ensuremath{\mathrm{MeV}}}$). For $k\lesssim\gamma_{t,s}$, eq. is not very sensitive to $k$, so all slopes should be small and near-identical. However, when $k\gg\gamma_{t,s}$ (but of course still $k\ll{\overline{\Lambda}_{\slashed{\pi}}}$, so that the EFT converges), they converge towards one region.
Indeed, the fits of $n$ to the nearly straight lines in the momentum range between $70$ and $100$ to $130\;{\ensuremath{\mathrm{MeV}}}<{\overline{\Lambda}_{\slashed{\pi}}}$ compare well to the PC prediction when $H_2$ is added at [N${}^{2}$LO]{} [@improve3body]: $$\label{eq:3Ntab}
\begin{tabular}{c@{{\hspace{0.5ex}}}|@{{\hspace{0.5ex}}{\hspace{0.5ex}}}c@{{\hspace{0.5ex}}{\hspace{0.5ex}}}c@{{\hspace{0.5ex}}{\hspace{0.5ex}}}c@{{\hspace{0.5ex}}{\hspace{0.5ex}}}|@{{\hspace{0.5ex}}{\hspace{0.5ex}}}c}
& LO&NLO&{N\ensuremath{{}^{2}}LO\xspace}&{N\ensuremath{{}^{2}}LO\xspace} {without $H_2$}\\ \hline
{$n+1$} fitted &${\sim 1.9}$&{$2.9$}&{$4.8$} [\emph{sic!}]&${3.1}$\\ {$n+1$} predicted &${2}$&{$3$}&{$4$}¬ renormalised
\end{tabular}$$ Without $H_2$ at [N${}^{2}$LO]{}, the slope does not improve from NLO. This is a clear signal that the PC is inconsistent without a momentum-dependent $3N$ interaction at [N${}^{2}$LO]{}: Its assumptions do not bear out in the functional behaviour of this observable on $k$. On the other hand, when $H_2$ is included, the slope is markedly steeper than at NLO. The general agreement between predicted and fitted slope is astounding, and actually quite stable against variation of the fit range or of the two cutoffs $\Lambda_1$ and $\Lambda_2$. Only the LO numbers are somewhat sensitive, and only to the upper limit [@improve3body].
It is somewhat surprising that the slope increases by two units from NLO to [N${}^{2}$LO]{} when one includes $H_2$. One would have expected the change from each order to the next to be by only one unit. This may stem from the “partially resummed formalism” used at that time. Since that resums some higher-order contributions, it may be worth revisiting this issue with J. Vanasse’s method to determine higher-order corrections in “strict perturbation” [@Vanasse:2013sda]. But we will see in the notes on “Assumptions of the Expansion” in Sect. \[sec:notesofnote\] that a fitted slope which is larger than predicted does not invalidate the power counting – the converse does.
Finally, one reads off a rough value of ${\overline{\Lambda}_{\slashed{\pi}}}\approx[120\dots150]\;{\ensuremath{\mathrm{MeV}}}$ as the region where the fitted lines coalesce. This is not in disagreement with the breakdown scale expected of [EFT($\slashed{\pi}$)]{}.
Notes of Note {#sec:notesofnote}
=============
With this example in mind, let us consider assumptions, strengths, extensions, features, caveats and limitations of such an analysis to assess the consistency of a PC proposal.
Matters of Principle {#sec:principle}
--------------------
[**Renormalisation Group Evolution**]{} Multiply eq. by $(\Lambda_1-\Lambda_2)/\Lambda_1$ and take $\Lambda_2\to\Lambda_1$: $$\label{eq:rge} \frac{\Lambda}{{\mathcal{O}}}\;\frac{{\mathrm{d}}{\mathcal{O}}}{{\mathrm{d}}\Lambda}=
\left(\frac{k,{p_\text{typ}}}{{\overline{\Lambda}_\text{EFT}}}\right)^{n+1}
\;\frac{{\mathrm{d}}\ln{\mathcal{C}}_n(\Lambda)}{{\mathrm{d}}\ln\Lambda}\;\;.$$ This is Wilson’s Renormalisation Group Equation for the observable ${\mathcal{O}}$. Note that eq. features a *total* derivative: LECs in ${\mathcal{O}}$ are readjusted as $\Lambda$ changes.
In practise, an EFT at finite order $n$ and with finite cutoff tolerates cutoff artefacts which are parametrically small, i.e. at least of order $n+1$. This also limits the rate of change in the residual ${\mathcal{C}}_n$: I call an observable “perturbatively renormalised” when the right-hand side of eq. is smaller than any term on the left-hand side. To some, this condition implies $\Lambda$ can only be varied in a range around ${\overline{\Lambda}_\text{EFT}}$; the functional dependence on $k$ and $n$ is then still a quantitative prediction. A double-logarithmic plot reveals quantitative aspects of the Renormalisation Group evolution and can be utilised to falsify claims of consistency in an EFT.
[**Extending the Expansion**]{} The order $n$ is not counted relative to LO. It is not even necessarily an integer, as Table \[tab:pc\] shows, and the first omitted order is not always $Q^{n+1}$, but more generally $Q^{n+\alpha}$, $\mathrm{Re}[\alpha]>0$. To replace $n+1\to n+\alpha$ in eqs. , , – and indeed throughout – is straightforward. In [EFT($\slashed{\pi}$)]{}, the slope-fit in eq. endorses that the $3N$ PC proceeds in integer steps. Including non-analytic dependencies of the residuals on $k$ or ${p_\text{typ}}$ is also straightforward. For the remainder of the presentation, all such replacements are implied, but we stick to the integer case for convenience.
[**Assumptions of the Expansion**]{} The assumptions on the residual ${\mathcal{C}}_n$ are endorsed if order $n$ and breakdown scale ${\overline{\Lambda}_\text{EFT}}$ follow indeed the functional form of eq. or its variant . Naïve Dimensional Analysis (NDA) sets the magnitude of ${\mathcal{C}}_n$ to the scale of its running [@NDA; @NDA2]. Its cutoff-dependence and other effects are eventually absorbed into higher-order LECs (see also below).
We can actually be somewhat more specific about the condition that the variation of the residual ${\mathcal{C}}_n$ variation with respect to $\Lambda$ should be larger than that for other parameters. Since $k,{p_\text{typ}}\ll\Lambda_1,\Lambda_2$, the dimensionless ratio on the left-hand side of eq. can be expanded as $$\label{eq:residualexpansion}
\frac{{\mathcal{C}}_n(\Lambda_1;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})-
{\mathcal{C}}_n(\Lambda_2;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})}
{{\mathcal{C}}_n(\Lambda_1;k,{p_\text{typ}},{\overline{\Lambda}_\text{EFT}})}= c_0(\Lambda_1,\Lambda_2;{\overline{\Lambda}_\text{EFT}})+
c_1(\Lambda_1,\Lambda_2;{\overline{\Lambda}_\text{EFT}})\;\frac{k,{p_\text{typ}}}{\Lambda_1,\Lambda_2}+\dots$$ If the first term dominates, then the dependence of eq. on $k$ and ${p_\text{typ}}$ is indeed indicative of the order $Q^{n+1}$. If subsequent terms dominate, the slope may be larger than $n+1$ – but never smaller.
[**Necessary but Not Sufficient**]{} This shows that a slope smaller than $n+1$ conclusively demonstrates failure of the PC to be consistent. However, the criterion is necessary rather than sufficient: Slopes $\ge n+1$ are proof neither of failure, nor of success. Indeed, a PC may be inconsistent but the coefficient of the terms with slope $<n+1$ may be anomalously small, leading to a “false positive”.
[**Estimating the Expansion Parameter**]{} When the cutoffs $\Lambda_1$ and $\Lambda_2$ are both varied over a wide range[^1], the analysis also gives a practical way to find the size of the expansion parameter as a function of $k$. Ratios between different orders estimate $Q(k,{p_\text{typ}})$, and hence residual theoretical uncertainties as function of $k$. This is of course only one way to assess $Q(k)$; within reason, the least optimistic and hence most conservative of several methods should be picked. For example, Ref. [@Griesshammer:2011md] combined this with the convergence pattern of the EFT series; see also [@Furnstahl:2015rha].
[**Choice of Expansion Parameter**]{} In Sect. \[sec:application\], $k$ is varied while the other scales ${p_\text{typ}}$ are fixed, but any combination of the low-energy scales may serve as variable(s). For example, scanning in the pion mass at fixed $k\ll m_\pi$ may elucidate the $m_\pi$-dependence of some couplings, with particular relevance to extrapolating lattice computations at non-physical pion masses. Here, I will continue to concentrate on variations with $k$, but most issues transfer straightforwardly to other variations.
[**Window of Opportunity**]{} One can read off slopes most easily in the range ${p_\text{typ}}<k<{\overline{\Lambda}_\text{EFT}}$. In [EFT($\slashed{\pi}$)]{}, that window is narrow but suffices: ${\overline{\Lambda}_{\slashed{\pi}}}/({p_\text{typ}}\sim\gamma_{t,s})\lesssim3$. In [$\chi$EFT]{}with dynamical $\Delta(1232)$ degrees of freedom, we expect a wider range: ${\overline{\Lambda}_{\chi}}/({p_\text{typ}}\sim m_\pi)\gtrsim4$. One may of course also fit the variables $n$ and ${\overline{\Lambda}_\text{EFT}}$ in eq. to the numerical results outside that window, but then one needs to specify ${p_\text{typ}}$ and determine its contribution relative to $k$.
[**Choice of Regulator**]{} Residual cutoff dependence comes naturally in numerical computations. This tests uses it as a tool to check consistency. The example used a “hard” cutoff, but ${\overline{\Lambda}_\text{EFT}}$ and $n$ do not depend on a specific regulator. If the theory can be renormalised exactly, all residual regulator dependence disappears by dimensional transmutation; cf. .
[**Choice of Cutoffs**]{} The functional dependencies of eqs. and on $n$ and ${\overline{\Lambda}_\text{EFT}}$ do not depend on $\Lambda_1$ and $\Lambda_2$. While any two cutoffs $\Lambda_1,\Lambda_2\gtrsim{\overline{\Lambda}_\text{EFT}}$ will do in principle, small leverage may lead to numerical artefacts. The larger $\Lambda_2-\Lambda_1$, the clearer the signal should be. For our example, Fig. \[fig:nonLepage900MeV\] shows that an upper cutoff of $900\;{\ensuremath{\mathrm{MeV}}}$ instead of $600\;{\ensuremath{\mathrm{MeV}}}$ leads to different curves but very similar slopes. Infinities, zeroes and oscillations of ${\mathcal{O}}$ with $k$ for any pair $\Lambda_1,\Lambda_2$ can lead to problems (see “*Observables: Accidental Zeroes and Infinities*” below) which are readily avoided by choosing a cutoff pair such that ${\mathcal{O}}(\Lambda_1)-{\mathcal{O}}(\Lambda_2)>0$ for all $k$. Even when one does not choose to take one of the cutoffs to infinity[^2], a reasonable range of allowed cutoffs exists. If $\Lambda_1\approx\Lambda_2$, one may of course directly consider the numerical derivative of eq. – over a range of cutoffs. \[To reiterate: exact cutoff independence ${\mathcal{O}}(\Lambda_1)\equiv{\mathcal{O}}(\Lambda_2)$ for any cutoff pair is not considered.\]
![\[fig:nonLepage2\] \[fig:nonLepage900MeV\] \[fig:EREvsZ\] *Left*: Thick coloured lines: Z-parametrisation of the $\mathrm{NN}$ amplitude as in Fig. \[fig:nonLepage\], but for $\Lambda_1=900\;{\ensuremath{\mathrm{MeV}}}$, not $600\;{\ensuremath{\mathrm{MeV}}}$; thin gray lines: Bethe’s Effective Range Parametrisation; from Ref. [@improve3body]. \[fig:toB3\] *Right*: Test when the leading $3N$ interaction is determined not by the $Nd$ scattering length as in Fig. \[fig:nonLepage\], but by the position of the triton pole. The [N${}^{2}$LO]{} fit uses again the $Nd$ scattering length and triton binding energy.](nonLepage.ZvsEREparameterisation.900MeV.pdf "fig:"){height="0.285\linewidth"} ![\[fig:nonLepage2\] \[fig:nonLepage900MeV\] \[fig:EREvsZ\] *Left*: Thick coloured lines: Z-parametrisation of the $\mathrm{NN}$ amplitude as in Fig. \[fig:nonLepage\], but for $\Lambda_1=900\;{\ensuremath{\mathrm{MeV}}}$, not $600\;{\ensuremath{\mathrm{MeV}}}$; thin gray lines: Bethe’s Effective Range Parametrisation; from Ref. [@improve3body]. \[fig:toB3\] *Right*: Test when the leading $3N$ interaction is determined not by the $Nd$ scattering length as in Fig. \[fig:nonLepage\], but by the position of the triton pole. The [N${}^{2}$LO]{} fit uses again the $Nd$ scattering length and triton binding energy.](nonLepage.Zparam-H0toB3-600.pdf "fig:"){height="0.285\linewidth"}
[**Decreasing Cutoff Dependence**]{} Equation is a variant of the Renormalisation Group evolution of ${\mathcal{O}}$, eq. , which in turn quantifies the fundamental EFT tenet that observables must become order-by-order less sensitive to loop contributions beyond ${\overline{\Lambda}_\text{EFT}}$, the range of applicability. Cutoff dependence in observables should therefore generically decrease from order to order, irrespective whether or not LECs are fitted. This does not apply to the ${\mathcal{C}}_n$ themselves, but to the entire left-hand side of eq. . Refitting LECs may of course help to absorb some cutoff dependence. Indeed, no new LECs enter at NLO in the example above ($H_0$ is just refitted), and the cutoff dependence decreases from LO to NLO. While it is conceivable that the residual ${\mathcal{C}}_n$ is sometimes somewhat larger than NDA predicts, NDA should apply “most of the time”, statistically speaking.
Still, a specific regulator form may produce a very small residual cutoff dependence at one order but a significantly larger one at a subsequent order, ${\mathcal{C}}_n(\Lambda_1)-{\mathcal{C}}_n(\Lambda_2)<
{\mathcal{C}}_{n+1}(\Lambda_1)-{\mathcal{C}}_{n+1}(\Lambda_2)$. This may for example occur if the regulator produces only corrections with even powers of $\Lambda$ and the numerics preserves this symmetry at least approximately (e.g. because $\Lambda_1\approx\Lambda_2$, allowing for a perturbative expansion). If this overwhelms the expansion in $Q$, ${\mathcal{O}}$ may indeed systematically become more dependent on $\Lambda$ between some orders, but not between all. Nonetheless, one should not just see some qualitatively improved cutoff dependence with increasing order, but one must see the quantitatively predicted slopes emerge for many orders: they must be $\ge n+1$; see eq. .
[**Constructing a PC by Trial-and-Error**]{} If the cutoff dependence of a given observable does not decrease consistently between subsequent orders, caution may be advisable. For example, $\Lambda$-dependence may increase from one order to the next, but then decrease markedly when another full order with a new LEC is included. This could signal that this LEC cures cutoff dependence already at a lower order – and hence that the PC is inconsistent. One should then study the convergence pattern as the LEC is promoted to a lower order such that the cutoff dependence decreases always between subsequent orders. This may help to construct a consistent PC by trial-and-error and iteration. Remember also that after a LEC starts contributing at a certain order, it is re-adjusted at each subsequent order to absorb both cutoff effects and still match its determining datum.
[**Calculating Higher Orders**]{} Traditionally, observables beyond LO have been found by “partially resumming” contributions, i.e. the power-counted potential is iterated like in Weinberg’s original suggestion. Since corrections to the LO potential are defined as parametrically small, they can be included in “strict perturbation”, avoiding potential problems with spurious deeply bound state which can be generated by iteration [@Vanasse:2013sda]. This may also provide clearer signals for the PC test.
Picking Observables
-------------------
[**Isolating Dynamical Effects**]{} While any observable could be chosen, those which are free from kinematic or other constraints (e.g. from symmetries) are preferred. Consider the scattering amplitude ${\mathcal{A}}_l$ in the $l$th partial wave (for simplicity, assume no mixing). Since it is complex, one could choose ${\mathcal{O}}=|{\mathcal{A}}_l|$. However, unitarity relates ${\mathcal{A}}_l=1/(k\cot\delta_l-{\mathrm{i}}k)$ to the phase shift $\delta_l$. This constraint dominates when $\delta_l$ is between about $\pi/4$ and $3\pi/4$ – which affects much of the $\mathrm{NN}$ S-wave phase shifts. Even outside this interval, the additional contribution to eq. is not sensitive to dynamics. In addition, analyticity dictates that phase shifts approach zero like $k^{2l+1}$ for $k\to0$ in the $l$th partial wave. Since both numerator and denominator in eq. are then zero, ${\mathcal{O}}=\delta_l$ is dominated by numerical uncertainties as $k\to0$. This may not be a problem if the region in which the slopes are determined is far away, but only a closer inspection could tell if that holds. Likewise, one eliminates phase-space factors in decay constants, production cross sections, etc.
A sensible choice for single-channel scattering appears thus to be ${\mathcal{O}}=k^{2l+1}\cot\delta_l$: It is only constrained to be real below the first inelasticity, and imaginary parts are usually small above it. Indeed, the S-wave example above kept track of the imaginary part by plotting $$\left|1-\frac{k\cot\delta_0(\Lambda_2)}{k\cot\delta_0(\Lambda_1)}\right|\;\;.$$ While factors of $k$ formally cancel, one computes ${\mathcal{A}}$ ($k\cot\delta_0$), so that numerics is more benign.
[**Partial-Wave Mixing**]{} In the $\mathrm{NN}$ system, two partial waves with total angular momentum $J$ mix. The corresponding unconstrained observables in the Stapp-Ypsilanti-Metropolis (SYM or “nuclear-bar”) parametrisation are $$k^{2\pm1-2J}\bar{\delta}_{J\pm1}\;\;\mbox{ and }
\;\;k^{-(2J+1)}\bar{\epsilon}_J\;\;.$$ In the Blatt-Biedenharn parametrisation, the same rules apply for the eigenphases, but $k^{-2}\epsilon_J$ is the unconstrained variable for the mixing angle; see e.g. [@deSwart:1995ui]. These choices do not suffer from unitarity constraints (except for being real below the first inelasticity) and can be used directly.
[**Dependence on Parameter Input**]{} Let us first consider processes in which ${\mathcal{O}}(k)$ is a parameter-free prediction, i.e. its LECs are all known from some other process(es). To what extent does the procedure depend on that choice? In the example, the two-nucleon interactions were determined to match the Z-parametrisation of $\mathrm{NN}$-scattering (fit to pole position and residue of the scattering amplitude) [@Phillips:1999hh]. Fig. \[fig:EREvsZ\] shows that results with Bethe’s Effective-Range parametrisation have a markedly different rate of convergence, but the extracted slopes and ${\overline{\Lambda}_{\slashed{\pi}}}$ agree very well [@improve3body].
[**Accidental Zeroes and Infinities**]{} Some observables may show additional structures which should be avoided. For example, the [${}^{3}\mathrm{P}_{0}$]{} phase shift in $\mathrm{NN}$ scattering is zero at a lab energy of about $150\;{\ensuremath{\mathrm{MeV}}}$, so that the relative deviation of ${\mathcal{O}}=\delta_l$ in eq. diverges. Likewise, ${\mathcal{O}}=k^{2l+1}\cot\delta_l$ diverges (approaches zero) at $\delta_l=0$ ($\pi/2$), e.g. in the [${}^{1}\mathrm{S}_{0}$]{} wave at $k\approx370\;{\ensuremath{\mathrm{MeV}}}$ and [${}^{3}\mathrm{S}_{1}$]{} wave at $k\approx90\;{\ensuremath{\mathrm{MeV}}}$ and $400\;{\ensuremath{\mathrm{MeV}}}$ [@Epelbaum:2014efa]. As the qualitative plot in Fig. \[fig:qualitative\] shows, the corresponding spikes may make it more difficult to determine slopes.
[**Fitting to a Point**]{} A “zero” in eq. is induced intentionally when the observable contains a LEC that is determined in the channel in which one tests the PC. If the observable is tuned to exactly reproduce a certain value at some point $(k_0,{p_\text{typ}})$, then ${\mathcal{O}}(k_0;\Lambda_1)-{\mathcal{O}}(k_0;\Lambda_2)=0$ – with all the problems mentioned just now. Obviously, one should choose the fit point to be outside the slope-region. In the example of Sect. \[sec:application\], the strength of the $3N$ interaction $H_0$ without derivatives was fixed at each order to the $Nd$ scattering length, i.e. using $k=0$ as fit point. That is far away from the slope-region. At [N${}^{2}$LO]{}, the momentum-dependent $3N$ interaction $H_2$ was in addition determined from the triton binding energy $B_3=8.48\;{\ensuremath{\mathrm{MeV}}}$, i.e. the pole in the amplitude is fixed to $k_0=\sqrt{-4M
B_3/3}\approx100\;{\mathrm{i}}\;{\ensuremath{\mathrm{MeV}}}$. If one chooses this fit point for $H_0$ at LO and NLO, instead of $k_0=0$, the pattern of the slopes is wiped out; see Fig. \[fig:toB3\]. It appears that fitting only at $k_0$ introduces a new low-energy scale ${p_\text{typ}}$ and leaves no window ${\overline{\Lambda}_{\slashed{\pi}}}\gg
k\gg|k_0|\approx100\;{\ensuremath{\mathrm{MeV}}}$, while the [N${}^{2}$LO]{} fit at both $k=0$ and $k_0$ does not suffer this limitation.
[**Fitting in a Region**]{} The issue is less transparent when the LEC is not determined by exactly reproducing some data, but by least-$\chi^2$ fitting over a whole region in $k$. That is the typical case in $\mathrm{NN}$ scattering; see e.g. Ref. [@Epelbaum:2014efa]. The deviation of the fitted result from data is more regular at any given cutoff $\Lambda$ than when it is exactly zero at $k_0$. A pronounced spike is therefore replaced by a more uniform behaviour inside the fit region. The comparison between two cutoffs in eq. is therefore also more uniform as a function of $k$. Since cutoff variations can now be balanced by adjusting LECs, the coefficients ${\mathcal{C}}_n$ are artificially small in that régime. One still expects the cutoff dependence to decrease order-by-order, but the characteristic slopes are harder to see since the observable is constrained by the fit. Just like in the neighbourhood of a fit point, an observable will first have to shed the fit constraints outside the fit region for pronounced slopes.
Such a fit region must of course be inside the applicability range of the EFT. Traditional fits do not take into account that the systematic uncertainties of an EFT increase with $k$ but assign a $k$-independent uncertainty weight. Eq. suggests that this is justified for $k\lesssim{p_\text{typ}}$ since the error varies only mildly. In that case, one can speculate that the impact on the slopes at higher $k$ is not too big. This limits a reasonable fit region to $k\lesssim\gamma_{t,s}$ in [EFT($\slashed{\pi}$)]{}; and to $k\lesssim{\ensuremath{m_\pi}}$ in [$\chi$EFT]{}. In addition, one expects clearer signals if the same fit region is used at each order. It is difficult to see how slopes can clearly be identified when the fit region extends far towards ${\overline{\Lambda}_\text{EFT}}$. Practical considerations, like insufficient or low-quality data at low momenta may well override this choice.
[**Fitting to Pseudo-Data**]{} As a recourse and in order to assess the impact of a fit region on the slopes, one may create an artificial, “exact datum” ${\mathcal{O}}_0(k_0)$ at very low $k\to0$ which agrees with low-energy data (e.g. a scattering length, effective range, etc); and then assess the dependence of the slope on reasonable variations of ${\mathcal{O}}_0(k_0)$. The goal is then not to find good agreement with actual data at higher energies, but to test the convergence pattern.
[**Summary: Choice of Observable**]{} Ideal candidates for ${\mathcal{O}}$ are positive-definite observables which are not subject to unitarity and other constraints, and which are nonzero and finite over a wide range in $k$ and $\Lambda$, including the régime $k\gtrsim{p_\text{typ}}$ where one hopes to determine the slope. EFT parameters/LECs should be determined at very low $k$. A good signal may need some creativity. The choices ${\mathcal{O}}=k^{2l+1}\cot\delta_l$, $k^{2\pm1-2J}\bar{\delta}_{J\pm1}$ and $k^{-(2J+1)}\bar{\epsilon}_J$, with effective-range parameters determining unknowns, appear suitable in most scattering cases.
Miscellaneous Notes {#sec:misc}
-------------------
[**Consistency Assessment vs. “Lepage Plots”**]{} Double-logarithmic convergence plots are not unfamiliar. Lepage compared to data in order to quantify how accurately the EFT reproduces experimental information [@Lepage:1997cs]. This triggered a series of influential studies of differences between approximations and “exact results” in toy-models, see e.g. [@Steele:1998un; @Steele:1998zc; @Kaplan:1999qa]. Recently, Birse perused similar techniques, after removing the strong influence of long-range Physics (One- and Two-Pion Exchange) from empirical phase shifts in a modified effective range expansion, allowing for a more detailed study of the residual short-distance interactions [@Birse:2007sx; @Birse:2010jr]. Such investigations assume that the correct PC is known and quantitative comparison to data is needed.
The test advocated here aims to answer different questions: Does the output match the assumptions? Is the theory consistent? Recall that an EFT may converge by itself, but not to data, if some dynamical degrees of freedom are incorrect or missing. For example, a [$\chi$EFT]{}without dynamical $\Delta(1232)$ at $k\approx300\;{\ensuremath{\mathrm{MeV}}}$ cannot reproduce Delta resonance properties – but it may well be consistent. In other words, an EFT may be consistent, but not consistent with Nature.
[**Insensitivity to Some LECs**]{} This procedure can only help determine if a LEC is correctly accounted for when it is needed to absorb residual cutoff dependence. Eq. then determines its running, and its initial condition is fixed by some input, for example data or results of a more fundamental theory. Some LECs do however start contributing just because of their natural size, and not to renormalise that order. For example, the magnetic moment of the nucleon enters the one-baryon Lagrangean of [$\chi$EFT]{}at NLO, albeit it is not needed to renormalise loops. Similarly, the contribution of a LEC to a particular observable may be unnaturally small (or even zero).
[**Numerics**]{} The analysis can be numerically indecisive. We would trust results only if $n$ and ${\overline{\Lambda}_\text{EFT}}$ can be determined quite robustly in a reasonably wide range to cutoffs (and, possibly, cutoff forms), parameter sets and fit-windows. None of this provides, however, sufficient excuse not report results.
[**Sampling Tests**]{} Finding that the slope at each order $Q^n$ is not smaller than $n+1$ is necessary but not sufficient for a consistent PC. We saw that fine-tuning, particular choices of regulator forms and observables, and anomalously small coefficients are some reasons which may hide signals of slopes $<n+1$ which violate the PC assumptions. If slopes are always $\ge n+1$ for a variety of independent observables, regulators etc., that may increase confidence in PC consistency – but cannot prove it.
Outlook
-------
The [$\chi$EFT]{}power-counting proposals differ most starkly in the attractive triplet partial waves of $\mathrm{NN}$ scattering since they reflect different philosophies on how to treat the non-selfadjoint, attractive $1/r^3$ potential at short distances which appears at leading order; see Table \[tab:pc\]. It would therefore be interesting to see this test applied to the [${}^{3}\mathrm{P}_{0}$]{} wave and to the [${}^{3}\mathrm{P}_{2}$]{}-[${}^{3}\mathrm{F}_{2}$]{} system. The test proposed here is not necessarily a silver bullet to endorse or reject a particular counting since its results may in the worst case be inconclusive. But that implies it is still worth a try.
I cordially thank the organisers of <span style="font-variant:small-caps;">Chiral Dynamics 2015</span> in Pisa for a stimulating atmosphere, and the participants for enlightening and entertaining discussions. These notes grew out of the inspirational and intense discourses at the workshops <span style="font-variant:small-caps;">Nuclear Forces from Effective Field Theory</span> at CEA/SPhN Saclay in 2013, and <span style="font-variant:small-caps;">Bound States and Resonances in Effective Field Theories and Lattice QCD Calculations</span> in Benasque (Spain) in 2014. I am most grateful to their organisers and participants. Since 2013, exchanges with M. Birse, B. Demissie, E. Epelbaum, R. Furnstahl, B. Long, M. Pavon Valderrama, D. R. Phillips, M. Savage, R. G. E. Timmermans, U. van Kolck and Ch.-J. Yang allowed me to develop these ideas into a sharper analysis tool. M. Birse, B. Demissie, E. Epelbaum, D. R. Phillips, and in particular the referee (DRP), suggested important improvements to this script. I am especially indebted to ceaseless questions by many emerging researchers. Finally, my colleagues may forgive mistakes and omissions in referencing work and historical precedents, and graciously continue to point out necessary corrections. This work was supported in part by the US Department of Energy under contract DE-FG02-95ER-40907, and by the Dean’s Research Chair programme of the Columbian College of Arts and Sciences of The George Washington University.
[99]{}
H. W. Grie[ß]{}hammer, forthcoming.
The Editors, *Editorial: Uncertainty Estimates*, Phys. Rev. A**83** (2011) 040001.
*Enhancing the Interaction between Nuclear Experiment and Theory Through Information and Statistics*, special issue J. Phys. **G42** number 3 (March 2015).
G. P. Lepage, nucl-th/9706029. A. Manohar and H. Georgi, Nucl. Phys. B**234** (1984), 189 \[n.b. Acknowledgement\].
H. Georgi and L. Randall, Nucl. Phys. B**276** (1986), 241.
P. F. Bedaque, U. van Kolck, Ann. Rev. Nucl. Part. Sci. [**52**]{}, 339-396 (2002). \[nucl-th/0203055\].
L. Platter, Few Body Syst. [**46**]{}, 139-171 (2009). \[arXiv:0904.2227 \[nucl-th\]\].
S. Weinberg, Nucl. Phys. B**363** (1991), 3.
M. C. Birse, Phys. Rev. C [**74**]{}, 014003 (2006) \[nucl-th/0507077\]. M. C. Birse, PoS CD [**09**]{}, 078 (2009) \[arXiv:0909.4641 \[nucl-th\]\]. M. Pavon Valderrama and E. Ruiz Arriola, Phys. Rev. C [**74**]{}, 054001 (2006) \[nucl-th/0506047\]. M. Pavon Valderrama and E. Ruiz Arriola, Phys. Rev. C [**74**]{}, 064004 (2006) \[nucl-th/0507075\]. M. P. Valderrama, Phys. Rev. C [**83**]{}, 024003 (2011) \[arXiv:0912.0699 \[nucl-th\]\]. M. Pavon Valderrama, Phys. Rev. C [**84**]{}, 064002 (2011) \[arXiv:1108.0872 \[nucl-th\]\]. B. Long and C. J. Yang, Phys. Rev. C [**85**]{}, 034002 (2012) \[arXiv:1111.3993 \[nucl-th\]\]. B. Long and C. J. Yang, Phys. Rev. C [**86**]{}, 024001 (2012) \[arXiv:1202.4053 \[nucl-th\]\]. H. W. Grie[ß]{}hammer, *Summary: Systematising the $NN$ System in Chiral Effective Field Theory*, remarks at <span style="font-variant:small-caps;">Nuclear Forces from Effective Field Theory</span>, CEA/SPhN Saclay (France) 2013.
D. R. Phillips, PoS CD [**12**]{}, 013 (2013) \[arXiv:1302.5959 \[nucl-th\]\]. D. B. Kaplan, M. J. Savage and M. B. Wise, Nucl. Phys. B [**478**]{}, 629 (1996) \[nucl-th/9605002\]. D. B. Kaplan, M. J. Savage and M. B. Wise, Nucl. Phys. B [**534**]{}, 329 (1998) \[nucl-th/9802075\]. S.R. Beane, P.F. Bedaque, M.J. Savage, U. van Kolck, Nucl. Phys.A [**700**]{}, 377 (2002) \[nucl-th/0104030\]. A. Nogga, R. G. E. Timmermans, U. van Kolck, Phys. Rev. C [**72**]{}, 054006 (2005) \[nucl-th/0506005\]. E. Epelbaum and U.-G. Meissner, Few Body Syst. [**54**]{}, 2175 (2013) \[nucl-th/0609037\]. S. R. Beane, D. B. Kaplan, A. Vuorinen, Phys. Rev. C [**80**]{}, 011001 (2009) \[arXiv:0812.3938 \[nucl-th\]\]. E. Epelbaum and J. Gegelia, Eur. Phys. J. A [**41**]{}, 341 (2009) \[arXiv:0906.3822 \[nucl-th\]\]. P. F. Bedaque, H. W. Grie[ß]{}hammer, G. Rupak and H.-W. Hammer, Nucl. Phys. **A714** (2003), 589 \[nucl-th/0207034\].
H. W. Grie[ß]{}hammer, Nucl. Phys. **A744** (2004), 192 \[nucl-th/0404073\].
H. W. Grie[ß]{}hammer, *Introduction to Effective Field Theories*, <span style="font-variant:small-caps;">National Nuclear Physics Summer School 2008</span>, Washington DC (USA); notes at [home.gwu.edu/ hgrie/](http://home.gwu.edu/~hgrie/lectures/lectures.0806.NNPSS2008-GW.djvu).
H. W. Grie[ß]{}hammer, *Testing a Power Counting*, remarks at <span style="font-variant:small-caps;">Bound States and Resonances in Effective Field Theories and Lattice QCD Calculations</span>, Benasque (Spain) 2014.
R. J. Furnstahl, D. R. Phillips, S. Wesolowski, J.Phys.G[**42**]{} 034028 (2015) \[arXiv:1407.0657\[nucl-th\]\]. E. Epelbaum, H. Krebs and U. G. Mei[ß]{}ner, Eur. Phys. J. A [**51**]{} (2015), 53 \[arXiv:1412.0142 \[nucl-th\]\]. H. W. Grie[ß]{}hammer and M. R. Schindler, Eur. Phys. J. A [**46**]{}, 73 (2010) \[arXiv:1007.0734 \[nucl-th\]\]. H. W. Grie[ß]{}hammer, Nucl. Phys. A**760** (2005), 110 \[nucl-th/0502039\].
L. Platter and D. R. Phillips, Few Body Syst. [**40**]{}, 35 (2006) \[cond-mat/0604255\]. C. Ji and D. R. Phillips, Few Body Syst. [**54**]{}, 2317 (2013) \[arXiv:1212.1845 \[nucl-th\]\]. J. Vanasse, Phys. Rev. C [**88**]{}, 044001 (2013) \[arXiv:1305.0283 \[nucl-th\]\].
H. W. Grie[ß]{}hammer, M. R. Schindler and R. P. Springer, Eur. Phys. J. A [**48**]{}, 7 (2012) \[arXiv:1109.5667 \[nucl-th\]\]. R. J. Furnstahl, N. Klco, D. R. Phillips and S. Wesolowski, Phys. Rev. C [**92**]{}, 024005 (2015) \[arXiv:1506.01343 \[nucl-th\]\].
J. J. de Swart, C. P. F. Terheggen and V. G. J. Stoks, nucl-th/9509032. D. R. Phillips, G. Rupak and M. J. Savage, Phys. Lett. B [**473**]{}, 209 (2000) \[nucl-th/9908054\]. J. V. Steele and R. J. Furnstahl, Nucl. Phys. A [**637**]{}, 46 (1998) \[nucl-th/9802069\]. J. V. Steele and R. J. Furnstahl, Nucl. Phys. A [**645**]{}, 439 (1999) \[nucl-th/9808022\]. D. B. Kaplan and J. V. Steele, Phys. Rev. C [**60**]{}, 064002 (1999) \[nucl-th/9905027\]. M. C. Birse, Phys. Rev. C [**76**]{}, 034002 (2007) \[arXiv:0706.0984 \[nucl-th\]\]. M. C. Birse, Eur. Phys. J. A [**46**]{}, 231 (2010) \[arXiv:1007.0540 \[nucl-th\]\].
[^1]: Some claim that renormalisability requires that ${\mathcal{O}}$ has a unique limit as $\Lambda\to\infty$.
[^2]: One could adhere to the philosophy that cutoffs and breakdown scales should be similar.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Cyclotron Institute and Department of Physics and Astronomy, Texas A&M University, College Station TX 77843, USA\
E-mail:
- |
Michael Kordell\
Cyclotron Institute, Texas A&M University, College Station TX 77843, USA\
E-mail:
title: Hybrid Hadronization
---
Hadronization of partons into hadrons is poorly understood from first principles. Yet most observables in high energy collisions with hadrons in the final state are to some degree sensitive to hadronization dynamics. Over the past decades phenomenological models of hadronization have been developed. They can be tuned to describe a large amount of data consistently. Lund string fragmentation [@Andersson:1983ia] and quark recombination [@Fries:2008hs] are among the most successful hadronization models available. Hybrid hadronization is proposed as a new hadronization model which incorporates both Lund string fragmentation and quark recombination. It interpolates smoothly between them based on physics principles discussed below. It recovers string fragmentation and pure quark recombination in limits of very dilute and very dense parton systems, respectively.
Lund string fragmentation has been developed to resemble properties of the QCD vacuum in which gluon field lines tend to create flux tubes over long distances. In Monte Carlo generators these QCD strings can be used to connect color charges over large distances to make color singlet structures which then break up into hadrons String formation and fragmentation have been incorporated into the PYTHIA 8 event generator [@Sjostrand:2014zea] . This model has been extremely successful in describing hadronization in dilute sytems like $e^++e^- \to \text{hadrons}$ and $p+p$. Quark recombination had first been considered in the 1970s and has seen a revival with the advent of data from high energy heavy ion colliders RHIC and LHC. It was proposed that in a phase space filled with partons, the valence quarks of hadrons do not have to be created by string fragmentation, but are simply taken from the available quarks and antiquarks. Several key observables, like enhanced baryon production in A+A and the constituent quark scaling of elliptic flow $v_2$ can be naturally explained through quark recombination [@Fries:2008hs]. The Hybrid Hadronization code has been created as part of the JETSCAPE effort [@Kauder:2018cdt]. Its goal is to describe both vacuum and in-medium hadronization consistently. Hybrid Hadronization should eventually be applicable to $p+p$, $p+$A and A$+$A systems alike.
We propose the following picture. The phenomenological $q\bar q$ potential in QCD is characterized by a Coulomb-like part at small distances and a linear part at large distances. The latter gives rise to QCD strings while at small distances the potential allows the direct formation of bound states. Thus in a given system of quarks one should loop over available quark-antiquark pairs and triplets of quarks and antiquarks, and calculate their probability to directly form a bound state or resonance. If recombination fails for a quark it has to rather form a string with a partner. This procedure will turn a system of quarks into a system of recombined hadrons and resonances plus QCD strings. The strings capture all partons with large separation from other colored objects. The length scale is set by the typical size of hadron wave functions. Strings can then be handed over to PYTHIA for fragmentation. The quark recombination step is suitable to describe medium effects for perturbative parton showers. Sampled thermal partons can simply be added to partons from a shower Monte Carlo. All partons are subsequently processed equally, allowing for recombination of mesons and baryons that contain both thermal and shower partons.
The quark recombination algorithm used here was first proposed in [@Han:2016uhh] based on the instantaneous recombination model [@Fries:2008hs]. It uses the quark model and hadron wave functions based on harmonic oscillator potentials for simplicity. The widths of the wave functions for ground states have been fitted to measured charge radii where available. The recombination probability for two wave packets, representing quark and antiquark, and assumed to be Gaussian for simplicity, can then be calculated in the Wigner formalism. The Wigner formalism is convenient to use information available in both space and momentum space. For example, for light quarks and antiquarks with wave packets centered around phase space coordinates $(\mathbf{x}_1,\mathbf{p}_1)$ and $(\mathbf{x}_2,\mathbf{p}_2)$, respectively, the recombination probability into a ground state meson with wave function width $\sigma$, at a particular time $t$, is $$P(u) = e^{-u} \quad \text{where}\quad u = \frac{1}{2 \sigma^2} \left( \mathbf{x}_2-\mathbf{x}_1\right)^2
+ \frac{\sigma^2}{8} \left( \mathbf{p}_2-\mathbf{p}_1\right)^2$$ if the quark wave packets are assumed to have width $\sigma/\sqrt{2}$. The latter is chosen for convenience.[^1] All probabilities have to be evaluated in the common rest frame of the quark-antiquark pair [@Han:2016uhh]. Spins of partons are treated purely statistical. Color can be treated in two different ways. Either purely statistical, which is a good approximation in a sample of mostly thermal partons. Or color can be traced through color tags, as used in PYTHIA 8, which is desirable in jet showers in $p+p$ collisions and other dilute systems [@kordellfries].
The Hybrid Hadronization code is written in C++ and needs PYTHIA 8 to run. Internally it consists of four different stages.
- Read in information about partons to be hadronized. The code accepts individual parton jet showers from Monte Carlo codes like PYTHIA or MATTER [@Majumder:2013re]. It also acceptes full events, like PYTHIA $p+p$ events. String structures and color tags, if available, are preserved. Thermal partons can be added at this stage, e.g. by sampling the $T=T_c$ hypersurface from a fluid dynamic simulation of A+A collisions. Gluons are always split into quark-antiquark pairs.
- All combinations of parton pairs and triplets are tested for recombination. For each the probability from color$\times$spin$\times$wave function overlap as described above is evaluated. A die role determines if a recombination under consideration actually happens, and into which of the available hadron or hadron resonance channels it proceeds. The formation of hadrons from thermal partons only is currently disabled. All other hadrons formed during this stage are written to the Hybrid Hadronization output, including resonances.
- As partons from strings recombine, holes will appear in the original QCD strings. They are repaired according to an algorithm which will be described in detail elsewhere [@kordellfries]. As an example, the simplest case of string repair happens when a quark or antiquark terminating a string has recombined with a thermal parton, in which case the missing string parton is replaced by a suitable thermal parton.
- In the last step all surviving string structures are surveyed if they can be handled by PYTHIA 8, and they are processed further if needed. For example, the previous string repair step might have created systems with multiple junctions. The output after step (IV) consists of a list of recombined hadrons and resonances, and the remnant string system. It can now be handed over to PYTHIA 8 which will fragment the strings and decay all resonances according to user settings.
![A preliminary result obtained from Hybrid Hadronization: The ratio of charged hadron yields in $p+p$ collisions at $\sqrt{s}=2.76$ TeV hadronized with Hybrid Hadronization and PYTHIA 8, as a function of hadron rapidity $y$ and for different $\hat p_T$ (25 GeV, 100 GeV, 500 GeV, 1 TeV) in PYTHIA 8. Deviations between the two calculations are generally less than 2 %. []{data-label="fig:1"}](Ratio_HHPY.png){width="0.7\columnwidth"}
![A preliminary result obtained from Hybrid Hadronization: The proton/pion ratio in Au+Au collisions at $\sqrt{s_\text{NN}}=200$ GeV and $p+p$ collisions at the same energy. Thermal-shower recombination is allowed in the Au+Au case but purely thermal bulk hadrons are not included. Thermal-shower recombination leads to the signature enhancement of baryons in the Au+Au case.[]{data-label="fig:2"}](Picture1.png){width="0.7\columnwidth"}
Testing of the Hybrid Hadronization code is ongoing but we can report preliminary results. Note that not tuning to data has been carried out. To restate the goals, we would like to interfere minimally with the success story of Lund strings in dilute systems. At the same time, we would like to reproduce signature effects of recombination in A+A collisions. The likelihood of quark recombination in jet showers has already been studied to some extent in [@Han:2016uhh] and we refer the reader to that reference for more details. The conclusion is that jets seem to naturally feature a bulk region in which several partons can be quite close together in phase space, in addition to a very dilute tail. The bulk part is associated with the low-$z$ part of the shower, where $z$ as usual is the momentum fraction in a jet, while the large-$z$ part is typically dilute. Note that this is a statement averaged over many showers. Fluctuations between showers are very large due to the relatively small number of partons. Fig. \[fig:1\] shows the ratio of rapitidy spectra in minimum bias $p+p$ collisions at 2.76 TeV obtained from Hybrid Hadronization and pure PYTHIA 8 string fragmentation. Full PYTHIA 8 parton events have been used in both cases as input to hadronization. The four different lines correspond to events reported by PYTHIA with an initial momentum transfer $\hat p_T$ in a bin around the value shown. Generally we find rather small deviations between Hybrid Hadronization and pure string fragmentation in minimum bias $p+p$. Fig. \[fig:2\] shows the ratio of protons vs pions as a function of transverse momentum $P_T$ around midrapidity for Au+Au collisions at 200 GeV. Enhanced baryon production has been a hallmark of quark recombination models, and we see this behavior reproduced by Hybrid Hadronization. Here individual jet showers created by MATTER [@Majumder:2013re] were embedded into fluid dynamic events. The resulting enhancement can be traced back to shower-thermal recombination into baryons. They are compared to MATTER showers in $p+p$. Note that no underlying event in $p+p$ and no thermal hadrons in A$+$A have been added, thus results below $P_T=2$ GeV/$c$ are not shown.
To summarize, Hybrid Hadronization is a new hadronization model which combines features of string fragmentation and quark recombination. The goal is a hadronization model that can describe the transition between very dilute systems like $p+p$ and even $e^++e^-$, and heavy ion collisions. The code is currently undergoing testing and will be publicly released in a future JETSCAPE version.
This work was supported by the US National Science Foundation under award nos. 1516590, 1550221 and 1812431.
[99]{}
B. Andersson [*et al.*]{}, Phys. Rept. [**97**]{}, 31 (1983). R. J. Fries, V. Greco and P. Sorensen, Ann. Rev. Nucl. Part. Sci. [**58**]{}, 177 (2008). T. Sjöstrand [*et al.*]{}, Comput. Phys. Commun. [**191**]{}, 159 (2015). K. Kauder \[JETSCAPE Collaboration\], arXiv:1807.09615 \[hep-ph\].
K. C. Han, R. J. Fries and C. M. Ko, Phys. Rev. C [**93**]{}, 045207 (2016). M. Kordell and R. J. Fries, [*to be published*]{}.
A. Majumder, Phys. Rev. C [**88**]{}, 014909 (2013).
[^1]: Monte Carlo generators usually do not provide realistic estimates for widths of wave packets.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we investigate a local to global principle for [é]{}tale K-theory of curves. More precisely, we show that the result obtained by G.Banaszak and the author in [@bk13] describing the sufficient condition for the local to global principle to hold is the best possible (i.e this condition is also necessary). We also give examples of curves that fulfil the assumptions imposed on the Jacobian of the curve. Finally, we prove the dynamical version of the local to global principle for [é]{}tale $K$-theory of a curve. The dynamical local to global principle for the groups of Mordell-Weil type has recently been considered by S.Bara[ń]{}czuk in [@b17]. We show that all our results remain valid for Quillen $K$-theory of ${\cal X}$ if the Bass and Quillen-Lichtenbaum conjectures hold true for ${\cal X}.$'
address: 'Department of Mathematics and Physics, University of Szczecin, ul. Wielkopolska 15, 70-451 Szczecin, Poland '
author:
- Piotr Krasoń
title: 'Note on linear relations in [é]{}tale $K$-theory of curves'
---
Introduction
============
The local to global type questions are of mathematical interest since the celebrated Hasse principle was proven. For the history of these type of problems in number theory and its extensions to the context of abelian varieties and linear algebraic groups see [@fk17].
In [@bk13] G.Banaszak and the author proved a sufficient condition for the local to global condition to hold for [é]{}tale $K$-theory of curves. The main result of the current paper is a proof that this condition is also necessary (cf. Theorem \[best\]).
To state this result let us recall the following definitions and notations from [@bk13]. Let $X/F$ be a smooth, proper and geometrically irreducible curve of genus $g$ defined over a number field $F$ and let $J$ be the Jacobian of $X.$
We call a finite field extension $F^{\prime} / F$ an [**isogeny splitting field**]{} of the Jacobian $J$ if $J$ is isogenous over $F^{\prime}$ to the product $A_{1}^{e_{1}}\times \dots \times A_{t}^{e_{t}}$ where $A_1, \dots, A_t$ are pairwise nonisogenous, absolutely simple abelian varieties defined over $F^{\prime}$. \[definition of an isogeny splitting field\]
Notice that $J$ is an abelian variety and therefore by Poincar[é]{} decomposition theorem the isogeny splitting field exists.
Let ${\tilde V}_{l,\, i}$ denote $T_{l}(A_{i})(n)\otimes_{\Z_l} {\mathbb Q}_{l}$ , where $T_{l}(A_{i})(n)$ is the $n$-th twist of the $l$-adic Tate module of $A_{i}$ (cf. [@t68]).
The main result of [@bk13] is the following theorem:
\[theorem 1.1\] Let $X/F$ be a smooth, proper and geometrically irreducible curve of genus $g.$ Let $F^{\prime}$ be an isogeny splitting field of the Jacobian $J$ and assume that for the corresponding product $A_{1}^{e_{1}}\times \dots \times A_{t}^{e_{t}},$ we have $End_{{\overline F}} \, A_{i}={\mathbb Z}$ for each $1 \leq i \leq t.$ Let $l > 2$ be a prime number which is coprime to the polarisation degrees of the abelian varieties $A_i.$ Let $S_{l}$ be a set of places of $F$ containing the places of bad reduction, the archimedean places and the primes above $l.$ Let ${\cal X}$ be a regular and proper model of $X$ over ${\cal O}_{F, S_l}.$ Assume that ${{dim}}_{{\mathbb Q}_{l}}{\tilde V}_{l,\, i} \, \, \geq e_{i}$ for each $1 \leq i \leq t.$ Let $\hat{P} \in K_{2n}^{et}({\cal X})$ and let $\hat{\Lambda}$ be a finitely generated ${\mathbb Z}_{l}$-submodule of $ K_{2n}^{et}({\cal X}).$ If $r_v (\hat{P}) \in
r_v (\hat{\Lambda})$ for almost all $v$ of ${\cal O}_{F,S_{l}}$ then $\hat{P} \in
\hat{\Lambda} + K_{2n}^{et}({\cal X}) _{tor}.$
We prove the following theorem
\[best\] Let ${X/F}$ be a curve as in Theorem \[theorem 1.1\] . Then the numerical bound in Theorem \[theorem 1.1\] is the best possible i.e if ${{dim}}_{{\mathbb Q}_{l}}{\tilde V}_{l,\, i} \, \, < e_{i}$ then local to global principle of Theorem \[theorem 1.1\] does not hold.
Theorem \[best\] shows that the condition of Theorem \[theorem 1.1\] is also necessary.
Without loss of generality in the proof of Theorem \[best\] we may assume $F^{\prime}=F$. From now on we assume this.
In section 2 we give some examples of Jacobians of curves satisfying the assumptions of Theorem \[theorem 1.1\]. Sections 3,4 and 5 contain necessary background from Galois cohomology, intermediate Jacobians and Kummer theory. In section 6 we prove our main theorem (Theorem \[best\]). We also show that we can replace [é]{}tale $K$-theory by Quillen theory of a curve provided Bass and Quillen-Lichtenbaum conjectures hold true. Section 7 is devoted to proof of the dynamical version of the local to global principle. This is done by checking the axioms for the dynamical local to global principle introduced in [@b17]. We check these axioms for Galois cohomology (cf. Lemma \[kg\]) and then pass to the [é]{}tale $K$-theory.
Examples
========
In this section we give examples of curves whose Jacobian decomposes as $A_1^{e_1}\times \dots A_t^{e_t},$ with ${\End}_{\overline F}A_i={\mathbb Z}.$ In general deciding whether the Jacobian of a curve splits is a difficult problem and has vast literature (cf. e.g. [@hn65], [@es93]) However, we have the following [@co12], p.589:
\[isogeny\] Any abelian variety $A$ defined over $\overline{\mathbb Q}$ of dimension $1\leq g\leq 3$ is isogenous to $J(C)$, where $C$ is a curve of genus $g$.
Genus $1$ case of Theorem \[isogeny\] is trivial, genus $2$ is [@w57] Satz 2 and $g=3$ is Theorem 4 of [@ou73].
\[ex1\] Let $A/F$ be principally polarized abelian surface with ${\End_{{\overline F}}}={\mathbb Z}$ then by Theorem \[isogeny\] there exists a curve $X$ defined over $F^{'} \subset {\overline Q}$ such J(X) is isogenous to $A.$
\[Ex2\] Let $B=A\times E$, where $A$ is as in Example \[ex1\] and $E$ an elliptic curve without complex multiplication then there exists a curve $X$ such that J(X) is isogenous to $B.$
A special case of Example \[ex1\] is given in the following (see [@bpp18] section 7.1 p. 39):
\[ex3\] Let $X/{\mathbb Q}$ be a smooth projective curve given by the following equation $$\label{ex4}
X: \quad y^2 +(x^3+x^2 +x+1)y=-x^2-x.$$ Then J(X) is a principally polarized abelian isotypical surface over ${\mathbb Q}$ of conductor 277.
Since the ${\mathbf j}$ invariant of a curve with complex multiplication is an algebraic integer (cf. [@c89], Theorem 11.1 ), in Example \[Ex2\] it is enough to pick an elliptic curve defined over ${\mathbb Q}$ whose ${\mathbf j}$ invariant is non-integral rational number.
The following lemma guarantees that there are many simple abelian surfaces $A$ with ${\End}_{\overline F} A={\mathbb Z}$
([@bpp18], Lemma 4.1.1) Let $A$ be a simple, semistable abelian surface over ${\mathbb Q}$ with non-square conductor then $A$ is isotypical, i.e. ${\End}_{\overline{\mathbb Q}}A={\mathbb Z}$.
necessary results concerning [é]{}tale $K$-theory of curves and cohomology
==========================================================================
We start with the definition of continuous cohomology [@j88] and [@df85].
Let $Y$ be a scheme over $\Z[\frac{1}{l}]$ and let $({\cal F}_m)$ be a projective system of ${\Bbb Z}/{l^{m}}$ - [' e]{}tale sheaves ${\cal F}_m$. The functor $({\cal F}_m) \rightarrow
\varprojlim H^{0}(Y,\, ({\cal F}_m))$ is left exact and its $i$-th right derived functor is by definition the continuous cohomology group $H^{i}_{cts}(Y,\, ({\cal F}_m)).$
The [' e]{}tale K-theory spaces and groups were defined in [@df85] by W.Dwyer and E.Friedlander. In [@bgk99] using the spectral sequences of [@df85]: $$E_{2}^{p,-q} = H^{p}_{cont} (Y, \, \Z_l ({q}\, {/}{2})) \Rightarrow K_{q-p}^{et} (Y)$$ $$E_{2}^{p,-q} = H^{p}_{et} (Y, \, \Z/l^k ({q}\, {/}{2})) \Rightarrow K_{q-p}^{et} (Y, \, \Z/l^k)\, .$$ it was shown that one has the following exact sequence connecting continuous cohomology of $\cal X$ and the Galois cohomology of the Galois group of the maximal unramified outside $S_{l}$ extension of $F$
$$\label{diagram 2.1}
0\rightarrow H^{2}(G_{S_{l}} ; {\Bbb Z}_{l}(k)) \rightarrow H^{2}_{cts}({\cal X}; {\Bbb Z}_{l}(k)) \rightarrow H^{1}(G_{S_{l}} ; {T}_{l}(J)(k-1))\rightarrow 0 \, .$$
as well as existence of the following commutative diagram
$$\label{diagram 2.4}
\xymatrix{
K_{2n}^{et}({\cal X}) \ar[d] \ar[r]^{r_v} & K_{2n}^{et}({\cal X}_{v}) \ar[d]^{\cong} \\
H^{1}(G_{S_{l}}; {T}_{l}(J)(n)) \ar[r]^{r_v}& H^{1}({g}_{v}; {T}_{l}(J_{v})(n ))\, .
}$$
The right hand vertical arrow in the diagram (\[diagram 2.4\]) is an isomorphism whereas the left vertical arrow is an epimorphism with finite kernel (cf. [@bk13], [@bgk99], [@bg08]).
One has the following Dwyer-Friedlander homomorphisms, for odd prime $l$ and $k\geq 1$, connecting Quillen $K$-theory with the [é]{}tale $K$-theory with coefficients (cf. [@df85] ) $$\label{dfr}
h_{l^k} : K_m({\cal X}, {\mathbb Z}/{l^k{\mathbb Z}}) \rightarrow K_m^{et}({\cal X}, {\mathbb Z}/{l^k{\mathbb Z}})$$ $$\label{dfr0}
{\overline{h_{l^k}}} : K_m({\cal X}_{v}, {\mathbb Z}/{l^k{\mathbb Z}}) \rightarrow K_m^{et}({\cal X}, {\mathbb Z}/{l^k{\mathbb Z}}).$$
In the sequel we assume the following
(Bass conjecture) The Quillen $K$-theory groups $K_m({\cal X})$ are finitely generated for $m>0$.
\[Bass\] Assuming Conjecture
\[Bass\] we obtain the following equality $$\varprojlim K_m({\cal X},{\mathbb Z}/{l^k{\mathbb Z}})=K_m({\cal X})\otimes {\mathbb Z}_l.$$
The Dwyer-Friedlander homorphisms (\[dfr\]) and (\[dfr0\]) induce the following homomorphisms $$\label{dfr1}
h : K_{2n}({\cal X})\otimes {\mathbb Z}_l \rightarrow K_{2n}^{et}({\cal X})$$ $$\label{dfr2}
{\overline{h}} : K_{2n}({\cal X}_v)\otimes {\mathbb Z}_l\rightarrow K_{2n}^{et}({\cal X}_{v})$$
The maps (\[dfr1\]) and (\[dfr2\]) are isomorphism if we assume that the Quillen-Lichtenbaum conjecture holds true for $\cal X.$ Thus we obtain the following commutative diagram $$\label{diagram 2.5a}
\xymatrix{
K_{2n}({\cal X}) \ar[d]_{h^{'} }\ar[r]^{r_v} & K_{2n}({\cal X}_{v}) \ar[d]^{\overline{h^{'}} } \\
K_{2n}^{et}({\cal X})\ar[r]^{r_v} & \, K_{2n}^{et}({\cal X}_v)
}$$ where the map $h^{'}$ (resp. ${\overline{h^{'}} }$ ) is the composition of the natural homomorphism $K_{2n}({\cal X})\rightarrow K_{2n}({\cal X})\otimes {\mathbb Z}_l$ (resp. $K_{2n}({\cal X}_v)\rightarrow K_{2n}({\cal X}_v)\otimes {\mathbb Z}_l$) with $h$ (resp. $\overline{h}$).
Concatenation of diagrams (\[diagram 2.4a\]) and (\[diagram 2.5a\]) yields the following commutative diagram $$\label{diagram 2.4a}
\xymatrix{
K_{2n}({\cal X}) \ar[d] \ar[r]^{r_v} & K_{2n}({\cal X}_{v}) \ar[d]\\
H^{1}(G_{S_{l}}; {T}_{l}(J)(n)) \ar[r]^{r_v} & H^{1}({g}_{v}; {T}_{l}(J_{v})(n ))\,
}$$ with the vertical maps having finite kernels.
Intermediate jacobian
=====================
Let $L$ be a finite extension of $F.$ For $w\notin S_{l} ,$ let $G_{w} :=
G({\overline L_{w}}/L_w)$ be the absolute Galois group of the completion $L_{w}$ of $L$ at $w.$ Let $k_{w}$ be the residue field at $w.$ Put $H^{1}_{f}(G_{w};{\tilde T}_{l,\, i})=i_{w}^{-1}H^{1}_{f}(G_w ;{\tilde V}_{l,\, i}).$ Here $i_{w}: H^{1}(G_{w};{\tilde T}_{l,\, i})\rightarrow H^{1}(G_{w};{\tilde V}_{l,\, i})$ and $H^{1}_{f}(G_{w};{\tilde V}_{l,\, i})=ker(\, res: H^{1}(G_{w}; {\tilde V}_{l,\, i})\rightarrow
H^{1}(I_{w}; {\tilde V}_{l,\, i}) )$ where $I_{w} \subset G_{w}$ is the inertia subgroup and $res$ is the restriction map. Let $$\label{equation 3.2}
H^{1}_{f,S_{l}}(G_{L}; {\tilde T}_{l,\, i}) = ker(\, H^{1}(G_{L}; {\tilde T}_{l,\, i})\rightarrow \prod_{w\notin S_{l}} H^1(G_{w}; {\tilde T}_{l,\, i})/H^1_{f}(G_{w}; {\tilde T}_{l,\, i}) \, )\, .$$ We define the intermediate Jacobian (see [@bgk05], [@bgk03]): $$\label{equation 3.3}
J_{f,S_{l}} ({\tilde T}_{l,\, i}) = \lim_ {{\rightarrow}\atop {L/F}}H^{1}_{f,S_{l}}(G_{L}; {\tilde T}_{l,\, i})\, .$$ In [@bgk05], [@bgk03], in the more general situation of any free ${\mathbb Z}_{l}$-module of finite rank $T_{l},$ we made the following:
\[assumption 3.2\] For any finite extension $L/F$ and any place $w\in {\mathcal O}_L$ such that $w\notin S_{l}\, ,$ we have $T_{l}^{Fr_{w}}=0.$
\[remm\] We have $H^{1}_{f,S_{l}}(G_{L}; {\tilde T}_{l,\, i}) \cong H^{1}(G_{L,S_{l}}; {\tilde T}_{l,\, i})$ and the assumption \[assumption 3.2\] is satisfied (cf. [@bgk05], p.5 ).
By [@bgk03 Prop. 2.14] we have the following isomorphisms $$J_{f,S_{l}}({\tilde T}_{l,\, i})_{l} \,\, \cong \,\,
{\tilde V}_{l,\, i} / {\tilde T}_{l,\, i} \,\, \cong \,\,
V_{l}(A_i) / T_{l}(A_i) \, (n) \,\, \cong \,\, A_i [l^{\infty}] (n) \, .
\label{torsion of interm Jacobian}$$
Let ${\tilde T}_{l} := \bigoplus_{i=1}^t \, {\tilde T}_{l,\, i}.$ Since each ${\tilde T}_{l,\, i}$ is a free ${\mathbb Z}_l (n)$-module we have the following natural isomorphism: $$J_{f,S_{l}} ({\tilde T}_{l}) \cong \bigoplus_{i=1}^t \, J_{f,S_{l}} ({\tilde T}_{l,\, i})\, .
\label{intermediate jacobian decomposition}$$
For a finite extension $L/F$ the following map corresponding to these used in the direct systems (\[equation 3.3\]) $$\label{i*}
i^{*}: H^{1}(G_{F, S_{l}}, {T}_{l}(J)(n)) \rightarrow H^{1}(G_{L, S_{l}}, {T}_{l}(J)(n))$$ is an injection. One readily verifies, using transfer in the Galois cohomology, that the map (\[i\*\]) has $l$-torsion kernel.. But the $l$-torsion part of (\[i\*\]) is just $$\label{lpart}
H^{0}(G_{F, S_{l}}, {V}_{l}(J)/{T}_{l}(J) \, (n)) \rightarrow H^{0}(G_{L, S_{l}}, {V}_{l}(J) / {T}_{l}(J) \, (n)),$$ which is clearly injective. Similarly, the reduction of the map (\[i\*\])
$$H^{1}({g}_{v}, {{T}_{l}(J_{v})}(n )) \rightarrow H^{1}({g}_{w}, {T}_{l}(J_{v})(n ))$$
is injective for any finite extension $L/F$ and any prime $w$ of $\mathcal{O}_L$ over $v,\, v\notin S_l.$ Notice that by Definition \[definition of an isogeny splitting field\] we have $$H^{1}(G_{F, S_{l}}; {T}_{l}(J)(n)) \cong
H^{1}(G_{F, S_{l}}; {T}_{l}(A) (n)) = \prod_{i=1}^{t}
H^{1}(G_{F, S_{l}}; {T}_{l}(A_i)(n))^{e_i}.
\label{The Zl module}$$ Let us also recall from the following
\[inj\]( [@bgk03], Lemma 2.13) For any finite extension $L/F$ and any prime $w \notin S_l$ in ${\cal O}_{L}$ the natural map $$r_{w}\,\,:\,\,H^1_{f, S_l}(G_L; T_l)_{tor}\rightarrow H^1(g_w;\, T_l)$$ is an imbedding.
Kummer theory
=============
Kummer theory in the context of abelian variety was developed in [@r79]. In this section we collect necessary facts which will be useful in section 7.
\[kummer\] Let $A$ be an abelian variety over a number field $F$. Let ${\cal R}=\End_FA$. For $\alpha\in {\cal R}$ we set $F_{\alpha}=F(A[\alpha]) $ and $ G_{\alpha} = Gal(F_{\alpha}/F)$. The Kummer map $$\label{psi}
{\psi}^{\alpha} : A(F )/{\alpha}A(F)\rightarrow {\Hom}_{G_{\alpha}} (G_{F_{\alpha}}, A[\alpha])$$ is defined as the composition $$A(F)/{\alpha}A(F)\hookrightarrow H^1( F,A[\alpha]) \xrightarrow{\res} H^1( F_{\alpha},A[\alpha])^{G_{\alpha}}$$ with the first map a coboundary map for the $G_F$-cohomology of the Kummer sequence $$\label{ks}
0\rightarrow A [ \alpha ] \rightarrow A ({\overline F} ) \xrightarrow{\alpha} A ({\overline F} ) \rightarrow 0$$ and the second map restriction to $F_{\alpha}$.
Explicitly, the map (\[psi\]) is given by the following formula $$\label{explicit}
{\psi}^{\alpha}({x} )(\sigma)={\sigma}(\frac{x}{\alpha})-\frac{x}{\alpha},$$ where $\frac{x}{\alpha}$ is a fixed “$\alpha$-root” of $x$ i.e. an element $y\in A({\overline F})$ such that ${\alpha}y=x.$
We are interested in Kummer maps where $l$ is a rational prime and ${\alpha}=l^k\in {\mathbb Z}\hookrightarrow {\cal R}.$ Thus we have the the family of Kummer maps: $$\label{ladic}
{\psi}^{l^k} : A(F )/{l^k}A(F)\rightarrow {\Hom}_{G_{l^k}} (G_{F_{l^k}}, A[l^k]).$$ The maps (\[ladic\]) are compatible with the natural maps induced by multiplication by $l$. Therefore taking the inverse limit of both sides and twisting it with ${\mathbb Z}_l(n)$ yields the map $$\label{ladic1}
A(F)\otimes {\mathbb Z}_l (n) \rightarrow {\Hom}(G_{F_{l^{\infty}}}, T_{l}(A)(n)),$$ where $F_{l^{\infty}}={\bigcup_k}F_{l^k}$ and $G_{F_{l^{\infty}}}=Gal({\overline F}/F_{l^{\infty}})$.
Proof of Theorem \[best\]
=========================
We have the following commutative diagram involving the Mordell-Weil groups of an abelian variety and the reduction maps $$\xymatrixcolsep{5pc}\xymatrix{
A(F){\otimes}{\mathbb Z}_l (n)\ar[d]_{{\mathrm{r}}_v} \ar[r]^{\quad{\psi}_{F,l}\otimes{\mathbb Z}_{l}(n)\qquad} & H^{1}(G_{F,S_{l}}; {T}_{l}(J)(n))\ar[d]^{{\mathrm{r}}_v} \\
A(k_v) {\otimes}{\mathbb Z}_l(n) \ar[r] & H^{1}({g}_{v}; {T}_{l}(J_{v})(n ))\, ,
}$$ where the map ${\psi}_{F,l}\otimes {\mathbb Z}_{l}(n)$ is a natural imbedding which comes from the Kummer map (\[ladic1\]) ( see [@bgk03] discussion on p.148 ). Our proof is a generalization of the counterexample to local - global principle for abelian varieties constructed by P. Jossen and A.Perucca in [@jp10] and later extended to the context of $t$-modules in [@bok18]. Because of (\[The Zl module\]) and the choice of $F$ we may assume that $A=A_1^{e_1}$ where $A_1$ is a geometrically simple abelian variety. So, let $e_1=d_1+1$ where ${d_1=\dim\tilde V}_{l,\, 1}.$ Let $P_{1},\dots ,P_{e}\in H^{1}(G_{F,S_{l}}; {T}_{l}(A)(n))=H^{1}(G_{F,S_{l}}; {T}_{l}(A_1)(n))^{e_1}$ be points linearly independent over ${\mathbb Z}_{l}(n)$ which are in the image of $A(F)\otimes {\mathbb Z} \hookrightarrow A(F)\otimes {\mathbb Z}_l(n) .$ Let $$\begin{aligned}
\label{cex}
P &:= \begin{bmatrix}
P_{1} \\
P_{2} \\
\vdots \\
P_{e}
\end{bmatrix},
\quad
{\Lambda}:= \{ MP \,: \quad M\in {\mathrm{Mat}}_{e\times e}({\mathbb Z}_{l}(n)), \quad {\tr}M=0 \}.
\end{aligned}$$ Let $v\in {\cal O_{F,S_{l}}}$ be a prime of good reduction for $A_1$ and $\overline{P}=[\overline{P}_{1}\,,\dots ,\overline{P}_{e}]^{T}$ , $\overline{P}_{i}\in H^{1}({g}_{v}; {T}_{l}(A_{v})(n ))$ be the reduction ${\mod}\,\, {v}$ of $P.$ Notice that $P\notin {\Lambda}$ since $P_{1},\dots ,P_{e}$ are ${\mathbb Z}_{l}(n)$-linearly independent. We will find a matrix $M\in {\mathrm{Mat}}_{e\times e}({\mathbb Z}_{l}(n))$ such that $\overline{P}=M{\overline{P}}.$ This will show that ${\mathrm{red}}_{v}P\in {\mathrm{red}}_{v}{\Lambda}.$ Since the group $H^{1}({g}_{v}; {T}_{l}(A_{v})(n ))$ is finite there exist ${\alpha}_{1},\dots , {\alpha}_{e}\in {\mathbb Z}$ with ${\alpha}_i$ minimal such that $$\begin{aligned}
{\alpha}_{1}{\overline P}_{1} + m_{1,2}{\overline P}_2 + \dots + m_{1,e}{\overline P}_{e} &= 0 \\
{m_{2,1}}{\overline P}_{1} + {\alpha}_{2}{\overline P}_2 + \dots + m_{2,e}{\overline P}_{e} &= 0 \\
\dots\dots\dots\dots\dots\dots\dots\dots\dots & \\
{m}_{e,1}{\overline P}_{1} + m_{e,2}{\overline P}_2 + \dots + {\alpha}_{e}{\overline P}_{e} &= 0 .\end{aligned}$$ We will show that $D={\gcd ({{\alpha}_1, \dots {\alpha}_e}})=1.$ Assume opposite. Choose a rational prime $p$ that divides $D.$ This means, by our choice of ${\alpha}_{1},\dots , {\alpha}_e,$ that $p$ divides coefficients of any linear combination of points ${\bar P}_{1},\dots, {\bar P}_e \in H^{1}({g}_{v}; {T}_{l}(A_{v})(n )).$ In particular $p$ divides the orders of ${\bar P}_i , i=1,\dots e.$ The $p$-torsion of $A_v(k_v)$ is generated by at most $d$-elements. Therefore the group $Y=( {\bar P}_{1},\dots, {\bar P}_e )\cap H^{1}({g}_{v}; {T}_{l}(A_{v})(n ))$ is generated by fewer than $e$ elements. We may assume without loss of generality that $Y=( {\bar P}_{2},\dots, {\bar P}_e )\cap H^{1}({g}_{v}; {T}_{l}(A_{v})(n ))$ Let ${\alpha}_1=x_1p$ and $$\label{cexd1}
{\alpha}_{1}{\bar P}_{1}+ {x}_{2}p{\bar P}_{2}+\dots +{x}_{e}p{\bar{P_{e}}}=0$$ be a linear relation. Therefore $${x}_{1}{ P}_{1}+ {x}_{2}{P}_{2}+\dots +{x}_{e}{{P_{e}}}=T.$$ But $T\in A(k_v)_p$ is generated by ${\bar P}_2,\dots {\bar P}_n$. Thus we obtain a contradiction with the minimality of ${\alpha}_{1}.$
Hence there exist $a_{1},\dots ,a_{e}\in {\mathbb Z}$ such that $$\label{cexd2}
e=a_{1}{\alpha}_{1}+\dots +a_{e}{\alpha}_{e}.$$ Put $m_{i,i}=1-a_{i}{\alpha}_{i }.$ Then $m_{1,1}+\dots +m_{e,e} =0$ and $$\label{fa1}
\begin{bmatrix}
m_{1,1} & \dots & m_{1,e}\\
\dots & \dots & \dots \\
m_{e,1} & \dots & m_{e,e}
\end{bmatrix}
\begin{bmatrix}
{\bar P}_{1}\\
\dots \\
{\bar P}_{e}
\end{bmatrix}
=
\begin{bmatrix}
{\bar P}_{1}\\
\dots \\
{\bar P}_{e}
\end{bmatrix} .
$$
Therefore ${\bar P}\in {\bar{\Lambda}}.$ We view the matrix M in (\[fa1\]) as the matrix with the ${\mathbb Z}_l(n)$ coefficients via the obvious map ${\mathbb Z}\rightarrow {\mathbb Z}\otimes {\mathbb Z}_l(n),\quad m\rightarrow m\otimes t^n,$ where $t^n$ is the generator of $Z_l(n)$.
Local to global principle and dynamical systems
===============================================
S. Bara[ń]{}czuk in [@b17] considers abelian groups satisfying the following two axioms.
\[ass1\] Let $B$ be an abelian group such that there are homomorphisms $r_{v}: B\rightarrow B_{v}$ for an infinite family $v$, whose targets $B_v$ are finite abelian groups.
1. Let $l$ be a prime number, $(k_{1},\dots , k_m)$ a sequence of nonnegative integers. If $P_1,\dots , P_m \in B$ are points linearly independent over ${\mathbb Z},$ then there is a family of primes $v$ in $F$ such that $l^{k_{i}}|| {\ord}_{v}P_{i}$ if $k_{i}>0$ and $l\nmid {\ord}_{v}P_i$ if $k_i=0.$
2. For almost all $v$ the map $B_{\text{\rm tors}}\rightarrow B_{v}$ is injective.
\[rmk\] Notice assumption (1) describes the assertion of the Reduction Theorem of [@bk11], whereas assumption (2) is one of the assumptions in [@bgk03].
Here ${\ord}_{v}P$ is the order of a reduced point $P\mod v$. Under the Assumptions \[ass1\], S.Bara[ń]{}czuk was able to prove the following dynamical version of the local lo global principle
\[b17\]( [@b17]) Let $\Lambda$ be a subgroup of $B$ and $P\in B$ be a point of infinite order and $\phi$ be a natural number. Then the following are equivalent:
1. For almost every $v$ $$O_{\phi}(P \mod v) \cap ({\Lambda} \mod v)\neq\emptyset ,$$
2. $O_{\phi}(P) \cap ({\Lambda})\neq\emptyset$.
Here $O_{\phi}(P)=\{ {\phi}^n(P) : n\geq 0\}.$ We have the following lemma
\[kg\] Let $A=A_{1}^{e_1}\times \dots \times A_{t}^{e_t}, \,\, i=1,\dots ,t$ be an abelian variety such that ${\End}A_i={\mathbb Z} $ for $i=1,\dots ,t.$ Let $B=H^{1}(G_{F, S_{l}}; {T}_{l}(A) (n)) = \prod_{i=1}^{t}
H^{1}(G_{F, S_{l}}; {T}_{l}(A_i)(n))^{e_i}.$ and $B_v=H^{1}({g}_{v}; {T}_{l}(J_{v})(n ))$ for $v\in {\cal O}_v$ fulfil Assumption \[ass1\].
Assumption (1) of \[ass1\] is a specialization to ${\cal R}_i={\mathbb Z}$ of Corollary 3.5 of [[@bk13]]{}. Assumption (2) is an assertion of Lemma \[inj\].
\[theorem 5.1\] Let $X/F$ be a smooth, proper and geometrically irreducible curve of genus $g.$ Let $F^{\prime}$ be an isogeny splitting field of the Jacobian $J$ and assume that for the corresponding product $A_{1}^{e_{1}}\times \dots \times A_{t}^{e_{t}},$ we have $End_{{\overline F}} \, A_{i}={\mathbb Z}$ for each $1 \leq i \leq t.$ Let $l > 2$ be a prime number which is coprime to the polarisation degrees of the abelian varieties $A_i.$ Let $S_{l}$ be a set of places of $F$ containing the places of bad reduction, the archimedean places and the primes above $l.$ Let ${\cal X}$ be a regular and proper model of $X$ over ${\cal O}_{F, S_l}.$ Let $\hat{P} \in K_{2n}^{et}({\cal X})$ and let $\hat{\Lambda}$ be a finitely generated ${\mathbb Z}_{l}$-submodule of $ K_{2n}^{et}({\cal X}).$ Let $O_{w}= \{ w^{n}(P): n\geq 0 \}$ Then the following are equivalent
1. For almost every prime ${v} \in {\cal O}_{F,S_l}$ $$O_{w}({\mathrm{r}}_{v}(P))\cap {\mathrm{r}}_{v}({\Lambda}) \neq \emptyset ,$$
2. $O_{w}(P)\cap {\Lambda}\neq\emptyset$.
The proof follows from Lemma \[kg\] , commutativity of the diagram (\[diagram 2.4\]) and finiteness of the kernel of the left vertical map in this diagram.
\[finrem\] If the Bass Conjecture and Quillen-Lichtenbaum Conjecture hold true for ${\cal X}$ then we obtain ( using diagram (\[diagram 2.4a\]) instead of (\[diagram 2.4\]) ) the corresponding to Theorem \[theorem 5.1\] statement for Quillen $K$-theory of ${\cal X}.$
<span style="font-variant:small-caps;">G. Banaszak, W. Gajda, and P. Kraso[' n]{}</span>, [*On Galois cohomology of some $p$-adic representations and [' e]{}tale $K$-theory of curves,*]{} Contemporary Math. AMS **241** (1999), 23–44.
<span style="font-variant:small-caps;">G. Banaszak, W. Gajda, P. Kraso[' n]{}</span>, [*On reduction map for ' etale $K$[-]{}theory of curves,*]{} Homology, Homotopy and Applications **7** (3), Proceedings of Victor’s Snaith 60-th Birthday Conference, (2005), 1–10.
<span style="font-variant:small-caps;">G. Banaszak, W. Gajda, P. Kraso[' n]{}</span>, [*Support problem for the intermediate Jacobians of $l$-adic representations,*]{} Journal of Number Theory **100** (1) (2003), 133–168.
<span style="font-variant:small-caps;">G. Banaszak, P. Krasoń,</span> [*On a local to global principle in [é]{}tale K-groups of curves,*]{} Journal of K-theory and its Applications to Algebra Geometry and Topology,12, (2013), pp.183-201.
<span style="font-variant:small-caps;">S. Bara[ń]{}czuk ,</span>[*On a dynamical local-global principle in Mordell-Weil type groups,*]{} Expo. Math. 35, No. 2, 206-211 (2017).
<span style="font-variant:small-caps;">S. Bara[ń]{}czuk, K.G[ó]{}rnisiewicz</span> ,[*On reduction maps for the [é]{}tale and Quillen K-theory of curves and applications,*]{} J. K-theory Vol. 2, Issue 1, 2008 , pp. 103-122.
<span style="font-variant:small-caps;">W. Bondarewicz, P.Kraso[ń]{}</span> ,[*On a reduction map for Drinfeld modules,*]{} preprint 2018, arXiv:1811.05631.
<span style="font-variant:small-caps;">A.Brumer, A.Pacetti, C.Poor, G.Tornaria, J.Voight, D.S.Yuen,</span> [*On paramodularity of typical abelian surfaces*]{}, arXiv: 1805.10873v.2, Aug. 2018.
<span style="font-variant:small-caps;">D.A Cox</span>, [*Primes of the form $x^2 +ny^2$*]{}, John Wiley and sons 1989.
<span style="font-variant:small-caps;">C-L, Chai, F.Oort</span>, [*Abelian varieties isogenous to a Jacobian*]{}, Annals of Mathematics **176** (2012), 589-635.
<span style="font-variant:small-caps;">W. Dwyer, E. Friedlander</span>, *Algebraic and ' etale K-theory*, Trans. Amer. Math. Soc. **292** (1985), 247-280.
<span style="font-variant:small-caps;">T, Ekedahl, J-P. Serre</span>, [*Examples of algebraic curves with totally split Jacobian*]{}, C. R. Acad. Sci., Paris, Sér. I 317, No. 5, 509-513 (1993).
<span style="font-variant:small-caps;">Y.Flicker, P.Kraso[ń]{}</span>, [*Multiplicative relations of points on algebraic groups.* ]{}, Bull. Pol. Acad. Sci., Math. 65, No. 2, 125-138 (2017).
<span style="font-variant:small-caps;">T. Hayashida, M. Nishi,</span> [*Existence of curves of genus two on a product of two elliptic curves*]{}, J. Math. Soc. Japan, vol. 17, no.1, (1965) pp. 1-16.
<span style="font-variant:small-caps;">U. Jannsen</span>, [*Continuous [é]{}tale cohomology,*]{} Math. Ann. 280 (1988), 207–245.
<span style="font-variant:small-caps;">P. Jossen, A. Perucca,</span> [*A counterexample to the local-global principle of linear dependence for abelian varieties.*]{}, C. R., Math., Acad. Sci. Paris 348, No. 1-2, pp. 9-10 (2010).
<span style="font-variant:small-caps;">F.Oort, K.Ueno</span> [*Principally polarized abelian varieties of dimension two or three are Jacobian varieties*]{}, Journ. Fac. Sc. Univ. Tokyo, Sec. IA 20 (1973), 377 - 381.
<span style="font-variant:small-caps;">K. A. Ribet,</span> [*Kummer theory on extensions of abelian varieties by tori*]{}, Duke Math. J. 46 (1979), pp. 745-761.
<span style="font-variant:small-caps;">J. Tate,</span> *p-divisible groups,* Proceedings of the Conference on local Fields, Springer-Verlag,1968.
<span style="font-variant:small-caps;">A. Weil,</span> [*Zum Beweis des Torellischen Satzes.*]{} Nachr. Akad. G[ö]{}ttingen, Math.-Phys. Kl. 1957, 33 – 53.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $K$ be a complete discretely valued field with residue field $\kappa$. If char$(K) = 0$, char$(\kappa) = 2$ and $[\kappa : \kappa^2] = d$, we prove that there exists an integer $N$ depending on $d$ such that the $u$-invariant of any function field in one variable over $K$ is bounded by $N$. The method of proof is via introducing the notion of uniform boundedness for the $p$-torsion of the Brauer group of a field and relating the uniform boundedness of the 2-torsion of the Brauer group to finiteness of the $u$-invariant. We prove that the 2-torsion of the Brauer group of function fields in one variable over $K$ are uniformly bounded.'
address: |
Department of Mathematics & Computer Science\
Emory University\
400 Dowman Drive NE\
Atlanta, GA 30322, USA
author:
- 'R. Parimala'
- 'V. Suresh'
title: ' On the $u$-invariant of function fields of curves over complete discretely valued fields'
---
Let $K$ be a complete discretely valued field with residue field $\kappa$ and $F$ a function field in one variable over $K$. Suppose char$(\kappa) \neq 2$. A bound for the $u$-invariant of $F$ in terms of the $u$-invariant of function fields in one variable over $\kappa$ were obtained by Harbater-Hartmann-Krashen [@hhk:patching_qf_csa] using patching techniques. This recovers the $u$-invariant of function fields of non-dyadic $p$-adic curves ([@ps:uinv_8]). Leep ([@leep]), using results of Heath-Brown ([@heath-brown]), proved that the $u$-invariant of function fields of all $p$-adic curves (including dyadic curves) is 8. An alternate proof for function fields of dyadic curves is given in ([@ps:period_index]). In fact more generally we proved that if char$(K) = 0$, char$(\kappa) =2$ and $\kappa$ is perfect, then $u(F) \leq 8$. If $[\kappa : \kappa^2]$ is infinite it is easy to construct anisotropic quadratic forms over $K$ and hence over $F$ of arbitrarily large dimension. The question remained open whether the $u$-invariant of $F$ is finite if char$(\kappa) = 2$ and $[\kappa : \kappa^2]$ is finite. The aim of this article is to give an affirmative answer to this question. More precisely we prove the following (\[theorem:u\_invariant\])
Let $K$ be complete discretely valued field with residue field $\kappa$ and $F$ a function field of a curve over $K$. Suppose that char$(K) = 0$, char$(\kappa) = 2$ and $[\kappa : \kappa^2]$ is finite. Then there exists an integer $M$ which depends only on $[\kappa :
\kappa^2]$ such that for any finite extension $F$ of $K(t)$, $u(F) \leq M$.
It was conjectured in ([@ps:period_index]) that $u(F)$ is at most $8[\kappa: \kappa^2]$. The bound we give for the $u(F)$ is not effective and far from the conjectural bound.
Let $L$ be a field of characteristic not equal to 2 with $H^M(L,
\mu_2^{\otimes M}) = 0$ for some $M \geq 1$. Suppose that there exists an integer $N$ such that for all finite extensions $E$ of $L$ and for any $\alpha \in H^n(E, \mu_2^{\otimes n}), n \geq 2$, there exists an extension $E'$ of $E$ of degree at most $N$ such that $\alpha \otimes_E E' = 0$. Then a theorem of Krashen (\[cor:krashen2\]) asserts that the $u$-invariant of $L$ is finite. Our aim is to prove that if $K$ is a complete discretely valued field with residue field $\kappa$ of characteristic 2 and $[\kappa : \kappa^2]$ finite and $F$ a function field in one variable over $K$, then such an integer $N$ exists for $F$, thereby proving the finiteness of the $u$-invariant of $F$.
We introduce the notion of uniform boundedness for the $\ell$-torsion of the Brauer group $\operatorname{\mathrm{Br}}(L)$ of $L$, where $L$ is any field. We say that the Brauer group of $L$ is [*uniformly $\ell$-bounded*]{} if there exists an integer $N$ such that for any finite extension $E$ of $L$ and for any set of finitely many elements $\alpha_1, \cdots , \alpha_n \in
{{}_\ell\mathrm{Br}}(E)$, there is a finite extension $E'$ of $E$ of degree at most $N$ such that $\alpha_i \otimes_E E' = 0$ for $1 \leq i \leq n$. Using the result of Krashen (\[cor:krashen2\]), we show that if $L$ is a field of characteristic not equal to 2 with $H^M(L, \mu_2^{\otimes M}) = 0$ for some $M \geq 1$ and with the Brauer group of $L$ is $2$-uniformly bounded, then $u(L)$ is finite (\[cor:uniformbound\_uinvariant\]). It looks plausible that there are fields $L$ of finite $u$-invariant with $L$ not uniformly 2-bounded.
The main result of the paper is to prove the uniform $p$-boundedness for the Brauer group of any function field $F$ in one variable over a complete discretely valued field $K$ with residue field $\kappa$, where char$(\kappa) = p$, char$(F) \neq p$ and $[\kappa :
\kappa^p]$ is finite. We also prove the uniform $\ell$-boundedness for the Brauer group of any function field in one variable over a complete discretely valued field with residue field $\kappa$ and $\ell \neq $char$(\kappa)$ under the assumption that the Brauer groups of $\kappa$ and $\kappa(t)$ are uniformly $\ell$-bounded. This result for function fields of $p$-adic curves ($p \neq \ell$) is due to Saltman ([@saltman:jrms]). To prove our theorems we use the patching techniques of Harbater-Hartmann-Krashen and results of ([@ps:period_index]).
We thank D. Harbater for his very useful comments on the text.
Galois cohomology, Symbol length, $u$-invariant and Uniform bound
==================================================================
In this section we recall the recent results of Krashen and Saltman connecting the symbol length and effective index in Galois cohomology with the $u$-invariant of a field.
Let $K$ be a field and $\ell$ a prime not equal to the characteristic of $K$. Let $\mu_\ell$ denote the Galois module of $\ell^{\rm th}$ roots of unity and $H^n(K, \mu_\ell^{\otimes m})$ denote the $n^{\rm
th}$ Galois cohomology group with values in $\mu_\ell^{\otimes m}$. We have $H^1(K, \mu_\ell) \simeq K^*/K^{*\ell}$. For $a \in K^*$, let $(a) \in H^1(K, \mu_\ell)$ denote the image of $aK^{*\ell}$. Let $a_1, \cdots , a_n \in K^*$. The cup product $(a_1) \cdot (a_2) \cdots (a_n) \in H^n(K, \mu_\ell^{\otimes n})$ is called a [*symbol*]{} in $H^n(K, \mu_\ell^{\otimes n})$. A theorem of Voevodsky ([@voevodsky]) asserts that every element in $H^n(K,
\mu_\ell^{\otimes n})$ is a sum of symbols.
Let $\alpha \in H^n(K,
\mu_\ell^{\otimes n})$. The [*symbol length*]{} of $\alpha$, denoted by $\lambda(\alpha)$, is defined as the smallest $m$ such that $\alpha $ is a sum of $m$ symbols in $H^n(K, \mu_\ell^{\otimes n})$. For any $\alpha \in H^n(K, \mu_\ell^{\otimes m})$, the [ *effective index*]{} of $\alpha$, denoted by $eind(\alpha)$, is defined to be the minimum of the degrees of finite field extensions $E$ of $K$ with $\alpha_E = 0$, where $\alpha_E$ is the image of $\alpha$ in $H^n(E, \mu_\ell^{\otimes m})$. Since $H^2(K, \mu_\ell) \simeq
{{}_\ell\mathrm{Br}}(K)$, for $\alpha \in {{}_\ell\mathrm{Br}}(K)$, eind$(\alpha)$ is equal to the index of a central simple algebra $A$ over $K$ representing $\alpha$. The following lemma asserts that this definition of effective index coincides with the definition in ([@krashen:symbol_length]).
\[lemma:sep\_ext\] Let $K$ be a field and $\ell$ a prime not equal to char$(K)$. Let $\alpha \in H^n(K, \mu_\ell^{\otimes
m})$. Suppose that exists an extension $L$ of $K$ of degree at most $N$ with $\alpha \otimes_K L = 0$. Then there exists a separable field extension $E$ of $K$ of degree at most $N$ such that $\alpha
\otimes_K E = 0$.
Let $L$ be an extension of $K$ of degree at most $N$ with $\alpha \otimes_KL = 0$. Let $E$ be the separable closure of $K$ in $L$. Let $p $ be the characteristic of $(K)$. Suppose $p > 0$. Then $L/E$ is of degree $p^r$ for some $r \geq 0$. Since $\ell \neq
p$, the restriction map $H^n(E, \mu_\ell^{\otimes m}) \to H^n(L,
\mu_\ell^{\otimes m})$ is injective ([@serre:gc Cor. on p.12]). Hence $\alpha \otimes_K E = 0$.
We know recall a theorem of Krashen ([@krashen:symbol_length 4.2], cf. [@saltman:symbol_length]).
\[theorem:krashen1\] Let $K$ be a field and $\ell$ a prime not equal to the characteristic of $K$. Let $n \geq 1$. Suppose that there exists an integer $N$ such that for every finite extension $L$ of $K$ and for every element $\beta \in H^d(L, \mu_\ell^{\otimes d})$, $ 1
\leq d \leq n-1$, eind$(\beta) \leq N$. Then for any $\alpha \in H^n(K,
\mu_\ell^{\otimes n})$, $\lambda(\alpha)$ is bounded in term of eind$(\alpha)$, $N$ and $n$.
The following is a consequence of the above theorem.
\[cor:krashen1\] Let $K$ be a field and $\ell$ a prime not equal to the characteristic of $K$. Let $n\geq 1$. Suppose that there exists an integer $N$ such that for all finite extensions $L$ of $K$ and for all $\alpha \in H^d(L, \mu_\ell^{\otimes d})$, $1\leq d \leq
n$, eind$(\alpha) \leq N$. Then there exists an integer $N'$ which depends only $N$ and $n$ such that $\lambda(\alpha) \leq N'$ for all finite extensions $L$ of $K$ and $\alpha
\in H^n(K, \mu_\ell^{\otimes n})$.
Let $K$ be a field of characteristic not equal to 2. The [*$u$-invariant*]{} of $K$ is defined to be the supremum of dimensions of anisotropic quadratic forms over $K$. The following theorem is a consequence of a theorem of Orlov, Vishik, and Voevodsky ([@ovv]) on the Milnor conjecture (cf. [@kahn], [@ps:symbol_length]).
\[theorem:symbol\_length\_u\_invariant\] Let $K$ be a field of characteristic not equal to 2. Suppose that there exist integer $M\geq 1 $ and $N$ such that $H^M(K,
\mu_2) =0$ and $\lambda(\alpha) \leq N$ for all $\alpha \in H^d(K,
\mu_2^{\otimes d})$, $ 1 \leq d < M$. Then the $u$-invariant is bounded by a function of $M$ and $N$.
The following follows from (\[theorem:symbol\_length\_u\_invariant\]) and (\[cor:krashen1\]) (cf. [@krashen:symbol_length 5.5]).
\[cor:krashen2\] Let $K$ be a field of characteristic not equal to 2. Suppose that there exist integers $N \geq 1$ and $M \geq
1$ such that for all finite extensions $L$ of $K$, $H^{M}(L,
\mu_2) = 0$ and for all $ n \geq 1$ and $\alpha \in H^n(L, \mu_2)$, eind$(\alpha) \leq N$. Then there exists an integer $N'$, which depends only on $N$ and $M$, such that $u(K) \leq N'$.
Let $K$ be a field and $\ell$ a prime not equal to the characteristic of $K$. We say that $K$ is $(n , \ell)$-[*uniformly bounded* ]{} if there exists an integer $N$ such that for any finite extension $L$ of $K$ and $\alpha_1, \cdots , \alpha_m
\in H^n(L, \mu_\ell^{\otimes n})$ there is an extension $E$ of $L$ with $[E : L] \leq N$ and $\alpha_{i} \otimes_L E = 0$ for $1 \leq i \leq m$. Such an $N$ is called an $(n, \ell)$-[*uniform bound*]{} for $K$. We note that if $N$ is an $(n, \ell)$-uniform boud for $K$, then $N$ is also an $(N, \ell)$-uniform bound for any finite extension $L$ of $K$.
In view of a theorem of Voevodsky ([@voevodsky]) on the Bloch-Kato conjecture, every element in $H^n(K, \mu_\ell^{\otimes n})$ is a sum of symbols. In particular $N$ is an $(n, \ell)$-uniform bound of $K$ if and only if for given symbols $\alpha_1, \cdots , \alpha_m \in H^n(K, \mu_\ell^{\otimes
n})$ there is an extension $L$ of $K$ with $[L : K] \leq N$ and $\alpha_{i_L} = 0$ for $1 \leq i \leq m$.
\[lemma:uniform\_bound\_d\_d+1\] Let $K$ be a field and $\ell$ a prime not equal to the characteristic of $K$. If $N$ is an $(n,\ell)$-uniform bound for $K $, then $N$ is also a $(d,
\ell)$-uniform bound for $K $ for all $d \geq n$.
Suppose $N$ is an $(n, \ell)$-uniform bound for $ K$. It is enough to prove the lemma for $d = n+1$. Let $L$ be a finite extension of $K$ and $\alpha_1, \cdots , \alpha_m
\in H^{n+1} (L, \mu_\ell^{\otimes (n+1)})$ be symbols. Then $\alpha_i =
\beta_i \cdot (a_i)$ for some symbols $\beta_i \in H^{n}(L,
\mu_{\ell}^{\otimes n})$ and $a_i \in L^*$, $1 \leq i \leq m$. Since $N$ is an $(n, \ell)$-uniform bound for $K$, there exists a field extension $E$ of $L$ with $[E : L] \leq N$ and $\beta_ {i_E} = 0$ for $1 \leq i \leq
m$. Then clearly $\alpha_{i_E} = 0$ for $1 \leq i \leq m$. Thus $N$ is also an $(n+1, \ell)$-uniform bound for $K$.
\[cor:symbol\_length\] Let $K$ be a field and $\ell$ a prime not equal to the characteristic of $K$. Suppose that $N$ is a $(2, \ell)$-uniform bound for $K$. Then for every $n \geq 2$, there exists an integer $N_n$, which depends only on $N$ and $n$ such that $\lambda(\alpha) \leq N_n$ for all $\alpha \in H^n(K,
\mu_\ell^{\otimes n})$.
Since $N$ is $(2, \ell)$-uniform bound for $K$, by ( \[lemma:uniform\_bound\_d\_d+1\]), $N$ is an $(n, \ell)$-uniform bound for $K$ for all $n \geq 2$. Let $\alpha \in H^n(K, \mu_\ell^{\otimes
n})$. Then, by (\[theorem:krashen1\]), $\lambda(\alpha)$ is bounded in terms of eind$(\alpha)$, $N$ and $n$. Since eind$(\alpha) \leq N$, $\lambda(\alpha)$ is bounded in terms of $N$ and $n$.
\[cor:uniformbound\_uinvariant\] Let $K$ be a field of characteristic not equal to 2. Suppose that there exists an integer $M$ such that for all finite extensions $L$ of $K$, $H^M(K,
\mu_2) = 0$ and $N$ is a $(2,2)$-uniform bound for $K$. Then there exists $N'$ which depends only on $N$ and $M$ such that for any finite extension $L$ of $K$, $u(L) \leq N'$.
Since the conditions on $K$ are also satisfied by any finite extensions of $K$, it is enough to find an $N'$ which depends only on $N$ and $M$ such that $u(K) \leq N'$.
Since $N$ is a $(2,2)$-uniform bound for $K$, by (\[cor:symbol\_length\]) there exist integers $N_n$ for $1 \leq n
< M$, which depends only on $N$ and $n$ such that for all $\alpha \in H^n(K, \mu_\ell^{\otimes n})$, $\lambda(\alpha)
\leq N$. Let $N'$ be the maximum of $N_n$ for $ 1 \leq n < M$. Thus corollary follows from (\[theorem:krashen1\]).
Let $R$ be a discrete valuation ring with field of fractions $K$ and residue field $\kappa$. Let $\ell$ be a prime not equal to char$(\kappa)$. Then there is a residue homomorphism $\partial : H^n(K, \mu_\ell^{\otimes m} ) \to H^{n-1}(\kappa,
\mu_\ell^{\otimes (m-1)})$ with kernel $H^n_{{\mathrm{\acute{e}t}}}(R,\mu_\ell^{\otimes
m})$.
Let $A$ be an integral domain with field of fractions $K$. Let $\ell$ be a prime which is a unit in $A$. We have the natural map $H^n_{\mathrm{\acute{e}t}}(A, \mu_\ell^{\otimes m}) \buildrel{\iota}\over{\to} H^n(K,
\mu_\ell^{\otimes m})$. An element $\alpha$ of $H^n(K, \mu_\ell^{\otimes m})$ is said to be [ *unramified*]{} on $A$ if $\alpha$ is in the image of $\iota$. Suppose that $A$ is a regular ring. For each height one prime ideal $P$ of $A$, we have the residue homomorphism $\partial_P : H^n(K, \mu_\ell^{\otimes m}) \to
H^{n-1}(\kappa(P), \mu_\ell^{\otimes (m-1)})$, where $\kappa(P)$ is the residue field at $P$. We have the following
([@auslander-goldman 7.4]) \[auslander-goldman\] Let $A$ be a regular two dimensional integral domain with field of fractions $K$ and $\ell$ a prime which is a unit in $A$. The sequence $$0 \to H^2_{\mathrm{\acute{e}t}}(A, \mu_\ell) \to H^2(K, \mu_\ell) \to \oplus_{P \in
Spec(A)^{(1)}} H^1(\kappa(P), {\mathbb Z}/ \ell{\mathbb Z})$$ is exact, where Spec$(A)^{(1)}$ is the set of height one prime ideals of $A$.
We now recall a few notation from ([@hhk:patching_qf_csa]). Let $R$ be a complete discrete valuation ring with field of fractions $K$ and residue field $\kappa$. Let $F$ be the function field of a curve over $K$. Let ${{\mathscr{X}}}$ be a regular proper model of $F$ over $R$ and $X$ its reduced special fiber. For any codimension one point $\eta$ of ${{\mathscr{X}}}$, let $F_\eta$ be the completion of $F$ at the discrete valuation of $F$ given by $\eta$ and $\kappa(\eta)$ the residue field at $\eta$. For a closed point $P$ of ${{\mathscr{X}}}$, let $F_P$ be the field of fractions of the completion of the local ring at $P$ and $\kappa(P)$ the residue field at $P$. Let $U$ be an open subset of $X$. Let $R_U$ be the ring of all those functions in $F$ which are regular on $U$. Then $R
\subset R_U$. Let $t$ be a parameter in $R$. Let $\hat{R}_U$ be the completion of $R_U$ at the ideal $(t)$. Let $F_U$ be the field of fractions of $\hat{R}_U$.
Let $A$ be a regular integral domain with field of fractions $F$. For a maximal ideal $m$ of $A$, let $\hat{A}_m$ denote the completion of the local ring $A_m$ and $F_m$ the field of fractions of $\hat{A}_m$.
Uniform bound - bad characteristic case
========================================
Let $K$ be a complete discretely valued field with residue field $\kappa$. Let $p = char(\kappa)$. In this section we show that there is a $(2, p)$-uniform bound for $K(t)$ which depends only on $[\kappa : \kappa^p]$.
First we recall the following two results from ([@ps:period_index]).
\[theorem:ps1\] [@ps:period_index 2.4] Let $R$ be a complete discrete valuation ring with field of fractions $K$ and residue field $\kappa$. Suppose that char$(K) = 0$, char$(\kappa)
= p >0$ and $[\kappa : \kappa^p] = p^d$. Let $\pi \in R$ be a parameter and $u_1,\cdots ,u_d \in R^*$ units such that $\kappa =
\kappa^p(\overline{u}_1, \cdots , \overline{u}_d)$, where for any $u \in R$, $\overline{u}$ denotes the image of $u$ in $\kappa$. Suppose that $K$ contains a primitive $p^{\rm th}$ root of unity. Then any $\alpha \in H^2(K, \mu_p) $ splits over $K(\sqrt[p]{\pi}, \sqrt[p^2]{u_1}, \cdots , \sqrt[p^2]{u_{d-1}}, \sqrt[p]{u_d})$. In particular if $d > 0$, $p^{2d}$ is a $(2, p)$-uniform bound for $K$ and if $d = 0$, $p$ is a $(2, p)$-uniform bound for $K$.
\[prop:ps2\] ([@ps:period_index 3.5]) Let $A$ be a complete regular local ring of dimension 2 with field of fractions $F$ and residue field $\kappa$. Suppose that char$(F) =
0$, char$(\kappa) = p > 0$ and $[\kappa : \kappa^p] = p^d$. Let $\pi, \delta \in A$ and $u_1, \cdots , u_d \in A^*$ such that the maximal ideal $m$ of $A$ is generated by $\pi$ and $\delta$, and $\kappa = \kappa^p(\overline{u}_1, \cdots ,
\overline{u}_d)$, where for any $u \in A$, $\overline{u}$ denotes the image of $u$ in $\kappa$. Suppose that $F$ contains a primitive $p^{\rm th}$ root of unity. Then any $\alpha \in H^2(F, \mu_p)$ which is unramified on $A$ except possibly at $(\pi)$ and $(\delta)$ splits over $F(\sqrt[p]{\pi}, \sqrt[p]{\delta}, \sqrt[p^2]{u_1},
\cdots , \sqrt[p^2]{u_d})$.
\[lemma:choice\_of\_f\_g\] Let ${{\mathscr{X}}}$ be a regular integral two dimensional scheme and $F$ its function field. Suppose $C$ and $E$ are regular curves on ${{\mathscr{X}}}$ with normal crossings. Let ${{\mathscr{P}}}$ be a finite set of closed points of ${{\mathscr{X}}}$ with ${{\mathscr{P}}}\subset C \cup E$. Then there exist $f, g \in F^*$ such that the maximal ideal at $P$ is generated by $f$ and $g$ for each $P \in {{\mathscr{P}}}$ and $f$ (resp. $g$) defines $C$ (resp. $E$) at each $P \in {{\mathscr{P}}}\cap C$ (resp. $P \in {{\mathscr{P}}}\cap E$).
Let $R$ be the semi-local ring at all $P \in
{{\mathscr{P}}}$. Since $R$ is a unique factorisation domain, there exist $f_1, g_1 \in R$ such that div$_{Spec(R)}(f_1) = C\mid _ {Spec(R)}$ and div$_{Spec(R)}(g_1) = E\mid _ {Spec(R)}$. For each $P \in {{\mathscr{P}}}$, let $m_P$ be the maximal ideal of the local ring $R_P$ at $P$. Then for each $P \in C\cap E \cap {{\mathscr{P}}}$, $m_P = (f_1, g_1)$. Let $P \in {{\mathscr{P}}}$. Suppose $P \not\in C$. Then $P \in E$. Since $E$ is regular on ${{\mathscr{X}}}$, by the choice of $g_1$, there exists $\theta_P \in m_P$ such that $m_P = (\theta_P, g_1)$. By the Chinese remainder theorem, there exists $\pi_P \in m_P$ such that $\pi_P
\not\in m_Q$ for all $Q \in {{\mathscr{P}}}$, $Q \neq P$ and $\pi_p = \theta_P$ module $m_P^2$. Then $m_P = (\pi_P, g_1)$. Similarly for each $P \not\in E$, choose $\delta_P
\in m_P$ such that $m_P = (f_1, \delta_P)$ and $\delta_P \not\in m_Q$ for all $Q \in {{\mathscr{P}}}$, $Q \neq P$. Let $$f_2 = \prod_{P \in {{\mathscr{P}}}\setminus C} \pi_P, ~~f = f_1f_2$$ and $$g_2 = \prod_{P \in {{\mathscr{P}}}\setminus E} \delta_P, ~~ g = g_1g_2.$$ Then $f_2$ and $g_2$ are units at all $P \in C\cap E$. We claim that $f$ and $g$ have the required properties. Let $P \in {{\mathscr{P}}}$. Suppose $P \in C \cap E$. Then by the choices $f_2$ and $g_2$, they are units at $P$ and $m_P = (f_1, g_1)$. In particular $m_P = (f, g)$ and $f, g$ define $C$ and $E$ respectively at $P$. Suppose that $P \not\in C$. Then $f_1$ and $g_2$ are units at $P$ and $f_2 = \pi_P u_P$ for some unit $u_P$ at $P$. Since $m_P = (\pi_P, g_1)$, we have $m_P = (f, g)$. Since $g_1$ defines $E$ at $P$ and $g_2$ is a unit at $P$, $g$ defines $E$ at $P$. Similarly if $P \not\in E$, then $m_P = (f, g)$ and $f$ defines $C$ at $P$.
\[theorem:bad\_char\] Let $K$ be a complete discretely valued field with residue field $\kappa$. Suppose that char$(K) = 0$ and char$(\kappa) = p > 0$ and $[\kappa : \kappa^p] = p^d$. Assume that $K$ contains a primitive $p^{th}$ root of unity. Then $p^{4d+4}$ is a $(2, p)$-uniform bound for $K(t)$.
Let $F$ be a finite extension of $K(t)$. Let $\alpha_1, \cdots , \alpha_m \in H^2(F,\mu_p)$. Let ${{{\mathscr{X}}}}$ be a regular proper model of $F$ over the ring of integers $R$ of $K$ such that the support of ram$(\alpha_i)$ for all $i$ and the special fiber is contained in $C \cup E$, where $C$ and $E$ are regular curves on ${{\mathscr{X}}}$ having only normal crossings. Let $\eta$ be the generic point of an irreducible component $X_\eta$ of the special fiber of ${{\mathscr{X}}}$. Then $\kappa(\eta)$ is a function field in one variable over $\kappa$. Since $[\kappa : \kappa^p] = p^d$, $[\kappa(\eta) : \kappa(\eta)^p] =
d+1$ ([@bourbaki:algebra; @ll A.V.135, Corollary 3]). Let $\pi_\eta$ be a parameter at $\eta$ and $u_{\eta, 1}, \cdots , u_{\eta, d+1} \in F^*$ be lifts of a $p$-basis of $\kappa(\eta)$. Then, by (\[theorem:ps1\]), $\alpha \otimes F_\eta(\sqrt[p^2]{u_{\eta, 1}},
\sqrt[p^2]{u_{\eta, d}}, \sqrt[p]{u_{\eta, d+1}}, \sqrt[p]{\pi_\eta}) = 0$ for all $i$. Let $f \in F^*$ be chosen such that $\nu_\eta(f) = 1$ for all $\eta$. By the Chinese remainder theorem, choose $u_1, \cdots , u_d \in F^*$ units at each $\eta$ such that $\overline{u}_j =
\overline{u}_{\eta, j} \in
\kappa(\eta)$. Then $\alpha_i \otimes F_\eta(\sqrt[p]{f},
\sqrt[p^2]{u_1}, \cdots , \sqrt[p^2]{u_{d}}, \sqrt[p]{u_{d+1}}) = 0$ for all $i$. By ([@hhk:lgp_torsors 5.8], [@kmrt 1.17]), there exists a non-empty open set $U_\eta$ of the component $X_\eta$ of the special fiber, such that $\alpha \otimes F_{U_\eta}(\sqrt[p]{f},
\sqrt[p^2]{u_1}, \cdots , \sqrt[p^2]{u_{d}}, \sqrt[p]{u_{d+1}})
= 0$ for all $i$.
Let ${{{\mathscr{P}}}}$ be the finite set of closed points of ${{{\mathscr{X}}}}$ which are not in $U_\eta$ for any $\eta$. Let $A$ be the semi-local ring at the points of ${{\mathscr{P}}}$. For $P \in {{\mathscr{P}}}$, let $A_P$ be the local ring at $P$. Since the ramifications of $\alpha_i$ for all $i$ are in normal crossings, for each $P \in {{{\mathscr{P}}}}$, the maximal ideal $m_p$ at $P$ is $(\pi_P, \delta_P)$ for some $\pi_P$ and $\delta_P$ such that $\alpha_i$ is unramified on $A_P$ except possibly at $(\pi_P)$ and $(\delta_P)$. Since the residue field $\kappa(P)$ at $P$ is a finite extension of $\kappa$ and $[\kappa : \kappa^p] = p^d$, $[\kappa(P) :
\kappa(P)^p] = p^d$ ([@bourbaki:algebra; @ll A.V.135, Corollary 3]). Let $v_{P, 1}, \cdots , v_{P, d} \in A_P^*$ be lifts of a $p$-basis of $\kappa(P)$. By the Chinese remainder theorem, choose $h_1, \cdots , h_d \in A^*$ such that $h_i = v_{P, i}$ modulo the maximal ideal at $P$ for all $P
\in {{\mathscr{P}}}$ and $1\leq i
\leq d$. By (\[lemma:choice\_of\_f\_g\]), there exist $g_1, g_2 \in F^*$ such that for any point $P \in {{\mathscr{P}}}$, we have $m_P =
(g_1, g_2)$ and $g_1$ defines $C$ at all $P \in {{\mathscr{P}}}\cap C$ and $g_2$ defines $E$ at all $P \in {{\mathscr{P}}}\cap E$. In particular, each $\alpha_i$ is unramified on $A_P$ except possibly at $(g_1)$ and $(g_2)$. Then, by (\[prop:ps2\]), $\alpha_i \otimes F_P(\sqrt[p^2]{h_1},
\cdots , \sqrt[p^2]{h_d}, \sqrt[p]{g_1}, \sqrt[p]{g_2}) = 0$ for all $i$.
Let $L = F(\sqrt[p]{f},
\sqrt[p^2]{u_1}, \cdots , \sqrt[p^2]{u_{d}}, \sqrt[p]{u_{d+1}}, \sqrt[p^2]{h_1}, \cdots ,
\sqrt[p^2]{h_d}, \sqrt[p]{g_1}, \sqrt[p]{g_2})$. We claim that $\alpha_i \otimes L = 0$ for all $i$. Let ${{{\mathscr{Y}}}}$ be a regular proper model of $L$ and $Y$ its special fiber. Let $y $ be a point of $Y$. If $y$ lies over a generic point of a component of the special fiber $X$ of ${{{\mathscr{X}}}}$, then by the choice of $f, u_1, \cdots , u_d$, $\alpha_i \otimes L_y = 0$. Suppose the image of $y$ lies over a closed point $P$ of $X$. Suppose $P \in U_\eta$ for some $\eta$. Then $F_{U_\eta} \subset F_P$ and hence once again by the choice of $f, u_1, \cdots , u_d$, $\alpha_i \otimes L_y = 0$ for all $i$. Suppose $P \not\in U_\eta$ for all $\eta$. Then $P \in {{{\mathscr{P}}}}$. Since $F_P \subset L_y$, by the choices of $h_1, \cdots , h_d, g_1, g_2$, $\alpha_i \otimes L_y = 0$ for all $i$.
Since $[L : F] \leq p^{4d + 4}$, the theorem follows.
\[cor:bad\_char\_2\] Let $K$ be a complete discretely valued field with residue field $\kappa$. Suppose that char$(K) = 0$ and char$(\kappa) = p > 0$ and $[\kappa : \kappa^p] = p^d$. Let $\zeta$ be a primitive $p^{th}$ root of unity. Then $[K(\zeta) : K]p^{4d+4}$ is a $(2, p)$-uniform bound for $K(t)$.
\[cor:bad\_char\_n\] Let $K' = K(\zeta)$. Then $K'$ is a complete discretely valued field with residue field $\kappa$. Let $F$ be a finite extension of $K(t)$. Let $\alpha_1, \cdots , \alpha_m \in H^2(F, \mu_p^{\otimes 2})$. Since $F'= F(\zeta)$ is also a function field over $K'$, by (\[theorem:bad\_char\]), there exists an extension $L$ of $F'$ of degree at most $p^{4d+4}$ such that $\alpha_i \otimes L = 0$ for all $i$. Since $[L : F] = [L :
F'][F' : F]$, the corollary follows.
The above corollary and (\[lemma:uniform\_bound\_d\_d+1\]) give the following
Let $K$ be a complete discretely valued field with residue field $\kappa$. Suppose that char$(K) = 0$ and char$(\kappa) = p > 0$ and $[\kappa : \kappa^p] = p^d$. Let $\zeta$ be a primitive $p^{th}$ root of unity. Then $[K(\zeta) : K]p^{4d+4}$ is an $(n, p)$-uniform bound for $K(t)$ for all $n \geq 2$.
Uniform bound - good characteristic case
=========================================
Let $F$ be the function field of a $p$-adic curve. In ([@saltman:jrms]), Saltman proved that if $\ell$ is a prime not equal to $p$ and $\alpha_1, \cdots , \alpha_m \in H^2(F, \mu_\ell)$, then there exists an extension $L$ of $F$ such that $[L : F] \leq \ell^2$ and $\alpha_{i_L} = 0 $ for all $i$, i.e., $F$ is $(2,
\ell)$-uniformly bounded.
Let $K$ be a complete discretely valued field with residue field $\kappa$. Let $\ell$ be a prime not equal to char$(\kappa)$. In this section we show that $K(t)$ is $(2,\ell)$-uniformly bounded under some conditions on $\kappa$.
\[theorem:good\_char\_dvr\] Let $K$ be a complete discretely valued field with residue field $\kappa$. Let $\ell$ be a prime not equal to char$(\kappa)$ and $n \geq 1$. If $N$ is an $(n, \ell)$-uniform bound for $\kappa$, then $\ell N$ is an $(n, \ell)$-uniform bound for $K$.
Let $L$ be a finite extension of $K$. Then $L$ is a complete discretely valued field with residue $\kappa'$ a finite extension of $\kappa$. Let $R$ be the valuation ring of $R$ and $\pi \in R$ be a parameter. Let $\alpha_1, \cdots, \alpha_m \in H^n(L,
\mu_\ell^{\otimes n})$. Let $S$ be the integral closure of $R$ in $L(\sqrt[\ell]{\pi})$. Then $S$ is also a complete discrete valuation ring with residue field $\kappa'$. Since $S/R$ is ramified, $\alpha_i \otimes
L(\sqrt[\ell]{\pi})$ is unramified at $S$ for each $i$. Hence there exists $\beta_i \in H^n_{{\mathrm{\acute{e}t}}}(S, \mu_\ell^{\otimes n})$ such that $\beta_i
\otimes_SK(\sqrt[\ell]{\pi}) = \alpha_i$. Since $N$ is an $(n,
\ell)$-uniform bound for $\kappa$, there exists an extension $L_0$ of $\kappa$ of degree at most $N$ such that $\beta_i \otimes_\kappa L_0 =
0$ for all $i$. Let $L$ be the extension of $K$ of degree equal to $[L_0 : \kappa]$ with residue field $L_0$. Let $T$ be the integral closure of $R$ in $L(\sqrt[\ell]{\pi})$. Then $T$ is a complete discrete valued ring with residue field $L_0$ and $S \subset T$. Since $\beta_i \otimes_\kappa L_0 = 0$, $\beta_i
\otimes_S T = 0$ for all $i$. In particular $\alpha_i \otimes_K
L(\sqrt[\ell]{\pi}) = 0$ for all $i$. Since the degree of $L$ over $K$ is equal to the degree of $L_0$ over $\kappa$ and $[L_0 :
\kappa] \leq N$, we have $[L(\sqrt[\ell]{\pi}) : K] \leq \ell N$.
\[lemma:good\_char\_dim2\] Let $A$ be a regular local ring with residue field $\kappa$ and maximal ideal $m = (\pi,\delta)$. Let $F$ be the field of fractions of $A$, $\ell$ a prime not equal to char$(\kappa)$. Let $B$ (resp. $B'$) be the integral closure of $A$ in $F(\sqrt[\ell]{\pi}, \sqrt[\ell]{\delta})$ (resp. $F(\sqrt[\ell]{\pi})$). Let $\alpha \in H^2(F, \mu_\ell^{\otimes 2})$. If $\alpha$ is unramified on $A$ except possibly at $(\pi)$ and $(\delta)$ (resp. except possibly at $(\pi)$), then $\alpha \otimes_F F(\sqrt[\ell]{\pi},
\sqrt[\ell]{\delta})$ is unramified on $B$ (resp. $\alpha \otimes_F
F(\sqrt[\ell]{\pi})$ is unramified on $B'$).
By ([@ps:period_index 3.3]), $B$ is a regular local ring of dimension 2 with residue field $\kappa$. Let $P$ be a height one prime ideal of $B$ and $Q = P \cap A$. Then $Q$ is a height one prime ideal of $A$. If $Q \neq (\pi)$ and $(\delta)$, then $\alpha$ is unramified at $Q$ and hence $\alpha \otimes F(\sqrt[\ell]{\pi}, \sqrt[\ell]{\delta})$ is unramified at $P$. Suppose $Q = (\pi)$ or $(\delta)$. Then $Q$ is ramified in $B$ and hence $\alpha \otimes F(\sqrt[\ell]{\pi}, \sqrt[\ell]{\delta})$ is unramified at $P$. Since $B$ is a regular local ring of dimension 2 and $\alpha \otimes F(\sqrt[\ell]{\pi}, \sqrt[\ell]{\delta})$ is unramified at every height one prime ideal of $B$, $\alpha \otimes F(\sqrt[\ell]{\pi}, \sqrt[\ell]{\delta})$ is unramified on $B$ (cf. \[auslander-goldman\]). The other case follows similarly.
\[theorem:good\_char\_killing\_ramification\] Let $R$ be a complete discrete valuation ring with field of fractions $K$ and residue field $\kappa$. Let $F$ be the function field of a curve over $K$. Let $\ell$ be a prime not equal to char$(\kappa)$ and $\alpha_1, \cdots ,
\alpha_r \in H^n(K, \mu_\ell^{\otimes m})$. Then there exist $f, g, h \in F^*$ such that each $\alpha_i \otimes F(\sqrt[\ell]{f}, \sqrt[\ell]{g},
\sqrt[\ell]{h})$ is unramified at all codimension one points of any regular proper model of $F(\sqrt[\ell]{f}, \sqrt[\ell]{g},
\sqrt[\ell]{h})$.
Let ${{\mathscr{X}}}$ be a regular proper model of $F$ over $R$ such that the union of the support of ramification locus of $\alpha_i$, for $1
\leq i \leq r$, is contained in the union of regular curves $C$ and $E$ with $C \cup E$ having only normal crossings. Let $f \in F^*$ be such that $${\rm div}_{{\mathscr{X}}}(f) = C + E + F$$ for some divisor $F$ on ${{\mathscr{X}}}$ which does not contain any irreducible component of $C \cup E$ and does not pass through any point of $C
\cap E$. Let $g \in F^*$ be such that $${\rm div}_{{\mathscr{X}}}(g) = C + G$$ for some divisor $G$ on ${{\mathscr{X}}}$ which does not contain any irreducible component of $C \cup E \cup F$ and does not pass through any point of $C
\cap E$, $C \cap F$ and $E \cap F$. Let $h \in F^*$ be such that $${\rm div}_{{\mathscr{X}}}(h) = E + H$$ for some divisor $H$ on ${{\mathscr{X}}}$ which does not contain any irreducible component of $C \cup E \cup F \cup G$ and does not pass through any point of $C
\cap E$, $C\cap F$, $C \cap G$, $E \cap F$, $E \cap G$ and $F \cap
G$.
Let $L = F(\sqrt[\ell]{f}, \sqrt[\ell]{g}, \sqrt[\ell]{h})$ and ${{\mathscr{Y}}}$ be a regular proper model of $L$ over $R$. We claim that $\alpha_i
\otimes_F L$ is unramified on ${{\mathscr{Y}}}$. Let $y \in {{\mathscr{Y}}}$ be a codimension one point of ${{\mathscr{Y}}}$. Then $y$ lies over a point $x$ of ${{\mathscr{X}}}$. If $x$ is not on $C$ or $E$, then each $\alpha_i$ is unramified at $x$ and hence $\alpha
\otimes_F L$ is unramified at $y$. Assume that $x \in C\cup E$.
Suppose that $x$ is a codimension one point of ${{\mathscr{X}}}$. Since $x$ is on $C$ or $E$, by the choice of $f$, $f$ is a parameter at $x$ and hence $L/K$ is unramified at $x$. In particular $\alpha_i \otimes_F L $ is unramified at $y$.
Suppose that $x$ is a closed point of ${{\mathscr{X}}}$. Suppose $x \in C$ and $x
\not\in E$. Let $A_x$ be the local ring of ${{\mathscr{X}}}$ at $x$ and $S_y$ be the local ring of ${{\mathscr{Y}}}$ at $y$. Suppose $x \not\in F$. Then the maximal ideal $m_x$ at $x$ is $(f, \delta_x)$ for some $\delta_x \in m_x$ and each $\alpha_i$ is unramified on the local ring at $x$ except possibly at $(f)$. Thus, by (\[lemma:good\_char\_dim2\]), each $\alpha_i \otimes_F
F(\sqrt[\ell]{f})$ is unramified at the integral closure of $A$ in $F[\sqrt[\ell]{f}]$. Since the integral closure of $A$ in $F[\sqrt[\ell]{f}]$ is contained in $S_y$, each $\alpha_i\otimes_F \otimes
L$ is unramified at $S_y$. Suppose $x \in F$. If $x \not\in G$, the as above each $\alpha_i \otimes_F L$ is unramified at $S_y$. If $x
\in F\cap G$, then by the choice $h$, $x \not\in H$ and hence as above, each $\alpha_i \otimes_F L$ is unramified at $S_y$. Similarly if $x \in E$ and $x \not\in C$, then each $\alpha_i
\otimes_F L$ is unramified at $S_y$.
Suppose $x \in C \cap E$. Then $x \not\in G$ and $x \not\in H$. In particular, $m_x = (g, h)$ and each $\alpha_i$ is unramified on $A_x$ except possibly at $(g)$ and $(h)$. As above each $\alpha_i \otimes_F
L$ is unramified at $S_y$.
\[lemma:crt\] Let $A$ be a semi-local regular domain with field of fractions $F$. For each maximal ideal $m$ of $A$, let $\mathfrak{s}(m)$ be a separable finite extension of the residue field $\kappa(m)$ at $m$ of degree $N_m$. Let $N$ be a common multiple of $N_m$, $m$ varying over all maximal ideals of $A$. Then there exists an extension $E$ of $F$ of degree at most $N$ such that for each maximal ideal $m$ of $A$ and for each maximal ideal $m'$ of the integral closure $B_m$ of $A_m$ in $E$, $B_m/m'$ contains a field isomorphic to $\mathfrak{s}(m)$.
Since $\mathfrak{s}(m)$ is a finite separable extension of $\kappa(m)$, there exists $\theta_m
\in \mathfrak{s}(m) $ such that $\mathfrak{s}(m) = \kappa(m)(\theta_m)$. Let $f_m(X) \in \kappa(m)[X]$ be the minimal polynomial of $\theta_m$ over $\kappa(m)$. Then the degree of $f_m(X)$ is $N_m$. Let $f(X) \in A[X]$ be a monic polynomial of degree $N$ such that $f(X) = f_m(X)^{N/deg(f_m)}$ modulo $m$ for each maximal ideal $m$ of $A$. Let $g(X)$ be any monic irreducible factor of $f(X)$ over $A$. Let $E = F[X]/(g(X))$. We claim that $E$ has the required property.
Let $m$ be a maximal ideal of $A$. By the choice of $f(X)$ and $g(X)$, we have $g(X) = f_m(X)^{r_m}$ modulo $m$ for some $r_m \geq
1$. Let $B_m$ be the integral closure of $A_m$ in $E$. Since $g(X)$ is monic, $A_m[X]/(g(X))$ is isomorphic to a subring of $B_m$. Let $g_m(X) \in A_m[X]$ be a monic polynomial with $g_m(X) = f_m(X)$ modulo $m$. Since $f(X) = f_m(X)^{r_m} = g_m(X)^{r_m}$ modulo $m$, the ideal $\tilde{m}$ of $ A_m[X]/(g(X))$ generated by $m$ and $g_m(X)$ is a maximal ideal with $(A_m[X]/(g(X)))/\tilde{m} \simeq \mathfrak{s}(m)$. Since $B_m$ is integral over a subring isomorphic to $A_m[X]/(g(X))$, for every maximal ideal $m'$ of $B_m$, $B_m/m'$ contains a subfield isomorphic to $(A_m[X]/(g(X)))/\tilde{m} \simeq \mathfrak{s}(m)$.
\[cor:crt\] Let $A$ be a semi-local regular domain with field of fractions $F$. Let $\ell$ be a prime. Suppose that $\ell$ is a unit in $A$. Let $\beta \in H^2_{\mathrm{\acute{e}t}}(A, \mu_\ell)$. Suppose that for every maximal ideal $m$ of $A$, there exists a finite separable extension $\mathfrak{s}(m)$ of $\kappa(m)$ of degree $N_m$ such that $\beta \otimes_A \mathfrak{s}(m) = 0$. Let $N$ be a common multiple of $N_m$, where $m$ varies over maximal ideals of $A$. Let $E$ be the field constructed in (\[lemma:crt\]). Then for any maximal ideal $m$ of $A$, $\beta \otimes_A (E \otimes_F F_m) = 0$.
Let $B$ be the integral closure of $A$ in $E$. Let $m$ be a maximal ideal of $A$. and $\hat{A}_m$ the completion of $A$ at $m$. Then, $B \otimes_A \hat{A}_m$ is complete and by the choice of $E$, $B \otimes_A \hat{A}_m$ modulo its radical is isomorphic to a product of fields with each factor containing a field isomorphic to $\mathfrak{s}(m)$. Since $\beta \otimes \mathfrak{s}(m) = 0$, it follows that $\beta
\otimes B \otimes \hat{A}_m = 0$. Since $E$ is the field of fractions of $B$, $\beta \otimes_A (E
\otimes_F F_m) = 0$.
\[theorem:good\_char\] Let $K$ be complete discretely valued field with residue field $\kappa$. Let $\ell$ be a prime not equal to char$(\kappa)$. Suppose that $N_1$ is a $(2, \ell)$-uniform bound for $\kappa(t)$ and $N_2$ is a $(2, \ell)$-uniform bound for $\kappa$. Then $\ell^3 (N_1!)(N_2!)$ is a $(2, \ell)$-uniform bound for $K(t)$.
Let $F$ be a finite extension of $K(t)$. Let $\alpha_1, \cdots , \alpha_m \in H^2(F, \mu_\ell^{\otimes 2})$. Then, by (\[theorem:good\_char\_killing\_ramification\]), there exist $f, g, h \in F^*$ such that $\alpha_i \otimes
F(\sqrt[\ell]{f}, \sqrt[\ell]{g}, \sqrt[\ell]{h})$ is unramified at every codimension one point of any regular proper model of $F(\sqrt[\ell]{f}, \sqrt[\ell]{g}, \sqrt[\ell]{h})$ over $R$. Let $L = F(\sqrt[\ell]{f}, \sqrt[\ell]{g}, \sqrt[\ell]{h})$. Let ${{{\mathscr{Y}}}}$ be a regular proper model of $L$ over $R$ and $Y$ the reduced special fiber of ${{\mathscr{Y}}}$.
Let $\eta$ be the generic point of an irreducible component $Y_\eta$ of $Y $. Let $A_\eta$ be the local ring at $\eta$. Since each $\alpha_i \otimes_F L$ is unramified at $\eta$, there exists $\beta_i \in H^2_{\mathrm{\acute{e}t}}(A_\eta, \mu_\ell^{\otimes 2})$ such that $\beta_i
\otimes_{A_\eta} F_\eta = \alpha_i$ for $1 \leq i \leq m$. Since $\kappa(\eta)$ is a finite extension of $\kappa(t)$ and $N_1$ is a $(2, \ell)$-uniform bund for $\kappa(t)$, there exists a finite extension $\mathfrak{s}(\eta)$ of degree at most $N_1$ such that $\beta_i \otimes_{A_\eta} \mathfrak{s}(\eta) = 0$ for all $i$. By (\[lemma:sep\_ext\]), we assume that $\mathfrak{s}(\eta)$ is separable over $\kappa(\eta)$. Let $A$ be the semi-local ring at the generic points of all irreducible components of the special fiber $Y$ of ${{\mathscr{Y}}}$. Then $A$ is a semi-local regular ring with field of fractions $L$. By (\[lemma:crt\]), there exists a field extension $L_1$ of $L$ of degree at most $N_1!$ such that for every maximal ideal $m'$ of the integral closure $B_m$ of $A_m$ in $L$, $B_m/m'$ contains a subfield isomorphic to $\mathfrak{s}(\eta)$. Hence, by (\[cor:crt\]), $\alpha_i \otimes (L_1 \otimes_{L} L_{\eta}) = \beta_i \otimes (L_1
\otimes_{L} L_{\eta}) = 0$ for all $i$. By ([@hhk:lgp_torsors 5.8], [@kmrt 1.17]), there exists a non-empty open set $U_\eta$ of the component $Y_\eta$ of the special fiber $Y$, such that $\alpha_i
\otimes L_1 \otimes L_{{U_\eta}} = 0$ for all $i$.
Let ${{\mathscr{P}}}$ be the finite set of closed points of ${{{\mathscr{Y}}}}$ which are not in $U_\eta$ for any $\eta$. Let $A_{{\mathscr{P}}}$ be the regular semi-local ring at the closed points of ${{\mathscr{P}}}$. Since each $\alpha_i $ is unramified on ${{\mathscr{Y}}}$, there exists $\beta_i \in H^2_{\mathrm{\acute{e}t}}(A_{{\mathscr{P}}}, \mu_\ell^{\otimes 2})$ such that $\beta_i \otimes L = \alpha_i \otimes L$. Let $P \in {{\mathscr{P}}}$. Since the residue field $\kappa(P)$ at $P$ is a finite extension of $\kappa$, by the assumption on $\kappa$, there exists an extension $\mathfrak{s}(P)$ of $\kappa(P)$ of degree at most $N_2$ such that $\beta_i \otimes \mathfrak{s}(P) = 0$ for all $i$. Once again, by (\[lemma:sep\_ext\]), we assume that each $\mathfrak{s}(P)$ is a separable extension of $\kappa(P)$. Let $L_2$ be as in (\[lemma:crt\]). Then, as above, by (\[cor:crt\]), $\alpha_i \otimes (L_2 \otimes L_{P}) = 0$ for all $i$.
Let $L = LL_1L_2$. Then as in (\[theorem:bad\_char\]), $L \otimes \alpha_i = 0$ for all $i$. Since $[L : F] \leq \ell^3(N_1!)(N_2!)$, the theorem follows.
\[cor:good\_char\_2\] Let $K$, $\kappa$, $\ell$, $N_1$ and $N_2$ be as in (\[theorem:good\_char\]). Let $\zeta$ be a primitive $\ell^{\rm th}$ root of unity. Then $[K(\zeta), K]\ell^3(N_1!)(N_2!)$ is a $(2, \ell)$-uniform bound of $K(t)$.
The above corollary and (\[lemma:uniform\_bound\_d\_d+1\]) gives the following
\[cor:good\_char\_n\] Let $K$, $\kappa$, $\ell$, $N_1$ and $N_2$ be as in (\[theorem:good\_char\]). Let $\zeta$ be a primitive $\ell^{\rm th}$ root of unity. Then $[K(\zeta), K]\ell^3(N_1!)(N_2!)$ is an $(n, \ell)$-uniform bound of $K(t)$ for all $n \geq 2$.
Symbol length and $u$-invariant
================================
\[theorem:symbol\_length1\] Let $K$ be a complete discretely valued field with residue field $\kappa$. Let $\ell$ be a prime not equal to char$(\kappa)$. Suppose that there exist integers $N_1$ and $N_2$ such that $\kappa(t)$ is $(2, \ell)$-uniformly bounded by $N_1$and $\kappa$ is $(2, \ell)$-uniformly bounded by $N_2$. Let $n \geq
2$. Then there exists an integer $M_n$ which depends only on $N_1$, $N_2$ and $n$ such that for every finite extension $F$ of $K(t)$ and for all $\alpha
\in H^n(F, \mu_\ell^{\otimes n})$, $\lambda(\alpha) \leq M_n$.
By (\[cor:good\_char\_2\]), $K(t)$ is $(2,
\ell)$-uniformly bounded by $N = (\ell-1)\ell^3(N_1!)(N_2!)$. Hence any finite extension $F$ of $K(t)$ is also $(2, \ell)$-uniformly bounded by $N$. The theorem follows from (\[cor:symbol\_length\]).
\[theorem:symbol\_length2\] Let $K$ be a complete discretely valued field with residue field $\kappa$. Let $p = $ char$(\kappa)$. Suppose that char$(K) = 0$, $p > 0$ and $[\kappa : \kappa^d] = p^d$. Then there exists an integer $M$, which depends only on $d$ such that for any finite extension $F$ of $K(t)$ and for all $\alpha \in H^n(F, \mu_p^{\otimes n})$, $n
\geq 1$, $\lambda(\alpha) \leq M$.
By (\[cor:bad\_char\_2\]), $K(t)$ is $(2,
p)$-uniformly bounded by $(p-1)p^{4d + 4}$. Let $F$ be a finite extension of $K(t)$. Then $F$ is also $(2, p)$-uniformly bounded by $(p-1)p^{4d+4}$. Let $n \geq 1$. By (\[cor:symbol\_length\]), there exists an integer $N_n$, which depends only on $d$ and $n$ such that for all $\alpha \in H^n(F, \mu_p^{\otimes n})$, $\lambda(\alpha) \leq N_n$. Since the $p$-cohomological dimension of $K$ is at most $d + 2$ ([@g-o]) and $F$ is a function field in one variable over $K$, the $p$-cohomological dimension of $F$ is $d +
3$. Hence $H^n(F, \mu_p^{\otimes n}) = 0$ for all $n \geq
d+4$. Let $N$ be the maximun of $N_n$ for $2 \leq n \leq d+4$. Then $\lambda(\alpha) \leq N$ for all $\alpha \in H^n(F,
\mu_p^{\otimes n})$ and $n \geq 2$.
\[theorem:u\_invariant\] Let $K$ be complete discretely valued field with residue field $\kappa$ and $F$ a function field of a curve over $K$. Suppose that char$(K) = 0$, char$(\kappa) = 2$ and $[\kappa : \kappa^2]$ is finite. Then there exists an integer $M$ which depends only on $[\kappa :
\kappa^2]$ such that for any finite extension $F$ of $K(t)$, $u(F) \leq M$.
The theorem follows from (\[theorem:symbol\_length2\]) and (\[cor:uniformbound\_uinvariant\]).
We end with the following
Let $L$ be a field of characteristic not equal to 2 with $u(L)$ finite. Is the Brauer group of $L$ uniformly 2-bounded?
[10]{}
A. A. Albert, *Structure of Algebras*, Amer. Math. Soc. Colloq. Publ., Vol. 24, Amer. Math. Soc., Providence, RI, 1961, revised printing.
Asher Auel, Eric Brussel, Skip Garibaldi, Uzi Vishne, *Open problems on central simple algebras*, Transformation Groups [**16**]{} (2011), 219-264.
\] N. Bourbaki, *Algebra II*, Springer-Verlag, New York, 1988.
M. Cipolla, *Remarks on the lifting of algebras over henselian pairs*, Mathematische Zeitschrift, [**152**]{} (1977), 253–257.
Auslander Maurice and Goldman Oscar, *The Brauer group of a commutative ring*, Trans. Amer. Math. Soc. [**97**]{} (1960) 367–409.
O. Gabber and F. Orgogozo, *Sur la $p$-dimension des corps*, Invent. Math. [**174**]{} (2008), 47-80.
D. Harbater, J. Hartmann and D. Krashen, *Applications of patching to quadratic forms and centrals simple algebras*, Invent. Math. [**178**]{} (2009), 231-263.
D. Harbater, J. Hartmann and D. Krashen, *Local-global principles for torsors over arithmetic curves*, arxive:1108.3323v2
Heath Brown, D.R., *Zeros of systems of p-adic quadratic forms*, Compos. Math. [**146**]{} (2010), 271-287.
Kahn, B., *On "horizontal” invariants attached to quadratic forms*, Algebra and number theory, 21–33, Hindustan Book Agency, Delhi, 2005.
M.-A. Knus, A.S. Merkurjev, M. Rost and J.-P. Tignol, *The Book of Involutions*, A.M.S, Providence RI, 1998.
Krashen, D., *Period and index, symbol lengths, and generic splittings in galois cohomology*, preprint 2013.
Leep, D.B., *The u-invariant of p-adic function fields*, J. Reine Angew. Math., to appear
Milne,J.S., *Étale cohomology*, Princeton Univ. Press, Princeton, N.J., 1980,
Orlov, D., Vishik, A., Voevodsky, V., *An exact sequence for $K^M_\ast/2$ with applications to quadratic forms*, Ann. of Math. [**165**]{} (2007), no. 1, 1–13.
Parimala, R. and Suresh, V., *On the length of a quadratic form*, Algebra and number theory, 147–157, Hindustan Book Agency, Delhi, 2005.
Parimala, R. and Suresh, V., *The $u$-invariant of the function fields of $p$-adic curves*, Annals of Mathematics [**172**]{} (2010), 1391-1405.
Parimala, R. and Suresh, V., *Period-index and $u$-invariant questions for function fields over complete discretely valued fields*, to appear in Invent. Math.
Saltman, D.J., *Division Algebras over $p$-adic curves*, J. Ramanujan Math. Soc. [**12**]{} (1997), 25-47.
Saltman, D.J., *Finite $u$ Invariant and Bounds on Cohomology Symbol Lengths*, preprint, 2011.
Serre, J.-P., *Galois Cohomology*, Springer monographs in Mathematics, 1996.
Voevodsky, V., *On motivic cohomology with ${\mathbb Z}/l$-coefficients*, Ann. of Math. (2) 174 (2011), no. 1, 401–438.
| {
"pile_set_name": "ArXiv"
} |
Introduction
============
Because of the nonlinear nature of the Einstein equations, linear fluctuations about a homogeneous and isotropic FRW cosmology will back-react on the background on which they live. The back-reaction of short wavelength gravitational waves was studied a long time ago by Brill, Hartle and Isaacson [@BHI]. In this approach, the back-reaction of fluctuations on the spatially averaged metric can be described by an effective energy-momentum tensor which contains the spatially averaged terms in the Einstein tensor which are second order in the amplitude of the fluctuations. More recently, Tsamis and Woodard [@WT] have initiated a detailed study of the back-reaction of long wavelength gravitational waves in a de Sitter background. Scalar-type metric fluctuations (called “cosmological fluctuations” in what follows) are also expected to back-react on the background space-time. Since in the context of inflationary and post-inflationary cosmology the scalar metric fluctuations are believed to dominate over the effects of gravitational waves, the back-reaction of these cosmological perturbations are expected to give the dominating back-reaction effect.
In [@bran2; @bran3], the formalism of [@BHI] was generalized to describe the back-reaction of cosmological perturbations on the spatially averaged metric. On this basis, it was argued [@bran2; @bran3; @aw0] that gravitational back-reaction in scalar field-driven inflationary models, calculated up to quadratic order in perturbations and to leading order in the long wavelength expansion and in the slow roll approximation could decrease the expansion rate of the universe and potentially solve the cosmological constant problem [@bran4] (see also [@atw] for similar results in the context of back-reaction studies of gravitational waves). The formalism of [@bran2; @bran3] is covariant under first order space-time diffeomorphisms [^1].
However, as emphasized by Unruh [@unruh] [^2], the approach of [@bran2; @bran3; @WT] is deficient in several respects. First of all, due to the nonlinear nature of the Einstein equations, calculating an “observable” from the spatially averaged metric will not in general give the same result as calculating the spatially averaged value of the observable. More importantly, the spatially averaged metric is not a local physical observable. Thus, to take into account the deficiencies of the previous work on gravitational back-reaction, we must identify a local physical variable which describes the expansion rate of the Universe, calculate the back-reaction of cosmological perturbations on this quantity, and then take the spatially averaged value of the result. It is important to fix the hypersurface of averaging by a clear physical prescription in order to remove the possibility of being misled by effects which are second order gauge artifacts.
In this paper, we propose an implementation of this approach. We focus on the variable which yields the general relativistic definition of the local expansion rate, calculate this quantity to second order in the amplitude of the cosmological fluctuations, in terms of a time variable defined by an unambiguous physical prescription. For simplicity, we assume that matter is described by a single field (either a hydrodynamical field or a single scalar field).
We study two examples, first a matter dominated Universe, and second the inflationary phase of a cosmology dominated by a single scalar field. In both cases we find that the leading infrared contributions to the back-reaction on the local expansion rate of the Universe vanish, in contrast to the findings of the initial work on gravitational back-reaction of cosmological fluctuations [@bran2; @bran3; @aw0], and confirming the analysis of [@unruh]. We thus confirm the conclusions reached recently in [@aw2; @aw3] where a different variable related to the local expansion rate is proposed, and different techniques to evaluate this variable are used.
Note that when evaluated at a fixed background time, our leading infrared back-reaction terms give a non-vanishing contribution. This leads us to the conjecture that in more realistic models in which a second field is present to determine time (e.g. the microwave background), the leading infrared back-reaction terms will not vanish.
A local observable
==================
For a general perfect fluid flow in a curved space-time we consider the velocity four-vector field $u^{\alpha}$ tangential to a family of world lines. In the context of cosmology, we can always define a preferred family of world lines representing the motion of a set of comoving observers. In the case of hydrodynamical matter, this is easy since the energy-momentum tensor is already defined in terms of a velocity four-vector field. Also in the case of scalar field matter, a corresponding four-vector field can be defined, although a bit more care is required to obtain a consistent definition. In both cases we have $$\label{norm}
u^{\alpha}u_{\alpha} \, = \, 1 \, ,$$ where $\alpha$ runs over the space-time indices. The projection tensor onto tangential three-surfaces orthogonal to $u^{\alpha}$ is $$h_{\alpha\beta} \, = \, g_{\alpha\beta} - u_{\alpha}u_{\beta} \, .$$
The first covariant derivative of the four-velocity can be decomposed as (see e.g. [@EB] for details ) $$u_{\alpha;\beta} \, = \, \omega_{\alpha\beta} + \sigma_{\alpha\beta}
+ \frac{1}{3}\Theta h_{\alpha\beta}-\dot{u}_{\alpha}u_{\beta} \, .$$ Here $$\Theta \, \equiv \, u^{\alpha}_{\,;\alpha}$$ is the local expansion rate of the tangential surfaces orthogonal to the fluid flow, $\omega$ is the vorticity tensor (with $\omega_{\alpha\beta}u^{\beta} = 0$), and $\sigma$ is the shear tensor (satisfying $\sigma_{\alpha\beta}u^{\beta}=0, \, \sigma^{\alpha}_{\alpha}=0$).
For a homogeneous Universe with scale factor $a(t)$ the Hubble expansion rate $H$ is $$\label{Hubble}
H \, \equiv \, \dot{a}/a \, = \, \frac{1}{3}\Theta \, .$$ For a cosmological model with fluctuations, $\Theta$ is local in space and time. We will use $\Theta$ to define the local expansion rate $\dot{S}/S$ via the local analog of (\[Hubble\]), namely via the equation $$\label{localHubble}
\dot{S}/S \, \equiv \,\frac{1}{3}\Theta \, .$$ The quantity $\dot{S}/S$ is a much better measure of the locally measured expansion rate in a Universe with fluctuations than the Hubble expansion rate used in [@bran2; @bran3] determined via the spatially averaged metric including back-reaction. It is a mathematically simpler object than the variable recently introduced in [@aw2] which involves the integral along the past light cone of the observation point. If we are interested in evaluating the expansion rate for a typical observer, we propose to take the spatial average of the local expansion rate defined via (\[localHubble\]).
Now that we have defined the observable we are interested in, the procedure will be as follows. First, we must determine the velocity four-vector field $u^{\alpha}$ for the models we are interested in. Then, we use the Einstein equations to express $u^{\alpha}$ in terms of the metric perturbation. Taking the relative amplitude of the metric fluctuations as the expansion parameter, we then calculate $\Theta$, our local measure of the Hubble expansion rate, to second order. After evaluating the result on a physically determined hypersurface we can then study the back-reaction of cosmological fluctuations on the locally measured Hubble expansion rate. In this paper we will focus on the leading infrared contributions to back-reaction, the terms found to dominate the back-reaction effects in [@bran2; @bran3; @aw0].
Deriving the Expansion Rate for Scalar Metric Perturbations
===========================================================
In this section we consider a model with hydrodynamical matter. Starting from the expression for the metric to linear order in the fluctuations $\Phi$ (see [@RA] for a detailed review), we determine the velocity four vector field $u^{\alpha}$ to the order required to analyze the leading infrared terms in the back-reaction to quadratic order. To obtain the full back-reaction terms (including terms which dominate in the ultraviolet but are negligible in the infrared) we should calculate $u^{\alpha}$ consistently up to second order. However, if we are only interested in the leading infrared terms, it is sufficient to keep all the terms quadratic in $\Phi$ but not containing any spatial gradients.
In order to obtain the complete result for gravitational back-reaction we would have to look at the Einstein equations for a perfect fluid with energy density $\rho$ and pressure $P$, $$G_{\mu\nu} =(P+\rho)u_{\mu}u_{\nu}-Pg_{\mu\nu}$$ (in units in which $8 \pi G = 1$), which, since $G^{\mu}_{\mu}=R$, will yield $$\begin{aligned}
R&=&\rho-3P \\
\rho&=&u^{\mu}G_{\mu\nu}u^{\nu} \, ,\end{aligned}$$ and lead to an equation that can be solved perturbatively to any desired order for $u_{i}$: $$\label{einstein}
G^{0}_{i}=\frac{4}{3}u^{\mu}G_{\mu\nu}u^{\nu}u^{0}u_{i}+Ru^{0}u_{i}$$ However, as mentioned above, here we just use the results for $u^{i}$ which are of linear order. Since we will calculate the divergence of $u^{\mu}$, our prescription implies that we are ignoring some of the extra second order gradient terms . The result can also be used for scalar fields if we define the $u$ vector field in a proper way.
For an unperturbed Robertson-Walker metric, the four-velocity field $u$ in comoving coordinates would be $$u^{\mu}=(1,0,0,0)$$ In linear perturbation theory, and in the case of simple forms of matter (such as a single fluid or a single scalar field) for which there is to linear order no anisotropic stress, the metric (in longitudinal gauge) can be written as $$\begin{aligned}
\label{metric}
ds^{2}&=& a(\eta)^2\bigl((1+2\Phi)d\eta^{2} -
(1-2\Psi)\gamma _{ij}dx^{i}dx^{j}\bigr),\\
\gamma_{ij}&=&\delta_{ij}[1+\frac{1}{4}{\cal
K}(x^{2}+y^{2}+z^{2})]\end{aligned}$$ where ${\cal K}=0,1,-1$ depending on whether the three-dimensional space corresponding to the hypersurface $t = {\rm const.}$ is flat, closed or open. In this paper we will take it to be zero in order to simplify the calculations. The time variable $\eta$ appearing in (\[metric\]) is conformal time and is related to the physical time $t$ via $d\eta=a^{-1}dt$. For the forms of matter considered here, $\Psi=\Phi$ at linear order [^3]. As discussed e.g. in [@RA], in longitudinal gauge the spatial components of the four-velocity vector field are related to $\Phi$ via $$\delta u_{i} \, = \, -a^{-2}({\cal H} ^{2}- {\cal H} ^{\prime}+ {\cal
K})^{-1}(a\Phi)^{\prime}_{,i}$$ where a prime denotes differentiation with respect to $\eta$ and ${\cal H}=a^{\prime}/a$. Using equation (\[norm\]) we can derive the expression for the time component of $u^{\alpha}$ in terms of $\Phi$: $$u^{0}(\eta)=a^{-1}(1-\Phi+\frac{3}{2}\Phi^{2})$$
Now that we have all components of $u^{\alpha}$, we take the covariant derivative of it and retain all $\Phi$ dependence up to second order [^4] These second order terms principally come from the Christoffel symbols. Other second order terms (which as mentioned before are gradient terms) could be added to the ones computed here if we were to solve the Einstein equations in the form (\[einstein\]) beyond linear order. A straightforward calculation yields $$\begin{aligned}
\label{expform}
\Theta \, &=& \, 3\frac{a^{\prime}}{a^{2}}(1-\Phi+\frac{3}{2}\Phi^{2})
-3\frac{\Phi^{\prime}}{a} \nonumber \\
&+& \, \frac{(a\partial_{i}\Phi)^{\prime}
(\partial_{i}\Phi)+(a\partial_{i}^{2}\Phi)^{\prime}}{a^{2}({\cal
H}^{\prime}-{\cal H}^{2})} \, .\end{aligned}$$ Since $\dot{a}/a=a^{\prime}/a^{2}$, we can immediately read off the extra terms contributing to the local expansion rate which result from the presence of cosmological fluctuations. Upon spatial averaging at a fixed conformal time, the terms linear in $\Phi$ drop out. Hence, it follows that if evaluated at a fixed conformal time, infrared modes on average lead to an increase in the expansion rate compared to what would be obtained at the same conformal time in the absence of metric fluctuations. Whether this is a physically measurable effect from an observational point of view will be discussed in more depth in following sections.
Expansion Rate for a Matter Dominated Universe
==============================================
Now let us use the result of the last section to (as an example) calculate the local Hubble expansion rate for a matter-dominated Universe. In this case, the scale factor $a$ and the scalar metric perturbation $\Phi$ have the following dependence on the conformal time $\eta$ : $$\begin{aligned}
a(\eta) \, &=& \, a_{m}\eta^{2}/2 \label{smfactor} \\
\Phi(\eta,x) \, &=& \, C_{1}(x)+C_{2}(x)\eta^{-5} \label{smpert} \, ,\end{aligned}$$ where $a_m$ is a constant, and $C_1$ and $C_2$ are time-independent. The second equation is valid in the long wavelength (super-Hubble-scale) limit and has been explicitly derived in [@RA]. Now using equations (\[expform\]), (\[smfactor\]) and (\[smpert\]) we obtain $\Theta $ in terms of conformal time: $$\begin{aligned}
\Theta \, &=& \, \frac{12}{a_{m}}\eta^{-3}(1-C_{1}(x)+\frac{3}{2}C_{1}^{2})
\nonumber \\
&-& \frac{2}{3}\frac{1}{a_{m}}\eta^{-1}((\partial_{i}C_{1}(x))^{2}
+ \partial_{i}^{2}C_{1}(x)) \nonumber \\
&+& \frac{1}{a_{m}}\eta^{-6}
(\frac{1}{3}(\partial_{i}C_{1}(x))(\partial_{i}C_{2}(x))
+ \partial_{i}^{2}C_{2}(x)) \\
&+& \frac{1}{a_{m}}\eta^{-8}(3C_{2}(x)+36C_{1}(x)C_{2}(x)) \nonumber \\
&+& \frac{1}{a_{m}}\eta^{-11}
(\partial_{i}C_{2}(x))^{2}+\frac{1}{a_{m}}\eta^{-13}(C_{2}(x))^{2} \nonumber\end{aligned}$$ Some of the terms are decreasing very fast as a function of time and thus we can ignore them. If we just keep the terms with the powers $-3$ and $-1$ of $\eta$ then we get: $$\begin{aligned}
\label{matterfin}
\Theta \, &=& \, \frac{12}{a_{m}}\eta^{-3}(1-C_{1}(x)+\frac{3}{2}C_{1}^{2}) \\
&-& \frac{2}{3}\frac{1}{a_{m}}\eta^{-1}[(\partial_{i}C_{1}(x))^{2}
+\partial_{i}^{2}C_{1}(x)] \nonumber\end{aligned}$$
If we take the average of $\Theta$ on a constant $\eta$ hypersurface, only the terms quadratic in the fluctuation variables survive. Thus, considering large values of $\eta$ and focusing on the second order terms, it appears from (\[matterfin\]) that infrared modes give a positive contribution to $\Theta$ and thus lead to a speeding-up of the expansion, whereas ultraviolet modes enter with a negative sign and thus yield a slowing effect, the latter becoming more significant (relative to the unperturbed expansion rate) for larger values of $\eta$.
However, before drawing definite physical conclusions from our analysis, we must take into account that the background time $\eta$ is not an observable quantity. To obtain results for back-reaction of any real physical significance we have to find an observable variable like proper time and evaluate the expansion rate in terms of this variable, so that we can discuss its evolution from an observer’s point of view.
If we use equation (\[metric\]) for the metric, equation (\[smpert\]) for $\Phi$, and equation (\[smfactor\]) for the scale factor, we can find the expression for the proper time $\tau$ in terms of conformal time. Since $$d\tau^{2}=a(\eta)^{2}(1+2\Phi)d\eta^{2} \, ,$$ a simple integration yields $$\begin{aligned}
\tau \, &=& \, \frac{a_{m}}{6}(1+C_{1}-\frac{1}{2}C_{1}^{2})\eta^{3} \\
&-& \frac{a_{m}}{4}(C_{2}-C_{1}C_{2})\eta^{-2}
+ \frac{a_{m}}{28}C_{2}^{2}\eta^{-7} \nonumber\end{aligned}$$ In the approximation of large values of $\eta$ we can ignore the second and the third term of this equation and thus obtain: $$\eta^{-3} \, = \, \frac{a_{m}}{6}(1+C_{1}-\frac{1}{2}C_{1}^{2} )\tau^{-1} \, .$$ Now we can use this relation and substitute into the first term of equation (\[matterfin\]). We see the effects of the dominant infrared terms on the local expansion rate cancel exactly up to second order in perturbations, when evaluating $\Theta$ at a fixed proper time: $$\Theta_{IR}=2\tau^{-1} \, .$$ This implies that at least in the approximation of keeping only the leading infrared terms, there is no local gravitational back-reaction of cosmological fluctuations on the local Hubble expansion rate in this matter-dominated universe.
Expansion Rate in Terms of Scalar Field as an Observable
=========================================================
We now move on to the example more relevant to the work of [@bran2; @bran3; @aw0], namely a Universe dominated by a single real scalar field $\varphi$,which is a toy model for inflationary cosmology. During inflation, fluctuations which are generated on sub-Hubble scales early on during the inflationary phase are red-shifted to scales much larger than the Hubble radius. Thus, in this context it is of great interest to consider the back-reaction of infrared modes.
In the following, we will generalize the previous analysis to be applicable to matter consisting of a single scalar field. In this case one can treat the scalar field as a perfect fluid and derive the velocity four-vector field. To do this, we need to write the energy-momentum tensor $T_{\mu\nu}$ of the scalar field (right hand side of the following equation) in the form of an energy-momentum tensor for a perfect fluid (left hand side of the following equation): $$(\rho+P)u_{\mu}u_{\nu} - P g_{\mu\nu} \,
= \, \partial_{\mu}\varphi\partial_{\nu}\varphi-{\cal L} \, .
g_{\mu\nu}$$ At the level of the background fields, the two expressions are identical if we take $P = {\cal L}$ and $u_{\mu}=A\partial_{\mu}\varphi$ with $$A \, = \, \bigl(\partial^{\nu}\varphi\partial_{\nu}\varphi\bigr)^{-1/2}
\, .$$
Now that we have shown that the energy-momentum tensor of a scalar field can be written in the form of that of a perfect fluid, we can use the expression (\[expform\]) which gives $\Theta$ in terms of the metric fluctuations to evaluate the local effect of gravitational back-reaction of cosmological fluctuations. Let us first for convenience rewrite $\Theta$ of (\[expform\]) in terms of the physical time $t$: $$\label{expform2}
\Theta \, = \, 3\frac{\dot{a}}{a}(1-\Phi+\frac{3}{2}\Phi^{2})
-3\dot{\Phi}+\frac{(\partial_{i}\dot{\Phi})(\partial_{i}\Phi)+\partial^{2}_{i}\dot{\Phi}}{a\ddot{a}-\dot{a}^{2}}$$
Theoretically, the scalar field $\varphi$ is an observable. In fact, in a system with a single matter field $\varphi$, it is this field which must be used as a clock. Hence, to obtain results with physical meaning, we must evaluate $\Theta$ on a surface of constant $\varphi$ and not constant $t$.
As discussed above, in the context of inflationary cosmology it is important to study the effects of infrared modes up to second order. We will also assume that we are in the slow rolling regime of inflation. In our case (as well as for a more general case), the prescription is to calculate $t$, $\Phi$, $a$ and $\frac{\partial}{\partial t}$ in terms of $\varphi$, and to insert the results into the general expression (\[expform2\]) for $\Theta$. The relation between $t$ and $\varphi$ can be derived starting from: $$\varphi(t)=\varphi_{0}(t)+\delta \varphi_{1}(t) \, ,$$ which can be written as: $$t \, = \, \varphi_{0}^{-1}[\varphi-\delta \varphi_{1}(t)] \, .$$ Thus, the equation $$\begin{aligned}
t \, &=& \, \varphi_{0}^{-1}(\varphi)-\frac{\partial\varphi_{0}^{-1}(\varphi)}{\partial\varphi}\delta\varphi_{1}(\varphi_{0}^{-1}(\varphi)) \nonumber \\
&+&(\frac{\partial\varphi_{0}^{-1}(\varphi)}{\partial\varphi})^{2}\frac{\partial\delta\varphi_{1}(\varphi_{0}^{-1}(\varphi))}{\partial
t}\delta\varphi_{1}(\varphi_{0}^{-1}(\varphi)) \\
&+& \frac{1}{2}\frac{\partial^{2}\varphi_{0}^{-1}(\varphi)}{\partial
\varphi^{2}}\delta \varphi_{1}^{2}(\varphi_{0}^{-1}(\varphi)) \nonumber\end{aligned}$$ relates $t$ and $\varphi$.
We now wish to express the value of $\Theta$ in terms of $\varphi$ (note that we are considering the local value of $\Theta$ in this analysis, and there is no need to perform a spatial averaging). In order to relate the metric, its fluctuations and the scale factor to the scalar field and its fluctuations we need to take an explicit form for the potential and make use of the Einstein constraint equations. The $G^{0i}$ and $G^{00}$ equations relate $\Phi$ to $\varphi$ (at the level of the first order fluctuations) and the Hubble parameter of the background to the background scalar field $\varphi_{0}$ (at the level of the unperturbed Friedmann equations). In our simple system we do not need to go through these calculations explicitly (as long as we are interested only in the leading infrared terms) since by dropping the gradient terms from the $G^{0i}$ and $G^{00}$ equations, it can be shown that [@niayesh] $$\frac{H}{\sqrt{1+2\Phi}} \, = \, {1 \over {\sqrt{3}}}\sqrt{V(\varphi)} \, .$$ If we expand the right hand side in terms of $\Phi$ we get the result $$H(1-\Phi+\frac{3}{2}\Phi^{2}) \,
= \, {1 \over {\sqrt{3}}}\sqrt{V(\varphi)} \, ,$$ and substituting this result into equation (\[expform2\]) and neglecting terms containing $\dot{\Phi}$ and $\partial_{i}\Phi$ (which are sub-dominant compared to other terms in the infrared and slow roll limit) leads to the final result for the local expansion rate: $$\label{final}
\Theta \, = \, \sqrt{3} \sqrt{V(\varphi)}$$ which as a function of $\varphi$ is the same as the relation for an unperturbed background.
Thus, again we do not see any back-reaction of cosmological perturbations on the local expansion rate in this approximation. In retrospect it is easy to understand this result, since [@niayesh] by neglecting gradient terms we have a Friedmann Universe whose expansion rate satisfies equation (\[final\]).
Conclusions
===========
In this paper, we have studied the back-reaction effects on a local observable which measures the local expansion rate of the Universe. The observable gives the rate at which neighboring comoving observers separate and coincides with the usual definition of expansion in the context of the fluid approach to cosmology. In order to obtain a physical quantity, we evaluated the observable at a fixed value of the scalar field.
We evaluated our observable, the [*local physical expansion rate*]{}, in a simple toy model of chaotic inflation consisting of a single scalar matter field coupled to gravity. We found that the leading infrared terms, the terms which dominate the effects discussed in [@bran2] and [@aw0], cancel if we evaluate the observable at a fixed value of the scalar field, whereas they do not vanish if we evaluate them at a fixed value of the background time. The former result is a physical result since it corresponds to a physical observable evaluated at a space-time point specified by a physical prescription, whereas the latter result (obtained by evaluating at a fixed background time) does not have a diffeomorphism-invariant meaning. Our analysis thus confirms the concern of [@unruh] that the results obtained in [@bran2] and [@aw0] are not invariant under second order gauge transformations. Our results confirm the conclusions of [@aw3] reached by means of a different method of analysis applied to a different physical observable.
Our result does [**not**]{} imply that there is no back-reaction of the infrared modes of cosmological perturbations. There is no reason to expect that the next to leading infrared terms in our result will cancel (they do not cancel in the analysis of [@aw3]). One of the advantages of our technique is that they can be evaluated without too much trouble. This is left to a future publication. So, even in single field models of inflation there might be some non-vanishing back-reaction of infrared modes.
We expect that back-reaction of infrared modes will be much more important in two field models of inflation. Let us assume that the matter sector of the theory contains both an inflation field $\varphi$ and a regular matter field $\chi$ (with nonvanishing and time-dependent spatial average) which, for example, could represent the cosmic microwave background. In this case, it is no longer true that long wavelength fluctuations have no physical effects on local observables. If the measurement point is (in an unambiguous physical way) determined by a fixed value of the field $\chi$, then the local expansion rate may sensitively depend on the amplitude of the long wavelength fluctuations in $\varphi$. Thus, the leading infrared terms may [**not**]{} cancel when evaluated according to the abovementioned physical prescription in the same way that they do not cancel in the analysis of this paper when the observable is evaluated at a fixed value of the background time.
There is a close analogy with the analysis of the parametric amplification of super-Hubble-scale cosmological fluctuations during inflationary reheating. From the point of view of the background space-time coordinates, it appears [@Kaiser] that the parametric amplification of matter fluctuations on super-Hubble scales in an unperturbed cosmological background (see e.g. [@tb; @kls] for a discussion of parametric resonance during reheating) would imply the parametric amplification of the cosmological fluctuations on these scales. However, it can be shown that in single field models physical observables measuring the amplitude of cosmological fluctuations do not feel any resonance [@fb1; @parry; @lin; @niayesh]. In contrast, in two field models of inflation there is [@bv; @fb2] parametric amplification of super-Hubble-scale cosmological fluctuations. In this case, there is a fluctuation mode corresponding to entropy fluctuations which cannot locally be gauged away. This mode is (in certain theories) parametrically amplified during reheating, and in turn drives the parametric resonance of the super-Hubble scale curvature fluctuations.
Our techniques allow us to calculate the back-reaction of cosmological fluctuations in two field models in a very similar way to what is presented here. Results will be presented in a followup publication [@Ghazal2].
Acknowledgments
We would like to thank Bill Unruh, Alan Guth and Andrei Linde for discussions which stimulated this research. We are grateful to Raul Abramo, Niayesh Afshordi, Yasusada Nambu and Richard Woodard for many useful insights during the course of this research. This research is supported in part (at Brown) by the US Department of Energy under Contract DE-FG0291ER40688, Task A. One of us (GG) is grateful to the Department of Astrophysical Sciences of Princeton University for hospitality during the course of this work. The other of us (RB) wishes to thank the CERN Theory Division and the Institut d’Astrophysique de Paris for hospitality and financial support.
D. Brill and J. Hartle, Phys. Rev. [**135**]{}, B271 (1964);\
R. Isaacson, Phys. Rev. [**166**]{}, 1272 (1968).
N. C. Tsamis and R. P. Woodard, Phys. Lett. B [**301**]{}, 351 (1993);\
N. C. Tsamis and R. P. Woodard, Nucl. Phys. B [**474**]{}, 235 (1996) \[arXiv:hep-ph/9602315\]. V. F. Mukhanov, L. R. Abramo and R. H. Brandenberger, Phys. Rev. Lett. [**78**]{}, 1624 (1997) \[arXiv:gr-qc/9609026\]. L. R. Abramo, R. H. Brandenberger and V. F. Mukhanov, Phys. Rev. D [**56**]{}, 3248 (1997) \[arXiv:gr-qc/9704037\]. L. R. Abramo and R. P. Woodard, Phys. Rev. D [**60**]{}, 044010 (1999) \[arXiv:astro-ph/9811430\]. R. H. Brandenberger, “Back reaction of cosmological perturbations,” arXiv:hep-th/0004016. L. R. Abramo, R. P. Woodard and N. C. Tsamis, Fortsch. Phys. [**47**]{}, 389 (1999) \[arXiv:astro-ph/9803172\]. T. Futamase, Phys. Rev. Lett. [**61**]{}, 2175 (1988).
T. Futamase, Mon. Not. R. Astr. Soc. [**237**]{}, 187 (1989).
S. Bildhauer and T. Futamase, Gen. Relativ. Gravit. [**23**]{}, 1251 (1991).
T. Futamase, Phys. Rev. D [**53**]{}, 681 (1996). U. Seljak and L. Hui, in ’Proceedings of Clusters, Lensing and the Future of the Universe’, College Park, Maryland, 1995 (unpublished).
H. Russ, M. H. Soffel, M. Kasai and G. Borner, Phys. Rev. D [**56**]{}, 2044 (1997) \[arXiv:astro-ph/9612218\]. J. P. Boersma, Phys. Rev. D [**57**]{}, 798 (1998) \[arXiv:gr-qc/9711057\]. T. Buchert, Gen. Rel. Grav. [**33**]{}, 1381 (2001) \[arXiv:gr-qc/0102049\]. T. Buchert and J. Ehlers, Astron. Astrophys. [**320**]{}, 1 (1997) \[arXiv:astro-ph/9510056\]. M. Takada and T. Futamase, Gen. Rel. Grav. [**31**]{}, 461 (1999) \[arXiv:astro-ph/9901079\]. Y. Nambu, Phys. Rev. D [**62**]{}, 104010 (2000) \[arXiv:gr-qc/0006031\]. Y. Nambu, Phys. Rev. D [**63**]{}, 044013 (2001) \[arXiv:gr-qc/0009005\]. Y. Nambu, “The back reaction and the effective Einstein’s equation for the Universe with ideal fluid cosmological perturbations,” arXiv:gr-qc/0203023. W. Unruh, “Cosmological long wavelength perturbations,” arXiv:astro-ph/9802323. L. R. Abramo and R. P. Woodard, Phys. Rev. D [**65**]{}, 043507 (2002) \[arXiv:astro-ph/0109271\]. L. R. Abramo and R. P. Woodard, Phys. Rev. D [**65**]{}, 063515 (2002) \[arXiv:astro-ph/0109272\]. G. F. Ellis, J. Hwang and M. Bruni, Phys. Rev. D [**40**]{}, 1819 (1989). V. F. Mukhanov, H. A. Feldman and R. H. Brandenberger, Phys. Rept. [**215**]{}, 203 (1992). N. Afshordi and R. H. Brandenberger, Phys. Rev. D [**63**]{}, 123505 (2001) \[arXiv:gr-qc/0011075\]. B. A. Bassett, D. I. Kaiser and R. Maartens, Phys. Lett. B [**455**]{}, 84 (1999) \[arXiv:hep-ph/9808404\]. J. H. Traschen and R. H. Brandenberger, Phys. Rev. D [**42**]{}, 2491 (1990). L. Kofman, A. D. Linde and A. A. Starobinsky, Phys. Rev. Lett. [**73**]{}, 3195 (1994) \[arXiv:hep-th/9405187\]. F. Finelli and R. H. Brandenberger, Phys. Rev. Lett. [**82**]{}, 1362 (1999) \[arXiv:hep-ph/9809490\]. M. Parry and R. Easther, Phys. Rev. D [**59**]{}, 061301 (1999) \[arXiv:hep-ph/9809574\]. W. B. Lin, X. H. Meng and X. M. Zhang, Phys. Rev. D [**61**]{}, 121301 (2000) \[arXiv:hep-ph/9912510\]. B. A. Bassett and F. Viniegra, Phys. Rev. D [**62**]{}, 043507 (2000) \[arXiv:hep-ph/9909353\]. F. Finelli and R. H. Brandenberger, Phys. Rev. D [**62**]{}, 083502 (2000) \[arXiv:hep-ph/0003172\]. R. Brandenberger, G. Geshnizjani, in preparation.
[^1]: The back-reaction of small-scale (i.e. smaller than the Hubble radius) cosmological perturbations has been considered (without taking into account the issues of gauge freedom) in [@Futamase1; @Futamase2; @BF; @Futamase3; @SH; @Ruus; @Jelle; @Buchert]. This problem has also been considered in the context of Newtonian cosmological perturbation theory in [@BE; @TF]. More recently, Nambu [@Nambu1; @Nambu2; @Nambu3] has initiated a program to compute back-reaction effects on the spatially averaged metric using the renormalization group method.
[^2]: We are also grateful to Andrei Linde and Alan Guth for detailed private discussions on these points.
[^3]: Even if we did not make the assumption $\Psi = \Phi$, it turns out that at second order all infrared terms depending on $\Psi - \Phi$ will drop out.
[^4]: This is in the philosophy of the general back-reaction approach in which it is assumed that the fluctuations of the metric and matter satisfy the linear perturbation equations, and we compute their back-reaction on physical quantities to second order. It is not a consistent second order perturbative formalism.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show how symmetry properties can be used to greatly increase the accuracy and efficiency in auxiliary-field quantum Monte Carlo (AFQMC) calculations of electronic systems. With the Hubbard model as an example, we study symmetry preservation in two aspects of ground-state AFQMC calculations, the Hubbard-Stratonovich transformation and the form of the trial wave function. It is shown that significant improvement over state-of-the-art calculations can be achieved. In unconstrained calculations, the implementation of symmetry often leads to shorter convergence time and much smaller statistical errors, thereby a substantial reduction of the sign problem. Moreover, certain excited states become possible to calculate which are otherwise beyond reach. In calculations with constraints, the use of symmetry can reduce the systematic error from the constraint. It also allows release-constraint calculations, leading to essentially exact results in many cases. Detailed comparisons are made with exact diagonalization results. Accurate ground-state energies are then presented for larger system sizes in the two-dimensional repulsive Hubbard model.'
author:
- Hao Shi
- Shiwei Zhang
bibliography:
- 'afqmc.bib'
title: 'Symmetry in Auxiliary-Field Quantum Monte Carlo Calculations '
---
Introduction {#sec:Introduction}
============
The study of interacting quantum many-body systems remains an outstanding challenge, especially systems with strong particle interactions, where perturbative approaches are ineffective. Numerical simulations provide a promising approach for studying such systems. One of the most general numerical approaches is quantum Monte Carlo (QMC) methods based on auxiliary-fields, which are applied in condensed matter physics, nuclear physics, high-energy physics, and quantum chemistry. These methods [@BSS1PhysRevD.24.2278; @BSS2PhysRevB.24.4295; @KooninSugiyama19861] allow essentially exact calculations of ground-state and finite-temperature equilibrium properties of interacting many fermion systems. As is well known, however, they suffer from the sign problem which severely limits their applicability [@SIGNPPRB1990; @PhysRevB.55.7464]. A key issue in algorithm development in this context is to reduce the convergence time or variance reduction for fixed computing cost.
Considerable progress has been achieved in circumventing this problem by constraining the random walks in sampling the space of auxiliary-fields. These methods have come under the name of Constrained Path Monte Carlo (CPMC) [@PhysRevB.55.7464; @PhysRevB.78.165101] for systems where there is a sign problem (for example, Hubbard-like models where the local interactions lead to auxiliary-fields that are real). For electronic systems where there is a phase problem (as the Coulomb interaction leads to complex fields), the methods [@PhysRevLett.90.136401; @PhysRevE.70.056702; @al-saidi:224101] have been referred to as phaseless or phase-free auxiliary-field QMC. In both cases, the idea is to constrain the sign or phase of the overlap of the sampled Slater determinants with a trial wave function. It eliminates the sign or phase instability and restores low-power (typically to the third power of system size) computational scaling. Applications to a variety of systems have shown that the methods are very accurate, even with simple trial wave wave functions taken directly from mean-field calculations (see, e.g. Refs [@PhysRevB.80.214116; @PhysRevLett.104.116402] and references therein). However, these methods are approximate. For example, open-shell situations often result in larger systematic errors. It is thus important to understand and develop ways to improve the quality of the constraint.
Symmetry properties and projection have a long history in quantum many body problems (see, e.g. Refs [@nuclearsym; @ImadaPRB2004; @PhysRevB.72.224518; @PhysRevB.72.085116; @PhysRevE.77.026705; @JPSJ.77.114701; @PhysRevB.85.245130]) In this paper, we discuss the imposition of symmetry properties and their effects in AFQMC calculations. It is shown that symmetry can often be rigorously preserved in the AFQMC framework despite the stochastic nature of the calculations, and taking advantage of this can lead to dramatic improvements. We address the issue from two aspects: symmetry preserved projection and the symmetry of the trial wave function.
In the symmetry preserved projection, we study the form of the Hubbard-Stratonovich (HS) transformation and its effect on the computational efficiency and accuracy. We focus on the comparison between charge-decompositions which preserve SU(2) symmetry and the standard form of Hirsch spin-decomposition. We discuss the optimal form of the charge decomposition which involves subtracting the proper “mean-field" background *prior to* making the HS transformation.
With a trial wave function which imposes exact symmetry properties, one is able to reduce the projection time as well as the statistical error bar. Symmetry alleviates the sign problem, since the convergence time to the ground state is reduced by ensuring the right symmetry in the trial wave function used to initiate the projection. Furthermore, the correct symmetry can help sustain projections of excited states which have different symmetry from that of the ground state. The most important feature of preserving symmetry in the trial wave function is in constrained AFQMC calculations, where we will show that the quality of the approximation improves significantly.
We consider the effect of symmetry in three different forms of ground-state AFQMC calculations, unconstrained Free Projection (FP), constrained path (CP) calculations, and released-constraint (RC) calculations. The first, FP, corresponds to the standard algorithm which is formally exact but has the sign problem. The third, RC, can be thought of in the present context as a more sophisticated version of FP, in which we start from a much better initial state, namely a converged CP run, and release (i.e., remove) the constraint and continue the projection. In FP and RC, the focus is on the use of symmetry to reduce the statistical fluctuations and prolong the imaginary-time projections. We show that improvements are seen in both methods. Especially in RC the use of symmetry can lead to a dramatic effects and significantly ameliorate the sign problem.
We are interested in studying the effect of symmetry in CP calculations, in order to further improve the constraints. We find that the use of trial wave functions which preserve symmetry can greatly reduce the systematic bias from the constraint. In open-shell systems, the use of simple trial wave functions which restore symmetry properties eliminates most of the bias from the constraint. Another key finding is that one can switch between different forms of the HS transformation in CP and the following RC calculations. This is motivated by considering the different behaviors of statistical errors and sensitivity to the constraint.
We will use the two-dimensional repulsive Hubbard model as an example in our discussions. The model provides an ideal test ground for the present study. On the one-hand, it is computationally simple where exact diagonalization results are available in non-trivial system sizes. On the other hand, it is a fundamental model which captures many of the key aspects of simulating correlated materials and which, in fact, contains many unanswered questions. Through the approach presented in this paper, significantly more accurate calculations can be done on larger systems. For example, essentially exact results can now be obtained for $> 100$ sites, as shown below.
The remainder of the paper is organized as follows. In section \[sec:AFQMC\], we summarize the relevant features of the AFQMC method in three different forms: the unconstrained FP, CP, and RC, in order to facilitate the ensuing discussions. In section \[sec:symmetry\], we describe separately the two aspects of symmetry improved AFQMC, namely symmetry-preserving HS transformation and symmetry in the trial wave function. Results are first shown for demonstration and benchmark, then followed by new, near-exact ground-state energies for larger system sizes. Finally we conclude in Sec. \[sec:conoutl\].
Auxiliary Field Quantum Monte Carlo Methods {#sec:AFQMC}
===========================================
We first summarize the key features of ground state AFQMC methods that are relevant to the symmetry study to follow. The three different forms of AFQMC are summarized in three subsections below. As mentioned, we will use the two-dimensional repulsive Hubbard model as a concrete example; however most of our discussions will apply to other Hamiltonians, including more realistic materials Hamiltonians.
The Hubbard model is written in second-quantized form as [ [@J.hubbard.1963]]{}: $$\hat{H}=\hat K +\hat V=-t\sum_{\langle i,j \rangle \sigma}^{L} c_{i\sigma}^{\dagger}c_{j\sigma}+U\sum_{i}^{L} n_{i\uparrow}n_{i\downarrow}\,.
\label{eq:HubHamil}$$ Here $L$ is the number of lattice sites, $c_{i\sigma}^{\dagger}$ and $c_{i\sigma}$ are creation and annihilation operators of an electron of spin $\sigma$ on the $i$-th lattice site, $t$ is the nearest-neighbor hopping energy and $U$ is the interaction strength. Throughout this paper, we will use $t$ as units of energy and set $t=1$. We assume that there are $N_{\uparrow}$ spin-up electrons and $N_{\downarrow}$ spin-down electrons on the lattice.
The Hamiltonian in Eq. (\[eq:HubHamil\]), whose Hilbert space grows exponentially in size with $L$, presents an enormous challenge. Questions remain open about its properties. Ground state projection with Monte Carlo (MC) sampling is one of the candidates to solve the problem. The projection method is: $$|\psi_{0}\rangle \propto \lim_{\beta\rightarrow\infty} e^{-\beta (\hat{H}-E_{T})} |\psi_{T}\rangle\,,$$ where $E_{T}$ and $|\psi_{T}\rangle$ are guesses of the ground state energy and wave function, and $\langle \psi_{0}|\psi_{T}\rangle\neq0$ in order for the projection to yield the ground state asymptotically. To target a lowest energy excited state of a different symmetry from $|\psi_{0}\rangle$ is similar to doing a ground-state calculation, except one would choose a $|\psi_{T}\rangle$ which is not orthogonal to the targeted excited state but satisfies $\langle \psi_{0}|\psi_{T}\rangle=0$.
The propagator can be evaluated using a Trotter-Suzuki breakup [ [@trotter1959; @MasuoSuzuki1976]]{}: $$(e^{-\varDelta \tau(\hat{K}+\hat{V})})^n=(e^{-\frac{1}{2}\varDelta \tau\hat{K}}e^{-\varDelta \tau\hat{V}}
e^{-\frac{1}{2}\varDelta \tau\hat{K}})^n+O(\varDelta \tau^2)\,.$$ Here we have $\varDelta \tau n=\beta$, and a Trotter error arises from the omission of the higher order terms. We will not be concerned with the Trotter error here, other than to note that it can be controlled by extrapolating to $\varDelta \tau\rightarrow0$ with separate calculations using different $\varDelta \tau$ values. In the results shown in this paper, we either perform such an extrapolation explicitly, or have checked via separate calculations that the Trotter error is within the statistical error. We also mention that, for the Hubbard interaction, decompositions without Trotter errors are possible [ [@PhysRevB.84.241110]]{}.
The two-body propagator is then decoupled into one-body propagator by auxiliary fields, using the HS transformation [ [@PhysRevLett.3.77; @PhysRevB.28.4059]]{}. The general form is: $$e^{-\varDelta \tau\hat{V}}=\sum_x p(x) e^{\hat{o}(x)}\,,
\label{eq:HS1}$$ where $\hat{o}(x)$ is a one-body operator that depends on the auxiliary field $x$, $p(x)$ is a probability density function with the normalization $\sum_{x} p(x)=1$. In general the sum in Eq. (\[eq:HS1\]) is an integral, and $x$ is a many-dimensional vector whose dimension is of the order of the size of the one-particle basis. In the Hubbard model, $x$ typically has $L$ components, one for each lattice site. By setting: $$\hat{B}(x)=e^{-\frac{1}{2}\varDelta \tau\hat{K}} e^{\hat{o}(x)} e^{-\frac{1}{2}\varDelta \tau\hat{K}},$$ we rewrite the projection as $$|\psi_{0}\rangle= \sum_{\overrightarrow{X}} P(\overrightarrow{X}) \prod_{i=1}^n \hat{B}(x_{i})|\psi_{T}\rangle\,.
\label{eq:projection0}$$ The vector$\overrightarrow{X}$ means $(x_{1},x_{2},\cdots,x_{n})$, and $P(\overrightarrow{X})=\prod_{i}p(x_{i})$.
Then the ground state properties can be evaluated by: $$\langle \hat{A}\rangle_{0}=\frac{\langle\psi_{0}|\hat{A}|\psi_{0}\rangle}{\langle\psi_{0}|\psi_{0}\rangle}\,,$$ which are many-dimensional integrals (e.g., $2nL$-dimensions in the Hubbard model). MC methods are used to calculate the high dimension integrals, by sampling the probability density function using the Metropolis algorithm[ [@metropolis:1087]]{} or a related method. The sign problem emerges because the integrand in the denominator, $P(\overrightarrow{X}) \langle \psi_{T}|\prod \hat B(x_i) | \psi_{T}
\rangle$ is not always positive, which causes the MC signal to be eventually lost in the sampling noise.
Unconstrained Free Projection (FP) {#ssec: method_FP}
----------------------------------
We carry out the FP calculation [@jacobi:43; @PhysRevLett.90.136401; @PhysRevB.80.214116] with an open-ended random walk similar to the CP approach [ [@PhysRevB.55.7464; @PhysRevLett.90.136401]]{}. However, the discussion on symmetry in FP calculations in this paper will apply directly to standard ground-state AFQMC calculations using Metropolis sampling of paths. In our approach, a population of $N_w$ random walkers is carried, which are typically initialized by the trial wave function. Each walker will have a weight $w$ whose value is set as one at the beginning of the projection: $$|\psi^{(0)}\rangle=\sum_{i}^{N_{w}} w_{i}^{(0)} |\phi_{i}^{(0)}\rangle
\label{eqn:fpwalk}$$
We apply the projection in Eq. (\[eq:projection0\]) by random walks in Slater determinant space, instead of using the usual approach of Metropolis sampling of auxiliary-field paths. In each step, we sample the auxiliary field $x$ according to $p(x)$ by MC, and apply $\hat{B}(x)$ to the Slater determinant wave function. Since the operators only contain one-body terms, they will generate another Slater determinant [ [@PhysRevB.41.11352]]{}: $$\begin{aligned}
|\psi^{(1)}\rangle &=&\sum_{i}^{N_w}\sum_{x_{i}} p(x_{i}) \hat{B}(x_{i}) w_{i}^{(0)} |\phi_{i}^{(0)}\rangle \nonumber \\
&=&\sum_{i}^{N_w} w_{i}^{(1)} |\phi_{i}^{(1)}\rangle
\end{aligned}$$
During the projection we multiply the constant (non-operator) values of the formula, e.g., the overall normalization $e^{\varDelta\tau E_{T}}$, to the weight of the walker. The weight of each walker will fluctuate in the random walk and after a few steps, some walkers can have large weights and some walkers will have small weights. We apply a population control procedure, splitting the walkers with large weights and eliminating walkers with small weights with the appropriate probability [ [@PhysRevB.57.11446]]{}, such that the overall probability distribution is preserved but the weights are made more uniform. It will introduce a population control bias [@PhysRevB.55.7464]. In our calculations, the population control bias is much smaller than the statistical error. This is easy to accomplish in CP calculations. In FP and RC calculations, large populations are necessary as discussed below. Modified Gram-Schmidt orthogonalization is applied to each walker periodically as well [ [@PhysRevB.40.506]]{}.
As mentioned, the FP calculations are done in the same framework as the CP calculations. This is different from standard Metropolis sampling [ [@BSS1PhysRevD.24.2278; @SorellaEUL; @ImadaJPSJ.65.189]]{}, which keeps an entire path of a *fixed length* $n$ as the object to sample. This framework does not have any ergodicity issues [ [@PhysRevB.44.10502]]{}, and it is straightforward to project to longer imaginary-time in order to approach the ground state. We typically turn off importance sampling in FP, sampling the fields according to $P(\overrightarrow{X})$ instead of using either the force bias [ [@PhysRevLett.90.136401; @PhysRevB.80.214116]]{} or direct importance-sampling of discrete fields [ [@PhysRevB.55.7464]]{} as is done in CP calculations. Empirically we find that this tends to give smaller statistical errors than invoking importance sampling and then lifting the constraint. The use of population control helps to reduce the noise but ultimately the shortcoming of this approach is that the lack of importance sampling will cause large noises as system size or $n$ increases. Since in these situations the sign problem, when uncontrolled, tends to overwhelm the calculation anyway, the shortcoming is not of major practical relevance, and we find this mode of sampling to often be the more efficient in practice.
At the $n^{\rm th}$ step in the propagation, we measure the energy by: $$E=\frac{\displaystyle\sum\limits_{i}^{N_w} w_{i}^{(n)}\langle \psi_{T} |H|\phi_{i}^{(n)}\rangle}
{\displaystyle\sum\limits_{i}^{N_w} w_{i}^{(n)}\langle \psi_{T}|\phi_{i}^{(n)}\rangle}\,.
\label{eqn:fpmear}$$ If the projection has equilibrated, we can combine the populations at multiple $n$ values in the estimator above to improve statistics on the ground-state energy. The energy measure in Eq. (\[eqn:fpmear\]) is variational if we set $|\psi^{(0)}\rangle=|\psi_{T}\rangle$, since $$E=\frac{ \langle \psi_{T} |H e^{-\beta H}|\psi_{T}\rangle}
{\langle \psi_{T}| e^{-\beta H}|\psi_{T}\rangle}=
\frac{ \langle \psi_{T} |e^{-\beta H/2 }H e^{-\beta H/2 }|\psi_{T}\rangle}
{\langle \psi_{T}| e^{-\beta H/2 }e^{-\beta H/2 }|\psi_{T}\rangle}\,.$$ To calculate the expectation value of an observable which does not commute with the Hamiltonian, we can use back-propogation [ [@PhysRevB.55.7464; @PhysRevE.70.056702]]{} using part of the path and projecting the trial wave function $\langle \psi_{T} |$. Because of the lack of importance sampling, back-propagation will tend to be very noisy in FP, and large population size will typically be needed.
The application of symmetry in either the trial wave function or the propagator $\hat B(x)$ is straightforward in FP. In Sec. \[sec:symmetry\] we show how the use of symmetry can improve FP calculations and allow longer projection time with smaller statistical fluctuations. Furthermore, some excited states can be accessed.
Constraint Path (CP)
--------------------
The constrained path (CP) approximation allows one to eliminate the sign problem present in FP. During the FP steps, the overlap between the ground state and the projected wave function, $\langle \psi_{0}|\psi^{(l)}\rangle$, will in general approach zero, because intrinsically the projection is symmetric about $|\psi^{(l)}\rangle$ and $-|\psi^{(l)}\rangle$. In other words, at any given imaginary time $l$, the projection would proceed identically if each random walker $|\phi^{(l)}\rangle$ were switched to $-|\phi^{(l)}\rangle$, for example by a permutation of two of its orbitals. This means that, unless the random walks are somehow strictly confined to only one kind of “sign”, it will invariably become a random and equal mixture of both, given sufficiently large $l$. Thus measurements from the MC sampling will eventually have infinite variance. This is how the sign problem appears in an FP calculation.
The CP approach is based on the observation that [ [@PhysRevB.55.7464]]{}, if any particular walker has the zero overlap with the ground state at imaginary time $\tau_l\equiv l\Delta \tau$ in the projection: $$\langle \psi_{0}|\phi^{(l)}\rangle=0$$ this walker will contribute zero at any future time $\beta> \tau_l$, because: $$\langle \psi_{0}|e^{-(\beta-\tau_{l})\hat{H}}|\phi^{(l)}\rangle=0\,.$$ Then we are able to discard the walker once its path reaches a point where the overlap becomes zero. With this constraint the sign problem is eliminated, and the projection will still lead to the exact ground state. However, we obviously do not know the exact ground state wave function. In CP calculations, a trial wave function, $|\psi_{T}\rangle$, is chosen for determining the sign of the overlap. A walker which develops a zero overlap with $|\psi_{T}\rangle$ during the projection is discarded.
Importance sampling can be introduced in CP calculations both as a natural way to impose the constraint and for variance reduction [ [@PhysRevB.55.7464; @PhysRevLett.90.136401; @PhysRevE.70.056702]]{}. With importance sampling, the wave function during the projection can be written as: $$|\psi^{(l)}\rangle=\sum_{i}^{N_w} w_{i}^{(l)} \frac{ |\phi_{i}^{(l)}\rangle}{\langle \psi_{T}|\phi_{i}^{(l)}\rangle}\,.
\label{eqn:cpwalk}$$ Instead of $p(x)$, one samples the auxiliary-fields from $$\widetilde{p}(x)=p(x) \frac{\langle \psi_{T}|\hat{B}(x)|\phi_{i}^{(l)}\rangle}{\langle \psi_{T}|\phi_{i}^{(l)}\rangle}\,,$$ which can be accomplished either directly for discrete fields using a heat bath-like approach [ [@PhysRevB.55.7464]]{} or more generally via a force bias [ [@PhysRevLett.90.136401; @PhysRevE.70.056702]]{}. This will automatically prevent the random walks from sampling any determinants with zero (or negative) overlap with the trial wave function. Those with larger overlap will be sampled more, although the weight from importance sampling will ensure that the exact distribution defined by Eq. (\[eqn:cpwalk\]) is sampled. The energy can be calculated by the mixed estimate similar to Eq. (\[eqn:fpmear\]), although now with importance sampling it has the form: $$E=\frac{\displaystyle\sum\limits_{i}^{N_w} w_{i}^{(n)} \frac{\langle \psi_{T} |H|\phi_{i}^{(n)}\rangle}{\langle \psi_{T}|\phi_{i}^{(n)}\rangle}}
{\displaystyle\sum\limits_{i}^{N_w} w_{i}^{(n)}}\,.
\label{eqn:CPmeas}$$ Following diffusion Monte Carlo (DMC), we refer to the quantity in the numerator, $E_L(\phi)\equiv \langle \psi_{T} |H|\phi\rangle/\langle \psi_{T}|\phi\rangle$ as the local energy. An important characteristic of the constrained path approximation is that the mixed estimate in Eq. (\[eqn:CPmeas\]) is not variational [@PhysRevB.59.12788].
The CP approximation has proved very accurate in the Hubbard model, especially for closed shell systems [ [@PhysRevB.55.7464]]{}. For instance, the energy at $U=4$ is typically within $< 0.5\%$ of the exact diagonalization result [@PhysRevB.78.165101]. It is, however, approximate. The systematic error in the energy tends to be larger in open-shell systems. Here we show that this error can be significantly reduced with trial wave functions which observe the correct symmetry.
Release Constraint (RC)
-----------------------
From a converged CP calculation, one can release the constraint and continue with the projection. Calculations of similar character have been done in the framework of DMC, under the name of released node [ [@ceperley:5833]]{}. Since the CP result is already very close to the ground state and FP in AFQMC tends to have a reduced sign problem in general, one can expect that releasing the constraint in AFQMC will be effective and will allow the removal or reduction of systematic bias in more systems.
In principle the idea of releasing the constraint is straightforward. As mentioned, the RC calculation can theoretically be viewed as an FP calculation, with a much better starting point. Technically, however, the implementation of RC can be challenging. The initial population, namely that from CP, is obtained with importance sampling, which automatically imposes the constraint. On the one hand, the importance function must be modified in RC to allow the random walks to have a significant chance to sample the region with $ \langle \psi_T|\phi^{(l)}\rangle$ being negative (or develop different phases in the more general case). Indeed, the more efficient approach in FP in AFQMC seems to be with no importance function, as discussed earlier. On the other hand, since the initial population from CP has the form of Eq. (\[eqn:cpwalk\]), “undoing" the importance sampling will cause large statistical fluctuations and accelerate the onset of the sign problem. Thus we need to find the best way to interface the RC part with CP so as to minimize the growth of statistical noise.
RC calculations within AFQMC have been published by Sorella [ [@PhysRevB.84.241110]]{} recently. Our approach is somewhat different and we will publish the method and further results elsewhere. The algorithmic details do not affect the issues that are the focus of the present paper, namely the use of symmetry properties to improve the calculations. A key aspect is our use of different forms of the HS transformation in the CP and the RC portions of the calculations. That is, we switch to a different HS decomposition in the RC in order to impose exact symmetry properties which drastically change the behavior of the RC calculations, as described below.
In RC calculations, we will use a mixed estimator similar to Eq. (\[eqn:fpmear\]) to measure the energy: $$E_{\rm RC}(\beta)=\frac{\displaystyle\sum\limits_{i}^{N_w} w_{i}^{\rm CP}\langle \psi_{T} |H e^{- \beta \hat{H}}|\phi_{i}^{\rm CP}\rangle}
{\displaystyle\sum\limits_{i}^{N_w} w_{i}^{\rm CP}\langle \psi_{T}|e^{-\beta \hat{H}}|\phi_{i}^{\rm CP}\rangle}\,.
\label{eq:E_mixed_RC}$$ Thus $E_{\rm RC}(\beta=0)=E_{\rm CP}$. As mentioned, the mixed estimate in CP is not variational. The RC energy will asymptotically converge to the exact ground-state energy. However, it can converge from below or above. Indeed, as we further discuss below, the convergence can be non-monotonic for poorer trial wave functions.
 Statistical error bar (log-scale) versus projection time for different HS transformations. FP calculations are shown. The error bars increase exponentially with projection time, but the optimal choice of the background $\bar{n}$ in Eq. (\[equ:bg-form\]) greatly reduces the fluctuation and improves efficiency. The system shown is a $4\times4$ lattice with $N_{\uparrow}=3$ and $N_{\downarrow}=3$, and $U=4$. ](\PLOTFILE{bg-error.eps})
Symmetry Improved AFQMC: Results and Discussion {#sec:symmetry}
===============================================
In this section we describe approaches to impose symmetry in AFQMC calculations, and study their effects on computational efficiency and, more importantly, on the sign problem. We divide the discussion into two parts: the choice of the Hubbard-Stratonovich transformation and its effect on symmetry, and symmetry in the trial wave function. While we will consider all three flavors of AFQMC introduced in Sec. \[sec:AFQMC\], our focus will be on CP and RC, since these are the most general methods which will allow calculations to scale to large system sizes. It is shown that the use of symmetry can lead to large reductions in statistical and systematic errors, and alleviate the sign problem.
Hubbard-Stratonovich Transformation {#ssec: symmetry_HS}
-----------------------------------
For each form of the two-body interaction, there are different ways to decompose the propagator, leading to different forms of Eq. (\[eq:HS1\]). Decompositions based on Hartree, Fock, and pairing mean-fields are all possible. Even within each mean-field framework, the details can affect the final form of the one-body propagator and the computational efficiency. When a constraint is imposed, the form of the HS transformation chosen can affect the systematic accuracy given a form of the trial wave function. For the Hubbard interaction, for example, the most commonly used HS transformation involves discrete auxiliary-fields, due to Hirsch [ [@PhysRevB.28.4059]]{}. The spin form of this decomposition is: $$\label{equ:hirsch-spin}
e^{-\varDelta \tau U n_{i\uparrow}n_{i\downarrow}}=e^{-\varDelta \tau U (n_{i\uparrow}+n_{i\downarrow})/2}
\sum_{x_{i}=\pm1}\frac{1}{2} e^{ \gamma x_{i} (n_{i\uparrow}-n_{i\downarrow})}\,,$$ which results in an Ising-like auxiliary-field for each lattice site. The constant $\gamma$ is determined by $$\cosh(\gamma)=e^{\varDelta \tau U/2}\,.
\label{eq:HS_posU_gamma}$$ It can also be mapped to a charge density form: $$e^{-\varDelta \tau U n_{i\uparrow}n_{i\downarrow}}=e^{-\varDelta \tau U (n_{i\uparrow}+n_{i\downarrow}-1)/2}
\sum_{x_{i}=\pm1}\frac{1}{2} e^{ \gamma x_{i} (n_{i\uparrow}+n_{i\downarrow}-1)}\,,
\label{eq:HS_negU_discrete}$$ with $$\cosh(\gamma)=e^{-\varDelta \tau U/2}\,.$$
A more general HS transformation [ [@PhysRevLett.3.77]]{}, $$\label{equ:con-hubb}
e^{\hat{A}^2}=\frac{1}{\sqrt{2\pi}} \int_{-\infty}^{\infty} e^{-x^2/2+\sqrt{2}x \hat{A}}dx,$$ applies to any two-body operators written in the form of a square. To apply this to the Hubbard interaction, we write: $$\label{equ:bg-form}
\begin{split}
n_{i\uparrow}n_{i\downarrow}=\frac{1}{2} [(n_{i\uparrow}+n_{i\downarrow}-\bar{n})^2-\bar{n}^2 \\
-(1-2\bar{n})(n_{i\uparrow}+n_{i\downarrow})
]\,,
\end{split}$$ where $\bar{n}$ can take any value (including acquiring a dependence on $i$). We then let $\hat A = \sqrt{-\Delta\tau U/2}\,(n_{i\uparrow}+n_{i\downarrow}-\bar{n})$ and use Eq. (\[equ:con-hubb\]) to obtain an HS transformation with continuous fields. The constant $\bar{n}$ can be thought of, physically, as a background term that one subtracts from the one-body operator *prior to* applying the HS transformation. This has been pointed out before for Hubbard interactions [@PhysRevA.72.053610] and for Coulomb interactions [@baer:6219; @al-saidi:224101]. As discussed below, the optimal choice for $\bar{n}$ is to remove all background interactions from the mean-field, by minimizing the quadratic first term on the right-hand side in Eq. (\[equ:bg-form\]). Over typically densities, this choice leads to significant improvement over other choices including the standard Hirsch discrete decomposition [ [@PhysRevB.28.4059]]{}.
 Illustration of the effect of preserving symmetry in the HS transformation: Hirsch spin (Eq. (\[equ:hirsch-spin\])) vs. Gaussian charge (Eqs. (\[equ:con-hubb\]) and (\[equ:bg-form\])). Panel (a) plots the energy from FP versus projection time, with the inset showing a magnified view of $\beta\in(3,3.5)$. Panel (b) shows the statistical error bar as a function of projection time in a semi-log plot. The system is the $8\times8$ Hubbard model, with $N_{\uparrow}=N_{\downarrow}=32$ and $U=8$. A Hartree-Fock trial wave function is chosen. The number of walkers was $10^5$, with a total 20 separate runs to obtain the final averages and estimate the error bars. ](\PLOTFILE{charge-vs-spin-4488.eps})
In Fig. \[fig:bg-error\], we illustrate the effect of the background $\bar{n}$ in the continuous charge decomposition of Eqs. (\[equ:con-hubb\]) and (\[equ:bg-form\]). The logarithm of the statistical error bar is plotted versus projection time for different values of the background $\bar{n}$. The calculations are all FP so the sign problem is present, as indicated by the growing error bars, which are essentially linear in the log-plot with projection time. It is seen that the minimum statistical error is achieved when $\bar{n}=\langle n_{i\uparrow}+n_{i\downarrow}\rangle_{\rm MF}=(N_{\uparrow}+N_{\downarrow})/L$. The efficiency of the HS decomposition decreases as $\bar{n}$ deviates from the optimal value. It is a symmetric function of the deviation: a background value which is larger or smaller than the optimal value by the same amount gives comparable results. We point out that, although we have illustrated this with the repulsive model, the same applies to the attractive case. For example, in dilute Fermi gas simulations, Eqs. (\[equ:bg-form\]) with a small $\bar{n}$ will be much more efficient than Eqs. (\[eq:HS\_negU\_discrete\]) which corresponds to $\bar{n}=1$ .
It is often thought that the use of a discrete HS field, compared to continuous fields, leads to significant performance advantages [@PhysRevB.33.3519]. This is not the case, as shown in Fig. \[fig:bg-error\]. The discrete charge decomposition of Eq. (\[eq:HS\_negU\_discrete\]) is shown in the figure. We see that it is almost the same as the continuous decomposition with $\bar{n}=1$. This is because the interaction term in the discrete charge decomposition is mapped to $(n_{i\uparrow}+n_{i\downarrow}-1)$, identical to the continuous transformation when $\bar{n}$ is set to $1$. The discrete decomposition is ideal near half-filling, but will be inefficient in dilute systems, for example, in Fermi gas simulations [ [@PhysRevA.84.061602]]{}.
 Comparison of discrete spin and Gaussian charge decomposition in the presence of a sign/phase problem. The logarithm of the statistical error bars from FP is plotted vs. projection time for a $4\times4$ lattice with $N_{\uparrow}=N_{\downarrow}=7$ and $U=8$. Symmetry multi-determinant trial wave function is used. The number of walkers is $5\times 10^5$, with total 100 separate runs to estimate the error bars.](\PLOTFILE{charge-vs-spin-4477-error.eps})
Because the decomposition in Eq. (\[equ:con-hubb\]) preserves SU(2) symmetry, it can be more efficient than the discrete spin decomposition of Eq. (\[equ:hirsch-spin\]), which is the most commonly used form in simulations of systems with repulsive interactions. This point is more subtle, however, as it is intertwined with the sign/phase problem. In Fig. \[fig:charge-spin-4488\], the two decompositions are compared in a situation free of the sign problem, namely the half-filled repulsive Hubbard model. It can be seen that the Gaussian charge decomposition leads to much smaller statistical fluctuations. This has also been pointed out by Meng [*et. al.*]{}[ [@Assadnature]]{}, using the discrete charge decomposition of Eq. (\[eq:HS\_negU\_discrete\]) at half filling.
In Fig. \[fig:charge-spin-4477-error\] we study the case when a phase problem is present (U$>$0, so $\hat{A}$ is imaginary): $4\times4$, with $N_{\uparrow}=N_{\downarrow}=7$ and $U=8$. This system has a severe sign problem for the discrete spin decomposition. The continuous Gaussian decomposition leads to a phase problem. The latter decomposition initially has smaller error bars, benefiting from the preservation of symmetry, but after some time, its error bars exceed that of the spin decomposition. So in systems with a sign/phase problem, the efficiency is a balance of two competing aspects. On the one hand, the charge decomposition has an advantage for preserving symmetry. On the other hand, the phase problem tends to result in fast deterioration of the statistical signal and is a disadvantage. Below we discuss how to exploit these characteristics in different calculations.
![\[fig:charge-spin-881313\] (Color online) Total energy versus projection time in FP calculations using discrete spin and continuous charge decompositions. The system is a $8\times8$ lattice under periodic boundary conditions, with $ N_{\uparrow}=N_{\downarrow}=13$ and $U=4$. A FE trial wave function is used. The number of walkers is $5\times10^5$, with total 60 runs to obtain the average and estimate the statistical error bars. The red horizontal line gives the final CPMC result. The inset shows a magnified view of the last part of the projection. The spin decomposition calculation provides a reliable upper bound to the ground-state energy.](\PLOTFILE{8-8-13-13.eps})
We can use the advantage of the spin decomposition in longer time projections, as shown in Fig. \[fig:charge-spin-881313\] in a FP calculation of the periodic $8\times8$ supercell with $N_{\uparrow}=N_{\downarrow}=13$ and $U=4$. The continuous charge decomposition has much larger noise and is not accurate enough. The discrete spin decomposition has rather small fluctuations and provides a useful estimate of an upper bound of the ground-state energy: $-66.855(2)$. The run took $\sim17$ hours on 100 AMD Opteron 2.4GHz cores. (For comparison, the CP calculation using a free-electron trial wave function gives $-66.857(2)$, as indicated by the line in Fig. \[fig:charge-spin-881313\], running for minutes on a single core. )
The charge decomposition can offer a significant advantage in short projection time, however. A main application of this is in RC calculations. Since we start from a population of a converged CP run, the initial state is close to the true ground state. One can expect a short projection in the RC calculation to recover a significant fraction of the correction to CP. In Fig. \[fig:charge-spin-4477-rcpmc\], we show an example RC calculation, for the same system as in Fig. \[fig:charge-spin-4477-error\]. In this calculation, the CPMC portion always uses the standard spin decomposition, which has a severe sign problem. As can be seen from the inset, the CP energy obtained from the mixed-estimate is not variational [ [@PhysRevB.59.12788]]{}. (Typical CP calculations will run to much larger $\beta$ than shown in the main figure, in order to collect statistics.) In the RC portion, two different calculations are shown, one continuing to use the discrete spin decomposition while the other switching to the symmetry charge-decomposition. It is seen that the latter leads to much smaller statistical fluctuations and allows the RC calculation to reach convergence. The RC with discrete spin decomposition has much larger errors, and also displays a population control bias [@PhysRevB.55.7464]. (The statistical error and the bias could, of course, be reduced by increasing the population size further.)
![\[fig:charge-spin-4477-rcpmc\] (Color online) Effect of symmetry decomposition in release constraint calculations. The main figure shows the convergence of CPMC energy with projection time, compared with the exact ground-state energy. The CPMC calculation uses the discrete spin decomposition. The inset shows RC calculations, starting from a converged CP state, using two different forms of the HS decomposition. The system is the same as in Fig. \[fig:charge-spin-4477-error\]: $4\times4$, $N_{\uparrow}=N_{\downarrow}=7$ and $U=8$, using a symmetry multi-determinant trial wave function. The number of walkers is $1\times 10^5$, with total 100 runs in the RC portion to collect statistics. ](\PLOTFILE{charge-vs-spin-4477-rcpmc.eps})
Symmetry of the Trial Wave Function {#ssec:symmetry_twf}
-----------------------------------
In this section we discuss the other aspect of symmetry in AFQMC calculations, the use of trial wave functions which preserve symmetry. We generate the trial wave function with particular symmetries: total spin $S^2$, total momentum $\overrightarrow{K}$, rotational symmetry in momentum space $R$, mirror reflection $\sigma$ along the line $K_y=K_x$ in momentum space. When the total momentum $\overrightarrow{K}=0$, we use the $C_{4v}$ point group irreducible representation to label different symmetry state. In the present paper, these properties are imposed in the trial wave function by a brute-force approach, making a linear combination of Slater determinants, as discussed in further details below.
### Trial Wave Function in FP Calculations
Imposing the proper symmetry in the trial wave function can accelerate convergence and reduce the equilibration time in the FP calculations. As mentioned in Sec. \[ssec: method\_FP\], the trial wave function is often also used to generate the initial population in FP. In all the FP calculations in this section, we use the HS transformation given in Eqs. (\[equ:con-hubb\]) and (\[equ:bg-form\]), using optimal background values from simple mean-field calculations.
![\[fig:symmetry-fp\] (Color online) The effect of symmetry trial wave functions in FP calculations, compared to exact results in $4\times4$ lattices with $U=4$: (a) $N_{\uparrow}=N_{\downarrow}=8$, no sign problem; (b) $N_{\uparrow}=N_{\downarrow}=7$, severe phase problem. The UHF trial wave function is generated by a UHF calculation with $U=0.5$, $0s0kx0ky1D$ is a single-determinant non-interacting trial wave function with $S^2=0$ and $k_x=k_y=0$, and $0s0kx0kyMD$ is a multi-determinant trial wave function which has rotational symmetry in momentum space in (a) and B1 symmetry in (b), in addition to $S^2=0,k_x=0,k_y=0$. ](\PLOTFILE{symmetry-fp.eps})
We illustrate the effect of the symmetry in the trial wave function in Fig. \[fig:symmetry-fp\]. In the top panel, we consider a half-filled system which is thus sign-problem-free. The effects of three trial wave functions are compared: the unrestricted Hartree-Fock (UHF) wave function, a single-determinant trial wave function formed by occupying $k$-states, and a multi-determinant trial wave function which preserves additional symmetry. The UHF trial wave function builds in correlation effect via its static anti-ferromagnetic order, and is an excellent starting point, as can be seen from the variational energy values at the beginning of the projection. However, it is contaminated by higher spin eigenstates, and the FP with UHF exhibits a long convergence time, as seen in the inset. (The effect of spin-contamination in AFQMC calculations has been discussed in continuum systems by Purwanto [*et. al.*]{}[ [@purwanto:114309]]{}.) The single-determinant FE trial wave function has a very high variational energy. Its statistical error bars are larger since the $|\psi_{T}\rangle$ in Eq. (\[eqn:fpmear\]) to evaluate the energy is much poorer, and preserves fewer symmetry properties. However, it eventually leads to a faster convergence than the UHF trial wave function because symmetry has properly removed certain excitations. The multi-determinant trial wave function with symmetry, which we obtain by diagonalizing in the subspace of the open-shell \[in the spirit of a small complete active-space self-consistent field (CASSCF) calculation\], leads to rapid convergence and small statistical errors. In the bottom panel, Fig. \[fig:symmetry-fp\]b, we show an example when there is a severe sign/phase problem. The same trends are seen, with the full symmetry trial wave function leading to rapid convergence of the projection.
With the mixed estimate, the symmetry trial wave function on the left will have zero overlap with any wave function component in a different symmetry space. This allows one to project out explicitly lower energy states of different symmetry, and thus an opportunity to study excited states. This has been used in QMC calculations before. In the AFQMC formalism, the walkers are full Slater determinants, so that the symmetry projection can be done rigorously and explicitly for each walker. We illustrate excited state calculations in Fig. \[fig:symmetry-excit\], where the converged FP results show excellent agreement with results from exact diagonalization.
 FP calculations for the ground state and three excited states. The energy is plotted versus projection time for a $4\times4$ lattice with $N_{\uparrow}=N_{\downarrow}=3$ and $U=4$. The exact results for the ground state and the three excited state energies are shown for comparison. The symmetry of each energy level is labeled. The trial wave functions are chosen with the correct symmetry using multi-determinant. The error bars are shown but are smaller than symbol size at most points. ](\PLOTFILE{4-4-3-3-4U-FP-excite-MDtw.eps})
### Trial Wave Function in CP Calculations
In CPMC, the sign or phase problem is controlled by the sign or gauge condition of the overlap with the trial wave function. The condition is approximate, and the resulting systematic error depends on the trial wave function. Thus the trial wave function has an especially important role in CP calculations. In this section, we study how trial wave functions which preserve symmetries impact the accuracy in CP calculations. As shown in prior studies [ [@PhysRevB.55.7464; @PhysRevB.78.165101]]{}, the systematic error from the constraint is in general small for Hubbard-like systems, even when an FE or UHF trial wave function is used. In closed-shell systems in particular, the error is often negligible, as seen in the example of $8\times 8$ system with $26$ electrons in Sec. \[ssec: symmetry\_HS\]. The systematic errors tend to be larger for open-shell systems. As we show below, the leading reason for the problem in open-shell systems seems to be symmetry in the trial wave function. The use of trial wave functions with proper symmetry often leads to a dramatic reduction in the CP error.
We use the discrete spin decomposition in Eq. (\[equ:hirsch-spin\]) in the CP calculations, which causes “only” a sign problem, even when a twist angle is applied in the boundary condition of the supercell [ [@PhysRevB.78.165101]]{}. We first focus on small system sizes where exact results are available to make detailed and systematic comparison. Larger systems are treated later, and compared with our best results from RC calculations. In Fig. \[fig:cpmc-symm-var-u\], we study the systematic error in the case of $4\times4$ with $N_{\uparrow}=N_{\downarrow}=7$, which has the most severe sign problem in systems that can be diagonalized presently. We study the systematic error as $U$ goes from $0$ to $12$, spanning weak to moderate to strong interactions. The CP results with an FE trial wave function is shown. (We use a small twist of opposite sign for $\uparrow$ and $\downarrow$ spins to generate the FE trial wave function, which breaks the SU(2) symmetry, but has translational symmetry and $\overrightarrow{K}=0$.) The CP systematic error tends to grow with $U$, reaching about $2$% of the total energy, or about $1$% of the correlation energy at $U=12$. As we see, the use of symmetry wave functions (obtained by diagonalizing the open-shell, leading to a total of 10 determinants) makes the CP systematic error very small across the range of $U$. Figure \[fig:charge-spin-4477-rcpmc\] contains a zoomed-in view, at $U=8$, of the CP/SYM run and the subsequent RC which leads to an exact result. In Table \[tab:cpmctwf\], we compile the results from a variety of systems where exact diagonalization can be done to provide a quantitative measure. The CP results are compared for FE (or UHF solution obtained from a weak $U$) trial wave functions and symmetry trial wave functions. Significant improvement is seen in open-shell systems, and accurate results are obtained from CP calculations.
 Accuracy of CPMC in the Hubbard model as a function of interaction strength. Results are shown for the $4\times4$ lattice with $N_{\uparrow}=N_{\downarrow}=7$ as a function of $U$, and compared with exact diagonalization. When the trial wave function preserves symmetry, the systematic bias in the calculated energy from the CP approximation is reduced.](\PLOTFILE{4-4-7-7-PBC-U-0-12-1.eps})
In Table \[tab:cpmctwf\], we have included a set of results for $4\times4$ and $N_{\uparrow}=N_{\downarrow}=8$. There is no sign problem at half-filling, As has been discussed before [ [@PhysRevB.55.7464; @PhysRevB.78.165101]]{}, the CP calculations can be easily made exact at half-filling (or for negative $U$ [ [@PhysRevA.84.061602]]{}) by re-defining the importance sampling to have a non-zero minimum. However, if this were ignored and the CP algorithm applied to half-filling literally, an artificial constraint would result because the random walk cannot tunnel from one side of $\langle \Psi_T|\phi\rangle=0$ to the other, even though both sides are positive. The calculated energy would then show a bias, which is visible in the results shown in the table. With symmetry trial wave functions, this bias is removed even when running CP unmodified, and the CP results at half-filling are accurate.
The improvement of CP calculations with the symmetry trial wave function is not just to the ground-state energy. The CP bias in the observable is also significantly reduced. An example is shown in Fig. \[fig:sk-with-sym\], in which we calculate the structure factor of the spin-spin correlation function in the ground state: $$S(K)= 1/N\sum_{ij}S^{z}_{i}S^{z}_{j} \exp[\imath K ({R_{i}-R_{j}})]\,.$$ As mentioned earlier, we use the back-propagation technique [ [@PhysRevB.55.7464; @PhysRevE.70.056702]]{} to calculate correlation functions. The result is plotted for the same $4\times 4$ systems for three different values of $U$. The peak at $(\pi,\pi)$ indicates strong anti-ferromagnetic correlations. We see that the CP result using UHF trial wave function shows a larger anti-ferromagnetic order, because the UHF state itself over-estimates the order. The symmetry trial wave function removes the bias and leads to results in agreement with exact diagonalization.
![\[fig:sk-with-sym\] (Color online) The structure factor $S(k)$ of the spin-spin correlation function for three interaction strengths. The system is $4\times4$ with $N_{\uparrow}=N_{\downarrow}=7$, and the horizontal axis labels of $K$ are in units of $\pi/2$. The symmetry trial wave function has $S^2=0$ and $K_{x}=K_{y}=0$ and B1 symmetry. CPMC has $10,000$ walkers, with back-propagation $\beta=1$. ](\PLOTFILE{sk-vU.eps})
L ($N_{\uparrow}, N_{\downarrow}, U$) CP/FE CP/SYM RC/SYM ED ($S^2, K_x, K_y$)
------------- ------------------------------------- ------------- ------------- ------------- ---------- -------------------
$2\times2$ (2,1,4) -1.60564(5) -1.60615(3) -1.60465(5) -1.60463 (0.75,0,1)
$2\times3$ (2,2,4) -1.38328(6) -1.40129(4) -1.40085(4) -1.40087 (2,0,0)
$2\times3$ (2,2,8) -1.2239(2) -1.2463(1) -1.2443(3) -1.2442 (2,0,0)
$2\times4$ (2,2,4) -1.36839(2) -1.37387(2) -1.37379(3) -1.37383 (0,0,2)
$2\times4$ (3,3,4) -1.56939(4) -1.56942(4) -1.56944(5) -1.56941 (0,0,0)
$3\times3$ (4,4,8) -0.7783(2) -0.8127(1) -0.8091(1) -0.8094 (0,0,0)
$4\times4$ (2,2,4) -0.72026(5) -0.72094(1) -0.72063(1) -0.72064 (0,0,0)
$4\times4$ (2,2,8) -0.7070(1) -0.7082(1) -0.7075(2) -0.7076 (0,0,0)
$4\times4$ (2,2,12) -0.6997(1) -0.7010(1) -0.7002(3) -0.7003 (0,0,0)
$4\times4$ (3,3,4) -0.93394(2) -0.94622(1) -0.94598(1) -0.94600 (6,0,0)
$4\times4$ (3,3,8) -0.9034(1) -0.9208(1) -0.9203(1) -0.9202 (6,0,0)
$4\times4$ (3,3,12) -0.8867(1) -0.9067(1) -0.9062(3) -0.9061 (6,0,0)
$4\times4$ (4,4,4) -1.09442(2) -1.09693(2) -1.09597(6) -1.09593 (0,0,0)
$4\times4$ (4,4,8) -1.0265(1) -1.0307(1) -1.0282(5) -1.0288 (0,0,0)
$4\times4$ (4,4,12) -0.9914(1) -0.9962(1) -0.9940(3) -0.9941 (0,0,0)
$4\times4$ (5,5,4) -1.22368(2) -1.22368(2) -1.22380(4) -1.22381 (0,0,0)
$4\times4$ (5,5,8) -1.0948(1) -1.0948(1) -1.0942(2) -1.0944 (0,0,0)
$4\times4$ (5,5,12) -1.0292(1) -1.0292(1) -1.0278(4) -1.0284 (0,0,0)
$4\times4$ (6,6,4) -1.1012(1) -1.1104(1) -1.1084(2) -1.1080 (0,0,0)
$4\times4$ (6,6,8) -0.9293(1) -0.9376(1) -0.9329(5) -0.9328 (0,0,0)
$4\times4$ (6,6,12) -0.8439(1) -0.8557(1) -0.8507(6) -0.8512 (0,0,0)
$4\times4$ (7,7,2) -1.19584(2) -1.19992(1) -1.19822(2) -1.19821 (0,2,2)
$4\times4$ (7,7,4) -0.9793(1) -0.9863(1) -0.9840(1) -0.9840 (0,0,0)
$4\times4$ (7,7,6) -0.8334(1) -0.8428(1) -0.8386(3) -0.8388 (0,0,0)
$4\times4$ (7,7,8) -0.7361(1) -0.7461(1) -0.7417(8) -0.7418 (0,0,0)
$4\times4$ (7,7,10) -0.6687(2) -0.6782(1) -0.673(2) -0.6754 (0,0,0)
$4\times4$ (7,7,12) -0.6202(2) -0.6296(2) -0.627(4) -0.6282 (0,0,0)
$4\times4$ (8,8,4) -0.84225(6) -0.85140(6) -0.85133(6) -0.85137 (0,0,0)
$4\times4$ (8,8,8) -0.5164(2) -0.5293(2) -0.5291(2) 0.5293 (0,0,0)
$4\times4$ (8,8,12) -0.364(3) -0.3741(2) -0.3739(4) -0.3745 (0,0,0)
### Trial Wave Function in RC Calculations
Formally the role of symmetry in the trial wave function in RC calculations is similar to that in FP. However it is intimately connected to the discussion in the previous section on CP, since the initial state in RC is the converged solution from CP. The symmetry trial wave functions improves the CP approximation and the quality of the wave function sampled from CP, as indicated by the improvement in the energy and in the calculated observables. This means symmetry trial wave functions also allow better RC calculations, by providing a better initial state *and* by giving a better trial wave function in the mixed estimate in Eq. (\[eq:E\_mixed\_RC\]). As discussed in Sec. \[ssec: symmetry\_HS\], we also impose symmetry with the HS transformation in RC, by switching from the Ising spin form in the CP calculation to the Gaussian form in the RC part. We find this combination to improve the quality of the RC calculations greatly in most cases. An example is shown in Fig. \[fig:charge-spin-4477-rcpmc\]. Results from RC/SYM calculations are also shown in Table \[tab:cpmctwf\] for systematic comparisons with CP and with exact diagonalization results.
Figure \[fig:sym-release\] illustrates the behavior of RC calculations using two different trial wave functions, the UHF versus a symmetry trial wave function. A small system size of $3\times3$ with $2\uparrow$ and $2\downarrow$ electrons is chosen such that the RC calculation can also be carried out explicitly to allow direct comparison. (In the explicit calculation, we propagate the CP population of $\{ |\phi_i^{\rm CP}\rangle \}$ directly by applying $e^{-\Delta \tau \hat H}$. The propagation is carried out by expanding each walker in terms of exact eigenstates of $\hat H$.) We see that CPMC/UHF gives an energy closer to the exact value ($\sim 0.1$% error) compared to CPMC/SYM ($\sim - 0.3$% error). The corresponding RC/UHF moves further away from the exact answer and shows no indication of convergence in the imaginary-time span in which RC/SYM is well-converged. The explicit RC calculation, as shown in the inset, reveals a highly non-monotonic behavior. The projection does converge to the correct ground-state energy, but requiring an imaginary time of $>100$. This would be impossible to reach in a QMC RC calculation because of the sign problem. Thus non-monotonic behaviors could be difficult to detect and would yield misleading results. The improvement with the symmetry trial wave function, which leads to rapid and monotonic convergence, is then especially valuable.
![\[fig:sym-release\] (Color online) RC calculations with symmetry trial wave functions and without. The system is $3\times 3$ with $N_{\uparrow}=
N_{\downarrow}=2$ and $U=4$. The symmetry trial wave function has $S^2=2,K_{x}=0,K_{y}=0$ while the UHF wave function breaks these symmetries. CP/UHF is very accurate, but RC/UHF has non-monotonic behavior and slow convergence, as shown by the explicit propagation. RC/SYM converges rapidly and monotonically. The explicit propagation (EP) result of RC/UHF is shown to large projection time in the inset. ](\PLOTFILE{sym-release.eps})
The use of proper symmetry can allow RC calculations of excited states, similar to the discussion in Fig. \[fig:symmetry-excit\]. Since CP allows one to start from an initial population much closer to the exact exact state, RC can be more accurate. An example is shown in Fig. \[fig:4-4-5-5-struct\], in which the many-body ground state and first excited state energies are calculated as a function of crystal momentum. Both CP and RC are done with the same trial wave function, in which the correct symmetry is imposed. Consistent with prior experience, CP is very accurate for the ground state, although systematic error is visible at larger twist angles. The CP result is less accurate for the excited state. With RC, the CP error is removed and the results are seen to be essentially exact.
 RC and CP results for ground and the first excited state energies versus crystal momentum. RC greatly improves the calculation of excited states and band structures. The system is $4\times 4$ with $N_{\uparrow}=
N_{\downarrow}=5$ and $U=4$. QMC statistical error bars are smaller than symbol size. The horizontal axis gives $|k|$ along a line cut $k_y=2k_x$. Exact diagonalization (ED) results are shown for comparison. The line is to aid the eye. ](\PLOTFILE{ED-RCPMC.eps})
L ($N_{\uparrow}$ ,$N_{\downarrow}$) CP/UHF CP/SYM RC/SYM
-------------- ------------------------------------ ------------- ------------- ------------- --
6$\times$6 (12 ,12) -1.18444(3) -1.18625(3) -1.18525(4)
6$\times$6 (24 ,24) 0.14889(2) 0.14709(3) 0.14809(4)
8$\times$8 (14 ,14) -1.07173(1) -1.07239(1) -1.07180(2)
8$\times$8 (22 ,22) -1.18580(2) -1.18673(2) -1.1858(2)
10$\times$10 (40,40) -1.11378(3) -1.11468(2) -1.1135(2)
12$\times$12 (58,58) -1.10912(3) -1.11015(3) -1.1089(2)
: \[tab:biglattice\] Energy per site in some open-shell Hubbard square lattices at $U=4$. The symmetry of the ground state, $S^2=0,K_x=K_y=0$, is preserved in the trial wave function used in CP/SYM and RC/SYM. The UHF trial wave function is generated with $U=0.5$ as has been done before [@PhysRevB.55.7464].
Discussion and Conclusion {#sec:conoutl}
=========================
In this paper, we have studied the role of symmetry in AFQMC calculations, and discussed the imposition of symmetry from two key aspects of an AFQMC calculation, namely the HS transformation and the trial wave function. It is shown that major improvements in efficiency and accuracy can be achieved. To allow detailed and systematic benchmark and analysis, we have used smaller lattice sizes extensively, where exact results are available. The method applies straightforwardly to larger systems. CP calculations will scale as a low power with system size; FP and RC will of course have a sign/phase problem, albeit at a much reduced level with the proposed symmetry improvements. In Table \[tab:biglattice\], we present CPMC energies in several open-shell systems up to $12\times 12$. (For closed shell systems, the single-determinant FE trial wave function already satisfies the symmetries. They are expected to be very accurate. We have verified this in several cases with RC runs). CP/SYM and the corresponding RC results are shown. The RC results in Table \[tab:biglattice\] are essentially exact ground-state energies. CP/UHF results are also shown, which are comparable and even closer to the exact answer than the CP/SYM using simple symmetry trial wave functions in these systems. (As discussed, the CP/SYM leads to much better convergence in RC calculations.) This confirms the accuracy of CP/UHF as has been previously asserted.
We have discussed the general continuous Gaussian charge decomposition, which preserves spin symmetry. It is shown that the proper choice of the background $\bar {n}$ can lead to large reduction of the statistical fluctuation. One advantage of the spin decomposition is that, for repulsive interactions, it results in a sign problem in contrast with a phase problem for the charge decomposition. CP calculations with the spin decomposition tend to perform much better. We have emphasized the idea that the different forms can have advantages in different situations. Generally, the merits of the discrete spin and continuous charge, or indeed other forms of the HS transformation, will depend on the actual problem and physics. However, preserving the right symmetry is highly valuable, as we have demonstrated. Especially worth noting is that the switch to charge-decomposition within a CP calculation of spin-decomposition makes much better RC performance over many parameter ranges.
We have shown the importance of having trial wave functions which preserve symmetry. These trial wave functions accelerate convergence, allow better calculations of excited states, can significantly reduce the CP systematic error, and make possible systematically improvable RC calculations. The approach we have taken to generate trial wave functions that preserve symmetry, the equivalent of a small CASSCF calculation in quantum chemistry, has provided a proof-of-concept. In addition, they have already allowed calculations in significant system sizes at physically important regimes (low doping), as shown in Table \[tab:biglattice\]. However, for general open-shell situations, the CASSCF approach does not scale well. The resulting number of Slater determinants in the trial wave function will grow rapidly with system size. Several alternatives are possible, including projected BCS wave functions [ [@PhysRevA.84.061602]]{} and projected Hartree-Fock wave functions [ [@PhysRevB.85.245130]]{}.
The development presented in this paper will allow many applications even in its current form. Although we have focused on zero-temperature methods, many of the ideas will also apply to finite-temperature calculations [@BSS1PhysRevD.24.2278; @PhysRevLett.83.2777]. The formula can be directly mapped from the U$>$0 case we discussed to U$<$0 with a particle-hole transformation. Indeed the principle works for any other two-body interactions. The CP/SYM calculations and the RC/SYM from it represent a major step forward, as we can now have internal checks and a systematically improvable computational method capable of reaching two- and three-dimensions and large system sizes.
We thank S. Chiesa, H. Krakauer, F. Ma, W. Purwanto, G. Scuseria, and C. Umrigar for helpful discussions. This research was supported by DOE (Grant No. DE-SC0008627) and NSF (Grant No. DMR-1006217). We also acknowledge a DOE CMCSN award for facilitating stimulating interactions, and an INCITE award for computing using resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new Monte-Carlo methodology to forecast the crude oil production of Norway and the U.K. based on a two-step process, (i) the nonlinear extrapolation of the current/past performances of individual oil fields and (ii) a stochastic model of the frequency of future oil field discoveries. Compared with the standard methodology that tends to underestimate remaining oil reserves, our method gives a better description of future oil production, as validated by our back-tests starting in 2008. Specifically, we predict remaining reserves extractable until 2030 to be $188 \pm 10$ million barrels for Norway and $98 \pm10$ million barrels for the UK, which are respectively $45\%$ and $66\%$ above the predictions using the standard methodology.'
address: 'Chair of Entrepreneurial Risks, ETH Zürich, Scheuchzerstrasse 7 (SEC F), CH-8092 Zürich, Switzerland'
author:
- 'L. Fiévet'
- 'Z. Forró'
- 'P. Cauwels'
- 'D. Sornette'
bibliography:
- 'bibliography.bib'
title: 'Forecasting future oil production in Norway and the UK: a general improved methodology'
---
Introduction
============
Forecasting future oil production has been a topic of active interest since the beginning of the past century because of oil central role in our economy. Its importance ranges from energy production, manufacturing to the pharmaceutical industry. As petroleum is a non-renewable and finite resource, it is primordial to be able to forecast future oil production. The fear of a global oil peak, beyond which production will inevitably decline, has been growing due to stagnating supplies and high oil prices since the crisis in 2008/2009 [@Murray2013]. As any industrialized country, Europe is strongly dependent on oil supply to maintain its economic power. In the nowadays difficult geopolitical environment, it is important to know how much of the oil needed in Europe will come from reliable sources. In the past, a big share has been coming from Norway and the U.K., two of Europe’s biggest exporters. However, the U.K. already became a net importer in 2005 and Norways production has been declining rapidly as well [@Hook2008].\
The methodology behind forecasting future oil production has not evolved much since M. King Hubbert, who in 1956 famously predicted that the U.S. oil production would peak around 1965-1970 [@Hubbert1956]. That prediction has proven itself to be correct. His main argument was based on the finiteness of oil reserves and to what amounts as the use of the logistic differential equation for the total quantity $P(t)$ of oil extracted up to time $t$ $$\frac{dP}{dt}=rP\left(1-\frac{P}{K}\right)~.
\label{logistic}$$ The logistic differential equation is characterized by an initial exponential growth, which then decreases to zero as the total oil extracted reaches saturation (no more oil is to be found). The parameter $r$ is commonly referred to as the growth rate, and $K$ as the carrying capacity (total quantity of oil that can be ultimately extracted). If $P(t)$ is the amount of oil extracted up to time $t$, then $f(t) := \frac{dP}{dt}$ is the oil production rate, the quantity that M. King Hubbert predicted with surprising accuracy to peak. From a methodological point of view, the Hubbert model has enjoyed a longstanding popularity in modeling future oil production given its simplicity. Various extensions have been studied by @Brandt2007 to account for multi-cycled or asymmetric production curves. The existing forecasts of future oil production use some form of the Hubbert model [@Brecha2012; @Laherrere2002; @Lynch2002] or some economical model applied to aggregate production [@Greiner2011], but none goes into the details of studying the underlying dynamics. The main reason for the lack of details is certainly the lack of available data.
In this article, a new methodology is introduced to forecast future oil production. Instead of taking the aggregate oil production profile and fitting it with the Hubbert curve or its variants (such as the multi-cyclic Hubbert curve), the production profile of each individual oil field is used. By extending their production into the future and extrapolating the future rate of discovery of new fields, the future oil production is forecasted by means of a Monte Carlo simulation. To demonstrate the generality of the methodology presented here, it is applied to two major oil producing countries with publicly available data: Norway and the U.K. [@GOVUK2014].
Methodology
===========
The idea behind the methodology is to model the future aggregate oil production of a country by studying the production dynamics of its individual constituents, the oil fields. The main benefit of this approach, compared to working directly with aggregate production data, is the possibility to forecast non-trivial oil production profiles arising from the combination of all the individual field dynamics. This ability reflects directly the fact that the total production is the sum of the contributions of each individual oil field in production. Thus modelling at the level of each field reflects more closely the reality and is likely to be over-performing and more reliable, as we show below. In order to implement our approach, one must be able 1) to **extend the oil production of each individual field** into the future and 2) to **extrapolate the rate of discoveries** of new oil fields.
Extending the oil production of individual fields
-------------------------------------------------
The first step to predicting the future oil production of a country is to extrapolate the future production of existing fields and to estimate the error on this extrapolation. The data of the fields developed in the past shows a repeating asymmetric pattern. A good example is the Oseberg field shown in figure \[regular\], with a quick ramp up once the field is being developed, and then a peak or plateau before the oil field production starts decaying. The decay can take many different shapes and is governed by a variety of geological and economical factors. The goal of the fitting procedure is to capture as much the impact of these different factors as possible.
### Regular, irregular and new fields\[sub:Regular,-irregular-and\]
To be able to forecast the oil production of each individual field, regularity had to be found in the production’s dynamics. Modeling the whole production profile from the beginning of extraction seems elusive due to the variety of the forms it can take. Fortunately, modeling the decay process is sufficient in order to extrapolate future oil production. A preliminary classification is necessary to achieve that goal. Figures \[regular\], \[irregular\] and \[new\] show that, independent of the country, oil fields can be classified into three main categories:
- **Regular fields** - Their decays show some regularity (see figure \[regular\] );
- **Irregular fields** - The ones that do not decay in a regular fashion (see figure \[irregular\]);
- **New fields** - The ones that do not decay yet. As such, there is no easy way to forecast their future oil production based on past data (see figure \[new\]).
All the fields have been fitted using an automated algorithm, but the results have been subsequently checked visually to sort out the irregular fields which could not be fitted. As of January 2014, regular fields make up 85% and 87% of the number of fields and 94% and 71% of the total produced oil volume in Norway and the U.K. respectively. As such, being able to model them is crucial. To capture as many different decay dynamics as possible, the decay part of the oil production rate $f(t) := \frac{dP}{dt}$ has been fitted by the stretched exponential $$f(t)=f_{0}~e^{-\left(\frac{t}{\tau}\right)^{\beta}}~.
\label{stretched_exponential}$$ The stretched exponential function has many advantages as it generalises the power law and can therefore capture a broad variety of distributions as shown by @Laherrere1998. Moreover, @Malevergne2005 showed that the power law function can be obtained as an asymptotic case of the stretched exponential family, allowing for asymptotically nested statistical tests. As can be seen in figure \[regular\], the stretched exponential (equation \[stretched\_exponential\]) is a good functional form to fit the decay process of regular fields.
For the minority of irregular fields, we assume no difference in the decay of production between giants and dwarfs, which has been modelled as follows: the decay time scale $\tau$ has been picked to be the average $\tau$ over the regular fields. Then $\beta$ has been fixed so that the sum of the field production over its lifetime be equal to the official ultimate recovery estimates, when such an estimate is available.
The minority of new fields, which did not yet enter their decay phase, cannot be extrapolated and will therefore be treated as new discoveries. The technical details of how to treat them as new discoveries are discussed in section \[sub:Future-production-from\].
### Back-testing & Error
To determine how well the extrapolation based on the stretched exponential predicts the future production, a complete back-testing has been performed on each field. A single back-test is made as follows:
- The production data $\{p_{0},\,\ldots,\, p_{N}\}$ of the field is truncated at a certain date in the past $T \in\{0,\,\ldots,\, N\}$, where $T$ is the time counted in months since the production start of the field.
- The extrapolation of the oil production rate $f(t) := \frac{dP}{dt}$ is made based on the truncated data $\{p_{0},\,\ldots,\, p_{T}\}$.
- The future production predicted by the extrapolation function $f(t)$ can be compared to the actual production from the date $T$ in the past up to the present $T_{f}=N$. The extrapolated total production can be computed as $$P_{e}(T)=\int_{T}^{T_{f}}f(t)dt$$ and the relative error is given by $$e(T)=\frac{P_{e}(T)-\sum_{i=T}^{T_{f}}p_{i}}{P_{e}(T)},
\label{ryjryukoik}$$ where both $P_{e}(T)$ and $e(T)$ are functions of the truncation time $T$.
Computing this back-test, for every month in the past since the field production started decaying, yields a plot showing the evolution of the relative error over time defined by expression (\[oesberg-error\]). By construction, the relative error will tend to zero as the truncation time $T$ approaches the present. Nonetheless, it is a useful indicator for the stability of the extrapolation. As can been seen for the Oseberg field in figure \[oesberg-error\], the relative error on future production remained fairly stable during the past decade.
From the complete back-test, we compute the average relative error $$\bar{e}=\frac{1}{N}\sum_{i=0}^{N}e(i)$$ of the extrapolation made on the future production. Assuming that the relative errors are normally distributed around the average relative error, the standard deviation on the average relative error is given by $$\sigma_{e}=\sqrt{\frac{1}{N}\sum_{i=0}^{N}\left(e(i)-\bar{e}\right)^{2}}.\label{eq:field-risk}$$ As the average relative error is often fairly constant, the extrapolation was corrected by the average relative error, that is, if the extrapolation consistently over-estimated the production by 10% during the back-test, the extrapolation was reduced by 10%. This results in an extrapolated production $p(t)$, including a 1$\sigma$ confidence interval, given by $$p(t)=(1-\bar{e})\, f(t)\pm\sigma_{e}.
\label{eq:extrapolation}$$ An example of such an extrapolation including a one standard deviation range is shown in figure \[regular\] for the Oseberg field.
### Aggregate error
Once the individual fields have been extrapolated using formula \[eq:extrapolation\], we compute the extrapolation of the oil production for the whole country. While it is straightforward to sum the extrapolations of the individual fields to obtain the expected production, some care has to be taken with respect to the confidence interval of the production at the country level.
As shown later in section \[sec:Results\], the same extrapolation including a complete monthly back-test of total future production has been performed at the country level and the resulting relative error is much smaller than the average error observed on the individual fields. To account for this observation, the assumption made is that the relative error between individual fields is uncorrelated. While one could imagine that some inter-dependence could result from a coordinated response of supply to a sharp increase/decrease of demand, we have not observed this to be the case at a significant level. Therefore, the fields can be considered as a portfolio of assets with a return given by their extrapolation $p(t)$ (eq. \[eq:extrapolation\]) and a risk given by $\sigma_{field}$ (eq. \[eq:field-risk\]). This means that the average standard deviation per field at the country level from the extrapolated production can be computed as $$\sigma_{country}^{2}=\frac{1}{\#fields}\sum_{field\in fields}\sigma_{field}^{2}.$$ Intuitively, this models well the fact that the uncorrelated errors among fields will mostly cancel out.
Discovery rate of new fields\[sub:Discovery-rate-of\]
-----------------------------------------------------
Knowing the future production rate of existing fields is not enough as new fields will be discovered in the future. The model describing the discovery rate of new fields should satisfy two fundamental observations.
1. The rate of new discoveries should tend to zero as time goes to infinity. This is a consequence of the finiteness of the number of oil fields.
2. The rate of new discoveries should depend on the size of the oil fields. As of today, giant oil fields are discovered much less frequently than dwarf oil fields.
### Discoveries modeled as logistic growth\[sub:Discoveries-modeled-as\]
A natural choice for such a model is a non-homogenous Poisson process. The Poisson process is a process that generates independent events at a rate $\lambda$. It is nonhomogeneous if the rate is time-dependent, $\lambda\rightarrow\lambda(t)$. The standard way to measure $\lambda(t)$ is to find a functional form for $N(t)$, the statistical average of the cumulative number of events (discoveries) up to time $t$. Then, $\lambda(t)$ is simply a smoothed estimation of the observed rate $\frac{dN(t)}{dt}$. Figure \[rate\_of\_discovery\] shows $N(t)$ for Norwegian fields classified according to their size in two classes, dwarfs and giant fields. The logistic curve is a good fit to the data (integral form of equation \[logistic\]). This implies that after an initial increase, the rate of new discoveries reaches a peak followed by a decrease until no more oil fields are to be found, consistent with our fundamental observations. This same approach has already been successfully applied by [@Forro2012] to estimate the number of daily active users on Zynga.
As the discovery and production dynamics are not independent of the field size, the fields have been split into two groups: dwarfs and giants. Unfortunately, the two logistic curves thus obtained are highly sensitive to the splitting size. This results from the major issue, when fitting a logistic curve to data, that the carrying capacity can not be determined if the data does not already exhibit the slowdown in growth towards the carrying capacity. However, it is mentioned in the literature that often dwarf fields have already been discovered a long time ago, but their production has been postponed for economical reasons [@Lynch2002 p. 378]. Therefore, it is expected that the large oil fields have mostly been found and produced, and that future discoveries will mostly be made up of dwarf fields. Consequently, the splitting size has been picked as small as possible in order to maximize the number of giant fields but large enough to avoid recent discoveries. Our definition is thus:
- **Dwarfs**: Fields which produced less than $50\cdot10^{6}$ barrels.
- **Giants**: Fields which produced more than $50\cdot10^{6}$ barrels.
We note that this definition differs by a factor 10 from the more standard one, for which oil fields with an ultimate recoverable resource of 0.5 billion barrels (Gb) or higher are classified as giants, while oil fields with smaller URR are considered to be dwarfs [@Hook2008].
The resulting plot shown in figure \[rate\_of\_discovery\] pictures the dynamics: giant oil fields have mostly been found while the discovery process for dwarf fields is still ongoing. The logistic growth curve fit to the giant discoveries is well constrained, however the fit for the dwarfs is poorly constrained. As can be seen in figure \[rate\_of\_discovery\], the carrying capacity $K$ of the logistic growth model is not well constrained by the available data. A large spectrum of values for $K$ can lead to an equally good fit of the data. We have not taken into account the possible effect that newly discovered dwarf fields could become smaller and smaller until the new fields have a size that is too small to be economical to drill.
There are in fact two competing effects that are likely to compensate each other. On the one hand, figure \[ccdffieldsizes\] shows the complementary cumulative distribution function (CCDF) of known oil field sizes $S$ from Norway and the UK. Two salient properties can be observed. First, the tails of the distributions are well described by power laws $${\rm CCDF}(S) \sim {1 \over S^{1+\alpha}}~,
\label{ccdfsizefield}$$ with exponent $\alpha = 1.2 \pm 0.1$ for Norway and $\alpha = 1.4 \pm 0.1$ for the UK. The fact that the estimations of the exponents $\alpha$ are larger than $1$ implies that the cumulative oil reserves are asymptotically controlled by the largest fields, and not the small ones [@Sornette2004]. However, the fact that the exponents $\alpha$ are rather close to $1$ (which is called “Zipf law”) would make the many small oil field contributing significantly in total. This brings us to the second important feature exhibited by figure \[ccdffieldsizes\], namely the roll-overs of the CCDFs for small oil fields, likely due to an under-sampling of the data. Indeed, as for most data sets involving broad distributions of sizes such as oil fields, the distributions are in general incomplete for the small events due to the non exhaustive sampling. This incompleteness raises the spectre that our extrapolations might be grossly underestimating the large potential contributions to the total reserves of the many small yet undiscovered small oil fields. Assuming that the power law (\[ccdfsizefield\]) would hold for smaller fields down to size of 1 Million barrels leads to a number of such fields 10 to 100 times larger than presently known.
But there is another key factor that needs to be considered, namely the fact that small oil fields are not economically viable for exploitation, which leads to an effective truncation in the distribution (\[ccdfsizefield\]) relevant for the estimation of recoverable oil. Taking the data from the providing the yearly investments broken by fields in Norway, let us consider the illustrative case of the field GAUPE. The investment spent to develop its exploitation was \$380M, while its estimated size is $\simeq 1.2Mb$, which corresponds approximately to a total market value of $\$120M$ at \$100 per barrel. It is thus not a surprise that investment to exploit this field was interrupted in 2013. Figure \[returnNorwayfield\] shows the total estimated revenue divided by investment as a function of the estimated ultimate recovery for a number of small oil fields. It can be inferred that oil fields of sizes smaller than about 10 Mb are not economically viable as long as the market oil price does not grow much higher than \$100 per barrel. This implies that, notwithstanding the large number of unknown small oil fields, economic considerations oblige us to neglect the small oil fields, therefore providing a justification of our procedure. In fact, economic constraints may lead to cap the carrying capacity at a value smaller than the one shown in figure \[rate\_of\_discovery\].
To address these issues from a more solid angle, the method described in the next section \[sub:Likelihood-function-for\] has been used to compute the probability of different carrying capacities.
### Likelihood function for the number of discoveries\[sub:Likelihood-function-for\]
To overcome the poor constraint on the carrying capacity $K$ obtained from the fitting procedure for dwarf fields, a method already used by [@Smith1980] has been implemented. This method makes the following two postulates:
1. “The discovery of reservoirs in a petroleum play can be modeled statistically as sampling without replacement from the underlying population of reservoirs.”
2. “The discovery of a particular reservoir from among the existing population is random, with a probability of discovery being dependent on (proportional to) reservoir size.”
The fields are split into $J$ size bins denoted $S_{1},\,\ldots,\, S_{J}$ occurring with frequency $n_{1},\,\ldots,\, n_{J}$. Each discovery is considered as a step $i$ at which a field of size $I(i)\in\{S_{1},\,\ldots,\, S_{J}\}$ is found and $m_{ij}$ denotes the number of fields of size $j\in\{1,\,\ldots,\, J\}$ discovered before the $i^{th}$ step. Then, the probability that the discovery at step $i$ is of size $j$ can be expressed as $$P\left(I(i)=S_{j}\right)=\frac{\left(n_{j}-m_{ij}\right)\cdot S_{j}}{\sum_{k=1}^{J}\left(n_{k}-m_{ik}\right)\cdot S_{k}}.$$ The likelihood $L$ for a complete sequence of $N$ discoveries $\{I(1),\,\ldots,\, I(N)\}$ can then be expressed as $$L=\prod_{i=1}^{N}\frac{\left(n_{I(i)}-m_{iI(i)}\right)\cdot S_{I(i)}}{\sum_{j=1}^{J}\left(n_{j}-m_{ij}\right)\cdot S_{j}}.\label{eq:likelihood}$$ The unknown parameters are the number of fields $n_{1},\,\ldots,\, n_{J}$, whose likelihood can now be estimated based on the existing discoveries. Using a brute force approach, the entire space of plausible values for the variables $n_{1},\,\ldots,\, n_{J}$ has been sampled. The values $n_{j}$ have been sampled between the number of existing fields $m_{Nj}$ in the bin $j$ and up to a value $n_{j}^{upper}$, such that the scenario with the largest likelihood according to equation (\[eq:likelihood\]) has $n_{j}^{max}=n_{j}^{upper}/2$ fields in the bin $j$. Subsequently, the likelihood of each scenario (value of the tuple $n_{1},\,\ldots,\, n_{J}$) has been normalized such that the total likelihood of all generated scenarios equals one.
For the analysis of the discoveries of the North Sea oil fields, the number of size bins has been fixed to $J=2$ splitting between dwarfs (1) and giants (2) as described in section \[sub:Discoveries-modeled-as\].
The results shown in table \[likelihood-results\] are coherent with the intuitive expectation that discovering a new giant field is unlikely and that future discoveries will mostly be made up of dwarf fields. The likelihoods obtained for the carrying capacities of dwarfs and giants have been used to constrain the logistic curve fitted to the discoveries. Figure \[rate\_of\_discovery\] shows a sample of fitted logistic curves, each curve being weighted by the likelihood of its carrying capacity given by equation (\[eq:likelihood\]).
### Future production from discoveries\[sub:Future-production-from\]
We now compute an expected oil production coming from future discoveries, which requires to combine the steps described in sections \[sub:Discoveries-modeled-as\] and \[sub:Likelihood-function-for\].
The method described in section \[sub:Likelihood-function-for\] yields probabilities for the total number of fields (including the not yet discovered fields) in each size bins (called a scenario). However, this likelihood method does not give the time distribution of future discoveries. We propose to use the likelihood function to generate scenarios with their respective occurrence probability.
For a given scenario, the carrying capacity $K$ (= total number of fields) is given for each size class. This is useful to resolve the instability in fitting the logistic curve to the number of discovered fields. The time distribution of the discoveries is then given for each size class by the fitted logistic curve.
The actual size of a newly discovered field is generated according to the size distribution of the existing fields in its size class. The probability distribution function of field sizes in a given size bin has been fitted by a stretched exponential function.
The production curve is computed based on the average production curve of all existing fields in the same size category. The production curves of the existing fields have all been normalized to a total production of one and then have been averaged. This yields the typical production profile including a one sigma confidence interval. For a new field, this typical production curve is than multiplied by the size of the field.
Superposing the production curves results in the expected production curve from future oil fields for a given scenario.
As the total parameter space is too large to be sampled entirely, a Monte Carlo technique is applied to compute the expected production with confidence interval from future discoveries. In a nutshell, the algorithm works as follows:
1. Draw a scenario (total number of fields in each size bin) based on its probability according to the likelihood function (\[eq:likelihood\]). This is done by generating a random number $r$ between 0 and 1, and computing the scenario that is mapped to $r$ by the cumulative distribution function of all scenarios.
2. Compute the time distribution of new discoveries by fitting a logistic curve for each size class.
3. For each discovery, generate a size and the resulting production curve based on the size distribution and production curves of existing fields.
4. Superpose all the production curves.
5. Repeat and average over all drawn scenarios.
The result is the expected production curve of future oil field discoveries. The distribution of generated scenarios yields the confidence interval.
Last but not least, it has to be defined how the expected production from these future oil field discoveries is added to the extrapolated production from existing fields. The start of the simulation of new discoveries does not match up with the date of the latest production data, the reason being that the new fields (defined in section \[sub:Regular,-irregular-and\]), which are already discovered but did not yet enter the decay phase, are not taken into account in the simulation as their final size is not known. The only meaningful way of treating the new fields is to consider them as a discovery. Therefore, the starting point in time of the simulated production resulting from new discoveries has been choosen as the date in the past where it matches the current production from new fields. In order words, the extrapolated production from regular and irregular fields added to the production from simulated future discoveries (which is shifted into the past as to match the production from new fields) must be equal to the current (latest available data) total production from all fields.
Results\[sec:Results\]
======================
Based on the methodology described in section \[sub:Discovery-rate-of\], simulating future oil production was straightforward. For each country, the existing oil field productions were extrapolated and the future discoveries were simulated. Figure \[no\_uk\_2013\] shows the average of 1000 simulations. For each country, the non-symmetric shape of the production dynamics, which contradicts the prediction based on Hubbert’s standard approach, is immediately noticeable.
The results in table \[forecast-14\] show a striking difference between the extrapolation of the fit and the Monte-Carlo model forecast. According to the fit (extrapolation of aggregate production), Norway’s future oil production would decay much faster than in the Monte-Carlo case. The remaining reserves estimated with the Monte-Carlo methodology are $45\%$ larger then the estimate from the fit. This difference originates from two different effects:
- The sum of the forecast of the individual existing fields is larger than the extrapolation of the aggregate production.
- The extrapolation of the aggregate production does not capture well the discovery process of dwarf fields.
In the U.K., oil production faced a change of regime during the early nineties due to technological innovation, giving rise to the inverted w shape of the oil production profile. However, this has not been an issue for us to extrapolate the decay of production starting at the second peak. The difference between the Hubbert-based methodology and the Monte-Carlo one is very similar to the Norwegian case. The former underestimates the remaining oil reserves by about $66\%$ compared to the latter.
Which of the two models is more trustworthy? Clearly, the implications in adopting one methodology over the other are significant. The only way to answer this question is to back-test them. In other words: What would each of the models have predicted, had they been used in the past? The next section addresses this question and presents the validation step of our approach.
Validation
==========
For both countries, namely Norway and the U.K., a back-test using the data truncated in 2008 has been made. Before that date, too many of the giant fields have not entered their decay phase for a sufficiently long time to apply the extrapolation algorithm. Figure \[model-back-testing\] shows the results of these tests. Comparing the forecast of both models, with the oil production of the subsequent 6 years, shows that the predictive power of the Monte-Carlo model is significantly better than a simple extrapolation of the aggregate past production. The Monte-Carlo model is found essentially right on target, while the extrapolation of past production (“fit”) is under-estimating the realised production by 19% and 16% respectively for Norway and the U.K. The following table \[backtest-08\] summarizes the difference between the two approaches for the back-testing period.
While the error of the extrapolation of the past oil production method (“fit”) is not dramatic over these six years from 2008 to 2014, the difference between it and the Monte-Carlo approach becomes huge from 2014 to 2030, as shown in table \[forecast-08\].
The simple extrapolation decays too fast and entirely misses the fat tails in the decay process of individual fields and the new discoveries. Moreover, it must be noted that the simple extrapolation changed massively between the back-test in table \[forecast-08\] and the current fit in table \[forecast-14\] (520% for Norway and 236% for the U.K.). In other words, the simple extrapolation is very unstable in its forecast, while the Monte-Carlo forecasts remains very consistent (less than 10% change).
As can be seen in figure \[model-back-testing\], the actual production of Norway during the back-testing period remained entirely within the quite narrow $1\sigma$ interval of the Monte-Carlo methodology, while totally breaking out of the $1\sigma$ interval of the simple extrapolation. For the U.K. the Monte-Carlo methodology only performs slightly better when considering the confidence interval, and the confidence interval is much larger due to the uncertainty on future discoveries and their production profile.
Conclusion
==========
We have presented a Monte-Carlo based methodology to forecast future oil production. By extending the oil production of current fields into the future and modeling the discovery rate of new fields, the future oil production of Norway and the U.K. could be forecasted. These forecasts are significantly different from the ones obtained with a standard extrapolation. Indeed, our model forecasts 45% to 66% more remaining oil reserves than the standard extrapolation. The back-test performed on the time period between 2008 and 2014 confirmed that the Monte-Carlo based model better captured the production dynamics.
The results suggest that it is highly likely that the decay of Norwegian and U.K. oil production will be much slower then one would expect from a standard extrapolation. Nonetheless, to maintain current levels of oil consumption in the European Union, more of it will have to be imported from outside Europe, as the imports from Norway will vanish (currently accounting for 11% of E.U. oil imports ) and the U.K. will need to import more oil.
As shown in table \[needs\], at constant consumption, the Monte-Carlo model predicts that in 2030 the E.U. with Norway will need to increase its oil imports by $1.3$ million barrels. These imports will most likely have to come from outside Europe, except for non-standard oil sources yet to be developed.
The present methodology can be applied to many other countries and geological areas, as well as updated at the level of the global oil production. Extensions to include the new wave of shale oil and non-standard oil can in principle be considered and constitute an interesting domain of application of our methodology for the future.
0.5cm [**Acknowledgements**]{}: This work derives from a Master thesis at ETH Zurich authored by Marc A. Del Degan (July 2012) under the supervision of the authors, entitled “Analysis of peak oil with focus on Norwegian oil production” (<http://www.er.ethz.ch/publications/MAS_Thesis_DelDegan_final_July12.pdf>), which itself derived from the methodology of pricing social network companies such as Zynga developed by three of the authors [@Forro2012].
![Example of the time evolution of the oil production per day of a regular field, parameterised beyond the peak in the decay regime by the stretched exponential function (\[stretched\_exponential\]) with $\beta=0.66\pm0.01$ and $\tau=-55\pm1$ months, shown with the black line. The one standard deviation given by expression (\[eq:extrapolation\]) is represented by the grey band.[]{data-label="regular"}](oseberg){width="1\columnwidth"}
![Example of a irregular field.[]{data-label="irregular"}](vale){width="1\columnwidth"}
![Example of a new field.[]{data-label="new"}](skarv){width="1\columnwidth"}
![Oesberg field - Relative error defined by expression (\[ryjryukoik\]) of the predicted total production from time $t$ indicated in the abscissa until 2014. One can observe that the predicted future total production is over-estimated by as much as 70% in 1999, then under-estimated by the same amount in 2002, while the forecast errors remain smaller than 20% since 2004.[]{data-label="oesberg-error"}](oesberg-error){width="1\columnwidth"}
![Logistic fit of the function solution of expression (\[logistic\]) to the number of discoveries for Norway. The discovery rate of new oil fields is dependent on their size as explained in the main text.[]{data-label="rate_of_discovery"}](extrapolating-no-discoveries){width="1\columnwidth"}
![Complementary cumulative distribution function (CCDF) of known oil field sizes $S$ from Norway and the UK. The two dashes lines visualise the power law behaviour (\[ccdfsizefield\]) of the tail of the distributions. []{data-label="ccdffieldsizes"}](ccdf-field-size){width="1\columnwidth"}
![Total estimated revenue divided by investment as a function of the estimated ultimate recovery for a number of small oil fields. []{data-label="returnNorwayfield"}](no-returns-size){width="1\columnwidth"}
![Monte-Carlo (green upper continuous line with standard deviation bank starting in 2014 onward) and fit forecast based on past production data (lower line and grey one standard deviation band) for Norway (top) and the U.K. (bottom). In both cases, the Monte-Carlo model forecasts a significantly slower decay than the fit by taking into account that new fields will come in production.[]{data-label="no_uk_2013"}](NO-future "fig:"){width="1\columnwidth"} ![Monte-Carlo (green upper continuous line with standard deviation bank starting in 2014 onward) and fit forecast based on past production data (lower line and grey one standard deviation band) for Norway (top) and the U.K. (bottom). In both cases, the Monte-Carlo model forecasts a significantly slower decay than the fit by taking into account that new fields will come in production.[]{data-label="no_uk_2013"}](UK-future "fig:"){width="1\columnwidth"}
![Monte-Carlo (green upper continuous line with standard deviation bank starting in 2014 onward) and Hubbert forecast based on past production data up to 2008 (lower line and grey one standard deviation band) for Norway (top) and the U.K. (bottom). The results can be compared with the subsequent oil production (blue area). In both cases, the Monte-Carlo methodology is more precise.[]{data-label="model-back-testing"}](NO-back-test "fig:"){width="1\columnwidth"} ![Monte-Carlo (green upper continuous line with standard deviation bank starting in 2014 onward) and Hubbert forecast based on past production data up to 2008 (lower line and grey one standard deviation band) for Norway (top) and the U.K. (bottom). The results can be compared with the subsequent oil production (blue area). In both cases, the Monte-Carlo methodology is more precise.[]{data-label="model-back-testing"}](UK-back-test "fig:"){width="1\columnwidth"}
$m_{N1}$ $n_{1}\pm\sigma_{1}$ $m_{N2}$ $n_{2}\pm\sigma_{2}$
-------- ---------- ---------------------- ---------- ----------------------
Norway $24$ $88.4\pm10.0$ $52$ $56.4\pm1.6$
U.K. $162$ $208\pm11$ $99$ $100\pm0.4$
: Likelihood estimation for the number of dwarf (1) and giant (2) fields. For Norway, our logistic fit suggests that up to two new giant fields could be ultimately discovered. For the U.K., the prediction is more bleak, suggesting that the most probable scenario is that one giant field will be found (which is most likely already discovered but classified as a new field).
\[likelihood-results\]
Fit (barrels) Model (barrels) $\Delta$
-------- ------------------ ------------------ ----------
Norway $130\cdot10^{6}$ $188\cdot10^{6}$ $45\%$
U.K. $59\cdot10^{6}$ $98\cdot10^{6}$ $66\%$
: Remaining oil reserves until 2030 in barrels predicted by the extrapolation of the **Fit** of the past country production, and predicted by the Monte-Carlo **Model**. The relative difference between these two predictions is defined by $\Delta=\frac{\textrm{Model}-\textrm{Fit}}{\textrm{Fit}}$.
\[forecast-14\]
Actual (barrels) Fit (barrels) Model (barrels)
-------- ------------------ ------------------ ------------------
Norway $133\cdot10^{6}$ $108\cdot10^{6}$ $130\cdot10^{6}$
U.K. $79\cdot10^{6}$ $66\cdot10^{6}$ $75\cdot10^{9}$
: Extrapolation of past oil production (“fit”) and prediction using the Monte-Carlo model are used on the data set truncated in 2008. Their forecast for the period 2008-2014 is compared to the actual realised production.
\[backtest-08\]
Fit (barrels) Model (barrels) $\Delta$
----------- ----------------- ------------------ ----------
Norway 08 $25\cdot10^{6}$ $171\cdot10^{6}$ $-584\%$
U.K. 08 $25\cdot10^{6}$ $91\cdot10^{9}$ $-264\%$
: Remaining oil reserves forecasted for the period 2014-2030 when using the data truncated in 2008, according to the extrapolation of past oil production (“fit”) and the Monte-Carlo model. The relative difference between these two predictions is defined by $\Delta=\frac{\textrm{Model}-\textrm{Fit}}{\textrm{Fit}}$.
\[forecast-08\]
2014 2020 2025 2030
------------- -------------------- -------------------- -------------------- --------------------
Norway $-1.23\cdot10^{6}$ $-0.88\cdot10^{6}$ $-0.58\cdot10^{6}$ $-0.43\cdot10^{6}$
U.K. $0.7\cdot10^{6}$ $0.9\cdot10^{6}$ $1.0\cdot10^{6}$ $1.1\cdot10^{6}$
E.U.+Norway $9.8\cdot10^{6}$ $10.45\cdot10^{6}$ $10.85\cdot10^{6}$ $11.1\cdot10^{6}$
: Oil import (bbl/day) at a constant consumption of 1.5M bbl/day for the U.K. and 0.22M bbl/day for Norway. The import for the E.U. and Norway is a lower bound based on the changes in the U.K and Norway. Negative numbers for Norway represent exports.
\[needs\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It has recently been shown that the problem of testing global convexity of polynomials of degree four is [strongly]{} NP-hard, answering an open question of N.Z. Shor. This result is minimal in the degree of the polynomial when global convexity is of concern. In a number of applications however, one is interested in testing convexity only over a compact region, most commonly a box (i.e., hyper-rectangle). In this paper, we show that this problem is also strongly NP-hard, in fact for polynomials of degree as low as three. This result is minimal in the degree of the polynomial and in some sense justifies why convexity detection in nonlinear optimization solvers is limited to quadratic functions or functions with special structure. As a byproduct, our proof shows that the problem of testing whether all matrices in an interval family are positive semidefinite is strongly NP-hard. This problem, which was previously shown to be (weakly) NP-hard by Nemirovski, is of independent interest in the theory of robust control.'
author:
- 'Amir Ali Ahmadi and Georgina Hall [^1]'
title: '**On the Complexity of Detecting Convexity over a Box**'
---
#### Keywords:
[Convexity detection, convex optimization, computational complexity, interval positive semidefiniteness.]{}
Introduction
============
In a relatively recent paper [@NPhard_Convexity_MathProg], Ahmadi, Olshevsky, Parrilo, and Tsitsiklis have shown that the problem of testing whether a quartic (multivariate) polynomial is globally convex is NP-hard. This result answers a question of N.Z. Shor that had appeared in 1996 on a list of seven open problems in computational complexity for numerical optimization [@pardalos1992open]. The significance of the question stems from the fact that in the theory of optimization, the divide between convex and nonconvex optimization has proved to be a fundamental one. Indeed, for many classes of optimization problems, stronger algorithmic, analytic, and geometric statements can be made when the functions involved in the description of the problem are convex compared to when they are not. Hence, it is natural to ask (as Shor did) whether one can test for convexity of functions in an efficient manner. Polynomial functions provide a convenient setting for a rigorous study of this question from a computational complexity viewpoint.
For a number of problems of applied nature, one is interested not in checking for convexity of a function globally, but only over a compact region. The most common region that arises in practical applications is a *box* (i.e., a hyper-rectangle). For example, the variables in many real-world optimization problems are lower and upper bounded a priori, allowing the algorithm designer to restrict attention to a box. Similarly, in branch-and-bound approaches to nonlinear and nonconvex optimization, one often recursively breaks the original problem into a number of subproblems over smaller boxes and aims to provide upper and lower bounds on the minimum of the objective function over these smaller boxes. In such a scenario, the subproblems for which the objective function is convex over the associated box can often be solved to global optimality.
The problem of detecting convexity over a box, or the related problem of *imposing* convexity over a box, appears in areas outside of optimization as well. In control theory for example, Lyapunov functions that certify properties of dynamical systems are sometimes required to be convex over the region of the space where the state variables lie [@chesi2008establishing; @ahmadi2018sos]. Hence, the need for algorithms that check convexity of a candidate Lyapunov function over a compact region—or those that find a Lyapunov function that meets the convexity requirement by construction—naturally arises. Similarly, in statistics, the well-known problem of *convex regression* [@xu2016faithful; @hannah2013multivariate; @lim2012consistency] is about finding a convex function (within a certain function class) that best agrees with observed data. In this situation again, imposing convexity of the function globally is often too restrictive. Indeed, one generally has access to a feature domain where the explanatory variables of the problem are known to lie. This domain, more often that not, is a box.
Motivated by these considerations, our goal in this paper is to study the complexity of testing convexity of a given function over a box (which can only be simpler than the problem of imposing convexity over a box). As is done in [@NPhard_Convexity_MathProg], we restrict our attention to polynomial functions since they appear routinely in applications and are amenable to complexity theoretic investigations due to their finite parametrization. Our main result (Theorem \[th:np.hardness.convex\]) shows that the problem of testing convexity of a polynomial over a box is strongly NP-hard already for cubic polynomials. This result completely classifies the complexity of testing convexity of polynomials of any degree over a box (cf. Proposition \[prop:degree\]). It also justifies, at least from a complexity viewpoint, why convexity detection in some of the most widely-used optimization packages such as BARON [@baron; @khajavirad2017hybrid], CVX [@cvx], Gurobi [@gurobi], is restricted to quadratic functions, or is replaced with the task of checking that the function in question be the output of certain convexity-preserving operations applied to an initial set of convex functions; see, e.g., [@BoydBook Section 3.2] .
The proof of has two parts, the first of which relies heavily on a clever gadget by Nemirovski [@nemirovskii1993several]. Some of our modifications to his arguments are of potential interest to the field of robust control (cf. Corollary \[cor:psd.interval\]).
The Main Result
===============
We begin with some basic definitions.
Given a set of scalars $l_1,\ldots,l_n$,$u_1,\ldots,u_n,$ with $l_i\leq u_i$ for $i=1,\ldots,n$, a *box* $B\subseteq \mathbb{R}^n$ is a set of the form $$B=\{x \in \mathbb{R}^n|~ l_i\leq x_i \leq u_i, i=1,\ldots,n\}.$$
A function $f:\mathbb{R}^n \rightarrow \mathbb{R}$ is said to be *convex over a set* $C\subseteq \mathbb{R}^n$ if for every $x, y \in C$ and for any $\lambda \in [0,1]$, we have $f(\lambda x+(1-\lambda)y)\leq \lambda f(x)+(1-\lambda)f(y)$.
We study the complexity of detecting convexity in the standard Turing model of computation (see, e.g., [@sipser2006introduction]) where the input to every problem instance must be defined by a finite number of bits. As a consequence, in the statement of Theorem \[th:np.hardness.convex\] below, the input to the problem—which consists of the coefficients of the polynomial $f$ and the scalars $l_1,\ldots,l_n,u_1,\ldots,u_n$ that define the box $B$—is taken to be rational.
We further remark that the problem in the statement of Theorem \[th:np.hardness.convex\] is *strongly* NP-hard. This is as opposed to problems that are *weakly* NP-hard such as the classical problems of KNAPSACK or PARTITION [@GareyJohnson_Book]. This distinction only occurs in problems whose instances involve numerical data[^2] and relates to how the numerical values that appear in an instance compare against the length of the instance. Roughly speaking, being NP-hard in the strong sense means that the problem remains NP-hard even when the numerical values in all problem instances are “small”. More rigorously, , define a Max (resp. Length) function which maps any instance $I$ of a decision problem to a nonnegative integer Max($I$) (resp. Length($I$)) that represents the largest numerator or denominator in magnitude of all rational numbers appearing in $I$ (resp. the number of bits required to write down $I$). A decision problem $\Pi$ is said to be *strongly NP-hard* if there exists a subproblem $\hat{\Pi}$ which is NP-hard and for which one can find a polynomial $p$ such that Max($I$) is upperbounded by $p$(Length($I$)) for all instances $I \in \hat{\Pi}$. The advantage of showing that a problem is strongly NP-hard (as opposed to weakly NP-hard) is that such a statement rules out, unless P=NP, the possibility of a *pseudo-polynomial time* algorithm. This is an algorithm whose running time is polynomial in the numerical value of the input but not necessarily in the bit length of the input; see [@GareyJohnson_Book Section 4.2.2] for more details. More concretely, if we had only shown weak NP-hardness of testing convexity of an $n$-variate cubic polynomial over a box, it could be that the problem is polynomial-time solvable (even assuming P$\neq$ NP) except for instances where some of the coefficients of the polynomial or the bounds of the box are of order $2^n$. Indeed, for the KNAPSACK and PARTITION problems for example, it is well known that one can solve, in polynomial time, all instances which do not contain such large numbers using dynamic programming [@GareyJohnson_Book p. 96].
\[th:np.hardness.convex\] Given a cubic polynomial $f$ and a box $B$, it is strongly NP-hard to test whether $f$ is convex over $B$.
The proof of is based on a reduction from the problem of testing whether a matrix $L(x)$—whose entries are affine polynomials in $x$—is positive semidefinite for all $x$ in a full-dimensional box $B$. This problem has already been shown to be NP-hard by Nemirovski [@nemirovskii1993several]. The connection between this problem and the problem of testing convexity of a cubic polynomial $f\mathrel{\mathop{:}}=f(x)$ over a box $B$ lies in the facts that (i) the Hessian $\nabla^2 f(x)$ of $f(x)$ is an affine polynomial matrix, and (ii) a twice continuously differentiable function $f$ is convex over a convex set $S$ with nonempty interior if and only if The proof of Theorem \[th:np.hardness.convex\] is split into two parts which we outline below.
In the first part of the proof, we concern ourselves with strengthening the NP-hardness result from [@nemirovskii1993several], which we use here, from weak to strong (Theorem \[th:strong.np.hard\]). Indeed, the reduction provided in [@nemirovskii1993several] is from the PARTITION problem which is only weakly NP-hard [@GareyJohnson_Book], and hence only proves weak NP-hardness of testing positive semidefiniteness of an affine polynomial matrix over a box. As a consequence, we give a new reduction, which still uses the insights that we gleaned from Nemirovski’s proof, but with two modifications. First, we give a reduction from a strongly NP-hard problem. Second, we bypass a step in the Nemirovski’s construction that involves matrix inversion. This is because it is possible for the inverse of a matrix to have entries that are exponential in the entries of the original matrix, thus causing the reduction to lose its strong NP-hardness implication. To bypass matrix inversion, we instead approximate the inverse by its first-order Taylor expansion and control the spectrum of the matrix in such a way that the higher-order terms in the Taylor expansion can be appropriately bounded (cf. Lemma \[lem:bounds\]).
In the second and core part of the proof, we give the reduction from the problem of testing positive semidefiniteness of an affine polynomial matrix over a box to the problem of testing convexity of a cubic polynomial over a box. The main obstacle that we need to overcome here is that not every affine (symmetric) polynomial matrix $L(x)$ is the Hessian of some cubic polynomial. For instance, if, for some integers $i,j,k \in \{1,\ldots,n\}$, the equality $$\frac{\partial L_{ij}(x)}{\partial x_k} =\frac{\partial L_{ik}(x)}{\partial x_j}$$ is violated, then $L(x)$ cannot be a valid Hessian matrix. This is because partial derivatives of polynomials must commute. Our goal is then to show that even with these additional constraints, the problem of checking positive semidefiniteness remains hard. To do this, we will introduce new variables $y$ and carefully construct a polynomial $f(x,y)$ whose Hessian can be related back to $L(x)$.
\[th:strong.np.hard\] Given a symmetric $n \times n$ matrix $A$ with entries in $\{0,1\}$ and a positive integer $k \leq n^2$, let $$\begin{aligned}
\label{def:gam.C.mu}
C\mathrel{\mathop{:}}=\frac{4}{(n+1)^3}\left(I_n+\frac{1}{(n+1)^3}A\right), \text{ and } \mu\mathrel{\mathop{:}}=\frac{n(n+1)^3}{4}+k-1-\frac{1}{4}e^TAe,
\end{aligned}$$ where $e$ is the $n \times 1$ vector of all ones and $I_n$ is the $n \times n$ identity matrix. It is strongly NP-hard to test whether $$\begin{aligned}
\label{def:Lx}
L(x)\mathrel{\mathop{:}}=\begin{bmatrix}
C & x \\ x^T & \mu +\frac14
\end{bmatrix} \succeq 0, \forall x \in \mathbb{R}^n \text{ with } ||x||_{\infty}\leq 1.
\end{aligned}$$ Furthermore, for any matrix $C$ and scalar $\mu$ thus defined, either (\[def:Lx\]) holds or there exists $x \in \mathbb{R}^n$ with $||x||_{\infty}\leq 1$ such that $x^TC^{-1}x \geq \mu +\frac34$.
To prove , we will make use of the following lemma.
\[lem:bounds\]
Let $A$ be a symmetric $n \times n$ matrix with entries in $\{0,1\}$ and let $C$ be defined as in (\[def:gam.C.mu\]). Then, for any $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1$, we have $$\begin{aligned}
\frac{1}{4}x^T\left((n+1)^3I_n-A\right)x -\frac{1}{4}\leq x^TC^{-1}x \leq \frac{1}{4}x^T\left((n+1)^3I_n-A\right)x +\frac{1}{4}.\end{aligned}$$
First, note that $C$ is by construction strictly diagonally dominant, and hence, by Gershgorin’s circle theorem [@gersh], positive definite. This implies that $C^{-1}$ is well defined. Furthermore, we have $$\begin{aligned}
\label{eq:C.identity}
C^{-1}=\frac{(n+1)^3}{4} \left(I_n- \frac{-A}{(n+1)^3}\right)^{-1}=\frac{(n+1)^3}{4} \sum_{k=0}^{\infty} \frac{(-A)^k}{(n+1)^{3k}},\end{aligned}$$ where the second equality follows from the identity $(I_n-M)^{-1}=\sum_{k=0}^\infty M^k$, Using (\[eq:C.identity\]), we have $$x^TC^{-1}x=\frac{1}{4} x^T((n+1)^3I_n -A)x +\frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{x^T(-A)^kx}{(n+1)^{3k}}.$$ It remains to show that $$\left| \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{x^T(-A)^kx}{(n+1)^{3k}} \right| \leq \frac14$$ for any $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1$ to conclude the proof. Letting $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1$, we have $$\begin{aligned}
\left| \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{x^T(-A)^kx}{(n+1)^{3k}} \right| &\leq \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{|x^T(-A)^kx|}{(n+1)^{3k}}\\
&\leq \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{||x||_2\cdot ||(-A)^kx||_2}{(n+1)^{3k}}\\
&\leq \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{||x||_2 \cdot ||A||^k_2 \cdot ||x||_2}{(n+1)^{3k}}\\
&\leq \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{n||A||^k_2}{(n+1)^{3k}}\\
&\leq \frac{n(n+1)^3}{4} \sum_{k=2}^\infty \frac{1}{(n+1)^{2k}},\end{aligned}$$ where the first inequality uses the triangle inequality, the second the Cauchy-Schwarz inequality, the third the definition and the submultiplicativity property of the matrix 2-norm, the fourth the facts that $||x||_2 \leq \sqrt{n}||x||_{\infty}$ and $||x||_{\infty} \leq 1$, and the fifth the fact that $||A||_2\leq n \leq n+1.$ Now, using the geometric series identity $$\sum_{k=2}^{\infty} \frac{1}{(n+1)^{2k}}=\frac{1}{(n+1)^4}\cdot \frac{1}{1-\frac{1}{(n+1)^2}},$$ we conclude that $$\begin{aligned}
\left| \frac{(n+1)^3}{4} \sum_{k=2}^\infty \frac{x^T(-A)^kx}{(n+1)^{3k}} \right| &\leq \frac{1}{4}\cdot \frac{n}{n+1} \cdot\frac{1}{1-\frac{1}{(n+1)^2}}\\
&=\frac{1}{4}\cdot \frac{n+1}{n+2} \\
&\leq \frac{1}{4}.\end{aligned}$$
We give a reduction from the SIMPLE MAX-CUT problem, which is the following decision problem: Given a simple[^3] graph $G=(V,E)$ and a positive integer $s \leq |V|^2$, decide whether there is a cut of size greater or equal to $s$ in $G$, i.e., a partition of $V$ into two disjoint sets $V_1$ and $V_2$ such that the number of edges from $E$ that have one endpoint in $V_1$ and one endpoint in $V_2$ is greater or equal to $s$. This problem is known to NP-hard [@GareyJohnson_Book p. 210]. Consider an instance of SIMPLE MAX-CUT given by a graph $G=(V,E)$ and a positive integer $s \leq |V|^2$. We construct an instance of the problem given in the statement of the theorem by taking $A$ to be the adjacency matrix of $G$ (i.e., a symmetric $|V| \times |V|$ matrix whose $(i,j)$-th entry is equal to one if $\{i,j\} \in E$ and equal to zero otherwise) and $k=s$. Clearly this reduction is polynomial in length. Now, let $C$ and $\mu$ be as defined in (\[def:gam.C.mu\]) with $n=|V|$.
We show that $L(x)$ as defined in (\[def:Lx\]) is positive semidefinite for all $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1$ if and only if there is no cut in $G$ of size greater or equal to $s$. This would show that the problem in the statement of the theorem is NP-hard. In fact, as explained previously, since $$\mbox{Max}(I)=k \leq n^2+\lceil \log_2(k) \rceil=\mbox{Length}(I)$$ for any instance $I$ of the problem, the problem is automatically strongly NP-hard; see e.g. [@GareyJohnson_Book p. 95]. Suppose first that there is no cut in $G$ of size greater or equal to $s$. As $k=s$, this implies that the largest cut in $G$ is of size less than or equal to $k-1$. It is straightforward to verify that the size of the largest cut in $G$ is given by $\max_{x \in \{-1,1\}^n} \frac{1}{4}\sum_{i,j}A_{ij} (1-x_ix_j)$. Hence, if $$\begin{aligned}
p^*\mathrel{\mathop{:}}=&\min_x &&\frac{1}{4}x^TAx\\
&\text{s.t. } &&x \in \{-1,1\}^n,
\end{aligned}$$ the size of the largest cut in $G$ is less than or equal to $k-1$ if and only if $\frac{1}{4}e^TAe -p^* \leq k-1 \Leftrightarrow p^{*} \geq \frac{1}{4} e^TAe -k+1.$ As $x^Tx=n$ when $x \in \{-1,1\}^n$, this is equivalent (after basic algebra) to the optimal value of $$\begin{aligned}
&\max_x &&\frac{1}{4} x^T((n+1)^3 I_n-A)x\\
&\text{s.t. } &&x \in \{-1,1\}^n
\end{aligned}$$ being less than or equal to $\frac{n(n+1)^3}{4} +k-1-\frac{1}{4}e^TAe=\mu.$ As the quadratic function $$x \mapsto \frac14 x^T((n+1)^3I_n-A)x$$ is convex, and as the maximum of a convex function over a box is attained at an extreme point of the box (see, e.g., [@bensonconcave Property 12]), the previous statement is equivalent to the optimal value of $$\begin{aligned}
&\max_x &&\frac{1}{4} x^T((n+1)^3I_n-A)x\\
&\text{s.t. } &&x \in [-1,1]^n
\end{aligned}$$ being less than or equal to $\mu.$ Hence we have shown that if the size of the largest cut in $G$ is less than or equal to $k-1$, then $\frac{1}{4} x^T ((n+1)^3 I_n-Ax)x \leq \mu$ for all $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1.$ From Lemma \[lem:bounds\], it follows that $$x^TC^{-1}x \leq \mu +\frac{1}{4}$$ for any $x \in \mathbb{R}^n$ with $||x||_{\infty}\leq 1.$ Using the Schur complement , this is equivalent to $L(x)$ being positive semidefinite for all $x \in \mathbb{R}^n$ with $||x||_{\infty}\leq 1.$
Suppose now that there is a cut in $G$ of size greater or equal to $s$. Let $\hat{x} \in \{-1,1\}^n$ be the indicator vector of this cut; i.e., $\hat{x}_i=1$ if node $i$ belongs to $V_1$ and $\hat{x}_i=-1$ if node $i$ belongs to $V_2$. It is easy to check that the number of edges with one endpoint in $V_1$ and one endpoint in $V_2$ is given by $\frac{1}{4} (e^TAe- \hat{x}^TA\hat{x}).$ Hence, as $k=s$, we have $$\frac{1}{4} (e^TAe- \hat{x}^TA\hat{x}) \geq k,$$ which is equivalent to $\frac{1}{4} n(n+1)^3-\frac{1}{4}\hat{x}^TA\hat{x} \geq \frac{1}{4} n(n+1)^3-\frac{1}{4}e^TAe+k=\mu+1$. As $\hat{x}^T\hat{x}=n$, it follows that $$\frac{1}{4}\hat{x}^T((n+1)^3I_n-A)\hat{x} \geq \mu +1.$$ Using Lemma \[lem:bounds\], we get $$\hat{x}^TC^{-1}\hat{x} \geq \mu +\frac34 >\mu +\frac14.$$ By the Schur complement, we conclude that $L(\hat{x}) \not\succeq 0.$
Note that the final claim of the theorem is already implied by the arguments given thus far. Indeed, given $C$ and $\mu$ as defined in (\[def:gam.C.mu\]), one can construct a graph whose adjacency matrix is $A$ and take $s=k$. Either this graph has no cut of size greater or equal to $s$, in which case we have shown that $x^TC^{-1}x \leq \mu+\frac14$ for all $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1$, or it has a cut of size greater or equal to $s$, in which case there exists $\hat{x}$ such that $||\hat{x}||_{\infty} \leq 1$ and $\hat{x}^T C^{-1} \hat{x} \geq \mu +\frac{3}{4}.$
We briefly provide an immediate corollary of this theorem, which we believe can be of independent interest as it relates to problems in robust control [@ben2003extended; @blondel1997np; @ben2002tractable]. This statement was first proven in [@nemirovskii1993several], except that the result there shows NP-hardness in the weak sense.
\[cor:psd.interval\] Given rational numbers $\hat{m}_{ij}, \bar{m}_{ij}, i=1,\ldots,N, j=1,\ldots,N,$ with $\hat{m}_{ij} \leq \bar{m}_{ij}$, $\hat{m}_{ij}=\hat{m}_{ji}$, and $\bar{m}_{ij}=\bar{m}_{ji}$ for $i=1,\ldots,N,$ $j=1,\ldots,N$, it is strongly NP-hard to test whether all $N \times N$ symmetric matrices $M$ with entries $m_{ij} \in [\hat{m}_{ij}, \bar{m}_{ij}]$ for $i=1,\ldots,N, j=1,\ldots,N$, are positive semidefinite.
The claim follows straightforwardly by a reduction from the problem in the statement of Theorem \[th:strong.np.hard\]. Let $A$ be a symmetric $n \times n$ matrix with entries in $\{0,1\}$ and $k \leq n^2$ be a positive integer. Construct the symmetric matrix $C$ and the scalar $\mu$ as in (\[def:gam.C.mu\]). Let $\hat{m}_{ij}=\bar{m}_{ij}=C_{ij}$ for $i=1,\ldots,n, j=1,\ldots,n,$ $\hat{m}_{i(n+1)}=\hat{m}_{(n+1)i}=-1$ and $\bar{m}_{i(n+1)}=\bar{m}_{(n+1)i}=1$ for $i=1,\ldots,n$, and $\hat{m}_{(n+1)(n+1)}=\bar{m}_{(n+1)(n+1)}=\mu+\frac{1}{4}$. It follows that all $(n+1) \times (n+1)$ symmetric matrices $M$ with entries $m_{ij} \in [\hat{m}_{ij}, \bar{m}_{ij}]$ for $i=1,\ldots,n+1, j=1,\ldots,n+1$, are positive semidefinite if and only if the matrix $L(x)$ defined in (\[def:Lx\]) is positive semidefinite for all $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1.$
Note that the reduction we have given leads to a strong NP-hardness result. Indeed, all instances $I$ of the interval positive semidefiniteness problem that arise from this construction have $\mbox{Length}(I)$ of order $n^2$ and $\mbox{Max}(I)$ of order $n^6$, and hence $\mbox{Max}(I)$ is upper bounded by a polynomial in $\mbox{Length}(I)$.
We now move on to the proof of the main theorem of this paper.
We prove the claim via a reduction from the NP-hard problem stated in Theorem \[th:strong.np.hard\]. Let $A$ be an $n \times n$ symmetric matrix with entries in $\{0,1\}$ and $k \leq n^2$ be a positive integer. Let $C$ and $\mu$ be as in (\[def:gam.C.mu\]) and, define $L(x)$ to be the following $(n+1) \times (n+1)$ symmetric matrix $$L(x) \mathrel{\mathop{:}}=\begin{bmatrix} C & x\\ x^T & \mu +\frac14 \end{bmatrix}.$$ Let $y\mathrel{\mathop{:}}=(y_1,\ldots,y_{n+1})^T$ be a new vector of variables and $H(y)$ be the following $n \times (n+1)$ matrix of mixed partial derivatives of the cubic polynomial $y^TL(x)y$: $$H_{ij}(y)=\frac{\partial^2 y^TL(x)y}{\partial x_i \partial y_j},~\text{for } i=1,\ldots,n, \text{ and }j=1,\ldots,n+1.$$ Using the structure of $L(x)$, after some basic algebra, we obtain: $$\begin{aligned}
\label{eq:Hy}
H(y)=\begin{bmatrix} 2y_{n+1} & 0 & \ldots & 0 & 2y_1 \\ 0 & \ddots & 0 & 0 & 2y_2\\ 0 & 0 & \ddots & 0 & \vdots \\ 0 & \ldots & 0 & 2y_{n+1} & 2y_n \end{bmatrix}.
\end{aligned}$$ Let $$\label{def:n.alpha.gamma}
\begin{aligned}
&\alpha\mathrel{\mathop{:}}=16\cdot n(1+ 16n^7),\\
\gh{\text{and }}&\eta\mathrel{\mathop{:}}=\frac{1}{4}\cdot \frac{1}{1+16n^7}.
\end{aligned}$$ Consider the following cubic polynomial in $2n+1$ variables $x_1,\ldots,x_n, y_1,\ldots,y_{n+1}$ $$f(x,y)=\frac{1}{2}y^TL(x)y+\frac{\alpha}{2} x^Tx+\frac{\eta}{2} y^Ty,$$ and the box $B=[-1,1]^{2n+1}.$ We claim that $f(x,y)$ is convex over $B$ if and only if $L(x)\succeq 0$ for all $x \in \mathbb{R}^n$ with $||x||_{\infty} \leq 1.$ This would imply the desired NP-hardness result as the construction of $f$ and $B$ from $A$ and $k$ can be carried out in polynomial time. Moreover, NP-hardness here is in the strong sense as all instances $I$ of the problem in the statement of Theorem \[th:np.hardness.convex\] that arise from this construction have $\mbox{Length}(I)$ of order $n^3$ (the number of coefficients of $f$) and $\mbox{Max}(I)$ of order $n^8$ (which is the magnitude of $\alpha$, which dominates all other coefficients). Hence, $\mbox{Max}(I)$ is upper bounded by a polynomial in To prove the claim, we begin by observing that the Hessian of $f$ has the following structure: $$\nabla^2 f(x,y)=\begin{bmatrix}\alpha I_n & \frac12 H(y) \\ \frac12 H(y)^T & L(x)+\eta I_{n+1}\end{bmatrix}.$$ Suppose first that $L(x)$ is not positive semidefinite for all $x\in [-1,1]^n$. Then, from the second claim in Theorem \[th:strong.np.hard\], there exists $\bar{x} \in [-1,1]^n$ such that $\bar{x}^TC^{-1}\bar{x} \geq \mu+\frac34.$ Let $\bar{y}=0_{(n+1)\times 1}$, i.e., the vector of zeros of length $n+1$, and observe that $(\bar{x},\bar{y}) \in [-1,1]^{2n+1}=B.$ Let $$z=\frac{1}{||(-C^{-1}\bar{x},1)^T||_2}\begin{pmatrix} 0_{n \times 1}\\ -C^{-1}\bar{x}\\ 1
\end{pmatrix}.$$ We will show that $z^T\nabla^2 f(\bar{x},\bar{y})z < 0.$ Indeed, we have $$\label{eq:expand.hessian}
\begin{aligned}
z^T\nabla^2f(\bar{x},\bar{y})z&= \frac{1}{1+||C^{-1}\bar{x}||_2^2} \cdot\begin{bmatrix} -C^{-1}\bar{x}\\ 1 \end{bmatrix}^T (L(\bar{x})+\eta I_{n+1})\begin{bmatrix} -C^{-1}\bar{x}\\ 1 \end{bmatrix} \\
&= \frac{1}{1+||C^{-1}\bar{x}||_2^2} \cdot\begin{bmatrix} -C^{-1}\bar{x}\\ 1 \end{bmatrix}^T \begin{bmatrix}
C+\eta I_n & \bar{x} \\ \bar{x}^T & \mu+\frac14 +\eta
\end{bmatrix}\begin{bmatrix} -C^{-1}\bar{x}\\ 1 \end{bmatrix}\\
&= \frac{1}{1+||C^{-1}\bar{x}||_2^2} \left( \mu +\frac14 -\bar{x}^TC^{-1}\bar{x} +\eta (1+||C^{-1}\bar{x}||_2^2) \right)\\
&=\frac{\mu+\frac14-\bar{x}^TC^{-1}\bar{x}}{1+||C^{-1}\bar{x}||_2^2}+\eta.
\end{aligned}$$ To show that this expression is negative, we upper bound the numerator and the denominator of its first summand. Using (\[eq:C.identity\]), we can write $$\begin{aligned}
||C^{-1}\bar{x}||_2 &\leq \frac{(n+1)^3}{4} \sum_{k=0}^{\infty} \frac{||A^k\bar{x}||_2}{(n+1)^{3k}}\\
&\leq \frac{(n+1)^3}{4} \sum_{k=0}^{\infty} \frac{||A||_2^k\cdot || \bar{x}||_2}{(n+1)^{3k}}\\
&\leq \frac{\sqrt{n}(n+1)^3}{4} \sum_{k=0}^{\infty} \frac{1}{(n+1)^{2k}}\\
&\leq \frac{(n+1)^4}{4\sqrt{n}}
\end{aligned}$$ where, much as was done in the proof of Lemma \[lem:bounds\], the first inequality uses the triangle inequality, the second the definition and the submultiplicativity property of the matrix 2-norm, the third the facts that $||\bar{x}||_2 \leq \sqrt{n}||\bar{x}||_\infty \leq \sqrt{n} $ and $||A||_2 \leq n+1$, and the last the formula for the sum of a geometric series and the fact that $\frac{n+1}{n+2} \leq 1$. Hence: $$\begin{aligned}
1+||C^{-1}\bar{x}||_2^2&{\gh \leq}1+ \frac{(n+1)^8}{16n} \leq 1+\frac{(2n)^8}{16n} {\gh =}1+16n^7.
\end{aligned}$$ Combining this with (\[eq:expand.hessian\]) and the fact that $$\begin{aligned}
\mu+\frac14 -\bar{x}^T C^{-1}\bar{x} \leq -\frac12,
\end{aligned}$$ we get $$z^T\nabla^2 f(\bar{x},\bar{y})z\leq -\frac{1}{2(1+16n^7)}+\eta.$$ Replacing $\eta$ by its expression in (\[def:n.alpha.gamma\]), we see that $z^T \nabla^2 f(\bar{x},\bar{y})z \leq -\frac{1}{4} \cdot \frac{1}{1+16n^7}<0.$ Hence $\nabla^2 f(x,y)$ is not positive semidefinite over $B$, and therefore $f(x,y)$ is not convex over $B$.
Suppose now that $L(x) \succeq 0$ for all $x \in [-1,1]^n.$ We prove that $f(x,y)$ is convex over $B$ by showing that $\nabla^2 f(x,y) \succeq 0$ for all $(x,y) \in B.$ As $\alpha>0$, this is equivalent to showing, using the Schur complement , that $$L(x)+\eta I_{n+1} -\frac{1}{4\alpha} H(y)^TH(y) \succeq 0, \text{ for all } (x,y) \in B.$$ We prove that $\eta I_{n+1}-\frac{1}{\alpha} H(y)^TH(y) \succeq 0$ for all $y \in [-1,1]^{n+1}$, and as $L(x) \succeq 0$ for any $x \in [-1,1]^n$, the claim would follow. From (\[eq:Hy\]), we deduce the following expression for the symmetric $(n+1) \times (n+1)$ matrix $H(y)^TH(y)$: $$H(y)^TH(y)=4\begin{bmatrix} y_{n+1}^2 I_n & y_{n+1} \cdot y \\ y_{n+1} \cdot y^T & \sum_{i=1}^n y_i^2 \end{bmatrix}.$$ By Gershgorin’s circle theorem [@gersh], for $y \in [-1,1]^{n+1}$, we can upper bound the largest eigenvalue of this matrix as follows: Using this bound and replacing $\alpha$ and $\eta$ by their expressions in (\[def:n.alpha.gamma\]), we conclude that $$\eta I_{n+1}-\frac{1}{4\alpha} H(y)^TH(y) \succeq \left(\eta-\frac{2n}{\alpha}\right)I_{n+1} = \frac{1}{8(1+16n^7)}I_{n+1} \succeq 0.$$
We end with the observation that Theorem \[th:np.hardness.convex\] immediately classifies the complexity of detecting convexity of a polynomial of degree $d$ over a box, for any integer $d \geq 1$.
\[prop:degree\] The problem of testing whether a polynomial $f$ of degree $d$ is convex over a box $B$ can be solved in polynomial time for $d=1$ and $d=2$ and is strongly NP-hard for any fixed integer $d\geq 3.$
When $d=1$, $f$ is always globally convex. Hence the problem of testing whether $f$ is convex over a box $B$ can trivially be solved in polynomial time (by answering ‘yes’ to all instances). When $d=2$, the Hessian of $f$ is constant, and so testing convexity of $f$ over $B$ is equivalent to testing whether its (constant) Hessian matrix $\nabla^2 f$ is positive semidefinite. This can be done in polynomial time, e.g. by performing Gaussian pivot steps along the main diagonal of $\nabla^2 f$ [@nonnegativity_NP_hard], or by computing the characteristic polynomial of $\nabla^2 f$ exactly and then checking that the signs of its coefficients alternate [@HJ_Matrix_Analysis_Book p. 403].
For $d=3$, Theorem \[th:np.hardness.convex\] establishes the claim. For $d\geq 4$, we give a reduction from the problem of testing convexity of a cubic polynomial $g(x)\mathrel{\mathop{:}}=g(x_1,\ldots,x_n)$ over a full-dimensional[^4] box $\tilde{B} \subseteq \mathbb{R}^n$. Given such $g, \tilde{B},$ let $$f(x_1,\ldots,x_n,x_{n+1})\mathrel{\mathop{:}}=g(x_1,\ldots,x_n)+x_{n+1}^d,$$ and $B\mathrel{\mathop{:}}=\tilde{B}\times [0,1]$. We have $$\nabla^2 f(x,x_{n+1})=\begin{bmatrix} \nabla^2 g(x) & 0\\ 0 & d(d-1)\cdot x_{n+1}^{d-2}\end{bmatrix},$$ and hence $\nabla^2 f(x,x_{n+1})\succeq 0$ over $B$ if and only if $\nabla^2 g(x)\succeq 0$ over $\tilde{B}$ (as $d(d-1)x_{n+1}^{d-2}$ is nonnegative over $[0,1]$). It follows that testing convexity of $f$ over $B$ is strongly NP-hard.
To conclude, we would like to emphasize that our result should not discourage researchers from seeking algorithms for testing convexity of polynomials over a box. a search for a pseudo-polynomial time algorithm that works on all cubic polynomials is hopeless unless P=NP,
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to two anonymous referees whose detailed and constructive feedback has improved this paper significantly.
[^1]: The authors are with the department of Operations Research and Financial Engineering at Princeton University. Email: {`a_a_a`, `gh4`}`@princeton.edu`. This work was partially supported by the DARPA Young Faculty Award, the CAREER Award of the NSF, the Google Faculty Award, the Innovation Award of the School of Engineering and Applied Sciences at Princeton University, and the Sloan Fellowship.
[^2]: Note that the instances of not all of NP-hard problems involve numerical data; consider, e.g., SATISFIABILITY [@GareyJohnson_Book]. For such problems, NP-hardness is by default in the strong sense.
[^3]: Recall that a graph is *simple* if it is unweighted, undirected, and has no self-loops or multiple edges.
[^4]: Note that the proof of Theorem \[th:np.hardness.convex\] established strong NP-hardness of this problem as the box that arose in the proof was the unit hypercube, which is full-dimensional.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- '**I.T. Habibullin and A.R. Khakimova**'
title: |
**On a direct algorithm for constructing recursion\
operators and Lax pairs for integrable models**
---
Introduction
============
In a series of our works we have suggested a method for construction the recursion operators and the Lax pairs for the nonlinear integrable equations (see [@HabKhaPo]-). Let us give a brief explanation of core of the method. For the sake of convenience we first take a class of evolutionary type nonlinear PDE although as it is shown below the algorithm can be applied to any kind of the integrable equations. Let us consider an integrable equation of the form $$\label{eq0}
u_t =f(u,u_1,u_2,...,u_k), \qquad \mbox{where} \quad u_j=D_x^{j} u.$$ Here $D_x$ stands for the operator of the total derivative with respect to the variable $x$. In what follows we use the linearization of the equation (\[eq0\]) around its arbitrary solution $u(x,t)$: $$\label{eq1}
U_t=F_*U, \quad F_*=\left(\frac{\partial f}{\partial u}+\frac{\partial f}{\partial u_1}D_x+\frac{\partial f}{\partial u_2}D_x^2+...+\frac{\partial f}{\partial u_k}D_x^k\right).$$ We look for an ordinary differential equation $$\label{eq2}
H(x,t,U,U_1,\dots,U_m;u,u_1,\dots,u_{m_1})=0$$ of the order $m\geq 1$ compatible with the linearized equation (\[eq1\]) for all values of the dynamical variables $u,u_1,u_2,...$ considered here as parameters. The compatibility condition of the equations (\[eq1\]) and (\[eq2\]) coincides with the equation $$\label{eq3}
D_t H(x,t,U,U_1,\dots,U_m;u,u_1,\dots,u_{m_1})=0 \quad\mbox{mod} \,(1),\, (2),\, (3).$$ It is supposed that in (\[eq3\]) all of the derivatives are expressed due to the equations (\[eq0\]) and (\[eq1\]) and the variables $U_{s+m}$ with $s\geq 0$ are expressed due to the equation (\[eq2\]).
In this case we say that equation (\[eq2\]) defines an invariant manifold in the space of the variables $U,\,U_1,\,U_2,\dots$. By the construction equations (\[eq1\]) and (\[eq2\]) are compatible if $u=u(x,t)$ is a solution of the equation (\[eq0\]). It is remarkable that for an appropriate choice of the equation (\[eq2\]) the converse is also true: the compatibility of the equations (\[eq1\]) and (\[eq2\]) implies (\[eq0\]). As it is commonly known just this property is principal for the Lax pairs. Studied examples convince that the ODE (\[eq2\]) might be of two different kinds. Namely function $H$ defining equation (\[eq2\]) is either linear or quadratic $$\begin{aligned}
&& H=\sum_{j=0}^{j=m}\alpha_j(u,u_1,\dots)U_j,\label{linear}\\
&& H=c+\sum_{i,j=0}^{j=s}\alpha_{i,j}(u,u_1,\dots)U_iU_j.\label{quadratic}\end{aligned}$$ Here $c$ is an arbitrary constant. We observed that the former case is connected with the recursion operator $R$ for the equation (\[eq0\]) while the latter corresponds to the usual Lax pair for $c=0$ and a nonlinear Lax pair for $c\neq0$ from which the well known Dubrovin-Weierstrass equations are derived connected with the finite-gap integration procedure.
Recall that the recursion operator is a solution of the equation (\[generalrecursion\]) (see below). In general it is a pseudo-differential (pseudo-difference for the lattices) operator generating symmetries of the equation (\[eq0\]). The Lax pair and the recursion operator are very close to each other. As it follows from (\[generalrecursion\]) the operators $R$ and $\frac{d}{dt}-F_*$ generate a Lax pair: $$\label{eq4}
R\Psi=\lambda\Psi, \quad \frac{d}{dt}\Psi=F_*\Psi$$ for the equation (\[eq0\]) (see [@ShabatIbragimov; @Sul]). Moreover, as it is demonstrated in [@HabKhaJPA17; @HabKhaTMP17] by numerous examples the Lax pair (\[eq4\]) is effectively converted into the classical one. A clear expression of this circumstance is that a linear invariant manifold is transformed into a nonlinear one by decreasing order in the corresponding ordinary differential equation (\[eq2\]).
Note that there is a great variety of approaches for searching the Lax pairs from the Zakharov-Shabat dressing [@ZakharovShabat74; @ZakharovShabat79] and the method of pseudopotentials by Wahlquist and Estabrook [@WahlquistEstabrook] to the singular manifold method [@Weiss; @Conte] of Painlev$\acute{e}$ analysis and 3D consistency approach developed in [@NijhoffWalker; @BobenkoSuris; @Nijhoff]. We mention also approaches proposed in [@Yam82; @Xenitidis; @ShabatIbragimov]. We stress that our method uses a close idea but differs from the well-known method of [@WahlquistEstabrook], where both of the Lax equations are assumed to be linear and unknown. In contrast we for a given integrable equation take its linearization as one of the Lax equations and look for the second one which is not assumed to be linear. Actually at the first stage we find a nonlinear Lax pair and then linearize it by an appropriate point transformation. Since we search only one of two equations our method is rather effective.
The problem of finding the recursion operator for an integrable equation is not easily solved. There are several methods to study the task. Some of them use the Lax representation (see, for instance, [@Gelfand; @Magri; @Sokolov; @Gurses; @Gardner; @ShabatIbragimov]). Others are based directly on studying the defining equation (\[generalrecursion\]). To solve this equation the most authors use the multi-Hamiltonian approach [@Krasilshchik; @Oevel; @Zhang; @Gurses2; @Svinolupov]. Their basic aim is to determine two Hamiltonian operators $H_1$ and $H_2$ to the given equation. Then the sought recursion operator is given by a simple formula $$\label{Hamilt}
R=H_2H_1^{-1}.$$
In the present article we use an alternative representation for the recursion operator $$\label{Symm}
R=L_1^{-1}L_2,$$ where $L_1$ and $L_2$ are differential (difference) operators chosen in such a way that the equations $L_1U=0$ and $L_2U=0$ define linear invariant manifolds . Evidently two representations (\[Hamilt\]) and (\[Symm\]) of the recursion operator are essentially different.
Let us briefly discuss on the content of the article. In §2 we study the close connection between the symmetries of the given nonlinear equation and the invariant manifolds of its linearization which provides an easy way to construct linear invariant manifolds. Then in §3 we give an explanation of the consecutive algorithm for defining the operators $L_1$ and $L_2$ used in the formula (\[Symm\]). In §§3.1-3.4 we illustrate the application of the algorithm with examples of the KdV equation, Volterra type autonomous and nonautonomous lattices and a system of two lattices. A direct method of finding and some applications of the nonlinear invariant manifolds in the integrability theory are demostrated in §4.
Invariant manifolds and symmetries
==================================
We observed that there exists a close connection between symmetries and the invariant manifolds.
[**Proposition 1.**]{} [*Any set of the symmetries $$\label{Symms}
u_{\tau_j}=U^{(j)},\quad j=1,2,\dots ,k$$ of the equation (\[eq0\]) defines an invariant manifold of the form (\[eq2\]) for the linearized equation (\[eq1\]). Here $U^{(j)}$ is a solution to the equation (\[eq1\]), depending on a finite number of the dynamical variables.*]{}
Proof. Let us consider a linear space spanned by the functions $U^{(j)}$. Then any element $U$ of the space is represented as the linear combination $$\label{SymmsComb}
U=c_1U^{(1)}+c_2U^{(2)}+\dots +c_kU^{(k)}.$$ Here we suppose that the functions $U^{(j)}$ are linearly independent. Let us consider equation (\[SymmsComb\]) as the general solution of a differential equation of the form $$\label{SymmDif}
L_1U:=(D_x^k+a^{(1)}D_x^{k-1}+\dots +a^{(k)})U=0.$$ By construction $U$ solves also the linearized equation (\[eq1\]), therefore (\[SymmDif\]) defines a linear invariant manifold. The coefficients of this linear differential equation are evidently found from the equation $$\label{SymmsMat}
L_1U=\det\left( \begin{array}{cccc}
U^{(1)}_k&U^{(1)}_{k-1}&\dots &U^{(1)}\\U^{(2)}_k&U^{(2)}_{k-1}&\dots &U^{(2)} \\\dots &\dots&\dots& \dots \\ U^{(k)}_k&U^{(k)}_{k-1}&\dots &U^{(k)}\\U_k&U_{k-1}&\dots &U \end{array} \right),$$ where $U_m=D^m_xU$ and $U^{(j)}_m=D^m_x U^{(j)}$. The proof is complete.
On the other side for any common solution $U$ of the equations (\[eq1\]) and (\[eq2\]) the equation $u_{\tau}=U$ defines a symmetry (possibly nonlocal) for the equation (\[eq0\]).
Algorithm for constructing linear invariant manifolds and the recursion operators for integrable equations
==========================================================================================================
Invariant manifolds can be found directly by solving the defining equation (\[eq2\]) which is highly overdetermined and usually effectively solved. However this way becomes rather labor consuming when applied to the systems or higher order equations.
To look for a linear invariant manifold we suggest a novel convenient algorithm. We observed that the linear invariant manifold is closely connected with the recursion operator $R$ of the equation (\[eq0\]). Let us consider the equation of the form $$\begin{aligned}
R U = \lambda U \label{rec}\end{aligned}$$ and then bring this equation by multiplying from the left by a differential operator $L_1$ to the form $$\begin{aligned}
L_2 U = \lambda L_1U \label{rec-inv}\end{aligned}$$ where $L_2$ is also a differential operator. Evidently the operator $L_1$ is not unique. We require that the order of $L_1$ is as small as possible. Then equation (\[rec-inv\]) defines an invariant manifold for arbitrary value of $\lambda$ including $\lambda=0$ and $\lambda=\infty$. Therefore we can easily obtain a linear invariant manifold for the known recursion operator. However, we are going to apply this scheme in the opposite direction.
Here we use an assumption that the recursion operator admits the representation $R=L_1^{-1}L_2$ as a ratio of two differential operators defining the invariant manifolds like $L_1U=0$ and $L_2U=0$. It is well known that usually the recursion operator satisfies the following equation [@Olver] $$\label{generalrecursion}
\frac{d}{dt}R=[F_*,R].$$ Let us replace $R=L_1^{-1}L_2$ in (\[generalrecursion\]) and get after a slight transformation $$\label{temp}
\frac{d}{dt}L_2=\left(\frac{d}{dt}(L_1)L_1^{-1}+L_1F_*L_1^{-1}\right)L_2-L_2F_*$$ where $F_*$ is the linearization operator (\[eq1\]). Operator $L_1$ can easily be found due to the formula (\[SymmsMat\]) above through the symmetries of the equation (\[eq0\]). Let us define a new operator $A$ according to the equation $$\label{generalA}
A:=\frac{d}{dt}(L_1)L_1^{-1}+L_1F_*L_1^{-1}.$$ Then (\[temp\]) immediately gives rise to the equation $$\label{generalL2}
\frac{d}{dt}L_2=AL_2-L_2F_*$$ which allows to find the operator $L_2$. As a result we obtain the recursion operator $R=L_1^{-1}L_2$ solving the equation (\[generalrecursion\]).
[**Remark**]{}. Equations (\[generalA\]) and (\[generalL2\]) generate two Darboux-type transformations $U\rightarrow V=L_1U$ and $U\rightarrow V=L_2U$ relating linearized equation (\[eq1\]) with one and the same linear equation $V_t=AV$.
Below we approve the efficiency of the algorithm with several examples.
Illustration with the KdV equation
----------------------------------
Consider the KdV equation $$\label{kdv_eq}
u_t=u_{xxx}+uu_x.$$ Its linearization looks as follows $$\label{lin_kdv}
U_t=U_{xxx}+uU_x+u_xU.$$ In this case $L_1$ can be defined by the formula (\[SymmsMat\]) from Proposition 1 with $k=1$ and with the symmetry $U^{(1)}=u_x$. We set $L_1=D_x\frac{1}{u_x}$ since this operator evidently annulates $u_x$.
Let us define the operator $A$ due to the formula $$\label{A}
A=D_t(L_1)L_1^{-1}+L_1F_*L_1^{-1}$$ where $F_*=D_x^3+uD_x+u_x$ is the linearization operator and evidently $L_1^{-1}=u_xD_x^{-1}$. Equation (\[A\]) implies $$\label{AA}
A=D_x^3+ \left(u-3\left(\frac{u_{xxx}}{u_x}- \frac{u_{xx}^2}{u_x^2}\right)\right)D_x+2 \frac{uu_{xx}}{u_x}-3 \frac{u_{xx}u_{xxx}}{u_x^2}-3 \frac{u_{xx}^3}{u_x^3}.$$ Now we have to look for an operator of the order $m>1$: $$\label{L2}
L_2=\alpha D^{m}_x+\alpha_1D_x^{m-1}+\cdots+\alpha_m$$ solving the following equation $$\label{L2L2}
D_t(L_2)=AL_2^{-1}+L_2F_*.$$ It is easily proved that for $m=2$ equation (\[L2L2\]) has no solution of the form (\[L2L2\]). For $m=3$ we have $$\label{L2found}
L_2= D^3_x - \frac{u_{xx}}{u_x} D^2_x + \frac{2 u}{3} D_x + u_x - \frac{2 u u_{xx}}{3 u_x}.$$ Direct computations based on the equalities (\[A\]), (\[L2L2\]) convince that the operator $R=L_1^{-1}L_2$ solves the equation $D_t(R)=[F_*,R]$ and therefore $R$ is the recursion operator for the KdV equation. Indeed we find from the formula $R=L_1^{-1}L_2$ that (cf. [@Gardner]) $$R=D_x^2+\frac{2}{3}u+\frac{1}{3}u_xD_x^{-1}.$$
Evaluation of the invariant manifolds and the recursion operators for non-autonomous Volterra type integrable lattices
----------------------------------------------------------------------------------------------------------------------
Due to the articles [@Yam; @Levi] we know that integrable lattices of the form $$\label{volterratype}
u_{n,t}=f(n,u_{n+1},u_n,u_{n-1})$$ satisfy a sequence of the integrability conditions. Below we mention some of these conditions which have important applications in the symmetry classification $$\begin{aligned}
&& D_t\log \frac{\partial f}{\partial u_1}=(D-1)q^{(1)},\label{ccl1}\\
&& D_t\log \frac{\partial f}{\partial u_{-1}}=(D-1)q^{(-1)}, \label{ccl2}\\
&& r^{(k)}=(D-1)s^{(k)},\quad k=1,2,\label{ccl3}\end{aligned}$$ where $D$ is the shift operator, such that $Dy(n)=y(n+1)$ and $q^{(1)}$, $q^{(-1)}$, $s^{(1)}$, $s^{(2)}$ are some functions depending on the dynamical variables $u_n$, $u_{n\pm 1}$, $u_{n\pm 2}$, …. In the formulas above we used the following commonly accepted abbreviated notations, which are used also everywhere below: $u:=u_{n},\,u_1:=u_{n+1},\, u_{-1}:=u_{n-1}$ and so on. Here $$\begin{aligned}
&&r^{(1)}=\log \left\{-\frac{f_{u_1}}{f_{u_{-1}}}\right\}, \label{r1}\\
&&r^{(2)}=s_{t}^{(1)}+2f_u \label{r2}.\end{aligned}$$
The problem of the complete classification for autonomous lattices of the form (\[volterratype\]) has been solved years ago (see [@Yam82]). Recursion operators for these equations in the autonomous case are discussed in [@Mikh]. Below we show that the method of invariant manifolds can be successfully applied to the integrable lattices in both autonomous and non-autonomous case. Moreover, we derive a general formula for the recursion operator which fits for a subclass of the integrable lattices (\[volterratype\]).
First we find the linearization of the equation (\[volterratype\]) $$\label{linvolterratype}
U_t=\frac{\partial f}{\partial u_1}U_1+\frac{\partial f}{\partial u}U+\frac{\partial f}{\partial u_{-1}}U_{-1}.$$
Notion of the invariant manifold introduced above for the evalutionary type PDE is easily adopted to other classes of equations.
We say that an ordinary difference equation of the form $$\label{difference}
H(n,t,U_{m},\dots U_{m_1};u_{l},\dots u_{l_1})=0$$ where $U=U(n,t)$ is considered as an unknown function and the variables $u_j$ are considered as parameters, determines an invariant manifold for the linearized equation (\[linvolterratype\]) if the equations (\[linvolterratype\]), (\[difference\]) are consistent for any choice of the solution $u=u(n,t)$ of the equation (\[volterratype\]).
The simplest invariant manifold is found very easily, to this end we use the classical symmetry $u_\tau=cu_t$ of the equation (\[volterratype\]). Evidently the function $U=cu_t$ determines the general solution of the linear difference equation $\displaystyle{(D-1)\frac{1}{u_t}U=0}$. Therefore equation $L_1U=0$ with $\displaystyle{L_1=(D-1)\frac{1}{u_t}}$ defines an invariant manifold for the linearized equation (\[linvolterratype\]).
Let us first consider only those equations for which the operators $L_1$ and $L_2$ are of the form $$\label{simpleL1L2}
L_1=(D-1)\frac{1}{u_t} \quad \mbox{and}\quad L_2=\alpha D^2+\beta D+\gamma+\delta D^{-1}$$ where the coefficients $\alpha, \beta, \gamma, \delta$ are functions of the dynamical variables ${u_j}$. Note that this class contains all of the lattices in the Yamilov’s list except the equation V4 for which $L_1$ is the second order difference operator which annulates $u_t$ and the next symmetry of the equation V4. Introduce a notation for the linearization operator of the lattice (\[volterratype\]) by setting $f^*=f_{u_1}D+f_{u}+ f_{u_{-1}}D^{-1}$. Operator $A$ is defined due to the formula $A:=\frac{d}{dt}(L_1)L_1^{-1}+L_1f^*L_1^{-1}$ (cf. (\[generalA\])) $$\label{volterrratypeA}
A=(D-1)\left(\frac{f_{u_1}f_1}{f}-\frac{f_{u_{-1}}f_{-1}}{f}D^{-1}\right).$$ Now we look for the operator $L_2$ from the following equation $$\label{volterratypeL2}
\frac{d}{dt}L_2=AL_2-L_2f^*.$$
Let us rewrite (\[volterratypeL2\]) in an explicit form $$\begin{aligned}
\label{volterratypeL2L2}
\alpha_t D^2+\beta_t D+\gamma_t+\delta_t D^{-1}&&=(D-1)\left(\frac{f_{u_1}f_1}{f}-\frac{f_{u_{-1}}f_{-1}}{f}D^{-1}\right)\left(\alpha D^2+\beta D+\gamma+\delta D^{-1}\right)\nonumber \\
&&- (\alpha D^2+\beta D+\gamma+\delta D^{-1})(f_{u_1}D+f_{u}+ f_{u_{-1}}D^{-1}).\label{+6}\end{aligned}$$
By comparing the coefficients in front of $D^3$ and $D^{-2}$ in this equation we find $$\label{+7}
\alpha=\left( \frac{f_{u_1}}{f}\right)_1,\quad \delta=\frac{f_{u_{-1}}}{f}.$$ Recall that $h_1$ means $D(h)$. Comparison of the coefficients at $D^2$ gives rise to the equation $$\label{volterrratypeD2}
\alpha_t=\left(\frac{f_{u_1}f_1}{f}\right)_1\beta_1-\left(\frac{f_{u_1}f}{f_1}+ \frac{(f_{u_{-1}})_1f}{f_1} \right)\alpha-\alpha(f_u)_2-\beta(f_{u_1})_1.$$ Now we divide (\[volterrratypeD2\]) by $\alpha$ and get $$D_t\log \alpha=(D-1)\beta f_1- \left(\frac{f_{u_1}f_1}{f}+ \frac{{(f_{u_{-1}})_1}f}{f_1} \right)-(f_u)_2.$$ Let us evaluate l.h.s. of the last equation due to (\[+7\]) and find
$$\label{intermediate}
D_t\log \alpha=D(D_t\log f_{u_1}-D_t\log f).$$
By combining the equations (\[volterrratypeD2\]), (\[intermediate\]) and replacing due to (\[ccl1\]) we obtain the equation $(D-1)(\beta f_1 -q^{(1)}_1+\frac{f_{u_1}f_1}{f}-(f_u)_1)=0$ which implies
$$\label{2beta}
\beta= \frac{1}{f_1}q^{(1)}_1-\frac{f_{u_1}}{f}-\frac{(f_u)_1}{f_1}.$$
In a similar way by gathering the coefficients at $D^{-1}$ and then applying the conservation law (\[ccl2\]) one can derive an explicit expression for $\gamma$: $$\label{gamma}
\gamma= -\frac{1}{f}q^{(-1)}_1+\frac{f_{u}}{f}-\left(\frac{(f_{u_{-1}})_1}{f_1}\right).$$
Thus all the coefficients of the sought operator $L_2$ are found, however to approve that equation (\[volterratypeL2L2\]) is completely satisfied we have to check the last two equations which are obtained by collecting the coefficients in front of $D$ and $D^0$
$$\label{3beta}
\beta_t= \left(\frac{f_{u_1}f_1}{f}\right)_1\gamma_1-\left(\frac{f_{u_1}f_1}{f}+ \frac{(f_{u_{-1}})_1f}{f_1} \right)\beta
+\frac{f_{u_{-1}}f_{-1}}{f}\alpha_{-1}
-\alpha(f_{u_{-1}})_2-\beta(f_{u})_1 -\gamma f_{u_1},$$
$$\label{3gamma}
\gamma_t= \left(\frac{f_{u_1}f_1}{f}\right)_1\delta_1-\left(\frac{f_{u_1}f_1}{f}+ \frac{(f_{u_{-1}})_1f}{f_1} \right)\gamma
+\frac{f_{u_{-1}}f_{-1}}{f}\beta_{-1}
-\beta(f_{u_{-1}})_1-\gamma(f_{u}) -\delta (f_{u_1})_{-1}.$$
Assume the coefficients $\alpha$, $\beta$, $\delta$, and $\gamma$ found above to satisfy equations (\[3beta\]) and (\[3gamma\]). Then the operator $R=L_1^{-1}L_2$: $$\label{rec2}
R=f(D-1)^{-1}(\alpha D^2+\beta D+\gamma+\delta D^{-1})$$ is the recursion operator for the lattice (\[volterratype\]). Indeed, it is easily verified that the found $R$ solves the following equation $$\label{rec-eq}
R_t=[f^*, R]$$ which is nothing else but the defining equation for the recursion operator (see the survey [@Yam]).
Let us take the Volterra chain $$\label{examplevolterra}
u_t=u(u_1-u_{-1})$$ as an illustrative example, then the operators $L_1$ and $L_2$ are as follows $$\begin{aligned}
L_1&=&(D-1)\frac{1}{u(u_1-u_{-1})},\nonumber \\
L_2&=&\frac{1}{u_2-u}D^2+\left(\frac{u_2+u_1}{u_1(u_2-u)}-\frac{1}{u_1-u_{-1}}\right)D+\frac{1}{u_2-u}- \label{volterraL12}\\
&&-\frac{u+u_{-1}}{u(u_1-u_{-1})}-\frac{1}{u_1-u_{-1}}D^{-1}.\nonumber\end{aligned}$$ Therefore the formula $R=L^{-1}_1L_2$ gives the well known recursion operator $$\label{RecVolterra}R=uD+u+u_1+uD^{-1}+u(u_1-u_{-1})(D-1)^{-1}\frac{1}{u}$$ found years ago in . Note that the operators $L_1$ and $L_2$ differ from the Hamiltonian operators $H_1$, $H_2$ used in to represent the recursion operator (\[RecVolterra\]) as a ratio $R=H_2H_1^{-1}$ where $$H_1=u(D-D^{-1})u, \quad H_2=u(DuD+uD+Du-uD^{-1}- D^{-1}u-D^{-1}uD^{-1})u.$$
We studied examples of the lattices above under assumption that $L_1$ is the first order operator (see (\[simpleL1L2\]) above). However for some lattices this assumption is too restrictive, such that the needed $L_2$ does not exist. Then we take $L_1$ of the order higher than one. The Proposition 1 discussed above can be applied to the case of lattices as well.
[**Proposition 2.**]{} [*Any set of the symmetries $$\label{SymmsD}
u_{\tau_j}=U^{(j)},\quad j=1,2,\dots ,k$$ of the equation (\[volterratype\]) defines an invariant manifold of the form (\[difference\]) for the linearized equation (\[linvolterratype\]). Here $U^{(j)}$ is a solution to the equation (\[linvolterratype\]), depending on a finite number of the dynamical variables.*]{}
We omit the proof of Proposition 2 since it is almost verbatim repeat of the proof of Proposition 1. The searched invariant manifold is defined by the following equation $$\label{SymmsMatD}
L_1U=\det\left( \begin{array}{cccc}
U^{(1)}_k&U^{(1)}_{k-1}&\dots &U^{(1)}\\U^{(2)}_k&U^{(2)}_{k-1}&\dots &U^{(2)} \\\dots &\dots&\dots& \dots \\ U^{(k)}_k&U^{(k)}_{k-1}&\dots &U^{(k)}\\U_k&U_{k-1}&\dots &U \end{array} \right)=0,$$ where $U_m=D^mU$ and $U^{(j)}_m=D^mU^{(j)}$.
Example of a recursion operator for the non-autonomous lattice
--------------------------------------------------------------
In this section we consider a non-autonomus lattice of the form $$\label{nonaut_eq}
u_t=\frac{(-1)^{n+m}u(u^2-1)(u_1u_{-1}+1)}{ww_1}, \quad w=uu_{-1}-u_{-1}+u+1$$ found in as a symmetry in the direction of $n$ of the quad equation $$\label{quad_nonaut_eq}
u_{n+1,m+1}u_{n,m}(u_{n+1,m}-1)(u_{n,m+1}+1)+(u_{n+1,m}+1)(u_{n,m+1}-1)=0$$ derived in . Here our goal is to find the recursion operator for the lattice (\[nonaut\_eq\]) according to the scheme above. Let us set for the simplicity $m=0$. We first find the linearization of (\[nonaut\_eq\]) $$\label{nonaut_eq_linform}
U_t=f^*U$$ where $f^*$ is the Frech$\acute{e}$t derivative of $f$ $$\begin{aligned}
\label{nonaut_eq_linop}
&f^*=(-1)^{n+1}\frac{u(u^2-1)}{w_1}D+\\
&(-1)^n\frac{u_1u_{-1}+1}{ww_1}\left(3u^2-1-\frac{u(u^2-1)(u_1-1)}{w_1}-\frac{u(u^2-1)(u_{-1}+1)}{w_{-1}}\right)+(-1)^n\frac{u(u^2-1)}{w^2}D^{-1}.\nonumber\end{aligned}$$ We take the operator $L_1$ as in (\[linvolterratype\]) and then find $A$ due to the formula (\[simpleL1L2\]): $$\label{nonaut_eq_A}
A=A^{(1)}D+A^{(0)}+A^{(-1)}D^{-1}$$ where $$\begin{aligned}
\label{nonaut_eq_Acoeff}
&&A^{(1)}=(-1)^{n+1}\frac{u_2(u^2_2-1)(u_3u_2+1)w_1}{(u_2u+1)w^2_2w_3}, \nonumber\\
&&A^{(0)}=(-1)^{n+1}\frac{u(u^2-1)(u_1u_{-1}+1)w_2}{(u_2u+1)w^2_1w}+(-1)^{n+1}\frac{u_1(u^2_1-1)(u_2u+1)w}{(u_1u_{-1}+1)w^2_1w_2}, \\
&&A^{(-1)}=(-1)^{n+1}\frac{u_{-1}(u^2_{-1}-1)(uu_{-2}+1)w_1}{(u_1u_{-1}+1)w^2w_{-1}}.\nonumber\end{aligned}$$ The next step is to look for the operator $L_2$ from the equation (\[volterrratypeA\]). We find the coefficients of $L_2$ by using the explicit formulas (\[volterratypeL2L2\]), (\[intermediate\]) and the first canonical conservation laws (\[volterratype\]), (\[ccl1\]). As a result we obtain $$\begin{aligned}
\label{nonaut_eq_L2}
&&L_2=-\frac{w_1}{w_2(uu_2+1)}D^2-\left(\frac{2(u+1)}{(u_1+1)(uu_2+1)}+\frac{w_1w_2}{2u(u_1^2-1)(uu_2+1)}+\right.\nonumber\\
&&\left.\frac{uu_1-1}{u_1(u_1^2-1)(uu_2+1)}-\frac{2u_1}{u_1^2-1}-\frac{1}{u_1(uu_{-1}+1)}+\frac{u_2-1}{w_2}\right)D+\nonumber\\
&&\left(\frac{2(u_1-1)}{(u-1)(u_1u_{-1}+1)}-\frac{ww_1}{2u(u^2-1)(u_1u_{-1}+1)}-\frac{(u_1u+1)}{u(u^2-1)(u_1u_{-1}+1)}+\right.\\
&&\left.\frac{2u}{u^2-1}+\frac{1}{u(u_2u+1)}-\frac{u_{-1}+1}{w}\right)+\frac{w_1}{(u_1u_{-1}+1)w}D^{-1}.\nonumber\end{aligned}$$ The important step is to verify whether the corresponding equations of the form (\[3beta\]), (\[3gamma\]) are satisfied. We checked that in this case these equations hold. Thus we have the recursion operator $R=L_1^{-1}L_2:$ $$\begin{aligned}
\label{nonaut_eq_R}
&&R=(-1)^{n+1}\frac{u(u^2-1)}{w_1^2}D+(-1)^{n+1}\left(\frac{u(u^2-1)(u_1-1)(u_1u_{-1}+1)}{ww^2_1}+\frac{w}{2u_{-1}w_1}-\right.\nonumber\\
&&\left.\frac{2(u_{-1}-1)(u+1)}{u_{-1}ww_1}+\frac{(u-1)(u_1u_{-1}+1)}{u_{-1}ww_1}+\frac{2u(u-1)(u_{-1}-1)(u_1u_{-1}+1)}{w_1w^2}+\right.\\
&&\left.\frac{u_{-1}-1}{u_{-1}w_1}-\frac{u+1}{u_{-1}w_1}-\frac{(u-1)(u_{-1}^2)(u^2-1)(u_1u_{-1}+1)}{2u_{-1}w_1w^2}\right)+(-1)^{n+1}\frac{u(u^2-1)}{w^2}D^{-1}-\nonumber\\
&&u_t(D-1)^{-1}\left(\frac{2(u_{-1}+1)}{w}+\frac{2(u_1-1)}{w_1}+\frac{3u^2-1}{u(u^2-1)}\right).\nonumber\end{aligned}$$
Recursion operator for a coupled lattice
----------------------------------------
Let us discuss the algorithm for constructing the recursion operator with the example of a system of lattices [@GHYan] Linearized system can be represented as where the linearization operator or the Fr$\acute{e}$chet derivative $F_n^*$ is given by where
In the coordinate representation the linearized equation looks as follows $$\label{linearizedcoupled}
U_t=-\frac{1}{v_1^2}V_1+2uvU+u^2V,\quad V_t=\frac{1}{u_{-1}^2}U_{-1}-2uvV-v^2U.$$ In order to construct the linear operator $L_1$ we use the classical symmetry $u_\tau=cu, v_\tau=-cv$ which is connected evidently with the following solution to the linearized equation $U=cu, V=-cv$. By excluding the constant parameter $c$ we arrive at the system of equations $(D-1)\frac{1}{v}V=0$, $U=\frac{u}{v}V,$ defining an invariant manifold. Thus we have We find the operator $A$ from the equation $A=D_t(L_1)L_1^{-1}+L_1F^*L_1^{-1}$ Then we look for the operator $L_2=aD+b+cD^{-1}$ from the equation $D_t(L_2)=AL_2+L_2F^*$. The answer is Therefore the recursion operator is given by $R=L_1^{-1}L_2$. Let us write it down in an explicit form [@HabKhaTMP17] where
Quadratic invariant manifolds, Lax pairs and Dubrovin-Weierstrass equations
===========================================================================
Let us illustrate the application of the nonlinear invariant manifolds for constructing the Lax pairs to integrable models with the example of the KdV equation $$\label{kdv2}
u_t=u_{xxx}+uu_x.$$ By the substitution $U=W_x$ we change the linearized equation $$\label{kdvlinW}
U_t=U_{xxx}+uU_x+u_xU \quad \Rightarrow \quad W_t=W_{xxx}+uW_x.$$ and find ODE compatible with the equations above (see also [@HabKhaJPA17]) $$\label{kdvFW}
W_{xx}=F(W_x,W,u)$$ i.e. we request that $$\label{kdvFWtlinWxx}
\left.\frac{d}{dt}(W_{xx})-\frac{d}{dx}(W_t)\right|_{(\ref{kdv2}),(\ref{kdvlinW}),(\ref{kdvFW})}=0.$$ Eq. (\[kdvFWtlinWxx\]) must satisfy identically for all values of $W$, $W_x$, $u$, $u_x$, $u_{xx}$, ... The consistency condition is reduced to a huge equation, which splits down into 7 equations and is effectively solved $$\begin{aligned}
\label{eq_found_F_compl_kdv}
&&3F_{uu}u_xu_{xx}+(3W_xF_{Wu}+W_x+3F_{W_xu}F)u_{xx}+(3F_{Wuu}W_x+3F_{W_xuu}F+3F_{W_xu}F_u)u_x^2 \nonumber\\
&&+F_{uuu}u_x^3+(3F_{W_x}F_{W_xu}F+2F+3F_{WWu}W_x^2-F_{W_x}W_x+3W_xF_{W_xu}F_W+6W_xF_{W_xWu}F\nonumber\\
&&+3W_xF_{W_xW}F_u+3F_{Wu}F+3FF_{W_xW_x}F_u+3F_{W_xW_xu}F^2)u_x+F_{WWW}W_x^3+F_{W_xW_xW_x}F^3 \\
&&+3F_{W_xW}F^2+3W_xF_{W_x}F_{W_xW}F+3W_xFF_{W_xW_x}F_W+3F_{W_x}F^2F_{W_xW_x}+3W_x^2F_{W_xWW}F\nonumber\\
&&+3W_x^2F_{W_xW}F_W+3W_xF_{WW}F+3W_xF_{WxWxW}F^2=0.\nonumber\end{aligned}$$
Compare the coefficients at the independent variables $u_{xx},u_x$:
1\. $u_{xx}u_x:$ $$\label{1}
\frac{d^2}{du^2}F(W_x,W,u)=0 \Rightarrow F(W_x,W,u)=F_1(W_x,W)u+F_2(W_x,W)$$ 2. $u_{xx}u:$ $$\label{2}
\frac{d}{dW_x}F_1(W_x,W)=0 \Rightarrow F_1(W_x,W)=F_1(W)$$ 3. $u_{xx}:$ $$\label{3}
3\frac{d}{dW}F_1(W)+1=0 \Rightarrow F_1(W)=-\frac{1}{3}W+c_1$$ 4. $u_{x}u:$ $$\begin{aligned}
\label{4}
\begin{aligned}
\frac{d^2}{dW^2_x}F_2(W_x,W)&-\frac{1}{W-3c_1}=0 \Rightarrow \\
&F_2(W_x,W)=\frac{W^2_x}{2(W-3c_1)}+F_4(W)W_x+F_3(W)
\end{aligned}\end{aligned}$$ 5. $u_{x}:$ $$\label{5}
(W-3c_1)\frac{d}{dW}F_4(W)+F_4(W)=0 \Rightarrow F_4(W)=\frac{c_2}{W-3c_1}$$ 6. $uW_x: \quad c_2=0$
7\. $u:$ $$\begin{aligned}
\label{7}
\begin{aligned}
(W-3c_1)^2\frac{d^2}{dW^2}F_3(W)&+(W-3c_1)\frac{d}{dW}F_3(W)-F_3(W)=0 \Rightarrow \\
&F_3(W)=\frac{c_3}{W-3c_1}+c_4(W-3c_1).
\end{aligned}\end{aligned}$$ Finally we obtain ODE $W_{xx}=F(W_x,W,u)$, where: $$\label{8}
F(W_x,W,u)=\left(-\frac{1}{3}W+c_1\right)u+\frac{W^2_x}{2(W-3c_1)}+\frac{c_3}{W-3c_1}+c_4(W-3c_1).$$
Parameter $c_1$ is easily removed by the shift $\bar W=W-3c_1$, therefore we set $c_1=0$. Denote $c_4=\lambda$. $$W_{xx}W-\left(\lambda-\frac{1}{3}u\right)W^2-\frac{1}{2}W^2_x-c_3=0.$$ We have two choices $ c_3=0$ and $ c_3\neq0$.
1\) If $ c_3=0$ then we get a nonlinear Lax pair: $$\begin{aligned}
\label{9}
\begin{aligned}
&W_{xx}=-\frac{1}{3}Wu+\frac{W^2_x}{2W}+\lambda W,\\
&W_t=\left(2\lambda -\frac{2}{3}u\right)W_x-\frac{1}{3}u_xW.
\end{aligned}\end{aligned}$$ It is linearized by taking $W=\varphi^2$ and reduces to the usual pair: $$\begin{aligned}
\label{10}
\begin{aligned}
&\varphi_{xx}=\left(-\frac{1}{6}u+\frac{1}{2}\lambda\right)\varphi,\\
&\varphi_t=\left(2\lambda -\frac{2}{3}u\right)\varphi_x-\frac{1}{6}u_x\varphi.
\end{aligned}\end{aligned}$$
2\) If $c_3\neq 0$, then we get the well-known equation, connected with the finite gap solutions (see [@Dubrovin]): $$W_{xx}W-\left(\lambda-\frac{1}{3}u\right)W^2-\frac{1}{2}W^2_x-c_3(\lambda)=0.$$ We assume that $W$ and $c_3=c_3(\lambda)$ are polynomials of the form: $$W=\prod^{n}_{j=1}(\lambda-r_j), \qquad c_3(\lambda)=-\prod^{2n+1}_{j=1}(\lambda-e_j).$$ Recall that functions $r_j$ satisfy the Dubrovin-Weierstrass equations $$r_{j,x}\prod_{i\neq j}(r_j-r_i)=\sqrt{2\prod^{2n+1}_{s=1}(r_j-e_s)}, j=1,...,n.$$ The potential $u$ is found as follows: $$u=3\sum^{2n+1}_{i=1}e_i-6\sum^{n}_{i=1}r_i.$$
Let us consider the Volterra lattice $$\label{V}
u_{n,t}=u_n(u_{n+1}-u_{n-1}).$$ Linearization of (\[V\]) $$\label{linvolterra}
U_{n,t}=U_n(u_{n+1}-u_{n-1})+u_n(U_{n+1}-U_{n-1})$$ is rather complicated, it contains $u_n$, $u_{n+1}$, $u_{n-1}$. Let us simplify it by the substitution $U_n=u_n(P_{n+1}-P_{n-1})$ which is the linearization of the substitution $\log u_n=p_{n+1}-p_{n-1}$ connecting the Volterra lattice with the equation $p_{nt}=e^{p_{n+1}-p_{n-1}}$. Thus we get from (\[linvolterra\]) the equation $$\label{mlinvolterra}
P_{n,t}=u_n(P_{n+1}-P_{n-1}),$$ which contains only one parameter $u_n$. Look for the invariant manifold in the form $$P_{n+1}=F(P_n,P_{n-1},u_n)$$ i.e. an ordinary difference equation with the parameter $u_n$. The answer is as follows $$\label{inva}
u_n(P_{n+1}+P_n)(P_n+P_{n-1})=\lambda P_n^2+c, \qquad c=const.$$ Assume that functions $P_n$ and $c$ are some polynomials on $\lambda$: $$P_n=\prod^{m}_{j=1}(\lambda-\gamma_j(n)), \qquad c(\lambda)=-\prod^{2m+1}_{j=1}(\lambda-e_j).$$ Then for $\gamma_j(n)$ and $u_n$ we obtain from (\[inva\]) difference equations $$u_n\prod^{m}_{i=1}(\gamma_j(n)-\gamma_i(n+1))(\gamma_j(n)-\gamma_i(n-1))=-2\prod^{2m+1}_{i=1}(\gamma_j(n)-e_i), $$ $$u_n=\frac{1}{4}\sum^{2m+1}_{i=1}e_i-\frac{1}{2}\sum^{m}_{i=1}\gamma_i(n),\qquad j=1,...,m.$$ which are the discrete versions of the Dubrovin-Weierstrass equations (see [@Dubrovin]).
Assume that $c=0$ then the invariant surface takes the form $$\label{V0}
u_n(P_{n+1}+P_n)(P_n+P_{n-1})=\lambda P_n^2.$$ Now from the linearized equation (\[V0\]) we obtain the equation $$\label{V1}
P_{n,t}=u_n\left(\frac{\lambda P_n^2}{P_n+P_{n-1}}-P_n-P_{n-1}\right)$$ describing the time evolution of $P$. Let us derive the Lax pair for the Volterra lattice. Since the equation (\[V0\]) is homogeneous it is reasonable to set $Z=\frac{P}{P_{-1}}$ and reduce (\[V0\]) to the discrete Riccati equation: $$\label{V2}
Z_{n+1}=\frac{Z_n(\frac{\lambda}{u_n}-1)-1}{Z_n+1}.$$ We linearize (\[V2\]) in the standard way by introducing $Z_n+1=\frac{\varphi_n}{\varphi_{n-1}}$: $$\label{V3}
\varphi_{n+1}=\frac{\lambda}{u_n}(\varphi_n-\varphi_{n-1}).$$ By differentiating the equality $\frac{P_n}{P_{n-1}}=\frac{\varphi_n}{\varphi_{n-1}}-1$ we get after some transformations the following equation $$\label{V4}
\frac{\varphi_{n,t}}{\varphi_n}-\frac{\varphi_{n-1,t}}{\varphi_{n-1}}=\lambda-u_n+u_{n-1}-\lambda\frac{\varphi_{n-1}}{\varphi_n}-u_{n-1}\frac{\varphi_n}{\varphi_{n-1}}.$$ Now we assume that the time evolution of the eigenfunctions is given by a linear system of the form $\varphi_{n,t}=a_n\varphi_n+b_n\varphi_{n-1}, \varphi_{n-1,t}=c_n\varphi_n+d_n\varphi_{n-1}$ and substitute this expression into (\[V4\]). By comparing the coefficients before the powers of the independent variables $\varphi_{n},\, \varphi_{n-1}$ we find equations for the searched functions $a_n, b_n, c_n, d_n$: $$\label{V5}
b_n=-\lambda,\, c_n=u_{n-1},\, a_n-d_n=\lambda-u_n+u_{n-1}.$$ In addition we have $D(\varphi_{n-1,t})=\varphi_{n,t}$ which implies that $a_n=\lambda+d_{n+1}$. Summarizing the reasonings above we get: $$\begin{aligned}
\label{V6}
\varphi_{n,t}=(\lambda-u_n)\varphi_n-\lambda\varphi_{n-1},\\
\varphi_{n-1,t}=u_{n-1}\varphi_n-u_{n-1}\varphi_{n-1}.\end{aligned}$$ Change of the variables $\varphi_n=\lambda^{n/2}\prod^{\infty}_{j=n}u_j\psi_{n-1}$, reduces the equations (\[V3\]), (\[V6\]) to the usual Lax pair for the Volterra equation $$\begin{aligned}
\label{V7}
&&\psi_{n+1}=\xi\psi_n-u_n\psi_{n-1}, \quad \xi=\sqrt{\lambda},\\
&&\psi_{n,t}=(\xi^2+u_n)\psi_n-\xi u_n\psi_{n-1}.\end{aligned}$$
[5]{}
I. T. Habibullin, A. R. Khakimova, M. N. Poptsova, “On a method for constructing the Lax pairs for nonlinear integrable equations”, Journal of Physics A: Mathematical and Theoretical, 49:3, 035202, (35pp) (2016).
E. V. Pavlova, I. T. Habibullin, A. R. Khakimova, “On one integrable discrete system”, Differential equations. Mathematical physics, Itogi Nauki i Tekhniki. Ser. Sovrem. Mat. Pril. Temat. Obz., 140, VINITI, M., 30–42 (2017).
I. T. Habibullin, A. R. Khakimova, “Invariant manifolds and Lax pairs for integrable nonlinear chains”, Theoret. and Math. Phys., 191:3, 793–810 (2017).
I. T. Habibullin and A. R. Khakimova, “On a method for constructing the Lax pairs for integrable models via a quadratic ansatz”, Journal of Physics A: Mathematical and Theoretical, Volume 50, Number 30 (2017).
N. H. Ibragimov, A. B. Shabat, “Evolution equations admitting a nontrivial Lie-Backlund group”, Funct. Anal. Appl, 14(1), 25-36 (1980).
B. I. Suleimanov “The “quantum” linearization of the Painleve equations as the component of their L-A pairs”, Ufa Math. J., V. 4, N 2, pp. 127-136 (2012).
V. E. Zakharov and A. B. Shabat, “A scheme for integrating the nonlinear equations of mathematical physics by the method of the inverse scattering problem I”, Funct. Anal. Appl. [**8**]{}:3, 226–235 (1974).
V. E. Zakharov and A. B. Shabat, “Integration of nonlinear equations of mathematical physics by the method of inverse scattering II”, Funct. Anal. Appl. [**13**]{}:3, 166–174 (1979).
H. D. Wahlquist and F. B. Estabrook, “Prolongation structures of nonlinear evolution equations”, JMP [**16**]{}:1, 1–7 (1975).
J. Weiss, M. Tabor, G. Carnevale, “The Painlevé property for partial differential equations”, Journal of Mathematical Physics, 24(3), 522-526 (1983).
M. Musette, R. Conte, “Algorithmic method for deriving Lax pairs from the invariant Painlevé analysis of nonlinear partial differential equations”, Journal of mathematical physics, 32(6), 1450-1457 (1991).
F. W. Nijhoff and A. J. Walker, “The discrete and continuous Painlevé VI hierarchy and the Garnier system”, Glasgow Math. J. [**43A**]{}, 109-23 (2001).
A. I. Bobenko and Yu B Suris, “Integrable systems on quad-graphs”, Int. Math. Res. Notes [**11**]{}, 573–611 (2002).
F. W. Nijhoff, “Lax pair for the Adler (lattice Krichever-Novikov) system”, Physics Letters A [**297**]{}:1–2, 49-58 (2002).
R. I. Yamilov, “On the classification of discrete equations Integrable Systems ed A B Shabat”, Ufa: Soviet Academy of Sciences, 95–114 (1982)(in Russian).
P. Xenitidis “Integrability and symmetries of difference equations: the Adler-Bobenko-Suris case”, Proc. 4th Workshop “Group Analysis of Differential Equations and Integrable Systems”, 226–42, arXiv:0902.3954 (2009).
M. G$\ddot{u}$rses, A. Karasu, V. V. Sokolov, “On construction of recursion operators from Lax representation”, Journal of Mathematical Physics, 40(12), 6473-6490 (1999).
C. S. Gardner, J. M. Greene, M. D. Kruskal, R. M. Miura, “Korteweg-devries equation and generalizations. VI. methods for exact solution”, Communications on pure and applied mathematics, 27(1), 97-133 (1974).
I. M. Gel’fand and L. A. Dikii, “Fractional powers of operators and Hamiltonian systems”, Funct. Anal. Appl., 10:4, 13-29 (1976).
F. Magri, “A simple model of the integrable Hamiltonian equation”, J. Math. Phys. 19:5, 1156-1170 (1978).
V. V. Sokolov, “On the Hamiltonian property of the Krichever-Novikov equation”, Sov. Math. Dokl., 30:1, 44-46 (1984).
M. G$\ddot{u}$rses, A. Karasu, “Integrable KdV systems: recursion operators of degree four”, Phys. Lett. A, A 251:4, 247-249 (1999).
S. I. Svinolupov, “Jordan algebras and generalized Korteweg–de Vries equations”, Theor. Math. Phys. 87:3, 391-403 (1991).
J. Krasil’shchik, “Cohomology background in geometry of PDE”, Contemp. Math., 219, 121-140, (1998).
H. Zhang, G. Z. Tu, W. Oevel, B. Fuchssteiner, “Symmetries, conserved quantities, and hierarchies for some lattice systems with soliton structure”, Journal of mathematical physics, 32(7), 1908-1918 (1991).
W. Oevel, H. Zhang, B. Fuchssteiner, “Mastersymmetries and multi-Hamiltonian formulations for some integrable lattice systems”, Progress of theoretical physics, 81(2), 294-308 (1989).
P. J. Olver, “Evolution equations possessing infinitely many symmetries”, Journal of Mathematical Physics, 18(6), 1212-1215 (1977).
R. I. Yamilov, “Symmetries as integrability criteria for differential difference equations”, J. Phys. A: Math. Gen. 39, R541–R623 (2006).
D. Levi, R. Yamilov, “Conditions for the existence of higher symmetries of evolutionary equations on the lattice”, Journal of Mathematical Physics, 38(12), 6648-6674 (1997).
F. Khanizadeh, A. V. Mikhailov, J. P. Wang, “Darboux transformations and recursion operators for differential-difference equations”, Theoretical and Mathematical Physics, 177(3): 1606–1654 (2013).
R. N. Garifullin, E. V. Gudkova, I. T. Habibullin, “Method for searching higher symmetries for quad-graph equations”, Journal of Physics A: Mathematical and Theoretical, 44(32), 325202 (2011).
P. E. Hydon and C. M. Viallet, “Asymmetric integrable quad-graph equations”, Appl. Anal. 89 493–506 2010.
R. Garifullin, I. Habibullin, M. Yangubaeva, “Affine and finite Lie algebras and integrable Toda field equations on discrete space-time”, SIGMA, Volume 8, 062, pp.33 (2012).
B. A. E. Dubrovin, V. B. Matveev, S. P. Novikov, “Non-linear equations of Korteweg-de Vries type, finite-zone linear operators, and Abelian varieties”, Russian mathematical surveys, 31(1), 59-146 (1976).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The phaseless auxiliary-field quantum Monte Carlo (AF QMC) method \[S. Zhang and H. Krakauer, [*Phys. Rev. Lett.*]{} [**90**]{}, 136401 (2003)\] is used to carry out a systematic study of the dissociation and ionization energies of second-row group 3A-7A atoms and dimers, Al, Si, P, S, Cl. In addition, the P$_2$ dimer is compared to the third-row As$_2$ dimer, which is also triply-bonded. This method projects the many-body ground state by means of importance-sampled random walks in the space of Slater determinants. The Monte Carlo phase problem, due to the electron-electron Coulomb interaction, is controlled via the phaseless approximation, with a trial wave function $|\Psi_T\rangle$. As in previous calculations, a mean-field single Slater determinant is used as $|\Psi_T\rangle$. The method is formulated in the Hilbert space defined by any chosen one-particle basis. The present calculations use a planewave basis under periodic boundary conditions with norm-conserving pseudopotentials. Computational details of the planewave AF QMC method are presented. The isolated systems chosen here allow a systematic study of the various algorithmic issues. We show the accuracy of the planewave method and discuss its convergence with respect to parameters such as the supercell size and planewave cutoff. The use of standard norm-conserving pseudopotentials in the many-body AF QMC framework is examined.'
author:
- Malliga Suewattana
- Wirawan Purwanto
- Shiwei Zhang
- Henry Krakauer
- 'Eric J. Walter'
bibliography:
- 'bibdb.bib'
title: 'Phaseless auxiliary-field quantum Monte Carlo calculations with planewaves and pseudopotentials—applications to atoms and molecules'
---
Introduction
============
Achieving accurate solutions of the electronic many-body Schrödinger equation is a challenging problem for calculations of the properties of real materials. For many systems, density functional theory (DFT), in a variety of approximations, has been applied with great success. In DFT, the many-body interactions are replaced by a single particle interacting with the mean-field generated by the other particles, similar in spirit to the Hartree-Fock (HF) method. Unfortunately, these methods have well-known limitations and often fail at describing the properties of materials with large electron-electron correlation.
A more accurate approach is the quantum Monte Carlo (QMC) method [@Ceperley1980; @Reynolds1982; @Foulkes2001; @sz-hk], which has been shown to be among the most effective methods for many-electron problems. Unlike other correlated methods, QMC calculation times scale as a low power of the system size [@WilliamsonEtAl]. The fixed-node diffusion Monte Carlo (DMC) approach, which samples the many-body wave function in real space, has been the most widely used QMC method in electronic structure calculations [@Reynolds1982; @Foulkes2001].
The recently developed phaseless auxiliary-field quantum Monte Carlo (AF QMC) method [@sz-hk] is an alternative and complementary QMC approach, which samples the many-body wave function in the space of Slater determinants. This method has several attractive features. The fermionic antisymmetry of the wave function is automatically accounted for, since it is sampled by Slater determinants. This provides a different route to controlling the sign problem [@Ceperley_sign; @kalos91; @Zhang1999_Nato] from fixed-node DMC, which has shown promise in reducing the dependence of the systematic errors on the trial wave function [@Al-Saidi_TMO; @Al-Saidi; @Al-Saidi-06JCP2]. The orbitals in the Slater determinants are expressed in terms of a chosen single-particle basis (*e.g.*, planewaves, Gaussians, etc.), so AF QMC shares much of the same computational machinery with DFT and other independent-particle type methods. AF QMC can thus straightforwardly incorporate many of the methodological advances from mean-field methods (such as pseudopotential and fast Fourier transforms) while systematically improving on mean-field accuracy.
Using a planewave basis, tests of the phaseless AF QMC method for a few simple atoms and molecules [@sz-hk; @AFQMC-CPC2005] as well as for the more correlated TiO and MnO molecules [@Al-Saidi_TMO] yielded excellent results. More systematic applications of the phaseless AF QMC method to atoms and molecules have been carried out using Gaussian basis sets. These include all-electron calculations for first-row systems [@Al-Saidi] and effective-core potential calculations in post-$d$ group elements [@Al-Saidi-06JCP2]. The results also showed excellent agreement with near-exact quantum chemistry results and/or experiment.
The planewave AF QMC method is well-adaped for correlated calculations of extended bulk systems, where planewave based methods have been the standard choice in traditional electronic structure calculations. It is therefore important to systematically study its algorithmic issues and to characterize its performance. In this paper, we use the planewave phaseless AF QMC method to carry out a systematic study of the dissociation and ionization energies of second-row atoms and dimers in Group 3A-7A, namely Al, Si, P, S, Cl. The interesting case of the triply-bonded P$_2$ dimer is also compared to the third-row As$_2$ dimer. The principal goal of this study is to further benchmark the AF QMC method across more systems and across different basis sets and to compare the results with those from other methods and experiment. While the use of localized basis sets, such as Gaussians, is generally more efficient for isolated atoms and molecules, it is straighforward to apply planewave methods using periodic boundary conditions and large supercells. Planewave methods have several desirable features. A planewave basis provides an unbiased representation of the wave functions, since convergence to the infinite basis limit is controlled by a single parameter, the kinetic-energy cutoff ${E_{\textrm{cut}}}$. Planewaves are algorithmically simple to implement, and operations with planewaves can be made very efficient, using fast Fourier techniques as in DFT methods. To keep the planewave basis size tractable, pseudopotentials must be used to eliminate the highly localized core electron states and to produce relatively smooth valence wave functions. The present choice of isolated atomic and molecular systems permits direct comparisons with Gaussian-based AF QMC and with well-establised quantum chemistry all-electron and pseudopotential calculations.
The remainder of this paper is organized as follows. In Section \[sec:compt\], we describe computational details of the phaseless AF QMC method in the planewave-pseudopotential framework. In Section \[sec:convergence\], we discuss calculation parameters such as supercell (simulation cell) size, cutoff energy, and AF QMC time step size, and the algorithm’s convergence behavior with respect to these parameters. Section \[sec:results\] presents the calculated dissociation and ionization energies and comparisons with other theoretical results and with experiment. In Section \[sec:discussion\], we discuss systematic errors due to the use of the phaseless approximation and use of norm-conserving pseudopotentials. We then conclude with some general remarks in Section \[sec:summary\].
Planewave AF QMC method: Computational Details {#sec:compt}
==============================================
Hamiltonian {#sec:Hamiltonian}
-----------
It is convenient to express, within the Born Oppenheimer approximation, the electronic Hamiltonian in second quantized form in terms of a chosen orthonormal one-particle basis $$\label{eq:H}
{{\hat{H}}}= \sum_{ij}^M {H^{(1)}_{ij} {c_{{i}}^\dagger}{c_{{j}}^{}}}
+ {\frac{1}{2}}\sum_{ijkl}^M
{H^{(2)}_{ijkl} {c_{{i}}^\dagger}{c_{{j}}^\dagger}{c_{{l}}^{}}{c_{{k}}^{}}}
+ {V_{\textrm{II}}}\,,$$ where $M$ is the number of basis functions, $c_i^\dagger$ and $c_i$ are the corresponding creation and annihilation operators, and the electron spins have been subsumed in the summations. $H^{(1)}_{ij}$ and $H^{(2)}_{ijkl}$ are the one- and two-body matrix elements, and ${V_{\textrm{II}}}$ is the classical Coulomb interaction of the point ions [@Yin_1982]. Atomic units are used throughout this paper. We use periodic boundary conditions and a planewave basis $${\langle{{{\mathbf{r}}}}|{{{\mathbf{G}}}}\rangle}
\equiv
{\langle{{{{\mathbf{r}}}}}|{{\,{c_{{{\mathbf{G}}}}^\dagger}}}|{{0}}\rangle}
= \frac{1}{{\sqrt \Omega}}\exp (i{{\mathbf{G}}} \cdot {{\mathbf{r}}})
\, ,$$ where $\Omega$ is the volume of the simulation cell and ${{\mathbf{G}}}$ is a reciprocal lattice vector. As in planewave-based density functional calculations, the number of planewaves $M$ in the basis is determined $G^2/2 \le {E_{\textrm{cut}}}$, where ${E_{\textrm{cut}}}$ is the cutoff kinetic energy.
The one-body operators in the Hamiltonian include the kinetic energy, $$\label{KE}
{{\hat{K}}}= {\frac{1}{2}}\sum_{{\mathbf{G}}} {G^2}\, {c_{{{\mathbf{G}}}}^\dagger} {c_{{{\mathbf{G}}}}^{}}
\,,$$ and nonlocal pseudopotential, which describes the electron-ion interaction [ $$\label{V_ei}
\begin{split}
{\hat{V}_{\textrm{ei}}}& = \sum_{{\mathbf{G}},{\mathbf{G}}'} {V_{\textrm{L}}}({{\mathbf{G}}} - {{\mathbf{G}}'})\, {c_{{{\mathbf{G}}}}^\dagger}{c_{{{\mathbf{G}}'}}^{}}
\\
&\,
+ \sum_{{\mathbf{G}},{\mathbf{G}}'} {V_{\textrm{NL}}}({{\mathbf{G}}}, {{\mathbf{G}}}')\, {c_{{{\mathbf{G}}}}^\dagger}{c_{{{\mathbf{G}}'}}^{}}
\,,
\end{split}$$ ]{} where ${V_{\textrm{L}}}({{\mathbf{G}}} - {{\mathbf{G}}'})$ and ${V_{\textrm{NL}}}({{\mathbf{G}}},{{\mathbf{G}}'})$ are the matrix elements of local and nonlocal parts of the pseudopotential, respectively. It is convenient to rewrite the local part of the pseudopotential and to define the the following quantities
$$\begin{aligned}
\label{V_ei_rho}
{\hat{V}_{\textrm{ei}}}& = {\hat{V}_{\textrm{ei,L}}}+ {\hat{V}_{\textrm{ei,NL}}}+ N{V_{\textrm{L}}}(\mathbf{0})
\,,
\\
{\hat{V}_{\textrm{ei,L}}}& = {\frac{1}{2}}\sum_{{\mathbf{Q}}\neq \mathbf{0}} {V_{\textrm{L}}}({\mathbf{Q}})
\left[
\hat{\rho}({{\mathbf{Q}}}) +
\hat{\rho}^\dag({{\mathbf{Q}}})
\right]
\,,
\\
{\hat{V}_{\textrm{ei,NL}}}& = \sum_{{\mathbf{G}},{\mathbf{G}}'}
{V_{\textrm{NL}}}({{\mathbf{G}}}, {{\mathbf{G}}'}) {c_{{{\mathbf{G}}}}^\dagger} {c_{{{\mathbf{G}}'}}^{}}
\,,\end{aligned}$$
where $N$ is the number of electrons, and the one-body density operator $\hat{\rho}({{\mathbf{Q}}})$ in this equation is given by $$\label{rho}
\hat{\rho}({{\mathbf{Q}}})
\equiv
\sum_{{\mathbf{G}},\lambda}
{c_{{{{\mathbf{G}}+ {\mathbf{Q}}},\lambda}}^\dagger} {c_{{{{\mathbf{G}}},\lambda}}^{}}
\,
\theta
\left(
{E_{\textrm{cut}}}- \left|{{\mathbf{G}}+{\mathbf{Q}}}\right|^2/2
\right)
,$$ where the step function ensures that $({{\mathbf{G}}+ {\mathbf{Q}}})$ lies within the planewave basis, and the summation over electron spins ($\lambda=1,2$) has been made explicit.
The electron-electron interaction is given by [ $$\label{Vee}
\begin{split}
{\hat{V}_{\textrm{ee}}}& = {\frac{1}{2}}N \xi
\\
& + \,
\frac{1}{{2\Omega}} {\sum_{ijkl}}'
\frac{4\pi}{\left| {\mathbf{G}}_i - {\mathbf{G}}_k \right|^2} \,
\delta_{{\mathbf{G}}_i - {\mathbf{G}}_k , {\mathbf{G}}_l - {\mathbf{G}}_j}^{}
\delta_{\lambda_i, \lambda_k}^{}
\delta_{\lambda_j, \lambda_l}^{}
\\
& \quad \, \times
{c_{{i}}^\dagger} {c_{{j}}^\dagger} {c_{{l}}^{}}{c_{{k}}^{}}
\;.
\end{split}$$ ]{} The primed summation indicates that the ${\mathbf{G}}_i = {\mathbf{G}}_k$ singular term is excluded, due to charge neutrality. The first term in this equation is a constant due to the self-interaction of an electron with its periodic images. It depends only on the number of electrons in the simulation cell and the Bravais lattice associated with the periodic boundary conditions. The standard Ewald expression for $\xi$ is given by [@Fraser96] [ $$\label{xi}
\begin{split}
\xi
& = \frac{1}{\Omega}
\sum_{{\mathbf{G}}\neq \mathbf{0}}
\frac{\exp(-\pi^2 G^2 / \kappa^2)}{\pi G^2}
- \frac{\pi}{\kappa^2 \Omega}
\\
& + \,
\sum_{{\mathbf{R}}\neq \mathbf{0}}
\frac{\operatorname{erfc}(\kappa R)}{R}
- \frac{2\kappa}{\sqrt{\pi}}
\,,
\end{split}$$ ]{} where ${\mathbf{R}}$ is a direct lattice vector, and $\xi$ is independent of the Ewald constant $\kappa$, which only controls the relative convergence rates of the direct and reciprocal space summations. For the discussion below, we rewrite the two-body contribution in Eq. (\[Vee\]): [ $$\label{V_rho-rho}
\begin{split}
{\hat{V}_{\textrm{ee}}}& = {\frac{1}{2}}N \xi
+ \frac{1}{2\Omega} \sum_{{\mathbf{Q}}\neq \mathbf{0}}
\frac{4\pi}{Q^2}\,\hat{\rho}^\dagger({\mathbf{Q}}) \hat{\rho}({\mathbf{Q}})
\\
& - \,
\frac{1}{2\Omega} \sum_\lambda \sum_{{\mathbf{G}},{\mathbf{G}}'}
\frac{4\pi}{\left|{\mathbf{G}}- {\mathbf{G}}'\right|^2}
{c_{{{\mathbf{G}},\lambda}}^\dagger} {c_{{{\mathbf{G}},\lambda}}^{}}
\,.
\end{split}$$ ]{} The third term in Eq. (\[V\_rho-rho\]) is a sum of diagonal one-body operators arising from the anticommutation of the fermion creation and destruction operators.
Finally, we can regroup the contributions to the Hamiltonian into constant, one-body, and two-body parts, $${{\hat{H}}}= H^{(0)} + {{\hat{H}}}^{(1)} + {{\hat{H}}}^{(2)},$$ where
$$\begin{aligned}
\label{eq:Hterms}
H^{(0)}
& = {\frac{1}{2}}N \xi
+ {V_{\textrm{II}}}+ N{V_{\textrm{L}}}(\mathbf{0})
\\
\begin{split}
{{\hat{H}}}^{(1)}
& = {{\hat{K}}}+ {\hat{V}_{\textrm{ei,L}}}+ {\hat{V}_{\textrm{ei,NL}}}\\
& - \,
\frac{1}{2\Omega} \sum_\lambda \sum_{{\mathbf{G}},{\mathbf{G}}'}
\frac{4\pi}{\left|{\mathbf{G}}- {\mathbf{G}}'\right|^2}
{c_{{{\mathbf{G}},\lambda}}^\dagger} {c_{{{\mathbf{G}},\lambda}}^{}}
\end{split}
\\
{{\hat{H}}}^{(2)}
& = \frac{1}{2\Omega} \sum_{{\mathbf{Q}}\neq \mathbf{0}}
\frac{4\pi}{Q^2}\,\hat{\rho}^\dagger({\mathbf{Q}}) \hat{\rho}({\mathbf{Q}})\end{aligned}$$
It is convenient to express the two-body part as a quadratic form of one-body operators (this can always be done, and such forms are not unique). We use the identity $\hat{\rho}(-{\mathbf{Q}}) = \hat{\rho}^\dagger({\mathbf{Q}})$ to write $$\hat{H}^{(2)}
= \sum_{{\mathbf{Q}}\neq \mathbf{0}}
\frac{\pi}{\Omega Q^2}
\left[
\hat{\rho}({\mathbf{Q}}) \hat{\rho}^\dagger({\mathbf{Q}})
+ \hat{\rho}^\dagger({\mathbf{Q}}) \hat{\rho}({\mathbf{Q}})
\right].
$$ Defining Hermitian operators ${{\hat{A}}}({{\mathbf{Q}}})$ and ${{\hat{B}}}({{\mathbf{Q}}})$ as
$$\begin{aligned}
\label{AB}
{{\hat{A}}}({\mathbf{Q}})
& \equiv
\sqrt{\frac{2\pi}{\Omega Q^2}}
\left[ \hat{\rho}({\mathbf{Q}}) + \hat{\rho}^\dagger({\mathbf{Q}}) \right]
\,,
\\
{{\hat{B}}}({\mathbf{Q}})
& \equiv
i \sqrt{\frac{2\pi}{\Omega Q^2}}
\left[ \hat{\rho}({\mathbf{Q}}) - \hat{\rho}^\dagger({\mathbf{Q}}) \right]
\,,\end{aligned}$$
the two-body contribution becomes a simple sum of quadratic operators, $$\begin{aligned}
\label{H2}
{{\hat{H}}}^{(2)}
& = \frac{1}{4} \sum_{{\mathbf{Q}}\neq \mathbf{0}}
\left[ {{\hat{A}}}^2({{\mathbf{Q}}}) + {{\hat{B}}}^2({{\mathbf{Q}}}) \right] \notag
\\
& = \frac{1}{2} \sum_{{\mathbf{Q}}> \mathbf{0}}
\left[ {{\hat{A}}}^2({{\mathbf{Q}}}) + {{\hat{B}}}^2({{\mathbf{Q}}}) \right]
\,,\end{aligned}$$ where we have used the ${{\mathbf{Q}}} \to -{{\mathbf{Q}}}$ symmetry to obtain the last expression.
Ground state projection and the Hubbard-Stratonovich Transformation {#sec:GS-projection}
-------------------------------------------------------------------
The ground state of ${{\hat{H}}}{|{{{\Psi_0}}}\rangle} = E_0 {|{{{\Psi_0}}}\rangle}$ is obtained by imaginary-time projection from a trial wave function ${|{{\Psi_T}}\rangle}$ [ $$\label{eq:QMC-proj}
\begin{split}
\lim_{n \to \infty}
{\left(e^{-\Delta \tau ({{\hat{H}}}- E_0)}\right)}^n {|{{\Psi_T}}\rangle}
=
{|{{{\Psi_0}}}\rangle}
\,,
\end{split}$$ ]{} provided ${\langle{{\Psi_T}}|{{{\Psi_0}}}\rangle} \neq 0$. In the present calculations, ${|{{\Psi_T}}\rangle}$ is a single Slater determinant obtained from a mean-field calculation. Expressing the imaginary-time projection in terms of the small discrete time step $\Delta \tau$ facilitates the separation of the one- and two-body terms, using the short-time Trotter-Suzuki decomposition [@Trotter1959; @Suzuki1976] [ $$\label{eq:Trotter}
\begin{split}
e^{-\Delta \tau {{\hat{H}}}}
& = e^{-(1/2) \Delta \tau {{\hat{H}}}^{(1)}}
e^{-\Delta \tau {{\hat{H}}}^{(2)}}
e^{-(1/2)\Delta \tau {{\hat{H}}}^{(1)}}
\\
& + \, O(\Delta \tau^3)
\,.
\end{split}$$ ]{} The application of the one-body propagator $e^{-(1/2)\Delta \tau {{\hat{H}}}^{(1)}}$ on a Slater determinant ${|{{\phi}}\rangle}$ simply yields another Slater determinant: $ {|{{\phi'}}\rangle} = e^{-(1/2) \Delta \tau {{\hat{H}}}^{(1)}} {|{{\phi}}\rangle}$. The two-body propagator is expressed as an integral of one-body propagators, using the Hubbard-Stratonovich transformation [@Hubbard; @Stratonovich] [ $$\label{eq:HS-xform0}
\begin{split}
& \exp{\left({-{\frac{1}{2}}\Delta\tau \sum_i\lambda_i {{\hat{b}}}_i^2}\right)}
\\
&\quad\;
=
\int \left( \prod_i \frac{d\sigma_i}{\sqrt{2\pi}} \right)
\exp{\left[\sum_i \left(-{\frac{1}{2}}\sigma_i^2
+\sigma_i \sqrt{-\Delta\tau\lambda_i} \, {{\hat{b}}}_i \right)
\right]}
\end{split}$$ ]{} for any one-body operators $\{\,{{\hat{b}}}_i\,\}$. Thus we have [ $$\label{eq:HS-xform}
\begin{split}
e^{-\Delta\tau{{\hat{H}}}^{(2)}}
= \left(\frac{1}{\sqrt{2\pi}}\right)^{\!\dim(\bm{\sigma})}
\!
\int d\bm{\sigma} \,
e^{-(1/2) \bm{\sigma} \cdot \bm{\sigma}}
e^{\sqrt{\Delta\tau}\, \bm{\sigma} \cdot \hat{\mathbf{v}} }
\, ,
\end{split}$$ ]{} where we introduce a vector $\bm{\sigma} \equiv \{\sigma_i\}$, whose dimensionality, $\dim(\bm{\sigma})$, is the number of *all* possible ${\mathbf{Q}}$-vectors satisfying ${\mathbf{Q}}= {\mathbf{G}}- {\mathbf{G}}'$ for two arbitrary wave vectors ${\mathbf{G}}$ and ${\mathbf{G}}'$ in the planewave basis. The operators $\hat{\mathbf{v}} \equiv \{ \sqrt{-\lambda_i}\,{{\hat{b}}}_i \}$ are given by the $i{{\hat{A}}}({{\mathbf{Q}}})$ or $i{{\hat{B}}}({{\mathbf{Q}}})$ one-body operators, since all the $\lambda_i = 1$.
In the original formulations of the AF QMC method [@Blankenbecler1981; @Koonin1986], the many-dimensional integral over the auxiliary fields $\bm{\sigma}$ in Eq. (\[eq:HS-xform\]) is evaluated by standard Metropolis or heat-bath algorithms. We instead apply an importance-sampling transformation [@sz-cpmc; @sz-hk; @ZhangB] to turn the projection into a branching random walk in an overcomplete Slater determinant space. The importance sampling helps guide the random walks according to the projected overlap with the trial wave function. More importantly, it allows the imposition of a constraint to control the phase problem.
A phase problem arises for a general repulsive two-body interaction, because the $\lambda_i$ cannot be made all negative. In other words, not all components of the operator $\hat{\mathbf{v}}$ can be made real. (Although this is in principle possible by an overall shift to the potential [@Koonin1986] or by introducing many more auxiliary fields, they both cause large fluctuations [@sz-hk; @Silvestrelli93].) As the random proceeds, the projection in Eq. (\[eq:HS-xform\]) $$\label{randwlk-stp}
{|{{\phi'}}\rangle}
\leftarrow
\exp(\sqrt{\Delta\tau}\, \bm{\sigma} \cdot \hat{\mathbf{v}}) {|{{\phi}}\rangle}$$ by a complex $\hat{\mathbf{v}}$ causes the orbitals in the Slater determinants $\left|\phi\right\rangle$ to become complex. For large imaginary projection times, the phase of each $\left|\phi\right\rangle$ becomes random, and the stochastic representation of the ground state ${|{{{\Psi_0}}}\rangle}$ becomes dominated by noise. This leads to the phase problem and the divergence of the fluctuations. The phase problem is of the same origin as the sign problem that occurs when the one-body operators $\hat{\mathbf{v}}$ are real, but is more severe because, instead of a $+\left|\phi\right\rangle$ and $-\left|\phi\right\rangle$ symmetry [@kalos91; @sz-cpmc], there is now an infinite set $\{ e^{i\theta} \left|\phi\right\rangle, \theta \in
[0,2\pi) \}$, among which the Monte Carlo sampling cannot distinguish.
The phaseless AF QMC method [@sz-hk] used in this paper controls the phase/sign problem in an approximate manner using a trial wave function. The method uses a [*complex*]{} importance function, the overlap $\langle \Psi_T|\phi\rangle$, to construct phaseless random walkers, $|\phi\rangle/\langle \Psi_T|\phi\rangle$, which are invariant under a phase gauge transformation. The resulting two-dimensional diffusion process in the complex plane of the overlap $\langle \Psi_T|\phi\rangle$ is then approximated as a diffusion process in one dimension. Additional implementation details can be found in Refs. . The phaseless constraint is different from the nodal condition imposed in fixed-node DMC, since the phaseless constraint confines the random walk in Slater determinant space according to its overlap with a trial wave function, which is a global property of $\left|\phi\right\rangle$. Thus, the phaseless approximation can behave differently from the fixed node approximation in DMC.
Finally, we describe the use of fast Fourier transform (FFT) with a planewave basis to efficiently implement the random walk projection given by Eqs. (\[AB\]), (\[H2\]), (\[eq:HS-xform\]), and (\[randwlk-stp\]). For example, [ $$\label{eq:proj-FFT}
\begin{split}
&\exp \left(
\sum_{{\mathbf{Q}}} \sqrt{\frac{\Delta\tau}{Q^2}} \,
\sigma({\mathbf{Q}}) \hat{\rho}({\mathbf{Q}})
\right) {|{{\phi}}\rangle} \\
&\qquad \simeq
\sum_{n = 0}^{n_{\max } } \frac{1}{n!}
{\left ( \sqrt { \frac{\Delta \tau} {Q^2} } \,
\sigma({\mathbf{Q}})\hat{\rho}({\mathbf{Q}}) \right)\!}^n {|{{\phi}}\rangle}
\, .
\end{split}$$ ]{} Terms in the series can be evaluated as an iterative FFT, since $\rho (\bf{Q}) {|{{\phi}}\rangle}$ is just a convolution. For typical values of $\Delta \tau$, we find that $n_{\max} \simeq 4$ accurately reproduces the propagator.
Ground-state mixed estimator {#sec:mixed-est}
----------------------------
The ground state energy $E_0$ can then be obtained by the mixed estimator $$\label{E_estimat}
E_0 = \frac{{\langle{{\Psi_T}}|{{{{\hat{H}}}}}|{{{\Psi_0}}}\rangle}}
{{\langle{{\Psi_T}}|{{{\Psi_0}}}\rangle}}
= \lim_{\beta \to \infty}
\frac{{\langle{{\Psi_T}}|{{{{\hat{H}}}e^{-\beta{{\hat{H}}}}}}|{{\Psi_T}}\rangle}}
{{\langle{{\Psi_T}}|{{e^{-\beta{{\hat{H}}}}}}|{{\Psi_T}}\rangle}}
\,,$$ which is evaluated periodically from the ensemble of Slater determinants generated in the course of the random walks. In the phaseless AF QMC method, an importance sampling transformation [@sz-hk] leads to a stochastic representation of the ground-state wave function in the form of $${|{{{\Psi_0}}}\rangle}
= \sum_\phi w_\phi
\frac{{|{{\phi}}\rangle}}{{\langle{{\Psi_T}}|{{\phi}}\rangle}}
\,.
\label{eq:wf_MC_imp}$$ This means the mixed estimate for the energy is given by $$E_0^{\textrm{MC}}
= \frac{\sum_\phi w_\phi E_L[\phi]}{\sum_\phi w_\phi}
\,,
\label{eq:mixed_w_EL}$$ where the local energy is defined as $$E_L[\phi]
\equiv
\frac{{\langle{{\Psi_T}}|{{{{\hat{H}}}}}|{{\phi}}\rangle}}
{{\langle{{\Psi_T}}|{{\phi}}\rangle}}
\,.
\label{eq:El}$$
Matrix elements of one-body terms in the local energy (and other similar estimators) can be expressed in terms of the one-body Green’s functions[@sz-cpmc; @ZhangB] $$\label{G1}
G_{ji}
= \langle {c_j^\dag c_i } \rangle
\equiv
\frac{{\langle{{\Psi_T}}|{{{c_{{j}}^\dagger} {c_{{i}}^{}}}}|{{\phi}}\rangle}}
{{\langle{{\Psi_T}}|{{\phi}}\rangle}}
\,.$$ The Green’s function can be expressed in terms of the one-particle orbitals in the ${|{{\Psi_T}}\rangle}$ and ${|{{\phi}}\rangle}$ Slater determinants as follows. A general Slater determinant ${|{{\phi}}\rangle}$ can be written as $$\label{phi}
| \phi \rangle \equiv \phi_1^\dag \phi_2^\dag \cdot \cdot
\cdot \phi_N^\dag | 0 \rangle$$ where the $\phi_i^\dag$ creates an electron in the orbital $i$ $$\label{phi2}
\phi_i^\dag \equiv \sum\limits_{j} c_j^\dag \Phi_{ji},$$ and $j$ labels the one-particle orthogonal basis functions, which are planewaves in the present case. The $\Phi_{ji}$ are the elements of a $M \times N$ dimensional matrix $\bm
\Phi$. Each column of the matrix $\bm{\Phi}$ represents a single-particle orbital expressed as a sum of planewaves. It is a well-known result that the overlap of two Slater determinants is given by the determinant of the overlap matrix of their one particle orbitals $$\label{overlap}
\left \langle \Psi_T | \phi \right \rangle =
{\rm det} \left ( {\bm \Psi}_T^\dag {\bm \Phi} \right ).$$ Finally, it can be shown that the Green’s function can be expressed as[@Loh1992] $$\label{G1_a}
G_{ji} =
\left [ {\bm \Phi}
\left ( {\bm \Psi}_T^\dag {\bm \Phi} \right )^{-1}
{\bm \Psi}_T^\dag
\right ]_{ij}$$ Hamiltonian matrix elements of two-body terms in the mixed estimator are expressed in terms of the two-body Green’s function, which can be written as products of one-body Green’s functions using the Fermion anticommutation properties, $$\begin{aligned}
\begin{split}
\label{G2G1G1}
\langle {{c_{{i}}^\dagger} {c_{{m}}^\dagger} {c_{{n}}^{}} {c_{{j}}^{}}} \rangle
& \equiv
\frac{{\langle{{\Psi_T}}|{{{c_{{i}}^\dagger} {c_{{m}}^\dagger} {c_{{n}}^{}} {c_{{j}}^{}}}}|{{\phi}}\rangle}}
{{\langle{{\Psi_T}}|{{\phi}}\rangle}}
\\
& = G_{ji} G_{nm} - G_{ni} G_{jm} \; .
\end{split}\end{aligned}$$ Rather than directly implementing Eq. (\[G2G1G1\]), it is more efficient to use fast Fourier transformations to take advantage of locality in real space. The computer time to calculate the mixed estimator then scales as $N^2 M \log(M)$, where N is the number of electrons and M is the number of planewaves.
Trial wave function {#sec:psiT}
-------------------
The trial wave function ${|{{\Psi_T}}\rangle}$ determines the systematic accuracy of our calculations, due to the use of the phaseless approximation. Its quality also affects the statistical precision. We use a single Slater determinant as the trial wave function, which is obtained either from a DFT calculation or HF calculation. The DFT wave functions were generated self-consistently with [@abinit], using a planewave basis and the local density approximation (LDA). The HF wave functions were obtained from an in-house planewave-based Hartree-Fock program. In both cases, identical setup is used in the independent-electron calculation as in the corresponding AF QMC calculations.
Pseudopotentials {#sec:pseudos}
----------------
Norm-conserving pseudopotentials [@KB] are used in the present calculations. Pseudopotentials are necessary to keep the basis size tractable by eliminating the highly-localized core states. Pseudopotential transferability is a source of potential errors, however, especially since the pesudopotentials used here are generated from independent-electron calculations. Such pseudopotentials are quite routinely employed in QMC and other many-body calculations and have proved very useful. But their transferability is not nearly as extensively quantified and studied as in standard independent-electron calculations. Thus one of our goals here is to examine the use of such pseudopotentials in the many-body AF QMC framework.
The pseudopotential has been adapted to take the Kleinman-Bylander (KB) [@KB] form suitable for planewave calculations,
$$\begin{aligned}
\label{eq:PSP-KB}
{\hat{V}_{\textrm{ei}}}^{\textrm{(KB)}}(\mathbf{r})
& = {\hat{V}_{\textrm{ei,L}}}^{\textrm{(KB)}}(\mathbf{r})
+ {\hat{V}_{\textrm{ei,NL}}}^{\textrm{(KB)}}(\mathbf{r})
\,,
\\
{\hat{V}_{\textrm{ei,NL}}}^{\textrm{(KB)}}(\mathbf{r})
& = \sum_{l,m}
\frac{{|{{V_l \varphi_l Y_{lm}}}\rangle} {\langle{{V_l \varphi_l Y_{lm}}}|}}
{{\langle{{\varphi_l}}|{{V_l}}|{{\varphi_l}}\rangle}}
\,,\end{aligned}$$
where $\varphi_l$ is the pseudoorbital for the $l$-th angular momentum component. In our case, we use the neutral atomic reference state (with an LDA-type Hamiltonian) to generate the pseudoorbitals. To examine the effects of pseudopotentials on the accuracy of the AF QMC calculations, we employ two pseudopotentials in this study: the optimized LDA-based pseudopotential [@Rappe] generated using the OPIUM [@opium] package, and the HF-based effective core potential developed by Ovcharenko, Aspuru-Guzik, and Lester [@LesterECP2; @LesterECP1]. We will subsequently refer to these pseudopotentials as OPIUM and [OAL]{}, respectively. The semilocal [OAL]{} pseudopotentials were converted to the fully nonlocal KB form, using the atomic LDA ground state wave function as the reference state. The [[OAL]{} pseudopotential]{} is not used in molecules, because it lacks a $d$ projector. An illustration of this point is given in Table \[tab:PSP-quality\].
Table \[Ecut+Psp\] gives parameters describing the OPIUM pseudopotentials. The second column shows the cutoff energy for each atomic species. The same cutoff energies are also used in our calculations with [[OAL]{} pseudopotentials]{}. (The parameters of [[OAL]{} pseudopotentials]{} have been published in Ref. .) These were tested for convergence with LDA and then verified with AF QMC calculations.
--------- --------------------------- --------- ----------- --------- ---------------------------------
Reference
Species ${E_{\textrm{cut}}}$ (Ha) $l = 0$ $l = 1$ $l = 2$ configuration
Al $7.50$
Si $6.13$ $2.20$ $2.20$ $2.50$ \[Ne\] $3s^2
3p^2$
P $18.00$ $1.75$ $1.75$ $2.50$ \[Ne\] $3s^2 3p^{2.5} 3d^{0.5}$
S $19.00$ $1.75$ $1.75$ $1.75$ \[Ne\] $3s^2 3p^{3.5} 3d^{0.5}$
Cl $18.00$ $1.75$ $1.75$ $2.50$ \[Ne\] $3s^2 3p^{4.5} 3d^{0.5}$
As $18.00$ $1.80$ $1.80$ $2.50$ \[Ar\] $4s^2 4p^{2.5} 4d^0$
--------- --------------------------- --------- ----------- --------- ---------------------------------
: \[Ecut+Psp\] Optimized ${E_{\textrm{cut}}}$ and OPIUM pseudopotential parameters used in the calculations. Each angular component ($l$) of the pseudopotential has its own cutoff radius ($r_c$).
Convergence studies {#sec:convergence}
===================
To achieve high accuracy and to minimize the computational cost, one should optimize the calculations with respect to the number of basis functions, the supercell size, and the magnitude of the Trotter time step. In this section, we illustrate the convergence of our method with respect to these parameters. Ionization energies are defined as ${\mathrm{IP}}\equiv E(N-1) - E(N)$ and ${\mathrm{IIP}}\equiv E(N-2) - E(N)$, for the singly- and doubly-ionized atoms, respectively, where $N$ is the number of electrons in the neutral atom. The dissociation energy $D_e$ is calculated as the difference between the total energy of the dimer at the experimental equilibrium bond length and the energy of the isolated atoms, ${D_e}\equiv 2E_{\textrm{atom}} - E_{\textrm{dimer}}$.
Planewave convergence {#sec:PWconverg}
---------------------
Convergence with respect to the planewave cutoff energy ${E_{\textrm{cut}}}$ depends on both ${{\hat{H}}}^{(1)}$ and ${{\hat{H}}}^{(2)}$ in Eq. (\[eq:Hterms\]). The ${{\hat{H}}}^{(1)}$ dependence is similar to that in independent-electron calculations. Convergence requires that ${E_{\textrm{cut}}}$ is sufficient for the “hardness” of the pseudopotential and the electronic density variations. The ${{\hat{H}}}^{(2)}$ dependence has to do with the scattering matrix elements in the two-body interaction. In the uniform electron gas, for example, ${{\hat{H}}}^{(1)}$ requires an ${E_{\textrm{cut}}}$ given by the Fermi energy $E_{\rm F}$ (for restricted HF), while ${{\hat{H}}}^{(2)}$ will lead to a finite convergence error for any finite ${E_{\textrm{cut}}}$, which decreases as ${E_{\textrm{cut}}}$ is increased and, for a fixed ${E_{\textrm{cut}}}/E_{\rm F}$, becomes more pronounced as the electronic density is decreased.
Fig. \[ecut\] shows the phosphorus atom total energy as a function of the planewave cutoff energy ${E_{\textrm{cut}}}$ for both AF QMC and LDA. The calculations were done for a fixed supercell size and a pseudopotential whose design cutoff energy is 18 Ha. In LDA the total energy was converged to within 5meV at this cutoff. The energy decreases monotonically with increasing ${E_{\textrm{cut}}}$ in both calculations. We see that the AF QMC convergence behavior is similar to LDA, indicating that the AF QMC convergence error from ${{\hat{H}}}^{(2)}$ is much smaller here than that from ${{\hat{H}}}^{(1)}$. This trend was found to be typical of the systems studied in this paper with the chosen pseudopotentials. Table \[Ecut+Psp\] shows the cutoff energy for each atomic species. In subsequent calculations for phosphorus, for example, we used ${E_{\textrm{cut}}}= 18$Ha in the AF QMC, as indicated by the vertical dashed line in Fig. \[ecut\].
![\[ecut\] Convergence of the total energy of a phosphorus atom in a $14 \times 14 \times 14 {\,a_0}^3$ unit cell. The OPIUM pseudopotential is used here. The convergence behavior is similar in both LDA and AF QMC methods. A constant shift is added to each data set for convenience, such that the converged energies are approximately $0$ eV and $0.5$ eV for LDA and AF QMC, respectively. ](\FIGDIR{}P_ecutConv.eps)
Supercell size convergence {#sec:Supercellconverg}
--------------------------
Due to the periodic boundary conditions imposed in the calculations, the interactions between electrons in the simulation cell and their periodic images give rise to finite-size errors. To study the behavior of these errors, a series of LDA and AF QMC calculations were performed using different system sizes for cubic (and some tetragonal) shaped cells. Results for the phosphorous atom are shown in Fig. \[PE\] as a function of supercell size. The AF QMC energies are seen to converge from below while the LDA energy converge from above. This is not surprising, since LDA treats the supercell Coulomb interaction differently from a many-body approach such as AF QMC. Fig. \[fig:S-Etotal-1/V\] shows that the total energy from AF QMC for the sulfur atom is nearly a linear function of $1/\Omega$ for this range of supercell sizes. At the largest $19 \times 19 \times 19 {\,a_0}^3$ supercell size, the total energy is converged to within about $\sim 0.1$ eV.
![\[PE\] Convergence of the phosphorus total energy with respect to the inverse of the simulation cell volume $\Omega$. The triangles denote the results of LDA calculations while the solid circles denote those of the AF QMC. The OPIUM pseudopotential is used here. ](\FIGDIR{}P_E_vsV_LDA_QMC.eps)
![\[fig:S-Etotal-1/V\] Convergence of the AF QMC total energy with simulation cell size for Sulfur, plotted as a function of the inverse volume, from box sizes $12 \times 12 \times 12 {\,a_0}^3$ through $19 \times 19 \times 19 {\,a_0}^3$. The [[OAL]{} pseudopotential]{} is used here. The dotted line shows the fitting of the total energy as a linear function of $1 / \Omega$. The data point shown at $1 / \Omega = 0$ shows the extrapolated energy at infinite box size. ](\FIGDIR{}S_atom_ECP_fzfx1.eps)
To demonstrate the supercell size effect on the energy differences, Fig. \[DE\_E\_2E\] shows the calculated dissociation energy of P$_2$. The top panel shows the supercell size dependence of the dissociation energy, and the bottom panel illustrates the convergence error of P and P$_2$ energies for supercells ranging from $14 \times 14 \times 14 {\,a_0}^3$ through $18 \times 18 \times 18 {\,a_0}^3$. The total atom energy from different supercells deviates no more than $0.2$ eV from that of the largest supercell. The dimer total energy, on the other hand, shows stronger finite-size effect. Most of the finite-size error in the dissociation energy thus arises from the dimer energy. To avoid any irregular dependence of the energy on the aspect ratio (cubic vs. tetragonal supercells), only the cubic supercells were used in the extrapolation.
![\[DE\_E\_2E\] Phosphorus dissociation energy (top panel) and the total energies of P (times two) and P$_2$ (bottom panel) for different supercell sizes, computed using the OPIUM pseudopotential. The dissociation energy obtained with the largest simulation cell, $18 \times 18 \times 18 {\,a_0}^3$ is within 0.1 eV of the experimental value of 5.08 eV (indicated by the horizontal dotted line in the top panel). ${E_{\textrm{conv}}}$ is the energy extrapolated to $1/\Omega \to 0$ (using cubic cells only). ](\FIGDIR{}P_DE_E_2E_eV.eps)
In the ionization energy calculations, the supercells are charged $+|e|$ and $+|2e|$ for the $X^+$ and $X^{++}$ species, respectively. Charged supercells are ill-defined under periodic boundary conditions, so an additional neutralizing background charge is introduced to maintain charge neutrality [@MakovPayne1; @MakovPayne2]. As discussed by Makov and Payne [@MakovPayne1; @MakovPayne2], a leading behavior, $q^2\alpha /2L$, arises from the self-interaction of the neutralizing charge with its periodic images, where $\alpha$ is the (supercell-dependent) Madelung constant, $q$ is neutralizing charge, and $L^3 = \Omega$. Correction of the total energy by the leading term leads to more rapid size convergence. Figure \[PECorr\] illustrates this effect. The bottom panel shows the slow convergence of the total energy with the system size in charged systems, while the upper panel shows the more rapid convergence after the correction has been made, i.e., the slowly convergent $q^2\alpha /2L$ contribution has been subtracted.
![\[PECorr\] An illustration of the size effect in the calculation of charged atoms. The top panel shows the corrected total energies for P, P$^+$ and P$^{++}$, while the bottom panel shows the uncorrected ones. Cell sizes used in this study are the same as in Fig. \[DE\_E\_2E\]. The errorbars are smaller than the point size. ](\FIGDIR{}P_E_Ecorr.eps)
Trotter time step error {#sec:Trotter}
-----------------------
The Trotter error arises from neglecting higher order terms of the imaginary-time propagator, $e^{-\Delta\tau{{{\hat{H}}}}}$, when we apply the Trotter-Suzuki decomposition in Eq. (\[eq:Trotter\]). The Trotter error can be eliminated by extrapolation, as demonstrated in Fig. \[trotter\_P\] for P, P$^+$, and P$^{++}$. The results reported in the next section use either a linear extrapolation a fixed $\Delta \tau$ ($\sim 0.05$ Ha$^{-1}$) which is sufficiently small so that the Trotter error is well within the statistical error.
![\[trotter\_P\] Trotter errors for P, P$^+$ and P$^{++}$ energies (shown in the top, middle, and bottom panels, respectively) in a $14 \times 14 \times 14 {\,a_0}^3$ simulation box. The data points at $\Delta \tau = 0$ represent the Trotter extrapolation. ](\FIGDIR{}P_Px_Pxx_regression.eps)
Results {#sec:results}
=======
Table \[tab:DE\] shows dissociation energies from planewave AF QMC calculations compared with experimental values and with LDA, GGA, HF, and diffusion Monte Carlo (DMC) calculated results. LDA trial wave functions were used in the molecular calculations, corresponding to the following electronic configurations: $\sigma ^2 _{\rm{3s}}
\sigma ^{*2} _{\rm{3s}}\sigma ^2 _{\rm{3p}}\pi ^4 _{\rm{3p}}$ for P$_2$, $\sigma ^2 _{\rm{3s}} \sigma ^{*2} _{\rm{3s}}\sigma ^2
_{\rm{3p}}\pi ^4 _{\rm{3p}}\pi ^{*2} _{\rm{3p}}$ for S$_2$, $\sigma ^2
_{\rm{3s}} \sigma ^{*2} _{\rm{3s}}\sigma ^2 _{\rm{3p}}\pi ^4
_{\rm{3p}}\pi ^{*4} _{\rm{3p}}$ for Cl$_2$, and $\sigma ^2 _{\rm{4s}}
\sigma ^{*2} _{\rm{4s}}\sigma ^2 _{\rm{4p}}\pi ^4 _{\rm{4p}}$ for As$_2$. All dimers except S$_2$ have “closed-shell” configurations in which all occupied orbitals are fully filled. All the calculations used OPIUM pseudopotentials.
LDA AF QMC (LDA) DMC [@GrossmanDMC] HF GGA [@PPP] Expt. [@HBconst]
-------- ------ --------------- -------------------- ------ ------------ ------------------
Si$_2$ 3.88 3.12(8)$\,\,$ 3.21(13)
P$_2$ 5.97 5.19(16) 4.73(1) 1.74 5.22 5.08
S$_2$ 5.61 4.63(17) 4.31(1) 2.29 4.94 4.41
4.48(19)
Cl$_2$ 3.12 2.78(10) 2.38(1) 0.67 2.76 2.51
As$_2$ 5.04 3.97(17) 3.96
: \[tab:DE\] Calculated dissociation energies (in eV) using LDA, AF QMC, DMC, HF, and GGA methods. Experimental values are in the last column (with the zero-point energy removed). The AF QMC calculations used LDA trial wave functions, except for sulfur, where we have also used unrestricted HF trial wave functions. Values from the largest supercell are shown here, since convergence has been reached. The statistical errors are given in parentheses. The HF results are obtained from an in-house planewave based code. The AF QMC result for Si$_2$ is taken from Ref. .
Both restricted and unrestricted trial wave functions were tested in the AF QMC calculations, but there were no difference within statistical errors. (In restricted trial wave functions, the orbitals with minority spin are identical to the corresponding majority-spin orbitals.) Unrestricted HF trial wave functions were also used to calculate the dissociation energy of S$_2$, which has an open-shell configuration. As shown in Table \[tab:DE\], there is no difference within statistical errors. The overall agreement between the AF QMC results and experiment is very good. The LDA and GGA slightly overestimate the dissociation energy, while the HF method significantly underestimates it. The heaviest dimer we calculated, As$_2$, is also in excellent agreement with experiment and compares favorably with results from other quantum chemistry methods [@SakaiMiyoshi; @Moshizuki].
---- --------- ---- ---- ---- ------- ------- -------- ---- ---- ---- ---- ------- ---- ------- ------- -------- ---- ----
Pseudo
Al [OAL]{} 5 88 5 88(2) 5 99 24 46 24 66(2) 24 81
Si OPIUM 8 18(2) 8 15 24 59(4) 24 50
P OPIUM 9 97 10 57 10 74(6) 10 49 29 41 30 42 30 79(6) 30 26
[OAL]{} 9 94 10 41 10 61(3) 29 11 29 97 30 28(6)
S OPIUM 9 33 10 45 10 09(7) 10 36 32 42 33 85 34 16(7) 33 67
[OAL]{} 9 21 10 35 10 08(2) 32 06 33 51 33 61(2)
Cl OPIUM 11 69 13 12 12 96(11) 12 97 34 36 36 92 36 76(10) 36 78
[OAL]{} 11 76 13 02 12 89(6) 34 30 36 64 36 25(6)
0 00 0 27
0 19 0 36
–0 09 –0 18
0 16 0 28
---- --------- ---- ---- ---- ------- ------- -------- ---- ---- ---- ---- ------- ---- ------- ------- -------- ---- ----
We show in Table \[tab:IP+IIP\] the first and second ionization energies of Al, Si, P, S and Cl. For comparison, experimental values and results from LDA and HF calculations are also shown. The AF QMC calculations are performed using both the OPIUM and [OAL]{} pseudopotentials together with LDA trial wave functions, except for sulfur with the [OAL]{} pseudopotential, where HF trial wave functions are also used. Again, no statistically significant dependence on the trial wave function is seen. The tabulated results are converged with respect to size effects, which are negligible compared to the QMC statistical error for supercells larger than $16 \times 16 \times 16 {\,a_0}^3$.
Ionization energies obtained using the LDA are generally in very good agreement with experiment, and the present results are consistent with this, with deviations typically within $0.1$ eV for IP and $0.3$ eV for IIP, regardless of which pseudopotential is used. The HF method results in larger deviations of $0.5$–$1.3$ eV compared to the experiment.
AF QMC results for Al and Si are in good agreement with experiment, and for Si the agreement is comparable to that obtained using DMC[@LeeNeeds], $8.166(14)$ eV and $24.444(22)$ eV for IP and IIP, respectively. For P, S, and Cl, however, the agreement between AF QMC and experiment is not as uniform. In particular, there is a significant dependence on the choice of the pseudopotential. For P and S, the AF QMC ionization energies are better estimated using [OAL]{}, while for Cl, OPIUM pseudopotential gives better results. We will discuss this dependence in the next section. Agreement between the best AF QMC and experiment values is in general very good.
Discussion {#sec:discussion}
==========
For dissociation energies, the agreement between AF QMC and experiment was uniformly very good. The appearance of larger discrepancies between AF QMC and experiment for ionization energies is somewhat surprising. The AF QMC calculations above were systematically converged with respect to finite size effects, Trotter time step error, and planewave basis size. Two remaining possibilities, errors arising from the use of the phaseless approximation and errors due to the use of pseudopotentials, are discussed in this section.
Our overall experience with the phaseless AF QMC[@sz-hk; @AFQMC-CPC2005; @Al-Saidi_TMO; @Al-Saidi; @Purwanto2004] suggests that the error due to the phaseless approximation itself is typically small. Recent AF QMC calculations using a Gaussian basis [@Al-Saidi] show that, in a variety of atoms and molecules, the AF QMC agrees well with experiment or high-level quantum chemistry methods such as the coupled cluster with single and double excitations and perturbative corrections for triple excitations \[CCSD(T)\][@CCSD_Cizek; @CCSD_Cizek_ACP; @CCSD_Purvis; @CCSD_T_pertb]. Open-shell systems such as P$^+$, P$^{++}$, S, S$^{++}$, Cl, and Cl$^+$, where the $p$ shell is neither half- nor fully-filled, tend to be more difficult to treat in general [@OpenShell]. In these cases a single-determinant trial wave function breaks the symmetry by using only one of the degenerate states, and the phaseless approximation could affect the accuracy of the results. Some indication of this may be present in our results (see Table \[tab:IP+IIP\]) where larger errors are observed for energy differences between half-filled and “open-shell” systems, such as the P$^+$ ionization energy. It is possible to use multideterminant trial wave functions in these cases. We have done several tests, in which we “symmetrize” the trial wave function, resulting in a linear combinations of three different determinants. This is designed to equally treat the $p_x$, $p_y$, and $p_z$ orbitals in the open shell. However, there seems to be no observable improvement over the single-determinant trial wave function at the level of statistical accuracy in this paper.
To isolate and quantitatively evaluate the errors due to the phaseless approximation on the second-row atoms studied here, we performed calculations with both AF QMC and CCSD(T) methods for Cl, Cl$^+$, and Cl$^{++}$, using identical Gaussian basis sets and the [[OAL]{} pseudopotential]{}. (The [[OAL]{} pseudopotential]{} was chosen since its form is already compatible with standard quantum chemistry programs.) Thus, both the AF QMC and CCSD(T) methods are applied to the same many-body Hamiltonian, expressed in the Hilbert space spanned by the selected Gaussian basis set. The CCSD(T) method is approximate but is known to be very accurate in atoms and in molecules near equilibrium. For the comparison, we employed an uncontracted aug-cc-pVDZ[@aug-cc-pVXZ; @PNL-basisform] basis set, where Gaussian functions with exponents larger than $98$ are removed, resulting in a $(7s7p2d)$ basis set. As shown in Table \[tab:GAFQMC-vs-CCSD(T)\], the AF QMC and CCSD(T) absolute total energies agree to within $0.07$ eV, and their ionization energies agree to within $0.04$ eV. These results indicate that the intrinsic error due to the phaseless approximation in the above planewave calculations is likely also small, consistent with previous experience with phaseless AF QMC[@sz-hk; @AFQMC-CPC2005; @Al-Saidi_TMO; @Al-Saidi; @Purwanto2004]. This suggests that errors due to the use of pseudopotentials are largely responsible for the deviations in ionization energies noted above.
------------- ------- ------- ------ ------- -------- -------- ------ ----- ------
Cl –403 91(1) –403 96
Cl$^+$ –391 37(1) 12 54(2) –391 44 12 51
Cl$^{++}$ –368 43(1) 35 49(2) –368 43 35 53
------------- ------- ------- ------ ------- -------- -------- ------ ----- ------
: \[tab:GAFQMC-vs-CCSD(T)\] Calculated total energies ($E$) and ionization energies (IP) for Cl, using AF QMC with a Gaussian basis together with the corresponding CCSD(T) results, using identical basis sets. The [[OAL]{} pseudopotential]{} is used with a double-zeta quality $(7s7p2d)$ Gaussian basis set. All energies are in eV. Since CCSD(T) is known to be accurate for atoms, differences compared to Gaussian-based AF QMC provide an estimate of errors due to the phaseless approximation in AF QMC.
The pseudopotentials used here were generated using independent-electron HF or DFT mean-field type calculations of atomic reference systems. While the transferability of these pseudopotentials to HF or DFT calculations of molecules or solids is well understood, their accuracy in many-body calculations is more problematic.[@Shirley1993; @LeeNeeds] The dependence of the AF QMC results in Table \[tab:IP+IIP\] on the choice of pseudopotentials is consistent with this.
To estimate the pseudopotential errors in our AF QMC calculations and to obtain insight into their origin, we have carried out several additional calculations. We first performed pseudopotential and all-electron (AE) CCSD(T) calculations to estimate the transferability of the pseudopotential in a many-body context. First and second ionization energies for P, S, and Cl atoms were calculated using both methods. The coupled-cluster calculations were performed using the and packages[@Gau98; @Gau03]. To eliminate basis set convergence errors, we performed a series of AE calculations using the aug-cc-p$w$CV$x$Z basis sets[@aug-cc-pVXZ; @aug-cc-pCVXZ; @PNL-basisform], where $x=$ D, T, Q for double, triple, and quadruple zeta basis sets, respectively. The infinite-basis estimate of the total energy, $E_{\infty}$, is then obtained using extrapolation[@Inf-basis] [ $$\label{eq:E-cbs}
\begin{split}
E_{\infty} \approx E_x - b e^{-cx}
\,,
\end{split}$$ ]{} where $x$ is $2, 3, 4$ for double, triple, and quadruple zeta basis sets, respectively, and $b$ and $c$ are fitting parameters. We then take the difference of the extrapolated energies as the ionization potential shown in Table \[tab:CCSDT-IP\]. For the pseudopotential calculations, the [OAL]{} ECP is used. Here we use the aug-cc-pV$x$Z basis sets[@aug-cc-pVXZ; @PNL-basisform] that are fully uncontracted, and again we use the extrapolation scheme in Eq. (\[eq:E-cbs\]).
---- -- ---- ---- ---- ---- ------ ---- ---- ---- -- ---- ---- ---- ---- ------ ---- ---- ----
P 10 48 10 47 0 01 10 49 30 13 30 18 –0 05 30 26
S 10 24 10 29 –0 05 10 36 33 64 33 64 0 00 33 67
Cl 12 92 13 05 –0 13 12 97 36 51 36 77 –0 26 36 78
---- -- ---- ---- ---- ---- ------ ---- ---- ---- -- ---- ---- ---- ---- ------ ---- ---- ----
: \[tab:CCSDT-IP\] Estimates of pseudopotential errors: CCSD(T) results for the first and second ionization energies of P, S, and Cl. All energies are in eV. [OAL]{} and AE results are shown together with the error in the ionization energy due to the pseudopotential, $\Delta\textrm{IP}_\mathrm{psp}$.
The CCSD(T) results for AE and [OAL]{} calculations are presented in Table \[tab:CCSDT-IP\]. AE CCSD(T) ionization energies are seen to be in excellent agreement with experimental values. \[We note that AE calculations using aug-cc-pCV$x$Z basis set (a variation of aug-cc-p$w$CV$x$Z) yield almost identical result, where the estimated IP differs by $\lesssim 0.03$ eV, and IIP by $\lesssim 0.07$ eV.\] The difference between the AE and [OAL]{} results provides an estimate of the error due to the pseudopotential in many-body calculations. The results in Table \[tab:CCSDT-IP\] indicate that the [OAL]{} pseudopotential tends to underestimate the ionization energies. The rms error attributable to the pseudopotential is about $0.12$ eV. This is not negligible compared to the overall AF QMC error, which is $\langle\Delta\textrm{IP}(\textrm{{OAL}{}})\rangle_{\mathrm{rms}} = 0.25$ eV for P, S, and Cl first and second ionization energies. An additional possible source of pseudopotential error in the planewave AF QMC calculations is the use of the fully nonlocal, separable Kleinman-Bylander (KB) construction of the pseudopotential. For example, the [OAL]{} ECP is defined in the usual semilocal form, which is used in quantum chemistry programs: [ $$\label{eq:PSP-semiloc}
\begin{split}
{\hat{V}_{\textrm{ei}}}^{\textrm{(SL)}}(\mathbf{r})
= \sum_{l,m} {|{{Y_{lm}}}\rangle} V_l(r) {\langle{{Y_{lm}}}|}
\,.
\end{split}$$ ]{} where $V_l(r)$ is the angular-momentum-dependent potential. For efficient use in the planewave calculations, it is common to express this pseudopotential in the fully nonlocal separable KB form shown in Eq. (\[eq:PSP-KB\]). While the KB and semilocal forms are identical when they act on the reference atomic state, the KB form can differ for other states.
To investigate the effect of pseudopotential KB formation, we perform LDA [@PerdewWang] calculations with the [[OAL]{} pseudopotential]{}: (1) planewave basis calculations with the KB form of the [OAL]{} using the package (calculations are converged with respect to ${E_{\textrm{cut}}}$ and box size), and (2) local basis calculations with the semilocal form of the [OAL]{} using (again we use the sequence of aug-cc-pV$x$Z basis sets to extrapolate to the infinite basis limit). The OAL pseudopotential was converted to the KB form using pseudo-orbitals obtained in an LDA calculation for the neutral atom. For the purpose of constructing the fully nonlocal projectors, the effects of using LDA rather than HF pseudo-orbitals are not expected to to be significant. In both methods, the total-energies are converged to within $0.5$ mHa ($\approx 0.01$ eV). Table \[tab:KB-test\] presents the results for the Cl ionization energies. The [[OAL]{} pseudopotential]{} expressed in the KB form tends to underestimate the LDA total energies, with the discrepancy increasing with the ionization state. Ionization energies are thus underestimated by up to $0.15$ eV. The same trend is observed in the calculated planewave AF QMC ionization energies compared to experiment, which indicates that the KB form may contribute errors of the order of $0.1-0.2$ eV for Cl using the [[OAL]{} pseudopotential]{}. It is clear that the quality of the pseudopotential is crucial for obtaining accurate results with AF QMC.
------------- -------- ----- ------ ----- -------- -------- ----- ------ -------- ---- -------- ----
System [ ]{}
Cl –403 798 –403 752 –0 05
Cl$^+$ –390 842 12 956 –390 701 13 051 –0 14
Cl$^{++}$ –367 217 36 581 –367 021 36 731 –0 20
------------- -------- ----- ------ ----- -------- -------- ----- ------ -------- ---- -------- ----
: \[tab:KB-test\] LDA calculations of chlorine energies, the Kleinman-Bylander (KB) and semilocal (SL) forms of the OAL pseudopotential. The programs used are indicated in parantheses, and all energies are in eV.
These test calculations are limited in scope. Small systematic errors may well be still present, for example from the planewave size-extrapolations, from Gaussian basis set extrapolations, and from approximations inherent in coupled cluster calculations at the CCSD(T) level. They suggest, however, a rather consistent picture for understanding the AF QMC results in Table \[tab:IP+IIP\]. Pseudopotential errors due to different origins appear to be the main cause for the discrepancies with experimental values. When such errors are removed, it seems that the accuracy of the planewave AF QMC is at the level of 0.1 eV.
PSP
--------- ------------------- ------------------ ---------------------- ---------------------- ------- ------- -------- ----
expt = 10 49
OPIUM [**10**&**57**]{} 9 97 10 74(6)
[OAL]{} 10 41 [**9**&**94**]{} [**10**&**61(3)**]{} 10 48
AE 10 53 9 91 10 47
expt = 30 26
OPIUM [**30**&**42**]{} 29 41 30 79(6)
[OAL]{} 29 97 [**29**&**11**]{} [**30**&**28(6)**]{} 30 13
AE 30 37 29 08 30 18
expt = 5 08
OPIUM [**5**&**97**]{} [**1**&**74**]{} [**5**&**19(16)**]{}
[OAL]{} 5 29 0 98 3 88(8) 4 39
AE 6 18 1 65 4 98[^1]
: \[tab:PSP-quality\] Phosphorus ionization and dissociation energies for P$^+$, P$^{++}$, and P$_2$ computed using LDA, HF, AF QMC, and CCSD(T) methods, shown below in eV. AE results are provided to benchmark the pseudopotentials. The numbers shown in boldface are those closest to the AE results (for LDA and HF) or the experimental values (for AF QMC).
Since the quality of a pseudopotential within HF and LDA can be easily determined (by its ability to reproduce AE results), it is interesting to see how well this correlates with the performance of the pseudopotential in a many-body calculation. In Table \[tab:PSP-quality\] we show various energies obtained from LDA, HF, AF QMC, and CCSD(T) calculations in phosphorus. (Similar trends also hold for sulfur and chlorine.) Results are shown using the OPIUM and [OAL]{} pseudopotentials, together with AE results for independent-electron and CCSD(T) methods. In the LDA calculations, the LDA-based OPIUM potential performs uniformly better for all quantities. In the HF calculations, the HF-based [OAL]{} pseudopotential performs well, except for the dissociation energy. (As mentioned earlier, the [[OAL]{} pseudopotential]{} lacks a $d$ projector in its construction, which results in poor molecular energies.) Overall, the HF method appears to give a better indication of pseudopotential performance with AF QMC. In the chlorine second ionization energy, for example, HF predicts that the OPIUM pseudopotential performs better than the OAL pseudopotential: IIP$_{\textrm{HF, OPIUM}} = 34.37$ eV, IIP$_{\textrm{HF, {OAL}{}}} = 34.30$ eV, compared to IIP$_{\textrm{HF, all-electron}} = 34.36$ eV. AF QMC results in Table \[tab:IP+IIP\] show a similar trend. We also note in Table \[tab:IP+IIP\] that the LDA-based OPIUM pseudopotentials tend to overestimate the ionization energies, while the OAL pseudopotentials do the opposite. The tabulated rms averages suggest that the performance of AF QMC with the LDA-generated OPIUM pseudopotential varies more widely across different atomic species, especially in the second ionization energies. The HF-generated [OAL]{} ECP, on the other hand, performs more consistently and yields better agreement with experiment in the majority of species studied here. Testing pseudopotentials using Hartree-Fock calculations may, therefore, be a useful predictor of their performance in the many-body AF QMC method.
Summary {#sec:summary}
=======
We have presented electronic structure calculations in atoms and molecules using the phaseless AF QMC method with a planewave basis and norm-conserving pseudopotentials. Various algorithmic issues and characteristics were described and discussed in some detail, and we have illustrated how the AF QMC method can be implemented by utilizing standard DFT planewave techniques. The structure of the AF QMC calculation is an independent collection of random walker streams. Each stream resembles an LDA calculation, which makes the overall computational scaling of the method similar to LDA calculation with a large prefactor. This makes the AF QMC approach more efficient than explicit many-body methods. All of the reported results were obtained using single-determinant trial wave functions, directly obtained from either LDA or HF calculations. This reduces the demand for wave function optimization in QMC and is potentially an advantage. The method also offers a different route to the sign problem by carrying out the random walks in Slater determinant space. Because our method is based in Slater determinant space, any single-particle basis can be used.
Results for the dissociation and ionization energies of second-row atoms and dimers in Group 3A-7A, Al, Si, P, S, Cl, as well as the As$_2$ dimer, were presented using the planewave-based phaseless AF QMC method. The effects of the phaseless approximation in AF QMC were studied, and the accuracy of the pseudopotentials were examined. Comparisons were made with experiment and with results from other methods including the DMC and CCSD(T). Errors due to the phaseless approximation were found to be small, but non-negligible pseudopotential errors were observed in some cases. In addition to pseudopotential errors that arise due to their construction in mean-field type DFT or HF calculations, possible errors in ionization energies arising from the separable Kleinman-Bylander form of the pseudopotentials were also observed. With the appropriate pseudopotentials, the method yielded consistently accurate results.
We acknowledge the support of DOE through the Computational Materials Science Network (CMSN), ARO (grant no. 48752PH), NSF (grant no. DMR-0535529), and ONR (grant no. N000140510055). Computing was done in the Center of Piezoelectric by Design (CPD) and National Center for Supercomputing Applications (NCSA). W. P. would also like to thank Richard Martin, Wissam Al-Saidi, and Hendra Kwee for many fruitful discussions.
[^1]: Frozen-core calculation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Diffusion is the main limitation of storage time in spatially multimode applications of warm atomic vapors. Precise knowledge of diffusional decoherence in the system is desired for designing most of vapor memory setups. Here we present a novel, efficient and direct method of measuring unbiased diffusional decoherence, clearly distinguished from all other decoherence sources. We found the normalized diffusion coefficients of rubidium atoms in noble gases to be as follows: neon 0.20 cm$^{2}$/s, krypton 0.068 cm$^{2}$/s and we are the first to give an experimental result for rubidium in xenon: 0.057 cm$^{2}$/s. Our method consists in creating, storing and retrieving spatially-varying atomic coherence. Raman scattering provides a necessary interface to the atoms that allows for probing many spatial periodicities of atomic coherence concurrently. As opposed to previous experiments the method can be used for any single sealed glass cell and it does not require any setup alterations during the measurements and therefore it is robust and repeatable.'
author:
- 'Rados[ł]{}aw Chrapkiewicz'
- Wojciech Wasilewski
- 'Czes[ł]{}aw Radzewicz'
title: 'How to measure diffusional decoherence in multimode rubidium vapor memories? '
---
Introduction {#intro}
============
In recent years warm atomic ensembles have been widely used in many applications in quantum optics and atomic physics. The most promising applications include quantum repeaters [@Duan2001b], quantum memories [@Appel2008] and ultraprecise magnetometry [@Chalupczak2012]. They have also been shown to be an effective medium for four-wave mixing processes [@McCormick2007], electromagnetically induced transparency (EIT) [@Fleischhauer2005] and slow light generation [@Matsko2001].
An advantage of using warm atomic gas is undoubtedly simplicity of performing experiments and obtaining large optical depths. However, it entails fast thermal motion of atoms. Atomic motion limits the efficiency of many modern systems that use multimode properties of atomic ensembles. At best this motion can be slowed down and made diffusive by addition of a suitable buffer gas. Since diffusion is usually the main source of decoherence in multimode vapor memory systems, ability to measure and control its speed is highly desired. In particular it would be very useful to distinguish pure diffusional decoherence from other decoherence effects.
There is a number of currently developing systems where diffusion is the main limitation. An important example of such a system is storing and retrieving transverse modes and images in gradient echo memory [@Glorieux2012; @Higginbottom2012; @Clark2013; @Luo2013], in collective Raman scattering [@Chrapkiewicz2012] or in EIT [@Firstenberg2012]. Typically diffusional motion of atoms in a buffer gas limits the storage time [@Shuker2008; @Vudyasetu2008], restricts the number of spatial modes retrieved [@Chrapkiewicz2012] or broadens the EIT spectrum [@Shuker2007]. Knowledge of the exact diffusion coefficient is particularly important for designing experiments with diffraction cancellation [@Firstenberg2008; @Firstenberg2009; @Firstenberg2012; @Yankelev2013]. Further progress would be significantly facilitated if one possessed convenient, robust and repeatable method for precise diagnostic of decoherence in the actual cell of a particular setup.
However, the available methods of measuring diffusion decoherence are indirect and require either variation of buffer gas pressure, prior knowledge of other sources of decoherence [@Franzen1959] or setup alterations within a single measurement [@Bicchi1980; @Glassner1996]. This makes them unsuitable for modern experiments where exact knowledge of diffusional decoherence of a single sealed glass cell in a specific setup is required.
Here we propose a novel, direct method which allows us to measure the diffusion in any given cell. The measurement provides more than enough data to verify its self-consistency and single out the diffusion from other motion-independent sources of decoherence. The method should be relatively easy to incorporate into a number of quantum memory setups.
As a demonstration we measure the diffusion coefficients of rubidium in neon, krypton and xenon in sealed glass cells at a pressure of a few torrs. These results will be useful for designing future experiments, since the data available till now is rather scarce and, most importantly, it was retrieved using indirect methods [@Franzen1959; @Bernheim1962; @McNeal1962; @Arditi1964; @Franz1965; @Gozzini1967; @Bouchiat1972; @Vanier1974; @Bicchi1980]. The data available for neon is inconsistent and that for krypton is hardly available [@Bouchiat1972; @Higginbottom2012].
We also recommend using xenon as a buffer gas, for which we provide the very first experimental data as far as we know. Despite the latest applications of hyperpolarized xenon [@Fink2005], the diffusion coefficient of rubidium in this gas has only been deduced from cross sections of velocity changing collisions [@Gibble1991] or interaction potentials [@Hamel1986].
This paper is organized as follows: in Sec. 2 we introduce the principles of our method, Sec. 3 describes in detail the experimental implementation, Sec. 4 contains the experimental results together with the reference data available. Finally, Sec. 5 concludes the paper.
Method
======
![(Color online) (a) Planewave components of spatially dependent atomic coherence $\rho_{01}(\mathbf{r})$ act as diffraction gratings deflecting laser beam at different angles. (b) Components of high periodicity are blurred faster due to diffusion, therefore the intensity of deflected light will decay faster for higher angles of deflection. During storage time $\tau$, the pattern component of specific periodicity corresponding to the wave vector $\mathbf{K}$ blurs with decay rate $D\mathbf{K}^{2}$, where $D$ is the diffusion coefficient. \[fig:Blurr\]](readout){width="8cm"}
General idea
------------
Quantification of the diffusive motion of atoms using optical methods can be done in three general steps. At first a group of atoms has to be chosen and distinguished from others by changing their internal state. Then the atoms are let to spread due to diffusion in the absence of light. In the third stage one probes the group and observes the effects of the spread. This general scheme has various implementations [@Franzen1959; @Bernheim1962; @McNeal1962; @Arditi1964; @Franz1965; @Gozzini1967; @Bouchiat1972; @Vanier1974; @Bicchi1980] which typically consisted in exciting and probing a pencil-shaped atomic group using light beams. Instead here we create and, after a certain delay, probe patterns of spatially dependent atomic coherence $\rho_{01}(\mathbf{r})$ between two long lived atomic levels ${|0\rangle}$ and ${|1\rangle}$.
Patterns of atomic coherence are created through spontaneous (Stokes) Raman scattering. Each pattern comprises many plane-wave components with different periodicities. Those components decay at different rates due to diffusive motion of the atoms. After a certain storage time relative contribution of each plane-wave component can be measured by driving the anti-Stokes scattering. Then each plane-wave component acts as a diffraction grating deflecting driving laser beam as illustrated in Fig. \[fig:Blurr\] (a). By measuring the intensity of the anti-Stokes scattering light as a function of deflection angle and time between pattern creation and readout, we can calculate the decay rate of different plane-wave components constituting atomic coherence. We rely on the fact that in the diffusion process periodic patterns of atomic coherence does not change their period but they are blurred over time as depicted in Fig. \[fig:Blurr\] (b). As atoms move, coherence at a specific point $\mathbf{r}_{0}$ will reshuffle its values with the neighboring points. Evolution of atomic coherence in the dark will be described by the equation of diffusion with a coefficient $D$ and homogeneous depolarization with a rate $\gamma_{0}$ [@Lowe1967; @Glorieux2012]: $$\frac{\partial}{\partial t}\rho_{01}(\mathbf{r},t)=D\nabla^{2}\rho_{01}(\mathbf{r},t)-\gamma_{0}\rho_{01}(\mathbf{r},t).\label{eq:diff}$$ This equation can be readily solved in Fourier domain: $$\rho_{01}(\mathbf{r},t)=e^{-\gamma_{0}t}\sum_{\mathbf{K}}\beta(K)e^{-DK^{2}t}e^{i\mathbf{K\cdot r}}.\label{eq:coherencedecay}$$
Evolution of each plane wave component of initial amplitude $\beta(K)$ and wave vector $K$ is described by a simple exponential decay at a rate $\gamma(\mathbf{K})=\gamma_{0}+DK^{2}$. As long as the evolution of $\rho_{01}$ can be described by Eq. with position-independent homogeneous depolarization with a rate $\gamma_{0}$, the measurement of decay rates $\gamma(K)$ is sufficient to calculate $D$ as a coefficient of the quadratic term of $\gamma(K)$.
Creation and probing of atomic coherence
----------------------------------------
![(Color online) (a) Atomic coherence $\rho_{01}(\mathbf{r})$ is created in spontaneous Stokes scattering. (b) The difference between wave vectors of laser field **$\mathbf{k_{1}}$** and the scattering field component **$\mathbf{k_{2}}$** is stored in the spatially dependent phase of the atomic coherence. (c) Atomic coherence can be converted back to light in anti-Stokes scattering. (d) The atomic coherence phase is imprinted on anti-Stokes scattering as momentum conservation. Unlike the Stokes process, anti-Stokes scattering deterministically retrieves created coherence pattern on demand. \[fig:Levels\]](levels){width="8cm"}
In Fig. \[fig:Levels\](a) we illustrated the atomic levels involved in Stokes scattering leading to creation of coherence $\rho_{01}(\mathbf{r})$ between levels ${|0\rangle}$ and ${|1\rangle}$. Upon spontaneous scattering of a laser beam detuned from the $|0\rangle\leftrightarrow|e\rangle$ transition both scattered light and atomic coherence are created. We shall consider a simple case where the laser beam and the scattered light are planewaves with wave vectors, **$\mathbf{k_{1}}$** and **$\mathbf{k_{2}}$**, respectively. The difference between the laser field wave vector $\mathbf{k_{1}}$ and the created photons wave vector $\mathbf{k_{2}}$ is accumulated in atoms as a spatial phase of atomic coherence. As illustrated in Fig. \[fig:Levels\](b) the atomic coherence created will be of a form $\rho_{01}(\mathbf{r})=\beta e^{i\mathbf{K\cdot r}}$, where **$\mathbf{K=k_{1}-k_{2}}$**.
Such periodic atomic coherence $\rho_{01}(\mathbf{r})$ can work as a diffraction grating and deflect a laser beam. This is realized in anti-Stokes scattering process presented in Fig. \[fig:Levels\](c) in which laser beam detuned from the $|1\rangle\leftrightarrow|e\rangle$ transition is scattered at an angle. The spatial phase of atomic coherence $\rho_{01}(\mathbf{r})$ is imprinted back onto the scattered photons as follows from Bragg condition. Provided driving laser beam is the plane wave with a wave vector $\mathbf{k_{1}'}$, the diffracted light wave vector will be $\mathbf{k_{2}'=k_{1}'-K}$, as illustrated in Fig. \[fig:Levels\](d). Therefore by observing the intensity of the light $I_{\mathrm{AS}}(\theta)$ scattered at an angle $\theta$, we register a signal which is proportional to the modulus square of the corresponding plane wave component of atomic coherence $|\beta(K)\exp(-\gamma(K)t)|^{2}$, with $\theta=K/k_{1}'=K\lambda/2\pi$.
Averaging and retrieving the diffusion coefficient
---------------------------------------------------
Spatially varying atomic coherence $\rho_{01}(\mathbf{r})$ is created in a spontaneous Stokes scattering process, which populates various planewave components randomly. Nonetheless, the average modulus square of the excitation amplitude $\langle|\beta(K)|^{2}\rangle$ created right after the scattering is set by the driving pulse parameters and can be kept constant between measurement series. Therefore, we can calculate the average intensity of the light scattered at a certain angle $\theta=K/k_{1}'$ and for a given storage time $\tau$, incorporating Eq. :
$$\langle I_{\mathrm{AS}}(\theta=K/k_{1}',\tau)\rangle=\eta(K)\langle|\beta(K)|^{2}\rangle e^{-2\gamma(K)\tau},\label{eq:ASamp}$$
where $\eta(K)$ is efficiency of readout, *a priori* dependent on $K$. The only factor that depends on the diffusion time $\tau$ is the intensity decay factor $e^{-2\gamma(K)\tau}$, which provides direct information about the decay rate $\gamma(K)$. Therefore for a given angle of observation $\theta=K/k_{1}'$ we can infer $\gamma(K)$ from an exponential fit to a series of experimental data taken for successive $\tau$.
By repeating the decay fits for many $K$ we can gather and then fit the expected functional dependence $\gamma(K)=\gamma_{0}+DK^{2},$ to obtain the diffusional coefficient $D$. In principle, we could use only two measurements of decay rates $\gamma$ corresponding to just two different directions. Note that thanks to a particular way of populating spatially varying atomic coherence we create and probe many wave vectors concurrently without altering the setup. Thence we obtain many points corresponding to a broad span of $K$ vectors in a single measurement sequence, which provides for a robust quadratic fit of $\gamma(K)$. The quality and reliability of the experimental data is directly reflected in this last fit.
Angular blurring at readout
---------------------------
So far we have assumed the driving beam to be a planewave. The finite size $w$ of the driving laser beam in anti-Stokes scattering results in limited resolution in probing a wave vector space. The angular spread of the laser beam driving the readout will be transferred onto the angular distribution of the scattered light due to momentum conservation even for planewave atomic coherence $\rho_{01}(\mathbf{r})$. Thus for any specific angle of observation $\theta=K/k_{1}'$ we detect the scattered light originating from several distinct Fourier components of the atomic coherence pattern. The contribution will come from the component of the wave vector $K$ and its vicinity to the spread $\sigma$. We expect the spread $\sigma$ to be of the order of the inverse of the driving laser beam size $1/w$.
The result of the limited resolution is an overall increase in the decay rates observed $\gamma_{\mathrm{obs}}$. It can be estimated by convolving the storage time-dependent Fourier distribution of atomic coherence $\rho_{01}(\mathbf{r})$ with a Gaussian of a spread $\sigma$, yielding:
$$\gamma_{\mathrm{obs}}(K)=(\gamma_{0}+2D\sigma^{2})+DK^{2}.\label{eq:gobs}$$
Note that the term quadratic in $K$ in the above formula did not change, therefore the procedure of obtaining the diffusion coefficient $D$ remains unchanged. We only have to assure that the term $2D\sigma^{2}$ is small as compared to $DK^{2}$. This can be done by increasing the beam size $w$.
In conclusion all measurements can be completed by varying only one parameter – the diffusion time $\tau$ equal to laser pulse separation while collecting scattered light on a camera. Data analysis require three straightforward steps: averaging, exponential fit and eventually quadratic fit to obtain $D$. This makes the whole procedure relatively quick and simple to repeat.
Additional sources of decoherence
---------------------------------
In Eq. we assumed that decoherence could be divided into two types: the $K$-dependent diffusional type and the homogenous type. The latter originates mostly from atomic collisions. This division applies sufficiently to typical experimental conditions; however, other types of processes may contribute to decoherence as well.
In Eq. we neglected the stray magnetic field. Taking such a field into account would require introducing an extra term in Eq. : $i\mu_{B}(g_{1}m_{1}-g_{0}m_{0})B(\mathbf{r})\rho_{01}(\mathbf{r},t)$ to represent additional, space-dependent build-up of the phase, where $\mu_{B}g_{i}m_{i}$ is the magnetic moment for the $i$-th level. Nevertheless, we estimated that our final results would change only by 10% in the presence of a magnetic field gradient of 0.5 Gauss/mm or with a quadratically changing magnetic field of 10 mGauss/mm$^{2}.$ These values are considerably higher than in normal experimental conditions even without applying magnetic shielding.
We also noted that decoherence due to spin-exchange collisions might lead to quite complicated effects if $\rho_{01}(\mathbf{r})$ was a fast varying function of position. These collisions could alter the state of the atoms at a different rate and at different points in space, leading to a nontrivial space dependence of $\gamma_{0}$. We calculated the rate of these collisions to be of the order of 1.5 kHz, which is negligible as compared to the diffusional decay caused by the atoms leaving the laser beams.
Experiment
==========
![(Color online) Central part of the experimental setup. Rubidium-87 atoms are mixed with the buffer gas in a glass cell inside a double magnetic shielding. Pump, write and read are the laser beams used to prepare atoms in the ground state ${|0\rangle}$, to create and probe atomic coherence patterns respectively. Stokes and anti-Stokes scattering are singled out on a polarizing beam splitter PBS and observed in the far field on an electron multiplying EM CCD camera. \[fig:setup\]](setup){width="8cm"}
In our case the levels ${|0\rangle}$ and ${|1\rangle}$ between which we create atomic coherence $\rho_{01}(r)$ are hyperfine split levels $5_{2}\mathrm{S}_{1/2}$, F=1 and F=2 respectively. Spatially dependent coherence $\rho_{01}(r)$ is created and read with the use of Raman transitions between this levels. The scattered light is separated from much stronger stimulating lasers by a polarizer and additionally filtered out spectrally by a rubidium-85 filtering cell. The scattered light is registered by an electron multiplying CCD (Hamamatsu) camera sensor which is placed in the far field. The main part of the experimental setup is shown in Fig. \[fig:setup\], further details can be found in [@Chrapkiewicz2012].
![(Color online) Single operational sequence. First rubidium atoms are pumped on D2 line to F=1 ground level ${|0\rangle}$. Then spontaneous Stokes scattering is driven to create spatially varying atomic coherence. Next the coherence is stored in the dark for a time $\tau$ and we fades due to diffusion. Finally the coherence driving anti-Stokes scattering is probed. \[fig:Sequence\]](sequence){width="8cm"}
The operational sequence is shown in Fig. \[fig:Sequence\]. Laser pulses were formed with the use of acousto-optical modulators. We began with optical pumping of rubidium atoms to the ground state F=1. The pump laser operates in resonance on the D2 line. Then we created random patterns of atomic coherence in the Stokes scattering process driven by 2-5 $\mu$s pulses of the write laser detuned from the F = 1$\rightarrow$F’ = 1 D1 transition line by 1 GHz to the red. The Stokes scattering was recorded with a camera for diagnostic purposes. Next the atoms were left to diffuse for time $\tau$. Finally we used 5 $\mu$s long pulses of read laser detuned 1 GHz from the F = 2$\rightarrow$F’ = 2 D1 transition line to the blue to probe blurred atomic coherence $\rho_{01}(\mathbf{r})$ and record anti-Stokes scattering intensity in the far field $I_{\mathrm{AS}}(\theta)$.
Beam diameters $1/e^{2}$ and powers for the Stokes and anti-Stokes drive laser were 5 mm, 4 mm and 16 mW, 7 mW respectively. They were chosen in order to achieve both good resolution as discussed above Eq. and a sufficient scattering rate [@Chrapkiewicz2012].
We repeated the create-store-read sequence multiple times and recorded random patterns of anti-Stokes scattering, changing the storage time $\tau$.
We used four rubidium-87 cells with different buffer gases. Those were: neon at the pressure of 2 Torr, krypton at 0.5 Torr and 1 Torr and xenon at 1 Torr. All cells were 10 cm long, 2.5 cm diameter cylinders made by Precision glassblowing. The longest times for anti-Stokes scattering observations were ca. $\tau=$50 $\mu$s, which corresponded to an RMS atomic displacement of about 1 mm, far less than the cell size.
The cells temperature was stabilized at ca. 70$^{\circ}$ C, which corresponded to an optical depth of 40 and a concentration of rubidium atoms $n=10^{12}\ \mathrm{cm^{-3}}$. Cells were heated with bifilar windings, but the heating current was interrupted for the time of impulse sequence inducing Raman scattering. The cells were placed inside a double magnetic shield. In the independent measurements we checked if the quality of our shielding is good and we estimated the decoherence rate due to imperfect shielding to be less than 1 kHz.
For each diffusion time $\tau$ we averaged 500 images of anti-Stokes scattering each time obtaining smooth symmetric profiles. We subtracted the averaged background. Given that the most important thing for us is the intensity as a function of azimuth angle $\theta$, we also carried out a polar averaging around the laser beam direction increasing signal to noise ratio. The result of the measurements was average scattering intensity $\langle I_{\mathrm{AS}}(\theta,\tau)\rangle$ as a function of angle $\theta$ and storage time $\tau$.
Results {#sec:results}
=======
![(Color online) Average number of photons per pixel $\langle I_{\mathrm{AS}}(\theta,\tau)\rangle$ in the anti-Stokes scattering process as a function of the storage time $\tau$ and angle of observation $\theta$. Data taken for a cell with 1 Torr xenon. The decay is conspicuously faster for higher angles of scattering. The contour lines show 1200, 800, 400, 50, 10 photons per pixel.\[fig:ASMap\] ](ASDecayMap){width="8cm"}
In Fig. \[fig:ASMap\] we present a typical map of the averaged angular profiles of anti-Stokes scattering in cell with xenon vs free depolarization time $\langle I_{\mathrm{AS}}(\theta,\tau)\rangle$. The results are given as a function of the scattering angle $\theta$, which is proportional to the wave vector of the corresponding Fourier components of atomic coherence patterns $\theta=K\lambda/2\pi$.
![(Color online) Average number of photons per pixel $\langle I_{\mathrm{AS}}(\theta,\tau)\rangle$ as a function of the storage time $\tau$ with exponential decay fits observed for four different angles $\theta$. The decays are visibly faster for higher scattering angles. The data corresponds to horizontal dashed lines in Fig. \[fig:ASMap\]. \[fig:ASDecay\]](ASDecayPlot){width="8cm"}
In Fig. \[fig:ASDecay\] we plot an average number of photons as a function of the storage time $\tau$ at a few angles $\theta$ marked with dashed lines in Fig. \[fig:ASMap\]. The decay rate is faster for higher angles of scattering. The error bars in Fig. \[fig:ASDecay\] correspond to 1$\sigma$ uncertainty and were calculated from the full statistics of camera counts. We fit exponential decays to the data taking the error bars into account. It is worth underlining that the data fits well to the curve at each scattering angle. As shown in Fig. \[fig:ASDecay\], for high scattering angles $\theta$ the average signal $\langle I_{\mathrm{AS}}(\theta,\tau)\rangle$ is at the level of one photon per shot which considerably increases the uncertainty of $\gamma$ .
Diffusion coefficients
----------------------
{width="15cm"}
Having obtained an exponential fits for decays in all directions, we can analyze decays rates $\gamma_{\mathrm{obs}}(\theta)$ as a function of the deflection angle $\theta$. In Fig. \[fig:XeNeKr\](a) we present fitted decay rates for measurements in cells filled with 1 Torr of xenon as well as with 2 Torr of neon . In Fig. \[fig:XeNeKr\] (b) we give measured data for cells with krypton at two pressure values: 0.5 Torr and 1 Torr.
As expected, the data fits to the quadratic dependence $\gamma_{\mathrm{obs}}(\theta)=\gamma_{0,\mathrm{obs}}+D\theta^{2}(2\pi/\lambda)^{2}$. Now our $1\sigma$ error bars correspond to respective confidence bound from the exponential fit. Note that for high decay rate values the uncertainties are large because they correspond to small and noisy signals.
The measurements described above were repeated several times in order to make sure the results were reliable and repeatable. We changed the amount of the light generated by altering pulse duration of the writing laser and by repeating measurements at different temperatures. The spread of diffusion coefficients obtained was about 5-12%, depending on the cell. We attribute this spread to the beam wander, laser power and frequency instability, and the drift of the temperature of the cell during each measurement sequence.
Buffer gas Pressure [\[]{}Torr[\]]{} $\gamma_{0,\mathrm{obs}}$ [\[]{}kHz[\]]{}
------------ --------------------------- ------ ------------------------------------------- ---- ----
Ne 2 91 $\pm$ 11 38
Kr 0.5 136 $\pm$ 9 71
Kr 1 57.5 $\pm$ 3 28
Xe 1 52 $\pm$ 3 33
: Fit parameters for measured cells. Temperature: 70$^{\circ}$ C . \[tab:Dgamma\]
In Tab. \[tab:Dgamma\] we summarize fitted values from the measured data from charts in Fig. \[fig:XeNeKr\]. Note, that for measurements in krypton the ratio of the obtained diffusion coefficients is close to the nominal pressure values.
To further verify the accuracy of our results, we carried out reference measurements for krypton at 1 Torr using write and read beams reduced ca. 3 times, so that their diameters $1/e^{2}$ were 1.6 mm i 1.4 mm respectively. This time the measurement was definitely less accurate due to the spread of wave vectors of the read beam and due to aberrations in the imaging system. The diffusion coefficient measured lay within the range of $40\ \mathrm{cm^{2}/s}$ to $65\ \mathrm{cm^{2}/s}$, which is consistent with other results .
Finally let us note, that the observed decay rate at $K=0$, $\gamma_{0,\mathrm{obs}}$ summarized in Tab. \[tab:Dgamma\] is dominated by excessive contribution due to the finite read beam size, $2D\sigma^{2}$ in Eq. . From respective collisional cross sections [@Franz1965] we estimate ca. $1.5$ kHz of decay rate due to Rb-Rb collisions and contributions less than 150 Hz from Rb – buffer gas collisions.
Normalized diffusion coefficients
---------------------------------
Buffer gas $D_{0}\mathrm{[cm^{2}}/\mathrm{s}]$ – this paper $D_{0}\mathrm{[cm^{2}}/\mathrm{s}]$ - previous results
------------ -------------------------------------------------- ---------------------------------------------------------------------------------------------
Ne 0.20$\pm$0.02 0.11 [@Shuker2008], 0.18 [@Vanier1974], 0.31 [@Franzen1959; @Arditi1964], 0.48 [@Franz1965]
Kr 0.068$\pm0.006$ 0.1 [@Bouchiat1972] 0.04 [@Higginbottom2012]
Xe 0.057$\pm$0.007 No experimental data
To compare our results with literature, we normalized the results to the standard conditions, i.e. at atmospheric pressure and at 0$^{\circ}$ C using the standard approximate formula [@Happer1972], however strict scaling in a wide range of temperatures and pressures would involve Chapman-Enskog theory [@Chapman1970]
$$D_{0}=D\Big(\frac{P}{760\ \textrm{Torr}}\Big)\sqrt{\frac{T_{0}}{T}},$$
where $P$ stands for the gas pressure at $T_{0}=0{}^{\circ}$C, and $T$ is the temperature upon measurement. The diffusion coefficient will scale as $T^{1/2}$ according to [@Hogervorst1971].
The temperature is known with good accuracy and the main error in determining $D_{0}$ results from inaccuracy of gas pressure in cells, specified by the manufacturer not to be worse than 10%, and from the spread of the measured $D$ values. The normalized diffusion coefficients are listed in Tab. \[tab:NormalizedD\] together with the data published previously.
The results we obtained do not differ from those obtained beforehand which – as we can see from Tab. \[tab:NormalizedD\]– were characterized by a noticeable spread.
Conclusions
===========
We have demonstrated a novel method for measuring diffusion coefficients of atoms tailored to atomic memory applications. It should be emphasized that the method allows for singling out contribution of the diffusion in any given cell without prior knowledge of other decoherence mechanisms. It is based on creation of spatially varying atomic coherence fields, letting them diffuse in the dark and probing them. The fields comprise various spatial periodicities evolving concurrently, created and probed with Raman scattering. Due to the diffusion distinct components decay at different rates. All other significant sources of decoherence are homogenous and contribute equally to the decay of all components. Therefore we can extract diffusion coefficients by measuring decay rates for different periodicities of different components of spatially varying atomic coherence. Distinct components are conveniently mapped on different angles of the scattered light which enables observations with a camera in the far field.
Our method does not require any setup alterations within the measurement. This leads to quite direct determination of the diffusion coefficient founded on a basic time and angle calibrations of the experimental setup.
We have made sure that the method is accurate and repeatable. The results are based on multiple independent measurements for a number of various periodicities which give almost the same values. We also checked that varying laser beam widths and detunings does not affect the final result and that it scales properly with gas pressure.
We suppose that our method could be incorporated into experiments in which diffusion is the limiting factor, such as EIT, quantum memories including gradient echo memory or collective Raman scattering, by relatively straightforward modifications, such as adding a pump laser and a camera in the far field. Other technical requirements are typically fulfilled since these experiments also rely on using lasers of a few MHz frequency stability and magnetic shielding.
We have measured diffusion coefficients of rubidium in neon, krypton and xenon. Reliable values of diffusion coefficients in these gases facilitate setup design and data interpretation in the multimode quantum storage experiments. Moreover we recommend the use of xenon as a buffer gas in case of Raman interaction and, to the best of our knowledge, we provide the first experimental value of diffusion coefficient in this gas. We believe that this value can also help develop experiments with hyperpolarized xenon.
Acknowledgments
===============
We acknowledge the generous support from Konrad Banaszek and Rafał Demkowicz-Dobrzański. This work was supported by the Foundation for Polish Science TEAM project, EU European Regional Development Fund and FP7 FET project Q-ESSENCE (Contract No. 248095), National Science Centre grant no. DEC-2011/03/D/ST2/01941 and by Polish NCBiR under the ERA-NET CHIST-ERA project QUASAR.
[35]{}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 (, , ) p.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The 2dF Galaxy Redshift Survey (2dFGRS) is designed to measure redshifts for approximately 250000 galaxies. This paper describes the survey design, the spectroscopic observations, the redshift measurements and the survey database. The 2dFGRS uses the 2dF multi-fibre spectrograph on the Anglo-Australian Telescope, which is capable of observing 400 objects simultaneously over a 2 diameter field. The source catalogue for the survey is a revised and extended version of the APM galaxy catalogue, and the targets are galaxies with extinction-corrected magnitudes brighter than =19.45. The main survey regions are two declination strips, one in the southern Galactic hemisphere spanning 80$\degr$$\times$15$\degr$ around the SGP, and the other in the northern Galactic hemisphere spanning 75$\degr$$\times$10$\degr$ along the celestial equator; in addition, there are 99 fields spread over the southern Galactic cap. The survey covers 2000deg$^2$ and has a median depth of $\bar{z}$=0.11. Adaptive tiling is used to give a highly uniform sampling rate of 93% over the whole survey region. Redshifts are measured from spectra covering 3600–8000Å at a two-pixel resolution of 9.0Å and a median $S/N$ of 13. All redshift identifications are visually checked and assigned a quality parameter Q in the range 1–5; Q$\ge$3 redshifts are 98.4% reliable and have an rms uncertainty of 85. The overall redshift completeness for Q$\ge$3 redshifts is 91.8%, but this varies with magnitude from 99% for the brightest galaxies to 90% for objects at the survey limit. The 2dFGRS database is available on the WWW at http://www.mso.anu.edu.au/2dFGRS.'
author:
- |
\
$^1$Research School of Astronomy & Astrophysics, The Australian National University, Weston Creek, ACT 2611, Australia\
$^2$Department of Physics, University of Oxford, Keble Road, Oxford OX1 3RH, UK\
$^3$School of Physics & Astronomy, University of Nottingham, Nottingham NG7 2RD, UK\
$^4$School of Physics and Astronomy, University of St Andrews, North Haugh, St Andrews, Fife, KY6 9SS, UK\
$^5$Department of Physics, University of Durham, South Road, Durham DH1 3LE, UK\
$^6$Anglo-Australian Observatory, P.O. Box 296, Epping, NSW 2121, Australia\
$^7$Astrophysics Research Institute, Liverpool John Moores University, Twelve Quays House, Birkenhead, L14 1LD, UK\
$^8$Department of Astrophysics, University of New South Wales, Sydney, NSW 2052, Australia\
$^9$Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK\
$^{10}$Department of Astronomy, California Institute of Technology, Pasadena, CA 91125, USA\
$^{11}$Department of Physics & Astronomy, Johns Hopkins University, Baltimore, MD 21218-2686, USA\
$^{12}$Department of Physics, University of Leeds, Woodhouse Lane, Leeds, LS2 9JT, UK\
$^{13}$Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh EH9 3HJ, UK\
date: 'Accepted —. Received —; in original form —.'
title: 'The 2dF Galaxy Redshift Survey: Spectra and redshifts'
---
surveys — galaxies: clustering — galaxies: distances and redshifts — cosmology: observations — cosmology: large scale structure of universe
\
INTRODUCTION {#sec:introduction}
============
The 2dF Galaxy Redshift Survey is designed to measure redshifts for approximately 250000 galaxies in order to achieve an order of magnitude improvement on previous redshift surveys, and provide a detailed and representative picture of the galaxy population and its large-scale structure in the nearby universe. The main goals of the 2dFGRS are:
1. To measure the galaxy power spectrum $P(k)$ on scales up to a few hundred Mpc, filling the gap between the small scales where $P(k)$ is known from previous galaxy redshift surveys and the largest scales where $P(k)$ is well-determined by observations of the cosmic microwave background (CMB) anisotropies. Particular goals are to determine the scale of the turnover in the power spectrum and to observe in the galaxy distribution the acoustic peaks detected in the CMB power spectrum (Percival 2001).
2. To measure the redshift-space distortion of the large-scale clustering that results from the peculiar velocity field produced by the mass distribution (Peacock 2001). This distortion depends on both the mass density parameter $\Omega$ and the bias factor $b$ of the galaxy distribution with respect to the mass distribution, constraining the combination $\beta=\Omega^{0.6}/b$.
3. To measure higher-order clustering statistics of the galaxy distribution in order to: (a) determine the bias parameter $b$, revealing the relationship between the distributions of mass and light and yielding a direct measure of $\Omega$; (b) establish whether the galaxy distribution on large scales is a Gaussian random field, as predicted by most inflationary models of the early universe; and (c) investigate the non-linear growth of clustering in the small-scale galaxy distribution.
4. To fully and precisely characterise the galaxy population in terms of the distributions of fundamental properties such as luminosity, surface brightness, spectral type and star-formation rate (Folkes 1999; Cole 2001; Madgwick 2001a; Cross 2001).
5. To quantify the relationships between the internal properties of galaxies (such as luminosity, spectral type and star-formation rate) and their external environment (the local density of galaxies and the surrounding large-scale structure) in order to constrain models of galaxy formation and evolution (Norberg 2001a).
6. To investigate the properties of galaxy groups and clusters, not just from existing cluster catalogues (De Propris 2001), but by defining a large, homogeneous sample of groups and clusters in redshift space, avoiding the problems of cluster catalogues defined in projection and allowing detailed study of the mass distributions and dynamical evolution in the most massive bound structures in the universe.
7. To provide a massive spectroscopic database for use in conjunction with other surveys, for finding rare and interesting types of object, and as a source for a wide variety of follow-up observations (Cole 2001; Sadler 2001, Magliocchetti 2001).
This paper provides an overview of the survey and detailed description of the survey observations. The layout of the papers is as follows: §2 summarises the main capabilities of the 2dF multi-fibre spectrograph; §3 describes the input source catalogue for the survey; §4 discusses the survey design, including the areas of sky covered by the survey and the tiling of the survey fields; §5 outlines the algorithm used to assign fibres to targets, and its uniformity and completeness; §6 describes the spectroscopic data obtained for the survey and the data reduction methods; §7 deals with the estimation of the redshifts and various internal and external checks on their reliability and precision; §8 describes the survey masks, which encapsulate the coverage, magnitude limits and redshift completeness of the survey; §9 outlines the main components and features of the survey database; §10 summarises some of the main results emerging from the survey; and §11 provides conclusions.
THE 2dF SPECTROGRAPH {#sec:instrument}
====================
The survey is designed around the Two-degree Field (2dF) multi-fibre spectrograph on the Anglo-Australian Telescope, which is capable of observing up to 400 objects simultaneously over a 2 diameter field of view. Here we summarise the aspects of the instrument relevant to the 2dFGRS; a full description is provided by Lewis (2001) and the 2dF User Manual (http://www.aao.gov.au/2df). The four main components of 2dF are the corrector optics, the fibre positioner, the fibres and the spectrographs.
The 2dF corrector optics are a four-element lens assembly which gives 1 images over a 2.1 diameter flat field of view at the prime focus of the AAT. The corrector incorporates an atmospheric dispersion compensator (ADC) which corrects for atmospheric dispersion at zenith distances less than about 70. The most significant remaining image degradation is a chromatic variation in distortion, which has the effect of dispersing images in the radial direction. This effect is largest halfway out from the field centre, where the wavelength range 0.35–1.0$\mu$m is dispersed over 2 radially. The radial distortion introduced by the corrector gives an image scale that varies over the field of view, from 15.5/mm at the centre to 14.2/mm at the edge; the corresponding change in focal ratio is from f/3.4 to f/3.7.
The 2dF fibre positioner X-Y robot takes about 6–7 seconds on average to position one fibre. Since approximately 550–580 fibre moves are required to re-configure a typical 400-fibre field, this means that a full re-configuration takes about 60–65 minutes. The actual configuration time varies depending on the number of fibres used and the complexity of the field. To avoid dead-time there are two field plates, each with 400 fibres. While one field plate is placed at the focal plane for observing, the other is being re-configured by the fibre positioner. The positions of the field plates are reversed by tumbling them about their horizontal axis.
The fibres are have 140$\mu$m diameter cores, corresponding to 2.16 at the field centre and 1.99 at the field edge. There are 400 fibres on each field plate. The internal precision with which the fibres are positioned is 11$\mu$m (0.16) on average, with no fibres outside 20$\mu$m (0.3). The fibres terminate in magnetic ‘buttons’ that attach to the steel field plates. Each button has a right-angle prism which directs the light from the focal plane into the fibre. To prevent fibre buttons coming into direct contact and protect the fragile prisms, the fibres cannot be placed too close together. The absolute minimum fibre separation is 800$\mu$m (approximately 12), and the required separation is generally larger (approximately 30), as it depends on the detailed geometry of the buttons and their relative orientations.
Half the fibres on each field plate go to one of the two identical spectrographs. Each spectrograph has an f/3.15 off-axis Maksutov collimator feeding a 150mm collimated beam to the grating and then, at an angle of 40, to an f/1.2 wide-field Schmidt camera. The detectors are Tektronix CCDs with 1024$\times$1024 24$\mu$m pixels. The 200 fibre spectra imaged onto each CCD are separated by approximately 5 pixels. The 2dFGRS used the 300B gratings, which are blazed at 4200Åand give a dispersion of 178.8Å/mm (4.3Å/pixel). For a typical spectrograph focus of 2.1 pixels (FWHM), this corresponds to a FWHM spectral resolution of 9.0Å. The 2dFGRS observations used a central wavelength around 5800Å, and covered the approximate range 3600–8000Å. Flexure is less than 0.2 pixel/hour (i.e. less than 40 over a typical integration time).
The overall system efficiency (source to detector) of 2dF with the 300B gratings used in the 2dFGRS is 2.8% at 4400Å, 4.3% at 5500Å and 4.7% at 7000Å. These figures were obtained from measurements of photometric standard stars, corrected to nominal 1 seeing.
SOURCE CATALOGUE {#sec:source}
================
The source catalogue for the survey (Maddox 2001, in preparation) is a revised and extended version of the APM galaxy catalogue (Maddox 1990a,b,c;1996). This catalogue is based on Automated Plate Measuring machine (APM) scans of 390 plates from the UK Schmidt Telescope (UKST) Southern Sky Survey. The extended version of the APM catalogue includes over 5 million galaxies down to =20.5 in both north and south Galactic hemispheres over a region of almost 10$^4$deg$^2$ (bounded approximately by declination $\delta$$\leq$+3 and Galactic latitude $b$$\ge$30). Small regions around bright stars, satellite trails and plate flaws are excluded; these are accounted for by the survey mask (see §\[ssec:maglimmask\]).
The magnitude system for the Southern Sky Survey is defined by the response of Kodak IIIaJ emulsion in combination with a GG395 filter. It is zeropointed to Vega—i.e. is equal to Johnson $B$ for an object with zero colour in the Johnson–Cousins system. The colour equation is normally taken to be $${\mbox{b$_{\rm\scriptscriptstyle J}$}}= B - 0.28(B-V) ~,$$ following Blair & Gilmore (1982). A larger coefficient ($-0.35$) has been suggested by Metcalfe (1995), but we measure $-0.27 \pm
0.02$ in comparison with the ESO Imaging Survey (Arnouts 2001), and we therefore retain the usual value of $-0.28$. The photometry of the catalogue is calibrated with numerous CCD sequences and, for galaxies with =17–19.45, has a 68% spread of approximately 0.15mag, but with a non-Gaussian tail to the error distribution. We emphasise that the calibration is to [*total*]{} CCD photometry, which absorbs any remaining correction to the thresholded APM magnitudes.
The star-galaxy separation is as described in Maddox (1990b), and the locus dividing stars and galaxies was chosen to exclude as few compact galaxies as possible while keeping the contamination of the galaxy sample by stars to about 5%. Spectroscopic identifications of the survey objects (see §\[sec:redshifts\]), show that the stellar contamination is in fact 6%.
The source catalogue is incomplete at all magnitudes due to various effects, including the explicit exclusion of objects classified by the APM as merged images, the misclassification of some galaxies as stars, and the non-detection (or misclassification as noise) of some low surface brightness objects. This incompleteness has been studied in comparisons with deeper wide-area CCD photometry by Pimbblet (2001) and Cross & Driver (2001, in preparation). The overall level of incompleteness is 10–15% and varies slightly with apparent magnitude, being largest for the brightest and faintest objects. The main classes of objects that are excluded are: (i) merged galaxy images that are explicitly excluded from the 2dFGRS source catalogue (about 60% of the missing objects); (ii) large galaxies that are resolved into components that are classified as stellar, merged or noise objects (20%); (iii) compact normal galaxies that are detected but classified as stars (15%); and (iv) low surface brightness galaxies that are either not detected or classified as noise objects (5%). Thus the main cause of incompleteness is misclassification of objects rather than their non-detection.
The target galaxies for the 2dFGRS were selected to have extinction-corrected magnitudes brighter than =19.45. Since the targets were selected, improvements to the photometric calibrations and revised extinction corrections have resulted in slight variations to the magnitude limit over the survey regions—this is precisely quantified by the magnitude limit mask for the survey (see §\[ssec:maglimmask\]). The extinction is taken to be $A_{b_J}$==4.035$E$($B$$-$$V$), where the coefficient, and the reddening $E$($B$$-$$V$) as a function of position, come from Schlegel (1998). The limit of =19.45 was chosen because: (i) The surface density of galaxies at =19.45 (approximately 165) is sufficiently larger than the surface density of 2dF fibres on the sky (127) to allow efficient use of all fibres—few fibres are unused even in low-density fields. (ii) The time taken to configure a typical field (60–65min) allows, with overheads, a sufficiently long exposure time to reach the desired signal-to-noise level of $S/N$$>$10 for galaxies with =19.45 even in rather poor conditions. This limiting magnitude corresponds to a median redshift for the survey of about $\bar{z}$=0.1, so that the 2dFGRS is essentially a survey of the [*local*]{} universe.
SURVEY DESIGN {#sec:design}
=============
Survey areas {#sec:areas}
------------
[ ]{}
The areas of the sky covered by the survey were chosen so as to satisfy a number of different requirements. The first goal was to cover as large a volume as possible, in order to closely approach a statistically representative sample of the universe on the largest possible scales. The second was to obtain near-complete sampling down to the survey limit in order to have the finest possible resolution of structure on small scales. The third requirement was to match the sample to the observational capabilities of the 2dF instrument in order to achieve high efficiency. The adopted geometry is an effective compromise between these requirements.
The survey consists of two separate declination strips of overlapping 2 fields plus 99 scattered ‘random’ 2 fields. One strip (the SGP strip) is in the southern Galactic hemisphere and covers approximately 80$\degr$$\times$15$\degr$ centred close to the South Galactic Pole ($21{\mbox{$^{\rm h}$}}40{\mbox{$^{\rm m}$}}<\alpha<03{\mbox{$^{\rm h}$}}40{\mbox{$^{\rm m}$}}$, $-37.5\degr<\delta<-22.5\degr$). The other strip (the NGP strip) is in the northern Galactic hemisphere and covers 75$\degr$$\times$10$\degr$ ($09{\mbox{$^{\rm h}$}}50{\mbox{$^{\rm m}$}}<\alpha<14{\mbox{$^{\rm h}$}}50{\mbox{$^{\rm m}$}}$, $-7.5\degr<\delta<+2.5\degr$). The 99 ‘random’ fields are chosen from the low-extinction region of the APM catalogue in the southern Galactic hemisphere outside the survey strip (the mean extinction over each field is required to be less than 0.2mag—see Figure 2 of Efstathiou & Moody (2001)). The fields are chosen pseudo-randomly within this region, except that the field centres are at least 3 apart. A map of the survey fields on the sky is shown in Figure \[fig:skymap\]; the locations of the fields with respect to the extinction map derived from Schlegel (1998) are shown in Figure \[fig:dustmap\]. All the survey fields lie at Galactic latitudes greater than $|b|$=30, and the whole of the SGP strip and most of the NGP strip and the random fields lie at Galactic latitudes greater than $|b|$=45.
The distribution of extinction corrections as a function of Galactic latitude, and the fraction of corrections larger than a given value, are shown in Figure \[fig:extndist\]. Overall, the median correction is 0.07mag, 90% are less than 0.16mag, and 99% are less than 0.26mag; the corresponding quantiles in the NGP are (0.12,0.19,0.28)mag, in the SGP (0.05,0.07,0.11)mag and in the random fields (0.07,0.13,0.30)mag.
The 2dFGRS target sample of galaxies contains [193550]{} galaxies in the SGP strip, [139144]{} galaxies in the NGP strip, and [57019]{} galaxies in the random fields. This gives a total of [389713]{} possible targets, significantly more than the survey goal of 250000 galaxies. Survey observations of the NGP and SGP strips are proceeding outwards in declination from the centre of each strip towards this goal. Note that the total number of galaxies listed in the survey source catalogue (and the survey database) is [467214]{}, which is larger than the number of possible survey targets because the source catalogues for the NGP and SGP strips conservatively includes galaxies fainter than the spectroscopic survey magnitude limit, down to =19.6.
At the median redshift of the survey ($\bar{z}$=0.11) the SGP strip extends over 400$\times$75, and the NGP strip over 375$\times$50. Out to the effective limit of the survey at $z$$\approx$0.3, the strips contain a volume of 1.2$\times$10$^8$(for $\Omega_m$=0.3, $\Omega_{\Lambda}$=0.7); the volume sparsely sampled including the random fields is between two and three times larger.
Tiling the survey {#ssec:tiling}
-----------------
The survey limit of =19.45 was chosen, in part, because it gives a good match between the surface density of galaxies and the surface density of 2dF fibres. Due to clustering, however, the number of galaxies in a given field varies considerably. The rms variation in the number of galaxies per randomly-placed 2 field is 140 at =19.5, and is largely independent of the choice of magnitude limit over the range considered here. To make efficient use of 2dF we therefore require an algorithm for tiling the sky with 2 fields that allows us to cover the survey area at a high, and nearly uniform, sampling rate with the minimum number of 2dF fields.
So long as the sampling of the source catalogue is not biased in any way that depends on the photometric or spectroscopic properties of the galaxies, we can always use the source catalogue to accurately determine the sampling rate as a function of position (see §\[sec:mask\]). The sampling can then be accounted for in any analysis. However to keep such corrections to a minimum, considerable effort has been invested in making the sampling as complete and uniform as possible.
There are a number of possible approaches to laying down target field centres. The simplest is to adopt a uniform grid of equally spaced centres and then either randomly sample each field with the number of available fibres or observe each field several times until all the galaxies have been observed. The second of these options is clearly inefficient, as it will give rise to a large number of fields being observed with significantly less than the full complement of fibres, while the first is undesirable as it gives a different sampling factor for each field. A more efficient solution is to use an adaptive tiling strategy, where we allow each field centre to drift from the regular grid so that we maximise the number of targets that are assigned to each field, subject to the constraint that the number of targets assigned to any one field should not exceed the number of available fibres, $N_f$.
We begin with a uniform grid with field centres equally spaced by $s=\sqrt{3}\degr$ in right ascension along rows 1.5 apart in declination. For each galaxy in the survey we then determine how many fields it lies within, and assign each field a weight $w_i$, where $$\begin{aligned}
w_i = & 0 & {\rm if}~N_i \ge N_f \\
= & 1 - \frac{N_i}{N_f} & {\rm if}~N_i < N_f \nonumber\end{aligned}$$ and $N_i$ is the number of galaxies in this field. These weights are normalised such that $\sum_1^{N_t}w_i = 1$, where $N_t$ is the number of fields that this galaxy could be assigned to. The galaxy in question is then randomly assigned to one of the fields using these weights, unless all the fields are already filled, in which case it is assigned to the first field in which it was found. Once all the field occupancies have been determined in this way we move each field in right ascension by an amount $$\begin{aligned}
\delta\alpha_i = & \Delta\alpha+0.05s\frac{N_i}{N_f} & {\rm if}~(N_i \ge N_f)\\
= & \Delta\alpha-s^\prime\left(1-\frac{N_i}{N_f}\right) & {\rm if}~(N_i<N_f) \nonumber\end{aligned}$$ where $s$ is the maximum allowed separation between adjacent fields, $s^\prime$ is the current distance to the neighbouring field centre, and $\Delta\alpha$ is the cumulative shift that has currently been applied to this row. New fields are added at the end of each row if the total length of the row has contracted enough to exclude any galaxies at the trailing edge. We found that using a fixed separation of 1.5 in declination, and adjusting the tile positions in right ascension only, provided sufficient flexibility to achieve uniform high completeness without a large increase in the total number of fields and without leaving gaps in the sky coverage.
In practice, it is found that the above prescription requires a further modification to account for the position of each object within the field. This additional constraint arises because of the physical restrictions on the positioning of individual fibres, both in terms of their deviation from the radial angle and their extension from the parked position at the edge of the field. We apply this constraint by dividing each field into 36 sub-fields and restricting the number of targets that can be assigned to each sub-field to 16. This is larger than the number of fibres whose park positions fall within the arc of the sector, but allows for the fact that the area of the field that can be reached by each fibre is increasing with the fibre extension. Without this extra constraint the algorithm tends to place large clusters in the overlapping areas of neighbouring fields where, although there are more available fibres because of the overlap, it rapidly becomes impossible to use these fibres because of the high density of targets close to the edge of each field. Limiting the number of targets within each sector effectively removes this problem, and so increases the uniformity of the survey completeness.
The adaptive tiling algorithm also needed to cope with the requirement that the galaxy redshift survey be merged with the concurrent survey of QSO candidates (the 2dF QSO Redshift Survey: Boyle 2000; Croom 2001). This results in a higher surface density of targets in the region of overlap, for which we compensate by reducing the separation in declination of the tiling strips in the QSO survey regions to 75% of the original value (i.e. to 1.125). As the QSO survey occupies the central declination strip of our SGP survey region this results in a 3-4-3 arrangement of tiling strips over the three rows of UKST fields used in the SGP survey. A similar consideration applied to the NGP survey region gives a 3-4 arrangement if we consider the full areas of the two rows of UKST fields used.
With these additional modifications, our tiling algorithm was able to achieve an overall sampling rate of 99.8%, dropping to 93% after the assignment of fibres to targets has been made (see §\[ssec:fibrecomplete\]). On average, fewer than 5% of the fibres in each field were unused. For comparison, the mean sampling rate that would be achieved for randomly distributed fields is 68%, with an rms variation of 15%.
FIBRE ASSIGNMENT {#sec:fibres}
================
The tiling procedure described above fixes the location of the 2dF target fields on the sky and provides a first pass priority scheme (based on the random field assignments) for use in assigning targets to fibres. However, the mechanical constraints of 2dF imply that we cannot usually allocate fibres to all the targets assigned to a particular field, and so we consider all targets that lie within each field boundary. Our tiling scheme implies that many of the survey targets are found on more than one field, so we adopt a priority scheme as follows: targets which are unique to the field in question are assigned the highest priority, then targets which are assigned to the field, but which can also be reached from neighbouring fields, then targets which fall within the field, but which are assigned to a different, overlapping field. The priorities of the QSO targets are increased one step to ensure that we do not imprint the strong clustering pattern of the galaxy distribution on the weak clustering expected from the QSO sample.
Single-field assignments {#ssec:singlefield}
------------------------
The allocation procedure for a single field can be divided into two distinct steps: an initial allocation pass and a fibre swapping pass. These two steps are repeated for each range of decreasing priorities for targets that have been assigned to this field. An attempt is then made to reduce the complexity of the configuration by identifying pairs of allocated fibres which are crossed, and which can safely be exchanged. The allocation and uncrossing steps are then repeated to allow consideration of the lower priority targets which were not assigned to this field. The details of these steps are as follows:
(1) [*Initial allocation pass.*]{} For each target, $t$, of a given priority we determine the set, ${\cal F}_t$, of all fibres that could be allocated to that target in the absence of any other allocations. We then determine the set, ${\cal T}_t$, of all targets which fall within the sector defined by the target and the two most extreme fibres in ${\cal F}_t$. We then use the difference $N({\cal T}_t)-N({\cal F}_t)$ between the number of targets in ${\cal T}_t$ and the number of fibres in ${\cal F}_t$ to determine which targets are hardest to access and hence should be given the highest priority: targets within the sector defined by the target with the largest value of $N({\cal T}_t)-N({\cal
F}_t)$ are allocated first. This procedure is then repeated for targets of successively lower priorities.
(2) [*Fibre-swapping pass.*]{} For each remaining unallocated target, $u$, we now consider each fibre in ${\cal T}_u$. For each fibre in this set that could still be allocated to $u$ given the current configuration we make a tentative re-allocation of this fibre from its initial assignment, $i$, to target $u$ and check to see if any unallocated fibres can be assigned to $i$. If no unallocated fibres can be assigned we repeat this process recursively until one of the following conditions have been satisfied: either (i) a previously unallocated fibre is allocated (implying that $u$ has been allocated and that all previously allocated targets remain allocated); or (ii) the search exceeds a depth of 10 iterations. In the latter case, if at any point in the search a fibre has been moved from a low-priority target such that a target of higher priority has been allocated then the search is unpacked to this point; if no such trade-off was found then the search is unpacked to the original configuration.
(3) [*Fibre-uncrossing pass.*]{} Each pair of fibres in the configuration is tested to see if the fibres cross. For each crossed pair an attempt is made to swap the allocations of fibres to targets. This process is iterated until no further swaps can be made. The net effect of this procedure is that the final configuration supplied to the positioner is simplified so that the transition to the [*following*]{} configuration will require fewer fibres to be moved twice. This simplification effectively reduces the time required for each re-configuration, and is achieved without constraint on the allocation.
Field-overlap assignments {#ssec:fieldoverlap}
-------------------------
While the above procedure provides an optimised method of allocating as many fibres as possible to a given individual field, it does not make optimal use of the flexibility provided by the overlapping fields. We therefore ran a second pass of the configuration algorithm which considered each pair of overlapping fields as an 800 fibre problem in which all objects found within the overlap area are matched up, so that an object which is allocated in one field can be allocated to the second field to allow acquisition of unallocated targets in the first field. Compared to the results of treating the two overlapping fields as separate configuration problems, this procedure proves more effective at reducing the overall incompleteness due to the instrumental constraints. Since the tiling algorithm tends to place local over-densities close to the edges of tiles, this procedure also serves to homogenise the distribution of incompleteness as a function of position within a field.
A significant fraction of this incompleteness arises due to close pairs of objects which lie within the unique regions of individual fields. The separation at which this occurs is a strong function of the relative orientation of the two fibre buttons (see Lewis 2001). The upper limit on the value of this separation is 11mm in the focal plane (corresponding to 2.2), while the mean separation of close pairs of objects for which one target cannot be assigned is $\sim$30.
Fibre assignment completeness {#ssec:fibrecomplete}
-----------------------------
The results of this procedure are that we are able to allocate fibres to 93% of the source catalogue objects. The distribution of the unallocated objects on the sky are shown in Figure \[fig:unallocsky\]. The most prominent features visible in these distributions are occasional localised clusters of unallocated objects. These are due to over-dense regions where the geometrical packing constraints imposed by the fibre button dimensions mean that it has not been possible to assign every fibre to a target, even though there are enough fibres available as a result of the tiling algorithm. This effect is enhanced by the relative increase in the number of close pairs in strongly clustered regions.
The distribution of these unallocated objects is also shown as a function of position within the 2dF field in Figure \[fig:unallocplate\]. This plot shows a slight overall gradient in the R.A. direction, which is due to the order in which the field overlaps are processed (see §\[ssec:fieldoverlap\]). This apparent gradient is misleading, since the missed objects are plotted according to the field plate coordinates on the plate to which they were assigned, whereas the majority of objects which contribute to this effect are actually missed on more than one 2dF field. The next most prominent feature of Figure \[fig:unallocplate\] is the imprint of the four fibres which are used for field acquisition and guiding. These fibres must be allocated with a higher priority than the targets if the field is to be observed successfully. They prevent the subsequent allocation of target objects lying along their length, and so produce a slight increase in the numbers of unallocated objects along the four cardinal axes of the field, in the regions of the plate where they can be placed.
It should be noted that these figures do not represent the final incompleteness in the survey, since individual target allocations are adjusted immediately prior to observation to account for the actual number of fibres available to 2dF on any given night (typically, about 1% of the fibres are broken at any given time, though they are continually replaced). The overall incompleteness at this stage, before any observations have been made, is 7%. The level of this incompleteness is such that the pattern is dominated by residuals due to clustering within individual two-degree fields, implying that an attempt to correct for this effect would effectively imprint the observed pattern onto the data. We therefore conclude that we have achieved our goal of reducing the mean incompleteness to an acceptably low level, comparable to the level of incompleteness due to the individual uncertainties in galaxy magnitudes and the incompleteness due to not measuring redshifts because of inadequate signal-to-noise. We emphasise that neither the tiling nor the fibre allocation depend on the photometric or spectroscopic properties of the objects, so that we can use the source catalogue to accurately determine the sampling of the survey, as discussed in §\[sec:mask\]).
Visual inspection {#ssec:inspection}
-----------------
Postage stamp images of each target galaxy and each candidate fiducial star were generated from the Digitized Sky Survey[^1] (DSS; Lasker 1998). These DSS images were used to examine the target galaxies for merged images and to remove unsuitable fiducial star candidates.
The APM image classification parameter $k$ (see Maddox 1990b) was used to find target galaxies which might possibly be merged images. Objects with $k$$>$1.15 in the SGP and $k$$>$1.2 in the NGP were analysed by a simple routine which attempts to relocate the fibre onto the local maximum of surface brightness closest to the nominal target position. This is achieved by starting from the central pixel of the DSS image and allowing the position to move in a direction of increasing intensity until a local maximum is found. The position corresponding to this local maximum is then inserted into the 2dF configuration file prior to observation. All objects which are tested in this way are flagged as merged images by changing the second character of the object’s [OBSNAME]{} parameter from [G]{} to [M]{}, even if the final position of the fibre is unchanged. The survey database records both the original source catalogue position (database parameters [RA]{} and [DEC]{}; see §\[ssec:database\])and the actual observed position (database parameters [OBSRA]{} and [OBSDEC]{}); the parameter [MATCH\_DR]{} gives the offset (in arcsec) between these two positions, and may be non-zero for objects flagged as mergers.
The DSS images of candidate fiducial stars are inspected visually to allow removal of the following types of object, which would compromise field acquisition: (i) stars which are merged with another object; (ii) stars which are close to other stars of comparable brightness; (iii) bright stars which are strongly saturated on the sky survey plates; (iv) spurious objects that are fragments of the diffraction spikes or halos of bright stars, or bright regions within large galaxies; (v) asteroids and portions of satellite trails; (vi) spurious objects due to noise.
Final adjustments {#ssec:adjustments}
-----------------
The configuration procedure described above determines the optimal assignment of fibres to targets over the whole survey. This procedure assumes a fixed set of parameters for the transformation of sky-coordinates to plate-coordinates, whereas in practice a slightly different transform is determined each time the instrument is mounted on the telescope. The above procedure also assumes a full complement of fibres, and does not allow for the small but significant attrition rate as fibres are disabled due to optical or mechanical failures. Fibres which are disabled in this way are repaired on a regular basis, but operational constraints require that disabled fibres are replaced in batches, and so there is generally some delay before a particular fibre is replaced.
We account for both the fibre attrition and the effects of changes to the coordinate transform by re-examining each configuration the day before that field is to be observed. All broken or inoperative fibres which had been allocated are deallocated and the targets to which they were assigned are flagged. An attempt is then made to recover each flagged target using the current list of available fibres, according to the same algorithm used for the original allocation. Again, low-priority targets may be lost at this stage in favour of high-priority targets. In practice, we usually find that the majority of the high-priority targets which were flagged can either be recovered or replaced.
A similar check is then made to allow for a further change to the coordinate transform due to the hour angle at which the field is to be observed. This correction is a combination of the effects of atmospheric refraction and flexure and misalignment of the telescope itself. Each configuration is checked for hour angles of $\pm$4 hours, and adjustments are made to ensure that no fibre conflicts will occur within this range. Again, we usually find that all such conflicts can be recovered without loss of target allocations.
After this correction to the configuration, a number of unallocated fibres (at least 10 for each spectrograph) are allocated to blank sky positions. This is done initially using an azimuthally symmetrical grid of positions around the field, but if insufficient fibres can be allocated to this grid, more are allocated to other blank sky regions manually. The position of each allocated sky position is then checked to ensure that it is genuinely blank by examining the surrounding area using the Digitized Sky Survey. Any sky fibres likely to be contaminated by objects visible on the DSS are reallocated to other blank sky regions.
SPECTROSCOPIC DATA {#sec:data}
==================
Observational procedure {#ssec:observing}
-----------------------
A target field is acquired using the four guide fibres, each of which consists of 6 fibres in a hexagonal pattern surrounding a central fibre. The individual fibres are 95$\mu$m (1.4) in diameter and have centre-to-centre separations of 1.8. These guide fibres are positioned on 14$<$$V$$<$15 stars with positions measured from the same UKST plates as the target galaxies. As the blue sky survey plates are often 20–25 years old, we use the more recent red sky survey plates to check for any proper motion, and we also apply a colour cut to reduce the number of nearby stars. On some fields, where red plates are not available, we have used fiducial stars from the USNO A1.0 astrometric catalogue (Monet 1996).
The guide fibres are viewed by the AAT’s standard acquisition and guiding TV system. The position for optimum acquisition of the fiducial stars, and subsequent guiding corrections, are determined by visual inspection of the locations of the stars in the guide fibres on the TV image. Acquisition of a new field typically requires less than 5 minutes.
For each field we first take a multi-fibre flat-field exposure using the quartz lamp in the calibration unit. This flat-field is used to trace the positions of the fibres on the image, to fit the spatial profile of each fibre as a function of wavelength, and to apply a 1-dimensional pixel-to-pixel flat-field correction to the extracted spectra. We next take a wavelength calibration exposure of helium and copper-argon arc lamps. We then typically take three 1100s exposures of the target objects. However the exposure time is varied to suit the conditions: for galaxy-only fields in good conditions, three 900s exposures are sufficient; in poor conditions, especially for fields including targets from the QSO survey, a longer total integration time is used. With the CCD readout time of 60s, the whole series of exposures typically takes just over an hour, so that the observations of one field are finished just as the configuration for the next field is completed.
Data reduction {#ssec:reduction}
--------------
The data are reduced using the 2dF data reduction pipeline software, , a full description of which is given by Bailey (2001; see also http://www.aao.gov.au/2df). The main steps in the process are as follows:
1\. [*Basic image processing.*]{} Each image undergoes the same initial processing, which consists of flagging bad pixels (both saturated pixels and those specified in the bad-pixel mask for each CCD). The bias level is computed as the median of the CCD overscan region and subtracted from the image. The image is then trimmed to remove the overscan region. Finally, a variance array is computed based on the pixel data values and the known properties of the CCDs. The variances are propagated through all subsequent reduction steps.
2\. [*Mapping the spectra.*]{} The next step is to map the locations and shapes of the spectra on the detector. The centroids of the fibre spectra along the slit are determined using the fibre flat field image, and the fibre identifications checked visually. This is necessary because the positioning of the slit block is not precisely reproducible, so that the locations of the fibres on the CCD can change by a few pixels, and occasionally the first or last fibre may fall off the edge of the detector. Visual checking ensures that spectra are correctly identified with fibres and thus with the objects that were observed. Given this starting location, finds the centroids of each fibre spectrum as a function of wavelength and fits a model based on the known optical properties of the spectrograph. For each subsequent frame, this model is adjusted (to account for flexure) by fitting an overall shift and rotation to the positions of the spectra on the detector.
3\. [*Subtracting the background.*]{} The background of scattered light is fitted with an empirically-determined model using the unilluminated portions of the detector. Fibres are so closely packed on the detector that the light level between transmitting fibres does not drop to the background level. Hence the only unilluminated areas are at the edges of the detector and at the positions of dead (broken, non-transmitting) fibres. A small number of fibres are broken at any one time and are continually being replaced. Occasionally the number of broken fibres becomes sufficiently small that it is not possible to reliably fit the background, and this step must be omitted.
4\. [*Fitting the spatial profiles.*]{} The close packing of the fibres on the detector means that their spatial profiles overlap, producing cross-talk between neighbouring spectra. In order to correct for this, it is necessary to know the spatial profile of each fibre spectrum as a function of wavelength. These profiles are determined from the fibre flat field image by performing simultaneous fits of 200 Gaussians to the fibre profiles at each of 20 wavelength ranges. These fits are then interpolated in the wavelength direction with cubic splines to give the spatial profile for each fibre at every wavelength on the detector.
5\. [*Spectrum extraction.*]{} Given the location and shape of each fibre’s spatial profile, optimal extraction consists of a weighted fit of the profiles to recover the amplitude of the spectrum at each wavelength. Because the spectra overlap, a simultaneous fit of the profile amplitudes of each fibre spectrum and its nearest neighbour on either side is carried out. The fit is performed using least-squares with variance weighting. 6. [*Flat-fielding.*]{} The extracted spectra are then ‘flat-fielded’ by dividing by the pixel-to-pixel variations in the extracted spectra from the fibre flat field image. This one-dimensional flat-fielding is used because there is presently no means of generating a full two-dimensional flat-field image.
7\. [*Wavelength calibration and linearisation.*]{} A predicted wavelength for each pixel, based on the nominal central wavelength and the optical model for the spectrograph, is first calculated. Lines in the helium/argon arc spectra are automatically identified by a peak-finding algorithm, then matched to a list of known lines, discarding both unidentified peaks and unmatched lines, as as well as known blends. The relation between the measured line positions and their true wavelengths is fit by a third-order polynomial. This fit iterated up to four times, with the most discrepant line excluded at each iteration. The final fits usually include 21–22 lines over the 4400Å spectral range, and have typical rms residuals of 0.3Å (0.07 pixels). Once the wavelength calibration is determined, the spectrum is re-binned onto a linear wavelength scale using quadratic interpolation.
8\. [*Fibre throughput calibration.*]{} The spectrum of the sky background is measured with at least 8 fibres (and usually more) on each of the two spectrographs. The spectra from the sky fibres are normalised by their mean fluxes and a median sky spectrum is calculated. The relative throughputs of the fibres are then derived from the relative fluxes in the strong sky lines as follows. All object and sky fibres and the median sky are continuum-subtracted using a continuum derived by median smoothing with a 201-pixel box. This removes the continuum leaving only the strong lines in the residual spectrum. A robust least-squares fit to the constant of proportionality between the counts in each pixel of the residual object+sky spectrum and the residual median sky spectrum then yields the relative throughput for each fibre (the sky lines in common to both spectra contribute to the fit, but strong emission lines found only in the object spectrum do not). The individual object and sky spectra are then normalised by dividing by their relative throughputs.
9\. [*Sky subtraction.*]{} The median sky spectrum is re-calculated from the individual normalised sky spectra, and subtracted from each normalised object and sky spectrum. The sky-subtraction precision, measured as the ratio of the total flux in the sky fibres before and after sky-subtraction, is typically 2–3% and rarely worse than 6%. Precise sky-subtraction is dependent on effective scattered light subtraction, wavelength calibration and throughput calibration.
10\. [*Combining spectra.*]{} The fully-reduced spectra from each of the multiple exposures on the target objects (typically three) are optimally combined in a final step. The combination algorithm accounts for overall flux variations between exposures (due to different integration times or observing conditions) and rejects cosmic ray events. The relative weighting of the exposures is computed from the total fluxes in the brightest 5% of the spectra (after they are median-smoothed over a 201-pixel box to eliminate cosmic rays and normalised to the first exposure). The maximum flux weight is unity, and the weights are applied to the individual exposures before they compared for cosmic rays. The cosmic ray rejection algorithm is similar to that of the IRAF [crreject]{} package. It flags likely cosmic ray events, which are taken to be those pixels that are more than 5$\sigma$ deviations from the median (over the set of flux-weighted exposures) together with some neighbouring pixels. The final pixel value in the combined spectrum is computed as the variance-weighted mean over the flux-weighted exposures of the unflagged pixels.
The final result of the reduction is a pair of data files (one for each CCD), both of which contain a multi-spectrum image (200 fibres $\times$ 1024 spectral pixels), the associated variance array (which is correctly propagated throughout the reduction process), and a copy of the median sky spectrum used in the background subtraction. The data files also contain the identifications of the spectra obtained from the fibre configuration file and detailed information about the instrument and observation.
The spectra are not flux-calibrated. Flux calibration of fibre spectra is very difficult to achieve for a number of reasons. Chief among these are: (i) astrometric/positioning errors, which, combined with the small fibre aperture, prevent reproducible sampling of each galaxy’s light distribution; (ii) chromatic variations in distortion (see §\[sec:instrument\]) and residual errors in the atmospheric dispersion correction, which produce variations in the flux calibration dependent on the attitude of the telescope and the position of the object in the field. These effects can produce significant variations in the continuum slope of the spectra, and are discussed in quantitative detail in Madgwick (2001b) and Madgwick (2001c, in preparation).
Various problems are exhibited by some of the spectra: (i) [*Poor sky-subtraction.*]{} A small but significant fraction of fields suffered sky-subtraction errors of more than 5%. The strong oxygen sky lines at 5577Å, 6300Å, 6363Å and the OH bands to the red of 6000Å leave significant residuals in the spectra. (ii) [*Fringing.*]{} Damaged fibres sometimes produce a fringing effect that results in a strong oscillation in the throughput with wavelength. (iii) [*Halation.*]{} Sometimes halation due to condensation on the field flattener lens just in front of the CCD results in a high level of scattered light which fills in absorption features; when present, this effect was worse in CCD camera \#2. (iv) [*ADC error.*]{} A software error meant that the atmospheric dispersion corrector (ADC) was incorrectly positioned for observations prior to 31 August 1999; some of the spectra obtained before this date suffer from significant atmospheric dispersion losses.
REDSHIFTS {#sec:redshifts}
=========
Redshift estimation {#ssec:zcode}
-------------------
The main parameter to be determined from the spectra was the redshift of each object. A sophisticated and highly-tuned redshift code was developed in order to achieve a high level of precision, reliability and automation in the measurement of redshifts from the 2dFGRS spectral data.
The redshift code (version 990505) uses two quasi-independent automated methods: ‘absorption’ redshifts are obtained by cross-correlation with template spectra (after clipping emission lines) and ‘emission’ redshifts are obtained by finding and fitting emission lines. These automated redshift estimates are followed by a visual check and (occasionally, where necessary) a ‘manual’ redshift obtained by fitting identified spectral features.
Before these redshift estimates are made, however, the spectra are pre-processed in a number of ways. Residual features from the strong atmospheric emission lines at 5577Å, 5893Å, 6300Å and 7244Åare masked by interpolating over a small wavelength range (20–30Å) centred on the sky line. The atmospheric absorption bands around 6870Å (B-band) and 7600Å (A-band), and the fibre absorption band around 7190Å, are removed by dividing by an approximate band correction. This correction is obtained by fitting a low-order polynomial to the mean of all the spectra in the reduced image, dividing the mean spectrum by this smooth fit, finding the region about each band centre where this ratio is less than unity, and setting the band correction to the value of this ratio in these regions and to unity elsewhere. The spectra, which are not flux-calibrated, are finally multiplied by a simple quadratic approximation to the mean flux calibration correction in order to give appropriate weighting across the whole spectral range.
An ‘absorption’ redshift is obtained by cross-correlating the galaxy spectrum against a suite of 8 template spectra following the general method of Tonry & Davis (1979). The templates that are used cover a broad range of spectral types, and include 5 galaxies and 3 stars. The galaxies (and their morphological types) are: 1. NGC3379 (E), 2. NGC4889 (cD), 3. NGC5248 (Sbc), 4. NGC2276 (Sc) and 5. NGC4485 (Sm/Im); the stars (and their spectral types) are: 6. HD116608 (A1V), 7. HD23524 (K0V) and 8. BD051668 (M5V). The galaxy spectra come from spectrophotometric atlas of Kennicutt (1992); the stellar spectra come from the library of Jacoby (1984). The full set of template spectra are shown in Figure \[fig:tempspec\].
The galaxy and template spectra are prepared for cross-correlation using the following steps: (1) continuum-subtraction using a sixth-order polynomial fit to the continuum; (2) removal of strong emission lines by patching out all points more than 5 times the rms variation above the mean; (3) re-binning to a logarithmic wavelength scale with 2048 pixels; (4) apodising (tapering the ends of the spectrum to zero) by multiplying the first and last 5% of the spectrum with a cosine bell; (5) Fourier transformation; and, finally, (6) multiplication by an exponential filter of the form $e^{-k/k_h}-e^{-k/k_l}$, where $k$ is wavenumber in inverse pixels, $k_h$=300 and $k_l$=15, in order to reduce the effects of both the residual continuum at low wavenumber and the noise at high wavenumber.
The cross-correlation function is then computed as the inverse Fourier transform of the complex product of the filter Fourier transforms of the processed galaxy and template spectra. The highest peak in cross-correlation function is fitted with a quadratic in order to obtain the position and height of the peak. The significance of the peak is measured in terms of the ratio, $R$, of the height of the peak to the noise in the cross-correlation function, which is estimated from its anti-symmetric part. The absorption redshift is taken to be the cross-correlation redshift obtained with the template giving the highest value of $R$. An estimate of the quality of the absorption redshift, Q$_a$, is based on the value of $R$, with Q$_a$=1,2,3,4 for $R$$>$3.5,4.0,4.5,5.0 and Q$_a$=0 otherwise (with the additional requirement that Q$_a$=3 and Q$_a$=4 also require, respectively, at least 4 and 6 of the 8 templates give the same redshift to within 600).
An ‘emission’ redshift is obtained by fitting Gaussians to significant spectral features and searching for a multi-line match. The galaxy spectrum is first continuum-subtracted using a sixth-order polynomial fit to the continuum and lightly smoothed with a Gaussian kernel having a dispersion of 0.8 pixels. Peaks in the smoothed spectrum above a threshold (set at 3.3 times the robustly-estimated rms variation) are flagged as candidate emission lines. For each of these, in descending order of strength, a Gaussian is then fitted to the unsmoothed data, using the peak in the smoothed spectrum as the initial guess for the line centre. The Gaussian fits to the stronger lines are subtracted in turn before the weaker lines are fitted (to minimize blending effects in the H$\alpha$/NII pair). Then lines with fitted FWHM between 0.7 and 7 pixels, and total ‘significance’ of the whole line above 3.5$\sigma$ are marked as ‘good’; this rejects unclipped cosmic rays and low-significance features.
The ‘good’ lines are then sorted by strength, and the three strongest tested to see if at least two can be identified as common emission lines (\[OII\]3727Å, H$\beta$4861Å, \[OIII\]5007Å, H$\alpha$6563Åor \[NII\]6584Å) at a single redshift (the two out of three criterion is used to allow for one spurious feature, such as a residual sky line or cosmic ray). If so, then all the other lines in this set (plus \[OIII\]4959Å) that match this redshift to better than 600 are found and the mean redshift from these matching lines is adopted as the emission redshift. If not, then up to two possible single-line redshifts are kept for comparison with the absorption redshift, assuming the single line is either \[OII\] or H$\alpha$ in the range 0$<$$z$$<$0.4, as appropriate. An estimate of the quality of the emission redshift, Q$_e$, is based on the number and strength of the lines from which it is determined, with Q$_e$=0 if there are no lines, Q$_e$=1 for one weak line, Q$_e$=2 for any two lines or one strong line, and Q$_e$=4 for three or more lines.
The penultimate step is the choice of the best automated redshift estimate, given the absorption and emission redshifts and their quality parameters. The best redshift is taken to be whichever of the absorption and emission redshifts has the higher quality parameter (with the absorption redshift preferred if Q$_a$=Q$_e$). The best estimate of the redshift quality is Q$_b$=max(Q$_a$,Q$_e$), with two exceptions: (i) if the difference between the absorption and emission redshifts is less than 600 then Q$_b$=max(Q$_a$,Q$_e$,3); and (ii) if Q$_a$$\ge$2 and Q$_e$$\ge$2 and the difference between the absorption and emission redshifts is greater than 600, then this discrepancy is flagged and we set Q$_b$=1.
The final step in measuring the redshift is a visual check of the best automatic redshift. The redshift code displays for the user both the cross-correlation function and the galaxy spectrum with all the common spectral features superposed at the best automatic redshift. The spectrum displayed is the version actually used for the redshift estimate, with bad pixels and the strongest night sky lines patched out, and with an approximate correction applied for the instrumental response. To assist in identifying spurious features the code also displays the mean sky spectrum (to show the positions of the sky features), the atmospheric absorption bands and the variance array associated with the spectrum. The user is given the emission and/or absorption line redshifts, the best-estimate redshift, whether this is an emission or absorption redshift (or both), and whether (if both are obtained) the emission and absorption redshifts agree or not. If neither of the automated estimates are correct, the user can manually estimate a redshift by identifying a particular spectral feature and fitting Gaussians at the positions of all the common features. These fits can be to emission or absorption features and are rejected if there is no detectable feature near the nominal position. The manual redshift is taken to be the mean of the redshifts estimated from each of the well-fitted features. Given all this information, the observer can visually inspect the spectrum, try alternative redshifts (absorption, emission and manual) and determine the best redshift estimate.
Once the final redshift estimate has been decided, the user assigns a quality, Q, to this redshift on a five-point scale, with the nominal interpretation (see below for further discussion): Q=1 means no redshift could be estimated; Q=2 means a possible, but doubtful, redshift estimate; Q=3 means a ‘probable’ redshift (notionally 90% confidence); Q=4 means a reliable redshift (notionally 99% confidence); Q=5 means a reliable redshift and a high-quality spectrum. Note that this quality parameter is determined entirely by the subjective judgement of the user, and is independent of the automatic quality parameter Q$_b$. Quality classes 1 and 2 are considered failures, so the redshift completeness is the number of Q=3,4,5 redshifts divided by the total number of galaxies in the field. The standard redshift sample comprises objects with Q$\ge$3, but (for applications where redshift reliability is more important than completeness) one may prefer to use the set of objects with Q$\ge$4.
For the objects ending up with ‘acceptable’ (Q=3,4,5) redshifts, the best automated redshift is the same as the best visual redshift about 93% of the time; about 5% of the time either the absorption or the emission redshift agrees with the best visual redshift but the code has preferred the other one; and about 2% of the time neither of the estimates matches the best visual redshift, and a manual redshift estimate is required.
While there is some element of subjectivity in deciding whether a redshift estimate is ‘acceptable’ or not, this visual assessment is found to improve the reliability of the final sample, since the observer can assess a number of factors that are hard to fully automate, such as the overall shape of the spectrum, line ratios, the general quality of sky subtraction, and the proximity of features to sky lines or absorption bands. For 15,000 spectra, two or more users have measured redshifts independently using the same redshift code on the same spectra; the pair-wise ‘blunder’ rate (see below) is 0.4%, which is several times smaller than that for different observations of the same object. The rms difference in redshift completeness (fraction of objects per field with Q$\ge$3) between two users for the same data is less than half of that between the [*same*]{} observer measuring redshifts for many different fields. The human variation is therefore substantially smaller than the unavoidable variation between fields due to weather and other effects, and will to first order be corrected by the completeness factors.
Quality of spectra {#ssec:quality}
------------------
A quantitative overall measure of spectral quality is the median signal-to-noise ratio (S/N) per pixel computed over all pixels in the wavelength range 4000–7500Å with fluxes greater than zero. This is computed directly from the spectrum and it associated variance array, and can be compared to the quality class Q assigned to the redshift identification. The distribution of S/N over the 5 quality classes for the best spectrum of each object (i.e. the spectrum on which the adopted redshift is based) is shown in Figure \[fig:snr\]. The fraction of objects in each quality class and the median S/N for each quality class are given in the figure legend. Objects with reliable redshifts (Q=3,4,5) make up 91.8% of the sample and have a median S/N of 13.1 (6.3Å$^{-1}$).
Example 2dFGRS spectra for quality classes Q=3, Q=4 and Q=5 are shown in Figure \[fig:galspec\]; for each quality class, one object with an absorption redshift and one object with an emission redshift are shown. In order to be representative, the spectra are chosen to have a S/N per pixel close to the median of their quality class. Residuals from the strong night sky lines at 5577Å and 6300Å have been interpolated over.
Repeat observations {#ssec:intred}
-------------------
Amongst the first 150000 galaxies in the survey with redshifts, there are 8480 repeat redshift measurements; i.e. a second, or even a third, reliable (Q=3,4,5) redshift measurement for an object. These repeat measurements are either objects that lie in the overlaps between fields (7444 cases) or are in fields which were observed more than once, usually because the first observation was of poor quality and low completeness (1036 cases). For these objects we can compare their best redshift measurement with the repeat measurements to estimate both the rate of incorrect redshift identifications in the survey and the uncertainties on the correct redshift identifications.
The top panel of Figure \[fig:repeats\] shows the distribution of the redshift differences (w.r.t. the best redshift measurement for the object) for all the repeat observations. The rms redshift difference (robustly estimated here and throughout as the half-width of the central 68% of the distribution) is 120, implying an overall single-measurement rms uncertainty of 85.
The lower panels of the figure show the trends in the redshift uncertainty as a function of redshift, magnitude, S/N and quality class Q. As expected, the uncertainties increase at higher redshifts (from 95 for $z$$<$0.05 to 150 at $z$$>$0.15), fainter magnitudes (from 100 for $<$18.5 to 140 for $>$19.3) and lower S/N (from 95 for S/N$>$25 to 145 for S/N$<$5). The strongest trends are with redshift quality class and the method of estimating the redshift. Table \[tab:repeats\] gives the rms precision of single measurements as function of redshift quality class or measurement method (ABS = cross-correlation of absorption features; EMI = automatic fit to emission lines; MAN = manual fit to features). The fact that manual redshift estimation has the largest uncertainties mainly reflects the fact that it is only employed when the automatic methods have failed, usually in cases of low S/N.
---------- ----------- ----------- --------- -- -- -- -- --
Quality Fraction RMS error Blunder
class of sample () rate
Q=1 2.5% — —
Q=2 4.8% 143 69.5%
Q=3 11.9% 123 10.2%
Q=4 57.3% 89 0.9%
Q=5 23.5% 64 0.2%
Redshift Fraction RMS error Blunder
method of sample () rate
Q=1,2 7.3% — —
ABS 72.0% 87 0.7%
EMI 18.3% 61 0.9%
MAN 2.4% 159 9.3%
---------- ----------- ----------- --------- -- -- -- -- --
: Summary statistics by quality class and method.[]{data-label="tab:repeats"}
We divide the repeat measurements into two categories: correct identifications, where the redshift difference is less than 600, and ‘blunders’, where it is greater than 600. The division at 600corresponds to 5 times the rms error in the redshift differences. It should be noted that these blunders are not necessarily due to incorrect redshift measurements for one (or both) of the observed spectra. They may also be due to two observations of a close pair of objects with slight offsets in the fibre position, resulting in a different member of the pair dominating the spectral flux in each observation.
There are 263 blunders, so that the pair-wise blunder rate amongst the Q=3,4,5 repeat measurements is 3.1%. This implies a single-measurement blunder rate of just 1.6%. Figure \[fig:rejects\] shows plots of the redshift differences for the blunders with respect to redshift, magnitude, S/N and quality class; it also shows the pair-wise fraction of blunders as a function of each of these quantities.
There is no increase in the blunder rate with redshift, but a significant increase at fainter magnitudes, lower S/N and lower quality class. The single-measurement blunder rates as functions of quality class and measurement method are given in Table \[tab:repeats\]. We find that Q=3 corresponds to individual redshifts having 90% reliability, and that Q=4 and 5 correspond to better than 99% reliability, as intended. The blunder rate for Q=2 redshifts is 70%, consistent with the definition of Q=2 as ‘possible but doubtful’. The blunder rates for the absorption and emission redshifts are less than 1%, but the blunder rate for manual redshifts is 9% (due to the fact that manual redshifts are only used in the 2.4% of cases where the automatic methods fail).
The distribution of blunders shows a slight peak around zero redshift difference, suggesting that the tail of the error distribution for correct redshift identifications is more extended than a Gaussian. Some of the blunders may thus be correct identifications with larger than normal errors, implying that we may be slightly over-estimating the blunder rate. However, the most notable feature in the distribution of blunders is the large number of instances (22% of the total) where one or other of the redshifts in the comparison is approximately zero. The redshift difference is defined as (other redshift $-$ best redshift), so the upper envelope corresponds to objects where the best redshift is zero and the lower envelope corresponds to the other redshift being zero. In the former case the best redshift indicates that the objects are stars and the other (non-zero) redshift is due to a blunder based on a poor-quality spectrum; in the latter case we have the reverse of this situation. There are 3 times as many objects in the upper envelope as the lower, so stars being mistaken for galaxies is a much more common error than galaxies being mistaken for stars—which is remarkable given that galaxies are 20 times more common than stars in our sample.
We have repeated the above analyses after separating the repeat observations into those cases where the repeat is due to re-observing a field (12%) and those where it is due to observing an object in overlapping fields (88%). We find no significant difference in either the single-measurement rms precision or the blunder rate for these two subsets of repeat measurements.
Examination of the individual object spectra in the cases where there is a blunder shows that in general the best redshift appears to be quite reliable and that the main reason for the discrepancy is that the other redshift measurement has been obtained from a poor-quality (low-Q) spectrum.
External comparisons {#ssec:extred}
--------------------
In order to compare our redshift measurements with those from other sources, we have matched our catalogue to the redshift catalogues from the second Center for Astrophysics redshift survey (CfA2; Huchra 1999), the Stromlo-APM redshift survey (SAPM; Loveday 1996), the PSCz survey (PSCz; Saunders 2000), the Las Campanas redshift survey (LCRS; Shectman 1996), and also to the heterogeneous redshift compilation of the ZCAT catalogue (version November 13, 2000; J.P. Huchra, priv.comm.). Objects in these catalogues were matched to objects in the 2dFGRS by position; positions separated by 4 arcsec or less were assumed to belong to the same object. This criterion was chosen to include most genuine matches and exclude most false matches between close but unrelated objects.
Table \[tab:extred\] and Figure \[fig:extred\] summarise the results of these comparisons. The table lists, for each external source catalogue: the number of matched objects in the comparison, $N_{comp}$; the number of cases, $N_{600}$, where the redshift difference between the 2dFGRS and the other source indicates a blunder ($|\Delta
cz|>600{\mbox{\,km\,s$^{-1}$}}$); the median redshift difference, $\langle \Delta cz
\rangle$; and the rms scatter of $\Delta cz$ about this median, $\sigma(\Delta cz)$. The median and rms scatter are estimated excluding the blunders.
-------- ------------ ----------- ----------------------------- ---------------------
Source $N_{comp}$ $N_{600}$ $\langle \Delta cz \rangle$ $\sigma(\Delta cz)$
CfA2 87 1 $-$22 93
SAPM 228 10 $+$11 123
PSCz 172 7 $-$11 120
LCRS 3135 11 $-$1 109
ZCAT 1593 103 $-$33 127
-------- ------------ ----------- ----------------------------- ---------------------
: Comparisons of redshifts with external sources.[]{data-label="tab:extred"}
In interpreting these comparisons we need to keep in mind the magnitude range of the objects that are in common. Figure \[fig:magcomp\] shows the magnitude distributions of the comparison samples, all of which are significantly brighter than the 2dFGRS itself. The CfA2, SAPM and PSCz catalogues only overlap with the bright tail, while the median magnitude of the LCRS is a magnitude brighter than that of the 2dFGRS. ZCAT has a broader magnitude range, but is also heavily weighted towards brighter objects.
The best comparison is with the LCRS: the comparison sample is the largest, the magnitude range most similar, and the catalogue is homogeneous (unlike ZCAT). The rms redshift difference of 109 is consistent with the single-measurement uncertainty of 76 for objects in the 2dFGRS brighter than =18.5 and the typical uncertainty of 67 on the LCRS redshifts (Shectman 1996). The pair-wise blunder rate in the comparison sample is 11/3135 = 0.4%, which in fact is rather lower than might be expected from the pair-wise blunder rate of 46/2920 = 1.6% for repeat measurements of objects in the 2dFGRS brighter than =18.5; this may be due to the high surface brightness selection criterion of the LCRS.
The worst comparison is with ZCAT: the rms redshift difference is 127 and the pair-wise blunder rate is 103/1593 = 6.5%, which is more than twice the pair-wise blunder rate for the 2dFGRS repeats. Given the single-measurement blunder rate for the 2dFGRS is 1.6%, the implied single-measurement blunder rate for ZCAT is 5.0%, which is presumably due to the heterogeneous nature of its source catalogues.
Although the samples in common with CfA2, SAPM and PSCz are relatively small, they give results that are consistent with the rms redshift errors and pair-wise blunder rate obtained from the 2dFGRS repeat observations.
Checking the 2dFGRS spectra of the objects that have redshift blunders, we find that 0/1 cases in the comparison with CfA2, 1/10 for SAPM, 1/7 for PSCz and 4/11 for LCRS might [*possibly*]{} be attributed to errors in the 2dFGRS redshift estimates. Examining the DSS images of these objects, we find that 2 of the discrepancies with SAPM, 1 of those with PSCz, and 6 of those with LCRS, occur in cases where there is a close companion to the target object and hence some potential for confusion.
SURVEY MASKS {#sec:mask}
============
\
For accurate statistical analysis of the 2dFGRS it is essential to fully understand the criteria that define its parent photometric catalogue and also the spatial and magnitude-dependent completeness of the redshift catalogue. For this purpose we have defined three maps or masks characterizing this information as a function of position on the sky:
\(1) The magnitude limit mask gives the extinction-corrected magnitude limit of the survey at each position.
\(2) The redshift completeness mask gives the fraction of measured redshifts at each position.
\(3) The magnitude completeness mask gives a parameter defining how the redshift success rate depends on apparent magnitude.
Each mask has its own use, but for some analyses it is necessary to make use of two or even all three masks. We now describe in more detail how each one of these masks is defined and briefly outline some of their uses.
Magnitude limit mask {#ssec:maglimmask}
--------------------
Although the 2dFGRS sample was originally selected to have a uniform extinction-corrected magnitude limit of =19.45, in fact the survey magnitude limit varies slightly with position on the sky. There are two reasons for this. First, the photometric calibrations now available are much more extensive than when the parent 2dFGRS catalogue was originally defined. This has enabled us to recalibrate the whole 2dFGRS parent catalogue (Maddox 2001, in preparation), and results in new zero-point offsets and linearity corrections for each of the UKST photographic plates. Second, the extinction corrections have been changed to use the final published version of the Schlegel (1998) extinction maps; the original extinction corrections came from a preliminary version of those maps.
The magnitude limit mask is therefore defined by the change in the photometric calibration of each UKST photographic plate and the change in the dust extinction correction at each position on the sky. The magnitude limit masks for the NGP and SGP strips are shown in Figure \[fig:maglimmask\]; note that the mask also accounts for the holes in the source catalogue around bright stars and plate flaws.
In the SGP, which is a subset of the APM galaxy survey (Maddox 1990a,b,c), the rms change in plate zero-point is only 0.03mag. However, in the NGP region the original calibration was less accurate and the change in zero-points have an rms of 0.08mag. The change in the dust corrections are also less in the SGP, as the extinction is generally lower in this region. In the SGP the rms magnitude change due to improved dust corrections is 0.01mag while in the NGP it is 0.02mag.
The magnitude limit distribution over the NGP and SGP strips is shown in Figure \[fig:maglimdist\]. In the SGP the median limiting magnitude is =19.40 with an rms about this value of 0.05mag; in the NGP the median limiting magnitude is =19.35 with an rms of 0.11mag.
For accurate statistical analysis of the 2dF survey the magnitude limits defined by this mask should be used. It is always possible to analyse the data with a fixed magnitude limit if one is prepared to omit both the areas of the survey that have magnitude limits brighter than the chosen limit and also all the galaxies in the remaining areas with magnitudes fainter than the chosen limit.
Simple redshift completeness mask {#ssec:redmask}
---------------------------------
\
The best way to define a redshift completeness mask is to make use of the geometry defined by the complete set of 2 fields that were used to tile the survey region for spectroscopic observations. Each region of the sky inside the survey boundary is covered by at least one 2 field, but more often by several overlapping fields. We define a sector as the region delimited by a unique set of overlapping 2fields. This is the most natural way of partitioning the sky, as it takes account of the geometry imposed by the pattern of 2 fields and the way in which the galaxies were targeted for spectroscopic observation. Within each sector, $\theta$, we define the redshift completeness, $R(\theta)$, as the ratio of the number of galaxies for which redshifts have been obtained, $N_z(\theta)$, to the total number of objects contained in the parent catalogue, $N_{\rm p}(\theta)$: $$R(\theta) = N_z(\theta)/N_{\rm p}(\theta) .$$ The redshift completeness of a given sector, $R(\theta)$, should be clearly distinguished from the redshift completeness of a given field, , since multiple overlapping fields can contribute to a single sector.
This simple redshift completeness mask, shown in Figure \[fig:zcompmask\], can be used to locate regions in which the redshift completeness is high. It can also be used as a first step in either applying weights to statistically correct for incompleteness or to construct a random unclustered catalogue (for use in estimating correlation functions) that have the same angular pattern of incompleteness as the redshift sample. For this latter purpose one should also take account of how the redshift completeness depends on position within a sector as a result of constraints on fibre positioning and other considerations. This is best done by using the parent catalogue to derive weights for each galaxy with a measured redshift (Norberg 2001b, in preparation). Also, as discussed in the next section, one should take account of how the redshift completeness depends on apparent magnitude.
Magnitude completeness mask {#ssec:magmask}
---------------------------
The success rate of measuring redshifts is generally very high. Fields for which the field completeness, , is less than 70% are re-observed, while of the remainder just over 76% have a completeness greater than 90%. As one approaches the magnitude limit of the survey it becomes increasingly difficult to obtain good-quality spectra from which reliable redshifts can be measured. Hence the success of measuring redshifts (the redshift completeness) is a function of apparent magnitude.
In Figure \[fig:magcomplete\] we show the redshift completeness as a function of apparent magnitude for four different intervals of field completeness, . We see that in all cases the completeness is a function of apparent magnitude, and that the magnitude at which the completeness begins to drop is brightest in the fields with lowest , which are generally those taken in marginal observing conditions. Overall, the magnitude-dependent incompleteness is well modelled by a function of the form $$c_z(m,\mu)=\gamma(1-\exp(m-\mu))
\label{eq:magcomp}$$ with $\gamma=0.99$ and the parameter $\mu$ depending on . For each field we have chosen to fix $\mu$ by combining equation (\[eq:magcomp\]) with a simple power-law model for the galaxy number counts, $N(m) \propto \exp(\alpha m )$, so that and $\mu$ are related by $${\mbox{$\rm c_{\rm F}$}}(\mu)=\frac{\int_{m_1}^{m_2}N(m)c_z(m,\mu)\,dm}{\int_{m_1}^{m_2}N(m)\,dm} ,
\label{eq:numcount}$$ where $m_1$ and $m_2$ are the bright and faint magnitude limits. This integral can be evaluated to give $$\begin{aligned}
{\mbox{$\rm c_{\rm F}$}}(\mu)=\gamma\,(1-g(\alpha)\,\exp(m_1-\mu)) ,
\label{eq:numcount1}\end{aligned}$$ where $$\begin{aligned}
g(\alpha)=\frac{\alpha}{\alpha+1}
\frac{\exp\left[(\alpha+1)(m_2-m_1)\right]-1}{\exp\left[\alpha(m_2-m_1)\right]-1}\end{aligned}$$ and we take $\alpha$=0.5 as appropriate for the galaxy number counts around =19. This equation can be inverted, yielding $$\mu=\ln\left[\frac{ g(\alpha)\,\exp(m_1)}{(1-{\mbox{$\rm c_{\rm F}$}}/\gamma)}\right] ,
\label{eq:numcount2}$$ which enables $\mu$ to be computed for each observed field.
Our goal is to define the value of $\mu$ characterising the magnitude-dependent completeness for each position in the sky. Since the 2 fields overlap it again makes sense to define $\mu(\theta)$ for each sector $\theta$. The value of $\mu(\theta)$ is defined by an appropriately-weighted average of the $\mu$ values of the $N_{\rm
F}(\theta)$ overlapping observed fields comprising the sector. Specifically we take $$\begin{aligned}
\mu(\theta)=-\ln\left[\sum_{i=1}^{N_{\rm F}(\theta)}f_i\,\exp(-\mu_i)\right] ,
\label{eq:mutheta}\end{aligned}$$ where $f_i$ is the fraction of the observed galaxies in this sector that were targeted in field $i$.
Finally, it is straightforward to combine the magnitude completeness given by $\mu(\theta)$ with the redshift completeness $R(\theta)$ to define an estimate of redshift completeness that depends on both position (sector) and magnitude: $$\begin{aligned}
S(\theta,m)=\frac{N_{\rm p}(\theta)}{N_{\rm e}(\theta)}\,R(\theta)\,c_z(m,\mu(\theta)) .
\label{eq:comps}\end{aligned}$$ In the first factor, $N_{\rm p}(\theta)$ is again the number of parent catalogue galaxies in this sector while $$\begin{aligned}
N_{\rm e}(\theta)=\sum_{i=1}^{N_p(\theta)}c_z(m_i,\mu(\theta))\end{aligned}$$ is an estimate of the number of galaxies one would expect to have measured redshifts for given the value of $\mu$ that has been assigned to this sector. For ease in calculating equation (\[eq:comps\]) we tabulate $N_{\rm p}(\theta)/N_{\rm e}(\theta)$ as well as $\mu(\theta)$ and $R(\theta)$, using, for convenience, a pixellated mask rather than the exact sectors.
SURVEY DATABASE {#sec:database}
===============
The 2dFGRS database consists of three main components: (i) a collection of FITS files, one per object in the source catalogue, that contains every piece of information about each object from the source catalogue, spectroscopic observations and subsequent analysis; (ii) an mSQL database (Jepson & Hughes 1998), that contains all of the parameters for each object and allows complex searching and subsetting of the survey objects and the retrieval of selected subsets of spectra and images; and (iii) a WWW interface, that provides a number of different modes for querying the mSQL database and a variety of ways for returning the results of such queries. The next three subsections describe these three database components.
Object FITS files {#ssec:fits}
-----------------
There are [467214]{} target objects in the FITS database, each with its own FITS file. This number is much larger than the number of objects surveyed, since it includes objects down to =19.6 in the NGP and SGP strips. Each object in the survey source catalogue has been given a serial number ([SEQNUM]{}), and the name of the FITS file for that object is this serial number. The serial numbers for objects in the SGP strip are 1–193550 and 416694–467214 for objects in the NGP strip 193551–332694 and 389714–416693, and for objects in the random fields 332695–389713.
Each FITS file has a primary part which contains all the source catalogue data about the object (as FITS keywords) and a DSS postage stamp image of the object. Spectra are appended as additional FITS extensions. Each spectrum extension contains the spectrum of the object, the variance (error) array for the object spectrum, the spectrum of the mean sky that was subtracted from the object spectrum, and FITS keyword data giving information about the spectroscopic observation and derived parameters such as the redshift and spectral quality. Many targets contain multiple spectrum extensions corresponding to multiple observations.
Table \[tab:fits0\] lists all the FITS keywords present in the primary part (extension 0) of the FITS files; Table \[tab:fits1\] lists all the keywords present in the spectrum extensions (1$\ldots$[spectra]{}). The tables give the names of the keywords, example values, and the keyword definitions.
--------------- ------------------- --------------------------------------------------------------------------------
Keyword Example Definition
[SIMPLE ]{} [T]{} [file does conform to FITS standard]{}
[BITPIX ]{} [16]{} [number of bits per data pixel]{}
[NAXIS ]{} [2]{} [number of data axes]{}
[NAXIS1 ]{} [49]{} [length of data axis 1]{}
[NAXIS2 ]{} [49]{} [length of data axis 2]{}
[EXTEND ]{} [T]{} [FITS dataset may contain extensions]{}
[BSCALE ]{} [1.0000]{} [REAL = (FITS $\times$ BSCALE) + BZERO]{}
[BZERO ]{} [0.0000]{} [zeropoint of conversion to REAL]{}
[SEQNUM ]{} [100100]{} [Serial number : database primary key]{}
[NAME ]{} [’TGS469Z164’ ]{} [2dFGRS assigned name]{}
[IMAGE ]{} [’SKYCHART’ ]{} [Existence of postage stamp image]{}
[RA ]{} [0.7943429758]{} [RA (B1950) in radians : 3 2 3.00]{}
[DEC ]{} [-0.5475286941]{} [DEC (B1950) in radians : -31 22 15.9]{}
[EQUINOX ]{} [1950.00]{} [equinox of RA and DEC]{}
[BJSEL ]{} [18.93]{} [final mag used in the object selection]{}
[PROB ]{} [2335.4]{} [psi star-galaxy classification parameter]{}
[PARK ]{} [0.910]{} [k star-galaxy classification parameter]{}
[PARMU ]{} [0.187]{} [mu star-galaxy classification parameter]{}
[IGAL ]{} [1]{} [final classification flag (1 for a galaxy)]{}
[JON ]{} [-1]{} [eyeball classification flag]{}
[ORIENT ]{} [91.0]{} [orientation in degrees clockwise from E-W]{}
[ECCENT ]{} [0.270]{} [eccentricity]{}
[AREA ]{} [308.0]{} [isophotal area in pixels]{}
[X\_BJ ]{} [2918.7]{} [plate $x_{{\mbox{b$_{\rm\scriptscriptstyle J}$}}}$ in 8 micron pixels]{}
[Y\_BJ ]{} [9123.1]{} [plate $y_{{\mbox{b$_{\rm\scriptscriptstyle J}$}}}$ in 8 micron pixels]{}
[DX ]{} [43.0]{} [corrected difference $100(x_{{\mbox{b$_{\rm\scriptscriptstyle J}$}}}-x_R)$]{}
[DY ]{} [49.0]{} [corrected difference $100(y_{{\mbox{b$_{\rm\scriptscriptstyle J}$}}}-y_R)$]{}
[BJG ]{} [18.99]{} [ without extinction correction]{}
[RMAG ]{} [10.35]{} [unmatched APM ‘total’ mag]{}
[PMAG ]{} [10.53]{} [unmatched raw APM profile integrated mag]{}
[FMAG ]{} [8.72]{} [unmatched raw APM 2 ‘fibre’ mag]{}
[SMAG ]{} [10.74]{} [unmatched raw stellar mag (from APMCAL)]{}
[IFIELD ]{} [417]{} [UKST field ]{}
[IGFIELD ]{} [2007]{} [galaxy number in UKST field]{}
[REGION ]{} [’S417 ’ ]{} [GSSS region name]{}
[OBJEQNX ]{} [2000.00]{} [equinox of the plate reference frame]{}
[OBJRA ]{} [0.8034094522]{} [RA (J2000) in radians : 03 04 07.673 ]{}
[OBJDEC ]{} [-0.5441390223]{} [DEC (J2000) in radians : -31 10 36.73]{}
[PLTSCALE ]{} [67.2000]{} [Plate scale in arcsec per mm]{}
[XPIXELSZ ]{} [25.2844500]{} [X pixel size in microns]{}
[YPIXELSZ ]{} [25.2844500]{} [Y pixel size in microns]{}
[OBJPLTX ]{} [7970.86]{} [object X on plate (pixels)]{}
[OBJPLTY ]{} [4148.11]{} [object Y on plate (pixels)]{}
[DATAMAX ]{} [14431]{} [Maximum data value]{}
[DATAMIN ]{} [4011]{} [Minimum data value]{}
[BJSELOLD ]{} [18.96]{} [original mag used in the object selection]{}
[BJG\_OLD ]{} [19.01]{} [original without extinction correction]{}
[END ]{} [End of FITS header]{}
--------------- ------------------- --------------------------------------------------------------------------------
: FITS keywords for the primary image (extension 0)[]{data-label="tab:fits0"}
--------------- ------------------------------- --------------------------------------
Keyword Example Definition
[XTENSION ]{} [’IMAGE ’ ]{} [IMAGE extension]{}
[BITPIX ]{} [-32]{} [number of bits per data pixel]{}
[NAXIS ]{} [2]{} [number of data axes]{}
[NAXIS1 ]{} [1024]{} [length of data axis 1]{}
[NAXIS2 ]{} [3]{} [length of data axis 2]{}
[PCOUNT ]{} [0]{} [required keyword; must = 0]{}
[GCOUNT ]{} [1]{} [required keyword; must = 1]{}
[CRVAL1 ]{} [5802.8979492]{} [coordinate value of axis 1]{}
[CDELT1 ]{} [4.3103027344]{} [coordinate increment on axis 1]{}
[CRPIX1 ]{} [512.0000000000]{} [reference pixel on axis 1]{}
[CUNIT1 ]{} [’Angstroms’ ]{} [units for axis 1]{}
[EXTNAME ]{} [’SPECTRUM’ ]{} [2dFGRS spectrum]{}
[OBSNAME ]{} [’TGS469Z164’ ]{} [observed object name]{}
[OBSRA ]{} [0.7943429758]{} [observed RA (B1950) in radians]{}
[OBSDEC ]{} [-0.5475286940]{} [observed DEC (B1950) in radians]{}
[MATCH\_DR]{} [0.0000]{} [position match error in arcsec]{}
[Z ]{} [0.178876]{} [raw measured redshift]{}
[Z\_HELIO ]{} [0.178860]{} [heliocentric redshift]{}
[QUALITY ]{} [5]{} [redshift measurement quality]{}
[ABEMMA ]{} [1]{} [redshift type: abs=1,emi=2,man=3]{}
[NMBEST ]{} [0]{} [\# emission lines for emission z]{}
[NGOOD ]{} [0]{} [number of good emission lines]{}
[Z\_EMI ]{} [-9.9990]{} [emission redshift]{}
[Q\_Z\_EMI]{} [0]{} [emission redshift quality]{}
[KBESTR ]{} [2]{} [cross-correlation template]{}
[R\_CRCOR ]{} [15.5600]{} [cross-correlation peak]{}
[Z\_ABS ]{} [0.1789]{} [cross-correlation redshift]{}
[Q\_Z\_ABS]{} [3]{} [cross-correlation quality]{}
[Q\_FINAL ]{} [3]{} [suggested quality for redshift]{}
[IALTER ]{} [0]{} [IALTER=1 if automatic z altered]{}
[Z\_COMM ]{} [’ ’ ]{} [observer’s comment]{}
[THPUT ]{} [0.96613]{} [fibre throughput]{}
[SPFILE ]{} [’sgp469\_991104\_1z.fits’]{} [2dF reduced data file]{}
[PLATE ]{} [1]{} [2dF plate number]{}
[PIVOT ]{} [302]{} [2dF pivot number]{}
[FIBRE ]{} [58]{} [2dF fibre number]{}
[OBSRUN ]{} [’99OCT ’ ]{} [observation run]{}
[GRS\_DATE]{} [’991104 ’ ]{} [2dF YYMMDD observed date]{}
[UTDATE ]{} [’1999:11:04’ ]{} [UT date of observation]{}
[SPECTID ]{} [’A ’ ]{} [2dF spectrograph ID]{}
[GRATID ]{} [’300B ’ ]{} [2dF grating ID]{}
[GRATLPMM ]{} [300]{} [2dF grating line per mm]{}
[GRATBLAZ ]{} [’COLLIMATOR’ ]{} [2dF grating blaze direction]{}
[GRATANGL ]{} [25.30000]{} [2dF grating angle]{}
[LAMBDAC ]{} [5782.700]{} [central wavelength]{}
[CCD ]{} [’TEKTRONIX\_5’ ]{} [CCD ID]{}
[CCDGAIN ]{} [2.790]{} [CCD inverse gain (e per ADU)]{}
[CCDNOISE ]{} [5.200]{} [CCD read noise (electrons)]{}
[OBJX ]{} [196833]{} [2dF object X position]{}
[OBJY ]{} [10401]{} [2dF object Y position]{}
[OBJXERR ]{} [6]{} [2dF object X position error]{}
[OBJYERR ]{} [14]{} [2dF object Y position error]{}
[OBJMAG ]{} [18.96]{} [2dF object magnitude]{}
[THETA ]{} [4.526]{} [2dF fibre angle]{}
[PTRTYPE ]{} [’P ’ ]{} [2dF ptrtype]{}
[PID ]{} [0]{} [2dF pid]{}
[OBSFLD ]{} [’sgp469 ’ ]{} [2dF observed field number]{}
[NCOMB ]{} [3]{} [number of frames combined]{}
[REFRUN ]{} [31]{} [AAT run number of reference]{}
[UTSTART ]{} [’16:37:59.48’ ]{} [UT start of reference exposure]{}
[UTEND ]{} [’16:57:59’ ]{} [UT end of reference exposure]{}
[REFEXP ]{} [1200.0]{} [reference run exposure (secs)]{}
[REFHASTA ]{} [’36.07264’ ]{} [HA start of reference exposure]{}
[REFHAEND ]{} [’41.08119’ ]{} [HA end of reference exposure]{}
[ETA\_TYPE]{} [-2.5934501E+00]{} [eta spectral type parameter]{}
[SNR ]{} [2.0299999E+01]{} [median S/N per pixel]{}
--------------- ------------------------------- --------------------------------------
: FITS keywords for the spectra (extensions 1$\ldots$[spectra]{})[]{data-label="tab:fits1"}
mSQL parameter database {#ssec:database}
-----------------------
The mSQL database can be thought of as a table. The rows of the table are labelled by the unique object serial number ([SEQNUM]{}) and the extension number ([extnum]{}), with multiple rows for each target object corresponding to each extension in the object’s FITS file. The columns of the table correspond to the object parameters, and are labelled by the name of the corresponding keyword. The object serial numbers ([SEQNUM]{}) provide the primary database key, but the objects are also indexed by their unique survey name ([NAME]{}), which has the format [TGhfffZnnn]{}, where [h]{} is the hemisphere (N for the NGP strip and S for the SGP strip and random fields), [fff]{} is the number of the primary field to which the object is assigned and [nnn]{} is the number of the galaxy within that field. Note that the observed name of the object (parameter [OBSNAME]{} in each spectrum extension) is the same as [NAME]{} except that: (i) if the field in which the object is observed (given by [OBSFLD]{}) is an overlapping field rather than its primary field (given by [fff]{}), then the first character of the name is changed from [T]{} to [X]{}; and (ii) if the object has been flagged as a possible merger, then the second character of the name is changed from [G]{} to [M]{}.
The first row for each object ([extnum]{}=0) contains the source catalogue data and the basic spectroscopic information from the best spectrum of that object. The keywords for that row are the FITS parameters for the primary image (Table \[tab:fits0\]) plus all the additional keywords listed in Table \[tab:msql\]. The best spectrum is the one with the highest redshift quality parameter; if there is more than one spectrum of the same quality, then the most recent of these spectra is used. Subsequent rows for the same object ([extnum]{}=1$\ldots$[spectra]{}, where [spectra]{} is the number of spectra obtained for that object) contain the FITS parameters pertaining to each spectroscopic observation (Table \[tab:fits1\]) plus the additional keywords in section (i) of Table \[tab:msql\]. If there is no spectrum for the object then [spectra]{}=0 and only the row corresponding to [extnum]{}=0 will exist. Note that some information is duplicated between rows and that not all parameters are defined for all rows; undefined parameters return a [NULL]{} value.
[lrl]{}\
Keyword & Example & Definition\
[serial ]{}&[100100]{}&[2dFGRS serial number]{}\
[name ]{}&[TGS469Z164]{}&[2dFGRS name]{}\
[UKST ]{}&[417]{}&[UKST sky survey field number]{}\
[spectra ]{}&[1]{}&[number of spectra for this object]{}\
[extnum ]{}&[1]{}&[extension number]{}\
[obsrun ]{}&[99OCT]{}&[observing run year and month]{}\
[TDFgg ]{}&[-469]{}&[2dFGRS field number (+NGP, -SGP)]{}\
[pivot ]{}&[302]{}&[2dF pivot]{}\
[plate ]{}&[1]{}&[2dF plate]{}\
[fiber ]{}&[58]{}&[2dF fibre]{}\
[z ]{}&[0.178876]{}&[observed redshift]{}\
[z\_helio]{}&[0.178860]{}&[heliocentric redshift]{}\
[abemma ]{}&[1]{}&[redshift type (abs=1,emi=2,man=3)]{}\
[quality ]{}&[5]{}&[redshift quality parameter]{}\
& &\
\
Keyword & Example & Definition\
[alpha ]{}&[0.7943429758]{}&[RA (B1950) in radians]{}\
[delta ]{}&[-0.5475286941]{}&[DEC (B1950) in radians]{}\
[ra ]{}&[3 2 3.00]{}&[RA (B1950) in HH MM SS.SS]{}\
[dec ]{}&[-31 22 15.9]{}&[DEC (B1950) in DD MM SS.S ]{}\
[ra2000 ]{}&[03 04 07.68]{}&[RA (J2000) in HH MM SS.SS]{}\
[dec2000 ]{}&[-31 10 36.8]{}&[DEC (J2000) in DD MM SS.S ]{}\
[l2 ]{}&[228.9258834424]{}&[Galactic longitude]{}\
[b2 ]{}&[-60.8572447739]{}&[Galactic latitude]{}\
Searches of the database use the mSQL query format (Jepson & Hughes 1998), which has the basic format
SELECT list\_of\_parameters FROM database\_name WHERE list\_of\_conditions
and [list\_of\_conditions]{} is a series of equalities and inequalities linked by Boolean relations.
An example is
SELECT name, extnum, ra, dec, BJSEL, Z, QUALITY, z, quality FROM public WHERE name=’TGS469Z164’
which selects the listed parameters for the object with 2dFGRS name [TGS469Z164]{} (note the single quotes around the character string) from both the summary row ([extnum]{}=0) and for each spectrum ([extnum]{}=1$\ldots$[spectra]{}). Note that the parameters with the same name in lower case and upper case are distinct: the former are generally from [extnum]{}=0, the latter from [extnum]{}$>$0 (parameters are returned as [NULL]{} in rows where they are not defined).
An example with a more complex list of conditions is
SELECT name FROM public WHERE extnum=0 AND ((BJSEL$<$15.5 AND quality$>$=3) OR quality$>$4)
which lists just the names of the objects which are either brighter than =15.5 with redshift quality at least 3, or have quality greater than 4, or both; the search is restricted just to the summary row by requiring [extnum]{}=0.
Simple searches on the two indexed parameters, [serial]{} and [name]{}), are quick—e.g. [WHERE serial=69656]{} or [WHERE name=’TGS203Z081’]{}; more complex searches take about 5 minutes. Further information about the mSQL database software and its structured query language is given in Yarger (1999) and on the WWW at http://www.hughes.com.au.
WWW interface {#ssec:www}
-------------
The 2dFGRS mSQL database can be searched via the WWW interface in a number of ways: (i) search via a standard mSQL query as described above—this is the most general method; (ii) perform a standard mSQL query restricted to a list of named objects; (iii) perform a standard mSQL query restricted to objects within a specified radius of a given position on the sky; (iv) match objects to a supplied catalogue of positions.
The results of a query can be returned in several forms: as an HTML table, as an ASCII table, as a gzipped ASCII file, as an email giving the URL of a gzipped ASCII file created with a background job, or as a tar file containing the FITS files for the selected objects. The HTML table has buttons allowing you to select a particular object and view the postage stamp DSS image and all the observed spectra for the object; if the spectra have measured redshifts, then the positions of prominent spectral features are indicated at the redshift associated with each individual spectrum. The FITS files for the objects in the HTML table can also be bundled up into a compressed tar file which can then be down loaded by anonymous ftp.
Full instructions for accessing the database are given on the survey website at http://www.mso.anu.edu.au/2dFGRS and at mirror sites given there.
RESULTS {#sec:results}
=======
The redshift distribution of the galaxies in the survey is shown in Figure \[fig:2dFnz\]. Also shown is a simple analytic approximation to the redshift distribution, following Efstathiou & Moody (2001), with the form $$dN \propto z^2 \exp \left[ - \left(
\frac{1.36z}{\bar{z}} \right)^{1.55} \right]\,{\rm d}z$$ where $\bar{z}$=0.11 is the median redshift of the survey.
The spatial distribution of the galaxies in the survey strips is shown in Figure \[fig:2dFzcone\]. This figure is the projection of the full width of the strips (10 in the NGP and 15 in the SGP). As not all fields have yet been observed, there are variations in the effective thicknesses of the strips in the directions of the missing fields that produce corresponding variations in the galaxy density. The clustering of galaxies revealed in this figure has been investigated by Peacock (2001), Percival (2001) and Norberg (2001a).
The 2dFGRS has provided the first clear detection of the redshift-space clustering anisotropy on large scales that is a key prediction of the gravitational instability paradigm for the growth of structure in the universe (Peacock 2001). Measurements of this distortion yield a precise estimate of the parameter $\beta = \Omega^{0.6}/b = 0.43 \pm
0.07$, where $\Omega$ is the total mass density of the universe and $b$ is the linear bias of the galaxies with respect to the mass. Combined with recent measurements of the anisotropies in the cosmic microwave background (Jaffe 2001), this result favours a low-density universe, with $\Omega$$\approx$0.3.
The power spectrum of the galaxy distribution has been determined from the survey using a direct FFT-based technique (Percival 2001). Over the range in wavenumber 0.02$<$$k$$<$0.15, the shape of the observed power spectrum will be close to that of the linear density perturbations convolved with the window function of the survey. Fitting convolved power spectrum models constrains the shape parameter $\Gamma =
\Omega h$ to be $0.20 \pm 0.03$ (68% confidence level) and shows that models containing baryon oscillations are preferred over models without baryons at the 95% confidence level. This is the first detection of baryon oscillations in the galaxy distribution, and yields an estimate for the baryon fraction $\Omega_b/\Omega_m = 0.15 \pm 0.07$, assuming scale-invariant primordial fluctuations.
The size of the 2dFGRS sample has allowed investigation of the variation in the strength of galaxy clustering with luminosity, using the projected two-point correlation function of galaxies in a series of volume-limited samples (Norberg 2001a). The clustering of $L^*$ ($M_{b_J}^* - 5\log h = -19.7$) galaxies in real space is well fitted by a power-law relation with exponent $\gamma = 1.71 \pm 0.06$ and a correlation length $r_0^* = (4.9\pm0.3){\mbox{$\,h^{-1}\,{\rm Mpc}$}}$. The exponent shows very little variation for galaxy samples with luminosities differing by a factor of 40. The clustering amplitude, however, increases with luminosity: slowly for galaxies fainter than $M^*$, but more strongly at brighter absolute magnitudes. This dependence of the correlation length on luminosity is in good agreement with the predictions of the hierarchical galaxy formation models of Benson (2001). In terms of the bias parameter, the relation is well-represented by $\delta b/b^*
= 0.15 \delta L/L^*$, where $\delta b = b-b^*$, $\delta L = L-L^*$ and the relative bias is given by $b/b^* = (r_0/r_0^*)^{\gamma/2}$.
The 2dFGRS has also been used to characterise the internal properties of the galaxy population. Folkes (1999) and Madgwick (2001a) have determined the luminosity function for galaxies, both overall and as a function of spectral type. Principal Component Analysis has been applied to the 2dFGRS spectra, and a linear combination of the first two principal components, $\eta$, has been used to parametrise the spectral type. Going from early types to late types, the luminosity functions appear to exhibit a systematic decrease in the characteristic luminosity (from $M_{b_J}^* - 5\log h = -19.6$ to $-19.0$) and a steepening of the faint-end slope (from $\alpha = -0.52$ to $-1.43$). However there is also evidence that, at the precision afforded by the 2dFGRS sample, the standard Schechter fitting function is no longer an adequate representation of the luminosity function.
Figure \[fig:etacones\] shows the redshift slices for each of the four spectral types defined in terms of the $\eta$ parameter by Madgwick (2001a): type 1 are the earliest types and type 4 the latest types. The association of earlier types with clusters and local density enhancements is apparent, as is the tendency for later types to be associated with lower-density regions. A quantitative study of the clustering of different spectral types will be the subject of a subsequent paper.
\
\
\
\
The large 2dFGRS sample also allows the generalisation of the luminosity function into the bivariate brightness distribution (BBD) over both luminosity and surface brightness (Cross 2001). The BBD derived from the 2dFGRS shows a strong surface brightness–luminosity relation, $M_{b_J} \propto (2.4\pm^{1.5}_{0.5}) \mu_e$. The luminosity density is dominated by normal giant galaxies and the peak of the BBD lies away from the survey selection boundaries, implying that the 2dFGRS is complete and that luminous low surface brightness galaxies are rare. By integrating over the BBD, the local luminosity density is estimated to be $j_B = (2.5\pm0.2)\times10^8\,h\,L_{\odot}\,{\rm Mpc}^{-3}$.
Cole (2001) have combined the 2dFGRS with the 2MASS extended source catalogue to produce an infrared-selected sample of over 17000 galaxies with redshifts. This sample has been used to determine the $J$ and $K_S$-band galaxy luminosity functions. The luminosity functions (in 2MASS Kron magnitudes) are fairly well fit by Schechter functions: in $J$ with $M_J^* - 5\log h = -22.36\pm0.02$, $\alpha_J = -0.93\pm0.04$, $\Phi_J^* = 0.0104\pm0.0016{\mbox{$\,h^3\,{\rm Mpc}^{-3}$}}$, and in $K_S$ by $M_K^* - 5\log
h = -23.44\pm0.03$, $\alpha_K = -0.96\pm0.05$, $\Phi_K^* =
0.0108\pm0.0016{\mbox{$\,h^3\,{\rm Mpc}^{-3}$}}$. From the distributions of $B$$-$$K$ and $J$$-$$K$ colours with absolute magnitude and models of the stellar populations, the galaxy stellar-mass function can be estimated. Integrated over all galaxy masses, it yields a total mass fraction in stars (in units of the critical density) of $\Omega_{\rm stars}h =
(1.6\pm0.2)\times10^{-3}$ for a Kennicutt IMF, and $\Omega_{\rm stars}h
= (2.9\pm0.4)\times10^{-3}$ for a Salpeter IMF. These values are consistent with estimates based on the time integral of the observed star formation history of the universe only if dust extinction corrections at high redshift are modest.
The 2dFGRS can add significant value to all-sky imaging surveys at all wavelengths, as in the studies of the radio galaxy population combining redshifts and optical spectra from the 2dFGRS with the NVSS radio survey (Sadler 2001) and the FIRST radio survey (Magliocchetti 2001). Using 20% of the full 2dFGRS area, Sadler (2001) find 757 optical counterparts for NVSS sources—the largest and most homogeneous set of radio-source spectra to date. These sources range from $z$=0.005 to $z$=0.438, and are a mixture of active galaxies (60%) and star-forming galaxies (40%). The local radio luminosity function at 1.4 GHz is determined for both active and star-forming galaxies, and yields an estimate for the local star-formation density of $(2.8\pm0.5)\times10^{-3}\,M_{\odot}\,{\rm yr}^{-1}{\mbox{$\,h^3\,{\rm Mpc}^{-3}$}}$.
CONCLUSIONS {#sec:conclusions}
===========
The 2dF Galaxy Redshift Survey (2dFGRS) is designed to measure redshifts for approximately 250000 galaxies. The survey uses the 2dF multi-fibre spectrograph on the Anglo-Australian Telescope, which is capable of observing up to 400 objects simultaneously over a 2 diameter field of view. The source catalogue for the survey is a revised and extended version of the APM galaxy catalogue, which is based on Automated Plate Measuring machine (APM) scans of 390 plates from the UK Schmidt Telescope (UKST) Southern Sky Survey. The target galaxies have extinction-corrected magnitudes brighter than =19.45, with extinctions derived from the maps of Schlegel (1998).
The main survey regions are two contiguous declination strips, one in the southern Galactic hemisphere (the SGP strip) covering 80$\degr$$\times$15$\degr$ centred close to the SGP, and the other in the northern Galactic hemisphere (the NGP strip) covering 75$\degr$$\times$10$\degr$. In addition, there are 99 random fields are spread uniformly over the entire region of the APM catalogue in the southern Galactic hemisphere outside the SGP strip. In total the survey covers some 2000deg$^2$ and has a median depth of $\bar{z}$=0.11. Out to the effective limit of the survey at $z$$\approx$0.3, the strips contain a volume of 1.2$\times$10$^8$; the volume sparsely sampled including the random fields is two to three times larger. An adaptive tiling algorithm is used to give a highly uniform sampling rate of 93% over the whole survey region.
Spectra are obtained over the wavelength range 3600–8000Å at two-pixel resolution of 9.0Å. The median $S/N$ is 13 over 4000–7500Å. Redshifts are measured both by cross-correlation with a range of template spectra and by fitting strong spectral features. All redshift identifications are visually checked and assigned a quality parameter Q in the range 1–5. From repeat measurements and comparisons with other redshift catalogues, we find that redshifts with Q=3 have a blunder rate (fraction of incorrect identifications) of just under 10% and Q=4 and 5 redshifts have blunder rates less than 1%—the overall blunder rate for reliable (Q$\ge$3) redshifts is 3%. The overall rms uncertainty in the Q$\ge$3 redshifts is 85. The completeness of the survey is computed as a function of both field and apparent magnitude. The overall redshift completeness is 91.8%, but this varies with magnitude from 99% for the brightest galaxies to 90% for objects at the survey limit.
The survey database has two components: a collection of FITS files, one per object, which contain all the parameters and spectra for each object, and a mSQL parameter database which can be used for sophisticated searching, matching and sorting of the survey data. The 2dFGRS database is available through the survey WWW site at http://www.mso.anu.edu.au/2dFGRS.
Acknowledgements {#acknowledgements .unnumbered}
================
The 2dF Galaxy Redshift Survey was made possible through the dedicated efforts of the staff of the Anglo-Australian Observatory, both in creating the 2dF instrument and in supporting the survey observations.
Arnouts S., , 2001, A&A, submitted Bailey J.A., Glazebrook K., Offer A., Taylor K., 2001, MNRAS, submitted Benson A.J., Frenk C.S., Baugh C.M., Cole S., Lacey C.G., 2001, MNRAS, submitted Blair M., Gilmore G., 1982, PASP, 94, 741 Boyle B.J., Shanks T., Croom S.M., Smith R.J., Miller L., Loaring N., Heymans C., 2000, MNRAS, 317, 1014 Cole S., (2dFGRS team), 2001, MNRAS, in press Colless M.M., 1999, Phil. Trans. Roy. Soc. Lond. A, 357, 105 Croom S.M., Smith R.J., Boyle B.J., Shanks T., Loaring N.S., Miller L., Lewis I.J., 2001, MNRAS, 322, 29 Cross N. (2dFGRS team), 2001, MNRAS, 324, 825 De Propris R., (2dFGRS team), 2001, MNRAS, submitted Efstathiou G., Moody S.J., 2001, MNRAS, in press Folkes S. (2dFGRS team), 1999, MNRAS, 308, 459 Huchra J.P., Vogeley M.S., Geller M.J., 1999, ApJS, 121, 287 Jacoby G.H., Hunter D.A., Christian C.A., 1984, ApJS, 56, 257 Jaffe A.H., , 2001, Phys. Rev. Lett., 86, 3475 Jepson B., Hughes D.J., 1998, The Official Guide to miniSQL 2.0, John Wiley & Sons Kennicutt R.C., 1992, ApJS, 79, 255 Lasker B.M. , 1998, BAAS, 192, 6403 Lewis I.J., , 2001, submitted Loveday J., 1996, MNRAS, 278, 1025 Loveday J., Maddox S.J., Efstathiou G., Peterson B.A., 1996, ApJS, 107, 201 Maddox S.J., Efstathiou G., Sutherland W.J., Loveday J., 1990a, MNRAS, 242, 43P Maddox S.J., Efstathiou G., Sutherland W.J., Loveday J., 1990b, MNRAS, 243, 692 Maddox S.J., Efstathiou G., Sutherland W.J., 1990c, MNRAS, 246, 433 Maddox S.J., Efstathiou G., Sutherland W.J., 1996, MNRAS, 283, 1227 Madgwick D.S., (2dFGRS team), 2001a, MNRAS, submitted Madgwick D.S., (2dFGRS team), 2001b, ‘Mining the Sky’, eds Banday A. , ESO Astrophysics Symposia, Springer-Verlag, in press Magliochhetti M., (2dFGRS team), 2001, MNRAS, submitted Metcalfe N., Fong R., Shanks T., 1995, MNRAS, 274, 769 Monet D., , 1996, USNO-A1.0: A Catalog of Astrometric Standards (Washington: US Nav. Obs.) Norberg P., (2dFGRS team), 2001a, MNRAS, submitted Peacock J.A., (2dFGRS team), 2001, Nature, 410, 169 Percival W.J., (2dFGRS team), 2001, MNRAS, submitted Pimbblet K.A., Smail I., Edge A.C., Couch W.J., O’Hely E., Zabludoff A.I., 2001, MNRAS, submitted Sadler E.M., (2dFGRS team), 2001, MNRAS, submitted Saunders W., , 2000, MNRAS, 317, 55 Schlegel D.J., Finkbeiner D.P., Davis M., 1998, ApJ, 500, 525 Shectman S.A., Landy S.D., Oemler A., Tucker D.L., Lin H., Kirshner R.P., Schechter P.L., 1996, ApJ, 470, 172 Taylor K., Cannon R.D., Watson F.G., 1997, Proc. SPIE, 2871, 145 Tonry J.L., Davis M., 1979, AJ, 84, 1511 Yarger R.J., Reese G., King T., 1999, MySQL and mSQL, O’Reilly & Associates
[^1]: The Digitized Sky Survey was produced at the Space Telescope Science Institute, and is based on photographic data obtained using the Oschin Schmidt Telescope on Palomar Mountain and the UK Schmidt Telescope.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We say that a family of $k$-subsets of an $n$-element set is [*intersecting*]{}, if any two of its sets intersect. In this paper we study properties and structure of large intersecting families.
We prove a conclusive version of Frankl’s theorem on intersecting families with bounded maximal degree. This theorem, along with its generalizations to cross-intersecting families, strengthens the results obtained by Frankl, Frankl and Tokushige, Kupavskii and Zakharov and others.
We study the structure of large intersecting families, obtaining some very general structural theorems which extend the results of Han and Kohayakawa, as well as Kostochka and Mubayi.
We also obtain an extension of some classic problems on intersecting families introduced in the 70s. We extend an old result of Frankl, in which he determined the size and structure of the largest intersecting family of $k$-sets with covering number $3$ for $n>n_0(k)$. We obtain the same result for $n>Ck$, where $C$ is an absolute constant. Finally, we obtain a similar extension for the following problem of Erdős, Rothschild and Szémeredi: what is the largest intersecting family, in which no element is contained in more than a $c$-proportion of the sets, for different values of $c$.
address: 'University of Birmingham, Moscow Institute of Physics and Technology; Email: [[email protected]]{}.'
author:
- Andrey Kupavskii
title: Structure and properties of large intersecting families
---
Introduction
============
For integers $a\le b$, put $[a,b]:=\{a,a+1,\ldots, b\}$, and denote $[n]:=[1,n]$ for shorthand. For a set $X$, denote by $2^{X}$ its power set and, for integer $k\ge 0$, denote by ${X\choose k}$ the collection of all $k$-element subsets ([*$k$-sets*]{}) of $X$. A is simply a collection of sets. We call a family , if any two of its sets intersect. A “trivial” example of an intersecting family is the family of all sets containing a fixed element. We call a family , if the intersection of all sets from the family is empty.
One of the oldest and most famous results in extremal combinatorics is the Erdős–Ko–Rado theorem:
Let $n\ge 2k>0$ and consider an intersecting family ${\mathcal F}\subset {[n]\choose k}$. Then $|{\mathcal F}|\le {n-1\choose k-1}$. Moreover, for $n>2k$ the equality holds only for the families of all $k$-sets containing a given element.
Answering a question of Erdős, Ko, and Rado, Hilton and Milner [@HM] found the size and structure of the largest non-trivial intersecting families of $k$-sets. For $k\ge 4$, up to a permutation of the ground set, it must have the form $\mathcal H_{k}$, where for integer $2\le u\le k$ $$\label{eqhu}\mathcal H_u:=\ \Big\{A\in {[n]\choose k}\ :\ [2,u+1]\subset A\Big\}\cup\Big\{A\in{[n]\choose k}\ :\ 1\in A, [2,u+1]\cap A\ne \emptyset\Big\}.$$ $\mathcal J_1$ has size ${n-1\choose k-1}-{n-k-1\choose k-1}+1$, which is much smaller than ${n-1\choose k-1}$, provided $n$ is large as compared to $k$.
For a family ${\mathcal F}\subset 2^{[n]}$ and $i\in [n]$, the $d_i({\mathcal F})$ of $i$ in ${\mathcal F}$ is the number of sets from ${\mathcal F}$ containing $i$. Let $\Delta({\mathcal F})$ stand for the of an element in ${\mathcal F}$. Frankl [@Fra1] proved the following far-reaching generalization of the Hilton–Milner theorem.
\[thmfr\] Let $n>2k>0$ and ${\mathcal F}\subset {[n]\choose k}$ be an intersecting family. If $\Delta({\mathcal F})\le {n-1\choose k-1}-{n-u-1\choose k-1}$ for some integer $3\le u\le k$, then $$|{\mathcal F}|\le {n-1\choose k-1}+{n-u-1\choose n-k-1}-{n-u-1\choose k-1}.$$
One can deduce the Hilton–Milner theorem from the $u=k$ case of Theorem \[thmfr\]. Theorem \[thmfr\] is sharp for integer values of $u$, as witnessed by the example . On a high level, it provides us with an upper bound on $|{\mathcal F}|$ in terms of the size of the largest trivial subfamily ([*star*]{}) in ${\mathcal F}$. Let us state a stronger version of Theorem \[thmfr\] in dual terms. For a family ${\mathcal F}$, the $\gamma({\mathcal F})$ is the quantity $|{\mathcal F}|-\Delta({\mathcal F})$. One may think of diversity as of the distance from ${\mathcal F}$ to the closest star.
The following strengthening of Theorem \[thmfr\] was obtained by Kupavskii and Zakharov [@KZ].
\[thm1\] Let $n>2k>0$ and ${\mathcal F}\subset {[n]\choose k}$ be an intersecting family. If $\gamma({\mathcal F})\ge {n-u-1\choose n-k-1}$ for some real $3\le u\le k$, then $$\label{eq01}|{\mathcal F}|\le {n-1\choose k-1}+{n-u-1\choose n-k-1}-{n-u-1\choose k-1}.$$
We note that the Hilton–Milner theorem, as well as Theorem \[thmfr\], is immediately implied by Theorem \[thm1\]. The deduction of Theorem \[thm1\] from Theorem \[thmfr\] for integer values of $u$ is possible, but not straightforward. Theorem \[thm1\] provides the strongest known stability result for the Erdős–Ko–Rado theorem for large intersecting families, more precisely, for the families of size at least ${n-2\choose k-2}+2{n-3\choose k-2}$. There are several other stability results for the Erdős–Ko–Rado theorem, see, e.g. [@DT; @EKL; @Fri].
In Section \[sec3\], we prove a conclusive version of Theorem \[thm1\], which gives the precise dependence of size of an intersecting family ${\mathcal F}$ of $k$-sets on (the lower bound on) $\gamma({\mathcal F})$. The result then is extended to cover the equality case, as well as the weighted case and the case of cross-intersecting families. In particular, it strengthens the results of [@Fra1], [@FT], [@KZ]. One of the main ingredients in the proof of the main result of Section \[sec3\] (as well as in the proofs of Theorems \[thmfr\], \[thm1\]) is the famous Kruskal–Katona theorem [@Kr; @Ka]. This theorem is a central result in extremal set theory and has numerous applications (see, e.g., [@BK; @BT]). Another key ingredient is the bipartite switching trick, which was introduced in [@KZ] (similar ideas appeared earlier in [@FK1]). In this paper, we exploit this trick to a much greater extent.
We do not state the aforementioned theorem in this section since it requires some preparations. Instead, we state the following corollary, which can be seen as a generalization of the Hilton–Milner phenomena.
\[corhm\] Let $n>2k\ge 8$ and ${\mathcal F}\subset {[n]\choose k}$ be an intersecting family. Suppose that $\gamma({\mathcal F})>{n-u-1\choose n-k-1}$ for some integer $4\le u\le k$. Then $$\label{eqhm}|{\mathcal F}|\le {n-1\choose k-1}+{n-u-1\choose n-k-1}-{n-u-1\choose k-1}-{n-k-2\choose k-2}+1.$$ Moreover, the same inequality holds for $n>2k\ge 6$ with $u=3$ if ${n-4\choose k-3}<\gamma({\mathcal F})\le {n-3\choose k-2}-{n-k-2\choose k-2}+1$ or ${n-3\choose k-2}< \gamma({\mathcal F})\le {n-3\choose k-2}+{n-4\choose k-2}-{n-k-2\choose k-2}+1$.
Compare with for integer $u$. The difference in the bounds is ${n-k-2\choose k-2}-1$, while the lower bounds on diversity differ by $1$. Thus, Corollary \[corhm\] states that the size of the largest family with diversity at least $\gamma$ has a big drop when $\gamma$ passes the point ${n-u-1\choose n-k-1}$ for integer $u$. We also note that Corollary \[corhm\] is sharp, as it will be clear from Section \[sec3\].\
Numerous authors aimed to determine [*precisely*]{}, what are the largest intersecting families in ${[n]\choose k}$ with certain restrictions. One of such questions was studied by Han and Kohayakawa, who determined the largest family of intersecting families that is neither contained in the Erdős–Ko–Rado family, nor in the Hilton–Milner family. In our terms, the question can be simply restated as follows: what is the largest intersecting family with $\gamma({\mathcal F})\ge 2$? The proof of Han and Kohayakawa is quite technical and long. Kruskal–Katona-based arguments allow for a very short and simple proof in the case $k\ge 5$. For $i\in[k]$ let us put $I_i:=[i+1,k+i]$ and $$\mathcal J_i:=\ \{I_1,I_i\}\cup \Big\{F\in{[n]\choose k}\ :\ 1\in F, F\cap I_1\ne \emptyset, F\cap I_i\ne \emptyset\Big\}.$$ We note that $\mathcal J_i\subset{[n]\choose k}$ and that $\mathcal J_i$ is intersecting for every $i\in [k]$. Moreover, $\gamma(\mathcal J_i) = 2$ for $i>1$ and $\mathcal J_1$ is the Hilton–Milner family. It is an easy calculation to see that $|\mathcal J_i|>|\mathcal J_{i+1}|$ for every $k\ge 4$ and $i\in [k-1]$.[^1]
\[thmhk\] Let $n>2k$, $k\ge 4$. Then any intersecting family ${\mathcal F}$ with $\gamma({\mathcal F})\ge 2$ satisfies $$\label{eqhk}|{\mathcal F}|\le {n-1\choose k-1}-{n-k-1\choose k-1}-{n-k-2\choose k-2}+2,$$ moreover, for $k\ge 5$ the equality is attained only on the families isomorphic to $\mathcal J_2$.
We note that Han and Kohayakawa also proved their theorem for $k= 3$, as well as described the cases of equality for $k=4$. These cases are more tedious and do not follow from our methods in a straightforward way. However, Theorem \[thmhk\] can be deduced without much effort. A slightly weaker version of the theorem above (without uniqueness) is a consequence of the main result in the paper by Hilton and Milner [@HM] (cf. also [@HK]).
Applying Corollary \[corhm\] with $u=k$, we conclude that holds for $k\ge 4$. The bound is sharp, as witnessed by $\mathcal J_2$. The uniqueness requires hardly more effort, but since it uses Theorem \[thmfull1\], stated in Section \[sec3\], we postpone its proof until Section \[sec4\].
For any set $X$, family ${\mathcal F}\subset 2^X$ and $i\in X$, we use the following standard notations $$\begin{aligned}
{\mathcal F}(\bar i):=&\ \{F\ :\ i\notin F\in{\mathcal F}\}\ \ \ \text{and}\\
{\mathcal F}(i):=&\ \{F\setminus\{i\}\ :\ i\in F\in{\mathcal F}\}.\end{aligned}$$ Note that ${\mathcal F}(\bar i)\subset {X\setminus \{i\}\choose k}$ and ${\mathcal F}(i)\subset {X\setminus \{i\}\choose k-1}$.
Denote by $\mathcal E_{l}$ the maximal intersecting family with $|\mathcal E_l(\bar 1)|=l$, $|\bigcap_{E\in \mathcal E_l(\bar 1)}E|=k-1$ (note that the family is defined up to isomorphism). Note that $\mathcal J_2$ is isomorphic to $\mathcal E_2$.[^2] It is not difficult to see that for $k-1<l<n-k$ we have $\mathcal E_l\subset \mathcal E_{n-k}$: we have $\mathcal E_l(1) = \mathcal E_{n-k}(1)$ for this range. The following theorem is one of the main results in [@KostM]:
\[thmko\] Let $k\ge 5$ and $n=n(k)$ be sufficiently large. If ${\mathcal F}\subset {[n]\choose k}$ is intersecting and $|{\mathcal F}|> |\mathcal J_3|$ then ${\mathcal F}\subset \mathcal E_l$ for $l\in \{0,\ldots k-1, n-k\}$.
We note that it is easy to verify that $|\mathcal J_3|<|\mathcal E_{n-k}|$, e.g., for $n>4k$. The authors of [@KostM] we using the delta-systems method of Frankl.[^3] Actually, Theorem \[thmko\] can be deduced from the results of Frankl [@F16] with little extra effort.
Many results in extremal set theory are much easier to obtain once one assumes that $n$ is sufficiently large in comparison to $k$. (The possibility to apply the delta-method is one of the reasons.) In particular, the bound on $n$ in Theorem \[thmko\] is doubly exponential in $k$. In this paper, we deduce Theorem \[thmko\] from a much more general result, which, additionally, holds without any restriction on $n$. A family is called with respect to some property, if none of its proper subfamilies possesses the property. For shorthand, we say that ${\mathcal M}$ is if, for any $M_l\in {\mathcal M}$, we have $|\bigcap_{M\in {\mathcal M}\setminus\{M_l\}}M|>|\bigcap_{M\in {\mathcal M}}M|$. The following theorem is one of the main results of this paper.
\[thmclass2\] Assume that $n>2k\ge 8$. Consider an intersecting family ${\mathcal F}\subset{[n]\choose k}$ with $\Delta({\mathcal F}) = d_1({\mathcal F})$. Take a subfamily $\mathcal M\subset {\mathcal F}(\bar 1)$, which is minimal w.r.t. common intersection and such that $|\bigcap_{M\in {\mathcal M}}M|= t$. Take the (unique) maximal intersecting family ${\mathcal F}'$, such that ${\mathcal F}'(\bar 1) = {\mathcal M}$. If $t\ge 3$ then we have $$\label{eqclass1} |{\mathcal F}|\le |{\mathcal F}'|,$$ and, for $k\ge 5$ equality is possible if and only if ${\mathcal F}$ is isomorphic to ${\mathcal F}'$.
Moreover, if ${\mathcal F}$ is as above and $|\bigcap_{F\in {\mathcal F}(\bar 1)}F|\le t$ for some $t\ge 3$, then $$\label{eqclass2} |{\mathcal F}|\le |\mathcal J_{k-t+1}|,$$ and, for $k\ge 5$, equality is possible only if ${\mathcal F}$ is isomorphic to $\mathcal J_{k-t+1}$.
This theorem generalizes Theorems \[thmhk\] and \[thmko\] and gives a reasonable classification of [*all*]{} large intersecting families. We also note that we cannot in general replace the condition on $t$ by $t\ge 2$. Indeed, one can see that the family $\mathcal H_2$ (cf. ) is much larger than $\mathcal J_{k-1}$ for large $n$.[^4]
\[cormk\] The statement of Theorem \[thmko\] is valid for any $n>2k\ge 10.$
Fix any ${\mathcal F}$ as in Theorem \[cormk\] and assume that $\Delta({\mathcal F}) = d_1({\mathcal F})$. First assume that $|\bigcap_{F\in{\mathcal F}(\bar 1)}|\le k-2$. Since $k\ge 5$, we are in position to apply the second part of Theorem \[thmclass2\] to ${\mathcal F}$, and get a contradiction with $|{\mathcal F}|>|\mathcal J_3|$. Therefore, $|\bigcap_{F\in {\mathcal F}(\bar 1)}F|=k-1$ and thus ${\mathcal F}(\bar 1)$ is isomorphic to $\mathcal E_l(\bar 1)$ for $l = |\mathcal F(\bar 1)|$, which concludes the proof.
The proof of Theorem \[thmclass2\] is given in Section \[sec41\]. The main tool of the proof of Theorem \[thmclass2\] is again the aforementioned bipartite switching trick. In this paper, we exploit it to much greater extent than in [@FK1] and [@KZ]. One key observation that allows us to prove Theorem \[thmclass2\] is that this bipartite switching is possible even in situations when we know practically nothing about the structure of the family.\
In the remaining part of the introduction, we present two results that extend some classic results of Frankl, Erdős Rothshild and Semeredi, and Furedi proven for $n>n_0(k)$ (with double-exponential dependency on $k$, coming from the aforementioned delta-method) to the range $n>Ck$, where $C$ is an absolute constant. Apart from the bipartite switching trick, we use as a main tool the junta approximation theorem due to Dinur and Friedgut [@DF]. The approach is resemblant of the recent paper [@Kup21] of the author, but here we apply it to a much wider class of problems. The rough framework of combining junta method to get approximate structure with combinatorial arguments for finer structure was used in an excellent recent paper [@KL]. One novel aspect in the use of junta method in the first problem below is that the actual extremal configuration is quite far from being a junta (i.e., the family is not defined by the intersection with a constant-size subset of the ground set, see precise definition in Section \[sec5\])! This, of course, poses additional complications.
For a family ${\mathcal F}$, let $\tau({\mathcal F})$ denote the of ${\mathcal F}$, that is, the minimum size of set $S$ that intersects all sets in ${\mathcal F}$. Each such $S$ we call a .
Intersecting families of $k$-sets with fixed covering number were studied in several classical works. The Erdős–Ko–Rado theorem shows that the largest intersecting family of $k$-element sets has covering number $1$. The aforementioned result of Hilton and Milner [@HM] determined the largest intersecting family with covering number $2$. It is clear that any $k$-uniform intersecting family ${\mathcal F}$ satisfies $\tau({\mathcal F})\le k$: indeed, any set of ${\mathcal F}$ is a hitting set for ${\mathcal F}$. In a seminal paper [@EL], Erdős and Lovász proved that an intersecting family ${\mathcal F}\subset {[n]\choose k}$ with $\tau({\mathcal F})=k$ has size at most $k^k$ (note that it is independent of $n$!) and provided a lower bound of size roughly $(k/e)^k$. Later, both lower [@FOT2] and upper [@Che; @AR; @Fra21] bounds were improved.
Let us define the following important family. $$\label{deft2}
\mathcal T_2(k):=\big\{[k]\big\}\cup \big\{\{1\}\cup [k+1,2k-1]\big\}\cup \big\{\{2\}\cup [k+1,2k-1]\big\}.$$ It is easy to see that $\mathcal T_2(k)$ is intersecting, moreover, $\tau(\mathcal T_2(k))=2$. In [@F16], Frankl studied the following question: what is the size $c(n,k,t)$ of the largest intersecting family ${\mathcal F}\subset {[n]\choose k}$ with $\tau({\mathcal F})\ge t$? Define $\mathcal C_3(n,k)\subset {[n]\choose k}$ to be the maximal intersecting family with $\mathcal C_3(n,k)(\bar 1)$ isomorphic to $\mathcal T_2(k)$. It is easy to see that $\tau(\mathcal C_3(n,k))=3$. Frankl managed to prove the following theorem.
\[thmtau3\] Let $k\ge 3$ and $n\ge n_0(k)$. Then $c(n,k,3)= |\mathcal C_3(n,k)|$. Moreover, for $k\ge 4$ the equality holds only for families isomorphic to $\mathcal C_3(n,k)$.
As in the case of Theorem \[thmko\], the bound on $n$ is doubly exponential in $k$. One of the main results of this paper is the proof of Theorem \[thmtau3\] under much milder restrictions.
\[thmtau3\] The conclusion of Theorem \[thmtau3\] holds for any $n>Ck$, where $C$ is an absolute constant, independent of $k$.
An important tool in the proof (as well as in the proof of Theorem \[thmfull2\]) is Lemma \[lemmin\], in which we found an elegant way to bound the sizes of families ${\mathcal F}$ satisfying $\tau({\mathcal F})=3$ and $\tau({\mathcal F}(\bar 1))=2$, and where ${\mathcal F}(\bar 1)$ is minimal w.r.t. this property.
In the paper [@FOT1], the authors managed to extend the result of [@F16] to the case $\tau =4$, determining the exact value of $c(n,k,4)$ and the structure of the extremal family for $n>n_0(k)$. The analysis in [@FOT1] is much more complicated than that in [@F16], and the problem for $\tau\ge 5$ is still wide open. The result of [@FOT1] may be extended to much smaller $n$ in a similar way, but, of course, progress on the case $\tau \ge 5$ would be more interesting.\
Erdős, Rothschild and Szemerédi (cf. [@Erd73]) raised the following question: how large can the intersecting family ${\mathcal F}\subset {[n]\choose k}$ be, given that $\Delta({\mathcal F})\le c|{\mathcal F}|$. They proved that, for any fixed $2/3<c<1$, there exists $n_0(k,c)$, such that for any $n>n_0(k,c)$ the largest intersecting family under this restriction is $\mathcal H_2$ (see ), up to isomorphism. Stronger results of this type were proven by Frankl [@F16] and by Füredi [@Fur]. See also a survey [@DeF], where the concise statement of the results is given.
The methods we developed in this paper allow us to extend these results to the range $n>Ck$, where $C$ depends on $c$ only. We are going to illustrate it for one such theorem, but we note that the same ideas would work for other cases. Let us state one of the theorems, proven in [@F16]. Recall that a (projective plane of order $2$) $\mathcal P$ is a family consisting of $7$ $3$-element sets $P_1,\ldots, P_7$, such that $|P_i\cap P_j|=1$ for each $i\ne j$.
\[thmfano\] Suppose that ${\mathcal F}\subset {[n]\choose k}$ is intersecting, and $\Delta({\mathcal F})\le c|{\mathcal F}|$ for some $c\in (3/7,1/2)$. Assume that ${\mathcal F}$ has the largest cardinality among such families. Then, for any $n>n_0(c,k)$, ${\mathcal F}$ is isomorphic to $$\mathcal D_{3/7}:=\ \Big\{F\in {[n]\choose k}\ :\ P\subset F\text{ for some }P\in \mathcal P\Big\}.$$
Our contribution here is as follows.
\[thmbounddeg\] There exists an absolute constant $C$, such that the following holds. In terms of Theorem \[thmfano\], if there exists $\epsilon>0$, such that $c\in [3/7+\epsilon, 1/2-\epsilon]$, then the conclusion of Theorem \[thmfano\] holds for $n>\frac{C}{\epsilon}k$.
The proof of this theorem is given in Section \[sec52\]. In Section \[sec6\], we discuss our results and methods, as well as pose some open problems.
The complete diversity version of Frankl’s theorem {#sec3}
==================================================
Let us begin with the statement of the Kruskal–Katona theorem. We shall state the Kruskal–Katona theorem in two different forms, but the most handy for our purposes is the form due to Hilton in terms of cross-intersecting families. Let us first give some definitions.
For a set $X$, (lex) $\prec$ on the sets from ${X\choose k}$ is a total order, in which $A\prec B$ iff the minimal element of $A\setminus B$ is smaller than the minimal element of $B\setminus A$. For $0\le m\le {|X|\choose k}$ let $\mathcal L(X,m,k)$ be the collection of $m$ largest $k$-element subsets of $X$ with respect to lex. We say that two families $\mathcal A,\mathcal B$ are if $A\cap B\ne \emptyset$ for any $A\in\mathcal A, B\in \mathcal B$.
\[thmHil\]If $\mathcal A\subset {[n]\choose a}, \mathcal B\subset {[n]\choose b}$ are cross-intersecting then the families $\mathcal L([n],|\mathcal A|,a),\mathcal L([n],|\mathcal B|, b)$ are also cross-intersecting.
In this section, we analyze in great details the relationship between the diversity of an intersecting family and its size. We first note that, if the value of diversity is given precisely, then it is easy to determine the largest intersecting family with such diversity. Indeed, the subfamilies of sets containing the element of the largest degree and not containing the element of the largest degree are cross-intersecting, and one can get exact bounds using Theorem \[thmHil\]. Studying the size of an intersecting family with given upper bounds on diversity is not interesting: the largest intersecting family has diversity 0.
In this section we obtain the concluding version of Theorem \[thm1\], which tells [*exactly*]{}, how large an intersecting family may be, given a [*lower*]{} bound on its diversity. We determine all “extremal” values of diversity and the sizes of the corresponding families.
The difficulty to obtain such a version of Theorem \[thm1\] is that, while Theorem \[thmHil\] gives a very strong and clear characterisation of families with fixed diversity, the size of the family is not monotone w.r.t. diversity (the size of the largest family with given diversity does not necessarily decrease as diversity increases, although it is true in “most” cases). Moreover, the numerical dependence between maximal possible sizes of ${{\mathcal A}}$ and ${{\mathcal B}}$ given by Theorem \[thmHil\] is complicated and difficult to work with, see [@FMRT] and [@FT7]. Thus, an effort is needed to find the right point of view on the problem.+0.1cm
We shall give two versions of the main theorem of this section. First, we give a “quantitative” version with explicit sharp bounds on the size of intersecting families depending on the lower bound on their diversity. It may be more practical to apply in some cases, but it is difficult to grasp what is hidden behind the binomial coefficients in the formulation. Thus, later in the section (and as an intermediate step of the proof), we shall give a “conceptual” version of our main theorem. We note that the proof that we present is completely computation-free. In Sections \[sec33\], \[sec34\] we present strengthenings and generalisations of our main result. The cases of equality in Theorems \[thmfull1\] and \[thmfull2\] are described in Theorem \[thmfulleq\]. We provide a weighted version of the main result and a generalization to the case of cross-intersecting families.
We note that the main results of the section are meaningful for any $k\ge 3$. (This is by no means a serious restriction since possible structure of intersecting families in ${[n]\choose k}$ for $k\le 2$ is trivial.)+0.1cm
The following representation of natural numbers is important for the (classic form of) the Kruskal–Katona theorem. Given positive integers $\gamma$ and $k$, one can always write down $\gamma$ [*uniquely*]{} in the : $$\gamma = {a_k\choose k}+{a_{k-1}\choose k-1}+\ldots +{a_s\choose s}, \ \ a_k>a_{k-1}>\ldots >a_s\ge 1.$$
For the sake of comparison, let us state the classical version of the Kruskal–Katona theorem (equivalent to Theorem \[thmHil\]).
\[thmkk\] Let ${\mathcal F}\subset {[n]\choose k}$ and put $$\partial({\mathcal F}):=\big\{F'\in {[n]\choose k-1}\ :\ F'\subset F \text{ for some }F\in {\mathcal F}\big\}.$$ If $|{\mathcal F}| = {a_k\choose k}+\ldots +{a_s\choose s}$, then $$|\partial({\mathcal F})|\ge {a_k\choose k-1}+{a_{k-1}\choose k-1}+\ldots +{a_s\choose s-1}.$$
Next, we start preparations to state the main result of this section. Given a number $\gamma<{n-1\choose k-1}$, let us write it in the $(n-k-1)$-cascade form: $$\gamma = {n-b_1\choose n-k-1}+{n-b_{2}\choose n-k-2}+\ldots+{n-b_{s_b}\choose n-k-s_b},$$ where $1< b_1<b_2<\ldots<b_{s_b}.$[^5] Define $T_{\gamma}:=\{b_1,\ldots,b_{s_b}\}$ and put $S_{\gamma}:=T_{\gamma}\,\oplus\, [2,b_{s_b}-1],$ where $\oplus$ stands for symmetric difference. Note that $S_{\gamma}\cup T_{\gamma} = [2,b_{s_b}]$ and $S_{\gamma}\cap T_{\gamma} = \{b_{s_b}\}$. Suppose that $S_{\gamma} = \{a_1,\ldots, a_{s_a}\}$, where $1< a_1<\ldots<a_{s_a} := b_{s_b}$.
\[defresnu\] We call a nonnegative integer $\gamma$ , if either $\gamma={n-4\choose k-3}$ or the following holds:
1. $s_a := |S_{\gamma}|\le k$ and $s_b := |T_{\gamma}|\le k-1$;
2. $b_i>2i+2$ for each $i\in [s_b]$.
In particular, any integer $\gamma>{n-4\choose k-3}$ has ${n-4\choose k-3}$ as one of the summands in the $(n-k-1)$-cascade form, and thus is not resistant.
Let $0= \gamma_0<\gamma_1<\ldots <\gamma_{m} = {n-4\choose k-3}$ be all the resistant numbers in increasing order.
\[thmfull1\] Let $n>2k\ge6$. Consider an intersecting family ${\mathcal F}\subset {[n]\choose k}$. Suppose that $\gamma_{l-1}<\gamma ({\mathcal F})\le\gamma_l$ for $l\in [m]$ and that the representation of $\gamma_l$ in the $(n-k-1)$-cascade form is $$\gamma_l = {n-b_1\choose n-k-1}+{n-b_{2}\choose n-k-2}+\ldots +{n-b_{s_b}\choose n-k-s_b},$$ then $$\label{eqfull1}|{\mathcal F}|\le {n-a_1\choose n-k}+{n-a_{2}\choose n-k-1}+\ldots +{n-a_{s_a}\choose n-k-s_a+1}+\gamma_l,$$ where $\{b_1,\ldots, b_{s_b}\} = T_{\gamma_l}$ and $\{a_1,\ldots, a_{s_a}\} = S_{\gamma_l}$.
The expression in the right hand side of strictly decreases as $l$ increases.
Moreover, the presented bound is sharp: for each $l =1,\ldots, m$ there exists an intersecting family with diversity $\gamma_l$ which achieves the bound in .
Theorem \[thm1\] and the concept of diversity was successfully used to advance in several problems concerning intersecting families [@Feg; @FK5; @IK; @Kup22]. However, sometimes the statement of Theorem \[thm1\] was not fine-grained enough for the applications, and some extra work was needed to deduce some rudimentary versions of Theorem \[thmfull1\]. (One example of an application of an easy application of Theorem \[thmfull1\] is Theorem \[thmhk\] in the introduction.) This is one of our motivations for proving Theorem \[thmfull1\]. The other motivation is the desire to have the conclusive version of Theorems \[thmfr\] and \[thm1\] and highlight the parallel with the original version of the Kruskal–Katona theorem.
Let us mention that we state Theorem \[thmfull1\] only for $\gamma({\mathcal F})\le {n-4\choose k-3}$, since Theorem \[thm1\] already gives us the bound $|{\mathcal F}|\le {n-2\choose k-2}+2{n-3\choose k-2}$ if $\gamma({\mathcal F})\ge {n-4\choose k-3}$, and we cannot get any better bound in general for larger $\gamma$. Indeed, the intersecting family $\mathcal H_2$ (cf. ) attains the bound above on the cardinality and has diversity ${n-3\choose k-2}$. The second part of Corollary \[corhm\] complements Theorem \[thm1\] in this respect, showing the aforementioned example is essentially the only exception. We also note that in a recent work [@Kup21] the author managed to prove that there exists $c$, such that for any $n,k$ satisfying $n\ge ck$ we have $\gamma({\mathcal F})\le {n-3\choose k-2}$ for any intersecting ${\mathcal F}\subset{[n]\choose k}$. Earlier, this was shown to be true for $n\ge 6k^2$ by Frankl [@Fra6]. Thus, for $n\ge ck$ Theorem \[thmfull1\] gives a complete answer to the question we address.
In Section \[sec33\], we shall analyze the cases when the equality in can be attained, as well as the cases of very large diversity. In particular, we shall prove the $u=3$ part of Corollary \[corhm\].
Some examples and Theorem \[thmfull1\] restated {#sec31}
-----------------------------------------------
Let us try to familiarize the reader with the statement of Theorem \[thmfull1\]. We have $\gamma_i = i$ for $i = 1,\ldots, k-3$. Indeed, for $1\le \gamma <n-k-1$ we have $\gamma = {n-k-1\choose n-k-1}+{n-k-2\choose n-k-2}+\ldots +{n-k-\gamma\choose n-k-\gamma}$. Thus, for any such $\gamma$ we have $T_{\gamma} = [k+1,k+\gamma]$ and $S_{\gamma} = [2,k]\cup \{\gamma\}$. Condition (1) in Definition \[defresnu\] is satisfied if $\gamma\le k-1$. Condition (2) is satisfied iff $k+\gamma>2\gamma+2$, which is equivalent to $\gamma\le k-3$. (Note also that $\gamma=1$ is resistant for $k=3$.) From the discussion above, we also conclude that, for $k>3$, we have $\gamma_{k-2} = {n-k\choose n-k-1}=n-k.$
The following observation is given for the sake of familiarizing the reader with the statement.
The bound in Theorem \[thmfull1\] is always at least as strong as the bound in Theorem \[thm1\] for intersecting ${\mathcal F}\subset {[n]\choose k}$ with diversity $\gamma({\mathcal F})\le {n-4\choose k-3}$.
Let us first compare the statement of Theorem \[thmfull1\] with the statement of Theorem \[thm1\] for $\gamma_l := {n-u-1\choose n-k-1}$ with integer $u$. Such $\gamma_l$ is resistant for any $u\in [3,k]$, and we have $A_{\gamma_l}=[2,u+1]$. Thus, if we substitute such $\gamma_l$ in , then we get the bound $$|{\mathcal F}|\le {n-2\choose n-k}+\ldots +{n-u-1\choose n-k-u+1}+\gamma_l = {n-1\choose n-k}-{n-u-1\choose n-k-u}+{n-u-1\choose n-k-1},$$ which is exactly the bound for $\gamma_l$.
Remark that the inequality , however, gives a stronger conclusion and is valid in weaker assumptions. Indeed, while we know that this bound is sharp for $\gamma({\mathcal F}) = \gamma_l$, Theorem \[thmfull1\] also tells us that ${\mathcal F}$ with $\gamma({\mathcal F})>\gamma_l$ has strictly smaller size (quantified in Corollary \[corhm\]). Moreover, even if $\gamma_{l-1}<\gamma({\mathcal F})\le \gamma_l$, we are still getting the same upper bound.
Returning to the proof of the observation, assume now that $u$ is a real number. The function in the right hand side of is monotone decreasing as $\gamma({\mathcal F})$ increases (or, equivalently, as $u$ decreases). Therefore, to show that the bound is stronger than , it is sufficient to verify it for all values of $\gamma_l$, $l=0,\ldots, m$. But for each of these values the bound is sharp, so can be only weaker than .
Here, we prove Corollary \[corhm\] for (integer) $u\ge 4$. Choose $l$ such that $\gamma_l={n-u-1\choose n-k-1}$. Assume first that $k\ge 5$. Then $\gamma_{l+1}={n-u-1\choose n-k-1}+{n-k-2\choose n-k-2}={n-u-1\choose n-k-1}+1=:\gamma$. To see this, we just have to check that $\gamma$ is a resistant number. In terms of Definition \[defresnu\], we have $T_{\gamma}=\{u+1,k+2\}$ and $S_{\gamma}=\{2,3,\ldots, u,u+2\ldots, k+2\}.$ We have $s_b=2\le k-1$ and $s_a=k$, thus, condition (1) of Definition \[defresnu\] is fulfilled. Moreover, $b_1=u+1>2\cdot 1+2$ and $b_2=k+2>2\cdot 2+2$, fulfilling condition (2) of the definition. Thus, implies that
$$\begin{aligned}
|{\mathcal F}|\le& {n-2\choose n-k}+\ldots+{n-u\choose n-k-u+2}+{n-u-2\choose n-k-u+1}+\ldots+{n-k-2\choose n-2k+1}+{n-u-1\choose n-k-1}+1\\
=&{n-1\choose n-k}-{n-u-1\choose n-u-k}+{n-u-1\choose n-k-1}+1-\epsilon,\end{aligned}$$
where
$$\epsilon={n-u-1\choose n-k-u+1}-\Big({n-u-2\choose n-k-u+1}+\ldots+{n-k-2\choose n-2k+1}\Big)={n-k-2\choose n-2k}={n-k-2\choose k-2}.$$
If $k=4$ then $u=4$ and $\gamma_{l+1}={n-4\choose k-3}={n-4\choose 1}$. We get that
$$|{\mathcal F}|\le {n-1\choose k-1}-{n-4\choose k-1}+{n-4\choose k-3}={n-1\choose k-1}-{n-5\choose k-1}+{n-5\choose k-4}-\epsilon,$$
where
$$\epsilon = {n-5\choose k-2}-{n-5\choose k-3}={n-5\choose 2}-{n-5\choose 1}={n-6\choose 2}-1.$$
The second part of the corollary is proved in Section \[sec33\].
Our next goal is to state the “conceptual” version of Theorem \[thmfull1\]. It requires certain preparations. We will use the framework and some of the ideas from [@FK1], as well as from [@KZ]. First of all, we switch to the cross-intersecting setting. Given an intersecting family ${\mathcal F}$ with $\Delta({\mathcal F})=\delta_1({\mathcal F})$, consider the families $$\begin{aligned}
{\mathcal F}(1):=&\{F\setminus\{1\}\ :\ 1\in F\in{\mathcal F}\}\ \ \ \ \ \text{and}\\
{\mathcal F}(\bar 1):=&\{F\ :\ 1\notin F\in{\mathcal F}\}.\end{aligned}$$ Remark that $\gamma({\mathcal F})=|{\mathcal F}(\bar 1)|$. Applying Theorem \[thmHil\], from now on and until the end of Section \[sec3\] we assume that ${\mathcal F}(1)=\mathcal L([2,n],|{\mathcal F}(1)|,k-1)$ and ${\mathcal F}(\bar 1) = \mathcal L([2,n],|{\mathcal F}(\bar 1)|, k)$. Note that ${\mathcal F}(1),{\mathcal F}(\bar 1)\subset 2^{[2,n]}$. For shorthand, we denote ${{\mathcal A}}:={\mathcal F}(1), {{\mathcal B}}:={\mathcal F}(\bar 1)$. While proving Theorem \[thmfull1\], we will work with the ground set $[2,n]$, in order not to confuse the reader and to keep clear the relationship between the diversity of intersecting families and the sizes of pairs of cross-intersecting families.+0.1cm
Both ${{\mathcal A}}$ and ${{\mathcal B}}$ are determined by their lexicographically last set. In this section, we use the lexicographical/containment order on $2^{[2,n]}$, which is defined as follows: $A\prec B$ iff $A\supset B$ or the minimal element of $A\setminus B$ is smaller than the minimal element of $B\setminus A$. Let us recall some notions and results from [@FK1] related to the Kruskal–Katona theorem and cross-intersecting families. For sets $S$ and $X$, $|S\cap X|\le a$, we define $$\mathcal L(X,S,a):=\Big\{A\in {X\choose a}\ :\ A\prec S\cap X\Big\}.$$ For example, the family $\{G\in{[2,n]\choose 10}\ :\ 2\in G, G\cap\{3,4\}\ne \emptyset\}$ is the same as the family $\mathcal L([2,n],S,10)$ for $S=\{1,2,4\}$. If ${\mathcal G}=\mathcal L(X,S,a)$ for a certain set $S$, then we say that $S$ is the of ${\mathcal G}$.
Note that, in what follows, we shall work with $X=[2,n]$ and will omit $X$ from the notation for shorthand. For convenience, we shall assume that $1\in S$ (motivated by the fact that $S$ will stand for the characteristic set for the subfamily of all sets containing 1 in the original family), while $T\subset [2,n]$.
\[defrespa\] We say that two sets $S\subset [n]$ and $T\subset [2,n]$ , if either $T =\{2,3,4\}$ and $S = \{1,4\}$, or the following holds: assuming that the largest element of $T$ is $j$, we have
1. $S\cap T = \{j\},\ S\cup T = [j]$, $|S|\le k,$ $|T|\le k$;
2. for each $ i\ge 4$, we have $|[i]\cap S|< |[i]\setminus S|$.
Condition (2), roughly speaking, states that in $[i]$ there are more elements in $T$ than in $S$. Note that 2 implies that $T\supset \{2,3,4\}$ for each resistant pair. There is a close relationship between this notion and the notion of a resistant number, which we discuss a bit later. Let us first give the characteristic set version of Theorem \[thmfull1\]. For convenience, we put $T_0 = [2,n]$ to be the characteristic set of the empty family and $S_0:=\{1,n\}$ to be the characteristic set of the family ${[2,n]\choose a}$.
\[thmfull2\] Let $n>2k\ge6$. Consider all resistant pairs $S_l\subset [n],\ T_l\subset [2,n]$, where $l\in [m]$. Assume that $T_0<T_1<T_2<\ldots <T_m$. Then $$\label{eqfull2} |\mathcal L(S_{l-1},k-1)|+|\mathcal L(T_{l-1},k)|>|\mathcal L(S_{l},k-1)|+|\mathcal L(T_{l},k)| \ \ \ \text{ for each }l \in [m],$$ and any cross-intersecting pair of families $\mathcal A\subset {[2,n]\choose k-1},\ \mathcal B\subset {[2,n]\choose k}$ with $|\mathcal L(T_{l-1},k)|<|{{\mathcal B}}|\le |\mathcal L(T_l,k)|$ satisfies $$\label{eqfull3} |{{\mathcal A}}|+|{{\mathcal B}}|\le |\mathcal L(S_l,k-1)|+|\mathcal L(T_l, k)|.$$
In terms of intersecting families, if ${\mathcal F}\subset {[n]\choose k}$ is intersecting and $|\mathcal L(T_{l-1},k)|<\gamma({\mathcal F})\le |\mathcal L(T_l,k)|$, then $|{\mathcal F}|\le |\mathcal L(S_l,k-1)|+|\mathcal L(T_l, k)|$.
First, we remark that the intersecting part is clearly equivalent to the second statement of the cross-intersecting part. Second, Proposition \[prop9\] below shows that the families $L(S_l,k-1)$ and $L(T_{l},k)$ are cross-intersecting and thus is sharp.
We say that two sets $S$ and $T$ in $[2,n]$ , if there exists a positive integer $j$ such that $S\cap T\cap[2,j]=\{j\}$ and $S\cup T\supset [2,j]$. The following proposition was proven in [@FK1]:
\[prop9\] Let $A$ and $B$ be subsets of $[2,n]$, $|A|\le a, |B|\le b$, and $|[2,n]|=n-1\ge a+b$. Then $\mathcal L(P,a)$ and $\mathcal L(Q,b)$ are cross-intersecting iff $P$ and $Q$ strongly intersect.
Now let us deduce Theorem \[thmfull1\] from Theorem \[thmfull2\].
Let us first compute the size of the family $\mathcal L(T, k-1)$ for a given set $T\in [2,n]$. Assume that $b_{s_b}$ is the largest element of $T$ and that $[2,b_{s_b}]\setminus T= \{b_1,\ldots, b_{s_b-1}\}$. Take the smallest element $b_1\in [2,n]\setminus T$ and consider the family with characteristic set $T_1:=(T\cap [b_1])\cup \{b_1\}$. Note that $T_1 = [2,b_1]$ and thus the size of this family is ${n-b_1\choose k-b_1+1} = {n-b_1\choose n-k-1}$. Since $T_1\prec T$, this family is a subfamily of $\mathcal L(T, k-1)$. Proceeding iteratively, at each step take the next smallest (not chosen yet) element $b_i$ from $[2,n]\setminus T$ and define the set $T_i:=(T\cap [b_i])\cup b_i$. Again, $T_i\prec T$. Count the sets that belong to $\mathcal L(T_i,k-1)\setminus L(T_{i-1},k-1)=\big\{F\in {[2,n]\choose k}\ :\ F\cap [b_i]=T_i\big\}$. Their number is precisely ${n-b_i\choose k-|T_i|} = {n-b_i\choose n-k-|[b_i]\setminus T_i|}$. Since we “stop” at every element that is not included in $T$, we get that $|[b_i]\setminus T_i| = |[b_{i-1}\setminus T_{i-1}|+1 = \ldots = i$. Therefore, ${n-b_i\choose n-k-|[b_i]\setminus T_i|}={n-b_i\choose n-k-i}$. We stop the procedure at the point when $T_i = T$, including the sets $F\in {[2,n]\choose k-1}$ that satisfy $F\cap [b_{s_b}] = T$ in the count. It should be clear that, in this counting procedure, we counted each set from $\mathcal L(T,k-1)$ exactly once. We get that $$|\mathcal L(T,k-1)| = {n-b_1\choose n-k-1}+{n-b_2\choose n-k-2}+\ldots +{n-b_{s_b}\choose n-k-s_b},$$ and the displayed formula gives the representation of $|\mathcal L(T,k-1)|$ in the $(n-k-1)$-cascade form! Moreover, we conclude that the set $\{b_1,\ldots, b_{s_b}\}$ is exactly the set $T_{\gamma}$ for $\gamma:=|\mathcal L(T,k-1)|$ (cf. the paragraph above Definition \[defresnu\]). We have $T_{\gamma} = ([2,b_{s_b}]\setminus T)\cup \{b_{s_b}\}$ and thus $T = S_{\gamma}$.
Therefore, if $S,T$ is a resistant pair, then, putting $\gamma:=|\mathcal L(T,k-1)|$, we get that $T = S_{\gamma}$ and $S\cap [2,n] = T_{\gamma}$. This immediately implies that $T_{\gamma}$ and $S_{\gamma}$ satisfy condition (1) of Definition \[defresnu\]. The implication in the other direction follows as well. Condition (2) of Definition \[defrespa\] is equivalent to the statement that $1+|T_{\gamma}\cap [i]|<|[2,i]\setminus T_{\gamma}|$ for each $i\ge 4$, which, in turn, is equivalent to $b_l>2l+2$ for each $l\ge 1$. Finally, it is clear that $\gamma = {n-4\choose k-3}$ correspond to the characteristic set $\{2,3,4\}$.
We conclude that $T_l$ and $S_l$ form a resistant pair if and only if $|\mathcal L(T_l,k-1)|$ is a resistant number. Doing calculations as above, one can conclude that $$|\mathcal L(S_l,k)| = {n-a_1\choose n-k}+{n-a_{2}\choose n-k-1}+\ldots +{n-a_{s_a}\choose n-k-s_a+1},$$ where $a_i$ are as in the statement of Theorem \[thmfull1\]. Given that, it is clear that the inequality is equivalent to the statement saying that the right hand side of is strictly monotone, and that is equivalent to .
Finally, the sharpness claimed in Theorem \[thmfull1\] immediately follows from the remark in the paragraph after Theorem \[thmfull2\].
Proof of Theorem \[thmfull2\]
-----------------------------
We say that $\mathcal A\subset {[2,n]\choose a}$ and $\mathcal B\subset {[2,n]\choose b}$ form a , if, whenever $\mathcal A'\subset {[2,n]\choose a}$ and $\mathcal B'\subset {[2,n]\choose b}$ are cross-intersecting with $\mathcal A'\supset \mathcal A$ and $\mathcal B'\supset \mathcal B$, then necessarily $\mathcal A = \mathcal A'$ and $\mathcal B = \mathcal B'$ holds.
The following proposition from [@FK1] is another important step in our analysis.
\[cross2\] Let $a$ and $b$ be positive integers, $a+b\le n-1$. Let $P$ and $Q$ be non-empty subsets of $[2,n]$ with $|P|\le a$, $|Q| \le b$. Suppose that $P$ and $Q$ strongly intersect in their largest element, that is, there exists $j$ such that $P\cap Q = \{j\}$ and $P\cup Q = [2,j]$. Then $\mathcal L([2,n],P,a)$ and $\mathcal L([2,n],Q,b)$ form a maximal pair of cross-intersecting families.
Inversely, if $\mathcal L([2,n],m,a)$ and $\mathcal L([2,n],r,b)$ form a maximal pair of cross-intersecting families, then there exist sets $P$ and $Q$ that strongly intersect in their largest element, such that $\mathcal L([2,n],m,a)=\mathcal L([2,n],P,a)$, $\mathcal L([2,n],r,b)=\mathcal L([2,n],Q,b)$.
Recall that we aim to maximize $|{{\mathcal A}}|+|{{\mathcal B}}|$ given a lower bound on $|{{\mathcal B}}|$. The proof is based on the following two lemmas.
\[lemdiv\] Consider a pair of cross-intersecting families ${{\mathcal A}}\subset {[2,n]\choose k-1},\ {{\mathcal B}}\subset {[2,n]\choose k}$. Suppose that ${{\mathcal A}}= \mathcal L(S,k-1),\ {{\mathcal B}}=\mathcal L(T,k)$ for some sets $S\subset [n]$, $T\subset [2,n]$ that strongly intersect in their last element $j$. Suppose also that $T\precneqq \{2,3,4\}$.
Assume that $S$ and $T$ do not form a resistant pair, that is, there exists $5\le i\le j$, such that $\big|[i]\cap S\big|\ge \big|[i]\setminus S\big|$. Put $T':=[i]\setminus S$ and choose $S'$ so that it strongly intersects with $T'$ in its largest element. Then the families ${{\mathcal A}}'\subset {[2,n]\choose k-1},\ {{\mathcal B}}'\subset {[2,n]\choose k}$ with characteristic sets $S', T'$ are cross-intersecting and satisfy $|{{\mathcal A}}'|+|{{\mathcal B}}'|\ge |{{\mathcal A}}|+|{{\mathcal B}}|$ and $|{{\mathcal B}}'|> |{{\mathcal B}}|$.
Moreover, if $\big|[i]\cap S\big|>\big|[i]\setminus S\big|$ then $|{{\mathcal A}}'|+|{{\mathcal B}}'|> |{{\mathcal A}}|+|{{\mathcal B}}|$.
First, recall that $1\in S$. Since $S'$ and $T'$ are strongly intersecting, the families ${{\mathcal A}}'$, ${{\mathcal B}}'$ are cross-intersecting. Next, clearly, $T'\subsetneq T$ and thus ${{\mathcal B}}'\supsetneq {{\mathcal B}}$. Therefore, we only have to prove that $|{{\mathcal A}}|+|{{\mathcal B}}|\le |{{\mathcal A}}'|+|{{\mathcal B}}'|$, and that the inequality is strict in the case indicated in the lemma.
Consider the following families: $$\begin{aligned}
\mathcal P_a:=&\big\{P\in {[2,n]\choose k-1}\ :\ P\cap [i] = [2,i]\cap S\big\},\\
\mathcal P_b:=&\big\{P\in {[2,n]\choose k}\ :\ P\cap [i] = [i]\setminus S\big\}.\end{aligned}$$ We have ${{\mathcal A}}\setminus \mathcal P_a = {{\mathcal A}}',$ ${{\mathcal B}}\cup \mathcal P_b = {{\mathcal B}}'$. Both equalities are proved in the same way, so let us show, e.g., the first one. We have $S'\prec S\prec S\cap [i]$, therefore ${{\mathcal A}}'\subset {{\mathcal A}}\subset \mathcal L(S\cap [i],k-1)$. On the other hand, we claim that $S'$ and $S\cap [i]$ are two consecutive sets in the lexicographic order on $[i]$. Indeed, assume that the largest element of $T'$ is $j'$. If $j' = i$, then $S'\supset S\cap [i]$, $(S'\setminus S)\cap [i] = \{i\}$, which proves it in this case. If $j'<i$, then $[j'+1,i]\subset S\cap [i]$, $j'\notin S\cap [i]$. It is easy to see that the set that precedes $S\cap [i]$ in the lexicographic order on $[i]$ “replaces” $[j'+1,i]$ with $\{j'\}$, that is, it is $S'$. Therefore, $${{\mathcal A}}\setminus {{\mathcal A}}' \subset \mathcal L(S\cap [i],k-1)\setminus {{\mathcal A}}' = \mathcal L(S\cap [i],k-1)\setminus \mathcal L(S',k-1) = \mathcal P_a,$$ which, together with the fact that $\mathcal P_a$ and ${{\mathcal A}}'$ are disjoint, is equivalent to the equality we aimed to prove.
Next, consider the bipartite graph $G$ with parts $\mathcal P_a,\mathcal P_b$ and edges connecting disjoint sets. Then, due to the fact that ${{\mathcal A}}$ and ${{\mathcal B}}$ are cross-intersecting, $({{\mathcal A}}\cap \mathcal P_a)\cup ({{\mathcal B}}\cap \mathcal P_b)$ is an independent set in $G$.
The graph $G$ is biregular, and therefore the largest independent set in $G$ is one of its parts. We have $|\mathcal P_a| = {n-i\choose s_a}$, $|\mathcal P_b| = {n-i\choose s_b}$, where $s_a = k-|[i]\cap S|,\ s_b = k-|[i]\setminus S|$. By the condition from the lemma, we have $s_b\ge s_a$, and, since $n-i> s_a+s_b$, we have $|\mathcal P_b|\ge |\mathcal P_a|$. We conclude that $|\mathcal P_b|$ is the largest independent set in $G$, so $|\mathcal P_b|\ge ({{\mathcal A}}\cap \mathcal P_a)\cup ({{\mathcal B}}\cap \mathcal P_b),$ and therefore $$|{{\mathcal A}}'|+|{{\mathcal B}}'|-(|{{\mathcal A}}|+|{{\mathcal B}}|) = |\mathcal P_b|-|{{\mathcal A}}\cap \mathcal P_a|-|{{\mathcal B}}\cap \mathcal P_b|\ge 0.$$ If $\big|[i]\setminus S\big|< \big|S\cap [i]\big|$, then $|\mathcal P_b|> |\mathcal P_a|$ and $\mathcal P_b$ is the [*unique*]{} independent set of maximal size in $G$. Thus, we have strict inequality in the displayed inequality above.
Slightly abusing notation, we say that ${{\mathcal A}}, {{\mathcal B}}$ form a resistant pair if the corresponding characteristic sets form a resistant pair. The second lemma describes how do the resistant pairs behave. More specifically, it shows that holds: the sum of sizes of resistant cross-intersecting families increases as the size of the second family decreases.
\[lemres\] Consider a resistant pair of cross-intersecting families ${{\mathcal A}}\subset{[2,n]\choose k-1},\ {{\mathcal B}}\subset {[2,n]\choose k}$, with characteristic sets $S\subset [n],T\subset [2,n]$, respectively, and another such resistant pair ${{\mathcal A}}'\subset{[2,n]\choose k-1},\ {{\mathcal B}}'\subset {[2,n]\choose k}$ with characteristic sets $S'\subset [n],T'\subset [2,n]$. If $T'\precneqq T$, then $|{{\mathcal B}}'|< |{{\mathcal B}}|$ and $|{{\mathcal A}}|+|{{\mathcal B}}|<|{{\mathcal A}}'|+|{{\mathcal B}}'|$.
Therefore, while for general lexicographic pairs of families the sum of sizes is not monotone w.r.t. the size of the second family, it is monotone for resistant pairs.
Remark that, since $T'\precneqq T$, we have $T'\ne \{2,3,4\}$ and thus $S',T'$ must satisfy Condition (2) of Definition \[defrespa\]. We also note that we do not use the property that $S,T$ form a resistant pair. The proof also works for $T' = T_0$ (recall that $T_0 =[2,n]$).
The proof of this lemma is based on biregular bipartite graphs and is very similar to the proof of Lemma \[lemdiv\], although is a bit trickier.
First, it is clear that, in the conditions of the lemma, we have $|{{\mathcal B}}'|< |{{\mathcal B}}|$. The rest of the proof is concerned with the inequality on the sums of sizes. We consider two cases depending on how do the sets $T$ and $T'$ relate.\
**Case 1.** Note that this condition, in particular, implies that $T\ne \{2,3,4\}$. Find the smallest $i\ge 5$, such that exactly one of the sets $T,T'$ contain $i$. Since $T'\prec T$, we clearly have $i\in T',\ i\notin T$. Consider the set $T'' = T'\cap [i]$. Then we clearly have $T'\precneqq T''\precneqq T$ and $T''\subset T'$. Accordingly, put $S''$ to be $\{i\}\cup ([i]\setminus T'')$, and consider the cross-intersecting families ${{\mathcal A}}''\subset {[2,n]\choose k-1},\ {{\mathcal B}}''\subset {[2,n]\choose k}$, which have characteristic vectors $S''$ and $T''$, respectively. Note that $S''=S\cap [i]$ and thus the pair ${{\mathcal A}}'', {{\mathcal B}}''$ is resistant. We claim that $|{{\mathcal A}}''|+|{{\mathcal B}}''|> |{{\mathcal A}}|+|{{\mathcal B}}|.$
We prove the inequality above as in Lemma \[lemdiv\], but the roles of $S$ and $T$ are now switched. Consider the bipartite graph $G$ with parts
$$\begin{aligned}
\mathcal P_a:=&\big\{P\in {[2,n]\choose k-1}\ :\ P\cap [i] = [2,i]\setminus T\big\},\\
\mathcal P_b:=&\big\{P\in {[2,n]\choose k}\ :\ P\cap [i] = [i]\cap T\big\},\end{aligned}$$
and edges connecting disjoint sets. Similarly, we have ${{\mathcal A}}\cup \mathcal P_a = {{\mathcal A}}'',$ ${{\mathcal B}}\setminus \mathcal P_b = {{\mathcal B}}''$. Indeed, let us verify, e.g., the second equality. All $k$-element sets $P$ such that $P\prec T''$ are in ${{\mathcal B}}$ and in ${{\mathcal B}}\setminus \mathcal P_b$, as well as in ${{\mathcal B}}''$, since $T''\precneqq T\cap [i]$. On the other hand, if we restrict to $[i]$, the sets $T\cap [i]$ and $T''$ are consecutive in the lexicographic order, and so any set $B$ from ${{\mathcal B}}$ such that $B\succneqq T''$ must satisfy $B\cap [i] = T\cap [i]$. Therefore, ${{\mathcal B}}\setminus {{\mathcal B}}''\subset \mathcal P_b$ and ${{\mathcal B}}\setminus \mathcal P_b = \mathcal B''$.
The families ${{\mathcal A}}$ and ${{\mathcal B}}$ are cross-intersecting, and so the set $({{\mathcal A}}\cap \mathcal P_a)\cup ({{\mathcal B}}\cap \mathcal P_b)$ is independent in $G$. On the other hand, the largest independent set in $G$ has size $\max\{|\mathcal P_a|, |\mathcal P_b|\}$. Since the pair ${{\mathcal A}}',$ ${{\mathcal B}}'$ is resistant, we have that $|[i]\cap T|=|[i]\setminus S''| > |[i]\cap S''|=|[i]\setminus T|$, which implies $|\mathcal P_a|={n-i\choose k-|[i]\setminus T|}> {n-i\choose k-|[i]\cap T|} = |\mathcal P_b|$, and thus $\mathcal P_a$ is the (unique) largest independent set in $G$. We have
$$|{{\mathcal A}}''|+|{{\mathcal B}}''|-(|{{\mathcal A}}|+|{{\mathcal B}}|) = |\mathcal P_a|-|{{\mathcal A}}\cap \mathcal P_a|-|{{\mathcal B}}\cap \mathcal P_b|>0,$$ and the desired inequality is proven. Therefore, when comparing $T'$ and $T$, we may replace $T$ with $T''$, or rather assume that $T\subset T'$. We have reduced Case 1 to the following case.\
**Case 2. [^6]** Assume that $t_1<t_2<\ldots <t_l$ are the elements forming the set $T'\setminus T$ and let $t_0$ be the largest element of $T$. Let us first show that, for each $i\in [l-1]$, we have $$|[t_i]\setminus S'|\ge |[t_i]\cap S'|+2.$$ Indeed, find the largest element $t'<t_i,$ such that $t'\in S'$. We have $|[t_i]\setminus S'|>|[t']\setminus S'|>|[t']\cap S'|=|[t_i]\cap S'|$. For $i=0,\ldots, l$, put $S_i:=[t_i]\cap S'\cup \{t_i\}$ and $T_i:=T'\cap [t_i]$. Observe that, for each $i\in[l]$, we have $|[t_i]\setminus S_i|\ge |[t_i]\cap S_i|$ by the displayed inequality. Moreover, the inequality is strict for $i=l$. For each $i=0,\ldots, l$, let ${{\mathcal A}}_i\subset {[2,n]\choose k-1},{{\mathcal B}}_i\subset {[2,n]\choose k}$ be the pair of cross-intersecting families defined by characteristic sets $S_i,T_i$. Observe that ${{\mathcal A}}_0={{\mathcal A}},{{\mathcal B}}_0={{\mathcal B}}$ and ${{\mathcal A}}_l={{\mathcal A}}',{{\mathcal B}}_l={{\mathcal B}}'$.
Finally, for each $i\in [l]$, we show that $|{{\mathcal A}}_i|+|{{\mathcal B}}_i|\ge |{{\mathcal A}}_{i-1}|+|{{\mathcal B}}_{i-1}|$, moreover, the inequality is strict for $i=l$. This is clearly sufficient to conclude the proof of the lemma. Consider the bipartite graph $G$ with parts $$\begin{aligned}
\mathcal P_a^i:=&\big\{P\in {[2,n]\choose k-1}\ :\ P\cap [i] = S_i\cap [2,i]\big\},\\
\mathcal P_b^i:=&\big\{P\in {[2,n]\choose k}\ :\ P\cap [i] = [i]\setminus S_i\big\},\end{aligned}$$ and edges connecting disjoint sets. As before, we have ${{\mathcal A}}_{i-1}\cup \mathcal P_a^i = {{\mathcal A}}_i,$ ${{\mathcal B}}_{i-1}\setminus \mathcal P_b^i = {{\mathcal B}}_i$. Using the fact that $|[i]\setminus S_i|\ge|[i]\cap S_i|$ and that this inequality is strict for $i=l$, we conclude the proof as in the previous case.
Now let us put the things together.
First, follows from Lemma \[lemres\]. Next, given a pair of cross-intersecting families ${{\mathcal A}},{{\mathcal B}}$ as in the theorem, we may assume using Theorem \[thmHil\] and Proposition \[cross2\] that ${{\mathcal A}}, {{\mathcal B}}$ are lexicographic families defined by characteristic sets $S,T$ that strongly intersect in their last coordinate. We may further assume that they do not form a resistant pair. Using Lemma \[lemdiv\] with the smallest $i$ satisfying its conditions, replace ${{\mathcal A}},{{\mathcal B}}$ with the corresponding pair ${{\mathcal A}}', {{\mathcal B}}'$ defined by characteristic sets $S',T'$. Remark that $T\precneqq T'$ and that, moreover, ${{\mathcal A}}',{{\mathcal B}}'$ form a resistant pair by the choice of $i$. (Note here that if $i=5$ then the resulting characteristic sets are $T'=T_m=\{2,3,4\}$ and $S'=S_m=\{1,4\}$.) Therefore, if $T_{l-1}\precneqq T$ then $T_l\prec T'$, and therefore $|\mathcal L(S_l,k-1)|+|\mathcal L(T_l,k)|\ge |{{\mathcal A}}'|+|{{\mathcal B}}'|\ge |{{\mathcal A}}|+|{{\mathcal B}}|$. This completes the proof of the theorem.
Equality in Theorems \[thmfull1\], \[thmfull2\] and families with large diversity {#sec33}
---------------------------------------------------------------------------------
In the notations of the previous section, let ${{\mathcal A}},{{\mathcal B}}$ be cross-intersecting and defined by the characteristic sets $S,T$, respectively, where $|S|,|T|\le k$. In this section, we determine, for which $T$, $T_{l-1}\prec T\prec T_l$, it is possible to have equality in .
\[defneutral\] We say that a pair of strongly intersecting sets $S,\ T$ as above is , if $T$ is obtained in the following recursive way:
1. $T_l$ is $T_l$-neutral;
2. If $T'$ is $T_l$-neutral, then the set $T:=T'\cup \{2|T'|\}$ is $T_l$-neutral.
In other words, to form all $T_l$-neutral sets, we start from the set $T_l$, add the element $2|T_l|$ and then continue adding every other element until we have $k$ elements in the set.
It is not difficult to see that any $T_l$-neutral pair $S, T$ actually satisfies $T_{l-1}\precneqq T\prec T_l$. Let us also note that, in terms of Definition \[defneutral\], each newly formed $T_l$-neutral set is different from the previous one: indeed, from Definition \[defrespa\], the largest element in $T_l$ is at most $2|T_l|-2$ (actually, it is at most $2|T_l|-3$ for all $l<m$ and equal to $2|T_l|-2$ in the case $l=m$), and every newly added element (by Condition (2) of Definition \[defneutral\]) is bigger by 2 than the previously added element.
\[thmfulleq\] Let $n>2k\ge6$. Consider a pair ${{\mathcal A}}\subset {[2,n]\choose k-1},\ {{\mathcal B}}\subset{[2,n]\choose k}$ defined by strongly intersecting sets $S, T$ that intersect in their largest element. If $T_{l-1}<T \le T_l$ for some $l\in[m]$, then equality in holds if and only if the pair $S, T$ is $T_l$-neutral.
First, let us show that, for any $T_l$-neutral pair, we have equality in . We prove it inductively. It is clear for the pair with characteristic sets $S,T$, where $T=T_l$. Assuming it holds for $T'$, let us prove that it holds for $T:=T'\cup \{2|T'|\}$.
Put $x:=2|T|$ and consider the pairs of cross-intersecting families ${{\mathcal A}}, {{\mathcal B}}$, ${{\mathcal A}}', {{\mathcal B}}'$, corresponding to the $T_l$-neutral pairs of sets $S, T$ and $S', T'$, respectively. By Definition \[defneutral\], we have $$|[x]\cap S| = |[x]\setminus S|.$$ Therefore, applying the argument of Lemma \[lemdiv\] with $$\begin{aligned}
\mathcal P_a:=&\big\{P\in {[2,n]\choose k-1}\ :\ P\cap [x] = [2,x]\cap S\big\},\\
\mathcal P_b:=&\big\{P\in {[2,n]\choose k}\ :\ P\cap [x] = [x]\setminus S\big\},\end{aligned}$$ We get that $k-1-|[2,x]\cap S| = k-|[x]\setminus S|$, which implies $|\mathcal P_a| = |\mathcal P_b|$. Moreover, ${{\mathcal A}}\setminus {{\mathcal A}}' = \mathcal P_a,$ ${{\mathcal B}}\setminus {{\mathcal B}}' = \mathcal P_b,$ since the sets $S'$ and $S$ are both subsets of $[x]$ and are consecutive in the lexicographical order on $[x]$ (and the same holds for $T, T'$). Therefore, $|{{\mathcal A}}'|+|{{\mathcal B}}'| = |{{\mathcal A}}|+|{{\mathcal B}}|$.\
In the other direction, take a set $T$, $T_{l-1}\precneqq T\precneqq T_l$, and its pair $S$. Consider the corresponding pair of cross-intersecting families ${{\mathcal A}}, {{\mathcal B}}$ and assume that they satisfy equality in . Then it is easy to see that $T\supset T_l$. (Otherwise, either $T\succ T_l$, or $T$ must contain and thus succeed some other resistant set, which succeeds$T_l$, again contradicting $T\precneqq T_l$.) Assuming that $x$ is the last element in $T$, we must have $$|[i]\cap S|<|[i]\setminus S| \ \ \ \ \ \ \ \text{for }5\le i\le x-1.$$ Indeed, otherwise, considering the bipartite graph $G$ with parts $\mathcal P_a$, $\mathcal P_b$ as displayed above for that $i$, we would get that $|\mathcal P_a|\le |\mathcal P_b|$ and both $\mathcal P_a\cap {{\mathcal A}}$ and $\mathcal P_b\cap {{\mathcal B}}$ are non-empty. In this case $|\mathcal P_a\cap {{\mathcal A}}|+|\mathcal P_b\cap {{\mathcal B}}|<|\mathcal P_a|$, which means that the pair ${{\mathcal A}}', {{\mathcal B}}'$ defined by the characteristic sets $T':=T\cap [i]$ and its pair $S'$ would satisfy $|{{\mathcal A}}'|+|{{\mathcal B}}'|>|{{\mathcal A}}|+|{{\mathcal B}}|$. Moreover, $T'\supset T_l$, so $T\precneqq T'\prec T_l$ and $|{{\mathcal A}}'|+|{{\mathcal B}}'|\le |\mathcal L(S_l,k-1)|+|\mathcal L(T_l,k)|$. This contradicts the equality for ${{\mathcal A}},{{\mathcal B}}$ in .
Therefore, since the pair $S, T$ is not resistant, we have $$|[x]\cap S|=|[x]\setminus S|.$$ (We cannot have “$>$”, since otherwise we would have “$\ge$” for $i=x-1$.) Removing $x$, we get a set $T'$, and conclude that $x = 2|T'|$. By induction on the size of the set $T$, we may assume that $T'$ is $T_l$-neutral. But then $T$ is $T_l$-neutral as well.
A slight modification of the argument above leads to the proof of the second part of Corollary \[corhm\].
In the framework of the previous proofs, replace ${\mathcal F}(1)$, ${\mathcal F}(\bar 1)$ with lexicographical families and put ${{\mathcal A}}:={\mathcal F}(1)$, ${{\mathcal B}}:={\mathcal F}(\bar 1)$. For $i\in [3]$, let ${{\mathcal A}}_i,{{\mathcal B}}_i$ be the pairs of cross-intersecting families defined by characteristic sets $T^i,S^i$, where $T^1=T_m=\{2,3,4\}$, $T^2=\{2,3\}$ and $T^3=\{2,4\}$. It is easy to see that $|{{\mathcal B}}_1|={n-4\choose k-3},\ |{{\mathcal B}}_2|={n-3\choose k-2}$ and $|{{\mathcal B}}_3|={n-3\choose k-2}+{n-4\choose k-2}$. If $\gamma({\mathcal F})<{n-3\choose k-2}$ then consider the graph $G$ with parts $\mathcal P_a,\mathcal P_b$, where the parts contains the $k$-element sets that intersect $[4]$ in $\{1,4\}$ and in $\{2,3\}$, respectively. Then $|\mathcal P_a|=|\mathcal P_b|$ and ${{\mathcal A}}_1\setminus {{\mathcal A}}_2=\mathcal P_a$, ${{\mathcal B}}_2\setminus {{\mathcal B}}_1 = \mathcal P_b$. Moreover, $(\mathcal P_a\cap {{\mathcal A}})\cup (\mathcal P_b\cap {{\mathcal B}})$ is independent in $G$. Thus, if ${{\mathcal A}}\notin \{{{\mathcal A}}_1,{{\mathcal A}}_2\}$ then the aforementioned independent set is strictly smaller than $\mathcal P_a,\mathcal P_b$, and thus $|\mathcal A|+|\mathcal B|<|\mathcal A_i|+|\mathcal B_i|$ for $i=1,2$.
Furthermore, provided that both $\mathcal P_a\cap {{\mathcal A}}$ and $\mathcal P_b\cap {{\mathcal B}}$ are non-empty, then $$|\mathcal P_a\cap {{\mathcal A}}|+|\mathcal P_b\cap {{\mathcal B}}|\le {n-4\choose k-2}-{n-k-2\choose k-2}+1.$$ This is easy to deduce from the Kruskal–Katona theorem and was proven already by Hilton and Milner [@HM]. For a more general statement, see Corollary \[corkz\] in Section \[sec34\]. Since $\gamma({\mathcal F})>{n-4\choose k-3}$, $\mathcal P_b\cap \mathcal {{\mathcal B}}$ is non-empty. If ${{\mathcal A}}$ is non-empty then the displayed bound holds, and thus is valid. If ${{\mathcal A}}$ is empty then, using $|{{\mathcal B}}|\le {n-4\choose k-2}-{n-k-2\choose k-2}+1$ (that immediately follows from $|{\mathcal F}(\bar 1)|\le {n-3\choose k-2}-{n-k-2\choose k-2}+1$), we conclude that the displayed holds trivially. Thus, is valid in any case.
If $\gamma({\mathcal F})>{n-3\choose k-2}$ then we repeat the proof, but using the parts $\mathcal P_a,\mathcal P_b$ that contain the $k$-element sets that intersect $[4]$ in $\{1,3\}$ and in $\{2,4\}$, respectively.
Further generalizations {#sec34}
-----------------------
In this subsection, we give several further generalizations and strengthenings. It includes weighted versions and full cross-intersecting version of our results. All of them we state without proof or with a sketch of proof since the proofs use the same ideas as the proof of Theorems \[thmfull2\] and \[thmfulleq\].
Our techniques allow us to give a weighted version of Theorems \[thmfull1\] and \[thmfull2\]. Assume that, given a lower bound on $\gamma({\mathcal F})$, we aim to maximise the expression $\Delta({\mathcal F})+c\gamma({\mathcal F})$ with some $c>1$. (In terms of cross-intersecting families, we are maximising the expression $|{{\mathcal A}}|+c|{{\mathcal B}}|$.) Then the following is true.
\[thmfullw\] Let $n>2k\ge 6$. Consider all resistant pairs $S_l\subset [n],\ T_l\subset [2,n]$, where $l=0,\ldots,m$. Assume that $T_0<T_1<T_2<\ldots <T_m$. Put $C:= \frac{n-k-2}{k-2}>1$. Then $$\label{eqfull4} |\mathcal L(S_{l-1},k-1)|+C|\mathcal L(T_{l-1},k)|\ge|\mathcal L(S_{l},k-1)|+C|\mathcal L(T_{l},k)| \ \ \ \text{ for each }l \in [m],$$ and any cross-intersecting pair of families $\mathcal A\subset {[2,n]\choose k-1},\ \mathcal B\subset {[2,n]\choose k}$ with $|\mathcal L(T_{l-1},k)|<|{{\mathcal B}}|\le |\mathcal L(T_l,k)|$ satisfies $|{{\mathcal A}}|+C|{{\mathcal B}}|\le |\mathcal L(S_l,k-1)|+C|\mathcal L(T_l, k)|$.+0.1cm
In terms of intersecting families, if ${\mathcal F}\subset {[n]\choose k}$ is intersecting and $|\mathcal L(T_{l-1},k)|<\gamma({\mathcal F})\le |\mathcal L(T_l,k)|$, then $\Delta({\mathcal F})+C\gamma({\mathcal F})\le |\mathcal L(S_l,k-1)|+C|\mathcal L(T_l, k)|$.
The proof of this theorem follows the same steps as that of Theorem \[thmfull2\]. We sketch the proof of the cross-intersecting version of the theorem. Using Lemma \[lemdiv\], we may assume that ${{\mathcal A}}$ and ${{\mathcal B}}$ form a resistant pair (indeed, otherwise, replacing ${{\mathcal A}},\ {{\mathcal B}}$ with ${{\mathcal A}}',$ ${{\mathcal B}}'$ which satisfy $|{{\mathcal A}}'|+|{{\mathcal B}}'|\ge |{{\mathcal A}}|+|{{\mathcal B}}|$, $|{{\mathcal B}}'|>|{{\mathcal B}}|$ definitely increases the value of $|{{\mathcal A}}|+C|{{\mathcal B}}|$). Then, looking at the proof of Lemma \[lemres\], we see that in each of the cases the bipartite graph $G$ had parts of sizes ${n-i\choose z_a}$ and ${n-i\choose z_b}$, where $z_a+z_b = 2k-i$ and $z_a>z_b$. Since $i\ge 5$, we have $z_b\le k-3$. Therefore, if we put weight $w$, $w\le {n-i\choose z_a}/{n-i\choose z_b}$, on each vertex of the part $\mathcal P_b$ and weight 1 on each vertex of the part $\mathcal P_a$, we can still conclude that $\mathcal P_a$ is the independent set of the largest weight in $G$. The rest of the argument works out as before. We have $$\frac{{n-i\choose z_a}}{{n-i\choose z_b}}\ge\frac{(n-i-z_b)}{z_b+1} = \frac{(n-2k+z_a)}{z_b+1}\ge \frac{(n-2k+z_b+1)}{z_b+1}\ge \frac{(n-k-2)}{k-2}.$$ Therefore, we may use $w=\frac{n-k-2}{k-2}$ as the weight of the vertices in $\mathcal P_b$.
\[corweight\] Let $n>2k\ge 6$. For any intersecting family ${\mathcal F}\subset{[n]\choose k},$ $\gamma({\mathcal F})\le {n-4\choose k-3}$, we have $|\Delta({\mathcal F})|+\frac{n-k-2}{k-2}\gamma({\mathcal F})\le {n-1\choose k-1}$.
If, additionally, ${\mathcal F}$ is non-trivial, then $|\Delta({\mathcal F})|+\frac{n-k-2}{k-2}\gamma({\mathcal F})\le {n-1\choose k-1}-{n-k-1\choose k-1}+\frac{n-k-2}{k-2}$.
+0.2cm
Is not difficult to extend the considerations of Section \[sec3\] to the case of cross-intersecting families ${{\mathcal A}}\subset {[n]\choose a}, {{\mathcal B}}\subset {[n]\choose b}$. The wording of Theorem \[thmfull2\] would stay practically the same. One just need to adjust Definition \[defrespa\]. Note that, unlike before in this section, here we will work with cross-intersecting families on $[n]$ (and not $[2,n]$), and so $\mathcal L(S, a)$ stands for $\mathcal L([n],S,a)$.
We say that two sets $S, T\subset [n]$ , if either $S = \{b-a+2\}, T =[b-a+2]$ or the following holds. Assuming that the largest element of $T$ is $j$, we have
1. $S\cap T =\{j\}$, $S\cup T= [j]$, $|S|\le a$ and $|T|\le b$;
2. for each $i\ge b-a+2$ we have $|[i]\cap S|-a< |[i]\setminus S|-b$.
Let $m$ be the number of resistant pairs. For convenience, put $T_0 = [2,n]$ to correspond to the empty family, as well as $T_{m+1}=[b-a+1]$ and $T_{m+2}$ to be the analogues of the sets $\{2,3\}$ and $\{2,4\}$, respectively. Below we state a theorem, which is the analogue of Theorems \[thmfull2\], \[thmfulleq\], \[thmfullw\] and (the second part of) Corollary \[corhm\] in the case of general cross-intersecting families. Its proof is a straightforward generalization of the proofs of the respective theorems, and thus we omit it.
\[thmfullcri\] Let $b, a>0$, $n>a+b$. Consider all $(a,b)$-resistant pairs $S_l\subset [n],\ T_l\subset [2,n]$, where $l\in [m]$. Assume that $T_0<T_1<T_2<\ldots T_m$.
1\. Then $$\label{eqfull5} |\mathcal L(S_{l-1},a)|+|\mathcal L(T_{l-1},b)|>|\mathcal L(S_{l},a)|+|\mathcal L(T_{l},b)| \ \ \ \text{ for each }l \in [m],$$ and any cross-intersecting pair of families $\mathcal A\subset {[n]\choose a},\ \mathcal B\subset {[n]\choose b}$ with $|\mathcal L(T_{l-1},b)|<|{{\mathcal B}}|\le |\mathcal L(T_l,b)|$ satisfies $$\label{eqfull6} |{{\mathcal A}}|+|{{\mathcal B}}|\le |\mathcal L(S_l,a)|+|\mathcal L(T_l, b)|.$$ If the families $\mathcal L(|{{\mathcal A}}|,a)$, $\mathcal L(|{{\mathcal B}}|,b)$ have characteristic sets $S,T$, then we have equality in if and only if $S,T$ is a $T_l$-neutral $(a,b)$-pair, where the notion of a $T_l$-neutral $(a,b)$-pair is a straightforward generalization of that of neutral pair.
2\. The same conclusion holds with $|\mathcal B|$, $|\mathcal L(T_{l-1},b)|$ $|\mathcal L(T_l,b)|$ replaced with $C|\mathcal B|$, $C|\mathcal L(T_{l-1},b)|$ and $C|\mathcal L(T_l,b)|$, where $C$ is a constant, $C<\frac{n-b-1}{a-1}.$
3\. Denote $t:=b+1-a$. Assume that ${n-i\choose b-i}<|{{\mathcal B}}|\le {n-t\choose a-1}+{n-t-1\choose a-1}$ for integer $i\in [t+1,b]$. If $i\ge t+2$ then we have
$$|{{\mathcal A}}|+|{{\mathcal B}}|\le {n\choose a}-{n-i\choose a}+{n-i\choose b-i} -{n-b-1\choose a-1}+1.$$
If $i=t+1$ and $|{{\mathcal B}}|\notin [{n-t\choose a-1}-{n-b-1\choose a-1}+2,{n-t\choose a-1}]$, $|{{\mathcal B}}|\le {n-t\choose a-1}+{n-t-1\choose a-1}-{n-b-1\choose a-1}+1$ then we have
$$|{{\mathcal A}}|+|{{\mathcal B}}|\le {n\choose a}-{n-t\choose a}+{n-t\choose b-t} -{n-b-1\choose a-1}+1.$$
Remark that we have $|{{\mathcal A}}|+|{{\mathcal B}}|={n\choose a}-{n-t\choose a}+{n-t\choose a-1}$ for ${{\mathcal A}}, {{\mathcal B}}$ defined by characteristic sets $S_i, T_i$, where $i=m,m+1,m+2$.
This theorem generalizes and strengthens many results on cross-intersecting families, in particular, the theorem for cross-intersecting families proven in [@KZ] and the following theorem due to Frankl and Tokushige [@FT]
\[lemft\] Let $n > a+b$, $a\le b$, and suppose that families $\mathcal F\subset{[n]\choose a},\mathcal G\subset{[n]\choose b}$ are cross-intersecting. Suppose that for some real number $\alpha \ge 1$ we have ${n-\alpha \choose n-a}\le |\mathcal F|\le {n-1\choose n-a}$. Then $$\label{eqft}|\mathcal F|+|\mathcal G|\le {n\choose b}+{n-\alpha \choose n-a}-{n-\alpha \choose b}.$$
One easy corollary of and , which also appeared in [@KZ] and other places, is as follows:
\[corkz\] Let $a,b>0$, $n>a+b$. Let ${{\mathcal A}}\subset {[n]\choose a},\ {{\mathcal B}}\subset {[n]\choose b}$ be a pair of cross-intersecting families. Denote $t:=b+1-a.$ Then, if $|{{\mathcal B}}|\le {n-t\choose a-1}$, then $$\label{eqcreasy} |{{\mathcal A}}|+|{{\mathcal B}}|\le {n\choose a}.$$ Moreover, the displayed inequality is strict unless $|{{\mathcal B}}|=0$.
If ${n-j\choose b-j}\le |{{\mathcal B}}|\le {n-t\choose a-1}$ for integer $j\in [t, b]$, then $$\label{eqcreasy2} |{{\mathcal A}}|+|{{\mathcal B}}|\le {n\choose a}-{n-j\choose a}+{n-j\choose b-j}.$$ Moreover, if the left inequality on ${{\mathcal B}}$ is strict, then the inequality in the displayed formula above is also strict, unless $j=t+1$ and $|{{\mathcal B}}| = {n-t\choose a-1}$.
We note that the results in [@KZ] did not explicitly treat the equality case. However, it is clear that strictness of follows from , and the equality case in follows from Theorem \[thmfullcri\] part 3.
Beyond Hilton–Milner. Proofs of Theorem \[thmhk\] and \[thmclass2\] {#sec4}
===================================================================
As we have already mentioned in the introduction, applying Corollary \[corhm\] with $u=k$, we conclude that holds for $k\ge 4$.
Next, in terms of Theorem \[thmfull1\], we know that $\gamma_i = i$ for $i\in[k-3]$, and $\gamma_{k-2} = n-k$. Thus, for $k\ge 5$, $\gamma_2= 2$ and, using Theorem \[thmfull1\], we conclude that any intersecting family ${\mathcal F}\subset {[n]\choose k}$ with $\gamma({\mathcal F})\ge 3$ is strictly smaller than the right hand side of . Therefore, if ${\mathcal F}$ with $\gamma({\mathcal F})\ge 2$ has size equal to the right hand side of , then $\gamma({\mathcal F})=2$. But then any maximal ${\mathcal F}$ must be isomorphic to $\mathcal J_i$ for some $i\ge 2$ (since the family is uniquely determined by the size of the intersection of the two sets contributing to the diversity). Finally, as we have mentioned in the introduction, $|\mathcal J_i|<|\mathcal J_2|$ for any $i>2$.
Let us recall the definition of shifting. For a given pair of indices $1\le i<j\le n$ and a set $A \subset [n]$, define its $S_{ij}(A)$ as follows. If $i\in A$ or $j\notin A$, then $S_{ij}(A) = A$. If $j\in A, i\notin A$, then $S_{ij}(A) := (A-\{j\})\cup \{i\}$. That is, $S_{ij}(A)$ is obtained from $A$ by replacing $j$ with $i$. The $(i,j)$-shift $S_{ij}(\mathcal A)$ of a family $\mathcal A$ is as follows: $$S_{ij}(\mathcal A) := \{S_{ij}(A)\ :\ A\in \mathcal A\}\cup \{A\ :\ A,S_{ij}(A)\in \mathcal A\}.$$
Proof of Theorem \[thmclass2\] {#sec41}
------------------------------
Let us deal with the first part of the statement first. Assume that the statement does not hold and choose ${\mathcal F}$, $t$, ${\mathcal M}$, and ${\mathcal F}'$ satisfying the requirements of the theorem, so that $|{\mathcal F}|>|{\mathcal F}'|$ and ${\mathcal F}$ has the smallest possible diversity under these conditions. First, we can clearly suppose that $\gamma({\mathcal F})\le {n-4\choose k-3}$. Indeed, the largest family with diversity bigger than ${n-4\choose k-3}$ is at most as large as $\mathcal H_4$ (cf. ), and $\mathcal H_4$ contains a copy of any ${\mathcal G}$ as in the statement of the theorem. Therefore, by the choice of ${\mathcal F}$, it cannot have diversity larger than $\mathcal H_4$, otherwise we should have replaced ${\mathcal F}$ with $\mathcal H_4$.
From now on, we suppose that $\gamma({\mathcal F})\le {n-4\choose k-3}$. Suppose that $\bigcap_{M\in {\mathcal M}}M = [2,t+1]$.
For $i=2,\ldots,t+1$, we may consecutively apply all the $S_{ij}$-shifts, where $j>i$, to ${\mathcal F}$. Note that the family ${\mathcal M}$ stays intact under these shifts, and the diversity and size of ${\mathcal F}$ is not affected. Thus, we may assume that ${\mathcal F}$ is invariant under these shifts. For any $j\ge 2$ and $S\subset [2,j]$, define $${\mathcal F}(S,[j]):=\big\{F\subset [j+1,n]\ :\ F\cup S\in {\mathcal F}(\bar 1)\big\}$$ and, for any family $\mathcal R\subset {[n]\choose k}$, let $\partial\mathcal R:=\bigcup_{R\in \mathcal R} {R\choose k-1}$ denote the of $\mathcal R$. (We gave an equivalent definition in Theorem \[thmkk\].)
An important consequence of the shifts we made is that, for any $i\in [2,t+1]$ and $S\subset [2,i-1]$, we have $$\label{eqcard} \big|{\mathcal F}(S\cup \{i\},[i])\big|\ge \big|\partial {\mathcal F}(S,[i])\big|.$$ Actually, we have ${\mathcal F}(S\cup \{i\},[i])\supset \partial {\mathcal F}(S,[i])$. Indeed, for any $F'\in \partial {\mathcal F}(S,[i])$, there exists $j\in [i+1,n]$ and $F\in {\mathcal F}(S,[i])$, such that $F'\cup \{j\}=F$, and, since we the $(i,j)$–shift, we have $F'\cup \{i\}\cup S\in {\mathcal F}$ and $F'\in {\mathcal F}(S\cup \{i\},[i])$.
We have $|{\mathcal F}(\bar 1)|\le {n-4\choose k-3}$ by assumption. Let us show that, for any $i\in [2,4]$, we have $$\label{eqcard2} |{\mathcal F}([2,i-1],[i])|\le {n-5\choose k-3}.$$ (Naturally, we put $[2,1]=\emptyset$.) Assume that does not hold. Then, using the Kruskal–Katona theorem, it is not difficult to see that $|\partial {\mathcal F}([2,i-1],[i])|> {n-5\choose k-4}$,[^7] and thus $$\big|{\mathcal F}(\bar 1)\big|\ge \big|{\mathcal F}([2,i-1],[i])\big|+\big|{\mathcal F}([2,i],[i])\big|\overset{\eqref{eqcard}}{\ge} \big|{\mathcal F}([2,i-1],[i])\big|+\big|\partial {\mathcal F}([2,i-1],[i])\big|>{n-4\choose k-3},$$ a contradiction. For each $i\in [2,4]$, consider the following bipartite graph $G_i$. The parts of $G_i$ are $$\begin{aligned}
\mathcal P_a(i) :=\ &\Big\{P\ :\ i\in P\in {[2,n]\choose k-1}\Big\},\\
\mathcal P_b(i) :=\ &\Big\{P \ :\ i\notin P \in {[2,n]\choose k}\Big\},\end{aligned}$$ and edges connect disjoint sets. We identify $\mathcal P_a(i)$ with ${X\choose k-2}$ and $\mathcal P_b(i)$ with ${X\choose k}$, where $X := [2,n]\setminus \{i\}$, $|X|=n-2>k+k-2$.
Due to , we have $|\mathcal P_b(i)\cap {\mathcal F}(\bar 1)|\le {n-5\choose k-3}={|X|-3\choose (k-2)-1}$ for $i=2$. Thus, we can apply to $$\begin{aligned}
{{\mathcal A}}:={\mathcal F}(1)\cap \mathcal P_a(i) \ \ \ \ \text{and}\ \ \ \
{{\mathcal B}}:={\mathcal F}(\bar 1)\cap \mathcal P_b(i),\end{aligned}$$ where $i=2$, and conclude that $|{{\mathcal A}}|+|{{\mathcal B}}|\le {|X|\choose k-2}$ (and the inequality is strict unless ${{\mathcal B}}=\emptyset$). Therefore, removing ${\mathcal F}(\bar 1)\cap \mathcal P_b(2)$ from ${\mathcal F}(\bar 1)$ and adding sets from $\mathcal P_a(2)$ to ${\mathcal F}(1)$, we get a pair of families with larger sum of cardinalities. Moreover, the new pair is cross-intersecting: all sets in ${\mathcal F}(\bar 1)\setminus \mathcal P_b(2)$, as well as the sets from $\mathcal P_a(2)$, contain $2$. Thus, by our choice of ${\mathcal F}$ we may assume that ${{\mathcal B}}=\emptyset$ and so all sets in ${\mathcal F}(\bar 1)$ must contain $2$. Repeat this argument for $i=3$ and $i=4$. Concluding, we may assume that all sets in ${\mathcal F}(\bar 1)$ contain $[2,4]$.
Next, for each $i\ge 5$, consider the following slightly different bipartite graph $G_i'$. The parts of $G_i$ are $$\begin{aligned}
\mathcal P_a'(i):=\ &\Big\{P\ :\ P\in {[2,n]\choose k-1},\ P\cap [2,i]=\{i\}\Big\},\\
\mathcal P_b'(i):=\ &\Big\{P\ :\ P\in {[2,n]\choose k},\ P\cap [2,i]=[2,i-1]\Big\},\end{aligned}$$ and edges connect disjoint sets. We identify $\mathcal P_a'(i)$ with ${[i+1,n]\choose k-2}$ and $\mathcal P_b'(i)$ with ${[i+1,n]\choose k-i+1}$.
We have $|\mathcal P_b'(i)\cap {\mathcal F}(\bar 1)|\le {n-i\choose k-i+1}\le {n-5\choose k-3}$. Thus, for each $i=5,\ldots, t+1$ we can apply to $$\begin{aligned}
{{\mathcal A}}:={\mathcal F}(1)\cap \mathcal P_a'(i) \ \ \ \ \text{and}\ \ \ \ \
{{\mathcal B}}:={\mathcal F}(\bar 1)\cap \mathcal P_b'(i),\end{aligned}$$ and conclude that $|{{\mathcal A}}|+|{{\mathcal B}}|\le {n-i\choose k-2}$ (and the inequality is strict unless ${{\mathcal B}}=\emptyset$). Arguing as before, we may assume that all sets in ${\mathcal F}(\bar 1)$ contain $[2,t+1]$. +0.1cm
Put ${\mathcal M}= \{M_1,\ldots, M_z\}$. Since ${\mathcal M}$ is minimal, for each $M_l\in {\mathcal M}$, $l\in [z]$, there is $$\label{eqil} i_l\in \Big(\bigcap_{M\in {\mathcal M}\setminus \{M_l\}}M\Big)\setminus \Big(\bigcap_{M\in {\mathcal M}}M\Big).$$ We assume that $i_l = t+l+1$, $l\in[z]$. In particular, $\{i_1,\ldots, i_z\} = [t+2,t+z+1]$.
Next, for each $i=t+2,\ldots, t+z+1$ consider the bipartite graph $G'_i$, defined above. We can apply with $k-2,k-i+1, k-i+1, 4-i$ playing roles of $a,b,j,t$, respectively. Indeed, we know that $|{{\mathcal B}}|\ge {n-i\choose (k-i+1)-(k-i+1)}=1$ since $M_{i-t-1}\in \mathcal P_b'(i)$ (note that $M_l\notin \mathcal P_b'(i)$ for $l\ne i-t-1$, since all of them contain $i_l$ due to the definition of $i_l$). Therefore, $|\mathcal A|+|\mathcal B|\le {n-i\choose k-2}-{n-k-i\choose k-2}+1$. For $k\ge 5$, the inequality is strict unless ${{\mathcal B}}=\{M_{i-t-1}\}$. For $k=4$, we apply it for $i=4$, thus $n-5,2,1,1,0$ play the roles of $n,a,b,j,t$, respectively, and so the inequality is strict unless $|{{\mathcal B}}| = {n-5\choose 1}$, i.e., ${{\mathcal B}}$ contains all possible $1$-element sets. Thus, we may replace ${{\mathcal A}},\ {{\mathcal B}}$ with ${{\mathcal B}}':=\{M_{i-t-1}\}$ and ${{\mathcal A}}':=\{A\in {[i+1,n]\choose k-2}\ :\ A\cap M_{i-t-1}\ne \emptyset\}$. The resulting family is cross-intersecting. Thus, by our choice of ${\mathcal F}$, all sets in ${\mathcal F}(\bar 1)\setminus {\mathcal M}$ must contain $i$. Iterating this procedure for $i=t+2,t+3,\ldots, t+z+1$, we may assume that any set in ${\mathcal F}(\bar 1)\setminus {\mathcal M}$ must contain the set $[2, t+z+1]$. Put $t':=t+z+1$. If ${\mathcal F}(\bar 1)\setminus {\mathcal M}$ is non-empty then, for each $l \in [z]$, the set $M_l\setminus [2,t']$ must be non-empty: otherwise, $t'-1>k$. If ${\mathcal F}(\bar 1)\setminus {\mathcal M}$ is empty then ${\mathcal F}(\bar 1)={\mathcal M}$, which contradicts $|{\mathcal F}|>|{\mathcal F}'|$. Note that $t'-1\ge z+3$.
For each $l\in [z]$, select one element $i_l\in M_l\cap [t'+1,n]$. Note that $i_l$ may coincide. Put $I:=\{i_l\ :\ l\in [z]\}$. Consider the bipartite graph $G(t',I)$ with parts $$\begin{aligned}
\mathcal P_a(t',I):=\ &\Big\{P \ :\ P\in {[2,n]\choose k-1},\ I\subset P,\ [2,t']\cap P=\emptyset\Big\},\\
\mathcal P_b(t',I):=\ &\Big\{P\ :\ P\in {[2,n]\choose k},\ [2,t']\subset P,\ I\cap P=\emptyset\Big\},\end{aligned}$$ and edges connecting disjoint sets. We identify $\mathcal P_a(t',I)$ with ${Y\choose k-z'}$, $z'\le z+1$, and $\mathcal P_b(t',I)$ with ${Y\choose k-z''}$, $z''=t'-1\ge z+3$, where $Y = [t'+1,n]\setminus I$, $|Y|=n-z''-z'$. In particular, $|Y|>k-z'+k-z''$. By the choice of $I$, we have $|{\mathcal M}\cap \mathcal P_b(t',I)| =\emptyset$. Denote $$\begin{aligned}
{{\mathcal A}}:={\mathcal F}(1)\cap \mathcal P_a(t',I) \ \ \ \ \text{and}\ \ \ \ \
{{\mathcal B}}:={\mathcal F}(\bar 1)\cap \mathcal P_b(t',I).\end{aligned}$$ We have $k-z'>k-z''$, and, therefore, we may apply with $a:=k-z',\ b:=k-z'',\ j:=k-z''$ (the upper bound on $|{{\mathcal B}}|$ becomes trivial in that case) and conclude that $|{{\mathcal A}}|+|{{\mathcal B}}|\le {|Y|\choose k-z'}$ (and the inequality is strict unless ${{\mathcal B}}=\emptyset$). As before, replacing ${{\mathcal A}}$ with $\mathcal P_a(I)$ and ${{\mathcal B}}$ with $\emptyset$ does not decrease the sum of sizes of the families and preserves the cross-intersecting property. Thus, by the choice of ${\mathcal F}$, we must have ${{\mathcal B}}=\emptyset$.
Repeating this for all possible choices of $I$, we arrive at the situation when any set from ${\mathcal F}(\bar 1)\setminus{\mathcal M}$ must intersect [*any*]{} such set $I$. Clearly, this is only possible for a set $F$ if $F\supset M_l\cap[t'+1,n]$. But this implies that $|F|>|M_l|$, which is impossible. Thus ${\mathcal F}(\bar 1) = {\mathcal M}$, which contradicts $|{\mathcal F}|>|{\mathcal F}'|$, so the proof of is complete.
Finally, uniqueness follows from the fact that the inequalities , are strict unless the family ${{\mathcal B}}$ has sizes $0$ and $1$, respectively, or $k=4$ (we mentioned it at every application). Therefore, if ${\mathcal F}(\bar 1)\ne {\mathcal M}$, then at some point we would have had a strict inequality in the application of , .\
Let us now prove the moreover part of the statement of Theorem \[thmclass2\]. First, if there is no family $\mathcal M\subset {\mathcal F}(\bar 1)$, minimal w.r.t. common intersection and such that $|\bigcap_{M\in{\mathcal M}}M|=t$, then we apply $(i,j)$-shifts to ${\mathcal F}$ for $2\le i<j\le n$ until such family appears. Since common intersection of any subfamily of ${\mathcal F}(\bar 1)$ may change by at most one after any shift, either we obtain the desired ${\mathcal M}$ or we arrive at a shifted family ${\mathcal F}(\bar 1)$ without such ${\mathcal M}$. But the latter is impossible. Indeed, for a shifted intersecting family ${\mathcal F}(\bar 1)$, we have $|\bigcap_{F\in {\mathcal F}(\bar 1)}F|=[2,j]$ for some $j\ge 2$, and in that case $[2,j]\cup [j+2,k+2]\in {\mathcal F}(\bar 1)$. But then the set $F_{j'}:=[2+j']\cup [j'+2,k+2]\in{\mathcal F}(\bar 1)$, where $j\le j'\le k+1$. It is clear that the sets $F_j,\ldots, F_{k-t+j+1}$ form a subfamily that has common intersection of size $t$.
Fix $\mathcal M\subset {\mathcal F}(\bar 1)$, minimal w.r.t. common intersection, such that $|\bigcap_{M\in{\mathcal M}}M|=t$. Applying the first part of Theorem \[thmfull2\], we may assume that ${\mathcal F}(\bar 1)={\mathcal M}$. Clearly, the number of sets in ${\mathcal F}(1)$ passing through $\bigcap_{M\in{\mathcal M}}M$ is always the same, independently of the form of ${\mathcal M}$. Thus, we need to analyze the sum of sizes of the family ${\mathcal F}':=\big\{F\in {\mathcal F}(1)\ :\ F\cap \bigcap_{M\in{\mathcal M}}M=\emptyset\big\}$ and ${\mathcal M}':= \big\{M\setminus \bigcap_{M\in{\mathcal M}}M\ :\ M\in {\mathcal M}\big\}.$ Note that ${\mathcal M}'$ and ${\mathcal F}'$ are cross-intersecting, moreover, $\tau({\mathcal M}')=2$ and ${\mathcal M}'$ is minimal w.r.t. this property.
The following lemma concludes the proof of the theorem. Let us first give some definitions. Given integers $m>2s$, let us denote by $\mathcal T_2'(s):=\{[s], [s+1,2s]\}$. Let ${\mathcal F}_2'(s)\subset {[m]\choose k-1}$ stand for the largest family, cross-intersecting with $\mathcal T_2'(s)$. Let ${\mathcal F}_2(s)\subset {[m]\choose k-1}$ stand for the largest family, cross-intersecting with $\mathcal T_2(s)$ (cf. ).
\[lemmin\] Let $k\ge s$ and $m\ge k+s$ be integers, $k\ge 4$. Given a family $\mathcal H\subset {[m]\choose s}$ with $\tau(\mathcal H)=2$ and minimal w.r.t. this property, consider the maximal family ${\mathcal F}\subset {[m]\choose k-1}$ that is cross-intersecting with $\mathcal H$. Then the unique maximum of $|{\mathcal F}|+|\mathcal H|$ is attained when $\mathcal H$ is isomorphic to $\mathcal T_2'(s)$ (and ${\mathcal F}$ is thus isomorphic to ${\mathcal F}_2'(s)$).
If we additionally require that $\mathcal H$ is intersecting[^8] then the maximum of $|{\mathcal F}|+|\mathcal H|$ is attained for $\mathcal H$ and ${\mathcal F}$ isomorphic to $\mathcal T_2(s)$ and ${\mathcal F}_2(s)$. The maximal configuration is unique if $s\ge k$.
We can apply the first part of Lemma \[lemmin\] in our situation with $[2,n]\setminus \bigcap_{M\in{\mathcal M}}M$ playing the role of $[m]$ and $k-t$ playing the role of $s$. Note that $|[2,n]\setminus \bigcap_{M\in{\mathcal M}}M| = n-t-1\ge 2k-t$, and so the condition on $m$ from the lemma is satisfied. This proves , moreover, we get that for $k\ge 5$ the inequality was strict unless $|{\mathcal F}(\bar 1)|=2$ in the first place. But then ${\mathcal F}$ is isomorphic to a subfamily of $\mathcal J_{i}$ for $i\ge i'$, and the equality is possible only if ${\mathcal F}$ is isomorphic to $\mathcal J_i$. Thus, to conclude the proof of Theorem \[thmfull2\], we only have to prove the lemma.
We note that the second, more complicated, part of Lemma‘\[lemmin\] is not needed for this application, however, it will be crucial for the completion of the proof of Theorem \[thmtau3\].
Let us first express $|{\mathcal F}'_2(s)|$. It is not difficult to see that
$$\begin{aligned}
\notag |{\mathcal F}'_2(s)|\ =\ & {m-1\choose k-2}-{m-s-1\choose k-2} + \\
\notag & {m-2\choose k-2}-{m-s-2\choose k-2}+\\
\notag &\cdots \\
\label{eqfs} &{m-s\choose k-2}-{m-2s\choose k-2}.\end{aligned}$$
Indeed, in the first line we count the sets containing $1$ that intersect $[s+1,2s]$, in the second line we count the sets not containing $1$, containing $2$ and intersecting $[s+1,2s]$ etc.
We can actually bound the size of ${\mathcal F}$ for any $\mathcal H$ in a similar way. Suppose that $z:=|\mathcal H|$ and $\mathcal H = \{H_1,\ldots, H_z\}$. Since $\mathcal H$ is minimal, for each $l\in[z]$ there exists an element $i_l$ such that $i_l\notin H_l$ and $i_l\in \bigcap_{j\in[z]\setminus \{l\}} H_l$. (All $i_l$ are of course different, cf. also .) Applying Bollobas’ set-pairs inequality [@Bol] to $\mathcal H$ and $\{i_{l}\ :\ l\in[z]\}$, we get that $|\mathcal H|\le {s+1\choose s}=s+1$.
For each $l=2,\ldots, z$, we count the sets $F\in{\mathcal F}$ such that $F\cap \{i_2,\ldots, i_l\}=\{i_l\}$. Such sets must additionally intersect $H_{l}\setminus \{i_2,\ldots, i_{l-1}\}$. Note that $H_1\supset \{i_2,\ldots, i_z\}$. Assuming that $H_1\setminus \{i_2,\ldots,i_z\} = \{j_1,\ldots, j_{s-z}\}$, for each $l=1,\ldots, s-z$ we further count the sets $F\in {\mathcal F}$ such that $F\cap \{i_2,\ldots, i_z, j_1,\ldots, j_l\}=\{j_l\}$. Such sets must additionally intersect $H_{i}\setminus \{i_2,\ldots, i_{z}\}$ for some $i\in [2,z]$.[^9] Since $F\cap H_1\ne \emptyset$ for any $F\in {\mathcal F}$ and given that the classes for different $l$ are disjoint, we clearly counted each set from ${\mathcal F}$ exactly once. (However, we may also count some sets that are not in ${\mathcal F}$.) Doing this count, we get the following bound on ${\mathcal F}$.
$$\begin{aligned}
\notag |{\mathcal F}|\ \le \ & {m-1\choose k-2}-{m-s-1\choose k-2} + \\
\notag & {m-2\choose k-2}-{m-s-1\choose k-2}+\\
\notag &\cdots\\
\notag & {m-z+1\choose k-2}-{m-s-1\choose k-2}+\\
\notag & {m-z\choose k-2}-{m-s-2\choose k-2}+\\
\notag &\cdots \\
\label{eqfz} &{m-s\choose k-2}-{m-2s-2+z\choose k-2}\ =:\ f(z).\end{aligned}$$
Remark that coincides with when substituting $z=2$. We have $f(z-1)-f(z)\ge {m-s-1\choose k-2}-{m-s-2\choose k-2}={m-s-2\choose k-3}> 1$ (here we use that $m\ge s+k$ and $k\ge 4$). Therefore, for any $z\ge z'$, $$\label{eqz'}|\mathcal H|+|{\mathcal F}|\le f(z')+z',$$ and the inequality is strict unless $z=|\mathcal H|=z'$.
At the same time, we have $|{\mathcal F}'_2(s)|+|\mathcal T'_2(s)|=f(2)+2$ and $|{\mathcal F}_2(s)|+|\mathcal T_2(s)|=f(3)+3$! (The former we have seen above, and the latter is easy to verify by doing exactly the same count.) Since, up to isomorphism, there is only one family $\mathcal H\subset {[m]\choose s}$ of size $2$ with $\tau(\mathcal H)=2$, we immediately conclude that the first part of the statement holds. To deduce the second part, we only need to show that, among all possible choices of $\mathcal H$ of size $3$, the only one (up to isomorphism) that attains equality in is $\mathcal H = \mathcal T_2(s)$.
Recall that, for uniqueness in the second part of the lemma, we have additional condition $s\ge k$. If there are two sets $H',H''\in \mathcal H$ such that $|H'\cap H''|=s-1$, then $\mathcal H$ is isomorphic to $\mathcal T_2(s)$. Therefore, in what follows we assume that $|H'\cap H''|\le s-2$ for any $H',H''\in\mathcal H$.
Let us deal with the case when $H_{l}\cap H_{l'} = i_{l''}$ for any $\{l,l',l''\}=[3]$. Note that this implies that $$\label{eqbign}
m\ge 3s-3.$$ Since $s\ge k \ge 4$, there are elements $j_l\in H_l\setminus (H_{l'}\cup H_{l''})$, $\{l,l',l''\}=[3]$. Perform the $(j_1,j_2)$-shift on ${\mathcal F}\cup \mathcal H$ and denote ${\mathcal F}':=S_{j_1j_2}({\mathcal F})$. Clearly, the sizes of the families stay the same and the resulting families are cross-intersecting. The family, $S_{j_1j_2}(\mathcal H)$ has covering number $2$. Moreover, we can add new sets to ${\mathcal F}'$ without violating the cross-intersecting property. Indeed, consider the families $$\begin{aligned}
\mathcal A\ :=&\ \Big\{F\in {[m]\choose k-1}\ :\ j_3\in F,\ F\cap (H_1\cup H_2)=\{j_2\}\Big\},\\
\mathcal A'\ :=&\ \Big\{F\in {[m]\choose k-1}\ :\ j_3\in F,\ F\cap (H_1\cup H_2)=\{j_1\}\Big\}.\end{aligned}$$ It is easy to see that actually ${{\mathcal A}}' = S_{j_1j_2}({{\mathcal A}})$. Moreover, ${{\mathcal A}}\cap {\mathcal F}= \emptyset$ since sets from ${{\mathcal A}}$ do not intersect $H_1$ and ${{\mathcal A}}'\cap {\mathcal F}' = \emptyset$ since ${{\mathcal A}}'\cap {\mathcal F}=\emptyset$ (again, since sets from ${{\mathcal A}}'$ do not intersect $H_2$) and ${{\mathcal A}}\cap {\mathcal F}= \emptyset$. At the same time, ${{\mathcal A}}'$ may be included into ${\mathcal F}'$ since sets from ${{\mathcal A}}'$ intersect all sets in $S_{j_1j_2}(\mathcal H)$. Finally, ${{\mathcal A}}= {[m]\setminus X\choose k-3}$, where $|X| = 2s$, and thus $|{{\mathcal A}}|\ge {k-3\choose k-3}$ due to $m\ge 3s-3\ge k+2s-3$, so ${{\mathcal A}}$ is non-empty, and thus $\mathcal H$ was not optimal.
Finally, we may assume that $|H_1\cap H_2|\in [2,s-2]$. Then we do the a similar count as for . The first two steps (with $i_2,i_3$) are the same. The part with $j_i$ is, however, slightly modified. Take $j'\in (H_1\cap H_2)\setminus \{i_3\}$ and $j'' \in H_1\setminus (H_2\cup \{i_2\}$. Such choices are possible due to $|H_1\cap H_2|\in [2,s-2]$. Count the sets $F\in {\mathcal F}$ such that $F\cap \{i_2,i_3,j'\}=j'$. They must intersect $H_3\setminus \{i_2\}$. Next, crucially, count the sets in $F\in {\mathcal F}$ such that $F\cap \{i_2,i_3,j',j''\}=j''$. They must intersect $H_2\setminus \{i_2,j'\}$ (note the size of this set is $s-2$ instead of $s-1$). The remaining count is the same: let $\{j_1,\ldots, j_{s-4}\}:=H_1\setminus \{i_2,i_3,j',j''\}$ and, for each $l\in[s-4]$, count the sets $F\in {\mathcal F}$ such that $F\cap \{i_2,i_3,j',j'',j_1,\ldots, j_{l}\} =j_l$. They must additionally intersect either $H_2\setminus \{i_2,j'\}$, or $H_3\setminus \{i_3\}$. Thus, we obtain the following bound.
$$\begin{aligned}
\notag |{\mathcal F}|\ \le \ & {m-1\choose k-2}-{m-s-1\choose k-2} + \\
\notag & {m-2\choose k-2}-{m-s-1\choose k-2}+\\
\notag & {m-3\choose k-2}-{m-s-2\choose k-2}+\\
\notag & {m-4\choose k-2}-{m-s-\textbf{2}\choose k-2}+\\
\notag & {m-5\choose k-2}-{m-s-4\choose k-2}+\\
\notag &\cdots\\
\label{eqfz} &{m-s\choose k-2}-{m-2s+1\choose k-2}\ =:\ f'(3).\end{aligned}$$
We have $f(3)-f'(3) = {m-s-2\choose k-2}-{m-s-3\choose k-2} = {m-s-3\choose k-3}\ge 1$ due to $m\ge s+k$, and thus $|{\mathcal F}|\le f'(3)<f(3) = |{\mathcal F}_2(s)|$. Thus, in the assumption $s\ge k$ and if $\mathcal H$, $|\mathcal H|\ge 3$, is not isomorphic to $\mathcal T_2(s)$, we have strict inequality in for $z'=3$. The lemma is proven.
Families with fixed covering number and maximum degree proportion {#sec5}
=================================================================
One of our main tools for this section is the following structural result due to Dinur and Friedgut [@DF]. We say that a family $\mathcal J\subset 2^{[n]}$ is a [*$j$-junta*]{}, if there exists a subset $J\subset [n]$ of size $j$ (the [*center*]{} of the junta), such that the membership of a set in ${\mathcal F}$ is determined only by its intersection with $J$, that is, for some family $\mathcal J^*\subset 2^{J}$ (the [*defining family*]{}) we have ${\mathcal F}=\{F\ :\ F\cap J\in \mathcal J^*\}$.
\[thmdf\] For any integer $r\ge 2$, there exist functions $j(r), c(r)$, such that for any integers $1 < j(r) < k < n/2$, if ${\mathcal F}\subset {[n]\choose k}$ is an intersecting family with $|{\mathcal F}|\ge c(r){n-r\choose k-r}$, then there exists an intersecting $j$-junta $\mathcal J$ with $j\le j(r)$ and $$\label{eqDF} |{\mathcal F}\setminus\mathcal J|\le c(r){n-r\choose k-r}.$$
Proof of Theorem \[thmtau3\] {#sec51}
----------------------------
Recall the expression of the size of $\mathcal C_3(n,k)$, obtained in the proof of Lemma \[lemmin\] (cf. ):
$$\begin{aligned}
\label{sizec}
\notag |\mathcal C_3(n,k)|\ =\ 3\ +\ &{n-2\choose k-2}-{n-k-2\choose k-2} + \\
\notag & {n-3\choose k-2}-{n-k-2\choose k-2}+\\
\notag & {n-4\choose k-2}-{n-k-3\choose k-2}+\\
\notag &\cdots \\
&{n-k-1\choose k-2}-{n-2k\choose k-2}.\end{aligned}$$
Let $C$ be a sufficiently large constant, which value shall be clear later.[^10] The case of $k\le C$ follows from the original result of Frankl. In what follows, we assume that $k\ge C$. Take any family ${\mathcal F}\subset {[n]\choose k}$ with $\tau({\mathcal F})=3$. Theorem \[thmdf\] implies that there exists a set $J$, $|J|\le c(5)$, and an intersecting family $\mathcal J^*\subset 2^{[J]}$, such that $|{\mathcal F}\setminus \mathcal J|\le j(5){n-5\choose k-5}\le_C {n-5\choose k-4}$, where $\mathcal J:=\{F\in {[n]\choose k}\ :\ F\cap J\in \mathcal J^*\}$.
The first step of the proof is to show that $\mathcal J$ is, in fact, a dictatorship: $\mathcal J^*$ consists of one singleton. Indeed, if $\mathcal J^*$ contains a singleton, then we may as well assume that $\mathcal J^*$ consists of that only singleton. Otherwise, any set in $\mathcal J\cap {\mathcal F}$ must intersect $J$ in at least $2$ elements. Moreover, for any two elements we choose, there is a set in $F\in{\mathcal F}$ that does not contain any of those two elements, so any set in $\mathcal J\cap {\mathcal F}$ must intersect $F$. This gives the bound $$\label{eqsizehun}|\mathcal J\cap {\mathcal F}|\le {|J|\choose 2}\Big({n-2\choose k-2}-{n-k-2\choose k-2}\Big),$$ and, given that $${n-|J|^2\choose k-2}-{n-|J|^2-k+1\choose k-2}>_C 0.99\Big({n-2\choose k-2}-{n-k-2\choose k-2}\Big),$$ we compare with and get that $|\mathcal J\cap {\mathcal F}|<\frac 12 |\mathcal C_3(n,k)|$ (remark that we also used that $k>C$ and thus that the number of lines in is bigger than $|J|^2$). Since $|{\mathcal F}\setminus \mathcal J|\le {n-5\choose k-4}<\frac 12 |\mathcal C_3(n,k)|$, we get that $|{\mathcal F}|<|\mathcal C_3(n,k)|$.
From now on, we may assume that $\mathcal J$ is a dictatorship, say, $J=\{1\}$ and $\mathcal J=\{\{1\}\}$.[^11] The next part of the proof will make use of the bipartite switching trick. We shall transform our family ${\mathcal F}$ into another family (denoted by ${\mathcal F}''$), which will satisfy $\tau({\mathcal F}'')=3$, $|{\mathcal F}''|\ge |{\mathcal F}|$ (with strict equality in case ${\mathcal F}'$ is not isomorphic to ${\mathcal F}$) and, moreover, ${\mathcal F}''(\bar 1)$ will have covering number $2$ and a much cleaner structure.
To that end, take any $\mathcal M=\{M_1,\ldots, M_z\}\subset {\mathcal F}'(\bar 1)$ such that $\tau (\mathcal M)=2$ and ${\mathcal M}$ is minimal w.r.t. this property. Remark that $z\ge 3$ due to the fact that $\tau({\mathcal M})=2$ and ${\mathcal M}$ is intersecting. The next part of the proof borrows notations and ideas of Theorem \[thmfull2\]. For each $l\in[z]$, we can find $i_l$ as in . W.l.o.g., assume that $\{i_1,\ldots, i_z\}=[2,z+1]$. Since $|{\mathcal F}(\bar 1)|\le {n-5\choose k-4}$, we may apply the same exchange argument via $G'_i$ as in Theorem \[thmfull2\] and get a family ${\mathcal F}'$, such that ${\mathcal F}'(\bar 1) = {\mathcal M}\cup \mathcal U$, where $\mathcal U$ contains only sets that contain $[2,z+1]$, and ${\mathcal F}(1)$ consists of all sets intersecting all sets in ${\mathcal F}'(\bar 1)$. We repeat the same exchange with any element contained in all but at one set from $\mathcal M$. At the end, we may assume that each $U\in\mathcal U$ contains $[2,t']$ for $t'\ge z+1$. As in the proof of Theorem \[thmfull2\], we get $|{\mathcal F}'|>|{\mathcal F}|$ unless ${\mathcal F}'$ is isomorphic to ${\mathcal F}$ (recall that $k\ge C$ by assumption).
If ${\mathcal M}$ is isomorphic to $\mathcal T_2(k)$ (cf. ), then the number of the elements contained in exactly two sets (all but one sets) is $k+1$, and thus we may immediately conclude that $\mathcal U=\emptyset$: no $k$-set can contain a subset of size $k+1$. Otherwise, ${\mathcal M}$ is not isomorphic to $\mathcal T_2(k)$.[^12]
Let us show that we may assume that $\mathcal U$ is empty. As we have said, it is clear if $t'\ge k+2$. Otherwise, consider the family ${\mathcal M}':=\{M\setminus [2,t']\ :\ M\in{\mathcal M}\}$ and note that sets in ${\mathcal M}'$ have size at least $1$. If there is no element $i'\in [t'+1,n]$ that is contained in at least $2$ sets of ${\mathcal M}'$, then take two elements $i\in M'$ and $j\in M''$, where $t'+1\le i<j\le n$ and $M',M''$ are distinct sets in $\mathcal M'$, and perform the $(i,j)$-shift on ${\mathcal F}'$. It is easy to see that $\tau(S_{ij}(\mathcal M))=2$, moreover, is $S_{ij}({\mathcal F})$ is intersecting. Thus, we may replace ${\mathcal F}'$ with $S_{ij}({\mathcal F}')$ and ${\mathcal M}$ with $S_{ij}({\mathcal M})$.
Next, we can assume that there is an element in $i'\in [t'+1,n]$ that is contained in at least $2$ sets of ${\mathcal M}'$. Now we may reuse the argument of Theorem \[thmclass2\] again. Consider the last exchange graph from the proof of Theorem \[thmclass2\], i.e., $G(t',I)$, where $I$ is formed as follows. We include $i'$ in $I$, as well as one element from each of the sets from ${\mathcal M}'$ that do not contain $i'$. We have $|I|\le z-1$. (Note that, as before, some elements we chose may coincide, making the last inequality strict.) In terms of the proof of Theorem \[thmclass2\], we have $z'=|I|+1\le z\le t'-1=z''$. Moreover, by the choice of $I$, we have $|{\mathcal M}\cap \mathcal P_b(t',I)| =\emptyset$. Therefore, we can perform the exchange operations as before for all possible choices of $I$, concluding that either all sets in $\mathcal U$ must contain $i'$, or they must contain some fixed set $M'\in \mathcal M'$. The latter is, however, impossible, since it would again imply that a $k$-element set from $\mathcal U$ contains a $(\ge k+1)$-element set $M'\cup [2,t']$.
Therefore, we may assume that all sets in $\mathcal U$ contain $i'$, and thus all contain $[2,t']\cup\{i'\}$. Finally, we may perform exactly the same exchange operations as in the proof of Theorem \[thmfull2\], but with $[2,t']$ replaced by $[2,t']\cup \{i'\}$ in the definition $\mathcal P_a(t', I)$ and $\mathcal P_b(t',I)$ (denoted by $\mathcal P_a'(t', I)$ and $\mathcal P_b'(t',I)$, respectively). The reason it will work now is the extra fixed element in $\mathcal P_b'(t',I)$, which makes the number of fixed elements in $\mathcal P_b'(t',I)$ at least as big as in $\mathcal P_a'(t',I)$.
After the switches, we obtain a family ${\mathcal F}''$, which is intersecting, satisfies $|{\mathcal F}''|\ge |{\mathcal F}|$ and ${\mathcal F}''(\bar 1) = \mathcal M$. Moreover, it is easy to see that $\tau ({\mathcal F}'')=3$.
Finally, we need to show that, among all [*minimal*]{} families, the choice of $\mathcal T_2(k)$ is the unique optimal. But this is a direct application of the second part of Lemma \[lemmin\] with $s=k$ and $[2,n]$ playing the role of $[m]$. Note that $n>2k$, and $m\ge 2k$. The proof of Theorem \[thmtau3\] is complete.
Proof of Theorem \[thmbounddeg\] {#sec52}
--------------------------------
Let $C$ be a sufficiently large constant, which value shall be clear later.[^13] The case of $k\le C$ follows from the original result of Frankl. In what follows, we assume that $k\ge C$. First of all, we remark that $\mathcal D_{3/7}$ satisfies the condition of the theorem, since $\Delta(\mathcal D_{3/7}) = \frac 37+O(\frac kn)<_C \frac 37+\epsilon.$
Take any family ${\mathcal F}\subset {[n]\choose k}$ satisfying the requirements. Theorem \[thmdf\] implies that there exists a set $J$, $|J|\le c(5)$, and an intersecting family $\mathcal J^*\subset 2^{[J]}$, such that $|{\mathcal F}\setminus \mathcal J|\le j(5){n-5\choose k-5}\le_C \epsilon{n-5\choose k-4}$, where $\mathcal J:=\{F\in {[n]\choose k}\ :\ F\cap J\in \mathcal J^*\}$. We have $$\label{eqbigsets}\big|{\mathcal F}\setminus \{F\ :\ |F\cap J|\le 3\}\big|\le 2^{c(5)}{n-4\choose k-4}\le_C \frac{\epsilon}2{n-3\choose k-3}.$$
If $\mathcal J^*$ contains a set $R$ of size at most $2$, then, since $\mathcal J^*$ is intersecting, one of the elements of $R$, say $i$, satisfies $d_i({\mathcal F}\cap \mathcal J)\ge \frac 12 |{\mathcal F}\cap \mathcal J|$, and thus $d_i({\mathcal F})\ge \frac 12\big(|{\mathcal F}|-|{\mathcal F}\setminus \mathcal J|\big)\ge \frac 12\big( |{\mathcal F}|-{n-5\choose k-4}\big)$. Given that $|{\mathcal F}|\ge |\mathcal D_{3/7}|>{n-3\choose k-3}$, we get $d_i({\mathcal F})\ge \frac 12|{\mathcal F}|\big(1-\frac kn\big)$, which is bigger than $c|{\mathcal F}|$ for $n>k/\epsilon$, a contradiction. Thus, $\mathcal J^*$ contains only sets of size at least $3$. Let us put $\mathcal J_3^*:=\mathcal J^*\cap {[n]\choose 3}$ and $\mathcal J_3:=\big\{F\in {[n]\choose k}\ :\ F\cap J\in \mathcal J_3^*\big\}$.
Recall that a $\tau^*(\mathcal G)$ of a family ${\mathcal G}\subset 2^{[n]}$ is the minimum value of $\sum_{i=1}^n w_i$, where $0\le w_i\le 1$ are real numbers, such that $\sum_{i\in G}w_i\ge 1$ for each $G\in {\mathcal G}$. Note that if we allow $w_i$ to take integer values only, then we get back to the definition of covering number. Clearly, $\tau({\mathcal G})\ge \tau^*({\mathcal G})$.
Arguing as above, we have $\tau^*(\mathcal J^*_3)>2$. Indeed, it is easy to see that, for any family ${\mathcal G}$, $\Delta({\mathcal G})\ge |{\mathcal G}|/\tau^*({\mathcal G})$. Therefore, if $\tau(\mathcal J^*_3)\le 2$ then, of course, $\tau^*(\mathcal J_3\cap {\mathcal F})\le 2$ and so at least one element $i\in J$ satisfies $d_i({\mathcal F}\cap\mathcal J_3)\ge \frac 12\big|{\mathcal F}\cap \mathcal J_3\big|$. Since the vast majority of sets from ${\mathcal F}$ are actually from ${\mathcal F}\cap \mathcal J_3$ (due to ), we again get $d_i({\mathcal F})>c|{\mathcal F}|$, a contradiction.
Füredi [@Fur] showed that, for an intersecting family ${\mathcal G}\subset {[n]\choose p}$, we have $\tau^*({\mathcal G})\le p-1+1/p$, and, moreover, if ${\mathcal G}$ is not a projective plane of order $p-1$, then $\tau^*({\mathcal G})\le p-1$. Thus, since $\tau^*(\mathcal J^*_3)>2$, it must be isomorphic to the Fano plane $\mathcal P$. In what follows, we assume that $\mathcal J^*_3 = \mathcal P$.
It is not difficult to check by simple case analysis that any set intersecting all sets in $\mathcal P$ must actually contain a set from $\mathcal P$. Thus, we get that $\mathcal J^*\subset \{S\subset J\ :\ P\subset S\text{ for some }P\in \mathcal P\}.$ We thus may also assume that $J = \bigcup_{P\in \mathcal P} P = [7]$ and $\mathcal J = \mathcal D_{3/7}$.
Summarizing the discussion above, we may assume that the junta that approximates ${\mathcal F}$ is isomorphic to $\mathcal D_{3/7}$. The last step is to show that ${\mathcal F}$ is actually isomorphic to $\mathcal D_{3/7}$. Assume that ${\mathcal F}':={\mathcal F}\setminus \mathcal J^*$ is non-empty. Then, for each $F\in {\mathcal F}'$, there exists $P\in \mathcal P$ such that $F\cap P = \emptyset$. For each $P\in \mathcal P$, put $$\begin{aligned}
\mathcal J(P)\ :=\ &\Big\{F\ :\ P\subset F\in {[n]\choose k}\Big\},\\
\mathcal {\mathcal F}(P)\ :=\ &\mathcal J(P)\cap {\mathcal F},\\
{\mathcal F}'(P)\ :=\ &\big\{F\ :\ F\in {\mathcal F}',\ F\cap P=\emptyset \big\}.\end{aligned}$$ Then ${\mathcal F}(P)$ and ${\mathcal F}'(P)$ are cross-intersecting, moreover, the former family is $(k-3)$-uniform, while the second one is $k$-uniform and satisfies $|{\mathcal F}'(P)|\le \epsilon{n-5\choose k-4}\le_C {n-7\choose k-4}$. Thus, we can apply with $k-3,k$ and $n-3$ playing the roles of $a,b$ and $n$, respectively, and get that replacing ${\mathcal F}(P)\cup {\mathcal F}'(P)$ with $\mathcal J(P)$ in ${\mathcal F}$ will strictly increase the size of ${\mathcal F}$ (given that it was not there in the first place). Repeating this for each $P\in\mathcal P$, we transform ${\mathcal F}$ into $\mathcal D_{3/7}$. Moreover, if there was some change made in the process, then $|{\mathcal F}|<|\mathcal D_{3/7}|$. Thus, we conclude that ${\mathcal F}$ is isomorphic to $\mathcal D_{3/7}$.
Conclusion {#sec6}
==========
In this paper, we found the largest intersecting family ${\mathcal F}\subset {[n]\choose k}$ with $\tau({\mathcal F})\ge 3$, provided $n>Ck$ for some absolute constant $C$. Actually, we proved more than that. Provided that $\gamma({\mathcal F})\le {n-5\choose k-4}$ and $\tau({\mathcal F})\ge 3$, we showed that $|{\mathcal F}|\le |\mathcal C_3(n,k)|$, with equality only if ${\mathcal F}$ is isomorphic to $\mathcal C_3(n,k)$, for any $n>2k\ge 8$. With some effort, we can replace the condition on diversity by $\gamma({\mathcal F})< {n-4\choose k-3}$. However, since we use juntas result to show that the diversity of any extremal ${\mathcal F}$ must be small, the condition $n>Ck$, $k>C$ is necessary for our approach. This motivates the following problem.
Determine $c(n,k,3)$ for all $n>2k$.
We believe that the family $\mathcal C_3(n,k)$ should be extremal for all $n>2k\ge 10$.
What can one say about $c(n,k,t)$ for $t\ge 4$? Using juntas, one can similarly show that the largest family must have very small diversity (say, at most ${n-2t\choose k-2t}$) for $n>C(t)k$. Using similar proof logic, this allows to deduce the result of [@FOT1] for $n>Ck^4$. The main difficulty (at least for $n>Ck^t$) for $t\ge 5$ lies in the following problem.
\[propbol\] Given an intersecting family ${\mathcal F}$ of $k$-sets with $\tau({\mathcal F})=t$, what is the maximum number of hitting sets of size $t$ it may have?
Finally, we state the following question:
\[propbol\] What is the maximum of $\gamma({\mathcal F})$ for intersecting ${\mathcal F}\subset {[n]\choose k}?$
I showed in [@Kup21] that $\gamma({\mathcal F})\le {n-3\choose k-2}$ for $n>Ck$ with some absolute constant $C$. In particular, this makes Theorem \[thmfull1\] complete: in that range, it covers all possible values of diversity. Following Frankl, I conjectured that the same bound should hold for $n\ge 3k$, however, a counterexample was provided by Huang [@Hua] for $n\le (2+\sqrt 3)k$. I still believe that $\gamma({\mathcal F})\le {n-3\choose k-2}$ should hold for moderately large $n$, say, for $n\ge 5k$. The case of smaller $n$ is also very interesting.\
[Acknowledgements:]{} I would like to thank Peter Frankl for introducing me to the area and for numerous interesting discussions we had on the topic.
[111]{} A. Arman and T. Retter, [*An upper bound for the size of a k-uniform intersecting family with covering number k*]{}, J. Comb. Theory Ser. A 147 (2017), 18–26.
A. Björner and G. Kalai, [*An extended Euler-Poincaré theorem*]{}, Acta Math. 161 (1988), N1, 279–303.
B. Bollobás, [*On generalized graphs*]{}, Acta Math. Acad. Sci. Hungar 16 (1965), 447–452.
B. Bollobás and A. Thomason, [*Threshold functions*]{}, Combinatorica 7 (1986) 35–38.
D. D. Cherkashin, [*On hypergraph cliques with chromatic number 3*]{}, Moscow J. Comb. Numb. Th. 1 (2011), N3.
S. Das and T. Tran, [*Removal and Stability for Erdős–Ko–Rado*]{}, SIAM J. Disc. Math. 30 (2016), 1102–-1114.
M. Deza and P. Frankl, [*Erdős–Ko–Rado Theorem — 22 Years Later,*]{} SIAM J. Algebraic Discrete Methods 4 (1983), N4, 419–431.
I. Dinur, E. Friedgut, [*Intersecting families are essentially contained in juntas*]{}, Comb. Probab. Comput. 18 (2009), 107–122.
D. Ellis, N. Keller and N. Lifshitz, [*Stability versions of Erd˝os–-Ko-–Rado type theorems, via isoperimetry*]{} (2016), arXiv:1604.02160
P. Erdős, [*Problems and results in combinatorial analysis*]{}, Théorie Combinatorie, Colloq. int. Roma 2 (1973), Acad. Naz. Lincei, Roma (1976), 3–17.
P. Erdős, C. Ko, R. Rado, *Intersection theorems for systems of finite sets*, The Quarterly Journal of Mathematics, 12 (1961), N1, 313–320.
P. Erdős, L. Lovász, [*Problems and results on 3-chromatic hypergraphs and some related questions*]{}, in: Infnite and Finite Sets, Proc. Colloq. Math. Soc. János Bolyai, Keszthely, Hungary (1973), North-Holland, Amsterdam (1974), 609–627. C. Feghali, [*Multiply Erdős–Ko–Rado Theorem*]{} (2017), arXiv:1712.09942.
P. Frankl, [*On intersecting families of finite sets*]{}, J. Comb. Theory Ser. A 24 (1978), N2, 146–161.
P. Frankl, [*On intersecting families of finite sets*]{}, Bull. Austral. Math. Soc. 21 (1980), 363–-372. P. Frankl, *Erdős–Ko–Rado theorem with conditions on the maximal degree*, J. Comb. Theory Ser. A 46 (1987), N2, 252–263.
P. Frankl, [*Antichains of fixed diameter*]{}, Moscow J. Comb. Numb. Th. 7 (2017), N3.
P. Frankl, [*A near exponential improvement on a bound of Erdős and Lov' asz*]{}, preprint, <https://users.renyi.hu/~pfrankl/2017-4.pdf>
P. Frankl, J. Han, H. Huang, Y. Zhao [*A degree version of the Hilton–Milner theorem*]{}, J. Comb. Theory Ser. A 155 (2018), 493–502.
P. Frankl, A. Kupavskii, [*Counting intersecting and pairs of cross-intersecting families*]{}, Comb. Probab. Comput. 27 (2018), N1, 60–68.
P. Frankl, A. Kupavskii, [*Erdős–Ko–Rado theorem for $\{0, \pm 1\}$-vectors*]{}, J. Comb. Theory Ser. A 155 (2018), 157–1792. P. Frankl, M. Matsumoto, I. Ruzsa and N. Tokushige, [*Minimum shadows in uniform hypergraphs and a generalization of the Takagi function*]{}, J. Comb. Theory Ser. A 69 (1995), 125–148.
P. Frankl, K. Ota and N. Tokushige, [*Uniform intersecting families with covering number four*]{}, J. Comb. Theory Ser. A 71 (1995), 127–145. P. Frankl, K. Ota and N. Tokushige, [*Covers in uniform intersecting families and a counterexample to a conjecture of Lovász*]{}, J. Comb. Theory Ser. A 74 (1996), 33–42.
P. Frankl, N. Tokushige, *Some best possible inequalities concerning cross-intersecting families*, J. Comb. Theory Ser. A 61 (1992), N1, 87–97.
P. Frankl, N. Tokushige, [*A note on Huang–Zhao theorem on intersecting families with large minimum degree*]{}, Discrete Mathematics 340 (2016), N5, 1098–1103.
P. Frankl and N. Tokushige, [*Invitation to intersection problems for finite sets*]{}, J. Comb. Theory Ser. A 144 (2016), 157–211.
E. Friedgut, [*On the measure of intersecting families, uniqueness and stability*]{}, Combinatorica, 28 (2008), 503–528.
Z. Füredi, [*Erdős–Ko–Rado type theorems with upper bounds on the maximum degree*]{}, Colloquia Math. Soc. J. Bolyai 25 (1978), Szeged, 177–207.
Z. Füredi, [*Maximum degree and fractional matchings in uniform hypergraphs,*]{} Combinatorica 1 (1981), N2, 155–162.
J. Han, Y. Kohayakawa, [*The maximum size of a non-trivial intersecting uniform family that is not a subfamily of the Hilton–Milner family*]{}, Proc. Amer. Math. Soc. 145 (2017) N1, 73–87. H. Hán, Y. Person, M. Schacht, [*On perfect matchings in uniform hypergraphs with large minimum vertex degree*]{}, SIAM J. Disc. Math. 23 (2009), N2, 732–748.
A.J.W. Hilton, E.C. Milner, *Some intersection theorems for systems of finite sets*, Quart. J. Math. Oxford 18 (1967), 369–384.
H. Huang, [*Two extremal problems on intersecting families*]{} (2018), arXiv:1804.11269
H. Huang and Y. Zhao, [*Degree versions of the Erdős–Ko–Rado theorem and Erdős hypergraph matching conjecture*]{}, J. Comb. Theory Ser. A 150 (2017), 233–247. to appear
F. Ihringer, A. Kupavskii, *Regular intersecting families* (2017), arXiv:1709.10462
G. Katona, *A theorem of finite sets*, “Theory of Graphs, Proc. Coll. Tihany, 1966”, Akad, Kiado, Budapest, 1968; Classic Papers in Combinatorics (1987), 381–401.
N. Keller, N. Lifshitz, [*The junta method for hypergraphs and the Erdős-Chvátal simplex conjecture*]{} (2017), arXiv:1707.02643
A. Kostochka, Dhruv Mubayi, [*The structure of large intersecting families*]{} Proc. Amer. Math. Soc. 145 (2017), N6, 2311–2321. J.B. Kruskal, *The Number of Simplices in a Complex*, Mathematical optimization techniques 251 (1963), 251–278.
A. Kupavskii, [*Diversity of intersecting families*]{}, European Journal of Combinatorics 74 (2018), 39–47.
A. Kupavskii, [*Degree versions*]{}, preprint
A. Kupavskii, D. Zakharov, [*Regular bipartite graphs and intersecting families*]{}, J. Comb. Theory Ser. A 155 (2018), 180–189.
[^1]: Indeed, the difference $|\mathcal J_i|-|\mathcal J_{i+1}|$ is ${n-k-2\choose k-2}-1$ for $i=1$ and ${n-k-i\choose k-1}-{n-k-i-1\choose k-1}={n-k-i-1\choose k-2}$ for $i\ge 2$.
[^2]: Cf. the discussion in the beginning of Section \[sec31\] and note the relation to the lexicographic families, defined in Section \[sec3\]: $\mathcal E_l(\bar 1)$ is isomorphic to $\mathcal L([2,n],l,k)$. Using Theorems \[thmfull2\] and \[thmfulleq\] (or by tedious direct calculation), we can conclude that $|\mathcal E_{k-1}|<|\mathcal E_{k-2}| = |\mathcal E_{n-k}|<|\mathcal E_{k-3}|<\ldots<|\mathcal E_{1}|$.
[^3]: The goal of their paper, was, in a sense, to draw the attention of the researchers to this method.
[^4]: The size of $\mathcal J_{k-1}$ can be bounded by ${n-2\choose k-2}+{n-3\choose k-2}+2+(k-2)\big({n-4\choose k-2}-{n-k-2\choose k-2}\big)<{n-2\choose k-2}+2{n-3\choose k-2}$ for $n>2k^2$, say.
[^5]: Note that if $b_1=1$ then $\gamma\ge {n-1\choose k}\ge {n-1\choose k-1}$, which contradicts our assumption on $\gamma$.
[^6]: We note that, in this case, we do not use the fact that ${{\mathcal A}},{{\mathcal B}}$ form a resistant pair (and thus the proof works for $T=\{2,3,4\}$).
[^7]: Assume that we have equality in . The Kruskal–Katona theorem in Lovasz’ form states that if $\mathcal K$ a family of $m$-element sets and, for some real $x\ge m$, we have $|\mathcal K|\ge {x\choose m}$, then $|\partial \mathcal K|\ge {x\choose m-1}$. In our case, $\mathcal K:={\mathcal F}([2,i-1],[i])\subset {[i+1,n]\choose k-i+1}$ and ${n-5\choose k-3}={x\choose k-i+1}$, where $x\le n-5$ for any $i\in [2,4].$ Finally, we have $\frac{{x\choose k-i+1}}{{x\choose k-i}}=\frac {x-k+i}{k-i+1}\le \frac {n-k-1}{k-3}=\frac{{n-5\choose k-3}}{{n-5\choose k-4}}$, and thus $|\partial \mathcal K|> {x\choose k-i}\ge {n-5\choose k-4}$.
[^8]: Note that this is equivalent to requiring that $|\mathcal H|>2$.
[^9]: Otherwise, $\mathcal H$ would have common intersection. Note that $H_{i}\setminus \{i_2,\ldots, i_{z}\}$ is a set of size $s-z+2$.
[^10]: We use a subscript “$C$” in the inequalities, i.e., $\ge_C$, in which we need that $C$ is sufficiently large.
[^11]: We note that the remaining part of the argument works for any $n>2k\ge 8$.
[^12]: This is only needed for the uniqueness of the extremal family $\mathcal C_3(n,k)$, since some of the exchanges we shall perform below may not necessarily strictly increase the size. But this does not pose problems since we will eventually arrive at a family ${\mathcal F}''$ with ${\mathcal F}''(\bar 1)={\mathcal M}$, which we will show to have strictly smaller size than that of $\mathcal C_3(n,k)$.
[^13]: As in the proof of Theorem \[thmtau3\], we use a subscript “$C$” in the inequalities, in which we need that $C$ is sufficiently large.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using first-principles calculations, we estimated the impact of large applied electric $E$ fields on the structural, dielectric, and ferroelectric properties of typical ferroelectrics. At large fields, the structural parameters change significantly, decreasing the strain between the different structural phases. This effect favours a polarization rotation model for ferroelectric switching in which the electronic polarization rotates between the directions of tetragonal, rhombohedral and orthorhombic phases. We estimate coercive fields $E_c$$\sim$31 MV/m and $\sim$52 MV/m at zero temperature for bulk ferroelectric monodomains of [BaTiO$_3$]{} and [PbTiO$_3$]{}, respectively. The dielectric permittivity and tunability of [BaTiO$_3$]{} are the least affected at large fields, making this material attractive for applications in electronics and energy storage.'
author:
- 'Daniel I. Bilc'
- Liviu Zarbo
- Sorina Garabagiu
- Eric Bousquet
- Liliana Mitoseriu
title: 'High field properties of typical perovskite ferroelectrics by first-principles modeling'
---
Understanding the behaviour of advanced oxide materials under applied finite electric field $E$ is very important for high performance technological applications. Many of such tunable oxide materials are composites with various nano/microstructural characteristics designed between typical ferroelectric (FE), like BaTiO$_3$, PbTiO$_3$ and [SrTiO$_3$]{}, and linear dielectric materials. In such compounds, the electric field $E$ can vary strongly at local level, and affect the local structural, dielectric and FE properties. Such properties as electrostriction, tunability, dielectric loss, polarization switching are of great interest and they enter into the coefficient of performance of many devices designed for advanced technological applications. In terms of wireless communication technologies, the tunability is one of the essential characteristics which must be well characterized and optimized at high values in these composite ceramics. The high interest in such oxide materials is also motivated by the more recent finding of coexistence of ferroelectricity with the metallic character of surface states in topological FE oxides, which opens new perspectives for electronic, spintronic and quantum technology applications.[@Felser; @Kee; @Kim; @Nagaosa]
In this respect, we studied the structural (unit cell volume and structural parameters), dielectric (dielectric permittivity $\epsilon$, and tunability), and FE (electronic polarization $P$ and polarization switching) properties of a few typical perovskite FE’s under finite applied $E$ using first principles methods based on density functional theory (DFT), which goes beyond the more common approach based on empirical models such Landau theory [@Landau; @Ginzburg2; @Devonshire1; @Devonshire3] or Johnson model.[@Johnson] Different first principles methods have been proposed to perform calculations under finite applied $E$ [@Fernandez; @Umari; @Cohen; @Sai; @Souza; @Bellaiche; @Dieguez] or electric displacement $D$ [@Hong; @Hong2013] fields. The practical calculations for periodic crystalline materials are more difficult to carry out than the zero field DFT calculations because the electric potential generated by $E$ field is linear, non-periodic and not compatible with the periodic boundary conditions. Moreover, the electric potential is unbounded giving rise to the Zener effect (interband tunneling). Thus, we have used the method developed by Vanderbilt [*et. al.*]{},[@Souza] which minimizes the electrical enthalpy functional with respect to Bloch type wave functions $\psi_{nk} (r) = e^{ikr} u_{nk} (r)$: $$\label{Eq1}
\mathcal{F}[u_{nk};E]=U_{KS} - \Omega PE$$ where $U_{KS}$ are the Kohn Sham energies routinely determined in DFT, $- \Omega PE$ is the coupling energy between polarization $P$ and $E$ field, $\Omega$ is the unit cell volume, and $u_{nk}$ are the Bloch wave functions. The two terms from electrical enthalpy are calculated considering the wave function $\psi_{nk}$ Bloch polarizable under the influence of external $E$ field. This method can be employed to perform atomic and structural relaxations in the presence of $E$ field for the materials of interest, and to estimate the field-dependent properties.
Our theoretical estimations of field-induced properties at zero absolute temperature were performed using the GGA approximation of Wu and Cohen, GGA-WC,[@GGA-WC] for the exchange and correlation energy functional as implemented in ABINIT code,[@ABINIT] and using optimized pseudopotentials generated with OPIUM.[@OPIUM] The electronic states considered as valence states are: 4$s$, 4$p$, and 5$s$ for Sr, 5$s$, 5$p$, and 6$s$ for Ba, 6$s$, 6$p$, and 5$d$ for Pb, 3$s$, 3$p$, and 3$d$ for Ti, and 2$s$ and 2$p$ for O. An energy cutoff of 45 hartree was used for the plane-wave expansion of the wave functions and the Brillouin zone integrations were performed using 14$\times$14$\times$14 and 14$\times$12$\times$12 grids of $k$ points for five atom cubic/tetragonal/rhombohedral, and ten atom orthorhombic structures, respectively. The self-consistent-field calculations were considered to be converged when the total energy changes were smaller than 10$^{-12}$ hartree. Atomic relaxations at fixed cubic lattice parameter for the cubic structures, and full relaxations (atomic and cell geometry) for the tetragonal, rhombohedral, and orthorhombic structures were performed until the forces were smaller than 5$\times$10$^{-7}$ hartree/bohr.
GGA-WC describes very well the structural and FE properties of typical ferroelectrics and does not have the supertetragonality problem (overestimation of $c/a$ ratio, atomic displacements and $P$) which usual GGA’s have for the tetragonal structures.[@Bilc2008] The values of quantities describing the spontaneous structural and FE properties (for $E=0$) for the tetragonal structures of BaTiO$_3$ and PbTiO$_3$ are included in Table \[Table1\]. The overestimation of spontaneous tetragonality $c/a$ ratio and polarization $P_s$ for BaTiO$_3$ are due to the accuracy of GGA-WC pseudopotentials used in these calculations. Note that our previous estimations of the structural and FE properties for BaTiO$_3$ and PbTiO$_3$ using the same GGA-WC functional and CRYSTAL code are in a better agreement with the experimental values.[@Bilc2008]
[0.48]{} [@ccccc]{} & &\
& GGA-WC & Exp. & GGA-WC & Exp.\
$a$(Å) & 3.973 & 3.986$^a$ & 3.89 & 3.88$^b$\
$c/a $ & 1.023 & 1.010$^a$ & 1.093 & 1.071$^b$\
$dz_{\rm Ti}$ & 0.015 & 0.015$^a$ & 0.038 & 0.040$^c$\
$dz_{\rm O_{\parallel}}$ & -0.03 & -0.023$^a$ & 0.119 & 0.112$^c$\
$dz_{\rm O_{\perp}}$ & -0.019 & -0.014$^a$ & 0.124 & 0.112$^c$\
$P$(C/m$^2$) & 0.34 & 0.27$^d$ & 0.97 &0.5-1.0$^e$\
\
\
\
\
\
The next step was to estimate the field-dependence of polarization $P(E)$ and of static dielectric permittivity $\epsilon_s = \epsilon(E)$ for the cubic structures of [SrTiO$_3$]{}, [BaTiO$_3$]{} and [PbTiO$_3$]{}. The cubic lattice parameter was kept constant, but the electronic and atomic positions have been relaxed in the presence of $E$ field, which was oriented in the crystallographic direction $z$. We followed this procedure to estimate $P(E)$ and $\epsilon(E)$, which correspond to the region of transition from the cubic to tetragonal structure in the phase vs temperature diagram, where the values of $\epsilon(E)$ are maximized at the transition zone boundary. The polarization $P(E)$ dependence on field is shown in Figure \[PscFig\]. In order to estimate $\epsilon(E\rightarrow0)$ for small fields, we considered the dependence of $P(E)$ on the dielectric susceptibilities including up to the cubic terms, $P(E) = \chi^{(1)}E + \chi^{(2)}E^{2} + \chi^{(3)}E^{3}$, and $\epsilon(E) = 1 + \chi^{(1)}(E)/\epsilon_0$, where $\chi^{(1)}(E)$ is the linear term of susceptibility and $\epsilon_0$ is the vacuum permittivity. By fitting $P(E)$ values for small $E$ fields, we estimate $\epsilon_s(E\rightarrow0)$ values of: $\sim$270, $\sim$46 and $\sim$56 for [SrTiO$_3$]{}, [BaTiO$_3$]{} and [PbTiO$_3$]{}, respectively. The $\epsilon_s$ value of $\sim$270 for [SrTiO$_3$]{} is smaller than the value of 391 estimated within LDA from Ref. \[\]. These theoretical values of $\epsilon_s$ for [SrTiO$_3$]{} compare well with the room temperature experimental value of $\sim$300, but they are largely underestimating the zero temperature experimental value of $\sim$20000.[@Sakudo] At low temperature the antiferrodistortive (AFD) oxygen octahedral rotations (a$^0$a$^0$c$^-$ in Glazer notations [@Glazer]) are present in [SrTiO$_3$]{}, which strongly influences the polar ferroelectric mode and $\epsilon_s$. The compressive inplane strain to [SrTiO$_3$]{}, increases these AFD rotations and $\epsilon_s$ which is comparable to the zero temperature experimental value for strain corresponding to transition zone boundary between paraelectric and FE tetragonal structures.[@Antons] The estimation of $\chi^{(1)}(E)$ at a given $E$ field can be made by the method of finite differences $\chi^{(1)}(E) = \Delta P(E)/\Delta E$. Using this method, we obtained the following $\epsilon_{s}(E)$ values of $\sim$235, $\sim$46, and $\sim$56 at 3 MV/m for [SrTiO$_3$]{}, [BaTiO$_3$]{} and [PbTiO$_3$]{}, respectively. These results show that [SrTiO$_3$]{} has a large relative tunability $n_{r}=[\epsilon_s(E\rightarrow0) - \epsilon_s(E)]/\epsilon_s(E\rightarrow0)$ of $\sim$ 13$\%$ at 3 MV/m, $n_{r}$ value that increases strongly at transition zone boundary,[@Antons] whereas the cubic phases of [BaTiO$_3$]{} and [PbTiO$_3$]{} do not show any tunability.
![\[PscFig\] (Color online) Electronic polarization $P$ dependence on electric $E$ field for the bulk cubic perovskite structure of [SrTiO$_3$]{}, [BaTiO$_3$]{} and [PbTiO$_3$]{}.](PvsE_sc_iu "fig:")\
For modeling the FE switching properties, we considered a model in which polarization rotates between the different polarization directions of the structural phases: tetragonal, orthorhombic, and rhombohedral, as the strength of applied field increases in a given direction (Figure \[SwitchModel\]). This approach was not often used in literature because the Ising model is believed to be favoured by the strain conditions to which the switching units ([*histerons*]{}) are subjected. Ising model involves the crossing of switching units through the high-symmetry structures (paraelectric structures) and the cancellation of polarization, $P = 0$. Previous theoretical studies for [BaTiO$_3$]{} showed that field-induced polarization rotation can occur,[@Garcia] and is responsible for the ultrahigh electromechanical response.[@Cohen] High piezoelectric effects and field-induced transition in [BaTiO$_3$]{} single-crystals were explained by a coherent polarization rotation within a Landau-based approach combined with crystalline anisotropy and field effects.[@Zang] Close to the structural FE-FE transitions of [BaTiO$_3$]{}, the polarization rotation instability was considered as responsible for the divergence of shear elastic compliance.[@Nanni]
![\[SwitchModel\] (Color online) The switching model through rotation of polarization $P$ in external $E$ field. The switching is achieved by $P$ rotation between the different phases: tetragonal (1) and (5), rhombohedral (2) and (4), and orthorhombic (3). $E$ field is oriented in \[0,0,1\] direction ($z$) and $P$ along the directions given by structural phases.](BaTiO3_Switching "fig:")\
We obtain the coercive fields $E_c$ from the dependence on $E$ of total energy corresponding to electric enthalpy given by Eq. \[Eq1\] for the different structural phases (see Figure \[SwitchModel\]), which were fully relaxed (atomic and stress relaxations) in the presence of field. In Figure \[EnFig\](a), we show this dependence for the total energy difference $d$E$_{tot}$ of a [BaTiO$_3$]{} monodomain relative to the orthorhombic structure for which $E \perp P$ (Figure \[SwitchModel\](3)). The orthorhombic structure is chosen as a reference for total energy since it is the least affected by $E$ field as $\Omega PE$ term is negligible. Structures with $d$E$_{tot}$ $<$ 0, are favourable in terms of structural stability (filled symbols in Figure \[EnFig\](a)). The coercive fields $E_c$ were estimated from the field values at which the three structural phases (tetragonal, rhombohedral and orthorhombic) are comparable in energy ($d$E$_{tot}$ $\sim$ 0 meV) and $P$ can easily rotate between the directions corresponding to the three [BaTiO$_3$]{} symmetries. These results suggest coercive fields $E_c \sim$ 31 MV/m for a [BaTiO$_3$]{} monodomain. This $E_c$ value agrees very well with the field value of 30 MV/m at which the large piezoelectric response was estimated within LDA for [BaTiO$_3$]{} [@Cohen; @Stolbov] and the values of $\sim$20 MV/m [@Waghmare] and $\sim$150 MV/m [@Stolbov] computed by molecular dynamics simulations at 100 K.
![\[EnFig\] (Color online) Electric field dependence of the total energy difference $d$E$_{tot}$ relative to orthorhombic structure for the different structural phases of a bulk monodomain: (a) [BaTiO$_3$]{} and (b) [PbTiO$_3$]{}. Negative $d$E$_{tot} <$ 0 values favourable in terms of structural stability are represented by filled symbols. Coercive $E_c$ fields, where all structural phases have comparable total energies are marked by arrows. The positive $E$ fields are oriented along \[0,0,1\] direction.](EdiffvsE_tet_iu_BTO "fig:")\
![\[EnFig\] (Color online) Electric field dependence of the total energy difference $d$E$_{tot}$ relative to orthorhombic structure for the different structural phases of a bulk monodomain: (a) [BaTiO$_3$]{} and (b) [PbTiO$_3$]{}. Negative $d$E$_{tot} <$ 0 values favourable in terms of structural stability are represented by filled symbols. Coercive $E_c$ fields, where all structural phases have comparable total energies are marked by arrows. The positive $E$ fields are oriented along \[0,0,1\] direction.](EdiffvsE_tet_iu_PTO "fig:")\
![\[PFig\] (Color online) Electronic polarization $P$ dependence on electric $E$ field for the different structural phases of a bulk monodomain: (a) [BaTiO$_3$]{} and (b) [PbTiO$_3$]{}. $P$ values corresponding to $d$E$_{tot} <$ 0 are represented by filled symbols. The positive $E$ fields are oriented along \[0,0,1\] direction.](PvsE_tet_iu_BTO "fig:")\
![\[PFig\] (Color online) Electronic polarization $P$ dependence on electric $E$ field for the different structural phases of a bulk monodomain: (a) [BaTiO$_3$]{} and (b) [PbTiO$_3$]{}. $P$ values corresponding to $d$E$_{tot} <$ 0 are represented by filled symbols. The positive $E$ fields are oriented along \[0,0,1\] direction.](PvsE_tet_iu_PTO "fig:")\
For the switching mechanism of a [PbTiO$_3$]{} monodomain, which has the polar tetragonal ground structure, we considered also the other polar orthorhombic and rhombohedral structures as “metastable”, similar to those of [BaTiO$_3$]{}. For $E$ values $\sim$52 MV/m, which are higher compared to those of [BaTiO$_3$]{}, the total energy of the three phases are comparable ($d$E$_{tot}$ = 0, Figure \[EnFig\](b)). The higher coercive field $E_c \sim$ 52 MV/m of [PbTiO$_3$]{} is due to larger $d$E$_{tot}$ values in the absence of field by comparison with the case of [BaTiO$_3$]{} ($|d$E$_{tot}|$ $\sim$ 20 meV for [PbTiO$_3$]{} and $|d$E$_{tot}|$ $\sim$ 5 meV for [BaTiO$_3$]{} at $E$ = 0). This $E_c$ value correspond to a linear piezoelectric response of [PbTiO$_3$]{} [@Roy], being smaller than the value of $\sim$270 MV/m estimated by molecular dynamics simulations at 100 K.[@Cohen2011] We would also like to emphasize that [PbTiO$_3$]{} has double well energy, $d$E$_{dw}$ (total energy difference between polar tetragonal and nonpolar cubic structures) and a much higher polarization than those of [BaTiO$_3$]{} ($d$E$_{dw}$ $\sim$ 100 meV, $P$ $\sim$ 0.98 C/m$^2$ for [PbTiO$_3$]{}, and $d$E$_{dw}$ $\sim$ 14 meV and $P$ $\sim$ 0.26 C/m$^2$ for [BaTiO$_3$]{}).[@Bilc2008] These results show that the required energy $d$E$_{tot}$ to be supplied externally by the electric field $E$ to a monodomain in the polarization rotation mechanism of switching are much lower compared with the necessary energy $d$E$_{dw}$ for the Ising model of polarization switching. The results suggest the polarization rotation as the dominant switching mechanism at high $E$ fields compared to the Ising model of switching. Our results are supported by the first principles DFT calculations performed for [PbTiO$_3$]{} under finite $D$ field within LDA, which find the polarization rotation through tetragonal, rhombohedral and orthorhombic phases to be more favourable than the Ising model of polarization switching that requires $E_c$ values of $\sim$250 MV/m.[@Hong]
![\[StFig\] (Color online) Electric field dependence of the unit cell volume and structural parameters for the different phases of bulk: (a) [BaTiO$_3$]{} and (b) [PbTiO$_3$]{}. ](Batio3_volumes "fig:")![\[StFig\] (Color online) Electric field dependence of the unit cell volume and structural parameters for the different phases of bulk: (a) [BaTiO$_3$]{} and (b) [PbTiO$_3$]{}. ](Pbtio3_volumes "fig:")\
The structural properties are significantly affected by the applied field (Figure \[StFig\]). The unfavourable energetic case of the tetragonal structure in which $E>$ 0 and $E$ is antiparallel to $P$ (Figure \[SwitchModel\](1)) leads to a decrease in the unit cell volume and $c_{tet}$ parameter such that the strain relative to $a_{ort}$ parameter of orthorhombic structure is reduced supporting $P$ rotation through which $c_{tet}$ becomes $a_{ort}$ and $a_{tet}$ becomes $c_{ort}$. Similarly, the favourable energetic case in which $E <$ 0 and $E$ is parallel with $P$ (Figure \[SwitchModel\](5)) enlarges the unit cell volume and $c_{tet}$ parameter, enlargements which are limited by the strain relative to $a_{ort}$ parameter.
At high $E$ fields ($E \sim E_c$), the polarization is also significantly affected. In the tetragonal structures, $P$ differs by $\sim 3\%$ for [BaTiO$_3$]{} at $E$ = 31 MV/m and respectively by $\sim 6\%$ for [PbTiO$_3$]{} at $E$ = 52 MV/m (Figure \[PFig\](a) and (b)). Considering the field-induced polarization variations from Figure \[PFig\] and the estimated coercivity $E_c$ values, we show in Figure \[HisteronFig\] the histeron model for tetragonal phase of [BaTiO$_3$]{} and [PbTiO$_3$]{} monodomains, which may be further used in multiscale modeling methods required to describe the properties of complex ferroelectric materials. Fitting $P(E)$ values from Figure \[PFig\], we estimate $\epsilon_s(E\rightarrow0)$ $\sim$48 and $\sim$110, for the tetragonal phases of [BaTiO$_3$]{} and [PbTiO$_3$]{}, respectively. The finite difference method gives the following $\epsilon_s(E)$ values of $\sim$49 and $\sim$104 at $E$ = 3 MV/m for the tetragonal FE phases of [BaTiO$_3$]{} and [PbTiO$_3$]{}, respectively. These results show an enhanced dielectric tunability for the tetragonal FE phase of [BaTiO$_3$]{} and [PbTiO$_3$]{}, comparing with that of the cubic paraelectric phase.
![\[HisteronFig\] (Color online) Histeron for the tetragonal phase of [BaTiO$_3$]{} and [PbTiO$_3$]{} bulk monodomains. Coercive $E_c$ fields are indicated by vertical lines. Polarization values corresponding to $d$E$_{tot} <$ 0 are represented by filled symbols.](PvsE_tet_Histeron_iu "fig:")\
In this work, we show the role of large external fields on the structural, dielectric and FE properties of a few typical ferroelectric perovskites. Whenever present in various types of material combinations (solid solutions, composites, multilayers), the properties of ferroelectrics are strongly affected by electric fields and this may significantly affect the device performance. We find [BaTiO$_3$]{} to have comparable static dielectric permittivities in the cubic and tetragonal bulk phases, which combined with the high dielectric tunability of [SrTiO$_3$]{}, make (BaSr)TiO$_3$ solid solutions among the most interesting materials for tunability applications, since they exhibit slight variation of dielectric permittivity in a large range of temperatures. [BaTiO$_3$]{} shows the smallest dielectric tunability, making it a good candidate to be used in composites for energy storage applications, for which large dielectric permittivity, small tunability and losses are require. The FE monodomain states of [BaTiO$_3$]{} and [PbTiO$_3$]{} show high coercive $E_c$ fields of $\sim$30-50 MV/m. These intrinsic $E_c$ fields fall between intrinsic values computed by other models or experiments considered to give intrinsic limits, since they were determined in ultrathin films such as: 5 MV/m in single crystal,[@Fridkins] 19.5 MV/m in Devonshire theory,[@Janovec] 70 MV/m in mean field theory,[@Fridkins] 25-160 MV/m in ultrathin polycrystalline films,[@Fridkins; @Fridkins2; @Jo] 196 MV/m in Ginzburg - Landau theory.[@Fridkins3] The large electric fields favour a FE switching model in which electronic polarization rotates between the different polarization orientations of the structural phases. This model is supported by the continuous polarization switching mechanism (Bloch type switching), which was recently predicted to be favorable for bulk [BaTiO$_3$]{} rhombohedral phase.[@Hlinka] We hope our findings will stimulate other more complex material studies under finite external $E$ field needed in many practical applications.
The authors acknowledge financial support from the Romanian National Authority for Scientific Research, CNCS-UEFISCDI, Project number PN-II-PT-PCCA-2013-4-1119. This work was supported by FRS-FNRS Belgium (E. B.) and E.B. acknowledges the Consortium des Equipements de Calcul Intensif (CECI), funded by the FRS-FNRS (Grant 2.5020.11).
[99]{}
B. Yan, M. Jansen, and C. Felser, Nat. Phys. [**9**]{}, 709 (2013). J. M. Carter, V. V. Shankar, M. A. Zeb, and H. Y. Kee, Phys. Rev. B [**85**]{}, 115105 (2012). B. J. Yang and Y. B. Kim, Phys. Rev. B [**82**]{}, 085111 (2010). A. Shitade, H. Katsura, J. Kunes, X. L. Qi, S. C. Zhang, and N. Nagaosa, Phys. Rev. Lett. [**102**]{}, 256403 (2009). L. D. Landau, I. M. Khalatnikov, Dokl. Acad. Nauk. SSSR [**96**]{}, 469 (1954). V. L. Ginzburg, Sov. Phys. Solid State [**2**]{},1824 (1960). A. F. Devonshire, Phyl. Mag. [**40**]{}, 1040 (1949). A. F. Devonshire, Adv. Phys. [**3**]{}, 85 (1954). K. Johnson, J. Appl. Phys. [**33**]{}, 2826 (1962). P. Fernández, A. Dal Corso, A. Baldereschi, and F. Mauri, Phys. Rev. B [**55**]{}, R1909 (1997). P. Umari and A. Pasquarello, Phys. Rev. Lett. [**89**]{}, 157602 (2000). H. Fu and R. E. Cohen, Nature [**403**]{}, 281 (2000). N. Sai, K. M. Rabe, and D. Vanderbilt, Phys. Rev. B [**66**]{}, 104108 (2002). I. Souza, J. Íñiguez, and D. Vanderbilt, Phys. Rev. Lett. [**89**]{}, 117602 (2002). H. Fu and L. Bellaiche, Phys. Rev. Lett. [**91**]{}, 057601 (2003). O. Dieguez and D. Vanderbilt, Phys. Rev. Lett. [**96**]{}, 056401 (2006). J. Hong and D. Vanderbilt, Phys. Rev. B [**84**]{}, 115107 (2011). J. Hong and D. Vanderbilt, Phys. Rev. B [**87**]{}, 064104 (2013). Z. Wu and R. E. Cohen, Phys. Rev. [**B 73**]{}, 235116 (2006). X. Gonze, [*et. al.*]{}, Comput. Phys. Commun. [**180**]{}, 2582 (2009). A. M. Rappe, K. M. Rabe, E. Kaxiras, and J. D. Joannopoulos, Phys. Rev. B [**41**]{}, 1227 (1990). D. I. Bilc, R. Orlando, R. Shaltaf, G. M. Rignanese, J. Iniguez, and Ph. Ghosez, Phys. Rev. B [**77**]{}, 165107 (2008). G. Shirane, H. Danner, and P. Pepinsky, Phys. Rev. [**105**]{}, 856 (1957). S. A. Mabud and A. M. Glazer, J. Appl. Crystallogr. [**12**]{}, 49 (1979). G. Shirane, P. Pepinsky, and B. C. Frazer, Acta Cryst. [**9**]{}, 131 (1956). H. H. Wieder, Phys. Rev. [**99**]{}, 1161 (1955). M. E. Lines and A. M. Glass [ *Principles and Applications of Ferroelectrics and Related Materials*]{} ( Oxford University Press, Oxford, 1977), Chap.8. A. Antons, J. B. Neaton, K. M. Rabe, and D. Vanderbilt, Phys. Rev. B [**71**]{}, 024102 (2005). H. Uwe and T. Sakudo, Phys. Rev. B [**13**]{}, 271 (1967). A. Glazer, Acta Cryst. B. [**28**]{}, 3384 (1972). A. Garcia, and D. Vanderbilt, Appl. Phys. Lett. [**72**]{}, 2981 (1998). H. Zang, J. Appl. Phys. [**113**]{}, 184111 (2013). F. Cordero, H. T. Langhammer, T. M�ller, V. Buscaglia, and P. Nanni, Phys. Rev. B [**93**]{}, 064111 (2016). S. Stolbov, H. Fu, R. E. Cohen, L. Bellaiche, and D. Vanderbilt. [*Fundamental Physics of Ferroelectrics 2000*]{}, (R.E. Cohen, ed.) AIP Conference Proceedings 535, New York, pp. 151-158 (2000). J. Paul, T. Nishimatsu, Y. Kawazoe, U.V. Waghmare, Phys. Rev. B [**80**]{}, 024107 (2009). A. Roy, M. Stengel, and D. Vanderbilt, Phys. Rev. B [**81**]{}, 014102 (2010). X. Zeng and R. E. Cohen, Appl. Phys. Lett. [**99**]{}, 142902 (2011). R. Gaynutdinov, M. Minnekaev, S. Mitko, A. Tolstikhina, A. Zenkevich, S. Ducharme, and V. Fridkin, Physica B, [**424**]{}, 8 (2013). V. Janovec, Czechosl. Journ. Phys. [**8**]{}, 3 (1958). R. Gaynutdinov, M. Minnekaev, S. Mitko, A. Tolstikhina, A. Zenkevich, S. Ducharme, and V. Fridkin, JETP Letters [**98**]{}, 339 (2013). J. Y. Jo, Y. S. Kim, T. W. Noh, Jong-Gul Yoon, and T. K. Song, Appl. Phys. Lett. [**89**]{}, 232909 (2006). V. M. Fridkin, and S. Ducharme, Phys. Solid State [**43**]{}, 1320 (2001). M. Taherinejad, D. Vanderbilt, P. Marton, V. Stepkova, and J. Hlinka, Phys. Rev. B [**86**]{}, 155138 (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the existence of low scale gravity at TeV scale could lead to a direct production of photons with energy above 10$^{22}$ eV due to annihilation of ultra high energy neutrinos on relic massive neutrinos of the galactic halo. Air showers initialized in the terrestrial atmosphere by these ultra energetic photons could be collected in near future by the new generation of cosmic rays experiments.'
author:
- |
**R.V.Konoplich**$^{1}$**, S.G.Rubin**$^{2,3}$\
\
$^{1}$New York University, New York, USA\
$^{2}$Moscow Engineering Physics Institute, Moscow,Russia\
$^{3}$Center for Cosmoparticle Physics ”Cosmion”, Moscow, Russia
title: '**POSSIBLE** **SIGNATURE OF LOW SCALE GRAVITY IN ULTRA HIGH ENERGY COSMIC RAYS**'
---
Recently it was proposed [@1] that the space is 4+n dimensional, with the Standard Model particles living on a brane. While the weakly, electromagnetically, and strongly interacting particles are confined to the brane in 4 dimensions, gravity can propagate also in extra n dimensions. This approach allows to save the gauge hierarchy problem by introducing a single fundamental mass scale (string scale) $M_{s}$ of the order of TeV. The usual Planck scale $M_{Pl}=1/\sqrt{G_{N}}\simeq 1.22\cdot 10^{19}$GeV is related to the new mass scale $M_{s}$ by Gauss’s law:
$$M_{Pl}^{2}\sim R^{n}M_{s}^{n+2} \label{1}$$
where $G_{N}$ is the Newton constant, $R$ is the size of extra dimensions. It follows from (\[1\]) that
$$R\sim 2\cdot 10^{-17}(\frac{TeV}{M_{s}})(\frac{M_{Pl}}{M_{s}})^{2/n},cm
\label{2}$$
gives at $n=1$ too large value, which is clearly excluded by present gravitation experiments. On the other hand $n\geq 2$ gives the value $R\lesssim 0.25$ cm, which is below the present experimental limit $\sim $1 cm but can be tested for the case $n=2$ in gravitational experiments in near future.
It can be shown that the graviton including its excitations in the extra dimensions, so-called Kaluza-Klein (KK) graviton emission, interacts with the Standard Model particles on the brane with an effective amplitude $\sim
M_{s}^{-1}$ instead of $M_{Pl}^{-1}$. Indeed, the graviton coupling to the Standard Model particle $\sim M_{Pl}^{-1}$, the rate [@2] of the graviton interaction $r\sim (M_{Pl}^{-1})^{2}N$, where $N$ is a multiplicity of KK-states. Since this factor is $\sim $ ($\sqrt{S}R)^{n}$, where $\sqrt{S}
$ is the c.m. energy, then substituting $R$ from (\[2\]) we get $r\sim
M_{s}^{-2}$. Thus the graviton interaction becomes comparable in strength with weak interaction at TeV scale.
This leads to the varieties of new signatures in particle physics, astrophysics and cosmology (see e.g. [@2; @3; @4; @5; @6] ) which have already been tested in experiments or can be tested in near future.
In this article we consider the possible signature of the low scale gravity in ultra high energy cosmic rays.
The detection [@7; @8] of cosmic rays with energy above Greisen-Zatsepin-Kuzmin (GZK) cut-off of $\sim 5\cdot 10^{19}eV$ presents a serious problem for interpretation. The origin of GZK cut-off [@9] is due to resonant photoproduction of pions by protons on cosmic microwave background radiation which leads to a significant degradation of proton energy (about 20% for 6 Mpc) during its propagation in the Universe. Of course, proton energy does not change by many orders of magnitude if high energy protons come from the distances 50 - 100 Mpc. However, no nearby sources like active galactic nuclei have been found up to now in the arrival direction.
It is difficult also to relate the observed ultra high energy events (Fig.1 [@8]) with the other particles. For example in the case of ultra high energy photons due to interaction with cosmic background radiation ($\gamma
+\gamma ^{\ast }\longrightarrow e^{+}+e^{-}$) the photon free mean path should be significantly less than 100 Mpc. A scenario based on direct cosmic neutrinos able to reach the Earth from cosmological distances can not reproduce the observed signatures of ultra high energy air showers occurred high in the atmosphere.
Different possibilities were considered (see e.g. [@10] and references therein) in order to solve this puzzle. In particular it was proposed [@11; @12; @13] that ultra high energy neutrinos reaching the Earth from cosmological distances interact with a halo of relic light neutrinos in the Galaxy, producing due to Z, W$^{\pm }$ boson exchange secondaries inside the galactic halo. Photons from $\pi ^{0}$ decays and nucleons can easily propagate to the Earth and be the source of the observed ultra high energy air showers. Critical elements of models [@11; @12; @13] are: the existence of neutrino mass in the range 0.1-10 eV and significant clustering of relic neutrinos in the halo up to 10$^{5}$n$_{\nu }$, where n$_{\nu }$ is the cosmological neutrino number density (n$_{\nu }$ $\sim 100cm^{-3}$). Also the existence of ultra high energy ( 10$^{21}$-10$^{23}$ eV) neutrino flux is necessary in order to produce multiple secondaries with energies above GZK cut-off.
However if the graviton interaction comparable in strength with weak interaction at TeV scale exists then photons can be produced directly in a reaction $$\nu +\overset{-}{\nu }\longrightarrow g\longrightarrow \gamma +\gamma
\label{3}$$
due to virtual graviton exchange (Fig.2). In Standard Model process (\[3\]) occurs via loop diagram and therefore is severely suppressed.
At high energies the cross section for the process (\[3\]) can be obtained immediately from that for the process $e^{+}e^{-}\longrightarrow \gamma
\gamma $ including graviton exchange (see for example [@3]) by substituting e = 0. Then $$\frac{d\sigma }{dz}=\frac{\pi }{16}\frac{S^{3}}{M_{s}^{8}}F^{2}(1-z^{4})
\label{4}$$
where $\sqrt{S}$ is c.m.s. energy, z = $\left| \cos \theta \right| $ is the polar angle of the outgoing photon. The factor F depends on the number of extra dimensions: $$F=
\begin{array}{c}
\log (M_{s}^{2}/S)\underset{}{} \\
2/(n-2)\overset{}{}
\end{array}
,
\begin{array}{c}
n=2\underset{}{} \\
n>2\overset{}{}
\end{array}
,$$
at $\sqrt{S}<<M_{s}$. In Eq.(\[4\]) it is also taken into account that primary beam of neutrinos is polarized.
Integrating (\[4\]) over the polar angle and including a symmetry factor for two $\gamma $ we get $$\sigma =\frac{\pi }{20}\frac{S^{3}}{M_{s}^{8}}F^{2}\approx 7\cdot
10^{-35}F^{2}\QOVERD( ) {\sqrt{S}}{TeV}^{6}\QOVERD( ) {TeV}{M_{s}}^{8}cm^{2}.
\label{5}$$
On can see from (\[5\]) that at TeV energies the rate of the reaction (\[3\]) is comparable with the rate of weak processes [@11].
Assuming $M_{s}\sim $ $\sqrt{S}\sim $ TeV we find for example for $n=3$ the following probability for the interaction of ultra high energy neutrinos inside the galactic halo: $P\approx \sigma n_{G}L_{G}\sim 10^{-3}$, where $L_{G}\sim 100$ Kpc is the size of the galactic neutrino halo, $n_{G}\sim
10^{5}n_{\nu }$ is the neutrino number density in the galactic halo. This probability is significantly greater than the probability of ultra high energy neutrino interaction in terrestrial atmosphere [@14].
Let us note that nearby galaxies also can be sources of additional ultra high energy photons due to neutrino interaction with relic neutrinos of galactic halos [@12; @13].
TeV range in c.m.s. corresponds to the energy of extragalactic neutrino flux $E\approx 10^{22}-10^{23}$ eV since $$E\approx \frac{S}{2m}\approx 5\cdot 10^{22}\QOVERD( ) {\sqrt{S}}{TeV}^{2}\QOVERD( ) {10eV}{m}eV \label{6}$$
where m is neutrino mass.
Photon distribution in reaction (\[3\]) in laboratory system is given by $$\frac{d\sigma }{d(\omega /E)}=8\pi F^{2}\frac{m^{3}E^{3}}{M_{s}^{8}}\frac{\omega }{E}(1-\frac{\omega }{E})[(1-\frac{\omega }{E})^{2}+(\frac{\omega }{E})^{2}] \label{7}$$
where $\omega >>m$ is photon energy. This distribution is shown in Fig.3. It follows from (\[7\]) that photons are produced in the reaction (\[3\]) mainly within the energy range $0.2E\lesssim \omega \lesssim 0.8E$ with an average energy $\approx E/2$.
Therefore existence of low scale gravity at TeV scale or above could lead to the direct production of photons with energy $\omega >10^{22}$ eV (at these energies the mean interaction length for pair production for photons in the radio background is $\approx 1-10Mpc$ [@15]). Such photons can be hardly produced in standard weak interaction processes because in last ones photons appear as a result of cascade processes significantly reducing photon energy in comparison with the initial neutrino energy. For example, as it was shown in [@11] final energy of photons produced due to cascade processes can be by 10-100 times less than the energy of the initial neutrino flux.
Of course photons with the energy $\sim 10^{23}$ eV could be produced in cascade processes induced by neutrinos of the energy $10^{24}-10^{25}$ eV but from the observations of cosmic rays we know that cosmic ray fluxes decrease with the energy as $E^{-3}$, and therefore the probability of such events is significantly suppressed.
Fluxes of ultra high energy cosmic rays at the Earth are very small $\Phi
\sim 0.03km^{-2}sr^{-1}yr^{-1}$. Until now only about 60 events were collected with energies above GZK cut-off. However in near future improved Fly’s Eye (7000 $km^{2}sr$) [@8] will allow to detect about 20 events/yr. It seems possible that such detector could collect rare ultra energetic photons ($\omega >10^{22}eV$). The detection of such events could be an indication that these ultra high energy photons were produced in $\nu
\overset{-}{\nu }$ annihilation in the galactic halo due to effects of low scale gravity at TeV scale.
Authors thank D.Fargion for interesting discussions on ultra high energy cosmic rays. One of us (RVK) is grateful to Physics Department of New York University for warm hospitality.
[99]{} N.Arkani-Hamed, S.Dimopoulos, G.Dvali, Phys.Lett.**B429**, 263,1998; I.Antoniadis, N.Arkani-Hamed, S.Dimopoulos, and G.Dvali, Phys.Lett.**B436**,257,1998; N.Arkani-Hamed, S.Dimopoulos, G.Dvali, hep-ph/9807344; N.Arkani-Hamed, S.Dimopoulos, J.March-Russel, hep-th/9809124.
S.Nussinov, R.Shrock, Phys.Rev.**D59**, 105002,1999.
K.Cheung, Phys.Rev.**D61**,105005,1999; K.Cheung, Wai-yee Keung, Phys.Rev.**D60**,112003,1999.
G.F.Giudice, R.Rattazzi, J.D.Wells, Nucl.Phys.**B554**,3,1999.
Tao Ham, J.Lykken, Ren-Jie Zhang, Phys.Rev.**D59**,105006,1999.
G.Dvali, A.Yu.Smirnov, hep-ph/9904211,1999; L.J.Hall, D.Smith, hep-ph/9904267,1999; V.Barger, T.Han, C.Kao, and R.-J.hang, hep-ph/9905474; P.Jain et al., hep-ph/0001031,2000.
D.J.Bird et al., Astrophys.J.,**441**,144,1995; S.Yoshida et al., Astropart.Phys.**3**,105,1995; B.N.Afanasiev et al., in Proceedings of the 24th International Cosmic Ray Conference, Rome,**2**,756,1995; M.A.Lawrence et al., J.Phys.**G17**,773,1991; N.Hayashida et al., Phys.Rev.Lett.**77**,1000,1996; M.Takeda et al., Phys.Rev.Lett.**81**,1163,1998.
J.W.Cronin, Rev.Mod.Phys.**71**,S165,1999.
K.Greisen, Phys.Rev.Lett.**16**,748,1966; G.T.Zatsepin, V.A.Kuzmin, Pisma Zh.Eksp.Teor.Fiz.**4**,114,1966.
J.W.Elbert, P.Sommers, Ap.J.**441**,151,1995; R.J.Protheroe, P.A.Johnston, Astropart.Phys.**4**,253,1996; N.Mohapatra, S.Nussinov, Phys.Rev.**D57**,1940,1998; C.Sigl, S.Lee, D.N.Schramm, and P.Coppi, Phys..Lett.**B392**,129,1997; V.Berezinsky, M.Kachelriess, A.Vilenkin, Phys.Rev.Lett.**79**,4302,1997; G.Burdman, F.Halzen, R.Gandhi, Phys.Lett.**B417,**107,1998; G.R.Farrar, T.Piran astro-ph/9906431,1999; V.Berezinsky, hep-ph/0001163,2000; A.V.Olinto, astro-ph/0002006,2000.
D.Fargion, B.Mele, A.Salis, astro-ph/9710029,1997.
T.J.Weiler, Astropart.Phys.**11**,303,1999.
T.J.Weiler, hep-ph/9910316,1999.
R.Gandhi et al., Astropart.Phys.,**5**,81,1996.
R.J.Protheroe, astro-ph/9812055,1998.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '**Abstract:** In this paper, we have studied the accretion of phantom energy on a (2+1)-dimensional stationary Banados-Teitelboim-Zanelli (BTZ) black hole. It has already been shown by Babichev et al that for the accretion of phantom energy onto a Schwarzschild black hole, the mass of black hole would decrease and the rate of change of mass would be dependent on the mass of the black hole. However, in the case of (2+1)-dimensional BTZ black hole, the mass evolution due to phantom accretion is independent of the mass of the black hole and is dependent only on the pressure and density of the phantom energy. We also study the generalized second law of thermodynamics at the event horizon and construct a condition that puts an lower bound on the pressure of the phantom energy.'
author:
- Mubasher Jamil
- 'M. Akbar'
title: Generalized second law of thermodynamics for a phantom energy accreting BTZ black hole
---
Introduction
============
It has been found by various astronomical and cosmological observations [@perl] that our universe is currently in the phase of accelerated expansion. In the framework of Einstein’s gravity, this accelerated expansion has been explained by the presence of a ‘cosmological constant’ bearing negative pressure which results in the stretching of the spacetime [@pad]. Many other theoretical models have been presented to explain the present accelerated expansion of the universe including based on homogeneous and time dependent scalar field like the quintessence [@essence], Chaplygin gas [@chaplygin] and phantom energy [@phantom], to name a few. The phantom energy is characterized by the equation of state $p=\omega\rho$, with $\omega<-1$. It possesses some weird properties: the cosmological parameters like energy density and scale factor become infinite in a finite time; all gravitationally bound objects lose mass with the accretion of phantom energy; the fabric of spacetime is torn apart at the big rip; and that it violates the standard relativistic energy conditions. The astrophysical data coming from the microwave background radiation categorically favors the phantom energy [@cald]. Motivated from the dark energy models, we model phantom energy by an ideal fluid with negative pressure.
The accretion of dark energy onto a black hole has been studied by many authors [@jamil] after the seminal work of Babichev et al [@babi] who have shown that the mass of the black hole will decrease with time when we consider the accretion of phantom energy. In the Einstein theory of gravity, the accretion of the phantom energy onto Schwarzschild black hole and evaporation of primordial black hole has been discussed [@babi; @jamil1]. It will be interesting to investigate the accretion dynamics in low and higher dimensional gravities. It is also important to investigate accretion dynamics in the extended theories of gravity.
In this paper we investigate the accretion of exotic phantom energy onto a static uncharged 3-dimensional BTZ black hole. We will show that the expression of the evolution of BTZ black hole mass is independent of its mass and dependents only on the energy density and pressure of the phantom energy. It is well-known that the horizon area of the black hole decreases with the accretion of phantom energy, hence it is essential to study the generalized second law of thermodynamics (GSL) in this case [@GSL]. We show that the validity of GSL in the present model yields an lower bound on the phantom energy pressure. We also demonstrate that the first law of thermodynamics holds in the present construction.
The plan of the paper is as follows: In second section we model the accretion of phantom energy onto three dimensional BTZ black hole. In third section, we study the GSL for BTZ black hole. Finally we conclude our results.
Model of accretion
==================
Consider the field equations for a (2+1)-dimensional spacetime with a negative cosmological constant $\Lambda$ $$G_{ab}+\Lambda g_{ab}=\pi T_{ab}, \ \ (a,b=0,1,2)$$ where $G_{ab}$ is the Einstein tensor in (2+1)-dimension while $T_{ab}$ is the stress energy tensor of the matter field. The units are chosen such that $c=1$ and $G_3=1/8$. Considering the stress-energy tensor to be vacuum, one can obtain the following spherically symmetric metric, a (2+1)-dimensional BTZ black hole [@ban] $$ds^2=-f(r)dt^2+\frac{dr^2}{f(r)}+r^2d\phi^2,$$ where $f(r)=-M+r^2/l^2$, $M$ is the dimensionless mass of the black hole and $l^2=-1/\Lambda$, is a positive constant. The coefficient $g_{00}$ is termed as the lapse function. The event horizon of the BTZ black hole is obtained by setting $f(r)=0$, which turns out, $r_e= l\sqrt{M}$. Also we have $\sqrt{|g|}=r$, where $g$ is the determinant of the metric. To analyze the accretion of phantom energy onto the BTZ black hole, we here employ the formalism from the work by Babichev et al [@babi]. The stress energy momentum tensor representing the phantom energy is the perfect fluid $$T^{ab}=(\rho+p)u^a u^b+pg^{ab},$$ Here $\rho$ and $p$ are the energy density and pressure of the phantom energy while $u^a=(u^0,u^1,0)$ is the velocity three vector of the fluid flow. Also $u^1=u$ is the radial velocity of the flow while the third component $u^2$ is zero due to spherical symmetry of the BTZ black hole. There are two important equations of motion in our model: one which controls the conservation of mass flux is $J^{a}_{;a}=0$, where $J^a$ is the current density and the other that controls the energy flux $T^{a}_{0;a}=0$, across the horizon. Since the black hole is stationary, the only component of stress energy tensor of interest is $T^{01}$. Thus the equation of energy conservation $T^{0a}_{;a}=0$ is $$ur(\rho+p)\sqrt{f(r)+u^2}=C_1,$$ where $C_1$ is an integration constant. Since the flow is inwards the black hole therefore $u<0$. Also the projection of the energy momentum conservation along the velocity three vector $u_a
T^{ab}_{;b}=0$ (the energy flux equation) is $$ur\exp\Big[\int\limits_{\rho_\infty}^{\rho_h}\frac{d\rho}{\rho+p}\Big]=-A_1.$$ Here $A_1$ is a constant and the associated minus sign is taken for convenience. Also $\rho_h$ and $\rho_\infty$ are the energy densities of phantom energy at the BTZ horizon and at infinity respectively. From Eqs. (4) and (5), we obtain $$(\rho+p)\sqrt{f(r)+u^2}\exp\Big[-\int\limits_{\rho_\infty}^{\rho_h}\frac{d\rho}{\rho+p}\Big]=C_2,$$ where $C_2=-C_1/A_1=\rho_\infty+p(\rho_\infty)$. The rate of change in the mass of black hole $\dot M=-2\pi r T^1_0$, is given by $$dM=2\pi A_1(\rho_\infty+p_\infty)dt.$$ Note that $\rho_\infty+p_\infty<0$ (violation of null energy condition) leads to decrease in the mass of the black hole. Moreover, the above expression is also independent of mass contrary to the Schwarzschild black hole and the Reissner-Nordström black hole [@jamil]. Further, the last equation is valid for any general $\rho$ and $p$ violating the null energy condition, thus we can write $$dM=2\pi A_1(\rho+p)dt.$$
Critical Accretion
==================
We are interested only in those solutions that pass through the critical point as these correspond to the material falling into the black hole with monotonically increasing speed. The falling fluid can exhibit variety of behaviors near the critical point of accretion, close to the compact object. The equation of mass flux or the continuity equation $J^a_{;a}=0$ is $$\rho u r=k_1.$$ Here $k_1$ is integration constant. From Eqs. (4) and (9), we have $$\Big(\frac{\rho+p}{\rho}\Big)^2\Big(f(r)+u^2
\Big)=\Big(\frac{C_1}{k_1}\Big)^2=C_3.$$ Taking differentials of (9) and (10) and after simplification, we obtain $$\frac{du}{u}\Big[-V^2 +\frac{u^2}{f(r)+u^2} \Big]+\frac{dr}{r}\Big[
-V^2+ \frac{r^2}{l^2\Big(f(r)+u^2\Big)} \Big]=0.$$ Here $$V^2\equiv\frac{d\text{ln}(\rho+p)}{d\text{ln}\rho}-1,$$ From (11) if one or the other bracket factor is zero, one gets a turnaround point corresponding double-valued solution in either $r$ or $u$. The only solution that passes through a critical point is feasible. The feasible solution will correspond to material falling into the object with monotonically increasing velocity. The critical point is obtained by taking the both bracketed factors in Eq. (11) to be zero. This will give us the critical points of accretion. We obtain $$\begin{aligned}
V_c^2&=&\frac{r_c^2}{(f(r_c)+u_c^2)l^2},\\
V_c^2&=&\frac{u_c^2}{f(r_c)+u_c^2}.\end{aligned}$$ Above the subscript $c$ refers to the critical quantity. On comparing Eqs. (13) and (14), we get $$u_c^2=\frac{r_c^2}{l^2},\ \ V_c^2=\frac{u_c^2}{-M+2u_c^2}.$$ Here $u_c$ is the critical speed of flow at the critical points which we determine below. For physically acceptable solution, we require $V_c^2>0$, hence we get the following restrictions on speeds and the location of the critical points $$u_c^2>\frac{M}{2},\ \ r_c^2>\frac{r_+^2}{2}.$$
Generalized second law of thermodynamics and BTZ black hole
============================================================
In this section we will discuss the thermodynamic of phantom energy accretion that crosses the event horizon of BTZ black hole. Let us first write the BTZ metric in the form $$ds^{2} = h_{mn}dx^{m}dx^{n}+r^{2}d\phi^{2}, \ \ \ m,n=0,1$$ where $h_{mn}= \text{diag}(-f(r), 1/f(r))$, is a 2-dimensional metric. From the condition of normalized velocities $u^{a}u_{a} =
-1$, one can obtain the relations $$u^{0} = f(r)^{-1}\sqrt{f(r) + u^{2}}, ~~~ u_{0} = -\sqrt{f(r) +
u^{2}}.$$ The components of stress energy tensor are $T^{00}=f(r)^{-1}[(\rho +
p)(\frac{f(r)+u^{2}}{f(r)})-p]$, and $T^{11}=(\rho + p)u^2+f(r)p$. These two components help us in calculating the work density which is defined by $W=-\frac{1}{2}T^{mn}h_{mn}$ [@cai]. In our case it comes out $$W=\frac{1}{2}(\rho - p).$$ The energy supply vector is defined by $$\Psi_{n}=T^{m}_{n}\partial_{m}r + W\partial_{n}r.$$ The components of the energy supply vector are $\Psi_{0}=
T^{1}_{0}=-u(\rho+p)\sqrt{f(r)+u^2}$, and $\Psi_{1}=
T^{1}_{1}+W=(\rho+p)\Big(\frac{1}{2}+\frac{u^2}{f(r)}\Big)$. The change of energy across the apparent horizon is determined through $-dE\equiv-A\Psi$, where $\Psi=\Psi_0dt+\Psi_1dr$. The energy crossing the event horizon of the BTZ black hole is given by $$dE=4\pi r_eu^2(\rho+p)dt.$$ Assuming $E=M$ and comparing Eqs. (8) and (21), we can determine the value of constant $A_1=2u^2l\sqrt{M}$.
The entropy of BTZ black hole is $$S_h=4\pi r_e.$$ It can be shown easily that the thermal quantities, change of phantom energy $dE$, horizon entropy $S_h$ and horizon temperature $T_h$ satisfy the first law $dE=T_hdS_h$, of thermodynamics. After differentiation of last equation w.r.t. $t$, and using Eq. (8), we have $$\dot{S}_h=8\pi^2l^2u^2(\rho+p).$$ Since all the parameters are positive in the above equation (23) except that $\rho+p<0$, it shows that the second law of thermodynamics is violated i.e. $\dot{S}_h<0$, as a result of accretion of phantom energy on a BTZ black hole.\
Now we proceed to the generalized second law of thermodynamics (GSL). It is defined by $$\dot{S}_{tot}=\dot{S}_h+\dot{S}_{ph}\geq0.$$ In other words, the sum of the rate of change of entropies of black hole horizon and phantom energy must be positive. We consider event horizon of the BTZ black hole as a boundary of thermal system and the total matter energy within the event horizon is the mass of the BTZ black hole. We also assume that the horizon temperature is in equilibrium with the temperature of the matter-energy enclosed by the event horizon, i.e. $T_h=T_{ph}=T$, where $T_{ph}$ is the temperature of the phantom energy. Similar assumptions for the temperatures $T_h$ and $T_{ph}$ has been studied in [@davies]. We know that the Einstein field equations satisfy first law of thermodynamics $T_hdS_h=pdA+dE$, at the event horizon [@azad]. We also assume that the matter-energy enclosed by the event horizon of BTZ black hole also satisfy the first law of thermodynamics given by $$T_{ph}dS_{ph}=pdA+dE.$$ Here the horizon temperature is given by $$T_h=\left.\frac{f^\prime(r)}{4\pi}\right\vert_{r=r_e}
=\frac{\sqrt{M}}{2\pi l}.$$ In this paper, we are assuming that $T_h=T_{ph}=T$. Therefore Eq. (24) gives $$T\dot{S}_{tot}=T(\dot{S}_h+\dot{S}_{ph})=4\pi
l^2u(\rho+p)(2\sqrt{M}+\pi lp).$$ From the above equation, note that $u<0$ and $\rho+p<0$ the GSL holds provided $2\sqrt{M}+\pi lp>0$ which implies $$p\geq-\frac{2\sqrt{M}}{\pi l}.$$ Since the pressure of the phantom energy is negative ($p<0$), therefore the GSL gives us the lower bound on the pressure of the phantom energy. $$-\frac{2\sqrt{M}}{\pi l}\leq p<0.$$ The GSL in the phantom energy accretion holds within the inequality (29). Otherwise GSL does not hold which forbid evaporation of BTZ black hole by the phantom accretion [@pavon]. In addition, it is not clear whether the GSL should be valid in presence of the phantom fluid not respecting the dominant energy condition [@pavon].
Conclusion
==========
In this paper, we have investigated the accretion of exotic phantom energy onto a BTZ black hole. The motivation behind this work is to study the accretion dynamics in low dimensional gravity. Our analysis has shown that evolution of mass of a BTZ black hole would be independent of its mass and will be dependent only on the energy density and pressure of the phantom energy in its vicinity. Due to spherical symmetry, the accretion process is simple since the phantom energy falls radially on the black hole. The accretion would be much more interesting when additional parameters like charge and angular momentum are also incorporated in the BTZ spacetime. Similarly, it would be of much interest to perform the above analysis in higher ($n+1$) dimensional black hole spacetimes.
We also discussed GSL in the BTZ black hole spacetime. We assumed that the event horizon of BTZ black hole acts as a boundary of the thermal system and the phantom energy crossing the event horizon will change the mass of the black hole. We assumed that the horizon temperature is in local equilibrium with the temperature of the matter energy at the event horizon. Under these constraints it is shown that the GSL holds provided the pressure of the phantom energy $p$ has an lower bound $p\geq-\frac{2\sqrt{M}}{\pi l}$, on the black hole parameters ($M$ and $l$).
### Acknowledgment {#acknowledgment .unnumbered}
We would like to thank NUST for providing us financial support to visit ICRANet, Pescara, Italy to present this paper at the Second Joint Italian-Pakistani Workshop on Relativistic Astrophysics. We would also thank anonymous referees for their useful comments on this work. Also MJ would thank Emmanuel N. Saridakis, H. Mohseni Sadjadi, Diego Pavon and Ahmad Sheykhi for enlightening discussions during this work.
[99]{} A.G. Riess et al, Astron. J. 116 (1998) 1009;\
S. Perlmutter et al, Astrophys. J. 517 (1999) 565;\
C.L. Bennett et al., Astrophys. J. Suppl. 148 (2003) 1;\
M. Tegmark et al, Phys. Rev. D 69 (2004) 103501;\
S.W. Allen, et al., Mon. Not. Roy. Astron. Soc. 353 (2004) 457. S. Weinberg, Rev. Mod. Phys. 61 (1999) 1;\
P.J.E. Peebles and B. Ratra, Rev. Mod. Phys. 75 (2003) 559;\
T. Padmanabhan, Phys. Rep. 380 (2003) 235;\
E.J. Copeland et al, Int. J. Mod. Phys. D 15 (2006) 1753. C. Wetterich, Nucl. Phys. B 302 (1988) 668;\
B. Ratra and P. J. Peebles, Phys. Rev. D 37 (1988) 3406;\
R.R. Caldwell et al, Phys. Rev. Lett. 80 (1998) 1582;\
P.F. Gonzalez-Diaz, Phys. Rev. D 62 (2000) 023513;\
Y. Fujii, Phys. Rev. D 62 (2000) 064004. M. Jamil and M.A. Rashid, Eur. Phys. J. C 60 (2009) 141;\
M. Jamil and M.A. Rashid, Eur. Phys. J. C 58 (2008) 111;\
R. Bean and O. Dore, Phys. Rev. D 68 (2003) 023515;\
A. Y. Kamenshchik et al, Phys. Lett. B 511 (2001) 265;\
V. Gorini et al, Phys. Rev. D 67 (2003) 063509;\
N. Bilic et al, Phys. Lett. B 535 (2002) 17;\
M. C. Bento et al, Phys. Rev. D 66 (2002) 043507;\
L. Amendola et al, J. Cosmo. Astropart. Phys. 0307 (2003) 005;\
M. Bouhmadi-Lopez and P.V. Moniz, Phys. Rev. D 71 (2005) 063521;\
M.R. Setare, Int. J. Mod. Phys. D 18 (2009) 419;\
M.R. Setare, Eur. Phys. J. C 52 (2007) 689;\
M.R. Setare, Phys. Lett. B 654 (2007) 1. S. M. Carroll et al, Phys. Rev. D 68 (2003) 023509;\
P. Singh et al, Phys. Rev. D 68 (2003) 023522;\
P. F. Gonzalez-Diaz, Phys. Rev. D 68 (2003) 021303;\
L.P. Chimento and R. Lazkoz, Phys. Rev. Lett. 91 (2003) 211301;\
V. Faraoni, Phys. Rev. D 68 (2003) 063508;\
S. Nojiri and S.D. Odintsov, Phys. Lett. B 571 (2003) 1;\
E.N. Saridakis et al, Class. Quant. Grav. 26 (2009) 165003. R. R. Caldwell et al, Phys. Rev. Lett. 91 (2003) 071301. M. Jamil et al, Eur. Phys. J. C 58 (2008) 325;\
P.F. Gonzalez-Diaz, Phys. Rev. D 70 (2004) 063530;\
R-G. Cai and A. Wang, Phys. Rev. D 73 (2006) 063005;\
J. A. Jiménez Madrid and P.F. Gonzalez-Diaz, Grav. Cosmol. 14 (2008) 213;\
X. Zhang, Eur. Phys. J. C 60 (2009) 661;\
D.C. Guariento et al, Gen. Relativ. Grav. 40 (2008) 1593. E. Babichev et al, Phys. Rev. Lett. 93 (2004) 021102;\
E. Babichev et al, arXiv:0806.0916v3 \[gr-qc\]. M. Jamil, Eur. Phys. J. C 62 (2009) 609. M. Jamil et al, Phys. Rev. D 81 (2010) 023007;\
M. Jamil et al, arXiv:1003.0876v1 \[hep-th\];\
H.M. Sadjadi and M. Jamil, arXiv:1002.3588v1 \[gr-qc\]. M. Banados et al, Phys. Rev. Lett. 69 (1992) 1849;\
M. Banados et al, Phys. Rev. D 48 (1993) 1506. R-G. Cai and S.P. Kim, JHEP 0502 (2005) 050. P.C.W. Davies, Class. Quant. Grav. 4 (1987) L225;\
H.M. Sadjadi, Phys. Rev. D 73 (2006) 063525;\
G. Izquierdo and D. Pavon, Phys. Lett. B633, 420(2006);\
M. Akbar, Int. J. Theor. Phys. 48 (2009) 2665;\
M. Akbar, Chin. Phys. Lett. 25 (2008) 4199. M. Akbar and A.A. Siddiqui, Phys. Lett. B 656 (2007) 217;\
M. Jamil and M. Akbar, arXiv:0911.2556 \[hep-th\].
D.C. Guariento et al, Gen. Relativ. Gravit. 40 (2008) 1593;\
G. Inqierdo and D. Pavon, Phys. Lett. B 633 (2006) 420
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Building on the locality conditions for first-order logic by Hanf and Gaifman, Barthelmann and Schwentick showed in 1999 that every first-order formula is equivalent to a formula of the shape $\exists x_1 \dotsc \exists x_k \forall y\,{\varphi}$ where quantification in ${\varphi}$ is relativised to elements of distance $\leq r$ from $y$. Such a formula will be called Barthelmann-Schwentick normal form ([<span style="font-variant:small-caps;">bsnf</span>]{}) in the following. However, although the proof is effective, it leads to a non-elementary blow-up of the [<span style="font-variant:small-caps;">bsnf</span>]{} in terms of the size of the original formula.
We show that, if equivalence on the class of all structures, or even only finite forests, is required, this non-elementary blow-up is indeed unavoidable. We then examine restricted classes of structures where more efficient algorithms are possible. In this direction, we show that on any class of structures of degree $\leq 2$, [<span style="font-variant:small-caps;">bsnf</span>]{} can be computed in 2-fold exponential time with respect to the size of the input formula. And for any class of structures of degree $\leq d$ for some $d\geq 3$, this is possible in 3-fold exponential time. For both cases, we provide matching lower bounds.
author:
- |
André Frochaux, Lucas Heimberg\
Humboldt-Universität zu Berlin,\
`{andre.frochaux,lucas.heimberg}@informatik.hu-berlin.de`
bibliography:
- 'bibliography.bib'
title: |
An Optimal Construction for the Barthelmann-Schwentick Normal Form on Classes of Structures of Bounded Degree\
[Preliminary Version]{}
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this letter, we investigate the anti-jamming defense problem in multi-user scenarios, where the coordination among users is taken into consideration. The Markov game framework is employed to model and analyze the anti-jamming defense problem, and a collaborative multi-agent anti-jamming algorithm (CMAA) is proposed to obtain the optimal anti-jamming strategy. In sweep jamming scenarios, on the one hand, the proposed CMAA can tackle the external malicious jamming. On the other hand, it can effectively cope with the mutual interference among users. Simulation results show that the proposed CMAA is superior to both sensing based method and independent Q-learning method, and has the highest normalized rate.'
author:
- 'Fuqiang Yao, Luliang Jia [^1] [^2]'
title: 'A Collaborative Multi-agent Reinforcement Learning Anti-jamming Algorithm in Wireless Networks'
---
Anti-jamming, multi-agent reinforcement learning, Q-learning, Markov game.
Introduction
============
Jamming attack is a serious threat in wireless networks, and various anti-jamming methods have been developed in recent years [@existwork1]-[@existworkadd]. Due to factors of the jammers’ activities, the quality of channels varies between “good" and “poor" dynamically. The Markov decision process (MDP) [@existwork8] is a suitable paradigm to model and analyze the anti-jamming defense problem. Unfortunately, it is difficult to obtain the state transition probability function in an adversarial environment. In these scenarios, reinforcement learning (RL) techniques are available, such as the classic Q-learning method [@existwork9]. Based on the Q-learning method, the anti-jamming decision-making problem in single-user scenarios were investigated in [@existwork10]-[@existwork12]. Then, the authors in [@existwork13]-[@existwork15] extended it to the multi-user scenarios, and they resorted to the Markov game framework [@existwork16], which is the extension of the Markov decision process and can characterize the relationship among multiple users. Moreover, the corresponding multi-user reinforcement learning anti-jamming algorithm was designed. However, each user employed a standard Q-learning method in [@existwork13]-[@existwork15], and the coordination among users was not considered.
In order to achieve better anti-jamming performance, the coordination among users is necessary. Through collaborative learning, on the one hand, it can tackle the external malicious jamming, and on the other hand, it can effectively cope with the mutual interference caused by competition among users. In this letter, a collaborative anti-jamming framework is formulated, in which the “coordination" and “competition" are simultaneously considered. To model and analyze the anti-jamming defense problem, the Markov game framework is adopted, and a collaborative multi-agent reinforcement learning anti-jamming algorithm is proposed. The main contributions of this letter are given as follows:
- Based on the Markov game, the anti-jamming defense problem is investigated in multi-user scenarios, and the coordination among users is considered.
- We develop a collaborative multi-agent reinforcement learning anti-jamming algorithm to obtain the optimal anti-jamming strategy.
System Model and Problem Formulation
====================================
System Model
------------
![System model.[]{data-label="Fig1"}](Fig1){width="3.5in"}
As illustrated in Fig. \[Fig1\], there are *N* users and one jammer in the considered model. The user set is denoted as ${\cal N} = \left\{ {1, \cdots ,N} \right\}$, and the available channel set is defined as ${\cal M} = \left\{ {1, \cdots ,M} \right\}$. The number of available channels is *M* ($N < M$). The coordination among users can be achieved through information exchange. It is noted that the jamming pattern is the sweep jamming, and one channel is jammed at each time slot. The jamming channel set is represented as ${\cal C} = \left\{ {1, \cdots ,C} \right\}$. In this letter, we assume that the available channel set is the same as the jamming channel set. If two or more users select the same channel, it will lead to the mutual interference. In order to realize the reliable transmission, it is necessary to simultaneously consider the external malicious jamming and mutual interference due to competition among users. In this letter, the mutual interference refers to the co-channel interference among users, and the strategy is the selection of available channels.
Problem formulation
-------------------
The anti-jamming defense problem can be formulated as a Markov game [@existwork16], which is the extension of the Markov decision process in multi-user scenarios. Mathematically, it can be expressed as ${\cal G} = \left\{ {{\cal S},{{\cal A}_1}, \cdots ,{{\cal A}_N},f,{r_1}, \cdots ,{r_N}} \right\}$, where ${\cal S}$ denotes the set of states, ${{\cal A}_n},n = 1, \cdots ,N$ is the set of the strategies, $f$ represents the state transition model, and ${r_n},n = 1, \cdots ,N$ is the reward. In this letter, referring to [@existwork10], [@existwork11], the state can be defined as $s = \{ {\bf{a}},{f_{jx}}\}$, where ${\bf{a}} = \left( {{a_1},{a_2}, \cdots ,{a_N}} \right)$ represents a joint action profile, and the set of the joint action profiles is ${\cal A}{\rm{ = }} \otimes {{\cal A}_n},n = 1, \cdots ,N$, where $\otimes$ represents the Cartesian product. Similar to [@existwork17], the global reward can be defined as:
$$\label{eq1}
R = \sum\nolimits_{n = 1}^N {{r_n}(s,{\bf{a}})} ,$$
where $s \in {\cal S}$ denotes a state. It is assumed that the jamming channel is denoted as ${f_{jx}}$, the selected channel of user *n* is represented as ${f_{n,x}}$, and the reward of user *n* at time slot *t* can be expressed as:
$${r_n}(s,{\bf{a}},t) = \left\{ \begin{array}{l}
1,\quad \text{if}\;{f_{n,x}} \ne {f_{jx}}\;\& {f_{n,x}} \ne {f_{m,x}}\left( {m \in {{\cal N} \mathord{\left/
{\vphantom {{\cal N} n}} \right.
\kern-\nulldelimiterspace} n}} \right),\\
0,\quad \text{otherwise}.
\end{array} \right.$$
Collaborative Multi-agent Anti-jamming Algorithm
================================================
In this letter, we consider the two characteristics “coordination" and “competition" among users simultaneously. A collaborative anti-jamming framework is shown in Fig. \[Fig2\]. In wireless network, the coordination has various meanings, such as relay and information exchange. Here, the coordination is realized by information exchange among users. Based on the coordination among users, the method of “decision-feedback-adjustment" is applied to obtain the optimal anti-jamming strategy.
![Illustration of the collaborative anti-jamming framework.[]{data-label="Fig2"}](Fig2){width="3.5in"}
To solve the formulated anti-jamming Markov game, a multi-agent Q-learning algorithm is proposed. Similar to [@existwork17], user *n* updates its Q values according to the following rules:
$$\label{eq3}
{Q_n}\left( {s,{\bf{a}}} \right) = (1 - \lambda ){Q_n}\left( {s,{\bf{a}}} \right) + \lambda \left[ {{r_n} + \gamma {V_n}(s')} \right],$$
$$\label{eq4}
{V_n}(s') = {Q_n}\left( {s',{{\bf{a}}^{\rm{*}}}} \right), \text{where}\;{{\bf{a}}^{\rm{*}}} \in \mathop {\arg \max }\limits_{{\bf{a}}'} \sum\limits_{n = 1}^N {{Q_n}\left( {s',{\bf{a}}'} \right)} ,$$
where $\lambda$ is the learning rate. It is noted that the multi-agent Q-learning algorithm in (3) is decentralized, and each user updates its Q values separately. However, for the problem in (4), it is necessary to solve a global coordination game, which has common payoff [@existwork17]:
$$\label{eq5}
\;Q\left( {s,{\bf{a}}} \right) = \sum\limits_{n = 1}^N {{Q_n}\left( {s,{\bf{a}}} \right)} .$$
------------------------------------------------------------------------
***Algorithm 1**: Collaborative Multi-agent Anti-jamming Algorithm (CMAA)*\
------------------------------------------------------------------------
**Initiate:** ${{\cal S}}$, ${Q_n},\;n \in N$.\
**Loop:** $t = 0, \cdots ,T$\
Each user observes its current state $s(t) = \{ {\bf{a}}(t),{f_{jx}}(t)\}$, and selects a channel according to the following rules:\
$\bullet$ User *n* randomly chooses a channel profile ${\bf{a}} \in {\cal A}$ with probability $\varepsilon$;\
$\bullet$ User *n* chooses a channel profile ${{\bf{a}}^{\rm{*}}} \in argmax\sum\nolimits_{n = 1}^N {{Q_n}\left( {s',{\bf{a}}'} \right)}$ with probability $1 - \varepsilon$.\
Each user measures its payoff ${r_n}(s,{\bf{a}})$.\
The state is transferred into $s(t + 1) = \{ {\bf{a}}(t + 1),{f_{jx}}(t + 1)\}$, and the Q values are updated according to the rules in (3).\
**End loop**\
------------------------------------------------------------------------
Each user broadcasts its current Q value to other users. The exploration rate $\varepsilon \in (0,1)$ is introduced to avoid falling into a local optimum. Users randomly select a joint action ${\bf{a}} \in {\cal A}$ with probability $\varepsilon$, and users select the joint action ${{\bf{a}}^{\rm{*}}} \in argmax\sum\nolimits_{n = 1}^N {{Q_n}\left( {s',{\bf{a}}'} \right)}$ with probability $1 - \varepsilon$. Based on the above analysis, a collaborative multi-agent anti-jamming algorithm (CMAA) is proposed, and its implementation procedure is shown in Algorithm 1.
Similar to [@existwork10], the wideband spectrum sensing is adopted to sense the jammer’s activities, and all Q values are updated simultaneously. A transmission slot structure diagram is presented in Fig. \[Fig3\]. At the end of current slot, each user obtains a reward, and updates its strategy according to the received reward.
![Illustration of the transmission slot structure.[]{data-label="Fig3"}](Fig3){width="3.5in"}
Numerical Results and Discussions
=================================
In this subsection, we present some simulation results. A system with two users and one jammer is considered, in which five channels are available. ${t_{Rx}}$, ${t_{WBSS}}$, ${t_{ACK}}$, and ${t_{Learning}}$ denote the transmission time, wideband sensing time, ACK time, and learning time, respectively. The jammer begins to jam the transmission at time slot $t = 0.2ms$. Referring to [@existwork11], the simulation parameters are given as: $\lambda = 0.8$, $\gamma = 0.6$, $\varepsilon = 0.2$, ${t_{Rx}} = 0.98ms$, ${t_{WBSS}}{\rm{ + }}{t_{ACK}}{\rm{ + }}{t_{Learning}} = 0.2ms$. Moreover, the dwelling time of the sweeping jammer on each channel is ${t_{dwell}} = 2.28ms$, the number of time slots for simulations is $K = 10000$, and the simulation time is $T = K * ({t_{Rx}} + {t_{WBSS}}{\rm{ + }}{t_{ACK}}{\rm{ + }}{t_{Learning}})$.
![Time frequency information at initial state.[]{data-label="Fig4"}](Fig4){width="3.5in"}
![Time-frequency information at convergent state.[]{data-label="Fig5"}](Fig5){width="3.5in"}
Fig. \[Fig4\] and Fig. \[Fig5\] respectively show the time-frequency information at the initial and convergent state. As indicated in Fig. \[Fig4\], at the initial stage, the users employ random actions, and the signals of users and jammer are overlapped. Moreover, the signals among users are also overlapped. Fig. \[Fig5\] shows the time-frequency information of the proposed CMAA at convergent stage. As can be seen from Fig. \[Fig5\], at convergent stage, the signals of users can avoid the signal of the jammer. Meanwhile, the signals among users can effectively cope with the mutual interference, and the actions of users are coordinated.
![Performance comparison of the normalized rate.[]{data-label="Fig6"}](Fig6){width="3.5in"}
To validate the proposed CMAA, we compare it with the following two methods:
- Sensing based method: In this method, the users cannot learn the actions of the jammer, and channels are selected based on the sensing results. Moreover, we resort to a coordination approach, as in [@existwork17], in which user *n* ($n > 1$) selects its channel $a_n^*$ until the previous users $1, \cdots ,n - 1$ broadcast their chosen channels in the ordering. Then, user *n* broadcasts its channel.
- Independent Q-earning [@existwork13]: Each user adopts a standard Q-learning method. The coordination among users is not considered, and other users are treated as part of its environment.
In this section, the normalized rate is introduced to validate the performance of the proposed CMAA, and it can be defined as $\rho = {{P{K_{succ}}} \mathord{\left/
{\vphantom {{P{K_{succ}}} {P{N_0}}}} \right.
\kern-\nulldelimiterspace} {P{N_0}}}$, where $P{K_{succ}}$ represents the number of packets for successful transmission, and $P{N_0}$ denotes the length of packet statistics, which means that the normalized rate $\rho$ is calculated after $P{N_0}$ packets are transmitted. In this simulation, we have $P{N_0} = 20$. Then, the following results are obtained by making 200 independent runs and then taking the mean.
Fig. \[Fig6\] shows the performance of the normalized rate, it can be seen that the proposed CMAA is superior to both sensing based method and independent Q-learning method. Moreover, the proposed CMAA has the highest normalized rate $\rho$. The reason is that the sensing based method cannot learn the actions of the jammer, and channels are chosen based on the sensing results. Meanwhile, the independent Q-learning method does not consider the coordination among users, and each user chooses its channel independently. For the proposed CMAA, it can not only learn the actions of the jammer, but also consider the coordination among users.
Conclusion
==========
In this letter, we consider the “coordination" and “competition" simultaneously, and the Markov game framework is employed to model and analyze the anti-jamming defense problem. Then, a collaborative multi-agent anti-jamming algorithm (CMAA) is proposed to obtain the optimal anti-jamming strategy. Through collaborative learning, it can cope with the external malicious jamming and the mutual interference caused by competition among users simultaneously. To validate the effectiveness of the proposed CMAA, simulation results are presented. Compared with the sensing based method and independent Q-learning method, the proposed CMAA has the highest normalized rate.
[1]{}
K. Grover, A. Lim, and Q. Yang,“Jamming and anti-jamming techniques in wireless networks: A survey," *Int. J. Ad Hoc and Ubiquitous Comput.*, vol. 17, no. 4, pp. 197-215, Dec. 2014.
L. Jia, Y. Xu, Y. Sun, S. Feng, and A. Anpalagan,“Stackelberg game approaches for anti-jamming defence in wireless networks," arXiv preprint arXiv: 1805.12308, 2018. (to be published in IEEE Wireless Communications)
D. Yang, G. Xue, J. Zhang, A. Richa, and X. Fang,“Coping with a smart jammer in wireless networks: A stackelberg game approach," *IEEE Trans. Wireless Commun.*, vol. 12, no. 8, pp. 4038-4047, Aug. 2013.
L. Xiao, T. Chen, J. Liu, and H. Dai,“Anti-jamming transmission stackelberg game with observation errors," *IEEE Commun. Lett.*, vol. 19, no. 6, pp. 949-952, Jun. 2015.
L. Jia, F. Yao, Y. Sun, Y. Niu, and Y. Zhu,“Bayesian Stackelberg game for anti-jamming with incomplete information," *IEEE Commun. Lett.*, vol. 20, no. 10, pp. 1991-1994, Oct. 2016.
L. Jia, F. Yao, Y. Sun, Y. Xu, S. Feng, and A. Anpalagan,“A hierarchical learning solution for anti-jamming Stackelberg game with discrete power strategies," *IEEE Wireless Commun. Lett.*, vol. 6, no. 6, pp. 818-821, Dec. 2017.
F. Yao, L. Jia, Y. Sun, Y. Xu, S. Feng, and Y. Zhu,“A hierarchical learning approach to anti-jamming channel selection strategies," *Wireless Netw.*, DOI: 10.1007/s11276-017-1551-9, to be published.
L. Jia, Y. Xu, Y. Sun, S. Feng, L. Yu, and A. Anpalagan,“A multi-domain anti-jamming defence scheme in heterogeneous wireless networks," *IEEE Access*, DOI: 10.1109/ACCESS.2018.2850879, to be published.
Q. Hu, and W. Yue, *Markov decision processes with their applications.* Springer US, 2007.
C. J. C. H. Watkins, and P. Dayan,“Q-learning," *Mach. Learn.*, vol. 8, pp. 279-292, 1992.
F. Slimeni, B. Scheers, Z. Chtourou, *et al.*,“Jamming mitigation in cognitive radio networks using a modified Q-learning algorithm," *in Proc. International Conference on Military Comunications and Information Systems (ICMCIS) 2015*, pp. 1-7.
F. Slimeni, Z. Chtourou, B. Scheers, *et al*.,“Cooperative Q-learning based channel selection for cognitive radio networks," *Wireless Netw.*, DOI:10.1007/s11276-018-1737-9, to be published.
S. Machuzak, and S. K. Jayaweera,“Reinforcement learning based anti-jamming with wideband autonomous cognitive radios," *in Proc. IEEE International Conference on Communications in China (ICCC) 2016*, pp. 1-5.
M. A. Aref, S. K. Jayaweera, and S. Machuzak,“Multi-agent reinforcement learning based cognitive anti-jamming," *in Proc. IEEE Wireless Communications and Networking Conference (WCNC) 2017*, pp. 1-6.
M. A. Aref, and S. K. Jayaweera,“A novel cognitive anti-jamming stochastic game," *in Proc. Cognitive Communications for Aerospace Applications Workshop (CCAA) 2017*, pp. 1-4.
M. A. Aref, and S. K. Jayaweera,“A cognitive anti-jamming and interference-avoidance stochastic game," *in Proc. IEEE International Conference on Cognitive Informatics and Cognitive Computing (ICCI\*CC) 2017*, pp. 520-527.
L. Busoniu, R. Babuska, and B. D. Schutter,“A comprehensive survey of multiagent reinforcement learning," *IEEE Trans. Systems, Man, and Cybernetics*, vol. 38, no. 2, pp. 156-172, Mar. 2008.
N. Vlassis, *A concise introduction to multiagent systems and distributed artificial intelligence.* Morgan and Claypool Publishers, 2007.
[^1]: F. Yao is with the Sixty-third/63rd Research Institute, National University of Defense Technology, Nanjing 210007, China (e-mail:[email protected]).
[^2]: L. Jia is with the College of Communications Engineering, Army Engineering University of PLA, Nanjing 210007, China. (e-mail: [email protected]).(Corresponding author: Luliang Jia.)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We exploit recent results on the stability and performance analysis of positive Markov jump linear systems (MJLS) for the design of interval observers for MJLS with and without delays. While the conditions for the $L_1$ performance are necessary and sufficient, those for the $L_\infty$ performance are only sufficient. All the conditions are stated as linear programs that can be solved very efficiently. Two examples are given for illustration.\
*Keywords.* Interval observation; Markov jump linear systems; Positive systems; Optimization
author:
- 'Corentin Briat[^1]'
title: 'A class of $L_1$-to-$L_1$ and $L_\infty$-to-$L_\infty$ interval observers for (delayed) Markov jump linear systems'
---
Introduction
============
Interval observers are a particular type of observers that aim at estimating upper and lower bounds on the state value at all times. They have been successfully designed for a wide variety of systems including systems with inputs [@Mazenc:11; @Briat:15g], linear systems [@Mazenc:12], delay systems [@Efimov:13c; @Briat:15g], LPV systems [@Efimov:13b; @Chebotarev:15], discrete-time systems [@Mazenc:13; @Briat:15g], impulsive systems [@Degue:16nolcos; @Briat:17ifacObs; @Briat:18_ImtImp] and switched systems [@Rabehi:17; @Ethabet:17; @Briat:18_ImtImp]. To the best of the author’s knowledge, no results have been obtained in the context of Markov jump linear systems albeit those systems are important for practical purposes. Those systems are a class of switched systems having the particularity that the switching rule is governed by a continuous-time Markov process with countable finite [@Costa:05; @Boukas:06] or infinite [@Todorov:08] state-space. The positive version of those systems have been studied in considered in [@Aitrami:09b; @Zhang:14; @Bolzern:15] whereas those subject to delays have been considered in [@Zhu:17b] where various necessary and sufficient conditions for their stability and performance analysis have been obtained. The importance of positive systems [@Farina:00] is that they are instrumental for solving the interval observation problem and that they benefit from very interesting theoretical properties, such as the existence of various necessary and sufficient conditions for their stability and performance characterizations; see e.g. [@Farina:00; @Briat:11h; @Rantzer:15b; @Ebihara:11; @Colombino:15].
The goal of this paper is to use state-of-the-art methods for the analysis of positive Markov jump linear systems (MJLS) for the design of interval observers for both MJLS with and without delays. We provide necessary and sufficient conditions for the design of a certain class of interval observers for Markov jump linear systems with delays. Interestingly, the observer can be designed in a way that minimizes the $L_1$-gain on the transfer from the disturbance to the estimation error. The obtained conditions can be checked using linear programming techniques that also allows for the consideration of structural constraints (bounds on the coefficients, zero pattern, etc) on the gains of the observers. Analogous conditions, albeit sufficient only, are provided in the context of the $L_\infty$-gain. Some examples are given for illustration.
*Outline:* The structure of the paper is as follows: in Section \[sec:preliminary\] preliminary definitions and results are given. Section \[sec:perf\] is devoted to the performance analysis of positive MJLS. Section \[sec:obs\] presents the main results of the paper on interval observation. Examples are given in Section \[sec:ex\].
*Notations:* The cone of positive and nonnegative vectors of dimension $n$ are denoted by $\mathbb{R}_{>0}^n$ and $\mathbb{R}_{\ge0}^n$, respectively. The notation $\operatorname*{col}(x_1,\ldots,x_n)$ denotes the column vector made by stacking the elements $x_1$ to $x_n$ on the top of each other. $\mathds{1}$ denotes the vector of ones.
Preliminaries {#sec:preliminary}
=============
Let us consider the following class of positive MJLS: $$\label{eq:mainsyst}
\begin{array}{rcl}
\dot{x}(t)&=&A_{r_t}x(t)+A_{h,r_t}x(t-h)+E_{r_t}w(t)\\
z(t)&=&C_{r_t}x(t)+C_{h,r_t}x(t-h)+F_{r_t}w(t)\\
x(t_0)&=&x_0
\end{array}$$ where $x,x_0\in\mathbb{R}_{\ge0}^n$, $u\in\mathbb{R}^{n_u}$, $w\in\mathbb{R}_{\ge0}^{n_w}$ and $z\in\mathbb{R}_{\ge0}^{n_z}$ are the state of the system, the initial condition, the control input, the exogenous input and the performance output, respectively. The disturbance signal $w$ can be either deterministic or stochastic (but independent of $x$ and $r$). This will be further explained when necessary. The stochastic switching signal $r_t\in\{1,\ldots,N\}$ is assumed to be governed by a continuous-time Markov process with discrete state-space. Let $P(\tau)$ defined as $[P(\tau)]_{ij}=p_{ij}(\tau):=\operatorname{ \mathbb{P}}[r_{s+\tau}=j|r_{s}=i]$. It is known that this matrix solves the forward Kolmogorov equation $$\label{eq:probability}
\dot{P}(\tau)= P(\tau)\Pi,\ P(0)=I_N$$ where the matrix $\Pi$ is Metzler and such that $\Pi\mathds{1}_N=0$.
The system is internally positive, i.e. for all $w(t)\ge0$, then we have that $x(t),z(t)\ge0$, if and only if the matrices $A_i$ are Metzler and the matrices $A_{h,i},E_i,C_i,C_{h,i}$ and $F_i$ are nonnegative for all $i=1,\ldots,N$.
We now define the moment system associated with that will play an important role in the rest of the paper:
Let $x_i(t):=\operatorname{ \mathbb{E}}[x(t)\mathds{1}_{r_t=i}]$, $z_i(t):=\operatorname{ \mathbb{E}}[z(t)\mathds{1}_{r_t=i}]$ and $w_i(t):=\operatorname{ \mathbb{E}}[w(t)\mathds{1}_{r_t=i}]$. Then, the moment system associated with is defined as $$\label{eq:congruent}
\begin{array}{lcl}
\dot{\bar x}(t)&=&\bar A \bar x(t)+\bar A_h \bar x(t-h)+\bar E \bar w(t)\\
\bar z(t)&=&\bar C \bar x(t)+\bar C_h \bar x(t-h)+\bar F \bar w(t)
\end{array}$$ where $\bar x:=\operatorname*{col}_i(x_i)\in\mathbb{R}^{Nn}_{\ge0}$, $\bar w:=\operatorname*{col}_i(z_i)\in\mathbb{R}^{n_w}_{\ge0}$, $\bar z:=\operatorname*{col}_i(z_i)\in\mathbb{R}^{Nn_z}_{\ge0}$ and $$\begin{array}{rclrcl}
\bar A&:=&\operatorname*{diag}_i(A_i)+\Pi^T\otimes I_n,&\bar A_h&:=&\operatorname*{diag}_i(A_{h,i})(P(h)^T\otimes I_n)\\
\bar C&:=&\operatorname*{diag}_i\{C_i\},& \bar C_h&:=&\operatorname*{diag}_i\{C_i\}(P(h)^T\otimes I_n)\\
\bar E&:=&\operatorname*{diag}_i\{E_i\},& \bar F&:=&\operatorname*{diag}_i\{F_i\}.
\end{array}$$ The transfer function of this system is given by $$\bar G(s):=(\bar C+\bar C_h)(sI-\bar A-\bar A_h)^{-1}\bar E+\bar F.$$
This reformulation is different from the one in [@Bolzern:15] where conditional moments are considered. The above formulation has the advantage that it does not depend on the value of the probability distribution of the Markov process when deterministic disturbances are considered.
Stochastic stability and performance of (delayed) positive Markov jump linear systems {#sec:perf}
=====================================================================================
Stochastic $L_1$ performance of delayed positive Markov jump linear systems
---------------------------------------------------------------------------
Let us first define the stochastic $L_1$-gain:
The $L_1$-gain of the system is defined as the smallest $\xi>0$ such that $$\int_0^\infty\mathds{1}_{q}^T \operatorname{ \mathbb{E}}\left[z(s)\right]ds\le\xi\int_0^\infty\mathds{1}_{p}^T\operatorname{ \mathbb{E}}\left[w(s)\right]ds$$ holds for all $w\in L_1$, $w\ge0$. When the input $w$ is deterministic, the expectation symbol can be removed in the right hand-side.
We then have the following result:
\[th:L1\] We assume here that the system is positive and that $w\in L_1$. Then, the following statements are equivalent:
(a) The system with is stochastically stable in the $L_1$-sense and the $L_1$-gain of the transfer $w\mapsto z$ is equal to $\gamma^*>0$.
(b) The $L_1$-gain $\gamma^*>0$ of the system is the optimal value of the linear program $$\gamma^*=\inf_{\gamma>0,\lambda_1>0,\ldots,\lambda_N>0}\gamma$$ such that $\lambda_i\in\mathbb{R}_{>0}^n$, $i=1,\ldots,N$, and $$\label{eq:L1cond1}
\begin{array}{rcl}
\hspace{-10mm}A_i^T\lambda_i+\sum_{j=1}^N\left(\pi_{ij}I_n+p_{ij}(h)A_{h,j}^T\right)\lambda_j\\
\qquad \quad+\sum_{j=1}^N\left(p_{ij}(h)C_{h,j}^T+C_i^T\right) \mathds{1}_{n_z}&<&0\\
E_i^T\lambda_i-\gamma \mathds{1}_{n_w}+F_i^T\mathds{1}_{n_z}&<&0
\end{array}$$ for all $i=1,\ldots,N$.
(c) The $L_1$-gain $\gamma^*>0$ of the system verifies the expression $ \gamma^*=||\bar G(0)||_1$. [$\vartriangle$]{}
Stochastic $L_\infty$ performance of positive Markov jump linear systems
------------------------------------------------------------------------
Let us consider now the computation of the $L_\infty$-gain which is given by:
The stochastic $L_\infty$-gain of the system is defined as the smallest $\xi>0$ such that $$\sup_{t\ge0}\{||\operatorname{ \mathbb{E}}[z(t)]||_\infty\}]\le\xi\sup_{t\ge0}\{||\operatorname{ \mathbb{E}}[w(t)]||_\infty\}$$ holds for all $w\in L_\infty$, $w\ge0$.
We then have the following result:
\[th:Linf\] Assume that the non-delayed version of the system (i.e. $A_{h,i}=0$ for all $i=1,\ldots,N$)) is positive and that $w\in L_\infty$ is a stochastic signal that is independent of $(x,r)$. Assume further that one of the following equivalent statements hold:
(a) The $L_\infty$-gain of the moment system is equal to $\gamma^*$.
(b) The $L_\infty$-gain $\gamma^*>0$ of the moment system is the optimal value of the linear program $$\gamma^*=\inf_{\gamma>0,\lambda_1>0,\ldots,\lambda_N>0}\gamma$$ such that $\lambda_i\in\mathbb{R}_{>0}^n$, $i=1,\ldots,N$, and $$\begin{array}{rcl}
A_i\lambda_i+\sum_{j=1}^N\pi_{ji}\lambda_j+E_i \mathds{1}_{n_w}&<&0\\
C_i\lambda_i-\gamma \mathds{1}_{n_z}+F_i\mathds{1}_{n_w}&<&0
\end{array}$$ hold for all $i=1,\ldots,N$.
(c) The $L_\infty$-gain $\gamma^*>0$ of the moment system verifies the expression $\gamma^*=||\bar G(0)||_\infty$.
Then, the $L_\infty$-gain of the system with $A_{h,i}=0$, $i=1,\ldots,N$, is at most $\gamma$.
This result is not tight in the sense that we only compute an upper-bound on the $L_\infty$-gain of the system .
Design of interval observers {#sec:obs}
============================
Let us consider now the following system $$\label{eq:systy}
\begin{array}{rcl}
\dot{x}(t)&=&A_{r_t}x(t)+A_{h,r_t}x^+(t-h)+E_{r_t}w(t),\ x(0)=x_0\\
y(t)&=&C_{r_t} x(t)+C_{h,r_t}x^+(t-h)+F_{r_t} w(t)
\end{array}$$ where $x,x_0\in\mathbb{R}^n$, $w\in\mathbb{R}^p$, $y\in\mathbb{R}^r$ are the state of the system, the initial condition, the persistent disturbance input and the measured output. Note that this system is not necessarily positive. We are interested in finding an interval-observer of the form $$\label{eq:obs}
\begin{array}{lcl}
\dot{x}^\bullet(t)&=&A_{r_t}x^\bullet(t)+A_{h,r_t}x^\bullet(t-h)+E_{r_t}w^\bullet(t)\\
&&\quad+L_{r_t}(y(t)-y^\bullet(t))\\
y^\bullet(t)&=&C_{r_t} x^\bullet(t)-C_{h,r_t}x^\bullet(t-h)-F_{r_t} w^\bullet(t)\\
x^\bullet(0)&=&x_0^\bullet
\end{array}$$ where $\bullet\in\{-,+\}$. Above, the observer with the superscript “$+$” is meant to estimate an upper-bound on the state value whereas the observer with the superscript “-” is meant to estimate a lower-bound, i.e. $x^-(t)\le x(t)\le x^+(t)$ for all $t\ge0$ provided that $x_0^-\le x_0\le x_0^+$. The signals $w^-,w^+\in L_\infty(\mathbb{R}_{\ge0},\mathbb{R}^p)$ are the lower- and the upper-bound on the disturbance $w(t)$ at any time, i.e. $w^-(t)\le w(t)\le w^+(t)$ for all $t\ge0$. We then accordingly define the following errors $e^+(t):=x^+(t)-x(t)$ and $e^-(t):=x(t)-x^-(t)$ that are described by the model $$\label{eq:error}
\begin{array}{rcl}
\dot{e}^\bullet(t)&=&(A_{r_t}-L^\bullet_{r_t}C_{r_t} )e^\bullet(t)\\
&&+(A_{h,r_t}-L_{r_t}C_{h,r_t} )e^\bullet(t-h)\\
&&+(E_{r_t}-L_{r_t}F_{r_t} )\delta^\bullet(t)\\
\zeta^\bullet(t)&=&M_{r_t}e^\bullet
\end{array}$$ where $\bullet\in\{-,+\}$, $\delta^+(t):=w^+(t)-w(t)\in\mathbb{R}_{\ge0}^p$ and $\delta^-(t):=w(t)-w^-(t)\in\mathbb{R}_{\ge0}^p$. The matrix $M_{r_t}\in\mathbb{R}^{q\times n}_{\ge0}$ is a nonzero matrix driving the errors $e^\bullet$ to the observed outputs $\zeta^\bullet$. It is assumed to be chosen a priori.
A class of $L_1$-to-$L_1$ interval observers
--------------------------------------------
With all the previous elements in mind, we can state the observation problem that is considered in this section:
\[problemL1\] Find an interval observer of the form such that
(a) The linear systems in are positive, i.e. $A_i-L_iC_i$ is Metzler and $E_i-L_i F_i $ and $A_{h,i}-L_i C_{h,i}$ are nonnegative for all $i=1,\ldots,N$;
(b) The linear systems in are stochastically stable in the $L_1$-sense;
(c) The $L_1$-gain of the transfers $\delta^\bullet\to\zeta^\bullet$, $\bullet\in\{-,+\}$, are minimum.
We then have the following result that provides a necessary and sufficient conditions for the existence of a solution to Problem \[problemL1\]:
\[th:L1obs\] The following statements are equivalent:
(a) There exists an optimal $L_1$-to-$L_1$ interval-observer of the form for the system that solves Problem \[problemL1\].
(b) There exist diagonal matrices $X_i\in\mathbb{R}^{n\times n}$, matrices $U_i\in\mathbb{R}^{n\times r_i}$, $i=1,\ldots,N$, and scalars $\gamma,\alpha$ such that the linear optimization problem [ $$\min_{X_1,\ldots,X_N,U_1,\ldots,U_N,\alpha,\gamma} \gamma$$ subject to the constraints $\alpha,\gamma>0$, $\bar X\mathds{1}>0$, $$\label{eq:L1:cond1}
\begin{array}{ccc}
\bar X\bar A-\bar U\bar C+\alpha\ge0, \bar X\bar A_h-\bar U\bar C_h\ge0,\bar X\bar E-\bar U\bar F\ge0
\end{array}$$ and $$\label{eq:L1:cond2}
\begin{bmatrix}
\mathds{1}\\
1
\end{bmatrix}^T\begin{bmatrix}
(\bar X\bar A-\bar U\bar C)+(\bar X\bar A_h-\bar U\bar C_h) && \bar X\bar E-\bar U\bar F\\
\mathds{1}^T(I_N\otimes M) && -\gamma \mathds{1}^T
\end{bmatrix}<0$$]{} where $\bar X:=\operatorname*{diag}_i(X_I)$ and $\bar U:=\operatorname*{diag}_i(U_I)$ is feasible.
Moreover, in such a case, if we define $(\bar X,\bar U^*,\alpha^*,\gamma^*)$ as the global minimizer of the above minimization problem, then the optimal gains $L_i^{*}$ are given by $L_i^{*}=(X_i^*)^{-1}U_i^*$.
The proof follows from algebraic manipulations using the change of variables $\lambda_i=X_i\mathds{1}$.
A class of $L_\infty$-to-$L_\infty$ interval observers
------------------------------------------------------
The following observation problem will be considered in this section:
\[problemLinf\] Find an interval observer of the form such that
(a) The linear systems in are positive, i.e. $A_i-L_iC_i$ is Metzler and $E_i-L_iF_i $ is nonnegative for all $i=1,\ldots,N$;
(b) The linear systems in are stochastically stable in the $L_\infty$-sense;
(c) The $L_\infty$-gain of the transfers $\delta^\bullet\to\zeta^\bullet$, $\bullet\in\{-,+\}$, is smaller than a certain level $\gamma$ that can be minimized.
\[th:Linfobs\] The following statements are equivalent:
(a) There exists a $L_\infty$-to-$L_\infty$ interval-observer of the form for the system that solves Problem \[problemLinf\].
(b) There exist diagonal matrices $X_i\in\mathbb{R}^{n\times n}$, matrices $U_i\in\mathbb{R}^{n\times r_i}$, $i=1,\ldots,N$, and scalars $\gamma,\alpha$ such that the linear optimization problem [ $$\min_{X_1,\ldots,X_N,U_1,\ldots,U_N,\alpha,\gamma} \gamma$$ subject to the constraints $\alpha,\gamma>0$, $\bar X\mathds{1}>0$, $$\label{eq:L1:cond1}
\begin{array}{ccc}
\bar X\bar A-\bar U\bar C+\alpha\ge0, \bar X\bar E-\bar U\bar F\ge0
\end{array}$$ and $$\label{eq:L1:cond2}
\begin{bmatrix}
\mathds{1}\\
1
\end{bmatrix}^T\begin{bmatrix}
(\bar X\bar A-\bar U\bar C) && (\bar X\bar E-\bar U\bar F)\mathds{1}\\
\mathds{1}^T && -\gamma \mathds{1}^T
\end{bmatrix}<0$$]{} where $\bar X:=\operatorname*{diag}_i(X_I)$ and $\bar U:=\operatorname*{diag}_i(U_I)$ is feasible.
Moreover, in such a case, if we define $(\bar X,\bar U^*,\alpha^*,\gamma^*)$ as the global minimizer of the above minimization problem, then the optimal gains $L_i^*$ are given by $L_i^*=(X_i^*)^{-1}U_i^*$ and is uniformly optimal over all the possible values for $M$ (i.e. it is independent of the values of $M$).
The proof follows from the same procedure as in [@Briat:15g]. As it is quite long, it is omitted here.
Examples {#sec:ex}
========
A system without delay
----------------------
Let us consider the system with the matrices $$\label{eq:ex1}
\begin{array}{r}
A_1=\begin{bmatrix}
-1 & 0\\1 0.1
\end{bmatrix}, A_2=\begin{bmatrix}
0.1 & 1\\0 && -2
\end{bmatrix}, E_1=\begin{bmatrix}
1 & 0.5\\0& 0.5
\end{bmatrix},\\
E_2=\begin{bmatrix}
0&0.5\\1&0.5
\end{bmatrix},
C_1=C_2=\begin{bmatrix}
1 & 1
\end{bmatrix},F_1=F_2=\begin{bmatrix}
0 & 0
\end{bmatrix}.
\end{array}$$ We also pick $A_{h,1}=A_{h,2}=0$, $C_{h,1}=C_{h,2}=0$ and $$\Pi=\begin{bmatrix}
-2&2\\2&-2
\end{bmatrix}.$$ Solving for the conditions of Theorem \[th:L1obs\], we get the gains $$\label{eq:Lex1}
L_{1} = \begin{bmatrix}
0\\1
\end{bmatrix}\textnormal{ and }L_{2} = \begin{bmatrix}
1\\0
\end{bmatrix}.$$ The inputs are given by $w_1(t)=\sin(t)$, $w_2(t)=\sin(t+\pi/2)$, $w_1^+(t)=w_2^+(t)=1$, $w_1^-(t)=w_2^-(t)=-1$. The $L_1$-gain of the transfer $\delta^\bullet\mapsto \zeta^\bullet$ is equal to 1.0426. Solving now for the conditions in Theorem \[th:Linfobs\], we obtain the same gains together with the value 1.1383 as an upper-bound on the $L_\infty$-gain of the transfer $\delta^\bullet\mapsto \zeta^\bullet$. The trajectories of the system and the observers are depicted in Fig. \[fig:L1nondel\].
![Trajectories of the system - and the observer with the gains .[]{data-label="fig:L1nondel"}](./Matlab/L1_observer_nondelayed.pdf){width="70.00000%"}
A system with delay
-------------------
Let us consider the system with the matrices $$\label{eq:ex2}
\begin{array}{r}
A_1 = \begin{bmatrix}
-7.364 &1.065 &1.255\\
1.809 &-9.3& 0\\
0.555 &0 &-7.086
\end{bmatrix}, A_{h,1} = \begin{bmatrix}
1.5&3&1.5\\
2.7 &3 &6.45\\
0 &1.5 &3
\end{bmatrix},\\
A_2 = \begin{bmatrix}
-7.469 &1.126 &1.3\\
1.851 & -9.222 &0\\
0.6180 &0& -7.171\\
\end{bmatrix}, A_{h,2} = \begin{bmatrix}
1.8 &3.44 &2.25\\
3 &3.45 &6.75\\
0 &0 &3.6
\end{bmatrix},\\
E_1 = \begin{bmatrix}
1& 0\\
0& 1\\
0& 1
\end{bmatrix}, E_{2} = \begin{bmatrix}
1& 0\\
1& 1\\
1& 0
\end{bmatrix},C_1=C_2=\begin{bmatrix}
1 & 0 & 0\\
0 & 1 & 0
\end{bmatrix}
\end{array}$$ together with $C_{h,1}=C_{h,2}=0$, $F_1=F_2=0$ and $$\Pi = \begin{bmatrix}
-1.5 & 1.5\\
0.3 & -0.3
\end{bmatrix}.$$ We consider the same inputs as in the other example. We now use Theorem \[th:L1obs\] to which we add the constraint that the coefficients of the observer gains must not exceed $20$ in absolute value. We get the gains $$\label{eq:Lex2}
L_{1} = \begin{bmatrix}
20& 1.0650\\
1.8090& 20\\
0.5550 & 0
\end{bmatrix}\textnormal{ and }L_{2} = \begin{bmatrix}
1.8& 1.126\\
1.851& 3.45\\
0 & 0
\end{bmatrix}$$ which yields the minimum $L_1$-gain 0.88892. The trajectories of the system and the observers are depicted in Fig. \[fig:L1del\].
![Trajectories of the system - and the observer with the gains .[]{data-label="fig:L1del"}](./Matlab/L1_observer_delayed.pdf){width="70.00000%"}
[10]{}
M. [Ait Rami]{} and J. Shamma. Hybrid positive systems subject to markovian switching. In [*Analysis and Design of Hybrid Systems (ADHS’09)*]{}, pages 138–143, Zaragoza, Spain, 2009.
P. Bolzern and P. Colaneri. Positive markov jump linear systems. , 2(3-4):275–427, 2015.
E. K. Boukas. . Birkh[ä]{}user, Boston, USA, 2006.
C. Briat. Robust stability and stabilization of uncertain linear positive systems via integral linear constraints - ${L_1}$- and ${L_\infty}$-gains characterizations. , 23(17):1932–1954, 2013.
C. Briat. ${L}_1/\ell_1$-to-${L}_1/\ell_1$ analysis of linear positive impulsive systems with application to the ${L}_1/\ell_1$-to-${L}_1/\ell_1$ interval observation of linear impulsive and switched systems. , 2018.
C. Briat and M. Khammash. Interval peak-to-peak observers for continuous- and discrete-time systems with persistent inputs and delays. , 74:206–213, 2016.
C. Briat and M. Khammash. Simple interval observers for linear impulsive systems with applications to sampled-data and switched systems. In [*20th IFAC World Congress*]{}, pages 5235–5240, Toulouse, France, 2017.
S. Chebotarev, D. Efimov, T. Ra[ï]{}ssi, and A. Zolghadri. Interval observers for continuous-time [LPV]{} systems with ${L}_1$/${L}_2$ performance. , 58:82–89, 2015.
M. Colombino and R. S. Smith. A convex characterization of robust stability for positive and positively dominated linear systems. , 61(7):1965–1971, 2016.
O. L. V. Costa, M. D. Fragoso, and R. P. Marques. . Springer-Verlag, London, UK, 2005.
K. H. Degue, D. Efimov, and J.-P. Richard. Interval observers for linear impulsive systems. In [*10th IFAC Symposium on Nonlinear Control Systems*]{}, 2016.
Y. Ebihara, D. Peaucelle, and D. Arzelier. ${L_1}$ gain analysis of linear positive systems and its applications. In [*50th Conference on Decision and Control, Orlando, Florida, USA*]{}, pages 4029–4034, 2011.
D. Efimov, W. Perruquetti, and J.-P. Richard. On reduced-order interval observers for time-delay systems. In [*12th European Control Conference*]{}, pages 2116–2121, Z[ü]{}rich, Switzerland, 2013.
D. Efimov, T. Ra[ï]{}ssi, and A. Zolghadri. Control of nonlinear and [LPV]{} systems: Interval observer-based framework. , 58(3):773–778, 2013.
H. Ethabet, T. Raissi, M. Amairi, and M. Aoun. Interval observers design for continuous-time linear switched systems. In [*20th IFAC World Congress*]{}, pages 6259–6264, Toulouse, France, 2017.
L. Farina and S. Rinaldi. . John Wiley & Sons, 2000.
F. Mazenc and O. Bernard. Interval observers for linear time-invariant systems with disturbances. , 47:140–147, 2011.
F. Mazenc, T. N. Dinh, and S.-I. Niculescu. Robust interval observers and stabilization design for discrete-time systems with input and output. , 49:3490–3497, 2013.
F. Mazenc, M. Kieffer, and E. Walter. Interval observers for continuous-time linear systems. In [*American Control Conference*]{}, pages 1–6, Montr[é]{}al, Canada, 2012.
D. Rabehi, D. Efimov, and J.-P. Richard. Interval estimation for linear switched systems. In [*20th IFAC World Congress*]{}, pages 6265–6270, Toulouse, France, 2017.
A. Rantzer. Scalable control of positive systems. , 24:72–80, 2015.
M. G. Todorov and M. D. Fragoso. Output feedback $\mathcal{H}_\infty$ control of continuous-time infinite markovian jump linear systems via [LMI]{} methods. , 950-974, 2008.
J. Zhang, Z. Han, and F. Zhu. Stochastic stability and stabilization of positive systems with markovian jump parameters (in press). , 2014.
S. Zhu, Q.-L. Han, and C. Zhang. ${L}_1$-stochastic stability and ${L}_1$-gain performance of positive markov jump linear systems with time-delays: Necessary and sufficient conditions. , 62(7):3634–3639, 2017.
[^1]: [email protected]; http://www.briat.info
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Existing face restoration researches typically relies on either the degradation prior or explicit guidance labels for training, which often results in limited generalization ability over real-world images with heterogeneous degradations and rich background contents. In this paper, we investigate the more challenging and practical “dual-blind” version of the problem by lifting the requirements on both types of prior, termed as “Face Renovation”(FR). Specifically, we formulated FR as a semantic-guided generation problem and tackle it with a collaborative suppression and replenishment (CSR) approach. This leads to HiFaceGAN, a multi-stage framework containing several nested CSR units that progressively replenish facial details based on the hierarchical semantic guidance extracted from the front-end content-adaptive suppression modules. Extensive experiments on both synthetic and real face images have verified the superior performance of HiFaceGAN over a wide range of challenging restoration subtasks, demonstrating its versatility, robustness and generalization ability towards real-world face processing applications.'
author:
- '$^{1,2}$Lingbo Yang, $^3$Chang Liu, $^2$Pan Wang, $^1$Shanshe Wang, $^2$Peiran Ren, $^1$Siwei Ma, $^1$Wen Gao'
bibliography:
- 'sample-base.bib'
title: |
HiFaceGAN: Face Renovation via Collaborative\
Suppression and Replenishment
---
Ł
<ccs2012> <concept> <concept\_id>10010147.10010178.10010224</concept\_id> <concept\_desc>Computing methodologies Computer vision</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
{width="\textwidth"}
Introduction
============
Face photographs record long-lasting precious memories of individuals and historical moments of human civilization. Yet the limited conditions in the acquisition, storage, and transmission of images inevitably involve complex, heterogeneous degradations in real-world scenarios, including discrete sampling, additive noise, lossy compression, and beyond. With great application and research value, face restoration has been widely concerned by industry and academia, with an plethora of works [@ir_review][@deblurring_survey][@com_art_reduction_survey] devoted to address specific types of image degradation. Yet under more generalized, unconstrained application scenarios, few existing works can report satisfactory restoration results.
For face restoration, most existing methods typically work in a “non-blind” fashion with specific degradation of prescribed type and intensity, leading to a variety of sub-tasks including super resolution[@sr_review][@srcnn][@edsr][@esrgan], hallucination [@fh_review][@face_hallu_4], denoising[@RIDNet][@VDNet], deblurring [@deblurring_survey][@DeblurGAN][@DeblurGANv2] and compression artifact removal [@com_art_reduction_survey][@ARCNN][@EPGAN]. However, task-specific methods typically exhibit poor generalization over real-world images with complex and heterogeneous degradations. A case in point shown in Fig. \[fig:teaser\] is a historic group photograph taken at the Solvay Conference, 1927, that super-resolution methods, ESRGAN [@esrgan] and Super-FAN [@SuperFAN], tend to introduce additional artifacts, while other three task-specific restoration methods barely make any difference in suppressing degradation artifacts or replenishing fine details of hair textures, wrinkles, etc., revealing the impracticality of task-specific restoration methods.
When it comes to blind image restoration [@BIR_original], researchers aim to recover high-quality images from their degraded observation in a “single-blind” manner without *a priori* knowledge about the type and intensity of the degradation. It is often challenging to reconstruct image contents from artifacts without degradation prior, necessitating additional guidance information such as categorial [@CategorySpecificID] or structural prior [@FSRNet] to facilitate the replenishment of faithful and photo-realistic details. For blind face restoration [@MultiScaleFR-AAAI2018][@SGEN], facial landmarks [@SuperFAN], parsing maps [@pix2pixHD], and component heatmaps [@faceSR_ECCV2018] are typically utilized as external guidance labels. In particular, Li *et.al.* explored the guided face restoration problem [@BlindFR-ECCV2018][@EnhancedBlindFR-CVPR2020], where an additional high-quality face is utilized to promote fine-grained detail replenishment. However, it often leads to limited feasibility for restoring photographs without ground truth annotations. Furthermore, for real-world images with complex background, introducing unnecessary guidance could lead to inconsistency between the quality of renovated faces and unattended background contents.
In this paper, we formally propose “Face Renovation”(FR), an extra challenging, yet more practical task for photorealistic face restoration in a “dual-blind” condition, lifting the requirements of both the degradation and structural prior for training. Specifically, we formulate FR as a semantic-guided face synthesis problem, and propose to tackle this problem with a collaborative suppression and replenishment(CSR) framework. To implement FR, we propose HiFaceGAN, a generative framework with several nested CSR units to implement FR in a multi-stage fashion with hierarchical semantic guidance. Each CSR unit contains a suppression module for extracting layered semantic features with content-adaptive convolution, which are utilized to guide the replenishment of corresponding semantic contents. Extensive experiments are conducted on both the synthetic FFHQ [@stylegan_ffhq] and real-world photographs against competitive degradation-specific baselines, highlighting the challenges in proposed face renovation and the superiority of our proposed HiFaceGAN. In summary, our contributions are threefold:
- We present a challenging, yet practical task, termed as “Face Renovation (FR)” to tackle unconstrained face restoration problem in a “dual-blind” fashion, lifting the requirements on the both degradation and structural prior.
- We propose a collaborative suppression and replenishment (CSR) framework “HiFaceGAN” with a nested architecture for multi-stage face renovation with hierarchical semantic guidance. Specifically, the extracted semantic hierarchy, the working mechanism of HiFaceGAN, and its advantages over existing restoration methods are thoroughly explained with illustrative examples.
- Extensive experiments are conducted on both synthetic and real face images with significant performance gain over a variety of “non-blind” and “single-blind” baselines, verifying the versatility, robustness and generalization capability of our proposed HiFaceGAN.
Related Works
=============
Non-Blind Face Restoration
--------------------------
Image restoration consists of a variety of subtasks, such as denoising [@RIDNet][@VDNet], deblurring [@DeblurGAN][@DeblurGANv2] and compression artifact removal [@ARCNN][@EPGAN]. In particular, image super resolution [@srcnn][@edsr][@srgan][@esrgan] and its counterpart for faces, hallucination [@fh_review][@face_hallu_1][@face_hallu_3][@face_hallu_4], can be considered as a specific type of restoration against downsampling. However, existing works often works in a “non-blind” fashion by prescribing the degradation type and intensity during training, leading to dubious generalization ability over real images with complex, heterogeneous degradations. In this paper, we perform face renovation by replenishing facial details based on hierarchical semantic guidance that are more robust against mixed degradations, and achieves superior performance over a wide range of restoration subtasks against state-of-the-art “non-blind” baselines.
Blind Face Restoration
----------------------
Blind image restoration [@BIR_original] [@BIR_1][@BIR_2] aims to directly learn the restoration mapping based on observed samples. However, most existing methods for general natural images are still sensitive to the degradation profile [@BlindIR-wo-prior] and exhibit poor generalization over unconstrained testing conditions. For category-specific [@CategorySpecificID] (face) restoration, it is commonly believed that incorporating external guidance on facial prior would boost the restoration performance, such as semantic prior [@ComponentSP], identity prior [@face_hallu_1], facial landmarks [@SuperFAN][@FSRNet] or component heatmaps [@faceSR_ECCV2018]. In particular, Li *et.al.* [@BlindFR-ECCV2018] explored the guided face restoration scenario with an additional high-quality guidance image to help with the generation of facial details. Other works utilize objectives related to subsequent vision tasks to guide the restoration, such as semantic segmentation [@whenID_meets_high] and recognition [@Joint_BIR_recog]. In this paper, we further explore the “dual-blind” case targeting at unconstrained face renovation in real-world applications. Particularly, we reveal an astonishing fact that with collaborative suppression and replenishment, the dual-blind face renovation network can even outperform state-of-the-art “single-blind” methods due to the increased capability for enhancing non-facial contents, which provides new insights for tackling unconstrained face restoration problem from a generative view.
Deep Generative Models for Face Images
--------------------------------------
Deep generative models, especially GANs [@gan] have greatly facilitated conditional image generation tasks [@pix2pix][@cyclegan], especially for high-resolution faces [@progressivegan][@stylegan_ffhq][@stylegan2]. Existing methods can be roughly summarized into two categories: semantic-guided methods, utilizing parsing maps [@pix2pixHD], edges [@vid2vid], facial landmarks [@SuperFAN] or anatomical action units [@GANimation] to control the layout and expression of generated faces, and style-guided generation [@stylegan_ffhq][@stylegan2], utilizing adaptive instance normalization [@adain] to inject style guidance information into generated images. Also, combining semantic and style guidance together leads to multi-modal image generation [@bicyclegan], enabling separable pose and appearance control of the output images. Inspired by SPADE [@spade] and SEAN [@sean] for semantic-guided image generation based on *external* parsing maps, our HiFaceGAN utilizes the SPADE layers to implement collaborative suppression and replenishment for multi-stage face renovation, which progressively replenishes plausible details based on hierarchical semantic guidance, leading to an automated renovation pipeline without external guidance.
Face Renovation {#sec:face_renovation}
===============
Generally, the acquisition and storage of digitized images involves many sources of degradations, including but not limited to discrete sampling, camera noise and lossy compression, as shown in Fig \[fig:intro\]. Non-blind face restoration methods typically focus on reversing a specific source of degradation, such as super resolution, denoising and compression artifact removal, leading to limited generalization capability over varying degradation types, Fig \[fig:teaser\]. On the other hand, blind face restoration often relies on the structural prior or external guidance labels for training, leading to quality inconsistency between foreground and background contents. To resolve the issues in existing face restoration works, we present face renovation to explore the capability of generative models for “dual-blind” face restoration without degradation prior and external guidance. Although it would be ideal to collect authentic low-quality and high-quality image pairs of real persons for better degradation modeling, the associated legal issues concerning privacy and portraiture rights are often hard to circumvent. In this work, we perturb a challenging, yet purely artificial face dataset [@stylegan_ffhq] with heterogeneous degradation in varying types and intensities to simulate the real-world scenes for FR. Thereafter, the methodology and comprehensive evaluation metrics for FR are analyzed in detail.
Degradation Simulation {#sec:train_setting}
----------------------
With richer facial details, more complex background contents, and higher diversity in gender, age, and ethnic groups, the synthetic dataset FFHQ [@stylegan_ffhq] is chosen for evaluating FR models with sufficient challenges. We simulate the real-world image degradation by perturbing the FFHQ dataset with different types of degradations corresponding to respective face processing subtasks, which will be also evaluated upon our proposed framework to demonstrate its versatility. For FR, we superimpose four types of degradation (except 16x mosaic) over clean images in random order with uniformly sampled intensity to replicate the challenge expected for real-world application scenarios. [^1] Fig. \[fig:deg\_egs\] displays the visual impact of each type of degradation upon a clean input face. It is evident that mosaic is the most challenging due to the severe corruption of facial boundaries and fine-grained details. Blurring and down-sampling are slightly milder, with the structural integrity of the face almost intact. Finally, JPEG compression and additive noise are the least conceptually obtrusive, where even the smallest details (such as hair bang) are clearly discernable. As will be evidenced later in Sec. \[sec:sota\], the visual impact is consistent with the performance of the proposed face renovation model. Finally, the full degradation for FR is more complex and challenging than all subtasks (except 16x mosaic), with both additive noises/artifacts and detail loss/corruption. We believe the proposed degradation simulation can provide sufficient yet still manageable challenge towards real-world FR applications.
Methodology
-----------
With the single dominated type of degradation, existing methods are devoted to fit an inverse transformation to recover the image content. When it comes to real-world scenes, the low-quality facial images usually contain unidentified heterogeneous degradation, necessitating a unified solution that can simultaneously address common degradations without prior knowledge. Given a severely degraded facial image, the renovation can be reasonably decomposed into two steps, 1) suppressing the impact of degradations and extracting robust semantic features; 2) replenishing fine details in a multi-stage fashion based on extracted semantic guidance. Generally speaking, a facial image can be decomposed into semantic hierarchies, such as structures, textures, and colors, which can be captured within different receptive fields. Also, noise and artifacts to be suppressed need to be adaptively identified according to different scale information. This motivates the design of HiFaceGAN, a multi-stage renovation framework consisting of several nested collaborative suppression and replenishment(CSR) units that is capable of resolving all types of degradation in a unified manner. Implementation details will be introduced in the following section.
![Visualization of degradation types and the corresponding face manipulation tasks. []{data-label="fig:deg_egs"}](images/degradation_examples.pdf){width="0.95\linewidth"}
{width="0.95\linewidth"}
Evaluation Criterion {#sec:eval}
--------------------
For real-world applications, the evaluation criterion for face renovation should be more consistent with human perception rather than machine judgement. Therefore, besides commonly-adopted PSNR and SSIM [@SSIM][@MultiscaleSSIM] metrics, the evaluation criterion for FR should also reflect the semantic fidelity and perceptual realism of renovated faces. For semantic fidelity, we measure the feature embedding distance (FED) and landmark localization error (LLE) with a pretrained face recognition model [@dlib_paper], where the average L2 norm between feature embeddings is adopted for both metrics. For perceptual realism, we introduce FID [@FID] and LPIPS [@LPIPS] to evaluate the distributional and elementwise distance between original and generated samples in the respective perceptual spaces: For FID it is defined by a pre-trained Inception V3 model [@inceptionv3], and for LPIPS, an AlexNet [@AlexNet]. Also, the NIQE [@niqe] metric adopted for the 2018 PIRM-SR challenge [@esrgan] is introduced to measure the naturalness of renovated results for in-the-wild face images. Moreover, we will explain the trade-off between statistical and perceptual scores with ablation study detailed in Sec. \[sec:ablation\].
The Proposed HiFaceGAN
======================
In this section, we detail the architectural design and the working mechanism of the proposed HiFaceGAN. As shown in Fig. \[fig:network\], the suppression modules aim to suppress heterogeneous degradation and encode robust hierarchical semantic information to guide the subsequent replenishment module to reconstruct the renovated face with corresponding photorealistic details. Further, we will illustrate the multi-stage renovation procedure and the functionality of individual units in Fig. \[fig:interp\] justifying the proposed methodology as well as providing new insights to the face renovation task.
Network Architecture {#sec:architecture}
--------------------
We propose a nested architecture containing several CSR units that each attend to a specific semantic aspect. Concretely, we cascade the front-end suppression modules to extract layered semantic features, in an attempt to capture the semantic “hierarchy” of the input image. Accordingly, the corresponding multi-stage renovation pipeline is implemented via several cascaded replenishment modules that each attend to the incoming layer of semantics. Note that the resulted renovation mechanism differs from the commonly-perceived “coarse-to-fine” strategy as in progressive GAN [@progressivegan][@progressive-face-sr]. Instead, we allow the proposed framework to automatically learn a reasonable semantic decomposition and the corresponding face renovation procedure in a completely data-driven manner, maximizing the collaborative effect between the suppression and replenishment modules. More evidence will be provided in Sec. \[sec:understanding\].
![Implementation of the suppression module.[]{data-label="fig:draft"}](images/draft.pdf){width="0.9\linewidth"}
**Suppression Module** A key challenge for face renovation lies in the heterogeneous degradation mingled within real-world images, where a conventional CNN layer with fixed kernel weights could suffer from the limited ability to discriminate between image contents and degradation artifacts. Specifically, we take a look at a conventional spatial convolution with kernel $W\in\mathbb{R}^{C\times C'\times S\times S}$:
$$\label{eqn:conv}
\mathrm{conv}(F;W)_{i} = (F * W)_{i} = \sum_{j\in\Omega(i)} w_{\Delta ji}f_j$$
where $i,j$ are 2D spatial coordinates, $\Omega(i)$ is the sliding window centering at $i$, $\Delta ji$ is the offset between $j$ and $i$ that is used for indexing elements in $W$. The key observation from Eqn. is that the conventional CNN layer shares the same kernel weights over the entire image, making the feature extraction pipeline *content-agnostic*. In other words, both the image content and degradation artifacts will be treated in an equal manner and aggregated into the final feature representation, with potentially negative impacts to the renovated image. Therefore, it is highly desirable to select and aggregate informative features with *content-adaptive* filters, such as LIP [@LIP] or PAC [@PixelAdaptiveCNN]. In this work, we implement the suppression module as shown in Fig. \[fig:draft\] to replace the conventional convolution operation in Eqn. , which helps select informative feature responses and filter out degradation artifacts through adaptive kernels. Mathematically,
$$\label{eqn:draft}
S(F;W)_{i} = \sum_{j\in\Omega(i)} \phi(f_j,f_i) w_{\Delta ji} f_j$$
where $\phi(\cdot,\cdot)$ aims to modulate the weight of convolution kernels with respect to the correlations between neighborhood features. Intuitively, one would expect a correlation metric to be symmetric, i.e. $\phi(f_i,f_j) = \phi(f_j,f_i),\forall f_i, f_j \in \mathbb{R}^{C}$, which can be fulfilled via the following parameterized inner-product function:
$$\label{eqn:phi}
\phi(f_i,f_j) = \mathcal{D}(\mathcal{G}(f_i)^\top \mathcal{G}(f_j) )$$
where $\mathcal{G}: \mathbb{R}^C \rightarrow \mathbb{R}^D$ carries the raw input feature vector $f_i \in \mathbb{R}^C$ into the D-dimensional correlation space, which can help reduce the redundancy of raw input features between channels, and $\mathcal{D}$ is a non-linear activation layer to adjust the range of the output, such as sigmoid or tanh. In practice, we implement $\mathcal{G}$ with a small multi-layer perceptron to learn the modulating criterion in an end-to-end fashion, maximizing the discriminative power of semantic feature selection for subsequent detail replenishment.
**Replenishment Module** Having acquired semantic features from the front-end suppression module, we now focus on utilizing the encoded features for guided detail replenishment. Existing works on semantic-guided generation has achieved remarkable progress with spatial adaptive denormalization (SPADE) [@spade], where semantic parsing maps are utilized to guide the generation of details that belongs to different semantic categories, such as the sky, sea, or trees. We leverage such progress by incorporating the SPADE block into our cascaded CSR units, allowing effective utilization of encoded semantic features to guide the generation of fine-grained details in a hierarchical fashion. In particular, the progressive generator contains several cascaded SPADE blocks, where each block receives the output from the previous block and replenish new details following the guidance of the corresponding semantic features encoded with the suppression module. In this way, our framework can automatically capture the global structure and progressively filling in finer visual details at proper locations even without the guidance of additional face parsing information.
Loss Functions {#sec:loss}
--------------
Most face restoration works aims to optimize the mean-square-error (MSE) against target images [@srcnn][@vdsr][@edsr], which often leads to blurry outputs with insufficient amount of details [@esrgan]. Corresponding to the evaluation criterion in Sec. \[sec:eval\], it is crucial that the renovated image exhibits high semantic fidelity and visual realism, while slight signal-level discrepancies are often tolerable. To this end, we follow the adversarial training scheme [@gan] with an adversarial loss $\L_{GAN}$ to encourage the realism of renovated faces. Here we adopt the LSGAN variant [@lsgan] for better training dynamics: $$\label{eqn:GAN}
\L_{GAN} = \mathrm{E}[\|\log(D(I_{gt}) - 1\|_2^2] + \mathrm{E}[\|\log(D(I_{gen})\|_2^2]$$
Also, we introduce the multi-scale feature matching loss $\L_{FM}$ [@pix2pixHD] and the perceptual loss $\L_{perc}$[@perceptual_loss] to enhance the quality and visual realism of facial details: $$\label{eqn:fm}
\L(\phi) = \sum_{i=1}^{L} \frac{1}{H_i W_i C_i} \|\phi_i(I_{gt}) - \phi_i(I_{gen})\|_2^2$$
where for the adversarial loss $\L_{FM}$, $\phi$ is implemented via the multi-scale discriminator $D$ in [@pix2pixHD] and for the perceptual loss $\L_{perc}$, a pretrained VGG-19 [@vgg] network. Finally, combining Eqn. and leads to the final training objective: $$\label{eqn:recon}
\L_{recon} = \L_{GAN} + \lambda_{FM}\L_{FM} + \lambda_{perc}\L_{perc}$$
![Visualization of (a) the working mechanism of HiFaceGAN and (b) its advantages against existing face restoration works. []{data-label="fig:interp"}](images/Fig-understanding_v6.pdf){width="\linewidth"}
Discussion {#sec:understanding}
----------
**Understanding of HiFaceGAN** To illustrate what each CSR unit can generate at the corresponding stage and how they work cooperatively to perform face renovation outstandingly, we provide an illustrative example shown in Fig. \[fig:interp\](a), where we ablate certain units by replacing the corresponding semantic feature map (blue dots) with a constant tensor (hollow circles), leading to a plain grey background to better isolate the contents generated at each individual stage. Given a 16x down-sampled low-quality facial image, we first sequentially utilize single semantic guidance from the inner stage to the outer stage, the upper row in Fig. \[fig:interp\], and then show the results of the accumulation of semantic guidance in the lower row of Fig. \[fig:interp\]. It is impressive that single semantic guidance from a specific stage leads the corresponding replenishment module to generate a hierarchical layer, which from the inner stage to the outer stage focuses on facial landmarks, edges and textures, shades and reflections, tune and illumination, colorization respectively. In details, by progressively adding semantic guidance, it can be found that with larger receptive field and high-level semantic features, our HiFaceGAN sketches the rough face boundary and localizes facial landmarks, allowing the subsequent CSR unit to replenish fine details upon the basic facial structure when the receptive field goes small and the resolution raises up. The step-by-step face renovation process acts like a hierarchical layer-by-layer overlaying of contents generated with replenishment modules in a semantic-guided fashion, which gradually enhances the visual quality and realism of the renovated image. So far, the progressive face renovation process with logically reasonable ordered steps has justified our heuristics in the network architecture design and illustrate the efficacy and interpretability of HiFaceGAN in a convincing manner.
Additionally, the advantages of the proposed HiFaceGAN can be summarized in three aspects:
- **Versatility** Without degradation prior and explicit guidance, our HiFaceGAN works in a “dual-blind” fashion which can resolve subtasks of face restoration in a unified fashion without tweaking the training configurations, Table \[tab:compare\_sota\_sr\].
- **Robustness** Our HiFaceGAN can withstand severe degradations even more intense than the training data, Fig. \[fig:pressure\_test\].
- **Generalization Ability** With content-adaptive suppression and hierarchical semantic guided replenishment, our HiFaceGAN performs well in real-world scenarios, Sec. \[sec:HPR\].
**Comparison with Blind Face Restoration** To better clarify the distinctions between our face renovation framework and existing restoration methods, we compare the residual maps generated with our HiFaceGAN and the state-of-the-art blind face restoration network GFRNet [@BlindFR-ECCV2018]. As shown in Fig. \[fig:big\_figure\](b), the residual map generated with GFRNet packs heavier noise and less semantically meaningful details, indicating a higher focus on “suppression” and insufficient attention to “replenishment”. This could be attributed to the PSNR-oriented optimization objective, where additive noises contribute a large proportion of the signal discrepancy. In contrast, HiFaceGAN can simultaneously suppress degradation artifacts and replenish semantic details, leading to semantic-aware residual maps and more refined renovation results. Also, HiFaceGAN can renovate background contents and foreground faces together, leading to consistent quality improvement across the entire image. This justifies the rationale of the “dual-blind” setting towards real-world applications with images containing rich non-facial contents.
{width="0.88\linewidth"}
Experiments {#sec:experiments}
===========
In this section, we demonstrate the versatility, robustness and generalization ability of our proposed HiFaceGAN over a wide range of related face restoration sub-tasks, both on synthetic images and real-world photographs. Furthermore, an thorough ablation study is performed to verify our major contributions and stimulate future research directions. Detailed configurations are provided in supplementary materials to facilitate the reproduction.
Comparison with state-of-the-art methods {#sec:sota}
----------------------------------------
We first evaluate our framework on five subtasks, including super resolution, hallucination, denoising, deblurring and compression artifact removal. For each subtask, the dataset is prepared by performing task-specific degradation upon raw images from FFHQ [@stylegan_ffhq], Fig. \[fig:deg\_egs\]. Finally, five most competitive task-specific methods, along with the state-of-the-art blind face restoration baseline [@BlindFR-ECCV2018], are chosen to compete with our HiFaceGAN over the most challenging and practical FR task.
**Comparison with Task-Specific Baselines** Overall, HiFaceGAN outperforms all baselines by a huge margin on perceptual performance, with 3-10 times gain on FID and 50%-200% gain on LPIPS, Table \[tab:compare\_sota\_sr\]. Furthermore, HiFaceGAN even outperform real images in terms of naturalness, as reflected by the NIQE metric. Generally, our generative approach is better suited for tasks with heavy structural degradation, such as face hallucination, denoising and deblurring. For super-resolution and JPEG artifact removal, the structural degradation is considerably milder(Fig. \[fig:deg\_egs\]), leading to narrowed gaps between task-specific solutions and our generalized framework, especially on statistical scores. This is reasonable since the training functions are more perceptually inclined for FR. Nevertheless, it is still possible to trade-off between perceptual and statistical performance, as will be discussed in ablation study.
For qualitative comparison, we showcase the representative results on corresponding tasks in Fig. \[fig:big\_figure\]. For all subtasks, our HiFaceGAN can replenish rich and convincing visual details, such as hair bangs, beards and wrinkles, leading to consistent, photo-realistic renovation results. In contrast, other task-specific methods either produce over-smoothed or color-shifted results (WaveletCNN, SIDNet), or incur severe systematic artifacts during detail replenishment (ESRGAN, Super-FAN). Moreover, our dual-blind setting is equally effective in enhancing details for non-facial contents, such as the interweaving grids on the microphone. In summary, HiFaceGAN can resolve all types of degradation in a unified manner with stunning renovation performances, verifying the efficacy of the proposed methodology and architectural design. More results are provided in the supplementary material.
**Dual-Blind vs. Single-Blind** To discuss the impact of external guidance, we compare our HiFaceGAN with state-of-the-art “single-blind” baseline GFRNet [@BlindFR-ECCV2018] over the fully-degraded FFHQ datset, where the ground truth image is provided as the high-quality guidance during testing. As shown in column 7-9 of Fig. \[fig:big\_figure\], even with the strongest guidance, GFRNet is still less effective in suppressing noises and replenishing fine-grained details than our network, indicating its limitation in feature utilization and generative capability. Consistent with our observation in Sec. \[sec:understanding\], the performance gain of GFRNet against other baselines is mainly statistical, where the semantic and perceptual scores are less competitive, Table \[tab:compare\_sota\_sr\]. Our empirical study suggests that 1) the lack of explicit guidance does not necessarily lead to inferior performance of face renovation; 2) the ability to replenish plausible details is most crucial for high-quality face renovation.
Historic Photograph Renovation {#sec:HPR}
------------------------------
The historic group photograph of famous physicists at the contemporary age taken at the 5th Solvay Conference in 1927 is utilized to evaluate generalization capability of state-of-the-art models for real-world face renovation, Fig. \[fig:teaser\]. We crop $64\times64$ face patches from the original image and resize them to $512\times512$ with bicubic interpolation for input. Apparently, compared to others, our HiFaceGAN can successfully suppress complex degradation in real old photos to generate faces with high definition, high fidelity, and fewer artifacts, while replenishing realistic details, such as facial luster, fine hair, clear facial features, and photo-realistic wrinkles. More outstanding renovation results are displayed in Fig. \[fig:big\_figure\]. Inevitably, the renovated faces contain minor artifacts that mostly occur at shading regions, where degradation artifacts have severely corrupted the underlying contents. Nevertheless, the renovated high-resolution person portraits still possess much better visual and artistic quality than the original input, which simultaneously demonstrates the capability of our model and the challenge in real-world applications.
Ablation Study {#sec:ablation}
--------------
![ Pressure test of the proposed HiFaceGAN. Red numbers indicate the degradation level has gone beyond the limit of training data. []{data-label="fig:pressure_test"}](images/pressure_test.pdf){width="0.93\linewidth"}
We perform an ablation study over the most challenging 16x face hallucination task. Four ablation methods are designed to verify our major contributions, as described below:
- **SPADE** The vanilla SPADE [@spade] network with semantic guidance being face parsing maps extracted from the original *high-resolution* images with a pretrained parsing model [@MaskGAN].
- **16xFace** replaces the semantic parsing map in SPADE with degraded faces containing 16-pixel mosaics.
- **FixConv** retains the nested CSR architectue of HiFaceGAN with the normal content-agnostic convolution layer in Eqn .
- **L1** adds an additional L1 loss upon default HiFaceGAN to adjust between statistical and perceptual scores.
The evaluation scores are reported in Table \[tab:ablation\]. Although face parsing maps provide much finer spatial guidance, it is evident that face renovation relies more on semantic features, as reflected by the huge performance gap between SPADE and 16xFace. Also, FixConv achieves visible performance gain by extracting hierarchical semantic features and applying multi-stage face renovation, verifying the proposed nested architecture. Moreover, incorporating the content-adaptive suppression module further improves the feature selection and degradation suppression ability, leading to substantial gain over FixConv on perceptual and semantic scores. Finally, adding an L1 loss term makes the model statistically inclined, with superior PSNR/SSIM and inferior FID/LPIPS/NIQE scores, verifying the flexibility of our framework to trading off between statistical and perceptual performances.
**Pressure Test** To verify the robustness of HiFaceGAN, we conduct two sets of pressure test targeting at the suppression and replenishment module respectively. For the suppression test, we add random noises upon clean images with up to 140% peak amplitude (twice the energy) of the training data; For the replenishment test, we evaluate a 16x super-resolution model with images downsampled up to 64x ratio. Fig. \[fig:pressure\_test\] displays the renovation results of our HiFaceGAN under extreme degradations. The proposed suppression module is effective for extracting robust semantic features under heavy noise, and the replenishment module can still recover plausible faces even for input beyond human recognition, where most signal discrepancies are smartly “hidden” in the more stochastic hair regions, thus mitigating the negative impact on the naturalness and perceptual realism of renovated faces. Overall, the pressure test demonstrates the impressive efficacy of the proposed collaborative suppression and replenishment framework.
Conclusion and Future Work
==========================
In this paper, we present a challenging, yet more practical task towards real-world photo repairing applications, termed as “face renovation”. Particularly, we propose HiFaceGAN, a collaborative suppression and replenishment framework that works in a “dual-blind” fashion, lifting the usual requirements of degradation prior or structural guidance for training. Extensive experiments on both synthetic face images and real-world historic photographs have demonstrated the versatility, robustness and generalization capability over a wide range of face restoration tasks, outperforming current state-of-the-art by a large margin. Furthermore, the working mechanism of HiFaceGAN, and the rationality of the “dual-blind” setting are justified in a convincing manner with illustrative examples, bringing fresh insights to the subject matter. In the future, we envision that the proposed HiFaceGAN would serve as a solid stepping stone towards the expectations of face renovation. Specifically, the severe degradation often lead to content ambiguity for renovation, like the Afro haircut appeared in Fig. \[fig:big\_figure\] where our method misjudged as normal straight hairs, which motivates us to increase the diversity and balance between different ethnic groups during data collection. Also, it is still a huge challenge for the renovation of objects with regular geometric shapes (such as glasses) and partially-occluded faces — a typical case where external structural guidance could be beneficial. Therefore, exploring multi-modal generation networks with both structural and semantical guidance is another possibility.
now.
[^1]: The python script will be provided in supplementary materials.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have performed a comprehensive investigation of the global integrated flux density of M33 from radio to ultraviolet wavelengths, finding that the data between $\sim$100GHz and 3THz are accurately described by a single modified blackbody curve with a dust temperature of $T_\mathrm{dust}$ = 21.67 $\pm$ 0.30K and an effective dust emissivity index of $\beta_\mathrm{eff}$ = 1.35 $\pm$ 0.10, with no indication of an excess of emission at millimeter/sub-millimeter wavelengths. However, sub-dividing M33 into three radial annuli, we found that the global emission curve is highly degenerate with the constituent curves representing the sub-regions of M33. We also found gradients in $T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ across the disk of M33, with both quantities decreasing with increasing radius. Comparing the M33 dust emissivity with that of other Local Group members, we find that M33 resembles the Magellanic Clouds rather than the larger galaxies, i.e., the Milky Way and M31. In the Local Group sample, we find a clear correlation between global dust emissivity and metallicity, with dust emissivity increasing with metallicity. A major aspect of this analysis is the investigation into the impact of fluctuations in the Cosmic Microwave Background (CMB) on the integrated flux density spectrum of M33. We found that failing to account for these CMB fluctuations would result in a significant over-estimate of $T_\mathrm{dust}$ by $\sim$5K and an under-estimate of $\beta_\mathrm{eff}$ by $\sim$0.4.'
author:
- 'C.T. Tibbs$^{1}$[^1], F.P. Israel$^{2}$[^2], R.J. Laureijs$^{1}$, J.A. Tauber$^{1}$, B. Partridge$^{3}$, M.W. Peel$^{4}$,'
- |
L. Fauvet$^{5}$\
$^{1}$Scientific Support Office, Directorate of Science, European Space Research and Technology Centre (ESA/ESTEC),\
Keplerlaan 1, 2201 AZ, Noordwijk, The Netherlands\
$^{2}$Leiden Observatory, Leiden University, P.O. Box 9513, 2300 RA, Leiden, The Netherlands\
$^{3}$Department of Astronomy, Haverford College, Haverford, PA 19041, USA\
$^{4}$Departamento de Física Matematica, Instituto de Física, Universidade de São Paulo, São Paulo, Brazil\
$^{5}$ARGANS Limited, Tamar Science Park, Plymouth, PL6 8BX, UK
title: Planck Observations of M33
---
\[firstpage\]
galaxies: individual: M33 – galaxies: ISM – galaxies: photometry – infrared: galaxies – submillimetre: galaxies – radio continuum: galaxies
Introduction {#Sec:Intro}
============
In the region between high-frequency radio waves ($\nu$$\gtrsim$10GHz) and long-wavelength infrared (IR) emission ($\lambda$$\gtrsim$100$\mu$m), thermal radiation from interstellar dust and ionized gas, as well as non-thermal synchrotron radiation, all contribute to the emission from cosmic objects. By unravelling the various contributions, we may obtain information on the ionising stars and the properties of interstellar dust in a variety of galactic environments. The observations provided by the *Planck* mission [@Tauber:10] allow us to sample the poorly observed far-IR to millimetre (mm) gap in the continuum emission spectrum of objects such as entire galaxies.
In the past, attempts have been made to extrapolate the incomplete IR continuum flux density spectrum (frequently, but incorrectly, referred to as a spectral energy distribution or SED[^3]) cutting off somewhere between 100 and 160$\mu$m (*IRAS, Spitzer Space Telescope*) by assuming a single effective integrated dust emissivity index of $\beta_\mathrm{eff}$ = 2 for the Rayleigh-Jeans extrapolation. Often, the values actually measured at wavelengths around 1mm significantly exceed such extrapolated flux densities. This so-called “millimetre excess” was readily interpreted as evidence for a large mass of colder dust [see, for instance, @Galliano:05]. However, both the increased far-IR wavelength coverage (up to 500$\mu$m) of the *Herschel Space Observatory* and the results of terrestrial laboratory experiments have subsequently indicated that the actual value of $\beta_\mathrm{eff}$ is generally $<$2. As a consequence, both the historic millimetre excess and the implied large mass of colder dust, can be reduced to an artefact of interpretation, and effectively disappear.
Thus far, reliable and complete continuum flux density spectra ranging from the radio to the mid-IR or even optical wavelengths, well-sampling the mm to far-IR range, have been published for a variety of Milky Way sources, but only for a few galaxies beyond. The wavelength coverage of *Planck* renders extrapolation superfluous; the value of $\beta_\mathrm{eff}$ can be measured directly. This measurement is complicated by the degeneracy between $\beta_\mathrm{eff}$ and the dust temperature, $T_\mathrm{dust}$, derived from flux densities that have finite instrumental noise: the two parameters are inversely correlated [e.g., @Shetty:09; @Juvela:12a; @Juvela:12b]. Nevertheless, the *Planck* spectra of the Local Group galaxies, the Large and Small Magellanic Clouds [LMC and SMC; @Planck_Early_Results_XVII:11], and M31 [@Planck_Intermediate_Results_XXV:15], imply galaxy-wide effective dust emissivities well below two. Similar emissivities have been found for other nearby galaxies [@Planck_Early_Results_XVI:11].
Surprisingly, the complete flux density spectra of the LMC and SMC, incorporating *WMAP* and *COBE* data, published by @Israel:10 and interpreted by @Bot:10, do show a pronounced excess of emission at mm to cm wavelengths. This “new” excess emission is not to be confused with the apparent historical millimetre excess discussed earlier as it does not result from an arbitrary assumption on the dust emissivity, but is a well-sampled spectral feature. Its existence was confirmed by @Planck_Early_Results_XVII:11, who explained the observed excess in the LMC as a fluctuation of the Cosmic Microwave Background (CMB), but admitted to the presence of a significant intrinsic excess in the SMC. @Draine:12 proposed that this intrinsic excess in the SMC could be explained if the interstellar dust includes magnetic nanoparticles, emitting magnetic dipole radiation resulting from the thermal fluctuations in the magnetisation.
Perhaps relatedly, an excess of emission at longer (cm) wavelengths has also been observed in many environments within the Milky Way [see @Planck_Intermediate_Results_XV:14 and references within]. This cm excess, more commonly known as anomalous microwave emission (AME), is typically observed at frequencies around 30GHz (or wavelengths of 1cm), is observed to be highly correlated with the IR dust emission [e.g., @Casassus:06; @Tibbs:10; @Tibbs:13; @Planck_Intermediate_Results_XV:14], and is believed to be due to electric dipole radiation from very small rapidly spinning dust grains [@Draine:98].
In this paper we present a study of the small Local Group spiral galaxy M33, using the most recent *Planck* data along with data from the literature, to produce a comprehensive continuum flux density spectrum from radio to ultraviolet (UV) wavelengths. We profit from the fact that, due to its proximity [$d$ = 840 kpc, @Freedman:91] and modest dimensions (approximately 70arcmin $\times$ 40arcmin - see Fig. \[Fig:M33\_PLCK857\]), M33 is an exceedingly well-studied object. In this analysis we will specifically address: (a) the shape of the Rayleigh-Jeans spectrum, (b) the magnitude of the effective dust emissivity spectral index, $\beta_\mathrm{eff}$, and (c) possible differences between the inner and outer regions of M33. Since the flux density spectra of both individual interstellar clouds and entire galaxies have a minimum close to the peak of the CMB, at these frequencies the CMB emission typically exceeds the interstellar contribution. Thus, our results depend critically on the reliability of the CMB subtraction. For this reason, we will pay special attention to an analysis of the CMB fluctuations, as these dominate the M33 spectrum at mm wavelengths.
This paper is organised as follows. In Section \[Sec:Data\] we describe the data used in this analysis, while in Section \[Sec:Analysis\] we produce a global continuum flux density spectrum for M33, accounting for contributions from both CMB fluctuations and CO line emission. We also spatially decompose M33 into three annuli, producing a flux density spectrum for each. In Section \[Sec:Discussion\] we discuss the results of our work, and we present our conclusions in Section \[Sec:Conclusions\].
Data {#Sec:Data}
====
[lcccc]{} Telescope/Instrument & $\nu_\mathrm{ref}$ & $\theta$ & $\epsilon_\mathrm{phot}$\
& (GHz) & (arcmin) &\
***Planck*** & & &\
LFI030 & 28.4 & 32.3 & 1$\%$\
LFI044 & 44.1 & 27.1 & 1$\%$\
LFI070 & 70.4 & 13.3 & 1$\%$\
HFI100 & 100 & 9.7 & 1$\%$\
HFI143 & 143 & 7.3 & 1$\%$\
HFI217 & 217 & 5.0 & 1$\%$\
HFI353 & 353 & 4.9 & 1$\%$\
HFI545 & 545 & 4.8 & 7$\%$\
HFI857 & 857 & 4.6 & 7$\%$\
\
***Herschel*** & & &\
SPIRE 500$\mu$m & 600 & 0.60 & 10$\%$\
SPIRE 350$\mu$m & 857 & 0.41 & 10$\%$\
SPIRE 250$\mu$m & 1200 & 0.30 & 10$\%$\
PACS 160$\mu$m & 1870 & 0.19 & 15$\%$\
PACS 100$\mu$m & 3000 & 0.12 & 15$\%$\
PACS 70$\mu$m & 4280 & 0.09 & 15$\%$\
\
***IRAS*/IRIS** & & &\
100$\mu$m & 3000 & 4.3 & 13.5$\%$\
60$\mu$m & 5000 & 4.0 & 10.4$\%$\
25$\mu$m & 12000 & 3.8 & 15.1$\%$\
12$\mu$m & 25000 & 3.8 & 5.1$\%$\
\
***Spitzer*** & & &\
MIPS 24$\mu$m & 12500 & 0.10 & 10$\%$\
IRAC 8$\mu$m & 37500 & 0.03 & 10$\%$\
IRAC 5.8$\mu$m & 51700 & 0.03 & 10$\%$\
IRAC 4.5$\mu$m & 66600 & 0.03 & 10$\%$\
IRAC 3.6$\mu$m & 83300 & 0.03 & 10$\%$\
\[Table:Data\]
 
Planck {#Subsec:Planck}
------
The *Planck* mission [@Tauber:10; @Planck_Early_Results_I:11] was the third cosmological satellite mission to observe the entire sky in a series of wide spectral passbands ($\Delta \nu/\nu\sim$0.3–0.6) designed to sample the CMB. It measured the emission from the sky with the Low Frequency Instrument (LFI) at 28.4, 44.1, and 70.4GHz (1.0–0.4cm) with amplifiers cooled to 20K between 2009 and 2013, and the High Frequency Instrument (HFI) at 100, 143, 217, 353, 545, and 857GHz (3.0–0.35mm), with bolometers cooled to 0.1K between 2009 and early 2012 (cf., Table \[Table:Data\]).
In this paper, we use the most recent *Planck* 2015 “full” data release [@Planck_2015_Results_I:16]. These data cover the full mission from 12 August 2009 to 23 October 2013 and are available to download from the *Planck* Legacy Archive.[^4] The full LFI/HFI data processing and calibration procedures are described in @Planck_2015_Results_II:16 [@Planck_2015_Results_III:16; @Planck_2015_Results_IV:16; @Planck_2015_Results_V:16; @Planck_2015_Results_VI:16; @Planck_2015_Results_VII:16; @Planck_2015_Results_VIII:16] with an overview provided in @Planck_2015_Results_I:16. The *Planck* full-sky maps are provided in <span style="font-variant:small-caps;">HEALPix</span> format [@Gorski:05], but for this analysis we extracted 2D projected maps centred on M33 using the <span style="font-variant:small-caps;">Gnomdrizz</span> package [@Paradis:12], which accurately conserves the photometry during the data re-pixelization. After performing this extraction for each of the nine *Planck* frequency maps, the maps were converted from units of CMB temperature to MJysr$^{-1}$ using the coefficients described in @Planck_2013_Results_IX:14. The nine *Planck* frequency maps are displayed in the left column of Fig. \[Fig:Planck\_maps\_all\].
The 2015 *Planck* data have been calibrated on the orbital modulation of the “cosmological dipole”, resulting in extremely high (sub-percent) calibration accuracy [see table 1 from @Planck_2015_Results_I:16]. However, it is important to note that the quoted accuracies are appropriate for diffuse emission at large angular scales, where the calibration signal appears. Additional uncertainties apply at smaller angular scales. For relatively compact sources such as M33, the main additional contributors are related to colour correction and to beam uncertainty. Colour correction uncertainties depend on the spectral shape of the source [@Planck_2015_Results_II:16; @Planck_2015_Results_VII:16], while the beam uncertainties depend on angular scale [@Planck_2015_Results_IV:16; @Planck_2015_Results_VII:16]; for M33 we conservatively assume the entire solid angle uncertainty to be applicable. Combining these uncertainties in quadrature, we conservatively assume a photometric uncertainty of 7% for the 545 and 857GHz bands, and 1% for the other seven bands. Ultimately, the uncertainty on the flux determination of compact sources is limited by fluctuations in both the physical backgrounds and foregrounds rather than photometry errors.
Herschel {#Subsec:Herschel}
--------
M33 was mapped with *Herschel* within the framework of the open-time key programme as part of the HerM33es KPOT\_ckrame01\_1 [@Kramer:10] and OT2\_mboquien\_4 [@Boquien:15] proposals. The @Kramer:10 observations (observation ID 1342189079 and 1342189080) were performed simultaneously with the PACS (100 and 160$\mu$m) and SPIRE (250, 350, and 500$\mu$m) instruments in parallel mode in two orthogonal directions to map a region of approximately 90arcmin $\times$ 90arcmin. The @Boquien:15 observations (observation ID 1342247408 and 1342247409) were performed solely with the PACS instrument in two orthogonal directions at 70 and 160$\mu$m, and covered a smaller area of approximately 50arcmin $\times$ 50arcmin. These maps have spatial resolutions ranging from approximately 6 to 11arcsec for the PACS maps and approximately 18 to 37arcsec for the SPIRE maps.
The fully reduced maps were made publicly available by the HerM33es team as *Herschel* User Provided Data Products, and we downloaded these data from the *Herschel* Science Archive.[^5] Full details of the PACS and SPIRE data reduction and map-making are described in detail by @Boquien:11 [@Boquien:15] and @Xilouris:12, respectively. Following @Boquien:11 and @Xilouris:12, we assume a 15 and 10% photometric uncertainty on the extended emission in these PACS and SPIRE maps, respectively.
IRAS {#Subsec:IRAS}
----
The original *IRAS* measurements of M33 were presented and discussed by @Rice:90. In this analysis we use the Improved Reprocessing of the *IRAS* Survey [IRIS; @Miville-Deschenes:05] data for all four *IRAS* bands at 12, 25, 60, and 100$\mu$m. These data have been reprocessed resulting in an improvement in the zodiacal light subtraction, destriping, and calibration. We use the photometric uncertainties estimated by @Miville-Deschenes:05 of 5.1, 15.1, 10.4, and 13.5% for the 12, 25, 60, and 100$\mu$m bands, respectively.
Spitzer {#Subsec:Spitzer}
-------
*Spitzer* mapped M33 as part of the Guaranteed Time Observations (PID 5, PI. R. Gehrz) and we use the MIPS 24$\mu$m data along with the IRAC 8.0, 5.8, 4.5, and 3.6$\mu$m data. The IRAC observations have spatial resolutions of $\sim$2.0arcsec, while the MIPS 24$\mu$m map has a resolution of 6arcsec. A detailed description of the *Spitzer* photometry of M33 is provided by @Verley:07 [@Verley:09]. For this analysis, we downloaded the data from the *Spitzer* Heritage Archive[^6] and reprocessed the data by subtracting the contribution from the zodiacal light, applying the extended emission correction, mosaicking the data, performing an overlap correction, and subtracting the brightest point sources. This processing was performed using <span style="font-variant:small-caps;">mopex</span> and <span style="font-variant:small-caps;">apex</span> in a similar manner to that discussed in @Tibbs:11, and we assume a calibration uncertainty of 10$\%$ on these maps.
Analysis And Results {#Sec:Analysis}
====================
![All of the *Planck* maps used in this analysis. The first column shows the nine standard *Planck* frequency maps, which contain a contribution from the CMB, while the second, third, fourth, and fifth columns show the CMB map, the CMB mask, and the nine corresponding frequency maps which have had the CMB contribution subtracted using the `SMICA`, `NILC`, `SEVEM`, and `Commander` methods, respectively. Each individual map is 2.5$^\circ$ across and orientated in the Galactic coordinate system as defined in Fig. \[Fig:M33\_PLCK857\].[]{data-label="Fig:Planck_maps_all"}](Fig2.pdf)
Aperture photometry {#Sec:Photometry}
-------------------
In order to determine the integrated flux densities for M33 we applied aperture photometry to the datasets summarised in Section \[Sec:Data\]. For our aperture photometry analysis we used an elliptical aperture with a semi-major axis of 45arcmin (11kpc at the distance of M33), a semi-major to semi-minor axis ratio of $10^{0.23}$ [@Paturel:03], and a position angle of 22.5degrees with respect to an equatorial reference frame [@Kramer:10], centred on M33 ($\ell$ = 133.60$^{\circ}$, $b$ = $-$31.34$^{\circ}$) as indicated in Fig. \[Fig:M33\_PLCK857\]. The size of the aperture was selected based on computing the integrated flux density in apertures of increasing size to determine when the computed flux density stopped growing. The unrelated background and foreground emission was estimated within an elliptical annulus with inner and outer semi-major axes of 1.15 and 1.50 times the aperture semi-major axis, respectively.
The estimated uncertainty on the computed flux densities is a combination of the photometric uncertainty, $\epsilon_\mathrm{phot}$, for which we have adopted the values listed in Table \[Table:Data\], and the flux measurement uncertainty, $\epsilon_\mathrm{bg}$, which contains two terms: the first term is the variance in the aperture flux, and the second term is due to the background/foreground subtraction. Both terms are estimated from the variance in the annulus surrounding the source, computed as
$$\epsilon_\mathrm{bg} = \sigma_\mathrm{bg} \left[ N_\mathrm{aper} + \frac{\pi N_\mathrm{aper}^{2}}{2 N_\mathrm{bg}} \right]^{0.5} ,
\label{equ:unc1}$$
[see also @Laher:12; @Hermelo:16], where $\sigma_\mathrm{bg}$ is the standard deviation of the pixels within the annulus, and $N_\mathrm{aper}$ and $N_\mathrm{bg}$ are the number of pixels within the aperture and the annulus, respectively. The total uncertainty on the computed flux densities is then estimated to be
$$\epsilon = \sqrt{\epsilon_\mathrm{phot}^{2} + \epsilon_\mathrm{bg}^{2}} .
\label{equ:unc3}$$
The aperture photometry described above takes into account the *average* CMB contribution, but not the effect of the CMB fluctuations, which may have a considerable impact on estimates of the M33 integrated flux density. Therefore, before determining the intrinsic M33 flux density spectrum, we need to consider the effect of the CMB on the shape of the spectrum.
CMB contribution {#Subsec:CMB_Contribution}
----------------
At the galactic latitude of M33, Milky Way foreground emission is relatively weak, even at the higher frequencies, and the main background affecting the source flux measurements is the CMB itself, especially at the lower observed frequencies. As can be seen from the left column in Fig. \[Fig:Planck\_maps\_all\], M33 only starts to become clearly visible above the CMB at frequencies $\gtrsim$143GHz. *Planck* has extracted the CMB over the whole sky using four different methods: `SMICA`, a non-parametric method that computes the CMB map by linearly combining all of the *Planck* maps with weights that vary with multipole in the spherical harmonic domain; Needlet Internal Linear Combination (`NILC`), which produces a CMB map using the *Planck* maps between 44 and 857GHz by applying the Internal Linear Combination technique in the needlet (wavelet) domain; `SEVEM`, which estimates a CMB map based on linear template fitting in the map domain using internal templates constructed using the *Planck* data; and `Commander`, which is a Bayesian parametric method that models all of the astrophysical signals in the map domain [@Planck_2015_Results_IX:16].
It is important to note that in these CMB estimations, bright sources in the input maps are masked and the resulting CMB maps contain “inpainted” values within the masked areas. These inpainted values are good estimations of the CMB near the edge of the masked area (to preserve continuity), but are only statistically representative of the CMB within the mask. For this reason, the CMB masks employed by each of the four CMB separation techniques, along with the resulting CMB maps for the vicinity of M33, are displayed in Fig. \[Fig:Planck\_maps\_all\], where it is apparent that M33 was only masked for the `SMICA` analysis. In addition to providing maps of the CMB, the *Planck* Legacy Archive also provide maps containing only foreground emission (i.e., the frequency maps with the CMB contribution subtracted). Since there are four different CMB maps, this produces four sets of CMB-subtracted maps. As recommended by @Planck_2015_Results_IX:16, it is not advisable to produce an analysis that depends solely on a single component separation algorithm, and therefore we investigate the impact of the CMB by incorporating all four of the CMB-subtracted datasets in our analysis. As for the standard *Planck* maps discussed in Section \[Subsec:Planck\], we used <span style="font-variant:small-caps;">Gnomdrizz</span> to extract 2D projected maps of the CMB-subtracted maps and subsequently converted them into MJysr$^{-1}$. These CMB-subtracted maps are displayed in Fig. \[Fig:Planck\_maps\_all\], where it is clear to see variations between the maps, reflecting the different approaches used to estimate the CMB. The differences are most obvious at the lower frequencies $\lesssim$200GHz, where the CMB and its fluctuations are strongest.
The *Planck* consortium recommends using the `SMICA` CMB map at small angular scales, as it results in the lowest foreground residuals. However, in this case, since M33 has been masked, it is the least reliable of the four CMB-subtraction techniques. Nonetheless, since all four of the CMB-subtracted maps are in principle statistically indistinguishable, we keep all four of them in our analysis.
In order to quantify the effect of the CMB as a contaminant of the integrated emission within M33, we compared the flux densities (estimated using aperture photometry as described in Section \[Sec:Photometry\]) in each of the nine *Planck* bands for the five different datasets (the standard *Planck* frequency maps and the `SMICA`, `NILC`, `SEVEM`, and `Commander` CMB-subtracted maps), which we plot in Fig. \[Fig:CMB\_Flux\_Dens\]. These plots clearly illustrate how the CMB impacts the estimate of the integrated flux density. Not only are there differences between the flux density computed using the standard (non CMB-subtracted) maps, but there are also variations between the different CMB estimation methods. These variations are systematic as they only depend on the map frequency and input CMB amplitude, with `SMICA` and `SEVEM` producing the highest and lowest flux densities, respectively, and they also reflect that M33 is masked for the `SMICA` analysis, but not for the other three methods. Fig. \[Fig:CMB\_Flux\_Dens\] also quantitatively shows what can be seen in Fig. \[Fig:Planck\_maps\_all\], which is that the CMB contamination is non-negligible at $\nu$$\lesssim$217GHz, while at higher frequencies the emission from M33 is dominant and hence the CMB contribution becomes increasing negligible.
Contamination by CO line emission {#Subsec:CO_Contribution}
---------------------------------
The wide passbands of the *Planck* instruments cover the rest frequencies of various molecular lines. Emission from these lines will contaminate the measured broadband continuum flux densities, causing the latter to be overestimated. As discussed by @Planck_2013_Results_XIII:14, Galactic line emission from carbon monoxide (CO) is strongly detected in the *Planck* bands. Specifically, only the $J$=1-0, $J$=2-1, and $J$=3-2 transitions of $^{12}$CO need to be considered, as only they are strong enough to have a significant effect on the *Planck* HFI100, HFI217, and HFI353 bands, respectively. Although the *Planck* data themselves have been used to produce maps of the $^{12}$CO emission [@Planck_2015_Results_X:16], these maps were produced for the velocity range of the Milky Way, which is substantially different to that of M33. We have inspected these maps and determined that they are largely unsuitable for this analysis. However, we can still investigate the magnitude of possible CO contamination by using the $^{12}$CO maps obtained with ground-based telescopes. M33 has been observed by @Heyer:04, who used the Five College Radio Astronomy Observatory (FCRAO) 14m telescope to map the $J $=1-0 CO emission, while @Druard:14 used the $J$=2-1 CO maps observed using the Institut de Radioastronomie Millimétrique (IRAM) 30m telescope. Based on these data, we computed a flux density of the $J$=1-0 CO emission in the HFI100 band to be $\sim$0.1Jy, which when compared to the total flux densities estimated in the Planck 100GHz band accounts for $\lesssim$9% of the emission. Likewise, from the IRAM data [see also @Hermelo:16] we estimated a $J$=2-1 CO flux density in the HFI217 band of $\sim$0.7Jy, which accounts for $\lesssim$4% of the emission in that band. Finally, the compilation of CO line ratios in spiral galaxies (Israel et al., in prep.) suggests that the integrated brightness temperature ratio between the $J$=3-2 and $J$=1-0 lines is $\sim$0.7, predicting a corresponding $J$=3-2 CO flux density of $\sim$1.9Jy, which is $\lesssim$3% of the emission in the HFI353 band.
Therefore, we use these flux densities to correct for the contribution from CO line emission. Throughout the rest of this analysis, we include small flux density corrections by subtracting the estimated CO flux density from the observed flux densities at 100, 217, and 353GHz in order to determine the intrinsic M33 continuum flux density spectrum as accurately as possible.
Global continuum flux density spectrum and spectral energy distribution {#Subsec:Full_SED}
-----------------------------------------------------------------------
To produce the global continuum flux density spectrum for M33, we performed aperture photometry on the *Planck*, *Herschel*, IRIS, and *Spitzer* maps at full angular resolution. The resulting flux densities are listed in Table \[Table:M33\_Flux\_Dens\]. As discussed in Section \[Subsec:CMB\_Contribution\], to account for the effect of the CMB, we computed the flux densities for the standard *Planck* maps along with the CMB-subtracted *Planck* maps. The flux densities listed in Table \[Table:M33\_Flux\_Dens\] have also been corrected for CO contamination as described in Section \[Subsec:CO\_Contribution\].
For a proper analysis of the M33 continuum spectrum, we expand its frequency (wavelength) range by adding results available in the literature, in addition to those listed in Table \[Table:M33\_Flux\_Dens\]. These include the radio flux densities from the compilation by @Israel:92 and from @Tabatabaei:07a, as well as the 2MASS $J$, $H$, and $K_{S}$ band flux densities by @Jarrett:03, the $U$, $B$, $V$ values from @DeVaucouleurs:91, and the *GALEX* far- and near-UV flux densities by @Lee:11. The resulting flux density spectra, with and without CMB-subtraction, are displayed in Fig. \[Fig:M33\_Full\_CFD\], while the corresponding SEDs, obtained by multiplying each flux density by its frequency, are shown in Fig. \[Fig:M33\_Full\_SED\].
--------------------------- ----------- ------------ ------------------ ------------------ ------------------ ------------------ ------------------ --
Instrument Frequency Wavelength Flux Density
`SMICA` `NILC` `SEVEM` `Commander`
\[GHz\] \[mm\] \[Jy\]
LFI030 28.4 10.6 0.46 $\pm$ 0.05 0.55 $\pm$ 0.03 0.51 $\pm$ 0.03 0.47 $\pm$ 0.03 0.49 $\pm$ 0.03
LFI044 44.1 6.80 0.61 $\pm$ 0.12 0.77 $\pm$ 0.08 0.63 $\pm$ 0.08 0.57 $\pm$ 0.08 0.64 $\pm$ 0.08
LFI070 70.4 4.26 1.83 $\pm$ 0.12 1.42 $\pm$ 0.14 1.07 $\pm$ 0.14 0.79 $\pm$ 0.14 0.99 $\pm$ 0.14
HFI100 100 3.00 2.93 $\pm$ 0.17 2.25 $\pm$ 0.09 1.54 $\pm$ 0.09 1.10 $\pm$ 0.08 1.45 $\pm$ 0.09
HFI143 143 2.10 7.41 $\pm$ 0.26 5.99 $\pm$ 0.08 4.86 $\pm$ 0.08 4.07 $\pm$ 0.07 4.64 $\pm$ 0.08
HFI217 217 1.38 21.1 $\pm$ 0.4 18.7 $\pm$ 0.2 17.2 $\pm$ 0.2 16.2 $\pm$ 0.2 17.0 $\pm$ 0.2
HFI353 353 0.849 76.6 $\pm$ 0.8 75.2 $\pm$ 0.8 74.0 $\pm$ 0.8 73.1 $\pm$ 0.8 74.0 $\pm$ 0.8
HFI545 545 0.550 241.0 $\pm$ 16.9 241.0 $\pm$ 16.9 239.0 $\pm$ 16.7 238.0 $\pm$ 16.7 239.0 $\pm$ 16.7
SPIRE 500$\mu$m 600 0.500 345.0 $\pm$ 34.5 344.0 $\pm$ 34.4 342.0 $\pm$ 34.2 341.0 $\pm$ 34.1 341.0 $\pm$ 34.1
SPIRE 350$\mu$m 857 0.350 768.0 $\pm$ 76.8 768.0 $\pm$ 76.8 766.0 $\pm$ 76.6 764.0 $\pm$ 76.5 766.0 $\pm$ 76.6
HFI857 857 0.350 693.0 $\pm$ 48.6 694.0 $\pm$ 48.7 693.0 $\pm$ 48.5 692.0 $\pm$ 48.5 693.0 $\pm$ 48.6
SPIRE 250$\mu$m 1200 0.250 1500 $\pm$ 150 1500 $\pm$ 150 1500 $\pm$ 150 1500 $\pm$ 150 1500 $\pm$ 150
PACS 160$\mu$m 1870 0.160 2230 $\pm$ 334 2250 $\pm$ 337 2250 $\pm$ 338 2250 $\pm$ 338 2260 $\pm$ 338
PACS 100$\mu$m 3000 0.100 1400 $\pm$ 210 1380 $\pm$ 208 1380 $\pm$ 207 1380 $\pm$ 208 1380 $\pm$ 207
IRIS 100$\mu$m 3000 0.100 1340 $\pm$ 181 1330 $\pm$ 179 1330 $\pm$ 179 1330 $\pm$ 179 1330 $\pm$ 179
PACS 70$\mu$m 4280 0.0700 535.0 $\pm$ 80.3 531.0 $\pm$ 79.7 529.0 $\pm$ 79.4 530.0 $\pm$ 79.6 527.0 $\pm$ 79.1
IRIS 60$\mu$m 5000 0.0600 464.0 $\pm$ 48.2 465.0 $\pm$ 48.3 464.0 $\pm$ 48.2 468.0 $\pm$ 48.7 460.0 $\pm$ 47.9
IRIS 25$\mu$m 12000 0.0250 53.0 $\pm$ 8.0 51.0 $\pm$ 7.7 50.6 $\pm$ 7.7 50.0 $\pm$ 7.6 50.8 $\pm$ 7.7
MIPS 24$\mu$m 12500 0.0240 53.7 $\pm$ 5.4 52.6 $\pm$ 5.3 52.4 $\pm$ 5.2 52.1 $\pm$ 5.2 52.5 $\pm$ 5.3
IRIS 12$\mu$m 25000 0.0120 43.5 $\pm$ 2.3 43.5 $\pm$ 2.3 43.5 $\pm$ 2.3 43.5 $\pm$ 2.3 43.5 $\pm$ 2.3
IRAC 8$\mu$m 37500 0.00800 77.0 $\pm$ 7.7 77.0 $\pm$ 7.7 77.0 $\pm$ 7.7 77.0 $\pm$ 7.7 77.0 $\pm$ 7.7
IRAC 5.8$\mu$m 51700 0.00580 62.5 $\pm$ 6.3 62.5 $\pm$ 6.3 62.5 $\pm$ 6.3 62.5 $\pm$ 6.3 62.5 $\pm$ 6.3
IRAC 4.5$\mu$m 66600 0.00450 14.2 $\pm$ 1.4 14.2 $\pm$ 1.4 14.2 $\pm$ 1.4 14.2 $\pm$ 1.4 14.2 $\pm$ 1.4
IRIS 3.6$\mu$m 83300 0.00360 18.8 $\pm$ 1.9 18.8 $\pm$ 1.9 18.8 $\pm$ 1.9 18.8 $\pm$ 1.9 18.8 $\pm$ 1.9
\[Table:M33\_Flux\_Dens\]
--------------------------- ----------- ------------ ------------------ ------------------ ------------------ ------------------ ------------------ --
Decomposition of the continuum flux density spectrum {#Subsec:Decomposition}
----------------------------------------------------
In order to quantify the resulting flux density spectra, we fitted each of them with a model simultaneously fitting contributions representing thermal dust emission, as a combination of two modified blackbodies at different temperatures (the use of two modified blackbodies is to insure an accurate fit to the peak of the cold dust component), but with identical dust emissivity indices,
$$S_{\nu}^\mathrm{dust} = \sum\limits_{i=1}^{2} C_\mathrm{dust,i} \left( \frac{\nu}{\nu_{0}} \right)^{\beta_\mathrm{eff}} B_{\nu}(T_\mathrm{dust,i}) ,
\label{equ:S_Td}$$
non-thermal synchrotron emission,
$$S_{\nu}^\mathrm{sync} = C_\mathrm{sync} \left( \frac{\nu}{\nu_{0}} \right)^{\alpha_\mathrm{sync}} ,
\label{equ:S_sync}$$
free-free emission,
$$S_{\nu}^\mathrm{ff} = C_\mathrm{ff} \left( \frac{\nu}{\nu_{0}} \right)^{\alpha_\mathrm{ff}} ,
\label{equ:S_ff}$$
and AME, assuming that it is due to spinning dust emission,
$$S_{\nu}^\mathrm{AME} = C_\mathrm{AME} j_{\nu} ,
\label{equ:S_AME}$$
where $j_{\nu}$ is the spinning dust emissivity for the warm ionised medium computed using the spinning dust model, <span style="font-variant:small-caps;">spdust</span> [@Ali-Haimoud:09; @Silsbee:11]. For each flux density spectrum we used the IDL fitting routine <span style="font-variant:small-caps;">mpfit</span> [@Markwardt:09], which employs the Levenberg-Marquardt least-squares minimisation technique, to fit the data between 1.4GHz and 24$\mu$m for $C_\mathrm{dust,i}$, $T_\mathrm{dust,i}$, $\beta_\mathrm{eff}$, $C_\mathrm{sync}$, $\alpha_\mathrm{sync}$, $C_\mathrm{ff}$, $\alpha_\mathrm{ff}$, and $C_\mathrm{AME}$. During the fitting process, $C_\mathrm{dust,i}$, $T_\mathrm{dust,i}$, $C_\mathrm{sync}$, $C_\mathrm{AME}$ were constrained to be physically realistic (i.e., $\ge$0), $\beta_\mathrm{eff}$ and $\alpha_\mathrm{sync}$ were unconstrained, while $\alpha_\mathrm{ff}$ was fixed to $-$0.1 and $C_\mathrm{ff}$, which as discussed below, was constrained based on additional observations. The estimated uncertainties on these fitted parameters were computed from the resulting covariance matrix.
 
[lccccc]{} Data & $\beta_\mathrm{eff}$ & $T_\mathrm{dust}$ & $\alpha_\mathrm{sync}$ & $\alpha_\mathrm{ff}$ & $\chi^{2}_\mathrm{red}$\
& & \[K\]\
standard & 0.93 $\pm$ 0.01 & 26.33 $\pm$ 0.18 & $-$1.14 $\pm$ 0.08 & $-$0.1 (fixed) & 3.00\
`SMICA` & 1.20 $\pm$ 0.01 & 22.16 $\pm$ 0.17 & $-$1.05 $\pm$ 0.06 & $-$0.1 (fixed) & 2.38\
`NILC` & 1.36 $\pm$ 0.01 & 21.59 $\pm$ 0.14 & $-$1.01 $\pm$ 0.05 & $-$0.1 (fixed) & 1.46\
`SEVEM` & 1.44 $\pm$ 0.01 & 21.60 $\pm$ 0.13 & $-$1.07 $\pm$ 0.06 & $-$0.1 (fixed) & 3.78\
`Commander` & 1.39 $\pm$ 0.01 & 21.51 $\pm$ 0.14 & $-$1.02 $\pm$ 0.05 & $-$0.1 (fixed) & 1.62\
\
Mean of CMB-subtracted & 1.35 $\pm$ 0.10 & 21.67 $\pm$ 0.30 & $-$1.03 $\pm$ 0.03 & $-$0.1 (fixed) & -\
\[Table:Fitted\_Parameters\]
In order to derive reliable thermal dust parameters, the unrelated contributions of the thermal (free-free) and non-thermal (synchrotron) emission components of the gas must be accurately determined. Unfortunately, the decomposition of the low-frequency radio continuum of galaxy flux density spectra is usually not clear-cut because of the degeneracy between the free-free and synchrotron contributions, especially since the intrinsic spectral index of the synchrotron emission is not known. This degeneracy specifically hampers the determination of any AME contribution to the observed emission spectrum and additional constraints are desirable. In the case of M33, such constraints exist. The sum of the directly measured H<span style="font-variant:small-caps;">ii</span> region flux densities corresponds to 235mJy at 10GHz [@Israel:74; @Israel:80]. Unfortunately, this does not include any contribution from the diffuse emission and the tally of H<span style="font-variant:small-caps;">ii</span> regions is incomplete, especially at large radii. It thus provides us only with a useful lower limit. However, a more accurate estimate of the total thermal radio emission may be obtained from the integrated H$\alpha$ emission ($I_\mathrm{H\alpha}$ = 3.6$\times$10$^{-13}$Wm$^{-2}$) measured by @Hoopes:00, after first correcting for global extinction. The M33 SED shown in Fig. \[Fig:M33\_Full\_SED\] exhibits a peak at optical wavelengths ($\sim$5$\times$10$^{5}$GHz) representing the directly observed integrated starlight, and another peak in the far-IR ($\sim$2$\times$10$^{3}$GHz) representing absorbed and re-emitted starlight. Since the second peak is less than the first one, a relatively minor fraction of all starlight is intercepted by dust, and the (small) global extinction can be estimated from the ratio of the peak fluxes. The M33 SED resembles those of the SMC and, in particular, the LMC [@Israel:10], with an optical luminosity exceeding the IR luminosity by a factor of $\sim$2.5. From the luminosity ratio of the far-IR and optical peaks in Fig. \[Fig:M33\_Full\_SED\], we estimate a visual extinction A$_{V}$ = 0.4 $\pm$ 0.1mag, of which 0.1mag is due to the Milky Way foreground [@DeVaucouleurs:91]. Assuming A$_\mathrm{H\alpha}$ = 0.81A$_{V}$, which is a typical Milky Way extinction curve [@Fitzpatrick:07], and at these wavelengths is very similar to typical SMC and LMC extinction curves [@Gordon:03], this corresponds to an H$\alpha$ extinction A$_\mathrm{H\alpha}$ = 0.24 $\pm$ 0.08mag *internal* to M33. Hence, the corrected H$\alpha$ flux is (4.5 $\pm$ 0.5) $\times$ 10$^{-13}$Wm$^{-2}$. Using
$$\frac{S^\mathrm{ff}_{\nu}}{I_\mathrm{H\alpha}} = 1.15\times10^{-14} [1-0.21\times\mathrm{log}\left(\frac{\nu}{\mathrm{GHz}}\right)]~\mathrm{Hz^{-1}} ,
\label{equ:S_Halpha}$$
we find that the free-free flux density at 10GHz is $S^\mathrm{ff}_\mathrm{10\,GHz}$ = 410 $\pm$ 45mJy. This is higher than the value of 280mJy inferred from the work by @Buczilowski:88, but in agreement with the value of 400mJy that follows from the determination by @Tabatabaei:07a [@Tabatabaei:07b]. Therefore, during the fitting process we constrained $C_\mathrm{ff}$ to be 410 $\pm$ 45mJy at 10GHz, and fixed $\alpha_\mathrm{ff}$ = $-$0.1. This is an important element in the decomposition of the observed continuum spectrum, essential for a reliable evaluation of the AME contribution in the 5–50GHz frequency range.
The full results of our fitting analysis are shown in Fig. \[Fig:M33\_Full\_CFD\], including the fitted parameters (which are also listed in Table \[Table:Fitted\_Parameters\]), the individual fitted components, the overall flux density spectrum fit, and the normalise residuals of the fits. It is clear that there is significant difference between the fit to the standard *Planck* data compared to the CMB-subtracted *Planck* data. Although there is a spread in the fitted $T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ values estimated from the CMB-subtracted data (blue, pink, red, and green curves in Fig. \[Fig:M33\_Full\_CFD\]), the fit to the standard data (black curve in Fig. \[Fig:M33\_Full\_CFD\]) is significantly outside this range. Focusing solely on the fits to the CMB-subtracted data, and combing these four fits, we find that the dust emission spectrum of M33 between $\sim$100GHz and 3THz is adequately described by a single modified blackbody curve, with a peak of $\sim$2000Jy, a mean dust temperature $T_\mathrm{dust}$ = 21.67 $\pm$ 0.30K, and a mean effective dust emissivity $\beta_\mathrm{eff}$ = 1.35 $\pm$ 0.10. There is also a warm dust component with a mean temperature of 61.89 $\pm$ 0.67K that was forced to have the same effective dust emissivity as the cold dust component. Since this warm component was simply included to insure an accurate fit to the peak of the cold dust component, we do not interpret this any further. The mean synchrotron radio continuum spectral index is $\alpha_\mathrm{sync}$ = $-$1.03 $\pm$ 0.03. The relevant mean values are also listed in Table \[Table:Fitted\_Parameters\]. For the individual entries we list the internal errors, whereas the errors given for the mean values reflect the rather larger dispersion of the individual values. Comparing the mean values to the standard values (i.e., comparing the first and last rows in Table \[Table:Fitted\_Parameters\]) we find that not correcting for the CMB contribution would result in a significant over-estimate of $T_\mathrm{dust}$ (by $\sim$5K) and under-estimate of $\beta_\mathrm{eff}$ (by $\sim$0.4), clearly highlighting the importance of correcting for the CMB.
We computed the mean fraction of thermal radio emission at 20cm and 3.6cm to be $\sim$16% and $\sim$49%, respectively, which are consistent with the estimates from @Tabatabaei:07b, confirming that our estimate of the thermal emission from the H$\alpha$ emission is reasonable. Not surprisingly, our estimated synchrotron spectral index is also consistent with the results obtained by @Tabatabaei:07b. Although our fitted free-free emission amplitudes match the limits of the estimate obtained from the H$\alpha$ observations, we confirmed that even without imposing this constraint on the free-free emission, we find consistent results, with a mean dust temperature of $T_\mathrm{dust}$ = 22.36 $\pm$ 0.69K and a mean spectral index of $\beta_\mathrm{eff}$ = 1.31 $\pm$ 0.10, but the fraction of free-free emission is decreased to $\sim$11% and $\sim$34% at 20cm and 3.6cm, respectively. Ignoring the radio data, and simply fitting the data at frequencies $\ge$100GHz, we find a slightly lower value of $\beta_\mathrm{eff}$ = 1.28 $\pm$ 0.10. The fact that these different approaches all yield consistent results confirms that our fit is not biased by the radio data, the decomposition between the thermal and non-thermal radio emission, nor the degeneracy between the free-free emission and the AME.
![SEDs for M33 for the five datasets (standard *Planck* data, along with the `SMICA`, `NILC`, `SEVEM`, and `Commander` CMB-subtracted data).[]{data-label="Fig:M33_Full_SED"}](Fig5.pdf)
We find that the AME in M33 is at best a minor component, both in an absolute sense and when compared to the free-free and synchrotron emission. Even though we modelled the AME using a spinning dust model for the warm ionised medium (as was used by @Planck_Intermediate_Results_XXV:15 for their M31 analysis), we obtained consistent results using the cold neutral medium, warm neutral medium, or molecular cloud spinning dust models. Based on observations of AME in the Milky Way, the ratio between the AME emission at 30GHz and the thermal dust emission at 100$\mu$m (often incorrectly referred to as an AME emissivity) is of the order of $\sim$2$\times$10$^{-4}$ [@Todorovic:10; @Planck_Intermediate_Results_XV:14]. Therefore, since we find a 100$\mu$m flux density of $\sim$1350Jy, this would lead us to expect $\sim$0.3Jy of AME at 30GHz, while we only estimate an AME flux density of $\lesssim$0.04Jy. Although this ratio between the AME and thermal dust emission is sensitive to dust temperature [as discussed by @Tibbs:12b], these results indicate that there is significantly less AME in M33 compared to our own Galaxy, which is consistent with what has been observed in M82, NGC253, and NGC4945 [@Peel:11]. On the other hand, in M31 the AME appears to be much more prominent, with a tentative detection that is comparable to what would be expected based on the AME level observed in our own Galaxy [@Planck_Intermediate_Results_XXV:15]. However, we note that for M31, the lack of observations between $\sim$1 and 20GHz could bias the fit, which is not the case for M33, M82, NGC253, and NGC4945, where the wavelength coverage is more complete.
Finally, we emphasise that a single dust temperature and a single effective emissivity index, i.e., a single curve, provides a reasonable fit to the observed data between $\sim$100GHz and 3THz. There is no spectral break, and there is no indication of an “excess” of any kind in the global (spatially integrated) flux density spectrum of M33.
![*Planck* 857GHz map convolved to 10arcmin angular resolution. Superimposed are the three independent elliptical apertures (white ellipses with semi-major axes, $a$ = 8, 5.33, and 2.67kpc) and the background annulus (green ellipses). The semi-major to semi-minor axis ratio was fixed to 10$^{0.23}$.[]{data-label="Fig:M33_PLCK857_10arcmin"}](Fig6_adobe.pdf)
 \
 \
 \
Radial variations {#Subsec:Annuli}
-----------------
In view of the radial changes in the brightness of the M33 disk, it is of interest to establish whether or not the flux density spectrum changes with radial distance from the center of M33. For this purpose, we now use the *Planck*, *Herschel*, and *IRAS* maps at the best common angular resolution (10arcmin), ignoring the three *Planck* LFI bands, which have beams larger than 10arcmin. Even so, the limited extent of M33 allows only three fully independent concentric ellipses to be constructed with semi-major axes of 8, 5.33, and 2.67kpc. As before, the background/foreground emission was estimated within an elliptical annulus with inner and outer semi-major axes of 1.15 and 1.50 times the semi-major axis of the 8kpc aperture, respectively (see Fig. \[Fig:M33\_PLCK857\_10arcmin\]).
@Tabatabaei:07a [@Tabatabaei:07b] have shown that across the disk of M33, the radial profiles of the far-IR and radio emission are very similar. Therefore, in addition to fitting the far-IR/sub-mm emission within each annulus, we also approximated the radio emission by scaling the curves from Fig. \[Fig:M33\_Full\_CFD\]. To do this, we computed the mean ratio between the 160$\mu$m flux density and the 4.8GHz flux density from the analysis in Section \[Subsec:Full\_SED\], and assuming that this is constant, we estimated the level of the radio emission within each annulus. We also assume that the fraction of free-free emission at 4.8GHz is fixed across M33, and fit the total radio emission with a synchrotron spectral index identical to the mean found for the entire galaxy. As before, we fit for synchrotron, free-free, AME, and thermal dust emissions. However, for this analysis we combine the four flux density measurements at each wavelength using the scatter as a measure of the uncertainty and perform a single fit, rather than a separate fit for each of the four CMB-subtracted maps. The resulting flux density spectra for each of the three elliptical annuli representing the inner, middle, and outer regions of M33 are displayed in Fig. \[Fig:M33\_radial\], along with the corresponding normalised residuals of the fits.
The radial dependence of both the dust temperature, $T_\mathrm{dust}$, and the effective emissivity, $\beta_\mathrm{eff}$, can be inferred from the plots, and it is clear that *both* dust temperature and the effective dust emissivity decrease with increasing radius. In the center of M33, $T_\mathrm{dust}$ = 22.36 $\pm$ 0.16K and $\beta_\mathrm{eff}$ = 1.53 $\pm$ 0.01. As the radius increases, both decrease to $T_\mathrm{dust}$ = 21.42 $\pm$ 0.16K and $\beta_\mathrm{eff}$ = 1.38 $\pm$ 0.01 in the middle annulus (semi-major axis between 2.67 and 5.33kpc) and to $T_\mathrm{dust}$ = 19.24 $\pm$ 0.15K and $\beta_\mathrm{eff}$ = 1.29 $\pm$ 0.02 in the outer annulus (semi-major axis between 5.33 and 8.00kpc). We emphasize that this result cannot be caused by the $T_\mathrm{dust}$, $\beta_\mathrm{eff}$ degeneracy referred to earlier as that would require $\beta_\mathrm{eff}$ to increase as $T_\mathrm{dust}$ decreases.
Our finding that both $T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ decrease with radius is qualitatively similar to the result obtained by @Tabatabaei:14, who used Monte Carlo simulations on more limited *Herschel* flux density spectra between 100 and 500$\mu$m to deduce a simultaneous decrease of dust temperature (from $\sim$24K to $\sim$18K) and emissivity (from $\sim$1.8 to $\sim$1.2) going from the center of M33 out to a radius of 6kpc. The more extended spectral coverage presented here allow quantitatively more robust results even though the spatial resolution is lower.
Discussion {#Sec:Discussion}
==========
The *global* continuum flux density spectrum of M33 is characterised by an overall emissivity $\beta_\mathrm{eff}$ = 1.35 $\pm$ 0.10, which is below the value of 1.5 estimated from combined *Herschel* and *Spitzer* observations down to 600GHz (500$\mu$m) by @Xilouris:12. The difference illustrates the bias introduced by the lack of low-frequency flux densities that most tightly constrain the Rayleigh-Jeans slope of the flux density spectrum. Even though we can fit the Rayleigh-Jeans part of the M33 global flux density spectrum with a single modified blackbody, it is a priori not likely that all of the dust in M33 radiates at a single temperature. However, a superposition of modified blackbodies representing grains with emissivities, $\beta_\mathrm{g}$, radiating at a range of temperatures may create a profile that is observationally hard to distinguish from a single-temperature modified blackbody profile with an apparent emissivity $\beta_\mathrm{eff}\leq\beta_\mathrm{g}$, especially when dust temperature and emissivity are negatively correlated as originally suggested by @Dupac:03 and @Desert:08, and later confirmed by @Planck_2013_Results_XI:14. Our analysis clearly shows that this is the case. The global flux density spectrum, whose Rayleigh-Jeans part is well-defined by a *single* modified blackbody is shown to be the sum of at least three different flux density spectra representing the inner, middle, and outer regions of M33, each with Rayleigh-Jeans sections equally well fitted by a single modified blackbody. As the number of sub-spectra is only limited by the available angular resolution, we expect that each of these in turn could be decomposed further.
Dust mass {#Subsec:Mass}
---------
Using
$$M_\mathrm{dust} = \frac{S_{\nu} d^{2}}{\kappa_{\nu} B_{\nu}(T_\mathrm{dust})} ,
\label{equ:Dust_Mass}$$
where $\kappa_{\nu}$ is the dust opacity, we estimated the global dust mass of M33, along with the dust mass in each of the three annuli. It is known that values of the dust opacity in the literature can vary by orders of magnitude [see @Clark:16], and in this work we adopt a value of $\kappa_{\nu}$ = 1.4m$^{2}$kg$^{-1}$ at 160$\mu$m taken from the “standard model” dust properties from @Galliano:11. Incorporating our results from Sections \[Subsec:Decomposition\] and \[Subsec:Annuli\] into Equation \[equ:Dust\_Mass\], we estimated a global dust mass for M33 of (2.3$\pm$0.4$)\times$10$^{6}$ M$_{\odot}$, and (0.8$\pm$0.1)$\times$10$^{6}$ M$_{\odot}$, (1.1$\pm$0.2)$\times$10$^{6}$ M$_{\odot}$, and (0.7$\pm$0.1)$\times$10$^{6}$ M$_{\odot}$ for the inner, middle, and outer regions of M33, respectively. We find that our global dust mass estimated assuming a single modified blackbody is consistent, within the uncertainties, with the sum of the three dust masses estimated for the sub-regions, suggesting that fitting the entirety of M33 is degenerate with fitting the three sub-regions.
Local Group sample {#Subsec:Local_Group}
------------------
The global dust emissivity, $\beta_\mathrm{eff}$ = 1.35 $\pm$ 0.10, for M33 may also be compared to those derived from *Planck* observations of other Local Group galaxies: $\beta_{eff}$ = 1.62 $\pm$ 0.10, 1.62 $\pm$ 0.11, 1.48 $\pm$ 0.25, and 1.21 $\pm$ 0.27 for the Milky Way [@Planck_2013_Results_XI:14], M31 [@Planck_Intermediate_Results_XXV:15], the LMC [@Planck_Early_Results_XVII:11], and the SMC [@Planck_Early_Results_XVII:11], respectively. We find that the M33 emissivity is significantly lower than that observed in the Milky Way and M31, and is more consistent with the values found in the Magellanic Clouds, with M33 actually falling between the LMC and the SMC values. Interestingly, these dust emissivities closely follow the mean metallicities of the Local Group galaxies: 12+log\[O/H\] = 8.32 $\pm$ 0.16, 8.67 $\pm$ 0.04, 8.72 $\pm$ 0.19, 8.43 $\pm$ 0.05, and 8.11 $\pm$ 0.03, for M33, the Milky Way, M31, the LMC, and the SMC, respectively [@Pagel:03; @Toribio:16]. To illustrate this, in Fig. \[Fig:M33\_beta\_z\] we plot the dust emissivity as a function of metallicity for these 5 galaxies (filled symbols), which clearly shows that the dust emissivity increases with increasing metallicity. This trend is also observed across the M33 disk itself, where our observed emissivity gradient follows the metallicity gradient [@Toribio:16], as can be seen when we plot the results from our three annuli within M33 (open squares) in Fig. \[Fig:M33\_beta\_z\].
$T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ radial variations {#Subsec:Radial_Variations}
------------------------------------------------------------
The apparent decrease in both $T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ with increasing M33 radius was discussed in some detail by @Tabatabaei:14. Without fully subscribing to their conclusion, we note that there are, in principle, two possible physical explanations for the observed radial decreases. The first involves dust grain composition and di-electric properties. For instance, the dust emissivity may decrease with the average interstellar energy density. Mechanical and radiative erosion of dust grains should be stronger in the more energetic inner regions than in the more quiescent outer regions. This would favour more delicate carbon/ice dust grains in the outer regions and more robust silicate-rich grains in the inner regions. The intrinsic dust grain composition may also undergo radial changes following radial gradients in the population of stellar dust producers. The second explanation involves large-scale dust cloud properties. Dust cloud heating and effective emissivity may decrease with the average radiation field, more specifically the mix of dust cloud temperatures within a specific temperature range many change as a function of irradiation. For instance, consider the possibility that each of the profiles in Fig. \[Fig:M33\_radial\] actually represent a collection of *dust clouds and filaments* with identical emissivities but different temperatures. In a radially decreasing *average* radiation field, clouds with temperatures at the high end would occur less frequently and have a smaller filling factor, resulting in a more skewed composite profile with a downward shift of apparent mean temperature and a consequent flattening of the Rayleigh-Jeans slope. However, the results presented in this analysis do not allow us to distinguish between these possibilities.
![The global effective emissivity index, $\beta_\mathrm{eff}$, of Local Group galaxies as a function of metallicity, 12+log\[O/H\]. Filled symbols mark the global values of the identified galaxies, while the open squares mark the three independent regions (inner, middle, and outer) within M33.[]{data-label="Fig:M33_beta_z"}](Fig8.pdf)
Comparison to previous studies {#Subsec:Hermelo}
------------------------------
Our analysis is not the first to produce a full flux density spectrum of M33, as @Hermelo:16 used the *Planck* 2013 “nominal” mission data along with a single CMB-subtraction method (`SMICA`) to derive the full flux density spectrum for M33. Using complex models [@Groves:08; @Popescu:11], they fitted their M33 flux density spectrum, deriving an excess of emission at mm/sub-mm wavelengths. However, in this analysis, we not only use the most recent *Planck* 2015 “full” mission data, but we also incorporate and evaluate four different CMB-subtraction techniques. As we have discussed in some depth, the contribution from the CMB fluctuations is significant and must be accurately accounted for. From Fig. \[Fig:M33\_Full\_CFD\] and Table \[Table:Fitted\_Parameters\] it is clear that there is scatter in both $T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ between each of the four CMB-subtraction methods, highlighting the dangers of adopting a single method.
In this analysis, we have chosen to fit our M33 flux density spectrum with a relatively simple model (i.e., a modified blackbody) and find no indication of any emission excess. A major advantage of fitting a modified blackbody to the data, rather than a complex model, is that this approach has been adopted by previous *Planck* analyses [e.g., @Planck_Early_Results_XVII:11; @Planck_2013_Results_XI:14; @Planck_Intermediate_Results_XXV:15], allowing direct comparisons to be made with other galaxies.
Conclusions {#Sec:Conclusions}
===========
We have performed a comprehensive analysis of the global continuum flux density spectrum of M33 over a very large wavelength range from radio to UV wavelengths. In the course of this analysis, we have demonstrated the importance of accurately accounting for the contribution of CMB fluctuations to the flux density spectrum, which if neglected, results in an over-estimate of $T_\mathrm{dust}$ of $\sim$5K and and under-estimate of $\beta_\mathrm{eff}$ of $\sim$0.4.
Surprisingly, we find that the global integrated emission of M33 between $\sim$100GHz and 3THz is adequately described by a single modified blackbody curve, with a mean dust temperature $T_{dust}$ = 21.67 $\pm$ 0.30K and a mean effective dust emissivity $\beta_{eff}$ = 1.35 $\pm$ 0.10, even though such constancy of emission over all of the galaxy and throughout the line of sight is physically unlikely. In order to investigate this, we split M33 into the three independent annuli that the available resolution allows. We find that *both* $T_\mathrm{dust}$ and $\beta_\mathrm{eff}$ decrease from the centre to the outskirts of M33. This correlation is not due to any observational effect and it confirms in a direct manner an earlier conclusion reached by @Tabatabaei:14. The dust emission spectrum between $\sim$100GHz and 3THz of each of the three sub-regions can be fitted with a single (but not identical) modified blackbody curve, and as the sum of the three curves (the global emission curve) itself is well-fitted with a single modified blackbody curve, we conclude that the combination of individual flux density spectra representing different parts of M33 is highly degenerate with the mean flux density spectrum.
Comparing the global far-IR emission of M33 to that of the other Local Group galaxies for which coverage is available, we find that M33 resembles the Magellanic Clouds rather than the larger spirals, the Milky Way and M31. Within this limited sample, there is a good correlation between the observed Rayleigh-Jeans slope, $\beta_\mathrm{eff}$, and the metallicity of these galaxies, with $\beta_\mathrm{eff}$ increasing with increasing metallicity. This global correlation for the Local Group galaxies is further strengthened by the finding that it also holds *within* M33. The internal M33 $\beta_\mathrm{eff}$ gradient follows its metallicity gradient, with the inner part of M33 much like the LMC, and the outer part much like the lower-metallicity SMC.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous referee for providing detailed comments that have improved the content of this paper. MWP acknowledges grant \#2015/19936-1, São Paulo Research Foundation (FAPESP).
Ali-Ha[ï]{}moud, Y., Hirata, C. M., & Dickinson, C. 2009, , 395, 1055
Bot, C., Ysard, N., Paradis, D., et al. 2010, , 523, A20
Boquien, M., Calzetti, D., Combes, F., et al. 2011, , 142, 111
Boquien, M., Calzetti, D., Aalto, S., et al. 2015, , 578, A8
Buczilowski, U. R. 1988, , 205, 29
Casassus, S., Cabrera, G. F., F[ö]{}rster, F., et al. 2006, , 639, 951
Clark, C. J. R., Schofield, S. P., Gomez, H. L., & Davies, J. I. 2016, , 459, 1646
D[é]{}sert, F.-X., Mac[í]{}as-P[é]{}rez, J. F., Mayet, F., et al. 2008, , 481, 411
de Vaucouleurs, G., de Vaucouleurs, A., Corwin, H. G., Jr., et al. 1991, Third Reference Catalogue of Bright Galaxies. Volume I: Explanations and references. Volume II: Data for galaxies between 0$^{h}$ and 12$^{h}$. Volume III: Data for galaxies between 12$^{h}$ and 24$^{h}$., by de Vaucouleurs, G.; de Vaucouleurs, A.; Corwin, H. G., Jr.; Buta, R. J.; Paturel, G.; Fouqu[é]{}, P.. Springer, New York, NY (USA), 1991, 2091 p., ISBN 0-387-97552-7, Price US\$ 198.00. ISBN 3-540-97552-7, Price DM 448.00. ISBN 0-387-97549-7 (Vol. I), ISBN 0-387-97550-0 (Vol. II), ISBN 0-387-97551-9 (Vol. III)., I,
Draine, B. T., & Lazarian, A. 1998, , 508, 157
Draine, B. T., & Hensley, B. 2012, , 757, 103
Druard, C., Braine, J., Schuster, K. F., et al. 2014, , 567, A118
Dupac, X., Bernard, J.-P., Boudet, N., et al. 2003, , 404, L11
Fitzpatrick, E. L., & Massa, D. 2007, , 663, 320
Freedman, W. L., Wilson, C. D., & Madore, B. F. 1991, , 372, 455
Galliano, F., Madden, S. C., Jones, A. P., Wilson, C. D., & Bernard, J.-P. 2005, , 434, 867
Galliano, F., Hony, S., Bernard, J.-P., et al. 2011, , 536, A88
Gordon, K. D., Clayton, G. C., Misselt, K. A., Landolt, A. U., & Wolff, M. J. 2003, , 594, 279
G[ó]{}rski, K. M., Hivon, E., Banday, A. J., et al. 2005, , 622, 759
Groves, B., Dopita, M. A., Sutherland, R. S., et al. 2008, , 176, 438-456
Hermelo, I., Rela[ñ]{}o, M., Lisenfeld, U., et al. 2016, , 590, A56
Heyer, M. H., Corbelli, E., Schneider, S. E., & Young, J. S. 2004, , 602, 723
Hoopes, C. G., & Walterbos, R. A. M. 2000, , 541, 597
Israel, F. P. 1980, , 90, 246
Israel, F. P., Mahoney, M. J., & Howarth, N. 1992, , 261, 47
Israel, F. P., & van der Kruit, P. C. 1974, , 32, 363
Israel, F. P., Wall, W. F., Raban, D., et al. 2010, , 519, A67
Jarrett, T. H., Chester, T., Cutri, R., Schneider, S. E., & Huchra, J. P. 2003, , 125, 525
Juvela, M., & Ysard, N. 2012a, , 539, A71
Juvela, M., & Ysard, N. 2012b, , 541, A33
Kramer, C., Buchbender, C., Xilouris, E. M., et al. 2010, , 518, L67
Laher, R. R., Gorjian, V., Rebull, L. M., et al. 2012, , 124, 737
Lee, J. C., Gil de Paz, A., Kennicutt, R. C., Jr., et al. 2011, , 192, 6
Markwardt, C. B. 2009, Astronomical Data Analysis Software and Systems XVIII, 411, 251
Miville-Desch[ê]{}nes, M.-A., & Lagache, G. 2005, , 157, 302
Pagel, B. E. J. 2003, CNO in the Universe, 304, 187
Paradis, D., Dobashi, K., Shimoikura, T., et al. 2012, , 543, A103
Paturel, G., Petit, C., Prugniel, P., et al. 2003, , 412, 45
Peel, M. W., Dickinson, C., Davies, R. D., Clements, D. L., & Beswick, R. J. 2011, , 416, L99
Planck Collaboration I. 2011, , 536, A1
Planck Collaboration XVI. 2011, , 536, A16
Planck Collaboration XVII. 2011, , 536, A17
Planck Collaboration XXV. 2015, , 582, A28
Planck Collaboration XV. 2014, , 565, A103
Planck Collaboration IX. 2014, , 571, A9
Planck Collaboration XI. 2014, , 571, A11
Planck Collaboration XIII. 2014, , 571, A13
Planck Collaboration XXVIII. 2014, , 571, A28
Planck Collaboration I. 2016, , 594, A1
Planck Collaboration II. 2016, , 594, A2
Planck Collaboration III. 2016, , 594, A3
Planck Collaboration IV. 2016, , 594, A4
Planck Collaboration V. 2016, , 594, A5
Planck Collaboration VI. 2016, , 594, A6
Planck Collaboration VII. 2016, , 594, A7
Planck Collaboration VIII. 2016, , 594, A8
Planck Collaboration IX. 2016, , 594, A9
Planck Collaboration X. 2016, , 594, A10
Popescu, C. C., Tuffs, R. J., Dopita, M. A., et al. 2011, , 527, A109
Rice, W., Boulanger, F., Viallefond, F., Soifer, B. T., & Freedman, W. L. 1990, , 358, 418
Shetty, R., Kauffmann, J., Schnee, S., & Goodman, A. A. 2009, , 696, 676
Silsbee, K., Ali-Ha[ï]{}moud, Y., & Hirata, C. M. 2011, , 411, 2750
Tabatabaei, F. S., Krause, M., & Beck, R. 2007a, , 472, 785
Tabatabaei, F. S., Beck, R., Kr[ü]{}gel, E., et al. 2007b, , 475, 133
Tabatabaei, F. S., Braine, J., Xilouris, E. M., et al. 2014, , 561, A95
Tauber, J. A., Mandolesi, N., Puget, J.-L., et al. 2010, , 520, A1
Tibbs, C. T., Watson, R. A., Dickinson, C., et al. 2010, , 402, 1969
Tibbs, C. T., Flagey, N., Paladini, R., et al. 2011, , 418, 1889
Tibbs, C. T., Paladini, R., & Dickinson, C. 2012, Advances in Astronomy, 2012, 124931
Tibbs, C. T., Scaife, A. M. M., Dickinson, C., et al. 2013, , 768, 98
Todorovi[ć]{}, M., Davies, R. D., Dickinson, C., et al. 2010, , 406, 1629
Toribio San Cipriano, L., Garc[í]{}a-Rojas, J., Esteban, C., Bresolin, F., & Peimbert, M. 2016, , 458, 1866
Verley, S., Hunt, L. K., Corbelli, E., & Giovanardi, C. 2007, , 476, 1161
Verley, S., Corbelli, E., Giovanardi, C., & Hunt, L. K. 2009, , 493, 453
Xilouris, E. M., Tabatabaei, F. S., Boquien, M., et al. 2012, , 543, A74
\[lastpage\]
[^1]: ESA Research Fellow
[^2]: E-mail: [email protected]
[^3]: An SED is a plot of energy as a function of frequency or wavelength, i.e., $\nu S_{\nu}$ vs $\nu$, or $\lambda S_\lambda$ vs $\lambda$, while a flux density spectrum is a plot of flux density as a function of frequency or wavelength i.e., $S_{\nu}$ vs $\nu$, or $S_{\lambda}$ vs $\lambda$.
[^4]: http://pla.esac.esa.int/pla/
[^5]: http://archives.esac.esa.int/hsa/whsa/
[^6]: http://sha.ipac.caltech.edu/applications/Spitzer/SHA/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper studies the role of dg-Lie algebroids in derived deformation theory. More precisely, we provide an equivalence between the homotopy theories of formal moduli problems and dg-Lie algebroids over a commutative dg-algebra of characteristic zero. At the level of linear objects, we show that the category of representations of a dg-Lie algebroid is an extension of the category of quasi-coherent sheaves on the corresponding formal moduli problem. We describe this extension geometrically in terms of pro-coherent sheaves.'
address: |
Mathematical Institute\
Utrecht University\
P.O. Box 80010\
3508 TA Utrecht\
The Netherlands.
author:
- Joost Nuiten
bibliography:
- '../bibliography\_thesis.bib'
title: Koszul duality for Lie algebroids
---
Introduction {#sec:intro}
============
Lie algebroids appear throughout algebraic and differential geometry as natural objects describing infinitesimal geometric structures, like foliations and actions of Lie algebras [@rin63; @mac87]. In this paper, we study the relation between Lie algebroids and deformation theory, which can informally be described as follows: suppose that $k$ has characteristic zero and that $$\xymatrix{
x\colon {{\mathrm{Spec}}}(A)\ar[r] & X
}$$ is a map from an affine (derived) scheme to a moduli space over $k$. Then a formal neighbourhood of $X$ around $x$ is controlled by a Lie algebroid over ${{\mathrm{Spec}}}(A)$. This Lie algebroid is given by the vector fields on ${{\mathrm{Spec}}}(A)$ that are (derived) tangent to the fibers of $x$.
When ${{\mathrm{Spec}}}(A)$ is a point, this recovers the well-known relation between deformation problems and dg-Lie algebras, which has been extensively studied and applied [@gol88; @kon03; @man02], and has been given a precise mathematical formulation in the work of Hinich [@hin98], Pridham [@pri10] and Lurie [@lur11X]. An important idea in these works is to describe the formal neighbourhood of a moduli space in terms of derived geometry, using its functor of points. More precisely, one can describe a formal neighbourhood of $X$ around $x$ by a functor $$\xymatrix{
X^\wedge\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A\ar[r] & {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}; \hspace{4pt} B\ar@{|->}[r] & X(B)\times_{X(A)} \{x\}
}$$ sending a *derived infinitesimal thickening* ${{\mathrm{Spec}}}(A){\rightarrow}{{\mathrm{Spec}}}(B)$ to the space of extensions of $x$ to a map $\tilde{x}\colon {{\mathrm{Spec}}}(B){\rightarrow}X$. For any reasonable moduli space $X$, the functor $X^\wedge$ satisfies a derived version of the Schlessinger conditions, which encodes the usual obstruction theory for existence and uniqueness of deformations [@sch68]. The works mentioned above provide an equivalence between the homotopy theory of such functors satisfying the Schlessinger conditions and the homotopy theory of dg-Lie algebras, at least when $A$ is a field (see also [@hen15] for an extension to dg-Lie algebras over more general dg-algebras $A$).
The purpose of this paper is to provide a similar identification of the homotopy theory of dg-Lie algebroids with the homotopy theory of *formal moduli problems*, in the sense of [@lur11X; @gai16]:
\[def:formalmoduli\] Let $A$ be a connective commutative $k$-algebra, where $k$ has characteristic zero. The $\infty$-category ${
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}_k^{\mathrm{sm}}/A$ of *small extensions* of $A$ is the smallest subcategory of commutative $k$-algebras over $A$ that contains $A$ and that is closed under square zero extensions by $A[n]$ with $n\geq 0$.
A *formal moduli problem* over $A$ is a functor $$\xymatrix{
X\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A\ar[r] & {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}}$$ satisfying the following two conditions:
- $X(A)\simeq \ast$ is contractible.
- $X$ preserves any pullback diagram of small extensions of $A$ of the form $$\label{diag:squarezerointro}\vcenter{\xymatrix{
B_\eta\ar[d]\ar[r] & A\ar[d]^{({\mathrm{id}}, 0)}\\
B\ar[r]_-\eta & A\oplus A[n+1].
}}$$ Such a pullback square realizes $B_\eta$ as a square zero extension of $B$ by $A[n]$ (for $n\geq 0$).
Geometrically, a formal moduli problem can be thought of as a map of stacks $x\colon {{\mathrm{Spec}}}(A){\rightarrow}X$ that realizes $X$ as an infinitesimal thickening of ${{\mathrm{Spec}}}(A)$. We show that such a stack determines a Lie algebroid $T_{A/X}$, which can be thought of as the fiberwise vector fields on ${{\mathrm{Spec}}}(A)$ over $X$. This construction is part of an equivalence, with inverse sending a Lie algebroid ${\mathfrak{g}}$ to the ‘quotient’ of ${{\mathrm{Spec}}}(A)$ by the infinitesimal ${\mathfrak{g}}$-action:
\[thm:intro1\] Suppose that $A$ is eventually coconnective. Then there is an equivalence of $\infty$-categories $$\xymatrix{
{\mathrm{MC}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\ar@<1ex>[r] & {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A\ar@<1ex>[l] \colon T_{A/}
}$$ between the $\infty$-category of Lie algebroids over $A$ and the $\infty$-category of formal moduli problems over $A$.
The $\infty$-category of Lie algebroids has an explicit description in terms of homological algebra, as the localization of the category of dg-Lie algebroids over $A$ at the quasi-isomorphisms. One can therefore think of the above result as a rectification result, showing that any formal moduli problem admits a rigid description in terms of chain complexes endowed with algebraic structure. For example, the formal moduli problem describing the deformations of a connective commutative $A$-algebra $R$ admits a simple chain-level description, as the dg-Lie algebroid of compatible derivations on both $A$ and $R$ (Proposition \[prop:defofalgebras\], cf. [@hin04; @toe16]).
The proof of Theorem \[thm:intro1\] follows the lines of [@lur11X], and relies on a version of Koszul duality: the small extensions of $A$ are Koszul dual to certain free Lie algebroids over $A$, by means of the functor sending a dg-Lie algebroid to its cohomology (as studied already in [@rin63]). To make efficient use of this result, it is useful to study Koszul duality at the level of *linear* objects as well. More precisely, given a quasi-coherent sheaf ${\mathcal{F}}$ on $X$, one expects the restriction $x^*{\mathcal{F}}$ along $x\colon {{\mathrm{Spec}}}(A){\rightarrow}X$ to carry a natural representation of the Lie algebroid $T_{A/X}$. Indeed, we prove the following result:
\[thm:intro2\] Let $A$ be eventually coconnective and let $X$ be a formal moduli problem over $A$ with associated Lie algebroid $T_{A/X}$. Then there is a fully faithful, symmetric monoidal left adjoint functor $$\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}(X)\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}
}$$ which induces an equivalence on connective objects.
It is well-known (cf. [@sai89; @kap91]) that there are convergence issues that prevent the functor $\Psi_X$ from being an equivalence. For example, for any map $f\colon X{\rightarrow}Y$ of formal moduli problems, the restriction functor $$\xymatrix{
f^!\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/Y}}\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}
}$$ preserves all colimits, while the restriction functor $f^*\colon {{\mathrm{QC}}}(Y){\rightarrow}{{\mathrm{QC}}}(X)$ does not. For this reason (among many others), a more refined geometric theory of sheaves has been introduced in [@gai13], in which quasi-coherent sheaves are replaced by *ind-coherent sheaves*. The work of Gaitsgory and Rozenblyum [@gai16] provides an extensive study of derived algebraic geometry in terms of ind-coherent sheaves. In particular, they *define* and study Lie algebroids purely in terms of formal moduli problems. More precisely, they consider a slight enlargement of the $\infty$-category of small extension of $A$, which also contains small extensions of $A$ by *coherent* connective $A$-modules. A Lie algebroid is then defined to be a functor $$\xymatrix{
X\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{sm, coh}}_k/A\ar[r] & {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}}$$ on this category satisfying conditions (a) and (b) of Definition \[def:formalmoduli\]. We will refer to such an object as a *pro-coherent formal moduli problem*.
In certain (somewhat restricted) situations, we show that the rectification results from Theorem \[thm:intro1\] and Theorem \[thm:intro2\] can also be applied to such pro-coherent formal moduli problems. In particular, this shows that pro-coherent formal moduli problems can indeed be interpreted at the point-set level as Lie algebroids. To obtain a simple point-set model for Lie algebroids in the ind-coherent setting, it will be more convenient to work with the $\infty$-category ${\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}})$ of *pro-coherent sheaves* on $A$. This $\infty$-category is often equivalent to the $\infty$-category of ind-coherent sheaves by Serre duality [@gai13 Section 9], but has the advantage of admitting an explicit presentation by a certain ‘tame’ model structure on dg-$A$-modules [@bec14].
This tame model structure can be transferred to a model structure on dg-Lie algebroids over $A$. Unfortunately, this only presents the correct $\infty$-category of Lie algebroids in ${\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}})$ when $A$ satisfies some rather strict technical assumptions at the chain level (see Warning \[war:boundecoh\]). As an important example, these technical conditions are met when $k$ is a coherent and eventually coconnective and when $${{\mathrm{Spec}}}(A) = {\mathbb{A}}_k^n\times_{{\mathbb{A}}_k^m}\{0\}.$$ In particular, for $A=k$ we obtain a simple point-set model for the $\infty$-category of pro-coherent Lie algebras over $A$.
In the cases where we have a good point-set model for Lie algebroids in ${\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}})$, we show (Theorem \[thm:formalmodulicoherent\]) that there is an equivalence $$\xymatrix{
T_{A/}\colon {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A^!\ar[r]^-\sim & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!
}$$ between the $\infty$-category of pro-coherent formal moduli problems over $A$ and the $\infty$-category of pro-coherent Lie algebroids over $A$. Theorem \[thm:intro2\] can then be refined (Theorem \[fm:thm:ind-coherentvsrepresentations\]) to an *equivalence* $$\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}^!(X)\ar[r]^-\sim & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}^!
}$$ between pro-coherent sheaves on $X$ and pro-coherent $T_{A/X}$-representations.
Outline {#outline .unnumbered}
-------
The paper is outlined as follows. In Section \[sec:dglie\], we recall the basic homotopy theory of dg-Lie algebroids over a commutative dg-algebra $A$, in which the weak equivalences are the quasi-isomorphisms. In Section \[sec:procoherent\] we provide model categorical descriptions of the homotopy theories of pro-coherent sheaves and pro-coherent Lie algebroids over $A$. Theorem \[thm:intro1\] and its analogue in the coherent case are proven in Section \[sec:formmod\], based on results about Lie algebroid cohomology that are discussed in Section \[sec:ce\] and \[sec:koszulduality\].
Section \[sec:representations\] is devoted to a proof of Theorem \[thm:intro2\]. As an application, we also show how Theorem \[thm:intro2\] can be used to provide a simple point-set model for the Lie algebroid classifying the deformations of a (connective) commutative algebra over $A$ (Proposition \[prop:defofalgebras\]). Finally, in Section \[fm:sec:indcoh\] we prove the equivalence of pro-coherent sheaves on formal moduli problems and representations of their associated Lie algebroid.
Conventions {#conventions .unnumbered}
-----------
Throughout, let ${\mathbb{Q}}\subseteq k$ be a fixed connective commutative dg-algebra of characteristic zero and let $A$ be a connective cdga over $k$. All differential-graded objects are homologically graded, so that connective objects are concentrated in non-negative degrees. Given a chain complex $V$, we denote its suspension and cone by $V[1]$ and $V[0, 1]$.
Acknowledgement {#acknowledgement .unnumbered}
---------------
The author was supported by NWO.
Recollections on DG-Lie algebroids {#sec:dglie}
==================================
In this section we recall the homotopy theory of dg-Lie algebroids over a commutative dg-algebra, based on the discussion in [@nui17b].
DG-Lie algebroids
-----------------
Recall that the tangent module of a commutative dg-$k$-algebra $A$ is the dg-$A$-module of $k$-linear derivations of $A$ $$T_A = {\mathrm{Der}}_k(A, A).$$ The commutator bracket endows this complex with the structure of a dg-Lie-algebra over $k$.
\[la:def:lralgebra\] A *dg-Lie algebroid* ${\mathfrak{g}}$ over $A$ (relative to $k$) is an unbounded dg-$A$-module ${\mathfrak{g}}$, equipped with a $k$-linear dg-Lie algebra structure and an *anchor map* $\rho\colon {\mathfrak{g}}{\rightarrow}T_A$ such that
1. $\rho$ is both a map of dg-$A$-modules and dg-Lie algebras.
2. the failure of the Lie bracket to be $A$-bilinear is governed by the Leibniz rule $$[X, a\cdot Y] = (-1)^{Xa} a[X, Y] + \rho(X)(a)\cdot Y.$$
Let ${{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A$ be the category of dg-Lie algebroids over $A$, with maps given by $A$-linear maps over $T_A$ that preserve the Lie bracket.
\[ex:atiyah\] Any dg-$A$-module $E$ gives rise to an *Atiyah dg-Lie algebroid* ${\mathrm{At}}(E)$ over $A$, which can be described as follows: a degree $n$ element of ${\mathrm{At}}(E)$ is a tuple $(v, \nabla_v)$ consisting of a derivation $v\colon A{\rightarrow}A$ (of degree $n$), together with a $k$-linear map $\nabla_v\colon E{\rightarrow}E$ (of degree $n$) such that $$\nabla_v(a\cdot e) = v(a)\cdot e + (-1)^{|a|\cdot n} a\cdot \nabla_v(e)$$ for all $a\in A$ and $e\in E$. This becomes a dg-$A$-module under pointwise multiplication and a dg-Lie algebra under the commutator bracket. The anchor map is the obvious projection ${\mathrm{At}}(E){\rightarrow}T_A$ sending $(v, \nabla_v)$ to $v$.
\[ex:atiyah2\] Similarly, suppose that $E\in {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A$ has the structure of an algebra over a $k$-linear dg-operad ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$. Then there is a sub dg-Lie algebroid ${\mathrm{At}}_{{
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}}(E)\subseteq {\mathrm{At}}(E)$ consisting of the tuples $(v, \nabla_v)$ where $\nabla_v$ is a ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$-algebra derivation.
The following is the main result of [@nui17b]:
\[prop:monadicity\] The category of dg-Lie algebroids over $A$ carries the *projective* semi-model structure, in which a map is a weak equivalence (fibration) if and only if it is a quasi-isomorphism (degreewise surjective). Furthermore, the forgetful functor to the projective model structure on dg-$A$-modules $$\smash{\xymatrix{
{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\ar[r] & {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A/T_A
}}$$ is a right Quillen functor that preserves all sifted homotopy colimits.
\[rem:equivalenttoquillen\] Let $0{\rightarrow}{\mathfrak{g}}{\rightarrow}T_A$ be a fibrant-cofibrant replacement of the initial dg-Lie algebroid over $A$. Then there is a Quillen equivalence $$\xymatrix{
{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\ar@<1ex>[r] & {\mathfrak{g}}/{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A .\ar@<1ex>[l]
}$$ Using the simplicial structure from [@vez13] and the fact that every object in ${\mathfrak{g}}/{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A$ is fibrant, one finds that ${\mathfrak{g}}/{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A$ is a genuine (combinatorial) model category.
\[def:liealgebroids\] Let $A$ be a cofibrant connective cdga over $k$. We define the *$\infty$-category of Lie algebroids over $A$* to be the $\infty$-categorical localization $${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A := {{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\Big[\{\text{quasi-iso}\}^{-1}\Big].$$ This is a locally presentable $\infty$-category since ${{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A$ is Quillen equivalent to a combinatorial model category (Remark \[rem:equivalenttoquillen\]).
\[rem:derivedtangent\] The condition that $A$ is cofibrant over $k$ guarantees that the tangent module $T_A$ has the correct homotopy type.
\[def:Q\] We will say that a dg-Lie algebroid ${\mathfrak{g}}$ is *$A$-cofibrant* when it is cofibrant as a dg-$A$-module. Every cofibrant dg-Lie algebroid is $A$-cofibrant. Conversely, if ${\mathfrak{g}}$ is $A$-cofibrant, then it has an explicit cofibrant replacement $Q({\mathfrak{g}}){\stackrel{\sim}{{\rightarrow}}} {\mathfrak{g}}$, described as follows [@nui17b Section 5]: without differential, $Q({\mathfrak{g}})$ is freely generated by the $A$-linear map $$\xymatrix{
\Big({\mathrm{Sym}}^{{{\scriptscriptstyle \geq}}1}_A{\mathfrak{g}}[1]\Big)[-1]\ar[r]^-{\pi} & {\mathfrak{g}}\ar[r]^-\rho & T_A
}$$ where $\pi$ is the obvious projection. As a dg-Lie algebra, $Q({\mathfrak{g}})$ is a quotient of the $A$-linear extension usual [@lod12] operadic bar-cobar resolution $\Omega B{\mathfrak{g}}$ of the dg-Lie algebra underlying ${\mathfrak{g}}$; this determines the differential.
Representations
---------------
Recall that a *representation* of a dg-Lie algebroid ${\mathfrak{g}}$ is a dg-$A$-module $E$, together with a Lie algebra representation $\nabla\colon {\mathfrak{g}}\otimes_k E{\rightarrow}E$ such that (without Koszul signs) $$\label{eq:reps}
\nabla_{a\cdot X}(e) = a\cdot \nabla_X(e) \qquad \qquad \nabla_X(a\cdot e) = a\cdot \nabla_X(e) + \rho(X)(a)\cdot e$$ for all $a\in A, X\in{\mathfrak{g}}$ and $e\in E$. Equivalently, a ${\mathfrak{g}}$-representation on $E$ is a map $$\xymatrix{
{\mathfrak{g}}\ar[r] & {\mathrm{At}}(E)
}$$ to the Atiyah Lie algebroid of $E$ (Example \[ex:atiyah\]). Such representations can be organized into a symmetric monoidal category ${{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}$, with tensor product given by $E\otimes_A F$, endowed with the ${\mathfrak{g}}$-representation $$\nabla_X(e\otimes f) = \nabla_X(e)\otimes f + e\otimes \nabla_X(f).$$ The internal hom is given by ${{\mathrm{Hom}}}_A(E, F)$, endowed with the ${\mathfrak{g}}$-representation by conjugation.
\[ex:algebrainrep\] Let ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$ be a dg-operad. Then a ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$-algebra in ${{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}$ is simply a ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$-algebra in $\smash{{{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A}$, equipped with a ${\mathfrak{g}}$-representation acting by ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$-algebra derivations. Equivalently, such a ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$-algebra structure on a ${\mathfrak{g}}$-representation $E$ is determined by a map of dg-Lie algebroids ${\mathfrak{g}}{\rightarrow}{\mathrm{At}}_{{
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}}(E)$ (Example \[ex:atiyah2\]).
One can identify ${{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}$ with the category of left modules over the *universal enveloping algebra* ${\mathcal{U}}({\mathfrak{g}})$ of ${\mathfrak{g}}$. In particular, it carries a model structure transferred from the projective model structure on dg-$A$-modules.
\[lem:pbw\] Any map between dg-Lie algebroids $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$ induces a Quillen adjunction $$\xymatrix{
f_*\colon {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}\ar@<1ex>[r] & {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{h}}\ar@<1ex>[l] \colon f^!
}$$ where $f^!$ restricts a representation along $f$. When ${\mathfrak{g}}$ and ${\mathfrak{h}}$ are $A$-cofibrant and $f$ is a weak equivalence, this Quillen adjunction is a Quillen equivalence.
The map $f$ induces a map of universal enveloping algebras ${\mathcal{U}}({\mathfrak{g}}){\rightarrow}{\mathcal{U}}({\mathfrak{h}})$, which gives rise to the Quillen pair $(f_*, f^!)$. The left adjoint $f_*$ is given by the functor $E\mapsto {\mathcal{U}}({\mathfrak{h}})\otimes_{{\mathcal{U}}({\mathfrak{g}})} E$.
For the second part, recall that the universal enveloping algebra ${\mathcal{U}}({\mathfrak{g}})$ has a natural PBW filtration, obtained by declaring generators from $A$ to have weight $0$ and generators from ${\mathfrak{g}}$ to have weight $1$ [@rin63]. The associated graded is a graded cdga and comes equipped with a surjective map of graded cdgas $$\xymatrix{
{\mathrm{Sym}}_A {\mathfrak{g}}\ar[r] & {\mathrm{gr}}\big({\mathcal{U}}({\mathfrak{g}})\big).
}$$ When ${\mathfrak{g}}$ is projective as a graded $A$-module, this map (or rather, the underlying map of graded graded-commutative algebras) is an isomorphism by the PBW theorem of [@rin63], which applies verbatim in the graded setting.
Now suppose that ${\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$ is a weak equivalence between $A$-cofibrant dg-Lie algebroids. The map ${\mathcal{U}}({\mathfrak{g}}){\rightarrow}{\mathcal{U}}({\mathfrak{h}})$ induces a quasi-isomorphism on the associated graded and is therefore a weak equivalence itself. This implies that $(f_*, f^!)$ is a Quillen equivalence.
The right Quillen functor $f^!$ induces a right adjoint functor of $\infty$-categories $f^!\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{h}}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$. This functor has a further right adjoint $f_!$, which sends $E$ to the derived coinduction ${{\mathrm{Hom}}}^h_{{\mathcal{U}}({\mathfrak{g}})}({\mathcal{U}}({\mathfrak{h}}), E)$.
\[rem:monoidalstructureonreps\] The tensor product $\otimes_A$ does not make ${{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}$ a monoidal model category. However, the tensor product $E\otimes_A (-)$ does preserve all weak equivalences whenever $E$ is cofibrant as a dg-$A$-module. When ${\mathfrak{g}}$ is $A$-cofibrant, every cofibrant left ${\mathcal{U}}({\mathfrak{g}})$-module is cofibrant as an $A$-module by the PBW theorem, so that the tensor product $\otimes_A$ descends to a tensor product on the $\infty$-category ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ [@lur16 Proposition 4.1.7.4]. Because the forgetful functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}{\rightarrow}{{\mathrm{Mod}}}_A$ detects equivalences and preserves all (co)limits, the resulting symmetric monoidal structure on ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ is closed.
By Lemma \[lem:pbw\] and Remark \[rem:monoidalstructureonreps\], any map of $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$ of dg-Lie algebroids induces a symmetric monoidal left and right adjoint $f^!\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{h}}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ between presentable (closed) symmetric monoidal $\infty$-categories. It follows from Lemma \[lem:pbw\] that this determines a functor $$\xymatrix{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar[r] & {
\StrLen{Pr^{\mathrm{L}}_{sym. mon.}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}_{sym. mon.}}
\else \mathrm{Pr^{\mathrm{L}}_{sym. mon.}}
\fi}\big/{{\mathrm{Mod}}}_A := {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({
\StrLen{Pr^L}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^L}
\else \mathrm{Pr^L}
\fi})/{{\mathrm{Mod}}}_A
}$$ to the $\infty$-category of presentable (closed) symmetric monoidal $\infty$-categories over ${{\mathrm{Mod}}}_A$, which sends ${\mathfrak{g}}$ to the forgetful functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}{\rightarrow}{{\mathrm{Mod}}}_A$.
\[fm:lem:algrepspreservessiftedcolim\] The functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}$ preserves all limits.
We can forget about the symmetric monoidal structures, since the forgetful functor ${
\StrLen{Pr^{\mathrm{L}}_{sym. mon.}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}_{sym. mon.}}
\else \mathrm{Pr^{\mathrm{L}}_{sym. mon.}}
\fi}={{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({
\StrLen{Pr^L}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^L}
\else \mathrm{Pr^L}
\fi}){\rightarrow}{
\StrLen{Pr^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}}
\else \mathrm{Pr^{\mathrm{L}}}
\fi}$ preserves limits and detects equivalences. For any map $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$, the restriction functor $f^!$ is both a left and a right adjoint. It therefore suffices to show that the functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}{\rightarrow}{
\StrLen{Pr^R}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^R}
\else \mathrm{Pr^R}
\fi}/{{\mathrm{Mod}}}_A$ preserves limits.
To see this, consider ${\mathcal{U}}({\mathfrak{g}})$ as a unital associative algebra in the category of dg-$A$-*bimodules* (over ${\mathbb{Q}}$), using the map of algebras $A{\rightarrow}{\mathcal{U}}({\mathfrak{g}})$. The category of dg-$A$-bimodules carries a monoidal model structure, whose weak equivalences and fibrations are transferred along the forgetful functor ${
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^{{\mathrm{dg}}}_A{\rightarrow}{
\StrLen{LMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LMod}
\else \mathrm{LMod}
\fi}^{{\mathrm{dg}}}_A$. Since we are working in characteristic zero, there exists a transferred model structure on the category of associative algebras in ${
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^{{\mathrm{dg}}}_A$ as well. By [@lur16 Theorem 4.1.8.4], this is a model for the $\infty$-category of algebras in the monoidal $\infty$-category ${
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}_A$ of $A$-bimodules (in loc. cit., the monoid axiom and symmetry of the monoidal structure are only used for the existence of a model structure on algebras).
The functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}$ now factors over the $\infty$-category of algebras in $A$-bimodules $$\xymatrix@C=3pc{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar[r]^-{{\mathcal{U}}} & {
\StrLen{Alg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Alg}
\else \mathrm{Alg}
\fi}\big({
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}_A\big)^{{\mathrm{op}}}\ar[r]^-{{\mathrm{LMod}}} & {
\StrLen{Pr^{R}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{R}}
\else \mathrm{Pr^{R}}
\fi}/{{\mathrm{Mod}}}_A.
}$$ The functor ${\mathrm{LMod}}$ preserves all limits by [@lur16 Theorem 4.8.5.11] (see the remarks just above Corollary 4.8.5.13). It remains to verify that ${\mathcal{U}}$ preserves limits. To see that it preserves sifted limits, it suffices to check that the composite $$\xymatrix{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar[r]^-{{\mathcal{U}}} & {
\StrLen{Alg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Alg}
\else \mathrm{Alg}
\fi}\big({
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}_A\big)^{{\mathrm{op}}}\ar[r]^-{{\mathrm{forget}}} & {
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^{{\mathrm{op}}}_A
}$$ preserves sifted limits. This was proven in [@nui17b Theorem 4.19].
To see that ${\mathcal{U}}$ preserves finite products, note that ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A$ is generated under sifted colimits by free Lie algebroids $F(V)$ on maps $0\colon V{\rightarrow}T_A$ of $A$-modules. It therefore suffices to verify that the composite $$\xymatrix{
{{\mathrm{Mod}}}_A\ar[r]^-{F} & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\ar[r]^-{{\mathcal{U}}} & {
\StrLen{Alg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Alg}
\else \mathrm{Alg}
\fi}\big({
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}_A\big)
}$$ preserves finite coproducts. Unraveling the definitions, one sees that for any cofibrant dg-$A$-module $F$, the dg-algebra ${\mathcal{U}}(F(V))$ is naturally equivalent to the $A$-linear tensor algebra $T_A(V)$. In other words, the above functor is naturally equivalent to the functor $$\xymatrix{
{{\mathrm{Mod}}}_A\ar[r]^\Delta & {
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}_A\ar[r]^-{{\mathrm{Free}}} & {
\StrLen{Alg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Alg}
\else \mathrm{Alg}
\fi}\big({
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}_A\big)
}$$ sending an $A$-module to the free algebra on $V$, considered as a symmetric $A$-bimodule. This functor clearly preserves colimits.
Pro-coherent sheaves and Lie algebroids {#sec:procoherent}
=======================================
The $\infty$-category of Lie algebroids over $A$ defined in the previous section was obtained by formally inverting the quasi-isomorphisms between dg-Lie algebroids over $A$. In this section, we discuss the *tame* homotopy theory of dg-Lie algebroids, which has *less* weak equivalences and is transferred from the tame model structure on dg-$A$-modules. We will start by recalling the tame model structure on dg-$A$-modules in Section \[sec:tamemodules\] and discuss the transferred homotopy theory on dg-Lie algebroids in Section \[sec:tameliealgebroids\].
Our main aim will be to show that in suitable cases, the tame model structure on dg-$A$-modules presents the $\infty$-category ${\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}})$ of *pro-coherent sheaves* on $A$ (or rather, its opposite). The tame homotopy theory theory on dg-Lie algebroids can then be considered as a model for the $\infty$-category of Lie algebroids in pro-coherent sheaves over $A$. In particular, the $\infty$-category of such pro-coherent Lie algebroids is *compactly generated*, which plays an important technical role in our discussion of Koszul duality. To make sure that the tame model structure on dg-$A$-modules indeed presents pro-coherent sheaves, we need to put some restrictions on $A$:
A connective commutative dg-$k$-algebra $A$ is called
1. (strictly) *bounded* if it is concentrated in degrees $[0, n]$, for some $n\geq 0$.
2. *coherent* if $\pi_0(A)$ is a coherent ring and each $\pi_n(A)$ is a finitely presented $\pi_0(A)$-module.
\[war:boundecoh\] The tame homotopy theory of dg-Lie algebroids described below is only well-behaved when $A$ is both cofibrant over $k$ (Remark \[rem:derivedtangent\]) and bounded coherent (Proposition \[prop:tamecompactlygenerated\]). There is some conflict between cofibrancy and being bounded: for example, $A$ cannot have any generators of degree $2$.
Suppose that $k$ is bounded coherent and that $A$ is obtained by adding finitely many generators in degrees $0$ and $1$. Then $A$ is cofibrant over $k$ and bounded coherent. Geometrically, $A$ describes the derived zero locus of a map ${\mathbb{A}}_k^n{\rightarrow}{\mathbb{A}}_k^m$ over ${{\mathrm{Spec}}}(k)$.
Tame dg-$A$-modules and pro-coherent sheaves {#sec:tamemodules}
--------------------------------------------
Recall that for any commutative dg-algebra $A$, the category of dg-$A$-modules can be endowed with the *tame*, or *contraderived* model structure on dg-$A$-modules (see e.g. [@bec14]). In this model structure, a map of dg-$A$-modules $E{\rightarrow}F$ is
- a fibration if it is degreewise surjective.
- a cofibration if it is a monomorphism, whose cokernel is projective as a graded $A$-module.
- a weak equivalence if for any graded-projective dg-$A$-module $P$, the map on hom-complexes ${{\mathrm{Hom}}}_A(P, E){\rightarrow}{{\mathrm{Hom}}}_A(P, F)$ is a quasi-isomorphism.
We will denote the associated $\infty$-category by ${{\mathrm{Mod}}}_A^!$ and refer to it as the $\infty$-category of *tame dg-$A$-modules*. This terminology is supposed to emphasize that ${{\mathrm{Mod}}}_A^!$ depends on the explicit cdga $A$, i.e. quasi-isomorphic cdgas may have non-equivalent $\infty$-categories of tame dg-$A$-modules.
The tame model structure is stable and symmetric monoidal for the usual tensor product over $A$, and the usual projective model structure is a (symmetric monoidal) right Bousfield localization of the tame model structure. It follows that there is a fully faithful, symmetric monoidal left adjoint functor $$\xymatrix{
{{\mathrm{Mod}}}_A\ar[r] & {{\mathrm{Mod}}}_A^!.
}$$
\[lem:tamecofgenerated\] Let ${
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}^{{{\scriptscriptstyle \leq}}n}$ be the set of graded-free dg-$A$-modules $T$ satisfying the following two conditions:
- $T$ has no generators in (homological) degree $>n$
- $T$ has finitely many generators in each individual degree.
Let ${
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}=\bigcup_n {
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}^{{{\scriptscriptstyle \leq}}n}$. Then the collection of cone inclusions $\big\{T{\rightarrow}T[0,1] : T\in{
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}\big\}$ is a set of generating cofibrations for the tame model structure. In particular, the tame model structure is cofibrantly generated.
A map $E{\rightarrow}F$ has the right lifting property against all $T{\rightarrow}T[0,1]$ if and only if the map ${{\mathrm{Hom}}}_A(T, E){\rightarrow}{{\mathrm{Hom}}}_A(T, F)$ is a trivial fibration for all $T\in {
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}$. Let ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\subseteq {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A$ be the full subcategory of graded-projective dg-$A$-modules $P$ for which ${{\mathrm{Hom}}}_A(P, E){\rightarrow}{{\mathrm{Hom}}}_A(P, F)$ is a trivial fibration. We have to show that ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ contains all graded-projective dg-$A$-modules. To see this, let us make the following observations:
1. ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ is closed under retracts.
2. Let $P{\rightarrow}Q$ be a cofibration such that $P$ and $Q/P$ are contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$. Then $Q$ is contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$.
3. If $\{P_\alpha\}$ is a (transfinite) sequence of cofibrations between objects in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$, then the colimit ${\operatornamewithlimits{\mathrm{colim}}}P_\alpha$ is contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ as well.
By (1), it suffices to verify that ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ contains all graded-free dg-$A$-modules. Suppose that $Q$ is a graded-free dg-$A$-module, with set of generators $S=\{y_i\}$. Let ${\mathcal{B}}\subseteq P(S)$ be the poset of subsets $S'\subseteq S$ which generate a sub dg-$A$-module of $Q$ that is contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$. It follows from (3) that any chain in ${\mathcal{B}}$ has an upper bound. By Zorn’s lemma, ${\mathcal{B}}$ admits a maximal element $S_0$. Let $P=A\cdot S_0$ and consider the quotient $Q/P$. We claim that $Q/P$ is trivial, so that $Q\in {
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$.
Indeed, $Q/P$ has a set of generators $S-S_0$. If there exists a generator $y\in S\setminus S_0$ of degree $n$, one can (inductively) find a set $S'\subseteq S-S_0$ containing $y$ such that
- $A\cdot S'$ is closed under the differential.
- $S'$ contains finitely many generators in each degree $\leq n$ and no generators in degrees $>n$.
In particular, such $A\cdot S'\subseteq Q/P$ is contained in ${
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}$. Using (2), it follows that $A\cdot (S_0\cup S')\subseteq Q$ is contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ as well. This contradicts maximality of $S_0$, so we conclude that $S-S_0$ is empty and $Q/P=0$.
\[rem:connectivemodules\] Let $E$ be a dg-$A$-module and consider its connective cover $$\tau_{{{\scriptscriptstyle \geq}}0} E= (\dots{\rightarrow}E_1{\rightarrow}Z_0(E){\rightarrow}0)$$ Then the map $\tau_{{{\scriptscriptstyle \geq}}0} E{\rightarrow}E$ is a tame weak equivalence if and only if $[T, E]=0$ for any $T\in {
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}^{{{\scriptscriptstyle \leq}}-1}$. We will say that $E$ is a *connective* tame dg-$A$-module if it satisfies these equivalent conditions. The full subcategory ${{\mathrm{Mod}}}_{A}^{!, {{\scriptscriptstyle \geq}}0}$ on the connective tame dg-$A$-modules determines an accessible $t$-structure on ${{\mathrm{Mod}}}_A^!$.
One can easily verify that a map between connective tame dg-$A$-modules is a tame weak equivalence if and only if it is a quasi-isomorphism. It follows that the fully faithful inclusion ${{\mathrm{Mod}}}_A{\rightarrow}{{\mathrm{Mod}}}_A^!$ induces an equivalence ${{\mathrm{Mod}}}_A^{{{\scriptscriptstyle \geq}}0}{\rightarrow}{{\mathrm{Mod}}}_A^{!,{{\scriptscriptstyle \geq}}0}$ on connective objects.
The remainder of this section is devoted to a proof of the following result, proven in e.g. [@nee08] in the case where $A$ is discrete:
\[prop:tamecompactlygenerated\] Suppose that $A$ is bounded and coherent and let ${\mathcal{K}}^{\leq 0}$ be the set of (weak equivalence classes of) dg-$A$-modules $E$ such that
- they are in ${
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}^{{{\scriptscriptstyle \leq}}0}$, i.e. graded-free, with generators of degree $\leq 0$ and finitely many generators in each degree.
- the $A$-linear dual $E^\vee = {{\mathrm{Hom}}}_A(E, A)$ is eventually coconnective, i.e. $\pi_i(E^\vee)$ vanishes for $i\gg 0$.
Then ${
\StrLen{K}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{K}
\else \mathrm{K}
\fi}^{\leq 0}$ is a set of compact generators for ${{\mathrm{Mod}}}_A^!$.
Before turning to the proof of Proposition \[prop:tamecompactlygenerated\], let us record the following consequence. Recall that for coherent $A$, an object $E\in {{\mathrm{Mod}}}_A$ is a *coherent* $A$-module if it has finitely many nontrivial homotopy groups $\pi_i(E)$, each of which is finitely presented over $\pi_0(A)$. Equivalently [@lur16 Proposition 7.2.4.17], it is eventually coconnective and almost finitely presented, i.e. presentable by a graded-free dg-$A$-module with finitely many generators in each degree and no generators in degrees $\ll 0$.
\[cor:tameasprocoh\] Let $A$ be bounded and coherent and consider the functor $$\smash{\xymatrix{
(-)^\vee\colon {{\mathrm{Mod}}}_A^!\ar[r] & {{\mathrm{Mod}}}_A^{{\mathrm{op}}}}}$$ sending a tame dg-$A$-module to its $A$-linear dual. This functor restricts to an equivalence between the full subcategory ${
\StrLen{K}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{K}
\else \mathrm{K}
\fi}^{\leq 0}\subseteq {{\mathrm{Mod}}}_A^!$ and the opposite of the full subcategory ${
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{{\scriptscriptstyle \geq}}0}_A\subseteq {{\mathrm{Mod}}}_A$ of connective coherent $A$-modules. In particular, there is an equivalence of stable $\infty$-categories $${{\mathrm{Mod}}}_A^!\simeq {\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}}).$$
The functor $(-)^\vee\colon {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A{\rightarrow}{{\mathrm{Mod}}}_A^{{{\mathrm{dg}}}, {{\mathrm{op}}}}$ is a left Quillen functor from the tame model structure to the projective model structure, whose right adjoint is the functor $(-)^\vee$ as well. For any (cofibrant) object $E\in{\mathcal{K}}^{\leq 0}$, the dual $E^\vee$ is graded-free, with finitely many generators in each nonnegative degree. In particular, it is cofibrant. Since $A$ is bounded, the derived unit map $E{\rightarrow}E^{\vee\vee}$ is an isomorphism. The full subcategory of ${{\mathrm{Mod}}}_A^!$ on the compact generators is therefore equivalent to its essential image under $(-)^\vee$ in ${{\mathrm{Mod}}}_A^{{\mathrm{op}}}$. Unwinding the definitions, this essential image is exactly the opposite of ${
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{{\scriptscriptstyle \geq}}0}$.
When $A$ is bounded coherent, we will refer to ${{\mathrm{Mod}}}_A^!={\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}})$ as the $\infty$-category of *pro-coherent sheaves* on $A$, even though this is technically its opposite category.
Let $f\colon A{\rightarrow}B$ be a weak equivalence of bounded coherent cdgas. Then there is a Quillen equivalence $f^!\colon {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A\leftrightarrows {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_B\colon f_!$ between the tame model structures, where $f^!$ sends $E$ to $B\otimes_A E$. Indeed, for every coherent $A$-module $E$, we have that $f^!(E)$ is a coherent $B$-module. Because $f^!(E^\vee) \cong f^!(E)^\vee$, the functor $f^!$ preserves compact objects and the composite $$\xymatrix{
{
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A^{{\mathrm{op}}}\ar[r]^{(-)^\vee} & {{\mathrm{Mod}}}_A^{!, \omega}\ar[r]^{f^!} & {{\mathrm{Mod}}}_B^{!, \omega}\ar[r]^{(-)^\vee} & {
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_B^{{\mathrm{op}}}}$$ is an equivalence. This implies that $f^!\colon {{\mathrm{Mod}}}_A^!{\rightarrow}{{\mathrm{Mod}}}_B^!$ is an equivalence.
\[lem:gradedfree\] Let $E$ be an object in ${\mathcal{K}}^{{{\scriptscriptstyle \leq}}0}$ and let $n_0$ be an integer such $\pi_i(E^\vee)=0$ for all $i\geq n_0$. Then the following holds:
1. For any indexing set $I$, the map $\bigoplus_I{{\mathrm{Hom}}}_A(E, A){\rightarrow}{{\mathrm{Hom}}}_A(E, \bigoplus_I A)$ is an isomorphism.
2. Let $F$ be a dg-$A$-module which is graded-free, with generators $x_i$. Let $F^{(n)}$ be the quotient of $F$ by the sub dg-$A$-module generated by the $x_i$ of degree $<n$. Then for all $n>1$: $$\pi_0{{\mathrm{Hom}}}_A(E, F^{(n)})=0.$$
3. For every $m\leq -n_0$, the map $$\xymatrix{
\pi_0{{\mathrm{Hom}}}_A(E, F)\ar[r] & \pi_0{{\mathrm{Hom}}}_A(E, F^{(m)})
}$$ is an isomorphism.
For (1), we can write $E=\bigoplus_{n{{\scriptscriptstyle \geq}}0} A^{\oplus i_n}[-n]$ at the level of graded $A$-modules. The map $\bigoplus_I{{\mathrm{Hom}}}_A(E, A){\rightarrow}{{\mathrm{Hom}}}_A(E, \bigoplus_I A)$ is then given by $$\xymatrix{
\bigoplus_I \prod_{n\geq 0} A^{\oplus i_n}[n]\ar[r] & \prod_{n\geq 0} \bigoplus_I A^{\oplus i_n}[n].
}$$ This map is an isomophism in each degree, since $A$ is concentrated in nonnegative degrees.
For assertion (2), note that $F^{(n)}$ is graded-free on the generators $x_i$ of degree $\geq n$. Since $E$ has generators in degree $\leq 0$, the complex ${{\mathrm{Hom}}}_A(E, F^{(n)})$ is trivial in degree $0$ when $n>1$.
For (3), consider the tower of fibrations between graded-free dg-$A$-modules $$\xymatrix{
F\ar[r] & \dots \ar[r] & F^{(-2)}\ar[r] & F^{(-1)}\ar[r] & F^{(0)}.
}$$ The natural map $F{\rightarrow}\lim_{n\leq 0} F^{(n)}$ to the (homotopy) limit of this tower can be identified with the map $$\xymatrix{
\bigoplus_{n} \bigoplus_{I_n} A[-n]\ar[r] & \prod_{n\leq 0} \bigoplus_{I_n} A[-n].
}$$ Because $A$ is concentrated in degrees $[0, n]$, this is an isomorphism.
For every $m\leq -n_0$, the fiber of the map ${{\mathrm{Hom}}}_A(E, F^{(m-1)}){\rightarrow}{{\mathrm{Hom}}}_A(E, F^{(m)})$ can be identified with $${{\mathrm{Hom}}}_A\Big(E, \bigoplus_{I_m} A[m-1]\Big)\cong \bigoplus_{I_m} {{\mathrm{Hom}}}_A\big(E, A[m-1]\big).$$ By assumption, the homotopy groups of this complex vanish in all degrees $\geq -1$. This implies that the map $$\xymatrix{
\pi_0{{\mathrm{Hom}}}_A(E, F)\cong \pi_0\lim_{n\leq 0}{{\mathrm{Hom}}}_A(E, F^{(n)})\ar[r] & \pi_0{{\mathrm{Hom}}}_A(E, F^{(m)})
}$$ is an isomorphism.
\[lem:generate\] Let ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\subseteq {{\mathrm{Mod}}}_A^!$ be the smallest full subcategory of ${{\mathrm{Mod}}}_A^!$ which contains the objects in ${\mathcal{K}}^{{{\scriptscriptstyle \leq}}0}$ and is closed under colimits and extensions. Then ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}={{\mathrm{Mod}}}_A^!$.
By Lemma \[lem:tamecofgenerated\], it suffices to show that ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ contains all objects $T\in {
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}$, which are of the form $\bigoplus_{n\geq n_0} A^{\oplus k_n}[-n]$ without differential. The dual of such a $T$ is given by the (projecively) cofibrant dg-$A$-module $\bigoplus_{n\geq n_0} A^{\oplus k_n}[n]$ and the map $T{\rightarrow}T^{\vee\vee}$ is an isomorphism. Consider the Postnikov tower of $T^\vee$ $$\xymatrix{
T^\vee\ar[r] & \dots\ar[r] & \tau_{{{\scriptscriptstyle \leq}}n}(T^\vee)\ar[r] & \tau_{{{\scriptscriptstyle \leq}}n-1}(T^\vee)\ar[r] & \cdots \ar[r] & \pi_0(T^\vee).
}$$ Because $A$ is coherent, each $\pi_i(T^\vee)$ is a coherent $A$-module and admits an almost finite resolution $Y_i{\stackrel{\sim}{{\rightarrow}}} \pi_i(T^\vee)$ [@lur16 Proposition 7.2.4.17]. One can use these resolutions to resolve the entire Postnikov tower by $$\smash{\xymatrix{
T^\vee\ar[r] & \dots \ar[r] & P_n\ar[r] & P_{n-1}\ar[r] & \dots \ar[r] & P_0
}}$$ where $P_n=\bigoplus_{i=0}^n Y_i[i]$, equipped with a certain differential. The sequence of $P_n$ becomes stationary in each individual degree, so that there is a homotopy equivalence $T^\vee{\rightarrow}P_\infty = \bigoplus_{i} Y_i[i]$. Taking the dual, one finds that $T=T^{\vee\vee}$ is homotopy equivalent to the colimit of the sequence $$\smash{\xymatrix{
P_0^\vee\ar[r] & P_1^\vee\ar[r] & \dots.
}}$$ This is a sequence of cofibrations whose associated graded consists of the $Y_i[i]^\vee$. Each $Y_i[i]^\vee$ is contained in ${\mathcal{K}}$, because its dual is $Y_i[i]\simeq \pi_i(T^\vee)[i]$. This implies that the (homotopy) colimit $T$ is contained in the category ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$.
By Lemma \[lem:generate\], the objects of ${\mathcal{K}}^{{{\scriptscriptstyle \leq}}0}$ generate ${{\mathrm{Mod}}}_A^!$. To prove that any $E\in {\mathcal{K}}^{{{\scriptscriptstyle \leq}}0}$ is compact, it suffices to show that for any set $S$ and any collection of graded-free dg-$A$-modules $\{P_\alpha\}_{\alpha\in S}$, the map $$\xymatrix{
\bigoplus_\alpha \pi_0{{\mathrm{Hom}}}_A(E, P_\alpha)\ar[r] & \pi_0{{\mathrm{Hom}}}_A\Big(E, \bigoplus_\alpha P_\alpha\Big)
}$$ is an isomorphism. Using the filtration of Lemma \[lem:gradedfree\], we obtain commuting squares $$\xymatrix@R=1.7pc{
\bigoplus_\alpha \pi_0{{\mathrm{Hom}}}_A(E, P_\alpha)\ar[d]\ar[r] & \pi_0{{\mathrm{Hom}}}_A\Big(E, \bigoplus_\alpha P_\alpha\Big)\ar[d]\\
\bigoplus_\alpha \pi_0{{\mathrm{Hom}}}_A(E, P^{(n)}_\alpha)\ar[r]^{\phi_n} & \pi_0{{\mathrm{Hom}}}_A\Big(E, \bigoplus_\alpha P^{(n)}_\alpha\Big)
}$$ for all $n\in {\mathbb{Z}}$. We will prove by decreasing induction on $n$ that $\phi_n$ is an isomorphism; this proves that the top map is an isomorphism, because the vertical maps become isomorphisms for all $m$ smaller than a certain $n_0$, by Lemma \[lem:gradedfree\]. We can start the induction at $n=2$, where both objects are zero by Lemma \[lem:gradedfree\].
For the inductive step, suppose that $\phi_n$ is an isomorphism. To prove that $\phi_{n-1}$ is an isomorphism as well, let $F_\alpha$ denote the fiber of the map $P_\alpha^{(n-1)}{\rightarrow}P_\alpha^{(n)}$. It suffices to check that the map $$\smash{\xymatrix{
\bigoplus_\alpha {{\mathrm{Hom}}}_A(E, F_\alpha)\ar[r] & {{\mathrm{Hom}}}_A\Big(E, \bigoplus F_\alpha\Big)
}}$$ is a quasi-isomorphism. But each $F_\alpha$ is just given by a direct sum $\bigoplus A[n-1]$, so the result follows from (a shift of) Lemma \[lem:gradedfree\].
Tame dg-Lie algebroids {#sec:tameliealgebroids}
----------------------
The tame model structure on dg-$A$-modules can be transferred to a semi-model structure on dg-Lie algebroids, by [@nui17b Remark 4.25]:
\[prop:monadicitytame\] The category of dg-Lie algebroids over $A$ carries the *tame* semi-model structure, in which a map is a weak equivalence (fibration) if and only if it is a tame weak equivalence (fibration). The forgetful functor ${{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A{\rightarrow}{{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A/T_A$ is a right Quillen functor to the tame model structure, which preserves all sifted homotopy colimits.
\[def:tameliealgebroids\] Let $A$ be a cofibrant connective cdga over $k$. The $\infty$-category of *tame* dg-Lie algebroids over $A$ is the $\infty$-categorical localization $${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A := {{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\Big[\{\text{tame w.e.}\}^{-1}\Big].$$ This is a locally presentable $\infty$-category by Remark \[rem:equivalenttoquillen\]. When $A$ is bounded coherent, one can also think of ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ as the $\infty$-category of *pro-coherent Lie algebroids* over $A$.
The projective model semi-model structure on dg-Lie algebroids is a right Bousfield localization of the tame semi-model structure, so that there is a fully faithful left adjoint functor $$\xymatrix{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\ar@{^{(}->}[r] & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!.
}$$ The discussion of the previous sections carries over verbatim to the tame case. For example, every (tame) $A$-cofibrant dg-Lie algebroid has an explicit cofibrant replacement $Q({\mathfrak{g}})$, which is graded-free on $\big({\mathrm{Sym}}_A{\mathfrak{g}}[1]\big)[-1]$. Furthermore, we have the following:
\[lem:propertiesoftamereps\] The category ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{{\mathrm{dg}}}$ carries a model structure transferred from the tame model structure on ${{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A$, whose associated $\infty$-category will be denoted ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^!$. This has the following properties:
1. For every $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$, restriction and induction form a Quillen pair between tame model structures $$\xymatrix{
f_*={\mathcal{U}}({\mathfrak{g}})\otimes_{{\mathcal{U}}({\mathfrak{h}})} -\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{{\mathrm{dg}}}\ar@<1ex>[r] & {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{h}}\ar@<1ex>[l]\colon f^!
}$$ which is a Quillen equivalence when $f$ is a tame weak equivalence between (tame) $A$-cofibrant dg-Lie algebroids.
2. ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_A^!$ is a closed symmetric monoidal category.
3. The induced functor of $\infty$-categories $$\xymatrix{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!, {{\mathrm{op}}}}\ar[r] & {
\StrLen{Pr^{\mathrm{L}}_{sym. mon.}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}_{sym. mon.}}
\else \mathrm{Pr^{\mathrm{L}}_{sym. mon.}}
\fi}/{{\mathrm{Mod}}}_A^!
}$$ preserves all limits.
The PBW filtration also applies to dg-Lie algebroids which are $A$-cofibrant for the tame model structure on dg-$A$-modules. The proof of Lemma \[lem:pbw\] can therefore be applied to show (1). Assertion (2) follows from Remark \[rem:monoidalstructureonreps\].
For (3), one applies the same proof as in Lemma \[fm:lem:algrepspreservessiftedcolim\]. Indeed, let ${
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^!_A$ be the $\infty$-category associated to the category of dg-$A$-bimodules (over ${\mathbb{Q}}$), endowed with the model structure transferred from the *tame* model structure on left dg-$A$-modules. This is a cofibrantly generated monoidal model category, with generating cofibrations with cofibrant domains. By [@lur16 Theorem 4.1.8.4], the $\infty$-category ${
\StrLen{Alg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Alg}
\else \mathrm{Alg}
\fi}({
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^!_A)$ can be modeled by the transferred model structure on associative algebras in this monoidal model category. Forgetting the symmetric monoidal structure, the functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}$ now decomposes as $$\xymatrix@C=3pc{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar[r]^-{{\mathcal{U}}} & {
\StrLen{Alg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Alg}
\else \mathrm{Alg}
\fi}\big({
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^!_A\big)^{{\mathrm{op}}}\ar[r]^-{{\mathrm{LMod}}} & {
\StrLen{Pr^{R}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{R}}
\else \mathrm{Pr^{R}}
\fi}/{{\mathrm{Mod}}}_A.
}$$ The second functor preserves limits by [@lur16 Theorem 4.8.5.11]. To see that the first functor preserves sifted limits, it suffices to show that ${\mathcal{U}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A{\rightarrow}{
\StrLen{BiMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{BiMod}
\else \mathrm{BiMod}
\fi}^!_A$ preserves sifted colimits. This follows from [@nui17b Theorem 4.19, Remark 4.23]. As in the proof of Lemma \[fm:lem:algrepspreservessiftedcolim\], the fact that ${\mathcal{U}}$ also preserves finite products follows from the fact that ${\mathcal{U}}(F(V))\cong T_A(V)$ for any cofibrant dg-$A$-module.
Lie algebroid cohomology {#sec:ce}
========================
The purpose of this section is to prove the following:
\[prop:deformationadjunction\] Let $A$ be a cofibrant commutative dg-$k$-algebra. There is an adjunction of $\infty$-categories $$\label{diag:deformationadjunction}\vcenter{\xymatrix{
C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A\ar@<1ex>[r] & \big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}\colon {\mathfrak{D}}\ar@<1ex>[l]
}}$$ between the $\infty$-category of tame dg-Lie algebroids over $A$ and the $\infty$-category of unbounded commutative $k$-algebras over $A$ (i.e. cdgas up to quasi-isomorphism). The right adjoint sends $B{\rightarrow}A$ to the dual in ${{\mathrm{Mod}}}_A^!$ of the map between cotangent complexes $L_A{\rightarrow}L_{A/B}$ (over $k$).
Composing with the fully faithful left adjoint ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A{\rightarrow}{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$, one obtains an adjunction $\smash{C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\leftrightarrows \big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}\colon {\mathfrak{D}}}$ as well. The right adjoint takes the dual of $L_A{\rightarrow}L_{A/B}$ inside ${{\mathrm{Mod}}}_A\subseteq {{\mathrm{Mod}}}_A^!$.
The left adjoint in is given by the *Chevalley-Eilenberg complex* (with trivial coefficients). Recall that for any representation $E$ of a dg-Lie algebroid ${\mathfrak{g}}$, this complex (with coefficients in $E$) is given by the graded vector space $$C^*({\mathfrak{g}}, E) := {{\mathrm{Hom}}}_A\Big({\mathrm{Sym}}_A{\mathfrak{g}}[1], E\Big)$$ with the usual Chevalley-Eilenberg (or de Rham) differential (without Koszul signs) $$\begin{aligned}
\label{fm:eq:cediff}
({\partial}\alpha)(X_1, \dots, X_n) & = {\partial}_E\big(\alpha(X_1, \dots, X_n)\big) - \sum_{i} \hspace{2pt} \alpha(X_1, \dots, {\partial}X_i, \dots, X_n) \nonumber\\
&+ \sum_{i} \hspace{2pt} \nabla_{X_{i}}\Big(\alpha(X_1, \dots, X_n)\Big) - \sum_{i<j} \hspace{2pt} \alpha\Big([X_i, X_j], X_1, \dots, X_n\Big).\end{aligned}$$ There is a natural $k$-linear augmentation map $C^*({\mathfrak{g}}, E){\rightarrow}E$, evaluating at the unit element of ${\mathrm{Sym}}_A{\mathfrak{g}}[1]$.
\[rem:celaxmonoidal\] The shuffle product of forms defines a lax symmetric monoidal structure on $C^*({\mathfrak{g}}, -)$, which is compatible with the augmentation in the sense that there is a commuting square $$\xymatrix{
C^*({\mathfrak{g}}, E)\otimes_k C^*({\mathfrak{g}}, F)\ar[r]^-{\times}\ar[d] & C^*({\mathfrak{g}}, E\otimes_A F)\ar[d]\\
E\otimes_k F\ar[r] & E\otimes_A F.
}$$ In particular, taking coefficients with values in the commutative algebra $A$, one obtains a functor $$\xymatrix{
C^*\colon {{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\ar[r] & \Big({{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{{{\mathrm{dg}}}}}}_k/A\Big)^{{\mathrm{op}}}; \hspace{4pt} {\mathfrak{g}}\ar@{|->}[r] & C^*({\mathfrak{g}}):=C^*({\mathfrak{g}}, A).
}$$ For every dg-Lie algebroid ${\mathfrak{g}}$, the Chevalley-Eilenberg complex yields a lax symmetric monoidal functor $C^*({\mathfrak{g}}, -)\colon {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}{\rightarrow}{{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_{C^*({\mathfrak{g}})}$.
The proof of Proposition \[prop:deformationadjunction\] is given at the very end of this section and uses some formal properties of the Chevalley-Eilenberg complex. To understand these properties, it will be useful to first give a slighty more model-categorical characterization of the Chevalley-Eilenberg complex.
The cotangent complex of a Lie algebroid
----------------------------------------
Consider the right Quillen functor $$\xymatrix{
{\mathfrak{g}}\oplus (-)\colon {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}\ar[r] & {{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A/{\mathfrak{g}}
}$$ sending a ${\mathfrak{g}}$-representation $E$ to the square zero extension of ${\mathfrak{g}}$ by $E$. If we denote the value of the left adjoint functor on ${\mathfrak{g}}$ itself by $\Upsilon_{\mathfrak{g}}$, then the other values of the left adjoint are given by $$\label{diag:kaehler}\xymatrix{
\Big(f\colon {\mathfrak{h}}{\rightarrow}{\mathfrak{g}}\Big)\ar@{|->}[r] & f_*\Upsilon_{\mathfrak{h}} = {\mathcal{U}}({\mathfrak{g}})\otimes_{{\mathcal{U}}({\mathfrak{h}})}\Upsilon_{{\mathfrak{h}}}.
}$$
\[fm:def:cotangentcomplexofliealgebroid\] Let ${\mathfrak{g}}$ be an $A$-cofibrant dg-Lie algebroid. The *cotangent complex* $L_{\mathfrak{g}}$ of ${\mathfrak{g}}$ is the value of the left derived functor of on the identity map of ${\mathfrak{g}}$. In other words, it is the universal ${\mathfrak{g}}$-representation classified by $${{\mathrm{Map}}}_{{\mathcal{U}}({\mathfrak{g}})}(L_{\mathfrak{g}}, E)\simeq {{\mathrm{Map}}}_{/{\mathfrak{g}}}\big({\mathfrak{g}}, {\mathfrak{g}}\oplus E\big).$$
\[fm:ex:cotangentcomplexoffreealgebroid\] Let ${\mathfrak{g}}=F(V)$ be the free dg-Lie algebroid generated by an $A$-linear map $V{\rightarrow}T_A$ whose domain is cofibrant. Then ${\mathfrak{g}}$ is a cofibrant dg-Lie algebroid and for any ${\mathfrak{g}}$-representation $E$, there is a natural bijection between sections ${\mathfrak{g}}{\rightarrow}{\mathfrak{g}}\oplus E$ and $A$-linear maps $V{\rightarrow}E$. It follows that $L_{\mathfrak{g}}={\mathcal{U}}({\mathfrak{g}})\otimes_A V$ is the free ${\mathfrak{g}}$-representation generated by the dg-$A$-module $V$.
More generally, one can use the cofibrant replacement $Q({\mathfrak{g}})$ of Definition \[def:Q\] to compute the cotangent complex:
\[prop:cotangentofliealgebroid\] Let ${\mathfrak{g}}$ be an $A$-cofibrant dg-Lie algebroid. Then the cotangent complex $L_{\mathfrak{g}}$ can be modeled by the cofibrant left ${\mathcal{U}}({\mathfrak{g}})$-module $$\label{eq:cotangentbycobar}
L_{\mathfrak{g}} = {\mathcal{U}}({\mathfrak{g}})\otimes_A \left({\mathrm{Sym}}_A^{{{\scriptscriptstyle \geq}}1}{\mathfrak{g}}[1]\right)[-1]$$ with differential given (modulo Koszul signs) by $$\begin{aligned}
\label{fm:eq:koszuldifferential}
{\partial}(u\otimes X_1\dots X_n) &= ({\partial}u)\otimes X_1\dots X_n + \sum_i u\otimes X_1\dots {\partial}(X_i)\dots X_n \nonumber\\
& \smash{\stackrel{(n>1)}{+}} \sum_{i} u\cdot X_k\otimes X_{1}\dots X_{n} + \sum_{i<j} u \otimes [X_i, X_j]X_1\dots X_n.\end{aligned}$$ The first term in the second row only applies when $n>1$.
Since ${\mathfrak{g}}$ is $A$-cofibrant, it suffices to find a ${\mathfrak{g}}$-representation corepresenting the functor $$\xymatrix{
{{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}\ar[r] & {{
\StrLen{Set}[\mystrleng]
\ifnum\mystrleng=1 \mathscr{Set}
\else \mathrm{Set}
\fi
}}; \hspace{4pt} E\ar@{|->}[r] & {{\mathrm{Hom}}}_{/{\mathfrak{g}}}(Q({\mathfrak{g}}), E).
}$$ By [@nui17b Corollary 6.13], there is a natural bijection between maps $Q({\mathfrak{g}}){\rightarrow}{\mathfrak{g}}\oplus E$ over ${\mathfrak{g}}$ and $0$-cycles in the kernel of the augmentation map $C^*({\mathfrak{g}}, E[1]){\rightarrow}E[1]$. On the other hand, unwinding the definition of the complex $L_{\mathfrak{g}}$ in , one sees that there is a short exact sequence $$\label{eq:ceismapsfromcot}\vcenter{\xymatrix{
{{\mathrm{Hom}}}{\hspace{1pt}}_{{\mathcal{U}}({\mathfrak{g}})}\big(L_{\mathfrak{g}}, E\big)\ar[r] & C^*({\mathfrak{g}}, E[1])\ar[r] & E[1].
}}$$ Passing to $0$-cycles, one finds that the complex $L_{\mathfrak{g}}$ given in indeed represents maps $Q({\mathfrak{g}}){\rightarrow}{\mathfrak{g}}\oplus E$ over ${\mathfrak{g}}$.
\[fm:rem:cofibersequenceoncotangent\] The cotangent complex comes with a ${\mathcal{U}}({\mathfrak{g}})$-linear map $$\xymatrix@C=1.8pc{
L_{\mathfrak{g}} = {\mathcal{U}}({\mathfrak{g}})\otimes_A \big({\mathrm{Sym}}_A^{{{\scriptscriptstyle \geq}}1}{\mathfrak{g}}[1]\big)[-1]\ar[r] & {\mathcal{U}}({\mathfrak{g}})
}$$ sending $u\otimes X_1\dots X_n$ to zero when $n>1$ and to $u\cdot X_1$ when $n=1$. The *Koszul complex* $K({\mathfrak{g}})$ of ${\mathfrak{g}}$ is the mapping cone of this map. It fits onto a cofiber sequence $$\xymatrix{
L_{\mathfrak{g}}\ar[r] & {\mathcal{U}}({\mathfrak{g}})\ar[r] & K({\mathfrak{g}}).
}$$ Unraveling the definitions, $K({\mathfrak{g}})$ can be identified with ${\mathcal{U}}({\mathfrak{g}})\otimes_A {\mathrm{Sym}}_A{\mathfrak{g}}[1]$, with differential given by formula , but where the term in the second line is also included when $n=1$. The Chevalley-Eilenberg complex $C^*({\mathfrak{g}}, E)$ can be identified with ${{\mathrm{Hom}}}\hspace{1pt}_{{\mathcal{U}}({\mathfrak{g}})}(K({\mathfrak{g}}), E)$, so that the above cofiber sequence induces (as shift of) the fiber sequence on mapping complexes.
\[rem:ceasext\] The composite map $$\label{fm:diag:cofiberoncotangent}\vcenter{\xymatrix@C=3pc{
L_{\mathfrak{g}}={\mathcal{U}}({\mathfrak{g}})\otimes_A \big({\mathrm{Sym}}_A^{{{\scriptscriptstyle \geq}}1}{\mathfrak{g}}[1]\big)[-1]\ar[r] & {\mathcal{U}}({\mathfrak{g}})\ar[r]^{u\mapsto u\cdot 1} & A
}}$$ is equal to zero, so that there is a ${\mathcal{U}}({\mathfrak{g}})$-linear map $K({\mathfrak{g}}){\rightarrow}A$. When ${\mathfrak{g}}$ is $A$-cofibrant, this map is a weak equivalence. Indeed, the PBW filtration on ${\mathcal{U}}({\mathfrak{g}})$ (see the proof of Lemma \[lem:pbw\]) and the filtration on ${\mathrm{Sym}}_A{\mathfrak{g}}[1]$ by polynomial degree determine a total filtration on the Koszul complex ${\mathcal{K}}({\mathfrak{g}})$. The map on the associated graded is the obvious projection $$\xymatrix@C=1.4pc{
{\mathrm{Sym}}_A({\mathfrak{g}}[0, 1])={\mathrm{Sym}}_A{\mathfrak{g}}\otimes_A {\mathrm{Sym}}_A {\mathfrak{g}}[1] \ar[r] & A
}$$ from the symmetric algebra on the cone ${\mathfrak{g}}[0, 1]$, which is a weak equivalence.
In other words, $C^*({\mathfrak{g}}, E)$ is a model for the derived mapping space ${{\mathrm{Hom}}}_{{\mathcal{U}}({\mathfrak{g}})}(A, E)$. The lax symmetric monoidal structure on $C^*({\mathfrak{g}}, -)$ arises from the fact that $A$ is a cocommutative coalgebra in ${{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}$.
The free case {#fm:sec:ceoffree}
-------------
When ${\mathfrak{g}}=F(V)$ is the free dg-Lie algebroid generated by a cofibrant dg-$A$-module over $T_A$, we now have two different (but weakly equivalent) descriptions of the cotangent complex $L_{\mathfrak{g}}$: Example \[fm:ex:cotangentcomplexoffreealgebroid\] simply evaluates the left Quillen functor on ${\mathfrak{g}}$ itself, while Proposition \[prop:cotangentofliealgebroid\] computes the value of on the ‘cobar’ resolution $Q({\mathfrak{g}})$. Of course, the first description is significantly smaller than the second.
When ${\mathfrak{g}}=F(V)$ is a free Lie algebroid, there is a canonical section of the map $Q({\mathfrak{g}}){\stackrel{\sim}{{\rightarrow}}} {\mathfrak{g}}$, induced by the canonical inclusion $$\xymatrix{
V\ar@{^{(}->}[r] & {\mathfrak{g}}\ar[r] & \Big({\mathrm{Sym}}_A{\mathfrak{g}}[1]\Big)[-1]\subseteq Q({\mathfrak{g}}).
}$$ Applying to this section produces a weak equivalence between the two models for the cotangent complex $L_{\mathfrak{g}}$, given by the ${\mathcal{U}}({\mathfrak{g}})$-linear extension of the above inclusion $$\label{diag:comparisonofcotangent}\smash{\xymatrix{
{\mathcal{U}}({\mathfrak{g}})\otimes_A V\ar[r]^-{\sim} & {\mathcal{U}}({\mathfrak{g}})\otimes_A \Big({\mathrm{Sym}}^{{{\scriptscriptstyle \geq}}1}_A {\mathfrak{g}}[1]\Big)[-1].
}}$$ Dually, restriction along this map induces a weak equivalence map to a significantly smaller complex $$\xymatrix{
\kappa\colon C^*({\mathfrak{g}}, A)\ar[r]^-\simeq & A^V = A\oplus_{\rho^\vee} {{\mathrm{Hom}}}_A\big(V[1], A\big); \hspace{4pt} \alpha\ar@{|->}[r] & \Big(\alpha(1), \alpha\big|_{V[1]}\Big).
}$$ The codomain $A^V$ has the natural structure of a commutative dg-algebra, the *square zero extension* of $A$ by $V[1]^\vee={{\mathrm{Hom}}}_A(V[1], A)$ classified by the map $$\xymatrix{
\rho^\vee\colon \Omega_A\ar[r] & V^\vee; \hspace{4pt} d_{{{\mathrm{dR}}}}a\ar@{|->}[r] & \Big(v\mapsto \rho(v)(a)\Big).
}$$ In other words, $A^V$ fits into a (homotopy) pullback square of cdgas $$\label{diag:squarezero}\vcenter{\xymatrix{
A^V= A\oplus_{\rho^\vee} V[1]^\vee\ar[r] \ar[d] & A\oplus V[0, 1]^\vee\ar[d]\\
A\ar[r]_{\rho^\vee} & A\oplus V^\vee.
}}$$ Using this pullback square, one can easily check that the functor $$\xymatrix{
A^{(-)}\colon {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A/T_A\ar[r] & \big({{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{{{\mathrm{dg}}}}}}_k/A\big)^{{\mathrm{op}}}; V\ar@{|->}[r] & A^V= A\oplus_{\rho^\vee} V[1]^\vee.
}$$ is a left Quillen functor from the tame model structure on dg-$A$-modules to the usual (projective) model structure on cdgas over $A$. Its right adjoint sends $B{\rightarrow}A$ to the mapping fiber of the map $$\xymatrix{
T_A = {\mathrm{Der}}_k(A, A)\ar[r] & {\mathrm{Der}}_k(B, A),
}$$ together with its natural projection to $T_A$. We may therefore summarize the previous discussion by the following result:
\[cor:ceoffree\] There is a natural transformation to a right Quillen functor $$\xymatrix@C=4pc{
{{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A/T_A \ar@/^1.5pc/[r]^{C^*\circ F}_{}="s"\ar@/_1.5pc/[r]_{A^{(-)}}^{}="t" \ar@{=>}"s";"t"^\kappa & \big({{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{{{\mathrm{dg}}}}}}_k/A\big)^{{\mathrm{op}}}}$$ which is a weak equivalence when restricted to cofibrant dg-$A$-modules over $T_A$.
Let us finally turn to the proof of Proposition \[prop:deformationadjunction\]:
Consider the functor $$\xymatrix{
{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\ar[r]^-{C^*} & \big({{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{{{\mathrm{dg}}}}}}_k/A\big)^{{\mathrm{op}}}\ar[r]^-\ker & \big({{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_k\big)^{{\mathrm{op}}}}$$ sending a dg-Lie algebroid to the kernel of the map $C^*({\mathfrak{g}}){\rightarrow}A$. By Proposition \[prop:cotangentofliealgebroid\], this functor can be identified with the composite $$\xymatrix{
{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\ar[r]^-{Q} & {{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A\ar[r]^-{\Upsilon} & {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{T_A}\ar[rr]^-{{{\mathrm{Hom}}}_{{\mathcal{U}}(T_A)}(-, A)} & & \big({{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_k\big)^{{\mathrm{op}}}.
}$$ The functor $Q$ sends an $A$-cofibrant dg-Lie algebroid to its cofibrant replacement, the functor $\Upsilon$ is the functor sending ${\mathfrak{g}}$ to ${\mathcal{U}}(T_A)\otimes_{{\mathcal{U}}({\mathfrak{g}})} \Upsilon_{\mathfrak{g}}$ and the last functor takes the maps from a $T_A$-representation to $A$. Since the last two functors are left Quillen functors, it follows that $C^*$ preserves weak equivalences between $A$-cofibrant dg-Lie algebroids. We therefore obtain functors of $\infty$-categories $$\xymatrix{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!\ar[r]^-{C^*} & \big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}\ar[r]^-\ker & \big({{\mathrm{Mod}}}_k\big)^{{\mathrm{op}}}.
}$$ The composition of these two functor preserves all colimits. Since the functor $\ker\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A{\rightarrow}{{\mathrm{Mod}}}_k$ detects equivalences and preserves all limits, it follows that $C^*$ preserves all colimits. By the adjoint functor theorem, it follows that $C^*$ admits a right adjoint ${\mathfrak{D}}\colon \big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}{\rightarrow}{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$.
It remains to describe this right adjoint ${\mathfrak{D}}$, at least at the level of the underlying tame $A$-modules. To this end, observe that the composite $$\label{diag:forgetafterkoszul}\xymatrix{
\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}\ar[r]^-{{\mathfrak{D}}} & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!\ar[r]^{U} & {{\mathrm{Mod}}}_A^!/T_A
}$$ is right adjoint to the functor $C^*\circ F\colon {{\mathrm{Mod}}}_A^!/T_A{\rightarrow}\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}$. Corollary \[cor:ceoffree\] provides a natural equivalence $\kappa\colon C^*\circ F{\rightarrow}A^{(-)}$, so that $U\circ {\mathfrak{D}}$ is equivalent to the derived right adjoint to the left Quillen functor $A^{(-)}$. Because $A$ is cofibrant over $k$, the discussion preceding Corollary \[cor:ceoffree\] shows that this derived functor sends $B{\rightarrow}A$ to the $A$-linear dual of $L_A{\rightarrow}L_{A/B}$.
\[fm:rem:dualityfunctor\] The functor ${\mathfrak{D}}$ does not admit a straightforward point-set description. However, when $B{\rightarrow}A$ is a *cofibration* of commutative dg-algebras, its image under ${\mathfrak{D}}$ can be identified with the sub-dg-Lie algebroid $T_{A/B}:={\mathrm{Der}}_B(A, A){\rightarrow}T_A$ of $B$-linear derivations of $A$. To see this, note that there is a natural diagram of commutative dg-$k$-algebras over $A$ $$\xymatrix{
B\ar[r]^-{f} \ar[rd]_-\phi & C^*(T_{A/B})\ar[r]\ar[d] & C^*(\tilde{T}_{A/B})\ar[ld]\\
& A &
}$$ where $\tilde{T}_{A/B}{\rightarrow}T_{A/B}$ is a cofibrant replacement. For any $b\in B$, its image under $f$ is the map $$\xymatrix@C=4pc{
{\mathrm{Sym}}_A\big(T_{A/B}[1]\big)\ar[r]^-{T_{A/B}\mapsto 0} & A\ar[r]^{\phi(b)\cdot (-)} & A
}$$ The map $f$ respects the differential because the derivations in $T_{A/B}$ are $B$-linear. The composition $B{\rightarrow}C^*(\tilde{T}_{A/B})$ is adjoint to a map $\tilde{T}_{A/B}{\rightarrow}{\mathfrak{D}}(B{\rightarrow}A)$. At the level of the underlying (tame) $A$-modules, this map is simply given by the composite map $$\xymatrix{
\tilde{T}_{A/B}\ar[r]^\sim & T_{A/B}\ar[r] & L_{A/B}^\vee
}$$ Because $B{\rightarrow}A$ was a cofibration, the second map is a weak equivalence. It follows that the dg-Lie algebroid $\tilde{T}_{A/B}$ (and therefore $T_{A/B}$) is weakly equivalent to ${\mathfrak{D}}(B{\rightarrow}A)$.
Koszul duality {#sec:koszulduality}
==============
The purpose of this section is to show that the adjunction from Proposition \[prop:deformationadjunction\] $$\xymatrix{
C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A\ar@<1ex>[r] & \big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}\colon {\mathfrak{D}}\ar@<1ex>[l]
}$$ is not too far from being an equivalence: when restricted to suitably *free* tame dg-Lie algebroids, the functor $C^*$ is fully faithful.
\[fm:prop:koszulduality\] Let $A$ be a cofibrant connective commutative dg-$k$-algebra and let ${\mathfrak{g}}$ be a cofibrant dg-Lie algebroid in the tame model structure. Suppose ${\mathfrak{g}}$ satisfies the following conditions:
- As a graded $A$-module, ${\mathfrak{g}}$ is free on a set of generators $x_i$.
- There are finitely many $x_i$ in each single degree, and no generators of (homological) degree $\geq 0$.
Then the unit map ${\mathfrak{g}}{\rightarrow}{\mathfrak{D}}C^*({\mathfrak{g}})$ can be identified with the map of dg-$A$-modules $$\xymatrix{
{\mathfrak{g}}\ar[r] & {\mathfrak{g}}^{\vee\vee}
}$$ from ${\mathfrak{g}}$ into its $A$-linear bidual.
\[cor:koszuldualitycoherent\] Let $A$ be a cofibrant commutative dg-$k$-algebra which is bounded. Then $C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!{\rightarrow}\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}$ is fully faithful on all tame dg-Lie algebroids satisfying conditions (i) and (ii) in Proposition \[fm:prop:koszulduality\].
Forgetting about the differential, we can write ${\mathfrak{g}} = \bigoplus_{n<0} A^{\oplus k_n}[n]$. The unit map ${\mathfrak{g}}{\rightarrow}{\mathfrak{g}}^{\vee\vee}$ can then be identified with the map $$\xymatrix{
\bigoplus_{n<0} A^{\oplus k_n}[n]\ar[r] & \prod_{n<0} A^{\oplus k_n}[n].
}$$ When $A$ is bounded, this map is an isomorphism.
\[cor:koszuldualityperfect\] Let $A$ be a cofibrant commutative dg-$k$-algebra which is eventually coconnective. Then $C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A{\rightarrow}\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\big)^{{\mathrm{op}}}$ is fully faithful on all Lie algebroids that can be modeled by *projectively* cofibrant dg-Lie algebroids satisfying conditions (i) and (ii) of Proposition \[fm:prop:koszulduality\].
When ${{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{{{\mathrm{dg}}}}}_A$ is equipped with the projective model structure, it suffices to verify that the map ${\mathfrak{g}}{\rightarrow}{\mathfrak{g}}^{\vee\vee}$ is a quasi-isomorphism for any projectively cofibrant dg-Lie algebroid satisfying conditions (i) and (ii). Since ${\mathfrak{g}}$ is a projectively cofibrant dg-$A$-module, ${\mathfrak{g}}^{\vee\vee}$ is a model for the derived bidual of ${\mathfrak{g}}$. It therefore suffices to pick a weak equivalence $A{\rightarrow}A'$ to a bounded cdga and verify that the map ${\mathfrak{g}}\otimes_A A'{\rightarrow}({\mathfrak{g}}\otimes_A A')^{\vee\vee}$ is a quasi-isomorphism. But because $A'$ is bounded, this map is an isomorphism, as in the previous proof.
To prove Proposition \[fm:prop:koszulduality\], let us start by considering the map of commutative dg-algebras over $A$ $$\label{diag:classifyingmapforunit}\xymatrix{
c\colon C^*({\mathfrak{g}})\ar[r] & A^{{\mathfrak{g}}} = A\oplus_{\rho^\vee} {\mathfrak{g}}[1]^\vee
}$$ which sends $\alpha\colon {\mathrm{Sym}}_A{\mathfrak{g}}[1]{\rightarrow}A$ to $\big(\alpha(1), \alpha\big|_{{\mathfrak{g}}[1]}\big)$. In the proof of Proposition \[prop:deformationadjunction\], we have seen that the functor ${\mathfrak{g}}\mapsto A^{\mathfrak{g}}$ was a right Quillen functor, whose derived left adjoint sent $$\xymatrix{
\big(B{\rightarrow}A\big)\ar@{|->}[r] & \big(L_{A/B}^\vee{\rightarrow}L_A^\vee=T_A\big).
}$$
\[lem:identificationunderadjunction\] Let ${\mathfrak{g}}$ be a tamely cofibrant dg-Lie algebroid over $A$ and consider the $A$-linear map ${\mathfrak{g}}{\rightarrow}L_{A/C^*({\mathfrak{g}})}^\vee$ adjoint to . This map is equivalent to the $A$-linear map underlying the unit map ${\mathfrak{g}}{\rightarrow}{\mathfrak{D}}C^*({\mathfrak{g}})$.
Let $F\colon {{\mathrm{Mod}}}_A^!/T_A\leftrightarrows {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!\colon U$ be the free-forgetful adjunction. The $A$-linear map $U({\mathfrak{g}}){\rightarrow}U{\mathfrak{D}}C^*({\mathfrak{g}})$ in ${{\mathrm{Mod}}}_A^!/T_A$ corresponds by adjunction to the map $$\xymatrix{
C^*({\mathfrak{g}})\ar[r] & C^*(F({\mathfrak{g}}))
}$$ in ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}/A$. The composition of this map with the equivalence $\kappa\colon C^*(F({\mathfrak{g}})){\rightarrow}A^{\mathfrak{g}}$ from Corollary \[cor:ceoffree\] is exactly the map . This means that the maps $${\mathfrak{g}}{\rightarrow}L_{A/C^*({\mathfrak{g}})}^\vee \quad \text{and} \quad U({\mathfrak{g}}){\rightarrow}U{\mathfrak{D}}C^*({\mathfrak{g}})$$ are identified under the adjoint equivalence between $U{\mathfrak{D}}$ and the functor sending $B{\rightarrow}A$ to $L_{A/B}^\vee$.
To use Lemma \[lem:identificationunderadjunction\], we will have to compute the relative cotangent complex of the map $C^*({\mathfrak{g}}){\rightarrow}A$. Unfortunately, $C^*({\mathfrak{g}})$ has the structure of a power series algebra, which means that $C^*({\mathfrak{g}})$ is not cofibrant and computing its cotangent complex requires some effort. Let us therefore introduce the following ‘global’ variant of the Chevalley-Eilenberg complex:
\[fm:cons:globalizedCE\] Let $A$ be a cofibrant commutative dg-$k$-algebra and let ${\mathfrak{g}}$ be a (tamely) cofibrant dg-Lie algebroid over $A$ satisfying conditions (i) and (ii) of Proposition \[fm:prop:koszulduality\]. Let $$C^*_{\mathrm{poly}}({\mathfrak{g}}) := {\mathrm{Sym}}_A \big({\mathfrak{g}}[1]^\vee\big) \subseteq C^*({\mathfrak{g}})$$ be the graded-subalgebra of $C^*({\mathfrak{g}})$ consisting of graded $A$-linear maps ${\mathrm{Sym}}_A{\mathfrak{g}}[1] {\rightarrow}A$ that vanish on some ${\mathrm{Sym}}^{{{\scriptscriptstyle \geq}}n}_A {\mathfrak{g}}[1]$. This graded subalgebra of $C^*({\mathfrak{g}})$ is closed under the Chevalley-Eilenberg differential of $C^*({\mathfrak{g}})$, which sends a function vanishing on ${\mathrm{Sym}}^{{{\scriptscriptstyle \geq}}n}_A {\mathfrak{g}}[1]$ to a function vanishing on ${\mathrm{Sym}}^{{{\scriptscriptstyle \geq}}n+1}_A {\mathfrak{g}}[1]$.
\[fm:ex:polynomialintopowerseries\] Let ${\mathfrak{g}}=A^{\oplus n}[-1]$ be the trivial dg-Lie algebroid on $n$ generators of degree $-1$. Then $C^*({\mathfrak{g}})$ is isomorphic to the ring of power-series $A[[x_1, ..., x_n]]$ and the inclusion $C^*_{\mathrm{poly}}({\mathfrak{g}})\subseteq C^*({\mathfrak{g}})$ is the inclusion of the polynomial algebra $A[x_1, ..., x_n]\subseteq A[[x_1, ..., x_n]]$.
The commutative dg-algebra $C^*_{\mathrm{poly}}({\mathfrak{g}})$ is not a homotopy invariant.
\[fm:lem:cotangentforglobalce\] Let $A$ be a cofibrant commutative dg-$k$-algebra and let ${\mathfrak{g}}$ be a dg-Lie algebroid over $A$ such that ${\mathfrak{g}}\cong \bigoplus_{n<0} A^{\oplus k_n}[n]$ as a graded $A$-module. Then the following hold:
1. $C^*_{\mathrm{poly}}({\mathfrak{g}})$ is a cofibrant commutative dg-$k$-algebra.
2. the map on Kähler differentials (relative to the base cdga $k$) $$\xymatrix{
\Omega_{C^*_{\mathrm{poly}}({\mathfrak{g}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{g}})} A\ar[r] & \Omega_A
}$$ can be identified with the projection map $\Omega_A\oplus {\mathfrak{g}}[1]^\vee{\rightarrow}\Omega_A$. Here $\Omega_A\oplus {\mathfrak{g}}[1]^\vee$ is the mapping fiber of the dual of the anchor map, i.e. it has differential $$\xymatrix{
{\partial}\big(d_{\mathrm{dR}}(a), \alpha\big) = \Big(d_{\mathrm{dR}}({\partial}_Aa), {\partial}_{{\mathfrak{g}}[1]^\vee}(\alpha) + \rho^\vee(d_{\mathrm{dR}}a)\Big)
}$$ where $\rho^\vee\colon \Omega_A {\rightarrow}{\mathfrak{g}}^\vee$ is the adjoint of the anchor map ${\mathfrak{g}}{\rightarrow}T_A$.
Since ${\mathfrak{g}}$ is given by $\bigoplus_{i<0} A^{\oplus k_i}[i]$ as a graded $A$-module, $C^*_{\mathrm{poly}}({\mathfrak{g}})$ is a polynomial algebra over $A$, generated by the free module ${\mathfrak{g}}[1]^\vee$. This module has generators in degrees $\geq 0$ and $A$ is cofibrant, so that $C^*_{\mathrm{poly}}({\mathfrak{g}})$ is the retract of a connective graded polynomial ring over $k$. This implies that $C^*_{\mathrm{poly}}({\mathfrak{g}})$ is cofibrant.
For the second assertion, note that $C^*_{\mathrm{poly}}({\mathfrak{g}})$ is freely generated over $A$ by the graded $A$-module ${\mathfrak{g}}[1]^\vee$. It follows that $$\Omega_{C^*_{\mathrm{poly}}({\mathfrak{g}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{g}})} A\cong \Omega_A\oplus {\mathfrak{g}}[1]^\vee$$ as a graded $A$-module. The map $C^*_{\mathrm{poly}}({\mathfrak{g}}){\rightarrow}A$ sends all ${\mathfrak{g}}[1]^\vee$ to zero, so that the induced map on Kähler differentials is just the projection $\Omega_A\oplus {\mathfrak{g}}[1]^\vee{\rightarrow}\Omega_A$. Furthermore, the Chevalley-Eilenberg differential is given by $$\begin{aligned}
A\ni a &\; \mapsto {\partial}_A a + \rho^\vee(d_{\mathrm{dR}}a)\\
{\mathfrak{g}}[1]^\vee \ni \alpha &\; \mapsto {\partial}_{{\mathfrak{g}}[1]^\vee}\alpha \quad \mod \big({\mathfrak{g}}[1]^\vee\big)^2.\end{aligned}$$ This shows that the differential on $ \Omega_A\oplus {\mathfrak{g}}[1]^\vee$ is given as in the lemma.
\[fm:lem:addingvariablestopowerseries\] Let $A$ be a nonnegatively graded commutative ${\mathbb{Q}}$-algebra, let $V$ be a finite dimensional ${\mathbb{Q}}$-vector space and let $W$ be a degreewise finite-dimensional graded ${\mathbb{Q}}$-vector space, concentrated in degrees $<0$. Then there is a natural isomorphism of graded-commutative $A$-algebras $$\xymatrix{
{{\mathrm{Hom}}}_{\mathbb{Q}}\big({\mathrm{Sym}}_{\mathbb{Q}}V, A\big)\otimes_{\mathbb{Q}} {\mathrm{Sym}}_{\mathbb{Q}}(W^\vee)\ar[r] & {{\mathrm{Hom}}}_{\mathbb{Q}}\big({\mathrm{Sym}}_{\mathbb{Q}}(V\oplus W), A\big)
}$$ where ${\mathrm{Sym}}_{\mathbb{Q}} W^\vee$ is the graded polynomial algebra on the dual vector space of $W$.
Observe that there is an isomorphism of graded cocommutative coalgebras ${\mathrm{Sym}}_{\mathbb{Q}}(V\oplus W)\cong {\mathrm{Sym}}_{\mathbb{Q}}V\otimes_{{\mathbb{Q}}} {\mathrm{Sym}}_{\mathbb{Q}}W$. There is a natural map of graded-commutative algebras $$\xymatrix{
{{\mathrm{Hom}}}_{\mathbb{Q}}\big({\mathrm{Sym}}_{\mathbb{Q}}V, A\big)\otimes_{\mathbb{Q}} {{\mathrm{Hom}}}_{\mathbb{Q}}({\mathrm{Sym}}_{\mathbb{Q}} W, {\mathbb{Q}})\ar[r]^-{\mu} & {{\mathrm{Hom}}}_{\mathbb{Q}}\big({\mathrm{Sym}}_{\mathbb{Q}}V\otimes_{{\mathbb{Q}}} {\mathrm{Sym}}_{\mathbb{Q}} W, A\big).
}$$ sending two maps $\alpha\colon {\mathrm{Sym}}_{\mathbb{Q}}V{\rightarrow}A$ and $\beta\colon {\mathrm{Sym}}_{\mathbb{Q}} W{\rightarrow}{\mathbb{Q}}$ to $\alpha\otimes \beta$. Using that ${\mathrm{Sym}}_{\mathbb{Q}} W$ is free on generators of degrees $<0$, with finitely many generators in each degree, one can identify ${\mathrm{Sym}}_{\mathbb{Q}}(W^\vee)\simeq {{\mathrm{Hom}}}_{\mathbb{Q}}({\mathrm{Sym}}_{\mathbb{Q}} W, {\mathbb{Q}})$. Using this, it follows that $\mu$ is an isomorphism.
\[fm:lem:polynomialintopower\] Let $A$ be a cofibrant connective commutative dg-$k$-algebra and let ${\mathfrak{g}}$ be as in Proposition \[fm:prop:koszulduality\]. Then the map $C^*_{\mathrm{poly}}({\mathfrak{g}}){\rightarrow}C^*({\mathfrak{g}})$ induces an equivalence on cotangent complexes over $k$ $$\xymatrix{
L_{C^*_{\mathrm{poly}}({\mathfrak{g}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{g}})} A\ar[r]^-\simeq & L_{C^*({\mathfrak{g}})}\otimes_{C^*({\mathfrak{g}})} A.
}$$
Consider the trivial cofibration, followed by a fibration $$\xymatrix{
0\ar[r]^-{\sim} & {\mathfrak{h}}=F({\mathfrak{g}}[0,-1])\ar@{->>}[r] & {\mathfrak{g}}
}$$ where ${\mathfrak{h}}$ is the free dg-Lie algebroid on the map ${\mathfrak{g}}[0, -1]{\rightarrow}{\mathfrak{g}}{\rightarrow}T_A$ from the path space of ${\mathfrak{g}}$. Let $V$ be the free graded ${\mathbb{Q}}$-vector space spanned by the generators $x_i$ of ${\mathfrak{g}}$, so that ${\mathfrak{g}}=A\otimes_{\mathbb{Q}} V$. As a graded Lie algebroid, ${\mathfrak{h}}$ is then freely generated by the graded ${\mathbb{Q}}$-vector space $V[0, -1]$. Consequently, the map ${\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$ is given without differentials by the $A$-linear extenion of a map from the free Lie algebra $$\xymatrix{
{\mathfrak{h}}=A\otimes{\mathrm{Lie}}\big(V[0, -1]\big)\ar[r] & A\otimes V= {\mathfrak{g}}
}$$ which sends $V$ to the generators of ${\mathfrak{g}}$ and $V[-1]$ to zero. This map has a splitting, induced by the inclusion $V{\rightarrow}{\mathrm{Lie}}(V[0, -1])$, so that ${\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$ can be identified with $$\xymatrix{
{\mathfrak{h}} = A\otimes (V\oplus W)\ar[r]^-{({\mathrm{id}}, 0)} & A\otimes V = {\mathfrak{g}}.
}$$ Here $W$ is a graded ${\mathbb{Q}}$-vector space isomorphic to ${\mathrm{Lie}}(V[0, -1])/V$, which is degreewise finite dimensional and concentrated in degrees $<-1$. Indeed, $V$ is degreewise finite dimensional and concentrated in degrees $<0$, so that ${\mathrm{Lie}}(V[0, -1])$ has these properties as well and $V{\rightarrow}{\mathrm{Lie}}(V[0, -1])$ is an isomorphism in degree $-1$.
Let us now consider the commutative diagram of cdgas associated to ${\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$ $$\label{fm:diag:thecrucialpushout}\vcenter{\xymatrix{
C^*_{\mathrm{poly}}({\mathfrak{g}})\ar[d]\ar[r] & C^*_{\mathrm{poly}}({\mathfrak{h}})\ar[d]\ar[r] & A\ar[d]\\
C^*({\mathfrak{g}})\ar[r] & C^*({\mathfrak{h}})\ar[r]_-\sim & A.
}}$$ The right bottom map is a weak equivalence since ${\mathfrak{h}}$ is cofibrant and weakly contractible. The map $C^*_{\mathrm{poly}}({\mathfrak{g}}){\rightarrow}C^*_{\mathrm{poly}}({\mathfrak{h}})$ can be identified with a map of polynomial algebras $$\xymatrix{
A\otimes_{\mathbb{Q}} {\mathrm{Sym}}_{\mathbb{Q}}(V[1]^\vee)\ar[r] & A\otimes_{\mathbb{Q}} {\mathrm{Sym}}_{\mathbb{Q}}\big((V\oplus W)[1]^\vee).
}$$ It follows that $C^*_{\mathrm{poly}}({\mathfrak{h}})$ is freely generated over $C^*_{\mathrm{poly}}({\mathfrak{g}})$ by $W[1]^\vee$, which is degreewise finite dimensional and concentrated in degrees $\geq 1$. In particular, $C^*_{\mathrm{poly}}({\mathfrak{g}}){\rightarrow}C^*_{\mathrm{poly}}({\mathfrak{h}})$ is a cofibration of cdgas.
On the other hand, the map $C^*({\mathfrak{g}}){\rightarrow}C^*({\mathfrak{h}})$ is given without differentials by the natural map $$\smash{\xymatrix{
{{\mathrm{Hom}}}_{\mathbb{Q}}\Big({\mathrm{Sym}}_{\mathbb{Q}} V[1], A\Big)\ar[r] & {{\mathrm{Hom}}}_{\mathbb{Q}}\Big({\mathrm{Sym}}_{\mathbb{Q}} (V[1]\oplus W[1]), A\Big).
}}$$ It now follows from Lemma \[fm:lem:addingvariablestopowerseries\] that the left square in is a (homotopy) pushout square of cdgas. Its image under $L_{(-)}\otimes_{(-)} A$ $$\label{fm:diag:pushoutofcotcomplexes}\vcenter{\xymatrix{
L_{C^*_{\mathrm{poly}}({\mathfrak{g}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{g}})} A\ar[d]\ar[r] & L_{C^*_{\mathrm{poly}}({\mathfrak{h}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{h}})} A\ar[d]\\
L_{C^*({\mathfrak{g}})}\otimes_{C^*({\mathfrak{g}})} A\ar[r] & L_{C^*({\mathfrak{h}})}\otimes_{C^*({\mathfrak{h}})} A
}}$$ is a homotopy pushout square as well. Since $C^*({\mathfrak{h}}){\rightarrow}A$ is a weak equivalence, the map $L_{C^*({\mathfrak{h}})}{\rightarrow}L_A$ is a weak equivalence. On the other hand, the map $L_{C^*_{\mathrm{poly}}({\mathfrak{h}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{h}})} A{\rightarrow}L_A$ is identified with the projection map $$\xymatrix{
\Omega_A\oplus {\mathfrak{h}}[1]^\vee\ar[r] & \Omega_A
}$$ by Lemma \[fm:lem:cotangentforglobalce\]. The kernel of this map is contractible, since ${\mathfrak{h}}$ is a cofibrant contractible dg-$A$-module. It follows that the right vertical map in Diagram is an equivalence, so that the left map is an equivalence as well.
By Lemma \[lem:identificationunderadjunction\], it suffices to show that the map ${\mathfrak{g}}{\rightarrow}L_{A/C^*}^\vee$ is adjoint to a weak equivalence $L_{A/C^*({\mathfrak{g}})}{\rightarrow}{\mathfrak{g}}^\vee$. This map fits into a sequence of maps $$\xymatrix{
L_{A/C^*_{\mathrm{poly}}({\mathfrak{g}})} \ar[r] & L_{A/C^*({\mathfrak{g}})} \ar[r] & {\mathfrak{g}}^\vee
}$$ classifying the composite map of commutative dg-algebras over $A$ $$\xymatrix{
C^*_{\mathrm{poly}}({\mathfrak{g}}) \ar[r] & C^*({\mathfrak{g}})\ar[r]^c & A^{\mathfrak{g}}
}$$ where $c$ is as in . The map $L_{A/C^*_{\mathrm{poly}}({\mathfrak{g}})}{\rightarrow}L_{A/C^*({\mathfrak{g}})}$ is an equivalence by Lemma \[fm:lem:polynomialintopower\], so it suffices to show that $L_{A/C^*_{\mathrm{poly}}({\mathfrak{g}})} {\rightarrow}{\mathfrak{g}}^\vee$ is an equivalence. This map can be computed explicitly: the map $$\xymatrix{
C_{\mathrm{poly}}^*({\mathfrak{g}})\ar[r] & A^{\mathfrak{g}}=A\oplus_{\rho^\vee} {\mathfrak{g}}[1]^\vee
}$$ is simply the quotient of $C^*_{\mathrm{poly}}({\mathfrak{g}})$ by the augmentation ideal $({\mathfrak{g}}[1]^\vee)^2$. Unwinding the definitions, e.g. using the pullback square , one finds the following description of the classifying map $L_{A/C^*_{\mathrm{poly}}({\mathfrak{g}})} {\rightarrow}{\mathfrak{g}}^\vee$: it is the canonical map from the mapping cone of $$\xymatrix{
\Omega_{C^*_{\mathrm{poly}}({\mathfrak{g}})}\otimes_{C^*_{\mathrm{poly}}({\mathfrak{g}})} A \cong \Omega_A\oplus {\mathfrak{g}}[1]^\vee\ar[r] & \Omega_A
}$$ to ${\mathfrak{g}}^\vee$. This map is a weak equivalence, which concludes the proof.
Formal moduli problems {#sec:formmod}
======================
We will now use the results of Section \[sec:koszulduality\] to establish an equivalence between Lie algebroids and formal moduli problems (Definition \[def:formalmoduli\]):
\[thm:formalmoduliperfect\] Let $A\in {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{{{\mathrm{dg}}}, {{\scriptscriptstyle \geq}}0}_k$ be cofibrant. Then there is an adjoint pair of functors $$\xymatrix{
{\mathrm{MC}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\ar@<1ex>[r] & {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A\ar@<1ex>[l] \colon T_{A/}
}$$ which is an equivalence whenever $A$ is eventually coconnective.
\[thm:formalmodulicoherent\] Suppose that $A\in {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{{{\mathrm{dg}}}, {{\scriptscriptstyle \geq}}0}_k$ is cofibrant, bounded and coherent. Then there is an equivalence of $\infty$-categories $$\xymatrix{
{\mathrm{MC}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A\ar@<1ex>[r] & {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}^!_A\ar@<1ex>[l] \colon T_{A/}
}$$ between the $\infty$-category of pro-coherent Lie algebroids over $A$ and the $\infty$-category of pro-coherent formal moduli problems ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{{\mathrm{sm, coh}}}/A{\rightarrow}{{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}$ (Section \[sec:intro\]).
These theorems follow formally from Corollary \[cor:koszuldualityperfect\] and \[cor:koszuldualitycoherent\], by means of a general procedure due to Lurie [@lur11X] that we will briefly recall.
Generators
----------
Categories of chain complexes or spectra endowed with a certain algebraic structure often admit a presentation in terms of generators and relations.
\[def:goodobjects\] Let $\Xi$ be a locally presentable $\infty$-category equipped with a collection of right adjoint functors $$\smash{\xymatrix{
e_\alpha\colon \Xi\ar[r] & {
\StrLen{Sp}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Sp}
\else \mathrm{Sp}
\fi}
}}$$ to the $\infty$-category of spectra. The left adjoint to $e_\alpha$ sends the map ${\mathbb{S}}^n{\rightarrow}0$ in ${
\StrLen{Sp}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Sp}
\else \mathrm{Sp}
\fi}$ to a map in $\Xi$ that we will denote by $K_{\alpha, n}{\rightarrow}\emptyset$. We will say that an object $X\in \Xi$ is *good* if it admits a finite filtration $$\smash{\xymatrix{
\emptyset=X^{(0)}\ar[r] & X^{(1)}\ar[r] & \dots\ar[r] & X^{(n)}
}}$$ where for each $i$, there is an $\alpha$ and a $n\leq -2$, together with a pushout square $$\label{diag:addingcell}\vcenter{\xymatrix{
K_{\alpha, n}\ar[r]\ar[d] & X^{(i-1)}\ar[d]\\
\emptyset\ar[r] & X^{(i)}.
}}$$ Let $\Xi^{\mathrm{good}}\subseteq \Xi$ be the full subcategory on the good objects; it is the smallest subcategory of $\Xi$ which contains $\emptyset$ and is closed under pushouts along the maps $K_{\alpha, n}{\rightarrow}\emptyset$ with $n\leq -2$.
\[prop:luriedefthy\] Let $(\Xi, e_\alpha)$ be as in Definition \[def:goodobjects\]. Suppose that each $e_\alpha$ preserves small sifted homotopy colimits and that a map $f$ in $\Xi$ is an equivalence if and only if each $e_\alpha(f)$ is an equivalence of spectra. Then the right adjoint functor $$\xymatrix{
\Psi\colon \Xi\ar[r] & {\mathrm{PSh}}(\Xi^{\mathrm{good}}); \hspace{4pt} X\ar@{|->}[r] & {{\mathrm{Map}}}_{\Xi}(-, X)
}$$ is fully faithful, with essential image consisting of those (space-valued) presheaves $F$ satisfying the following two conditions:
- $F(\emptyset)$ is contractible.
- For any $\alpha$ and $n\leq -2$, $F$ sends a pushout square of the form to a pullback square of spaces.
This is exactly [@lur11X Theorem 1.3.12], replacing the category $\Upsilon^{\mathrm{sm}}$ from loc. cit. by the opposite of $\Xi^{\mathrm{good}}$.
\[ex:obvious1\] Let $\Xi$ be a compactly generated stable $\infty$-category. The collection of functors ${{\mathrm{Hom}}}_{\Xi}(K, -)\colon \Xi{\rightarrow}{
\StrLen{Sp}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Sp}
\else \mathrm{Sp}
\fi}$, for all compact objects $K$, satisfies the condition of Proposition \[prop:luriedefthy\]. The good objects are exactly the compact objects and a functor $F\colon \Xi^{\omega, {{\mathrm{op}}}}{\rightarrow}{{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}$ satisfies conditions (a) and (b) if and only if it is left exact. Proposition \[prop:luriedefthy\] then reproduces the equivalence $$\Xi\simeq {\mathrm{Ind}}(\Xi^\omega).$$
\[ex:obvious2\] Let $A$ be a commutative dg-algebra (over a field of characteristic zero) and let ${{\mathrm{Mod}}}_A$ be the $\infty$-category of $A$-modules. The single functor $e\colon {{\mathrm{Mod}}}_A{\rightarrow}{
\StrLen{Sp}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Sp}
\else \mathrm{Sp}
\fi}$, forgetting the $A$-module structure, satisfies the conditions of Proposition \[prop:luriedefthy\]. In this case, the good $A$-modules can be presented by the dg-$A$-modules whose underlying graded $A$-module is free on finitely many generators $x_i$ of degree $<0$.
There is an equivalence of $\infty$-categories $$\xymatrix{
{{\mathrm{Mod}}}_A^{\mathrm{good, op}}\ar[r] & {{\mathrm{Mod}}}_A^{{\mathrm{f.p.}}, {{\scriptscriptstyle \geq}}0}; \hspace{4pt} E\ar[r] & E[1]^\vee
}$$ to the $\infty$-category of finitely presented connective $A$-modules, i.e. dg-$A$-modules generated by finitely many generators of degree $\geq 0$. Combining this equivalence with Proposition \[prop:luriedefthy\], one finds that ${{\mathrm{Mod}}}_A$ is equivalent to the $\infty$-category of functors $$\xymatrix{
F\colon {{\mathrm{Mod}}}_A^{{\mathrm{f.p.}}, {{\scriptscriptstyle \geq}}0}\ar[r] & {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}}$$ that send $0$ to a contractible space and preserve pullbacks along the maps $0{\rightarrow}A[n]$ with $n\geq 1$.
Proof of Theorem \[thm:formalmoduliperfect\]
--------------------------------------------
Let $A$ be a cofibrant commutative dg-$k$-algebra. By Proposition \[prop:monadicity\] and Example \[ex:obvious2\], the composite forgetful functor $$\xymatrix{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\ar[r]^U & {{\mathrm{Mod}}}_A/T_A\ar[r]^-{\ker} & {{\mathrm{Mod}}}_A\ar[r]^-{e} & {
\StrLen{Sp}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Sp}
\else \mathrm{Sp}
\fi}
}$$ preserves small limits and sifted colimits, and detects equivalences.
The corresponding notion of a *good Lie algebroid* can be described in terms of the (projective) model structure on dg-Lie algebroids as follows: let us say that a dg-Lie algebroid ${\mathfrak{g}}$ is *very good* if it admits a finite sequence of cofibrations $$\xymatrix{
0={\mathfrak{g}}^{(0)}\ar[r] & \cdots \ar[r] & {\mathfrak{g}}^{(n)}={\mathfrak{g}},
}$$ each of which is the pushout of a (generating) cofibration with $n_i\leq -2$ $$\label{fm:rem:verygoodliealgebroids}\xymatrix{
{\mathrm{Free}}\big({\partial}\phi\colon A[n_i] {\rightarrow}T_A\big)\ar[r] & {\mathrm{Free}}\big(\phi\colon A[n_i, n_i+1]{\rightarrow}T_A\big).
}$$ Here $\phi$ is a map from the cone of $A[n_i]$ to $T_A$, which is determined uniquely by a degree $(n_i+1)$ element of $T_A$. Then the good Lie algebroids can be presented by the very good dg-Lie algebroids over $A$.
Good Lie algebroids may have nontrivial anchor maps, even though the Lie algebroids in have null-homotopic anchor maps.
\[fm:lem:verygoodlralgebroids\] Let ${\mathfrak{g}}$ be a very good dg-Lie algebroid over $A$. Then the following hold:
1. ${\mathfrak{g}}$ has a cofibrant underlying dg-$A$-module.
2. Without the differential, ${\mathfrak{g}}$ is freely generated by a negatively graded finite-dimensional vector space over $T_A$.
3. ${\mathfrak{g}}$ is isomorphic as a graded $A$-module to $\bigoplus_{n<0} A^{\oplus k_n}[n]$ for some sequence of $k_n\in {\mathbb{N}}_{\geq 0}$.
Assertion (1) is obvious and (3) follows immediately from (2). For (2), note that each pushout along a map freely adds a single generator of degree $<0$ at the level of graded Lie algebroids.
Let us denote the $\infty$-category of presheaves satisfying conditions (a) and (b) of Proposition \[prop:luriedefthy\] by $${
\StrLen{E}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{E}
\else \mathrm{E}
\fi}\subseteq {\mathrm{PSh}}({{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{\mathrm{good}}).$$ We then have an equivalence ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\simeq {
\StrLen{E}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{E}
\else \mathrm{E}
\fi}$, so that it suffices to produce the required adjunction (equivalence) between ${
\StrLen{E}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{E}
\else \mathrm{E}
\fi}$ and the $\infty$-category of formal moduli problems. To this end, recall that the good Lie algebroids form the smallest subcategory of ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A$ that contains $0$ and is closed under pushouts against the maps $$\label{diag:generatinggoodmaps}\xymatrix{
F(0\colon A[n]{\rightarrow}T_A)\ar[r] & 0 & n\leq -2.
}$$ The functor $C^*$ preserves colimits (Proposition \[prop:deformationadjunction\]) and sends the above maps to the maps $({\mathrm{id}}, 0)\colon A{\rightarrow}A\oplus A[-1-n]$, for $n\leq -2$. It follows that $C^*$ restricts to a functor $$\label{diag:cerestrictedtogood}\xymatrix{
C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{\mathrm{good}}\ar[r] & \Big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{sm}}_k/A\Big)^{{\mathrm{op}}}.
}$$ The restriction of a formal moduli problem along $C^*$ is a presheaf contained in ${
\StrLen{E}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{E}
\else \mathrm{E}
\fi}$. We therefore obtain a right adjoint functor $$\xymatrix{
T_{A/}\colon {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A\ar[r]^-{(C^*)^*} & {
\StrLen{E}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{E}
\else \mathrm{E}
\fi}\ar[r]^-{\simeq} & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A.
}$$ If $A$ is eventually coconnective, Corollary \[cor:koszuldualityperfect\] and Lemma \[fm:lem:verygoodlralgebroids\] show that is fully faithful. The essential image of $C^*$ contains $A$ and is closed under pullbacks along the maps $({\mathrm{id}}, 0)\colon A{\rightarrow}A\oplus A[n]$ for $n\geq 1$. Indeed, such pullbacks can dually be computed as pushouts along the images of the maps . The small extension of $A$ form the smallest subcategory of ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A$ with these two closure properties, so is essentially surjective as well. This implies that the functor $T_{A/}$ is an equivalence.
\[ex:tangenttomapofaffines\] Let $f\colon B{\rightarrow}A$ be a map of connective commutative $k$-algebras and let $$\xymatrix{
{{\mathrm{Spec}}}(B)^\wedge\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A \ar[r] & {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}; \hspace{4pt} \tilde{A}\ar@{|->}[r] & {{\mathrm{Map}}}(B, \tilde{A})\times_{{{\mathrm{Map}}}(B, A)} \{f\}
}$$ be the formal completion of ${{\mathrm{Spec}}}(B)$ at $A$. This is a formal moduli problem over $A$. Unwinding the definitions, one sees that the associated Lie algebroid is given by ${\mathfrak{D}}(B)$, which can be identified with the Lie algebroid $T_{A/B}{\rightarrow}T_A$ of (derived) $B$-linear derivations of $A$ by Remark \[fm:rem:dualityfunctor\].
\[rem:mcintermsofD\] Suppose that $A$ is eventually coconnective, so that ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A$ is equivalent to the $\infty$-category of good Lie algebroids over $A$. If ${\mathfrak{g}}$ is a Lie algebroid over $A$, then the formal moduli problem ${\mathrm{MC}}_{\mathfrak{g}}$ is given (up to a natural equivalence) by $${\mathrm{MC}}_{\mathfrak{g}}(B) = {{\mathrm{Map}}}_{{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A}\big({\mathfrak{D}}(B), {\mathfrak{g}}\big).$$ By Remark \[ex:tangenttomapofaffines\], one can think of this as the space of flat ${\mathfrak{g}}$-valued connections on the fiberwise tangent bundle of $A$ over $B$.
Proof of Theorem \[thm:formalmodulicoherent\] {#sec:tamecase}
---------------------------------------------
Theorem \[thm:formalmodulicoherent\] is proven in exactly the same way: let ${\mathcal{K}}^{{{\scriptscriptstyle \leq}}0}\subseteq {{\mathrm{Mod}}}_A^!$ be the full subcategory on the compact generators provided by Proposition \[prop:tamecompactlygenerated\]. For each $K_\alpha\in {\mathcal{K}}$, there is a functor $$\xymatrix{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!\ar[r]^U & {{\mathrm{Mod}}}_A^!/T_A\ar[r]^-\ker & {{\mathrm{Mod}}}_A^!\ar[rr]^-{{{\mathrm{Hom}}}_A(K_\alpha, -)} & & {
\StrLen{Sp}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Sp}
\else \mathrm{Sp}
\fi}
}$$ which preserves limits and sifted colimits by Proposition \[prop:monadicitytame\]. All these functor collectively detect equivalences between tame Lie algebroids.
The good tame Lie algebroids then have the following point-set description: they are given by those dg-Lie algebroids obtained from finitely many pushouts along $$\label{fm:rem:verygoodtameliealgebroids}\xymatrix{
{\mathrm{Free}}\big({\partial}\phi\colon K[-2] {\rightarrow}T_A\big)\ar[r] & {\mathrm{Free}}\big(\phi\colon K[-2, -1]{\rightarrow}T_A\big)
}$$ with $K\in{\mathcal{K}}$. The assertions of Lemma \[fm:lem:verygoodlralgebroids\] apply to such dg-Lie algebroids as well, so that Corollary \[cor:koszuldualitycoherent\] provides fully faithful functor $$\xymatrix{
C^*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!, {\mathrm{good}}}\ar[r] & \Big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k/A\Big)^{{\mathrm{op}}}.
}$$ The images of the maps under $C^*$ are exactly the maps $A{\rightarrow}A\oplus E[1]$, where $E$ is a connective coherent $A$-module. Arguing as in the proof of Theorem \[thm:formalmoduliperfect\], it follows that $C^*$ induces an equivalence between good tame Lie algebroids over $A$ and the $\infty$-category ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A$ of small extensions of $A$ by coherent $A$-modules.
Restriction along this equivalence provides an equivalence between ${
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A^!$ and the $\infty$-category of presheaves on ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!, {\mathrm{good}}}$ satisfying conditions (a) and (b) from Proposition \[prop:luriedefthy\]. In turn, the $\infty$-category of such presheaves is equivalent to ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ by Proposition \[prop:luriedefthy\].
Representations and quasi-coherent sheaves {#sec:representations}
==========================================
In the previous section, we have seen that there is an equivalence between formal moduli problems over $A$ and Lie algebroids over $A$. Geometrically, one can think of a formal moduli problem as a map of stacks $x\colon {{\mathrm{Spec}}}(A){\rightarrow}X$ that realizes $X$ as an infinitesimal thickening of ${{\mathrm{Spec}}}(A)$. Any such stack $X$ gives rise to an $\infty$-category of *quasi-coherent sheaves* on $X$ in the usual way (Definition \[def:qc\]): a quasi-coherent sheaf $F$ is given by a collection of $B$-modules $F_y$ for every $B\in {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A$ and every $y\in X(B)$, together with a coherent family of equivalences $$F_{\alpha(y)}\simeq B'\otimes_B F_y$$ for every $\alpha\colon B{\rightarrow}B'$. In particular, a quasi-coherent sheaf $F$ on $X$ determines an $A$-module $F_x$, by restricting to the canonical point $x\in X(A)$. We will see that $F_x$ carries a representation of $T_{A/X}$. In fact, we will prove the following:
\[fm:thm:quasi-coherentvsrepresentations\] Let $A$ be eventually coconnective and let $X$ be a formal moduli problem over $A$ with associated Lie algebroid $T_{A/X}$. Then there is a fully faithful, symmetric monoidal left adjoint functor $$\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}(X)\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}
}$$ where $\Psi_X(F)$ has underlying $A$-module given by the restriction $F_x$ to the canonical basepoint $x\in X(A)$. Furthermore, the functor $\Psi_X$ induces an equivalence $${{\mathrm{Mod}}}(X)^{{{\scriptscriptstyle \geq}}0}\simeq {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}^{{{\scriptscriptstyle \geq}}0}$$ between the connective quasi-coherent sheaves on $X$ and the $T_{A/X}$-representations whose underlying $A$-module is connective.
Our proof closely follows the discussion in [@lur11X Section 2.4]. We will begin by considering representations of good Lie algebroids in Section \[fm:sec:representationsofgood\]. In fact, for later purposes it will be convenient to work with the tame model structure on ${\mathfrak{g}}$-representations. Theorem \[fm:thm:quasi-coherentvsrepresentations\] then follows from a formal argument described in Section \[fm:sec:qcoh\].
Representations of good Lie algebroids {#fm:sec:representationsofgood}
--------------------------------------
Recall from Remark \[rem:celaxmonoidal\] that there is a lax monoidal functor sending a ${\mathfrak{g}}$-representation $E$ to its Chevalley-Eilenberg complex $C^*({\mathfrak{g}}, E)$. This functor is part of an adjoint pair $$\label{diag:koszulquillenpair}\xymatrix{
K({\mathfrak{g}})\otimes_{C^*({\mathfrak{g}})} (-)\colon {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_{C^*({\mathfrak{g}})} \ar@<1ex>[r] & {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}\colon C^*({\mathfrak{g}}, -)\ar@<1ex>[l]
}$$ were $K({\mathfrak{g}})$ is the Koszul complex of ${\mathfrak{g}}$ (see Remark \[fm:rem:cofibersequenceoncotangent\]).
When ${\mathfrak{g}}$ has a graded-projective underlying dg-$A$-module, this adjoint pair is a Quillen pair between the *projective* model structure on dg-$C^*({\mathfrak{g}})$-modules and the *tame* model structure on ${\mathfrak{g}}$-representations, transferred from the tame model structure on dg-$A$-modules. We let $\Phi_{\mathfrak{g}}\colon {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}$ denote the left derived functor between $\infty$-categories. One can think of $\Phi_{\mathfrak{g}}$ as the functor $F\mapsto A\otimes_{C^*({\mathfrak{g}})} F$, since $K({\mathfrak{g}})$ is equivalent to the ${\mathfrak{g}}$-representation $A$.
When the underlying dg-$A$-module of ${\mathfrak{g}}$ is projectively cofibrant, the adjoint pair is also a Quillen pair for the projective model structure on dg-$A$-modules. In terms of $\infty$-categories, this means that the functor $\Phi_{\mathfrak{g}}$ factors as $$\xymatrix{
\Phi_{\mathfrak{g}}\colon {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}\ar@{^{(}->}[r]& {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^!.
}$$
\[fm:lem:fullyfaithfulonqcoh\] Suppose that ${\mathfrak{g}}\in {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ is compact. Then $\Phi_{\mathfrak{g}}\colon {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{{\mathfrak{g}}}$ is fully faithful.
Let ${\mathcal{C}}$ be the class of objects in ${{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}$ for which the derived unit map is an equivalence. Then ${\mathcal{C}}$ is closed under finite colimits and retracts and contains $C^*({\mathfrak{g}})$ because $\Phi_{\mathfrak{g}}(C^*({\mathfrak{g}}))\simeq A$. It follows that the unit map is an equivalence for all compact $C^*({\mathfrak{g}})$-modules. Since ${{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}\simeq {\mathrm{Ind}}({{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^\omega)$ is the ind-completion of the category of compact $C^*({\mathfrak{g}})$-modules, it suffices to show that the right adjoint $C^*({\mathfrak{g}}, -)$ preserves filtered colimits. Equivalently, it suffices to verify that $\Phi_{\mathfrak{g}}(C^*({\mathfrak{g}}))\simeq A$ is a compact object of ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$.
By Remark \[fm:rem:cofibersequenceoncotangent\], $A$ fits into a cofiber sequence of ${\mathcal{U}}({\mathfrak{g}})$-modules $$\xymatrix{
L_{\mathfrak{g}}\ar[r] & {\mathcal{U}}({\mathfrak{g}})\ar[r] & A.
}$$ It suffices to show that $L_{\mathfrak{g}}$ is compact. This follows from the fact that the cotangent complex functor preserves compact objects, since its right adjoint (taking square zero extensions) preserves filtered colimits, which are computed at the level of tame dg-$A$-modules (Proposition \[prop:monadicitytame\]).
\[fm:lem:generatorsforconnectivemodules\] Suppose that ${\mathfrak{g}}\in {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ admits a finite filtration $$\xymatrix{
0={\mathfrak{g}}^{(0)}\ar[r] & \dots \ar[r] & {\mathfrak{g}}^{(n)}={\mathfrak{g}}
}$$ with the following property: for each $i$, there is a dg-$A$-module $V$ contained in ${
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}^{{{\scriptscriptstyle \leq}}-2}$ (see Lemma \[lem:tamecofgenerated\]), such that ${\mathfrak{g}}^{(i-1)}{\rightarrow}{\mathfrak{g}}^{(i)}$ is a pushout of the map $F(V){\rightarrow}0$ from the free dg-Lie algebroid on $0\colon V{\rightarrow}T_A$.
Let $E$ be a dg-${\mathfrak{g}}$-representation whose underlying tame dg-$A$-module is connective (Remark \[rem:connectivemodules\]). Then there exists a map $\bigoplus_\alpha A {\rightarrow}E$ in the $\infty$-category ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}$ which induces a surjection on $\pi_0$.
Pick representatives $e_\alpha\in E$ for the generators of $\pi_0(E)$ and consider the associated map of ${\mathfrak{g}}$-representations $\bigoplus_\alpha {\mathcal{U}}({\mathfrak{g}}){\rightarrow}E$. This map is clearly surjective on $\pi_0$, so it suffices to prove that it factors (up to homotopy) as $$\xymatrix{
\bigoplus_\alpha {\mathcal{U}}({\mathfrak{g}})\ar[r] & \bigoplus_\alpha A\ar[r] & E.
}$$ Using the cofiber sequence $L_{\mathfrak{g}}{\rightarrow}{\mathcal{U}}({\mathfrak{g}}){\rightarrow}A$, we therefore have to provide a null-homotopy of each composite map $$\xymatrix{
L_{\mathfrak{g}}\ar[r] & {\mathcal{U}}({\mathfrak{g}})\ar[r]^{e_\alpha} & E.
}$$ By the assumption on ${\mathfrak{g}}$, the cotangent complex $L_{\mathfrak{g}}$ admits a filtration by ${\mathcal{U}}({\mathfrak{g}})$-modules $$\xymatrix{
0=L_{\mathfrak{g}}^{(0)}\ar[r] & \dots \ar[r] & L_{\mathfrak{g}}^{(n)}=L_{\mathfrak{g}}
}$$ where each $L_{\mathfrak{g}}^{(i-1)}{\rightarrow}L_{\mathfrak{g}}^{(i)}$ has cofiber of the form ${\mathcal{U}}({\mathfrak{g}})\otimes_A V$, with $V\in {
\StrLen{T}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{T}
\else \mathrm{T}
\fi}^{{{\scriptscriptstyle \leq}}-1}$. An inductive application of Remark \[rem:connectivemodules\] now shows that any map $L_{\mathfrak{g}}{\rightarrow}E$ to a connective ${\mathfrak{g}}$-representation is null-homotopic, which concludes the proof.
\[fm:cor:equivonconnmodules\] Suppose that $A$ is eventually coconnective and let ${\mathfrak{g}}\in{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A$ be good. Then $\Phi_{\mathfrak{g}}\colon{{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ is fully faithful and induces an equivalence $$\xymatrix{
\Phi_{\mathfrak{g}}\colon {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^{{{\scriptscriptstyle \geq}}0} \ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{g}}}^{{{\scriptscriptstyle \geq}}0}
}$$ between the full subcategories consisting of connective $C^*({\mathfrak{g}})$-modules and representations whose underlying $A$-module is connective.
\[fm:cor:equivonconnprocoherent\] Suppose that $A$ is bounded coherent and let ${\mathfrak{g}}\in{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ be good. Then $\Phi_{\mathfrak{g}}\colon{{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ is fully faithful and induces an equivalence $$\xymatrix{
\Phi_{\mathfrak{g}}\colon {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^{{{\scriptscriptstyle \geq}}0} \ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{g}}}^{!,{{\scriptscriptstyle \geq}}0}
}$$ between the full subcategories consisting of connective $C^*({\mathfrak{g}})$-modules and representations whose underlying pro-coherent sheaf over $A$ is connective.
Both assertions are proven in the same way. Since ${\mathfrak{g}}$ is good, it is compact in ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ (or ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A$), so that $\Phi_{\mathfrak{g}}$ is fully faithful by Lemma \[fm:lem:fullyfaithfulonqcoh\].
The subcategory ${{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^{{{\scriptscriptstyle \geq}}0}\subseteq {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}$ is the smallest subcategory of ${{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}$ which is closed under colimits and extensions and which contains $C^*({\mathfrak{g}})$. As a consequence, the essential image $${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}:= \Phi_{\mathfrak{g}}\Big({{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^{{{\scriptscriptstyle \geq}}0}\Big)\subseteq {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^!$$ under $\Phi_{\mathfrak{g}}$ is the smallest subcategory of ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}$ which is closed under colimits and extensions and which contains $A$. Clearly ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ is contained in ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{!, {{\scriptscriptstyle \geq}}0}$, so it suffices to prove the reverse inclusion.
To this end, let $E$ be a left ${\mathcal{U}}({\mathfrak{g}})$-module whose underlying (tame) dg-$A$-module is connective. We will inductively construct a sequence of left ${\mathcal{U}}({\mathfrak{g}})$-modules $0=E^{(-1)}{\rightarrow}E^{(0)}{\rightarrow}\cdots {\rightarrow}E$ such that each $E^{(n)}\in {
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ and such that each map $E^{(n)}{\rightarrow}E$ induces an isomorphism on homotopy groups in degrees $<n$ and a surjection on $\pi_n$. It follows that the map ${\operatornamewithlimits{\mathrm{colim}}}E^{(n)}{\rightarrow}E$ is a weak equivalence, so that $E\in {
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$.
To construct this sequence, suppose we have constructed $E^{(n-1)}$ and let $F$ be the fiber of the map $E^{(n-1)}{\rightarrow}E$. Then $F$ is a left ${\mathcal{U}}({\mathfrak{g}})$-module whose underlying (tame) dg-$A$-module is $(n-2)$-connective. Since ${\mathfrak{g}}$ is good, a shift of Lemma \[fm:lem:generatorsforconnectivemodules\] shows that there exists a map $\bigoplus_\alpha A[n-2]{\rightarrow}F$ which induces a surjection on $\pi_{n-2}$. Now let $E^{(n)}$ be the cofiber of the map $\bigoplus_\alpha A[n-2]{\rightarrow}F{\rightarrow}E^{(n-1)}$. This cofiber is contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ and the five lemma shows that $E^{(n)}{\rightarrow}E$ induces an isomorphism on homotopy groups in degrees $<n$ and a surjection on $\pi_n$.
Naturality
----------
Let us now address the functoriality of $\Phi_{\mathfrak{g}}$ in the dg-Lie algebroid ${\mathfrak{g}}$. This is somewhat delicate, because the Quillen adjunction $$\xymatrix{
K({\mathfrak{g}})\otimes_{C^*({\mathfrak{g}})} (-)\colon {{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_{C^*({\mathfrak{g}})} \ar@<1ex>[r] & {{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}}}_{\mathfrak{g}}\colon C^*({\mathfrak{g}}, -)\ar@<1ex>[l]
}$$ does not strictly intertwine restriction of ${\mathfrak{g}}$-representations with induction of $C^*({\mathfrak{g}})$-modules. However, this does become true at the level of $\infty$-categories. To see this, let us make the following definitions:
\[fm:con:smcatofdgmodules\] Let ${{\mathrm{Mod}}}^{{{\mathrm{dg}}}, \otimes}$ be the category in which
an object is a tuple $(B, M_1, \dots, M_m)$, where $B$ is a cdga and each $M_i$ is a dg-$B$-module.
a morphism $(B, M_1, \dots, M_m){\rightarrow}(C, N_1, \dots, N_n)$ is a map of finite pointed sets $\alpha\colon \big<m\big>{\rightarrow}\big<n\big>$, a map of cdgas $B{\rightarrow}C$ and $B$-multinear maps $\bigotimes^{\alpha(i)=j} M_i{\rightarrow}N_j$.
Similarly, let ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}, \otimes}$ be the category in which
an object is a tuple $({\mathfrak{g}}, E_1, \dots, E_m)$, where ${\mathfrak{g}}$ is a (tame) $A$-cofibrant dg-Lie algebroid and each $E_i$ is a ${\mathfrak{g}}$-representation.
a morphism $({\mathfrak{g}}, E_1, \dots, E_m){\rightarrow}({\mathfrak{h}}, F_1, \dots, F_n)$ is a map of finite pointed sets $\alpha\colon \big<m\big>{\rightarrow}\big<n\big>$, a map of dg-Lie algebroids ${\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$ and maps of ${\mathfrak{h}}$-representations $\bigotimes^{\alpha(i)=j}_A E_i{\rightarrow}F_j$.
These categories fit into a commuting square $$\label{diag:ceatdglevel}\vcenter{\xymatrix@R=1.7pc{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}, \otimes}\ar[d]\ar[r]^-{C^*} & {{\mathrm{Mod}}}^{{{\mathrm{dg}}}, \otimes}\ar[d]\\
\left({{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{{\mathrm{dg}}}, A-{\mathrm{cof}}}\right)^{{\mathrm{op}}}\times \Gamma^{{\mathrm{op}}}\ar[r]_-{C^*} & {{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{\mathrm{{{\mathrm{dg}}}}}}_k\times \Gamma^{{\mathrm{op}}}.
}}$$ The vertical functors are the obvious projections, which are both cocartesian fibrations. The top functor sends $({\mathfrak{g}}, E_1, \dots, E_m)$ to $\big(C^*({\mathfrak{g}}), C^*({\mathfrak{g}}, E_1), \dots, C^*({\mathfrak{g}}, E_m)\big)$.
After inverting the tame weak equivalences on the left and quasi-isomorphisms on the right of , one obtains a commuting square of $\infty$-categories $$\xymatrix@R=1.7pc{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{!, \otimes}\ar[r]^{C^*}\ar[d] & {{\mathrm{Mod}}}^\otimes\ar[d]\\
\big({{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!}\big)^{{{\mathrm{op}}}}\times\Gamma^{{\mathrm{op}}}\ar[r]_-{C^*} & {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k\times \Gamma^{{\mathrm{op}}}.
}$$ in which the vertical functors are cocartesian fibrations.
All functors in preserve weak equivalences, so that they descend to functors between localizations. It suffices to verify that the vertical projections remain cocartesian fibrations after inverting the quasi-isomorphisms. To see this, let ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\subseteq {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}, \otimes}$ be the full subcategory on $({\mathfrak{g}}, E_1, \dots, E_m)$ where each $E_i$ is tamely cofibrant as a dg-$A$-module. The inclusion ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\subseteq {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\mathrm{dg}}}, \otimes}$ induces an equivalence on localizations, with inverse provided by a cofibrant replacement functor. The projection $$\xymatrix{
{
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\ar[r] & \Big({{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{{\mathrm{dg}}}, A-{\mathrm{cof}}}\Big)^{{\mathrm{op}}}\times \Gamma^{{\mathrm{op}}}}$$ is a cocartesian fibration. For any map $\alpha\colon \big<m\big>{\rightarrow}\big<n\big>$ and any $f\colon {\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$, the induced functor between fibers is given by $$\xymatrix{
\left({{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{{{\mathrm{dg}}}, A-{\mathrm{cof}}}\right)^{\times m}\ar[r] & \left({{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{h}}^{{{\mathrm{dg}}}, A-{\mathrm{cof}}}\right)^{\times n}; \hspace{4pt} (E_j)_{j\leq m}\ar[r] & \left(\bigotimes_{\alpha(j)=i} f^*E_j\right)_{i\leq m}.
}$$ This functor preserves all quasi-isomorphisms and induces an equivalence of $\infty$-categories whenever $\alpha$ is a bijection and $f$ is a quasi-isomorphism. It follows from [@hin16 Proposition 2.1.4] that the induced functor of $\infty$-categories $$\smash{\xymatrix{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^\otimes\ar[r] & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\times\Gamma^{{\mathrm{op}}}}}$$ is a cocartesian fibration. A similar argument shows that ${{\mathrm{Mod}}}^\otimes{\rightarrow}{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k\times\Gamma^{{\mathrm{op}}}$ is a cocartesian fibration.
\[rem:cocartesianedges\] Fix a map $\alpha\colon \big<m\big>{\rightarrow}\big<n\big>$, a map $f\colon {\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$ and a collection $E_1, \dots, E_m$ of ${\mathfrak{g}}$-representations that are tamely cofibrant dg-$A$-modules. By [@hin16 Proposition 2.1.4], the cocartesian lift of $(f, \alpha)$ in ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^\otimes$ with domain $({\mathfrak{g}}, E_1, \dots, E_m)$ is the image of $$\xymatrix{
({\mathfrak{g}}, E_1, \dots, E_m)\ar[r] & \left({\mathfrak{h}}, \bigotimes_{\alpha(j)=1} f^*E_j, \dots, \bigotimes_{\alpha(j)=n} f^*E_j\right)
}$$ in the $\infty$-categorical localization of ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{{{\mathrm{dg}}}, \otimes}$.
\[fm:lem:functorialleftadjoint\] Let $\smash{{{\mathrm{Mod}}}^\otimes_{C^*}{\rightarrow}\big({{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A\big){}^{{\mathrm{op}}}\times\Gamma^{{\mathrm{op}}}}$ denote the base change of the projection ${{\mathrm{Mod}}}^\otimes {\rightarrow}{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k\times\Gamma^{{\mathrm{op}}}$ along the functor $C^*$, so that there is a functor of cocartesian fibrations $$\xymatrix@C=1.6pc@R=1.6pc{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{!, \otimes}\ar[rr]^{C^*}\ar[rd] & & {{\mathrm{Mod}}}^\otimes_{C^*}\ar[ld]\\
& \big({{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A\big)^{{\mathrm{op}}}\times\Gamma^{{\mathrm{op}}}. &
}$$ This functor admits a left adjoint $\Phi$, which preserves cocartesian edges.
For each $\big<m\big>$ and each dg-Lie algebroid ${\mathfrak{g}}$, the functor between the fibers $$\smash{\xymatrix{
C^*\colon \big({{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}\big){}^{\times m}\ar[r] & {{\mathrm{Mod}}}^{\times m}_{C^*({\mathfrak{g}})}
}}$$ admits a left adjoint $\Phi_{\mathfrak{g}}$: this is just the $m$-fold product of the left derived functor of the Quillen pair . By Remark \[rem:cocartesianedges\] and [@lur16 Proposition 7.3.2.11], the existence of the global left adjoint $\Phi\colon {{\mathrm{Mod}}}_{C^*}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}$, as well as the fact that it preserves cocartesian edges, follows once we verify the following: for any map of dg-Lie algebroids $f\colon {\mathfrak{h}}{\rightarrow}{\mathfrak{g}}$ and any collection of $C^*({\mathfrak{g}})$-modules $M_i$, the natural map $$\xymatrix{
\Phi_{\mathfrak{h}}\big(C^*({\mathfrak{h}})\otimes_{C^*({\mathfrak{g}})}M_1\otimes_{C^*({\mathfrak{g}})}\dots\otimes_{C^*({\mathfrak{g}})} M_m\big)\ar[r] & \Phi_{{\mathfrak{g}}}(M_1)\otimes_A \dots \otimes_A \Phi_{\mathfrak{g}}(M_m)
}$$ is an equivalence. Since both functors preserve colimits of modules in each variable, we can reduce to the case where each $M_i$ is equivalent to $C^*({\mathfrak{g}})$. In that case, the map can be identified with a map $$\smash{\xymatrix{
K({\mathfrak{h}})\ar[r] & K({\mathfrak{g}})\otimes_A \dots \otimes_A K({\mathfrak{g}})
}}$$ between Koszul complexes. Since both $K({\mathfrak{g}})$ and $K({\mathfrak{h}})$ were resolutions of the canonical representation $A$, the result follows.
In other words, the functors $\Phi_{\mathfrak{g}}$ from Section \[fm:sec:representationsofgood\] determine a natural (symmetric monoidal) transformation between diagrams of symmetric monoidal $\infty$-categories.
Quasicoherent sheaves {#fm:sec:qcoh}
---------------------
When restricted to ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A\subseteq {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$, the left adjoint $\Phi$ of Lemma \[fm:lem:functorialleftadjoint\] corresponds under straightening to a natural transformation $$\xymatrix@C=3pc{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar@/^1.5pc/[r]^{{{\mathrm{Mod}}}_{C^*}}_{}="s" \ar@/_1.5pc/[r]_{{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}}^{}="t" & \hspace{50pt}{
\StrLen{Pr_{sym. mon.}^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr_{sym. mon.}^{\mathrm{L}}}
\else \mathrm{Pr_{sym. mon.}^{\mathrm{L}}}
\fi}={{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({
\StrLen{Pr^L}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^L}
\else \mathrm{Pr^L}
\fi}) \ar@{=>}"s";"t"^{\Phi}
}$$ between diagrams of presentable (closed) symmetric monoidal $\infty$-categories, with symmetric monoidal left adjoint functors between them. This natural transformation is given pointwise by the left derived functor $\Phi_{\mathfrak{g}}$ of .
We can precompose with the duality functor ${\mathfrak{D}}$ and obtain a natural transformation of functors $$\label{fm:diag:psinattransformation}\vcenter{\xymatrix{
{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A\ar[d]\ar[rr] \ar[d] & & {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k\ar[d]^{{{\mathrm{Mod}}}}\\
{{
\StrLen{Fun}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Fun}
\else \mathrm{Fun}
\fi}}({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A, {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}})^{{\mathrm{op}}}\ar[r]_-{{\mathfrak{D}}_!} & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar[r]_-{{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}} & {
\StrLen{Pr_{sym. mon.}^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr_{sym. mon.}^{\mathrm{L}}}
\else \mathrm{Pr_{sym. mon.}^{\mathrm{L}}}
\fi} \ar@{=>}[-1,0]+<-50pt, -10pt>;[0, -2]+<50pt, 15pt>^\Psi
}}$$ Here ${\mathfrak{D}}_!$ is the unique limit-preserving functor which restricts to ${\mathfrak{D}}$ on the corepresentable functors. For each $B$ in ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A$, the functor $\Psi_B$ is the composite $$\label{fm:diag:compositeformodules}\vcenter{\xymatrix{
\Psi_{B}\colon {{\mathrm{Mod}}}_{B}\ar[r] & {{\mathrm{Mod}}}_{C^*{\mathfrak{D}}(B)} \ar[r]^{\Phi_{{\mathfrak{D}}(B)}} & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{D}}(B)}
}}$$ where the first functor arises from the unit map $B{\rightarrow}C^*{\mathfrak{D}}(B)$.
\[def:qc\] The presentable, symmetric monoidal $\infty$-category ${{\mathrm{QC}}}(X)$ of *quasi-coherent sheaves* on a functor $X\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A{\rightarrow}{{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}$ is the value on $X$ of the right Kan extension $$\xymatrix{
{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A\ar[r]^{{\mathrm{Mod}}}\ar[d] & {
\StrLen{Pr_{sym. mon.}^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr_{sym. mon.}^{\mathrm{L}}}
\else \mathrm{Pr_{sym. mon.}^{\mathrm{L}}}
\fi}\\
{{
\StrLen{Fun}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Fun}
\else \mathrm{Fun}
\fi}}({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A, {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}})^{{\mathrm{op}}}\ar@{..>}[ru]_{{{\mathrm{QC}}}}
}$$ This right Kan extension exists by [@lur09 Lemma 5.1.5.5], since $ {
\StrLen{Pr_{sym. mon.}^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr_{sym. mon.}^{\mathrm{L}}}
\else \mathrm{Pr_{sym. mon.}^{\mathrm{L}}}
\fi}$ has all small limits [@lur16 Proposition 4.8.1.15].
By the universal property of ${{\mathrm{QC}}}$, we obtain a natural transformation of presentable symmetric monoidal $\infty$-categories $\Psi_X\colon{{\mathrm{QC}}}(X){\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{D}}_!(X)}$. When $X$ is corepresentable, this is simply the functor $\Psi$ from . If $A$ is eventually coconnective and $X$ is a formal moduli problem over $A$, then ${\mathfrak{D}}_!(X)$ is naturally equivalent to $T_{A/X}$, by Remark \[rem:mcintermsofD\]. We thus obtain a natural symmetric monoidal functor $$\label{fm:diag:psi_X}\vcenter{\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}(X)\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}
}}$$ for any formal moduli problem $X$. By naturality, the composition $$\smash{\xymatrix{
{{\mathrm{QC}}}(X)\ar[r]^{\Psi_X} \ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}} \ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_0 = {{\mathrm{Mod}}}_A
}}$$ is naturally equivalent to the functor $x^*\colon {{\mathrm{QC}}}(X){\rightarrow}{{\mathrm{Mod}}}_A$ that restricts $F$ to the basepoint $x\in X(A)$.
We have to prove that for any formal moduli problem $X$, $\Psi_X$ is fully faithful and restricts to an equivalence between ${{\mathrm{QC}}}^{{{\scriptscriptstyle \geq}}0}(X)$ and ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}^{{{\scriptscriptstyle \geq}}0}$. When $X$ is representable by an object $B\in {
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}_k^{\mathrm{sm}}/A$, this functor is given by the composite . The first functor is an equivalence by Theorem \[thm:formalmoduliperfect\], so that the result follows from Corollary \[fm:cor:equivonconnmodules\].
The functor ${{\mathrm{QC}}}$ sends sifted colimits in ${
\StrLen{Fun}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Fun}
\else \mathrm{Fun}
\fi}({
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}^{\mathrm{sm}}/A, {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}})$ to limits of symmetric monoidal $\infty$-categories by construction. Lemma \[fm:lem:algrepspreservessiftedcolim\] implies that the same assertion holds for $X\mapsto {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{D}}_!(X)}$. Every formal moduli problem is a sifted colimit of representable functors, since every Lie algebroid is a sifted colimit of good Lie algebroids. It follows that $\Psi_X$ is fully faithful, being a limit of fully faithful functors.
It remains to identify the essential image of ${{\mathrm{QC}}}(X)^{{{\scriptscriptstyle \geq}}0}$. Recall that a quasi-coherent sheaf $F$ is connective if and only if $f^*F\in {{\mathrm{Mod}}}_{B}$ is connective for any $f\in X(B)$. In terms of its image $\Psi_X(E)$, this means for any map $f\colon {\mathfrak{D}}(B){\rightarrow}T_{A/X}$, the restricted representation $f^!\Psi_X(E)$ is the image of a connective $B$-module. By Corollary \[fm:cor:equivonconnmodules\], this is equivalent to $\Psi_X(E)$ being a connective $T_{A/X}$-representation.
Let $A$ be eventually coconnective and let $X$ be a formal moduli problem such that $T_{A/X}$ is connective. Then there is an equivalence $$\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}(X)\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}.
}$$
Since $T_{A/X}$ is connective, its enveloping algebra ${\mathcal{U}}(T_{A/X})$ is connective as well. It follows that $T_A$-representations carry a *right complete* $t$-structure, where ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{{{\scriptscriptstyle \geq}}0}_{T_{A/X}}$ consists of ${\mathfrak{g}}$-representations whose underlying chain complex is connective [@lur16 Example 2.2.1.3]. Similarly, ${{\mathrm{QC}}}(X)$ carries a right complete $t$-structure where $F\in {{\mathrm{QC}}}(X)^{{{\scriptscriptstyle \geq}}0}$ if and only if $f^*F$ is a connective chain complex for all $f\in X(B)$.
The functor $\Psi_X$ fits into a sequence of locally presentable $\infty$-categories and left adjoint functors between them $$\xymatrix{
{{\mathrm{QC}}}(X)^{{{\scriptscriptstyle \geq}}0}\ar[d]_{\Psi_X^{{{\scriptscriptstyle \geq}}0}}\ar[r] & {{\mathrm{QC}}}(X)^{{{\scriptscriptstyle \geq}}-1}\ar[r]\ar[d]_{\Psi_X^{{{\scriptscriptstyle \geq}}-1}} & \dots \ar[r] & {{\mathrm{QC}}}(X)\ar[d]^{\Psi_X}\\
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}^{{{\scriptscriptstyle \geq}}0}\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}^{{{\scriptscriptstyle \geq}}-1}\ar[r] & \dots \ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}.
}$$ Since ${{\mathrm{QC}}}(X)$ and ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{T_{A/X}}$ are right complete, the horizontal sequences are colimit diagrams, so that the result follows Theorem \[fm:thm:quasi-coherentvsrepresentations\].
Deformations of algebras
------------------------
Theorem \[fm:thm:quasi-coherentvsrepresentations\] can be used to study the deformation theory of connective (commutative) $A$-algebras. Suppose that $R$ is a cofibrant commutative dg-algebra in ${{\mathrm{Mod}}}^{{{\mathrm{dg}}}, {{\scriptscriptstyle \geq}}0}_A$, so that $R$ determines an object in the $\infty$-category ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}^{{{\scriptscriptstyle \geq}}0}_A)$. Consider the functor $$\xymatrix{
{\mathrm{Def}}_R\colon {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A^{{\mathrm{op}}}\ar[r] & \widehat{{
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}}_\infty; \hspace{4pt} X\ar@{|->}[r] & {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{QC}}}(X))\times_{{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_A)}\{R\}
}$$ sending each formal moduli problem $X$ to the (locally small) $\infty$-category of commutative algebras in ${{\mathrm{QC}}}(X)$, equipped with an equivalence between $R$ and their restriction to the canonical basepoint $x\in X(A)$. One can think of a point in ${\mathrm{Def}}_R(X)$ as a *deformation* of the commutative algebra along the map of stacks ${{\mathrm{Spec}}}(A){\rightarrow}X$.
Every such deformation is necessarily connective, since $R$ itself is connective. It follows from Theorem \[thm:formalmoduliperfect\] and Theorem \[fm:thm:quasi-coherentvsrepresentations\] that this functor can be identified with the functor $$\smash{\xymatrix{
{\mathrm{Act}}_R\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{{\mathrm{op}}}\ar[r] & \widehat{{
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}}_\infty; \hspace{4pt} {\mathfrak{g}}\ar@{|->}[r] & {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}})\times_{{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_A)} \{R\}
}}$$ sending each Lie algebroid ${\mathfrak{g}}$ to the category of commutative algebras in ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ whose underlying commutative $A$-algebra is equivalent to $R$. In light of Example \[ex:algebrainrep\], one an think of ${\mathrm{Act}}_R({\mathfrak{g}})$ as the category of action of ${\mathfrak{g}}$ on $R$ by derivations.
It follows from Lemma \[fm:lem:algrepspreservessiftedcolim\] that ${\mathrm{Act}}_R$, and therefore ${\mathrm{Def}}_R$, preserves all limits. In particular, the restriction of ${\mathrm{Def}}_R$ along the Yoneda embedding gives rise to a formal moduli problem $$\xymatrix{
{\mathrm{Def}}_R\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A\ar[r] & {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}; \hspace{4pt} B\ar@{|->}[r] & {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_B)\times_{{{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_A)} \{R\}.
}$$ This takes values in (small) spaces because ${\mathrm{Def}}_R(A\oplus A[n])\simeq \Omega {\mathrm{Def}}_R(A\oplus A[n+1])$ and all $\infty$-categories involved are locally small. In particular, ${\mathrm{Def}}_R$ is classified by a certain Lie algebroid ${\mathfrak{h}}$.
Recall from Example \[ex:atiyah2\] that associated to the commutative dg-algebra $R$ in ${{{\mathrm{Mod}}}^{{{\mathrm{dg}}}}}_A$ is a dg-Lie algebroid ${\mathrm{At}}_{{\mathbb{E}}_\infty}(R)$ of compatible derivations of $A$ and $R$. Since $R$ is a cofibrant commutative dg-algebra, ${\mathrm{At}}_{{\mathbb{E}}_\infty}(R)$ is a fibrant dg-Lie algebroid. There is an obvious representation of ${\mathrm{At}}_{{\mathbb{E}}_\infty}(R)$ on $R$ itself by means of derivations, which is classified by a map $$\smash{\xymatrix{
\phi\colon {\mathrm{At}}_{{\mathbb{E}}_\infty}(R)\ar[r] & {\mathfrak{h}}.
}}$$ We have the following (folklore) result, cf. [@toe16] (see also [@hin04]):
\[prop:defofalgebras\] The map $\phi$ is an equivalence for any cofibrant connective dg-$A$-algebra $R$. In other words, ${\mathrm{At}}_{{\mathbb{E}}_\infty}(R)$ classifies the formal moduli problem ${\mathrm{Def}}_R$.
Let ${\mathfrak{s}}_n$ be the free Lie algebroid on the map $0\colon A[n]{\rightarrow}T_A$, with natural maps $i\colon 0{\rightarrow}{\mathfrak{s}}_n$ and $r\colon {\mathfrak{s}}_n{\rightarrow}0$. It suffices to verify that the maps $\phi_*$ in the following commuting diagram are bijections for all $n$: $$\xymatrix@R=1.9pc{
\pi_1{{\mathrm{Map}}}\big({\mathfrak{s}}_{n}, {\mathrm{At}}_{{\mathbb{E}}_\infty}(R)\big) \ar[r]^{\phi_*}\ar[d]^{\cong} & \pi_1{{\mathrm{Map}}}\big({\mathfrak{s}}_{n}, {\mathfrak{h}}\big) \ar[r]^-{\cong}\ar[d]^{\cong} & \pi_1{\mathrm{Act}}_R({\mathfrak{s}}_{n})=: M_1\ar[d]_\psi^{\cong}\\
\pi_0{{\mathrm{Map}}}({\mathfrak{s}}_{n+1}, {\mathrm{At}}_{{\mathbb{E}}_\infty}(R))\ar[r]_{\phi_*} & \pi_0{{\mathrm{Map}}}\big({\mathfrak{s}}_{n+1}, {\mathfrak{h}}\big) \ar[r]_-{\cong} & \pi_0{\mathrm{Act}}_R({\mathfrak{s}}_{n+1})=: M_0.
}$$ The dg-Lie algebroid ${\mathrm{At}}(R)$ is fibrant and the kernel of its anchor map is the Lie algebra of $A$-linear derivations of $R$. The abelian group $\pi_1{{\mathrm{Map}}}\big({\mathfrak{s}}_{n}, {\mathrm{At}}_{{\mathbb{E}}_\infty}(R)\big)$ can therefore be identified with $\pi_{n+1}{\mathrm{Der}}_A(R, R)$.
On the other hand, $M_1$ can *also* be identified with $\pi_{n+1}{\mathrm{Der}}_A(R, R)$. To see this, let $r^!R\in {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{s}}_n})$ be the trivial representation of ${\mathfrak{s}}_{n}$ on $R$. We can identify $M_1$ with the set of homotopy classes of maps $\alpha\colon r^!R{\rightarrow}r^!R$ such that the restriction $i^!(\alpha)$ is the identity map in ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_A)$. Using that the restriction functors $r^!$ and $i^!$ have a right adjoints $r_!$ and $i_!$, such maps can be identified with sections of $$\xymatrix{
r_!r^!(R)\ar[r]^-{r_!\eta r^!} & r_!i_!i^!r^!(R) = R
}$$ in the $\infty$-category ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_A)$. The right adjoint $r_!$ can be computed as $$r_!(r^!R) \simeq {{\mathrm{Hom}}}^{\mathbb{R}}_{{\mathcal{U}}({\mathfrak{s}}_n)}(A, r^!R) \simeq C^*\big({\mathfrak{s}}_n, r^!R\big).$$ In fact, since ${\mathfrak{s}}_n$ is a free Lie algebroid, the discussion of Section \[fm:sec:ceoffree\] shows that restriction along $A[n]{\rightarrow}{\mathfrak{s}}_n$ induces a weak equivalence of commutative dg-$A$-algebras $$\xymatrix{
C^*({\mathfrak{s}}_n, r^!R)\ar[r]^-\sim & R\oplus {{\mathrm{Hom}}}_A\big(A[n+1], r^!R\big) \cong R\oplus R[-n-1]
}$$ to the *square zero extension* of $R$ by a shifted copy of itself. We conclude that $M_1$ is isomorphic to the set of homotopy classes of sections in ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{\mathrm{Mod}}}_A)$ of $$\smash{\xymatrix{
R\oplus R[-n-1]\ar[r] & R.
}}$$ The set of such sections is indeed isomorphic to $\pi_{n+1}{\mathrm{Der}}_A(R, R)$, as asserted.
Consider an element $v\in{\mathrm{Der}}_A(R, R)\cong M_1$, corresponding to a map $v\colon r^!R{\rightarrow}r^!R$ in ${{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}({{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{s}}_n})$. Unwinding the definitions, the image $\psi(v)\in M_0$ is given by (the equivalence class of) the ${\mathfrak{s}}_{n+1}$-action on $R$ where the generator of ${\mathfrak{s}}_{n+1}$ acts by $v$. This representation is exactly the image of the map $v\colon A[n+1]{\rightarrow}{\mathfrak{s}}_{n+1}{\rightarrow}{\mathrm{At}}(R)$ under the map $\phi_*$. We conclude that $\phi_*$ is indeed an isomorphism.
The same proof shows that for any (simplicial, coloured) operad ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$ and a cofibrant ${
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}$-algebra $R$, the formal moduli problem ${\mathrm{Def}}_R$ of deformations of $R$ is classified by the Atiyah Lie algebroid ${\mathrm{At}}_{{
\StrLen{P}[\mystrlengt]
\ifnum\mystrlengt=1 \mathscr{P}
\else \mathrm{P}
\fi
}}(R)$ of Example \[ex:atiyah\]. For example, this provides an explicit dg-Lie algebroid classifying the deformations of modules, associative algebras or diagrams thereof.
Representations and pro-coherent sheaves {#fm:sec:indcoh}
========================================
Let $X$ be a formal moduli problem over $A$ with associated Lie algebroid $T_{A/X}$. We have seen in the previous section that the $\infty$-category of $T_{A/X}$-representation is an *extension* of the $\infty$-category of quasi-coherent sheaves on $X$. The purpose of this section is to provide a geometric description of this extension when $A$ is bounded coherent, by identifying it with the $\infty$-category of *pro-coherent sheaves* on $X$ (Definition \[def:procohonformmod\]).
\[fm:thm:ind-coherentvsrepresentations\] Let $A$ be a cofibrant, bounded and coherent commutative dg-$k$-algebra and let $X$ be a pro-coherent formal moduli problem over $A$. Then there is an equivalence $$\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}^!(X)\ar[r]^-\sim & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{T_{A/X}}
}$$ between the $\infty$-categories of *pro-coherent sheaves* on $X$ and pro-coherent $T_{A/X}$-representations. For any map $f\colon X{\rightarrow}Y$ of pro-coherent formal moduli problems, this equivalence identifies the functor $f^!\colon {{\mathrm{QC}}}^!(Y){\rightarrow}{{\mathrm{QC}}}^!(X)$ with the restriction functor from $T_{A/Y}$-representations to $T_{A/X}$-representations.
Suppose that $A$ is coherent and eventually coconnective and let $X\in {
\StrLen{FormMod}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{FormMod}
\else \mathrm{FormMod}
\fi}_A$ be an ordinary formal moduli problem over $A$, with Lie algebroid $T_{A/X}$. The $\infty$-category of (quasi-coherent) $T_{A/X}$-representations can then be identified with the full subcategory of pro-coherent sheaves on $X$ whose restriction to $A$ is quasi-coherent.
We will begin by discussing pro-coherent representations over good pro-coherent Lie algebroids in Section \[sec:koszulformodules\]. In Section \[sec:indcohonformmod\] we use this to prove Theorem \[fm:thm:ind-coherentvsrepresentations\].
Koszul duality for modules {#sec:koszulformodules}
--------------------------
Recall that the left adjoint $\Phi$ from Lemma \[fm:lem:functorialleftadjoint\] corresponds under straightening to a natural transformation between diagrams $$\smash{\xymatrix{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!, {{\mathrm{Mod}}}_{C^*}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!, {{\mathrm{op}}}}\ar[r] & {
\StrLen{Pr}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr}
\else \mathrm{Pr}
\fi}^{\mathrm{L}}
}}$$ of locally presentable $\infty$-categories and left adjoint functors between them. When evaluated on an arrow $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$, this natural transformation is given by the commuting square of left adjoints $$\hspace{-60pt}\xymatrix{
{{\mathrm{Mod}}}_{C^*({\mathfrak{h}})}\ar[d]_{f^*:=C^*({\mathfrak{g}})\otimes_{C^*({\mathfrak{h}})}-} \ar[r]^-\Phi & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{h}}\ar[d]^{f^!}\\
{{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}\ar[r]_-\Phi & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}.
}$$ Let us pass to the associated diagram of right adjoint functors and next take opposite categories. The natural transformation $\Phi$ then determines a natural transformation between two diagrams of large $\infty$-categories and left adjoint functors between them $$\smash{\xymatrix{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{!, {{\mathrm{op}}}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!, {{\mathrm{op}}}}\ar[r] & \widehat{{
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}}{}^{\mathrm{L}}_\infty & {{\mathrm{Mod}}}_{C^*}^{{\mathrm{op}}}\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^{!, {{\mathrm{op}}}}\ar[r] & \widehat{{
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}}{}^{\mathrm{L}}_\infty.
}}$$ The value of this natural transformation on an arrow $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$ is given by $$\label{diag:coinduction}\vcenter{\xymatrix@C=3pc{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{!,{{\mathrm{op}}}}_{\mathfrak{g}}\ar[d]_{f_!}\ar[r]^-{C^*({\mathfrak{g}}, -)} & {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^{{{\mathrm{op}}}}\ar[d]^{f_*}\\
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{!,{{\mathrm{op}}}}_{\mathfrak{h}}\ar[r]_-{C^*({\mathfrak{h}}, -)} & {{\mathrm{Mod}}}_{C^*({\mathfrak{h}})}^{{{\mathrm{op}}}}
}}$$ where $f_!$ is the right adjoint to $f^!$, given by coinduction. For every $A$-cofibrant dg-Lie algebroid ${\mathfrak{g}}$, taking the $A$-linear dual of a ${\mathfrak{g}}$-representation determines a left Quillen functor between the tame model structures on dg-${\mathfrak{g}}$-representations $$\smash{\xymatrix{
(-)^\vee\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{{{\mathrm{dg}}}}\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{{{\mathrm{dg}}}, {{\mathrm{op}}}}
}}$$ whose right adjoint is given by $(-)^\vee$ as well. For every map $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$, this functor intertwines induction and coinduction, i.e. there is a natural commuting diagram $$\label{diag:inductionandcoinduction}\vcenter{\xymatrix@C=2.5pc{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}\ar[r]^-{(-)^\vee}\ar[d]_{f_*} & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{!, {{\mathrm{op}}}}\ar[d]^{f_!}\\
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{h}}\ar[r]_-{(-)^\vee} & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{h}}^{!, {{\mathrm{op}}}}
}}$$ at the level of $\infty$-categories. We therefore obtain a natural transformation $$\label{diag:mu}\xymatrix{
\mu\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_* \ar[r]^-{(-)^\vee} & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^{!, {{\mathrm{op}}}}\ar[r]^{C^*} & {{\mathrm{Mod}}}^{{{\mathrm{op}}}}_{C^*}
}$$ between functors ${{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A{\rightarrow}\widehat{{
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}}{}^{\mathrm{L}}_\infty$. Here ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_*$ denotes the functor sending a map of Lie algebroids $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$ to the induction functor $f_*\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}{\rightarrow}{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{h}}$.
\[prop:dualityforreps\] Let ${\mathfrak{g}}\in {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A^!$ be a good pro-coherent Lie algebroid over $A$. Then the left adjoint functor $\mu$ restricts to an equivalence $$\smash{\xymatrix{
\mu\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^{!,\omega}\ar[r] & {
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_{C^*({\mathfrak{g}})}^{{\mathrm{op}}}}}$$ between the $\infty$-category of compact pro-coherent ${\mathfrak{g}}$-representations and the opposite of the $\infty$-category of coherent $C^*({\mathfrak{g}})$-modules.
\[lem:coherentoversmallext\] Let ${\mathfrak{g}}$ be a good pro-coherent Lie algebroid over $A$. Consider the thick subcategory ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\subseteq {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}$ generated by the modules of the form $f_*F$, where $F\in {
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_A$ is a coherent $A$-module and $f\colon C^*({\mathfrak{g}}){\rightarrow}A$ is the canonical map. Then ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ coincides with ${
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_{C^*({\mathfrak{g}})}$.
To see that ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}\subseteq {
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_{C^*({\mathfrak{g}})}$, we have to show that for every coherent $A$-module $F$, $f_*F$ is a coherent $C^*({\mathfrak{g}})$-module. Since $C^*({\mathfrak{g}}){\rightarrow}A$ is a small extension, the map $\pi_0(C^*({\mathfrak{g}})){\rightarrow}\pi_0(A)$ is a surjection of coherent commutative rings. It follows that any finitely generated $\pi_0(A)$-module, such as $\pi_n(F)$, is also finitely generated over $\pi_0(C^*({\mathfrak{g}}))$. This means that $f_*F$ is indeed a coherent $C^*({\mathfrak{g}})$-module.
It remains to be verified that ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ exhausts ${
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_{C^*({\mathfrak{g}})}$. Using Postnikov towers, one sees that ${
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_{C^*({\mathfrak{g}})}$ is the thick subcategory generated by the (discrete) finitely presented $\pi_0(C^*({\mathfrak{g}}))$-modules. It therefore suffices to show that ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$ contains any such a finitely generated $\pi_0(C^*({\mathfrak{g}}))$-module $E$. Since $C^*({\mathfrak{g}}){\rightarrow}A$ is a small extension, there is a finite sequence of square zero extensions by finitely generated $\pi_0(A)$-modules $$\xymatrix{
\pi_0(C^*({\mathfrak{g}}))= R_n\ar[r] & R_{n-1}\ar[r] & \dots \ar[r] & R_0=\pi_0(A).
}$$ We prove by induction that all (discrete) finite generated $R_i$-modules are contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$. This clearly holds for finitely generated $\pi_0(A)$-modules, which are coherent $A$-modules. Let $R_i{\rightarrow}R_{i-1}$ be a square zero extension of $R_{i-1}$ by a finitely generated $\pi_0(A)$-module $I$ and let $E$ be a finitely generated $R_i$-module. Then there is an exact sequence $$\xymatrix{
I\otimes_{R_{0}} (R_0\otimes_{R_i} E) \ar[r] & E\ar[r] & R_{i-1}\otimes_{R_n} E\ar[r] & 0.
}$$ Since $I\otimes_{R_{0}} (R_0\otimes_{R_i} E)$ and $R_{i-1}\otimes_{R_n} E$ are contained in ${
\StrLen{C}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{C}
\else \mathrm{C}
\fi}$, so is $E$.
The right adjoint of the functor $\mu$ is given by $$\xymatrix{
\nu\colon {{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}^{{\mathrm{op}}}\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}; \hspace{4pt} M\ar[r] & \Phi(M)^\vee.
}$$ We will first show that for any compact ${\mathfrak{g}}$-representation $E$, the unit map $E{\rightarrow}\nu\mu(E)$ is an equivalence. To this end, let $f\colon 0{\rightarrow}{\mathfrak{g}}$ be the initial map and note that the forgetful functor $f^!\colon {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}{\rightarrow}{{\mathrm{Mod}}}_A$ detects equivalences and preserves filtered colimits. It follows that ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^\omega\subseteq {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}$ is the thick subcategory generated by the ${\mathfrak{g}}$-representations $$f_*(F) = {\mathcal{U}}({\mathfrak{g}})\otimes_A F$$ where $F\in {{\mathrm{Mod}}}_A^{!, \omega}$ is a compact object in the $\infty$-category of pro-coherent sheaves on $A$. It therefore suffices to check that the unit map $f_*F{\rightarrow}\nu\mu(f_*F)$ is an equivalence for any such free ${\mathfrak{g}}$-representation $f_*F$. To see this, note that the unit map is the composite $$\xymatrix{
f_*F \ar[r] & \big((f_*F)^{\vee}\big){}^\vee\ar[r]^-{\epsilon} & \Phi\big(C^*({\mathfrak{g}}, (f_*F)^\vee)\big)^\vee
}$$ where $\epsilon$ is the unit of the adjoint pair $(\Phi, C^*)$. The PBW theorem and Lemma \[fm:lem:verygoodlralgebroids\] imply that ${\mathcal{U}}({\mathfrak{g}})\cong \bigoplus_{n{{\scriptscriptstyle \geq}}0} A^{\oplus k_n}[-n]$ as a graded $A$-module. By Diagram \[diag:inductionandcoinduction\], we find that $$(f_*F)^\vee = f_!(F^\vee) = {{\mathrm{Hom}}}_A\big({\mathcal{U}}({\mathfrak{g}}), F^\vee\big) = \prod_{n{{\scriptscriptstyle \geq}}0} (F^\vee)^{\oplus k_n}[-n].$$ Since $F^\vee\in {{\mathrm{Mod}}}^!_A$ is a coherent $A$-module, hence in particular eventually connective, $(f_*F)^\vee$ is eventually connective as well. The map $\epsilon$ is then an equivalence by Corollary \[fm:cor:equivonconnprocoherent\]. Furthermore, the map $f_*F{\rightarrow}(f_*F)^{\vee\vee}$ can be identified with $$\xymatrix{
f_*F = \bigoplus_{n{{\scriptscriptstyle \geq}}0} F^{\oplus k_n}[-n]\ar[r] & (f_!F^\vee){}^\vee = \prod_{n{{\scriptscriptstyle \geq}}0} (F^{\vee\vee})^{\oplus k_n}[-n].
}$$ The description of the compact objects from Proposition \[prop:tamecompactlygenerated\] shows that $F{\rightarrow}F^{\vee\vee}$ is an isomorphism. Since $A$ is a bounded cdga, $F$ is trivial in sufficiently high (homological) degrees and the above map is an isomorphism.
We conclude that $\mu$ is fully faithful on the compact objects of ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{\mathfrak{g}}^!$. The essential image of the compact objects is the smallest thick subcategory of ${{\mathrm{Mod}}}_{C^*({\mathfrak{g}})}$ containing the images $\mu(f_*F)$. But $$\mu(f_*F) \simeq C^*({\mathfrak{g}}, (f_*F)^\vee) \simeq C^*({\mathfrak{g}}, f_!(F^\vee)) \simeq f_*(F^\vee)$$ by the commuting diagrams and . The compact objects of ${{\mathrm{Mod}}}_A^!$ are precisely the preduals of the coherent $A$-modules, so that Lemma \[lem:coherentoversmallext\] shows that the essential image of ${{\mathrm{Mod}}}_A^{!, \omega}$ under $\mu$ coincides with the coherent $C^*({\mathfrak{g}})$-modules.
Pro-coherent sheaves {#sec:indcohonformmod}
--------------------
We will now deduce Theorem \[fm:thm:ind-coherentvsrepresentations\] from the functoriality of the equivalence provided by Proposition \[prop:dualityforreps\].
\[cons:indcohonformalmoduli\] It follows from Lemma \[lem:coherentoversmallext\] that for any map $f\colon B{\rightarrow}B'$ between small extensions of $A$ (by coherent $A$-modules), the functor $$\xymatrix{
f_*\colon {{\mathrm{Mod}}}_{B'}\ar[r] & {{\mathrm{Mod}}}_B
}$$ preserves coherent modules. Consequently, there is a functor $$\xymatrix{
{
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}\colon \big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}^{{\mathrm{sm, coh}}}/A\big)^{{{\mathrm{op}}}}\ar[r] & {
\StrLen{StCat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{StCat}
\else \mathrm{StCat}
\fi}^{{\mathrm{ex}}}_\infty
}$$ sending each small extension $B{\rightarrow}A$ to the stable $\infty$-category of coherent $B$-modules, and each map $f\colon B{\rightarrow}B'$ to the exact functor $f_*$. Consider the composite functor $$\xymatrix{
\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A\big)^{{\mathrm{op}}}\ar[r]^-{{
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}}} & {
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}_\infty\ar[r]^{{\mathrm{Ind}}} & {
\StrLen{Pr}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr}
\else \mathrm{Pr}
\fi}^{\mathrm{L}}
}$$ sending each small extension $B{\rightarrow}A$ to the locally presentable stable $\infty$-category ${\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}}_B)$ of pro-coherent sheaves on $B$. Every map $f\colon B{\rightarrow}B'$ of small extensions over $A$ induces an adjoint pair $$\xymatrix{
f_* = {\mathrm{Ind}}(f_*)\colon {\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}}_{B'})\ar@<1ex>[r] & {\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}_{B}^{{\mathrm{op}}}) \ar@<1ex>[l] \colon f^!.
}$$ By construction, the functor $f_*$ preserves compact objects. It follows that the right adjoint $f^!$ preserves all colimits and admits a further right adjoint $f_!$ (this is discussed in much greater generality in [@gai13]).
\[def:procohonformmod\] For any functor $X\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A{\rightarrow}{{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}$, define the stable presentable $\infty$-category ${{\mathrm{QC}}}^!(X)$ of *pro-coherent sheaves* on $X$ to be the value on $X$ of the left Kan extension $$\xymatrix@C=3pc{
\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A\big)^{{\mathrm{op}}}\ar[r]^-{{\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}})}\ar[d] & {
\StrLen{Pr^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}}
\else \mathrm{Pr^{\mathrm{L}}}
\fi}\\
{{
\StrLen{Fun}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Fun}
\else \mathrm{Fun}
\fi}}({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A, {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}})\ar@{..>}[ru]_-{{{\mathrm{QC}}}^!}
}$$ This exists by [@lur09 Lemma 5.1.5.5], since $ {
\StrLen{Pr^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}}
\else \mathrm{Pr^{\mathrm{L}}}
\fi}$ has all small colimits [@lur09]. Informally, a pro-coherent sheaf on $X$ is given by a collection of pro-coherent sheaves $F_y$ over $B\in {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A$ for every $y\in X(B)$, together with a coherent family of equivalences for every $f\colon B{\rightarrow}B'$ $$F_{f(y)}\simeq f^!F_y.$$
By Proposition \[prop:dualityforreps\], the natural transformation $\mu$ induces a natural equivalence $$\xymatrix@C=3pc{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{!, {\mathrm{good}}}_A\ar@/^1.5pc/[r]^{{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_\ast^{!,\omega}}_{}="s" \ar@/_1.5pc/[r]_{{
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{{\mathrm{op}}}} \circ C^*}^{}="t" & {
\StrLen{Cat}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Cat}
\else \mathrm{Cat}
\fi}_\infty \ar@{=>}"s";"t"^{\mu}
}$$ to the $\infty$-category of small $\infty$-categories. Here ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_\ast^{!, \omega}$ sends a good pro-coherent Lie algebroid ${\mathfrak{g}}$ to the $\infty$-category of compact ${\mathfrak{g}}$-representations and a map $f\colon {\mathfrak{g}}{\rightarrow}{\mathfrak{h}}$ to the induction functor $f_*$. Since each ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{\mathfrak{g}}$ is a compactly generated stable $\infty$-category, we obtain an induced natural equivalence $$\xymatrix@C=3pc{
{{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^{!, {\mathrm{good}}}_A\ar@/^1.5pc/[r]^{{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_\ast}_{}="s" \ar@/_1.5pc/[r]_{{\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{{\mathrm{op}}}}) \circ C^*}^{}="t" & {
\StrLen{Pr}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr}
\else \mathrm{Pr}
\fi}^{\mathrm{L}} \ar@{=>}"s";"t"^{\mu}
}$$ between Ind-completions. Precomposing with the duality functor ${\mathfrak{D}}$ , we obtain a natural diagram $$\vcenter{\xymatrix{
\big({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A\big)^{{{\mathrm{op}}}}\ar[d]\ar@/^1.5pc/[rrd]^-{{\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}})}_{}="s" \ar[d] & & \\
{{
\StrLen{Fun}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Fun}
\else \mathrm{Fun}
\fi}}({{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm, coh}}/A, {{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}})\ar[r]_-{{\mathfrak{D}}_!} & {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}^!_A\ar[r]_-{{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_\ast} & {
\StrLen{Pr^{\mathrm{L}}}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr^{\mathrm{L}}}
\else \mathrm{Pr^{\mathrm{L}}}
\fi} \ar@{=>}"s"+<-5pt, -5pt>;[0, -2]+<50pt, 15pt>^\Psi
}}$$ commuting up a natural equivalence $\Psi$, given by *inverse* of the composite $$\vcenter{\xymatrix{
{{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}^!_{{\mathfrak{D}}(B)}\ar[r]^-{\mu} & {\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}}_{B}) \ar[r] & {\mathrm{Ind}}({
\StrLen{Coh}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Coh}
\else \mathrm{Coh}
\fi}^{{\mathrm{op}}}_{C^*{\mathfrak{D}}(B)})
}}$$ where the second functor arises from the unit map $B{\rightarrow}C^*{\mathfrak{D}}(B)$. By the universal property of ${{\mathrm{QC}}}^!(X)$, we obtain a natural transformation $$\xymatrix{
\Psi_X\colon {{\mathrm{QC}}}^!(X)\ar[r] & {{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_{{\mathfrak{D}}_!(X)}
}$$ for any $X\colon {{
\StrLen{CAlg}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{CAlg}
\else \mathrm{CAlg}
\fi}}_k^{\mathrm{sm}}/A{\rightarrow}{{
\StrLen{S}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{S}
\else \mathrm{S}
\fi}}$. The proof of Theorem \[fm:thm:quasi-coherentvsrepresentations\] can now be repeated verbatim: when $X$ is a formal moduli problem, ${\mathfrak{D}}_!(X)$ is equivalent to $T_{A/X}$. Since $T_{A/X}$ is a sifted colimit of good pro-coherent Lie algebroids, $X$ is a sifted colimit of corepresentable functors. The functor ${{
\StrLen{Rep}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Rep}
\else \mathrm{Rep}
\fi}}_*\colon {{
\StrLen{LieAlgd}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{LieAlgd}
\else \mathrm{LieAlgd}
\fi}}_A{\rightarrow}{
\StrLen{Pr}[\mystrlen]
\ifnum\mystrlen=1 \mathscr{Pr}
\else \mathrm{Pr}
\fi}^{\mathrm{L}}$ preserves colimits by Lemma \[lem:propertiesoftamereps\], so that $\Psi_X$ is a sifted colimit of equivalences and hence an equivalence itself.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the spectral properties of classical and quantum Markovian processes that are reset at random times to a specific configuration or state with a reset rate that is independent of the current state of the system. We demonstrate that this simple reset dynamics causes a uniform shift in the eigenvalues of the Markov generator, excluding the zero mode corresponding to the stationary state, which has the effect of accelerating or even inducing relaxation to a stationary state. Based on this result, we provide expressions for the stationary state and probability current of the reset process in terms of weighted sums over dynamical modes of the reset-free process. We also discuss the effect of resets on processes that display metastability. We illustrate our results with two classical stochastic processes, the totally asymmetric random walk and the one-dimensional Brownian motion, as well as two quantum models: a particle coherently hopping on a chain and the dissipative transverse field Ising model, known to exhibit metastability.'
author:
- 'Dominic C. Rose'
- Hugo Touchette
- Igor Lesanovsky
- 'Juan P. Garrahan'
title: Spectral properties of simple classical and quantum reset processes
---
Introduction
============
The dynamics of stochastic processes, such as animals foraging for food in the wilderness or a person searching for car keys, often include random resets in time, taking the form of returns to past locations where food was successfully located or the last place a person remembers seeing their keys [@benichou2011]. Recently, there has been a renewed interest in these processes, due to the fact that they can improve the efficiency of certain random search processes and algorithms in terms of mean hitting or first-passage time [@evans2011; @evans2011b; @evans2013; @evans2014; @reuveni2016; @pal2017]. Reset processes have also been studied from a more physical point of view, as they provide a simple model of nonequilibrium processes breaking detailed balance [@manrubia1999; @gupta2014; @roldan2016; @eule2016], as well as of processes showing dynamical phase transitions in their relaxation dynamics [@majumdar2015b], mean first-passage time [@kusmierz2014], or large deviations [@meylahn2015b; @Harris2017; @hollander2018].
These studies follow many previous works in mathematics, in queuing theory and in population dynamics, in particular, on stochastic processes involving some form of random resets, variously referred to as failures, catastrophes, disasters or decimations; see, e.g., [@brockwell1982; @luby1993; @pakes1997; @kumar2000; @cairns2004; @janson2012; @avrachenkov2013; @dharmaraja2015; @montero2017]. Most of these works, as well as those from physics mentioned above, make use of the correspondence that exists between resets and renewals to obtain renewal representations of both time-dependent and stationary distributions, in addition to first-passage statistics. Modified Fokker–Planck and Feynman–Kac equations with additional source and sink terms describing the evolution of these distributions and statistics have also been obtained (see, e.g., [@evans2011b; @dharmaraja2015; @meylahn2015b]) and can be solved explicitly for some simple models, including reset versions of Brownian motion [@evans2011] and the Ornstein-Uhlenbeck process [@pal2015].
In this paper, we present a different approach to reset processes based on the spectral properties of their generator or master operator. Our main result is a relation between the spectrum of the generator of a reset process and that of its reset-free counterpart. More precisely, we show that the real part of the eigenvalues of the generator are shifted for all non-stationary modes by the reset rate, while the corresponding eigenstates, representing the dynamical modes, are not modified. We also provide explicit expressions for the stationary state and current of reset processes involving weighted sums over the dynamical modes, which are applied to two prototypical models, namely, the totally asymmetric random walk in one dimension, related to queuing, and the one-dimensional Brownian motion. The results obtained clearly explain how resets can accelerate or even induce relaxation to a stationary state by opening a spectral gap, and how non-zero stationary currents can be created without having complex eigenvalues in the spectrum of the generator. The eigenvalue result can also be used within the spectral theory of metastability [@GaveauSchulman1987Meta; @Gaveau1998; @Gaveau2006] to demonstrate that weak resetting can modify the weighting of metastable states without modifying those states as such.
While resets have been extensively considered in classical nonequilibrium physics, a relatively unexplored area is the addition of resets to quantum systems, either closed or interacting with an environment. This case has been considered recently for a quantum walker subjected to continuous measurements on a particular site, resulting in a random collapse of the wavefunction followed by an evolution starting from the measured site [@Thiel2017; @Friedman2017]. We conclude our study by generalising our spectral results to this type of open quantum systems described in general by the Lindblad master equation, providing a natural link with the recent extension of the spectral theory of metastability to quantum systems [@Macieszczak2016]. We illustrate this generalization by computing numerically the stationary state of a model of coherent hopping in one dimension realizing the reset quantum random walker, and by applying our method to a dissipative transverse field Ising model [@Ates2012], known to display metastability [@Rose2016].
Reset Markov processes {#ResetMarkov}
======================
We consider a classical stochastic process evolving according to a continuous-time Markov chain. The master equation describing the evolution of the probability $P(C,t)$ for the process to be in state $C$ at time $t$ is given by $$\partial_t P(C,t)=\sum_{C'\neq C}W(C'\rightarrow C)P(C',t)-R(C)P(C,t),$$ where $W(C'\rightarrow C)$ is the transition rate from $C'$ to $C$ and $$R(C)=\sum_{C'\neq C}W(C\rightarrow C')$$ is the escape rate from $C$. Following the notation commonly used in physics [@schutz2001], this can be written more compactly as $$\label{MasterEq}
\partial_t \ket{P(t)} = {\mathcal{L}}\ket{P(t)},$$ where $$\ket{P(t)}=\sum_C P(C,t)\ket{C}$$ is the probability vector expressed in terms of ket states $\ket{C}$, such that $\left\langle C|C' \right\rangle = \delta_{CC'}$, and $${\mathcal{L}}=\sum_{C,C'\neq C}W(C\rightarrow C')\ket{C'}\bra{C}-\sum_C R(C)\ket{C}\bra{C}$$ is the master operator.
Since this operator is non-Hermitian, it has two sets of eigenvectors, right and left, given by $${\mathcal{L}}\ket{r_i}=\lambda_i\ket{r_i}$$ and $$\bra{l_i}{\mathcal{L}}=\lambda_i\bra{l_i},$$ respectively. These two sets of eigenvectors form a complete basis, are dual to each other, and can be normalised in a such a way that $\left\langle l_i | r_j \right\rangle=\delta_{ij}$.
We assume here that the process is ergodic and, therefore, that it has a unique stationary state $\ket{P_\text{ss}}$, corresponding from to the eigenvalue $\lambda_1=0$, so that $\ket{P_\text{ss}}=\ket{r_1}$. We also assume that ${\mathcal{L}}$ contains no Jordan blocks, corresponding to non-exponentially decaying modes, so we do not need to consider generalized eigenvectors [^1]. The normalization of the stationary state can be expressed as $$\sum_C \left\langle C|P_\text{ss}\right\rangle=\left\langle-|P_\text{ss}\right\rangle=1,$$ where we have introduced the “flat” state $$\bra{-}=\sum_C \bra{C}.$$ Conservation of probability also requires $\bra{-}{\mathcal{L}}=0$, which implies $\bra{l_1}=\bra{-}$ and hence $\left\langle -|r_i \right\rangle=0$ for all $ i \neq 1$. From the ergodicity assumption, all other eigenvalues are possibly complex but have real parts less than zero, that is, $\text{Re}(\lambda_i)<0$ for all $i\neq1$. Both this and the Jordan block assumption can be relaxed to arrive at similar but slightly more general results.
The generator ${\mathcal{L}}$ defines our original process. The reset version of that process is constructed simply by adding new transitions at a rate $\Gamma$ from every configuration to a target or reset state, denoted by $C_0$. The generator of the reset process is thus given by $$\begin{aligned}
{\mathcal{L}}^\Gamma={\mathcal{L}}+\Gamma\sum_{C\neq C_0} \ket{C_0}\bra{C}-\Gamma\sum_{C\neq C_0}\ket{C}\bra{C},\end{aligned}$$ where the additional terms can be absorbed into the old jump operators and escape rate operator to give shifted transition and escape rates. Note that we can add $$0=\Gamma\ket{C_0}\bra{C_0}-\Gamma\ket{C_0}\bra{C_0}$$ to ${\mathcal{L}}^\Gamma$ to obtain the simpler form $$\begin{aligned}
{\mathcal{L}}^\Gamma&={\mathcal{L}}+\Gamma\sum_{C} \ket{C_0}\bra{C}-\Gamma\sum_{C}\ket{C}\bra{C}\nonumber\\
&={\mathcal{L}}+\Gamma\ket{C_0}\bra{-}-\Gamma I,
\label{LR}\end{aligned}$$ where $I$ is the identity operator. In this form, it is clear that the reset adds transitions from all states to $C_0$, contributing to an extra escape rate $\Gamma$ in the diagonal, which keeps the conservation condition $\bra{-}{\mathcal{L}}^\Gamma=0$.
The dynamics generated by is arguably the simplest form of reset process. As defined, reset events are Poissonian with rate $\Gamma$, which is independent of the state of the system at the time of the reset.
Results {#SecRes}
=======
We study in this section the spectral properties of the reset process, deriving the new eigenvalues, left eigenvectors and right eigenvectors of ${\mathcal{L}}^\Gamma$ in terms of those of ${\mathcal{L}}$. The results are then used to obtain spectral representations of the stationary state and current of the reset process, and to discuss the effect of resets on metastable states.
Eigenvalues and eigenvectors {#ResetSpec}
----------------------------
We begin our analysis with the right eigenvectors by noting that, due to conservation of probability and the fact that $\left\langle -|r_i \right\rangle=0$ for all $i\neq1$, we have $$\begin{aligned}
{\mathcal{L}}^\Gamma\ket{r_i} &=\big({\mathcal{L}}+\Gamma\ket{C_0}\bra{-}-\Gamma I\big)\ket{r_i}\nonumber\\
&=(\lambda_i-\Gamma)\ket{r_i}\end{aligned}$$ for $i\neq 1$. Consequently, the right eigenvectors of the reset process are the same as those of the original process, while the eigenvalues are shifted down by $\Gamma$: $$\lambda^\Gamma_i = \lambda_i-\Gamma,\qquad i\neq 1.$$ This applies, as noted, to all modes except the stationary state, discussed below, which is still such that $\lambda^\Gamma_1=0$.
To determine the left eigenvectors, we act from the left with the original eigenmodes $$\begin{aligned}
\bra{l_i}{\mathcal{L}}^{\Gamma}&=\bra{l_i}\big({\mathcal{L}}+\Gamma\ket{C_0}\bra{-}-\Gamma I\big)\nonumber\\
&=(\lambda_i-\Gamma)\bra{l_i}+\Gamma\left\langle l_i|C_0\right\rangle\bra{-}\nonumber\\
&=(\lambda_i-\Gamma)\left(\bra{l_i}+\frac{\Gamma\left\langle l_i|C_0\right\rangle}{\lambda_i-\Gamma}\bra{-}\right).\nonumber\end{aligned}$$ But by conservation of probability $\bra{-}\mathcal{L}^\Gamma=0$, we also have $$\bra{l_i}{\mathcal{L}}^{\Gamma}=\left(\bra{l_i}+\frac{\Gamma\left\langle l_i|C_0\right\rangle}{\lambda_i-\Gamma}\bra{-}\right){\mathcal{L}}^{\Gamma}$$ for all $i\neq 1$. Thus we see that the new left eigenvectors are given by $$\begin{aligned}
\label{NewLeft}
\bra{l^\Gamma_i}=\bra{l_i}+\frac{\Gamma\left\langle l_i|C_0\right\rangle}{\lambda_i-\Gamma}\bra{-},\qquad i\neq1.\end{aligned}$$ For $i=1$, we have as before $\bra{l_1^\Gamma}=\bra{-}$.
Stationary state {#StatStateResets}
----------------
The stationary state $\ket{P^\Gamma_\text{ss}}$ of the reset process, corresponding to $\ket{r_1^\Gamma}$, is obtained from the results above by noting that the new left and right eigenvectors are orthonormal to each other, so that $\left\langle{l^\Gamma_i}|P^\Gamma_\text{ss}\right\rangle=\delta_{i1}$. Substituting this condition in Eq. , we find $$\left\langle{l_i}|P^\Gamma_\text{ss}\right\rangle=-\frac{\Gamma\left\langle l_i|C_0\right\rangle}{\lambda_i-\Gamma},$$ for $i\neq1$, and thus $$\label{NewSteadyState}
\ket{P^\Gamma_\text{ss}}=\ket{P_\text{ss}}-\sum_{i=2}^{D}\frac{\Gamma\left\langle l_i|C_0\right\rangle}{\lambda_i-\Gamma}\ket{r_i}.$$
We see that in addition to the stationary state of the process without resets, the resetting stationary state contains a contribution of the dynamical modes, weighted according to how significant they are in the evolution of the reset state in the original dynamics (the overlaps $\left\langle l_i|C_0\right\rangle$), and the reset rates magnitude relative to the corresponding eigenvalue $\lambda_i$. This agrees with intuition: since the eigenvalues are related to the lifetimes of the dynamical modes, if the average time between resets is larger than the lifetime of a mode, it will not make a significant contribution to the new steady state. We note that this equation also has the expected limits of $\ket{P^0_\text{ss}}=\ket{P_\text{ss}}$ and $\ket{P^\infty_\text{ss}}=\ket{C_0}$.
This result applies for processes with a finite number $D$ of states, but also to infinite-dimensional processes, provided that they possess a well-defined spectrum of eigenvalues with corresponding left and right eigenmodes. This latter fact will be illustrated in the next section with the example of Brownian motion.
We should also note that the result holds if the reset state $\ket{C_0}$ is replaced by a reset probability distribution $\ket{P_0}$ over configurations, giving the probability of reaching different states after a reset event. In this case, $\ket{C_0}$ in Eq. is simply replaced by the “mixed” reset state $\ket{P_0}$. This follows since the reset state is never referred to above as anything more than as a vector in state space.
Finally we mention the modification resulting from a lack of ergodicity, i.e., when $\mathcal{L}\ket{r_i}=0$ and $\bra{l_i}\mathcal{L}=0$ for more than one pair of states, causing an initial state dependence of the stationary state. Generically, we can still choose a basis such that $\bra{l_1}=\bra{-}$ and $\left\langle l_i|r_j\right\rangle=\delta_{ij}$ within this null eigenspace, allowing us to use nearly the same proof scheme as above. The reset rates break the zero eigenvalues degeneracy and provides a unique stationary state, with the remainder of the null eigenspace shifted to an eigenvalue of $-\Gamma$. In Eq. the unique steady state $\ket{P_{\text{ss}}}$ is replaced by the state-dependent steady state that would be reached from the reset state under the original dynamics.
Stationary current {#DetailedBalance}
------------------
The current associated with the stationary state $\ket{P^\Gamma_\text{ss}}$ of the reset process is defined, for any given link or transition $c\rightarrow c'$, by $$J^\Gamma_{c\rightarrow c'}=\left\langle c|P^\Gamma_{\text{ss}}\right\rangle{\mathcal{L}}^\Gamma_{c'c}-\left\langle c'|P^\Gamma_{\text{ss}}\right\rangle{\mathcal{L}}^\Gamma_{cc'},$$ where ${\mathcal{L}}^\Gamma_{ij}$ is the $(i,j)$ component of ${\mathcal{L}}^\Gamma$. Substituting the expression of the stationary state, found in Eq. , together with the expression of the generator ${\mathcal{L}}^\Gamma$, we can decompose the current into three parts as $$\begin{aligned}
J^\Gamma_{c\rightarrow c'}&=J_{c\rightarrow c'}
+\sum_{i=2}^{D}\frac{\Gamma\left\langle l_i|C_0\right\rangle}{\Gamma-\lambda_i}J^{i}_{c\rightarrow c'}\nonumber\\
&\quad+\Gamma\left(\left\langle c|P^\Gamma_{\text{ss}}\right\rangle{\delta}_{c'c_0}-\left\langle c'|P^\Gamma_{\text{ss}}\right\rangle{\delta}_{c_0c'}\right),
\label{eqcurr1}\end{aligned}$$ where $$J^{i}_{c\rightarrow c'}=\left\langle c|r_i\right\rangle{\mathcal{L}}_{c'c}-\left\langle c'|r_i\right\rangle{\mathcal{L}}_{cc'}.$$ The first term on the right-hand side of Eq. is the current of the original reset-free process, while the second is the weighted contribution of new currents coming from the non-stationary modes of the original process. Finally, the third term is the current coming from the reset transitions. From the signs appearing in the last term, we see that there is a current loop from all the states to the reset state $C_0$ and then back from $C_0$ to all other states, so that $C_0$ acts both as a sink and source.
From this result, it is clear that the reset process will violate the condition of detailed balance, i.e., $J^\Gamma\neq 0$, if the original process satisfies the condition of detailed balance, i.e., $J=0$. In this case, it is known that the original process has a real spectrum, which implies from our results that the spectrum of the reset process must also be real, even though $J^\Gamma\neq 0$. This shows that currents are not necessarily associated with complex spectra of the dynamical generator. In fact, detailed balance is only a sufficient condition for the generator spectrum to be real, not a necessary condition.
In principle, it is also possible to have a reset process satisfying detailed balance ($J^\Gamma=0$) if the original process violates detailed balance ($J\neq 0$). However, this is a rather peculiar case, requiring that the added reset transitions with rate $\Gamma$ exactly counterbalance all the non-zero currents arising in the reset-free process. In this case, the original process must again have a real spectrum in order for the spectrum of the reset process to be real.
Metastability {#Metastability}
-------------
We close this section by discussing the effect of resets on metastable states that arise when the evolution of the probability distribution $P(C,t)$ exhibits two distinct time scales: a fast evolution towards long-lived metastable states, followed by a slow relaxation to the final stationary state. These metastable states typically reside in a reduced subset of the full state space, called the *metastable manifold* (MM), with the later relaxation to the stationary state occurring within the MM (see, e.g., [@Macieszczak2016] and references therein for definitions and nomenclature that applies both to classical and quantum metastability).
Much work has been done on Markovian processes to understand metastability [@GaveauSchulman1987Meta; @Gaveau1998; @Bovier2002; @Gaveau2006; @larralde2005; @olivieri2005], based on the presence of large gaps in the spectrum of the master operator, which are necessary for the occurrence of distinct timescales. The MM in this context is understood to correspond to the reduced set of eigenmodes defined by these gaps, with the late time relaxation given by a projection of the master operator onto the MM.
To illustrate this phenomenon in the simplest way possible, let us consider a Markov process with a unique stationary state and a large gap between the second and third eigenvalues, i.e., $|\text{Re}(\lambda_2)|\ll|\text{Re}(\lambda_3)|$. The MM of long-lived states corresponds in this case to a one-dimensional manifold of linear combinations of the stationary state and $\ket{r_2}$, with the coefficients of $\ket{r_2}$ bounded by the maximum and minimum values of $\bra{{l_2}}$ on the space of probability distributions, i.e., the maximum and minimum components of this vector in the configuration basis $c_2^{\text{max}}$ and $c_2^{\text{min}}$. These two values define the so-called *extreme metastable states* (eMSs) on the boundary of the manifold: $$\begin{aligned}
\ket{\tilde{P}_1}&=\ket{P_\text{ss}}+c_2^{\text{max}}\ket{r_2}\nonumber\\
\ket{\tilde{P}_2}&=\ket{P_\text{ss}}+c_2^{\text{min}}\ket{r_2},
\label{ems}\end{aligned}$$ in terms of which we can write the stationary state as $$\begin{aligned}
\ket{P_\text{ss}}&=\frac{1}{\Delta c_2}\left(-{c}_{2}^{\rm min}\ket{\tilde{P}_1}+{c}_{2}^{\rm max}\ket{\tilde{P}_2}\right)\label{SteadyStateMMExpansion}\\
&=p_1^\text{ss}\ket{\tilde{P}_1}+p_2^\text{ss}\ket{\tilde{P}_2},\nonumber\end{aligned}$$ where $\Delta c_2={c}_{2}^{\rm max}-{c}_{2}^{\rm min}$. Note that $c_2^\text{min}\leq0$ as $\bra{l_2}$ is orthogonal to the stationary state which has purely positive components, so that the coefficients $p_1^\text{ss}$ and $p_2^\text{ss}$ in this expansion can be viewed as the probability weight of the two eMSs. Finally, we can construct an effective evolution on this subspace by projecting the master operator to find an effective master operator given in terms of $\lambda_2$ and the maximum and minimum components of $\bra{l_2}$ $${\mathcal{L}}_{\text{eff}}=\frac{-\lambda_{2}}{\Delta{c}_{2}}
\begin{pmatrix}
-{c}_{2}^{\rm max} & -{c}_{2}^{\rm min}\\
{c}_{2}^{\rm max} & {c}_{2}^{\rm min}
\end{pmatrix}.
\label{EffectiveMMDynamics}$$
The meaning of the above is the following. Consider the system starting in an initial state $\ket{P_0}$ (either a specific configuration or a probability over configurations). At some time $t$, the state of the system will read in terms of the spectrum of ${\mathcal L}$, $$\ket{P_0} = \ket{P_\text{ss}} + e^{t \lambda_2} \langle l_2 | P_0 \rangle \ket{r_2} + \sum_{i \geq 3}
e^{t \lambda_i} \langle l_i | P_0 \rangle \ket{r_i} .
\label{pot}$$ Due to the separation of timescales, for times $t$ such that $ 1/|\text{Re}(\lambda_3)|\ll t \ll 1/|\text{Re}(\lambda_2)|$, all but the first two terms in will be negligible, assuming that the overlap of the initial state with the modes $i\geq 3$ is small so that these terms are suppressed by the decaying modes for $i\geq 3$. Within these timescales, the initial state $\ket{P_0}$ evolves to a state in the one-dimensional MM, subsequently evolving within the MM and eventually reaching the unique stationary state for $t \gg 1/|\text{Re}(\lambda_2)|$, i.e., schematically $$\ket{P_0} \to p_1(t) \ket{\tilde{P}_1} + p_2(t) \ket{\tilde{P}_2} \to \ket{P_{\rm ss}} .$$ The evolution for $t \gg 1/|\text{Re}(\lambda_3)|$, prior to reaching the stationary state, is within the one-dimensional MM, as it corresponds to the evolution of $p_{1,2}(t) \geq 0$ in the linear combination above, with $p_1(t)+p_2(t)=1$, and is described by the effective generator .
We now add resets to this metastable dynamics, focusing for simplicity on the one-dimensional metastable manifold case. First, we note that we can rewrite Eq. as $$\ket{P^\Gamma_\text{ss}}=\ket{P_\text{ss}}-\sum_{i=2}^{D}\frac{\left\langle l_i|C_0\right\rangle}{\frac{\lambda_i}{\Gamma}-1}\ket{r_i}.$$ As a result, we see that, if we consider $\Gamma \approx |\lambda_2|\ll|\lambda_3|$, then the coefficients in the sum for the terms $i\geq3$ are small compared to the coefficient for $i=2$, so we can truncate to only the first two terms: $$\label{MetaTruncation}
\ket{P^\Gamma_\text{ss}}\approx\ket{P_\text{ss}}-\frac{\Gamma\left\langle l_2|C_0\right\rangle}{\lambda_2-\Gamma}\ket{r_2}.$$ Since the spectrum of the process with resets is simply a real shift by $\Gamma$, values on this scale preserve the gap in the spectrum required for metastability, as for $\Gamma\approx|\lambda_2|$ we still have $\text{Re}(\lambda_2)-\Gamma\ll \text{Re}(\lambda_3)-\Gamma$. Physically, this regime corresponds to the average time between resets being comparable to the lifetime of the metastable phases, in which the short timescale dynamical modes represented by $i\geq3$ are averaged out and make negligible contribution to the stationary state.
Using the definitions of the eMSs with the modified left eigenmodes from Sec. \[ResetSpec\], the eMSs of the model with resets are given by $$\begin{aligned}
\label{MetaeMS}
\ket{\tilde{P}^\Gamma_{1}}=&\ket{{P}^\Gamma_{\rm ss}}+{c}_{2}^{\Gamma,\rm max}\ket{{r}_{2}}\nonumber\\
\ket{\tilde{P}^\Gamma_{2}}=&\ket{{P}^\Gamma_{\rm ss}}+{c}_{2}^{\Gamma,\rm min}\ket{{r}_{2}},\end{aligned}$$ where, since the modification to the left eigenmodes is simply a shift by the flat state, the coefficients are now given by $${c}_{2}^{\Gamma,\rm min/max}={c}_{2}^{\rm min/max}-\frac{\Gamma\left\langle l_2|C_0\right\rangle}{\Gamma-\lambda_2}.$$ Substituting this and Eq. into Eq. , we see that the eMSs with reset are approximately equal to the original eMSs, as the effect of the modifications to the steady state and left eigenmodes cancel. Applying Eq. to the model with reset and using $\ket{\tilde{P}^\Gamma_{i}}\approx\ket{\tilde{P}_{i}}$, we thus obtain $$\begin{aligned}
\label{ResetMetastableDecomp}
\ket{P^\Gamma_\text{ss}}&\approx&\left(p_1^\text{ss}+\Delta p\right)\ket{\tilde{P}_1}+\left(p_2^\text{ss}-\Delta p\right)\ket{\tilde{P}_2}\nonumber\\
&=&\ket{P_{ss}}+\Delta p \ket{\tilde{P}_1}-\Delta p\ket{\tilde{P}_2},\nonumber\end{aligned}$$ where $$\Delta p = \frac{\Gamma\left\langle l_2|C_0\right\rangle}{\Delta c_2(\Gamma-\lambda_2)}.$$ We see that, depending on the overlap of the reset state with $\ket{l_2}$, this coefficient can cause a notable modification of the steady state mixture even for small $\Gamma$. This means physically that resets will make whichever eMS is closer to the reset state more likely to occur in the stationary state, as expected. Finally, the modified effective dynamics can be constructed simply by replacing the coefficients and eigenvalues with the new ones in Eq. .
Applications {#Applications}
============
We apply in this section our formula for the stationary state of the reset process for two exactly-solvable models. The applicability of this formula is obviously limited by the fact that it requires the full spectrum of the reset-free process. For this reason, we expect it to be more useful for approximating the stationary state than for calculating that state exactly, either by truncating the sum involved to a limited number of modes or by expanding the sum perturbatively in $\Gamma$. Moreover, while exact results can be hard to find, the formula can be useful numerically when applied to processes in which the resets break symmetries of the original, reset-free process. Such symmetries can indeed be used to diagonalise the original process for system sizes much larger than would otherwise be possible, with the resulting spectrum then being used in to derive the stationary state with resets. This is demonstrated later in the context of adding resets to a closed quantum system breaking time-reversal symmetry.
Totally asymmetric random walk {#AsymHop}
------------------------------
The first model that we consider is a particle hopping on a one-dimensional lattice of length $L$ with periodic boundary conditions, so the states are $\ket{x}$ with $x\in\{1,2,\ldots,L\}$ and $\ket{L+1}=\ket{1}$. We take the particle to hop only to the right with rate $\gamma$, so that the master operator is $${\mathcal{L}}=\gamma\sum_{x=1}^{L}\ket{x+1}\bra{x}-\gamma I.
\label{HoppingMastOp}$$ This operator is translation invariant and can be diagonalized by discrete Fourier transform, with left and right eigenvectors given by $$\ket{r_n}=\frac{1}{L}\sum_{x=1}^{L}e^{i\frac{2\pi n}{L}(x-1)}\ket{x}$$ and $$\bra{l_n}=\sum_{x=1}^{L}e^{-i\frac{2\pi n}{L}(x-1)}\bra{x},$$ respectively, where $n\in\{0,...,L-1\}$. As before, these eigenvectors are normalized such that $\left\langle l_n | r_m \right\rangle = \delta_{nm}$. Moreover, the eigenvalues are given by $$\lambda_n = \gamma(e^{-i2\pi n/L}-1).$$
We now add resets at rate $\Gamma$ onto the reset site $\ket{1}$. The new eigenvalues and eigenvectors can then be calculated exactly using the results of the previous section, with Eq. leading to $$\left\langle x|P_{\text{ss}}^\Gamma\right\rangle = \frac{\Gamma}{\gamma+\Gamma}S_L(x),$$ where $$S_L(x)=\frac{1}{L}\sum_{n=0}^{L-1}\frac{e^{i\frac{2\pi n}{L}x}}{e^{i\frac{2\pi n}{L}}-\frac{\gamma}{\gamma+\Gamma}}.$$ Considering the sum $S_L(x)$ on different sites, we find $$S_L(x+1)-\frac{\gamma}{\gamma+\Gamma}S_L(x)=\delta_{xL},$$ and so $S_L(x+1)=\gamma/(\gamma+\Gamma)S_L(x)$ when $x\neq L$. Consequently, $$\left\langle x|P_{\text{ss}}^\Gamma\right\rangle = \frac{\Gamma}{\gamma}{\left(\frac{\gamma}{\gamma+\Gamma}\right)}^{x}S_L(1).$$ At this point, rather than explicitly calculate the sum $S_L(1)$, we can just normalise the stationary state to find $$S_L(1)=\frac{1}{1-{\left(\frac{\gamma}{\gamma+\Gamma}\right)}^{L}},$$ thus giving $$\label{NewHoppingStationary}
P_\text{ss}^\Gamma(x)=\left\langle x|P_{\text{ss}}^\Gamma\right\rangle = \frac{\Gamma}{\gamma+\Gamma}\frac{1}{1-{\left(\frac{\gamma}{\gamma+\Gamma}\right)}^{L}}{\left(\frac{\gamma}{\gamma+\Gamma}\right)}^{x-1}.$$
This result is interesting because, while the infinite-size limit of the initial model neither has a stationary state or the ability to reach one (the spectral gap tends to zero), the infinite-size limit of the model with resets gains a gap of exactly $\Gamma$, with the corresponding stationary state given by the limit of Eq. , with probabilities $$\label{AsymHopSS}
P_\text{ss}^\Gamma(x) = \frac{\Gamma}{\gamma+\Gamma}{\left(\frac{\gamma}{\gamma+\Gamma}\right)}^{x-1},$$ where normalization can be checked via the geometric series. In this way, we see that resets localise the particle near the reset state, with a localisation length of $$\epsilon = \frac{1}{\ln\left(\frac{\gamma+\Gamma}{\gamma}\right)}.$$
Brownian motion {#BMReset}
---------------
We now show how to apply our results to continuous-state models by seeing them as the limit of a sequence of finite-dimensional models which have a discrete spectrum and well-defined eigenvectors. We consider for this purpose the reset Brownian motion in one dimension, first studied in [@evans2011] via a modified Fokker-Planck equation.
The master, or Fokker-Planck operator in this case, without reset is the Laplacian $${\mathcal{L}}p(x)=D\frac{d^2 p(x)}{d x^2}$$ on the real line, with $D$ as the diffusion constant and $p(x)$ the probability density. The spectrum of this operator is trivial, but it has no normalised stationary state, nor is it possible to define the orthogonality relation between left and right eigenfunctions. To address this problem, we restrict the system, as commonly done in physics, to the finite interval $[-L/2,L/2)$ with periodic boundary conditions. The spectrum of this restricted model is simply given by $$\lambda_n=-D\left(\frac{2\pi n}{L}\right)^2,\qquad n\in\mathbb{Z},$$ with equal right and left eigenmodes, due to the Hermiticity of ${\mathcal{L}}$, given by $$r_n(x)=\frac{1}{\sqrt{L}}e^{i\frac{2\pi n}{L}x},$$ and normalised according to $$\begin{aligned}
\left\langle r_n|r_m \right\rangle = \int_{-L/2}^{L/2}r^*_n(x)r_m(x)\, dx=\delta_{nm}.\end{aligned}$$
Since the spectrum exists and can be normalized appropriately, we may use our results of Sec. \[SecRes\]. Adding resets at a rate $\Gamma$ to the position 0, and defining the states $\ket{0}$ and $\bra{-}$ by $\left\langle g|0 \right\rangle=g(0)$ and $$\left\langle -|f \right\rangle=\left\langle l_0|f \right\rangle=\int_{-L/2}^{L/2}f(x)\, dx,$$ respectively, we find the new stationary state, which is now a probability density, to be given by the sum $$\begin{aligned}
p_{\text{ss}}^\Gamma(x)=\frac{\Gamma}{L}\sum_{n=-\infty}^{\infty}\frac{e^{i\frac{2\pi n}{L}x}}{D\left(\frac{2\pi n}{L}\right)^2+\Gamma}.\end{aligned}$$ Taking the infinite-size limit, we then find $$p_{\text{ss}}^\Gamma(x)=\Gamma\int_{-\infty}^{\infty}\frac{dk}{2\pi}\frac{e^{ikx}}{Dk^2+\Gamma},$$ which can be solved using residues to give $$p_{\text{ss}}^\Gamma(x)=\frac{1}{2}\sqrt{\frac{\Gamma}{D}}e^{-\sqrt{\frac{\Gamma}{D}}|x|}.$$ This agrees with the result of [@evans2011] and is similar to the distribution found for the discrete random walk.
Open quantum systems with resets {#OQS}
================================
We next consider adding resets to open quantum systems interacting with an environment. Under appropriate conditions on the timescales of the dynamics in the environment and the strength of the interactions, the environment can be suitably viewed as memoryless, allowing us to consider the system to be a quantum generalization of the Markovian systems considered earlier .
Theory {#secresqtheory}
------
We consider a quantum system in a Hilbert space ${\mathcal{H}}$ of dimension $\text{dim}({\mathcal{H}})$ with density matrix $\rho$, whose evolution is given by $$\frac{d\rho}{dt}={\mathcal{L}}(\rho),
\label{eqlindeq}$$ where $$\label{LindbladOp}
{\mathcal{L}}(\rho) = -i[H,\rho]+\sum_{j}\left[{J}_{j}\rho{J}_{j}^{\dagger}-\frac{1}{2}\{{J}_{j}^{\dagger}{J}_{j},\rho\}\right],$$ is the Lindblad master operator. Here $H$ is the Hamiltonian of the system and the jump operators ${J}_{j}$ mediate the system-bath interaction, providing coupling of the system to the surrounding environment.
Since ${\mathcal{L}}$ acts linearly on the density matrix, the evolution can be understood in terms of its eigenvalues and eigenmatrices. Let us denote the eigenvalues of ${\mathcal{L}}$ by ${\lambda}_{k}$ and order them such that $\text{Re}({\lambda}_{k})\geq\text{Re}({\lambda}_{k+1})$. As in the classical case, we have $\text{Re}(\lambda_k)\leq 0$, with $\lambda_1=0$ corresponding to the stationary state, due to the fact that ${\mathcal{L}}$ is completely positive and trace-preserving. Moreover, as ${\mathcal{L}}$ is in general not Hermitian, it has right and left eigenmatrices denoted by $${\mathcal{L}}(R_k)=\lambda_k R_k$$ and $${\mathcal{L}}^\dagger(L_k)=\lambda_k^* L_k,$$ respectively. These are normalised such that $$\text{Tr}({L}^\dagger_{k}{R}_{k'})={\delta}_{kk'}.$$ Generally, the stationary state ${\rho}_{\rm ss}$ is unique. Normalising it, as usual, by $\text{Tr}({\rho}_{\rm ss})=1$, we then have $L_1=\mathds{1}$. Defining $${c}_{k}=\text{Tr}[{L}^\dagger_{k}\rho(0)]$$ for an initial state $\rho(0)$, the system’s state at time $t$ is given by $$\label{Expansion}
\rho(t)={e}^{t{\mathcal{L}}}[\rho(0)]={\rho}_{\rm ss}+\sum_{k}{c}_{k}{e}^{t{\lambda}_{k}}{R}_{k}.$$
For the classical stochastic processes discussed in the previous sections, reset occurred with equal probability from every state at times distributed exponentially with rate $\Gamma$. We can construct a similar kind of reset dynamics for quantum open systems by adding jump operators $$J^\Gamma_i=\sqrt{\Gamma}\ket{\psi}\bra{\phi_i}$$ to the Lindbladian . Here $\ket{\phi_i}$ form a complete orthonormal basis, i.e., $\left\langle\phi_i|\phi_j\right\rangle=\delta_{ij}$, and $\ket{\psi}$ is the reset state. This modifies the Lindblad generator to $$\label{ModifiedLindblad}
{\mathcal{L}}_{\Gamma}(\rho) = {\mathcal{L}}(\rho)+\mathcal{V}_{\Gamma}(\rho)-\Gamma\rho,$$ where $$\mathcal{V}_{\Gamma}(\rho) = \Gamma\,\text{Tr}(\rho)\ket{\psi}\bra{\psi}.$$ To check that this construction has state-independent resets, as desired, we can consider the quantum jump Monte Carlo (QJMC) approach to simulating individual trajectories of the system’s evolution (see, e.g., [@plenio1998]). When the system undergoes a stochastic dissipative change (a “jump”), the probability of the change being a reset is $$P_\Gamma(\ket{\phi})\propto\sum_k^{\text{dim}(\mathcal{H})}\bra{\phi}{J^\Gamma_k}^\dagger J^\Gamma_k\ket{\phi}=\Gamma\left\langle\phi|\phi\right\rangle,$$ and is thus state independent as required.
To analyse the spectrum of the model with resets, we make the same assumptions as in Sec. \[ResetMarkov\]: we assume that the stationary state of the model without resets is unique, and that it is diagonalizable (i.e., there are no non-trivial Jordan blocks). The uniqueness of the stationary state implies $\text{Tr}(R_i)=0$ and hence $${\mathcal{L}}_\Gamma(R_i)=(\lambda_i-\Gamma)R_i$$ for $i\in\{2,...,\text{dim}({{\mathcal{H}}}^2)\}$. This shows, similarly to the classical case, that the $R_i$’s remain eigenmodes of the model with resets, with modified eigenvalues $$\lambda^\Gamma_i=\lambda_i-\Gamma.$$ For $i=1$, we have again $\lambda_1=0$ and the stationary state $R_1=\rho_{\rm ss}$, which we calculate below using an analogous method to that of Sec. \[ResetSpec\].
Next, we consider the corresponding left eigenmodes dual to the above. The adjoint equation is given by $${\mathcal{L}}_{\Gamma}^\dagger(\rho) = {\mathcal{L}}^\dagger(\rho)+\mathcal{V}_{\Gamma}^\dagger(\rho)-\Gamma\rho,$$ where $\mathcal{V}_{\Gamma}^\dagger(\rho) = \Gamma\bra{\psi}\rho\ket{\psi}I$. Note that the identity $I$ remains an eigenmode with eigenvalue 0, as expected. Inserting the original left eigenmodes, we find $${\mathcal{L}}_\Gamma^\dagger(L_k)=(\lambda^*_k-\Gamma)L_k+\Gamma\bra{\psi}L_k\ket{\psi}{I}.$$ Defining $$L^\Gamma_k=L_k+\frac{\Gamma\bra{\psi}L_k\ket{\psi}}{\lambda^*_k-\Gamma}I,$$ we see that, since the identity is annihilated by the adjoint operator, we have $${\mathcal{L}}_\Gamma^\dagger\left(L^\Gamma_k\right)=(\lambda^*_k-\Gamma)L^\Gamma_k,$$ demonstrating that the new left eigenmodes are $L^\Gamma_k$. It can be checked that $\text{Tr}({L^\Gamma_i}^\dagger R_j)=\delta_{ij}$ for $j\neq 1$ and for all $i$, as expected, since $\text{Tr}(R_j)=0$. Finally, requiring $\text{Tr}({L^\Gamma_i}^\dagger \rho^\Gamma_{ss})=\delta_{i1}$ for the new stationary state $\rho^\Gamma_{ss}$ gives the expansion coefficients of that state in the original right eigenmode basis as $$\label{SteadyStateResets}
\rho_{ss}^\Gamma=\rho_{ss}^0-\sum_{j=2}^{\text{dim}({\mathcal{H}})^2}\frac{\Gamma\bra{\psi}L^\dagger_j\ket{\psi} }{\lambda_j-\Gamma}R_j,$$ which, as can be checked, gives ${\mathcal{L}}_\Gamma\rho_{ss}^\Gamma=0$.
The practical applicability of these results depends, as in the classical case, on the system studied and whether, in particular, it has symmetries simplifying the spectral problem. There is an additional benefit in the case of closed quantum systems coming from the fact that the Lindblad equation reduces in that case to the von Neumann equation, allowing for much larger system sizes to be studied compared to a direct spectral solution of the open quantum problem. We demonstrate this next.
Coherent hopping on a chain {#CoherentHop}
---------------------------
The first example that we consider to illustrate our results is a simple model of coherent hopping on a closed periodic chain, described by the Hamiltonian $$\begin{aligned}
H=&\gamma\sum_{x=1}^{L-1} \left( \ket{x+1}\bra{x}+\ket{x}\bra{x+1} \right)
\nonumber \\
&+ \gamma \left( \ket{1}\bra{L}+\ket{L}\bra{1} \right),\end{aligned}$$ and no jump operators, so that the reset-free system is closed. The dynamics generated by this Hamiltonian is similar to that considered in . The main difference is that we consider state-independent resets whereas consider resets induced by continuously measuring the system’s state on a particular site, leading to a single jump operator proportional to the projective measurement on that site.
![(a,b) Magnitude of the stationary density matrix on a log scale. (c,d) Occupation probability for each site of the chain compared with the probability of a classical hopping model with exponential distribution. Plots are for (a,c) $\Gamma=1$ and (b,d) $\Gamma=5$.[]{data-label="CoherentHoppingFig"}](CoherentHopping.pdf){width="1\linewidth"}
For closed quantum systems, the spectrum of the corresponding Lindblad equation is given by all possible outer products of the eigenvectors and corresponding differences of eigenvalues of $H$. Given $$H\ket{e_k}=\lambda_k\ket{e_k},$$ the matrices $R_{kk'}=\ket{e_k}\bra{e_{k'}}$ then provide the right eigenmodes of the Lindblad equation $${\mathcal{L}}(R_{kk'})=-i(\lambda_{k}-\lambda_{k'})R_{kk'},$$ with the corresponding left eigenmodes also given by $L_{kk'}=\ket{e_k}\bra{e_{k'}}$. In this context, our results of Sec. \[secresqtheory\] can be modified to handle degenerate modes with ${\mathcal{L}}(R_i)=0$, as is the case for closed quantum systems, analogously to the procedure described at the end of Sec. \[StatStateResets\]. The resulting modification to Eq. for coherent dynamics contains a sum over the non-zero eigenvalue modes as before, with the unique stationary density matrix replaced by a diagonal matrix in the energy eigenbasis of probabilities for the reset state to be measured in each eigenstate of the Hamiltonian.
Using the outer product structure of the eigenstates, we can rewrite this sum more compactly as the matrix product $$\rho_{ss}^{\Gamma}=E\Lambda^\Gamma E^\dagger,$$ where $E$ is the matrix of eigenvectors defined by $HE=E\Lambda$ with $\Lambda_{ii}=\lambda_i$, and $\Lambda^\Gamma$ has elements defined by $$\Lambda^\Gamma_{ij}=\frac{\Gamma\left\langle\psi|e_j\right\rangle\left\langle e_i|\psi\right\rangle}{\Gamma+i(\lambda_i-\lambda_j)}.$$ These matrices can be efficiently constructed numerically and used to calculate the stationary state of a closed system after the addition of resets.
For the coherent hopping model we have $$\ket{e_n}=\frac{1}{\sqrt{L}}\sum_{x=1}^{L}e^{i\frac{2\pi n}{L}(x-1)}\ket{x},$$ and $\lambda_k=2\Omega \text{cos}(2\pi n/L)$. Choosing resets to the state $\ket{0}$ for a chain of $L=2001$ and hopping rate $\gamma=1000\gg\Gamma$, the numerically calculated stationary state is given for two different reset rates in Fig. \[CoherentHoppingFig\]. As expected, the magnitude of the components of the stationary state decay away from the reset state (Fig. \[CoherentHoppingFig\]). In Fig. \[CoherentHoppingFig\], we plot the probability for the system to be found in each site against the distance from the reset state. For comparison, we also plot the probability for a classical random walk with resets, with parameters fixed by equating the probability of the two distributions at the reset state. We see that the coherent dynamics allows particles to move away from the reset state at a faster rate than the dissipative dynamics, leading to a crossing point beyond which there is a higher probability of locating the particle in the coherent model compared to the dissipative model.
Open quantum Ising model
------------------------
We consider as a second example the transverse field Ising model $$\label{Hamiltonian}
H = \Omega\sum_{j=1}^{N}{S}^{(j)}_{x}+V\sum_{i=1}^{N}{S}^{(j)}_{z}{S}^{(j+1)}_{z}$$ with periodic boundary conditions, where the spin operators are ${S}^{(j)}_{\alpha}=\frac{1}{2}{\sigma}^{(j)}_{\alpha}$ with $\alpha=\{x,y,z\}$ and the jump operators are given by $$\label{Jumps}
{J}_{j}=\sqrt{\kappa} \, {S}^{(j)}_{-}=\sqrt{\kappa} \, ({S}^{(j)}_{x}-i{S}^{(j)}_{y}).$$ Unlike the models previously considered, this system is not exactly solvable without resets; however, it possesses a translation symmetry, which can be used with Eq. to numerically diagonalise the model for larger system sizes than if we simply tried to diagonalize the symmetry-lacking Lindblad equation with resets.
An interesting feature of this model is the presence of metastability located in a region around a crossover of the stationary properties [@Rose2016]. This metastability takes the form of a decomposition of the system’s state after a long evolution into a linear combination of a paramagnetic phase and ferromagnetic phase on either side of the crossover. This is followed by an eventual relaxation to a particular mixture of these two phases. We thus use this model to study the effect of adding resets to a model with metastability explicitly, comparing the quantum generalization of the results from Sec. \[Metastability\] with the stationary states given by Eq. .
We start by studying the reset rate dependence of the system’s magnetisation in the $z$ direction: $$\begin{aligned}
M=\frac{1}{N}\sum_{i=1}^{N}S_z^i.\end{aligned}$$ Considering a system of $N=7$ spins, in Fig. \[OpenIsingFig\] we plot the magnetisation of both the full reset stationary state given by Eq. (solid lines) and the approximate decomposition of the reset stationary state into the original metastable phases given by Eq. (dashed lines). This is done for two different reset states, $\ket{\psi_1}=\ket{\uparrow\uparrow\uparrow\uparrow\uparrow\uparrow\uparrow}$ and $\ket{\psi_2}=\ket{\uparrow\downarrow\uparrow\downarrow\uparrow\downarrow\uparrow}$, both of which have a high probability of evolving into the paramagnetic state after a time in the metastable regime.
Without resets the stationary state is dominated by the ferromagnetic phase. However, for both reset states, there is a larger probability of evolving into the paramagnetic phase than the ferromagnetic phase on metastable timescales. When the reset rate is increased, we expect the stationary state to become more biased towards the paramagnetic phase. This is clearly demonstrated in Fig. \[OpenIsingFig\], with strong agreement between the exact result and the approximation for reset rates up to the order of $|\lambda_2|$. For $\Gamma$ beyond this scale, metastability is lost and the approximation fails, with the magnetisation approaching that of the reset states for large $\Gamma$.
To quantify the agreement, we show in Fig. \[OpenIsingFig\] the trace distance between the truncated metastable state, as given by Eq. , and the full reset stationary state, as given by Eq. . We can see that this distance is close to zero up until $\Gamma=|\lambda_2|$, after which it increases rapidly, demonstrating a strong accuracy of Eq. when the average time between resets is equal to or longer than the metastable timescale, as assumed in Sec. \[Metastability\]. This change in the stationary magnetisation for smaller reset rates corresponds directly to the changing mixture of metastable phases in the stationary state. This is seen in Fig. \[OpenIsingFig\], which shows a higher probability of the system being found in the paramagnetic phase as $\Gamma$ is increased. The same behavior can also be seen at the trajectory level in the two plots of Fig. \[OpenIsingFig\], which show sample trajectories without resets and with resets to the state $\ket{\uparrow\uparrow\uparrow\uparrow\uparrow\uparrow\uparrow}$ at a rate $\Gamma$, respectively. As expected, we see that resets induce more periods of paramagnetic phase dropping back into the ferromagnetic phase.
![(a) Magnetisation in the $z$ direction of the stationary state (solid) compared with the metastable approximation (dashed) as a function of the reset rate $\Gamma$. (b) Probability of the two extreme metastable states as a function of $\Gamma$. (c) Trace distance between the stationary state and the metastable approximation as a function of $\Gamma$. (d-e) Sample trajectories of the magnetisation over time for $\Gamma=0$ (d) and $\Gamma = 0.05$ (e). Purple lines are for resets to the “all up” state $\psi_1$, while blue lines are for resets to the “alternating” state $\psi_2$. Arrows indicate the point where $\Gamma = |\lambda_2|$.[]{data-label="OpenIsingFig"}](OpenIsingResets.pdf){width="1\linewidth"}
We note that it may be possible to conduct an experiment to study the effect of resets on this Ising model (and similar many-body spin models) using ultra-cold atoms confined in optical latices [@Muller2008; @Saffman2010; @Viteau2011; @Lee2012]. In such experiments, the $\ket{\downarrow}$-state is associated with the atomic ground state, while the $\ket{\uparrow}$-state is represented by a Rydberg nS-state. These states are coupled coherently by laser, leading to Rabi oscillations at a frequency $\Omega$ and a detuning $\Delta$ relative to the energy difference between the two states. Excited atoms at lattice sites with position $r_i$ and $r_j$ interact via a van der Waals potential $V_{ij}=C_6/|r_i-r_j|^6$, where $C_6$ is the dispersion coefficient characterizing the interaction strength. Altogether this gives a Hamiltonian of the form $$H = \Omega\sum_iS_x^i+\Delta\sum_iS_z^i+\frac{1}{2}\sum_{i\neq j}V_{ij}n_in_j,$$ where $n_i=1/2I+S^i_z$. For sufficiently large lattice spacing $a$, the interaction decays so rapidly that it can be approximated as a nearest-neighbour interaction. Applying a laser detuning of the form $\Delta=-C_6/a^6$ then leads for a periodic chain to the Ising Hamiltonian , up to an overall energy shift that can be discarded. Dissipation occurs naturally via photon-emitted decay of the Rydberg states.
To simulate resets in this system, we can force it into a high magnetisation state at random intervals determined externally from an exponential distribution. While it is difficult in practice to place the system in a specific pure state with high magnetisation reliably, the above results simply generalize to a probability distribution of pure reset states (i.e., a reset “density matrix”). If such a density matrix has a large positive expectation value for the magnetisation, evolution after reset will have a high probability of leading to the paramagnetic state on the metastable timescale, resulting in similar observations to those of Fig. \[OpenIsingFig\]. These resets could be implemented via a strong laser pulse, such that the system can be momentarily approximated as non-interacting and the pulse modelled as instantaneous.
Conclusion
==========
We have developed in this work a general spectral approach to investigate the properties of Markov processes that are reset to a fixed state at random exponentially distributed times. Our main result shows that the spectrum of the generator of a reset Markov process is globally shifted by the reset rate compared with the spectrum of the corresponding reset-free process, except for the stationary mode, which stays at zero. We have also provided an explicit formula for the stationary distribution of the reset process, based on the spectrum of the reset-free process.
This spectral approach can be applied not only to classical stochastic processes but also, as we have shown, to closed and open quantum systems modelled by Lindblad-type equations. In both cases, the approach provides a natural way to study how resets can create a stationary state by opening a gap in the spectrum and how it affects metastable states. This was illustrated using various classical and quantum processes, including reset Brownian motion and the transverse field Ising model reset to a paramagnetic or ferromagnetic state.
For future work, it would be interesting to develop a similar approach for the large deviations of reset processes, based on the spectrum of the tilted generator [@meylahn2015b; @Harris2017]. There does not seem to be, a priori, a direct extension of our results for this generator, as the basic property used to prove our results, namely, that the non-stationary eigenstates have zero norm, does not hold in general for the eigenstates of the tilted generator. However, it might be possible to obtain partial results when the addition of reset to the master operator mixes only a small subset of the non-reset spectrum.
There also remains much work to be done on quantum systems, for which resets can be induced either by measurements, as in , or through other external perturbations. Some results on these systems have also been obtained very recently in [@mukherjee2018] for the stationary state of closed quantum systems with reset, which form, as we have seen, a specific case of the quantum systems considered here.
This work was supported by ERC under the EU Seventh Framework Programme (Grant No. FP/2007-2013), the ERC project ESCQUMA (Grant No. 335266), and by EPSRC (Grants No. EP/M014266/1 and No. EP/R04340X/1). H.T. is supported by NRF South Africa (Grant No. 96199) and by Stellenbosch University. I.L. gratefully acknowledges funding through the Royal Society Wolfson Research Merit Award. This work was also supported in part by the International Centre for Theoretical Sciences (ICTS) during a visit for participating in the program “Large deviation theory in statistical physics: Recent advances and future challenges” (Code: ICTS/Prog-ldt/2017/8).
[50]{} 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
10.1103/RevModPhys.83.81) [****, ()](\doibase 10.1103/PhysRevLett.106.160601) [****, ()](http://stacks.iop.org/1751-8121/44/i=43/a=435001) [****, ()](http://stacks.iop.org/1751-8121/46/i=18/a=185001) [****, ()](http://stacks.iop.org/1751-8121/47/i=28/a=285001) [****, ()](\doibase 10.1103/PhysRevLett.116.170601) [****, ()](\doibase 10.1103/PhysRevLett.118.030603)[****, ()](\doibase 10.1103/PhysRevE.59.4945)[****, ()](\doibase 10.1103/PhysRevLett.112.220601)[****, ()](\doibase 10.1103/PhysRevE.93.062411)[****, ()](http://stacks.iop.org/1367-2630/18/i=3/a=033006)[****, ()](\doibase 10.1103/PhysRevE.91.052131)[****, ()](\doibase 10.1103/PhysRevLett.113.220602)[****, ()](\doibase 10.1103/PhysRevE.92.062148)[****, ()](\doibase 10.1088/1751-8121/aa5734)[ ()](http://arxiv.org/abs/1801.09909), [****, ()](\doibase 10.2307/1427020)[****, ()](\doibase https://doi.org/10.1016/0020-0190(93)90029-9)[****, ()](\doibase 10.1080/15326349708807425)[****, ()](\doibase
http://dx.doi.org/10.1016/S0898-1221(00)00234-0)[****, ()](\doibase 10.1239/jap/1101840567)[****, ()](\doibase 10.1137/100796352)[****, ()](\doibase 10.1239/jap/1389370093)[****, ()](\doibase 10.1007/s10955-015-1336-4)[****, ()](\doibase
10.1140/epjb/e2017-80348-4)[****, ()](\doibase 10.1103/PhysRevE.91.012113)[****, ()](\doibase 10.1088/0305-4470/20/10/031)[****, ()](\doibase 10.1063/1.532394)[****, ()](\doibase 10.1103/PhysRevE.73.036124)[****, ()](\doibase 10.1103/PhysRevLett.120.040502)[****, ()](\doibase 10.1088/1751-8121/aa5191)[****, ()](\doibase
10.1103/PhysRevLett.116.240404)[****, ()](\doibase 10.1103/PhysRevA.85.043620)[****, ()](\doibase 10.1103/PhysRevE.94.052132)in @noop [**]{}, Vol. , (, )[****, ()](\doibase 10.1007/s002200200609)[****, ()](\doibase 10.1103/PhysRevLett.94.160201)[**](\doibase 10.1017/CBO9780511543272), , Vol. (, , )[****, ()](\doibase 10.1007/BF01608499)@noop [**]{}, ed., Complexity (, )[**](\doibase 10.1093/acprof:oso/9780199213900.001.0001) (, )@noop [****, ()]{}[****, ()](http://stacks.iop.org/1367-2630/10/i=9/a=093009)[****, ()](\doibase 10.1103/RevModPhys.82.2313)[****, ()](\doibase
10.1103/PhysRevLett.107.060402)[****, ()](\doibase 10.1103/PhysRevLett.108.023602)[ ()](https://arxiv.org/abs/1806.00019),
[^1]: The left eigenvectors may not be calculable in the case where there are Jordan blocks. The right eigenvectors and their corresponding eigenvalues are however the same and the stationary state can still be constructed using an alternative approach based on the renewal representation of trajectory ensembles.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
There has been renewed interest in recent years in McKinsey and Tarski’s interpretation of modal logic in topological spaces and their proof that S4 is the logic of any separable dense-in-itself metric space. Here we extend this work to the modal mu-calculus and to a logic of tangled closure operators that was developed by Fernández-Duque after these two languages had been shown by Dawar and Otto to have the same expressive power over finite transitive Kripke models. We prove that this equivalence remains true over topological spaces.
We establish the finite model property in Kripke semantics for various tangled closure logics with and without the universal modality $\forall$. We also extend the McKinsey–Tarski topological ‘dissection lemma’. These results are used to construct a representation map (also called a d-p-morphism) from any dense-in-itself metric space $X$ onto any finite connected locally connected serial transitive Kripke frame.
This yields completeness theorems over $X$ for a number of languages: (i) the modal mu-calculus with the closure operator $\Diamond$; (ii) $\Diamond$ and the tangled closure operators $\langle t \rangle$; (iii) $\Diamond,\forall$; (iv) $\Diamond,\forall,\langle t \rangle$; (v) the derivative operator $\langle d \rangle$; (vi) $\langle d \rangle$ and the associated tangled closure operators $\langle dt \rangle$; (vii) $\langle d \rangle,\forall$; (viii) $\langle d \rangle,\forall,\langle dt \rangle$. Soundness also holds, if: (a) for languages with $\forall$, $X$ is connected; and (b) for languages with $\langle d \rangle$, $X$ validates the well known axiom $\mathrm{G}_1$. For countable languages without $\forall$, we prove strong completeness. We also show that in the presence of $\forall$, strong completeness fails if $X$ is compact and locally connected.
author:
- 'Robert Goldblatt[^1] and Ian Hodkinson[^2]'
bibliography:
- 'paper.bib'
date: 6 March 2016
title: 'Spatial logic of modal mu-calculus and tangled closure operators'
---
Introduction
============
Modal logic can be given semantics over topological spaces. In this setting, the modality $\di$ can be interpreted in more than one way. The first and most obvious way is as *closure.* Writing $\sem\varphi$ for the set of points (in a topological model) at which a formula $\varphi$ is true, $\sem{\di\varphi}$ is defined to be the *closure* of $\sem\varphi$, so that $\di\varphi$ holds at a point $x$ if and only if every open of $x$ contains a point $y$ satisfying $\varphi$. Then, $\bo$ becomes the interior operator: $\sem{\bo\varphi}$ is the interior of $\sem\varphi$. Early studies of this semantics include [@Tang38; @Tar38; @McKi42; @McKiT44; @McKiT46].
In a seminal result, McKinsey and Tarski [@McKiT44] proved that the logic of any given separable[^3] space in this semantics is S4: it can be axiomatised by the basic modal Hilbert system $\axK$ augmented by the two axioms $\bo \varphi\to \varphi$ (T) and $\bo \varphi\to\bo\bo \varphi$ (4).
Motivated perhaps by the current wide interest in spatial logic, a wish to present simpler proofs in ‘modern language’, growing awareness of the work of particular groups such as Esakia’s and Shehtman’s, or involvement in new settings such as dynamic topology, interest in McKinsey and Tarski’s result has revived in recent years. A number of new proofs of it have appeared, some for specific spaces or embodying other variants [@Mints:S4no1; @GuramMai02:S4; @Aiello-vB-Guram:space; @Mints05aproof; @Slavnov05; @Lando-Sarenac:11; @Hodk:mckt_R]. Very recently, strong completeness (every countably infinite S4-consistent set of modal formulas is satisfiable in every space) was established by Kremer [@Kremer2010:stro].
In this paper, we seek to extend McKinsey and Tarski’s theorem to more powerful languages. We will extend the modal syntax in two separate ways: first, to the mu-calculus, which adds least and greatest fixed points to the basic modal language, and second, by adding an infinite sequence of new modalities $\di_n$ of arity $n$ $(n\geq1)$ introduced in the context of Kripke semantics by Dawar and Otto [@DO09]. The semantics of $\di_n$ is given by the mu-calculus formula $$\di_n(\vec\varphi n)\equiv\nu q\bigwedge_{1\leq i\leq n}\di(\varphi_i\wedge q),$$ for a new atom $q$ not occurring in $\vec\varphi n$. The order and multiplicity of arguments to a $\di_n$ is immaterial, so we will abbreviate $\di_n(\vec\gamma n)$ to $\dit\{\vec\gamma n\}$. Fernández-Duque used this to give the modalities topological semantics, dubbed them *tangled closure modalities* (this is why we use the notation $\dit$), and studied them in [@FD:SL11; @FD:ijcai11; @FD:apal12; @FD:jsl12b].
Dawar and Otto [@DO09] showed that, somewhat surprisingly, the mu-calculus and the tangled modalities have exactly the same expressive power over finite Kripke models with transitive frames. We will prove that this remains true over topological spaces. So the tangled closure modalities offer a viable alternative to the mu-calculus in both these settings.
We go on to determine the logic of an arbitrary space $X$ in these languages. We will show that in the mu-calculus, the logic of $X$ is axiomatised by a system called $\axS4\mu$ comprising Kozen’s basic system for the mu-calculus augmented by the S4 axioms, and the tangled logic of $X$ is axiomatised by a system called $\axS4t$ similar to one in [@FD:ijcai11]. We will establish strong completeness for countable sets of formulas.
We will also consider the extension of the tangled language with the *universal modality,* ‘’. (Earlier work on the universal modality in topological spaces includes [@Sheh:everywhere99; @LucBry11].) This language can express connectedness: there is a formula $\axC$ valid in precisely the connected spaces. Adding this and some standard machinery for to the system $\axS4t$ gives a system called ‘$\axS4t.\axU\axC$’. We will show that every $\axS4t.\axU\axC$-consistent formula is satisfiable in every space. Thus, the logic of an arbitrary connected space is $\axS4t.\axU\axC$. We also show that strong completeness fails in general, even for the modal language plus the universal modality.
A second and more powerful spatial interpretation of $\di$ is as the *derivative operator.* Following tradition, when considering this interpretation we will generally write the modal box and diamond as $\bod$ and $\did$. In this interpretation, $\sem{\did\varphi}$ is defined to be the set of *strict limit points* of $\sem\varphi$: so $\did\varphi$ holds at a point $x$ precisely when every open of $x$ contains a point $y\neq x$ satisfying $\varphi$. The original closure diamond is expressible by the derivative operator: $\di\varphi$ is equivalent in any topological model to $\varphi\vee\did\varphi$, and $\bo\varphi$ to $\varphi\wedge\bod\varphi$. So in passing to $\did$, we have not reduced the power of the language.
Already in [@McKiT44 Appendix I], McKinsey and Tarski discussed the derivative operator and asked a number of questions about it. It has since been studied by, among others, Esakia and his Tbilisi group ([@Esa04:apal; @GEG05], plus many other publications), Shehtman [@Sheh:d90; @Sheh:hab], Lucero-Bryan [@LucBry11], and Kudinov–Shehtman [@KudShe14], section 3 of which contains a survey of results.
In the derivative semantics, determining the logic of a given space is not a simple matter, for the logic can vary with the space. As McKinsey and Tarski observed, $\did\big((x\wedge\did\neg x)\vee(\neg x\wedge\did x)\big)\leftrightarrow\did x\wedge\did\neg x$ is valid in $\R^2$ but not in $\R$. This formula is valid in the same topological spaces as the formula $\axG_1$, where for each integer $n\geq1$, $$\axG_n=\Big(\bod\bigvee_{0\leq i\leq n}\bo Q_i\Big)
\to\bigvee_{0\leq i\leq n}\bod\neg Q_i.$$ Here, $\Vec pn$ are pairwise distinct atoms, and for $i=0,\ldots,n$, $$Q_i=p_i \land \bigwedge_{i\ne j\leq n}\neg p_j.$$ In [@Sheh:d90], Shehtman proved that the logic of $\R^n$ for finite $n\geq 2$ is $\axK\axD4\axG_1$, axiomatised by the basic system $\axK$ together with the axioms $\did\top$ (D), $\bod p\to\bod\bod p$ (4), and $\axG_1$. The logic of $\R$ was shown by Shehtman [@Sheh:hab] and Lucero-Bryan [@LucBry11] to be $\axK\axD4\axG_2$. The logic of every separable zero-dimensional space (such as ${\mbox{\(\mathbb Q\)}}$ and the Cantor space) is just $\axK\axD4$ [@Sheh:d90], the smallest possible logic of a space in the derivative semantics. [@GBezLucBry12] proves that there are continuum-many logics of subspaces of the rationals in the language with $\bod$.
It is plain that $\axG_1\vdash \axG_2\vdash \axG_3\vdash\cdots$, so the logics $\axK\axD4\axG_1\supseteq \axK\axD4\axG_2\supseteq\cdots$ form a decreasing chain, and by [@LucBry11 corollary 3.11], its intersection is $\axK\axD4$. Shehtman [@Sheh:d90 problem 1] asked if $\axK\axD4\axG_1$ is the largest possible logic of a space in the derivative semantics.
In this paper, we answer Shehtman’s question affirmatively: every $\axK\axD4\axG_1$-consistent formula of the language with $\did$ is satisfiable in every space. Thus, the logic of every space that validates $\axG_1$ is exactly $\axK\axD4\axG_1$. We also establish strong completeness for such spaces.
Adding the tangled closure operators, we prove similarly that the logic of every space that validates $\axG_1$ is axiomatised by $\axK\axD4\axG_1t$ (including the tangle axioms). We also prove strong completeness.
Further adding the universal modality, we show similarly that $\axK\axD4\axG_1t.\axU\axC$ (and $\axK\axD4\axG_1.\axU\axC$ if the tangle closure operators are dropped) axiomatises the logic of every connected space that validates $\axG_1$. Strong completeness fails in general, as a consequence of the proof that it already fails for the weaker language with $\bo$ and .
The reader can find a summary of our results in table \[table1\] in section \[sec:the end\].
Our proof works in a fairly familiar way, similar in spirit to McKinsey and Tarski’s original argument in [@McKiT44] — indeed, we use some results from that paper. There are three main steps.
1. We establish the *finite model property* for the various logics, in Kripke semantics. This work may be of independent interest: earlier related results were proved in [@Sheh:d90; @FD:ijcai11].
2. We then prove a topological theorem that establishes Tarski’s ‘dissection lemma’ [@Tar38 satz 3.10], [@McKiT44 theorem 3.5] and a variant of it.
3. These topological results are used to construct a map from an arbitrary space onto any finite connected $\axK\axD4\axG_1$ Kripke frame, that preserves the required formulas.
Putting the three steps together proves completeness for all the languages, which is then lifted by a separate argument to strong completeness for languages without .
It can be seen that our results concern the logic of each individual space within a large class of spaces (the spaces), rather than the logic of a large class of spaces, or of particular spaces such as $\R$. This is as in [@McKiT44]. We do not assume separability, we consider languages that have not previously been much studied in the topological setting, and we obtain some results on strong completeness, a matter that has only recently been investigated in this setting.
The paper is divided into two parts of roughly equal length. Part 1 is devoted to proving the finite model property in Kripke semantics for the logics of concern in the paper. Part 2 covers topology and spatial completeness results, and can be read independently by taking the results of part 1 on trust. The short section \[sec:basics 1\] preceding part 1 contains foundational material needed in both parts.
Basic definitions {#sec:basics 1}
=================
In this section, we lay out the main definitions, notation, and some basic results.
Notation for sets and binary relations
--------------------------------------
Let $X,Y,Z$ be sets. We let $\wp(X)$ denote the power set (set of all subsets) of $X$. We write $X\setminus Y$ for $\{x\in X:x\notin Y\}$. Note that $(X\cap Y)\setminus Z=X\cap(Y\setminus Z)$, so we may omit the parentheses in such expressions. For a partial function $f:X\to Y$, we let $\dom f$ denote the domain of $f$, and $\rng f$ its range.
A *binary relation* on a set $W$ is a subset of $W\times W$. Let $R$ be a binary relation on $W$. We write any of $R(w_1,w_2)$, $Rw_1w_2$, and $w_1Rw_2$ to denote that $(w_1,w_2)\in R$. We say that $R$ is *reflexive* if $R(w,w)$ for all $w\in W$, and *transitive* if $R(w_1,w_2)$ and $R(w_2,w_3)$ imply $R(w_1,w_3)$. We write $R^*$ for the *reflexive transitive closure* of $R$: the smallest reflexive transitive binary relation that contains $R$. We also write $$\begin{array}{cclcl}
R^{-1}&=&\{(w_2,w_1)\in W\times W:R(w_1,w_2)\},
\\
R^\circ&=&\{(w_1,w_2)\in W\times W: R(w_1,w_2)\wedge R(w_2,w_1)\}
&=&R\cap R^{-1},
\\
R^\bullet&=&\{(w_1,w_2)\in W\times W:R(w_1,w_2)\wedge\neg R(w_2,w_1)\}
&=&R\setminus R^{-1}.
\end{array}$$
For $w\in W$, we let $R(w)$ denote the set $\{w'\in W:R(w,w')\}$, sometimes called the set of *$R$-successors* or *$R$-alternatives* of $w$. For $W'\subseteq W$, we write $R\restriction W'$ for the binary relation $R\cap(W'\times W')$ on $W'$.
We write $\R$ for the set of real numbers, On for the class of ordinals, and $\omega$ for the first infinite ordinal.
Kripke frames {#ss:kripke frames}
-------------
A *(Kripke) frame* is a pair $\c F=(W,R)$, where $W$ is a non-empty set of ‘worlds’ and $R$ is a binary relation on $W$. We attribute properties to a frame by the usual extrapolation from the frame’s components. So, we say that $\c F$ is *finite* if $W$ is finite, *reflexive* if $R$ is reflexive, and *transitive* if $R$ is transitive. Two frames are said to be *disjoint* if their respective sets of worlds are disjoint. And so on.
A *root of $\c F$* is an element $w\in W$ such that $W=R^*(w)$. Roots of a frame may not exist, nor be unique when they do. We say that $\c F$ is *rooted* if it has a root. At the other end, an element $w\in W$ is said to be *$R$-maximal* if $R^\bullet(w)=\emptyset$. Such an element has no ‘proper’ $R$-successors, of which it is not itself an $R$-successor.
A *subframe* of $\c F$ is a frame of the form $\c F'=(W',R\restriction W')$, for non-empty $W'\subseteq W$. It is simply a sub of $\c F$ in the usual model-theoretic sense. We call $\c F'$ the *subframe of $\c F$ based on $W'$.* We say that $\c F'$ is a *generated* or *inner subframe* of $\c F$ if $R(w)\subseteq W'$ for every $w\in W'$ — equivalently, $R\restriction W'=R\cap(W'\times W)$. For $w\in W$, we write:
- $\c F(w)$ for the subframe $(R(w),R\restriction R(w))$ of $\c F$ based on $R(w)$,
- $\c F^*(w)$ for the subframe $(R^*(w),R\restriction R^*(w))$ of $\c F$ generated by $w$.
For an integer $n\geq1$, we say that $\c F$ is *$n$-connected* if it is not the union of $n+1$ pairwise disjoint generated subframes (recall that subframes are non-empty), *connected* if it is 1-connected, and *locally $n$-connected* if for each $w\in W$, the subframe $\c F(w)$ is $n$-connected. Note that $\c F$ is $n$-connected iff the equivalence relation $(R\cup R^{-1})^*$ on $W$ has at most $n$ equivalence classes. Every rooted frame is connected. Connectedness will be discussed in more detail in section \[sec:path conn\].
Fixed points {#ss:lfp}
------------
Let $X$ be a set and $f:\wp(X)\to\wp(X)$ be a map. We say that $f$ is *monotonic* if $f(S)\subseteq f(S')$ whenever $S\subseteq S'\subseteq X$. By a well known theorem of Knaster and Tarski [@T-Knaster55], actually formulated for complete lattices, every monotonic $f:\wp(X)\to\wp(X)$ has *least and greatest fixed points* — there is a unique $\subseteq$-minimal subset $L\subseteq X$ such that $f(L)=L$, and a unique $\subseteq$-maximal $G\subseteq X$ such that $f(G)=G$. We write $L=LFP(f)$ and $G=GFP(f)$.
There are a couple of useful ways to ‘compute’ these fixed points. First, define by recursion a subset $S_\alpha\subseteq X$ for each ordinal $\alpha$, by $S_0=\emptyset$, $S_{\alpha+1}=f(S_\alpha)$, and $S_\delta=\bigcup_{\alpha<\delta}S_\alpha$ for limit ordinals $\delta$. The $S_\alpha$ form an increasing chain terminating in $LFP(f)$, so $$LFP(f)=\bigcup_{\alpha\in\rm On}S_\alpha.$$ A similar expression can be given for $GFP(f)$. Second, a subset $S\subseteq X$ is said to be a *pre-fixed point* of $f$ if $f(S)\subseteq S$, and a *post-fixed point* if $f(S)\supseteq S$. In [@T-Knaster55] it is proved that $LFP(f)$ is the intersection of all pre-fixed points of $f$, and dually for $GFP(f)$: $$\begin{array}{rcl}
LFP(f)&=&\bigcap\{S\subseteq X:f(S)\subseteq S\},
\\
GFP(f)&=&\bigcup\{S\subseteq X:f(S)\supseteq S\}.
\end{array}$$
For $f:\wp(X)\to\wp(X)$, define $f':\wp(X)\to\wp(X)$ by $f'(S)=X\setminus f(X\setminus S)$. It is an exercise to check that $f$ is monotonic iff $f'$ is, and in that case, $GFP(f)=X\setminus LFP(f')$.
Least fixed points are used in the semantics of the mu-calculus, coming up next.
Languages
---------
We assume some familiarity with modal languages and the mu-calculus. We fix a set $\Var$ of *propositional variables,* or *atoms.* Sometimes we may make assumptions on $\Var$ — for example, that it is finite. We will be considering various logical languages. The biggest of them is denoted by $\Lbig$, which is a set of formulas defined as follows:
1. each $p\in\Var$ is a formula (of $\Lbig$),
2. $\top$ is a formula,
3. if $\varphi,\psi$ are formulas then so are $\neg\varphi$, $(\varphi\wedge\psi)$, $\bo\varphi$, $\bod\varphi$, and $\forall\varphi$,
4. if $\Delta$ is a non-empty finite set of formulas then $\dit\Delta$ and $\didt\Delta$ are formulas,
5. if $q\in \Var$ and $\varphi$ is a formula that is *positive in $q$* (that is, every free occurrence of $q$ as an atomic subformula of $\varphi$ is in the scope of an even number of negations in $\varphi$; *free* means ‘not in the scope of any $\mu q$ in $\psi$’), then $\mu q\varphi$ is a formula, in which all occurrences of $q$ are *bound*. Bound atoms arise only in this way.
For formulas $\varphi,\psi$, and $q\in\Var$, the expression $\varphi(\psi/q)$ denotes the result of replacing every free occurrence of $q$ in $\varphi$ by $\psi$, where the result is well-formed — that is, all of its subformulas of the form $\mu p\theta$ are such that $\theta$ is positive in $p$. For example, if $\varphi=\mu p\, q$ then $\varphi(\neg p/q)=\mu p\,\neg p$ is not well-formed.
We use standard abbreviations: $\bot$ denotes $\neg\top$, $(\varphi\vee\psi)$ denotes $\neg(\neg\varphi\wedge\neg\psi)$, $(\varphi\to\psi)$ denotes $\neg(\varphi\wedge\neg\psi)$, $(\varphi\leftrightarrow\psi)$ denotes $(\varphi\to\psi)\wedge(\psi\to\varphi)$, $\di\varphi$ denotes $\neg\bo\neg\varphi$, $\did\varphi$ denotes $\neg\bod\neg\varphi$, $\exists\varphi$ denotes $\neg\forall\neg\varphi$, and if $\varphi$ is positive in $q$ then $\nu q\varphi$ denotes $\neg\mu q\neg\varphi(\neg q/q)$ (this is well-formed). For a non-empty finite set $\Delta=\{\vec\delta n\}$ of formulas, we let $\bigwedge\Delta$ denote $\delta_1\wedge\ldots\wedge\delta_n$ and $\bigvee\Delta$ denote $\delta_1\vee\ldots\vee\delta_n$ (the order and bracketing of the conjuncts and disjuncts will always be immaterial). We set $\bigwedge\emptyset=\top$ and $\bigvee\emptyset=\bot$. Parentheses will be omitted where possible, by the usual methods.
The connectives $\dit,\didt$ are called *tangle connectives,* or (more fully) *tangled closure operators.*
We will be using various *sublanguages* of $\Lbig$, and they will be denoted in the obvious way by omitting prohibited operators from the notation. So for example, $\c L^{\mu}_{\bo\forall}$ denotes the language consisting of all $\Lbig$-formulas that do not involve $\bod,$ $\dit$, or $\didt$.
Kripke semantics {#ss:Kripke sem}
----------------
An *assignment* or *valuation* into a frame $\c F=(W,R)$ is a map $h:\Var\to\wp(W)$. A *Kripke model* is a triple $\c M=(W,R,h)$, where $(W,R)$ is a frame and $h$ an assignment into it. The *frame of $\c M$* is $(W,R)$, and we say that $\c M$ is finite, reflexive, transitive, etc., if its frame is.
For every Kripke model $\c M=(W,R,h)$ and every world $w\in W$, we define the notion $\c M,w\models\varphi$ of a formula $\varphi$ of $\Lbig$ being *true at $w$ in $\c M$.* The definition is by induction on $\varphi$, as follows:
1. $\c M,w\models p$ iff $w\in h(p)$, for $p\in\Var$.
2. $\c M,w\models\top$.
3. $\c M,w\models\neg\varphi$ iff $\c M,w\not\models\varphi$.
4. $\c M,w\models\varphi\wedge\psi$ iff $\c M,w\models\varphi$ and $\c M,w\models\psi$.
5. $\c M,w\models\bo\varphi$ iff $\c M,v\models\varphi$ for every $v\in R(w)$.
6. The truth condition for $\bod\varphi$ is exactly the same as for $\bo\varphi$.
7. $\c M,w\models\forall\varphi$ iff $\c M,v\models\varphi$ for every $v\in W$.
8. \[item: sem tangle kripke\] $\c M,w\models\dit\Delta$ iff there are worlds $w=w_0,w_1,\ldots\in W$ with $R(w_n,w_{n+1})$ for each $n<\omega$ and such that for each $\delta\in\Delta$ there are infinitely many $n<\omega$ with $\c M,w_n\models\delta$.
9. The truth condition for $\didt\Delta$ is exactly the same as for $\dit\Delta$.
10. The truth condition for $\mu q\varphi$ takes longer to explain. For an assignment $h:\Var\to\wp(W)$ and $S\subseteq W$, define a new assignment $h[S/q]:\Var\to\wp(W)$ by $$h[S/q](p)=
\begin{cases}
S,&\mbox{if }p=q,
\\
h(p),&\mbox{otherwise,}
\end{cases}$$ for $p\in\Var$. Inductively, the set $\sem\varphi_h=\{w\in W:(W,R,h),w\models\varphi\}$ is well defined, for every assignment $h$ into $(W,R)$. Define a map $f:\wp(W)\to\wp(W)$ by $$f(S)=\sem\varphi_{h[S/q]}\quad\mbox{for }S\subseteq W.$$ Since $\varphi$ is positive in $q$, it can be shown that $f$ is monotonic, so it has a least fixed point, $LFP(f)$ (see section \[ss:lfp\]). We define $\c M,w\models\mu q\varphi$ iff $w\in LFP(f)$.
In the notation of the last clause, it can be checked that $\c M,w\models\nu q\varphi$ iff $w\in GFP(f)$.
A word on the semantics of $\did$ and $\didt$. Let us temporarily write $\varphi\equiv\psi$ to mean that $\c M,w\models\varphi\leftrightarrow\psi$ for every *transitive* Kripke model $\c M=(W,R,h)$ and every $w\in W$. Then it can be checked that for every non-empty finite set $\Delta$ of formulas, $$\label{e:mu and tangle}
\begin{array}{rcl}
\dit\Delta&\equiv&\displaystyle\nu q\bigwedge_{\delta\in\Delta}\di(\delta\wedge q),
\\[16pt]
\didt\Delta&\equiv&\displaystyle\nu q\bigwedge_{\delta\in\Delta}\did(\delta\wedge q),
\end{array}$$ if $q\in\Var$ is a ‘new’ atom that does not occur in any formula in $\Delta$. For more details, see lemma \[lem:mu trans\]. In a sense, is the ‘official’ definition of the semantics of the tangle connectives, which boils down to clause \[item: sem tangle kripke\] above in the case of transitive Kripke models.
Kripke semantics in generated submodels
---------------------------------------
Let $\c M=(W,R,h)$ be a Kripke model. A *generated submodel* of $\c M$ is a model of the form $\c M'=(W',R',h')$, where $(W',R')$ is a generated subframe of $(W,R)$ and $h':\Var\to\wp(W')$ is given by $h'(p)=h(p)\cap W'$ for $p\in\Var$. The following is an easy extension to $\c L^{\mu\did\didt}_{\bo\bod}$ of a well known result in modal logic:
\[lem:gen submodels\] Let $\c M'=(W',R',h')$ be a generated submodel of $\c M=(W,R,h)$. Then for each $\varphi\in\c L^{\mu\did\didt}_{\bo\bod}$ and $w\in W'$, we have $$\c M,w\models\varphi\iff\c M',w\models\varphi.$$
There is no distinction between $\bo$ and $\bod$ or between $\dit$ and $\didt$ in Kripke semantics. This is not so in topological semantics, to be studied in part 2.
Hilbert systems
---------------
These are familiar, and we will be informal. A *Hilbert system $H$* in a given language $\c L\subseteq\Lbig$ is a set of *axioms,* which are $\c L$-formulas, and *inference rules,* which have the form $$\label{e:inf rule}
\infrule{\vec\varphi n}\psi,$$ for $\c L$-formulas $\vec\varphi n,\psi$. A *derivation in $H$ (of length $l$)* is a sequence $\vec\varphi l$ of $\c L$-formulas such that each $\varphi_i$ $(1\leq i\leq l$) is either an $H$-axiom or is derived from earlier $\varphi_j$ by an $H$-rule — that is, there are $1\leq\vec jn<i$ such that $$\infrule{\varphi_{j_1},\ldots,\varphi_{j_n}}{\varphi_i}$$ is an instance of a rule of $H$.
A *theorem of $H$* is a formula that occurs in some derivation in $H$. An *$H$-logic* is a set of $\c L$-formulas that contains all $H$-axioms and is closed under all $H$-rules. The set of theorems of $H$ is the smallest $H$-logic. Sometimes we identify (notationally) $H$ with this set, or present $H$ implicitly by defining an $H$-logic.
A formula $\varphi$ is *consistent* with $H$ if $\neg\varphi$ is not a theorem of $H$. A set $\Gamma$ of formulas is *consistent* with $H$ if $\bigwedge\Gamma_0$ is consistent with $H$, for every finite $\Gamma_0\subseteq\Gamma$.
Some familiar Hilbert systems used later are:
$\axK$:
: the axioms comprise (i) all instances of propositional tautologies (e.g., $\varphi\to(\psi\to\varphi)$, etc.) and (ii) all formulas of the form $\bo(\varphi\to \psi)\to(\bo \varphi\to\bo \psi)$ (the so-called ‘normality’ schema). The inference rules are:
- modus ponens: $\infrule{\varphi,\;\varphi\to\psi}\psi$
- $\bo$-generalisation: $\infrule \varphi{\bo\varphi}$
$\axK4$:
: this is $\axK$ plus all instances of the ‘4’ schema: $\bo \varphi\to\bo\bo \varphi$.
$\axS4$:
: this is $\axK$ plus all instances of the S4 schemata: $\bo \varphi\to \varphi$ and $\bo \varphi\to\bo\bo \varphi$.
The well known substitution rule $\infrule\varphi{\varphi(\psi/q)}$ is not always sound in the mu-calculus and is not needed in other systems, so we omit it.
As usual, we denote particular Hilbert systems by sequences of letters and numbers indicating the axioms present. For example, $\axS4.\axU\axC$ denotes the extension of $\axS4$ by the axioms generated by two schemes $\axU$ and $\axC$ to be seen later. The letter $t$ will denote the schemata for the tangle operator given in section \[ss:tangle logics\].
Satisfiability, validity, equivalence {#ss:validity}
-------------------------------------
Let $\c F=(W,R)$ be a Kripke frame. A set $\Gamma$ of $\Lbig$-formulas is said to be *satisfiable in $\c F$* if there exist an assignment $h$ into $\c F$ and a world $w\in W$ such that $(W,R,h),w\models\gamma$ for every $\gamma\in\Gamma$.
Let $\varphi$ be an $\Lbig$-formula. We say that $\varphi$ is *satisfiable in $\c F$* if the set $\{\varphi\}$ is so satisfiable. We say that $\varphi$ is *valid in $\c F$* if $\neg\varphi$ is not satisfiable in $\c F$. We may also say in this case that $\c F$ *validates* $\varphi$.
We also say that $\varphi$ is *equivalent* to a formula $\psi$ in $\c F$ if $\varphi\leftrightarrow\psi$ is valid in $\c F$.
Logics
------
Let $\c K$ be a class of Kripke frames. In the context of a given language $\c L\subseteq\Lbig$, the *($\c L$)-logic of $\c K$* is the set of all $\c L$-formulas that are valid in every member of $\c K$. A Hilbert system $H$ for $\c L$ whose set of theorems is $T$, say, is said to be
- *sound over $\c K$* if $T$ is a subset of the logic of $\c K$ (all $H$-theorems are valid in $\c K$),
- *weakly complete*, or simply *complete, over $\c K$* if $T$ contains the logic of $\c K$ (all $\c K$-valid formulas are $H$-theorems),
- *strongly complete over $\c K$* if every countable $H$-consistent set $\Gamma$ of $\c L$-formulas is satisfiable in some in $\c K$. (The restriction to countable sets will be discussed at the beginning of subsection \[subsec:strongcomp\].)
The logic of a single frame $\c F$ is defined to be the logic of the class $\{\c F\}$; similar definitions are used for the other terms here.
We say that a Kripke frame $\c F$ is an *$H$-frame,* or that $\c F$ *validates $H$,* if $H$ is sound over $\c F$. To establish this, it is enough to check that each axiom of $H$ is valid in $\c F$, and that each rule of $H$ preserves $\c F$-validity (in the notation in above, this means that if $\vec\varphi n$ are valid in $\c F$ then so is $\psi$). We assume familiarity with basic results about modal validity: for example, that a frame is a $\axK4$-frame iff it is transitive, and an $\axS4$-frame iff it is reflexive and transitive.
It can be checked that $H$ is weakly complete over $\c K$ iff every *finite* $H$-consistent set of formulas is satisfiable in some in $\c K$. Hence, every strongly complete Hilbert system is also weakly complete.
A system $H$ is said to have the *finite model property over $\c K$* if each $H$-consistent formula is satisfiable in some *finite* member of $\c K$. Equivalently, this means that $H$ is weakly complete over the class of finite members of $\c K$ (i.e. any formula valid in all finite members of $\c K$ is an $H$-theorem).
In this part of the paper, we look briefly at Hilbert systems for the mu-calculus, but mainly we establish the finite model property for the logics of concern in the paper.
Hilbert systems for mu-calculus {#ss:hs mu}
===============================
We now present a very brief diversion on a Hilbert system for the mu-calculus that is sound and complete over the class of finite reflexive transitive Kripke frames. It will be used to translate $\mu$ to $\dit$ and to axiomatise the $\c L^\mu_\bo$-logic of spaces. In this section, all formulas are $\c L^\mu_\bo$-formulas, all Hilbert systems are for this language, and we assume that $\Var$ is infinite.
\[def:mu systems\] Consider the two Hilbert systems:
$\axK\mu$:
: standard modal logic $\axK$ with the axioms comprising all instances of propositional tautologies and of normality ($\bo(\varphi\to \psi)\to(\bo \varphi\to\bo \psi)$), and the inference rules modus ponens, $\bo$-generalisation, plus the following for each formula $\varphi$ positive in $q$:
- fixed point axiom: $\varphi(\mu q\varphi/q)\to\mu q\varphi$, provided that no free occurrence of an atom in $\mu q\varphi$ gets bound in $\varphi(\mu q\varphi/q)$ — consequently, $\varphi(\mu q\varphi/q)$ is well formed (the idea is roughly that $\mu q\varphi$ is a pre-fixed point of $\varphi$)
- fixed point rule: $\infrule{\varphi(\psi/q)\to\psi}{\mu q\varphi\to\psi}$, provided that no free occurrence of an atom in $\psi$ gets bound in $\varphi(\psi/q)$ — hence, $\varphi(\psi/q)$ is well formed (the idea this time is roughly that $\mu q\varphi$ is the least pre-fixed point of $\varphi$).
We write $\axK\mu\vdash\varphi$ if $\varphi$ is a theorem of this system. It is well known (see, e.g., [@BS-mucalc07 §6]) that the system is equivalent to the original equational system of Kozen [@Koz:mu82].
$\axS4\mu$:
: this is $\axK\mu$ plus the S4 schemata $\bo \varphi\to\varphi$, $\bo\varphi\to\bo\bo\varphi$. We write $\axS4\mu\vdash\varphi$ if $\varphi$ is a theorem of this system.
The following combines some famous and difficult work in the mu-calculus.
\[fact:major mu\] $\axK\mu$ is sound and complete over the class of all finite Kripke frames.
We are going to extend it to show that $\axS4\mu$ is sound and complete over the class of finite reflexive transitive frames (and, in Part 2, over every space). First, a form of the substitution rule can be established.
\[lem:limited sub\] Suppose $\varphi,\psi$ are formulas such that for each atom $s$ occurring free in $\psi$, there is no subformula of $\varphi$ of the form $\mu s\theta$. If $\axS4\mu\vdash\varphi$, then $\axS4\mu\vdash\varphi(\psi/p)$ for any atom $p$.
Let $\varphi,\psi,p$ be as stipulated. For a formula $\alpha$, write $\alpha^\dag=\alpha(\psi/p)$. We show that $\axS4\mu\vdash\alpha\Rightarrow \axS4\mu\vdash\alpha^\dag$ (when the stipulation holds) by induction on the length of a derivation of $\varphi$ in $\axS4\mu$.
Suppose that $\varphi$ is an instance $\alpha(\mu q\alpha/q)\to\mu q\alpha$ of the fixed point axiom. Then $\varphi^\dag$ is valid in all Kripke frames, so by fact \[fact:major mu\], $\axK\mu\vdash\varphi^\dag$ and hence certainly $\axS4\mu\vdash\varphi^\dag$.
Suppose that $\varphi$ is derived by the fixed point rule, so that $\varphi=\mu q\alpha\to\beta$ for some $\alpha,\beta,q$ meeting the condition of the rule, and $\alpha(\beta/q)\to\beta$ occurs earlier in the derivation. If $s$ occurs free in $\psi$ then there is no $\mu s$ in $\mu q\alpha\to\beta$, so none in $\alpha(\beta/q)\to\beta$ either. So the inductive hypothesis applies, to give $\axS4\mu\vdash(\alpha(\beta/q)\to\beta)^\dag$. Let us evaluate this. If $p=q$, it is $\axS4\mu\vdash\alpha(\beta^\dag/q)\to\beta^\dag$. By our stipulation, the fixed point rule applies, giving $\axS4\mu\vdash\mu q\alpha\to\beta^\dag$. But $(\mu q\alpha)^\dag=\mu q\alpha$. So $\axS4\mu\vdash\varphi^\dag$ as required. If instead $p\neq q$, then it is $\axS4\mu\vdash\alpha^\dag(\beta^\dag/q)\to\beta^\dag$. Again, the rule applies, to give $\axS4\mu\vdash\mu q\alpha^\dag\to\beta^\dag$. But this is exactly $\axS4\mu\vdash\varphi^\dag$.
All other cases of the induction are easy and left to the reader.
\[def:phi\*\] For a formula $\varphi$, define a new formula $\varphi^*$ by induction:
- $p^*=p$ for $p\in\Var$;
- $-^*$ *commutes* with the boolean connectives and $\mu$. That is, $\top^*=\top$, $(\neg\varphi)^*=\neg\varphi^*$, $(\varphi\wedge\psi)^*=\varphi^*\wedge\psi^*$, and $(\mu q\varphi)^*=\mu q\varphi^*$.
- $(\bo\varphi)^*=\nu q(\varphi^*\wedge\bo q)$, where $q\in\Var$ is a ‘new’ atom not occurring in $\varphi^*$.
The formula $\varphi^*$ is plainly well formed, for all $\varphi\in\c L^\mu_\bo$.
\[lem:semantic \*\] Let $\varphi$ be any formula. Then for every Kripke model $(W,R,h)$ and $w\in W$, we have $(W,R,h),w\models\varphi^*$ iff $(W,R^*,h),w\models\varphi$, where (recall) $R^*$ is the reflexive transitive closure of $R$.
The proof is by induction on $\varphi$. The atomic and boolean cases are easy. Assuming the result for $\varphi$, it is a well known exercise in the mu-calculus to check that $(W,R,h),w\models(\bo\varphi)^*$ iff $(W,R,h),u\models\varphi^*$ for every $u\in R^*(w)$. Inductively, this is iff $(W,R^*,h),u\models\varphi$ for every $u\in R^*(w)$, iff $(W,R^*,h),w\models\bo\varphi$ as required.
Finally assume that the result holds for $\varphi$, positive in $q$, for every Kripke model. For a formula $\psi$ and Kripke model $(W,R,h)$, write $\sem\psi_{(W,R,h)}=\{w\in W:(W,R,h),w\models\psi\}$. Then $(W,R,h),w\models(\mu q\varphi)^*$ iff $(W,R,h),w\models\mu q\varphi^*$, iff $w$ is in the least fixed point of the map $f:\wp(W)\to\wp(W)$ given by $f(S)=\sem{\varphi^*}_{(W,R,h[S/q])}$. But inductively, $f(S)=\sem{\varphi}_{(W,R^*,h[S/q])}$. So this is iff $(W,R^*,h),w\models\mu q\varphi$ as required.
\[lem:\*\] $\axS4\mu\vdash\varphi\leftrightarrow\varphi^*$ for every $\varphi$.
Again, the proof is by induction on $\varphi$. We write just ‘$\vdash$’ for ‘$\axS4\mu\vdash$’ in the proof. We also write $\alpha\equiv\beta$ for $\vdash\alpha\leftrightarrow\beta$. First, replace all bound atoms in $\varphi$ by fresh ones, to give a formula $\overline\varphi$. More formally, $\overline\psi$ is defined for each subformula $\psi$ of $\varphi$ by induction: $\overline{\mu q\psi}=\mu s(\overline\psi(s/q))$, where $s$ is a new atom associated with $\psi$ and not occurring in $\varphi$, and $\overline{\,\cdot\,}$ commutes with all other operators. By fact \[fact:major mu\], $\overline\varphi\equiv\varphi$ and $(\overline\varphi)^*\equiv\varphi^*$. So, replacing $\varphi$ by $\overline\varphi$, we can suppose without loss of generality that for each atom $q$ that occurs free in $\varphi$, there is no subformula of $\varphi$ of the form $\mu q\theta$. The $-^*$ operator preserves this condition, so it holds for $\varphi^*$ as well.
For atomic $\varphi$, the result is trivial since $\varphi^*=\varphi$, and booleans are fine.
Assume inductively that $\varphi\equiv\varphi^*$ and consider $\bo\varphi$. We need to show that $\bo\varphi\equiv\nu q(\varphi^*\wedge\bo q)$, for ‘new’ $q$ — that is, $\bo\varphi\equiv\neg\mu q\neg(\varphi^*\wedge\bo \neg q)$. By a tautology, it is enough to show $\neg\bo\varphi\equiv\mu q\neg(\varphi^*\wedge\bo \neg q)$. By fact \[fact:major mu\], $\neg\bo\varphi\equiv\di\neg\varphi$ and $\mu q\neg(\varphi^*\wedge\bo \neg q)\equiv\mu q(\neg\varphi^*\vee\di q)$. So, letting $\psi=\neg\varphi$, it is enough to prove $$\label{e:mu to prove}
\di\psi\equiv\mu q\chi,\mbox{ where }\chi=\psi^*\vee\di q.$$ Note that the inductive hypothesis gives $\psi\equiv\psi^*$, and that $\chi(\theta/q)$ is well-formed for any well-formed $\theta$. Let $\chi^0=\bot$, and $\chi^{n+1}=\chi(\chi^n/q)$ for $n<\omega$. The following claim, needed only for $n=2$, is an instance of a more general result.
$\vdash \chi^n\to\mu q\chi$ for each $n<\omega$.
By induction on $n$. For $n=0$, it is $\vdash\bot\to\mu q\chi$, a tautology. Assume inductively that $\vdash \chi^n\to\mu q\chi$. We desire $\vdash \psi^*\vee\di \chi^n\to\mu q\chi$. By the fixed point axiom, it is enough to prove that $\vdash \psi^*\vee\di \chi^n\to\chi(\mu q\chi/q)$ — that is, $\vdash \psi^*\vee\di \chi^n\to\psi^*\vee\di\mu q\chi$. But the inductive hypothesis plus standard uses of generalisation and normality yield $\vdash \di\chi^n\to\di\mu q\chi$, and the result follows using tautologies and modus ponens. This proves the claim.
Towards , we first show that $\vdash\di\psi\to\mu q\chi$. Observe that inductively, $\chi^1=\psi^*\vee\di\bot\equiv\psi$ and $\chi^2=\psi^*\vee\di\chi^1\equiv \psi\vee\di\psi$. By the claim for $n=2$, and tautologies, $\vdash \psi\vee\di\psi\to\mu q\chi$ and applying more tautologies yields $\vdash\di\psi\to\mu q\chi$.
Now we show $\vdash\mu q\chi\to\di\psi$. By the fixed point rule, it is enough to show $\vdash\chi(\di\psi/q)\to\di\psi$. That is, $\vdash\psi^*\vee\di\di\psi\to\di\psi$. But given the inductive hypothesis, this is just what the S4 axioms say. This proves and completes the case of $\bo\varphi$.
Finally assume the result for $\varphi$ positive in $q$, and consider the case $\mu q\varphi$. All formulas below meet all necessary conditions because of our initial assumption on $\varphi$. By the inductive hypothesis and lemma \[lem:limited sub\] we get $\vdash\varphi(\mu q\varphi^*/q)\to\varphi^*(\mu q\varphi^*/q)$. The fixed point axiom gives $\vdash\varphi^*(\mu q\varphi^*/q)\to\mu q\varphi^*$. Putting the two together gives $\vdash\varphi(\mu q\varphi^*/q)\to\mu q\varphi^*$. This says that $\mu q\varphi^*$ is a pre-fixed point of $\varphi$, so the fixed point rule gives $\vdash\mu q\varphi\to\mu q\varphi^*$. The converse, $\vdash\mu q\varphi^*\to\mu q\varphi$, is similar.
\[thm:S4mu sc\] The system $\axS4\mu$ is sound and complete over the class of finite reflexive transitive Kripke frames (finite S4 frames).
Soundness is easily checked. Conversely, assume that $\varphi$ is consistent with $\axS4\mu$. By lemma \[lem:\*\], $\varphi^*$ is consistent with $\axS4\mu$ and hence with $\axK\mu$ as well. By fact \[fact:major mu\], there is a finite Kripke model $\c M=(W,R,h)$ in which $\varphi^*$ is satisfied at $w$, say. We do not know that $(W,R)$ is reflexive or transitive. However, by lemma \[lem:semantic \*\] we have $(W,R^*,h),w\models\varphi$ as well, and $R^*$ is reflexive and transitive.
Finite model property {#sec:fmp}
=====================
In the second part of the paper, we prove topological completeness theorems for the logics discussed in part 1. The results in part 1 will of course be used, but much of part 2 can be read independently — indeed, nearly all of it, if the reader takes the results of part 1 on trust.
Further basic definitions
=========================
In this section, the main definitions, notation, and basic results needed in Part 2 are developed.
Topological spaces {#ss:top spaces}
------------------
We will assume some familiarity with topology, but we take some time to reprise the main concepts and notation. A *topological space* is a pair $(X,\tau)$, where $X$ is a set and $\tau\subseteq\wp(X)$ satisfies:
1. if $\c S\subseteq\tau$ then $\bigcup\c S\in\tau$,
2. if $\c S\subseteq\tau$ is finite then $\bigcap\c S\in\tau$, on the understanding that $\bigcap\emptyset=X$.
So $\tau$ is a set of subsets of $X$ closed under unions and finite intersections. By taking $\c S=\emptyset$, it follows that $\emptyset,X\in\tau$. The elements of $\tau$ are called *open subsets* of $X$, or just *open sets.* An *open* of a point $x\in X$ is an open set containing $x$. A subset $C\subseteq X$ is called *closed* if $X\setminus C$ is open. The set of closed subsets of $X$ is closed under intersections and finite unions. If $O$ is open and $C$ closed then $O\setminus C$ is open and $C\setminus O$ is closed.
We use the signs $\int$, $\cl$, $\did$ to denote the *interior,* *closure,* and *derivative* operators, respectively. So for $S\subseteq X$,
- $\int S=\bigcup\{O\in\tau:O\subseteq S\}$ — the largest open set contained in $S$,
- $\cl S=\bigcap\{C\subseteq X:C$ closed, $S\subseteq C\}$ — the smallest closed set containing $S$; we have $\cl S=\{x\in X:S\cap O\neq\emptyset
\mbox{ for every open \nhd\ } O\mbox{ of } x\}$,
- $\did S=\{x\in X:S\cap O\setminus\{x\}\neq\emptyset
\mbox{ for every open \nhd\ } O\mbox{ of } x\}$.
Then $\int S\subseteq S\subseteq\cl S\supseteq\did S$. For all subsets $A,B$ of $X$, we have $$\begin{array}{rcl}
\cl(A\cup B)&=&\cl A\cup\cl B,
\\
\did(A\cup B)&=&\did A\cup\did B,
\\
\int(A\cap B)&=&\int A\cap\int B.
\end{array}$$ That is, *closure and $\did$ are additive and interior is multiplicative.*
We follow standard practice and identify (notationally) the space $(X,\tau)$ with $X$. The reader should note that we do allow empty topological spaces, where $X=\emptyset$. This is particularly useful when dealing with subspaces.
A *subspace* of $X$ is a topological space of the form $(Y,\{O\cap Y:O\in\tau\})$, for (possibly empty) $Y\subseteq X$. It is a subset of $X$, made into a topological space by endowing it with what is called the *subspace topology.* It is said to be an *open subspace* if $Y$ is an open subset of $X$. As with $X$, we identify (notationally) the subspace with its underlying set, $Y$. We write $\int_Y,\cl_Y$ for the operations of interior and closure in the subspace $Y$. It can be checked that for every $S\subseteq Y$ we have $\cl_YS=Y\cap\cl S$, and if $Y$ is an open subspace then $\int_YS=\int S$.
We will be considering various properties that a topological space $X$ may have. We leave most of them for later, but we mention now that $X$ is said to be *dense in itself* if no singleton subset is open, *connected* if it is not the union of two disjoint non-empty open sets, and *separable* if it has a countable subset $D$ with $X=\cl D$. $X$ is *T1* if every singleton subset $\{x\}$ is closed, and *T$_D$* if the derivative $\did\{x\}$ of every singleton is closed, which is equivalent to requiring $\did\did\{x\}\sub\did\{x\}$. The T$_D$ property, introduced in [@aull:sepa62], is strictly weaker than T1.
Metric spaces {#ss:metric spaces}
-------------
A *metric space* is a pair $(X,d)$, where $X$ is a set and $d:X\times X\to\R$ is a ‘distance function’ (having nothing to do with the modal operator $\did$) satisfying, for all $x,y,z\in X$,
1. $d(x,y)\geq0$,
2. $d(x,y)=0$ iff $x=y$,
3. $d(x,y)=d(y,x)$,
4. $d(x,z)\leq d(x,y)+d(y,z)$ (the ‘triangle inequality’).
We assume some experience of working with this definition, in particular with the triangle inequality. Examples of metric spaces abound and include the real numbers $\R$ with the standard distance function $d(x,y)=|x-y|$, $\R^n$ with Pythagorean distance, etc. As usual, we often identify (notationally) $(X,d)$ with $X$.
Let $(X,d)$ be a metric space, and $x\in X$. For non-empty $S\subseteq X$, define $$d(x,S)=\inf\{d(x,y):y\in S\}.$$ We leave $d(x,\emptyset)$ undefined. For a real number $\varepsilon>0$, we let $N_\varepsilon(x)$ denote the so-called ‘open ball’ $\{y\in X:d(x,y)<\varepsilon\}$. A metric space $(X,d)$ gives rise to a topological space $(X,\tau_d)$ in which a subset $O\subseteq X$ is declared to be open (i.e., in $\tau_d$) iff for every $x\in O$, there is some $\varepsilon>0$ such that $N_\varepsilon(x)\subseteq O$. In other words, the open sets are the unions of open balls. We frequently regard a metric space $(X,d)$ equally as a topological space $(X,\tau_d)$. So, we will say that a metric space has a given topological property (such as being dense in itself) if the associated topological space has the property. As an example, every metric space is T$_D$, since it has the stronger Hausdorff (or T2) property.
A *subspace* of a metric space $(X,d)$ is a pair of the form $(Y,d\restriction Y\times Y)$, where $Y\subseteq X$. It is plainly a metric space, and the topological space $(Y,\tau_{d\restriction Y\times Y})$ is a subspace of $(X,\tau_d$).
Topological semantics
---------------------
Given a topological space $X$, an *assignment* into $X$ is simply a map $h:\Var\to\wp(X)$. A *topological model* is a pair $(X,h)$, where $X$ is a topological space and $h$ an assignment into $X$. We will also be considering topological models where $\Var$ is replaced by some other set of atoms. Details will be given later.
As with Kripke models, we attribute a topological property to a topological model if the underlying topological space has the property.
For every topological model $(X,h)$ and every point $x\in X$, we define $(X,h),x\models\varphi$, for a $\Lbig$-formula $\varphi$, by induction on $\varphi$:
1. $(X,h),x\models p$ iff $x\in h(p)$, for $p\in\Var$.
2. $(X,h),x\models\top$.
3. $(X,h),x\models\neg\varphi$ iff $(X,h),x\not\models\varphi$.
4. $(X,h),x\models\varphi\wedge\psi$ iff $(X,h),x\models\varphi$ and $(X,h),x\models\psi$.
5. $(X,h),x\models\bo\varphi$ iff there is an open $O$ of $x$ with $(X,h),y\models\varphi$ for every $y\in O$.
6. $(X,h),x\models\bod\varphi$ iff there is an open $O$ of $x$ with $(X,h),y\models\varphi$ for every $y\in O\setminus\{x\}$. We do not require $\varphi$ to hold at $x$ itself.
7. $(X,h),x\models\forall\varphi$ iff $(X,h),y\models\varphi$ for every $y\in X$.
8. \[topsem clause 8\] For a non-empty finite set $\Delta$ of formulas for which we have inductively defined semantics, write $\sem\delta=\{x\in X:(X,h),x\models\delta\}$, for each $\delta\in\Delta$. Then define:
- $(X,h),x\models\dit\Delta$ iff there is some $S\subseteq X$ such that $x\in S\subseteq\bigcap_{\delta\in\Delta}\cl(\sem\delta\cap S)$,
- $(X,h),x\models\didt\Delta$ iff there is some $S\subseteq X$ such that $x\in S\subseteq\bigcap_{\delta\in\Delta}\did(\sem\delta\cap S)$.
9. Suppose inductively that $\sem\varphi_h=\{x\in X:(X,h),x\models\varphi\}$ is well defined, for every assignment $h$ into $X$. Define a map $f:\wp(X)\to\wp(X)$ by $$f(S)=\sem\varphi_{h[S/q]}\quad\mbox{for }S\subseteq X,$$ where $h[S/q]$ is defined as in Kripke semantics (section \[ss:Kripke sem\]). Again, $f$ is monotonic, and we define $(X,h),x\models\mu q\varphi$ iff $x\in LFP(f)$.
The definition makes sense but has no content if $X$ is empty: there are no points $x\in X$ to evaluate at. Writing $\sem\varphi_h=\{x\in X:(X,h),x\models\varphi\}$, we have $\sem{\bo\varphi}_h=\int(\sem\varphi_h)$, $\sem{\di\varphi}_h=\cl(\sem\varphi_h)$, and $\sem{\did\varphi}_h=\did(\sem\varphi_h)$ for each $\varphi,h$. Again, $\sem{\nu q\varphi}=GFP(f)$, where $\varphi,f$ are as in the last clause.
\[rmk:sem of tangle\]Again we briefly discuss the semantics of $\dit$ and $\didt$ (see clause \[topsem clause 8\] above). With $\varphi\equiv\psi$ redefined to mean that $(X,h),x\models\varphi\leftrightarrow\psi$ for every topological model $(X,h)$ and $x\in X$, the equivalences in above continue to hold, and indeed they motivate clause \[topsem clause 8\]. However, there is a perhaps more intuitive meaning for $\dit$ and $\didt$ in terms of *games,* which are used extensively in the mu-calculus. Let players , play a game of length $\omega$ on $X$. Initially, the position is $x$. In each round, if the current position is $y\in X$, player chooses an open $O$ of $y$ and a formula $\delta\in\Delta$. Player must select a point $z\in O$ at which $\delta$ is true (and with $z\neq y$ in the case of $\didt$). If she cannot, player wins. That is the end of the round, and the next round commences from position $z$. Player wins if she survives every round. It can be checked that $(X,h),x\models\dit\Delta$ (respectively, $(X,h),x\models\didt\Delta$) iff has a in this game (respectively, the game where she must additionally choose $z\neq y$).
Topological semantics in open subspaces
---------------------------------------
Let $X$ be a topological space and $Y$ a subspace of $X$. Each assignment $h:\Var\to\wp(X)$ into $X$ induces an assignment $h_Y$ into $Y$, via $h_Y(p)=Y\cap h(p)$, for each $p\in\Var$. Thus, we can evaluate formulas at points in $Y$ in both $(X,h)$ and $(Y,h_Y)$. Because the semantics of the connectives $\bo,\bod,\dit,\didt$ depend on only arbitrarily small open s of the evaluation point, it is easily seen that if $Y$ is an *open* subspace of $X$, we get the same result for every formula not involving $\forall$. That is, the following analogue of lemma \[lem:gen submodels\] holds:
\[lem:open subspaces\] Whenever $Y$ is an open subspace of $X$, we have $(X,h),y\models\varphi$ iff $(Y,h_Y),y\models\varphi$, for every $y\in Y$ and $\varphi\in\c L^{\mu\dit\didt}_{\bo\bod}$.
(This holds vacuously if $Y$ is empty.)
Satisfiability, validity, equivalence {#ss:validity 2}
-------------------------------------
Let $X$ be a topological space. A set $\Gamma$ of $\Lbig$-formulas is said to be *satisfiable in $X$* if there exist an assignment $h$ into $X$ and a point $x\in X$ such that $(X,h),x\models\gamma$ for every $\gamma\in\Gamma$.
Let $\varphi$ be an $\Lbig$-formula. We say that $\varphi$ is *satisfiable in $X$* if the set $\{\varphi\}$ is so satisfiable. We say that $\varphi$ is *valid in $X$,* or that *$X$ *validates* $\varphi$,* if $\neg\varphi$ is not satisfiable in $X$. We also say that $\varphi$ is *equivalent* to a formula $\psi$ in $X$ if $\varphi\leftrightarrow\psi$ is valid in $X$.
In any space $X$, the ‘4’ schema: $\bo \varphi\to\bo\bo \varphi$ is valid under the interpretation $\sem{\bo\varphi}=\int\sem\varphi$. But the schema $\bod \varphi\to\bod\bod \varphi$, or equivalently $\did\did \varphi\to \did \varphi$, is valid under the interpretation $\sem{\did\varphi}=\did\sem\varphi$ if, and only if, $X$ is a T$_D$ space. This is because in any space the derivatives of all subsets are closed iff the derivatives of all singletons are closed (see [@aull:sepa62 Theorem 5.1]).
Logics
------
Let $\c K$ be a class of topological spaces. In the context of a given language $\c L\subseteq\Lbig$, the *($\c L$)-logic of $\c K$* is the set of all $\c L$-formulas that are valid in every member of $\c K$. Exactly as for Kripke semantics, a Hilbert system $H$ for $\c L$ with set of theorems $T$ is said to be
- *sound over $\c K$* if $T$ is a subset of the logic of $\c K$ (all $H$-theorems are valid in $\c K$),
- *weakly complete*, or simply *complete, over $\c K$* if $T$ contains the logic of $\c K$ (all $\c K$-valid formulas are $H$-theorems),
- *strongly complete over $\c K$* if every countable $H$-consistent set $\Gamma$ of $\c L$-formulas is satisfiable in some in $\c K$.
For example the $\c L_\bod$-system K4 is sound and complete over the class of all T$_D$-spaces, a result due to Esakia (see [@Esa04:apal]).
The logic of a single space $X$ is defined to be the logic of the class $\{X\}$; similar definitions are used for the other terms here.
We say that a topological space $X$ *validates $H$* if $H$ is sound over $X$. To establish this, it is enough to check that each axiom of $H$ is valid in $X$, and that each rule of $H$ preserves $X$-validity.
It can be checked that $H$ is weakly complete over $\c K$ iff every *finite* $H$-consistent set of formulas is satisfiable in some space in $\c K$. Hence, every strongly complete Hilbert system is also weakly complete. The main aim of this part of the paper is to provide Hilbert systems that are (where possible) sound and strongly complete over various topological spaces, with respect to various sublanguages of $\Lbig$.
Translations {#sec:translations}
============
The language $\Lbig$ has some redundancy. We can express $\bo$ with $\bod$, and $\dit$ with $\didt$ (but not vice versa). We can also express $\dit,\didt$ with $\mu$ — and often vice versa, using results of Dawar and Otto [@DO09].
Later, we will need translations that work in both topological spaces and (possibly restricted) Kripke models. In this section, we will explore translations — but only to the extent needed for later work. We will again assume that $\Var$ is infinite.
Translating $\did$ and $\didt$ to $\mu$
---------------------------------------
This is the simplest case. We have already seen the idea, in the equivalence of $\dit$- and $\dit$-formulas to $\nu$-formulas given in in section \[ss:Kripke sem\].
For each $\Lbig$-formula $\varphi$, we define a $\c L^{\mu}_{\bo\bod\forall}$-formula $\varphi^\mu$ as follows:
1. $p^\mu=p$ for $p\in\Var$.
2. $-^\mu$ commutes with the boolean connectives, $\bo$, $\bod$, $\forall$, and $\mu$ (cf. definition \[def:phi\*\]).
3. $(\dit\Delta)^\mu=\nu q\bigwedge_{\delta\in\Delta}\di(\delta^\mu\wedge q)$, where $q\in\Var$ does not occur in any $\delta^\mu$ ($\delta\in\Delta$).
4. $(\didt\Delta)^\mu=\nu q\bigwedge_{\delta\in\Delta}\did(\delta^\mu\wedge q)$, where $q\in\Var$ does not occur in any $\delta^\mu$ ($\delta\in\Delta$).
These formulas can be checked to be well formed. The translation simply replaces $\dit$ by an expression using $\mu$ and $\bo$, and similarly for $\didt$. So if $\varphi\in\c L_{\bo}^{\dit}$ then $\varphi^\mu\in\c L_{\bo}^{\mu}$, if $\varphi\in\c L_{\bod}^{\didt}$ then $\varphi^\mu\in\c L_{\bod}^{\mu}$, etc.
This translation is faithful in all relevant semantics:
\[lem:mu trans\] Let $\varphi$ be any $\Lbig$-formula. Then $\varphi$ is equivalent to $\varphi^\mu$ in every transitive Kripke frame and in every topological space. (See sections \[ss:validity\] and \[ss:validity 2\] for the definition of equivalence.)
An easy induction on $\varphi$. We consider only the case $\dit\Delta$ (for finite $\Delta\neq\emptyset$), and only in Kripke semantics (the case $\didt\Delta$ is of course identical). Assume the lemma for each $\delta\in\Delta$. Take any transitive Kripke model $\c M=(W,R,h)$ and any $w\in W$. Inductively, $\c M,w\models(\dit\Delta)^\mu$ iff $\c M,w\models\nu q\bigwedge_{\delta\in\Delta}\di(\delta\wedge q)$. By the post-fixed point characterisation of greatest fixed points, this holds iff $(*)$ there is $S\subseteq W$ with $w\in S$ and such that for every $s\in S$ and $\delta\in\Delta$, there is $t\in S$ with $sRt$ and $\c M,t\models\delta$.
Assuming $(*)$, it is easy to choose a sequence $w=s_0Rs_1Rs_2\ldots$ in $S$ by induction so that $\{n<\omega:\c M,s_n\models\delta\}$ is infinite for every $\delta\in\Delta$. It follows that $\c M,w\models\dit\Delta$. Conversely, if $\c M,w\models\dit\Delta$ then there are worlds $w=w_0Rw_1Rw_2\ldots$ in $W$ with $\{n<\omega:\c M,w_n\models\delta\}$ infinite for every $\delta\in\Delta$. Let $S=\{w_n:n<\omega\}$. Then $w\in S$, and for each $w_n\in S$ and $\delta\in\Delta$, there is $m>n$ with $\c M,w_m\models\delta$. Then $w_m\in S$, and by transitivity of $R$ we have $w_nRw_m$. So $(*)$ holds.
Translating $\bo$ to $\bod$ and $\dit$ to $\didt$
-------------------------------------------------
Just replacing $\bo$ by $\bod$ and $\dit$ by $\didt$ in a formula $\varphi\in\Lbig$ yields an $\c L^{\mu\didt}_{\bod\forall}$-formula equivalent to $\varphi$ in all Kripke frames. But the two are not equivalent in topological spaces, so we seek a better translation that works in both semantics.
For each $\Lbig$-formula $\varphi$, we define a $\c L^{\mu\didt}_{\bod\forall}$-formula $\varphi^d$ as follows:
1. $p^d=p$ for $p\in\Var$.
2. $-^d$ commutes with the boolean connectives, $\bod$, $\didt$, $\forall$, and $\mu$.
3. $(\bo\varphi)^d=\varphi^d\wedge\bod\varphi^d$.
4. $(\dit\Delta)^d=(\bigwedge\Delta^d)\vee
\did(\bigwedge\Delta^d)\vee(\didt\Delta^d)$, where $\Delta^d=\{\delta^d:\delta\in\Delta\}$.
Again, $\varphi^d$ is always well formed. The translation $-^d$ is pretty good:
\[lem:-d in refl frames\] Each $\Lbig$-formula $\varphi$ is equivalent to $\varphi^d$ in every reflexive Kripke frame.
An easy induction on $\varphi$. To show, e.g., that $\bo\varphi$ implies $(\bo\varphi)^d$, we need reflexivity. We also note that $\bigwedge\Delta$ and $\did\bigwedge\Delta$ both imply $\dit\Delta$ in reflexive Kripke models.
\[lem:trans equivalent top\] Each $\Lbig$-formula $\varphi$ is equivalent to $\varphi^d$ in a topological space $X$ if, and only if, $X$ is T$_D$.
Let $X$ be a T$_D$ topological space. We prove by induction on $\varphi$ that each $\Lbig$-formula $\varphi$ is equivalent to $\varphi^d$ in $X$. We consider only two cases: $\bo\varphi$ and $\dit\Delta$. Inductively assume the result for $\varphi$ and each formula in the finite set $\Delta$ of formulas, let $h$ be an assignment into $X$, and let $x\in X$. In the proof, we write ‘$x\models{}$’ as short for ‘$(X,h),x\models$’, and for a formula $\varphi$, we write $\sem\varphi=\{y\in X:y\models\varphi\}$.
We prove that $x\models\bo\varphi\leftrightarrow(\bo\varphi)^d$. We have $x\models\bo\varphi$ iff for some open $O$ of $x$, we have $(X,h),y\models \varphi$ for every $y\in O$. This is plainly iff $x\models\varphi\wedge\bod\varphi$. Inductively, this is iff $x\models\varphi^d\wedge\bod\varphi^d$ — i.e., iff $x\models(\bo\varphi)^d$.
Now we prove that $x\models\dit\Delta\leftrightarrow(\dit\Delta)^d$. Recall that $$(\dit\Delta)^d=(\bigwedge\Delta^d)\vee
\did(\bigwedge\Delta^d)\vee(\didt\Delta^d).$$ First we prove that $x\models(\dit\Delta)^d\to\dit\Delta$. Suppose that $x\models(\dit\Delta)^d$. To show that $x\models\dit\Delta$, we need to find $S\subseteq X$ with $x\in S\subseteq\bigcap_{\delta\in\Delta}\cl(\sem\delta\cap S)$. If $x\models\bigwedge\Delta^d$, take $S=\{x\}$. If $x\models\did\bigwedge\Delta^d$, take $S=\{x\}\cup\sem{\bigwedge\Delta^d}$. And if $x\models\didt\Delta^d$, there is $S\subseteq X$ with $x\in S\subseteq\bigcap_{\delta\in\Delta}\did(\sem\delta\cap S)$; then $x\in S\subseteq\bigcap_{\delta\in\Delta}\cl(\sem\delta\cap S)$ as required.
It remains to prove that $x\models\dit\Delta\to(\dit\Delta)^d$. So suppose that $x\models\dit\Delta$. If $x\models(\bigwedge\Delta^d)\vee\did(\bigwedge\Delta^d)$, we are done.
So suppose not. Thus, there is an open $U$ of $x$ with $y\models\neg\bigwedge\Delta^d$ for every $y\in U$. So for every $y\in U$, there is $\delta_y\in\Delta$ with $y\models\neg\delta_y^d$.
We prove that $x\models\didt\Delta^d$.
Since $x\models\dit\Delta$, there is $S\subseteq X$ with $x\in S\subseteq\bigcap_{\delta\in\Delta}\cl(\sem\delta\cap S)$.
Put $S'=U\cap S$. Then $x\in S'\subseteq\bigcap_{\delta\in\Delta}\did(\sem{\delta^d}\cap S')$.
Plainly, $x\in S'$. For the other half, let $y\in S'$ and $\delta\in\Delta$ be arbitrary; we show that $y\in\did(\sem{\delta^d}\cap S')$. So let $O$ be any open of $y$. As $X$ is T$_D$, $\did\{y\}$ is closed, so since it does not contain $y$, $O\cap U\setminus \did\{y\}$ is an open of $y$ too. As $y\in S'\subseteq S\subseteq\cl(\sem{\delta_y}\cap S)$, there is some $z\in O\cap U\cap S\setminus \did\{y\}$ with $z\models\delta_y$. But $y\models\neg\delta_y^d$, so inductively, $y\models\neg\delta_y$. It follows that $z\neq y$.
Now we have $z\notin\{y\}\cup \did\{y\}=\cl\{y\}$, so $O\cap U\setminus\cl\{y\}$ is an open of $z$. Since $z\in S\subseteq\cl(\sem\delta\cap S)$, there is some $t\in O\cap U\cap S\setminus\cl\{y\}=O\cap S'\setminus\cl\{y\}$ with $t\models\delta$. Then $t\ne y$. Since $O$ was arbitrary, this shows that $y\in\did(\sem\delta\cap S')$. Since inductively, $\sem\delta=\sem{\delta^d}$, this proves the claim.
By definition of the semantics, the claim immediately yields $x\models\didt\Delta^d$ as required. This completes the induction and the proof that each $\varphi$ is equivalent to $\varphi^d$. (The reader may like to construct an alternative proof using the games described in remark \[rmk:sem of tangle\].)
Conversely, to show that the T$_D$ hypothesis is necessary, we first prove
> \[dxclosed\] In any space $X$, for any $x\in X$, $\cl\did\{x\}\setminus \did\{x\}\sub\{x\}$. Hence $\did\{x\}$ is closed iff $x\notin \cl\did\{x\}$.
>
> For the first part, since $\did\{x\}\sub \cl\{x\}$ and the latter is closed, $\cl\did\{x\}\sub \cl\{x\}=\did\{x\}\cup \{x\}$. This implies $\cl\did\{x\}\setminus \did\{x\}\sub\{x\}$.
>
> For the second part, $\did\{x\}$ is closed iff $\cl\did\{x\}\setminus \did\{x\}=\emptyset$. By the first part, this holds iff $x\notin\cl\did\{x\}\setminus \did\{x\}$. But $x\notin\did\{x\}$, so $x\notin\cl\did\{x\}\setminus \did\{x\}$ iff $x\notin \cl\did\{x\}$.
Now suppose the space $X$ is not T$_D$. Then there is some point $x$ of $X$ with $\did\{x\}$ not closed. By Lemma \[dxclosed\], $x\in \cl\did\{x\}$. Hence $\cl\{x\}\sub\cl\did\{x\}$. Let $p\in\Var$ and $h:\Var\to\wp X$ satisfy $h(p)=\{x\}$ for some (arbitrary) $x$. Then $(X,h),x\models\dit\{p,\did p\}$, but $(X,h),x\not\models(\dit\{p,\did p\})^d$, i.e. $(X,h),x\not\models (p\wedge \did p)\vee\did(p\wedge\did p)\vee\didt\{p,\did p\}$, giving a case of $\ph$ not being equivalent to $\ph^d$. That $x\not\models (p\wedge \did p)\vee\did(p\wedge\did p)$ follows because $\sem{p\wedge \did p}=\{x\}\cap\did\{x\}=\emptyset$. That $x\not\models\didt\{p,\did p\}$ follows as no punctured $O\setminus\{x\}$ contains a point of $\sem{p}=\{x\}$. To see that $x\models\dit\{p,\did p\}$, let $S=\cl\{x\}$. Then $S$ is included in both $\cl(\sem{p}\cap S)=\cl\{x\}=S$ and $\cl(\sem{\did p}\cap S)=\cl(\did\{x\})$ (because $\cl\{x\}\sub\cl\did\{x\}$ as noted above). Since $x\in S$, it follows that $x\models\dit\{p,\did p\}$ .
Translating $\mu$ to $\dit$ {#ss:DO09}
---------------------------
We use this translation only to prove strong completeness for $\c L_\bo^\mu$ in theorem \[thm:str compl boxes\](\[stro compl boxes part2\]). Fortunately, most of the hard work involved has already been done by others. We will need only the fact below, but its proof was a major enterprise.
\[fact:DO\] For each formula $\varphi$ of $\c L^\mu_{\bo}$, there is a formula $\varphi^t$ of $\c L^{\dit}_{\bo}$ that is equivalent to $\varphi$ in every finite transitive Kripke frame.
To lift this to topological spaces, we will use the proof theory from section \[ss:hs mu\].
\[cor:t transl top equiv\] Each $\c L_\bo^\mu$-formula $\varphi$ is equivalent to $\varphi^t$ in every topological space.
By fact \[fact:DO\] and lemma \[lem:mu trans\], $\varphi\leftrightarrow(\varphi^t)^\mu$ is an $\c L_\bo^\mu$-formula valid in every finite transitive Kripke frame. By theorem \[thm:S4mu sc\], $\axS4\mu\vdash\varphi\leftrightarrow(\varphi^t)^\mu$.
Now it is easy to check that $\axS4\mu$ is sound over every topological space. (The S4 axioms are sound by definition of the topological semantics of $\bo$, and the fixed point axiom and rule are sound by the semantics of $\mu$.) Hence, $\varphi\leftrightarrow(\varphi^t)^\mu$ is valid in every topological space. But by lemma \[lem:mu trans\], $(\varphi^t)^\mu$ is equivalent to $\varphi^t$ in every topological space. We conclude that $\varphi$ is equivalent to $\varphi^t$ in every topological space, as required.
By the corollary and lemma \[lem:mu trans\], $\c L_\bo^\mu$ and $\c L_\bo^\dit$ uniformly have the same expressive power in every topological space.
Since $\bo,\bod$ and $\dit,\didt$ are indistinguishable in Kripke semantics, a similar analysis would give a translation from $\c L_\bod^\mu$ to $\c L_\bod^\didt$ valid in every topological space. (For this purpose, the T axiom $\bo\varphi\to\varphi$ would be dropped in section \[ss:hs mu\], and the translation in definition \[def:phi\*\] adapted to represent transitive closure.) The translation would show that $\c L_\bod^\mu$ and $\c L_\bod^\didt$ are equally expressive over all T$_D$ topological spaces. We could use it to lift weak completeness for $\c L_\bod^\mu$ to strong completeness. Unfortunately, we do not have a weak completeness result for $\c L_\bod^\mu$ to lift.
More topology
=============
The finite model property theorems of Part 1 will be instrumental in our completeness theorems for (some) topological spaces. Not surprisingly, we will also need some simple and standard topological definitions and results, together with some more substantial ones. The first one is very simple.
\[lem:inf\] Let $X$ be a dense-in-itself T$_D$ topological space. Then every non-empty open subset of $X$ is infinite.
It suffices to show that every non-empty open subset $O$ has a non-empty open *proper* subset $O'$, since infinitely many iterations of that fact will produce an infinite sequence of distinct points in $O$. Take any $x\in O$. Then $x$ belongs to $ O\setminus\did\{x\}$, which is open as $\did\{x\}$ is closed in the T$_D$-space. Since $X$ is dense-in-itself, there must then be some $y\ne x$ with $y\in O\setminus\did\{x\}$. As $y\ne x$ and $y\notin \did\{x\}$, $y$ has an open $U$ with $x\notin U$. Put $O'=O\cap U$ to get that $O'$ is open, non-empty as it contains $y$, and a proper subset of $O$ as it does not contain $x$.
The $\did$ operator on sets
---------------------------
Let $X$ be a topological space. For a set $S\subseteq X$, recall that $\did S=
\{x\in X:S\cap O\setminus\{x\}\neq\emptyset$ for every open $O$ of $x\}$, the set of strict limit points of $S$. The $\did$ operator has the following basic properties.
\[lem:did cl\] Let $S,T\subseteq X$.
1. $\cl S=S\cup\did S$.
2. \[didcl additive\] $\did$ is additive: $\did(S\cup T)=\did S\cup\did T$.
3. \[didcl part5\] If $X$ is dense in itself, then (i) $\int S\subseteq\did S$, and (ii) if $S$ is open then $\did S=\cl S$.
Easy.
Regular open sets
-----------------
Let $X$ be a topological space. A subset of $X$ is one equal to the interior of its closure. We will mainly be interested in subsets of open subspaces of $X$, so we give definitions directly for such situations.
\[def:ro subset\] Let $U$ be an open subset of $X$. A subset $S$ of $X$ is said to be a * subset of $U$* if $S=\int(U\cap\cl S)$.
As ‘$\int$’ is multiplicative and $U$ is open, it is equivalent to say that $S=U\cap\int\cl S$, and we sometimes prefer this formulation. In such a case, $S\subseteq U$ and $S$ is open. So $S=\int_U\cl_US$: $S$ is a subset of the subspace $U$ of $X$. It is worth noting that if $S\subseteq U$ is arbitrary then $\int_U\cl_U S$ is a subset of $U$.
It is known (see, e.g., [@GiHa:ba chapter 10]) that for every open subset $U$ of $X$, the set $RO(U)$ of subsets of $U$ is closed under the operations $+,\cdot,-,0,1$ defined by
- $S+S'=U\cap\int\cl(S\cup S')$
- $S\cdot S'=S\cap S'$
- $-S=U\setminus\cl S$
- $0=\emptyset$ and $1=U$,
and $(RO(U),+,\cdot,-,0,1)$ is a (complete) . We will also use the notation $RO(U)$ to denote this . The standard boolean ordering $\leq$ on $RO(U)$ coincides with set inclusion, because for $S,T\in RO(U)$ we have $S\leq T$ iff $S\cdot T=S$, iff $S\cap T=S$, iff $S\subseteq T$. We will need the following general lemma.
\[lem:ro1\] Let $V\subseteq U$ be open subsets of $X$, and $S,S'$ be subsets of $U$.
1. \[lem:ro:part1\] If $T=U\setminus\cl S$, then $T$ is also a subset of $U$, with $S=U\setminus\cl T$ and $U\setminus S\subseteq\cl T$.
2. \[lem:ro:part2\] If $U\cap\cl S\cap\cl S'=\emptyset$, then $S+S'=S\cup S'$.
3. \[lem:ro:part3\] If $S\subseteq V$, then $S$ is a subset of $V$.
4. \[lem:ro:part4\] Every subset of $S$ is a subset of $U$.
<!-- -->
1. The first two points follow from considerations, and can easily be shown directly. The third point, $U\setminus S\subseteq\cl T$, follows from $U\setminus\cl T=S$.
2. Since $S,S'\leq S+S'$ and $\leq$ coincides with $\subseteq$, we obtain $S,S'\subseteq S+S'$ and so $S\cup S'\subseteq S+S'$. Conversely, it is easy to check[^4] that $$\int\cl(S\cup S')\subseteq \int\cl S\cup\int\cl S'\cup(\cl S\cap\cl S').$$ Since $U\cap\cl S\cap\cl S'=\emptyset$, $$S+S'=U\cap\int\cl(S\cup S')\subseteq (U\cap\int\cl S)\cup(U\cap\int\cl S')=S\cup S',$$ as required.
3. $V\cap\int\cl S=(V\cap U)\cap \int\cl S=V\cap (U\cap \int\cl S)=V\cap S=S$.
4. Let $T$ be a subset of $S$. Clearly, $\int\cl T\subseteq\int\cl S$. So $U\cap\int\cl T=U\cap(\int\cl S\cap\int\cl T)=(U\cap\int\cl S)\cap\int\cl T=S\cap\int\cl T=T$.
Normal spaces
-------------
\[def:normal\] A topological space $X$ is said to be *Hausdorff* (or T2) if for every two distinct points $x_0,x_1\in X$, there are disjoint open sets $O_0,O_1$ with $x_0\in O_0$ and $x_1\in O_1$, and *normal* (or T4) if it is Hausdorff and for every two disjoint closed subsets $C_0,C_1$ of $X$, there are disjoint open sets $O_0,O_1$ with $C_0\subseteq O_0$ and $C_1\subseteq O_1$.
Equivalently, $X$ is normal iff it is Hausdorff and if $C\subseteq O\subseteq X$, $C$ closed, and $O$ open, then there is open $P$ with $C\subseteq P\subseteq\cl P\subseteq O$.
\[lem:norm++\] Let $C_0,C_1$ be disjoint closed subsets of a normal topological space $X$. Then there are subsets $O_0,O_1$ of $X$ with disjoint closures, such that $C_0\subseteq O_0$ and $C_1\subseteq O_1$.
By normality, there are disjoint open sets $O^-_0\supseteq C_0$ and $U\supseteq C_1$. Then $O_0^-\subseteq X\setminus U$, a closed set. So $O_0=\int\cl O_0^-$ is a subset of $X$ disjoint from $U$. We have $C_0\subseteq O_0^-\subseteq O_0\subseteq\cl O_0\subseteq X\setminus U$, so $\cl O_0$ and $C_1$ are disjoint closed sets. By normality again, there are disjoint open sets $V\supseteq \cl O_0$ and $O_1^-\supseteq C_1$. Let $O_1=\int\cl O_1^-$, a subset of $X$ disjoint from $V$. Then $C_1\subseteq O_1^-\subseteq O_1\subseteq\cl O_1\subseteq X\setminus V$, so $\cl O_0\cap\cl O_1=\emptyset$. Now $O_0,O_1$ are as required.
The following is well known (see, e.g., [@RS:mm III, 6.1]), but is so important for us that we include a quick proof.
\[lem:ms\] Every metric space is normal.
Let $X$ be a metric space. It is easy to check that $X$ is Hausdorff, and we leave this to the reader. Let $C,D$ be disjoint closed subsets of $X$. By symmetry, it is enough to show that there is open $O\supseteq C$ with $\cl(O)\cap D=\emptyset$. If $C=\emptyset$, take $O=\emptyset$. If $D=\emptyset$ take $O=X$. So we can suppose $C,D\neq\emptyset$, and thus define $$O=\{x\in X:d(x,C)<d(x,D)/2\}$$ (recall from section \[ss:metric spaces\] that $d(x,S)=\inf\{d(x,s):s\in S\}$ for non-empty $S\subseteq X$). Then $C\subseteq O$, because if $x\in C$ then $d(x,C)=0$, while $x\notin D$, so $d(x,D)>0$ as $D$ is closed. It is easily seen that $O$ is open and $\cl(O)\subseteq \{x\in X:d(x,C)\leq d(x,D)/2\}$, so it is enough to show that this latter set is disjoint from $D$. If $x$ is in both, then $d(x,C)\leq d(x,D)/2=0$ so $x\in C$ as $C$ is closed. This contradicts the assumption that $C\cap D=\emptyset$.
Tarski’s theorem and relatives
------------------------------
The primary topological results needed later (for representing finite Kripke frames in proposition \[prop:di rep\]) are provided by the next theorem. A recent related result is [@KudShe14 proposition 11].
\[thm:Tarski\]\[thm:Tarski++\] Let $X$ be a space.
1. \[Tarski thm 1\] Let $\setT,\setU$ be open subsets of $X$, with $\emptyset\neq\setT\subseteq\setU$. Let $k<\omega$. Then there are pairwise disjoint non-empty subsets $\Vec\setI k\subseteq\setT$ satisfying $$\did\setI_i=\cl(\setT)\setminus\setU\quad\mbox{for each }i\leq k.$$
2. \[Tarski thm 2\]Let $\setG$ be a non-empty open subset of $X$, and let $r,s<\omega$. Then $\setG$ can be partitioned into non-empty open subsets $\vec \setG r$ and other non-empty sets $\Vec \setB s$ such that, letting $$D=\cl(\setG)\setminus \bigcup_{1\leq l\leq r}\setG_l,$$ we have $\cl(\setG_i)\setminus \setG_i=D$ for each $i=1,\ldots,r$, and $\did\setB_j=D$ for each $j=0,\ldots,s$.
Part \[Tarski thm 2\] above is essentially known. Paraphrasing slightly, Tarski [@Tar38 satz 3.10] proved the following. Let $X$ be a dense-in-itself normal topological space with a countable basis of open sets (see below). Then for every $r<\omega$, every non-empty open subset $\setG$ of $X$ can be partitioned into non-empty open sets $\vec\setG r$ and a non-empty set $\setB_0$ such that $\cl(\setG)\setminus\setG\subseteq\cl\setB_0
\subseteq\cl\setG_1\cap\ldots\cap\cl\setG_r$. Here and below, the empty intersection (when $r=0$) is taken to be $X$. This statement is equivalent to the statement in part \[Tarski thm 2\] of theorem \[thm:Tarski++\] above in the case $s=0$ and with $\did \setB_j$ replaced by $\cl\setB_j$.
A topological space $(X,\tau)$ has a countable basis of open sets iff there is countable $\tau_0\subseteq\tau$ such that $\tau$ is the smallest topology on $X$ containing $\tau_0$. Given this and normality, Urysohn’s theorem [@Urysohn25:met] yields that $\tau=\tau_d$ for some metric $d$ on $X$. Any metric space is normal, and has a countable basis of open sets iff it is separable (see section \[ss:top spaces\]). So Tarski’s stipulation on $X$ boils down to stipulating that $X$ is a separable space.
Removing the restriction to $s=0$ but with the same hypotheses on $X$, McKinsey and Tarski [@McKiT44 theorem 3.5] proved that for every $r,s<\omega$, every non-empty open set $\setG$ can be partitioned into non-empty open sets $\vec\setG r$ and non-empty sets $\Vec\setB s$ with $\cl(\setG)\setminus\setG\subseteq\cl\setB_0=\cdots=\cl\setB_s\subseteq\cl\setG_1\cap\ldots\cap\cl\setG_r$. This statement is equivalent to the statement of theorem \[thm:Tarski++\](\[Tarski thm 2\]) above, with $\did \setB_j$ replaced by $\cl\setB_j$. It was used in [@McKiT44] to prove (in our terminology) that the $\c L_\bo$-logic of $X$ is S4.
Removing the assumption of separability, Rasiowa and Sikorski [@RS:mm III, 7.1] proved theorem \[thm:Tarski++\](\[Tarski thm 2\]) as formulated above, but with $\did \setB_j$ replaced by $\cl\setB_j$. Our use of $\did\setB_j$ is only a formal strengthening of [@RS:mm III, 7.1], since the same effect can be achieved by first obtaining disjoint sets $\setB_j^i$ with $\cl\setB_j^i=D$ for $j=0,\ldots,s$ and $i=0,1$, and then defining $\setB_j=\setB_j^0\cup\setB_j^1$ for each $j$. As $\setB_j^0\cap\setB_j^1=\emptyset$, using lemma \[lem:did cl\] we have $$D\subseteq (D\setminus\setB_j^0)\cup(D \setminus\setB_j^1)
=(\cl\setB_j^0\setminus\setB_j^0)\cup(\cl\setB_j^1\setminus\setB_j^1)
\subseteq\underbrace{\did\setB_j^0\cup\did\setB_j^1}_{\did\setB_j}
\subseteq
\cl\setB_j^0\cup\cl\setB_j^1=D,$$ so $\did\setB_j=D$ as required. Given this, the reader may ask why we give a proof of part 2 at all. The answer is that we wish to make clear the affinity between the two parts of the theorem, as well as make our paper more self contained and explicit as to the topological arguments needed in our completeness proof.
We will get to the theorem shortly, but first, fix $k<\omega$. We define a game, $\c G_k$, to build pairwise disjoint subsets $\Vec\setI k$ of $X$. The game has two players, (male) and (female), and $\omega$ rounds, numbered $0,1,2,\ldots.$ At the start of round $n$ (for each $n<\omega$), pairwise disjoint sets $\Vec{I^n}k\subseteq X$ are in play, satisfying $$\label{e:did=0}
\did I_i^n=\emptyset\quad\mbox{ for each }i\leq k.$$ Observe that each $I_i^n$ is closed, because by lemma \[lem:did cl\], $\cl I^n_i=I^n_i\cup\did I^n_i=I^n_i$. Also, $$\label{e:UIjn empty int}
\int\Big(\bigcup_{j\leq k}I^n_j\Big)=\emptyset.$$ For if $U\subseteq\bigcup_{j\leq k}I^n_j$ is open, then by lemma \[lem:did cl\] and , $$U\subseteq\cl U=\did U\subseteq\did\bigcup_{j\leq k}I_j^n
=\bigcup_{j\leq k}\did I_j^n=\emptyset.$$ The game starts off with all of the sets empty: $I_0^0=\cdots=I_{k}^0=\emptyset$. Round $n$ is played as follows. Player moves first, by playing a triple $
(\varepsilon_n,i_n,O_n),
$ of his choice, where $\varepsilon_n>0$ is a real number, $i_n\leq k$, and $O_n$ is a non-empty open subset of $X$. Let $$\label{e:Pn}
P_n=O_n\setminus\bigcup_{j\leq k}I_j^n.$$ Then $P_n\neq\emptyset$: for otherwise, $\emptyset\neq O_n\subseteq\bigcup_{j\leq k}I_j^n$, contradicting . Player responds to ’s move by using Zorn’s lemma to choose a maximal subset $Z_n\subseteq P_n$ such that $d(x,y)\geq\varepsilon_n$ for each distinct $x,y\in Z_n$. Observe that
1. \[cond Z1\] $\did Z_n=\emptyset$ (because for all $x\in X$, the set $N_{\varepsilon_n/2}(x)\cap Z_n$ has at most one element). Just as with $I^n_i$ above, it follows that $Z_n$ is closed.
2. \[cond Z2\] $Z_n$ is non-empty (because $P_n$ is non-empty and any singleton subset of $P_n$ satisfies the $\varepsilon_n$-condition).
3. \[cond Z3\] $d(x,Z_n)<\varepsilon_n$ for every $x\in P_n$ (else $x$ can be added to $Z_n$, contradicting its maximality). Recall again that $d(x,Z_n)=\inf\{d(x,z):z\in Z_n\}$, which is defined because $Z_n$ is non-empty.
Player then extends $I^n_{i_n}$ by $Z_n$, leaving the other sets $I^n_i$ unchanged. Formally, she defines $$\begin{array}{rcll}
I_{i_n}^{n+1}&=&I_{i_n}^n\cup Z_n,
\\[4pt]
I_i^{n+1}&=&I_i^n&\mbox{for each }i\leq k\mbox{ with }i\neq i_n.
\end{array}$$ This completes the round, and the sets $\Vec{I^{n+1}}k$ are passed to the start of round $n+1$. Note that holds for these sets, since $\did I_{i_n}^{n+1}=\did I_{i_n}^n\cup\did Z_n=\emptyset$ by lemma \[lem:did cl\], for $I_{i_n}^n$, and \[cond Z1\] above. Also, by , $Z_n$ is disjoint from each $I^n_i$, so the $I^{n+1}_i$ $(i\leq k)$ are pairwise disjoint.
At the end of the game, we define $\setI_i=\bigcup_{n<\omega}I_i^n$ for each $i\leq k$. Plainly, $\Vec\setI k$ are pairwise disjoint.
We say that *plays well* in $\c G_k$ if his choices of $\varepsilon_n$ tend to zero, the set $\{n<\omega:i_n=i\}$ is infinite for each $i\leq k$, and his choices of $O_n$ form a descending chain: $O_0\supseteq O_1\supseteq\cdots$.
It is clear by condition \[cond Z2\] above that if plays well then $\Vec\setI k$ are all non-empty.
In any play (match?) of the game in which plays well, for each $i\leq k$ we have $$\did\setI_i=\bigcap_{n<\omega}\cl O_n.$$
Let $n<\omega$. Define $I_i^{>n}=\setI_i\setminus I_i^n$. This is the set of points that added to $\setI_i$ in or after round $n$. By the game rules and because played well, $I_i^{>n}\subseteq\bigcup_{n\leq m<\omega}Z_m
\subseteq\bigcup_{n\leq m<\omega}O_m=O_n$. Obviously, $\setI_i=I_i^n\cup I_i^{>n}$. So by lemma \[lem:did cl\] and , $$\did\setI_i=\did(I_i^n\cup I_i^{>n})=\did I_i^n\cup\did I_i^{>n}
=\did I_i^{>n}\subseteq\did O_n\subseteq\cl O_n.$$ This holds for all $n$, so $\did\setI_i\subseteq\bigcap_{n<\omega}\cl O_n$.
Conversely, let $x\in\bigcap_{n<\omega}\cl O_n$. Let a real number $\varepsilon>0$ be given. Since plays well, we can pick a round, say $n$, such that chose $\varepsilon_n\leq\varepsilon$ and $i_n=i$, and such that if $x\in\setI_i$ then already $x\in I_i^n$. Since $x\in\cl O_n$, the set $N_\varepsilon(x)\cap O_n$ is non-empty, and plainly it is open. As before, implies that $N_\varepsilon(x)\cap O_n\setminus\bigcup_{j\leq k}I^n_j$ is non-empty as well. Fix a point $y$ in this set. Then $y\in P_n$ and $d(x,y)<\varepsilon$.
In round $n$, player picks $Z_n\subseteq P_n$ satisfying conditions \[cond Z1\]–\[cond Z3\] above. Observe that $x\notin Z_n$, because otherwise, $x\in Z_n\subseteq\setI_i$ (since $i_n=i$), so by assumption on $n$ we have $x\in I_i^n$, so by , $x\notin P_n\supseteq Z_n$, a contradiction. Since $y\in P_n,$ by \[cond Z3\] we have $d(y,Z_n)<\varepsilon_n$. Since $d(x,y)<\varepsilon$, we have $d(x,Z_n)<\varepsilon+\varepsilon_n\leq 2\varepsilon$. So there is $z\in Z_n\subseteq\setI_i$ with $z\neq x$ (since $x\notin Z_n$) and $d(x,z)<2\varepsilon$. This holds for all $\varepsilon>0$, and it follows that $x\in\did\setI_i$, proving the claim.
Now we prove part 1 of the theorem. Suppose first that $\cl(\setT)\setminus\setU=\emptyset$. Noting that $\setT$ is infinite (by lemma \[lem:inf\]), we can take $\Vec\setI k$ to be disjoint singleton subsets of $\setT$. Plainly, all requirements are met.
So suppose that $\cl(\setT)\setminus\setU\neq\emptyset$. Let and play the game $\c G_k$. We suppose that plays well, and also so that for each $n<\omega$, $$O_n=\setT\cap\bigcup_{x\in \cl(\setT)\setminus\setU}N_{\varepsilon_n}(x).$$ Note that $O_n$ is open, and non-empty because $\cl(\setT)\setminus\setU\neq\emptyset$, so can legally play it. Then $\Vec\setI k$ are pairwise disjoint, and non-empty since plays well. We have $Z_n\subseteq O_n\subseteq\setT$ for each $n$, so $\Vec\setI k$ are subsets of $\setT$. By the claim, $\did\setI_i=\bigcap_{n<\omega}\cl O_n$ for each $i\leq k$, so it suffices to show that $\bigcap_{n<\omega}\cl O_n=\cl(\setT)\setminus\setU$.
Certainly, each $x\in\cl(\setT)\setminus\setU$ lies in $\cl O_n$ for each $n$, because for every $\varepsilon>0$, $$O_n\cap N_\varepsilon(x)
\supseteq
\Big(\setT\cap\bigcup_{y\in \cl(\setT)\setminus\setU}N_{\varepsilon_n}(y)\Big)
\cap N_{\min(\varepsilon,\varepsilon_n)}(x)
=\setT\cap N_{\min(\varepsilon,\varepsilon_n)}(x)\neq\emptyset.$$ So $\cl(\setT)\setminus\setU\subseteq\bigcap_{n<\omega}\cl O_n$. Conversely, first note that $O_0\subseteq\setT$, so $\bigcap_{n<\omega}\cl O_n\subseteq\cl O_0\subseteq\cl\setT$. It remains to show that $\setU\cap\bigcap_{n<\omega}\cl O_n=\emptyset$. Suppose for contradiction that there is some $x\in\setU\cap\bigcap_{n<\omega}\cl O_n$. As $\setU$ is open, we can choose $\delta>0$ with $N_\delta(x)\subseteq\setU$. As played well, we can pick $n<\omega$ such that $\varepsilon_n\leq\delta$. Then $x\in\cl O_n$, so $d(x,O_n)=0$. By definition of $O_n$, for each $y\in O_n$ we have $d(y,\cl(\setT)\setminus\setU)<\varepsilon_n$. So $d(x,\cl(\setT)\setminus\setU)<\varepsilon_n$ as well. As $\varepsilon_n\leq\delta$ and $N_\delta(x)\subseteq\setU$, this is a contradiction. We conclude that indeed $\setU\cap\bigcap_{n<\omega}\cl O_n=\emptyset$, so $\bigcap_{n<\omega}\cl O_n\subseteq\cl(\setT)\setminus\setU$, as required. We have proved part 1 of the theorem.
To prove part 2, let and play $\c G_{s+r}$. As we will see, will play so that $\Vec\setI{s+r}\subseteq\setG$. In the end, $\vec\setB s$ will be $\vec\setI s$, $\vec\setG r$ will be ‘fattened’ versions of $\setI_{s+1},\ldots,\setI_{s+r}$, and $\setB_0$ will be the rest of $\setG$ (we will have $\setB_0\supseteq\setI_0$). For the fattening, at the start of round $n$ (for each $n<\omega$), for each $j=s+1,\ldots,s+r$, defines an auxiliary open set $G^n_j$ such that $$\begin{aligned}
I^n_{j}\subseteq G^n_{j}
\label{e:G sets 1}
\\
G^0_j\subseteq G^{1}_j\subseteq\cdots
\label{e:G sets 2}
\\
I^n_0,\ldots,I^n_s,\cl G^n_{s+1},\ldots,\cl G^n_{s+r}
\mbox{ are pairwise disjoint subsets of }\setG.
\label{e:G sets 3}\end{aligned}$$ The sets $G^n_j$ are for ’s own private use and are not formally part of the game. (If $r=0$, there are no $j$ in range and he does nothing.) At the start of round 0, he simply puts $G^0_{s+1}=\cdots=G^0_{s+r}=\emptyset$. Suppose we are at the start of round $n$, for arbitrary $n<\omega$, and that has defined open $G^n_{j}\supseteq I^n_{j}$ ($s+1\leq j\leq s+r$) satisfying –. In round $n$ he plays $(\varepsilon_n,i_n,O_n)$, where $i_0=0$, $$\label{e:pa's O_n in Tarski}
O_n=\setG\setminus\bigcup_{s+1\leq j\leq s+r}\cl G^n_j,$$ and the $\varepsilon_n,i_n$ are chosen so that overall, he plays well. By , $O_0\supseteq O_1\supseteq\cdots$, as required for him to play well. (We remark that if $r=0$ then $O_n=\setG$ for all $n$.)
We check that this is always a legal move for . Certainly, $O_n$ is open. We show that it is always non-empty. For $n=0$ we plainly have $O_0=\setG\neq\emptyset$. In round $0$, plays $i_0=0$, and defines $I^1_0=Z_0\neq\emptyset$ by condition \[cond Z2\] above. Since the $I^n_0$ form a chain, $I^n_0\supseteq I^1_0\neq\emptyset$ for all $n>0$, and by and , $I^n_0\subseteq O_n$. So $O_n\neq\emptyset$ for all $n$.
Player continues round $n$ by selecting $Z_n\subseteq P_n$ and defining $I^{n+1}_{i_n}=I^n_{i_n}\cup Z_n$ according to the rules.
It is now time for to define $G^{n+1}_j$ for $j=s+1,\ldots,s+r$. If $i_n\leq s$, he leaves the sets unchanged, defining $G^{n+1}_j=G^n_j$ for all $j$. Trivially, conditions – continue to hold. We check . First, $Z_n\subseteq P_n$, so $I^{n+1}_{i_n}$ is disjoint from $I^{n+1}_j$ for $i_n\neq j\leq s$. Second, if $s+1\leq j\leq s+r$ then $I^{n+1}_{i_n}=I^n_{i_n}\cup Z_n\subseteq I^n_{i_n}\cup O_n$; by , $I^n_{i_n}$ is disjoint from $\cl G^n_j=\cl G^{n+1}_j$, and by , $O_n$ is disjoint from $\cl G^{n+1}_j$ as well.
If instead, $i_n>s$, then defines $G^{n+1}_j=G^n_j$ for $j\neq i_n$, and uses normality of $X$ to choose an open set $G^{n+1}_{i_n}$ satisfying $$\label{e:A's set}
\overbrace{\cl(G_{i_n}^n)\cup Z_n}^{\rm closed}\subseteq G^{n+1}_{i_n}\subseteq \cl(G^{n+1}_{i_n})\subseteq
\overbrace{\setG\setminus
\big(\bigcup_{j\leq s}I^n_j\;\;\cup\bigcup_{\genfrac{}{}{0pt}{1}{s+1\leq j\leq s+r}{j\neq i_n}}\cl(G^n_{j})\big)}^{\rm open}.$$ We need to check some things here. First, by condition \[cond Z1\] above, $Z_n$ is closed and so the left-hand side of is closed. Similarly, we saw just after that each $I^n_j$ is closed, so the right-hand side of is open. Second, it follows from that $\cl(G_{i_n}^n)$ is contained in the right-hand side of . Also $Z_n\subseteq P_n\subseteq O_n$, and it follows from and that $Z_n$ is contained in the right-hand side of as well. So $G^{n+1}_{i_n}$ can be found as stated.
We also need to check – for the $G^{n+1}_j$. Condition holds because $I^{n+1}_{i_n}=I^n_{i_n}\cup Z_n\subseteq G^n_{i_n}\cup Z_n\subseteq G^{n+1}_{i_n}$, and for $j\neq i_n$ we have $G^{n+1}_j=G^n_j\supseteq I^n_j=I^{n+1}_j$. Conditions and are clear from the definitions and .
As promised, at the end of play we define $$\begin{array}{rcll}
\setG_i&=&\displaystyle\bigcup_{n<\omega}G^n_{s+i}&\mbox{for }1\leq i\leq r,
\\[14pt]
\setB_j&=&\setI_j&\mbox{for }1\leq j\leq s,
\\[4pt]
\setB_0&=&\displaystyle\setG\setminus\Big(\bigcup_{1\leq i\leq r}\setG_i
\cup\bigcup_{1\leq j\leq s}\setB_j
\Big)
\\[14pt]
D&=&\displaystyle\cl(\setG)\setminus\bigcup_{1\leq l\leq r}\setG_l.
\end{array}$$ Note that $\setI_{s+i}\subseteq\setG_i$ for $1\leq i\leq r$ by , and $\setI_j\subseteq\setB_j$ for $j\leq s$ by the definitions. Because played well, the $\setG_j$ are non-empty (and plainly open) and the $\setB_j$ are non-empty. It follows from that together they partition $\setG$.
For the final piece of the theorem, there are two preliminaries. First, we observe that each set $\setG_i$ ($1\leq i\leq r$) has a nice property. Each time plays $i_n= s+i$ in some round $n$, by , , and the definition of $\setG_i$, for every $m\leq n$ we have $\cl G^m_{s+i}\subseteq\cl G^n_{s+i}\subseteq G^{n+1}_{s+i}\subseteq\setG_i$. Since played $i_n=s+i$ infinitely often, it follows that $$\label{e:closure inside}
\cl G^m_{s+i}\subseteq\setG_i\quad\mbox{for each }
m<\omega\mbox{ and } 1\leq i\leq r.$$ Second, we use this to show that $$\label{e:value of D}
D=\bigcap_{n<\omega}\cl O_n.$$ Note that if $C\subseteq S\subseteq X$ and $C$ is closed, then $S= C\cup(S\setminus C)\subseteq C\cup\cl(S\setminus C)$; the right-hand side is closed, so $\cl S\subseteq C\cup\cl(S\setminus C)$, whence $\cl(S)\setminus C\subseteq\cl(S\setminus C)$. Now, for each $n<\omega$ we have $$\begin{array}{rcll}
D&=&\cl(\setG)\setminus\bigcup_{1\leq i\leq r}\setG_i
&\mbox{by definition}
\\
&\subseteq&\cl(\setG)\setminus\bigcup_{1\leq i\leq r}\cl G^n_{s+i}
&\mbox{by \eqref{e:closure inside}}
\\
&\subseteq&\cl\big(\setG\setminus\bigcup_{1\leq i\leq r}\cl G^n_{s+i}\big)
\quad
&
\mbox{by the observation above}
\\
&=&\cl O_n&\mbox{by \eqref{e:pa's O_n in Tarski}.}
\end{array}$$ So $D\subseteq\bigcap_{n<\omega}\cl O_n$. Conversely, we certainly have $\bigcap_{n<\omega}\cl O_n\subseteq\cl O_0=\cl\setG$ since $O_0=\setG$. Now fix $i$ with $1\leq i\leq r$. By , for each $n<\omega$ we have $G^n_{s+i}\cap O_n=\emptyset$, so as $G^n_{s+i}$ is open, $G^n_{s+i}\cap\cl O_n=\emptyset$. It follows that $$\setG_i\cap\bigcap_{n<\omega}\cl O_n=(\bigcup_{n<\omega}G^n_{s+i})\cap\bigcap_{n<\omega}\cl O_n=\emptyset.$$ This holds for each $i$, so $\bigcap_{n<\omega}\cl O_n\subseteq\cl(\setG)\setminus\bigcup_{1\leq i\leq r}\setG_i=D$, proving .
Now we can finish easily. For each $0\leq j\leq s$, we plainly have $\setB_j\subseteq\setG\setminus\bigcup_{1\leq l\leq r}\setG_l
\subseteq D$. Since $D$ is closed, $\did\setB_j\subseteq\cl\setB_j\subseteq D$. Conversely, by and the claim, $D=\bigcap_{n<\omega}\cl O_n=\did\setI_j\subseteq\did\setB_j$.
Similarly, take $i$ with $1\leq i\leq r$. Since the $\setG_l$ ($1\leq l\leq r$) are pairwise disjoint open subsets of $\setG$, we have $\cl\setG_i\subseteq\cl(\setG)\setminus\bigcup_{l\neq i}\setG_l$ and hence $\cl(\setG_i)\setminus\setG_i\subseteq
\cl(\setG)\setminus\bigcup_{1\leq l\leq r}\setG_l=D$. Conversely, by , the claim, and lemma \[lem:did cl\] we have $D=\bigcap_{n<\omega}\cl O_n=\did\setI_{s+i}\subseteq\did\setG_i\subseteq\cl\setG_i$. By definition, $D\cap\setG_i=\emptyset$. So $D\subseteq\cl(\setG_i)\setminus\setG_i$, as required.
\[cor:Tarski++opens\] Let $\setU$ be an open subspace of a space $X$, and suppose that $\setS_0$, $\setS_1$ are open subsets of $\setU$ such that $\setU\cap\cl \setS_0\cap\cl \setS_1=\emptyset$ and $\setT=\setU\setminus\cl(\setS_0\cup \setS_1)\neq\emptyset$. Then there are subsets $\setU_0,\setU_1$ of $\setU$ such that $\setU\cap\cl\setU_0\cap\cl\setU_1=\emptyset$, and for each $i=0,1$:
1. $\setU\cap\cl \setS_i\subseteq\setU_i$,
2. writing $\setT_i=\setU_i\setminus\cl \setS_i$, we have $\setT_i\neq\emptyset$ and $\cl(\setT)\setminus\setU\subseteq\cl\setT_i$.
Since $\setT$ is a non-empty open subset of $\setU$, we can use theorem \[thm:Tarski++\] to choose disjoint non-empty subsets $\setI_0,\setI_1\subseteq\setT$ such that $\did\setI_0=\did\setI_1=\cl(\setT)\setminus\setU$.
We now work in the subspace $\setU$. Recall that $\cl_{\setU}$ denotes the closure operator in the subspace topology on $\setU$, so $\cl_{\setU} K=\setU\cap\cl K$ for subsets $K\subseteq\setU$. The sets $$\cl\nolimits_{\setU} \setS_0,
\;
\cl\nolimits_{\setU}\setS_1,
\;
\setI_0,
\;
\setI_1$$ are pairwise disjoint (by assumptions) and closed in $\setU$. (Each $\setI_i$ is closed in $\setU$ because by lemma \[lem:did cl\], $\cl_{\setU}\setI_i=\setU\cap\cl\setI_i=\setU\cap(\setI_i\cup\did\setI_i)
=\setU\cap(\setI_i\cup(\cl(\setT)\setminus\setU))=\setU\cap\setI_i=\setI_i$.) Hence, $\setI_0\cup\cl\nolimits_{\setU}\setS_0$ and $\setI_1\cup\cl\nolimits_{\setU}\setS_1$ are disjoint closed subsets of $\setU$. The subspace $\setU$ is a metric space in its own right, and so, by lemma \[lem:ms\], normal. Using lemma \[lem:norm++\] in $\setU$, we can find subsets $\setU_0,\setU_1$ of $\setU$ with $$\label{cor:pf:cond1}
\setI_i\cup\cl\nolimits_{\setU}\setS_i\subseteq\setU_i\subseteq\setU
\quad\mbox{for } i=0,1,$$ and $\cl\nolimits_{\setU}\setU_0\cap\cl\nolimits_{\setU}\setU_1=\emptyset$. Working back in $X$ again, this says that $$\label{cor:pf:cond2}
\setU\cap\cl\setU_0\cap\cl\setU_1=\emptyset.$$ Now for each $i=0,1$, write $\setT_i=\setU_i\setminus\cl \setS_i$. By definition, $\setI_i\subseteq\setU_i$. Also, $\setI_i\cap(\setU\cap\cl \setS_i)=\emptyset$, and since $\setI_i\subseteq\setU$, this gives $\setI_i\cap\cl \setS_i=\emptyset$. Hence, $\setI_i\subseteq\setT_i$, so $\setT_i\neq\emptyset$. We now obtain $$\label{cor:pf:cond3}
\cl(\setT)\setminus\setU=\did\setI_i\subseteq\cl\setI_i\subseteq\cl\setT_i.$$ Lines , , and , together with $\setT_i\neq\emptyset$, establish the corollary.
Representations of frames over topological spaces
=================================================
Our next aim is to use the results of the preceding section to construct a ‘’ from an arbitrary space to any given finite connected locally connected KD4 Kripke frame. The notion of is chosen so as to preserve $\c L_{\bod\forall}^\mu$-formulas, and this will allow us to prove completeness theorems in the next two sections.
Until the end of section \[ss:full rby\], we fix a topological space $X$ and a finite Kripke frame $\c F=(W,R)$. We will frequently regard the elements of $W$ as propositional atoms.
Representations
---------------
The following definition seems to originate with Shehtman: see equation (71) in [@Sheh:d90 §5, p.25].
A map $\rho:X\to W$ is said to be a * of $\c F$ over $X$* if for every $x\in X$ and $w\in W$ we have $$(X,\rho^{-1}),x\models\did w\iff R(\rho(x),w).$$
Here, $\rho^{-1}$ assigns an atom $w\in W$ to the possibly empty subset $\{x\in X:\rho(x)=w\}$ of $X$. The condition says that for every $x\in X$, the set of points of $W$ with preimages under $\rho$ in every open of $x$ but distinct from $x$ itself is precisely $R(\rho(x))$. Equivalently, $\did\rho^{-1}(w)=\rho^{-1}(R^{-1}(w))$ for every $w\in W$, where $R^{-1}$ is the converse relation of $R$.
Note that $\rho$ need not be surjective. Indeed, the empty map is vacuously a of $\c F$ over the empty space — and we definitely do allow empty s.
It can be checked that if $\rho:X\to W$ is a then $R\restriction\rng\rho$ is transitive. Endow $W$ with the topology generated by $\{R(w):w\in W\}$ (so the open sets are those $A\subseteq W$ such that $a\in A$ implies $R(a)\subseteq A$). Then every of $\c F$ over $X$ is an interior map from $X$ to $W$: that is, a map that is both continuous and open. (Many other topological completeness proofs use interior maps.) The converse, however, does not hold in general. See [@GEG05; @LucBry11] for more information.
Although Shehtman uses the term ‘d-p-morphism’ (when $\rho$ is surjective), here we will call $\rho$ a ‘’ because it is closely related to the s of algebras of relations seen in algebraic logic. Indeed, if $\rho$ is a surjective of $(W,R)$ over $X$ then $\rho^{-1}$ induces an embedding from $\wp(W)$ into $\wp(X)$ that preserves the algebraic structure with which these power sets can be naturally endowed.
Representations over subspaces
------------------------------
Our main interest is in s over $X$ itself, but s over subspaces are also useful in proofs. Given a subspace $U$ of $X$, a map $\rho:U\to W$ induces a well defined assignment $\rho^{-1}:W\to\wp(X)$ by $\rho^{-1}(w)=\{x\in X:x\in U$ and $\rho(x)=w\}$, for $w\in W$. Put simply, preimages under $\rho$ of elements of $W$ are obviously subsets of $U$, but they are also subsets of $X$, and so $\rho^{-1}$ can be regarded equally as an assignment into $U$ or $X$, as appropriate. The following easy lemma gives some connections between the two views.
\[lem:bod reps invar\] Let $U$ be a subspace of $X$ and let $\rho:U\to W$ be a map. Let $x\in U$ and $w\in W$ be arbitrary.
1. If $(U,\rho^{-1}),x\models\did w$ then $(X,\rho^{-1}),x\models\did w$.
2. If $U$ is open in $X$, then $(U,\rho^{-1}),x\models\did w$ iff $(X,\rho^{-1}),x\models\did w$.
For the first part, assume that $(U,\rho^{-1}),x\models\did w$ and let $O$ be any open of $x$ in $X$. Then $O\cap U$ is an open of $x$ in $U$, so by assumption, there is $y\in O\cap U\setminus\{x\}$ with $(U,\rho^{-1}),y\models w$. Then $y\in O\setminus\{x\}$ and $(X,\rho^{-1}),y\models w$. Hence, $(X,\rho^{-1}),x\models\did w$.
For the second part, assume that $(X,\rho^{-1}),x\models\did w$. Let $N$ be an arbitrary open of $x$ in $U$, so that $N=O\cap U$ for some open $O$ of $x$ in $X$. As $U$ is assumed open in $X$, we see that $N$ is also open in $X$, so by assumption, there is $y\in N\setminus\{x\}$ with $(X,\rho^{-1}),y\models w$. Plainly, $(U,\rho^{-1}),y\models w$. This shows that $(U,\rho^{-1}),x\models\did w$, and the converse follows from the first part.
By part 2 of the lemma, if $\rho$ is a of $\c F$ over an open subspace $U$ of $X$, then $(X,\rho^{-1}),x\models\did w$ iff $R(\rho(x),w)$ for every $x\in U$ and $w\in W$. So we can work in $(X,\rho^{-1})$ instead of $(U,\rho^{-1})$. To avoid too much jumping around between subspaces, we will do this below, often without mention. Part \[lem:simple:3\] of the next lemma makes it a little more explicit. The lemma gives some general information on how s of different generated subframes of $\c F$ over different subspaces of $X$ are related.
\[lem:reps simple\] Let $\c G=(W',R')$ be a generated subframe of $\c F$. Let $T$, $U$, and $U_i$ $(i\in I)$ be open subspaces of $X$, with $T\subseteq U=\bigcup_{i\in I}U_i$. Finally, let $\rho:U\to W'$ be a map. Then:
1. \[lem:simple:1\] $\rho$ is a of $\c F$ over $U$ iff it is a of $\c G$ over $U$.
2. \[lem:simple:2\] $\rho$ is a of $\c F$ over $U$ iff for each $i\in I$, the restriction $\rho\restriction U_i$ is a of $\c F$ over $U_i$.
3. \[lem:simple:3\] If $\rho\restriction T$ is a of $\c F$ over $T$, then $(X,\rho^{-1}),x\models\did w$ iff $R(\rho(x),w)$, for each $x\in T$ and $w\in W$.
Simple.
Representations preserve formulas
---------------------------------
Here, we will show that surjective s preserve all formulas of $\c L_{\bod\forall}^{\mu}$. Since s are like p-morphisms, albeit between different kinds of , this is entirely expected and the proof is essentially quite standard — see [@Sheh:d90 lemma 20] and [@GEG05 corollary 2.9], for example. We do need, however, that $\c F$ is finite. We will be able to handle larger sublanguages of $\Lbig$ by using the translations of section \[sec:translations\].
Let us explain the setting. Suppose we are given a $\rho:X\to W$ of $\c F$ over $X$. Recall that $\Var$ is our fixed base set of propositional variables, or atoms. For each assignment $h:\Var\to\wp(W)$ of atoms in $\Var$ into $W$, the map $\rho^{-1}\circ h:\Var\to\wp(X)$ is an assignment of atoms into $X$, given of course by $$(\rho^{-1}\circ h)(p)=\{x\in X:\rho(x)\in h(p)\}, \quad\mbox{ for each }
p\in \Var.$$ So $\rho$, or rather $\rho^{-1}$, gives us a way to transform an assignment into $\c F$ to one into $X$, and then to evaluate a formula in the resulting model on $X$. The following definition encapsulates when we get the same result as in the original model on $\c F$:
Let $\rho:X\to W$ be a map, and let $\varphi$ be a formula of $\c L_{\bo\bod}^{\mu\dit\didt}$. We say that *$\rho$ preserves $\varphi$* if for every assignment $h:\Var\to\wp(W)$ and every $x\in X$, $$\label{e:rho pres phi}
(X,\rho^{-1}\circ h),x\models\varphi\quad\mbox{iff}\quad(W,R,h),\rho(x)\models\varphi.$$
We are now ready for our main preservation result.
\[prop:fmla pres\] Let $\rho:X\to W$ be a surjective of $\c F$ over $X$. Then $\rho$ preserves every formula of $\c L_{\bod\forall}^\mu$.
The proof is by induction on $\varphi$. The atomic and boolean cases are easy and left to the reader. Let $\varphi$ be a formula, and inductively assume for every assignment $h:\Var\to\wp(W)$ and every $x\in X$. It is sufficient to consider the cases $\did\varphi,$ $\forall\varphi$, and $\mu q\varphi$.
First, consider $\did\varphi$. Fix $h,x$. Suppose that $(W,R,h),\rho(x)\models\did\varphi$. Choose $w\in R(\rho(x))$ with $(W,R,h),w\models\varphi$. As $\rho$ is a , $(X,\rho^{-1}),x\models\did w$. So for every open $O$ of $x$, there is $y\in O\setminus\{x\}$ with $\rho(y)=w$. Since $(W,R,h),w\models\varphi$, for any such $y$ we inductively have $(X,\rho^{-1}\circ h),y\models\varphi$. It follows that $(X,\rho^{-1}\circ h),x\models\did\varphi$.
Conversely, suppose that $(X,\rho^{-1}\circ h),x\models\did\varphi$. Let $\sem\varphi=\{y\in X:(X,\rho^{-1}\circ h),y\models\varphi\}$. As $\c F$ is finite and $\did$ is additive (lemma \[lem:did cl\](\[didcl additive\])), we have $$\begin{aligned}
x\in \did \sem\varphi
&=\did(\sem\varphi\cap X)
=\did\Big(\sem\varphi\cap\bigcup_{w\in W}\rho^{-1}(w)\Big)
\\
&=
\did\Big(\bigcup_{w\in W}\big(\sem\varphi\cap\rho^{-1}(w)\big)\Big)
=\bigcup_{w\in W}\did(\sem\varphi\cap\rho^{-1}(w)).\end{aligned}$$ So we can take $w\in W$ with $x\in\did(\sem\varphi\cap\rho^{-1}(w))$. Then $(X,\rho^{-1}),x\models\did w$, so as $\rho$ is a , $R(\rho(x),w)$. Moreover, $\sem\varphi\cap\rho^{-1}(w)\neq\emptyset$. Take any $y\in\sem\varphi\cap\rho^{-1}(w)$. Then $(X,\rho^{-1}\circ h),y\models\varphi$ and $\rho(y)=w$. Inductively, $(W,R,h),w\models\varphi$. By Kripke semantics, $(W,R,h),\rho(x)\models\did\varphi$, as required.
Next, consider $\forall\varphi$. Then $(X,\rho^{-1}\circ h),x\models\forall\varphi$ iff $(X,\rho^{-1}\circ h),y\models\varphi$ for all $y\in X$, iff $(W,R,h),\rho(y)\models\varphi$ for all $y\in X$ (by the inductive hypothesis ), iff $(W,R,h),w\models\varphi$ for all $w\in W$ (since $\rho$ is surjective), iff $(W,R,h),\rho(x)\models\forall\varphi$.
Finally consider the case $\mu q\varphi$, assumed well formed. Fix arbitrary $h:\Var\to\wp(W)$. We define an assignment $h^\alpha:\Var\to\wp(W)$ for each ordinal $\alpha$. For each atom $p\neq q$, we set $h^\alpha(p)=h(p)$. We define $h^\alpha(q)$ by induction on $\alpha$ as follows:
- $h^0(q)=\emptyset$,
- $h^{\alpha+1}(q)=\{w\in W:(W,R,h^\alpha),w\models\varphi\}$,
- $h^\delta(q)=\bigcup_{\alpha<\delta}h^\alpha(q)$ for limit ordinals $\delta$.
Of course, $W$ is finite, but we need all ordinals for the argument below. Let $\eta=\rho^{-1}\circ h:\Var\to\wp(X)$. Define an assignment $\eta^\alpha:\Var\to\wp(X)$ in the same way as for $h^\alpha$: let $\eta^\alpha(p)=\eta(p)$ for all atoms $p\neq q$ and all $\alpha$, and
- $\eta^0(q)=\emptyset$,
- $\eta^{\alpha+1}(q)=\{x\in X:(X,\eta^\alpha),x\models\varphi\}$,
- $\eta^\delta(q)=\bigcup_{\alpha<\delta}\eta^\alpha(q)$ for limit ordinals $\delta$.
$\eta^\alpha(q)=\rho^{-1}(h^\alpha(q))$ for each ordinal $\alpha$.
By induction on $\alpha$. For $\alpha=0$ this is saying that $\rho^{-1}(\emptyset)=\emptyset$, which is true. Assume the result for $\alpha$ inductively. So $\eta^\alpha=\rho^{-1}\circ h^\alpha$. We now obtain $$\begin{array}{rcll}
\eta^{\alpha+1}(q)
&=&\{x\in X:(X,\eta^\alpha),x\models\varphi\}&\mbox{ by definition of }\eta^{\alpha+1}
\\
&=&\{x\in X:(X,\rho^{-1}\circ h^\alpha),x\models\varphi\}&\mbox{ since }\eta^\alpha=\rho^{-1}\circ h^\alpha
\\
&=&\{x\in X:(W,R, h^\alpha),\rho(x)\models\varphi\}&\mbox{ by inductive hypothesis }\eqref{e:rho pres phi}
\\
&=&\{x\in X:\rho(x)\in h^{\alpha+1}(q)\}&\mbox{ by definition of }h^{\alpha+1}
\\
&=&\rho^{-1}(h^{\alpha+1}(q)).
\end{array}$$ For limit $\delta$ we have $$\rho^{-1}(h^\delta(q))
=\rho^{-1}(\bigcup_{\alpha<\delta}h^\alpha(q))
=\bigcup_{\alpha<\delta}\rho^{-1}(h^\alpha(q))
=_{IH}\bigcup_{\alpha<\delta}
\eta^\alpha(q)=\eta^\delta(q).$$ This completes the induction on $\alpha$, and proves the claim.
By semantics of $\mu$, we have $(X,\eta),x\models\mu q\varphi$ iff $x\in\bigcup_{\alpha\in \rm On}\eta^\alpha(q)$, iff $x\in \bigcup_\alpha \rho^{-1}(h^\alpha(q))$ by the claim, iff $\rho(x)\in\bigcup_\alpha h^\alpha(q)$, iff $(W,R,h),\rho(x)\models\mu q\varphi$. This completes the induction and proves the proposition.
Basic s
-------
Certain very primitive s called *basic s* will play an important role later, because they can easily be extended to more interesting s.
Let $S,U$ be open subspaces of $X$, with $S\subseteq U$, and let $\sigma:S\to W$ be a of $\c F$ over $S$. We say that $\sigma$ is *$U$-basic* if for every $x\in U$ and $w,v\in W$, if $(X,\sigma^{-1}),x\models\di w\wedge\di v$ then $Rwv$.
Note that we use $\di$ and not $\did$ here.
\[rmk:basic reps\]In the setting of this definition:
1. Vacuously, if $\sigma$ is empty then it is $U$-basic.
2. More generally, but equally trivially, if $\rng\sigma$ is contained in a nondegenerate cluster $C$ in $\c F$, then $\sigma$ is $U$-basic. For, $(X,\sigma^{-1}),x\models\di w\wedge\di v$ implies that $w,v\in\rng\sigma\subseteq C$, and so $Rwv$ as $C$ is a nondegenerate cluster.
We remark (but will not formally use) that $\sigma$ is $U$-basic iff $\rng\sigma$ is a (possibly empty) union of $R$-maximal clusters in $\c F$ whose preimages under $\sigma$ have pairwise disjoint closures within $U$. Moreover, each such preimage is a subset of $S$.
Full s
------
In induction proofs, one often needs a stronger inductive hypothesis than formally required for the final result. This will be the case in proposition \[prop:di rep\] below, and the notion of *$T$-full* will be used to formulate it.
\[def:Tfull rep\] Let $T\subseteq U$ be open subspaces of $X$. A $\rho:U\to W$ of $\c F$ over $U$ is said to be *$T$-full* if:
1. for every $x\in \cl(T)\setminus U$ and $w\in W$, we have $(X,\rho^{-1}),x\models\did w$,
2. if $T$ is non-empty then $\rho:U\to W$ is surjective.
Every is vacuously $\emptyset$-full.
Full representability {#ss:full rby}
---------------------
\[def:fully rb did case\] We say that $\c F$ is *fully (over $X$)* if whenever
1. $U\subseteq X$ is open,
2. $S$ is a subset of $U$,
3. $\sigma:S\to W$ is a $U$-basic of $\c F$ over $S$,
4. $T=U\setminus\cl S$,
then $\sigma$ extends to a $T$-full $\rho:U\to W$ of $\c F$ over $U$.
Notice that in the $RO(U)$ of subsets of $U$, we have $T=-S$, so $\{S,T\}$ is a partition of 1. That is, $S,T\in RO(U)$, $S\cdot T=0$, and $S+T=1$.
In proposition \[prop:di rep\] below, we will fulfil our main aim, to prove (surjective) representability of every finite connected locally connected $\axK\axD4$-frame. We are going to do it by induction on the size of the frame; we appear to need a stronger inductive hypothesis, namely full representability, than is needed for the conclusion; $T$-fullness and extending $\sigma$ are mainly to do with this, but the $\sigma$ part is also helpful in the proof of strong completeness in theorem \[thm:strong did\] later. Note that if $\c F$ is fully over $X$, and $X\neq\emptyset$, then by taking $U=X$ and $S=\sigma=\emptyset$, we see that there exists a surjective of $\c F$ over $X$. So we do obtain our desired conclusion from the stronger hypothesis of full representability.
Main proposition {#ss:main rep prop}
----------------
The following proposition has relatives in the literature: see, e.g., [@McKiT44 theorem 3.7], [@Sheh:d90 proposition 22], [@LucBry11 lemma 4.4], and [@KudShe14 lemma 16]. It actually holds for any dense-in-itself topological space $X$ for which theorem \[thm:Tarski\] and corollary \[cor:Tarski++opens\] can be proved.
\[prop:di rep\] Suppose that $X$ is a space. Then every finite connected locally connected $\axK\axD4$ frame $\c F=(W,R)$ is fully over $X$.
The proof is by induction on the number of worlds in $\c F$. Let $\c F=(W,R)$ be a finite connected locally connected $\axK\axD4$ frame, and assume the result inductively for all smaller frames. Note that $R$ is transitive. Recall that we write
- $R^\circ=\{(w,v)\in W^2: Rwv\wedge Rvw\},$
- $R^\bullet=\{(w,v)\in W^2:Rwv\wedge\neg Rvw\},
$
and for $w\in W$,
- $\c F(w)$ for the subframe $(R(w),R\restriction R(w))$ of $\c F$ with domain $R(w)$,
- $\c F^*(w)$ for the subframe $(R^*(w),R\restriction R^*(w))=(R(w)\cup\{w\}, R\restriction R(w)\cup\{w\})$ of $\c F$ generated by $w$.
Let $U\subseteq X$ be open, let $S$ be a subset of $U$, and let $\sigma:S\to W$ be a $U$-basic of $\c F$ over $S$. Write $$T=U\setminus\cl S.$$ We need to extend $\sigma$ to a $T$-full $\rho:U\to W$ of $\c F$ over $U$.
If $T=\emptyset$, then $U\subseteq\cl S$, so $S=\int(U\cap\cl S)=\int U=U$. Thus, $\sigma:S\to W$ is already a of $\c F$ over $U$, and it is vacuously $T$-full. So we can take $\rho=\sigma$. We are done.
So assume from now on that $T\neq\emptyset$. There are three cases.
#### Case 1: $\c F=\c F^*(w_0)$ for some reflexive ${w_0}\in W$
Choose such a ${w_0}$ (it may not be unique). Then $R(w_0)=W$ and ${w_0}\in R^\circ({w_0})$ since $w_0$ is reflexive. So $R^\circ(w_0)\neq\emptyset$. Since $T$ is clearly a non-empty open set, we can use theorem \[thm:Tarski\](\[Tarski thm 2\]) to partition $T$ into non-empty open sets $G_{v^\bullet}$ (${v^\bullet}\in R^\bullet({w_0})$) and other non-empty sets $B_{v^\circ}$ $({v^\circ}\in R^\circ({w_0})$) such that for each ${v^\bullet}\in R^\bullet({w_0})$ and ${v^\circ}\in R^\circ({w_0})$ we have $$\label{e:tarski cond1}
\cl (G_{v^\bullet})\setminus G_{v^\bullet}=\did B_{v^\circ}=\cl( T)\setminus\bigcup_{v\in R^\bullet({w_0})}G_{v}=D,\mbox{ say.}$$
For each ${v^\bullet}\in R^\bullet({w_0})$, the frame $\c F^*(v^\bullet)$ is connected (as it is rooted) and locally connected $\axK\axD4$ (as it is a generated subframe of $\c F$). Since ${w_0}$ is a world of $\c F$ but not of $\c F^*(v^\bullet)$, the frame $\c F^*(v^\bullet)$ is smaller than $\c F$. By the inductive hypothesis, $\c F^*(v^\bullet)$ is fully over $X$. So, taking the subset ‘$S$’ of $G_{v^\bullet}$ to be $\emptyset$ and ‘$T$’ to be $G_{v^\bullet}\setminus\cl\emptyset=G_{v^\bullet}$, we can find a $G_{v^\bullet}$-full $\rho_{v^\bullet}$ of $\c F^*(v^\bullet)$ over $G_{v^\bullet}$.
Define $\rho:U\to W$ by: $$\rho(x)=
\begin{cases}
\rho_{v^\bullet}(x),&\mbox{if }x\in G_{v^\bullet}\mbox{ for some (unique) }{v^\bullet}\in R^\bullet({w_0}),
\\
{v^\circ},&\mbox{if }x\in B_{v^\circ}\mbox{ for some (unique) }{v^\circ}\in R^\circ({w_0}),
\\
\sigma(x),&\mbox{if }x\in S,
\\
{w_0},&\mbox{otherwise,}
\end{cases}$$ for each $x\in U$. The map $\rho$ is well defined because the $G_{v^\bullet}$, the $B_{v^\circ}$, and $S$ are pairwise disjoint, and plainly it is total and extends $\sigma$.
We aim to show that $\rho$ is a $T$-full of $\c F$ over $U$. The following claim will help.
Let $x\in D$ (see ). Then $(X,\rho^{-1}),x\models\did w$ for every $w\in W$.
Let $x\in D$ and $w\in W$ be given. There are two cases. The first is when $w\in R^\bullet({w_0})$. Now gives $x\in\cl G_w\setminus G_w$. As $\rho_w$ is a $G_w$-full of $\c F^*(w)$, a frame of which $w$ is a world, we have $(X,\rho_w^{-1}),x\models\did w$, and hence $(X,\rho^{-1}),x\models\did w$ (since $\rho_w\subseteq\rho$).
The second case is when $w\notin R^\bullet({w_0})$. Since $w\in W=R(w_0)=R^\bullet({w_0})\cup R^\circ({w_0})$, we have $w\in R^\circ({w_0})$. By , $x\in\did B_w$ (since $x\in D$). Since $\rho\restriction B_w$ has constant value $w$, we obtain again that $(X,\rho^{-1}),x\models\did w$. This proves the claim.
We now check that $\rho$ is a of $\c F$ over $U$. Let $x\in U$ and $w\in W$. We require $(X,\rho^{-1}),x\models\did w$ iff $R(\rho(x),w)$. There are four cases.
1. Suppose that $x\in G_{v^\bullet}$ for some ${v^\bullet}\in R^\bullet({w_0})$. Since $G_{v^\bullet}$ is open and $\rho\restriction G_{v^\bullet}=\rho_{v^\bullet}$, a over $G_{v^\bullet}$ of the generated subframe $\c F^*(v^\bullet)$ of $\c F$, lemma \[lem:reps simple\] yields $(X,\rho^{-1}),x\models\did w$ iff $R(\rho(x),w)$.
2. Suppose that $x\in B_{v^\circ}$ for some ${v^\circ}\in R^\circ({w_0})$. Then $\rho(x)={v^\circ}$. As ${v^\circ}\in R^\circ({w_0})$, we have $R{v^\circ}{w_0}$. By transitivity of $R$, we have $R(\rho(x),w)$ for every $w\in W$. So we need to prove that $(X,\rho^{-1}),x\models\did w$ for every $w\in W$. But $x\in B_{v^\circ}\subseteq D$ by definition of $D$ , so this follows from the claim.
3. If $x\in S$, then since $S$ is open and $\rho\restriction S=\sigma$, a of $\c F$ over $S$, the result follows from lemma \[lem:reps simple\] again.
4. Suppose finally that $x\in U\setminus(S\cup T)$. Then $\rho(x)={w_0}$. Since $R({w_0},w)$ for all $w\in W$, we require that $(X,\rho^{-1}),x\models\did w$ for all $w\in W$ as well.
Now as $S$ is a subset of $U$, by lemma \[lem:ro1\] we obtain $U\setminus S= \cl T$. Hence, $x\in \cl T\setminus T\subseteq D$ by . As in case 2, the claim now gives $(X,\rho^{-1}),x\models\did w$ for all $w\in W$.
So $\rho$ is indeed a of $\c F$ over $U$. We check that it is $T$-full. First let $x\in\cl T\setminus U$. Then $x\in \cl T\setminus T\subseteq D$ by . By the claim, $(X,\rho^{-1}),x\models\did w$ for every $w\in W$, as required.
We also need that $\rho$ is surjective. Take any $x\in B_{w_0}$. Then $x\in D$ by definition of $D$ in . By the claim, $(X,\rho^{-1}),x\models\did w$, and so $\rho^{-1}(w)\neq\emptyset$, for every $w\in W$. Hence, $\rho$ is surjective.
#### Case 2: $\c F=\c F^*(w_0)$ for some irreflexive ${w_0}\in W$
Choose such a ${w_0}$ (it is unique this time). Then $W$ is the disjoint union of $\{w_0\}$ and $R(w_0)$. Using theorem \[thm:Tarski\](\[Tarski thm 1\]), select non-empty $I\subseteq T$ with $$\label{e:I set}
\did I=\cl T\setminus U.$$ Write $$\begin{array}{rcl}
U'&=&U\setminus I,
\\
T'&=&T\setminus I.
\end{array}$$ We aim to use the inductive hypothesis on these sets and $\sigma:S\to\c F({w_0})$, so we check the necessary conditions.
1 $U'$ is open, $S$ is a subset of $U'$, and $T'=U'\setminus\cl S$.
First, $U'$ is open. For, by lemma \[lem:did cl\] and , $$U\setminus\cl I=U\setminus(I\cup\did I)=U\setminus(I\cup(\cl(T)\setminus U))=U\setminus I=U',$$ and the left-hand side is open.
We are given that $S$ is a subset of $U$. Since $S\subseteq U$ and $I\subseteq T=U\setminus\cl S$, we have $S\subseteq U\setminus I=U'$. By lemma \[lem:ro1\](3), $S$ is a subset of $U'$.
Finally, $U'\setminus\cl S=(U\setminus I)\setminus\cl S=(U\setminus\cl S)\setminus I=T\setminus I=T'$. This proves the claim.
2 $\sigma$ is a $U'$-basic of $\c F({w_0})$ over $S$.
First we show that $\sigma:S\to R({w_0})$. We know that $\sigma:S\to W=\{{w_0}\}\cup R({w_0})$. Assume for contradiction that there is some $x\in S$ with $\sigma(x)={w_0}$. Then plainly, $x\in U$ and $(X,\sigma^{-1}),x\models\di{w_0}$. As $\sigma$ is a $U$-basic of $\c F$ over $S$, we obtain $R{w_0}{w_0}$, contradicting the choice of $w_0$ as irreflexive. So indeed, $\rng\sigma\subseteq W\setminus\{{w_0}\}= R({w_0})$. Since $\sigma$ is a of $\c F$ over $S$, by lemma \[lem:reps simple\] it is also a (over $S$) of the generated subframe $\c F({w_0})$ of $\c F$. It is trivially $U'$-basic, since if $x\in U'$, $w,v\in R({w_0})$, and $(X,\sigma^{-1}),x\models\di w\wedge\di v$, then $x\in U$ and $w,v\in W$ as well, so $Rwv$ since $\sigma$ is $U$-basic. This proves the claim.
In summary, $U'$ is open, $S$ is a subset of $U'$, $\sigma$ is a $U'$-basic of $\c F({w_0})$ over $S$, and $T'=U'\setminus\cl S$.
Now $\c F({w_0})$ is smaller than $\c F$ (since ${w_0}\notin R({w_0})$), connected (since $\c F$ is locally connected), and locally connected $\axK\axD4$ (since it is a generated subframe of $\c F$). By the inductive hypothesis, $\c F({w_0})$ is fully over $X$.
So $\sigma$ extends to a $T'$-full $\rho':U'\to R({w_0})$ of $\c F({w_0})$ over $U'$. By $T'$-fullness, $$\label{e:2 cases in 1}
(X,\rho'^{-1}),x\models\did v\mbox{ for every } v\in R({w_0}) \mbox{ and }
x\in\cl T'\setminus U'.$$ We extend $\rho'$ to a map $\rho:U\to W$ by defining $$\rho(x)=
\begin{cases}
\rho'(x),&\mbox{if }x\in U',
\\
w_0,&\mbox{if }x\in I,
\end{cases}$$ for $x\in U$. This is plainly well defined and total. Since $\rho$ extends $\rho'$, it also extends $\sigma$. We will show that $\rho$ is a $T$-full of $\c F$ over $U$. To do it, we need another claim.
3 $\cl T\setminus U\subseteq\cl I\subseteq\cl T'\setminus U'$.
By and lemma \[lem:did cl\], we have $\cl T\setminus U=\did I\subseteq\cl I$.
Using openness of $T=T'\cup I$, the assumption that $X$ is dense in itself, and lemma \[lem:did cl\](\[didcl part5\],\[didcl additive\]), we have $I\subseteq T\subseteq\cl T=\did T=\did T'\cup\did I$. But by , $I\cap\did I\subseteq U\cap\cl T\setminus U=\emptyset$. So in fact, $I\subseteq\did T'\subseteq\cl T'$. Hence, $\cl I\subseteq\cl T'$. Since $I\cap U'=\emptyset$ and $U'$ is open (claim 1), we have $\cl I\cap U'=\emptyset$. So $\cl I\subseteq\cl T'\setminus U'$, proving the claim.
4 $\rho$ is a of $\c F$ over $U$.
Let $x\in U$. We require $(X,\rho^{-1}),x\models\did w$ iff $R(\rho(x),w)$, for each $w\in W$.
There are two cases here. The first is when $x\in I$. Then $\rho(x)={w_0}$, so we require first that $(X,\rho^{-1}),x\models\did w$ for each $w\in R({w_0})$. So pick any $w\in R({w_0})$. By claim 3, $x\in I\subseteq\cl I\subseteq\cl T'\setminus U'$, so by , $(X,\rho'^{-1}),x\models\did w$. As $\rho'\subseteq\rho$, the result follows.
We also require that $(X,\rho^{-1}),x\not\models\did w$ for each $w\in W\setminus R({w_0})$ — that is, $(X,\rho^{-1}),x\not\models\did {w_0}$. But as $x\in U$, we have $x\notin\cl T\setminus U=\did I$ by . Since $\rho^{-1}({w_0})=I$, we do indeed have $(X,\rho^{-1}),x\not\models\did {w_0}$.
The second case is when $x\notin I$. In this case, $x\in U'$, an open set, and $\rho\restriction U'=\rho'$, a over $U'$ of the generated subframe $\c F({w_0})$ of $\c F$. By lemma \[lem:reps simple\], $(X,\rho^{-1}),x\models\did w$ iff $R(\rho(x),w)$ for every $w\in W$, as required. The claim is proved.
5 $\rho$ is $T$-full.
Let $x\in\cl T\setminus U$ and $w\in W$. We require $(X,\rho^{-1}),x\models\did w$.
Suppose first that $w={w_0}$. By , $x\in\did I$. Since $I=\rho^{-1}({w_0})$, we obtain $(X,\rho^{-1}),x\models\did {w_0}$. Suppose instead that $w\in R({w_0})$. By claim 3, $x\in\cl T'\setminus U'$. So by , $(X,\rho'^{-1}),x\models\did w$. As $\rho'\subseteq\rho$, we obtain $(X,\rho^{-1}),x\models\did w$ as required.
We must also show that $\rho(U)=W$. Well, $I\neq\emptyset$, and it follows from claim 3 that $T'\neq\emptyset$ as well. As $\rho'$ is $T'$-full, $\rho'(U')=R({w_0})$. So $$\rho(U)=\rho'(U')\cup\rho(I)=R({w_0})\cup\{{w_0}\}=W,$$ as required. This proves the claim and completes case 2 of proposition \[prop:di rep\]. Only case 3 remains, but this is the hardest case.
#### Case 3: otherwise
As $\c F$ is finite and connected, we can choose worlds $a_0,b_0,a_1,b_1,\ldots,\allowbreak b_{n-1},\allowbreak a_{n}\in W$, for some least possible $n<\omega$, such that $Ra_ib_i$ and $Ra_{i+1}b_i$ for each $i<n$, each $b_i$ is $R$-maximal (so that $R^\bullet(b_i)=\emptyset$), and $W=\bigcup_{i\leq n}R^*(a_i)$. By the case assumption, $n\geq1$.
Write $\c F^*({a_0})$ as $\c F_0=(W_0,R_0)$, say. Let $\c F_1=(W_1,R_1)$ be the smallest generated subframe of $\c F$ containing $a_1,\ldots,a_{n}$. We have $W_{0}\cup W_1=W$ and $b_0\in W_0\cap W_1$. Plainly, $\c F_0$ and $\c F_1$ are connected generated subframes of $\c F$. Therefore, they are locally connected $\axK\axD4$ frames. By minimality of $n$, they are proper subframes of $\c F$. By the inductive hypothesis, $\c F_0$ and $\c F_1$ are fully over $X$. Our plan is to combine suitable s of them to give a of $\c F$ over $U$.
Recall that $S$ is a subset of $U$ and $\sigma:S\to W$ is a $U$-basic of $\c F$. We use $W_0,W_1$ to split $S$ (and, later, $\sigma$) in two. Let $$\begin{array}{rcl}
S_0&=&\sigma^{-1}(W_0) =\{x\in S:\sigma(x)\in W_0\},
\\
S_1&=&S\setminus S_0.
\end{array}$$ So $\sigma(S_0)\subseteq W_0$ and $\sigma(S_1)\subseteq W\setminus W_0\subseteq W_1$. Also, $S_0=S\setminus S_1$.
1 $S_0$ and $S_1$ are subsets of $U$, and $U\cap\cl(S_0)\cap\cl(S_1)=\emptyset$.
We prove the last point first. Suppose for contradiction that there is some $x\in U\cap\cl(S_0)\cap\cl(S_1)$. As $x\in\cl S_0$, we have $(X,\sigma^{-1}),x\models\di\bigvee_{w\in W_0}w$. As $\di$ is additive, it follows that there is some $w_0\in W_0$ such that $(X,\sigma^{-1}),x\models\di w_0$. Similarly, as $x\in\cl S_1$ and $\sigma(S_1)\subseteq W\setminus W_0$, there is some $w_1\in W\setminus W_0$ with $(X,\sigma^{-1}),x\models\di w_1$. As $\sigma$ is a $U$-basic , we obtain $Rw_0w_1$. Since $\c F_0$ is a generated subframe of $\c F$, this implies that $w_1\in W_0$, a contradiction. So $U\cap\cl(S_0)\cap\cl(S_1)=\emptyset$ as required.
Now let $i<2$. We show that $S_i$ is in $U$. First note that $S_i$ is open. To see this, observe that $$\begin{array}{rcll}
S_i&\subseteq& S\cap U\cap\cl S_i
&\mbox{as }S_i\subseteq S\subseteq U\mbox{ by definition and assumption}
\\
&\subseteq& S\cap U\setminus\cl S_{1-i}
&\mbox{by the first part}
\\
&=& S\setminus\cl S_{1-i}
&\mbox{as }S\subseteq U\mbox{ by assumption}
\\
&\subseteq&
S\setminus S_{1-i}
&\mbox{as }S_{1-i}\subseteq\cl S_{1-i}
\\
&=&S_i
&\mbox{by definition of }S_i.
\end{array}$$ Hence, $S_i=S\setminus\cl S_{1-i}$, an open set.
It follows that $\cl(S_i)\cap S_{1-i}=\emptyset$, so $S_i\subseteq S\cap\cl S_i\subseteq S\setminus S_{1-i}=S_i$. Thus, $S\cap\cl S_i=S_i$, and so $\int(S\cap\cl S_i)=\int S_i=S_i$ as $S_i$ is open. So $S_i$ is in $S$, and as $S$ is in $U$, lemma \[lem:ro1\](\[lem:ro:part4\]) yields that $S_i$ is in $U$. The claim is proved.
The claim and the assumption at the outset that $T\neq\emptyset$ are more than enough to apply corollary \[cor:Tarski++opens\], to obtain open subsets $U_i,T_i$ of $U$, for $i=0,1$, satisfying the following conditions:
1. \[ROsets cond1\] $U\cap\cl U_0\cap\cl U_1=\emptyset$,
2. \[ROsets cond2\] $U\cap\cl S_i\subseteq U_i$,
3. \[ROsets cond3\] $T_i=U_i\setminus\cl S_i\neq\emptyset$,
4. \[ROsets cond4\] $\cl(T)\setminus U\subseteq\cl(T_i)$,
5. \[point:Ui ro\] $U_i$ is a subset of $U$.
We now work in the $RO(U)$ of subsets of $U$. By \[point:Ui ro\], we have $U_0,U_1\in RO(U)$. We define further elements of $RO(U)$:
1. $M=-(U_0+U_1)$,
2. $V_i=M+U_i$ for $i=0,1$.
The main property of these sets is as follows.
2 $\{M,S_0,S_1,T_0,T_1\}$ is a partition of $1$ in the $RO(U)$. That is, the five elements are pairwise disjoint subsets of $U$, with $$\label{partition of U}
U=
\setbox0\hbox{$\overbrace{S_0+T_0\strut}^{U_0}+\overbrace{M+S_1+T_1\strut}^{V_1}$}\wd0=0pt\box0\hskip-2pt
\underbrace{\phantom{\strut S_0+T_0+M}}_{V_0}\phantom{+}\underbrace{\phantom{S_1+T_1\strut}}_{U_1}.$$
Let $i<2$. By claim 1 and condition \[point:Ui ro\] above, $S_i,U_i\in RO(U)$. By this and condition \[ROsets cond3\], $$\label{e:char Ti}
T_i=U_i\setminus\cl S_i=U_i\cap U\setminus\cl S_i
=U_i\cdot-S_i\in RO(U).$$ So $S_i\cdot T_i=\emptyset$ and, since $S_i\subseteq U_i$ by condition \[ROsets cond2\], also $U_i=U_i\cdot S_i+U_i\cdot-S_i=S_i+T_i$. Condition \[ROsets cond1\] above gives $U_0\cdot U_1=\emptyset$. By definition, $M=-(U_0+U_1)$, so $M\in RO(U)$ and $M$ is disjoint from $T_i,S_i$. Also, $U=U_0+U_1+M=S_0+T_0+S_1+T_1+M$. It is now plain that $M+S_i+T_i=M+U_i=V_i$. This proves the claim.
We aim to apply the inductive hypothesis to $V_i,M+S_i,T_i,\c F_i$, for each $i=0,1$. We will need a $V_i$-basic of $\c F_i$ over $M+S_i$, and the next claim helps us get one.
3 For each $i<2$ we have $U\cap \cl M\cap \cl S_i=\emptyset$, and $M+S_i=M\cup S_i$ in $RO(U)$.
By definition, $M=-(U_0+U_1)=U\setminus\cl(U_0+U_1)\subseteq U\setminus U_i$. Since $U_i$ is open, $\cl M\cap U_i=\emptyset$. But $U\cap\cl S_i\subseteq U_i$ by condition \[ROsets cond2\] above, so $U\cap \cl M\cap \cl S_i=\emptyset$. By lemma \[lem:ro1\], $M+S_i=M\cup S_i$. This proves the claim.
So all we need is to find suitable s over $M$ and $S_i$ and take their union.
Clearly, $\c F^*(b_0)$ is a subframe of $\c F_0$, and so a *proper* subframe of $\c F$. It is obviously connected (since rooted), and a generated subframe of $\c F$, so a locally connected $\axK\axD4$ frame. By the inductive hypothesis, it is fully over $X$. So we can find an ($M$-full) $\mu:M\to R(b_0)$ of $\c F^*(b_0)$ over $M$.
For each $i<2$ let $$\sigma_i=(\sigma\restriction S_i)
\;:\;S_i\to W_i.$$
4 For each $i<2$, $\mu\cup\sigma_i: M\cup S_i\to W_i$ is a well defined $V_i$-basic of $\c F_i$ over $M\cup S_i$.
Since $\c F^*(b_0)$ is a generated subframe of $\c F_i$, it follows from lemma \[lem:reps simple\](\[lem:simple:1\]) that $\mu$ is a of $\c F_i$ over $M$. Similarly, $\sigma_i$ is a of $\c F_i$ over $S_i$. Since $M$ and $S_i$ are disjoint open sets, $\mu\cup\sigma_i:M\cup S_i\to W_i$ is well defined and, by lemma \[lem:reps simple\](\[lem:simple:2\]), a of $\c F_i$ over $M\cup S_i$.
To prove that it is $V_i$-basic, let $x\in V_i$ and $v,w\in W_i$ be given, and suppose that $(X,(\mu\cup\sigma_i)^{-1}),x\models\di w\wedge\di v$. We require $Rwv$.
Plainly, $x\in\cl(M\cup S_i)=\cl M\cup\cl S_i$, and $x\in V_i\subseteq U$. But $U\cap\cl M\cap\cl S_i=\emptyset$ by claim 3. So there are two possibilities.
The first one is that $x\notin\cl M$. In this case, we must have $(X,\sigma_i^{-1}),x\models\di w\wedge\di v$. As $\sigma_i\subseteq\sigma$, we also have $(X,\sigma^{-1}),x\models\di w\wedge\di v$. As $\sigma$ is $U$-basic, we obtain $Rwv$.
The other possibility is that $x\notin\cl S_i$. So $(X,\mu^{-1}),x\models\di w\wedge\di v$. Since $\mu$ is a of $\c F^*(b_0)$, we have $w,v\in R(b_0)$. But $b_0$ is $R$-maximal, so $R^\bullet(b_0)=\emptyset$. Hence, $w\in R^\circ(b_0)$, so $Rwb_0$, and since $Rb_0v$, we deduce $Rwv$ by transitivity. (Essentially we are using that $\c F^*(b_0)$ is a non-degenerate cluster.) This proves the claim.
In summary, for each $i<2$ we have:
- $V_i$ is open (by claim 2)
- $M+S_i,V_i\in RO(U)$ and $M+S_i\subseteq V_i$, so by lemma \[lem:ro1\], $M+S_i$ is a subset of $V_i$
- working in $RO(U)$, we have $V_i=(M+S_i)+T_i$ and $(M_i+S_i)\cdot T_i=\emptyset$ by claim 2. So $T_i=V_i\cdot-(M+S_i)=V_i\cap U\setminus\cl(M+S_i)=V_i\setminus\cl(M+S_i)$.
- $M+S_i=M\cup S_i$ (by claim 3), and $\mu\cup\sigma_i:M\cup S_i\to W_i$ is a $V_i$-basic of $\c F_i$ over $M+S_i$ (by claim 4)
So for each $i<2$, recalling that $\c F_i$ is fully , we see that $\mu\cup\sigma_i:M\cup S_i\to W_i$ extends to a $T_i$-full $\rho_i:V_i\to W_i$ of $\c F_i$ over $V_i$. We have $$\label{e:T_ifull gives}
(X,\rho_i^{-1}),x\models\did w\quad\mbox{for every }
w\in W_i\mbox{ and }x\in \cl T_i\setminus V_i.$$
Finally define $$\rho=\rho_0\cup\rho_1:U\to W.$$ We check first that $\rho$ is well defined and total. Working in $RO(U)$ again, we have $\dom\rho_0\cap\dom\rho_1=V_0\cap V_1=V_0\cdot V_1=M$ by . But $\rho_0\restriction M=\mu=\rho_1\restriction M$. So $\rho$ is well defined. Also, $V_i=-U_{1-i}=U\setminus\cl U_{1-i}$ (for $i=0,1$) by , and $U\cap\cl U_0\cap\cl U_1=\emptyset$ by condition \[ROsets cond1\] above, so $$\label{D0+D1=U}
\dom\rho=V_0\cup V_1=(U\setminus\cl U_1)\cup(U\setminus\cl U_0)=U\setminus(\cl U_1\cap\cl U_0)=U.$$ Hence, $\rho$ is total. Plainly, $\rho$ extends $\sigma$, since $\rho=\rho_0\cup\rho_1
\supseteq(\mu\cup\sigma_0)\cup(\mu\cup\sigma_1)=\mu\cup\sigma$.
5 $\rho$ is a of $\c F$ over $U$.
Let $i<2$. Then $\rho\restriction V_i=\rho_i$, a of $\c F_{i}$ over $V_i$. By lemma \[lem:reps simple\](\[lem:simple:1\]), this is also a of $\c F$ over $V_i$, which is an open set by claim 2. By , $U=V_0\cup V_1$, so by lemma \[lem:reps simple\](\[lem:simple:2\]), $\rho$ is a of $\c F$ over $U$, proving the claim.
6 $\rho$ is $T$-full.
Let $x\in\cl T\setminus U$. We require $(X,\rho^{-1}),x\models\did w$ for every $w\in W$.
For each $i<2$, as $\cl T\setminus U\subseteq\cl T_i$ by condition \[ROsets cond4\] above, and $x\notin U\supseteq V_i$, we have $x\in\cl T_i\setminus V_i$. Since $\rho_i\subseteq\rho$, it follows from that $(X,\rho^{-1}),x\models\did w$ for every $w\in W_i$. This holds for each $i=0,1$. Since $W_0\cup W_1=W$, we have $(X,\rho^{-1}),x\models\did w$ for every $w\in W$.
Finally, we show that $\rho(U)=W$. Since each $\rho_i$ is a $T_i$-full of $\c F_i$ over $V_i$, and $T_i\neq\emptyset$ by condition \[ROsets cond3\], by we obtain $\rho(U)=\rho(V_0)\cup\rho(V_1)=\rho_0(V_0)\cup\rho_1(V_1)=W_0\cup W_1=W$. This proves the claim, and with it, proposition \[prop:di rep\].
We end with some technical remarks on the definition of ‘fully ’ (definition \[def:fully rb did case\]) and its relation to the proof just completed. They are not needed later, and the reader can of course skip them if desired.
It is very helpful throughout the proof that $U$ is open — see, e.g., lemma \[lem:reps simple\]. However, we cannot assume in definition \[def:fully rb did case\] that $U$ is in $X$. For if we did, then in case 2 of the proof, we have $\cl I\subseteq\cl T'\subseteq\cl U'$ by claim 3 and $T'\subseteq U'$, so $U'\neq U=\int\cl U=\int(\cl U'\cup\cl I)=\int\cl U'$. Therefore, $U'$ is not in $X$, and we can not apply the inductive hypothesis to it. We use that $X$ is dense in itself to show that $I\subseteq\cl T'$.
At least according to the construction we gave, $S$ should be open. In case 1, if $S$ is not open then there is $x\in S\setminus\int S\subseteq \cl (U\setminus S)$, and a little thought shows that $(X,\rho^{-1}),x\models\did {w_0}$ for any such $x$. For $\rho$ to be a , we would need $R(\rho(x),w_0)$. Since $\rho\supseteq\sigma$ and $x\in S$, this says that $R(\sigma(x),w_0)$, which we have no reason to suppose is true.
The problem if $S$ is not in $U$ is that, again in case 1, we used that $U\setminus S=\cl T$. If this were to fail, there may be points $x\in U\setminus (S\cup\cl T)$ (so $x\in U\cap\int\cl S$). We have to define $\rho$ on these $x$, and defining $\rho(x)={w_0}$ as in the proof may not give a . However, as $\sigma$ is $U$-basic, it is possible to define $\rho(x)$ using $\sigma$ instead. This effectively extends $\sigma$ to $U\cap\int\cl S$. So we can assume without loss of generality that $S$ is in $U$. It is therefore easier to do so and avoid the problem completely.
We could just suppose in definition \[def:fully rb did case\] that $S$ is in $X$, but we cannot suppose this of $U$, and we have to work in $RO(U)$, so there is little gain in doing so.
We need that $\sigma$ is $U$-basic in order that in case 3, the subsets $S_0,S_1$ have disjoint closures in $U$. This in turn is needed to apply normality in the proof of corollary \[cor:Tarski++opens\].
We cannot assume instead in definition \[def:fully rb did case\] that $\sigma$ is $X$-basic, because in case 3, we cannot guarantee that $\mu\cup\sigma_i$ is $X$-basic. This is because we do not know that $M\cap\cl S_i=\emptyset$, but only that $U\cap M\cap\cl S_i=\emptyset$. We could solve this problem by assuming further that $\cl S\subseteq U$ (which implies that $S$ is in $X$), but this weakens the proposition sufficiently to cause trouble in theorem \[thm:strong did\] later, where we would need to ensure that $\cl S_n\cup\cl S_{n+1}\subseteq U_n$ for each $n$.
Finally, we mention that actually $\rho(T)=W$ when $T\neq\emptyset$ — not only $\rho$ but also $\rho\restriction T$ is surjective. We might try to drop the second, surjectivity part of definition \[def:Tfull rep\] and simply prove it from the first part, as in cases 1 and 2 of the proof, but it is not clear how to do this in case 3.
Weak completeness
=================
We are now ready to prove our first tranche of main results, showing that Hilbert systems for various sublanguages of $\Lbig$ are sometimes sound and always complete over any non-empty space. Several of the proofs use the translations $-^d$ and $-^\mu$ of section \[sec:translations\]. We establish only weak completeness. We will discuss strong completeness later, in section \[sec:strong c\].
Here and later, we include ‘$t$’ in the name of a Hilbert system to indicate that it includes the tangle axioms [**Fix**]{} and [**Ind**]{} of section \[ss:tangle logics\]. Recall that by lemma \[lem:ms\], metric spaces, regarded as topological spaces, are Huasdorff and hence T$_D$.
Weak completeness for $\c L_\bo^\mu$ and $\c L_\bo^\dit$ {#ss:weak comp bo case}
--------------------------------------------------------
The pioneering result in this field was the theorem of [@McKiT44] that the $\c L_\bo$-logic of every separable space is S4. The assumption of separability was removed in [@RS:mm]. We begin by generalising this theorem, establishing (weak) completeness results for $\c L_\bo^\mu$ and $\c L_\bo^\dit$ over any space. We will go on to prove strong completeness in theorem \[thm:str compl boxes\].
\[thm:compl S4mu, S4t\] Let $X$ be a non-empty space.
1. The Hilbert system $\axS4\mu$ is sound and complete over $X$ for $\c L_\bo^\mu$-formulas.
2. The Hilbert system $\axS4t$ is sound and complete over $X$ for $\c L_\bo^\dit$-formulas.
For part 1, soundness is easy to check and indeed we have already mentioned it in corollary \[cor:t transl top equiv\]. For completeness, let $\varphi$ be an $\c L_\bo^\mu$-formula that is not a theorem of $\logicbomu$. By theorem \[thm:S4mu sc\], we can find a finite S4 frame $\c F=(W,R)$, an assignment $h$ into $\c F$, and a world $w\in W$ with $(W,R,h),w\models\neg\varphi$. By replacing $\c F$ by $\c F(w)$, we can suppose that $w$ is a root of $\c F$ — this can be justified in a standard way using lemma \[lem:gen submodels\]. Since $\c F$ is rooted, it is clearly connected. Since it is reflexive and transitive, it is a locally connected $\axK\axD4$ frame. So by proposition \[prop:di rep\], $\c F$ is fully over $X$. So, taking $U=X$ and $S=\sigma=\emptyset$ in the definition of ‘fully ’ (definition \[def:fully rb did case\]), we may choose an $X$-full, hence surjective, $\rho$ of $\c F$ over $X$. Choose $x\in X$ with $\rho(x)=w$. Then $$\begin{array}{rcll}
(W,R,h),w\models\varphi
&\mbox{iff}&
(W,R,h),w\models\varphi^d&\mbox{by lemma~\ref{lem:-d in refl frames},
since }\c F\mbox{ is reflexive,}
\\
&\mbox{iff}&(X,\rho^{-1}\circ h),x\models\varphi^d
&\mbox{by proposition~\ref{prop:fmla pres}, since $\varphi^d\in\c L_{\bod\forall}^\mu$,}
\\
&\mbox{iff}&(X,\rho^{-1}\circ h),x\models\varphi
&\mbox{by lemma~\ref{lem:trans equivalent top}, since $X$ is T$_D$.}
\end{array}$$ We obtain $(X,\rho^{-1}\circ h),x\models\neg\varphi$. Thus, $\varphi$ is not valid over $X$, proving completeness.
The proof of part 2 is similar. For the soundness of the tangle axioms see [@FD:apal12 Theorem 6.1]. For completeness, the differences are: $\varphi$ is assumed to be an $\c L_\bo^\dit$-formula that is not a theorem of $\axS4t$; we use the results of section \[sec:fmp S4\] in place of theorem \[thm:S4mu sc\] to obtain a finite S4 Kripke model satisfying $\neg\varphi$ at a root; and having obtained a surjective $\rho$ of $\c F$ over $X$ and $x\in X$ with $\rho(x)=w$, we use the additional translation $-^\mu$ from section \[sec:translations\], as follows. Note that $\varphi\in\c L_\bo^\dit$, $\varphi^d\in\c L_\bod^\didt$, and $(\varphi^d)^\mu\in \c L_\bod^\mu\subseteq\c L_{\bod\forall}^\mu$. $$\begin{array}{rcll}
(W,R,h),w\models\varphi
&\mbox{iff}&
(W,R,h),w\models\varphi^d&\mbox{by lemma~\ref{lem:-d in refl frames},
since }\c F\mbox{ is reflexive,}
\\
&\mbox{iff}&(W,R,h),w\models(\varphi^d)^\mu
&\mbox{by lemma~\ref{lem:mu trans}, since $\c F$ is transitive,}
\\
&\mbox{iff}&(X,\rho^{-1}\circ h),x\models(\varphi^d)^\mu
&\mbox{by proposition~\ref{prop:fmla pres},
since $(\varphi^d)^\mu\in\c L_{\bod\forall}^\mu$,}
\\
&\mbox{iff}&(X,\rho^{-1}\circ h),x\models\varphi^d
&\mbox{by lemma~\ref{lem:mu trans} again,}
\\
&\mbox{iff}&(X,\rho^{-1}\circ h),x\models\varphi
&\mbox{by lemma~\ref{lem:trans equivalent top}, since $X$ is T$_D$.}
\end{array}$$
Weak completeness for $\c L_{\bo\forall}$ and $\c L_{\bo\forall}^\dit$ {#ss:weak comp bo forall}
----------------------------------------------------------------------
Completeness for languages with $\forall$ follows the same lines, although soundness requires that the space be connected.
\[thm:compl S4.UC, S4t.UC\] Let $X$ be a non-empty space.
1. The Hilbert system $\axS4.\axU\axC$ is complete over $X$ for $\c L_{\bo\forall}$-formulas, and sound if $X$ is connected.[^5]
2. The Hilbert system $\axS4t.\axU\axC$ is complete over $X$ for $\c L_{\bo\forall}^\dit$-formulas, and sound if $X$ is connected.
For part 1, soundness when $X$ is connected is again clear: connectedness is needed so that the $C$ axiom is valid in $X$. For completeness, even when $X$ is not connected, suppose that $\varphi\in\c L_{\bo\forall}$ is not a theorem of $\axS4.\axU\axC$. By the results of section \[sec:path conn\], or by [@Sheh:everywhere99 theorem 10], $\axS4.\axU\axC$ has the finite model property, so we can find a finite connected S4 frame $\c F=(W,R)$, an assignment $h$ into $\c F$, and a world $w\in W$ such that $(W,R,h),w\models\neg\varphi$. The proof that $\varphi$ is not valid in $X$ is now exactly as in theorem \[thm:compl S4mu, S4t\].
Part 2 is proved similarly, using the results of section \[sec:path conn\] to obtain a finite model.
We have no results for $\c L_{\bo\forall}^\mu$ because we are not aware of any completeness theorem for this language with respect to finite connected S4 frames. If one is proved in future, we could take advantage of it.
Weak completeness for $\c L_{\bod}$ and $\c L_{\bod}^\didt$ {#ss:weak comp bod and tangle}
-----------------------------------------------------------
In one way this is even easier, as we do not need the translation $\varphi^d$. But again, soundness requires a condition on the space.
\[thm:compl KD4G1(t)\] Let $X$ be a non-empty space.
1. The Hilbert system $\axK\axD4\axG_1$ is complete over $X$ for $\c L_{\bod}$-formulas, and sound if $\axG_1$ is valid in $X$.
2. The Hilbert system $\axK\axD4\axG_1t$ is complete over $X$ for $\c L_{\bod}^\didt$-formulas, and sound if $\axG_1$ is valid in $X$.
For part 1, soundness is clear. For completeness, even when $X$ does not validate $\axG_1$, suppose that $\varphi\in\c L_\bod$ is not a theorem of $\axK\axD4\axG_1$. As we mentioned in section \[ss:weak models\], $\axK\axD4\axG_1$ has the finite model property [@Sheh:d90 theorem 15], so we can find a finite $\axK\axD4\axG_1$ frame $\c F=(W,R)$, an assignment $h$ into $\c F$, and a world $w\in W$ such that $(W,R,h),w\models\neg\varphi$. As usual, by replacing $\c F$ by $\c F(w)$, we can suppose that $\c F$ is connected. It is also locally connected because it validates $\axG_1$ (see fact \[fact:Gn = loc nconn\]). Using proposition \[prop:di rep\], let $\rho$ be a surjective of $\c F$ over $X$. Let $x\in X$ satisfy $\rho(x)=w$. Then $(X,\rho^{-1}\circ h),x\models\neg\varphi$ by proposition \[prop:fmla pres\]. So $\varphi$ is not valid in $X$.
The proof of part 2 is similar, except that we use the results of section \[sec:fmp Gn\] to obtain a finite model, and in order to apply proposition \[prop:fmla pres\], we first use the translation $-^\mu$ to turn $\varphi\in\c L_\bod^\didt$ into an $\c L_\bod^\mu$-formula $\varphi^\mu$ equivalent to $\varphi$ in transitive frames and in $X$.
\[rmk:shehtman\]Theorem \[thm:compl KD4G1(t)\](1) is related to earlier work of Shehtman [@Sheh:d90]. In [@Sheh:d90 theorem 23, p.39], the following is proved for the language $\c L_\bod$:
> 1. Let $X$ be a topological space having an open set homeomorphic to some $\R^n$, $n>0$. Then $L(D(X))\subseteq D4\axG_1$ \[the $\c L_\bod$-logic of $X$ is contained in $\axK\axD4\axG_1$\].
>
> 2. If additionally $X$ satisfies conditions of lemma 2 then $L(D(X))= D4\axG_1$.
>
Lemma 2 [@Sheh:d90 p.3] states the following.
> Let $X$ be a topological space satisfying the following condition: for any open $U$ and any $x\in U$ there is open $V\subseteq U$ such that $x\in V$ and $(V\setminus\{x\})$ is connected \[as a subspace of $X$\]. Then $X\models \axG_1$.
Shehtman’s results (i), (ii) above follow from theorem \[thm:compl KD4G1(t)\](1). We remark that the converse of his lemma 2 fails in general — a counterexample is given by the subspace $X=
\R^2\setminus\{(1/n,y):n\mbox{ a positive integer,}\;y\in\R\}
$ of $\R^2$. [@LucBry11 theorems 3.12, 3.14] give a characterisation of when a topological space validates $\axG_n$, for $n\geq1$.
Shehtman [@Sheh:d90 p.43] also states two open problems:
1. To describe all \[$\c L_\bod$-\]logics \[of\] dense-in-itself metric spaces $X$. In particular, is $[K]D4\axG_1$ the greatest of them?
2. Is theorem 23(ii) extended to the infinite dimensional case? In particular, does it hold for Hilbert space $\ell_2$ (with the weak or with the strong topology)?
Theorem \[thm:compl KD4G1(t)\](1) appears to resolve problem 2 and the second part of problem 1, both positively.
Shehtman also proved in [@Sheh:d90 theorem 29] that the $\c L_\bod$-logic of every zero-dimensional separable space is $\axK\axD4$. This does not follow from theorem \[thm:compl KD4G1(t)\].
Weak completeness for $\c L_{\bod\forall}$ and $\c L_{\bod\forall}^\didt$ {#ss:weak comp bod forall and tangle}
-------------------------------------------------------------------------
The following is now purely routine.
\[thm:compl KD4G1(t).UC\] Let $X$ be a non-empty space.
1. The Hilbert system $\axK\axD4\axG_1.\axU\axC$ is complete over $X$ for $\c L_{\bod\forall}$-formulas, and sound if $X$ is connected and validates $\axG_1$.
2. The Hilbert system $\axK\axD4\axG_1t.\axU\axC$ is complete over $X$ for $\c L_{\bod\forall}^\didt$-formulas, and sound if $X$ is connected and validates $\axG_1$.
The finite model property for $\axK\axD4\axG_1.\axU\axC$ and $\axK\axD4\axG_1t.\axU\axC$ follows from the results of section \[sec:fmp Gn\]. There are no other new elements in the proof, so we leave it to the reader.
Strong completeness {#sec:strong compl}
===================
Here, we will prove that $\axK\axD4\axG_1t$ is strongly complete over any non-empty space $X$: any countable $\axK\axD4\axG_1t$-consistent set of $\c L_\bod^\didt$-formulas is satisfiable over $X$. The analogous results for $\c L_\bo^\mu$ and the weaker languages $\c L_\bod$ and $\c L_\bo^\dit$ will follow. The analogous result for $\c L_\bo$ also follows, but this is a known result, proved recently by Kremer [@Kremer2010:stro]. We will then show that strong completeness frequently fails for languages with $\forall$.
The problem
-----------
Let us outline a naïve approach to the problem. It does not work, but it will illustrate the difficulty we face and motivate the formal proof later.
Let $\Gamma$ be a countable $\axK\axD4\axG_1t$-consistent set of $\c L_\bod^\didt$-formulas. For simplicity, assume that $\Gamma$ is maximal consistent. Write $\Gamma$ as the union of an increasing chain $\Gamma_0\subseteq\Gamma_1\subseteq\cdots$ of finite sets. Fix $x\in X$. By weak completeness (theorem \[thm:compl KD4G1(t)\]), each $\Gamma_n$ ($n<\omega$) is satisfiable at $x$, so we can find an assignment $g_n$ on $X$ with $(X,g_n),x\models\Gamma_n$. Suppose we could build a new assignment $g$ that *behaves like $g_n$ for larger and larger $n$,* as we approach $x$. Then we might hope that $(X,g),x\models\Gamma_n$ for all $n$, and so $(X,g),x\models\Gamma$.
To define such a $g$, we choose a countable sequence $X=S_0\supseteq S_1\supseteq\cdots$ of open s of $x$, such that
1. every open of $x$ contains some $S_n$ (that is, the $S_n$ form a ‘base of open s’ of $x$).
$X$ is a metric space, so we can do this. Since we can make the $S_n$ as small as we like, and the $\Gamma_n$ are finite sets, we can suppose that for each $n<\omega$:
1. \[naive suppose 1\] for each $\bod\varphi\in\Gamma_n$, we have $(X,g_n),y\models\varphi$ for every $y\in S_n\setminus\{x\}$,
2. \[naive suppose 2\] for each $\did\varphi\in\Gamma_n$, there is $y\in S_n\setminus\cl S_{n+1}$ with $(X,g_n),y\models\varphi$.
We can now define a new assignment $g$ by ‘using $g_n$ within $S_n$’, for each $n<\omega$. More precisely, we let $$g(p)\cap(S_n\setminus S_{n+1})=g_n(p)\cap(S_n\setminus S_{n+1})$$ for each atom $p$ and each $n<\omega$. We also need to define $g$ at $x$ itself, but we can use $\Gamma$ to determine truth values of atoms there.
Now we try to prove that $\varphi\in\Gamma$ iff $(X,g),x\models\varphi$ for all formulas $\varphi$, by induction on $\varphi$. The atomic and boolean cases are easy. Consider the case $\did\varphi$.
If $\did\varphi\in\Gamma$, then $\did\varphi\in\Gamma_n$ for all large enough $n$, so by \[naive suppose 2\], there is $y\in S_n\setminus\cl S_{n+1}$ with $(X,g_n),y\models\varphi$. As $S_n\setminus\cl S_{n+1}$ is open and $g_n$ agrees with $g$ on it, it follows that $(X,g),y\models\varphi$. This holds for cofinitely many $n$, so $(X,g),x\models\did\varphi$.
Conversely, if $(X,g),x\models\did\varphi$, then for infinitely many $n$, there is $y\in S_n\setminus S_{n+1}$ with $(X,g),y\models\varphi$. *If we could find such a $y\in S_n\setminus\cl S_{n+1}$,* then as above, $(X,g_n),y\models\varphi$, and it would follow by \[naive suppose 1\] and maximality of $\Gamma$ that $\did\varphi\in\Gamma$.
But it may be that we can only find such $y\in\cl S_{n+1}$. The truth of $\varphi$ at such $y$ may not be preserved when we change from $g$ to $g_n$, because it may depend on points in $S_{n+1}$, and at such points, $g$ agrees with $g_{n+1}$, not $g_n$. (We cannot just make $S_{n+1}$ smaller to take the witnesses $y$ out of $\cl S_{n+1}$, because $g$ will then change, and we may no longer have $(X,g),y\models\varphi$.)
So we would like to arrange a *smooth transition* between $g_n$ and $g_{n+1}$, avoiding unpleasant discontinuities. It would be sufficient if there is some closed $T_{n+1}\subseteq S_{n+1}$ such that $g_n$ and $g_{n+1}$ agree on the ‘buffer zone’ $S_{n+1}\setminus T_{n+1}$. Much of the formal proof below is aimed at achieving something like this for atoms occurring in $\Gamma_n$ — see claim 3 especially.
Strong completeness for $\c L_\bod^\didt$
-----------------------------------------
\[thm:strong did\] Let $X$ be a non-empty space. Then the Hilbert system $\axK\axD4\axG_1t$ is strongly complete over $X$ for $\c L_{\bod}^\didt$-formulas, and sound if $\axG_1$ is valid in $X$.
For soundness, see theorem \[thm:compl KD4G1(t)\]. For strong completeness, let $\Gamma$ be a countable $\axK\axD4\axG_1t$-consistent set of $\c L_\bod^\didt$-formulas. We show that $\Gamma$ is satisfiable over $X$. We can suppose without loss of generality that $\Gamma$ is maximal consistent. Since $\Gamma$ is countable, we can write it as $\Gamma=\bigcup_{n<\omega}\Gamma_n$, where $\Gamma_0\subseteq\Gamma_1\subseteq\cdots$ is a chain of finite sets. Let $L_n$ be the finite set of atoms occurring in formulas in $\Gamma_n$, for each $n<\omega$. So $L_0\subseteq L_1\subseteq\cdots$. For each $n<\omega$, as $\Gamma_n$ is $\axK\axD4\axG_1t$-consistent, by the results of section \[sec:fmp Gn\] there is a finite Kripke model $\c M_n=(W_n,R_n,h_n)$ whose frame $(W_n,R_n)$ validates $\axK\axD4\axG_1$, and a world $w_n\in W_n$ with $$\c M_n,w_n\models\Gamma_n.$$ We can assume without loss of generality that the $W_n$ $(n<\omega)$ are pairwise disjoint. For each $n$, fix an arbitrary $e_n\in W_n$ with $R_nw_ne_n$ and such that $e_n$ is $R_n$-maximal — that is, $R^\bullet_n(e_n)=\emptyset$.
For $i\leq j<\omega$ and $w\in W_j$ write $$\begin{array}{rcll}
\tp_i(w)&=&\{p\in L_i:\c M_j,w\models p\}&\in\wp L_i
\\
\tau{}^j_i&=&\{\tp_i(w):w\in R_j(e_j)\}&\in\wp\wp L_i
\end{array}$$ So $\tp_i(w)$ is the ‘atomic type’ of $w$ in $\c M_j$ with respect to the finite set $L_i$ of atoms. We do not need to write $\tp_i^j(w)$ since the $W_n$ are pairwise disjoint so $j$ is determined by $w$. And $\tau{}^j_i$ is the set of such types that occur as types of points in the cluster $R_j(e_j)$.
1 We can suppose without loss of generality that $\tau{}^j_i=\tau{}^i_i$ whenever $i\leq j<\omega$.
Essentially König’s tree lemma. We will define by induction infinite sets $\omega=I_{-1}\supseteq I_0\supseteq I_1\supseteq\cdots$. We let $i_n=\min I_n$, and we will arrange that $0=i_{-1}<i_0<i_1<\cdots$ and $i_n\geq n$ for all $n$. Let $n<\omega$ and suppose that we are given $I_{n-1}$ and $i_{n-1}=\min I_{n-1}\geq n-1$ inductively. Using that $\wp\wp L_{n} $ is finite, choose infinite $I_{n}\subseteq I_{n-1}\setminus\{i_{n-1}\}$ such that $\tau{}^i_n\in\wp\wp L_n$ is constant for all $i\in I_{n}$. The term $\tau{}^i_n$ is defined for all $i\in I_n$, because $i\geq\min I_n>i_{n-1}\geq n-1$ and so $i\geq n$. Of course define $i_n=\min I_n$. Then $i_n>i_{n-1}$ and $i_n\geq n$ as required. This completes the definition. Now replace $\c M_n,w_n,e_n$ by $\c M_{i_n},w_{i_n},e_{i_n}$ for each $n<\omega$. Do not change $\Gamma_n$ or $L_n$. Since $n\leq i_n$, we have $\Gamma_n\subseteq\Gamma_{i_n}$, and consequently we still have $\c M_n,w_n\models\Gamma_n$ for each $n$. And if $r\leq s<\omega$ we have $i_r,i_s\in I_r$, so $\tau{}^{i_r}_r=\tau{}^{i_s}_r$, and consequently after replacement, $\tau{}^r_r=\tau{}^{s}_r$. This proves the claim.
For each $n<\omega$, define the frames $$\begin{array}{rcl}
\c F_n&=&(R_n(w_n),R_n\restriction R_n(w_n)),
\\
\c C_n&=&(R_n(e_n),R_n\restriction R_n(e_n)).
\end{array}$$ $\c F_n$ is a generated subframe of $(W_n,R_n)$, so also a $\axK\axD4\axG_1$-frame; it is connected since $(W_n,R_n)$ validates $\axG_1$. As $e_n$ is $R_n$-maximal, $\c C_n$ is a nondegenerate cluster, so trivially a connected $\axK\axD4\axG_1$-frame, and (as $R_n$ is transitive) a generated subframe of $\c F_n$. We conclude from proposition \[prop:di rep\] that $\c F_n$ and $\c C_n$ are fully over $X$, for all $n<\omega$.
Now fix arbitrary $x_0\in X$. Let $O$ be an open of $x_0$. Since $X$ is a metric space, all singletons are closed, and since it is dense in itself, lemma \[lem:inf\] tells us that $O$ is infinite, so we can pick $y\in O\setminus\{x_0\}$. Then $O\setminus\{y\}$ is open, $\{x_0\}\subseteq O\setminus\{y\}$, and $\{x_0\}$ is closed. By lemma \[lem:ms\], $X$ is normal, so there is open $P$ with $x_0\in P\subseteq\int\cl P\subseteq \cl P\subseteq O\setminus\{y\}\subset O$ (the last inclusion being strict). Note that $\int\cl P$ is in $X$. So *every open of $x_0$ properly contains the closure of some of $x_0$.* Using this repeatedly, we may choose subsets $O_n,P_n$ of $X$ (for $n<\omega$) containing $x_0$, with $O_0=X$, and with the following properties:
1. $\cl O_{n+1}\subset P_n$ and $\cl P_n\subset O_n$ (the inclusions are strict) for each $n<\omega$.
2. $O_n\subseteq N_{1/n}(x_0)$ for each $n>0$.
It follows that for every open $O$ of $x_0$, there is $n<\omega$ with $O_n\subseteq O$. That is, the $O_n$ form a base of open s of $x_0$.
For each $n<\omega$ define open sets $$\begin{array}{rcl}
U_n&=&O_n\setminus\cl P_{n+1},
\\
S_n&=&O_n\setminus\cl P_n.
\end{array}$$ See figure \[fig:cone\].
![rough guide to the sets $O_n,P_n,U_n,S_n$[]{data-label="fig:cone"}](strongcomplfig1.ps){width="11cm"}
It is easily seen that $$\begin{aligned}
\bigcup_{n<\omega}(O_n\setminus O_{n+1})&=&X\setminus\{x_0\},\label{e:fact1 on On}
\\
\bigcup_{n\leq m<\omega}U_m&=& O_n\setminus\{x_0\}
\quad\mbox{ for each }n<\omega.\label{e:fact2 on On}\end{aligned}$$ The following claim lists some more basic facts about our situation.
2 For each $n<\omega$:
1. $U_n\cap U_{n+1}=S_{n+1}\neq\emptyset$.
2. $S_n\cup S_{n+1}\subseteq U_n$,
3. $\cl S_n\cap\cl S_{n+1}=\emptyset$,
4. $S_n$, $S_{n+1}$, and $S_n\cup S_{n+1}$ are subsets of $U_n$,
5. $U_n\setminus\cl(S_n\cup S_{n+1})\neq\emptyset$.
<!-- -->
1. Easy.
2. From the definitions we have $S_n=O_n\setminus\cl P_n\subseteq O_n\setminus\cl P_{n+1}=U_n$ and $S_{n+1}=O_{n+1}\setminus\cl P_{n+1}\subseteq O_n\setminus\cl P_{n+1}=U_n$.
3. It is clear that $$\label{e:clBn}
\cl S_n\subseteq\cl O_n\setminus P_n.$$ Applying this for $n+1$ and $n$ gives $\cl S_{n+1}\cap\cl S_n\subseteq
\cl O_{n+1}\setminus P_n\subseteq P_n\setminus P_n=\emptyset$.
4. $O_n$ and $P_n$ are subsets of $X$, so by lemma \[lem:ro1\], $S_n=O_n\setminus\cl P_n$ is a subset of $X$ too. Since $\cl S_n\cap\cl S_{n+1}=\emptyset$ by part 2, lemma \[lem:ro1\](\[lem:ro:part2\]) yields that $S_n\cup S_{n+1}$ is also a subset of $X$. Since each of these three sets is a subset of $U_n$ by part 2, by lemma \[lem:ro1\](\[lem:ro:part3\]) it is also in $U_n$.
5. By (for $n$ and $n+1$), $\cl S_n$ and $\cl S_{n+1}$ are disjoint from $P_n\setminus\cl O_{n+1}$, so by additivity of closure, $U_n\setminus\cl(S_n\cup S_{n+1})
=U_n\setminus(\cl S_n\cup\cl S_{n+1})\supseteq P_n\setminus\cl O_{n+1}\neq\emptyset$.
3 There are surjective s $\rho_n$ of $\c F_n$ over $U_n$ ($n<\omega$) such that
1. $\rho_n\restriction S_{n+1}$ is a of $\c C_n$ over $S_{n+1}$,
2. $\tp_n(\rho_n(x))=\tp_{n}(\rho_{n+1}(x))$ for all $x\in S_{n+1}$.
We define the $\rho_n$ by induction on $n$. First let $n=0$. Since $\c C_0$ is fully over $X$, we can choose a $\sigma:S_1\to\c C_0$. Because $\c C_0$ is a nondegenerate cluster, $\sigma$ is actually a $U_0$-basic (see remark \[rmk:basic reps\]). By claim 2, $S_1$ is a subset of $U_0$, and $U_0\setminus\cl S_1\neq\emptyset$. Now $\c F_0$ is also fully over $X$, so $\sigma$ extends to a surjective $\rho_0$ of $\c F_0$ over $U_0$. Clearly, condition 1 above is met.
Let $n<\omega$ and assume inductively that for each $m\leq n$, a surjective $\rho_m$ of $\c F_m$ over $U_m$ has been constructed, such that $\rho_m\restriction S_{m+1}$ is a of $\c C_m$ over $S_{m+1}$ and $tp_m(\rho_m(x))=tp_{m}(\rho_{m+1}(x))$ for all $x\in S_{m+1}$ whenever $m<n$. We will define $\rho_{n+1}$ to continue the sequence.
Note first that since $\c C_n$ is a non-degenerate cluster, $\rho_n\restriction S_{n+1}$ is $U_n$-basic — see remark \[rmk:basic reps\]. It is also surjective. For, let $w \in R_n(e_n)$ be given. Take $x\in S_{n+1}$ (note that $S_{n+1}$ is non-empty by claim 2). As $\c C_n$ is a non-degenerate cluster, $R_n(\rho_n(x),w)$, so as $\rho_n\restriction S_{n+1}$ is a , $(S_{n+1},(\rho_n\restriction S_{n+1})^{-1}),x\models\did w$. This certainly implies that $\rho_n(y)=w$ for some $y\in S_{n+1}$.
For each $w\in R_n(e_n)$, define $$\begin{array}{rclcl}
D_w&=&\{x\in S_{n+1}:\rho_n(x)=w\}&\subseteq&S_{n+1},
\\
H_w&=&\{v\in R_{n+1}(e_{n+1}):\tp_n(v)=\tp_n(w)\}&\subseteq& W_{n+1},
\\
\c H_w&=&(H_w,R_{n+1}\restriction H_w).
\end{array}$$ See figure \[fig:maps\].
![illustration for claim 3[]{data-label="fig:maps"}](strongcomplfig2.ps){width="11cm"}
Because $\rho_n\restriction S_{n+1}$ is surjective onto $\c C_n$, each set $D_w$ is non-empty, and plainly, $S_{n+1}$ is partitioned by the $D_w$ ($w\in R_n(e_n)$). Because $\tau{}^{n+1}_n=\tau{}^n_n$, each $H_w$ is non-empty and $\bigcup_{w\in R_n(e_n)}H_w=R_{n+1}(e_{n+1})$. (The sets $H_w$ may not be pairwise disjoint, but any two of them are equal or disjoint.)
Let $w\in R_n(e_n)$ and consider $D_w$ as a subspace of $X$. We show that it is dense in itself. Let $x\in D_w$ and suppose for contradiction that $\{x\}$ is open in $D_w$. So there is open $O\subseteq X$ with $O\cap D_w=\{x\}$, and as $S_{n+1}$ is open, we can suppose that $O\subseteq S_{n+1}$. Now by the inductive hypothesis, $\rho_n\restriction S_{n+1}$ is a of $\c C_n$ over $S_{n+1}$. Because $\c C_n$ is a non-degenerate cluster, $R_nww$, so $(X,(\rho_n\restriction S_{n+1})^{-1}),x\models\did w$. So there is $y\in O\setminus\{x\}$ with $\rho_n(y)=w$. But then $y\in O\cap D_w=\{x\}$, a contradiction.
So $D_w$ is a space in its own right. Since $\c C_{n+1}$ is a nondegenerate cluster, so is its subframe $\c H_w$. Hence, $\c H_w$ is trivially a finite connected $\axK\axD4\axG_1$ frame. So by proposition \[prop:di rep\], there is a surjective $$\sigma_w:D_w\to H_w$$ of $\c H_w$ over $D_w$. We have $(D_{w},\sigma_{w}^{-1}),x\models\did v$ for every $x\in D_w$ and $v\in H_w$. By lemma \[lem:bod reps invar\], $$\label{e:subspace rep}
(X,\sigma_{w}^{-1}),y\models\did v\quad
\mbox{ for every }x\in D_w\mbox{ and }v\in H_w.$$
Now let $$\sigma=\Big(\bigcup_{w\in R_n(e_n)}\sigma_w\Big)
\;:\;S_{n+1}\to R_{n+1}(e_{n+1}).$$ The sets $D_w$ partition $S_{n+1}$, so $\sigma$ is a well defined and total map. It has the following property. Let $x\in S_{n+1}$. Writing $\rho_n(x)=w$, say, we have $x\in D_w$ and $\sigma(x)=\sigma_w(x)\in H_w$, so $\tp\nolimits_n(\sigma(x))=\tp\nolimits_n(w)$ by definition of $H_w$. That is, $$\label{e:types pres by sig}
\tp\nolimits_n(\sigma(x))=\tp\nolimits_n(\rho_n(x))\quad
\mbox{ for each }x\in S_{n+1}.$$
We show that $\sigma$ is a of $\c C_{n+1}$ over $S_{n+1}$. Since $\c C_{n+1}$ is a non-degenerate cluster, we need show only that $(X,\sigma^{-1}),x\models\did v$ for every $x\in S_{n+1}$ and $v\in R_{n+1}(e_{n+1})$.
So take such $x,v$. Suppose that $\rho_n(x)=w$, say, so $x\in D_w$. Choose $w'\in R_n(e_n)$ such that $v\in H_{w'}$ (it may not be unique). As $\c C_n$ is a cluster, $R_n(w,w')$. As $\rho_n\restriction S_{n+1}$ is a of $\c C_n$ over $S_{n+1}$, we have $(X,(\rho_n\restriction S_{n+1})^{-1}),x\models\did w'$. That is, $x\in\did D_{w'}$. But by , $(X,\sigma^{-1}),y\models\did v$ for every $y\in D_{w'}$. It follows that $(X,\sigma^{-1}),x\models\did\did v$, and hence $(X,\sigma^{-1}),x\models\did v$ as required.
So $\sigma$ is indeed a of $\c C_{n+1}$ over $S_{n+1}$. As $\c C_{n+1}$ is fully over $X$, we may choose a $\sigma'$ of $\c C_{n+1}$ over $S_{n+2}$. By claim 2, $S_{n+1}\cap S_{n+2}=\emptyset$, so by lemma \[lem:reps simple\], $\sigma\cup\sigma'$ is a well defined of $\c C_{n+1}$ over the subset $S_{n+1}\cup S_{n+2}$ of $U_{n+1}$. Also, $U_{n+1}\setminus\cl(S_{n+1}\cup S_{n+2})\neq\emptyset$. And since $\c C_{n+1}$ is a nondegenerate cluster, $\sigma\cup\sigma'$ is $U_{n+1}$-basic (see remark \[rmk:basic reps\] again). We can now use the fact that $\c F_{n+1}$ is fully over $X$ to extend $\sigma\cup\sigma'$ is to a surjective $\rho_{n+1}$ of $\c F_{n+1}$ over $U_{n+1}$. Then $\rho_{n+1}\restriction S_{n+2}=\sigma'$ is a of $\c C_{n+1}$ over $S_{n+2}$, and by , $tp_n(\rho_n(x))=\tp_n(\sigma(x))=tp_{n}(\rho_{n+1}(x))$ for all $x\in S_{n+1}$. This proves claim 3.
Let $n<\omega$. Define an assignment $g_n$ on $U_n$ by $$\label{e:def g}
g_n(p)=\rho_n^{-1}(h_n(p))\mbox{ for each atom }p.$$ By the claim, if $p\in L_n$, then for each $x\in S_{n+1}$ we have $x\in g_n(p)$ iff $\rho_n(x)\in h_n(p)$, iff $p\in\tp_n(\rho_n(x))=\tp_n(\rho_{n+1}(x))$, iff $\rho_{n+1}(x)\in h_{n+1}(p)$, iff $x\in g_{n+1}(p)$. So $$\label{eq:agree}
S_{n+1}\cap g_n(p)=S_{n+1}\cap g_{n+1}(p)\quad\mbox{ for each }p\in L_n.$$ Finally, define an assignment $g$ on $X$ as follows. Let $p$ be an atom.
- For $x\in X\setminus\{x_0\}$, define $x\in g(p)$ iff $x\in g_n(p)$, where $x\in O_n\setminus O_{n+1}$.
Since the $O_n\setminus O_{n+1}$ are pairwise disjoint, and $\bigcup_{n<\omega}(O_n\setminus O_{n+1})=X\setminus\{x_0\}$ by , this is well defined.
- Define $(X,g),x_0\models p$ iff $p\in\Gamma$.
4 Let $n<\omega$, let $x\in U_n$, and let $\varphi$ be a formula whose atoms lie in $L_n$. Then $(X,g),x\models\varphi$ iff $\c M_n,\rho_n(x)\models\varphi$.
Let $p\in L_n$ be arbitrary. Recall that $U_n=O_n\setminus\cl P_{n+1}$. By definition of $g$, if $x\in O_n\setminus O_{n+1}$ then $x\in g(p)$ iff $x\in g_n(p)$. If instead $x\in O_{n+1}$, then $x\in O_{n+1}\setminus\cl P_{n+1}=S_{n+1}\subseteq O_{n+1}\setminus O_{n+2}$, and since $p\in L_{n+1}$ too, the definition of $g$ gives $x\in g(p)$ iff $x\in g_{n+1}(p)$. But by , this is iff $x\in g_n(p)$ again. So $g$ and $g_n$ agree on $U_n$ as far as atoms in $L_n$ are concerned, and as $U_n$ is open, it follows easily that $(X,g),x\models\varphi$ iff $(U_n,g_n),x\models\varphi$. Since $\rho_n$ is a over $U_n$ of the generated subframe $\c F_n$ of $(W_n,R_n)$, by lemma \[lem:reps simple\] it is also a of $(W_n,R_n)$ over $U_n$. So by and proposition \[prop:fmla pres\], $(U_n,g_n),x\models\varphi$ iff $\c M_n,\rho_n(x)\models\varphi$. This proves the claim.
5 For all $\varphi$ we have $(X,g),x_0\models\varphi$ iff $\varphi\in\Gamma$.
By induction on $\varphi$. For atoms, the result follows from the definition of $g$. The boolean operators are handled in the usual way by induction, using the maximal consistency of $\Gamma$; they are the only cases in which the inductive hypothesis is used.
We now tackle the case $\bod\varphi$. It is sufficient (and seems more intuitive) to deal with $\did\varphi$ instead. Suppose first that $\did\varphi\in \Gamma$. Choose $n<\omega$ such that $\did\varphi\in\Gamma_n$. Let $i\geq n$ be arbitrary. Then $\did\varphi\in\Gamma_i$, so $\c M_i,w_i\models\did\varphi$, and hence there is $v\in R_i({w_i})$ with $\c M_i,v\models\varphi$. As $\rho_i:U_i\to R_i(w_i)$ is surjective (see claim 3), there is $x\in U_i$ with $\rho_i(x)=v$. Since $\did\varphi\in\Gamma_i$, the atoms of $\varphi$ lie in $L_i$, so claim 4 applies: $(X,g),x\models\varphi$. We conclude that for every $i\geq n$ there is $x\in U_i$ with $(X,g),x\models\varphi$. As $U_i\subseteq O_i\setminus\{x_0\}$ and the $O_i$ form a base of s of $x_0$, it follows that $(X,g),x_0\models\did\varphi$.
Conversely, suppose that $(X,g),x_0\models\did\varphi$. For each $n<\omega$, $O_n$ is an open of $x_0$, so there is $x\in O_n\setminus\{x_0\}$ with $(X,g),x\models\varphi$. Since $O_n\setminus\{x_0\}=\bigcup_{n\leq i<\omega}U_i$ by , we have $x\in U_i$ for some $i\geq n$. It follows that there are infinitely many $i<\omega$ such that $(X,g),x\models\varphi$ for some $x\in U_i$. Since the atoms of $\varphi$ lie in $L_i$ for cofinitely many $i$, there must be infinitely many $i$ with $\c M_i,v\models\varphi$ for some $v\in R_i({w_i})$ (by claim 4), and so $\c M_i,w_i\models\did\varphi$ (by Kripke semantics), and so $\neg\did\varphi\notin\Gamma_i$ (since $\c M_i,w_i\models\Gamma_i$). Since $\Gamma$ is the union of the chain $\Gamma_0\subseteq\Gamma_1\subseteq\cdots$, we have $\neg\did\varphi\notin\Gamma$. As $\Gamma$ is maximal consistent, it follows that $\did\varphi\in\Gamma$.
Finally, consider the case $\didt\Delta$, where $\Delta$ is any non-empty finite set of formulas. Suppose first that $\did\Delta\in\Gamma$. We only sketch the proof here, referring the reader to the case of $\did\varphi$ for more details. Pick any $\delta\in\Delta$. Then as in the case of $\did\varphi$, each of the following holds for cofinitely many $i<\omega$:
- $\didt\Delta\in\Gamma_i$
- $\c M_i,w_i\models\didt\Delta$
- there is $v\in R_i(w_i)$ with $\c M_i,v\models\delta\wedge\didt\Delta$
- there is $x\in U_i$ with $(X,g),x\models\delta\wedge\didt\Delta$.
As the latter holds for every $\delta\in\Delta$, it follows that $(X,g),x_0\models\didt\Delta$.
Conversely, suppose $(X,g),x_0\models\didt\Delta$. Then as in the $\did\varphi$ case, there are infinitely many $i<\omega$ such that $(X,g),x\models\didt\Delta$ for some $x\in U_i$. Since the atoms of $\didt\Delta$ lie in $L_i$ for cofinitely many $i<\omega$, it follows by claim 4 that there are infinitely many $i$ such that there is $v\in R_i(w_i)$ with $\c M_i,v\models\didt\Delta$, and hence — by the semantics of $\didt$ — $\c M_i,w_i\models\didt\Delta$. As in the $\did\varphi$ case, we obtain $\neg\didt\Delta\notin\Gamma_i$ for infinitely many $i$, so $\neg\didt\Delta\notin\Gamma$, and so $\didt\Delta\in\Gamma$ by maximal consistency of $\Gamma$. The claim is proved, and the theorem with it.
Strong completeness for $\c L_\bod$ {#ss:compactness bod}
-----------------------------------
We can now easily derive the analogous result for ‘modal’ $\c L_\bod$-formulas, essentially by showing that $\axK\axD4\axG_1t$ is a conservative extension of $\axK\axD4\axG_1$.
\[thm:sc bod\] Let $X$ be a non-empty space. Then the Hilbert system $\axK\axD4\axG_1$ is strongly complete over $X$ for $\c L_{\bod}$-formulas, and sound if $\axG_1$ is valid in $X$.
For soundness, see theorem \[thm:compl KD4G1(t)\]. For strong completeness, let $\Gamma$ be a countable $\axK\axD4\axG_1$-consistent set of $\c L_\bod$-formulas. Let $\Gamma_0\subseteq\Gamma$ be finite and put $\gamma=\bigwedge\Gamma_0$. Then $\gamma$ is $\axK\axD4\axG_1$-consistent, so by the results of section \[sec:fmp Gn\] it is satisfied in some finite $\axK\axD4\axG_1$-frame $\c F$. Plainly, $\c F$ is also a $\axK\axD4\axG_1t$-frame, and it follows that $\gamma$ is $\axK\axD4\axG_1t$-consistent. So $\Gamma$ is $\axK\axD4\axG_1t$-consistent. By theorem \[thm:strong did\], $\Gamma$ is satisfiable over $X$.
Strong completeness for $\c L_\bo^\dit$ and $\c L_\bo^\mu$ {#sec:strong c}
----------------------------------------------------------
This also follows, using the translations $-^d$ and $-^t$ of section \[sec:translations\].
\[thm:str compl boxes\] Let $X$ be any space.
1. The Hilbert system $\axS4t$ is sound and strongly complete over $X$ for $\c L_\bo^\dit$-formulas.
2. \[stro compl boxes part2\] The Hilbert system $\logicbomu$ is sound and strongly complete over $X$ for $\c L_\bo^\mu$-formulas.
3. (Kremer, [@Kremer2010:stro]) The Hilbert system $\axS4$ is sound and strongly complete over $X$ for $\c L_\bo$-formulas.
Soundness is clear in all cases: cf. theorem \[thm:compl S4mu, S4t\]. We prove strong completeness. For part 1, let $\varphi$ be an $\axS4t$-consistent $\c L_\bo^\dit$-formula. By the results of section \[sec:fmp S4\], $\varphi$ is satisfiable in some finite S4 Kripke frame $\c F$. Recall from section \[sec:translations\] the translation $-^d$: it takes $\c L_\bo^\dit$-formulas to $\c L_\bod^\didt$-formulas. Since $\c F$ is reflexive, it follows from lemma \[lem:-d in refl frames\] that $\varphi^d$ is equivalent to $\varphi$ in $\c F$. So $\varphi^d$ is satisfiable in $\c F$. Plainly, $\c F$ is also a $\axK\axD4\axG_1t$ frame, so $\varphi^d$ is $\axK\axD4\axG_1t$-consistent.
Since $-^d$ commutes with $\wedge$, it is now easily seen that if $\Gamma\subseteq\c L_\bo^\dit$ is a countable $\axS4t$-consistent set then $\Gamma^d=\{\gamma^d:\gamma\in\Gamma\}\subseteq\c L_\bod^\didt$ is a countable $\axK\axD4\axG_1t$-consistent set. By theorem \[thm:strong did\], $\Gamma^d$ is satisfiable over $X$. Since $X$ is T$_D$, by lemma \[lem:trans equivalent top\] each $\gamma\in\Gamma$ is equivalent to $\gamma^d$ in $X$, so $\Gamma$ is also satisfiable over $X$.
For part 2, for a set $\Gamma\subseteq\c L_\bo^\mu$ we write $\Gamma^t=\{\gamma^t:\gamma\in\Gamma\}\subseteq\c L_\bo^\dit$, where the translation $-^t:\c L_\bo^\mu\to\c L_\bo^\dit$ is as in section \[ss:DO09\]. Let $\Gamma\subseteq\c L_\bo^\mu$ be a countable $\axS4\mu$-consistent set. Let $\Gamma_0\subseteq\Gamma$ be any finite subset. By assumption, the formula $\bigwedge\Gamma_0$ is $\axS4\mu$-consistent. So by theorem \[thm:S4mu sc\], there is a finite $\axS4$ frame $\c F$ in which $\bigwedge\Gamma_0$ is satisfied. By fact \[fact:DO\], $\varphi^t$ is equivalent to $\varphi$ in $\c F$, for each $\varphi\in\c L_\bo^\mu$. So $\bigwedge(\Gamma_0^t)$ is also satisfied in $\c F$. Since $\c F$ is plainly an $\axS4t$ frame, it follows that $\bigwedge(\Gamma_0^t)$ is $\axS4t$-consistent. As $\Gamma_0$ was arbitrary, $\Gamma^t$ is $\axS4t$-consistent.
By part 1, $\Gamma^t$ is satisfied in $X$. But by corollary \[cor:t transl top equiv\], each $\gamma\in\Gamma$ is equivalent to $\gamma^t$ in $X$. So $\Gamma$ is also satisfied in $X$.
Part 3 can be proved similarly, by showing in the same way that for $\c L_\bo$-formulas, $\axS4$-consistency implies $\axS4t$-consistency, and then appealing to part 1.
Universal modality
------------------
We do not include the universal modality in our strong completeness results, for good reason.
\[thm:comp fails\] There is a set $\Sigma$ of $\c L_{\bo\forall}$-formulas such that for every non-empty compact locally connected space $X$, each finite subset of $\Sigma$ is satisfiable in $X$, but $\Sigma$ as a whole is not.
*Compact* means that if $\c S$ is a set of open sets with $\bigcup\c S=X$, then $X=\bigcup\c S_0$ for some finite $\c S_0\subseteq\c S$. *Locally connected* means that every open neighbourhood of a point $x$ contains a connected (in the subspace topology) open neighbourhood of $x$. An example of a compact locally connected space is the subspace $[0,1]$ of $\R$.
The proof is based on the following model $\c M=(W,R,h)$, where we suppose that $\Var=\{\r,\g,\bb\}\cup\{p_i:i<\omega\}$.
1. $W=\{a_n,b_n:n<\omega\}$, where the $a_n$ and $b_n$ are pairwise distinct
2. $R$ is the reflexive closure of $\{(a_n,b_n),(a_n,b_{n+1}):n<\omega\}$
3. $h(\r)=\{b_{3n}:n<\omega\}$, $h(\g)=\{b_{3n+1}:n<\omega\}$, $h(\bb)=\{b_{3n+2}:n<\omega\}$, and $h(p_n)=\{b_{3n},b_{3n+1}\}$ for each $n<\omega$.
(100,140)(-60,-45) (0,0)
(0,0)(80,0)[6]{}[ (42,77)[(1,-2)[37]{}]{} (38,77)[(-1,-2)[37]{}]{} (40,80) (80,0) ]{}
(15,80)[$a_0$]{} (95,80)[$a_1$]{} (175,80)[$a_2$]{} (255,80)[$a_3$]{} (335,80)[$a_4$]{} (415,80)[$a_5$]{}
(-25,0)[$b_0$]{} (55,0)[$b_1$]{} (135,0)[$b_2$]{} (215,0)[$b_3$]{} (295,0)[$b_4$]{} (375,0)[$b_5$]{} (455,0)[$b_6$]{}
(0,-40)[$p_0$]{} (80,-40)[$p_0$]{} (240,-40)[$p_1$]{} (320,-40)[$p_1$]{} (480,-40)[$p_2$]{}
(0,-20)[$\r$]{} (240,-20)[$\r$]{} (480,-20)[$\r$]{}
(80,-20)[$\g$]{} (320,-20)[$\g$]{}
(160,-20)[$\bb$]{} (400,-20)[$\bb$]{}
(500,41)[(-1,-2)[19]{}]{} (520,40)[$\cdots$]{}
The model is shown in figure \[fig:w\] — it goes off to the right forever, roughly repeating after every three steps. Of course $R$ is reflexive. Note that the underlying frame is connected.
We let $\Sigma$ be the set comprising the following formulas:
1. \[sigma2\] $\exists( \di p_i\wedge\di\r\wedge\di\g)$ for each $i<\omega$
2. \[sigma3a\] $\forall\neg(\di p_i\wedge\di p_j)$ for $i<j<\omega$
3. \[sigma10\] $\forall\neg(\di\r\wedge\di\g\wedge\di\bb)$
4. \[sigma15\] $\forall(\di p_i\wedge\bo\neg\bb\to\bo\di p_i)$ for $i<\omega$.
They are plainly valid in $\c M$. Hence $\Sigma$ is satisfied in $\c M$, at every point. Moreover, any finite subset $\Sigma_0\subseteq\Sigma$ is satisfied in a finite submodel of $\c M$ obtained by taking a large enough ‘initial segment’ of $\c M$ ending on the right at a $b$-world. Check especially formulas of the form $\forall\exists$. In particular, \[sigma15\] is valid in such a submodel. Or one can use that it is a generated submodel. The submodel is finite and its frame validates $\axS4.\axU\axC$, so every formula satisfied in it — for example, $\bigwedge\Sigma_0$ — is $\axS4.\axU\axC$-consistent. Hence, by theorem \[thm:compl S4.UC, S4t.UC\], every finite subset of $\Sigma$ is satisfiable in $X$.
Assume for contradiction that $\Sigma$ is satisfied in some model $(X,h)$ on $X$. Below, we will write $x\models\varphi$ instead of $(X,h),x\models\varphi$. By \[sigma2\], for each $i<\omega$ there is $x_i\in X$ with $x_i\models \di p_i\wedge\di\r\wedge\di\g$. As $X$ is compact, it contains a point $z$ such that for every open $N$ of $z$, the set $\{i<\omega:x_i\in N\}$ is infinite. Then $z\models\di\r\wedge\di\g$ as well. By \[sigma10\], $z\models\bo\neg\bb$. As $X$ is locally connected, there is a connected open neighbourhood $N$ of $z$ with $y\models\neg\bb$ for all $y\in N$.
Take $i<j<\omega$ with $x_i,x_j\in N$. Let $U=\{x\in N:x\models\di p_i\}$. Then $U$ is an open subset of $N$, because for every $u\in U$ we have $u\models\di p_i\wedge\bo\neg\bb$, and \[sigma15\] gives $u\models\bo\di p_i$. And $N\setminus U$ is also open, because $U'=\{x\in X:x\models\di p_i\}$ is closed and $N\setminus U=N\setminus U'$. We have $x_i\in U$, but by \[sigma3a\], $x_j\in N\setminus U$. So $N$ is the union of two disjoint non-empty open sets ($U$ and $N\setminus U$), contradicting its connectedness.
\[cor:stro compl fails\] Let $X$ be a non-empty compact locally connected space, and $\c L\subseteq\Lbig$ a language containing $\c L_{\bo\forall}$ or $\c L_{\bod\forall}$. Then no Hilbert system for $\c L$ is sound and strongly complete over $X$.
Assume for contradiction that the Hilbert system $H$ is sound and strongly complete over $X$. Let $\Sigma$ be as in theorem \[thm:comp fails\] (use the translation $-^d$ if necessary to ensure it is a set of $\c L$-formulas). Since every finite subset of $\Sigma$ is satisfiable in $X$, and $H$ is sound over $X$, it follows that $\Sigma$ is $H$-consistent. But $H$ is strongly complete over $X$, so $\Sigma$ is satisfiable over $X$, contradicting the theorem.
Conclusion {#sec:the end}
==========
This paper has presented some completeness theorems for various spatial logics over spaces. Table \[table1\] summarises them. The numbers in parentheses refer to our earlier results. The first line of the table is of course known, included here to give a more complete picture. For handy reference, table \[table2\] summarises the ingredients of each logic.
Language Logic sound complete strongly complete
------------------------------------ ----------------------------- --------------------------------------- ---------------- ------------------------
$\c L_\bo$ $\axS4$ yes yes [@McKiT44] yes [@Kremer2010:stro]
\[3pt\] $\c L_\bo^\mu$ $\axS4\mu$ yes yes yes
\[3pt\] $\c L_\bo^\dit$ $\axS4t$ yes yes yes
\[3pt\] $\c L_{\bo\forall}$ $\axS4.\axU\axC$ if $X$ connected yes not in general
\[3pt\] $\c L_{\bo\forall}^\dit$ $\axS4t.\axU\axC$ if $X$ connected yes not in general
\[3pt\] $\c L_\bod$ $\axK\axD4\axG_1$ if $\axG_1$ valid in $X$ yes yes
\[3pt\] $\c L_{\bod}^\didt$ $\axK\axD4\axG_1t$ if $\axG_1$ valid in $X$ yes yes
\[3pt\] $\c L_{\bod\forall}$ $\axK\axD4\axG_1.\axU\axC$ if $X$ connected & validates $\axG_1$ yes not in general
\[3pt\] $\c L_{\bod\forall}^\didt$ $\axK\axD4\axG_1t.\axU\axC$ if $X$ connected & validates $\axG_1$ yes not in general
\[5pt\]
: Soundness and completeness for a non-empty space $X$[]{data-label="table1"}
---------- --------------------------------------------------------------------------------------------------------------
S4 $\bo\varphi\to\varphi$, $\bo\varphi\to\bo\bo\varphi$
$\mu$ fixed point axiom and rule: see definition \[def:mu systems\]
$t$ tangled closure axioms from section \[ss:tangle logics\]
U $\forall\varphi\to\bo\varphi$, S5 axioms for $\forall$, $\forall$-generalisation rule
C $\forall(\Box^*\ph\lor\Box^*\neg\ph)\to(\forall\ph\lor\forall\neg\ph)$, where $\bo^*\varphi=\ph\land\Box\ph$
$\axG_1$ all uniform substitution instances of $\big(\bod\bigvee_{i=0}^1\bo Q_i\big)
\to\bigvee_{i=0}^1\bod\neg Q_i$,
1em where $Q_i=p_i\wedge\neg p_{1-i}$ ($i=0,1$)
\[2pt\]
---------- --------------------------------------------------------------------------------------------------------------
: Parts of the logics[]{data-label="table2"}
There are of course many problems left open by our work, and we present some of them here.
Extensions
----------
Can the results be extended to more general topological spaces?
For example, consider the topological space $T$ defined as follows. For ordinals $\alpha,\beta$ write ${}^\alpha\beta$ for the set of all maps $f:\alpha\to\beta$. The set of points of $T$ is $\bigcup_{n\leq\omega}{}^n2$, and the open sets are unions of sets of the form $\{f\in T:f\supseteq g\}$ for some $g\in\bigcup_{n<\omega}{}^n2$. This space is not even T$_D$, though it is T0 (that is, no two distinct points have the same open s) and dense in itself.
What is the logic of $T$ in the various languages discussed above?
Can the results be extended to stronger languages, for example, the mu-calculus with $\bod$ and/or , languages with the difference modality or graded modalities, hybrid languages, and so on? Results of Kudinov [@Kud06:topdiff; @Kud08:Rnfa] are relevant. Recently, Kudinov and Shehtman [@KudShe14] proved numerous results about logics of topology with $\bo$, $\bod$, , and the ‘difference modality’ $[\neq]$. In particular, they determine the logic of $\R^n$ for $n\geq2$ in the language with $\bod$ and $[\neq]$. However, results for general spaces appear to be lacking.
Strong completeness {#subsec:strongcomp}
-------------------
Our definition of strong completeness is limited to countable sets of formulas. We have not investigated the extent to which the strong completeness results in section \[sec:strong compl\] generalise to uncountable sets, but an argument based on the Erdős–Rado theorem [@ErdRad56] will show that for any given dense-in-itself topological space $X$ and any Hilbert system $H$ that is sound over $X$, there is an (uncountable) cardinal $\kappa$ such that the set $\{\di p_i:i<\kappa\}\cup\{\bo\neg(p_i\wedge p_j):i<j<\kappa\}$ is $H$-consistent but not satisfiable in $X$. So strong completeness will fail over any given $X$, for large enough sets of formulas.
Let $X$ be a space. For which uncountable cardinals $\kappa$ can our strong completeness results for $X$ be extended to sets of at most $\kappa$ formulas?
Our strong completeness results for languages with $\bod$ are limited to logics with $\axG_1$. We could ask for more:
Let $X$ be a space and let $\c L$ be $\c L_\bod$ or $\c L_{\bod}^{\didt}$. Is the $\c L$-logic of $X$ strongly complete over $X$?
By theorems \[thm:strong did\] and \[thm:sc bod\], the answer is ‘yes’ if $X$ validates $\axG_1$.
We saw in corollary \[cor:stro compl fails\] that in the language $\c L_{\bo\forall}$, there are many dense-in-themselves metric spaces over which $\axS4.\axU\axC$ is not strongly complete. So we ask:
Can strong completeness for languages with be proved for each space in some reasonably large class, and for $\R^n$ for $n\geq1$?
Is $\axS4.\axU\axC$ strongly complete for Kripke semantics in the language $\c L_{\bo\forall}$?
Even without , the example in section \[ss:can frame\] can be used to show that strong completeness fails in Kripke semantics for all our systems for languages containing $\c L_{\bo}^{\dit}$. But we saw that strong completeness does hold for some of these systems over dense-in-themselves metric spaces. Taking the example of $\axS4t$ for $\c L_{\bo}^\dit$, it is striking that this logic is sound and complete for two different semantics (the class of finite S4 frames, and any non-empty space), but strongly complete for only the latter.
Is there any general connection between strong completeness for topological semantics and for Kripke semantics?
Complexity
----------
Decidability of the logics in table \[table1\] follows from the finite model property results of section \[sec:fmp\] and their finite (schema) axiomatisations. But we have not investigated their complexity.
What is the complexity of the logics discussed in this paper?
Of course, the complexity of some are known (e.g., S4 is [PSpace]{}-complete).
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Nick Bezhanishvili for stimulating conversations and for graciously sharing his interest in spatial semantics for the mu-calculus, and Valentin Shehtman for some inspiring lectures on spatial logic and help with references. The second author thanks the School of Mathematics and Statistics at the Victoria University of Wellington for their warm and generous hospitality during his visit in January–March 2014 when the authors worked on the research presented here, and the UK EPSRC for an overseas travel grant, EP-L020750-1, which supported this visit.
[^1]: School of Mathematics and Statistics, Victoria University of Wellington, New Zealand. [sms.vuw.ac.nz/\~rob/]{}
[^2]: Department of Computing, Imperial College London, UK. [www.doc.ic.ac.uk/\~imh/]{}
[^3]: The separability assumption was removed in [@RS:mm].
[^4]: Indeed, $\bo\di(p\vee q)\to\bo\di p\vee\bo\di q\vee(\di p\wedge\di q)$ is valid in $\axS4$ frames, so provable in $\axS4$. Since $\axS4$ is sound over $X$, the formula is valid in $X$.
[^5]: In [@Sheh:everywhere99 theorem 18], Shehtman states this result when $X$ is additionally assumed separable. However, [@KudShe14 footnote 7] states that [@Sheh:everywhere99] “contains a stronger claim: \[the $\c L_{\bo\forall}$-logic of $X$ is $\axS4.\axU\axC$\] for any connected dense-in-itself separable metric $X$. However, recently we found a gap in the proof of Lemma 17 from that paper. Now we state the main result only for the case $X=\R^n$; a proof can be obtained by applying the methods of the present Chapter, but we are planning to publish it separately.”
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $A \subseteq {\mathbb{F}}_2^n$ be a set with $|2A| = K|A|$. We prove that if
1. for at least a fraction $1-K^{-9}$ of all $s \in 2A$, the set $(A+s) \cap A$ has size *at most* $L\cdot|A|/K$, or
2. for at least a fraction $K^{-L}$ of all $s \in 2A$, the set $(A+s) \cap A$ has size *at least* $|A|\cdot(1- K^{-1/L})$,
then there is a subset $B \subseteq A$ of size $|A|/K^{O_L(1)}$ such that $\operatorname*{span}(B) \leq K^{O_L(1)}\cdot|A|$.
author:
- Thomas Holenstein
bibliography:
- 'bibliography.bib'
title: |
The PFR Conjecture Holds for\
Two Opposing Special Cases
---
Introduction
============
If $A$ is a subset of an abelian group $G$ (which will always be ${\mathbb{F}}_2^n$ in this paper), we let $A+A := 2A := \{a_1+a_2 | a_1,a_2 \in A\}$. Suppose that $|A+A| = K|A|$ (where one should think of $K$ as “small” compared to $|A|$). Intuitively, one expects $A$ to have a lot of structure in this case. The Freiman-Ruzsa Theorem (which we quote in a a version implied by [@EveZoh12]), gives such structure.
Let $A \subseteq {\mathbb{F}}_2^{n}$ be a set with $|2A| = K|A|$. Then, $\operatorname*{span}(A) \leq 2^{2K}|A|$.
The exponential loss in $K$ is necessary, as can be seen by considering the set $A \subseteq {\mathbb{F}}_2^{n}$, which contains all $a$ with exactly one $1$ in the first $t$ positions for an appropriately chosen $t$. In fact, the theorem given in [@EveZoh12] describes $\operatorname*{span}(A)/|A|$ correctly in dependence of $|2A|/|A|$.
One can conjecture that the exponential loss is unneccessary if one is allowed to take a large subset of $A$. This is called the Polynomial Freiman-Ruzsa Conjecture, and seems to be due to Marton (see [@Ruzsa93]).
\[con:pfr\] There exists $c \in {\mathbb{N}}$ such that, for any set $A \subseteq {\mathbb{F}}_2^{n}$ with $|2A| = K|A|$ there is a subset $A' \subseteq A$ such that $|A'| \geq K^{-c} |A|$ and $\operatorname*{span}(A') \leq K^{c}|A|$.
The strongest result in this direction was given by Sanders [@Sanders12].
\[thm:sanders\] For any set $A \subseteq {\mathbb{F}}_2^{n}$ with $|2A| = K|A|$ there is a subset $A' \subseteq A$ such that $|A'| \geq 2^{-\log^{4}(K)} |A|$ and $\operatorname*{span}(A') |A|$.
The proof of Theorem \[thm:sanders\] is based on a result by Crook and Sisask [@CroSis10].
We remark that the validity of Conjecture \[con:pfr\] was proven in the special case that $A$ is a downset [@GreTao09].
Our Contributions
-----------------
For a condition $C$, we use ${\langle}C{\rangle}$ as Iverson bracket, i.e., ${\langle}C{\rangle}$ is $1$ if $C$ is satisfied, and 0 otherwise.
For $s \in 2A$, let $A(s) := A \cap (s+A) = \{a \in A :
\exists a' \in A: s = a+a'\}$. Because $\sum_{s\in 2A} |A(s)| =
\sum_{s \in 2A} \sum_{a_1 \in A} {\langle}\exists a_2: a_1 + a_2 = s{\rangle}=\sum_{s \in 2A} \sum_{a_1,a_2 \in A} {\langle}a_1 + a_2 = s{\rangle}=\sum_{a_1,a_2 \in A}1 = |A|^2$ we see that $\operatorname*{E}_{s \in 2A}[A(s)] = |A|/K$.
We show that the PFR conjecture holds for two special cases. In Theorem \[thm:unstructured\], we show that in case most (namely, a fraction $1-K^{-9}$) of the sets $A(s)$ have size close to the minimum they can have on average (i.e., $|A|/K$), then the PFR conjecture holds. In Theorem \[thm:structured\], we prove that if a large fraction of the sets $A(s)$ have size close to the maximum they can have (namely $|A|$), then the PFR conjecture holds.
For both theorems it is easy to find examples of sets $A$ which satisfy the conditions of the theorem.
Preliminaries
-------------
We use the Balog-Szemerédi-Gowers Theorem, whose formulation we take from [@BeLoZe12].
\[thm:bsg\] There exist polynomials $f(x,y)$, $g(x,y)$ such that the following holds for every subset $A$ of an abelian group $G$. If $A$ satisfies $\Pr_{a_1,a_2 \in A}[a_1 + a_2 \in S] \geq \frac{1}{K}$ for some $S\subseteq G$, $|S| \leq C|A|$, then there is $A' \subseteq A$ such that $|A'| \geq |A|/f(K,C)$ and $|A'+A'| \leq g(K,C)|A|$.
The Unstructured Case
=====================
\[thm:unstructured\] Let $A \subseteq {\mathbb{F}}_2^n$, $|A+A| = K|A|$. If $$\begin{aligned}
\label{eq:2}
\Pr_{a_1,a_2\in A}
\Bigl[|A(a_1+a_2)| \geq \frac{L|A|}{K}\Bigr] \leq \frac{1}{K^{8}}\end{aligned}$$ then there exists $B \subseteq A$ with $|B| \geq |A|/{\mathrm{poly}}(L)$ such that $\operatorname*{span}(B) \leq |A+A|2^{{\mathrm{poly}}(L)}$.
We note that statement (1) in the abstract is implied by Theorem \[thm:unstructured\]. To see this, we observe that $\Pr_{a_1,a_2\in A}
\bigl[|A(a_1+a_2)| \geq \frac{L|A|}{K}\bigr]
\leq
K\cdot \Pr_{s \in 2A}
\bigl[|A(s)| \geq \frac{L|A|}{K}\bigr]$ because $\Pr_{a_1,a_2 \in A}[a_1 + a_2 = s] \leq \frac{1}{|A|} \leq
K \Pr_{s' \in 2A}[s = s']$ for any $s$.
The idea of the proof of Theorem \[thm:unstructured\] is to pick values $\beta_1,\beta_2,\beta_3,\beta_4$ uniformly at random from $A$, and show that (\[eq:2\]) implies that with probability depending only on $L$ we have $$\begin{aligned}
\label{eq:3}
A(\beta_1+\beta_2) \cap A(\beta_3+\beta_4) \neq \emptyset.\end{aligned}$$ Suppose now that indeed (\[eq:3\]) holds, and let $a$ be in the intersection. Then, because $\beta_1+\beta_2 = a+a_1$ and $\beta_3+\beta_4 = a+a_2$, we see that $\beta_1+\beta_2+\beta_3+\beta_4 = a_1 + a_2 \in 2A$. Thus, in this case we get that $\Pr_{s_1,s_2 \in 2A}[s_1+s_2 \in 2A]$ only depends on $L$ (actually this is not quite true, as the distribution is somewhat wrong, but we ignore this for this informal discussion). Thus, Theorem \[thm:bsg\] gives us a subset of $2A$ whose doubling constant only depends on $L$. By the Freiman-Ruzsa Theorem, the span of this subset is as small as required, and with somewhat more work we can get a large subset of $A$ which also has small span.
In the remainder of this section, we show that (\[eq:3\]) happens with probability dependent only on $L$. After this, we make the above intuition formal to get the result.
To prove that (\[eq:3\]) happens with probability dependent only on $L$, we define the random variable $$\begin{aligned}
Y &:= |A(\beta_1+\beta_2) \cap A(\beta_3+\beta_4)|\nonumber\\
&\phantom{:}=
\sum_{a \in A} {\langle}a \in A(\beta_1+\beta_2){\rangle}{\langle}a \in A(\beta_3+\beta_4){\rangle}.\end{aligned}$$ and use $\Pr[Y>0] \geq \frac{\operatorname*{E}^2[Y]}{\operatorname*{E}^2[Y]}$.
For technical reasons (the “somewhat more work” above), we later need actually that not only (\[eq:3\]) happens, but at the same time the sets $A(\beta_1+\beta_2)$ and $A(\beta_3+\beta_4)$ are not too small. Because of this, we actually work with the the random variable $Z$: $$\begin{aligned}
Z := \begin{cases}
0 & \text{if $|A({\beta_1+\beta_2})| \leq \frac{|A|}{2K}$ or
$|A({\beta_3+\beta_4})| \leq \frac{|A|}{2K}$}\\
Y & \text{otherwise}.
\end{cases}\end{aligned}$$ We will show that (\[eq:2\]) implies $\Pr[Z>0] \geq \Omega(L^{-4})$. For this, we first show that a simple application of Cauchy-Schwarz gives a good lower bound on $\operatorname*{E}[Z]$.
\[lem:lowerBoundOnZ\] If $A \subseteq {\mathbb{F}}_2^n$, $|A+A| = K|A|$, then $\operatorname*{E}[Z] \geq \frac{|A|}{16 K^2}$ (where the expectation is over the uniform random choice of $\beta_1$ to $\beta_4$ in $A$).
We see that $$\begin{aligned}
\operatorname*{E}[Z] &=
\operatorname*{E}\Bigl[ \sum_{a \in A}
{\langle}a \in A({\beta_1+\beta_2}) \land |A({\beta_1+\beta_2})|\geq\tfrac{|A|}{2K} {\rangle}\times\nonumber\\
&\qquad\qquad\qquad\qquad
{\langle}a \in A({\beta_3+\beta_4}) \land |A({\beta_3+\beta_4})|\geq\tfrac{|A|}{2K}{\rangle}\Bigr]
\\
& =
\sum_{a \in A}
\Pr\Bigl[a \in A({\beta_1+\beta_2}) \land |A({\beta_1+\beta_2})|\geq\frac{|A|}{2K}\Bigr]^2\\
&\geq
\frac{
\Bigl(\sum_{a \in A}
\Pr[a \in A({\beta_1+\beta_2}) \land |A({\beta_1+\beta_2})|\geq\frac{|A|}{2K}]\Bigr)^2}{|A|}\label{eq:1}\end{aligned}$$
We now observe that $$\begin{aligned}
\Pr_{\beta_1,\beta_2 \in A}\Bigl[|A({\beta_1+\beta_2})| \leq \frac{|A|}{2K}\Bigr]
\leq
\frac{|A+A| \frac{|A|}{2K}}{|A|^2} = \frac12.\end{aligned}$$ and so $$\begin{aligned}
\MoveEqLeft{\sum_{a \in A}
\Pr\Bigl[a \in A({\beta_1+\beta_2}) \land |A({\beta_1+\beta_2})|
\geq \tfrac{|A|}{2K}
\Bigr]}\\
&\geq\frac{1}{|A|^2}
\sum_{\beta_1,\beta_2: |A({\beta_1+\beta_2})| \geq |A|/2K}
\sum_{a} {\langle}a \in A({\beta_1+\beta_2}){\rangle}\\
&\geq\frac{1}{|A|^2}
\sum_{\beta_1,\beta_2: |A({\beta_1+\beta_2})| \geq |A|/2K}
\frac{|A|}{2K}
\geq \frac{|A|}{4K}\end{aligned}$$ Inserting into (\[eq:1\]) gives the lemma.
We next prove that $\operatorname*{E}[Z^2]$ is small; of course, it suffices to show that $\operatorname*{E}[Y^2]$ is small.
\[lem:YsquareIsSmall\] Let $A \subseteq {\mathbb{F}}_2^n$, $|A+A| = K|A|$. Suppose that $$\begin{aligned}
\Pr_{a_1,a_2 \in A}\Bigl[|A(a_1+a_2)| \geq \frac{L|A|}{K}\Bigr] \leq
\frac{1}{K^{8}}.\end{aligned}$$ Then, $E[Y^2] \leq 6\frac{L^4|A|^2}{K^4}$.
We sketch the proof of Lemma \[lem:YsquareIsSmall\]: linearity of expectation gives $$\begin{aligned}
\operatorname*{E}[Y^2] &= \operatorname*{E}\Bigl[\sum_{a_1,a_2 \in A} {\langle}\{a_1,a_2\} \subseteq A({\beta_1+\beta_2}){\rangle}{\langle}\{a_1,a_2\} \subseteq A({\beta_3+\beta_4}){\rangle}\Bigr]\nonumber\\
&=
\sum_{a_1,a_2 \in A} \Pr[\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})]^2 \label{eq:4}\end{aligned}$$ In order to give an upper bound on the probabilities in this expression, the following lemma is key.
\[lem:keyStep\] For any $A \subseteq {\mathbb{F}}_2^n$ and any $a_1,a_2 \in A$ we have $$\begin{aligned}
&\bigl|\{(\beta_1,\beta_2) \in A\times A : \{a_1,a_2\} \subseteq A({\beta_1+\beta_2})\}\bigr|\nonumber\\
&\qquad =
\bigl|\{(c_1,\beta_2) \in A \times A : c_1 \in A({a_1+a_2}) \land \beta_2 \in A({a_1+c_1})\bigr|\;.\label{eq:222}\end{aligned}$$
We show that the map $(\beta_1,\beta_2) \mapsto (a_1+\beta_1+\beta_2,\beta_2)$ is a bijection of the two sets. Clearly, the map is bijective as a map in ${\mathbb{F}}_2^{*}$.
Now, suppose that $(\beta_1,\beta_2)$ is in the first set. Let $c_1 := a_1+\beta_1+\beta_2$, and also set $c_2 := a_2+\beta_1+\beta_2$. We see that both $c_1$ and $c_2$ are in $A$ (because $a_1 \in A({\beta_1+\beta_2})$, and $a_2 \in A({\beta_1+\beta_2})$). Since also $c_1+c_2 = a_1+a_2$ we see that $c_1 \in A({a_1+a_2})$. Furthermore, because $\beta_1+\beta_2 = a_1+c_1$ we see that $\beta_2 \in A({a_1+c_1})$. Thus, $(a_1+\beta_1+\beta_2,\beta_2)$ is in the second set.
Next, suppose that $(c_1,\beta_2)$ is in the second set. Let $\beta_1 := a_1+\beta_2+c_1$, so that the preimage of $(c_1,\beta_2)$ is $(\beta_1,\beta_2)$. We first notice that $\beta_1 = a_1+\beta_2+c_1 \in A$ because $\beta_2 \in A({a_1+c_1})$. Next, we clearly have $a_1 \in A({\beta_1+\beta_2}) = A({a_1+c_1})$. Finally, we also have $a_2 \in A({a_1+c_1})$: this is the same condition as $c_1 \in A({a_1+a_2})$.
Continuing our proof sketch of Lemma \[lem:YsquareIsSmall\], we suppose for a moment that $|A(s)| \leq L|A|/K$ for all $s \in 2A$. Then, the set on the right hand side in (\[eq:222\]) is clearly at most of cardinality $L^2|A|^2/K^2$, and by (\[eq:4\]) we see that $\operatorname*{E}[Y^2] \leq L^4|A|^2/K^4$, as we want to show.
Thus, all which remains to do is to show that in case $|A(s)| \leq L|A|/K$ with probability $1-K^{-8}$ we can still give the bound; we have chosen the $8$ in the exponent to make this possible.[^1]
We start with $$\begin{aligned}
\label{eq:5}
\operatorname*{E}[Y^2] = \sum_{a_1,a_2} \Pr[\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})]^2\end{aligned}$$ and $$\begin{aligned}
\Pr[\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})]
=
\frac{1}{|A|^2} \sum_{\beta_1,\beta_2 \in A} {\langle}\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})\}{\rangle}\;.\end{aligned}$$ By Lemma \[lem:keyStep\], $$\begin{aligned}
\Pr[\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})]
=\frac{1}{|A|^2} \sum_{c_1 \in A({a_1+a_2})}|A({a_1+c_1})|\;.\end{aligned}$$
For $a \in A$ let $B(a) := \{ b \in A : |A({a+b})| \geq L|A|/K\}$. Then, we let $A'$ be the set of elements for which $|B(a)|$ is large. $$\begin{aligned}
A'= \{ a : |B(a)| \geq |A|/K^4 \}\end{aligned}$$ Note that $|A'| \leq |A|/K^4$, as otherwise $\Pr_{a_1,a_2}[|A({a_1+a_2})| \geq L|A|/K] \geq \frac{1}{K^{8}}$.
To now upper bound the right hand side of (\[eq:5\]), we first fix $a_1 \in A \setminus A'$ and $a_2 \in A \setminus B(a_1)$. Then, we have $$\begin{aligned}
\lefteqn{|A|^4 \cdot \Pr[\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})]^2}\\
&=
\Bigl(
\sum_{c_1 \in A({a_1+a_2}) \cap B(a_1)} |A({a_1+c_1})|
+ \sum_{c_1 \in A({a_1+a_2}) \setminus B(a_1)} |A({a_1+c_1})|
\Bigr)^2\\
&\leq
\Bigl(
\sum_{c_1 \in B(a_1)} |A|
+ \sum_{c_1 \in A({a_1+a_2})} \frac{L|A|}{K}
\Bigr)^2\\
&\leq
\Bigl(
|B(a_1)| |A|
+ |A({a_1+a_{2}})| \frac{L|A|}{K}
\Bigr)^2\\
&\leq
\Bigl(
\frac{|A|^2}{K^4}
+ \frac{L^2|A|^2}{K^2}
\Bigr)^2 \leq \frac{4L^4|A|^4}{K^4}\;.\end{aligned}$$
Thus, $$\begin{aligned}
\sum_{a_1,a_2 \in A} \Pr[&\{a_1,a_2\} \subseteq A({\beta_1+\beta_2})]^2\\
&=
\sum_{a_1 \notin A', a_2 \notin B(a_1)} 4\frac{L^{4}}{K^4}
+
\sum_{a_1 \notin A', a_2 \in B(a_1)} 1
+
\sum_{a_1 \notin A', a_2 \in A} 1\\
&\leq
4|A|^2\frac{L^4}{K^4} + |A|^2 \frac{1}{K^4} +
|A|^2 \frac{1}{K^4}
\\
&\leq
6|A|^2\frac{L^4}{K^4}\;.\end{aligned}$$
We can now prove Theorem \[thm:unstructured\].
Let $C \subseteq 2A$ be the elements $s$ of $2A$ for which $A(s)$ has “typical” size, formally $C := \{s : \frac{|A|}{2K} \leq |A(s)| \leq L \frac{|A|}{K}\}$. The total number of pairs $(a_1,a_2) \in A\times A$ is at most $\frac{|2A|}{K^8} \cdot |A| +
|C|\cdot L\frac{|A|}{K} + |2A|\cdot \frac{|A|}{2K}
=
|A|^2 \Bigl(\frac{1}{K^7} + L\frac{|C|}{|2A|} + \frac{1}{2}\Bigr)$. Because the number of pairs equals $|A|^2$, and we can assume $K \geq 2$ (otherwise the theorem is easily seen to hold), we have $L\frac{|C|}{|2A|} + \frac{3}{4} \geq 1$, i.e., $\frac{|C|}{|2A|} \geq \frac{1}{4L}.$
By Lemmas \[lem:lowerBoundOnZ\] and \[lem:YsquareIsSmall\] we have $$\begin{aligned}
\label{eq:6}
\Pr[Z >0]
\geq \frac{(E[Z])^2}{E[Z^2]}
\geq \frac{(E[Z])^2}{E[Y^2]} \geq
\Omega\Bigl(\frac{1}{L^4}\Bigr)\;.\end{aligned}$$
We want to count the number of tuples $(\beta_1,\beta_2,\beta_3,\beta_4)$ such that $\beta_1+\beta_2+\beta_3+\beta_4 \in 2A$, and further $\beta_1+\beta_2 \in C$ and $\beta_3+\beta_4 \in C$. By (\[eq:6\]) there are at least $\Omega(\frac{1}{L^4}) |A|^4$ tuples with $\beta_1+\beta_2+\beta_3+\beta_4 \in 2A$ and $|A(\beta_1+\beta_2)|, |A(\beta_1+\beta_2)| \geq \frac{|A|}{2K}$. Of those, at most $\frac{|A|^4}{K^8}$ also satisfy $|A(\beta_1+\beta_2)| \geq \frac{L|A|}{K}$, and so there are still at least $\Omega(\frac{1}{L^4}) |A|^4$ tuples with $\beta_1+\beta_2+\beta_3+\beta_4 \in 2A$, $\beta_1+\beta_2 \in C$, and $\beta_3+\beta_4 \in C$ (we can assume that $L \leq K/2$, as otherwise the theorem is trivial).
This implies that there are at least $\Omega(\frac{1}{L^6})
K^2 |A|^2$ pairs $(\gamma_1,\gamma_2) \in C \times C$ which satisfy $\gamma_1 + \gamma_2 \in 2A$: each such pair can be split into at most $\bigl(\frac{L|A|}{K}\bigr)^2$ four-tuples.
Thus, if we pick $\gamma_1$ and $\gamma_2$ uniformly in $C$ we have $$\begin{aligned}
\Pr[\gamma_1 + \gamma_2 \in 2A] \geq \Omega\Bigl(\frac{1}{L^{6}}\Bigr)\;.\end{aligned}$$
By the Balog-Szemerédi-Gowers Theorem (Theorem \[thm:bsg\]) applied on the set $A_{\mathrm{Thm\ref{thm:bsg}}}=C$ and $S_{\mathrm{Thm\ref{thm:bsg}}}=2A$ we get $C' \subseteq C$ such that $|C'| \geq |C|/{\mathrm{poly}}(L)$ and $|C'+C'| \leq {\mathrm{poly}}(L)|C|$. By the Freiman-Ruzsa Theorem we get that $\operatorname*{span}(C') \leq |C| 2^{{\mathrm{poly}}(L)}$.
Consider now the graph $G$ whose vertices are the elements of $A$, and where we $a_1$ and $a_2$ is connected by an edge iff $a_1+a_2 \in C'$. If two vertices $a_1$ and $a_2$ are in the same connected component of this graph, then $a_1+a_2$ is in $\operatorname*{span}(C')$: $a_1+a_2$ equals the sum of the elements in $C'$ which correspond to the edges of a path from $a_1$ to $a_2$.
The graph has at least $|A|^2/{\mathrm{poly}}(L)$ edges, because each element in $C'$ gives rise to at least $\frac{|A|}{2K}$ edges, and $|C'| \geq
\frac{K|A|}{{\mathrm{poly}}(L)}$. Thus, the graph must have a connected component with at least $|A|/{\mathrm{poly}}(L)$ vertices (if the largest connected component takes fraction $\delta$ of the vertices, we get at most $\delta^2 \cdot 1/\delta$ fraction of the edges). Let $A' \subseteq A$ be the elements of such a component. Because $2A' \subseteq \operatorname*{span}(C')$ we also have $\operatorname*{span}(2A') \subseteq
\operatorname*{span}(\operatorname*{span}(C')) = \operatorname*{span}(C')$. Furthermore, for every fixed $a' \in A'$ we have $\operatorname*{span}(A') = \operatorname*{span}(\{a'\} \cup 2A') \leq 2 \cdot |\operatorname*{span}(2A')| \leq
2\cdot |\operatorname*{span}(C')| \leq |C|\cdot 2^{{\mathrm{poly}}(L)}$.
The Structured Case
===================
\[thm:structured\] Let $A \subseteq {\mathbb{F}}_2^n$, $|A+A|= K|A|$. Suppose that there is an element $a^*$, $L \in {\mathbb{N}}$, and $\epsilon > 10/\log(K)$ such that the set $$\begin{aligned}
B = \{b \in A : |A({a^*+b}) | > |A|(1-K^{-\epsilon})\}\end{aligned}$$ has size at least $K^{-L} |A|$. Then, $B$ has span at most $2^{(4^{L/\epsilon})}|A+A|$.
Define $$\begin{aligned}
S(\delta) := \{s \in A+A: |A(s)| > |A|(1-\delta)\}\;.\end{aligned}$$ Clearly, $S(1) = |A+A|$. Furthermore, we see that $B+B \subseteq S(2K^{-\epsilon})$, and also $S(\delta)+S(\delta) \subseteq S(2\delta)$.
Consider the sequence $2K^{-\epsilon}, 2\cdot2K^{-\epsilon}, 4\cdot2K^{-\epsilon},
\ldots, 2^{r}K^{-\epsilon}$ where $r$ is chosen such that the last term is $(\frac{1}{2},1]$. We see that $r = \lceil\log_2(K^\epsilon/2)\rceil \geq \frac{\epsilon\log_2(K)}{2}$.
Since $K^{-L}|A| \leq |S(2K^{-\epsilon})|$ and $|S(2^{r}K^{-\epsilon})|
\leq |A+A| \leq K|A|$, there is some $j \in \{1,\ldots,r-1\}$ for which $$\begin{aligned}
|S(2^{j+1} K^{-L})| \leq K^{(L+1)/(r-1)}
|S(2^{j}K^{-L})|\;,\end{aligned}$$ which implies that $S(2^j K^{-L})$ has doubling constant at most $K^{(L+1)/(r-1)}$.
Since $K^{(L+1)/(r-1)} \leq 2^{2L\log_2(K)/\epsilon\log_2(K)} = 4^{L/\epsilon}$, by the Freiman-Ruzsa theorem, $S(2^{j}K^{-L})$ has span at most $2^{2\cdot 4^{L/\epsilon}}|S(2^j K^{-L})| \leq
2^{2\cdot 4^{L/\epsilon}}|A+A|$. This in particular means that $2B$ has at most this span. Since for any $b \in B$ we have that $b+B \subseteq 2B$, we see that $\operatorname*{span}(B) \leq 2 \operatorname*{span}(2B)$, and so we get the result.
Finally, we note that the theorem implies the statement given by item (2) in the abstract: since $\Pr_{a_1,a_2 \in A}[|A(a_1+a_2)| \geq |A|(1-K^{-1/L})] \geq
K^{-L-1}$, there must be some $a_1$ for which $\Pr_{a_2\in A}[|A(a_1+a_2)| \geq |A|(1-K^{-1/L})] \geq K^{-L-1}$. Applying Theorem \[thm:structured\] gives the claim, as long as $L$ is smaller than some function of $K$, which (using the Freiman-Ruzsa Theorem again) is clearly sufficient.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank Eli Ben-Sasson and Noga Ron-Zewi for inspiring discussions and helpful pointers.
[^1]: We remark that the assumption $\forall s: |A(s)| \leq L|A|/K$ does not really make sense, because $A(0) = A$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quaternion-valued solutions to the non-commutative KdV equation are produced using determinants. The solutions produced in this way are (breather) soliton solutions, rational solutions, spatially periodic solutions and hybrids of these three basic types. A complete characterization of the parameters that lead to non-singular 1-soliton and periodic solutions is given. Surprisingly, it is shown that such solutions are never singular when the solution is essentially non-commutative. When a 1-soliton solution is combined with another solution through an iterated Darboux transformation, the result behaves asymptotically like a combination of different solutions. This “non-linear superposition principle” is used to find a formula for the phase shift in the general 2-soliton interaction. A concluding section compares these results with other research on non-commutative soliton equations and lists some open questions.'
address: 'Department of Mathematics, College of Charleston, Charleston SC 29401'
author:
- 'John Cobb, Alex Kasman, Albert Serna, and Monique Sparkman'
title: 'Quaternion-Valued Breather Soliton, Rational, and Periodic KdV Solutions'
---
-.5in.5in-1in-1in1.21.2
Introduction
============
The KdV Equation
----------------
The Korteweg-deVries (KdV) Equation $$u_t=\frac{3}{2}uu_x+\frac{1}{4}u_{xxx}\label{eqn:stdKdV}$$ was originally derived in order to better understand the solitary waves observed in 1834 by John Scott Russell on Union Canal in Scotland [@KdV; @KdV-hist; @GOST]. It is unusual among nonlinear partial differential equations in that it is completely integrable and so it is possible to write many of its solutions in closed form. Moreover, among those solutions are the multi-soliton solutions that behave asymptotically like localized disturbances traveling at constant speeds which exhibit a phase shift upon interaction [@ZK; @BKY]. Even among other completely integrable differential equations with soliton solutions, the KdV Equation holds a special place because it was historically the first one recognized as having these properties.
In the case that $u(x,t)$ takes values in some non-commutative algebra, a natural generalization of the KdV Equation is the symmetrized form: $$u_t=\frac{3}{4}uu_x+\frac{3}{4}u_xu+\frac{1}{4}u_{xxx}.\label{eqn:KdV}$$
The purpose of this paper is to carefully study certain quaternion-valued solutions to . Although solutions to integrable equations such as KdV have been previously explored both in more general non-commutative settings [@EGR; @NC1; @NC2; @NC3] and in the quaternionic case [@QKdV], the specific breather soliton, rational and periodic solutions investigated below, their construction in terms of quaternionic determinants, and their nonlinear superpositions have not previously been described.
Quaternions
-----------
The quaternions were first studied by William Rowan Hamilton as a number system which generalized the complex numbers [@Hamilton]. Although their non-commutativity was a novelty in 1844, since the quaternions can be embedded into a matrix group, they may not seem particularly interesting to a modern mathematical physicist. For many years they were seen as being “old-fashioned”, merely a historical stepping stone on the way to more general non-commutative algebras. However, recently they have received an increasing amount of attention in relation to differential equations and dynamical systems [@QDiff1; @QDiff2; @QDiff3; @QDiff4], for their uses in mathematical physics and engineering [@QMP1; @QMP2; @QMP3; @QMP4], and even for their unique algebraic structure [@Chen; @QMatInv; @Kyrchei]. This resurgence of interest in the quaternions shows that some important properties of quaternionic solutions are not immediately evident when they are viewed in the more general context of matrix algebras and justifies the current investigation into the quaternion-valued solutions of the KdV equation.
This section will briefly review some key properties of the quaternions and set up the terminology and notation to be used in the remainder of the paper. For additional information, readers should consult References [@quaternion1; @quaternion2].
### Notation and Arithmetic
The quaternions are the 4-dimensional real vector space $$\H=\{q_0+q_1\qi+q_2\qj+q_3\qk\ :\ q_i\in\R\}$$ with multiplication satisfying the usual distributive and associative laws along with the identities $$\qi^2=\qj^2=\qk^2=\qi\qj\qk=-1$$ which William Rowan Hamilton famously carved into Brougham Bridge in 1843. However, the multiplication is not commutative because $$\qi\qj=-\qj\qi,\ \qj\qk=-\qk\qj,\ \hbox{and}\ \qi\qk=-\qk\qi.$$
If a letter is used to index a quaternion, then the subscripts $0$, $1$, $2$, and $3$ on the same letter will denote the real numbers which are its coefficients relative to the basis $\{1,\qi,\qj,\qk\}$ of $\H$. The length of a quaternion $q\in\H$ is defined to be $|q|=\sqrt{q_0^2+q_1^2+q_2^2+q_3^2}$, its quaternionic conjugate is $q^*=q_0-q_1\qi-q_2\qj-q_3\qk$, and if $q\not=0$ then it has a unique multiplicative inverse $
q^{-1}=\frac{1}{|q|}q^*
$.
It is often convenient to separate a quaternion $q$ into its real part $q_0$ and vector part $\vec q=q_1\qi+q_2\qj+q_3\qk$ where the latter is thought of as being an element of $\R^3$. Then, conjugating one element of $\H$ by another has an interpretation as a *rotation* in 3-dimensional space in the following sense: If $q$ and $g\not=0$ are quaternions then $q$ and $
r=gqg^{-1}
$ are related by the facts that $q_0=r_0$ and that $\vec r\in\R^3$ is a vector obtained by rotating $\vec q$ through an angle depending only on the choice of $g$. This sets up a well-known correspondence between unit quaternions and rotations by which every rotation corresponds to either of two quaternions of length $1$. The details of this correspondence will not be needed in this paper. However, a certain consequence of its existence will be:
\[prop:conj\] Two quaternions $q$ and $r$ satisfy $r=gqg^{-1}$ for some quaternion $g$ if and only if $q_0=r_0$ and $|\vec q|=|\vec r|$.
Exponential functions involving quaternions will be needed later in this paper. It is therefore useful to note that by writing $e^q$ as a power series one can easily show that $$e^q=e^{q_0}\left(\cos(|\vec q\,|)+\frac{\sin(|\vec q\,|)}{|\vec q\,|}\vec q\right)\label{eqn:exp}$$ for any quaternion $q=q_0+\vec q$ with non-zero vector part. Moreover, if $q$ and $r$ are quaternions that commute (i.e. $[q,r]=0$) then $
e^{q}e^{r}=e^{q+r}
$.
### Quaternion-Valued Functions
Throughout the remainder of this paper, $x$ and $t$ will be real-valued variables and functions of these variables will take values in $\H$. Together, such quaternion-valued functions will be taken to form a right module over the quaternions. (Hence, any reference to linear combinations of such functions will be considered to be a sum of functions with quaternionic coefficients on the right.)
When a quaternion-valued function $f(x,t)$ is to be represented graphically, it will be illustrated by graphing each component function $f_i(x,t)$ ($0\leq i
\leq 3$) separately on the same set of axes for some fixed value of $t$. So, for instance, the function $f(x,t)=(x+t)^{-2}+\sin(x)\qi + \cos(t)\qj+(x^2+1)^{-1}\qk$ is defined for all $(x,t)\in\{(x,t)\in R^2\ :\ x\not=-t\}$ and its graphical representation at a fixed value of $t$ would look like the superposition of four graphs, one having a pole at $x=-t$, a trig function, a horizontal line, and a function with one “peak” at $x=0$.
### Determinants of Quaternionic Matrices {#sec:Chen}
Interestingly, although there is no useful generalization of the determinant to arbitrary non-commutative settings[^1], there *are* definitions for a determinant of a square matrix of quaternions that generalizes the usual determinant and have corresponding Cramer-like theorems [@Chen; @Kyrchei]. By setting up notation and summarizing some prior results, this section lays the foundation for the construction of quaternion-valued KdV solutions using these determinants in Theorem \[thm:kdv\].
\[def:cycles\] Let $S_n$ denote the group of permutations on $n$ elements. A permutation $\sigma\in S_n$ is a cycle $$\sigma=(c_1c_2\cdots c_k)$$ if $\sigma(j)=j$ for $j\not\in\{c_1,\ldots,c_k\}$ and $\sigma(c_i)=c_{i+1}$ for $i<k$ and $\sigma(c_k)=c_1$. If a permutation $\sigma$ in the group $S_n$ of permutations on the set $\{1,\ldots,n\}$ is a cycle, then it can be written in the normalized form $\sigma=(c_1c_2\cdots c_k)$ where $c_1>c_j$ for $j>1$. Any permutation $\sigma\in S_n$ has a unique factorization into normalized cycles $$\sigma=\sigma_1\cdots\sigma_r$$ where for each $j$ one has $\sigma_j=(c^j_1c^j_2\cdots)$ and $c^j_1>c^{j+1}_1$ (i.e. the sequence of first terms in the cycles is decreasing) and where each element of $\{1,\ldots,n\}$ appears exactly once (which requires including cycles of length $1$ for fixed points of $\sigma$).
\[def:MatrixStuff\] Let $\Mat=[\mat_{ij}]=[\vec \mat_i]$ be an $n\times n$ matrix with entries $\mat_{ij}$ from some non-commutative ring and column vectors $\vec \mat_i$. We denote by $\Mat_{\langle i\rangle}$ the matrix obtained by exchanging the $i^{th}$ and $n^{th}$ columns of $\Mat$: $$\Mat_{\langle i\rangle}=\left[\vec \mat_1\ \vec \mat_2\ \cdots
\vec \mat_{i-1}\ \vec
\mat_n\ \vec\mat_{i+1}\ \cdots\ \vec\mat_{n-1}\ \vec \mat_{i}\right]$$ and let $\Mat_{\langle i,j\rangle}$ denote the matrix obtained by replacing the $i^{th}$ column of $\Mat$ by its $n^{th}$ column and replacing the $n^{th}$ column by the $n$-vector $\vec e_j$ whose only non-zero entry is a $1$ in the $j^{th}$ position: $$\Mat_{\langle i,j\rangle}=\left[\vec \mat_1\ \vec \mat_2\ \cdots
\vec \mat_{i-1}\ \vec
\mat_n\ \vec\mat_{i+1}\ \cdots\ \vec\mat_{n-1}\ \vec e_{j}\right]$$ For a cycle $\sigma=(c_{1}\cdots
c_{k})\in S_n$ the symbol $M_{\sigma}$ denotes the ordered product $$M_{\sigma}=m_{c_1c_2}m_{c_2c_3}\cdots m_{c_{k-1}c_{k}} m_{c_kc_1}.$$
\[def:chen\] For an $n\times n$ matrix $\Mat=[\mat_{ij}]$ whose elements are from some non-commutative ring, define the Chen Determinant $\cdet(\Mat)$ to be $$\cdet(\Mat)=\sum_{\sigma\in S_n}(-1)^{n-r}
M_{\sigma_1}M_{\sigma_2}\cdots M_{\sigma_r}$$ where for each permutation $\sigma=\sigma_1\cdots\sigma_r$ is the decomposition into normalized cycles in Definition \[def:cycles\] and $M_{\sigma_j}$ is defined in Definition \[def:MatrixStuff\].
Note that if the elements of the matrix mutually commute, then $\cdet(\Mat)=\det(\Mat)$ is the ordinary determinant of the matrix, but if they do not then this definition specifies a unique ordering of the factors. If the elements $\mat_{ij}\in\H$ are quaternions, then it is possible to solve the vector equation $\Mat v = w$ or to write the inverse matrix $\Mat^{-1}$ in terms of these Chen determinants [@Chen; @Kyrchei]. This construction involves not only the matrix $M$ but also its conjugate transpose $\Mat^{\dagger}=[\mat_{ji}^*]$.
\[prop:inv\] An $n\times n$ matrix $\Mat$ of quaternions is invertible if and only if the real number $\cdet(\Mat^{\dagger}\Mat)$ is non-zero. If it is, then the $(i,j)$ entry of the matrix $\Mat^{-1}$ is $$\Mat_{ij}^{-1}=\frac{1}{\cdet(\Mat^{\dagger}\Mat)}\cdet\!\left(\Mat_{\langle
i\rangle}^\dagger\Mat_{\langle i,j\rangle}\right).$$
\[rem:notnew\] Definition \[def:chen\] and Proposition \[prop:inv\] can be found in References [@Chen; @Kyrchei]. However, they have been rewritten in the notation set up by Definitions \[def:cycles\] and \[def:MatrixStuff\] into a form that is more convenient for their use in this paper.
Construction of Quaternion-Valued Solutions
===========================================
KdV-Darboux Kernels
-------------------
\[def:KdVDarbouxKernel\] Let $\F=\{\phi_1(x,t),\ldots,\phi_n(x,t)\}$ be a set of functions $\phi_i:\R^2\to\H$ depending on the real variables $x$ and $t$ and taking values in the set $\H$ of quaternions. We will call $\F$ a *KdV-Darboux Kernel* if it has the following properties:
- **Dispersion:** For each $1\leq i\leq n$, $\phi_i$ satisfies the linear equation $$\frac{\partial^3\phi_i}{\partial x^3}=\frac{\partial\phi_i}{\partial t}.\label{eqn:dispersion}$$
- **Closure:** For each $1\leq i\leq n$, the second derivative $(\phi_i)_{xx}$ is in $\span(\F)$, the right $\H$-module generated by $\F$: $$\frac{\partial^2\phi_i}{\partial x^2}\in\span(\F)=\left\{\sum_{j=1}^n\phi_j\alpha_j\ :\ \alpha_j\in\H\right\}.\label{eqn:closure}$$
- **Independence:** The $n\times n$ Wronskian matrix $$\Wr=[\wr_{ij}]\hbox{ with }
\wr_{ij}=\frac{\partial^{i-1}\phi_j}{\partial x^{i-1}}$$ satisfies $\cdet(\Wr^{\dagger}\Wr)\not\equiv0$ and hence is an invertible matrix by Proposition \[prop:inv\].
For the purposes of this paper, a differential operator of order $n$ is a polynomial $Q$ in the symbol $\partial$ of the form $$Q=\sum_{i=0}^n c_i(x,t)\partial^i$$ where $c_i(x,t)$ are meromorphic, infinitely differentiable functions from $\R^2$ to $\H$ with $c_n(x,t)\not=0$. We say that this operator is monic if $c_n\equiv1$. These operators act on infinitely differentiable functions by the formula $$Q[f(x,t)]=\left(\sum_{i=0}^n c_i(x,t)\partial^i\right)[f]=\sum_{i=0}^n c_i(x,t)\frac{\partial^i f}{\partial x^i}.$$ The product $Q_1\circ Q_2$ of two differential operators is defined to coincide with their composition as operators: $Q_1\circ Q_2[f(x,t)]=Q_1\left[ Q_2[f(x,t)]\right]$.
KdV Solution Associated to a KdV-Darboux Kernel
-----------------------------------------------
\[thm:kdv\] Let $\F=\{\phi_1,\ldots,\phi_n\}$ be a KdV-Darboux Kernel with Wronskian matrix $\Wr$. Then the quaternion-valued function $$u_{\F}(x,t)=\left[\frac{2}{\cdet(\Wr^{\dagger}\Wr)}\sum_{i=1}^n\frac{\partial^n\phi_i}{\partial
x^n}\,\cdet\!\left(\Wr_{\langle i\rangle}^{\dagger}\Wr_{\langle i,n\rangle}\right)\right]_x
\label{eqn:uF}$$ is a solution to the non-commutative KdV Equation . In the special case that $\F=\{\phi\}$ contains only one element, the formula simplifies to $$u_{\F}(x,t)=2\phi_{xx}\phi^{-1}-2\phi_x\phi^{-1}\phi_x\phi^{-1}.\label{eqn:uFone}$$
As a consequence of the independence property of Definition \[def:KdVDarbouxKernel\], it follows from Theorem 3.6 in Reference [@KasOp] that $$K=\partial^n-\sum_{i=1}^n\sum_{j=1}^n\frac{\partial^n\phi_i}{\partial
x^n}\Wr
^{-1}_{ij}\partial^{j-1}.\label{eqn:preKF}$$ is the unique monic operator of order $n$ such that $\ker(K)=\span(\F)$. Using Proposition \[prop:inv\], equation can be rewritten in terms of Chen determinants as $$K=\partial^n-\frac{1}{\cdet(\Wr^{\dagger}\Wr)}\sum_{i=1}^n\sum_{j=1}^n\frac{\partial^n\phi_i}{\partial x^n}\cdet\!\left(\Wr_{\langle
i\rangle}^\dagger\Wr_{\langle i,j\rangle}\right)
.\label{eqn:KF}$$
The closure property of the KdV-Darboux kernel implies that each element of $\F$ is in the kernel of the operator $K\circ \partial^2$. Then, Theorem 5.1 in Reference [@KasOp] implies the existence of a differential operator $L$ satisfying the intertwining relationship $$K\circ \partial^2=L\circ K.\label{eqn:intertwining}$$ Setting up notation for the coefficients of the operators $K$ and $L$, let us write $$K=\partial^n+\sum_{i=0}^{n-1}c_i(x,t)\partial^i\qquad
\hbox{and}
\qquad
L=\partial^2+v(x,t)\partial+u_{\F}(x,t).$$ Equating coefficients on each side of one finds that $v(x,t)=0$ and $u_{\F}(x,t)=(-2c_{n-1})_x$. (N.B.That the potential in the Schrödinger operator $L$ is $-2$ times the $x$-derivative of the coefficient of $\partial^{n-1}$ in $K$ is a useful observation which will be referred to in several of the other proofs in this paper.) The formula for $u_{\F}(x,t)$ in the claim can then be recovered by isolating the coefficient $c_{n-1}$ from .
In the case where $\F$ contains only one element, it is clear that $K=\partial-\phi_x\phi^{-1}$ since this is a monic differential operator of order $1$ having $\phi$ in its kernel. But, by the argument above, this means that $u_{\F}=(2\phi_x\phi^{-1})_x$, which expands to the claimed formula.
All that remains is to demonstrate that $u_{\F}$ satisfies the KdV equation, a fact that follows from the dispersion property of the KdV-Darboux kernel using a standard technique in soliton theory which is only briefly outlined below.
Differentiating $K(\phi_i)=0$ with respect to $t$, using the dispersion relation to rewrite $t$ derivatives as $x$ derivatives and again applying Theorem 5.1 from Reference [@KasOp], one concludes that $\dot K+K\circ \partial^3=M\circ K$ for some differential operator $M$. Equating coefficients again one determines that $M=\partial^3+\frac{3}{2}u\partial+\frac{3}{4}u_x$. Furthermore, differentiating with respect to $t$ results in the Lax equation $\dot L=M\circ L -L\circ M$. Finally, expanding out these products of differential operators that Lax equation is seen to be equivalent to the non-commutative KdV equation .
\[rem:EGR\] This method of producing solutions to and the arguments in the proof are not very different from those used in the seminal paper by Etingof, Gelfand and Retakh [@EGR] where non-commutative solutions were produced using *quasi-determinants*. However, the formula in Theorem \[thm:kdv\] works for all KdV-Darboux kernels, even those for which the Wronskian matrix contains many zero entries which impose obstacles to computing the quasi-determinant. In addition, we wanted to take advantage of the extra algebraic structure of the quaternions that allows for solution of linear systems using the Chen determinant.
\[examp:ratl\] Let $\F$ be the KdV-Darboux kernel $$\F=\left\{x^3+x^2\qi+6t+\qk,6x+2\qi\right\}$$ and let $\Wr$ be its $2\times 2$ Wronskian matrix. There are two terms in the sum in formula , but the second term will be zero since it has the second derivative of $6x+2\qi$ as a factor. So, we multiply $$\cdet(\Wr_{\langle 1\rangle}^{\dagger}\Wr_{\langle 1,2\rangle})=
72 x^4-216 x t
+\left(-72 t-48 x^3-8 x
\right)\qi
+12
\qj
+36x
\qk$$ on the left by $(x^3+x^2\qi+6t+\qk)_{xx}=6x+2\qi$, and multiply that by the real-valued function $$\frac{2}{\cdet(\Wr^{\dagger}\Wr)}=\frac{1}{648 t^2-432 x^3 t+144 t
x+72 x^6+24 x^4+8 x^2+18}$$ to get $$\frac{-2592 x^2 t+288 t+864 x^5+192 x^3+32 x
+\left(-1728 xt-288 x^4-96 x^2\right)\qi+\left(432 x^2+48\right)\qk
}{324 t^2-216 x^3t+72 xt+36 x^6+12 x^4+4 x^2+9}$$ The solution $u_{\F}(x,t)$ of is the $x$-derivative of the expression above.
Lemmas Relating Different KdV-Darboux Kernels
---------------------------------------------
The map associating a KdV-Darboux kernel $\F$ to the corresponding solution $u_{\F}$ actually depends only on $\span(\F)$:
\[lem:samespan\]If $\F$ and $\hat\F$ are KdV-Darboux kernels which span the same right $\H$ module, then they produce the same KdV solution.
In particular, if $\F=\{\phi_1,\ldots,\phi_n\}$ and $\hat\F=\{\phi_1q_1,\ldots,\phi_nq_n\}$, for some $q_i\in\H$ such that $q_i\not=0$, then $
u_{\F}(x,t)=u_{\hat\F}(x,t)$.
Suppose $\F$ and $\hat\F$ are KdV-Darboux kernels such that $\span(\F)=\span(\hat\F)$. By the Independence property (cf. Definition \[def:KdVDarbouxKernel\]), we know that this is a *free* module and since $\H$ is a division ring, they have the same dimension, which we will call $n$. As can be seen in the proof of Theorem \[thm:kdv\], $u_{\F}=(-2c_{n-1})_x$ where $c_{n-1}$ is the coefficient of $\partial^{n-1}$ in the unique monic differential operator of order $n$ having the elements of $\F$ in its kernel. However, by assumption, the elements of $\hat\F$ are linear combinations of those elements of $\F$ with constant coefficients on the right and hence are also in the kernel of this same operator. Consequently, the same operator is associated to $\hat\F$ and the solution $u_{\hat\F}$ produced from it is also the same.
However, spanning the same right $\H$ module is not the only way two KdV-Darboux kernels can correspond to the same solution. The following lemma shows that they do not even have to have the same number of elements:
\[lem:dropone\] If $\F=\{\phi_1,\ldots,\phi_n\}$ is a KdV-Darboux kernel and $\phi_n=\alpha e^{\lambda x+\lambda^3t}$ for some $\alpha,\lambda\in\H$ then $u_{\F}(x,t)=u_{\QF}(x,t)$ where $$\QF=\{Q(\phi_1),\ldots,Q(\phi_{n-1})\}\qquad \hbox{and}\qquad
Q(f)=f_x-\alpha \lambda\alpha^{-1}f.$$
Let $\QK=\partial^{n-1}+\sum_{i=0}^{n-2}\Qc_i(x,t)\partial^i$ be the unique monic differential operator of order $n-1$ having the elements of $\QF$ in its kernel. Then we know from the proof of Theorem \[thm:kdv\] that $u_{\QF}=(-2c_{n-2})_x$.
Let $Q=\partial-\alpha \lambda \alpha^{-1}$ be the monic differential operator of order $1$ with $\phi_n$ in its kernel. Define $K=\QK\circ Q$ and note that $K=\partial^n+\sum_{i=0}^{n-1}c_n(x,t)\partial^i$ is a monic differential operator of order $n$. Now consider $$K(\phi_i)=\QK\circ Q(\phi_i)=\QK(Q(\phi_i)).$$ For $i<n$ it is zero since $\QK$ was constructed so that $Q(\phi_i)$ is in its kernel and for $i=n$ this is zero because $Q(\phi_n)=0$. Then $K$ must be the unique monic differential operator of order $n$ having the elements of $\F$ in its kernel and $u_{\F}=(-2c_{n-1})_x$.
Expanding the product $\QK\circ Q$ we find that the coefficient of $\partial^{n-1}$ is $c_{n-1}=\Qc_{n-2}-\alpha^{-1}\lambda\alpha$. Since the second term is constant and has derivative equal to zero, we conclude that $u_{\F}=(-2c_{n-1})_x=(-2\Qc_{n-2})_x=u_{\QF}$.
For example, for any quaternions $\alpha$ and $\lambda$ (with $\alpha\not=0$) the two-element KdV-Darboux kernel $\F=\{x,\alpha e^{\lambda x+\lambda^3t}\}$ and the single-element KdV-Darboux kernel $\hat\F=\{1-\alpha^{-1}\lambda\alpha x\}$ produce the same solution.
Finally, we note that multiplying every element of the KdV-Darboux kernel on the left by the same non-zero quaternion has the effect of *rotating* the corresponding solution:
\[lem:rot\] Let $\F=\{\phi_1,\ldots,\phi_n\}$ be a KdV-Darboux kernel, $q\in\H$ a non-zero quaternion, and $\hat\F=\{q\phi_1,\ldots,q\phi_n\}$. Then the solutions $u_{\F}(x,t)$ and $u_{\hat\F}(x,t)$ are related by the formula $$u_{\hat\F}=qu_{\F}q^{-1}.$$
Let $K$ be the monic differential operator of order $n$ having the elements of $\F$ in its kernel. Note that $\hat K=qKq^{-1}$ is a monic differential operator of order $n$ and that $\hat
K(q\phi_i)=qKq^{-1}(q\phi_i)=qK(\phi_i)=0$. Consequently, $\hat K$ is the unique monic differential operator of order $n$ having the elements of $\hat \F$ in its kernel. Letting $c_{n-1}(x,t)$ and $\hat
c_{n-1}(x,t)$ denote the coefficient of $\partial^{n-1}$ in $K$ and $\hat K$ respectively we have by the definition of $\hat K$ that $\hat
c_{n-1}=qc_{n-1}q^{-1}$. Then $$u_{\hat \F}=(-2\hat c_{n-1})_x=(-2q c_{n-1}q^{-1})_x=q(-2c_{n-1})_xq^{-1}=qu_{\F}q^{-1}.$$
Basic Solution Types
====================
There are three kinds of non-trivial quaternion-valued solutions to that can be produced using a KdV-Darboux kernel with one element: localized breather solitons, translating periodic solutions, and rational solutions.
$1$-Soliton and Translating Periodic Solutions {#sec:uabl}
----------------------------------------------
Section \[sec:uabl\] will consider the solutions associated to KdV-Darboux kernels of the form $\{\phi_{\alpha,\beta,\lambda}\}$ where $$\phi_{\alpha,\beta,\lambda}(x,t)=\alpha e^{\lambda x+\lambda^3t}+\beta
e^{-\lambda x-\lambda^3t}\label{eqn:phiabl}$$ for some choice of $\alpha$, $\beta$ and $\lambda$ in $\H$. For convenience, we will write $u_{\alpha,\beta,\lambda}(x,t)$ for the corresponding KdV solution $$u_{\alpha,\beta,\lambda}(x,t)=u_{\{\phi_{\alpha,\beta,\lambda}\}}.$$
In fact, it is not necessary to consider *all* possible combinations of quaternions $\alpha$, $\beta$ and $\lambda$. First, we will assume that $\alpha\beta\lambda \not=0$. This both guarantees that $\{\phi_{\alpha,\beta,\lambda}\}$ is a KdV-Darboux kernel (which fails to be the case when $\alpha=\beta=0$) and eliminates the cases in which $u_{\alpha,\beta,\lambda}(x,t)\equiv0$ is the trivial solution.
Furthermore, one can greatly restrict the selection of the parameter $\lambda$ without losing any corresponding KdV solutions.
\[lem:complexlambda\] Let $\alpha$, $\beta$ and $\lambda$ be quaternions such that $\alpha\beta \lambda\not=0$. Then there are quaternions $\hat\alpha$ and $\hat\beta$ and a complex number $\hat \lambda=\hat \lambda_0+\hat \lambda_1\qi$ with $\hat\lambda_0\geq 0$ and $\hat\lambda_1\geq0$ such that $$u_{\alpha,\beta,\lambda}(x,t)=u_{\hat \alpha,\hat\beta,\hat \lambda}(x,t).$$
Since $\phi_{\alpha,\beta,\lambda}=\phi_{\beta,\alpha,-\lambda}$ we may assume, without loss of generality, that $\lambda_0\geq0$.
Let $\lambda=\lambda_0+\vec \lambda$ be the decomposition of $\lambda$ into its real and vector parts. It will be shown that the same solution can be constructed using the complex number $\hat \lambda=\lambda_0+|\vec \lambda|\qi$ which has a non-negative real and imaginary components.
By Proposition \[prop:conj\], because $\lambda$ and $\hat \lambda$ have the same real part and vector parts of the same length, there is a non-zero quaternion $g$ which satisfies $$\hat \lambda=g\lambda g^{-1}.$$ Define $\hat\alpha=\alpha g^{-1}$ and $\hat \beta=\beta g^{-1}$. Now, note that $$\begin{aligned}
\phi_{\hat\alpha,\hat\beta,\hat \lambda}
&=&\hat \alpha e^{\hat \lambda x+\hat \lambda^3t}+\hat \beta e^{-\hat
\lambda x-\hat \lambda^3t}\\
&=& \alpha g^{-1} e^{g \lambda g^{-1} x+g\lambda^3g^{-1}t}+ \beta g^{-1} e^{- g \lambda
g^{-1}x-g \lambda^3g^{-1}t}\\
&=& \alpha g^{-1}(g e^{ \lambda x+\lambda^3t}g^{-1})+ \beta g^{-1}(g e^{- \lambda x-
\lambda^3t}g^{-1})\\
&=& \left(\alpha e^{ \lambda x+\lambda^3t}+ \beta e^{- \lambda x-
\lambda^3t}\right)g^{-1}
=\phi_{\alpha,\beta,\lambda}g^{-1}.\end{aligned}$$ It then follows from Lemma \[lem:samespan\] that they generate the same solutions.
Consequently, no non-trivial solutions will be lost by the fact that we will henceforth limit our attention only to the case in which $\alpha\beta\lambda \not=0$ and $\lambda=\lambda_0+\lambda_1\qi$ is a complex number with $\lambda_0,\lambda_1\geq 0$.
The real numbers $$\vr=\lambda_0^2-3\lambda_1^2\qquad \hbox{and}\qquad
\vi=3\lambda_0^2-\lambda_1^2$$ will then be useful in understanding the solution any of the solutions $u_{\alpha,\beta,\lambda}$ since by Equation \[eqn:exp\] $$e^{\lambda x+\lambda^3t}=e^{\lambda_0(x+\vr t)}\left(\cos(\lambda_1(x+\vi
t))+\sin(\lambda_1(x+\vi t))\qi\right).\label{eqn:velocities}$$ As one might guess from , $\vr$ and $\vi$ will play the role of two separate *velocities*. Considering the graph of $u_{\alpha,\beta,\lambda}$ as a function of $x$ with $t$ playing the role of a time parameter, the *periodic* features coming from the trigonometric functions will translate to the left with velocity $\vi$ while the localized soliton has a *center* which translates with velocity $\vr$.
The next two sections separately handle the cases $\lambda_0=0$ and $\lambda_0\not=0$ which are qualitatively very different.
### Translating Periodic Solutions
Consider the case in which $\lambda_0=0$ (so that $\lambda=\lambda_1\qi$ is a purely imaginary complex number). Then the corresponding solution to the KdV equation is a spatially periodic solution that translates at a constant speed in time.
If $\lambda=\lambda_1\qi$ then the associated KdV solution has a graph that is invariant under a horizontal translation in $x$ by $2\pi/\lambda_1$ units and viewing $t$ as a time parameter this periodic waveform translates to the right at constant speed $\lambda_1^2$.
Using Equation \[eqn:velocities\] and the fact that $\vi=3\lambda_0^2-\lambda_1^2=-\lambda_1^2$ one finds that $$\phi_{\alpha,\beta,\lambda}(x,t)=
(\alpha+\beta)\cos(\lambda_1(x-\lambda_1^2
t))+(\alpha-\beta)\sin(\lambda_1(x-\lambda_1^2 t))\qi.$$
Substituting this into and using trigonometric identities, one can determine that the corresponding KdV solution has the form $$u_{\alpha,\beta,\lambda}(x,t)=R(\cos(2\lambda_1(x-\lambda_1^2t)),\sin(2\lambda_1(x-\lambda_1^2t)))$$ where $R(\xi,\eta)$ is a certain rational function.
Since $u_{\alpha,\beta,\lambda}$ can be written as a function in $x-\lambda_1^2t$, we know that its graph as a function of $x$ will translate to the right with speed $\lambda_1^2$ in the time parameter $t$. And, since the translation $x\mapsto x+\pi/\lambda_1$ shifts the arguments of the trigonometric functions by $2\pi$, it leaves the graph unchanged.
![The periodic translating solution $u_{\alpha,\beta,\lambda}(x,t)$ with $\alpha=1+\qk$, $\beta=1$ and $\lambda=\qi$ at time $t=0$.[]{data-label="fig:periodic"}](periodic.eps){width="3in" height="2in"}
\[examp:periodic\] If $\alpha=1+\qk$, $\beta=1$ and $\lambda=\qi$ then $$\phi_{\alpha,\beta,\lambda}(x,t)=2\cos(x-t)+\sin(x-t)\qj+\cos(x-t)\qk$$ and $u_{\alpha,\beta,\lambda}(x,t)=w(x-t)$ where $$w(\xi)=
-\frac{8 (3 \cos (2 (\xi))+2)}{(2 \cos (2 (\xi))+3)^2}
-\frac{8 \sin (2 (\xi))}{(2 \cos (2 (\xi))+3)^2}\qi
+\frac{16 \sin (2 (\xi))}{(2 \cos (2 (\xi))+3)^2}\qj.$$ The four components of this solution at time $t=0$ are shown in Figure \[fig:periodic\]. As expected, an animation shows the solution translating to the right at constant speed $1$ and a horizontal spatial translation by $\pi$ units leaves the graph of $w(\xi)$ invariant.
### Localized Breather Solitons
\[thm:breathers\] If $\lambda_0\not=0$ then for any fixed $t$ the graph of the solution $u_{\alpha,\beta,\lambda}$ as a function of $x$ will be localized in a small neighborhood of $$x=c_{\alpha,\beta,\lambda}(t)=\frac{\ln|\alpha^{-1}\beta|}{2\lambda_0}-\vr t\label{eqn:center}$$ and hence this disturbance is moving to the left with velocity $\vr=\lambda_0^2-3\lambda_1^2$. Moreover, if $\lambda_1\not=0$ then the solution will also exhibit fluctuations moving to the left at velocity $\vi=3\lambda_0^2-\lambda_1^2$, giving it the “throbbing” appearance of a *breather soliton*.
The squared amplitude of the solution $u_{\alpha,\beta,\lambda}$ can be written in the form $$|u_{\alpha,\beta,\lambda}(x,t)|^2=
\frac{C_1}{\left(|\alpha|^2e^{\Lambda}+|\beta|^2e^{-\Lambda}+C_2
\cos(\Theta)+C_3\sin(\Theta)\right)^2}$$ where $\Lambda=2\lambda_0(x+\vr t)$, $\Theta=2 \lambda_1(x+\vi t)$ and $C_1$, $C_2$ and $C_3$ are some constants.
If $\lambda_0>0$ then for sufficiently large values of $|x|$ this amplitude converges quickly to $0$. In this sense, we can already see that the solution is *localized* when $\lambda$ has a non-zero real part. Moreover, if we “average out” the small variation from the trigonometric functions by setting them both equal to zero, then this amplitude function has a unique local maximum located at $x=c_{\alpha,\beta,\lambda}(t)$.
So, in the case $\lambda_0>0$ an animation of the solution $u_{\alpha,\beta,\lambda}(x,t)$ as a function of $x$ with $t$ playing the role of time will show a localized disturbance centered at $x=c_{\alpha,\beta,\lambda}(t)$ and traveling to the left with velocity $\vr$. However, if $\lambda_1>0$ as well, then the formula for the solution will also involve at least one of the functions $\cos(\theta)$ or $\sin(\theta)$. Since $\theta$ is a function of $x+\vi t$, these features will be moving to the left at velocity $\vi$. If it were the case that $\vr=\vi$, then the waveform would simply translate in time. However, there are no real solutions to $3\lambda_0^2-\lambda_1^2=\lambda_0^2-3\lambda_1^2$ and so whenever $\lambda_0$ and $\lambda_1$ are both non-zero, an animation of the solution will exhibit the “breathing” phenomenon.
\[examp:1sol\] Consider the case $\alpha=\qj$, $\beta=1+\qk$ and $\lambda=1+\sqrt{3}\qi$. We expect to see a localized disturbance traveling to the left with velocity $\vr=-8$ (which means it will move to the right with speed $8$ as $t$ increases). Moreover, since $\vi=0$ all of the trigonometric function in the formula for $u_{\alpha,\beta,\lambda}(x,t)$ have arguments that are independent of $t$. At time $t$ we would expect to see the center of the localized hump being located at $ x=8t+\ln(2)/4 $ (just slightly to the right of $x=8t$). And this is what we see in the figures showing this solution at times $t=-1$ and $t=1$.
### Singularities
Either the periodic or breather soliton solutions can exhibit singularities. The following result completely identifies the values of the parameters $\alpha$, $\beta$ and $\lambda$ that produce entirely non-singular solutions. A surprising corollary is that any of these solutions exhibiting singularities must be inherently *commutative* in that it is conjugate to a complex-valued solution of the standard KdV equation.
\[thm:sing\] The quaternion-valued KdV solution $u_{\alpha,\beta,\lambda}(x,t)$ is non-singular for all $x$ and $t$ if any of these three conditions involving the number $q=\alpha^{-1}\beta$ is satisfied:
- $q_2^2+q_3^2>0$ (i.e. $q$ is not a complex number)
- $\lambda_0=0$ and $|q|\not=1$
- or
- $\lambda_1=0$ and $q$ is not a negative real number.
Moreover, the solution $u_{\alpha,\beta,\lambda}(x,t)$ is undefined for some $(x,t)\in\R^2$ if none of the conditions is satisfied.
Throughout Section \[sec:uabl\] it has been assumed that $\alpha\beta\lambda\not=0$. Hence, we know that $\alpha$ is invertible. Define $q=\alpha^{-1}\beta$ and note that by Lemma \[lem:rot\] we know that $u_{1,q,\lambda}=\alpha^{-1}
u_{\alpha,\beta,\lambda}\alpha$. One of these solutions is singular if and only if the other is. Consequently, it is sufficient to determine when the solution $u_{1,q,\lambda}$ is singular.
Since $\phi_{1,q,\lambda}$ is infinitely differentiable for all $x$ and $t$ regardless of the values of the parameters, the only way that $u_{1,q,\lambda}$ given by could fail to be defined and differentiable at $(x_0,t_0)$ is if $|\phi_{1,q,\lambda}|(x_0,t_0)$ is zero. Using the previous notation that $\vi=3\lambda_0^2-\lambda_1^2$, $\vr=\lambda_0^2-3\lambda_1^2$ and now defining $\theta=2\lambda_1(x+\vi t)$ we find $$|\phi_{1,q,\lambda}(x,t)|^2=(q_0+e^{2\lambda_0(x+\vr
t)}\cos(\theta))^2+(q_1+e^{2\lambda_0(x+\vr t)}\sin(\theta))^2+q_2^2+q_3^2.
\label{eqn:lensqphi}$$ This sum of the squares of four real numbers can only be zero if all four of them are equal to zero.
If condition I is met then the last two terms in are non-zero and the solution must be non-singular.
Now suppose condition II is met. Since $\lambda_0=0$ the exponential terms are equal to $1$ and reduces to $(q_0+\cos(\theta))^2+(q_1+\sin(\theta))^2+q_2^2+q_3^2$. However, this is only zero if $(-q_0,-q_1)$ are the coordinates of the point at angle $\theta$ radians on the unit circle. Since $|q|\not=1$ that cannot be true.
And, if condition III is met then $\theta=0$ and becomes $$|\phi_{1,q,\lambda}(x,t)|^2=(q_0+e^{2\lambda_0(x+\vr
t)})^2+q_1^2+q_2^2+q_3^2.$$ If $q$ is not a real number then $q_1^2+q_2^2+q_3^2>0$ and the expression is non-zero And if $q$ is a non-negative real number then the first term is positive for all values of $x$ and $t$.
This shows that the solution is entirely non-singular if any one of the three conditions is met. Now, assume that none of the conditions is satisfied. So, we know that $q_2^2+q_3^2=0$, that either (i) $\lambda_0\not=0$ or that (ii) $|q|=1$, and that either (iii) $\lambda_1\not=0$ or that (iv) $q$ is negative.
Assuming (i) and (iii) ensures that the function in the exponent and the trigonometric argument $\theta$ are linearly independent linear functions of the variables $x$ and $t$ and hence can be simultaneously solved to take any desired value. The point $(-q_0,-q_1)$ is on a circle of radius $|q|$ around the origin and hence can be written as $(|q|\cos(\theta_0),|q|\sin(\theta_0))$ for some value of $\theta$. Then one can simultaneously find values of $x$ and $t$ such that $e^{2\lambda_0(x+\vr t)}=|q|$ and $\theta=\theta_0$ thereby making the entire expression equal to zero.
If (i) and (iv) are assumed to be true then $q$ is a negative real number and we want to show that $$(q+e^{2\lambda_0(x+\vr
t)}\cos(\theta))^2+(e^{2\lambda_0(x+\vr t)}\sin(\theta))^2$$ is zero for some choice for $x$ and $t$. If $\theta=0$ then the second term is already zero and if not then a value of $t$ can be found to make it zero. Either way the expression then reduces to $(q+e^{2\lambda_0(x+ C)})^2$ for some constant $C$ and the expression is equal to zero at $x=(\ln(-q)-C)/(2\lambda_0)$ (which is a real number as a consequence of (i) and (iv)).
Now suppose that (ii) and (iii) are true. We can further assume that $\lambda_0=0$ because otherwise (i) is true and we already handled that case. But then the expression reduces to $(q_0+\cos(\theta))^2+(q_1+\sin(\theta))^2$. By assumption (ii), we know that the point $(-q_0,-q_1)$ lies on the unit circle and hence there is a number $\theta_0$ such that it is equal to $(\cos(\theta_0),\sin(\theta_0))$. Since $\lambda_1\not=0$ it is possible to choose $x$ and $t$ so that $\theta=\theta_0$ and the expression then becomes zero.
Finally, consider the case in which (ii) and (iv) are both true. If $\lambda_0\not=0$ is also true then that would mean (i) and (iv) are true, and it has already been demonstrated that the solution is singular in that case. On the other hand, if $\lambda_0=0$ then $\lambda_1\not=0$ (because $\lambda\not=0$ is assumed throughout this section), but then (ii) and (iii) are true which has also already been handled.
\[rem:non-sing-hard\] One might guess from Figure \[fig:1sol\] that the breather soliton solution in Example \[examp:1sol\] is singular because it appears to have a pole in the bottom figure. However, $\alpha^{-1}\beta=-\qi-\qj$ is not a complex number and hence according to Theorem \[thm:sing\] it is not. (In fact, redrawing the graph at time $t=1$ over a larger vertical range confirms that there is simply a local maximum that is outside of the viewing window in Figure \[fig:1sol\].)
\[singular\] The periodic solution shown in Example \[examp:ratl\] is non-singular because $\alpha=1+\qk$, $\beta=1$ and $\lambda=\qi$ so $q=\alpha^{-1}\beta=1/2-1/2\qk$ which satisfies criterion II. On the other hand, choosing $\alpha=1$, $\beta=1/\sqrt{5}-2/\sqrt{5}\qi$ and $\lambda=\qi$ results in a singular solution $u_{\alpha,\beta,\lambda}(x,t)$ since none of the conditions are satisfied. This particular solution may seem uninteresting as it is complex-valued and therefore not inherently non-commutative, but it will play an important role in Example \[examp:onesidesing\] below.
Surprisingly, it turns out that $u_{\alpha,\beta,\lambda}$ has singularities *only* when the solution is really commutative:
\[cor:sing\] If the solution $u_{\alpha,\beta,\lambda}(x,t)$ is singular then $u(x,t)=\alpha^{-1}u_{\alpha,\beta,\lambda}\alpha$ is a complex-valued function, and $u_{\alpha,\beta,\lambda}$ is a solution of the usual KdV equation .
If $u_{\alpha,\beta,\lambda}(x,t)$ is singular then $q=\alpha^{-1}\beta\in\C$ must be a complex number (else Condition I of Theorem \[thm:sing\] is met and the solution would be non-singular). Note that $\phi_{1,q,\lambda}=\alpha^{-1}\phi_{\alpha,\beta,\lambda}$. By Lemma \[lem:rot\] $$u_{1,q,\lambda}(x,t)=\alpha^{-1}u_{\alpha,\beta,\lambda}(x,t)\alpha=u(x,t)$$ is another solution to . However, since $q$ and $\lambda$ are both complex numbers, $\phi_{1,q,\lambda}$ and therefore $u_{1,q,\lambda}$ which can be computed from it using are complex-valued functions. Consequently $u$ and $u_x$ commute. Commutativity is preserved by conjugation so $u_{\alpha,\beta,\lambda}$ also commutes with its derivative. Then, both of these functions solve .
Rational Solutions
------------------
\[def:Delta\] Let $\psi_0(x,t,z)=e^{xz+tz^3}$ and for $m=0,1,2,\ldots$ define $\Delta_m(x,t)$ to be $$\Delta_m(x,t)=\frac{\partial^m\psi_0}{\partial z^m}\bigg|_{z=0}.$$ Note that $\Delta_m(x,t)\in\Z[x,t]$ is a polynomial in $x$ and $t$ with integer coefficients and that it has degree $m$ as a polynomial in $x$.
\[thm:rat\] For any $n\in\N$ and $\alpha_j\in\H$, $\F=\{\phi_0,\ldots,\phi_n\}$ is a KdV-Darboux kernel where $$\phi_i(x,t)=\frac{\partial^{2i}}{\partial
x^{2i}}\left(\Delta_{2n+1}(x,t)+\sum_{j=0}^{n}\Delta_{2j}(x,t)\alpha_j\right).$$ Each component function $u_i(x,t)$ of corresponding quaternion-valued solution $u_{\F}(x,t)=u_0(x,t)+u_1(x,t)\qi+u_2(x,t)\qj+u_3(x,t)\qk$ to is a rational function.
Because $$\frac{\partial^3}{\partial x^3}\Delta_m(x,t)
=
\frac{\partial^3}{\partial x^3}\frac{\partial^i}{\partial z^i}\psi_0\bigg|_{z=0}
=
\frac{\partial^i}{\partial z^i}\frac{\partial^3}{\partial x^3}\psi_0\bigg|_{z=0}
=
\frac{\partial^i}{\partial z^i}z^3\psi_0\bigg|_{z=0}$$ $$=
\frac{\partial^i}{\partial z^i}\frac{\partial}{\partial t}\psi_0\bigg|_{z=0}
=
\frac{\partial}{\partial t}\frac{\partial^i}{\partial z^i}\psi_0\bigg|_{z=0}
=\frac{\partial}{\partial t}\Delta_m(x,t),$$ by linearity each function $\phi_i$ also satisfies the dispersion condition of Definition \[def:KdVDarbouxKernel\].
For $i<n$, $(\phi_i)_{xx}=\phi_{i+1}\in\F$. On the other hand, $\phi_n$ is the $(2n)^{th}$ derivative of a polynomial of degree $2n+1$, and so it is a linear function of $x$. Then $(\phi_n)_{xx}=0\in\span(\F)$. Thus $\F$ satisfies the closure property for KdV-Darboux Kernels.
Finally, to demonstrate the invertibility of the Wronskian matrix $W$, we consider a linear combination $$\sum_{i=0}^n\phi_i\alpha_i$$ of the entries in its first row with quaternionic coefficients on the right. Suppose that not all of the coefficients in this linear combination are zero and let $j$ be the smallest value of $i$ for which $\alpha_i\not=0$. Then, since $\phi_i$ is a polynomial in $x$ of degree $2(n-i)+1$ the $x^{2(n-j)+1}$ term which comes from $\phi_j\alpha_j$ cannot be cancelled by any other terms in the sum. Then the linear combination is not equal to zero unless all of the coefficients are zero and the homogenous vector equation $Wv=0$ has no non-trivial solutions which implies the invertibility of the matrix [@QMatInv].
Since $\F$ is a KdV-Darboux kernel we know that $u_{\F}$ is a KdV solution and shows that it can be found through products, and sums of derivatives of these polynomials followed by division by a real-valued polynomial and hence each component function is a rational function of $x$ and $t$.
\[examp:ratl2\] The first example of a quaternion-valued KdV solution above in Example \[examp:ratl\] was an instance of this construction in the case $n=1$, $\alpha_0=\qk$ and $\alpha_1=\qi$.
\[rem:Serna\] In fact, one may choose any finite linear combination of the polynomials $\Delta_m(x,t)$ and create a KdV-Darboux kernel out of this polynomial and all of its non-zero, even order derivatives in $x$. However, due to Lemmas \[lem:samespan\] and \[lem:dropone\], no new KdV solutions would be gained by considering even degree polynomials in $\F$ or including lower order odd degree terms in the formula for $\phi_i$. (See [@Serna] for details.)
Unions of KdV-Darboux Kernels
=============================
Since the dispersion and closure properties of Definition \[def:KdVDarbouxKernel\] are preserved under the taking of unions, it follows immediately that:
If $\F_1$ and $\F_2$ are KdV-Darboux kernels then $\F=\F_1\cup\F_2$ is also a KdV-Darboux kernel as long as its Wronskian matrix is invertible.
This way of combining KdV-Darboux kernels allows for the creation of $n$-soliton solutions or hybrids that exhibit features of more than one of the basic solution types described in the previous section. For example, there is a rational-periodic hybrid solution coming from the union of the KdV-Darboux kernels in Examples \[examp:ratl\] and \[examp:periodic\].
The main focus of this section will be the case in which one additional function of the form $\phi_{\alpha,\beta,\lambda}$ with $\lambda_0>0$ is added to a KdV-Darboux kernel. It is a consequence of Lemma \[lem:dropone\] that the resulting solution will look like two different solutions “glued” together, one visible to the left of the localized disturbance that has been added and other other to the right of it. This general fact will be both proved and illustrated in Section \[sec:LR\]. Then in Section \[sec:2-sol\] it will be used to derive a formula for the phase shift of an arbitrary $2$-soliton solution.
Asymptotics to the Left and Right of a Localized Disturbance {#sec:LR}
------------------------------------------------------------
\[prop:LR\] Let $\F=\{\phi_1,\ldots,\phi_n\}$ be a KdV-Darboux kernel where $n\geq2$ and $\phi_n$ has the form $$\phi_n(x,t)=\phi_{\alpha,\beta,\lambda}(x,t)=\alpha e^{\lambda x+
\lambda^3 t}+\beta e^{-\lambda x - \lambda^3 t}$$ with $\lambda=\lambda_0+\lambda_1\qi,\ \lambda_0>0$. Then for each fixed $t$ and far enough to the left of $x=c_{\alpha,\beta,\lambda}(t)$ the graph of $u_{\F}(x,t)$ looks like[^2] the graph of $u_{\F^L}(x,t)$ where $$\F^L=\{Q^L(\phi_1),\ldots,Q^L(\phi_{n-1})\}$$ with $Q^L(f)=f_x+\beta\lambda\beta^{-1}f$. Similarly, for each fixed $t$ and far enough to the right of $x=c_{\alpha,\beta,\lambda}(t)$ the solution $u_{\F}(x,t)$ looks like $u_{\F^R}(x,t)$ where $$\F^R=\{Q^R(\phi_1),\ldots,Q^R(\phi_{n-1})\}$$ with $Q^R(f)=f_x-\alpha\lambda\alpha^{-1}f$.
The “center” $c_{\alpha,\beta,\lambda}(t)$ is the value of $x$ for which there is a balance between the magnitude of the two exponential terms in $\phi_n$. For $x$ much smaller than it, $\alpha e^{\lambda x + \lambda^3 t}$ is negligibly small. For those values of $x$, the solution $u_{\F}(x,t)$ will not look noticeably different than it would if $\phi_n$ was equal to $\beta e^{-\lambda x - \lambda^3
t}$, which according to Lemma \[lem:dropone\] is precisely $u_{\F^L}$. Similarly, when $x$ is much larger than $c_{\alpha,\beta,\lambda}(t)$ the term $\beta e^{-\lambda x -
\lambda^3 t}$ is negligibly small and the solution would not look noticeably different than it would if that term was not there, which is $u_{\F^R}$ according to Lemma \[lem:dropone\].
\[examp:LRratl\] Consider the “hybrid” rational/soliton solution $u_{\F}$ that comes from the choice $$\F=\{x+3\qk,\phi_{\alpha,\beta,\lambda}\}\hbox{ with
}\alpha=1,\ \beta=\qj,\ \lambda=2+\qi.$$ According to Proposition \[prop:LR\] the left side of this solution should look like $$u_{\F^L}=\frac{-50 x^2-40 x+444}{\left(5 x^2+4 x+46\right)^2}
+\frac{ (5 x+2)}{\left(5 x^2+4 x+46\right)^2}(4\qi+48\qj+36\qk)$$ and the right side should like look $$u_{\F^R}=\frac{-50 x^2+40 x+444}{\left(5 x^2-4 x+46\right)^2}
+\frac{(5 x-2)}{\left(5 x^2-4 x+46\right)^2}(4\qi-48\qj+36\qk)$$ Each of these is a stationary ($t$-independent) quaternion-valued KdV solution. They are shown in the left-most and right-most images of Figure \[fig:LRratl\] respectively. The middle two images of that figure show $u_{\F}$ at times $t=-5$ and $t=5$. Then we can see that $u_{\F}$ looks like $u_{\F^L}$ to the left of the incoming soliton and looks like $u_{\F^R}$ to the right of it.
to
\[rem:phitophi\] Since Proposition \[prop:LR\] will mostly be applied in the case that each $\phi_i\in\F$ is a function of the form , it is worth noting that the differential operator $Q$ defined by $Q(f)=f_x-\gamma f$ preserves that form. In particular, it can easily be computed that for any $\hatalpha,\hatbeta,\hatlambda\in\H$ $$Q(\phi_{\hatalpha,\hatbeta,\hatlambda})=\phi_{\tildealpha,\tildebeta,\hatlambda}\
\hbox{where}\ \tildealpha=\hatalpha\hatlambda-\gamma\hatalpha,\ \tildebeta=-\hatbeta\hatlambda-\gamma\hatbeta.
\label{eqn:Qphi}$$
\[examp:onesidesing\] Let $\F_1$ be the KdV-Darboux kernel $$\F_1=\{\phi_{\alpha_1,\beta_1,\lambda_1}\}
\qquad\hbox{where}\qquad
\alpha_1=1,\ \beta_1=\frac{1}{\sqrt{5}},\ \lambda_1=\qi$$ then $u_{\F_1}$ is a complex-valued, non-singular, periodic solution to . And let $\F_2$ be the KdV-Darboux kernel $$\F_2=\{\phi_{\alpha_2,\beta_2,\lambda_2}\}\qquad
\hbox{where}\qquad
\alpha_2=\qi+\qj,\ \beta_2=\qj+\qk,\ \lambda_2=1+\qi$$ so that $u_{\F_2}$ is a breather soliton solution traveling to the right at speed $2$. What will the solution coming from $\F=\F_1\cup\F_2$ look like? According to Proposition \[prop:LR\] and Remark \[rem:phitophi\], to the left of $x=c_{\alpha_2,\beta_2,\lambda_2}=2t$ should look like $u_{\alpha_L,\beta_L,\lambda_1}$ with $$\alpha_L=1,\
\beta_L=\frac{1}{\sqrt{5}}-\frac{2}{\sqrt{5}}\qi$$ while on the right it should look like $u_{\alpha_R,\beta_R,\lambda_1}$ where $$\alpha_R=-1+\qi-\qj,\
\beta_R=-\frac{1}{\sqrt{5}}-\frac{1}{\sqrt{5}}\qi-\frac{1}{\sqrt{5}}\qj.$$ The interesting thing about this is that since $|\beta_L|=1\not=|\beta_R|$ the solution it looks like on the left is *singular* while the one on the right is not. Figure \[fig:LR\] shows that the solutions appear as predicted. Moreover, this solution is very interesting to watch as an animation because the localized disturbance traveling to the right seems to transform a non-singular periodic solution into a singular one as it passes.
to
Phase Shift of the General 2-soliton {#sec:2-sol}
------------------------------------
Suppose $\alpha$, $\beta$, $\hat\alpha$ and $\hat\beta$ are non-zero quaternions and that $\lambda$ and $\hat\lambda$ are complex numbers such that: $$\lambda_0>0,\qquad\hat\lambda_0>0,\qquad\hbox{and}\qquad
\hat\lambda_0^2-3\hatlambda_1^2<\lambda_0^2-3\lambda_1^2.$$ Then $u_{\alpha,\beta,\lambda}$ and $u_{\hatalpha,\hatbeta,\hatlambda}$ are each quaternion-valued 1-soliton solutions to . Moreover, because of the last inequality they have different velocities. In particular, the center $c_{\alpha,\beta,\lambda}(t)$ moves to the left more quickly than $c_{\hatalpha,\hatbeta,\hatlambda}(t)$ (or, equivalently, moves to the right more slowly).
This section will use Proposition \[prop:LR\] to analyze the quaternion-valued KdV solution $u_{\F}$ where $$\F=\{\phi_{\alpha,\beta,\lambda},\phi_{\hatalpha,\hatbeta,\hatlambda}\}.$$ As will be seen below, there are constants $\alpha_-,\beta_-,\hatalpha_-,\hatbeta_-\in\H$ so that an observer watching an animation of $u_{\F}$ for sufficiently negative values of $t$ would see two localized disturbances traveling with the same velocities as $c_{\alpha,\beta,\lambda}$ and $c_{\hatalpha,\hatbeta\hatlambda}$. In particular, there are quaternions $\alpha_-,\beta_-,\hatalpha_-,\hatbeta_-\in\H$ so that the observer would see what appeared to be the localized disturbances from the solution $u_{\alpha_-,\beta_-,\lambda}$ and $u_{\hatalpha_-,\hatbeta_-,\hatlambda}$ together in this $2$-soliton solution.
There are also numbers $\alpha_+,\beta_+,\hatalpha_+,\hatbeta_+\in\H$ so that an animation of $u_{\F}$ for sufficiently positive values of $t$ will look like a sum of the localized solutions $u_{\alpha_+,\beta_+,\lambda}$ and $u_{\hatalpha_+,\hatbeta_+,\hatlambda}$. So, the observer would still see two localized disturbances traveling with the same velocities. However, they might not be in the locations that the observer would have predicted. Suppose the observer identifies the localized disturbances of the same velocities at different times. Then, since at negative times the disturbance with greater velocity appears to be following the linear trajectory $c_{\alpha_-,\beta_-,\lambda}(t)$, the observer may expect it to be found on that same line at positive times as well. Indeed, since $c_{\alpha_-,\beta_-,\lambda}$ and $c_{\alpha_+,\beta_+,\lambda}(t)$ are moving to the left at the same velocity, their difference is a constant. If it is zero, then the observer will find that the disturbance at later times is right where it would have been expected to be. However, if this difference is non-zero, then it will be *shifted* from its expected position. Hence, the number $c_{\alpha_+,\beta_+,\lambda}-c_{\alpha_-,\beta_-,\lambda}$ is called the *phase shift* of that localized disturbance and it is generally interpreted as being a lasting consequence of its collision with the other localized disturbance (though there are other interpretations as well)[@ZK; @BKY].
The following result provides formulas for the parameters for the corresponding $1$-soliton solutions and the phase shifts of each of the localized disturbances. As the examples after it demonstrate, unlike the real case, the phase shift of the solitary wave with greater velocity in the quaternion-valued KdV $2$-soliton can be positive, negative, *or* zero. Additionally, the solitary waves traveling at the same velocities at positive and negative may differ in *shape* as well as being horizontally shifted relative to each other.
\[prop:PhaseShift\] Choose constant $\alpha$, $\beta$, $\lambda$, $\hatalpha$, $\hatbeta$, and $\hatlambda$ as described above and let $\F$ be the KdV-Darboux kernel $$\F=\{\phi_{\alpha,\beta,\lambda},\phi_{\hat\alpha,\hat\beta,\hatlambda}\}.$$
- For sufficiently negative values of $t$, the solution $u_{\F}$ will look like the $1$-soliton $u_{\alpha_-,\beta_-,\lambda}$ near its center $x=c_{\alpha_-,\beta_-,\lambda}(t)$ and will also look like the $1$-soliton $u_{\hatalpha_-,\hatbeta_-,\hatlambda}$ near its center $x=c_{\hatalpha_-,\hatbeta_-,\hatlambda}(t)$ where $$\alpha_-= \alpha\lambda-\hatalpha\hatlambda\hatalpha^{-1}\alpha,
\qquad
\beta_-=-\beta\lambda-\hatalpha\hatlambda\hatalpha^{-1}\beta,$$ $$\hatalpha_-= \hatalpha\hatlambda+\beta\lambda\beta^{-1}\hatalpha,
\qquad \hbox{and}\qquad
\hatbeta_-=-\hatbeta\hatlambda+\beta\lambda\beta^{-1}\hatbeta.$$
- For sufficiently positive values of $t$, the solution $u_{\F}$ will look like the $1$-soliton $u_{\alpha_+,\beta_+,\lambda}$ near its center $x=c_{\alpha_+,\beta_+,\lambda}(t)$ and will also look like the $1$-soliton $u_{\hatalpha_+,\hatbeta_+,\hatlambda}$ near its center $x=c_{\hatalpha_+,\hatbeta_+,\hatlambda}(t)$ where $$\alpha_+=\alpha\lambda+\hatbeta\hatlambda\hatbeta^{-1}\alpha
\qquad
\beta_+=-\beta\lambda+\hatbeta\hatlambda\hatbeta^{-1}\beta$$ $$\hatalpha_+= \hatalpha\hatlambda-\alpha\lambda\alpha^{-1}\hatalpha
\qquad \hbox{and}\qquad
\hatbeta_+=-\hatbeta\hatlambda-\alpha\lambda\alpha^{-1}\hatbeta$$
- An observer looking at the solution $u_{\F}$ at very negative values of time would see two localized disturbances that appear to be moving at constant speeds. Looking again at a very positive time the observer would see two localized disturbances with the same velocities, but they might not be in the locations that would have been expected if they had continued to follow a linear trajectory. The phase shifts experienced by the interacting solitary waves are $$c_{\alpha_+,\beta_+,\lambda}(t)-c_{\alpha_-,\beta_-,\lambda}(t)=\frac{\ln(\gamma)}{2\lambda_0}%\\
\qquad\hbox{ and }\qquad
c_{\hatalpha_+,\hatbeta_+,\hatlambda}(t)-c_{\hatalpha_-,\hatbeta_-,\hatlambda}(t)=-\frac{\ln(\gamma)}{2\hatlambda_0}$$ where $$\gamma=\frac{|\alpha\lambda\alpha^{-1}-\hatalpha\hatlambda\hatalpha^{-1}||
\beta\lambda\beta^{-1}-\hatbeta\hatlambda\hatbeta^{-1}|}{|\alpha\lambda\alpha^{-1}+\hatbeta\hatlambda\hatbeta^{-1}||\beta\lambda\beta^{-1}+\hatalpha\hatlambda\hatalpha^{-1}|}.$$
First, we will apply Proposition \[prop:LR\] to $\F$ using $\phi_{\hatalpha,\hatbeta,\hatlambda}$ in the role of $\phi_n$. The proposition says that for all $t$, far enough to the right of $c_{\hatalpha,\hatbeta,\hatlambda}(t)$, the graph of the solution $u_{\F}$ will look like $u_{\alpha_-,\beta_-,\lambda}$ using the values from above. Because the center $c_{\alpha_-,\beta_-,\lambda}(t)$ is moving to the left more quickly, for any sufficiently negative value of $t$, it will be far enough to the right of $c_{\hatalpha,\hatbeta,\hatlambda}$ so that it is in the region where $u_{\F}$ looks like $u_{\alpha_-,\beta_-,\lambda}$. Therefore for very negative values of $t$, $u_{\F}$ has a localized disturbance that looks like the one in the one soliton $u_{\alpha_-,\beta_-,\lambda}$.
On the other hand, when $t$ is very positive then $c_{\alpha_+,\beta_+,\lambda}(t)$ will be far to the right of $c_{\hatalpha,\hatbeta,\hatlambda}(t)$. In particular because it is moving to the left at a faster constant speed when $t$ is sufficiently large it will be located in the region where according to Proposition \[prop:LR\] the solution $u_{\F}$ will look like $u_{\alpha_+,\beta_+,\lambda}$. Since that is the location around which the soliton is localized in that $1$-soliton, the solution $u_{\F}$ will look like that 1-soliton near that point.
The other parameters come from repeating this same process but now using $\phi_{\alpha,\beta,\lambda}$ in the role of $\phi_n$ when applying Proposition \[prop:LR\].
Now, $c_{\alpha_+,\beta_+,\lambda}(t)$ and $c_{\alpha_-,\beta_-,\lambda}(t)$ are two linear trajectories with the same velocity. The difference between them is the phase shift, how much further to the right the disturbance is after the collision than it would have been if it had continued to look like $u_{\alpha_-,\beta_-,\lambda}$. Using the formula for the center, properties of logarithms and properties of the length operator on quaternions, we can see that $$\begin{aligned}
c_{\alpha_+,\beta_+,\lambda}(t)-c_{\alpha_-,\beta_-,\lambda}(t)
&=&\frac{\ln|\alpha_+^{-1}\beta_+|}{2\lambda_0}-\frac{\ln|\alpha_-^{-1}\beta_-|}{2\lambda_0}
=\frac{1}{2\lambda_0}\ln\left(\frac{|\alpha_+^{-1}\beta_+|}{|\alpha_-^{-1}\beta_-|}\right)\\
&=&\frac{1}{2\lambda_0}\ln\left(\frac{|\alpha_-||\beta_+|}{|\alpha_+||\beta_-|}\right)\\
&=& \frac{1}{2\lambda_0}\ln\left(\frac{|\alpha\lambda-\hatalpha\hatlambda\hatalpha^{-1}\alpha||
-\beta\lambda+\hatbeta\hatlambda\hatbeta^{-1}\beta|}{|\alpha\lambda+\hatbeta\hatlambda\hatbeta^{-1}\alpha||-\beta\lambda-\hatalpha\hatlambda\hatalpha^{-1}\beta|}\right)\\
&=&\frac{1}{2\lambda_0}\ln\left(\frac{|\alpha\lambda\alpha^{-1}-\hatalpha\hatlambda\hatalpha^{-1}||
\beta\lambda\beta^{-1}-\hatbeta\hatlambda\hatbeta^{-1}|}{|\alpha\lambda\alpha^{-1}+\hatbeta\hatlambda\hatbeta^{-1}||\beta\lambda\beta^{-1}+\hatalpha\hatlambda\hatalpha^{-1}|}\right)=\frac{\ln(\gamma)}{2\lambda_0}.\end{aligned}$$ A similar calculation for $c_{\hatalpha_+,\hatbeta_+,\hatlambda}-c_{\hatalpha_-,\hatbeta_-,\hatlambda}$ results in the same formula but with the numerator and denominator switched in the argument of the logarithm with the effect of changing the sign.
to
to
to
\[examp:phaseshift\] Imagine a naive observer watching an animation of the 2-soliton solution $u_{\F}$ where $\F=\{\phi_{\alpha,\beta,\lambda},\phi_{\hatalpha,\hatbeta,\hatlambda}\}$ with $$\alpha=\beta=\hatalpha=1,
\
\hatbeta=20\qi+\qk,
\
\lambda=2+\qi,
\
\hatlambda=3/2+\sqrt{3}/2\qi.$$ Watching for some negative values of $t$, the observer would see a stationary breather soliton sitting just a bit to the right of $x=0$ and then another breather soliton approaching from the right at speed $1$. In fact, as predicted by Proposition \[prop:PhaseShift\] the moving soliton looks like $u_{\alpha_-,\beta_-,\lambda}$ as shown at the top-left of Figure \[fig:phaseshift\]. In particular, for these negative times, the moving localized disturbance in $u_{\F}$ is located at $$x=c_{\alpha_-,\beta_-,\lambda}(t)=\frac{\log \left(31+16 \sqrt{3}\right)}{8}-t.$$ So, the naive observer might assume that this will continue to be true in the future. However, as the images on the right in Figure \[fig:phaseshift\] show, it does not. Although there is still a disturbance moving left at speed $1$ at times $t=3$ and $t=4$, it no longer looks like $u_{\alpha_-,\beta_-,\lambda}$. Instead, it looks like $u_{\alpha_+,\beta_+,\lambda}$, a 1-soliton whose center is: $$x=c_{\alpha_+,\beta_+,\lambda}(t)=\frac{1}{8} \log
\left(\frac{711433-365712 \sqrt{3}}{4434199}\right)-t.$$ Since they are both traveling to the left at the same speed, their difference is a constant $$\begin{aligned}
c_{\alpha_+,\beta_+,\lambda}-c_{\alpha_-,\beta_-,\lambda}&=&\frac{1}{8}
\log \left(\frac{39608599-22720000
\sqrt{3}}{855800407}\right)\approx-1.01413,\end{aligned}$$ which means that the observer would find the localized disturbance after the collision is about $1.014$ units farther to the left than expected. (Similarly, the phase shift for the stationary soliton would be positive, which is why it has moved to the right after the interaction.)
\[rem:phaseshift\] The phase shift experienced by each of the two solitary waves experience have opposite signs. Like the classical commutative KdV solitons, if one is shifted forwards then the other shifts backwards. However, unlike the original KdV solitons studied by Zabusky and Kruskal [@ZK], the phase shift is not completely determined by the velocities, as the next example illustrates.
\[examp:phaseshiftrange\] A one parameter family of interesting examples is the case in which $$\alpha=1 \qquad
\beta=1\qquad
\lambda=1+3\qi$$ $$\hatalpha=\qj\qquad
\hatbeta=-\qi+c\qk\qquad\hbox{and}\qquad
\hatlambda=1+4\qi.$$ For any value of $c\in\R$ this represents a 2-soliton solution with one traveling to the right at speed $26$ and another at speed $47$. However, the phase shift that they experience depends on $c$: $\ln(\gamma)$ is positive for $|c|>1/\sqrt{11}$, zero if $|c|=1/\sqrt{11}$ and negative if $|c|<1/\sqrt{11}$. This dramatically demonstrates the fact that in the non-commutative case the phase shift depends on the coefficients $\alpha$, $\beta$, $\hatalpha$ and $\hatbeta$. That does not happen in the commutative case, as one can observe by noting that all dependence on these coefficients cancel from the formula for $\gamma$ in Proposition \[prop:PhaseShift\] if the parameters commute.
Concluding Remarks
==================
Although it is surely no more than a coincidence that the quaternions and the existence of solitary waves were both famously discovered beside British canals in the 19th century, this paper has found interesting results by studying quaternion-valued solutions to the KdV-equation that can be produced using the Chen determinant. This is both a generalization of and a special case of other published research, as will be explained further below.
When the functions in the KdV-Darboux kernel $\F$ are all either complex-valued or real-valued, then the solution $u_{\F}$ satisfies and many of the new results in this paper reduce to well-known results. For instance, singularities of soliton and periodic complex-valued solutions to KdV have been studied in Reference [@complex] much as we considered the singularities of the quaternion-valued solutions in Theorem \[thm:sing\]. However, the generalization to the non-commutative case handled here is non-trivial. Without Corollary \[cor:sing\] it was not at all obvious that the singularities that can be found in complex-valued KdV solutions would necessarily fail to exist in their non-commutative quaternionic counterparts. Moreover, as noted in Example \[examp:phaseshiftrange\], the phase shift in the quaternion-valued 2-soliton depends on the coefficients as well as on the exponents, something that is not true in the commutative case.
On the other hand, there are also many published papers which address non-commutative solutions to integrable PDEs in more general settings. The KdV equation is merely one equation in the KdV hierarchy, which is a reduction of the KP hierarchy. Quaternions can be viewed as being a special four-dimensional subspace of larger matrix groups, which are then special cases of abstract non-commutative rings. With all of that in mind, the methods utilized herein can be seen as simply being a special case of the more general approaches found in papers such as References [@NC3; @EGR; @NC1; @NC2]. However, limiting ourselves to this manageable situation allows us to study details that would be difficult to notice and demonstrate in those more general settings. For instance, we were able to show that it was sufficient to consider exponential functions of the form $e^{\lambda x
+ \lambda^3 t}$ where $\lambda=\lambda_0+\lambda_1\qi$ is a complex number with non-negative components $\lambda_i$ (cf. Lemma \[lem:complexlambda\]). Doing so was essential in being able to state Theorem \[thm:sing\] (the result about singularities) in an easily understandable way. And, since $\H$ is a four-dimensional vector space, we were able to graph the corresponding solutions as a super-position of graphs of four real-valued functions. Moreover, it is interesting to know that quaternion-valued solutions to KdV can be written in terms of the Chen determinant, a result that presumably would not generalize to solutions with values in arbitrary non-commutative rings.
There is one relatively recent paper by Huang [@QKdV] which, like this one, specifically addresses quaternion-valued soliton solutions to KdV. However, Huang’s paper only considers a small subset of the solution types that were addressed above. In particular, it only looks at solutions that would come from KdV-Darboux kernels made up of functions of the form $\phi_{\alpha,\beta,\lambda}$ where $\alpha,\beta\in\H$ and $\lambda\in\R$. Thus, it does not include breather, rational, or periodic solutions. Finally, although Reference [@QKdV] does discuss “interactions” of the solutions, that term has a very different meaning in that paper. Here, Proposition \[prop:LR\] is viewed as a means to understand the interaction of different solutions, with special emphasis on the $2$-soliton solutions as representing the interaction between two separate $1$-solitons (cf. Proposition \[prop:PhaseShift\]). But in Reference [@QKdV], “interaction” refers to an algebraic structure that Huang studies whereby two $n$-solitons can be combined to produce another $n$-soliton (for the same fixed value of $n$).
There are many interesting examples which can be made using the methods described above that we did not have the time or space to present here. For instance, there are $2$-soliton solutions that look like a combination of non-singular $1$-solitons at negative times and then like a pair of singular $1$-solitons for positive times (as if the collision produced the singularities).
There are also open problems that we have not been able to fully address. Theorem \[thm:sing\] completely determines when a solution of the form $u_{\alpha,\beta,\lambda}$ is singular. However, it is not entirely clear when combinations of such solutions are singular. Although Propositions \[prop:LR\] and \[prop:PhaseShift\] may tell us when they *look* singular, as Remark \[rem:non-sing-hard\] shows, that is not quite the same as actually being singular. We do not yet have any prediction for what $u_{\F}$ will look like if $\F=\{\phi_{\alpha_1,\beta_1,\lambda_1},\ldots,\phi_{\alpha_n,\beta_n,\lambda_n}\}$ with $n>1$ and $\lambda_i=\lambda_{i1}\qi$ purely imaginary. Most intriguingly, since the solutions above were written in terms of the Chen determinant of a Wronskian matrix, it would be interesting to know whether there is a quaternionic analogue of the $\tau$-function and Hirota’s bilinear approach to soliton equations.
Acknowledgments {#acknowledgments .unnumbered}
===============
This paper grew out of a student research project at the College of Charleston conducted in Summer 2018. The first author’s research was supported by a grant from the School of Sciences and Mathematics, the third author’s research was supported by the University of Charleston, South Carolina (the Graduate School at the College of Charleston), and the fourth author’s research was supported by a Summer Undergraduate Research with Faculty (SURF) grant from the Office of Undergraduate Research and Creative Activities. Assistance from all of these offices and from the Department of Mathematics at the College of Charleston is greatly appreciated.
[10]{}
[^1]: The quasi-determinant [@EGR] is useful in non-commutative settings. However, it is not a *generalization* of the determinant in that a quasi-determinant of a matrix which happens to have commuting entries is not equal to the determinant of that matrix.
[^2]: We are being intentionally vague about what it means for one solution to “look like another” to the right or left of some value of $x$ here because the notation and proofs both become unwieldy otherwise. A rigorous mathematical definition might include an arbitrarily small maximum amplitude for the difference of the two functions when $x$ is is more than a certain distance to the right or left of $c_{\alpha,\beta,\lambda}(t)$. It is hoped that Examples \[examp:LRratl\] and \[examp:onesidesing\] illustrate both the meaning and significance of Proposition \[prop:LR\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the one-loop corrected decay widths for the decays of the neutral Higgs bosons $h^0$, $H^0$ and $A^0$ into a neutralino pair ${{\tilde{\chi}}}^0_m\, {{\tilde{\chi}}}^0_n$ $(m, n = 1,\ldots,4)$ and to the decay ${{\tilde{\chi}}}^0_m \to (h^0,\,H^0,\,A^0)\,+\,{{\tilde{\chi}}}^0_n$. The corrections contain the one-loop contributions of all fermions and sfermions. All parameters are taken on-shell. This requires a proper treatment of the neutralino mass and mixing matrix. The dependence on the SUSY parameters is discussed. The corrections can be large in certain regions of the parameter space.'
---
HEPHY-PUB 749/01\
TU-639\
hep-ph/0111303
**One-loop corrections to\
neutral Higgs boson decays into neutralinos**
[H. Eberl$^a$, M. Kincel$^{a,\,b}$, W. Majerotto$^a$, Y. Yamada$^c$]{}
[l]{} $^a$[*Institut für Hochenergiephysik der Österreichischen Akademie der Wissenschaften,*]{}\
\
$^b$[*Department of Theoretical Physics FMFI UK, Comenius University, SK-84248*]{}\
\
$^c$[*Department of Physics, Tohoku University, Sendai 980–8578, Japan*]{}
Introduction
============
The Minimal Supersymmetric Standard Model (MSSM) [@mssm] is considered the most attractive extension of the Standard Model. The MSSM requires the existence of two isodoublets of scalar Higgs fields, implying three neutral Higgs bosons, two CP-even bosons ($h^0$, $H^0$), one CP-odd ($A^0$), and two charged Higgs bosons ($H^{\pm}$). Searching for these Higgs bosons is one of the main goals of all future colliders as the Tevatron, LHC, and an $e^+ e^-$ linear collider. The search strategies very much depend on the way these Higgs bosons decay. It is therefore mandatory to have a clear picture of the decay modes. Thus it is necessary to calculate the widths and branching ratios of the various decays as precisely as possible.
The lightest Higgs boson $h^0$ with a mass of at most $140$ GeV will decay mainly into $b\bar{b}$ and to a lesser extent into ${\tau}^+{\tau}^-$. It is, however, possible that it also decays as $$h^0\to {{\tilde{\chi}}}^0_1 + {{\tilde{\chi}}}^0_1 \, ,
\label{eq:h0ch1ch1}$$ where ${{\tilde{\chi}}}^0_1$ is the lightest neutralino. In the case of $R$-parity conservation, this decay is invisible, and its appearance would reduce the branching ratios of the other decay modes. The heavier neutral Higgs bosons $H^0$ and $A^0$ may decay into a pair of neutralinos $$(H^0, A^0) \to {{\tilde{\chi}}}^0_m + {{\tilde{\chi}}}^0_n \, ,
\label{eq:Hkchmchn}$$ with $(m,n=1,\ldots,4)$. At tree level, the decays occur by higgsino-gaugino-Higgs boson couplings [@gunion], and are therefore sensitive to the components of neutralinos. The decays (\[eq:h0ch1ch1\]) and (\[eq:Hkchmchn\]) as well as those of $H^{\pm}\to \tilde\chi^{\pm}_{i}{{\tilde{\chi}}}^0_m,\;(i=1,2)$ have been numerically analyzed in [@tree1; @tree2] at tree level. Electroweak corrections to the widths of $H^{\pm}\to
\tilde\chi^{\pm}_{i} {{\tilde{\chi}}}^0_m$ due to one-loop exchanges of the third generation quarks and squarks were recently calculated in [@Yi]. The one-loop corrections, involving fermions and sfermions, to the invisible width of $(h^0, H^0, A^0)\to{{\tilde{\chi}}}^0_1 + {{\tilde{\chi}}}^0_1$ have been calculated in the higgsino limit of ${{\tilde{\chi}}}^0_1,\,(|\mu |\ll M_1
,M_2)$ in [@higgsinolike], and in the gaugino limit of ${{\tilde{\chi}}}^0_1,\,(|\mu |\gg M_1 ,M_2)$ very recently in [@Djouadi-Drees]. (Here $M_1$ and $M_2$ are the $U(1)$ and $SU(2)$ gaugino mass parameters, respectively, and $\mu$ is the higgsino mass parameter.) In these limiting cases, the wave-function corrections can be neglected and no renormalization is necessary. The couplings of $(h^0,H^0, A^0)$ to ${{\tilde{\chi}}}^0_1$ ${{\tilde{\chi}}}^0_1$ also enter in the neutralino-quark interaction [@Djouadi-Drees], a process which is very important for the dark matter search [@dark1; @dark2], where one looks for the elastic scattering of neutralinos ${{\tilde{\chi}}}^0_1$ off nuclei in a detector. Moreover, since the decays (\[eq:h0ch1ch1\],\[eq:Hkchmchn\]) are generated by gaugino-higgsino-Higgs boson couplings at tree level, they can be also useful to probe the components of the neutralinos, complementary to the pair production process $e^+e^-\hspace{-3pt}\rightarrow{{\tilde{\chi}}}^0_m{{\tilde{\chi}}}^0_n$ [@neuproduction].
In this paper, we present the one-loop corrections to the widths of the decays (\[eq:h0ch1ch1\]) and (\[eq:Hkchmchn\]) due to the exchange of all fermions (quarks and leptons) and their superpartners (sfermions).
The decays (\[eq:h0ch1ch1\]) and (\[eq:Hkchmchn\]) are particularly interesting because the calculation of their radiative corrections requires corrections to the neutralino mass matrix and mixing matrix in addition to the conventional wave-function and vertex corrections with counter terms. The one-loop corrections to the neutralino mass and mixing matrix in the on-shell renormalization scheme were already worked out in [@chmasscorr] and they will be used here.\
Related to these decays are the decays of neutralinos into Higgs bosons, $${{\tilde{\chi}}}^0_m \to (h^0, H^0, A^0) + {{\tilde{\chi}}}^0_n \, .
\label{eq:chidecay}$$ These decays are also important as they occur in the cascade decays of gluinos and/or squarks, $\tilde{g}\to
q\bar{q}\tilde\chi^0_m$ and $\tilde{q}\to q\tilde\chi^0_m$, with $\tilde\chi^0_m$ then decaying according to (\[eq:chidecay\]). The decays (\[eq:chidecay\]) with a real Higgs boson emission [@neutree1; @neutree2] as well as three-body decays due to an off-shell Higgs boson [@neutree3] have been studied at tree level. In this paper, we also present the formulae for the decays (\[eq:chidecay\]) including the one-loop corrections.
Tree-level widths
=================
Throughout this paper, we will use the notations $m_{{{\tilde{\chi}}}^0_i}
\equiv m_i$ and $H_k^0 \equiv \{h^0, H^0, A^0, G^0 \}$. In a non-unitary gauge we have the ghost $G^0$. The momenta are assigned as $(k = 1,2,3$; $m,n=1,\ldots,4)$ $$H_k^0(p) \to {{\tilde{\chi}}}^0_m (k_1) + {{\tilde{\chi}}}^0_n (k_2) \, .
\label{eq:Hk0neu}$$ All couplings are given in the Appendix \[sec:appA\] (or it is referred to previous works).
The tree-level widths for a neutral Higgs decaying into two neutralinos is [@tree1] $$\begin{aligned}
&& \hspace{-1cm}
\Gamma^{\rm tree}_H = \Gamma^{\mbox{\footnotesize
tree}}(H_{k}^0 \to {{\tilde{\chi}}}^{0}_{m}\, {{\tilde{\chi}}}^{0}_{n}) = \nonumber\\[2mm]
&& \hspace{-0.5cm}
\frac{g^{2}}{8 \pi\, m^{3}_{H_{k}^0}\,
(1 + \delta_{mn})} \, \kappa(m^{2}_{H_{k}^0},m^{2}_{m},m^{2}_{n}) \,
|F^0_{mnk}|^{2}\, \left[ m^{2}_{H_{k}^0} - m^{2}_m - m^{2}_n
- 2 (-1)^{\delta_{k3}}\, m_m m_n \right]\, ,
\label{eq:gammatree}\end{aligned}$$ with $\kappa(x,y,z)\equiv ((x-y-z)^2-4yz)^{1/2}$. The couplings $F^0_{mnk}$ are given in the Appendix, eqs. (\[eq:F0lag\]-\[eq:dudd\]).
For the decay of a neutralino into a lighter one and $H_k^0$, we get [@neutree1] $$\begin{aligned}
&& \hspace{-1.5cm} \Gamma^{\rm tree}_{{{\tilde{\chi}}}^0} = \Gamma^{\mbox{\footnotesize
tree}}({{\tilde{\chi}}}^{0}_m \to
H_{k}^0\, {{\tilde{\chi}}}^{0}_n) = \nonumber\\[2mm]
&& \hspace{0.5cm}
\frac{g^{2}}{16 \pi\, m^{3}_m} \, \kappa(m^{2}_m, m^{2}_{H_{k}^0}, m^{2}_n)
\, |F^0_{mnk}|^{2}\,\left[ m^{2}_m - m^{2}_{H_{k}^0} + m^{2}_n + 2
(-1)^{\delta_{k3}}\, m_m m_n \right]\, .
$$ In our convention, the $4 \times 4$ neutralino mixing matrix $Z$, which diagonalizes the neutralino mass matrix $Y$, is real. Therefore, the neutralino mass parameters $m_m$ and $m_n$ can be positive or negative.
One-loop corrections
====================
We calculate the one-loop corrections to the amplitudes of the decays (\[eq:Hk0neu\]) stemming from fermion and sfermion exchange. The renormalization is done in the on-shell scheme. All one-, two-, and three-point functions [@thooft] used for calculating the loop integrals are given in the convention [@denner].\
The correction to the coupling $F^0_{mnk}$ is $$\label{eq:Fren}
F^{0\, {\rm corr.}}_{mnk} = F^0_{mnk} + \Delta F^0_{mnk}\, ,$$ with the ultraviolet (UV) finite one-loop correction $$\label{eq:DFmnk}
\Delta F^0_{mnk} = \sum_{\rm flavors}\, N_c^f\,
\left({\delta}F^{0\,(v)}_{mnk} + {\delta}F^{0\,(w)}_{mnk} +
{\delta}F^{0\,(c)}_{mnk} \right)\, ,$$ with the color factor $N_c^f = 1$ for (s)lepton and $N_c^f = 3$ for (s)quark exchange. $\sum_{\rm flavors}$ stands for the summation over all (s)fermion flavors, e. g. (top, stops), (bottom, sbottoms), (tau, staus), etc.. For convenience, the color factor $N_c^f$ is given only in the total correction term eq. (\[eq:DFmnk\]).
In our convention, both $F^0_{mnk}$ and $\Delta F^0_{mnk}$ are real. Therefore, the corrected widths can be written as $$\label{eq:Gcorr}
\Gamma^{\rm corr.}_p = \Gamma^{\rm tree}_p \, \left(1 + \frac{\Delta
F^0_{mnk}}{F^0_{mnk}} \right)^2\, ,$$ with the decaying particle $p = H_k^0$ or ${{\tilde{\chi}}}^0_m$.\
The [**vertex correction**]{} stems from the two diagrams shown in Figs. \[fig:feynman\]a and \[fig:feynman\]b. Because of the Majorana nature of the neutralinos the charge conjugated (s)fermion fields denoted by the superscript “$c$” can also circulate in the loop.\
For $h^0$ and $H^0$ ($a = 1,2$) we have $$\begin{aligned}
&&\hspace*{-1.1cm} g\, {\delta}F^{(v)}_{mna} =
\frac{s^f_a}{(4 \pi)^2}
\sum_{i=1}^2
\Bigg\{m_f\Big(a^{\tilde{f}}_{im}a^{\tilde{f}}_{in}+b^{\tilde{f}}_{im}
b^{\tilde{f}}_{in}\Big) \bigg[\Big(m_{m} + m_{n}\Big) C_0^{i} +
2\Big(m_{m} C_1^{i} + m_{n}C_2^{i}\Big)\bigg] +
\nonumber\\
&&\hspace*{-1.1cm}+
\Big(a^{\tilde{f}}_{im}b^{\tilde{f}}_{in}+
a^{\tilde{f}}_{in}b^{\tilde{f}}_{im}\Big)
\bigg[\Big(m_f^2+m_{\sf_i}^2+m_{m}m_{n}\Big)C_0^{i}+
\Big(m_{m}+m_{n}\Big)\Big(m_{m}C_1^{i}+m_{n}C_2^{i}\Big)
+B_0^{}\bigg]
\Bigg\}
\nonumber\\
&&\hspace*{-1.1cm}+
\frac{1}{(4 \pi)^2}
\sum_{i,j\,= 1}^2 G_{ija}^{\tilde{f}}\,
\Bigg[\Big(a^{\tilde{f}}_{im}b^{\tilde{f}}_{jn}+a^{\tilde{f}}_{jn}
b^{\tilde{f}}_{im}\Big)m_f C_0^{ij}-
\Big(a^{\tilde{f}}_{im}a^{\tilde{f}}_{jn}+
b^{\tilde{f}}_{im}b^{\tilde{f}}_{jn}\Big)
\Big(m_{m} C_1^{ij} + m_{n} C_2^{ij}\Big)
\Bigg]
\, . \label{eq:dvgFSa}\end{aligned}$$ For $A^0$ the vertex correction reads $$\begin{aligned}
&&\hspace*{-1.1cm} g\,
{\delta}F^{(v)}_{mn3} =
\frac{i\,s^f_3}{\left(4\pi\right)^2}
\sum_{i=1}^2
\Bigg\{m_f
\Big(a^{\tilde{f}}_{im}a^{\tilde{f}}_{in}+
b^{\tilde{f}}_{im}b^{\tilde{f}}_{in}\Big)
\Big(m_{m}+m_{n}\Big)\,C_0^{i}+
\nonumber\\
&&\hspace*{-1.1cm} +
\Big(a^{\tilde{f}}_{im}b^{\tilde{f}}_{in}+
a^{\tilde{f}}_{in}b^{\tilde{f}}_{im}\Big)\bigg[
\Big(m_{m} m_{n}-m_{\sf_i}^2+m_f^2\Big)C_0^{i}+
\Big(m_{n}-m_{m}\Big)\Big(m_{m}C_1^{i}-m_{n}C_2^{i}\Big)
-B_0\bigg]
\Bigg\}\nonumber\\
&&\hspace*{-1.1cm}+
\frac{i}{(4 \pi)^2}\,\sum_{i,j\,= 1}^2 G_{ij3}^{\tilde{f}}\,
\Bigg[\Big(a^{\tilde{f}}_{im}b^{\tilde{f}}_{jn}-a^{\tilde{f}}_{jn}
b^{\tilde{f}}_{im}\Big) \,m_f \,C^{ij}_0 +
\Big(a^{\tilde{f}}_{im}a^{\tilde{f}}_{jn}-
b^{\tilde{f}}_{im}b^{\tilde{f}}_{jn}\Big)\,
\Big(m_{m} C^{ij}_1 - m_{n} C^{ij}_2\Big)
\Bigg] . \,\label{eq:dvgFSc}
\end{aligned}$$
The abbreviations $B_0 = B_0(m_{H^0_k}^2,m_f^2,m_f^2)$, $C^i_{..} =
C_{..}(m_m^2,m_{H^0_k}^2,m_n^2,m_{\sf_i}^2,m_f^2,m_f^2)$, and $C^{ij}_{..}= C_{..}(m_m^2,m_{H^0_k}^2,
m_n^2,m_f^2,m_{\sf_i}^2,m_{\sf_j}^2)$ have been used.
The [**wave-function correction**]{} is given by $$\delta{F}_{mnk}^{0\,(w)}\,=\, \frac{\,1}{\,2}\,\bigg[\,
\delta{Z}^{H^0}_{lk}{F}_{mnl}^{0}+ \delta{Z}_{qm}^{{{\tilde{\chi}}}}{F}_{qnk}^{0}
+\delta{Z}_{qn}^{{{\tilde{\chi}}}}{F}_{mqk}^{0}\bigg]
\,,
\label{eq:dFw}$$ with the implicit summations over $l = 1,2$ for k = 1 or 2, $l =
3,4$ for $k = 3$, and $q = 1,\ldots,4$. $\delta{Z}_{}^{{{\tilde{\chi}}}}$ are the wave-function constant terms for the neutralinos given in (\[eq:dZchpp\]), (\[eq:dZchps\]). The wave-function constant terms for the Higgs bosons $\delta{Z}^{H^0}$ are $$\begin{aligned}
\label{eq:dZHkk}
\delta{Z}^{H^0}_{kk}&=&
-\;{\rm Re}\,\dot{\Pi}^{H^0}_{kk}(m_{H_k^0}^2)\, , \\ \label{eq:dZHlk}
\delta{Z}^{H^0}_{kl}&=& \frac1{m^2_{H_k^0}-m^2_{H_l^0}}\, {\rm
Re}\,\bigg\{\Pi^{H^0}_{kl}(m^2_{H_l^0}) - \Pi^{H^0}_{kl}(m^2_{H_k^0})\bigg\} \,
, \hspace{20mm}l\neq{k}\,,\end{aligned}$$ with $k,l = 1,2$ for the system $(h^0, H^0)$ and $k,l = 3,4$ for $(A^0, G^0)$. Eq. (\[eq:dZHlk\]) has been symmetrized with respect to $(k,l)$. This is due to the on-shell renormalization of the Higgs mixing angle $\alpha$ ($k,l=1,2$) or $\beta$ ($k,l=3,4$). In this scheme ([@chmasscorr], extending [@earlier] for quark and lepton mixing) the counter terms for the mixing angles are determined by the requirement that they cancel the antisymmetric parts of the wave-function corrections.
The decays of $A^0$ are a little complicated by the contribution of the $A^0-Z^0$ mixing in addition to the $A^0-G^0$ mixing in eq. (\[eq:dZHlk\]). Moreover, both depend on the gauge parameter $\xi$. However, the sum of these two contributions is independent of $\xi$, as it is shown in Appendix C. Here we work in the $\xi = 0$ (Landau) gauge, where the contribution of the $A^0-Z^0$ mixing vanishes, and use (\[eq:dZHlk\]) with $m_{H^0_4}=0$. The resulting on-shell $\tan\beta$ agrees with the one defined by the $A^0$–$Z^0$ mixing [@pokorski; @dabelstein; @chmasscorr].
The Higgs self-energy contributions due to fermions and sfermions are written as $$\Pi^{H^0}_{kl}(k^2) = \Pi^{H^0\,(a)}_{kl}(k^2) +
\Pi^{H^0\,(b)}_{kl}(k^2) + \Pi^{H^0\,(c)}_{kl} + T_{kl}\,.
\label{eq:TotalHiggsSE}$$ The fermion contribution $\Pi^{H^0\,(a)}$ (Fig. \[fig:feynman\]c) is $$\begin{aligned}
\Pi^{H^0\,(a)}_{ab}(k^2)&=&\frac{2\,s^f_a\,s^f_b}{\left(4\pi\right)^2}\,
\bigg[\Big(k^2-4m_f^2\Big)\,B_0(k^2,m_f^2,m_f^2)-2A_0(m_f^2)\bigg]\,,
\label{eq:PiHa1}
\\
\Pi^{H^0\,(a)}_{cd}(k^2)&=&-\frac{2\,s^f_c\,s^f_d}{\left(4\pi\right)^2} \,
\bigg[k^2\,B_0(k^2,m_f^2,m_f^2)-2A_0(m_f^2)\bigg]\,,
\label{eq:PiHa2}
\end{aligned}$$ with $a,b = 1,2$ and $c,d = 3,4$. The sfermion contributions $\Pi^{H^0\,(b)}$ (Fig. \[fig:feynman\]d) and $\Pi^{H^0\,(c)}$ (Fig. \[fig:feynman\]e) are $$\begin{aligned}
\Pi^{H^0\,(b)}_{kl}(k^2) & = &\frac{1}{\left(4\pi\right)^2}\;
\sum_{i,j=1}^2
G^{\sf}_{jik}\, G^{\sf}_{ijl}\;
B_0(k^2,m_{\sf_i}^2,m_{\sf_j}^2)\,,
\label{eq:PiHb}
\\
\Pi^{H^0\,(c)}_{kl}& = & {{\frac1{\left(4\pi\right)^2}}}\;
\sum_{i=1}^2
\bigg[h_f^2\,c_{kl}^\sf\,+\,g^2d_{kl}\,e_{ii}^\sf\bigg]
A_0(m_{\sf_i}^2)\,,
\label{eq:PiHc}\end{aligned}$$ where $k,l = 1,2$ or $3,4$. $T_{kl}$ in eq. (\[eq:TotalHiggsSE\]) represent momentum-independent contributions from the tadpole shifts [@pokorski; @dabelstein] and leading higher-order corrections. We include the latter by the renormalization group improvement as in Ref. [@twopointfunctions]. Since the zero-momentum contribution $\Pi^{H^0}_{kl}(0)$, including $T_{kl}$, is very large it is often resummed as in Refs. [@dabelstein; @carena]. In practice, we calculate the effective $m_{H^0_a}(a=1,2)$ and $\alpha$ obtained from the effective mass matrix, which includes the $\Pi^{H^0}_{ab}(k^2=0)$ contribution with $m_{A^0}$, ${\tan\beta}$, and the (s)quark parameters, and regard them as the lowest-order parameters. If one is replacing $\alpha$ in all the previous formulae with the effective one, the self energies $\Pi^{H^0}_{kl}(k^2)$ in the wave-function correction and $\delta\alpha$ must be replaced by $\Delta\Pi^{H^0}_{kl}(k^2)=\Pi^{H^0}_{kl}(k^2)-\Pi^{H^0}_{kl}(0)$. Nevertheless, the form of their sums eqs. (\[eq:dZHkk\],\[eq:dZHlk\]) is not affected by the elimination of $\Pi^{H^0}_{kl}(0)$.
The neutralino wave-function terms read $$\begin{aligned}
\hspace*{-1cm} \delta{Z}^{{{\tilde{\chi}}}}_{pp} & = & - \;{\rm Re}\,
\bigg\{\Pi^{{{\tilde{\chi}}}}_{pp}(m_{p}^2)+ 2m_{p}\,
\Big[m_{p}\dot{\Pi}^{{{\tilde{\chi}}}}_{pp}(m_{p}^2)+
m_f\dot{\Pi}^{{{\tilde{\chi}}}_S}_{pp}(m_{p}^2)\Big]\bigg\}\, ,
\label{eq:dZchpp}
\\
\label{eq:dZchps}
\hspace*{-1cm} \delta{Z}^{{{\tilde{\chi}}}}_{ps}&=&\frac1{m_{p}-m_{s}}\;
{\rm Re}\bigg\{
m_{s}\left[\Pi^{{{\tilde{\chi}}}}_{ps}(m_{s}^2) - \Pi^{{{\tilde{\chi}}}}_{ps}(m_{p}^2)\right] + m_f
\left[\Pi^{{{\tilde{\chi}}}\,S}_{ps}(m_{s}^2) - \Pi^{{{\tilde{\chi}}}\,S}_{ps}(m_{p}^2) \right]\bigg\} \,,\end{aligned}$$ $p\neq s$. As before, $\delta{Z}^{{{\tilde{\chi}}}}_{ps}$ in (\[eq:dZchps\]) has been symmetrized by subtracting the counter term for the rotation matrix $Z$ of the neutralinos [@chmasscorr]. The neutralino self-energies due to the sfermion-fermion loop (Fig. \[fig:feynman\]f ) are $$\begin{aligned}
\label{eq:neuself} \Pi^{{\tilde{\chi}}}_{ps}(k^2) & = & - \frac{1}{(4\pi)^2}\,
\sum_{i=1}^2(a^{\tilde{f}}_{ip} a^{\tilde{f}}_{is}+ b^{\tilde{f}}_{ip}
b^{\tilde{f}}_{is})\, B_1(k^2,m_f^2,m_{\sf_i}^2)\, ,
\\
\label{eq:neuselfS} \Pi^{{{\tilde{\chi}}}\,S}_{ps}(k^2) & = & \frac{1}{(4\pi)^2}\,
\sum_{i=1}^2(a^{\tilde{f}}_{ip} b^{\tilde{f}}_{is}+ a^{\tilde{f}}_{is}
b^{\tilde{f}}_{ip})\, B_0(k^2,m_f^2,m_{\sf_i}^2)\, ,\end{aligned}$$ see also [@twopointfunctions].
We need the [**counter term**]{} for the couplings $gF^0_{mnk}$, which is a function of the gauge couplings $g$, $g'$, the Higgs boson mixing angle $\alpha$ (for $k=1,2$) or $\beta$ (for $k=3$), and the neutralino rotation matrix $Z$, as shown in eq. (\[eq:F0tree\]). The counter terms for ($\alpha$, $\beta$) and $Z$ in the on-shell scheme [@chmasscorr] are already included in the wave-function corrections (\[eq:dZHlk\]) and (\[eq:dZchps\]), respectively. The remaining counter term of ${\delta}g$ and ${\delta}g'$ is, after being absorbed into the correction to $F^0_{mnk}$, $$\begin{aligned}
{\delta}F^{0\, (c)}_{mnk} &=& \sum_{x = 3,4}\,
d^x_k \left[\left(Z_{mx}Z_{n2}+Z_{m2}Z_{nx}\right)\, \frac{{\delta}g}{g}
- \left(Z_{mx}Z_{n1}+Z_{m1}Z_{nx}\right)\, \tan\theta_W\, \frac{{\delta}g'}{g'}\right]
\,.
\label{eq:dgF0c}\end{aligned}$$ We fix the electroweak gauge boson sector by $m_Z$, $m_W$, and $e$. One gets from the relations $g=e/s_W$, $g'=e/c_W$, and $c_W={m_W}/{m_Z}$ ($c_W\equiv \cos\theta_W$, $s_W \equiv \sin\theta_W$) [@sirlin; @denner] $$\frac{\delta g}{g}= \frac{\delta e}{e} +\frac{c_W^2}{2\,s_W^2}
\left(\frac{\delta m_W^2}{m_W^2}-\frac{\delta m_Z^2}{m_Z^2}\right)\, ,
\hspace{1.5cm}
\frac{\delta g'}{g'}= \frac{\delta e}{e} -
\frac12\left(\frac{\delta m_W^2}{m_W^2}-
\frac{\delta m_Z^2}{m_Z^2}\right)\,.
\label{eq:dggdgpgp}$$ The formulae for $\delta m_W$ and $\delta m_Z$ can be also found in [@chmasscorr] and for ${\delta}e/e$ in the Appendix \[sec:appB\].
Now all parts are given which are needed in order to calculate the (UV finite) one-loop contribution to the neutral Higgs boson-neutralino-neutralino coupling, eq. (\[eq:DFmnk\]). The vertex correction part ${\delta}F^{0\,(v)}_{mnk}$ is given by eqs. (\[eq:dvgFSa\]) and (\[eq:dvgFSc\]), the wave-function correction term ${\delta}F^{0\,(w)}_{mnk}$ by eq. (\[eq:dFw\]), and ${\delta}F^{0\,(c)}_{mnk}$ by eq. (\[eq:dgF0c\]).
Further, one has to note that the on-shell masses and the mixing of the neutralinos are not independent of each other. In fact, when the gauge and Higgs boson sectors are fixed, the neutralino sector is determined by three free parameters only. Here we follow the method given by [@chmasscorr]: The on-shell mass parameters $M$ and $\mu$ are defined as the elements of the on-shell mass matrix $X$ of charginos, and the on-shell mass parameter $M'$ is defined as the element of the on-shell mass matrix $Y$ of neutralinos . The finite correction $\Delta Y=Y-Y^{\rm tree}$, where $Y^{\rm tree}$ is the tree-level mass matrix in terms of the on-shell parameters $(M,\mu,M',m_Z,\sin\theta_W,\tan\beta)$, is calculated by eqs. (42–51) in [@chmasscorr]. The one-loop corrected on-shell masses $m_i$ and mixing matrix $Z=Z^{\rm tree}+\Delta Z$ are then obtained by diagonalizing $Y$. For a proper treatment of the loop corrections, the resulting shifts of the masses and the mixing matrix from the tree-level values have to be taken into account.
Numerical results
=================
For simplicity, we will take in the following (if not specified otherwise) for the soft breaking sfermion mass parameters of the first, second and third generation $M_{\tilde Q_{1,2}}=M_{\tilde
U_{1,2}}=M_{\tilde D_{1,2}}=M_{\tilde L_{1,2}}=M_{\tilde
E_{1,2}}=M_{\tilde Q_{3}}=\frac{10}9M_{\tilde U_{3}}
=\frac{10}{11}M_{\tilde D_{3}}=M_{\tilde L_{3}}=M_{\tilde E_{3}} =
M_{\tilde Q}=500$ GeV and for the trilinear couplings $A_t=A_b=A_{\tau}=A=300$ GeV. We take $m_Z=91.2$ GeV, $m_W=80$ GeV, $m_{A^0}=500$ GeV, $m_t=175$ GeV, $m_b=5$ GeV, and $m_{\tau}=1.8$ GeV. Masses of all other SM fermions are neglected. We use the GUT relations $M'=\frac53\tan^2{\theta_W}\,M$ and for the gluino mass $m_{\tilde{g}}=({\alpha_S}/{\alpha_{EW}})\sin^2{\theta_W}\,M$. The other input parameters are $\{\,\tan{\beta},\,M,\,\mu\}$ (all as on-shell parameters). For the values of the Yukawa couplings of the $3^{\rm rd}$ quark generation ($h_t$, $h_b$), we take the running ones at the scale of the decaying particle mass.
In our numerical analysis we have discussed four cases: the tree-level width, the corrections (\[eq:Fren\]–\[eq:Gcorr\]) with the tree-level $Z$ and $m_i$ (“conventional correction”), the corrections (\[eq:Fren\]–\[eq:Gcorr\]) with the one-loop corrected $Z$ and tree-level $m_i$ (“conventional $+$ $\Delta Z$ correction”), and the corrections (\[eq:Fren\]–\[eq:Gcorr\]) with the one-loop corrected $Z$ and one-loop corrected $m_i$ (full correction). The “conventional correction” corresponds to the correction to the gaugino-higgsino-Higgs boson coupling, “conventional $+$ $\Delta Z$ correction” includes the correction to the neutralino components, and the correction due to the shift of $m_i$ is added in the full correction.
In Fig. \[fig:mudependence\]a we show, as a function of $\mu$, the tree-level widths of $H^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_2$, $H^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_3$ and $H^0\to{{\tilde{\chi}}}^0_2+{{\tilde{\chi}}}^0_2$, respectively, for $\tan{\beta}=10$ and $M=150$ GeV. The $H^0$ mass is $m_{H^0}\sim500$ GeV. The widths vary with the gaugino and higgsino components of the various neutralino states. Fig. \[fig:mudependence\]b exhibits the corrections to the width of $H^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_2$: The “conventional”, “conventional $+\Delta Z$”, and full corrections are shown. One can see that, compared to the “conventional” correction, the corrections by the shifts $\Delta Z$ and $m_i$ cannot be neglected. Figs. \[fig:mudependence\]c and \[fig:mudependence\]d show the corrections to the widths of $H^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_3$ and $H^0\to{{\tilde{\chi}}}^0_2+{{\tilde{\chi}}}^0_2$, respectively. While the “conventional” correction is dominant for $\mu<0$ in Fig. \[fig:mudependence\]c, the “$\Delta Z$” correction is dominant in Fig. \[fig:mudependence\]d. The full corrections amount to several %.
\
In Fig. \[fig:MsQdependence\], we show the tree and corrected widths of (a) $H^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_2$ decay with $\tan\beta=10$, $M=500$ GeV and $\mu=150$ GeV, and those of (b) $A^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_3$ decay with $\tan\beta=50$ and $M=\mu=300$ GeV, as functions of $M_{\tilde{Q}}$. In Fig. \[fig:MsQdependence\]a, the decay is suppressed due to the small gaugino components of ${{\tilde{\chi}}}^0_1$ and ${{\tilde{\chi}}}^0_2$. The “conventional$+\Delta Z$” correction is close to the full correction and therefore not shown here. We see that the “conventional” correction is dominant. In contrast, the “$\Delta m$” correction in Fig. \[fig:MsQdependence\]b is large and negative (up to $-16$%), which dominates over the positive “conventional” correction (up to $+4$%). This is because the decay in Fig. \[fig:MsQdependence\]b is kinematically suppressed and sensitive to the shift of $m_i$. We note that the sfermion loop corrections do not decouple in large $M_{\tilde{Q}}$ limit, due to the supersymmetry breaking corrections [@superoblique] to the gaugino-higgsino-Higgs boson couplings.
Figs. \[fig:Adependence\]a and \[fig:Adependence\]b show the dependence of the widths on the trilinear coupling $A$ for the same decays modes and parameter sets as in Figs. \[fig:MsQdependence\]a and \[fig:MsQdependence\]b, respectively. The $A$ dependence is mainly caused by $A_t$ and numerically important in general.
We also discuss the related decays (\[eq:chidecay\]) of the neutralinos. In Figs. \[fig:NeuDecay\]a and \[fig:NeuDecay\]b we show the corrections to the width of the decays ${{\tilde{\chi}}}^0_3\to h^0+{{\tilde{\chi}}}^0_1$ and ${{\tilde{\chi}}}^0_4\to h^0+{{\tilde{\chi}}}^0_1$ as functions of $M_{\tilde{Q}}$ and $A$, respectively. The parameters are as in Fig. \[fig:MsQdependence\]a for Fig. \[fig:NeuDecay\]a and $\tan{\beta}=10$ and $M=\mu=300$ GeV for Fig. \[fig:NeuDecay\]b. The total correction can go up to 25%.
Finally, Fig. \[fig:lighthiggsdecay\] shows the $\mu$ dependence of the width of the decay $h^0\to{{\tilde{\chi}}}^0_1+{{\tilde{\chi}}}^0_1$ (\[eq:h0ch1ch1\]), both the tree-level value (Fig. \[fig:lighthiggsdecay\]a) and the relative one-loop full correction (Fig. \[fig:lighthiggsdecay\]b). This decay occurs when $\tilde{\chi}^0_1$ is sufficiently light and is mainly a U(1) gaugino to escape from the present direct search. In order to realize this case, we consider very small $M'$ and take the following parameters which are similar to those in Ref. [@Djouadi-Drees]: $M_{\tilde Q_{1,2}}=M_{\tilde U_{1,2}}=M_{\tilde D_{1,2}}=
M_{\tilde Q_{3}}=M_{\tilde U_{3}}=M_{\tilde D_{3}}=
2M_{\tilde L_{1,2}}=2M_{\tilde E_{1,2}}=
2M_{\tilde L_{3}}=2M_{\tilde E_{3}}=500$ GeV, $A_t=1000$ GeV, $A_b=A_{\tau}=0$ GeV, $M'=30$ GeV, $M=120$ GeV, $m_{\tilde{g}}=500$ GeV, ${\tan\beta}=20$, and $m_{A^0}=300$ GeV. The loop correction can be comparable to or even larger than the tree-level width as observed in Ref. [@Djouadi-Drees]. Although the decay width is much smaller than the other modes, the effect of the loop correction might be seen in precision studies of $h^0$ at a linear collider [@Djouadi-Drees] and in neutralino dark matter search [@Djouadi-Drees; @dark1; @dark2].
Conclusions {#sec:concl}
============
We have presented the calculation of the one-loop corrections to the decays$(h^0, H^0, A^0) \to {{\tilde{\chi}}}^0_m + {{\tilde{\chi}}}^0_n$ and ${{\tilde{\chi}}}^0_m \to (h^0, H^0, A^0) + {{\tilde{\chi}}}^0_n$, $(m,n=1,\ldots,4)$, with all fermions and sfermions in the loop. These decays are special in the sense that they require particular care in the treatment of the neutralino mixing and mass matrix in a scheme, where all parameters in the neutralino mass matrix $Y$ and mixing matrix $Z$ are defined on-shell. We have shown the importance of the corrections to these matrices in addition to the conventional corrections (vertex and wave-function corrections with counter terms). We have studied the dependence on the parameters $M$, $\mu$, $A$, $M_{\tilde Q}$, and $\tan\beta$. The corrections to the widths of the decays $(h^0, H^0, A^0) \to {{\tilde{\chi}}}^0_m + {{\tilde{\chi}}}^0_n$ can go up to $\sim$ 15%, those of the decays ${{\tilde{\chi}}}^0_m \to (h^0,
H^0, A^0) + {{\tilde{\chi}}}^0_n$ to $\sim$ 25%. For the invisible decay $h^0
\to {{\tilde{\chi}}}^0_1 + {{\tilde{\chi}}}^0_1$, giving up the GUT relation for $M'$, one even gets corrections up to 140%.
Acknowledgements {#acknowledgements .unnumbered}
================
The work of Y.Y. was supported in part by the Grant–in–aid for Scientific Research from Japan Society for the Promotion of Science, No. 12740131. The work was also supported by the “Fonds zur Förderung der wissenschaftlichen Forschung” of Austria, project no. P13139-PHY and the EU TMR Network Contract HPRN-CT-2000-00149.
Appendix {#sec:app .unnumbered}
========
In the following we give the formulae for the couplings and for ${\delta}e$. Furthermore, the proof of the gauge independence of the processes considered will be given.
Coupling parameters\[sec:appA\]
===============================
The $H^0_k\,{{\tilde\chi^0}}_m\,{{\tilde\chi^0}}_n$ interaction is given by $${\cal L} = - \frac12\, g\, H^0_a\,\bar{{\tilde{\chi}}}^0_m\,F^0_{mna}\,{{\tilde\chi^0}}_n +
\frac{i}{2}\, g\, H^0_c\,\bar{{\tilde{\chi}}}^0_m\,F^0_{mnc}\,\gamma^5\,{{\tilde\chi^0}}_n\, ,
\label{eq:F0lag}$$ with $a = 1,2$, $c = 3,4$, and $$\begin{aligned}
F^0_{mnk}&=& \hphantom{+} \frac{d^u_k}{2} \,
\bigg[Z_{m4}Z_{n2}+Z_{n4}Z_{m2}-
{{\tan{\theta_W}}}(Z_{m4}Z_{n1}+Z_{n4}Z_{m1})\bigg]
\nonumber\\&&
+ \frac{d^d_k}{2} \, \bigg[Z_{m3}Z_{n2}+Z_{n3}Z_{m2}-
{{\tan{\theta_W}}}(Z_{m3}Z_{n1}+Z_{n3}Z_{m1})\bigg]\,.
\label{eq:F0tree}\end{aligned}$$ The $Z_{mn}$ are the elements of the neutralino mixing matrix which diagonalizes the neutralino mass matrix and $$\begin{aligned}
d_k^4 = d^u_k&=&
\Big(-{{\cos{\alpha}}},\,-{{\sin{\alpha}}},\,\hphantom{-}{{\cos{\beta}}},\,\hphantom{-}{{\sin{\beta}}}\Big)_k\,,
\nonumber\\
d_k^3 = d^d_k&=&
\Big(-{{\sin{\alpha}}},\,\hphantom{-}{{\cos{\alpha}}},
\,-{{\sin{\beta}}},\,\hphantom{-}{{\cos{\beta}}}\Big)_k\,.
\label{eq:dudd}
\end{aligned}$$ The superscript “$u$” denotes an up-type and “$d$” a down-type fermion.
The neutral Higgs boson-fermion-fermion couplings, defined by ${\cal L} = s^f_k\,
H_k^0\, \bar f \, f$ are $$s^{u}_a = \frac{h_{u}}{\sqrt{2}}\;d^{u}_a\,,
\hspace{10mm}
s^{u}_c = i\,\frac{h_{u}}{\sqrt{2}}\;d^{u}_c\,,
\hspace{10mm}
s^{d}_a = -\,\frac{h_{d}}{\sqrt{2}}\;d^{d}_a\,,
\hspace{10mm}
s^{d}_c\;=\;-\,i\,\frac{h_{d}}{\sqrt{2}}\;d^{d}_c\,,
\label{eq:sfk}$$ with $a = 1,2$ and $c = 3,4$, using the Yukawa couplings $$h_u = \frac{g\, m_u}{\sqrt2\, m_W \sin\beta}\,,
\hspace{10mm}
h_d = \frac{g\, m_d}{\sqrt2\, m_W \cos\beta}\, .
\label{eq:yuk}$$
The fermion-sfermion-neutralino coupling parameters $a^{\sf}_{ip}$ and $b^{\sf}_{ip}$ ($i = 1,2,\, p = 1,\ldots,4$) have the form $$\begin{aligned}
\label{eq:neufsfcoup}
a^{\sf}_{ip} &=& g f_{Lp}^f R^\sf_{i1} +
h_f Z_{px} R^\sf_{i2}\,,
\\
\label{eq:neufsfcoup1}
b^{\sf}_{ip} &=&
g f_{Rp}^f R^\sf_{i2} +
h_f Z_{px}R^\sf_{i1}\,,\end{aligned}$$ with $x=3$ for down-type and $x=4$ for up-type fermions, $R^\sf$ the $2 \times 2$ sfermion rotation matrix, $$\begin{aligned}
\label{eq:fLi}
f_{Lp}^f &=& \sqrt2\, \bigg[ \Big(e_f-{{I_f^{3L}}}\Big) {{\tan{\theta_W}}}Z_{p1}+{{I_f^{3L}}}Z_{p2}\bigg]\,,
\\
\label{eq:fRi}
f_{Rp}^f &=& -\sqrt2\, e_f \,{{\tan{\theta_W}}}\, Z_{p1}\,.\end{aligned}$$ ${{I_f^{3L}}}$ denotes the SU(2)$_L$ isospin and $e_f$ the charge of the fermion $f$.
The $H^0_k\, \sf_i^*\, \sf_j$ couplings [@gunion] are $$\begin{aligned}
\label{eq:7}
&& \hspace*{-2cm}
G_{ij1}^\su \equiv G(h^0\, \su_i^*\, \su_j) =
\left(R^\su\, G^\su_{LR1}\, (R^{\su})^T \right)_{ij} = \nonumber\\
&& \hspace*{-2cm}
-g R^{\su} \left(
\begin{array}{cc} -\frac{m_Z}{c_W}(I_u^{3 L} - e_u s_W^2) {{\mbox{s}}}_{{\alpha}+ {\beta}} +
\frac{m_u^2}{m_W{{\mbox{s}}}_{\beta}} {{\mbox{c}}}_{\alpha}& \frac{m_u}{2 m_W {{\mbox{s}}}_{\beta}} (A_u {{\mbox{c}}}_{\alpha}+ \mu
{{\mbox{s}}}_{\alpha})\\ \frac{m_u}{2 m_W {{\mbox{s}}}_{\beta}} (A_u {{\mbox{c}}}_{\alpha}+ \mu {{\mbox{s}}}_{\alpha}) &
-\frac{m_Z}{c_W} e_u s_W^2 {{\mbox{s}}}_{{\alpha}+ {\beta}} + \frac{m_u^2}{m_W {{\mbox{s}}}_{\beta}} {{\mbox{c}}}_{\alpha}\end{array}
\right)(R^{\su})^T\, ,\end{aligned}$$ $$\begin{aligned}
\label{eq:8}
&& \hspace*{-2.9cm}
G_{ij1}^\sd \equiv G(h^0\, \sd_i^*\, \sd_j) = \left(R^\sd\, G^\sd_{LR1} \,
(R^{\sd})^T \right)_{ij} = \nonumber\\
&& \hspace*{-2.9cm}
g R^{\sd}
\left( \begin{array}{cc} \frac{m_Z}{c_W}(I_d^{3 L} - e_d s_W^2) {{\mbox{s}}}_{{\alpha}+ {\beta}}
+ \frac{m_d^2} {m_W{{\mbox{c}}}_{\beta}} {{\mbox{s}}}_{\alpha}& \frac{m_d}{2 m_W {{\mbox{c}}}_{\beta}} (A_d {{\mbox{s}}}_{\alpha}+
\mu {{\mbox{c}}}_{\alpha})\\ \frac{m_d}{2 m_W {{\mbox{c}}}_{\beta}} (A_d {{\mbox{s}}}_{\alpha}+ \mu {{\mbox{c}}}_{\alpha}) &
\frac{m_Z}{c_W} e_d s_W^2 {{\mbox{s}}}_{{\alpha}+ {\beta}} + \frac{m_d^2}{m_W {{\mbox{c}}}_{\beta}} {{\mbox{s}}}_{\alpha}\end{array}
\right)(R^{\sd})^T\, ,\end{aligned}$$ $$\begin{aligned}
& G_{ij2}^\sf \equiv G(H^0\, \sf_i^*\, \sf_j) = G(h^0 \sf_i^* \sf_j)
\mbox{ with } {\alpha}\to {\alpha}- \frac\pi2\,,
\label{eq:9}\end{aligned}$$ (i. e. $\ \sin{\alpha}\equiv {{\mbox{s}}}_{\alpha}\to -{{\mbox{c}}}_{\alpha}\,,\ \cos{\alpha}\equiv {{\mbox{c}}}_{\alpha}\to
{{\mbox{s}}}_{\alpha}\,,\ \mbox{and}\; \sin({\alpha}+{\beta}) \equiv {{\mbox{s}}}_{{\alpha}+{\beta}} \to
-{{\mbox{c}}}_{{\alpha}+{\beta}})$,
$$\begin{aligned}
\label{eq:10}
G_{ij3}^\su &=& G(A^0\, \su_i^*\, \su_j) = \frac{i g}{2 m_W}
\left(
\begin{array}{cc} \hphantom{-}0 & 1\\ -1 & 0
\end{array}
\right)_{ij}\, m_u (A_u \cot{\beta}+ \mu)\, ,\end{aligned}$$
$$\begin{aligned}
\label{eq:11}
G_{ij3}^\sd &=& G(A^0 \sd_i^* \sd_j) = \frac{i g}{2 m_W} \left(
\begin{array}{cc} \hphantom{-}0 & 1\\ -1 & 0
\end{array}
\right)_{ij}\,m_d (A_d \tan{\beta}+ \mu)\, ,\end{aligned}$$
$$\begin{aligned}
& G_{ij4}^\sf \equiv G(G^0\, \sf_i^*\, \sf_j) = G(A^0 \sf_i^* \sf_j)
\mbox{ with } {\beta}\to {\beta}- \frac\pi2\,,
\label{eq:12}\end{aligned}$$
(i. e. $\ \tan{\beta}\leftrightarrow -\cot{\beta}$). The superscript “$\su$” (“$\sd$”) denotes an up-type (down-type) sfermion.
The $H^0_k\,H^0_l\,\sf_i^*\,\sf_j$ interaction is given by $${\cal L} = - \frac12\,\left( h_f^2\, c_{kl}^\sf\, {\delta}_{ij} + g^2\,d_{kl}\, e_{ij}^\sf\right)\,
H^0_k\,H^0_l\,\sf_i^*\,\sf_j\, ,$$ with $$\begin{aligned}
c^\su_{kl}&=&
\left(\begin{array}{cccc}
{{\cos^2{\alpha}}}& \hphantom{-}{{\cos{\alpha}}}\,{{\sin{\alpha}}}& 0 & 0\\
\hphantom{-}{{\cos{\alpha}}}\,{{\sin{\alpha}}}& {{\sin^2{\alpha}}}& 0 & 0 \\
0 & 0 & {{\cos^2{\beta}}}& \hphantom{-}{{\cos{\beta}}}\,{{\sin{\beta}}}\\
0 & 0 & \hphantom{-}{{\cos{\beta}}}\,{{\sin{\beta}}}& {{\sin^2{\beta}}}\\
\end{array}\right)_{kl}\,,\\[3mm]
c^\sd_{kl}&=&
\left(\begin{array}{cccc}
{{\sin^2{\alpha}}}& -{{\cos{\alpha}}}\,{{\sin{\alpha}}}& 0 & 0\\
-{{\cos{\alpha}}}\,{{\sin{\alpha}}}& {{\cos^2{\alpha}}}& 0 & 0 \\
0 & 0 & {{\sin^2{\beta}}}& -{{\cos{\beta}}}\,{{\sin{\beta}}}\\
0 & 0 & -{{\cos{\beta}}}\,{{\sin{\beta}}}& {{\cos^2{\beta}}}\\
\end{array}\right)_{kl}\,,\\[3mm]
d_{kl} & = & c^\sd_{kl} - c^\su_{kl}\,,\\[3mm]
e_{ij}^\sf & = & \frac{1}{2\, c_W^2}\,\left[(I^{3L}_f - e_f
s_W^2)\, R_{i1}^\sf\, R_{j1}^\sf +
e_f\, s_W^2\, R_{i2}^\sf\, R_{j2}^\sf \right]\, .\end{aligned}$$
Counter term ${\delta}e$\[sec:appB\]
====================================
When we give the renormalized electric charge in the Thomson limit with the measured fine structure constant $\alpha=e_0^2/(4\pi)=1/137$, the counter term $\delta e_0$ is given by the general form [@denner] $$\frac{{\delta}e_0}{e_0} =\frac12\;\dot{\Pi}^{AA}_T(0)-\frac{s_W}{c_W}\,
\frac{\Pi^{AZ}_T(0)}{m_Z^2}\, , \label{eq:dZe}$$ with the momentum derivative of the transverse photon self-energy $\dot{\Pi}^{AA}_T$ and the $\gamma-Z^0$ mixing self-energy, both for the on-shell photon ($p^2=0$). Fermions and sfermions do not contribute to $\Pi^{AZ}_T(0)$ as a consequence of the fact that the physical photon is massless to all orders. However, the contribution of light hadrons to $\dot{\Pi}^{AA}_T(0)$ has a large theoretical uncertainty [@alphamz1; @denner]. To avoid this problem, in this work we use the running coupling at $Q=m_Z$, $\alpha(m_Z)=e^2(m_Z)/(4\pi)=1/129$ as input. The counter term $\delta e(m_Z)$ then becomes $$\frac{{\delta}e}{e} = \frac{e^2\, e_f^2}{(4\pi)^2}\, \left[
\frac{2}{3}\left(
\Delta + \log\frac{Q^2}{x_f^2} \right) +
\sum_{i=1}^2\frac{1}{6}\left(
\Delta + \log\frac{Q^2}{m_{\sf_i}^2} \right)\right] \, ,
\label{eq:de/e2}$$ with $x_f=m_Z$ for all $m_f<m_Z$ and $x_t=m_t$. Here $\Delta$ denotes the UV divergence factor.
Proof of the $\xi$ independence
===============================
We investigate the wave-function corrections to the process $$A^0(p) \to {{\tilde{\chi}}}^0_m (k_1) + {{\tilde{\chi}}}^0_n (k_2) \, .
\label{eq:A0neu}$$ Both the contributions of the transitions $A^0 \to G^0 \to {{\tilde{\chi}}}^0_m + {{\tilde{\chi}}}^0_n$ and $A^0 \to Z^0 \to {{\tilde{\chi}}}^0_m + {{\tilde{\chi}}}^0_n$ have a dependence on the gauge parameter $\xi= \xi_Z$ in the propagators of $(G^0,Z^0)$. We show that the sum of these contributions is independent of $\xi$.\
We start from the matrix elements in a general $R_{\xi}$ gauge, $$\begin{aligned}
{{\cal M}}^G & = & ( i\, \Pi_{AG}) \frac{i}{p^2 - \xi m_Z^2} (- g F_{mn4}^0 )
\bar u(k_1) \gamma^5 v(k_2)\, ,\\
\label{eq:MG1}
{{\cal M}}^Z & = & ( -i p^\mu \Pi_{AZ}) \frac{i}{p^2 - m_Z^2}
\left( - g_{\mu\nu} + (1 - \xi)
\frac{p_\mu p_\nu}{p^2 - \xi m_Z^2}\right) (-i \frac{g}{c_W} O^{''L}_{mn})
\bar u(k_1) \gamma^\nu \gamma^5 v(k_2) \, .\nonumber\\
\label{eq:MZ1}\end{aligned}$$ The self-energies $\Pi_{AG}$ and $\Pi_{AZ}$ by (s)fermion one-loop contributions are $\xi$ independent. The $Z^0{{\tilde{\chi}}}^0_m{{\tilde{\chi}}}^0_n$ couplings $O^{''L}_{mn}$ are $$\hspace*{-1cm} O^{''L}_{mn} = \frac{1}{2} (-Z_{m3} Z_{n3} + Z_{m4} Z_{n4})\, .
\label{eq:Oppdef}$$ As limiting cases, ${{\cal M}}^{G} = 0$ in the physical unitary gauge $\xi \to \infty$ and ${{\cal M}}^{Z} = 0$ in the $\xi = 0$ (Landau) gauge. Note that the tadpole contributions have to be included [@pokorski; @dabelstein] in $\Pi_{AG}$.
We can write ${{\cal M}}^G$ directly as $${{\cal M}}^G = \frac{1}{p^2 - \xi m_Z^2}\,g\, \Pi_{AG}(p^2)\, F_{mn4}^0\,
\bar u(k_1) \gamma^5 v(k_2)\, .
\label{eq:MG2}$$ For ${{\cal M}}^Z$ we first contract the Lorentz indices, $$\begin{aligned}
&&\hspace*{-1cm} p^\mu \,\left( - g_{\mu\nu} + (1 - \xi)
\frac{p_\mu p_\nu}{p^2 - \xi m_Z^2}\right)
\bar u(k_1) \gamma^\nu \gamma^5 v(k_2)
= \left( \frac{(1 - \xi)\, p^2}{p^2 - \xi m_Z^2} - 1\right)\, \bar u(k_1)\psla
\gamma^5 v(k_2) \, ,\nonumber
\label{eq:tmp1}\end{aligned}$$ and use $\bar u(k_1)\psla \gamma^5 v(k_2) = \bar u(k_1) (\ksla_1 +
\ksla_2) \gamma^5 v(k_2) = (m_m + m_n) \bar u(k_1) \gamma^5
v(k_2)$. So we get $${{\cal M}}^Z = - \frac{i}{p^2 - m_Z^2}
\left( \frac{(1 - \xi)\, p^2}{p^2 - \xi m_Z^2} - 1\right)
\frac{g}{c_W} \, \Pi_{AZ}(p^2)\, O^{''L}_{mn}\,
(m_m + m_n)\,
\bar u(k_1) \gamma^5 v(k_2)\, . \label{eq:MZ2}$$ We use the Slavnov-Taylor identity (see also [@dabelstein], eq. (3.7)) $$p^2\, \Pi_{AZ}(p^2) + i\, m_Z\, \Pi_{AG}(p^2) = 0\,,
\label{eq:wardID}$$ and split the sum ${{\cal M}}^G + {{\cal M}}^Z$ in an obviously $\xi$ independent and possibly dependent part, $$\begin{aligned}
\hspace*{-0.8cm} {{\cal M}}^{G + Z}
& = & \bigg\{ \frac{i}{p^2 - m_Z^2}\frac{g}{c_W} \,
\Pi_{AZ}(p^2)\, O^{''L}_{mn}\,
(m_m + m_n) \nonumber\\
&&\hspace*{-2.2cm}+\, i\,g \, \Pi_{AZ}(p^2) \frac{p^2}{p^2 - \xi m_Z^2}\left(
\frac{F_{mn4}^0}{m_Z} - \frac{(1 - \xi)}{p^2 - m_Z^2} \frac{O^{''L}_{mn}}{c_W}
(m_m + m_n) \right)\bigg\} \, \bar u(k_1) \gamma^5 v(k_2) \, .
\label{eq:MGplusZ1}\end{aligned}$$ With the relation (proved later) $$m_Z\, c_W\, F^0_{mn4} = - O^{''L}_{mn}\, (m_m + m_n)\,,
\label{eq:F02Opp}$$ we get for the part written in the brackets in eq. (\[eq:MGplusZ1\]) $$- (m_m + m_n) \frac{O^{''L}_{mn}}{c_W} \left( \frac{1}{m_Z^2}
+ \frac{1 - \xi}{p^2 - m_Z^2}\right) =
- (m_m + m_n) \frac{O^{''L}_{mn}}{m_Z^2\,c_W}
\frac{p^2 - \xi m_Z^2}{p^2 - m_Z^2}\, ,
\label{eq:tmp2}$$ and therefore the final result $${{\cal M}}^{G + Z} =- i\, \frac{g}{m_Z^2\,c_W}\,\Pi_{AZ}(p^2)\, O^{''L}_{mn}\,
(m_m + m_n) \, \bar u(k_1) \gamma^5 v(k_2)\, .
\label{eq:res1}$$ The $\xi$ dependence is completely cancelled in (\[eq:res1\]).
Finally, we prove (\[eq:F02Opp\]). With the abbreviation $A_{ij} = (Z_{mi} Z_{nj} + Z_{mj} Z_{ni})$ and knowing the entries of the neutralino tree-level mass matrix $Y$ (see e. g. eq. (35) in [@chmasscorr]), one can write $F^0_{mn4}$ as $$2\,m_Z\,c_W\,F^0_{mn4} = Y_{13} A_{31} + Y_{23} A_{32}
- Y_{14} A_{41} - Y_{24} A_{42} \,.
\label{eq:tmp6}$$ Next we add and subtract the terms $Y_{33} A_{33} + Y_{43} A_{34}$ and $Y_{34}
A_{43} + Y_{44} A_{44}$. Exploiting the fact that $Y_{33} = Y_{44} = 0$ and $Y_{34} = Y_{43}$, we get $$2\,m_Z\,c_W\,F^0_{mn4} = \sum_k \left( Y_{k3} A_{3k} - Y_{k4} A_{4k} \right)
\,.
\label{eq:tmp7}$$ Writing the entries of $Y$ in terms of neutralino masses, $Y_{kj} = \sum_l m_l
Z_{lk} Z_{lj}$, and using $\sum_k Z_{ik} Z_{jk} = \delta_{ij}$ we get $$\begin{aligned}
2\,m_Z\,c_W\,F^0_{mn4} & = & \sum_{k,l} m_l Z_{lk} \left[Z_{l3}\left(Z_{m3}
Z_{nk} + Z_{mk} Z_{n3}\right) - Z_{l4}\left(Z_{m4} Z_{nk} + Z_{mk}
Z_{n4}\right)\right]
\nonumber\\ & = & \sum_{l} m_l \left( \delta_{nl} Z_{l3} Z_{m3} +
\delta_{ml} Z_{l3} Z_{n3} - \delta_{nl} Z_{l4} Z_{m4} -
\delta_{ml} Z_{l4} Z_{n4} \right) \nonumber\\ & = &
m_n Z_{n3}Z_{m3} + m_m Z_{m3}Z_{n3} - m_n Z_{n4}Z_{m4} - m_m Z_{m4}Z_{n4}
\nonumber \\
& = & \left(Z_{m3} Z_{n3} - Z_{m4}Z_{n4}\right) (m_m + m_n) =
- 2\, O^{''L}_{mn}\, (m_m + m_n)\,.\label{eq:tmp8}\end{aligned}$$ Therefore, eq. (\[eq:F02Opp\]) is proven.
However, from the Slavnov-Taylor identity one can prove in general that the same cancellation of the gauge dependent parts in $G^0$ and $Z^0$ propagators occurs for any one-loop two-body decay of $A^0$.
[99]{}
H. P. Nilles, Phys. Rep. [**110**]{}, 1 (1984); H. E. Haber and G. L. Kane, Phys. Rep. [**117**]{}, 75 (1985); R. Barbieri, Riv. Nuov. Cim. [**11**]{}, 1 (1988).
J. F. Gunion, H. E. Haber, G. L. Kane, and S. Dawson, The Higgs Hunter’s Guide, Addison-Wesley (1990); J. F. Gunion and H. E. Haber, Nucl. Phys. [**B272**]{}, 1 (1986); [**B402**]{}, 567(E) (1993). H. Baer, D. Dicus, M. Drees, and X. Tata, Phys. Rev. D [**36**]{}, 1363 (1987); J. F. Gunion and H. E. Haber, Nucl. Phys. [**B307**]{}, 445 (1988); [**B402**]{}, 569(E) (1993); K. Griest and H. E. Haber, Phys. Rev. D [**37**]{}, 719 (1988).
A. Djouadi, J. Kalinowski, and P. M. Zerwas, Z. Phys. C [**57**]{}, 569 (1993); A. Djouadi, P. Janot, J. Kalinowski, and P. M. Zerwas, Phys. Lett. B [**376**]{}, 220 (1996); A. Djouadi, J. Kalinowski, P. Ohmann, and P. M. Zerwas, Z. Phys. C [**74**]{}, 93 (1997); A. Djouadi, Mod. Phys. Lett. A [**14**]{}, 359 (1999); G. Bélanger, F. Boudjema, F. Donato, R. Godbole, and S. Rosier-Lees, Nucl. Phys. [**B581**]{}, 3 (2000). Wang Lang–Hui, Ma We–Gan, Zhang Ren–You, and Jiang Yi, Phys. Rev. D [**64**]{}, 115004 (2001).
M. Drees, M. M. Nojiri, D. P. Roy, and Y. Yamada, Phys. Rev. D [**56**]{}, 276 (1997); [**64**]{}, 039901(E) (2001).
A. Djouadi, M. Drees, P. Fileviez Perez, and M. Mühlleitner, hep-ph/0109283
M. Drees and M. M. Nojiri, Phys. Rev. D [**47**]{}, 4226 (1993); [**48**]{}, 3483 (1993).
G. Jungman, M. Kamionkowski, and K. Griest, Phys. Rep. [**267**]{}, 195 (1996); A. B. Lahanas, D. V. Nanopoulos, and V. C. Spanos, Mod. Phys. Lett. A [**16**]{}, 1229 (2001); Phys. Lett. B [**518**]{}, 94 (2001); M. Drees, Y. G. Kim, T. Kobayashi, and M. M. Nojiri, Phys. Rev. D [**63**]{}, 115009 (2001).
A. Bartl, H. Fraas, and W. Majerotto, Nucl. Phys. [**B278**]{}, 1 (1986); S. Ambrosanio and B. Mele, Phys. Rev. D [**52**]{}, 3900 (1995).
H. Eberl, M. Kincel, W. Majerotto, and Y. Yamada, Phys. Rev. D [**64**]{}, 115013 (2001).
J. F. Gunion and H. E. Haber, Phys. Rev. D [**37**]{}, 2515 (1988).
S. Ambrosanio and B. Mele, Phys. Rev. D [**53**]{}, 2541 (1996).
A. Bartl, W. Majerottok, and W. Porod, Phys. Lett. B [**465**]{}, 187 (1999); A. Djouadi, Y. Mambrini, and M. Mühlleitner, Eur. Phys. J. C [**20**]{}, 563 (2001).
G. ’t Hooft and M. Veltman, Nucl. Phys. [**B153**]{}, 365 (1979); G. Passarino and M. Veltman, Nucl. Phys. [**B160**]{}, 151 (1979).
A. Denner, Fortschr. Phys. [**41**]{}, 307 (1993).
A. Denner and T. Sack, Nucl. Phys. [**B347**]{}, 203 (1990); B. A. Kniehl and A. Pilaftsis, Nucl. Phys. [**B474**]{}, 286 (1996).
P. H. Chankowski, S. Pokorski, and J. Rosiek, Phys. Lett. B [**274**]{}, 191 (1992); Nucl. Phys. [**B423**]{}, 437; 497 (1994).
A. Dabelstein, Z. Phys. C [**67**]{}, 495 (1995); Nucl. Phys. [**B456**]{}, 25 (1995).
D. Pierce and A. Papadopoulos, Phys. Rev. D [**50**]{}, 565 (1994); Nucl. Phys. [**B430**]{}, 278 (1994).
M. Carena, M. Quirós, and C. E. M. Wagner, Nucl. Phys. [**B461**]{}, 407 (1996).
A. Sirlin, Phys. Rev. D [**22**]{}, 971 (1980); K.-I. Aoki, Z. Hioki, R. Kawabe, M. Konuma, and T. Muta, Prog. Theor. Phys. Suppl. [**73**]{}, 1 (1982); M. Böhm, H. Spiesberger, and W. Hollik, Fortschr. Phys. [**34**]{}, 687 (1986).
P. H. Chankowski, Phys. Rev. D [**41**]{}, 2877 (1990); H.-C. Cheng, J. L. Feng, and N. Polonsky, Phys. Rev. D [**57**]{}, 152 (1998); S. Kiyoura, M. M. Nojiri, D. M. Pierce, and Y. Yamada, Phys. Rev. D [**58**]{}, 075002 (1998).
H. Burkhardt, F. Jegerlehner, G. Penzo, and C. Verzegnassi, Z. Phys. C [**43**]{}, 497 (1989).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nanoscale photonic crystal cavity optomechanical devices enable detection of nanomechanical phenomena with a sensitivity sufficient to observe quantum effects. Here we present the design of a one-dimensional air-mode photonic crystal cavity patterned in a silicon nitride nanobeam, and show that it forms the basis for cavity optomechanical split-beam and paddle nanocavity devices useful for force detection and nonlinear quantum sensing. The air-mode of this device is advantageous for optomechanical coupling, while also having ultrahigh optical quality factor $Q_o\sim 10^6$ despite its proximity to the light-line and the relatively low refractive index of silicon nitride. Paddle nanocavities realized from this device have a quadratic coupling coefficient $g^{(2)}/2\pi$ = 10 MHz/nm$^{2}$, and their performance within the context of quantum optomechanics experiments is analyzed.'
address: |
Institute For Quantum Science And Technology, University of Calgary, AB, Canada, T2N 1N4\
NRC National Institute For Nanotechnology, 11421 Saskatchewan Drive NW, Edmonton, AB, Canada, T6G 2M9
author:
- 'Chris Healey, Hamidreza Kaviani, and Paul E. Barclay'
bibliography:
- 'squidrefs.bib'
title: Air mode silicon nitride photonic crystals and their application to nonlinear quantum optomechanical sensing
---
Introduction
============
One-dimensional photonic crystal cavities (PCC’s) patterned in nanobeams have many uses in both fundamental and practical applications of cavity optomechanics [@ref:aspelmeyer2014co; @ref:favero2014fom; @ref:metcalfe2014aco], for instance to observe quantum correlations at room temperature [@purdy2017quantum], cooling of a nanomechanical oscillator into its quantum ground state [@Chan2011], ultra-sensitive torque magnetometry [@Wu:17], and to realize laser [@Zhang:10] and electro-optic modulator technologies [@Shakoor:14]. Many methodologies for designing defects within nanobeam waveguides typically patterned with holes to define PCCs have been discussed. One of the most successful approaches relies on tailoring the envelope of the electric field in the central region of the nanobeam by gradually varying the hole shape or spacing [@Chan:09]. By following a deterministic design procedure based on gradually varying the bandgap of each unit cell of the nanobeam [@Quan10; @Quan11], optical modes with ultra-high quality factors ($Q_o$) can be created with fields concentrated in the high-index dielectric or lower-index air/vacuum regions, referred to as dielectric- and air-modes, respectively.
Cavities supporting air-modes are advantageous for applications such as sensing and optomechanics due to their high concentration of optical energy in the lower index regions surrounding the mechanically moving boundaries of the device [@Johnson02]. Air-mode PCC’s have been demonstrated with $Q_o$ > $10^5$ at optical [@Liang2015] frequencies, and have applications such as temperature and refractive index sensing [@Yang:15] and detecting single nanoparticles [@Liang2015]. They are also the basis for optomechanical devices such as split-beam cavities (SBC’s) [@Hryciw:13; @ref:wu2014ddo; @zhang2016chip; @lin2015design], where a complete gap is introduced into the centre of the photonic crystal, and paddle nanocavities (PNC’s) [@Healey:15; @Kaviani:15], where a nano-mechanical resonator is inserted into the gap of the split-beam cavity and interacts strongly with the optical mode. In these devices, the electric fields concentrated in the air and overlapping with the boundaries of the nanomechanical resonator allow sensitive transduction of the mechanical motion and enable studies of magnetometry [@Wu:17] and nonlinear optomechanics [@Kaviani:15].
To date, one dimensional air-mode PCC’s have typically been fabricated from high refractive index silicon (Si, $n_\text{Si}~\sim~3.5$). There is significant interest however in creating air mode silicon nitride devices (Si$_{3}$N$_{4}$, $n_\text{Si$_{3}$N$_{4}$}~\sim~2.0$) [@Baets:16]. The large internal tensile stress of Si$_{3}$N$_{4}$ films on Si ($\sim$ 1 GPa) allows for realization of high mechanical quality factor $Q_m$ > $10^6$ nanobeam mechanical resonators [@Verbridge:06; @Verbridge:08; @ghadimi2018elastic]. It also enables creating optomechanical structures [@Camacho:09; @Krause:12; @Krause:15; @norte2016mechanical; @reinhardt2016ultralow; @tsaturyan2017ultracoherent; @ghadimi2018elastic] suspended over large length scales without warping or stiction compared with materials such as Si. This allows fabrication of long, thin supports suitable for suspending nanomechanical elements with enhanced sensitivity to external forces owing to a combination of low mechanical frequency $\omega_m$, effective mass $m_{eff}$, and spring-constant [@reinhardt2016ultralow]. For example, zipper cavities consisting of pairs of side-coupled dielectric-mode Si$_{3}$N$_{4}$ PCC’s [@Chan:09] have been fabricated with high aspect ratio tethers (115 $\mu$m x 130 nm) for ultrasensitive accelerometers [@Krause:12] and studying feedback cooling [@Krause:15]. As well, Si$_{3}$N$_{4}$ is compatible with standard silicon and CMOS manufacturing processes, has low optical absorption from telecom through most visible wavelengths with nominally no two-photon absorption at telecommunication wavelengths, and can be integrated with lithium niobate [@chang2017heterogeneous]. However, its relatively low refractive index is detrimental to high-$Q_o$ air-mode PCC design [@Barth:08; @McCutcheon:08] due to its reduced optical energy confinement and resulting increased coupling to vertically radiating modes.
Air-mode PCC’s locally confine the higher energy conduction band mode of a photonic crystal waveguide through tapering a decreasing fill factor towards the centre of the crystal, and vice-versa for the dielectric mode. Encapsulated air-mode Si$_{3}$N$_{4}$ PCC’s with elliptical holes designed for 740-1000 nm wavelengths have been used to show the deterministic positioning of nanoparticles [@panettieri2016control; @fryett2018encapsulated; @chen2018deterministic]. However, most demonstrations of Si$_{3}$N$_{4}$ PCC’s to date have utilized the dielectric mode, ensuring that the bandedge frequency of the photonic crystal waveguide from which the PCC is formed is far from the light-line $\omega~=~ck_{x}$, where $k_x$ is component of the wavevector $k$ parallel to the PCC’s waveguiding axis [@PCbook]. Lowering the bandedge frequency from which the PCC mode is formed, either by increasing the effective refractive index $n_{eff}$ or basing a PCC on a dielectric mode, reduces coupling between the optical mode and radiation modes above the light-line, and thus reduces loss. Dielectric mode Si$_{3}$N$_{4}$ PCC’s have been shown to have high $Q_o$ [@Khan:11], have been demonstrated in the resolved-sideband regime of cavity optomechanics [@Davanco:14], and have been used to study the emission properties of SiN [@Barth:08] and to modify spontaneous emission [@Barth:07]. Dielectric mode zipper cavities have also been used for cavity optomechanics [@Eichenfield:09], and side-coupled optical and mechanical beams can form slot-mode optomechanical crystals that have been used in multimode optomechanics [@Grutter:15]. These devices are often designed at target wavelengths for specific applications, for instance at 637 nm for interfacing with diamond color centres [@McCutcheon:08].
In contrast, free-standing air-mode Si$_{3}$N$_{4}$ PCC’s at telecommunication wavelengths have not been reported to the best of the authors’ knowledge. Here we show that air-mode Si$_{3}$N$_{4}$ nanobeam PCC’s based on tapered round holes can be designed with ultra-high $Q_o$. This is achieved by increasing the thickness of the devices, thus increasing $n_{eff}$ and lowering the air-mode band-edge frequency away from the light-line. Motivated by high $Q_o$ optomechanical SBC’s and PNC’s [@Hryciw:13; @Healey:15], we then design devices based on elliptical holes that mode-match to a physical gap introduced within the PCC. Due to the higher optical bandedge frequencies of these devices, the standard deterministic design approach fails to predict the highest $Q_o$ device. However, by substituting a highly elliptical hole for the centre hole and tapering to circular holes we can lower the band frequencies and achieve an order of magnitude higher $Q_o$ than predicted with the deterministic procedure. We then further extend this highly elliptical hole design to add mechanical degrees of freedom and create high-$Q_o$ SBC and PNC designs. These devices combine the excellent mechanical attributes of Si$_{3}$N$_{4}$ with the sensitive cavity-optomechanical transduction of earlier Si based designs [@Kaviani:15]. This is predicted to allow their room-temperature thermal motion to be detected in a nonlinear optomechanical measurement whose power spectral density (PSD) is observable above system noise levels. We also show that in cryogenic conditions, observations of thermally-driven nonlinear optomechanical phononic Fock state quantum jumps and shot noise should be detectable. Notwithstanding these developments in Si$_{3}$N$_{4}$, we note that the results presented here can be used for device design in any low-loss material with a similar index of refraction, such as diamond [@Mouradian:17].
Photonic crystal nanocavity design
==================================
![(a) Schematic of the Si$_{3}$N$_{4}$ photonic crystal dielectric patterning with the cavity section shown enlarged and dotted line denoting the cavity centre. Dimensions: width $w$, lattice constant $a$, semi-minor and -major axes ($R_x,R_y$). (b) The TE-polarization bandstructure for the centre hole unit cell of the device in (a) for varying nanobeam thicknesses. The air (dielectric) modes are denoted with solid (dotted) lines, and the bandgaps (bg’s) of the mirror holes for the respective coloured bands are shown to the right of the plot area. (c) The quality factor $Q_o$ of the PCC’s symmetric fundamental mode determined from 3D FDTD simulations for thickness $t$. (d) The 3D FDTD electric field $E_y$ mode profile for a 500 nm thick device.[]{data-label="figureOne"}](figure1v6.pdf){width="12cm"}
We begin by discussing the design of the Si$_{3}$N$_{4}$ air-mode PCC shown in Fig. \[figureOne\](a). The hole dimensions and tapering were calculated with the deterministic design recipe in Refs. [@Quan10; @Quan11], as discussed below. We set the nanobeam width $w$ = 760 nm and lattice constant $a$ = 670 nm such that the target PCC mode wavelength $\lambda_o$ $\sim$ 1550 nm ($\omega_o~\sim~200$ THz, $a/\lambda_o~\sim$ 0.432) can be achieved for the range of hole sizes considered here. Initially the nanobeam thickness was set to $t = 350~\text{nm}$, then varied to study its effect on the PCC modes. As part of this design recipe, the band structure and band-edge frequencies of unit cells of the waveguide for varying hole shape were calculated using a 3D frequency-domain eigensolver [@Johnson:01] with high spatial resolution. From these bandstructure calculations, we first find the central cavity hole dimensions that brings the unit-cell air-mode band-edge frequency close to the target $\omega_o$. We then find the outer mirror hole dimensions to maximize the mirror strength $\gamma$: $$\gamma=\left( \frac{(\omega_2-\omega_1)^2}{(\omega_2+\omega_1)^2}-\frac{(\omega_\textrm{o}-\omega_{mid})^2}{(\omega_{mid})^2} ) \right) ^{1/2},
% \gamma=[(\omega_2-\omega_1)^2/(\omega_2+\omega_1)^2-(\omega_\textrm{o}-\omega_{mid})^2/(\omega_{mid})^2]^{1/2},$$ experienced by the mode at $\omega_\textrm{o}$ that is within the mirror hole bandgap, defined by the dielectric and air band-edge frequencies $\omega_1$ and $\omega_2$, respectively, of the mirror hole unit cell [@Hryciw:13]. Here $\omega_{mid} = (\omega_2+\omega_1)/2$ is the mid bandgap frequency, and $\gamma$ is proportional to the inverse attenuation length.
![(a) The dielectric pattern and $E_y$ field profile of the elliptical hole air-mode PCC following the standard deterministic design procedure. (b) The band structures of a unit cell containing the three centre hole shapes and gap discussed: 150 nm round (purple), standard design ellipse (yellow), highly elliptical (red), 112 nm gap (blue). (c) Fill factors (FF) of ellipses with dimensions ($R_x,R_y$). The black line denotes elliptical hole dimensions with a fill factor similar to the round centre hole. Fill factors of the three centre hole shapes are noted, with the gap FF matching that of the highly elliptical hole. (d) Schematic of the PCC with a highly elliptical centre hole and hole axes quadratically tapering to round mirror holes, and the resultant $E_y$ field profile. (e) Mirror strength (MS) trajectories of the hole tapering of the devices in (a) (dotted) and (d) (dashed). The hole dimensions that have a mirror strength at the target frequency $\omega_o$ are shown with the solid white line.[]{data-label="figureTwo"}](figure2v6.pdf){width="12cm"}
Initially we restricted ourselves to circular holes. A cavity centre hole with radius 150 nm was found to have an air-mode band-edge frequency $\omega_2 \sim \omega_o$. To form an air-mode nanocavity, the surrounding hole radii were tapered quadratically over $N_c$ = 12 holes symmetrically on either side of the nanobeam to the first mirror holes with radius 215 nm. $N_m$ = 15 mirror holes were then added to both sides. The bandstructure for the centre hole unit cell plotted in Fig. \[figureOne\](b) shows the fundamental (TE-like, $y$-odd parity, $z$-even parity) air and dielectric mode frequencies in normalized units ($\omega a/2\pi c~=~a/\lambda$) as a function of $k_x$ for varying $t$. From the bandstructure, we see that the air-mode band-edge frequency for the $t=350~\text{nm}$ device is relatively close to the light-line. When we compute the fundamental optical mode of the full $t=350~\text{nm}$ PCC using a 3D finite-difference time-domain (FDTD) simulation [@ref:oskooi2010mff] we find a $Q_o \sim 10^4$. A similar $Q_o$ is observed if we keep the cavity hole radii constant and quadratically decrease the lattice constant from the centre of the cavity to the mirrors [@Grutter:15IEEE].
To increase $Q_o$ we then studied the effect of changing $t$. In nanobeam PCC’s, light is index-guided in the $z$ direction, and typically $t \le \lambda / 2n$ is chosen so that the underlying waveguide remains single-mode over the frequency range of interest [@Zhang:09]. Figure \[figureOne\](b) shows that increasing the thickness pushes the band structure lower in frequency, away from the light-line: as $t$ is increased to 610 nm, the air-mode band-edge falls from a normalized frequency of 0.441 to 0.409. This corresponds to increasing $n_{eff}$, which decreases $\omega_{o}$ and reduces coupling to radiation modes, and is reflected by $Q_o$ increasing towards $10^6$ as shown in Fig. \[figureOne\](c). A simulated optical mode electric field profile ($E_y$) for a $t$ = 500 nm device is shown in Fig. \[figureOne\](d), whose corresponding $Q_o$ is $\sim 10^5$. In fabricated Si$_{3}$N$_{4}$ PCC devices, $t$ generally ranges from 200-450 nm [@Khan:11; @Grutter:15IEEE]. In consideration of the single-mode condition and with a desire for a high-$Q_o$, we will use a $t \sim 500~\text{nm}$ for the remainder of this analysis.
To design a nanobeam PCC with elliptical holes that will form the basis of the optomechanical devices discussed below, we relax the constraint of requiring holes to be circular, leave the other nanobeam dimensions unchanged, and again apply the deterministic design procedure [@Hryciw:13]. There are a continuum of hole shapes with varying ellipticity that maintain the same $\omega_o$ as the round hole device, as discussed below. We begin by choosing the centre hole with maximum ellipticity as constrained by the waveguide width. The resulting device is shown in Fig. \[figureTwo\](a). The cavity centre hole with semi-minor and semi-major axis dimensions ($R_{x_\textrm{c}}$, $R_{y_\textrm{c}}$) = (119.4, 350) nm has $a/\lambda~=~0.432$, and Fig. \[figureTwo\](b) shows how the bandstructure of this unit cell compares with that of the 150 nm radius round centre hole unit cell. If we quadratically taper to a mirror hole ($R_{x_\textrm{m}}$, $R_{y_\textrm{m}}$) = (210, 285) nm with maximum of the mirror strength following: $$R_{x_j,y_j}= R_{x_\textrm{c},y_\textrm{c}}+ (j/N_\textrm{c})^2(R_{x_\textrm{m},y_\textrm{m}}-R_{x_\textrm{c},y_\textrm{c}})$$ for integer $j\in[-N_\textrm{c},N_\textrm{c}]$, we find a simulated 3D $Q_o$ = $4.4~\times~10^4$.
For the purpose of comparison, we also simulated a device with a highly elliptical centre hole with the same fill factor as the round centre hole of the device in Fig. \[figureOne\](a). In Fig. \[figureTwo\](c), the fill factors of ellipses of dimensions ($R_x,R_y$) are calculated; the black line shows hole dimensions with the same fill factor as the round centre hole. The bandstructure for the hole with dimensions ($R_{x_c},R_{y_c}$) = (64.3, 350) nm is shown in Fig. \[figureTwo\](b) (red line) to have much lower bandedge frequencies; substituting this highly elliptical hole for the round cavity centre hole and tapering quadratically to the 215 nm round mirror holes results in the device shown in Fig. \[figureTwo\](d) with a 3D simulated $Q_o$ = $5.3~\times~10^5$. This is more than an order of magnitude greater than the $Q_o$ of the PCC designed following the standard deterministic recipe based on bandedge matching. This can be explained by this design’s lower mode frequency $a/\lambda=0.40$ that is further from the air light-line and as a result has less vertical radiation loss. For comparison, the trajectories of the hole mirror strengths of both the band-edge matching design (dotted line) and the highly elliptical to round hole fill-factor design (dashed line) are shown in Fig. \[figureTwo\](e). Also note the solid white equi-frequency line at the target frequency in this Figure. We expect that numerical optimization in future work, e.g. using a genetic algorithm [@goh2007genetic], would allow $Q_o$ to be increased further.
Creating split-beam and paddle nanocavity optomechanical devices
================================================================
We now show how the highly elliptical hole device from Fig. \[figureTwo\](d) can be used as the basis for the split-beam cavity shown in Fig. \[figureThree\](a). A split-beam cavity is formed by making a clear cut through the centre of the nanobeam, creating nanocantilevers whose mechanical modes are similar to those used in many sensing and metrology applications. Placing the cut in the centre of PCC ensures that the nanocantilever motion interacts strongly with the optical cavity mode. In order to maintain high-$Q_o$, tapering to highly elliptical holes is beneficial, as they better match the rectangular shape of the gap [@Hryciw:13]. When designing SBC’s, it is critical to understand the optical mode structure to know what gap width $g$ to replace the centre hole with. With SBC’s in Si, it was previously found that the dielectric mode of the gap had to be matched to the air mode of the centre hole due to a crossing of the waveguide band edges [@Hryciw:13]. As shown in Fig. \[figureThree\](b), due to the lower refractive index of the Si$_{3}$N$_{4}$, the dielectric and air mode bandedge frequencies of the gap unit cell as a function of nanobeam width in Si$_{3}$N$_{4}$ do not cross, even with the same $w/t$ ratio as in Ref. [@Hryciw:13]. Knowing this, we then substitute a gap with $g$ = 110 nm that has the same fill factor as the cavity centre hole that it replaces (given from Fig. \[figureTwo\](c)). After local optimization of $g$, we find that for $g$ = 112 nm the SBC supports an air-mode with $Q_o$ = $2.2~\times~10^5$. The bandstructure of a unit cell containing this gap is shown in Fig. \[figureTwo\] (blue line) to overlap very closely with that of the highly elliptical hole.
![(a) Schematic of the split-beam cavity with gap width $g$ and corresponding $E_y$ field profile. (b) Gap unit cell band edge frequencies (green) and $E_y$ field profiles for the optical modes at the points marked (and in blue at the same $w/t$ ratio as in Ref. [@Hryciw:13]) showing that the modes for the gap for each device do not cross. (c,d) Design and field profile of the paddle nanocavity with the axial motion of the paddle indicated (supports not shown in (d)).[]{data-label="figureThree"}](figure3v2.pdf){width="11.5cm"}
![(a) Schematic of the paddle nanocavity with $L$ = 50 $\mu$m supports. (b) Mechanical frequency $\omega_m$ (red) and effective mass $m_{eff}$ (blue) and (c) the single-photon to two-phonon coupling rate $\Delta\omega_{o}$ (red) given from the zero-point fluctuations $x_{zpf}$ (blue).[]{data-label="figureFour"}](figure4v7.pdf){width="11cm"}
A similar process can be followed to create a Si$_3$N$_4$ paddle nanocavity such as the device shown in Figs. \[figureThree\](c,d). In this device a Si$_3$N$_4$ paddle is inserted between the split-beam cavity nanocantilevers, with the same gap $g = 112\text{nm}$ between each nanocantilever and the paddle. The paddle length is chosen following the procedure used to design Si paddle nanocavities in Ref. [@Healey:15]. When simulated with a paddle length of $\sim$538 nm and without any supports, as shown in Fig. \[figureThree\](d), a $Q_o$ = $1.1~\times~10^5$ is found. In a realistic device, the paddle is suspended from the unpatterned device layer by thin supports as shown in Fig. \[figureThree\](c). Adding 100 nm wide supports connected to the center of the paddle reduces $Q_o$ to $1.7~\times~10^4$.
The support and paddle dimensions determine the properties of the paddle’s mechanical resonances, which can interact optomechanically with the optical modes of the device. Here we are interested in the axial sliding mechanical resonance shown in Fig. \[figureThree\](c) that mimics membrane-in-the-middle devices [@Thompson2008; @sankey2010strong; @mason2018continuous; @rossi2018observing; @rossi2018measurement]. For the sliding resonance, the paddle nanocavity has a vanishing linear optomechanical coupling owing to the odd symmetry of the mechanical mode and the even symmetry of the optical mode intensity [@Kaviani:15]. However, it has a nonlinear quadratic optomechanical coupling that is of particular interest for quantum optomechanics experiments [@Thompson2008; @sankey2010strong; @paraiso2015position]. This coupling strength is determined by the change in local dielectric constant $\Delta\epsilon(\mathbf{r};x)$ as a function of position $\mathbf{r}$ of the paddle’s normal mechanical mode displacement amplitude $x$, overlapped and cross-coupling the electric fields $E_{\omega}$ of the nanocavity optical mode spectrum at frequencies $\omega$ [@Kaviani:15]. The quadratic optomechanical coupling coefficient: $$g^{(2)}=\frac{\omega}{2} \frac{|\langle E_{\omega}|\frac{\delta \epsilon}{\delta x}|E_{\omega}\rangle|^2}{|\langle E_{\omega}| \epsilon |E_{\omega}\rangle|^2} -\sum\limits_{\omega^{\prime}\neq\omega}\left(\frac{\omega^{3}}{\omega^{\prime 2}-\omega^{2}}\right) \frac{|\langle E_{\omega^{\prime}}|\frac{\delta \epsilon}{\delta x} |E_{\omega}\rangle|^2}{\langle E_{\omega^{\prime}}| \epsilon |E_{\omega^{\prime}}\rangle \langle E_{\omega}| \epsilon |E_{\omega}\rangle}
\label{eq:second_order}$$ describes the strength of the nonlinear photon-phonon interactions in the system [@Johnson02; @Rodriguez:11; @Eichenfield:09OE]. As discussed above, the first self-term describing the nonlinear contribution from the linear optomechanical coupling vanishes [@Kaviani:15]. However, in practice there is some amount of linear coupling [@Healey:15] generated by imperfections in the fabricated device. Linear coupling can also be introduced with an external waveguide used to couple light into and out of the cavity [@Hryciw:15]. Following Ref. [@Kaviani:15], a $g^{(2)}/2\pi\approx$ 10 MHz/nm$^{2}$ was calculated by inputting the mechanical sliding resonance profile and the optical field profiles of the fundamental ($E_\omega$) and higher order ($E_{\omega'}$) optical modes of the paddle nanocavity into Eq. \[eq:second\_order\]. These calculations where performed using finite element method simulations (COMSOL), and contributions from higher order modes over a spectral range of tens of terahertz were considered. Including this extended higher order mode spectrum was aided by automating the calculations that were performed manually in Ref. [@Kaviani:15].
![PSD of the mechanical resonance for (a) low $Q_o$ = $4.0~\times~10^3$, $Q_m~=~10^2$ and (b) high $Q_o$ = $4.0~\times~10^4$, $Q_m~=~10^5$ ((a,b) top plots are with $L \sim 50~\mu$m long supports). The detector noise (lower, red dashed line) and shot noise (upper, black dashed line) are also shown. (c) The PSD at $L$ = 50 $\mu$m and $\omega/\omega_m$ = 2. (d) PSD for high $Q_o$, low $Q_m$ as given in (a,b).[]{data-label="figureFive"}](figure5v5.pdf){width="11.5cm"}
Although $g^{(2)}$ is reduced in Si$_3$N$_4$ paddle nanocavities in comparison to their Si implementation, due in part to the lower refractive index of Si$_3$N$_4$, applications in quantum optomechanics can benefit from the exceptionally high aspect-ratio mechanical structures realizable in Si$_3$N$_4$. To illustrate this potential, we now investigate the effect of increasing the paddle’s sliding mode per-phonon displacement amplitude by increasing the support length $L$ to reduce the mode’s spring constant. Figure \[figureFour\](a) shows a schematic of the device with a 50 $\mu$m $\times$ 100 nm support cross section. The mechanical frequency $\omega_m$ and effective mass $m_{eff}$ of the sliding mode (computed with COMSOL) as a function of support length are shown in Fig. \[figureFour\](b). For $L$ ranging 5-50 $\mu$m, $\omega_m/2\pi$ varies from $\sim10^7-10^5$ Hz, and has a $L^{-1.9}$ dependence slightly shifted from $L^{-2}$ as expected from Euler-Bernoulli nanobeam theory [@Cleland:02] due to the mass distribution of the paddle. The corresponding $m_{eff}$ varies from $\sim1-7$ pg. Shown in Fig. \[figureFour\](c), the zero-point fluctuation amplitude $x_{zpf}~=~\sqrt{\hbar/2m_{eff}\omega_{m}}$ varies from 20-120 fm (where $\hbar$ is the reduced Planck’s constant), and the corresponding single-photon to two-phonon coupling rate $\Delta\omega_{o}\equiv|g^{(2)}x^{2}_{zpf}|$ varies from 2$\pi~\times$ (0.25-0.9) Hz. These plots illustrate that increasing $L$ results in a larger per-phonon nonlinear optomechanical coupling due to the increase in $x_{zpf}$.
Given these simulated parameters we can make predictions regarding the suitability of Si$_3$N$_4$ paddle nanocavities for nonlinear optomechanical detection of motion driven by thermal and quantum effects. If photons are coupled into and out of the cavity with an external waveguide (e.g. a fiber taper or end-fire coupled waveguide) and measured with a photodetector, the optical response of the paddle nanocavity transducing the nonlinear mechanical displacement can be predicted [@Kaviani:15]. For input power $P_i$, the photodetected optical power spectral density (PSD) $S_{P}^{(2)}(\omega) = \frac{1}{4}P_i^2 G_2^{2} S_{x^{2}}(\omega)$, where $G_2$ is the quadratic coefficient of transduction of the fluctuating waveguide output, and $S_{x^{2}}$ is the PSD of the $x^2$ mechanical motion, given in Ref. [@Kaviani:15] for the case of a thermally driven mechanical resonator. Here $G_2$ describes the cavity optical response measured by the photodetector, and includes the influence of the photodiode quantum efficiency, $Q_o$, $g^{(2)}$, and waveguide transmission losses. Both $G_2$ and $S_{x^{2}}$ assume there is no backaction and the device is operating in the sideband-unresolved regime [@ref:aspelmeyer2014co].
Figures \[figureFive\](a,b) show the thermally driven $S_{P}^{(2)}(\omega)$ of the sliding resonance at an operating temperature of 300 K, and indicate that it can be observed above technical noise. The signal strength varies with $L$ and the mechanical detuning $\omega/\omega_m$; traces in Fig.’s \[figureFive\](a,b) correspond to signal for $L$ = 50 $\mu$m in their respective spectrographs $S_{P}^{(2)}(\omega,L)$, which are also shown. These spectra assume reasonable experimental parameters: $P_i$ = 100 $\mu$W chosen so as to not saturate the detector, the input laser detuning from the nanocavity resonance set to maximize $G_2$, and either relatively low $Q_o$ = $4.0~\times~10^3$ and $Q_m$ = $10^2$ (Fig.’s \[figureFive\](a)) or attainably high $Q_o$ = $4.0~\times~10^4$ and $Q_m$ = $10^5$ (Fig.’s \[figureFive\](b)). The power spectrum is strongest at $\omega = 0$ and $\omega/\omega_m$ = $\pm$ 2, with the latter being of interest in practical experiments so as to minimize the influence of $1/\omega$ technical noise not considered here. The signal generally increases with $Q_o$ and $Q_m$, and the FWHM (full-width, half-maximum) of the peaks narrows with increasing $Q_m$. As shown in Fig. \[figureFive\](b), for the high $Q_o$ and $Q_m$ values, the nonlinear signal can exceed the electronic noise (red dashed line) of a Newport 1811 photoreceiver (NEP = 2.5 pW/Hz) and the optical shot noise (black dashed line) by over 30 dBm/$\sqrt{\text{Hz}}$.
The importance of increasing $L$ can be seen from the spectrographs in Figs. \[figureFive\](a,b), whose contours indicate that the frequency range where the signal exceeds the noise sources becomes larger as $L$ increases. This is a result of the increase in thermal occupancy (due to decreasing $\omega_m$) and increase in $g^{(2)}$ of the device with increasing $L$. The impact of reducing mechanical and optical dissipation is illustrated by Fig. \[figureFive\](c), which shows how the signal strength at $L$ = 50 $\mu$m and $\omega/\omega_m$ = 2 depends on $Q_o$ ranging over $10^3-10^5$ and $Q_m$ over $10^2-10^6$. At the maximum of these values, the signal strength reaches $\sim-35$ dBm/$\sqrt{\text{Hz}}$, approximately 50 dBm/$\sqrt{\text{Hz}}$ over the predicted system noise levels. Figure \[figureFive\](d) shows that increasing $Q_o$ while maintaining low $Q_m$ raises the signal level but maintains a broad bandwidth, making the signal accessible at all mechanical frequencies between $\omega/\omega_m$ = $\pm$3 for support lengths > 40 $\mu$m.
Finally, we discuss how the Si$_{3}$N$_{4}$ paddle nanocavity is expected to perform for phononic quantum non-demolition (QND) measurements [@Kaviani:15]. Although the $g^{(2)}$ for this device is almost two orders of magnitude lower than that of similar Si devices, the high internal stress of Si$_{3}$N$_{4}$ enables longer supports that in practice lower $\omega_m$. Stress engineering can be used to increase $Q_m$ and has resulted in record $Q_m \omega_m$ products in recently reported work [@ghadimi2018elastic; @tsaturyan2017ultracoherent]. This figure of merit is a measure of the ability of a mechanical resonator to maintain coherence in the presence of a thermal bath. Considering measurement shot noise and the expected change in cavity transmission due to a change in phonon number, as in Ref. [@Thompson2008], it is predicted that the signal-to-noise ratio $\Sigma^{(0)}$ = $\tau_{tot}^{(0)}\Delta\omega_{o}^{2}/S_{w_{o}}$ for a quantum jump measurement from the motional ground state to a single phonon Fock state could exceed unity for sufficiently high $Q_m$ and $L$, as shown in Fig. \[figureSix\](a). The $\Sigma^{(0)}$ for $Q_m$ = $10^2$, $10^5$ and $10^8$ (blue, orange and green curves respectively) is presented, assuming the device and its surroundings are cryogenically pre-cooled to a bath temperature $T_{b}~=~10$ mK, and that $Q_o$ = $10^6$. Here $\tau_{tot}^{(0)}~=~\hbar Q_m~/~k_{B}T_{b}$ (for Boltzmann’s constant $k_B$) is the thermal lifetime quantifying the rate of decoherence of the ground-state-cooled nanomechanical resonator due to bath phonons, and $S_{w_{o}}~=~\hbar\omega_{o}\kappa^{2}/16P_{i}$ is the shot-noise-limited sensitivity of an ideal Pound-Drever-Hall detector for cavity loss rate $\kappa$ (cooling the device to the ground state will be further discussed below). With $Q_m$ = $10^8$ and $L$ = 500 $\mu$m, $\Sigma^{(0)}$ > 1, however this can also be achieved with lower $Q_m$ and longer $L$. We note that calculating the signal-to-noise ratio as in Ref. [@paraiso2015position] gives similar results.
For measurements of phonon shot noise (the granularity of large phonon number coherent states) [@ClerkPRL2010] with signal-to-noise ratio $S=8\bar{n}_{d}n_{th}\Sigma^{(0)}$ > 1, the required drive amplitude $x_d$ and its correspondence in units of phonon number $\bar{n}_d$, assuming a thermal phonon number $n_{th}=1/4$, is shown in Fig. \[figureSix\](b). At $L$ = 500 $\mu$m, the number of required coherent drive phonons is < 400, a relatively small value that is more than three orders of magnitude less than what is required to drive the Si device previously discussed [@Kaviani:15]. This leaves significant available headroom in drive strength to compensate for non-optimal devices parameters and reduced $\Sigma^{(0)}$.
However, the above figures of merit do not consider the effect of parasitic linear optomechanical coupling between the optical cavity modes that is discussed in detail in Refs. [@miao2009standard; @ludwig2012enhanced; @paraiso2015position]. This coupling, which originates from the non-zero cross-coupling terms in Eq. , is a form of decoherence that imposes a strong coupling requirement for QND measurements [@miao2009standard]. Applying the theory from Ref. [@miao2009standard] to our system and assuming that the nonlinear coupling is dominated by a single cross-coupling term, this condition roughly requires $\kappa_1\kappa_2/|\Delta\omega_o (\omega'~-~\omega)|~<~1$, where $(\omega'~-~\omega)$ is the detuning between the cavity modes whose cross-coupling is the dominant contributor to $g^{(2)}$ in Eq. , and $\kappa_i = \omega_i/Q_i$ are the optical loss rates of the modes. For our system, $|\omega' - \omega|/2\pi \sim 10$ THz, $\kappa_{1,2}/2\pi \sim 1$ GHz ($Q_o~\sim~10^5)$, and $g^{(2)}/2\pi \sim 1$ Hz, indicating that the device is four to five orders of magnitude away from satisfying this condition, and that increasing $Q_o$ as well as $g^{(2)}$ and/or $|\omega - \omega'|$ is required before QND measurements will be possible. Increasing $Q_o$ to $10^7$, as achieved in silicon photonic crystals [@sekoguchi2014photonic], is one such path towards this goal. In parallel, further increasing $g^{(2)}$ while maintaining a large $|\omega~-~\omega'|$ would contribute to reaching this regime.
![(a) Predicted phononic Fock state measurement signal-to-noise ratio $\Sigma^{(0)}$ for $Q_m$ = $10^8$ (top, green), $Q_m$ = $10^5$ (middle, orange), and $Q_m$ = $10^2$ (bottom, blue) at ultrahigh $Q_o$ = $10^6$. (b) The number of phonons $n_{th}$ in the device at $T_b$ = 10 mK (blue) and the drive amplitude $x_d$ (red, and in units of phonon number $n_d$) to achieve a signal-to-noise ratio $S$ > 1 when measuring phonon shot noise.[]{data-label="figureSix"}](figure6v2.pdf){width="12.5cm"}
Although the longer supports make the quantized phononic energy fluctuations of the oscillator observable, there will be increased thermal decoherence associated with the lower $\omega_m$. There is also a trade-off in the increase in thermal phonon occupation number at achievable cryogenic operation temperature, requiring development of the ability to further cool the mechanical resonance from $T_b$ to an effective temperature corresponding to the quantum ground state. The number of phonons $n_{th}$ remaining in the device at $T_{b}$ and $\omega_m$ given by $L$ (according to Bose-Einstein statistics) is shown in Fig. \[figureSix\](b) (blue). Removing these phonons in order to demonstrate the quantum effects described above would require sideband-unresolved nonlinear optomechanical cooling [@Clark2017] to extract mechanical energy by scattering incident photons to higher energies (anti-Stokes scattering), as recently achieved in a low-frequency optomechanical system [@Clark2017]. By interacting squeezed light with the optomechanical cavity, one can coherently null the Stokes processes which prevent cooling, and the mechanical system can in principle be cooled arbitrarily close to the motional quantum-mechanical ground state [@Clark2017]. Techniques to increase the $Q_m$, such as the utilization of phononic crystals [@Chan2011] and nanostring strain engineering [@ghadimi2018elastic; @ghadimi2017radiation; @fedorov2018evidence], could also help reach this goal.
Conclusion
==========
Designing a silicon nitride air-mode PCC is a delicate task due to the sensitivity of the device to optical loss through radiating modes. High optical quality factor devices can be realized by reducing the air-mode bandedge frequencies through increasing the device thickness. These devices can form the basis for optomechanical devices that can be used for both fundamental and practical research and applications.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors are grateful to Roohollah Ghobadi for helpful discussions. This work is supported by Alberta Innovates (AI), the Natural Science and Engineering Research Council of Canada (NSERC), and the Canada Foundation for Innovation (CFI).
Disclosures {#disclosures .unnumbered}
===========
The authors declare that there are no conflicts of interest related to this article.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper deals with grouping of entities in a fleet based on their behavior. The behavior of each entity is characterized by its historical dataset, which comprises a dependent variable, typically a performance measure, and multiple independent variables, typically operating conditions. A regression model built using this dataset is used as a proxy for the behavior of an entity. The validation error of the model of one unit with respect to the dataset of another unit is used as a measure of the difference in behavior between two units. Grouping entities based on their behavior is posed as a graph clustering problem with nodes representing regression models and edge weights given by the validation errors. Specifically, we find communities in this graph, having dense edge connections within and sparse connections outside. A way to assess the goodness of grouping and finding the optimum number of divisions is proposed. The algorithm and measures proposed are illustrated with application to synthetic data.'
author:
-
-
-
bibliography:
- 'grouping.bib'
title: Grouping Entities in a Fleet by Community Detection in Network of Regression Models
---
Introduction
============
Monitoring of infrastructure such as aircrafts, turbines and vehicles is becoming increasingly important. The sensors on these entities generate a wealth of data, which is analysed to predict failures in them. Typically the analysis of historical data has been on an entity-by-entity basis [@nielsen06]. However, given a fleet of units, it is possible to pose questions about the collective behavior of the fleet and answer them using the tools of network analysis. One such question pertains to finding units in the fleet which behave similarly.
Specifically, our problem is to cluster sets of measurements $Y$ measured as a function of dependent variable(s) $x$. Data, in the form of discrete observations, is given for $N$ different entities and for each entity, $Y$ is assumed to be a smooth function of $x$. The length of dataset need not be same for all entities. In cases where $x$ is time measurement, the data is referred to as *longitudinal* data or *repeated measures* data. In a general context, where $x$ need not be time, the term *functional* or *trajectory* data is used. Given trajectory data, we are interested in determining if the data can be naturally clustered into groups.
Major functional data clustering approaches can be classified in one of three groups: two-stage methods, non-parametric clustering and model-based clustering [@jacques13]. These algorithms aim to cluster *curves*, where it is implied that the dimensionality of data is very large but no relation between $Y$ and $x$ is utilized. Hence, these algorithms do not characterize the entities by their behavior or by the way in which $Y$ depends on $x$.
Catez *et al.*[@cadez00] have clustered trajectories by explicitly using the depedency of **Y** on $x$ using the expectation-maximization (EM) algorithm to find the grouping. It is based on probabilistic modeling of a set of trajectories generated from a finite mixture model consisting of regression model components. The algorithm proceeds by making an assumption about the functional form of the component models. Our method, on the other hand, does not involve such an assumption, which may be difficult make a priori in several cases.
Given the historical data of a unit, our method first computes a regression model for each unit. This model stands as a proxy for the behavior of the unit. The fleet is represented by an ensemble of such models, one for each unit. We propose a suitable dissimilarity measure between two units and form a graph of the fleet. Each vertex of the graph represents the model of a unit and the edge weight between two vertices is given by the value of dissimilarity measure between two models. The trajectory clustering problem is thus translated to a graph clustering problem. We also propose a measure to characterize the goodness of grouping of similarly-behaving units and hence find the best division.
Section II gives a formal description of the problem and precise meaning of terms used above - behavior, adequate modeling accuracy and similarity of behavior. It also gives the method used to find optimal grouping of the units. Section III demonstrates the application of our approach for performance modeling of a fleet of steam turbines. Section IV discusses potential alternative approaches and directions for future work.
Method
======
Model Estimation of Entities
----------------------------
Given the hisorical dataset of the fleet in the form
$$\left\{\{X^{i}(t), Y^{i}(t)\}_{t=1}^{T_{i}}\right\}_{i=1}^{N}$$
where $i$ numbers the unit in the fleet, and the integer $t$ denotes the occasion (time) when the data from unit $i$ is recorded. The dataset of each unit has length $T_{i}$ occasions (time samples). The independent variables $X(t)$ $\in$ $\textbf{R}^{p}$ are the input measurements - measurements of the operating conditions of unit $i$ at time $t$. The dependent variable $Y(t)$ $\in$ $\textbf{R}$ is the output measurement - typically, a performance metric of unit $i$ at time $t$ [@chu11].
The historical data of unit $i$ is denoted as $D_{i}$. The *behavior* of unit $i$ implies the relationship between $Y$ and $X = (X_{1}, X_{2}, ......, X_{p})$ which can be written in the general form
$$Y = f(X) + \epsilon$$
Here $f$ is some fixed but unknown function of $X_{1}, ....., X_{p}$ and $\epsilon$ is a random error term which is independent of $X$ and has zero mean. Since $f$ is unknown, we compute an estimate $\hat{f}$ of $f$ from historical data. With $\hat{Y} = \hat{f}(X)$, we have
$$\begin{aligned}
E(Y - \hat{Y})^{2} &= E[f(X) + \epsilon - \hat{f}(X)]^{2} \notag \\
&= \underbrace{E[f(X) - \hat{f}(X)]^{2}}_{\text{Reducible}} + \underbrace{Var(\epsilon)}_\text{Irreducible}\end{aligned}$$
where $E(Y - \hat{Y})^2$ represents the average, or *expected value*, of the squared difference between the predicted and actual value of $Y$, and Var($\epsilon$) represents the variance associated with the error term $\epsilon$ [@james13].
Given $D_{i}$, $f_{i}$ can be estimated by either parametric or non-parametric statistical methods. If a parametric method is used, the functional form of $f_{i}$ may be assumed linear or non-linear, depending on the application and the dataset $D_{i}$. For a fleet of $N$ units, we have an ensemble of $N$ models - $f_{1}$, $f_{2}$...$f_{N}$. To assess the accuracy of the models, residual standard error (RSE) can be used and is given by
$$RSE = \sqrt{\frac{1}{T_{i}}\sum_{t=1}^{T_{i}} (Y(t) - \hat{Y}(t))^{2}}$$
Constructing Fleet Network of Regression Models
-----------------------------------------------
The RSE is a measure of the lack of fit of the model to the data. The same measure can be used to represent the dissimilarity between two models. Let $r_{ij}$ stand for the RSE when model $f_{i}$ is tested on the dataset $D_{j}$. Given the estimated models $f_{1}$, $f_{2}$...$f_{N}$ and the datasets $D_{1}$, $D_{2}$...$D_{N}$, a $N \times N$ matrix $\textbf{R}$ of RSE errors can be constructed
$$\textbf{R} = \begin{pmatrix}
r_{11} & r_{12} & \cdots & r_{1N} \\
r_{21} & r_{22} & \cdots & r_{2N} \\
\vdots & \vdots & \ddots & \vdots \\
r_{N1} & r_{N2} & \cdots & r_{NN}
\end{pmatrix}$$
**R** represents a weighted multigraph. For simplicity of analysis, we would like to consider weighted, simple graphs. This can be accomplished by
$$\begin{aligned}
\textbf{R}' &= \frac{\textbf{R} + \textbf{R}^T}{2} \\ \notag \\
&= \begin{pmatrix}
r'_{11} & r'_{12} & \cdots & r'_{1N} \\ \notag
r'_{21} & r'_{22} & \cdots & r'_{2N} \\ \notag
\vdots & \vdots & \ddots & \vdots \\ \notag
r'_{N1} & r'_{N2} & \cdots & r'_{NN} \notag
\end{pmatrix}\end{aligned}$$
$\textbf{R}'$ is a symmetric matrix, that is, $r'_{ij}$ = $r'_{ji}$. Also, $r'_{ii}$ represents the training error of $f_{i}$. $\textbf{R}'$ can be considered as a dissimilarity matrix, where $r'_{ij}$ and $r'_{ji}$ can be considered as a measure of difference between units $i$ and $j$.\
Consider a transformation of $\textbf{R}'$ to $\textbf{A}(\lambda)$ given by
$$\textbf{A}(\lambda) = \begin{pmatrix}
0 & H(\lambda - r'_{12}) & \cdots & H(\lambda - r'_{1N}) \\
H(\lambda - r'_{21}) & 0 & \cdots & H(\lambda - r'_{2N})\\
\vdots & \vdots & \ddots & \vdots \\
H(\lambda - r'_{N1}) & H(\lambda - r'_{N2}) & \cdots & H(\lambda - r'_{NN})
\end{pmatrix}$$
where $\lambda$ is a given scalar value and H(x) is the unit step function given by
$$H(x) =
\begin{cases}
0 & \text{if } n < 0, \notag \\
1 & \text{if } n \geq 0
\end{cases}$$
$\textbf{A}(\lambda)$ is a binary, symmetric matrix. It can be considered as an adjacency matrix of a graph. This graph has an edge between a pair of vertices $V_{i}$ and $V_{j}$ (representing units $i$ and $j$ respectively) only if $r'_{ij}$, or equivalently $r'_{ji}$, is less than the given threshold $\lambda$ with self-loops being removed.
Grouping of Entities in Network
-------------------------------
The notion of groups in a graph may be represented by three concepts: (a) clique (b) community (c) connected component.
The kind of networks we are interested in possess community structure. This implies that the network divides naturally into groups of nodes with dense connections internally and sparser connections between groups. This same structural feature is an important property of social networks, such as those of dolphins illustrated in Fig. 1 [@newman03].
Detecting cliques is analogous to complete-linkage clustering and finding connected components is akin to single-linkage clustering. For the networks being studied here, cliques are too strong a criterion and connected components too weak. In either case, we do not arrive at a meaningful grouping or obtain non-trivial knowledge. Community detection, being representative of average-linkage clustering, provides just the right balance between the two extremes. This fact is highlighted in Fig. 1.\
### Detection of Communities
Community structure detection algorithms try to find dense subgraphs in directed or undirected graphs, by optimizing a criterion particular to the algorithm. Several community detection algorithms have been developed, which are based on centrality measures, flow models, random walks, resistor networks, optimization and other approaches [@danon05].
For our application, we have used the leading eigenvector method which has proven very effective [@newman06]. The heart of the method is the definition of a modularity matrix **B**
$$\textbf{B} = \textbf{A}(\lambda) - \textbf{P}$$
where $\textbf{A}(\lambda)$ is the adjacency matrix obtained in (7), and **P** contains the probability that certain edges are present according to the ‘configuration model’. In other words, a $\textbf{P}[i,j]$ element of **P** is the probability that there is an edge between vertices $i$ and $j$ in a random network in which the degrees of all vertices are the same as $\textbf{A}(\lambda)$.
The leading eigenvector method works by calculating the eigenvector of the modularity matrix for the largest positive eigenvalue and then separating vertices into two community based on the sign of the corresponding element in the eigenvector. If all elements in the eigenvector are of the same sign that means that the network has no underlying comuunity structure. Further details can be found in [@newman06].\
### Finding Optimal Grouping
Given a grouping by the preceding method, a measure to assess the accuracy of the grouping is required to ascertain the goodness of grouping. Such a measure, called average meta-validation accuracy, is described now.
For a given $\lambda$, let the community detection algorithm result in $k$ groups. We shall use the prefix *meta* in reference to this grouping, thereby implying a higher level of abstraction and analysis. The datasets of all the units belonging to the $i$-th group are merged in a single meta-dataset $G_{i}$. Hence, we have $k$ meta-datasets $G_{1}, G_{2}, ...., G_{k}$. Each $G_{i}$ is divided into a training set $G_{Ti}$ and a validation set $G_{Vi}$ in an suitable proportion.
Each $G_{i}$ can be considered to be generated by an underlying function $g(X)$, analogous to $f(X)$ in (2). Given the training set $G_{Ti}$ a meta-model $g_{i}$ can be estimated. The form of $g(X)$ can be parametric or non-parametric and can differ from that of $f(X)$. However, for the sake of comparison, the form of $g_{i}$ at the meta-level must be consistent for varying values of $\lambda$. Thus, we have an ensemble of meta-models $g_{1}, g_{2}, ...., g_{k}$.
For each model $g_{i}$, the validation RMSE $e_{i}$ is calculated using (4). Let $|G_{i}|$ represent the number of units in the group $G_{i}$. The average meta-validation accuracy, $\eta(\lambda)$ is defined as
$$\eta(\lambda) = 100 - \frac{\sum\limits_{i = 1}^{k}|G_{i}| \times e_{i}}{N}$$
A plot of $\eta(\lambda)$ versus the number of communities detected (henceforth referred to as ’accuracy plot’) tends to be a monotonically increasing curve. The point where the slope of the curve levels off (the ’elbow point’) indicates the suitable number of communities present in the fleet.
Results
=======
We analysed the historical data for a fleet of 65 GE steam turbines to model the power output of the units as a function of the operating conditions in the steady-state. The data consisted of sensor readings of temperature and pressure at various stages of the turbine and the power output. The data was sampled at one hour interval over a period of 5-10 years. Due to its proprietary nature, more description of the turbine data and corresponding results cannot be given here.
In lieu of analysis of operational fleet data, the results of the algorithm on synthetic data are illustrated in Fig. \[fig:no\_groups\], \[fig:fuzzy\_groups\], \[fig:clear\_groups\]. A population of 30 curves is analyzed. The independent variable of each curve (x) has been randomly sampled in the interval 0-100. The dependent variable of each curve (y) is generated using a linear model with additive gaussian noise having zero mean and variance of one. To state one application, the dependent variable may be the response of individuals to varying stimulus (independent variable) in a psychology experiment.
The three cases considered are meant to verify whether we can use the algorithm to distinguish the cases where meaningful grouping exists from those where it does not and to find the groups in the latter case. Curves belonging to the same group have been assigned the same color.
Figure \[fig:no\_groups\] represents the data generated by sampling the slope of each generative linear model from a uniform distribution in an appropriate interval. As a result, the curves are uniformly spread apart. The corresponding accuracy plot does not level-off as the number of communities increases. Hence, all the curves have been assigned different colors, implying that curves have not been grouped at all.
Figure \[fig:fuzzy\_groups\] was generated from data where the slopes of the linear model were sampled from a mixture of 6 gaussian distributions with large variance. The plot suggests that the curves are spread apart, yet they fuzzily group together in pockets. The accuracy plot does not provide a single number for the clusters in the data. The color coding of the curves is shown for the case when the number of communities is 16. Curves in vicinity do share the same color, yet nothing conclusive can be said about the number of communities.
Figure \[fig:clear\_groups\] considers the case where the curves group themselves distinctly. The slopes were sampled from a mixture of 5 gaussian distributions with small variance. The accuracy plot has a distinct elbow for the number of community as 5. The corresponding grouping of the curves matches with the data.
{width="65.00000%"}
{width="65.00000%"}
{width="65.00000%"}
{width="65.00000%"}
{width="65.00000%"}
{width="65.00000%"}
Discussion
==========
Grouping of entities in the aforementioned manner provokes new questions: why do units group themselves? What similarities do entities belonging to the same group exhibit? If meta-data about the entities is present, it may be possible to perform root-cause anlaysis and find the factors which cause such grouping. The detection of such factors may reveal previously unknown insights. These questions have been bypassed here by use of simulated data.
A suitable measure needs to be devised which can reflect the confidence of grouping by the algorithm. For example, our grouping has low confidence in Fig. 3, but high confidence in Fig. 4. Though such a conclusion can be drawn from the accuracy plot, a numerical value to suggest this will be convenient. Besides, the order/form of the generative curve models is typically unknown a priori. The model estimation procedure may result in overfitting or underfitting. The effect of model selection on the accuracy of the groups needs to be investigated.
An alternative way of grouping regression models may be to cluster the coefficients of the models of individual units. For small datasets, the unreliability of coefficient estimates due to multicollinearity may be a challenge. Even if multicollinearity is absent, we need to factor the variable importance of coefficients, possibly by scaling the respective axes by variance values.
Grouping regression models may hold promise for other applications, such as reducing the number of trees in a random forest without compromising the radomness of the forest. Another application may be to analyse unbalanced datasets by sampling multiple datasets and analysing the relations among the resulting models. These applications will be explored in a later work.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank Subhankar Ghosh and Subodh Kolwankar for their valuable suggestions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quantum state tomography is the task of inferring the state of a quantum system by appropriate measurements. Since the frequency distributions of the outcomes of any finite number of measurements will generally deviate from their asymptotic limits, the estimates computed by standard methods do not in general coincide with the true state, and therefore have no operational significance unless their accuracy is defined in terms of error bounds. Here we show that quantum state tomography, together with an appropriate data analysis procedure, yields reliable and tight error bounds, specified in terms of confidence regions|a concept originating from classical statistics. Confidence regions are subsets of the state space in which the true state lies with high probability, independently of any prior assumption on the distribution of the possible states. Our method for computing confidence regions can be applied to arbitrary measurements including fully coherent ones; it is practical and particularly well suited for tomography on systems consisting of a small number of qubits, which are currently in the focus of interest in experimental quantum information science.'
author:
- Matthias
- Renato
date: 'November 7, 2012'
title: Reliable Quantum State Tomography
---
The state of a classical system can in principle be determined to arbitrary precision by applying a single measurement to it. Any imprecisions are due solely to inaccuracies of the measurement technique, but not of fundamental nature. This is different in quantum theory. It follows from Heisenberg’s uncertainty principle that measurements generally have a random component and that individual measurement outcomes only give limited information about the state of the system|even if an ideal measurement device is used. To illustrate this difference, it is useful to take an information-theoretic perspective. Assume, for instance, that we are presented with a two-level system about which we have no prior information except that it has been prepared in a pure state, and our task is to determine this state. If the system was classical, there are only two possible pure states, and one single bit of information is therefore sufficient for its full description. Furthermore, a single measurement of the system suffices to retrieve this bit. If the system was quantum, however, the situation becomes more interesting. A two-level quantum system (a qubit) admits a continuum of pure states that can, for example, be parameterized by a point on the Bloch sphere. To determine this point to a given accuracy $\Delta$, at least $\log_2 (4/\Delta^2)$ bits of information are necessary [^1]. Conversely, according to Holevo’s bound [@Holevo73], any measurement applied to a single qubit will provide us with at most one bit of information. And even if $n$ identically prepared copies of the qubit were measured, at most $\log_2 (n+1)$ bits of information about their state can be obtained [^2]. Hence the accuracy, $\Delta$, to which the state can be determined always remains finite ($\Delta \geq
\frac{2}{\sqrt{n+1}}$), necessitating the specification of error bars.
The impact that randomness in measurement data has on the accuracy of estimates has been studied extensively in statistics and, in particular, estimation theory [@Kay93]. The latter is concerned with the general problem of estimating the values of parameters from data that depend probabilistically on them. The data may be obtained from measurements on a quantum system with parameter-dependent state, as considered in quantum estimation theory [@Helstrom76]. Quantum state tomography can be seen as a special instance of quantum estimation, where one aims to estimate a set of parameters large enough to determine the system’s state completely [@Fano57; @VogRis89; @Raymeretal93; @LePaAr95; @Hradil97; @Arianoetal03; @Sugiyama11].
An obvious choice of parameters are the matrix elements of a density operator representation of the state. Due to the finite accuracy, however, the individual estimates for the matrix elements do not generally correspond to a valid density operator (for instance, the matrix may have negative eigenvalues). This problem is avoided with other techniques, such as maximum likelihood estimation (MLE) [@Hradil97; @Banaszeketal99; @Hradiletal04], which has been widely used in experiments [@Kwiatetal01; @Blattetal04; @Zeilingeretal05; @BlattWineland08; @Wallraffetal09; @Home09; @ZollerBlattetal11], or Bayesian estimation [@Helstrom76; @Jones91; @Buzeketal98; @Schacketal01; @TanKom05; @Blume10; @kalman].
In MLE, an estimate for the error bars can be obtained from the width of the likelihood function, which is approximated by the Fisher information matrix [@Banaszeketal99; @usami; @hradilmogil; @burgh; @hradil-diag; @Sugiyama11]. In current experiments one also uses numerical plausibility tests known as “bootstrapping” or, more generally, “resampling” [@Efron93; @Home09] in order to obtain bounds on the errors. However, despite being reasonable in many practical situations, these bounds are not known to have a well-defined operational interpretation and, in the case of the resampling method, may lead to an underestimate of the errors [@Blume12].
In contrast, Bayesian methods can be used to calculate “credibility regions”, i.e., subsets of the state space in which the state is found with high probability. This probability, however, depends on the choice of a “prior”, corresponding to an assumption about the distribution of the states before the measurements (in particular, the assumption can not be justified by the experimental data). Furthermore, we remark that most known techniques are based on the assumption of independent and identical measurements (a notable exception is the one-qubit adaptive tomography analysis of [@Sugiyama12]). We refer to [@Blume10] for a further discussion of currently used approaches to quantum state tomography, including pedagogical examples illustrating their limitations.
In this Letter we introduce a method to obtain *confidence regions*, that is, regions in state space which contain the true state with high probability. A point in the region may then serve as estimate and the maximal distance of the point to the border of the region as error bar. Our method allows to analyse data obtained from arbitrary quantum measurements, including fully coherent ones. The method does not rely on any assumptions about the prior distribution of the states to be measured. This makes it highly robust so that it can, for instance, be applied in the context of quantum cryptography, where the states to be estimated are chosen adversarially.
The remainder of this Letter is organized as follows. We first describe a very general setup for tomography of quantum states prepared in a sequence of experiments, where we do not make the typical assumption that the states are independent and identically distributed (i.i.d.). We then show that, nevertheless, properties of the states can be inferred reliably using a suitable tomographic data analysis procedure (Theorem \[thm:main\]). In motivation and spirit, this result relates to recent research efforts on quantum de Finetti representations. We then specialise our setup to the case where, in principle, the experiments may be run an arbitrary number of times (while still only finitely many runs are used to generate data). This special case is (by the quantum de Finetti theorem) equivalent to an i.i.d. preparation of the states, thereby justifying the common i.i.d. assumption in data analysis. The theorem, applied to this special case, then results in a construction for confidence regions for quantum state tomography (Corollary \[cor:main\]).
*General Scenario.|*Consider a collection ${\mathcal{S}}_1, \ldots,
{\mathcal{S}}_{n+k}$ of finite-dimensional quantum systems with associated Hilbert space ${\mathcal{H}}$, as depicted in Fig. \[fig\_setup\] (see also [@Renner07] and [@chiri-defin], where a similar setup is considered). We denote by $d$ the dimension of $\mathcal{H}$. For example, one may think of $n+k$ particles prepared in a series of experiments, where ${\mathcal{H}}$ could correspond to the spin degree of freedom. From this collection, a *sample* consisting of $n$ systems is selected at random and measured according to an (arbitrary) Positive Operator Valued Measure (POVM) $\{B^n\}$, a family of positive semi-definite operators $B^n$ on ${\mathcal{H}}^{\otimes n}$ such that $\sum_{B^n} B^n = {\mathbf{1}}_{{\mathcal{H}}}^{\otimes n}$. That is, each POVM element $B^n$ corresponds to a possible sequence of outcomes resulting from (not necessarily independent) measurements on the $n$ systems. The goal of quantum state tomography is to infer the state of the remaining $k$ systems, using the outcomes of these measurements.
Note that the $k$ extra systems are not measured during data acquisition. Nevertheless, they play a role in the above scenario, as they are used to define operationally what state we are inferring. (In the special case of i.i.d. states, the extra systems are simply copies of the measured systems|see below.) We also remark that, instead of measuring a sample of $n$ systems chosen at random, one may equivalently permute the initial collection of $n+k$ systems at random and then measure the first $n$ of them, i.e., ${\mathcal{S}}_1,
\ldots, {\mathcal{S}}_{n}$. We will use this alternative description for our theoretical analysis.
In order to describe our main results, we imagine that the measurement outcomes $B^n$ are processed by a data analysis routine that outputs a probability distribution $\mu_{B^n}$ on the set of mixed states, defined by $$\begin{aligned}
\mu_{B^n}(\sigma)d\sigma = \frac{1}{c_{B^n}} {\mathrm{tr}}[\sigma^{\otimes n}
B^n] d\sigma \end{aligned}$$ (see Fig. \[fig\_estimateillustration\] for an illustration). Here $d\sigma$ denotes the Hilbert-Schmidt measure with $\int
d\sigma=1$. Furthermore, $c_{B^n} ={\mathrm{tr}}[B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes
n} \cdot {\mathbf{1}}_{{\mathrm{Sym}}^n({\mathcal{H}}\otimes {\mathcal{K}})}] / \binom{n+d^2-1}{d^2-1}$ is a normalisation constant, where ${\mathcal{K}}\cong {\mathcal{H}}\cong {\mathbb{C}}^d$ and where ${\mathbf{1}}_{{\mathrm{Sym}}^n({\mathcal{H}}\otimes {\mathcal{K}})}$ is the projector onto the symmetric subspace of $({\mathcal{H}}\otimes {\mathcal{K}})^{\otimes n}$. Note that, in Bayesian statistics, $\mu_{B^n}(\sigma)d\sigma$ corresponds to the *a posteriori* distribution when updating a Hilbert-Schmidt prior $d \sigma$. Furthermore, in MLE, $\sigma \mapsto {\mathrm{tr}}[\sigma^{\otimes
n} B^n]$ is known as the *likelihood function*. Since our work is not based on either of these approaches, however, we will not use this terminology and simply refer to $\mu_{B^n}$.
*Reliable Predictions.|*We now show that $\mu_{B^n}$ contains all information that is necessary in order to make reliable predictions about the state of the remaining systems ${\mathcal{S}}_{n+1},
\ldots, {\mathcal{S}}_{n+k}$. To specify these predictions, we consider *hypothetical tests*, a quantum version of a similar concept used in classical statistics. Any such test acts on the joint system consisting of ${\mathcal{S}}_{n+1}, \ldots, {\mathcal{S}}_{n+k}$ (see Fig. \[fig\_setup\]). Mathematically, a test is simply a measurement with binary outcome, “success” or “failure”, specified by a joint POVM $\{T^{{\mathrm{fail}}},{\mathbf{1}}_{{\mathcal{H}}}^{\otimes k} -T^{{\mathrm{fail}}}\}$ on ${\mathcal{H}}^{\otimes
k}$ [^3]. Note that the state of ${\mathcal{S}}_{n+1}, \ldots, {\mathcal{S}}_{n+k}$ could be inferred if we knew which hypothetical tests it would pass. Hence, instead of estimating this state directly, we can equivalently consider the task of predicting the outcomes of the hypothetical tests.
Assume now that we carry out a test ${\mathcal{T}}_{\mu_{B^n}} =
\{T^{{\mathrm{fail}}}_{\mu_{B^n}},{{\mathbf{1}}_{{\mathcal{H}}}^{\otimes k}
-T^{{\mathrm{fail}}}_{\mu_{B^n}}}\}$ depending on $\mu_{B^n}$. We denote by $\rho^{n+k}$ the (unknown) joint state of the systems ${\mathcal{S}}_1, \ldots,
{\mathcal{S}}_{n+k}$ before the tomographic measurements. (As described above, we can assume without loss of generality that the systems are permuted at random, so that $\rho^{n+k}$ is permutation invariant.) If the outcome of the tomographic measurement is $B^n$, then the post-measurement state of the remaining systems is given explicitly by $\rho_{B^n}^k = \frac{1}{{\mathrm{tr}}[B^n \rho^{n}]} {\mathrm{tr}}_{{\mathcal{H}}^{\otimes n}}[B^n
\otimes {\mathbf{1}}_{{\mathcal{H}}}^{\otimes k} \cdot\rho^{n+k}], $ where ${\mathrm{tr}}_{{\mathcal{H}}^{\otimes n}}$ denotes the partial trace over the $n$ measured systems. Hence, the probability that the test ${\mathcal{T}}_{\mu_{B^n}}$ fails for the above state $\rho_{B^n}^k $ equals ${\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} \rho_{B^n}^k]$. The following theorem now provides a criterion under which this failure probability is upper bounded by any given ${\varepsilon}> 0$. Crucially, the criterion only depends on $\mu_{B^n}$, which is obtained by the tomographic data analysis. In other words, $\mu_{B^n}$ allows us to determine which hypothetical tests the state $\rho_{B^n}^k$ would pass.
![[**Illustration of $\mu_{B^n}$.**]{} \[fig\_estimateillustration\] The graphs show $\mu_{B^n}$ for measurements performed on $n=20$ and $n=240$ qubits (for illustration purposes, we only depict the density $\mu_{B^n}$ on the surface of the Bloch ball). Half of the qubits have been measured in the $z$ direction and half in the $y$ direction with relative frequencies of $(0.2, 0.8)$ and $(0.7, 0.3)$, respectively. One observes a rapid decrease in the size of the bright regions (which are connected by a bright tube inside the Bloch ball), which correspond to large values of $\mu_{B^n}$.](Qubitn240new){width="0.44\columnwidth"}
\[theorem:operational\] \[thm:main\] For all $B^n$ let $T^{{\mathrm{fail}}}_{\mu_{B^n}}$ be a POVM element on ${\mathcal{H}}^{\otimes
k}$ such that $$\int \mu_{B^n}(\sigma) {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} \sigma^{\otimes k}] d \sigma
\leq {\varepsilon}c_{n+k, d}^{-1} \ ,$$ where $c_{N, d} = {N+ d^2-1\choose
d^2-1}$. Then, for any $\rho^{n+k}$, $$\begin{aligned}
\big\langle {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} \rho_{B^n}^k]\big\rangle_{B^n}
\leq {\varepsilon}\ ,\end{aligned}$$ where $ \big\langle\cdot\big\rangle_{B^n}$ denotes the expectation taken over all possible measurement outcomes $B^n$ when measuring $\rho^n$ (i.e., outcome $B^n$ has probability ${\mathrm{tr}}[B^n \rho^{n}]$).
As we shall see, the tests are typically chosen such that the integral over $d \sigma$ decreases exponentially with $n$. The additional factor $c_{n+k, d}^{-1}$, which is inverse polynomial in $n+k$, plays therefore only a minor role in the criterion. We also emphasize that the theorem is valid independently of how the systems ${\mathcal{S}}_1, \ldots,
{\mathcal{S}}_{n+k}$ have been prepared. In particular, the (commonly made) assumption that they all contain identical copies of a single-system state is not necessary.
The proof of the theorem, together with a slightly more general formulation, is provided in the *Supplemental Information*. It makes crucial use of the following fact, which has also been used in quantum-cryptographic security proofs: there exists a so-called *de Finetti state* $\tau^N$, i.e., a convex combination of tensor products, such that $\rho^{N}\leq c_{N, d} \cdot \tau^N$ holds for all permutation-invariant states $\rho^{N}$ on ${\mathcal{H}}^{\otimes
N}$ [@HayashiApprox; @ChKoRe09].
*Confidence Regions.|*A *confidence region* is a subset of the single-particle state space which is likely to contain the “true” state. In order to formalize this, we consider the practically relevant case of an experiment that can *in principle* be repeated arbitrarily often. Within the above-described general scenario, this corresponds to the limit where $k$ approaches infinity while $n$, the number of actual runs of the experiment (whose data is analyzed), is still finite and may be small.
Since the initial state $\rho^{n+k}$ of all $n+k$ systems can without loss of generality be assumed to be permutation invariant (see above), the Quantum de Finetti Theorem [@HudMoo76; @RagWer89; @CaFuSc02; @CKMR06; @Renner07] implies that, for fixed $n, k' \in \mathbb{N}$, the marginal state $\rho^{n+k'}$ on $n+k'$ systems is approximated by a mixture of product states, i.e., $$\begin{aligned}
\label{eq_iid}
\rho^{n+k'} = {\mathrm{tr}}_{k-k'}(\rho^{n+k}) \approx \int P(\sigma) \sigma^{\otimes (n+k')}
d \sigma \ ,\end{aligned}$$ for some probability density function $P$ and approximation error proportional to $1/k$. In the limit of large $k$, the marginal state $\rho^{n+k'}$ is thus fully specified by $P$. We can therefore equivalently imagine that all systems were prepared in the same unknown “true” state $\sigma$, which is distributed according to $P$ (see Fig. \[fig\_iid\]). This corresponds to the i.i.d. assumption commonly made in the literature on quantum state tomography, which is therefore rigorously justified within our general setup.
As before, we assume that tomographic measurements are applied to the systems ${\mathcal{S}}_1, \ldots, {\mathcal{S}}_n$, whereas the remaining systems, ${\mathcal{S}}_{n+1}, \ldots, {\mathcal{S}}_{n+k'}$, undergo a test (depending on the output $\mu_{B^n}$ of the data analysis procedure). We may now consider tests that are passed if and only if the true state $\sigma$ is contained in a given subset $\Gamma^\delta_{\mu_{B^n}}$ of the state space. The following corollary provides a sufficient criterion under which the tests are passed, so that $\Gamma^\delta_{\mu_{B^n}}$ are confidence regions. (Note that the criterion refers to additional sets $\Gamma_{\mu_{B^n}}$ that are related to the confidence regions $\Gamma^\delta_{\mu_{B^n}}$; see the *Supplemental Information* for an illustration.)
\[cor:main\] For all $B^n$ let $\Gamma_{\mu_{B^n}}$ be a set of states on ${\mathcal{H}}$ such that $$\begin{aligned}
\int_{\Gamma_{\mu_{B^n}}} \mu_{B^n}(\sigma) d\sigma \geq 1- \frac{{\varepsilon}}{2} c_{2n, d}^{-1} \ .\end{aligned}$$ Then, for any $\sigma$, $$\mathrm{Prob}_{B^n}[\sigma \in \Gamma^{\delta}_{\mu_{B^n}}]\geq
1-\epsilon \ ,$$ where $\mathrm{Prob}_{B^n}$ refers to the distribution of the measurement outcomes $B^n$ when measuring $\sigma^{\otimes n}$ (i.e., outcome $B^n$ has probability ${\mathrm{tr}}[B^n
\sigma^{\otimes n}]$) and where $$\begin{aligned}
\label{eq_Gammadelta} \Gamma^\delta_{\mu_{B^n}}=
\{\sigma: \exists \sigma' \in \Gamma_{\mu_{B^n}} \text{ with }
F(\sigma, \sigma')^2 \geq 1- \delta^2 \} \ ,\end{aligned}$$ with $\delta^2 =\frac{2}{n} (\ln \frac{2}{{\varepsilon}}+ 2 \ln c_{2n, d})$ and $F(\sigma, \sigma')= \|\sqrt{\sigma} \sqrt{\sigma'}\|_1$ the fidelity.
The main idea for the proof of the corollary is to apply the above theorem to tests (acting on $k' = n$ systems) derived from Holevo’s optimal covariant measurement [@Holevo]. We refer to the *Supplemental Information* for the technical proof.
Note that $1-{\varepsilon}$ can be interpreted as the *confidence level* of the statement that the true state $\sigma$ is contained in the set $\Gamma^\delta_{\mu_{B^n}}$. Crucially, the claim is valid for all $\sigma$. In particular, it is independent of any initial probability distribution, $P$, according to which $\sigma$ may have been chosen (see Eq. \[eq\_iid\]). In other words, the operational interpretation of the sets $\Gamma_{\mu_{B^n}}^\delta$ as confidence regions does not depend on any extra assumptions about the preparation procedure or on the specification of a prior. In fact, $\sigma$ could even be chosen “maliciously”, for example in a quantum cryptographic context, where an adversary may try to pretend that a system has certain properties (e.g., that its state is entangled while in reality it is not).
Obviously, the assertion that a state $\sigma$ is contained in a certain set $\Gamma_\mu^{\delta}$ can only be considered a good approximation of $\sigma$ if the set $\Gamma_\mu^{\delta}$ is small. This is indeed the case for reasonable choices of the measurement $\{B^n\}$. For instance, in the practically important case where each system is measured independently and identically with POVM $\{E_i\}$, the confidence region is, for generic states, asymptotically of size proportional to $\frac{1}{\sqrt{n}}$ in the (semi-)norm on the set of quantum states induced by the POVM: $\|\cdot
\|_{\{E_i\}}=\sum_i |{\mathrm{tr}}(E_i \cdot) |$ (see *Supplemental Information* and [@MattewsWehnerWinter]).
*Conclusion.|*Despite conceptual differences, our technique is not unrelated to MLE and Bayesian estimation. As mentioned before, $\mu(\sigma)$ is proportional to the likelihood function and, therefore, methods to construct confidence regions with our technique are likely to use adaptations of techniques from MLE. Also, $\mu_{B^n}(\sigma)d\sigma$ corresponds to the probability measure obtained from applying Bayes’ updating rule to the Hilbert-Schmidt measure; a fact that implies near optimality [^4] of our method in the practically most relevant case of independent tomographically complete measurements [@TanKom05].
Recently, another novel approach to quantum state tomography has been proposed [@CompressedTomo; @EfficientTomo], which yields reliable error bounds similar to ours. A central difference between this approach and ours is the level of generality. In [@CompressedTomo; @EfficientTomo] a specific sequence of measurement operations is proposed, which is adapted to systems whose states are fairly pure. Under this condition, the estimate converges fast and, in addition, can be computed efficiently. In contrast, our method can be applied to arbitrary measurements (i.e., any tomographic data may be analyzed). Accordingly, the convergence of the confidence region depends on the choice of these measurements. However, we do not propose any specific algorithm for the efficient computation of confidence regions.
Finally, we refer to the very recent work of Blume-Kohout [@Blume12] for an excellent discussion of the notion of confidence regions in quantum state tomography. In particular, he shows that confidence regions, as considered here, can be defined via likelihood ratios.
*Acknowledgements.|*We thank Robin Blume-Kohout for useful comments on earlier versions of this work. We acknowledge support from the Swiss National Science Foundation (grants PP00P2-128455 and 200020-135048, and through the National Centre of Competence in Research ‘Quantum Science and Technology’), the German Science Foundation (grant ), and the European Research Council (grant 258932).
[46]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, **, vol. (, ).
, ** (, , ).
, ****, ().
, ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , in **, edited by (, ), vol. , pp. .
, , , ****, ().
, , , , ****, ().
, , , , in **, edited by (, ), pp. .
, , , , ****, ().
, , , , , , , , , , ****, ().
, , , ****, ().
, ****, ().
, , , , , , , , , , , ****, ().
, , , , , , ****, ().
, , , , , , , , , , ****, ().
, ****, (), ISSN .
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , ****, ().
, ** (, ).
(), .
, , , ****, ().
, ****, ().
, in **, edited by , , (, ), vol. of **, pp. .
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, ** (, , ).
, , , ****, ().
, , , , , ****, ().
, , , , , , , , , ****, ().
[Supplemental Information]{} {#supplemental-information .unnumbered}
============================
[**Note: This document contains additional material (Appendices A, B, C, and D) that is not included in the “Supplemental Information” accompanying the journal version.**]{}
In the first part, we give precise statements and proofs of our technical results (Section 1), discuss the practically important case of independent measurements (Section 2) and present further remarks (Section 3). The second part explains how to represent states on the symmetric subspace (Appendix A) as well as functions on the state space (Appendices B and C), and concludes with examples (Appendix D).
1 $\quad$ Statements and Proofs
===============================
Let ${\mathcal{H}}$ be a Hilbert space of finite dimension $d$, i.e., ${\mathcal{H}}\cong
{\mathbb{C}}^d$. We denote the set density matrices on ${\mathcal{H}}$ by ${\mathcal{S}}({\mathcal{H}})$ and the subset of pure states by ${\mathcal{P}}({\mathcal{H}})$. Note that ${\mathcal{P}}({\mathcal{H}})$ can be identified with ${\mathbb{C}P^{d-1}}$, the complex projective space of dimension $d-1$. ${\mathbb{C}P^{d-1}}$ carries a natural action of the unitary group $U(d)$. The Haar measure on $U(d)$ therefore descends to a measure on ${\mathcal{P}}({\mathcal{H}})$ which is invariant under the action of $U(d)$. We denote this measure by $d\phi$ and fix the normalisation so that $\int d\phi =1$. The symmetric subspace ${\mathrm{Sym}}^n({\mathcal{H}})$ of ${\mathcal{H}}^{\otimes n}$ is defined as the space of vectors that are invariant under the action of the symmetric group $S_n$ that permutes the tensor factors. Since the action of $S_n$ commutes with the action of the unitary group on ${\mathcal{H}}^{\otimes n}$, $U(d)$ acts on ${\mathrm{Sym}}^n({\mathcal{H}})$ as well. We denote the dimension of ${\mathrm{Sym}}^n({\mathcal{H}})$ by $\dim(n, d)$. The following lemma [@HayashiApproxSupp; @ChKoRe09Supp] is crucial in the derivation of the main results.
\[lem:postselect\] Let $\rho^n \in {\mathcal{S}}({\mathrm{Sym}}^n({\mathbb{C}}^d))$. Then $$\rho^n\leq \dim(n, d) \int \phi^{\otimes n} d\phi.$$ Furthermore, $\dim(n, d) = {n+d-1 \choose n} \leq (n+1)^{d-1}$.
The state defined by the integral on the right hand side is sometimes called *de Finetti state*. Note that the corresponding statement mentioned in the Letter for general permutation-invariant density operators $\rho^n$ is obtained from this lemma by considering a purification of $\rho^n$ in the symmetric subspace (see [@ChKoRe09Supp]).
The space ${\mathrm{Sym}}^n({\mathbb{C}}^{d})$ is irreducible under the action of the unitary group $U(d)$ [@FultonHarris91]. The operator $\int \phi^{\otimes n} d\phi$ is supported on ${\mathrm{Sym}}^n({\mathbb{C}}^d)$ and invariant under the action of $U(d)$. By Schur’s lemma we therefore have $$\dim(n, d) \int \phi^{\otimes n} d\phi = {\mathbf{1}}_{{\mathrm{Sym}}^n({\mathbb{C}}^d)}.$$ The claim follows since $$\rho^n \leq {\mathbf{1}}_{{\mathrm{Sym}}^n({\mathbb{C}}^d)}$$ holds for any density operator. $\dim(n, d)$ equals ${n+d-1 \choose n}$ and is easily seen to be upper bounded by $(n+1)^{d-1}={\mathrm{poly}}(n)$.
Consider now the measure $d\phi$ on a tensor product space ${\mathcal{H}}\otimes {\mathcal{K}}$, where ${\mathcal{K}}\cong {\mathcal{H}}\cong {\mathbb{C}}^d $ and perform the partial trace operation over system ${\mathcal{K}}$. We denote the resulting measure by $d\sigma$ on ${\mathcal{S}}({\mathcal{H}})$ and note that it may also be defined as the measure induced by the Hilbert-Schmidt metric [@ZycSom01]. (In the Letter, we refer to $d \sigma$ as the Hilbert-Schmidt measure.) For any POVM element $B^n$ on ${\mathcal{H}}^{\otimes
n}$, our data analysis procedure produces a probability distribution $$\mu_{B^n}(\sigma)d\sigma:= \frac{1}{c_{B^n}} {\mathrm{tr}}[ B^n \sigma^{\otimes n}]d\sigma,$$ where $c_{B^n}= \int {\mathrm{tr}}B^n \sigma^{\otimes n} d\sigma= {\mathrm{tr}}[B^n
\otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n}\cdot {\mathbf{1}}_{{\mathrm{Sym}}^n({\mathcal{H}}\otimes {\mathcal{K}})}]$. Let $\rho^{n+k}$ be a permutation-invariant density operator on ${\mathcal{H}}^{\otimes n+k}$ and $\varrho^{n+k}$ a purification with support on ${\mathrm{Sym}}^{n+k}({\mathcal{H}}\otimes {\mathcal{K}})$ (see e.g. [@CKMR06Supp]). That is, $\varrho^{n+k}$ is pure and ${\mathrm{tr}}_{{\mathcal{K}}^{\otimes
n}}\varrho^{n+k}=\rho^{n+k}$. Denote by $\varrho^{k}_{B^n}=\frac{1}{{\mathrm{tr}}[B^n \rho^n]}{\mathrm{tr}}_n [ ( B^n\otimes
{\mathbf{1}}_{{\mathcal{K}}}^{\otimes n} ) \otimes {\mathbf{1}}_{{\mathcal{H}}\otimes {\mathcal{K}}}^{\otimes k}
\cdot \varrho^{n+k}] $ the post-measurement state and note that it appears with probability ${\mathrm{tr}}[B^n \rho^n]$. Furthermore, we define $\nu_{B^n}(x):= \frac{1}{c_{B^n}} {\mathrm{tr}}[B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes
n} \cdot {{| x \rangle}\!{\langle x |}}^{\otimes n}]$ for $x \in {\mathcal{H}}\otimes {\mathcal{K}}$. The following theorem holds for any POVM element $T^{{\mathrm{fail}}}_{\mu_{B^n}}$ on $({\mathcal{H}}\otimes {\mathcal{K}})^{\otimes k}$.
If for all ${B^n}$ $$\begin{aligned}
\label{eq_criterion} \int \nu_{B^n}(x) {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} {{| x \rangle}\!{\langle x |}}^{\otimes k} ]dx
\leq {\varepsilon}{n+k+ d^2-1\choose d^2-1}^{-1},
\end{aligned}$$ then $$\begin{aligned}
\big\langle {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} \varrho_{B^n}^k]\big\rangle_{B^n}
\leq {\varepsilon}\ ,\end{aligned}$$ where $ \big\langle\cdot\big\rangle_{B^n}$ denotes the expectation taken over all possible measurement outcomes $B^n$ according to the probability distribution ${\mathrm{tr}}[B^n \rho^{n}]$.
Note that the probability distribution $\mu_{B^n}(\sigma)d\sigma$ is obtained from the probability distribution $\nu_{B^n}(x) dx$ by taking the partial trace over the purifying system ${\mathcal{K}}$. The above theorem therefore immediately implies the theorem in the Letter, which corresponds to the specialisation where $T^{{\mathrm{fail}}}_{\mu_{B^n}}$ acts only on ${\mathcal{H}}^{\otimes k}$.
$$\begin{aligned}
\big\langle {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} \cdot & \varrho_{B^n}^k]\big\rangle_{B^n}
= \sum_{B^n} {\mathrm{tr}}[(B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n}) \otimes T^{{\mathrm{fail}}}_{\mu_{B^n}} \cdot \varrho^{n+k}]\\
& \leq \dim(n+k, d^2) \sum_{B^n} \int {\mathrm{tr}}[(B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n} ) \otimes T^{{\mathrm{fail}}}_{\mu_{B^n}} \cdot {{| x \rangle}\!{\langle x |}}^{\otimes n+k}] dx\\
&= \dim(n+k, d^2) \sum_{B^n} c_{B^n} \int \nu_{B^n}(x) {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\mu_{B^n}} {{| x \rangle}\!{\langle x |}}^{\otimes k}]dx\\
& \leq {\varepsilon},\end{aligned}$$
where we used Lemma \[lem:postselect\] in the first inequality and the assumption, , in the second.
For a subset $\Gamma_{\mu}$ of ${\mathcal{S}}({\mathcal{H}})$, we define $$\Gamma^\delta_\mu= \{\sigma: \exists \sigma' \in \Gamma_\mu \text{ with } P(\sigma, \sigma')\leq \delta \},$$ where $P(\sigma, \sigma')$ is the purified distance defined as $\sqrt{1-F(\sigma, \sigma')^2}$, where $F(\sigma, \sigma')= \|
\sqrt{\sigma} \sqrt{\sigma'} \|_1 = {\mathrm{tr}}\sqrt{\sqrt{\sigma} \sigma' \sqrt{\sigma}}$ is the fidelity. For more details regarding the purified distance as well as its relation to the trace distance see [@dualities].
For all $B^n$, let $\Gamma_{\mu_{B^n}}$ be such that $$\begin{aligned}
\label{eq:gamma}
\int_{\Gamma_{\mu_{B^n}}} \mu_{B^n}(\sigma) d\sigma \geq 1- \frac{{\varepsilon}}{2} {2n+d^2-1\choose d^2-1}^{-1} \ \end{aligned}$$ and let $\delta :=\sqrt{\frac{2}{n} (\ln \frac{2}{{\varepsilon}}+ 2 \ln {2n+d^2-1\choose d^2-1} )}$. Then for all $\sigma$, $$\text{Prob}_{B^n}[\sigma \in \Gamma^{\delta}_{\mu_{B^n}}]\geq 1-\epsilon$$ where the probability is with respect to the measurement outcomes $B^n$ according to the probability distribution ${\mathrm{tr}}[B^n
\sigma^{\otimes n}]$.
See Figure \[fig\_convergence\] for an illustration of $\Gamma^{\delta}_{\mu_{B^n}}$ and its relation to $\mu_{B^n}$.
![[**Construction of the confidence region.**]{} The graph schematically illustrates a possible distribution $\mu \equiv
\mu_{B^n}$ (blue line), a high probability set $\Gamma_{\mu}$ of $\mu$ (blue bar), and the set $\Gamma_{\mu}^{\delta}$ which includes a $\delta$-region around $\Gamma_{\mu}$ (orange bar). In the scenario depicted by Fig. 3 in the Letter, the state $\sigma$ chosen by the preparation procedure is with high probability (at least $1-{\varepsilon}$) in $\Gamma_{\mu}^{\delta}$, which can therefore be seen as a confidence region. []{data-label="fig_convergence"}](PlotGamma2){width="0.7\columnwidth"}
The failure probability of the test is given by $$\begin{aligned}
P^{{\mathrm{fail}}}(P)& := \int P(\sigma) \sum_{B^n} {\mathrm{tr}}[B^n \cdot \sigma^{\otimes n}] \Theta_{\overline{\Gamma_{\mu_{B^n}}^\delta}}(\sigma) d\sigma\end{aligned}$$ where $\Theta_{\overline{\Gamma_{\mu_{B^n}}^\delta}}(\sigma)$ equals one for $\sigma$ in the set $\overline{\Gamma_{\mu_{B^n}}^\delta}$ and zero otherwise ($\overline{A}$ denotes the complement of a subset $A$ of ${\mathcal{S}}({\mathcal{H}})$). This can be rewritten more conveniently as $$\begin{aligned}
P^{{\mathrm{fail}}}(P)= \sum_{B^n} c_{B^n} \int_{\overline{\Gamma_{\mu_{B^n}}^\delta}} P(\sigma) \mu_{B^n}(\sigma) d\sigma.\end{aligned}$$ Instead of the a priori probability density $P(\sigma)$ we consider a probability density $Q(\psi)$ on the pure states ${\mathcal{P}}({\mathcal{H}}\otimes {\mathcal{K}})$, where ${\mathcal{K}}\cong {\mathcal{H}}$, that gives rise to $P(\sigma)$. That is, $Q(\psi)$ satisfies $$\int_{A} P(\sigma) d\sigma = \int_{\psi: {\mathrm{tr}}_{{\mathcal{K}}} \psi \in A} Q(\psi) d\psi$$ for all measurable subsets $A$ of ${\mathcal{S}}({\mathcal{H}})$. Note that such a probability density $Q(\psi)$ always exists as we can purify the state of the particle with a purifying space ${\mathcal{K}}\cong
{\mathcal{H}}$. $\mu_{B^n}(\sigma)$ is replaced by $$\nu_{B^n}(\psi):=\frac{1}{c_{B^n}} {\mathrm{tr}}[( B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n}) \cdot \psi^{\otimes n}].$$ Furthermore we consider the following extension of the set $\Gamma_{\mu}$: $$\Omega_{\mu} = \{\psi \in {\mathcal{P}}({\mathcal{H}}\otimes {\mathcal{K}}): {\mathrm{tr}}_{{\mathcal{K}}} \psi \in \Gamma_\mu \}$$ and of $\Gamma^{\delta}_{\mu}$: $$\Omega^{\delta}_{\mu} = \{\psi \in {\mathcal{P}}({\mathcal{H}}\otimes {\mathcal{K}}): {\mathrm{tr}}_{{\mathcal{K}}} \psi \in \Gamma^\delta_\mu \}.$$ The failure probability of the test can then be expressed as $$\begin{aligned}
\label{eq:fail}
P^{{\mathrm{fail}}}(P)=\sum_{B^n} c_{B^n} \int_{\overline{\Omega_{\mu_{B^n}}^\delta}} Q(\psi) \nu_{B^n}(\psi) d\psi.\end{aligned}$$ Let $k' \in {{\mathbb{N}}}$ be the number of systems that we use to approximate the test by a test procedure that is given by a POVM $\{T^{{\mathrm{fail}}}, {\mathbf{1}}- T^{{\mathrm{fail}}}\}$. Defining $$T^{{\mathrm{fail}}}_{\Omega_\mu^{\delta/2}}:= \dim(k', d^2)\int_{\overline {\Omega_\mu^{\delta/2}}} \phi^{\otimes k'} d\phi.$$ we see that for all $\psi \in \overline{\Omega_{\mu}^\delta}$ $$\begin{aligned}
{\mathrm{tr}}T^{{\mathrm{fail}}}_{\Omega_\mu^{\delta/2}} \psi^{\otimes k'}& =1- \dim(k', d^2)\int_{\Omega_\mu^{\delta/2}} {\mathrm{tr}}[\phi^{\otimes k'} \psi^{\otimes k'} ]d\phi \\
& \geq 1- \dim(k', d^2)\max_{\phi \in \Omega_\mu^{\delta/2}} F({\mathrm{tr}}_{{\mathcal{K}}} \phi, {\mathrm{tr}}_{{\mathcal{K}}} \psi)^{k'} \\
& \geq 1- \underbrace{ \dim(k', d^2)e^{-\frac{\delta^2}{2} k'}}_{=:{\varepsilon}''}.\end{aligned}$$ Inserting this estimate into leads to $$\begin{aligned}
P^{{\mathrm{fail}}}(P)& \leq \sum_{B^n} c_{B^n} \int_{\overline{\Omega_{\mu_{B^n}}^\delta}} Q(\psi) \nu_{B^n}(\psi) ({\mathrm{tr}}[ T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}}\cdot \psi^{\otimes k'}]+{\varepsilon}'') d\psi.\end{aligned}$$ We now remove the restriction in the integral, thereby further weakening the estimate and obtain $$\begin{aligned}
P^{{\mathrm{fail}}}(P)& \leq {\varepsilon}'' + \sum_{B^n} c_{B^n}\int Q(\psi) \nu_{B^n}(\psi) {\mathrm{tr}}[ T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}}\cdot \psi^{\otimes k'}]d\psi,\\
& ={\varepsilon}'' + \sum_{B^n} \int Q(\psi) {\mathrm{tr}}[(B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n}) \cdot \psi^{\otimes n}] {\mathrm{tr}}[ T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}}\cdot \psi^{\otimes k'}] d\psi,\\
& ={\varepsilon}'' + \sum_{B^n}{\mathrm{tr}}[(B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n})\otimes T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}}\cdot ( \int Q(\psi) \psi^{\otimes n+k'} d\psi) ].\end{aligned}$$ Applying Lemma \[lem:postselect\] to the state $\int Q(\psi)
\psi^{\otimes n+k'}d \psi$ we can find an upper bound on this quantity which is independent of the initial distribution $P(\sigma)$ (or $Q(\psi)$): $$\begin{aligned}
P^{{\mathrm{fail}}}(P)& \leq {\varepsilon}'' + \dim(n+k', d^2) \sum_{B^n}{\mathrm{tr}}[ (B^n \otimes {\mathbf{1}}_{{\mathcal{K}}}^{\otimes n}) \otimes T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}} \cdot( \int \psi^{\otimes n+k'} d\psi)],\\
&={\varepsilon}''+ \dim(n+k', d^2) \sum_{B^n}c_{B^n}\int_{\Omega_{\mu_{B^n}}} \nu_{B^n}(\psi) {\mathrm{tr}}[T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}}\cdot \psi^{\otimes k'}] d\psi \\
& \quad + \dim(n+k', d^2)\sum_{B^n}c_{B^n}\int_{\overline{\Omega_{\mu_{B^n}}}} \nu_{B^n}(\psi) {\mathrm{tr}}[ T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}}\cdot \psi^{\otimes k'} ] d\psi. \end{aligned}$$ Since for $\psi \in \Omega_{\mu_{B^n}}$ $$\begin{aligned}
{\mathrm{tr}}[ T^{{\mathrm{fail}}}_{\Omega_\mu^{\delta/2}}\cdot \psi^{\otimes k'}]& \leq \dim(k', d^2)\max_{\phi \in\overline{\Omega^{\delta/2}}} F( {\mathrm{tr}}_{{\mathcal{K}}} \phi, {\mathrm{tr}}_{{\mathcal{K}}} \psi)^{k'}\leq \dim(k', d^2)e^{-\frac{\delta^2}{2} k'}={\varepsilon}''\end{aligned}$$ and since $${\mathrm{tr}}[ T^{{\mathrm{fail}}}_{\Omega^{\delta/2}_{\mu_{B^n}}} \cdot \psi^{\otimes k'} ]\leq 1$$ we find $$\begin{aligned}
P^{{\mathrm{fail}}}(P)&\leq {\varepsilon}''+ \dim(n+k', d^2){\varepsilon}'' + \dim(n+k', d^2)\sum_{B^n}c_{B^n}\int_{\overline{\Omega_{\nu_{B^n}}}} \nu_{B^n}(\psi) d\psi\\
& = {\varepsilon}''+ \dim(n+k', d^2){\varepsilon}'' + \dim(n+k', d^2)\sum_{B^n}c_{B^n}\int_{\overline{\Gamma_{\mu_{B^n}}}} \mu_{B^n}(\sigma) d\sigma.\end{aligned}$$ We now set $k'=n$ and use the assumption $$\int_{\Gamma_{\mu_{B^n}} }\mu_{B^n}(\sigma) d\sigma \geq 1- \frac{{\varepsilon}}{2} \dim(2n, d^2)^{-1}.$$ for all $\mu_{B^n}$. This results in $$\begin{aligned}
P^{{\mathrm{fail}}}(P)&\leq {\varepsilon}''+ \dim(2n, d^2){\varepsilon}'' + \frac{{\varepsilon}}{2}\leq \dim(2n, d^2)^2 e^{-\frac{\delta^2}{2} n} +\frac{{\varepsilon}}{2}.\end{aligned}$$ Choosing $\delta=\sqrt{ \frac{2}{n} (\ln{\frac{2}{{\varepsilon}}}+ 2 \ln \dim(2n, d^2))}$ ensures that $$\begin{aligned}
P^{{\mathrm{fail}}}(P)&\leq {\varepsilon}.\end{aligned}$$ Inserting for $P$ a Dirac delta distribution concludes the proof.
2 $\quad$ Independent Measurements
==================================
We now restrict our discussion to the case where $B^n$ is of product form, i.e. $$B^n = \prod_{i=1}^r E_i^{\otimes f^{(i)}}$$ where $E_i$ are the elements of a POVM, i.e., $E_i \geq 0$ and $\sum_iE_i={\mathbf{1}}$. $f=(f^{(1)}, \ldots, f^{(r)})$ is the vector containing the frequencies with which the outcomes occur, i.e. $f^{(i)} \in {{\mathbb{N}}}$ and $\sum_{i=1}^n f^{(i)}=n$. We find $$\mu_{B^n}(\sigma) =\frac{1}{c_{B^n}} \prod_{i=1}^r ({\mathrm{tr}}E_i \sigma)^{f^{(i)}},$$ where $c_{B^n}=\int {\mathrm{tr}}B^n \sigma^{\otimes n} d \sigma$. We now want to investigate the maximum of this function which we assume for simplicity to be unique, i.e. we assume that the POVM is tomographically complete. Since the log function is monotonically increasing, the $\sigma$ which maximises this function can also be expressed as $$\begin{aligned}
\label{eq:MLE}
\sigma_{\max}:= \mathrm{argmax}_{\sigma} (\sum_i \bar{f}^{(i)} \log {\mathrm{tr}}E_i \sigma),\end{aligned}$$ where $\bar{f}^{(i)}=\frac{f^{(i)}}{n}$ are the relative frequencies. This shows that the value at which $\mu_{B^n}$ is maximised coincides with the density matrix that the Maximum Likelihood Estimation method infers since $\sum_i f^{(i)} \log {\mathrm{tr}}E_i \sigma$ is the so-called ”log likelihood function” [@jezek-hradil Eqs. (2) and (3)].
The MLE method is therefore consistent with our method and our work can be seen as a theoretical justification of it. We emphasize that in contrast to MLE, our work shows how to compute reliable error bars. This implies in particular that most of the likely states (i.e. the states within the error bars) are not on the boundary of the state space, even though the maximum might lie on the boundary. Our method can therefore be seen as a resolution of the “problem” that the states predicted by MLE are unphysical because they lie on the boundary.
We now want to study the decay of $\mu_{B^n}$ around its maxima. In general this is not easy, as the maxima might lie on the boundary of the set. Useful statements can be made however, when the maxima are in the interior of the set of states. We consider the decay exponent (i.e. $\frac{-1}{n}$ times the natural logarithm) of the function ${\mathrm{tr}}B^n \sigma^{\otimes n}$ which is $$\begin{aligned}
\label{eq:decay}
- \sum_i \bar{f}^{(i)}\ln {\mathrm{tr}}E_i \sigma.\end{aligned}$$ We then search for the extreme points of the function $$- \sum_i \bar{f}^{(i)}\ln {\mathrm{tr}}E_i \sigma + c ({\mathrm{tr}}\sigma -1)$$ where we introduced the Lagrange multiplier $c$ in order to take care of the normalisation of the density matrix. The extreme points are characterised by the equations $$- \sum_i \bar{f}^{(i)}\frac{E_i}{{\mathrm{tr}}E_i \sigma}= c{\mathbf{1}}$$ $${\mathrm{tr}}\sigma =1.$$ Let us for simplicity restrict to the case, where the $E_i$ are linearly independent, then, since the $E_i$ form a POVM, we have $$\begin{aligned}
\label{eq:extreme}\bar{f}^{(i)}= {\mathrm{tr}}E_i \sigma \end{aligned}$$ as a condition for an extreme point.
In order to carry over these findings to the estimate density, we need to understand the behaviour of the normalisation constant $c_{B^n}=\int {\mathrm{tr}}B^n \sigma^{\otimes n}d\sigma$. Since $$\int {\mathrm{tr}}B^n \sigma^{\otimes n}d\sigma \leq \max_\sigma {\mathrm{tr}}B^n \sigma^{\otimes n}$$ and by Lemma \[lem:postselect\] $$\int {\mathrm{tr}}B^n \sigma^{\otimes n}d\sigma \geq \frac{1}{{\mathrm{poly}}(n)} \max_\sigma {\mathrm{tr}}B^n \sigma^{\otimes n}$$ we find that $\frac{1}{n}\ln c_{B^n}$ approaches the maximum of , $- \sum_i \bar{f}^{(i)}\ln \bar{f}^{(i)}$, for large $n$. The decay exponent, $-\frac{1}{n}\ln \mu_{B^n}$, is therefore asymptotically equal to the relative entropy (in units of the natural logarithm) $$D(\bar{f}\| E(\sigma))= \sum_i \bar{f}^{(i)}(\ln \bar{f}^{(i)}-\ln {\mathrm{tr}}E_i\sigma)$$ where $E(\sigma)=({\mathrm{tr}}E_1 \sigma, \ldots, {\mathrm{tr}}E_r \sigma)$ (this shows in particular that the extreme points are maxima since the relative entropy is nonnegative). Pinsker’s inequality $$D(\bar{f}\| E(\sigma))\geq \frac{1}{2}\|\bar{f}- E(\sigma)\|_1^2.$$ implies that the error bars around the maxima are therefore given by $$\epsilon = O\left(\frac{1}{\sqrt{n}}\right)$$ in the distance on the set of density matrices induced by the norm [@MattewsWehnerWinterSupp] $$\|X\|:= \|E(X)\|_1=\sum_i |{\mathrm{tr}}E_i X|.$$
3 $\quad$ Remarks
=================
We remark that it is possible to adapt our method to the case where additional information, e.g., about the rank of the state or its symmetry, is available, and thereby to establish confidence regions also in these situations. For example, if it is known that the state $\rho^{n+k}$ is invariant under local actions (on the individual systems ${\mathcal{H}}$) of unitaries from a given set ${\mathcal{U}}$, then the integral in can be restricted to a subset ${\mathcal{S}}$ of states $\sigma$ on ${\mathcal{H}}$ such that $U \sigma U^{\dagger} = \sigma$ for any $U
\in {\mathcal{U}}$. Accordingly, the confidence region is given by ${{\mathcal{S}}\cap
\Gamma_{\mu_{B^n}}^{\delta}}$, provided the criterion of the corollary is satisfied.
We also note that the output of the data analysis, $\mu_{B^n}$, can be specified using a representation in terms of generalised spherical harmonics. Data obtained from measurements on $n$ systems specify exactly the moments of degree less than $n$ of this representation. In particular, these moments contain all information that is needed for a later update of $\mu_{B^n}$ based on additional measurement data (see Appendices). While in the case of i.i.d. measurements (with a finite number of outcomes) this representation of the measurement data is costly ($O(\text{poly}(n))$ bits) when compared to simply storing the frequencies ($O(\log n)$ bits), it may be useful in the case of non-i.i.d. measurements or measurements with an unbounded (or even continuous) set of outcomes, since it does not depend on the number of outcomes.
A $\quad$ Quasi-Probability Distributions
=========================================
In this section we derive quasi-probability distribution representations for operators on ${\mathrm{Sym}}^n({\mathbb{C}}^d)$ similar to the P- and Q-representations that are well-known from quantum optics.
\[thm:Qrep\] Let $B$ be an operator on ${\mathrm{Sym}}^{n}({\mathbb{C}}^{d})$. Then $B$ is uniquely determined by its Q-representation, the function $$Q_{B}(x)= {\langle x |}^{\otimes n} B {| x \rangle}^{\otimes n},$$ where ${| x \rangle}=\sum_i x_i {| i \rangle}$, $x=(x_1, \ldots, x_d)^T \in {\mathbb{C}}^d$ and $\sum_i |x_i|^2=1$. When convenient we will view $Q$ as a function on ${\mathbb{C}P^{d-1}}$.
We adopt an argument very similar to the one used in [@Perelomov86 p. 30] in the context of Glauber coherent states. Note that the values $${\langle x |}^{\otimes n} B {| x \rangle}^{\otimes n}$$ for $x \in {\mathbb{C}}^d$ are determined by ${\langle x |}^{\otimes n} B {| x \rangle}^{\otimes n}$ for $x \in {\mathbb{C}}^d$ with $\sum_i |x_i|^2=1$. It therefore suffices to show that the values $ {\langle x |}^{\otimes n} B {| x \rangle}^{\otimes n}$, $x \in \mathbb{C}^{d}$ determine $ {\langle x |}^{\otimes n} B {| x' \rangle}^{\otimes n}$, $x, x' \in \mathbb{C}^{d}$ uniquely. Let $\{ {| m \rangle} \}$ be the Gelfand-Zetlin basis for ${\mathrm{Sym}}^{n} ({\mathbb{C}}^{d})$ (see Appendix B) and write $$B = \sum_{m,m'} B_{m,m'} {| m \rangle} {\langle m' |}.$$ The function $ {\langle m' | x' \rangle}^{\otimes n}$ is a polynomial in $x' \in
{\mathbb{C}}^{d}$ and ${\langle x |}^{\otimes n}{| m \rangle}$ is a polynomial in $\bar{x}$, the complex conjugate of $x$. Defining $$\alpha = \frac{\bar{x}+x'}{2},\quad \beta = i \frac{\bar{x}-x'}{2}$$ we see that $$x' = \alpha + i \beta,\quad \bar{x} = \alpha - i \beta$$ and hence ${\langle x |}^{\otimes n} B {| x' \rangle}^{\otimes n}=\sum_{m, m'}
B_{m, m'}{\langle x |}^{\otimes n}{| m \rangle}{\langle m' | x' \rangle}^{\otimes n}$ is a polynomial in $\alpha$, $\beta$. Note that every polynomial (in fact every entire function) is determined by its values for real parameters, i.e. by $\alpha, \beta \in {\mathbb{R}}^d$ in our case. This can be seen by writing the polynomial in the form of a Taylor series (around a real point, e.g. 0). The coefficients in this series are partial derivatives (evaluated at 0), which can be taken in real directions without losing generality and are therefore only dependent on real values of the polynomial. Since for real $\alpha, \beta$ $$\begin{aligned}
\mathrm{Im} (\alpha) &= \frac{1}{2} (-\mathrm{Im} (x) + \mathrm{Im}( x') ) = 0
\\
\mathrm{Im} (\beta) &= \frac{1}{2} ( \mathrm{Re}( x) - \mathrm{Re} (x') ) = 0\end{aligned}$$ or, in other words $x=x'$, it follows that ${\langle x |}^{\otimes n} B
{| x' \rangle}^{\otimes n}$, as a function of $x, x'$, is wholly determined by the values ${\langle x |}^{\otimes n} B {| x \rangle}^{\otimes n},
x \in {\mathbb{C}}^d$.
\[thm:Prep\] Let $B$ be an operator on ${\mathrm{Sym}}^{n}({\mathbb{C}}^{d})$. Then $B$ may be represented in the form $$B = \int_{{\mathbb{C}}P^{d-1}} P_{B}(x) {{| x \rangle}\!{\langle x |}}^{\otimes n} dx$$ where $$P_{B}(x) = \sum_{\ell \in {{\mathbb{N}}}} \sum_{m} p_{B}(\ell, m) y_{\ell,m}(x)$$ with the second sum extending over the Gelfand-Zetlin basis for the irreducible representation of $U(d)$ with highest weight $\underbrace{(\ell, 0, \ldots, 0, -\ell)}_{d}$ (see Appendix B). The constants $p_{B}(\ell, m)$ are uniquely determined by $B$ for $\ell
\leq n$ and are arbitrary otherwise.
Two lemmas will be needed in order to prove the theorem.
*([@KlaSka85 p. 35])* Let $D$ be the space of operators on ${\mathrm{Sym}}^n({\mathbb{C}}^d)$ that can be represented in the form $$B = \int_{{\mathbb{C}P^{d-1}}} P_B(x) {{| x \rangle}\!{\langle x |}}^{\otimes n} dx$$ for some $P_B \in L^2({\mathbb{C}P^{d-1}})$. Furthermore let $E$ be the space of operators on ${\mathrm{Sym}}^n({\mathbb{C}}^d)$ with vanishing Q-representation. Then $D^{\bot} = E$, where $D^{\bot} = \{ A : {\mathrm{tr}}A B^{\dagger} = 0\quad \forall B \in D \}$. \[KlaSkalem\]
If $A \in E$, then for all $B \in D$ $${\mathrm{tr}}A B^{\dagger} = {\mathrm{tr}}A \int \overline{P_B(x)} {{| x \rangle}\!{\langle x |}}^{\otimes n} dx = \int \overline{P_B(x)} {\langle x |}^{\otimes n} A {| x \rangle}^{\otimes n} dx = 0,$$ hence $A \in D^{\bot}$. Conversely let $A \in D^{\bot}$, then $${\mathrm{tr}}A B^{\dagger} = 0 \quad \forall B \in D.$$ Writing this out results in $\int \overline{P_B(x)} {\langle x^{\otimes n} |} A {| x^{\otimes n} \rangle} dx = 0$, for all functions $P_B(x)$ on ${\mathbb{C}P^{d-1}}$. This implies $${\langle x^{\otimes n} |} A {| x^{\otimes n} \rangle} =0 \quad \forall x \in {\mathbb{C}P^{d-1}},$$ since only the identically vanishing function is orthogonal to all square integrable functions on ${\mathbb{C}P^{d-1}}$ (and, in particular, to itself).
\[lem:onbasis\] The operators $ \int dx \ y_{\ell, m}(x) {{| x \rangle}\!{\langle x |}}^{\otimes n}$ are non-vanishing and orthogonal with respect to the Hilbert-Schmidt inner product for $\ell \leq n$ and $m$ a corresponding Gelfand-Zetlin pattern. For $\ell >n$, $ \int dx \ y_{\ell, m}(x) {{| x \rangle}\!{\langle x |}}^{\otimes n}=0$.
We calculate $$\begin{aligned}
{\mathrm{tr}}[ \int dx \ y_{\ell, m}(x) & {{| x \rangle}\!{\langle x |}}^{\otimes n} \int dz \ \overline{y_{\ell, m}(z)} {{| z \rangle}\!{\langle z |}}^{\otimes n}]\\
& =
\int \int y_{\ell,m}(x) \overline{y_{\ell',m'}(z)} |{\langle x | z \rangle}|^{2n} dxdz\\
& =
\int \int y_{\ell,m}(x) \overline{y_{\ell',m'}(z)} \\
& \qquad \times \frac{1}{\dim(n, d)^2} \sum_{\ell''} {\left[\begin{array}{ccc} \nu & \ \nu^* & \lambda'' \\ 0 &
0 & 0 \end{array}\right]} \sum_{m''} \overline{y_{\ell'', m''}(x)}y_{\ell'', m''}(z) dxdz\\
& = \delta_{\ell, \ell'} \delta_{m, m'} \frac{1}{\dim(n, d)^2} {\left[\begin{array}{ccc} \nu & \ \nu^* & \lambda'' \\ 0 &
0 & 0 \end{array}\right]} \ , \end{aligned}$$ where we have used Corollary \[cor:productexpansion2\] (Appendix B) in the second equality sign and the orthonormality of the $y$ functions in the third equality sign. Since $ {\mathrm{mult}}(\ell, n, d)$ is nonzero for $\ell\leq n $ and vanishes for $\ell>n$ (see Corollary \[cor:productexpansion2\]), this concludes the proof.
By Theorem \[thm:Qrep\], ${\langle x |}^{\otimes n} A {| x \rangle}^{\otimes n} = 0$ for all $x$ implies $A=0$. Therefore the operator space $E$ from Lemma \[KlaSkalem\] contains only the identically vanishing operator. As a consequence, the space $D$ of operators that have a P-representation equals the space of operators on ${\mathrm{Sym}}^n({\mathbb{C}}^d)$ which proves the first part of the claim.
The Fourier decomposition of $P_B$ (see Appendix B) $$P_B(x)= \sum_{\ell, m} p_B(\ell, m) y_{\ell, m}(x)$$ implies the decomposition $$B= \sum_{\ell, m} p_B(\ell, m)\left( \int dx \ y_{\ell, m}(x) {{| x \rangle}\!{\langle x |}}^{\otimes n} \right).$$ From Lemma \[lem:onbasis\] we see that the coefficients $p_B(\ell, m)$ are determined by the operator $B$ for $\ell \leq n$ and are arbitrary for $\ell>n$.
B $\quad$ Spherical Harmonics for Higher Dimensions
===================================================
As we have seen, the functions on ${\mathcal{S}}({\mathcal{H}})$, ${\mathcal{P}}({\mathcal{H}})$ and therefore ${\mathbb{C}P^{d-1}}$ play a central role in the present work. In this section, we perform a Fourier decomposition of the functions defined on ${\mathbb{C}P^{d-1}}$ and derive properties that will enable us to work with these functions very effectively. Whereas this may be considered standard by some readers, we include it for the benefit of completeness. Our construction of an orthonormal basis of functions on ${\mathbb{C}P^{d-1}}$ uses the representation theory of $U(d)$ and its subgroup $U(d-1)\times U(1)$. As general references on representation theory of the unitary group we recommend [@FultonHarris91; @CarterSegalMacDonald95]. A (complex) representation $V$ of a group $G$ is a finite-dimensional complex vector space $V$, equipped with an action of $G$ preserving the group operation. $V$ is irreducible if the only invariant subspaces of $V$ are the empty subspace and $V$ itself. For $H$ a subgroup of a group $G$, let $V \downarrow^{G}_{H}$ denote the restriction of a representation of $G$ to $H$.
Let $G=U(d)$, $V$ a (holomorphic) representation of $U(d)$, i.e. a representation whose representing matrices have entries that are holomorphic functions in the variables of $U(d)$, and let $H=T(d)$ be the torus of diagonal matrices in $U(d)$. $V$ decomposes according to $$V\downarrow^{U(d)}_{T(d)}\cong \bigoplus_w W_w,$$ where $W_w$ are the isotypic components of the irreducible representations of $T(d)$. Since $T(d)$ is abelian its irreducible representations are one-dimensional. Vectors in $W_w$ are known as *weight vectors* with *weight* $w=(w_1, \ldots, w_d)$, $w_i \in {\mathbb{Z}}$, that is, for all ${| v \rangle} \in W_w$: $$T(d) \ni t : {| v \rangle} \mapsto t {| v \rangle}=t_1^{w_1} \cdots t_d^{w_d} {| v \rangle},$$ where $t={\text{diag}}(t_1, \ldots, t_d)$. The $w$ with $\dim W_w>0$ are called *weights* of $V$. The lexicographical ordering on the set of weights is the relation $w> w'$ if for the smallest $i$ with $w_i\neq w'_i$, $w_i> w'_i$. It turns out that every irreducible representation $V$ of $U(d)$ has a unique highest weight $\lambda$ satisfying $\dim W_\lambda=1$. $\lambda$ is furthermore *dominant*, i.e. $\lambda=(\lambda_1, \ldots, \lambda_d)$ satisfies $\lambda_i \geq \lambda_{i+1}$. To every dominant $\lambda$, there also exists an irreducible representation denoted by $V_\lambda$. Two irreducible representations $V_\lambda$ and $V_{\lambda'}$ are equivalent if and only if $\lambda =\lambda'$.
In the case where $G = U(d)$ and $H = U(d-1)$ (embedded as $H\ni h \mapsto \left(\begin{array}{cc} h & 0 \\ 0 & 1 \end{array}\right)\in G$ and using the definition ${| i \rangle}=(\underbrace{0, \ldots, 0, 1}_{i}, 0, \ldots, 0)^T$ for all $i$) and where the representation of $U(d)$ is irreducible with highest weight $\lambda$ one has the following decomposition, known as the *branching rule* for $U(d)$: $$\begin{aligned}
\label{eq:branching}
V_\lambda \downarrow^{U(d)}_{U(d-1)}
\cong
\bigoplus_{\mu} V_{\mu}
\end{aligned}$$ where the sum extends over dominant weights $\mu=(\mu_1, \ldots, \mu_{d-1})$ that are *interlaced* by $\lambda$, i.e. that satisfy $$\lambda_{i+1} \leq \mu_i \leq \lambda_i
\quad \forall i \in \{1, \ldots, d-1\} \ .
\label{betcond}$$ Iteratively using the branching rule allows us to define an orthonormal basis of the representation $V_\lambda$, called *Gelfand-Zetlin basis*, where any basis vector is labeled by a sequence of $d$ diagrams $\underbrace{\lambda^{(d)}}_{=\lambda}, \underbrace{\lambda^{(d-1)}, \ldots, \lambda^{(1)}}_{=:m}$ such that $\lambda^{(i+1)}$ is interlaced by $\lambda^{(i)}$. $m$ is called a *Gelfand-Zetlin* pattern for $\lambda$. The state with Gelfand-Zetlin pattern $m=( (0^{d-1}), (0^{d-2}), \ldots, (0) )$, where $(0^i)=(\underbrace{0, \ldots, 0}_i)$, will be abbreviated by $m=0$. The corresponding state is ${| \lambda, 0 \rangle}$.
We denote by $dg$ the volume element of the Haar measure on $U(d)$ with normalisation $\int dg =1$. We now consider the Hilbert space of square integrable functions on $U(d)$ with the inner product $$\int \overline{\alpha(g)}\beta(g) dg,$$ for two functions $\alpha(g)$ and $\beta(g)$. $L^2(U(d))$ carries a representation of $U(d)\times U(d)$ when equipped with the action $$U(d)\times U(d) \ni (g_1, g_2): \alpha(g) \mapsto \alpha(g_1^{-1} g g_2).$$ Let $$t_{\lambda,m,m'}(g) :=d_{\lambda}{\langle \lambda,m |} g {| \lambda,m' \rangle}$$ be the characteristic (or *representative*) functions, i.e. the matrix elements of the irreducible representations of $U(d)$ (multiplied by $d_{\lambda}:=\dim V_\lambda$). Note that these functions are orthonormal with respect to the above defined inner product and — for fixed $\lambda$ — span an irreducible representation of $U(d)\times U(d)$ with a pair of highest weights $(\lambda^*, \lambda)$, where the $ \lambda^*$ denotes the highest weight of the $V_\lambda^*$, the representation dual to $V_\lambda$. It is not difficult to check that $\lambda^*=(-\lambda_d, \ldots, -\lambda_1)$. The Peter-Weyl theorem asserts that these functions are dense in the $L^2(U(d))$. Note that one can interpret this theorem as a Fourier theorem on $U(d)$ as it shows that any square integrable function can be expressed as a linear combination of the basis functions $t_{{\lambda}, m, m'}$. In the following we want to derive a similar statement for functions on ${\mathbb{C}P^{d-1}}$.
When a group $G$ acts transitively on a set $X$ one can identify $X$ with the set $G \slash H$ of left-cosets of the stabiliser group $H$ of a point $x_0 \in X$, i.e., the group $H := \{g \in G: g x_0 = x_0\}$ and the isomorphism $G \slash H \rightarrow X$ is $g H \mapsto g x_0$. \[See [@CarterSegalMacDonald95 p. 59]\]. In the following we consider the transitive action of $U(d)$ on ${\mathbb{C}P^{d-1}}$ and let $x_{0}$ be the point with homogeneous coordinates $[0: \cdots: 0: 1]$. Then $H= U(d-1)\times U(1)$ and ${\mathbb{C}P^{d-1}} \cong U(d) / [ U(d-1)\times U(1)]$, [@VilKlimVol2 p. 278].
We will show below that the vectors ${| \lambda, 0 \rangle}$ for $$\begin{aligned}
\label{eq:special}
\lambda = (\ell, 0, \ldots, 0, -\ell)\end{aligned}$$ are exactly the ones being stabilized by $U(d-1)\times U(1)$. For such $\lambda$, we can therefore define the functions $y_{\ell, m}$ on ${\mathbb{C}P^{d-1}}$ by $$y_{\ell, m}(x) := t_{\lambda, m, 0}(g)$$ for $g \in x$. The index $\ell$ is sometimes called a *moment*. Since the measure $dg$ on $U(d)$ descends to a measure $dx$ on ${\mathbb{C}P^{d-1}}$, these functions are also square integrable and orthonormal with respect to the standard inner product. The following theorem, the main statement of this section, asserts that these functions span $L^2({\mathbb{C}P^{d-1}})$ densely.
\[thm:decomposition\] Let $\mu \in L^2({\mathbb{C}P^{d-1}})$. Then $$\mu(x) = \sum_{\ell \in {{\mathbb{N}}}} \sum_{m} \mu(\ell, m) y_{\ell,m}(x)$$ where the second sum ranges over Gelfand-Zetlin pattern $m$ associated to the irreducible representation $(\ell, 0, \cdots, 0, -\ell)$ of $U(d)$. The constants $\mu(\ell, m)$ are square summable.
The proof is based on the following extension of the Peter-Weyl theorem. Define $$V_\lambda^H
:=
\{v \in V_\lambda: \, h {| v \rangle} = {| v \rangle}, \, \forall h \in H\}$$ as the $H$ invariant subspace of $V_{\lambda}$.
\[thm:PeterWeyl\] $$L^2(G/H) \cong \bigoplus_\lambda^\wedge V_{\lambda}^* \otimes {V_{\lambda}}^H$$ where ${\displaystyle \bigoplus^\wedge}$ is the completion of the direct sum. A basis for $V_{\lambda}^* \otimes V_{\lambda}$ is given by $t_{\lambda,m,m'}$.
See e.g. [@CarterSegalMacDonald95 Corollary 9.14].
The following lemma characterises the components in the direct sum in terms of the functions $y_{\ell, m}$.
\[lem:Uinvariant\] We have $$V_\lambda^* \otimes {V_\lambda}^{U(d-1) \times U(1)}
=
\begin{cases}
{\mathrm{span}}\{ y_{\ell, m}\}& \lambda = (\ell, 0, 0, \ldots, 0, -\ell) \\
0 & \text{otherwise} \ .
\end{cases}$$
Since $t_{\lambda, m, m'}(g)=d_{\lambda}{\langle \lambda, m |} g {| \lambda, m' \rangle}$, it suffices to show that the vectors ${| \lambda, 0 \rangle}$ with $\lambda$ as in the statement are exactly the vectors fixed by $U(d-1)\times U(1)$. The claim is therefore equivalent to $${V_\lambda}^{U(d-1) \times U(1)}
=
\begin{cases}
{\mathrm{span}}\{ {| \lambda, 0 \rangle}\}& \lambda = (\ell, 0, 0, \ldots, 0, -\ell) \\
0 & \text{otherwise} \ .
\end{cases}$$
It follows from the branching rule and simple counting of degrees of the polynomials that $$V_\lambda \downarrow^{U(d)}_{U(d-1)\times U(1)}\cong \bigoplus_\mu V_\mu \otimes V_{|\lambda|-|\mu|}$$ where the sum extends over $\mu$ that are interlaced by $\lambda$ and $V_{|\lambda|-|\mu|}$ is the one-dimensional representation of $U(1)$ with weight $|\lambda|-|\mu|$. Since $V_\lambda^H = (V_\lambda \downarrow_H^G)^H$, we find $$V_\lambda^{U(d-1)\times U(1)}= \left( V_\lambda \downarrow^{U(d)}_{U(d-1)\times U(1)}\right)^{U(d-1)\times U(1)}\cong \bigoplus_\mu V_\mu^{U(d-1)} \otimes V_{|\lambda|-|\mu|}^{U(1)}.$$ $V_\mu^{U(d-1)}$ is exactly nonzero when $V_\mu$ is the trivial representation, i.e. $\mu=(0^{d-1})$. Likewise, $V_{|\lambda|-|\mu|}^{U(1)}$ is non-vanishing only when $V_{|\lambda|-|\mu|}$ is the trivial representation of $U(1)$, i.e. $|\lambda|-|\mu|=0$. Note that $(0)^{d-1}$ interlaces $\lambda$ only when $\lambda=(\lambda_1, 0, \ldots, 0, \lambda_{d})$ and that $|\lambda|=|\mu|$ furthermore implies $\lambda_1+\lambda_d=0$. Setting ${\lambda}_1=\ell$ completes the proof.
We apply Theorem \[thm:PeterWeyl\] to $G=U(d)$ and $H=U(d-1)\times U(1)$. Recalling that in this case $G/H\cong {\mathbb{C}P^{d-1}}$ the left hand side becomes $L^2({\mathbb{C}P^{d-1}})$. According to Lemma \[lem:Uinvariant\], the right hand side equals the space spanned by the functions $y_{\ell, m}$. This concludes the proof.
We now want to relate the multiplication of the $t$ functions to the Clebsch-Gordan coefficients of $U(d)$. In general we have the decomposition $$V_\lambda \otimes V_{\lambda'}\downarrow^{U(d)\times U(d)}_{U(d)}\cong \bigoplus_{\lambda''} {\mathbb{C}}^{c_{\lambda, \lambda'}^{\lambda''}} \otimes V_{\lambda''},$$ where $U(d)$ is embedded diagonally into $U(d)\times U(d)$, i.e. $U(d) \ni g \mapsto g\times g \in U(d)\times U(d)$. The multiplicities $c_{{\lambda}, {\lambda}'}^{{\lambda}''}$ are the well-known Littlewood-Richardson coefficients (see e.g. [@fulton97]). In terms of a basis transform this isomorphism reads $${| {\lambda}, m \rangle} {| {\lambda}', m' \rangle}=\sum_{{\lambda}'', m'', r} {\langle {\lambda}, {\lambda}', r, {\lambda}'', m'' | {\lambda}, m \rangle}{| {\lambda}', m' \rangle}{| {\lambda}, {\lambda}', r, {\lambda}'', m'' \rangle}$$ with the $U(d)$ Clebsch-Gordan coefficients ${\langle {\lambda}, {\lambda}', r, {\lambda}'', m'' | {\lambda}, m \rangle}{| {\lambda}', m' \rangle}$, where $r$ counts the different copies of $V_{{\lambda}''}$. The following lemma relates the product of two functions $t_{\lambda, m, 0}$ and $t_{\lambda', m',0}$ to the $U(d)$ Clebsch-Gordan coefficients. More generally, such a formula can be derived for the product of $t_{\lambda, m, \tilde{m}}$ and $t_{\lambda, m', \tilde{m}'}$ functions (see [@VilKlimVol3 Chapter 18.2.1]).
\[lem:productexpansion\] $$\begin{aligned}
t_{{\lambda}, m, 0}(g)t_{{\lambda}', m', 0}(g) =\sum_{{\lambda}'', m''} {\left[\begin{array}{ccc} {\lambda}& \ {\lambda}' & {\lambda}'' \\ m &
m' & m'' \end{array}\right]}t_{{\lambda}'', m'', 0}(g)\end{aligned}$$ where $$\begin{aligned}
\label{eq:symbol}
&{\left[\begin{array}{ccc} {\lambda}& \ {\lambda}' & {\lambda}'' \\ m &
m' & m'' \end{array}\right]} := \int t_{{\lambda}, m, 0}(g)t_{{\lambda}', m', 0}(g)\overline{t_{{\lambda}'', m'', 0}(g)}dg\\
&\qquad =\frac{d_{\lambda}d_{{\lambda}'}}{d_{{\lambda}''}}\left(\sum_{r}{\langle {\lambda}, {\lambda}', r, {\lambda}'', 0 | {\lambda}, 0 \rangle}{| {\lambda}', 0 \rangle}{\langle {\lambda}, m |}{\langle {\lambda}', m' | {\lambda}, {\lambda}', r, {\lambda}'', m'' \rangle} \right). \nonumber\end{aligned}$$
$$\begin{aligned}
t_{{\lambda}, m, 0}(g)t_{{\lambda}', m', 0}(g) & = d_{\lambda}d_{{\lambda}'} {\langle {\lambda}, m |} g {| {\lambda}, 0 \rangle} {\langle {\lambda}', m' |} g {| {\lambda}', 0 \rangle} \\
& = d_{\lambda}d_{{\lambda}'} ({\langle {\lambda}, m |} {\langle {\lambda}', m' |}) g ({| {\lambda}, 0 \rangle} {| {\lambda}', 0 \rangle}) \\
& =d_{\lambda}d_{{\lambda}'}\sum_{{\lambda}'', m'', r}{\langle {\lambda}, {\lambda}', r, {\lambda}'', 0 | {\lambda}, 0 \rangle}{| {\lambda}',0 \rangle}\\
& \qquad \qquad \times {\langle {\lambda}, m |}{\langle {\lambda}', m' | {\lambda}, {\lambda}', r, {\lambda}'', m'' \rangle} {\langle {\lambda}'', m'' |} g {| {\lambda}'', 0 \rangle} \\
& =\sum_{{\lambda}'', m''} {\left[\begin{array}{ccc} {\lambda}& \ {\lambda}' & {\lambda}'' \\ m &
m' & m'' \end{array}\right]}t_{{\lambda}'', m'', 0}(g)\end{aligned}$$
This leads to an important product formula which we use, for instance, in the update rule.
\[cor:productexpansion\] $$\begin{aligned}
y_{\ell, m}(x)y_{\ell', m'}(x) =\sum_{\ell''}^{\ell +\ell'} \sum_{m''} {\left\{\begin{array}{ccc} \ell & \ \ell' & \ell'' \\ m & m' & m'' \end{array}\right\}}y_{\ell'', m''}(x)\end{aligned}$$ where $ {\left\{\begin{array}{ccc} \ell & \ \ell' & \ell'' \\ m & m' & m'' \end{array}\right\}}:= {\left[\begin{array}{ccc} {\lambda}& \ {\lambda}' & {\lambda}'' \\ m &
m' & m'' \end{array}\right]} $ for ${\lambda}=(\ell, 0, \ldots, 0, -\ell)$ and similarly for ${\lambda}'$ and ${\lambda}''$ (see ).
\[cor:productexpansion2\] $$\begin{aligned}
\dim(n, d)|{\langle d | x \rangle}|^{2n}=\frac{1}{\dim(n, d)}\sum_{\ell} {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} y_{\ell, 0}(x).\end{aligned}$$ where $$\begin{aligned}
\label{eq:CGmult}
{\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} =\frac{\dim(n, d)^2}{d_{{\lambda}}}|{\langle \nu, \nu^*, {\lambda}, 0 | \nu, 0 \rangle}{| \nu^*, 0 \rangle}|^2 \end{aligned}$$ for ${\lambda}=(\ell, 0, \ldots, 0, -\ell)$ with $\ell\leq n$. Furthermore, ${\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} \neq 0$ for $\ell \leq n$ and vanishes for $\ell >n$.
Note that $\dim(n, d)=d_\nu=d_{\nu^*}$. Then $$\begin{aligned}
d_{\nu}^2|{\langle d | x \rangle}|^{2n}& = d_{\nu}^2{\langle d |}^{\otimes n} g^{\otimes n} {| d \rangle}^{\otimes n} \overline{{\langle d |}^{\otimes n} g^{\otimes n} {| d \rangle}^{\otimes n}}\\
& = d_\nu^2 {\langle \nu, 0 |} g {| \nu, 0 \rangle}\overline{{\langle \nu, 0 |} g {| \nu, 0 \rangle}}\\
& = t_{\nu, 0, 0} (x) \overline{t_{\nu, 0, 0} (x)} = t_{\nu, 0, 0} (x) t_{\nu^*, 0, 0} (x) \end{aligned}$$ since $ {| d \rangle}^{\otimes n}$ is a weight vector in $\nu=(n, 0, \ldots,
0)$ that is invariant with respect to the subgroup $U(d-1)$ (embedded into $U(d)$ by inclusion into the top left corner), and therefore has a Gelfand-Zetlin pattern $m=0$. The invariance with respect $U(d-1)$ follows from the tensor production action of the group $U(d)$ as well as the fact that the stabilizer of ${| d \rangle}$ contains $U(d-1)$. holds since the Littlewood-Richardson coefficient $c_{\nu, \nu^*}^{\lambda}$ equals one for $\ell\leq n$ and vanishes for larger values of $\ell$ [@fulton97]. This implies in particular that $ {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]}$ vanishes for $\ell>n$. In order to see that $ {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]}$ does not vanish for smaller values of $\ell$, note that the projection of ${| \nu, 0 \rangle}{| \nu^*, 0 \rangle}$ onto the irreducible representation $\lambda=(\ell, 0, \ldots, 0, -\ell)$ is given by $$\begin{aligned}
P_\lambda {| \nu, 0 \rangle}{| \nu^*, 0 \rangle}&=\sum_m ({\langle \nu, \nu^*, \lambda, m | \nu, 0 \rangle}{| \nu^*, 0 \rangle}){| \nu, \nu^*, \lambda, m \rangle}\\
&= ({\langle \nu, \nu^*, \lambda, 0 | \nu, 0 \rangle}{| \nu^*, 0 \rangle}){| \nu, \nu^*, \lambda, 0 \rangle}\end{aligned}$$ where we used the fact that the sum can only contain $H$ invariant vectors. The claim follows since we know that this projection cannot vanish, as the Littlewood-Richardson coefficient is nonzero for all $\ell \leq n$.
\[lem:multiply\] For $g \in zH$, where $H=U(d-1)\times U(1)$, we have $$\begin{aligned}
y_{\ell, 0}(g^\dagger x)= \sum_m \overline{y_{\ell, m}(z)} y_{\ell, m}(x)\end{aligned}$$
$$\begin{aligned}
\int y_{\ell, 0}(g^\dagger x) \overline{y_{\ell, m}(x)}dx & = d_{\lambda}d_{{\lambda}'}\int {\langle {\lambda}, 0 |} g^{\dagger} \tilde{g}{| {\lambda}, 0 \rangle} {\langle {\lambda}', 0 |} \tilde{g}^\dagger {| {\lambda}', m \rangle}d\tilde{g}\\
& = d_{\lambda}\delta_{{\lambda}, {\lambda}'} {\langle {\lambda}, 0 |} g^\dagger {| {\lambda}, m \rangle}= \delta_{{\lambda}, {\lambda}'} \overline{y_{\ell, m}(z)}\end{aligned}$$
The next corollary generalises the decomposition in Corollary \[cor:productexpansion2\]. It is needed in some technical aspects of the paper as well as the examples.
\[cor:productexpansion3\] $$\begin{aligned}
\dim(n, d) |{\langle z | x \rangle}|^{2n}& =\frac{1}{\dim(n, d)} \sum_{\ell}^n {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} \sum_m \overline{y_{\ell, m}(z)}y_{\ell, m}(x), \end{aligned}$$ where $\nu=(n, 0, \ldots, 0)$ and ${\lambda}=(\ell, 0, \ldots, 0, -\ell)$. The coefficients are defined in .
By Corollary \[cor:productexpansion2\] and we have $$\begin{aligned}
d_\nu |{\langle z | x \rangle}|^{2n}& = \frac{1}{d_{\nu}}\sum_{\ell}^n {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} y_{\ell, 0}(g^\dagger x) \nonumber \\
& =\frac{1}{d_{\nu}} \sum_{\ell}^n {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} \sum_m \overline{y_{\ell, m}(z)}y_{\ell, m}(x) \end{aligned}$$ where $g \in zH$ (${| x \rangle}=g {| d \rangle}$)and where we used Lemma \[lem:multiply\] in the last equation.
C $\quad$ Recovering the Spherical Harmonics on the Bloch Sphere
================================================================
In the following we restrict our attention to the special case $d=2$. The complex projective space ${\mathbb{C}P^{1}}$ can be viewed as the sphere $S^2$ with $x \in {\mathbb{C}P^{1}}$ being represented as a point on the sphere parameterised by angles $\theta \in [0, \pi]$ and $\phi \in [0, 2\pi)$. The measure $dx$ turns into $\frac{1}{4\pi} \sin \theta d\theta d\phi$. As a unitary representative for $x$, $g \in xH$, we choose $$\begin{aligned}
\label{eq:g}
g=e^{i \frac{\phi}{2} \sigma_z} e^{i \frac{\theta}{2} \sigma_x} =\begin{pmatrix} e^{i\frac{\phi}{2}} & 0 \\ 0 &e^{-i\frac{\phi}{2}} \end{pmatrix} \begin{pmatrix} \cos \frac{\theta}{2} & i \sin \frac{\theta}{2} \\ i \sin \frac{\theta}{2} &\cos \frac{\theta}{2} \end{pmatrix} .\end{aligned}$$ This implies ${| x \rangle}=g {| 2 \rangle}=i e^{i\frac{\phi}{2}}\sin \frac{\theta}{2} {| 1 \rangle} + e^{-i\frac{\phi}{2}}\cos \frac{\theta}{2} {| 2 \rangle} $. where we used $${| 1 \rangle}=\begin{pmatrix} 1 \\ 0 \end{pmatrix}\qquad {| 2 \rangle}=\begin{pmatrix} 0 \\ 1 \end{pmatrix}$$ We may think of $\theta=0$ as the south pole and $\theta=\pi$ as the north pole of the sphere (when the $z$ direction is the rotation axis of the earth).
It is then natural to expect that the $y_{\ell, m}$ are related to the ordinary spherical harmonics on the sphere. The next lemma provides us with the precise dependence. Thereafter we will find a formula for the coefficients ${\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ * &
* & * \end{array}\right]}$ that govern the multiplication of two functions.
Before we start, note that for $d=2$ the possible Gelfand-Zetlin patterns $m$ for $\lambda=(\ell, -\ell)$ lie in the interval $-\ell
\leq m \leq \ell$ and that the spin projection in $z$-direction (i.e. the eigenvalue of the (Lie algebra) representation of the operator $\frac{1}{2}\sigma_z$) of the state ${| \lambda, m \rangle}$ equals $m$ (cf. Lemma \[lem:SU2\]).
For $d=2$ $$\begin{aligned}
y_{\ell, m}(x)&=(-i)^m \sqrt{2 \ell+1} \sqrt{\frac{(\ell-m)!}{(\ell+m)!}}P^m_\ell(\cos \theta) e^{im\phi}=(-i)^m\sqrt{4\pi} Y_{\ell}^{m}(\theta, \phi)\end{aligned}$$ where $P_\ell^m$ are the associated Legendre polynomials and $$Y_{\ell}^{m}(\theta, \phi):= \sqrt{\frac{2 \ell+1}{4\pi}} \sqrt{\frac{(\ell-m)!}{(\ell+m)!}}P^m_\ell(\cos \theta) e^{im\phi}$$ are the spherical harmonics [^5].
By [@VilKlimVol1 eq (1) in Chapter 6.3.1] $${\langle \ell, m |} g {| \ell, 0 \rangle}= {\bf t}_{-m, 0}(g).$$ Using and [@VilKlimVol1 eq (2) & (4) in Chapter 6.3.3 and eq. (3) in Chapter 6.3.7.] we see that for positive $m$ $$e^{i m \phi} (-i)^m\sqrt{ \frac{(\ell-m)!}{(\ell+m)!}}P^m_\ell(\cos \theta)$$ holds, where $P^m_\ell$ denote the associated Legendre polynomials. The same formula can be seen to hold for negative $m$ by use of [@VilKlimVol1 eq (2’) in Chapter 6.3.6 and eq (3’) in Chapter 6.3.7].
We will now find formulae for the ${\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ * &
* & * \end{array}\right]}$ by relating them to the Clebsch-Gordan coefficients of $SU(2)$ for which closed formulae are known. We start by relating the Clebsch-Gordan coefficients of $SU(2)$ and $U(2)$.
\[lem:SU2\] Let $d=2$. If $c_{{\lambda}, {\lambda}'}^{{\lambda}''} \neq 0$ then $${\langle {\lambda}, {\lambda}', {\lambda}'', m'' | {\lambda}, m \rangle}{| {\lambda}', m' \rangle}_{U(2) }= {\langle L, L', L'', M'' | L, M \rangle}{| L', M' \rangle}_{SU(2)}$$ where $${\lambda}=({\lambda}_1, {\lambda}_2) \qquad \qquad L=\frac{{\lambda}_1-{\lambda}_2}{2} \qquad \qquad M=m-\frac{{\lambda}_1+{\lambda}_2}{2}$$ and likewise for the primed variables.
If $V_{{\lambda}''} \subset V_{{\lambda}} \otimes V_{{\lambda}'}$, then $$V_{{\lambda}''}\downarrow^{U(2)}_{SU(2)} \subset V_{{\lambda}}\downarrow^{U(2)}_{SU(2)} \otimes V_{{\lambda}'}\downarrow^{U(2)}_{SU(2)},$$ Since $V_{{\lambda}}\downarrow^{U(2)}_{SU(2)}$ is equivalent to a $\mathrm{spin-} L$ representation (with $L$ as in the claim) we can obtain the Clebsch-Gordan coefficients for $U(2)$ from those of $SU(2)$. This works as follows. The mapping of the basis state in the irreducible representation ${\lambda}= ({\lambda}_1, {\lambda}_2)$ with Gelfand-Zetlin pattern $(m)$ is $${| {\lambda}, m \rangle}_{U(2)} \rightarrow {| L, M \rangle}_{SU(2)},$$ where $L$ and $M$ are defined as in the statement of the claim since the weight of a Gelfand-Zetlin pattern $m$ in a representation ${\lambda}$ equals $(w_1, w_2)=(m, {\lambda}_1+{\lambda}_2-m)$ and the spin projection $M$ along the $z$-direction equals $\frac{w_1-w_2}{2}$. This concludes the proof.
\[lem:qubitCG\] Let $d=2$ and ${\lambda}=(\ell, -\ell)$, $\nu=(n, 0)$ and $\ell \leq n$. If $\ell$ is even, then $$\begin{aligned}
\label{eq:deltafourier}
{\langle \nu, \nu^*, {\lambda}, 0 | \nu, 0 \rangle}{| \nu^*, 0 \rangle}_{U(2) }= \frac{\sqrt{2\ell+1}}{\sqrt{n+\ell+1}}\frac{n!}{\sqrt{(n-\ell)!(n+\ell)!}} \end{aligned}$$ and zero otherwise. If $n$ and $\ell$ are even $${\langle \nu, \nu^*, {\lambda}, 0 | \nu, \frac{n}{2} \rangle}{| \nu^*, -\frac{n}{2} \rangle}_{U(2)}=\frac{(-1)^{\frac{n-\ell}{2}} }{\sqrt{n+\ell+1}}\frac{ \frac{n+\ell}{2}!\ell!\sqrt{(n-\ell)!}}{\frac{n-\ell}{2}!\frac{\ell}{2}!^2\sqrt{(n+\ell)!}}$$ and zero otherwise.
By Lemma \[lem:SU2\] $${\langle \nu, \nu^*, {\lambda}, 0 | \nu, 0 \rangle}{| \nu^*, 0 \rangle}={\langle \frac{n}{2}, \frac{n}{2}, \ell, 0 | \frac{n}{2}, -\frac{n}{2} \rangle}{| \frac{n}{2}, \frac{n}{2} \rangle}.$$ Using the formula [@VilKlimVol1 eq (4) in Chapter 8.2.4] $${\langle \ell, \ell', \ell'', \ell-\ell' | \ell, \ell \rangle}{| \ell', -\ell' \rangle}_{SU(2)}=\sqrt{\frac{(2\ell''+1)(2\ell)!(2\ell')!}{(\ell+\ell'-\ell'')!(\ell+\ell'+\ell''+1)!}}$$ we find $${\langle \frac{n}{2}, \frac{n}{2}, \ell, 0 | \frac{n}{2}, \frac{n}{2} \rangle}{| \frac{n}{2}, -\frac{n}{2} \rangle}= \frac{\sqrt{2\ell+1}}{\sqrt{n+\ell+1}}\frac{n!}{\sqrt{(n-\ell)!(n+\ell)!}}$$ This concludes the proof of . By Lemma \[lem:SU2\] $${\langle \nu, \nu^*, {\lambda}, 0 | \nu, \frac{n}{2} \rangle}{| \nu^*, -\frac{n}{2} \rangle}_{U(2)}={\langle \frac{n}{2}, \frac{n}{2}, \ell, 0 | \frac{n}{2}, 0 \rangle}{| \frac{n}{2}, 0 \rangle}_{SU(2)}.$$ Using the formula [@VilKlimVol1 eq (8) in Chapter 8.2.6] $$\begin{aligned}
{\langle \ell, \ell', \ell'', 0 | \ell, 0 \rangle}{| \ell', 0 \rangle} =\frac{(-1)^{g-\ell''}g!\Delta(\ell, \ell', \ell'')\sqrt{2\ell''+1}}{(g-\ell)!(g-\ell')!(g-\ell'')!}\end{aligned}$$ where $2g:=\ell+\ell'+\ell''$ is even (the coefficient vanishes for odd $2g$) and (see [@VilKlimVol1 eq (3) in Chapter 8.1.3] $$\Delta(\ell, \ell', \ell'')= \sqrt{\frac{(\ell+\ell'-\ell'')!(\ell-\ell'+\ell'')!(\ell'-\ell+\ell'')!}{(\ell+\ell'+\ell''+1)!}}.$$ If $\ell=\ell'$ the coefficient vanishes unless $\ell''$ is even. If $n$ and $\ell$ are even we find $${\langle \frac{n}{2}, \frac{n}{2}, \ell, 0 | \frac{n}{2}, 0 \rangle}{| \frac{n}{2}, 0 \rangle}_{SU(2)}= \frac{(-1)^{\frac{n-\ell}{2}} }{\sqrt{n+\ell+1}}\frac{ \frac{n+\ell}{2}!\ell!\sqrt{(n-\ell)!}}{\frac{n-\ell}{2}!\frac{\ell}{2}!^2\sqrt{(n+\ell)!}}.$$ Otherwise the coefficient vanishes.
\[cor:clebsch-gordon-estimate2\] For ${\lambda}=(\ell, -\ell)$ and $\nu=(n, 0)$ and $\ell$ and $n$ even we have $$\begin{aligned}
\frac{1}{d_\nu}{\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ \frac{n}{2} &
-\frac{n}{2} & 0 \end{array}\right]}&=(-1)^{n-\frac{\ell}{2}} (\frac{1}{2})^{\ell} \binom{\ell}{\frac{\ell}{2}} \prod_{i=1}^{\ell}\left( 1-\frac{\ell-i}{n+2+i}\right)\end{aligned}$$ $$\begin{aligned}
\frac{1}{d_\nu}{\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]}&=\frac{n!(n+1)!}{(n-\ell)!(n+\ell+1)!}\end{aligned}$$
By Lemma \[lem:qubitCG\] we have $$\begin{aligned}
(-1)^{\frac{n-\ell}{2}} &
\frac{\dim V_\nu}{\dim V_\lambda}{\langle \nu, \nu^*, {\lambda}, 0 | \nu, 0 \rangle}{| \nu^*, 0 \rangle}{\langle \nu, \frac{n}{2} |}{\langle \nu^*, -\frac{n}{2} | \nu, \nu^*, {\lambda}, 0 \rangle}\\
&=\frac{(\frac{n+\ell}{2})!\ell! (n+1)!}{(\frac{n-\ell}{2})!(\frac{\ell}{2})!^2 (n+1+\ell)!}\\
&=\frac{ (\frac{n}{2}-\frac{\ell}{2}+1) \cdots (\frac{n}{2}+\frac{\ell}{2}) \ell!}{(n+2) \cdots (n+\ell+1) (\frac{\ell}{2})!^2}\\
&= (\frac{1}{2})^{\ell} \frac{(n-\ell+2)(n-\ell+4)\cdots (n+\ell) \ell!}{(n+2) \cdots (n+\ell+1) \frac{\ell}{2}!^2}\\
&= (\frac{1}{2})^{\ell} \frac{\ell!}{ \frac{\ell}{2}!^2} (1-\frac{\ell}{n+2}) (1-\frac{\ell-1}{n+3})\cdots (1-\frac{1}{n+2+\ell}),\end{aligned}$$ which proves the first formula. The second formula follows from Corollaries \[lem:SU2\] and \[lem:qubitCG\]. The estimate derives from $$\begin{aligned}
\frac{(n+1)!n!}{(n-\ell)!(n+1+\ell)!}= \frac{(n-\ell+1)\cdots n}{(n+2)\cdots (n+1+\ell)}&\geq \left(\frac{n-\ell+1}{n+2}\right)^{\ell}\geq 1-\frac{\ell(\ell+1)}{n+2}.\end{aligned}$$
Finally, we compute the Fourier coefficients of the distribution that is uniform on the equator of the Bloch sphere.
\[lem:equator\] Let $d=2$ and let $\mu(x)$ be the distribution that is uniformly concentrated on the equator of the Bloch sphere, i.e. $$\int \mu(x) f(x) dx = \frac{1}{2\pi }\int_{[0, 2\pi)} f(x_{\phi}) d\phi$$ for all test functions $f(x)$, where $$x_\phi= \frac{(e^{i\phi/2}{| 1 \rangle}+e^{-i\phi/2}{| 2 \rangle})(e^{-i\phi/2}{\langle 0 |}+e^{-i\phi/2}{\langle 1 |})}{2}$$ are the points on the equator. Then $$\begin{aligned}
\mu(x)= \sum_{\ell} (-1)^{\frac{\ell}{2}}(\frac{1}{2})^{\ell} \binom{\ell}{\frac{\ell}{2}} y_{\ell,0}(x) \end{aligned}$$
Note that $\mu(x)=\mu(hx)$ for all $h\in H=U(1)\times U(1)$, i.e. $\mu$ is $H$ invariant. The non-vanishing Fourier components must therefore also be $H$ invariant, which implies that only the ones where $m= 0$ can be nonzero.
The remaining coefficients are $$\begin{aligned}
\int \mu(x) \overline{y_{\lambda,0}(x)} dx &=\frac{1}{2\pi} \int_\phi d\phi \overline{y_{\lambda, 0}(x_\phi)}\\
&=\frac{1}{2\pi} \int_\phi d\phi {\langle \lambda, 0 |} g_\phi {| \lambda, 0 \rangle}\\
&= {\langle \lambda, 0 |} \left(\begin{array}{cc}\frac{1}{\sqrt{2}} & i\frac{1}{\sqrt{2}} \\ i\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \end{array}\right) {| \lambda, 0 \rangle}\end{aligned}$$ where we chose a representative $g_\phi$ from the coset $x_\phi H$: $$g_\phi=\left(\begin{array}{cc}e^{i \phi/2} & 0 \\ 0 & e^{-i \phi/2} \end{array}\right) \left(\begin{array}{cc}\frac{1}{\sqrt{2}} & i\frac{1}{\sqrt{2}} \\ i\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \end{array}\right)$$ and used ${\langle \lambda, 0 |}= {\langle \lambda, 0 |}\left(\begin{array}{cc}e^{i \phi/2} & 0 \\ 0 & e^{-i \phi/2} \end{array}\right)$ in the last line. The calculation of the last term is a combinatorial feast: $$\begin{aligned}
{\langle \lambda, 0 |}& \left(\begin{array}{cc}\frac{1}{\sqrt{2}} &i \frac{1}{\sqrt{2}} \\ i \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \end{array}\right) {| \lambda, 0 \rangle} ={\langle \lambda, 0 |} \left(\begin{array}{cc}\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ \frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} \end{array}\right) {| \lambda, 0 \rangle} \\
& ={\langle \lambda, 0 |} \begin{pmatrix} e^{-i\pi/4} & 0 \\ 0 & e^{i\pi/4} \end{pmatrix}\left(\begin{array}{cc}\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ \frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} \end{array}\right) \begin{pmatrix} e^{-i\pi/4} & 0 \\ 0 & e^{i\pi/4} \end{pmatrix}{| \lambda, 0 \rangle} \\
& =\frac{1}{\binom{2\ell}{\ell}}\sum_{z, z'} {\langle z |} \left(\begin{array}{cc}\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\\frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} \end{array}\right)^{\otimes 2\ell} {| z' \rangle}\\
&=\sum_{z} {\langle z |} \left(\begin{array}{cc}\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\\frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} \end{array}\right)^{\otimes 2\ell} {| 1 \cdots 1 2 \cdots 2 \rangle}\\
&=\left(\frac{1}{\sqrt{2}}\right)^{2\ell} \sum_{z} {\langle z |} {| 1+2 \rangle}^{\otimes \ell}{| 1-2 \rangle}^{\otimes \ell}\\
&=\frac{1}{2^{\ell}} \frac{1}{\ell!^2}\left( \frac{d}{d\alpha}\right)^{\ell} \left( \frac{d}{d\beta}\right)^{\ell}(\alpha+\beta)^{\ell} (\alpha-\beta)^{\ell}\big|_{\alpha=\beta=0} \\
&=\frac{1}{2^{\ell}} \frac{1}{\ell!^2}\left( \frac{d}{d\alpha}\right)^{\ell} \left( \frac{d}{d\beta}\right)^{\ell}(\alpha^2-\beta^2)^{\ell}\big|_{\alpha=\beta=0} \\
&=\frac{1}{2^{\ell} } \sum_{z} {\langle z |} {| 11-22 \rangle}^{\otimes \ell}\\
&=\frac{1}{2^{\ell}} (-1)^{\frac{\ell}{2}}\binom{\ell}{\frac{\ell}{2}}\end{aligned}$$ if $\ell$ is even. Otherwise the last formula vanishes. The sums over $z$ and $z'$ extend over all binary strings (with symbols $1$ and $2$) of length $ 2\ell$ with Hamming weight $\ell$.
D $\quad$ Examples
==================
Holevo’s Covariant Measurement
------------------------------
It was shown by Holevo that an optimal measurement procedure (in terms of the fidelity) for state estimation is given by the POVM $\{{{| y \rangle}\!{\langle y |}}^{\otimes n}dy\}$ [@Holevo p. 163]. We now want to analyse this measurement with our methods. Let us start by assuming that we have measured the effect ${{| d \rangle}\!{\langle d |}}^{\otimes n}$, giving rise to an estimate density $$\begin{aligned}
\mu_{{{| d \rangle}\!{\langle d |}}^{\otimes n}}(x)& = \dim(n, d)|{\langle x | d \rangle}|^{2n}\end{aligned}$$ We find $$\begin{aligned}
\label{eq:convergence}
\lim_{n\rightarrow \infty}\mu_{\rho^k_{{{| d \rangle}\!{\langle d |}}^{\otimes n}}}(x)=\delta(x),\end{aligned}$$ since $\dim(n, d)|{\langle x | d \rangle}|^{2n}$ converges to the $\delta$-distribution. This, as expected, reflects the fact that the scheme is asymptotically correct. If we measured ${{| z \rangle}\!{\langle z |}}^{\otimes n}$ instead of ${{| d \rangle}\!{\langle d |}}^{\otimes n}$, we find $$\begin{aligned}
\label{eq:optimalmeasurement}
\mu_{{{| z \rangle}\!{\langle z |}}^{\otimes n}}(x)& =\dim(n, d) |{\langle x | z \rangle}|^{2n}\end{aligned}$$ with $$\lim_{n\rightarrow \infty} \mu_{\rho^k_{{{| z \rangle}\!{\langle z |}}^{\otimes n}}}(x)= \delta(z-x).$$
Let us now do the analysis in terms of Fourier coefficients: Comparing with $\delta(x)= \sum_{
\ell} y_{\ell, 0}(x)$ we see that the Fourier coefficients of $\dim(n, d)|{\langle d | x \rangle}|^{2n}$ must all converge to one. Explicitly, the latter are given by (see Corollary \[cor:productexpansion2\]) $$\begin{aligned}
\dim(n, d)|{\langle d | x \rangle}|^{2n}=\frac{1}{\dim(n, d)}\sum_{\ell}^n {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} y_{\ell, 0}(x),\end{aligned}$$ where ${\lambda}=(\ell, 0, \ldots, 0, -\ell)$, $\nu=(n, 0, \ldots, 0)$ and $\nu^*$ denotes highest weight dual to $\nu$. More generally, we have (see Corollary \[cor:productexpansion3\]) $$\begin{aligned}
\dim(n, d) |{\langle z | x \rangle}|^{2n}& =\frac{1}{\dim(n, d)} \sum_{\ell}^n
{\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} \sum_m \overline{y_{\ell, m}(z)}y_{\ell,
m}(x)\ .\end{aligned}$$
We conclude this example with a formula for the Fourier coefficients for the qubit case and derive from it explicit bounds on the convergence the value one (Corollary \[cor:clebsch-gordon-estimate2\]): $$\frac{1}{\dim(n, 2)} {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} =\frac{n!(n+1)!}{(n-\ell)!(n+\ell+1)!},$$ for $\ell\leq n$ and zero otherwise. For small $n$ and $\ell$, we have
----- ----- --------------- ---------------- ----------------- ------------------
$n$ $1$ $2$ $3$ $4$
$l$
$0$ $1$ $1$ $1$ $1$
$1$ $\frac{1}{3}$ $\frac{5}{10}$ $\frac{21}{35}$ $\frac{84}{126}$
$2$ $\frac{1}{10}$ $\frac{7}{35}$ $\frac{36}{126}$
$3$ $\frac{1}{35}$ $\frac{9}{126}$
$4$ $\frac{1}{126}$
----- ----- --------------- ---------------- ----------------- ------------------
and in general there is the bound (Corollary \[cor:clebsch-gordon-estimate2\]) $$1-\frac{\ell(\ell+1)}{n+2}\leq\frac{1}{\dim(n, 2)} {\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ 0 &
0 & 0 \end{array}\right]} \leq 1.$$ This example shows how one may perform a convergence analysis of a tomographic measurement in terms of the Fourier coefficients of the of the estimate density. The convergence of the Fourier coefficients to a constant value (for fixed $\ell$ and $n\rightarrow \infty$) is also in agreement with our intuition about the duality of the Fourier transform: more information about $x$ corresponds to less information about $(\ell, m)$.
Basis Measurements
------------------
We will now analyse the case where a product measurement is carried out with the measurement in a single system given by an orthonormal basis.
Assume that the basis in which we measure is the computational basis $\{{{| i \rangle}\!{\langle i |}}\}_{i}^d$. Since the state we are measuring lives in the symmetric subspace, we can, without loss of generality, project the effect $B_f^n$ onto this subspace and obtain the projector onto the vector with weight $f$ in the representation $V_\nu$. This vector, denoted by ${| \nu, m \rangle}$ is unique and has Gelfand-Zetlin pattern $m=(m^{(d-1)}, \ldots, m^{(1)})$ for $m^{(i)}=(\sum_{j=d+1-i}^d f^{(j)}, 0, \ldots, 0)$. The estimate density is therefore given by $$\mu_{B^n_f}(x)=\dim(n, d)|{\langle x | \nu, m \rangle}|^2,$$ where $\nu=(n, 0, \ldots, 0)$. By Lemma \[lem:productexpansion\] we have $$\mu_{B^n_f}(\ell, m')=\delta_{m', 0} \frac{1}{\dim(n, d)}{\left[\begin{array}{ccc} \nu & \ \nu^* & {\lambda}\\ m &
-m & 0 \end{array}\right]},$$ where ${\lambda}=(\ell, 0, \ldots, 0, -\ell)$.
We will now compute these coefficients for qubits, $d=2$, in the case where we have measured an equal number, namely, $\frac{n}{2}$, 1s and 2s (i.e. the case $m=\frac{n}{2}$). We will use this formula to show that estimate density converges to the uniform distribution on the equator of the Bloch sphere. It follows from Corollary \[cor:clebsch-gordon-estimate2\] that for $\ell$ and $n/2$ even: $$\begin{aligned}
\mu_{B^n_f}(\ell, m)=\delta_{m, 0}(-1)^{\frac{\ell}{2}} (\frac{1}{2})^{\ell} \binom{\ell}{\frac{\ell}{2}} \prod_{i=1}^{\ell}\left( 1-\frac{\ell-i}{n+2+i}\right).\end{aligned}$$ For large $n$, these coefficients turn into $$(-1)^{\frac{\ell}{2}} (\frac{1}{2})^{\ell} \binom{\ell}{\frac{\ell}{2}}$$ which are the Fourier coefficients of the uniform distribution on the equator of the Bloch sphere by Lemma \[lem:equator\]. The estimate density therefore concentrates on the equator just as expected, since we do not obtain any information on the phase of the state from this measurement.
This example shows that Fourier analysis is able to trace a complicated convergence behaviour in a compact way. When several bases are used (such as in the BB84 or the six-state protocols for quantum key distribution) one can use the just derived formula together Lemma \[lem:multiply\] — which allows to rotate the basis — in the update rule.
[1]{}
M. Hayashi. Universal approximation of multi-copy states and universal quantum lossless data compression. , 293:171–183, 2010.
M. Christandl, R. K[ö]{}nig, and R. Renner. Post-selection technique for quantum channels with applications to quantum cryptography. , 102:020504, 2009.
W. Fulton and J. Harris. . Springer, New York, 1991.
K. Życzkowski and H.-J. Sommers. Induced measures in the space of mixed quantum states. , 34:7111–7125, 2001.
M. Christandl, R. König, G. Mitchison, and R. Renner. One-and-a-half quantum de [Finetti]{} theorems. , 273:473–498, 2007.
M. Tomamichel, R. Colbeck, and R. Renner. Duality between smooth min- and max-entropies. , 56:4674 –4681, 2010.
M. Je žek, J. Fiurá šek, and Z. Hradil. Quantum inference of states and processes. , 68:012305, 2003.
W. Matthews, S. Wehner, and A. Winter. Distinguishability of quantum states under restricted families of measurements with an application to quantum data hiding. , 291:813–843, 2009.
, [*Generalized coherent states and their application*]{}, Texts and Monographs in Physics, Springer-Verlag, Berlin, 1986.
, [*Coherent states: applications in physics and mathematical physics*]{}, World Scientific, Singapore, 1985.
, [*Lectures on [L]{}ie Groups and [L]{}ie Algebras*]{}, vol. 32 of London Mathematical Society Student Texts, Cambridge University Press, 1st ed., September 1995.
, [*Representation of [L]{}ie Groups and Special Functions*]{}, vol. 1, Kluwer Academic Publishers Group, Dordrecht, 1991.
height 2pt depth -1.6pt width 23pt, [*Representation of [L]{}ie Groups and Special Functions*]{}, vol. 2, Kluwer Academic Publishers Group, Dordrecht, 1993.
height 2pt depth -1.6pt width 23pt, [*Representation of [L]{}ie Groups and Special Functions*]{}, vol. 3, Kluwer Academic Publishers Group, Dordrecht, 1993.
, [*Young Tableaux*]{}, Cambridge University Press, 1997.
[^1]: A disc with (great-circle) radius $\Delta$ on the Bloch sphere has area $2 \pi (1- \cos \Delta) \leq \pi\Delta^2 $, whereas the full Bloch sphere has area $4 \pi$. Consequently, there are at least $(4
\pi) /(\pi\Delta^2) = 4/\Delta^2$ such discs. Note also that the (great-circle) distance $\Delta$ between two pure states $\phi$ and $\psi$ is related to their fidelity, $F(\phi, \psi) =
|{\langle \phi | \psi \rangle}| = |\cos \frac{\Delta}{2}|$, as well as to their trace distance, $\| {{| \psi \rangle}\!{\langle \psi |}} - {{| \phi \rangle}\!{\langle \phi |}}\|_1 = 2 |\sin
\frac{\Delta}{2}| \approx \Delta$.
[^2]: The bound follows from the fact that the joint state of $n$ identically prepared copies of a pure state in ${\mathcal{H}}=
\mathbb{C}^2$ lies in the symmetric subspace of ${\mathcal{H}}^{\otimes n}$, which has dimension $n+1$.
[^3]: For our technical treatment (see *Supplemental Information*), we also consider tests that act on a larger space, $({{\mathcal{H}}\otimes {\mathcal{K}}})^{\otimes k}$, which includes purifications of the systems.
[^4]: More precisely, the bound on the parameter ${\varepsilon}$, which is usually exponentially decreasing in the size of the confidence region $\Gamma_{\mu_{B^n}}$, is tight up to a polynomial factor.
[^5]: Note that we are using a standard convention also used in Mathematica.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Inhomogeneous strain applied to lattice systems can induce artificial gauge fields for particles moving on this lattice. Here we demonstrate how to engineer a novel state of matter, namely an antiferromagnet with a Landau-level excitation spectrum of magnons. We consider a honeycomb-lattice Heisenberg model and show that triaxial strain leads to equally spaced pseudo-Landau levels at the upper end of the magnon spectrum, with degeneracies characteristic of emergent supersymmetry. We also present a particular strain protocol which induces perfectly quantized magnon Landau levels over the whole bandwidth. We discuss experimental realizations and generalizations.'
author:
- Mary Madelynn Nayga
- Stephan Rachel
- Matthias Vojta
title: Magnon Landau levels and emergent supersymmetry in strained antiferromagnets
---
Artificial gauge fields have become a powerful tool to engineer states of matter. Originally discussed in the context of strain applied to carbon nanotubes [@suzuura02], they have later been utilized to induce Landau levels in strained graphene [@guinea09; @levy10; @strain-arc; @voz_rev10], with a spacing corresponding to a magnetic field exceeding 300T, far beyond what is reachable with laboratory magnetic fields. More fundamentally, artificial gauge fields can be used to induce effects akin to that of orbital magnetic fields for *charge-neutral* particles, resulting in novel states which cannot be generated otherwise. This has been proposed for ultracold gases as well as solids [@cold_gauge; @aidelsburger18], with strain-induced Landau levels for Bogoliubov quasiparticles of nodal superconductors [@wachtel17; @franz18] and for Majorana excitations of a Kitaev spin liquid [@rachel16a] being two prominent examples.
In this Letter, we propose a novel setting for strain engineering, namely ordered quantum antiferromagnets. We show that inhomogeneous exchange couplings in simple Néel antiferromagnets, generated by applying a suitable strain pattern, transform the conventional magnon spectrum into a sequence of magnonic pseudo-Landau levels. There are three key differences to previous condensed-matter realizations of pseudo-Landau levels: (i) The magnon Landau levels emerge in the high-energy part of the spectrum, i.e., starting from the upper band edge, not at low energies as in the graphene case. (ii) The magnon Landau levels derive from a mode which disperses quadratically in the absence of strain, not from a linear Dirac-like dispersion. As a result, the magnon Landau levels are equally spaced, with a spacing scaling linearly with the pseudo-magnetic field, as opposed to the square-root dependence of Dirac Landau levels. (iii) Perhaps most remarkably, the Landau-level spectrum displays emergent supersymmetry: Its degeneracies are equivalent to that of a system with a boson and a fermion of equal energy, thus forming one of the rare condensed-matter realizations of supersymmetry [@kulish80; @grover14; @jian17], which in general refers to systems where each boson has a fermionic superpartner and vice versa.
We present numerical spin-wave results for the Heisenberg antiferromagnet on finite honeycomb lattices subject to triaxial strain, and we derive the corresponding continuum field theory which is based on an expansion about the *maximum* of the magnon dispersion. Following earlier work [@rachel16b], we also show that the combined limit of strong magneto-elastic coupling and weak lattice deformations leads to perfectly quantized magnon Landau levels over the entire range of magnon energies. We argue that magnon Landau levels can be detected in high-resolution spectroscopic experiments on strained honeycomb antiferromagnets, and we discuss generalizations to other lattices.
We note that very different types of magnon Landau levels have appeared in earlier work: Ref. considered gapped antiferromagnets subject to electric field gradients, and Ref. studied Dirac magnons in ferromagnets under strain.
*Model and spin-wave theory.* We consider a standard nearest-neighbor antiferromagnetic Heisenberg model of spins $S$ placed on the sites of a honeycomb lattice. The Hamiltonian with spatially varying couplings reads $$\label{hk}
\mathcal{H} = \sum_{\langle ij\rangle} J_{ij} \vec{S}_i \cdot \vec{S}_j\,.$$ On this bipartite lattice, the antiferromagnetic couplings are unfrustrated, such that the ground state of the homogeneous system, $J_{ij}\equiv J$, is a simple Néel antiferromagnet for any $S$. In the following we will assume that a Néel state is also realized in the strained system; this applies in the semiclassical ($S\to\infty$) limit [@classnote] as well as, by continuity, for weak strain and any $S$.
The excitation spectrum can be obtained using spin-wave theory. Using Holstein-Primakoff bosons $a$ and $b$ on sublattices $A$ and $B$, respectively, which describe fluctuations above the collinear Néel state, the bilinear piece of the Hamiltonian reads $$\label{hsw}
{\mathcal{H}_{\rm SW}}= S \sum_{\langle ij\rangle} J_{ij} \left(a_i^\dagger a_i + b_j^\dagger b_j + a_i b_j + a_i^\dagger b_j^\dagger \right).$$ In the homogeneous (i.e., unstrained) case, the Hamiltonian can be transformed by subsequent Fourier and Bogoliubov transformations into a system of non-interacting bosons with energies $$\label{specsw}
{\omega_{\vec q}}= 3JS \sqrt{1 - |{\gamma_{\vec q}}|^2}$$ where ${\gamma_{\vec q}}= (1/3) \sum_j e^{i\vec{q}\cdot\vec{\delta_j}}$ with $\vec{\delta}_j$ being the three nearest-neighbor vectors on the honeycomb lattice. This spectrum displays gapless (Goldstone) modes at $\vec{q}=0$ and is maximum at the momenta $\vec{q} = \vec K= 2\pi/(3a_0)(1/\sqrt{3},1)$ and $\vec K'= -2\pi/(3a_0)(1/\sqrt{3},1)$, with $a_0$ the lattice constant.
![Left: Distorted honeycomb-lattice antiferromagnet, with displacements from triaxial strain, Eq. ; the undistorted lattice is shown in light gray. Longer (shorter) bonds correspond to weaker (stronger) exchange couplings $J_{ij}$. The system size is $N=8$. Right: Magnon spectrum of the unstrained honeycomb Heisenberg antiferromagnet, showing a quadratic maximum at momenta $\vec K$ and $\vec K'$. []{data-label="fig:setup"}](lattice_setup_N8 "fig:"){width="0.45\columnwidth"}![Left: Distorted honeycomb-lattice antiferromagnet, with displacements from triaxial strain, Eq. ; the undistorted lattice is shown in light gray. Longer (shorter) bonds correspond to weaker (stronger) exchange couplings $J_{ij}$. The system size is $N=8$. Right: Magnon spectrum of the unstrained honeycomb Heisenberg antiferromagnet, showing a quadratic maximum at momenta $\vec K$ and $\vec K'$. []{data-label="fig:setup"}](spin-wave_AFM-Hberg.jpg "fig:"){width="0.5\columnwidth"}
*Strain-induced Landau levels.* In the context of graphene, it has been theoretically shown [@guinea09; @voz_rev10] that a spatial modulation of hopping energies mimics the effect of a vector potential. Near the Dirac energy, this emergent vector potential can be expressed through the strain tensor ${u}_{ij}$ as $\vec A \propto \pm ( {u}_{xx} - {u}_{yy} , -2 {u}_{xy} )^T$. If the resulting pseudo-magnetic field, $\vec B={\rm rot}\vec A$, is sufficiently homogeneous – this applies, e.g., to triaxial strain with the displacement vector given by [@guinea09; @peeters13] $$\label{displace}
\vec{{U}}(x,y) = \bar{C} \big(2xy, x^2-y^2 \big)^T$$ – it can induce single-particle pseudo-Landau levels very similar to Landau levels in a physical magnetic field. In Eq. $\bar{C}$ (measured in units of $1/a_0$) parameterizes the distortion, and ${u}_{ij} = (\partial_i {U}_j + \partial_j {U}_i)/2$. Adapting the idea of strain-induced artificial gauge fields to antiferromagnets, we shall study the Heisenberg model with spatially modulated exchange couplings. We choose $$\label{ourjij}
J_{ij} = J \left[ 1- \beta (|\vec{\delta}_{ij}|/a_0 - 1)\right]$$ where we calculate the distance $\vec{\delta}_{ij}=\vec{R}_i+\vec{U}_i-\vec{R}_j-\vec{U}_j$ using the displacement $\vec{U}(x,y)$ evaluated at the lattice positions $\vec{R}_i$ of the undistorted honeycomb lattice. The factor $\beta$ encodes the strength of magneto-elastic coupling, and the dimensionless parameter (dubbed “strain” below) $C = \bar{C}\beta a_0$ will enter our simulations as a measure of the modulations of the $J_{ij}$. Note that Eq. represents a linear approximation to the full dependence of the exchange constant on the bond length [@peeters13]. The inhomogeneous Heisenberg model with couplings given by Eq. is expected to display magnon excitations (on top of its collinear ground state) which are influenced by artificial gauge fields. As we demonstrate now, triaxial strain will produce a homogeneous pseudo-magnetic field for magnons with momenta near $\vec{K}$ and $\vec{K'}$ leading to equidistant pseudo-Landau levels.
*Field theory for strained magnons.* We derive a continuum field theory which captures the effect of the strain-induced gauge field. To this end, we expand the magnon Hamiltonian near $\vec{K}$. We introduce the two-vector $\Psi(\vec{r})=(\psi_a(\vec{r}),\psi_b^\dagger(\vec{r}))^T$, representing the real-space Fourier transform of $(a_{\vec{K}+\vec{k}},b_{-\vec{K}+\vec{k}}^\dagger)$ for small $\vec k$. Working to linear order in the lattice distortion and to linear order in $\vec k$, we find that the magnon Hamiltonian can be written as [@suppl] $$\label{hstrain}
{\mathcal{H}_{\rm SW}}= 3JS\! \int\! d^2r \Psi^\dagger\!
\begin{pmatrix}
1+a & -\frac{a_0}{2} (\Pi_x \!-\! i \Pi_y) \\
-\frac{a_0}{2} (\Pi_x \!+\! i \Pi_y) & 1+a
\end{pmatrix}
\!\Psi$$ where $\vec{\Pi} = \vec{p} + \vec{A}$ and $p_x=-i\partial_x$, $p_y=-i\partial_y$. Here, $a = - \beta ({u}_{xx}+{u}_{yy})/2$ encodes the strain-induced change of the magnon bandwidth, and the strain-induced gauge field $\vec{A}$ has the form $$\label{eqn: pseudo A}
\vec{A} = \dfrac{\beta}{2a_0}
\begin{pmatrix}
{u}_{xx}-{u}_{yy} \\
-2{u}_{xy}
\end{pmatrix} \,
+ \vec{A}_K \,.$$ This consists of a $\beta$-dependent part responsible for Landau-level physics [@guinea09; @voz_rev10] and a $\beta$-independent lattice correction $\vec{A}_K$ [@kitt12; @masir13] which will be neglected in the following, for details see Ref. . In the unstrained case, $\vec A=0$ and $a=0$, the spectrum of the Hamiltonian can be obtained by Fourier and Bogoliubov transformations as ${\omega_k}= 3JS (1- k^2 a_0^2/8)$, consistent with Eq. for $\vec q=\vec K+\vec k$ and small $\vec k$.
The Hamiltonian differs in three crucial points from other realizations of strain-induced gauge fields: (i) It describe bosonic quasiparticles, not fermionic electrons. (ii) Particle number is not conserved. As a result, a bosonic Bogoliubov transformation is required to diagonalize , technically different from the diagonalization of a Hermitian matrix. (iii) The diagonal elements in are non-zero, as a result of magnon on-site energies.
We now solve the problem explicitly for triaxial strain where $a=0$ by symmetry, and the $\beta$-dependent part of the vector potential reads $\vec A = 2 \beta \bar{C} (y/a_0,-x/a_0)^T$ corresponding to symmetric gauge.
We introduce two types of bosonic ladder operators $$\begin{aligned}
\alpha &= \left[p_x + y \frac{2C}{a_0^2} + i (p_y - x \frac{2C}{a_0^2}) \right] \frac{a_0}{\sqrt{8C}} \,,\\
\gamma &= \left[p_x - y \frac{2C}{a_0^2} - i (p_y + x \frac{2C}{a_0^2}) \right] \frac{a_0}{\sqrt{8C}}\end{aligned}$$ which obey $[\alpha,\alpha^\dagger]_-=[\gamma,\gamma^\dagger]_-=1$ and $[\alpha^{(\dagger)},\gamma^{(\dagger)}]_-=0$, and we have assumed $C>0$. The Hamiltonian now reads $$\label{hstrain2}
{\mathcal{H}_{\rm SW}}= 3JS
\int \! d^2r
\Psi^\dagger(\vec r)
\begin{pmatrix}
1 & -{\sqrt{2C}} \alpha^\dagger \\
-{\sqrt{2C}} \alpha & 1
\end{pmatrix}
\Psi(\vec r) \,.$$ As usual for Landau levels in symmetric gauge, the spectrum is determined by the $\alpha$ boson operators, whereas the $\gamma$ boson operators commute with the $\alpha$ and the Hamiltonian, such that all single-particle levels display an extensive degeneracy corresponding to $m = \gamma^\dagger\gamma$.
The single-particle (magnon) eigenstates can be obtained from by a bosonic Bogoliubov transformation [@wessel05]. The solutions of the corresponding differential equations can be written in real space as [@suppl] $$\begin{aligned}
\label{solu}
\Phi_{n,m}^\pm =
\begin{pmatrix}
c^\pm_n \phi_{n,m} \\
\phi_{n-1,m} \\
\end{pmatrix}
,~~
\Phi_{0,m}^+ =
\begin{pmatrix}
\phi_{0,m} \\
0 \\
\end{pmatrix}
,\end{aligned}$$ with integer quantum numbers $n\geq1$ and $m\geq0$. The wavefunctions $\phi_{n,m}$ are harmonic-oscillator eigenstates with $\phi_{n,m} = (n!m!)^{-1/2} (\alpha^\dagger)^n (\gamma^\dagger)^m \phi_0$ with $\alpha \phi_0=\gamma \phi_0=0$, in real space $\phi_0(r) \propto \exp(-Cr^*r/(4a_0^2))$ where $r=x+iy$, and the coefficients are $c_n^+=\sqrt{2/(Cn)}$ and $c_n^-=\sqrt{Cn/2}$. Hence, we have a single set of $n=0$ solutions and paired sets of solutions with $n=1, 2, 3, \ldots$ The Hamiltonian is diagonalized by magnon operators $A_{n,m}^\pm = \int d^2 r \Psi^T(r) \Phi_{n,m}^\pm(r)$ and takes the Landau-level form $${\mathcal{H}_{\rm SW}}= \sum_m \left( E_0 (A_{0,m}^+)^\dagger A_{0,m}^+ + \sum_{n=1,\pm}^\infty E_n (A_{n,m}^\pm)^\dagger A_{n,m}^\pm \right)$$ up to a global constant. The magnon energies are $$\label{magen}
E_n = 3JS \sqrt{1 - 2 C n} \approx 3JS (1 - C n),~~n=0,1,2,\ldots$$ valid near the upper end of the spectrum, i.e., $Cn\ll1$.
{width="\textwidth"}
Eq. is our key result: The $E_n$ represent the energies of highly degenerate magnon pseudo-Landau levels. In contrast to strained graphene, the Landau-level energy decreases with increasing Landau-level index $n$ and the levels are equally spaced – the latter is related to the unstrained spectrum being quadratic (as opposed to linear) in momentum.
*Emergent supersymmetry.* Given that the solutions for $n=1, 2, \ldots$ occur in pairs, the degeneracy of the $n>0$ magnon Landau levels is twice as large as that of the $n=0$ level. This implies an emergent supersymmetry, i.e., the single-particle states can be mapped to that of a supersymmetric harmonic oscillator.
We can make this apparent by noting that the single-particle states , with energies , are the eigenstates of the following Hamiltonian $$\label{susyosc}
h = 3JS \left[1 - C (F^\dagger F + B^\dagger B)\right]$$ where we have introduced the operators $$\begin{aligned}
F &= \sum_m \sum_{n=0}^\infty |\Phi^+_{n,m}\rangle \langle\Phi^-_{n+1,m}|\,,\\
B &= \sum_m \left( \sum_{n=0}^\infty \!\sqrt{n} |\Phi^+_{n,m}\rangle \langle\Phi^+_{n+1,m}|
+ \sum_{n=1}^\infty \!\sqrt{n} |\Phi^-_{n,m}\rangle \langle\Phi^-_{n+1,m}| \right)\end{aligned}$$ which act in the single-particle Hilbert space, see Fig. S1 [@suppl]. The $|\Phi_{0,m}^+\rangle$ and $|\Phi_{n,m}^\pm\rangle$ are (normalized) states corresponding to the real-space representation , with their scalar product defined via a $\Sigma$ norm [@suppl]. These states form a complete single-particle basis, and the operators $F$ and $B$ obey the canonical commutation relations for fermions and bosons, respectively, $[F,F^\dagger]_+ = 1$ and $[B,B^\dagger]_- = 1$ [@suppl].
Since both the boson ($B$) and the fermion ($F$) have the same excitation energy, Eq. resembles an inverted supersymmetric harmonic oscillator. We emphasize that the supersymmetry arises from the two-sublattice structure combined with the non-conservation of the bosonic particle number: While the fermionic two-sublattice case yields a spectrum which is particle-hole-symmetric, the bosonic case requires non-negative excitation energies, then implying pairs of degenerate solutions for $n>0$.
*Numerical results.* To verify our field-theoretic calculation, we now turn to a numerical analysis of the magnon Hamiltonian on finite strained systems. In principle, arbitrarily shaped systems with open boundary conditions can be considered; guided by the insights of Ref. we choose triangular-shaped systems with linear size $N$ and $N^2$ sites, as shown in Fig. \[fig:setup\](a). We employ exchange couplings according to Eqs. and , yielding a spin-wave Hamiltonian as in Eq. . We solve the bosonic Bogoliubov problem, amounting to the diagonalization of a non-Hermitian matrix of size $N^2 \times N^2$, using the algorithm outlined in the appendix of Ref. .
For given system size $N$, the strain $C=\bar C\beta a_0$ has a maximum value ${C_{\rm max}}$ beyond which some of the exchange couplings become negative due to the linearization in . Hence, we parameterize the strain by its strength $C/{C_{\rm max}}$ and consider different values of the magneto-elastic coupling $\beta$. In general, the pseudo-magnetic field will be spatially inhomogeneous, with variations near the edges, and clear Landau-level signatures are expected in the bulk of the sample [@peeters13].
Results for the local magnetic susceptibility, equivalent to the local magnon density of states (DOS), measured near the center of the sample, are shown in Fig. \[fig:ldos1\]. Equally spaced Landau levels are clearly visible at the upper end of the spectrum in perfect agreement with Eq. , with the $n=0$ level being present only on the A sublattice [@suppl]. The Landau-level spacing scales linearly with applied strain [@suppl] as anticipated. Note that the states at the lower spectrum end display small degeneracies only, see also Fig. \[fig:dos1\] below, and do not correspond to Landau levels [@suppl], except for the special case in Fig. \[fig:dos1\](c).
{width="\textwidth"}
As observed in Ref. and discussed in detail in Ref. , the combined limit of large $\beta$ and small $\bar C$ reduces non-linearities in the pattern of bond strengths, such that pseudo-Landau levels can exist over a large range of energies. For this (artificial) $\beta\to\infty$ limit we show results for the global magnon DOS in Fig. \[fig:dos1\]. Most remarkable is the result for $C={C_{\rm max}}$, Fig. \[fig:dos1\](c), where we find perfectly degenerate and equally spaced magnon levels over the full bandwidth. For a system of linear size $N$ there are $N+1$ magnon energies, $E_n=3JS(1-n/N)$, with degeneracies $d_0=N-1$, $d_N=1$, and $d_n=2(N-n)$ for $n=1,\ldots,N-1$. For large $N$ and small $n$, these degeneracies agree with the field-theoretic result derived above and reflect the supersymmetric nature of the spectrum. For $C<{C_{\rm max}}$, Fig. \[fig:dos1\](a,b), the energy range of clearly visible magnon Landau levels decreases. Moreover additional peaks appear in between the Landau levels which correspond to magnon states localized near the sample edges, i.e., reflect finite-size effects [@suppl].
*Discussion.* Having established the existence of magnon Landau levels in strained honeycomb antiferromagnets, we turn to broader aspects. First, our calculation is controlled in the semiclassical limit of large spin $S$. Quantum (i.e. $1/S$) corrections lead to higher-order magnon terms in the Hamiltonian. In the present case with collinear order, cubic vertices are forbidden. Self-energy effects from quartic magnon interactions will shift all excitation modes at the upper end of the spectrum in a similar fashion, such that the Landau-level structure remains intact. Broadening effects arising from quartic vertices are typically small. While a full calculation is beyond the scope of this paper, we expect our results to be robust down to $S=1/2$. Second, the Landau quantization implies that the semiclassical trajectory of a magnon wavepacket centered at $\vec K$ or $\vec K'$ is a circle [@unpub].
The experimental realization of Fig. \[fig:setup\](a) requires to prepare a strained quasi-2D antiferromagnet and to measure its dynamic spin susceptibility. Although challenging, we believe this can be achieved with present-day technology: van-der-Waals-bonded layered magnets such as $\alpha$-RuCl$_3$, CrCl$_3$, or CrI$_3$ can be prepared as ultrathin (even monolayer) samples. Those can be placed on a curved substrate surface to realize suitable strain, akin to what has been done for graphene [@levy10]. The excitation spectrum of such a sample may be probed using inelastic Raman scattering or electron-spin-resonance techniques. Realistic values of $\beta=3\ldots 5$ and $\bar C \lesssim 0.2/(Na_0)$ imply a Landau-level spacing of $10^{-3}J$ for a system of $500^2$ atoms. Alternatively, a strained antiferromagnet may be realized in a cold-atom quantum simulator. This requires to generate a Hubbard model in the Mott regime [@greif16] on a suitable inhomogeneous optical lattice. More generally, all quantum simulators for fermionic Hubbard models could in principle be used [@hensgens17], in particular when an arbitrary two-dimensional lattice can be simulated [@salfi16].
*Summary.* We have demonstrated how to engineer a novel state of matter, where magnon excitations of a collinearly ordered antiferromagnet display highly degenerate Landau levels at the top of the spectrum. This realization of strain-induced Landau levels fundamentally differs from previous ones, as the starting spectrum displays a quadratic (instead of linear) dispersion. Consequently, the spectrum features equidistant Landau levels, and moreover realizes emergent supersymmetry. We have also discovered that the combined limit of strong magneto-elastic coupling and weak lattice deformations leads to perfectly quantized magnon Landau levels over the full range of magnon energies.
Generalizations to other lattice structures will be discussed elsewhere. We anticipate that adapting the scheme developed in Ref. will enable us to engineer magnon Landau levels also in collinear diamond-lattice antiferromagnets. Strain engineering of non-collinear antiferromagnets [@attig17] will not only modify the excitation spectrum, but also lead to strain-dependent reference states, possibly generating entirely new types of order – this exciting phenomenology is left for future work.
We thank D. Arovas, J. Attig, L. Fritz, I. Goethel, A. Rosch, and S. Trebst for discussions as well as collaborations on related work. We acknowledge financial support from the DFG through SFB 1143 and the Würzburg-Dresden Cluster of Excellence on Complexity and Topology in Quantum Matter – *ct.qmat* (EXC 2147, project-id 39085490) as well as by the IMPRS on Chemistry and Physics of Quantum Materials. SR acknowledges an ARC Future Fellowship (FT180100211).
H. Suzuura and T. Ando, Phys. Rev. B **65**, 235412 (2002).
F. Guinea, M. I. Katsnelson, and A. K. Geim, Nat. Phys. [**6**]{}, 30 (2009).
N. Levy, S. A. Burke, K. L. Meaker, M. Panlasigui, A. Zettl, F. Guinea, A. H. Castro-Neto, and M. F. Crommie, Science [**329**]{}, 544 (2010).
F. Guinea, A. K. Geim, M. I. Katsnelson, and K. S. Novoselov, Phys. Rev. B [**81**]{}, 035408 (2010).
M. Vozmediano, M. Katsnelson, and F. Guinea, Phys. Rep. [**496**]{}, 109 (2010).
J. Dalibard, F. Gerbier, G. Juzeli|[u]{}ūnas, and P. Öhberg, Rev. Mod. Phys. [**83**]{}, 1523 (2011).
M. Aidelsburger, S. Nascimbene, and N. Goldman, Comptes Rendus Physique **19**, 394 (2018).
G. Massarelli, G. Wachtel, J. Y. T. Wei, and A. Paramekanti, Phys. Rev. B **96**, 224516 (2017).
E. M. Nica and M. Franz, Phys. Rev. B **97**, 024520 (2018).
S. Rachel, L. Fritz, and M. Vojta, Phys. Rev. Lett. [**116**]{}, 167201 (2016).
P. Kulish and E. Sklyanin, Zap. Nauchn. Sem. LOMI **95**, 129 (1980); J. Soviet Math. **19**, 1956 (1982).
T. Grover, D. N. Sheng, and A. Vishwanath, Science **344**, 280 (2014).
S.-K. Jian, C.-H. Lin, J. Maciejko, and H. Yao, Phys. Rev. Lett. **118**, 166802 (2017).
S. Rachel, I. Göthel, D. P. Arovas, and M. Vojta, Phys. Rev. Lett. [**117**]{}, 266801 (2016).
K. Nakata, S. K. Kim, J. Klinovaja, and D. Loss, **96**, 224414 (2017).
Y. Ferreiros and M. A. H. Vozmediano, **97**, 054404 (2018).
In the semiclassical limit, $S\to\infty$, the Hamiltonian on the honeycomb lattice has a Néel ground state for any choice of $J_{ij}>0$ which follows from energy minimization.
M. Neek-Amal, L. Covaci, K. Shakouri, and F. M. Peeters, Phys. Rev. B [**88**]{}, 115428 (2013).
See supplemental material for a derivation of the continuum field theory, an illustration of the supersymmetric nature of the spectrum, a discussion of the effect of strain on the Goldstone modes, and for additional numerical results.
A. Kitt, V. Pereira, A. Swan, and B. Goldberg, Phys. Rev. B **85**, 115432 (2012).
M. R. Masir, D. Moldovan, and F. M. Peeters, Solid State Comm. **175**, 76 (2013).
S. Wessel and I. Milat, Phys. Rev. B **71**, 104427 (2005).
C. Poli, J. Arkinstall, and H. Schomerus, Phys. Rev. B **90**, 155418 (2014).
M. M. Nayga *et al.*, to be published.
D. Greif, M. F. Parsons, A. Mazurenko, C. S. Chiu, S. Blatt, F. Huber, G. Ji, and M. Greiner, Science **351**, 953 (2016).
T. Hensgens, T. Fujita, L. Janssen, X. Li, C. J. Van Diepen, C. Reichl, W. Wegscheider, S. Das Sarma, and L. M. K. Vandersypen, Nature **548**, 70 (2017).
J. Salfi, J. A. Mol, R. Rahman, G. Klimeck, M. Y. Simmons, L. C. L. Hollenberg, and S. Rogge, Nat. Commun. **7**, 11342 (2016).
J. Attig and S. Trebst, Phys. Rev. B **96**, 085145 (2017).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the spin-Hall conductance fluctuations in ballistic mesoscopic systems. We obtain universal expressions for the spin and charge current fluctuations, cast in terms of current-current autocorrelation functions. We show that the latter are conveniently parametrized as deformed Lorentzian shape lines, functions of an external applied magnetic field and the Fermi energy. We find that the charge current fluctuations show quite unique statistical features at the symplectic-unitary crossover regime. Our findings are based on an evaluation of the generalized transmission coefficients correlation functions within the stub model and are amenable to experimental test.'
author:
- 'J. G. G. S. Ramos'
- 'A. L. R. Barbosa'
- 'D. Bazeia'
- 'M. S. Hussein'
- 'C. H. Lewenkopf'
title: Generalized correlation functions for conductance fluctuations and the mesoscopic spin Hall effect
---
Introduction {#sec:introduction}
============
The discovery of the spin Hall effect (SHE) [@mote1; @mote2; @mote3; @mote4] in both metal and semiconductor structures has opened an important new possibility to control the effects of non-equilibrium spin accumulation. [@jungwith12] The basic idea underlying the SHE is to generate spin currents transverse to the longitudinal transport of charge by creating an imbalance between the spin up and spin down states. [@mote5; @mote6]
The detection of spin-Hall conductance fluctuations is a major goal of semiconductor spintronics. [@revisao1] It is, however, a hard endeavor. The main reason is the difficulty to efficiently connect ferromagnets leads to two-dimensional semiconductor structures. [@revisao2] For ballistic systems coupled metallic leads, it is in principle possible to detect the signal when scattering by impurities induce a separation of the spin states.
Through this mechanism, universal spin-Hall conductance fluctuations (USCF) can lead to accumulation of spin at the electron reservoirs. The USCF appear in the transverse current measured in multi-terminal devices in the presence of a sufficiently large magnetic field. [@mote5] Signals of the spin accumulation can be inferred, for instance, from time-dependent fluctuations of the spectral currents (noise power) [@nos1] or from the analysis of universal conductance fluctuations. [@simulacao; @jacquod1; @Nazarov2007; @Krich2008] Spin-Hall conductance fluctuations have been theoretically studied for mesoscopic systems in the diffusive [@simulacao] as well as in the ballistic regime. [@jacquod1] In the absence of both spin rotation symmetry and magnetic field, these studies predict universal spin-Hall conductance fluctuations with a root mean square amplitude of about $0.18 (e/ 4 \pi)$.
So far, a direct detection of spin-Hall currents by analyzing transverse current fluctuations has not been made. In this paper, we propose an alternative way to infer spin-Hall conductance fluctuations, based on the universal relation between spin and charge current fluctuations in chaotic quantum dots. We find that the change and spin current-current correlation functions show a quite unique dependence on the ratio of open modes between transversal and longitudinal terminals. This dependence allows one to infer the magnitude of the spin current. From the technical point of view, we adapt the diagrammatic technique developed to describe the electronic transport in two-terminal chaotic quantum dots in the presence of a spin-orbit interaction, at the symplectic-unitary corssover, [@brouwer2; @Cremers03] to the case of multi-terminal spin resolved currents.
The paper is organized as follows. In Sec. \[sec:theory\] we review the Landauer-Büttiker approach used to calculate the multi-terminal charge and spin currents. Next, in Sec. \[sec:disorder\], we present the diagrammatic theory we employ to calculate the universal spin Hall current fluctuations. The phenomenological implications of our findings are discussed in Sec. \[sec:results\]. Finally, in Sec. \[sec:conclusions\], we present our conclusions.
Theoretical framework {#sec:theory}
=====================
In this section we review the scattering matrix formalism that describes the spin Hall effect in ballistic conductors. We follow the approach put forward in Refs. . We find helpful for the reader to have, in a nutshell, the expressions for the charge and spin currents with their explicit dependence on the electron energy and the presence of an external magnetic field. The latter is of particular importance in our study: The magnetic field breaks time-reversal symmetry and drives the symplectic-unitary crossover. The description of the universal spin current fluctuations at the crossover is one of the key results of this paper, as discussed in Sec. \[sec:disorder\].
We consider multi-terminal two-dimensional systems where the electrons flow under the influence of a spin-orbit interaction of the Rashba or/and Dresselhaus type. The device is connected by ideal point contacts to ${\cal N}$ independent electronic reservoirs, denoted by $i=1,\cdots,{\cal N}$. The electrodes are subjected to voltages denoted by $V_i$. We use the Landauer-Büttiker approach to write the $\alpha$ direction spin resolved current through the $i$th terminal as [@Buttiker86] $$I^{\alpha\sigma}_i = \frac{e^2}{h}\sum_{j,\sigma^\prime} \sum_{\scriptsize{\small m\in i}\atop{n\in j}} | S_{m,\sigma;n,\sigma^\prime}^{\alpha}|^2 (V_i - V_j)$$ where $\sigma$ and $\sigma^\prime$ are the spin projections in the $\alpha=x,y$ or $z$ direction and $S$ is the quaternionic scattering matrix of order $2N_T\times 2N_T$ that describes the transport of electrons through the system. The total number of open orbital scattering channels is $N_T=\sum^{\cal N}_{i=1}N_i$, where $N_i$ is the number of open channels in $i$th lead point contact.
The electric current at the $i$th terminal is $I^{(0)}_i=I_{i}^{\alpha\uparrow}+I_{i}^{\alpha\downarrow}$, for any direction $\alpha=x,y,z$ of the electron spin projection. Similarly, the $\alpha$-axis component of the spin current $I_i^{(\alpha)}$ is defined as the difference between the two spin projections along the $\alpha$-axis, namely, $I_{i}^{(\alpha)}=I_{i}^{\alpha\uparrow}-I_{i}^{\alpha\downarrow}$.
Let us consider a set up with ${\cal N}=4$ terminals. An applied bias voltage between electrodes $i=1$ and 2 gives rise to a longitudinal electronic current $I$ and due to the spin Hall effect to spin currents at the transversal contacts. [@simulacao] Charge conservation imposes that $I\equiv I_{1}^{(0)}=-I_{2}^{(0)}$. Moreover, due to the absence of a transversal voltage bias, $I_i^{(0)}=0$ for $i=3,4$. Using these constraints, it was shown [@jacquod1] that the transversal spin currents are given by $$J_{i}^{(\alpha)}=\frac{1}{2}\left(\mathcal{T}_{i2}^{(\alpha)}-\mathcal{T}_{i1}^{(\alpha)}\right)-\mathcal{T}_{i3}^{(\alpha)} \widetilde{V}_{3}- \mathcal{T}_{i4}^{(\alpha)} \widetilde{V}_{4}
\label{eq:spincurrent}$$ where $i=3,4$. Likewise, the longitudinal charge current reads $$\begin{aligned}
J^{(0)}_i=&\,\frac{1}{4}\! \left(2N_{1}-\mathcal{T}_{11}^{(0)}+2N_{2}-\mathcal{T}_{22}^{(0)}+\mathcal{T}_{12}^{(0)}+\mathcal{T}_{21}^{(0)}\right)\nonumber\\& -\frac{1}{2}\!\left(\mathcal{T}_{23}^{(0)}-\mathcal{T}_{13}^{(0)}\right)\widetilde{V}_{3} -\frac{1}{2}\!\left(\mathcal{T}_{24}^{(0)}-\mathcal{T}_{14}^{(0)}\right)\widetilde{V}_{4} \;,
\label{eq:chargecurrent}\end{aligned}$$ where $i=1,2$. For notational convenience one introduces the dimensionless currents $J=h/e^2(I/V)$. The effective voltages $\widetilde{V}_{i}$ (in units of $V$), are given by rather lengthy expressions of the generalized transmission coefficients $\mathcal{T}_{ij}^{(0)}$, which can be found in Ref. . Finally, the generalized transmission coefficients read $$\begin{aligned}
\mathcal{T}_{ij}^{(\alpha)}(E,E'\!,B,B')={\rm tr}\!\left[\left(1_i \otimes \sigma^\alpha\right) S^{\dagger}(E,B)1_j S(E'\!,B')\right],\end{aligned}$$ where $E$ and $B$ stand for the electron energy and the magnitude of an external magnetic field. The trace is taken over the scattering channels that belong to the $i$th and $j$th point contact. The matrix $1_i$ stands for a projector of the scattering amplitudes into $i$th point contact channels. The matrices $ \sigma^\alpha$, with $\alpha \in \{x,y,z \} $, are the Pauli matrices, while $ \sigma^0 $ is the $ 2 \times 2 $ identity matrix characterizing an unpolarized (charge) transport.
Universal spin-Hall conductance fluctuations {#sec:disorder}
============================================
We assume that the electron dynamics in the quantum dot is ballistic and ergodic, [@Beenakker97; @Alhassid00] and model the system statistical properties using the random matrix theory. In this section, we describe the procedure to obtain universal spin and charge current-current autocorrelation functions given by Eqs. and . We compare our analytical expressions the with results from numerical simulations.
Let us assume that spin-orbit interaction is sufficiently strong to fully break the system spin rotational symmetry. In other words, the spin-orbit scattering time is much smaller than the electron dwell time in the system, that is, $\tau_{\rm s.o.} \ll \tau_{\rm dwell}$. Hence, in the absence of an external magnetic field, the scattering matrix has symplectic symmetry. By increasing $B$, the system is driven through a crossover from the symplectic to the unitary symmetry.
The resonance $S$-matrix can be parameterized as [@brouwer1] $$\begin{aligned}
S(E,B)=TU \! \left[1-Q^\dagger R(E,B) Q U\right]^{-1} \!T^\dagger,
\label{SMatriz}\end{aligned}$$ where $U$ is a matrix of order $2M \times 2M $ of quaternionic form. [@Mehta91] $M$ stands for the number of resonances of the quantum dot. We take $M \gg N_T$. The matrices $ Q $ and $ T $ describe projector operators of order $(2M-2N_T) \times 2M$ and $2N_T \times 2M$, respectively. Their matrix elements read $ Q_{i, j} = \delta_{i +2 N_T, j} $ and $ T_{i, j} = \delta_{i, j} $. The matrix $R$ of order $ (2M-2N_T) \times (2M-2N_T) $ reads $$\begin{aligned}
R(\epsilon,x)=\exp\left(\frac{i\epsilon}{M} \sigma^0+\frac{x}{M}X\sigma^0 \right) ,
\label{stube}\end{aligned}$$ where $X$ is an anti-hermitian Gaussian distributed random matrix, while $\epsilon = (\tau_{\rm dwell}/ \hbar)E$ and $x^2 = \tau_{\rm dwell}/\tau_ {B}$ are dimensionless parameters representing $E$ and $B$.
Both the dwell time $\tau_{\rm dwell}$ and the magnetic scattering time $\tau_B$ are a system specific quantities. $\tau_{\rm dwell}$ is usually expressed in terms of the decay width, namely $\Gamma = \hbar/\tau_{\rm dwell}= N_T\Delta /2\pi$, where $\Delta$ is the system mean level spacing. In turn, $\tau_B^{-1}$ is the rate by which the electron trajectory accumulates magnetic flux in the quantum dot. For chaotic systems, $\tau_B^{-1}=\kappa ({\cal A} B)/\phi_0$, where $\phi_0$ is the unit flux quantum, ${\cal A}$ is the quantum dot lithographic area, and $\kappa$ is a diffusion coefficient that depends on the quantum dot geometry. [@Pluhar95]
The $ S $-matrix given by Eq. can be formally expanded in powers of $U$, namely, $$S(\epsilon,x)=\sum_{m=0}^{\infty}TU\left[Q^\dagger R(\epsilon,x) Q U\right]^{m}T^\dagger,\nonumber$$ and used to calculate the ensemble average of the transmission coefficients $\langle \mathcal{T}_{ij}^{(\alpha)}\rangle$ for $M \gg N_{T}$ Following this algebraic procedure, we write the sample transmission coefficient as a two-point function of the $S$-matrix, $$\begin{aligned}
\mathcal{T}_{ij}^{(\alpha)}(\epsilon,\epsilon',x,x')=&\!\! \sum_{m,n=0}^{\infty} \!\!{\rm Tr}\!\left\{\left(1_i \otimes \sigma^\alpha\right)T
\left[U^{\dagger}Q^\dagger R^{\dagger}\!(\epsilon,x) Q \right]^{m}
\right.\nonumber\\
&\hskip-0.3cm\times\!\left.
U^\dagger T 1_jTU
\left[Q^\dagger R(\epsilon',x') Q U\right]^{n}T^\dagger\right\}.
\label{Tij}\end{aligned}$$
For chaotic quantum dots, as standard, [@brouwer2] we assume the matrix elements of $U$ as Gaussian variables, with zero mean and variance $1/M$. This allows one to express the calculation of moments and cumulants of $\mathcal{T}_{ij}^{(\alpha)}$ by an integration over the unitary group leading to a diagrammatic expansion in powers of $N_T^{-1}$ in terms of diffusons (ladder) and cooperons (maximally crossed) diagrams. The method is described for the Dyson ensembles in Ref. . This approach has been extended to treat the crossover between symmetry classes [@brouwer2; @GeneralCrossover] and is shown to render the same results as the quantum circuit theory. [@GeneralCrossover]
The diagrammatic technique [@brouwer1] allows one to calculate moments up to arbitrary order of the transmission coefficients. It consists in grouping and integrating over the Haar measure all the independent powers of $U$ in Eq. . Figure \[fig:diagrams\] shows the diagrams that represent the leading order contributions to all possible contractions of the $U$ matrices. In the sum of Eq. , we verify that, after taking the average over $U$, only the powers with $m=n$ contribute to the average transmission coefficients. The white and black dots in Fig. \[fig:diagrams\] stand for the indices of the matrix $U$, with elements $U_{ij}$ in the channels space, while the Greek symbols represent the Pauli matrices indices $\sigma_{\rho,\sigma}$ in the spin space.
![Diagrams representing the leading contributions to the average transmission coefficient $\langle \mathcal{T}_{ij}^{(\alpha)}\rangle$ of Eq. . Vertical lines represent contractions of the $U$ matrix elements. The white and black dots in stand for the channel indices of the matrix $U$ and the Greek symbols represent the Pauli matrices indices $\sigma_{\rho,\sigma}$ in the spin space. The first diagram (top) is known as diffuson and accounts for the semiclassical diffusive processes. The two others diagrams are known as cooperons and give the mean quantum correction, namely, the weak (anti)localization corrections.[]{data-label="fig:diagrams"}](diagramas){width="\columnwidth"}
-0.5cm
The diagram at the top row of Fig. \[fig:diagrams\] is called diffuson, in analogy to the ladder diagram that appears in diagrammatic expansion to calculate the conductivity in disordered diffusive mesoscopic systems. [@mesoscopic] The diffuson contribution to the average transmission coefficient reads $$\begin{aligned}
\langle\mathcal{T}_{ij}^{(\alpha)}(\epsilon,\epsilon',x,x')\rangle^{({\rm d})} =&\frac{1}{2}\sum_{\rho\sigma}\Big[\text{Tr}\left( 1_i\otimes \sigma^\alpha\right)\mathcal{D}\text{Tr}\left(1_j\right)\nonumber\\
&-\text{Tr}\left( 1_i\otimes \sigma^\alpha\right)\mathcal{D}^2\text{Tr}\left(1_j\right)\Big]_{\rho\sigma;\sigma\rho},
\label{Tijm}\end{aligned}$$ where, for clarity, we make explicit the spin degree freedom of the symplectic structure. We use the properties $\text{Tr}\left( 1_i\otimes \sigma^\alpha\right)=2N_i \delta_{\alpha 0}$ and $\text{Tr}\left(1_j\right)=2N_j$ to write $$\begin{aligned}
\mathcal{D}^{-1}= 2 M \sigma^0 \otimes\sigma^0 - \text{Tr} \left(R \otimes R'^\dagger\right)
\label{eq:defD}\end{aligned}$$ with $$\begin{aligned}
\label{eq:defR}
\text{Tr} &\left(R \otimes R'^\dagger\right) = \sigma^0 \otimes\sigma^0 \\
&\;\;\;\;\times \left[2M- 2 N_T-2i(\epsilon-\epsilon') - (x-x')^2\right], \nonumber\end{aligned}$$ where $R'$ is a shorthand notation for $R(\epsilon',x')$. The tensor products follow the “backward algebra" , [@brouwer2; @Cremers03] namely, $( \sigma^i \otimes \sigma^j)( \sigma^k \otimes \sigma^l)= (\sigma^i\sigma^k) \otimes (\sigma^l\sigma^j)$.
The diffuson contribution to the generalized transmission coefficient is obtained by evaluating Eq. using the expressions and . It reads $$\begin{aligned}
\langle\mathcal{T}_{ij}^{(\alpha)}(\epsilon,\epsilon',x,x')\rangle^{({\rm d})} = \delta_{\alpha 0}\frac{N_iN_j}{N_D} \left(2+\frac{1}{N_D}\right),\end{aligned}$$ where $N_D=N_T[1- i\left(\epsilon'-\epsilon\right)+(x'-x)^2/2]$.
We are now ready to evaluate the two maximally crossed diagrams at the bottom row of Fig. \[fig:diagrams\]. They are known as cooperons [@mesoscopic] and represent the main quantum interference correction to the conductance in chaotic systems, responsible for the weak localization peak.
Following the procedure described above, we obtain the cooperon contribution for the generalized average average transmission, namely, $$\begin{aligned}
\langle\mathcal{T}_{ij}^{(\alpha)}(\epsilon,&\epsilon', x,x')\rangle^{\rm (c)} =\,\frac{1}{2}\sum_{\rho\sigma} \bigg[\,
\text{Tr}\Big(F_i^{(\alpha)}(\mathcal{T}f_{TT}\mathcal{T})F_j\Big) \nonumber\\
& -\frac{1}{(2M)^3}{\rm Tr}\Big(F_i^{(\alpha)}(\mathcal{T}f_{TT}\mathcal{T})\Big){\rm Tr}(F_j) \bigg]_{\rho\sigma;\rho\sigma}.
\label{Tijmd}\end{aligned}$$ The operator $\mathcal{T} = \sigma^0 \otimes \sigma^y$ is related with the time-reversed of path in the cooperon channel of the dual space. We also define the matrices $$\begin{aligned}
F_i ^{(\alpha)}=&\, 1_i \otimes \sigma^{\alpha}+\text{Tr}(1_i \otimes \sigma^{\alpha}) \mathcal{D} \,(R'^\dagger\otimes R), \nonumber \\
F_j =&\, 1_j +\text{Tr}(1_j) \mathcal{D}\, (R'^\dagger\otimes R),\end{aligned}$$ where $$\begin{aligned}
f_{UU}=&\left[2M\sigma^0\otimes\sigma^0-\text{Tr}(R\otimes R^*)\right]^{-1},\nonumber\\
f_{TT}=&\,(2M\sigma^0\otimes\sigma^0)\,{\rm Tr}(R\otimes R'^*)f_{UU}.\end{aligned}$$ The superscript $*$ denotes the quaternion complex conjugation. Using the quaternionic conjugation rules and taking the limit $M \to \infty$, we obtain $$\begin{aligned}
f_{UU}^{-1}=& \, 2 N_C \sigma^0 \otimes\sigma^0.\nonumber\\
f_{TT}=&\,(2 M \sigma_0 \otimes \sigma_0)[(2 M - 2N_C) \sigma_0 \otimes \sigma_0] f_{UU},\end{aligned}$$ where $N_C=N_T[1 - i\left(\epsilon'-\epsilon\right)+(x'+x)^2/2]$.
Summing up the diffuson and the cooperon contributions to the generalized transmission coefficients we obtain $$\begin{aligned}
\langle\mathcal{T}_{ij}^{(\alpha)}(\epsilon,\epsilon',x,x')\rangle &=\delta_{\alpha 0}\left\{\frac{2N_iN_j}{N_D} + \frac{N_i}{N_C} \right.\nonumber\\
&\times \left.\left[ \frac{N_j\left[N_D+i (\epsilon-\epsilon')\right]}{N_D^2}-\delta_{ij}\right]\right\}.
\label{condg}\end{aligned}$$ For $\alpha=0$, Eq. reproduces the average electron transmission reported in the literature. [@brouwer1]. As expected in this case, the average spin transmission coefficients are zero.
Let us now analyze the transmission coefficient fluctuations. We use the same diagramatic procedure described above for all the $32$ diagrams characteristic of the usual covariance calculations. [@brouwer1] To address relevant physical situations, it is sufficient to consider transmission coefficients with single energy and magnetic field arguments, that is, $\mathcal{T}_{ij}^{(\alpha)}(\epsilon,x)\equiv\mathcal{T}_{ij}^{(\alpha)}(\epsilon,\epsilon,x,x)$. Following the diagrammatic approach, [@GeneralCrossover] we calculate the covariance ${\rm cov}[\mathcal{T}_{ij}^{(\alpha)}(\epsilon,x),\mathcal{T}_{kl}^{(\beta)}(\epsilon',x')] \equiv \langle\mathcal{T}_{ij}^{(\alpha)}(\epsilon,x)\mathcal{T}_{kl}^{(\beta)}(\epsilon',x')\rangle- \langle\mathcal{T}_{ij}^{(\alpha)}(\epsilon,x)\rangle \langle\mathcal{T}_{kl}^{(\beta)}(\epsilon',x')\rangle $ and obtain
$$\begin{aligned}
{\rm cov}\!\left[\mathcal{T}_{ij}^{(\alpha)}(\epsilon,x),\mathcal{T}_{kl}^{(\beta)}(\epsilon',x')\right] &=& \delta_{\alpha 0}\frac{N_iN_jN_kN_l}{N_T^2}\left(\frac{1}{\left|N_D\right|^2}+\frac{1}{\left|N_C\right|^2}\right)
+\delta_{\alpha\beta}\frac{\delta_{ik}\delta_{jl}N_iN_j}{\left|N_D\right|^2}+\delta_{\alpha 0}\frac{\delta_{il}\delta_{jk}N_iN_k}{\left|N_C\right|^2}
\nonumber\\
&-&\frac{N_iN_k}{N_T}\left(\delta_{\alpha 0}\frac{\delta_{jl}N_j}{\left|N_D\right|^2}+\delta_{\alpha\beta}\frac{\delta_{ik}N_jN_l}{N_k\left|N_D\right|^2}+\delta_{\alpha 0}\frac{\delta_{jk}N_l+\delta_{il}N_j}{\left|N_C\right|^2}\right)
\label{covg}\end{aligned}$$
![(Color online) Generalized average transmission coefficient $\langle \mathcal{T}_{ij}^{(\alpha)}(\epsilon,x=0)\rangle $ versus energy $\epsilon = E/\Gamma$ for different spin and terminal indices. The analytical results (stub model) are represented by the solid and doted lines, while the results of numerical simulations (RMT Hamiltonian model) are represented by the symbols. The figure inset describes the different $\mathcal{T}_{ij}^{(\alpha)}$. []{data-label="fig:THall"}](ghall){width="0.95\columnwidth"}
-0.5cm
Support to our analytical findings is provided by numerical simulations. For that purpose, we find convenient to employ the Hamiltonian approach to the $S$-matrix, [@mahaux69]. The latter is more amenable for numerical simulations than the $S$ matrix parametrization of Eq. and both are statistically equivalent. [@Lewenkopf91]
The Hamiltonian parametrization of the $S$ matrix reads $$\label{eq:SHeidelberg}
S (E,X) = \mbox{$\openone$} - 2\pi i W^\dagger (E - H(B) + i \pi W W^\dagger)^{-1} W \;,$$ where $E$ is the electron propagation energy and $H(B)$ is the matrix of dimension $2 M \times 2 M$ that describes the resonant states ($M$ orbital states times the 2 spin projections). In general, $H$ depends on one (or more) external parameters $X$. As discussed before, we are interested in the case where by increasing $B$ one breaks time-reversal symmetry, driving the system from the symplectic to the unitary symmetry. In our numerically simulations, we consider $B \gg B_c$, namely, the specific case of pure ensembles. Accordingly, $H$ is taken as a member of the Gaussian unitary ensemble corresponding to the case of broken time-reversal symmetric case, usually denoted by $\beta=2$. The matrix $W$ of dimension $M \times (2N_T)$ contains the channel-resonance coupling matrix elements. Since the $H$ matrix is statistically invariant under unitary transformations, the statistical properties of $S$ depend only on the mean resonance spacing $\Delta$, determined by $H$, and $W^\dagger W$. We assume a perfect coupling between channels and resonances, which corresponds to maximizing the average transmission following a procedure described in Ref. .
For simplicity, we take the case of $N\equiv N_i$. The results presented in Figs. \[fig:THall\] and \[fig:covT\] correspond to the systems with $N=5$ perfectly coupled modes and $M=400$ resonant levels. Hence, the $S$-matrix has $2N_T =40$ open channels. The ensemble averages are taken over $N_{\rm r} = 10^5$ realizations within an energy interval around the band center, comprising about $M/4$ resonances.
Figure \[fig:THall\] compares the average transmission $\langle T_{ij}^{(\alpha)}(\epsilon)\rangle$ obtained from the numerical simulations with the analytical expression for a number of different cases. The agreement is very good, with accuracy of the order of $N_{\rm r}^{-1/2}$. The simulations indicate that the average transmission in stationary in $\epsilon=E/\Gamma$, as it should.
Figure \[fig:covT\] contrasts transmission coefficient covariances calculated using Eq.÷ with those obtained from numerical simulations for a number of different cases. As before, the discrepancies are very small and stay within the statistical precision $N_{\rm r}^{-1/2}$. The random matrix theory [@vwz85] predicts an autocorrelation length $\Gamma = N_T \Delta/(2\pi)$ for a two-terminal geometry. Our results for the correlation function extend the latter to four-terminal geometries with (or without) spin polarization.
0.3cm
![(Color online) Transmission coefficient covariance ${\rm cov}[\mathcal{T}_{ij}^{(\alpha)}(\epsilon,x=0)\mathcal{T}_{kl}^{(\beta)}(\epsilon^\prime,x=0)] $ as a function of the energy difference $\Delta\epsilon = (E-E^\prime)/\Gamma$ for different spin and terminal indices. The analytical results (stub model) are represented by the solid and doted lines, while the results of numerical simulations (RMT Hamiltonian model) are represented by the symbols. The figure inset describes the different ${\rm cov}[\mathcal{T}_{ij}^{(\alpha)}(\epsilon)\mathcal{T}_{kl}^{(\beta)}(\epsilon^\prime)]$ considered. []{data-label="fig:covT"}](covghall){width="0.95\columnwidth"}
-0.5cm
Let us now return to the problem of spin and charge current and effective potential. As mentioned, both are combinations of transmission coefficients. Fortunately, it is possible to calculate average currents and current-current correlation functions in terms of the average and the transmission coefficients correlation functions already calculated and confirmed numerically. The effective voltages $\widetilde{V}_{3}$ and $\widetilde{V}_{4}$ show sample-to-sample fluctuations that depend both on the energy and magnetic field. On the other hand, as discussed in Ref. , their ensemble averages depend only on the number of open channels, namely, $ \langle \widetilde{V}_i (\epsilon, x) \rangle = 1 / 2 (N_1-N_2) / (N_1 + N_2) $, with $ i = 3,4$. We also note that the ensemble average of the spin current is always zero, $\langle J^{(\alpha)}_{i}(\epsilon,x) \rangle = 0$, with $i=3,4$ and $\alpha\neq 0$, independently of the energy and magnetic field.
The USCF do not depend on the device geometry (nor on the positions of the terminals), but rather on the number of open channels at each terminal. Without loss of generality, let us analyze the spin current covariance for the case $N_1=N_2=N$ and $N_3=N_4=nN$, a setting that is easily realized in experiments. Here, $n$ is a real positive number that we call “channel factor". [@assimetria]
For the spin currents, for which $\alpha\neq 0$, we obtain $$\begin{aligned}
{\rm cov}\!\left[ J^{(\alpha)}_i(\epsilon, x),J^{(\alpha)}_i(\epsilon', x')\right] = \frac{1}{8}\frac{n/(1+n)^2}{(1+\delta x^2)^2+\delta \epsilon^2}
\label{covspin}\end{aligned}$$ where $\delta \epsilon = \epsilon - \epsilon'$ and $\delta x = x - x'$. It is worth noticing that, for $n=1$ and $\delta \epsilon=\delta x = 0$, Eq. perfectly reproduces the recent reported results [@simulacao; @jacquod1] for the universal fluctuations of the transverse spin conductance, namely, $\textrm{rms}[G_{\rm sH}]=e/4 \pi \{{\rm cov}[ J^{(\alpha)}_i(\epsilon,x),J^{(\alpha)}_i(\epsilon,x)]\}^{1/2} \approx 0.18 e/4 \pi$.
Equation shows that the spin current correlation functions do not depend on the cooperon channels, that give rise to terms containing $N_C$. Hence, these quantities do not depend on the magnetic field, represented by $x$, but rather on its variations, $\delta x$. As a consequence, in the set up we consider, the spin current fluctuations are invariant in the symplectic-unitary crossover regime, a quite remarkable property.
The charge current fluctuations, on the other hand, depend both on the cooperon and diffuson channels, leading to
$$\begin{aligned}
{\rm cov}[ J^{(0)}_i(\epsilon,x),J^{(0)}_i(\epsilon',x')] = \frac{1}{16}\!
\left\{\frac{1+2n}{(1+n)^3}\frac{1}{\left(1+\delta x^2\right)^2+\delta \epsilon^2}
+\frac{1}{(1+n)^2}\frac{1}{[1+( 2 x+\delta x)^2]^2+\delta \epsilon^2}\right\},\label{covcarga}\end{aligned}$$
where $i=1,2$. The magnetic field, represented by $x$, drives the symplectic-unitary crossover. For $x=0$, one recovers the symplectic limit, while the unitary one is attained when $x\gg 1$. Note that in the absence of “transverse" leads, or $n=0$, Eq. reproduces the two-terminal result found in Ref. .
From Eq. it follows that $${\rm cov} \{[J^{(0)}_i (\epsilon,x)-J^{(0)}_i(\epsilon,- x)]^{2}\} = \frac{16nx (1 +2 x^2)}{(1 + n) (1 +4 x^2)^ 2},$$ which demonstrates that, except for the two-terminal case where $n \neq 0$, the charge currents are not even functions of the magnetic field $x$. [@Buttiker86]
Alternative statistical measures {#sec:results}
================================
Equations and are the main results of this paper. Unfortunately, the statistical sampling required to confirm our predictions for the dimensionless currents is rather large, making the experimental requirements quite daunting. An easier accessible statistical measure has been recently proposed: [@densidade2] The dimensionless current $J_i^{(\alpha)}$ fluctuates as $\epsilon$ and $x$ are varied. Let us call the external parameter $z$. Useful statistical information can be extracted from the number of maxima (or minima) of the $J_i^{(\alpha)}(z)$ in a given interval $[z,z+\delta z]$. Using a scale invariance and maximum entropy principle, we relate the joint probability of $J_i^{(\alpha)}(z)$ and its derivatives to a general equation for the density of maxima, for spin and/or charge transport. The average densities of maxima, $ \langle \rho_ {z}^{(\alpha)} \rangle$ of the fluctuating current $J^{(\alpha)}_i$ are given by [@densidade2] $$\begin{aligned}
\left< \rho_{z}^{(\alpha)} \right> =&\frac{1}{2\pi}\sqrt{\frac{T_4}{T_2}} \label{densidade}\\
T_2=&-\frac{d^2}{d(\delta z)^2}{\rm cov}\!\left[ J^{(\alpha)}_i(\epsilon,x),J^{(\alpha)}_i(\epsilon',x')\right]\bigg|_{\delta z=0}\nonumber\\
T_4=&\frac{d^4}{d(\delta z)^4}{\rm cov}\!\left[ J^{(\alpha)}_i(\epsilon,x),J^{(\alpha)}_i(\epsilon',x')\right]\bigg|_{\delta z=0}\nonumber\end{aligned}$$ where $\delta z$ is $\delta \epsilon$ or $\delta x$.
It is convenient to write the charge current covariance as a deformed Lorentzian. For parametric variations of $z$, we set ${\rm cov}[ J^{(0)}_i(\epsilon,x),J^{(0)}_i(\epsilon,x')]=\alpha_{z}(n,x)/ [1+(\delta x)^2]^{h_{z}(n,x)}$, where $\alpha_{z}(n,x)$ is a crossover function and $h_z(n,x)$ characterizes the Lorentzian shape deformation of the charge current correlation as a function of $\delta x$. In terms of the factor $h_{z}$, the average density of maxima reads $$\left< \rho_{z}^{(\alpha)} \right>=\frac{1}{2 \pi} \sqrt{6\big[h_{z}(n,x)+1\big]}, \label{densidadeh}$$ where $z$ can be either $\epsilon$ of $x$.
![\[Color online\] Contour plot of $h_{x}(n, x)$ as a function of the magnetic field, represented by $x$, and the channel factor $n$. The color code is explained at the strip on the right panel.[]{data-label="figuramagnetico"}](f1.eps){width="\columnwidth"}
-0.5cm
Using Eqs. and , we obtain an exact analytical expression for $h_{x}(n,x)$. Its explicit form is not presented here, since it is rather lengthy. Figure \[figuramagnetico\] illustrates its general features. For the unitary symmetry limit, it is well stablished [@Efetov; @Caio] that the electronic conductance correlation function shows a square Lorentzian behavior Accordingly, we find $h_{\rm sqL}\equiv\lim_{x \rightarrow \infty} h_{x}(n,x)=\lim_{x \rightarrow 0}h_{x}(n,x)=2$, for the pure circular unitary and symplectic ensembles, respectively. The symplectic-unitary crossover shows a much richer behavior. Figure \[figuramagnetico\] exhibits a remarkable crossover between sub-Lorentzian, for which $h_{x}<1$, with a minimum value of $h_x \approx 0.64$, and super-Lorentzian, for which $h_{x}>1$ with a maximum value of $h_x \approx 2.92$.
![\[Color online\] Contour plot of $h_{\epsilon}(n, x)$ as a function of the magnetic field, represented by $x$, and the channel factor $n$. The color code is explained at the strip on the right panel.[]{data-label="figuraenergia"}](f2.eps){width="\columnwidth"}
-0.5cm
Parametric variations of $\epsilon$ were first studied in nuclear scattering at low energies and known as Ericson fluctuations [@Ericson]. As it is well-known, their characteristic correlation function versus energy has a Lorentzian shape. In the presence of a perpendicular magnetic field and the channel factor, we obtained a unitary-simplectic crossover of Lorentzian-type shapes, generalizing the correlation functions of Ericson fluctuations. For parametric variations of $ \epsilon $, we set ${\rm cov}[ J^{(0)}_i(\epsilon,x),J^{(0)}_i(\epsilon',x)]=\alpha_{\epsilon}(n,x)/ [1+(\delta \epsilon)^2]^{h_{\epsilon}(n,x)}$, where $\alpha_{\epsilon}(n,x)$ is a crossover parameter and $h_{\epsilon}(n,x)$ characterizes the deformation of the Lorentzian shape. As in the previous case, we also obtain a lengthy analytical expression for $h_{\epsilon}(n,x)$. Its main features are displayed in Fig. \[figuraenergia\]. As expected, $h_{\rm L} \equiv \lim_{x \rightarrow \infty} h_{\epsilon}(n,x)=\lim_{x \rightarrow 0}h_{\epsilon}(n,x)$=1, for pure circular unitary and symplectic ensembles, respectively. Figure \[figuraenergia\] exhibits another remarkable crossover from a sub-Lorentzian, for which $h_{\epsilon}<1$, to a Lorentzian behavior.
According to Eq. , the density of maxima corresponding to pure ensembles, namely, $x=0$ or $x \gg 1$, is $ \langle \rho_ {x}^{(\alpha)} \rangle= \sqrt{6 (h_{\rm sqL} +1)} \approx 0.68$ and $\langle \rho_{\epsilon}^{(\alpha)} \rangle= \sqrt{6 (h_{\rm L} +1)} \approx 0.55$ [@densidade1; @densidade2], for both spin and charge currents. We emphasize that for the case of the spin correlation function, Eq. , $ h_{\epsilon} = h_{\rm L} = 1 $ and $h_{x} = h_{\rm sqL} = 2 $ [*even*]{} in the crossover regime (any value of $ n $ and $x$).
Let us now focus on the longitudinal (charge) correlation function, Eq. . For a given value of the channel asymmetry factor, $\langle \rho_x^{(0)}(n,x)\rangle$ has a unique global maximum, $\langle\rho_{x}^{(0)}(n,x_{max}) \rangle$, and minimum, $\langle\rho_{x}^{(0)}(n,x_{min}) \rangle$ The difference, $\Delta \langle \rho_{x} (n)\rangle \equiv \langle\rho_{x}^{(0)}(n,x_{max}) \rangle-\langle\rho_{x}^{(0)}(n,x_{min}) \rangle $ increases with $n$ until it saturates at $ n \approx 3$. In the absence of spin leads, we find the difference $ \Delta \left <\rho_ {x}(0) \right> \approx 0.27$. In the presence of spin leads, we get $ \Delta \left <\rho_{x}(0.5) \right> \approx 0.19$, $\Delta \left<\rho_ {x} (1)\right> \approx 0.17 $, $\Delta \left< \rho_ {x} (2)\right> \approx 0.15$, and $\Delta \left< \rho_{x}(5) \right> \approx 0.14$. Thus, in measurements made with a perpendicular magnetic field, for symmetric channels ($n=1$), the spin terminals lead to a reduction in the signal of about $37\%$, which becomes even larger with increasing $n$. Interestingly, the maximum and minimum of $ \left< \rho_{x} \right>$ correspond to magnetic field strengths $ x_{min} \approx 0.20$ and $x_{max} \approx 0.47$, for $ n \in [0,5] $, a rather narrow range of values which is accessible experimentally.
In contrast to $\left< \rho_{x}(n) \right>$, the energy variation generates a density of conductance peaks containing a single global minimum, $ \left<\rho_ {\epsilon} (n, x_{min})\right> $, and no global maximum. This minimum is located in a very narrow range of values of $ x_{min} \approx 0.26$ as a function of $ n $. The minimum value of the density at these points is of the order of $ \left <\rho_ {\epsilon} \right> \approx 0.52 $.
Summary and Conclusion {#sec:conclusions}
======================
In this paper, we have investigated the spin-Hall conductance fluctuations in a chaotic open quantum dot with spin-orbit interaction. Both the electronic and the spin-Hall conductance fluctuations are universal functions, with autocorrelation functions that depend on the magnitude of the external magnetic field $B$ and the channel asymmetry factor $n$. A clear intermediate case of symplectic-unitary transitional behavior is found and can be tested experimentally. In particular, the spin current can be measured by using the charge current density of maxima. The results of this Letter extend the understanding of mesoscopic fluctuations to spin- and charge currents in the symplectic-unitary crossover, characteristic of quantum dots subjected to an external magnetic field.
This work is supported in part by the Brazilian funding agencies CAPES, CNPq, FACEPE, FAPERJ, FAPESP, and the Instituto Nacional de Ciência e Tecnologia de Informação Quântica-MCTI.
[10]{}
Y. K. Kato, R. C. Myers, A. C. Gossard, and D. D. Awschalom, Science [**306**]{}, 1910 (2004).
S. O. Valenzuela and M. Tinkham, Nature [**442**]{}, 176 (2006).
T. Kimura, Y. Otani, T. Sato, S. Takahashi, and S. Maekawa, Phys. Rev. Lett. [**98**]{}, 156601 (2007).
D. D Awschalom and M. E. Flatté, Nature Phys. [**3**]{}, 153 (2007).
T. Jungwirth, J. Wunderlich, and K. Olejník, Nature Mat. [**11**]{}, 382 (2012).
J. E. Hirsch, Phys. Rev. Lett. [**83**]{}, 1834 (1999).
J. Sinova, [*et al.*]{}, Phys. Rev. Lett. [**92**]{}, 126603 (2004).
I. Zutić, J. Fabian, and S. Das Sarma, Rev. Mod. Phys. [**76**]{}, 323 (2004).
J. Fabian, A. Matos-Abiaguea, C. Ertler, P. Stano, and I. Zutić, Acta Physica Slovaca [**57**]{}, 565 (2007).
J. G. G. S. Ramos, A. L. R. Barbosa, D. Bazeia, and M. S. Hussein, Phys. Rev. B [**85**]{}, 115123 (2012) .
W. Ren, Z. Qiao, J. Wang, Q. Sun, and H. Guo, Phys. Rev. Lett. [**97**]{}, 066603 (2006).
J. H. Bardarson, I. Adagideli, and Ph. Jacquod, Phys. Rev. Lett. [**98**]{}, 196601 (2007).
Y. V. Nazarov, New J. Phys. [**9**]{}, 352 (2007).
J. J. Krich and B. I. Halperin, Phys. Rev. B [**78**]{} 035338 (2008).
P. W. Brouwer, J. H. Cremers, and B. I. Halperin, Phys. Rev. B [**65**]{}, 081302(R) (2002).
J.-H. Cremers, P. W. Brouwer, and V. I. FalÕko, Phys. Rev. B [**68**]{}, 125329 (2003).
I. Adagideli, J. H. Bardarson, and P. Jacquod, J. Phys.: Condens. Matter [**21**]{}, 155503 (2009).
P. Jacquod, R. S. Whitney, J. Meair, and M. Büttiker, Phys. Rev. B [**86**]{}, 155118 (2012).
M. Büttiker, Phys. Rev. Lett. [**57**]{}, 1761 (1986).
J. G. G. S. Ramos, D. Bazeia, M. S. Hussein, and C. H. Lewenkopf, Phys. Rev. Lett. [**107**]{}, 176807 (2011).
D. M. Brink and R. O. Stephen, Phys. Lett. [**5**]{}, 77 (1963).
C. W. J. Beenakker, Rev. Mod. Phys. [**69**]{}, 731 (1997).
Y. Alhassid, Rev. Mod. Phys. [**72**]{}, 895 (2000).
Z. Pluhař, H. A. Weidenmüller, J. A. Zuk, C. H. Lewenkopf, and F. J. Wegner, Ann. Phys. (N. Y.) [**243**]{}, 1 (1995).
P. W. Brouwer and C. W. J. Beenakker, J. Math. Phys. [**37**]{}, 4904 (1996).
M. L. Mehta, [*Random Matrices*]{} (Academic, New York, 1991).
J. G. G. S. Ramos, A. L. R. Barbosa, and A. M. S. Macêdo, Phys. Rev. B [**84**]{}, 035453 (2011).
For a review, see [*Mesoscopic Phenomena in Solids*]{}, edited by B. L. Altshuler, P. A. Lee, and R. A. Webb (North-Holland, Amsterdam, 1991).
C. Mahaux and H. A. Weidenmüller, [*Shell model approach to nuclear reactions*]{} (North Holland, Amsterdam, 1969).
C. H. Lewenkopf and H. A. Weidenmüller, Ann. Phys. (N.Y.) [**212**]{}, 53 (1991).
J. J. M. Verbaarschot, H. A. Weidenmüller, and M. R. Zirnbauer, Phys. Rep. [**129**]{}, 367 (1985).
A. L. R. Barbosa, J. G. G. S. Ramos, and D. Bazeia, Phys. Rev. B [**84**]{}, 115312 (2011).
P. Stano and Ph. Jacquod, Phys. Rev. Lett. [**106**]{}, 206602 (2011).
K. B. Efetov, Phys. Rev. Lett. [**74**]{}, 2299 (1995).
R. O. Vallejos and C. H. Lewenkopf, J. Phys. A [**34**]{}, 2713 (2001); E. R. P. Alves and C. H. Lewenkopf, Phys. Rev. Lett. [**88**]{}, 256805 (2002).
T. Ericson, Ann. Phys. (NY) [**23**]{}, 390 (1963).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The antiferromagnetic (AF) model is generalized for the quasielectron system composed of identical ionic-covalent dimers. The density-fluctuation and covalent-correlation operators are constructed based on the extended AF density matrices, and the quasielectron system is decomposed into 4-level subsystems for the electron ionization and affinity. By considering the nearest-neighbor hopping near the covalent limit, we can see the importance of the bonding coefficients to the effective mass of the excited carrier in the crystal of the zincblende structure.'
author:
- 'C. F. Huang'
title: 'An extended antiferromagnetic model for quasielectrons in identical ionic-covalent bonds'
---
Introduction
============
Different quasiparticles are taken into account for many-electron systems. Based on the Bardeen Cooper Schrieffer (BCS) theory [@James], the Bogoliubov-BCS quasiparticles are constructed under the superconducting (SC) order by considering Bogoliubov-deGennes or Hartree-Bogoliubov equations [@Ghosal; @Barankov; @Paar]. The antiferromagnetic (AF) quasielectrons are introduced when the orbital antiferromagnetism or the d-density wave (DDW) order becomes important [@Gerami; @Chakravarty], and the quasiparticles under multiple orders are discussed in the literature [@Bena; @Kee; @Huang1; @Laughlin; @Ramshaw]. The Hubbard model [@Mahan] may help us to clarify the SC and AF behaviors, and the 4-level Hubbard dimer [@Matlak1; @Matlak2] is constructed by including the up- and down-spin orbitals at the two sites. The ionic-covalent chemical bond can be approximated by such a dimer when the two sites correspond to the atomic orbitals, and we have the Heitler-London state [@Fulde; @Soos], which is denoted by $| \Psi _{c} \rangle$ in this manuscript, for the covalent limit. The ionic state $| \Psi _{i} \rangle$ may become dominant in the hetero-diatom bond, in which we shall consider different on-site energies introduced in the ionic Hubbard model [@Kampf; @Buzatu].
The Bloch states [@Fulde; @Grosso; @Grafenstein] are important to construct the quasiparticle orbitals, including the plane-wave ones for the nearly free carriers in crystals. Such states can be introduced based on the mean-field methods such as the Hartree-Fock (HF) approach, which yields the self-consistent-field (SCF) solutions [@Fulde; @Grafenstein; @Stoyanova]. It has been discussed in the literature how to include the correlation energy beyond the HF approach by considering the coupled-cluster corrections such as those due to coupled-cluster doubles (CCDs) [@Fulde; @Grafenstein; @Stoyanova; @Doll; @Talukdar]. In addition, my group [@Huang1] has used the AF part of the extended Bogoliubov-BCS quasiparticles to construct the energy form of the one-bond system by considering [@Prasad] $$\begin{aligned}
| \Psi _{b} \rangle = \alpha _{i} | \Psi _{i} \rangle + \alpha _{c} | \Psi _{c} \rangle.\end{aligned}$$ Here $| \Psi _{b} \rangle$ denotes the bonding wavefunction of the half-filled ionic-covalent bond, and the complex numbers $ \alpha _{i}$ and $\alpha _{c}$ are the bonding coefficients satisfying $| \alpha _{i}| ^{2} +|\alpha _{c}| ^{2} =1$. In this manuscript, the AF model is generalized for the compound in which the chemical bonds are identical to the red one in Fig. 1 (a). To include the bonding correlation, the extended AF density matrix $$\begin{aligned}
\rho _{ea} = \left[
\begin{array}{c}
\rho \text{ \ \ } \Delta \\ \Delta \text{ \ \ } \rho
\end{array}
\right] \end{aligned}$$ is introduced to construct the correlation operators $$\begin{aligned}
\begin{cases}
d ^{(1)} = ( \sqrt{2} -1) [ \rho ^{(1)} \rho ^{(2)} ( I - \rho ^{(1)}) + ( I - \rho ^{(1)}) \rho ^{(2)} \rho ^{(1)} ] \\
d ^{(2)} = ( I - \rho ^{(1)}) \rho ^{(2)} ( I - \rho ^{(1)}) \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }
\end{cases}\end{aligned}$$ for the electron pair in the bonding region $\Omega$. Here $\rho$ and $\Delta$ are self-adjoint operators, $I$ denotes the identity operator, and the operators $\rho ^{(1)} = \rho + \Delta$ and $\rho ^{(2)} = \rho - \Delta$ can serve as the density matrices for the quasielectrons in Fig. 2 (a), which shows the 4-level dimer corresponding to the one-bond system. The correlation matrices $d ^{(1)}$ and $d ^{(2)}$ are denoted as the density-fluctuation and covalent-correlation operators because they represent the fluctuating charge and the correlation due to the covalent component, respectively. For convenience, the background is mentioned in section II, and the operators $d ^{(1)} $ and $d ^{(2)}$ are introduced in subsection III-A by considering the non-interacting and interacting parts of the one-bond Hamiltonian. The assumptions about the one-bond dimer are discussed in Appendix A, and an orbital transformation is mentioned in Appendix B to improve my model. The ionized and affinitive processes for the one-bond system are taken into account in subsection IV-A.
The quasielectron model for the binary compound is constructed in section III-B by considering the identical chemical bonds connecting the anions and cations. Figure 1 (b) shows such bonds in the zincblende-structure crystal as an example. The density-fluctuation and covalent-correlation operators are extended as $d ^{( \text{I} )}$ and $d ^{( \text{II} )}$ for the bonding correlation in the quasielectron system of the considered compound. I decompose such a system into the 4-level subsystems shown in Fig. 3 (a), and the energy difference to excite a carrier is obtained in section IV-B by considering the ionized/affinitive process. When the compound forms an ideal crystal following the periodic boundary condition, each subsystem may correspond to a Bloch-type function. Near the ionic limit, as shown in Appendix C, my model can be supported by the coupled-cluster theory. On the other hand, we can see the importance of the bonding coefficients to the excited carrier near the covalent limit under the strong e-e repulsive strength, which is responsible for the Mott insulating behaviors in some AF systems [@Lee1; @Liu]. The Hamiltonian family, which can correspond to the random Schr$\ddot{o}$dinger operators in the random-matrix theory [@Kirsc; @Erdos; @Lee2; @Huang2], is discussed in section V. Actually we may generalize Eq. (2) to include a set of Hamiltonians by constructing the multiple-component quasielectrons. I note that the multiple-component functions can be used to introduce the vector bundles [@Bohm; @Friedman; @Banks] for the gauge theory. The compound system composed of different ionic-covalent dimers are discussed in Appendix D. The summary is made in section VI.
Background
==========
The DDW Hamiltonian $H _{ddw} = \sum _{ {\bf k} \sigma } \chi _{ {\bf k} \sigma } ^\dag B _{ {\bf k} } \chi _{ {\bf k} \sigma } $ [@Gerami] has been introduced for the AF quasielectrons with $$\begin{aligned}
B _{ {\bf k} } = \left[
\begin{array}{c}
\varepsilon _{ {\bf k} } - \mu \text{ \ \ \ \ \ \ \ } \Delta _{ {\bf k} } \text{ \ \ \ \ } \\ \text{ \ \ \ } \Delta _{ {\bf k} } ^{\ast} \text{ \ \ \ \ } \varepsilon _{ {\bf k} + {\bf Q} } - \mu
\end{array}
\right]\end{aligned}$$ and $\chi _{ {\bf k} \sigma } ^{\dag} = ( c _{ {\bf k} \text{ } \sigma } ^{\dag} , - i c _{ {\bf k}+{\bf Q} \text{ } \sigma } ^{\dag} )$. Here $\mu$ denotes the chemical potential, the wave-vector [**k**]{} belongs to the reduced Brillouin zone (RBZ), ${\bf Q }$ equals the DDW ordering wavevector, $\sigma$ denotes the spin orientation $\uparrow$ or $\downarrow$, $\varepsilon _{ {\bf k} }$ and $ \varepsilon _{ {\bf k} + {\bf Q} }$ are the energy coefficients, $ \Delta _{ {\bf k} }$ is the DDW order parameter, and $ c _{ {\bf k} \text{ } \sigma }$ and $ c _{ {\bf k}+{\bf Q} \text{ } \sigma }$ are to annihilate electrons at $({\bf k} , \sigma )$ and $({\bf k} +{\bf Q}, \sigma )$, respectively. At half-filling, the zero-temperature chemical potential may lie in the gap between the two DDW energy bands, under which the gapless quasielectrons appear at the nodal points [@Gerami; @Chakravarty]. The lower band is filled with the AF quasielectrons while the upper one is empty. Each filled eigenket $\varphi_{f}$ can be denoted by the two-component wavefunction $$\begin{aligned}
| \varphi _{ f } \rangle = \left[
\begin{array}{c}
\varphi ^{\prime} _{ f } \\ \varphi ^{\prime \prime} _{ f }
\end{array}
\right]\end{aligned}$$ with the plane waves $\varphi ^{\prime} _{ f } = a _{f} ^{\prime} exp (i {\bf k}_{f} \cdot {\bf r})$ and $\varphi ^{\prime \prime} _{ f } = a _{f} ^{ \prime \prime} exp (i ({\bf k}_{f} + {\bf Q} ) \cdot {\bf r})$. Here ${\bf k}_{f} \in $ RBZ, and the coefficients $ a _{f} ^{\prime}$ and $ a _{f} ^{ \prime \prime}$ satisfying $( a _{f} ^{ \prime \ast }$, $ -i a _{f} ^{ \prime \prime \ast} ) B _{ {\bf k} } \propto ( a _{f} ^{ \prime \ast}$, $ -i a _{f} ^{ \prime \prime \ast} )$ are normalized such that $\langle \varphi _{f} | \varphi _{f} \rangle =1$. We may exchange the components in Eq. (5) to obtain $$\begin{aligned}
T | \varphi _{ f } \rangle = \left[
\begin{array}{c}
\varphi ^{\prime \prime} _{ f } \\ \varphi ^{\prime} _{ f }
\end{array}
\right],\end{aligned}$$ which also represents the ket $\varphi _{ f }$, by the operator $$\begin{aligned}
T = \left[
\begin{array}{c}
0 \text{ \ \ } I \\ I \text{ \ \ } 0
\end{array}
\right]. \end{aligned}$$ The extended AF density matrix $ \sum ( | \varphi _{ f } \rangle \langle \varphi _{ f } | + T | \varphi _{ f } \rangle \langle \varphi _{ f } | T ^{\dag} ) $ is of the form of $\rho _{ea}$ given by Eq. (2) because it commutes with $T$. We can decompose $\rho _{ea}$ as $$\begin{aligned}
\rho _{ea} = \rho ^{(1)} \otimes e _{1} e _{1} ^{\dag} + \rho ^{(2)} \otimes e _{2} e _{2} ^{\dag}. \end{aligned}$$ such that $\rho ^{(1)} = \rho + \Delta = e _{1} ^{\dag} \rho _{ea} e _{1} $ and $\rho ^{(2)} = \rho - \Delta = e _{2} ^{\dag} \rho _{ea} e _{2} $, where $e _{1} ^{\dag} = \frac{1}{ \sqrt{2} } (1,1) $ and $e _{2} ^{\dag} = \frac{1}{ \sqrt{2} } (1,-1) $. The effective Hamiltonian $\hat{H_{D}}$ given by Eq. (49) in Ref. [@Huang1] can be obtained by generalizing $H _{ddw}$ for the extended AF density matrix. In addition to the orbital antiferromagnetism, the order parameter $\Delta _{ {\bf k} }$ is taken into account for d-wave superconductivity [@Chakravarty; @Bena; @Kee]. Moreover, we may use $\rho ^{(1)}$ and $\rho ^{(2)}$ to represent the quasielectrons in the one-bond system [@Huang1].
The 4-level dimer composed of the two spatial sites with up- and down-spin orientations has been introduced by considering the dimer Hamiltonian $H _{t-U} = \sum _{\sigma} (t _{AB} c _{A \sigma} ^{\dag} c _{B \sigma} +t _{AB} ^{*} c _{B \sigma} ^{\dag} c _{A \sigma}) + \hat{U}$ [@Matlak1]. Here $A$ and $B$ represent the two spatial sites, $t _{AB}$ is the hopping coefficient, the annihilators $c _{A \sigma}$ and $c _{B \sigma}$ follow $\{ c _{B \sigma} ,c _{B \sigma ^{\prime} } \} = \{ c _{A \sigma} ,c _{A \sigma ^{\prime} } \} = \{ c _{A \sigma} ,c _{B \sigma ^{\prime} } \} = \{ c _{A \sigma} ,c _{B \sigma ^{\prime} } ^{\dag} \} =0$ and $\{ c _{A \sigma} ,c _{A \sigma ^{\prime} } ^{\dag} \} = \{ c _{B \sigma} ,c _{B \sigma ^{\prime} } ^{\dag} \} = \delta _{ \sigma \sigma ^{\prime} } $, the factor $\sum _{\sigma} (t _{AB} c _{A \sigma} ^{\dag} c _{B \sigma} +t _{AB} ^{*} c _{B \sigma} ^{\dag} c _{A \sigma})$ is responsible for the intra-bond hopping, and $\hat{U}$ denotes effective e-e interaction potential. To model the AF states, we shall note that such states occur as the up- and down-spin electrons repel each other when they are at the same sites. We can take $$\begin{aligned}
\hat{U} = \int d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{ \prime} - {\bf r} ^{ \prime \prime} | ) \psi _{ \uparrow } ^{\dag} ({\bf r} ^{\prime}) \psi _{ \uparrow } ( {\bf r} ^{\prime} ) \psi _{ \downarrow } ^{\dag} ({\bf r} ^{ \prime \prime}) \psi _{ \downarrow } ({\bf r} ^{ \prime \prime})\end{aligned}$$ as the short-range repulsive potential to reduce the double occupancy [@Lee1] such that the electrons prefer $c _{A \uparrow} ^{\dag} c _{B \downarrow} ^{\dag} | 0 \rangle $ and $c _{B \uparrow} ^{\dag} c _{A \downarrow} ^{\dag} | 0 \rangle $ at half filling. Here $| 0 \rangle $ denotes the vacuum state, the nonnegative function $U(|{\bf r}|)$ equals zero as $|{\bf r}|$ exceeds the short interacting length $a$, and $ \psi _{ \sigma } ( {\bf r} )$ is to annihilate the electron with the spin orientation $\sigma$ at position ${\bf r}$. Such a dimer can be used to model the ionic-covalent chemical bond, which separates atom $A$ from atom $B$ in Fig. 1 (a). Assume that the bonding electrons almost concentrate in $\Omega$ such that we can consider the approximation $ \langle {\bf r} | A \rangle = \langle {\bf r} | B \rangle =0$ as $ {\bf r} \notin \Omega$, where the $A$- and $B$-site orbitals $| A \rangle$ and $|B \rangle$ denote the normalized spatial parts of $c _{A \sigma} ^{\dag} | 0 \rangle$ and $c _{ B \sigma} ^{\dag} | 0 \rangle$, respectively. In this manuscript, I assume that $| A \rangle$ and $|B \rangle$ are localized near the corresponding atoms in the considered bond, as shown in Fig. 4. We can construct these site orbitals by considering the linear combinations of the atomic orbitals after truncating the tails outside the bonding region $\Omega$ and performing suitable orthogonalization to have the zero overlap integral [@Levine]. Let $ \mathbb{C} ^{2} _{\Omega} $ be the vector space composed of the linear combinations of $| A \rangle$ and $|B \rangle$. The space $ \mathbb{C} ^{2} _{\Omega} $ is isomorphic to $ \mathbb{C} ^{2} $, and is a subspace of the Hilbert space $ L^{2} _{\Omega} $ composed of the square-integrable functions which equal zero outside $\Omega$. The Heitler-London correlated state $| \Psi _{c} \rangle = \frac{1}{ \sqrt{2} } ( c_{A \uparrow} ^{\dag} c_{B \downarrow} ^{\dag} + c_{B \uparrow} ^{\dag} c_{A \downarrow} ^{\dag}) | 0 \rangle $, a linear combination of the AF states $c_{A \uparrow} ^{\dag} c_{B \downarrow} ^{\dag} | 0 \rangle $ and $c_{B \uparrow} ^{\dag} c_{A \downarrow} ^{\dag} | 0 \rangle $, has been taken into account in the covalent limit [@Fulde; @Soos] when the 4-level bond is half-filled. To include the ionic part, we may consider the ionic Hubbard model [@Kampf; @Buzatu] and modify $H _{t-U}$ as the one-bond Hamiltonian $$\begin{aligned}
H _{b} = \sum _{\sigma} (t _{AB} c _{A \sigma} ^{\dag} c _{B \sigma} + t _{AB} ^{*} c _{B \sigma} ^{\dag} c _{A \sigma} + \varepsilon _{A} c _{A \sigma} ^{\dag} c _{A \sigma} + \varepsilon _{B} c _{B \sigma} ^{\dag} c _{B \sigma}) + \hat{U},\end{aligned}$$ which is suitable to model the polar molecule [@Prasad], to introduce the on-site energies $\varepsilon _{A}$ and $\varepsilon _{B}$. In the above equation, $\sum _{\sigma} (t _{AB} c _{A \sigma} ^{\dag} c _{B \sigma} + t _{AB} ^{*} c _{B \sigma} ^{\dag} c _{A \sigma} + \varepsilon _{A} c _{A \sigma} ^{\dag} c _{A \sigma} + \varepsilon _{B} c _{B \sigma} ^{\dag} c _{B \sigma}) = H _{b} - \hat{U} $ is the non-interacting part of $H _{b}$ while the e-e potential $\hat{U}$ serves as the interacting part. In the following, assume that $\varepsilon _{B} > \varepsilon _{A}$ such that atom $A$ has the higher electronegativity. The two-electron wavefunction becomes the uncorrelated state $ | \Psi _{i} \rangle = c _{A \uparrow} ^{\dag} c _{A \downarrow} ^{\dag} | 0 \rangle$ in the ionic limit, so we shall take both $ | \Psi _{i} \rangle$ and $ | \Psi _{c} \rangle$ in general and consider the bonding wavefunction given by Eq. (1) at half filling. The assumptions about Eq. (10) are discussed in Appendix A. While there exists another state $ | \Psi _{BB} \rangle = c ^{\dagger} _{B \uparrow} c ^{\dagger} _{B \downarrow} |0 \rangle$, as mentioned in Appendix B, its contribution is small and we can include it just by performing an orbital transformation. For convenience, I introduce my model without considering $ | \Psi _{BB} \rangle$ in the main text, and include its contribution in Appendix B by using such a transformation to preserve the form of Eq. (1).
Because the bonding electrons in the covalent limit are described by the linear combination of the AF states $ c_{A \uparrow} ^{\dag} c_{B \downarrow} ^{\dag} | 0 \rangle$ and $ c_{B \uparrow} ^{\dag} c_{A \downarrow} ^{\dag} | 0 \rangle$, it is natural to try the extended AF density matrix $\rho _{ea}$ to construct the quasielectron orbitals of the one-bond system in Fig. 1 (a). In each AF state one electron is located at $ | A \rangle$ while the other one is located at $ |B \rangle$, so we shall take the matrices $\rho ^{(1)}$ and $\rho ^{(2)}$ in Eq. (8) as $| A \rangle \langle A |$ and $| B \rangle \langle B |$ in the covalent limit. On the other hand, both electrons occupy $ | A \rangle$ in the ionic limit and thus these two matrices should equal $| A \rangle \langle A |$ as $| \Psi _{i} \rangle$ is dominated. By taking [@Huang1] $$\begin{aligned}
\rho ^{(1)} = | A \rangle \langle A | \text{ and } \rho ^{(2)} = | L \rangle \langle L | \end{aligned}$$ with $ | L \rangle = \alpha _ {i} | A \rangle + \alpha _{c} | B \rangle$ in $\mathbb{C} ^{2} _{ \Omega }$, the matrix $\rho ^{(1)}$ just represents the quasielectron occupying $| A \rangle$ under both limits. In addition, the matrix $\rho ^{(2)}$ corresponds to the other one jumping to $| A \rangle$ from $| B \rangle$ as $| \Psi _{i} \rangle$ becomes significant. The operator $\rho _{sb} = \frac{1}{2} (\rho ^{(1)} + \rho ^{(2)} + d ^{(1)} )$ satisfies $\langle {\bf r} ^{\prime} | \rho _{sb} | {\bf r} ^{ \prime \prime } \rangle = \langle \Psi _{b} | \psi _{ \uparrow } ^{ \dagger } ({\bf r} ^{ \prime } ) \psi _{ \uparrow } ({\bf r} ^{\prime \prime} ) | \Psi _{b} \rangle = \langle \Psi _{b} | \psi _{ \downarrow } ^{ \dagger } ({\bf r} ^{ \prime } ) \psi _{ \downarrow } ({\bf r} ^{\prime \prime} ) | \Psi _{b} \rangle $ and thus serves as the one-electron density matrix, which is spin-degenerate because $\langle \Psi _{b} | \psi _{ \sigma } ^{ \dagger } ({\bf r} ^{\prime}) \psi _{ - \sigma } ({\bf r}^{\prime \prime} ) | \Psi _{b} \rangle=0$ for $\sigma = \uparrow$ and $\downarrow$.
Quasielectrons in ionic-covalent bonds
======================================
In this section, the energy forms are constructed for the quasielectrons in the half-filled bonding systems. For convenience, I discuss the one-bond system based on the 4-level dimer [@Matlak1] in subsection III-A, and extend the results to the compound system composed of identical dimers [@Matlak2] in subsection III-B.
III-A One-bond system {#iii-a-one-bond-system .unnumbered}
=====================
Consider the two uncorrelated states $| \Phi _{b} \rangle = c _{A \uparrow} ^{\dag} c _{ L \downarrow} ^{\dag} | 0 \rangle = \alpha _{i} | \Psi _{i} \rangle + \alpha _{c} c _{ A \uparrow} ^{\dag} c _{B \downarrow} ^{\dag} | 0 \rangle $ and $| \Phi _{b} ^{\prime} \rangle = c _{L \uparrow} ^{\dag} c _{ A \downarrow} ^{\dag} | 0 \rangle = \alpha _{i} | \Psi _{i} \rangle + \alpha _{c} c _{ B \uparrow} ^{\dag} c _{A \downarrow} ^{\dag} | 0 \rangle $, which can be obtained from $| \Psi _{b} \rangle$ by substituting the AF states $ c _{ A \uparrow} ^{\dag} c _{B \downarrow} ^{\dag} | 0 \rangle $ and $c _{ B \uparrow} ^{\dag} c _{A \downarrow} ^{\dag} | 0 \rangle$, respectively, for the covalent component $| \Psi _{c} \rangle$ in Eq. (1). Here $c _{L \sigma}$ is the operator to annihilate the electron with the spin orientation $\sigma$ at $| L \rangle$ in the bonding region $\Omega$ in Fig. 1 (a). The matrices $\rho ^{(1)}$ and $\rho ^{(2)}$ in Eq. (11) correspond to the up- and down-spin electrons in $| \Phi _{b} \rangle$, and correspond to the down- and up-spin ones in $ | \Phi _{b} ^{\prime} \rangle$. So $\rho ^{(1)}$ and $\rho ^{(2)}$ are for the quasielectrons with the opposite spin orientations in the one-bond system, and can represent the occupied levels of the half-filled 4-level dimer in Fig. 2 (a) if $| 1 \rangle \otimes | \sigma \rangle = |A \rangle \otimes | \sigma \rangle $ and $| 2 \rangle \otimes | - \sigma \rangle = | L \rangle \otimes | - \sigma \rangle $. The two unoccupied levels $| \bar{1} \rangle \otimes | \sigma \rangle$ and $| \bar{2} \rangle \otimes | - \sigma \rangle$ in Fig. 2(a) serve as $ | B \rangle \otimes | \sigma \rangle$ and $| \bar{L} \rangle \otimes | - \sigma \rangle$ in such a one-bond system, where $| \bar{L} \rangle = \alpha _{c} ^{*} | A \rangle - \alpha _{i} ^{*} | B \rangle \bot | L \rangle$ is a ket in $\mathbb{C} ^{2} _{\Omega}$. The uncorrelated energy $\langle \Phi _{b} | H _{b} | \Phi _{b} \rangle = \langle \Phi ^{\prime} _{b} | H _{b} | \Phi _{b} ^{\prime} \rangle$ equals $$\begin{aligned}
tr (\rho ^{(1)} + \rho ^{(2)}) H_{sb} + \int _{ {\bf r} ^{ \prime },{\bf r} ^{ \prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{ \prime \prime } U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime} \rangle \langle {\bf r} ^{\prime \prime} | \rho ^{(2)} | {\bf r} ^{\prime \prime} \rangle,\end{aligned}$$ where $H_{sb} = t _{AB} | A \rangle \langle B | + t _{AB} ^{*} | B \rangle \langle A | + \varepsilon _{A} | A \rangle \langle A | + \varepsilon _{B} | B \rangle \langle B |$ results from the non-interacting part of $H_{b}$. The factors $t _{AB} | B \rangle \langle A | + t _{AB} ^{*} | A \rangle \langle B |$ and $\varepsilon _{A} | A \rangle \langle A | + \varepsilon _{B} | B \rangle \langle B |$ of $H _{sb}$ are responsible for the intra-bond hopping and on-site energy difference, respectively.
While we may construct the qausielectron density matrices by Eq. (11), the wavefunction $| \Psi _{b} \rangle$ in Eq. (1) is different from the uncorrelated functions $| \Phi _{b} \rangle$ and $| \Phi _{b} ^{\prime} \rangle$ when $ \alpha _{c} \neq 0$. To obtain the bonding energy $E _{b} = \langle \Psi _{b} | H _{b} | \Psi _{b} \rangle$, therefore, we shall include the correlation contributions corresponding to the blue dash curve in Fig. 2 (a) by calculating the difference $\langle \Psi _{b} | H _{b} | \Psi _{b} \rangle - \langle \Phi _{b} | H _{b} | \Phi _{b} \rangle = \langle \Psi _{b} | H _{b} | \Psi _{b} \rangle - \langle \Phi ^{\prime} _{b} | H _{b} | \Phi _{b} ^{\prime} \rangle$. The non-interacting part of $H_{b}$ induces the correlation energy $ tr H_{sb} d ^{(1)} = \langle \Psi _{b} | (H _{b} - \hat{U}) | \Psi _{b} \rangle - \langle \Phi _{b} | (H _{b} - \hat{U}) | \Phi _{b} \rangle = \langle \Psi _{b} | (H _{b} - \hat{U}) | \Psi _{b} \rangle - \langle \Phi ^{\prime} _{b} | (H _{b} - \hat{U}) | \Phi ^{\prime} _{b} \rangle $. We can take $I = | A \rangle \langle A| + | B \rangle \langle B | $, which is the identity operator on $\mathbb{C} ^{2} _{\Omega}$, in Eq. (3) to introduce the density-fluctuation operator $ d ^{(1)} $. The operator $d ^{(1)}$ results from the factor $\alpha _{i} \alpha _{c} ^{*} \langle \Psi _{c} | (H _{b} - \hat{U}) | \Psi _{i} \rangle + \alpha _{c} \alpha _{i} ^{*} \langle \Psi _{i} | (H _{b} - \hat{U}) | \Psi _{c} \rangle $ when we calculate $\langle \Psi _{b} | (H _{b} - \hat{U}) | \Psi _{b} \rangle $, and becomes zero if the ionic and covalent components do not coexist. The up- or down-spin density $\langle {\bf r} | \rho _{sb} | {\bf r} \rangle $ at position ${\bf r}$ includes the factor $\langle {\bf r} | d ^{(1)} | {\bf r} \rangle$, which follows $\int _{ {\bf r} \in \Omega} d ^{3} r \langle {\bf r} | d ^{(1)} | {\bf r} \rangle=tr d ^{(1)} = ( \sqrt{2} -1) tr [ ( I - \rho ^{(1)}) \rho ^{(1)} \rho ^{(2)} + \rho ^{(2)} \rho ^{(1)} ( I - \rho ^{(1)}) ]=0 $ because $( I - \rho ^{(1)}) \rho ^{(1)}= \rho ^{(1)} ( I - \rho ^{(1)})=0$ under Eq. (11). Therefore, $\langle {\bf r} | d ^{(1)} | {\bf r} \rangle$ has no contribution to the total charge $2 \int _{ {\bf r} \in \Omega} d ^{3} r \langle {\bf r} | \rho _{sb} | {\bf r} \rangle $, and represents the fluctuating charge density due to the coexistence of the ionic and covalent components. I note that $\delta \rho$, the deviation from the average density [@Zhang] due to the correlation in the quantum Hall effect [@DHLee; @You], also has no contribution to the total charge.
To include the correlation energy due to $\hat{U}$ in the quasielectron space, we can further introduce $d ^{(2)}$ to rewrite $ \langle \Psi _{b} | \hat{U} | \Psi _{b} \rangle - \langle \Phi _{b} | \hat{U} | \Phi _{b} \rangle$ or $\langle \Psi _{b} | \hat{U} | \Psi _{b} \rangle - \langle \Phi ^{\prime} _{b} | \hat{U} | \Phi _{b} ^{\prime} \rangle$ as $\int _{ {\bf r} ^{\prime},{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) [ \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime \prime } \rangle \langle {\bf r} ^{\prime \prime} | d ^{(2)} | {\bf r} ^{ \prime } \rangle + \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime } \rangle \langle {\bf r} ^{\prime \prime} | d ^{(1)} | {\bf r} ^{\prime \prime } \rangle] $. We can take $I= |A\rangle \langle A| + | B \rangle \langle B| $ in Eq. (3) to introduce $d ^{(2)}$ on $\mathbb{C} ^{2} _{ \Omega} $. The operator $d ^{(2)}$ comes from the factor $ | \alpha _{c} | ^{2} \langle \Psi _{c} | \hat{U} | \Psi _{c} \rangle$ when we calculate $\langle \Psi _{b} | \hat{U} | \Psi _{b} \rangle$, and becomes zero when $\alpha _{c} =0$. Hence $d ^{(2)}$ represents the covalent correlation, and we shall include both $d ^{(1)}$ and $d ^{(2)}$ for the blue dash curve in Fig. 2 (a). By including the correlation contributions, we have $$\begin{aligned}
E_{b} = tr (\rho ^{(1)} + \rho ^{(2)}) H_{sb} + \int _{ {\bf r} ^{ \prime },{\bf r} ^{ \prime \prime } \in \Omega} d ^{3} r ^{ \prime } d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime} \rangle \langle {\bf r} ^{\prime \prime} | \rho ^{(2)} | {\bf r} ^{\prime \prime} \rangle +\end{aligned}$$ $$tr H_{sb} d ^{(1)} + \int _{ {\bf r} ^{\prime},{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime} \rangle \langle {\bf r} ^{\prime \prime} | d ^{(1)} | {\bf r} ^{\prime \prime} \rangle + \text{ \ \ \ \ \ }$$ $$\int _{ {\bf r} ^{ \prime },{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime \prime} \rangle \langle {\bf r} ^{\prime \prime} | d ^{(2)} | {\bf r} ^{\prime} \rangle \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }$$ from Eq. (12) when the ionic-covalent chemical bond in Fig. 1 (a) is half-filled. In the above equation, the factors in the second and third lines are due to the bonding correlation. The above equation provides the energy form of the density matrix $\rho _{ea}$, which can be decomposed into $\rho ^{(1)} = e _{1} ^{\dag} \rho _{ea} e _{1}$ and $\rho ^{(2)} = e _{2} ^{\dag} \rho _{ea} e _{2}$, for the one-bond system.
III-B Compound system {#iii-b-compound-system .unnumbered}
=====================
Consider the binary compound where the anions and cations are connected by the identical ionic-covalent bonds, and assume that all the bonds are well-separated without overlap. Figure 1 (b) shows such bonds in the zincblende-structure crystal [@Grosso], in which each atom provides 4 site orbitals, as an example. For convenience, I parameterize these bonds by the integer parameter $j=1 \sim N$ and denote the bonding region of the $j$-th bond as $\Omega _{j}$, where $N$ is the total number of the bonds. Each bond in the compound is a 4-level dimer just as the one-bond system in Fig. 1 (a). Assume that there exists the one-to-one mapping ${\bf R} _{j}$ to relate any position ${\bf r} _{j} \in \Omega _{j}$ in the $j$-th bond to ${\bf r} \in \Omega$ in Fig. 1 (a) by ${\bf R} _{j} ({\bf r}) = {\bf r} _{j}$ and ${\bf R} _{j} ^{-1} ({\bf r} _{j}) = {\bf r}$ such that the distance $|{\bf r}_{1}-{\bf r} _{2}|$ between ${\bf r}_{1}$ and ${\bf r}_{2} \in \Omega$ equals $|{\bf R} _{j} ( {\bf r} _{1} )- {\bf R} _{j}({\bf r} _{2})| $ for all $j$. Therefore, every ionic-covalent chemical bond in the compound is identical to that in Fig. 1 (a). Let $| A, j \rangle $ and $| B, j \rangle $ as the kets mapped from $| A \rangle $ and $| B \rangle $ under ${\bf R}_{j}$, respectively. The space $\mathbb{C} ^{2} _{ \Omega _{j} }$ spanned by $| A, j \rangle $ and $| B, j \rangle $ is a subspace of the Hilbert space $L^{2} _{ \Omega _{j} }$ composed of square-integrable functions which equal zero outside $\Omega _{j}$, and any two kets in $L^{2} _{ \Omega _{l} }$ and $L^{2} _{ \Omega _{p} }$ are orthogonal to each other when $l \neq p$. We can introduce the space $\mathbb{C} ^{N} \otimes \mathbb{C} ^{2} _{\Omega} = \mathbb{C} ^{2} _{\Omega _{1} } \oplus \mathbb{C} ^{2} _{\Omega _{2} } \oplus \textellipsis \oplus \mathbb{C} ^{2} _{\Omega _{N} }$ for the compound, and choose a set of orthonormal basis $\{ | w _{j} \rangle \}$ in $\mathbb{C} ^{N}$ to represent $| A, j \rangle $ and $| B, j \rangle $ by $|w _{j} \rangle \otimes | A \rangle $ and $|w _{j} \rangle \otimes | B \rangle $, respectively. The space $\mathbb{C} ^{N} \otimes \mathbb{C} ^{2} _{\Omega}$ is a subspace of $\mathbb{C} ^{N} \otimes L ^{2} _{\Omega} = L ^{2} _{\Omega _{1} } \oplus L ^{2} _{\Omega _{2} } \oplus \textellipsis \oplus L ^{2} _{\Omega _{N} }$, and the position ket $| {\bf r} _{j} \rangle$ corresponding to the position ${\bf r} _{j} \in \Omega _{j} $ is taken as $|w _{j} \rangle \otimes | {\bf R} _{j} ^{-1} ({\bf r} _{j}) \rangle $ to perform the integral in $\mathbb{C} ^{N} \otimes L ^{2} _{\Omega}$. For any function $F _{\Omega}$ defined on $\Omega$ and the position ${\bf r} _{j _{1}} \in {\Omega}_{j _{1}} $, we have $(\langle {\bf r} _{j _{1} } |)(| w _{ j _{2} } \rangle \otimes | F _{\Omega} \rangle) = ( \langle w _{ j _{1} } | \otimes \langle {\bf R} _{ j _{1} } ^{ -1 } ( {\bf r} _{ j _{1} } ) | )( | w _{ j _{2} } \rangle \otimes | F _{\Omega} \rangle ) = \delta _{ j _{1} , j _{2} } \langle {\bf R} _{ j _{1} } ^{ -1 } ( {\bf r} _{ j _{1} } ) | F _{\Omega} \rangle = \delta _{ j _{1} , j _{2} } F _{\Omega} ({\bf R} _{ j _{1} } ^{ -1 } ( {\bf r} _{ j _{1} } ) )$.
By mapping the one-bond system in Fig. 1 (a) to the identical bonds in the compound, at half filling we can transfer $ \rho ^{(1)}$ and $ \rho ^{(2)}$ in Eq. (11) to the $j$-th bond and obtain $ \rho _{ w _{j} } ^{(1)} = | w _{j} \rangle \langle w _{j} | \otimes |A \rangle \langle A |$ and $\rho _{ w _{j} } ^{(2)} = | w _{j} \rangle \langle w _{j} | \otimes |L \rangle \langle L|$ for the quasielectrons at $ | A, j \rangle$ and $ | L, j \rangle $. Here $ | L, j \rangle = \alpha _{i} | A, j \rangle + \alpha _{c} | B, j \rangle$. The matrices $$\begin{aligned}
\rho ^{( \text{I} )} = \sum _{j=1} ^{N} \rho _{ w _{j} } ^{(1)} \text{ \ \ and \ \ } \rho ^{( \text{II} )} = \sum _{j=1} ^{N} \rho _{ w _{j} } ^{(2)}\end{aligned}$$ for the half-filled compound are of the opposite spin orientations just as $ \rho ^{(1)}$ and $ \rho ^{(2)}$ in the one-bond system, and the corresponding extended AF density matrix is $\rho ^{( \text{I} )} e _{1} e _{1} ^{\dag} + \rho ^{( \text{II} )} e _{2} e _{2} ^{\dag}$. In fact, we can take $I _{ \mathbb{C} ^{N} } $ as the identity operator on $\mathbb{C} ^{N}$ and rewrite Eq. (14) by $\rho ^{( \text{I} )} = I _{ \mathbb{C} ^{N} } \otimes \rho ^{(1)} $ and $\rho ^{( \text{II} )} = I _{ \mathbb{C} ^{N} } \otimes \rho ^{(2)} $ to see that $\rho ^{( \text{I} )}$ and $\rho ^{( \text{II} )}$ are the natural extensions of $ \rho ^{(1)}$ and $ \rho ^{(2)}$, respectively. To include the bonding correlation due to the $j$-th bond, we shall substitute $ d _{ w _{j} } ^{(1)} = | w _{j} \rangle \langle w _{j} | \otimes d ^{(1)}$ and $ d _{ w _{j} } ^{(2)} = | w _{j} \rangle \langle w _{j} | \otimes d ^{(2)}$ for the operators in Eq. (3). The correlation operators $d ^{( \text{I} )}= \sum _{j} d _{ w _{j} } ^{(1)} = I _{ \mathbb{C} ^{N} } \otimes d ^{(1)}$ and $d ^{( \text{II} )}= \sum _{j} d _{ w _{j} } ^{(2)} = I _{ \mathbb{C} ^{N} } \otimes d ^{(2)}$ satisfy $$\begin{aligned}
\begin{cases}
d ^{( \text{I} )} = ( \sqrt{2} -1) [ \rho ^{( \text{I} )} \rho ^{( \text{II} )} ( I - \rho ^{( \text{I} )}) + ( I - \rho ^{( \text{I} )}) \rho ^{( \text{II} )} \rho ^{( \text{I} )} ] \\
d ^{( \text{II} )} = ( I - \rho ^{( \text{I} )}) \rho ^{( \text{II} )} ( I - \rho ^{( \text{I} )}) \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }
\end{cases},\end{aligned}$$ and the matrix $\rho _{sC} = \frac{1}{2}(\rho ^{( \text{I} )} + \rho ^{( \text{II} )} +d ^{( \text{I} )} ) = I _{ \mathbb{C} ^{N} } \otimes \rho _{sb}$ yields the density $Q({\bf r} _{j} ) = 2 \langle {\bf r} _{j} | \rho _{sC} | {\bf r} _{j} \rangle = 2 \langle {\bf R} _{j} ^{-1} ({\bf r} _{j}) | \rho _{sb} | {\bf R} _{j} ^{-1} ({\bf r} _{j}) \rangle $ at ${\bf r} _{j} \in \Omega _{j}$. In the above equation, the matrix $I$ denotes the identity operator on $\mathbb{C} ^{N} \otimes \mathbb{C} ^{2} _{\Omega} $, and we can interpret $d ^{( \text{I} )}$ and $d ^{( \text{II} )}$ as the density-fluctuation and covalent-correlation operators of the quasielectron system in the compound because they serve as $d ^{(1)}$ and $d ^{(2)}$. For the ideal crystal, we may discuss the correlation under the crystal symmetry imposed on $ \rho ^{( \text{I} )}$ and $ \rho ^{( \text{II} )}$ based on Eq. (15).
When the distances between different bonds are larger than $a$, the interacting length of $\hat{U}$, there is no inter-bond e-e interaction in the half-filled compound. Therefore, the two quasielectrons in a specific bond only interact with each other just as those in the one-bond system in Fig. 1 (a), and the e-e energy term is composed of $$\text{(i) \ \ } \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime} _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | \rho ^{( \text{II} )} | {\bf r} ^{\prime \prime} _{ j } \rangle$$ $$\begin{aligned}
\text{(ii) \ } \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime} _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | d ^{( \text{I} )} |{\bf r} ^{\prime \prime} _{ j } \rangle \end{aligned}$$ $$\text{(iii) } \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime \prime } _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | d ^{( \text{II} )} | {\bf r} ^{\prime} _{ j } \rangle. \text{ \ }$$ Equation (16)-(i) corresponds to the second term in Eq. (12) and can be obtained without considering the bonding correlation. On the other hand, Eqs. (16)-(ii) and (16)-(iii) correspond to the last two terms in Eq. (13) and provide the correlation contributions.
To include the energy resulting from the non-interacting term $H _{sb}$, we shall consider the factor $2 tr \rho _{sC} ( \sum _{j=1} ^{N} | w _{j} \rangle \langle w _{j} | \otimes H_{sb} ) = 2 tr \rho _{sb} H _{sb} \times N$ in the compound system. In addition to the intra-bond hopping in $H _{sb}$, the inter-bond hopping $$\begin{aligned}
H _{hop} = \sum _{ j \neq j ^{\prime} } t _{ j \xi , j ^{\prime} \xi ^{\prime} } | w _{ j } \rangle \langle w _{j ^{\prime} } | \otimes | \xi \rangle \langle \xi ^{\prime} |\end{aligned}$$ should be taken into account to relate different chemical bonds [@Matlak2]. Here $ | \xi \rangle \text{ and } | \xi ^{\prime} \rangle \in \{ | A \rangle , |B \rangle \} $, and each coefficient $t _{ j \xi , j ^{\prime} \xi ^{\prime} } = t^{*} _{j ^{\prime} \xi ^{\prime}, j \xi } $ is for the jump from $| w _{ j ^{\prime} } \rangle \otimes | \xi ^{\prime} \rangle$ to $| w _{ j } \rangle \otimes | \xi \rangle$. Therefore, we shall introduce the non-interacting Hamiltonian $H _{sC} = H _{hop}+ \sum _{j=1} ^{N} | w _{j} \rangle \langle w _{j} | \otimes H_{sb} $ and include the energy $2 tr \rho _{sC} H _{sC}$. In this manuscript I consider the short-range hopping, so $t _{ j \xi , j ^{\prime} \xi ^{\prime} }=0$ if the distance between the $j$-th and $j^{\prime}$-th bonds is longer than a specific length. The energy for the half-filled compound system is $$E_{Cr} = tr ( \rho ^{ ( \text{I} ) } + \rho ^{ ( \text{II} ) } ) H _{sC} + \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime} _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | \rho ^{( \text{II} )} | {\bf r} ^{\prime \prime} _{ j } \rangle +$$ $$\begin{aligned}
trH _{sC} d^{ ( \text{I} )} + \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime} _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | d ^{( \text{I} )} |{\bf r} ^{\prime \prime} _{ j } \rangle +\end{aligned}$$ $$\sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime \prime } _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | d ^{( \text{II} )} | {\bf r} ^{\prime} _{ j } \rangle. \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }$$
In addition to $\{ | w _{j} \rangle \}$, we can choose another orthonormal complete set $ \{ | \eta _{j} \rangle \}$ in $ \mathbb{C} ^{N}$ to describe the quasielectron system of the compound. For an example, it is important to choose the set composed of the Bloch-type functions in $\mathbb{C} ^{N}$ when the considered compound is an ideal crystal following the periodic boundary condition. The matrices in Eqs. (14) and (15) can be rewritten as $\rho ^{( \text{I} )} = \sum _{j=1} ^{N} \rho _{ \eta _{j} } ^{(1)}$, $\rho ^{( \text{II} )} = \sum _{j=1} ^{N} \rho _{ \eta _{j} } ^{(2)}$, $d ^{( \text{I} )} = \sum _{j=1} ^{N} d_{ \eta _{j} } ^{(1)}$, and $d ^{( \text{II} )} = \sum _{j=1} ^{N} d_{ \eta _{j} } ^{(2)}$, where $\rho _{ \eta _{j} } ^{(1)} = | \eta _{j} \rangle \langle \eta _{j} | \otimes | A \rangle \langle A |$, $\rho _{ \eta _{j} } ^{(2)} = | \eta _{j} \rangle \langle \eta _{j} | \otimes | L \rangle \langle L |$, $d _{ \eta _{j} } ^{(1)} = | \eta _{j} \rangle \langle \eta _{j} | \otimes d ^{(1)}$, and $d _{ \eta _{j} } ^{(2)} = | \eta _{j} \rangle \langle \eta _{j} | \otimes d ^{(2)}$. Every $| \eta _{j} \rangle$ can correspond to the 4-level dimer in Fig. 2 (a) if we take $\rho _{ \eta _{j} } ^{(1)} $ and $\rho _{ \eta _{j} } ^{(2)}$ as the two quaielectrons at $| 1\rangle = | \eta _{j} \rangle \otimes | A \rangle $ and $| 2 \rangle = | \eta _{j} \rangle \otimes | L \rangle $. The spatial parts of the two empty orbitals are $| \bar{1} \rangle = | \eta _{j} \rangle \otimes |B \rangle $ and $ | \bar{2} \rangle = | \eta _{j} \rangle \otimes | \bar{L} \rangle $, respectively. The operators $d _{ \eta _{j} } ^{(1)}$ and $d _{ \eta _{j} } ^{(2)}$ are determined by $\rho _{ \eta _{j} } ^{(1)} $ and $\rho _{ \eta _{j} } ^{(2)}$ because $d _{ \eta _{j} } ^{(1)} = ( \sqrt{2} -1) [ \rho _{ \eta _{j} } ^{(1)} \rho _{ \eta _{j} } ^{(2)} ( I - \rho _{ \eta _{j} } ^{(1)}) + ( I - \rho _{ \eta _{j} } ^{(1)}) \rho _{ \eta _{j} } ^{(2)} \rho _{ \eta _{j} } ^{(1)} ]$ and $d _{ \eta _{j} } ^{(2)} = ( I- \rho _{ \eta _{j} } ^{(1)} )\rho _{ \eta _{j} } ^{(2)} ( I - \rho _{ \eta _{j} } ^{(1)}) $, and we can interpret $d _{ \eta _{j} } ^{(1)}$ and $d _{ \eta _{j} } ^{(2)}$ as the correlation contributions of the two quasielectrons in subsystem $\eta _{j}$. Therefore, the half-filled quasielectron system to model the compound are decomposed into the 4-level subsystems as shown in Fig. 3 (a), where the blue dash curves denote the corresponding correlation contributions. In addition, the two quasielectrons in each subsystem $\eta _{j}$ are correlated just as those described by $\rho ^{(1)}$ and $\rho ^{(2)}$ in the one-bond system.
Electron affinity and ionization
================================
To model the ionized and affinitive processes, we shall consider how to remove and/or add one quasielectron to excite the carrier. For convenience, first I focus on the one-bond system in subsection IV-A. The assumptions about the one-bond Hamiltonian $H _{b}$ are discussed in Appendix A. Secondly, I consider only the change of the 4-level subsystem $\eta ^{\prime}$ in subsection IV-B to add and/or remove one quasielectron in the compound system, and discuss the excited carrier near the covalent limit to see the importance of bonding coefficients.
IV-A Electron affinity and ionization of the one-bond system {#iv-a-electron-affinity-and-ionization-of-the-one-bond-system .unnumbered}
============================================================
The one-bond system in Fig. 1 (a) is taken as 4-level dimer to model the ionic-covalent bonding. When one quasielectron is removed from such a dimer, the remained one occupying $| 1 \rangle \otimes | \sigma \rangle $ in Fig. 2 (b) has no correlated partner. Because atom A has the higher electro-negativity, we can approximate the remained quasielectron by $\rho ^{(1)}$ in Eq. (11) and obtain the energy $E_{b} ^{(-)} = tr H _{sb} \rho ^{(1)}$ as $| \Psi _{b} \rangle$ becomes the uncorrelated one-electron state $| \Psi _{b,1, \uparrow } \rangle = c _{ A \uparrow } ^{\dag} | 0 \rangle$ or $| \Psi _{b,1, \downarrow } \rangle = c _{ A \downarrow } ^{\dag} | 0 \rangle$. On the other hand, there are three quasielectrons when we change $| \Psi _{b} \rangle$ to a three-electron state, which corresponds to the left-hand side of Fig. 2(c), by the affinitve process. We can see from Fig. 2 (c) that such a three-electron state is equivalent to the one-hole state because there are only 4 levels. The remained quasihole is located at $| \bar{1} \rangle \otimes | \sigma \rangle$, and its spatial part $| \bar{1} \rangle$ can be approximated by $ | B \rangle $ in $\mathbb{C}^{2} _{\Omega}$ because atom $B$ is of the lower electro-negativity. So we can take $| \Psi _{b,3, \sigma } \rangle = c _{ B \sigma} | Fb \rangle \propto c _{ B \underline{ \sigma } } ^{ \dag } c _{ A \sigma } ^{\dag} c _{ A \underline{\sigma} } ^{\dag} | 0 \rangle \propto c _{ L \underline{\sigma} } ^{\dag} c _{ A \sigma } ^{\dag} c _{ \bar{L} \underline{\sigma} } ^{\dag} | 0 \rangle$ as the wavefunction of the three-electron or one-hole state for $\sigma = \uparrow$ or $\downarrow$, and the added electron is located at the spatial ket $| \bar{L} \rangle \in \mathbb{C}^{2} _{\Omega}$. Here $ \underline{ \sigma } = - \sigma$, $c _{ \bar{L} \underline{ \sigma } } $ is to annihilate the electron at $ | \bar{L} \rangle \otimes | - \sigma \rangle $, and $| Fb \rangle$ denotes the four-electron state for the filled one-bond system. The state $| \Psi _{b,3, \sigma } \rangle$ is uncorrelated, and its density matrices for $\sigma$ and $- \sigma $ are $\rho ^{(1)} = | A \rangle \langle A |$ and $ \rho ^{(2)} _{+} = | A \rangle \langle A | + | B \rangle \langle B | = | L \rangle \langle L | + | \bar{L} \rangle \langle \bar{L} |$, respectively. Hence the energy for the one-bond system becomes $E ^{(+)} _{b} =tr (\rho ^{(1)} + \rho ^{(2)} _{+})H_{sb} + \int _{ {\bf r} ^{ \prime },{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime} \rangle \langle {\bf r} ^{\prime \prime} | \rho ^{(2)} _{+} | {\bf r} ^{\prime \prime} \rangle$ after we add one electron.
By taking $ \rho ^{(2)} _{-} = d ^{(1)} _{ \pm } = d ^{(2)} _{ \pm } =0$, we can rewrite $E^{( \pm )} _{b}$ as $$\begin{aligned}
E_{b} ^{( \pm )}= tr (\rho ^{(1)} + \rho ^{(2)} _{ \pm }) H_{sb} + \int _{ {\bf r} ^{\prime},{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime} \rangle \langle {\bf r} ^{\prime \prime} | \rho ^{(2)} _{ \pm } | {\bf r} ^{\prime \prime} \rangle +\end{aligned}$$ $$tr H_{sb} d ^{(1)} _{ \pm } + \int _{ {\bf r} ^{\prime},{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime} \rangle \langle {\bf r} ^{\prime \prime} | d ^{(1)} _{ \pm } | {\bf r} ^{\prime \prime} \rangle + \text{ \ \ \ \ \ }$$ $$\int _{ {\bf r} ^{\prime},{\bf r} ^{\prime \prime} \in \Omega} d ^{3} r ^{\prime} d ^{3} r ^{\prime \prime} U ( | {\bf r} ^{\prime} - {\bf r} ^{\prime \prime} | ) \langle {\bf r} ^{\prime} | \rho ^{(1)} | {\bf r} ^{\prime \prime} \rangle \langle {\bf r} ^{\prime \prime} | d ^{(2)} _{ \pm } | {\bf r} ^{\prime} \rangle. \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }$$ The above equation can be obtained from Eq. (13) by substituting $ \rho ^{(2)} _{ \pm } $, $ d ^{(1)} _{ \pm }$, and $ d ^{(2)} _{ \pm } $ for $ \rho ^{(2)}$, $ d ^{(1)} $, and $ d ^{(2)} $, respectively. The meaning of $d ^{(1)} _{ \pm } = d ^{(2)} _{ \pm } =0$ is that there is no fluctuating charge or covalent correlation after the ionized and affinitive processes. The remained quasielectron in Fig. 2 (b) has no correlated partner, so it is natural that $d ^{(1)} _{ - } = d ^{(2)} _{ - } =0$. On the other hand, only one quasihole is left at the right-hand side of Fig. 2 (c), and the corresponding one-hole state should be similar to the one-electron state in Fig. 2 (b) based on the electron-hole symmetry. Hence it is reasonable that $d ^{(1)} _{ + } = d ^{(2)} _{ + } =0$.
IV-B Electron affinity and ionization of the compound system {#iv-b-electron-affinity-and-ionization-of-the-compound-system .unnumbered}
============================================================
In subsection III-B, the quasielectron system to model the considered compound is decomposed into 4-level subsystems, as shown in Fig. 3 (a). To remove (add) one quasielectron from (to) the subsystem characterized by $| \eta ^{ \prime } \rangle \in \{ | \eta _{j} \rangle \}$, I note that $| \eta ^{\prime } \rangle \otimes | A \rangle$ and $| \eta ^{\prime } \rangle \otimes | B \rangle$ in subsystem $\eta ^{\prime}$ serve as $| A \rangle$ and $|B \rangle$ in the one-bond system, respectively. Therefore, we shall introduce $\rho ^{(2)} _{ \eta ^{ \prime} , - } =0$ and $\rho ^{(2)} _{ \eta ^{ \prime} , + } = | \eta ^{ \prime } \rangle \langle \eta ^{ \prime } | \otimes ( | A \rangle \langle A | + | B \rangle \langle B |) $ for the electron ionization and affinity just as how we introduce $\rho ^{(2)} _{ - }$ and $\rho ^{(2)} _{+} $ according to the electronegativities in subsection IV-A. The ionized quasielectron is removed from $| \eta ^{\prime } \rangle \otimes | L \rangle$ while one quasielectron enters $| \eta ^{\prime } \rangle \otimes | \bar{L} \rangle$ in the affinitive process, and the removed/added charge in the j-th chemical bond equals $| \langle w _{j} | \eta ^{\prime} \rangle | ^{2}$. Together with $\rho ^{(1)} _{ \eta ^{\prime}} $, the matrix $\rho ^{(2)} _{ \eta ^{ \prime} , - } $ corresponds to the one-electron state in Fig. 2 (b) while $\rho ^{(2)} _{ \eta ^{ \prime} , + } $ corresponds to the three-electron or one-hole state in Fig. 2 (c). Because the one- and three-electron states are both uncorrelated, we shall take $d ^{(1)} _{ \eta ^{\prime} , \pm } = d ^{(2)} _{ \eta ^{\prime} , \pm } =0$ to remove the correlation contribution of the subsystem $ \eta ^{\prime}$. If each subsystem characterized by $ \eta _{j} \neq \eta ^{\prime} $ remains unchanged, as shown in Fig. 3 (b), we shall replace $\rho ^{( \text{II} )}$, $d ^{( \text{I} )}$, and $d ^{( \text{II} )}$ by $ \rho ^{( \text{II} )} _{ \eta ^{\prime} , \pm } = \sum _{ \eta _{j} \neq \eta ^{\prime} } \rho ^{(2)} _{ \eta _{ j } } +\rho ^{(2)} _{ \eta ^{ \prime } , \pm } $, $d ^{( \text{I} )} _{ \eta ^{\prime} , \pm } = \sum _{ \eta _{j} \neq \eta ^{\prime} } d ^{(1)} _{ \eta _{ j } } +d ^{(1)} _{ \eta ^{ \prime} , \pm } $, and $d ^{( \text{II} )} _{ \eta _{\prime} , \pm } = \sum _{ \eta _{j} \neq \eta ^{\prime} } d ^{(2)} _{ \eta _{j} } +d ^{(2)} _{ \eta ^{ \prime} , \pm } $, respectively. The energy $E _{Cr} $ becomes $$E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } = tr ( \rho ^{ ( \text{I} ) } + \rho ^{ ( \text{II} ) } _{ \eta ^{\prime}, \pm } ) H _{sC} + \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime} _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | \rho ^{( \text{II} )} _{ \eta ^{\prime}, \pm} | {\bf r} ^{\prime \prime} _{ j } \rangle +$$ $$\begin{aligned}
trH _{sC} d^{ ( \text{I} )} _{ \eta ^{\prime}, \pm} + \sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime} _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | d ^{( \text{I} )} _{ \eta ^{\prime}, \pm} |{\bf r} ^{\prime \prime} _{ j } \rangle +\end{aligned}$$ $$\sum _{j=1} ^{N} \int _{ {\bf r} ^{\prime} _{ j },{\bf r} ^{\prime \prime} _{ j } \in \Omega _{j} } d ^{3} r ^{\prime} _{ j } d ^{3} r ^{\prime \prime} _{ j } U ( | {\bf r} ^{\prime} _{ j } - {\bf r} ^{\prime \prime} _{ j } | ) \langle {\bf r} ^{\prime} _{ j } | \rho ^{( \text{I} )} | {\bf r} ^{\prime \prime } _{ j } \rangle \langle {\bf r} ^{\prime \prime} _{ j } | d ^{( \text{II} )} _{ \eta ^{\prime}, \pm} | {\bf r} ^{\prime} _{ j } \rangle \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }$$ after we change the number of electrons in subsystem $\eta ^{\prime} $.
An effective carrier is excited in the ionized/affinitive process, and we can obtain its excitation energy by calculating the difference $E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } -E_{Cr}$ based on Eqs. (18) and (20). It is convenient to rewrite $\rho ^{( \text{II} )} _{ \eta ^{\prime} , \pm }$, $d ^{( \text{I} )} _{ \eta ^{\prime} ,\pm }$, and $d ^{( \text{II} )} _{ \eta ^{\prime} ,\pm }$ as $$\begin{aligned}
\begin{cases}
\rho ^{( \text{II} )} _{ \eta ^{\prime} , + } = \rho ^{( \text{II} )} + | \eta ^{ \prime } \rangle \langle \eta ^{ \prime } | \otimes | \bar{L} \rangle \langle \bar{L} | \\
\rho ^{( \text{II} )} _{ \eta ^{\prime} , - } = \rho ^{( \text{II} )} -\rho ^{(2)} _{ \eta ^{\prime} } \\
d ^{( \text{I} )} _{ \eta ^{\prime} ,\pm } = ( \sqrt{2} -1) [ \rho ^{( \text{I} )} \rho ^{( \text{II} )} _{ \eta ^{\prime} , -} ( I - \rho ^{( \text{I} )}) + ( I - \rho ^{( \text{I} )}) \rho ^{( \text{II} )} _{ \eta ^{\prime} , - } \rho ^{( \text{I} )} ] \\
d ^{( \text{II} )} _{ \eta ^{\prime} ,\pm } = ( I - \rho ^{( \text{I} )}) \rho ^{( \text{II} )} _{ \eta ^{\prime} , - } ( I - \rho ^{( \text{I} )}) \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }
\end{cases},\end{aligned}$$ in Eq. (20) to obtain the result irrelevant to $| \eta _{j} \rangle $ for all $ \eta _{j} \neq \eta ^{\prime}$. Direct calculation yields $$\begin{aligned}
E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } -E_{Cr}=\langle \eta ^{ \prime} | H _{\pm } | \eta ^{ \prime } \rangle + E _{ b} ^{ (\pm)} - E _{b} \text{ with }
\begin{cases}
H_{+} = tr ^{ \prime } H _{hop} (| \bar{L} \rangle \langle \bar{L} | - d ^{(1)}) \\
H_{-} = - tr ^{ \prime } H _{hop} (| L \rangle \langle L | + d ^{(1)})
\end{cases},\end{aligned}$$ where $tr ^{\prime} $ denotes the trace with respect to $\mathbb{C} ^{2} _{ \Omega}$. Near the ionic limit, it is shown in Appendix C that $E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } -E_{Cr}$ can be close to the energy difference obtained by considering the coupled-cluster corrections after we improve my model based on Eq. (25).
The operators $| \bar{L} \rangle \langle \bar{L} |$, $| L \rangle \langle L |$, and $ d ^{(1)}$ in Eq. (22) include the bonding coefficients $\alpha _{i}$ and $\alpha _{c}$, which depend on the repulsive strength of the e-e interaction potential $\hat{U}$ as mentioned in Appendix B. Therefore, we can obtain the interaction-dependent electron ionization and affinity for the excited carrier. To obtain the quantitative results, I consider the nearest-neighbor hopping in the zincblende-structure crystal, in which the atom $A$ located at ${\bf R} _{A} = n _{1} (l_{c}/2, l_{c}/2, 0) + n _{2} (0, l_{c}/2,l_{c}/2) + n _{3} (l_{c}/2, 0, l_{c}/2)$ is accompanied by the atom $B$ at ${\bf R} _{B} = {\bf R} _{A} + (l_{c}/4, l_{c}/4, l_{c}/4) $. Here $n _{1}$, $n _{2}$, and $n _{3}$ are integers, and $l _{c}$ is the length of the crystal lattice. For convenience, I denote ${\bf n} = (n _{1}, n _{2}, n _{3}) $ for the parameters $n _{1}$, $n _{2}$, and $n _{3}$ of ${\bf R} _{A}$, and take $ m =1 \sim 4$ to parameterize the 4 ionic-covalent bonds around the same atom $A$ such that each $| w _{j} \rangle$ can be re-parameterized as $| w _{ {\bf n}, m } \rangle$. Assume that the hopping coefficients equal $t _{A}$ and $t _{B}$ for the adjacent A- and B-site orbitals, respectively. By adding one quasielectron to the s-like Bloch-type orbital $ | \eta ^{\prime} \rangle =\frac{1}{2 \sqrt{N} } \sum _{ {\bf n} , m } e ^{ i {\bf k} \cdot {\bf R}_{A} } | w_{ {\bf n}, m } \rangle \in \mathbb{C} ^{N}$ near the covalent limit, we can obtain $$\begin{aligned}
E_{Cr} ^{ ( \eta ^{\prime} , + ) } -E_{Cr}=3 t _{A} | \alpha _{c} | ^{2} + E _{b} ^{(+)} - E _{b} -\gamma \varepsilon _{dis} ( {\bf k}) \text{ with } \end{aligned}$$ $$\varepsilon _{dis} ( {\bf k}) = cos \frac{ l _{c} k _{x} }{2} cos \frac{ l _{c} k _{y} }{2} + cos \frac{ l _{c} k _{y} }{2} cos \frac{ l _{c} k _{z} }{2} + cos \frac{ l _{c} k _{z} }{2} cos \frac{ l _{c} k _{x} }{2}$$ as the energy dispersion curve [@Grosso] for the excited carrier in the tight-binding scheme. Here $ \gamma = - | \alpha _{i}| ^{2} t _{B} $ serves as the overlap energy [@Kittel], ${\bf k}= ( k _{x} , k _{y}, k _{z}) $ denotes the wavevector, and $\varepsilon _{dis} ( {\bf k})$ can be obtained by considering the twelve nearest-neighbor vectors [@Grosso]. The effective mass $m ^{*} ({\bf k})$ [@Grundmann] follows $ [ m ^{*-1}({\bf k}) ] _{i,j} = - \frac{ \gamma }{\hbar ^2} \frac{ \partial ^{2} }{\partial k_{i} \partial k _{j}} \varepsilon _{dis} ( {\bf k}) \propto |\alpha _{i}| ^{2} = 1- |\alpha _{c}| ^{2} $ at each ${\bf k}$, which reveals the importance of the bonding coefficients to the excited carrier. In addition, the carrier becomes immobile in the covalent limit because the bandwidth equals zero as $| \alpha _{i} | = 0$. The coefficients $\alpha _{i}$ and $\alpha _{c}$ are determined by the e-e repulsive strength as mentioned in Appendix B, so $m ^{*} ({\bf k})$ depends on the e-e interaction potential $\hat{U}$ in my model. The zero bandwidth in the covalent limit is due to the lack of the double occupancy at half filling under the strong repulsive strength of $\hat{U}$, which induces the Mott insulating behaviors in some AF systems [@Lee1; @Liu].
Discussion
==========
In the last three sections, the non-negative function $U(| {\bf r} |)$ is taken into account to introduce $\hat{U}$ without considering the inter-bond e-e correction. Actually there should exist the inter-bond e-e energy $\frac{1}{2} \sum _{ j_{1} \neq j_{2} } \int _{ {\bf r} _{1} \in \Omega _{ j _{1} } } d ^{3} r _{1} \int _{ {\bf r} _{2} \in \Omega _{ j _{2} } } d ^{3} r _{2} U ^{\prime} (| {\bf r} _{1} - {\bf r} _{2} |) Q ({\bf r} _{1})Q({\bf r} _{2})$ in the quasielectron system to model the considered compound, where $U ^{\prime}$ represents the long-range e-e correction. Because the electron density $Q({\bf r} _{j} )= 2 \langle {\bf r} _{j} | \rho _{sC} | {\bf r} _{j} \rangle$ at ${\bf r} _{j} \in \Omega _{j}$ in the $j$-th bond and $\rho _{sC} = \frac{1}{2}( \rho ^{ ( \text{I} ) } + \rho ^{ ( \text{II} ) } + d ^{ ( \text{I} ) }) $, we shall include $$\text{(i) } \frac{1}{2} \sum _{ q = \text{I}, \text{II} } \text{ } \sum_{ q ^{\prime} = \text{I}, \text{II} } \text{ } \sum _{ j_{1} \neq j_{2} } \int _{ {\bf r} _{1} \in \Omega _{ j _{1} } } d ^{3} r _{1} \int _{ {\bf r} _{2} \in \Omega _{ j _{2} } } d ^{3} r _{2} U ^{\prime} (| {\bf r} _{1} - {\bf r} _{2} |) \langle {\bf r} _{1} | \rho ^{ (q) } | {\bf r} _{1} \rangle \langle {\bf r} _{2} | \rho ^{(q^{\prime})} | {\bf r} _{2} \rangle$$ $$\begin{aligned}
\text{(ii) \ } \sum _{ q = \text{I}, \text{II} } \text{ } \sum _{ j_{1} \neq j_{2} } \int _{ {\bf r} _{1} \in \Omega _{ j _{1} } } d ^{3} r _{1} \int _{ {\bf r} _{2} \in \Omega _{ j _{2} } } d ^{3} r _{2} U ^{\prime} (| {\bf r} _{1} - {\bf r} _{2} |) \langle {\bf r} _{1} | \rho ^{ (q) } | {\bf r} _{1} \rangle \langle {\bf r} _{2} | d ^{ (I) } | {\bf r} _{2} \rangle \text{ \ \ \ \ \ \ } \end{aligned}$$ $$\text{(iii) } \frac{1}{2} \sum _{ j_{1} \neq j_{2} } \int _{ {\bf r} _{1} \in \Omega _{ j _{1} } } d ^{3} r _{1} \int _{ {\bf r} _{2} \in \Omega _{ j _{2} } } d ^{3} r _{2} U ^{\prime} (| {\bf r} _{1} - {\bf r} _{2} |) \langle {\bf r} _{1} | d^{( \text{I} )} | {\bf r} _{1} \rangle \langle {\bf r} _{2} | d^{( \text{I} )} | {\bf r} _{2} \rangle \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ }$$ for the inter-bond e-e correction. Since each chemical bond in the compound system is identical to the one-bond system discussed in subsection III-A, the charge fluctuation $\langle {\bf r} _{j} | d ^{( \text{I})} | {\bf r} _{j} \rangle$ at ${\bf r} _{j} \in \Omega _{j}$ for any $j$ has no contribution to the total electron charge just as $\langle {\bf r} | d ^{(1)} | {\bf r} \rangle$. I note that the deviation $\delta \rho$ responsible for the density-density interaction [@You] in the quantum Hall theory [@Zhang; @DHLee] also has no contribution to the total charge, and Eq. (24)-(iii) shows the universality of such interaction. The charge fluctuation due to $ d ^{( \text{I} )}$ may interact with the charge density given by $\rho ^{( \text{I} )}$ and $\rho ^{( \text{II} )}$, and Eq. (24)-(ii) just provides the corresponding energy together with Eq. (16)-(ii). Equations (16)-(i) and (24)-(i) yield the Hartree-potential energy resulting from $\rho ^{( \text{I} )}$ and $\rho ^{( \text{II} )}$, and we shall include the Fock-potential term [@Nelson] because of the lack of the self-interaction in Eq. (16)-(i).
When the compound is an ideal crystal following the periodic boundary condition, it is important to consider the case that each $ | \eta _{j} \rangle \in \mathbb{C} ^{N}$ in Fig. 3 (a) corresponds to one Bloch-type function. The Bloch wavefunction, which consists of its Bloch-type part in $ \mathbb{C} ^{N}$ and the bonding part in $ \mathbb{C} ^{2} _{ \Omega}$, can be extended to the form $e ^{ i {\bf k} \cdot {\bf r} } u _{ {\bf k} } ({\bf r}) $ [@Kittel] to include the small density due to the electron tails in the shaded region in Fig. 1 (c) as $H _{ sC} $ is replaced by $\frac{{\bf p}^{2}}{2m_{0}} +V _{cr}({\bf r})$. Here $V_{cr}$ denotes the one-electron periodic potential in the crystal, $m _{0} $ is the electron mass in vacuum, ${\bf p}$ is for the momentum operators, and $u _{ {\bf k} } ( {\bf r})$ represents the periodic part of the corresponding Bloch state. To determine the hopping coefficients $t _{ j \xi, j^{\prime} \xi ^{\prime}}$ in Eq. (17), principally we can transform the Bloch wavefunctions to the Wannier ones [@Marzari], which serve as the localized atomic orbitals in the tight-binding model [@Mahan]. For the well-developed ionic-covalent bonds, in Fig. 1 (c) the electron density in the shaded region must be so low that the bonding electrons almost concentrate in the red region, where the high density induces the bonding correlation representing by Eq. (15). After introducing the inter-bond hopping, therefore, we may approximate the $j$-th bond’s Wannier function as zero outside $\Omega _{j} $ for all $j$ to calculate the e-e energy. Actually Eqs. (18) and (20) can be valid in the systems composed of different ionic-covalent dimers such as the those in the chalcopyrite-structure [@Grundmann] compound when all the bonds are well-separated, as shown in Appendix D. The Wannier basis, however, depends on the gauge freedom [@Marzari] and is not unique. When the ionic-covalent bonds in the compound are not identical, the decomposition in Fig. 3 (a) can become invalid because of the non-constant bonding coefficients. More studies are necessary to clarify how to exactly include the bonding correlation beyond the compound model developed in subsections III-B and IV-B.
It is shown in subsection IV-B that bandwidth can become zero because of the strong e-e repulsive strength, which is responsible for the Mott insulator in some AF systems [@Lee1; @Liu]. It is known that the random fields [@Kirsc; @Erdos; @Lee2; @Huang2] modeled by a family of parameterized Hamiltonians can result in the disorder leading to different insulators, and both the disorder and e-e interaction effects have been observed in the quantum Hall systems [@TYH; @Wang]. The transition between insulating phases has been studied by considering the disordered interacting systems. [@Byczuk; @Braganc] To include a Hamiltonian family, we may replace $H_{sC}$ by the random-matrix set $H_{sC} ^{\omega}$ parameterized by $\omega$ and consider $\sum _{\omega} (\rho ^{( \text{I} )}_{\omega} e _{1,\omega} e _{1,\omega} ^{\dag} + \rho ^{( \text{II} )}_{\omega} e _{2,\omega} e _{2,\omega} ^{\dag})$. Here the set $\{ e _{1, \omega} , e _{2,\omega} \}$ is an orthonormal one in the corresponding vector space, and for each $\omega$ the matrices $\rho ^{( \text{I} )}_{\omega}$ and $\rho ^{( \text{II} )}_{\omega}$ serve as $\rho ^{( \text{I} )}$ and $\rho ^{( \text{II} )}$. If the one-bond system in Fig. 1 (a) is asymmetric with respect to the bonding axis, the rotation centered on such an axis is important to the mapping ${\bf R} _{j}$ in subsection III-B for each $j$ and we need to introduce the parameter $\omega$ for the rotation degrees of freedom [@Huang3]. In the Born-Oppenhemier method [@Bohm] (BOM), we also need to consider a family of Hamiltonians to determine the electron wavefunctions parameterized by the relative position of the nuclei. For any two $2^{n} \times 2^{n}$ matrices $\rho ^{\prime} _{ea}$ and $\rho ^{\prime \prime} _{ea}$, actually we can construct a $ 2 ^{n+1} \times 2 ^{n+1}$ matrix $\rho ^{\prime \prime \prime} _{ea} = \rho ^{\prime} _{ea} \otimes e _{1} e _{1} ^{\dag} + \rho ^{\prime \prime} _{ea} \otimes e _{2} e _{2} ^{\dag}$ and take Eq. (8) as the case for $n=0$, where the integer $n$ is non-negative. By this way we can construct AF-type quasielectrons with $2^{n+1}$ components for the Hamiltonian family parameterized by $\omega = 1 \sim 2 ^{n}$. The multiple-component orbitals can be used to include the multiple CCDs, which are briefly discussed in Appendix B after including $| \Psi _{BB} \rangle$, in the quasielectron space [@Huang3]. I note that the multiple-component functions are introduced to develope the vector bundles [@Bohm; @Friedman; @Banks]. While the hole components [@Huang1] do not appear in the density matrices in my ionic-covalent model, they may become important when both the particle-particle and particle-hole channels [@Yu] are taken into account for the Bogoliubov-BCS quasiparticles. By considering the fractal structures [@Schwalm] to extend such quasiparticles [@Huang1; @Huang4], in fact, we can obtain the form of Eq. (2) from the electron components of the extended ones.
Summary
=======
The extended AF quasielectrons are introduced for the compound where the ionic-covalent bonds are identical to the one-bond dimer. The density-fluctuation and covalent-correlation operators are constructed for the bonding correlation, and my quasielectron model shows the universality of the density-density interaction. The quasielectron system is decomposed into the 4-level subsystems for the electron ionization and affinity in the compound, and such a model can be supported by the coupled-cluster theory near the ionic limit. For the ideal crystal, each subsystem may correspond to one Bloch-type function. By considering the nearest-neighbor hopping in the zincblende-structure crystal, we can see the importance of the bonding coefficients to the effective mass near the covalent limit.
Acknowledgment {#acknowledgment .unnumbered}
==============
The author thanks Profs. I.-H. Tsai, Keh-Ning Huang, and Hrong-Tzer Yau for the valuable discussions about the vector bundles, coupled-cluster corrections, and random-matrix theory, respectively.
Appendix A {#appendix-a .unnumbered}
==========
For the one-bond system in Fig. 1 (a), I assume that the distance $\ell$ separating the peaks of $| \langle {\bf r } | A \rangle | ^{2}$ and $| \langle {\bf r } | B \rangle | ^{2}$ in Fig. 4 is longer than the interacting length $a$ of $\hat{U}$. In addition, assume that $\int _{ {\bf r} ^{\prime} \in \Omega} d ^{3} r ^{\prime} \int _{ {\bf r ^{\prime \prime} } \in \Omega} d^{3} r ^{ \prime \prime} U (|{\bf r} ^{\prime} -{\bf r} ^{ \prime \prime } |) | \langle {\bf r } ^{\prime} | A \rangle | ^{2} | \langle {\bf r ^{\prime \prime } } | A \rangle | ^{2} \sim \int _{ {\bf r} ^{\prime} \in \Omega} d ^{3} r ^{\prime} \int _{ {\bf r ^{\prime \prime } } \in \Omega} d^{3} r ^{ \prime \prime} U (|{\bf r} ^{\prime} -{\bf r} ^{ \prime \prime} |) | \langle {\bf r }^{\prime} | B \rangle | ^{2} | \langle {\bf r ^{\prime \prime} } | B \rangle | ^{2} \sim U _{0} >0$, where the positive parameter $U _{ 0} $ represents the repulsive strength of $\hat{U}$. Hence the effective e-e potential $\hat{U}$ in Eqs. (9) and (10) is dominated by $U _{0} c _{A \uparrow} ^{\dag} c _{A \uparrow} c _{A \downarrow} ^{\dag} c _{A \downarrow} + U _{0} c _{B \uparrow} ^{\dag} c _{B \uparrow} c _{B \downarrow} ^{\dag} c _{B \downarrow} \equiv \hat{U} _{0} $, which corresponds to the intrasite Coulomb repulsion [@Matlak1; @Matlak2] in the t-U model. Moreover, I assume that the difference $\varepsilon _{B} - \varepsilon_{A} $ is high enough for us to take $\sum _{\sigma} (t _{AB} c _{A \sigma} ^{\dag} c _{B \sigma} + t _{AB} ^{*} c _{B \sigma} ^{\dag} c _{A \sigma} ) + \hat{U}- \hat{U}_{0}$ as the perturbation part of $H _{b}$.
Under the above assumptions, the remained quasielectron and quasihole in the one-bond system are roughly located at $| A \rangle$ and $| B \rangle$ as the one-bond wavefunction becomes the one- and three-electron ground states, respectively. Hence the spatial ket $| 1 \rangle $ of the occupied level in Fig. 2 (b) is close to $| A \rangle $ when the 4-level dimer in Fig. 2 represents such a one-bond system, and the ket $| \bar{1} \rangle$ for the quasihole at the right-hand side of Fig. 2 (c) can be approximated by $| B \rangle$. In addition, we may neglect the small change on $\rho ^{(1)}$ as one electron is removed/added. For the one- and three-electron states of subsystem $ \eta ^{\prime} $ in Fig. 3 (b), on the other hand, we shall take $| 1 \rangle \sim | \eta ^{\prime} \rangle \otimes | A \rangle $ and $| \bar{1} \rangle \sim | \eta ^{\prime} \rangle \otimes | B \rangle$ in Figs. 2 (b) and (c), respectively, and approximate $\rho ^{(1)} _{\eta ^{\prime} }$ as $( | \eta ^{\prime} \rangle \otimes | A \rangle)( \langle \eta ^{\prime} | \otimes \langle A |)$. We can tune $\rho ^{(1)}$ and $\rho ^{(1)} _{ \eta ^{\prime} }$ in section IV to modify the orbitals of the remained quasiparticles in the corresponding spaces $\mathbb{C} ^{2} _{\Omega}$ and $| \eta ^{\prime} \rangle \langle \eta ^{\prime} | \otimes \mathbb{C} ^{2} _{\Omega}$. The charge background due to the half-filled subsystems, which are characterized by $\eta _{j} \neq \eta ^{\prime}$ in Fig. 3 (b), should be taken into account to perform the modification for the compound system.
Appendix B {#appendix-b .unnumbered}
==========
In the one-bond system in Fig. 1 (a), the wavefunction $| \Psi _{b} ^{SCF} \rangle = ( \sqrt{1- | \lambda _{1} | ^{2} } c _{ A \uparrow} ^{\dag} + \lambda _{1} c _{ B \uparrow} ^{\dag} ) ( \sqrt{1- | \lambda _{1} | ^{2} } c _{ A \downarrow} ^{\dag} + \lambda _{1} c _{ B \downarrow} ^{\dag} ) | 0 \rangle$ can serve as the effective SCF state at half filling near the ionic limit if the small parameter $\lambda _{1}$ is determined by minimizing $\langle \Psi _{b} ^{SCF} | H _{b} | \Psi _{b} ^{SCF} \rangle$. The wavefunction $| \Psi _{b} ^{SCF} \rangle$ is a linear combination of $| \Psi _{i} \rangle$, $| \Psi _{c} \rangle$, and $| \Psi _{BB} \rangle$, so principally we should take $| \Psi _{BB} \rangle$ into account in addition to the ionic and covalent parts. The ket $| \Psi _{b} ^{CCD} \rangle = ( \sqrt{1- | \lambda _{1} | ^{2} } c _{ B \uparrow} ^{\dag} - \lambda _{1} ^{*} c _{ A \uparrow} ^{\dag} ) ( \sqrt{1- | \lambda _{1} | ^{2} } c _{ B \downarrow} ^{\dag} - \lambda _{1} ^{*} c _{ A \downarrow} ^{\dag} ) | 0 \rangle$ is the only allowed CCD for the bonding electrons. When Brillouin theorem [@Manninen; @Pople] is valid near the ionic limit, the coupled-cluster method is applicable and we may take $| \Psi _{b} ^{Br} \rangle = \sqrt{1- | \lambda _{2} | ^{2} } | \Psi _{b} ^{SCF} \rangle + \lambda _{2} | \Psi _{b} ^{CCD} \rangle $ as $| \Psi _{b} \rangle $ to model the ground state. Here the small parameter $\lambda _{2}$ is determined by minimizing $\langle \Psi _{b} ^{Br} | H _{b} | \Psi _{b} ^{Br} \rangle$. The single substitution [@Pople] is neglected in $| \Psi _{b} ^{Br} \rangle $.
While Brillouin theorem may become invalid, we have the bonding wavefunction $ | \Psi _{b} \rangle = \tau _{i} | \Psi _{i} \rangle + \tau _{c} | \Psi _{c} \rangle+ \tau _{BB} | \Psi _{BB} \rangle $ [@Prasad; @Havenith] in general when the one-bond system is half-filled. Here $ \tau _{i} $, $ \tau _{c} $, and $ \tau _{BB} $ are the coefficients satisfying $ | \tau _{i} | ^{2} + | \tau _{c} | ^{2} + | \tau _{BB} | ^{2} =1$. The energies of $| \Psi _{i} \rangle$ and $| \Psi _{c} \rangle$ are close to $2 \varepsilon _{A} + U _{0}$ and $\varepsilon _{A}+\varepsilon _{B}$, respectively, and are both lower than the energy of $| \Psi _{BB} \rangle$ under the assumptions mentioned in Appendix A. Hence $| \tau_{BB} | $ should be small, and we can approximate $\tau _{i}$ and $\tau _{c}$ as the coefficients $\alpha _{i}$ and $\alpha _{c}$ in Eq. (1) if it is suitable to neglect the small contribution of $ | \Psi _{BB} \rangle$. The bonding wavefunction $| \Psi _{b} \rangle \rightarrow | \Psi _{i} \rangle$ near the ionic limit as the repulsive strength $ U_{0} << \varepsilon _{B}- \varepsilon _{A}$. With increasing the e-e repulsive strength, $| \alpha _{i} |$ decreases and $\alpha _{c}$ becomes significant. The wavefunction $|\Psi _{b} \rangle \rightarrow |\Psi _{c} \rangle$ near the covalent limit when $U _{0} >> \varepsilon _{B}- \varepsilon _{A} $, under which the double occupancy is forbidden.
When it is inappropriate to neglect $| \Psi _{BB} \rangle$, we can perform the orbital transformation $$\begin{aligned}
\begin{cases}
c _{ A^{\prime} \sigma } =\sqrt{ 1 - | \lambda _{3} | ^{2} } c _{ A \sigma } + \lambda _{3} c _{ B \sigma } \\
c _{ B^{\prime} \sigma } = - \lambda _{3} ^{*} c _{ A \sigma } + \sqrt{ 1 - | \lambda _{3} | ^{2} } c _{ B \sigma }
\end{cases},\end{aligned}$$ to rewrite $| \Psi _{b} \rangle$ as $\alpha_{i} ^{\prime} | \Psi _{i} ^{\prime} \rangle + \alpha_{c} ^{\prime} | \Psi _{c} ^{\prime} \rangle + \tau_{BB} ^{\prime} | \Psi _{BB} ^{\prime} \rangle $ with $ | \Psi _{i} ^{\prime} \rangle = c _{ A^{\prime} \uparrow } ^{\dag} c _{ A^{\prime} \downarrow } ^{\dag} | 0 \rangle $, $| \Psi _{c} ^{\prime} \rangle = \frac{1}{ \sqrt{2} } (c _{ A^{\prime} \uparrow } ^{\dag} c _{ B^{\prime} \downarrow } ^{\dag} + c _{ B^{\prime} \uparrow } ^{\dag} c _{ A^{\prime} \downarrow } ^{\dag}) | 0 \rangle$, and $ | \Psi _{BB} ^{\prime} \rangle = c _{ B^{\prime} \uparrow } ^{\dag} c _{ B^{\prime} \downarrow } ^{\dag} | 0 \rangle $. Here $\lambda _{3}$ is a complex number following $0 \leq | \lambda _{3} | \leq 1 $, and $\alpha_{i} ^{\prime}$, $\alpha_{c} ^{\prime}$, and $\tau_{BB} ^{\prime}$ are the coefficients determined by $\tau _{i}$, $\tau _{c}$, $\tau _{BB}$ and $\lambda _{3}$. The bonding wavefunction $| \Psi _{b} \rangle$ becomes $\alpha_{i} ^{\prime} | \Psi _{i} ^{\prime} \rangle + \alpha_{c} ^{\prime} | \Psi _{c} ^{\prime} \rangle$, the ionic-covalent form given by Eq. (1), if the parameter $\kappa \equiv \lambda ^{*} _{3} / \sqrt{1 - | \lambda ^{*} _{3} | ^{2} } $ follows $\tau _{i} \kappa ^{2} - \sqrt{2} \tau _{c} \kappa + \tau _{BB} =0$ such that $\tau_{BB} ^{\prime} =0 $. There are two solutions to $\kappa$, and we can choose the solution with the smaller absolute value while the other one is important to the spontaneous symmetry breaking [@Huang3]. To improve my model by including $| \Psi _{BB} \rangle$, we shall replace $|A\rangle$ and $| B \rangle $ by $| A ^{\prime} \rangle \equiv \sqrt{ 1 - | \lambda _{3} | ^{2} } | A \rangle + \lambda _{3} ^{*} | B \rangle $ and $| B ^{\prime} \rangle \equiv - \lambda _{3} | A \rangle + \sqrt{ 1 - | \lambda _{3} | ^{2} } | B \rangle$ such that $\rho ^{(1)} \rightarrow | A ^{\prime} \rangle \langle A ^{\prime} | $ and $\rho ^{(2)} \rightarrow | L ^{\prime} \rangle \langle L ^{\prime} | $ in Eq. (11), where $ | L ^{\prime} \rangle = \alpha _{i} ^{\prime} | A ^{\prime} \rangle + \alpha _{c} ^{\prime} | B ^{\prime} \rangle$. In addition, the ket $| \bar{L} \rangle$ in subsection III-A should be replaced by $ | \bar{L} ^{\prime} \rangle = \alpha _{c} ^{\prime \ast} | A ^{\prime} \rangle - \alpha _{i} ^{\prime \ast} | B ^{\prime} \rangle $. Based on the mapping ${\bf R}_{j}$, for Eq. (14) we have $ \rho _{ w _{j} } ^{(1)} = | w _{j} \rangle \langle w _{j} | \otimes |A ^{\prime} \rangle \langle A ^{\prime} |$ and $ \rho _{ w _{j} } ^{(2)} = | w _{j} \rangle \langle w _{j} | \otimes |L ^{\prime} \rangle \langle L ^{\prime} |$, under which $ \rho _{ \eta _{j} } ^{(1)} = | \eta _{j} \rangle \langle \eta _{j} | \otimes |A ^{\prime} \rangle \langle A ^{\prime} |$ and $ \rho _{ \eta _{j} } ^{(2)} = | \eta _{j} \rangle \langle \eta _{j} | \otimes |L ^{\prime} \rangle \langle L ^{\prime} |$. The matrix $\rho ^{(2)} _{+}$ in Eq. (19) and the operator $\rho ^{(2)} _{ \eta ^{\prime} +}$ for subsystem $\eta ^{\prime}$ in subsection IV-B both remain unchanged because $| A \rangle \langle A| + | B \rangle \langle B | = | A ^{\prime} \rangle \langle A ^{\prime} | + | B ^{\prime} \rangle \langle B ^{\prime} | $. The one- and three-electron states $| \Psi _{b,1,\sigma} \rangle$ and $| \Psi _{b,3,\sigma} \rangle$ in subsection IV-A should be replaced by $| \Psi ^{\prime} _{b,1,\sigma} \rangle \equiv c ^{\dagger} _{ A^{\prime} \sigma } | 0 \rangle$ and $ | \Psi ^{\prime} _{b,3,\sigma} \rangle \equiv c _{ B^{\prime} \sigma } | Fb \rangle $ if we neglect the small change on $\rho ^{(1)}$ when one electron is removed/added. In Eqs. (19) and (20), we can perform the modification discussed at the end of Appendix A by tuning $\rho ^{(1)}$ and $\rho ^{(1)} _{ \eta ^{\prime} }$.
When we take $|\Psi ^{Br} _{b} \rangle $ as the bonding wavefunction according to the coupled-cluster method near the ionic limit, the coefficient $\tau _{c}$ can be small and become comparable with $\tau _{BB}$. The bonding wavefunction $| \Psi _{b} \rangle $ is dominated by $| \Psi _{i} \rangle$, but we cannot consider only the ionic part to probe the bonding correlation. Therefore, it is important to improve my model near such a limit by using Eq. (25) to include $| \Psi _{BB} \rangle$ in addition to the covalent part if we hope to exactly probe the bonding correlation.
In the BOM, a family of Hamiltonians are taken into account by considering the variation on the positions of the nuclei. While $| \Psi _{b} ^{CCD} \rangle$ is the only allowed CCD in the 4-level dimer for the one-bond system in Fig. 1 (a), it depends on the positions of the nuclei and thus can generate a CCD family $ \{ | \Psi _{b} ^{CCD} (\omega) \rangle \}$. Here the parameter $\omega$ is to parameterize such a family. Multiple CCDs, in fact, can be incorporated in the quasiparticle space by considering the corresponding family [@Huang3], and we may extend the BOM to develop the quasiparticles including both the electron-correlation and nucleus-vibration effects.
Appendix C {#appendix-c .unnumbered}
==========
In subsection III-B, I consider the compound system where the identical bonds are parametrized by $j$. For convenience, let $c_{jA \sigma}$ and $c_{jB \sigma}$ as the annihilators to remove electrons with the spin orientation $\sigma$ in $| A, j \rangle$ and $| B, j \rangle$, respectively. The compound Hamiltonian ${\cal H}= \sum _{ j \neq j ^{ \prime}} H_{hop} ^{(j, j^{\prime})} + \sum _{ j } (H_{sb} ^{(j)} + \hat{U _{j} }) $, where $ H_{hop} ^{(j, j^{\prime})} = \sum _{\sigma} \sum _{ \xi , \xi ^{\prime} \in \{ A,B \} } t _{ j \xi , j ^{\prime} \xi ^{\prime} } c_{j \xi \sigma} ^{\dag} c_{ j ^{\prime} \xi ^{\prime} \sigma} $, $H_{sb} ^{(j)} = \sum _{\sigma} (t _{AB} c _{j A \sigma} ^{\dag} c _{ j B \sigma} + t _{AB} ^{*} c _{jB \sigma} ^{\dag} c _{jA \sigma} + \varepsilon _{A} c _{jA \sigma} ^{\dag} c _{jA \sigma} + \varepsilon _{B} c _{jB \sigma} ^{\dag} c _{jB \sigma})$, and $\hat{U _{j} } = \int _{ {\bf r} _{1},{\bf r} _{2} \in \Omega _{j} } d ^{3} r _{1} d ^{3} r _{2} U ( | {\bf r} _{1} - {\bf r} _{2} | ) \psi _{ \uparrow } ^{\dag} ({\bf r} _{1}) \psi _{ \uparrow } ({\bf r} _{1}) \psi _{ \downarrow } ^{\dag} ({\bf r} _{2}) \psi _{ \downarrow } ({\bf r} _{2})$.
Assume that all the hopping coefficients $ t _{ j \xi , j ^{\prime} \xi ^{\prime} }$ are so small that every bond in the considered compound is almost independent and is mapped from the one-bond system in Fig. 1 (a). When $| \Psi ^{Br} _{b} \rangle$ is taken as the one-bond wavefunction near the ionic limit as mentioned in Appendix B, we shall take $ | \Psi ^{Br} _{j} \rangle= \sqrt{1- | \lambda _{2} | ^{2} } | \Psi _{j} ^{SCF} \rangle + \lambda _{2} | \Psi _{j} ^{CCD} \rangle$ with $| \Psi ^{SCF} _{j} \rangle=c _{j A^{\prime \prime} \uparrow } ^ {\dag} c _{j A^{\prime \prime} \downarrow } ^ {\dag} | 0 \rangle$ and $| \Psi ^{CCD} _{j} \rangle = c _{j B^{\prime \prime} \uparrow } ^ {\dag} c _{j B^{\prime \prime} \downarrow } ^ {\dag} | 0 \rangle$ for the $j$-th bond based on the mapping ${\bf R}_{j}$. Here $c _{j A^{\prime \prime} \sigma } ^ {\dag}= \sqrt{1- | \lambda _{1} | ^{2} } c _{ j A \sigma} ^{\dag} + \lambda _{1} c _{ j B \sigma} ^{\dagger} $ and $c _{j B^{\prime \prime} \sigma } ^ {\dag} = \sqrt{1- | \lambda _{1} | ^{2} } c _{ j B \sigma} ^{\dag} - \lambda _{1} ^{*} c _{ j A \sigma} ^{\dag} $. We may use Eq. (25) to rewrite $| \Psi ^{Br} _{b} \rangle$ and $ | \Psi ^{Br} _{j} \rangle$ by the ionic-covalent form, and approximate the one- and three-electron states of the one-bond system by $| \Psi ^{\prime} _{b,1,\sigma} \rangle $ and $ | \Psi ^{\prime} _{b,3,\sigma} \rangle $, which are introduced in Appendix B. Let $| \Psi ^{ \prime (j)} _{b,1,\sigma} \rangle = c _{j A^{\prime} \sigma } ^{\dagger} |0 \rangle$ and $| \Psi ^{\prime (j)} _{b,3,\sigma} \rangle = c _{j B^{\prime} \sigma } | Fb ^{(j)} \rangle $ as the $j$-th bond’s states mapped from $| \Psi ^{ \prime } _{b,1,\sigma} \rangle$ and $| \Psi ^{ \prime} _{b,3,\sigma} \rangle$. Here the 4-electron state $|Fb ^{(j)} \rangle$ describes the fully occupied $j$-th bond, and the annihilators $c _{j A^{\prime} \sigma } = \sqrt{ 1 - | \lambda _{3} | ^{2} } c _{ j A \sigma } + \lambda _{3} c _{ j B \sigma }$ and $c _{j B^{\prime} \sigma } = - \lambda _{3} ^{*} c _{ j A \sigma } + \sqrt{ 1 - | \lambda _{3} | ^{2} } c _{ j B \sigma }$. The effective SCF state for the compound is $| \Psi ^{SCF} _{Cr} \rangle = \prod _{j=1} ^{ N } c _{ j A^{\prime \prime} \uparrow } ^ {\dag} c _{j A^{\prime \prime} \downarrow } ^ {\dag} | 0 \rangle$, and $E _{Cr} ^{SCF} =\langle \Psi ^{SCF} _{Cr} | {\cal H} | \Psi ^{SCF} _{Cr} \rangle$ is the SCF value of $E _{Cr}$.
For the electron affinity and ionization discussed in subsection IV-B, in the SCF calculation [@Stoyanova; @Grafenstein2] we shall calculate $ \langle \Psi ^{SCF} _{ \eta ^{\prime} \sigma, + } | {\cal H} | \Psi ^{SCF} _{ \eta ^{\prime} \sigma, + } \rangle$ and $\langle \Psi ^{SCF} _{ \eta ^{\prime} \sigma, - } | {\cal H} | \Psi ^{SCF} _{ \eta ^{\prime} \sigma, - } \rangle$ when the carrier excitations occur in the subsystem corresponding to $| \eta ^{\prime} \rangle$. Here $$\begin{aligned}
| \Psi ^{SCF} _{ \eta ^{\prime} \sigma, \pm } \rangle=
\begin{cases}
\sum _{j=1} ^{N} \langle w _{j} | \eta ^{ \prime} \rangle c _{ j B^{ \prime \prime} \underline{ \sigma } } ^{\dag} | \Psi ^{SCF} _{Cr} \rangle \text{ \ \ for \ \ } + \\
\sum _{j=1} ^{N} \langle w _{j} | \eta ^{ \prime} \rangle ^{*} c _{ j A^{ \prime \prime} \underline{ \sigma } } | \Psi ^{SCF} _{Cr} \rangle \text{ \ for \ \ } -
\end{cases}\end{aligned}$$ with $\underline{\sigma} = - \sigma$. We can denote the spin-independent values $ \langle \Psi ^{SCF} _{ \eta ^{\prime} \sigma, + } | {\cal H} | \Psi ^{SCF} _{ \eta ^{\prime} \sigma, + } \rangle$ and $\langle \Psi ^{SCF} _{ \eta ^{\prime} \sigma , - } | {\cal H} | \Psi ^{SCF} _{ \eta ^{\prime} \sigma , - } \rangle$ as $E ^{SCF} _{ \eta ^{\prime} , +} $ and $E ^{SCF} _{ \eta ^{\prime} , -} $, respectively, and choose $\sigma= \downarrow$ in Eq. (26) without loss of generality. In the SCF calculation, the added/removed charge in the $j$-th bond equals $| \langle w _{j} | \eta ^{\prime} \rangle |^{2}$ just as that in subsection IV-B, and $E ^{ ( \eta ^{\prime} ,\pm ) }_{Cr} - E _{Cr}$ is approximated as $$\begin{aligned}
E ^{SCF} _{ \eta ^{\prime} , \pm} - E ^{SCF}_{Cr} = {\cal K} ^{SCF} _{ \eta ^{\prime} , \pm} + {\cal B}^{SCF} _{ \eta ^{\prime} , \pm} \text{ \ with } \end{aligned}$$ $$\begin{cases}
{\cal B}^{SCF} _{ \eta ^{\prime} , \pm} = \sum _{j=1} ^{ N } | \langle w _{j} | \eta ^{\prime} \rangle |^{2} [ \langle j , \pm| (H_{sb} ^{(j)} + \hat{ U } _{j})| j , \pm \rangle - \langle \Psi ^{SCF}_{j} | (H_{sb} ^{(j)} + \hat{ U } _{j}) | \Psi ^{SCF} _{j} \rangle ] \\
{\cal K} ^{SCF} _{ \eta ^{\prime} , +} = \sum _{ j \neq j ^{\prime} } \sum _{ \xi \xi ^{\prime} } t _{ j \xi , j ^{\prime} \xi ^{\prime} } \langle w _{ j ^{\prime} } | \eta ^{\prime} \rangle \langle \eta ^{ \prime} | w _{ j } \rangle \langle j ,+| c _{j \xi \uparrow } ^{\dag} | \Psi ^{SCF} _{j} \rangle \langle \Psi ^{SCF} _{j ^{\prime} } | c _{j ^{\prime} \xi ^{\prime} \uparrow } | j ^{\prime} , + \rangle \\
{\cal K} ^{SCF} _{ \eta ^{\prime} , -} = - \sum _{ j \neq j ^{\prime} } \sum _{ \xi \xi ^{\prime} } t _{ j \xi , j ^{\prime} \xi ^{\prime} } \langle w _{ j ^{\prime} } | \eta ^{\prime} \rangle \langle \eta ^{ \prime} | w _{ j } \rangle \langle \Psi ^{SCF} _{j}| c _{j \xi \uparrow } ^{\dag} | j , - \rangle \langle j ^{\prime} , - | c _{j ^{\prime} \xi ^{\prime} \uparrow } |\Psi ^{SCF} _{j ^{\prime} } \rangle
\end{cases}.$$ Here $| j , - \rangle = c _{j A^{\prime \prime} \uparrow } | \Psi ^{SCF} _{j} \rangle$ and $| j , + \rangle = c _{j B^{\prime \prime} \uparrow } ^{\dag} | \Psi ^{SCF} _{j} \rangle$. The energy factors ${\cal B}^{SCF} _{ \eta ^{\prime} , \pm}$ and ${\cal K} ^{SCF} _{ \eta ^{\prime} , \pm}$ are due to $\sum _{ j } (H_{sb} ^{(j)} + \hat{U _{j} })$ and $ \sum _{ j \neq j ^{ \prime}} H_{hop} ^{(j, j^{\prime})} $, respectively.
To include the coupled-cluster corrections, in Eq. (27) we shall consider the two-particle excitation [@Stoyanova; @Grafenstein2] ${\cal S}^{(2)}_{J}= \frac{\lambda _{2}}{ \sqrt{ 1 - |\lambda _{2}| ^{2} } } c _{J B^{\prime \prime} \uparrow } ^ {\dag} c _{ J A^{\prime \prime} \uparrow } c _{J B^{\prime \prime} \downarrow } ^ {\dag} c _{J A^{\prime \prime} \downarrow } $ to replace $| \Psi ^{SCF} _{J} \rangle $ by $ | \Psi _{J} ^{Br} \rangle = \sqrt{ 1 - |\lambda _{2}| ^{2} } \times exp( {\cal S} ^{(2)} _{J}) | \Psi ^{SCF} _{J} \rangle $ for any $J=j$ or $j ^{\prime}$. In addition, we may introduce the one-particle excitation [@Stoyanova; @Grafenstein2] ${\cal S}^{(1)}_{J}= \frac{\lambda _{4}}{ \sqrt{ 1 - |\lambda _{4}| ^{2} } } c _{ J B ^{\prime \prime} \downarrow } ^{ \dag } c _{J A^{\prime \prime} \downarrow } $ to modify $| J, - \rangle$ and $| J, + \rangle$ as $| \Psi ^{\prime} _{b,1,\downarrow} \rangle = \sqrt{ 1 - |\lambda _{4}| ^{2} } \times exp ( {\cal S}^{(1)}_{J} ) | J , - \rangle $ and $| \Psi ^{\prime} _{b,3,\downarrow} \rangle= \sqrt{ 1 - |\lambda _{4}| ^{2} } \times exp( {\cal S}^{(1)}_{J} ) | J , + \rangle $, respectively, in the $J$-th bond. Here $\lambda _{4} = \lambda _{3} ^{\ast} \sqrt{ 1 - | \lambda _{1} | ^{2} } - \lambda _{1} \sqrt{ 1 - | \lambda _{3} | ^{2} } $. The generator for the ground-state correlation [@Grafenstein] equals $\sum _{J} {\cal S} ^{(2)} _{J}$, and the operator ${\cal S} ^{(1)} _{j} + \sum _{j \neq J} {\cal S} ^{(2)} _{J}$ generates the correlated states corresponding to $c _{ j B^{ \prime \prime} \uparrow } ^{\dag} | \Psi ^{SCF} _{Cr} \rangle$ and $c _{ j A^{ \prime \prime} \uparrow } | \Psi ^{SCF} _{Cr} \rangle$ [@Grafenstein; @Stoyanova; @Grafenstein2].
When the coupled-cluster method is applicable near the ionic limit, as mentioned in Appendix B, it is important to perform the orbital transformation in Eq. (25) to improve my model. Rewriting $| \Psi _{b} ^{Br} \rangle $ as the ionic-covalent form by performing such a transformation, in the improved model we can re-obtain the difference $ E _{b} ^{(\pm)} - E _{b}$ in Eq. (22) by using the coupled-cluster method to modify ${\cal B}^{SCF} _{ \eta ^{\prime} , \pm}$. The coupled-cluster corrections to $ {\cal K} ^{SCF} _{ \eta ^{\prime} , \pm}$, in fact, are small and we have $\langle \eta ^{ \prime} | H _{\pm } | \eta ^{ \prime } \rangle \simeq {\cal K} ^{SCF} _{ \eta ^{\prime} , \pm}$ under the assumption about the small hopping coefficients. So we can use the coupled-cluster method to correct $E ^{SCF} _{ \eta ^{\prime} , \pm} - E ^{SCF}_{Cr}$ and obtain the difference close to $E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } -E_{Cr}$. In addition, the added/removed charge in the $j$-th bond equals $| \langle w _{j} | \eta ^{\prime} \rangle |^{2}$, which is the same as that in subsection IV-B, because the above one- and two-particle excitations do not change the number of electrons in each bond. Therefore, my improved model can be supported by the coupled-cluster theory near the ionic limit.
Appendix D {#appendix-d .unnumbered}
==========
The matrix $\rho ^{(2)} _{ \omega _{j} }$ in Eq. (14) represents the quasielectron at the orbital $ | \omega _{j} \rangle \otimes | L \rangle = \alpha _{i} | A, j \rangle + \alpha _{c} | B, j \rangle$. When the ionic-covalent bonds in the compound are not identical to each other, the bonding coefficients can depend on $j$ and we shall replace $\alpha _{i}$ and $\alpha _{c}$ by $\alpha _{i} ^{(j)}$ and $\alpha _{c} ^{(j)}$ in the $j$-th bond. Here the coefficients $\alpha _{i} ^{(j)}$ and $\alpha _{c} ^{(j)}$ satisfy $| \alpha _{i} ^{(j)} | ^{2} + | \alpha _{c} ^{(j)} | ^{2} =1$ for all $j=1 \sim N$. The matrix $\rho ^{(2)} _{ \omega _{j} }$ should be modified as $(\alpha _{i} ^{(j)} | A, j \rangle + \alpha _{c} ^{(j)} | B, j \rangle ) (\alpha _{i} ^{(j)} \langle A, j | + \alpha _{c} ^{(j)} \langle B, j |) $ and we can still take $\rho ^{(1)} _{ \omega _{j} } = | A, j \rangle \langle A, j |$. Equations (14) and (15) remain valid after the modification, and the energy $E _{Cr}$ can still be obtained based on Eq. (18).
In subsection IV-B, the quasielectron at $| \eta ^{\prime } \rangle \otimes | L \rangle = \sum _{j} \langle \omega _{j} | \eta ^{\prime} \rangle (\alpha _{i} | A, j \rangle + \alpha _{c} | B, j \rangle)$ is ionized from subsystem $\eta ^{ \prime}$ while one quasielectron enters $| \eta ^{\prime } \rangle \otimes | \bar{L} \rangle = \sum _{j} \langle \omega _{j} | \eta ^{\prime} \rangle (\alpha _{c} ^{*} | A, j \rangle - \alpha _{i} ^{*} | B, j \rangle)$ in the affinitive process. When the coefficients for the ionic and covalent parts depend on $j$, the orbital of the quasielectron to be ionized should be modified as $ \sum _{j} \langle \omega _{j} | \eta ^{\prime} \rangle (\alpha _{i} ^{(j)} | A, j \rangle + \alpha _{c} ^{(j)} | B, j \rangle) \equiv| \eta ^{\prime} _{L} \rangle $. In addition, the orbital for the added quasielectron becomes $ \sum _{j} \langle \omega _{j} | \eta ^{\prime} \rangle (\alpha _{c} ^{(j) \ast} | A, j \rangle - \alpha _{i} ^{(j) \ast} | B, j \rangle) \equiv | \eta ^{\prime} _{ \bar{L} } \rangle$. So we need to modify the first two matrices in Eq. (21) as $\rho ^{( \text{II} )} _{ \eta ^{\prime} , + } = \rho ^{( \text{II} )} + | \eta ^{\prime} _{ \bar{L} } \rangle \langle \eta ^{\prime} _{ \bar{L} } |$ and $\rho ^{( \text{II} )} _{ \eta ^{\prime} , - } = \rho ^{( \text{II} )} - | \eta ^{\prime} _{L} \rangle \langle \eta ^{\prime} _{L} |$. The energy $E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } $ can be calculated based on Eq. (20), where $d ^{( \text{I} )} _{ \eta ^{\prime} ,\pm } $ and $d ^{( \text{II} )} _{ \eta ^{\prime} ,\pm } $ can still be obtained from the last two lines in Eq. (21). In the compound composed of different chemical bonds, therefore, Eqs. (18) and (20) may yield the difference $E_{Cr} ^{ ( \eta ^{\prime} , \pm ) } - E_{Cr}$ for the electron ionization and affinity under the suitable modification.
[99]{} James F. Annett, $Superconductivity$, $Superfluids$ and $Condensates$ (Oxford Universality Press, 2004). A. Ghosal and H.-Y. Kee, Phys. Rev. B [**69**]{}, 224513 (2004). R. A. Barankov, L. S. Levitov, and B. Z. Spivak, Phys. Rev. Lett. [**93**]{},160401 (2004). N. Paar, P. Ring, T. Niksic, and D. Vretenar, Phys. Rev. C [**67**]{}, 034312 (2003). R. Gerami and C. Nayak, Phys. Rev. B [**73**]{}, 024505 (2006). S. Chakravarty, R. B. Laughlin, D. K. Morr, and C. Nayak, Phys. Rev. B [**63**]{}, 094503 (2001). C. Bena, S. Chakravarty, J. Hu, and C. Nayak, Phys. Rev. B [**69**]{}, 134517 (2004). H.-Y. Kee and Y. B. Kim, Phys. Rev. B [**66**]{}, 052504 (2002). C. F. Huang and K.-N. Huang, arXiv:1066.1088v2 (2014). R. B. Laughlin, Phys. Rev. Lett. [**112**]{}, 017004 (2014). B. J. Ramshaw, S. E. Sebastian, R. D. McDonald, J. Day, B. S. Tan, Z. Zhu, J. B. Betts, R. Liang, D. A. Bonn, W. N. Hardy, and N. Harrison, Science [**348**]{}, 317 (2015). G. D. Mahan, $Many$-$Particle$ $Physics$ (Kluwer Academic/Plenum Publishers, 2000). M. Matlak, T. Slomska, and B. Grabiec, Phys. Stat. Sol. (b) [**242**]{}, 317 (2005). M. Matlak, J. Aksamit, B. Grabiec, and W. Nolting, Ann. Phys. [**12**]{}, 304 (2003). P. Fulde, Adv. Phys. [**51**]{}, 909 (2002). Z. G. Soos, S. Mazumdar, and T. T. P. Cheung, Mol. Cryst. Liq. Cryst. [**52**]{}, 93 (1979). A. P. Kampf, M. Sekania, G. I. Japaridze, and Ph Brune, J. Phys.: Condens. Matt. [**15**]{}, 5895 (2003). F. D. Buzatu and D. Buzatu, Rom. Rep. Phys. [**59**]{}, 351 (2007). G. Grosso and G. P. Parravicini, $Solid$ $State$ $Physics$, (Academic Press, A Harcourt Science and Technology Company 1999). J. Grafenstein, H. Stoll, and P. Fulde, Chem. Phys. Lett. [**215**]{}, 611 (1993). A. Stoyanova, L. Hozoi, P. Fulde, and H. Stoll, J. Chem, Phys. [**131**]{}, 044119 (2009). K. Doll and H. Stoll, Phys. Rev. B [**56**]{}, 10121 (1997). K. Talukdar, S. Sasmal, M. K. Nayak, N. Vaval, and S. Pal, Phys. Rev. A [**98**]{}, 022507 (2018). R. K. Prasad, $Quantum$ $Chemistry$ (New Age International, 2010). P. A. Lee, N. Nagaosa, and X.-G. Wen, Rev. Mod. Phys. [**78**]{}, 17 (2006). H.-D. Liu, Y.-H. Chen, H.-F. Liu, H.-S. Too, and W.-M. Liu, Sci. Rep. [**4**]{}, 4829 (2014). W. Kirsch, P. Stollmann, and G. Stolz, Commun. Math. Phys. [**195**]{}, 495 (1998). L. Erdos and H.-T. Yau, Commun. Pure Appl. Math. LIII, 0667 (2000). J. O. Lee and J. Yin, Duke Math J. [**163**]{}, 117 (2014). C. F. Huang, arXiv:quant-ph/0305146v2. A. Bohm, A. Mostafazadeh, H. Koizumi, Q.Niu, and J. Zwanziger, $The$ $Geometric$ $Phase$ $in$ $Quantum$ $Systems:$ $Foudations$, $Mathematical$ $Concepts$, $and$ $Applications$ $in$ $Molecular$ $and$ $Condensed$ $Matter$ $Physics$ (Springer-Verlag Berlin Heidelberg, 2003). R. Friedman, J. Morgan, and E. Witten, Commun. Math. Phys. [**187**]{}, 679 (1997). T. Banks, $Modern$ $Quantum$ $Field$ $Theory:$ $A$ $Concise$ $Introduction$ (Cambridge University Press, 2008) I. N. Levine, $Quantum$ $chemistry$ (Pearson Education, Inc. 2009). S. C. Zhang, Int. J. Mod. Phys. B [**6**]{}, 25 (1992). D.-H. Lee and M. P. A. Fisher, Int. J. Mod. Phys. B [**5**]{}, 2675 (1991). Y. You, G. Y. Cho, and E. Fradkin, Phys. Rev. X [**4**]{}, 041050 (2014). C. Kittel, $Introduction$ $to$ $Solid$ $State$ $Physics$, (John Wiley and Sons Inc., 2005). M. Grundmann, $The$ $Physics$ $of$ $Semiconductors:$ $An$ $Introduction$ $Including$ $Nanophysics$ $and$ $Applications$, (Springer-Verlag Berlin Heidelberg 2010). W. Nelson, P. Bokes, P. Rinke, and R. W. Godby, Phys. Rev. A [**75**]{}, 032505 (2007). N. Marzari, A. A. Mostofi, J. R. Yates, I. Souza, and D. Vanderbilt, Rev. Mod. Phys. [**84**]{}, 1419 (2012). T.-Y. Huang, C.-T. Liang, G-H. Kim, C. F. Huang, C.-P. Huang, J.-Y. Lin, H.-S. Goan, and D. A. Ritchie, Phys. Rev. B [**78**]{}, 113305 (2008). Y-T. Wang, G.-H. Kim, C. F. Huang, S.-T. Lo, W.-J. Chen, J. T. Nicholls, L.-H. Lin, D. A. Ritchie, Y. H. Chang, C.-T. Liang and B. P. Dolan, J. Phys.: Condens. Matter [**24**]{}, 405801 (2012). K. Byczuk, W. Hofstetter, and D. Vollhardt, Phys. Rev. Lett. [**94**]{}, 056404 (2005). H. Braganca, M. C. O. Aguiar, J. Vucicevic, D. Tanaskovic, and V. Dobrosavljevic, Phys. Rev. B [**92**]{}, 125143 (2015) C. F. Huang, unpublished. Z.-D. Yu, Y. Zhou,W.-G. Yin, H.-Q. Lin, and C.-D. Gong, Phys. Rev. B [**96**]{}, 045110 (2017). W. A. Schwalm and B. J. Moritz, Phys. Rev. B [**71**]{},134207 (2005). C. F. Huang and K.-N. Huang, arXiv:quant-ph/0604054v3 (2007). P. Manninen, $Lecture$ $notes$ $for$ $the$ $course$ $554017$ $advanced$ $quantum$ $chemistry$ (2012). J. A. Pople, M. HeadGordon, and K. Raghavachari, J. Chem. Phys. [**87**]{}, 5968 (1987). R. W. A. Havenith, J. H. Lenthe, L. W. Jenneskens, and J. J. Engelberts, Faraday Discuss. [**135**]{}, 299 (2007). J. Grafenstein, H. Stoll, and P. Fulde, Phys. Rev. B [**55**]{}, 13588 (1997).
{width="5in"}
{width="5in"}
{width="5in"}
{width="5in"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The inadequacy of Liénard-Wiechert potentials is demonstrated as one of the examples related to the inconsistency of the conventional classical electrodynamics. The insufficiency of the Faraday-Maxwell concept to describe the whole electromagnetic phenomena and the incompleteness of a set of solutions of Maxwell equations are discussed and mathematically proved. Reasons of the introduction of the so-called “electrodynamics dualism concept" (simultaneous coexistence of instantaneous Newton long-range and Faraday-Maxwell short-range interactions) have been displayed. It is strictly shown that the new concept presents itself as the direct consequence of the complete set of Maxwell equations and makes it possible to consider classical electrodynamics as a self-consistent and complete theory, devoid of inward contradictions. In the framework of the new approach, all main concepts of classical electrodynamics are reconsidered. In particular, a limited class of motion is revealed when accelerated charges do not radiate electromagnetic field.'
address: |
Escuela de Física, Universidad Autónoma de Zacatecas\
Apartado Postal C-580 Zacatecas 98068, ZAC., México
author:
- '[**Andrew E. Chubykalo and Roman Smirnov-Rueda**]{} [^1]'
date: 'August 31, 1996'
title: 'Action at a distance as a full-value solution of Maxwell equations: basis and application of separated potential’s method'
---
= 7.0in
Introduction
============
In the last century, the understanding of the nature of electromagnetic phenomena was proceeding in a constant rivalry between two concepts of interaction: namely, [*Newton instantaneous long-range interaction*]{} (NILI) and [*Faraday-Maxwell short-range interaction*]{} (FMSI). At the first moments, owing to the fundamental works of Gauss and Ampére, all electromagnetic phenomena were related to NILI. In other words, it was understood that the interaction forces between both unmoving and moving charges at some specific time were determined by their distribution and the character of their motion at the same instant (implicit time-dependence). As a matter of fact, the concept of field was merely subsidiary (it was considered in a limited sense only as an external-force field) and could be omitted at all. On the contrary to this, the concept of field is primary for FMSI, but charges and currents come to be auxiliary. More fundamentally, a field is a system in its own right (has physical reality), carries energy and fills the whole space. In accordance with Faraday-Maxwell’s idea, the interaction between charged particles can be described only by the intermediary of a field as an energy-carrying physical system. Any electromagnetic perturbation must be spread through the space continuously from point to point during a certain amount of time (finite spread velocity). Finally, the discovery of Faraday’s law of induction (explicit time-dependence of electromagnetic phenomena) and the experimental observation of electromagnetic waves seemed to confirm the field concept. Nevertheless, the idea of NILI still have many supporters in our century. Among the physicists who have developed some theories based, in any case, on this concept, we can find names such as Tetrode and Fokker, Frenkel and Dirac, Wheeler and Feynman, Hoyle and Narlikar \[1\]. This interest to the concept of NILI is explained by the fact that classical theory of electromagnetism is an unsatisfactory theory all by itself, and so there have been many attempts to modify either the Maxwell equations or the principal ideas of electromagnetism. In connection to this, we only mention some works which have tried to unify the advantage of the NILI concept with the conventional theory of field. They are the so-called “retarded action-at-a-distance" theories \[2-6\]. The fact that all new general solutions are represented by half the retarded plus half the advanced Liénard-Wiechert solutions \[7-8\] of the Maxwell’s equations makes it familiar with the conventional FMSI concept. On the other hand, these theories suggest the primacy of charge and use the notion of field as an external-force field like the action-at-a-distance theories. A single charged particle, in this approach, does not produce a field of its own, hence has no self-energy. Thus the classical theory can be saved from some difficulties like self-reaction force (self-interaction), the idea of a whole electromagnetic mass etc. It turns out, however, that no one effort to straighten out the classical difficulties has ever succeeded in making a self-consistent electromagnetic theory. Moreover, the principal difficulties in Maxwell’s theory do not disappear still after the quantum mechanics modifications are made. In spite of the great variety of methods applied to arrange the situation, no one theory dealing with electromagnetism had ever admitted the possibility of simultaneous and independent co-existence of two types of interaction: NILI and FMSI. A new approach, based on this idea, has no need to modify neither Maxwell equations nor the basis ideas of the classical electromagnetic theory. In this work we take a complete set of Maxwell equations as a correct one and show that dualism of electromagnetic phenomena is an intrinsic feature. Physical and mathematical ground for that will be given in the next sections.
Inadequacy of Liénard-Wiechert potentials. A paradox
====================================================
The presence of a paradox in some theory no always means its inconsistency but often indicates the cause of difficulties. In this section we show one of the confusion’s of classical electrodynamics in describing an electromagnetic field of an accelerated charge. The attractiveness of this example consists in the way it lightens the main conventional theory difficulties and the way it leads to the dualism idea. Let us consider a charge $q$ moving in a laboratory reference system with a constant acceleration $a$ along the positive direction of the $X$-axis. An electric field created by an arbitrarily moving charge is given by the following expression obtained directly from Liénard-Wiechert potentials \[9\]: $${\bf E}(x,y,z,t)=q\frac{({\bf R}-R\frac{{\bf
V}}{c})(1-\frac{V^{2}}{c^{2}})}{(R-{\bf R}\frac{{\bf
V}}{c})^{3}}+q\frac{[{\bf R},[({\bf R}-R\frac{{\bf
V}}{c}),\frac{{\bf{\dot{V}}}}{c^{2}}]]}{(R-{\bf R}\frac{{\bf
V}}{c})^{3}}.$$ We remind here that all values in the right-hand of (1) are taken in the moment of time $t_0=t-\tau$, where $\tau$ is the “retarded time“. We shall see that formula (1) satisfies the D’Alembert’s equation along the $X$-axis at any time. To begin with, we note that in a free space along the $X$-axis (except the site of a charge) an electric field component $E_x$ satisfies the homogeneous wave equation: $$\Delta E_x
-\frac{1}{c^2}\frac{{\partial}^{2} E_x}{{\partial t}^{2}}=0.$$ To find the value $E_x$ at the moment of time $t$, one must take all the values in $rhs$ of (1) at the previous instant $t_0$ derived from the condition: $$t_0=t-\tau=t-\frac{R(t_0)}{c};\quad\{(R^2=(x-x_0)^2+(y-y_0)^2+(z-z_0)^2)\}$$ (here $(x_0,y_0,z_0)$ is the site of the charge at instant $t_0$) or from implicit function: $$F(x,y,z,t,t_0)=t-t_0-\frac{R}{c}=0.$$ Then, we have the following expression for $E_x(x,y,z,t)$: $$E_x(x,y,z,t)=q\frac{\Bigl(x-x_0-R\frac{at_0}{c}\Bigr)\Bigl(1-\frac{a^2t_0^2}
{c^2}\Bigr)}{\Bigl(R-(x-x_0)\frac{at_0}{c}\Bigr)^3}-q\frac{a\Bigl((y-y_0)^2+
(z-z_0)^2\Bigr)}{c^2\Bigl(R-(x-x_0)\frac{at_0}{c}\Bigr)^3}.$$ Substituting $E_x$ given by (5) in the wave equation (2), one ought to calculate in any case $\frac{
\partial E_x}{\partial t_0}$, $\frac{\partial t_0}{\partial t}$ and $\frac{\partial t_0}{\partial x_i}$ using differentiation rules for implicit function: $$\frac{\partial t_0}{\partial t}=-\frac{{\partial F}/{\partial t}}
{{\partial F}/{\partial t_0}};\qquad \frac{\partial t_0}{\partial
x_i}=-\frac{{\partial F}/{\partial x_i}} {{\partial F}/{\partial t_0}}.$$ As a result of the substitution of (5) in (2) one obtains (one tends $y,y_0,z,z_0$ to zero after the differentiation): $$\Delta E_x
-\frac{1}{c^2}\frac{{\partial}^{2} E_x}{{\partial t}^{2}}=
0.$$ This result is not reasonable if we remember that wave equation (2) describes only transverse modes. In this particular case, the $x$ component of electric field turns out to be the longitudinal one and, according to an ordinary wisdom, is inconsistent with the wave equation (2). Thus, the Li[é]{}nard-Wiechert potentials, as a solution of the complete set of Maxwell equations are inadequate for describing the properties of electromagnetic field along the direction of an arbitrarily moving charge. We note here that inadequacy of Liénard-Wiechert potentials for describing the properties of relativistic fields was also shown by C.K. Whitney (see, e.g. \[10\]). The same singular behavior along $X$-axis direction displays another important quantity. The Poynting vector represents the electromagnetic field energy flow per unit area per unit time across a given surface: $${\bf
S}=\frac{c}{4\pi}[{\bf E,H}];\qquad {\bf P}=\frac{1}{c^2}{\bf S},$$ where [**S**]{} is the Poynting vector, [**P**]{} is the momentum density, [**E**]{} and [**H**]{} are the electric and magnetic field strength, respectively. One can easily see that expressions (8) are identically zero along the whole $X$-axis. On the other hand, from the energy conservation law: $$w=\frac{E^2+H^2}{8\pi},\qquad \qquad \frac{\partial w}{\partial
t}=-\nabla\cdot{\bf S}$$ we conclude that $w$ and $\frac{\partial w}{\partial t}$ must differ from zero everywhere along $X$ and there is a linear connection between $w$ and $E^2$. The conflict takes place, if, for instance, the charge is vibrating in some mechanical way along the $X$-axis, then the value of $w$ (which is a point function like $E$) on the same axis will be also oscillating. Then the question arises: [*how does the point of observation, lying at some fixed distance from the charge on continuation of X-axis, “know” about the charge vibration?*]{} The presence of ”retarded time" $\tau$ in (1) indicates that along the $X$-axis the longitudinal perturbation should be spread with the energy transfer (on the contrary to (8)). Since the vector [**S**]{} is a product of the energy density and its spreading velocity [**v**]{}: $${\bf S}=w{\bf v}$$ then either the spreading velocity [**v**]{} or the energy density $w$ must be zero along the $X$-axis. The first assumption puts aside the possibility of any interaction transfer. It is necessary to examine carefully the second one $(w=0)$. The Maxwell’s equations state that time-varying fields are transverse. In electro- and magnetostatics (as correct stationary approximations of Maxwell’s theory), the static fields are longitudinal, in the sense that the fields are derived from scalar potentials \[11\]. Consequently, we can assume the spreading of only longitudinal mods along the singular $X$-axis direction of our example, capable to change the field value in any point along this axis. In this case, according to (10), the energy of the longitudinal mods cannot be stored locally in space $(w=0)$ but the spread velocity may be whatever. On the other hand, FMSI concept forbids the spreading (not the presence) of any longitudinal electromagnetic field component in vacuum. Hence, this paradox can not be resolved in the framework of Faraday-Maxwell electrodynamics. This simplest example underlines the insufficiency of only transverse solutions of Maxwell’s equations to describe full properties of electromagnetic field and leads directly to the dualism idea of [*simultaneous and constant*]{} co-existence of longitudinal ([*action-at-a-distance*]{}) and transverse electromagnetic interactions. In the next sections one can find mathematical and physical reasons for the dualism concept which permits to build up a self-consistent classical electrodynamics. As a final remark, we make a reference to P.A.M. Dirac, who writes \[12\]: “
*As long as we are dealing only with transverse waves, we cannot bring in the Coulomb interactions between particles. To bring them in, we have to introduce longitudinal electromagnetic waves... The longitudinal waves can be eliminated by means of mathematical transformation. Now, when we do make this transformation which results in eliminating the longitudinal electromagnetic waves, we get a new term appearing in the Hamiltonian. This new term is just the Coulomb energy of interaction between all the charged particles: $$\sum_{(1,2)} \frac{e_1 e_2}{r_{12}}$$*
... This term appears automatically when we make the transformation of the elimination of the longitudinal waves
".
Reasons and foundations of the method of separated potentials
=============================================================
Let us recall that a complete set of Maxwell equations is: $$\begin{aligned}
&& \nabla\cdot{\bf E}=4\pi\varrho,\\
&& \nabla\cdot{\bf B}=0,\\
&& \nabla\times{\bf H}=\frac{4\pi}{c}{\bf j}+\frac{1}{c}\frac{\partial
{\bf E}}{\partial t},\\
&& \nabla\times{\bf E}=-\frac{1}{c}\frac{\partial {\bf B}}{\partial t}.\end{aligned}$$ If this system of equations is really complete, it must describe all electromagnetic phenomena without exceptions.
It is often convenient to introduce potentials, satisfying Lorentz condition: $$\nabla\cdot{\bf A}+\frac{1}{c}\frac{\partial\varphi}{\partial t}=0.$$ As a result, the set of coupled first-order partial differential equations (11)-(14) can be reduced to the equivalent pair of uncoupled inhomogeneous D’Alembert’s equations: $$\begin{aligned}
&&\Delta\varphi-\frac{1}{c^2}\frac{{\partial}^{2}{\varphi}}{{\partial
t}^{2}}= -4\pi\varrho({\bf r},t),\\
&&\Delta{\bf A}-\frac{1}{c^2}\frac{{\partial}^{2}{\bf A}}{{\partial
t}^{2}}= -\frac{4\pi}{c}{\bf j}({\bf r},t).\end{aligned}$$ Differential equations have, generally speaking, an infinite number of solutions. An uniquely determined solution is selected by laying down sufficient additional conditions. Different forms of additional conditions are possible for the second order partial differential s equations: initial value and boundary conditions. Usually, a general solution of D’Alembert’s equation is considered as explicit time-dependent function $g({\bf r},t)$. In the stationary state the D’Alembert equation is transformed into the Poisson’s equation which solution is everything an implicit-time dependent function $f({\bf R}(t))$. Nevertheless, the conventional theory does not explain in details how the function $g({\bf r},t)$ is converted into implicit time-dependent function $f({\bf R}(t))$ (and vice versa) when the steady-state problems are studied.
Further we shall demonstrate that former solutions of Maxwell’s equations are incomplete and do not ensure a continuous transition between the D’Alembert and Poisson’s equations solutions, respectively. As a matter of fact, it will be shown that a mathematically complete solution of Maxwell’s equations must be written as a linear combination of two non-reducible functions with implicit and explicit time-dependence: $$f({\bf R}(t))+g({\bf r},t).$$
In the classical Faraday-Maxwell electrodynamics the Poisson’s equation is mathematically exact for the steady-state problems. Based on the idea of a continuous nature of electromagnetic phenomena, one could suppose that the general solution of Poisson’s equation should be continuously transformed to the D’Alembert’s equation solution (and vice versa) when the explicit time-dependence appears (disappears). This requirement can also be formulated as a mathematical condition on the continuity of the general solutions of Maxwell’s equations at every moment of time. By force of the uniqueness theorem for the second order partial differential equations, only one solution can exist satisfying the given initial and boundary conditions. Consequently, the continuous transition from the D’Alembert’s equation solution into the Poisson’s one (and vice versa) must be ensured by the continuous transition between the respective initial and boundary conditions. This is the point where FMSI concept fails. Really, only implicit time-dependence function $f({\bf R}(t))$ can be unique solution of Poisson’s equation and boundary conditions for external problem are to be formulated in the infinity. On the other hand, the D’Alembert’s equation solution is looking for only as explicit time-dependent function $g({\bf r},t)$ since only that one corresponds to the classical FMSI concept as a physically reasonable solution. The boundary conditions in this case are given in a finite region. It has no sense to establish them at the infinity if it cannot be reached by any perturbation with finite spread velocity. Dealing with large external region when the effect of the boundaries is still insignificant over a small interval of time, it is possible to consider the limiting problem with initial conditions for an infinite region (initial Cauchy’s problem).
Let us consider carefully the formulation of respective boundary-value problems in a region extending to infinity \[13\]. There are three external boundary-value problems for Poisson’s equation. They are known as Dirichlet problem, Neumann problem and its combination. The mathematical problem, for instance, for the Dirichlet boundary conditions is formulated as follows. It is required to find the function $u({\bf r})$ satisfying:
\(i) Laplace’s equation $\Delta u=0$ everywhere outside the given system of charges (currents);
\(ii) $u({\bf r})$ is continuous everywhere in the given region and takes the given value $G$ on the internal surface $S$: $u\vert_S = G$;
\(iii) $u({\bf r})$ converges uniformly to $0$ at infinity: $u({\bf r})\rightarrow 0$ as $\vert{\bf r}\vert\rightarrow\infty$.
[*The final condition*]{} (iii) [*is essential for a unique solution!*]{} In the case of D’Alembert’s equation the mathematical problem is formulated in a different manner. Obviously, we are interested only in the problem for an infinite region (initial Cauchy’s problem). So it is required to find the function $u({\bf r},t)$ satisfying:
\(j) homogeneous D’Alembert’s equation everywhere outside the given system of charges (currents) for every moment of time $t\geq 0$;
(jj) initial conditions in all infinite region as follows: $$u({\bf r},t)\vert_{t=0}=G_1({\bf r});\qquad u_t({\bf r},t)\vert_{t=0} =
G_2({\bf r}).$$
The condition (iii) about the uniform convergence at the infinity is not mentioned. We remind here that Cauchy’s problem is considered when one of the boundaries is insignificant over all process time. This condition (iii) will never affect the problem and, hence, cannot be taken into account for the correct solution selecting. However, it may be formally included into the mathematical formulation of D’Alembert’s equation boundary-value problem to fulfil the formal continuity with Poisson’s equation solution at the initial moment of time. Nevertheless, this condition is already meaningless the next instant of time since only explicit time-dependent solutions as $g({\bf r},t)$ (retarded solutions with finite spreading velocity) are considered.
Thus, we underline here that the absence of the condition (iii) for every moment of time in the initial Cauchy’s problem does not ensure the continuous transition into external boundary-value problem for Poisson’s equation and, as a result, mutual continuity between the corresponding solutions cannot be expected by force of the uniqueness theorem. However, there is a way to solve the problem: to satisfy the continuous transition between the D’Alembert’s and Poisson’s equation solutions, one must look for a general solution in form of separated functions (18) non-reducible to each other. When applied to the potentials ${\bf A}$ and $\varphi$ this statement takes a form: $$\begin{aligned}
&& {\bf A}={{\bf A}_0}({\bf R}(t))+{{\bf A}^{\ast}}({\bf r},t),\\
&& \varphi={{\varphi}_0}({\bf R}(t))+{{\varphi}^{\ast}}({\bf r},t).\end{aligned}$$ In this case, the presence of the condition (iii) in the Cauchy’s problem turns out to be meaningful for any instant of time, and the corresponding boundary conditions keeps continuity in respect of mutual transformation.
As an additional remark, we conclude that the traditional solution of D’Alembert’s equation cannot be complete, since the Faraday-Maxwell concept does not allow to take into account the first term in (18) as a valuable one at any moment of time. Turning to the previous section, we see that new solution in form (20) is able to change the electric field component $E_x$ along the $X$-axis at any distance and at any time. It is quite obvious now why Li‚nard-Wiechert potentials (as only explicit time-dependent solution of Cauchy’s problem) turned out not to be the complete solutions of Maxwell equations, and why they are not adequate to describe the whole electromagnetic field.
Let us consider again the set of Maxwell’s equations (11)-(14). A pair of uncoupled differential equations can be obtained immediately for the new general solution in form of separated potentials (19)-(20) (we omit boundary conditions premeditatedly): $$\begin{aligned}
&& \Delta\varphi_0=-4\pi\varrho({\bf r},t),\\
&& \Delta{\bf A}_0=-\frac{4\pi}{c}{\bf j}({\bf r},t)\end{aligned}$$ and $$\begin{aligned}
&&\Delta\varphi^{\ast}-\frac{1}{c^2}\frac{{\partial}^{2}{{\varphi}^{\ast}}}
{{\partial t}^{2}}= 0,\\
&&\Delta{\bf A}^{\ast}-\frac{1}{c^2}\frac{{\partial}^{2}{{\bf
A}^{\ast}}}{{\partial t}^{2}}= 0.\end{aligned}$$ The initial set of Maxwell’s equation has been decomposed into two independent sets of equations. The first one (21)-(22) answers for the instantaneous aspect (“[*action-at-a-distance"*]{}) of electromagnetic nature while the second one (23)-(24) is responsible for explicit time-dependent phenomena. The dualism as an intrinsic feature of Maxwell’s equations is evident. The potential separation (19)-(20) implies the same with respect to the field strengths: $$\begin{aligned}
&& {\bf E}={\bf E}_0({\bf R}(t))+{\bf E}^{\ast}({\bf r},t),\\
&& {\bf B}={\bf B}_0({\bf R}(t))+{\bf B}^{\ast}({\bf r},t),\end{aligned}$$ where ${\bf E}_0$ and ${\bf B}_0$ are instantaneous (NILI) fields. If we see again the formula (1) based on Liénard-Wiechert potentials, then in accordance with (25) the first term must be considered without “retarded time" (at a given instant of time $t$) and the whole expression will be as follows: $${\bf E}[{\bf R}(t),{\bf R}_0(t_0),t_0]=q\frac{{\bf
R}(1-\frac{V^2}{c^2})}{R^2(1-\frac{V^2}{c^2}{\sin}^2\Theta)^{3/2}}+{\bf
E}^{\ast}[{\bf a}({\bf R}_0,t)],$$ here $\Theta$ is the angle between the vectors [**V**]{} and [**R**]{}, ${\bf
a}({\bf R}_0,t_0)$ is the acceleration of the charge $q$ in the previous moment of time $t_0=t-\tau$, $\tau$ is the “retarded time". We note that the first term in (1) is mathematically equivalent to that in (27) (see \[9\]). In the steady state ([**a**]{}=0), the second term ${\bf
E}^{\ast}$ must be zero, so (27) can be consistent with the requirements of the Lorentz transformation. The same approach is applicable to the Liénard-Wiechert potentials. We leave out the complete modification of L.-W. potentials as well as an exact expression for ${\bf E}^{\ast}$ which, while of interest in itselves, have not direct connection with the following material.
To finish this section we conclude that NILI [*must exist as a direct consequence of Maxwell equations*]{}. According to this, both pictures, the NILI and the FMSI, have to be considered as two [*supplementary descriptions of one and the same reality*]{}. Each of the descriptions is only [*partly*]{} true. In other words, both Faraday and Newton in their external argument about the nature of interaction at a distance turned out right: instantaneous long-rang interaction takes place, not [*instead of*]{}, but [*along with*]{} the short-range interaction in the classical field theory.
Relativistic non-invariance of the concept of energy of self-field of a charge (self-energy concept). Mechanical analogy of Maxwell’s equations
===============================================================================================================================================
As a matter of fact, Maxwell’s equations lend themselves to covariant description and are in agreement with the requirements of relativity. In the previous section we have not modified the set of Maxwell’s equation, we have only separated two non-reducible parts in the general solution. Hence, the usual four-vector form of the basis equations can be used. For four-vectors of separated potentials we have automatically the following expressions: $$\Box(A_{0\mu}+A^{\ast}_{\mu})=-\frac{4\pi}{c}j_{\mu};
\qquad\qquad (\mu=0,1,2,3),$$ where $$A_{0\mu}+A^{\ast}_{\mu}=(\varphi_0+\varphi^{\ast},{\bf A}_0+{\bf
A}^{\ast});\qquad j_{\mu}=(c\varrho,{\bf j}).$$
To give some substance to the above formalism we exhibit explicitly the Poisson’s equation for instantaneous four-vector ${\bf A}_{0\mu}$: $$\Box A_{0\mu}=\Delta A_{0\mu}=-\frac{4\pi}{c}j_{\mu},$$ where $$A_{0\mu}=(\varphi_0({\bf R}(t)),{\bf A}_0({\bf R}(t))).$$ The Eq.(30) is an covariant also under Lorentz transformations. This is an exact consequence of (28) in the steady approximation. It is supported by the well-known fact that covariance is not necessary (it is sufficient) for the relativistic invariance. Nevertheless, in the Faraday-Maxwell electrodynamics this fact was always perceived as a quite odd one. Actually, potentials of an unmoving charge do not have explicit time-dependence. For a general Lorentz transformation from a reference system $K$ to an inertial system $K^{\prime}$ moving with the velocity ${\bf v}$ relative to $K$, the explicit time-dependence does not appear. Why those potentials keep implicit time-dependence under the Lorentz transformation? Without any approximation, the influence of a possible retarded effect is cancelled itself at any time and at any distance from the moving charge. From the other hand, the conventional theory is unable to describe correctly the transition from an uniform movement of a charge into an arbitrary one and then again into uniform over a limited interval of time. In this case, the first and the latter solutions at large distances can be given exactly by the Lorentz transformation. Furthermore the question arises: what mechanism changes this potentials at the distance unreachable for retarded Liénard-Wiechert fields? The lack of continuity between the corresponding solutions is obvious. It has the same nature as discussed in the above sections, due to incompleteness of existent solutions.
The new approach also highlights the invariant deficiency of the self-energy concept in the framework of relativity theory. We confine our reasoning to the example of the electrostatic. The total potential energy of $N$ charges due to all the forces acting between them is: $$W=\frac{1}{2}\sum^{N}_{i=1}\sum_{j\neq i}\frac{q_i q_j}
{\vert r_i-r_j\vert}.$$ Here, the infinite self-energy terms $(i=j)$ are omitted in the double sum. The expression obtained by Maxwell for the energy in an electric field, expressed as a volume integral over the field, is \[14\]: $$W=\frac{1}{2}\int_{\cal V} E^2 d{\cal V}.$$ This corresponds to the Maxwell’s idea that the system energy must be stored somewhere in space. The expression (33) includes self-energy terms and in the case of point charges they make infinite contributions to the integral. The introduction of a finite radius for the elementary charges enables to get rid of that difficulty but breaks down the possibility to see the classical electrodynamics as a self-consistent theory (Poincaré’s non-electrical forces \[15\]).
In spite of introducing the self-energy concept long before the special relativity principle had arisen, there was no much alarm about the fact that it did not satisfy the relativity invariance condition. Strictly speaking, the Einstein’s theory refutes the invariance of energy. The law of energy conservation cannot be maintained in its classical form. In a relativistically covariant formulation the conservation of energy and the conservation of momentum are not independent principles. In particular, the local form of energy-momentum conservation law can be written in a covariant form, using the energy-momentum tensor: $$\frac{{\partial T}^{\mu\nu}}{{\partial x}^{\nu}}=0.$$ For an electromagnetic field, it is well-known that (34) can be strictly satisfied only for a free field (when a charge is not taken into account), whereas, for the total field of a charge this is not true, since (34) is not satisfied mathematically (four-dimensional analogy of Gauss’s theorem). As everyone knows in classical electrodynamics, this fact gives rise to the “electromagnetic mass" concept, which violates the exact relativistic mass-energy relationship $({\cal E}=mc^2)$. Let us examine this problem in a less formal manner. The equivalent three-dimensional form of (34) is the formula (9). The amount of electrostatic energy of an unmoving charge in a given volume ${\cal V}$ is proportional to $E^2$ (see (33)). According to (34) (or (9)), in a new inertial frame $K^{\prime}$, this value $W$ must be, generally speaking, an explicit time-dependent function ($\partial w/\partial t\neq
0$). Furthermore, this means also the explicit time-dependence for the electric field ($\partial {\bf E}/\partial t\neq 0$). On the other hand, the electric field strength of an unmoving charge keeps its implicit time-dependent behavior under the Lorentz transformation ($\partial {\bf
E}/\partial t=0$). The conflict with the relativistic invariance condition is obvious. The analogous reasoning can be applied for Coulomb’s electrostatic energy of a system of charged particles. In this case, if one is thinking that electrostatic energy can be stored locally in space, the conflict with the relativity principle is inevitable. However, in the framework of the above-purposed separated potential’s method it is possible to avoid those difficulties. Actually, in the new general solution (25) it is ${\bf E}_0$ the only term exclusively linked to the charges. According to the above speculation, no local energy conservation law can be written for this field ${\bf E}_0$. The mathematical form (32) must be saved for it. But there is no cause to reject the local form for the time-dependent free field ${\bf
E}^{\ast}$. In fact, the mathematical expression (33) is adequate for it. Thus, if one wishes not to get into trouble with the relativity principle, one must distinguish two different terms in the total electric field energy: $$W=\frac{1}{2}\sum^{N}_{i=1}\sum_{j\neq i}\frac{q_i q_j}
{\vert r_i-r_j\vert}+\frac{1}{2}\int_{\cal V} E^{{\ast}2} d{\cal V}.$$
We should make one further remark about this energy formula. In first place, the dualism concept reveals the dual nature of the electromagnetic field energy. So, for instance, the total electric energy is the electrostatic energy plus the electric energy of the free electromagnetic field. The first term is non-zero if the system consists of at least two interacting charged particles. The second term is taken as an integral over the region of ${\cal V}$ where the local value of $E^{\ast}$ is not equal to zero. In the next section the correctness of this energy representation for all electromagnetic field will be strictly verified by applying the principle of least action. The introduction of the self-energy concept in XIX century’s physics can be explained historically. Maxwell considered the total electromagnetic field to be a uniform physical object on its own rights.
Removing the self-energy concept, a valuable mechanical analogy of the Maxwell’s equations in form of (21)-(24) can be used to understand why their general solution must be as separated potential’s (19)-(20). From the mathematical point of view, the two equations (21) and (22) correspond to the electro- and magnetostatic approximations, respectively, may be considered as wave equations with infinite spread velocity of longitudinal perturbations. If there is no local energy transfer, the Einstein’s theory does not limit the signal spreading velocities. In this case, the set of differential equations for elastic waves in an isotropic media (see \[16\]) can be treated as mechanical analogy of (21)-(24): $$\begin{aligned}
&& \frac{{\partial}^2 u_{\ell}}{{\partial t}^2}-
c^{2}_{\ell}\Delta u_{\ell}=0,\\
&& \frac{{\partial}^2 u_t}{{\partial t}^2}-c^{2}_{t}\Delta u_t=0.\end{aligned}$$ The general solution of (36)-(37) is the sum of two independent terms correspond to longitudinal $u_{\ell}$ and transverse $u_t$ waves: $$u=u_{\ell}+u_t.$$ If the longitudinal spreading velocity approaches formally to infinity ($c_{\ell}\rightarrow\infty$) then (36) transforms into Laplace’s equation whereas the function $u_{\ell}$ turns out to have an implicit time-dependence. Thus, the formula (38) takes the form of separated potential’s solution (19)-(20).
To end this section, we note that the idea of non-local interactions can be immediately derived from the Maxwell’s equations as an exact mathematical result. On the other hand, some of the quantum mechanical effects like Aharonov-Bohm effect \[17\], violation of the Bell’s inequalities \[18-19\] etc. point out indirectly to the possibility of non-local interactions in electromagnetism. Nevertheless, in this work we prefer to confine themselves to the classical theory.
Hamiltonian form of Maxwell’s equations from the point of view of separated potential’s method
==============================================================================================
In the latter section we have introduced the prototype for a new electromagnetic energy interpretation. In this section we shall discuss general field equations for arbitrary fields from the standpoint of the principle of least action and the change in their interpretation due to the new dualism concept. In extending the separated potential’s method no modifications at all are necessary in the set of Maxwell’s equations to make them agree with the requirements of the covariant formulation. Hence, in the steady approximation ($\varphi^{\ast}=0, {\bf A}^{\ast}=0$) a relativistic action for a system of interacting charged particles can be written in conventional form \[9\]: $$S_{m}+S_{mf}=\int\Bigl(-\sum^{N}_{a=1}m_acds_a-
\sum^{N}_{a=1}\frac{e_a}{c}\sum_{\mu=0}A_{0(\mu a)}dx^{\mu}_{a}\Bigr),$$ where $A_{0(ma)}$ is the [*instantaneous*]{} potential ($\varphi_0,
{\bf A}_0$) in the four-point on the world-line of the particle with the number “$a$" created by other particles. This expression is sufficient to derive the first couple of equations (21)-(22) from the least action principle. It can be proved directly rewriting the second term in (39) as: $$S_{mf}=-\frac{1}{c}\int\sum_{\mu}A_{0\mu}j^{\mu}d{\cal V}dt,$$ using the Dirac’s expression for four-current: $$j_{\mu}({\bf r},t)=\sum_{a}\Bigl[\frac{e_a}{4\pi}\Delta
\Bigl(\frac{1}{\vert {\bf r}-{\bf r}_a\vert}\Bigr)\Bigr]u_{\mu a},$$ where $u_{\mu a}$ - four-velocity of the charged particle “$a$". Generally, for a system of arbitrary moving charges, the time-dependent potentials ($\varphi^{\ast},{\bf A}^{\ast}$) appear in the general solution. It means that an additional term corresponding to the free electromagnetic field must be added to (39). In the first place, it must vanish under the transition to the steady approximation ($\varphi^{\ast}=0,{\bf A}^{\ast}=0$). On the other hand, the variation of this term has to lead to the second pair of equations (23)-(24). As a result, it is easy to see that conventional Hamiltonian form can be adopted to describe the presence of the free electromagnetic field \[9\]: $$S_f=-\frac{1}{16\pi}\int\sum_{\mu,\nu}F_{\mu\nu}F^{\mu\nu}d{\cal V}dt,$$ where $$F_{\mu\nu}=\frac{\partial A^{\ast}_{\nu}}{\partial x^{\mu}}-
\frac{\partial A^{\ast}_{\mu}}{\partial x^{\nu}}.$$ Finally, it remains to be proved that from the variation derivative: $$\delta S_f=-\int\sum_{\mu}\Bigl(\frac{1}{4\pi}\sum_{\nu}\frac
{\partial F^{\mu\nu}}{\partial x^{\nu}}\Bigr)\delta A^{\ast}_{\mu}
d{\cal V}dt$$ one obtains the covariant analogous of (23)-(24) in the following form: $$\sum_{\nu}\frac{\partial}{\partial x^{\nu}}F^{\mu\nu}=
\sum_{\nu}\frac{\partial}{\partial x^{\nu}}
\Bigl[\frac{\partial A^{\ast\nu}}{\partial x_{\mu}}-
\frac{\partial A^{\ast\mu}}{\partial x_{\nu}}\Bigr]=0.$$ The only difference with the classical field interpretation consists in the way how electromagnetic potentials take part in this Hamiltonian formulation. Actually, the second term in (39) contains only instantaneous potentials whereas $S_f$ is related with time-varying field components. Consequently, contrary to traditional interpretation, the quantity $F^{\mu\nu}$ can be defined as a free electromagnetic field tensor.
In the light of the new approach, the electromagnetic energy-momentum tensor demands some corrections in the interpretation of its formal mathematical formulation \[9\]: $$T^{\mu\nu}=-\frac{1}{4\pi}\sum_{\rho}F^{\mu}_{\rho}F^{\nu\rho}+
\frac{1}{16\pi}{\sl g}^{\mu\nu}
\sum_{\beta,\gamma}F_{\beta\gamma}F^{\beta\gamma}.$$ As a consequence of (43), in this form it can describe the energy-momentum conservation law for, exclusively, free electromagnetic field as follows: $$\sum_{\nu}\frac{\partial T^{\mu\nu}}{\partial x^{\nu}}=0$$ which supports the new interpretation of electric field energy given in the previous section. Strictly speaking, from the point of view of the dualism concept, the total field energy $W$ must consist of two non-compatible parts: on one hand, the energy $W_{mf}$ of electro- and magnetostatic interaction between charges and currents ([*non-local*]{} term), on the other hand, the energy $W_f$ of the free electromagnetic field ([*local*]{} term): $$W=W_{mf}+W_f.$$ It contradicts considerably the FMSI concept about the unique nature of electromagnetic field energy. Summarizing these results we see that the concept of potential ([*non-local*]{}) energy and potential forces must be conserved in the classical electrodynamics as valid. So, the system of charges and currents in the absence of free electromagnetic field must be considered as conservative system without any idealization. As an important remark we note the physical meaning of Poynting vector has been changed notably. So far the classical theory dealt with it as a quantity attributed to all dynamic properties of the total electromagnetic field. From the new point of view, it can be non-zero only in the presence of [*free*]{} field. The great problem of the classical electrodynamics, the indefiniteness in the location of the field energy, does not exist anymore. In particular, the flux of the electromagnetic energy in the steady state has no sense since no presence of the free electromagnetic field is supposed in this case.
Non-radiation condition for free electromagnetic field
======================================================
In this section we shall discuss the energy balance between the system of interacting charged particles and free electromagnetic field, namely, energy and momentum lost by radiation. Turning to the latter section results we must examine carefully one essential difference in electromagnetic energy interpretation. Let us write the total relativistic action as: $$S=S_m+S_{mf}+S_f.$$ Although we adopt the same denominations used in the conventional theory, the physical essence of the last two terms has changed significantly. Usually, the interaction between particles and electromagnetic field was attributed to $S_{mf}$ whereas the properties of electromagnetic field manifested itselves by the additional term $S_f$.
In the new approach, no concept of field as intermediary is needed to describe the interaction between charges (currents). Hence, $S_{mf}$ cannot be treated in terms of particle-field interaction. Such interaction as well as the intrinsic properties of a free electromagnetic field are enclosed now in the last term $S_f$. The possible free field interaction with the system of charges (currents) depends entirely on its location in space. This reasoning makes it possible to consider the isolated system of charged particles and free field as consisting of two corresponding subsystems. Each of the subsystems may be completely independent if there is no mutual interaction (for instance, free electromagnetic field is located far from the given region of charges and currents). In the steady approximation the first subsystem (charges and currents) can be considered as conservative. In other words, it means the total Hamiltonian of the whole isolated system can be decomposed into two corresponding parts: $${\cal H}={\cal H}_1+{\cal H}_2,$$ where ${\cal H}_1$ is the Hamiltonian of the conservative system of charges and currents. It involves apart from electro- and magnetostatic energy also mechanical energy of particles (corresponds to the action $S_m + S_{mf}$). ${\cal H}_2$ is the Hamiltonian of the free electromagnetic field (corresponds to the action $S_f$).
We remind here that in the relativistic case, the energy is the zero component of the momentum. However, if we deal with the isolated system, the total Hamiltonian is not time-dependent and the energy conservation law as well as the momentum conservation may be treated independently. It is important to note that such separation into two subsystems is valid only in the new approach. The conventional interpretation of $S_f$ did not allow to consider it apart. Actually, in the steady approximation $S_f$ was not zero, and corresponded to the self-energy of field \[9\]: $$S_f=\int^{t_1}_{t_2}{\cal L}_{f}dt,$$ where $${\cal L}_f=\frac{1}{8\pi}\int_{\cal V}(E^2-B^2)d{\cal V}.$$ Here ${\bf E}$ and ${\bf B}$ are the total electro- and magnetic field strengths, respectively. Thus, the fact that $S_f$ is responsible solely for free field, turns out to be a meaningful argument in separating into two subsystems. It is often possible to extract a large amount of information about the physical nature of the system using conservation laws, even when complete solutions cannot be obtained. Let us now consider the case when charges (currents) and free electromagnetic field are located in the same region and become interacting. Internal forces of mutual reaction between two subsystems are usually named as internal dissipative forces. They carry out the energy exchange inside the total isolated system. In terms of the Hamiltonian formalism it can be expressed as a corresponding Hamiltonian evolution (see, for instance, \[20\]): $$\frac{d{\cal H}_{1,2}}{dt}=\frac{\partial {\cal H}_{1,2}}{\partial t}+
{\cal P}^{\tt ext}_{1,2}+{\cal P}^{\tt int}_{1,2},$$ where ${\cal P}_{1,2}^{\tt ext}({\cal P}_{1,2}^{\tt int})$ is the power of the external (internal) forces acting on two the subsystems, respectively. In our case ${\cal P}_{1}^{\tt ext}$ and ${\cal P}_{2}^{\tt ext}$ appear as a result of the mutual interaction. On the other hand, any internal non-potential force in the first subsystem can also cause energy dissipation (${\cal P}_{1}^{\tt int}$). Even in the absence of a real mechanical friction, other internal non-potential forces (for example, inhomogeneous gyroscopic forces) can still act in this subsystem and dissipate energy. In other words, if initially there is no free electromagnetic field (${\cal H}_2=0$), it can be created by internal non-potential forces (${\cal P}_{1}^{\tt int}$) acting in the first subsystem (${\cal H}_2$ stops to be zero). It means that energy is lost by radiation in the subsystem of charges and currents. In mathematical language the corresponding energy balance can be written as follows: $$\frac{d}{dt}({\cal H}_1+{\cal H}_2)=
\dot{\cal H}_1+\dot{\cal H}_2=0,$$ where $\dot{\cal H}_1$ and $\dot{\cal H}_2$ are energy change rates for the first and the second subsystems respectively. It might be easily noted that the energy balance (54) is symmetrical in respect to time reversion which is in accordance with the time symmetry of Maxwell’s equations. The real direction of the energy exchange process may be determined by some subsidiary conditions. On the contrary to this, the energy balance in the conventional electrodynamics was always irreversible in time. From the other hand, the former class of theories based on the action-at-a-distance principle (for example, the electrodynamics of Wheeler and Feynman) did not consider at all the third term $S_f$ in (49), corresponding to radiation reaction. As a matter of fact, there were no radiation effects in those theories, but only interactions of a number of particles.
To end the section we formulate the previous statement about the energy conservation as [**the condition of non-radiation of the free electromagnetic field**]{}:
[*If in an isolated system of charges (currents) in the absence of free electromagnetic field (${\cal H}_2=0$), all internal non-potential forces are compensated or do not exist then this system will not produce (radiate) free electromagnetic field (${\cal H}_2$ remains zero) and will keep conservative system itself*]{}.
This implies not only an equilibrium between radiation and absorption but no radiation at all. As a simple example of a non-radiating system we can consider two charged particles moving with acceleration along a direct line under mutual Coulomb interaction. The absence of other frictional forces is supposed. The presence of any inhomogeneous gyroscopic (Lorentz-type) forces here are not expected due to the one-dimensional character of motion. Some mention should also be made of the many-particle system. It is possible that there is some limited class of motion when all non-potential (for, example, internal gyroscopic forces) can be compensated due to the own magnetic moment of charged particles. This possibility would be of particular interest in the attempt to understand the quantum mechanics principles.
In the present work the question of interaction of free fields with sources (charges and currents) is given in perfunctory manner and should be studied carefully. It should be compared with the older non-radiation theories based on the [*extended*]{} Dirac electron models (see, for instance, \[21\]). Furthermore, emission, absorption and, for instance, scattering processes can be caused by the interaction of matter fields with the $B(3)$ spin field. It is created by transverse left- and right- circular polarized waves, as found by Evans and Vigier \[22-25\]. On the other hand, the existence of the longitudinal $B(3)$ field may hint on non-zero photon mass. Theoretical constructs of such a type were introduced and developed by Einstein, Schrödinger, Deser, de Broglie and Vigier (see, e.g. \[26\]). However, relations between $B(3)$ and other longitudinal solutions of Maxwell equations, as well as the problem of photon mass, must be studied more carefully.
Conclusions
===========
Finally, we conclude that the FMSI concept could not give a complete and adequate description of the great variety of electromagnetic phenomena. It has been shown that other concept (the so-called [*dualism concept*]{}), consistent with the full set of Maxwell’s equations, can be accepted as a correct description of electromagnetism. In other words, the new concept states that there is a [*simultaneous*]{} and [*independent*]{} coexistence of Newton instantaneous long-range (NILI) and Faraday-Maxwell short-range interactions (FMSI) which cannot be reduced to each other. The reasons are based on the mathematical method (so-called [*separated potential’s method*]{}) proposed in this work for a complete general solution of Maxwell’s equations. As a result, the incompleteness of former solutions of Maxwell’s equations is proved.
In the framework of the new approach, all main concepts of the classical electrodynamics have been reconsidered. In particular, it has been shown that the dual nature of the total electromagnetic field must be taken into consideration. On one hand, there is a free electromagnetic field ${\bf E}^{\ast} ({\bf B}^{\ast})$ which has no direct connection with charges and currents, and can be transferred [*locally*]{}. On the other hand, there is a field ${\bf E}_0 ({\bf B}_0)$ linked exclusively to charges (currents) and responsible for interparticle interaction which [*cannot be transferred locally*]{} in space. However, in total, this two kinds of electromagnetic field ${\bf E}_0+{\bf E}^{\ast}$ $({\bf
B}_0+{\bf B}^{\ast})$ as a superposition satisfy Maxwell’s equations and are observed experimentally as an unique electromagnetic field. Other quantities of the classical electrodynamics such as electromagnetic field tensor, electromagnetic energy-momentum tensor etc. have also changed their physical meanings. In particular, the Poynting vector can be associated [*only with the free electromagnetic field*]{}. In the light of this result, the problem of the indefiniteness in the field energy location has no place and no flux of electromagnetic energy in steady state can be derived from the theory. Also problems such as [*self-force, infinite contribution of self-energy, the concept of electromagnetic mass, radiation irreversibility in time with respect of time symmetry of Maxwell equations*]{} have been removed in the new approach. A new interpretation of the energy conservation law is possible as a non-radiation condition which states that [*a limited class of motion exists when accelerated charged particles do not produce electromagnetic radiation*]{}.
[Acknowledgments]{}
We are grateful to Dr. V. Dvoeglazov and Professor J. V. Narlikar for many stimulating discussions. We acknowledge papers of Professor M. W. Evans, which gave support to our belief that ideas presented here have sufficient grounds. Authors are indebted for financial support, R. S.-R., to the Comunidad de Madrid, Spain, for the award of a Postgraduate Grant, A. Ch., to the Zacatecas University, México, for a Full Professor position.
J.V. Narlikar: [*Astrofisica e Cosmologia Gravitazione Quanti e Relativita (Negli sviluppi del pensiero scientifico di Albert Einstein. “Centenario di Einstein" 1879-1979*]{}. (Giunti Barbera, Firenze 1979). K.Schwarzchild: [*Gottinger Nachrichten*]{}, [ **128**]{}, 132 (1903). H.Tetrode: [*Zeits. f. Physik*]{}, [**10**]{}, 317 (1922). Frenkel: [*Zeits. f. Physik*]{}, [**32**]{}, 518 (1925). J.A. Wheeler and R.P.Feynman: [*Revs. of Mod. Phys.*]{}, [**17**]{}, 157-181 (1945). J.A. Wheeler and R.P.Feynman: [*Revs. of Mod. Phys.*]{}, [**21**]{}, 425-433 (1949). A. Liénard: [*L’Éclairage Électrique*]{}, [**16**]{}, pp.5,53,106 (1898). E. Wiechert: [*Ann. d. Physik*]{}, [**4**]{}, 676 (1901). L.D. Landau and E.M. Lifshitz: [*Teoria polia*]{} (Nauka, Moscow 1973), (English translation: [*Classical Theory of Field*]{}, Pergamon Press, Oxford/New York, 1985). C.K. Whitney: [*Hadronic J.*]{}, [**11**]{}, pp.107,257 (1988). J.D. Jackson: [*Classical Electrodynamics*]{}, (John Wiley & Sons, New York/London, 1963). P.A.M. Dirac: [*Directions in Physics*]{}, (John Wiley & Sons, New York, 1978). A.N. Tikhonov and A.A. Samarski: [*Equations of Mathematical Physics*]{}, (Pergamon Press, Oxford/New York, 1963). J.C. Maxwell: [*Treatise on Electricity and Magnetism*]{}, (Oxford Univ. Press, London, 1892). H. Poincaré: [*Rend. Palermo*]{}, [**21**]{}, 165 (1906). L.D. Landau and E.M. Lifshitz: [*Theory of Elasticity*]{}, (Pergamon Press, Oxford/New York, 1970). Y. Aharonov and D.Bohm: [*Phys. Rev.*]{}, [**115**]{}, 115 (1959). J.S. Bell: [*Revs. Mod. Phys.*]{}, [**38**]{}, 447 (1966). B. D’Espagnat: [*Phys. Repts*]{}, [**110**]{}, 201 (1984). H. Goldstein: [*Classical Mechanics*]{}, (Addison-Wesley Publishing Company Inc., London, 1959). D. Gutkowski, M. Moles and J.-P. Vigier: [*Il Nuovo Cimento*]{}, [**39B**]{}, N 1, 193-224 (1977). M.W. Evans: [*Physica B*]{}, [**182**]{}, 227 (1992). M.W. Evans: [*Mod. Phys. Lett.*]{}, [**7**]{}, 1247 (1993); [*Found. Phys. Lett.*]{}, [**7**]{}, 67 (1994). M.W. Evans and S. Kielich eds.: [*Modern Nonlinear Optics*]{}, [**85**]{}(1),[**85**]{}(2),[**85**]{}(3) [*of Advances in Chemical Physics*]{}, (Wiley Interscience, New York, 1993-1994). M.W.Evans and J.-P. Vigier: [*“The enigmatic photon"*]{}, [**1**]{},[**2**]{}, (Kluwer Academic, 1994,1995). L. de Broglie and J.-P. Vigier: [*Phys. Rev. Lett.*]{}, [**28**]{}, N 15, 1001-1004 (1972).
[^1]: Instituto de Ciencia de Materiales, C.S.I.C., Madrid, Spain
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Large scale, homogeneous quasi-free standing monolayer graphene is obtained on cubic silicon carbide, i.e. the 3C-SiC(111) surface, which represents an appealing and cost effective platform for graphene growth. The quasi-free monolayer is produced by intercalation of hydrogen under the interfacial, [[(6$\sqrt{3}$$\times$6$\sqrt{3}$)R30$^{\circ}$]{}]{}-reconstructed carbon layer. After intercalation, angle resolved photoemission spectroscopy reveals sharp linear [[$\pi$-bands]{}]{}. The decoupling of graphene from the substrate is identified by X-ray photoemission spectroscopy and low energy electron diffraction. Atomic force microscopy and low energy electron microscopy demonstrate that homogeneous monolayer domains extend over areas of hundreds of square-micrometers.'
author:
- 'C. Coletti'
- 'K.V. Emtsev'
- 'A.A. Zakharov'
- 'T. Ouisse'
- 'D. Chaussende'
- 'U. Starke'
title: 'Large area quasi-free standing monolayer graphene on 3C-SiC(111)'
---
The unique two-dimensional electron gas properties of graphene make it a potential candidate for future electronics. Currently, the main techniques adopted to produce single or few-layer graphene are mechanical exfoliation of graphite [@Novoselov2004], graphitization of silicon carbide (SiC) [@Berger2004], chemical vapor deposition on transition metals [@Sutter2008], and chemical synthesis [@Gomez2007]. By enabling growth of large area graphene directly on a semi-insulating substrate, thermal decomposition of SiC is the most promising route towards a future of carbon-based nano-electronics [@Berger2004; @Emtsev2009]. Hexagonal SiC crystals, namely 4H- and 6H-SiC, provide an ideal template for graphene growth and thus have evolved as the substrates of choice in the past years [@First2010]. In contrast, limited attention has been given to graphene growth on cubic SiC (3C-SiC), although the \[111\] orientation of this crystal would also naturally accommodate the six-fold symmetry of graphene. With its extreme robustness and proven biocompatibility [@Coletti2007IEEE], cubic SiC is an appealing platform for the growth of graphene that could lead to a new generation of microelectromechanical systems and advanced biomedical devices. Moreover, cubic SiC can be epitaxially grown on Si crystals and - provided the process temperatures can be sufficiently lowered - this could reduce the production costs of graphene.
To date, the growth of graphene on 3C-SiC(111) was attempted by adopting the classical ultra-high-vacuum (UHV) thermal decomposition process [@Ouerghi2010APL96]. The structural and electronic properties of such graphene were comparable to those of graphene on SiC(0001) [@Ouerghi2010PRB; @Ouerghi2010APL97]. The first carbon layer that grows on top of SiC(111) is known as zero-layer graphene (ZLG) [@Emtsev2009]. It is in part covalently bound to the SiC substrate and hence electronically inactive. The second carbon layer grows on top of the ZLG without interlayer bonds and acts like monolayer graphene. Despite the good quality on the nanometer scale, the lateral size of homogenous graphene domains produced to date on 3C-SiC(111) is limited to about 1 $\mu$m [@Ouerghi2010PRB; @Ouerghi2010APL97].
In the present work we report a truly large-scale production of epitaxial graphene on 3C-SiC(111). We demonstrate that on cubic substrates it is indeed possible to obtain homogenous monolayer graphene with domains extending over areas of hundreds of square-micrometers. To achieve this result, we combine the method of atmospheric pressure (AP) growth [@Emtsev2009] with the hydrogen intercalation technique recently developed by our group [@Riedl2009PRL]. In the first part of this work, the morphologies of 3C-SiC(111) and of ZLG obtained using different processing parameters are probed using atomic force microscopy (AFM). Subsequently, ZLG is hydrogen intercalated in order to obtain quasi free-standing monolayer graphene (QFMLG). The chemical, electronic and structural properties of the resulting films are investigated by means of X-ray photoemission spectroscopy (XPS), angle resolved photoemission spectroscopy (ARPES), and low energy electron diffraction (LEED) and microscopy (LEEM). XPS measurements were performed using photons from a non-monochromatic Mg K$_{\alpha}$ source (h$\nu$ = 1253.6 eV). ARPES analysis was carried out using monochromatic He II radiation (h$\nu$ = 40.8 eV) from a UV discharge source with a display analyzer oriented for momentum scans perpendicular to the [[$\overline{\Gamma \mbox{K}}$]{}]{}-direction. LEEM experiments were performed using the ELMITEC LEEM III instrument at beamline I311 of the MAX radiation laboratory in Lund (Sweden).
![(Color online) AFM micrographs and related line profile for: (a) as-grown 3C-SiC(111), (b) UHV grown ZLG on 3C-SiC(111), (c) hydrogen etched 3C-SiC(111), (d) AP-grown ZLG on 3C-SiC(111). Inset axes: nm (horizontal) and $\mu$m (vertical).[]{data-label="AFM"}](fig1.eps){width="6.0"}
In order to obtain large-area homogeneous graphene, high-quality 3C-SiC(111) substrates are required. The 3C-SiC(111) samples used in this work were grown by hetero-epitaxy on a 4H-SiC(0001) substrate with the continuous-feed physical-vapour transport method [@Chaussende2005]. In this way, cubic epilayers with a thickness of $\sim$390 $\mu$m were obtained and subsequently rendered free-standing by polishing away the 4H-SiC substrate. The high quality and crystallinity of the epilayers was verified in optical microscopy which revealed an almost complete absence of double positioning boundaries [@Chaussende2005]. No anti-phase domain boundaries, as those found on 3C-SiC(001) [@Coletti2007APL], were observed. AFM analysis confirmed a high degree of order at an atomic level. The surface consists of atomically flat terraces with 0.5 to 1 $\mu$m widths separated by steps with multiple unit cell heights (1 to 3 nm), cf. Fig. \[AFM\](a). For comparison purposes, growth of ZLG was first attempted using the UHV graphitization method described in [@Riedl2007PRB]. The resulting surface morphology is shown in Fig. \[AFM\](b). Although the original step structure remains visible, preferential Si desorption sites make their initial appearance, thus contributing to a rough and inhomogeneous morphology typical of UHV prepared graphene. Notably, the zero-layer could be removed by etching in a hydrogen atmosphere at a temperature of 1250[[$^{\circ}$C]{}]{} for 15 minutes. However, the terraces of the reverted SiC surfaces present zig-zag edges with facets oriented 60 with respect to each other, indicating etching along two preferential directions, cf. Fig. \[AFM\](c).
![(Color online) (a) LEED pattern of the [[(6$\sqrt{3}$$\times$6$\sqrt{3}$)R30$^{\circ}$]{}]{}-reconstructed ZLG (188 eV) (b) $\mu$-LEED image of QFMLG on 3C-SiC(111) (98 eV) (c) C1s core level emission region for as-grown ZLG (bottom curve) and QFMLG on 3C-SiC(111) (top curve). The experimental data are displayed in black dots. Different components, accordingly labeled, are fitted by a line shape analysis. The gray solid line is the envelope of the fitted components. (d) Dispersion of the [[$\pi$-bands]{}]{} measured with ARPES for QFMLG on 3C-SiC(111).[]{data-label="XPS"}](fig2.eps){width="8.7"}
The typical morphology of ZLG obtained annealing 3C-SiC(111) at 1400[[$^{\circ}$C]{}]{} in 800 mbar argon for 10 minutes is displayed in Fig. \[AFM\](d). The surface is highly homogeneous, with 2 to 4 $\mu$m large terraces and 6 to 15 nm high steps, which are indicative of an increased step bunching (note the image’s larger scale). The improved morphology in panel (d) can be ascribed to the AP growth method used [@Emtsev2009]. It requires higher growth temperatures which favor diffusion of the atoms so that the restructuring of the surface is completed before graphene is formed. Notably, the surface morphology shown in panel (d) could be obtained when starting from both as-grown (panel (a)) and etched (panel (c)) surfaces. The LEED pattern of the AP grown ZLG displayed in Fig. \[XPS\](a) shows the fractional order spots of the [[(6$\sqrt{3}$$\times$6$\sqrt{3}$)R30$^{\circ}$]{}]{} reconstruction around the first order diffraction spots of the SiC(111) substrate. At 188 eV electron energy the extinction of three of these six first order spots corroborates the complete absence of twinning domains in the SiC substrate.
Hydrogen intercalation of the ZLG samples was achieved by annealing at 850[[$^{\circ}$C]{}]{} in a molecular hydrogen atmosphere of 1 bar [@Riedl2009PRL] and controlled by ARPES and XPS. Before intercalation the ARPES analysis reveals the surface states characteristic for the [[(6$\sqrt{3}$$\times$6$\sqrt{3}$)R30$^{\circ}$]{}]{} reconstructed ZLG [@Riedl2009PRL; @Emtsev2008] (not shown). Also the corresponding C1s core level spectrum from XPS measurements, shown in the bottom part of Fig. \[XPS\](c), is typical for ZLG. It can be well fitted with three components [@Emtsev2008], namely the SiC bulk component positioned at 283.5 eV and the two zero-layer components indicated in the literature as S1 and S2 located at 284.8 eV and 285.4 eV, respectively. The decoupling of the graphene layer after hydrogen intercalation (weak superstructure [@double]) and its long-range order (sharp spots) are visible from the $\mu$-LEED-pattern shown in Fig. \[XPS\](b).
![(Color online) (a) LEEM micrograph with FOV of $\sim$57 $\mu$m recorded with an electron energy of 19 eV. (b) LEEM micrograph with FOV of $\sim$23 $\mu$m recorded with an electron energy of 4 eV and labeled representative region A, B and C. (c) Electron reflectivity spectra measured for A, B and C (the curves are shifted on the y-axis for better display).[]{data-label="LEEM"}](fig3.eps){width="8.7"}
After the intercalation process sharp monolayer [[$\pi$-bands]{}]{} develop as shown in the ARPES dispersion plot in Fig. \[XPS\](d). In the XPS spectrum (see top curve of Fig. \[XPS\](c)) the disappearance of the S1 and S2 components is accompanied by a shift of the SiC component to lower binding energies (i.e., 283.2 eV) and the emergence of a peak located at 284.7 eV imputable to monolayer graphene (indicated “g” in the figure). These results indicate [@Riedl2009PRL] that hydrogen atoms migrate under the ZLG and bind to the topmost Si atoms of the 3C-SiC(111) substrate. Hence the ZLG, freed from the covalent bonds and decoupled from the substrate, becomes what has been named QFMLG [@Riedl2009PRL]. However, different from the case of hexagonal SiC crystals [@Riedl2009PRL], QFMLG on 3C-SiC(111) is found to be slightly n-doped. Both ARPES and XPS data indicate the electron doping to be roughly 10$^{12}$ cm$^{-2}$. The origin for the doping is currently under investigation.
Evaluation of the homogeneity of the QFMLG was carried out using LEEM. Figure \[LEEM\](a) displays a characteristic LEEM micrograph recorded with a field of view (FOV) of $\sim$57 $\mu$m. Three different gray levels are visible. These domains - which may be areas of different thickness - are magnified (23 $\mu$m FOV) in panel (b), and labeled A, B and C. For these representative regions LEEM reflectivity spectra are plotted in panel (c). The number of minima in such spectra in the energy regime below $\sim$6 eV identifies the number of graphene layers [@Hibino2008; @Riedl2009PRL]. The spectra indicate that the sample is homogenously covered by monolayer graphene (regions A and B) while bilayer graphene is only present at the step edges (region C). Note, that in contrast to pristine epitaxial graphene, an additional minimum around 6-7 eV appears for H-intercalated graphene [@Forti_unpubl]. Hence, the lighter gray contrast observed along certain crystal orientations (region B) is not indicative of thickness inhomogeneities. We speculate that it could be caused by defects in the substrate, which might be mediated by strain. LEEM analysis also reveals that the large atomically flat macro-terraces homogenously covered by QFMLG run uninterrupted for hundreds of micrometers. This thickness homogeneity is highly remarkable considering that, until now, the maximum lateral dimension reported for homogenous graphene domains on SiC(111) was roughly 1 $\mu$m [@Ouerghi2010APL97].
In conclusion, this work demonstrates that large area QFMLG can be produced on 3C-SiC(111) substrates. The morphological, structural and electronic properties of such layers are fully investigated. The high quality of the graphene obtained suggests that 3C-SiC(111) might be an appealing and cost effective platform for the future development of graphene technology.
C.C. acknowledges the Alexander von Humboldt Foundation for financial support. This research was partially funded by the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 226716. We are indebted to the staff at MAX-Lab (Lund, Sweden) for their advice and support.
[00]{}
K.S. Novoselov, A.K. Geim, S.V. Morozov, D. Jiang, Y. Zhang, S.V. Dubonos, I.V. Grigorieva, A.A. Firsov, Science **306**, 666 (2004). C. Berger, Z. Song, T. Li, X. Li, A. Y. Ogbazghi, R. Feng, Z. Dai, A. N. Marchenkov, E. H. Conrad, P. N. First, W. A. De Heer, J. Phys. Chem. B **108**, 19912 (2004). P.W. Sutter, J.-I. Flege, E.A. Sutter, Nat. Mater. **7**, 406 (2008).
C. Gómez-Navarro, R.T. Weitz, A.M. Bittner, M. Scolari, A. Mews, M. Burghard, K. Kern, Nano Lett. **7**, 3499 (2007)
K.V. Emtsev, A. Bostwick, K. Horn, J. Jobst, G.L. Kellogg, L. Ley, J.L. McChesney, T. Ohta, S.A. Reshanov, E. Rotenberg, A.K. Schmid, D. Waldmann, H.B. Weber, T. Seyller, Nat. Mater. **8**, 203 (2009). P.N. First, W.A. de Heer, T. Seyller, C. Berger, J.A. Stroscio, J-S. Moon, MRS Bulletin, **35**, 296 (2010). C. Coletti, M. J. Jaroszeski, A. Pallaoro, A. M. Hoff, S. Iannotta, S.E. Saddow, Conf. Proc. IEEE Eng. Med. Biol. Soc. 2007, 5850 (2007). A. Ouerghi, A. Kahouli, D. Lucot, M. Portail, L. Travers, J. Gierak, J. Penuelas, P. Jegou, A. Shukla, T. Chassagne, M. Zielinski, Appl. Phys. Lett. **96**, 191910 (2010). A. Ouerghi, M. Marangolo, R. Belkhou, S. El Moussaoui, M. G. Silly, M. Eddrief, L. Largeau, M. Portail, B. Fain, F. Sirotti, Phys. Rev. B **82**, 125445 (2010). A. Ouerghi, R. Belkhou, M. Marangolo, M.G. Silly, S. El Moussaoui, M. Eddrief, L. Largeau, M. Portail, F. Sirotti, Appl. Phys. Lett. **97**, 161905 (2010). C.Riedl, C. Coletti, T. Iwasaki, A.A. Zakharov, U. Starke, Phys. Rev. Lett. **103**, 246804 (2009). D. Chaussende, L. Latu-Romain, L. Auvray, M. Ucar, M. Pons, R. Madar, Mat. Sci. Forum **483-485**, 225 (2005). C. Coletti, C.L. Frewin, S.E. Saddow, M. Hetzel, C. Virojanadara, U. Starke, Appl. Phys. Lett. **91**, 061914 (2007). C. Riedl, U. Starke, J. Bernhardt, M. Franke, K. Heinz, Phys. Rev. B **76**, 245406 (2007). K.V. Emtsev, F. Speck, Th. Seyller, L. Ley, J.D. Riley, Phys. Rev. B **77**, 155303 (2008). The weak fractional order spots remaining can be attributed to double diffraction between SiC and graphene.
H. Hibino, H. Kageshima, F. Maeda, M. Nagase, Y. Kobayashi, H. Yamaguchi, Phys. Rev. B **77**, 075413 (2008). S. Forti, K.V. Emtsev, C. Coletti, A.A. Zakharov, C. Riedl, and U. Starke, to be published.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a minimal model of one-dimensional discrete-time random walk with step-reinforcement, introduced by Harbola, Kumar, and Lindenberg (2014): The walker can move forward (never backward), or remain at rest. For each $n=1,2,\cdots$, a random time $U_n$ between $1$ and $n$ is chosen uniformly, and if the walker moved forward \[resp. remained at rest\] at time $U_n$, then at time $n+1$ it can move forward with probability $p$ \[resp. $q$\], or with probability $1-p$ \[resp. $1-q$\] it remains at its present position. For the case $q>0$, several limit theorems are obtained by Coletti, Gava, and de Lima (2019). In this paper we prove limit theorems for the case $q=0$, where the walker can exhibit all three forms of asymptotic behavior as $p$ is varied. As a byproduct, we obtain limit theorems for the cluster size of the root in percolation on uniform random recursive trees.'
address:
- 'Graduate School of Engineering Science, Yokohama National University, Yokohama, Japan'
- 'Department of Applied Mathematics, Faculty of Engineering, Yokohama National University, Yokohama, Japan'
author:
- Tatsuya Miyazaki
- Masato Takei
title: 'Limit theorems for the ‘laziest’ minimal random walk model of elephant type'
---
Introduction {#intro}
============
The elephant random walk, introduced by Schütz and Trimper [@SchutzTrimper04], is defined as follows:
- The first step $Y_1$ of the walker is $+1$ with probability $s$, and $-1$ with probability $1-s$.
- For each $n=1,2,\cdots$, let $U_n$ be uniformly distributed on $\{1,\cdots,n\}$, and $$\begin{aligned}
Y_{n+1} &= \begin{cases}
Y_{U_n} &\mbox{with probability $p$}, \\
-Y_{U_n} &\mbox{with probability $1-p$}. \\
\end{cases}\end{aligned}$$
Each of choices in the above procedure is made independently. The sequence $\{Y_i\}$ generates a one-dimensional random walk $\{S_n\}$ by $$S_0:=0,\quad \mbox{and} \quad S_n:= \sum_{i=1}^n Y_i \quad \mbox{for $n=1,2,\cdots$.}$$ It admits a phase transition from diffusive to superdiffusive behavior at the critical value $p_c=3/4$. Several limit theorems are obtained in [@BaurBertoin16; @Bercu18; @Collettietal17a; @Collettietal17b; @KubotaTakei19JSP]. Variations of elephant random walks studied mainly from mathematical viewpoint are found in [@BercuLaulin19; @Bertoin18; @Bertoin19; @Businger18; @GutStadtmuller18; @GutStadtmuller19].
Kumar, Harbola, and Lindenberg [@KumarHarbolaLindenberg10PRE] proposed a random walk model of elephant type, which exhibits asymptotic subdiffusion, normal diffusion, and superdiffusion as a single parameter is swept. An even simpler model of this kind was introduced by Harbola, Kumar, and Lindenberg [@HarbolaKumarLindenberg14PRE]. Assume that $p \in (0,1)$, $q \in [0,1)$ and $s \in [0,1]$. Define a sequence $\{X_i\}$ of $\{0,1\}$-valued random variables as follows:
- $P(X_1=1)=1-P(X_1=0)=s$.
- For each $n=1,2,\cdots$, let $U_n$ be uniformly distributed on $\{1,\cdots,n\}$.
- If $X_{U_n}=1$, then $X_{n+1} = \begin{cases}
1 &\mbox{with probability $p$}, \\
0 &\mbox{with probability $1-p$}. \\
\end{cases}$
- If $X_{U_n}=0$, then $X_{n+1} = \begin{cases}
1 &\mbox{with probability $q$}, \\
0 &\mbox{with probability $1-q$}. \\
\end{cases}$
A random walk $\{H_n\}$ on $\mathbb{Z}_+ := \{0,1,2,\cdots\}$ is defined by $$H_0:=0,\quad \mbox{and} \quad H_n:=\sum_{i=1}^n X_i \quad \mbox{for $n=1,2,\cdots$.}$$ Note that for $n=1,2,\cdots$, the conditional distribution of $X_{n+1}$ given the history up to time $n$ is $$\begin{aligned}
&P(X_{n+1} = 1 \mid X_1,\cdots,X_n) = 1- P(X_{n+1} = 0 \mid X_1,\cdots,X_n)\\
&= p \cdot \dfrac{\#\{i=1,\cdots,n : X_i=1\}}{n} + q \cdot \dfrac{\#\{i=1,\cdots,n : X_i=0\}}{n}. $$ Since $$\begin{aligned}
\#\{i=1,\cdots,n : X_i=1\} = H_n \quad \mbox{and} \quad \#\{i=1,\cdots,n : X_i=0\} = n-H_n,\end{aligned}$$ the conditional expectation of $X_{n+1}$ is $$\begin{aligned}
E[X_{n+1} \mid X_1,\cdots,X_n] = P(X_{n+1} = 1 \mid X_1,\cdots,X_n) &=\alpha \cdot \dfrac{H_n}{n} + q,
\label{eq:HarbolaKumarLindenberg14PREtrans}\end{aligned}$$ where $n=1,2,\cdots$ and $\alpha:=p-q \in (-1,1)$. Solving $$\begin{aligned}
E[H_{n+1}] = \left(1+\dfrac{\alpha}{n} \right) E[H_n] + q,\end{aligned}$$ we have $$\begin{aligned}
E[H_n] &= \dfrac{qn}{1-\alpha} + \left(s -\dfrac{q}{1-\alpha} \right)\cdot \dfrac{\Gamma(n+\alpha)}{\Gamma(1+\alpha) \Gamma(n)} \label{eq:CGL19(3)} \\
&\sim \begin{cases}
\dfrac{qn}{1-\alpha}&(q>0), \\[2mm]
\dfrac{s}{\Gamma(1+p)} \cdot n^p &(q=0),
\end{cases} \notag\end{aligned}$$ where $x_n \sim y_n$ means that $x_n/y_n$ converges to $1$ as $n \to \infty$. The walker is ballistic if $q>0$. On the other hand, in the case $q=0$, which we call the [*‘laziest’ minimal random walk model of elephant type*]{}, all three phases of asymptotic behavior are observed as $p \in (0,1)$ varies (see [@HarbolaKumarLindenberg14PRE]).
Coletti, Gava, and de Lima [@CollettiGavadeLima19JSM] proved several limit theorems for this model. The strong law of large numbers holds for any $q \in [0,1)$: $$\lim_{n \to \infty} \dfrac{H_n-E[H_n]}{n}= 0 \qquad \mbox{a.s..}$$ In particular, together with , $$\begin{aligned}
\lim_{n \to \infty} \dfrac{H_n}{n}= \dfrac{q}{1-\alpha}\qquad \mbox{a.s..}
\label{eq:Colettietal19sLLN}\end{aligned}$$ For $q>0$, further limit theorems are proved in [@CollettiGavadeLima19JSM]. Let $\rho := q/(1-\alpha) \in (0,1)$ and $\phi(t):=\sqrt{2t\log\log t}$.
- If $0 \leq \alpha <1/2$, then $$\begin{aligned}
\dfrac{H_n-E[H_n]}{\sqrt{\tfrac{\rho(1-\rho)}{1-2\alpha}n}} \stackrel{d}{\to} N\left(0,1\right),\mbox{ and }
\limsup_{n \to \infty} \pm \dfrac{H_n-E[H_n]}{\phi\left(\tfrac{\rho(1-\rho)}{1-2\alpha}n\right)}=1 \mbox{ a.s..}
$$
- If $\alpha=1/2$, then $$\begin{aligned}
\dfrac{H_n-E[H_n]}{\sqrt{\rho(1-\rho)n\log n}} \stackrel{d}{\to} N\left(0,1\right), \mbox{ and }
\limsup_{n \to \infty} \pm \dfrac{H_n-E[H_n]}{\phi\left(\rho(1-\rho)n\log n\right)}=1\mbox{ a.s..}
$$
- If $1/2<\alpha<1$, then there exists a random variable $W$ with positive variance such that $$\begin{aligned}
\displaystyle \lim_{n \to \infty} \dfrac{H_n-E[H_n]}{a_n} = W \quad \mbox{a.s. and in $L^2$,}\end{aligned}$$ where $a_n:=\dfrac{\Gamma(n+\alpha)}{\Gamma(1+\alpha) \Gamma(n)}$. Moreover, essentially the same calculation as in [@KubotaTakei19JSP] gives that $$\begin{aligned}
\dfrac{H_n-E[H_n]-W \cdot a_n}{\sqrt{\tfrac{\rho(1-\rho)}{2\alpha-1}n}} \stackrel{d}{\to} N\left(0,1\right),
\intertext{ and }
\limsup_{n \to \infty} \pm \dfrac{H_n-E[H_n]-W \cdot a_n}{\phi \left( \tfrac{\rho(1-\rho)}{2\alpha-1}n \right)}
=1 \mbox{ a.s..}\end{aligned}$$
If $s=\rho \in (0,1)$, then the minimal random walk model is equivalent to the “correlated Bernoulli process" introduced by Drezner and Farnum [@DreznerFarnum93]. In the latter context, the central limit theorem was proved by Heyde [@Heyde04].
Although the $q=0$ case is the most interesting, only partial results are obtained in Coletti, Gava, and de Lima [@CollettiGavadeLima19JSM]. The aim of this article is to prove several limit theorems for this case.
Results {#sec:Results}
=======
In this section we consider the ‘laziest’ minimal random walk model of elephant type ($q=0$). Note that if $X_1=0$, then $X_n=0$ for all $n$. Hereafter we assume that $s=1$ in addition. The dynamics can be summarized as follows: Let $p \in (0,1)$.
- The first step $X_1$ of the walker is $1$ with probability one.
- For each $n=1,2,\cdots$, let $U_n$ be uniformly distributed on $\{1,\cdots,n\}$, and $$\begin{aligned}
X_{n+1} &= \begin{cases}
X_{U_n} &\mbox{with probability $p$}, \\
0 &\mbox{with probability $1-p$}. \\
\end{cases}\end{aligned}$$
The equation becomes $$\begin{aligned}
E[X_{n+1} \mid {\mathscr}{F}_n]=P(X_{n+1}= 1 \mid {\mathscr}{F}_n)
= p \cdot \dfrac{H_n}{n}, \label{eq:elephantRWCondDistp}\end{aligned}$$ where ${\mathscr}{F}_n$ is the $\sigma$-algebra generated by $X_1,\cdots,X_n$. Noting that $$E[H_{n+1} \mid {\mathscr}{F}_n]=\left(1+\dfrac{p}{n}\right) H_n,$$ we introduce $$\begin{aligned}
\label{eq:ElephantRWa_nDef}
a_0:=1, \quad \mbox{and} \quad
a_n := \prod_{k=1}^{n-1} \left(1+\dfrac{p}{k}\right)= \dfrac{\Gamma(n+p)}{\Gamma(n)\Gamma(1+p)}\quad \mbox{for $n=1,2,\cdots$,}\end{aligned}$$ and set $$\begin{aligned}
\widehat{M_n}:=\dfrac{H_n}{a_n}\quad \mbox{for $n=0,1,2,\cdots$.}\end{aligned}$$ Then $\{ \widehat{M_n} \}$ satisfies a martingale property $E[\widehat{M_{n+1}} \mid {\mathscr}{F}_n]=\widehat{M_n}$. Since $ \widehat{M_n} $ is nonnegative, Doob’s convergence theorem implies that $$\begin{aligned}
\lim_{n \to \infty} M_n = \lim_{n \to \infty}\dfrac{H_n}{a_n} = \widehat{W} \quad \mbox{a.s..}
\label{eq:LaziestMartLim1}\end{aligned}$$ In Corollary \[cor:MiyazakiTakeiMoments\] below, we show that $\widehat{W}$ is positive with probability one.
Moments of the position
-----------------------
For $k=1,2,\cdots$, let $$a_n^{(k)} := \dfrac{\Gamma(n+kp)}{\Gamma(n)\Gamma(1+kp)}.$$ Note that $a_n^{(1)}=a_n$. The moments of the position $H_n$ up to the fourth are calculated in section 4.6 of Coletti, Gava, and de Lima [@CollettiGavadeLima19JSM]: $$\begin{aligned}
\begin{aligned}
E[H_n] &= a_n^{(1)}, \\
E[(H_n)^2] &= 2a_n^{(2)}-a_n^{(1)}, \\
E[(H_n)^3] &= 6a_n^{(3)}-6a_n^{(2)}+a_n^{(1)}, \\
E[(H_n)^4] &= 24a_n^{(4)}-36a_n^{(3)}+14a_n^{(2)}-a_n^{(1)}.
\end{aligned}
\label{eq:ColettideLimaGavaSec4.6}\end{aligned}$$ We could not find a simple way to describe the coefficients. Let $(x)_1:=x$ and $(x)_k:=x(x-1)\cdots(x-k+1)$ for $k=2,3,\cdots$. The $k$-th factorial moment of a random variable $X$ is defined by $E[(X)_k]$. Using , we can see that $$\begin{aligned}
E[(H_n)_2] &= E[(H_n)^2]-E[H_n] = 2( a_n^{(2)}-a_n^{(1)}), \\
E[(H_n)_3] &= E[(H_n)^3]-3E[(H_n)^2]+2E[H_n] = 6(a_n^{(3)}-2a_n^{(2)}+a_n^{(1)}),\\
E[(H_n)_4] &= E[(H_n)^4]-6E[(H_n)^3]+11E[(H_n)^2]-6E[H_n] \\
&= 24(a_n^{(4)}-3a_n^{(3)}+3a_n^{(2)}-a_n^{(1)}).\end{aligned}$$
The following theorem, which will be proved in section \[sec:Moments\], gives the general solution for $E[(H_n)_k]$.
\[thm:MiyazakiTakeiMoments\] Assume that $p \in (0,1)$, $q=0$, and $s=1$. For any $k=1,2,\cdots$ and $n=1,2,\cdots$, $$\begin{aligned}
E[(H_n)_k] = k! \cdot \sum_{i=1}^k (-1)^{k-i} \binom{k-1}{i-1} a_n^{(i)}. \label{eq:MiyazakiTakeiMoments}\end{aligned}$$
Since $a_n^{(k)} \sim \dfrac{n^{kp}}{\Gamma(1+kp)}$ as $n \to \infty$, we have $$\lim_{n \to \infty} \dfrac{E[(H_n)^k]}{a_n^{(k)}} = \lim_{n \to \infty} \dfrac{E[(H_n)_k]}{a_n^{(k)}} = k!
\quad \mbox{for any $k=1,2,\cdots$}.$$ The following corollary is a much more precise result than for $q=0$.
\[cor:MiyazakiTakeiMoments\] For any $k=1,2,\cdots$, $$\lim_{n \to \infty} E \left[\left(\dfrac{H_n}{n^p}\right)^k\right] = \dfrac{k!}{\Gamma(1+kp)}.$$ Thus the martingale $\{\widehat{M}_n\}$ is $L^k$-bounded for any $k$, and the almost sure limit $$\begin{aligned}
{\mathscr}{W}:= \lim_{n \to \infty}\dfrac{H_n}{n^p} = \dfrac{\widehat{W}}{\Gamma(1+p)} \label{eq:LaziestMartLim2}\end{aligned}$$ has a Mittag–Leffler distribution with parameter $p$ (see Appendix \[app:MLdistr\]). In particular, $P({\mathscr}{W}>0)=1$.
Limit theorems for the laziest case
-----------------------------------
By Corollary \[cor:MiyazakiTakeiMoments\], the limit $\widehat{W}$ in satisfies $P(\widehat{W}>0)=1$. Based on this fact, we obtain central limit theorems in the following form.
\[thm:MiyazakiTakeiCLT\] Assume that $q=0$ and $s=1$. For $0<p<1$, $$\begin{aligned}
\dfrac{H_n-\widehat{W} \cdot a_n}{\sqrt{\widehat{W} \cdot a_n}} \stackrel{d}{\to} N\left(0,1\right),
\intertext{and}
\dfrac{H_n-\widehat{W} \cdot a_n}{\sqrt{n^p}} \stackrel{d}{\to} \sqrt{{\mathscr}{W}'} \cdot Z,\end{aligned}$$ where $Z$ is distributed as $N(0,1)$, and $ {\mathscr}{W}'$ is independent of $Z$ and has the same distribution as $ {\mathscr}{W}$.
This situation is quite different from $q>0$: The central limit theorem holds for whole regions of parameter space, with random centering and random norming.
We also prove the law of the iterated logarithm.
\[thm:MiyazakiTakeiLIL\] Assume that $q=0$ and $s=1$. For $0<p<1$, $$\begin{aligned}
\limsup_{n \to \infty} \pm \dfrac{H_n-\widehat{W} \cdot a_n}{\phi \left( \widehat{W} \cdot a_n \right)}
=1 \mbox{ a.s.,}\end{aligned}$$ where $\phi(t):=\sqrt{2t\log \log t}$.
Theorems \[thm:MiyazakiTakeiCLT\] and \[thm:MiyazakiTakeiLIL\] will be proved in section \[sec:ProofLimitThm\].
Applications to percolation on random recursive trees {#sec:ApplPercRRT}
-----------------------------------------------------
Kürsten [@Kursten16] found important connections between (several variations of) elephant random walks and percolation on random recursive trees.
Consider the following procedure for obtaining a sequence $\{T_i\}$ of recursive trees: The first graph $T_1$ consists of a single vertex labeled 1. For each $i=2,3,\cdots$, the graph $T_i$ is evolved from $T_{i-1}$ by joining a new vertex labeled $i$ to a uniformly chosen vertex labeled $u_{i-1}$ from $T_{i-1}$. We perform Bernoulli bond percolation on $T_n$: Each edge of $T_n$ is independently removed with probability $1-p$, and otherwise retained. Then we can see that the size $\# {\mathscr}{C}_{1,n}$ of the cluster containing the vertex labeled 1 has the same distribution as the position $H_n$ of the laziest minimal random walk model of elephant type (this relation is implicitly mentioned before eq. (43) in [@Kursten16]).
Thus all of our results described above have counterparts for $\# {\mathscr}{C}_{1,n}$. To our best knowledge, Theorems \[thm:MiyazakiTakeiMoments\], \[thm:MiyazakiTakeiCLT\] and \[thm:MiyazakiTakeiLIL\] are new also in this context. We remark that the $\# {\mathscr}{C}_{1,n}$-version of Corollary \[cor:MiyazakiTakeiMoments\] is Lemma 3 in Businger [@Businger18], where it plays a crucial role in analyzing the shark random swim, and is proved using a connection with the Yule process. Our proof of Corollary \[cor:MiyazakiTakeiMoments\] based on Theorem \[thm:MiyazakiTakeiMoments\] is a short alternative.
In Appendix \[sec:minRWandPerc\], we give a precise description of the relation between the minimal random walk model with $0\leq q < p <1$ and percolation on random recursive trees. As an easy and useful application of this, we derive new expressions of the expectation and the variance of $H_n$, in terms of size of percolation clusters.
Factorial moments of the position {#sec:Moments}
=================================
To prove Theorem \[thm:MiyazakiTakeiMoments\], we use the probability generating functions: Let $$f_n(x) := E[x^{H_n}] \qquad \mbox{for $n=1,2,\cdots$.}$$ Recall that $f_n^{(k)}(1)=E[(H_n)_k]$, where $f_n^{(k)}(1)$ denotes the $k$-th derivative of $f_n(x)$ at $x=1$. By , we can see $$\begin{aligned}
E[x^{H_{n+1}} \mid {\mathscr}{F}_n] &= x^{H_n} \cdot \left\{ x^1 \cdot p \cdot \dfrac{H_n}{n} + x^0 \cdot \left(1-p \cdot \dfrac{H_n}{n}\right)\right\} \\
&= x^{H_n} \cdot \left\{ \dfrac{p(x-1)}{n} \cdot H_n +1\right \},\end{aligned}$$ and $$\begin{aligned}
f_{n+1}(x) &= E\left[ x^{H_n} \cdot \left\{ \dfrac{p(x-1)}{n} \cdot H_n +1 \right\} \right] \\
&= \dfrac{p(x-1)}{n} \cdot E[H_n x^{H_n} ] + E[x^{H_n}]
= \dfrac{px(x-1)}{n} \cdot f'_n(x) + f_n(x).\end{aligned}$$ The Leibniz rule yields $$\begin{aligned}
f_{n+1}^{(k)}(x) &= \dfrac{px(x-1)}{n} \cdot f_n^{(k+1)}(x) + \binom{k}{1} \cdot \dfrac{2px-p}{n} \cdot f_n^{(k)}(x)\\
&\quad + \binom{k}{2} \cdot \dfrac{2p}{n} \cdot f_n^{(k-1)}(x)+f_n^{(k)}(x).\end{aligned}$$ Thus we have $$\begin{aligned}
E[(H_{n+1})_k] = f_{n+1}^{(k)}(1)
&= \dfrac{kp}{n} \cdot f_n^{(k)}(1)+ \dfrac{k(k-1)p}{n} \cdot f_n^{(k-1)}(1)+f_n^{(k)}(1) \\
&= \dfrac{n+kp}{n} \cdot E[(H_n)_k] + \dfrac{k(k-1)p}{n} \cdot E[(H_n)_{k-1}].\end{aligned}$$
We prove by induction. For each $n=1,2,\cdots$, $E[H_n] = a_n^{(1)}$ in clearly satisfies . On the other hand, $$E[H_1]=1,\quad \mbox{and} \quad E[(H_1)_k] = 0 \quad \mbox{for $k=2,3,\cdots$}$$ satisfy , as $a_1^{(1)}=1$. Assume that $k>2$, and that $E[(H_n)_{k-1}]$ and $E[(H_n)_k]$ satisfy . We can see that $$\begin{aligned}
E[(H_{n+1})_k]&= \dfrac{n+kp}{n} \cdot k! \cdot \sum_{i=1}^k (-1)^{k-i} \binom{k-1}{i-1} a_n^{(i)} \\
&\quad + \dfrac{k(k-1)p}{n} \cdot (k-1)! \cdot \sum_{i=1}^{k-1} (-1)^{k-1-i} \binom{k-2}{i-1} a_n^{(i)} \\
&= \dfrac{n+kp}{n} \cdot k! \cdot a_n^{(k)} \\
&\quad + k! \cdot \sum_{i=1}^{k-1} (-1)^{k-i} \binom{k-1}{i-1} a_n^{(i)} \left(\dfrac{n+kp}{n}-\dfrac{(k-1)p}{n} \cdot \dfrac{k-i}{k-1} \right) \\
&= k! \cdot a_{n+1}^{(k)} + k! \cdot \sum_{i=1}^{k-1} (-1)^{k-i} \binom{k-1}{i-1} a_n^{(i)} \cdot \dfrac{n+ip}{n} \\
&= k! \cdot \sum_{i=1}^k (-1)^{k-i} \binom{k-1}{i-1} a_{n+1}^{(i)}.\end{aligned}$$ This completes the proof of Theorem \[thm:MiyazakiTakeiMoments\].
Limit theorems {#sec:ProofLimitThm}
==============
The structure of our proof of Theorems \[thm:MiyazakiTakeiCLT\] and \[thm:MiyazakiTakeiLIL\] is similar to that of [@KubotaTakei19JSP]. For $n=0,1,2,\cdots$, we set $$M_n:=\dfrac{H_n-E[H_n]}{a_n}(=\widehat{M_n}-1),$$ where $a_n$ is defined in . Clearly $\{M_n \}$ is a martingale with mean zero. As in the proof of Lemma 1 of [@KubotaTakei19JSP], we can obtain $$\begin{aligned}
M_{n+1}-M_n = \dfrac{X_{n+1}-E[X_{n+1} \mid {\mathscr}{F}_n]}{a_{n+1}} \label{eq:ElephantRWmartdiff1}\end{aligned}$$ for each $n=1,2,\cdots$. Noting that $X_n^2=X_n$, we have $$\begin{aligned}
E[(M_{n+1}-M_n)^2 \mid {\mathscr}{F}_n] &=\dfrac{E[(X_{n+1}-E[X_{n+1} \mid {\mathscr}{F}_n])^2 \mid {\mathscr}{F}_n]}{(a_{n+1})^2} \notag \\
&=\dfrac{E[X_{n+1}^2 \mid {\mathscr}{F}_n] -(E[X_{n+1} \mid {\mathscr}{F}_n])^2}{(a_{n+1})^2} \notag \\
&=\dfrac{E[X_{n+1} \mid {\mathscr}{F}_n] \cdot (1 -E[X_{n+1} \mid {\mathscr}{F}_n])}{(a_{n+1})^2}.\label{eq:ElephantRWmartdiff2}\end{aligned}$$ Note that and hold also for $n=0$, where ${\mathscr}{F}_0$ is the trivial $\sigma$-algebra.
For $k=1,2,\cdots$, let $$d_k:=M_k-M_{k-1}=\dfrac{X_k-E[X_k \mid {\mathscr}{F}_{k-1}]}{a_k}.$$ Note that $|d_k| \leq \dfrac{1}{a_k} \leq 1$. Using , , and , $$\begin{aligned}
E\left[ (d_k)^2 \mid {\mathscr}{F}_{k-1} \right] &= \dfrac{E[X_k \mid {\mathscr}{F}_{k-1}] \cdot (1 -E[X_k \mid {\mathscr}{F}_{k-1}])}{(a_k)^2} \\
&= p \cdot \dfrac{H_{k-1}}{k-1} \cdot \left(1-p \cdot \dfrac{H_{k-1}}{k-1}\right) \cdot \dfrac{1}{(a_k)^2} \\
&= p \cdot \dfrac{H_{k-1}}{a_{k-1}} \cdot \dfrac{a_{k-1}}{k-1} \cdot \left(1-p \cdot \dfrac{H_{k-1}}{k-1}\right) \cdot \dfrac{1}{(a_k)^2} \\
&\sim \dfrac{p \cdot \widehat{W}}{ka_k} \qquad \mbox{as $k \to \infty$ a.s..}\end{aligned}$$ As $|d_k| \leq 1$, the bounded convergence theorem yields that $$\begin{aligned}
E[ (d_k)^2 ] \sim \dfrac{p \cdot E[\widehat{W}]}{ka_k} =\dfrac{p}{ka_k} \qquad \mbox{as $k \to \infty$.}\end{aligned}$$ Thus we have $$\begin{aligned}
V_n^2&:= \sum_{k=n}^{\infty} E\left[ (d_k)^2 \mid {\mathscr}{F}_{k-1} \right] \\
&\sim p \cdot \widehat{W} \cdot \Gamma(1+p) \sum_{k=n}^{\infty} \dfrac{1}{k^{1+p}} \\
&\sim p \cdot \widehat{W} \cdot \Gamma(1+p) \cdot \dfrac{1}{pn^p}
=\frac{\widehat{W} \cdot \Gamma(1+p)}{n^p} \sim \dfrac{\widehat{W}}{a_n}\qquad \mbox{as $n \to \infty$ a.s.,} \end{aligned}$$ and $$\begin{aligned}
s_n^2&:=\sum_{k=n}^{\infty} E[(d_k)^2] \sim \dfrac{1}{a_n}\qquad \mbox{as $n \to \infty$,} \end{aligned}$$ which imply that $$\begin{aligned}
\lim_{n \to \infty} \dfrac{V_n^2}{s_n^2}=\widehat{W} \quad \mbox{a.s..} \label{eq:KubotaTakeiSupercriticalCLTLILV_nLLN}\end{aligned}$$
We check the conditions of Theorem \[thm:Heyde77Theorem1b\] (ii) and (iii) in Appendix \[sec:MartLimitThm\] are satisfied.
To prove conditions a) and a’) with $\eta^2=\widehat{W}$ are satisfied, we will show $$\lim_{n \to \infty} \dfrac{1}{s_n^2}\sum_{k=n}^{\infty} \{ (d_k)^2 - E[(d_k)^2 \mid {\mathscr}{F}_{k-1}]\} = 0 \quad \mbox{a.s..}$$ By the tail version of Kronecker’s lemma (see Lemma 1 (ii) in Heyde [@Heyde77]), it is sufficient to show $$\begin{aligned}
\sum_{k=1}^{\infty} \dfrac{1}{s_k^2} \{ (d_k)^2 - E[(d_k)^2 \mid {\mathscr}{F}_{k-1}]\}<+\infty \quad \mbox{a.s..} \label{eq:WnVnKroneckerTail}\end{aligned}$$ Let $\widetilde{d}_k$ denote the summand. Note that $$\sum_{k=1}^{\infty} E[ (\widetilde{d}_k)^2 \mid {\mathscr}{F}_{k-1}] \leq \sum_{k=1}^{\infty} \dfrac{1}{s_k^4} E[(d_k)^4 \mid {\mathscr}{F}_{k-1}].$$ Since $$\begin{aligned}
E[(d_k)^4 \mid {\mathscr}{F}_{k-1}]
&\leq \dfrac{1}{(a_k)^2} \cdot E[(d_k)^2 \mid {\mathscr}{F}_{k-1}] \\
&\sim s_k^4 \cdot \dfrac{p \cdot \widehat{W}}{ka_k} \sim s_k^4 \cdot \dfrac{p \cdot \widehat{W} \cdot \Gamma(1+p)}{k^{1+p}}\quad \mbox{as $k \to \infty$,}
$$ the series in the right hand side converges a.s.. Theorem \[thm:HallHeyde80Thm215\] implies . For $\varepsilon>0$, noting that $$E[(d_k)^2 : |d_k| > \varepsilon s_n] \leq \dfrac{1}{\varepsilon^2 s_n^2} E[(d_k)^4],$$ and $$E[(d_k)^4] \leq \dfrac{1}{(a_k)^2} \cdot E[(d_k)^2] \sim \dfrac{p \cdot \Gamma(1+p)^3}{k^{1+3p}}\quad \mbox{as $k \to \infty$,}$$ we have $$\begin{aligned}
\dfrac{1}{s_n^2} \sum_{k=n}^{\infty} E[(d_k)^2 : |d_k| > \varepsilon s_n] &\leq \dfrac{1}{\varepsilon^2 s_n^4} \sum_{k=n}^{\infty} E[(d_k)^4] \\
&\leq C_1 n^{2p} \cdot \dfrac{1}{n^{3p}} = \dfrac{C_1}{n^p} \to 0\quad \mbox{as $n \to \infty$}.\end{aligned}$$ In view of Remark \[rem:Heyde77Theorem1b\], condition b) is also satisfied.
Similarly, for $\varepsilon>0$ we have $$\begin{aligned}
\dfrac{1}{s_k}E[|d_k| : |d_k| > \varepsilon s_k] &\leq \dfrac{1}{s_k} \cdot \dfrac{1}{\varepsilon^3 s_k^3} E[(d_k)^4] \\
&\leq C_2 k^{2p} \cdot \dfrac{1}{k^{1+3p}} = \dfrac{C_2}{k^{1+p}},\end{aligned}$$ which implies that condition c) holds.
Condition d) is implied by $$\displaystyle \sum_{n=1}^{\infty} \dfrac{1}{s_n^4} E[(d_n)^4]<+\infty.$$
To deduce the conclusions of Theorems \[thm:MiyazakiTakeiCLT\] and \[thm:MiyazakiTakeiLIL\] from Theorem \[thm:Heyde77Theorem1b\] (ii) and (iii), note that $$\begin{aligned}
W-M_n = \dfrac{a_n \cdot W - (H_n-E[H_n])}{a_n} = \dfrac{a_n \cdot \widehat{W} - H_n}{a_n},
\intertext{and}
a_n \cdot \widehat{\phi}\left(\dfrac{\widehat{W}}{a_n} \right) \sim \widehat{\phi}\left( a_n \cdot \widehat{W} \right)\quad \mbox{as $n \to \infty$},\end{aligned}$$ where $\widehat{\phi}(t)=\sqrt{2t\log |\log t|}$.
The minimal random walk model and percolation on random recursive trees {#sec:minRWandPerc}
=======================================================================
We explore a relation between the minimal random walk model by Harbola, Kumar, and Lindenberg [@HarbolaKumarLindenberg14PRE], explained in the Introduction, and percolation on random recursive trees. Throughout this section we assume that $0 \leq q < p < 1$, and set $\alpha=p-q$ and $\rho = q/(1-\alpha)$.
The sequence $\{T_i\}$ of random recursive trees is defined in section \[sec:ApplPercRRT\]. Consider bond percolation on $T_n$ with parameter $\alpha$. The expectation regarding this model is denoted by $E_{\alpha} [ \,\cdot\,]$. There are at most $n$ clusters, which are denoted by ${\mathscr}{C}_{1,n},\,{\mathscr}{C}_{2,n},\,\cdots,\,{\mathscr}{C}_{n,n}$ (for convenience we regard ${\mathscr}{C}_{j,n}=\emptyset$ if $j$ is larger than the number of clusters). We quote some of results in Kürsten [@Kursten16].
\[lem:Kursten16summary\] For bond percolation on $T_n$ with parameter $\alpha \in (0,1)$, $$\begin{aligned}
E_{\alpha} [\# {\mathscr}{C}_{1,n} ] &= \dfrac{\Gamma(n+\alpha)}{\Gamma(1+\alpha)\Gamma(n)}, \label{eq:Kuersten16PRE(13)} \\
E_{\alpha} [(\# {\mathscr}{C}_{1,n})^2 ] &= \dfrac{2\Gamma(n+2\alpha)}{\Gamma(1+2\alpha)\Gamma(n)} - \dfrac{\Gamma(n+\alpha)}{\Gamma(1+\alpha)\Gamma(n)}, \label{eq:Kuersten16PRE(44)} \\
\sum_{j=1}^n E_{\alpha} [(\# {\mathscr}{C}_{j,n})^2 ] &= \begin{cases}
\dfrac{1}{1-2\alpha} \cdot n+ \dfrac{1}{2\alpha-1} \cdot \dfrac{\Gamma(n+2\alpha)}{\Gamma(2\alpha)\Gamma(n)} &(\alpha \neq 1/2), \\[3mm]
\displaystyle n\sum_{\ell=1}^n \dfrac{1}{\ell}&(\alpha=1/2).
\end{cases} \label{eq:Kuersten16PRE(17)}\end{aligned}$$
In [@Kursten16], and are derived from basic results on the original elephant random walk, found in [@SchutzTrimper04]. In view of the connection with the laziest minimal random walk model explained in section \[sec:ApplPercRRT\], and are paraphrases of . Those are obtained by solving relatively easy difference equations.
Let $\xi_1,\xi_2,\cdots,\xi_n$ be a sequence of independent random variables, which is also independent from bond percolation, satisfying $$\begin{aligned}
&P_{p,q,s} (\xi_1=1)=1-P_{p,q,s}(\xi_1=0)=s,\quad \mbox{and} \\
&P_{p,q,s}(\xi_j=1)=1-P_{p,q,s}(\xi_j=0)=\rho \quad \mbox{for $j>1$.}\end{aligned}$$ The transition probability in can be interpreted as follows: For each step $n=2,3,\cdots$,
- with probability $\alpha$, the walker repeats the behavior at a uniformly chosen time, and
- with probability $1-\alpha$, the walker moves forward with probability $\rho$, or remains at rest otherwise.
Similarly to [@Kursten16], we can see that $H_n$ has the same distribution as $$\begin{aligned}
\sum_{j=1}^n \xi_j \cdot (\# {\mathscr}{C}_{j,n}). \label{eq:Kuersten16(11)bis}\end{aligned}$$
In the case $q>0$, computation of the second moment (and the variance) of $H_n$ by solving difference equations is straightforward but quite tedious, as is imagined from very complicated equations (8), (9) and (10) in [@HarbolaKumarLindenberg14PRE]. Using the above connection with percolation, we can easily obtain concise formulae described in terms of the moments of the size of open clusters.
\[thm:MiyazakiTakei20AppB\] Let $E_{p,q,s}$ and $V_{p,q,s}$ denote the expectation and the variance for the minimal random walk model. Assume that $0 \leq q < p < 1$, and set $\alpha=p-q$ and $\rho = q/(1-\alpha)$. Then we have the following. $$\begin{aligned}
E_{p,q,s} [H_n] &= \rho n + (s-\rho) E_{\alpha} [\# {\mathscr}{C}_{1,n} ], \label{eq:MiyazakiTakei20AppB-E} \\
V_{p,q,s} [H_n]
&= \rho (1-\rho) \sum_{j=1}^n E_{\alpha}[(\# {\mathscr}{C}_{j,n})^2] \notag \\
&\quad + (1-2\rho)(s-\rho) E_{\alpha} [(\# {\mathscr}{C}_{1,n})^2 ] - (s-\rho)^2 (E_{\alpha} [\# {\mathscr}{C}_{1,n}])^2. \label{eq:MiyazakiTakei20AppB-V}\end{aligned}$$
Since $E_{p,q,s} [\xi_1]=s$ and $E_{p,q,s} [\xi_j]=\rho$ for $j>1$, we have $$\begin{aligned}
E_{p,q,s} [H_n] &= \sum_{j=1}^n E_{p,q,s} [\xi_j] \cdot E_{\alpha} [\# {\mathscr}{C}_{j,n}]
= s E_{\alpha}[\# {\mathscr}{C}_{1,n}]+\rho \sum_{j=2}^n E_{\alpha}[\# {\mathscr}{C}_{j,n}] \\
&= \rho \sum_{j=1}^n E_{\alpha}[\# {\mathscr}{C}_{j,n}] + (s-\rho) E_{\alpha} [\# {\mathscr}{C}_{1,n}]
= \rho n + (s-\rho) E_{\alpha}[\# {\mathscr}{C}_{1,n}].\end{aligned}$$ Turning to the mean square displacement, similarly as eq. (17) in [@Kursten16], $$\begin{aligned}
&E_{p,q,s}[(H_n)^2]\notag \\
&= \sum_{j=1}^n E_{p,q,s} [(\xi_j)^2] \cdot E_{\alpha} [(\# {\mathscr}{C}_{j,n})^2] \notag\\
&\quad + 2\sum_{1\leq j<k \leq n} E_{p,q,s}[\xi_j] \cdot E_{p,q,s}[\xi_k] \cdot E_{\alpha} [(\# {\mathscr}{C}_{j,n}) \cdot (\# {\mathscr}{C}_{k,n})] \notag \\
&= \rho \sum_{j=1}^n E_{\alpha}[(\# {\mathscr}{C}_{j,n})^2] + 2\rho^2 \sum_{1 \leq j<k \leq n} E_{\alpha}[(\# {\mathscr}{C}_{j,n}) \cdot (\# {\mathscr}{C}_{k,n})] \notag \\
&\quad +(s-\rho) E_{\alpha}[(\# {\mathscr}{C}_{1,n})^2]+ 2(s - \rho)\rho \sum_{1<k \leq n} E_{\alpha}[(\# {\mathscr}{C}_{1,n}) \cdot (\# {\mathscr}{C}_{k,n})]. \label{eq:MiyazakiTakei20AppB-Vtemp}\end{aligned}$$ The first two terms in are $$\begin{aligned}
&\rho^2 E_{\alpha}\left[ \left(\sum_{j=1}^n \# {\mathscr}{C}_{j,n}\right)^2 \right]+(\rho-\rho^2) \sum_{j=1}^n E_{\alpha}[(\# {\mathscr}{C}_{j,n})^2] \\
&=\rho^2n^2 + \rho(1-\rho) \sum_{j=1}^n E_{\alpha}[(\# {\mathscr}{C}_{j,n})^2].\end{aligned}$$ Noting that $$\begin{aligned}
\sum_{1<k \leq n} E_{\alpha}[(\# {\mathscr}{C}_{1,n}) \cdot (\# {\mathscr}{C}_{k,n})]
&= E_{\alpha}\left[(\# {\mathscr}{C}_{1,n})\cdot \sum_{1<k \leq n} (\# {\mathscr}{C}_{k,n}) \right] \\
&= E_{\alpha}[(\# {\mathscr}{C}_{1,n})\cdot (n-\# {\mathscr}{C}_{1,n} )],\end{aligned}$$ the other two terms in are $$\begin{aligned}
&(s-\rho) E_{\alpha}[(\# {\mathscr}{C}_{1,n})^2]+ 2(s-\rho)\rho E_{\alpha}[(\# {\mathscr}{C}_{1,n})\cdot (n-\# {\mathscr}{C}_{1,n} )] \\
&=(1-2\rho)(s-\rho) E_{\alpha}[(\# {\mathscr}{C}_{1,n})^2]+ 2(s-\rho)\rho n E_{\alpha}[\# {\mathscr}{C}_{1,n}] .\end{aligned}$$ Using $$\begin{aligned}
(E_{p,q,s}[H_n])^2 &= \rho^2 n^2 +2(s-\rho) \rho n E_{\alpha} [\# {\mathscr}{C}_{1,n} ]+ (s-\rho)^2 (E_{\alpha} [\# {\mathscr}{C}_{1,n} ])^2,\end{aligned}$$ we have the conclusion.
Combining with Lemma \[lem:Kursten16summary\], we can obtain the asymptotics of the variance.
\[cor:MiyazakiTakei20AppB\] When $q>0$, $$\begin{aligned}
V_{p,q,s} [H_n]
&\sim \begin{cases}
\dfrac{\rho (1-\rho)}{1-2\alpha}n &(\alpha<1/2), \\
\rho(1-\rho) n \log n &(\alpha=1/2), \\
\left[ \dfrac{\rho(1-\rho)}{(2\alpha-1)\Gamma(2\alpha)} + \dfrac{(1-2\rho)(s-\rho)}{\Gamma(1+2\alpha)} - \dfrac{(s-\rho)^2}{\Gamma(1+\alpha)^2}\right] n^{2\alpha} &(\alpha>1/2) \\
\end{cases}\end{aligned}$$ as $n \to \infty$.
To close this section, we give a remark on phase transition of the [*biased elephant random walk*]{} $\{S_n\}$ on $\mathbb{Z}$:
- With probability $\alpha$, the walker repeats one of previous steps.
- With probability $1-\alpha$, the walker performs like a simple random walk, which jumps to the right with probability $\rho$, or to the left with probability $1-\rho$ (The unbiased case $\rho=1/2$ is the original elephant random walk explained in the Introduction, where $p \geq 1/2$ and $\alpha=2p-1$.)
This is obtained from the minimal random walk model as follows: Let $$Y_i :=2X_i-1 \quad \mbox{and} \quad S_n=\sum_{i=1}^n Y_i = 2H_n-n.$$ Then $P(Y_1=+1)=1- P(Y_1=-1)=s$, and by , $$P(Y_{n+1}=\pm 1 \mid {\mathscr}{F}_n) = \alpha \cdot \dfrac{\#\{i=1,\cdots,n : Y_i=\pm 1\}}{n} + (1-\alpha) \cdot \rho.$$ By , we have $$\begin{aligned}
\lim_{n \to \infty} \dfrac{S_n}{n}= 2\rho-1\qquad \mbox{a.s..} \label{eq:Colettietal19sLLNBiasedRW}\end{aligned}$$
Consider bond percolation on $T_n$ with parameter $\alpha$, and assign ‘spin’ $m_j:=2\xi_j-1\in \{+1,-1\}$ to each of percolation clusters ${\mathscr}{C}_{j,n}$, independently for different clusters. By , $S_n$ has the same distribution as $$\begin{aligned}
\sum_{j=1}^n m_j \cdot (\# {\mathscr}{C}_{j,n}).\end{aligned}$$ The above procedure is essentially the same as the [*“Divide and Color" model*]{} introduced by Häggström [@Haggstrom01SPA]. When $s=\rho=1/2$, the resulting model resembles the Ising model with zero external field, and increasing $\alpha$ corresponds to lowering the temperature. The parameter $\varepsilon:=2\rho-1$ plays a similar role to the external field in the Ising model. By , when $\varepsilon \neq 0$, the asymptotic speed of the walker remains unchanged regardless of the value of $\alpha$. On the other hand, when $\varepsilon = 0$, the walker admits a phase transition from diffusive to superdiffusive behavior. This is reminiscent of the fact that the Ising model admits a phase transition only when the external field is zero.
Martingale limit theorems {#sec:MartLimitThm}
=========================
Lemmas from calculus
--------------------
\[lem:KroneckerHeyde\] Consider a positive real sequence $\{a_n\}$ which monotonically diverges to $+\infty$, and another real sequence $\{b_n\}$.
- (Kronecker’s lemma) If $\displaystyle \sum_{k=1}^{\infty} \dfrac{b_k}{a_k}$ converges, then $\displaystyle \lim_{n \to \infty} \dfrac{1}{a_n}\sum_{k=1}^n b_k =0$.
- (Heyde [@Heyde77], Lemma 1 (ii)) If $\displaystyle \sum_{k=1}^{\infty} a_k b_k$ converges, then$\displaystyle \lim_{n \to \infty} a_n\sum_{k=n}^{\infty} b_k =0$.
\[thm:HallHeyde80Thm215\] Suppose that $\{M_n\}$ is a square-integrable martingale with mean $0$. Let $d_k=M_k-M_{k-1}$ for $k=1,2,\cdots$, where $M_0=0$. On the event $$\left\{ \sum_{k=1}^{\infty} E[(d_k)^2 \mid {\mathscr}{F}_{k-1}] <+\infty \right\},$$ $\{M_n\}$ converges a.s..
\[thm:Heyde77Theorem1b\] Suppose that $\{M_n\}$ is a square-integrable martingale with mean $0$. Let $d_k=M_k-M_{k-1}$ for $k=1,2,\cdots$, where $M_0=0$. If $$\displaystyle \sum_{k=1}^{\infty} E[(d_k)^2] < +\infty$$ holds in addition, then we have the following: Let $$\displaystyle W_n^2 :=\sum_{k=n}^{\infty} (d_k)^2
\quad \mbox{and} \quad
s_n^2 := \sum_{k=n}^{\infty} E[ (d_k)^2].$$
- The limit $M_{\infty}:=\sum_{k=1}^{\infty} d_k$ exists a.s., and $M_n \stackrel{L^2}{\to} M_{\infty}$.
- Assume that
- $\displaystyle \dfrac{W_n^2}{s_n^2} \to \eta^2$ as $n \to \infty$ in probability, and
- $\displaystyle \lim_{n \to \infty} \dfrac{1}{s_n^2}E\left[ \sup_{k \geq n} (d_k)^2\right]=0$,
where $\eta^2$ is some a.s. finite and non-zero random variable. Then we have $$\begin{aligned}
\dfrac{M_{\infty} - M_n}{W_{n+1}} &= \dfrac{ \sum_{k=n+1}^{\infty} d_k}{W_{n+1}}\stackrel{d}{\to} Z,\quad {and} \\
\dfrac{M_{\infty} - M_n}{s_{n+1}} &= \dfrac{ \sum_{k=n+1}^{\infty} d_k}{s_{n+1}}\stackrel{d}{\to} \widehat{\eta} \cdot Z, $$ where $Z$ is distributed as $N(0,1)$, and $\widehat{\eta}$ is independent of $Z$ and distributed as $\eta$.
- Assume that the following three conditions hold:
- $\displaystyle \dfrac{W_n^2}{s_n^2} \to \eta^2$ as $n \to \infty$ a.s.,
- $\displaystyle \sum_{k=1}^{\infty} \dfrac{1}{s_k} E[ |d_k| : |d_k| > \varepsilon s_k] < +\infty$ for any $\varepsilon > 0$, and
- $\displaystyle \sum_{k=1}^{\infty} \dfrac{1}{s_k^4} E[ (d_k)^4 : |d_k| \leq \delta s_k] < +\infty$ for some $\delta > 0$.
Then $\displaystyle \limsup_{n \to \infty} \pm \dfrac{M_{\infty} - M_n}{\widehat{\phi}(W_{n+1}^2)} =1$ a.s., where $\widehat{\phi}(t):=\sqrt{2t\log |\log t|}$.
\[rem:Heyde77Theorem1b\] A sufficient condition for b) in Theorem \[thm:Heyde77Theorem1b\] is that $$\displaystyle \lim_{n \to \infty}\dfrac{1}{s_n^2} \sum_{k=n}^{\infty} E[ (d_k)^2 : |d_k| > \varepsilon s_n] =0$$ for any $\varepsilon > 0$. (See the proof of Corollary 1 in Heyde [@Heyde77].)
The Mittag–Leffler disribution {#app:MLdistr}
==============================
The [*Mittag–Leffler function*]{} is defined by $$E_{\alpha}(z) := \sum_{k=0}^{\infty} \dfrac{z^k}{\Gamma(k + \alpha)}\qquad (\alpha,z \in \mathbb{C}).$$ Note that $E_1(z)=e^z$ (See e.g. [@BinghamGoldieTeugels89], p. 315).
The random variable $X$ is [*Mittag–Leffler distributed*]{} with parameter $p \in [0,1]$ if $$E[e^{\lambda X}] = E_p (\lambda) = \sum_{k=0}^{\infty} \dfrac{\lambda^k}{\Gamma(1+ kp)}\qquad \mbox{for $\lambda \in \mathbb{R}$}.$$ Thus the $k$-th moment of $X$ is $ \dfrac{k!}{\Gamma(1+ kp)}$, and this distribution is determined by moments (see [@BinghamGoldieTeugels89], p. 329 and p. 391). If $p=0$ (resp. $p=1$), then $X$ has the exponential distribution with mean one (resp. $X$ concentrates on $\{1\}$). For $p \in (0,1)$, the probability density function $f_p(x)$ of $X$ is $$f_p(x)= \dfrac{\rho_p(x^{-1/p})}{px^{1+1/p}}\qquad \mbox{for $x>0$},$$ where $\rho_p(x)$ is the density function of the one-sided $\mbox{stable$(p)$}$ distribution. See [@Pollard48BAMS] for details. In particular, $f_{1/2}$ is the density function of the standard half-normal distribution: $$f_{1/2} (x) = \sqrt{\dfrac{2}{\pi}} e^{-x^2/2}\qquad \mbox{for $x>0$}.$$
[99]{} Baur, E. and Bertoin, J. (2016). Elephant random walks and their connection to Pólya-type urns, [*Phys. Rev. E*]{}, [**94**]{}, 052134 Bercu, B. (2018). A martingale approach for the elephant random walk, [*J. Phys. A: Math. Theor.*]{}, [**51**]{}, 015201 Bercu, B. and Laulin, L. (2019). On the multi-dimensional elephant random walk, [*J. Statist. Phys.*]{}, [**175**]{}, 1146–1163. Bertoin, J. (2018). Noise reinforcement for Lévy processes, to appear in [*Ann. Inst. Henri Poincaré Probab. Stat.*]{}, arXiv:1810.08364 Bertoin, J. (2020). Universality of noise reinforced Brownian motions, arXiv:2002.09166 Bingham, N. H., Goldie, C. M., and Teugels, J. L. (1989). Regular variation, [*Encyclopedia of Mathematics and its Applications*]{}, [**27**]{}, Cambridge University Press. Businger, S. (2018). The shark random swim (Lévy flight with memory), [*J. Statist. Phys.*]{}, [**172**]{}, 701–717. (See also arXiv:1710.05671v3) Coletti, C. F., Gava, R. J., and de Lima, L. R. (2019). Limit theorems for a minimal random walk model, [*J. Stat. Mech.*]{}, 083206 Coletti, C. F., Gava, R. J., and Schütz, G. M. (2017a). Central limit theorem for the elephant random walk, [*J. Math. Phys.*]{}, [**58**]{}, 053303 Coletti, C. F., Gava, R. J., and Schütz, G. M. (2017b). A strong invariance principle for the elephant random walk, [*J. Stat. Mech.*]{}, 123207 Drezner, Z. and Farnum, N. (1993). A generalized binomial distribution, [*Comm. Statist. Theory Methods*]{}, [**22**]{}, 3051–3063. Gut, A. and Stadtmüller, U. (2018). Variations of the elephant random walk, arXiv:1812.01915 Gut, A. and Stadtmüller, U. (2019). Elephant random walks with delays, arXiv:1906.04930 Häggström, O. (2001). Coloring percolation clusters at random, [*Stoch. Proc. Appl.*]{}, [**96**]{}, 213–242. Hall, P. and Heyde, C. C. (1980). Martingale limit theory and its application, [*Probability and Mathematical Statistics*]{}, Academic Press. Harbola, U., Kumar, N., and Lindenberg, K. (2014). Memory-induced anomalous dynamics in a minimal random walk model, [*Phys. Rev. E*]{}, [**90**]{}, 022136 Heyde, C. C. (1977). On central limit and iterated logarithm supplements to the martingale convergence theorem, [*J. Appl. Probab.*]{}, [**14**]{}, 758–775. Heyde, C. C. (2004). Asymptotics and criticality for a correlated Bernoulli process, [*Aust. N. Z. J. Stat.*]{}, [**46**]{}, 53–57. Kubota, N. and Takei, M. (2019). Gaussian fluctuation for superdiffusive elephant random walks, [*J. Statist. Phys.*]{}, [**177**]{}, 1157–1171. Kumar, N., Harbola, U., and Lindenberg, K. (2010). Memory-induced anomalous dynamics: Emergence of diffusion, [*Phys. Rev. E*]{}, [**82**]{}, 021101 Kürsten, R. (2016). Random recursive trees and the elephant random walk, [*Phys. Rev. E*]{}, [**93**]{}, 032111 Pollard, H. (1948). The completely monotonic character of the Mittag–Leffler function $E_a(-x)$, [*Bull. Amer. Math. Soc.*]{}, [**54**]{}, 1115–1116. Schütz, G. M. and Trimper, S. (2004). Elephants can always remember: Exact long-range memory effects in a non-Markovian random walk, [*Phys. Rev. E*]{}, [**70**]{}, 045101
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Xiangwen Wang
- Michel Pleimling
title: 'Online Gambling of Pure Chance: Wager Distribution, Risk Attitude, and Anomalous Diffusion'
---
Introduction {#introduction .unnumbered}
============
Today, gambling is a huge industry with a huge social impact. According to a report by the American Gaming Association [@AGA2018], commercial casinos in the United States alone made total revenue of over 40 billion US dollars in 2017. On the other hand, different studies reported that $0.12\%-5.8\%$ of the adults and $0.2\%-12.3\%$ of the adolescents across different countries in the world are experiencing problematic gambling [@Calado2016; @Calado2017]. Studying the gamblers’ behavior patterns not only contributes to the prevention of problematic gambling and adolescent gambling, but also helps to better understand human decision-making processes. Researchers have put a lot of attention on studying gambling-related activities. Economists have proposed many theories about how humans make decisions under different risk conditions. Several of them can also be applied to model gambling behaviors. For example, the prospect theory introduced by Kahneman and Tversky [@Kahneman1979] and its variant cumulative prospect theory [@Tversky1992] have been adopted in modeling casino gambling [@Barberis2012]. In parallel to the theoretical approach, numerous studies focus on the empirical analysis of gambling behaviors, aiming at explaining the motivations behind problematic gambling behaviors. However, parametric models that quantitatively describe empirical gambling behaviors are still missing. Such models can contribute to evaluating gambling theories proposed by economists, as well as yield a better understanding of the gamblers’ behaviors. Our goal is to provide such a parametric model for describing human wagering activities and risk attitude during gambling from empirical gambling logs. However, it is very difficult to obtain gambling logs from traditional casinos, and it is hard to collect large amounts of behavior data in a lab-controlled environment. Therefore in this paper we will focus on analyzing online gambling logs collected from online casinos.
Whereas historically the development of probability theory, which then became the foundation of statistics, was tied to chance games, today we use statistical tools to analyze gamblers’ behaviors.
Recent years have seen an increasing trend of online gambling due to its low barriers to entry, high anonymity and instant payout. For researchers of gambling behaviors, online gambling games present two advantages: simple rules and the availability of large amounts of gambling logs. In addition to the usual forms of gambling games that can be found in traditional casinos, many online casinos also offer games that follow very simple rules, which makes analyzing the gambling behavior much easier as there are much fewer degrees of freedom required to be considered. On the other hand, many online casinos have made gambling logs publicly available on their websites, mainly for verification purposes, which provides researchers with abundant data to work on. Due to the high popularity of online gambling, in a dataset provided by an online casino there are often thousands or even hundreds of thousands of gamblers listed. Such a large scale of data can hardly be obtained in a lab environment. Prior research has begun to make use of online gambling logs. For example, Meng’s thesis [@Meng2018] presented a pattern analysis of typical gamblers in Bitcoin gambling. It is worth arguing that although our work only focuses on the behaviors of online gamblers, there is no reason to think that our conclusions cannot be extended to traditional gamblers.
Naturally, we can treat the changing cumulative net income of a player during their gambling activities as a random walk process [@Wang2018]. We are particularly interested in the diffusive characteristics of the gambler’s net income. This is another reason why we want to analyze the wager distribution and risk attitude of gamblers, since both distributions are closely related to the displacement distribution for the gambler’s random walks. Within this paper, we will mainly focus on the analysis at the population level. Physicists have long been studying diffusion processes in different systems, and recently anomalous diffusive properties have been reported in many human activities, including human spatial movement [@Rhee2011; @Brockmann2008b; @Kim2010], and information foraging [@Wang2017]. In a previous study of skin gambling [@Wang2018], we have shown that in a parimutuel betting game (where players gamble against each other), a gambler’s net income displays a crossover from superdiffusion to normal diffusion. We have reproduced this crossover in simulations by introducing finite and overall conserved gamblers’ wealth (see [@Toscani2019] for a different way of modeling this using kinetic equations of Boltzmann and Fokker-Planck type). However, this explanation cannot be used in other types of gambling games where there is no interaction among gamblers (e.g., fixed-odds betting games, which will be introduced below), as they violate the conservation of gamblers’ overall wealth. In this paper, we want to expand the scope of our study to more general gambling games, check the corresponding diffusive properties, and propose some explanations for the observed behaviors.
One of our goals is to uncover the commonalities behind the behavior of online gamblers. To implement this, we analyze the data from different online gambling systems. The first one is skin gambling, where the bettors are mostly video game players and where cosmetic skins from online video games are used as virtual currency for wagering [@Wang2018; @Holden2018]. The other system is crypto-currency gambling, where the bettors are mostly crypto-currency users. Different types of crypto-currencies are used for wagering. Commonly used crypto-currencies include Bitcoin, Ethereum, and Bitcoin Cash, whose basic units are BTC, ETH and BCH, respectively. As the overlap of these two communities, video game players and crypto-currency users, is relatively small for now, features of gambling patterns common between these two gambling systems are possibly features common among all online gamblers.
Not only do we consider different gambling systems, but we also discuss different types of gambling games. In this paper, we discuss four types of solely probability-based gambling games (Roulette, Crash, Satoshi Dice and Jackpot), whose outcomes in theory will not benefit from the gamblers’ skill or experience when the in-game random number generators are well designed. In general, there are two frameworks of betting in gambling: fixed-odds betting, where the odds is fixed and known before players wager in one round; and parimutuel betting, where the odds can still change after players place the bets until all players finish wagering. In fixed-odds betting, usually players bet against the house/website, and there is no direct interaction among players; and in parimutuel betting, usually players bet against each other. The four types of games we discuss in this paper will cover both betting frameworks (see the Methods section).
When a player attends one round in any of those games, there are only two possible outcomes: either win or lose. When losing, the player will lose the wager they placed during that round; whereas when winning, the prize winner receives equals their original wager multiplied by a coefficient. This coefficient is generally larger than $1$, and in gambling terminology, it is called odds in decimal format [@Buhagiar2018; @Rodriguez2017]. Here we will simply refer to it as odds. Note that the definition of odds in gambling is different than the definition of odds in statistics, and in this paper we follow the former one. When a player attends one round, their chance of winning is usually close to, but less than the inverse of the odds. The difference is caused by the players’ statistical disadvantage in winning compared to the house due to the design of the game rules. In addition, the website usually charges the winner with a site cut (commission fee), which is a fixed percentage of the prize.
We further define the [*payoff*]{}, $o_p$, to be the net change of one player’s wealth after they attend one round. Although the four types of games are based on different rules, the payoffs all follow the same expression $$\label{eq:payoff}
\displaystyle o_p = \left\lbrace \begin{split}
\displaystyle &-b, &\text{with\ probability\ } &p= 1- \frac{1}{m} + f_\text{m}, \\
\displaystyle &(1-\eta)(m-1) b, &\text{with\ probability\ } &q=1-p = \frac{1}{m} - f_\text{m},
\end{split}\right.$$ where $b>0$ is the wager the player places, $m>1$ is the odds, $1>\eta\geq 0$ corresponds to the site cut, and $f_\text{m}$ is a non-negative value based on the odds representing the players’ statistical disadvantage in winning, as mentioned earlier. At least either $\eta$ or $f_\text{m}$ are non-zero.
From Eq. (\[eq:payoff\]), we can obtain the expected payoff of attending one round $$\label{eq:expected-payoff}
\displaystyle E\left( o_p|m, b\right)
= \Big(- (1-1/m +f_m) + (1-\eta)(m-1)(1/m - f_m)\Big)b
= - \Big((1-\eta)m f_m + (1-1/m + f_m)\eta \Big)b\equiv -\xi b,$$ which is always negative since either $\eta$ or $f_\text{m}$ are non-zero. In gambling terminology, $\xi$ is called the house edge, from which the websites make profits. The house edge represents the proportion the website will benefit on average when players wager. In the four types of games we discuss, the house edge $\xi$ ranges from $1\%$ to $8\%$. If there is no house edge $\xi=0$, that means it is a fair game. In a fair game or when we ignore the house edge, the expected payoff would be 0. In the Results section, we begin with an analysis of wager distribution and log-ratios between successive wagers, which helps us to understand the gamblers’ wagering strategy. We then focus on an analysis of risk attitude by studying the distribution of the odds players choose to wager with. We conclude by extending our discussion to the analysis of net incomes of gamblers viewed as random walks. This allows us to gain insights into the gamblers’ behaviors by computing quantities like the ensemble/time-averaged mean-squared displacement, the first-passage time distribution, ergodicity breaking parameter, and Gaussianity. Detailed information about the games and datasets discussed in this paper can be found in the Methods section.
Results {#results .unnumbered}
=======
Wager distribution {#wager-distribution .unnumbered}
------------------
From the viewpoint of the interaction among players, the games discussed in this paper can be grouped into two classes: in Roulette, Crash, and Satoshi Dice games, there is little or no interaction among players, whereas in Jackpot games, players need to gamble against each other. At the same time, from the viewpoint of wager itself, the games can also be grouped into two classes: In games (A-G), the wagers can be an arbitrary amount of virtual currencies, such as virtual skin tickets or crypto-currency units, whereas in game (H), the wagers are placed in the form of in-game skins, which means the wager distribution further involves the distributions of the market price and availability of the skins.
Furthermore, from the viewpoint of the odds, considering the empirical datasets we have, when analyzing the wager distribution, there are three situations: i) For Roulette and Satoshi Dice games, the odds are fixed constants, and wagers placed with the same odds are analyzed to find the distribution. ii) For Crash games, the odds are selected by the players, and wagers placed with different odds are mixed together during distribution analysis. iii) For the Jackpot game, the odds are not fixed at the moment when the player wagers.
\[table:wager-distribution\]
![ \[fig:wager\_distribution\] In games (A-G), where players are allowed to choose arbitrary bet values, the wager distribution can be best fitted by log-normal distributions (\[eq:log-normal\]). In game (D), the log-normal distribution is truncated at its maximum bet value, indicated by $^*$. The fitting lines represent the log-normal fittings. Wagers placed under the different maximum allowed bet values are discussed separately, e.g., in game (A), (A$_1$) has a maximum bet value of $500,000$, and (A$_2$) has a maximum bet value of $50,000$. On the other hand, in game (H) where wagers can only be in-game skins, the wager distribution is best described by a pairwise power law with an exponential transition, see Eq. (\[eq:powerlaw\_exp\_powerlaw\]). The red dotted line represents the log-normal fitting and the blue solid line represents the fitting of a pairwise power law with an exponential transition.](figure1.pdf){width="\columnwidth"}
In Table \[table:wager-distribution\] we categorize the 8 datasets based on the above information. At the same time, for each dataset we perform a distribution analysis of wagers at the aggregate level. Within the same dataset wagers placed under different maximum allowed bet values are discussed separately. We plot the complementary cumulative distribution function (CCDF) of the empirical data and the fitted distribution to check the goodness-of-fit, see Fig. \[fig:wager\_distribution\]. CCDF, sometimes also referred to as the survival function, is given by $\bar{F}(x) = P(X>x) = 1 - P(X\le x)$.
It turns out that when players are allowed to place arbitrary wagers (games A-G in Table \[table:wager-distribution\]), the wager distributions can in general be best-fitted by log-normal distributions. In particular, in games (A, B, C, E, F, G), the wager distribution can be approximated by the following expression $$\label{eq:log-normal}
\displaystyle P(x) =
\frac{\displaystyle\Phi\left(\frac{\ln(x+1)-\mu}
{\sigma}\right)-\Phi\left(\frac{\ln(x)-\mu}{\sigma}\right)}
{\displaystyle 1-\Phi\left(\frac{\ln(x_\text{min})-\mu}{\sigma}\right)},$$ with $x_\text{min}\le x$ and $\sigma>0$. $\Phi(\cdot)$ is the cumulative distribution function of the standard normal distribution. Meanwhile in game (D), the fitted log-normal distribution is truncated at an upper boundary $x_\text{max}$, which might result from the maximum allowed small bet value and the huge variation of the market price of crypto-currencies.
During model selection, we notice that when we select different $x_\text{min}$, occasionally a power-law distribution with exponential cutoff is reported to be a better fit, but often it does not provide a decent absolute fit on the tail, and overall the log-normal distribution provides smaller Kolmogorov-Smirnov distances, see the Methods section.
On the other hand, as we have pointed out in the previous study [@Wang2018], when players are restricted to use in-game skins as wagers for gambling, the wager distribution can be best fitted by a shifted power law with exponential cutoff. Now, with a similar situation in game (H), where wagers can only be in-game skins, we find that the early part of the curve can be again fitted by a power law with exponential cutoff, as shown in Fig. \[fig:wager\_distribution\](H). However, this time it does not maintain the exponential decay of its tail; instead, it changes back to a power-law decay. The overall distribution contains six parameters, given by the expression
$$\label{eq:powerlaw_exp_powerlaw}
\displaystyle P(x) = \left\lbrace \begin{split}
\displaystyle
&\frac{1}{c_1+c_2 c_3}\
\frac{(x-\delta)^{-\alpha}}{1+e^{\lambda(x-\beta)}},\
&\text{for}\ x \le x_\text{trans}, \\
\displaystyle
&\frac{c_3}{c_1+c_2 c_3}\ x^{-\eta}, \
&\text{for}\ x > x_\text{trans},
\end{split}\right.$$
where $\displaystyle c_1=\sum\limits_{x=x_\text{min}}^{x_\text{trans}}
\frac{(x-\delta)^{-\alpha}}{1+e^{\lambda(x-\beta)}}$, $\displaystyle c_2=\zeta\left(\eta, x_\text{trans}\right)$, and $\displaystyle c_3=x_\text{trans}^\eta
\frac{\left(x_\text{trans}-\delta\right)^{-\alpha}}
{1+e^{\lambda\left(x_\text{trans}-\beta\right)}}$.
We believe that when players are restricted to use in-game skins as wagers, the decision to include one particular skin in their wager is further influenced by the price and availability of that skin. These factors make the wager distribution deviate from the log-normal distribution, which is observed in games (A-G). This is very clear when comparing the wager distributions of games (G) and (H) as both games are jackpot games of skin gambling, and the only difference is whether players are directly using skins as wagers or are using virtual skin tickets obtained from depositing skins. The power-law tail, which was not observed in the previous study [@Wang2018], might result from the increment of the maximum allowed skin price (from \$$400$ to \$$1 800$).
The above discussions, including the results for games (A-G) in Table \[table:wager-distribution\], show that the wager distributions in pure probability-based gambling games, no matter whether the game follows parimutuel betting or fixed-odds (preset/player-selected) betting, stay log-normal as long as the players are allowed to place arbitrary amounts of wagers. This commonality of log-normal distribution no longer holds when this arbitrariness of wager value is violated, e.g., in the scenario where the player can only wager items (in-game skins).
Log-normal distribution has been reported in a wide range of economic, biological, and sociological systems [@LIMPERT2001], including income, species abundance, family size, etc. Economists have proposed different kinds of generative mechanisms for log-normal distributions (and power-law distributions as well). One particular interest for us is the multiplicative process [@Gibrat1930; @mitzenmacher2003]. Starting from an initial value $X_0$, random variables in a multiplicative process follow an iterative formula $X_{i+1}=\exp(\nu_i) X_i$ or $\ln X_{i+1}= \ln X_i + \nu_i.$ If the $v_i$ has finite mean and variance, and is independent and identically distributed, then according to the central limit theorem, for large $i$, $\ln X_i$ will follow a normal distribution, which means $X_i$ will follow a log-normal distribution.
If we want to check whether gamblers follow multiplicative processes when they wager, we can first check the correlation between consecutive bets $(b_i, b_{i+1})$. Due to the large variances of the wager distributions, Pearson’s correlation coefficient may perform poorly. Instead, we adopt two rank-based correlation coefficients, Kendall’s Tau [@Kendall1990] $\tau_K$ and Spearman’s Rho [@Taylor1987] $\rho_S$. At the same time, we also check the mean and variance of the log-ratios $\ln(b_{i+1}/b_{i})$ between consecutive bets. These statistics can be found in Table \[table:log\_ratio\]. The results reveal that the values of consecutive bets exhibit a strong positive correlation, with all the correlation coefficients larger than $0.5$. It shows that players’ next bet values are largely dependent on their previous bet values. At the same time, the bet values are following gradual changes, rather than rapid changes. These conclusions can be confirmed by the small mean values and small variances of log-ratios between consecutive bets.
\[table:log\_ratio\]
![ \[fig:log\_ratio\] The distribution of the logarithmic of the ratio (log-ratio) between consecutive bet values. For games (A, B, C), the log-ratio can be described by a Laplace distribution. For games (D, F, G, H), the log-ratio presents bell-shaped distribution. In general, the distributions are symmetric with respect to the y-axis, except in games (D) and (F). The $x$-coordinate $\log_{10}(b_{i+1}/b_i)$ is proportional to the parameter $\nu$. ](figure2.pdf){width="\columnwidth"}
Further analysis of the distribution of $\nu$ shows an exponential decay on both of its tails, see Fig. \[fig:log\_ratio\]. This means that $\nu$ approximately follows a Laplace distribution. However, compared to a Laplace distribution, the empirical log-ratio distribution shows a much higher probability at $\nu=0$, whose value can be found in the last column of Table \[table:log\_ratio\]. We also observe that $\nu$ presents higher probability densities around small integers/half-integers and their inverses. Due to the existence of these differences, we will skip the parameter fitting for the distribution of $\nu$. The high probability of staying on the same wager indicates that betting with fixed wager is one of the common strategies adopted by gamblers.
Meanwhile, the high positive auto-correlations, along with the higher probability densities at small integers/half-integers and their inverses, provide evidence that gamblers often follow a multiplicative process when wagering. The multiplication process can be explained by the wide adoption of multiplicative betting systems. “Betting system” here refers to the strategy of wagering where the next bet value depends on both the previous bet value and the previous outcome [@Dubins2014; @Epstein2012]. Although betting systems will not provide a long-term benefit, as the expected payoff will always be $0$ in a fair game, still they are widely adopted among gamblers. A well-known multiplicative betting system is the Martingale (sometimes called geometric progression) [@Epstein2012]. In Martingale betting, starting with an initial wager, the gambler will double their wager each time they lose one round, and return to the initial wager once they win. Martingale is a negative-progression betting system where the gambler will increase their wager when they lose and/or decrease their wager when they win.
Apart from multiplicative betting, there are many other types of betting systems, such as additive betting and linear betting [@Epstein2012]. The reasons why multiplicative betting systems are dominant in our datasets are: 1) Martingale is a well-known betting system among gamblers; 2) Many online gambling websites provide a service for changing the bet value in a multiplicative way. For example, for the Crash games csgofast-Crash (C) and ethCrash (D), both websites provide a simple program for automatically wagering in a multiplicative way. For the Roulette games and Coinroll (F), the websites provide an interface with which the gambler can quickly double or half their wager. However, for Satoshi Dice (E) and csgospeed-Jackpot (G), no such function is provided, yet we still observe similar results, indicating that gamblers will follow a multiplicative betting themselves.
Fig. \[fig:log\_ratio\] provides us with the distribution of $\nu$, however, it will not tell us whether the gamblers adopt the negative/positive-progression betting systems. Therefore we further analyze the effect on the bet values of winning/losing a round. How the gamblers adjust their wager after winning/losing rounds is shown in Table \[table:win-lose\]. We can see that although there is a high probability for sticking to the same bet values, the most likely outcome after losing a round is that the gambler increases their wager. When winning one round, gamblers are more likely to decrease their wager. This means that negative-progression strategies are more common among gamblers than positive-progression strategies.
\[table:win-lose\]
Risk attitude {#risk-attitude .unnumbered}
-------------
We now turn to the following question: When a player is allowed to choose the odds themselves in a near-fair game, how would they balance the risk and potential return? Higher odds means a lower chance of winning and higher potential return, for example, setting odds of $10$ means that the winning chance is only $1/10$, but the potential winning payoff equals $9$ times the original wager. In our analysis, we can examine such behaviors based on the gambling logs from Crash and Satoshi Dice games. For the Crash game only CSGOFAST.COM provides the player-selected odds even when players lose that round, whereas for the Satoshi Dice game only Coinroll accepts player-selected odds. We will therefore focus on the data collected on these two websites. For the Crash game on CSGOFAST.COM, the odds can only be set as multiples of $0.01$, whereas for the Satoshi Dice game on Coinroll the odds can be set to $0.99 \cdot 65536/i$ where $i$ is a positive integer less than $64000$. To simplify our modeling work, we will convert the odds on Coinroll to be multiples of $0.01$ (same as for the Crash game).
![\[fig:odds\] Odds distributions can be well-fitted by truncated shifted power-law distributions. ](figure3.pdf){width=".75\columnwidth"}
It turns out that in both cases the odds can be modeled with a truncated shifted power-law distribution, $$\displaystyle P(m) = \left\lbrace \begin{split}
\displaystyle &\frac{\left(\ m - \delta \right)^{-\alpha}}{\zeta(\alpha, \ m_\text{min} -\delta)},
\ &\text{for}\ m_\text{min}\le m < m_\text{max}, \\
\displaystyle &\frac{\zeta(\alpha,\ m_\text{max} -\delta)}{\zeta(\alpha, \ m_\text{min} -\delta)},
\ &\text{for}\ m = m_\text{max},
\end{split}\right.$$ where $\zeta(\cdot,\cdot)$ is the incomplete Zeta function, and $m_\text{max}$ is the upper truncation. Note that there is a jump at $m_\text{max}$, meaning that the players are more likely to place bets on the maximum allowed odds than on a slightly smaller odds. The estimated parameters $\alpha=1.881$, $\delta=0.849$, and $m_\text{min}=1.15$ for csgofast-Jackpot on CSGOFAST.COM, whereas for Coinroll the parameters are found to be $\alpha = 1.423$, $\delta = 2.217$, and $m_\text{min} = 2.58$. From the comparison between the CCDFs of empirical data and fitting curves, as shown in Fig. \[fig:odds\], we can see that the truncated shifted power law can capture the overall decaying trends of odds distribution. The stepped behavior results from the gamblers’ preference of simple numbers. A distribution that is close to a power law indicates that a gambler’s free choice of odds displays scaling characteristics (within the allowed range) in near-fair games. It also means that when gamblers are free to determine the risks of their games, although in most times they will stick to low risks, showing a risk-aversion attitude, they still present a non-negligible probability of accepting high risks in exchange for high potential returns. The scaling properties of risk attitude might not be unique to gamblers, but also may help to explain some of the risk-seeking behaviors in stock markets or financial trading.
We now re-examine the distributions from the point of view of estimating the crash point $m_C$ (Satoshi Dice games can be explained with the same mechanism). The true distribution of $m_C$ generated by the websites follow a power-law decay with an exponent of $2$ (with some small deviation due to the house edge). Meanwhile, a closer look at the fitted exponents listed above gives us two empirical exponents of $1.423$ and $1.881$, both of which are smaller than $2$. The smaller exponents reveal that gamblers believe that they have a larger chance to win a high-odds game than they actually do. Or equivalently, it means the gamblers over-weight the winning chance of low-probability games. At the same time, the “shifted” characteristics here lead to more bets on small odds, which also indicates that the gamblers over-estimate the winning chance of high-probability games. As a result, they under-weight the winning chances of mild-probability games. These are clear empirical evidence of probability weighting among gamblers, which is believed to be one of the fundamental mechanisms in economics [@Barberis2012].
Wealth distribution {#wealth-distribution .unnumbered}
-------------------
In the previous study of skin gambling [@Wang2018], we pointed out that the wealth distribution of skin gamblers shows a pairwise power-law tail. This time, by considering the players’ deposits to their wallets on a gambling site as the wealth data, we find that the pairwise power-law tails are also observed for bitcoin gambling. We find that on the gambling website Coinroll, starting from $5660$ cents, the players’ wealth distribution follows a pairwise power-law distribution, with the power of the first regime to be $1.585$, and the power of the second regime to be $3.258$, see Fig. \[fig:wealth\]. The crossover happens at $1.221\times 10^5$ cents. As both wealth distributions of skin gambling and bitcoin gambling can be approximated by a pairwise power distribution, we believe that it is a good option for modeling the tails of gambler wealth distribution in different scenarios.
![ \[fig:wealth\] The tail of the wealth distribution of Bitcoin gamblers follows a pairwise power-law distribution. ](figure4.pdf){width=".5\columnwidth"}
Removing effects due to inequality in the number of bets {#removing-effects-due-to-inequality-in-the-number-of-bets .unnumbered}
--------------------------------------------------------
In the above sections, we have analyzed the distributions of several quantities at the population level. However, there is a huge inequality of the number of placed bets among gamblers. We therefore wonder whether those distributions we obtain result from the inequality of number of bets among individuals. To remove the effects of this inequality, we randomly sample in each dataset the same number of bets from heavy gamblers. We re-analyze the wager distribution and odds distribution with the sample data to see if we obtain the same distribution as before. In each dataset we randomly sample $500$ bets from each of those gamblers who placed at least $500$ bets above $b_\text{min}$ given in Table \[table:wager-distribution\]. Some datasets are excluded here as either they do not have enough data or we cannot identify individual gamblers. When re-analyzing the odds distribution, to ensure we have enough data, we respectively sample $100$ and $2000$ bets from each of those gamblers in games (C) and (F) who have at least $100$ and $2000$ valid player-selected odds above $m_\text{min}$. According to the results in Fig. \[fig:wager\_distr\_top\], after removing the inequality the wager distributions can still be approximated by log-normal distributions, but some deviation can be observed. Similarly, the odds distributions again follow truncated shifted power-law distributions after removing the inequality. These results demonstrate that the shape of the distributions we obtained in the above sections is not a result of the inequality of the number of bets.
![\[fig:wager\_distr\_top\] The wagers obtained from random sampling of top gamblers’ bets still present log-normal distributions, although there are some observable deviations. ](figure5.pdf){width="\columnwidth"}
Now our question becomes whether the conclusion regarding the distribution at the population level can be extended to the individual level. Here due to the limitation of data, we will only discuss the wager distribution. Analyzing the individual distribution of top gamblers, we find that although heavy-tailed properties can be widely observed at the individual level, only a small proportion of top gamblers presents log-normal distributed wagers. Other distributions encountered include log-normal distributions, power-law distributions, power-law distributions with exponential cutoff, pair-wise power-law distributions, irregular heavy-tailed distributions, as well as distributions that only have a few values. The diversity of the wager distributions at the individual level suggests a diversity of individual betting strategies. Also, it indicates that a gambler may not stick to only one betting strategy. It follows that the log-normal wager distribution observed at the population level is very likely an aggregate result.
Diffusive process {#diffusive-process .unnumbered}
-----------------
For an individual player’s gambling sequence we define “time” $t$ as the number of bets one player has placed so far, and define as net income the sum of the payoffs of those bets. In all the games we analyze, there are only two possible outcomes: a win or a loss. The player’s net income will change each time they place a bet in a round, with the step length to be the payoff from that bet. We can treat the change of one player’s net income as a random walk in a one-dimensional space (see Fig. 1 in Ref. [@Wang2018] for an example of such a trajectory). The time $t$ will increase by $1$ when the player places a new bet, therefore the process is a discrete-time random walk.
Now, let us focus on the analysis of the diffusive process of the gamblers’ net incomes, starting with the analysis of the change of the mean net income with the number of rounds played (time), $\left\langle \Delta w(t) \right\rangle = \left\langle w(t) - w_0 \right\rangle =
\left\langle \sum\limits_{i=1}^t o_p(i) \right\rangle,$ where $w_0$ is the player’s initial wealth, $w(t)$ is the player’s wealth after attending $t$ rounds, and $o_p(i)$ is the payoff from the $i_{th}$ round the player attended. $\langle \cdot \rangle$ represents an ensemble average over a population of players placing bets. In the rest of this paper, $\langle \cdot \rangle$ will always be used for representing an ensemble average. In Fig. \[fig:netincome\] we show the change of $\left\langle \Delta w(t) \right\rangle$ over time. In most of the datasets, players’ mean net income decreases over time, which suggests that in general players will lose more as they gamble more. At the same time, in some datasets such as Ethcrash (D) and Coinroll (F), large fluctuations can be observed.
![ \[fig:netincome\] Change of the mean net income with time for the different datasets. Most of the datasets present a decreasing net income as time $t$ increases. Each point is obtained from an average over at least $200$ players.](figure6.pdf){width="\columnwidth"}
An useful tool for studying the diffusive process is the ensemble-averaged mean-squared displacement (MSD), defined as $$\left\langle\Delta w^2(t) \right\rangle =
\left\langle\left(w(t)-w_0\right)^2\right\rangle =
\left\langle \left(\sum\limits_{i=1}^t o_p(i)\right)^2 \right\rangle,$$ For a normal diffusive process, $\left\langle\Delta w^2(t) \right\rangle \sim t$, otherwise an anomalous diffusive behavior prevails. More specifically, when the MSD growth is faster (respectively, slower) than linear, superdiffusion (respectively, subdiffusion) is observed.
![ \[fig:msd\] The growth of ensemble-averaged mean-squared displacement in different datasets presents different diffusive behaviors. In the figures, the error bars represent $95\%$ confidence intervals, blue dashed lines follow linear functions (slope $=1$), and green dotted lines follow quadratic functions (slope $=2$).](figure7.pdf){width="\columnwidth"}
In Fig. \[fig:msd\], we present the growth of the ensemble-averaged MSD against time for each of the datasets. To reduce the coarseness, MSD curves are smoothed with log-binning technique. The error bars in Fig. \[fig:msd\] represent $95\%$ confidence intervals computed with bootstrapping using $2000$ independent re-sampling runs. It is interesting to see that for different datasets we observe different diffusive behaviors. For games csgofast-Crash (C) we observe that the MSD grows faster than a linear function, suggesting superdiffusive behavior. Meanwhile, for games csgofast-Double (A), ethCrash (D), csgospeed (G), and csgofast-Jackpot (H), the MSD first presents a superdiffusive regime, followed by a crossover to a normal diffusive regime. For games csgofast-X50 (B) and Coinroll (F), although the ensemble-averaged MSD roughly presents a linear / sublinear growth, a careful inspection shows that both curves consist of several convex-shaped regimes, indicating a more complex behavior. Convex-shaped regimes can also be observed in csgofast-Crash games (C).
In Ref. [@Wang2018] we argued that the crossover from a superdiffusive regime to a normal diffusive regime in a parimutuel game is due to the limitation of individuals’ wealth and the conservation of total wealth. Similar crossovers are observed in games (G) and (H), two parimutuel betting games, where the same explanation can be applied. On the other hand, this crossover is also found in a Roulette game and in a Crash game, where there is no interaction among gamblers. The limitation of an individual’s wealth can still be a partial explanation, but the conservation of total wealth no longer holds. A different explanation needs to be proposed to model this crossover.
In the following we briefly discuss how we can obtain from gambling models the different diffusive processes observed in the data. We will not attempt to reproduce the parameters we obtained from the gambling logs, but rather try to explore the possible reasons for the anomalous diffusion we reported.
For a gambling process, if the gambler’s behavior is independent among different rounds, i.e., the wager and odds are respectively independent and identically distributed (IID), with no influence from the previous outcomes, and if the wager $b$ has finite variance and the odds $m$ has finite mean, then MSD’s growth will be a linear function of time $t$: $$\label{eq:linear-msd}
\displaystyle \left\langle \Delta w^2(t) \right\rangle =
\left\langle\left(w(t)-w_0\right)^2\right\rangle
=(\langle m\rangle-1)\left\langle b^2\right\rangle t,$$ where $\langle m\rangle$ is the mean value of odds distribution and $\left\langle b^2\right\rangle$ is the second moment of the wager distribution. But normal diffusion is only found in few datasets, the remaining datasets presenting anomalous diffusion which conflicts with the IID assumption.
Having shown the popularity of betting systems among gamblers, we would like to check how different betting systems affect diffusive behaviors. First, we simulate gamblers that follow Martingale strategies in a Crash game. We assume that the selection of odds follows a power-law distribution with an exponent $\alpha$, with a minimum odds of $1$ and a maximum odds of $50$, where the maximum odds is set to ensure a finite mean of the odds distribution. Starting from a minimum bet of $1$, we multiply wagers by a ratio $\gamma$ each time the gamblers lose one round and return to the minimum bet each time they win. Once the wager reaches a preset maximum bet value $10000$, we reset the gambler with a minimum bet. MSD obtained from 10 billion individual simulations is shown in Fig. \[fig:martingale\]. Different curves correspond to different exponents in odds distribution. We can see that the MSD initially presents an exponential-like growth, before the growths reduce to a linear function. It is easy to explain the exponential growth since many gamblers lose the rounds and therefore increase their wager by the factor $\gamma$, which leads to an increase in the average bet value. The superdiffusion here suggests that Martingale strategy increases gamblers’ risks of huge losses. Considering the wide adoption of Martingale among gamblers, this could be a reason for the superdiffusion as well as the crossover to normal diffusion we found in several datasets. Comparison of the MSD curves of different $\alpha$ suggests that a more aggressive risk attitude leads to a higher risk of huge losses (as well as higher potential winnings).
![ \[fig:martingale\] A betting system similar to Martingale will lead to a crossover from superdiffusion to normal diffusion according to the growth of mean-squared displacement. Comparison between curves of different parameters shows that higher $\gamma$ and lower $\alpha$ both will lead to a higher chance of huge losses/winnings. ](figure8.pdf){width=".75\columnwidth"}
Next we examine the ergodicity of the random walk process of net income by computing the time-averaged mean-squared displacement and the ergodicity breaking parameter. The time-averaged MSD is defined as $$\displaystyle \overline{\delta^2}(t) = \frac{1}{T-t} \sum\limits_{k=1}^{T-t} \left(w(k + t) - w(k)\right)^2,$$ where $T$ is the length of the player’s betting history, i.e. total number of rounds they attend, and $\overline{\cdots}$ is used for representing a time average. To calculate the time-averaged MSD, we need to make sure the player has played enough rounds so that we have a long enough series of net income data, therefore in each dataset we filter out the players who played less than $T=1000$ rounds. As shown in Fig. \[fig:msd\_t\] the time-averaged MSD shows huge deviations from player to player, suggesting diverse betting behaviors at the individual level. At the same time, comparison between the ensemble-averaged time-averaged MSD $\left\langle \overline{\delta^2}(t) \right\rangle$ and the ensemble-averaged MSD $\left\langle\Delta w^2(t)\right\rangle$ shows clear deviations in most datasets, except in the Coinroll (F), csgospeed (G) and csgofast-Jackpot (H) games. To further examine breaking of ergodicity, we have calculated the ergodicity breaking parameter $\rm{EB}$ [@Cherstvy2013; @Cherstvy2014; @Cherstvy2019] defined as $$\displaystyle \rm{EB}(t) = \left\langle \left(\overline{\delta^2}(t)\right)^2\right\rangle \bigg/ \left\langle\overline{\delta^2}(t)\right\rangle^2 - 1.$$
For an ergodic process, the parameter $\rm{EB}$ should be close to $0$. However, as shown in Fig. \[fig:eb\], in most datasets, with the exception of csgospeed (G) and csgofast-Jackpot (H), $\rm{EB}$ is large. It follows that non-ergodicity is observed in most games and that gambling processes indeed often deviate from normal diffusion, which further highlights the complexity of human gambling behavior.
![ \[fig:msd\_t\] The growth of the time-averaged MSD for individual gamblers, presented as thin lines, suggests diverse betting behaviors at the individual level. The comparison between $\left\langle \overline{\delta^2}(t) \right\rangle$ (thick dashed black lines) and $\left\langle\left( \Delta w(t) \right)^2 \right\rangle$ (thick full red lines) reveals that these quantities are different for most games, with the exception of the Coinroll (F), csgospeed (G) and csgofast-Jackpot (H) games. Players who played less than $1000$ rounds are filtered out in each dataset. ](figure9.pdf){width=".75\columnwidth"}
![ \[fig:eb\] The change of the ergodicity breaking parameter with time. For all games, with the exception of the games csgospeed (G) and csgofast-Jackpot (H), ${\rm{EB}}$ is found to be much larger than $0$, suggesting non-ergodic behavior.](figure10.pdf){width=".75\columnwidth"}
Another way to examine the diffusive behavior of a process is through the analysis of the first-passage time distribution. The first-passage time $t_{FP}$ is the time required for a random walker at location $w$ to leave the region $[w-V_{FP},w +V_{FP}]$ for the first time, where $V_{FP}$ is the target value or first-passage value. The first-passage time distribution $P(t_{FP})$ [@Inoue2007; @Wang2018], defined as the survival probability that the random walker, who is located at $w$ at time $t_0$, stays within range $[w-V_{FP},w +V_{FP}]$ up to time $t = t_0 + t_{FP}$, can be calculated from the expression $$P(t_{FP}) = \lim\limits_{T \longrightarrow \infty} \frac{1}{T} \sum\limits_{k=1}^T \Theta \left( \left|
w(k+t_{FP}) - w(k) \right| - V_{FP} \right) - \lim\limits_{T \longrightarrow \infty} \frac{1}{T} \sum\limits_{k=1}^T \Theta \left( \left|
w(k+t_{FP}-1) - w(k) \right| - V_{FP} \right)~,$$ where $\Theta(\cdot)$ is the Heaviside step function. We use $V_{FP} = 200$ (US cents), with the exception of csgofast-Jackpot (H) for which $V_{FP}$ is chosen to be 5000. For a normal diffusive process, the tail of $P(t_{FP})$ should decay with an exponent of $3/2$. In Fig. \[fig:fpt\] we plot the first-passage time distribution for each dataset, where again diverse diffusive behaviors are observed. In the games csgofast-Double (A) and csgofast-Jackpot (H), the tails of $P(t_{FP})$ approximately decay with an exponent of $3/2$ (see the thin green lines), indicating normal diffusive processes. For the game csgospeed (G), the exponent is found to be larger than $3/2$, indicating a superdiffusive process. And in games csgofast-X50 (B), csgofast-Crash (C), ethCrash (D), and Coinroll (F), the exponents are clearly smaller than $3/2$, indicating a subdiffusive behavior. We note that the results obtained from ensemble-averaged MSD sometimes differ from the results obtained from the first-passage time distributions. Nonetheless, anomalous diffusive behavior is widely observed.
![ \[fig:fpt\] The tails of first-passage time distributions for the different datasets indicate different diffusive behaviors. The green lines represent a power-law decay with an exponent $3/2$. The blue error bars indicate $95\%$ confidence intervals. Only gamblers who attended more than 1000 rounds of games have been included in these calculations. ](figure11.pdf){width=".75\columnwidth"}
To confirm our conclusion about the wide existence of anomalous diffusive behavior in gambling activities, we further calculate the non-Gaussian parameter (NGP) [@Rahman1964; @Toyota2011; @Cherstvy2019] $$\displaystyle NGP(t) = \frac{\left\langle \Delta w^4(t)\right\rangle}{3\left\langle \Delta w^2(t)\right\rangle^2} - 1~.$$ For a Gaussian process, the NGP should approach $0$ when $t$ gets large. In Fig. \[fig:ngp\] we show the NGP as a function of time. In most of the games, except Coinroll (F), NGP shows a clear decreasing trend as $t$ increases. In the game Coinroll (F), a decrease is not apparent, and most likely this game does not follow a Gaussian process. In the other games, although the NGP is still decreasing, we can not discriminate whether for large $t$ this quantity will tend to $0$ or instead reach a plateau value larger than zero. For example, for the game csgospeed (G) the NGP seems to reach a plateau $NGP(t)\approx 1.5$ instead of continuing to decrease, but this could also be the consequence of insufficient data. Still, our analysis does not provide clear evidence for the presence of Gaussianity in gambling behaviors.
![ \[fig:ngp\] In most datasets, except Coinroll (F), the non-Gaussian parameter shows a decreasing trend as $t$ increases. However, in none of the studied cases does the non-Gaussian parameter fall below the value 1. ](figure12.pdf){width=".75\columnwidth"}
To sum up our analysis of the players’ net incomes viewed as random walks, the diverse diffusive behaviors found in the datasets indicate that human gambling behavior is more complex than random betting and simple betting systems. Further studies are required in order to fully understand the observed differences. At the individual level, as has been pointed out by Meng [@Meng2018], gamblers show a huge diversity of betting strategies, and even individual gamblers constantly change their betting strategy. Differences in the fractions of gamblers playing specific betting strategies could be a reason why we see a variety of diffusive behaviors in the datasets.
Discussion {#discussion .unnumbered}
==========
The quick development of the video gaming industry has also resulted in an explosive growth of other online entertainment. This is especially true for online gambling that has evolved quickly into a booming industry with multi-billion levels. Every day million of bets are placed on websites all around the globe as many different gambling games are available online for gamblers.
Analysing different types of gambling games (ranging from Roulette to Jackpot games), we have shown that log-normal distributions can be widely used to describe the wager distributions of online gamblers at the aggregate level. The risk attitude of online gamblers shows scaling properties too, which indicates that although most gamblers are risk-averse, they sometime will take large risks in exchange for high potential gains.
Viewing the gamblers’ net income as a random walk in time (where for each gambler time is increased by one unit every time they play a game), we can analyze the mean-squared displacement of net income and related quantities like the ergodicity breaking parameter or the non-Gaussian parameter with the goal to gain an understanding of the gamblers’ betting strategies through the diffusive behaviors emerging from the datasets. For some games the mean-squared displacement and the first-passage time distribution reveal a transition from superdiffusion to normal diffusion as time increases. For all games the ergodicity breaking parameter and the non-Gaussian parameter reveal deviations from normal diffusion. All this indicates that gamblers’ behaviors are very diverse and more complex than what would be expected from simple betting systems. We speculate that one of the reasons for the observed diverse diffusive behaviors at the aggregate level can be found in the differences in the fractions of gamblers playing specific betting strategies, but more work is required to fully understand the gamblers’ complex behaviors.
Methods {#methods .unnumbered}
=======
Detailed rules of the different games {#detailed-rules-of-the-different-games .unnumbered}
-------------------------------------
### Roulette {#roulette .unnumbered}
We focus on a simplified version of Roulette games that appears in online casinos, where a wheel with multiple slots painted with different colors will be spun, after which a winning slot will be selected. The Roulette table of a traditional Roulette game is composed of $38$ slots, among which $18$ slots are painted in black, $18$ slots are painted in red, and two slots (“0” and “00”) are painted in green. The online Roulette games are similar to the traditional ones, except that the number of colors and the number of slots for each color might be different. Each slot has the same probability to be chosen as the winning slot. Players will guess the color of the winning slot before the game starts. The players have a certain time for wagering, after which the game ends and a winning slot is selected by the website. Those players who successfully wagered on the correct color win, the others lose. As the chance of winning and odds for each color are directly provided by the website, roulette is a fixed-odds betting game.
### Crash {#crash .unnumbered}
“Crash” describes a type of gambling games mainly hosted in online casinos. Before the game starts, the site will generate a crash point $m_C$, which is initially hidden to the players. With a lower boundary of $1$, the crash point is distributed approximately in an inverse square law. The players need to place their wager in order to enter one round. After the game starts, on the player’s user interface a number, called multiplier, will show up and gradually increase from $1$ to the predetermined crash point $m_C$, after which the game ends. During this process, if the player “cash-outs” at a certain multiplier $m$, before the game ends, they win the round; otherwise they lose. This multiplier $m$ they cashed out at is the odds, which means when winning, the player will receive a prize that equals his wager multiplied by $m$. When $m_C$ is generated with a strict inverse-square-law distribution, the winning chance exactly equals the inverse of the player-selected odds $m$. The player can also set up the cash-out multipliers automatically before the game starts, to avoid the possible time delay of manual cash-out. Since in a manual cash-out scenario, after the game starts, the multiplier will show up on the screen, at a given moment the decision of the cash-out multiplier is based on the player’s satisfaction with the current multiplier, and involves more complicated dynamics of decision-making processes. Meanwhile, in an auto cash-out scenario, the multiplier $m$ is chosen before the game starts, which means the decision making is more “static.” Crash is also a fixed-odds betting game where the odds are player-selected.
### Satoshi Dice {#satoshi-dice .unnumbered}
Satoshi Dice is one of the most popular games in crptocurrency gambling. In 2013, the transactions resulting from playing Satoshi Dice games accounted for about $60\%$ of overall Bitcoin transactions [@Meiklejohn2013]. When playing Satoshi Dice, the player needs to pick a number $A$ within a range $(0, U)$ provided by the website. The odds can be calculated with the expression $m=U/A$. Once the player finishes wagering, the website will pick another number $B$ which is uniformly distributed on $(0, U)$. If $B$ is less than $A$, then the player wins the round, otherwise they lose. Satoshi Dice is a fixed-odds betting game. In some online casinos, players cannot choose $A$ arbitrarily, but instead, they have to select $A$ from a preset list provided by the gambling website. Since the odds $m$ is determined from $A$, we are more interested in the case where the players can choose $A$ arbitrarily, from which we can obtain a more detailed distribution of the odds $m$, which helps us to understand the players’ risk attitude. According to the rules of Satoshi Dice games, the maximum allowed bet is proportional to the inverse of $A$, which means the accepted range of wager is directly related to the odds.
### Jackpot {#jackpot .unnumbered}
Unlike the games discussed above, Jackpot is a parimutuel betting game, where players gamble against each other. During the game, each player attending the same round will deposit their wager to a pool. The game-ending condition varies across different websites, it could be a certain pool size, a certain amount of players, or a preset time span. When the game ending condition is reached, each player’s winning chance will be determined by the fraction of their wager in the wager pool, based on which one player will be chosen as the winner by the website. The winner will obtain the whole wager pool as the prize, after excluding the site cut. The odds can be calculated by the pool size divided by the player’s wager, but it is unknown to the players at the moment they wager. In the previous study [@Wang2018], we have already discussed the player’s behavior in Jackpot games of skin gambling where in-game skins are directly used as wagers. In this paper, we extend the analysis to a case where wagers can be arbitrary amounts of virtual skin tickets (players need to first exchange in-game skins into virtual skin tickets).
Data summary {#data-summary .unnumbered}
------------
For each type of game, we collect two datasets. In total, we analyze 8 datasets collected from 4 different online gambling websites, and the number of bet logs contained in each dataset ranges from $0.3$ million to $19.2$ million. Due to the high variation of market prices of crypto-currencies and in-game skins, the wager and deposits are first converted into US cents based on their daily market prices.
### CSGOFAST {#csgofast .unnumbered}
From the skin gambling website CSGOFAST [@csgofast] we collected four datasets on the Roulette, Crash and Jackpot games ([*csgofast-Double*]{}, [*csgofast-X50*]{}, [*csgofast-Crash*]{}, [*csgofast-Jackpot*]{}) it provides.
[*csgofast-Double*]{} (A) is a Roulette game in which players can bet on 3 different colors (Red, Black, Green), which respectively provide odds of (2, 2, 14). The data were collected in two different time periods, and the only difference between them is a change of the maximum allowed bet values. [*csgofast-X50*]{} (B) is also a Roulette game in which players can bet on 4 different colors (Blue, Red, Green, Gold) with odds (2, 3, 5, 50).
[*csgofast-Crash*]{} (C) is a Crash game. As we mentioned earlier, when analyzing the risk attitude of gamblers in Crash game, we are more interested in how players set up the odds (multiplier) with the automatically cash-out option. On CSGOFAST, under the automatically cash-out option, players can only setup odds ranging from $1.10$ to $50$. The interesting point about this dataset is that even if the player loses the round, if they used the automatically cash-out option, it still displays the player-selected odds (which is set before the game starts); meanwhile if they used the manually cash-out option, no odds is displayed. Therefore in early-crashed games ($m_C < 1.10$), all the displayed odds that are larger than $1.10$ were placed with automatically cash-out option. These displayed odds will be used in odds distribution analysis. The data are also collected in two different periods, where the only difference is still a change of the maximum allowed bet value. Roulette and Crash games on CSGOFAST all use virtual skin tickets for wagering.
[*csgofast-Jackpot*]{} (H) is a Jackpot game, where in-game skins are directly placed as wagers. Each skin has a market value that ranges from $3$ to $180 000$ US cents. A player can place at most $10$ skins in one round.
### CSGOSpeed {#csgospeed .unnumbered}
From the skin gambling website CSGOSpeed [@csgospeed] we collected one dataset from its Jackpot game [*csgospeed-Jackpot*]{} (G), in which arbitrary amounts of virtual skin tickets can be used as wagers. The difference between datasets (H) and (G) focuses on whether the wagers are in-game skins or virtual skin tickets.
### ethCrash {#ethcrash .unnumbered}
ethCrash [@ethcrash] is a cryto-currency gambling website providing a Crash game [*ethCrash*]{} (D). Players need to place wagers in Ethereum (ETH), one type of crypto-currency.
### SatoshiDice {#satoshidice .unnumbered}
SatoshiDice [@satoshidice] is a cryto-currency gambling website which accepts Bitcoin Cash (BCH) as wagers. It provides a Satoshi Dice game [*satoshidice*]{} (E), where only 11 preset odds can be wagered on, ranging from $1.05$ to $1013.74$. Among the preset odds, we find that more than $30\%$ of the bets are placed under the odds $1.98$, and we will analyze those bets for wager distribution.
### Coinroll {#coinroll .unnumbered}
Coinroll [@coinroll] is a cryto-currency gambling website which accepts Bitcoin (BTC) as wagers. It provides a Satoshi Dice game [*Coinroll*]{} (F), where players can either wager on the 8 preset odds listed by the website, or choose an odds of their own. When further analyzing the data, we find that a few players placed an unusual large amount of bets, where the top player placed more than 11 million bets. Although these large number of bets prove the heavy-tailed distribution of the number of bets of individuals, we have doubts that these players are playing for the purpose of gambling. As we have pointed out, all the games discussed in this paper have negative expected payoffs. Indeed, prior studies have raised suspicion about the use of crypo-currency gambling websites as a way for money laundering [@fiedler2013]. We will therefore exclude from our analysis gamblers who placed more than half a million bets. For bets wagered on the preset odds, we find that more than $57\%$ are placed under the odds $1.98$, and we use these bets to analyze the wager distribution. On the other hand, since player-selected odds show a broader spectrum regarding the risk attitude of gamblers, we focus on the odds distribution of the player-selected odds. As already mentioned, we will exclude the bets from those players who placed at least half a million bets from our odds distribution analysis.
Although crypto-currency has gained decent popularity in the financial and technological world, in this paper we still measure the wager/wealth deposited in forms of crypto-currencies in US dollars, since the wagers in skin gambling are measured in US dollars. The historical daily price data of crypto-currencies (Bitcion, Ethereum, Bitcoin Cash) are obtained from CoinDesk [@coindesk] (for Bitcoin) and CoinMetrics [@coinmetrics] (for Ethereum and Bitcoin Cash).
Ethics for data analysis {#ethics-for-data-analysis .unnumbered}
------------------------
The data collected and analyzed in this paper are all publicly accessible on the internet, and we collect the data either with the consent of the website administrators or without violating the terms of service or acceptance usage listed on the hosting website. The data we use do not include any personally identifiable information (PII), and we further anonymize account-related information before storing them into our databases to preserve players’ privacy. In addition, our data collection and analysis procedures are performed solely passively, with absolutely no interaction with any human subject. To avoid abusing the hosting websites (i.e., the gambling websites), the request rates of data-collecting are limited to $1$ request per second. Considering the legal concerns and potential negative effects of online gambling [@Martinelli2017; @Millar2018; @Kairouz2012; @Gonzalez2017; @Gainsbury2015; @Banks2017; @Redondo2015; @Macey2018], our analysis aims only to help better prevent adolescent gambling and problem gambling.
Parameter estimation and model selection {#parameter-estimation-and-model-selection .unnumbered}
----------------------------------------
In our analysis, the parameters of different distribution models are obtained by applying Maximum Likelihood Estimation (MLE) [@Bauke2007]. To select the best-fit distribution, we compare the models’ Akaike weights [@Burnham2002] derived from Akaike Information Criterion (AIC). Note that analyzing the fitting results, we constantly found that players show a tendency of using simple numbers when allowed to place wagers with arbitrary amounts of virtual currency. As a result, the curves of probability distribution functions appear to peak at simple numbers, and the corresponding cumulative distribution function shows a stepped behavior. This makes the fitting more difficult, especially for the determination of the start of the tail. To address this issue, we choose the start of the tail $x_{min}$ such that we obtain a small KolmogorovSmirnov (KS) distance between the empirical distribution and the fitting distribution, while maintaining a good absolute fit between the complementary cumulative distribution functions (CCDF) of the empirical distribution and the best-fitted distribution. Candidate models for model selection in this paper include exponential distribution, power-law distribution, log-normal distribution, power-law distribution with sharp truncation, power-law distribution with exponential cutoff, and pairwise power-law distribution. More details about parameter fitting and model selection can be found in the article by Clauset et al. article [@Clauset2009] as well as in the previous paper by the authors [@Wang2018].
Data availability {#data-availability .unnumbered}
=================
The datasets generated and/or analysed during the current study are available from the authors on reasonable request.
[99]{} American Gaming Association. State of the states: The AGA survey of the commercial casino industry (2018). Accessed October 1, 2018. Calado, F. & Griffiths, M. D. Problem gambling worldwide: An update and systematic review of empirical research (2000–2015). [*J. Behav. Addict.*]{} 5, 592-613 (2016). Calado, F., Alexandre, J. & Griffiths, M. D. Prevalence of adolescent problem gambling: A systematic review of recent research. [*J. Gambl. Stud.*]{} 33, 397-424, DOI: 10.1007/s10899-016-9627-5 (2017). Kahneman, D. & Tversky, A. Prospect theory: An analysis of decision under risk. [*Econometrica*]{} 47, 263-291 (1979). Tversky, A. & Kahneman, D. Advances in prospect theory: Cumulative representation of uncertainty. [*J. Risk Uncertain.*]{} 5, 297-323, DOI: 10.1007/BF00122574 (1992). Barberis, N. A model of casino gambling. [*Manag. Sci.*]{} 58, 35-51, DOI: 10.1287/mnsc.1110.1435 (2012). Meng, J. [*Understanding Gambling Behavior and Risk Attitudes Using Massive Online Casino Data*]{}. Bachelor’s thesis, Dartmouth College, Hanover, New Hampshire, USA (2018). Wang, X. & Pleimling, M. Behavior analysis of virtual-item gambling. [*Phys. Rev. E*]{} 98, 012126, DOI: 10.1103/PhysRevE. 98.012126 (2018). Rhee, I. [*et al*]{}. On the levy-walk nature of human mobility. [*IEEE/ACM Transactions on Netw.*]{} 19, 630-643, DOI: 10.1109/TNET.2011.2120618 (2011). Brockmann, D. Anomalous diffusion and the structure of human transportation networks. [*Eur. Phys. J. Special Top.*]{} 157, 173-189, DOI: 10.1140/epjst/e2008-00640-0 (2008). Kim, S., Lee, C.-H. & Eun, D. Y. Superdiffusive behavior of mobile nodes and its impact on routing protocol performance. [*IEEE Transactions on Mob. Comput.*]{} 9, 288–304, DOI: 10.1109/TMC.2009.124 (2010). Wang, X. & Pleimling, M. Foraging patterns in online searches. [*Phys. Rev. E*]{} 95, 032145, DOI: 10.1103/PhysRevE.95. 032145 (2017). Toscani, G., Tosin, A. & Zanella, M. Multiple-interaction kinetic modelling of a virtual-item gambling economy. [*arXiv:1904.07660*]{} (2019). Holden, J. T. Trifling and gambling with virtual money. [*UCLA Entertain. Law Rev.*]{} 25, 41 (2018). Buhagiar, R., Cortis, D. & Newall, P. W. Why do some soccer bettors lose more money than others? [*J. Behav. Exp. Finance*]{} 18, 85-93 (2018). Rodríguez, P., Humphreys, B. R. & Simmons, R. [*The Economics of Sports Betting*]{} (Edward Elgar Publishing, 2017). Limpert, E., Stahel, W. A. & Abbt, M. Log-normal distributions across the sciences: Keys and clues. [*BioScience*]{} 51, 341-352, DOI: 10.1641/0006-3568(2001)051\[0341:LNDATS\]2.0.CO;2 (2001). Gibrat, R. Une loi des réparations économiques: l’effet proportionnel. [*Bull. de Stat. Gen. France*]{} 19, 419-513 (1930). Mitzenmacher, M. A brief history of generative models for power law and lognormal distributions. [*Internet Math.*]{} 1, 226-251 (2003). Kendall, M. & Gibbons, J. D. [*Rank Correlation Methods*]{} (Oxford University Press, 1990). Taylor, J. M. Kendall’s and Spearman’s correlation coefficients in the presence of a blocking variable. [*Biometrics*]{} 43 409-416 (1987). Dubins, L. E., Savage, L. J., Sudderth, W. & Gilat, D. [*How to Gamble If You Must: Inequalities for Stochastic Processes*]{}. Dover Books on Mathematics (Dover Publications, 2014). Epstein, R. A. [*The Theory of Gambling and Statistical Logic*]{} (Academic Press, 2012). Cherstvy, A. G., Chechkin, A. V. & Metzler, R. Anomalous diffusion and ergodicity breaking in heterogeneous diffusion processes. [*New J. Phys.*]{} 15, 083039, DOI: 10.1088/1367-2630/15/8/083039 (2013). Cherstvy, A. G., Chechkin, A. V. & Metzler, R. Particle invasion, survival, and non-ergodicity in 2d diffusion processes with space-dependent diffusivity. [*Soft Matter*]{} 10, 1591-1601, DOI: 10.1039/C3SM52846D (2014). Cherstvy, A. G., Thapa, S., Wagner, C. E. & Metzler, R. Non-gaussian, non-ergodic, and non-fickian diffusion of tracers in mucin hydrogels. [*Soft Matter*]{} 15, 2526-2551, DOI: 10.1039/C8SM02096E (2019). Inoue, J.-i. & Sazuka, N. Crossover between Lévy and Gaussian regimes in first-passage processes. [*Phys. Rev. E*]{} 76, 021111, DOI: 10.1103/PhysRevE.76.021111 (2007). Rahman, A. Correlations in the motion of atoms in liquid argon. [*Phys. Rev.*]{} 136, A405-A411, DOI: 10.1103/PhysRev.136. A405 (1964). Toyota, T., Head, D. A., Schmidt, C. F. & Mizuno, D. Non-gaussian athermal fluctuations in active gels. [*Soft Matter*]{} 7, 3234-3239, DOI: 10.1039/C0SM00925C (2011). Meiklejohn, S. [*et al*]{}. A fistful of bitcoins: Characterizing payments among men with no names. In [*Proceedings of the 2013 Conference on Internet Measurement Conference*]{}, 127-140, DOI: 10.1145/2504730.2504747 (ACM, New York, 2013). CSGOFAST. https://csgofast.com/. Accessed on April 20, 2018. CSGOSpeed. https://csgospeed.com/home. Accessed on April 20, 2018. ethCrash. https://www.ethcrash.io/play. Accessed on August 1, 2018. SatoshiDICE. https://www.satoshidice.com/. Accessed on August 1, 2018. Coinroll. https://coinroll.com/home. Accessed on August 1, 2018. Fiedler, I. Online gambling as a game changer to money laundering? (2013). Accessed on January 8, 2019. Coindesk. Bitcoin (usd) price. https://www.coindesk.com/price/bitcoin/. Accessed on August 1, 2018. CoinMetrics. Data downloads. https://coinmetrics.io/data-downloads/. Accessed on August 1, 2018. Martinelli, D. Skin gambling: Have we found the millennial goldmine or imminent trouble? [*Gaming Law Rev.*]{} 21, 557-565, DOI: 10.1089/glr2.2017.21814 (2017). Millar, S. I. Cryptocurrency expands online gambling. [*Gaming Law Rev.*]{} 22, 174-174, DOI: 10.1089/glr2.2018.2232 (2018). Kairouz, S., Paradis, C. & Nadeau, L. Are online gamblers more at risk than offline gamblers? [*Cyberpsychology. Behav. Soc. Netw.*]{} 15, 175-180, DOI: 10.1089/cyber.2011.0260 (2012). González-Roz, A., Fernández-Hermida, J. R., Weidberg, S., Martínez-Loredo, V. & Secades-Villa, R. Prevalence of problem gambling among adolescents: A comparison across modes of access, gambling activities, and levels of severity. [*J. Gambl. Stud.*]{} 33, 371-382, DOI: 10.1007/s10899-016-9652-4 (2017). Gainsbury, S. M. Online gambling addiction: the relationship between internet gambling and disordered gambling. [*Curr. Addict. Reports*]{} 2, 185-193, DOI: 10.1007/s40429-015-0057-8 (2015). Banks, J. [*Gambling, Crime and Society*]{} (Palgrave Macmillan, London, 2017). Redondo, I. Assessing the risks associated with online lottery and casino gambling: A comparative analysis of players’ individual characteristics and types of gambling. [*Int. J. Mental Heal. Addict.*]{} 13, 584-596, DOI: 10.1007/s11469-014-9531-0 (2015). Macey, J. & Hamari, J. esports, skins and loot boxes: Participants, practices and problematic behaviour associated with emergent forms of gambling. [*New Media & Soc.*]{} 21, 20-41, DOI: 10.1177/1461444818786216 (2019). Bauke, H. Parameter estimation for power-law distributions by maximum likelihood methods. [*Eur. Phys. J. B*]{} 58, 167 (2007). Burnham, K. P. & Anderson, D. R. [*Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach*]{} (Springer, New York, 2002). Clauset, A., Shalizi, C. & Newman, M. Power-law distributions in empirical data. [*SIAM Rev.*]{} 51, 661-703, DOI: 10.1137/070710111 (2009).
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by the US National Science Foundation through grant DMR-1606814.
Author contributions statement {#author-contributions-statement .unnumbered}
==============================
X.W. and M.P. conceived the study, X.W. wrote the computer codes and conducted the data analysis, X.W. and M.P. discussed the results and wrote the manuscript.
Additional information {#additional-information .unnumbered}
======================
**Competing Interests**\
The authors declare no competing interests.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'D. Gandolfi'
- 'G. Hébrard'
- 'R. Alonso'
- 'M. Deleuil'
- 'E.W. Guenther'
- 'M. Fridlund'
- 'M. Endl'
- 'P. Eigmüller'
- 'Sz. Csizmadia'
- 'M. Havel'
- 'S. Aigrain'
- 'M. Auvergne'
- 'A. Baglin'
- 'P. Barge'
- 'A. S. Bonomo'
- 'P. Bordé'
- 'F. Bouchy'
- 'H. Bruntt'
- 'J. Cabrera'
- 'S. Carpano'
- 'L. Carone'
- 'W. D. Cochran'
- 'H. J. Deeg'
- 'R. Dvorak'
- 'J. Eislöffel'
- 'A. Erikson'
- 'S. Ferraz-Mello'
- 'J.-C. Gazzano'
- 'N. P. Gibson'
- 'M. Gillon'
- 'P. Gondoin'
- 'T. Guillot'
- 'M. Hartmann'
- 'A. Hatzes'
- 'L. Jorda'
- 'P. Kabath'
- 'A. Léger'
- 'A. Llebaria'
- 'H. Lammer'
- 'P. J. MacQueen'
- 'M. Mayor'
- 'T. Mazeh'
- 'C. Moutou'
- 'M. Ollivier'
- 'M. Pätzold'
- 'F. Pepe'
- 'D. Queloz'
- 'H. Rauer'
- 'D. Rouan'
- 'B. Samuel'
- 'J. Schneider'
- 'B. Stecklum'
- 'B. Tingley'
- 'S. Udry'
- 'G. Wuchterl'
date: 'Received ; accepted '
title: |
Transiting exoplanets from the [*CoRoT*]{} space mission XIV.\
CoRoT-11b: a transiting massive “hot-Jupiter" in a\
prograde orbit around a rapidly rotating F-type star[^1]
---
Introduction
============
Discovering and studying extrasolar planets, and in general planetary systems other than ours, aims at understanding whether the solar system is peculiar and unique or usual and unremarkable. In this context, the discovery of a large population of Jupiter-like planets with a semi-major axis $\lesssim0.1$ AU (i.e., hot-Jupiters), as well as the detection of massive planets in very eccentric orbits [@Moutou09a; @OToole09] or even strongly misaligned with the stellar spin axis [@Hebrard08; @Pont09], has proven how the properties of extrasolar planets can be surprisingly different from those observed in the solar system’s planets.
From this point of view, studies of transiting extrasolar planets are cornerstones for understanding the nature of planets beyond the solar system, since a wealth of precious information can be gained [@Winn10]. Indeed, the peculiar geometry of transiting planets makes them very special targets for obtaining direct measurements of the planet-to-star radius [e.g., @Rosenblatt71; @Borucki84]. By combining time-series photometric observations acquired during the transit with radial velocity (RV) measurements of the host star, it is possible to derive the radius and mass of the planet, and therefore its mean density, once the mass and radius of the star has been determined. Transits also offer a unique opportunity to measure the sky-projected angle ($\lambda$) between the orbital angular momentum vector and the spin axis of the star [@Gaudi07]. This can be done by detecting the Rossiter-McLaughlin (RM) effect, i.e., the spectral distortion observed in the line profile of a rotating star as a second object passes in front of the stellar disc.
Furthermore transit surveys have the potential to enlarge the parameter space of planet host stars by detecting planets around stars that are usually not observed in radial velocity surveys. As is well known, Doppler surveys typically focus on slowly rotating solar-like stars because high RV precision can easily be achieved. They usually discard more massive main-sequence stars for which accurate RV measurements are rendered unfeasible by the rapid stellar rotation rate and the paucity of spectral lines. Indeed, a few RV searches have been conducted up to now around A- and F-type stars [e.g., @Lagrange09; @Guenther09; @Bowler10; @Johnson10]. Transit detections are not affected by the stellar rotation and can lead to the discovery of planets around rapidly rotating F-type stars.
Space missions like [*CoRoT*]{} [Co*nvection*, Ro*tation, and planetary* T*ransits*, @Baglin06; @Auvergne09] and *Kepler* [@Borucki10; @Koch10] are crucial to increase the number of planets with well-known orbital and physical parameters, and consequently improve the database that is needed to investigate all the aspects of the exoplanets population, down to the Earth-like mass-regime. The recent discoveries announced by the [*CoRoT*]{} exoplanet science team fully demonstrate the capability of the mission to determine the radius and the mean density of the transiting extrasolar planets, from the “transition desert regime” between brown dwarfs and planets [@Deleuil08], across hot and temperate Jupiter-like objects [@Barge08; @Alonso08; @Bouchy08; @Aigrain08; @Moutou08; @Rauer09; @Fridlund10; @Deeg10], and down to the Earth-like radius regime [@Leger09; @Queloz09].
In the present paper the *CoRoT Exoplanet Science Team* announces its eleventh transiting planet, namely , a fairly massive Jupiter-like planet in a relatively short-period orbit (about three days) around a rapidly rotating F6 dwarf star. By combining the high-precision photometric data from [*CoRoT*]{} with RV measurements and high signal-to-noise spectroscopy from the ground, we fully characterised the planet’s orbit and derived the main physical parameters of the planet-star system. Thanks to time-series RV measurements acquired during the transit we observed part of the RM effect and confirm the planetary transit event.
[*CoRoT*]{} observations, data reduction, and analysis {#CoRoT-Obs}
======================================================
Satellite observations {#Sec:Lightcurve}
----------------------
CoRoT-11b was discovered during the [*CoRoT*]{}’s second long run towards the Galactic centre direction, i.e., the [*CoRoT*]{} *LRc02* run.[^2] The observations lasted 145 days, from 2008 April 15 to September 7. The *LRc02* field is centred at $\alpha\approx18^h\,42^m$ and $\delta\approx6\degr\,39\arcmin$ (J2000), between the *Ophiuchus* and the *Serpens Cauda* constellations. The planet was detected transiting the [*CoRoT*]{} star[^3] with ID=0105833549. The main designations of the planet host star along with its equatorial coordinates and optical and near-infrared photometry, are reported in Table \[StarTable\] as retrieved from the *ExoDat* database [@Deleuil09] and *2MASS* catalogue [@Cutri03].
[lll]{}
\
[*CoRoT*]{} ID & 0105833549\
GSC2.3 ID & N1RO000587\
USNO-A2 ID & 0900-13499974\
2MASS ID & 18424494+0556156\
\
RA (J2000) & $18^h\,42^m\,44\fs95$\
Dec (J2000) & $05\degr\,56\arcmin\,16\farcs12$\
\
Filter & Mag & Error\
B & 13.596 & 0.024\
V & 12.939 & 0.019\
$r^\prime$ & 12.638 & 0.019\
$i^\prime$ & 12.283 & 0.053\
J & 11.589 & 0.021\
H & 11.416 & 0.029\
Ks & 11.248 & 0.022\
\[StarTable\]
The transiting planet was detected after 51 days of observations in the so-called [*CoRoT*]{} *alarm-mode* [@Quentin06; @Surace08]. This observing strategy consists in processing and analysing a first set of photometric data in order to single out planetary transits while the [*CoRoT*]{} run is still on-going. This enabled us to switch the time sampling of the light curve from 512 to 32 seconds and trigger the ground-based follow-up observations (see Sect. \[Follow-ups\]). Thanks to an objective prism in the optical path of the [*CoRoT*]{} exoplanet channel [@Auvergne09], CoRoT-11 was observed in three broad-band colours (red, green, and blue), according to the specific photometric mask selected at the beginning of the run. This usually allows us to remove false positives that mimick planetary transit events, such as stellar activity or eclipsing binaries. A total of 261917 photometric data-points were collected for each colour channel, 8349 of those were obtained with a time sampling of 512 sec, and 253568 with 32 sec. The transit signal was detected in all three colour channels with the similar depth and the same duration and ephemeris as expected for a *bona-fide* planetary transit.
At the end of the *LRc02* observing run, the whole photometric data-set of CoRoT-11 was processed using the [*CoRoT*]{} reduction and calibration package described in @Auvergne09. The pipeline also flags bad photometric data-points collected during the entrance into and exit from the Earth’s shadow or data-points that are strongly affected by hits of high-energy particles resulting from the crossing of the South Atlantic Anomaly (SAA).
[*CoRoT*]{} light curve and transit fit {#Transit_Fit}
---------------------------------------
In order to increase the signal-to-noise (S/N) ratio, the analysis of the photometric data was performed using the white light curve, as derived by adding together the signal from the three colour channels. CoRoT-11 has a close neighbour star at about $2\arcsec$ northwest, falling well inside the [*CoRoT*]{} photometric aperture, as well as a handful of faint nearby stars that are spatially located around the [*CoRoT*]{} mask (see also Sect. \[Phot-FU\] and Fig. \[CoRoT11b\_imagete\]). According to the $ExoDat$ database, the closest neighbour star is 2.1 magnitudes fainter than CoRoT-11 in the $r^\prime$-band. Following the method described in previous [*CoRoT*]{} papers [e.g., @Alonso08], we estimated that it contributes about $12.2$ % of the total flux of the photometric mask of CoRoT-11. Taking also into account the light coming from the fainter neighbours, the total contamination fraction rises up to $13.0\pm1.5$ %. This fraction was removed from the whole data-set prior to analysing the white light curve.
A first cleaning of the raw [*CoRoT*]{} data was performed applying an iterative sigma-clipping algorithm. Most of the photometric points rejected according to this criterion (about 7.8 %) resulted from the crossing of the SAA and matched the photometric data previously flagged as outlier by the [*CoRoT*]{} automatic pipeline [@Auvergne09]. The cleaned white light curve is plotted in Fig. \[Lightcurve\]. It shows 49 transits with a depth of $\sim1$ %, occurring about every three days. The transits are clearly visible, as shown in the inset of Fig. \[Lightcurve\]. Out of 49 transits, 17 were detected with a 512 seconds time sampling, whereas 32 were observed with the oversampled rate (32 seconds). The light curve is moderately affected by sudden “jumps” in intensity resulting from the impact of high-energy particles onto the [*CoRoT*]{} CCDs [@Pinheiro08]. The high-frequency scatter of the light curve (Fig. \[Lightcurve\]) is compatible with other sources of similar brightness observed by [*CoRoT*]{} @Aigrain09.
The [*CoRoT*]{} light curves are affected by a modulation of the satellite orbital period, which changes its shape and amplitude during a long run observation. Following the prescription of @Alonso08, the orbital signal of each $j^{th}$ orbit was corrected with the signals from the previous and the following 30 orbits. The data points acquired during the transits were not considered in the estimate of the mean orbital signal of the $j^{th}$ orbit.
In order to determine the period $P$ and transit epoch $T_\mathrm{tr}$, we first used an approximate ephemeris to build a phase-folded curve of the transit. A simple trapezoidal model was fitted to this curve to get the parameters of the average transit, i.e., depth, duration, centre, and ingress/egress time. This model was then fitted to each individual transit, leaving only the centre as free parameter. A linear fit to the final observed-calculated ($O-C$) diagram of the transit centres served to refine the ephemeris, and we iterated the process until the fitted line had no significant slope. Once the orbital period and transit epoch were derived (Table \[Par\_table\]), we constructed a combined phase-folded transit curve to this ephemeris by successively normalising each transit to the regions surrounding it. In this case, because the star is not very active (see Fig. \[Lightcurve\]), we used simple line fits to the phases between $[\,-0.06,-0.02\,]$, and $[\,0.02,0.06\,]$, using the ephemeris of Table \[Par\_table\], and applied this normalisation to the whole section from phases $[\,-0.06,0.06\,]$. We used a Savitzky-Golay filter to recognize a few remaining outliers before binning in phase. Taking into account the photometry rejected according to the first sigma-clipping algorithm, this process removed about $1.6$ % of the remaining data points, leading to a final duty cycle of $\sim90$ %.
Finally, the data-points were binned in blocks of 0.0005 in phase, and the error bars were estimated as the standard deviation of the data points inside each bin divided by the square root of the number of points inside the bin. The phase-folded curve of the transit is shown in Fig. \[Transit\].
The transit was fitted to a model using the formalism of @Gimenez06. To find the solution that best matches our data, we minimized the using the algorithm AMOEBA [@Press92]. The fitted parameters were the centre of the transit, the phase of start of the transit $\theta_1$, the planet-to-star radius ratio $k=R_\mathrm{p}/R_{*}$, the orbital inclination $i$ and the two non-linear limb darkening coefficients $u_{+}=u_a+u_b$ and $u_{-}=u_a-u_b$. We used a quadratic law for the limb darkening, given by $I(\mu)=I(1)[1-u_a(1-\mu)-u_b(1-\mu)^2]$, where I is the distribution of brightness over the star and $\mu$ is the cosine of the angle between the normal to the local stellar surface and the line of sight. The use of $u_{+}$ and $u_{-}$ is a better choice to avoid correlations between the two limb darkening coefficients $u_a$ and $u_b$, as described in @Gimenez06. To estimate the errors in each of the parameters we performed the minimization to five hundred different sets of data. These data-sets were constructed with different values for the contamination factor (estimated at $13.0\pm1.5$ %), and different starting values to the AMOEBA minimization. To build each set, we first subtracted the best solution to the data, and then shifted circularly the residuals by a random quantity to keep the low-frequency content of the noise. The best-fit solution was finally added to these new residuals. For each of the fitted parameters in all data-sets, we calculated the standard deviation as an estimate of the errors. In order to take into account the effect of the non-Gaussian distributions of the parameters, we also forced a Gaussian fit to the measured distributions. The adopted (conservatively larger) error bar was the biggest error among the standard deviation of the fitted parameters and the $\sigma$ of the fitted Gaussian distribution. The parameters and associated errors are listed in Table \[Par\_table\], along with the scaled semi-major axis $a/R_{*}$, as derived using Eq. 12 in @Gimenez06. Assuming a circular orbit (i.e., $e=0$) and combining the scaled semi-major axis $a/R_{*}$ with the orbital period $P$ via the Kepler’s laws, we derived the parameter $M^{1/3}_{*}/R_{*}=0.787\pm0.010$ in solar units, which leads to a mean stellar density $\rho_{*}=0.62\pm0.02$ g/cm$^3$ [see @Seager03; @Winn10 for the relevant formulas]. The transit fit yielded the limb darkening coefficients $u_+=0.61\pm0.06$ and $u_-=0.02\pm0.04$ (Table \[Par\_table\]), in very good agreement with the theoretical values $u_+=0.64\pm0.01$ and $u_-=0.06\pm0.01$ predicted by @Sing10 for a star with the same fundamental parameters as CoRoT-11 (Sect. \[HR-spec\]).
Note that the standard deviation of the residuals outside the transit phase is 230 ppm, which is within the uncertainties identical to the mean error bar of each of the bins (239 ppm), thus revealing the small low-frequency noise in the phase-folded light curve after the analysis described above.
Planetary eclipse upper limits
------------------------------
We searched for the eclipse of the planet in the [*CoRoT*]{} light curve with the same techniques as described in @Alonso09a [@Alonso09b] and @Fridlund10. To account for a possible eccentric orbit, we mapped the levels of a fit to a trapezoid (with the shape parameters estimated from the transit) for different orbital phases. We did not obtain any significant detection above 100 ppm in depth. We can thus set an upper 3$\sigma$ limit for the planetary eclipse depth of 100 ppm, which we translated into an upper limit to the brightness temperature of 2650 K [@Alonso09b].
Ground-based observations, data reduction, and analysis {#Follow-ups}
=======================================================
As already described in previous [*CoRoT*]{} discovery papers, intensive ground-based observations are mandatory to establish the planetary nature of the transiting candidates detected by [*CoRoT*]{}. These follow-ups are crucial to rule out possible false positives, i.e., physical configurations that mimick planetary transits, which cannot be excluded on the basis of meticulous light curve analyses only. Out of about $50$ promising candidates detected per [*CoRoT*]{} long run field [see @Carpano09; @Moutou09b; @Cabrera09; @Carone10], usually only a handful turns out to be *bona fide* planetary objects. Furthermore, ground-based observations are needed to assess the planetary nature of the transiting object, derive the true mass of the planet, and measure the stellar parameters of the host star needed to accurately compute the planet radius[^4].
Follow-up campaigns of the planetary transit candidates detected by the *alarm-mode* in the *LRc02* field started in early Summer 2008. In the following subsection we will describe the complementary photometric and spectroscopic ground-based observations of CoRoT-11.
Photometric follow-ups {#Phot-FU}
----------------------
The objective prism placed along the optical path of the [*CoRoT*]{} exoplanet channel spreads the light of the observed targets over about 50 pixels on the CCDs, corresponding to a projected sky area of about $20\arcsec\times10\arcsec$. The total flux of each [*CoRoT*]{} target is then computed by integrating the pixel signal over a preselected photometric mask elongated along the dispersion direction and covering most of the point spread function (PSF). As already described in Sect. \[Transit\_Fit\], this increases the possibility that the light of neighbour stars could contaminate the flux of the [*CoRoT*]{} target. Furthermore, what is believed to be a “good” planetary transit might actually turn out to be the eclipse of a faint nearby binary system, whose light is diluted by the [*CoRoT*]{} target star.
To reproduce the observed $\sim$$1.1$% deep transit, a generic contaminant star cannot be more than $\sim$$5$ mag fainter than CoRoT-11. As already mentioned above, CoRoT-11 has a close, 2.1 mag fainter neighbour star that might thus actually be the [*potential*]{} source of false alarm (Fig. \[CoRoT11b\_imagete\]). In order to exclude this scenario, we took advantage of the [*CoRoT*]{} ephemeris to perform the so-called “on-off” photometry. In this procedure, candidates are photometrically observed with ground-based facilities at higher spatial resolution than [*CoRoT*]{}during the transit (on-observation) and out of the transit (off-observation). The brightness of the candidates, as well as that of any nearby stars, is then monitored to unveil any potential background eclipsing binary. Full details of this method are described in @Deeg09.
According to this observing strategy, $R$-band photometry was carried out on 2008 July 4, using the CCD camera mounted on the Swiss Leonard Euler 1.2m telescope at La Silla Observatory (Chile), under photometric sky condition. Four sets of five consecutive exposures of 45 seconds each were acquired. The first set was obtained during the predicted transit (on-observation), whereas the remaining three were taken out of transit (off-observations). The data were reduced with standard IRAF routines[^5]; aperture photometry was performed with the DAOPHOT package under the IRAF environment. Differential photometry was obtained for CoRoT-11, as well as for the nearby contaminants, using a set of nearby comparison stars. The on-off Euler observations clearly show that CoRoT-11 is the source of the transit events detected by [*CoRoT*]{} (Fig. \[EulerCam\_LC\]). The first photometric data-set shows a dimming of the light of CoRoT-11 at the expected time and with roughly the same depth at mid-transit. The 2.1 mag fainter star located at about $2\arcsec$ northwest of the target should undergo eclipses with depths of about $0.1$ mag, something which the ground-photometry clearly excluded. By centring and co-adding the best-seeing Euler images, we excluded the presence of a third nearby object down to $R\approx17.5$ mag and up to $1.5\arcsec$ from CoRoT-11.
More Johnson $R$-band photometric observations of CoRoT-11 were carried out using the 30cm TEST telescope at the Thüringer Landessternwarte (TLS), Tautenburg (Germany), on 2008 September 7. Full details of the instrument, observing strategy, and data reduction can be found in @Eisloffel07 and @Eigmuller09. Although these observations were performed at higher time-sampling than those at Euler, they were affected by poor weather conditions, especially in the second half of the night. Nevertheless, we succeeded to observe the transit ingress of CoRoT-11b at the expected [*CoRoT*]{} ephemeris and exclude significant photometric variations in the nearby contaminant stars.
Reconnaissance low-resolution spectroscopy {#LowRes_FU}
------------------------------------------
Low-resolution reconnaissance spectroscopy of the planet host star was performed with the long-slit spectrograph mounted at the Nasmyth focus of the 2m Alfred Jensch telescope of the TLS observatory, Tautenburg, Germany. These observations were part of an intensive programme devoted to the spectroscopic “snap-shot” of the planetary candidates detected by the *alarm-mode* in the *LRc02* [*CoRoT*]{} field. They were useful to quickly classify the stars and derive a first estimate of their photospheric parameters. Furthermore, they allowed us to identify and remove giant stars, for which the transiting object would result in a low-mass stellar companion, as well as B-type objects and rapidly rotating early-type stars, for which high-precision radial velocity measurements cannot be achieved.
CoRoT-11 was observed on 2008 August 8, under clear and stable weather conditions. Three consecutive exposures of 20 minutes each were acquired and subsequently combined to remove cosmic ray hits and improve the S/N ratio. The data reduction was performed with a semi-automatic pipeline developed under the IDL[^6] environment. Relative flux calibration was performed observing two spectro-photometric standard stars. The final extracted and co-added spectrum covers the wavelength range 4950–7320 Å, with a resolving power $R\approx2100$ and an average $S/N\approx60$. The spectral type and the luminosity class of CoRoT-11 was derived by fitting the observed spectrum with a grid of suitable template spectra, as described in @Frasca03 and @Gandolfi08 and shown in Fig. \[Fig\_TLS\]. We found that CoRoT-11 is an F6V star, with an accuracy of about $\pm\,1$ sub-class (Table \[Par\_table\]).
Radial velocity observations {#RV_FU}
----------------------------
The RV follow-up of the host star CoRoT-11 was started in summer 2008 by acquiring two high-resolution spectra with the SOPHIE spectrograph [@Bouchy09] attached to the 1.93m telescope of the Haute-Provence Observatory (France). The instrument was set in its high efficiency (HE), leading to a resolving power of $R\approx40\,000$. These observations revealed a rapidly rotating star with relatively broad spectral lines, corresponding to a projected rotational velocity () of $\sim40$ km/s. According to the [*CoRoT*]{} ephemeris, the SOPHIE spectra were secured around the extreme orbital phases (i.e., phase 0.25 and 0.75), and showed a RV variation of $\sim450$ m/s in phase with the [*CoRoT*]{} ephemeris. Because the F6V spectral type of the host star translates into a stellar mass of about 1.3 , the measured RV variation is compatible with a $\sim2$ Jupiter planet. Twelve additional RV measurements were obtained in summer 2008 and 2009 using the HARPS spectrograph [@Mayor03] mounted at the 3.6m ESO telescope on La Silla (Chile). The spectra were acquired at different orbital phases, under good weather conditions and without strong moonlight contamination. The data were acquired setting the spectrograph both in the EGGS and HARPS standard modes, yielding a resolving power of $R\approx70\,000$ and $R\approx115\,000$, respectively.
The extraction of both the SOPHIE and HARPS spectra was performed using the respective pipelines. Following the techniques described by @Baranne96 and @Pepe02, the radial velocities were measured from a weighted cross-correlation of the spectra with a numerical mask. We used a standard G2 mask that includes more than 3500 lines. Cross-correlations with F0 and K5 masks gave similar results. One SOPHIE spectrum and three HARPS spectra were corrected for small moonlight contamination following the method described in @Pollacco08 and @Hebrard08, which uses a reference background sky spectrum obtained through a second fiber spatially located near the target. This led to a radial velocity correction of $700\pm80$ m/s and below $450\pm50$ m/s for the SOPHIE and HARPS spectra, respectively.
[crrrrc]{} HJD & RV &$\sigma_{RV}$ & Bisector & Texp & S/N per resolution\
(days) & (km/s) & (km/s) & (km/s) & (sec) & element at 5500 Å\
& & & km/s\
2454643.60252 & -1.164 & 0.218 & & 1402 & 42\
2454683.42036 & -0.714 & 0.206 & & 1607 & 50\
& & & km/s\
2454731.52836 & -0.932 & 0.067 & -0.531 & 2700 & 88\
2454742.51115 & -1.364 & 0.104 & -0.149 & 2700 & 54\
2454745.51100 & -1.370 & 0.087 & -0.476 & 2700 & 66\
2454746.51455 & -0.986 & 0.071 & -0.583 & 2700 & 81\
2454747.51953 & -1.509 & 0.115 & -0.892 & 1800 & 48\
& & & km/s\
2455023.66206 & -1.663 & 0.149 & -0.286 & 3600 & 34\
2455024.63997 & -1.405 & 0.111 & -0.705 & 3600 & 48\
2455045.69605 & -1.013 & 0.114 & -0.747 & 3600 & 52\
2455064.60739 & -1.166 & 0.083 & -0.408 & 3300 & 64\
2455067.51277 & -1.085 & 0.092 & -0.204 & 3300 & 59\
2455068.49041 & -1.547 & 0.130 & 0.472 & 3300 & 42\
2455069.51299 & -1.384 & 0.110 & -0.497 & 3300 & 48\
& & & km/s\
2455035.46055 & -0.951 & 0.174 & & $2\times1800$ & 25\
2455057.40390 & -0.541 & 0.175 & & $2\times1800$ & 25\
& & & km/s\
2455012.80417 & -0.033 & 0.050 & & 1200 & 61\
2455013.06216 & 0.181 & 0.086 & & 1200 & 71\
2455013.78975 & -0.165 & 0.089 & & 900 & 54\
2455013.80073 & -0.083 & 0.090 & & 900 & 56\
2455013.81181 & -0.170 & 0.135 & & 900 & 56\
2455013.82275 & 0.173 & 0.111 & & 900 & 57\
2455013.83360 & -0.012 & 0.142 & & 900 & 57\
2455013.84473 & 0.099 & 0.073 & & 900 & 56\
2455013.85572 & 0.037 & 0.119 & & 900 & 56\
2455013.86677 & 0.274 & 0.108 & & 900 & 55\
2455013.87785 & 0.240 & 0.078 & & 900 & 55\
2455013.88881 & 0.017 & 0.100 & & 900 & 56\
2455013.89978 & 0.089 & 0.080 & & 900 & 56\
2455013.91097 & -0.263 & 0.086 & & 900 & 54\
2455014.05389 & -0.386 & 0.119 & & 900 & 58\
Two complementary RV measurements were acquired using the echelle spectrograph mounted on the 2m Alfred Jensch telescope of the TLS observatory, Tautenburg (Germany), in July and August 2009. For each observing night, two consecutive exposures of 30 minutes each were recorded to increase the S/N ratio and remove cosmic ray hits. The adopted instrument set-up yielded a spectral resolution of about $R\approx30\,000$. The data were reduced using standard IRAF routines. The wavelength solution was improved acquiring ThAr spectra immediately before and after each stellar observation. After accounting for instrumental shifts with telluric lines, the radial velocities were measured cross-correlating the target spectra with a spectrum of the RV standard star observed with the same instrument set-up.
As part of NASA’s key science programme in support of the [*CoRoT*]{} mission, more RV measurements were obtained with the HIRES spectrograph [@Vogt94] mounted on the Keck I 10m telescope, at the Keck Observatory (Mauna Kea, Hawai’i). With the aim of detecting the Rossiter-McLaughlin effect, 13 RV measurements were secured during the expected transit on 2009 July 1 (UT). The observations were performed with the red cross-disperser and the $I_2$ absorption cell to correct for instrumental shifts of the spectrograph. The $0\farcs861$ wide slit together with the 14$\arcsec$ tall decker was employed to allow proper background subtraction, yielding a resolving power of $R\approx50\,000$. In order to adequately sample the RM anomaly, the exposure time was set to 900 seconds. Two extra spectra of 1200 seconds each were also acquired out of transit, on 2009 June 30 (UT). The spectra were reduced with IRAF standard routines. The HIRES RV measurements were derived with the iodine data modelling code “Austral" [@Endl00].
The final RV measurements of CoRoT-11 are reported in Table \[Table\_rv\], along with error bars, exposure times, and S/N ratio. In spite of the good RV stability of the spectrographs used in the present work, the relatively high of CoRoT-11 strongly affected the RV precision of our measurements and led to an accuracy in the range $\sim100-200$ m/s, with a typical error bar of about $100$ m/s even for the HARPS and HIRES data. The five data-sets, i.e., the SOPHIE, HIRES, and TLS data, and the two HARPS modes, were simultaneously fitted with a Keplerian model, assuming a circular orbit. The HIRES RV measurements acquired during the transit were not used in the fit. Both period and transit central time were fixed according to the [*CoRoT*]{} ephemeris. An RV shift was let free to vary in the fit between the five data sets.
The RV measurements are plotted in Fig. \[Fig\_omc\] together with the best-fitting circular orbit. The derived orbital parameters are reported in Table \[Par\_table\], along with error bars that were computed from variations and Monte Carlo experiments. The RV measurements led to a semi-amplitude $K=280\pm40$ m/s. The standard deviation of the residuals to the fit is $\sigma_{O-C}=88$ m/s, in agreement with the expected accuracy of the RV measurements. The reduced is 1.1 for the 22 RV measurements used in the fit.
We explored the possibility that the observed RV variations of CoRoT-11 do not result from the planet’s orbital motion, but are instead caused by a periodic distortions in the spectral lines caused by either stellar magnetic activity or the presence of a hypothetical unresolved eclipsing binary, whose light is diluted by CoRoT-11. In order to exclude these scenarios, we performed an analysis of the cross-correlation function (CCF) profile. Using the highest resolution spectra in our data-set (i.e., the HARPS measurements) and following the line-bisector technique described in @Queloz01, we derived the difference in velocity space between the lower and upper part of the HARPS CCFs (i.e., bisector span). The value of the bisector span velocities are listed in Table \[Table\_rv\]. The uncertainty was set to twice that of the corresponding HARPS radial velocity measurements. We found that the CCFs show a systematic asymmetric profile, translating into a negative value of the bisector span velocities (Table \[Table\_rv\] and Fig. \[Bisector\]), which is usually observed in rapidly rotating F-type stars [@Gray86; @Gray89]. Nevertheless, the CCF bisector spans show neither significant variations nor any trend as a function of both RV measurements and orbital phases (Fig. \[Bisector\]). The linear Pearson correlation coefficient between the HARPS RV measurements and the corresponding CCF bisector spans is $-0.25$. Removing the only outlier point with positive bisector span (i.e., $0.472$ km/s), the correlation coefficient approaches zero, being $-0.03$. Thus the observed RV variations seem not to be caused by spectral line profile variations to any significant degree, but are mainly due to the Doppler shift induced by the orbital motion of CoRoT-11b. The RV observations and the transit-signal detected by [*CoRoT*]{} point to a hot-Jupiter-sized planet that orbits the star.
![Bisector spans versus radial velocity measurements (top panel) and orbital phases (bottom panel) as derived from the HARPS data (blue downward and upward triangles for EGGS and HARPS modes, respectively). The horizontal dotted line marks the average negative value of the CCF bisector span, i.e. $-0.417$ km/s.[]{data-label="Bisector"}](15132fg7.eps "fig:") ![Bisector spans versus radial velocity measurements (top panel) and orbital phases (bottom panel) as derived from the HARPS data (blue downward and upward triangles for EGGS and HARPS modes, respectively). The horizontal dotted line marks the average negative value of the CCF bisector span, i.e. $-0.417$ km/s.[]{data-label="Bisector"}](15132fg8.eps "fig:")
The phase-folded RV measurements are plotted in Fig. \[Fig\_orb\_phas\]. As already described, the orbit was assumed to be circular, which is a reasonable assumption for close-in hot-Jupiters. The radial velocities are not accurate enough to constrain the eccentricity with the orbital fit only. Indeed, a Keplerian fit with an eccentricity of about 0.6 provides a solution that agrees with the [*CoRoT*]{} ephemeris, with a RV semi-amplitude $K$ which is 15 % larger than the one obtained for a circular orbit. The standard deviation of the residuals to this eccentric fit ($\sigma_{O-C}=95$ m/s) is marginally higher than the circular fit. Only extremely eccentric orbits with $e>0.7$ produce low-quality fits, with dispersions larger than 120 m/s. Nevertheless, we put some constraints on the possible value for the planet eccentricity, taking advantage of the transit fit parameter $M^{1/3}_{*}/R_{*}$, from which the mean stellar density can be inferred (Sect. \[Transit\_Fit\]). The obtained value of $M^{1/3}_{*}/R_{*}$ depends on the eccentricity of the orbit. We found that for $e\gtrsim0.2$ the mean stellar density would be incompatible with a F6 dwarf star, because it is significantly higher than the expected value [@Cox00].
The on-transit RV data of CoRoT-11 are plotted in Fig. \[Fig\_RM\]. Unfortunately, the HIRES observations were performed according to an old, slightly incorrect, transit ephemeris based on the analysis of the *alarm-mode* data only. Nevertheless, although the HIRES measurements cover only the first half of the transit, they clearly show that the Rossiter-McLaughlin (RM) anomaly has been detected, which also confirms the occurrence of the transit events. The RM amplitude is large ($\sim400$ m/s), because of the fast stellar rotation. This also proves that the transiting object has a planetary size. The first part of the spectroscopic transit shows radial velocities that are blue-shifted compared to the Keplerian fit, which clearly indicates that the orbit is prograde. In addition, systematics seem to be present in the data at a level above the expected uncertainties for some measurements. It is thus difficult to constrain the spin-orbit angle with the current data. In Fig. \[Fig\_RM\] we show a model with $\lambda=0\degr$ and $\vsini=40$ km/s, using the analytical approach developed by @Ohta05. The fit is not satisfying, suggesting in particular a value higher than the one we derived from the SOPHIE RV data and the spectral analysis (Sect. \[HR-spec\]), in order to have a larger amplitude for the anomaly. It is known that a discrepancy could be found between the values measured from the RM effect and from the spectral modelling of line broadening, especially for fast rotators [see e.g., @Simpson10]. Concerning the spin-orbit angle, the data are compatible with $\lambda = 0\degr$. However, as for , the moderate quality of the data-set prohibits accurate measurements. Additional RM observations of CoRoT-11 should be performed, with a full coverage of the event.
High-resolution spectroscopy and stellar analysis {#HR-spec}
-------------------------------------------------
To derive the fundamental atmospheric parameters of the planet host star, we observed CoRoT-11 with the high-resolution spectrograph UVES mounted at the 8.2m Very Large Telescope (ESO-VLT; Paranal Observatory, Chile). Two consecutive spectra of 2380 seconds each were acquired in service mode on 2008 September 17, under the ESO programme 081.C-0413(C). The star was observed through a $0\farcs5$ wide slit, setting the UVES spectrograph to its Dic-1 mode (346+580). The adopted configuration yielded a resolving power of about $65\,000$, with a spectral coverage $\lambda\approx3000-6800$ Å. The spectra were extracted and combined with standard IRAF packages, giving a final S/N ratio of about 160 at 5500 Å.
The effective temperature ($T_\mathrm{eff}$), surface gravity (log$g$), metallicity ($[\rm{Fe/H}]$), and projected rotational velocity () of CoRoT-11 were derived following the procedure already adopted for other [*CoRoT*]{} host stars [e.g., @Deleuil08; @Fridlund10; @Bruntt10]. We took advantage of different spectral analysis packages applied independently by different teams within the [*CoRoT*]{} community, e.g., the SME 2.1 [@Valenti96; @Valenti05], the VWA [@Bruntt04; @Bruntt08; @Bruntt10] software. We found that the estimated values of the above mentioned physical parameters agree within the error bars. The final adopted values are $T_\mathrm{eff}=
6440\pm120$ K, log$g=4.22\pm0.23$, $[\rm{Fe/H}]=-0.03\pm0.08$, and $\vsini=40\pm5$ km/s (Table \[Par\_table\]), with the latter value in perfect agreement with the one derived from the RV data (Sect. \[RV\_FU\]).
We also used the VWA software package to perform a detailed abundance analysis of the UVES spectrum of CoRoT-11, by iteratively fitting reasonably isolated spectral lines. Atmosphere models were interpolated in a grid of MARCS models [@Gustafsson08] and atomic data were extracted from VALD [@Kupka99]. However, owing to the relatively high of the star, only 71 lines turned out to be sufficiently isolated and thus suitable for spectral analysis. A small section of the observed and fitted synthetic spectra is shown in Fig. \[Fig\_VWA\]. Abundances were computed relative to the Sun to correct the oscillator strengths [see @Bruntt08; @Bruntt09]. We determined the atmospheric parameters by adjusting them to minimise the correlation of iron (Fe) with equivalent width (EW) and excitation potential (EP). Furthermore, we required that Fe[i]{} and Fe[ii]{} have the same mean abundance within the uncertainty. To evaluate the uncertainty on the atmospheric parameters, we perturbed them to determine when the correlations of Fe[i]{} with EW or EP become significant or the Fe[i]{} and Fe[ii]{} abundances deviate by more than $1\,\sigma$ . In Table \[tab:ab\] we list the abundances relative to the Sun for the five elements Na, Si, Ca, Fe, and Ni.
Results {#Res}
=======
Stellar parameters of the parent star CoRoT-11 {#Star_Param}
----------------------------------------------
To determine the mass and radius of the CoRoT-11 host star we took advantage of the stellar parameter ($M^{1/3}_{*}/R_{*}$) as derived from the [*CoRoT*]{} light curve analysis (Sect. \[Transit\_Fit\]), and of the effective temperature and metallicity ($T_\mathrm{eff}$ and $[\rm{Fe/H}]$) as obtained from the spectral analysis (Sect. \[HR-spec\]). We thus compare the location of the star on a $log(M^{1/3}_{*}/R_{*})\,vs.\,log(T_\mathrm{eff})$ H-R diagram with evolutionary tracks computed with the CESAM code [@Morel08]. According to these theoretical models we obtained a stellar mass of $M_{\star}=1.27\pm0.05$ and a stellar radius of $R_{\star}=1.37\pm0.03$ , with an age of about $2.0\pm1.0$ Gyr (Table \[Par\_table\]). From these results we derived a surface gravity of log$g=4.26\pm0.06$, in good agreement with the spectroscopically determined value of log$g=4.22\pm0.23$. We also checked whether the high rotation rate of the star can account for a significant flattening at the poles. According to the equation by @Claret00 and assuming that the star is seen almost edge-on, the equatorial and polar radii should differ by only $\sim0.6$ %, making the flattening effect negligible.
[lrr]{} Element & $[A/H]$ & $N$\
& $ -0.10\pm0.15 $ & 2\
[Si i]{} & $ 0.10\pm0.09 $ & 5\
[Si ii]{} & $ 0.19\pm0.18 $ & 2\
[Ca i]{} & $ 0.03\pm0.12 $ & 4\
[Fe i]{} & $ -0.04\pm0.08 $ & 44\
[Fe ii]{} & $ -0.02\pm0.08 $ & 6\
[Ni i]{} & $ -0.08\pm0.11 $ & 8\
We computed the interstellar extinction to the star following the general guidelines described in @Gandolfi08. The seven $BVr^\prime i^\prime JHKs$ broad-band magnitudes as retrieved from the *ExoDat* database enabled us to construct the spectral energy distribution (SED) of CoRoT-11, covering a wide spectral range, from optical to near-infrared wavelengths (see Table \[StarTable\]). Simultaneously using all the photospheric colours encompassed by the SED, we derived the interstellar extinction to the star ($A_{\mathrm V}$) by fitting the observed SED with a theoretical one progressively reddened with an increasing value of $A_{\mathrm V}$. The theoretical SEDs were computed with the *NextGen* stellar atmosphere model [@Hauschildt99] with the same $T_\mathrm{eff}$, log$g$, and $[\rm{Fe/H}]$ as the star, the response curve of the $ExoDat$ photometric system, and the extinction law by @Cardelli89. Assuming a total-to-selective extinction $R_{\mathrm V}=A_{\mathrm V}/E_{\mathrm {B-V}}=3.1$ (typical of the diffuse interstellar medium in our Galaxy), as well as a black body emission at the star’s effective temperature and radius, we derived an extinction $A_{\mathrm V}=0.55\pm0.10$ mag and a distance to the star $d=560\pm30$ pc (Table \[Par\_table\]).
We attempted to derive the rotation period of the star from the [*CoRoT*]{} light curve. The Lomb-Scargle periodogram [@Scargle82] applied to the out-of transit data-points shows only one significant broad peak at about $8.5\pm2$ days, with a light curve amplitude of $\sim0.25$ %. This period is not compatible with the maximum rotation period of $1.73\pm0.22$ days derived from the projected rotational velocity and radius of the planet host star. The detected signal at $8.5\pm2$ days might be due to a $0.02$ mag periodic variation of the nearby contaminant star located at about $2\arcsec$ from CoRoT-11 (see Sect. \[Transit\_Fit\] and \[Phot-FU\]). On the other hand, a low-level of magnetic activity of CoRoT-11 might account for no significant signals at $\lesssim2.0$ days, as confirmed from the absence of emission features in the core of the Ca[ii]{} H & K and Balmer lines.
Planetary parameters of CoRoT-11b {#Planet_Param}
---------------------------------
Based on the stellar mass and radius (Sect. \[Star\_Param\]), the RV curve semi-amplitude (Sect. \[RV\_FU\]), the planet-to-star radius ratio, and the planet orbit inclination (Sect. \[Transit\_Fit\]), we derived a mass for CoRoT-11b of $=2.33\pm0.34$ and a radius of $=1.43\pm0.03$ , yielding a mean planetary density $\rho_\mathrm{p}=0.99\pm0.15$ g/cm$^3$. The planet orbits its host star at a distance of $a=0.0436\pm0.005$ AU in $2.994339\pm0.000011$ days. The planetary mass has been obtained assuming an eccentricity $e=0$. According to the results presented in Sect. \[RV\_FU\], we cannot exclude a slightly eccentric orbit, with $0\lesssim e \lesssim 0.2$. Nevertheless, for $e=0.2$ the planetary mass would decrease of about 4 %, i.e., well within the error bar of our estimation. A summary of the planetary parameters derived in the present work is reported in Table \[Par\_table\].
Discussion {#Disc}
==========
Together with [@Guenther09], [@Snellen09], and [@CollierCameron10] orbiting a F6V ($\vsini=39$ km/s), F3V ($\vsini=39.3$ km/s), and A5V star ($\vsini=90\pm10$ km/s), respectively, CoRoT-11b is the fourth extrasolar planet discovered around a rapidly rotating main sequence star ($\vsini=40\pm5$ km/s; see Fig. \[FIG\_VsiniHist\]). Furthermore, the planet host star CoRoT-11, with its effective temperature of $T_\mathrm{eff}=6440\pm120$ K, is one of the hottest stars known to harbour an extrasolar planet.
Most of the bulk of known extrasolar planets have been detected with the RV method. Although this technique has dramatically increased the number of discoveries in the last fifteen years, it suffers from a strong selection bias, because it is mostly restricted to planets around slowly rotating stars ($\vsini\lesssim10$ km/s). This observational bias limits our knowledge of extrasolar planets to mainly late-type solar-like stars. One of the big advantages of the transit method is that it is insensitive to the stellar rotation, enabling us to single out planets even around intermediate-mass stars [@CollierCameron10]. This allows us to enlarge the parameter space of planet host stars and gives us a chance to study the planet formation around A and F stars. Even if CoRoT-11b has been confirmed and studied thanks to a complementary and intensive RV campaign, it would have likely been rejected from any RV search sample because of the fast rotation of its parent star. Indeed, about 20 RV measurements were needed to assess the planetary nature of CoRoT-11b and constrain its mass within $\sim15$ %. But the RV signature of CoRoT-11b has been detected because of its high mass. Taking into account the accuracy of our RV measurements ($100-200$ m/s), if the mass of CoRoT-11b had been $\lesssim0.5$ , it would not have been detected by the RV survey.
According to our planetary mass determinations ($=2.33\pm0.34$), CoRoT-11b is among the most massive transiting hot-Jupiters discovered so far. It actually belongs to the poorly populated sub-group of objects with planetary masses around 2 . As already noticed by @Torres10, there seems to be a lack of transiting hot-Jupiters with masses larger than about 2 . Based on the list of currently known transiting planets[^7], hot-Jupiters with masses in the range $0.5\lesssim$$\lesssim1.5$ seem to be $\sim5$ times more numerous than those with masses $1.5\lesssim$$\lesssim2.5$ . This trend is also confirmed by the number of planets discovered with the RV method. In Fig. \[FIG\_AvsM\] the semi-major axis of the planets detected in radial velocity and transit surveys is plotted as a function of the planetary mass. Let us consider only the objects with $a\lesssim0.1$ AU and $\gtrsim0.2$ , i.e., hot-Saturn and Jupiter planets. There is a clear clump of hot-giant planets with masses between the mass of Saturn ($\sim0.30$ ) and Jupiter, orbiting their parent star at about $0.04-0.06$ AU. Starting from $\sim1.0$ , the number of hot-Jupiters seems to drop off, whereas the spread in the semi-major axis increases. For $\gtrsim2$ the number of hot-Jupiters falls off significantly. The same trend is not seen for planets with $1\lesssim a \lesssim 5$ AU. The lack of Saturn planets orbiting their parent at such a distance might be owing to an observational bias of the RV technique. Nevertheless, since both the Doppler and transit methods are strongly sensitive in detecting close-in massive planets, we conclude that hot-Jupiters with $>2$ are significantly less common than “normal” hot-Saturn and Jupiter planets.
By assuming that CoRoT-11b is a hydrogen-rich gas giant we estimated the planet’s thermal mass loss by applying the method outlined in detail in @Lammer09. Because the planet orbits a F6-type star with an age between 1.0 and 3.0 Gyr, we used the soft X-rays and EUV flux scaling law of Eq. 12 of @Lammer09 and integrated the thermal mass loss during the planet’s history up to the two age values given above. By using the stellar and planetary parameters and a heating efficiency $\eta$ for hydrogen-rich thermospheres, which can be considered between 10 - 25 % [@Lammer09; @Murray09], we obtained a present time mass loss rate for CoRoT-11b of about $2.0\times10^{10}$ g/s or an integrated loss of 0.07 % of its present mass ($\eta=10$ %), and about $5.0\times10^{10}$ g/s, or 0.18 % ($\eta=25$ %) if the host star and planet are 1.0 Gyr old. If the the star/planet system is 3.0 Gyr old, we estimated a mass loss rate of about $3.0\times10^9$ g/s, or $0.1$ % ($\eta=10$ %) and about $7.7\times10^9$ g/s, or 0.25 % ($\eta=25$ %) during the planet lifetime. These loss rates agree well with hydrodynamic escape model results for typical hot Jupiters [@Yelle08]. Although the planet radius is $1.43\pm0.03$ , the main reason why the thermal mass loss of CoRoT-11b is not significant is related to the large mass of the planet of $2.33\pm0.34$ . According to @Lammer09, only hot gas giant planets with $\rho_\mathrm{p}<<1$ g/cm$^3$ should experience large thermal mass loss.
In order to investigate whether a standard model for an irradiated planet can account for the density of CoRoT-11b, we computed stellar and planetary evolution models using CESAM [@Morel08] and CEPAM [@Guillot95], as described in @Borde10 and @Guillot10. The results are shown in Fig. \[Fig\_Model\] where the evolution of the size of CoRoT-11b is plotted as a function of the system age. The colours indicate the distance in standard deviations from the inferred effective temperature and mean stellar density, i.e., less than $1\sigma$ (red), $2\sigma$ (blue) or $3\sigma$ (green). These constraints are compared to planetary evolution models for a homogeneous solar-composition hydrogen-planet, with different hypotheses: (1) using a “standard model”, i.e., without additional sources of heat; (2) by increasing interior opacities by a factor 30; (3) by adding a fraction ($\sim 1$%) of the incoming stellar energy and dissipating it at the centre; (4) and (5) by dissipating $5\times 10^{27}$ and $5\times 10^{28}$ erg/s at the centre of the planet. The first three cases correspond to standard recipes used to explain the inflated giant exoplanets [@Guillot08]. The last two cases correspond to higher dissipation levels that are required to explain the planet size for the oldest ages.
Interestingly, as for CoRoT-2b [@Guillot10], two classes of solutions are found: (i) the standard solution for which the host star is on the main sequence (with an age of about $2.0\pm1.0$ Gyr) and the planet requires a high level of dissipation in its interior in order to account for its large size; (ii) a very young class of solutions in which the star is still on the pre-main sequence (PMS) phase (with an age of about $12\pm2$ Myr) and the planet size can be quite naturally explained with a “standard model”. However, the latter scenario is in contradiction with the absence of the detectable Li[i]{} $\lambda$6708 Å line in the spectra of CoRoT-11. According to our effective temperature determination ($T_\mathrm{eff}=6440\pm120$ K), CoRoT-11 would belong to the narrow class of F-stars, which have suffered strong surface lithium depletion during the first billion years of their life. Indeed, studies of the lithium content in the photosphere of F-type stars in galactic clusters and field stars have revealed the presence of a narrow dip in the lithium abundance for effective temperature between 6500 and 6800 K [@Mallik03; @Bohm04]. While the so-called “lithium-dip” is absent in the [@Pilachowski87] and in general in all the young cluster ($\lesssim100$ Myr), this dip is well observed in older cluster like the [$700$ Myr; @Boesgaard86a], [$1.7$ Gyr; @Hobbs86], as well as in many field F-stars [e.g., @Boesgaard86b]. We thus believe that an age of $2.0\pm1.0$ Gyr is more plausible for the planet-star system.
[l r]{}\
Planet orbital period $P$ \[days\] & $2.994330\pm0.000011$\
Planetary transit epoch $T_\mathrm{tr}$ \[HJD-2400000\] & $54597.6790\pm0.0003$\
Planetary transit duration $d_\mathrm{tr}$ \[h\] & $2.5009\pm0.0144$\
\
Orbital eccentricity $e$ & 0 (fixed)\
Radial velocity semi-amplitude $K$ \[m/s\] & $ 280.0\pm40.0 $\
O-C residuals \[m/s\] & 88\
\
Planet-to-star radius ratio $k=R_\mathrm{p}/R_{*}$ & $ 0.1070\pm0.0005 $\
Linear limb darkening coefficient $u_+$ & $ 0.61\pm0.06 $\
Linear limb darkening coefficient $u_-$ & $ 0.02\pm0.04 $\
Orbital phase of planetary transit ingress ($\theta_1$) & $ -0.0174\pm0.0001 $\
\
Scaled semi-major axis $a/R_{*}$ & $ 6.890\pm0.080 $\
$M^{1/3}_{*}/R_{*}$ \[solar units\] & $ 0.787\pm0.010 $\
Mean stellar density $\rho_{*}$ \[g/cm$^3$\] & $ 0.69\pm0.02 $\
Inclination $i$ \[deg\] & $ 83.170\pm0.150 $\
Impact parameter $b$ & $ 0.818\pm0.008 $\
\
Effective temperature $T_\mathrm{eff}$ \[K\] & $ 6440\pm120 $\
Surface gravity log$g$ \[dex\] & $ 4.22\pm0.23 $\
Surface gravity log$g$ \[dex\] & $ 4.26\pm0.06 $\
Metallicity $[\rm{Fe/H}]$ \[dex\] & $ -0.03\pm0.08 $\
Stellar rotational velocity \[km/s\] & $ 40.0\pm5.0 $\
Spectral type & F6V\
\
Star mass $M_{\star}$ \[\] & $ 1.27\pm0.05 $\
Star radius $R_{\star}$ \[\] & $ 1.37\pm0.03 $\
Age of the star $t$ \[Gyr\] & $ 2.0\pm1.0 $\
Interstellar extinction $A_{\mathrm V}$ \[mag\] & $ 0.55\pm0.10 $\
Distance of the system $d$ \[pc\] & $ 560\pm30 $\
\
Planet mass $M_\mathrm{p}$ \[M$_\mathrm{J}$ \] & $ 2.33\pm0.34 $\
Planet radius $R_\mathrm{p}$ \[R$_\mathrm{J}$\] & $ 1.43\pm0.03 $\
Planet density $\rho_\mathrm{p}$ \[g/cm$^3$\] & $ 0.99\pm0.15 $\
Orbital semi-major axis $a$ \[AU\] & $ 0.0436\pm0.005$\
Equilibrium temperature $T_\mathrm{eq}$ \[K\] & $ 1657\pm55 $\
\[Par\_table\]
We thank the anonymous referee for his/her careful reading, useful comments, and suggestions, which helped to improve the manuscript.
This paper is based on observations carried out at the European Southern Observatory (ESO), La Silla and Paranal (Chile), under observing programs numbers 081.C-0388, 081.C-0413, and 083.C-0186. The authors are grateful to the staff at ESO La Silla and ESO Paranal Observatories for their support and contribution to the success of the HARPS and UVES observing runs.
This paper is also based on observations performed with SOPHIE at the Observatoire de Haute-Provence, France, under observing program PNP.08A.MOUT.
Part of the data presented herein were obtained at the W.M. Keck Observatory from telescope time allocated to the National Aeronautics and Space Administration through the agency’s scientific partnership with the California Institute of Technology and the University of California. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation. The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain.
The team at IAC acknowledges support by grant ESP2007-65480-C02-02 of the Spanish Ministerio de Ciencia e Innovaci´on. The German [*CoRoT*]{} Team (TLS and the University of Cologne) acknowledges DLR grants 50OW0204, 50OW0603, and 50QP07011.
This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
Aigrain, S., Collier Cameron, A., Ollivier, M., et al. 2008, , 488, 43 Aigrain, S., Pont, F., Fressin, F., et al. 2009, , 506, 425 Alonso, R., Auvergne, M., Baglin, A., et al. 2008, , 482, L21 Alonso, R., Guillot, T., Mazeh, T. , et al. 2009, , 501, L23 Alonso, R., Alapini, A., Aigrain, S. et al. 2009, , 506, 353 Anderson, D. R., Hellier, C., Gillon, M., et al. 2010,709, 159 Auvergne, M., Bodin, P., Boisnard, L., et al. 2009, , 506, 411 Baglin, A., Auvergne, M., Boisnard, L., et al. 2006, 36th COSPAR Scientific Assembly, 36, 3749 Böhm-Vitense E. 2004, , 128, 2435 Baranne, A., Queloz, D., Mayor, M., et al. 1994, , 119, 373 Barge, P., Baglin, A., Auvergne, M., et al. 2008, , 482, L17 Boesgaard, A. M. & Tripicco, M. J. 1986, , 302, 49 Boesgaard, A. M. & Tripicco, M. J. 1986, , 303, 724 Bordé P., Bouchy F., Deleuil M., et al. 2010, , submitted Borucki, W. J. & Summers, A. L. 1984, , 58, 121 Borucki, W. J., et al. 2010, , 713, L126 Bouchy, F., Queloz, D., Deleuil, M., et al. 2008, , 482, L25 Bouchy, F., Hébrard, G., Udry, S., et al. 2009, , 505, 853 Bowler, B. P., Johnson, J. A., Marcy, G. W., et al. 2010, , 709, 396 Bruntt, H., Bikmaev, I. F., Catala, C., et al. 2004, , 425, 683 Bruntt, H., De Cat, P. & Aerts, C. 2008, , 478, 487 Bruntt, H. 2009, , 506, 235 Bruntt, H., Deleuil, M., Fridlund M. et al. 2010, eprint arXiv:1005.3208 Cabrera, J., Fridlund, M., Ollivier, M., et al. 2009, 506, 501 Collier Cameron, A., Günther, E., Smalley, B., et al. 2010, , 407, 507 Carpano, S., Cabrera, J., Alonso, R., et al. 2009, 506, 491 Cardelli J. A., Clayton G. C. & Mathis J. S. 1989, 345, 245 Carone, L., Gandolfi, D. et al., submitted Claret, A. 2000, , 359, 289 Cox, A.N. 2000, “Allen’s astrophysical quantities”, 4$^{th}$ ed., edited by A.N. Cox, AIP Press Cresswell, P., Dirksen, G., Kley, W. nad Nelson, R.P. 2007, 473, 329 Cutri R. M., Skrutskie M. F., van Dyk S. et al. 2003, *2MASS All-Sky Catalog of Point Sources*, NASA/IPAC Infrared Science Archive Deeg, H. J., Gillon, M., Shporer, A., et al. 2009, , 506, 343 Deeg, H. J., Moutou, C., Erikson, A., et al. 2010, , 464, 384 Deleuil, M., Deeg, H. J., Alonso, R., et al. 2008, , 491, 889 Deleuil, M., Meunier, J. C., Moutou, C., et al. 2009, , 138, 649 Eigmüller. P. & Eislöffel, J. 2009, Transiting Planets, Proceedings of the International Astronomical Union, IAU Symposium “Trasiting Planet”, 253, 340 Eislöffel, J., Hatzes, A. P., Rauer, H., et al. 2007, Solar and Stellar Physics Through Eclipses ASP Conference Series, 370, 91 Endl, M., Kürster, M., Els, S. 2000, , 362, 585 Frasca A., Alcalá J.M., Covino E. et al. 2003, 405, 149 Fridlund, M., Hebrard, G., Alonso, R., et al. 2010, , 512, A14 Gandolfi, D., Alcalá, J. M., Leccia, S., et al. 2008, , 686, 1303 Gaudi, B. S. & Winn J. N 2007, , 655, 550 Guenther, E. W., Hartmann, M., Esposito, M., et al. 2009, , 507, 1659 Guillot, T. & Morel, P. 1995, , 109, 109 Guillot, T. 2008, Physica Scripta, 130, 014023 Guillot, T. & Havel, M. 2010, , submitted Gray D. F., , 98, 319 Gray D. F., , 101, 832 Giménez, A. 2006, , 450, 1231 Gustafsson, B., Edvardsson, B., Eriksson, K., et al. 2008, , 486, 951 Hauschildt P.H., Allard F. & Baron E. 1999, , 512, 377 Hébrard, G., Bouchy, F., Pont, F., et al. 2008, , 481, 52 Hobbs, L. M. & Pilachowski, C. 1986, , 309, 17 Johnson, J. A., Howard, A. W., Bowler, B. P., et al. 2010, , 122, 701 Kallrath, J. & Milone, E.F. 2009, “Eclipsing Binary Stars: Modeling and Analysis”, Springer-Verlag New York Koch, D., et al. 2010, ApJ, 713, L79 Kupka, F., Piskunov, N., Ryabchikova, T. A., Stempels, H. C., Weiss, W. W., 1999, , 138, 119 Lammer, H., Odert, P., Leitzinger, M. et al. 2009, , 506, 399 Lagrange, A.-M., Desort, M., Galland, F., et al. 2009, , 495, 335 Léger, A., Rouan, D., Schneider, J., et al. 2009, , 506, 287 Lubow, S.H. & Ogilvie, G.I. 2001 ,ApJ 560, 997 Mayor, M., Queloz, D. 1995, , 378, 355 Mayor, M., Pepe, F., Queloz, D., et al. 2003, Msngr., 114, 20 Mallik, Sushma V., Parthasarathy, M. & Pati, A. K. 2003,, 409, 251M Morel, P. & Lebreton, Y. 2008, , 316, 61 Moutou, C., Bruntt, H., Guillot, T., et al. 2008, , 488, L47 Moutou, C., Hébrard, G., Bouchy, F., et al. 2009, , 498, L5 Moutou, C., Pont, F., Bouchy, F., et al. 2009, , 506, 321 Murray-Clay, R. A., Chiang, E. I., & Murray, N. 2009, , 693, 23 O’Toole, S. J., Tinney, C. G., Jones, H. R. A., et al. 2009, , 392, 641 Ohta, Y., Taruya, A., and Suto, Y. 2005, , 622, 1118 Papaloizou, J.C.B. & Terquem, C. 2006, Reports on Progress in Physics, 69, 119 Pepe, F., Mayor, M., Galland, F., et al. 2002, , 388, 632 Pilachowski, C. A., Booth, J. & Hobbs, L. M. 1987, , 99, 1288 Pinheiro da Silva, L., Rolland, G., Lapeyrere, V. and Auvergne, M. 2008, , 384, 1337 Pollacco, D., Skillen, I., Collier Cameron, A., et al. 2007, , 385, 1576 Pont, F., Endl, M., Cochran, W. D., et al., 2009, , 402, 1 Press, W.H., Teukolsky, S.A., Vetterling, W.T. and Flannery, B.P. 1992, “Numerical recipes in C. The art of scientific computing”, Cambridge: University Press, 2$^{nd}$ edition Queloz, D., Henry, G. W., Sivan, J. P., et al. 2001, , 379, 279 Queloz, D., Bouchy, F., Moutou, C., et al. 2009, , 506, 303 Quentin, C. G., Barge, P., Cautain, R., et al. 2006, in ESA Special Publication, 1306, 409 Rauer, H., Queloz, D., Csizmadia, S., et al. 2009, , 506, 281 Rosenblatt F. 1971, , 14, 71 Scargle, J. D. 1982, , 263, 835 Seager, S., Mallén-Ornelas, G. 2003, , 585, 1038 Simpson, E. K., Pollacco, D., Hébrard, G., et al. 2010, , 405, 1867 Sing D.K 2010, å, 510, A21 Snellen, I. A. G., Koppenhoefer, J., van der Burg, R. F. J., et al. 2009, , 497, 545 Surace, C., Alonso, R., Barge, P., et al. 2008, SPIE, 7019, 111 Torres, G., Bakos, G.Á., Hartman, J. et al. 2010, , 715, 458 Valenti, J. A. & Piskunov, N., 1996, , 118, 595 Valenti, J. A. & Fischer, D. A., 2005, , 159, 141 Vogt, S.S., Allen, S.L. & Bigelow, B.C., et al. 1994, SPIE, 2198, 362 Winn J. N. 2010, eprint arXiv:1001.2010 Yelle, R., Lammer, H. & Ip, W.-H. 2008, Space Sci. Rev., 139, 437
[^1]: The [*CoRoT*]{} space mission, launched on 2006 December 27, has been developed and is operated by CNES, with the contribution of Austria, Belgium, Brazil, ESA (RSSD and Science Programme), Germany and Spain.
[^2]: The ‘*LR*’ prefix means that the field is a long-run field (typically 150 days of observations). The letter ‘*c*’ refers to the Galactic centre direction. The last two digits ’*02*’ indicate that the observed field is the second [*CoRoT*]{} long-run towards the Galaxy centre.
[^3]: See @Carpano09 for a full description of the [*CoRoT*]{} target nomenclature.
[^4]: We remind the reader that transits provide the direct measurement of only the planet-to-star radius ratio.
[^5]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA), inc., under cooperative agreement with the National Science Foundation.
[^6]: IDL is distributed by ITT Visual Information Solutions, Boulder, Colorado.
[^7]: We refer the reader to the Extra Solar Planets Encyclopedia for a constantly updated list of known extra solar planets (http://exoplanet.eu/).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Preparing an integrable system in a mixed state described by a thermal density matrix , we subject it to a sudden quench and explore the subsequent unitary dynamics. Defining a version of the generalised Loschmidt overlap amplitude (GLOA) through the purifications of the time evolved density matrix, we claim that non-analyiticies in the corresponding “dynamical free energy density" persist and is referred to as mixed state dynamical quantum phase transitions (MSDQPTs). Furthermore, these MSDQPTs are uniquely characterised by a topological index constructed by the application of the Pancharatnam geometry on the purifications of the thermal density matrix; the quantization of this index however persists up to a critical temperature. These claims are corroborated analysing the non-equilibrium dynamics of a transverse Ising chain initially prepared in a thermal state and subjected to a sudden quench of the transverse field.'
author:
- |
Utso Bhattacharya, Souvik Bandopadhyay and Amit Dutta\
Department of Physics, Indian Institute of Technology, Kanpur-208016, India
- |
Utso Bhattacharya, Souvik Bandopadhyay and Amit Dutta\
Department of Physics, Indian Institute of Technology, Kanpur-208016, India
title:
- Mixed state dynamical quantum phase transition and emergent topology
- Mixed state dynamical quantum phase transition and emergent topology
---
Recent experimental advances in realisation of closed condensed matter systems via cold atoms in optical lattices, especially studies of the real time evolution of closed quantum systems in cold atomic gases [bloch08,lewenstein12,jotzu14,greiner02,kinoshita06,gring12,trotzky12,cheneau12,schreiber15]{} and also light-induced non-equilibrium superconducting and topological systems [fausti11,rechtsman13]{} have resulted in an upsurge in related theoretical works [calabrese06,rigol08,oka09,kitagawa10,lindner11,bermudez09,patel13,thakurathi13,mukherjee09,das10,Russomanno\_PRL12,bukov16,pal10,nandkishore15]{}. [(For review, we refer to [dziarmaga10,polkovnikov11,dutta15,eisert15,alessio16]{}.)]{} One of the emerging areas of research on non-equilibrium (quenched) closed quantum systems is the dynamical quantum phase transitions (DQPTs) as has been proposed by Heyl $et~al.$ [heyl13]{} in connection to thermal phase transitions. These phase transitions are characterised by non-analyticities in the so called dynamical free energy during the subsequent temporal evolution of the quenched system when the initial state is temporally orthogonal to the time-evolved state; these non-analyticities are manifested in the logarithm of the Loschmidt echo (or the rate function of the return probability). Remarkably, in some situations, these non-equilibrium transitions can be characterised by a dynamical topological order parameter (DTOP) [budich15]{} extracted from the “gauge-invariant" Pancharatnam phase [@pancharatnam56; @berry84] (see \[\] for review) extracted from the Loschmidt overlap, i.e., the overlap between the initial state and the time evolved state.
Let us first elaborate on the basic notion of a DQPT focussing on the sudden quenching case [heyl13]{}. Denoting the ground state of the initial Hamiltonian as $|\psi_0\rangle$ and the final Hamiltonian reached through the quenching process as $H_f$, the Loschmidt overlap amplitude (LOA) is defined as $G(t)=\langle\psi_0|e^{-iH_ft}|\psi_0\rangle$. Generalizing $G(t)$ to $G(z)$ defined in the complex time ($z$) plane such that $z=R+it$, one can introduce the notion of a dynamical free energy density, $f(z)=-\lim_{L\to \infty} \ln{G(z)}/L^d$, where $L$ is the linear dimension of a $d$-dimensional system. One then looks for the zeros of the $G(z)$ (or non-analyticities in $f(z)$) to define a DQPT. For a transverse Ising chain, it has been observed that when the system is suddenly quenched across the quantum critical point (QCP), the lines of Fisher zeros (FZs) cross the imaginary time axis at instants $t_n^{*}$; at these instants [the rate function of the return probability]{} defined as $I(t) = - \ln |G(t)|^2/L$ shows sharp non-analyticities. Several subsequent studies [karrasch13,kriel14,andraschko14,canovi14,heyl14,vajna14,sharma15,heyl15,palami15,vajna15,schmitt15,budich15,sharma16,divakaran16,huang16,puskarov16,zhang16,heyl16,bhattacharya16,zunkovic16, bhattacharya17,zvyagin17,sei17,fogarty17,halimeh171,homrighausen17]{} established that similar DQPTs are observed for sudden quenches across the QCP for both integrable and non-integrable models, although crossing the QCP is not essential in some situations [vajna14,sharma15]{}. DQPTs have also been observed when the final state, evolving with the time-independent final Hamiltonian, is prepared through a slow ramping of a parameter of the Hamiltonian [sharma16,pollmann10]{}. Furthermore, on the experimental side, non-analyticities have been detected in the dynamical evolution of a fermionic many-body state after a quench [flaschner16]{} and also in the non-equilibrium dynamics of a string of ions simulating interacting transverse-field Ising models [jurcevic16]{}. We note in passing that the rate function $I(t)$ is related to the Loschmidt echo which has been studied in the context of equilibrium quantum phase transitions and associated dynamics [quan06,rossini07,cucchietti07,venuti10, sharma12,nag12,mukherjee12,dora13,zanardi07\_echo,gambassi11,dorner12]{}.
All these afore mentioned studies concern the subsequent unitary dynamics of the initially prepared pure state $|\psi_0\rangle$ following a quench of the system. The questions we address are the following: does the fascinating aspects of DQPTs arising due to quantum coherence get wiped out if the initial state happens to be a thermal (mixed) state. To achieve this goal, we prepare the initial state in a thermal Gibbs state described by a density matrix $\rho_0$ characterised by an inverse temperature $\beta$; a parameter of the Hamiltonian is then suddenly changed and we propose a generic form of the LOA (that reduces to the conventional definition given above in the pure state limit) to track the subsequent unitary evolution. We then explore the following questions: do mixed state DQPTs (MSDQPTs), manifested in the zeros of the generalized LOA (GLOA), occur in this situation? If so, can these MSDQPTs be characterised by an appropriate topological index? Interestingly, we find affirmative answers for all these questions. MSDQPTs indeed exist. Furthermore, these non-equilibrium transitions are characterized by a gauge-invariant topological index up to a sufficiently high initial temperature (denoted by $\beta_c$) below which the quantisation gets destroyed. In passing, we note that there have been several studies in recent years which explore the interconnection between temperature and topology of equilibrium systems [@delgado13; @arovas14] using Uhlmann phase [uhlmann86]{}. However, the topological index that we use here to characterise the MSDQPTs manifests out of the interferometric phase [sjoqvist00]{}.
[*Quenches, GLOA and MSDQPTs:*]{} Let us consider a many-body integrable Hamiltonian that can be decoupled into $2\times2$ Hamiltonians for each momentum mode $k$, i,e. $\hat{H}_k=\vec{d}_k \cdot {\vec {\sigma}}$, where ${\sigma}_i$s are the Pauli matrices. Henceforth, we set $\hbar$ and the Boltzmann constant $k_B$ to unity. The mixed state density matrix at time $t=0$ describing the system at thermal equilibrium with a bath corresponding to the initial Hamiltonian $\hat{H}^i_{k}=\vec{d}^i_{k}\cdot {\vec{\sigma}}$ can be written as
$${\hat \rho}_k(0)=\frac{e^{-\beta {\hat H}^i_{k}}}{{\rm Tr}(e^{-\beta {\hat H}^i_{k}})}=\frac{1}{2}(\mathbb{I}_k-n{\hat{{d}}^i}_{k}\cdot {\vec {\sigma}})$$
where $\beta$ is the inverse temperature, $ n=\tanh(\beta \epsilon^i_k)$, $\hat{{d}}^i_{k}= \vec{d}^i_k/\epsilon^i_k$ with $\epsilon^i_k=|\vec{d^i_{k}}|$ and $\mathbb{I}_k$ is the $2\times 2$ identity matrix.
Having prepared the initial density matrix we decouple the system from the bath [dorner12]{} at $t=0$ and subject it to a sudden change in a parameter $h$ of the Hamiltonian from an initial value $h_i$ to a final value $h_f$; we then probe the subsequent non-equilibrium dynamics of the system generated by the final time-independent Hamiltonian $ \hat{H}^f_{k}=\vec{d}^f_{k}\cdot \hat{\vec{\sigma}}$, so that the density matrix ${\hat \rho}_k(t)$ at a time $t$ after the quench is given by ${\hat \rho}_k(t)={\hat U}_k(t) \hat {\rho}_k(0){\hat U}_k^{\dagger}(t)$.
![ The lines of FZs in the complex $z$ plane following a quench from the initial value of the transverse $h_i=0.5$ and to the final value $h_f=2$ across the QCP $h=1$ for the Hamiltonian corresponding to the initial $\beta=10$ and $\beta =20$. The lines of FZs cut the imaginary (real time) axis at critical times $t_j^*$ given in Eq. corresponding to $j=0,\pm 1$.[]{data-label="fig_FZs"}](Fz10.pdf "fig:"){width=".20\textwidth" height="3cm"} ![ The lines of FZs in the complex $z$ plane following a quench from the initial value of the transverse $h_i=0.5$ and to the final value $h_f=2$ across the QCP $h=1$ for the Hamiltonian corresponding to the initial $\beta=10$ and $\beta =20$. The lines of FZs cut the imaginary (real time) axis at critical times $t_j^*$ given in Eq. corresponding to $j=0,\pm 1$.[]{data-label="fig_FZs"}](Fz20.pdf "fig:"){width=".20\textwidth" height="3cm"}
Since we are dealing with the dynamics of mixed state, one needs to define a generalised Loschmidt overlap amplitude (GLOA) defined for each $k$ mode as the overlap between the purifications [uhlmann86]{} as (for a discussion on the purifications, refer to the supplementary materials (SM)) $$\begin{aligned}
L_k&=&\langle w_k(0)|w_k(t)\rangle=Tr ({\hat {\rho}}_k(0) {\hat U}_k(t)) ~~\\
&{\rm where}&~~{\hat U}_k(t)=e^{-i {\hat H}^f_{k}t}~~~{\rm and}~~{\cal L} =\prod L_k,
\end{aligned}$$ where ${\cal L}$ is called the total GLOA obtained through a product over all the modes $k$ from $0$ to $\pi$. Notably, the GLOA defined for the mixed state situation reduces to the conventional form of the LOA, $\langle \psi_0|e^{-i{\hat H}_f t}|\psi_0\rangle$, in the pure state limit when the system is in the initial state $|\psi_0\rangle$, before the quenching is applied.
The complete gauge invariant phase obtained from the GLOA for each $k$ mode is given by [sjoqvist00]{}, $$\begin{aligned}
\phi^k_{g}&=&\phi^k-\phi^k_{dyn}{\nonumber}\\
&=&{\rm Arg} [{\rm Tr}(\hat{\rho_k}(0) {{\hat U}_k}(t))]+\int_0^t {\rm Tr}(\hat{\rho_k}(t)\hat{H}^f_k)dt
\label{eq_geo_phase}\end{aligned}$$ where $\phi^k={\rm Arg} [{\rm Tr}
(\hat{\rho_k}(0)\hat{U_k}(t))]$ and $\phi^k_{dyn}= -\int_0^t {\rm Tr}(\hat{\rho_k}(t)\hat{H}^f_k)dt$. See the SM for a thorough discussion. For the sudden quench protocol from an initial value of the parameter $h_i$ to a final value $h_f$ with initial state characterised by the initial inverse temperature $\beta$, we find
$$\phi^k=\tan ^{-1}\left[\frac{n\vec{d}^i_{k} \cdot \vec{d}^f_{k}}{\epsilon^i_k.\epsilon_f^k}\tan (\epsilon_f^kt)\right];~~~
\phi^k_{dyn}=\frac{n\vec{d}^i_{k} \cdot \vec{d}^f_{k}t}{\epsilon^i_k},
\label{eq_phases}$$
where $\epsilon_{k}^{f}=|\vec{d}^f_{k^{}}|$.
{width="37.00000%"} {width="37.00000%"}
Let us now proceed to define the dynamical counterpart of the free energy density in the thermodynamic limit analogous to the pure state situation [heyl13]{},
$$f(t)=-\frac{1}{2\pi}{\rm Re} \left[\int_{0}^{\pi}\log[{\rm Tr}(\hat{\rho}_k(0) {\hat U}_k(t))]dk\right].
\label{eq_free_energy}$$
\
The position of the non-analyticities in the free energy density can be determined solely by the zeros of the partition function $L_{k}(z)=Tr[\hat{\rho}_{k}(0)e^{-\hat{H}^f_{k} z}]$ (where $z\in C$); these are reflected in the real time evolution of the dynamical free energy density when the zeros lie on the imaginary (real time axis). The zeros of $L_k(z)$ can be evaluated for the present case as $$z_j(k)=\frac{1}{2\epsilon^f_{k}}\ln\left(\frac{1-n\hat{d}^i_{k}\cdot \hat{d}^f_{k}}{1+n\hat{d}^i_{k}\cdot \hat{d}^f_{k}}\right)+\frac{i(2j+1)\pi}{2\epsilon^f_{k}}$$ where ${\hat d}^f_{k} = \vec{d}^f_{k}/\epsilon^f_{k}$ and $j \in \mathbb{Z}$. The line of FZs in the complex plane cuts the imaginary axis for the mode $k^{*}$ when $\hat{d}^i_{k^{*}}\cdot \hat{d}^f_{k^{*}}=0$; consequently, one observes non-analyticities in the dynamical free energy density $f(t)$ at instants of real time given by:
$$t_{j}^{*}=\frac{(2j+1)\pi}{2\epsilon_{k*}^{f}}
\label{eq_time}$$
where $\epsilon_{k*}^{f}=|\vec{d}^f_{k^{*}}|$. Evidently, the critical times given in Eq. do not depend upon the initial temperature of the system.
[*Mixed state dynamical topological order parameter (MSDTOP):*]{} Let us now address the question whether the MSDQPTs occurring at critical times $t_j^*$ can be characterised by an appropriate topological index. To address this question, in a spirit similar to the pure state situation [budich15]{}, we invoke upon the gauge invariant geometric phase defined in to construct the MSDTOP as $$\nu (t)=\frac{1}{2\pi}\int_{0}^{\pi}\frac{\partial \phi_g^k}{\partial k}dk
\label{eq_dtop}$$ The MSDTOP will remain fixed between two critical times showing discontinuous jumps at every MSDQPT thereby appropriately characterising them as long as the initial temperature does not exceed the minimum characteristic energy scale of the problem. In the pure state limit evidently $\phi^k_g$ reduces to the Pancharatnam geometric phase and one can readily check that for the modes $k=0$ and $k=\pi$, $\phi^{k}_g$ is pinned to zero [budich15]{}. Even when $\beta$ is finite but $\epsilon^i_k \gg T$ (so that $n \to 1$), $\phi^k_g$ still remains periodic in the interval $k\in[0,\pi]$; consequently, $\nu(t)$ cannot change in the region $k\in[0,\pi]$ except for the instants of MSDQPTs when, as seen from Eq. , the phase $\phi^k$ becomes ill-defined. We therefore find that MSDTOP $\nu(t)$ stays quantised and jumps by a factor of unity at every MSDQPT. However if $\epsilon^i_k\lesssim T$, so that $n \lesssim 1$ the phase $\phi^k_g$ is no longer periodic in the interval $k\in[0,\pi]$, rather we have $$\begin{aligned}
\nonumber\phi_g^{k=0,\pi}&=&\tan ^{-1}\left[n(k=0,\pi)\tan (\epsilon^f_{k=0,\pi}t)\right]-\\
&&n(k=0,\pi)\epsilon^f_{k=0,\pi}t
\end{aligned}$$ where the quantity $n$ also depends on $k$ rendering $\phi^{k=0}_g \neq \phi^{k=\pi}_g$ and hence destroying the periodicity in $\phi^{k}_g$. Thus, we can define a critical initial temperature $T_c$ above which the quantization of the MSDTOP gets destroyed, determined by the condition, $$\tanh\left(\frac{\epsilon_i^{k^{'}}}{T_c}\right)\sim1
\label{eq_critical}$$ where $k^{'}$ is the mode for which the minimum gap in the spectrum of $\epsilon^i_k$ is $\epsilon^i_{k'}$.
[*Illustration with the transverse field Ising chain:* ]{} We shall now illustrate the above claims considering the integrable one-dimensional (1D) transverse Ising Hamiltonian [@sachdev96; @sei13; @dutta15]:
$$\hat{H}=-J\sum_{\langle i,j\rangle} {\sigma}^x_i {\sigma}^x_j-h\sum_i {\sigma}^z_i
\label{eq_ham}$$
where $J$ (henceforth set equal to $1$) is the nearest neighbour ferromagnetic coupling strength and $h$ is the non-commuting transverse magnetic field. Employing a Fourier transformation and a subsequent Jordan-Wigner transformation, one can map the many-body Hamiltonian into decoupled $2\times2$ Hamiltonians for each momentum mode $k$, i,e. $\hat{H}_k=\vec{d}_k \cdot {\vec{\sigma}}$, where the ${\vec d}_k$ has components $(0,\cos k-h,\sin k)$. Analysing the spectrum of the model, one can show that the TFIM has two Quantum Critical Points (QCPs) at $h = \pm 1$ separating the ferromagnetic phase from the paramagnetic phases for $|h| >1$.\
Having prepared the system in a thermal state, we employ a sudden change in the field $h$ from $h_i$ to $h_f$ and hence in the present example, $\vec{d}^i_{k}\equiv(0,\cos k-h_i,\sin k)$ and $\vec{d}^f_{k}\equiv(0,\cos k-h_f,\sin k)$. In the following, we choose $h_i=0.5$ and $h_f=2$ so that the spin chain is quenched across the QCP at $h=1$. We indeed find that the lines of FZs cross the real time axis at critical times $t_{j}^{*}$ as shown in Fig. \[fig\_FZs\] where the dynamical free energy density $f(t)$ defined in Eq. becomes non-analytic (as shown in Fig. \[fig\_dtop\]). Moreover, here we also show that the MSDTOP defined in Eq. remains fixed between two critical times showing discontinuous jumps by a factor of unity at every MSDQPT, thereby appropriately characterising them. However, if the initial temperature is relatively high, as we illustrate in Fig. \[fig\_finite\_T\], the perfect quantisation of MSDTOP gets destroyed. The critical condition follows from Eq. with $k^{'}=0$ in the present example.
![The perfect quantisation of $\nu(t)$ between two successive critical times for the same quenching protocol as shown in Fig. \[fig\_dtop\] gets destroyed when the initial temperature is relatively higher. []{data-label="fig_finite_T"}](Beta_1_5.png){width="37.00000%"}
[*Experimental possibilities:*]{} The first experimental observation of quantum geometric phases for mixed states was performed by Du $et~al.$ [du03]{}. The quantum holonomies for a mixed state of spin half nucleus was measured and observed in accordance with the theoretical predictions by Sjoqvist $et~al$ [sjoqvist00]{}. Interestingly, for a pure state non-equilibrium dynamics, DQPTs were observed using state tomography technique, where Flaschner $et~al.$ [flaschner16]{} measured the time evolution of the pure state for each quasi-momenta mode on the Bloch sphere to reconstruct the instantaneous many-body state. Combining this technique with the experimental procedure of Ref. \[\], a similar approach may be undertaken to map a time evolving mixed state on a Bloch sphere to measure the geometric phase acquired and observe MSDQPTs. Furthermore, the successful measurement of DQPTs in an interacting many-body system such as the transverse field Ising model with long range interactions, opens up the possibility of the observation of MSDQPTs and MSDTOPs in such a model.
[*Concluding comments:*]{} We have defined a GLOA for a thermal mixed state undergoing unitary dynamics using purifications and have proposed a MSDTOP similar to the construction of the gauge invariant interferometric phase. Our GLOA and MSDTOP comes out of a generic framework which is independent of the underlying integrable model involved. To substantiate our claims we take the example of the 1-D Ising model and show the existence of MSDQPTs and a MSDTOP. We observe that although the thermal nature of the mixed state completely destroys the quantization of the MSDTOP when the initial temperature scale becomes comparable to the temperature associated with the initial minimum energy gap, the critical modes and times of the MSDQPTs which depend only on the unitarity of the quenching process involved, remain unaffected. This firmly establishes the existence of a critical temperature below which MSDQPTs exist and are characterized by the MSDTOP just as in the $T\rightarrow0$ limit. Since, our formalism holds true for integrable quantum systems, the existence or non-existence of MSDQPTS for any general unitary dynamics involved can be confirmed. It also has a natural well defined reduction to the pure state limit. Although the definition of GLOA is indeed independent of the dimensionality of the system, the MSDTOP defined here is strictly valid for a one dimensional system though may even be generalised to higher dimensions [bhattacharya17]{}.
[11]{}
[bloch08]{} I. Bloch, J. Dalibard, and W. Zwerger, Rev. Mod. Phys. [**80**]{}, 885 (2008). [lewenstein12]{} M. Lewenstein, A. Sanpera, and V. Ahufinger, (Oxford University Press, Oxford (2012)).
G. Jotzu , M. Messer, R. Desbuquois, M. Lebrat, T. Uehlinger, D. Greif and T. Esslinger, Nature [**515**]{}, 237 (2014).
[greiner02]{} M. Greiner , O. Mandel, T. W. Hansch and I. Bloch, Nature 419, 51 (2002).
[kinoshita06]{} T. Kinoshita, T. Wenger and D. S. Weiss, Nature [**440**]{}, 900 (2006).
[gring12]{} M. Gring, M. Kuhnert, T. Langen, T. Kitagawa, B. Rauer, M. Schreitl, I. Mazets1, D. Adu Smith, E. Demler, and J. Schmiedmayer, Science [**337**]{}, 1318 (2012).
[trotzky12]{} S. Trotzky, Y-A. Chen, A. Flesch, I. P. McCulloch, U. Schollwšck, J. Eisert and I. Bloch, Nature [**8**]{}, 325 (2012).
[cheneau12]{} M. Cheneau, P. Barmettler, D. Poletti, M. Endres, P. Schauss, T. Fukuhara, C. Gross, I. Bloch, C. Kollath and S. Kuhr, Nature [**481**]{}, 484 (2012).
[schreiber15]{} M. Schreiber, S. S. Hodgman, P. Bordia, Henrik P. LŸschen, M. H. Fischer, R. Vosk, E. Altman, U. Schneider, I. Bloch, Science [**349**]{}, 842 (2015).
[fausti11]{} D. Fausti, R. I. Tobey, , N. Dean, S. Kaiser, A. Dienst, M. C. Hoffmann, S. Pyon, T. Takayama, H. Takagi, A. Cavalleri, Science [**331**]{}, 189 (2011).
[rechtsman13]{} M. C. Rechtsman, J. M. Zeuner, Y. Plotnik, Y. Lumer, D.Podolsky, F. Dreisow, S. Nolte, M. Segev and A. Szameit, Nature [**496**]{} 196 (2013).
[calabrese06]{} P. Calabrese, and J. Cardy, Phys. Rev. Lett. [**96**]{}, 136801 (2006); J. Stat. Mech, P06008 (2007).
[rigol08]{} M. Rigol, V. Dunjko and M. Olshanii, Nature [**452**]{}, 854 (2008).
T Oka, H Aoki, Phys. Rev. B [**79**]{} 081406 (2009).
T. Kitagawa, E. Berg, M. Rudner, and E. Demler, Phys. Rev. B [**82**]{}, 235114 (2010).
N. H. Lindner, G. Refael and V. Galitski, Nat. Phys. [**7**]{}, 490-495, (2011).
[bermudez09]{} A. Bermudez, D. Patane, L. Amico, M. A. Martin-Delgado, Phys. Rev. Lett. [**102**]{}, 135702, (2009).
[patel13]{} A. A. Patel, S. Sharma, A. Dutta, Eur. Phys. Jour. B [**86**]{}, 367 (2013); A. Rajak and A. Dutta, Phys. Rev. E 89, 042125, 2014. P. D. Sacramento, Phys. Rev. E [**90**]{} 032138, (2014); M. D. Caio, N. R. Cooper and M. J. Bhaseen, Phys. Rev. Lett. [**115**]{}, 236403 (2015).
M. Thakurathi, A. A. Patel, D. Sen, and A. Dutta, Phys. Rev. B 88, 155133 (2013).
V. Mukherjee V. and A. Dutta, J. Stat. Mech. P05005 (2009).
A. Das, Phys. Rev. B [**82**]{}, 172402 (2010).
, Phys. Rev. Lett. [**109**]{}, 257201 (2012); S. Sharma, A. Russomanno, G. E. Santoro and A. Dutta, EPL [**106**]{}, 67003 (2014).
[bukov16]{} M. Bukov, L. D’Alessio and A. Polkovnikov, Adv. Phys. [**64**]{} , No. 2, 139-226 (2016).
A Pal and D. A. Huse, Phys. Rev. B [**82**]{}, 174411 (2010).
[nandkishore15]{} R. Nandkishore, D. A. Huse, Annual Review of Condensed Matter Physics, [**6**]{}, 15-38 (2015).
J. Dziarmaga, Advances in Physics [**59**]{}, 1063 (2010).
A. Polkovnikov, K. Sengupta, A. Silva, and M. Vengalattore, Rev. Mod. Phys. [**83**]{}, 863 (2011).
[dutta15]{} A. Dutta, G. Aeppli, B. K. Chakrabarti, U. Divakaran, T. Rosenbaum and D. Sen, *Quantum Phase Transitions in Transverse Field Spin Models: From Statistical Physics to Quantum Information* (Cambridge University Press, Cambridge, 2015).
[eisert15]{} J. Eisert, M. Friesdorf and C. Gogolin, Nat. Phys. [**11**]{}, 124 (2015).
[alessio16]{} L. D’Alessio, Y. Kafri, A. Polkovnikov, M. Rigol, Adv. Phys. [**65**]{}, 239 (2016).
[heyl13]{} M. Heyl, A. Polkovnikov, and S. Kehrein, Phys. Rev. Lett., [**110**]{}, 135704 (2013).
[budich15]{} J. C. Budich and M. Heyl, Phys. Rev. B 93, 085416 (2016).
[pancharatnam56]{} S. Pancharatnam, Proc. Indian Acad. Sci. A [**44**]{}, 247 (1956)
[berry84]{} M. V. Berry, Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences [**392**]{}, 45 (1984).
[zella12]{} Francisco. De Zela, [*The Pancharatnam-Berry Phase: Theoretical and Experimental Aspects*]{}, INTECH Open Access Publisher, 2012.
[karrasch13]{} C. Karrasch and D. Schuricht, Phys. Rev. B, [**87**]{}, 195104 (2013).
[kriel14]{} N. Kriel, C. Karrasch, and S. Kehrein, Phys. Rev. B [**90**]{}, 125106 (2014).
[heyl14]{} M. Heyl, Phys. Rev. Lett., [**113**]{}, 205701 (2014).
[andraschko14]{} F. Andraschko, J. Sirker, Phys. Rev. B [**89**]{}, 125120 (2014).
[canovi14]{} E. Canovi, P. Werner, and M. Eckstein, Phys. Rev. Lett. [**113**]{}, 265702 (2014).
[heyl15]{} M. Heyl, Phys. Rev. Lett., [**115**]{}, 140602 (2015) .
[palami15]{} T. Palmai, Phys. Rev. B [**92**]{}, 235433 (2015).
[vajna14]{} S. Vajna and B. Dora, Phys. Rev. B [**89**]{}, 161105(R) (2014).
[sharma15]{} S. Sharma, S. Suzuki and A. Dutta, Phys. Rev. B [**92**]{}, 104306 (2015).
[vajna15]{} S. Vajna and B. Dora, Phys. Rev. B [**91**]{}, 155127 (2015).
[schmitt15]{} M. Schmitt and S. Kehrein, Phys. Rev. B [**92**]{}, 075114 (2015).
[sharma16]{} S. Sharma, U. Divakaran, A. Polkovnikov and A. Dutta, Phys. Rev. B [**93**]{}, 144306 (2016).
[divakaran16]{} U. Divakaran, S. Sharma and A. Dutta, Phys. Rev. E [**93**]{}, 052133 (2016).
[huang16]{} Z. Huang, and A. V. Balatsky, Phys. Rev. Lett. [**117**]{}, 086802 (2016).
[puskarov16]{} T. Puskarov and D. Schuricht, arXiv: 1608.05584 (2016).
[zhang16]{} J. M. Zhang abd H.-T. Yang, arXiv: 1605.05403 (2016).
[heyl16]{} M. Heyl, Phys. Rev. B [**95**]{}, 060504 (2017).
[bhattacharya16]{} U. Bhattacharya and A. Dutta, arXiv:1610.02674 (2016); to appear in Phys. Rev. B (2017).
[zunkovic16]{} Bojan Zunkovic, Markus Heyl, Michael Knap, Alessandro Silva, arXiv:1609.08482 (2016).
[bhattacharya17]{} U. Bhattacharya and A. Dutta, arXiv: 1701.03911 (2017).
[zvyagin17]{} A.A. Zvyagin, arXiv:1701.08851 (2017).
[sei17]{} T. Obuchi, S. Suzuki, K. Takahashi, arXiv:1702.05396 (2017).
[fogarty17]{} Thom‡s Fogarty, Ayaka Usui, Thomas Busch, Alessandro Silva, John Goold, arXiv:1704.07659 (2017).
[halimeh171]{} J. C. Halimeh and V. Zauner-Stauber, arXiv:1610.02019 (2017).
[homrighausen17]{} I. Homrighausen, N. O. Abeling, V. Zauner-Stauber, and J. C. Halimeh, arXiv:1703.09195 (2017).
[dutta17]{} A. Dutta and A. Dutta, arXiv: 1705.03770 (2017).
[pollmann10]{} F. Pollmann, S. Mukerjee, A. G. Green, and J. E. Moore, Phys. Rev. E [**81**]{}, 020101(R) (2010).
[flaschner16]{} N. Flaschner, D. Vogel, M. Tarnowski, B, S. Rem, D.-S. LuŸhmann, M. Heyl, J. Budich, L. Mathey, K. Sengstock, C. Weitenberg, arXiv:1608.05616 (2016).
[jurcevic16]{} P. Jurcevic, H. Shen, P. Hauke, C. Maier, T. Brydges, C. Hempel, B. P. Lanyon, M. Heyl, R. Blatt, C. F. Roos, arXiv:1612.06902 (2016).
[quan06]{} H.T. Quan, Z. Song, X.F. Liu, P. Zanardi, and C.P. Sun, Phys.Rev.Lett. [**96**]{}, 140604 (2006).
[rossini07]{} D. Rossini, T. Calarco, V. Giovannetti, S. Montangero, R. Fazio, Phys. Rev. A [**75**]{}, 032333 (2007).
[cucchietti07]{} F. M. Cucchietti, $et~al$, Phys. Rev. A [**75**]{}, 032337 (2007); C. Cormick and J. P. Paz, Phys. Rev. A [**77**]{}, 022317 (2008).
[venuti10]{} Lorenzo C Venuti and P. Zanardi, Phys. Rev. A [**81**]{}, 022113 (2010); Lorenzo C. Venuti, N. T. Jacobson, S. Santra, and P. Zanardi, Phys. Rev. Lett. [**107**]{}, 010403 (2011).
[sharma12]{} S. Sharma, V. Mukherjee, and A. Dutta, Eur. Phys. J. B, [**85**]{}, 143 (2012).
[mukherjee12]{} V. Mukherjee, S. Sharma, A. Dutta, Phys. Rev. B [**86**]{}, 020301 (R) (2012).
[nag12]{} T. Nag, U. Divakaran and A. Dutta, Phys. Rev. A [**93**]{}, 012112(2016)
[dora13]{} B. Dora, F. Pollmann, J. Fort‡gh, G. Zarand, Phys. Rev. Lett. [**111**]{}, 046402 (2013); R. Sachdeva, T. Nag, A. Agarwal, A. Dutta, Phys. Rev. B [**90**]{}, 045421 (2014).
[gambassi11]{} A. Gambassi and A. Silva, arXiv: 1106.2671 (2011); Phys. Rev. Lett. [**109**]{}, 250602 (2012); P. Smacchia and A. Silva, Phys. Rev. E [**88**]{}, 042109, (2013); A. Russomanno, S. Sharma, A. Dutta and G. E. Santoro, J. Stat. Mech., P08030 (2015).
[zanardi07\_echo]{} P. Zanardi. H. T. Quan, X. Wang and C. P. Sun, Phys. Rev. A [**75**]{}, 032109 (2007).
[dorner12]{} R. Dorner, J. Goold, C. Cormick, M. Paternostro and V. Vedral, Phys. Rev. Lett. [**109**]{}, 160601 (2012);S. Sharma and A. Dutta, Phys. Rev. E [**92**]{}, 022108 (2015).
[delgado13]{} O. Viyuela, A. Rivas, M.A. Martin-Delgado, Phys. Rev. Lett. [**112**]{}, 130401 (2014); Phys. Rev. Lett. 113, 076408 (2014).
[arovas14]{} Zhoushen Huang, Daniel P. Arovas, Phys. Rev. Lett. [**113**]{}, 076407 (2014).
[uhlmann86]{} A. Uhlmann, Rep. on Math. Phys. [**24**]{}, 229 (1986); A. Uhlmann, Annalen der Physik [**501**]{}, 63 (1989).
[sjoqvist00]{} E.Sjoqvist, A. K. Pati, A. and Ekert, and J. S. Anandan, M. Ericsson, D. K. L. Oi, and V. Vedral, Phys. Rev. Lett. [**85**]{}, 2845 (2000).
S. Sachdev, *Quantum Phase Transitions* (Cambridge University Press, Cambridge, UK, 2010).
S. Suzuki, J-i Inoue, and B. K. Chkarabarti, [*Quantum Ising Phases and Transitions in Transverse Ising Models*]{}, Springer Lecture Notes in Physics, Vol. [**862**]{} (Springer, Berlin, 2013).
[du03]{} J. Du, $et ~al.$, Phys. Rev. Lett. [**91**]{} 100403 (2003).
**“Supplementary Material on “Mixed state dynamical quantum phase transition and emergent topology””**\
\
\
The mathematical extension of the Pancharatnam-Berry phase [@pancharatnam56], [@berry84] and the whole underlying geometry of pure quantum states (see for review) can be extended to mixed states by a redefined parallel transport criteria for mixed state density matrices, defined in the extended Hilbert Space of the purifications of the density matrix [@uhlmann86], [@uhlmann89]. In general there are two approaches i) gauge invariant phase for density matrices defined by Uhlmann [@uhlmann86] which has been used to study finite temperature topologies for thermal mixed states at equilibrium [@viyuela14]. (ii) The gauge invariant interferometric phase defined more recently based on purifications of density matrices [@sjoqvist00].
Purifications:
==============
Purification of a density matrix is defined on an extended Hilbert space constructed by taking the direct product of the original Hilbert space ${\cal H}_0$ with an [*ancillary*]{} Hilbert space ${\cal H}_A$ (whose states do not transform under the operators defined on ${\cal H}_0$). Let there be a density operator $ {\rho}$ defined to act on the Hilbert space ${\cal H}_0$. It can be expanded in it’s eigen-basis as: $ {\rho}=\sum_{i} p_i|\psi_i\rangle\langle\psi_i|$. Now, a normalized state (purification) $|w\rangle\in {\cal H}_p (={\cal H}_{0} \otimes {\cal H}_A)$ of $ {\rho}$ is defined as [uhlmann86]{}
$$|w\rangle=\sum_{i}\sqrt{p_i}| \psi_i\rangle\otimes |\psi_i' \rangle, ~~~{\rm where}~~~ |\psi_i'\rangle\in H_A,$$
so that the original density matrix can be obtained by tracing out the ancillary states $ {\rho}=Tr_A(|w\rangle\langle|w|)$. Under a unitary transformation $U$, the state $|\psi_i(t)\rangle\rightarrow U_0(t)|\psi_i(0)\rangle$. On the contrary, the purifications transform as $$|w(t)\rangle=\sum_{i}\sqrt{p_i}| \psi_i(t)\rangle\otimes |\psi_i' \rangle=U(t)\sum_{i}\sqrt{p_i}| \psi_i(0)\rangle\otimes |\psi_i' \rangle
\label{eq_purification}$$ where $U(t)=U_0(t)\otimes I_A$ and $I_A$ is the identity operator of $H_A$. Comparing with the main text, the index $i$ refers to the two states of the density matrix corrresponding to each mode $k$.
The topological structure:
==========================
Imposing the constraint that $|w(t)\rangle$ are normalized leads to the condition $\frac{d}{dt}\langle w(t)|w(t)\rangle=0$. Now the complete formulation of Pancharatnam and Berry[@pancharatnam56; @berry84] can be extended to these purifications as was shown by Sjöqvist et al [sjoqvist00]{} which leads to a topological index different from that introduced by Uhlmann [uhlmann86,uhlmann89]{}. Let the states $|w(t)\rangle$ be parametrized by a continuous parameter $t$, with $|w(t)\rangle$ describing a curve in the Hilbert space ${\cal H}_p$. A metric is defined in $H_p$ as the measure of distance between two states as $d=|| |w(t_1)\rangle-|w(t_2)\rangle||$.
Let us note that two states are said to be parallel if the distance between them is minimum. But, the purification states $|w(t)\rangle$ also have a phase ambiguity or an $U(1)$ gauge freedom as a gauge transformation $ |w(t)\rangle\rightarrow e^{i\phi}|w(t)\rangle$ produces the same density matrix and preserves inner products in the space ${\cal H}_p$. Now, this phase ambiguity or a $U(1)$ gauge choice needs to be fixed in order to define unique trajectories of $|w(t)\rangle$ in ${\cal H}_p= {\cal H}_0 \bigotimes {\cal H}_A$. In the parallel transport criteria imposed by Pancharatnam[@pancharatnam56; @berry84] for pure states, one fixes the gauge at every point such that two infinitesimally separated states are parallel to each other. For the purifications, the corresponding parallel transport condition can be recast to the form [sjoqvist00]{}: $$\langle w(t)|\dot{w}(t)\rangle=0, \implies Tr( {\rho}(0)U^{\dagger}\dot{U})=0,$$ where in the second step we have used Eq. and $ {\rho}(0)=\sum_{i} p_i | \psi_i (0) \rangle\langle \psi_i (0)|$.
The parallel transport condition fixes the phase at every point of the trajectory in ${\cal H}_p $ and for a transport in time from $0$ to $t$ is given by
$$\phi={\rm Arg} \langle w(t_2)|w(t_1)\rangle \implies \phi={\rm Arg} \left[ {\rm Tr}\left({ \rho}(0) {U}(t)\right) \right],$$
which is the associated Pancharatnam phase. We would like to emphasize, that this phase becomes gauge invariant only when the unitary operator $ {U(t)}$ ensures a parallel transport. Now, consider an [*open*]{} but unitary trajectory being generated by $U(t)$ which can be viewed as the unitary time evolution generated by $H$ (which in the main text becomes the time-independent final Hamiltonian). As in the pure case [berry84,zella12]{}, the unitary operation $U$ introduces a [*dynamical phase*]{} to the purifications in addition to the geometric phase; this dynamical phase must be subtracted from the total phase rendering the geometric part gauge invariant. The dynamical phase accumulated by the purifications over a time $t$ can easily be written as: $$\phi_{\rm dyn}=-\int_0^t \langle w(t)| {H}|w(t)\rangle dt = -\int_0^t {\rm Tr}( {\rho(t) } {H})dt$$ This amounts to a weighted average of dynamical phases of the pure eigenstates of the density matrix and hence, evidently correctly reproduces the pure state limit. The gauge invariant (purely) geometrical phase following a unitary time evolution of a density matrix $$\phi_{g}=\phi-\phi_{\rm dyn}={\rm Arg} \left[ {\rm Tr}\left({ \rho}(0) {U}(t)\right) \right]+\int_0^t {\rm Tr}( {\rho(t)} {H})dt
\label{s_eq_phiG}$$ The conclusion that this phase is gauge invariant follows from the fact that it is induced by a manifestly gauge invariant phase, namely, the Pancharatnam-Berry phase for the purifications $|w(t)\rangle$ and reduces to the pure geometric phase if the parallel transport condition is met (as in this case, the second term in Eq. vanishes). In the pure state limit, the above phase reduces to the Pancharatnam-Berry phase $$\phi_{\rm pure}={\rm Arg}\langle\psi(0)|\psi(t)\rangle+\int_{0}^{t}\langle\psi(t)| {H}|\psi(t)\rangle dt$$ where the pure state density operator, $ {\rho}(t)=|\psi(t)\rangle\langle\psi(t)|$. In the main text, we have considered the temporal evolution of the initial density matrix (characterising a mixed state) generated by the final time-independent Hamiltonian and calculated $\phi$ and $\phi_{\rm dyn}$ for each momentum mode $k$.
[99]{} [pancharatnam56]{} S. Pancharatnam, Proc. Indian Acad. Sci. A [**44**]{}, 247 (1956)
[berry84]{} M. V. Berry, Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences [**392**]{}, 45 (1984).
[zella12]{} Francisco. De Zela, [*The Pancharatnam-Berry Phase: Theoretical and Experimental Aspects*]{}, INTECH Open Access Publisher, 2012.
[uhlmann86]{} A. Uhlmann, Rep. on Math. Phys. [**24**]{}, 229 (1986).
[uhlmann89]{} A. Uhlmann, Annalen der Physik [**501**]{}, 63 (1989).
[viyuela14]{} O. Viyuela, and A. Rivas, and M. A. Martin-Delgado, 2D Mater. 2 034006 (2015) , focus issue on “Artificial Graphene”. Edited by Maciej Lewenstein, Vittorio Pellegrini, Marco Polini and Mordechai (Moti) Segev. [sjoqvist00]{} E.Sjoqvist, A. K. Pati, A. and Ekert, and J. S. Anandan, M. Ericsson, D. K. L. Oi, and V. Vedral, Phys. Rev. Lett. [**85**]{}, 2845 (2000).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recently, Gallego *et.al.* \[[Phys. Rev. Lett [**107**]{}, 210403 (2011)](http://dx.doi.org/10.1103/PhysRevLett.107.210403)\] proved that any future information principle aiming at distinguishing between quantum and post-quantum correlation must be intrinsically multipartite in nature. We establish similar result by using device independent success probability of Hardy’s nonlocality argument for tripartite quantum system. We construct an example of a tri-partite Hardy correlation which is post-quantum but satisfies not only all bipartite information principle but also the GYNI inequality.'
author:
- Subhadipa Das
- Manik Banik
- Ashutosh Rai
- MD Rajjak Gazi
- Samir Kunkri
title: 'Hardy’s nonlocality argument as witness for post-quantum correlation'
---
=1
Introduction
============
Recently, understanding the correlations among distant observers which are compatible with our current description of nature based on quantum mechanics has generated much interest. It has been shown that, some general physical principles can restrict the set of no-signalling correlations among distant observers to a significant degree. Information theoretic principles like information causality [@pawlowski] and nontrivial communication complexity [@vandam1; @vandam2] are novel proposals to single out the quantum correlations, from rest of the no-signalling correlations, when two distant observers (or in some cases even when more than two observers [@yang1]) are involved. However, by applying the known information principles to the bipartite case, it has not been possible to derive the full set of quantum correlations resulting from the Hilbert space structure of quantum mechanics.
For a multipartite (more than two subsystems) scenario, the situation becomes even more complex and extremely challenging. Very recently, interesting results [@gallego; @yang1] have been produced when more than two distant observers are involved. In [@gallego], Gallego *et.al.* provide an example of tripartite no-signalling correlation which is *time-ordered-bi-local* (TOBL) [@pironio; @gallego1; @barrett] and therefore respects any bipartite information principle, yet this correlation is nonquantum (unphysical) since they violate the *guess-your-neighbor’s-input* (GYNI) inequality [@almeida]. Thus, this result demonstrates that in general any bi-principle is insufficient for deriving the set of all multi-party physical (quantum) correlations. A similar example is also provided in the work by Yang *et.al.*[@yang1], where an extremal point of the tripartite non-signaling polytope is proved to be non-quantum. The non-quantumness of this correlation is shown through a violation of an inequality (eqn.-(A1) in [@yang1]) which is satisfied by all quantum correlations; however, in contrast with [@gallego], this correlation respects the GYNI inequality.
In the present work, we give a TOBL correlation which is non-quantum since it exceeds the maximum success probability of Hardy argument for tripartite quantum correlations. To show this, first we prove that in quantum mechanics the maximum success probability of Hardy argument for a tripartite system is $0.125$, which was earlier known only for projective measurement on three qubits system [@ghosh]. Then we explicitly construct a tripartite correlation in a general probabilistic theory with the following properties: (i) the correlation is TOBL and hence satisfies all bi-partite information principle, (ii) the correlation shows Hardy nonlocality with success probability taking the value $0.2$ which is greater than the maximum value $0.125$ that can be achieved within quantum mechanics. Interestingly, this correlation respects all the GYNI inequalities, indicating that any multipartite information principle based on GYNI type of games may still be insufficient in distinguishing physical correlations.
The paper is organized as follows. In section (\[hardy\]) we prove that the device independent success probability for Hardy nonlocality for tripartite quantum system is $\frac{1}{8}$. In section (\[tobl\]), we briefly described the TOBL correlations. In section (\[nonquantum\]), we present a tripartite no signaling probability distribution which belong to TOBL set with success probability for Hardy nonlocality argument greater than that for quantum system, and finally in section (\[conclusion\]) we give our conclusions.
Tripartite Quantum Systems and Hardy Argument {#hardy}
==============================================
Lucian Hardy first provided an argument which reveals nonlocality within quantum mechanics without using any inequality [@hardy]. In order to present Hardy’s argument in the general probabilistic theories, we consider the set of tripartite no-signalling correlations with binary input and binary output for each party—the set of such correlation are known to be points of a polytope in a $26$-dimensional space with $53,856$ extremal points [@pironio]. A tripartite two-input-two-output Hardy correlation is defined by some restrictions on a certain choice of $5$ out of $64$ joint probabilities in the correlation matrix [@kar; @cereceda; @boschi; @kunkri]. The following five conditions, for example, define a tripartite Hardy correlation:
$$\begin{aligned}
P(A_1=+1, ~B_1=+1, ~C_1=+1)> 0 \nonumber \\
P(A_2=+1, ~B_1=+1, ~C_1=+1)=0 \nonumber\\
P(A_1=+1, ~B_2=+1, ~C_1=+1)=0 \\
P(A_1=+1, ~B_1=+1, ~C_2=+1)=0 \nonumber\\
P(A_2=-1, ~B_2=-1, ~C_2=-1)=0 \nonumber\end{aligned}$$
where, $\{A_1,A_2\}$, $\{B_1,B_2\}$ and $\{C_1,C_2\}$ are outcome of respective local observables corresponding to measurements performed by three distant parties, say Alice, Bob and Charlie, and $\pm1$ are the possible measurement outcomes. One can easily show that the above correlation can not be reproduced by any local realistic model.
The joint probability appearing in the first condition, inequality (1), is the success probability for Hardy’s nonlocality argument. In quantum mechanics, for three qubits systems subjected to local projective measurements, the maximum value of the success probability of Hardy argument has been shown to be $0.125$ [@ghosh]. In view of a recent result providing a device independent bound for the success probability of a bipartite Hardy argument [@scarani], one can ask what is the maximum probability of success for Hardy nonlocality for tripartite quantum systems. We extend this result for the tripartite case [@proof]:
**Proposition**: *The maximum value of the success probability of the Hardy argument for tripartite quantum systems is $\frac{1}{8}$.*
**Proof:** To prove the proposition we use an important lemma in [@masanes], which states that, given four projectors $P$, $I-P$, $Q$, $I-Q$ acting on a Hilbert space $\mathbb{H}$, there exists an orthogonal basis which allows to decompose $\mathbb{H}$ as a direct sum of subspaces $\mathbb{H}^i$ of dimension $d\leq2$ for each $i$, such that each of the four projectors can be written as $\Pi=\bigoplus_i \Pi^i$, where $\Pi^i$ acts on $\mathbb{H}^i$ and $\Pi\in\{P, I-P,Q, I-Q\}$.
In quantum mechanics joint probabilities for the outcomes of measurements on a tripartite system are given by $$P(A,B,C|\hat{a},\hat{b},\hat{c})=Tr(\rho M_{A|\hat{a}}\otimes M_{B|\hat{b}}\otimes M_{C|\hat{c}})$$ where $\rho$ is the state of the system and $M_{A|\hat{a}}$, $M_{B|\hat{b}}$, $M_{C|\hat{c}}$ are the measurement operators associated to outcomes $A$, $B$, $C$ of measurements $\hat{a}$, $\hat{b}$, $\hat{c}$ respectively. These measurements operators are POVM, in general; but as we are not restricting the dimension of the Hilbert space, without loss of generality Neumark’s theorem allows us to consider only projective measurements. Each of Alice, Bob and Charlie choose to perform from two measurements, say $\hat{a}=\{\hat{a}_1,\hat{a}_2\}$, $\hat{b}=\{\hat{b}_1,\hat{b}_2\}$ and $\hat{c}=\{\hat{c}_1,\hat{c}_2\}$ respectively with measurement results $\pm1$ for each measurements. Now, $\hat{a}_1=M_{+1|\hat{a}_1}-M_{-1|\hat{a}_1}$ and $\hat{a}_2=M_{+1|\hat{a}_2}-M_{-1|\hat{a}_2}$, where $M_{\pm1|\hat{a}_{1(2)}}$ are projection operators. According to the above mentioned lemma [@masanes], it follows that $M_{A|\hat{a}}=\bigoplus_iM^i_{A|\hat{a}}$ for all $A$ and $\hat{a}$. Let us denote $M^i=M^i_{+1|\hat{a}}+M^i_{-1|\hat{a}}$ for all $\hat{a}$. Also this lemma is valid for Bob’s and Charlie’s end. Using analogous notations for Bob’s and Charlie’s operators, we can write $$\begin{aligned}
P(A,B,C|\hat{a},\hat{b},\hat{c})&=&\sum_{i,j,k}q_{ijk}Tr(\rho_{ijk}M^i_{A|\hat{a}}\otimes M^j_{B|\hat{b}}\otimes M^k_{C|\hat{c}})\nonumber \\
&=&\sum_{i,j,k}q_{ijk}P_{ijk}(A,B,C|\hat{a},\hat{b},\hat{c})\end{aligned}$$ where, $q_{ijk}=Tr(\rho M^i_{A|\hat{a}}\otimes M^j_{B|\hat{b}}\otimes M^k_{C|\hat{c}})$ and $\rho_{ijk}=(M^i\otimes M^j \otimes M^k \rho M^i\otimes M^j \otimes M^k)/q_{ijk}$ is a three qubit state. As $q_{ijk}\geq0$ for all $i,j,k$ and $\sum_{i,j,k}q_{ijk}=1$ the Hardy constraints (last four condition of eqn. (1)) are satisfied for $P$ if and only if they are satisfied for each of the $P_{ijk}$. Which implies that $$\begin{aligned}
P(A_1=+1, ~B_1=+1, ~C_1=+1)\nonumber\\=\sum_{i,j,k}q_{ijk}P_{ijk}(A^{i}_1=+1, ~B^{j}_1=+1, ~C^{k}_1=+1)\end{aligned}$$ Being a convex sum, the success probability Hardy argument is therefore less or equal to the largest element in the combination, which is $\frac{1}{8}$. $\Box$
In the remaining part of this paper, for convenience we adopt another notation for the joint correlation probabilities expressed as $P(abc|xyz)$ with input $x,y,z\in\{0,1\}$ and output $a,b,c\in\{0,1\}$.
Time-ordered Bi-local(TOBL) Correalions {#tobl}
=======================================
A tripartite no-signaling probability distribution $P(abc|xyz)$ belongs to TOBL [@pironio; @gallego1; @barrett] if it can be written as $$\begin{aligned}
P(abc|xyz)=\sum_{\lambda}p_{\lambda}P(a|x,\lambda)P_{B \rightarrow C}(bc|yz,\lambda)\\
=\sum_{\lambda}p_{\lambda}P(a|x,\lambda)P_{B \leftarrow C}(bc|yz,\lambda)\end{aligned}$$ and analogously for $B|AC$ and $C|AB$, where $p_{\lambda}$ is the distribution of some random variable $\lambda$, shared by the parties. The distributions $P_{B \rightarrow C}$ and $P_{B \leftarrow C}$ respect the conditions $$\begin{aligned}
P_{B \rightarrow C}(b|y,\lambda)=\sum_{c}P_{B \rightarrow C}(bc|yz,\lambda)\\
P_{B \leftarrow C}(c|z,\lambda)=\sum_{b}P_{B \leftarrow C}(bc|yz,\lambda)\end{aligned}$$ From these equations it is clear that the distributions $P_{B \rightarrow C}$ allow signaling from Alice to Bob and $P_{B \leftarrow C}$ allow signaling from Bob to Alice. If a tripartite no-signaling probability distribution $P(abc|xyz)$ belongs to the set of TOBL distributions, all possible bipartite distributions derived by applying any local wirings on $P(abc|xyz)$ are local, i.e., the probability distribution $P(abc|xyz)$ respects all bi-partite physical principle [@gallego1].
non quantum correlation satisfying bipartite principle {#nonquantum}
======================================================
Now we show that, there exist a tripartite Hardy correlation which lies in the TOBL set. Since, this correlation belongs to the TOBL set, it must satisfy all bipartite information principle. The probability distribution is given in the TABLE-I.
[lllllllll]{} $ xyz\backslash abc$ & $000$ & $001$ & $010$ & $011$ & $100$ & $101$ & $110$ & $111$\
\
$000$ & $\frac{1}{5}$ & $0$ & $0$ & $\frac{1}{5}$ & $0$ & $\frac{1}{5}$ & $\frac{1}{5}$ & $\frac{1}{5}$\
\
$001$ & $0$ & $\frac{1}{5}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{2}{5}$ & $0$\
\
$010$ & $0$ & $\frac{1}{10}$ & $\frac{1}{5}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{2}{5}$ & $\frac{1}{10}$ & $0$\
\
$011$ & $0$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{5}$ & $\frac{2}{5}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $0$\
\
$100$ & $0$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{2}{5}$ & $\frac{1}{5}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $0$\
\
$101$ & $0$ & $\frac{1}{10}$ & $\frac{2}{5}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{5}$ & $\frac{1}{10}$ & $0$\
\
$110$ & $0$ & $\frac{2}{5}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{10}$ & $\frac{1}{5}$ & $0$\
\
$111$ & $\frac{2}{5}$ & $0$ & $0$ & $\frac{1}{5}$ & $0$ & $\frac{1}{5}$ & $\frac{1}{5}$ & $0$\
\
\[TABLE I\]
In this table, $ P(000|000)=\frac{1}{5},~P(000|001)=0,~P(000|100)=0,~P(000|010)=0,~P(111|111)=0$. The success probability of Hardy argument for this correlation is $\frac{1}{5}$ which is strictly larger than the maximum achievable quantum value $\frac{1}{8}$ and thus this is a nonquantum correlation.
To prove that the distribution $P(abc|xyz)$ belongs to the TOBL set we show that it admits TOBL decomposition for all bi-partition. Note that the correlation considered in the TABLE-I is symmetric under any permutation of the parties, so here it it is sufficient to provide a TOBL model in any one bipartition, say $A|BC$.
Probability distribution appearing in the TOBL decomposition for the bi-partition $A|BC$, are such that for a given $\lambda$ Alice’s outcome $a$ depends only her measurement settings $x$. Also, for given $\lambda$, $P_{B \rightarrow C}(b|y,\lambda)$ is independent of $z$ but for $B \rightarrow C$, $c$ depends on both $ y$ and $z$. Similarly, for given $\lambda$, $P_{B\leftarrow C}(c|z,\lambda)$ is independent of $y$ but for $B\leftarrow C$, $b$ depends on both $ y$ and $z$. Let $a_x$, $b_y$ and $c_{z}$ denote the outcomes for Alice, Bob and Charlie for the respective inputs $x$, $y$ and $z$. In the following TABLE-II and TABLE-III , the outputs are deterministic and the weights $p_\lambda$ are same. For any given $\lambda$, the outcome assignments of $A$ in both the tables are same.
----- ----------- ----------------- -------- ------- ------- ------- ---------- ---------- ---------- ----------
$ $ $\lambda$ $p_\lambda $ $a_0$ $a_1$ $b_0$ $b_1$ $c_{00}$ $c_{01}$ $c_{10}$ $c_{11}$
$ $ $1$ $\frac{1}{10}$ $0$ $0$ $1$ $1$ $1$ $0$ $1$ $1$
$ $ $2$ $\frac{1}{10}$ $0$ $0$ $1$ $1$ $1$ $1$ $0$ $1$
$ $ $3$ $\frac{1}{10}$ $1$ $0$ $0$ $0$ $1$ $1$ $1$ $0$
$ $ $4$ $\frac{1}{10}$ $1$ $0$ $1$ $0$ $0$ $0$ $1$ $0$
$ $ $5$ $\frac{1}{5}$ $1$ $0$ $1$ $0$ $1$ $0$ $1$ $0$
$ $ $6$ $\frac{1}{10}$ $0$ $1$ $0$ $0$ $0$ $1$ $1$ $1$
$ $ $7$ $\frac{1}{10}$ $0$ $1$ $0$ $1$ $0$ $1$ $0$ $0$
$ $ $8$ $\frac{1}{10}$ $1$ $1$ $1$ $0$ $0$ $0$ $0$ $1$
$ $ $9$ $\frac{1}{10}$ $1$ $1$ $0$ $1$ $1$ $0$ $0$ $0$
----- ----------- ----------------- -------- ------- ------- ------- ---------- ---------- ---------- ----------
: TOBL decomposition for the case $A|B \rightarrow C$.
\[TABLE II\]
----- ----------- ----------------- -------- ------- ---------- ---------- ---------- ---------- --------- ---------
$ $ $\lambda$ $p_\lambda $ $a_0$ $a_1$ $b_{00}$ $b_{01}$ $b_{10}$ $b_{11}$ $c_{0}$ $c_{1}$
$ $ $1$ $\frac{1}{10}$ $0$ $0$ $1$ $0$ $1$ $1$ $1$ $1$
$ $ $2$ $\frac{1}{10}$ $0$ $0$ $1$ $1$ $0$ $1$ $1$ $1$
$ $ $3$ $\frac{1}{10}$ $1$ $0$ $1$ $1$ $1$ $0$ $0$ $0$
$ $ $4$ $\frac{1}{10}$ $1$ $0$ $0$ $1$ $0$ $0$ $1$ $0$
$ $ $5$ $\frac{1}{5}$ $1$ $0$ $1$ $1$ $0$ $0$ $1$ $0$
$ $ $6$ $\frac{1}{10}$ $0$ $1$ $0$ $1$ $1$ $1$ $0$ $0$
$ $ $7$ $\frac{1}{10}$ $0$ $1$ $0$ $0$ $1$ $0$ $0$ $1$
$ $ $8$ $\frac{1}{10}$ $1$ $1$ $0$ $0$ $0$ $1$ $1$ $0$
$ $ $9$ $\frac{1}{10}$ $1$ $1$ $1$ $0$ $0$ $0$ $0$ $1$
----- ----------- ----------------- -------- ------- ---------- ---------- ---------- ---------- --------- ---------
: TOBL decomposition for the case $A|B \leftarrow C$.
\[TABLE III\]
It is interesting to observe that the correlation in TABLE-I satisfies the most general GYNI inequality [@almeida] (see also [@almeida1]) whose violation certifies non-quantumness of a correlation. Thus the post-quantum nature (which is guaranteed by the larger success probability for Hardy’s argument compared to quantum result) of the correlation given in TABLE I can not be witnessed by violating the GYNI inequality as happened for the correlation given in [@gallego]. Therefore the nonlocality of this post-quantum correlation is qualitatively different from the nonlocality of the post-quantum correlation that appeared in [@gallego].
Conclusions {#conclusion}
===========
Distinguishing physically realized correlations from unphysical ones by some fundamental principle is an active area of research in the foundational perspective. Rather it has been proved that nonlocality is useful resource for device independent cryptography [@acin]. So it is very important to know which nonlocal correlations can be obtained by physical means. Like celebrated Bell inequality [@bell], the elegant argument of Hardy [@hardy] reveals the nonlocality of quantum mechanics. Again like the device independent value of Bell violation for bipartite quantum mechanical system i.e. Cirel’son bound [@cirel], the optimal success probability for Hardy’s nonlocality argument [@scarani] in quantum mechanics could be a potential witness for detecting post-quantum no signaling correlations. We derive the device independent success probability of Hardy’s argument for tripartite quantum system. Then we provide an explicit tripartite correlation, which satisfies any bipartite information principles, but shows Hardy nonlocality with probability which is post-quantum. In this way we establish that this device independent value is a potential witness for tripartite post-quantum correlations. Our example also satisfies most general GYNI inequality indicating insufficiency of multipartite information principle based on GYNI game in identifying physical correlations. It shows the intricate structure of physically allowed correlations when more than two spatially separated observers are involved.
It is a pleasure to thank Guruprasad Kar for motivating us in this direction. We also thank Sibasish Ghosh and Ramij Rahaman for many stimulating discussion regarding Hardy’s nonlocality argument. SD and AR acknowledge support from the DST project SR/S2/PU-16/2007.
[99]{} M. Pawlowski,T. Paterek, D. Kaszlikowski, V. Scsrani, A.Winter and M. Zukowski, [Nature [**461**]{}, 1101 (2009).](http://dx.doi.org/10.1038/nature08400) W. van Dam, [Nonlocality and Communication complexity, Ph.D. thesis, University of Oxford (2000).](http://cs.ucsb.edu/~vandam/oxford_thesis.pdf) W. van Dam, [arXiv:0501159 \[quant-ph\] (2005).](http://arxiv.org/pdf/quant-ph/0501159.pdf) R. Gallego, L. Erik Wurflinger, A. Acin and M. Navascues, [Phys. Rev. Lett [**107**]{}, 210403 (2011).](http://dx.doi.org/10.1103/PhysRevLett.107.210403) T. H. Yang, D. Cavalcanti, M. L. Almeida, C. Teo and V. Scarani, [New J. Phys. [**14**]{}, 013061 .](http://dx.doi.org/10.1088/1367-2630/14/1/013061) S. Pironio, J. D. Bancal and V. Scarani, [J. Phys. A:Math. Theor. [**44**]{} 065303 (2011).](http://dx.doi.org/doi:10.1088/1751-8113/44/6/065303 ) R. Gallego, L. Erik Wurflinger, A. Acin and M. Navascues, [Phys. Rev. Lett. 109, 070401 (2012)](http:10.1103/PhysRevLett.109.070401) J. Barrett, S. Pironio, J. Bancal and N. Gisin, [arXiv:1112.2626 \[quant-ph\] (2011).](http://arxiv.org/abs/1112.2626) M. L. Almeida *et.al.*, [Phys. Rev. Lett.[**104**]{}, 230404 (2010).](http://dx.doi.org/10.1103/PhysRevLett.104.230404) S. K. Choudhary *et.al.*, [Quantum Information and Computation, [**Vol. 10**]{}, No. 9 and 10 (2010) 0859–0871.](http://arxiv.org/abs/0807.4414) L. Hardy, [Phys. Rev. Lett. [**68**]{}, 2981 (1992)](http://dx.doi.org/10.1103/PhysRevLett.68.2981); L. Hardy, [Phys. Rev. Lett. [**71**]{}, 1665 (1993).](http://dx.doi.org/10.1103/PhysRevLett.71.1665) G. Kar, [Phys. Rev. A, [**56**]{}, 1023 (1997)](http://dx.doi.org/10.1103/PhysRevA.56.1023); J. Cereceda,[ Phys. Lett. A [**327**]{}, 433 (2004)](http://dx.doi.org/10.1016/j.physleta.2004.06.004); D. Boschi, S. Branca, F. De Martini, L. Hardy, [Phys.Rev. Lett. [**79**]{}, 2755 (1997)](http://dx.doi.org/10.1103/PhysRevLett.79.2755); S. Kunkri, S. K. Choudhary, [Phys. Rev. A [**72**]{}, 022348(2005).](http://dx.doi.org/10.1103/PhysRevA.72.022348) R. Rabelo, L. Y. Zhi and V. Scarani, [arXiv:1205.3280v2 \[quant-ph\] (2011).](http://arxiv.org/abs/1205.3280v2) In fact, this result can be stated in a more general way—If an N-qubits system subjected to local projective measurements exhibit Hardy’s nonlocality with the maximum success probability, say $\alpha$, then for any N-partite quantum system subjected to general dichotomic local measurements, the maximum success probability of Hardy’s nonlocality argument cannot exceeds the value $\alpha$. For $N=2$ and $N=3$ the respective values of $\alpha$ are $0.09$ and $\frac{1}{8}$. L. Masanes, [Phys. Rev. Lett. [**97**]{}, 050503 (2006).](http://dx.doi.org/10.1103/PhysRevLett.97.050503) A. Acin, M. L. Almeida, R. Augusiak and N. Brunner, [arXiv:1205.3076v1 \[quant-ph\] (2012).](http://arxiv.org/abs/1205.3076v1) A. Acin, N. Brunner, N. Gisin, S. Massar, S. Pironio and V. Scarani, [Phys. Rev. Lett. [**98**]{}, 230501 (2007).](http://dx.doi.org/10.1103/PhysRevLett.98.230501) J. S. Bell, Speakable and Unspeakable in Quantum Mechanics: Collected papers on quantum philosophy, Cambridge University Press. B. S. Cirel’son, [Lett. Math. Phys. [**4**]{}, 93 (1980).](http://www.springerlink.com/content/l57053g573430450/)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recent work has validated the importance of subword information for word representation learning. Since subwords increase parameter sharing ability in neural models, their value should be even more pronounced in low-data regimes. In this work, we therefore provide a comprehensive analysis focused on the usefulness of subwords for word representation learning in truly low-resource scenarios and for three representative morphological tasks: fine-grained entity typing, morphological tagging, and named entity recognition. We conduct a systematic study that spans several dimensions of comparison: **1)** *type of data scarcity* which can stem from the lack of task-specific training data, or even from the lack of unannotated data required to train word embeddings, or both; **2)** *language type* by working with a sample of 16 typologically diverse languages including some truly low-resource ones (e.g. Rusyn, Buryat, and Zulu); **3)** the choice of the *subword-informed word representation method*. Our main results show that subword-informed models are universally useful across all language types, with large gains over subword-agnostic embeddings. They also suggest that the effective use of subwords largely depends on the language (type) and the task at hand, as well as on the amount of available data for training the embeddings and task-based models, where having sufficient in-task data is a more critical requirement.'
author:
- |
Yi Zhu^1^[^1], Benjamin Heinzerling^2,3$\ast$^, Ivan Vulić^1^,\
[**Michael Strube^4^, Roi Reichart^5^, Anna Korhonen^1^**]{}\
^1^Language Technology Lab, University of Cambridge\
^2^RIKEN AIP, ^3^Tohoku University\
^4^Heidelberg Institute for Theoretical Studies, ^5^Technion, IIT\
[{yz568,iv250,alk23}@cam.ac.uk]{}, [[email protected]]{}\
[[email protected]]{}, [[email protected]]{}
bibliography:
- 'references.bib'
title: |
On the Importance of Subword Information for\
Morphological Tasks in Truly Low-Resource Languages
---
Introduction and Motivation {#s:introduction}
===========================
Methodology {#s:methodology}
===========
Evaluation Tasks {#s:tasks}
================
Experimental Setup {#s:experimental}
==================
Results and Discussion {#s:results}
======================
Conclusions and Future Work {#s:conclusion}
===========================
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the ERC Consolidator Grant LEXICAL: Lexical Acquisition Across Languages (no 648909) and the Klaus Tschira Foundation, Heidelberg, Germany. We thank the three anonymous reviewers for their helpful suggestions.
[^1]: Equal contribution, work partly done while at HITS.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'refs.bib'
nocite:
- '[@Cohen:2003xe]'
- '[@Cohen:2003qw]'
- '[@Giedt:2003ve; @Onogi:2005cz]'
- '[@Ohta:2006qz]'
- '[@Catterall:2004np]'
- '[@Sugino:2004qd]'
- '[@Sugino:2004uv]'
- '[@D''Adda:2005zk]'
- '[@Kaplan:1983sk]'
- '[@Maru:1997kh]'
- '[@Neuberger:1997bg]'
- '[@Kaplan:1999jn]'
- '[@Fleming:2000fa]'
- '[@Cohen:2003xe]'
- '[@Cohen:2003qw]'
---
[**Exact Vacuum Energy of Orbifold Lattice Theories**]{}
\
[*The Niels Bohr Institute,\
The Niels Bohr International Academy,\
Blegdamsvej 17, DK-2100 Copenhagen, Denmark*]{}
[**Abstract**]{}
We investigate the orbifold lattice theories constructed from supersymmetric Yang-Mills matrix theories (mother theories) with four and eight supercharges. We show that the vacuum energy of these theories does not receive any quantum correction perturbatively.
Introduction
============
Recently, there has been a rapid development in supersymmetric lattice gauge theories. A systematic way to construct supersymmetric lattice formulations is developed in [@Kaplan:2002wv]–[@Kaplan:2005ta], where a space-time lattice is generated by an orbifold projection of a supersymmetric Yang-Mills matrix theory (mother theory), and a lattice spacing is introduced by “deconstruction” [@Arkani-Hamed:2001ca]. By choosing the orbifold projection properly, one can make at least one supercharge or BRST charge preserved on the lattice. These formulations are further analysed in [@Giedt:2003xr]–[@Damgaard:2007be][^1]. A prescription to generate a lattice theory from a topologically twisted continuum supersymmetric gauge theory is proposed by Catterall [@Catterall:2003wd]–[@Catterall:2005fd]. In these formulations, the BRST charge of the continuum theory is preserved on the lattice. A characteristic feature of these formulations is that all the degrees of freedom on the lattice except for site variables are doubled by a complexification and the path-integral is performed along “the real line”. Numerical simulations are carried out for the model of two-dimensional ${{\cal N}}=(2,2)$ supersymmetric gauge theory [@Catterall:2006jw], which reproduce the Ward-Takahashi identities in fairly good accuracy. Other formulations constructed from topologically twisted supersymmetric gauge theories are developed by Sugino [@Sugino:2003yb]–[@Sugino:2006uf], where it is shown that the BRST transformation for the continuum fields can also be defined for lattice variables. The lattice action is straightforwardly generated from the $Q$-exact form of the continuum action by replacing all the fields by the lattice variables. A common feature of the above three formulations is that they possess at least one preserved supercharge or BRST charge. Alternative approach (the link approach) has been developed in [@D'Adda:2004jb]–[@D'Adda:2007ax], where it is claimed that all the supersymmetry of the continuum theory is preserved on the lattice. They first explicitly construct a supersymmetry algebra on a lattice and next make a lattice action based on the algebra, although there are some discussions on this approach [@Bruckmann:2006ub][@Bruckmann:2006kb]. For conventional but useful approaches to supersymmetric lattice gauge theories, see [@Nishimura:1997vg]–[@Montvay:2001aj] in which the theories do not have any supersymmetry on a lattice but they flow to supersymmetric theories without fine-tuning because of a discrete chiral symmetry on the lattice. See also [@Suzuki:2007jt] for a recent lattice approach to two-dimensional ${{\cal N}}=(2,2)$ supersymmetric gauge theory.
The above seemingly different supersymmetric lattice formulations with a supercharge on the lattice are related to the orbifold lattice theories. In fact, the prescription given by Catterall can be reproduced using the orbifolding procedure [@Damgaard:2007xi]. Sugino’s formulations can be obtained from Catterall’s formulations by restricting the degrees of freedom of the complexified fields with preserving the supercharge [@Takimi:2007nn]. Furthermore, the formulations given by the link approach have been shown to be equivalent to those given by orbifolding [@Damgaard:2007eh]. In this sense, it seems important to examine quantum mechanical properties of the orbifold lattice theories. In the next section, we examine the vacuum energy of the orbifold lattice theories constructed from $Q=4$ and $Q=8$ mother theories. We show that the vacuum energy exactly vanishes to all orders of the perturbation theory and the flat directions of these theories are never lifted up by any perturbative effect. The final section is devoted to conclusion and discussion.
Quantum Corrections to Vacuum Energy
====================================
Orbifold lattice theories from $Q=4$ mother theory {#Q=4}
--------------------------------------------------
As discussed in detail in [@Kaplan:2002wv]–[@Kaplan:2005ta], an orbifold lattice theory is obtained by performing an appropriate orbifold projection to a supersymmetric Yang-Mills matrix theory (mother theory) followed by deconstruction, that is, by expanding the orbifolded matrix theory around a classical vacuum. Let us start with the orbifold lattice theories constructed from the dimensionally reduced four-dimensional ${{\cal N}}=1$ supersymmetric Yang-Mills theory [@Cohen:2003xe]. As discussed in [@Damgaard:2007be], the lattice gauge theory obtained from this mother theory is essentially unique to be a lattice formulation for two-dimensional ${{\cal N}}=(2,2)$ supersymmetric Yang-Mills theory[^2]. The action of the orbifolded matrix theory (before deconstruction) is given by $$\begin{aligned}
S_{\rm orb} &=
\frac{1}{g^2}{\rm Tr}\sum_{{{\mathbf n}}\in {\mathbb{Z}}_N^2}\Biggl(
\frac{1}{4}\Bigl|z_m({{{\mathbf n}}})
z_n({{\mathbf n}}+{{\mathbf e}}_m) - z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr|^2 \cr
&
+ \frac{1}{8}\Bigl(z_m({{\mathbf n}})\bar{z}_m({{\mathbf n}})-\bar{z}_m({{\mathbf n}}-{{\mathbf e}}_m)z_m({{\mathbf n}}-{{\mathbf e}}_m)\Bigr)^2 \cr
& + \psi_m({{\mathbf n}})\Bigl(\bar{z}_m({{\mathbf n}})\eta({{\mathbf n}})-\eta({{\mathbf n}}+{{\mathbf e}}_m)
\bar{z}_m({{\mathbf n}})\Bigr)\cr
& - \frac{1}{2}\chi_{mn}({{\mathbf n}})\Bigl(z_m({{\mathbf n}})\psi_n({{\mathbf n}}+{{\mathbf e}}_n)
-\psi_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)
-(m\leftrightarrow n)\Bigr)\Biggr) ~,
\label{orbifold action Q=4}\end{aligned}$$ where $m,n=1,2$, ${{\mathbf e}}_m$ are two linearly independent integer valued two-vectors, and all the fields are complex matrices with the size $M$. Although this action does not contain any lattice spacing nor kinetic terms, we can regard it as a lattice action by identifying ${{\mathbf n}}$ as the label of a site on a two-dimensional square lattice with the size $N$. In this sense, the variables $z_m({{\mathbf n}})$ and ${\bar{z}}_m({{\mathbf n}})$ are bosonic fields living on the links $({{\mathbf n}},{{\mathbf n}}+{{\mathbf e}}_m)$ and $({{\mathbf n}}+{{\mathbf e}}_m,{{\mathbf n}})$, respectively, and $\eta({{\mathbf n}})$, $\psi_m({{\mathbf n}})$ and $\chi_{12}({{\mathbf n}})=-\chi_{21}({{\mathbf n}})$ are fermionic fields living on the site ${{\mathbf n}}$, the link $({{\mathbf n}},{{\mathbf n}}+{{\mathbf e}}_m)$ and the link $({{\mathbf n}}+{{\mathbf e}}_1+{{\mathbf e}}_2,{{\mathbf n}})$, respectively. Note the action (\[orbifold action Q=4\]) is invariant under a $U(M)$ “gauge transformation” $z_m({{\mathbf n}})\to g^{-1}({{\mathbf n}})z_m({{\mathbf n}})g({{\mathbf n}}+{{\mathbf e}}_m)$ $(g({{\mathbf n}})\in U(M))$, and so on. As mentioned above, kinetic terms and a lattice spacing $a$ are introduced by expanding $z_m({{\mathbf n}})$ and ${\bar{z}}_m({{\mathbf n}})$ as $$z_m({{\mathbf n}}) = \frac{1}{a}{\mathbf 1}_M + z'_m({{\mathbf n}}),
\quad {\bar{z}}_m({{\mathbf n}}) = \frac{1}{a}{\mathbf 1}_M + {\bar{z}}'_m({{\mathbf n}}),
\label{shift}$$ then we obtain a lattice formulation for two-dimensional ${{\cal N}}=(2,2)$ supersymmetric Yang-Mills theory. Since the potential terms of this theory are given by $$\frac{1}{4}\Bigl|z'_m({{{\mathbf n}}})
z'_n({{\mathbf n}}+{{\mathbf e}}_m) - z'_n({{\mathbf n}})z'_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr|^2
+ \frac{1}{8}\Bigl(z'_m({{\mathbf n}})\bar{z'}_m({{\mathbf n}})-\bar{z'}_m({{\mathbf n}}-{{\mathbf e}}_m)z'_m({{\mathbf n}}-{{\mathbf e}}_m)\Bigr)^2,$$ the classical moduli space (the flat directions) of this theory is parametrized by the vacuum expectation values of $z'_m({{\mathbf n}})$ and ${\bar{z}}'_m({{\mathbf n}})$, $$z'_m({{\mathbf n}})=\left(
\begin{matrix}b_m^1 & & \\ & \ddots & \\ && b_m^M\end{matrix}
\right)\equiv b_m, \qquad
{\bar{z}}'_m({{\mathbf n}})=\left(
\begin{matrix}\bar{b}_m^1 & & \\ & \ddots & \\ && \bar{b}_m^M\end{matrix}
\right)\equiv \bar{b}_m,
\label{lattice vacuum}$$ with $b_m^i\in {\mathbb{C}}$ ($i=1,\cdots,M$) up to gauge transformations.
In this paper, we are interested in quantum corrections to this classical moduli space. To examine them, we will estimate the vacuum energy at the point (\[lattice vacuum\]) in the classical moduli space. Perturbatively, this is achieved by expanding the lattice action (after deconstruction) around the vacuum (\[lattice vacuum\]) and summing up all 1PI vacuum graphs. However, recalling that the action of the lattice gauge theory is obtained by substituting (\[shift\]) into the action of the orbifolded matrix theory (\[orbifold action Q=4\]), we see that the same result is obtained by directly replacing $z_m({{\mathbf n}})$ and ${\bar{z}}_m({{\mathbf n}})$ in the action (\[orbifold action Q=4\]) with $$\begin{aligned}
z_m({{\mathbf n}})&\to z_m({{\mathbf n}})+\frac{1}{a}{\mathbf 1}_M + b_m
\equiv z_m({{\mathbf n}})+ a_m, {\nonumber}\\
{\bar{z}}_m({{\mathbf n}})&\to {\bar{z}}_m({{\mathbf n}})+\frac{1}{a}{\mathbf 1}_M + {\bar{b}}_m
\equiv {\bar{z}}_m({{\mathbf n}})+{\bar{a}}_m,
\label{comment}\end{aligned}$$ respectively. In the following calculation, we will use this notation and estimate the vacuum energy as a function of $a_m^i$ $(i=1,\cdots,M)$.
We first calculate the 1-loop vacuum energy. It is convenient to fix the gauge by imposing a gauge condition, $$D^-_m {\bar{z}}_m({{\mathbf n}}) - \bar{D}^-_m z_m({{\mathbf n}}) =0,
\label{gauge fix}$$ where $$\begin{aligned}
D^-_m f({{\mathbf n}})\equiv
a_m f({{\mathbf n}}) - f({{\mathbf n}}-{{\mathbf e}}_m) a_m, \qquad
\bar{D}^-_m f({{\mathbf n}}) \equiv
-{\bar{a}}_m f({{\mathbf n}}-{{\mathbf e}}_m) + f({{\mathbf n}}) {\bar{a}}_m.
\label{D-}\end{aligned}$$ For the purpose of the later discussion, we also define $$\begin{aligned}
D^+_m f({{\mathbf n}})\equiv
a_m f({{\mathbf n}}+{{\mathbf e}}_m) - f({{\mathbf n}}) a_m, \qquad
\bar{D}^+_m f({{\mathbf n}})\equiv
-{\bar{a}}_m f({{\mathbf n}}) + f({{\mathbf n}}+{{\mathbf e}}_m) {\bar{a}}_m.
\label{D+}\end{aligned}$$ By introducing gauge fixing terms and FP ghost fields corresponding to the gauge condition (\[gauge fix\]) in a standard way, the second-order action is obtained as $$\begin{aligned}
S^{(2)} = \frac{1}{g^2}{{\rm Tr}\,}\sum_{{{\mathbf n}}\in{\mathbb{Z}}_N^2}\Biggl(
&\frac{1}{2}\bar{D}^-_n z_m({{\mathbf n}}) D^-_n z_m({{\mathbf n}})
+\frac{1}{2}\bar{D}^-_n B({{\mathbf n}}) D^-_n C({{\mathbf n}}) {\nonumber}\\
&+\eta({{\mathbf n}}) \bar{D}^-_m \psi_m({{\mathbf n}})
-\frac{1}{2}\chi_{mn}({{\mathbf n}})\Bigl(D^+_m \psi_n({{\mathbf n}})- D^+_n \psi_m({{\mathbf n}}) \Bigr)
\Biggr),
\label{2nd order action}\end{aligned}$$ where $B({{\mathbf n}})$ and $C({{\mathbf n}})$ are FP ghost fields. By integrating over the fields, we get the 1-loop contribution to the partition function as[^3] $$\begin{aligned}
Z\Bigr|_{\rm 1-loop} &= \int \prod_{{{\mathbf n}}}d\Phi({{\mathbf n}})
e^{-S^{(2)}[\Phi({{\mathbf n}})]} {\nonumber}\\
&= \frac{\det \Delta}{\det \Delta} =1,
\label{1-loop result}\end{aligned}$$ where $\Delta\equiv \sum_m \bar{D}_m^+ D_m^-$ is the Laplacian and the lattice variables have been abbreviated as $\Phi({{\mathbf n}})$. The denominator of the second line comes from the contributions from the bosonic fields and the ghost fields and the numerator comes from the fermionic fields. The result (\[1-loop result\]) means that the vacuum energy is equal to zero and the classical flat directions remain flat at the 1-loop level. Note that the same calculation is carried out at the origin of the moduli space in [@Onogi:2005cz]. We can reproduce it by setting $b_m^i=0$ (or $a_m^i=1/a$) in our calculation.
One might think that, even though the 1-loop contribution to the vacuum energy is zero, higher-loop contributions would give non-trivial corrections to the vacuum energy, since the supersymmetry is almost broken except for the only one preserved supercharge (or BRST charge). However, we can show that it is not the case and the above 1-loop result is exact. The key point is that the action (\[orbifold action Q=4\]) can be written in a $Q$-exact form [@Cohen:2003xe]: $$\begin{aligned}
S_{\rm orb} =
\frac{1}{2g^2}{\rm Tr}\sum_{{{\mathbf n}}\in {\mathbb{Z}}_N^2}Q\Biggl(
& \eta({{\mathbf n}})\Bigl({z}_m({{\mathbf n}}) {\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m){z}_m({{\mathbf n}}-{{\mathbf e}}_m)+d({{\mathbf n}})\Bigr) {\nonumber}\\
& - \chi_{mn}({{\mathbf n}})\Bigl(z_m({{\mathbf n}})z_n({{\mathbf n}}+{{\mathbf e}}_n)
-z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr)\Biggr) ~,
\label{Q-exact action Q=4}\end{aligned}$$ where $Q$ is a BRST charge that acts on the fields as $$\begin{aligned}
Q z_m({{\mathbf n}}) &= \psi_m({{\mathbf n}}), \quad Q {\bar{z}}_m({{\mathbf n}}) = 0, {\nonumber}\\
Q d({{\mathbf n}}) &= \psi_m({{\mathbf n}}){\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m)\psi_m({{\mathbf n}}-{{\mathbf e}}_m), {\nonumber}\\
\label{BRST Q=4}
Q\eta({{\mathbf n}}) &=
\frac{1}{4}\Bigl(z_m({{\mathbf n}}){\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m)z_m({{\mathbf n}}-{{\mathbf e}}_m)-d({{\mathbf n}})\Bigr), \\
Q\chi_{mn}({{\mathbf n}}) &= \frac{1}{2}\Bigl({\bar{z}}_m({{\mathbf n}}+{{\mathbf e}}_n){\bar{z}}_n({{\mathbf n}})
-{\bar{z}}_n({{\mathbf n}}+{{\mathbf e}}_m){\bar{z}}_m({{\mathbf n}})\Bigr), {\nonumber}\end{aligned}$$ and $d({{\mathbf n}})$ is an auxiliary bosonic field which makes $Q$ be nilpotent off-shell. Recalling the discussion in topological field theory [@Witten:1988ze], we see that the partition function of this theory does not depend on the coupling constant $g$. In fact, if we write the partition function as $Z(g)=\int {{\cal D}}\Phi e^{\frac{1}{g^2}Q\Xi[\Phi]}$, the derivative of the partition function by $g$ gives $$\frac{d}{dg} Z(g) \propto \Bigl\langle Q\Xi[\Phi] \Bigr\rangle=0,
\label{g-independence}$$ where $\langle {{\cal O}}\rangle$ denotes the expectation value of an operator ${{\cal O}}$ and we have used the fact that, as long as the BRST symmetry is not broken spontaneously, the expectation value of a $Q$-exact operator vanishes when the action is $Q$-exact. This means that the partition function evaluated in the weak coupling limit, that is, the 1-loop result given above is exact. In particular, we can expect that all the higher-loop contributions to the vacuum energy vanish.
Note that one might think that the partition function given above expresses not the vacuum energy but the Witten index of the theory since we impose the periodic boundary condition to the fermionic fields in the time direction. Although it is actually the case, the boundary conditions do not affect the perturbative contributions in the limit that the period of the time direction goes to infinity. Therefore we can conclude that there is no perturbative correction to the vacuum energy from (\[g-independence\]).[^4]
Another note is that we can apply the same analysis to a deformed theory given by[^5] $$\begin{aligned}
S_{\rm orb} =
\frac{1}{2g^2}{\rm Tr}\sum_{{{\mathbf n}}\in {\mathbb{Z}}_N^2}Q\Biggl(
& \eta({{\mathbf n}})\Bigl({z}_m({{\mathbf n}}) {\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m){z}_m({{\mathbf n}}-{{\mathbf e}}_m)+d({{\mathbf n}})\Bigr) {\nonumber}\\
& -\beta \chi_{mn}({{\mathbf n}})\Bigl(z_m({{\mathbf n}})z_n({{\mathbf n}}+{{\mathbf e}}_n)
-z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr)\Biggr) ~,
\label{deformed Q-exact action Q=4}\end{aligned}$$ where $\beta \in {\mathbb{R}}$ and the BRST transformation is given by (\[BRST Q=4\]). By construction, this deformation does not spoil the $Q$-exactness of the action and it becomes identical with the original orbifolded matrix theory (\[orbifold action Q=4\]) by setting $\beta=1$. We can show that the vacuum energy of this deformed theory also vanishes at the 1-loop level. Therefore, repeating the same argument above, we can conclude that there is no perturbative correction to the vacuum energy of this theory.
Orbifold lattice theories from $Q=8$ mother theory {#Q=8}
--------------------------------------------------
Next we consider the lattice theories constructed from the mother theory with eight supercharges, that is, the dimensionally reduced six-dimensional ${{\cal N}}=1$ supersymmetric Yang-Mills theory [@Cohen:2003qw]. By performing an orbifold projection to the mother theory, we obtain the action of the orbifolded matrix theory [@Damgaard:2007be]: $$\begin{aligned}
S_{\rm orb} &= \frac{1}{g^2}{\rm Tr}\sum_{{{\mathbf n}}\in{\mathbb{Z}}_N^d}\Biggl(
\frac{1}{4}\Bigl| z_m({{{\mathbf n}}})
z_n({{\mathbf n}}+{{\mathbf e}}_m) - z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr|^2 \cr
& + \frac{1}{8}\Bigl(z_m({{\mathbf n}}){\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m)z_m({{\mathbf n}}-{{\mathbf e}}_m)\Bigr)^2 \cr
& - \psi_m({{\mathbf n}})\Bigl({\bar{z}}_m({{\mathbf n}})\eta({{\mathbf n}})-\eta({{\mathbf n}}+{{\mathbf e}}_m)
{\bar{z}}_m({{\mathbf n}})\Bigr)\cr
& +\frac{1}{2} \xi_{mn}({{\mathbf n}})\Bigl(z_m({{\mathbf n}})\psi_n({{\mathbf n}}+{{\mathbf e}}_m)
-\psi_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)-(m\!\leftrightarrow\!n)\Bigr) \cr
& -\frac{1}{2}\chi_{lmn}({{\mathbf n}})\Bigl(
{\bar{z}}_l({{\mathbf n}}+{{\mathbf e}}_m+{{\mathbf e}}_n)\xi_{mn}({{\mathbf n}})
-\xi_{mn}({{\mathbf n}}+{{\mathbf e}}_l){\bar{z}}_l({{\mathbf n}})
\Bigr)\Biggr)~,
\label{orbifold action Q=8}\end{aligned}$$ where $l,m,n=1,2,3$, ${{\mathbf e}}_m$ are integer valued three-component vectors, $d$ is the number of linearly independent vectors in $\{{{\mathbf e}}_m\}$, and again we assume that all the fields are complex matrices with the size $M$. Note that $d$ is the maximal dimensionality of the lattice theory obtained after deconstruction. The fields $z_m({{\mathbf n}})$ and ${\bar{z}}_m({{\mathbf n}})$ are bosonic fields living on links $({{\mathbf n}},{{\mathbf n}}+{{\mathbf e}}_m)$ and $({{\mathbf n}}+{{\mathbf e}}_m,{{\mathbf n}})$, respectively, and $\eta({{\mathbf n}})$, $\psi_m({{\mathbf n}})$, $\xi_{mn}({{\mathbf n}})$ and $\chi_{lmn}({{\mathbf n}})$ are fermionic fields on the site ${{\mathbf n}}$, the link $({{\mathbf n}},{{\mathbf n}}+{{\mathbf e}}_m)$, the link $({{\mathbf n}}+{{\mathbf e}}_m+{{\mathbf e}}_n,{{\mathbf n}})$ and the link $({{\mathbf n}},{{\mathbf n}}+{{\mathbf e}}_l+{{\mathbf e}}_m+{{\mathbf e}}_n)$, respectively. The fields $\xi_{mn}({{\mathbf n}})$ and $\chi_{lmn}({{\mathbf n}})$ are antisymmetric in terms of a permutation of the indices.
In this case, we can construct several kinds of supersymmetric lattice gauge theories with a different dimensionality, with a different number of preserved supercharges and with a different lattice structure by changing the vectors ${{\mathbf e}}_m$ and the number of bosonic fields to shift as (\[shift\]) [@Damgaard:2007be]. Recalling the discussion around (\[comment\]), however, we can estimate the vacuum energy of these theories at once by directly expanding the orbifolded matrix theory (\[orbifold action Q=8\]) around $$z_m({{\mathbf n}})=\left(
\begin{matrix}a_m^1 & & \\ & \ddots & \\ && a_m^M\end{matrix}
\right)\equiv a_m, \qquad
{\bar{z}}_m({{\mathbf n}})=\left(
\begin{matrix}\bar{a}_m^1 & & \\ & \ddots & \\ && \bar{a}_m^M\end{matrix}
\right)\equiv \bar{a}_m.
\label{vacuum Q=8}$$ By fixing the gauge by the gauge condition (\[gauge fix\]), we obtain the second-order action, $$\begin{aligned}
S^{(2)} = \frac{1}{g^2}{{\rm Tr}\,}\sum_{{{\mathbf n}}\in{\mathbb{Z}}_N^d}\Biggl(
&\frac{1}{2}\bar{D}^-_n z_m({{\mathbf n}}) D^-_n z_m({{\mathbf n}})
+\frac{1}{2}\bar{D}^-_n B({{\mathbf n}}) D^-_n C({{\mathbf n}})
+\eta({{\mathbf n}}) \bar{D}^-_m \psi_m({{\mathbf n}}){\nonumber}\\
&-\frac{1}{2}\xi_{mn}({{\mathbf n}})\Bigl(D^+_m \psi_n({{\mathbf n}})- D^+_n \psi_m({{\mathbf n}})
\Bigr)
+\frac{1}{2}\xi_{mn}({{\mathbf n}})\bar{D}_l^- \chi_{lmn}({{\mathbf n}})
\Biggr).
\label{2nd order action Q=8}\end{aligned}$$ From this expression, it is easy to show that the 1-loop contribution to the vacuum energy vanishes again.
As for the case of the $Q=4$ orbifold lattice theories, the lattice theory (\[orbifold action Q=8\]) possesses a BRST charge $Q$ that acts on the fields as [@Cohen:2003qw] $$\begin{aligned}
Q z_m({{\mathbf n}}) &= \psi_m({{\mathbf n}}), \quad Q {\bar{z}}_m({{\mathbf n}}) = 0, {\nonumber}\\
Q d({{\mathbf n}}) &= \psi_m({{\mathbf n}}){\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m)\psi_m({{\mathbf n}}-{{\mathbf e}}_m), {\nonumber}\\
\label{BRST Q=8 old}
Q\eta({{\mathbf n}}) &=
\frac{1}{4}\Bigl(z_m({{\mathbf n}}){\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m)z_m({{\mathbf n}}-{{\mathbf e}}_m)-d({{\mathbf n}})\Bigr), \\
Q\xi_{mn}({{\mathbf n}}) &= \frac{1}{2}\Bigl({\bar{z}}_m({{\mathbf n}}+{{\mathbf e}}_n){\bar{z}}_n({{\mathbf n}})
-{\bar{z}}_n({{\mathbf n}}+{{\mathbf e}}_m){\bar{z}}_m({{\mathbf n}})\Bigr), {\nonumber}\\
Q\chi_{lmn}({{\mathbf n}})&=0, {\nonumber}\end{aligned}$$ where $d({{\mathbf n}})$ is again an auxiliary field to make $Q$ nilpotent off-shell. Here we can extend (\[BRST Q=8 old\]) by supplementing the fields with an additional bosonic field $f_{lmn}({{\mathbf n}})$ satisfying $$\begin{aligned}
Qf_{lmn}({{\mathbf n}})&=\chi_{lmn}({{\mathbf n}}).
\label{new field}\end{aligned}$$ Then the action of the orbifolded matrix theory (\[orbifold action Q=8\]) can be equivalently expressed in a $Q$-exact form: $$\begin{aligned}
S_{\rm orb} =
\frac{1}{2g^2}{\rm Tr}\sum_{{{\mathbf n}}\in {\mathbb{Z}}_N^d}Q\Biggl(
& \eta({{\mathbf n}})\Bigl({z}_m({{\mathbf n}}) {\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m){z}_m({{\mathbf n}}-{{\mathbf e}}_m)+d({{\mathbf n}})\Bigr) {\nonumber}\\
& - \chi_{mn}({{\mathbf n}})\Bigl(z_m({{\mathbf n}})z_n({{\mathbf n}}+{{\mathbf e}}_n)
-z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr){\nonumber}\\
&-\frac{1}{2}f_{lmn}({{\mathbf n}})\Bigl(
{\bar{z}}_l({{\mathbf n}}+{{\mathbf e}}_m+{{\mathbf e}}_n)\xi_{mn}({{\mathbf n}})
-\xi_{mn}({{\mathbf n}}+{{\mathbf e}}_l){\bar{z}}_l({{\mathbf n}})
\Bigr)\Biggr) ~.
\label{Q-exact action Q=8}\end{aligned}$$ Note that, although the partition function diverges by integration over $f_{lmn}({{\mathbf n}})$, it is irrelevant for the vacuum energy. Therefore, the 1-loop result given above is shown to be exact by repeating the argument in the previous subsection, and the vacuum energy is expected to be zero in all order of the perturbative expansion.
In summary, we can conclude that [*the flat directions of the orbifold lattice theories constructed from the mother theory with four and eight supersymmetries do not receive any quantum correction perturbatively.*]{}
Conclusion and Discussion
=========================
In this paper, we examined quantum corrections to the classical moduli space of orbifold supersymmetric lattice theories constructed from the $Q=4$ and $Q=8$ mother theories. We showed that the classical moduli space does not receive any quantum correction perturbatively, namely, the flat directions of these theories remain flat even if we take into account quantum effects. We also modified the action of the $Q=4$ orbifolded matrix theory without spoiling the $Q$-exactness and showed that the classical moduli space of the deformed theory does not receive any perturbative correction either.
We conclude this paper by making some comments on other orbifold lattice theories. Let us first consider an orbifolded matrix theory, $$\begin{aligned}
S_{\rm orb} =
&\frac{1}{g^2}{\rm Tr}\sum_{{{\mathbf n}}}\Biggl(
\frac{1}{4}\Bigl|z_m({{{\mathbf n}}})
z_n({{\mathbf n}}+{{\mathbf e}}_m) - z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr|^2 {\nonumber}\\
&+\frac{1}{8}\Bigl(z_m({{\mathbf n}})\bar{z}_m({{\mathbf n}})
-\bar{z}_m({{\mathbf n}}-{{\mathbf e}}_m)z_m({{\mathbf n}}-{{\mathbf e}}_m)\Bigr)^2 {\nonumber}\\
& + \eta({{\mathbf n}})
\Bigl(\bar{z}_m({{\mathbf n}}+{{\mathbf a}}-{{\mathbf e}}_m)\psi_m({{\mathbf n}}+{{\mathbf a}}-{{\mathbf e}}_m)
-\psi_m({{\mathbf n}}+{{\mathbf a}})\bar{z}_m({{\mathbf n}}+{{\mathbf a}})\Bigr)\cr
& - \frac{1}{2}\chi_{mn}({{\mathbf n}})
\Bigl(z_m({{\mathbf n}})\psi_n({{\mathbf n}}+{{\mathbf e}}_m)
-\psi_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf a}}_n) -(m \leftrightarrow n)
\Bigr)\Biggl),
\label{general orbifold action}\end{aligned}$$ where ${{\mathbf e}}_m$, ${{\mathbf a}}$, ${{\mathbf a}}_m$ and ${{\mathbf a}}_{12}$ are three-component vectors satisfying $${{\mathbf a}}+{{\mathbf a}}_m = {{\mathbf e}}_m, \quad {{\mathbf a}}_{12}+{{\mathbf a}}_m=-|{\epsilon}_{mn}|{{\mathbf e}}_n,
\quad {{\mathbf a}}+{{\mathbf a}}_1+{{\mathbf a}}_2+{{\mathbf a}}_{12}=0,
\label{a-relations}$$ $z_m({{\mathbf n}})$ and ${\bar{z}}_m({{\mathbf n}})$ are the same bosonic fields as in (\[orbifold action Q=4\]) but $\eta({{\mathbf n}})$, $\psi_m({{\mathbf n}})$ and $\chi_{12}({{\mathbf n}})$ are fermionic fields living on the links $({{\mathbf n}},{{\mathbf n}}+{{\mathbf a}})$, $({{\mathbf n}},{{\mathbf n}}+{{\mathbf a}}_m)$ and $({{\mathbf n}}-{{\mathbf a}}_{12},{{\mathbf n}})$, respectively. In particular, we assume that any of the vectors ${{\mathbf a}}$, ${{\mathbf a}}_m$ and ${{\mathbf a}}_{12}$ is not zero. The action (\[general orbifold action\]) has been first given in [@D'Adda:2005zk] and is shown to be obtained from $Q=4$ mother theory by an orbifold projection with no preserved supercharge in any usual sense [@Damgaard:2007eh]. It is easy to show that the vacuum energy of this theory again vanishes at the 1-loop level. However, in this case, there seems to be no guarantee that higher-loop contributions to the vacuum energy vanish, since there is no usual BRST symmetry in this theory. It would be interesting, however, to investigate quantum corrections to this theory from the view point of the supersymmetry algebra on lattice discussed in [@D'Adda:2004jb][@D'Adda:2005zk].
Interesting orbifold lattice theories are those constructed from $Q=16$ mother theory [@Kaplan:2005ta], that is, IKKT matrix theory [@Ishibashi:1996xs]. The action of the corresponding orbifolded matrix theory is written as $$\begin{aligned}
S_{\rm orb} &= \frac{1}{g^2}{\rm Tr}\sum_{{{\mathbf n}}\in{\mathbb{Z}}_N^d}\Biggl(
\frac{1}{4}\Bigl| z_m({{{\mathbf n}}})
z_n({{\mathbf n}}+{{\mathbf e}}_m) - z_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)\Bigr|^2 \cr
& + \frac{1}{8}\Bigl(z_m({{\mathbf n}}){\bar{z}}_m({{\mathbf n}})
-{\bar{z}}_m({{\mathbf n}}-{{\mathbf e}}_m)z_m({{\mathbf n}}-{{\mathbf e}}_m)\Bigr)^2 \cr
& - \psi_m({{\mathbf n}})\Bigl({\bar{z}}_m({{\mathbf n}})\eta({{\mathbf n}})-\eta({{\mathbf n}}+{{\mathbf e}}_m)
{\bar{z}}_m({{\mathbf n}})\Bigr)\cr
& +\frac{1}{2} \xi_{mn}({{\mathbf n}})\Bigl(z_m({{\mathbf n}})\psi_n({{\mathbf n}}+{{\mathbf e}}_m)
-\psi_n({{\mathbf n}})z_m({{\mathbf n}}+{{\mathbf e}}_n)-(m\!\leftrightarrow\!n)\Bigr) \cr
& -\frac{1}{2}{\epsilon}_{mnpqr}\xi_{mn}({{\mathbf n}})\Bigl(
{\bar{z}}_p({{\mathbf n}}+{{\mathbf e}}_q+{{\mathbf e}}_r)\xi_{pq}({{\mathbf n}})
-\xi_{pq}({{\mathbf n}}+{{\mathbf e}}_p){\bar{z}}_p({{\mathbf n}})
\Bigr)\Biggr)~,
\label{orbifold action Q=16}\end{aligned}$$ where $m,n=1,\cdots,5$, ${{\mathbf e}}_m$ are five-component vectors satisfying $\sum_{m=1}^5 {{\mathbf e}}_m=0$ and $d$ is the number of the linearly independent vectors in $\{{{\mathbf e}}_m\}$. Again the classical vacua are parametrized as (\[lattice vacuum\]) with $m=1,\cdots 5$, and it is straightforward to show that the vacuum energy is zero at the 1-loop level. However, we cannot apply the same argument in the previous section since the last term of the action (\[orbifold action Q=16\]) is not $Q$-exact but $Q$-closed [@Kaplan:2005ta]. Thus, there is a possibility that the classical flat directions would be lifted up by quantum effects. In fact, from the viewpoint of the superstring theory, we can expect that non-trivial quantum corrections to the vacuum energy exist in this case. Recalling that the mother theory with sixteen supercharges is identical with the low energy effective theory on D-instantons on a ten-dimensional flat space-time, the orbifolded matrix theory (\[orbifold action Q=16\]) can be regarded as the low energy effective theory on D-instantons in the background of an orbifold[^6]. In this interpretation, the background (\[lattice vacuum\]) can be regarded as the positions of D-instantons. The point is that this orbifold background breaks the supersymmetry on the ten-dimensional space-time, so (\[lattice vacuum\]) or (\[vacuum Q=8\]) gives a non-BPS configuration of D-branes. Therefore, it seems that there should be some force between the separated D-instantons. In terms of the theory on the D-instantons, this means the classical flat directions parametrized by $a_m$ are no longer flat if we take into account quantum corrections to the orbifolded matrix model. It would be interesting to analyse these theories along this way [@DM].
Acknowledgments {#acknowledgments .unnumbered}
===============
The author would like to thank P. H. Damgaard and S. Hirano for useful discussions and valuable comments. He would also like to thank D. B. Kaplan, M. Ünsal, H. Suzuki, F. Sugino and T. Takimi for useful comments. This work is supported by JSPS Postdoctoral Fellowship for Research Abroad.
[^1]: For a nice review, see [@Giedt:2006pd].
[^2]: In this paper, we restrict ourselves to consider gauge theories in $d$ dimensional space-time with $d\ge 2$.
[^3]: In this calculation, the constant modes are treated by shifting the difference operators (\[D-\]) and (\[D+\]) as $D_m^\pm \to D_m^\pm + i\mu$, which corresponds to adding mass terms as done in [@Cohen:2003xe]. Although this modification breaks the BRST symmetry, the final result of the following discussion still holds in the limit of $\mu\to 0$ since the breaking of the symmetry is soft.
[^4]: The author would like to thank to H. Suzuki for discussing this point.
[^5]: The physical interpretation of this deformation is still unclear. In fact, the continuum limit of this deformed theory is not Lorentz invariant, though it has a BRST symmetry generated by $Q$. The author would like to thank M. Ünsal for pointing it out.
[^6]: For a related work, see [@Unsal:2005us].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the factorization formula for non-leptonic $B$ decays to two light flavor non-singlet mesons derived by Bauer et al. in the context of soft-collinear effective theory is equivalent to the corresponding formula in the QCD factorization approach. The apparent numerical differences in the analysis of $B\to\pi\pi$ data performed by these authors, as compared to previous QCD factorization analyses, can largely be attributed to the neglect of known perturbative and power corrections.'
author:
- 'M. Beneke${}^1$, G. Buchalla${}^2$, M. Neubert${}^{3,4}$, and C.T. Sachrajda${}^5$'
date: 'September 19, 2004'
title: |
Comment on “$B\to M_1 M_2$: Factorization, charming penguins,\
strong phases, and polarization”
---
The extent to which hadronic decays of $B$ mesons to two light hadrons can be computed from first principles in QCD has been the subject of many investigations, following the statement [@Beneke:1999br; @Beneke:2000ry] that the decay amplitudes factorize in the limit of very large $B$-meson mass. Taking the final state to consist of two pions, the amplitude can be represented schematically in the form $$\label{ff}
A(B\to \pi\pi) = F^{B\to\pi}\,T^{\rm I}\star\Phi_\pi
+ T^{\rm II}\star\Phi_B\star\Phi_\pi\star\Phi_\pi \,.$$ In this equation, $F^{B\to\pi}$ denotes a physical form factor, and $\Phi_B$ and $\Phi_\pi$ are the leading-twist light-cone distribution amplitudes of the mesons. The quantities $T^{\rm I,II}$ are perturbative hard-scattering kernels involving the two scales $m_b$ (hard) and $\sqrt{m_b\Lambda}$ (hard-collinear), which are linked to the other elements of the formula by convolution integrals (indicated by an asterisk).
The authors of [@Bauer:2004tj] suggest a factorization formula (their Eq. (24)) similar to (\[ff\]) in the framework of soft-collinear effective theory (SCET) and imply that it is conceptually different. They also argue that, even at leading order in the $1/m_b$ expansion, there may be an additional term on the right-hand side of (\[ff\]), corresponding to long-distance contributions from $c\bar c$ penguins. They do not disprove factorization of charm-penguin loops by providing a counter-example to factorization; rather, they state that they were not able to demonstrate factorization.
In this Comment, we explain why the formula given in [@Bauer:2004tj] is identical in content to the QCD factorization formula (\[ff\]), and why non-factorizable charm-penguin contributions are of higher order in the $1/m_b$ expansion. We also point out that the phenomenological analysis of [@Bauer:2004tj] neglects important perturbative and power corrections which are already known. Once these are included, there is little room for significant additional contributions to the QCD penguin amplitude.
Equivalence of the SCET and QCD factorization formulae
======================================================
We first note that the coefficient function of the spectator-scattering term can be represented as a convolution $T^{\rm II}=C^{\rm II}\star J$ of hard and hard-collinear coefficient functions. The formula quoted in [@Bauer:2004tj] follows from (\[ff\]) by rewriting $$\label{rearrange}
T^{\rm II}\star\Phi_B\star\Phi_\pi\star\Phi_\pi
= C^{\rm II}\star\zeta_J^{B\pi}\star\Phi_\pi \,,$$ with $\zeta_J^{B\pi}$ defined as $J\star\Phi_B\star\Phi_\pi$, and $C^{\rm II}$ defined by the decomposition of $T^{\rm II}$ shown above. In addition, in [@Bauer:2004tj] the SCET form factor $\zeta^{B\pi}$ rather than the physical QCD form factor $F^{B\to\pi}$ is used. As discussed in [@Beneke:2000ry], this implies another rearrangement of this type.
In general SCET provides a powerful tool to simplify factorization proofs (a task that has not yet been completed for the case of $B\to\pi\pi$ considered here), but the resulting QCD factorization formulae can also be obtained using traditional factorization methods, as is frequently done in practice.
The authors of [@Bauer:2004tj] entertain the possibility that the hard-collinear scale may be non-perturbative, and hence they choose not to factorize $\zeta_J^{B\pi}$ into $J\star\Phi_B\star\Phi_\pi$ as indicated above. This is a logical possibility in the QCD factorization approach. However, we show below that it is not supported by theoretical calculations.
We disagree with [@Bauer:2004tj] on the statement that $\zeta_J^{B\pi}\ll F^{B\to\pi}$ is a prediction of QCD factorization, whereas the SCET treatment suggests $\zeta_J^{B\pi}\sim F^{B\to\pi}$. The factorization formula states that both terms in (\[ff\]) are of the same order in $1/m_b$ power counting and that $T^{\rm I}$ and $T^{\rm II}$ start at $O(\alpha_s^0)$ and $O(\alpha_s(\sqrt{m_b\Lambda}))$, respectively. However, it does not predict the relative size of the two terms, and as we explain below, the numerical value of $\zeta_J^{B\pi}$ depends on several hadronic input parameters, which are rather uncertain at present.
The authors of [@Bauer:2004tj] point out that the hard-collinear kernel $J$ is universal, so that only a single function $\zeta_J^{B\pi}$ appears in (\[rearrange\]), which is the same that appears in the factorization of the $B\to\pi$ form factors. This fact is important for phenomenology when one opts to treat the hard-collinear scale as non-perturbative, but does not by itself represent a conceptual difference between the formulae given in [@Beneke:1999br; @Beneke:2000ry] and [@Bauer:2004tj]. Furthermore, the usefulness of the universality of $J$ is limited to the approximation where one neglects radiative corrections to the hard-scattering kernels $C^{\rm II}$, since only then does the function $\zeta_J^{B\pi}$ reduce to a single number [@Beneke:2003pa; @Hill:2004if]. In other words, a phenomenological treatment of the hard-collinear scale as non-perturbative relies on the approximation that the kernels are restricted to their tree-level approximations, whereas one of the key features of QCD factorization (as opposed to naive factorization) is that one can consistently include radiative corrections.
Charm-penguin loops
===================
In [@Bauer:2004tj] it is claimed that there is a possible exception to factorization from diagrams with charm-quark loops (see Figure \[fig:penguin\]). The argument is based on the observation that when the gluon virtuality is near the $c\bar c$ threshold, $q^2\approx 4 m_c^2$, the non-relativistic scales $m_c v$ and $m_c v^2$ become important. Since $m_c v^2\approx\Lambda$ numerically, this appears to introduce a sensitivity to non-perturbative scales without power-suppression in $1/m_b$. The authors of [@Bauer:2004tj] suggest that these diagrams do not factorize, in the sense that the long-distance physics at leading-power in the $1/m_b$ expansion cannot be factored into form factors or light-cone distribution amplitudes. We emphasize that the question of non-factorizable $c\bar c$ effects at leading order in the heavy-quark expansion is a different issue than the one raised in [@Ciuchini:2001gv], where it is speculated that power corrections to the QCD charm-penguin amplitudes may be numerically large.
Factorization statements, be they derived diagrammatically or with soft-collinear effective theory, always concern properties of the amplitude in certain asymptotic limits, here an expansion in $1/m_b$, independent of the actual size of the expansion parameter. It is therefore important to clearly distinguish the issue of factorization at leading order in the $1/m_b$ expansion from the question of whether there are non-factorizable contributions which are formally power-suppressed, but which nevertheless may be numerically significant. In the following discussion, we focus on the question of factorization in the formal heavy-quark limit.
The most intuitive way of understanding why the threshold region does not require special treatment is based on quark-hadron duality [@Beneke:2000ry]. The integration over the gluon virtuality in the range $0\le q^2\le m_b^2$ is weighted by the pion distribution amplitude, which is smooth over the entire integration region. This provides the necessary smearing of the loop amplitude, which ensures that the result is given by a simple partonic calculation up to power corrections, in complete analogy with the standard justification of the partonic interpretation of inclusive heavy-meson decays, or cross sections in general. The smearing also ensures that one can apply the same power-counting arguments that demonstrate factorization of other diagrams to charm-penguin diagrams with no need to single out the threshold region. In particular, in the non-relativistic situation implied in [@Bauer:2004tj] there is no need to sum Coulomb ladder diagrams, since these do not result in large perturbative corrections after the integration over $q^2$.
Even without invoking the duality argument, the fact that the charm threshold region comprises only a parametrically small portion of the entire integration implies a phase-space suppression. This fact has been neglected in the argument of [@Bauer:2004tj]. More precisely, writing $q^2=\bar x m_b^2$, where $\bar x$ denotes the longitudinal momentum fraction of the anti-quark in one of the pions, this region is $\Delta\bar x\sim v^2\,(m_c/m_b)^2$ or $\Delta\bar x\sim\Lambda m_c/m_b^2$, whichever is larger. In order to study the question of whether long-distance $c\bar c$ loop effects are of leading order or not, it is necessary to decide how the limit $m_b\to\infty$ is to be taken. If we define the heavy-quark limit by $m_b\to\infty$ with $m_c$ fixed, one may distinguish several possibilities such as $m_c\sim\Lambda$, $m_c v^2\sim\Lambda$, or $m_c v^2\gg\Lambda$. While the physics of the threshold region is very different for all these cases, they share the common feature that, for the purposes of power counting, the charm quark can be considered to be a light quark, and the suppression of long-distance $c\bar{c}$ effects has the same origin as that for the corresponding diagrams with light-quark loops, which implies $\Delta\bar x\sim 1/m_b^2$. In addition, since in this region $\bar x\sim m_c^2/m_b^2$, there is a further suppression due to the end-point behavior of the pion distribution amplitude (which vanishes linearly as $\bar x\to 0$).
If we define the heavy-quark limit as $m_{b,c}\to\infty$ with the ratio $m_c/m_b$ fixed, then there is no power suppression due to the phase-space or end-point behavior of the distribution amplitudes, but the threshold region is perturbative, up to a small non-perturbative contribution of order $v^2\cdot v^2\,(\Lambda/(m_c v^2))^4\sim(\Lambda/m_b)^4$ [@Voloshin:1979uv].
We conclude that charm-penguin diagrams factorize at leading power in $1/m_b$. The argument for factorization remains valid also for more complicated higher-order penguin graphs whenever the threshold region is phase-space (and end-point) suppressed or the charm quark is heavy so that perturbation theory is applicable.
Validity of perturbation theory at the hard-collinear scale
===========================================================
In applying the QCD factorization formula to phenomenology the authors of [@Bauer:2004tj] treat the hard-collinear scale $\sqrt{m_b\Lambda}$ as non-perturbative, and hence the quantity $\zeta_J^{B\pi}$ as an unknown phenomenological function. This is justified [*a posteriori*]{} following the result of a phenomenological fit (on which we comment below). This line of argument ignores the fact that perturbation theory at scales of order $\sqrt{m_b\Lambda}\sim m_c$ has been used successfully in many important applications in $B$-physics, including all determinations of $|V_{ub}|$ from inclusive $B$-decays, and studies of the hadronic decay rate of the $\tau$ lepton. As already mentioned, a serious drawback of treating the hard-collinear scale as non-perturbative is that it renders the factorization approach unpredictive beyond the tree approximation, because only the integral over $\zeta_J^{B\pi}$ and not its functional form can be extracted from measurements.
A systematic way to address the question of the perturbativity of the hard-collinear scale is to calculate higher-order corrections in $\alpha_s$ to the jet function $J$ in the product $T^{\rm II}=C^{\rm II}\star J$, defined as the Wilson coefficient function arising in the matching of certain (type-B) SCET$_{\rm I}$ current operators onto four-quark operators of SCET$_{\rm II}$. The next-to-leading order terms have been computed recently and were found to be small [@Hill:2004if]. Specifically, for the case of light pseudoscalar mesons and an asymptotic light-cone distribution amplitude $\Phi_\pi(x)=6x(1-x)$, one finds that the convolution integrals over the jet function give rise to the series $$\label{nlocorrection}
\frac{\alpha_s(\mu_i)}{\lambda_B}
\left[ 1 + {\alpha_s(\mu_i)\over\pi} \left( \frac{\langle L^2\rangle}{3}
- 1.31 \langle L\rangle + 1.00 \right) + \dots \right] ,$$ where $\mu_i\sim\sqrt{m_b\Lambda}$ is the hard-collinear scale, $\lambda_B$ is the first inverse moment of the $B$-meson distribution amplitude $\Phi_B(\omega,\mu_i)$, $L=\ln(m_b\omega/\mu_i^2)$, and $\langle\dots\rangle$ denotes an average over $\Phi_B(\omega,\mu_i)$ with measure $d\omega/\omega$. While the precise form of the $B$-meson distribution amplitude is unknown, the fact that $\omega\sim\Lambda$ ensures that $L$ cannot be large, giving a small coefficient to the next-to-leading term. For example, using the results of [@Braun:2003wx] for the moments $\langle L^2\rangle$, $\langle L\rangle$ the coefficient of $\alpha_s/\pi$ in (\[nlocorrection\]) is $2.2\pm 0.6$ for $\mu_i^2=0.5\,\mbox{GeV}\,m_b$. There is thus no evidence that perturbation theory cannot be applied at the hard-collinear scale. We also note in this context that the power corrections from the hard-collinear scale are $1/m_b$ suppressed (and not $1/\sqrt{m_b}$) just as those from the hard scale.
Since the perturbative corrections to the jet function are well behaved, the quantity $\zeta_J^{B\pi}$ can be factorized and expressed in terms of convolution integrals over light-cone distribution amplitudes. The question of the numerical value of $\zeta_J^{B\pi}$, and whether it is a small contribution to the physical form factor $F^{B\to\pi}$, rests on the properties of these amplitudes, as well as on other parameters such as the strange-quark mass. At leading order in perturbation theory, we obtain for $\zeta_J^{B\pi}$ the result $$\zeta_J^{B\pi} = \frac{3\pi\alpha_s C_F}{N_c^2}\,
\frac{f_B f_\pi}{M_B\lambda_B}\left(\langle \bar y^{-1}\rangle_\pi
+ r_\chi^\pi X_H\right)$$ where the notations of [@Beneke:2003zv] have been used. Taking the default values and uncertainties of the input parameters from this reference, and adding errors in quadrature, yields $\zeta_J^{B\pi}=0.016\mbox{--}0.064$, which is small compared with typical values $F^{B\to\pi}=0.24\mbox{--}0.30$. ($F^{B\to\pi}=\zeta^{B\pi}+\zeta_J^{B\pi}$ when hard matching corrections are neglected.) Taking some correlated parameter variations so as to reproduce the data on $B\to\pi\pi$ decays, scenario S2 of [@Beneke:2003zv] yields the somewhat increased value $\zeta_J^{B\pi}=0.080$. To obtain significantly larger results would require a very small value of the hadronic parameter $\lambda_B$. While this is a logical possibility, a recent QCD sum-rule calculation of $\lambda_B$ gives a value around $0.45\,$GeV [@Braun:2003wx], which is in fact somewhat larger than the estimate adopted in [@Beneke:1999br; @Beneke:2000ry].
These estimates are to be compared to the fit result $\zeta_J^{B\pi}=0.11\pm 0.03$ obtained by the authors of [@Bauer:2004tj], who also find that the bulk of the $B\to\pi$ form factor comes from $\zeta_J^{B\pi}$, giving the very small result $F^{B\to\pi}=0.17\pm 0.02$. This picture contradicts the QCD sum rules for heavy-to-light form factors, in which $\zeta_J^{B\pi}$ must be associated with a radiative correction [@Bagan:1997bp]. The preference of the data for a smaller $B\to\pi$ form factor together with an increase of the hard-spectator scattering contribution to the color-suppressed tree amplitude $a_2$ has already been discussed in [@Beneke:2003zv], which however did not arrive at a similarly extreme conclusion. This discrepancy can be traced to a few omissions in the calculation of [@Bauer:2004tj], each of which has a minor effect: the absence of radiative corrections, the absence of phases in tree amplitudes, the absence of the scalar up-penguin amplitude in $T_c$, the use of asymptotic wave functions, and finally, a larger value of $|V_{ub}|$. When these effects are taken into account and combined with the most recent experimental data, one finds a significantly smaller value of $\zeta_J^{B\pi}$ and a larger value of $F^{B\to\pi}$, in qualitative agreement with theoretical expectations.
The QCD penguin amplitude
=========================
We now turn to the discussion of the phenomenological analysis of the $B\to\pi\pi$ data performed in [@Bauer:2004tj]. Our principal criticism in addition to what has already been described concerns the evaluation of the QCD penguin amplitude. It may be written as $$P = a_4 + r_\chi^\pi a_6 + \beta_3 \approx -0.09 \,,$$ where $a_4\approx -0.023\,[\alpha_s^0]-0.002\,[\alpha_s^1]$ represents the vector penguin contribution, $r_\chi^\pi a_6\approx -0.038\,[\alpha_s^0]-0.014\,[\alpha_s^1]$ the $1/m_b$ suppressed scalar penguin contribution, and $\beta_3\approx -0.011$ a power suppressed and rather uncertain penguin annihilation term. (The numbers are based on the analysis in [@Beneke:2003zv]. Without errors they should be taken only for illustration purposes. In particular, we neglected all phases, since they are unimportant for the following discussion.)
In the calculation of [@Bauer:2004tj] the lowest order ($\alpha_s^0$) and some of the $\alpha_s$ contributions to $a_4$ (those included in the phenomenological parameters for hard scattering and charm penguins) are taken into account. The term $r_\chi a_6+\beta_3$ is dropped, because it is power-suppressed. Now while it is true that the factorization properties of power-suppressed contributions in general, and scalar penguin contributions in particular, have not yet been investigated to all orders in perturbation theory, the large tree-level contribution to $r_\chi a_6$ suggests that if one neglects power corrections entirely (as done in [@Bauer:2004tj]) one is certain to obtain a poor approximation. We emphasize that this is unrelated to the charm-quark loops discussed above, which appear only in the small $\alpha_s$ corrections. By dropping the scalar penguin amplitude, the authors of [@Bauer:2004tj] are forced to erroneously assign the QCD penguin amplitude almost entirely to the charm-quark loops.
There is considerable phenomenological evidence that the scalar penguin amplitude is in approximate agreement with our theoretical expectations. The suppression of the pseudoscalar-vector and vector-pseudoscalar penguin amplitudes relative to the pseudoscalar-pseudoscalar penguin amplitude [@Beneke:2003zv], as well as the pattern of drastically different branching fractions for the decay modes $B\to \eta^{(\prime)} K^{(*)}$ [@Beneke:2002jn], can be attributed directly to the different size and sign of the $r_\chi a_6$ term relative to $a_4$ in the QCD penguin amplitude. We are unaware of any other theoretical framework that can explain these facts. From such studies of penguin dominated $B$-decays we are therefore led to the conclusion that there is little room for extra contributions to the QCD penguin amplitude.
We are grateful to Iain Stewart for useful discussions. M.B. and M.N. would like to thank the KITP, Santa Barbara, for hospitality during the preparation of this note. The work of M.B. is supported in part by the DFG Sonderforschungsbereich/Transregio 9 “Computer-gestützte Theoretische Teilchenphysik”. The research of M.N. is supported by the National Science Foundation under Grant PHY-0355005, and by the Department of Energy under Grant DE-FG02-90ER40542. The work of C.T.S. is supported by PPARC Grant PPA/G/0/2002/00468.
[99]{}
M. Beneke, G. Buchalla, M. Neubert and C. T. Sachrajda, Phys. Rev. Lett. [**83**]{}, 1914 (1999) \[hep-ph/9905312\]. M. Beneke, G. Buchalla, M. Neubert and C. T. Sachrajda, Nucl. Phys. B [**591**]{}, 313 (2000) \[hep-ph/0006124\]. C. W. Bauer, D. Pirjol, I. Z. Rothstein and I. W. Stewart, Phys. Rev. D [**70**]{}, 054015 (2004) \[hep-ph/0401188\]. M. Beneke and T. Feldmann, Nucl. Phys. B [**685**]{}, 249 (2004) \[hep-ph/0311335\]. R. J. Hill, T. Becher, S. J. Lee and M. Neubert, JHEP [**0407**]{}, 081 (2004) \[hep-ph/0404217\]. M. Ciuchini, E. Franco, G. Martinelli, M. Pierini and L. Silvestrini, Phys. Lett. B [**515**]{}, 33 (2001) \[hep-ph/0104126\]. M. B. Voloshin, Sov. J. Nucl. Phys. [**36**]{}, 143 (1982) \[Yad. Fiz. [**36**]{}, 247 (1982)\].
V. M. Braun, D. Y. Ivanov and G. P. Korchemsky, Phys. Rev. D [**69**]{}, 034014 (2004) \[hep-ph/0309330\]. M. Beneke and M. Neubert, Nucl. Phys. B [**675**]{}, 333 (2003) \[hep-ph/0308039\]. E. Bagan, P. Ball and V. M. Braun, Phys. Lett. B [**417**]{}, 154 (1998) \[hep-ph/9709243\]. M. Beneke and M. Neubert, Nucl. Phys. B [**651**]{}, 225 (2003) \[hep-ph/0210085\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The scale invariance of natural images suggests an analogy to the statistical mechanics of physical systems at a critical point. Here we examine the distribution of pixels in small image patches and show how to construct the corresponding thermodynamics. We find evidence for criticality in a diverging specific heat, which corresponds to large fluctuations in how ‘surprising’ we find individual images, and in the quantitative form of the entropy vs. energy. The energy landscape derived from our thermodynamic framework identifies special image configurations that have intrinsic error correcting properties, and neurons which could detect these features have a strong resemblance to the cells found in primary visual cortex.'
author:
- 'Greg J Stephens,$^{a,}$[^1] Thierry Mora,${}^{a,*}$ Gasper Tkačik$^b$ and William Bialek$^{a,c}$'
title: Thermodynamics of natural images
---
Introduction
============
From the familiar faces of our friends and family to objects of almost every size in environments of every type, the world that we see is full of structure. Although this structure seems obvious when we look at the world, providing a precise mathematical description has proven more difficult. One way to formulate this problem is to ask for a probability distribution of images such that, if we draw at random out of this distribution, the resulting images resemble those that we see in the natural environment. Such a probabilistic or generative model would provide a rigorous basis for practical algorithms in image coding, processing and recognition [@mvBook]. It is also reasonable to hypothesize that our brains have learned at least an approximation to this probabilistic model, allowing us to form more efficient representations of the visual world and to find efficient solutions of many seemingly difficult computational problems. In this view, aspects of vision ranging from the responses of individual neurons to gestalt perceptual rules would be seen not as artifacts of the brain’s circuitry but rather as matched to the statistical structure of the physical world [@barlow61; @atick; @simoncelli+olshausen_01; @bialek_02].
One statistical feature of natural images that provides a clue about the nature of the underlying probability distribution is scale invariance. In particular, Field observed that the spatial patterns of image intensity from reasonably natural environments have power spectra that approximate $S_{I} \propto 1/k^2$, which is what one would expect from the hypothesis of scale invariance and simple dimensional analysis, and he suggested that this scaling behavior may have a direct connection to the distribution of receptive field parameters across neurons in visual cortex [@field87]. The intuition that scale invariance is a strong constraint on the form of the probability distribution comes from statistical mechanics. We recall that for systems in thermal equilibrium, the probability that we observe the system in state $\rm s$ is given by the Boltzmann distribution, $p_{\rm s} \propto \exp(-E_{\rm s}/T)$, where $E_s$ is the energy of the state and $T$ is the absolute temperature [@note1]. For most physical systems at generic values of the temperature and other parameters, correlations and power spectra are [*not*]{} scale invariant; rather there is some characteristic length $\xi$ that determines the distance beyond which structures approach statistical independence. Scale invariance emerges only when we tune the temperature to a special value $T_c$, the critical point which marks a second order phase transition between two different phases (liquid and gas, ferromagnet and paramagnet, ... ) [@cardy_96]. The modern theory of critical phenomena teaches us that such scale invariance can occur while violating the naive expectations of dimensional analysis, so that power spectra can acquire “anomalous dimensions,” $S\propto 1/k^{2-\eta}$. Further, scaling extends beyond low order statistics, so that the full probability distributions are predicted to be invariant (but non–Gaussian) under appropriate scaling transformations. Both anomalous scaling and invariant non–Gaussian distributions for local features have been observed in an ensemble of natural scenes .
The analogy between scaling in natural images and the behavior of physical systems at their critical point point raises the question of whether there are analogs to the [*thermodynamic*]{} features of a critical point. Can we, for example, generalize a given natural image ensemble to a family of ensembles indexed by a “temperature,” and show that there is something special (i.e., critical) about the temperature of the real ensemble? If there is an analog of the diverging specific heat at $T_c$, what does this say about the nature of images? What are the order parameters that characterize the underlying phase transition? Here we report some preliminary results on these and related questions.
{width="0.9\linewidth"}
The image ensemble and scaling
==============================
As an initial data set we returned to the image ensemble of Ref . We focus here on the 45 images taken at lower spatial resolution, corresponding to $256\times 256$ pixel regions covering $\sim 15^\circ\times15^\circ$ scenes in the woods of Hacklebarney State Park in New Jersey; an example is shown in Fig 1a. Our path to the construction of a thermodynamics involves sampling the distribution of images in small patches. To make this problem manageable, we quantize the grey scale images into just two levels, with the quantization threshold chosen so that the numbers of black and white pixels are exactly equal over the ensemble.
It is important to verify that the rather harshly quantized images preserve interesting structures of the original scenes. First, by inspection of Fig 1b we see that objects and even parts of objects (branches and leaves on the trees) are recognizable. More quantitatively, if the original image is $\phi(\vec x )$, then we have constructed a discrete image $\sigma (\vec x) = {\mbox{sgn}}[\phi(\vec x ) - \theta ]$, with $\theta$ chosen so that $\langle \sigma \rangle = 0$, where $\langle \cdots \rangle$ represents an average over the image ensemble. Power spectra are defined by $$\langle \phi(\vec x ) \phi(\vec x ')\rangle = \int {{d^2 k}\over{(2\pi )^2}}
S_\phi (\vec k ) e^{i\vec k \cdot (\vec x - \vec x ')} ,$$ and similarly for $S_\sigma$. In Fig 1c we show the spectra $S_\phi$ and $S_\sigma$, averaged over the orientation of the ‘momentum’ vector $\vec k$ and normalized by the total variance. We see that both spectra exhibit scaling, with very similar exponents. As discussed in Ref , there is excess power at high frequencies because of aliasing, and more compelling evidence for scaling is obtained by combining these data with an ensemble of images from the same environment at higher angular resolution, so that the full range of $|\vec k|$ spans 2.5 decades.
In the quantized images, an $L\times L$ pixel region can take on $2^{L^2}$ possible states, and our data set provides $\sim 3\times 10^6$ samples of these states, although these are not independent. Thus we can expect to provide a good sampling of the distribution of discretized image patches for $L=3$ or even $L=4$, since $2^{16} \ll 3\times 10^6$, but $L=5$ is out of reach with this data set. A direct estimate of the entropy shows that $S(4\times 4) = 11.154\pm 0.002\,{\rm bits}$, much less than the 16 bits we would obtained from random pixels; similarly, we find $S(3\times 3)=6.580\pm0.003<9\,{\rm bits}$. This quantifies our impression that a substantial amount of local structure is preserved in the discretized images.
We can also test for scaling more generally by asking how distribution of image states in $L\times L$ patches evolves when we coarse–grain the images, and we can do this in two ways. First, we can take the original grey scale images $\phi (\vec x )$ and create a new image such that the value of $\phi$ in each pixel of the new image is the average over a $2^{n}\times 2^{n}$ block of pixels in the original image, and then we can quantize these images. When we look at $3\times 3$ patches in these filtered and quantized images, we again have $2^9$ possible states, and we call the distribution over these states $P_n$, where $P_0$ is the distribution obtained from the original images without any blocking. In the same spirit (and following the original approach in Ref [@kadanoff]), we can take the quantized image $\sigma (\vec x )$ and directly create new quantized images by applying majority rule to the pixels in $3\times 3$ blocks, and this can be iterated; we’ll call the resulting distributions of states in images patches $\bar P_n$, where again $\bar P_0 = P_0$ is what we obtain without blocking. Scale invariance is the claim that all the $P_n$ and $\bar P_n$ will be the same, independent of $n$, because the distribution of states is at a fixed point of this “renormalization” transformation [@cardy_96]. In Fig 1d we test this prediction, showing that it is obeyed with good accuracy over four decades in probability. There are more significant deviations in the first step of coarse–graining ($n=1$, at left in the figure), presumably because of the effects of aliasing noted above. We emphasize that this test of scale invariance involves the full, joint distribution of image intensities in $3\times 3$ patches, and thus goes beyond checking the power law behavior of the spectrum (a second order moment) or the invariance of distributions of features (e.g., the outputs of local filters) evaluated at a single point. Similar results are obtained for $4\times 4$ patches.
Temperature and specific heat
=============================
Small patches of our discrete images are described by a set $\vec \sigma$ of binary variables. Let us imagine that the distribution of these image patches is really the Boltzmann distribution for some physical system at temperature $T=1$, with some “energy” function $E(\vec \sigma )$ describing each possible patch, $$P(\vec \sigma ) = {1\over {Z}} e^{- E(\vec \sigma )} .
\label{basic}$$ Then, following the methods used in the analysis of dynamical systems , we can define the distribution at any temperature $T$, since $$\begin{aligned}
\label{eq:pt}
P_T (\vec \sigma ) &\equiv& {1\over {Z(T)}} e^{-E(\vec \sigma )/T} =
{1\over {{\cal Z} (T)}}[P(\vec \sigma )]^{1/T}\\
{\cal Z}(T) &=& \sum_{\vec\sigma} [P(\vec \sigma )]^{1/T} .\end{aligned}$$ We can define an entropy at each value of the temperature, $S(T) = -\sum_{\vec\sigma} P_T(\vec\sigma ) \log P_T (\vec\sigma )$, and then the usual thermodynamic relations tell us that the heat capacity is $C(T) = T \partial S(T)/\partial T$. It is also useful to note that the heat capacity is proportional to the variance in energy or log probability, $C(T) = \langle [\delta E (\vec \sigma )]^2 \rangle_T/ T^2$, where $\langle \cdots \rangle_T$ denotes an average in the distribution $P_T(\vec \sigma )$.
{width="\linewidth"}
In a system with a critical point, the specific heat should diverge at $T=T_c$. Of course this is true only in the thermodynamic limit of large systems, corresponding here to image patches containing many pixels. Can we see precursors of this divergence in the small patches that we can actually sample? Figure 2a shows the specific heat for $2\times 2$, $3\times 3$ and $4\times 4$ patches in our image ensemble, calculated directly from our sampling of the distributions $P(\vec \sigma )$. We see that, even when we normalize by the number of pixels $N=L^2$ in each patch (since we expect that the heat capacity is extensive), looking at larger patches reveals a larger specific heat with a clear peak as a function of temperature, and this peak is shifting toward $T=1$.
To calibrate our intuition about the specific heat estimated from small patches, we have done precisely analogous computations on the nearest neighbor ferromagnetic Ising model in two dimensions, defined by $E(\vec \sigma ) = -J \sum_{({\rm ij})} \sigma_{\rm i} \sigma_{\rm j}$, where $\sum_{({\rm ij})}$ denotes a sum over neighboring pairs of pixels. Monte Carlo simulations of this model generate binary “images,” and so many of the practical sampling questions are very similar to those in our problem. We see in Fig 2b that the specific heat again shows a peak which grows and moves toward the true critical temperature $T_c =1$ as we look at larger patches. Quantitatively the behavior is actually less dramatic than in the images, perhaps because the divergence of the specific heat in the thermodynamic limit is very gentle (logarithmic). Although this Ising spin system certainly is much simpler than the ensemble of images, comparison of Fig 2a and 2b supports the idea that we what we see in the images is consistent with an underlying divergence of the specific heat at a critical temperature close the to real temperature $T=1$.
Entropy vs. energy and Zipf’s law
=================================
A complementary perspective on thermodynamics is the microcanonical ensemble, corresponding to fixed energy rather than fixed temperature. The end result of the discussion will be an attempt to measure, for our image ensemble, the entropy as a function of energy. We begin with some standard results on how thermodynamic quantities are encoded in the plot of entropy vs. energy and on how we can identify a critical point in this plot.
{width="\linewidth"}
All thermodynamic quantities can be recovered from the partition function, $$Z(T) = \sum_{\vec \sigma} e^{-E(\vec\sigma ) /T} .$$ We can rewrite this sum by grouping together all states that have the same energy,
$$Z(T) = \sum_{\vec \sigma} e^{-E(\vec\sigma ) /T}
= \int dE \left[ \sum_{\vec \sigma} \delta \left( E - E(\vec\sigma )\right) \right] e^{-E/T} = \int dE\,\rho(E) e^{-E/T} ,$$
which defines the density of states $\rho(E)$. For a large system (patches with many pixels), the density of states becomes a smooth function, and we can define an entropy $S(E)$ at fixed energy as the log of the number of states in a narrow range of energies, so that $\rho(E) = (1/\Delta ) e^{S(E)}$. Then the partition function is $$Z(T) = {1\over \Delta} \int dE \,e^{S(E) - E/T} .$$ Further, both the energy and entropy are extensive variables which should be proportional to the size of the system, $N$; here $N$ will be the number of pixels in a patch. Then we define $\epsilon = E/N$ and $s(\epsilon ) = S(E = N\epsilon )/N$, and the partition function becomes $$Z(T) = {N\over \Delta} \int d\epsilon \,e^{N[s(\epsilon ) - \epsilon/T]} .$$ Now it is clear that, as $N$ becomes large, the integral will be dominated by the point where exponent is maximal, that is an energy such that $ds(\epsilon )/d\epsilon = dS(E)/dE = 1/T$. This connects the (microcanonical) description at fixed energy with the (canonical) description at fixed temperature. In addition, one can show that the specific heat is (inversely) related to the second derivative of the entropy, $$C = {N\over {T^2}}\left[ - {{d^2 s(\epsilon )}\over{d\epsilon^2}}\right]^{-1} .$$ In this language, the divergence of the specific heat occurs where the second derivative of the entropy vs. energy vanishes. This is the hallmark of a second order phase transition.
It is important to note that we can define $E$ for every state that we observe simply as the log of the probability, from Eq (\[basic\]), $E(\vec\sigma ) = -\ln P(\vec\sigma ) + c$, where $c$ defines the (arbitrary) zero of energy, which we choose so that the most probable state has zero energy. While it is tricky to measure the density of states or distribution of energies, it is easy to define the cumulative distribution, $${\cal N}(E) = \int_0^E dE'\,\rho(E' ) ,$$ which just counts the number of possible image patches for which the observed log probability is greater than $-E+c$. If $S(E)$ is increasing, then this integral is dominated by the behavior near its upper limit, so that $$\begin{aligned}
{\cal N}(E) &=& {N\over \Delta } \int_0^{E/N} d\epsilon \,e^{Ns(\epsilon )} \\
&\approx& {N\over \Delta} \left[ N{{ds(\epsilon )}\over{d\epsilon}}\right]^{-1} e^{Ns(\epsilon = E/N)}\\
\Rightarrow s(\epsilon ) &=& {1\over N} \ln {\cal N}(E =N\epsilon) + {{\ln(T/\Delta )}\over N} .\end{aligned}$$ Notice that the second term in this equation vanishes for large $N$, and so we approximate the entropy per pixel as a function of energy per pixel by the first term.
The results of the previous paragraph imply that if we just count the number of possible image patches with probability greater than a certain level, then we can construct the entropy vs. energy and hence derive all other thermodynamic functions. This is what we do in Fig. 3a for our image ensemble, using rectangular $L\times L'$ patches of size from 8 pixels up to 50 pixels; in Fig 3b we do the same thing for our Monte Carlo simulations of the Ising model. As a practical matter it is important to note that sampling problems are less serious at low energies (states with high probability), so we expect that even if we look at regions where we can’t sample the whole distribution we will get the correct low energy behavior.
We first note that the results on image patches of different sizes are remarkably consistent with one another, suggesting that we are seeing signs of the thermodynamic limit despite the small size of the regions that we can explore fully. Next, since the real image ensemble is at $T=1$, we want to pick out the energy at which $dS/dE = 1$; this seems very difficult, since the plot is very nearly a straight line with unit slope. But we know what this means: if the point where $dS/dE = 1$ is also a place where $d^2 S/dE^2 =0$, then $T=1$ is a critical point. Thus, [*the fact that $S/N$ vs. $E/N$ is very nearly a straight line of unit slope is direct evidence that the ensemble of natural images is at criticality.*]{}
{width="\linewidth"}
It is interesting that this approach to the thermodynamics of images is connected to Zipf’s law [@zipf]. We recall that Zipf estimated the probability distribution from which words are drawn in an English text, and argued that if we put the words in rank order ($r=1$ is the most common word), then $p_r \propto 1/r$ up to some maximum rank $r= N_w$ corresponding to the number of different words $N_w$ used in the text. Subsequently, other authors have considered generalized Zipf–like distributions [@newman], $p_r \propto 1/r^\alpha$, and there has been much discussion about the meaning of these relationships. Suppose we identify the Zipf–like distribution $p_r = A/r^\alpha$ with a Boltzmann distribution at $T=1$, $p_r = (1/Z)e^{-E_r}$. Then the energy of the state at rank $r$ is $E_r = \alpha \ln r - \ln(AZ)$. In the limit of a large system with many possible states, we can approximate the density of states by realizing the variable $r$ has a uniform distribution, and hence $\rho(E) \approx | dE_r /dr|^{-1}$; this gives $\rho(E) = r/\alpha$. But we also have $r = (AZ)^{1/\alpha} e^{E_r/\alpha}$, so we find $$\rho(E) = {1\over\alpha} (AZ)^{1/\alpha} e^{E/\alpha} \Rightarrow S_{\rm Zipf} (E) = E/\alpha + {\rm constant} .$$ Thus a generalized Zipf’s law is equivalent to an entropy that is (exactly!) linear in the energy. The original Zipf’s law ($\alpha =1$) corresponds to a unit slope, as we have found for image patches. Further, we have seen that this simple linear relation corresponds exactly to what we find for a thermodynamic system at a critical point [@note2].
Why does it matter that the ensemble of natural images is at a critical point? The signature of criticality is the divergence of the specific heat, and the specific heat is the variance in the energy, which is the log probability. Thus, being at a critical point means that the log probability has an enormously broad distribution, with a formally divergent second moment even once we normalize by the number of pixels. One consequence is that the approach toward typicality in the sense of information theory [@shannon48] will be much slower than one would find away from the critical point, which may be related to difficulties in compressing large natural images, or even in estimating their entropy (see, for example, Ref ). The large variance in log probability also means that there are large fluctuations in how surprised we should be by any given scene or segment of a scene, which perhaps quantifies our common experience.
The energy landscape
====================
Critical points mark the transition between phases characterized by different forms of order: liquid vs. gas, ferromagnet vs. paramagnet, and so on. What is the ordering that would emerge if somehow the distribution of natural images could be “cooled” from $T=1$ down toward $T=0$? This ultimately is a question about the nature of the image patches that correspond to the low energy states. Certainly the lowest energy states of small patches are solid black or white blocks, as in a ferromagnet where all the spins can align up or down, and these states will dominate at $T=0$. But, searching through all $4\times 4$ patches, we find $\sim 100$ states that are local minima of the energy, in the sense that flipping any single pixel from black to white (or vice versa) results in increased energy or reduced probability. In Figure 4a we show 49 of these states, ordered in decreasing probability. We see that many of these states are interpretable, for example as edges between dark and light regions, and that much of the multiplicity arises from the different ways of realizing these patterns (e.g., the six possible cases of a single vertical edge). We can think of these local minima in the energy landscape [@spin_glass] as being like the attractors in the Hopfield model of neural networks [@hopfield82], or like the code words in statistical mechanics approaches to error–correcting codes [@sourlas]. Usually we think of error–correcting coding as a construct, but here it seems that the signals which the world presents to us have some intrinsic error–correcting properties.
Although there are many reasons why edges may be important for vision, it is interesting to take seriously the idea that such image features acquire their importance because of their intrinsic properties of error correction, as if these are the signals that the world is “trying” to send us in the most fault tolerant fashion. If this is the case, then the visual system might build feature detecting neurons which serve to identify the basins of attraction defined by these local minima in energy. If such cells respond only when the original grey scale image corresponds to a discrete image within a particular basin of attraction, then it is easy to compute the response–triggered average within our natural image ensemble, with the results shown in Fig 4b. These results have a strong resemblance to the spike–triggered average responses of neurons in visual cortex to natural scenes [@sta]. Interestingly, if we look more closely at the problem of identifying the basin of attraction, we find that perceptron–like models based on filtering through a single receptive field do rather poorly. Thus, if visual cortex really builds a representation of the world based on the identification of these local minima in the energy landscape, the computations involved necessarily involve nonlinear combinations of multiple filters, as observed [@rust+al_05]. Much remains to be done to see if this really is a path to a theory of these more complex neural responses.
We thank D Chigirev, SE Palmer & E Schneidman for helpful discussions, and DL Ruderman for recovering the original data from Ref . This work was supported in part by National Science Foundation Grants IIS–0613435, IBN–0344678, and PHY–0650617, by National Institutes of Health Grant T32 MH065214, by the Human Frontier Science Program, and by the Swartz Foundation.
[99]{}
ER Davies, [*Machine Vision: Theory, Algorithms, Practicalities.*]{} (Morgan Kaufmann 2005).
HB Barlow, Possible principles underlying the transformation of sensory messages. In [*Sensory Communication*]{} W Rosenblith, ed, pp 217–234 (MIT Press, Cambridge, 1961).
JJ Atick, Could information theory provide an ecological theory of sensory processing? [*Network*]{} [**3,**]{} 213–251 (1999).
E Simoncelli & B Olshausen, Natural image statistics and neural representation. [*Ann Rev Neurosci*]{} [**24,**]{} 1193–1216 (2001).
W Bialek, Thinking about the brain. In [*Physics of Biomolecules and Cells: Les Houches Session LXXV,*]{} H Flyvbjerg, F Jülicher, P Ormos & F David, eds, pp 485–577 (EDP Sciences, Les Ulis; Springer–Verlag, Berlin, 2002); arXiv:physics/0205030.
DJ Field, Relations between the statistics of natural images and the response properties of cortical cells. [*J. Opt. Soc. Am. A.*]{} [**4,**]{} 2379–2394 (1987).
We simplify our notation by measuring temperature and energy in the same units.
J Cardy, [*Scaling and Renormalization in Statistical Physics*]{} (Cambridge University Press, Cambridge, 1996).
DL Ruderman & W Bialek, Statistics of natural images: Scaling in the woods, [*Phys Rev Lett*]{} [**73,**]{} 814–817 (1994).
DL Ruderman, The statistics of natural images, [*Network*]{} [**5,**]{} 517–548 (1994).
LP Kadanoff, Scaling laws for Ising models near $T_c$. [*Physics*]{} [**2**]{}, 263 (1966).
MJ Feigenbaum, MH Jensen & I Procaccia, Time ordering and the thermodynamics of strange sets: Theory and experimental tests. [*Phys Rev Lett*]{} [**57,**]{} 1503–1506 (1986). MJ Feigenbaum, Some characterizations of strange sets. [*J Stat Phys*]{} [**46,**]{} 919–924 (1987).
C Beck & F Schlögl, [*Thermodynamics of Chaotic Systems*]{} (Cambridge University Press, NY 1993).
L Onsager, Crystal statistics. I. A two–dimensional model with an order–disorder transition. [*Phys Rev*]{} [**65,**]{} 117–149 (1944).
CE Shannon, A mathematical theory of communication. 379–423 & 623–655 (1948).
D M Chandler & D J Field, Estimates of the information content and dimensionality of natural scenes from proximity distributions. [*J Opt Soc Am A*]{} [**24,**]{} 922–941 (2007).
GK Zipf, [*Selected Studies of the Principle of Relative Frequency in Language*]{} (Harvard University Press, Cambridge, 1932).
For a review see MEJ Newman, Power laws, Pareto distributions and Zipf’s law. [*Contemp phys*]{} [**46,**]{} 323–351 (2005).
Interestingly, if $\alpha\neq 1$ then there is no correspondence with thermodynamics.
M Mezard, G Parisi & M Virasoro [*Spin Glass Theory and Beyond*]{} (World Scientific, Singapore, 1987).
JJ Hopfield, Neural networks and physical systems with emergent collective computational abilities. [*Proc Nat’l Acad Sci (USA)*]{} [**79,**]{} 2554–2558 (1982).
N Sourlas, Spin glass models as error correcting codes, [*Nature*]{} [**339,**]{} 693-695 (1989).
D Smyth, B Willmore, GE Baker, ID Thompson & DJ Tolhurst, The receptive–field organization of simple cells in primary visual cortex of ferrets under natural scene stimulation. [*J Neurosci*]{} [**23,**]{} 4746–4759 (2003). FE Theunissen, SV David, NC Singh, A Hsu, WE Vinje & JL Gallant, Estimating spatio–temporal receptive fields of auditory and visual neurons from their responses to natural stimuli. [*Network*]{} [**12,**]{} 289–316 (2001).
NC Rust, O Schwartz, JA Movshon & EP Simoncelli, Spatiotemporal elements of macaque V1 receptive fields. [*Neuron*]{} [**46,**]{} 945–956 (2005).
[^1]: These authors contributed equally.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We extend and apply a recent theory of the dynamical spin response of Anderson lattice systems to interpret ESR data on YbRh$_{2}$Si$_{2}$. Starting within a semiphenomenological Fermi liquid description at low temperatures $T<T_{x}$ (a crossover temperature) and low magnetic fields $B\ll B_{x},$ we extend the description to the non-Fermi liquid regime by adopting a quasiparticle picture with effective mass and spin susceptibility varying logarithmically with energy/temperature, as observed in experiment. We find a *sharp* ESR resonance line slightly shifted from the local $f$-level resonance and broadened by quasiparticle scattering (taking unequal $g$-factors of conduction and $f$ electrons) and by spin-lattice relaxation, both significantly reduced by the effect of ferromagnetic fluctuations. A detailed comparison of our theory with the data shows excellent agreement in the Fermi liquid regime. In the non-Fermi liquid regime we find a close relation of the $T$-dependence of the specific heat/spin susceptibility with the observed $T$-dependence of line shift and linewidth.'
author:
- 'Peter W[" o]{}lfle'
- Elihu Abrahams
title: 'Phenomenology of ESR in heavy fermion systems: Fermi liquid and non-Fermi liquid regime '
---
Introduction
============
In several recent experiments [@YRS1; @Ce], low-temperature ESR has been observed in some heavy-fermion metals, in particular YRh$_{2}$Si$_{2}$ (YRS) [@YRS1; @YRS2]. The phase diagram of YRS has a magnetic-field induced quantum critical point and is a model system for the study of quantum criticality in the Kondo lattice. Consequently, the observation of a narrow ESR resonance in this compound aroused great interest, especially since it was commonly believed that heavy-fermion ESR would be unobservable due to an enormous intrinsic linewidth $\Delta B$ of order $k_{B}T_{K}/g\mu _{B}$ [YRS1]{}. Here $T_{K}$ is the lattice coherence (“Kondo") temperature for the onset of heavy-fermion behavior and $g\mu_{B}$ is the gyromagnetic ratio for the resonance. These were the first observations of ESR in Kondo lattice systems at $T<T_{K}$.
A common feature of the compounds in which ESR has been observed appears to be the existence of ferromagnetic fluctuations [@Ce; @FM] These findings challenge our understanding of heavy fermion compounds: How does a sharp electron spin resonance emerge despite Kondo screening and spin lattice relaxation, and why is this process influenced by ferromagnetic fluctuations? In a recent paper (AW") [@AW], we discussed the background of these questions and answered them in the framework of Fermi-liquid theory. An alternative explanation based on localized spins was subsequently proposed by Schlottmann [@schlottmann]. The general derivation of Fermi liquid theory from the microscopic theory for a two-band Anderson lattice model has been given by Yip [@Yip]
In YRS, the observed narrow dysonian [@feher] ESR line shape was originally interpreted [@YRS1] as indicating that the resonance was due to local spins at the Yb sites. Therefore, initially the authors speculated that the narrow ESR line might indicate the suppression of the Kondo effect near the quantum critical point, since, as explained above, carrying over Kondo impurity physics to the Kondo lattice, one might expect the local spins to be screened by the Kondo effect, giving rise only to a broad spin excitation peak, too wide to be observed in ESR experiments. However, a closer look [@ea] revealed that itinerant (heavy) electron ESR could give rise to a similar line shape since the carrier diffusion in YRS is quite slow. Thus, whether the resonance is that of localized or itinerant spins remained an open question.
In this paper, we extend our previous work [@AW] to the non-Fermi liquid (NFL) region of the YRS phase diagram and make a detailed comparison with the data. Excellent agreement is obtained for the Fermi-liquid (FL) regime. In particular, the ratio of the contributions $\propto T^{2}$ and $\propto
B^{2} $ to the linewidth in the FL region is very well reproduced. In addition, we account for the anomalous behavior observed in the NFL region for the resonance line shift and the linewidth, One absolutely essential aspect of our theory is the lattice coherence of the quasiparticles in the Anderson or Kondo lattice model: it is this lattice coherence that is responsible for the absence in the lattice case of the strong local spin relaxation that is observed in single Kondo impurity physics. Attempts to account for the observed logarithmic temperature dependence of the lineshift as arising from single Kondo ion physics above the Kondo temperature are therefore problematic, since lattice coherence is lost in that case
ESR in the Kondo-screened Anderson lattice model: Fermi liquid regime.
======================================================================
This was analyzed in Sec. III of AW. The Hamiltonian of the simplest Anderson lattice model, assuming momentum independent hybridization is given by $$\begin{aligned}
H &=& \sum_{\mathbf{k},\sigma }\epsilon _{\mathbf{k}\sigma }c_{\mathbf{k}\sigma
}^{+}c_{\mathbf{k}\sigma }+\sum_{i,\sigma }\epsilon _{f\sigma }n_{fi\sigma
}+U\sum_{i}n_{fi\uparrow }n_{fi\downarrow } \notag \\
&+& V\sum_{i,\mathbf{k},\sigma }(e^{i\mathbf{{k\cdot R}_{i}}}f_{i\sigma }^{+}c_{\mathbf{k}\sigma }+h.c.),\end{aligned}$$where $\epsilon _{\mathbf{k}\sigma }=\epsilon _{\mathbf{k}}-\omega
_{c}\sigma /2,$ $\sigma =\pm 1,$ is the conduction-electron energy spectrum and $\omega _{c}=g_{c}\mu _{B}B$ is its Zeeman splitting; $c_{\mathbf{k}\sigma }^{+},f_{i\sigma }^{+}$ are creation operators of the conduction electrons in momentum and spin eigenstates $(\mathbf{k}\sigma )$, and of electrons in the local $f$ level at site $\mathbf{{R}_{i}}$, respectively. The operator $n_{fi\sigma }=f_{i\sigma }^{+}f_{i\sigma }$ counts the number of electrons on the local level, and $\epsilon _{f\sigma }=\epsilon
_{f}-\omega _{f}\sigma /2$. $V$ and $U$ are the hybridization amplitude and the Coulomb interaction matrix element. We take the Zeeman splittings $\omega _{c}$ and $\omega _{f}$ to be unequal, as they are in real materials. We consider the limit $\omega _{f}-\omega _{c}\rightarrow 0$ in the Appendix.
We now review the results for the dynamical spin susceptibility $\chi
^{+-}(\Omega )$ obtained in AW [@AW]. We find a single resonance peak
$$\chi ^{+-}(\Omega )=\chi ^{+-}(0){\frac{-\omega _{r}+i\Gamma}{\Omega -\omega
_{r}+i\Gamma}},$$
where the resonance frequency is given by [@expl]
$$\omega _{r}=\omega _{f}-\frac{m}{m^{\ast }}(\omega _{f}-\omega _{c}),
\label{shift}$$
Here $m^{\ast }/m$ is the quasiparticle effective mass ratio. We note that for equal $g$-factors the line position is not shifted. In the Appendix, we discuss the complete result, showing that even in the case of unequal $g$-factors there is only a single resonance peak. We also show that the residual Fermi liquid interaction effects drop out of the resonance frequency.
The linewidth $\Gamma$ has contributions from quasiparticle scattering and from the conduction electron spin lattice relaxation $\gamma $
$$\Gamma =A[\alpha (\pi T)^{2}+\frac{1}{4}(R\omega _{f})^{2}]\frac{1}{R}+2\gamma \frac{m}{m^{\ast }}\frac{1}{R}. \label{width}$$
Here $R=[1+\widetilde{U}\chi ^{+-}(0)]$ is identified as the Wilson ratio, $\widetilde{U}$ is the Fermi-liquid spin-exchange interaction [@FL] and $\chi ^{+-}(0)=M/B$ is the static transverse spin susceptibility ($M$ is the spin polarization). The numerical coefficient $\alpha $ depends on the band structure and is of order unity. In the case of a sizeable ferromagnetic interaction, ($\widetilde{U}>0$), $R>>1,$ the linewidth gets narrowed by a factor $1/R.$ We suggest that this effect is responsible for the fact that so far an ESR line has only been observed in compounds that exhibit signatures of ferromagnetic fluctuations.
Magnetic anisotropy of ESR line
-------------------------------
The magnetic response of YRS is strongly anisotropic, largely because of the single ion anisotropy. The Zeeman Hamiltonian of the $f$-electron ground state doublet of Yb in tetragonal symmetry has the form $H_{Z}=-\mu
_{B}g_{f\perp }(S_{x}B_{x}+S_{y}B_{y})-\mu _{B}g_{f\parallel }S_{z}B_{z}$ , where the $z$-axis is along the crystallographic $c$-axis. The anisotropy of the $g$-factor is about a factor of $20$, $g_{f\perp }=3.6$ and $g_{f\parallel }=0.17$ . We assume for the present that the anisotropy of the interaction is negligible. The Hamiltonian is then diagonal in the coordinate system in the spin space that diagonalizes $H_{Z}$. The eigenvalues of $H_{Z}$ are found as
$$\omega _{f}(\phi )=\mp \mu _{B}B\sqrt{g_{f\parallel }^{2}\cos ^{2}\phi
+g_{f\perp }^{2}\sin ^{2}\phi }\ , \label{anis}$$
where $\phi $ is the angle between the magnetic field ${\mathbf{B}}$ and the $c$-axis.
The above results, Eq. (2), for the dynamical spin susceptibility $\chi
^{+-}(\Omega )$ obtained in [@AW] for the isotropic model may then be generalized to the anisotropic model by replacing $\omega _{f}$ by $\omega
_{f}(\phi )$ and taking the tensor of spin susceptibility projected onto the direction of the static magnetic field. According to Ref. [@ganis], the angle dependence of the resonance frequency is well-represented by Eq. ([anis]{}). This indicates that the anisotropic part of the residual Fermi-liquid spin-exchange interaction is small. As we shall see later, the temperature dependence of the line shift in the non-Fermi liquid regime suggests that there may be a small anisotropic interaction component. We shall explore the consequences of such a non-spin rotation invariant term in Sec. IIC, below.
ESR line shift and linewidth in the Fermi liquid regime
-------------------------------------------------------
In Fig. 1, we show a sketch of the phase diagram of YRS, including the $B,\; T$ ranges in which ESR experiments have been carried out. Here, the crossover to the Fermi liquid (FL) regime is determined by the onset of FL behavior in thermodynamic quantities [@see]. The $T^*$ crossover is primarily determined by Hall effect results [@hall; @sci] that can be interpreted as a transition (from left to right) to a large Fermi surface.
{width="0.8\linewidth"}
-1.5 cm
The high magnetic field ESR data reported in [@YRS2] show a crossover from a low-temperature Fermi-liquid (FL) like regime to a higher temperature non-Fermi liquid behavior at a temperature $T_{x}\simeq 5$ Kelvin. In the FL regime the line shift appears to be temperature independent. Relative to the ionic $g$-factor of $3.86$ [@YRS1], the resonance is shifted to lower values $g_{\perp}\simeq 3.42$ independent of magnetic field in the range $5.15$ to $7.45$ Tesla. Estimating the effective mass ratio, which is also temperature independent below $T_{x}$ in the magnetic field range considered, as $m^{\ast }/m\sim 40$, we obtain from Eq. (\[shift\]) a $g$-shift $\Delta g\simeq 0.04$, which is an order of magnitude too small. This discrepancy may point to an additional small anisotropic spin interaction, which we consider in the following subsection.
As for the linewidth, in the FL regime Schaufuß, et al [@YRS2] find a linewidth that follows the law $\Gamma (T,B)\sim T^{2}$, extrapolating to $\Gamma (0,B)\sim B^{2}$ as $T\rightarrow 0$. The experimental ratio of the prefactors of the $T^{2}$ and $B^{2}$ terms, $r_{exp}=B^{2}[\Gamma
(T,B)-\Gamma (0,B)]/[\Gamma (0,B)T^{2}]$ turns out to be $r_{exp}\sim 2$. Estimating the Wilson ratio from the available specific heat data [trov,NJP]{} at $T_{x}=5$ K and in magnetic fields $B\approx 6$ Tesla, $\Delta
C/T=0.032$ K$^{-1}$Yb$^{-1}$ and spin susceptibility data [@NJP; @JPS] $\Delta \chi =M/B=0.224\mu _{B}^{2}$ K$^{-1}$Yb$^{-1}$ as $R=[\chi /(g_{f}\mu
_{B}/2)^{2}](\pi ^{2}T/3\Delta C)\simeq 7.5$, we calculate from Eq. ([width]{}) the theoretical ratio $r_{th}\sim 1.2\alpha $, in good agreement with the experimental value. Note that the large enhancement of the single particle Zeeman splitting by the Fermi liquid interaction (a factor $R$) is essential in obtaining this agreement.
Effect of non spin-rotation invariant Fermi liquid interaction.
---------------------------------------------------------------
The spin-orbit interaction in conjunction with the tetragonal lattice anisotropy may be expected to lead to a small admixture of a non-spin-symmetric component to the Fermi liquid interaction of the form $-4I(\overrightarrow{S}\cdot \hat{c})^{2}$ , where $\hat{c}$ is the unit vector along the $c$-axis of the tetragonal lattice. Taking the magnetic field along the $b$-axis, we employ a coordinate system in spin space, in which the $z$-axis is oriented along the magnetic field and the $x$-axis along the $c$-axis (see Fig. 2).
-3cm
The ESR oscillating transverse magnetic field is circularly polarized in the $x-y$ plane, which is the $a-c$ plane of the crystal, perpendicular to the static magnetic field. The screening of the static magnetic field is effected in linear order in $I$ only when the component of the magnetic field along $\hat{c}$ is nonvanishing. In turn, the dynamic screening is changed at linear order in $I$, for any component of $\mathbf{B}_{static}$ perpendicular to $\hat{c} $ . Thus, the static and dynamic screening are effected differently by $I$, which gives rise to a resonance line shift, which we now calculate. As we show in the Appendix, the dynamical screening of the $ff$ component of the dynamical susceptibility is modified in the presence of $I$ to $$\chi _{ff}^{+-}(i\Omega _{m})=\chi _{ff,H}^{+-}(i\Omega _{m})[1+\widetilde{U}_{d}\chi _{ff}^{+-}(i\Omega _{m})]$$Thus, in the transverse spin response, the Fermi-liquid interaction is changed to $\widetilde{U}_{d}=\widetilde{U}+I\sin ^{2}\phi $. The notation is as in AW [@AW], Eqs. (16-18): $\chi _{ff,H}$ is the susceptibility of Fermi-liquid quasiparticles in the absence of vertex corrections (bubble diagram only) and the subscript $H$ indicates that the bare Zeeman energy $\omega _{f}$ is replaced everywhere by $$\widetilde{\omega }_{f}=\omega _{f}[1+\widetilde{U}_{s}\chi
_{ff}^{+-}(0)]=\omega _{f}[1-\widetilde{U}_{s}\chi _{ff,H}^{+-}(0)]^{-1}.$$Here ${\widetilde{U}}_{s}= {\widetilde{U}}+4I\cos^2\phi$ and ${\widetilde{U}}
$ is the renormalized onsite $ff$ repulsion that appears in the effective low-energy Hamiltonian (see AW for further details). The resonance position is therefore shifted as
$$\omega _{r}=\widetilde{\omega }_{\mathbf{k}_{F}}^{-}[1-\widetilde{U}_{d}\chi
_{ff,H}^{+-}(0)]=\omega _{k_{F}}^{-}\frac{1-\widetilde{U}_{d}\chi
_{ff,H}^{+-}(0)}{1-\ \widetilde{U}_{s}\chi _{ff,H}^{+-}(0)}$$
Substituting $\omega _{k_{F}}^{-}$ as obtained in the Appendix and using the definition of $R$, we find that the resonance frequency is modified from Eq. (3) to $$\begin{aligned}
\omega _{r}&\simeq&\omega _{k_{F}}^{-}[1-I\chi _{ff,H}^{+-}(0)R] \notag \\
&\simeq&\omega _{f}-\frac{m}{m^{\ast }}(\omega _{f}-\omega _{c})-\omega
_{f}I(1-5\cos ^{2}\phi )\chi _{ff}^{+-}(0), \notag\end{aligned}$$ or $$g_{r} \simeq g_{f}-\frac{m}{m^{\ast }}(g_{f}-g_{c})-g_{f}I(1-5\cos ^{2}\phi )\chi _{ff}^{+-}(0).$$
Since most of the ESR data have been taken in the configuration of magnetic field perpendicular to the $c$-axis, *i.e*. $\phi =\pi /2$ , we concentrate on this case from now on. We may try to determine $I$ by fitting the low-temperature line shift [@YRS2]. In the following, we assume $I$ to be independent of temperature and magnetic field. In the NFL regime, experimentally $\chi _{ff}^{+-}(0;T)$ is a decreasing function of temperature, such that the $g$-shift increases as observed in experiment, provided $I>0$. We may relate $I$ to $\chi _{ff}^{+-}(0;T_{1})$ at a reference point $\ T_{1}=4$K, $B=0.2$T, where $\chi
_{ff}^{+-}(0;T_{1})\approx 1.3\times 10^{-6}$m$^{3}$/mol [@trov2] as $I=(g_{f}^{ion}-g_{r})/(g_{f}\chi _{ff}^{+-})$ \[the reference $g$-factor $g_{f}^{ion}$ is actually reduced by the factor $(1-m/m^{\ast }$)\]. The data at low magnetic fields, $B=0.18$T and $B=0.68$T show a $g$-factor of $g\simeq 3.5$ at the lowest temperature, $T=2$K, whereas the high-field data show $g=3.42$ in the Fermi liquid regime. It follows that $I\approx
0.075\times 10^{6}$ m$^{-3}$mol. From a comparison with the temperature dependence of the resonance frequency we determine in the next Section a value of $I=0.063\times 10^{6}$ m$^{-3}$mol, which agrees very well with the independently-obtained value above. We observe in passing that the magnetic susceptibility data indicate that in the non-Fermi liquid regime $\widetilde{U}$ appears to depend on both temperature and magnetic field.
ESR in the non-Fermi liquid regime.
===================================
We now attempt to phenomenologically relate the framework we have set up to the ESR data in the NFL regime by using the observed specific heat and susceptibility $T$ and $B$ dependences. The non-Fermi liquid behavior in the temperature range $T>T_{x}$ appears in the ESR data as a nearly logarithmic increase of the $g$-factor with temperature and a change in the temperature dependence of the linewidth from $T^{2}$ to $T$ . This change into the NFL regime occurs at about the same temperature as the observed changes in the specific heat and spin susceptibility.
Resonance shift
---------------
In the theoretical resonance shift, Eq. (6), the $T$- and $B$- dependences enter in two ways, if we continue to assume that the temperature dependence of the anisotropic Fermi-liquid interaction parameter $I$ may be neglected: 1) through the susceptibility $\chi _{ff}^{+-}(0)$, which we get from experiment, and 2) through the effective mass ratio, which we extract from the measured specific heat $\gamma $-coefficient ($\Delta C=\gamma T$), by taking $\gamma \propto m^{\ast }/m$. Using these experimentally determined quantities, we shall use Eq. (6) to evaluate the theoretical resonance shift and compare it to the observed one. We shall use Eq. (6) to calculate the $g$-shift $\delta g=g-g_{f}^{ion}$ at two reference temperatures $T=4$K and $T=10$K. The inputs are the values of $m/m^{\ast }$ from the observed specific heat [@trov], taking $m^{\ast
}/m=40$ at $T=5$ K and $B=6$ T as a reference point, the observed susceptibility [@JPS; @trov2] and the value of the anisotropic FL interaction $I$. These data and the calculated $g$-shifts are collected in Table I. Since we have assumed that $I$ is independent of $T$ and $B$, we can evaluate it from Eq. (6) using experimental data at $T=4$K, $B=0.2$T as discussed in the previous Section. The result is $I=0.063\times 10^{6}$ mol/m$^{3}$ The data on the ESR line shift given in Fig. 2 of [@YRS2] show an approximately linear $\ln T$ dependence in the $T$-range $4K<T<10K$. Therefore, to check the accuracy of our theoretical result, Eq. (6), we fit the two calculated $\delta g$ values to a linear $\ln T$ function and give the resulting theoretical slope in the last column of Table I. The comparison of the calculated and observed values of the slope $\Delta
g/\Delta \ln T$ is shown in Fig. 3. It is seen that the agreement is quite good, supporting our assumption of a constant interaction $I$. The theory explains the rather strong dependence of the $\ln T$ term on magnetic field, decreasing by approximately a factor of $8$ as the magnetic field is stepped up from $0.19$ T to $7.45$ T. The present theory predicts that the slope of the $\ln T$ dependence of the $g$-shift depends sensitively on the magnetic field orientation (the angle $\phi $) and it reverses sign when $\mathbf{B}_{static}$ is oriented along the ${\hat{c}}$-axis of the crystal.
$B$(T) $\chi(T_1)$ $\chi(T_2)$ $\gamma(T_1)$ $\gamma(T_2)$ $\delta g (T_1)$ $\delta g (T_2)$ $(\Delta g/\Delta\ln T)_{th}$
-------- ------------- ------------- --------------- --------------- ------------------ ------------------ -------------------------------
7.5 0.82 0.65 0.24 0.185 -0.222 -0.197 0.027
6.0 0.91 0.70 0.26 0.18 -0.239 -0.210 0.032
5.0 0.98 0.74 0.27 0.17 -0.254 -0.222 0.035
1.85 1.13 0.78 0.29 0.15 -0.283 -0.238 0.050
1.0 1.2 0.80 0.29 0.15 -0.298 -0.243 0.060
0.68 1.23 0.80 0.29 0.15 -0.305 -0.243 0.068
0.5 1.25 0.80 0.29 0.15 -0.309 -0.243 0.072
0.2 1.3 0.80 0.29 0.15 -0.320 -0.243 0.085
: Experimental values of susceptibility and specific heat coefficient and calculated $g$-shift \[from Eq. (6)\] at different $B$ and two temperatures $T_{1}=4$K, $T_{2}=10$K. We chose $I=0.063\times 10^{6}$ mol/m$^{3}$. Units: $\protect\chi $ in $10^{-6}$m$^{3}/$mol, $\protect\gamma
$ in J/(mol K$^{2}$)
{width="0.9\linewidth"}
-1.4 cm
Linewidth
---------
Turning now to the linewidth, we use the analyticity properties of the self energy $\Sigma (\omega )$ to infer the linewidth from the temperature dependence of the specific-heat coefficient $\gamma (T)$. Here one has to observe that only part of the specific heat enhancement is coming from the nonanalytic contribution of the self energy $\Sigma $. An additional part is coming from the regular (analytic) contribution to $\Sigma $. Therefore one may split the effective mass into two components, $m^{\ast }/m=(m^{\ast
}/m)_{reg}+(m^{\ast }/m)_{sing}$. The specific heat data show a $\ln T$ variation over a wide range on top of a background. If we identify the background with $(m^{\ast }/m)_{reg}$, the singular part at the reference point $T=5$ K and $B=6$ T is about $60\%$ of the total, i.e. $(m^{\ast
}/m)_{s}=24$, taking the FL value $m^{\ast }/m=40$. The singular part can now be associated with the non-Fermi-liquid logarithmic temperature dependence of $\gamma $. Thus, $\gamma _{sing}\propto (m^{\ast }/m)_{sing}$. To account for the crossover from NFL to FL behavior at $T=T_{x}$, we adopt an interpolation formula $\gamma _{sing}=-c\ln [(T^{2}+T_{x}^{2})/T_{0}^{2}]$. Since $(m^{\ast }/m)=[1-\mathrm{Re}\{\partial \Sigma /\partial \omega
\}|_{0}]$, we may write $(m^{\ast }/m)_{sing}=-\mathrm{Re}\{\partial \Sigma
_{sing}/\partial \omega \}|_{0}$. The temperature dependence $(m^{\ast
}/m)_{sing}=-a\ln [(T^{2}+T_{x}^{2})/T_{0}^{2}]$ may be approximately converted into a frequency dependence $\mathrm{Re}\{\partial \Sigma
_{sing}/\partial \omega \}=a\ln [(\omega ^{2}+T_{x}^{2})/T_{0}^{2}]$ of the nonanalytic real part of the self energy. The self energy in the complex plane may be inferred as
$$\Sigma _{sing}(\omega )=2a\omega \ln [(-i\omega +T_{x})/T_{0}] \notag$$
From this approximate model the imaginary part of the self energy and hence the quasiparticle contribution \[the first term in Eq. (4)\] to the ESR line width follows as
$$\Gamma _{qp}=2\frac{m}{m^{\ast }}\frac{1}{R}\mathrm{Im}\Sigma _{s}(\omega
=T)=pT\tan ^{-1}(T/T_{x}) \notag$$
In the limit of $T\ll T_{x}$ the above expression recovers the Fermi liquid result $\Gamma _{qp}\propto T^{2}$, as discussed above, while at $T\gg T_{x}$ the non-Fermi liquid result $\Gamma _{qp}\propto T$ is obtained. This is in qualitative agreement with experiment. It is worth pointing out that a similar structure of the self energy has been proposed for the strange metal“ phase of the cuprates under the name marginal Fermi liquid theory” [@mfl]. By comparison with the effective mass ratio we find $a_{th}\approx 6$ and using $R\approx 7
$ we get $p_{th}\approx 0.1$. The experimental value is $p_{ex1}\approx 0.02$, which is quite a bit smaller. Similarly, from the experimentally observed coefficient of the $T^{2}$ term of the linewidth, $\Gamma /T^{2}\approx 0.004
$ K$^{-1}$ in the Fermi liquid regime, one extracts again a value $p_{ex2}\approx 0.02$ . The discrepancy may come from our assumption that $\Sigma _{sing}$ is entirely due to spin-flip scattering and from our very approximate determination of $\Sigma _{sing}$. Vertex corrections will for example remove any non-spin interaction contribution to $\Sigma _{sing}$ from the linewidth $\Gamma $. If this is correct, it would imply that the fluctuations contributing most to the $\ln T$ term in the specific heat are nonmagnetic in origin. Finally we comment on the possible contribution to $\Gamma $ caused by the regular part of $\Sigma _{reg}(\omega )$. In this case the prefactors $c_{r},c_{i}$ of the low energy limiting forms $\mathrm{Re}\Sigma _{reg}(\omega )=\Sigma _{reg}(0)+c_{r}\omega $ and $\mathrm{Im}\Sigma _{reg}(\omega )=c_{i}\omega ^{2}$ are not directly related. The Kramers-Kronig relations imply in this case that, e.g., the higher frequency parts of $\mathrm{Re}\Sigma _{reg}(\omega )$ will predominantly determine the coefficient $c_{i}$, while the coefficient $c_{r}$ has little influence in this. In the present case, the resulting imaginary part and coefficient $c_{i}$ is apparently small.
Conclusion
==========
We extended and applied our recent theory [@AW] of the dynamical spin response of Anderson lattice systems to interpret ESR data on YbRh$_{2}$Si$_{2}$. Starting with a semiphenomenological Fermi-liquid description at low temperatures $T<T_{x}$ (a crossover temperature) and low magnetic fields $B\ll B_{x},$ we extended the description to the non-Fermi liquid regime by adopting a quasiparticle picture with effective mass and spin susceptibility varying logarithmically with energy/temperature, as observed in experiments. We find a *sharp* ESR resonance line that is broadened by quasiparticle scattering and spin lattice relaxation, both significantly reduced by the effect of ferromagnetic fluctuations. A more complete evaluation of the results presented in our first paper shows that the ESR-line position is shifted by an amount $\propto (g_{f}-g_{c})$ , thus reducing to zero for equal $g$-factors. In the case of different $g$-factors there is only one sharp resonance line at $\omega \simeq \omega _{f}$, the local $f$-resonance frequency. The observed strong anisotropy of the ESR response is shown to follow from the single-ion spin anisotropy, assuming an approximately spin-conserving exchange interaction.
A detailed comparison of our theory with the data shows excellent agreement in the Fermi-liquid regime, when the model is amended by a small anisotropic part of the spin exchange interaction, induced by spin-orbit coupling. We assumed the strength of the latter to be independent of temperature and magnetic field throughout the regime considered in the experiments. In particular, the ratio of the contributions $\propto T^{2}$ and $\propto
B^{2} $ to the linewidth in the FL region is very well reproduced by theory.
In the non-Fermi liquid regime we find a close relation of the $T$-dependences of the specific heat and spin susceptibility with the observed $T$-dependences of the line shift and linewidth. There are two terms contributing to the temperature dependence of the lineshift \[Eq. (6)\] in opposite ways. The first and dominant one is proportional to the spin susceptibility $\chi $ and leads to a resonance frequency increasing with temperature, while the second and smaller one is proportional to the inverse specific heat coefficient $1/\gamma $, leading to a decreasing behavior. The observed approximately linear $\ln T$ dependence of the $g$-shift is determined by $\chi $ and $1/\gamma $, (in the restricted temperature regime, where the remaining curvature in both $\chi $ and $1/\gamma $ tends to compensate), while the magnitude of the shift is fitted by adjusting the anisotropic exchange interaction constant $I$. The observed rather strong magnetic field dependence of the prefactor of $\ln T$ is very well accounted for by $\chi $. Finally, we attempted to relate the linewidth to the singular part of the self energy $\Sigma _{sing}$, by identifying the $\ln T$ contribution to the specfic heat coefficient with the effective mass deduced from $\Sigma _{sing}$. On a qualitative level, the observed crossover from $T^{2}$ to linear $T$ behavior of the linewidth upon entering the non-Fermi liquid regime is reproduced. However, the line width is found to be approximately a factor of $5$ too large compared to experiment. The most likely explanation of this discrepancy is our neglect of vertex corrections, which would remove any non-magnetic contribution to $\Sigma _{sing}$ from the linewidth.
Overall the extended quasiparticle picture used to account for the ESR-properties in the non-Fermi liquid regime appears to work quite well. It would be interesting to compare with data taken at much lower temperatures, when the present theory would predict, e.g. nonlinear variation of the $g$-shift with $\ln T$, as exhibited by $\chi $. Also, a cleaner identification of the linear $T$ dependence of the linewidth would be essential to corroborate the extended quasiparticle picture. Finally, our theory makes definite predictions for the anisotropy of the lineshift.
We thank the Aspen Center for Physics, where part of this work was completed. P.W. acknowledges partial support from the DFG-Center for Functional Nanostructures and the DFG-Forschergruppe “Quantum Phase Transitions".
Derivation of dynamical susceptibility: spin-rotation invariant Fermi liquid interaction
----------------------------------------------------------------------------------------
In this Appendix we derive the quasiparticle properties and the dynamical spin susceptibility in a more detailed and complete way than was done in AW [@AW]. We start from the Dyson equation for the single particle Green’s functions,
$$\mathcal{G}^{-1}\mathcal{G} =
\begin{pmatrix}
i\omega _{n}-\epsilon _{f\sigma }-\Sigma _{f\sigma }(i\omega _{n},\mathbf{k})
& -V \\
-V & i\omega _{n}-\epsilon _{\mathbf{k\sigma }}-\Sigma _{c\sigma }(i\omega_n
,\mathbf{k})\end{pmatrix}\begin{pmatrix}
G_{\mathbf{k}\sigma }^{ff} & G_{\mathbf{k}\sigma }^{cf} \\
G_{\mathbf{k}\sigma }^{fc} & G_{\mathbf{k}\sigma }^{cc}\end{pmatrix}=\mathbf{1}.$$
We approximate the conduction electron retarded self-energy by $\Sigma
_{c\sigma }(\omega +i0,\mathbf{k})=-i\gamma $, where $\gamma$ is the conduction-electron spin-lattice relaxation rate. To carefully derive the quasiparticle Zeeman energies, we make use of the conservation of total spin in the model considered here, to remove the $f$-electron Zeeman term of the Hamiltonian by performing a gauge transformation that shifts the zero of energy of $\uparrow $-spins and $\downarrow $-spins by $\mp \omega _{f}/2$ respectively. As a consequence $i\omega _{n}$ is replaced by $i\omega
_{n}+\sigma \omega _{f}/2$ , $\epsilon _{f\sigma }$ by $\epsilon _{f}$ and the conduction electron Zeeman energy is changed to $-\sigma (\omega
_{f}-\omega _{c})/2$. For an isotropic band structure ($\epsilon _{\mathbf{k}}=\epsilon _{k}$), the magnetic field dependence of the self energy (neglecting small band edge terms) is then of the form $$\Sigma _{f\sigma }(\omega +i0,\mathbf{k})\simeq \Sigma _{f}(\omega +\sigma
\omega _{f}/2+i0,\epsilon _{k}-\sigma (\omega _{f}-\omega _{c})/2).$$ In $\mathcal{G}^{-1}_{11}$, we expand this $f$-self energy about the Fermi energy:
$$\begin{aligned}
\omega +\sigma \omega _{f}/2-\epsilon _{f}-\Sigma _{f\sigma }(\omega +i0,\mathbf{k}) &=&(\omega +\sigma \omega _{f}/2)(1-\partial \Sigma
_{f}/\partial \omega |_{0})-\epsilon _{f}-\Sigma _{f}(i0,\epsilon _{k_{F}})
\notag \\
&-&(\partial \Sigma _{f}/\partial \epsilon _{k}|_{k_{F}})(\epsilon
_{k}-\epsilon _{k_{F}}-\sigma (\omega _{f}-\omega _{c})/2)+i\mathrm{Im}\Sigma _{f}(\omega +i0,\epsilon _{k}) \notag \\
&=&z^{-1}[\omega -\widetilde{\epsilon }_{fk\sigma }+i\gamma _{fk}],\end{aligned}$$
where $z^{-1}=1-\partial \Sigma_f /\partial \omega |_{0}$, $\widetilde{\epsilon }_{fk\sigma }=\widetilde{\epsilon }_{fk}-\sigma \omega
_{f}/2+\sigma (\partial \Sigma _{f}/\partial \epsilon _{k}|_{k_{F}})(\omega
_{f}-\omega _{c})/2$, $\widetilde{\epsilon }_{fk}=z[\epsilon _{f}+\Sigma
_{f}(i0,\epsilon _{k_{F}})+(\partial \Sigma _{f}/\partial \epsilon
_{k}|_{k_{F}})(\epsilon _{k}-\epsilon _{k_{F}})]$, and $\gamma _{fk}=z\mathrm{Im}\Sigma _{f}(\omega +i0,\epsilon _{k})$
Then for low energies one has a quasiparticle description, with $G_{\mathbf{k}\sigma }^{ff}(\omega )=z_{\sigma }\widetilde{G}_{\mathbf{k}\sigma }^{ff}$, $G_{\mathbf{k}\sigma }^{cf}=\sqrt{z_{\sigma }}\widetilde{G}_{\mathbf{k}\sigma }^{cf}$ and the renormalized hybridization amplitude $\widetilde{V}^{2}$ $=z_{\sigma }V^{2}$. The complex energy eigenvalues are given by $$\zeta _{\mathbf{k\sigma }}^{\pm }=\frac{1}{2}(\widetilde{\epsilon }_{f\sigma
}-i\gamma _{fk}+\epsilon _{\mathbf{k\sigma }}-i\gamma )\pm \sqrt{\frac{1}{4}(\widetilde{\epsilon }_{f\sigma }-i\gamma _{fk}-\epsilon _{\mathbf{k\sigma }}+i\gamma )^{2}+\widetilde{V}^{2}}=\epsilon _{\mathbf{k}}^{\pm }-\frac{1}{2}\sigma \omega _{\mathbf{k}}^{\pm }-i\gamma _{\mathbf{k}}^{\pm },$$where, assuming $|\epsilon _{\mathbf{k}_{F}}|>>$ $|\widetilde{\epsilon }_{fk}|,\widetilde{V}$, $$\begin{aligned}
\epsilon _{\mathbf{k}}^{\pm }&=&\epsilon _{k_{F}}^{\pm }+\frac{1}{2}(\epsilon _{k}-\epsilon _{k_{F}})(1+z(\partial \Sigma _{f}/\partial \epsilon
_{k}|_{k_{F}}))[1\pm {\frac{\epsilon _{\mathbf{k}_{F}}-\widetilde{\epsilon }_{f}}{\sqrt{(\widetilde{\epsilon }_{f}-\epsilon _{\mathbf{k}_{F}})^{2}+4\widetilde{V}^{2}}}]},\cr \omega _{\mathbf{k}}^{\pm }&=&\omega _{f}+\frac{1}{2}[(\omega _{c}-\omega _{f})(1+z(\partial \Sigma _{f}/\partial \epsilon
_{k}|_{k_{F}}))[1\pm {\frac{\epsilon _{\mathbf{k}_{F}}-\widetilde{\epsilon }_{f}}{\sqrt{(\widetilde{\epsilon }_{f}-\epsilon _{\mathbf{k}_{F}})^{2}+4\widetilde{V}^{2}}}]}\cr \gamma _{\mathbf{k\sigma }}^{\pm }&=&\frac{1}{2}(\gamma _{fk}+\gamma )\mp \frac{1}{2}(\gamma _{fk}-\gamma )\frac{\epsilon _{\mathbf{k\sigma }}-\widetilde{\epsilon }_{f\sigma }}{\sqrt{(\widetilde{\epsilon }_{f}-\epsilon _{\mathbf{k}_{F}})^{2}+4\widetilde{V}^{2}}}\end{aligned}$$
From now on we can safely neglect the term involving $(\partial \Sigma
_{f}/\partial \epsilon _{k}|_{k_{F}})$, since it is small - of order $z<<1$.
Using partial fraction decomposition, we construct the retarded Green function $$\widetilde{G}_{\mathbf{k}\sigma }^{ff}(\omega +i0)={\frac{a_{\mathbf{k}\sigma }^{ff,+}}{\omega -\zeta _{\mathbf{k\sigma }}^{+}}}+{\frac{a_{\mathbf{k}\sigma }^{ff,-}}{\omega -\zeta _{\mathbf{k\sigma }}^{-}}}$$and similar expressions for $\widetilde{G}_{\mathbf{k}\sigma }^{cf}$ and $G_{\mathbf{k}\sigma }^{cc}$, where, with $u_{\mathbf{k}\sigma }=\zeta _{\mathbf{k}\sigma }^{+}-\zeta _{\mathbf{k}\sigma }^{-}$, $$\begin{aligned}
a_{\mathbf{k}\sigma }^{ff,\pm }&=&\pm (\zeta _{\mathbf{k\sigma }}^{\pm }-\widetilde{\epsilon }_{\mathbf{k}\sigma })/u_{\mathbf{k}\sigma }, \\
a_{\mathbf{k}\sigma }^{cc,\pm }&=&\pm (\zeta _{\mathbf{k\sigma }}^{\pm }-\epsilon
_{f\mathbf{\sigma }})/u_{\mathbf{k}\sigma }\\
a_{\mathbf{k}\sigma
}^{cf,\pm }&=&\pm \widetilde{V}/u_{\mathbf{k}\sigma }.\end{aligned}$$For sufficiently small imaginary parts, $\gamma \ll (\widetilde{V},\widetilde{\epsilon }_{f\sigma })$, we may neglect them in the weight factors $a_{\mathbf{k}\sigma }^{ff,\pm },...$ and replace $\zeta _{\mathbf{k\sigma }}^{\pm }$ by $\epsilon _{\mathbf{k\sigma }}^{\pm }$.
The quasiparticles interact via the residual Fermi liquid interaction. For ESR, the relevant component of the Fermi liquid interaction is the spatially isotropic spin-antisymmetric part described by the Landau parameter $F_{0}^{a}=-2N_{0}\widetilde{U}$ . Here $\widetilde{U}$ is the coupling constant of a spin isotropic exchange interaction $H_{ex}=-$ $\widetilde{U}\overrightarrow{S}\cdot \overrightarrow{S}$ , which leads to a quasiparticle energy shift $\delta \omega _{\mathbf{k}}=$ $\widetilde{U}M$, where the spin polarization $M$ is given by $M=\chi _{H}\widetilde{\omega }_{k_{F}}$, with $\chi _{H}$ the unscreened static spin susceptibility (in the absence of the Fermi liquid interaction) and $\widetilde{\omega }_{k_{F}}$ the fully renormalized quasiparticle Zeeman splitting at the Fermi energy.
For definiteness in the following we assume a band filling of somewhat less than two electrons per site, such that the Fermi level lies in the lower quasiparticle band (energy $\zeta _{\mathbf{k\sigma }}^{\pm }$), and $\widetilde{\epsilon }_{f\sigma }$ is close to the Fermi energy. The screened Zeeman splitting is then obtained by solving the self-consistent equation $\widetilde{\omega }_{k_{F}}=\omega _{k_{F}}^{-}+$ $\widetilde{U}\chi _{H}\widetilde{\omega }_{k_{F}}$. The solution is $$\widetilde{\omega }_{k_{F}}=\frac{\omega _{k_{F}}^{-}}{1-\ \widetilde{U}\chi
_{H}}=\omega _{k_{F}}^{-}R,$$ where $R=1/(1-\ \widetilde{U}\chi _{H})$ is the Wilson ratio. One observes that in the case of ferromagnetic correlations, when $R>>1$, the single particle Zeeman splitting is enhanced by a large factor. We conjecture that this effect should be observable in tunneling experiments. In the two-particle spectrum probed by electron spin resonance, the enhancement is completely removed by dynamical screening (see AW).
As derived in AW, the dynamical transverse susceptibility $\chi ^{+-}(\Omega
)$, where $\Omega $ is the frequency of an a.c. electromagnetic field polarized transverse to the static magnetic field, is given by $$\chi ^{+-}(\Omega )=\mu _{B}^{2}[g_{c}^{2}\chi _{cc}^{+-}(\Omega
)+g_{f}^{2}\chi _{ff}^{+-}(\Omega )+2g_{c}g_{f}\chi _{cf}^{+-}(\Omega )].$$ The partial susceptibilities are obtained by evaluating Feynman bubble diagrams dressed by vertex corrections of the ladder type referring to the Fermi liquid interaction (local electrons) and the spin-orbit interaction (impurity correlation lines for the conduction electrons). The final result obtained in AW may be reexpressed as
$$\begin{aligned}
\chi ^{+-}(\Omega +i0) &=&\mu _{B}^{2}\{[(g_{f}\chi _{ff}^{+-}(0)+g_{c}\chi
_{cf}^{+-}(0))^{2}/\chi _{ff}^{+-}(0)]\frac{-\omega _{r}+i\Gamma}{\Omega
-\omega _{r}+i\Gamma} \\
&+& g_{c}^{2}\frac{-\omega _{\mathbf{k}_{F}}^{-}+2i\gamma _{\mathbf{k}_{F}}^{-}}{\Omega -\omega _{\mathbf{k}_{F}}^{-}+2i\gamma _{\mathbf{k}_{F}}^{-}}[\chi _{cc,H}^{+-}(0)-(\chi _{cf,H}^{+-}(0))^{2}/\chi
_{ff,H}^{+-}(0)]\}\end{aligned}$$
There appear to be two different resonance denominators in the above expression. Only the resonance at $\omega _{r}$ has been observed in ESR experiments The resonance at $\omega _{\mathbf{k}_{F}}^{-}$, is shifted to much higher frequencies (the factor $R$ !). However, a closer look reveals that the weight of the resonance at $\omega _{\mathbf{k}_{F}}^{-}$ is zero in the regime considered. Indeed, by inserting the quasiparticle weight factors, $\chi _{ab,H}^{+-}(0)\simeq 2N_{0}(a_{\mathbf{k}_{F}\sigma
}^{ab,-})^{2}$ , where $a,b$ is $c\ $or $f$, one finds that the prefactor of the second resonance is zero. In the limit of equal g-factors, the prefactor of the first resonance simplifies to $2N_{0}R$ and its position is unshifted, $\omega _{r}=\omega _{f}$ . In the absence of spin lattice relaxation, the linewidth shrinks to zero in this case. This is not reproduced in the above calculation, since we did not take into account the vertex corrections belonging to the imaginary part of the self energy (the scattering-in term in Boltzmann equation language).
Non-spin rotation invariant Fermi liquid interaction
----------------------------------------------------
The relatively large $g$-shift observed in experiment suggests the presence of a small additional spin-symmetry breaking Fermi liquid interaction $\mathbf{I}$. Since this interaction can only be mediated by the spin-orbit interaction it should have preferred direction given by the lattice symmetry. We therefore assume the form $\mathbf{I}_{\alpha\beta
;\gamma\delta}=-I(\mathbf{c\cdot \boldsymbol{\tau } }_{\alpha\beta })(\mathbf{c\cdot \boldsymbol{\tau } }_{\gamma\delta})$, where $\boldsymbol{\tau }$ is the vector of Pauli matrices and $\mathbf{c}$ is a unit vector in the direction of the crystallographic $c$-axis of the tetragonal lattice. The screened and unscreened tensor susceptibilities $\mathbf{X},\mathbf{X}_{H}$, where $\mathbf{X} _{ij}=\langle\langle S_{i};S_{j}\rangle\rangle$, $i=$ $x,y,z$, are connected by the Bethe-Salpeter equation
$$\mathbf{X} = \mathbf{X}_{H}+\widetilde{U}(\mathbf{X}_{H} \mathbf{X})+4I(\mathbf{X}_{H}\cdot \mathbf{c})(\mathbf{c\cdot }\mathbf{X}).$$
The solution is given by
$$\mathbf{X} =(1-\widetilde{U} \mathbf{X}_{p})^{-1} \mathbf{X}_{p}$$
where $\mathbf{X}_{p}$ is the projected unscreened susceptibility
$$\mathbf{X}_{p}=\mathbf{X}_{H}+\frac{4I}{1-4I\chi _{H}^{cc}}(\mathbf{X}_{H}\cdot \mathbf{c})(\mathbf{c\cdot }\mathbf{X}_{H}).,\text{ \ \ \ \ \ \ \
\ }\chi _{H}^{cc}=(\mathbf{c\cdot }\mathbf{X}_{H}\cdot \mathbf{c})\text{\ \
\ }$$
To linear order in $I$ the general expression simplifies to $$\mathbf{X}=[1-\widetilde{U}\mathbf{X}_{H}-4I\,\mathbf{X}_{H}^{-1}(\mathbf{X}_{H}\cdot \mathbf{c})(\mathbf{c\cdot }\mathbf{X}_{H})]^{-1}\mathbf{X}_{H}$$In the main configuration of the experiments, the static magnetic field is oriented parallel to the $ab$-plane, say along the $a$-axis. We take this to be the $z$-axis in spin space and identify the $c$-axis with the $x$-axis. Then we see that the screening of the static field is not changed to linear order in $I,$ as $\mathbf{X}_{+z}=\mathbf{X}_{z+}=0,$ etc. The dynamical response with the time-dependent magnetic field oriented perpendicular to the $z$-axis is, however modified. Using $\chi ^{+x}\chi ^{x-}=\frac{1}{4}(\chi ^{+-})^{2}$ we find
$$\chi ^{+-}=\chi _{H}^{+-}/[1-(\widetilde{U}+I)\chi _{H}^{+-}]$$
[99]{} J. Sichelschmidt, V.A. Ivanshin, J. Ferstl, C. Geibel and F. Steglich, Phys. Rev. Lett. **91**, 156401-1 (2003).
C. Krellner, T. Forster, H. Jeevan, C. Geibel and J. Sichelschmidt, Phys. Rev. Lett. 100, 066401 (2008).
U. Schaufu[ß]{}, V. Kataev, A.A. Zvyagin, B. B[" u]{}chner, J. Sichelschmidt, J. Wykhoff, C. Krellner, C. Geibel, and F. Steglich, arXiv0804.4105v1.
P. Gegenwart, J. Custers, Y. Tokiwa, C. Geibel, and F. Steglich, Phys. Rev. Lett. **94**, 076402 (2005)
Elihu Abrahams and Peter W[" o]{}lfle, Phys. Rev, B **78**, 104423 (2008).
P. Schlottmann, Phys. Rev. B **79**, 045104 (2009).
S.-K. Yip, Phys. Rev. B **38**, 8785 (1988)
George Feher and A.F. Kip, Phys. Rev. **98**, 337 (1955).
E. Abrahams, unpublished.
The term proportional to $m/m^*$ was inadvertently misevaluated in Eq. (20) of AW.
In the familiar Fermi liquid notation, $F_0^a = -2N_0{\widetilde U}$.
J. Sichelschmidt, J. Wykhoff, H.-A. Krug von Nidda, J. Ferstl, C. Geibel, and F. Steglich J. Phys. Condens. Matter **19** 116204 (2007).
See for example, Fig. 1 of Ref. [@YRS2].
S. Paschen, T. L[" u]{}hmann, S. Wirth, P. Gegenwart, O. Trovarelli, C. Geibel, F. Steglich, P. Coleman, Q. Si, Nature **413**, 804 (2004)
P. Gegenwart, T. Westerkamp, C. Krellner, Y. Tokiwa, S. Paschen, C. Geibel, F. Steglich, E. Abrahams, Q. Si, Science **315**, 969 (2007).
O. Trovarelli, C. Geibel, S. Mederle, C. Langhammer, F.M. Grosche, P. Gegenwart, M. Lang, G. Sparn, and F. Steglich, Phys. Rev. Lett. **85**, 626 (2000).
P. Gegenwart, Y. Tokiwa, T. Westerkamp, F. Weickert, J. Custers, J. Ferstl, C. Krellner, C. Geibel, P. Kerschl, K-H. M[" u]{}ller, and F. Steglich, New J. Phys. **8**, 171 (2006).
P. Gegenwart, Y. Tokiwa, J. Custers, C. Geibel, and F. Steglich, J. Phys. Soc. Jpn. **75** (Suppl.) 155 (2006)
A.C. Hewson, *The Kondo Problem to Heavy Fermions*. (Cambridge University Press, 1993)
J. Kondo, Prog.Theor. Phys. **32**, 37 (1961).
K. G. Wilson, Rev. Mod. Phys. **47**, 773 (1975).
For references to the early experimental and theoretical papers, see W. Brenig, W. Götze and P. Wölfle, Phys. Rev. B **2**, 4533 (1970); J. Sweer, D.C. Langreth and J.W. Wilkins, Phys. Rev B **13**, 192 (1976).
U. Walter, E. Holland-Moritz and Z. Fisk, Phys Rev B **43**, 320 (1991)
P. Nozi[\` e]{}res, J. Low Temp. Phys. **17**, 31 (1974).
A.C. Hewson, J. Bauer and W. Koller, Phys. Rev. B **73**, 045117 (2006).
S.E. Barnes and J. Zitkova-Wilcox, Phys. Rev. B **7**, 2163 (1973).
O. Trovarelli et al., Physica B 284-288, 1507 (2000)
C. M. Varma, P. B. Littlewood, S. Schmitt-Rink, E. Abrahams, and A. E. Ruckenstein, Phys. Rev. Lett. 63, 1996 (1989).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Atomic clocks use atomic transitions as frequency references. The susceptibility of the atomic transition to external fields limits clock stability and introduces systematic frequency shifts. Here, we propose to realize an atomic clock that utilizes an entangled superposition of states of multiple atomic species, where the reference frequency is a sum of the individual transition frequencies. The superposition is selected such that the susceptibilities of the respective transitions, in individual species, destructively interfere leading to improved stability and reduced systematic shifts. We present and analyze two examples of such combinations. The first uses the optical quadrupole transitions in a [$^{40}$[Ca]{}$^+$]{} - [$^{174}$[Yb]{}$^+$]{} two-ion crystal. The second is a superposition of optical quadrupole transitions in one [$^{88}$[Sr]{}$^+$]{} ion and three [$^{202}$[Hg]{}$^+$]{} ions. These combinations have reduced susceptibility to external magnetic fields and blackbody radiation.'
author:
- Nitzan Akerman
- Roee Ozeri
bibliography:
- 'combinationBib.bib'
title: Atomic Combination Clocks
---
Atomic clocks are composed of three main components. An electromagnetic oscillator, an atomic reference to which the oscillator locks; i.e. a narrow atomic transition, and a frequency counter which translates the number of oscillator periods to a time unit. The atomic reference is usually carefully chosen to have as little as possible coupling to varying conditions in the environment which can lead to systematic frequency shifts. These systematic shifts can be generally divided into two categories. The first is due to the susceptibility of the atomic transition to different kinds of environmental electromagnetic fields such as the ambient magnetic field or blackbody radiation [@Rosenband2006Blackbody; @Safronova2011Blackbody; @dube2013evaluation]. The second kind of clock shifts are relativistic and emanate from the fact that the atoms are not investigated in an inertial frame. Here examples include second-order Doppler shifts or the gravitational red-shift [@chou2010optical]. Relativistic shifts can be only mitigated by maintaining the atoms in an inertial frame. The first kind however can be reduced not only by trying to minimize background electromagnetic fields, but also by reducing the differential susceptibility of the clock transition to different electromagnetic perturbations. The blackbody radiation shift, for example, can be reduced by choosing a particular atomic transition with a small differential polarizability [@Rosenband2006Blackbody]. Alternatively, it was recently proposed to suppress blackbody radiation shift by using two different atomic clocks and a frequency comb to generate a “synthetic” frequency which is largely insensitive to the presence of electric fields [@yudin2011atomic].
Reducing the susceptibility of atomic superpositions to electromagnetic perturbations has been the focus of intensive research in the context of quantum computing. Here, the motivation is to improve the coherence of multi-qubit superpositions for the purpose of metrology. One of the most successful methods for prolonging the coherence times of superpositions is the use of decoherence-free subspaces [@lidar1998decoherence]. In this approach entangled superpositions of different states have reduced susceptibility to noise, leading to multi-second coherence times [@Roos2006Designer; @Kotler2014Measurement].
We propose to use several atomic transitions in different species, each ticking with it own environmentally-sensitive frequency. These transitions are chosen such that their susceptibilities compensate each other in an entangled superposition leading to significantly reduced sensitivity to environmental fields. The idea of combining different materials with opposite susceptibilities in clocks dates back to the work of John Harrison in the eighteenth century [@usher2013history]. Harrison revolutionized the fields of timekeeping and navigation by building a marine chronometer based on a bi-metallic balance spring to compensate for temperature variations of the spring constant.
The phase of a multi-species entangled state will evolve at a frequency which is a linear sum of the different frequencies of individual transitions. It is thus possible to construct an atomic clock superposition that will evolve at a frequency of an ultra-violate transition, however, no ultra-violate laser will be involved in the clock operation. Instead, the superposition phase would be compared to the linear sum of frequencies of the clock lasers of the different transitions using an optical frequency comb. Fig.\[experimentScheme\](A) shows a schematics of our method. Inquiry the two-specie superposition phase with an optical frequency comb serves as a ”mechanical differential” in analogy with the device used in vehicles which allows two wheels to rotate at different rotation frequencies while keeping the average rotation frequency fixed.
Maximally entangled multi-atom states can, in principle, be used to surpass the frequency estimation accuracy attainable with uncorrelated atoms (standard quantum limit) and reach the Heisenberg limit [@Wineland1994Squeezed]. However, as pointed by Huelga et.al [@Huelga1997Improvement], in practice, clock stability is ultimately limited by the finite coherence time of the system. For uncorrelated Markovian noise the two cases will give the same result. Nevertheless, in the some cases of correlated noise, one can sill gain by cleverly engineering the proper entangled state [@Roos2006Designer; @Kessler2014heisenberg; @Louchet2010entanglement].
The Combination clock
=====================
![(A) schematic illustration of a two-species combination clock. The clock operates using the two clock lasers of the individual transitions. An optical frequency comb is used to compare the phase of the frequency sum of the two lasers to that of an entangled superposition. (B) The quantum circuit of gates for preparation of the clock entangled state and the final parity check allowing for clock interrogation. []{data-label="experimentScheme"}](Figure1v2.pdf){width="8.5"}
We begin by describing our general scheme and discuss in detail the quantum protocol that can be used in the case of a two-ion combination clock . Our scheme is a generalization of the Ramsey spectroscopy method similar to the case of identical ions in a maximally entangled state [@Bollinger1996Optimal; @Leibfried2004Toward].
An $N$-ion, multi-species, atomic register is prepared in a superposition of two states $\Psi=\frac{1}{\sqrt{2}}(\ket{G}+e^{i\Omega t}\ket{E})$. Where $\ket{G}=\prod_{i}^{N}\ket{f^{k,m}_i}$ and $\ket{E}=\prod_{i}^{N}\ket{s^{k,m'}_i}$ are tensor products of states in the ground and excited state manifolds of the different ions. Here $s$ and $f$ can be either the ground or excited states of the clock transition in ion $i$ of species $k$, and $(m,m')$ are the quantum numbers of the exact state in the ground or excited states manifolds of the transition (e.g. Zeeman or hyperfine state). The clock frequency is $\Omega=\sum_{{i}}^{N}\omega^{k,m,m'}_i$, where $\hbar\omega^{k,m,m'}_i = E(\ket{s^{k,m}_i}) - E(\ket{f^{k,m'}_i})$ and $E$ is the energy of the state. Notice that with our definition $\omega^{k,m}_i$ can be negative if ion $i$ is prepared such that $\ket{f^{k,m}_i}$ is the excited state and $\ket{s^{k,m}_i}$ is the ground state of the clock transition. The total frequency is thus a linear sum of the clock transition frequencies of all the ions in the crystal.
After an interrogation time $t=\tau$ the phase between the atomic superposition and corresponding linear sum of local oscillators (i.e. clock lasers), $\delta\tau$, is obtained in the following way. A $\pi/2$ carrier pulse is applied to all ions in the crystal with their corresponding clock lasers followed by state detection and parity analysis. The probability of measuring even parity (i.e. an even number of ions in the excited state) will be given by $P_{even} = \cos{\delta\tau}$. A following change to the frequency of one of the clock lasers can lock the linear sum of frequencies to the entangled superposition.
Comparing with the case of a single clock-ion species, the use of multi ion-species increases experimental complexity due to the need of having a set of all the lasers, including narrow-linewidth clock lasers, that are necessary for each of the ion-species. On the other hand, some of the technical aspects such as single-ion addressing and detection are simplified owing to the spectral distinguishably between species. Comparing with logic spectroscopy schemes [@Schmidt2005Spectroscopy; @Chou2010Frequency], the combination clock does not involve motion-sensitive operations after the initial preparation stage and therefore continuous cooling or very-low heating rates are not a prerequisite for implementing this scheme. The experimental sequence for the specific example of a two species two-ion crystal is depicted in Fig.\[experimentScheme\](B). The entanglement of two different ion species was recently demonstrated [@Tan2015multi; @Ballance2015hybrid]. As a simple example, we begin with the two ions in the ground state of their internal state $\ket{g^A},\ket{g^B}$ as well as in the ground state of one of their normal modes of motion $\ket{0}$. After applying a blue-sideband (BSB) $\pi/2$-pulse to ion A, we entangled the internal state of ion A with the motion of the two-ion crystal. Next we apply a red-sideband (RSB) $\pi$-pulse on ion B which entangles the internal states of the ions while disentangling the internal states from motion. In the rotating-frame with respect to the sum of atomic transition frequencies $\omega_A+\omega_{B}$, the resulting state is,
$$\begin{aligned}
\ket{g^Ag^B}&\ket{0} \overset{BSB}{\underset{\pi/2}{\Longrightarrow}}
\frac{1}{\sqrt{2}}\left(\ket{g^Ag^B}\ket{0}+\ket{e^Ag^B}\ket{1}\right)\\
&\overset{RSB}{\underset{\pi}{\Longrightarrow}}\frac{1}{\sqrt{2}} \left( \ket{g^Ag^B}+e^{i\phi_{tot}}\ket{e^Ae^B} \right)\ket{0}\nonumber\end{aligned}$$
Here $\phi_{tot}= \phi_{0}+(\delta_A+\delta_B)t_{int}$ with $\delta_{A,B}$ as the detunings of the two lasers from their corresponding atomic transitions and $\phi_{0}$ is a constant phase for zero interrogation time resulting from finite pulse times, light shifts, etc. The two-ion entangled state acquires a phase which is the sum of the two transition frequencies times the interrogation time $\tau$. This phase difference with respect to the sum of both clock laser frequencies is measured by a $\pi/2$ carrier pulse on each ion with its own clock laser, followed by state detection and parity analysis.
The combined two-ion clock is equivalent to a single atom with $\omega_0=\omega_A+\omega_B$ and with measured error signal $\delta=\delta_A+\delta_B$. In the case of two uncorrelated lasers it is enough to feedback on only one of the lasers while making sure that each of the lasers is sufficiently close to resonance such that it can execute the necessary pulses for the protocol; i.e. to within their Rabi frequency. One way to experimentally retrieve the stable clock frequency is by sum frequency generation using a nonlinear crystal. A more practical way is to work with correlated lasers which may be derived from a frequency comb. In this case we can write the two lasers frequencies as $\omega_A=N\cdot\omega_R$ and $\omega_B=M\cdot\omega_R$ where $N$ and $M$ are integers and $\omega_R$ is the comb repetition rate. Here the comb offset frequency is assumed to be set to zero. The error signal from the experimental interrogation $\delta_0=\delta_R(n+m)$ is then fed-back to stabilize $\omega_R$ to the value that locks $\delta= 0$. The extension of the locking of the comb to the linear sum in the case of more than two species is straightforward. .
Below we analyze two examples where the combination of two optical transitions with differential scalar polarizabilities of opposite sign such that the BBR shift is considerably reduced. This is one of the more difficult systematic shifts to mitigate because, for a given polarizability, it requires the evaluation of the BBR fields with high accuracy. As shown below, other differential shifts due to electromagnetic moments can be either reduced (e.g. first order Zeeman) or enhanced (e.g. electric quadrupole shifts). It should be mentioned that this method is not limited to the combination of optical transitions only. It is also possible to combine optical clock transitions with a microwave hyperfine transition in an entangled superposition. Here the resulting effective clock frequency is very similar to that of the optical component, however the magnetic susceptibility of the combination superposition can be controlled, especially when working at a high magnetic field in the Paschen-Back regime. In this regime the effective g-factor of the hyperfine transition can be smoothly tuned to null the superposition first order magnetic susceptibility.
We note that the use of entangled states is not always necessary. In many cases one can start with non-entangled states which have a significant overlap with the desired entangled state. If the part orthogonal to the desired clock superposition dephases quickly it will yield no parity signal. In this case, the measurement of parity will yield the same information as before, however at the cost of lower contrast and thus lower signal-to-noise ratio [@chwalla2007precision]. The advantage here would be the lower technical complexity as an entangling operation will not be necessary.
[$^{40}$[C]{}$^+$]{} - [$^{174}$[Y]{}$^+$]{} combination clock
==============================================================
The first combination clock we examine is composed of the electric quadrupole transition $^4S_{1/2} \rightarrow$$ ^3D_{3/2}$ in [$^{40}$[Ca]{}$^+$]{} and the same electric quadrupole transition; $^6S_{1/2} \rightarrow$$ ^5D_{3/2}$; in [$^{174}$[Yb]{}$^+$]{}. This seems like a particularly nice combination since the two transitions are driven using fairly accessible wavelengths of $732\ nm$ and $436\ nm$ correspondingly; while the effective frequency at which this clock superposition rotates corresponds to an effective wavelength of 273 nm (1090 THz) and lifetime of $\approx 50$ ms limited by the [$^{}$[Yb]{}$^+$]{} transition. This effective wavelength is very close to that of the [$^{27}$[Al]{}$^+$]{} optical clock (267 nm) with the advantage that no UV light has to be used in order to drive this transition. Furthermore the magnetic moments of these two transitions are the same to within a fractional difference of $10^{-3}$, so superpositions with states containing an equal number of states with opposite projections along the magnetic field direction can have almost zero magnetic susceptibility. But most importantly, the differential static polarizability of these two transitions are very similar in magnitude but have opposite signs [@Ludlow2015Optical] and therefore a superposition of the two species can have a significantly smaller susceptibility to blackbody radiation. Up to normalization, the specific superposition we consider here is, $${\ket{\text{Ca};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}+
{\ket{\text{Ca};\text{D}_{{{\frac{3}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{D}_{{{\frac{3}{2}}},{-{\frac{1}{2}}}}}}.$$ Below is an analysis of the different systematic shifts for this combination clock.
First-order Zeeman
------------------
The first-order Zeeman shift due to a uniform magnetic field for this superposition is almost zero as both superposition parts have a total zero magnetic moment, up to a small ($\approx10^{-3}$) difference between the bound electron $g$-factor in the two cases. Magnetic-field gradients, however, will still contribute a first order sensitivity to magnetic field. Here the magnetic field gradient can be measured, and subsequently nulled, by preparing the superposition, $${\ket{\text{Ca};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}+ {\ket{\text{Ca};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}.$$ A similar superposition has been previously used to estimate and compensate magnetic field gradients to the 20 mHz level [@Kotler2014Measurement] which translates into a fractional uncertainty of $2\times 10^{-17}$. One can further improve on this uncertainty by interlacing experiments in which the spin direction of both ions is reversed, thus time-averaging out the gradient effect in the same manner that first order Zeeman shift is commonly averaged over [@Bernard1998laser].
Second order Zeeman
-------------------
Other optical clocks such as [$^{171}$[Yb]{}$^+$]{} benefit from the possibility of a clock transition which is first order insensitive to the magnetic field. In most of these cases this happens due to the hyperfine interaction between the electron and the nucleus. This first-order insensitivity then comes at the expense of a relatively large second-order Zeeman shift (typically $10^3$ or more larger than in the first-order sensitive case). Here, since the first order insensitivity results from careful engineering of the clock superposition, the clock superposition amplitudes do not depend on the magnetic field value. A change in the magnetic field will not result in a change of the magnetic moment of the superposition states and hence no second order Zeeman shift. The small second order Zeeman shift will be, as in first order sensitive clocks, due to the fine-structure coupling in each of the transitions. The order of the second order Zeeman will be, $$\Delta f_{M2} \approx \frac{(g\mu_B mB)^2}{h\Delta E_{FS}},
\label{secondZeemanEq}$$ where $g$ is the electron dimensionless magnetic moment, $\mu_B$ is Bohr magneton, $B$ is the magnetic field, $h$ is Plank’s constant and $\Delta E_{FS}$ is the fine-structure (FS) splitting in the D manifold. The FS splitting in [$^{}$[Ca]{}$^+$]{} is $\Delta E_{FS}/h = 1.82$ THz and in [$^{}$[Yb]{}$^+$]{} it is $\Delta E_{FS}/h =$ 42 THz. This means that the second order Zeeman will be dominated by [$^{}$[Ca]{}$^+$]{}. Working at a field of $10^{-4}$ T we get, $$\Delta f_{M2} \approx \frac{(1.4\times 10^6)^2}{1.82\times 10^{12}} \approx 2\ Hz,$$ One can measure the average magnetic field by preparing the superposition, $${\ket{\text{Ca};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}+
{\ket{\text{Ca};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}$$ Here it is likely that the magnetic field will be measured with at least $10^{-3}$ accuracy, allowing for a clock frequency error as low as 4 mHz which translates into a fractional uncertainty below $2\times 10^{-18}$.
Electric Quadrupole shift
-------------------------
The quadrupole shift (QS) of this clock superposition will be due to the quadrupole moment of the $D$ level of both ions since the $S$ orbital has no electric quadrupole moment. The quadrupole shift of $\left|D_{\frac{3}{2}},m_{j}\right\rangle$ level is given by [@Itano2006Quadrupole], $$\Delta f_{Q}=\frac{1}{4h}\frac{\partial{E_{z}}}{\partial{z}}\Theta\left(D,\frac{3}{2}\right)\frac{5-4m_{j}^{2}}{4}\left(3\mbox{cos}^{2}\left(\beta\right)-1\right).$$ Here, $\Theta\left(D,\frac{3}{2}\right)$ is the $D_{\frac{3}{2}}$ level quadrupole moment and $\beta$ is the angle between the quantization axis set by the magnetic field and the quadrupole axis set by the trap geometry, where we assume full radial (cylindrical) symmetry. In the case of two ions in the trap, the equilibrium positions of the ions are $\pm z_0$, symmetrically positioned around the potential minimum along the trap axial direction. This results in an electric field gradient that is three times larger than the case of a single ion. Using $\frac{\partial{E_{z}}}{\partial{z}}= 1000 V/mm^2$ and taking the case where $\beta=0$ as a worst case scenario, the QS for the $|D, m=\pm \frac{1}{2}\rangle$ will be $$\Delta f_{Q}=\frac{ea_0^2}{2h}\frac{1000}{10^{-6}}( 2.08 + 1.34) = 1156 \ Hz.$$ This result is rather large due to the constructive addition of the entangled superposition and the position of the ions in the trap. In order to calibrate the QS one can place the ions in the superposition, $${\ket{\text{Ca};\text{D}_{{{\frac{3}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Yb};\text{D}_{{{\frac{3}{2}}},{-{\frac{1}{2}}}}}}+
{\ket{\text{Ca};\text{D}_{{{\frac{3}{2}}},{{\frac{3}{2}}}}}}{\ket{\text{Yb};\text{D}_{{{\frac{3}{2}}},{-{\frac{3}{2}}}}}}$$ Which, in the absence of magnetic moment evolves solely due to the QS [@Roos2006Designer]. In this case the QS will be factor of two larger as compared with that of the clock superposition above. Here a value of $\beta$ can be chosen to minimize (null) the QS. Moreover the QS can be eliminated by averaging over different magnetic sub-levels [@dube2013evaluation]. For the case of the D$_{3/2}$ level, only the m$_j=1/2$ and m$_j=3/2$ states are required.
Blackbody radiation shift
-------------------------
The blackbody shift is due to the scalar differential polarizability of the clock transition; $\Delta \alpha _s$. The blackbody radiation creates a varying electric field at the ion position $E_{rms}$. The shift of the clock transition is given by, $$\hbar\Delta\Omega = \frac{1}{2}E_{rms}^2\Delta \alpha _s(1 + \eta),$$ Here $\Delta \alpha _s$ is the differential static polarizability between the two clock transition states. The dynamic correction factor $\eta$ takes into account the effect of BBR fields not being static but rather peaked at around $10\ \mu m$ (30 THz) at 300 K. This dynamic correction factor is typically small ($\approx 10^{-1}-10^{-2}$) and we will therefore neglect it in the following. The effective differential static polarizability of the superposition above (for the purpose of frequency shift calculation) will therefore be the sum of the two polarizabilities of the two transitions involved.
The differential polarizability for the above $4S_{1/2} \rightarrow 3D_{3/2}$ transition in [$^{}$[Ca]{}$^+$]{} is $\Delta\alpha_s = -44(1)\ a_o^3$ \[$-73(2)\ \frac{Jm^2}{V^2}$\] [@Arora2007Blackbody-radiation; @Safronova2011Blackbody]. This value is similar to the [$^{}$[Ca]{}$^+$]{} $4S_{1/2}\rightarrow3D_{5/2}$ clock transition which is investigated as an optical clock in several labs [@Chwalla2009Absolute; @Wolf2011Direct; @Matsubara2012Direct; @Huang2016Frequency]. The corresponding differential polarizability on the same clock transition in [$^{}$[Yb]{}$^+$]{} is $\Delta\alpha_s = 42(9)\ a_o^3$ \[$69(14)\ \frac{Jm^2}{V^2}$\] [@Schneider2005Sub-Hertz]. The differential polarizability of the [$^{}$[Ca]{}$^+$]{}-[$^{}$[Yb]{}$^+$]{} combination clock is $\Delta\alpha_s = -2 (9)\ a_o^3$ \[$-4(14)\ \frac{Jm^2}{V^2}$\]. This is a significant reduction in the expected BBR shift of at least a factor of $5$ if not more. The expected shift to the clock transition at $T = 300\ K$ is $20 (70)\ mHz$ corresponding to a fractional uncertainty of $6\times 10^{-17}$ dominated by the uncertainty in the differential polarizability. With better measurement of the this polarizability and a $10\ \%$ accuracy in estimating the BBR field power, the BBR uncertainty of this clock will be in the $10^{-18}$ range.
Second order Stark shift due to trap fields
-------------------------------------------
Trap fields will induce frequency changes due to the differential polarizability of the transition. Trap fields are mostly due to uncompensated micromotion or inherent micromotion that results from ion thermal motion in the pseudo-potential. Here, since the trap fields have a well-defined direction both the scalar and the tensor part of the differential polarizability will play a role.
To the extent that the trap rf fields are identical on both ions then the scalar polarizability of the transition is suppressed by the same factor that is mentioned in the BBR section. A difference in the rf amplitude the two ions experience is likely. Firstly, any stray field orthogonal to the trap axis will tilt the two ion crystal and displace the ions into different rf regions. This is because the radial pseudo-potential depends on mass and therefore the two ions experience different radial restoring forces. Secondly, any inhomogeneous axial rf fields will lead to a difference too. However, with the minimization of micromotion this residual shift should be at least as small as in the case of each of the two clocks when operated on a single ion.
The tensor polarizability of [$^{}$[Ca]{}$^+$]{} is $-17.43(23)\times 10^{-41}\ Jm^2/V^2$ [@Safronova2011Blackbody]. The tensor polarizability of $Yb^+$ is $-13.6(2.2)\times 10^{-41}\ Jm^2/V^2$ [@Ludlow2015Optical]. Therefore, the combined tensor polarizability of this clock transition is $-31.0(2.2)\times 10^{-41}\ Jm^2/V^2$. The amplitude of rf fields that the ions experience can be minimized to the $10\ V/m$ level. The order of magnitude of tensor shifts expected is therefore in the few $10's\ \mu Hz$ corresponding to fractional uncertainty in the $10^{-19}$ level. Furthermore, the tensor shift depend on the $m$ level and magnetic field direction, here the projection of the electric field, rather than the electric field gradient along the magnetic field direction is used. This means that the same average of measurements along different magnetic field directions or different $m$ levels that nulls the quadrupole shift will null the tensor polarizability contribution as well.
[$^{88}$[S]{}$^+$]{} - 3$\times$[$^{202}$[H]{}$^+$]{} combination clock
=======================================================================
The second combination we look into here is motivated primarily by the desire to reduce BBR shifts as much as possible by reducing the combination differential scalar polarizability. The differential polarizability $\Delta\alpha_s=-47.938(71)\times 10^{-41}\ Jm^2/V^2$ [@Ludlow2015Optical] of the $S_{1/2}\rightarrow D_{5/2}$ electric quadrupole transition at $674\ nm$ in [$^{88}$[Sr]{}$^+$]{} is about 3 times larger and opposite in sign than the differential polarizability $\Delta\alpha_s=15\times 10^{-41}\ Jm^2/V^2$ [@Ludlow2015Optical] of the same transition in [$^{202}$[Hg]{}$^+$]{} at $282\ nm$. Thus, a superposition of a single [$^{88}$[Sr]{}$^+$]{} and three [$^{202}$[Hg]{}$^+$]{} ions would significantly reduce $\Delta\alpha_s$ to $\approx-3\times 10^{-41}\ Jm^2/V^2$ . $$\begin{aligned}
&{\ket{\text{Sr};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Hg};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}{\ket{\text{Hg};\text{S}_{{{\frac{1}{2}}},{-{\frac{1}{2}}}}}}{\ket{\text{Hg};\text{S}_{{{\frac{1}{2}}},{{\frac{1}{2}}}}}}+\\
&{\ket{\text{Sr};\text{D}_{{{\frac{5}{2}}},{{\frac{1}{2}}}}}}{\ket{\text{Hg};\text{D}_{{{\frac{5}{2}}},{-{\frac{1}{2}}}}}}{\ket{\text{Hg};\text{D}_{{{\frac{5}{2}}},{-{\frac{1}{2}}}}}}{\ket{\text{Hg};\text{D}_{{{\frac{5}{2}}},{\nonumber}}}}\end{aligned}$$ The effective frequency at which this superposition rotates is $82\ nm$ ($3634\ THz$) - deep in the extreme UV. The expected shift to the clock transition at T=300 K is below 20 mHz corresponding to a fractional uncertainty of $5\times10^{-18}$. With better measurement of the differential polarizability and 10% accuracy in estimating the BBR field power, the BBR uncertainty of this clock will be in the $10^{-19}$ range.
The rest of the systematic shift analysis here is rather similar to the one presented above for the [$^{40}$[Ca]{}$^+$]{} - [$^{174}$[Yb]{}$^+$]{} combination clock. In short: The first-order Zeeman shift due to a uniform magnetic field for this superposition is zero as both superposition parts have a total zero magnetic moment. Moreover, with this particular spatial ordering of spin states, there would not be first order sensitivity to magnetic field gradients as well and only second order spatial magnetic field variation would cause a shift. The FS splitting in [$^{}$[Hg]{}$^+$]{} is $\Delta E_{FS}/h = 451$ THz and in [$^{}$[Sr]{}$^+$]{} it is $\Delta E_{FS}/h = 8.4$ THz and therefore second order Zeeman shifts will be dominated largely by [$^{}$[Sr]{}$^+$]{} and can therefore be estimated using Eq. \[secondZeemanEq\] to be $\Delta f_{M2} \approx0.2$ Hz. Here as well, an entangled superposition can be used to measure the average magnetic field on the clock array. Evaluating it at the $10^{-3}$ level, leads to a systematic shift of the order of $2\times 10^{-19}$. The quadrupole shift in this case will be rather large due to the large crystal and is estimate to be around 5 kHz for the same conditions considered above.
In conclusion, we propose and analyze the use of entangled states composed of multi ion-species for realizing an optical clock with reduced systematic shifts and improved stability. As concrete examples we proposed two different combinations. A [$^{40}$[C]{}$^+$]{} - [$^{174}$[Y]{}$^+$]{} two ions clock and a [$^{88}$[S]{}$^+$]{} - 3$\times$[$^{202}$[H]{}$^+$]{} four-ion crystal. Both these examples are expected to operate with a significantly reduced BBR shift. Our approach opens a new dimension to choosing suitable atomic clock references. Instead of merely scanning the periodic table for a suitable transition in different atomic species, we propose to search for the suitable linear combination of transitions, in different species which would have the lowest susceptibilities to environmental conditions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With the rich dynamics studies of single-state processes, the two-state processes attract more and more interests of people, since they are widely observed in complex system and have effective applications in diverse fields, say, foraging behavior of animals. This report builds the theoretical foundation of the process with two states: Lévy walk and Brownian motion, having been proved to be an efficient intermittent search process. The sojourn time distributions in two states are both assumed to be heavy-tailed with exponents $\alpha_\pm\in(0,2)$. The dynamical behaviors of this two-state process are obtained through analyzing the ensemble-averaged and time-averaged mean squared displacements (MSDs) in weak and strong aging cases. It is discovered that the magnitude relationship of $\alpha_\pm$ decides the fraction of two states for long times, playing a crucial role in these MSDs. According to the generic expressions of MSDs, some inherent characteristics of the two-state process are detected. The effects of the fraction on these observables are detailedly presented in six different cases. The key of getting these results is to calculate the velocity correlation function of the two-state process, the techniques of which can be generalized to other multi-state processes.'
author:
- Xudong Wang
- Yao Chen
- Weihua Deng
bibliography:
- 'Reference.bib'
title: 'Aging two-state process with Lévy walk and Brownian motion'
---
Searching a target is a natural demand in the real world. At the same time, many physical or biological problems can be regarded as the search processes, describing how a searcher finds a target located in an unknown position. At the macroscopic scale, it is exemplified as animals searching for food or a shelter [@Bell:1991]. At the microscopic scale, one can cite the localization by a protein of a specific DNA sequence or the active transport of vesicles in cells [@BenichouLoverdoMoreauVoituriez:2011]. In these examples, the search time is generally a limiting quantity which has to be optimized by choosing different search strategies. Intermittent search strategies have been proved to play a crucial role in optimizing the search time of randomly hidden targets [@BenichouCoppeyMoreauSuetVoituriez:2005; @LomholtKorenMetzlerKlafter:2008]. This kind of search behavior could be extended to broader research domains such as the theory of stochastic processes [@BartumeusCatalanFulcoLyraViswanathan:2002], applied mathematics [@Stone:1975], and molecular biology [@CoppeyBenichouVoituriezMoreau:2004]; and it also motivates some new interesting research topics [@XuDeng:2018; @*XuDeng:2018-2].
For the intermittent search process, it switches between two phases — local Brownian search phase and ballistic relocation phase (Lévy walk). The searcher displays a slow reactive motion in the first phase, during which the target can be detected. The latter fast phase aims at relocating into unvisited regions to reduce oversampling, during which the searcher is unable to detect the target. In the situation of rare targets, it has been shown that the search process with Lévy distributed relocations significantly outperforms that with exponentially distributed relocation [@LomholtKorenMetzlerKlafter:2008]. While the two-state process effectively models intermittent strategy, it is also observed in the transport of the neuronal messenger ribonucleoproteins delivered to their target synapses [@SongMoonJeonPark:2018], where a type of Lévy walk process is interrupted by the emerging of rest. The rest period can be very long, characterized by power-law distribution without finite mean. This phenomenon becomes a striking feature of the RNA transport in neuronal systems.
The intermittent strategy has been verified to be optimum for searching targets in some specific macroscopic and microscopic situations. But generally it is hard to believe that the intermittent strategy is always the best one in all the foraging behaviors of animals and the intracellular transport in microscopic scale. A question naturally comes up: How about the field of its application? Based on this motivation, it is necessary to build a complete theoretical foundation for this kind of two-state processes for dealing with data observed in experiments. In this report, we consider the two-state process mentioned earlier (i.e., the standard Lévy walk and Brownian motion) and mainly investigate their statistical behaviors, such as ensemble-averaged mean square displacement (EAMSD) and time-averaged mean square displacement (TAMSD). In particular, we carefully examine the aging behaviors of the two-state process, while the aging continuous-time random walk (CTRW) [@BarkaiCheng:2003], aging renewal theory [@SchulzBarkaiMetzler:2014] and aging ballistic Lévy walks [@MagdziarzZorawik:2017] have been fully discussed. Since the observation time might not be the beginning of a process in experiments, aging behavior should be paid some attention and it may display interesting phenomena in anomalous diffusion processes [@SokolovBlumenKlafter:2001; @Allegrini--etal:2002].
Lévy walk dynamics describe enhanced transport phenomena in many systems. Within the CTRW framework, originally introduced by Montroll and Weiss [@MontrollWeiss:1965], the significant feature of Lévy walk is the underlying spatiotemporal coupling, which penalizes long jumps and leads to a finite EAMSD [@ZaburdaevDenisovKlafter:2015]. While the uncoupled process, Lévy flight [@ShlesingerZaslavskyFrisch:1995; @MetzlerKlafter:2000], has divergent EAMSD. The diffusion behavior of Lévy walk depends on the exponent $\alpha$ of the power-law distributed running time. It displays ballistic diffusion for $\alpha<1$ and sub-ballistic superdiffusion for $1<\alpha<2$. We assume the particle switches between Lévy walk phase and Brownian phase, denoted as states ‘$+$’ and ‘$-$’, respectively. The velocities of the two-state process are, respectively, $v_+(t)$ for Lévy walk and $v_-(t)$ for Brownian motion. The PDF of $v_+(t)$ is $\delta(|v|-v_0)/2$, while $v_-(t)=\sqrt{2D}\xi(t)$ with $\xi(t)$ being a Gaussian white noise satisfying $\langle\xi(t)\rangle=0$ and $\langle\xi(t_1)\xi(t_2)\rangle=\delta(t_1-t_2)$. By taking the diffusivity $D=0$, the Brownian phase becomes a trap event and we immediately obtain the process – Lévy walk interrupted by rest.
Let the sojourn times $t$ in the two states ‘$\pm$’ be random variables obeying power-law distribution: $$\psi_\pm(t)\simeq \frac{a_\pm}{|\Gamma(-\alpha_\pm)|t^{1+\alpha_\pm}}$$ for large $t$, where $a_\pm$ are scale factors and $\Gamma(\cdot)$ is the Gamma function. We assume that the exponents $\alpha_\pm\in(0,2)$ in two states and the sojourn times in two sates are mutually independent. As usual, we apply the approach of Laplace transform $\hat{\psi}_\pm(s):=\int_0^\infty dt e^{-st}\psi_\pm(t)$ and obtain the asymptotic behavior of the sojourn time distribution for small $s$: $$\begin{aligned}
&\hat{\psi}_\pm(s) \simeq 1-a_\pm s^{\alpha_\pm} ~~\qquad\qquad {\rm for}~~ \alpha_\pm\in(0,1), \\
&\hat{\psi}_\pm(s) \simeq 1-\mu_\pm s + a_\pm s^{\alpha_\pm} ~~~~~ {\rm for}~~ \alpha_\pm\in(1,2),\end{aligned}$$ where $\mu_\pm$ is the mean sojourn time in state ‘$\pm$’, being finite when $\alpha_\pm\in(1,2)$. The survival probability that the sojourn time in state ‘$\pm$’ exceeds $t$ is defined as $\Psi_\pm(t)=\int_t^\infty dt' \psi_\pm(t')$ with Laplace transform $\hat{\Psi}_\pm(s)=[1-\hat{\psi}_\pm(s)]/s$. Note that the dynamic behaviors of standard Lévy walk are significantly different for exponent less or larger than $1$ [@ZaburdaevDenisovKlafter:2015]. We will fully discuss the EAMSD and TAMSD of the two-state process for different sets of $\alpha_\pm$ in the following. Although the mean sojourn time is finite (i.e., $\alpha_\pm>1$) in most cases, such as the intermittent search process, there are still some circumstances presenting scale free dynamics with $\alpha_\pm<1$, for example, the RNA transport in neuronal systems. Here we make uniform discussions with $\alpha_\pm\in(0,2)$ for comprehensive understanding of the two-state process.
*Propagator and occupation fraction of two states*. Suppose that the particles are initialized at the origin. The propagator $p(x,t)$ represents the PDF of finding the particle at position $x$ at time $t$. For the two-state process, it is natural to concern which state the particles are located in at time $t$. Here we denote the joint PDF of finding the particle at position $x$ and state ‘$\pm$’ at time $t$ as $p_\pm(x,t)$, which is associated with the propagator by the relation $p(x,t)=p_+(x, t)+p_-(x,t)$. The subscript ‘$\pm$’ will imply an identical meaning for other quantities.
The integral equations for $p_\pm(x,t)$ can be similarly obtained as the master equations for CTRWs. Besides the sojourn time distribution $\psi_\pm(t)$ and survival probability $\Psi_\pm(t)$, we introduce the notation $G_\pm(x,t)$ to represent the conditional probability that a particle makes a displacement $x$ during sojourn time $t$ at one step in state ‘$\pm$’. Their expressions are given by $$\begin{aligned}
&G_+(x,t)=\delta(|x|-v_0t)/2, \\
&G_-(x,t)=\frac{1}{\sqrt{4\pi Dt}}\exp\left(-\frac{x^2}{4Dt}\right),\end{aligned}$$ since the state ‘$+$’ represents Lévy walk and state ‘$-$’ denotes Brownian motion, respectively. Then the transport equation governing flux of particles $\gamma_\pm(x,t)$, which defines how many particles leave the position $x$ and change from state ‘$\mp$’ to state ‘$\pm$’ per unit time, satisfies, $$\begin{split}
\gamma_\pm(x,t)&= \int_0^tdt'\int_{-\infty}^{\infty}dx'\psi_\mp(t')G_\mp(x',t')\gamma_\mp(x-x',t-t') \\
&~~~+p^0_\mp\psi_\mp(t)G_\mp(x,t),
\end{split}$$ where the constant $p^0_\pm$ is the initial fraction of two states, that is $p_\pm(x,t=0)=p^0_\pm\delta(x)$. The current density $p_\pm(x,t)$ of particles is connected to the flux $\gamma_\pm(x,t)$ $$\begin{split}
p_\pm(x,t)&=\int_0^t dt'\int_{-\infty}^\infty dx' \Psi_\pm(t')G_\pm(x',t')\gamma_\pm(x-x',t-t') \\
&~~~+p^0_\pm \Psi_\pm(t)G_\pm(x,t).
\end{split}$$ By means of the techniques of Laplace and Fourier transform, $\tilde{\hat{p}}_\pm(k,s)=\int_0^\infty dt\int_{-\infty}^\infty dx e^{-st}e^{ikx}p_\pm(x,t)$ can be obtained (see Supplemental Material). Besides, the occupation fraction of two states $p_\pm(t)$, as the marginal density of finding the particles in state ‘$\pm$’ at time $t$, can be obtained by taking $k = 0$ in $\tilde{\hat{p}}_\pm(k,s)$. The expression of $p_\pm(t)$ in Laplace space $(t\rightarrow s)$ is $$\label{ps}
\hat{p}_{\pm}(s)=\frac{p_{\pm}^0+p_{\mp}^0\hat{\psi}_{\mp}(s)}{1-\hat{\psi}_+(s)\hat{\psi}_-(s)}
\cdot\frac{1-\hat{\psi}_{\pm}(s)}{s},$$ the normalization of which can be confirmed by verifying $\hat{p}_+(s)+\hat{p}_-(s)=1/s$.
*EAMSD and TAMSD*. If one is eager for more information of a process, such as the TAMSD, the propagator $p(x,t)$ at a single point is not enough. Instead, the two-point velocity correlation function $\langle v(t_1)v(t_2)\rangle$ plays a crucial role. We will calculate it firstly and then show the generic results of EAMSD and TAMSD for the aging process $x_{t_a}(t)$. The age $t_a$ means that this process has evolved for a time period $t_a$ before we start to observe it, and $t$ is the measurement time.
Since the model we considered contains two states: Lévy walk and Brownian motion, represented by symbols ‘$+$’ and ‘$-$’, respectively. The velocity correlation function could be written as a sum of four possible cases in terms of different states: $$\label{vv}
\begin{split}
\langle v(t_1)v(t_2) \rangle&=\langle v_+(t_1)v_+(t_2) \rangle+\langle v_-(t_1)v_-(t_2) \rangle\\
&~~~+\langle v_+(t_1)v_-(t_2) \rangle+\langle v_-(t_1)v_+(t_2) \rangle.
\end{split}$$ The first term on the right-hand side represents the case that the velocity process $v(t)$ are in Lévy walk phase at both time points $t_1$ and $t_2$; other terms stand for similar parts of the correlation function. For the first term, the velocity is correlated only when there is no renewal happens between $t_1$ and $t_2$. Thus, we have $$\label{vv1}
\langle v_+(t_1)v_+(t_2) \rangle = v_0^2 p_+(t_1) p_{+,0}(t_1,t_2),$$ where $p_+(t)$ has been given in and $p_{+,0}(t_1,t_2)$ is the PDF that no renewal happens between times $t_1$ and $t_2$ in state ‘$+$’. Similarly, the second term on the right hand side of is $$\label{vv2}
\begin{split}
\langle v_-(t_1)v_-(t_2) \rangle = 2D\delta(t_1-t_2) p_-(t_1) p_{-,0}(t_1,t_2),
\end{split}$$ where $p_{-,0}(t_1,t_2)=1$ for $t_1=t_2$, since there must be no renewals within a zero time lag. The two states at times $t_1$ and $t_2$ are different in the last two terms on . Therefore, the velocity at $t_1$ and $t_2$ are independent. Considering the velocity is unbiased at any time, the last two terms are void.
Note that the PDFs $p_\pm(t)$ and $p_{+,0}(t_1,t_2)$ should be calculated firstly to obtain the velocity correlation function in . The former one has been given in , while the double Laplace transform $(t\rightarrow s,\tau\rightarrow u)$ of the latter PDF $f_{+,0}(t,\tau)=p_{+,0}(t,t+\tau)$ is [@GodrecheLuck:2001] $$\label{ps1s20}
\hat{f}_{+,0}(s,u)= \frac{s-u+u\hat{\psi}_{+}(s)-s\hat{\psi}_{+}(u)}{s(s-u)(1-\hat{\psi}_{+}(s))u}.$$ It seems not easy to perform the inverse Laplace transform on $\hat{f}_{+,0}(s,u)$. Instead, we can obtain the expression of $p_{+,0}(t_1, t_2)$ in Laplace space ($t_1\rightarrow s_1, t_2\rightarrow s_2$) by substituting variables (see Supplemental Material): $$\label{ps1s2}
\begin{split}
\hat{p}_{+,0}(s_1,s_2)= \frac{1+\hat{\psi}_+(s_1+s_2)-\hat{\psi}_+(s_1)-\hat{\psi}_+(s_2)}{s_1s_2(1-\hat{\psi}_+(s_1+s_2))}.
\end{split}$$ Taking inverse Laplace transform on becomes doable. Based on and , the velocity correlation function $\langle v(t)v(t+\tau)\rangle$ in can be obtained for different sojourn time distributions $\psi_\pm(t)$. Noticing the asymptotic forms of $p_\pm(t)$ and $p_{+,0}(t,t+\tau)$ for large $t$, the velocity correlation function can be rewritten in the scaling form as $$\label{vtvttau}
\begin{split}
\langle v(t)v(t+\tau)\rangle &=\langle v_+(t)v_+(t+\tau)\rangle+\langle v_-(t)v_-(t+\tau)\rangle\\
&\simeq C_1t^{\nu_1-2}\rho\left(\frac{\tau}{t}\right)+C_2t^{\nu_2-1}\delta(\tau),
\end{split}$$ where the parameters $\nu_1,\nu_2$ and the scaling function $\rho(\cdot)$ are determined by $p_\pm(t)$ and $p_{+,0}(t,t+\tau)$. The scaling form helps to show different scaling behaviors of $\langle v(t)v(t+\tau)\rangle$ for different sojourn time distributions $\psi_\pm(t)$, and brings convenience to give a generic expressions of MSDs [@DechantLutzKesslerBarkai:2014; @MeyerBarkaiKantz:2017].
Now we focus on the aging process $x_{t_a}(t)$. The EAMSD of this aging process is defined as $\langle x_{t_a}^2(t)\rangle=\langle (x(t_a+t)-x(t_a))^2\rangle$, which can be obtained through the scaling form in . For weak aging $t_a\ll t$ and strong aging $t_a\gg t$ cases (see Supplemental Material), it behaves as $$\label{EA}
\langle x_{t_a}^2(t)\rangle\simeq
\left\{
\begin{array}{ll}
K_1 t^{\nu_1}+K_2 t^{\nu_2}, & t_a\ll t, \\
K_3 t_a^\beta t^{\nu_1-\beta}+C_2 t_a^{\nu_2-1}t, & t_a\gg t,
\end{array}
\right.$$ where the coefficients $K_1=2C_1/\nu_1\int_0^\infty dt (t+1)^{-\nu_1 }\rho(t)$, $K_2=C_2/\nu_2$ and $K_3=2c_1C_1[(\nu_1-\beta-1)(\nu_1-\beta)]^{-1}$. Here $c_1$ depends on the asymptotic form of scaling function $\rho(z)\simeq c_1 z^{-\delta_1}$ for small $z$, and $\beta$ is the exponent of the variance of velocity in the Lévy walk phase for large $t$ [@DechantLutzKesslerBarkai:2014], i.e., $$\langle v_+^2(t)\rangle=v_0^2p_+(t)\propto t^{\beta}.$$
When constructing single particle tracking experiments, the aging process $x_{t_a}(t)$ is evaluated in terms of its TAMSD, which is defined as $\overline{\delta_{t_a}^2(\Delta)}=1/(T-\Delta)\int_{t_a}^{t_a+T-\Delta}dt [x(t+\Delta)-x(t)]^2$ with $\Delta$ denoting the lag time and $T$ the total measurement time [@MetzlerJeonCherstvyBarkai:2014]. The TAMSD is calculated in the limit $\Delta\ll T$ to obtain good statistics. Weak ergodicity breaking is the common phenomenon of a majority of anomalous diffusion. Similarly to the procedure of calculating EAMSD, we obtain the ensemble-averaged TAMSD as (see Supplemental Material): $$\label{TA}
\langle\overline{\delta_{t_a}^2(\Delta)}\rangle\simeq
\left\{
\begin{array}{ll}
\frac{K_3}{1+\beta} \,T^{\beta}\Delta^{\nu_1-\beta}+K_2 T^{\nu_2-1}\Delta, & t_a\ll T, \\
K_3 t_a^\beta \Delta^{\nu_1-\beta}+C_2 t_a^{\nu_2-1}\Delta, & t_a\gg T.
\end{array}
\right.$$
There are at least four findings being worth to report from the observations of the generic results of EAMSDs in and TAMSDs in . (i) All the four mentioned formulae consist of two parts (one from Lévy walk phase and another one from Brownian phase). The exponents of evolution time $t$ or time lag $\Delta$ in these two parts might be different from the ones of the corresponding individual Lévy walk and Brownian motion. This is because the PDF $p_\pm(t)$ in plays a weighted role on Lévy walk and Brownian motion. Besides, the sums of exponents of the time variables (including $t,t_a,T,\Delta$) in individual two parts are $\nu_1$ and $\nu_2$, respectively, whatever it is EAMSD or TAMSD, and weak or strong aging cases. (ii) The exponents of time variables in weak and strong aging cases are closely related for TAMSD in . While keeping the exponents of $\Delta$ invariant and replacing measurement time $T$ by age $t_a$, the result for strong aging case is obtained from the one of weak aging case. In other words, the TAMSD for weak aging case only depends on $T$ and $\Delta$, while in the same way it counts on $t_a$ and $\Delta$ for strong aging cases. (iii) The EAMSD and TAMSD in weak aging case do not depend on the age $t_a$, the results of which are identical to the non-aging case $t_a = 0$. In contrast, they explicitly depend on $t_a$ for strong aging case, which implies that the exponents $\beta$ and $\nu_2-1$ of $t_a$ must be zero if the equilibrium initial ensemble (i.e., $t_a\rightarrow\infty$ discussed in last section) of this system exists (see specific case $2$ in Table \[table\]). And in this case, the TAMSD will be the same for weak and strong aging cases, and only depends on $\Delta$. (iiii) Comparing the strong aging EAMSD and the mean of TAMSD , it can be noted that $$\langle x_{t_a}^2(\Delta)\rangle = \langle\overline{\delta_{t_a}^2(\Delta)}\rangle \quad {\rm for }~ t_a\gg T,$$ which shows that the aging seemingly makes the weak ergodicity breaking system to be ergodic. It is clear that for any $\alpha_-$ Brownian motion is ergodic in its own phase. However, for TAMSD in Lévy walk phase, there are some differences between $\alpha_+<1$ and $1<\alpha_+<2$. For $1<\alpha_+<2$, the mean sojourn time in Lévy walk phase is finite, individual trajectories become self-averaging at sufficiently long (infinite) times, such that there will be no difference between $\overline{\delta_{t_a}^2(\Delta)}$ obtained from different trajectories and ensemble-averaged quantity $\langle\overline{\delta_{t_a}^2(\Delta)}\rangle$ [@GodecMetzler:2013; @FroembergBarkai:2013]. While for $\alpha_+<1$, the characteristic time scale is infinite, then the individual TAMSD $\overline{\delta_{t_a}^2(\Delta)}$ is irreproducible and inequivalent with the corresponding EAMSD.
*Specific cases*. Since both $\alpha_+$ and $\alpha_-$ go through the range $(0,2)$, it can be divided into six cases as shown in Table \[table\]. See the detailed derivations of parameters $\nu_1,\nu_2$, and $\beta$ for these cases in (Supplemental Material). It seems tedious to discuss the EAMSDs and TAMSDs individually for six different cases of $\alpha_\pm$. In fact, they can be organized into three categories to deepen understandings of the two-state process by considering the properties of its ingredients — Lévy walk and Brownian motion. It is well-known that the standard Lévy walk performs ballistic diffusion when the exponent of the distribution of running times $\alpha<1$ and sub-ballistic superdiffusion when $1<\alpha<2$, which is faster than the normal diffusion of Brownian motion. Based on this understanding, the Brownian phase undoubtedly suppresses the diffusion behavior of Lévy walk. This effect may be durable or transient, which is completely determined by the fraction of two states $p_\pm(t)$, or more essentially, the magnitude of the exponents $\alpha_\pm$. From this point of view, the three categories are: (i) $\alpha_+$ and $\alpha_-$ are comparable, including the first two cases in Table \[table\]; (ii) $\alpha_+$ is smaller, including the middle two cases in Table \[table\]; (iii) $\alpha_-$ is smaller, including the last two cases in Table \[table\].
specific cases $\nu_1$ $\nu_2$ $\beta$
----------------------------------- ----------------------- ----------------------- ---------------------
1\. $\alpha_+=\alpha_-<1$ $2$ $1$ $0$
2\. $1<\alpha_\pm<2$ $3-\alpha_+$ $1$ $0$
3\. $\alpha_+<\alpha_-<1$ $2$ $\alpha_+-\alpha_-+1$ $0$
4\. $\alpha_+ < 1 < \alpha_- < 2$ $2$ $\alpha_+$ $0$
5\. $\alpha_-<\alpha_+<1$ $\alpha_--\alpha_++2$ $1$ $\alpha_--\alpha_+$
6\. $\alpha_-<1<\alpha_+<2$ $\alpha_--\alpha_++2$ $1$ $\alpha_--1$
: Values of several major parameters of EAMSD and TAMSD in and for six cases with different $\alpha_\pm$.[]{data-label="table"}
As representatives of the above three situations, we choose three sets of parameters: (i) $\alpha_+=1.5,\alpha_-=1.8$, (ii) $\alpha_+=0.6,\alpha_-=1.5$, and (iii) $\alpha_+=1.5,\alpha_-=0.8$. The corresponding TAMSDs for weak and strong aging cases are simulated and shown in Fig. \[figure\]. The TAMSDs for other cases and EAMSDs are presented in (Supplemental Material). The theoretical TAMSDs for these three cases can be obtained from as: $$\label{case1}
\textrm{(i)}~ \langle \overline{\delta^2_{t_a}(\Delta)}\rangle \simeq
\left\{
\begin{array}{ll}
2D_2\Delta^{3-\alpha_+}+\frac{2D\mu_-}{\mu_++\mu_-}\Delta, & t_a\ll T, \\
2D_2\Delta^{3-\alpha_+}+\frac{2D\mu_-}{\mu_++\mu_-}\Delta, & T\ll t_a,
\end{array}
\right.$$ $$\label{case2}
\textrm{(ii)}~\langle \overline{\delta^2_{t_a}(\Delta)}\rangle \simeq
\left\{
\begin{array}{ll}
v_0^2\Delta^2+\frac{2D\mu_-}{a_+\Gamma(1+\alpha_+)}T^{\alpha_+-1}\Delta, & t_a\ll T, \\
v_0^2\Delta^2+\frac{2D\mu_-}{a_+\Gamma(\alpha_+)}t_a^{\alpha_+-1}\Delta, & T\ll t_a,
\end{array}
\right.$$ $$\label{case3}
\textrm{(iii)}~\langle \overline{\delta^2_{t_a}(\Delta)}\rangle \simeq
\left\{
\begin{array}{ll}
2D_4T^{\alpha_--1}\Delta^{3-\alpha_+}+2D\Delta, & t_a\ll T, \\
2D_4\alpha_-t_a^{\alpha_--1}\Delta^{3-\alpha_+}+2D\Delta, & T\ll t_a.
\end{array}
\right.$$ For the first category (i), a stationary of the fractions of two states $p_\pm(t)$ can be achieved for long times, that is, $p_\pm(t)$ tends to a constant not equal to $0$ or $1$ (see Supplemental Material). Then the EAMSD and TAMSD are the combination of the fraction of analogues of individual Lévy walk and Brownian motion whether it is weak aging or strong aging. For the second category (ii) with $\alpha_+<\alpha_-$ where $p_+(t)\rightarrow1$ as $t\rightarrow\infty$, the Lévy walk phase in state ‘$+$’ tends to occupy the whole time. Then the results are naturally similar to an individual Lévy walk, except for the small asymptotic form $\Delta$ resulting from Brownian phase. For the third category (iii) with $\alpha_+>\alpha_-$, by contrast, now $p_-(t)\rightarrow1$ as $t\rightarrow\infty$ and the Lévy walk phase in state ‘$+$’ gradually withdraws from the two states in a power-law way. This power-law way suppresses the diffusion of Lévy walk phase and gives the opportunity to Brownian motion to be the leading term when $\alpha_+-\alpha_->1$. In conclusion, compared to the EAMSD and TAMSD of individual aging Lévy walk [@MagdziarzZorawik:2017] and Brownian motion, it can be found that the fraction $p_\pm(t)$ in a two-state process plays a crucial role. It contributes a power term of $\Delta$ to weak aging EAMSD, a power term of $T$ to weak aging TAMSD, and a power term of $t_a$ to strong aging EAMSD and TAMSD.
The model Lévy walk interrupted by rest has attracted considerable attention in physics [@SolomonWeeksSwinney:1993; @KlafterZumofen:1994] and biology [@SongMoonJeonPark:2018]. The EAMSD and TAMSD for this model can be obtained by taking the diffusivity $D$ in Brownian phase to be zero. It has been pointed that all the results above consist two parts corresponding to Lévy walk and Brownian motion, respectively. Taking $D = 0$ just eliminates the latter part and brings no effect on the former part of Lévy walk phase. For Lévy walk interrupted by rest, the asymptotic behavior of small $\Delta$ in TAMSD disappears and subdiffusion behavior might exist if $\alpha_+-\alpha_->1$.
*Initial ensemble*. In general, the standard Lévy walk model is a non-Markovian process and so is the two-state process alternating between Lévy walk and Brownian motion with power-law distributed sojourn time. It is natural to consider the effects of the initial ensembles of the particles. It is called a nonequilibrium initial ensemble [@Cox:1962; @KlafterZumofen:1993] if all particles are introduced to the system at $t = 0$ without any prehistories. In contrast, if the particles have been evolving for time $t_1$ before we start to measure this system, we call this system with equilibrium initial ensemble when $t_1\rightarrow\infty$ [@Cox:1962; @KlafterZumofen:1993]. The EAMSD of standard Lévy walk has been shown to be different for different initial ensemble [@ZaburdaevDenisovKlafter:2015; @WangChenDeng:2019]. Note that the equilibrium initial ensemble exists only if the sojourn times in two states ‘$\pm$’ both have finite first moments, i.e., $1 <\alpha_\pm < 2$ in our concerned model.
For nonequilibrium initial ensemble, the corresponding EAMSD $\langle x^2(\Delta)\rangle$ and TAMSD $\overline{\delta^2(\Delta)}$ can be obtained by taking $t_a=0$ in previous section, i.e., $$\begin{split}
\langle x^2(\Delta)\rangle = \langle x_{t_a}^2(\Delta)\rangle |_{t_a=0}, \quad
\overline{\delta^2(\Delta)} = \overline{\delta_{t_a}^2(\Delta)} |_{t_a=0}.
\end{split}$$ Since the results of the weak aging case (i.e., $t_a\ll\Delta$) with different sojourn time pairs $\psi_\pm(t)$ in Eqs. (\[EA\]) and (\[TA\]) are independent of $t_a$, they are indeed the results for nonequilibrium initial ensemble. When $1<\alpha_\pm<2$, the results of the strong aging case (i.e., $t_a\gg T$) in are independent of $t_a$. Therefore, the EAMSD $\langle x^2_{\textrm{eq}}(\Delta)\rangle$ and TAMSD $\langle\overline{\delta^2_{\textrm{eq}}(\Delta)}\rangle$ for equilibrium initial ensemble ($t_a\rightarrow\infty$) are $$\langle x^2_{\textrm{eq}}(\Delta)\rangle=\langle\overline{\delta^2_{\textrm{eq}}(\Delta)}\rangle
\simeq2D_2\Delta^{3-\alpha_+}+ \frac{2D\mu_-}{\mu_++\mu_-}\Delta.$$ If the sojourn times are so long that the mean sojourn time diverges, there is no sense in talking about the equilibrated initial ensemble. However, the asymptotic behaviors of strong aging case $t_a\gg\Delta$ can still be investigated (see Supplemental Material). There is a special case $0<\alpha_+=\alpha_-<1$, where the particles reach a balance that each half of them are located in each of the two states and the EAMSD and TAMSD are both independent on the age $t_a$, that is, $$x^2_{t_a}(\Delta)\rangle = \langle\overline{\delta^2_{t_a}(\Delta)}\rangle
\simeq \frac{v_0^2}{2}\Delta^2 +D\Delta$$ for sufficiently large $t_a$. If $\alpha_+\neq\alpha_-$ and at least one of them less than $1$, then neither an equilibrium initial ensemble nor a balance for long time exists. The state with small exponent $\alpha_\pm$ of sojourn time distribution will dominate the MSD for long times. One can see this phenomenon in the last four cases in Table \[table\]. In these cases, the EAMSD and TAMSD for strong aging cases all consist of two parts corresponding to Lévy walk and Brownian motion. One of the parts is independent on $t_a$ while another part contains a power term of $t_a$ with a negative exponent. The latter part tends to zero as $t_a\rightarrow\infty$ and the former one dominates, which corresponds to the state with smaller exponent $\alpha_\pm$ of sojourn time distributions.
*Conclusion*. It often happens that a single-state process cannot sufficiently describe the observed physical and biological phenomena. Two-state process is a kind of simple but important model to characterize some of these phenomena. A Langevin equation with two diffusion modes (fast and slow diffusion modes) has been investigated in [@MiyaguchiAkimotoYamamoto:2016], where a transient subdiffusion and the non-Gaussian propagator for short time are observed for a nonequilibrium ensemble. In this report, we consider a two-state process with fast phase (Lévy walk) and slow phase (Brownian motion), which is also the intermittent search process for finding rare hidden targets. It is not easy to model the process with two completely different phases by a Langevin equation. By contrast, we resort to the velocity process $v(t)$, which also consists of two states. Based on the velocity correlation function, we obtain the generic expressions of the EAMSD and TAMSD for different sojourn time distributions.
One of the key contributions of this report is to explicitly discuss the relation between EAMSD and TAMSD. In particular, the weak and strong aging cases are also considered for these MSDs since the measurement in experiments might not begin at the start of the process concerned. It is found that the occupation fraction plays a weighed role in Lévy walk phase and Brownian phase, and the MSDs are just a combination of these two parts. The meticulous discussions on the aging MSDs are helpful to understand the two-state process and to analyze the experimental data. If taking the diffusivity $D$ to be zero in Brownian phase, we obtain another important process — Lévy walk interrupted by rest. Taking $D=0$ just eliminates the contributions from Brownian phase. From another aspect of the two-state process, we find the fact that the slow phase, whether it is rest or Brownian motion, suppresses the diffusion behavior of Lévy walk if its sojourn time is longer than that of Lévy walk phase. The mechanism is similar to the trap event [@GoldingCox:2006] in CTRW models. Compared to them, there exist some other models describing the suppression of the diffusion of Lévy walk with different mechanism, such as the Lévy walk with memory in running time [@ChenWangDeng:2019] and the walker moving in a heterogeneous medium [@KaminskaSrokowski:2018].
*Acknowledgments*. This work was supported by the National Natural Science Foundation of China under grant no. 11671182, and the Fundamental Research Funds for the Central Universities under grant no. lzujbky-2018-ot03.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a directed search for continuous gravitational waves in data from the sixth initial LIGO science run. The target was the nearby globular cluster NGC 6544 at a distance of $\approx 2.7$ kpc. The search covered a broad band of frequencies along with first and second frequency derivatives for a fixed sky position. The search coherently integrated data from the two LIGO interferometers over a time span of 9.2 days using the matched-filtering $\mathcal{F}$-statistic. We found no gravitational-wave signals and set 95% confidence upper limits as stringent as $6.0 \times 10^{-25}$ on intrinsic strain and $8.5 \times 10^{-6}$ on fiducial ellipticity. These values beat the indirect limits from energy conservation for stars with characteristic spindown ages older than 300 years and are within the range of theoretical predictions for possible neutron-star ellipticities. An important feature of this search was use of a barycentric resampling algorithm which substantially reduced computational cost; this method will be used extensively in searches of Advanced LIGO and Virgo detector data.'
bibliography:
- 'S6GlobClust.bib'
title: Search for continuous gravitational waves from neutron stars in globular cluster NGC 6544
---
Introduction
============
The and Virgo Collaboration have undertaken numerous searches for continuous . None has yet detected a signal, but many have placed interesting upper limits on possible sources. These searches have generally been drawn from one of three types.
Targeted searches are aimed at a single known pulsar, with a known precise timing solution. The first search for continuous waves, using data from , was of this type [@S1pulsar], and subsequent searches have probed the Crab and Vela pulsars, among others [@S2pulsars; @S3S4pulsars; @S5Crab; @VirgoVela; @S5pulsars; @S6pulsars]. A number of these most recent searches have been able to set direct upper limits on emission comparable to or stricter than the indirect “spin-down limits” (derived from energy conservation, as well as the distance from Earth of the target, its gravitational-wave frequency, and the frequency’s first derivative, the “spin-down”) for a few of the pulsars searched.
All-sky searches, as their name suggests, survey the entire sky for neutron stars not seen as pulsars. These are very computationally costly, searching over wide frequency bands and covering large ranges of spin-down parameters [@S2Hough; @S2Fstat; @S4PSH; @S4Einstein; @S5PowerFlux; @S5Einstein; @S5PowerFlux2; @S5Einstein2; @S5Hough; @VirgoAllSky]. The latest of these have incorporated new techniques to cover possible binary parameters as well [@AllSkyBinary]. Recent all-sky searches have set direct upper limits close to indirect upper limits derived from galactic neutron-star population simulations [@KnispelAllen].
Directed searches sit between these two extremes. As in the all-sky case, their targets are neutron stars not seen as pulsars, so that the frequency and other parameters are unknown. They focus, however, on a known sky location (and therefore a known detector-frame Doppler modulation). This directionality allows for searching over a wide range of frequencies and frequency derivatives while remaining much cheaper computationally than an all-sky search without sacrificing sensitivity. This approach was first used in a search for the accreting neutron star in the low-mass X-ray binary Sco X-1 [@S2Fstat; @S4radiometer; @S5Stochastic].
The search for the in the Cassiopeia A (Cas A)[@S5CasA] was the first directed search for a young neutron star without electromagnetically detected pulsation, motivated by the idea that young neutron stars might be promising emitters of continuous . The Cas A search [@S5CasA] set upper limits on strain which beat an indirect limit derived from energy conservation and the age of the remnant [@Wette2008] over a wide frequency band. Other directed searches have since followed in its footsteps, using different data analysis methods, for supernova 1987A and unseen neutron stars near the galactic center [@S5Stochastic; @S5GalacticCenter]. Most methodologically similar to this search and the S5 Cas A search was a recent search for nine supernova remnants [@S6CasAFriends], which also used fully coherent integration over observation times on the order of 10 days.
In this article, we describe a search of data from for potential young isolated neutron stars with no observed electromagnetic pulsations in the nearby ($d \approx 2.7$ kpc) globular cluster NGC 6544. Globular clusters are unlikely to contain young neutron stars, but in these dense environments older neutron stars may be subject to debris accretion (see Sec. \[s:TS\]) or other events which could render them detectable as gravitational wave sources. This particular globular cluster was chosen so that a computationally-feasible coherent search similar to [@S5CasA] could beat the age-based indirect limits on emission.
The search did not find a signal, and hence the main result is a set of upper limits on strain amplitude, fiducial ellipticity, and $r$-mode amplitude $\alpha $, similar to those presented in [@S5CasA]. An important new feature of the search described here was use of a barycentric resampling algorithm which substantially reduced computational cost, allowing a search over a larger parameter space using a longer coherence time (see Sec. \[s:SPS\]). This barycentric resampling method will be used extensively in searches of Advanced LIGO and Virgo detector data.
This article is structured as follows: In Sec. \[s:searches\] we present the method, implementation, and results of the search. The upper limits set in the absence of a signal are presented in Sec. \[s:uls\], and the results are discussed in Sec. \[s:disc\].
Searches {#s:searches}
========
Data selection
--------------
The sixth initial LIGO science run (S6) extended from July 7 2009 21:00:00 UTC (GPS 931035615) to October 21 2010 00:00:00 UTC (GPS 971654415) and included two initial LIGO detectors with 4-km arm lengths, H1 at near Hanford, Washington and L1 at near Livingston, Louisiana.
After optimization at fixed computing cost determined an optimum coherence time of 9.2 days (see Sec. \[s:SPS\]), two different methods were used to determine which data would be searched, producing two different 9.2-day stretches. Both were searched, allowing for the comparison of search results between them.
The first method was to look for the most sensitive average data from S6. This was done by taking nine week-long data samples from each detector spaced roughly 55 days apart, giving nine evenly spaced weeks throughout the duration of S6. The data samples used are shown in Table \[SampleTable\]. We chose four representative frequencies ($100$ Hz, $200$ Hz, $400$ Hz, $600$ Hz) and generated joint-detector strain noise power spectral densities (PSDs) in $1$-Hz bands about these frequencies, using $0.01$-Hz binning. The sensitivity $h_{\textup{sens}}$ was then taken to be
$$\label{SensitivityEquation}
h_{\textup{sens}}^{j} = \left ( \frac{1}{\sqrt{(1/100) \cdot \sum_{i = 0}^{100}(S_{h}^{i}(f_{i}))^{-1}}} \right )_{j}$$
where $S_{h}^{i}(f)$ represents the PSD value of the $i^{th}$ bin, at frequency $f_{i}$, and the index $j$ runs from 1 through 4 and represents the four representative frequencies (note that this is not an actual estimate of detectable strain). Based on this figure of merit, the final nine days of S6 yielded the most sensitive data stretch for all four frequencies: October 11-20, 2010 (GPS 970840605 – 971621841).
[| c | c | c | c |]{}\
Label & GPS Start & GPS End & Dates (UTC)\
Week 1 & 931053000 & 931657800 & Jul 8-15, 2009\
Week 2 & 936053000 & 936657800 & Sep 3-10, 2009\
Week 3 & 941053000 & 941657800 & Oct 31-Nov 7, 2009\
Week 4 & 946053000 & 946657800 & Dec 28, 2009-Jan 4, 2010\
Week 5 & 951053000 & 951657800 & Feb 24-Mar 3, 2010\
Week 6 & 956053000 & 956657800 & Apr 23-30, 2010\
Week 7 & 961053000 & 961657800 & Jun 20-27, 2010\
Week 8 & 966053000 & 966657800 & Aug 17-24, 2010\
Week 9 & 971053000 & 971657800 & Oct 14-21, 2010\
An alternate data selection scheme [@S6CasAFriends; @S5CasA], which takes detector duty cycle into account is to maximize the figure of merit
$$\label{FOM}
\sum_{k,f} \frac{1}{S_h(f)}$$
where $S_h(f)$ represents the strain noise power spectral density at frequency $f$ in the $k^{th}$ , and the sum is taken across all frequencies $f$ in the search band and all SFTs in a given 9.2-day (see Sec. \[s:SPS\] below) observation time. The format is science-mode detector data split into 1800s segments, band-pass filtered from 40–2035 Hz, Tukey windowed in the time domain, and Fourier transformed. This method favored a different data stretch: July 24–August 2, 2010 (GPS 964007133 – 964803598). This second data stretch had slightly worse average sensitivity than the first, but a higher detector livetime: our first (October) data set contained 374 SFTs (202 from Hanford and 172 from Livingston) with average sensitivity $h_{sens}^{200\textup{Hz}} = 1.92 \times 10^{-23}$; the second (July-August) data set contained 642 SFTs (368 from Hanford and 274 from Livingston) with average sensitivity $h_{sens}^{200\textup{Hz}} = 1.95 \times 10^{-23}$.
Analysis method {#s:A_method}
---------------
The analysis was based on matched filtering, the optimal method for detecting signals of known functional form. To obtain that form we assumed that the potential target neutron star did not glitch (suffer an abrupt frequency jump) or have significant timing noise (additional, possibly stochastic, time dependence of the frequency) [@LyneSmith2006] during the observation. We also neglected third and higher derivatives of the frequency, based on the time span and range of $\dot{f}$ and $\ddot{f}$ (the first two derivatives) covered. The precise expression for the interferometer strain response $h(t)$ to an incoming continuous also includes amplitude and phase modulation by the changing of the beam patterns as the interferometer rotates with the earth. It depends on the source’s sky location and orientation angles, as well as on the parameters of the interferometer. The full expression can be found in [@Jaranowski1998].
The detection statistic used was the multi-interferometer $\mathcal{F}$-statistic [@Cutler2005], based on the single-interferometer $\mathcal{F}$-statistic [@Jaranowski1998]. This combines the results of matched filters for the signal in a way that is computationally fast and nearly optimal [@Prix2009]. Assuming Gaussian noise, $2\mathcal{F}$ is drawn from a $\chi^2$ distribution with four degrees of freedom.
We used the implementation of the $\mathcal{F}$-statistic in the LALSuite package [@lalsuite]. In particular most of the computing power of the search was spent in the `ComputeFStatistic_v2` program. Unlike the version used in preceding methodologically similar searches [@S5CasA; @S6CasAFriends], this one implements an option to use a barycentric resampling algorithm which significantly speeds up the analysis.
The method of efficiently computing the $\mathcal{F}$-statistic by barycentering and Fast-Fourier-transforming the data was first proposed in [@Jaranowski1998]. Various implementations of this method have been developed and used in previous searches, such as [@Krolak2010; @PinkeshMethods; @PatelThesis]. Here we are using a new LALSuite [@lalsuite] implementation of this method, which evolved out of [@PatelThesis], and which will be described in more detail in a future publication. It converts the input data into a heterodyned, downsampled timeseries weighted by antenna-pattern coefficients, and then resamples this timeseries at the solar system barycenter using an interpolation technique. The resampled time series is then Fourier-transformed to return to the frequency domain, and from there the $\mathcal{F}$-statistic is calculated. For this search, both single-detector and multi-detector $\mathcal{F}$-statistics were calculated (see Vetoes section below).
Timing tests run on a modern processor (ca. 2011) showed that the resampling code was more than 24 times faster in terms of seconds per template per SFT. This improvement, by more than an order of magnitude, was used to perform a deeper search over a wider parameter space than previously possible for the computational cost incurred (see target selection and search parameters below).
Target selection {#s:TS}
----------------
Unlike previous directed searches, this one targets a globular cluster. Since stars in globular clusters are very old, it is unlikely that a young neutron star will be found in such an environment. However, some neutron stars are known to be accompanied by debris disks [@Wang] and even planets [@WolszczanFrail; @Thorsett; @Sigurdsson2003]. In the densely populated core of a globular cluster, close encounters may stimulate bombardment episodes as debris orbits are destabilized, akin to cometary bombardments in our solar system when the Oort cloud is perturbed [@Sigurdsson1992]. A neutron star which has recently accreted debris could have it funneled by the magnetic field into mountains which relax on timescales of $10^{5}$–$10^{8}$ years [@Vigelius] and emit gravitational waves for that duration. Other mechanisms are likely to last a few years at most [@Haskell]. Hence an old neutron star could be a good gravitational wave source with a low spin-down age.
The first step in picking a globular cluster is a figure of merit based on that for directed searches for supernova remnants [@Wette2008], an indirect upper limit on gravitational wave strain based on energy conservation and the age of the object. Here the inverse of the object age is replaced by the interaction rate of the globular cluster, which scales like density$^{(3/2)}$ times core radius$^{2}$ [@VerbuntHut; @Sigurdsson1992], reflecting the mean time since last bombardment. It is hard to know when the most recent bombardment episode was, and thus the constant factor out in front, but globular clusters can be ranked with respect to each other by a maximum-strain type figure of merit
$$\label{GCFOM}
h_{0} \propto \rho_{c}^{3/4} r_{c}/d,$$
where $\rho_{c}$ is the globular cluster core density, $r_{c}$ is the core radius, $d$ is the distance to the cluster, and thus $r_{c}/d$ is the angular radius of the core. We ranked the Harris catalog of globular clusters [@Harris1; @Harris2] by this figure of merit and looked at the top few choices, which were mainly nearby core-collapsed clusters. The closest is NGC 6397 at $\approx 2.2$ kpc, but it is at high declination. This lessens the Doppler modulation of any gravitational wave signal, making it harder to distinguish from stationary spectral line artifacts, which tend to contaminate searches at high declination near the ecliptic pole. Hence we chose the next closest, NGC 6544, which is at a declination of less than 30 degrees and only slightly further away at $\approx 2.7$ kpc.
We restrict the search described below to sources for which the bombardment history corresponds to a characteristic spindown age older than 300 years. The figure of 300 years is mainly a practical consideration: the cost of a search rises steeply for lower spin-down ages, and 300 years proved tractable for the Cas A search [@S5CasA].
Search parameter space {#s:SPS}
----------------------
An iterative method was used to generate the parameter space to be searched. Starting with an (assumed) spin-down age no younger than 300 years, a braking index $n$ = 5 (see below), and the known distance to the globular cluster, we calculated the age-based indirect upper limit. This is an optimistic limit on the gravitational wave strain $h_0$ which assumes that all energy lost as the target neutron star spins down is radiated away as gravitational waves[@Wette2008]:
$$\label{indirect1}
h_0 \leq \frac{1}{d} \sqrt{\frac{5 G I}{2 {c}^{3}\tau (n-1)}}.$$
Here $d$ is the distance to the target, $\tau $ the assumed age of the target object, and $I$ a fiducial moment of inertia for a neutron star ($10^{38} \textup{kg} \cdot \textup{m}^{2}$). $G$ and $c$ are the gravitational constant and the speed of light, respectively. This age-based limit was then superimposed on a curve of expected upper limits in the absence of signal for the LIGO detectors, obtained from the noise harmonically averaged over all of and both interferometers. A running median with a 16-Hz window was further applied to smooth the curve. The curve is given by:
$$\label{senscurve}
h_{0}^{95\%} = \Theta \sqrt{ \frac{S_h} {T_\mathrm{data}} }$$
where $S_{h}$ is the harmonically averaged noise, $T_\mathrm{data}$ is the coherence time (the total data livetime searched coherently), initially estimated at two weeks, and $\Theta$ is a sensitivity factor that includes a trials factor, or number of templates searched, and uncertainty in the source orientation [@Wette2008]. For a directed search like ours, $\Theta$ is approximately 35 [@Wette2008; @WetteSens]. The intersection of this coherence-time adjusted upper limit curve and our indirect limit (Eq. (\[indirect1\])) gives an initial frequency band over which the indirect limit can be beaten. The braking index is related to the frequency parameters by the definition:
$$\label{braking_index}
n = \frac{f\ddot{f}}{\dot{f}^{2}}.$$
Assuming a braking index $n$ between 2 and 7 covers most accepted neutron star models ($n=5$, the neutron star radiating all energy as gravitational waves via the mass quadrupole, is used to obtain the indirect limit). We allow the braking indices in these expressions to range from 2 to 7 independently, to reflect the fact that in general multiple processes are operating and $\dot{f}$ is not a simple power law. This constraint on the braking indices then produces limits on the frequency derivatives given by [@Wette2008]
$$\label{spindown range}
\frac{f} {\tau } \leq -\dot{f} \leq \frac{f} {6\tau}$$
for the spindown at each frequency and
$$\label{second spindown range}
\frac{2 \dot{f}^2} {f} \leq \ddot{f} \leq \frac{7 \dot{f}^2} {f}$$
for the second spindown at each $(f,\dot{f})$. The step sizes for frequency and its derivatives are given by the equations [@WhitbeckThesis2006; @PatelThesis; @ReinhardStep]
$$\label{freqstep}
df = \frac{2\sqrt{3m}}{\pi }\frac{1}{T_\mathrm{data}},$$
$$\label{fdotstep}
d\dot{f} = \frac{12\sqrt{5m}}{\pi }\frac{1}{T^{2}_\mathrm{data}},$$
and
$$\label{fddotstep}
d\ddot{f} = \frac{20\sqrt{7m}}{\pi }\frac{1}{T^{3}_\mathrm{data}}.$$
where $m$ is the mismatch parameter, the maximum loss of $2\mathcal{F}$ due to discretization of the frequency and derivatives [@Owen1996; @Brady1998]. This search used a mismatch parameter $m = 0.2$.
From these relations the total number of templates (points in frequency parameter space) to be searched can be calculated, and with knowledge of the per-template time taken by the code (obtained from timing tests), the total computing time can be obtained. Limiting the target computing time, in our case to 1000 core-months, then allows us to solve for the coherence time $T_\mathrm{data}$, which we then feed back into Eq. (\[senscurve\]) to begin the process anew until it iteratively converges on a parameter space and accompanying coherence time. The iterative algorithm thus balances the computational gains from resampling between the use of a longer coherence time (giving better sensitivity) and the expansion of the parameter space over which the indirect limit can be beaten (caused by the improved sensitivity). The result for the globular cluster NGC 6544 is a search over the frequency range 92.5 Hz to 675 Hz, with a coherence time of 9.2 days.
The peculiar velocities of globular clusters are negligible, as they represent an essentially constant Doppler shift of order $1 \times 10^{-3}$; so is velocity dispersion, which is an order of magnitude smaller. Since we search down to 300-year timescales, the acceleration of the cluster is also not an issue [@MillionBody].
Implementation
--------------
All searches were run on the LIGO-Caltech Computing Cluster at the California Insitute of Technology in Pasadena, CA, under the control of the Condor queuing system for parallel processing. The search process was split into 5825 individual Condor jobs, each of which searched over a $0.1$-Hz subband and corresponding swathes of $(\dot{f},\ddot{f})$. The number of templates searched by each job thus varied as a function of frequency.
Each search job produced three distinct outputs. First, a record was made of all candidates with $2\mathcal{F}$ above 45.0, a choice of recording different from the search which recorded the loudest $0.01\%$ of events. This was needed because of the contamination of the noise by detector artifacts, as well as limits on the disk space available and the input/output capability of the cluster filesystem. Second, a histogram of $2\mathcal{F}$ values for all templates searched was produced to verify that the data matched the expected chi-square distribution (described in Subsec. \[s:A\_method\] above). Last, each job produced a record of the loudest (highest-$2\mathcal{F}$-valued) candidate in its 0.1-Hz band, regardless of threshold. This data was used in the setting and validation of upper limits (see Section \[s:uls\] below).
Vetoes
------
A high value of $2\mathcal{F}$ is not enough to claim a detection, since instrumental artifacts lead to non-Gaussian and/or non-stationary noise in many narrow frequency bands. A variety of veto techniques were used to trim down the initial list of candidates and arrive at a final list of outliers.
Six 0.1 Hz sub-bands (see Table (\[t:BadBands\])) had to be manually aborted in both searches, with a seventh aborted in the July-August search, as even with the threshold in place, they produced an excessive number of candidates. Each of these subbands was compared to records of known noise artifacts and disturbances in the detector, and in each case a known instrumental line was confirmed. These sub-bands were later rerun with the record of candidates disabled in order to produce histograms and loudest-outlier files for upper limit validation.
To protect against spurious noise lines, a second veto based on the $\mathcal{F}$-statistic consistency veto introduced in [@S5Einstein2] was used. This uses the fact that an astrophysical signal should have a higher joint value of $2\mathcal{F}$ (combining data from the two interferometers) than in either interferometer alone. Recorded candidates that violate this inequality were vetoed. This is a simpler version of the more recent line veto [@Keitel2014].
Finally, to enforce coincidence between detectors, a single-detector threshold was employed. Since a true astrophysical signal should be present in both detectors at a significant level, any candidates passing the initial joint-detector detection criteria (see \[s:DCR\]) also had to pass an additional threshold on the individual-detector values of $2\mathcal{F}$.
The 0.1-Hz band between 200 and 200.1 Hz was arbitrarily chosen as a test band. The joint-detector $2\mathcal{F}$ values were taken from the loudest-candidate files and used to semi-analytically compute [@Wette2009] an estimate of the 95% upper limit for that subband using the SFTs employed by the search. Sets of 1,000 software injections were performed with strengths of 100%, 80%, 60%, 40% and 20% of this estimated upper limit. The results were used to set a threshold of $2\mathcal{F} \geq 20$ in each individual detector, leading to an additional false dismissal rate of 1.5% of injections at the 95% confidence upper limit estimate. Candidates failing to meet this criterion were vetoed.
Detection criteria and results {#s:DCR}
------------------------------
The results of a mock data challenge were used to set a detection criterion for the joint-detector $2\mathcal{F}$ value. The mock data challenge consisted of a set of 1577 artificial signals injected into a set of real detector data from S6, which were then searched for using the same resampled $\mathcal{F}$-Statistic used in the search. A survey of the loudest joint-detector $2\mathcal{F}$ value reported for background subbands known to be free of injected signals for the band between 200 Hz and 240 Hz (used in a pilot run) gave a mean loudest joint-detector $2\mathcal{F} \approx 55$. Given this background level, the detection criterion was chosen to be joint-detector $2\mathcal{F} \geq 60$ to maintain high efficiency and low false-alarm rate (the false-alarm rate was 3.17% in these pilot subbands).
With these detection criteria, a search was carried out in data. The lists of all templates with joint-detector $2\mathcal{F}$ greater than 45.0 were filtered for the individual detector threshold and the consistency veto, both singly and in tandem. If the loudest template failed either check, the list was used to move to the next-loudest template until the loudest template passing all thresholds and vetoes was identified. This created three sets of results (threshold-only, veto-only, and threshold+veto) which could all be queried independently.
The joint $2\mathcal{F}$ values for the loudest single template (passing all thresholds and vetoes) in each 0.1 Hz subband were collated into lists spanning 10Hz (100 joint $2\mathcal{F}$ values per list). These lists were then parsed, and any joint $2\mathcal{F}$ values greater than the joint $2\mathcal{F}$ threshold of 60 were identified. Each such entry’s corresponding template was then added to a list of outliers. This method produced a list of 168 outliers for the entirety of the search band in the October data, and a list of 155 outliers for the entirety of the search band in the July-August data.
These outliers were then tested using *time shifts* and *extended looks*. In a time shift, the frequency parameters of the outliers from each data stretch (October and July-August) were evolved forwards or backwards in time, as appropriate, and sought in the opposite data stretch, under the assumption that a true astrophysical signal should be present in both data sets for the implicitly long-lived signals searched for here. A set of 1000 software injections (simulated signals with randomly generated parameters) underwent the same treatment to provide a baseline $2\mathcal{F}$ threshold for signal detection, and outliers surpassing the threshold were considered present.
In an extended look, each outlier was sought in an expanded 20-day coherence time encompassing the original nine-day coherence time; the same assumption of signal continuity would predict, roughly, a doubling of the $2\mathcal{F}$ value for a doubling of coherence time. These cases as well were tested with software injections to determine a threshold.
In both time shifts and extended looks, the searches were conducted over a parameter space envelope obtained by starting at the outlier frequency parameters $(f, \dot{f}, \ddot{f}) \pm 2$ bins, and evolving those ranges backwards or forwards in time using the extremum values of the next derivative (e.g., $f$ evolved at maximum $\dot{f}$, $\dot{f}$ evolved at maximum $\ddot{f}$) to achieve a conservatively wide envelope.
Outliers detected with joint $2\mathcal{F}$ greater than the threshold established by the software injections were labeled candidates and received manual followup. These tests were not cumulative; an outlier needed only to survive any one test, not all of them, to persist as a candidate. The software injection threshold for both types of test was placed at a value for joint $2\mathcal{F}$ yielding 80% injection recovery; because each outlier would receive further consideration if it passed either test, the false dismissal probability for the first follow-up stage was $\approx 4\%$. The combined 323 outliers produced only seven candidates, listed in Table \[CandidateTable\].
These candidates were subject to manual followup. They were compared to strain histograms of run-averaged (i.e., over all of S6) spectra from each detector, to identify instrumental noise lines which could be responsible. In five of the seven cases, the strain histograms gave clear evidence of an instrumental noise line responsible for the candidate, and in these cases records of prior detector characterization studies were consulted to provide explanations for the noise artifacts. In those cases the artifact is listed in Table \[CandidateTable\] as well. Two of the artifacts arose from hardware injections at different points in the sky, used to test interferometer response [@S6CasAFriends].
The remaining two candidates were given another round of followup, with a time shift and extended look performed in data from June 2010, the farthest removed (in the time domain) available data of comparable sensitivity. The large time separation creates a large difference in the Doppler corrections needed to reconstruct an astrophysical source, making these corrections unlikely to reinforce instrumental or environmental artifacts. Both outliers failed to pass the $2\mathcal{F}$ thresholds established by software injections in any of their June tests.
The loudest $2\mathcal{F}$ value expected in the absence of signal depends on the number of templates searched [@Wette2009];[^1] for our search, the largest expected $2\mathcal{F}$ value lies in the range $72 \leq 2\mathcal{F} \leq 80$ with 90% confidence. The $2\mathcal{F}$ values associated with the two remaining candidates, outliers 79 and 131, were joint-detector $2\mathcal{F} = 61.3$ and joint-detector $2\mathcal{F} = 61.9$, respectively. The outliers’ failure to pass the June tests and their marginal $2\mathcal{F}$ values led us to dismiss them as noise fluctuations.
Thus no credible gravitational wave signals were detected by our search. In the absence of a detection, we can set upper limits on the possible strength of gravitational waves in our data.
Upper limits {#s:uls}
============
Methods
-------
The method of setting upper limits was a variation on that used in [@S5CasA]and [@S6CasAFriends]. This upper limit determination is based only of the $\mathcal{F}$-statistic and does not include additional criteria involved in candidate followup. We split the frequency band into 0.1-Hz subbands, and for each of these used a semi-analytic Monte Carlo method to estimate a 95% upper limit, defined as the strain $h_0$ at which our detection criterion would successfully detect 95% of signals. Due to the high computational cost of individually verifying all 5800 such subbands, these 0.1-Hz upper limit bands were consolidated into 1-Hz subbands. For each such 1-Hz band, we performed 1,000 software injections, split into eight groups of 125 signals. The strain $h_0$ of each group was $\pm 5\%, \pm 10\%, \pm 15\%, \textup{and} \pm 20\%$ of the semi-analytic , respectively. A software injection was considered validated if it returned a value of $2\mathcal{F}$ greater than or equal to the loudest outlier in its 0.1-Hz subband, thus maintaining the original granularity. For each 1-Hz band, these 1,000 injections thus produced eight points on a detection efficiency curve. We then used a least-squares method to produce a sigmoid fit to the data points, and from this curve determined a true 95% upper limit, defined as the value of $h_0$ at which the fitting curve intersected 95% efficiency. Figure \[ExamplePlot\] shows such a plot for a sample band. In cases where the 95% point was extrapolated (as opposed to interpolated) from the eight points, and in cases where the uncertainty on the 95% point was greater than 5%, a new set of eight points were generated using the 95% point as the initial $h_{0}$ estimate and a 95% point was determined from the combined sets (e.g., a curve was fit to 16 points after one rerun, 24 points after two reruns, etc.).
![A demonstration of the upper limit validation technique for a sample band (101 Hz; October data). The x-axis is $h_{nom}$, $h_0$ divided by the for this upper-limit band; the eight points represent detection efficiencies for eight sets of 125 software injections. These eight points are then fit to a sigmoid curve (in black); the 95% upper limit can then be read off from the point where the curve crosses 95% detection efficiency.[]{data-label="ExamplePlot"}](SigmoidLeastSquareFit_101_0.pdf){width="49.00000%"}
A small number of $0.1$-Hz bands had outliers so large that the semi-analytic method failed to converge to an estimate for $h_{0}$. Instrumental artifacts at these frequencies were identified using run-averaged spectra and the respective 1-Hz bands were then rerun with the disturbed $0.1$-Hz subband excluded. The excluded bands are detailed in Table \[ExclBand\].
Results
-------
{width="98.00000%"}
![Upper limits at 95% confidence (red circles) compared to the initial age-based limit on fiducial ellipticity $\epsilon $ (black line).[]{data-label="Ellip"}](FullSpectrum95ULCorrectLivetimeMethodPSDSkyPosCorrectedAgeLimitP2Ellipticity.pdf){width="49.00000%"}
![Upper limits at 95% confidence (red circles) compared to the initial age-based limit on $r$-mode amplitude $\alpha $ (black line).[]{data-label="RMode"}](FullSpectrum95ULCorrectLivetimeMethodPSDSkyPosCorrectedAgeLimitP2RModes.pdf){width="49.00000%"}
Figure \[ULE\] shows the 95% confidence upper limits (ULs) over the full band for the July-August data set, which was the more sensitive of the two because of its much greater livetime (642 SFTs vs. 374 SFTs for the October data set). The blue curve represents the expected sensitivity of the search for this data set, computed from the power spectral density at each frequency; there is good agreement with the ULs. The black line represents the age-based limit derived when first considering the parameter space. Its intersection with the ULs at either end of the plot is a confirmation that we correctly estimated the frequency band to search over.
Figure \[Ellip\] is a similar plot converting the upper limits on $h_{0}$ to upper limits on fiducial ellipticity $\epsilon $, using the formula [@Saulson1994]
$$\label{EllipForm}
\epsilon = \frac{c^{4}}{4\pi ^{2}G}\frac{d}{I_{zz}f^{2}}h_{0}.$$
The black curve represents the age-based limit on fiducial ellipticity, using the same assumptions (braking index $n = 5$, age $\tau = 300$ years) used in the parameter space calculations.
The amplitude $\alpha $ of $r$-mode oscillations in a neutron star is related to the gravitational wave strain amplitude $h_{0}$ by [@OwenRModes]
$$\label{RModeForm}
\alpha = 0.028\left ( \frac{h_{0}}{10^{-24}} \right )\left ( \frac{r}{1 ~\textup{kpc}} \right )\left ( \frac{100 ~\textup{Hz}}{f} \right )^{3}.$$
Figure \[RMode\] uses this formula to convert the upper limits on $h_{0}$ to upper limits on the $r$-mode amplitude $\alpha $. The black curve represents the age-based limit on $\alpha $, under the same age assumptions used for $h_{0}$ and ellipticity, but with $n = 7$, which characterizes $r$-mode emission. In all three plots, the age-based limit is beaten everywhere the upper limits lie below the black curve.
Discussion {#s:disc}
==========
This search has placed the first explicit upper limits on continuous gravitational wave strength from the nearby globular cluster NGC 6544 for spin-down ages as young as 300 years, and is the first directed CW search for any globular cluster. The most stringent upper limits on strain ($h_{0}^{\textup{UL}}$) obtained were $h_{0}^{\textup{UL}} = 6.7 \times 10^{-25}$ for the 173-174 Hz band in the October data set, and $h_{0}^{\textup{UL}} = 6.0 \times 10^{-25}$ for the 170-171 Hz band in the July-August data.
The best upper limit is comparable to the best upper limit of $7 \times 10^{-25}$ at $150$ Hz obtained by the Cas A search [@S5CasA]; the recent search over nine supernova remnants, done without resampling, [@S6CasAFriends] set upper limits as low as $3.7 \times 10^{-25}$ for the supernova remnant G93.3+6.9, but used a coherence time of over 23 days (and a frequency band of only 264 Hz). The same analysis set a comparable 95% upper limit of $6.4 \times 10^{-25}$ for the supernova remnant G1.9+0.3, as expected given its similar declination and the search’s similar coherence time (9.2 days for NGC 6544 vs. 9.1 days for G1.9+0.3). Note, however, that the G1.9+0.3 search was limited to a 146 Hz search band, compared to 583 Hz for NGC 6544. The search reported here was carried out at substantially less computational cost because of barycentric resampling, and could thus search over a much larger parameter space.
The best upper limit on fiducial ellipticity, established using the July-August data set, was $\epsilon = 8.5 \times 10^{-6}$, for the $1$-Hz band starting at $670$ Hz. This is comparable to the best upper limit ($4 \times 10^{-5}$) obtained by the Cas A search [@Wette2009]; the supernova remnant search [@S6CasAFriends] set a comparable upper limit on fiducial ellipticity at $7.6 \times 10^{-5}$ for the supernova remnant G1.9+0.3.
These ellipticities are within the range of maximum theoretical ellipticities predicted for stars with some exotic phases in the core [@OwenStrangeStars; @OwenJ-McD2013], and the lowest of them is achievable for purely nucleonic stars with a sufficiently stiff equation of state and low mass [@OwenJ-McD2013]. Hence the search could have detected some exotic stars if they were supporting close to their maximum possible ellipticity; however, the lack of a detection cannot be used to infer constraints on the composition of any star, since the deformation could be much less than its maximum supportable value.
The first observing run of the Advanced LIGO detectors began in September 2015 [@Prospects] and the sensitivity of the detectors is already three times or more better than that used in this search, with an order of magnitude improvement over expected eventually [@AdvLIGO]. The barycentric resampling algorithm first implemented in this search is being used extensively in CW searches in the advanced detector era; it has been integrated into the search codes for both coherent searches (like the supernova remnant search) [@IdrisyThesis] and semi-coherent searches (Einstein@Home).
Acknowledgments
===============
This document has been assigned LIGO Laboratory document number `LIGO-P1500225`.
[^1]: The $N_{T}$ templates used in our searches are not completely independent, but can be represented by $N$ statistically independent templates where $N \approx 0.88N_{T}$. See section 8.7 of [@Wette2009].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an imaging and spectroscopic survey of galaxies in fields around QSOs HE0226$-$4110, PKS0405$-$123, and PG1216$+$069. The fields are selected to have ultraviolet echelle spectra available, which uncover 195 absorbers and 13 absorbers along the three sightlines. We obtain robust redshifts for 1104 galaxies of rest-frame absolute magnitude $M_R-5\log\,h{\lesssim}-16$ and at projected physical distances $\rho{\lesssim}4\ h^{-1}$ Mpc from the QSOs. HST/WFPC2 images of the fields around PKS0405$-$123 and PG1216$+$069 are available for studying the optical morphologies of absorbing galaxies. Combining the absorber and galaxy data, we perform a cross-correlation study to understand the physical origin of and absorbers and to constrain the properties of extended gas around galaxies. The results of our study are (1) both strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\ge 14$ and absorbers exhibit a comparable clustering amplitude as emission-line dominated galaxies and a factor of $\approx 6$ weaker amplitude than absorption-line dominated galaxies on co-moving projected distance scales of $r_p<3\ h^{-1}$ Mpc; (2) weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})<13.5$ appear to cluster very weakly around galaxies; (3) none of the absorption-line dominated galaxies at $r_p\le 250\ h^{-1}$ kpc has a corresponding absorber to a sensitive upper limit of $W(1031){\lesssim}0.03$ Å, while the covering fraction of absorbing gas around emission-line dominated galaxies is found to be $\kappa\approx 64$%; and (4) high-resolution images of five absorbing galaxies show that these galaxies exhibit disk-like morphologies with mildly disturbed features on the edge. Together, the data indicate that absorbers arise preferentially in gas-rich galaxies. In addition, tidal debris in groups/galaxy pairs may be principally responsible for the observed absorbers, particularly those of $W(1031)>70$ mÅ.'
author:
- 'Hsiao-Wen Chen and John S. Mulchaey'
title: |
PROBING THE IGM-GALAXY CONNECTION AT z $<$ 0.5 I :\
A GALAXY SURVEY IN QSO FIELDS AND A GALAXY-ABSORBER CROSS-CORRELATION STUDY$^{1,2}$
---
INTRODUCTION
============
Traditional galaxy surveys trace large-scale structures visible through stellar light or radio emission of cold neutral gas. Stars and known gaseous components account for roughly 12% of all baryons in the local universe and the rest are thought to reside in ionized gaseous halos around galaxies or in intergalactic space (e.g.Fukugita 2004). In principle, the diffuse halo gas and intergalactic medium (IGM) can be probed by the forest of absorption line systems observed in the spectra of background QSOs. Studies of QSO absorption-line systems are therefore expected to provide important constraints for theoretical models that characterize the growth of large scale structures.
Indeed, the forest is found to account for ${\gtrsim}95$% of the total baryons at redshift $z=2-3$ (e.g. Rauch [$\mbox{et~al.}$]{} 1997). In the nearby universe, Penton, Stocke, & Shull (2002, 2004) have argued that absorbers of neutral hydrogen column density $N({\mbox{${\rm H\,I}$}}) \le
10^{14.5}$ may contain between $20-30$% of the total baryons based on a simple assumption for the cloud geometry. Various numerical simulations have further suggested that approximately 40% of the total baryons reside in diffuse intergalactic gas of temperature $T < 10^{5-6}$ K (e.g. Davé [$\mbox{et~al.}$]{} 2001; Cen & Ostriker 2006). This temperature range makes high-ionization transitions, such as OVI$\lambda\lambda$1031,1037 and NeVIII$\lambda\lambda$770, 780, ideal tracers of this warm-hot intergalactic medium (see e.g. Verner [$\mbox{et~al.}$]{} 1994; Mulchaey [$\mbox{et~al.}$]{}1996; Tripp [$\mbox{et~al.}$]{} 2000). However, recent studies have shown that roughly 50% of OVI absorbers are better explained by photo-ionized gas of cooler temperatures (e.g. Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b). Therefore, the nature of these high-ionization systems does not appear to be unique.
Whether or not QSO absorption-line systems trace the typical galaxy population bears directly on the efforts to locate the missing baryons in the present-day universe (Persic & Salucci 1992; Fukugita, Hogan, & Peebles 1998), and to apply known statistical properties of the absorbers for constraining statistical properties of faint galaxies. This issue remains, however, unsettled (Lanzetta [$\mbox{et~al.}$]{} 1995; Stocke [$\mbox{et~al.}$]{} 1995; Chen [$\mbox{et~al.}$]{} 1998, 2001a; Penton [$\mbox{et~al.}$]{} 2002; Churchill [$\mbox{et~al.}$]{} 2007). While nearly all galaxies within a projected physical distance of $\rho=100\ h^{-1}$ kpc from a background QSO have associated absorbers produced by CIV$\lambda\lambda$1548,1550 and MgII$\lambda\lambda$2796,2803 (Chen [$\mbox{et~al.}$]{} 2001b; Chen & Tinker 2008) and nearly all galaxies within $\rho=180\ h^{-1}$ kpc have associated absorbers of $N({\mbox{${\rm H\,I}$}}) \ge 10^{14}$ (Chen [$\mbox{et~al.}$]{}1998, 2001a), not all absorbers have a galaxy found within 1 $h^{-1}$ Mpc physical distance to a luminosity limit of $0.5\, L_*$ (Morris [$\mbox{et~al.}$]{} 1993; Tripp [$\mbox{et~al.}$]{} 1998; Stocke [$\mbox{et~al.}$]{} 2006). It is not clear whether these absorbers are associated with fainter galaxies or not related to galaxies at all. In addition, the origin of OVI absorbers in terms of their galactic environment is also poorly quantified. While Stocke [$\mbox{et~al.}$]{} (2006) found 95% of OVI absorbers at $z<0.1$ are located at $\rho{\lesssim}560\ h^{-1}$ kpc from an $L_*$ galaxy, only three of the six OVI absorbers studied in Prochaska [$\mbox{et~al.}$]{} (2006) have an $L_*$ galaxy found at $\rho<1\ h^{-1}$ Mpc.
To understand the origin of QSO absorption-line systems, Chen [$\mbox{et~al.}$]{}(2005) presented a pilot study of the galaxy– absorber cross-correlation function based on the absorbers and galaxies identified along a single sightline toward PKS0405$-$123. The two-point cross-correlation analysis provides a quantitative measure of the origin of the absorbers based on their clustering amplitude. More massive systems arise in relatively higher overdensity environments and are expected to exhibit stronger clustering amplitude than low-mass objects arising in lower-overdensity environments. A primary result of Chen [$\mbox{et~al.}$]{} (2005) was that the cross-correlation function $\xi_{ga}$ including only emission-line dominated galaxies and strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\ge 14$ showed a comparable strength to the galaxy auto-correlation function $\xi_{gg}$ on co-moving, projection distance scales $r_p\le 1\,h^{-1}$ Mpc, while there remained a lack of cross-correlation signal when using only absorption-line dominated galaxies. Absorption-line dominated galaxies have red colers and are presumably evolved early-type galaxies, while emission-line dominated galaxies are blue and presumably younger star-forming systems. Early-type galaxies are found to cluster more strongly than younger, star-formig galaxies (e.g. Madgwick [$\mbox{et~al.}$]{} 2003; Zehavi [$\mbox{et~al.}$]{} 2005), and are therefore expected to reside in regions of higher matter overdensity. The comparable correlation amplitudes of emission-line galaxies and strong absorbers therefore suggested that strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\ge 14$ reside in the same overdensity regions as emission-line dominated galaxies. It appeared that these absorbers do not trace regions where more massive galaxies with dominant absorption spectral features reside. This result of Chen [$\mbox{et~al.}$]{} (2005) provided the first quantitative constraint on the origin of absorbers in terms of the significance of the underlying matter density fluctuations around the regions where they reside. It also offered a physical explanation for the on-average weaker clustering amplitude of absorbers relative to the clustering amplitude of the luminous galaxy population reported earlier (e.g. Morris [$\mbox{et~al.}$]{} 1993). Wilman [$\mbox{et~al.}$]{} (2007) performed a similar analysis based on a galaxy and absorber pair sample established over 16 QSO sightlines. While these authors concluded that a spectral-type dependent galaxy and cross-correlation function was not confirmed by their analysis, the two-dimensional cross-correlation function presented in Figure 4 of Wilman [$\mbox{et~al.}$]{}clearly displays a strong signal between absorbers and emission-line galaxies that is absent in the cross-correlation function measured using only absorption-line galaxies.
To examine whether the initial results of Chen [$\mbox{et~al.}$]{} (2005) obtained based on a single field are representative of the general absorber population, we have been conducting a deep, wide-area survey of galaxies in fields around QSOs at $z = 0.3 - 0.6$. The QSO fields are selected to have ultraviolet echelle spectra available from the Far Ultraviolet Spectroscopic Explorer (FUSE) and the Space Telescope Imaging Spectrograph (STIS) on board the Hubble Space Telescope (HST). The high-resolution UV spectra are necessary for finding intervening hydrogen and absorbers to form a statistically representative sample (see e.g. Thom & Chen 2008a; Tripp [$\mbox{et~al.}$]{} 2008), as well as identifying their associated metal-line transitions for constraining the ionization state of the gas. The galaxy sample from our survey program therefore also allows us to expand upon the initial absorber study to understanding the galactic environment of ionized gas traced by the OVI absorbers. The primary objectives of our galaxy survey program are (i) to examine the physical origin of and absorbers based on their respective clustering amplitudes, and (ii) to constrain the properties of extended gas around galaxies.
This paper is organized as follows. In Section 2, we describe the design of our galaxy survey program. In Section 3, we describe our observing program that includes both imaging observations, the selection of candidate galaxies, and multi-slit spectroscopic observations of three QSO fields. In Section 4, we summarize the results of the spectroscopic survey and examine the survey completeness. In Section 5, we summarize the properties of known absorption-line systems along individual QSO sightlines and present a new absorber catalog for the sightline toward PG1216$+$069. Descriptions of individual fields are presented in Section 6. Results of a galaxy–absorber cross-correlation study are presented in Section 7, and their implications are discussed in Section 8. Finally, we summary the main results of the paper in Section 9. We adopt a $\Lambda$CDM cosmology, $\Omega_{\rm M}=0.3$ and $\Omega_\Lambda =
0.7$, with a dimensionless Hubble constant $h = H_0/(100 \ {\rm km} \
{\rm s}^{-1}\ {\rm Mpc}^{-1})$ throughout the paper.
GOALS AND DESIGN OF THE GALAXY SURVEY
=====================================
The primary objectives of our galaxy survey program are (i) to examine the physical origin of and absorbers based on their clustering amplitudes, and (ii) to constrain the properties of extended gas around galaxies. To reach the goals, we have been conducting a wide-area survey of galaxies in fields around QSOs at $z
= 0.3 - 0.6$, for which echelle spectra at ultraviolet wavelengths are available from FUSE and HST/STIS. The high-resolution UV spectra are necessary for finding intervening hydrogen and absorbers to form a statistically representative sample (see e.g. Thom & Chen 2008a; Tripp [$\mbox{et~al.}$]{} 2008), as well as identifying their associated metal-line transitions for constraining the ionization state of the gas. The clustering amplitudes of and absorbers are then determined based on their cross-correlation signals with the galaxies identified in the wide-field redshift survey.
To characterize the cross-correlation function of galaxies and / absorbers, we aim to establish a statistically representative sample of $L_*$ galaxies on scales of $\approx 1-4\
h^{-1}$ co-moving Mpc from the QSO lines of sight over a redshift range of $z=0.1-0.5$. These luminous galaxies are thought to reside at the peaks of the underlying dark matter density fluctuations with a small fraction (${\lesssim}20$%) arising in group/cluster environments (Zheng [$\mbox{et~al.}$]{} 2007). Therefore, they offer a unique tracer of the large-scale matter overdensity in the distant universe. Under the $\Lambda$CDM paradigm, the relative clustering amplitudes of QSO absorption-line systems with respect to these galaxies provide quantitative measures of the overdensities of the regions where the absorbers reside (Mo & White 2002; Tinker [$\mbox{et~al.}$]{} 2005).
To examine the properties of extended gas around galaxies, we aim to establish a complete sample of sub-$L_*$ galaxies at $z=0.1-0.5$ within a projected co-moving distance of $r_p\approx 250\ h^{-1}$ kpc of the QSO lines of sight. The galaxy and absorber pair sample allows us to study the incidence and extent of ionized gas around galaxies of a range of luminosity and stellar content.
The short camera in the IMACS multi-object imaging spectrograph (Dressler [$\mbox{et~al.}$]{} 2006) on the Magellan Baade telescope contains eight CCDs and has a plate scale of $0.2''$. It observes a sky area of $\approx 15\arcmin$ radius, corresponding to a co-moving distance of $3.6\ h^{-1}$ Mpc at $z=0.3$. In a single setup, IMACS provides the necessary field coverage to efficiently carry out our galaxy survey program. To reach the scientific objectives outlined above, we have selected three QSOs, HE0226$-$4110 ($z_{\rm QSO}=0.495$), PKS0405$-$123 ($z_{\rm QSO}=0.573$), and PG1216$+$069 ($z_{\rm
QSO}=0.331$), all of which are accessible from the Las Campanas Observatory. Together, these QSO sightlines provide the largest redshift pathlength available in the STIS echelle sample for the absorber survey (see Figure 3 of Thom & Chen 2008a), and allow us to examine possible field to field variation in the properties of absorbers and galaxies.
We have targeted the spectroscopic survey to reach (1) $>80$% completeness for galaxies brighter than $R=22$ and at an angular radius less than $\Delta\,\theta = 2'$ from the QSOs and (2) ${\gtrsim}50$% completeness for galaxies brighter than $R=20$ and at $\Delta\,\theta>2'$. At $z=0.2$, $R=22$ corresponds to $\approx
0.04\,L_*$ galaxies and $\Delta\,\theta=2'$ corresponds to $\rho=280\
h^{-1}$ kpc. At $z=0.5$, $R=22$ corresponds to $\approx 0.3\,L_*$ galaxies and $\Delta\,\theta=2'$ corresponds to $\rho=510\ h^{-1}$ kpc. We note that the field around PKS0405$-$123 has been studied by Chen [$\mbox{et~al.}$]{} (2005) and Prochaska [$\mbox{et~al.}$]{} (2006). The previous survey targeted galaxies brighter than $R=20$, which was only sensitive to roughly $L_*$ galaxies at $z=0.5$. Our program has been designed to expand upon the earlier spectroscopic survey for measuring redshifts of fainter galaxies along the QSO line of sight.
[p[1.75 in]{}rcccl]{}\[h\] HE0226$-$4110 ($z_{\rm QSO}=0.495$) & IMACS & $B$ & 1200 & 0.8 & 2004-09-19 & IMACS & $R$ & 1200 & 0.7 & 2004-09-19 & IMACS & $I$ & 3000 & 0.7 & 2004-09-19 PKS0405$-$123 ($z_{\rm QSO}=0.573$) & HST/WFPC2 & F702W & 2400 & 0.1 & 1996-01-08 & HST/WFPC2 & F702W & 2100 & 0.1 & 1998-09-16 & HST/WFPC2 & F702W & 2100 & 0.1 & 1998-09-23 PG1216$+$069 ($z_{\rm QSO}=0.331$) & IMACS & $B$ & 1200 & 0.9 & 2005-03-19 & IMACS & $R$ & 1200 & 0.7 & 2005-03-19 & IMACS & $I$ & 3000 & 0.8 & 2005-03-19 & HST/WFPC2 & F702W & 2100 & 0.1 & 1998-03-30 & HST/WFPC2 & F702W & 2100 & 0.1 & 1999-04-18
OBSERVING PROGRAM
=================
The observing program of the galaxy survey consists of three phases: (1) imaging observations of the QSO fields to identify candidate galaxies, (2) object selection for follow-up faint galaxy spectroscopy, and (3) multi-slit spectroscopic observations to measure the redshifts of the selected galaxies. In this section, we describe each of the three phases.
Imaging Observations and Data Reduction
---------------------------------------
Imaging observations of the fields around HE0226$-$4110 and PG1216$+$069 were necessary in order to identify galaxies for follow-up spectroscopy[^1]. Faint galaxies in the field around PKS0405$-$123 have been published in Prochaska [$\mbox{et~al.}$]{} (2006), which are complete to $R=22.5$. We have therefore selected our spectroscopic targets of this field from this galaxy catalog.
Optical images of the field around HE0226$-$4110 were obtained on 2004 September 19, using the IMACS short camera and the $B$, $R$, and $I$ filters on the Magellan Baade Telescope. Optical images of the field around PG1216$+$069 were obtained on 2005 March 19, using the IMACS short camera and the $B$, $R$, and $I$ filters. To increase the efficiency in the spectroscopic follow-up, the multi-bandpass $B$, $R$, and $I$ photometric measurements were necessary for selecting candidate galaxies at $z<0.5$, excluding contaminating red stars and possible luminous high-redshift galaxies. The observations were carried out in a series of four to six exposures of between 300 and 500 s each, and were dithered by between 20 and 30 arcsec in space to cover the chip gaps between individual CCDs. All the imaging data were obtained under photometric conditions. We also observed standard star fields selected from Landolt (1992) through each of the $B$, $R$, and $I$ filters on each night in order to calibrate the photometry in our targeted fields. Flat-field images were taken both at a flat screen at the secondary mirror of the Magellan Baade telescope and at a blank sky during evening twilight. A journal of the IMACS imaging observations is presented in Table 1, which lists in columns (1) through (6) the object field, instrument, filter, total exposure time, mean FWHM, and Date of observations, respectively.
The reduction and processing of IMACS imaging data were complicated because of the multi-CCD format and because of a substantial geometric distortion across the field of the IMACS short camera. Before stacking individual exposures, we first formed a single, geometrically-dewarped frame of every exposure according to the following steps. First, we subtracted the readout bias of individual CCDs using the overscan regions of the chips. Next, the bias subtracted frames were flattened using a super sky flat formed from median-filtering unregistered science exposures. We found that in the $B$ and $R$ images a super sky flat works more effectively than either dome flats or twilight flats in removing the gain variations between individual pixels. For images obtained through the $I$ filter, however, the fringes in individual science exposures made it impossible to obtain an accurate flat-field image from the data. We therefore flattened the $I$-band exposures by first correcting the pixel-to-pixel variation using a median dome flat. The remaining fringes were then removed by subtracting a median $I$-band sky image formed from individual flattened science exposures.
Next we registered individual CCDs of each exposure using known stars found in the USNO A2.0 catalog. We formed a geometrically corrected mosaic frame of eight chips using the IRAF MSCRED package. The resulting mosaic frame contained calibrated astrometry with a typical r.m.s. scatter in stellar positions of $0.4''$. Finally, we combined individual mosaic frames to form a single stacked image of the QSO field for each of the $BRI$ bandpasses. The final stacked images covered a contiguous sky area of $\approx 28\times 28$ arcmin$^2$ and have mean point spread functions (PSFs) in the range of ${\rm FHWM}=0.7''-0.9''$ in the center of the IMACS field. Photometric calibrations were obtained using Landolt standard stars observed on the same night. The combined mosaic images reached 5 $\sigma$ limiting Vega magnitudes of $B=24.5$, $R=24.5$, and $I=23.5$. False color images of the central $4'\times 4'$ regions around HE0226$-$4110 and PG1216$+069$ are presented in Figures 1 and 2, respectively.
Additional optical images of PKS0405$-$123 and PG1216$+$069 obtained with HST using the Wide Field and Planetary Camera2 (WFPC2) with the F702W filter were accessed from the Hubble Space Telescope (HST) data archive. Individual exposures of each field were reduced using standard pipeline techniques, registered to a common origin, filtered for cosmic-ray events, and stacked to form a final combined image. These high spatial resolution images allowed us to resolve faint galaxies close to the QSO line of sight and to study their morphology in detail (see Figures 8 & 9 below). A journal of the HST/WFPC2 observations is presented in Table 1 as well.
Selection of Candidate Galaxies for Spectroscopic Follow-up
-----------------------------------------------------------
To select galaxies in the fields around HE0226$-$4110 and PG1216$+$069 for spectroscopic follow-up, we first identified objects in the stacked $R$-band mosaic image using the SExtractor program (Bertin & Arnout 1996). Next, we measured object magnitudes in the $B$, $R$, and $I$ images using the isophotal apertures of individual objects determined by SExtractor. This procedure yielded 948 and 1619 objects of $R=16-22$ within $\Delta\,\theta=11'$ angular radius from HE0226$-$4110 and PG1216$+$069, respectively. The difference in the object surface density underscores the significance of field to field variation. The sightline toward PG1216$+$069 is known to pass through the outskirts of the Virgo cluster at $z=0.004$ (and a projected physical distance $\rho\approx 1.4\ h^{-1}$ Mpc). As discussed below, the results of our spectroscopic survey have also uncovered multiple large-scale overdensities along the sightline that contribute to the high surface density of galaxies in this field.
Separating stars and galaxies was challenging in these images, because of non-negligible PSF distortions toward the edge of the IMACS field. To optimize the efficiency of the spectroscopic observations, we applied the observed $B-R$ versus $R-I$ colors to exclude likely contaminations from faint red stars and high-redshift galaxies. Figure 3 shows the $B-R$ vs. $R-I$ color distribution of the objects found in the field around HE0226$-$4110 (dots), together with stellar colors calculated using the stellar library compiled by Pickles (1998; star symbols) and simulated galaxy colors from $z=0$ (lower left) to $z=1$ (upper right) in steps of $\Delta\,z=0.1$ for E/S0 galaxies (filled circles), Sab (filled triangles), Scd (filled squares), and Irr (filled pentagon). The galaxy templates are adopted from Coleman [$\mbox{et~al.}$]{} (1980). The straight line indicates our color selection, above which objects are likely to arise at $z{\lesssim}0.7$ and represent candidate galaxies for follow-up spectroscopic observations with IMACS. This procedure yielded 776 and 1551 objects of $R=16 -
22$ and $B-R+ 1.7 > 3.5 (R-I)$ within $\Delta\,\theta=11'$ angular radius from HE0226$-$4110 and PG1216$+$069, respectively.
The goals of our galaxy survery, as outlined in § 2, are (1) to establish a statistically representative sample of $L_*$ galaxies for tracing the large-scale matter overdensities along the lines of sight toward the QSOs and (2) to identify a highly complete sample of sub-$L_*$ galaxies within projected co-moving distance $r_p\approx
250\ h^{-1}$ kpc from the QSO lines of sight for studying extended gas around individual galaxies and galaxy groups. In designing the multi-slit masks, we therefore aimed to include nearly all galaxies within $\Delta\,\theta=2'$ of the background QSO. Because of spectral collisions between closely located objects, this goal dictates the number of masks required to reach a high survey completeness. We were able to reach the goal in five IMACS masks and included 693 and 750 objects in the multi-slit spectroscopic observations of HE0226$-$4110 and PG1216$+$069, respectively[^2].
For the field around PKS0405$-$123, we extracted 524 galaxies within $\Delta\,\theta=11'$ of the QSO and brighter than $R=22$ from the photometric catalog of Prochaska [$\mbox{et~al.}$]{} (2006). These galaxies had not been observed in the previous survey. We aimed to complete the redshift survey of galaxies to $R\le 22$ in this QSO field with IMACS observations. We designed five IMACS masks to observe 450 objects in this field.
Spectroscopic Observations and Data Reduction
---------------------------------------------
Multi-slit spectroscopic observations of objects selected from the imaging program described in § 3.2 were carried out using IMACS and the short camera during three different runs in November 2004, May 2006, and April 2007. We were able to complete the observations of five masks for the field around HE0226$-$4110, four masks in the field around PKS0405$-$123, and five masks in the field around PG1216$+$069. Each mask contained between 103 to 218 slitlets of $1.2''$ in width. Typical slit lengths were five arcseconds for compact sources. For extended objects, we adopted the Kron radius estimated in SExtractor and expanded by a factor of three to allow accurate sky subtraction. We used the 200 l/mm grism that offers a spectral coverage of $\lambda=5000-9000$ Å with $\approx 2$ Åper pixel resolution. Observations of faint galaxies were carried out in a series of two to three exposures of between 1200 to 1800 s each. Observations of He-Ne-Ar lamps and a quartz lamp were performed every hour for wavelength and flat-field calibrations. A journal of the IMACS spectroscopic observations is presented in Table 2, which lists in columns (1) through (7) the object field, instrument set-up, spectral resolution (FWHM), mask number, number of objects observed per mask, total exposure time, and Date of observations, respectively.
Additional spectroscopic observations of 27 objects at $\Delta\,\theta<2'$ from HE0226$-$4110 and eight objects from PKS0405$-$123 were attempted in October 2007 and December 2008, using the Low Dispersion Survey Spectrograph 3 (LDSS3) on the Magellan Clay telescope. We included galaxies fainter than $R=22$ and clearly visible near the QSO sightlines in the LDSS3 observations, to further improve the survey depth in the immediate vicinity of the QSOs. LDSS3 observes a field of $8.3'$ diameter at a pixel scale of $0.189''$. We used the volume phase holographic (VPH) [*blue*]{} and VPH [*all*]{} grisms that cover a spectral range of $\lambda=4000-6500$ Å and $\lambda=6000-9000$ Å, respectively, and $1.0''\times 7''$ slitlets. The additional LDSS3 multi-slit spectroscopic observations allow us to reach 100% survey completeness of faint galaxies near the QSOs. A journal of the LDSS3 observations is included in Table 2 as well.
[p[1.75 in]{}lcrrcl]{} HE0226$-$4110 ($z_{\rm QSO}=0.495$) & IMACS/f2/200l & 12 & 1 & 187 & 3600 & 2004 November & IMACS/f2/200l & 12 & 2 & 168 & 3600 & 2004 November & IMACS/f2/200l & 12 & 3 & 119 & 3600 & 2004 November & IMACS/f2/200l & 12 & 4 & 116 & 2400 & 2004 November & IMACS/f2/200l & 12 & 5 & 103 & 3600 & 2004 November & LDSS3/VPH Blue & 4.3 & ... & 27 & 3600 & 2007 October & LDSS3/VPH All & 12 & ... & 27 & 3600 & 2008 December PKS0405$-$123 ($z_{\rm QSO}=0.573$) & IMACS/f2/200l & 12 & 1 & 159 & 3600 & 2004 November & IMACS/f2/200l & 12 & 2 & 141 & 4200 & 2004 November & IMACS/f2/200l & 12 & 3 & 134 & 2400 & 2004 November & IMACS/f2/200l & 12 & 5 & 48 & 2400 & 2004 November & LDSS3/VPH All & 12 & ... & 8 & 3600 & 2008 December PG1216$+$069 ($z_{\rm QSO}=0.331$) & IMACS/f2/200l & 12 & 1 & 218 & 3600 & 2006 May & IMACS/f2/200l & 12 & 2 & 199 & 7200 & 2007 April & IMACS/f2/200l & 12 & 3 & 182 & 3600 & 2006 May & IMACS/f2/200l & 12 & 4 & 142 & 5400 & 2007 April & IMACS/f2/200l & 12 & 5 & 151 & 3600 & 2006 May
The multi-slit spectroscopic data were processed and reduced using the Carnegie Observatories System for MultiObject Spectroscopy (COSMOS) program[^3] that has been developed and tested by A. Oemler, K. Clardy, D.Kelson, and G. Walth. The program adopts the optical model of the spectrograph and makes initial guesses both for the slit locations in the 2D spectral frame and for the wavelength calibrations of individual spectra. These initial gueses are then further refined using known spectral lines in the He-Ne-Ar frame obtained close in time with the science data. The rms scatter of the wavelength solution is typically a fraction of a pixel, i.e. ${\lesssim}1$ Å. Each raw frame of our spectral data was processed for bias subtraction and corrected for flat-fielding variations, following the standard procedure. Spectral extraction was performed using optimal weights determined according to the noise of relevant pixels. An error spectrum was generated from each raw frame, following noise counting statistics and the spectrum extraction procedure, and propagated through the pipeline. Finally, the extracted 1D spectra and their associated errors were flux-calibrated using a crude spectral sensitivity function calculated from the spectra of the alignment stars.
RESULTS OF THE SPECTROSCOPIC SURVEY
===================================
Redshift Measurements and Spectral Classification
-------------------------------------------------
The redshifts of objects were determined independently by each of us in two steps. First, we performed separate automatic $\chi^2$ fitting routines that determine a best-fit redshift based on cross-correlating the flux-calibrated object spectrum with an input model template. One of the cross-correlation routines was written by D. Kelson. This routine adopts SDSS spectral templates[^4] for early-type (SDSS template \# 24) and late-type galaxies (\#28) as input models, and determines a best-fit redshift based on matching absorption or emission line features. The other cross-correlation software was written by one of us (HWC), which employs four galaxy eigen spectra kindly provided by S. Burles. These eigen templates were obtained from analyzing $>100,000$ SDSS galaxy spectra, using a principal component analysis (see e.g. Yip [$\mbox{et~al.}$]{} 2004). The first two eigen spectra are characterized by predominantly absorption features and predominantly emission features, respectively. The last two eigen spectra offer additional modifications in the continuum slope. A linear combination of the four eigen function formed a model template to be compared with a flux-calibrated object spectrum, and a $\chi^2$ routine was performed to determine the best-fit linear coefficients and redshift for each object.
Next, the best-fit redshifts returned from the cross-correlation routines were then visually inspected by each of us to determine a final and robust redshift for every object. In some cases, the cross-correlation routines failed to identify a correct redshift due to contaminating residuals of bright skylines or imperfect background sky subtraction. We were able to recover the redshifts during this visual inspection process based on the presence of CaII H&K, or the presence of H$\alpha$ and \[NII\]. A comparison between redshifts determined independently by the two of us shows a typical scatter of $|\Delta\,z|\approx 0.0003$, corresponding to 1-$\sigma$ uncertainty of 70 [$\rm{km\,s}^{-1}$]{}.
The procedure described above led to robust redshift measurements for 432 and 448 galaxies in the fields around HE0226$-$4110 and PG1216$+$069, respectively. In addition, we measured 224 new redshifts for galaxies in the field around PKS0405$-$123, in comparison to 91 galaxy redshifts in the same area from Prochaska [$\mbox{et~al.}$]{} (2006). The remaining galaxies turned out to have insufficient $S/N$ in their spectra for a robust redshift measurement.
Spectral classification was guided by the relative magnitudes between the best-fit coefficients of the eigen spectra described above. These coefficients provide a quantitative evaluation of the fractional contributions from different spectral components in a galaxy, such as absorption spectra due to low-mass stars or emission spectra due to HII regions. The contribution due to absorption components may be determined by the best-fit coefficient of the first eigen spectrum ($C_1$). The contribution due to emission components may be determined by the best-fit coefficient of the second eigen spectrum, corrected for the continuum slope by subtracting off the best-fit coefficients of the remaining two eigen spectra ($C_2$). Figure 4 shows that despite a large scatter, there exists a linear distribution between emission $C_2$ and absorption $C_1$ components for 224 galaxies found in the field around PKS0405$-$123. We classified galaxies with $C_1>0.09$ and $C_2<0.3$ as absorption-line dominated early-type galaxies and the rest as emission-line dominated late-type galaxies.
Survey Completeness and Galaxy Redshift Distribution
----------------------------------------------------
The results of our spectroscopic survey program in three QSO fields are summarized in Figure 5. For the field around HE0226$-$4110, the open circles indicate the sky positions of all objects brighter than $R=23$ and the closed circles indicate the ones with available spectroscopic redshifts. The QSO is at ${\rm RA(J2000)}=37.0633$ deg and ${\rm Dec(J2000)}=-40.9544$ deg. For the fields around PKS0405$-$123 and PG1216$+$069, the open circles indicate the sky positions of all objects brighter than $R=22$ and the closed circles indicate the ones with available spectroscopic redshifts. The QSOs are at ${\rm
RA(J2000)}=61.9518$ deg and ${\rm Dec(J2000)}=-12.1935$ deg, and ${\rm
RA(J2000)}=184.8372$ deg and ${\rm Dec(J2000)}=6.6440$ deg, respectively.
We estimate the completeness of our survey by calculating the fraction of photometrically identified galaxies that have available spectroscopic redshifts. The calculations were performed for different minimum brightness in the $R$ band and at different angular distances $\Delta\theta$ to the QSO. The results are presented in the top panels of Figure 6. We find that our survey is most complete in the field around HE0226$-$4110, reaching 100% completeness for galaxies brighter than $R=23$ at angular distances $\Delta\theta\le
2'$. For PKS0405$-$123 and PG1216$+$069, we have also reached 100% completeness for galaxies brighter than $R=20$ (solid lines), $>80$% at $R=21$ (dotted lines) and $>60$% at $R=22$ (dash-dotted lines) in the inner $2'$ radius. The survey completeness becomes $\approx 50$% at larger radii.
The bottom panels of Figure 6 show the redshift distributions of the spectroscopically identified galaxies in the three fields. The solid curve in each panel represents the model expectation from convolving a non-evolving rest-frame $R$-band galaxy luminosity function of Blanton [$\mbox{et~al.}$]{} (2003), which is characterized by $M_{R_*}-5\,\log\,h=-20.4$, $\alpha=-1.1$, and $\phi_*=0.015\,h^{3}\,{\rm Mpc}^{-3}$, with the completeness functions displayed in the top panels. Redshift spikes indicate the presence of large-scale galaxy overdensities along the QSO lines of sight. The comparison between observations and model expectations in HE0226$-$4110 and PG1216$+$069 further demonstrates that the color selection criterion described in § 3.2 and Figure 3 has effectively excluded most galaxies at $z>0.6$.
CATALOGS OF QSO ABSORPTION-LINE SYSTEMS
=======================================
A necessary component of the galaxy–absorber cross-correlation analysis is an absorber catalog. A catalog of 57 absorbers along the sightline toward HE0226$-$4110 has been published by Lehner [$\mbox{et~al.}$]{} (2006). These absorbers exhibit a range of column density from $\log\,N({\mbox{${\rm H\,I}$}})=12.5$ to $\log\,N({\mbox{${\rm H\,I}$}})=15.1$ at $z_{{\mbox{${\rm Ly}\alpha$}}}=0.017-0.4$. We adopt their catalog in our cross-correlation analysis in this field. To summarize, the left panels of Figure 7 display the redshift distribution (top) and the cumulative $N({\mbox{${\rm H\,I}$}})$ distribution function (bottom) of these absorbers.
The sightline toward PKS0405$-$123 has been studied by Williger [$\mbox{et~al.}$]{} (2006) and Lehner [$\mbox{et~al.}$]{} (2007). We adopt the revised catalog of 76 absorbers from Lehner [$\mbox{et~al.}$]{} (2007) in our analysis. These absorbers exhibit a range of column density from $\log\,N({\mbox{${\rm H\,I}$}})=12.8$ to $\log\,N({\mbox{${\rm H\,I}$}})=16.3$ at $z_{{\mbox{${\rm Ly}\alpha$}}}=0.012-0.410$. The middle panels of Figure 7 display their redshift distribution (top) and cumulative $N({\mbox{${\rm H\,I}$}})$ distribution.
A systematic survey of absorbers for the sightline toward PG1216$+$069 has only been conducted for strong systems by Jannuzi [$\mbox{et~al.}$]{} (1998) in low-resolution (${\rm FWHM}\,\approx\,250$ [$\rm{km\,s}^{-1}$]{}) spectra obtained using the Faint Object Spectrograph. These authors identified nine strong absorbers with $\log\,N({\mbox{${\rm H\,I}$}})>14$. To complete the study of low-column density gas ($\log\,N({\mbox{${\rm H\,I}$}})<14$) along the sightline, we have conducted our own search of intervening lines in available echelle spectra of the QSO obtained by STIS.
Details regarding the reduction and processing of the STIS echelle spectra can be found in Thom & Chen (2008a). The final combined spectrum of PG1216$+$069 covers a spectral range of $\lambda=1160 -
1700$ Å with a spectral resolution of $\approx 6.8$ [$\rm{km\,s}^{-1}$]{} and has SNR of $\approx 7$ per resolution element in most of the spectral region. To establish a complete catalog of absorbers, we first identify absorption features that are at $>5$-$\sigma$ level of significance. Next, we identify known features due to either interstellar absorption of the Milky Way, or metal absorption lines and higher order Lyman series associated with known strong absorbers. We consider the remaining unidentified absorption lines as intervening absorbers. Finally, we perform a Voigt profile analysis using the VPFIT package[^5] for estimating the underlying $N({\mbox{${\rm H\,I}$}})$ and the Doppler parameter $b$ (e.g. Carswell [$\mbox{et~al.}$]{} 1991). This procedure has yielded 66 absorbers of column density $\log\,N({\mbox{${\rm H\,I}$}})=12.6-19.3$ at $z_{{\mbox{${\rm Ly}\alpha$}}}=0.003-0.321$. The results of our search are presented in Table 3, which for each absorber lists the absorber redshift, $N({\mbox{${\rm H\,I}$}})$ and the associated error, and $b$. The redshift distribution and cumulative $N({\mbox{${\rm H\,I}$}})$ distribution of these absorbers are presented in the right panels of Figure 7.
[lrrclrr]{} 0.00362 & $13.32\pm0.17$ & $49\pm24$ & | & 0.13503$^b$ & $14.41\pm0.04$ & $35\pm 2$ 0.00630$^{a,b}$ & $19.32\pm0.03$ & ... & | & 0.15475 & $13.07\pm0.09$ & $29\pm 7$ 0.01264 & $13.94\pm0.03$ & $40\pm 3$ & | & 0.15512 & $12.75\pm0.16$ & $24\pm 11$ 0.01493 & $13.07\pm0.10$ & $34\pm 9$ & | & 0.15576 & $12.77\pm0.13$ & $18\pm 7$ 0.01682 & $13.00\pm0.14$ & $59\pm 22$ & | & 0.15601 & $13.10\pm0.07$ & $14\pm 3$ 0.01752 & $12.62\pm0.18$ & $20\pm 10$ & | & 0.16076 & $13.14\pm0.10$ & $67\pm 19$ 0.02393 & $13.58\pm0.03$ & $26\pm 2$ & | & 0.17993$^b$ & $13.97\pm0.03$ & $34\pm 2$ 0.05725 & $12.87\pm0.12$ & $35\pm 12$ & | & 0.18026 & $13.23\pm0.10$ & $34\pm 9$ 0.06508 & $12.89\pm0.11$ & $25\pm 7$ & | & 0.18071 & $13.48\pm0.04$ & $33\pm 3$ 0.07381 & $12.83\pm0.12$ & $23\pm 8$ & | & 0.18698 & $13.26\pm0.06$ & $40\pm 6$ 0.07421 & $13.61\pm0.03$ & $35\pm 3$ & | & 0.19007 & $12.96\pm0.11$ & $23\pm 7$ 0.07606 & $13.02\pm0.08$ & $31\pm 7$ & | & 0.20040$^b$ & $13.98\pm0.03$ & $93\pm 7$ 0.08035 & $13.79\pm0.24$ & $45\pm 11$ & | & 0.20114 & $13.17\pm0.09$ & $23\pm 5$ 0.08053$^b$ & $13.99\pm0.15$ & $32\pm 4$ & | & 0.20442 & $12.94\pm0.12$ & $25\pm 8$ 0.08253 & $12.61\pm0.19$ & $16\pm 8$ & | & 0.22189 & $13.34\pm0.06$ & $42\pm 7$ 0.08278 & $12.83\pm0.16$ & $35\pm 16$ & | & 0.22578 & $12.99\pm0.13$ & $22\pm 9$ 0.08313 & $12.77\pm0.12$ & $16\pm 5$ & | & 0.22728 & $13.09\pm0.10$ & $32\pm 9$ 0.08348 & $13.22\pm0.07$ & $37\pm 7$ & | & 0.23035 & $12.67\pm0.18$ & $17\pm 9$ 0.08400 & $12.86\pm0.18$ & $64\pm 32$ & | & 0.23641 & $13.45\pm0.12$ & $43\pm 15$ 0.08813 & $13.22\pm0.07$ & $54\pm 10$ & | & 0.25035 & $12.93\pm0.14$ & $17\pm 6$ 0.09137 & $12.98\pm0.08$ & $30\pm 6$ & | & 0.26267 & $13.12\pm0.12$ & $26\pm 9$ 0.09415 & $13.07\pm0.10$ & $40\pm 12$ & | & 0.26624 & $12.93\pm0.14$ & $28\pm 10$ 0.09578 & $13.61\pm0.03$ & $39\pm 3$ & | & 0.26763$^b$ & $14.01\pm0.04$ & $32\pm 2$ 0.09727 & $12.72\pm0.20$ & $44\pm 26$ & | & 0.27154 & $12.81\pm0.12$ & $12\pm 4$ 0.09999 & $13.12\pm0.05$ & $20\pm 3$ & | & 0.27352 & $14.43\pm0.07$ & $22\pm 3$ 0.10234 & $13.00\pm0.15$ & $78\pm 32$ & | & 0.27865 & $12.98\pm0.15$ & $42\pm 17$ 0.10309 & $12.67\pm0.15$ & $15\pm 7$ & | & 0.28108 & $12.72\pm0.18$ & $16\pm 8$ 0.10440 & $13.09\pm0.08$ & $32\pm 6$ & | & 0.28226$^b$ & $16.34\pm0.10$ & $21\pm 2$ 0.12357 & $14.36\pm0.08$ & $29\pm 2$ & | & 0.28298 & $12.88\pm0.19$ & $21\pm 10$ 0.12388$^b$ & $14.53\pm0.14$ & $28\pm 3$ & | & 0.30394 & $13.40\pm0.09$ & $78\pm 20$ 0.12461$^b$ & $14.30\pm0.02$ & $53\pm 3$ & | & 0.31377 & $12.97\pm0.13$ & $47\pm 17$ 0.12493 & $14.06\pm0.07$ & $20\pm 2$ & | & 0.31857 & $13.23\pm0.11$ & $79\pm 24$ 0.12700 & $12.82\pm0.09$ & $20\pm 5$ & | & 0.32102 & $12.83\pm0.12$ & $21\pm 7$
For absorbers along the sightlines toward the three QSOs in our survey, we adopt the catalog of Thom & Chen (2008a,b) and include absorbers at $z<0.14$ from Tripp [$\mbox{et~al.}$]{} (2008). There are five absorbers known along the sightline toward HE0226$-$4110, six absorbers known along the sightline toward PKS0405$-$123 (see also Prochaska [$\mbox{et~al.}$]{} 2004), and two absorbers known along the sightline toward PG1216$+$069. The absorber at $z=0.1829$ along the sightline toward PKS0405$-$123 shows two distinct components separated by $\approx 87$ [$\rm{km\,s}^{-1}$]{} (see Thom & Chen 2008b; Prochaska [$\mbox{et~al.}$]{} 2004). The absorber at $z=0.1236$ along the sightline toward PG1216$+$069 also shows two dominant components well separated by $\approx 350$ [$\rm{km\,s}^{-1}$]{} (see § 6.3 and Figure 12 below). We consider these components separate objects in the cross-correlation study presented in § 7. The redshifts of these absorbers range from $z_{{\mbox{${\rm O\,VI}$}}}=0.017$ to $z_{{\mbox{${\rm O\,VI}$}}}=0.495$. The column densities of these absorbers range from $\log\,N({\mbox{${\rm O\,VI}$}})=13.4$ to $\log\,N({\mbox{${\rm O\,VI}$}})=14.7$. The redshift distribution of these absorbers is shown in shaded histograms in the top panels of Figure 7.
DESCRIPTION OF INDIVIDUAL FIELDS
================================
In this section, we review the galaxy and absorber properties in each individual field.
The Field toward HE0226$-$4110 at $z_{\rm QSO}=0.495$
-----------------------------------------------------
The sightline toward HE0226$-$4110 exhibits 57 absorbers (Lehner [$\mbox{et~al.}$]{} 2006) and five absorbers (Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b) over the redshift range from $z_{{\mbox{${\rm Ly}\alpha$}}}=0.017$ to $z_{{\mbox{${\rm Ly}\alpha$}}}=0.4$. The column density of these absorbers span a range from $\log\,N({\mbox{${\rm H\,I}$}})=12.5$ to $\log\,N({\mbox{${\rm H\,I}$}})=15.1$, and the column density of the absorbers span a range from $\log\,N({\mbox{${\rm O\,VI}$}})=13.6$ to $\log\,N({\mbox{${\rm O\,VI}$}})=14.4$. This field is particularly interesting because it exhibits a NeVIII$\lambda\lambda$770,780 absorber at $z=0.20701$ (Savage [$\mbox{et~al.}$]{} 2005), which is so far the only known detection of such high-ionization species in the IGM. The high ionization potential required to produce Ne$^{7+}$ (207.28 e.V.) and the relatively large cosmic abundance of Ne make the NeVIII doublet transitions a sensitive probe of warm-hot gas of $T=(0.5-1)\times
10^6$ K. Savage [$\mbox{et~al.}$]{} (2005) performed a detailed analysis of the ionization state of the gas, taking into account the relative abudances of additional highly ionized species such as SVI and OVI, and concluded that the observations are best explained by a collisional ionized gas of $T=5.4\times 10^5$ K. While the detection of NeVIII at $z=0.207$ provides exciting support for the presence of warm-hot gas, this system remains unique and continuing efforts to search for more NeVIII absorbers have uncovered no new systems (e.g. Lehner [$\mbox{et~al.}$]{} 2006).
Our galaxy survey is most complete in the field around HE0226$-$4110, reaching 100% completeness for galaxies brighter than $R=23$ at angular distances $\Delta\theta\le 2'$ (upper-left panel of Figure 6). The redshift distribution shown in the bottom-left panel of Figure 6 displays clear galaxy overdensities at $z=0.27$ and $z=0.4$ in front of the QSO. While there are absorbers present at these redshifts, no absorbers are found to coincide with these large-scale galaxy overdensities to the limit of $\log\,N({\mbox{${\rm O\,VI}$}})=13.5$.
Figure 8 shows the redshifts of galaxies in the inner $2.7'\times
2.7'$ field around the QSO. We have been able to obtain high $S/N$ spectra of galaxies at angular distance as close as $\Delta\theta{\lesssim}3''$ to the background bright QSO. Galaxies with redshift coincide with known OVI absorbers ($|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{}) are marked by a rectangular box. A complete photometric and spectroscopic catalog of galaxies with $R\le 23$ and at angular distance $\Delta\theta\le 11'$ of the QSO is available electronically at http://lambda.uchicago.edu/public/cat/cat\_0226.html. An example of the first 30 targets in the catalog is presented in Table 4, which lists from columns (1) through (13) the object ID, the right ascension (RA) and declination (Dec), the position offsets in RA ($\Delta\alpha$) and Dec ($\Delta\delta$) of the galaxy from the QSO, the angular distance to the QSO ($\Delta\theta$), the projected distance in physical $h^{-1}$ kpc, the $BRI$ magnitudes and uncertainties, the spectroscopic redshift $z_{\rm spec}$ ($-1$ indicates an absence of spectroscopic redshift measurement), spectral type, and rest-frame $R$-band absolute magnitude[^6].
Of the five OVI absorbers found along the sightline toward HE0226$-$4110, we have identified coincident galaxies at velocity offsets $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} and projected physical distances $\rho\le 250\ h^{-1}$ kpc for two absorbers, including the NeVIII absorber at $z=0.207$. No galaxies are found near the absorber at $z=0.01746$ (Lehner [$\mbox{et~al.}$]{} 2006; Tripp [$\mbox{et~al.}$]{} 2008), although our spectroscopic survey only covers $\rho<160\ h^{-1}$ physical kpc at this low redshift.
Three galaxies of $R=20.3-21.9$ are found within $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} and $\rho<200\ h^{-1}$ physical kpc of the and NeVIII absorber at $z=0.207$, all of which are sub-$L_*$ galaxies with rest-frame $R$-band absolute magnitudes spanning a range from $M_R-5\log\,h=-17.2$ to $M_R-5\log\,h=-18.9$. The survey completness rules out the presence of additional galaxies that are more luminous than $M_R-5\log\,h=-16.1$ within $\rho=285\ h^{-1}$ kpc. A detailed analysis of the galactic environment of the NeVIII absorber is presented in Mulchaey & Chen (2009).
The nearest galaxy to the absorber at $z=0.32629$ (Thom & Chen 2008a,b) has $R=23.0$ and $\rho=381\ h^{-1}$ physical kpc, with a corresponding rest-frame $R$-band absolute magnitude of $M_R-5\log\,h=-17.7$. The galaxy spectrum is dominated by absorption features. Three more galaxies are found at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{}from the absorber redshift, but they are over $\rho=840\ h^{-1}$ kpc to $\rho=2.6\ h^{-1}$ Mpc physical distances away. The survey completness rules out the presence of additional star-forming galaxies that are more luminous than $M_R-5\log\,h=-17.3$ within $\rho=396\
h^{-1}$ kpc.
The nearest galaxy to the absorber at $z=0.34034$ (Lehner [$\mbox{et~al.}$]{}2006; Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b) has $R=22.1$ and $\rho=213\ h^{-1}$ physical kpc, with a corresponding rest-frame $R$-band absolute magnitude of $M_R-5\log\,h=-18.3$. The galaxy spectrum is dominated by emission features. Two more galaxies are found at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} from the absorber redshift, and have $\rho=432\ h^{-1}$ kpc and $\rho=643\ h^{-1}$ kpc physical distances, respectively. The survey completness rules out the presence of additional galaxies that are more luminous than $M_R-5\log\,h=-17.4$ within $\rho=407\ h^{-1}$ kpc.
The nearest galaxy to the absorber at $z=0.35529$ (Lehner [$\mbox{et~al.}$]{}2006; Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b) has $R=22.0$ and $\rho=306\ h^{-1}$ physical kpc, with a corresponding rest-frame $R$-band absolute magnitude of $M_R-5\log\,h=-18.5$. The galaxy spectrum is dominated by emission features. Ten more galaxies are found at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} from the absorber redshift over a physical distance range of $\rho=387\ h^{-1}$ kpc to $\rho=1.8\
h^{-1}$ Mpc. The survey completness rules out the presence of additional galaxies that are more luminous than $M_R-5\log\,h=-17.5$ within $\rho=419\ h^{-1}$ kpc.
[lccrrrrcccrcr]{} 00443 & 02:29:21.173 & $-40$:54:57.72 & $ 743.2$ & $ 132.2$ & $ 754.9$ & $ -1.00$ & $ 22.740\pm 0.034$ & $ 20.199\pm 0.006$ & $ 18.865\pm 0.002$ & $ -1.0000$ & 0 & $ 0.00$ 00542 & 02:29:19.348 & $-40$:53:34.44 & $ 723.2$ & $ 215.3$ & $ 754.5$ & $ -1.00$ & $ 23.831\pm 0.066$ & $ 22.208\pm 0.024$ & $ 21.471\pm 0.021$ & $ -1.0000$ & 0 & $ 0.00$ 00551 & 02:29:19.107 & $-40$:52:55.99 & $ 720.7$ & $ 253.5$ & $ 764.0$ & $ -1.00$ & $ 23.595\pm 0.048$ & $ 22.316\pm 0.022$ & $ 21.853\pm 0.025$ & $ -1.0000$ & 0 & $ 0.00$ 00562 & 02:29:19.090 & $-40$:53:31.52 & $ 720.3$ & $ 218.2$ & $ 752.6$ & $ -1.00$ & $ 25.019\pm 0.200$ & $ 22.435\pm 0.031$ & $ 21.368\pm 0.020$ & $ -1.0000$ & 0 & $ 0.00$ 00578 & 02:29:19.162 & $-40$:56:50.16 & $ 719.9$ & $ 20.5$ & $ 720.2$ & $ -1.00$ & $ 22.480\pm 0.024$ & $ 20.453\pm 0.006$ & $ 19.700\pm 0.005$ & $ -1.0000$ & 0 & $ 0.00$ 00595 & 02:29:19.137 & $-40$:57:35.04 & $ 719.4$ & $ -24.1$ & $ 719.8$ & $ -1.00$ & $ 23.495\pm 0.066$ & $ 21.802\pm 0.023$ & $ 20.779\pm 0.015$ & $ -1.0000$ & 0 & $ 0.00$ 00604 & 02:29:18.789 & $-40$:57:29.05 & $ 715.5$ & $ -18.1$ & $ 715.7$ & $ -1.00$ & $ 24.723\pm 0.147$ & $ 22.970\pm 0.048$ & $ 22.043\pm 0.036$ & $ -1.0000$ & 0 & $ 0.00$ 00676 & 02:29:18.104 & $-40$:54:59.59 & $ 708.6$ & $ 130.6$ & $ 720.6$ & $ -1.00$ & $ 22.600\pm 0.035$ & $ 19.784\pm 0.004$ & $ 18.270\pm 0.002$ & $ -1.0000$ & 0 & $ 0.00$ 00685 & 02:29:17.937 & $-40$:54:45.17 & $ 706.8$ & $ 145.0$ & $ 721.6$ & $ 2422.79$ & $ 22.052\pm 0.021$ & $ 20.544\pm 0.009$ & $ 19.855\pm 0.008$ & $ 0.3347$ & 2 & $ -19.81$ 00686 & 02:29:18.299 & $-40$:55:16.06 & $ 710.7$ & $ 114.2$ & $ 719.9$ & $ 2838.03$ & $ 23.014\pm 0.047$ & $ 21.567\pm 0.021$ & $ 21.286\pm 0.021$ & $ 0.4329$ & 2 & $ -19.45$ 00693 & 02:29:17.587 & $-40$:57:16.51 & $ 702.0$ & $ -5.6$ & $ 702.0$ & $ -1.00$ & $ 25.241\pm 0.267$ & $ 22.480\pm 0.035$ & $ 21.038\pm 0.016$ & $ -1.0000$ & 0 & $ 0.00$ 00694 & 02:29:17.554 & $-40$:58:39.99 & $ 701.1$ & $ -88.6$ & $ 706.7$ & $ 2640.43$ & $ 23.287\pm 0.051$ & $ 21.406\pm 0.014$ & $ 20.639\pm 0.012$ & $ 0.3957$ & 1 & $ -19.87$ 00699 & 02:29:17.643 & $-40$:55:33.34 & $ 703.3$ & $ 97.1$ & $ 709.9$ & $ -1.00$ & $ 24.376\pm 0.160$ & $ 22.318\pm 0.038$ & $ 22.169\pm 0.047$ & $ -1.0000$ & 0 & $ 0.00$ 00731 & 02:29:17.061 & $-40$:56:26.82 & $ 696.4$ & $ 43.9$ & $ 697.8$ & $ -1.00$ & $ 24.646\pm 0.137$ & $ 22.940\pm 0.045$ & $ 22.294\pm 0.036$ & $ -1.0000$ & 0 & $ 0.00$ 00733 & 02:29:17.169 & $-40$:57:12.21 & $ 697.3$ & $ -1.2$ & $ 697.3$ & $ -1.00$ & $ 22.885\pm 0.041$ & $ 19.942\pm 0.005$ & $ 18.311\pm 0.002$ & $ -1.0000$ & 0 & $ 0.00$ 00734 & 02:29:16.844 & $-40$:55:31.43 & $ 694.3$ & $ 99.1$ & $ 701.3$ & $ -1.00$ & $ 24.860\pm 0.220$ & $ 22.249\pm 0.031$ & $ 21.780\pm 0.028$ & $ -1.0000$ & 0 & $ 0.00$ 00766 & 02:29:16.521 & $-40$:52:29.86 & $ 691.7$ & $ 279.7$ & $ 746.1$ & $ 1806.70$ & $ 21.877\pm 0.016$ & $ 19.504\pm 0.003$ & $ 18.689\pm 0.002$ & $ 0.2125$ & 1 & $ -20.00$ 00768 & 02:29:16.668 & $-40$:55:35.82 & $ 692.2$ & $ 94.7$ & $ 698.7$ & $ -1.00$ & $ 24.080\pm 0.137$ & $ 21.605\pm 0.021$ & $ 20.676\pm 0.012$ & $ -1.0000$ & 0 & $ 0.00$ 00808 & 02:29:15.823 & $-40$:58:33.06 & $ 681.7$ & $ -81.6$ & $ 686.6$ & $ -1.00$ & $ 24.985\pm 0.197$ & $ 21.960\pm 0.020$ & $ 20.525\pm 0.009$ & $ -1.0000$ & 0 & $ 0.00$ 00811 & 02:29:15.782 & $-40$:54:39.60 & $ 682.6$ & $ 150.7$ & $ 699.0$ & $ -1.00$ & $ 23.763\pm 0.060$ & $ 22.377\pm 0.029$ & $ 21.780\pm 0.028$ & $ -1.0000$ & 0 & $ 0.00$ 00820 & 02:29:15.782 & $-40$:59:06.36 & $ 681.0$ & $ -114.7$ & $ 690.6$ & $ 2935.54$ & $ 23.482\pm 0.060$ & $ 21.809\pm 0.020$ & $ 20.971\pm 0.016$ & $ 0.4951$ & 2 & $ -19.58$ 00830 & 02:29:15.386 & $-40$:52:03.47 & $ 679.0$ & $ 306.1$ & $ 744.8$ & $ -1.00$ & $ 24.118\pm 0.095$ & $ 22.542\pm 0.036$ & $ 21.686\pm 0.026$ & $ -1.0000$ & 0 & $ 0.00$ 00855 & 02:29:15.283 & $-40$:59:43.43 & $ 675.2$ & $ -151.6$ & $ 692.0$ & $ -1.00$ & $ 24.084\pm 0.095$ & $ 22.935\pm 0.049$ & $ 21.812\pm 0.029$ & $ -1.0000$ & 0 & $ 0.00$ 00861 & 02:29:14.925 & $-40$:57:10.49 & $ 672.1$ & $ 0.6$ & $ 672.1$ & $ -1.00$ & $ 24.847\pm 0.172$ & $ 22.933\pm 0.046$ & $ 21.701\pm 0.025$ & $ -1.0000$ & 0 & $ 0.00$ 00862 & 02:29:14.974 & $-40$:57:27.68 & $ 672.5$ & $ -16.5$ & $ 672.7$ & $ -1.00$ & $ 25.048\pm 0.182$ & $ 22.919\pm 0.042$ & $ 21.489\pm 0.018$ & $ -1.0000$ & 0 & $ 0.00$ 00863 & 02:29:14.971 & $-40$:52:05.68 & $ 674.3$ & $ 303.9$ & $ 739.7$ & $ -1.00$ & $ 24.107\pm 0.099$ & $ 22.633\pm 0.041$ & $ 22.033\pm 0.036$ & $ -1.0000$ & 0 & $ 0.00$ 00871 & 02:29:15.132 & $-40$:57:02.04 & $ 674.4$ & $ 9.0$ & $ 674.5$ & $ -1.00$ & $ 24.085\pm 0.081$ & $ 22.600\pm 0.034$ & $ 21.911\pm 0.029$ & $ -1.0000$ & 0 & $ 0.00$ 00875 & 02:29:15.616 & $-40$:59:51.01 & $ 678.9$ & $ -159.1$ & $ 697.3$ & $ -1.00$ & $ 22.372\pm 0.037$ & $ 20.852\pm 0.015$ & $ 20.062\pm 0.012$ & $ -1.0000$ & 0 & $ 0.00$ 00879 & 02:29:15.087 & $-40$:55:48.33 & $ 674.4$ & $ 82.4$ & $ 679.4$ & $ -1.00$ & $ 23.696\pm 0.076$ & $ 22.361\pm 0.035$ & $ 21.305\pm 0.022$ & $ -1.0000$ & 0 & $ 0.00$ 00885 & 02:29:14.503 & $-40$:52:28.29 & $ 668.9$ & $ 281.4$ & $ 725.7$ & $ -1.00$ & $ 25.539\pm 0.282$ & $ 22.831\pm 0.038$ & $ 21.314\pm 0.014$ & $ -1.0000$ & 0 & $ 0.00$
The Field toward PKS0405$-$123 at $z_{\rm QSO}=0.573$
-----------------------------------------------------
The sightline toward PKS0405$-$123 exhibits 76 absorbers (Lehner [$\mbox{et~al.}$]{} 2006) and six absorbers (Prochaska [$\mbox{et~al.}$]{} 2004; Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b) over the redshift range from $z_{{\mbox{${\rm Ly}\alpha$}}}=0.012$ to $z_{{\mbox{${\rm Ly}\alpha$}}}=0.538$. The column density of these absorbers span a range from $\log\,N({\mbox{${\rm H\,I}$}})=12.5$ to $\log\,N({\mbox{${\rm H\,I}$}})=16.3$, and the column density of the absorbers span a range from $\log\,N({\mbox{${\rm O\,VI}$}})=13.5$ to $\log\,N({\mbox{${\rm O\,VI}$}})=14.7$.
The field around PKS0405$-$123 has been surveyed by Prochaska [$\mbox{et~al.}$]{}(2006). This previous survey has yielded robust redshifts for 95% of all galaxies brighter than $R=20$, corresponding to roughly $L_*$ galaxies at $z=0.5$. Our program expands upon the earlier spectroscopic survey for uncovering fainter galaxies along the QSO line of sight, reaching 100% completeness at $R=20$ and $> 70$% at $R=22$ within $\Delta\theta\le 2'$ (upper-middle panel of Figure 6). The redshift distribution shown in the bottom-middle panel of Figure 6 displays numerous galaxy overdensities in front of the QSO, but only one absorber is found to coincide with the large-scale galaxy overdensity at $z=0.0966$.
Figure 9 displays a combined image of the field around PKS0405$-$123, covering roughly a $2.7'\times 2.7'$ area. Galaxies with known spectroscopic redshifts are indicated by their redshifts to the left. Blue values represent redshift measurements obtained by previous authors (Spinrad [$\mbox{et~al.}$]{} 1993; Prochaska [$\mbox{et~al.}$]{} 2006). Red values represent new redshifts obtained in our survey. Galaxies with redshift coincident with known OVI absorbers ($|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{}) are marked by a rectangular box. While this field contains a relatively high surface density of galaxies, the majority of the galaxies turn out to reside in the QSO host environment. A complete photometric and spectroscopic catalog of galaxies with $R\le 22$ and at angular distance $\Delta\theta\le 11'$ of the QSO is available electronically at http://lambda.uchicago.edu/public/cat/cat\_0405.html. An example of the first 30 targets in the catalog is presented in Table 5.
Of the six OVI absorbers found along the sightline toward PKS0405$-$123, four are identified with coincident galaxies at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} and $\rho\le 250\ h^{-1}$ kpc. For the absorber at $z=0.0918$ (Prochaska [$\mbox{et~al.}$]{} 2004), the closest galaxy found in previous surveys was at $\rho=306\ h^{-1}$ kpc with $R=19.7$ and $\Delta\,v=-341$ [$\rm{km\,s}^{-1}$]{} (Prochaska [$\mbox{et~al.}$]{} 2006). We have identified three new galaxies of $R=21.3-21.9$ at $|\Delta\,v|=27-275$ [$\rm{km\,s}^{-1}$]{} and $\rho=74-209 \ h^{-1}$ kpc from the absorber, including one that is observed in the combined HST/WFPC2 image (left panel of Figure 10). All three galaxies are faint dwarfs with emission-line dominated spectral features (see the top three panels of Figure 15 below). The corresponding rest-frame $R$-band absolute magnitudes span a range from $M_R-5\log\,h^{-1}=-15.4$ to $M_R-5\log\,h^{-1}=-16.1$. The HST image presented in Figure 10 shows that the dwarf galaxy at $\rho=73.5\ h^{-1}$ kpc exhibits a compact core with low-surface brightness emission extended to the north of the galaxy.
Only one galaxy is found within $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} and $\rho<200\
h^{-1}$ physical kpc of the absorber at $z=0.09658$ (Prochaska [$\mbox{et~al.}$]{} 2004). The galaxy has $R=19.0$, corresponding to $M_R-5\log\,h=-18.2$ (Prochaska [$\mbox{et~al.}$]{} 2006). Eight additional galaxies are found at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} from the absorber redshift and between $\rho=257\ h^{-1}$ kpc to $\rho=945\ h^{-1}$ kpc physical distances away.
The strong absorber of $\log\,N({\mbox{${\rm O\,VI}$}})=14.7$ at $z=0.1671$ has been studied extensively by Chen & Prochaska (2000) and Prochaska [$\mbox{et~al.}$]{} (2004). Two galaxies of $R=17.43$ and $R=21.0$ are found within $|\Delta\,v|<200$ [$\rm{km\,s}^{-1}$]{} and $\rho<100\ h^{-1}$ physical kpc of the absorber at $z=0.1670$ (see the discussions in Chen & Prochaska 2000 and Prochaska [$\mbox{et~al.}$]{} 2006). No additional galaxies are found within $\rho=1\ h^{-1}$ Mpc of the absorber. The luminous galaxy at $\Delta\,\delta=40''$ (corresponding to $\rho=80.9\ h^{-1}$ kpc) from the QSO has $M_R-5\log\,h=-20.9$ and exhibits post starburst spectral features (Prochaska [$\mbox{et~al.}$]{} 2006). The WFPC2 image presented in Figure 9 covers only part of the galaxy. A detailed examination of the WFPC2 image shows regular spiral structures with several compact HII regions along the spiral arms (center-left panel of Figure 10). The dwarf galaxy at $\rho=67.8\ h^{-1}$ kpc has $M_R-5\log\,h=-18.1$. The disk-like morphology appears to be mildly distburbed by a faint companion at the northeast edge (center-right panel of Figure 10).
No galaxies are found near the absorber at $z=0.1829$ (Prochaska [$\mbox{et~al.}$]{} 2004; Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b). The high completness of our survey to $R=22$ within $\Delta\theta=2'$ suggests that the galaxies associated with the absorber is likely fainter than $M_R-5\log\,h=-16.6$.
No galaxies are found near the absorber at $z=0.3633$ (Prochaska [$\mbox{et~al.}$]{} 2004; Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b). The high completness of our survey to $R=22$ within $\Delta\theta=2'$ suggests that the galaxies associated with the absorber is likely fainter than $M_R-5\log\,h=-18.5$.
The strong absorber of $\log\,N({\mbox{${\rm O\,VI}$}})=14.5$ at $z=0.4951$ has been studied extensively by Howk [$\mbox{et~al.}$]{} (2009). The early search by Prochaska [$\mbox{et~al.}$]{} (2006) did not uncover any galaxies within $\rho=2.6\
h^{-1}$ Mpc. Our spectroscopic survey has uncovered a galaxy of $R=22.63$ at $\Delta\,v=-181$ [$\rm{km\,s}^{-1}$]{} and $\rho=77\ h^{-1}$ physical kpc from the absorber, with a corresponding rest-frame $R$-band absolute magnitude of $M_R-5\log\,h=-18.8$. The spectrum is dominated by emission-line features due to \[OII\], H$\gamma$, H$\beta$, and \[OIII\] (see the 4th panel in Figure 15 below). The morphology appears to be extended in the HST image in Figure 10 (right panel). No additional galaxies are found at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} and $\rho<1\
h^{-1}$ Mpc.
[lccrrrrcrcr]{} 00071 & 04:08:33.10 & $-12$:10:48.0 & $ 655.0$ & $ 48.7$ & 656.8 & $ -1.00$ & $ 21.340\pm 0.110$ & -1.0000 & 0 & $ 0.00$ 00078$^c$ & 04:08:32.80 & $-12$:13: 9.0 & $ 650.5$ & $ -92.3$ & 657.0 & $ 2980.30$ & $ 20.510\pm 0.090$ & 0.5639 & 1 & $ -22.03$ 00089 & 04:08:32.50 & $-12$:13:42.0 & $ 646.0$ & $ -125.3$ & 658.1 & $ -1.00$ & $ 20.470\pm 0.080$ & -1.0000 & 0 & $ 0.00$ 00107 & 04:08:32.10 & $-12$:14:17.0 & $ 640.2$ & $ -160.3$ & 659.9 & $ -1.00$ & $ 19.740\pm 0.070$ & -1.0000 & 0 & $ 0.00$ 00108 & 04:08:32.30 & $-12$:11:06.0 & $ 643.2$ & $ 30.7$ & 643.9 & $ -1.00$ & $ 21.230\pm 0.100$ & -1.0000 & 0 & $ 0.00$ 00114 & 04:08:32.20 & $-12$:10:39.0 & $ 641.8$ & $ 57.7$ & 644.3 & $ -1.00$ & $ 21.290\pm 0.090$ & -1.0000 & 0 & $ 0.00$ 00120 & 04:08:32.00 & $-12$:09:42.0 & $ 638.9$ & $ 114.7$ & 649.1 & $ -1.00$ & $ 16.970\pm 0.060$ & -1.0000 & 0 & $ 0.00$ 00123 & 04:08:31.80 & $-12$:12:32.0 & $ 635.8$ & $ -55.3$ & 638.2 & $ -1.00$ & $ 18.090\pm 0.060$ & -1.0000 & 0 & $ 0.00$ 00129 & 04:08:31.70 & $-12$:14:01.0 & $ 634.3$ & $ -144.3$ & 650.5 & $ -1.00$ & $ 20.810\pm 0.090$ & -1.0000 & 0 & $ 0.00$ 00138 & 04:08:31.40 & $-12$:10:44.0 & $ 630.0$ & $ 52.7$ & 632.2 & $ -1.00$ & $ 17.290\pm 0.060$ & -1.0000 & 0 & $ 0.00$ 00142$^c$ & 04:08:32.00 & $-12$:14:10.0 & $ 638.7$ & $ -153.3$ & 656.8 & $ 2981.03$ & $ 20.890\pm 0.090$ & 0.5645 & 2 & $ -20.86$ 00144 & 04:08:31.20 & $-12$:13:59.0 & $ 627.0$ & $ -142.3$ & 642.9 & $ -1.00$ & $ 21.320\pm 0.090$ & -1.0000 & 0 & $ 0.00$ 00145 & 04:08:31.20 & $-12$:11:14.0 & $ 627.1$ & $ 22.7$ & 627.5 & $ -1.00$ & $ 17.570\pm 0.060$ & -1.0000 & 0 & $ 0.00$ 00149 & 04:08:31.00 & $-12$:14:55.0 & $ 624.0$ & $ -198.3$ & 654.8 & $ -1.00$ & $ 21.580\pm 0.120$ & -1.0000 & 0 & $ 0.00$ 00153 & 04:08:31.00 & $-12$:13:02.0 & $ 624.1$ & $ -85.3$ & 629.9 & $ -1.00$ & $ 21.600\pm 0.110$ & -1.0000 & 0 & $ 0.00$ 00162 & 04:08:31.00 & $-12$:11:01.0 & $ 624.2$ & $ 35.7$ & 625.2 & $ -1.00$ & $ 21.890\pm 0.130$ & -1.0000 & 0 & $ 0.00$ 00163 & 04:08:30.70 & $-12$:13:09.0 & $ 619.7$ & $ -92.3$ & 626.5 & $ -1.00$ & $ 21.370\pm 0.100$ & -1.0000 & 0 & $ 0.00$ 00165 & 04:08:30.80 & $-12$:11:30.0 & $ 621.2$ & $ 6.7$ & 621.2 & $ -1.00$ & $ 17.600\pm 0.060$ & -1.0000 & 0 & $ 0.00$ 00169 & 04:08:30.30 & $-12$:15:00.0 & $ 613.7$ & $ -203.3$ & 646.5 & $ -1.00$ & $ 16.240\pm 0.060$ & -1.0000 & 0 & $ 0.00$ 00174 & 04:08:30.60 & $-12$:10:56.0 & $ 618.3$ & $ 40.7$ & 619.6 & $ -1.00$ & $ 19.140\pm 0.070$ & -1.0000 & 0 & $ 0.00$ 00175 & 04:08:30.80 & $-12$:08:27.0 & $ 621.3$ & $ 189.7$ & 649.6 & $ -1.00$ & $ 21.220\pm 0.090$ & -1.0000 & 0 & $ 0.00$ 00179 & 04:08:30.60 & $-12$:11:14.0 & $ 618.3$ & $ 22.7$ & 618.7 & $ -1.00$ & $ 21.590\pm 0.110$ & -1.0000 & 0 & $ 0.00$ 00190 & 04:08:30.60 & $-12$:08:09.0 & $ 618.4$ & $ 207.7$ & 652.3 & $ -1.00$ & $ 21.760\pm 0.120$ & -1.0000 & 0 & $ 0.00$ 00191 & 04:08:30.00 & $-12$:13:17.0 & $ 609.4$ & $ -100.3$ & 617.6 & $ -1.00$ & $ 21.980\pm 0.130$ & -1.0000 & 0 & $ 0.00$ 00203 & 04:08:29.60 & $-12$:13:26.0 & $ 603.5$ & $ -109.3$ & 613.4 & $ -1.00$ & $ 20.830\pm 0.080$ & -1.0000 & 0 & $ 0.00$ 00209 & 04:08:29.80 & $-12$:09:00.0 & $ 606.6$ & $ 156.7$ & 626.5 & $ -1.00$ & $ 19.400\pm 0.070$ & -1.0000 & 0 & $ 0.00$ 00210 & 04:08:29.90 & $-12$:08:04.0 & $ 608.1$ & $ 212.7$ & 644.2 & $ -1.00$ & $ 20.800\pm 0.090$ & -1.0000 & 0 & $ 0.00$ 00213 & 04:08:29.40 & $-12$:15:16.0 & $ 600.5$ & $ -219.3$ & 639.3 & $ -1.00$ & $ 21.170\pm 0.100$ & -1.0000 & 0 & $ 0.00$ 00214 & 04:08:29.50 & $-12$:11:19.0 & $ 602.2$ & $ 17.7$ & 602.4 & $ -1.00$ & $ 21.560\pm 0.110$ & -1.0000 & 0 & $ 0.00$ 00221$^d$ & 04:08:29.00 & $-12$:13:28.0 & $ 594.7$ & $ -111.3$ & 605.1 & $ 2079.39$ & $ 19.400\pm 0.070$ & 0.3466 & 1 & $ -20.41$
The Field toward PG1216$+$069 at $z_{\rm QSO}=0.3313$
-----------------------------------------------------
The search of absorbers described in § 5 has identified 66 absorbers along the sightline toward PG1216$+$069 over the redshift range from $z_{{\mbox{${\rm Ly}\alpha$}}}=0.0036$ to $z_{{\mbox{${\rm Ly}\alpha$}}}=0.321$ (Table 3). The column density of these absorbers span a range from $\log\,N({\mbox{${\rm H\,I}$}})=12.6$ to $\log\,N({\mbox{${\rm H\,I}$}})=19.3$. Two absorbers are found at $z=0.1242$ and $z=0.2823$ with $\log\,N({\mbox{${\rm O\,VI}$}})=14.7$ and $\log\,N({\mbox{${\rm O\,VI}$}})=13.4$, respectively (Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a,b)[^7].
Our galaxy survey in the field around PG1216$+$069 in the least incomplete of all three, reaching $>80$% completeness for galaxies brighter than $R=21$ at angular distances $\Delta\theta\le 2'$ (upper-right panel of Figure 6). The redshift distribution shown in the bottom-right panel of Figure 6 displays a significant galaxy overdensity at $z=0.26-0.28$ in front of the QSO. One of the two absorbers does coincide with this large-scale galaxy overdensity. Figure 11 shows the redshifts of galaxies in the inner $2.7'\times 2.7'$ field around the QSO. We have not been able to obtain high quality spectra of galaxies at angular distance $\Delta\theta{\lesssim}5''$ due to the presence of a bright star and the background QSO. Only one foreground galaxy is found at $\rho<200\
h^{-1}$ kpc. A complete photometric and spectroscopic catalog of galaxies with $R\le 22$ and at angular distance $\Delta\theta\le 11'$ of the QSO is available electronically at http://lambda.uchicago.edu/public/cat/cat\_1216.html. An example of the first 30 targets in the catalog is presented in Table 6.
[lccrrrrcccrcr]{} 00539 & 12:18:48.159 & $+$06:32:41.11 & $ -484.7$ & $ -358.0$ & $ 602.6$ & $ -1.00$ & $ 18.723\pm 0.023$ & $ 16.923\pm 0.005$ & $ 16.654\pm 0.005$ & $-1.0000$ & 0 & $ 0.00$ 01375 & 12:19:20.821 & $+$06:27:37.63 & $ 0.9$ & $ -657.5$ & $ 657.5$ & $ 2023.15$ & $ 21.500\pm 0.014$ & $ 20.237\pm 0.005$ & $ 19.765\pm 0.006$ & 0.2943 & 2 & $ -19.79$ 01419 & 12:19:21.951 & $+$06:27:38.90 & $ 17.7$ & $ -656.2$ & $ 656.4$ & $ -1.00$ & $ 21.148\pm 0.012$ & $ 18.744\pm 0.002$ & $ 17.645\pm 0.001$ & $-1.0000$ & 0 & $ 0.00$ 01454 & 12:19:21.826 & $+$06:27:48.01 & $ 15.8$ & $ -647.1$ & $ 647.3$ & $ -1.00$ & $ 22.316\pm 0.019$ & $ 21.763\pm 0.015$ & $ 21.491\pm 0.021$ & $-1.0000$ & 0 & $ 0.00$ 01567 & 12:19:14.802 & $+$06:27:54.42 & $ -88.4$ & $ -641.2$ & $ 647.3$ & $ 2551.46$ & $ 22.493\pm 0.026$ & $ 21.123\pm 0.010$ & $ 20.820\pm 0.012$ & 0.4328 & 2 & $ -19.89$ 01591 & 12:19:13.861 & $+$06:27:58.57 & $ -102.4$ & $ -637.2$ & $ 645.3$ & $ -1.00$ & $ 23.090\pm 0.046$ & $ 20.541\pm 0.006$ & $ 19.378\pm 0.004$ & $-1.0000$ & 0 & $ 0.00$ 01610 & 12:19:27.817 & $+$06:27:50.45 & $ 104.6$ & $ -644.3$ & $ 652.7$ & $ 1357.74$ & $ 20.959\pm 0.015$ & $ 19.128\pm 0.004$ & $ 18.510\pm 0.004$ & 0.1753 & 2 & $ -19.67$ 01619 & 12:19:10.432 & $+$06:27:56.42 & $ -153.2$ & $ -639.5$ & $ 657.6$ & $ -1.00$ & $ 22.830\pm 0.040$ & $ 21.748\pm 0.019$ & $ 21.337\pm 0.024$ & $-1.0000$ & 0 & $ 0.00$ 01621 & 12:19:24.357 & $+$06:27:49.33 & $ 53.3$ & $ -645.6$ & $ 647.8$ & $ 2044.82$ & $ 21.010\pm 0.011$ & $ 19.615\pm 0.004$ & $ 19.123\pm 0.004$ & 0.3054 & 2 & $ -20.50$ 01693 & 12:19:28.263 & $+$06:27:56.16 & $ 111.2$ & $ -638.6$ & $ 648.2$ & $ -1.00$ & $ 22.168\pm 0.038$ & $ 20.307\pm 0.009$ & $ 19.929\pm 0.012$ & $-1.0000$ & 0 & $ 0.00$ 01740 & 12:19:27.316 & $+$06:27:58.82 & $ 97.1$ & $ -636.0$ & $ 643.4$ & $ -1.00$ & $ 18.688\pm 0.002$ & $ 17.676\pm 0.001$ & $ 17.384\pm 0.001$ & $-1.0000$ & 0 & $ 0.00$ 01785 & 12:19:26.316 & $+$06:28:10.08 & $ 82.3$ & $ -624.9$ & $ 630.2$ & $ -1.00$ & $ 22.555\pm 0.034$ & $ 20.485\pm 0.006$ & $ 19.921\pm 0.005$ & $-1.0000$ & 0 & $ 0.00$ 01794 & 12:19:12.163 & $+$06:28:13.40 & $ -127.6$ & $ -622.5$ & $ 635.5$ & $ -1.00$ & $ 22.901\pm 0.029$ & $ 21.660\pm 0.013$ & $ 21.428\pm 0.020$ & $-1.0000$ & 0 & $ 0.00$ 01814 & 12:19:23.459 & $+$06:27:58.51 & $ 40.0$ & $ -636.6$ & $ 637.8$ & $ 1979.71$ & $ 21.963\pm 0.028$ & $ 20.423\pm 0.009$ & $ 19.919\pm 0.010$ & 0.2981 & 2 & $ -19.64$ 01819 & 12:19:18.003 & $+$06:28:07.61 & $ -41.0$ & $ -627.9$ & $ 629.2$ & $ 1953.51$ & $ 22.342\pm 0.034$ & $ 20.752\pm 0.010$ & $ 20.270\pm 0.012$ & 0.2981 & 2 & $ -19.31$ 01852 & 12:19:15.205 & $+$06:28:10.43 & $ -82.5$ & $ -625.3$ & $ 630.7$ & $ 2756.59$ & $ 23.582\pm 0.111$ & $ 20.399\pm 0.008$ & $ 19.560\pm 0.006$ & 0.5220 & 1 & $ -21.84$ 01875 & 12:19:25.887 & $+$06:28:17.02 & $ 75.9$ & $ -618.0$ & $ 622.6$ & $ -1.00$ & $ 22.975\pm 0.035$ & $ 21.821\pm 0.017$ & $ 21.236\pm 0.016$ & $-1.0000$ & 0 & $ 0.00$ 01897 & 12:19:32.437 & $+$06:28:15.65 & $ 173.0$ & $ -618.9$ & $ 642.6$ & $ -1.00$ & $ 22.797\pm 0.046$ & $ 21.590\pm 0.020$ & $ 21.289\pm 0.024$ & $-1.0000$ & 0 & $ 0.00$ 01898 & 12:19:09.029 & $+$06:28:19.40 & $ -174.1$ & $ -616.8$ & $ 640.9$ & $ -1.00$ & $ 23.362\pm 0.065$ & $ 21.496\pm 0.015$ & $ 20.783\pm 0.015$ & $-1.0000$ & 0 & $ 0.00$ 01915 & 12:19:33.091 & $+$06:28:20.09 & $ 182.7$ & $ -614.4$ & $ 641.0$ & $ -1.00$ & $ 23.848\pm 0.101$ & $ 21.792\pm 0.019$ & $ 21.147\pm 0.017$ & $-1.0000$ & 0 & $ 0.00$ 01921 & 12:19:13.717 & $+$06:28:24.51 & $ -104.6$ & $ -611.4$ & $ 620.3$ & $ -1.00$ & $ 22.682\pm 0.022$ & $ 21.873\pm 0.015$ & $ 21.701\pm 0.025$ & $-1.0000$ & 0 & $ 0.00$ 01972 & 12:19:17.545 & $+$06:28:20.58 & $ -47.8$ & $ -615.0$ & $ 616.9$ & $ 1746.19$ & $ 22.692\pm 0.045$ & $ 21.316\pm 0.016$ & $ 20.763\pm 0.020$ & 0.2618 & 2 & $ -18.43$ 01996 & 12:19:35.051 & $+$06:28:26.57 & $ 211.7$ & $ -607.8$ & $ 643.6$ & $ -1.00$ & $ 22.038\pm 0.018$ & $ 21.173\pm 0.010$ & $ 20.940\pm 0.014$ & $-1.0000$ & 0 & $ 0.00$ 02007 & 12:19:19.594 & $+$06:28:27.37 & $ -17.5$ & $ -608.1$ & $ 608.4$ & $ -1.00$ & $ 22.915\pm 0.042$ & $ 19.955\pm 0.004$ & $ 18.404\pm 0.002$ & $-1.0000$ & 0 & $ 0.00$ 02025 & 12:19:16.440 & $+$06:28:31.24 & $ -64.3$ & $ -604.5$ & $ 607.9$ & $ 2661.45$ & $ 23.481\pm 0.064$ & $ 21.396\pm 0.014$ & $ 20.673\pm 0.012$ & 0.5237 & 2 & $ -20.14$ 02028 & 12:19:13.644 & $+$06:28:23.06 & $ -105.7$ & $ -612.8$ & $ 621.9$ & $ -1.00$ & $ 21.271\pm 0.016$ & $ 20.161\pm 0.008$ & $ 19.790\pm 0.010$ & $-1.0000$ & 0 & $ 0.00$ 02044 & 12:19:34.051 & $+$06:28:29.44 & $ 196.9$ & $ -605.1$ & $ 636.3$ & $ -1.00$ & $ 23.178\pm 0.052$ & $ 21.980\pm 0.022$ & $ 21.772\pm 0.031$ & $-1.0000$ & 0 & $ 0.00$ 02048 & 12:19:17.171 & $+$06:28:25.84 & $ -53.4$ & $ -609.8$ & $ 612.2$ & $ -1.00$ & $ 23.015\pm 0.061$ & $ 21.382\pm 0.018$ & $ 20.722\pm 0.018$ & $-1.0000$ & 0 & $ 0.00$ 02108 & 12:19:37.598 & $+$06:28:37.06 & $ 249.4$ & $ -597.2$ & $ 647.2$ & $ -1.00$ & $ 23.376\pm 0.058$ & $ 21.848\pm 0.020$ & $ 21.412\pm 0.021$ & $-1.0000$ & 0 & $ 0.00$ 02111 & 12:19:13.945 & $+$06:28:38.92 & $ -101.3$ & $ -597.0$ & $ 605.6$ & $ -1.00$ & $ 21.601\pm 0.011$ & $ 20.708\pm 0.006$ & $ 20.529\pm 0.010$ & $-1.0000$ & 0 & $ 0.00$
A galaxy is identified at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} and $\rho\le 250\
h^{-1}$ kpc from the absorber at $z=0.1242$ (see Chen [$\mbox{et~al.}$]{} 2001a). The galaxy has $R=19.3$, corresponding to $M_R-5\log\,h=-20.0$, and $\rho=64\ h^{-1}$ kpc. The morphology of the galaxy is indicative of a galaxy (at the southern edge) being tidally torn by a more massive galaxy during the process of merging (top-left panel of Figure 12). The spectrum of the system shown in the bottom left panel of Figure 12 exhibits prominent emission features, such as H$\alpha$, \[NII\], and \[SII\], suggesting that the ISM is chemically enriched to solar metallicity. The large H$\alpha$ to H$\beta$ line ratio also suggests the presence of substantial dust in the ISM.
The absorber at $z=0.1242$ exhibits a complex absorption profile. We show in the right panel of Figure 12 the absorption features identified in the QSO spectra obtained using both HST/STIS and FUSE. Reduction and processing of the STIS spectra are described in Thom & Chen (2008). Reduction and processing of the FUSE spectra are described in Scott [$\mbox{et~al.}$]{} (2004). The combined STIS spectrum has a spectral resolution of $\delta\,v\approx 6.8$ [$\rm{km\,s}^{-1}$]{}; the combined FUSE spectrum has a spectral resolution of $\delta\,v\approx 30$ [$\rm{km\,s}^{-1}$]{}.
At least four absorption components ($\Delta\,v=-158, -82, +120$ and $+188$ [$\rm{km\,s}^{-1}$]{}) are seen in the HI, CIII, and OVI absorption transitions (see also Tripp [$\mbox{et~al.}$]{} 2008 for a brief discussion of this system). Narrow absorption due to the SiIII$\lambda$1206 transition is also detected, but only in three components. All four components of the absorption are saturated. We have performed a Voigt profile analysis to determine the column densities of individual components. The results are presented in Table 7, which lists from columns (2) through (9) the best-fit column density, Doppler parameter ($b$), and their associated uncertainties. Column (10) of Table 7 gives the total estimated column densities of individual transitions. Given the saturated profiles of the transition and relatively low resolution in the observation of , we place conservative limits for the underlying $N({\mbox{${\rm H\,I}$}})$. The SiIII components all appear to be very narrow and unresolved in the STIS echelle data. We estimate the column densities based on a fixed $b$ value of $b=2.4$ [$\rm{km\,s}^{-1}$]{} that corresponds to gas of $\approx 10^4$ K. The best-fit Voigt profiles are shown in the right panel of Figure 12 as smooth curves.
While the blue-shifted and redshifted components of and show nearly symmetric kinematic signatures (and possibly in the CIII absorption as well) that are indicative of an origin in expanding supperbubble shells (see e.g. Bond [$\mbox{et~al.}$]{} 2001 and Simcoe [$\mbox{et~al.}$]{}2002, although see also Kawata & Rauch 2007 who showed that such absorption features could also be produced by filamentary accretion onto a central galaxy or galaxy group), there exists a strong abundance gradient in SiIII and OVI from $\Delta\,v=-158$ [$\rm{km\,s}^{-1}$]{} to $\Delta\,v=+188$ [$\rm{km\,s}^{-1}$]{} that suggests a significant variation in the underlying gas density (assuming a uniform background radiation field). Combining the optical morphology of the galaxy and the differential relative abundances of different ionization species indicates that the absorber is likely to originate in disrupted tidal tails as a result of a merger.
The nearest galaxy to the absorber at $z=0.28232$ (Tripp [$\mbox{et~al.}$]{}2008; Thom & Chen 2008a,b) has $R=19.9$ and $\rho=374\ h^{-1}$ physical kpc, with a corresponding rest-frame $R$-band absolute magnitude of $M_R-5\log\,h=-20.3$. The galaxy spectrum is dominated by absorption features. Three more galaxies are found at $|\Delta\,v|<300$ [$\rm{km\,s}^{-1}$]{} from the absorber redshift, but they are over $\rho=835\ h^{-1}$ kpc to $\rho=2.0\ h^{-1}$ Mpc physical distances away.
[p[0.75in]{}crrcrrcrrcrrrr]{} HI 1215$^a$ & & $>14.8$ & ... & & $> 15.0$ & ... & & $>15.0$ & ... & & $14.3\pm 0.1$ & $12\pm 2$ & & $>15.5$ SiIII 1206$^b$ & & $12.5\pm 0.2$ & $2.4$ & & $12.4\pm 0.2$ & $2.4$ & & $12.5\pm 0.1$ & $2.4$ & & $<11.9$ & ... & & $12.9\pm 0.2$ CIII 977 & & $13.6\pm 0.2$ & $17\pm 7$ & & $13.7\pm 0.1$ & $25\pm 5$ & & $13.3\pm 0.1$ & $26\pm10$ & & $13.4\pm 0.1$ & $36\pm 12$ & & $14.1\pm 0.2$ OVI 1031,1037$^a$ & & $14.5\pm 0.2$ & $13\pm 4$ & & $14.0\pm 0.1$ & $24\pm 9$ & & $13.8\pm 0.2$ & $29\pm 17$ & & $14.1\pm 0.1$ & $42\pm 13$ & & $14.8\pm
0.2$
ANALYSIS
========
We have obtained a spectroscopic sample of galaxies in fields around three QSOs, for which ultraviolet echelle spectra from FUSE and HST/STIS are available for identifying intervening hydrogen and absorbers. The galaxies span a broad range in the rest-frame $R$-band magnitude from $M_R-5\log\,h>-16$ to $M_R-5\log\,h<-22$ and a broad range in the projected physical distance from $\rho<30\ h^{-1}$ kpc to $\rho>4\ h^{-1}$ Mpc. We have shown in § 4.2 that the completeness of the spectroscopic survey is well understood and characterized by the angular selection function presented in Figure 6. Together with a complete sample of intervening and absorbers found at $z=0.01-0.5$ in these fields (§ 5 and Figure 7), the galaxy sample offers a unique opportunity for studying the origin of and absorbers based on their cross-correlation amplitude with known galaxies, and for investigating the gas content in halos around galaxies.
The Galaxy–Absorber Cross-Correlation Functions
-----------------------------------------------
To quantify the origin of and absorbers, we first measure the projected two-point galaxy auto-correlation function using a flux-limited sample that has been assembled from our spectroscopic survey. The flux-limited galaxy sample contains 670 spectroscopically identified intervening galaxies of $R\le 22$ within $\Delta\,\theta=11'$ of the background QSOs, 222 of which show absorption-line dominated spectral features. The median rest-frame $R$-band absolute magnitude of the 448 emission-line dominated galaxies is $\langle M_R\rangle-5\log\,h=-19.2$ (corresponding to $\approx 0.33\,L_*$ for $M_R*-5\log\,h=-20.44$ from Blanton [$\mbox{et~al.}$]{} 2003), while the 222 absorption-line dominated galaxies have $\langle M_R\rangle-5\log\,h=-20.43$ (corresponding to $\approx
L_*$). We calculate the projected two-point correlation function $\omega_{gg}(r_p)$ versus co-moving projected distance $r_p$, using the Landy & Szalay (1993) estimator $$\omega_{gg}(r_p)=\frac{D_g\,D_g-2\,D_g\,R_g+R_g\,R_g}{R_g\,R_g},$$ where $D_g$ represents the input galaxy sample and $R_g$ represents a random galaxy sample. The random galaxy sample is generated following the completeness function presented in Figure 6 for each field. To minimize possible counting noise introduced by the random galaxy sample, we have produced a random galaxy sample per field that is a factor of ten larger than the true galaxy sample. The two-point function is then calculated by counting the appropriate $D_g\,D_g$, $D_g\,R_g$, and $R_g\,R_g$ pairs within different $r_p$ intervals, from $r_p<250\ h^{-1}$ kpc, to $r_p=250-800\ h^{-1}$ kpc, to $r_p=800-1500\ h^{-1}$ kpc, and to $r_p=1.5-3\ h^{-1}$ Mpc.
We present in panel (a) of Figure 13 the auto-correlation functions intergrated over $40\,h^{-1}$ co-moving Mpc in redshift space for all galaxies (open circles), absorption-line dominated galaxies (open triangles), and emission-line dominated galaxies (open squares) in the flux-limited sample. Errorbars represent poisson counting uncertainties[^8]. These data points for each of these subsamples are also repeated in panels (b), (c), and (d), respectively, to be compared with their corresponding galaxy–absorber cross-correlation functions. Panel (a) of Figure 13 confirms that absorption-line dominated galaxies indeed cluster more strongly than emission-line dominated galaxies. The observed clustering amplitude of absorption-line dominated galaxies in our sample is comparable to what is found by Zehavi [$\mbox{et~al.}$]{} (2005) for SDSS galaxies of $M_R-5\log\,h<-19$ at $z\sim 0.1$. The stronger clustering amplitude indicates that on average these absorption-line galaxies originate in higher overdensity regions and presumably more massive dark matter halos.
Next, we measure the projected two-point galaxy–absorber cross-correlation function using the flux-limited galaxy sample discussed above and the absorber catalogs discussed in § 5. We exclude absorbers that are within $\Delta\,v=3000$ [$\rm{km\,s}^{-1}$]{} of the background QSO to reduce contaminations due to QSO associated absorbers. There are 195 absorbers and 15 absorbers included in our analysis. We calculate the projected two-point cross-correlation function $\omega_{ga}(r_p)$ versus co-moving projected distance $r_p$ between galaxies and absorbers, using the Landy & Szalay (1993) estimator $$\omega_{ga}(r_p)=\frac{D_g\,D_a-D_g\,R_a-R_g\,D_a+R_g\,R_a}{R_g\,R_a},$$ where $D_g$ and $D_a$ represent the input galaxy and absorber samples, and $R_g$ and $R_a$ represent random galaxy and absorber samples. We have produced random and absorber catalogs for each field, assuming a flat absorber selection function in the redshift interval between $z=0.01$ and $z_{\rm QSO}$. The random absorber catalogs are also ten times larger than the true absorber catalogs. The two-point cross-correlation function is then calculated by counting the appropriate $D_g\,D_a$, $D_g\,R_a$, $R_g\,D_a$, and $R_g\,R_a$ pairs within the same $r_p$ intervals of the auto-correlation function calculation.
Panels (b), (c), and (d) of Figure 13 show the galaxy–absorber and galaxy– absorber cross-correlation functions intergrated over $40\,h^{-1}$ co-moving Mpc in redshift space. Comparisons between galaxy–absorber cross-correlation and galaxy auto-correlation functions of different galaxy type show four interesting features.
First, while both strong and weak absorbers exhibit on average weaker clustering amplitudes than the galaxies as a whole (solid points in panel b), strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\ge 14$ appear to exhibit a comparable clustering amplitude on large scales of $r_p=0.25-3\ h^{-1}$ Mpc and a higher clustering amplitude on small scales of $r_p\le 250\ h^{-1}$ kpc with the emission-line dominated galaxies (solid points with solid errorbars panel c). The comparable clustering signal on large scales indicates that strong absorbers and emission-line galaxies share common halos. The higher galaxy–absorber cross-correlation signal relative galaxy auto-correlation signal on small scales is understood if the gas covering fraction is high in halos around these galaxies. Taking into account previous findings of Chen [$\mbox{et~al.}$]{} (1998; 2001a), we conclude that emission-line dominated galaxies are surrounded by extended gas of nearly unity covering fraction and that strong absorbers primarily probe extended gaseous halos around young star-forming galaxies.
Second, these strong absorbers exhibit on average $\approx 6$ times weaker amplitude with the absorption-line dominated galaxies (solid points with solid errorbars panel d) at all separations of $r_p<3\ h^{-1}$ Mpc. The weaker clustering amplitude between strong absorbers and absorption-line dominated galaxies is qualitatively consistent with the clustering amplitude of MgII absorbers and luminous red galaxies (LRGs) found by Gauthier [$\mbox{et~al.}$]{}(2009) using a flux-limited LRG sample. It suggests that the incidence of strong absorbers is very low around these absorption-line dominated galaxies.
Third, weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})<13.5$ exhibit only a weak clustering signal to galaxies of all type at separations $r_p<3\
h^{-1}$ Mpc (solid points with dotted errorbars in Figure 13). The weak clustering amplitude indicates that the majority of these weak absorbers do not share the same dark matter halos as the galaxies. This result extends the finding of Grogin & Geller (1998) to ${\lesssim}0.3\,L_*$ galaxies that weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})<13.5$ occur far more frequently in underdense regions in comparison to these dwarf galaxies. Clustering amplitudes measured at large separations of $r_p \approx 10\ h^{-1}$ Mpc are necessary to constrain the mean overdensities where these weak absorbers reside.
Finally, absorbers exhibit similar clustering amplitudes as strong absorbers. Specifically, they show a comparable clustering amplitude (crosses in panel c of Figure 13) with emission-line galaxies on scales of $r_p=0.25- 1.5\ h^{-1}$ Mpc but a factor of $\approx 6$ times lower (petagon points in panel d of Figure 13) than absorption-line dominated galaxies. Despite large uncertainties, the differential clustering amplitudes suggest a direct association between absorbers and star-forming galaxies and a lack of physical connections between the majority of absorbers and the gaseous halos around these massive, early-type galaxies.
We have also attempted to investigate possible dependence of the cross-correlation amplitude on the strength of OVI absorbers using a mark two-point correlation statistic (see e.g. Sheth et al. 2005). Applying the absorber column density as a mark, we could not distinguish whether such dependence is present. It appears that the sample size is still too small for such study to yield statistically significant results.
Nevertherless, the comparable clustering amplitude among emission-line galaxies, strong absorbers, and absorbers is, however, difficult to interpret, because the number density of absorbers is found to be $n_{\displaystyle{\mbox{${\rm O\,VI}$}}}\approx 10-17$ per unit redshift interval per line of sight (e.g. Tripp [$\mbox{et~al.}$]{} 2008; Thom & Chen 2008a) and the number density of strong absorbers is found to be $n_{\displaystyle{\mbox{${\rm Ly}\alpha$}}}\approx 25-30$ (Weymann [$\mbox{et~al.}$]{} 1998; Dobrzycki [$\mbox{et~al.}$]{} 2002). It shows that on average roughly half of the strong absorbers have associated . We defer a more detailed discussion of this discrepancy to § 8.2.
Incidence and Covering Fraction of OVI Absorbing Gas Around Galaxies
--------------------------------------------------------------------
The galaxy sample also allows us to examine the incidence and covering fraction of absorbing gas around galaxies. We have identified a total of 20 foreground galaxies of $R\le 22$ in our spectroscopic sample that occur at co-moving projected distances $r_p
\le 250\ h^{-1}$ kpc from the lines of sight toward HE0226$-$4110, PKS0405$-$123, and PG1216$+$069. Six of the galaxies are absorption-line dominated galaxies and eight are “isolated” galaxies without additional neighboring galaxies found at $r_p\le 250\ h^{-1}$ kpc from the QSO sightline and within velocity offsets of $\Delta\,v\le 300$ [$\rm{km\,s}^{-1}$]{}. The adopted search volume is typical of the halo size around $L_*$ galaxies. Galaxies and absorbers that occur in this small volume are likely to share a common halo.
For each galaxy in this sample, we search for corresponding absorption features in the available echelle spectra of the QSOs and measure the rest-frame absorption equivalent width $W(1031)$. In the absence of an absorption feature, we determine a 2-$\sigma$ upper limit to $W(1031)$ for the galaxy. Figure 14 shows $W(1031)$ versus $r_p$ for these galaxies of $R\le 22$ in our sample. Emission-line dominated galaxies are marked by squares and absorption-line dominated galaxies are marked by triangles. Arrows indicate the 2-$\sigma$ upper limit in $W(1031)$ for galaxies that do not have an associated absorber. For those galaxies with neighbors, we include only the galaxy at smallest $r_p$. These galaxies are marked by an additional open circle (indicating at least one neighboring galaxy is present at $r_p\le 250\ h^{-1}$ kpc from the absorber) in Figure 14.
![Observed absorption strength for all galaxies of $R\le
22$ found at $z=0.01-0.5$ and co-moving projected distances $r_p\le
250\ h^{-1}$ kpc in the fields around HE0226$-$4110, PKS0405$-$123, and PG1216$+$069. Squares represent emission-line dominated galaxies, and triangles represent absorption-line dominated galaxies. Arrows indicate the 2-$\sigma$ upper limit to the rest-frame 1031 Å absorption equivalent width for galaxies that do not have an associated absorber. Note that there are two non-absorbing galaxies occur at $r_p\approx 157\ h^{-1}$ kpc. Several galaxies are found to have [*at least*]{} one neighboring galaxy located within $r_p=250\ h^{-1}$ kpc and $|\Delta\,v|\le 300$ [$\rm{km\,s}^{-1}$]{}from the absorber, in which cases we include only the closest galaxy in the plot. Points with open circles represent galaxies that either have additional neighboring galaxies or exhibit disturbed morphologies indicative of a merger event.](f14.eps)
It is clear that none of the absorption-line dominated galaxies in Figure 14 has a corresponding absorber to a sensitve upper limit, consistent with the lower clustering amplitude seen in panel (d) of Figure 13. Considering only emission-line dominated galaxies in our sample, we find that nine of the 14 galaxies at $r_p\le 250\
h^{-1}$ kpc have an associated absorber. This translates to a covering fraction of $\kappa \approx 64$% for absorbing gas within $250\ h^{-1}$ co-moving projected distance of star-forming galaxies. The covering fraction appears to be $\kappa\approx 100$% at $r_p\le 100\ h^{-1}$ kpc, as all four emission-line dominated galaxies have an associated absorber. The observed $\kappa\approx 64$% covering fraction at $r_p\le 250\ h^{-1}$ kpc from emission-line galaxies is consistent with the known number density statistics of strong and absorbers, but it further underscores the difficulty in interpreting the comparable clustering measurements presented in panel (c) of Figure 13.
We note that the absorbing galaxy at $r_p=72\ h^{-1}$ in Figure 14 is the complex absorber at $z=0.1242$ toward PG1216$+$069. As shown in § 6.3 and Figure 12, the galaxy exhibits a disturbed morphology that suggests an on-going merger. It is therefore likely that this galaxy is located in a multiple galaxy environment. In addition, the only galaxy at $r_p<100\ h^{-1}$ that does not have a corresponding absorber to a sensitive upper limit is an absorption-line dominated galaxy at $z=0.2678$. At this redshift, an emission-line dominated galaxy of $R=21.6$ at $r_p \approx 140\
h^{-1}$ and an absorption-line dominated galaxy of $R=18.6$ at $r_p=251\ h^{-1}$ kpc are also found from the QSO sightline. Although our sample is still small and the results have large uncertainties because of systematic bias from possible field to field variations, it is interesting to find that star-forming galaxy “groups” at $r_p{\lesssim}100\ h^{-1}$ kpc appear to show a higher incidence of absorbers.
DISCUSSION
==========
Using a flux-limited ($R\le 22$) sample of 670 intervening galaxies spectroscopically identified at $z<0.5$ in fields around three QSOs, HE0226$-$4110, PKS0405$-$123, and PG1216$+$069, we have calculated the projected two-point correlation functions of galaxies and QSO absorption-line systems. Our analysis confirms early results of Chen [$\mbox{et~al.}$]{} (2005; see also Figure 4 of Wilman [$\mbox{et~al.}$]{} 2007) that strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})> 14$ share comparable clustering amplitude with emission-line dominated galaxies and that weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\le 13.5$ cluster only very weakly around galaxies. Combining the observed $\approx 100$% covering fraction around galaxies (Chen [$\mbox{et~al.}$]{} 1998, 2001a) and the comparable clustering amplitude, we conclude that strong absorbers at $z<0.5$ primarily probe extended gaseous halos around star-forming galaxies and that a large fraction of weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\le 13.5$ originate in low overdensity regions that occur more frequently than $\approx 0.3\,L_*$ galaxies. The differential clustering amplitude of strong absorbers with different types of galaxies at projected co-moving distances $r_p\le
250\ h^{-1}$ kpc further indicates that the incidence of strong absorbers is very low around absorption-line dominated galaxies.
We have also studied the correlation between galaxies and absorbers. While our spectroscopic survey has uncovered multiple large-scale galaxy overdensities in the three fields, Figures 6&7 show that only two absorbers (at $z=0.0966$ toward PKS0405$-$123 and at $z=0.2823$ toward PG1216$+$069) occur at the redshifts of these overdensities. An interesting result from the two-point correlation analysis is that while the majority of absorbers do not probe the gaseous halos around massive, early-type galaxies, they appear to share common halos with emission-line dominated star-forming galaxies (and therefore with strong absorbers as well). The interpretation of the clustering analysis is, however, complicated by additional observations that the covering fraction of absorbing gas is $\kappa\approx 64$% around emission-line dominated galaxies. While our sample is still small, this low gas covering fraction is qualitatively consistent with the observed difference in the number densities of strong and absorbers. It indicates that only a subset of star-forming galaxies are associated with the observed absorption features. The discrepancy between the observed clustering amplitude and covering fraction of absorbers implies that additional variables need to be accounted for.
In this section, we review the known properties of individual absorbing galaxies, and discuss the implications of our results on the origin of absorbers and on the extended gaseous envelopes around galaxies.
The Properties of Absorbing Galaxies
-------------------------------------
There are 13 absorption systems (15 well-separated components) identified at $z=0.017-0.495$ along the sightlines toward HE0226$-$4110, PKS0405$-$123, and PG1216$+$069 (Tripp [$\mbox{et~al.}$]{}2008; Thom & Chen 2008a,b), including one at $z=0.207$ with associated NeVIII features (Savage [$\mbox{et~al.}$]{} 2005). Our spectroscopic survey, together with previous searches in these fields, has uncovered 11 galaxies at projected co-moving distances of $r_p<250\ h^{-1}$ kpc and velocity offsets of $\Delta\,v\le 300$ [$\rm{km\,s}^{-1}$]{} from six of the 13 absorbers[^9]. Although half of the absorbers do not yet have an associated galaxy identified, this small sample of 11 galaxies allows us to examine the common properties of absorbing galaxies, including the luminosity, galaxy environment, morphology, and spectral features.
First, the rest-frame $R$-band absolute magnitude of the galaxies associated with six absorbers span a broad range from $M_R-5\log\,h=-15.4$ to $M_R-5\log\,h=-20.9$. Adopting the magnitude limit at which our spectroscopic survey is 100% complete, we further place a conservative upper limit on $M_R$ for the underlying absorbing galaxies of six remaining absorbers[^10]. A summary of galaxies found in the vicinity of absorbers is presented in Table 8, which lists from colums (2) through (11) the absorber redshift $z_{\rm abs}$, the velocity centroid relative to $z_{\rm
abs}$ of the OVI feature $\Delta\,v_{\rm OVI}$, the OVI absorbing gas column density $N({\rm OVI})$, the rest-frame absorption equivalent width $W_{\rm rest}(1031)$, the velocity centroid relative to $z_{\rm abs}$ of the HI feature $\Delta\,v_{\rm HI}$, the HI absorbing gas column density $N({\rm HI})$, references of the absorber measurements, the projected co-moving distance $r_p$ and absolute magnitude $M_R$ of the galaxies, and references of the galaxy measurements. We find that while the known absorbing galaxies exhibit a broad range of intrinsic luminosity, from $<0.01\,L_*$ to $\approx L_*$, a clear correlation between $N({\mbox{${\rm O\,VI}$}})$ and $M_R$ is absent.
At the same time, four of the six absorbers that have known associated galaxies are surrounded by either a merging galaxy pair or by more than one galaxy in the small volume of $r_p<250\ h^{-1}$ kpc and $\Delta\,v\le 300$ [$\rm{km\,s}^{-1}$]{}. In contrast, deep spectroscopic surveys to search for the galaxies producing absorption features of low-ionization species such as MgII$\lambda\lambda\,2796, 2803$ have yielded ${\lesssim}16$% of MgII absorbers originating in multiple galaxy environment (e.g. Steidel [$\mbox{et~al.}$]{} 1997). On the other hand, not all galaxy “groups” within a similar volume produce an associated absorber in the spectrum of the background QSO. We have shown in § 7.2 that a group of three galaxies at $z\approx
0.268$ and a group of two galaxies at $z\approx 0.199$ within $r_p\le
250\ h^{-1}$ kpc from the sightline toward HE0226$-$4110 do not have an associated absorber. The most luminous members of both groups exhibit absorption-line dominated spectral features (see e.g.the bottom panel of Figure 15). It appears that originates preferentially in groups of gas-rich galaxies. A detailed investigation using a statistically representative sample of galaxy groups close to QSO sightlines is necessary to confirm this result (see Thom [$\mbox{et~al.}$]{} 2009, in preparation).
Using available high-resolution images in the HST archive, we are also able to examine the detailed optical morphology of five galaxies in the sample of 11 absorbing galaxies. Individual images presented in Figures 10 & 12 show that four of the galaxies (responsible for three absorbers) exhibit disk-like morphology with mildly disturbed features on the edge. This asymmetric disk morphology suggests that tidal disruption may be in effect. The absorbing galaxy at $z=0.4942$ (right panel of Figure 10) appears to be a regular disk, although the available image depth is insufficient to reveal more details.
Finally, all but one of these 11 absorbing galaxies show prominent emission-line features that suggest a range of ISM metallicity and star formation history. The exception is a galaxy at $z=0.2077$ and projected co-moving distance $r_p=238\ h^{-1}$ kpc from the sightline toward HE0226$-$ 4110, which exhibits primarily absorption features in the spectrum (see Mulchaey & Chen 2009). The spectra of three absorbing galaxies, one at $z=0.0965$ and two at $z=0.1670$ toward PKS0405$-$123, have been published by previous authors (see Spinrad [$\mbox{et~al.}$]{} 1993; Prochaska [$\mbox{et~al.}$]{} 2006). The spectra of three galaxies identified for the $+$ NeVIII absorber at $z=0.207$ toward HE0226$-$4110 are presented in Mulchaey & Chen (2009). Two of these galaxies show emission-line dominated spectral features. The spectrum of the absorbing galaxy at $z=0.1239$ toward PG1216$+$069 is presented in Figure 12. It exhibits strong emission features due to H$\alpha$, \[NII\], and \[SII\] that imply roughly solar metallicity in the ISM. In the top four panels of Figure 15, we present the spectra of four additional absorbing galaxies identified in our survey, three of which are for the absorber at $z=0.0918$ toward PKS0405$-$123 and one for the absorber at $z=0.4951$ along the same sightline. Emission line features, such as H$\beta$ and \[OIII\], are present in all four galaxies. While the absorbing galaxy at $z=0.4942$ exhibits Balmer absorption series that implies a post-starburst nature, the lack of \[NII\] emission features together with the presence of strong H$\alpha$ emission in the three absorbing galaxies at $z=0.091$ places 2-$\sigma$ upper limits to the ISM metallicity at $\approx
10$% to 30% of solar values. In contrast to the spectral features of absorbing galaxies, we include in the bottom panel of Figure 15 the spectrum of an absorption-line dominated galaxy at $z=0.2678$ and $r_p=62\ h^{-1}$ co-moving kpc from the sightline toward HE0226$-$4110. The galaxy has $M_R-5\log\,h=-19.9$ and does not have a corresponding absorber to a 2-$\sigma$ upper limit of $W(1031)<0.02$ Å.
[lrrrrrrrcrrr]{} HE0226$-$4110 & 0.01746 & 0 & $13.6\pm 0.1$ & $40\pm 10$ & 0 & $13.28\pm 0.06$ & (1) & & ... & ... & & 0.20701 & 0 & $14.37\pm 0.03$ & $169\pm 15$ & $-24$ & $15.06\pm 0.04$ & (2) & & 32.0 & $-17.2$ & (3) & & & & & $+5$ & $14.89\pm 0.05$ & (2) & & 92.3 & $-18.9$ & (3) & & & & & & & & & 238.1 & $-18.9$ & (3) & 0.32639 & 0 & $13.6\pm 0.2$ & $43\pm 8$ & 0 & $< 12.5$ & (4) & & ... & $>-17.3$ & (3) & 0.34034 & 0 & $13.9\pm 0.1$ & $62\pm 7$ & 0 & $13.6\pm 0.1$ & (4) & & ... & $>-17.4$ & (3) & 0.35529 & 0 & $13.7\pm 0.1$ & $46\pm 8$ & 0 & $13.6\pm 0.2$ & (4) & & ... & $>-17.5$ & (3) PKS0405$-$123 & 0.09180 & $+20$ & $13.80\pm 0.04$ & $73\pm 8$ & 0 & $14.52\pm 0.04$ & (5) & & 80.2 & $-16.0$ & (3) & & & & & & & & & 101.2 & $-16.1$ & (3) & & & & & & & & & 228.5 & $-15.4$ & (3) & 0.09658 & 0 & $13.7\pm 0.2$ & $71\pm 9$ & 0 & $14.65\pm 0.05$ & (5) & & 203.5 & $-18.2$ & (6) & 0.16710 & 0 & $14.78\pm 0.07$ & $361\pm 41$ & 0 & $16.45\pm 0.05$ & (5),(7) & & 94.3 & $-20.9$ & (8) & & & & & & & & & 79.1 & $-18.1$ & (8) & 0.18291 & $-87$ & $13.7\pm 0.2$ & $42\pm 13$ & $-82$ & $14.90\pm 0.05$ & (4),(5) & & ... & $>-18.9$ & (3) & & 0 & $14.0\pm 0.2$ & $65\pm 15$ & 0 & $14.1\pm 0.1$ & (4),(5) & & ... & ... & & 0.36332 & 0 & $13.5\pm 0.1$ & $30\pm 5$ & 0 & $13.6\pm 0.1$ & (4),(5) & & ... & $>-20.6$ & (3) & 0.49510 & 0 & $14.5\pm 0.1$ & $213\pm 16$ & 0 & $14.3\pm 0.2$ & (4),(5),(9) & & 114.9 & $-18.8$ & (3) PG1216$+$069 & 0.12420 &$-158$ & $14.5\pm 0.2$ & $225\pm 25$ &$-158$ & $>14.8$ & (3) & & 72.2 & $-20.0$ & (3) & & $-82$ & $14.0\pm 0.1$ & & $-82$ & $>15.0$ & (3) & & ... & ... & & &$+120$ & $13.8\pm 0.2$ & $194\pm 28$ &$+120$ & $>15.0$ & (3) & & ... & ... & & &$+188$ & $14.1\pm 0.1$ & &$+188$ & $14.3\pm 0.1$ & (3) & & ... & ... & & 0.28232 & 0 & $13.4\pm 0.2$ & $26\pm 6$ & 0 & $16.70\pm 0.04$ & (4),(10) & & ... & $>-19.9$ & (3)\
Excluding the absorber at $z=0.01746$ toward HE0226$-$4110 for which our galaxy survey does not cover beyond $160\ h^{-1}$ kpc, we note that four of the five strong absorbers with $\log\,N({\rm O\,VI})\ge 14$ have been identified with galaxies at $r_p<250\ h^{-1}$ kpc and $\Delta\,v\le 300$ [$\rm{km\,s}^{-1}$]{}, while only two of the seven weaker absorbers with $\log\,N({\rm O\,VI})< 14$ have been identified with galaxies at $r_p<250\ h^{-1}$ kpc and $\Delta\,v\le 300$ [$\rm{km\,s}^{-1}$]{}. Given the small sample size, we cannot rule out possible correlations between absorber strengths and galaxy properties or the possibility that some of the weaker OVI absorbers may originate in underdense IGM. For absorbers that do not have galaxies found at $r_p<250\ h^{-1}$ kpc, we estimate the magnitude limit of the absorbing galaxies based on our survey depth and present these numbers in Column (9) of Table 8.
The Origin of Absorption Systems
---------------------------------
The galaxy– absorber cross-correlation analysis presented in § 7.1 has clearly identified emission-line dominated galaxies to be principally responsible for the observed absorbers, but the physical mechanism that distributes the metal-enriched gas to large galactic distances is unclear. Two competing scenarios to explain the origin of absorbers in regions around emission-line galaxies are (1) starburst driven outflows (e.g. Heckman [$\mbox{et~al.}$]{} 2001; Kawata & Rauch 2007; Oppenheimer & Davé 2009) and (2) satellite accretion (e.g. Wang 1993; Bournaud [$\mbox{et~al.}$]{} 2004; Elmegreen [$\mbox{et~al.}$]{} 2007). Given that these emission-line galaxies are presumably low-mass (given the sub-$L_*$ nature) and star-forming (given the presence of strong emission lines), the strong correlation seems qualitatively consistent with the expectations of originating in starburst outflows (c.f. Kawata & Rauch 2007; Oppenheimer & Davé 2009). On the other hand, satellite accretions or galaxy interactions are also expected to induce star formation (see e.g. Li [$\mbox{et~al.}$]{} 2008). Here we examine whether additional insights for the origin of absorbers can be gleaned based on known properties of the absorbing galaxies.
We have noted an apparent discrepancy between the observed partial covering fraction of absorbing gas, $\kappa\approx 64$%, at projected co-moving distances $r_p{\lesssim}250\ h^{-1}$ kpc and the comparable clustering amplitudes of emission-line galaxies and absorbers on scales $r_p\le 1.5\ h^{-1}$ Mpc. While the partial covering fraction indicates that not all emission-line galaxies contribute to the observed absorber statistics, the comparable clustering amplitude suggests otherwise. Namely, every emission-line galaxy has a corresponding absorber. Likewise, the comparable clustering amplitudes of and strong absorbers also imply that the two absorber populations share common halos, but absorption-line surveys show that not all strong absorbers are accompanied by an absorber (see e.g. Tables 3&8). Although both starburst outflows and tidal debris due to galaxy interactions may explain the observed partial covering fraction, it is not clear that they can both explain the comparable clustering amplitude at $r_p{\lesssim}3\ h^{-1}$ Mpc.
To understand these apparent discrepancies, we first note that the correlation amplitude on scales $r_p < 1\ h^{-1}$ Mpc is dominated by galaxies that share common halos (e.g. Zheng [$\mbox{et~al.}$]{} 2007; Tinker [$\mbox{et~al.}$]{} 2007). The comparable clustering amplitudes of and emission-line galaxies (and strong absorbers) found in our small sample from three QSO fields may be understood, if absorbers arise preferentially in groups of emission-line galaxies. This hypothesis is supported by two features found in Figure 14. First, four of the five galaxy– absorber pairs are indeed found in an environment of multiple emission-line galaxies. Second, only one of the four “isolated” emission-line galaxies has a correponding absorber. Available observations are indeed consistent with the expectations that absorbers arise primarily in gas-rich galaxy groups. Taking into account the mildly disturbed disk-like morphologies of absorbing galaxies, we further argue that the absorbers originate in tidal debris produced by galaxy interactions in a group or pair environment and that tidal interactions may be principally responsible for distributing chemically enriched gas to large galactic distances.
A larger sample is necessary to improve the uncertainties in the measurements of $\kappa$ for both group/pair and “isolated” galaxies, and to address possible bias due to field to field variation. Extending the cross-correlation analysis to scales of $r_p\approx 10\
h^{-1}$ Mpc is also necessary for constraining the mean halo mass of the absorbers based on their large-scale clustering amplitudes.
Extended Gas Around Galaxies
----------------------------
Previous studies have shown that luminous galaxies are surrounded by extended gaseous envelopes that may be probed by CIV and MgII absorption transitions out to projected physical distances of $\rho\approx 100\ h^{-1}$ kpc (e.g. Steidel [$\mbox{et~al.}$]{} 1994; Chen [$\mbox{et~al.}$]{}2001b; Tinker & Chen 2008; Chen & Tinker 2008) and by strong absorption out to $\rho\approx 200\ h^{-1}$ kpc (Lanzetta [$\mbox{et~al.}$]{} 1995; Chen [$\mbox{et~al.}$]{} 1998,2001a). The covering fraction of extended gas probed by the presence of these transitions is found to be $>80$%. Our analysis extends these studies to absorbing gas. We find extended gas is indeed present around emission-line galaxies out to projected co-moving distance $r_p\approx 250\ h^{-1}$ kpc (corresponding to a projected physical distance $\rho\approx 200\
h^{-1}$ kpc at $z=0.3$) with a mean covering fraction of $\kappa\approx 64$%. While and MgII absorption features are found to be progressively weaker in galaxies at larger impact parameters (Chen [$\mbox{et~al.}$]{} 1998, 2001a; Chen & Tinker 2008), we find a lack of such correlation in either CIV (Chen [$\mbox{et~al.}$]{} 2001b) or (Figure 14). Including known absorbing galaxies from Cooksey [$\mbox{et~al.}$]{} (2008) and Lehner [$\mbox{et~al.}$]{} (2009) further increases the scatter in the $W(1031)$ versus $r_p$ distribution at $r_p<250\ h^{-1}$ kpc.
In a simple two-phase medium model, CIV and MgII absorbers arise in photo-ionized cold clumps pressure-confined in a hot medium. The lack of correlation between absorber strength and galaxy impact parameter can therefore be understood by higher gas pressure (and therefore reduced abundances of C$^{3+}$) at smaller radii (e.g. Mo & Miralda-Escudé 1996). The same model cannot explain the presence of around galaxies of a broad range of luminosity due to the high ionization potential. Instead, the mildly disturbed disk morphologies observed in absorbing galaxies suggest that tidal debris in small groups or close galaxy pairs may be principally responsible for the observed absorption features.
Additional insights may be learned from comparisons with known properties of the Milky Way halos. While exquisite details, including tidal streams, infalling clouds, and outflows, have been recorded for extended gas around the Milky Way (e.g. Sembach [$\mbox{et~al.}$]{} 2003), applications of these observations for constraining models of galaxy growth has been limited due to a lack of knowledge for the distances to these Halo clouds. In contrast, observations of distant galaxy–absorbers pairs offer direct measurements of the distances between absorbing gas and star-forming regions, but the origin of the absorbing clouds (either due to accretion, outflow, or tidal stripping) is often more uncertain.
A systematic survey of absorbers toward 100 extragalactic sources has yielded positive detections in the Galactic thick disk and halo along $60-85$% of these sightlines (Sembach [$\mbox{et~al.}$]{} 2003). Because some of these detections are associated with known high velocity clouds at $\approx 10$ kpc distances (e.g. Thom [$\mbox{et~al.}$]{}2008), the observed incidence of $60-85$% represents an upper limit for the covering fraction of absorbing gas over $\sim 250$ kpc radius to an external observer. This partial covering of gas in the halo of the Milky Way, representing typical $\sim\,L_*$ galaxies in the nearby universe, is qualitatively consistent with the partial covering fraction found in our study for halos around distant sub-$L_*$ galaxies. The lack of evidence for large-scale outflows in the Milky Way halo based on detections of absorbers also appears to be consistent with our interpretation that the majority of absorbers are not produced in starburst outflows around distant star-forming galaxies. A larger sample of galaxy– absorber pairs that covers a broad range of impact parameters will allow a conclusive characterization of the nature of the gas, which in turn will shed light for the halo gas content around the Milky Way.
Finally, our analysis has also shown that the majority of strong and absorbers do not probe the gaseous halos around early-type galaxies. To study the warm-hot gas around early-type galaxies or galaxy groups would require a different probe.
SUMMARY
=======
We have carried out a deep, wide-area survey of galaxies in fields around HE0226$-$4110, PKS0405$-$123, and PG1216$+$069. The QSO fields are selected to have ultraviolet echelle spectra available from FUSE and HST/STIS. The high-resolution UV spectra have uncovered 195 absorbers of $\log\,N({\mbox{${\rm H\,I}$}})=12.5-16.3$ and 13 absorbers of $\log\,N({\mbox{${\rm O\,VI}$}})=13.4-14.7$ (15 individual components) along the sightlines toward the background QSOs. The spectroscopic program has yielded redshift measurements for 1104 galaxies in three fields. The rest-frame $R$-band magnitudes of these galaxies range from $M_R-5\log\,h>-16$ to $M_R-5\log\,h{\lesssim}-22$. The projected physical distances of these galaxies range from $\rho<30\ h^{-1}$ kpc to $\rho>4\ h^{-1}$ Mpc. While our spectroscopic survey has uncovered multiple large-scale galaxy overdensities in the three fields, only two absorbers (at $z=0.0966$ toward PKS0405$-$123 and at $z=0.2823$ toward PG1216$+$069) occur at the redshifts of these overdensities. Including previously found absorbing galaxies, we have collected a sample of ten galaxies that coincide with six absorbers found along the three sightlines. Adopting our spectroscopic survey limit, we further place upper limits to the intrinsic luminosities of six remaining absorbing galaxies that are still missing (three of these are expected to be fainter than $M_R-5\log\,h=-17.5$). Two of the fields, PKS0405$-$123 and PG1216$+$069, have high quality HST/WFPC2 images available that cover the $2'\times 2'$ area roughly centered at the QSO. These images also allow us to examine the optical morphologies of individual absorbing galaxies.
Combining various absorber and galaxy data, we have performed a cross-correlation study to understand the physical origin of and absorbers and to constrain the properties of extended gas around galaxies. The main results of our study are summarize below.
1\. Using a flux-limited sample of 670 foreground galaxies within $\Delta\,\theta=11'$ of the QSO, we find based on a cross-correlation analysis that both strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\ge 14$ and absorbers exhibit a comparable clustering amplitude as emission-line dominated galaxies on scales of $r_p<3\ h^{-1}$ co-moving Mpc. The clustering amplitudes of these absorbers are found to be $\approx 6$ times lower than those of absorption-line dominated galaxies. At the same time, weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})<13.5$ exhibit only a weak clustering signal to galaxies of all type. These findings confirm early results of Chen [$\mbox{et~al.}$]{}(2005; see also Figure 4 of Wilman [$\mbox{et~al.}$]{} 2007) that strong absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\le 14$ probe extended halos of emission-line dominated galaxies and that a large fraction of weak absorbers of $\log\,N({\mbox{${\rm H\,I}$}})\le 13.5$ are likely to originate in low-overdensity regions that occur more frequently than $\approx
0.3\,L_*$ galaxies.
2\. absorbers exhibit a similar behavior as strong absorbers. These absorbers also show a comparable clustering amplitude as emission-line galaxies but a factor of six lower clustering amplitude relative to absorption-line dominated galaxies on scales of $r_p\le 3\ h^{-1}$ Mpc. The results imply that the majority of absorbers do not probe the gaseous halos around massive, early-type galaxies, and that they probe primarily halos around emission-line dominated star-forming galaxies.
3\. Using a small sample of 20 $R\le 22$ galaxies found at co-moving projected distances $r_p \le 250\ h^{-1}$ kpc from the lines of sight toward the three QSOs, we find that none of the absorption-line dominated galaxies in the sample has a corresponding absorber to a sensitive upper limit of $W(1031){\lesssim}0.03$ Å, and that the covering fraction of absorbing gas around emission-line dominated galaxies is $\kappa\approx 64$%.
4\. Four of the six absorbers that have associated galaxies identified are surrounded by either a merging galaxy pair or by more than one galaxy in the small volume of $r_p<250\ h^{-1}$ kpc and $\Delta\,v\le 300$ [$\rm{km\,s}^{-1}$]{}. On the other hand, two galaxy “groups” found at $r_p\le 250\ h^{-1}$ kpc from a QSO sightline do not have an associated absorber. A more detailed examination of galaxies in these latter two groups shows that the most luminous members of both groups exhibit absorption-line dominated spectral features, suggesting that may originate preferentially in gas-rich galaxy groups.
5\. Available high-resolution HST/WFPC2 images of five absorbing galaxies show that these galaxies exhibit disk-like morphology with mildly disturbed features on the edge, suggesting that tidal disruption may be in effect.
JSM would like to acknowledge the visitor’s program at the Kavli Institute for Cosmological Physics, where part of the work presented here was completed. We thank helpful discussions with J. Tinker and C. Thom. We thank L. Matthews, M. Rauch, A. Kravtsov for helpful comments on an early version of the manuscript. We thank R. Simcoe and A. Bolton for assistance on obtaining part of the IMACS images presented in this paper and R. Marzke for helpful discussions on reducing and assembling geometrically distorted multi-CCD imaging data. We are grateful to D. Kelson and G. Walth for assistance on the reduction of the IMACS spectra. We also thank J. Scott for providing the combined FUSE spectrum of PG1216$+$069. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. H.-W.C. acknowledges partial support from NASA Long Term Space Astrophysics grant NNG06GC36G and an NSF grant AST-0607510. JSM acknowledges partial support for this work from NASA grant NNG04GC846.
0.2in
Bertin, E. & Arnouts, S. 1996, A&AS, 117, 393
Blanton, M. R. [$\mbox{et~al.}$]{} 2003, ApJ 592, 819
Bond, N. A., Churchill, C. W., Charlton, J. C., & Vogt, S. S. 2001, ApJ, 557, 761
Bournaud, F., Duc, P.-A., Amram, P., Combes, F., & Gach, J.-L. 2004, A&A, 425, 813
Carswell, R. F., Lanzetta, K. M., Parnell, H. C., & Webb, J. K. 1991, ApJ, 371, 36
Cen, R. & Ostriker, J. P. 2006, 650, 573
Chen, H.-W., Lanzetta, K. M., Webb, J. K., & Barcons, X. 1998, ApJ, 498, 77
[.]{} 2001a, ApJ, 559, 654
Chen, H.-W. & Prochaska, J.X. 2000, ApJ, 543, L9
Chen, H.-W., Lanzetta, K. M., & Webb, J. K. 2001b, ApJ, 556, 158
Chen, H.-W., Prochaska, J. X., Weiner, B. J., Mulchaey, J. S., & Williger, G. M. 2005, ApJ, 629, L25
Chen, H.-W. & Tinker, J. L. 2008, ApJ, 687, 745
Churchill, C. W., Kacprzak, G. G., Steidel, C. C., & Evans, J. L. 2007, ApJ, 661, 714
Coleman, G. D., Wu, C. C., & Weedman, D. W. 1980, ApJS, 43, 393
Cooksey, K. L., Prochaska, J. X., Chen, H.-W., Mulchaey, J. S., & Weiner, B. J. 2008, ApJ, 676, 262
Dav[' e]{}, R., Cen, R., Ostriker, J. P., Bryan, G. L., Hernquist, L., Katz, N., Weinberg, D. H., Norman, M. L., & O’Shea, B. 2001, ApJ, 552, 473
Dobrzycki, A., Bechtold, J., Scott, J., & Morita, M. 2002, ApJ, 571, 654
Dressler, A., Hare, T., Bigelow, B. C., & Osip, D. J. 2006, Proc. SPIE, 6269, 13
Elmegreen, D. M., Elmegreen, B. G., Ferguson, T., & Mullan, B. 2007, ApJ, 663, 734
Fukugita, M., Hogan, C.J, & Peebles, P.J.E. 1998, Nature, 503, 518
Fukugita, M. 2004, in IAU Symp. 220, Dark Matter in Galaxies, ed. S. D. Ryder et al. (San Francisco: ASP), 227
Grogin, N. A. & Geller, M. J. 1998, 505, 506
Howk, J. C., Ribaudo, J. S., Lehner, N., Prochaska, J. X., & Chen, H.-W. 2009, ApJ submitted
Jannuzi, B. T. [$\mbox{et~al.}$]{} 1998, ApJS, 118, 1
Kawata, D. & Rauch, M. 2007, ApJ, 663, 38
Landolt, A.U. 1992, AJ, 104, 340
Landy, S. D. & Szalay, A. S. 1993, ApJ, 412, 64
Lanzetta, K. M., Bowen, D. V., Tytler, D., & Webb, J. K. 1995, ApJ, 442, 538
Lehner, N., Savage, B. D., Wakker, B. P., Sembach, K. R., & Tripp, T. M. 2006, ApJS, 164, 1
Lehner, N., Savage, B. D., Richter, P., Sembach, K. R., Tripp, T. M., & Wakker, B. P. 2007, ApJ, 658, 680
Lehner, N., Prochaska, J. X., Kobulnicky, H. A., Cooksey, K. L., Howk, J. C., Williger, G. M., & Cales, S. L. 2009, ApJ in press (arXiv:0812,4231)
Li, C., Kauffmann, G., Heckman, T. M., White, S. D. M., & Jing, Y. P. 2008, MNRAS, 385, 1915
Madgwick, D. S. [$\mbox{et~al.}$]{} 2003, MNRAS, 344, 847
, A. H. & [Bullock]{}, J. S. 2004, MNRAS, 355, 694
, H. J. & [Miralda-Escudé]{}, J. 1996, ApJ, 469, 589
Mo, H. J. & White, S. D. M. 2002, MNRAS, 336, 112
Morris, S. L., Weymann, R. J., Dressler, A., McCarthy, P. J., Smith, B. A., Terrile, R. J., Giovanelli, R., & Irwin, M. 1993, ApJ, 419, 524
Mulchaey, J. S., Mushotzky, R. F., Burstein, D., & Davis, D. S. 1996, ApJ, 456, L5
Mulchaey, J. S. & Chen, H.-W. 2009, ApJ submitted
Oppenheimer, B. D. & Davé, R. 2009, MNRAS submitted (arXiv:0806.2866)
Penton, S. V., Stocke, J. T., & Shull, J. M. 2002, ApJ, 565, 720
[.]{} 2004, ApJS, 152, 29
Persic, M. & Salucci, P. 1992, MNRAS, 258, 14
Pickles, A. J. 1998, PASP, 110, 863
Prochaska, J. X., Chen, H.-W., Howk, J. C., Weiner, B. J., & Mulchaey, J. S. 2004, ApJ, 617, 718
Prochaska, J. X., Weiner, B. J., Chen, H.-W., & Mulchaey, J. S. 2006, ApJ, 642, 989
Rauch, M., Miralda-Escudé, J., Sargent, W. L. W., Barlow, T. A., Weinberg, D. H., Hernquist, L., Katz, N., Cen, R., Ostriker, J. P. 1997, ApJ, 489, 7
Rauch, M. 1998, ARA&A, 36, 267
Savage, B. D., Lehner, N., Wakker, B. P., Sembach, K. R., & Tripp, T. M. 2005, ApJ, 626, 776
Scott, J. E., Kriss, G. A., Brotherton, M., Green, R. F., Hutchings, J., Shull, J. M., & Zheng, W. 2004, ApJ, 615, 135
Sembach, K. R., Wakker, B. P., Savage, B. D., Richter, P., Meade, M., Shull, J. M., Jenkins, E. B., Sonneborn, G., & Moos, H. W. 2003, ApJS, 146, 165
Sheth, R. K., Connolly, A. J., & Skibba, R. 2005, astro-ph/0511773
Simcoe, R. A., Sargent, W. L. W., & Rauch, M. 2002, ApJ, 578, 737
Spinrad, H., Filippenko, A. V., Yee, H. K., Ellingson, E., Blades, J. C., Bahcall, J. N., Jannuzi, B. T., Bechtold, J., & Dobrzycki, A. 1993, AJ, 106, 1
, C. C., [Dickinson]{}, M., & [Persson]{}, S. E. 1994, ApJL, 437, L75
Stocke, J. T., Shull, J. M., Penton, S., Donahue, M., & Carilli, C. 1995, ApJ, 451, 24
Stocke, J. T., Penton, S. V., Danforth, C. W., Shull, J. M., Tumlinson, J., & McLin, K. M. 2006, ApJ, 641, 217
Thom, C. & Chen, H.-W. 2008a, ApJ, 683, 22
[.]{} 2008b, ApJS, 179, 37
Thom, C., Peek, J. E., Putman, M. E., Heiles, C., Peek, K. M. G., & Wilhelm, R. 2008, ApJ, 684, 364
Tinker, J. L., Weinberg, D. H., Zheng, Z., & Zehavi, I. 2005, ApJ, 631, 41
Tinker, J. L., Norberg, P., Weinberg, D. H., & Warren, M. S. 2007, ApJ, 659, 877
Tinker, J. L. & Chen, H.-W. 2008, ApJ, 679, 1218
Tripp, T. M., Savage, B. D., & Jenkins, E. B. 2000, ApJ, 534, L1
Tripp, T. M., Sembach, K. R., Bowen, D. V., Savage, B. D., Jenkins, E. B., Lehner, N., & Richter, P. 2008, ApJS, 177, 39
Verner, D.A., Tytler, D., & Barthel, P.D. 1994, ApJ, 430, 186
Wang, B. 1993, ApJ, 415, 174
Weymann, R. J. [$\mbox{et~al.}$]{} 1998, ApJ, 506, 1
Williger, G. M. [$\mbox{et~al.}$]{} 2006, ApJ, 636, 631
Wilman, R. J., Morris, S. L., Jannuzi, B. T., Davé, R., & Shone, A. M., 2007, MNRAS, 375, 735
Yip, C. W. [$\mbox{et~al.}$]{} 2004, AJ, 128, 585
York, D. G. [$\mbox{et~al.}$]{} 2000, AJ, 120, 1579
Zehavi, I. [$\mbox{et~al.}$]{} 2005, ApJ, 630, 1
Zheng, Z., Coil, A. L., & Zehavi, I. 2007, ApJ, 667, 760
[^1]: We note that the field around PG1216$+$069 is covered by the Sloan Digital Sky Survey (SDSS; York [$\mbox{et~al.}$]{} 2000), which has a targeted imaging depth of $r'=22.2$ (in the AB magnitude system). However, because of relatively large point spread functions (PSFs) with a mean full-width-half-maximum $\langle {\rm FWHM}\rangle=1.4''$, the available SDSS imaging data become significantly incomplete for detecting distant faint galaxies near the magnitude limit and for resolving objects near the QSO. Additional imaging data are therefore necessary for the purpose of our study.
[^2]: Some objects were duplicated in different masks.
[^3]: http://www.lco.cl/lco/telescopes-information/magellan/magellan/instruments/imacs/cosmos/cosmos.html/view
[^4]: See http://www.sdss.org/dr7/algorithms/spectemplates/index.html
[^5]: see http://www.ast.cam.ac.uk/’176rfc/vpfit.html.
[^6]: The rest-frame $R$-band magnitude of each spectroscopically identified object was estimated based on the observed $R$-band magnitude and spectral type. For absorption-line dominated galaxies, we evaluate the $k$-corretion using the early-type E/S0 and Sab galaxy templates from Coleman [$\mbox{et~al.}$]{} (1980). For emission-line dominated galaxies, we evaluate the $k$-correction using the Irr templates.
[^7]: We note that Tripp [$\mbox{et~al.}$]{} (2008) reported one more absorber at $z=0.26768$ along this sightline. However, the expected OVI 1037 transition falls at a wavelength that has been contaminated by a saturated feature at $z=0.2819$. No other metal absorption lines are found for the absorber to confirm the nature of the presumed OVI 1031 feature. We therefore consider this absorption feature as at $z=0.0760$.
[^8]: We note that on small scales ($r_p{\lesssim}1\
h^{-1}$ Mpc) the correlation amplitude is dominated by satellite galaxies (e.g. Zheng [$\mbox{et~al.}$]{} 2007). Errors in the measured clustering signals on these scales are expected to be roughly poisson counting errors. On larger scales, however, field-to-field variations are expected to dominate the errors in the observed clustering signals and the reported poisson errors represent a lower limit to the true uncertainties. Given that our survey covers only three QSO fields, we are unable to evaluate the uncertainties due to field-to-field variations using a jackknife method.
[^9]: In the absence of peculiar velocity field, $\Delta\,v = \pm 300$ [$\rm{km\,s}^{-1}$]{} would correspond to a co-moving pathlength of $\pm 3.3\ h^{-1}$ Mpc. Adopting the luminosity function of Blanton et al. (2003) and the two-point galaxy auto-correlation function of Zehavi et al. (2005), we estimate that the probability of finding a random galaxy of $L > 0.1\,L_*$ at $r_p < 100\ h^{-1}$ co-moving kpc due to large-scale overdensity is $\approx 3$%. The probability increases to 20% for $r_p < 250\ h^{-1}$ kpc.
[^10]: We have excluded the absorber at $z=0.01746$ toward HE0226$-$4110, because at this low redshift our spectroscopic survey area only covers $\approx
170\ h^{-1}$ kpc co-moving radius of the absorber.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $K$ be a field and let $A=K[X_1, \ldots, X_n]$ be the polynomial ring in $X_1, \ldots, X_n$ with coefficients in the field $K$. We study the universal squarefree lexsegment ideals. We put our attention on their combinatorics computing some invariants. Moreover we study the link between such special class of squarefree lexsegment ideals and the so called $s$-sequences.\
address:
- |
University of Messina, Department of Mathematics and Computer Science\
Viale Ferdinando Stagno d’Alcontres, 31\
98166 Messina, Italy
- |
University of Messina, Department of Mathematics and Computer Science\
Viale Ferdinando Stagno d’Alcontres, 31\
98166 Messina, Italy
author:
- Marilena Crupi
- Monica La Barbiera
title: Algebraic properties of universal squarefree lexsegment ideals $ ^1$
---
[^1]
Introduction {#sec0 .unnumbered}
============
Let $K$ be a field and let $A=K[X_1, \ldots, X_n]$ be the polynomial ring in $X_1, \ldots, X_n$ with coefficients in the field $K$. Set $A_{[m]} = K[X_1, \cdots, X_n,X_{n+1},\cdots, X_{n+m}]$, where $m$ is a positive integer. A squarefree lexsegment ideal $I$ of $A$ is called *universal squarefree lexsegment ideal* (abbreviated USLI), if for any integer $m\geq 1$, the squarefree monomial ideal $IA_{[m]}$ of the polynomial ring $A_{[m]}$ is a squarefree lexsegment ideal. Let $M^s$ denote the set of all squarefree monomials in the variables $X_1, \cdots, X_n$. A squarefree lexsegment ideal $I$ of $A$ with $G(I) = \{u_1, \ldots, u_{\ell}, u_{{\ell}+1}\}$, $u_1> \ldots> u_{\ell}> u_{{\ell}+1}$ with respect to the homogeneous lexicographic order on $M^s$, is called *almost universal squarefree lexsegment ideal* (abbreviated AUSLI), if $I$ is not an USLI of $A$ but the ideal $J = (u_1, \ldots, u_{\ell})$ is an USLI of $A$. $G(I)$ is the unique minimal set of monomial generators of the monomial ideal $I$. These definitions were introduced by Babson, Novik and Thomas in [@BNT] in order to study the simmetric version of algebraic shifting. The algebraic shifting is an algebraic operation introduced by Kalai ([@BK], [@GK]) that transforms a simplicial complex into a simpler complex that preserves important combinatorial, topological and algebraic invariants.
In this paper we put our attention on the structure of universal squarefree lexsegment ideals (Characterization \[char\]). We analyze their combinatorics in order to compute some invariants as the projective dimension, the Castelnuovo-Mumford regularity and the depth (Corollary \[inv\]).
In [@HRT], the authors introduced the concept of $s$-sequences in order to study the symmetric algebra of a module $M$ on a noetherian ring $R$. One of their motivation was that is a difficult problem to compute standard algebraic invariants of the graded algebra ${{\operatorname{Sym}}}_R(M)$. Their proposal was to determine these invariants in terms of the corresponding invariants of special quotients of the ring $R$. The $s$-sequences are an important tool for this computation.
In this paper we analyze the problem when a squarefree lexsegment ideal $I$ of degree $d$ of the polynomial ring $A$ is generated by an $s$-sequence. We are able to state that this happens if $I$ is an USLI or an AUSLI (Theorem \[T\*\]). Consequently their symmetric algebra is studied (Theorems \[inv:USLI\] and \[inv:ASLI\]).\
The structure of the paper is organized as follows.\
In section \[sec1\], we recall some notions that we will use during the paper.\
In section \[sec2\], we describe in a suitable way the USLIs (Characterization \[char\]). Hence we state a characterization of an USLI of degree $d$ (Proposition \[samedeg\]). Moreover, we analyze some invariants associated to the universal squarefree lexsegment ideals. The main result states that an USLI $I\varsubsetneq A$ has a unique extremal Betti number whose value is $1$ (Proposition \[exU\]). This fact allows us to compute ${{\operatorname{proj}}}_A(I)$, ${{\operatorname{reg}}}_A(I)$ and ${{\operatorname{depth}}}_A(A/I)$ (Corollary \[inv\]).\
Section \[sec3\] is dedicated to the symmetric algebra of USLIs and AUSLs of the polynomial ring $A$. More precisely let $I$ be a lexsegment ideal of $A$ generated by squarefree monomials in a same degree, we establish that the ideal $I$ is generated by an $s$-sequence if and only if $I$ is an USLI or an AUSLI (Theorem \[T\*\]). This result is proved using the characterization of the monomial $s$-sequences by the Gröbner bases. As a consequence of this result we study the problem of computing standard algebraic invariants of the graded algebra ${{\operatorname{Sym}}}_A(I)$ when $I$ is a squarefree lexsegment ideal generated by an $s$-sequence. More precisely, we give a formula for the dimension and the multiplicity of ${{\operatorname{Sym}}}_A(I)$ when $I$ is an AUSLI (Theorem \[inv:ASLI\]). A formula for the Castelnuovo-Mumford regularity and for the depth of ${{\operatorname{Sym}}}_A(I)$ when $I$ is an USLI (Theorem \[inv:USLI\]) is also stated.
Preliminaries and notations {#sec1}
===========================
Let $K$ be a field and let $A=K[X_1, \ldots, X_n]$ be the polynomial ring in $X_1, \ldots, X_n$ with coefficients in the field $K$. We consider $A$ as an $\NN$-graded ring and each deg$X_i$ = $1$. We denote by $M_d$ the set of all monomials of degree $d$ of the polynomial ring $A$. If $I\subsetneq A$ is a monomial ideal we denote by $G(I)$ the unique minimal set of monomial generators of $I$ and by $G(I)_d$ the set $G(I)_d = \{u \in G(I)\,:\, \deg u = d\}$, for $d>0$.
For a monomial $1 \neq u \in A$, we set $${{\operatorname{supp}}}(u) = \{i \,:\, X_i \,\, \textrm{divides} \,\, u\}.$$ $${{\operatorname{m}}}(u) = \max\{i : X_i \,\, \textrm{divides}\,\, u\}.$$
Recall that a squarefree monomial ideal $I$ of $A$ is called *squarefree stable* if for all $u\in G(I)$, one has $(X_ju)/X_{{{\operatorname{m}}}(u)} \in I$ for all $j <{{\operatorname{m}}}(u)$ with $j \notin {{\operatorname{supp}}}(u)$ ([@Ahh:square]).
Now let $M^s_d$ denote the set of all squarefree monomials of degree $d\geq 1$ in the variables $X_1, \cdots, X_n$. We write $>_{\textrm{slex}}$ for the *lexicographic order* on the finite set $M^s_d$, that is, if $u = X_{i_1}\cdots X_{i_d}$ and $v = X_{j_1}\cdots X_{j_d}$ are squarefree monomials belonging to $M^s_d$ with $1 \leq i_1 < i_2 < \cdots < i_d \leq n$ and $1 \leq j_1 < j_2 < \cdots < j_d \leq n$, then $u>_{\textrm{slex}} v$ if $i_1 = j_1$, $\ldots$, $i_{s-1} = j_{s-1}$ and $i_s < j_s$ for some $1 \leq s \leq d$ ([@Ahh:square]).\
Let $M^s$ be the set of all squarefree monomials in the variables $X_1, \cdots, X_n$. We denote by $>_{\textrm{hslex}}$ the *homogeneous lexicographic order* on $M^s$, that is, if $u,v \in M^s$, then $u >_{\textrm{hslex}} v$ if $\deg u > \deg v$, or if $\deg u = \deg v$ and $u>_{\textrm{slex}} v$.\
A monomial ideal $I \subsetneq A$ is called a *squarefree lexsegment ideal* if $I$ is generated by squarefree monomials, and for all squarefree monomials $u \in I$ and all squarefree monomials $v\in A$ with $\deg u = \deg v$ and $v >_{\textrm{slex}} u$, then $v \in I$. Every squarefree lexsegment ideal of $A$ is obviously a squarefree stable ideal.
Set $A_{[m]} = K[X_1, \cdots, X_n,X_{n+1},\cdots, X_{n+m}]$, where $m$ is a positive integer.
We quote the next definitions from [@BNT].
A squarefree lexsegment ideal $I$ of $A$ is called *universal squarefree lexsegment ideal* (USLI, for short), if for any integer $m\geq 1$, the squarefree monomial ideal $IA_{[m]}$ of the polynomial ring $A_{[m]}$ is a squarefree lexsegment ideal.
In other words a universal squarefree lexsegment ideal of $A$ is a squarefree lexsegment ideal $I$ of $A$ which remains being squarefree lexsegment if we regard $I$ as an ideal of the polynomial ring $A_{[m]}$ for all $m \geq 1$.
\[expls\] (1) The squarefree lexsegment ideal $I = (X_1X_2, X_1X_3X_4)$ of $A = K[X_1,X_2,X_3,X_4]$ is an USLI. Indeed $I$ is a squarefree lexsegment ideal of the polynomial ring $A_{[m]}$ for all $m \geq 1$.\
(2) The squarefree lexsegment ideal $I = (X_1X_2, X_1X_3X_4, X_2X_3X_4)$ of $A = K[X_1,X_2,$$X_3,X_4]$ is not an USLI. Indeed $I$ is not a squarefree lexsegment ideal of the polynomial ring $A_{[1]} = K[X_1,X_2,X_3,X_4,X_5]$. In fact $X_1X_4X_5 >_{\textrm{slex}} X_2X_3X_4$ and $X_1X_4X_5 \notin IA_{[1]}$.
A squarefree lexsegment ideal $I$ of $A$ with $G(I) = \{u_1, \ldots, u_\ell, u_{\ell+1}\}$, $u_1>_{\textrm{hslex}} \ldots>_{\textrm{hslex}} u_{\ell}>_{\textrm{hslex}} u_{{\ell}+1}$, is called *almost universal squarefree lexsegment ideal* (AUSLI, for short), if $I$ is not an USLI of $A$ but the ideal $J = (u_1, \ldots, u_{\ell})$ is an USLI of $A$.
The squarefree lexsegment ideal $I = (X_1X_2, X_1X_3X_4, X_2X_3X_4)$ of $A = k[X_1,X_2,X_3,X_4]$ is an AUSLI. Indeed $I$ is not an USLI of $A$, but the ideal $J = (X_1X_2, X_1X_3X_4)$ is an USLI of $A$ (Example \[expls\], (1)).
We finish this section recall the notion of extremal Betti numbers of a graded ideal $I$ of the polynomial ring $A$.
If $I$ is a graded ideal of $A$, then $I$ has a minimal graded free $A$-resolution $$F. : 0 \rightarrow F_s \rightarrow \cdots \rightarrow F_1
\rightarrow F_0 \rightarrow I \rightarrow 0$$ where $F_i = \oplus_{j \in \ZZ}A(-j)^{\beta_{i,j}}$.
The integers $\beta_{i,j} = \beta_{i,j}(I) = \textrm{dim}_K
\textrm{Tor}_i(K, I)_j $ are called the graded Betti numbers of $I$, while $\beta_i(I) = \sum_j \beta_{i,j}(I)$ are called the total Betti numbers of $I$.
To a graded ideal $I$ two invariants can be associated the [*projective dimension*]{} and the [*Castelnuovo-Mumford regularity*]{} ([@BH], [@Eis]) that are defined, respectively, as follows: $${{\operatorname{proj}}}_A(I) = \max\{i : \beta_i(I) \neq 0\},$$ $$\begin{array}{lll}
\textrm{reg}_A(I) = \max\{j-i : \beta_{i, j}(I) \neq 0\} = \max\{j
: \beta_{i,i+j}(I) \neq 0,\,\, \mbox{for some $i\in \NN$}\}.
\end{array}$$
Bayer, Charalambous and Popescu introduced in [@BCP] a refinement of the invariants above defined, giving the notion of extremal Betti numbers.
\[def:extr\] A Betti number $\beta_{k,k+\ell}(I) \neq 0$ is called [*extremal*]{} if $\beta_{i, i+j}(I) = 0$ for all $i \geq k$, $j \geq
\ell$, $(i, j) \neq (k, \ell)$.
If $\beta_{k_1,k_1+\ell_1}(I), \dots,
\beta_{k_t,k_t+\ell_t}(I),\,\,\,k_1 > \dots > k_t, \ell_1 <\dots <
\ell_t$, are all extremal Betti numbers of $I$, then $k_1 =
{{\operatorname{proj}}}_A(I)$ and $\ell_t = \textrm{reg}_A(I)$.
The following characterization of the extremal Betti numbers of squarefree stable ideals was given in [@CU2 Proposition 4.1].
\[pro:extr\] Let $I\subsetneq A$ be a squarefree stable ideal. The following conditions are equivalent:
1. $\beta_{k,\,k+\ell}(I)$ is extremal.
2. $k+\ell = \max\{{{\operatorname{m}}}(u)\,:\, u \in G(I)_{\ell}\}$ and ${{\operatorname{m}}}(u) < k +j$ for all $j>\ell$ and $u \in G(I)_j$.
As a consequence of the above result, we obtain the following.
\[value\] Let $I\subsetneq A$ be a squarefree stable ideal.
1. If $\beta_{k, \, k+ \ell}(I)$ is an extremal Betti number of $I$, then $$\beta_{k,\,k+\ell}(I) = \vert \{u \in G(I)_{\ell}\,:\,{{\operatorname{m}}}(u) = k +\ell\}\vert.$$
2. Set $d = \max\{j\,:\, G(I)_j\neq \emptyset\}$ and $m = \max\{{{\operatorname{m}}}(u)\,:\,u\in G(I)\}$, then $\beta_{m - d,\,m - d+d}(I)$ is the unique extremal Betti number of $I$ if and only if $m=\max\{{{\operatorname{m}}}(u)\,:\,u\in G(I)_d\}$ and for every $w\in G(I)_j$, $j<d$, ${{\operatorname{m}}}(w)<m$.
Universal squarefree lexsegment ideals {#sec2}
======================================
In this section we discuss the combinatorics of universal squarefree lexsegment ideals. Moreover we compute some standard invariants.
In [@BNT Definition 4.1] there is a characterization of USLIs. In order to reformulate it for our purpose, we need to introduce some notations.
For a sequence of non negative integers $(k_i)_{i\in \NN}$, we define the following set: $${{\operatorname{supp}}}(k_i)_{i\in \NN} = \{i \in \NN\,:\, k_i \neq 0\}.$$
If ${{\operatorname{supp}}}(k_i)_{i\in \NN} = \{d_1, \ldots, d_t\}$, with $d_1 < d_2 < \ldots < d_t$, then we associate to $(k_i)_{i\in \NN}$ the following integers: $$R_j = j + \sum_{i=1}^j k_i$$ for $j= 1,\ldots, d_t$. We set $R_j = 0$, for $j >d_t$.
Hence we can reformulate the characterization contained in [@BNT Definition 4.1], as follows:
\[char\] Let $I\subsetneq A$ be an ideal generated in degrees $d_ 1< d_2 < \ldots < d_t$. Then $I$ is an USLI of $A$ if and only if $$\mbox{$G(I)_{d_i} = \left\{\left(\prod_{j=1}^{d_i-1}X_{R_j}\right)X_{\ell}\,:\, \ell \in [R_{d_i-1} +1, R_{d_i}-1]\right\}$, for \, $i=1, \ldots, t$,}$$ where $R_j = j + \sum_{i=1}^j \vert G(I)_{d_i}\vert$, for $j= 1,\ldots, d_t$.
The characterization above follows from the statement contained in [@BNT Definition 4.1], choosing $(k_i)_{i\in \NN}$ as the sequence of non negative integers such that ${{\operatorname{supp}}}(k_i)_{i\in \NN} = \{d_1, \ldots, d_t\}$ and $k_{d_i} = \vert G(I)_{d_i}\vert$, for $i = 1, \ldots, t$.
\[seq\] Assume that $(k_i)_{i\in \NN}$ is a sequence of non negative integers such that $${{\operatorname{supp}}}(k_i)_{i\in \NN} = \{d_1, \ldots, d_t\},\quad d_ 1< d_2 < \ldots < d_t.$$ Then there exists an USLI $I\subsetneq A= K[X_1, \ldots, X_n]$ generated in degrees $d_1, \ldots, d_t$ such that $\vert G(I)_{d_i}\vert = k_{d_i}$, for $i = 1, \ldots, t$ if and only if $n \geq d_t + \sum_{i=1}^{d_t}k_i-1$.
Thanks to the above statements we can give the following characterizations of an USLI generated in a same degree $d$.
\[samedeg\] Let $I$ be a squarefree lexsegment ideal of $A = K[X_1, \ldots, X_n]$ generated in degree $d$. Then $I$ is an USLI of $A$ if and only if $\vert G(I)\vert\leq n-d+1$.
Let $(k_i)_{i\in \NN}$ be the sequence of non negative integers such that ${{\operatorname{supp}}}(k_i)_{i\in \NN} = \{d\}$, with $k_{d} = \vert G(I)_{d}\vert = \vert G(I)\vert$. From Remark \[seq\], $I\subsetneq A$ is an USLI generated in degree $d$ if and only if $n \geq d + k_d -1 = d + \vert G(I)\vert-1$ that is if and only if $\vert G(I)\vert\leq n-d+1$.
Note that if $I\subsetneq A = K[X_1, \ldots, X_n]$ is an USLI generated in degree $d$ then $$\label{degree1}
G(I) = \{X_1X_2\cdots X_{d-1}X_d, X_1X_2\cdots X_{d-1}X_{d+1}, \ldots, X_1X_2\cdots X_{d-1}X_k\},$$ with $d \leq k \leq n$.
Moreover if $I$ is an AUSLI generated in degree $d$, then $$\label{degree2}
G(I) = \{X_1X_2\cdots X_{d-1}X_d, \ldots, X_1X_2\cdots X_{d-1}X_n,X_1X_2\cdots X_{d-2}X_dX_{d+1}\}.$$
\[ASLI\] It is clear that a squarefree lexsegment ideal $I\varsubsetneq A$ generated in degree $d$ is an AUSLI if and only if $\vert G(I) \vert = n-d+2$.
We finish this section computing some invariants of an USLI $I$ by its extremal Betti numbers. In general squarefree lexsegment ideals may have more than just one extremal Betti number ([@CU1],[@CU2]).\
Consider, for example the squarefree ideal $$I=(X_1X_2,X_1X_3,X_1X_4,X_1X_5, X_1X_6,X_1X_7,X_2X_3X_4,X_2X_3X_5,X_2X_3X_6, X_2X_3X_7,$$ $$X_2X_4X_5X_6,X_2X_4X_5X_7, X_3X_4X_5X_6X_7)$$ of $K[X_1,\ldots, X_7]$. It is a squarefree lexsegment ideal with $\beta_{5,5+2}=1$, $\beta_{4,4+3}=1$, $\beta_{3,3+4}=1$, $\beta_{2,2+5}=1$ as extremal Betti numbers.\
\
For an USLI, we can state.
\[exU\] Let $I\subsetneq A$ be an USLI. Then $I$ has an unique extremal Betti number whose value is equal to $1$.
Let $I$ be an USLI generated in degrees $d_ 1< d_2 < \ldots < d_t$. From Theorem \[char\], $\ell = R_{d_t} -1 = \max\{{{\operatorname{m}}}(u)\,:\,u\in G(I)_{d_t}\}$, where $d_t = \max\{j\,:\, G(I)_j\neq \emptyset\}$ and $\ell = \max\{{{\operatorname{m}}}(u)\,:\,u\in G(I)\}$. Hence, from Corollary \[value\], $\beta_{\ell - d_t,\,\ell - d_t+d_t}(I)$ is its unique extremal Betti number and its value is $1$.
\[inv\] Let $I\subsetneq A$ be an USLI generated in degrees $d_ 1< d_2 < \ldots < d_t$. Then
1. ${{\operatorname{proj}}}_A(I) = \vert G(I)\vert -1$ and ${{\operatorname{reg}}}_A(I) = d_t$.
2. ${{\operatorname{depth}}}_A(A/I) = n - \vert G(I) \vert$.
(1). From Proposition \[exU\], $\beta_{\ell - d_t,\,\ell - d_t+d_t}(I)$, with $\ell =\max\{{{\operatorname{m}}}(u)\,:\,u\in G(I)_{d_t}\}$, is the unique extremal Betti number of $I$. Hence $\ell-d_t = {{\operatorname{proj}}}_A(I)$ and $d_t ={{\operatorname{reg}}}_A(I)$. With the same notations of Characterization \[char\], we have that $${{\operatorname{proj}}}_A(I) = \ell-d_t = d_t + \sum_{i=1}^{t}\vert G(I)_{d_i} \vert- d_t-1= \vert G(I)\vert-1.$$ (2). It follows from the Auslander-Buchsbaum formula.
Recall that for a squarefree stable ideal $I\varsubsetneq A$, ${{\operatorname{reg}}}_A(I) = \max\{\deg u: $$u \in G(I)\}$ [@Ahh:square Corollary 2.6].
USLIs, AUSLIs and $s$-sequences {#sec3}
===============================
In this section we study the strict link between USLIs (resp. ASLIs) and $s$-sequences. We compute standard algebraic invariants of the graded algebra $Sym_A (I)$ when $I$ is an USLI or an AUSLI of degree $d$ in terms of the annihilator ideals of the $s$-sequence that generates $I$.
Let $A$ be a noetherian ring, $M$ be a finitely generated $A$-module with generators $f_1, \ldots, f_q$. For every $i = 1, \dots , q$, we set $M_{i-1} = Af_1 + \cdots + Af_{i-1}$ and let $I_i =
M_{i-1}:_A f_i$ be the colon ideal. We set $I_0 = (0)$. Since $M_i/M_{i-1} \simeq A/I_i$, so $I_i$ is the annihilator of the cyclic module $A/I_i$. $I_i$ is called an annihilator ideal of the sequence $f_1, \dots, f_q$.\
Let $Sym_A(M)$ be the symmetric algebra of $M$. Let $(a_{ij})$, for $i=1,\ldots,q$, $j=1,\ldots,p$, be the relation matrix of $M$. It is known that the symmetric algebra $Sym_A(M)$ has a presentation $A[T_1,\ldots,T_q]/J$, with $J=(g_1,\ldots,g_p)$ where $g_j=\sum_{i=1}^q a_{ij}T_i$ for $j=1, \ldots, p$.\
We consider $S=A[T_1, \dots, T_q]$ a graded ring by assigning to each variable $T_i$ degree $1$ and to the elements of $A$ degree $0$. Then $J$ is a graded ideal and the natural epimorphism $S\rightarrow Sym_A(M)$ is a homomorphism of graded $A$-algebras.\
Let $<$ be a monomial order on the monomials in the variables $T_i$ such that $T_1 < T_2 < \cdots < T_q$. With respect to this term order, for any polynomial $f = \sum a_{\alpha}
\underline{T}^{\alpha} \in S$, where $\underline{T}^{\alpha} =
T_1^{\alpha_1} \cdots T_q^{\alpha_q} $ and $\alpha=(\alpha_1,\ldots,
\alpha_q) \in \NN^q$, we put $ \textit{in}_{<} (f) =
a_{\alpha}\underline{T}^{\alpha}$, where $\underline{T}^{\alpha}$ is the largest monomial in $f$ such that $a_{\alpha} \neq 0$.\
So we can define the monomial ideal $\textit{in}_{<}(J)=(\textit{in}_{<}(f)| f\in J)$. Notice that $(I_1T_1, I_2 T_2, \dots, I_q T_q) \subseteq
\textit{in}_{<}(J)$ and the two ideals coincide in degree $1$.
The generators $f_1, \dots, f_q$ of $M$ are called an $s$-sequence (with respect to an admissible term order $<$) if $$(I_1T_1, I_2 T_2, \dots, I_q T_q) = in_{<}(J).$$ If $I_1 \subseteq I_2 \subseteq \cdots \subseteq I_q$, the sequence is a strong $s$-sequence.
Now, let $A = K[X_1,\dots, X_n]$ be the polynomial ring over a field $K$ and let $<$ any term order on $K[X_1,\dots, X_n;T_1,\dots, T_q]$ with $X_1 >
\cdots > X_n$, $T_1 < T_2 < \cdots < T_q$, $X_i< T_j$ for all $i$ and $j$. Then for any Gröbner basis $G$ of $J \varsubsetneq K[X_1, \dots, X_n; T_1,
\dots,T_q]$ with respect to $<$, we have $\textrm{in}_{<}(J)
=
(\textrm{in}_{<} (f) | f \in G)$. If the elements of $G$ are of degree $1$ in the $T_i$, it follows that $f_1, \dots, f_q$ is an $s$-sequence of $M$.
Let $f_1,\ldots,f_q$ be monomials of $A$. Set $f_{ij}=\displaystyle{\frac{f_i}{[f_i,f_j]}}$ for $i \neq j$, where $[f_i,f_j]$ is the greatest common divisor of the monomials $f_i$ and $f_j$. $J$ is generated by $g_{ij}=f_{ij}T_j - f_{ji}T_i$ for $1\leq i< j \leq q$. The monomial sequence $f_1,\ldots,f_q$ is an $s$-sequence if and only if $g_{ij}$, for $1\leq i< j \leq q$, is a Gröbner basis for $J$ for any term order which extends an admissible term order on the $T_i$ in $S= K[X_1,\dots, X_n;T_1,\dots, T_q]$. Note that the annihilator ideals of the monomial sequence $f_1,\ldots,f_q$ are the ideals $I_i=(f_{1i},f_{2i}, \ldots,
f_{i-1,i})$ for $i=1,\ldots, q$.
Let $I$ be an ideal of $A$ generated by an $s$-sequence $f_1,\dots,
f_q$ of monomials with respect to some admissible term order $<$. From the theory of Gröbner bases, one has that $f_1, \dots, f_q$ is an $s$-sequence with respect to any other admissible term order ([@HRT], Lemma 1.2).
For more details on this subject see [@HRT].
Since the property to be an $s$-sequence may depend on the order on the sequence, if $I$ is a squarefree lexicographic ideal when we write $$I=(f_1,f_2, \ldots, f_q),$$ we suppose $$f_1>_{\textrm{hslex}} f_2 >_{\textrm{hslex}} \cdots >_{\textrm{hslex}} f_q.$$ In order to simply the notations we will denote $>_{\textrm{hslex}}$ by $>$.\
For any positive integer $q$ we set $[q] = \{1, \ldots, q\}$.
The following lemma will be crucial in the sequel.
\[L1\] Let $I=(f_1,\ldots,f_q)\varsubsetneq A$ be a squarefree lexsegment ideal generated in degree $d$ with $G(I) \varsubsetneq M^s_d$.
The following conditions are equivalent:
1. $[f_{ij},
f_{h\ell}]=1$, for $i<j$, $h<\ell$, $i\neq h$, $j\neq \ell$, $i,j,h,\ell \in [q]$.
2. $|G(I)|\leq n-d+2$.
3. $I$ is an USLI or $I$ is an ASLI of $A$.
From Corollary \[samedeg\] and Remark \[ASLI\], conditions (2) and (3) are equivalent, then we have only to prove that (1)$\Leftrightarrow$(2).\
(1)$\Rightarrow$(2). Note that the monomial generators $f_i$ of $I$ are described by (\[degree2\]), for $i=1, \ldots, n-d+2$. Suppose $|G(I)|>n-d+2$. Since $I$ is a squarefree lexsegment ideal of degree $d$, then $X_1X_2\cdots
X_{d-2}X_dX_{d+2}\in G(I)$.\
From (\[degree2\]), set $t = n-d+2$ and $t' = t+1$, then $f_t
=X_1X_2\cdots X_{d-2}X_dX_{d+1}$ and $f_{t'} = X_1X_2\cdots X_{d-2}X_dX_{d+2}$. Hence $f_{tt'} = X_{d+1}$. Again from (\[degree2\]), $f_{23} = X_{d+1}$. Hence $[f_{23}, f_{tt'}] = X_{d+1}$. A contradiction.\
(2)$\Rightarrow$(1). Let $|G(I)|= q \leq n-d+2$ and $I =(f_1, \ldots, f_q)$.\
If $q < n-d+2$, then from (\[degree1\]): $$G(I) = \{X_1X_2\cdots
X_{d-1}X_d, X_1X_2\cdots X_{d-1}X_{d+1},\ldots, X_1X_2\cdots
X_{d-1}X_k\},$$ with $d \leq k \leq n$. Hence: $$f_{12}=X_{d},\,\,
f_{13}=X_{d},\,\, \ldots , \,\,f_{1q}=X_{d},$$ $$f_{23}=X_{d+1},\,\, \ldots,\,\,
f_{2q}=X_{d+1},$$ and so on.\
By the structure of $G(I)$, this computation implies that $f_{ij} \neq f_{h\ell}$ for $i\neq h$ and $j\neq \ell$. Hence $[f_{ij}, f_{h\ell}]=1$ for $i<j$, $h<\ell$, $i\neq h$, $j\neq \ell$ and $i,j,h,\ell \in
[q]$.\
If $q = n-d+2$, then from (\[degree2\]): $$G(I) = \{X_1X_2\cdots
X_{d-1}X_d, \ldots,
X_1X_2\cdots X_{d-1}X_n, X_1X_2\cdots X_{d-2}X_dX_{d+1}\}.$$ We have: $$f_{12}=X_{d},\,\,
f_{13}=X_{d},\,\, \ldots, \,\,f_{1,n-d+1}=X_{d},\,\,
f_{1,n-d+2}=X_{d-1},$$ $$f_{23}=X_{d+1},\,\, \ldots,\,\,
f_{2,n-d+1}=X_{d+1},\,\, f_{2,n-d+2}=X_{d-1}, \ldots,\,\,$$ $$f_{1q}=X_{d-1}, \,\,f_{2q}=X_{d-1}, \,\,f_{3q}=X_{d-1}X_{d+2},\,\,
\ldots, \,\,f_{q-1,q}=X_{d-1}X_{n}.$$ Hence $[f_{ij},f_{h\ell}]=1$ for $i<j$, $h<\ell$, $i\neq h$, $j\neq \ell$ and $i,j,h,\ell \in
[q]$.
\[T\*\] Let $I\varsubsetneq A$ be a squarefree lexsegment ideal generated in degree $d$ with $G(I) \varsubsetneq M^s_d$. Then $I$ is generated by an $s$-sequence if and only if $|G(I)| \leq n-d+2$.
Let $I=(f_1,f_2, \ldots, f_q)$ be a squarefree lexsegment ideal and suppose that $f_1,f_2, \ldots, f_q$ is an $s$-sequence. We prove that $[f_{ij}, f_{h\ell}]=1$, for $i<j$, $h<\ell$, $i\neq h$, $j\neq \ell$, with $i,j,h,\ell \in [q]$.\
The $s$-sequence property implies that $G=\{g_{ij}=f_{ij}T_j -
f_{ji}T_i \mid 1\leq i < j \leq q \}$ is a Gröbner basis for $J$. In particular, $S(g_{ij}, g_{h\ell})$ has a standard expression with respect $G$ with remainder $0$. Note that, to get a standard expression of $S(g_{ij}, g_{h\ell})$ is equivalent to find some $g_{st} \in G$ whose initial term divides the initial term of $S(g_{ij}, g_{h\ell})$ and substitute a multiple of $g_{st}$ such that the remaindered polynomial has a smaller initial term and so on up to the remainder is $0$. We have: $$S(g_{ij}, g_{h\ell})= \frac{f_{ij}f_{\ell
h}}{[f_{ij},f_{h\ell}]}T_jT_h -
\frac{f_{h\ell}f_{ji}}{[f_{ij},f_{h\ell}]}T_iT_\ell.$$ First observe that $[f_{ij},f_{j\ell}]=1$ as $f_1, \ldots, f_q$ are squarefree monomials.\
Now we consider the other cases. Suppose that $i<j$, $h<\ell$, $i\neq h$, $j\neq \ell$. As $S(g_{ij}, g_{h\ell})$ has a standard expression with respect $G$ there exists $g_{st}$ such that $in_<(g_{st})$ divides $in_< (S(g_{ij}, g_{h\ell}))$.\
We distinguish two cases: $\ell>j$, $\ell<j$.\
Let $\ell>j$, then $in_<(g_{st}) \mid
\frac{f_{h\ell}f_{ji}}{[f_{ij},f_{h\ell}]}$.\
If $f_{h\ell} \mid \frac{f_{h\ell}f_{ji}}{[f_{ij},f_{h\ell}]}$, then $[f_{ij},f_{h\ell}] \mid f_{ji}$. But, since $[[f_{ij},f_{h\ell}],
f_{ji}]=1$, it follows $[f_{ij},f_{h\ell}]=1$.\
Consider $f_{s\ell} \mid
\frac{f_{h\ell}f_{ji}}{[f_{ij},f_{h\ell}]}$, where $f_{s\ell}=in_<(g_{s\ell})$ with $s<j$ and $s<h$. We can write: $$S(g_{ij}, g_{h\ell})= -
\frac{f_{ji}f_{h\ell}}{f_{s\ell}[f_{ij},f_{h\ell}]}g_{s\ell} T_i+
\frac{f_{ij}f_{\ell h}}{[f_{ij},f_{h\ell}]}T_jT_h -
\frac{f_{ji}f_{h\ell}f_{\ell
s}}{f_{s\ell}[f_{ij},f_{h\ell}]}T_iT_s.$$ Hence $\frac{f_{ij}f_{\ell
h}}{[f_{ij},f_{h\ell}]}T_jT_h$ is divided by $f_{ij}$ and consequently $[f_{ij},f_{h\ell}] \mid f_{\ell h}$. But, since $[[f_{ij},f_{h\ell}],
f_{\ell h}]=1$, then it follows $[f_{ij},f_{h\ell}]=1$.\
Let $\ell<j$, then $in_<(g_{st}) \mid
\frac{f_{ij}f_{\ell h}}{[f_{ij},f_{h\ell}]}$.\
If $f_{ij} \mid \frac{f_{ij}f_{\ell h}}{[f_{ij},f_{h\ell}]}$, then $[f_{ij},f_{h\ell}] \mid f_{\ell h}$. But, as we have $[[f_{ij},f_{h\ell}],
f_{\ell h}]=1$, then it follows $[f_{ij},f_{h\ell}]=1$.\
Consider $f_{sh} \mid \frac{f_{ij}f_{\ell h}}{[f_{ij},f_{h\ell}]}$, where $f_{sh}=in_<(g_{sh})$. We can write: $$S(g_{ij}, g_{h\ell})= \frac{f_{ij}f_{\ell
h}}{f_{sh}[f_{ij},f_{h\ell}]}T_jg_{sh} -
\frac{f_{h\ell}f_{ji}}{[f_{ij},f_{h\ell}]}T_iT_{\ell} +
\frac{f_{ij}f_{\ell h}f_{hs}}{f_{sh}[f_{ij},f_{h\ell}]}T_jT_s.$$ Hence $\frac{f_{ij}f_{\ell
h}f_{hs}}{f_{sh}[f_{ij},f_{h\ell}]}T_jT_s$ is divided by $f_{ij}$. Therefore $f_{sh}[f_{ij},f_{h\ell}] \mid f_{\ell h}f_{hs}$. But as we have $[[f_{ij},f_{h\ell}], f_{\ell h}]=1$, then $f_{sh} \mid
f_{\ell h}$ and $[f_{ij},f_{h\ell}] \mid f_{hs}$. By the structure of the monomials $f_1, \ldots, f_q$, if $[f_{ij},f_{h\ell}] \mid
f_{hs}$, with $s<h$, then $[f_{ij},f_{h\ell}]=1$.\
Hence in any case we have $[f_{ij},f_{h\ell}]=1$ for $i<j$, $h<\ell$, $i\neq h$, $j\neq \ell$, with $i,j,h,\ell \in [q]$. It follows $|G(I)|=q
\leq n-d+2$ by Lemma \[L1\].\
\
Now suppose $|G(I)| \leq n-d+2$. Hence from Lemma \[L1\], $[f_{ij}, f_{h\ell}]=1$, for $i<j$, $h<\ell$, $i\neq h$, $j\neq
\ell$, $i,j,h,\ell \in [q]$ and the assert follows from [@HRT] (Proposition 1.7).
$A=K[X_1,X_2,X_3,X_4]$, $I=(X_1X_2, X_1X_3, X_1X_4, X_2X_3,
X_2X_4)$.\
Set $f_1=X_1X_2$, $f_2=X_1X_3$, $f_3=X_1X_4$, $f_4=X_2X_3$, $f_5=X_2X_4$.\
$G=\{f_{ij}T_j - f_{ji}T_i | \ \ 1\leq i< j \leq 5
\}$ is not a Gröbner basis for $J$. In fact $J$ does not admit a linear Gröbner basis $BG(J)$ for any term order in $A[T_1,
\ldots,T_5]$: $BG(J)= \{X_2T_4 - X_3T_5, X_1T_2 - X_3T_5, X_2T_3 -
X_4T_5, X_1T_1 - X_4T_5, X_3T_3 - X_4T_4,$\
$X_3T_1 - X_4T_2,
X_4T_2T_3 - X_4T_1T_4\}$ [@CNR]. Hence $f_1, \ldots, f_5$ is not an $s$-sequence.\
[If $G(I) = M^s_d$, then $I=\mathcal{I}_d$, $2\leq d \leq n$, where $\mathcal{I}_d$ is the Veronese ideal of $A$ generated by all the squarefree monomials of degree $d$ in the variables $X_1,\dots, X_n$. The ideal $\mathcal{I}_d$ is generated by an $s$-sequence if and only if $d=n-1$ ([@LR], Theorem 2.3). Hence $I=(M^s_d)$ is generated by an $s$-sequence if and only if $d=n-1$.]{}\
\[P1\] Let $I\varsubsetneq A$ be a squarefree lexsegment ideal generated in degree $d$ such that $|G(I)|\leq n-d+2$. Then the annihilator ideals of the sequence of the monomial generators of $I$ are:$$I_1=(0), \
\ I_i=(X_d, \ldots, X_{d+i-2}) \ \ for \ \ i=2, \ldots,n-d+1,$$ $$I_i=(X_{d-1}) \ \ for \ \ i=n-d+2.$$
Set $|G(I)|=q$. Let $I= (f_1,\ldots,f_q)$ with $f_1> \ldots
> f_q$.\
Set $f_{ij}= \displaystyle{\frac{f_i}{[f_i,f_j]}}$ for $i < j$, $i,j\in [q]$. Then the annihilator ideals of the monomial sequence $f_1,\ldots,f_q$ are $I_i=(f_{1i},f_{2i}, \ldots, f_{i-1,i})$, for $i\in [q]$.\
For $i=1$, $I_1=(0)$ and by the structure of these monomials, it follows: $$I_2=(f_{12})=(X_{d}),\,\, I_3=(f_{13},
f_{23})=(X_{d}, X_{d+1}),\,\, \ldots,\,\,$$ $$I_{n-d+1}=(f_{1,n-d+1},f_{2,n-d+1}, \ldots,f_{n-d,n-d+1})=(X_{d},
X_{d+1}, \ldots, X_{n-1})$$ and $$I_{n-d+2}=(f_{1,n-d+2}, \ldots,
f_{n-d+1,n-d+2})=(X_{d-1},X_{d-1}X_{d-2},\ldots,
X_{d-1}X_{n})=(X_{d-1}).$$ Hence the assert follows.
If $I$ is an USLI of $A$ generated in degree $d$, then $I$ is generated by a strong $s$-sequence. In fact $I_1 \varsubsetneq I_2 \varsubsetneq \cdots
\varsubsetneq I_{n-d+1}$ for $|G(I)|< n-d+2$.
\[inv:USLI\] Let $I\varsubsetneq A$ be an USLI generated in degree $d$. Then
1. $\dim (Sym_A(I)) = n+1$;
2. ${{\operatorname{e}}}(Sym_A(I))=|G(I)|$;
3. ${{\operatorname{reg}}}_A (Sym_A(I))= 1$;
4. ${{\operatorname{depth}}}_A (Sym_A(I)) = n+1$.
\(1) $I$ is generated by a strong $s$-sequence. Hence, by [@T] (Theorem 4.8), $Sym_A(I)$ has dimension $\dim (A) + 1=n+1$.\
(2) Let $|G(I)|=q$. By [@HRT] (Proposition 2.4) we have ${{\operatorname{e}}}(Sym_A(I))= \sum_{i=1}^q {{\operatorname{e}}}(A/I_{i})$. By Proposition \[P1\], the annihilator ideals $I_i$ are generated by a regular sequence, then, by [@T] (Theorem 4.8), ${{\operatorname{e}}}(A/I_i)=1$, for $i=2,\ldots,q$ and ${{\operatorname{e}}}(A/(0))=1$. Hence ${{\operatorname{e}}}(Sym_A(I))=\sum_{i=1}^q{{\operatorname{e}}}(A/I_i)=q$.\
(3) By [@T] (Theorem 4.8): $$\begin{aligned}
{{\operatorname{reg}}}(Sym_A(I)) &=& {{\operatorname{reg}}}_A (A[T_1,\ldots,T_{q}]/J) \\
&\leq & {{\operatorname{reg}}}_A (A[T_1,\ldots,T_{q}]/in_{<}(J)) = {{\operatorname{reg}}}_A (A[T_1,\ldots,T_{q}]/(I_1T_1,\ldots, I_qT_q)) \\
&\leq & \max_{2\leq i\leq q} \{\sum_{j=1}^{i-1} \deg X_j- (i-2)\}=(i-1)-(i-2)=1,
$$ for $i=2, \ldots, q$. Since $J$ is generated by the linear forms of degree two $X_iT_j -
X_jT_i$, for $i,j=1, \ldots,q$, then ${{\operatorname{reg}}}_A (A[T_1,\ldots,T_{q}]/J)
\geq 1$. It follows that ${{\operatorname{reg}}}_A (Sym_A(I))=1$.\
(4) ${{\operatorname{depth}}}_A (Sym_A(I))\geq \dim(A) +1 = n+1$ ([@T], Theorem 4.8).\
Since ${{\operatorname{depth}}}_A (Sym_A(I)) \leq \dim Sym_A(I)) = n+1$, the assert follows.
If $I$ is an USLI of $A$ generated in degree $d$, then $Sym_A(I)$ is Cohen-Macaulay.
\[inv:ASLI\] Let $I\varsubsetneq A$ be an AUSLI generated in degree $d$. Then
1. $\dim (Sym_A(I)) = n+1$.
2. ${{\operatorname{e}}}(Sym_A(I))=2|G(I)| - 2$.
\(1) By [@HRT] (Proposition 2.4), we have $\dim (Sym_A(I))=
\textrm{max} \{ \dim (A/(I_{i_1} + \ldots + I_{i_r})) + r \},$ for $1\leq i_1 < \cdots < i_r \leq n-d+2$.\
Hence, by Proposition \[P1\]: $$\dim (Sym_A(I))= (n-r+ 1) + r = n+1.$$ (2) We have that $|G(I)|=n-d+2$.\
From [@HRT] (Proposition 2.4), $${{\operatorname{e}}}(Sym_A(I))= \sum_{1\leq i_1 < \cdots < i_r
\leq n-d+2} {{\operatorname{e}}}(A/(I_{i_1}+\ldots+ I_{i_r}))$$ with $\dim
(A/(I_{i_1}+\ldots + I_{i_r}))=\dim (Sym_A(I))-r = n+1-r$, $1\leq r \leq n-d+2$.\
By Proposition \[P1\], $I_i=(X_d, \ldots, X_{d+i-2})$, for $i=2,
\ldots,n-d+1$, $I_i=(X_{d-1})$ for $i=n-d+2$.\
Set $H=I_{i_1}+\ldots+ I_{i_r}$. Hence $A/H$ is Cohen-Macaulay and has a linear resolution with projective dimension equal to the number of the generators of $H$ ([@HK]). Then $e(A/H)=1$, by Huneke-Miller formula ([@HM]).\
Set $|G(I)|=n-d+2=q$ and $d'=\dim (A/H)=n+1-r$. Then ${{\operatorname{e}}}(Sym_A(I))$ is given by the sum of the following terms:\
${{\operatorname{e}}}(A/(0))=1$, for $r=1$ and $d'=n$,\
${{\operatorname{e}}}(A/(I_1+I_2))=1$, ${{\operatorname{e}}}(A/(I_1+I_q))=1$, for $r=2$ and $d'=n-1$,\
${{\operatorname{e}}}(A/(I_{1}+ I_{2}+I_3))=1$, ${{\operatorname{e}}}(A/(I_{1}+ I_{2}+I_q))=1$, for $r=3$ and $d'=n-2$,\
${{\operatorname{e}}}(A/(I_{1}+I_{2}+I_{3}+I_4))=1$, ${{\operatorname{e}}}(A/(I_{1}+I_{2}+I_{3}+I_q))=1$, for $r=4$ and $d'=n-3$,\
and so on up to\
${{\operatorname{e}}}(A/(I_{1}+I_{2}+\ldots + I_{q-1}))=1$, ${{\operatorname{e}}}(A/(I_{1}+\ldots +I_{q-2}+I_q))=1$, for $r=n-d+1$ and $d'=d-2$,\
${{\operatorname{e}}}(A/(I_1+I_2+\cdots +I_q))=1$, for $r=n-d+2$ and $d'=d-1$.\
Hence $$\begin{aligned}
{{\operatorname{e}}}(Sym_A(I)) &=& e(A/(0)) + 2 e(A/(I_1+I_2))+ \cdots + 2
e(A/(I_{1}+I_{2}+\ldots + I_{q-1}))\\
&&+ e(A/(I_1+I_2+ \cdots +I_q))=
2(q-2)+2=2q-q.
\end{aligned}$$
[99]{}
Aramova, A., Herzog, J., Hibi, T.: Squarefree lexsegment ideals. Math. Z. 228, 353–378 (1998).
Babson, E., Novik, I., Thomas R.: Reverse lexicographic shifting. J. Algebraic Combin. 23, 107–123 (2006). Bayer, D., Charalambous, H., Popescu, S.: Extremal Betti numbers and Applications to Monomial Ideals. J. Algebra 221, 497–512 (1999). Björner,A., Kalai, G.: An extended Euler-Poincarè. Acta Math. 161, 279-303 (1988). Bruns, W., Herzog, J.: Cohen-Macaulay rings. Cambridge University Press, 1996.
Crupi, M., Utano, R: Extremal Betti numbers of lexsegment ideals. In: Lecture Notes in Pure and Applied Math., Geometric and combinatorial aspects of Commutative algebra 217, 159–164 (2000).
Crupi, M., Utano R.: Extremal Betti numbers of graded ideals. Results Math. 43, 235-244 (2003).
Eisenbud, D.: Commutative Algebra with a view towards Algebraic Geometry. Springer-Verlag 1995.
Herzog, J., Kühl, M.: On Betti numbers of finite pure and linear resolutions. Comm. Algebra 12, 1627-1646 (1984).
Herzog, J., Restuccia, G., Tang, Z.: $s$-sequences and symmetric algebras. Manuscripta Math. 104, 479–501 (2001).
Huneke, C., Miller, M.: A note on the multiplicity of Cohen-Macaulay algebras with pure resolutions. Canad. J.Math. 37, 1149-1162 (1985).
Kalai, G.: The diameter of graphs of convex polytopes and $f$-vector theory. In: Applied Geometry and Discrete Mathematics - The Victor Klee Festschrift (Eds:P.Gritzmann and B.Sturmfels), DIMACS Series in Discrete Mathematics and Theoretical Computer Science 4, 387-411. Amer. Math.Soc., Providence, RI (1991).
La Barbiera, M., Restuccia, G.: Mixed product ideals generated by s-sequences, Algebra Colloq. 18(4), 553-570 (2011).
Tang, Z.: On certain monomial sequences. J. Algebra 282, 831–842 (2004).
[^1]: **$^1$ To appear in Algebra Colloquium**
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.