text
stringlengths
4
2.78M
meta
dict
--- abstract: 'The strong no loop conjecture states that a simple module of finite projective dimension over an artin algebra has no non-zero self-extension. The main result of this paper establishes this well known conjecture for finite dimensional algebras over an algebraically closed field.' address: - 'Kiyoshi Igusa, Department of Mathematics, Brandeis University, Waltham, MA 02453, United States' - 'Shiping Liu, Département de Mathématiques, Université de Sherbrooke, Sherbrooke, Québec, Canada, J1K 2R1' - 'Charles Paquette, Département de Mathématiques, Université de Sherbrooke, Sherbrooke, Québec, Canada, J1K 2R1' author: - 'Kiyoshi Igusa, Shiping Liu, and Charles Paquette' title: A proof of the strong no loop conjecture --- Introduction {#introduction .unnumbered} ============ Let ${\Lambda}$ be an artin algebra, and denote by ${\rm mod}\hskip 0.4pt{\Lambda}$ the category of finitely generated right ${\Lambda}$-modules. It is an important problem in the representation theory of algebras to determine whether ${\Lambda}$ has finite or infinite global dimension, and more specifically, whether a simple ${\Lambda}$-module has finite or infinite projective dimension. For instance, the derived category $D^b({\rm mod}\hskip 0.4pt {\Lambda})$ has Auslander-Reiten triangles if and only if ${\Lambda}$ has finite global dimension; see [@Hap1; @Hap2]. One approach to this problem is to consider the extension quiver of ${\Lambda}$, which has vertices given by a complete set of non-isomorphic simple ${\Lambda}$-modules and single arrows $S\to T$, where $S$ and $T$ are vertices such that ${\rm Ext}^1_\Lambda(S, T)$ is non-zero. Then the [*no loop conjecture*]{} affirms that the extension quiver of ${\Lambda}$ contains no loop if ${\Lambda}$ is of finite global dimension, while the [*strong no loop conjecture*]{}, which is due to Zacharia, strengthens this to state that a vertex in the extension quiver admits no loop if it has finite projective dimension; see [@ARS; @I]. The no loop conjecture was first explicitly established for artin algebras of global dimension two; see [@GGZ]. For finite dimensional elementary algebras, as shown in [@I], this can be easily derived from an earlier result of Lenzing on Hochschild homology in [@Le]. Lenzing’s technique was to extend the notion of the trace of endomorphisms of projective modules, defined by Hattori and Stallings in [@Ha; @Sta], to endomorphisms of modules over a noetherian ring with finite global dimension, and apply it to a particular kind of filtration for the regular module. In contrast, up to now, the strong no loop conjecture has only been verified for some special classes of algebras such as monomial algebras; see [@BF; @I], special biserial algebras; see [@LM], and algebras with at most two simple modules and radical cubed zero; see [@Je]. Many other partial results can be found in [@BuS; @DK; @GSZ; @MP; @Paq; @Zac]. Most recently, Skorodumov generalized and localized Lenzing’s filtration to indecomposable projective modules. This allowed him to prove this conjecture for finite dimensional elementary algebras of finite representation type; see [@De]. In this paper, we shall localize Lenzing’s trace function to endomorphisms of modules in ${\rm mod}\hskip 0.4pt {\Lambda}$ with an $e$-bounded projective resolution, where $e$ is an idempotent in ${\Lambda}$. The key point is that every module in ${\rm mod}\hskip 0.4pt {\Lambda}$ has an $e$-bounded projective resolution if the semi-simple module supported by $e$ has finite injective dimension. This will enable us to solve the strong no loop conjecture for a large class of artin algebras including finite dimensional elementary algebras, and particularly, for finite dimensional algebras over an algebraically closed field. Localized trace function and Hochschild homology ================================================ Throughout, $J$ will stand for the Jacobson radical of ${\Lambda}$. The additive subgroup of ${\Lambda}$ generated by the elements $ab-ba$ with $a, b\in {\Lambda}$ is called the *commutator group* of ${\Lambda}$ and written as $[{\Lambda},{\Lambda}]$. One defines then the Hochschild homology group ${{\rm H\hspace{-1pt}H}_0}({\Lambda})$ to be ${\Lambda}\hskip -0.8pt /\hskip -0.2pt[{\Lambda},{\Lambda}]$. We shall say that ${{\rm H\hspace{-1pt}H}_0}({\Lambda})$ is *radical-trivial* if $J\subseteq [{\Lambda},{\Lambda}]$. To start with, we recall the notion of the trace of an endomorphism ${{\varphi}}$ of a projective module $P$ in ${\rm mod}\hskip 0.4pt {\Lambda}$, as defined by Hattori and Stallings in [@Ha; @Sta]; see also [@I; @Le]. Write $P=e_1{\Lambda}\oplus \cdots \oplus e_r{\Lambda}$, where the $e_i$ are primitive idempotents in ${\Lambda}$. Then $\varphi=(a_{ij})_{r\times r}$, where $a_{ij} \in e_i{\Lambda}e_j$. The [*trace*]{} of $\varphi$ is defined to be $${{\rm tr}}(\varphi)= {\sum}_{i=1}^r a_{ii} + [{\Lambda}, {\Lambda}]\in {{\rm H\hspace{-1pt}H}_0}({\Lambda}).$$ We collect some well known properties of this trace function in the following proposition, in which the property (2) is the reason for defining the trace to be an element in ${{\rm H\hspace{-1pt}H}_0}({\Lambda})$. \[tr\] Let $P, P'$ be projective modules in ${\rm mod}\hskip 0.5pt{\Lambda}$. 1. If $\varphi, \psi\in {\rm End}_{\Lambda}(P)$, then ${{\rm tr}}({{\varphi}}+ \psi)={{\rm tr}}({{\varphi}})+{{\rm tr}}(\psi)$. 2. If ${{\varphi}}: P \to P'$ and $\psi: P' \to P$ are ${\Lambda}$-linear, then ${{\rm tr}}({{\varphi}}\psi) = {{\rm tr}}(\psi{{\varphi}})$. 3. If ${{\varphi}}=(\varphi_{ij})_{2\times 2}: P \oplus P' \to P \oplus P'$, then ${{\rm tr}}({{\varphi}}) = {{\rm tr}}({{\varphi}}_{11}) + {{\rm tr}}({{\varphi}}_{22})$. 4. If $\psi: P\to P'$ is an isomorphism and ${{\varphi}}\in {\rm End}_{\Lambda}(P)$, then ${{\rm tr}}(\psi{{\varphi}}\psi^{-1})={{\rm tr}}({{\varphi}})$. 5. If ${{\varphi}}: {\Lambda}\to {\Lambda}$ is the left multiplication by $a \in {\Lambda}$, then ${{\rm tr}}({{\varphi}}) = a + [{\Lambda},{\Lambda}]$. Next, we recall Lenzing’s extension of this notion to endomorphisms of modules of finite projective dimension. For $M\in {\rm mod}\hskip 0.5pt {\Lambda}$, let $\mathcal{P}_M$ denote a projective resolution $$\cdots \longrightarrow P_i\stackrel{d_i}{\longrightarrow} P_{i-1}\rightarrow \cdots \rightarrow P_1 \stackrel{d_1}{\longrightarrow} P_0 \stackrel{d_0}{\longrightarrow} M\rightarrow 0 \vspace{4pt}$$ of $M$ in ${\rm mod}\hskip 0.5pt {\Lambda}$. For each ${{\varphi}}\in {\rm End}_{\Lambda}\hskip -0.5pt (M)$, one can construct a commutative diagram $$\xymatrix{ \cdots \ar[r] & P_i\ar[r]^{d_i} \ar[d]^{{{\varphi}}_i}& P_{i-1}\ar[r] \ar[d]^{{{\varphi}}_{i-1}} & \cdots \ar[r]& P_1 \ar[r]^{d_1}\ar[d]^{{{\varphi}}_1}& P_0\ar[r]^{d_0} \ar[d]^{{{\varphi}}_0}& M\ar[r] \ar[d]^{{{\varphi}}}& 0\\ \cdots \ar[r] & P_i\ar[r]^{d_i}& P_{i-1}\ar[r] & \cdots \ar[r]& P_1 \ar[r]^{d_1} & P_0\ar[r]^{d_0} & M\ar[r] & 0} \vspace{2pt}$$ in ${\rm mod}\hskip 0.4pt{\Lambda}$. We shall call $\{{{\varphi}}_i\}_{i\ge 0}$ a [*lifting*]{} of ${{\varphi}}$ to $\mathcal{P}_M$. If $M$ is of finite projective dimension, then one may assume that $\mathcal{P}_M$ is bounded and define the [*trace*]{} of ${{\varphi}}$ by $${{\rm tr}}({{\varphi}})={\sum}_{i=0}^{\infty}(-1)^i\,{{\rm tr}}({{\varphi}}_i)\in {{\rm H\hspace{-1pt}H}_0}({\Lambda}), \vspace{3pt}$$ which is independent of the choice of $\mathcal{P}_M$ and $\{{{\varphi}}_i\}$; see [@Le], and also [@I]. Our strategy is to localize this construction. Let $e$ be an idempotent in ${\Lambda}$. Set $${\Lambda}_{\hskip 0.4pt e} = {\Lambda}/{\Lambda}(1-e){\Lambda}.\vspace{0pt}$$ The canonical algebra projection ${\Lambda}\to {\Lambda}_{\hskip 0.4pt e}$ induces a group homomorphism $$H_e : {{\rm H\hspace{-1pt}H}_0}({\Lambda}) \to {{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4pt e}). \vspace{2pt}$$ For an endomorphism ${{\varphi}}$ of a projective module in ${\rm mod}\hskip 0.5pt {\Lambda}$, we define its [*$e$-trace*]{} by $$\vspace{3pt} {{\rm tr}}_e({{\varphi}})=H_e({{\rm tr}}({{\varphi}}))\in {{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4pt e}).$$ It is evident that this $e$-trace function has the properties (1) to (4) stated in Proposition \[tr\]. More importantly, we have the following result. \[etr\] Let $e$ be an idempotent in $\Lambda$, and let $P$ be a projective module in ${\rm mod}\hskip 0.4pt{\Lambda}$ whose top is annihilated by $\hskip 0.2pt e$. If ${{\varphi}}\in {\rm End}_{\Lambda}(P)$, then ${{\rm tr}}_e({{\varphi}})=0$. [*Proof.*]{} We may assume that $P$ is non-zero. Then $1-e=e_1+\cdots+e_r$, where the $e_i$ are pairwise orthogonal primitive idempotents in ${\Lambda}$. Let ${{\varphi}}\in {\rm End}_{\Lambda}(P)$. By Proposition \[tr\](3), we may assume that $P$ is indecomposable. Then $P\cong e_{\hskip -1pt s} {\Lambda}$ for some $1\le s\le r$. By Proposition \[tr\](4), we may assume that $P=e_{\hskip -1pt s} {\Lambda}$. Then ${{\varphi}}$ is the left multiplication by some $a\in e_{\hskip -1pt s}{\Lambda}e_{\hskip -1pt s}$. By Proposition \[tr\](5), $${{\rm tr}}_e({{\varphi}})=H_e(a+[{\Lambda}, {\Lambda}])=\bar a+[{\Lambda}_{\hskip 0.4pt e}, {\Lambda}_{\hskip 0.4pt e}],$$ where $\bar a=a+{\Lambda}\hskip 0.2pt (1-e){\Lambda}$. Since $a=e_{\hskip -1pt s}ae_{\hskip -1pt s}=(1-e)a(1-e)\in {\Lambda}\hskip 0.2pt (1-e){\Lambda}$, we get ${{\rm tr}}_e({{\varphi}})=0$. The proof of the lemma is completed. To extend the $e$-trace function, we shall call a projective resolution $\mathcal{P}_M$ of $M$ *$e$-bounded* if $e$ annihilates the tops of all but finitely many terms in $\mathcal{P}_M$. In this case, if ${{\varphi}}\in {\rm End}_{\Lambda}(M)$ with a lifting $\{{{\varphi}}_i\}_{i\ge 0}$ to $\mathcal{P}_M$ then, by Lemma \[etr\], ${{\rm tr}}_e({{\varphi}}_i)=0$ for all but finitely many $i$. This allows us to define the [*$e$-trace*]{} of ${{\varphi}}$ by $${{\rm tr}}_e({{\varphi}})={\sum}_{i=0}^{\infty}(-1)^i \,{{\rm tr}}_e({{\varphi}}_i) \in {{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4pt e}).$$ Let $e$ be an idempotent in ${\Lambda}$. The $e$-trace is well defined for endomorphisms of modules in ${\rm mod}\hskip 0.4pt{\Lambda}$ having an $e$-bounded projective resolution. [*Proof.*]{} Let $M$ be a module in ${\rm mod}\hskip 0.4pt{\Lambda}$ having an $e$-bounded projective resolution $$\mathcal{P}_M: \qquad \cdots \to P_i \stackrel{d_i}{\longrightarrow} P_{i-1} \to \cdots \to P_1 \stackrel{d_1}{\longrightarrow} P_0 \stackrel{d_0}{\longrightarrow} M \to 0. \vspace{3pt}$$ Fix ${{\varphi}}\in {\rm End}_{\Lambda}(M)$. We first show that ${{\rm tr}}_e({{\varphi}})$ is independent of the choice of its lifting to $\mathcal{P}_M$. By Proposition \[tr\](1), it amounts to proving that $\sum_{i=0}^\infty (-1)^i\,{{\rm tr}}_e({{\varphi}}_i)=0$ for any lifting $\{{{\varphi}}_i\}_{i\ge 0}$ of the zero endomorphism of $M$. Indeed, let $h_i : P_i \to P_{i+1}$ be morphisms such that ${{\varphi}}_0 = d_1h_0$ and ${{\varphi}}_i= d_{i+1}h_i + h_{i-1}d_i$, for $i \ge 1.$ Applying Proposition \[tr\], we get $${{\rm tr}}_e({{\varphi}}_i) = {{\rm tr}}_e(d_{i+1}h_i) + {{\rm tr}}_e(h_{i-1}d_i)= {{\rm tr}}_e(d_{i+1}h_i) + {{\rm tr}}_e(d_ih_{i-1}),$$ for $i\ge 1$. On the other hand, by assumption, there exists some $m \ge 0$ such that $e$ annihilates the top of $P_i$ for $i \ge m$. By Lemma \[etr\], ${{\rm tr}}_e(d_{m+1}h_{m})=0$ and ${{\rm tr}}_e({{\varphi}}_i)=0$ for $i \ge m$. This yields $$\begin{aligned} {\sum}_{i=0}^{\infty}(-1)^i \,{{\rm tr}}_e({{\varphi}}_i) &=& {{\rm tr}}_e({{\varphi}}_0) + {\sum}_{i=1}^{m}(-1)^i \,{{\rm tr}}_e({{\varphi}}_i) \\ &=& {{\rm tr}}_e(d_1h_0) + {\sum}_{i=1}^{m}(-1)^i\left({{\rm tr}}_e(d_{i+1}h_i) + {{\rm tr}}_e(d_ih_{i-1})\right) \\ &=& (-1)^{m}\,{{\rm tr}}_e(d_{m+1}h_{m})\\ &=&0.\end{aligned}$$ Next, we show that ${{\rm tr}}_e({{\varphi}})$ is independent of the choice of the $e$-bounded projective resolution $\mathcal{P}_M$. Suppose that $M$ has another $e$-bounded projective resolution $$\mathcal{P}'_M: \qquad \cdots \to P'_i \stackrel{d'_i}{\longrightarrow} P'_{i-1} \to \cdots \to P'_1 \stackrel{d'_1}{\longrightarrow} P'_0 \stackrel{d'_0}{\longrightarrow} M \to 0.$$ Considering ${{\varphi}}$, we get morphisms $u_i : P_i \to P'_i$ with $i\ge 0$ such that $d'_0u_0= {{\varphi}}d_0$ and $d'_iu_i=u_{i-1}d_i$ for $i \ge 1$. Similarly, considering $1_M$, we obtain maps $v_i: P_i'\to P_i$ with $i\ge 0$ such that $d_0v_0= d'_0$ and $d_iv_i=v_{i-1}d'_i$ for $i \ge 1$. Observe that $\{v_iu_i\}_{i\ge 0}$ and $\{u_iv_i\}_{i\ge 0}$ are liftings of ${{\varphi}}$ to $\mathcal{P}_M$ and $\mathcal{P}_M'$, respectively. By Proposition \[tr\](2), we have $$\qquad {\sum}_{i=0}^{\infty}(-1)^i\,{{\rm tr}}_e(u_iv_i)= {\sum}_{i=0}^{\infty}(-1)^i \,{{\rm tr}}_e(v_iu_i).$$ The proof of the lemma is completed. In the sequel, $S_e$ will stand for the semi-simple ${\Lambda}$-module $e{\Lambda}/eJ$. Suppose that $S_e$ has finite injective dimension. If $M$ is a module in ${\rm mod}\hskip 0.3pt{\Lambda}$, then ${{\rm Ext}}_{{\Lambda}}^{i}(M,S_e)=0$ for all sufficient large integers $i$, that is, the minimal projective resolution of $M$ is $e$-bounded. Therefore, the $e$-trace is defined for every endomorphism in ${\rm mod}\hskip 0.4pt{\Lambda}$. In particular, if ${\Lambda}$ is of finite global dimension, then we recover Lenzing’s trace function by taking $e=1_A$. \[additivity\] Let $e$ be an idempotent in ${\Lambda}$. Consider a commutative diagram $$\xymatrixrowsep{18pt} \xymatrix{0 \ar[r] & L\ar[r]^u\ar[d]^{{{\varphi}}_{_L}} & M \ar[r]^v\ar[d]^{{{\varphi}}_{_M}} & N\ar[d]^{{{\varphi}}_{_N}} \ar[r] & 0\\ 0 \ar[r] & L \ar[r]^u & M \ar[r]^v & N \ar[r] &0}$$ in ${\rm mod}\hskip 0.4pt{\Lambda}$ with exact rows. If $L, N$ have $e$-bounded projective resolutions, then so does $M$ and ${{\rm tr}}_e({{\varphi}}_{_M}) = {{\rm tr}}_e({{\varphi}}_{_L}) + {{\rm tr}}_e({{\varphi}}_{_N}).$ [*Proof.*]{} Assume that $L$ and $N$ have $e$-bounded projective resolutions as follows$\,:$ $$\mathcal{P}_L: \qquad \cdots \to P_i \stackrel{d_i}{\longrightarrow} P_{i-1} \to \cdots \to P_1 \stackrel{d_1}{\longrightarrow} P_0 \stackrel{d_0}{\longrightarrow} L \to 0$$ and $$\mathcal{P}_N:\qquad \cdots \to P'_i \stackrel{d'_i}{\longrightarrow} P'_{i-1} \to \cdots \to P'_1 \stackrel{d'_1}{\longrightarrow} P'_0 \stackrel{d'_0}{\longrightarrow} N \to 0. \vspace{4pt}$$ By the Horseshoe lemma, there exists in ${\rm mod}\hskip 0.5pt {\Lambda}$ a commutative diagram $$\xymatrixrowsep{18pt} \xymatrixcolsep{18pt}\xymatrix{ \cdots \ar[r]& P_i\ar[r]^-{d_i} \ar[d]^{q_i}& P_{i-1} \ar[r] \ar[d]^{q_{i-1}}& \cdots \ar[r] & P_0 \ar[r]^{d_0}\ar[d]^{q_{_0}}& L \ar[r]\ar[d]^u& 0\\ \cdots \ar[r]& P_i\oplus P_i'\ar[r]^-{d''_i} \ar[d]^{p_i}& P_{i-1} \oplus P'_{i-1}\ar[r] \ar[d]^-{p_{i-1}}& \cdots \ar[r] & P_0\oplus P'_0 \ar[r]^-{d''_0}\ar[d]^{p_{_0}}& M \ar[r]\ar[d]^v& 0\\ \cdots \ar[r]& P'_i\ar[r]^-{d'_i} & P'_{i-1} \ar[r] & \cdots \ar[r] & P'_0 \ar[r]^-{d'_0}& N \ar[r]& 0\vspace{5pt}}$$ with exact rows, where $q_i={1 \choose {\hskip 0.4pt 0\hskip 0.5pt}}$, $p_i=(0,1)$ for all $i\ge 0$. In particular, the middle row is an $e$-bounded projective resolution of $M$ which we denote by ${\mathcal P}_M$. Choose a lifting $\{f_i\}_{i\ge 0}$ of ${{\varphi}}_{_L}$ to $\mathcal{P}_L$ and a lifting $\{g_i\}_{i\ge 0}$ of ${{\varphi}}_{_N}$ to $\mathcal{P}_N$. It is well known; see, for example, [@We p. 46] that there exists a lifting $\{h_i\}_{i\ge 0}$ of ${{\varphi}}_{_M}$ to $\mathcal{P}_M$ such that $$\xymatrixrowsep{18pt} \xymatrix{0 \ar[r] & P_i \ar[r]^-{q_i} \ar[d]^{f_i} & P_i \oplus P'_i \ar[r]^-{p_i}\ar[d]^{h_i} & P'_i \ar[r] \ar[d]^{g_i} & 0 \\ 0 \ar[r] & P_i \ar[r]^-{q_i} & P_i \oplus P'_i \ar[r]^-{p_i} & P'_i \ar[r] & 0 } \vspace{3pt}$$ is commutative, for every $i\ge 0$. Since $h_iq_i=q_if_i$ and $g_ip_i=p_ih_i$, we can write $h_i$ as a $(2 \times 2)$-matrix whose diagonal entries are $f_i$ and $g_i$. Thus ${{\rm tr}}_e(h_i) = {{\rm tr}}_e(f_i) + {{\rm tr}}_e(g_i)$ by Proposition \[tr\](3). As a consequence, ${{\rm tr}}_e({{\varphi}}_M) = {{\rm tr}}_e({{\varphi}}_N) + {{\rm tr}}_e({{\varphi}}_L).$ The proof of the proposition is completed. Finally, we shall describe the Hochschild homology group ${{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4pt e})$ in case $S_e$ has finite injective dimension. \[maintheo\] Let ${\Lambda}$ be an artin algebra, and let $e$ be an idempotent in ${\Lambda}$. If $S_e$ has finite injective dimension, then ${{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4pt e})$ is radical-trivial. [*Proof.*]{} Suppose that $S_e$ has finite injective dimension. Then the $e$-trace is defined for every endomorphism in ${\rm mod}\hskip 0.5pt {\Lambda}$. Let $x \in {\Lambda}$ be such that $\bar x=x+{\Lambda}(1-e)\hskip 0.3pt{\Lambda}\,$ lies in the radical of ${\Lambda}_{\hskip 0.4pt e}$, which is $(J + {\Lambda}(1-e){\Lambda}) /{\Lambda}(1-e){\Lambda}$. Hence, $\bar x = \bar a$ for some $a\in J$. Let $r>0$ be such that $a^r=0$, and consider the chain $$0=M_r\subseteq M_{r-1}\subseteq \cdots \subseteq M_1\subseteq M_0={\Lambda},$$ of submodules of ${\Lambda}$, where $M_i = a^i{\Lambda}$, $i=0, \ldots, r$. Let ${{\varphi}}_0: {\Lambda}\to {\Lambda}$ be the left multiplication by $a$. Since ${{\varphi}}_0(M_i)\subseteq M_{i+1}$, we see that ${{\varphi}}_0$ induces morphisms ${{\varphi}}_i : M_i \to M_i$, $i=1, \ldots, r$, such that $$\xymatrixrowsep{17pt} \xymatrix{0 \ar[r] & M_{i+1}\ar[r]\ar[d]^{{{\varphi}}_{i+1}} & M_i \ar[r]\ar[d]^{{{\varphi}}_i} & M_i/M_{i+1}\ar[d]^{0} \ar[r] & 0\\ 0 \ar[r] & M_{i+1} \ar[r] & M_i \ar[r] & M_i/M_{i+1} \ar[r] &0}$$ commutes, and hence ${{\rm tr}}_e({{\varphi}}_i) = {{\rm tr}}_e({{\varphi}}_{i+1})$ by Proposition \[additivity\], for $i=0, 1, \ldots, r-1$. Applying Proposition \[tr\](5), we get $$\bar a+[{\Lambda}_{\hskip 0.4pt e}, {\Lambda}_{\hskip 0.4pt e}]=H_e(a+[{\Lambda}, {\Lambda}])=H_e({\rm tr}({{\varphi}}_0))={{\rm tr}}_e({{\varphi}}_0) ={{\rm tr}}_e({{\varphi}}_r)=0,$$ that is, $\bar x = \bar a\in [{\Lambda}_{\hskip 0.4pt e},{\Lambda}_{\hskip 0.4pt e}]$. The proof of the theorem is completed. Taking $e=1_A$, we recover the following well known result; see, for example, [@Le]. If ${\Lambda}$ is an artin algebra of finite global dimension, then ${{\rm H\hspace{-1pt}H}_0}({\Lambda})$ is radical-trivial. Indeed, if ${\Lambda}$ is a finite dimensional algebra of finite global dimension over a field of characteristic zero, then all the Hochschild homology groups ${\rm H\hspace{-1pt}H}_i({\Lambda})$ with $i \ge 1$ vanish; see [@Le]. However, in the situation as in Theorem \[maintheo\], the higher Hochschild homology groups of ${\Lambda}_{\hskip 0.2pt e}$ do not necessarily vanish and ${\Lambda}_{\hskip 0.4pt e}$ may be of infinite global dimension. [Example.]{} Let ${\Lambda}= kQ/I$, where $k$ is a field, $Q$ is the quiver $$\xymatrixrowsep{19pt}\xymatrixcolsep{19pt}\xymatrix{1 \ar[r]^{\alpha}\ar[d]_{\gamma} & 2 \ar[d]^{\beta}\\ 4 \ar[r]^{\delta} & 3 \ar[ul]_{\varepsilon} }$$ and $I$ is the ideal in $kQ$ generated by $\alpha \beta - \gamma \delta, \beta \varepsilon, \delta \varepsilon, \varepsilon\alpha$. One can show that ${\Lambda}$ has finite global dimension. Now, let $e$ be the sum of the primitive idempotents in ${\Lambda}$ corresponding to the vertices $1,2,3$. Then ${\Lambda}_{\hskip 0.4pt e}$ is a Nakayama algebra with radical squared zero, which clearly has infinite global dimension. By Theorem \[maintheo\], ${{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4pt e})$ is radical-trivial. However, a direct computation shows that ${\rm H\hspace{-1pt}H}_2({\Lambda}_{\hskip 0.4pt e})$ is non-zero; see also [@IZ]. Main results ============ The main objective of this section is to apply the previously obtained result to solve the strong no loop conjecture for finite dimensional algebras over an algebraically closed field. We start with an artin algebra ${\Lambda}$ with a primitive idempotent $e$. We shall say that ${\Lambda}$ is [*locally commutative*]{} at $e$ if $e{\Lambda}e$ is commutative and that ${\Lambda}$ is [*locally commutative*]{} if it is locally commutative at every primitive idempotent. Moreover, $e$ is called [*basic*]{} if $e{\Lambda}$ is not isomorphic to any direct summand of $(1-e){\Lambda}$. In this terminology, ${\Lambda}$ is basic if and only if all its primitive idempotents are basic. \[main\] Let ${\Lambda}$ be an artin algebra, and let $e$ be a basic primitive idempotent in ${\Lambda}$ such that ${\Lambda}/J^2$ is locally commutative at $e+J^2$. If $S_e$ has finite projective or injective dimension, then ${{\rm Ext}}^1_{\Lambda}(S_e,S_e)=0$. [*Proof.*]{} Firstly, we assume that $S_e$ is of finite injective dimension. For proving that ${{\rm Ext}}^1_{\Lambda}(S_e,S_e)=0$, it suffices to show that $e J e /eJ^2e = 0$. Let $a \in eJe$. Then $a + {\Lambda}(1-e){\Lambda}\in [{\Lambda}_{\hskip 0.4pt e}, {\Lambda}_{\hskip 0.4pt e}]$ by Theorem \[maintheo\]. Since $e$ is basic, $e{\Lambda}(1-e){\Lambda}e \subseteq eJ^2e$. This yields an algebra homomorphism $$f: {\Lambda}_e \to e{\Lambda}e/eJ^2e: x+{\Lambda}(1-e){\Lambda}\mapsto exe + eJ^2e.$$ Thus, $a + eJ^2e=f(a + {\Lambda}(1-e){\Lambda})$ lies in the commutator group of $e{\Lambda}e/eJ^2e$. On the other hand, $e{\Lambda}e/eJ^2e\cong (e+J^2)({\Lambda}/J^2)(e+J^2)$, which is commutative. Therefore, $a + eJ^2e=0$, that is, $a \in eJ^2e$. The result follows in this case. Next, assume that $S_e$ has finite projective dimension. Let $D$ be the standard duality between ${{\rm mod}}{\Lambda}$ and ${{\rm mod}}\hskip 0.3pt {\Lambda}^{\, \rm op}$. Then $D(S_e)$ is the simple ${\Lambda}^{\,\rm op}$-module supported by the idempotent $e^{\rm o}$ corresponding to $e$, which is of finite injective dimension. Observe that the quotient of ${\Lambda}^{\rm op}$ modulo its radical square is also locally commutative at the class of $e^{\rm o}$ modulo the radical square. By what we have proven, ${{\rm Ext}}_{{\Lambda}}^1(S_e,S_e)\cong {{\rm Ext}}_{{\Lambda}^{\, \rm op}}^1(D(S_e),D(S_e))=0.$ The proof of the theorem is completed. [Remark.]{} The preceding result establishes the strong no loop conjecture for basic artin algebras ${\Lambda}$ such that ${\Lambda}/J^2$ is locally commutative. Now we shall specialize this result to finite dimensional algebras over a field. Recall that such an algebra is called [*elementary*]{} if its simple modules are all one dimensional over the base field; see [@ARS]. \[main2\] Let ${\Lambda}$ be a finite dimensional algebra over a field $k$, and let $S$ be a simple ${\Lambda}$-module which is one dimensional over $k$. If $S$ has finite projective or injective dimension, then ${{\rm Ext}}^1_{\Lambda}(S,S)=0$. [*Proof.*]{} Let $e\in {\Lambda}$ be the primitive idempotent supporting $S$. Then ${\Lambda}$ has a complete set $\{e_1, \ldots, e_n\}$ of primitive orthogonal idempotents with $e=e_1$. We may assume that $e_1{\Lambda}, \ldots, e_r{\Lambda}$, with $1\le r\le n$, are the non-isomorphic indecomposable projective modules in ${\rm mod}\hskip 0.5pt {\Lambda}$. Then $${\Lambda}/J \cong M_{n_1}(D_1) \times \cdots \times M_{n_r}(D_r),$$ where $D_i={\rm End}_{\Lambda}(e_i{\Lambda}/e_iJ)$ and $n_i$ is the number of indices $j$ with $1\le j\le n$ such that $e_j\hskip 0.4pt {\Lambda}\cong e_i{\Lambda}$, for $i=1, \ldots, r.$ Now $S$ is a simple $M_{n_1}(D_1)$-module, and hence $S\cong D_1^{n_1}$. Since $S$ is one dimensional over $k$, it is one dimensional over $D_1$. In particular, $n_1=1$. That is, $e$ is a basic primitive idempotent. Moreover, $e{\Lambda}e/eJe \cong Se \cong k$. Thus, for $x_1,x_2\in e {\Lambda}e$, we can write $x_i=\lambda_ie + a_i$, where $\lambda_i\in k$ and $a_i\in eJe$, $i=1, 2$. This yields $x_1x_2-x_2x_1=a_1a_2-a_2a_1\in eJ^2e.$ Therefore, $e{\Lambda}e/ eJ^2e$ is commutative, and so is $(e+J^2)({\Lambda}/J^2)(e+J^2)$. The result follows immediately from Theorem \[main\]. The proof of the theorem is completed. [Remark.]{} The preceding theorem establishes the strong no loop conjecture for finite dimensional elementary algebras, and hence for finite dimensional algebras over an algebraically closed field. We shall extend our results in this direction. Let ${\Lambda}$ be a finite dimensional elementary algebra over a field $k$. We may assume that ${\Lambda}=kQ/I$, where $Q$ is a finite quiver, $kQ$ is the path algebra, and $I$ is an admissible ideal in $kQ$; see [@ARS]. Recall that $I$ is [*admissible*]{} if $(kQ^+)^n\subseteq I\subseteq (kQ^+)^2$ for some $n\ge 2$, where $kQ^+$ is the ideal in $kQ$ generated by the arrows, and [*monomial*]{} if $I$ in addition is generated by some paths. In this setting, the extension quiver of ${\Lambda}$ is isomorphic to the quiver obtained from $Q$ by shrinking the possible multiple arrows. If $p_1, \ldots, p_r$ are distinct paths in $Q$ of length $\ge 2$ from one vertex to another, then a $k$-linear combination $$\rho=\lambda_1p_1+\cdots+\lambda_rp_r\vspace{3pt}$$ is called a *minimal relation* for ${\Lambda}$ if $\rho\in I$ and $\sum_{i \in J}\lambda_ip_i \not \in I$ for any $J\subset \{1,\cdots,r\}$. Moreover, let $\sigma = \alpha_1 \alpha_2 \cdots \alpha_r$ be an oriented cycle in $Q$, where the $\alpha_i$ are arrows. The support of $\sigma$, written as ${\rm supp}(\sigma)$, is the set of vertices in $Q$ occurring as starting points of $\alpha_1, \ldots, \alpha_r$. The *idempotent supporting $\sigma$* is the sum of all primitive idempotents in ${\Lambda}$ associated to the vertices in ${\rm supp}(\sigma)$. Write $$\sigma_1 = \sigma, \; \sigma_i = \alpha_{i} \cdots \alpha_r\alpha_1 \cdots \alpha_{i-1}, \; i=2, \ldots, r,$$ called the [*cyclic permutations*]{} of $\sigma$. We shall say that $\sigma$ is [*cyclically free*]{} in ${\Lambda}$ if none of the $\sigma_i$ with $1\le i\le r$ is a summand of a minimal relation for ${\Lambda}$, and [*cyclically non-zero*]{} in ${\Lambda}$ if none of the $\sigma_i$ lies in $I$. \[main2\] Let ${\Lambda}= kQ/I$ with $Q$ a finite quiver and $I$ an admissible ideal in $kQ$, and let $\sigma$ be an oriented cycle in $Q$ with supporting idempotent $e\in {\Lambda}$. If $\sigma$ is cyclically free in ${\Lambda}$, then $S_e$ has infinite projective and injective dimensions. [*Proof.*]{} Suppose that $\sigma$ is cyclically free in ${\Lambda}$. If $\sigma$ is a power of a shorter oriented cycle $\delta$, then it is easy to see that $\delta$ is also cyclically free in ${\Lambda}$ and ${\rm supp}(\delta)={\rm supp} (\sigma)$. Hence, we may assume that $\sigma$ is not a power of any shorter oriented cycle. Let $\sigma_1, \ldots, \sigma_r$, where $\sigma_1=\sigma$, be the cyclic permutations of $\sigma$. It is then well known that the $\sigma_i$ with $1\le i\le r$ are pairwise distinct. For any $p\in kQ$, denote by $\tilde p$ its class in ${\Lambda}$ and by $\bar{p}$ the class of $\tilde p$ in ${\Lambda}_{\hskip 0.4pt e}$. Let $W$ be the vector subspace of ${\Lambda}_{\hskip 0.4pt e}$ spanned by the classes $\bar p$, where $p$ ranges over the paths in $Q$ different from $\sigma_1, \ldots, \sigma_r$. Then, there exist paths $p_1, \ldots, p_m$ in $Q$ such that $\{\bar p_1, \ldots, \bar p_m\}$ is a $k$-basis of $W$. We claim that $\{\bar \sigma_1, \ldots, \bar \sigma_r, \bar p_1, \ldots, \bar p_m\}$ is a $k$-basis of ${\Lambda}_{\hskip 0.4pt e}$. Indeed, it clearly spans ${\Lambda}_{\hskip 0.4pt e}$. Assume that $${\sum}_{i=1}^r\lambda_i \bar \sigma_i + {\sum}_{j=1}^m\nu_j\bar p_j=\bar 0, \; \lambda_i, \nu_j\in k.$$ That is, $\sum \lambda_i \tilde \sigma_i + \sum \nu_j \tilde p_j \in {\Lambda}(1-e) {\Lambda}$. Then $${\sum}_{i=1}^r\lambda_i \,\tilde \sigma_i + {\sum}_{j=1}^m\nu_j\,\tilde p_j = {\sum}_{l=1}^s \mu_l \,\tilde q_l, \;\mu_l\in k,$$ where $q_1, \ldots, q_s$ are distinct paths in $Q$ passing through a vertex not in ${\rm supp}(\sigma)$. Fix some $t$ with $1\le t\le r$. Letting $\varepsilon_t$ be the trivial path in $Q$ associated to the starting point $a_t$ of $\sigma_t$, we get $${\sum}_{i=1}^r \lambda_i \varepsilon_t\,\sigma_i\,\varepsilon_t + {\sum}_{j=1}^m \nu_j \,\varepsilon_t\, p_j \,\varepsilon_t- {\sum}_{l=1}^s \mu_l \varepsilon_t\, q_l \,\varepsilon_t \in I.$$ Note that the non-zero elements of the $\varepsilon_t{\sigma}_i\varepsilon_t$, $\varepsilon_t p_j \varepsilon_t$, $\varepsilon_t \, q_l \,\varepsilon_t\in kQ$ are distinct oriented cycles from $a_t$ to $a_t$. Since $\sigma$ is cyclically free in ${\Lambda}$, we have $\lambda_j=0$ whenever $\varepsilon_t\sigma_j\varepsilon_t$ is non-zero. In particular, $\lambda_t=0$. Therefore, the $\lambda_i$ are all zero, and so are the $\nu_j$. This proves our claim. Suppose now that $\bar \sigma \in [{\Lambda}_{\hskip 0.4pt e},{\Lambda}_{\hskip 0.4pt e}]$. Then $$\bar{\sigma} = {\sum}_{i=1}^n\eta_i(\bar{u}_i\bar{v}_i - \bar{v}_i\bar{u}_i)$$ where $\eta_i \in k$ and $u_i, v_i\in \{\sigma_1, \ldots, \sigma_r, p_1, \ldots, p_m\}$. For each $1\le i\le n$, we see easily that $u_iv_i\not\in \{\sigma_1, \ldots, \sigma_r\}$ if and only if $v_iu_i\not\in \{\sigma_1, \ldots, \sigma_r\}$, and in this case, $\bar{u}_i\bar{v}_i- \bar v_i\bar u_i\in W$. Therefore, the equation $(1)$ becomes $$\bar{\sigma} = {\sum} \,\eta_{ij}(\bar{\sigma}_i - \bar{\sigma}_j) + w,$$ where $\eta_{ij}\in k$ and $w\in W$. Let $L$ be the linear form on ${\Lambda}_{\hskip 0.4pt e}$, which sends each of $\bar{\sigma}_1, \ldots, \bar \sigma_r$ to $1$ and vanishes on $W$. Since $\sigma=\sigma_1$, applying $L$ to the equation $(2)$ yields $1 = 0$, a contradiction. Therefore, the class of $\bar \sigma$ in ${{\rm H\hspace{-1pt}H}_0}({\Lambda}_{\hskip 0.4 pt e})$ is non-zero. Since $\bar \sigma$ lies in the radical of ${\Lambda}_{\hskip 0.4pt e}$, by Theorem \[maintheo\], $S_e$ has infinite projective and injective dimensions. The proof of the theorem is completed. [Example.]{} Let ${\Lambda}=kQ/I$, where $Q$ is the following quiver $$\xymatrixrowsep{25pt}\xymatrixcolsep{25pt} \xymatrix{1 \ar@/^1pc/[r]^(0.5){\alpha} \ar[r]_{\varepsilon} & 2 \ar@/^1pc/[r]^(0.5){\gamma} \ar@/^1pc/[l]^(0.5){\beta} & 3 \ar@/^1pc/[l]^{\delta} \ar@/^1pc/[r]^{\mu} & 4 \ar@/^1pc/[l]^{\nu} } \vspace{-3pt}$$ and $I$ is the ideal in $kQ$ generated by $\alpha\beta,\delta\gamma,\beta\varepsilon,\varepsilon\beta, \nu\delta,\nu\mu,\mu\nu,\gamma\mu$, $\alpha\gamma\delta\beta\alpha\gamma-\varepsilon\gamma$. It is easy to see that the oriented cycle $\beta\alpha \gamma\delta$ is cyclically free in ${\Lambda}$. By Theorem \[main2\], one of the simple modules $S_1, S_2, S_3$ has infinite projective dimension. \[cor1\] Let ${\Lambda}= kQ/I$ with $Q$ a finite quiver and $I$ an admissible ideal in $kQ$. If $Q$ contains an oriented cycle which is cyclically free in ${\Lambda}$, then ${\Lambda}$ has infinite global dimension. If $I$ is a monomial ideal in $kQ$, then an oriented cycle in $Q$ is cyclically free in ${\Lambda}$ if and only if it is cyclically non-zero in ${\Lambda}$. This yields the following consequence, which can also be derived from results in [@IZ]. Let $\,{\Lambda}=kQ/I$ with $Q$ a finite quiver and $I$ a monomial ideal in $kQ$. If $Q$ contains an oriented cycle which is cyclically non-zero in ${\Lambda}$, then ${\Lambda}$ has infinite global dimension. To conclude, we would like to draw the reader’s attention to an even stronger version of the no loop conjecture as follows. Let $S$ be a simple module over an artin algebra. If ${{\rm Ext}}^1(S,S)$ is non-zero, then ${{\rm Ext}}^i(S,S)$ is non-zero for infinitely many integers $i$. This conjecture was originally posed under the name of [*extreme no loop conjecture*]{} in [@LM]. It remains open except for monomial algebras and special biserial algebras; see [@GSZ; @LM]. [Acknowledgment.]{} This research was supported in part by the Natural Sciences and Engineering Research Council of Canada and was carried out during the first author’s visit to the Université de Sherbrooke. [99]{} , “Representation Theory of Artin Algebras", Cambridge Studies in Advanced Mathematics 36 (Cambridge University Press, Cambridge, 1995). , “Left monomial rings-a generalization of monomial algebras", Osaka J. Math. (3) 30 (1993) 543–558. , “Homological aspects of semigroup gradings on rings and algebras", Canad. J. Math. (3) 51 (1999) 488–505. , “Cohomological reduction by split pairs", J. Pure Appl. Algebra (3) 212 (2008) 471–485. , “Artin rings of global dimension two", J. Algebra 92 (1985) 375 – 379. , “Minimal projective resolutions", Trans. Amer. Math. Soc. (7) 353 (2001) 2915 – 2939. , “On the derived category of a finite dimensional algebra", Comment. Math. Helvetici. 62 (1987) 339 – 389. , “Auslander-Reiten triangles in derived categories of finite-dimensional algebras", Proc. Amer. Math. Soc. 112 (1991) 641 – 648. , “Rank element of a projective module", Nagoya Math. J. 25 (1965) 113 – 120. , “Notes on the no loops conjecture", J. Pure Appl. Algebra (2) 69 (1990) 161 – 176. , “On the cyclic homology of monomial relation algebras", J. Algebra (2) 151 (1992) 502 – 521. , “Strong no-loop conjecture for algebras with two simples and radical cube zero", Colloq. Math. (1) 102 (2005) 1 – 7. , “Nilpotente Elemente in Ringen von endlicher globaler Dimension", Math. Z. 108 (1969) 313 – 324. , “The strong no loop conjecture for special biserial algebras", Proc. Amer. Math. Soc. (12) 132 (2004) 3513 – 3523. , “Extensions of abelian categories and the strong no-loops conjecture", J. Algebra (1) 178 (1995) 1 – 20. , “Strictly stratified algebras revisited", Comm. Algebra (8) 37 (2009) 2591 – 2600. , “The strong no loop conjecture is true for mild algebras", preprint (arXiv:1011.1143v2). , “Centerless groups-an algebraic formulation of Gottlieb’s theorem", Topology 4 (1965) 129 – 134. , “[0.5pt]{}An introduction to homological algebra", Cambridge Studies in Advanced Mathematics 38 (Cambridge University Press, Cambridge, 1994). , “[0.5pt]{}Special monomial algebras of finite global dimension", NATO Adv. Sci. Inst. 233 (Kluwer, Dordrecht, 1988) 375 – 378.
{ "pile_set_name": "ArXiv" }
--- author: - Emma Storm - Christoph Weniger - Francesca Calore bibliography: - 'skyfact.bib' date: 'Compiled: ' title: 'SkyFACT: High-dimensional modeling of gamma-ray emission with adaptive templates and penalized likelihoods' --- LAPTH-015/17 Introduction {#sec:intro} ============ Two important activities in the analysis of astronomical images are parametric modeling and image reconstruction. Identifying adequate parametric models plays a central role in the astrophysical interpretation of observations, in particular for model discrimination and parameter regression. In the case of gamma rays, modeling observational data is very often done using template regression techniques, where fits with linear combinations of physically or observationally motivated spatial templates are used to perform component separation, often as function of energy [@Bennett:2003ca; @Finkbeiner:2003im; @Dobler:2007wv; @Ade:2012nxf; @Selig:2014qqa; @Calore:2014xka]. An orthogonal approach uses spectral templates to infer spatial characteristics of the various emission components [@Selig:2014qqa; @Huang:2015rlu; @deBoer:2016esu]. Template regression does in general allow a straightforward incorporation of observations at other frequencies or sky regions. A more ambitious approach is the full spatio-spectral modeling of gamma-ray emission, using numerical codes that simulate the generation, propagation and interaction of a multitude of cosmic-ray species in the Galaxy. This approach was pioneered with GALPROP [see, [*e.g.*]{}, @Strong:1998fr; @Ackermann:2012pya; @TheFermi-LAT:2015kwa]; modern independent codes like DRAGON [@Gaggero:2014xla; @Gaggero:2015nsa; @Evoli:2016xgn] and PICARD [@Kissmann:2014sia; @Werner:2014sya] build on that success (for a description of similarities and differences between the codes see ref. [@Evoli:2016xgn]). Both template regression and the full modeling of gamma-ray emission are used for parameter regression, usually of up to a few dozen parameters [@Ackermann:2012pya]. The overall performance of the models can be considered as reasonably good, with residuals typically at the level of $\sim30\%$ [@Ackermann:2012pya]. However, given the uncertainties in various emission components, which are, [*e.g.*]{}, related to errors in estimated gas densities [@1976ApJ...208..346G; @1988ApJ...324..248B; @Pohl:2007dz; @2016PASJ...68....5N], poorly constrained interstellar radiation field , simplifying assumptions in the physical cosmic-ray propagation models [[*e.g.*]{}, @Gaggero:2014xla; @Evoli:2012ha; @Acero:2016qlg], it is not a surprise that the models are far away from providing formally acceptable fits to the data [@Calore:2014xka]. One possible consequence of imperfect modeling is that estimators can be severely biased. At high Galactic latitudes, where the spatial characteristics of the background are reasonably well constrained and the relevant emission components (with the exception of, [*e.g.*]{}, Loop I [@1962MNRAS.124..405L; @2009arXiv0912.3478C]) are reasonably well known [see, [*e.g.*]{} @Ackermann:2014usa], one could argue that biases are under control and comparable to statistical uncertainties. However, close to and in the Galactic disk, where modeling the Galactic emission becomes increasingly important and statistical errors are small, bias can be rather significant. Strong residuals along the Galactic disk [see, [*e.g.*]{} @Calore:2014xka; @Acero:2016qlg] can potentially lead to wrong conclusions in particular about extended emission features. One possible approach to incorporate the characteristic magnitude of residuals as systematic error in statements about extended emission components was presented in [@Calore:2014xka], and based on a principal component analysis. Model-agnostic image reconstruction is a central tool to aid the visual interpretation of complex data. It is hence no surprise that the use of image reconstruction techniques spans disciplines. Their relevance extends across many research fields, including optical interferometry [[*e.g.*]{}, @Schutz:2014a], the analysis of gamma rays [[*e.g.*]{}, @1979MNRAS.187..145S; @Selig:2014qqa], and image reconstruction in medical positron emission tomography (PET) [[*e.g.*]{}, @Shepp:1982a]. Interestingly, the underlying formal problems are often very similar, which allows the transfer of algorithms and techniques. Image reconstruction is usually based on penalized likelihood maximization (or, in a Bayesian interpretation, maximum-a-posteriori reconstruction), often using Poisson or Gaussian likelihoods or one of their bias-reducing variants [@Slambrouck:2015a]. The number of parameters is usually very high ($\geq10^4$, often $\gg10^5$), as parameters correspond to individual pixels in the image reconstruction domain. This requires additional regularization conditions to (a) reduce shot noise, (b) prevent over-fitting of the data, and (c) define unique solutions in potentially under-constrained problems [[*e.g.*]{}, @Sra:2008a; @Afonso:2011a; @2013JOSAA..30..160T]. The regularization has the form of penalty terms in the likelihood, and its impact on the fit is controlled by various hyper-parameters. In a Bayesian interpretation, they correspond to priors on the regression parameters. Depending on the application, on the desired image properties, and on the available optimization algorithms, the regularization terms can take very different forms  [@Afonso:2011a; @2014MNRAS.439.3591C]. The arguably simplest regularizer is the $\ell_2$ norm, sometimes called ‘least square error method’ or ‘energy penalty’, which directly penalizes the variance of the model parameters [@Afonso:2011a; @Sra:2008a]. A traditional regularization technique popular in astronomy is the maximum entropy method (MEM) [@1979MNRAS.187..145S; @Cornwell:1985a; @Strong:1992a; @Buscher:1994a; @Hobson:1998a; @Lalush:2000a; @Bennett:2003ca; @Sra:2008a]. It aims to minimize the configurational entropy, providing a reconstructed image that is as ‘uninformative’ and featureless as possible, while still being compatible with the data. However, non-smooth and/or sparsity-enhancing regularizers have become increasingly important, since they aid in feature selection. In particular the simple $\ell_1$ regularization enhances the sparsity of features in the image, and is connected to compressed sensing [@Zhang:2013a]. The similar ‘total variation regularization’ enhances the sparsity of the image in the gradient domain [@Afonso:2011a; @2014MNRAS.439.3591C]. More complex variants of this approach enforce sparsity in the wavelet space [@Schutz:2015a]. However, explicit smoothing regularizers, based on finite differences and the $\ell_2$ norm, are also used. The induced effective smoothing has an exponential profile and is in general non-uniform, since it depends on the varying signal-to-noise ratio of the image [@Fessler:1996a; @Nuyts:2002a; @Asma:2004a; @Schutz:2014a]. In gamma-ray data analysis, a combination of various aspects of image reconstruction and parametric modeling is desirable for a number of reasons. In the case of existing spectral templates or models, fits in the spatial domain are practically an image reconstruction problem; a typical example is the [*Fermi*]{}bubbles [@2010ApJ...724.1044S; @Fermi-LAT:2014sfa]. But even in situations where spatial information about an emission component is available (from, [*e.g.*]{}, observations at other frequencies), the associated predictions have in general correlated uncertainties that can be fully modeled only with a very large number of nuisance parameters. An analysis of these nuisance parameters resembles again an image reconstruction problem. Indeed, various approaches that effectively combine aspects of image reconstruction and modeling exist in the literature. One example is the D$^3$PO algorithm [@Selig:2014qqa; @Huang:2015rlu], which is based on a Bayesian analysis using information field theory. The authors of this technique perform component separation, based on spectral information as well as priors on the smoothness of diffuse emission, to separate point source and diffuse components. Another example is ref. [@Bennett:2003ca], where the authors perform an analysis of multi-frequency WMAP data, using spectral models and the MEM for spatial regularization. In most gamma-ray analyses to date, nuisance parameters for spatial and/or spectral components of the models are not accounted for. Instead, a common approach is to ‘bracket uncertainties’ by focusing on a discrete set of plausible astrophysical configurations and analysis decisions. Each of these scenarios leads then to a different result for the observable of interest, and the envelope of all results defines the ‘systematic uncertainty band’ [[*e.g.*]{}, @Ackermann:2012pya; @Ackermann:2014usa; @Calore:2014xka; @TheFermi-LAT:2017vmf]. This approach is problematic in a number of ways. Usually, none of the models provides a statistically acceptable fit to the data. This raises the question why the consideration of various deficient models should lead to correct conclusions. Furthermore, it greatly hinders any rigorous model comparison, because the typically adopted statistical approaches (comparing the goodness-of-fit, considering likelihood ratios, or even Bayesian model comparison) exhibit undefined behaviour. In many cases, heuristic arguments can be made to support conclusions, but this does not replace a robust statistical approach. Lastly, the nuisance parameters are not only accounting for uncertainties in the model predications, but usually provide valuable information about the underlying astrophysical cause for the model deviations. This information is otherwise lost, or has to be guessed from residual maps. With the present work, we aim at providing a *first step* to overcome these limitations. In this article, we propose a new hybrid approach, which attempts to generalize and combine the benefits of many of the above techniques. Our approach incorporates intrinsic uncertainties of various model components as penalty terms on spatial and spectral template modulation parameters. It effectively allows the application of image reconstruction techniques to excesses above partially constrained backgrounds. In particular, we aim to (a) facilitate component separation in scenarios where only partial knowledge about the spatio-spectral characteristics of the components is available, and (b) introduce a sufficient number of nuisance parameters in the analysis such that we can obtain formally good fits. A good fit to the data is a necessary criterion for reliable model comparison and bias-free parameter estimation. Our approach is based on a penalized Poisson likelihood [[*e.g.*]{}, @Asma:2004a; @He:2016a] optimization with adaptive spatio-spectral (2-dim + 1-dim) templates. The spatial and spectral modulation parameters for the adaptive templates can be in general regularized with MEM, $\ell_2$ or explicit first- or second-order smoothing terms. Baring avoidable component degeneracies, we show that the underlying optimization problem is convex and hence has a unique (local and global) solution, despite the potentially millions of parameters. We perform analytic calculations of the gradient, which allows the use of the efficient optimization algorithm L-BFGS-B [@Byrd:1995a; @Zhu:1997a; @Morales:2011a]. Uncertainties and covariances of the fit parameters are sampled from the inverse of the Fisher information matrix, using sparse Cholesky decomposition [@Davis:2009a]. We study the properties of our approach with synthetic data. As the first application of our method, we analyze gamma-ray data from the [*Fermi*]{}Large Area Telescope (LAT), focusing on a Region of Interest (RoI) along the Galactic plane including the Galactic center. Our analysis is based on predictions from the cosmic-ray propagation code DRAGON [@Evoli:2016xgn], as well as a number of *ad hoc* templates that absorb residual emission in the inner Galaxy. We focus here on the technical aspects of obtaining a good fit to gamma-ray data, and will present a detailed physical interpretation of the results elsewhere. One of the purposes of this work is hence to define a simple reference Galactic diffuse emission model that we can refer to (and further improve) in future work. This paper is organized as follows: In section \[sec:skyfact\], we describe the adaptive template model, regularization terms in the likelihood function, the fitting strategy as well as our treatment of the covariance matrix of the problem. In section \[sec:gde\], we describe some details of the modeling of Galactic diffuse emission that we use in this work. In section \[sec:tomin\] we provide a step-by-step example for how to construct a model for gamma-ray emission in our RoI. In section \[sec:min\] we will discuss selected aspects of this model. In section \[sec:discussion\], we discuss the fit quality of the reference model. In section \[sec:conclusions\], we present our conclusions. In appendix \[apx:synthetic\] we present several studies using synthetic data to estimate possible biases of the method. SkyFACT – Concepts and definitions {#sec:skyfact} ================================== SkyFACT (Sky Factorization with Adaptive Constrained Templates) is based on penalized maximum likelihood regression, using spatial and spectral templates with additional nuisance degrees of freedom that account for uncertainties of these templates. Penalization terms act as priors on the nuisance parameters. Errors are estimated by sampling from the inverse Fisher information matrix of the best fit point. We discuss model definition, penalization terms, the optimization procedure and error calculations in detail in this section. Multi-linear modeling with adaptive templates --------------------------------------------- The diffuse component of the photon flux (per time and area) in energy bin $b$ and spatial pixel $p$ is given by a tri-linear model, $$\phi_{pb} = \sum_k T_{p}^{(k)} \tau^{(k)}_{p} \cdot S_{b}^{(k)} \sigma^{(k)}_{b} \cdot \nu^{(k)}\;, \label{eqn:diffmodel}$$ where $T_p^{(k)}$ and $S_b^{(k)}$ refer, respectively, to the spatial and spectral template for emission component $k$. Furthermore, $\tau_p^{(k)}$ and $\sigma_b^{(k)}$ are spatial and spectral modulation parameters, respectively, and $\nu^{(k)}$ is an overall component normalization factor. Note that, in general, the overall normalization factor cannot be absorbed into the spatial and/or spectral modulation parameters: In some cases the spatial and spectral uncertainties might be very small, while only the overall normalization remains as unconstrained fitting parameter. The only *strict* physical constraint on the modulation parameters and the normalization is non-negativity, $$\sigma_b^{(k)}, \tau_p^{(k)}, \nu^{(k)} \geq 0\;.$$ Further constraints are introduced by regularization terms in the likelihood, and will be discussed below. For diffuse components, the expected photon count in pixel $p$ and energy bin $b$ is given by $$\mu_{pb}^{\rm D} = \sum_{p'} \mathcal{P}_{bpp'} \mathcal{E}_{bp'} \phi_{p'b}\;,$$ where $\mathcal{E}_{bp}$ denotes the exposure in the energy bin $b$ and $\mathcal{P}_{bpp'}$ the effect of the instrument Point Spread Function, PSF (the probability that a photon from pixel $p'$ is measured in pixel $p$). We neglect here the effects of energy dispersion, although there is no conceptual difficulty in adding this when required. The modeling of point sources does not require spatial modulation parameters, so the model is only bi-linear. The expected number of signal photons in this case is given by $$\mu_{pb}^{\rm P} = \sum_{s} \mathcal{P}_{bp}({\boldsymbol{\Omega_s}}) \mathcal{E}_b({\boldsymbol{\Omega_s}}) \cdot S_b^{(s)} \sigma_b^{(s)} \cdot \nu^{(s)}\;,$$ where the sum is over point sources $s$ at angular locations ${\boldsymbol{\Omega_s}}$, and the PSF and exposure are evaluated at the source position. Note that the PSF matrix depends on the exact position of the source within the analysis pixel (whether it is at the center or at the edges). The total expected photon count is given by $\mu_{pb} = \mu_{pb}^{\rm D} + \mu_{pb}^{\rm P}$. To keep the notation compact, we use the symbol of the upper index ($s$ vs $k$) to discriminate between parameters related to point sources and related to diffuse components. Poisson likelihood and regularization terms ------------------------------------------- ![Various parameter regularization functions, as shown in Eqs. , and , assuming $\lambda=1$. We plot $\sqrt{\mathcal{R}_X}$ rather than $\mathcal{R}_X$, since this allows a simple interpretation of the associated penalty in terms of standard deviations. Throughout the paper, we adopt the convex MEM regularizer, which is motivated by the Maximum Entropy Method, as discussed in the text. It is a compromise between the classical $\ell_2$ regularizer, and the non-convex $\ln^2$ regularizer.[]{data-label="fig:MEM"}](figures/MEM.pdf){width="0.6\linewidth"} The likelihood function used in this work has two major parts, $$\ln\mathcal{L} = \ln\mathcal{L}_{\rm P}+ \ln\mathcal{L}_{\rm R}\;,$$ where $\mathcal{L}_{\rm P}$ corresponds to the Poisson likelihood that confronts model predictions with observations, and $\mathcal{L}_{\rm R}$ contains various regularization terms for the modulation parameters. The Poisson likelihood reads $$\ln\mathcal{L}_{\rm P} = \sum_{pb} c_{pb} - \mu_{pb} + c_{pb}\ln \frac{\mu_{pb}}{c_{pb}}\;, \label{eqn:lnLP}$$ where $\mu_{pb}$ and $c_{pb}$ refer, respectively, to the expectation value and the observed number of events. We emphasize that we use the full Poisson likelihood in our fits, and do not assume Gaussian errors. The definition is slightly different from the commonly adopted Cash-statistic [@Cash:1979vz], since we are actually considering the likelihood ratio conditioned on the observed data (in the sense that $\ln \mathcal{L}_P = 0$ for $\mu_{pb} = c_{pb}$). This version of the Poisson likelihood is commonly used in X-ray astronomy and is called the “c-statistic” in the widely-used spectral fitting tool XSPEC[^1] [@1996ASPC..101...17A]. It facilitates goodness-of-fit calculations, but has no effect on parameter regression. A central ingredient of our method are the large number of modulation parameters. They can either act as nuisance parameters that effectively account for systematic uncertainties in the adopted templates, or they permit functionality that is akin to image reconstruction. However, even in the case of image reconstruction, it remains important to regularize the modulation parameters in order to reduce the effects of Poisson noise and avoid over-fitting of the data. In the present work, the regularization terms are defined as $$\begin{gathered} -2\ln\mathcal{L}_R = \sum_{k} \lambda_k \mathcal{R}_X({\boldsymbol{\tau}}^{(k)})+ \lambda'_k \mathcal{R}_X({\boldsymbol{\sigma}}^{(k)})+ \lambda''_k \mathcal{R}_X(\nu^{(k)})+ \eta_k \mathcal{S}_1({\boldsymbol{\tau}}^{(k)})+ \eta'_k \mathcal{S}_2({\boldsymbol{\sigma}}^{(k)})\\ +\sum_{s} \lambda'_s \mathcal{R}_X({\boldsymbol{\sigma}}^{(s)})+ \lambda''_s \mathcal{R}_X(\nu^{(s)})+ \eta'_s \mathcal{S}_2({\boldsymbol{\sigma}}^{(s)})\;, \label{eqn:lnLR}\end{gathered}$$ where $X$ refers to the type of regularization, for each term separately. The first three regularization hyper-parameters, $\lambda_k$, $\lambda'_k$ and $\lambda''_k$, control respectively the constraints on the spatial, spectral and overall modulation parameters, for the diffuse components. The regularization hyper-parameters $\eta_k$ and $\eta'_k$ control the strength of the spatial and spectral smoothing, respectively. Explicit smoothing is for instance useful to reduce the impact of Poisson noise on the spatial modulation parameters, or to enforce a physically motivated smoothness of a given emission component. The point source regularization hyper-parameters, $\lambda'_s$, $\lambda''_s$ and $\eta'_s$, have an analogous functionality for point sources. We use the vector notation, ${\boldsymbol{\tau}}^{(k)}$ etc, to indicate that the regularization and smoothing terms depend in general on all parameters simultaneously. We briefly discuss three choices for the form of the regularization function $\mathcal{R}_X$, although only one of them (MEM) is used in the rest of this paper. The three choices are equivalent for large values of the regularization parameters, but differ in the tails, which is relevant if the regularization hyper-parameters are small and therefore the modulation parameters not tightly constrained. The $\ell_2$-norm regularization is quite commonly used in the literature [@Afonso:2011a; @Sra:2008a]. For our parameter choices, it has the simple form $$\lambda \mathcal{R}_{\ell_2}({\boldsymbol{x}}) = \lambda \sum_i (x_i-1)^2\;, \label{eqn:R_l2}$$ where $i$ runs over the elements of ${\boldsymbol{x}}$. This regularization is a convex function, with the minimum at $x_i=1$; the associated variance is $1/\sqrt{\lambda}$. It can be used to constrain the element-wise variance of ${\boldsymbol{x}}$, and corresponds to Gaussian priors in a Bayesian interpretation. However, it behaves poorly when parameters are supposed to lie within a logarithmic range, [*e.g.*]{}, one or two dex. For that reason we do not use this regularization in the present work. A naive modification of the above $\mathcal{R}_{\ell_2}(\cdot)$ regularization that constrains parameters on logarithmic scales is obtained by applying the $\ell_2$ in log-space of the modulation parameters. This yields, $$\lambda\mathcal{R}_{\ln^2}({\boldsymbol{x}}) = \lambda \sum_i \ln^2 x_i\;, \label{eqn:R_ln2}$$ and has a global minimum at $x_i=1$. The second derivative is given by $2\lambda(1-\ln x_i)/x_i^2$. This means that again for large values of $\lambda$, this regularization is equivalent to the above linear $\ell_2$ one. For small values of $\lambda$, it corresponds to a quadratic penalty term in the log-space of $x_i$, which is useful to constrain parameter ranges within a few dex, if necessary. However, a disadvantage of the regularization term is that it is not convex (the second derivative becomes non-positive for $\ln x_i \geq 1$). This spoils the convexity of the entire optimization problem, potentially introducing multiple local minima. For this reason, we do not consider this regularizer further in the present work. When the magnitude of an emission component is largely unknown and should be inferred from the data, a well motivated regularization is based on the maximum entropy method [@1979MNRAS.187..145S; @Strong:1992a; @Buscher:1994a]. It reads $$\lambda\mathcal{R}_{MEM}({\boldsymbol{x}}) = 2\lambda\sum_i 1- x_i + x_i \ln x_i\;. \label{eqn:R_MEM}$$ The MEM regularizer is convex, and has a minimum at $x_i=1$. The variance for large values of $\lambda$, as derived from the second derivative at $x_i=1$, is $1/\sqrt{\lambda}$. It hence becomes equivalent to the $\ell_2$ and $\ln^2$ regularization in the large-$\lambda$ limit. For smaller values of $\lambda$, the main effect of this regularization term is to smooth out regions with low flux levels, while still permitting pronounced bright regions if they are preferred by the data. A good intuition can be obtained by noticing that its first and second derivatives are given by $2\lambda\ln x_i$ and $2\lambda/x_i$, respectively. For very small values of $x_i$, $\mathcal{R}({\boldsymbol{x}})_{MEM}$ hence steepens infinitely and prevents $x_i$ from completely approaching zero. For very large values of $x_i$, on the other hand, the Poisson part of the likelihood can take control, such that compact bright regions remain relatively unconstrained. Due to its numerous practical properties, we use the MEM regularizer in the current work, not only for the spatial modulation parameters, but for all parameters. In figure \[fig:MEM\], we show how the MEM regularizer behaves relative to the other regularizers discussed above. The smoothness of components can be imposed by constraining the *gradient* of the modulation parameters instead of their range. For the spatial modulation parameters, this is enforced by terms of the form $$\eta\mathcal{S}_1({\boldsymbol{x}}) = \eta\sum_{(p,p')\in \mathcal{N}} (\ln x_p - \ln x_{p'})^2\;,$$ where $\mathcal{N}$ refers to the set of nearest-neighbour pixel pairs. Here, only pairs are included where both pixels are actually part the template (which can only cover a fraction of the sky). Note that we constrain the gradient of the *logarithm* of the modulation parameters instead of the modulation parameters directly. This has the advantage that the overall normalization of the modulation parameters drops out. The disadvantage is a loss of convexity, like for the above $\ln^2$ regularization. This will be discussed below. The spectral smoothing terms that we adopt in this work, on the other hand, constrain the *second* derivative of the logarithm of the modulation parameter, and not the gradient. The corresponding regularization term has the form $$\eta\mathcal{S}_2({\boldsymbol{x}}) = \eta \sum_b (\ln x_{b-1} - 2 \ln x_b + \ln x_{b+1})^2\;.$$ Assuming that the energy bins are logarithmically separated, this regularization favours power-law spectra. Since many astrophysical spectra actually have the form of a (potentially rolling) power-law, this is a reasonable behaviour that favours in many cases ‘more physical’ spectra. However, it has to be used with care since it will cause an automatic ‘power-law extrapolation’ when the shot-noise level increases at high energies. The strength of all regularization terms are controlled with the regularization hyper- parameters $\lambda_k$, $\lambda'_k$, $\lambda''_k$, $\eta_k$ and $\eta'_k$ for diffuse components, and $\lambda'_s$, $\lambda''_s$, $\eta'_s$ for point sources. We will discuss below, and in appendix \[apx:synthetic\], a few examples of typical values and their effects on the resulting reconstruction. Parameter optimization with L-BFGS-B and convergence conditions --------------------------------------------------------------- Internally, the tri-linear model for the calculation of expected photon counts from diffuse emission components is represented by three matrices $A^{(1)}$, $A^{(2)}$ and $A^{(3)}$. These map model parameters ${\boldsymbol{\theta}} \equiv ({\boldsymbol{\tau}}^{(k)}, {\boldsymbol{\sigma}}^{(k)}, \nu^{(k)}, {\boldsymbol{\sigma}}^{(s)}, \nu^{(s)})^T$ onto the expected photon fluxes ${\boldsymbol{\phi}}^D \equiv (\phi_{bp})$ via $$({\boldsymbol{\phi}}^D)_i = (A^{(1)} {\boldsymbol{\theta}})_i (A^{(2)} {\boldsymbol{\theta}})_i (A^{(3)} {\boldsymbol{\theta}})_i\;.$$ The transformed vectors are multiplied element-wise, as indicated by the index $i$. Although the matrices $A^{(1,2,3)}$ are very large (in our example of the order $10^5\times10^5$), they are rather sparse (in our example with a sparsity well above $99\%$) and can be efficiently stored in the compressed sparse column format. The use of optimized sparse matrix routines leads then to very fast evaluation times. Another advantage of the above representation is its simple analytical form, which facilitates the calculation of gradients and the Fisher information matrix. The final expected number of photons, ${\boldsymbol{\mu}}^D = (\mu_{bp}^D)$ are obtained from the matrix product $${\boldsymbol{\mu}}^D = \sum_j P_{ij} (\phi^D)_j ({\boldsymbol{E}})_j\;,$$ where $P$ and ${\boldsymbol{E}}$ represent the PSF and the exposure map convolution in data space respectively. For the PSF convolution, we assume periodic boundary conditions for simplicity, which leads to minor artefacts at the borders of the RoI, but is not relevant for the present discussion (although it will some impact on the formal goodness-of-fit). Point sources are calculated in a similar way, but here two matrices $B^{(1)}$ and $B^{(2)}$ are sufficient, since the corresponding model is only bi-linear. For the models below, we find that the evaluation of the likelihood function and the gradient takes less than a second on one core of a typical modern computer. Sparse matrix multiplication is not straightforward to parallelize on shared or distributed memory systems, which makes it somewhat challenging (but certainly not impossible [[*e.g.*]{}, @ballard2013communication]) to reduce the computation time further, even on multi-core systems. Parameter optimization in SkyFACT is performed with the L-BFGS-B method [@Byrd:1995a; @Zhu:1997a; @Morales:2011a]. The L-BFGS-B algorithm is based on the quasi-Newtonian method of Broyden, Fletcher, Goldfarb and Shanno (BFGS) [@Wright:1999a]. Its limited-memory version (L-BFGS) does not directly store numerical approximations to the Hessian matrix, but rather a number of previously calculated gradients to construct approximate Hessian matrices on the fly. This makes it possible to use it for very large number of parameters ($10^6$ and more). Furthermore, the version used here (L-BFGS-B) supports parameter boundaries, using the active-set method. This is essential in order to be able to impose the non-negativity constraints. We use here the implementation from Refs. [@Zhu:1997a; @Morales:2011a; @Jones:2001--a]. The algorithm takes both the objective function and its derivative as input, which are both straightforward to calculate based on the above analytical expressions. There are two main convergence criteria that we consider in this work. In the simplest case, one can just require that the vertical step-size of the last iteration is below a certain threshold value (this is one of the criteria of the adopted L-BFGS-B algorithm). A more informative quantity is however the estimated (vertical) distance to the minimum (EDM), which is known from the popular Minuit algorithm [@James:1975dr]. To estimate the EDM properly requires information about the full Hessian, which is however not easily available outside the L-BFGS-B implementation that we use. We therefore adopt an extremely simple heuristic definition of an ‘effective EDM’ that is motivated by the one-dimensional case. It is defined as $$\text{EDM} \equiv \frac{||{\boldsymbol{\nabla}} f ||_2^2}{2}\frac{||\delta{\boldsymbol{x}}||_2} {||\delta {\boldsymbol{\nabla}} f||_2}\;,$$ where $\delta {\boldsymbol{x}}$ refers to the parameter changes in the last step, $\delta {\boldsymbol{\nabla}} f$ to the change of the gradient, and $||\cdot||_2$ is the $\ell_2$ norm. We find that although it is only a crude (noisy, not monotonically decreasing) estimate of the vertical distance to the minimum, it is a useful dimensionless estimator to determine how close the fit is to the global minimum. Error estimation with Cholesky decomposition -------------------------------------------- Estimating parameter uncertainties requires knowledge of the full covariance matrix of the model. Here, we approximate the covariance matrix by the inverse of the Fisher information matrix. The Fisher matrix encodes the *expected* error, and is independent of the data (it is numerically faster to calculate than the full Hessian matrix of the likelihood function, and numerically more stable even when the minimum of the system has not been exactly reached). Formally, it is given by $$\mathcal{I}_{ij} = -\left\langle\frac{\partial^2}{\partial \theta_i \partial \theta_j} \ln\mathcal{L}\right\rangle_{\mathcal{D}({\boldsymbol{\theta}})}\;,$$ where the average is taken over mock data from the best-fit model, $\mathcal{D}({\boldsymbol{\theta}})$. The average can be taken analytically, and the resulting expressions are relatively simple to handle (see, [*e.g.*]{}, ref. [@Fisher] for a detailed discussion in context of Poisson likelihoods). Since the Fisher matrix is composed of sparse matrices, it is sparse as well, at the 99% level in the examples discussed below. We estimate the errors of individual model parameters as well as predicted fluxes and other quantities using a sampling method. This circumvents the need to explicitly calculate the covariance matrix (this would be computationally very costly, since the inverse of a sparse matrix is not sparse itself). The procedure is based on sparse Cholesky decomposition [@Davis:2009a]: The Fisher information matrix is decomposed in the form $P\mathcal{I}P^T = L D L^T$, where $L$ is a lower unit triangular matrix, $D$ is a positive-definite diagonal matrix, and $P$ is a fill-reducing permutation matrix that maximizes the sparsity of $L$. Based on the matrices $L$, $D$ and $P$, one can efficiently sample model parameter vectors $\delta{\boldsymbol{\theta}}$ such that their covariance is given by the inverse, $\mathcal{I}^{-1}$. These model parameters describe then deviations from the best-fit model, similar to samples from Bayesian posterior distributions. The sampling procedure works as follows. We select a vector of standard normal distributed parameters, ${\boldsymbol{x}} \sim \mathcal{N}(0, 1)$, with the same length as the number of model parameters. Next, we obtain ${\boldsymbol{y}}$ as solution of $L^T {\boldsymbol{y}} = D^{-1/2} {\boldsymbol{x}}$, which can be done by back-insertion. Finally, we calculate $\delta {\boldsymbol{\theta}} = P^T {\boldsymbol{y}}$. We repeat the process as many times as necessary to obtain a reasonably-sized sample. As mentioned above, the full covariance matrix, $\Sigma_{ij}$, of the model parameters ${\boldsymbol{\theta}}$ can be estimated from $\langle \delta \theta_i \delta\theta_j \rangle\approx \Sigma_{ij}$, where the average is taken over many samples. In practice, we are usually interested in mean values and variances of model fluxes, which are functions of ${\boldsymbol{\theta}}$. Mean values for fluxes and other model predictions are derived from the best-fit value of ${\boldsymbol{\theta}}$. Variances are derived by calculating model predictions for ${\boldsymbol{\theta}}+\delta{\boldsymbol{\theta}}$ and averaging over many samples. Convex optimization and non-degeneracy conditions ------------------------------------------------- Convex optimization is the minimization of convex functions, and has a number of useful properties. In particular, any local minimum is also a global minimum, meaning that there is only one best-fit solution to the problem. We will show here that – as long as the model components remain sufficiently non-degenerate – the objective function is indeed convex. We will point out qualitatively under what conditions problems can occur (one can also derive quantitative conditions, but the resulting expressions are somewhat lengthy and we do not reproduce them here). We first consider the Poisson likelihood $\mathcal{L}_P$. To simplify the notation, it is useful to write the $\ln \mathcal{L}_P$ in the unbinned form (we refer to ref. [@Fisher] for a discussion of unbinned Poisson likelihoods). For simplicity, we neglect point source degrees of freedom (which are very sparse in the sky and not expected to be a problem in the present scenario), as well as the overall normalization factor. We are in particular interested in the differential version of the Hessian. It is given by $$-\sum_{ij} \delta \theta_i \delta \theta_j \frac{\partial^2}{\partial \theta_i\partial \theta_j} \ln\mathcal{L}_P = \int d\Omega\, dE\, \frac{\mathcal{C}}{\Phi^2}\delta \Phi^2 + \frac{\Phi-\mathcal{C}}{\Phi} \sum_k \frac{\delta\tau_k}{\tau_k}\frac{\delta\sigma_k}{\sigma_k}\Phi_k\;. \label{eqn:hessianP}$$ Here, we defined the ‘unbinned counts map’ $\mathcal{C}(\Omega,E) \equiv \sum_i \delta^{(2)}(\Omega-\Omega_i)\delta(E-E_i)$, where $E_i$ and $\Omega_i$ are the energy and position of photon $i$, and $\delta(\cdot)$ refers to Dirac delta distributions. Furthermore, $\delta \Phi \equiv \delta{\boldsymbol{\theta}} \cdot {\boldsymbol{\nabla}}_\theta \Phi$, and $\delta \tau_k$ and $\delta\sigma_k$ are defined analogously. Finally, $\Phi_k$ is the contribution to $\Phi$ from component $k$. It is clear that the first term in Eq.  is always non-negative, and it is strictly positive unless two or more model components are exactly degenerate (in that case some change in component $k$ can be exactly undone by the opposite change in component $k'$, leading to $\delta\Phi=0$ in that particular direction of the parameter vector $\delta{\boldsymbol{\theta}}$). The second term can, depending on $\mathcal{C}$, however become negative and spoil the convexity of the objective function. However, it depends on the simultaneous change of spectral and spatial modulation parameters of the *same* component. It is exactly zero for any component that has either spatial or spectral degrees of freedom, but not both. As long as the model consists only of components that each have either a fixed distinct morphology or spectrum, the optimization problem is strictly convex. However, if both spectral and spatial variations are allowed for at least two components in the fit, one can find in general degenerate transformations with $\delta{\boldsymbol{\Phi}} = 0$ while $\delta\tau_k\neq0$ and $\delta\sigma_k\neq0$. In that case, the second term of Eq.  can potentially become negative, and the objective function could have multiple local minima. This corresponds to the case where components have so much spectral and spatial freedom that they can effectively exchange their roles in the fit. Situations like that are usually not desired, and should be prevented by the regularization of the modulation parameters. As discussed above, both the MEM and $\ell_2$ regularizations are completely convex, and hence cannot spoil the convexity of the problem. Since we use the MEM regularization, increasing the value of the corresponding hyper-parameters $\lambda$ can only increase the convexity of the problem, while breaking the degeneracy between model components. Note, however, that the smoothing terms $\mathcal{S}$ are *not* generally convex. This is caused by the fact that the logarithm of the modulation parameters ([*e.g.*]{}, $\ln \tau_p^{(k)}$) enters their definition, which is similar to the above $\ln^2$ regularization. For this reason, it is advisable to check explicitly that the global minimum is actually reached. This can be done by confirming that reducing the smoothing hyper-parameters $\eta$ and $\eta'$ does not qualitatively change the result (although it will reduce the smoothness of the components in the fit; see appendix \[apx:run5\].) Modeling Galactic diffuse emission {#sec:gde} ================================== Data selection -------------- ![Counts map of data used in the analysis.[]{data-label="fig:data"}](figures/LAT_map_counts.pdf){width="0.99\linewidth"} We use $7.6$ years of [*Fermi*]{}-LAT Pass 8 data from 2008 August 4 and 2016 March 3. We apply standard data cuts to events recommended by the [*Fermi*]{}Science Support Center (“`(DATA_QUAL>0) && (LAT_CONFIG==1)`” and zenith angles $>90^{\circ}$). We select only ULTRACLEAN events (evclass=512) and use both FRONT and BACK converted events (evtype=3). We use the [*Fermi*]{}Science Tools[^2] (v10r0p5) to build count and exposure maps binned into 25 log-spaced energy slices in the energy range $0.34-228.65$ GeV. Our RoI spans $|\ell| \leq 90^\circ$ and $|b|\leq 20.25^\circ$. The data are binned in Cartesian coordinates with a pixel size of $0.5^\circ$. In figure \[fig:data\], we show the data used in this analysis. Note that the adopted pixelization ensures that the bright Galactic disk falls onto the central row of pixels. Since we use a relatively small number of pixels in the analysis (the number of spatial nuisance parameters would otherwise explode in our current treatment), using a pixelization scheme that takes into account the characteristic emission gradient towards the disk is particularly important. Foreground modeling ------------------- We model the foreground components following standard methods and codes. In particular, we rely on the prescriptions of the GALPROP [@Moskalenko:2001ya] and DRAGON [@Evoli:2008dv] codes, and refer to the [*Fermi*]{}-LAT Collaboration works, [*e.g.*]{} [@Ackermann:2012pya]. Details of the modeling are explained below. The gamma-ray emission from interactions of cosmic rays (CR) with the interstellar medium (ISM) and low-energy ambient photons is the so-called Galactic diffuse emission (GDE). At GeV energies, the GDE originates mainly from three processes: (1) the decay of neutral pions produced by collisions of CR protons with ISM; (2) the bremsstrahlung radiation of electrons in the electric field of ISM charged particles; and (3) the up-scattering to GeV energies of the low-energy ambient photons of the interstellar radiation field (ISRF) due to Inverse Compton scattering (ICS) off CR electrons. The GDE depends on the production of CR is the Galaxy, their distribution and spectra, as well as on the propagation mechanisms in the ISM and the interactions with the environment. In the present work, we model the GDE in a relatively minimalistic way, and explore what ingredients we need to obtain a reasonably good fit to the data (after nuisance parameters are taken in account). #### Hadronic $\pi^0$ emission. The spatial distribution of the hadronic emission from $\pi^0$ decay traces closely the target material distribution, i.e. the distribution of hydrogen in the Galaxy. The hydrogen in the Galaxy is in the form of atomic (HI), molecular (H$_2$), and ionized (HII) gas, traced by different observables, namely the 21 cm line, the 2.6 mm CO emission line, and pulsar dispersion measurements respectively [@Ferriere:2001rg]. The most abundant component is represented by the molecular and atomic gas (almost 70% by mass), while ionized H is subdominant in mass (only few percent on average). The total H column density (in cm$^{-2}$) can be approximated by the sum of atomic (HI) and molecular hydrogen (H$_2$): $$N_{\rm H} \simeq N_{\rm HI} + 2 N_{\rm H_2} = N_{\rm HI} + 2 X_{\rm CO} W_{\rm CO} \, ,$$ where the molecular H$_2$ column density is assumed to be proportional to the integrated line intensity of CO, $W_{\rm CO}$ (in K km/s ), through the $X_{\rm CO}$ coefficient (in cm$^{-2}$/(K km/s)). The $X_{\rm CO}$ is typically considered to be constant thorough the Galaxy. However, there are no *a priori* reasons for it to be such and, indeed, gamma-ray analyses have shown that, most likely, it is spatial dependent [@Strong:2004td; @2016PASJ...68....5N; @Yang:2016jda; @Acero:2016qlg]. In the present work, we build the $\pi^0$ template as the sum of the gas column densities for atomic and molecular hydrogen, available within the GALPROP public release.[^3] The HI column density and $W_{\rm CO}$ integrated line intensity maps are based on the corresponding 2D analytical models developed in ref. [@Moskalenko:2001ya], renormalised to the Leiden-Argentine-Bonn 21 cm line survey [@Kalberla:2005ts] and to the 2.6 mm CO line survey [@Dame:2000sp], respectively. For the HI, the spin temperature used to correct for opacity of the 21 cm line is $T_S = 125$ K. No dust-reddening correction is applied to the column density. Furthermore, *we do not apply a dark gas correction*, since we are instead interested in seeing how dark gas affects the spatial modulation parameters associated with hadronic emission . The GALPROP gas maps are binned in Galacto-centric radial annuli, using the distance information extracted from the gas velocity and rotation curve (for details about the methodology see appendix B of ref. [@Ackermann:2012pya]). In practice, we obtain the gas template by adding the hydrogen column densities as: $N_{\rm HI} + 2 N_{\rm H_2}$, assuming a constant $X_{\rm CO} = 1.9 \times 10^{20}$ cm$^{-2}$/(K km/s). We build three gas templates using the following radial binning: the first gas ring corresponds to the emission within 0–3.5 kpc from the GC, the second ring to 3.5–6.5 kpc, and the third one to 6.5–19 kpc. The advantage of using radial rings is that we can combine them arbitrarily to get maps for different spherical shells about the GC and study how this affects our conclusions. #### Leptonic ICS emission. To model the ICS emission we use the publicly available DRAGON code [@Evoli:2008dv] and GammaSky [@Evoli:2012ha; @2013JCAP...03..036D], a custom numerical package interfaced to DRAGON to compute the gamma-ray emission from the CR distribution output of DRAGON. We use the “Ferrière” model for the source distribution from the primary components [@Ferriere:2001rg], and propagation parameters corresponding approximately to the “KRA4” model in ref. [@2013JCAP...03..036D].[^4] The ISRF model adopted by DRAGON/GammaSky is documented in [@Porter:2005qx] and available for download from the GALPROP public website (v54). #### Leptonic bremsstrahlung emission. Starting from the same DRAGON model file and using the GALPROP gas maps as target for interactions, we model the bremsstrahlung emission of electrons and positrons off interstellar ions and nuclei. The gamma-ray emission from bremsstrahlung is typically suppressed with respect to that from $\pi^0$ and ICS. However, it might play a significant role in the very inner region of the Galaxy; see for example ref. [@2013ApJ...762...33Y]. We tested the effect of including this additional template in our baseline model, without improvement of the fit that would be relevant for the current work (for our purposes, it is enough that bremsstrahlung can be absorbed by the $\pi_0$ templates). Therefore, we do not include the bremsstrahlung template in our baseline model. ![Initial model templates for fits of 3–5. For 1 and 2, only one gas emission template is used, which is equal to the sum of the three gas templates shown here. Gas ring I corresponds to the emission from the innermost shell, 0–3.5 kpc from the Galactic center. Gas ring II correponds to emission from 3.5–6.5 kpc, and gas ring III to emission from 6.5–19 kpc.[]{data-label="fig:templates"}](figures/r5_initial_map_mean_initial_pi0ii.pdf "fig:"){width="0.49\linewidth"} ![Initial model templates for fits of 3–5. For 1 and 2, only one gas emission template is used, which is equal to the sum of the three gas templates shown here. Gas ring I corresponds to the emission from the innermost shell, 0–3.5 kpc from the Galactic center. Gas ring II correponds to emission from 3.5–6.5 kpc, and gas ring III to emission from 6.5–19 kpc.[]{data-label="fig:templates"}](figures/r5_initial_map_mean_initial_pi0i.pdf "fig:"){width="0.49\linewidth"} ![Initial model templates for fits of 3–5. For 1 and 2, only one gas emission template is used, which is equal to the sum of the three gas templates shown here. Gas ring I corresponds to the emission from the innermost shell, 0–3.5 kpc from the Galactic center. Gas ring II correponds to emission from 3.5–6.5 kpc, and gas ring III to emission from 6.5–19 kpc.[]{data-label="fig:templates"}](figures/r5_initial_map_mean_initial_pi0o.pdf "fig:"){width="0.49\linewidth"} ![Initial model templates for fits of 3–5. For 1 and 2, only one gas emission template is used, which is equal to the sum of the three gas templates shown here. Gas ring I corresponds to the emission from the innermost shell, 0–3.5 kpc from the Galactic center. Gas ring II correponds to emission from 3.5–6.5 kpc, and gas ring III to emission from 6.5–19 kpc.[]{data-label="fig:templates"}](figures/r5_initial_map_mean_initial_ICS.pdf "fig:"){width="0.49\linewidth"} The spatial templates used as starting points in the fits are shown in figure \[fig:templates\]. The energy range shown is 0.96 – 12.98 GeV, chosen to highlight both the $\pi^0$ bump as well as a significant fraction of the Galactic center excess emission (discussed below). Although DRAGON also produces the energy spectra of the various emission components, we do not use them in the current work. The energy spectra depend critically on the energy spectra of hadronic and leptonic CRs, and hence on propagation and injection. Instead, we use the $\pi^0$ and ICS component spectra from [@Ackermann:2012pya] as references. These spectra are obtained from the gamma-ray analysis of CR propagation models that are also in agreement with local CR measurements, and hence represent for us a good starting point. However, CR spectra in the Galaxy might be in general quite different from what we measure locally. For this reason, we use the reference spectra only as starting points in the fit, and allow considerable freedom such that they can adapt to the gamma-ray data if needed. Towards a simple model for the Galactic disk emission {#sec:tomin} ===================================================== In this section, we derive a simple reference model for Galactic diffuse emission in our main RoI. Since even this reference model will have $10^5$ parameters, the meaning of the term ‘simple’ is not automatically evident. What we actually aim to find is a minimal set of physical components, with distinct spatial or spectral characteristics (these contribute a very low number of parameters), augmented with realistically sized nuisance parameters (this number is necessarily very large). The goal is to remove most of the residuals in our RoI in a physically motivated way. Our reference model is not meant to be the final answer to modeling all of the gamma-ray emission in our RoI, but rather as a base for a number of follow-up studies. The starting point for our reference model will be the numerical predictions for hadronic and leptonic emission in the Galaxy from DRAGON, based on locally measured CR fluxes and HI and CO gas maps, as discussed in section \[sec:gde\]. We will study how much of the observed gamma-ray emission is already explained by these *a priori* predictions, both directly and when augmented with additional nuisance parameters that account for expected spatial and spectral uncertainties in the predictions. For a number of irreducible residuals, particularly in the bulge region, we will introduce additional templates. Conventional diffuse emission components with nuisance parameters ----------------------------------------------------------------- ---------------------------------------------------- -- --------- ------------------ ------------------ ------------------ ------------------ -- -- Components 1 2 3 4 5 IGRB 00 00 00 00 00 \[0.1cm\] 3FGL PSC 00 00 00 00 00 \[0.1cm\] Gas (0–19 kpc) 000 0[25]{}0 — — — \[0.1cm\] Gas ring I (0–3.5 kpc) — — 0[25]{}0 0[25]{}0 0[25]{}0 \[0.1cm\] Gas ring II (3.5–6.5 kpc) — — 0[25]{}0 0[25]{}0 0[25]{}0 \[0.1cm\] Gas ring III (6.5–19 kpc) — — 0[25]{}0 0[25]{}0 0[25]{}0 \[0.1cm\] Extended sources — — — 0140 0140 \[0.1cm\] Inverse Compton 000 1[16]{}0[100]{}0 1[16]{}0[100]{}0 1[16]{}0[100]{}0 1[16]{}0[100]{}0 \[0.1cm\] [*Fermi*]{}bubbles — — — 0[400]{}40 0[400]{}40 \[0.1cm\] 511 keV template — — — — 000 \[0.1cm\] Naive model parameters, $N_\text{param}$ 20253 78573 97838 104596 107639 Naive DOF 708747 650427 631162 624404 621361 Eff. model parameters, $N_\text{param}^\text{eff}$ 1900 11800 10200 12700 12800 Eff. data bins, $N_\text{data}^\text{eff}$ 624700 622700 620200 618800 619000 Eff. DOF, $k$ 622800 610900 610000 606100 606200 $-2\ln\mathcal{L}_P$ 1016041 637742 633334 627206 626998 $-2\ln\mathcal{L}_R$ 14152 24652 23709 21640 20988 Model fidelity, $\mathcal{F}$ 627 164 153 145 144 ---------------------------------------------------- -- --------- ------------------ ------------------ ------------------ ------------------ -- -- : Summary of components and associated regularization hyper-parameters used for the models 1 to 5 in this work. For each component, we show in the compact matrix notation the adopted regularization hyper-parameters for the spatial ($\lambda$), spectral ($\lambda'$) and overall ($\lambda''$) modulation parameters, and for the spatial ($\eta$) and spectral ($\eta'$) smoothing parameters. The definition of hyper-parameters is shown in Eq. . For the spatial and spectral hyper-parameters, an ‘$\infty$’ indicates that the corresponding modulation parameters are kept fixed to one, while for the smoothing parameters, a ‘0’ indicates no smoothing on that component. We also show the total number of fit parameters $N_\text{param}$, the naive number of DOF ($N_\text{ebin}\times N_\text{pix} - N_\text{par}$), the effective number of DOF as estimated from mock data along with the effective number of model parameters and data bins, the values of the Poisson and regularization parts of the likelihood function, and the model fidelity, $\mathcal{F}$ as an indicator for the goodness-of-fit (see section \[sec:gof\]). Effective fit parameters are rounded to the nearest hundred. []{data-label="tab:fits"} In table \[tab:fits\], we list five increasingly complex models, their components and the associated regularization parameters (1 to 5). They provide a step-by-step illustration for how additional components and nuisance parameters account for residuals when modeling the gamma-ray emission observed by [*Fermi*]{}-LAT in our RoI. We will discuss each of the five models separately. ![Baseline spectra in our initial 1, in two different sky regions. As show in table \[tab:fits\], only the normalization of the various components is left free to vary. Residuals are in units of standard deviations, defined as $\sigma = (\text{data}-\text{model})/\sqrt{\text{model}}$.[]{data-label="fig:run1_spectra"}](figures/r1_spectra_fullROI.pdf "fig:"){width="0.49\linewidth"} ![Baseline spectra in our initial 1, in two different sky regions. As show in table \[tab:fits\], only the normalization of the various components is left free to vary. Residuals are in units of standard deviations, defined as $\sigma = (\text{data}-\text{model})/\sqrt{\text{model}}$.[]{data-label="fig:run1_spectra"}](figures/r1_spectra_CentralRegion.pdf "fig:"){width="0.49\linewidth"} 1\. This is our initial model, mostly based on the *a priori* information that enters the modeling with DRAGON. It includes just one gas and one ICS component, along with a component for the isotropic gamma-ray background (IGRB) as measured in ref. [@Ackermann:2014usa] and the 3FGL point sources [@Acero:2015hja]. In the case of the IGRB, we allow spectral uncertainties with a regularization of $\lambda = 16$ (this corresponds to $1/\sqrt{\lambda}=25\%$ uncertainty) that are somewhat smaller than the systematic errors quoted in ref. [@Ackermann:2014usa]. This prevents large deviations from the measured spectrum. The spectra of the gas and ICS components are constrained to be close to the spectra corresponding to the locally measured CRs (again within $25\%$), and are taken from ref. [@Ackermann:2012pya]. For all of the components of 1, except the IGRB, we keep the overall normalization, $\nu^{(k)}$, free in the fit. The same is the case for the point sources. Furthermore, we use a weak regularization to keep the source spectrum close to its original spectrum in the 3FGL (to within $20\%$). The total number of model parameters is 20253 (77 for the diffuse components, and the remaining 20176 for the 776 3FGL point sources). We show the resulting spectra and the spectral residuals in figure \[fig:run1\_spectra\]. It can be clearly seen that very significant residuals remain in particular in the central region. The overall spectrum integrated over the full RoI is better recovered, however only at the expense of a strongly distorted IGRB spectrum. ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r1_map_residuals_0s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r1_map_residuals_1s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r2_map_residuals_0s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r2_map_residuals_1s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r3_map_residuals_0s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r3_map_residuals_1s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r4_map_residuals_0s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r4_map_residuals_1s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r5_map_residuals_0s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>0.34$ GeV (left column) and $>1.24$ GeV (right column).[]{data-label="fig:residuals_low"}](figures/r5_map_residuals_1s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r1_map_residuals_5s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r1_map_residuals_20s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r2_map_residuals_5s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r2_map_residuals_20s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r3_map_residuals_5s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r3_map_residuals_20s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r4_map_residuals_5s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r4_map_residuals_20s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r5_map_residuals_5s.pdf "fig:"){width="0.49\linewidth"} ![Significance of residuals for 1 (top row) to 5 (bottom row), for energies $>4.57$ GeV (left column) and $>16.85$ GeV (right column).[]{data-label="fig:residuals_high"}](figures/r5_map_residuals_20s.pdf "fig:"){width="0.49\linewidth"} ![Histogram of residual significance for 1 (left panel) and 5 (right panel), for five different energy ranges. The black line shows the appropriately normalized PDF of a standard normal distribution for comparison.[]{data-label="fig:residual_histogram"}](figures/r1_hist_residuals.pdf "fig:"){width="0.47\linewidth"} ![Histogram of residual significance for 1 (left panel) and 5 (right panel), for five different energy ranges. The black line shows the appropriately normalized PDF of a standard normal distribution for comparison.[]{data-label="fig:residual_histogram"}](figures/r5_hist_residuals.pdf "fig:"){width="0.47\linewidth"} Furthermore, in figures \[fig:residuals\_low\] and \[fig:residuals\_high\], we show in the top panels the significance of the residuals when subtracting the best-fit 1 model from [*Fermi*]{}-LAT data. As expected, we find that this simple scenario reproduces observations, with fractional residuals around $\sim30\%$ in many regions of the sky. In particular, large and very significant residuals remain along the bright Galactic plane, and at higher energies in the [*Fermi*]{}bubble regions and towards Cygnus X (at $\ell \sim 80^\circ$). A histogram of the residuals is shown in figure \[fig:residual\_histogram\], and will be discussed further below in context of 5. 2\. As a next step, we introduce also spatial modulation for all components, except the IGRB, with values for the regularization hyper-parameters listed in table \[tab:fits\]. We introduce $33\%$ ($\lambda=10$) uncertainties for the morphology of the gas component, and $100\%$ ($\lambda=1$) uncertainties for the morphology of the ICS component. We also now include smoothing regularization terms, with stronger smoothing for the ICS component than for the gas. The ICS component is smoothed with a hyper-parameter of $\eta=100$, which corresponds to roughly $10\%$ variations between neighboring pixels, while the gas component is smoothed with $\eta=25$ ($20\%$ variations). The total number of parameters is now 78548. As in 1, 20176 of these parameters correspond to point sources and their spectra. Most of the remaining parameters correspond to spatial modulation parameters for the ICS and gas components. Looking at figures \[fig:residuals\_low\] and \[fig:residuals\_high\], it is apparent that accounting for spatial and spectral nuisance parameters in the fit reduces the amount of residuals in the sky drastically. Typical fractional residuals are now of the order $\lesssim 10\%$ at energies above 1.24 GeV. Fractional residuals above the minimum energy of 0.34 GeV are practically zero (this can also be seen in figure \[fig:residuals\_low\]), because there are enough degrees of freedom in the fit to remove the most significant residuals. They happen to be at the lowest energies due to the larger number of low-energy photons. One effect of the nuisance parameters is that *reducible* residuals, in the sense that they can be absorbed by small changes of the local intensity of model components without significantly changing their spectra, are largely removed, whereas irreducible residuals remain. Structures that become now evident are the bubble-shaped positive residuals that appear from the second energy band on, and, at high energies, numerous localized residuals along the Galactic plane. 3\. This model is similar to the previous one, except that we split the gas component into three rings that cover the radial ranges 0–3.5 kpc, 3.5–6.5 kpc and 6.5–19.0 kpc. Details about the construction of gas rings can be found in section \[sec:gde\]. The main effect is that residuals along the Galactic disk are further reduced, as can be seen in the third row of figures \[fig:residuals\_low\] and \[fig:residuals\_high\]. However, localized residuals remain along the Galactic plane, both close to and at the Galactic center, as well as further away. Most of these residuals grow stronger at higher energies, indicating rather hard fluxes. Indeed, many of the residuals correspond to extended emission associated with the positions of various extended sources observed by [*Fermi*]{}-LAT. Below, we will be able to absorb most of these sources with extended source templates from the [*Fermi*]{}3FGL. ![Spatial modulation parameters for the four diffuse components (three gas and one ICS) of 3.[]{data-label="fig:run3_rescaling"}](figures/r3_map_rescaling_pi0ii.pdf "fig:"){width="0.49\linewidth"} ![Spatial modulation parameters for the four diffuse components (three gas and one ICS) of 3.[]{data-label="fig:run3_rescaling"}](figures/r3_map_rescaling_pi0i.pdf "fig:"){width="0.49\linewidth"} ![Spatial modulation parameters for the four diffuse components (three gas and one ICS) of 3.[]{data-label="fig:run3_rescaling"}](figures/r3_map_rescaling_pi0o.pdf "fig:"){width="0.49\linewidth"} ![Spatial modulation parameters for the four diffuse components (three gas and one ICS) of 3.[]{data-label="fig:run3_rescaling"}](figures/r3_map_rescaling_ICS.pdf "fig:"){width="0.49\linewidth"} In figure \[fig:run3\_rescaling\], we show the modulation parameters that correspond to the four diffuse components in 3. For the gas ring I and gas ring II, the modulation parameters stay close to one except for a few bright spots along the Galactic disk. Most of these correspond to extended sources in the [*Fermi*]{}3FGL. However, in gas ring III, which reaches up to high latitudes, we can see the effects of ‘dark gas’, which is not included in our gas templates and typically forms extended halos around CO and HI emission regions [@2005Sci...307.1292G]. This gas component is not captured by the usual CO line or 21 cm observations, but can be inferred from gas extinction maps to good accuracy. We did not include such a correction in our initial maps shown in figure \[fig:templates\]. Observing that the nuisance parameters are in fact recovering the dark gas from the gamma-ray data alone indicates that our method works as expected. In the ICS template, we also clearly see an additional component in the modulation parameters: the [*Fermi*]{}bubbles. The spectrum for the [*Fermi*]{}bubbles is somewhat similar to the ICS spectrum, but it is, of course, a distinct component that we will include as such in subsequent fits. Additional spectral and spatial components ------------------------------------------ 4\. Up to now, we considered only the most conventional ingredients in the modeling of the GDE, although augmented with nuisance parameters. We found in 3 that various significant residuals still remain. We will now add additional components that account for these residuals. We note that this does not mean that this is the only valid way to decompose the observed gamma-ray emission from the Galaxy in individual components. We first add a spectral template for the [*Fermi*]{}bubbles. The [*Fermi*]{}bubbles component has an unconstrained morphology, while the spectrum is fixed, to within $5\%$ ($\lambda=400$), to the [*Fermi*]{}bubble spectrum measured in ref. [@Fermi-LAT:2014sfa]. We furthermore use a weak spatial smoothing regularization with $\eta=4$. In this case, we use non-zero smoothing regularization terms to reduce the effect of Poisson noise on the reconstructed morphology, although this is also expected to somewhat wash out potentially sharp features in the bubble edges. We constrain the bubble template to be non-zero only in an hour-glass shaped region defined by $|\ell| + 4\cos(b/6.4^\circ)< 16^\circ$. Furthermore, we add templates for the many spots of localized extended high-energy emission that we saw in 3 along the Galactic disk. Many of these regions are colocated with the extended sources in the 3FGL. 18 such sources, out of 25 total in the 3FGL catalog, fall within our RoI. We therefore introduce new diffuse components that are only non-zero in regions that correspond to the extended sources in the 3FGL catalog that fall within our RoI. We initialize the spectra for these components to the extended source spectra from ref. [@Acero:2015hja], but leave considerable freedom for the spectra in the fit. Furthermore, the morphologies of the extended sources are left unconstrained, except for some smoothing. The templates of the extended sources are restricted to circular regions centered on their 3FGL locations with radii varying from 1.5–5$^{\circ}$, depending on the size of the residuals left by the source in 3. We use a total of 16 templates for the extended sources. We exclude two sources, $\gamma$-Cygni and HESS J1616-508, because their templates overlap with the templates of other extended sources, and they are therefore completely degenerate with them, given the freedom allowed in both the spectral and spatial modulation parameters. Details about the adopted hyper-parameters can be found in table \[tab:fits\]. We can see in figures \[fig:residuals\_low\] and \[fig:residuals\_high\] that the additional components remove much of the residual emission that remained in 3. Besides removing the [*Fermi*]{}bubbles at high latitudes, most of the bright high-energy emission along the Galactic disk is removed. However, not all of the residuals disappear. The most notable one that remains at high energies is localized at $\ell\approx -48^\circ$ is extended by a few degrees. This residual is most likely associated with an extended source or sources not present the 3FGL catalog. In fact, there are two extended sources within $|b|<0.5^{\circ}$ near $\ell=-48^{\circ}$ in the recently released 3FHL catalog [@2017arXiv170200664T], 3FHL J1420.3-6046e and 3FHL J1409.1-6121e (see also the FGES catalog, built from a dedicated search for extended emission above 10 GeV along the Galactic plane: [@2017arXiv170200476T]). Overall there are 19 extended emission sources in the 3FHL catalog that are within our RoI that are also not detected as extended emission in the 3FGL catalog. These sources may very well be contributing to some of the high-energy residuals we find along the disk in our fits. However, we restrict ourselves to the 3FGL catalog for now and save a full analysis of the extended emission sources along the disk for a future study. Another residual that remains is some extended emission close to the Galactic center. It has the peculiar feature that it is strongest in the second energy band, and appears almost as a negative residual in the third and fourth energy band. We show this feature in figure \[fig:run4\_residratio\]. The top two figures are the same as in figures \[fig:residuals\_low\] and \[fig:residuals\_high\], but more heavily smoothed with a gaussian kernel to reduce the visual effect of noise and enhance the larger residual present close to the Galactic center. In the bottom panel of figure \[fig:run4\_residratio\], we show the ratio of the higher-energy residual map divided by the lower-energy map, after the maps were smoothed, to highlight the strength of this residual (the sharp positive-negative edges in the ratio map are the result of the smoothing). We address this residual in 5. ![Top two figures: heavily smoothed residual maps for 4 for two energy bands, $>~1.24$ GeV and $>~16.84$ GeV. Bottom: ratio of the heavily smoothed residual maps, where the ratio is the $>~16.84$ GeV map divided by the $>~1.24$ GeV map.[]{data-label="fig:run4_residratio"}](figures/r4_map_residuals_1_5s.pdf "fig:"){width="0.49\linewidth"} ![Top two figures: heavily smoothed residual maps for 4 for two energy bands, $>~1.24$ GeV and $>~16.84$ GeV. Bottom: ratio of the heavily smoothed residual maps, where the ratio is the $>~16.84$ GeV map divided by the $>~1.24$ GeV map.[]{data-label="fig:run4_residratio"}](figures/r4_map_residuals_20_5s.pdf "fig:"){width="0.49\linewidth"} ![Top two figures: heavily smoothed residual maps for 4 for two energy bands, $>~1.24$ GeV and $>~16.84$ GeV. Bottom: ratio of the heavily smoothed residual maps, where the ratio is the $>~16.84$ GeV map divided by the $>~1.24$ GeV map.[]{data-label="fig:run4_residratio"}](figures/r4_map_residuals_ratio_3div1.pdf "fig:"){width="0.49\linewidth"} 5\. This is the last and most complex model in the present analysis. On top of the components of 4, it includes a template that accounts for the central bulge excess that we found above. It corresponds to the [*Fermi*]{}Galactic center excess (GCE) extensively studied in the literature [[*e.g.*]{}, @Goodenough:2009gk; @Hooper:2010mq; @Boyarsky:2010dr; @Abazajian:2012pn; @Macias:2013vya; @Daylan:2014rsa; @Gaggero:2015nsa; @Calore:2014xka; @deBoer:2016esu; @TheFermi-LAT:2015kwa; @Karwin:2016tsw]. As template for the GCE, instead of an analytic DM-inspired model as typically done in the literature, we use here, for the first time, the profile of the 511 keV flux as measured by INTEGRAL  [e.g.  @Leventhal_1978; @Knodlseder:2005yq; @Weidenspointner:2006nua; @Siegert:2015knp]. The positron annihilation signal (for a comprehensive review see ref. [@Prantzos:2010wi]) in the Galactic bulge shows a few common features to the GCE: both are almost spherically symmetric around the Galactic center, show a spectrum uniform in an extended RoI, and peaks in the direction of the center of the Galaxy. For a more thorough study about the connection between the two signals we refer the reader to a forthcoming publication [@511keV]. We here report the main results of introducing such an additional template component. We only include the Galactic center and bulge component of the 511 keV flux, and neglect the disk for the present purpose. These components are modeled following ref. [@Siegert:2015knp]. Furthermore, we keep the morphology of the component constrained to within $20\%$, while the spectrum remains completely unconstrained, see table \[tab:fits\]. The corresponding residuals in figures \[fig:residuals\_low\] and \[fig:residuals\_high\], show that this component is enough to remove the remaining positive and negative residuals in the central region. In figure \[fig:residual\_histogram\], we show the distribution of residuals in units of standard deviations for 1 and 5. Comparing 1 to 5, we find that the introduction of nuisance parameters in the fit reduces the number of pixels with highly significant residuals (above/below $\pm3\sigma$) by over two orders of magnitude. In fact, the most drastic reduction of residuals appears from 1 to 2, which can be seen in the drop of the Poisson $-2\ln\mathcal{L}_P$ as shown in table \[tab:fits\]. However, even in 5 there are residuals above $\pm5\sigma$ that remain. On closer inspection, we find that at energies $\lesssim10{\ensuremath{{\rm\,GeV}}}$, they are usually related to regions around bright point sources which are over- or under-subtracted. A likely cause is that we do not refit the positions of the 3FGL sources in our analysis; another may be due to potential, small numerical imperfections in our treatment of the [*Fermi*]{}PSF. At high energies, the remaining residuals often correspond to the regions that are also visible in the residual maps in figure \[fig:residuals\_high\]: extended regions along the Galactic disk with hard emission. The modeling of this emission component is still not optimal in the current analysis, and could be further improved by, [*e.g.*]{}, increasing the size of the patches that we use for extended sources in our 5 or including templates for additional known extended sources in the 3FHL. Lastly, the distribution of residuals in figure \[fig:residual\_histogram\] strongly deviates from a naive normal distribution, illustrated by the black line, which would be approximately expected if the residuals were due to statistical fluctuations only. Additionally, there is a strong tilt towards positive residuals, particularly for energies above 60 GeV, for both 1 and 5. This can be understood by realizing that the smallest possible negative residual in a pixel with an expected number of $\mu$ photons is, in our adopted schema, $-\sqrt{\mu}$. At high energies, $\mu$ can be of order one, leading to the apparent tilt. In fact, we observe the same behaviour for our mock results that are discussed in appendix \[apx:synthetic\]. Characteristics of the reference model {#sec:min} ====================================== ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_pi0ii.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_pi0i.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_pi0o.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_ICS.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_bubbles.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_511.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_extended.pdf "fig:"){width="0.49\linewidth"} ![All extended emission components for the reference model 5 (see table \[tab:fits\] for details about the model). In the bottom row, we show the combined emission from extended sources (left panel), although their templates and spectra are kept independent in the fit. We also show the combined emission of point sources (right panel), which also have independent normalizations and spectra in the fits.[]{data-label="fig:run5_components"}](figures/r5_map_mean_PSC.pdf "fig:"){width="0.49\linewidth"} In this section, we discuss in detail the aspects of 5 that we motivated in the previous section. In particular, we are interested in the plausibility of the values and variances of the modulation parameters in the analysis. Spatial modulation parameters ----------------------------- ![Modulation parameters for the conventional diffuse emission components of 5, which includes multiple disk components.[]{data-label="fig:run5_rescaling"}](figures/r5_map_rescaling_pi0ii.pdf "fig:"){width="0.49\linewidth"} ![Modulation parameters for the conventional diffuse emission components of 5, which includes multiple disk components.[]{data-label="fig:run5_rescaling"}](figures/r5_map_rescaling_pi0i.pdf "fig:"){width="0.49\linewidth"} ![Modulation parameters for the conventional diffuse emission components of 5, which includes multiple disk components.[]{data-label="fig:run5_rescaling"}](figures/r5_map_rescaling_pi0o.pdf "fig:"){width="0.49\linewidth"} ![Modulation parameters for the conventional diffuse emission components of 5, which includes multiple disk components.[]{data-label="fig:run5_rescaling"}](figures/r5_map_rescaling_ICS.pdf "fig:"){width="0.49\linewidth"} ![Scatter plot and histogram of rescaling factors of the spatial modulation parameters of the conventional diffuse emission components of 5 (see also figure \[fig:run5\_rescaling\]). The scatter plot shows the rescaling factor vs. the template flux in each pixel. The histogram in the lower panel is weighted by the pixel flux.[]{data-label="fig:run5_rescaling_hist"}](figures/r5_hist_rescaling_pi0ii.pdf "fig:"){width="0.49\linewidth"} ![Scatter plot and histogram of rescaling factors of the spatial modulation parameters of the conventional diffuse emission components of 5 (see also figure \[fig:run5\_rescaling\]). The scatter plot shows the rescaling factor vs. the template flux in each pixel. The histogram in the lower panel is weighted by the pixel flux.[]{data-label="fig:run5_rescaling_hist"}](figures/r5_hist_rescaling_pi0i.pdf "fig:"){width="0.49\linewidth"} ![Scatter plot and histogram of rescaling factors of the spatial modulation parameters of the conventional diffuse emission components of 5 (see also figure \[fig:run5\_rescaling\]). The scatter plot shows the rescaling factor vs. the template flux in each pixel. The histogram in the lower panel is weighted by the pixel flux.[]{data-label="fig:run5_rescaling_hist"}](figures/r5_hist_rescaling_pi0o.pdf "fig:"){width="0.49\linewidth"} ![Scatter plot and histogram of rescaling factors of the spatial modulation parameters of the conventional diffuse emission components of 5 (see also figure \[fig:run5\_rescaling\]). The scatter plot shows the rescaling factor vs. the template flux in each pixel. The histogram in the lower panel is weighted by the pixel flux.[]{data-label="fig:run5_rescaling_hist"}](figures/r5_hist_rescaling_ICS.pdf "fig:"){width="0.49\linewidth"} The various best-fit fluxes for the extended components of 5 are shown in figure \[fig:run5\_components\]. The spatial modulation parameters of 5 are shown in figure \[fig:run5\_rescaling\], for the four conventional diffuse emission components. Histograms for the same modulation parameters are shown in figure \[fig:run5\_rescaling\_hist\], weighted by pixel intensity. The original morphology of the four components is shown in figure \[fig:templates\], where the spatial modulation parameters are set to one. We find that for the innermost gas ring I, the nuisance parameters remain rather close to one. The weighted $90\%$ central quantile covers the range \[0.88, 1.04\], indicating typically less than $10\%$ deviations from the original template. Weighting these parameters is important, since a large number of pixels of the template correspond to very small fluxes, which are not constrained by the data and hence not rescaled. This does not, however, necessarily reflect that the original template was very accurate, but more likely that measuring this template is difficult in the bright Galactic disk. The gas rings II and III have modulation parameters with a broader distribution. Their weighted $90\%$ central quantiles cover \[0.80, 1.46\] and \[0.70, 1.45\], respectively. Note that there are also a few outliers, which likely correspond to isolated point sources that get incorrectly absorbed by the template. These do not affect the quantiles, which are relatively resilient towards outliers. The larger variations indicate that the templates are well constrained by the data, and give an estimate for the quality of gas maps in describing the data. The gas ring III, which includes the position of the Sun and is responsible for the structured emission at mid-latitudes, exhibits an interesting pattern in the nuisance parameters. The spatial modulation parameters are in some regions as large as 2–3 (in the tail of the distribution); in other regions we observe a suppression of similar magnitude. As for 3, the emerging structures are very similar to the ‘dark gas’ corrections [@2005Sci...307.1292G] (see discussion for 3). In figure \[fig:run5\_rescaling\], we also show the rescaling of the ICS emission w.r.t. the initial template from figure \[fig:templates\]. The weighted $90\%$ quantile covers here the values \[0.84, 1.21\], which is a smaller range than for the gas emission. However, there are some similarities between the large-scale features of the gas ring III and the ICS rescaling map. This suggests that both components are not completely separable, and probably some of the ICS emission is falsely absorbed by the gas map and vice versa. This is expected if, for instance, the spectra of the two components are not perfectly uniform throughout the entire RoI (this effect could be mitigated by further splitting up these diffuse components). On the other hand, looking at the total ICS emission shown in figure \[fig:run5\_components\] suggests that the emission absorbed by our ICS template might actually have a bi-modal ring-like structure, with enhancements towards $\ell\approx\pm25^\circ$. This corresponds to the edges of the molecular ring and could be potentially explained by the 3D distribution of CR sources, or yet by the asymmetry expected in the case of 3D interstellar radiation field. Spectral modulation parameters ------------------------------ ![Spectra from 5, in the same regions as used in figure \[fig:run1\_spectra\]. In the upper panels, we show the various components of the fit, as well as their sum compared to the data from that region. The lower panels show the residuals in units of standard deviations, where $\sigma = (\text{data}-\text{model})/\sqrt{\text{model}}$.[]{data-label="fig:run5_spectra"}](figures/r5_spectra_fullROI.pdf "fig:"){width="0.49\linewidth"} ![Spectra from 5, in the same regions as used in figure \[fig:run1\_spectra\]. In the upper panels, we show the various components of the fit, as well as their sum compared to the data from that region. The lower panels show the residuals in units of standard deviations, where $\sigma = (\text{data}-\text{model})/\sqrt{\text{model}}$.[]{data-label="fig:run5_spectra"}](figures/r5_spectra_CentralRegion.pdf "fig:"){width="0.49\linewidth"} The best-fit spectra from 5 are shown in figure \[fig:run5\_spectra\], both for the full RoI considered in this work as well as for a smaller central region that covers the Galactic bulge. In both regions, the measured spectrum is well reproduced. The error bars on the ICS, gas ring II, and gas ring III components are very small, which is a consequence of the large sky region that they cover. The same is true for the combined emission of the 776 point sources in our analysis, although some $20\%$ uncertainties becomes visible at high energies in the central region. Gas ring I, however, has sizable uncertainties, as it only covers a small sky region and overlaps along the line-of-sight with the other gas rings. We remark that our results are relatively stable w.r.t. choices of the adopted smoothing parameters, as discussed in appendix \[apx:run5\]. We find that the inner gas ring II has a somewhat harder spectrum at higher energies than the outer gas ring III (the effect is not visible is the poorly constrained ring I). This hardening can also be observed in figure \[fig:run5\_tree\], where we show the strength of the spectral modulation parameters for the diffuse emission components in 5. There is no strong drag for gas ring I and the ICS components, which indicates self-consistency in the results. However, the spectra of the gas rings II and III harden by more than 3$\sigma$. Although (part of) the indicated spectral change could be also caused by some unmodeled extended sources, this hardening of the spectrum towards the inner Galaxy could be a real feature of the interaction of CR protons with the gas, and point towards a concentration of CR accelerators or specific propagation effects in that region [@Gaggero:2014xla]. Although a more systematic analysis would be required in this respect, we find results qualitatively in agreement with Refs. [@Yang:2016jda; @Acero:2016qlg]. The Galactic Center Excess -------------------------- Lastly, the 511 keV template, which we use to model the [*Fermi*]{}Galactic center excess (GCE), exhibits a spectrum that peaks, as usual, at 1–3 GeV, and falls quickly off at lower and higher energies. The error bars are here at the level of $10-20\%$ at the peak. As can be seen in the right panel of figure \[fig:run5\_spectra\], at energies above 10 GeV, the template does not absorb any significant emission, which shows that the high-energy emission found in previous studies [@Calore:2014xka; @Linden:2016rcf] is here absorbed by other components, in particular the [*Fermi*]{}bubbles. Here, the [*Fermi*]{}bubbles absorb the high-energy emission as a consequence of not having imposed a uniform brightness bubble template. Very marginal negative residuals remain at energies between 10 and 50 GeV, most likely because of the spectral constraint on the [*Fermi*]{}bubbles, which could be reduced by relaxing the constraints on that component. The only change in our model between 4 and 5 is the addition of the GCE component. We can in theory determine its significance straightforwardly with the standard likelihood ratio test; the difference in the poisson likelihood between 4 and 5 is approximately $208$. However, the difference in the degrees of freedom between the two models cannot be read off from table \[tab:fits\], because the difference between 4 and 5 is smaller than the uncertainty in the number of effective degrees of freedom (see section \[sec:gof\]). We therefore performed a test where the difference in degrees of freedom is obvious: instead of allowing some small spatial modulation in the GCE component, we fix the spatial template completely, but allow the spectral modulation to be completely free. We kept all other parameters the same as in 5. For this run, the difference in the degrees of freedom between 4 and 5 is equal to $25$, or the number of spectral bins. The Poisson likelihood decreases such that the signifcance for the GCE is about $12\sigma$. The overall fit is for this run is quantitatively very similar to our original version of 5, as expected. We can therefore conclude that the GCE component is highly significant. ![Spectral modulation parameters, $\sigma_i^{(k)}$, for the four ‘conventional’ diffuse emission components in 5. The rescaling is w.r.t. the nominal input spectra, which are taken from ref. [@Ackermann:2012pya]. The different gray shaded regions show the 1, 2 and 3-$\sigma$ ranges corresponding to the MEM regularization (namely the values of $\sigma_i^{(k)}$ where $\lambda=16$). Lines correspond respectively, from bottom to top, to the energy bins from $0.3$ to $230$ GeV.[]{data-label="fig:run5_tree"}](figures/r5_tree.pdf){width="0.6\linewidth"} Discussion {#sec:discussion} ========== The goodness-of-fit {#sec:gof} ------------------- We are interested in estimating the goodness-of-fit for our final 5 model. The traditional Pearson’s chi-squared goodness-of-fit test cannot be applied to the current problem for two reasons. First, the data is Poisson and not normally distributed. Second, given that we introduced a number of penalization constraints on the model parameters, as well as smoothing constraints, it is not clear what the effective number of free model parameters really is. We use an approach that fully takes into account the Poisson nature of the problem as well as the unclear effective number of degrees of freedom (DOF), and which is based on the mock results obtained in appendix \[apx:run5\]. First, we consider here only the Poisson part of the likelihood function, and define $\chi^2\equiv-2\ln\mathcal{L}_P$. We assume that the $\chi^2$ function is chi-squared distributed under repeated data realizations (this is true to very good approximation, since for the very high number of DOF the central limit theorem dominates the shape of the distribution). We estimate the effective number of DOF from fits to mock data that are based on the best-fit models for each of the five runs in table \[tab:fits\]. It can be estimated to be $k \approx \langle -2\ln\mathcal{L}_P\rangle_{\rm mock}$; its associated uncertainty is simply the square root of the variance of the chi-square distribution, $\sqrt{2k}$, about $\sim1100$ for our five runs. The number of effective DOF along with the final Poisson likelihood and regularization terms, $-2\ln\mathcal{L}_P$ and $-2\ln\mathcal{L}_R$, can be found in table \[tab:fits\] for each fit. Naively, the number of model parameters is just given by the number of free parameters in the fit, as listed in table \[tab:fits\]. The number of independent data points would normally be simply equal to the number of pixels times the number of energy bins, $$N_\text{data} = N_\text{pix}\times N_\text{ebin}\;.$$ For the current analysis, we have $N_\text{data} = 360\times81\times25 = 729000$. However, these estimates are only valid if we are in the Gaussian regime in all data bins, and if all model parameters are free to vary and can independently and without degeneracies improve the fit to the data. However, these conditions are not satisfied in our current setup. We therefore estimate the *effective* number of data bins, model parameters, and DOF in a way that makes these estimates useful for goodness-of-fit calculations. ![Effective number of DOF per data bin, $d(\mu)$, as function of the Poisson expectation value $\mu$, see Eq. . In the Gaussian regime, $\mu\gg1$, the function converges to one; in the deep Poissonian regime, $\mu\ll1$, it is well approximated by $2\mu\ln(1/\mu)$ (red dashed lines).[]{data-label="fig:d"}](figures/dof.pdf){width="0.6\linewidth"} The effective number of data bins is calculated by averaging the $-2\ln\mathcal{L}_P$ over mock realizations of the best fit model, without refitting model parameters. This can be written as $$N_\text{data}^\text{eff} \equiv \left\langle -2\ln\mathcal{L}_P({\boldsymbol{\theta}})\right\rangle_{\mathcal{D}({\boldsymbol{\theta}})} = \sum_{i=1}^{N_\text{data}} 2\left\langle c \ln\frac{\mu_i({\boldsymbol{\theta}})}{c}\right\rangle_{c\sim P(\mu_i({\boldsymbol{\theta}}))} = \sum_{i=1}^{N_\text{data}} d(\mu_i({\boldsymbol{\theta}})) \;, \label{eqn:Neffdata}$$ where we explicitly show that the average can be written as a sum over all data bins. We furthermore defined a function $d(\mu)$, which can be interpreted as the effective number of “degrees of freedom” that correspond to a data bin with expectation value $\mu$, assuming Poisson noise. The sum of this function over all data bins gives the total number of *effective* data bins, $N_\text{data}^\text{eff}$. We show that function in figure \[fig:d\]. For $\mu\ll1$, one can show that it follows $2\mu\ln(1/\mu)$; for $\mu\gg1$ it converges to one. We use a tabulated version of the function $d(\mu)$ to calculate $N_\text{data}^\text{eff}$ for each of the five runs. The results are similar for all runs, with values around 620000, and shown in table \[tab:fits\]. The number of effective model parameters can now be estimated in a similar way. We define it as $$N_\text{param}^\text{eff} \equiv N_\text{data}^\text{eff} -\left\langle -2\ln\mathcal{L}_P({\boldsymbol{\hat\theta}})\right\rangle_{\mathcal{D}({\boldsymbol{\theta}})}\;,$$ where ${\boldsymbol{\hat\theta}}$ are the maximum-likelihood estimators for the model parameters, and different for each data set the average is taken over. Note that $ \left\langle -2\ln\mathcal{L}_P({\boldsymbol{\hat\theta}})\right\rangle_{\mathcal{D}({\boldsymbol{\theta}})}$ is simply equal to the DOF, $k$, which we define above. For each of the five runs, instead of performing the average, we in practice use only one mock realization to obtain $N_\text{param}^\text{eff}$ (the expected standard deviation is $\sim1100$ and not relevant for our discussion). The resulting values are always significantly smaller than the nominal number of free model parameters, $N_\text{param}$, and summarized in table \[tab:fits\]. We estimate the fit-quality of the five runs by using some ‘model fidelity’ parameter, $$\mathcal{F} = \sqrt{-2\ln\mathcal{L}_P({\boldsymbol{\hat\theta}})-k}\;.$$ It can be interpreted as the expected statistical significance (in standard deviations) of any additional *single-parameter* component that would bring the model at hand in perfect agreement with the data. Since the expected significance is lower for a multi-parameter component, our model fidelity parameter $\mathcal{F}$ provides an upper limit on the significance at which additional unmodeled components could be identified in the data by improvement of the model. We consider this quantity to be more useful than the standard $p$-values, which remain extremely small for all models considered in this work ($p \lll 10^{-315}$ for 1 and $p=10^{-77}$ for 5). The model fidelities $\mathcal{F}$ for 1–5 are summarized in table \[tab:fits\]. We find that the most drastic improvement is between 1 and 2, where the fit fidelity drops from 627 to 164. Subsequent refinements of the model appear to lead only to a relatively mild improvement of the fit quality. The still relatively large value for ${\textsc{Run5}}$, 144, is likely dominated by residuals around bright point sources and boundary effects related to our treatment of the PSF. We stress that the statistically significance for including additional model components with many parameters cannot be directly read off by comparing the improvement in $\mathcal{F}$ alone. Future directions and potential applications {#sec:potential} -------------------------------------------- In this paper we present a novel method for fitting the gamma-ray sky that can help us in deepening our understanding of gamma-ray emission mechanisms. We discuss here a number of possible or planned extensions of the current framework, and some of the science questions that it can be used for to address. On the *scientific* side, we are interested in full parameter scans over GALPROP/DRAGON predictions for cosmic-ray diffusion and gamma-ray emission. To this end, we would leave both the propagation model parameters as well as the relevant template modulation parameters free to vary. One could then study constraints on GALPROP/DRAGON parameters (like the halo height, the diffusion coefficient, source distributions, etc) while simultaneously accounting for, e.g., hard-to-model components like the [*Fermi*]{}bubbles. However, such a simultaneous fit is currently hindered by the long run-times of GALPROP/DRAGON when generating high-resolution gamma-ray predictions for large regions of the sky. A combined analysis could be used, e.g., to study potential cosmic-ray gradient, or substructure in the inverse Compton emission. This is for instance relevant to understand whether there is really a hardening of the proton spectrum suggested by the data, as previously claimed [@Yang:2016jda; @Acero:2016qlg]. One particularly interesting aspect is the potential observability of spiral arm structures in the gamma-ray emission of the Milky Way. It is well known that the distribution of CR sources as modeled in numerical codes might not be the most realistic one. For example, we do know that CR sources are expected to be distributed in spiral arms. The impact of spiral arms modeling has been discussed in the literature in the context of CR observables [@PhysRevLett.103.111302; @Gaggero:2013rya; @2015APh....64...18W]. Also, the spiral arm dynamic can have an impact of secondary-to-primary ratios [@Benyamin:2016xcq]. However, so far, the implications for gamma rays observables remain largely unexplored, with a pioneering work [@2017MNRAS.466.3674N] showing that the spiral arm dynamic might have a clear signature on gamma-ray pion spectrum. Fitting both propagation model parameters and nuisance parameters simultaneously would be likely important to make sure that a detection of spiral arm structures is not just an effect of mismodeling of gas. Finally, an obvious application is the characterization of [*Fermi*]{}bubbles at low latitudes and of possible degeneracies with the Galactic center excess. As we saw above, we are able to reconstruct the bubbles template given their spectral distribution, without including a specific spatial template. A more detailed analysis of the degeneracy between the spectral and spatial characteristics of the bubbles and the GCE would enable us to characterize very large scale emission down to the center of the Galaxy and might provide a way to disentangle various emission processes. On the *technical* side, a full Bayesian extension of the current analysis would be interesting and desirable. To this end, Hamiltonian Monte Carlos are particularly promising, since they make use of the available gradient information. In this framework, our sampling from the inverse Fisher information to estimate errors could be replaced by directly drawing from the proper posterior distributions of the model parameters. However, a sufficiently fine sampling of the posterior would require speeding up the code significantly and/or using variational inference for the posterior distributions. Another interesting technical extension would be the parallelization of the code, which is, as mentioned above, currently hindered by the difficulties of parallelizing sparse matrix multiplications. Conclusions {#sec:conclusions} =========== We presented a new hybrid approach for studying, modeling and decomposing diffuse gamma-ray emission, which we dubbed <span style="font-variant:small-caps;">SkyFACT</span> (**Sky** **F**actorization with **A**daptive **C**onstrained **T**emplates). Our approach combines methods of image reconstruction and adaptive spatio-spectral template regression in one coherent framework, based on penalized Poisson likelihood regression. We discussed in detail the implementation of our approach, as well as various solutions to technical challenges related to the high dimensionality of the optimization problem. In particular, we showed how the L-BFGS-B algorithm can be used to find unique solutions in high-dimensional parameter spaces. The largest example in this work has $1\times10^5$ parameters. We showed that the optimization problem is, for scenarios considered here (and for sufficiently soft smoothing), convex, and hence has no non-global minima. With conventional desktop computers, convergence can be reached in a few dozens of core hours. We demonstrated how Cholesky decomposition of the Fisher matrix, which we use here as approximation to the inverse covariance matrix, can be used to sample parameter errors. From these, uncertainties and correlations of various model components can be inferred. This is an important step towards going beyond the typically adopted strategy of ‘bracketing uncertainties’ with discrete model choices, although model assumptions still have to be made. We applied our approach to the gamma-ray emission from the inner Galactic disk, $|\ell|<90^\circ$ and $|b|<90^\circ$, using 7.6 years of [*Fermi*]{}-LAT data binned in $0.5^\circ$ pixels and covering the energy range 0.34–228.65 GeV. We presented a series of five increasingly complex models that show the challenges and typical residuals for conventional fits to the data, and the reduction of residuals when realistic nuisance parameters are included in the fit. The final, most complex model includes three gas rings, an ICS template, the IGRB, a spectral template for the [*Fermi*]{}bubbles, a spatial template for the [*Fermi*]{}Galactic center excess (based on INTEGRAL measurements of the 511 keV line), as well as [*Fermi*]{}-LAT extended and point sources. As a result, we can remove most of the residuals over much of the RoI and across all energies in the analysis. We also performed a series of tests with synthetic data to explore the performance of the regularization used in our fits; see appendix \[apx:synthetic\]. In general, we found that moderate regularization in the form of either modulation or smoothing can effectively recover model components without signficant overfitting. The main physics results of our analysis will be presented elsewhere. However, we summarize a few findings: - We do not include dark gas corrections in the gas maps, but instead recover the missing dark gas from the nuisance parameters in the fit. In general, the spatial nuisance parameters rescale the *a priori* gas and ICS templates that we use from DRAGON by less than $\pm40\%$, which is enough to provide good fits to the data, and within the uncertainties of these components. - We furthermore do not include a spatial template for the [*Fermi*]{}bubbles, but instead recover their flux at high and lower Galactic latitudes. We find that the bottom of the [*Fermi*]{}bubbles is not coincident with the Galactic center, but lies at around $\ell\approx-2^\circ$, which was also observed by [@TheFermi-LAT:2017vmf]. However, despite being bright, the emission is not very significant, and we cannot exclude that this is an extended unmodeled component of the foreground. - For the Galactic center excess we use a spatial template based on the morphology of the Galactic center and bulge emission of the 511 keV line measured by INTEGRAL. We find that the template is capable of removing the inner [*Fermi*]{}Galactic center excess and is detected at about the $12\sigma$ level, and that it picks up a spectrum compatible with the combined emission of (young or millisecond) pulsars. #### Acknowledgements. {#acknowledgements. .unnumbered} We are indebted to D. Gaggero for providing the package GammaSKY and helping with the GDE templates. F.C. thanks E. Charles for valuable discussion. C.W. and E.S. acknowledge funding from an NWO Vidi fellowship. Tests with synthetic data {#apx:synthetic} ========================= Image reconstruction, MEM and smoothing {#apx:circles} --------------------------------------- ![Counts map for the mock data set built from flat rings and an exponentially-varying background. The number of counts in each section, from left to right, total $10^5$, $10^6$, and $10^7$.[]{data-label="fig:ring_counts"}](figures/ring_logB_3exp_mockdata_map_counts.pdf){width="0.75\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_nosm_MEM_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_nosm_MEM_hist_residuals.pdf "fig:"){width="0.25\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_sm1p0_MEM_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_sm1p0_MEM_hist_residuals.pdf "fig:"){width="0.25\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_sm0p5_MEM_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_sm0p5_MEM_hist_residuals.pdf "fig:"){width="0.25\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_sm0p1_MEM_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and four levels of smoothing applied to the ring template (from top to bottom: $\eta = 0$, $1$, $4$, $100$; $\lambda$ is set to $0$ for these runs.) The left 4 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist"}](figures/ring_logB_3exp_sm0p1_MEM_hist_residuals.pdf "fig:"){width="0.25\linewidth"} ![Results of fits to rings with three different total counts in each map, and three values of the template modulation parameter for ring template (from top to bottom: $\lambda = 1$, $10^{-2}$, $10^{-4}$; $\eta$ is set to $0$ for these runs.) The left 3 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist_MEM"}](figures/ring_logB_3exp_MEM_t1_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and three values of the template modulation parameter for ring template (from top to bottom: $\lambda = 1$, $10^{-2}$, $10^{-4}$; $\eta$ is set to $0$ for these runs.) The left 3 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist_MEM"}](figures/ring_logB_3exp_MEM_t1_hist_residuals.pdf "fig:"){width="0.25\linewidth"} ![Results of fits to rings with three different total counts in each map, and three values of the template modulation parameter for ring template (from top to bottom: $\lambda = 1$, $10^{-2}$, $10^{-4}$; $\eta$ is set to $0$ for these runs.) The left 3 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist_MEM"}](figures/ring_logB_3exp_MEM_t10_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and three values of the template modulation parameter for ring template (from top to bottom: $\lambda = 1$, $10^{-2}$, $10^{-4}$; $\eta$ is set to $0$ for these runs.) The left 3 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist_MEM"}](figures/ring_logB_3exp_MEM_t10_hist_residuals.pdf "fig:"){width="0.25\linewidth"} ![Results of fits to rings with three different total counts in each map, and three values of the template modulation parameter for ring template (from top to bottom: $\lambda = 1$, $10^{-2}$, $10^{-4}$; $\eta$ is set to $0$ for these runs.) The left 3 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist_MEM"}](figures/ring_logB_3exp_MEM_t100_map_mean_ring.pdf "fig:"){width="0.74\linewidth"} ![Results of fits to rings with three different total counts in each map, and three values of the template modulation parameter for ring template (from top to bottom: $\lambda = 1$, $10^{-2}$, $10^{-4}$; $\eta$ is set to $0$ for these runs.) The left 3 panels are the recovered “flux” maps for the ring template, in arbitrary units. The right four panels are histograms of the residuals. Each histogram plot contains three histograms that represent the residuals for each of the three exposures. The histograms marked “low” correspond to the left-most panel with $10^5$ counts, “med” correspond to the middle panel with $10^6$ counts, and “high” correspond to the right-most panel with $10^7$ counts.[]{data-label="fig:ring_mean_hist_MEM"}](figures/ring_logB_3exp_MEM_t100_hist_residuals.pdf "fig:"){width="0.25\linewidth"} We first test both the effect of smoothing and MEM on image reconstruction for varying signal-to-background ratios and number of counts. We build a mock data set that consists of two spatial templates: a background that varies exponentially with latitude and is constant in longitude, and a set of three rings with flat, uniform intensity. No energy dependence is included in the data or fits. The ring and background templates are normalized such that the signal-to-background ratio is roughly equal to $10^{-2}$ at the top of the map ($b=20.25^{\circ}$), $10^{-1}$ at $b=10.125^{\circ}$, $1$ at $b=0^{\circ}$, $10$ at $b=-10.125^{\circ}$, and $10^{2}$ at the bottom ($b=-20.25^{\circ}$). Poisson noise is added and the data is divided into three equal-longitude sections of $60^{\circ}$. The exposure in each section varies such that the total number of counts is different in each section. There are approximately $10^5$ in the left section, $10^6$, in the middle, and $10^7$ in the right section. In the real [*Fermi*]{}-LAT data we use in our fits, there are roughly $5 \times 10^6$ photons in the lowest energy bin. At the top of the map, where the signal-to-background is small, the average number of background counts is $10^2$ in the left section, $10^3$ in the middle, and $10^4$ in the right section. Correspondingly, at the bottom where the signal-to-background is large, the average number of background counts is 0.01, 0.1, and 1 in the left, middle, and right sections. We show the counts map used for this analysis in figure \[fig:ring\_counts\]. To perform fits, we fix the background component completely. For the ring component, we initialize the template to be completely flat, and set the template modulation parameter to be essentially free ($\lambda = 0$). We perform four fits with increasing amounts of smoothing applied to the ring template ($\eta = 0$, $1$, $4$, and $100$). We show the results of these runs in figure \[fig:ring\_mean\_hist\]. The top-left image shows the recovered “flux” map where no smoothing was used in the template reconstruction. It is obvious even by eye that some amount of background is being recovered in the ring template; the histograms of residuals for this run also indicate this. In general, the rings are more accurately recovered in regions of high signal-to-background and large number of counts, as seen in the bottoms of right-most sections of the “flux” maps. For regions with fewer photons, and low signal-to-background, the rings are less well-recovered, seen in the tops of the middle and left sections of the flux maps. However, moderate amounts of smoothing can help to recover the rings and suppress too much fitting of the Poisson flucuations; compare, for example, the left sections of the middle two flux maps. If the smoothing parameter is too large, the ring shape is washed out, as seen in bottom flux map, especially for low photon counts. The strong positive and negative residuals seen in the histogram for this run with the strongest smoothing are mostly the result of the edge of ring getting smeared out: there are strong positive residuals on the inside edge of the ring and strong negative residuals on the outside edge. We perform a second set of runs without smoothing, but with varying template modulation parameters, to illustrate how MEM regularization alone influences image reconstruction. Again, we fix the background component and initialize the ring template to be flat. We vary the template modulation parameter in each fit ($\lambda = 1, 10^{-2}, 10^{-4}$). If $\lambda>1$, there is insufficient variation permitted in the template parameters to recover the ring flux. These results are shown in figure \[fig:ring\_mean\_hist\_MEM\]. In the top panel, where the template modulation is strongest, there are still significant remaining residuals. Essentially, this template modulation is almost too strong, especially in the low-photon section, to allow a flat template to completely reconstruct the ring flux. For weaker template modulation, there are very few positive residuals, indicating that the ring flux is fully recovered (this is evident from the residual maps, not shown, as well.) The negative residuals that remain are the result of the flat template not being able to go completely to zero; the rescaling for most of those pixels is at their smallest allowed value ($10^{-10}$; this is also true for the runs with smoothing). However, also evident is some amount of Poisson noise being picked up by this template, seen especially in the high-background-count upper regions of the flux maps. As seen in the previous set of runs, even a weak smoothing constraint applied here would drastically reduce the amount of fitted Poisson noise, especially in low-photon regions. Component separation and Poisson-induced bias with two components {#apx:two} ----------------------------------------------------------------- --------------------------- ------ ------------- ----------- Components free constrained smoothed Gas 000 000 0[100]{}0 \[0.1cm\] Inverse Compton 000 000 0[100]{}0 --------------------------- ------ ------------- ----------- : Hyper-parameters for mock 2-component runs.[]{data-label="tab:pi0_ICS_mock"} ![Top: counts map, summed over all energies. Bottom: spectral model components.[]{data-label="fig:2comp_initial"}](figures/pi0_ICS_mock_MEM_initial_map_data_counts.pdf "fig:"){width="0.6\linewidth"} ![Top: counts map, summed over all energies. Bottom: spectral model components.[]{data-label="fig:2comp_initial"}](figures/pi0_ICS_mock_MEM_initial_spectra_fullROI-crop.pdf "fig:"){width="0.4\linewidth"} For the next set of mock data runs, we wish to test how fits perform at separating model components (both templates and spectra) with different values for smoothing and template modulation parameters. For these tests, we build a mock data set with two components: the gas and ICS components described in section \[sec:gde\]. Here we also include energy dependencies for both components, using the spectra described in section \[sec:gde\] for the gas and ICS components. The energy range we fit over is the same as for the real LAT data: $0.34-228.65$ GeV. We choose these two components because they are roughly similar in overall brightness in our RoI, have similar spectral shapes at high energies, and share some morphological characteristics. The diffuse emission from the gas and ICS also dominate the real LAT data we consider. We show the mock data counts map and the initial spectral energy distribution in figure \[fig:2comp\_initial\]. ![Results for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with moderate template and spectra regularization terms and no smoothing ($\lambda,\lambda' = 1$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t1_s1"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_spectra_fullROI.pdf){width="0.8\linewidth"} ![Results for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with moderate template and spectra regularization terms and no smoothing ($\lambda,\lambda' = 1$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t1_s1"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_map_mean_pi0.pdf "fig:"){width="0.99\linewidth"} ![Results for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with moderate template and spectra regularization terms and no smoothing ($\lambda,\lambda' = 1$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t1_s1"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_map_mean_ICS.pdf "fig:"){width="0.99\linewidth"} ![Residuals for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 1$, $\eta = 0$). Residual maps and histograms are in units of signficance.[]{data-label="fig:2comp_sm0p0_t1_s1_res"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_hist_residuals_masked.pdf){width="0.8\linewidth"} ![Residuals for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 1$, $\eta = 0$). Residual maps and histograms are in units of signficance.[]{data-label="fig:2comp_sm0p0_t1_s1_res"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_map_residuals_0s.pdf "fig:"){width="0.99\linewidth"} ![Residuals for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 1$, $\eta = 0$). Residual maps and histograms are in units of signficance.[]{data-label="fig:2comp_sm0p0_t1_s1_res"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_map_residuals_5s.pdf "fig:"){width="0.99\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t1_s1_rsc"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_tree.pdf "fig:"){width="0.32\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t1_s1_rsc"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_hist_rescaling_pi0.pdf "fig:"){width="0.32\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “free” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t1_s1_rsc"}](figures/pi0_ICS_mock_MEM_sm0p0_t1_s1_Ns30_fullROI_hist_rescaling_ICS.pdf "fig:"){width="0.32\linewidth"} ![Results for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 4$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t0p5_s0p5"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_spectra_fullROI.pdf){width="0.8\linewidth"} ![Results for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 4$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t0p5_s0p5"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_map_mean_pi0.pdf "fig:"){width="0.99\linewidth"} ![Results for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 4$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t0p5_s0p5"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_map_mean_ICS.pdf "fig:"){width="0.99\linewidth"} ![Residuals for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 4$, $\eta = 0$). Residual maps and histograms are in units of signficance.[]{data-label="fig:2comp_sm0p0_t0p5_s0p5_res"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_hist_residuals_masked.pdf){width="0.8\linewidth"} ![Residuals for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 4$, $\eta = 0$). Residual maps and histograms are in units of signficance.[]{data-label="fig:2comp_sm0p0_t0p5_s0p5_res"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_map_residuals_0s.pdf "fig:"){width="0.99\linewidth"} ![Residuals for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with stronger template and spectra regularization, and no smoothing ($\lambda,\lambda' = 4$, $\eta = 0$). Residual maps and histograms are in units of signficance.[]{data-label="fig:2comp_sm0p0_t0p5_s0p5_res"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_map_residuals_5s.pdf "fig:"){width="0.99\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 4$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t0p5_s0p5_rsc"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_tree.pdf "fig:"){width="0.32\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 4$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t0p5_s0p5_rsc"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_hist_rescaling_pi0.pdf "fig:"){width="0.32\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “constrained” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 4$, $\eta = 0$).[]{data-label="fig:2comp_sm0p0_t0p5_s0p5_rsc"}](figures/pi0_ICS_mock_MEM_sm0p0_t0p5_s0p5_Ns30_fullROI_hist_rescaling_ICS.pdf "fig:"){width="0.32\linewidth"} ![Results for mock 2-component (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) run with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$).[]{data-label="fig:2comp_sm0p1_t1_s1"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_spectra_fullROI.pdf){width="0.8\linewidth"} ![Results for mock 2-component (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) run with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$).[]{data-label="fig:2comp_sm0p1_t1_s1"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_map_mean_pi0.pdf "fig:"){width="0.99\linewidth"} ![Results for mock 2-component (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) run with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$).[]{data-label="fig:2comp_sm0p1_t1_s1"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_map_mean_ICS.pdf "fig:"){width="0.99\linewidth"} ![Residuals for mock 2-component run (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$). Residual maps and histograms are in units of significance.[]{data-label="fig:2comp_sm0p1_t1_s1_res"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_hist_residuals_masked.pdf){width="0.8\linewidth"} ![Residuals for mock 2-component run (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$). Residual maps and histograms are in units of significance.[]{data-label="fig:2comp_sm0p1_t1_s1_res"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_map_residuals_0s.pdf "fig:"){width="0.99\linewidth"} ![Residuals for mock 2-component run (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$). Residual maps and histograms are in units of significance.[]{data-label="fig:2comp_sm0p1_t1_s1_res"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_map_residuals_5s.pdf "fig:"){width="0.99\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$).[]{data-label="fig:2comp_sm0p1_t1_s1_rsc"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_tree.pdf "fig:"){width="0.32\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$).[]{data-label="fig:2comp_sm0p1_t1_s1_rsc"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_hist_rescaling_pi0.pdf "fig:"){width="0.32\linewidth"} ![Rescaling parameters for spectra (left panel), and templates (center and right panels) for mock 2-component run (labeled “smoothed” in table \[tab:pi0\_ICS\_mock\]) with essentially free templates and spectra, and strong smoothing ($\lambda,\lambda' = 1$, $\eta = 100$).[]{data-label="fig:2comp_sm0p1_t1_s1_rsc"}](figures/pi0_ICS_mock_MEM_sm0p1_t1_s1_Ns30_fullROI_hist_rescaling_ICS.pdf "fig:"){width="0.32\linewidth"} For all of the following fits, we allow the overall normalization of each of the templates to freely vary ($\lambda'' = 0$). The templates and spectra are initialized to their original components described in section \[sec:gde\]. All runs were performed with MEM regularization. We show the hyper-parameter values for each fit in table \[tab:pi0\_ICS\_mock\]. We first show the results of a run with moderate template and spectra modulation parameters, and no smoothing ($\lambda,\lambda' = 1$, $\eta,\eta' = 0$). As seen in figure \[fig:2comp\_sm0p0\_t1\_s1\], the fit fails to properly recover the two components, instead separating into a “low-energy” component and a “high-energy” component. If the modulation is set to be weaker, the same behavior is seen (the error bars on the spectra become larger). While a histogram of residuals and map of residuals summed over all energies indicates no substantial deviations, there is clear structure remaining in at least the higher energy residual map, all shown in figure \[fig:2comp\_sm0p0\_t1\_s1\_res\]. Additionally, as shown in figure \[fig:2comp\_sm0p0\_t1\_s1\_rsc\], there are large deviations from the original spectral and spatial components. While this is formally not a bad fit (its likelihood is similar to those of the following runs), it is obvious from the figures that it is an unphysical result. Next, we show the results of a run with stronger template and spectra modulation parameters, still with no smoothing ($\lambda,\lambda' = 4$, $\eta,\eta' = 0$); these regularization values are near those considered in fits to real data. In this case, the original model components are recovered well, as shown in figure \[fig:2comp\_sm0p0\_t0p5\_s0p5\]. We show the residuals for this fit in figure \[fig:2comp\_sm0p0\_t0p5\_s0p5\_res\]. From the rescaling parameters shown in figure \[fig:2comp\_sm0p0\_t0p5\_s0p5\_rsc\], we see very little deviation from the original spectra, and small deviations in the template rescaling Strong smoothing can act to effectively constrain the allowed variation in the templates, even if the modulation parameters are relatively weak. In figure \[fig:2comp\_sm0p1\_t1\_s1\], we show the results from a fit with the same modulation parameters as in figure \[fig:2comp\_sm0p0\_t1\_s1\], but with strong smoothing applied to both templates (but still no smoothing on the spectra; $\lambda,\lambda' = 1$, $\eta=0.1$, $\eta'=0$). There are no strong residuals obvious in the maps or histograms in figure \[fig:2comp\_sm0p1\_t1\_s1\_res\]. As shown in figure \[fig:2comp\_sm0p1\_t1\_s1\_rsc\], there are essentially no deviations from the original spectra or spatial templates. Tests with 5 mock data {#apx:run5} ---------------------- For our final set of tests, we build a mock data set that is composed of the best-fit model components to run5, described in section \[sec:min\]. We show the counts map and spectra used for this set of tests in figure \[fig:run5\_mock\_initial\]. ![Top: counts map, summed over all energies. Bottom: spectral data components.[]{data-label="fig:run5_mock_initial"}](figures/rp5_initial_map_data_counts.pdf "fig:"){width="0.6\linewidth"} ![Top: counts map, summed over all energies. Bottom: spectral data components.[]{data-label="fig:run5_mock_initial"}](figures/rp5_initial_spectra_fullROI-crop.pdf "fig:"){width="0.4\linewidth"} ![Comparison of spectra for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_spectra"}](figures/rp5_1_spectra_fullROI.pdf "fig:"){width="0.4\linewidth"} ![Comparison of spectra for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_spectra"}](figures/r5_mock_nosm_spectra_fullROI.pdf "fig:"){width="0.4\linewidth"} ![Comparison of spectra for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_spectra"}](figures/r5_mock_tempdbl_spectra_fullROI.pdf "fig:"){width="0.4\linewidth"} ![Comparison of spectra for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_spectra"}](figures/r5_mock_temphlf_spectra_fullROI.pdf "fig:"){width="0.4\linewidth"} ![Comparison of histograms for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_hist"}](figures/rp5_1_hist_residuals.pdf "fig:"){width="0.4\linewidth"} ![Comparison of histograms for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_hist"}](figures/r5_mock_nosm_hist_residuals.pdf "fig:"){width="0.4\linewidth"} ![Comparison of histograms for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_hist"}](figures/r5_mock_tempdbl_hist_residuals.pdf "fig:"){width="0.4\linewidth"} ![Comparison of histograms for the various tests performed with the run5 mock data set. Top-left: post-fit, original template modulation and smoothing parameters. Top-right: no smoothing. Bottom-left: weaker template modulation parameters. Bottom-right: stronger template modulation parameters.[]{data-label="fig:run5_mock_hist"}](figures/r5_mock_temphlf_hist_residuals.pdf "fig:"){width="0.4\linewidth"} We first fit the mock data set, with the same model components as in 5, but starting from the best-fit parameters for 5 (see section  \[sec:min\]). As shown in in the top left panel of figure \[fig:run5\_mock\_spectra\], the spectral components are almost perfectly recovered, to within $\sim30\%$ on average (except for a few obvious energy bins in the 511 keV spectrum, which is not smoothed). We show histograms of the residuals for this mock data set after fitting in the top left panel of figure \[fig:run5\_mock\_hist\]; this can be compared to the original fit for 5 in figure \[fig:residual\_histogram\]. The Poisson tail is clearly seen in the highest energy bin. To test the robustness of our results for 5 on real LAT data, we vary the template modulation and smoothing parameters for this mock data set. In our first test, we set the smoothing on all components to zero. The results of these tests are shown in the top-right panels of figures \[fig:run5\_mock\_spectra\] and \[fig:run5\_mock\_hist\]. The most obvious feature in the spectra plot is the deviation in the ICS spectrum at one high energy bin. The mockdata set is Poisson-dominated at this energy, and there are few photons actually being fit. Additionally, the spatial and spectral modulation parameters are somewhat weaker for the ICS component than the other components, especially the gas, which, in the absence of of smoothing, means that the ICS component is permitted more variation. This deviation is clearly unphysical, and demonstrates the need for moderate regularization in low photon regimes, as was also seen in the previous two sections. Interestingly, the overall deviation from the data is also quite high, at the $\sim2\sigma$ level for three energy bins, indicating that these energies are not well fit. The overall likelihood is not significantly different, however, from any of the other mock data runs in this section. We also perform fits where the spatial modulation parameters for the ICS and three gas components are set to be weaker and stronger by a factor of 4 relative to those in table \[tab:fits\]. The other modulation parameters and all smoothing parameters remain unchanged from the original 5 set-up. The results of these tests are shown in the bottom panels of figures \[fig:run5\_mock\_spectra\] and \[fig:run5\_mock\_hist\]. Again, in the case of the weaker modulation (bottom left), the ICS component is considerably suppressed while the outer gas ring III is enhanced relative to the original best-fit parameters; the 511 keV template is also somewhat enhanced. This illustrates how weakly constrained the ICS component is relative to the other components. The residual histograms in figure \[fig:run5\_mock\_hist\] all show reasonable overall fits. [^1]: https://heasarc.nasa.gov/docs/xanadu/xspec/ [^2]: http://fermi.gsfc.nasa.gov/ssc/data/analysis/ [^3]: <http://galprop.stanford.edu> [^4]: We use the run\_2D\_KRA.xml model file available at <https://github.com/cosmicrays/DRAGON/examples>.
{ "pile_set_name": "ArXiv" }
--- author: - | Aleksey M. Tikhonov$^a$[^1], Viktor E. Asadchikov$^b$, Yurii O. Volkov$^b$, Boris S. Roshchin$^b$, Yurii A. Ermakov$^c$\ $^a$Kapitza Institute for Physical Problems, Russian Academy of Sciences, Moscow, Russia\ $^b$Shubnikov Institute of Crystallography, Federal Research Center Crystallography and Photonics,\ Russian Academy of Sciences, Moscow, Russia\ $^c$Frumkin Institute of Physical Chemistry and Electrochemistry, Russian Academy of Sciences, Moscow, Russia title: '**Study of DMPS Monolayers on a Water Substrate with Laboratory X-ray Reflectometer**' --- Phospholipids on a water surface form an insoluble monomolecular layer, a film that is a two-dimensional thermodynamic system with parameters $(\Pi, T)$. Under certain conditions, the structure of the layer is described by a symmetry axis perpendicular to the water–air interface \[1, 2\]. In particular, the Langmuir monolayer formed by dimyristoyl phosphatidylserine (DMPS) molecules is such a system (Fig. 1). Under normal conditions, the phase transition from a two-dimensional liquid to a gel structure (liquid crystal) occurs in this film as the surface pressure $\Pi$ rises at constant temperature $T$ \[3, 4\]. However, simulations of such systems even using molecular dynamics methods leave the question about the molecular nature of electrostatic effects in monolayers open. Considerable help in solving this question may be expected from the use of direct methods of recording the structural changes in a monolayer. In particular, X-ray scattering was successfully used previously in \[2, 5\] to study the behavior of zwitterionic lipid monolayers. In this paper we propose to use such a technique to study the monolayer structures of anionic DMPS lipids with a pronounced phase transition. Indeed, under DMPS monolayer compression there are two distinctly different regions of change in electric potential (potential drop in the lipid monolayer): a comparatively small and smooth change in potential in the liquid state of the monolayer gives way to its sharp increase ($\sim 200$mV) when the lipid passes into a solid gel phase. Various hypotheses \[4\], for example, a change in the hydration state of the polar phospholipid groups \[5, 6\], are proposed to explain this fact. We think that X-ray reflectometry data can be useful for testing these hypotheses. In this paper, based on our measurements of the X-ray reflectivity with a photon energy of $\sim 8$keV, we have reconstructed the electron density profile across the surface of a DMPS monolayer in its different phase states. Two approaches to analyzing the experimental data were used to extract the structural information: with a priori information (“model” approach) and without any assumptions about the transverse surface structure (“modelless” approach). The samples of DMPS phospholipid monolayers were prepared and studied in an airtight cell with X-ray-transparent windows in accordance with the technique described in \[7, 8\]. Some volume of a phospholipid solution with a concentration of 0.5mg/ml in a 5 : 1 chloroform–methanol mixture was spread with a syringe over the surface of a liquid substrate (a KCl solution in deionized water with a concentration of 10mmol/L and pH=7) placed in a fluoroplastic dish with a diameter $D = 100$mm. The solution volume required for our experiments was calculated for three values of the finite area $A$ per molecule chosen for different lipid phase states in the monolayer. The dependence of the surface pressure, $\Pi(A) = \gamma_0 - \gamma(A)$, measured on DMPS layers in \[3, 4\] is uniquely specified by $A$ (see Fig. 2). Here, the surface tension of pure water under normal conditions is $\gamma_0 = 72.5$mN/m, while $\gamma(A)$ is the surface tension in the presence of a lipid film. According to this dependence, at $A \approx 100$Å$^2$ the monolayer is in an “expanded liquid” state I, while at $A \approx 50$Åthe monolayer is a two-dimensional inhomogeneous structure II and contains an equilibrium mixture of liquid and solid phase domains. Finally, at $A\approx 34$Å$^2$ the DMPS monolayer is in a condensed state III that is usually defined as a liquid crystal \[5, 9\]. The transverse structure of the lipid monolayer was investigated by X-ray reflectometry on a versatile laboratory diffractometer with a movable emitter–detector system \[10\]. An X-ray tube with a copper anode is used as the emitter. The K.1 line (photon energy $E =8048$eV, wavelength $\lambda = 1.5405 \pm 0.0001$Å) is chosen from the tube emission spectrum using a single-crystal monochromator Si (111). The vertical and horizontal beam sizes are $\sim 0.1$ and $\sim 8$mm, respectively. The three-slit collimation system forms a probing X-ray beam with an angular width of $\sim 10^{–4}$rad in the plane of incidence. The angular resolution of the point detector $2\Delta\beta = 1.7 \times 10^{–3}$rad is determined by the entrance slit with a gap of 1mm at a distance of $\sim 570$mm from the sample center. Vacuum paths with X-ray-transparent windows are used to reduce the absorption and scattering of emission in air. Let ${\bf k}_{in}$ and ${\bf k}_{sc}$ be the wave vectors with an amplitude $k_0 = 2\pi/\lambda$ for the incident and scattered beams, respectively. It is convenient to introduce a coordinate system in which the origin $O$ lies at the center of the illumination region, the $xy$ plane coincides with the water boundary, the $x$ axis is perpendicular to the beam direction, and the $z$ axis is directed along the normal to the surface oppositely to the force of gravity (see the inset in Fig. 3). The scattering vector ${\bf q} = {\bf k}_{sc}-{\bf k}_{in}$ upon mirror reflection has only one nonzero component $q_z = 2 k_0 \sin\alpha$, where $\alpha$ is the grazing angle in a plane normal to the surface. The angle of total external reflection for a water surface $\alpha_c$ ($q_c = 2k_0\sin\alpha_c$) is fixed by the electron density in water $\rho_w=0.333$ [*e-/*]{}[Å]{}$^3$, $\alpha_c\approx\lambda\sqrt{r_e\rho_w/\pi}$$\approx0.15^\circ$, where $r_e = 2.814 \times 10^{-5}$Åis the classical electron radius. The diffractometer software allows a variable angular step, a detector slit width, and an exposure time to be specified, which makes it possible to optimize the measurement of the reflectivity $R$ rapidly decreasing with increasing $\alpha$. For a beam incident on the sample at an angle $\alpha$ the linear size of the illumination region along the beam is approximately $\propto 1/\sin\alpha$. As a consequence, at $\alpha \approx \alpha_c$ the beam section in the lateral $y$ direction (parallel to the sample surface) turns out to be appreciably larger than the sample diameter $D$, which leads to an incorrect determination of $R$. The correcting factor corresponding to the ratio of the total intensity of the direct beam to the intensity of its fraction falling within the sample surface is calculated before each measurement. The calculation of such a factor is similar to that in \[11\]. Figure 3 shows the dependence $R(q_z)$ measured in three independent experiments for a DMPS monolayer on a water surface at various areas per molecule A near the phase transition: for $A \approx 100$ (state I), 50 (state II), and 34 (state III) Å$^2$. At $q_z < q_c \approx 0.022$Å$^{–1}$ the incident beam undergoes total external reflection, $R \approx 1$. Thus, the data for the reflectivity $R(q_z)$ collected on the diffractometer are comparable in spatial resolution $2\pi/q_z^{max} \approx 10$Å ($q_z^{max} \approx 0.7$Å$^{-1}$ is the maximum value of $q_z$ in our experiment) to the data obtained previously for various planar systems using synchrotron radiation \[12-18\]. The information about the surface structure in our experiment is averaged over a large illumination area $A_0\approx 0.5$cm$^2$ and, therefore, the structure of the near-surface layer of our samples may be considered in the approximation of an ideal layer-inhomogeneous structure \[19\]. The electron density profile across the surface, $\rho(z)$, was reconstructed from the reflectometry data by invoking two different approaches. The first approach is based on qualitative models with a minimum number of adjustable parameters using a priori information about the molecular structure of a lipid film \[17, 18\]. For simplicity, below this approach is called the “model” one. In the distorted wave Born approximation (DWBA) the reflectivity for a flat surface is \[20\] $$R(q_z)=\left|\frac{\displaystyle q_z-q_z^t}{\displaystyle q_z+q_z^t}\right|^2\left|\Phi(\sqrt{q_zq_z^t})\right|^2,$$ where $q_z^t=\sqrt{q_z^2 - q_c^2}$. Thus, interpreting the reflectometry data is reduced to finding some complex function of the structure factor $\Phi(q)$ that generally has the following form $$\Phi(q)=\frac{1}{\rho_w}\int^{+\infty}_{-\infty}\left\langle\frac{d\rho(z)}{dz}\right\rangle e^{iqz} dz,$$ where the electron density gradient is averaged over the area $A_0$. In the case under consideration, for example, to achieve good agreement of the model curves with the experimental data, it will suffice to divide the near-surface structure into two layers. In accordance with the structure of the DMPS molecule, the first layer of thickness $L_1$ and electron density $\rho_1$ is formed by the polar phosphatidylserine groups, while the second layer of thickness $L_2$ and electron density $\rho_2$ is formed by the aliphatic tails -C$_{14}$H$_{27}$. Next, we construct the model profile $\rho(z)$ for the monolayer based on the error function ${\rm erf}(x)$ by assuming that all boundaries between the layers and bulk phases have the same width $\sigma_0$ \[21\]: $$\begin{array}{c} \displaystyle \rho=\frac{1}{2}\rho_{0}+\frac{1}{2}\sum_{j=0}^2(\rho_{j+1}-\rho_j) {\rm erf}\left(\frac{l_j}{\sigma_0\sqrt{2}}\right), \\ \\ \displaystyle l_j=z+\sum_{n=0}^{j}L_n, \\ \\ \displaystyle {\rm erf}(x)=\frac{2}{\sqrt{\pi}}\int_0^x\exp(-y^2)dy, \end{array}$$ where $\rho_0 \equiv \rho_w$ is the electron density in water, $L_0 \equiv 0$ is the position of the water - polar group layer interface $(z=0)$, and $\rho_3\approx 0$ is the bulk electron density in air. Thus, we have \[22\] $$\begin{array}{c} \displaystyle R(q_z)=\frac{\exp(-\sigma^2_0q_zq_z^t)}{\rho^2_w}\left|\frac{\displaystyle q_z-q_z^t}{\displaystyle q_z+q_z^t}\right|^2 \\ \\ \displaystyle \times \left|\sum_{j=0}^{j=2}{(\rho_{j+1}-\rho_{j})\exp\left(-i\sqrt{q_zq_z^t}\sum_{n=0}^{j}L_n \right)}\right|^2. \end{array}$$ In our calculations the parameter $\sigma_0$ was fixed to be equal to the “capillary width” $$\sigma_{0}^2 = \frac{k_BT}{2\pi\gamma(A)}\ln\left(\frac{Q_{max}}{Q_{min}}\right)$$ (where $k_B$ is the Boltzmann constant). The latter is specified by the short-wavelength limit in the spectrum of capillary waves $Q_{max} = 2\pi/a$ ($a\approx 10$Å is the intermolecular distance in order of magnitude) and the angular resolution of the detector $Q_{min}=q_z^{max}\Delta\beta$. This method of allowance for the contribution of capillary waves to the observed structure on a liquid surface was proposed in \[23-26\] and turned out to be useful in interpreting numerous experiments \[12, 14, 24-28\]. Note that if $\sigma_0$ is used as an adjustable parameter, then its values for the systems being investigated lie in the range $2.8-3.1$Å and coincide, within the error limits, with the calculated values of $\sigma_0 = 2.7-3.0$Å from the dependence $\gamma(A)$, which also defines the compression diagram $\Pi(A)$. The second approach is based on an extrapolation of the asymptotic behavior of the reflectivity curve $R(q_z)$ to large $q_z$ without using any a priori assumptions about the transverse surface structure \[29, 30\]. This approach can be arbitrarily called the “modelless” one. In this approach the polarizability distribution in depth $\delta(z)$ is assumed to contain singular points $z_j$ at which the polarizability (or its $n$th derivative) changes abruptly: $$\Delta^{n}(z_j) \equiv \frac{d^n\delta\left(z_j + 0\right)}{dz^n} - \frac{d^n\delta\left(z_j - 0\right)}{dz^n}.$$ A combination of such singular points uniquely determines the asymptotic behavior of the amplitude reflectivity $r(q_z)$ when $q_z \to \infty$ ($R(q_z) \equiv |r(q_z)|^2$). The arrangement of points $z_j$ can be determined from the experimental curve $R(q_z)$ measured in a limited range of values for $q_z$ using the procedure of a modified Fourier transform described in detail in \[29\]. Generally, there exist only two physically reasonable distributions $\delta(z)$ that simultaneously satisfy the experimental values of the reflectivity $R(q_z)$ and the specified combination of singular points $\Delta^n(z_j)$ in the polarizability profile and that differ only by the order of their arrangement relative to the substrate. For each of the measured curves we found a pair of points with mutually opposite signs of the jumps in the first derivative: the first corresponds to the air–sample interface, while the second corresponds presumably to the electron density maximum near the glycerin base of the polar group. The distance between them was 16.4, 23.5, and 25.4Å for the films with $A$ equal to 100, 50, and 34Å$^2$, respectively. The sought-for profile $\delta(z)$ was represented as a step function with fixed positions of the singular points $\Delta^1(z_j)$ and was divided into a large number $M$ $(M \approx 100)$ of thin layers: $\delta(z)= \sum_{m=1}^{M}{\Delta(z_m)H(z-z_m)}$, where $H(z)$ is the Heaviside step function \[31\]. The reflectivity curve $R(q_z, \delta(z1), …,\delta(zM))$ for such a structure was calculated in accordance with Parratt’s recurrence relations \[32\]. Thus, the polarizability profile was found by numerically optimizing the residual between the experimental reflectivity curve and the calculated one regularized by the smoothness condition for the sought-for profile and by the positions of the singular points using the standard Levenberg–Marquardt algorithm \[16, 33\]. Finally, for weakly absorbing materials in the hard X-ray spectral range the electron density profile $\rho(z) \simeq \pi\delta(z)/(r_0\lambda^2)$ can be calculated from the reconstructed polarizability distribution in depth $\delta(z)$ \[34\]. Analysis of the data obtained confirms that the DMPS molecules are arranged on the water surface in the form of a monolayer. In Fig. 4 the solid lines indicate the profiles $\rho(z)$ for the two-layer model (3), while the dashed lines indicate the profiles reconstructed within the modelless approach. The dependences $R(q_z)$ corresponding to these curves are represented in Fig. 3 by the solid and dashed lines. In Fig. 4 the difference between the two approaches becomes noticeable at large glancing angles, at which the experimental error increases significantly. To a first approximation, the measured and calculated curves presented in Figs. 3 and 4 show satisfactory agreement between the two approaches in describing the experimental results. The model profile is characterized by four adjustable parameters (see the Table 1). As the surface pressure rises, the electron density in the layer of polar phosphatidylserine heads directly in contact with the water increases from $\rho_1\approx1.2\rho_w$ in state I to $\rho_1\approx1.4\rho_w$ in state III. At the same time, its thickness $L_1$ is virtually constant and lies in the range 10 - 12Å for all states. In contrast, the thickness of the layer formed by the hydrocarbon chains increases noticeably from $L_2\approx 10$Å (state I) to $L_2\approx 15$Å (state III). Concurrently, the electron density also increases from $\rho_2\approx0.9\rho_w$ in the liquid phase to $\rho_2\approx0.95\rho_w$ in the solid phase. The total thickness $L$ of the monolayer is $L = (20 \pm 3)$Å in state I and $L = (28 \pm 2)$Å in state III. In intermediate state II the monolayer thickness is $L = (25 \pm 3)$Å. For the modelless description of the structures it will suffice to use only the first-order singular points, because all experimental curves decrease approximately as $R(q_z)\sim q_z^{-6}$. The deviation from a strict power law is apparently attributable to the scattering by sample surface roughnesses with an effective height $\sigma$. Its value can be estimated within the Nevot-Croce formalism from the requirements imposed on the asymptotics $$R(q_z)q_z^6\exp\left(\sigma^2q_z\sqrt{q_z^2-4k_0^2\delta_+}\right)\to {\rm const}$$ when $q_z \to \infty$, where $\delta_+ \approx 7.5\times 10^{–6}$ is the water polarizability for $\lambda \approx 1.54$Å \[19, 35\]. Thus, we obtain $\sigma \approx 3.2$Å, which agrees well with the calculated value of $\sigma_0 \approx 3$Å given above. Note that the estimation of the integral roughness parameters from the reflectometry curves alone is highly ambiguous \[36, 37\]. For a more proper analysis of the statistical roughness properties of the sample being investigated, it is necessary to additionally invoke the angular distributions of diffuse scattering, for example, within the procedure described in \[30\]. Next, given $\rho(z)$, the specific surface density $\Psi$ of structural units (ions, molecules, chemical groups) in a layer of thickness $d = z^{\prime\prime} - z^{\prime}$ can be estimated: $$\Psi=\frac{1}{\Gamma}\int^{z^{\prime\prime}}_{z^{\prime}}\rho(z)dz,$$ where $\Gamma$ is the number of electrons in one structural unit. For example, $\Gamma = 390$ for potassium salt C$_{34}$H$_{65}$NO$_{10}$4PK, $\Gamma_t/2 = 111$ for one -C$_{14}$H$_{27}$ chain, and $\Gamma_h = 168$ for the phosphatidylserine group. For state I the thickness $L \approx 20$Å and the distance between the singular points in the profile $\rho(z)$ of the modelless approach, approximately 16Å, are appreciably smaller than the length of the lipid molecule, approximately 27Å. This suggest that the hydrocarbon chains of molecules in the liquid phase of the lipid ($A\approx100$Å$^2$) are disordered relative to the normal to the surface. For state III ($A\approx34$Å$^2$), the thickness of the second layer $L_2\approx15$Å roughly corresponds to the calculated length of the hydrocarbon tails -C$_{14}$H$_{27}$, in the DMPS molecule, 16.7[Å]{} ($\approx 12\times 1.27$[Å]{}(C-C) + 1.5[Å]{}(-CH$_3$)). The density $\rho_2 \approx 0.95\rho_w$ and the area per hydrocarbon chain $A/2 \approx 17$Å$^2$ correspond to one of the crystalline phases of long-chain saturated hydrocarbons \[38\]. Thus, the deflection angle of the molecular tail axis from the normal to the surface can be estimated:$\theta = \arccos(15/16.7)\approx 26^\circ$ ($26^\circ \pm 8^\circ$). Note that the following imbalance in the number of electrons per structural unit is observed for all states. For example, in state III the number of electrons accounted for by the polar part of the DMPS molecule and the aliphatic tails is $\rho_1AL_1 \approx 206$[*e$^-$*]{} and $\rho_2AL_2 \approx 161$[*e$^-$*]{}, respectively. The excess electron density in the layer of heads is $A(\rho_1L_1 - \rho_2L_2\Gamma_h/\Gamma_t) \approx 84$[*e$^-$*]{} per DMPS molecule, which is equivalent to approximately eight H$_2$O molecules. Such a degree of hydration was reported previously for the gel phase of phospholipids in \[5\]. If the electron density is taken as a rough estimate of the degree of hydration, then it rises more than twofold as the area per molecule increases to $A\approx100$Å$^2$ and is $\sim 20$ H$_2$O molecules per polar group. According to the X-ray reflectometry data (Fig. 3), as the surface pressure rises, the phase transition from an expanded two-dimensional liquid to a solid state becomes noticeable in the DMPS monolayer. The main, and quite unexpected, result of our analysis of the experimental data is that the chosen two-layer model of the structure (model approach) describes the electron density profile predicted within the modelless approach in a good approximation. This fact is illustrated by the parameters of the gel phase established within these approaches and presented in Fig. 4 and the table. Both methods of analyzing the experimental curves give a pretty authentic idea of the behavior of the lipid monolayer as the lateral pressure changes. The model approach allows its important structural components to be identified, while the modelless approach allows one to independently confirm the correctness of the electron density distribution found and, thus, to reduce the ambiguity in interpreting the structural model. Thus, we investigated the molecular structure of a dimyristoyl phosphatidylserine (DMPS) monolayer on a water substrate in different phase states based on our X-ray reflectometry data. According to our analysis of the reflectivity curves, as the surface pressure rises, the transition from a two-dimensional expanded liquid state to a solid gel state accompanied by the ordering of the hydrocarbon tails -C$_{14}$H$_{27}$ occurs in the monolayer, while the thickness of the polar DMPS region remains virtually constant. The monolayer thickness is $(20 \pm 3)$ and $(28 \pm 2)$Å in the liquid and solid phases, respectively. In the gel phase the deflection angle of the tail axis from the normal to the surface is $26^\circ \pm 8^\circ$. At least a twofold decrease in the degree of hydration of the polar lipid groups occurs under two-dimensional monolayer compression. It is important to note that the decrease in the number of water molecules associated with the polar heads of lipids per se cannot lead to the positive change in electric potential observed in our experiments. Judging by the molecular dynamics data, the water and adsorbed cations are responsible for the positive changes in this potential \[39\]. Most likely, not the change in the number of water molecules and the degree of hydration but the orientation of their dipole moments and the adsorption of cations should be taken into account to explain the electrostatic effects in the monolayer. Detailed information about the molecules structures that are involved in such effects can be established by molecular dynamics methods in combination with comprehensive experimental studies, including the measurements of the reflectivity curves in a wide range of areas per molecule in the monolayer. We proved that a quantitative analysis of the X-ray reflectometry data is fundamentally possible using several examples given above. [49]{} L.D.Landau and E.M.Lifshitz, [*Course of Theoretical Physics*]{}, Vol. 5: [*Statistical Physics*]{} (Nauka, Moscow, 1995; Pergamon, Oxford, 1980). V.M.Kaganer, H.Mohwald, and P.Dutta, Rev. Mod. Phys. 71, 779 (1999). Yu.A.Ermakov, Biochemistry (Moscow) Suppl. Ser. A: Membr. Cell Biol. 5, 379 (2011). Y.A.Ermakov, K.Kamaraju, K.Sengupta, S.Sukharev, Biophys J. 98, 1018 (2010). H.Mohwald, Annu. Rev. Phys. Chem. 41, 441 (1990). H.Binder, Applied Spectroscopy Reviews 38, 15 (2003). A. M. Tikhonov, JETP Lett. 92, 356 (2010). V.E.Asadchikov, V.V.Volkov, Yu.O.Volkov, K.A.Dembo, I.V.Kozhevnikov, B.S.Roshchin, D.A.Frolov, and A.M.Tikhonov, JETP Lett. 94, 585 (2011). H.M.McConnell, Annu. Rev. Phys. Chem. 42, 171 (1991). V.E.Asadchikov, V.G.Babak, A.V.Buzmakov, Yu.P.Dorokhin, I.P.Glagolev, Yu.V.Zanevskii, V.N.Zryuev, Yu.S.Krivonosov, V.F.Mamich, L.A.Moseiko, N.I.Moseiko, B.V.Mchedlishvili, S.V.Savel’ev, R.A.Senin, L.P.Smykov, et al., Instrum. Exp. Tech. 48, 364 (2005). A.Gibaud, G.Vignaud, S.K.Sinha, Acta Cryst. A49, 642 (1993). M.L.Schlossman, M.Li, D.M.Mitrinovic, A.M.Tikhonov, High Performance Polymers 12, 551 (2000). L.Hanley, Y.Choi, E.R.Fuoco, F.A.Akin, M.B.J.Wijesundara, M.Li, A.M.Tikhonov, M.L.Schlossman, Nucl. Instrum. Methods Phys. Res. B 203, 116 (2003). A.M.Tikhonov, J. Phys. Chem. C 111, 930 (2007). A.M.Tikhonov, J. Chem. Phys. 130, 024512 (2009). A.M.Tikhonov, V.E.Asadchikov, and Yu.O.Volkov, JETP Lett. 102, 478 (2015). A.M.Tikhonov, J. Phys. Chem. B 110, 2746 (2006). A.M.Tikhonov, M.L.Schlossman, J. Phys.: Condens. Matter 19, 375101 (2007). A.M.Tikhonov, V.E.Asadchikov, Yu.O.Volkov, B.S.Roshchin, I.S.Monakhov, and I.S.Smirnov, JETP Lett. 104, 873 (2016). S.K.Sinha, E.B.Sirota, S.Garoff, and H.B.Stanley, Phys. Rev. B 38, 2297 (1988). F. P. Buff, R. A. Lovett, F. H. Stillinger, Phys. Rev. Lett. 15, 621 (1965). J.Daillant, L.Bosio, B.Harzallah, and J.J.Benattar, J. Phys. II 1, 149 (1991). J. D. Weeks, J. Chem. Phys. 67, 3106 (1977). A.Braslau, M.Deutsch , P.S.Pershan, A.H.Weiss, J.Als-Nielsen, J.Bohr, Phys. Rev. Lett. 54, 114 (1985). A.Braslau, P.S.Pershan, G.Swislow, B.M.Ocko, and J.Als-Nielsen, Phys. Rev. A 38, 2457 (1988). D.K.Schwartz, M.L.Schlossman, E.H.Kawamoto, G.J.Kellogg, P.S.Pershan, B.M.Ocko, Phys. Rev. A 41, 5687 (1990). D.M.Mitrinovic, A.M.Tikhonov, M.Li, Z.Huang, and M.L.Schlossman, Phys. Rev. Lett. 85, 582 (2000). A.M.Tikhonov, J. Chem. Phys 124, 164704 (2006). I.V.Kozhevnikov, Nuclear Instruments and Methods in Physics Research A 508, 519 (2003). I.V.Kozhevnikov and L.Peverini and E.Ziegler, Phys. Rev. B 85, 125439 (2012). R.Kanwal, [*Generalized Functions: Theory and Technique*]{}, 2nd ed., Birkhauser, 1998. L.G.Parratt, Phys. Rev. 95, 359 (1954). J.Nocedal, S.Wright, [*Numerical Optimization*]{}, 2nd ed., Springer, 2006. B.L.Henke, E.M.Gullikson, J.C.Davis, Atomic Data and Nuclear Data Tables 54, 181 (1993). L.Nevot, P.Croce, Rev. Phys. Appl. 15, 761 (1980). O.Gilev, V.Asadchikov, A.Duparr, N.Havronin, I.Kozhevnikov, Yu.Krivonosov, S.Kuznetsov, V.Mikerov, V.Ostashev, V.Tukarev, Proc. SPIE 2000, 4099:279-289. V.E.Asadchikov, I.V.Kozhevnikov, Yu.S.Krivonosov, R.Mercier, T.H.Metzger, C.Morawe, E.Ziegler, Nucl. Instrum. Meth. Phys. Res. A 530, 575 (2004). D.M. Small, “The Physical Chemistry of Lipids”, NY, Plenum Press, 1986. A.M.Nesterenko and Yu.A.Ermakov, Biochemistry (Moscow) Suppl. Ser. A: Membr. Cell Biol. 6, 320 (2012). [^1]: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'For two arithmetical functions $f$ and $g$, we study the convolution sum of the form $\sum_{n \le N} f(n) g(n+h)$ in the context of its asymptotic formula with explicit error terms. Here we introduce the concept of finite Ramanujan expansion of an arithmetical function and extend our earlier works in this setup.' address: - | Giovanni Coppola\ Universitá degli Studi di Napoli, Complesso di Monte S. Angelo-Via Cinthia 80126 Napoli (NA), Italy - | M. Ram Murty\ Department of Mathematics, Queen’s University, Kingston, Ontario, K7L 3N6, Canada - | Biswajyoti Saha\ School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Navy Nagar, Mumbai, 400 005, India author: - 'Giovanni Coppola, M. Ram Murty and Biswajyoti Saha' title: Finite Ramanujan expansions and shifted convolution sums of arithmetical functions --- [^1] Introduction ============ In 1918, Ramanujan studied [@SR] the following sum of roots of unity. For positive integers $r,n$, $$c_r(n):=\sum_{a\in ({{\mathbb Z}}/r{{\mathbb Z}})^*}\zeta_r^{an},$$ where $\zeta_r$ denotes a primitive $r$-th root of unity. These sums are now known as Ramanujan sums. It is also possible to write $c_r(n)$ in terms of the Möbius function $\mu$ (see [@RM]). One has $$\label{crn-exp1} c_r(n)=\sum_{d|n, d|r} \mu(r/d) d$$ for any positive integers $r,n$. Ramanujan studied these sums in the context of point-wise convergent series expansions of the form $\sum_r a_rc_r(n)$ for various arithmetical functions. Such expansions are now known as Ramanujan expansions. More precisely: We say an arithmetical function $f$ admits a Ramanujan expansion (in the sense of Ramanujan) if for each $n$, $f(n)$ can be written as a convergent series of the form $$f(n)= \sum_{r \ge 1} \hat f(r)c_r(n)$$ for appropriate complex numbers $\hat f(r)$. The number $\hat f(r)$ is said to be the $r$-th Ramanujan coefficient of $f$ with respect to this expansion. Shifted convolution sums are ubiquitous in number theory and recently such sums have been studied for functions with absolutely convergent Ramanujan expansions. It has been done systematically in [@GMP; @MS; @CMS; @BS2]. For two arithmetical functions $f$ and $g$ we study the convolution sum of the form $\sum_{n \le N} f(n) g(n+h)$. In this article we introduce the concept of finite Ramanujan expansion of an arithmetical function. This idea particularly enables us to avoid technical infinite sums and obtain an asymptotic formula with explicit error terms for the convolution sum $C_{f,g}(h):=\sum_{n \le N} f(n) g(n+h)$ for some fixed positive integer $N$ and non-negative integer $h$. Let us write our functions $f$ and $g$ as $$f(n)=\sum_{d|n}f'(d) \ \ \text{and} \ \ g(n)=\sum_{d|n}g'(d),$$ where $f' := f \ast \mu$ and $g' := g \ast \mu$. Here $\mu$ denotes the Möbius function and $\ast$ denotes the Dirichlet convolution. Then $$C_{f,g}(h)=\sum_{n\le N} \sum_{d|n}f'(d) \sum_{q|n+h}g'(q) =\sum_{d\le N}f'(d)\sum_{q\le N+h}g'(q) \sum_{{n\le N}\atop {{n\equiv 0\bmod d}\atop {n+h\equiv 0\bmod q}}}1.$$ Hence from the point of view of studying the convolution sums, we may put $$\label{f} f(n)=\sum_{d|n, d\le N}f'(d)$$ and $$\label{g} g(n+h)=\sum_{d|n+h, d\le N+h}g'(d),$$ i.e. we are enforcing that $f'(n)$ vanishes if $n>N$ and $g'(n)$ vanishes if $n>N+h$. This will definitely change the values our $f(n),g(n)$ that we started with, but only for $n>N$ and $n>N+h$ respectively. Hence this will not alter our convolution sum $\sum_{n \le N} f(n) g(n+h)$. At this point we note the following interesting property satisfied by Ramanujan sums. \[l1\] $$\frac{1}{d} \sum_{r|d} c_r(n)= \begin{cases} 1 & \text{if } \ d|n,\\ 0 & \text{otherwise}. \end{cases}$$ For a proof see Section \[lemmas\]. Now, using [Lemma \[l1\]]{}, we get $$f(n) = \sum_{d|n, d\le N}f'(d) = \sum_{d\le N} f'(d) \frac{1}{d} \sum_{r|d} c_r(n) = \sum_{r \le N} c_r(n) \left( \sum_{r|d, d\le N} \frac{f'(d) }{d} \right) = \sum_{r \le N} \hat f(r) c_r(n),$$ where $$\label{hat f} \hat f(r) := \sum_{r|d, d\le N} \frac{f'(d) }{d}.$$ Similarly $$g(n+h) = \sum_{s \le N+h} \hat g(s) c_s(n+h),$$ where $$\label{hat g} \hat g(s) := \sum_{s|d, d\le N+h} \frac{g'(d) }{d}.$$ Thus we obtain a finite series expansion for our functions $f$ and $g$ which is more like a Ramanujan expansion. This we refer to as finite Ramanujan expansion relative to $N$ and $h$. Note that such kind of an expansion depends on the fixed parameters $N$ and $h$. This helps us to avoid dealing with infinite sums, which was not the case in [@MS; @CMS; @BS2]. From now on, all these above notations will be used freely without referring to them. Using the dual Möbius inversion formula (see page 4 of [@CM]) it is also possible to express $f'$ in terms of $\hat f$. We have $$\label{f'} f'(r) = r \sum_{r|d, d\le N} \mu(d/r) \hat f(d)$$ and $$\label{g'} g'(s) = s \sum_{s|d, d\le N+h} \mu(d/s) \hat g(d).$$ For arithmetical functions with usual Ramanujan expansion that are absolutely convergent, the following theorem was proved in [@CMS]. \[CMS\] Suppose that $f$ and $g$ are two arithmetical functions with absolutely convergent Ramanujan expansions (in the sense of Ramanujan): $$f (n) = \sum_{r \ge 1} \hat f(r) c_r(n), \phantom{mm} g(n) = \sum_{s \ge 1} \hat g(s)c_s(n)$$ respectively. Further suppose that $$\big|\hat f(r)\big|,\big|\hat g(r)\big| \ll \frac{1}{r^{1 +\delta}}$$ for some $\delta > 0$ and $h$ is a non-negative integer. Then we have $$\sum_{n \le N} f(n) g(n+h) = \begin{cases} N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O(N^{1-\delta}(\log N)^{4-2\delta}) & ~ \mbox{ if } ~ \delta < 1,\\ N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O(\log^3 N) & ~ \mbox{ if } ~ \delta = 1,\\ N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O(1) & ~ \mbox{ if } ~ \delta > 1. \end{cases}$$ Now suppose we impose the conditions $$\label{hypo-1} \big|\hat f(r)\big|,\big|\hat g(r)\big| \ll \frac{1}{r^{1 +\delta}} \ \ \text{for some} \ \ \delta>0$$ on the coefficients of the finite Ramanujan expansions of the arithmetic functions $f$ and $g$, defined in and . Using the dual Möbius inversion formula, these conditions can be rewritten equivalently as $$\label{hypo-2} \big|f'(r)\big|,\big|g'(r)\big| \ll \frac{1}{r^{\delta}} \ \ \text{for some} \ \ \delta>0.$$ With these conditions in place we can derive a theorem that is analogous to [Theorem \[CMS\]]{} and we are also able to improve the error term in the case of $\delta \le 1$ by certain exponents of $\log N$. \[finite\] Let $N$ be a positive integer and $f$ and $g$ be two arithmetical functions for which we want to estimate the shifted convolution sums $$\sum_{n \le N} f(n) g(n+h)$$ for a positive integer $h$. Further suppose that $$\big|\hat f(r)\big|,\big|\hat g(r)\big| \ll \frac{1}{r^{1 +\delta}} \ \ \text{for some} \ \ \delta > 0,$$ where $\hat f(r),\hat g(r)$ are as in and . Then we have $$\sum_{n \le N} f(n) g(n+h) = N \sum_{r=1}^{\infty}\hat{f}(r)\hat{g}(r)c_r(h) + O_{\delta,h}\Big(N^{1-\delta} \log^2 N+1\Big).$$ The study of shifted convolution sums in the context of arithmetical functions with absolutely convergent Ramanujan expansions was initiated by Gadiyar, Murty and Padma in [@GMP]. The authors in [@GP] showed that if we ignore convergence questions, a Ramanujan expansion of the function $\frac{\phi(n)}{n}\Lambda(n)$, which is due to Hardy, can be used to derive the Hardy-Littlewood conjecture about prime tuples. Later in [@GMP] it was investigated whether the work in [@GP] can be justified for arithmetical functions with absolutely convergent Ramanujan expansion, under certain hypothesis on the Ramanujan coefficients. The objective of [@BS2] was to reach the minimality of such hypothesis. In that quest the last author proved the following theorem in [@BS2]. \[BS\] Let $f,g$ be two arithmetical functions with absolutely convergent Ramanujan expansions (in the sense of Ramanujan) $$f (n) = \sum_{r \ge 1} \hat f(r) c_r(n), \phantom{mm} g(n) = \sum_{s \ge 1} \hat g(s)c_s(n).$$ Further suppose that there exists $\alpha > 4$ such that $$\big|\hat f(r)\big|, \big|\hat{g}(r) \big| \ll \frac{1}{r \log^\alpha r}$$ and $h$ is a positive integer. Then for a positive integer $N$, we have $$\sum_{n \le N} f(n) g(n+h) = N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O \left( \frac{N}{(\log N)^{\alpha-4}} \right).$$ Now for the Ramanujan coefficients coming from finite Ramanujan expansions we first observe the following. \[l2\] Suppose that the Ramanujan coefficients $\hat f(r)$ coming from the finite Ramanujan expansion of $f$ satisfy $$\big|\hat f(r)\big| \ll \frac{1}{r \log^{\alpha}r} \ \ \text{for some} \ \ \alpha>1.$$ Then we have $$\big|f'(r)\big| \ll \frac{1}{\log^{\alpha-1}r}.$$ Similarly if we assume $$\big|f'(r)\big| \ll \frac{1}{\log^{\beta}r} \ \ \text{for some} \ \ \beta>1,$$ then we have $$\big|\hat f(r)\big| \ll \frac{1}{r \log^{\beta-1}r}.$$ The proof uses the conversion formulas and and partial summation formula. For a detailed proof see Section \[lemmas\]. Next we prove the following theorem. \[finite3\] Let $N$ be a positive integer and $f$ and $g$ be two arithmetical functions for which we want to estimate the shifted convolution sums $$\sum_{n \le N} f(n) g(n+h)$$ for a positive integer $h$. Further suppose that $$\big|f'(d)\big|, \big|g'(d)\big| \ll \frac{1}{\log^{\beta}d} \ \ \text{for some} \ \ \beta>2,$$ where $f'(d),g'(d)$ are as in and . Then we have $$\sum_{n \le N} f(n) g(n+h) = N \sum_{r=1}^{\infty}\hat{f}(r)\hat{g}(r)c_r(h) + O_{\beta,h}\Big(\frac{N}{\log^{\beta-2} N} \Big).$$ As an immediate corollary of [Theorem \[finite3\]]{}, we can now derive (using [Lemma \[l2\]]{}) the following theorem, which is an analogoue of [Theorem \[BS\]]{}, in the setting of finite Ramanujan expansions. \[finite2\] Let $N$ be a positive integer and $f$ and $g$ be two arithmetical functions for which we want to estimate the shifted convolution sums $$\sum_{n \le N} f(n) g(n+h)$$ for a positive integer $h$. Further suppose that $$\big|\hat f(r)\big|,\big|\hat g(r)\big| \ll \frac{1}{r \log^{\alpha}r} \ \ \text{for some} \ \ \alpha>3,$$ where $\hat f(r),\hat g(r)$ are as in and . Then we have $$\sum_{n \le N} f(n) g(n+h) = N \sum_{r=1}^{\infty}\hat{f}(r)\hat{g}(r)c_r(h) + O_{\alpha,h}\Big(\frac{N}{\log^{\alpha-3} N} \Big).$$ Proofs of the lemmas {#lemmas} ==================== The lemma follows from the known identity $$\frac{1}{d} \sum_{a=1}^d \zeta_d^{an}= \begin{cases} 1 & \text{if } \ d|n,\\ 0 & \text{otherwise}, \end{cases}$$ where $\zeta_d$ denotes a primitive $d$-th root of unity. We then partition the sum in the left hand side in terms of $\gcd$ and using the definition we write $$\frac{1}{d} \sum_{a=1}^d \zeta_d^{an} = \frac{1}{d} \sum_{r|d} \sum_{a=1 \atop (a,d)=r}^d \zeta_d^{an} = \frac{1}{d} \sum_{r|d} c_{d/r} (n) = \frac{1}{d} \sum_{r|d} c_r (n).$$ We use the formula and write $$\big|f'(d)\big|=\left| d\sum_{j\le {N\over d}}\mu(j)\hat{f}(jd) \right| \ll \sum_{j\le {N\over d}}{1\over {j\log^{\alpha}(jd)}}.$$ Without loss of generality we take $d>1$ and break the above sum in two different cases: one for $d\le \sqrt{N}$ and the other for $d>\sqrt{N}$. When $d\le \sqrt{N}$, we have $d\le {N\over d}$. Thus we have $$\sum_{j\le N/d}{1\over {j\log^{\alpha}(jd)}} \ll \sum_{j\le d}{1\over {j\log^{\alpha} d}} +\sum_{d<j\le {N\over d}}{1\over {j\log^{\alpha} j}} \ll_{\alpha} {1\over {\log^{\alpha-1} d}} + {1\over {\log^{\alpha-1}(N/d)}} \ll_{\alpha} {1\over {\log^{\alpha-1} d}}.$$ Next, if $d>\sqrt{N}$ then ${N\over d}<d$. Thus $$\sum_{j\le N/d}{1\over {j\log^{\alpha}(jd)}} \le \sum_{j\le d}{1\over {j\log^{\alpha} d}} \ll {1\over {\log^{\alpha-1} d}}.$$ This completes the proof of the first part. For the second part we use the formula and write $$\big| \hat f(r) \big| = \left| \sum_{r|d, d\le N} \frac{f'(d) }{d} \right| \ll {1\over r}\sum_{n\le N/r}{1\over {n\log^{\beta}(rn)}}.$$ Again we take $r>1$ and break the above sum in two different cases: one for $r \le \sqrt{N}$ and the other for $r>\sqrt{N}$. If $ r\le \sqrt{N}$, then $r \le {N\over r}$ and hence $$\sum_{n\le N/r}{1\over {n\log^{\beta}(rn)}} \ll \sum_{n\le r}{1\over {n\log^{\beta} r}} + \sum_{r<n\le {N\over r}}{1\over {n\log^{\beta} n}} \ll_{\beta} {1\over {\log^{\beta-1} r}} + {1\over {\log^{\beta-1}(N/r)}} \ll_{\beta} {1\over {\log^{\beta-1} r}}.$$ Further if $r>\sqrt{N}$, then ${N\over r}<r$ and thus $$\sum_{n\le N/r}{1\over {n\log^{\beta}(rn)}} \ll \sum_{n\le r}{1\over {n\log^{\beta} r}} \ll {1\over {\log^{\beta-1} r}}.$$ This completes the proof. Proofs of the theorems ====================== We start with $$\sum_{n \le N} f(n) g(n+h)=\sum_{n\le N} \sum_{d|n}f'(d) \sum_{q|n+h}g'(q) =\sum_{d\le N}f'(d)\sum_{q\le N+h}g'(q) \sum_{{n\le N}\atop {{n\equiv 0\bmod d}\atop {n+h\equiv 0\bmod q}}}1.$$ Now gathering by gcd and changing variables $d,q$ we get $$\sum_{n \le N} f(n) g(n+h) =\sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}f'(ld) \sum_{q\le {{N+h}\over l}}g'(lq) \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1.$$ Here and from now on $\overline{d}$ denotes an inverse of $d$ modulo $q$. Next we split the summations with conditions $dq\le N/l$ and $dq>N/l$ (the $\ast$ in $q-$sums abbreviates $(q,d)=1$ hereafter): $$\begin{aligned} \sum_{n \le N} f(n) g(n+h) = & \sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}f'(ld) {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}g'(lq) \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1\\ & + O \Bigg( \sum_{{l|h}\atop {b:=-{h\over l}}}{1\over {l^{2\delta}}} \sum_{d\le {N\over l}}{1\over d^{\delta}} \sum_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over q^{\delta}} \sum_{{m\le {N\over {ld}}}\atop {md\equiv b \bmod q}}1 \Bigg),\end{aligned}$$ where the last sum is (thanks to condition $dq>N/l$) $$\ll_{\delta} \sum_{l|h}{1\over {l^{2\delta}}}\Big({l\over N}\Big)^{\delta} \sum_{n\le {N\over l}}d(n)d(n+h/l) \ll_{\delta,h} N^{1-\delta}(\log N)^2.$$ Here we have used the equivalent form of our hypothesis as per and . We now have the term $$\sum_{{l|h}\atop {b:=-{h\over l}}}\sum_{d\le {N\over l}}f'(ld) {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}g'(lq) \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1 =\sum_{l|h}\sum_{d\le {N\over l}}f'(ld) {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}g'(lq) \Big({N\over {ldq}}+O(1)\Big),$$ where the part with $O(1)$ contributes $$\ll_{\delta} \sum_{l|h}{1\over {l^{2\delta}}}\sum_{d\le {N\over l}}{1\over {d^\delta}}\sum_{q\le {N\over {ld}}}{1\over {q^\delta}} \ll_{\delta} \sum_{l|h}{1\over {l^{2\delta}}}\sum_{d\le {N\over l}}{1\over {d^\delta}}\Big({N\over {ld}}\Big)^{1-\delta} \ll_{\delta} N^{1-\delta}(\log N)$$ if $\delta \neq 1$ and $\ll \log^2 N$ if $\delta=1$. The main term is then coming from $$N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{f'(ld)}\over d} {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}{{g'(lq)}\over q}$$ which is written as $$\begin{aligned} & N \sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{f'(ld)}\over d} {\mathop{\enspace{\sum}^{\ast}}}_{q\le {{N+h}\over l}}{{g'(lq)}\over q} +O_{\delta}\Big( N \sum_{l|h}{1\over l^{1+2\delta}} \sum_{d\le {N\over l}}{1\over {d^{1+\delta}}} \sum_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over {q^{1+\delta}}}\Big)\\ &= N \sum_{l|h}l\sum_{d}{{f'(ld)}\over ld} {\mathop{\enspace{\sum}^{\ast}}}_{q}{{g'(lq)}\over lq} +O_{\delta}\Big(N^{1-\delta}\sum_{l|h}{1\over l^{1+\delta}} \sum_{d\le {N\over l}}{1\over d}\Big)\\ &= N \sum_{l|h}l\sum_{d}{{f'(ld)}\over ld} \sum_{{q}\atop {(q,d)=1}}{{g'(lq)}\over lq} +O_{\delta}\Big(N^{1-\delta} \log N\Big).\end{aligned}$$ Next we use the following fundamental property of the Möbius function (See page 3 of [@CM]). \[Mobius\] $$\sum_{d|n} \mu(d) = \begin{cases} 1 & \text{if } \ n=1,\\ 0 & \text{otherwise}. \end{cases}$$ Using [Lemma \[Mobius\]]{}, for $n=(q,d)$, we write $$\begin{aligned} \sum_{l|h}l\sum_{d}{{f'(ld)}\over ld} \sum_{{q}\atop {(q,d)=1}}{{g'(lq)}\over lq} &= \sum_{l|h}l\sum_{t}\mu(t)\sum_{d'}{{f'(ltd')}\over {ltd'}} \sum_{q'}{{g'(ltq')}\over {ltq'}}\\ &=\sum_{l|h}l \sum_{t}\mu(t)\hat{f}(lt)\hat{g}(lt)\\ &=\sum_{r=1}^{\infty}\hat{f}(r)\hat{g}(r)\sum_{{l|h}\atop {l|r}}l \mu\left({r\over l}\right) \\ &=\sum_{r=1}^{\infty}\hat{f}(r)\hat{g}(r)c_r(h).\end{aligned}$$ Thus we get $$\sum_{n \le N} f(n) g(n+h) = N \sum_{r=1}^{\infty}\hat{f}(r)\hat{g}(r)c_r(h) + O_{\delta,h}\Big(N^{1-\delta} \log^2 N+1\Big).$$ One can also follow steps of [@MS] and [@CMS]. The proof obtained in this way will be a little shorter. However, this will only prove a weaker version of this result. We briefly sketch it below. Keeping the principle of [@MS] and [@CMS] in mind, we consider a parameter $U$ tending to infinity which is to be chosen later. Then we write $$\sum_{n \le N} f(n) g(n+h) = A+B,$$ where $$A:= \sum_{n \le N} \sum_{\substack{r,s\\ rs \le U}} \hat f(r) \hat g(s) c_r(n)c_s(n+h) ~\mbox{ and }~ B:=\sum_{n \le N} \sum_{\substack{r,s\\ rs > U}} \hat f(r) \hat g(s) c_r(n)c_s(n+h).$$ As per our derivation in [@CMS], we have $$A =\begin{cases} N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O_h\left(\frac{N}{U^{1/2+\delta}}\right) + O(U^{1-\delta} \log^2 U) & ~ \mbox{ if } ~ \delta < 1,\\ N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O_h\left(\frac{N}{U^{3/2}}\right) + O(\log^3 U) & ~ \mbox{ if } ~ \delta = 1,\\ N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O_h\left(\frac{N}{U^{1/2+\delta}}\right) + O(1) & ~ \mbox{ if } ~ \delta > 1. \end{cases}$$ Using and the hypotheses on $\hat f(r)$ and $\hat g(r)$ we write $$|B| \ll \sum_{\substack{r \le N ,s \le N+h\\ rs > U}}\frac{1}{(rs)^{1+\delta}} \sum_{r' | r} r' \sum_{s' | s} s' \sum_{\substack{ n \le N \\ r'|n, s'|n+h}} 1.$$ Next we put $r= r' n_r$ and $s=s' n_s$. Hence $$\begin{aligned} |B| & \ll \sum_{r' \le N ,s' \le N+h} \frac{1}{(r's')^{\delta}} \sum_{\substack{n_r ,n_s \\ n_r n_s > U/r's'}} \frac{1}{(n_r n_s)^{1+\delta}} \sum_{\substack{ n \le N \\ r'|n,s'|n+h}} 1\\ & = \sum_{r' \le N ,s' \le N+h} \frac{1}{(r's')^{\delta}} \sum_{t > U/r's'} \frac{d(t)}{t^{1+\delta}} \sum_{\substack{ n \le N \\ r'|n, s'|n+h}} 1\\ & \ll \sum_{r' \le N ,s' \le N+h} \frac{1}{(r's')^{\delta}} \frac{\log (U/r's')}{(U/r's')^\delta} \sum_{\substack{ n \le N \\ r'|n, s'|n+h}} 1\\ & \ll \frac{\log(UN^2)}{U^\delta} \sum_{n \le N} d(n) d(n+h)\\ & \ll_h \frac{N \log^2 N \log(UN^2)}{U^\delta}.\end{aligned}$$ To optimize the error terms, we choose $$U=\begin{cases} N \log N & ~\mbox{ if }~\delta < 1,\\ N & ~\mbox{ if }~\delta = 1,\\ N^{1/\delta} (\log N)^{3/\delta} & ~\mbox{ if }~\delta > 1. \end{cases}$$ These choices yield $$\sum_{n \le N} f(n) g(n+h) = \begin{cases} N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O(N^{1-\delta} (\log N)^{3-\delta}) & ~ \mbox{ if } ~ \delta < 1,\\ N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O(\log^3 N) & ~ \mbox{ if } ~ \delta = 1,\\ N \sum_{r \ge 1} \hat f(r) \hat g(r) c_r(h) + O(1) & ~ \mbox{ if } ~ \delta > 1. \end{cases}$$ This proof starts off similarly. We just rewrite the hypotheses on $f'$ and $g'$ as $$\big| f'(d) \big|, \big| g'(d) \big| \ll {1\over {1+\log^\beta d}}$$ and obtain $$\begin{aligned} \sum_{n \le N} f(n) g(n+h) = & \sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}f'(ld) {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}g'(lq) \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1\\ & + O \Bigg( \sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}{1\over {1+\log^\beta(ld)}} {\mathop{\enspace{\sum}^{\ast}}}_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over {1+\log^\beta(lq)}} \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1 \Bigg)\\ = & \sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}f'(ld) {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}g'(lq) \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1 + O(R_1),\end{aligned}$$ where $$R_1:= \sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}{1\over {1+\log^\beta(ld)}} {\mathop{\enspace{\sum}^{\ast}}}_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over {1+\log^\beta(lq)}} \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1.$$ To estimate $R_1$ we separate according to the cases $d \le \sqrt{N}$ and $d > \sqrt{N}$. If $d \le \sqrt{N}$ then $$\log^{\beta}(lq)\gg_{\beta} \log^{\beta} N$$ for all $q>N/ld$, while $d > \sqrt{N}$ implies $$\log^{\beta}(ld)\gg_{\beta} \log^{\beta} N.$$ This yields $$\begin{aligned} R_1 & \ll_{\beta} \sum_{{l|h}\atop {b:=-{h\over l}}} \frac{1}{\log^\beta N} \sum_{d\le {N\over l}}\sum_{{N\over {ld}}<q\le {{N+h}\over l}} \sum_{{m\le {N\over {ld}}}\atop {md\equiv b \bmod q}}1\\ & \ll_{\beta} {1\over {\log^\beta N}}\sum_{l|h}\sum_{n\le {N\over l}} d(n) d(n+h/l)\\ & \ll_{\beta,h} {1\over {\log^{\beta} N}} \sum_{l|h} {N\over l} \log^2 {N\over l}\\ & \ll_{\beta,h} {N\over {\log^{\beta-2} N}}.\end{aligned}$$ Here we used the asymptotic estimate $$\sum_{n \le N/l} d(n) d(n+h/l) \sim \frac{6}{\pi^2} \sigma_{-1}(h/l) \frac{N}{l} \log^2(N/l),$$ due to Ingham [@AEI]. So now we are left to estimate $$\sum_{{l|h}\atop {b:=-{h\over l}}} \sum_{d\le {N\over l}}f'(ld) {\mathop{\enspace{\sum}^{\ast}}}_{q\le {N\over {ld}}}g'(lq) \sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b \bmod q}}1 = N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{f'(ld)}\over d} {\mathop{\enspace{\sum}^{\ast}}}_{{q\le {N\over {ld}}}}{{g'(lq)}\over q} +O(R_2),$$ where $$R_2:=\sum_{l|h}\sum_{d\le {N\over l}}|f'(ld)|\sum_{q\le {N\over {ld}}}|g'(lq)|.$$ Here we used the fact that $$\sum_{{m\le {N\over {ld}}}\atop {m\equiv \overline{d}b\bmod q}}1 = {N\over {ldq}}+O(1).$$ Using the hypothesis we get that $$R_2 \ll \sum_{l|h}\sum_{d\le {N\over l}}{1\over {1+\log^\beta(ld)}} \sum_{q\le {N\over {ld}}}{1\over {1+\log^\beta(lq)}}.$$ To treat the $q$-sum on the right hand side we split as follows: $$\sum_{q\le {N\over {ld}}}{1\over {1+\log^\beta(lq)}} \ll_{\beta} \sum_{q\le {1\over l}\sqrt{N\over d}}1 +{1\over {1+\log^{\beta}(N/d)}}\sum_{{1\over l}\sqrt{N\over d}<q\le {N\over {ld}}}1,$$ where we used that $$q>{1\over l}\sqrt{N\over d} \Rightarrow {1\over {1+\log^\beta(lq)}}\ll_{\beta} {1\over {1+\log^{\beta}(N/d)}}.$$ Hence $$\begin{aligned} R_2 & \ll_{\beta} \sum_{l|h}\sum_{d\le {N\over l}}{1\over {1+\log^\beta(ld)}} \ {N\over {ld(1+\log^{\beta}(N/d))}}\\ & \ll_{\beta} {N\over {\log^\beta N}}\sum_{l|h}{1\over l} \left(1+\sum_{1<d\le \sqrt{N}}{1\over {d\log^\beta(ld)}} +\sum_{\sqrt{N}<d\le {N\over l}}{1\over {d(1+\log^{\beta}(N/d))}}\right)\\ & \ll_\beta {N\over {\log^\beta N}}\sum_{l|h}{1\over l} \sum_{d\le {N\over l}}{1\over d}\\ & \ll_{\beta,h} \frac{N}{\log^{\beta-1} N}.\end{aligned}$$ Thus so far we have obtained that $$\sum_{n \le N} f(n) g(n+h) = N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{f'(ld)}\over d} {\mathop{\enspace{\sum}^{\ast}}}_{{q\le {N\over {ld}}}}{{g'(lq)}\over q} + O_{\beta,h}\left( \frac{N}{\log^{\beta-2} N} \right).$$ Now we essentially repeat what we did in the proof of [Theorem \[finite\]]{} and write $$N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{f'(ld)}\over d} {\mathop{\enspace{\sum}^{\ast}}}_{{q\le {N\over {ld}}}}{{g'(lq)}\over q} = M + O(R_3),$$ where $$M:=N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{f'(ld)}\over d} {\mathop{\enspace{\sum}^{\ast}}}_{{q\le {{N+h}\over l}}}{{g'(lq)}\over q} =N\sum_{r = 1}^\infty\hat{f}(r)\hat{g}(r)c_r(h)$$ and $$R_3:=N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{{|f'(ld)|}\over d} \sum_{{N\over {ld}}<q\le {{N+h}\over l}}{{|g'(lq)|}\over q}.$$ Using the hypothesis we get that $$\begin{aligned} R_3 & \ll N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{1\over d(1+\log^\beta(ld))} \sum_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over q(1+\log^\beta(lq))}\\ & \ll N\sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{1\over d(1+\log^\beta(ld)) (1+\log^\beta(N/d))} \sum_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over q}\\ & \ll \frac{N}{\log^\beta N} \sum_{l|h}{1\over l}\sum_{d\le {N\over l}}{1\over d} \sum_{{N\over {ld}}<q\le {{N+h}\over l}}{1\over q}\\ & \ll_h \frac{N}{\log^{\beta-2} N}.\end{aligned}$$ This completes the proof. Concluding remarks ================== The method outlined in this paper will undoubtedly have further applications as the theory moves forward. It offers us yet another way to approach these general convolution sums. The technical issues regarding absolute convergence of infinite series that complicated our earlier work have now been simplified through the use of finite Ramanujan expansions. As demonstrated in the paper, [*all*]{} arithmetical functions now afford a [*finite*]{} Ramanujan expansion. Convolution sums lie at the heart of analytic number theory. Earlier, there have been attempts to study such sums. Our paper offers yet another route to this study. We expect to investigate in future work further refinements of the theory.\ [**Acknowledgements:**]{} We would like to thank the referee for useful remarks which improved the presentation of this article. We would also like to thank the organisers of conferences ‘Leuca 2016’ and ‘CNTA 2016’ for their kind hospitality, where the final part of this work was done. [100]{} H.G. Gadiyar and R. Padma, Ramanujan-Fourier series, the Wiener-Khintchine formula and the distribution of prime pairs, [*Physica A.*]{} [**269**]{} (1999), 503-510. A.C. Cojocaru and M.R. Murty, [*An introduction to sieve methods and their applications*]{}, London Mathematical Society Student Texts [**66**]{}, Cambridge University Press, Cambridge, 2006. G. Coppola, M.R. Murty, B. Saha, [On the error term in a Parseval type formula in the theory of Ramanujan expansions II]{}, [*J. Number Theory*]{} [**160**]{} (2016), 700-715. H.G. Gadiyar, M.R. Murty and R. Padma, [Ramanujan - Fourier series and a theorem of Ingham]{}, [*Indian J. Pure Appl. Math.*]{} [**45**]{} (2014), no. 5, 691-706. A.E. Ingham, [ Some asymptotic formulae in the theory of numbers]{}, [*J. London Math. Soc.*]{} [**2**]{} (1927), no. 3, 202-208. M.R. Murty, [ Ramanujan series for arithmetical functions]{}, [*Hardy-Ramanujan J.*]{} [**36**]{} (2013), 21-33. M.R. Murty and B. Saha, [ On the error term in a Parseval type formula in the theory of Ramanujan expansions]{}, [*J. Number Theory*]{} [**156**]{} (2015), 125-134. S. Ramanujan, [On certain trigonometrical sums and their applications in the theory of numbers]{}, [*Trans. Cambridge Philos. Soc.*]{} [**22**]{} (1918), no. 13, 259-276. B. Saha, A note on arithmetical functions with absolutely convergent Ramanujan expansions, [*Int. J. Number Theory*]{} [**12**]{} (2016), no. 6, 1595-1611. [^1]: Research of the second author was partially supported by an NSERC Discovery grant.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Weyl semimetals are a three-dimensional topological phase of matter with isolated band touchings in the Brillouin Zone. These points have an associated chirality, and many of the proposals to detect the Weyl semimetal state rely on the chiral anomaly. A consequence of the chiral anomaly is that under the application of an ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field, charge is transferred between points of opposite chirality. In this paper we propose an optical absorption experiment that provides evidence for the chiral anomaly. We use the Kubo formula, and find that an applied ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ induces the formation of step-like features at finite frequency in the interband optical conductivity. We study the effect of scattering, and finite temperatures on this feature and find that it should be observable at low temperatures in pure samples. Finally we discuss how the application of an ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field can be used to map out the frequency dependence of the scattering rate.' author: - 'Phillip E. C. Ashby' - 'J. P. Carbotte' bibliography: - 'biblio.bib' title: Chiral anomaly and optical absorption in Weyl semimetals --- Introduction ============ Weyl semimetals are a novel topological phase of matter and have attracted considerable interest.[@Wan:2011fk; @Hosur:2012fk; @Witczak-Krempa:2012ve; @vafek:2013fk] A Weyl semimetal is a three-dimensional system whose band structure contains pairs of bands crossings (called Weyl points) at isolated points in the Brillouin zone (BZ). For such a band crossing to occur, the Weyl semimetal state must break time-reversal or inversion symmetry. Each Weyl point can be assigned a chirality, $\chi$, that takes values $\pm1$. The Nielsen-Ninomiya theorem[@Nielsen:1981fv] shows that the number of Weyl points in the BZ must be even, with half the points of each chirality. Near a Weyl point with chirality $\chi$, the Hamiltonian takes the form $$\begin{aligned} \label{eq:hamil} \mathcal{H} = \chi v_F {\ensuremath{\boldsymbol{\mathrm{k}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{\sigma}}}},\end{aligned}$$ where ${\ensuremath{\boldsymbol{\mathrm{k}}}}$ is the momentum measured from the Weyl point, and ${\ensuremath{\boldsymbol{\mathrm{\sigma}}}}$ is the vector of Pauli matrices. The three-dimensional nature ensures that the Weyl points are stable against perturbations. In fact, the only way to annihilate a Weyl point is if two points of opposite chirality meet in the BZ. There are many candidate materials for the Weyl semimetal, yet there is still no compelling experimental evidence for the observation of one. The pyrochore iridates[@Wan:2011fk; @Witczak-Krempa:2012ve], as well as topological insulator heterostructes[@Burkov:2011kx; @Burkov:2011ys; @Zyuzin:2012zr; @Halasz:2012ly] were among the first systems proposed to host the Weyl semimetal state. It is also possible that certain quasicrystals may be host to the Weyl semimetal state[@Timusk:2013fk]. The observation of linear conductivity over a wide frequency range is a sign of Dirac physics. When combined with the lack of inversion symmetry there are sufficient conditions for the existence of the Weyl semimetal state. Recently there has been evidence for three dimensional Dirac physics in both Cd$_3$As$_2$ [@Neupane:fk2013; @Borisenko:fk2013]and Na$_3$Bi[@xu:2013nb]. The discovery of three dimensional Dirac materials is a promising first step towards the discovery of a Weyl semimetal. Ideally, to correctly identify the Weyl semimetal, one needs as many probes as possible that can uniquely identify it from other phases (such as three dimensional Dirac semimetals). Most of the research to date has focused on anomalous properties that can be traced back to the chiral anomaly[@Zyuzin:2012fkk; @Grushin:2012fk; @Aji:2012fk; @Heon:2013fk; @Liu:2013fkk; @Landsteiner:2014fk]. The chiral anomaly is a peculiar non-conservation of chiral charge and has been mostly discussed in the context of high-energy physics. The Weyl semimetal is a condensed matter realization of the chiral anomaly and adds to the growing list of high-energy phenomenon in condensed matter systems[@Volovik]. In the presence of external fields ${\ensuremath{\boldsymbol{\mathrm{E}}}}$ and ${\ensuremath{\boldsymbol{\mathrm{B}}}}$ the continuity equation for a Weyl point of chirality $\chi$ takes the form $$\begin{aligned} \label{eq:contin} {\ensuremath{\frac{\partial n}{\partial t}}}+{\ensuremath{{\ensuremath{\nabla}}\cdot}}{{\ensuremath{\boldsymbol{\mathrm{j}}}}} = \frac{\chi}{4\pi^2}{\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}.\end{aligned}$$ That is, the charge density at a single Weyl point is not conserved in the presence of parallel ${\ensuremath{\boldsymbol{\mathrm{E}}}}$ and ${\ensuremath{\boldsymbol{\mathrm{B}}}}$ fields. The missing (extra) charge at a given Weyl point is compensated at another Weyl point of the opposite chirality to ensure the overall conservation of charge in the system. Thus the application of parallel ${\ensuremath{\boldsymbol{\mathrm{E}}}}$ and ${\ensuremath{\boldsymbol{\mathrm{B}}}}$ fields can be used to drive charge between Weyl points of opposite chirality. This charge pumping would continue until it is cutoff by some relaxation time, $\tau$ that corresponds to scattering between the two Weyl points. As these scattering processes generically involve large momenta, $\tau$ is expected to be large[@hosur:2014fk]. Most of the proposals to detect a Weyl semimetal have focused around experiments hoping to detect the chiral anomaly in some form. One of the transport properties tied to the chiral anomaly is a large longitudinal magnetoconductivity[@Gorbar:2014fk]. Other transport predictions have focused on an anomalous non-quantized Hall effect[@Xu:2011kx; @Goswami:2012fk; @Burkov:2011kx; @Zyuzin:2012zr], that is proportional to the separation of the Weyl points in momentum space. The chiral magnetic effect is another consequence of the chiral anomaly wherein current flows parallel to an applied magnetic field[@Jian-Hui:2013kx; @Chen:2013uq]. Another transport experiment was proposed in which chiral charge pumping could be measured as a voltage drop over long distances[@Parameswaran:fk2013]. There has also been a proposal to measure the chiral anomaly through the non-vanishing gyrotropy induced by external fields[@hosur:2014fk; @Goswami:2014fk]. Most recently, the density response of Weyl semimetals was studied[@Pesin:2014fk], showing that both the compressibility and plasmon modes contain signatures unique to the Weyl semimetal state. In this paper, we propose an optical absorption experiment that measures the chiral anomaly. Our proposal takes advantage of the charge pumping induced by the chiral anomaly. The charge pumping leaves different Weyl points at different chemical potentials and causes measurable effects in both the Drude peak, as well as in the interband portion of the conductivity. In particular, we identify sharp step-like features in the interband conductivity that should prove as another ‘smoking gun’ for the Weyl semimetal state. Single point conductivity ========================= ![(Color online) The finite frequency optical conductivity for a clean Weyl semimetal at $T=0$. In black we show the optical conductivity for a doped Weyl semimetal, we have used a small broadening, $\gamma = 0.01$, to the Drude for graphing purposes. After the application of the applied ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field charge is pumped from one Weyl point to the other, and step-like signatures appear in the interband portion of the optical conductivity. The missing interband spectral weight is transferred to the Drude peak. The measurement of these interband features is tied to the chiral anomaly and would be a direct signature of a Weyl semimetal.[]{data-label="fig:main"}](finite_mu7.eps){width="0.8\linewidth"} Our starting point is the Kubo formula for the optical conductivity. Written in terms of the spectral functions, $A$, and for a chemical potential, $\mu$, the $xx$ component of the real part of the optical conductivity is given by $$\begin{aligned} {\ensuremath{\nonumber}}\sigma_{xx}(\Omega)= \frac{e^2\pi}{\Omega}&\int d\omega [f(\omega-\mu) - f(\omega-\mu+\Omega)]\\ \label{eq:kubo} &\times\int \frac{d^3k}{(2\pi)^3}\mathrm{Tr}\left[v_x\hat{A}({\ensuremath{\boldsymbol{\mathrm{k}}}},\omega)v_x\hat{A}({\ensuremath{\boldsymbol{\mathrm{k}}}},\omega+\Omega)\right].\end{aligned}$$ Here $f(x) = 1/(e^{x/T}+1)$ is the Fermi function, and $v_x = {\ensuremath{\frac{\partial \mathcal{H}}{\partial k_x}}} = \sigma_x$ are the velocity operators. We work in units where $\hbar = v_F = k_B = 1$ and all photon energies and temperatures are in meV (in Appendix \[sec:apa\] we restore the factors of $v_F$ and $\hbar$ for clarity). The spectral functions are found from the decomposition of the Green’s function $$\begin{aligned} \hat{G}({\ensuremath{\boldsymbol{\mathrm{k}}}},\omega) = \int d\omega' \frac{\hat{A}({\ensuremath{\boldsymbol{\mathrm{k}}}},\omega')}{\omega-\omega'}.\end{aligned}$$ After evaluating the trace and converting the ${\ensuremath{\boldsymbol{\mathrm{k}}}}$-space integration to an integral over energy, the optical conductivity appears as the sum of two terms, $\sigma_{xx} = \sigma^\textrm{D}+\sigma^\textrm{IB}$. Details of this derivation are provided in Appendix \[sec:apa\]. The first term is a Drude (or intraband) term, $\sigma^\textrm{D}$, and the second is an interband term, $\sigma^{\textrm{IB}}$. We find $$\begin{aligned} {\ensuremath{\nonumber}}\sigma^{\textrm{D}}_{xx}(\Omega) = \frac{e^2}{6\pi}\int_{-\infty}^\infty d\omega \frac{f(\omega-\mu)-f(\omega-\mu+\Omega)}{\Omega}\\\times\int_0^\infty d\epsilon \epsilon^2\left[A(\epsilon,\omega)A(\epsilon,\omega+\Omega)+A(-\epsilon,\omega)A(-\epsilon,\omega+\Omega)\right],\end{aligned}$$ and $$\begin{aligned} {\ensuremath{\nonumber}}\sigma^{\textrm{IB}}_{xx}(\Omega) = \frac{e^2}{3\pi}\int_{-\infty}^\infty d\omega \frac{f(\omega-\mu)-f(\omega-\mu+\Omega)}{\Omega}\\\times\int_0^\infty d\epsilon \epsilon^2\left[A(\epsilon,\omega)A(-\epsilon,\omega+\Omega)+A(-\epsilon,\omega)A(\epsilon,\omega+\Omega)\right].\end{aligned}$$ We would like to point out the factor of two difference between the interband term and the Drude term. This factor of two arises when preforming the angular integration in Eq. (\[eq:kubo\]). In the presence of a self energy $\Sigma(\omega)$ the spectral functions are given by $$\begin{aligned} A(\pm\epsilon,\omega) = \frac{1}{\pi}\frac{-\textrm{Im}\Sigma(\omega)}{\left(\omega-\textrm{Re}\Sigma(\omega)\mp\epsilon\right)^2+\left(\textrm{Im}\Sigma(\omega)\right)^2}.\end{aligned}$$ We can now perform the integration over $\epsilon$ to obtain expressions for the conductivity. An essential feature of our expressions is that they retain the energy dependence of $\Gamma(\omega)$. We adopt the shorthand $\Gamma(\omega) = \Gamma$ and $\Gamma(\omega+\Omega) = \Gamma'$ for the frequency dependent scattering rate. In the limit of a small impurity scattering rate, and neglecting the real part of the self energy we obtain $$\begin{aligned} {\ensuremath{\nonumber}}\sigma^\textrm{D}(\Omega)=\frac{e^2}{6\pi^2}\int_{-\infty}^\infty d\omega\frac{f(\omega-\mu)-f(\omega+\Omega-\mu)}{\Omega}\\ \label{eq:drude}\times\frac{\omega^2\Gamma'+(\omega+\Omega)^2\Gamma}{(\Gamma+\Gamma')^2+\Omega^2}\end{aligned}$$ for the Drude piece, and $$\begin{aligned} {\ensuremath{\nonumber}}\sigma^\textrm{IB}(\Omega)= \frac{e^2}{3\pi^2}\int_{-\infty}^\infty d\omega\frac{f(\omega-\mu)-f(\omega+\Omega-\mu)}{\Omega}\\ \label{eq:ib}\times\frac{\omega^2\Gamma'+(\omega+\Omega)^2\Gamma}{(\Gamma+\Gamma')^2+(2\omega+\Omega)^2}\end{aligned}$$ for the interband piece. In the strict $\Gamma = 0$ and $T = 0$ limit our formula reduces to the well known result $$\begin{aligned} \label{eq:clean}\sigma_{xx}(\Omega)= \frac{e^2\mu^2}{6\pi v_F}\delta(\Omega)+\frac{e^2\Omega}{24\pi v_F}\Theta(\Omega-2|\mu|),\end{aligned}$$ where we have restored the factor of the Fermi velocity that defines the Weyl Fermions (see also Eq. (\[eq:a14\]) and (\[eq:a16\])). This makes it clear how the relativistic Hamiltonian Eq. (\[eq:hamil\]) impacts the conductivity. For a moment let us consider only the intraband term, $\sigma^\textrm{D}$. In the small $\Omega$ limit (appropriate for this term) we have $\Gamma = \Gamma'$ and for $\Omega \ll T$ the difference $[f(\omega-\mu)-f(\omega+\Omega-\mu)]/\Omega$ can be replaced by $-{\ensuremath{\frac{\partial f}{\partial \omega}}}$. In this limit we have $$\begin{aligned} \sigma^\textrm{D}(\Omega) = \frac{e^2}{3\pi}\int_{-\infty}^\infty d\omega\left(-{\ensuremath{\frac{\partial f}{\partial \omega}}}\right)\omega^2\frac{\Gamma}{4\Gamma^2+\Omega^2},\end{aligned}$$ which is precisely the form of the Boltzmann equation for the conductivity used elsewhere in the literature[@Burkov:2011ys]. Weyl semimetal -------------- Now that we have obtained formulae for a single Weyl point we can examine the consequences in the Weyl semimetal state. A real Weyl semimetal contains an even number of Weyl points, with half of each chirality. For our purposes it will be sufficient to consider the case of two Weyl points. The chiral anomaly implies that the application of a constant ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ will induce a charge difference between the pair of Weyl points. This change in charge density is captured through a change in the chemical potential at each Weyl point. The charge density continues to change until one reaches a steady state characterized by some relaxation time, $\tau$. The pumping of the chemical potential due to the applied field is given by $\mu_p^3 = \frac{3e^2\hbar v_F^3}{2}{\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}\tau$ (we have restored the factors of $\hbar$ and $v_F$ in this equation for clarity). For a Weyl point of chirality $\chi$ and chemical potential $\mu$ before the application of the ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field, the resulting chemical potential after pumping is given by[@hosur:2014fk] $$\begin{aligned} \mu_\chi^3 = \left(\mu^3+\chi\mu_p^3\right),\end{aligned}$$ and it is understood that the real root should always be chosen. To estimate the size of the chemical potential shift we take $v_F = 4.3\times10^5$m/s, $B = 1$T, $E = 10^3$V/m and $\tau = 10^{-11}$s which gives $\mu_p =9.18$meV. Such a chemical potential shift should be observable in a low frequency optical experiment. The value of $v_F$ stated above is typical of 2D Dirac systems[@Liu14:PRB; @ZhouCarb:fk14] and is conservative for 3D Dirac materials. An experiment by Orlita [*et al.*]{}[@Orlita:2014fk] presents spectroscopic evidence for 3D Dirac Fermions with a velocity $v_F = 10^6$m/s. Timusk [*et al.*]{}[@Timusk:2013fk] recently pointed out that the quasicrystal AlCuFe and its related approximant Al$_2$Ru show a conductivity that is remarkably linear over a large energy range ($>$0.5eV). Their best estimate of the Fermi velocity is $4.3\times10^6$m/s. Such large Fermi velocities ($v_F >10^6$m/s) leads to a chemical potential shift 10 times larger than our estimated energy. In Figure \[fig:main\] we show the finite frequency optical conductivity for two Weyl points in the clean limit ($\Gamma=0$) at $T=0$. In this case the optical conductivity is simply given by $$\begin{aligned} \label{eq:clean2}\sigma_{xx}(\Omega)= \frac{e^2}{6\pi}\sum_\chi\left[\mu_\chi^2\delta(\Omega)+\frac{\Omega}{4}\Theta(\Omega-2|\mu_\chi|)\right].\end{aligned}$$ The Weyl points pictured have a chemical potential initially at $7$meV. The optical conductivity is given by Eq. (\[eq:clean2\]) (with $\mu_\pm =$ 7meV) and is shown in black. After the application of the ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field the chemical potentials change at the two Weyl points. The resulting optical conductivity after charge pumping is pictured in orange. As can be seen in Figure \[fig:main\] the pumping has two effects: a step-like feature has appeared in the interband conductivity, and spectral weight has been transferred to the Drude peak. Careful measurement of these two features as a function of applied ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ would be a direct signal of the Weyl semimetal state since this phenomenon is intimately linked with the chiral anomaly. There are two special cases that we now mention. The first is an undoped Weyl semimetal. In this case, the charge pumping simply transforms it into a doped Weyl semimetal with doping $\mu_p$ (one Weyl cone is electron doped, and the other is hole doped). In this case there is a single step in the interband conductivity, and the missing spectral weight is transferred to the Drude. The second case is when the chemical potential due to pumping, $\mu_p$ exactly matches the initial doping, $\mu$. In this case, one Weyl point is completely drained of its charge density. The point that sits at charge neutrality has no Drude and contributes linearly at all frequencies. The resulting optical conductivity has a peculiar shape (see the blue curve in Fig \[fig:effects\]). ![(Color online) Top: The finite frequency conductivity for several different values of the charge pumping. In this figure we show how disorder smears out the step-like features in the interband. The pictured curves are for a residual scattering rate of $\Gamma = 0.1\mu$. The step-like features are most prominent when $\mu_p \approx \mu$ and are clearly visible as long as the scattering is not an appreciable fraction of $\mu$. Also notice the anomalous case, $\mu_p = \mu$ which contains a linear piece all the way to zero frequency. Bottom:The finite frequency conductivity for several different values of the charge pumping at finite temperature. Finite temperature shifts the chemical potential downward, as well as introducing thermal broadening. This figure has $T=0.1\mu$, the same energy as we used in the residual scatting figure. However, the effect of temperature is much more noticeable, and the step-like features have almost been completely washed out.[]{data-label="fig:effects"}](fig1.eps "fig:"){width="0.8\linewidth"} ![(Color online) Top: The finite frequency conductivity for several different values of the charge pumping. In this figure we show how disorder smears out the step-like features in the interband. The pictured curves are for a residual scattering rate of $\Gamma = 0.1\mu$. The step-like features are most prominent when $\mu_p \approx \mu$ and are clearly visible as long as the scattering is not an appreciable fraction of $\mu$. Also notice the anomalous case, $\mu_p = \mu$ which contains a linear piece all the way to zero frequency. Bottom:The finite frequency conductivity for several different values of the charge pumping at finite temperature. Finite temperature shifts the chemical potential downward, as well as introducing thermal broadening. This figure has $T=0.1\mu$, the same energy as we used in the residual scatting figure. However, the effect of temperature is much more noticeable, and the step-like features have almost been completely washed out.[]{data-label="fig:effects"}](fig2.eps "fig:"){width="0.8\linewidth"} The presence of these steps can be understood from an inspection of Eq. (\[eq:clean\]). Most simply, the steps are a consequence of the Pauli principle for the two absorption scales $\mu_\chi$. We would like to point out that Eq. (\[eq:clean\]) was obtained in the analysis the Faraday and Kerr rotations presented by Hosur and Qi[@hosur:2014fk]. Our calculation of the Kubo bubble is essentially the same as theirs except that we have not taken the $T\rightarrow0$ or $\Gamma\rightarrow0$ limit. Indeed, the Faraday and Kerr effects can be derived from the optical conductivity. Here we have chosen to focus on the direct measurement of the changes in the optical conductivity, rather than a derived quantity. The experiment outlined by Hosur and Qi[@hosur:2014fk] is both technically challenging, and produces an incredibly small signal (picoradians for the Kerr effect). On the other hand, the chemical potential shift of 9meV occurs at frequencies routinely measured in infrared spectroscopic experiments[@RevModPhys.77.721]. Thus, we expect that a direct measurement of the chiral chemical potential shift through optical absorption will be favorable. Impurities ---------- We now turn to the effect of impurities on the features that we saw in Figure \[fig:main\]. It is important to understand if these features will still be observable in the presence of disorder. As a first approximation we take a constant residual scattering rate in Eqns. (\[eq:drude\]) and (\[eq:ib\]). In this case, at $T=0$ we obtain simple expressions for the single node conductivity: $$\begin{aligned} \sigma^\textrm{D}(\Omega,T=0)=\frac{e^2\mu^2\Gamma}{3\pi^2(4\Gamma^2+\Omega^2)},\end{aligned}$$ and $$\begin{aligned} {\ensuremath{\nonumber}}\sigma^\textrm{IB}(\Omega,T=0)=\frac{e^2}{24\pi^2}\left[4\Gamma+\Omega\textrm{arccot}\left(\frac{2\Gamma}{2\mu+\Omega}\right)\right.\\ \left.-\Omega\textrm{arccot}\left(\frac{2\Gamma}{2\mu-\Omega}\right)\right].\end{aligned}$$ In the formula for $\sigma^\textrm{D}$ we have assumed $\Omega \ll \mu$. We see that in this case $\sigma^\textrm{D}$ takes the form of a simple Drude peak. The conductivity is plotted as a function of frequency in the top panel of Figure \[fig:effects\] for a constant residual scattering rate $\Gamma = 0.1\mu$ for several different values of the charge pumping. Notice the peculiar shape of the conductivity when $\mu_p = \mu$. At all values of $\mu_p$ the step-like feature at finite frequency has been smoothed out by disorder, but is still clearly visible. The step-like feature is most pronounced when $\mu_p \approx \mu$ due to the way the chemical potentials add, away from this region the larger of the two energies, $\textrm{max}[\mu,\mu_p]$, dominates the shape of the conductivity. Finally we considered the Born approximation, where the scattering rate is proportional to the density of states, $\Gamma(\omega) \propto g(\omega)$. We found that the form of scattering had little effect on the interband features in the conductivity. The impact of the Born approximation on the Drude will be discussed later. ![(Color online) Here we show the effect of finite temperature on the shape of the Drude peak for two different types of scattering. Constant residual scattering is pictured on the left, and finite temperature simply causes the Drude peak to broaden. On the right we show the case of Born limit scattering. In this case the scattering rate is frequency dependent and the lineshape depends on which energy is dominant. For $T<|\mu|$ the lineshape has a Drude form, while for $T > |\mu|$ the lineshape develops a sharp cusp.[]{data-label="fig:drude"}](fig3.eps){width="0.8\linewidth"} Finite temperature ------------------ Until now we have shown results at $T=0$. A Weyl semi-metal has a non-constant density of states, which results in the chemical potential being strongly temperature dependent. Since we are interested in the effects of finite temperature as well as finite doping, we include the shift in $\mu$ due to finite $T$. To find the chemical potential as a function of temperature we require that the charge density remain constant as we change T. The charge density is given by $$\begin{aligned} n = \int \frac{d^3k}{(2\pi)^3}\left[f(\epsilon-\mu)-f(\epsilon+\mu)\right].\end{aligned}$$ We use the identity $$\begin{aligned} \frac{df(\epsilon\pm\mu)}{dT} = \left(\frac{\epsilon\pm\mu}{T}\mp\frac{d\mu}{dT}\right)\left(-{\ensuremath{\frac{\partial f}{\partial \epsilon}}}\right)\end{aligned}$$ and integrate over $k$. Thus, we obtain the following differential equation for $\mu$: $$\begin{aligned} \frac{d\mu}{dT}\left(\mu^2+\frac{\pi^2T^2}{3}\right)+\frac{2\pi^2T}{3}\mu = 0.\end{aligned}$$ Combining this with the boundary value $\mu(0) = \mu_0$ gives the following solution $$\begin{aligned} \mu(T) = \frac{2^{1/3}\left(9\mu_0^3+\sqrt{81\mu_0^6+12\pi^6T^6}\right)^{2/3}-2\pi^23^{1/3}T^2}{6^{2/3}\left(9\mu_0^3+\sqrt{81\mu_0^6+12\pi^6T^6}\right)^{1/3}}.\end{aligned}$$ This equation gives the chemical potential at finite $T$, which is rapidly suppressed as $T$ increases. Thus, finite $T$ has two effects on the conductivity: The first is the usual thermal broadening from the Fermi functions. The second is the finite $T$ shift, which moves the $\mu_p$ split chemical potentials closer to one another. Both of these effects tend to smear out the step-like features that identify the Weyl semimetal. The conductivity at $T=0.1\mu$ is plotted as a function of frequency for several values of $\mu_p$ in the bottom panel of Figure \[fig:effects\]. The effect of temperature is much more drastic than disorder, even though the energy scale is similar. Although the step-like feature is almost completely smeared out, the onset of spectral weight transfer from the interband to the Drude due to finite $\mu_p$ is still a clear signature of the Weyl semimetal state. Finite temperature also has a large effect on the shape of the Drude peak at $\Omega = 0$. In Figure \[fig:drude\] we show the Drude peak for both constant residual scattering (left) and Born limit scattering (right). For constant residual scattering, the peak always has a Drude from and the effect of finite temperature only broadens the peak further. In the Born limit the scattering rate is proportional to the density of states: $\Gamma(\omega) = \gamma\omega^2/\pi$, resulting in an unusual lineshape. For a constant scattering rate, the conductivity as a function of $\omega$ is concave down for small $\omega$ and approaches $\omega=0$ with zero slope. This is in stark contrast to the dashed red curve in the right hand frame of Figure \[fig:drude\] which shows a cusp at $\omega=0$ and is concave up. The unusual lineshape associated with Born limit scatterers was first pointed out by Burkov and Balents for the case of a $\mu = 0$ Weyl semimetal[@Burkov:2011kx]. They showed that in this case the line shape had a cusp of the form $$\begin{aligned} \sigma^\textrm{D}(\Omega) = \frac{e^2}{6\pi\gamma}\left(1-\frac{1}{8}\sqrt{\frac{2\pi^3\Omega}{\gamma T^2}}\right).\end{aligned}$$ This lineshape follows from our expressions in the Born limit provided that $\Omega \ll T$. If the chemical potential is increased by the application of an ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field so that $|\mu| > T$ then the lineshape for a single Weyl point takes the form $$\begin{aligned} \sigma^\textrm{D}(\Omega) = \frac{e^2\gamma}{3\pi^2}\left[\frac{\mu^4}{\frac{4\gamma^2}{\pi^2}\mu^4+\Omega^2}\right],\end{aligned}$$ which has the form of a Drude peak. The width of the Drude peak can be increased simply by increasing the applied ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field. Using the chiral anomaly to change $\mu_p$ allows one to change the energy scale at which $\Gamma(\omega)$ is probed. In this way the frequency dependence of a general $\Gamma(\omega)$ can be completely mapped out, since $\Gamma(\omega)$ is responsible for the shape of $\sigma^\textrm{D}$. The spectral weight in the Drude peak characterizes its width. At finite $T$ and $\mu$ we find that the spectral weight is given by $$\begin{aligned} \int_0^\infty d\omega\sigma(\omega) = \frac{1}{12\pi}\left(\mu^2+\frac{T^2\pi^2}{3}\right).\end{aligned}$$ The Drude weight does not vanish at $T=0$ since it is experimentally impossible to arrange $\mu=0$. Conclusions =========== We have described a method for detecting the chiral anomaly using an optical absorption experiment. The chiral anomaly is one way in which the Weyl semimetal is distinct from its 3D-Dirac cousin. Direct measurement of the chiral anomaly is therefore a sign of bulk Weyl points. We show that there are signatures in the finite frequency optical response that can be controlled through the application of an applied ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field. After the application of an ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field a pair of Weyl points have different chemical potentials and step-like features appear in the interband portion of the conductivity. An important result is that the signature of this anomaly in the absorptive part of the conductivity is large and should be easily detected with presently available optical absorption techniques. In addition to this, spectral weight is transferred to the Drude peak at $\Omega = 0$. We showed that these features remain as long as the scattering rate remains small compared to the chemical potential. We estimated that the pumping provided by the ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ term in Eqn. (\[eq:contin\]) leads to a pump chemical potential on the order of 9meV and thus the impurity scattering rate should be kept below a few meV. This impurity scattering rate is both realistic and achievable. The effect of finite temperature had a more dramatic effect on the interband transitions, so it seems likely that low temperatures will be required to resolve the steps cleanly. Optical experiments are routinely carried out at a few Kelvin and so we expect this should not pose a technical challenge. Finally, we discussed how the application of the ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field can be used to trace out the frequency dependence of the scattering rate. Since the ${\ensuremath{\boldsymbol{\mathrm{E}}}}\cdot{\ensuremath{\boldsymbol{\mathrm{B}}}}$ field controls the chemical potential, it can be used to probe $\Gamma(\omega)$ at many different energy scales. It is natural to wonder about the effects that the gapless surface states of Weyl semimetals (known as Fermi arcs) would have on the measurement of the optical conductivity. Since the optical conductivity is a bulk probe, we expect our results to be largely unchanged by the presence of the Fermi arc surface states. The features in the interband occur at energies too high to be affected by the low energy surface states. The Fermi arcs may enhance the low energy absorption (i.e. the Drude), however, analysis of the changes to the low frequency absorption would require the study of the Weyl semimetal in the presence of a boundary. This boundary value problem is beyond the scope of the results presented here. We would like to address one final point. In this work we considered the case of a weak magnetic field where Landau level quantization was unimportant. If one applies a ${\ensuremath{\boldsymbol{\mathrm{B}}}}$ field strong enough that Landau level formation is important, the interband lineshapes discussed here should be replaced instead by the magneto-optical conductivity lineshapes[@Ashby:2013ys]. The step-like features will still appear in the interband, but occur from a superposition of the magneto-optical lineshapes instead of the free Fermion lineshapes discussed here. This work was supported by the Natural Sciences and Engineering Research Council of Canada and the Canadian Institute for Advanced Research. Calculation of Optical Conductivity {#sec:apa} =================================== The Kubo formula reads $$\begin{aligned} \textrm{Re}\sigma_{ij}(\Omega) = \frac{e^2\pi}{\Omega}\int_{-\infty}^\infty d\omega [f(\omega-\mu) - f(\omega+\Omega)]\int \frac{d^3k}{(2\pi)^3} \mathrm{Tr}\left[v_i\hat{A}({\ensuremath{\boldsymbol{\mathrm{k}}}},\omega)v_j\hat{A}({\ensuremath{\boldsymbol{\mathrm{k}}}},\omega+\Omega)\right]\end{aligned}$$ The velocity operators are given by ${\ensuremath{\boldsymbol{\mathrm{v}}}} = {\ensuremath{\frac{\partial \mathcal{H}}{\partial {\ensuremath{\boldsymbol{\mathrm{k}}}}}}} = v_F{\ensuremath{\boldsymbol{\mathrm{\sigma}}}}$. We can find the spectral densities from the Greens function. The inverse Greens function is given by $$\begin{aligned} \mathcal{G}^{-1}(z) = \left(\begin{matrix} z-k_z & -k_x+ik_y \\ -k_x-ik_y & z+k_z \\ \end{matrix}\right)\end{aligned}$$ Inverting gives $$\begin{aligned} \mathcal{G}(z) = \frac{1}{|k|^2-z^2}\left(\begin{matrix} -k_z-z & -k_x+ik_y \\ -k_x-ik_y & k_z-z \\ \end{matrix}\right)\end{aligned}$$ The spectral densities follow from the relation $$\begin{aligned} \mathcal{G}(z) = \int_{-\infty}^\infty d\omega \frac{\hat{A}(\omega)}{z-\omega}\end{aligned}$$ We can read off the components of $\hat{A}$ directly. For example, $$\begin{aligned} \mathcal{G}_{11} = \frac{-k_z-z}{|k|^2-z^2} = \frac{-k_z-|k|}{2|k|(|k|-z)}+\frac{-k_z+|k|}{2|k|(|k|+z)}.\end{aligned}$$ Now defining $$\begin{aligned} u^2 = \frac{1}{2}\left(1+\frac{k_z}{|k|}\right)\\ v^2 = \frac{1}{2}\left(1-\frac{k_z}{|k|}\right),\end{aligned}$$ we have that $A_{11} = u^2\delta(\omega-|k|)+v^2\delta(\omega+|k|)$. Similarly, $A_{22} = v^2\delta(\omega-|k|)+u^2\delta(\omega+|k|)$. Now for $\sigma_{xx}$ the trace takes the form $\mathrm{Tr}\left[\sigma_x\hat{A}\sigma_x\hat{A}'\right] = A_{12}A_{12}'+A_{21}A_{21}'+A_{22}A_{11}'+A_{11}A_{22}'$. The terms proportional to $A_{12}$ and $A_{21}$ will vanish once the angular integration is carried out since they are proportional to $k_x\pm ik_y$. So we have $\mathrm{Tr}\left[\sigma_x\hat{A}\sigma_x\hat{A}'\right] = A_{22}A_{11}'+A_{11}A_{22}'$. Now we define $A_{\pm} = \delta(\omega\mp\epsilon_k)$ and $A_{\pm}' = \delta(\omega+\Omega\mp\epsilon_k)$. Expanding out the trace we have $\mathrm{Tr}\left[\sigma_x\hat{A}\sigma_x\hat{A}'\right] = 2u^2v^2\left(A_+A_+'+A_-A_-'\right)+\left(u^4+v^4\right)\left(A_-A_+'+A_+A_-'\right)$. We will now perform the angular integration. We need $u^4+v^4 = \frac{1}{2}\left(1+\frac{k_z^2}{|k|^2}\right)$ and $2u^2v^2 = \frac{1}{2}\left(1-\frac{k_z^2}{|k|^2}\right)$. The relevant integrals are thus $$\begin{aligned} &\int\frac{d^3k}{(2\pi)^3}\frac{1}{2}\left(1\pm\frac{k_z^2}{|k|^2}\right)\\ &=\frac{1}{2}\int d\epsilon\frac{\epsilon^2}{2\pi^2v_F^3} \pm \frac{1}{2}\int d\epsilon\frac{\epsilon^2}{(2\pi)^3v_F^3}2\pi\int_0^\pi d\theta \sin(\theta)\cos^2(\theta)\\ & = \frac{3\pm1}{6v_F^3}\int d\epsilon\frac{\epsilon^2}{2\pi^2}\end{aligned}$$ Finally we obtain $$\begin{aligned} \textrm{Re}\sigma_{xx}(\Omega) = \frac{e^2}{6\pi v_F}\int_{-\infty}^\infty d\omega\frac{f(\omega-\mu)-f(\omega+\Omega-\mu)}{\Omega}\int_0^\infty d\epsilon \epsilon^2\left[A_+A_+'+A_-A_-'+2\left(A_+A_-'+A_-A_+'\right)\right].\end{aligned}$$ In the presence of impurities, after performing an impurity average over a random distribution which restores translation invariance, the spectral functions can be written as[@Mahan] $$\begin{aligned} A(\pm\epsilon,\omega) = \frac{1}{\pi}\frac{-\textrm{Im}\Sigma(\omega)}{\left(\omega-\textrm{Re}\Sigma(\omega)\mp\epsilon\right)^2+\left(\textrm{Im}\Sigma(\omega)\right)^2},\end{aligned}$$ where $\Sigma(\omega)$ is the self energy. It is this self energy which in general depends on $\omega$, and carries the information on the detailed properties of the impurity potential associated with a single scattering center. It also differs depending on the strength of the impurity scattering: if the scattering is weak the self energy can be treated in the Born approximation, but if it is strong a full T-matrix approach is required, as in the unitary limit. In this paper we consider only two cases. The first is a constant residual scattering rate. This is appropriate for a weak delta function potential with constant density of states. The second we consider is the case considered by Burkov and Balents[@Burkov:2011kx] of Born scattering. In this case the scattering rate depends on the density of states and is quadratic in energy. Using the above form for the spectral functions one can recover our formulae that contain the impurity scattering rate presented in the main text. Now we check the clean limit at $T = 0$. In that limit we have the intraband piece $$\begin{aligned} \textrm{Re}\sigma_{xx}^{\textrm{D}}(\Omega) =& \frac{e^2}{6\pi v_F\Omega}\int_{\mu-\Omega}^\mu d\omega\int_0^\infty d\epsilon \epsilon^2\left[\delta(\omega-\epsilon)\delta(\omega+\Omega-\epsilon)+\delta(\omega+\epsilon)\delta(\omega+\Omega+\epsilon)\right]\\ \label{eq:a14}& = \frac{e^2\mu^2}{6\pi v_F}\delta(\Omega) = \frac{e^2\mu^2}{3 h \hbar v_F}\delta(\Omega). \end{aligned}$$ We also have the interband piece $$\begin{aligned} \textrm{Re}\sigma_{xx}^{\textrm{IB}}(\Omega) =& \frac{e^2}{3\pi v_F\Omega}\int_{\mu-\Omega}^\mu d\omega\int_0^\infty d\epsilon \epsilon^2\left[\delta(\omega-\epsilon)\delta(\omega+\Omega+\epsilon)+\delta(\omega+\epsilon)\delta(\omega+\Omega-\epsilon)\right]\\ \label{eq:a16}& = \frac{e^2}{24\pi v_F}\Omega\Theta(\Omega-2\mu)= \frac{e^2}{12h \hbar v_F}\Omega\Theta(\Omega-2\mu).\end{aligned}$$ These two results give the well known formula Eq. (\[eq:clean\]). In the final equalities we have restored the factors of $\hbar$ to make the physical units clear.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Image registration, especially the quantification of image similarity, is an important task in image processing. Various approaches for the comparison of two images are discussed in the literature. However, although most of these approaches perform very well in a two image scenario, an extension to a multiple images scenario deserves attention. In this article, we discuss and compare registration methods for multiple images. Our key assumption is, that information about the singular values of a feature matrix of images can be used for alignment. We introduce, discuss and relate three recent approaches from the literature: the Schatten $q$-norm based [$\mathrm{S}q\mathrm{N}$]{} distance measure, a rank based approach, and a feature volume based approach. We also present results for typical applications such as dynamic image sequences or stacks of histological sections. Our results indicate that the [$\mathrm{S}q\mathrm{N}$]{} approach is in fact a suitable distance measure for image registration. Moreover, our examples also indicate that the results obtained by [$\mathrm{S}q\mathrm{N}$]{} are superior to those obtained by its competitors.' author: - Kai Brehmer - Hari Om Aggrawal - Stefan Heldmann - Jan Modersitzki bibliography: - '2019-SSVM.bib' title: 'Variational registration of multiple images with the SVD based [$\mathrm{S}q\mathrm{N}$]{} distance measure' ---
{ "pile_set_name": "ArXiv" }
--- --- IFT-2005/2\ CERN-PH-TH/2005-011\ DESY-05-023\ [****]{} [**Flavour violation in general supergravity**]{}\ $^1$, [**Oleg Lebedev**]{}$^2$ and [**Stefan Pokorski**]{}$^{1,3}$\ $^1$ Institute of Theoretical Physics, Warsaw University, Hoża 69, 00-681, Warsaw, Poland\ $^2$ DESY, Theory Group, Notkestra$\beta$e 85, 22603 Hamburg, Germany\ $^3$ Theoretical Physics Division, Physics Department, CERN, CH-1211 Geneva 23, Switzerland\ We reappraise the flavour changing neutral currents (FCNC) problem in string–derived supergravity models. We overview and classify possible sources of flavour violation and find that the problem often does not arise in classes of models which generate hierarchical Yukawa matrices. In such models, constraints from the $K-$ and $D-$meson systems leave room for substantial flavour non-universality of the soft terms. The current $B-$physics experiments only begin to probe its natural range. Correlations among different observables can allow one to read off the chirality structure of flavour violating sources. We briefly discuss the lepton sector where the problem of FCNC is indeed serious and perhaps points at an additional symmetry or flavour universality. Introduction ============ The purpose of this paper is to classify different sources of flavour violation in supergravity (SUGRA) theories and study the FCNC problem associated with these sources. Concerning the latter, it is important to ask the “right” question, that is, how problematic are the FCNC in models that successfully generate the Yukawa matrices rather than in some “generic” framework. The strategy we pursue in this paper is as follows. First, we classify various sources of flavour violation in general supergravity models. We concentrate on certain “benchmark” textures of the soft terms. Then, using a number of representative Yukawa textures, we evolve the soft terms to low energies and study implications of the generated flavour changing neutral currents. In some cases such as models with flavour violation through the Kähler potential or democratic Yukawa textures, the problem is severe. On the other hand, when flavour structures are due to different localizations of matter fields in the compact space, the FCNC are well suppressed. This is in fact a common situation in string models which produce the Yukawa coupling hierarchy (e.g. heterotic string or intersecting brane models). In string theory, the mechanisms that can generate the fermion mass hierarchy are quite constraining. As a result, the soft supersymmetry (SUSY) breaking terms in superstring derived SUGRA models are of special forms and the problem of FCNC becomes mild or simply disappears. A natural consequence of such models is that the mixing between the second and third generation squarks is expected to be substantial and to lead to effects potentially detectable at $B-$factories. We briefly discuss the lepton sector in which a different picture emerges and the expected flavour violation exceeds the experimental limits. This perhaps signals an additional symmetry or flavour universality of the soft terms, which will be probed further in the upcoming round of $\mu\rightarrow e\gamma$ experiments. Finally, we discuss how correlations among various $B-$physics observables would allow one to pinpoint the source of flavour violation. Classification of flavour–violating sources in supergravity =========================================================== The $N=1$ supergravity Lagrangian is determined by 3 functions: the Kähler potential $K$, the superpotential $W$ and the gauge kinetic function $f$ (see [@Nilles:1983ge] for a review). These are functions of the hidden and observable sector fields. Since the characteristic mass scale of the observable fields $\phi^\alpha$ is much smaller than the Planck scale, one can expand $K$ and $W$ as $$\begin{aligned} &&K=\hat K + \tilde K_{\bar\alpha\beta}\phi^{\ast\bar\alpha}\phi^\beta + \left({1\over2}Z_{\alpha\beta}\phi^\alpha\phi^\beta + {\rm H.c.}\right)+...~,\nonumber\\ &&W= \hat W + {1\over2}\mu_{\alpha\beta}\phi^\alpha\phi^\beta + {1\over 6} {\cal Y}_{\alpha\beta\gamma}\phi^\alpha\phi^\beta\phi^\gamma +...~,\end{aligned}$$ where all parameters are generally functions of the hidden sector fields. Once the hidden sector fields develop (large) vacuum expectation values, these parameters will play a role of various couplings in the observable sector. Generally, the resulting kinetic terms for the observable sector fields are neither canonical nor diagonal, so in order to obtain physical fields, further diagonalization and rescaling are required. The soft SUSY breaking terms are obtained from the general supergravity scalar potential by fixing the gravitino mass $m_{3/2}$ and the VEVs of the hidden sector fields, while sending the Planck mass to infinity, $M_{\rm Pl}\rightarrow\infty$ [@Soni:1983rm; @Kaplunovsky:1993rd]. The fermion Yukawa couplings are then rescaled as $${\cal Y}_{\alpha\beta\gamma}^\prime={\hat W^\ast\over\vert\hat W\vert} e^{\hat K/2}{\cal Y}_{\alpha\beta\gamma} \;.$$ The resulting soft SUSY breaking potential relevant to flavour physics is given by [@Brignole:1997dp] $$\begin{aligned} V_{\rm soft}={\mathpzc m}_{\bar\alpha\beta}^2 \phi^{\ast\bar\alpha} \phi^\beta+\biggl({1\over6}{\cal A}_{\alpha\beta\gamma} \phi^\alpha\phi^\beta\phi^\gamma +{\rm H.c.}\biggr)~, \label{eqn:susybreakingl}\end{aligned}$$ where $$\begin{aligned} {\mathpzc m}_{\bar\alpha\beta}^2&=&(m_{3/2}^2+V_0) \tilde K_{\bar\alpha\beta}- \bar F^{\bar m} F^n (\partial_{\bar m}\partial_n\tilde K_{\bar\alpha\beta} -\tilde K^{\gamma\bar\delta}~\partial_{\bar m} \tilde K_{\bar\alpha\gamma} ~ \partial_{n} \tilde K_{\bar\delta\beta})~,\nonumber\\ {\cal A}_{\alpha\beta\gamma}&=& {\hat W^\ast\over\vert\hat W\vert} e^{\hat K/2}F^m \Bigl[\hat K_m {\cal Y}_{\alpha\beta\gamma} +\partial_m {\cal Y}_{\alpha\beta\gamma}\nonumber\\ &-&\Bigl(\tilde K^{\delta\bar\rho}~ \partial_m \tilde K_{\bar\rho\alpha}~ {\cal Y}_{\delta\beta\gamma} + (\alpha\leftrightarrow\beta) + (\alpha\leftrightarrow\gamma)\Bigr)\Bigr]~. \label{eqn:mA}\end{aligned}$$ Here $V_0$ is the vacuum energy, $\partial_m$ denotes differentiation with respect to the $m$-th hidden sector field ($\hat K_m\equiv\partial_m\hat K$), $F^m$ are the SUSY breaking $F-$terms and $\tilde K^{\gamma\bar\delta}$ is the inverse of the Kähler metric $\tilde K_{\bar\delta\gamma}$. Let us now make our notation more transparent. Greek indices $\alpha,\dots$ run over all MSSM [^1] superfields. However, only fields with appropriate quantum numbers can couple together. For instance, the Kähler mixing is allowed only for fields with the same SM quantum numbers, i.e. only intergenerational mixings of $Q_i$, $U_i$, $D_i$ ($i$=1,2,3) are permitted. The allowed Yukawa couplings in the quark sector are of the type $Q_iD_jH_1$ or $Q_iU_jH_2$. In the Yukawa matrices and the $A-$terms, it is convenient to fix the notation as follows: the first index is to refer to the quark doublets, the second to the quark singlets, and the last to the Higgs field, e.g. $Y_{Q_iD_jH_1}\equiv Y_{ij}^d$. Physical fields are obtained by diagonalizing and rescaling the Kähler metric responsible for kinetic terms of the observable sector fields. The canonically normalized superfields $\varphi^a$ ($a=Q,D,U$) are given by $$\phi^a= H^a\varphi^a$$ (no summation over $a$), with $$H^a= V^a~{\rm diag}\left({1\over\sqrt{\tilde K^a_1}}, {1\over\sqrt{\tilde K^a_2}}, {1\over\sqrt{\tilde K^a_3}} \right)~.$$ Here $V^a$ are 3$\times$3 unitary matrices diagonalizing the appropriate subblocks $\tilde K^a_{\bar\alpha\beta}$ of the Kähler metric and $\tilde K^a_{1-3}$ are the corresponding eigenvalues, $$(V^a)^\dagger \tilde K^a V^a = {\rm diag}\left( \tilde K^a_1, \tilde K^a_2, \tilde K^a_3\right)~.$$ The Yukawa couplings and the soft terms transform correspondingly, such that the seven $physical$ 3$\times$3 flavour structures are $$\begin{aligned} Y^u={1\over\sqrt{\tilde K_{H_2}}}(H^q)^T{\cal Y}^u H^u~,\phantom{aaa} Y^d={1\over\sqrt{\tilde K_{H_1}}}(H^q)^T{\cal Y}^d H^d~,\nonumber\\ A^u={1\over\sqrt{\tilde K_{H_2}}}(H^q)^T{\cal A}^u H^u~,\phantom{aaa} A^d={1\over\sqrt{\tilde K_{H_1}}}(H^q)^T{\cal A}^d H^d~,\end{aligned}$$ $$\begin{aligned} m^2_Q = H^{q\dagger} {\mathpzc m}^2_Q H^q~,\phantom{aa} m^2_U = H^{u\dagger} {\mathpzc m}^2_U H^u~,\phantom{aa} m^2_D = H^{d\dagger} {\mathpzc m}^2_D H^d~,\nonumber\end{aligned}$$ where the factors $1/\sqrt{\tilde K_{H_{1,2}}}$ account for rescaling of the Higgs fields. These structures are the sources of flavour violation in the MSSM. Clearly, flavour dependence in the Yukawa matrices is mandatory, whereas the soft terms serve as additional sources of flavour violation. The underlying reason for these additional sources is the Kähler potential and Yukawa couplings dependence on the hidden sector fields: indeed, as seen from Eq. (\[eqn:mA\]), if the Kähler metric for observables fields and the Yukawa couplings were mere constants, the soft masses would be proportional to the unit matrix in the physical basis and the $A-$terms would be proportional to the Yukawa matrix. Generally, flavour structures of the soft masses and the $A-$terms are independent. In particular, unlike the soft masses, the $A-$terms receive a contribution from $\partial_m{\cal Y}_{\alpha\beta\gamma}$. Thus, even if the Kähler potential is trivial, the $A-$term structure can be quite rich. It is also manifest from Eq. (\[eqn:mA\]) that a non-trivial Kähler potential generally induces flavour violation in both the soft masses and the $A-$terms. Yet, it is conceivable that such flavour violating terms may cancel out in the $A-$terms, due to some deeper dynamical reason. The soft masses would then be the only source of flavour violation (in addition to the Yukawa couplings). In general, the soft breaking terms violate CP. CP violating phases have two sources. Firstly, they are induced by complex SUSY breaking $F-$terms, $F^m$, which also generate CP phases in flavour–independent quantities such as gaugino masses, the $\mu$-term, etc. Secondly, CP phases appear due to complex SUSY preserving quantities such as the Yukawa couplings and the Kähler metric[^2]. Both sources are problematic for phenomenology, which will be discussed below. Clearly, there are many possibilities which have different motivations and distinct phenomenology. Below we classify them. Flavour violation through the Yukawa couplings only --------------------------------------------------- It is possible that flavour dependence does not appear in the soft terms [@Nath:1983fp]. This occurs, for example, when $$F^m \partial_m\tilde K_{\bar\alpha\beta} = F^m \bar F^{\bar n}\partial_m \partial_{\bar n} \tilde K_{\bar\alpha\beta}= F^m \partial_m {\cal Y}_{\alpha\beta\gamma} = 0\;.$$ This essentially means that the hidden sector fields that generate flavour dependence do not break supersymmetry. The most common example of this situation is the dilaton dominated SUSY breaking scenario [@Kaplunovsky:1993rd]. In this case, the only SUSY breaking field is the dilaton $S$ which produces no flavour dependence, $$F^S \neq0,~ F^m =0 ~{\rm{for}} ~m\neq S \;.$$ Then, the soft terms are universal at the string scale and we have a version of the minimal supergravity (minimal SUGRA) model with the sfermion soft terms parametrized by $m^2_0$ and $A_0$: $$(m^2_{Q,U,D})_{ij} = m^2_0 ~\delta_{ij}~,\phantom{aaa} A^{u,d}_{ij} = A_0~Y^{u,d}_{ij}~.$$ Additional flavour violation in $A-$terms only ----------------------------------------------- If, for example, $$F^m \partial_m {\cal Y}_{\alpha\beta\gamma}\not\propto {\cal Y}_{\alpha\beta\gamma} \;, \label{eqn:A-non}$$ but $$F^m \partial_m \tilde K_{\bar\alpha\beta}\propto F^m \bar F^{\bar n}\partial_m \partial_{\bar n}\tilde K_{\bar\alpha\beta} \propto\tilde K_{\bar\alpha\beta} \;,$$ flavour dependence appears only in the $A-$terms [@Abel:2001cv]. This means that the hidden sector fields responsible for flavour dependence of the Kähler potential (if present at all) do not break supersymmetry and, moreover, are different from those generating flavour dependence of the Yukawa couplings. This is a rather common situation in string models. Indeed, the $A-$terms are trilinear parameters and are closely related to the Yukawa couplings, whereas the soft masses are bilinear and more akin to the Kähler potential. Thus, generally, they are not directly related to each other. For example, in the heterotic string models the Yukawa hierarchies are naturally produced if the matter fields are twisted [@Hamidi:1986vh]-[@Casas:1989qx], i.e. localized at special points in the compactified space. In such models, the hidden sector field that enters the Yukawa coupling is the $T-$modulus, ${\cal Y}_{\alpha\beta\gamma}={\cal Y}_{\alpha \beta \gamma}(T)$. Generally, $$F^T\neq0$$ and $\partial_m {\cal Y}_{\alpha \beta \gamma}$ are $not$ proportional to ${\cal Y}_{\alpha\beta\gamma}$, which leads to non-trivial $A-$terms. On the other hand, the Kähler metric is diagonal for twisted states and depends on modular weights of these states: $$\tilde K_{\bar\alpha\beta} =\delta_{\bar\alpha\beta}(T+\bar T)^{n_{\bar\alpha}} ~, \label{eqn:K}$$ The modular weights $n_\alpha$ are constrained by the string selection rules for the Yukawa couplings (see e.g. [@Brax:1994ae]) and are typically generation-independent. The reason is that to obtain a non-trivial structure of the Yukawa couplings and/or CP violation at the renormalizable level often requires the quark fields of different generations to belong to the same twisted sector[^3] (see e.g. [@Casas:1989qx; @Lebedev:2001qg; @Khalil:2001dr]). Consequently, these quark fields have the same modular weights. (In any case, the modular weights can only be $-1$ or $-2$ for non-oscillator states [@Ibanez:1992hc]. Oscillator states usually correspond to SM singlets.) As a result, in these models the Kähler metric is generation–independent. The Yukawa couplings at $\langle T\rangle\sim1$ are given by $${\cal Y}_{\alpha\beta\gamma}\sim e^{-\kappa_{_{\alpha\beta\gamma}}T}$$ with order one coefficients $\kappa_{\alpha\beta\gamma}$, and a nontrivial flavour structure of the $A-$terms results from $$\Delta{\cal A}_{\alpha\beta\gamma}\sim\kappa_{{\alpha\beta\gamma}}~ {\cal Y}_{\alpha\beta\gamma} F^T\;.$$ Analogous results hold for the Yukawa matrices generated by the Froggatt–Nielsen mechanism, in which case $U(1)$ charges play the role of $\kappa_{\alpha\beta\gamma}$. Similarly, in semirealistic brane models (for a review, see [@Blumenhagen:2005mu]) the Kähler metric for matter fields is often diagonal and generation–independent. For instance, replication of families naturally appears in intersecting brane models with different generations located at different intersections of the same branes [@Blumenhagen:2000wh]-[@Cremades:2003qj]. The Kähler metric is then diagonal and depends on the intersection angles (and moduli) [@Lust:2004cx] but is the same for fields of the same type belonging to different generations. Thus, the situation here is similar to the one in the heterotic string case. The resulting soft masses are generation–independent at the string scale (although they can generally be different for up and down squarks, and for left and right squarks), while the $A-$terms can have a rich flavour structure due to Eq. (\[eqn:A-non\]). Thus, the string scale soft breaking lagrangian is parametrized by $$(m^2_{Q,U,D})_{ij} = m^2_{Q,U,D}~\delta_{ij}~,\phantom{aaa}A_{ij}^{u,d} \;, \label{eqn:univ_masses}$$ in addition to flavour-independent parameters. Additional flavour violation in the Kähler potential only --------------------------------------------------------- If the Kähler metric has a non-trivial generation dependence, e.g. $$F^m \partial_m \tilde K_{\bar\alpha\beta}\not\propto \tilde K_{\bar\alpha\beta}~,$$ both the scalar masses and the $A-$terms have a non-trivial flavour structure. This situation can occur, for example, in the heterotic string if the quark field modular weights $n_\alpha$ are generation–dependent (see Eq. (\[eqn:K\])). The non–trivial flavour structures arise then from $$\begin{aligned} &&\Delta{\mathpzc m}_{\bar\alpha\beta}^2 = n_{\bar\alpha} \tilde K_{\bar\alpha\beta} ~{\vert F^T\vert^2\over(T+\bar T)^2} \;,\nonumber\\ &&\Delta{\cal A}_{\alpha\beta\gamma} = -(n_\alpha+n_\beta+n_\gamma) {\cal Y}_{\alpha\beta\gamma}~ {F^T\over T+\bar T} ~{\hat W^\ast\over\vert\hat W\vert}e^{\hat K/2} \;. \label{eqn:m-A}\end{aligned}$$ Here the modular weights are order one integers. Their typical values are $-1$ and $-2$ due to non–oscillator nature of the SM matter. The Kähler metric is the only source of SUSY flavour violation if $$F^m \partial_m{\cal Y}_{\alpha\beta\gamma}\propto {\cal Y}_{\alpha\beta\gamma}~.$$ For example, the contribution $F^m \partial_m {\cal Y}_{\alpha\beta\gamma}$ to the $A-$terms vanishes if the Yukawa structure is generated as in the Froggatt-Nielsen models [@Froggatt:1978nt] by non-renormalizable couplings through the vacuum expectation value of a scalar field $\phi$ which does not break supersymmetry. One then has $${\cal Y}_{\alpha\beta\gamma}={\cal Y}_{\alpha\beta\gamma}(\phi)~, \phantom{aaa}F^\phi=0 \;. \label{eqn:FN}$$ Since in the case discussed above the Kähler metric is diagonal, the soft mass terms are also diagonal, but generically nonuniversal. In contrast, the $A-$terms can be quite complicated due to the generation dependence of the modular weights (see e.g. [@Khalil:1999zn]). More complicated, non-diagonal soft mass terms can be obtained, for instance, in compactifications of the ten dimensional heterotic string on $(T_2/Z_3)^3$ orbifolds, where $T_2/Z_3$ are compact complex spaces (“planes”) obtained by dividing complex tori $T_2$ by a discrete group $Z_3$. In this case, the three generations of untwisted matter superfields can be associated with the three “planes” $T_2/Z_3$ according to their holomorphic indices. The Kähler metric is then non-diagonal and is given by $$\tilde K_{\bar\alpha\beta}=({\rm Re}M_{\bar\alpha\beta})^{-1} \;,$$ where $\alpha$, $\beta=1,2,3$ and the 9 moduli $M_{\bar\alpha\beta}$ parametrize the sizes of the compactification tori and the angles between the three “planes” $T_2/Z_3$ (see e.g. [@Casas:1989qx]). In such models, the soft terms will have a richer non-diagonal flavour structure depending on specific values of the moduli. An interesting possibility is that non-trivial flavour dependence drops out of the $A-$terms, but remains in the soft mass terms. This occurs, for instance, if the vacuum expectation values of the moduli fields take on special values such that $$\begin{aligned} F^m\partial_m \tilde K_{\bar\alpha\beta}\propto \tilde K_{\bar\alpha\beta} ~, \phantom{aaa} F^m\partial_m {\cal Y}_{\alpha\beta\gamma}\propto {\cal Y}_{\alpha\beta\gamma}\nonumber \;,\end{aligned}$$ $$\begin{aligned} F^m \bar F^{\bar n} \partial_m \partial_{\bar n}\tilde K_{\bar\alpha\beta}\not\propto \tilde K_{\bar\alpha\beta} \label{eqn:specialK}\end{aligned}$$ and the flavour dependence appears only through the second derivatives. Then, the $A-$terms are proportional to the Yukawa couplings while the soft masses are general. We conclude that flavour violation through the Kähler potential generally leads to a complicated non-diagonal structure of the soft terms. The squark mass matrices and the $A-$terms are correlated, although this correlation can be far from transparent. The string scale soft breaking lagrangian is parametrized by $$(m^2_{Q,U,D})_{ij}~,\phantom{aa}A_{ij}^{u,d} \;.$$ Two interesting special cases are: (i) diagonal soft masses with general $A-$terms, $$(m^2_{Q,U,D})_{ij} = (m^2_{Q,U,D})_i~\delta_{ij}~, \phantom{aa}A_{ij}^{u,d} \;,$$ as in Eq. (\[eqn:m-A\]), and (ii) general soft masses with universal $A-$terms, $$(m^2_{Q,U,D})_{ij}~,\phantom{aa}A_{ij}^{u,d}=A^{u,d}~Y_{ij}^{u,d} \;, \label{zzz}$$ as in Eq. (\[eqn:specialK\]). Both additional sources present ------------------------------- This is a general case and not much can be said here apart from what already appears in Eq. (\[eqn:mA\]). Special cases have been covered in previous subsections. The string scale soft breaking lagrangian is general and is parametrized by $$(m^2_{Q,U,D})_{ij}~,\phantom{aa}A_{ij}^{u,d} \;.$$ Summary of the textures ----------------------- The above discussion leads us to the following supergravity benchmark textures: $$\begin{aligned} {\rm (A)}:~&&{\rm complete~universality:} \nonumber\\ &&m^2_0,~A_0 \nonumber\\ {\rm (B)}:~&&{\rm generation~independent~ scalar~masses~and~general}~A- {\rm terms:}\nonumber\\ &&m^2_Q,~m^2_U,~m^2_D,~ A_{ij}^u,~A_{ij}^d \nonumber\\ {\rm (C)}:~&&{\rm diagonal~scalar~masses~and~universal}~ A-{\rm terms:}\nonumber\\ &&(m^2_Q)_i,~(m^2_U)_i,~(m^2_D)_i,~ A^u,~A^d \nonumber\\ {\rm (D)}:~&&{\rm diagonal~scalar~masses~and~general}~ A-{\rm terms:}\nonumber\\ &&(m^2_Q)_i,~(m^2_U)_i,~(m^2_D)_i,~ A_{ij}^u,~A_{ij}^d \; \nonumber\\ {\rm (E)}:~&&{\rm general~soft~terms:}\nonumber\\ &&(m^2_Q)_{ij},~(m^2_U)_{ij},~(m^2_D)_{ij},~A_{ij}^u,~A_{ij}^d \;. \nonumber\end{aligned}$$ Textures (D) and (E) also include the possibility that the soft masses and the $A-$terms are correlated, as in the case of flavour violation through the Kähler potential (Eq. (\[eqn:m-A\])). As texture (C) we choose a restricted version of the Ansatz (\[zzz\]) with diagonal squark masses. Here we neglect supergravity radiative corrections which can be considerable in certain models [@Choi:1997de]. In what follows, we will study experimental constraints on the above textures and discuss how to distinguish them. They serve as boundary conditions at high energies and evolve with the energy scale. At the electroweak scale, each texture leads to a specific pattern of the mass insertions (or, more generally, flavour matrices at the interaction vertices). The main features of the resulting patterns can be summarized as follows: 0.2cm (A): very little flavour changing (B): small LL, RR and significant LR, RL flavour changing (C): substantial LL, RR and small LR, RL flavour changing (D): substantial LL, RR, LR, RL flavour changing (E): substantial LL, RR, LR, RL flavour changing 0.2cm Here LL and RR refer to chirality conserving flavor changing transitions in the left– and right–handed sectors, respectively. LR and RL refer to chirality flipping flavor changing transitions. For textures (C) and (D), LL/RR flavour changing results from non–universality of the soft masses at the string scale, that is, their departure from the form (\[eqn:univ\_masses\]). Note that order one non–universality applies to $m^2_i$ rather than $m_i$, which makes a considerable difference for the FCNC analysis. Since different physical processes are sensitive to different types of mass insertions, the above textures (perhaps except for (D) and (E)) are distinguishable given enough experimental information. Low energy effects of the textures {#sec:textures} ================================== In this section, we recall the steps which are necessary to obtain the low energy manifestations of textures (A) to (E) and discuss their main consequences. First of all, as we have already mentioned, each texture evolves with the energy scale and this evolution is described by the renormalization group (RG) equations. Different soft supersymmetry breaking terms evolve differently. The evolution from the GUT scale down to the electroweak scale mainly amounts to adding flavour-universal contributions to the squark mass matrices and the A-terms. These are due to gluino loops and grow with the gluino mass. The main effect of these contributions is that the average squark masse $\tilde M$ increases significantly. This has two important consequences: firstly, the mass insertions decrease as $1/\tilde M^2$ and, secondly, the bounds on the mass insertions relax as $\tilde M$ or $\tilde M^2$. Both of these effects make the FCNC problem milder [@Choudhury:1994pn]. To deal with complicated flavour structures it is convenient to employ the mass insertion approximation [@Hall:1985dx] (although sometimes it may not be precise enough). The mass insertions are defined in the super-CKM basis, i.e. the basis in which the quark mass matrices are diagonal and positive, $$\begin{aligned} &&Y^u \longrightarrow V_L^{u\dagger}~ Y^u~ V_R^{u}={\rm diag}(h_u,h_c,h_t)\;,\nonumber\\ &&Y^d \longrightarrow V_L^{d\dagger} ~Y^d~ V_R^{d}={\rm diag}(h_d,h_s,h_b)\;,\end{aligned}$$ where $h_i$ denote the physical quark Yukawa couplings. To preserve the diagonal flavour structure of the supergauge vertices, the squark fields are rotated in the same fashion as the quark fields. Thus, we have the following superfield transformations: $$\begin{aligned} && \hat U_{L,R} \longrightarrow V^u_{L,R}~ \hat U_{L,R} \;, \nonumber\\ && \hat D_{L,R} \longrightarrow V^d_{L,R}~ \hat D_{L,R} \;.\end{aligned}$$ In this basis, the mass insertions at the electroweak scale are given by $$(\delta^u_{XY})_{ij}\equiv {\left({\cal M}^u_{XY}\right)^2_{ij}\over\tilde M^2}~,\phantom{aa} (\delta^d_{XY})_{ij}\equiv {\left({\cal M}^d_{XY}\right)^2_{ij}\over\tilde M^2}~,\phantom{aa} i\neq j~,$$ where ${\cal M}^{u,d}_{LL}$, ${\cal M}^{u,d}_{RR}$, ${\cal M}^{u,d}_{LR}$ and ${\cal M}^{u,d}_{RL}$ are the $3\times3$ blocks of the full up and down squark mass squared matrices (see e.g. [@Misiak:1997ei]) and $\tilde M$ is the “average” squark mass appropriate for a given mass insertion. The mass insertion approximation works well when the mass insertions are significantly smaller than unity and the splittings among the eigenvalues of the mass matrix are significantly smaller than the eigenvalues themselves. The squark propagator in the mass insertion approximation has an expansion $$\begin{aligned} \langle\tilde q_\alpha\tilde q_\beta^\ast\rangle &=& {i\over\Bigl(k^2{\bf 1}-\tilde M^2{\bf 1}-\delta m^2\Bigr)_{\alpha\beta}} \\ &=&{i\over k^2-\tilde M^2}~{\bf 1}_{\alpha\beta}~+~ {i\over(k^2-\tilde M^2)^2}~ \delta m^2_{\alpha\beta} ~+~ {i\over(k^2-\tilde M^2)^3}~\delta m^2_{\alpha \gamma} ~ \delta m^2_{\gamma\beta}~+~\dots,\nonumber\end{aligned}$$ where $\alpha,\beta$ are indices of the 6$\times$6 mass matrices. If the linear in $\delta m^2_{\alpha\beta}$ term happens to vanish for some ${\alpha,\beta}$, the leading contribution is provided by the “effective” mass insertion $\tilde \delta_{\alpha\beta}\sim \delta_{\alpha\gamma}\delta_{\gamma\beta}$. However, this combination is not completely equivalent to a single mass insertion $\tilde\delta_{\alpha\beta} $ due to a different momentum dependence of the relevant loop integral. It is important to note that since the super-CKM basis is defined only up to a phase, one must also fix the CKM phase convention [@Lebedev:2002wq], which we take to be of the Wolfenstein type. Let us now discuss some features of SUSY flavour structures in the super-CKM basis. $A-$terms in the super-CKM basis -------------------------------- In order to study the LR sector, it is convenient to factor out the Yukawa couplings from the $A-$terms (see Eq. (\[eqn:mA\])), $$A^{u,d}_{ij}\equiv\tilde A^{u,d}_{ij} ~Y^{u,d}_{ij}\;. \label{atilde}$$ Deviations from universality are then encoded in the matrix $\tilde A^{u,d}_{ij}$, which in the universal case has all entries equal: $\tilde A^{u,d}_{ij}=\tilde A^{u,d}$. Unlike $A^{u,d}_{ij}$, matrix elements of $\tilde A^{u,d}_{ij}$ are typically ${\cal O}(1)$ times an overall scale factor [@Abel:2001cv], which is determined by the SUSY breaking scale and is usually of the order of the gravitino mass $m_{3/2}$. For instance, in the heterotic string $Y_{ij}\sim e^{-\alpha_{ij}T}$ with $\alpha_{ij}$ of order one and $T$ being the vacuum expectation value of the $T-$modulus. The non-universality of $\tilde A_{ij}$ is in this case given by $\partial_T\ln Y_{ij}\sim\alpha_{ij}$, or, for non-universal modular weights $n_i$, by $n_i+n_j$. In the case of the Froggatt-Nielsen mechanism, the role of $\alpha_{ij}$’s is played by the $U(1)_X$ charges, with the same conclusion [@Dudas:1995eq]-[@Ross:2002mr]. The amount of non-universality may reduce for democratic Yukawa textures, due to smaller values of the effective $\alpha_{ij}$’s, but this is a model-dependent issue. The $A-$terms undergo RG evolution to low energies with the dominant contribution coming from gluino loops, which has an “aligning” effect similar to those for the squark mass matrices. Upon going over to the super-CKM basis, the $A-$terms transform just as the Yukawa matrices do, $$\begin{aligned} &&A^u_{ij} \longrightarrow\left(V_L^{u\dagger}~A^u~V_R^{u}\right)_{ij} \;,\nonumber\\ &&A^d_{ij} \longrightarrow\left(V_L^{d\dagger}~A^d~V_R^{d}\right)_{ij}\;.\end{aligned}$$ In this basis, $A^{u,d}_{ij}$ are generally non-diagonal and the resulting flavour-changing mass insertions are ($i\not=j$): $$\left(\delta_{LR}^{u,d}\right)_{ij}= {v_{u,d}~A_{ij}^{u,d} \Bigl\vert_{SCKM}\over\tilde M^2} \;,$$ where $v_{u,d}$ are the Higgs VEVs and $\tilde M$ is the appropriate average squark mass at the electroweak scale. To discuss orders of magnitude of various insertions it is instructive to write $$A^u_{ij}\Bigl\vert_{SCKM}= {\rm scale ~factor}\times\Bigl(\alpha_{ij}~h_u + \beta_{ij}~h_c +\gamma_{ij}~ h_t\Bigr) \;, \label{eqn:Asckm}$$ and similarly for the down sector. Here $\alpha_{ij}$, $\beta_{ij}$, $\gamma_{ij}$ are model dependent $\leq{\cal O}(1)$ coefficients parametrizing departure from universality. In the universal case, $$\begin{aligned} && \alpha_{11}=1 {\rm ~~and~~0~otherwise} \;,\nonumber\\ && \beta_{22}=1 {\rm ~~and~~0~otherwise} \;,\label{eqn:abg}\\ && \gamma_{33}=1 {\rm ~~and~~0~otherwise} \;.\nonumber\end{aligned}$$ This representation of the $A-$terms is useful for estimating typical magnitudes of the LR mass insertions. Since $\tilde A_{ij}^{u,d}$ are of order unity, $A_{ij}^{u,d}$ have a similar structure to that of the Yukawa matrices and the misalignment is characterized by deviation of $\alpha_{ij}$, $\beta_{ij}$, $\gamma_{ij}$ from the universal limit (\[eqn:abg\]). This deviation is expected to be small if the Yukawa matrices and the $A-$terms are diagonalized by small angle rotations, which is often the case for hierarchical Yukawa textures. This ceases to be true in the case of democratic textures. An interesting limiting case is [*matrix-factorizable*]{} $A-$terms, i.e. such that they can be written as $$A^{u,d}\equiv\tilde B^{u,d}\cdot Y^{u,d}+Y^{u,d}\cdot\tilde C^{u,d}$$ in the matrix sense with $\tilde B^{u,d}_{ij}$ and $\tilde C^{u,d}_{ij}\simlt{\cal O}(1)$ times a scale factor. In this case, $$A^u\Bigl\vert_{SCKM}= \tilde B^u\Bigl\vert_{SCKM}\cdot ~{\rm diag}(h_u,h_c,h_t) + {\rm diag}(h_u,h_c,h_t)\cdot\tilde C^u\Bigl\vert_{SCKM}\;$$ and similarly for $A^d$. Here $\tilde B^u\Bigl\vert_{SCKM}$ and $\tilde C^u\Bigl\vert_{SCKM}$ have elements of order $\simlt{\cal O}(1)$, again up to an overall scale. This implies, for instance, that the (12) element contains contributions only from $h_u$ and $h_c$, the (11) element – only from $h_u$, etc., and $$A^u_{ij}\Bigl\vert_{SCKM} \leq{\rm scale ~factor}\times({\cal O}(1)~ h_i +{\cal O}(1)~ h_j)\;. \label{eqn:eq1}$$ This limits the magnitude of the LR mass insertions for the first two generations and makes the SUSY FCNC problem less severe. Obviously, this situation occurs in the universal case. Other examples include models in which the Yukawa hierarchy is produced via a Froggatt-Nielsen field and models with non–universal modular weights (Eq.\[eqn:m-A\]), such that $\tilde A_{ij}=a_i+b_j$ (see also [@Kobayashi:2000br]). We also note that the form (\[eqn:eq1\]) is favoured by the absence of charge and color breaking minima in the scalar potential [@Casas:1996de]. Finally, a useful estimate of the mass insertions is obtained by setting the overall scale of the $A-$terms (and the $\mu$–term) to be equal to the average squark mass. Then, for the up sector, $$(\delta_{LR}^u)_{ij}\sim\alpha_{ij} {m_u\over\tilde M} + \beta_{ij}{m_c\over\tilde M} + \gamma_{ij}{m_t\over\tilde M} \;. \label{eqn:LRest}$$ LL and RR sectors in the super–CKM basis ---------------------------------------- Upon going to the super-CKM basis, the LL and RR blocks of the squark mass matrix are rotated as $$\begin{aligned} ({\cal M}^{u,d})^2_{LL} \longrightarrow V_L^{u,d~\dagger}({\cal M}^{u,d})^2_{LL}~V_L^{u,d}\;,\nonumber\\ ({\cal M}^{u,d})^2_{RR} \longrightarrow V_R^{u,d~\dagger}({\cal M}^{u,d})^2_{RR}~V_R^{u,d}\;.\end{aligned}$$ The squark mass squared matrices $({\cal M}^{u,d})^2_{LL}$ and $({\cal M}^{u,d})^2_{RR}$ at the electroweak scale are determined by the original textures (A)-(E) and by the RG evolution. For textures (A)-(D), these matrices are diagonal and therefore remain approximately diagonal after the RG evolution. Flavour violation at the electroweak scale is due to the rotations to the super-CKM basis. We can easily estimate the order of magnitude of the expected effects. Let us assume for a moment that the first two and the third generations do not communicate and consider the (12) block, $${\cal M}^2_{LL}=\left(\matrix{m^2_1 & 0\cr0 & m^2_2}\right) \;. \label{eqn:mLL12}$$ Parametrizing the orthogonal rotation matrix $V_L$ by $\cos\theta$ and $\sin\theta$ in the usual fashion, we get $$(\delta_{LL})_{12}= \cos\theta ~\sin\theta~ {\Delta m^2\over\tilde M^2} \label{deg}$$ in the super-CKM basis, with $\Delta m^2_{}\equiv m^2_1-m^2_2$ and $\tilde M^2\equiv{1\over2}(m^2_1+m^2_2)$. Thus, small mass insertions are obtained for nearly degenerate masses $m_1$ and $m_2$ and/or for a small rotation angle $\theta$. Consider now the case of 3 generations with the first two being degenerate, $${\cal M}^2_{LL} = \left(\matrix{m^2_1 & 0 & 0 \cr 0 & m^2_1 & 0 \cr 0 & 0 & m^2_3}\right) \;.$$ When the rotation matrix $V_L$ is well approximated by the CKM matrix, we have $$\begin{aligned} &&(\delta_{LL})_{12}\sim10^{-4}~e^{i{\cal O}(1)}~ {\Delta m^2_{}\over\tilde M^2} \;,\nonumber\\ &&(\delta_{LL})_{13}\sim10^{-3}~e^{i{\cal O}(1)}~ {\Delta m^2_{}\over\tilde M^2} \;,\\ &&(\delta_{LL})_{23}\sim 10^{-2}~e^{i{\cal O}(10^{-1})}~ {\Delta m^2_{}\over\tilde M^2} \;,\nonumber\end{aligned}$$ where $\Delta m^2_{}\equiv m^2_1-m^2_3$ and $\tilde M$ is the average squark mass. This gives a good idea of the expected magnitude of mass insertions in the case of a small angle rotation, but can be drastically different for textures requiring a large angle rotation. As we will see, FCNC constraints require mass insertions in the LL and RR sectors to be quite small but, nevertheless, leave room for departures from degeneracy of the eigenvalues, particularly for small rotation angles. In the most general texture (E), the flavour off-diagonal entries of the squark mass squared matrices are present already at the high energy scale. Their RG evolution is not important, so at the electroweak scale they remain of similar order of magnitude. Furthermore, barring accidental cancellations, the rotations to the super-CKM basis do not change the qualitative picture. The only important effect is the increase of the flavour–diagonal entries due to the RG running, which reduces the magnitude of the mass insertions. Taking this into account, experimental constraints on the off–diagonal LL and RR mass insertions can be applied directly to the high–energy texture (E). The result is that such insertions have to be small and generic textures (E) are inconsistent with experiment. It is possible that the soft breaking terms and the Yukawa matrices align due to some horizontal symmetry, resulting in suppressed mass insertions. In this paper, we will take a conservative view and will not pursue this option further. Yukawa textures --------------- An important issue to address is dependence of SUSY FCNC on Yukawa textures. To cover both ends of the spectrum, we take a few representative examples with both small and large angle rotations. 0.3cm [**(i). The simplest texture.**]{}\ The simplest texture contains no extra parameters beyond those already present in the CKM matrix and quark masses:[^4] $$\begin{aligned} &&Y^u = {1\over v \sin\beta} ~{\rm diag} (m_u,m_c,m_t) \;,\nonumber\\ &&Y^d = {1\over v \cos\beta} ~V_{CKM}^\dagger~{\rm diag} (m_d,m_s,m_b) \;,\end{aligned}$$ with $v^2=v_u^2+v_d^2$ and $\tan\beta=v_u/v_d $. To diagonalize these matrices, only a left handed rotation $V_{CKM}$ in the down sector is required. 0.3cm [**(ii). A hierarchical texture.**]{}\ Hierarchical textures encode the hierarchy of the quark masses in a transparent manner and arise in various models. For example, they appear in string models due to the exponential dependence of the twisted sector couplings on moduli (for a recent analysis, see [@Ko:2004ic]) or due to a Froggatt-Nielsen type mechanism [@Faraggi:1993su]. For definiteness, we use a set of textures from Ref. [@Chankowski:2005qp]. This is an $SU(5)\times U(1)$ model with order ${\cal O}(1)$ coefficients chosen so that a good fit to the fermion masses and mixings is assured. The $U(1)_X$ charges are $q=\bar u=\bar e=(3,2,0)$ and $\bar d= l=(4,2,2)$. The up-quark Yukawa has a structure $$\begin{aligned} &&Y^u=\left(\matrix{\epsilon^6 & \epsilon^5 & \epsilon^3 \cr \epsilon^5 & \epsilon^4 & \epsilon^2 \cr \epsilon^3 & \epsilon^2 & \epsilon^0}\right).\end{aligned}$$ The down quark Yukawa matrix $Y^d$ is given by $$\begin{aligned} &&Y^d_{ij} = \left(C^{(1)} _{ij}+ \kappa\, C^{(2)}_{ij}\right) \epsilon^{q_i+\bar d_j} ~, \label{eqn:YdYe}\end{aligned}$$ where $\kappa=0.3$, $\epsilon=0.22$ and matrices $C^{(1,2)}$ have a structure of the form $$\begin{aligned} &&C_{1,2}=\left(\matrix{\epsilon^7 & \epsilon^5 & \epsilon^5 \cr \epsilon^6 & \epsilon^4 & \epsilon^4 \cr \epsilon^4 & \epsilon^2 & \epsilon^2}\right)~.\end{aligned}$$ Diagonalization of the Yukawa matrices in this model requires small angle rotations in the left-handed sector and large angle rotations in the right-handed sector. The latter are correlated with the large neutrino mixing angles in GUT models [@Chang:2002mq]. 0.2cm [**(iii). A democratic texture.**]{}\ A strictly democratic texture predicts one massive and two massless quarks in the up and down sectors. Realistic quark masses and mixings can be produced by a perturbation around this texture [@Branco:1990fj; @Fritzsch:1989qm]. Democratic textures can also arise in string models, e.g. when the exponential suppression of the twisted sector couplings is not significant (see e.g. Abel [*et al.*]{} in [@Cremades:2003qj]). In our numerical analysis, we use democratic textures similar to those of Ref. [@Abel:2002zg], $$\begin{aligned} Y^u={m_t\over 3 v \sin\beta} \left(\matrix{1.013&0.987&0.999\cr 0.987&1.013&0.999\cr 0.999&0.999&0.998}\right)\;, \nonumber\end{aligned}$$ $$\begin{aligned} Y^d={m_b\over3v\cos\beta} ~K^\dagger\cdot \left(\matrix{0.987&0.905&0.968\cr 0.903&1.212&1.008\cr 0.967&1.008&1}\right)\cdot K \;, \end{aligned}$$ $$\begin{aligned} K={\rm diag(1,e^{-0.01i},e^{0.01i} )} \;.\nonumber\end{aligned}$$ These matrices are diagonalized by a large angle rotation. Experimental constraints {#sec:expconstr} ======================== Before we present our numerical analysis, in this section we collect experimental constraints on various mass insertions and also discuss qualitatively their implications for textures (A)-(E). 0.5cm [*Electric dipole moments*]{} The electric dipole moments of the neutron and mercury atom (and of the electron, in the lepton sector) are especially sensitive to flavour conserving LR mass insertions for the first generation. In particular, the current bounds $$\begin{aligned} &&\vert d_n\vert<6\times10^{-26} ~e~{\rm cm} \;, \nonumber\\ &&\vert d_{\rm Hg}\vert < 2\times10^{-28} ~e~{\rm cm} \end{aligned}$$ impose the constraints [@Gabbiani:1996hi]-[@Abel:2001vy]: $$\vert{\rm Im}(\delta_{11}^{\{ u,d \}})_{LR}\vert \leq10^{-7}\div10^{-6} \;. \label{edmbound}$$ This implies that the corresponding CP phases of the $\mu-$ and $A-$terms have to be small. Clearly, this SUSY CP problem [@Ellis:1982tk] arises for all supergravity textures, including the flavour universal one. The unwanted CP phases can be suppressed in some special cases, e.g. in the dilaton dominated SUSY breaking scenario with an axionic symmetry [@Lebedev:2002zt] or with the help of the CP phase alignment [@Ibanez:2004iv], but this requires additional assumptions. In the non-universal case, the problem becomes more severe due to additional CP phases which appear in the process of diagonalizing the quark mass matrices and the enhanced magnitude of the mass insertions [@Khalil:2001dr; @Abel:2001cv]. For example, $$(\delta_{LR}^u)_{11}\sim {\tilde A~(m_u+\varepsilon m_c+\varepsilon^\prime m_t)\over\tilde M^2}\;,$$ where $\tilde A\sim\tilde M$ is the “overall” scale of the $A-$terms and $\varepsilon$, $\varepsilon^\prime$ are model-dependent coefficients. This makes it more difficult to satisfy the bound (\[edmbound\]). In this sense, the EDMs prefer some sort of universality, at least in the $A-$terms, although they are problematic in any case. The problem could be partly solved by decoupling the first two sfermion generations (although the third generation still contributes to the EDMs through the Weinberg operator). However, it appears difficult to realize this possibility in specific models. Since the overall scale of the soft terms in SUGRA is given by $m_{3/2}$, a very large splitting between the masses of the first two and the third sfermion generations would require some sort of a singularity in the derivatives of the Kähler potential. In practice, this does not happen and the non-universality is governed, for instance, by order one modular weights (Eq. (\[eqn:K\])). Another option would be to assume a large $m_{3/2}$ and obtain light, as required by naturallness, third generation sfermions via RG evolution to infrared fixed points. However, in this case it would be problematic to get light enough gauginos. Given these difficulties, we will not attempt to resolve the SUSY CP problem and, in what follows, will simply treat the EDMs as a constraint on all of the SUGRA textures. 0.5cm [*Kaon observables*]{} The most important observables in the Kaon system are the Kaon mass splitting, $\Delta M_K$, and the parameters $\epsilon$ and $\epsilon^\prime$ measuring CP violation in Kaon decays, $$\begin{aligned} &&\Delta M_K=M_{K_L}-M_{K_S}\simeq3.5\times10^{-15}{\rm ~GeV}\;, \nonumber\\ &&\epsilon={A(K_L\rightarrow\pi\pi) \over A(K_S \rightarrow\pi\pi)} \simeq 2.3\times10^{-3} \; , \nonumber\\ &&\epsilon^\prime/\epsilon=-{\omega\over\sqrt2\vert\epsilon\vert {\rm Re}A_0}\Bigl({\rm Im}A_0 -{1\over\omega}{\rm Im}A_2\Bigr) \simeq1.9\times10^{-3}\;, \\end{aligned}$$ where $A_{0,2}$ are the $K\rightarrow\pi\pi$ amplitudes for the $\Delta I=1/2,3/2$ transitions and $\omega\equiv{\rm Re }A_2/{\rm Re}A_0\simeq1/22$. These observables place severe constraints on new physics flavour structures. In particular, SUSY contributions mediated by gluinos and squarks generally lead to $|\Delta M_K|$, $|\epsilon|$ and $|\epsilon^\prime/\epsilon|$ orders of magnitude too large. For gluino and squark masses of 500 GeV, and assuming no accidental cancellations between contributions of different mass insertions, the measured $\Delta M_K$ imposes the following bounds on the mass insertions [@Gabbiani:1996hi], $$\begin{aligned} &&\sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^d\right)_{LL}^2\Bigr\vert} < 4\times 10^{-2} \;, \phantom{aaa} \sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^d\right)_{LL} \left(\delta_{12}^d\right)_{RR}\Bigr\vert} < 3\times10^{-3} \;, \nonumber\\ &&\sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^d\right)_{LR}^2 \Bigr\vert}< 4\times 10^{-3} \;,\end{aligned}$$ while the measured value of the $\epsilon$ parameter [@Gabbiani:1996hi] imposes the bounds $$\begin{aligned} &&\sqrt{\Bigl\vert{\rm Im}\left(\delta_{12}^d\right)_{LL}^2\Bigr\vert} <3\times10^{-3} \;, \phantom{aaa} \sqrt{\Bigl\vert{\rm Im}\left(\delta_{12}^d\right)_{LL} \left(\delta_{12}^d\right)_{RR}\Bigr\vert} <2\times10^{-4} \;, \nonumber\\ &&\sqrt{\Bigl\vert{\rm Im}\left(\delta_{12}^d\right)_{LR}^2\Bigr\vert} <4\times 10^{-4} \;.\end{aligned}$$ The bounds on the RR mass insertion are the same as the ones on the LL insertion. All these bounds scale as $\tilde M$ when the squark masses are changed. Furthermore, the chargino-squark loop contributions to $\Delta M_K$ and $\epsilon$ impose analogous (although somewhat weaker) constraints on the up squark sector mass insertions [@Khalil:2001wr]. For $\mu=M_2=350$ GeV and the squark masses of 500 GeV, they read $$\begin{aligned} \sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^u\right)_{LL}^2\Bigr\vert} <1\times 10^{-1}~,\phantom{aaa} \sqrt{\Bigl\vert{\rm Im}\left(\delta_{12}^u \right)_{LL}^2\Bigr\vert} < 1\times 10^{-2}\;.\end{aligned}$$ Similar constraints involving RR and LR mass insertions are much weaker because of the factors $m_q/M_Z$ suppressing the couplings of the right-chiral squarks to charginos and quarks. Finally, for the same gluino and down type squark masses as above, the measured value of the $\epsilon^\prime$ parameter sets the rough bounds [@Gabbiani:1996hi] $$\begin{aligned} && \Bigl\vert{\rm Im}\left(\delta_{12}^d \right)_{LL}\Bigr\vert <5\times10^{-1} ~,\phantom{aaa} \Bigl\vert{\rm Im}\left(\delta_{12}^d\right)_{LR}\Bigr\vert<2\times10^{-5} \;.\end{aligned}$$ In addition, from the chargino–up type squark contributions to $\epsilon^\prime$ one obtains a rather weak limit $\Bigl\vert{\rm Im}\left(\delta_{12}^u\right)_{LL}\Bigr\vert<0.3$ [@Khalil:2001wr] and essentially no bound on other mass insertions. Let us now discuss implications of these constraints. Clearly, at the electroweak scale only little mixing between squarks of the first two generations is allowed. The strongest bounds on the chirality conserving mass insertions come from $\Delta M_K$ and $\epsilon$, while those on the chirality changing mass insertions are due to $\epsilon^\prime$. In the LL and RR sectors, the allowed mass insertions are of order $10^{-2}$ or smaller. This means that the soft masses in the original basis are almost diagonal (barring alignment) and the diagonal entries are almost degenerate at low energies. In the down sector, the most conservative bounds imply that this degeneracy is at a percent level for democratic Yukawa textures and at about 10% level for hierarchical textures. In the up sector, the mass splittings can be larger[^5]. This implies that the LL and RR sectors are to some extent universal at low energies. As we will see in the next section, the RG running from the GUT scale to the electroweak scale has an important “aligning” effect. As a result, constraints on the high energy values of the soft parameters are milder. The situation is very different in the LR sector. Consider the mixing of the first two generation squarks and assume that their mixing with the third generation squarks is small, $\leq$$10^{-2}$. A natural magnitude of the mass insertion is then (see Eq. (\[eqn:LRest\])) $$\left(\delta_{12}^d\right)_{LR}\sim\alpha_{12}{m_d\over \tilde M} + \beta_{12} {m_s\over \tilde M}\simlt{\cal O}(10^{-4})\;, \label{eqn:K-K}$$ where $\alpha_{12}$, $\beta_{12}$ are $\simlt{\cal O}(1)$ model dependent coefficients. It follows that all the bounds except for that from $\epsilon^\prime$ are satisfied automatically. $\epsilon^\prime$ imposes a rather mild constraint on the imaginary part of $A_{12}^d$ (see, e.g. [@Masiero:1999ub]). The same considerations also apply to the up squark sector. This means that order one non–universality is allowed in the $(12)$ block of $\tilde A^d_{ij}$. The above estimate holds in a wide class of models including those with hierarchical textures, matrix–factorizable A–terms, etc. However, it may not apply to the case of democratic textures, which we study below numerically. Finally, we note that there also exist bounds on products of the mass insertions when one goes beyond a single mass insertion approximation. To give an example, chargino–squark penguin diagrams with two mass insertions on the squark line modify, in particular, the effective $Z^0\bar ds$ vertex. The resulting SUSY contribution to BR($K^+\rightarrow\pi^+\nu\bar\nu$) does not exceed its measured value provided $\vert{\rm Re}(\delta_{LR}^u)_{32}^\ast(\delta_{LR}^u)_{31}\vert<0.2$ [@Colangelo:1998pm]. In SUGRA models, this product is expected to be bounded by $(m_t/\tilde M)^2\sim10^{-1}$ and the constraint is satisfied automatically. However, in other scenarios, large SUSY effects in the $Z^0\bar ds$ vertex are possible. They can lead, in particular, to the branching ratio of the $K_L\rightarrow\pi^0\nu\bar\nu$ decay by up to two orders of magnitude larger than the SM prediction [@Colangelo:1998pm]. 0.5cm [*$D^0$-$\bar D^0$ mixing*]{} The experimental bound on the $D^0$-$\bar D^0$ mixing is $$\Delta M_D < 4.8 \times 10^{-14}~ {\rm GeV} \;.$$ For the gluino and squark masses of 500 GeV, the constraints on the up type squark mass insertions read [@Chang:2001ah]: $$\begin{aligned} &&\sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^u\right)_{LL}^2\Bigr\vert} <5\times10^{-2} \;, \phantom{aaa} \sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^u\right)_{LL} \left(\delta_{12}^u\right)_{RR}\Bigr\vert} <1\times10^{-2} \;, \nonumber\\ &&\sqrt{\Bigl\vert{\rm Re}\left(\delta_{12}^u\right)_{LR}^2\Bigr\vert} <2\times10^{-2} \;,\end{aligned}$$ and they scale as $\tilde M$ when the squark masses are changed. The above bounds result from the gluino-up type squark contribution to $\Delta M_D$ and are comparable to the ones stemming from the chargino-squark contributions to the Kaon observables discussed in the previous section. Concerning the LR mass insertions in the up sector, an estimate analogous to Eq. (\[eqn:K-K\]) holds and we reach the same conclusion that order one non-universality in the (12) block of the $A-$terms is allowed. 0.5cm [*$B$ meson observables*]{} The most important constraints on the flavour changing transitions involving the $b$ quark are $$\begin{aligned} &&\Delta M_{B_d} \simeq 3.4\times10^{-13} ~{\rm GeV} \;,\nonumber\\ &&{\rm BR}(\bar B\rightarrow X_s\gamma) \simeq 3.3\times 10^{-4} \;, \\ &&S_{B^0_d\rightarrow\psi K_S} \simeq 0.73 \;,\nonumber\end{aligned}$$ where $S_{B^0_d\rightarrow\psi K_S}$ measures the CP violating asymmetry in the $B^0_d \rightarrow\psi K_S$ decay, proportional to $\Gamma(\bar B^0_d\rightarrow\psi K_S)-\Gamma(B^0_d\rightarrow\psi K_S)$. For the gluino and squark masses of 500 GeV, gluino–squark loop contributions to $\Delta M_{B_d}$ lead to the following constraints on the down type squark mass insertions [@Gabbiani:1996hi]: $$\begin{aligned} &&\sqrt{\Bigl\vert{\rm Re}\left(\delta_{13}^d\right)_{LL}^2\Bigr\vert} <1\times10^{-1} \;, \phantom{aaa} \sqrt{\Bigl\vert{\rm Re}\left(\delta_{13}^d\right)_{LL} \left(\delta_{13}^d\right)_{RR}\Bigr\vert} <2\times10^{-2} \;, \nonumber\\ &&\sqrt{\Bigl\vert{\rm Re}\left(\delta_{13}^d\right)_{LR}^2\Bigr\vert} <3\times10^{-2} \;.\end{aligned}$$ These limits scale as $\tilde M$ when the squark masses are changed. $S_{B^0_d\rightarrow \psi K_S}$ imposes similar constraints on the imaginary parts of the same combinations of the mass insertions [@Gabrielli:2002fr]. As the value of $\Delta M_{B_s}$ is still not bounded from above by experiment, there are no similar limits on the $23$ mass insertions. The experimental value of ${\rm BR}(\bar B\rightarrow X_s\gamma)$ sets a limit on the absolute value of the $(\delta^d_{23})_{LR}$ mass insertion [@Gabbiani:1996hi]: $$\bigl\vert(\delta^d_{23})_{LR}\bigr\vert < 1.5\times10^{-2} \label{eqn:bsg}$$ for the gluino and squark masses of 500 GeV and the bound scales as $\tilde M^2$ when the squark masses are changed. Constraints on the up sector mass insertions are quite weak: the chargino-squark contribution to $\Delta M_{B_d}$ leads to the bound on $\left(\delta_{13}^u\right)_{LL}$ [@Gabrielli:2002fr]: $\vert\left(\delta_{13}^u\right)_{LL}\vert\leq {\cal O} (10^{-1})$. The insertions $\left(\delta_{13}^u\right)_{RR}$, $\left(\delta_{13}^u\right)_{LR}$, $\left(\delta_{13}^u\right)_{RL}$ as well as all $\left(\delta_{23}^u\right)_{XY}$ are essentially unconstrained. In SUGRA models, the LR mass insertions connecting the third generation with the other two are generically of the order $$\begin{aligned} &&\left(\delta_{i3}^d\right)_{LR}\simlt{\cal O}\left({m_b\over \tilde M}\right)\sim10^{-2}\;, \nonumber\\ &&\left(\delta_{i3}^u\right)_{LR}\simlt{\cal O}\left({m_t\over \tilde M}\right)\sim0.1\div1\; \label{deltai3}\end{aligned}$$ for $i=1,2$ and the experimental constraints are satisfied automatically. Thus, the current bounds allow for order one non-universality in the $A-$terms involving the third generation. Some degree of universality is required in the chirality conserving sectors, although the constraints are much weaker than those on the mixing of the first two generations. To conclude this section, the above simple estimates show that order one nonuniversality in the $A-$terms is consistent with the current data, whereas the soft mass terms are required to be essentially diagonal and somewhat degenerate at low energies. As explained earlier, the corresponding constraints on the high energy parameters are significantly weaker due to the aligning effect of the RG evolution. Below we confirm these conclusions numerically. Numerical results ================= In this section, we present results of our numerical analysis for textures (A), (B) and (C) and compare them with the experimental constraints listed in section \[sec:expconstr\]. Barring accidental cancellations, these results also apply to texture (D) which is a combination of textures (B) and (C). For texture (E), no numerical analysis is needed to see that it is inconsistent with the FCNC constraints unless it reduces to one of the special cases (A)-(D). The numerical analysis of this section is necessary to support our qualitative conclusions of section 4. Moreover, it illustrates the dependence of the FCNC constraints on the chosen Yukawa textures and on squark and gluino masses. It is also important to study at a quantitative level the still remaining room for flavour dependence in the Kähler potential and the $A-$terms in supergravity models, so that prospects for further experimental investigations can be assessed. In our analysis, we use the Yukawa textures described in section \[sec:textures\] and the standard 1-loop RG equations for the evolution of the soft terms from the high energy scale down to the electroweak scale. Our results are presented as a function of the high (string) scale values of the parameters and compared with the limits on mass insertions. The limits shown in the plots are properly rescaled to account for the actual values of low energy mass parameters obtained from the RG evolution. 0.3cm [*Texture (A)*]{} Flavour violating effects are very small, especially in the RR sector. For completeness, in Table \[table77\] we provide representative mass insertions generated by the RG running. mass insertion 12 13 23 ---------------- ------------ ------------ ----------- d, LL $10^{-4}$ $10^{-3}$ $10^{-2}$ d, RR $10^{-13}$ $10^{-10}$ $10^{-8}$ d, LR $10^{-8}$ $10^{-5}$ $10^{-5}$ u, LL $10^{-7}$ $10^{-5}$ $10^{-4}$ u, RR $10^{-16}$ $10^{-12}$ $10^{-9}$ u, LR $10^{-11}$ $10^{-8}$ $10^{-6}$ : RG generated mass insertions for $\tan\beta=3$, $m_0=200$ GeV, $A_0=50$ GeV and $M_{1/2}=100$ GeV. \[table77\] [*Texture (B)*]{} The most important effects are in the LR/RL sector. Figures (\[fig:lrydg2\])-(\[fig:lrdemg5\]) display the relevant mass insertions and the experimental bounds for Yukawa textures (i)-(iii), respectively. At the GUT scale the gaugino masses are fixed to 200 GeV for textures (i), (ii), and to 500 GeV for texture (iii). The horizontal axis corresponds to the sfermion masses at the GUT scale: $m_Q^2=m_U^2=m_D^2\equiv m^2$. The mass scale of the $A-$terms is taken as $\tilde A=m/2$ and order one non–universal entries of $\tilde A_{ij}$ are generated randomly. Each panel shows combinations of both the LR and the RL mass insertions, e.g. $|{\rm Re}((\delta^u_{LR})_{12}(\delta^u_{LR})_{12})|^{1/2}$ and $|{\rm Re}((\delta^u_{RL})_{12}(\delta^u_{RL})_{12})|^{1/2}$. This explains the presence of two distinct bands in some of the panels. We see that for the hierarchical Yukawa textures (i) and (ii), the only problematic observables are the EDMs, representing the SUSY CP problem. The constraint stemming from $\epsilon^\prime$ is rather mild. The situation is much worse for the democratic texture (iii), for which all the limits imposed by the Kaon observables are exceeded. 0.3cm [*Texture (C)*]{} Significant LL and RR mass insertions are induced. In Figures (\[fig:yd1kkg2\])-(\[fig:dem1bbg5\]), relevant mass insertions are shown for the GUT scale boundary condition $$m^2_Q=m^2_U=m^2_D= \left(\matrix{m^2_1 & 0 & 0 \cr 0 & m^2 & 0 \cr 0 & 0 & m^2}\right)$$ and Yukawa textures (i)-(iii). As before, the gaugino masses are fixed to 200 GeV for textures (i), (ii), and to 500 GeV for texture (iii). $m_1$ is generated randomly in the range ${1\over2}m\div m$. (In fact, we allow for larger departures from universality than usually exists in typical semirealistic models, see Section 2). Figures (\[fig:yd1kkg2\])-(\[fig:dem1bbg5\]) show that the mass insertions grow with $m$, which is perhaps counterintuitive. The reason for that is the gluino loop renormalization effect which is more important for small values of $m$ [@Choudhury:1994pn]. The low–energy degeneracy parameter $(m^2-m^2_1)/(m^2 + \Delta_{\tilde g} m^2)$, where $\Delta_{\tilde g} m^2$ is induced by the RG running, grows with $m^2$ leading to larger mass insertions. This “aligning” gluino effect is very important and can reduce a mass insertion by up to an order of magnitude for similar squark and gluino masses. For larger gluino masses it is even more important. For $m \sim M_{1/2}$ and small rotation angles (texture (i)), we see that $no$ significant FCNC problem exists. Even if the mass splitting at the GUT scale is of order one, $\delta_{LL,RR}^{12}$ are suppressed by both the gluino loop RG effect and the small rotation angle (see Eq. (\[deg\])). For large rotation angles in the right-handed sector (texture (ii)), the problem of FCNC becomes more acute, mainly due to the simultaneous presence of large LL and RR mass insertions. The most constraining observable here is $\epsilon$. The problem disappears eventually as $M_{1/2}$ increases, yet it still persists for $M_{1/2}=500$ GeV. This is also true for texture (iii) as seen in Figs. (\[fig:dem1kkg5\],\[fig:dem1bbg5\]). An interesting “focusing” effect is seen in Figures (\[fig:yd1bbg2\]) and (\[fig:lav1bbg2\]). The values of $(\delta^d_{23})_{LL}$ are independent of $m_1$ and of the specific Yukawa texture as long as it is diagonalized by a small angle rotation in the left–handed sector. This is because the dominant contribution to $(\delta^d_{23})_{LL}$ is due to the top Yukawa RG effect which depends on the CKM matrix only. Figures (\[fig:yd3kkg2\])-(\[fig:dem3bbg5\]) show our results for the case of degenerate first two generations, $$m^2_Q=m^2_U=m^2_D= \left(\matrix{m^2 & 0 & 0 \cr 0 & m^2 & 0 \cr 0 & 0 & m^2_3}\right) \;,$$ with $m_3$ generated randomly in the range ${1\over2}m\div m$. Again, no problem with FCNC arises for texture (i). For texture (ii), there is some tension with $\epsilon$ due to large rotations in the right–handed sector. The problem becomes milder with increasing gluino mass and dissappears for $M_{1/2}=500$ GeV. The FCNC problem is serious for the Yukawa texture (iii), in which case the limits stemming from the Kaon and $D-$meson observables are exceeded for $M_{1/2}=200$ GeV (the $B^0_d$-$\bar B^0_d$ mixing imposes only a mild constraint). However, no significant FCNC problem exists for heavier gaugino masses, $M_{1/2}=500$ GeV, as shown in Figures (\[fig:dem3kkg5\]) and (\[fig:dem3bbg5\]). For texture (D), similar conclusions can be drawn by combining the results for textures (B) and (C). The main point is that the $A-$term nonuniversality is essentially unconstrained (ignoring the CP problem), while some degeneracy of the diagonal soft masses may be required. Observable Kähler potential flavour violation $A-$term flavour violation -------------------------------------- ------------------------------------ ---------------------------- $\Delta M_K$ problem no problem $\epsilon$ problem no problem $\epsilon'$ no problem no problem BR($K \rightarrow \pi \nu \bar \nu$) no problem no problem $\Delta M_D$ problem no problem $\Delta M_{B_d}$ problem no problem BR($ b \rightarrow s \gamma$) no problem no problem $A_{\rm CP}(B\rightarrow \psi K_s)$ problem no problem : Observables and their sensitivity to the source of flavour violation. Here “Kähler potential flavour violation” refers to misalignment between the Kähler potential and the soft scalar masses of Eq. (\[eqn:mA\]) (texture (E)), while “$A-$term flavour violation” refers to misalignment between the $A-$terms and the Yukawa matrices. The entries indicate whether order one non-universality at the high energy scale is in conflict with the particular observable when hierarchical Yukawa matrices are assumed. \[table1\] Finally, we compare in Table 2 SUSY flavour violation resulting from misalignment between the Kähler potential and the soft scalar masses (texture (E)) with flavour violation resulting from misalignment between the $A-$terms and the Yukawa matrices. Clearly, the former scenario is strongly constrained. This implies that the Kähler potential and the soft scalar mass terms are diagonal (to a good approximation) in the same basis. Generally, there are further constraints on the diagonal entries of the soft mass squared matrices. These constraints strongly depend on the Yukawa textures. If the Yukawa matrices are diagonalized by large angle rotations, the diagonal entries must be degenerate to a large extent. On the other hand, if the Yukawa matrices are diagonalized by rotation matrices similar to the CKM one, order one splittings among the diagonal entries are allowed. Comments on the lepton sector ============================= So far we have been focusing on the squark sector. In the lepton sector, the analysis becomes more involved due to unknown origin and nature of neutrino masses. For example, if they originate from the seesaw mechanism operating at some high scale $M_R<M_{\rm string}$, one should also take into account effects of the additional Yukawa couplings which generate off–diagonal entries in the slepton mass matrices during the RG evolution. We do not undertake such an analysis here. Instead, we only make some qualitative remarks (neglecting the RG effects). The most restrictive observables are the $l_i\rightarrow l_j\gamma$ branching ratios [@Lavignac:2003tk]: $$\begin{aligned} && {\rm BR}(\mu \rightarrow e \gamma) <1.2 \times 10^{-11} \;, \nonumber\\ && {\rm BR}(\tau \rightarrow e \gamma) <2.7 \times 10^{-6} \;, \nonumber\\ && {\rm BR}(\tau \rightarrow \mu \gamma) <5.0 \times 10^{-7} \;.\end{aligned}$$ The corresponding constraints on the mass insertions can be obtained by rescaling the original results of Ref.[@Gabbiani:1996hi], $$\begin{aligned} &&\Bigl\vert\left(\delta^l_{12}\right)_{LL}\Bigr\vert <4\times10^{-3}~,\phantom{aaa} \Bigl\vert\left(\delta^l_{12}\right)_{LR}\Bigr\vert <1\times10^{-6}~,\nonumber\\ &&\Bigl\vert\left(\delta^l_{13}\right)_{LR}\Bigr\vert <2\times10^{-2}~,\phantom{aaa} \Bigl\vert\left(\delta^l_{23}\right)_{LR}\Bigr\vert <1\times10^{-2} \label{eqn:lepton_bounds}\end{aligned}$$ for the photino and slepton masses of 100 GeV. The constraint on $\left(\delta^l_{12}\right)_{RR}$ is the same as that on $\left(\delta^l_{12}\right)_{LL}$, while the other insertions are essentially unconstrained. The bounds (\[eqn:lepton\_bounds\]) immediately tell us that the LL and RR blocks of the slepton mass matrix in the (12) sector are proportional to the unit matrix with good accuracy. As to the LR sector, a simple estimate of $\left(\delta^l_{12}\right)_{LR}$ gives $$(\delta_{12}^l)_{LR} \sim \alpha_{12}{m_e\over\tilde M} + \beta_{12}{m_\mu\over\tilde M} + \gamma_{12}{m_\tau\over\tilde M} \;$$ with $\alpha_{12}$, $\beta_{12}$, $\gamma_{12}\simlt{\cal O}(1)$. The bounds then imply $$\alpha_{12}\simlt 10^{-1}~,\phantom{aa}\beta_{12}\simlt10^{-3}~, \phantom{aa}\gamma_{12}\simlt 10^{-4}\;.$$ Thus, the alignment between the $A-$terms and the lepton Yukawa matrices has to be very precise,[^6] in sharp contrast with the squark sector. This suggests that the charged lepton Yukawa matrix is diagonal at the high energy scale, presumably due to some symmetry such as the lepton number, or that the $A-$terms are universal. The bound on ${\rm BR}(\mu \rightarrow e \gamma) $ is expected to be improved by 3 orders of magnitude at PSI. The above estimates suggest that a non–zero signal is expected. A negative result would mean that the Yukawa matrices and the $A-$terms are extremely well aligned indicating a special symmetry or a special nature of the SUSY breaking. Prospects ========= Different processes are sensitive to different chirality types of mass insertions, making it possible (at least in principle) to distinguish various supergravity textures. Below we give a few examples supporting this point. The estimates (\[deltai3\]) mean that SUSY flavour effects grow with quark masses. In particular, the magnitude of the LR mass insertions involving the third generation is enhanced in SUGRA models, which renders their effect potentially observable. They can significantly affect processes sensitive to LR mass insertions. Such processes are loop suppressed in the Standard Model and are usually due to penguin–type diagrams. Apart from $\epsilon^\prime$ and Br$(\bar B\rightarrow X_s\gamma)$, a good example is the $B^0_d\rightarrow\phi K_S$ decay (for a recent discussion, see [@Kane:2003zi]). In this case, the decay is due to the $b\rightarrow s \bar s s$ transition and the ratio of the SUSY and SM decay amplitudes for squark and gluino masses of 500 GeV is given by [@Khalil:2003bi] $$\left({A^{\rm SUSY}\over A^{\rm SM}}\right)_{B\rightarrow \phi K_S} \simeq 100\times[(\delta_{LR}^d)_{23}+(\delta_{RL}^d)_{23}]+ 0.2\times[(\delta_{LL}^d)_{23}+(\delta_{RR}^d)_{23}] \;.$$ Clearly, for $(\delta_{LR}^d)_{23}\sim 10^{-2}$ consistent with the $b\rightarrow s \gamma$ constraint, the SUSY and SM contributions are of similar magnitude. On the other hand, the chirality conserving insertions contribute far less significantly. Since in the SM the CP asymmetry in this decay coincides with that in the $B^0_d\rightarrow\psi K_S$ decay, it provides a sensitive probe for new physics contributions to the $b\rightarrow s\bar ss$ transition and, in particular, for supersymmetric contributions (see e.g. [@Nir:2002gu]). In the pseudoscalar channel of the $b\rightarrow s \bar s s$ transition, $B^0_d\rightarrow\eta^\prime K_S$, the SM predicts the same (within 5-10%) CP asymmetry, whereas supersymmetry gives [@Khalil:2003bi] $$\left({A^{\rm SUSY}\over A^{\rm SM}}\right)_{B^0_d\rightarrow\eta^\prime K_S} \simeq100\times[(\delta_{LR}^d)_{23}-(\delta_{RL}^d)_{23}]+ 0.2\times[(\delta_{LL}^d)_{23}-(\delta_{RR}^d)_{23}] \;.$$ Thus, in SUGRA models one expects (correlated) deviations in both decays. Another well known example of a process sensitive to LR mass insertions is the $b\rightarrow s \gamma$ transition (see e.g. [@Gabbiani:1996hi]). As is clear from the bound (\[eqn:bsg\]), $(\delta^d_{23})_{LR}$ of order $10^{-2}$, which is natural in SUGRA models, is sufficient to produce significant deviations from the SM prediction. On the other hand, to have a similar effect from the LL sector, a large $(\delta^d_{23})_{LL} = {\cal O}(1)$ would be required. We note that the observation of a direct CP asymmetry in the $b\rightarrow s \gamma$ transition of order few percent would be a clean signal of new physics, since the well controlled SM prediction yields the asymmetry of less than one percent. An example of a process particularly sensitive to LL and RR insertions is the $B^0_s$-$\bar B^0_s$ mixing. For moderate $\tan\beta$, the SUSY to SM ratio of the mixing amplitudes is [@Ball:2003se] $$\begin{aligned} \left({M^{\rm SUSY}_{12}\over M^{\rm SM}_{12}}\right)_{B^0_s-\bar B^0_s} &\simeq& 1\times\left[(\delta_{LL}^d)_{23}^2+(\delta_{RR}^d)_{23}^2\right] + 30\times\left[(\delta_{LR}^d)_{23}^2+(\delta_{RL}^d)_{23}^2\right] \nonumber\\ &-& 45\times\left[(\delta_{LR}^d)_{23}(\delta_{RL}^d)_{23}\right] -175\times\left[(\delta_{LL}^d)_{23}(\delta_{RR}^d)_{23}\right] \end{aligned}$$ for gluino and squark masses of 500 GeV. At large $\tan\beta$, other contributions become important [@Buras:2002wq]. Given the limits on $(\delta_{LR}^d)_{23}$ from $b\rightarrow s\gamma$, the contribution of chirality changing mass insertions is negligible, while even a small chirality conserving insertions $(\delta_{LL}^d)_{23}$ and $(\delta_{RR}^d)_{23}$ of order ${\cal O}(10^{-1})$ can induce a large departure of $\Delta M_{B_s}$ from the value predicted in the SM. Essential for identifying the sources of flavour violation are correlations among different observables. To give an example, suppose that significant deviations from the SM predictions are found in the $B^0_d\rightarrow\phi K_S$ and $B^0_d\rightarrow\eta^\prime K_S$ decays. If they are due to LR/RL mass insertions, no deviation is expected in the $B^0_s$-$\bar B^0_s$ mixing. On the other hand, if the “anomaly” is due to LL/RR mass insertions, the $B^0_s$-$\bar B^0_s$ mixing should also be significantly affected. Thus, different SUGRA textures lead to different signatures. In principle, a more sophisticated network of correlations, including various other observables, can and should be developed. This, however, is beyond the scope of the present paper. Significant SUSY effects can also be present in the top quark decays. The magnitude of $(\delta_{LR}^u)_{i3}$ mass insertions is enhanced due to a large top Yukawa coupling. Since BR($t\rightarrow c \gamma$) is particularly sensitive to LR insertions, BR($t\rightarrow c \gamma$)$\sim 10^{-6} (\delta_{LR}^u)_{23}^2 $ [@Delepine:2004hr], large departures from the SM prediction $\sim 10^{-13}$ are expected. To conclude, we see that the current $B-$physics experiments are beginning to probe a natural range of supergravity non–universality in the A–terms. Some processes can further probe non–universality in the soft scalar masses, yet in this case it would be difficult to define a “natural” range due to larger model-dependence. Correlations among various observables can allow one to identify the source of flavour violation. Conclusions =========== In this paper, we have presented a classification and analysis of flavour violating sources in general supergravity models. The current flavour physics data lead us to the following conclusions: 1. Flavour violation through the Kähler potential is disfavored, but room for flavour non–universality remains if the Kähler metric is diagonal. This often occurs in string models in which the Kähler potential is protected by string selection rules. 2. Departures from universality of order unity are allowed in the squark sector $A-$terms. Such departures are expected in typical string models. 3. $A-$terms in the charged slepton sector must be very well aligned with the lepton Yukawa matrix, which points at either a diagonal Yukawa matrix or universal $A-$terms. 4. A common and rather serious problem in supergravity models is the SUSY CP problem. 5. The FCNC problem depends strongly on the Yukawa texture and is much milder for hierarchical Yukawa matrices. For diagonal squark masses, the problem essentially disappears if the Yukawa matrices are diagonalized by small angle rotations or if the CKM matrix derives entirely from rotations of the up type quarks. Thus, there exist varieties of textures ensuring sufficient suppression of FCNC. 6. Importance of SUSY flavour effects grows with quark masses. 7. Current $B-$physics experiments are beginning to probe a natural range of flavour non–universality in SUGRA models. Correlations among different observables can allow one to identify the source of flavour violation. Texture Hierarchical Yukawas Democratic Yukawas Mode --------- ---------------------- -------------------- ------------------------------------------------- (A) no problem no problem ? (B) no problem problem $b\rightarrow s\gamma$, $b\rightarrow s\bar ss$ (C) no serious problem problem $B^0_s$-$\bar B^0_s$ (D) no serious problem problem see (B), (C) (E) problem problem see (B), (C) : Compatibility of the supergravity soft term textures with the Yukawa textures. Possible detection modes are also indicated. \[table2\] It is important to note the difference between the flavour structure of the Standard Model and that of supergravity soft terms. In the former, we encounter a hierarchical pattern of the Yukawa couplings. In supergravity, the soft terms are logarithmic derivatives of hierarchical quantities. As a result, $m^2_i$ and $\tilde A_{ij}$ are expected to be of the same order, in sharp contrast with the Yukawa couplings. Finally, from the point of view of avoiding the FCNC problem, Ansätze (A) and (B) are clearly preferred. Ansatz (B) represents a rather typical prediction of string models (e.g. heterotic string, intersecting branes, etc.). Textures (C) and (D) are also allowed under the condition of some degeneracy in the (12) block. Texture (E) is clearly disfavored. These conclusions are summarized in Table 3. The main message is that the supergravity textures are not necessarily restricted to the universal one and can be quite rich, depending on the mechanism generating the Yukawa matrices. These flavour structures can be probed experimentally. 0.3cm [**Acknowledgements**]{}\ This work has been supported in part by the RTN European Program MRTN-CT-2004-503369. P.H.Ch. and S.P. were supported by the Polish State Committee for Scientific Research Grants 2 P03B 040 24 for 2003-2005 and 2 P03B 129 24 for 2003-2005, respectively. P.H.Ch. would like to thank the CERN Theory Group for hospitality during the completion of this work. This work was started during S.P.’s visit to the University of Hamburg. The visit was possible owing to “Forschung–Preis” of the A. von Humboldt Foundation. S.P. thanks Jan Louis for his hospitality. This collaboration was stimulated by an ENTApP sponsored visitor’s programme on dark matter at CERN, 17 January - 4 February 2005. 0.3cm [99]{} H. P. Nilles, Phys. Rept.  [**110**]{} (1984) 1; S. Weinberg, “The quantum theory of fields. Vol. 3: Supersymmetry,” Cambridge University Press, Cambridge, 2000. S. K. Soni and H. A. Weldon, Phys. Lett. B [**126**]{}, 215 (1983). V. S. Kaplunovsky and J. Louis, Phys. Lett. B [**306**]{}, 269 (1993). A. Brignole, L. E. Ibanez and C. Munoz, “Soft supersymmetry-breaking terms from supergravity and superstring models,” hep-ph/9707209. D. J. H. Chung, L. L. Everett, G. L. Kane, S. F. King, J. Lykken and L. T. Wang, “The soft supersymmetry-breaking Lagrangian: Theory and applications,” hep-ph/0312378. O. Lebedev, Phys. Rev. D [**67**]{}, 015013 (2003). P. Nath, R. Arnowitt and A. H. Chamseddine, “Applied N=1 Supergravity,” NUB-2613, [*Lectures given at Summer Workshop on Particle Physics, Trieste, Italy, Jun 20 - Jul 29, 1983*]{}. S. Abel, S. Khalil and O. Lebedev, Phys. Rev. Lett.  [**89**]{}, 121601 (2002). S. Hamidi and C. Vafa, Nucl. Phys. B [**279**]{}, 465 (1987); L. J. Dixon, D. Friedan, E. J. Martinec and S. H. Shenker, Nucl. Phys. B [**282**]{}, 13 (1987). L. E. Ibanez, Phys. Lett. B [**181**]{}, 269 (1986). J. A. Casas and C. Munoz, Nucl. Phys. B [**332**]{}, 189 (1990) \[Erratum-ibid. B [**340**]{}, 280 (1990)\]. P. Brax, U. Ellwanger and C. A. Savoy, Phys. Lett. B [**347**]{}, 269 (1995). O. Lebedev, Phys. Lett. B [**521**]{}, 71 (2001). S. Khalil, O. Lebedev and S. Morris, Phys. Rev. D [**65**]{}, 115014 (2002). L. E. Ibanez and D. Lust, Nucl. Phys. B [**382**]{}, 305 (1992). R. Blumenhagen, M. Cvetic, P. Langacker and G. Shiu, “Toward Realistic Intersecting D-Brane Models,” hep-th/0502005. R. Blumenhagen, L. Goerlich, B. Kors and D. Lust, JHEP [**0010**]{}, 006 (2000). G. Aldazabal, S. Franco, L. E. Ibanez, R. Rabadan and A. M. Uranga, J. Math. Phys.  [**42**]{}, 3103 (2001). D. Cremades, L. E. Ibanez and F. Marchesano, JHEP [**0307**]{}, 038 (2003). For realistic examples, see S. A. Abel, O. Lebedev and J. Santiago, Nucl. Phys. B [**696**]{}, 141 (2004). D. Lust, P. Mayr, R. Richter and S. Stieberger, Nucl. Phys. B [**696**]{}, 205 (2004). C. D. Froggatt and H. B. Nielsen, Nucl. Phys. B [**147**]{}, 277 (1979). S. Khalil, T. Kobayashi and A. Masiero, Phys. Rev. D [**60**]{}, 075003 (1999). K. Choi, J. S. Lee and C. Munoz, Phys. Rev. Lett.  [**80**]{}, 3686 (1998). A. Brignole, L. E. Ibanez and C. Munoz, Nucl. Phys. B [**422**]{}, 125 (1994) \[Erratum-ibid. B [**436**]{}, 747 (1995)\]; D. Choudhury, F. Eberlein, A. Konig, J. Louis and S. Pokorski, Phys. Lett. B [**342**]{}, 180 (1995); P. Brax and C. A. Savoy, Nucl. Phys. B [**447**]{}, 227 (1995). L. J. Hall, V. A. Kostelecky and S. Raby, Nucl. Phys. B [**267**]{}, 415 (1986). M. Misiak, S. Pokorski and J. Rosiek, Adv. Ser. Direct. High Energy Phys.  [**15**]{}, 795 (1998). E. Dudas, S. Pokorski and C. A. Savoy, Phys. Lett. B [**369**]{}, 255 (1996); E. Dudas, C. Grojean, S. Pokorski and C. A. Savoy, Nucl. Phys. B [**481**]{}, 85 (1996). S. A. Abel and G. Servant, Nucl. Phys. B [**597**]{}, 3 (2001). G. G. Ross and O. Vives, Phys. Rev. D [**67**]{}, 095013 (2003). T. Kobayashi and O. Vives, Phys. Lett. B [**506**]{}, 323 (2001). J. A. Casas and S. Dimopoulos, Phys. Lett. B [**387**]{}, 107 (1996). For a recent study, see P. Ko, T. Kobayashi and J. h. Park, Phys. Lett. B [**598**]{}, 263 (2004). A. E. Faraggi and E. Halyo, Nucl. Phys. B [**416**]{}, 63 (1994); T. Kobayashi, Phys. Lett. B [**358**]{}, 253 (1995). P. H. Chankowski, K. Kowalska, S. Lavignac and S. Pokorski, “Update on fermion mass models with an anomalous horizontal U(1) symmetry,” hep-ph/0501071. D. Chang, A. Masiero and H. Murayama, Phys. Rev. D [**67**]{}, 075013 (2003). G. C. Branco, J. I. Silva-Marcos and M. N. Rebelo, Phys. Lett. B [**237**]{}, 446 (1990). H. Fritzsch and J. Plankl, Phys. Lett. B [**237**]{}, 451 (1990). S. Abel, G. C. Branco and S. Khalil, Phys. Lett. B [**569**]{}, 14 (2003). F. Gabbiani, E. Gabrielli, A. Masiero and L. Silvestrini, Nucl. Phys. B [**477**]{}, 321 (1996); J. S. Hagelin, S. Kelley and T. Tanaka, Nucl. Phys. B [**415**]{}, 293 (1994). T. Falk, K. A. Olive, M. Pospelov and R. Roiban, Nucl. Phys. B [**560**]{}, 3 (1999). S. Abel, S. Khalil and O. Lebedev, Nucl. Phys. B [**606**]{}, 151 (2001). J. R. Ellis, S. Ferrara and D. V. Nanopoulos, Phys. Lett. B [**114**]{}, 231 (1982); W. Buchmuller and D. Wyler, Phys. Lett. B [**121**]{}, 321 (1983); J. Polchinski and M. B. Wise, Phys. Lett. B [**125**]{}, 393 (1983). O. Lebedev and S. Morris, JHEP [**0208**]{}, 007 (2002). L. E. Ibanez, “The fluxed MSSM,” hep-ph/0408064. S. Khalil and O. Lebedev, Phys. Lett. B [**515**]{}, 387 (2001). A. Masiero and H. Murayama, Phys. Rev. Lett.  [**83**]{}, 907 (1999). G. Colangelo and G. Isidori, JHEP [**9809**]{}, 009 (1998). D. Chang, W. F. Chang, W. Y. Keung, N. Sinha and R. Sinha, Phys. Rev. D [**65**]{}, 055010 (2002). E. Gabrielli and S. Khalil, Phys. Rev. D [**67**]{}, 015008 (2003). For an update and prospects, see e.g. S. Lavignac, “Flavour and CP violation in the lepton sector and new physics,” eConf [**C030603**]{}, VEN04 (2003), hep-ph/0312309. G. L. Kane, P. Ko, H. b. Wang, C. Kolda, J. h. Park and L. T. Wang, Phys. Rev. Lett.  [**90**]{}, 141803 (2003). S. Khalil and E. Kou, Phys. Rev. Lett.  [**91**]{}, 241602 (2003). Y. Nir, “CP violation: The CKM matrix and new physics,” Nucl. Phys. Proc. Suppl.  [**117**]{}, 111 (2003); Y. Grossman, “Footprints of new physics in the B system,” AIP Conf. Proc.  [**722**]{} (2004) 255. P. Ball, S. Khalil and E. Kou, Phys. Rev. D [**69**]{}, 115011 (2004). A. J. Buras, P. H. Chankowski, J. Rosiek and L. Sławianowska, Phys. Lett. B [**546**]{} (2002) 96; Nucl. Phys. B [**659**]{} (2003) 3; P. H. Chankowski and J. Rosiek, Acta Phys. Polon. B [**33**]{} (2002) 2329; A. Dedes, Mod. Phys. Lett. A [**18**]{} (2003) 2627. D. Delepine and S. Khalil, Phys. Lett. B [**599**]{}, 62 (2004). [^1]: For a review, see [@Chung:2003fi]. [^2]: The reparametrization invariant measures of CP violation are given by quantities of the type Arg$({\cal A}_{\alpha\beta\gamma}^\ast{\cal Y}_{\alpha \beta \gamma})$ [@Lebedev:2002wq]. These can be non-vanishing even if all $F^m$ are real. [^3]: This is always true in prime orbifolds since there is only one twisted sector. [^4]: An alternative texture of this sort, $Y^u\propto V_{CKM}^\dagger{\rm diag}~(m_u,m_c,m_t)$, $Y^d\propto{\rm diag}(m_d,m_s,m_b)$ would lead to smaller FCNC effects in the down type quark sector, which is constrained by experiment stronger than the up sector. [^5]: If the CKM matrix is entirely due to the up sector and the Yukawa textures are hierarchical, the constraints are particularly weak and no significant degeneracy is required. This, however, appears to be a rather special case. [^6]: We note that the constraints on $\left(\delta^l_{13} \right)_{LR}$ and $\left(\delta^l_{23} \right)_{LR}$ are trivially satisfied since their magnitude is bounded by $m_\tau/\tilde m$.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We interpret the “explicit formula” in the sense of analytic number theory for the zeta function of an ordinary abelian variety of dimension $g$ over a finite field as a transversal index theorem on a Riemannian foliated space of dimension $2g+1$. This generalizes a work of Deninger for elliptic curves.' address: - 'Mathematisches Institut, Westfälische Wilhelms-Universität, Einsteinstrasse 62, 48149 Münster, Germany.' - 'Institut mathématique de Jussieu-Paris Rive Gauche, UMR 7586, Bâtiment Sophie Germain, Case 7012, 75205 Paris Cedex 13.' author: - Ouidad Filali - Francesco Lemma title: Abelian varieties and transversal index theorems --- Introduction ============ In the search for an understanding of the properties of zeta and $L$-functions, Deninger has developed analogies between the theory of dynamical systems on certain foliated spaces and arithmetic geometry [@deninger2], [@deninger1], [@deninger3]. For example, to a $d$-dimensional regular scheme $\mathcal{X}$ of finite type over $\operatorname{Spec}\operatorname{\mathbb{Z}}$ should correspond a triple $(X, \mathcal{F}, \phi^t)$ where $X$ is a certain $(2d+1)$-dimensional space endowed with a one codimensional lamination $\mathcal{F}$ and a flow $\phi^t$ whose orbits are transversal to the leaves and such that the closed points $x \in \mathcal{X}$ should correspond to the closed orbits $\gamma$ of $\phi^t$. In the simple case where $\mathcal{X}$ is an elliptic curve over a finite field, Deninger constructed such a three dimensional foliated dynamical system $(X, \mathcal{F}, \phi^t)$ and interpreted the explicit formula for the zeta function of the elliptic curve as a transversal index theorem on $(X, \mathcal{F}, \phi^t)$ [@deninger]. In this paper, we will be concerned in generalizing Deninger’s construction to higher dimension.\ Let $A_0/k$ be an abelian variety of dimension $g$ over the finite field $k=\operatorname{\mathbb{F}}_q$, let $|A_0|$ the set of closed points of $A_0$ and, for $x \in |A_0|$, let $\deg(x)=[k(x):\operatorname{\mathbb{F}}_q]$ be the degree of the residue field of $x$. The zeta function $$\zeta_{A_0}(s)=\prod_{x \in |A_0|} \frac{1}{1-q^{-s\deg(x)}}$$ converges for $\operatorname{Re}s > g$. Let $l$ be a prime different from the characteristic of $k$, let $T_l(A_0)$ be the $l$-adic Tate module of $A_0$, a free $\operatorname{\mathbb{Z}}_l$-module of finite rank $2g$, and let $\mu_1, \ldots, \mu_g$ be the eigenvalues of the arithmetic Frobenius endomorphism of $T_l(A_0)$. Then we have $$\zeta_{A_0}(s)=\prod_{j=0}^{2g} P_j(q^{-s})^{(-1)^j}$$ where $ P_j(X)=\prod(1-\mu_{i_1}\ldots \mu_{i_j}X)$, the product being taken over all $j$-uples $(i_1, \ldots, i_j)$ such that $1 \leq i_1 < \ldots < i_j \leq 2g$. Moreover, the functional equation $$\zeta_{A_0}(s)=\pm \zeta_{A_0}(g-s)$$ is satisfied. The explicit formula we are interested in is the following. Given a test function $\alpha \in \mathcal{C}^{\infty}_0(\operatorname{\mathbb{R}})$ we have $$\label{EF} \sum_{j=1}^{2g} \sum_{\rho_j}(-1)^{j} \Phi(\rho_j) = \log q \sum_{x \in |A_0|}\deg(x) \sum_{k \geq 1} \alpha(k \deg(x) \log q)$$ $$\begin{aligned} \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,+ \log q \sum_{x \in |A_0|}\deg(x) \sum_{k \leq -1} q^{kg\deg(x)} \alpha(k \deg(x) \log q)\end{aligned}$$ where $\Phi(s)$ is the function defined by the integral $$\Phi(s)=\int_{\operatorname{\mathbb{R}}}e^{ts} \alpha(t) dt$$ and where the sums $\sum_{\rho_j}$ in the left hand term are indexed by the zeroes of $P_j(q^{-s})$. This equality can be proved in a similar, and in fact easier way as in the standard case [@barner] 4. Under the assumption that $A_0$ is ordinary, we are going to interpret the explicit formula as a transversal index theorem on a $(2g+1)$-dimensional Riemannian foliated space $S(A_0)$. See Cor. \[main\] for a precise statement.Transversal index theory is concerned with differential operators or complexes of such operators, which are elliptic in the directions transversal to the orbits of an action by a Lie group, which in our case will simply be $\operatorname{\mathbb{R}}^\times_+$. For a short introduction to such a theory, the reader is referred to [@deninger] 2.\ **Acknowledgements.** The main ideas of the present article have their origin in the 2007 PhD thesis of the first named author under supervision of Christopher Deninger [@filali]. Recently, the second named author worked on the subject idependently and brought a new idea. The second named author would like to thank Eric Urban for his invitation to Columbia University, where part of this work has been done and Michael Harris for support. Finally, we would like to thank Christopher Deninger, Frans Oort and Felipe Voloch for enlightening correspondence. Preliminaries on abelian varieties {#preliminaires} ================================== In this section, we would like to recall some classical facts about abelian varieties, the Frobenius endomorphism and ordinarity.\ Let $A/k$ be an abelian variety over a field $k$, of dimension $g>0$ and let $\phi \in {\mathrm}{End}(A)$ be an endomorphism. There exists a polynomial $P_\phi \in \operatorname{\mathbb{Z}}[X]$ of degree $2g$ such that for any $t \in \operatorname{\mathbb{Z}}$ we have $P_\phi(t)=\deg(\phi-[t])$ where $[t]$ denotes the endomorphism of $A$ given by multiplication by $t$ (see [@cornell-silvermann] p. 125). For any prime number $l$, denote by $T_l(A)$ the $l$-adic Tate module of $A$. It follows from [@mumford] IV §19 Thm. 4 that if $l \neq {\mathrm}{char}(k)$ the polynomial $P_\phi$ is the characteristic polynomial of the endomorphism $T_l(\phi)$ induced by $\phi$ on $T_l(A) \otimes_{\operatorname{\mathbb{Z}}_l} \operatorname{\mathbb{Q}}_l$. Assume that $k={\mathbb}{F}_q$ is a finite field with $q$ elements and that $\phi$ is the $q$-th power Frobenius endomorphism, then $\deg(\phi)=q^g$ and hence $$\label{det} \det T_l(\phi)=q^g$$ for any prime $l$ different from the characteristic of $k$.\ Let $A_0/k$ be an abelian variety of dimension $g$ over a finite field $k=\operatorname{\mathbb{F}}_q$ of characteristic $p$. For any integer $n \geq 0$, let us denote by $G_n$ the kernel of the multiplication by $p^n$ on $A_0$. One says that $A_0$ is ordinary if the following equivalent conditions are satisfied:\ 1. $G_1(\overline{k}) \simeq (\operatorname{\mathbb{Z}}/p\operatorname{\mathbb{Z}})^g$,\ 2. $G_n(\overline{k}) \simeq (\operatorname{\mathbb{Z}}/p^n\operatorname{\mathbb{Z}})^g$ for all $n \geq 0$.\ A proof of the equivalence of these two conditions can be found in [@mumford] §15 “the $p$-rank”. Let $G_n^0$ be the connected component of the identity in $G_n$ and let $G_n^{et}$ be the largest étale quotient of $G_n$. Then, there is an exact sequence $$0 \longrightarrow G_n^0 \longrightarrow G_n \longrightarrow G_n^{et} \longrightarrow 0$$ which is split by [@tate] (3.7) IV. If $A_0/k$ is ordinary, then by the Serre-Tate theorem [@serre-tate], it admits a canonical lift $\mathcal{A}$ over the ring of Witt vectors $W$ of $k$. This canonical lift is characterized by the following equivalent conditions (see [@deligne] 3):\ 1. The $p$-divisible group of $\mathcal{A}$ is the product of the $p$-divisible groups lifting the connected component of the identity and the largest étale quotient of the $p$-divisible group of $A$,\ 2. Every endomorphism of $A_0$ lifts uniquely to $\mathcal{A}$. The Riemannian foliated dynamical system $(S(A_0), \mathcal{F}, \phi^t)$ ======================================================================== Let $A_0$ be a $g$-dimensionnal ordinary abelian variety over $k=\operatorname{\mathbb{F}}_q$, let $p$ be the characteristic of $k$ and let $\phi_0: A_0 \longrightarrow A_0$ be its $q$-th power Frobenius endomorphism. Let $\mathcal{A}$ be the Serre-Tate canonical lift of $A_0$ to $W$ and let $\phi: \mathcal{A} \longrightarrow \mathcal{A}$ be the endomorphism lifting $\phi_0$. Let $K=W[1/p]$ and let $A=\mathcal{A} \otimes_W K$ be the generic fibre of $\mathcal{A}$. In what follows, we fix an embedding $\iota: K \longrightarrow \operatorname{\mathbb{C}}$ so that we can consider the complex analytic abelian variety $A(\operatorname{\mathbb{C}})$. Let us denote by $\Gamma$ the first integral homology group $\Gamma=H_1(A(\operatorname{\mathbb{C}}), \operatorname{\mathbb{Z}})$, which is a free $\operatorname{\mathbb{Z}}$-module of rank $2g$. Via the natural map $$\Theta_{\iota}: {\mathrm}{End}_K(A) \longrightarrow {\mathrm}{End}_{\operatorname{\mathbb{C}}}(A(\operatorname{\mathbb{C}})),$$ the Frobenius lift $\phi \otimes K$ induces the endomorphism $$\xi=\Theta_{\iota}(\phi \otimes K)_* \in {\mathrm}{End}_{\operatorname{\mathbb{Z}}}(\Gamma).$$ For any prime $l \neq p$, we have a functorial isomorphism $\Gamma \otimes \operatorname{\mathbb{Z}}_l = T_l(A)$ of $\operatorname{\mathbb{Z}}_l$-modules. Hence, it follows from (\[det\]) that $$\label{determinant} \det(\xi)=q^g$$ As the $\operatorname{\mathbb{R}}$-vector space $\Gamma \otimes \operatorname{\mathbb{R}}$ is identified to the Lie algebra of $A(\operatorname{\mathbb{C}})$, it has a complex structure for which the endomorphism $\xi \otimes {\mathrm}{id}_{\operatorname{\mathbb{R}}}$ is $\operatorname{\mathbb{C}}$-linear. In what follows, we will identify the Lie algebra $\Gamma \otimes \operatorname{\mathbb{R}}$ to $\operatorname{\mathbb{C}}^g$ and we denote by $\xi_{\operatorname{\mathbb{R}}} \in {\mathrm}{End}_{\operatorname{\mathbb{C}}}(\operatorname{\mathbb{C}}^g)$ the endomorphism $\xi \otimes {\mathrm}{id}_{\operatorname{\mathbb{R}}}$. Let us introduce the following $\xi$-adic “Tate modules”: $$\begin{aligned} T_{\xi} \Gamma &=& \underleftarrow{\lim}_\nu \Gamma/\xi^\nu \Gamma,\\ V_{\xi} \Gamma &=& T_{\xi} \Gamma \otimes \operatorname{\mathbb{Q}}.\end{aligned}$$ Note that multiplication by $\xi$ defines an automorphism of $V_\xi \Gamma$. We also define an additive subgroup of $\operatorname{\mathbb{C}}^g$ by $$\begin{aligned} V &=& \bigcup_{\nu \geq 0} \xi^{-\nu}_{\operatorname{\mathbb{R}}} \Gamma\end{aligned}$$ The group $V$ acts on $\operatorname{\mathbb{C}}^g \times V_\xi \Gamma$ by $v.(z, \hat{v})=(z+v, \hat{v}-v)$ and we denote by $\operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma$ the quotient space. The group $q^{\operatorname{\mathbb{Z}}}$ acts on $(\operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma) \times \operatorname{\mathbb{R}}^\times_+$ by $q^\nu ([z, \hat{v}], x)=([\xi^{-\nu}_{\operatorname{\mathbb{R}}}z, \xi^{-\nu} \hat{v}], q^\nu x)$ and we denote by $$S(A_0)=(\operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma) \times_{q^{\operatorname{\mathbb{Z}}}} \operatorname{\mathbb{R}}^\times_+$$ the quotient space. The images of the sets $\operatorname{\mathbb{C}}^g \times \{\hat{v}\} \times \{x\}$ by the natural projection map $\pi: \operatorname{\mathbb{C}}^g \times V_\xi \Gamma \times \operatorname{\mathbb{R}}^\times_+ \longrightarrow S(A_0)$ form a partition $\mathcal{F}$ of $S(A_0)$ such that $(S(A_0), \mathcal{F})$ is a foliated space (see [@cc] Def. 11.2.12). The tangent spaces to the leaves form a $\operatorname{\mathbb{R}}$-vector budle $T\mathcal{F}$ over $S(A_0)$. We define a flow $\phi^t$ on $(S(A_0), \mathcal{F})$ by $$\phi^t([z, \hat{v}, x])=[z, \hat{v}, e^tx].$$ Clearly, this flow sends each leaf to another leaf.\ As the complex torus $\operatorname{\mathbb{C}}^g/\Gamma$ is the set of complex points of an abelian variety, it admits a non-degenerate Riemann form $H$ (see [@cornell-silvermann] Thm. A p. 85). This means that there exists a positive definite hermitian form $H$ on $\operatorname{\mathbb{C}}^g$ such that the restriction of the imaginary part ${\mathrm}{Im} H$ to $\Gamma$ is integral valued. Let us fix such an hermitian form $H$ once and for all and let us denote by $\Psi: \Gamma \otimes \Gamma \longrightarrow \operatorname{\mathbb{Z}}$ the induced pairing. Note that, for any $\eta_1, \eta_2 \in \operatorname{\mathbb{C}}^g$, we have ${\mathrm}{Re} H(\eta_1, \eta_2)={\mathrm}{Im} H(i\eta_1, \eta_2)$. \[metrique\] The Riemannian metric on the bundle $T\operatorname{\mathbb{C}}^g \times V_\xi \Gamma \times \operatorname{\mathbb{R}}^\times_+$ over the space $\operatorname{\mathbb{C}}^g \times V_\xi \Gamma \times \operatorname{\mathbb{R}}^\times_+$ given by the formula $$\tilde{g}_{[z, \hat{v}, x]}(\eta_1, \eta_2)=x {\mathrm}{Re} H(\eta_1, \eta_2)$$ induces a metric $g$ along the leaves of $(S(A_0), \mathcal{F})$ such that $$(\phi^t)^*g=e^t g.$$ To prove the first statement, we need to show that for every $\nu \in \operatorname{\mathbb{Z}}$ one has $${\mathrm}{Re} H(\xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_1), \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_2))=q^{\nu}{\mathrm}{Re} H (\eta_1, \eta_2).$$ Let $l \neq p$ be a prime. Identify $\operatorname{\mathbb{Z}}_l$ to $\operatorname{\mathbb{Z}}_l(1)$ via the compatible system of primitive $l^n$-th roots of unity $(e^{2\pi i/l^n})$. Then the Weil pairing $$\Psi_l: T_l(A) \otimes T_l(A) \longrightarrow \operatorname{\mathbb{Z}}_l(1)$$ is minus the $\operatorname{\mathbb{Z}}_l$-linear extension of $\Psi$ ([@mumford] Thm. 1 p. 237). As a consequence, for any $\gamma_1, \gamma_2 \in \Gamma$, we have $$\Psi(\xi^{\nu}(\gamma_1), \xi^{\nu}(\gamma_2))=-\Psi_l(\xi^{\nu}(\gamma_1), \xi^{\nu}(\gamma_2))=-q^{\nu}\Psi_l(\gamma_1, \gamma_2)=q^{\nu}\Psi(\gamma_1. \gamma_2)$$ This implies that ${\mathrm}{Im} H(\xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_1), \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_2))=q^{\nu} {\mathrm}{Im} H(\eta_1, \eta_2)$ for any $\eta_1, \eta_2 \in \operatorname{\mathbb{C}}^g=\Gamma \otimes \operatorname{\mathbb{R}}$. As a consequence we have $$\begin{aligned} {\mathrm}{Re} H(\xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_1), \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_2)) &=& {\mathrm}{Im} H(i \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_1), \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_2))\\ &=& {\mathrm}{Im} H(\xi_{\operatorname{\mathbb{R}}}^{\nu}(i \eta_1), \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_2))\\ &=& q^{\nu} {\mathrm}{Im} H(i \eta_1, \eta_2)\\ &=& q^{\nu} {\mathrm}{Re} H(\eta_1, \eta_2).\end{aligned}$$ Let us prove the second statement. Let $t \in \operatorname{\mathbb{R}}$, $x \in \operatorname{\mathbb{R}}^\times_+$ and assume that $e^tx=q^\nu x'$ for some $\nu \in \operatorname{\mathbb{Z}}$ and $x' \in \operatorname{\mathbb{R}}^\times_+$. Then for any $z \in \operatorname{\mathbb{C}}^g$ and $\hat{v} \in V_\xi \Gamma$, we have $$\phi^t([z, \hat{v}, x])=[\xi_{\operatorname{\mathbb{R}}}^{\nu}(z), \xi^{\nu}(\hat{v}), x']$$ and the tangent map $T_{[z, \hat{v}, x]}\phi^t: T_{[z, \hat{v}, x]} \mathcal{F} \longrightarrow T_{[\xi_{\operatorname{\mathbb{R}}}^{\nu}(z), \xi^{\nu}(\hat{v}), x']} \mathcal{F}$ sends $\eta \in \operatorname{\mathbb{C}}^g$ to $\xi^{\nu}(\eta)$. As a consequence $$\begin{aligned} (\phi^{t*}g)(\eta_1, \eta_2) &=& x'\operatorname{Re}H(T_{[z, \hat{v}, x]}\phi^t(\eta_1), T_{[z, \hat{v}, x]}\phi^t(\eta_2))\\ &=& x' \operatorname{Re}H(\xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_1), \xi_{\operatorname{\mathbb{R}}}^{\nu}(\eta_2))\\ &=& q^{\nu}x' \operatorname{Re}H(\eta_1, \eta_2)\\ &=& e^tg(\eta_1, \eta_2).\end{aligned}$$ \[basis\] There exists an orthonormal basis $ \left( \frac{\partial}{\partial x_1}, \frac{\partial}{\partial y_1}, \ldots, \frac{\partial}{\partial x_{g}}, \frac{\partial}{\partial y_{g}} \right)$ of $(\Gamma \otimes \operatorname{\mathbb{R}}, \operatorname{Re}H)$ such that the vectors $\frac{\partial}{\partial z_j}=\frac{\partial}{\partial x_j}+\frac{\partial}{\partial y_j} \otimes i \in \Gamma \otimes \operatorname{\mathbb{C}}$ are eigenvectors of $\xi_{\operatorname{\mathbb{C}}}=\xi \otimes \operatorname{id}_{\operatorname{\mathbb{C}}}$. We have explained in the proof of the lemma above that $q^{-\frac{1}{2}} \xi_{\operatorname{\mathbb{R}}}$ is an orthogonal automorphism of $(\Gamma \otimes \operatorname{\mathbb{R}}, \operatorname{Re}H)$. By elementary linear algebra, there exists an orthonormal basis $ \left( \frac{\partial}{\partial x_1}, \frac{\partial}{\partial y_1}, \ldots, \frac{\partial}{\partial x_{g}}, \frac{\partial}{\partial y_{g}} \right)$ of $(\Gamma \otimes \operatorname{\mathbb{R}}, \operatorname{Re}H)$ in which the matrix of $q^{-\frac{1}{2}} \xi_{\operatorname{\mathbb{R}}}$ is diagonal by blocks of size one of the form $(\pm 1)$ and of size two of the form $\begin{pmatrix} \cos \theta & -\sin \theta\\ \sin \theta & \cos \theta \\ \end{pmatrix}$. In fact, the blocks of size one do not occur because $\pm q^{\frac{1}{2}}$ is not an eigenvalue of $\xi$ by [@deligne] proof of Thm. 7. (A). Then the basis $ \left( \frac{\partial}{\partial x_1}, \frac{\partial}{\partial y_1}, \ldots, \frac{\partial}{\partial x_{g}}, \frac{\partial}{\partial y_{g}} \right)$ satisfies the statement of the lemma. The following results will be useful later: let us denote by $\Gamma_p$ the $p$-adic Tate module $\Gamma \otimes \operatorname{\mathbb{Z}}_p$. According to [@deligne] 6.1 and 6.2, we have a decomposition into $\operatorname{\mathbb{Z}}_p$-modules $\Gamma_p=\Gamma'_p \oplus \Gamma''_p$ such that $\xi(\Gamma'_p)=\Gamma'_p$ and $\xi(\Gamma''_p)=q \Gamma''_p$. It follows from (\[determinant\]) that $\Gamma/\xi^\nu \Gamma$ is a finite abelian $p$-group for every $\nu \geq 0$ and hence that $T_{\xi} \Gamma=\underleftarrow{\lim}_\nu \Gamma_p/\xi^\nu \Gamma_p$. As a consequence, we have $$\label{technique2} T_\xi \Gamma=\underleftarrow{\lim}_\nu \Gamma''_p/\xi^\nu \Gamma''_p= \underleftarrow{\lim}_\nu \Gamma''_p/q^\nu \Gamma''_p=\Gamma''_p.$$ We define the quotient space $\operatorname{\mathbb{C}}^g \times_{\Gamma} T_\xi \Gamma$ similarly as $\operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma$. \[technique\] The natural inclusion $\operatorname{\mathbb{C}}^g \times T_\xi \Gamma \longrightarrow \operatorname{\mathbb{C}}^g \times V_\xi \Gamma$ induces a canonical isomorphism $\operatorname{\mathbb{C}}^g \times_\Gamma T_\xi \Gamma \simeq \operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma$ which is equivariant with respect to the diagonal action of $\xi$ on both sides. The natural inclusion $\Gamma[1/p] \longrightarrow V_\xi \Gamma$, which is dense, factors through the inclusion $\Gamma''_p[1/p] \longrightarrow V_\xi \Gamma$. Hence, for any $\hat{v} \in V_\xi \Gamma$ there exists $p^{-\alpha} \gamma$, where $\alpha \in \operatorname{\mathbb{Z}}$ and $\gamma \in \Gamma$ such that $\hat{v}-p^{-\alpha} \gamma \in T_\xi \Gamma$ and we can assume that $\gamma \in \Gamma \cap \Gamma''_p$. Hence we have $\xi_{\operatorname{\mathbb{R}}}^\alpha(p^{-\alpha} \gamma) \in \Gamma$ which means that $p^{-\alpha} \gamma \in V$. This concludes the proof. \[compact\] The topological space $S(A_0)$ is compact. The natural projection $\operatorname{\mathbb{C}}^g \times_{\Gamma} T_{\xi} \Gamma \longrightarrow \operatorname{\mathbb{C}}^g/\Gamma$ is locally trivial with compact fiber and compact target. This implies that $\operatorname{\mathbb{C}}^g \times_\Gamma T_\xi \Gamma$ is compact. Similarly, we have that $(\operatorname{\mathbb{C}}^g \times_{\Gamma} T_\xi \Gamma) \times_{q^{\operatorname{\mathbb{Z}}}} \operatorname{\mathbb{R}}^\times_+$ is compact. But this space is nothing but $S(A_0)$ by the previous lemma. $L^2$-harmonic forms along the leaves ===================================== In this section, we work with a fixed orthonormal basis of $(\operatorname{\mathbb{C}}^g, \operatorname{Re}H)$ given by Lem. \[basis\] and we denote by $(dx_1, dy_1, \ldots, dx_g, dy_g)$ the dual basis. \[mesure\] Let $\mu_\xi$ be a Haar measure on the locally compact abelian group $V_\xi \Gamma$.\ $(i).$ The measure $\prod_{j=1}^g dx_j dy_j \otimes \mu_\xi \otimes \frac{dx}{x}$ induces a measure $\mu$ on $S(A_0)$.\ $(ii).$ The measure $\mu$ is invariant under the action of $\phi^t$. (i). We need to check that for any $\nu \in \operatorname{\mathbb{Z}}$ and any Borel subset $A$ of $V_\xi \Gamma$ we have $$\mu_{\xi}(\xi^{\nu}(A))=|\det{\xi}^{\nu}|^{-1}\mu_{\xi}(A)=q^{-g\nu}\mu_{\xi}(A).$$ As $A \longmapsto \mu_{\xi}(\xi^{\nu}(A))$ is also a Haar measure on $V_{\xi} \Gamma$, it is enough to verify the above equality for $A=T_{\xi} \Gamma$. But it follows from the fact that, for any $\nu \geq 0$, we have $T_\xi \Gamma/\xi^\nu T_\xi \Gamma=\Gamma/\xi^\nu \Gamma$ has $\det(\xi^\nu)=q^{g\nu}$ elements. (ii). Trivial. For $0 \leq j \leq 2g$ let $\mathcal{A}^j_{\mathcal{F}}(S(A_0))$ be the $\operatorname{\mathbb{R}}$-vector space of sections of the vector bundle $\bigwedge^j T^*\mathcal{F}$ which are smooth along the $\mathcal{F}$-leaves and continuous transversally (see [@moore-schochet] III Def. 3.2). Denote by $$d^j_{\mathcal{F}}: \mathcal{A}^j_{\mathcal{F}}(S(A_0)) \longrightarrow \mathcal{A}^{j+1}_{\mathcal{F}}(S(A_0))$$ the leafwise exterior derivative. These form the de Rham complex along $\mathcal{F}$ $$\begin{CD} 0 @>>> \mathcal{A}^0_{\mathcal{F}}(S(A_0)) @>>> \ldots @>>> \mathcal{A}^{2g}_{\mathcal{F}}(S(A_0)) @>>> 0. \end{CD}$$ The metric $g$ induces a metric $g_j$ on the vector bundle $\bigwedge^j T^*\mathcal{F}$ in the following standard way: if $e_1, \ldots, e_{2g}$ is an orthonormal basis of $T_{[z, \hat{v}, x]} \mathcal{F}$ and $e_1^*, \ldots, e_{2g}^*$ is the dual basis, then the vectors $e_{i_1}^* \wedge \ldots \wedge e_{i_j}^*$ for $1 \leq i_1 < \ldots < i_j \leq 2g$ form an orthonormal basis of $\bigwedge^j T^*_{[z, \hat{v}, x]}\mathcal{F}$. Because $S(A_0)$ is compact by Cor. \[compact\] we can define the scalar product on $\mathcal{A}^j_{\mathcal{F}}(S(A_0))$ by $$\langle \omega, \omega' \rangle=\int_{S(A_0)} g_j(\omega, \omega') d\mu.$$ Let $d_{\mathcal{F}}^{j\dagger}: \mathcal{A}^j_{\mathcal{F}}(S(A_0)) \longrightarrow \mathcal{A}^{j-1}_{\mathcal{F}}(S(A_0))$ be the formal adjoint of $d^j_{\mathcal{F}}$. We also denote as usual by $\Delta^j_\mathcal{F}=d^j_\mathcal{F}d^{j \dagger}_\mathcal{F}+d^{j \dagger}_\mathcal{F} d^j_\mathcal{F}$ the Laplacian. \[formulae\] Let $$\omega=\sum_{1 \leq i_1< \ldots < i_j \leq 2g} \alpha_{i_1 \ldots i_j} du_{i_1} \wedge \ldots \wedge du_{i_j} \in \mathcal{A}^j_{\mathcal{F}}(S(A_0)).$$ Then $$\begin{aligned} d_{\mathcal{F}}^j \omega &=& \sum_{1 \leq i_1 < \ldots < i_{j+1} \leq 2g} \sum_{1 \leq k \leq j+1}(-1)^{k-1} \frac{\partial \alpha_{i_1 \ldots \hat{i}_k \ldots i_{j+1}}}{\partial u_{i_k}} du_{i_1} \wedge \ldots \wedge du_{i_{j+1}},\\ d_{\mathcal{F}}^{j\dagger} \omega &=& -x^{-1} \sum_{1 \leq i_1< \ldots < i_j \leq 2g} \sum_{1 \leq k \leq j}(-1)^{k-1}\frac{\partial \alpha_{i_1 \ldots i_j}}{\partial u_{i_k}} du_{i_1} \wedge \ldots \wedge \hat{du}_{i_k} \ldots \wedge du_{i_j},\\ \Delta^j_\mathcal{F} \omega &=& -x^{-1} \sum_{1 \leq i_1< \ldots < i_j \leq 2g} \sum_{1 \leq k \leq j} \frac{\partial^2 \alpha_{i_1 \ldots i_j}}{\partial u_k^2} du_{i_1} \wedge \ldots \wedge du_{i_j}.\end{aligned}$$ The first statement follows from an easy standard computation. To prove the second statement let $$\eta=\sum_{1 \leq i_1<\ldots<i_{j-1} \leq 2g}\beta_{i_1 \ldots i_{j-1} } du_{i_1} \wedge \ldots \wedge du_{i_{j-1}} \in \mathcal{A}^{j-1}_{\mathcal{F}}(S(A_0)).$$ Then, one has $$\langle d_{\mathcal{F}}^{j\dagger} \omega, \eta \rangle = \langle \omega, d^{j-1}_{\mathcal{F}} \eta \rangle$$ $$\begin{aligned} &=& \sum_{1 \leq i_1< \ldots < i_j \leq 2g} \sum_{1 \leq k \leq j}(-1)^{k-1} \langle \alpha_{i_1 \ldots i_j} du_{i_1} \wedge \ldots \wedge du_{i_j}, \frac{\partial \beta_{i_1 \ldots \hat{i}_k \ldots i_{j}}}{\partial u_{i_k}} du_{i_1} \wedge \ldots \wedge du_{i_{j}} \rangle\\ &=& \sum_{1 \leq i_1< \ldots < i_j \leq 2g} \sum_{1 \leq k \leq j}(-1)^{k-1} \int_{S(A_0)} x^{-j} \alpha_{i_1 \ldots i_j}\frac{\partial \beta_{i_1 \ldots \hat{i}_k \ldots i_{j}}}{\partial u_{i_k}} d\mu\\ &=& -\sum_{1 \leq i_1< \ldots < i_j \leq 2g} \sum_{1 \leq k \leq j}(-1)^{k-1} \int_{S(A_0)} x^{-j} \frac{\partial \alpha_{i_1 \ldots i_j}}{\partial u_{i_k}} \beta_{i_1 \ldots \hat{i}_k \ldots i_{j}} d\mu \\ &=& \left \langle -x^{-1} \sum_{1 \leq i_1< \ldots < i_j \leq 2g} \sum_{1 \leq k \leq j}(-1)^{k-1}\frac{\partial \alpha_{i_1 \ldots i_j}}{\partial u_{i_k}} du_{i_1} \wedge \ldots \wedge \hat{du}_{i_k} \ldots \wedge du_{i_j}, \eta \right \rangle\end{aligned}$$ which proves the statement. The proof of the last statement is left to the reader. Let $\mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))$ be the $L^2$-completion of $\mathcal{A}^j_{\mathcal{F}}(S(A_0))$, view $d_{\mathcal{F}}^j$ as an unbounded operator on $\mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))$ and define $\tilde{d}_{\mathcal{F}}^j$ to be the closed unbounded operator $\tilde{d}_{\mathcal{F}}^j={d}_{\mathcal{F}}^{j\dagger *}$. Define $${\mathrm}{Harm}^j_{L^2}(S(A_0))=\ker \tilde{d}^j_{\mathcal{F}} \cap \ker \tilde{d}_{\mathcal{F}}^{j*}.$$ This is a Hilbert space, as a closed subspace of $\mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))$.\ We introduce the following notation: $$\begin{aligned} \mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))_{\operatorname{\mathbb{C}}} &=& \mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0)) \otimes_{\operatorname{\mathbb{R}}} \operatorname{\mathbb{C}},\\ {\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}} &=& {\mathrm}{Harm}^j_{L^2}(S(A_0)) \otimes_{\operatorname{\mathbb{R}}} \operatorname{\mathbb{C}}.\end{aligned}$$ Let us denote by $(d\tau_1, \ldots, d\tau_{2g})$ the basis of $(\Gamma \otimes \operatorname{\mathbb{C}})^*$ which is dual to the basis $(\frac{\partial}{\partial z_1}, \frac{\partial}{\partial \overline{z}_1}, \ldots, \frac{\partial}{\partial z_g}, \frac{\partial}{\partial \overline{z}_g})$ (see Lem. \[basis\]). We endow ${\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ with the hermitian scalar product induced by the scalar product on ${\mathrm}{Harm}^j_{L^2}(S(A_0))$. Note that any element $\omega$ of $\mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ is written as $$\omega=\sum_{1 \leq i_1< \ldots < i_j \leq 2g} \alpha_{i_1 \ldots i_j} d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j}$$ where $\alpha_{i_1 \ldots i_j}$ are $\operatorname{\mathbb{C}}$-valued $L^2_{loc}$ functions on $\operatorname{\mathbb{C}}^g \times V_{\xi} \Gamma \times \operatorname{\mathbb{R}}^\times_+$ satisfying the following invariance properties: $$\begin{aligned} \alpha_{i_1 \ldots i_j}(z+v, \hat{v}-v, x) &=& \alpha_{i_1 \ldots i_j}(z, \hat{v},x),\\ \alpha_{i_1 \ldots i_j}(z, \hat{v},q^\nu x) d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j} &=& \alpha_{i_1 \ldots i_j}(z, \hat{v},x) \xi^{\nu *}_{\operatorname{\mathbb{C}}}(d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j})\end{aligned}$$ The proof of the following result is inspired from the beginning of the one of [@deninger] Thm. 4.1. \[harmo\] Let $$\omega=\sum_{1 \leq i_1< \ldots < i_j \leq 2g} \alpha_{i_1 \ldots i_j} d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j} \in \mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))_{\operatorname{\mathbb{C}}}.$$ Then $\omega \in {\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ if and only if for any $x \in \operatorname{\mathbb{R}}^\times_+$ the functions defined by $\operatorname{\mathbb{C}}^g \times V_\xi \Gamma \longrightarrow \operatorname{\mathbb{C}}$, $(z, \hat{v}) \longmapsto \alpha_{i_1 \ldots i_j}(z, \hat{v}, x)$ are constant. The fact that the condition is sufficient follows from the first two statements of Lem. \[formulae\]. Let us prove that the condition is necessary. Fix $x \in \operatorname{\mathbb{R}}^\times_+$. Denote by $\overline{M}$ the abelian group $\operatorname{\mathbb{C}}^g \times_V V_{\xi} \Gamma$. As $\alpha_{i_1 \ldots i_j}$ is in $L^2_{loc}$ and $\overline{M}$ is compact (Cor. \[compact\]), the function $(z, \hat{v}) \longmapsto \alpha_{i_1 \ldots i_j}(z, \hat{v}, x)$ is in $L^2(\overline{M})$. The character group of $\overline{M}$ is $$\left\{ \chi \otimes \chi' \,|\, \chi \in (\operatorname{\mathbb{C}}^g)^\vee, \chi' \in (V_{\xi} \Gamma)^\vee, \chi|_V=\chi'|_V \right\}$$ hence, by Fourier theory, one has the equality $$\alpha_{i_1 \ldots i_j}(z, \hat{v}, x) = \sum_{\chi|_V=\chi'|_V} a_{\chi, \chi'} \chi(z) \chi'(\hat{v})$$ in $L^2(\overline{M})$. Any character $\chi$ of $\operatorname{\mathbb{C}}^g$ is of the form $$\chi(z)=\chi_w(z)=\prod_{1 \leq k \leq g} \exp(w_k z_k- \overline{w}_k \overline{z}_k)$$ for a uniquely determined $w=(w_1, \ldots, w_g) \in \operatorname{\mathbb{C}}^g$. Because $$\omega \in {\mathrm}{Harm}^j_{L^2}(S(A_0)),$$ we have $\langle \omega, \Delta^j_{\mathcal{F}} \beta \rangle =0$ for every $\beta \in \mathcal{A}^j_{\mathcal{F}}(S(A_0))$. Since distributional derivatives commute with convergent series of distributions it follows that $$\sum_{\chi|_V=\chi'|_V} a_{\chi, \chi'} \Delta(\chi(z)) \chi'(\hat{v})=0$$ where $\Delta$ is the usual Laplacian on $\operatorname{\mathbb{C}}^g=\operatorname{\mathbb{R}}^{2g}$ (see the third statement of Lem. \[formulae\]). As $\Delta \chi_w(z)=-|w|^2 \chi_w(z)$, the coefficients $a_{\chi, \chi'}$ corresponding to the non-trivial $\chi$ are zero. This implies that the $\chi'$ such that $a_{\chi, \chi'} \neq 0$ are trivial on $V$ hence are trivial on $V_{\xi} \Gamma$ by density of $V$ in $V_{\xi} \Gamma$. This shows that $\alpha_{i_1 \ldots i_j}(z, \hat{v}, x)$ does not depend on $(z, \hat{v}) \in \operatorname{\mathbb{C}}^g \times V_{\xi} \Gamma$. Recall that we work with a basis of the $\operatorname{\mathbb{R}}$-vector space $\Gamma \otimes \operatorname{\mathbb{R}}=\operatorname{\mathbb{C}}^g$ given by Lem. \[basis\]. In particular, each $d\tau_i$ is an eigenvector of $\xi_{\operatorname{\mathbb{C}}}$. Let us denote by $\mu_i$ the corresponding eigenvalue and let us fix a branch $\log_q$ of the complex logarithm for the basis $q$. The Hilbert space ${\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ has an orthonormal basis consisting of the $$x^{\frac{2 \pi i \nu}{\log q}+\log_q(\prod_{k=1}^j \mu_{i_k})} d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j}$$ where $\nu \in \operatorname{\mathbb{Z}}$ and $1 \leq i_1 < \ldots< i_j \leq 2g$. According to Lem. \[harmo\] an element $$\alpha_{i_1 \ldots i_j} d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j}$$ of $\mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ belongs to ${\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ if and only if $\alpha_{i_1 \ldots i_j}$ does not depend on the variables in $\operatorname{\mathbb{C}}^g$ and $V_{\xi} \Gamma$. By (6), we have $$\alpha_{i_1 \ldots i_j}(q^\nu x)=\left(\prod_{k=1}^j \mu_{i_k} \right)^\nu \alpha_{i_1 \ldots i_j}(x)$$ for any $\nu \in \operatorname{\mathbb{Z}}$ and $x \in \operatorname{\mathbb{R}}^\times_+$. As a consequence, the function $$x \longmapsto \left(\prod_{k=1}^j \mu_{i_k} \right)^{-\log_q x} \alpha_{i_1 \ldots i_j}(x)=x^{-\log_q(\prod_{k=1}^j \mu_{i_k})} \alpha_{i_1 \ldots i_j}(x)$$ is in $L^2(\operatorname{\mathbb{R}}_+^\times/q^{\operatorname{\mathbb{Z}}}, \operatorname{\mathbb{C}})$. Then the fact that the family $$\left(x^{\frac{2 \pi i \nu}{\log q}+\log_q(\prod_{k=1}^j \mu_{i_k})} d\tau_{i_1} \wedge \ldots \wedge d\tau_{i_j} \right)_{\nu \in \operatorname{\mathbb{Z}}, 1 \leq i_1<\ldots<i_j \leq 2g}$$ is a basis follows from the fact that $\left(x^{\frac{2 \pi i \nu}{\log q}} \right)_{\nu \in \operatorname{\mathbb{Z}}}$ is a basis of $L^2(\operatorname{\mathbb{R}}_+^\times/q^{\operatorname{\mathbb{Z}}}, \operatorname{\mathbb{C}})$. The orthogonality statement is obvious. The fact that the vectors have norm equal to one follows from the fact that the eigenvalues $\mu$ of $\xi$ verify $\mu \overline{\mu}=q$ by Weil. Since $\mu$ is $\phi^t$-invariant, for $\omega, \omega' \in \mathcal{A}^j_{\mathcal{F}}(S(A_0))$ we have $$\langle \phi^{t *}\omega, \phi^{t *}\omega' \rangle=e^{jt}\langle \omega, \omega' \rangle.$$ Hence $e^{-jt/2}\phi^{t *}$ is an orthogonal operator on the real Hilbert space $\mathcal{A}^j_{\mathcal{F}}(S(A_0))$. It follows that $\phi^{t*}$ has a unique extension to $\mathcal{A}^j_{\mathcal{F}, L^2}(S(A_0))$, still denoted by $\phi^{t*}$, such that $e^{-jt/2} \phi^{t*}$ is orthogonal. As a consequence $\phi^{t*}$ commutes with $\tilde{d}^j_{\mathcal{F}}$ and with $\tilde{d}_{\mathcal{F}}^{j*}$. In particular $\phi^{t *}$ leaves ${\mathrm}{Harm}^j_{L^2}(S(A_0))$ invariant. Let $\alpha \in \mathcal{C}^\infty_0(\operatorname{\mathbb{R}})$ be a test function on $\operatorname{\mathbb{R}}$. Consider the bounded operator $$S_j(\alpha)=\int_{\operatorname{\mathbb{R}}} \alpha(t) \phi^{t*}_{\operatorname{\mathbb{C}}}dt$$ on the complex Hilbert space ${\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}}$ where $\phi^{t*}_{\operatorname{\mathbb{C}}}=\phi^{t*} \otimes \operatorname{id}_{\operatorname{\mathbb{C}}}$. \[eigenvalues\] Let $0 \leq j \leq 2g$. For every $\alpha \in \mathcal{C}^\infty_0(\operatorname{\mathbb{R}})$, the operator $S_j(\alpha)$ is of trace class and its trace is given by $$\operatorname{Tr}\left( S_j(\alpha)\,|\, {\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}} \right)=\sum_{\rho_j} \Phi(\rho_j)$$ where the sum is indexed by the zeroes of $P_j(q^{-s})$ and where $$\Phi(s)=\int_{\operatorname{\mathbb{R}}}e^{ts}\alpha(t)dt.$$ For every $\nu \in \operatorname{\mathbb{Z}}$ and every $1 \leq i_1<\ldots<i_j \leq 2g$, the basis vector $$x^{\frac{2 \pi i \nu}{\log q}+\log_q(\prod_{k=1}^j \mu_{i_k})} du_{i_1} \wedge \ldots \wedge du_{i_j}$$ is an eigenvector of $\phi^{t*}$ with eigenvalue $e^{t\rho_j}$ where $$\rho_j=\frac{2 \pi i \nu}{\log q}+\log_q(\prod_{k=1}^j \mu_{i_k}).$$ These $\rho_j$ are precisely the zeroes of $P_j(q^{-s})$. Hence we need to show that $$\sum_{\nu \in \operatorname{\mathbb{Z}}, 1 \leq i_1<\ldots<i_j \leq 2g} \left| \int_{\operatorname{\mathbb{R}}} e^{t \rho_j} \alpha(t) dt \right| < +\infty.$$ But this follows from straightforward estimates of Fourier coefficients. A transversal index computation on $(S(A_0), \mathcal{F}, \phi^t)$ ================================================================== The transversal index of the de Rham complex of $S(A_0)$ along $\mathcal{F}$ is defined as the distribution $${\mathrm}{Ind}_t(d_{\mathcal{F}}): \alpha \longmapsto {\mathrm}{Ind}_t(d_{\mathcal{F}})(\alpha)=\sum_{j=0}^{2g}(-1)^j \operatorname{Tr}\left( S_j(\alpha)\,|\, {\mathrm}{Harm}^j_{L^2}(S(A_0))_{\operatorname{\mathbb{C}}} \right).$$ Then by Prop. \[eigenvalues\], the left hand side of the explicit formula (\[EF\]) for $\alpha$ equals ${\mathrm}{Ind}_t(d_{\mathcal{F}})(\alpha)$. In this section, we will interpret the left hand term of (\[EF\]) in terms of the dynamical system $(S(A_0), \mathcal{F}, \phi^t)$ by proving the existence of a natural bijection between the set of primitive compact orbits of $\phi^t$ and the set of closed points of the abelian variety $A_0$. Furthermore, according to Deninger’s analogy [@deninger3], if a primitive compact orbit $\gamma$ corresponds to the closed point $x$, then the length of $\gamma$ should be $\log N(x)$ where $N(x)=q^{\deg(x)}$ denotes the cardinality of the residue field of $x$. To prove these facts, we need a preliminary lemma.\ For any compact orbit $\gamma$ of $\phi^t$ on $S(A_0)$, denote by $l(\gamma)$ the length of $\gamma$. Note that if $[z, \hat{v}, x] \in S(A_0)$ is such that $\phi^t[z, \hat{v}, x]=[z, \hat{v}, x]$ then necessarily $t=\nu \log q$ for some integer $\nu \geq 0$. This shows that the length of any compact orbit is of the form $\nu \log q$ for some integer $\nu \geq 0$. \[technique-pour-noeuds1\] There is a natural bijection between the set of primitive compact orbits of $\phi^t$ of length $\nu \log q$ and the set of $\xi$-orbits on $\operatorname{\mathbb{C}}^g/\Gamma$ of order $\nu$. Let us denote by $\overline{M}$ be the quotient space $\operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma$. Then $q^{\operatorname{\mathbb{Z}}}$ acts on $\overline{M}$ by $q^\nu[z, \hat{v}]=[\xi^{-\nu}_{\operatorname{\mathbb{R}}}z, \xi^{-\nu} \hat{v}]$. Identify $\overline{M}$ with its image $\overline{M} \times \{1\}$ in $S(A_0)$. Then the map $\gamma \longmapsto \gamma \cap \overline{M}$ is a bijection between the set of primitive compact orbits of $\phi^t$ on $S(A_0)$ of length $\nu \log q$ and the finite orbits of $q^{-1}$ on $\overline{M}$ of order $\nu$. According to Lem. \[technique\], there is a natural bijection $\operatorname{\mathbb{C}}^g \times_\Gamma T_\xi \Gamma \simeq \operatorname{\mathbb{C}}^g \times_V V_\xi \Gamma$ which is equivariant with respect to the diagonal action of $\xi$ on both sides. Furthermore, under the projection $\operatorname{\mathbb{C}}^g \times_\Gamma T_\xi \Gamma \longrightarrow \operatorname{\mathbb{C}}^g/\Gamma$, the orbits of order $\nu$ of the diagonal action of $\xi$ are mapped bijectively onto the $\xi$-orbits of order $\nu$ on $\operatorname{\mathbb{C}}^g/\Gamma$. The inverse map sends the orbit of $z+\Gamma$ to the orbit of $[z, \hat{\gamma}]$ where $\hat{\gamma}=(1-\xi^\nu)^{-1} \gamma$ if $\gamma=\xi^{\nu}z-z \in \Gamma$. Note that this makes sense because $1-\xi^{\nu}$ is invertible on $T_\xi \Gamma$: indeed if we denote by $V$ the Verschiebung endomorphism of $\Gamma$ we have $(1-\xi^\nu)^{-1}=V^{\nu}(1-q^{\nu})^{-1}$. The proof of the next result is a direct generalisation of the one of [@deninger] Prop. 3.3. \[noeuds\] There is a natural bijection between the set of closed points of $A_0$ and the set of primitive compact orbits of $\phi^t$ on $S(A_0)$ such that if $x$ corresponds to $\gamma$, then $$l(\gamma)=\log N(x),$$ where $N(x)=q^{\deg(x)}$ denotes the cardinality of the residue field of $x$. The closed points $x$ of $A_0$ such that $N(x)=n$ are in bijection with the $\phi_0$-orbits on $A_0(\overline{\operatorname{\mathbb{F}}}_q)$ of order $n$. Hence, according to Lem. \[technique-pour-noeuds1\], it remains to construct a natural bijection between the $\xi$-orbits of order $n$ on $\operatorname{\mathbb{C}}^g/\Gamma$ and the $\phi_0$-orbits of the same order on $A_0(\overline{\operatorname{\mathbb{F}}}_q)$. Let $\overline{K}$ be the algebraic closure of $K$ in $\operatorname{\mathbb{C}}$ and let $\overline{W}$ be the integral closure of $W$ in $\overline{K}$. Fix a maximal ideal $\overline{\mathfrak{m}}$ over $\mathfrak{m}$. Then $\overline{W}/\overline{\mathfrak{m}}=\overline{\operatorname{\mathbb{F}}}_q$. By the valuative criterion of properness, the natural map $\mathcal{A}(\overline{W}) \longrightarrow A(\overline{K})$ is an isomorphism. As the torsion points of $A(\operatorname{\mathbb{C}})$ are algebraic over $K$, the inclusion $A(\overline{K}) \longrightarrow A(\operatorname{\mathbb{C}})$ induces an isomorphism on the torsion subgroups. As a consequence, we obtain a reduction map $${\mathrm}{red}: A(\operatorname{\mathbb{C}})_{tors} \simeq \mathcal{A}(\overline{W})_{tors} \longrightarrow \mathcal{A}(\overline{W}/\overline{\mathfrak{m}})_{tors}=A_0(\overline{\operatorname{\mathbb{F}}}_q).$$ This map is equivariant with respect to the actions of $\xi$ on the left and $\phi_0$ on the right. The group $$(A(\operatorname{\mathbb{C}})_{tors})^{\xi^n=1}=A(\operatorname{\mathbb{C}})^{\xi^n=1}=(\operatorname{\mathbb{C}}^g/\Gamma)^{\xi^n=1}=(\xi^n-1)^{-1} \Gamma/\Gamma=\Gamma/(\xi^n-1)\Gamma$$ is finite and has order $\det(\xi^n-1)$. The group $A_0(\overline{\operatorname{\mathbb{F}}}_q)^{\phi_0^n=1}=A_0(\operatorname{\mathbb{F}}_{q^n})$ has the same order according to the first statement of [@mumford] IV §21 Thm. 4. Moreover, for any integer $N$ prime to $p$, the restriction $${\mathrm}{red}: A(\operatorname{\mathbb{C}})_N \longrightarrow A_0(\overline{\operatorname{\mathbb{F}}}_q)_N$$ is an isomorphism (see [@deligne] (3.1)). As $\mathcal{A}/W$ is the Serre-Tate canonical lift of $A_0/k$, we have $$\mathcal{A}(\overline{W})_{p^n}=\mathcal{G}^0_{p^n} \times (\operatorname{\mathbb{Z}}/p^n \operatorname{\mathbb{Z}})^g$$ for some group $\mathcal{G}^0_{p^n}$ and where the natural projection $\mathcal{A}(\overline{W})_{p^n} \longrightarrow (\operatorname{\mathbb{Z}}/p^n \operatorname{\mathbb{Z}})^g$ is identified to the reduction $${\mathrm}{red}: \mathcal{A}(\overline{W})_{p^n} \longrightarrow \mathcal{A}(\overline{W}/\overline{\mathfrak{m}})_{p^n}=A_0(\overline{\operatorname{\mathbb{F}}}_q)_{p^n} \simeq (\operatorname{\mathbb{Z}}/p^n \operatorname{\mathbb{Z}})^g.$$ As a consequence, via the identification $\mathcal{A}(\overline{W})_N=A(\overline{K})_N$ we obtain $$A(\overline{K})_N=\mathcal{G}^0_{p^{v_p(N)}} \times A_0(\overline{\operatorname{\mathbb{F}}}_q)_N$$ for any integer $N \geq 0$. Taking $N=\det(\xi^n-1)=|A_0(\operatorname{\mathbb{F}}_{q^n})|$ and passing to $\phi^n-1$ fixed modules we obtain a natural surjection $$A(\operatorname{\mathbb{C}})^{\xi^n=1}=A(\overline{K})^{\phi^n=1} \longrightarrow A_0(\operatorname{\mathbb{F}}_{q^n})$$ which is a bijection because the left and right hand groups have the same order. Combining (\[EF\]), Prop. \[eigenvalues\] and Prop. \[noeuds\] we obtain the following index theoretic way to write the explicit formula for an ordinary abelian variety of dimension $g$. \[main\] The following equality holds in the space of distributions $\mathcal{D}'(\operatorname{\mathbb{R}})$: $${\mathrm}{Ind}_t(d_{\mathcal{F}})=\sum_{\gamma} l(\gamma) \sum_{k \geq 1} \delta_{k l(\gamma)}+\sum_{\gamma} l(\gamma) \sum_{k \leq -1} e^{gkl(\gamma)}\delta_{kl(\gamma)}.$$ There is a perfect analogy between the formula above and the transversal index formula [@deninger] Thm. 2.2 except for the factor $e^{gkl(\gamma)}$ in front of Dirac distribution $\delta_{kl(\gamma)}$ for $k \leq -1$. For an explanation of this dissymetry, see Rem (2) p. 213 in [@leichtnam1] and [@deninger1] p. 18. [99]{} A. Candel, L. Conlon, [*Foliations I*]{}, Graduate Studies in Mathematics 23, AMS Providence, RI, (2003), xiv+545. G. Cornell, J. H. Silvermann (editors), [*Arithmetic geometry*]{}, Springer-Verlag, New York, (1986), xvi+353. K. Barner, [*On A. Weil’s explicit formula*]{}, J. Reine Angew. Math. 323, (1981), 139-152. P. Deligne, [*Variétés abéliennes ordinaires sur un corps fini*]{}, Invent. Math. 8, (1969), 238-243. C. Deninger, [*Some analogies between number theory and dynamical systems on foliated spaces*]{}, Doc. Math. J. DMV, Extra Volume ICM I, (1998), 23-46. C. Deninger, [*Number theory and dynamical systems on foliated spaces*]{}, In: Jber. d. dt. Math.-Verein 103, (2001), 79-100. C. Deninger, [*On the nature of explicit formulas in analytic number theory, a simple example*]{}, Number theoretic methods (Izuka 2001), Dev. Math. 8, Kluwer Acad. Publ., Dordrecht, (2002), 97-118. C. Deninger, [*Analogies between analysis on foliated spaces and arithmetic geometry*]{}, Groups and Analysis, London Math. Soc. Lecture Notes Series, Cambridge Univ. Press, Cambridge, (2008), 174-190. O. Filali, [*Über abelsche Varietäten und den transversalen Indexsatz*]{}, PhD thesis, Münster Univ., (2007), 1-84. E. Leichtnam, [*An invitation to Deninger’s work on arithmetic zeta functions*]{}, In: Geometry, Spectral Theory, Groups and Dynamics, Contemp. Math., vol. 387, Amer. Math. Soc., Providence, RI, (2005), 201-236. C.C. Moore, C. Schochet, [*Global analysis on foliated spaces*]{}, MSRI Publications 9, Springer-Verlag, New York, (1988), vi+337. D. Mumford, [*Abelian varieties*]{}, Oxford Univ. Press, London, (1970), vii+242. J.P. Serre, J. Tate, [*Good reduction of abelian varieties*]{}, Ann. of Math. Vol. 88, no. 3, (1968), 492-517. J. Tate, [*Finite flat group schemes*]{}, In: Modular forms and Fermat’s last theorem, Springer, New York, (1997) 121-154.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Gossip monoids form an algebraic model of networks with exclusive, transient connections in which nodes, when they form a connection, exchange all known information. They also arise naturally in pure mathematics, as the monoids generated by the set of all equivalence relations on a given finite set under relational composition. We prove that a number of important decision problems for these monoids (including the membership problem, and hence the problem of deciding whether a given state of knowledge can arise in a network of the kind under consideration) are NP-complete. As well as being of interest in their own right, these results shed light on the apparent difficulty of establishing the cardinalities of the gossip monoids: a problem which has attracted some attention in the last few years.' title: 'NP-Completeness in the Gossip Monoid' --- PETER FENNER[^1], MARIANNE JOHNSON[^2] and MARK KAMBITES[^3] School of Mathematics,  University of Manchester,\ Manchester M13 9PL,  England. [^4] Introduction ============ *Gossip problems* are concerned with the flow of information through networks with exclusive, transient connections in which nodes, when connected, exchange all known information. The *gossip monoid* (of rank $n$) forms an algebraic model of such a network (with $n$ nodes). It is a semigroup whose elements correspond to possible states of knowledge across the network, and in which the natural multiplication action of a generating set simulates information flow through the establishment of connections. Gossip monoids are also of considerable interest for purely mathematical reasons: they arise naturally in semigroup theory as the monoids generated by the set of equivalence relations on a finite set, under the operation of relational composition. Since equivalence relations are idempotent, this means gossip monoids are a natural and interesting family of *idempotent-generated monoids*, the latter being an area of great current interest in semigroup theory (see for example [@BriMarMea; @GraRus1; @GraRus2; @DolGra; @YanDolGou].) Despite the obvious importance of gossip monoids, both within semigroup theory and for applications, remarkably little is known about them. In the 1970s, a number of authors, including Tijdeman [@Tijdeman], Baker and Shostak [@BakSho], Hajnal, Milner and Szemeredi [@HajMilSze], independently computed the minimum number of two-way connections required to ensure permeation of all information throughout the network; in semigroup-theoretic terms this is the length of the zero element as a word with respect to a particularly natural set of generators. More recently, Brouwer, Draisma and Frenk [@BroDraFre; @Frenk] have introduced a continuous (tropical) analogue, termed the *lossy gossip monoid*, which has some intriguing connections to composition of metrics and the tropical geometry of the orthogonal group. Perhaps the most obvious question in this area is: how many elements are there in the gossip monoid of rank $n$? In other words, how many distinct knowledge configurations can arise in an $n$-node network of the type discussed above? In [@BroDraFre] this number was calculated computationally for $n \leq 9$, but no obvious pattern emerges from their results, and the hope of extending the sequence much further by similar brute-force computations seems remote. It remains open whether there is an explicit formula for, or even a markedly faster way of counting, the cardinality of the gossip monoid of rank $n$. In this paper, we consider the complexity of decision problems concerning gossip monoids. Elements of gossip monoids are naturally represented by boolean matrices, and an important problem is to decide, given a boolean matrix, whether it represents a gossip element. One of our main results is that this problem is NP-complete. As well as being of interest in its own right, this sheds some light on the apparent difficulty of determining the cardinality of the gossip monoid, by suggesting that there is no simple combinatorial characterisation of elements, of the kind which might be used to count them. We also establish NP-completeness for a number of other important problems, including Green’s $\mathcal{J}$-order. Besides this introduction, the paper is organised into five sections. Section \[sec\_gossip\] recalls the definition and basic properties of gossip monoids. Section \[sec\_problems\] introduces the algorithmic problems we study, and describes in outline our strategy for establishing their complexity. Sections 4-6 are concerned with the proofs of NP-hardness for these problems. The Gossip Monoid {#sec_gossip} ================= In this section we briefly introduce the gossip monoids and their relationship to knowledge distribution in networks, and discuss some semigroup-theoretic properties of these monoids. Consider $n$ people (“gossips”) each of whom knows a unique piece of information (a “scandal”) initially unknown to the others. The people communicate by telephone and in each call the two participants tell each other every scandal they know. The ‘gossip problem’ (what is the minimum number of calls required before every person knows every scandal?) attracted the attention of a number of researchers in the 1970s, including Tijdeman [@Tijdeman], Baker and Shostak [@BakSho], Hajnal, Milner and Szemeredi [@HajMilSze], who proved in a variety of different ways that the minimal number of calls required is: $$\begin{aligned} 0 \;\text{ if }\; n = 1, \qquad 1 \;\text{ if }\; n = 2, \qquad 3 \;\text{ if }\; n = 3, \qquad \text{ and } 2n-4 \;\text{ if }\; n \geq 4.\end{aligned}$$ Whilst the gossip problem concerns only the most efficient means to create one particular state of knowledge between the participants, it is clear that understanding the ‘gossip state-space’ (in other words all possible states of knowledge that can occur in this system) is both a more complicated and more important problem; it is equivalent to understanding the distribution of knowledge within an $n$-node communication network with transient connections, assuming that initially each node possesses a unique piece of information, and that whenever a connection is made nodes exchange all known information. We begin by showing that this system can be modelled by considering the right action of a particular monoid (the gossip monoid) on the set of all $n \times n$ boolean matrices, which we use to record these states of knowledge. We write $\mathbb{B}$ for the *boolean semiring*, that is, the algebraic structure comprising the set $\lbrace 0, 1 \rbrace$ equipped with the operations of maximum (logical “or”) and multiplication (logical “and”.) Algebraically, this structure is somewhat like a commutative ring, with maximum playing the role of addition, the fundamental difference being that the addition operation is non-invertible. The set $\mathbb{B}_n$ consisting of all $n \times n$ boolean matrices (that is, matrices with entries from $\mathbb{B}$) forms a monoid under the matrix multiplication induced from the operations in $\mathbb{B}$. (In other words, for $A, B \in \mathbb{B}_n$ the $(i,j)$th entry of the product $AB$ is obtained by taking the *maximum* over $k$ of the products $a_{i,k}b_{k,j}$. It is easy to see that the ‘usual’ identity matrix also behaves as an identity with respect to this new multiplication.) Moreover, there is a natural partial order on $\mathbb{B}_n$ induced from the natural order on $\mathbb{B}$; for $A, B \in \mathbb{B}_n$ we say that $A \preceq B$ if and only if $a_{i,j} \leq b_{i,j}$ for all indices $i,j$. We will then write $A \prec B$ to mean $A \preceq B$ but $A \neq B$. Each matrix in $\mathbb{B}_n$ can be thought of as representing a state of knowledge distribution within an $n$-node network: the entry in row $i$ and column $j$ being $1$ exactly if node $j$ (or “person $j$”, in gossip problem terminology) has learnt the knowledge initially possessed by node $i$ (“scandal $i$”.) In particular, the identity matrix $I_n$ corresponds to the initial state of knowledge (each node knowing only what it knows initially.) For $i, j \in \lbrace 1, \dots, n \rbrace$ the *call matrix* $C[i,j]$ is the matrix with $1$s on the main diagonal, and also in the $(i,j)$ and $(j,i)$ positions, and $0$s elsewhere. It is readily verified that if $K \in \mathbb{B}_n$ represents a state of knowledge, then the product $K C[i,j]$ represents the state of knowledge resulting by starting in the state represented by $K$ and exchanging all information (a “call”) between node $i$ and node $j$. The *gossip monoid* $G_n$ of rank $n$ is the submonoid of $\mathbb{B}_n$ generated by the set of all call matrices, $$\lbrace C[i,j] \mid i, j \in \lbrace 1, \dots, n \rbrace \rbrace.$$ Its elements are exactly those matrices modelling states of knowledge which can actually arise in a network of the kind under consideration, starting from the initial state of knowledge and proceeding through calls in which two nodes exchange all information available to them. In matrix terms, the effect of right multiplication by $C[i,j]$ is to replace columns $i$ and $j$ with their element-wise maximum. Notice in particular that this action is *monotonic*, in the sense that $K \preceq KC[i,j]$; this corresponds to the fact that nodes do not forget things once learnt. (Dually, one can see that the effect of *left* multiplication by $C[i,j]$ is to replace *rows* $i$ and $j$ with their element-wise maximum, and so it is clear that this left action is also monotonic, that is $K \preceq C[i,j]K$.) Notice also that this multiplication respects the partial order, in that $K \preceq L$ implies $KC[i,j] \preceq LC[i,j]$ and $C[i,j]K \preceq C[i,j]L$. Repeated use of these two properties gives us $AC \preceq ABC$ for all $A,B,C \in G_n$. Recall that *Green’s relations* [@Green; @CliPre] are five equivalence relations (${\mathcal{R}}$, ${\mathcal{L}}$, ${\mathcal{H}}$, ${\mathcal{D}}$ and ${\mathcal{J}}$) and three pre-orders (${\mathrel{\leq_\mathcal{R}}}$, ${\mathrel{\leq_\mathcal{L}}}$ and ${\mathrel{\leq_\mathcal{J}}}$) which can be defined on any monoid (or semigroup) and play a fundamental role in understanding its structure. In the case of $G_n$, the five equivalence relations are easily seen to be trivial — we say that $G_n$ is a *${\mathcal{J}}$-trivial* monoid — but the pre-orders remain important and we recall their definitions. We define $A {\mathrel{\leq_\mathcal{R}}}B$ \[respectively, $A {\mathrel{\leq_\mathcal{L}}}B$\] if there exists a $C \in G_n$ with $A = BC$ \[respectively, $A = CB$\], and $A {\mathrel{\leq_\mathcal{J}}}B$ if there exists $D, E \in G_n$ with $A = DBE$. Notice that $A {\mathrel{\leq_\mathcal{R}}}B$ if and only if the state of knowledge represented by matrix $A$ can be obtained by starting with the state of knowledge represented by matrix $B$ and applying a sequence of calls. It is straightforward to verify (as we shall explain below) that the call matrices satisfy the following relations for all distinct values $i,j,k,l$: $$\begin{aligned} \label{idmpt}C[i,j] \ C[i,j] &\ = \ & C[i,j]\\ \label{commute}C[i,j] \ C[k,l]&\ = \ & C[k,l]C[i,j]\\ \label{transfer}C[i,j] \ C[j,k] \ C[i,j]& \ = \ & C[j,k] \ C[i,j] \ C[j,k].\end{aligned}$$ In terms of gossip, the first set of relations simply record the fact if two people make a repeat call when nothing has occurred in the interim then no information is exchanged in the second call. The second set of relations corresponds to the fact that if the two pairs of callers $\{i,j\}$ and $\{k,l\}$ are disjoint, then it doesn’t matter in what order the calls between these pairs take place. (In the situation we are modelling, these communications could in fact occur concurrently.) The third set of relations, which are a kind of *braid relation*, can easily be verified by matrix multiplication. It follows immediately from the relations above that $G_n$ is the homomorphic image of an infinite 0-Hecke monoid (see [@MazSte]) corresponding to the Coxeter presentation with $n\choose 2$ involutions $c_{i,j}$ satisfying relations analogous to and above. We also note that the *double Catalan monoid* $DC_n$ studied by Mazorchuk and Steinberg [@MazSte] is the submonoid of $G_n$ generated by the call matrices $C[i,i+1]$ for $i=1, \ldots, n-1$; in fact the latter can be thought of as an algebraic model of *linear* networks with exclusive, transient connections in which nodes, when they form a connection, exchange all known information. The set of all call matrices is a particularly natural set of idempotent generators, which may be generalised as follows. If $S \subseteq \lbrace 1, \dots, n \rbrace$ then the *conference call matrix* $C[S]$ is the matrix with $1$s on the main diagonal and in the $(i,j)$ position for all $i, j \in S$, and $0$s elsewhere. In terms of knowledge distribution, right multiplication by $C[S]$ models a complete exchange of knowledge between the nodes in $S$; in matrix terms it replaces each column whose index is in $S$ with the element-wise maximum of all columns whose indices are in $S$. (The action of *left* multiplication by $C[S]$, corresponding to replacing each *row* whose index is in $S$ with the maximum of all *rows* whose indices are in $S$, is algebraically dual but less easy to visualise in terms of networks.) It is easy to see that every conference call matrix is an idempotent element of $G_n$; indeed it follows from the solution to the original gossip problem [@BakSho; @HajMilSze; @Tijdeman] that $C[S]$ is a product of $3$ call matrices if $|S| = 3$ and $2 |S|-4$ call matrices if $|S| \geq 4$, these numbers of call matrices being the minimum possible. We may associate to each element of $\mathbb{B}_n$ the binary relation on the set $\lbrace 1, \dots, n \rbrace$ of which it is the adjacency matrix, that is, the relation where $i$ is related to $j$ if and only if the $(i,j)$ entry of the matrix is $1$; matrix multiplication in $\mathbb{B}_n$ then corresponds to *relational composition* of binary relations. It is easy to see that every call matrix (and indeed every conference call matrix) is the adjacency matrix of an equivalence relation. In fact, equivalence relations correspond exactly to *idempotent* elements of $G_n$: \[Prop\] The idempotents in $G_n$ are precisely the matrices in $\mathbb{B}_n$ which correspond to equivalence relations on $\{1,\ldots, n\}$. (In other words, $A \in G_n$ is an idempotent if and only if there exists an equivalence relation $\sim$ on $\{1, \ldots, n\}$ such that $a_{i,j} = 1 \Leftrightarrow i \sim j$.) First let $A \in G_n$ be an idempotent and let $\sim$ denote the binary relation on $\{1, \ldots, n\}$ given by $i \sim j$ if and only if $a_{i,j} = 1$. Since $A$ is a product of call matrices, it is clear that $a_{i,i}=1$ and hence $i \sim i$ for all $i$. Since $A$ is idempotent we note that $a_{i,j} = \max_k\{a_{i,k}a_{k,j}\}$, and hence whenever $i\sim k$ and $k \sim j$, we must also have $i\sim j$. To see that $\sim$ is symmetric we note that $A$ can be written as a product of call matrices, $A = A_1 \cdots A_m$ for some $m$, and since $A$ is idempotent we have $A^m = A$. We can therefore write $A$ as a product which contains $A_m, \ldots , A_1$ as a scattered subsequence. Since right multiplication by call matrix $C[i,j]$ replaces columns $i$ and $j$ with their maximum and left multiplication by the same matrix replaces rows $i$ and $j$ with their maximum, we see that $A_m \cdots A_1 = A^T$. We therefore have $A^T \preceq A$, and so if $i \sim j$ then $j \sim i$. For the converse, let $\sim$ be an equivalence relation on $\{1, \ldots, n\}$ and let $A$ be the corresponding adjacency matrix. It is easy to see that $A$ is the product (in any order) of the conference call matrices corresponding to the equivalence classes of $\sim$, so that $A \in G_n$. It remains to show that $A$ is an idempotent. It follows from reflexivity of $\sim$ that $B := A^2 \succeq A$, since $b_{i,j} \geq a_{i,j}a_{j,j}$. Finally we note that if $b_{i,j}=1$, then there exists $k$ such that $a_{i,k}a_{k,j}=1$. By transitivity of $\sim$ we conclude that $a_{i,j}=1$, hence showing that $A^2=A$. $G_n$ is the submonoid of $\mathbb{B}_n$ generated by the adjacency matrices of equivalence relations. One containment follows from the fact that call matrices (which by definition generate $G_n$) are adjacency matrices of equivalence relations, and the other from the fact that, by Proposition \[Prop\], $G_n$ contains all adjacency matrices of equivalence relations. Throughout this paper, when the sizes are understood we write $1$ for the matrix consisting entirely of ones, $0$ for the zero matrix and $I_n$ for the identity matrix. Similarly, we shall write $\underline{0}$ to denote the vector of zeros and $\underline{1}$ to denote the vector of all ones. Algorithmic Problems in the Gossip Monoid {#sec_problems} ========================================= We shall consider the following decision problems concerning the gossip monoids: - the *Gossip Membership Problem* (GMP) is to determine whether a given $A \in \mathbb{B}_n$ is a member of $G_n$; - the *Gossip $\mathcal{J}$-Order Problem* (GJP) is to determine, given matrices $X, Y \in G_n$, whether $X \leq_\mathcal{J} Y$, that is, whether there exist matrices $U,V \in G_n$ with $UYV = X$; - the *Gossip Transformation Problem* (GTP) is to determine, given matrices $X, Y \in \mathbb{B}_n$, whether there exists a matrix $G \in G_n$ such that $X G = Y$; - the *Maximal Gossip Transformation Problem* (MGTP) is the restriction of GTP to pairs of matrices $X$, $Y$ where the matrix $X$ satisfies the *maximal column condition*: $X$ is non-zero and the set of distinct columns of $X$ form an anti-chain. (In other words, every column is non-zero and maximal among the set of columns.) Of these problems, the first three are of clear importance in their own right. GMP is critical for understanding the gossip monoid, and also has an obvious application to deciding whether a given knowledge configuration is (absolutely) reachable in a network. GTP corresponds to understanding orbits under the natural action of $G_n$ on $\mathbb{B}_n$, and hence whether one given knowledge configuration is reachable from another. The $\mathcal{J}$-order is the key to understanding the ideal structure of a $\mathcal{J}$-trivial monoid, and so GJP is essential for understanding the semigroup-theoretic structure of $G_n$. The final problem, MGTP, is less natural but is included because it functions as a stepping stone in the proof of NP-hardness for GMP. By monotonicity, and using the fact that there are only $n(n-1)$ zeros in the identity matrix, we see that there is a polynomial bound on the length of elements of $G_n$ as a product of the (quadratically many) call generators. (In [@BroDraFre] it is shown that this polynomial bound can be lowered to $n(n-1)/2$.) It is therefore possible for a non-deterministic polynomial time computation to guess an element of $G_n$. This clearly suffices to show that all of the above problems are in NP. The remainder of the article is therefore concerned with establishing NP-hardness. We first show that MGTP (and hence also GTP) is NP-hard, by a relatively straightforward reduction from the *Dominating Set Problem*. Then we show that GJP is NP-hard by a reduction from GTP. Finally, GMP is shown to be NP-hard by a (rather more complex) reduction from MGTP. We briefly recall the definition of the Dominating Set Problem. Given an undirected graph $H$ with vertex set $V$ and edge set $E \subseteq V \times V$, we say that $D \subseteq V$ is a *dominating set* for $H$ if every vertex is either in $D$ or adjacent to a vertex in $D$. Given a graph $H$ with vertex set $V$ and a positive integer $k \leq |V|$, the Dominating Set Problem is to decide whether $H$ admits a dominating set of size at most $k$. The problem is known to be NP-complete [@GarJoh79 Problem GT2, Appendix A1] via a transformation from Vertex Cover. (See for example [@Jungnickel Theorem 2.8.6] for details.) NP-completeness of the Gossip Transformation Problems ===================================================== In this section we show that the Gossip Transformation Problem and the Maximal Gossip Transformation Problem are NP-complete. We show the NP-hardness of these problems with a polynomial time reduction from the Dominating Set Problem. \[transthm\] GTP and MGTP are NP-complete. We have already seen that both are in NP, and since MGTP is a restricted version of GTP it suffices to show that MGTP is NP-hard. As discussed above, we do this by reduction from the Dominating Set Problem. Let $H$ be a graph, say with vertex set $V = \{1, \ldots, n\}$, and let $1 \leq k \leq n$. We will construct matrices $A,B \in \mathbb{B}_{3n}$, with $A$ satisfying the maximal column condition, such that $AG = B$ for some $G \in G_{3n}$ if and only if $H$ admits a dominating set of size at most $k$. If $k = n$, then $V$ is a dominating set with size at most $k$, so we let $A = B = I_{2n}$. Otherwise, define a matrix $M \in \mathbb{B}_n$ by $$m_{i,j} = 1 \ \iff \ (i = j) \textrm{ or } (i \textrm{ and } j \textrm{ are adjacent in } H.)$$ It is easy to see that $D \subseteq \lbrace 1, \dots, n \rbrace$ is a dominating set for $H$ if and only if $\max\limits_{j \in D} (m_{i,j}) = 1$ for every $i$, so there exists a dominating set of size at most $k$ if and only if there exists a set of $k$ or fewer columns of $M$ whose maximum is $\underline{1}$. Clearly, such a set of columns exists if and only if there is a set of $k$ or fewer *maximal* columns of $M$ whose maximum is $\underline{1}$. $M$ has at least one maximal column, $m$. Replace each non-maximal column of $M$ with $m$ and call the resulting matrix $M'$. This ensures that $M'$ satisfies the maximal column condition but no columns are added which are not already columns of $M$. Note that finding the non-maximal column vectors only requires $n(n-1)/2$ vector comparisons, so $M'$ can be computed in polynomial time. Let (1mat) [ & & & & &\ & & & & &\ & & & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-6.south east); (1mat-2-1.south west) – (1mat-2-6.south east); (1mat-1-2.north east) – (1mat-3-2.south east); (1mat-1-4.north east) – (1mat-3-4.south east); at (1mat-1-1.east) [$M'$]{}; at (1mat-1-3.east) [$0$]{}; at (1mat-1-5.east) [$0$]{}; at (1mat-2-1.east) [$0$]{}; at (1mat-2-3.east) [$1$]{}; at (1mat-2-5.east) [$1$]{}; at (1mat-3-1.east) [$0$]{}; at (1mat-3-3.east) [$0$]{}; at (1mat-3-5.east) [$0$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-5.north west) – node\[above=8pt\] [$n$]{} (1mat-1-6.north east); (1mat-1-6.north east) – node\[right=15pt\] [$n$]{} (1mat-1-6.south east); (1mat-2-6.north east) – node\[right=15pt\] [$n$]{} (1mat-2-6.south east); (1mat-3-6.north east) – node\[right=15pt\] [$n$]{} (1mat-3-6.south east); at (1mat-2-1.west) [$A = $]{}; (2mat) [ & & & & &\ & & & & &\ & & & & &\ ]{}; (2mat-1-1.south west) – (2mat-1-6.south east); (2mat-2-1.south west) – (2mat-2-6.south east); (2mat-1-2.north east) – (2mat-3-2.south east); (2mat-1-4.north east) – (2mat-3-4.south east); (2mat-1-5.north east) – (2mat-1-5.south east); at (2mat-1-1.east) [$M'$]{}; at (2mat-1-3.east) [$M'$]{}; at (2mat-1-5) [$1$]{}; at (2mat-1-6) [$0$]{}; at (2mat-2-1.east) [$1$]{}; at (2mat-2-3.east) [$1$]{}; at (2mat-2-5.east) [$1$]{}; at (2mat-3-1.east) [$0$]{}; at (2mat-3-3.east) [$0$]{}; at (2mat-3-5.east) [$0$]{}; (2mat-1-1.north west) – node\[above=8pt\] [$n$]{} (2mat-1-2.north east); (2mat-1-3.north west) – node\[above=8pt\] [$n$]{} (2mat-1-4.north east); (2mat-1-5.north west) – node\[above=8pt\] [$k\;$]{} (2mat-1-5.north east); (2mat-1-6.north west) – node\[above=8pt\] [$\;\;\;n-k$]{} (2mat-1-6.north east); (2mat-1-6.north east) – node\[right=15pt\] [$n$]{} (2mat-1-6.south east); (2mat-2-6.north east) – node\[right=15pt\] [$n$]{} (2mat-2-6.south east); (2mat-3-6.north east) – node\[right=15pt\] [$n$]{} (2mat-3-6.south east); at (2mat-2-1.west) [and $B = $]{}; at (2mat-2-6.east) [.]{}; Note that the matrix $A$ satisfies the maximal column condition since $M'$ does, and $A$ and $B$ give an instance of the problem $MGTP$. If $H$ admits a dominating set of size $k$ or less then there is a set of $k$ or fewer columns of $M'$ whose maximum is $\underline{1}$. We can multiply matrix $A$ on the right by call matrices which copy each of these columns into a unique column between $2n+1$ and $2n+k$. These call matrices will not affect the first $n$ rows of columns $1, \dots, n$ but will put a 1 in each of the first $n$ rows in columns $2n+1, \dots, 2n+k$. Multiplying this product on the right by the conference call matrix $C[\lbrace 2n+1, \dots, 2n+k\rbrace]$ and by each of the call matrices $C[i, n+i]$ for $1 \leq i \leq n$, in any order, will therefore result in matrix $B$. The product of these call matrices is thus an element $G \in G_{3n}$ such that $AG = B$. Conversely, if there is a $G \in G_{3n}$ such that $AG = B$, then there is a sequence of call matrices whose product is $G$, say $G = C_1 C_2 \cdots C_q$. Without loss of generality we may clearly assume that $A C_1 \cdots C_p \prec A C_1 \cdots C_p C_{p+1}$ for all $p$. Since right multiplication by a call matrix has the effect of replacing two columns with their maximum, it is clear that each column of each product $A C_1 \cdots C_p$ must be equal to the maximum of some collection of columns of $A$. In particular, the first $n$ rows of each column must equal the maximum of some collection of columns of $M'$. In the product $B = A C_1 \cdots C_q$ the first $n$ rows of column $n+i$, for $1 \leq i \leq n$, equal the $i$th column of $M'$. By monotonicity and the fact that $M'$ satisfies the maximal column condition, the first $n$ rows of column $n+i$ must equal either $0$ or the $i$th column of $M'$ for every product $A C_1 \cdots C_p$. For each $r \in \{2n + 1, \ldots, 2n + k\}$, there is at most one matrix $C_p$ of the form $C[i,r]$ such that $i \leq 2n$ and the two products $A C_1 \cdots C_{p-1}$ and $A C_1 \cdots C_p$ differ on column $r$. Assume for contradiction that the sequence $C_1,\dots,C_q$ contains call matrices $C_s = C[i,r]$ and $C_t = C[j,r]$, in that order, with $i,j \leq 2n$ and $2n+1 \leq r \leq 2n+k$, such that the products $AC_1\cdots C_{s-1}$ and $AC_1\cdots C_s$ differ on column $r$ and the products $AC_1\cdots C_{t-1}$ and $AC_1\cdots C_t$ also differ on column $r$. We observe that the last $2n$ rows of column $r$ are identical in $A$ and $B$, so by monotonicity the differences in column $r$ occur in the first $n$ rows. Let $v_s, v_t \in \mathbb{B}^n$ be the first $n$ rows of column $r$ in the matrices $AC_1 \cdots C_s$ and $AC_1 \cdots C_t$, respectively. Since the differences occur in the first $n$ rows, and by monotonicity, we have $0 \prec v_s \prec v_t$. As the last call matrix in the product $AC_1 \cdots C_s$ is $C_s = C[i,r]$, columns $i$ and $r$ must be identical in this matrix, and so $v_s$ is also equal to the first $n$ rows of column $i$. If $i \leq n$ then $v_s$ is equal to column $i$ of $M'$. Otherwise $n < i \leq 2n$ and since $v_s$ is non-zero it is equal to the $(i-n)$th column of $M'$. Either way, $v_s$ is equal to a column of $M'$ and similar observations show that $v_t$ is also equal to a column of $M'$. Now $v_s \prec v_t$ contradicts the maximal column condition on $M'$, and this proves the claim. It follows from the claim that right-multiplying $A$ by the product $C_1 \cdots C_q$ copies at most $k$ of the columns of $M'$ into the columns $2n+1, \dots, 2n+k$. But since the result of this right multiplication is to place $1$s in the first $n$ rows of these columns, it must be that $\underline{1}$ is a maximum of at most $k$ columns of $M'$. It follows from our observations that $H$ admits a dominating set of size at most $k$. NP-Completeness of the Gossip $\mathcal{J}$-order Problem ========================================================= To show GJP is NP-hard we will give a polynomial time reduction from GTP. This reduction uses the following lemma, which allows us to “nest” an arbitrary boolean matrix inside a gossip matrix with polynomially larger size: \[boolingoss\] Let $n\geq 2$. For any $A \in \mathbb{B}_n$, the following matrix lies in $G_{n(n+1)}$: (1mat) [ & &\ & &\ & &\ ]{}; (1mat-1-1.south west) – (1mat-1-3.south east); (1mat-1-2.north east) – (1mat-3-2.south east); at (1mat-1-2.west) [$1$]{}; at (1mat-1-3) [$A$]{}; at (1mat-3-2.north west) [$1$]{}; at (1mat-3-3.north) [$1$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n^{2}$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n$]{} (1mat-1-3.north east); (1mat-1-3.north east) – node\[right=15pt\] [$n$]{} (1mat-1-3.south east); (1mat-2-3.north east) – node\[right=15pt\] [$n^{2}$]{} (1mat-3-3.south east); at (1mat-2-1.west) [$X = $]{}; We shall show that $X$ can be written as $X = X_1 X_2 X_3 X_4$, where the matrices $X_i$ are defined in terms of conference call matrices as follows $$\begin{aligned} X_1 = C[\{n+1, \ldots, n(n+1)\}],\qquad\;\;\;\;\; && X_4 = C[\{1, \ldots, n^2\}],\\ X_2 = \prod_{i = 1}^n C[\{i + n(j-1) : 1 \leq j \leq n\}], && X_3 = \prod_{j = 1}^n C[\{i + n(j-1) : a_{i,j} = 1\} \cup \{n^2 + j\}], \qquad\end{aligned}$$ and where the products taken in $X_2$ and $X_3$ can be in any order. Each matrix in the product $X_2$ is of the form $C[\{i + n(j-1) : 1 \leq j \leq n\}]$ for some $i$. This matrix has a $1$ in the $a,b$ position if and only if either $a=b$, or $a,b \leq n^2$ and $a$ and $b$ are both congruent to $i$ modulo $n$. The product $X_2$ therefore has a $1$ in the $a,b$ position if and only if either $a=b$, or $a,b \leq n^2$ and $a$ and $b$ are congruent modulo $n$. The result is that the top-left $n^2 \times n^2$ block of $X_2$ consists of an $n \times n$ array of copies of $I_n$: (1mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-3.south east); (1mat-3-1.south west) – (1mat-3-4.south east); (1mat-1-1.north east) – (1mat-3-1.south east); (1mat-1-3.north east) – (1mat-4-3.south east); (1mat-1-2.north east) – (1mat-3-2.south east); (1mat-2-1.south west) – (1mat-2-3.south east); at (1mat-1-1) [$I_n$]{}; at (1mat-1-2) […]{}; at (1mat-1-3) [$I_n$]{}; at (1mat-2-1) [$\vdots$]{}; at (1mat-2-2) [$\ddots$]{}; at (1mat-2-3) [$\vdots$]{}; at (1mat-2-4) [$0$]{}; at (1mat-3-1) [$I_n$]{}; at (1mat-3-2) [$\ldots$]{}; at (1mat-3-3) [$I_n$]{}; at (1mat-4-2) [$0$]{}; at (1mat-4-4) [$I_n$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n^2$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-4.north east) – node\[right=15pt\] [$n^2$]{} (1mat-3-4.south east); (1mat-4-4.north east) – node\[right=15pt\] [$n$]{} (1mat-4-4.south east); at (1mat-3-1.north west) [$X_2 = $]{}; at (1mat-3-4.north east) [$.$]{}; For $1 \leq j \leq n$ let $S_j = \{i : a_{i,j} = 1\} \subseteq \{1, \ldots, n\}$ and let $A_{[j]}$ be the matrix which is the same as $A$ on column $j$ but zero elsewhere. For each $j$ we have (1mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-2.south east); (1mat-1-4.south west) – (1mat-1-4.south east); (1mat-2-2.south west) – (1mat-2-4.south east); (1mat-3-1.south west) – (1mat-3-4.south east); (1mat-1-1.north east) – (1mat-2-1.south east); (1mat-4-1.north east) – (1mat-4-1.south east); (1mat-2-2.north east) – (1mat-4-2.south east); (1mat-1-3.north east) – (1mat-4-3.south east); at (1mat-1-1) [$I_{n(j-1)}$]{}; at (1mat-1-3) [$0$]{}; at (1mat-1-4) [$0$]{}; at (1mat-2-2) [$C[S_j]$]{}; at (1mat-2-4) [$A_{[j]}$]{}; at (1mat-3-1) [$0$]{}; at (1mat-3-3) [$\;I_{n(n-j)}$]{}; at (1mat-3-4) [$0$]{}; at (1mat-4-1) [$0$]{}; at (1mat-4-2) [$(A_{[j]})^T$]{}; at (1mat-4-3) [$0$]{}; at (1mat-4-4) [$I_n$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n(j-1)$]{} (1mat-1-1.north east); (1mat-1-2.north west) – node\[above=8pt\] [$n$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n(n-j)$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-4.north east) – node\[right=15pt\] [$n(j-1)$]{} (1mat-1-4.south east); (1mat-2-4.north east) – node\[right=15pt\] [$n$]{} (1mat-2-4.south east); (1mat-3-4.north east) – node\[right=15pt\] [$n(n-j)$]{} (1mat-3-4.south east); (1mat-4-4.north east) – node\[right=15pt\] [$n$]{} (1mat-4-4.south east); at (1mat-3-1.north west) [$C[\{n(j-1) + i : i \in S_j\} \cup \{n^2 + j\}] = $]{}; at (1mat-3-4.north east) [$.$]{}; Notice that, as $j$ varies, the corresponding conference call matrices are between disjoint sets of nodes. The product of all such matrices, as $j$ ranges between $1$ and $n$, is therefore (1mat) [ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-2.south east); (1mat-1-5.south west) – (1mat-1-5.south east); (1mat-2-2.south west) – (1mat-2-3.south east); (1mat-2-5.south west) – (1mat-2-5.south east); (1mat-3-3.south west) – (1mat-3-5.south east); (1mat-4-1.south west) – (1mat-4-5.south east); (1mat-1-1.north east) – (1mat-2-1.south east); (1mat-5-1.north east) – (1mat-5-1.south east); (1mat-2-2.north east) – (1mat-3-2.south east); (1mat-5-2.north east) – (1mat-5-2.south east); (1mat-3-3.north east) – (1mat-5-3.south east); (1mat-1-4.north east) – (1mat-5-4.south east); at (1mat-1-1) [$C[S_1]$]{}; at (1mat-1-3.south east) [$0$]{}; at (1mat-1-5) [$A_{[1]}$]{}; at (1mat-2-2) [$C[S_2]$]{}; at (1mat-2-5) [$A_{[2]}$]{}; at (1mat-3-1.south east) [$0$]{}; at (1mat-3-3) [$\ddots$]{}; at (1mat-3-5) [$\vdots$]{}; at (1mat-4-4) [$C[S_n]$]{}; at (1mat-4-5) [$A_{[n]}$]{}; at (1mat-5-1) [$(A_{[1]})^T$]{}; at (1mat-5-2) [$(A_{[2]})^T$]{}; at (1mat-5-3) [$\ldots$]{}; at (1mat-5-4) [$(A_{[n]})^T$]{}; at (1mat-5-5) [$I_n$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n^2$]{} (1mat-1-4.north east); (1mat-1-5.north west) – node\[above=8pt\] [$n$]{} (1mat-1-5.north east); (1mat-1-5.north east) – node\[right=15pt\] [$n^2$]{} (1mat-4-5.south east); (1mat-5-5.north east) – node\[right=15pt\] [$n$]{} (1mat-5-5.south east); at (1mat-3-1.west) [$X_3 = $]{}; at (1mat-3-5.east) [.]{}; Consider the effect of multiplying $X_3$ on the left by $X_2$. This leaves the last $n$ rows unchanged. The remaining $n^2$ rows can be split into $n$ blocks of $n$ rows each. The structure of the identity matrices in $X_2$ means that in $X_2 X_3$ these blocks are all identical to each other and equal to the element-wise maximum of the $n$ corresponding blocks in $X_3$. Since the element-wise maximum of $A_{[1]}$ to $A_{[n]}$ is $A$, we get (1mat) [ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-5.south east); (1mat-2-1.south west) – (1mat-2-5.south east); (1mat-3-1.south west) – (1mat-3-5.south east); (1mat-4-1.south west) – (1mat-4-5.south east); (1mat-1-1.north east) – (1mat-5-1.south east); (1mat-1-2.north east) – (1mat-5-2.south east); (1mat-1-3.north east) – (1mat-5-3.south east); (1mat-1-4.north east) – (1mat-5-4.south east); at (1mat-1-1) [$C[S_1]$]{}; at (1mat-1-2) [$C[S_2]$]{}; at (1mat-1-3) [$\ldots$]{}; at (1mat-1-4) [$C[S_n]$]{}; at (1mat-1-5) [$A$]{}; at (1mat-2-1) [$C[S_1]$]{}; at (1mat-2-2) [$C[S_2]$]{}; at (1mat-2-4) [$C[S_n]$]{}; at (1mat-2-5) [$A$]{}; at (1mat-3-1) [$\vdots$]{}; at (1mat-3-3) [$\ddots$]{}; at (1mat-3-5) [$\vdots$]{}; at (1mat-4-1) [$C[S_1]$]{}; at (1mat-4-2) [$C[S_2]$]{}; at (1mat-4-4) [$C[S_n]$]{}; at (1mat-4-5) [$A$]{}; at (1mat-5-1) [$(A_{[1]})^T$]{}; at (1mat-5-2) [$(A_{[2]})^T$]{}; at (1mat-5-3) [$\ldots$]{}; at (1mat-5-4) [$(A_{[n]})^T$]{}; at (1mat-5-5) [$I_n$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n^2$]{} (1mat-1-4.north east); (1mat-1-5.north west) – node\[above=8pt\] [$n$]{} (1mat-1-5.north east); (1mat-1-5.north east) – node\[right=15pt\] [$n^2$]{} (1mat-4-5.south east); (1mat-5-5.north east) – node\[right=15pt\] [$n$]{} (1mat-5-5.south east); at (1mat-3-1.west) [$X_2 X_3 = $]{}; at (1mat-3-5.east) [.]{}; Between them, the last $n^2$ rows of $X_2 X_3$ contain a $1$ in each column (each conference call has only $1$s on its diagonal.) Therefore, multiplying on the left by $X_1 = C[\{n+1, \ldots, n(n+1)\}]$ fills each of these rows with $1$s. Now, between them, the first $n^2$ columns of $X_1 X_2 X_3$ contain a $1$ in each row, so multiplying on the right by $X_4 = C[\{1, \ldots, n^2\}]$ fills each of these columns with $1$s. We therefore obtain $X_1X_2X_3X_4=X$, as required. GJP is NP-complete. We have already seen that the problem is in NP. We show the problem is NP-hard by reduction from MGTP. Given matrices $A, B \in \mathbb{B}_n$, let (1mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (1mat-2-1.south west) – (1mat-2-4.south east); (1mat-1-2.north east) – (1mat-4-2.south east); (1mat-1-3.south west) – (1mat-1-4.south east); (1mat-1-3.north east) – (1mat-2-3.south east); at (1mat-1-2.south west) [$1$]{}; at (1mat-1-3) [$A$]{}; at (1mat-1-4) [$I_n$]{}; at (1mat-2-3) [$0$]{}; at (1mat-2-4) [$1$]{}; at (1mat-3-2.south west) [$1$]{}; at (1mat-3-3.south east) [$1$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$(2n)^2$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-4.north east) – node\[right=15pt\] [$n$]{} (1mat-1-4.south east); (1mat-2-4.north east) – node\[right=15pt\] [$n$]{} (1mat-2-4.south east); (1mat-3-4.north east) – node\[right=15pt\] [$(2n)^2$]{} (1mat-4-4.south east); at (1mat-3-1.north west) [$X = $]{}; (2mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (2mat-2-1.south west) – (2mat-2-4.south east); (2mat-1-2.north east) – (2mat-4-2.south east); (2mat-1-3.south west) – (2mat-1-4.south east); (2mat-1-3.north east) – (2mat-2-3.south east); at (2mat-1-2.south west) [$1$]{}; at (2mat-1-3) [$B$]{}; at (2mat-1-4) [$I_n$]{}; at (2mat-2-3) [$0$]{}; at (2mat-2-4) [$1$]{}; at (2mat-3-2.south west) [$1$]{}; at (2mat-3-3.south east) [$1$]{}; (2mat-1-1.north west) – node\[above=8pt\] [$(2n)^2$]{} (2mat-1-2.north east); (2mat-1-3.north west) – node\[above=8pt\] [$n$]{} (2mat-1-3.north east); (2mat-1-4.north west) – node\[above=8pt\] [$n$]{} (2mat-1-4.north east); (2mat-1-4.north east) – node\[right=15pt\] [$n$]{} (2mat-1-4.south east); (2mat-2-4.north east) – node\[right=15pt\] [$n$]{} (2mat-2-4.south east); (2mat-3-4.north east) – node\[right=15pt\] [$(2n)^2$]{} (2mat-4-4.south east); at (2mat-3-1.north west) [$\mathrm{and} \quad Y = $]{}; at (2mat-3-4.north east) [.]{}; These can clearly be constructed in polynomial time, and by Lemma \[boolingoss\] we know that $X,Y \in G_{2n(2n+1)}$. If there exists $G \in G_n$ such that $AG = B$ then a simple calculation shows that $UXV = Y$ (and hence $Y \leq_{\mathcal{J}} X$) where (1mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (1mat-2-1.south west) – (1mat-2-4.south east); (1mat-1-2.north east) – (1mat-4-2.south east); (1mat-3-3.south west) – (1mat-3-4.south east); (1mat-3-3.north east) – (1mat-4-3.south east); at (1mat-1-1.south east) [$I_{(2n)^2}$]{}; at (1mat-1-3.south east) [$0$]{}; at (1mat-3-1.south east) [$0$]{}; at (1mat-3-3) [$G$]{}; at (1mat-3-4) [$0$]{}; at (1mat-4-3) [$0$]{}; at (1mat-4-4) [$I_n$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$(2n)^2$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-4.north east) – node\[right=15pt\] [$(2n)^2$]{} (1mat-2-4.south east); (1mat-3-4.north east) – node\[right=15pt\] [$n$]{} (1mat-3-4.south east); (1mat-4-4.north east) – node\[right=15pt\] [$n$]{} (1mat-4-4.south east); at (1mat-3-1.north west) [$U = I_{2n(2n+1)}, \qquad V = $]{}; at (1mat-3-4.north east) [.]{}; Conversely, if there exist $U,V \in G_{2n(2n+1)}$ such that $UXV = Y$, then we consider what structure the matrices $U$ and $V$ could possibly have by considering them as products of call matrices $C[i,j]$. Recall that left multiplication of $X$ by $C[i,j]$ replaces rows $i$ and $j$ of $X$ with their (element-wise) maximum, whilst right multiplication by $C[i,j]$ replaces columns $i$ and $j$ of $X$ with their maximum. Thus, by regarding $U$ and $V$ as products of call generators, we see that $Y$ can be built from $X$ by successively replacing either two rows or two columns with their maximum. The last $(2n)^2$ rows of $X$ all contain $1$s in each of the last $2n$ columns, whilst the first $2n$ rows of $Y$ each have a $0$ in one of the last $2n$ columns. It follows that $U$ cannot contain $C[i,j]$ as a factor if $i \leq 2n$ and $j > 2n$. Each of the rows $n+1, \ldots, 2n$ of $X$ contain $1$s in each of the last $n$ columns, whilst the first $n$ rows of $Y$ each have a $0$ in one of the last $n$ columns. Thus $U$ cannot contain $C[i,j]$ as a factor if $i \leq n$ and $n < j \leq 2n$. The factors of $U$ are therefore call matrices $C[i,j]$ such that $i$ and $j$ are either both less than $n$, both between $n$ and $2n$, or both greater than $2n$. Both $X$ and $Y$ contain the $n \times n$ identity matrix in the top right corner. It follows that $U$ cannot contain $C[i,j]$ as a factor if $i,j \leq n$ are distinct, as this would result in an off-diagonal zero in this $n \times n$ submatrix of $Y$. Therefore, for some $D \in G_n, E \in G_{(2n)^2}$ we must have (1mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (1mat-2-1.south west) – (1mat-2-4.south east); (1mat-1-2.north east) – (1mat-4-2.south east); (1mat-1-1.south west) – (1mat-1-2.south east); (1mat-1-1.north east) – (1mat-2-1.south east); at (1mat-1-1) [$I_n$]{}; at (1mat-1-2) [$0$]{}; at (1mat-2-1) [$0$]{}; at (1mat-2-2) [$D$]{}; at (1mat-1-3.south east) [$0$]{}; at (1mat-3-1.south east) [$0$]{}; at (1mat-3-3.south east) [$E$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n$]{} (1mat-1-1.north east); (1mat-1-2.north west) – node\[above=8pt\] [$n$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$(2n)^2$]{} (1mat-1-4.north east); (1mat-1-4.north east) – node\[right=15pt\] [$n$]{} (1mat-1-4.south east); (1mat-2-4.north east) – node\[right=15pt\] [$n$]{} (1mat-2-4.south east); (1mat-3-4.north east) – node\[right=15pt\] [$(2n)^2$]{} (1mat-4-4.south east); at (1mat-3-1.north west) [$U = $]{}; at (1mat-3-4.north east) [.]{}; It follows that $X$ and $UX$ are identical in the first $n$ rows. But we have $X \preceq UX \preceq UXY = Y$ and $X$ and $Y$ are identical in all the other rows, so it must be that $X = UX$, and $XV = UXV = Y$. Consider now the action of $V$ by right multiplication on $X$. The first $(2n)^2$ columns of $X$ all contain $1$s in each of the first $2n$ rows, whilst the last $2n$ columns of $Y$ each have a $0$ in one of the first $2n$ rows, and so $V$ cannot contain $C[i,j]$ as a factor if $i \leq (2n)^2$ and $j > (2n)^2$. Each of the last $n$ columns of $X$ contains a $1$ in row $n+1$, whilst columns $(2n)^2+1, \ldots, (2n)^2+n$ of $Y$ each contain a $0$ in row $n+1$, and so $V$ cannot contain $C[i,j]$ as a factor if $(2n)^2 < i \leq (2n)^2 + n$ and $j > (2n)^2 + n$. It now follows that there must be some $F \in G_{(2n)^2}, G, H \in G_n$ such that (2mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (2mat-2-1.south west) – (2mat-2-4.south east); (2mat-1-2.north east) – (2mat-4-2.south east); (2mat-3-3.south west) – (2mat-3-4.south east); (2mat-3-3.north east) – (2mat-4-3.south east); at (2mat-1-1.south east) [$F$]{}; at (2mat-1-3.south east) [$0$]{}; at (2mat-3-1.south east) [$0$]{}; at (2mat-3-3) [$G$]{}; at (2mat-3-4) [$0$]{}; at (2mat-4-3) [$0$]{}; at (2mat-4-4) [$H$]{}; (2mat-1-1.north west) – node\[above=8pt\] [$(2n)^2$]{} (2mat-1-2.north east); (2mat-1-3.north west) – node\[above=8pt\] [$n$]{} (2mat-1-3.north east); (2mat-1-4.north west) – node\[above=8pt\] [$n$]{} (2mat-1-4.north east); (2mat-1-4.north east) – node\[right=15pt\] [$(2n)^2$]{} (2mat-2-4.south east); (2mat-3-4.north east) – node\[right=15pt\] [$n$]{} (2mat-3-4.south east); (2mat-4-4.north east) – node\[right=15pt\] [$n$]{} (2mat-4-4.south east); at (2mat-3-1.north west) [$V = $]{}; at (2mat-3-4.north east) [,]{}; and then $XV=Y$ tells us that $AG=B$ as required. NP-Completeness of the Gossip Membership Problem ================================================ In this section we prove that the Gossip Membership Problem is NP-complete. \[mainthm\] GMP is NP-complete. We have already seen that GMP is in NP. We show NP-hardness with a polynomial time reduction from MGTP. Let $A, B \in \mathbb{B}_n$ with $A$ satisfying the maximal column condition. Specifically, we show that for each instance $(A,B)$ of the Maximal Gossip Transformation Problem, there exists a Boolean matrix $C \in B_{n(n+4)}$, which can be constructed in polynomial time, for which the problem of deciding membership in the gossip monoid is equivalent to the decision problem MGTP for the pair $(A,B)$. If $n = 1$ then we simply set $C = [1]$ if $A = B$ and $C = [0]$ if $A \neq B$. Suppose, then, that $n \geq 2$. Let $C$ be the $n(n + 4) \times n(n + 4)$ matrix (1mat) [ & & & & &\ & & & & &\ & & & & &\ & & & & &\ & & & & &\ & & & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-6.south east); (1mat-3-1.south west) – (1mat-3-6.south east); (1mat-4-1.south west) – (1mat-4-6.south east); (1mat-5-1.south west) – (1mat-5-6.south east); (1mat-1-2.north east) – (1mat-6-2.south east); (1mat-1-3.north east) – (1mat-6-3.south east); (1mat-1-4.north east) – (1mat-6-4.south east); (1mat-1-5.north east) – (1mat-6-5.south east); at (1mat-1-2.west) [$1$]{}; at (1mat-1-3) [$A$]{}; at (1mat-1-4) [$A$]{}; at (1mat-1-5) [$0$]{}; at (1mat-1-6) [$B$]{}; at (1mat-3-2.north west) [$1$]{}; at (1mat-3-3.north) [$1$]{}; at (1mat-3-4.north) [$1$]{}; at (1mat-3-5.north) [$0$]{}; at (1mat-3-6.north) [$1$]{}; at (1mat-4-2.west) [$0$]{}; at (1mat-4-3) [$0$]{}; at (1mat-4-4) [$1$]{}; at (1mat-4-5) [$I_n$]{}; at (1mat-4-6) [$1$]{}; at (1mat-5-2.west) [$0$]{}; at (1mat-5-3) [$I_n$]{}; at (1mat-5-4) [$1$]{}; at (1mat-5-5) [$I_n$]{}; at (1mat-5-6) [$1$]{}; at (1mat-6-2.west) [$0$]{}; at (1mat-6-3) [$I_n$]{}; at (1mat-6-4) [$1$]{}; at (1mat-6-5) [$I_n$]{}; at (1mat-6-6) [$1$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n^{2}$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-5.north west) – node\[above=8pt\] [$n$]{} (1mat-1-5.north east); (1mat-1-6.north west) – node\[above=8pt\] [$n$]{} (1mat-1-6.north east); (1mat-1-6.north east) – node\[right=15pt\] [$n$]{} (1mat-1-6.south east); (1mat-2-6.north east) – node\[right=15pt\] [$n^{2}$]{} (1mat-3-6.south east); (1mat-4-6.north east) – node\[right=15pt\] [$n$]{} (1mat-4-6.south east); (1mat-5-6.north east) – node\[right=15pt\] [$n$]{} (1mat-5-6.south east); (1mat-6-6.north east) – node\[right=15pt\] [$n$]{} (1mat-6-6.south east); at (1mat-3-1.south west) [$C = $]{}; at (1mat-3-6.south east) [.]{}; This matrix can clearly be constructed in polynomial time. We claim that $C \in G_{n(n+4)}$ if and only if there is a $G \in G_n$ such that $AG = B$. For ease of reference during the proof, we shall label the blocks of the matrix (and other matrices of the same size) as follows: (1mat) [ & & & & &\ & & & & &\ & & & & &\ & & & & &\ & & & & &\ & & & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-6.south east); (1mat-3-1.south west) – (1mat-3-6.south east); (1mat-4-1.south west) – (1mat-4-6.south east); (1mat-5-1.south west) – (1mat-5-6.south east); (1mat-1-2.north east) – (1mat-6-2.south east); (1mat-1-3.north east) – (1mat-6-3.south east); (1mat-1-4.north east) – (1mat-6-4.south east); (1mat-1-5.north east) – (1mat-6-5.south east); at (1mat-1-2.west) [$a1$]{}; at (1mat-1-3) [$b1$]{}; at (1mat-1-4) [$c1$]{}; at (1mat-1-5) [$d1$]{}; at (1mat-1-6) [$e1$]{}; at (1mat-3-2.north west) [$a2$]{}; at (1mat-3-3.north) [$b2$]{}; at (1mat-3-4.north) [$c2$]{}; at (1mat-3-5.north) [$d2$]{}; at (1mat-3-6.north) [$e2$]{}; at (1mat-4-2.west) [$a3$]{}; at (1mat-4-3) [$b3$]{}; at (1mat-4-4) [$c3$]{}; at (1mat-4-5) [$d3$]{}; at (1mat-4-6) [$e3$]{}; at (1mat-5-2.west) [$a4$]{}; at (1mat-5-3) [$b4$]{}; at (1mat-5-4) [$c4$]{}; at (1mat-5-5) [$d4$]{}; at (1mat-5-6) [$e4$]{}; at (1mat-6-2.west) [$a5$]{}; at (1mat-6-3) [$b5$]{}; at (1mat-6-4) [$c5$]{}; at (1mat-6-5) [$d5$]{}; at (1mat-6-6) [$e5$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$n^2$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$n$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$n$]{} (1mat-1-4.north east); (1mat-1-5.north west) – node\[above=8pt\] [$n$]{} (1mat-1-5.north east); (1mat-1-6.north west) – node\[above=8pt\] [$n$]{} (1mat-1-6.north east); (1mat-1-6.north east) – node\[right=15pt\] [$n$]{} (1mat-1-6.south east); (1mat-2-6.north east) – node\[right=15pt\] [$n^{2}$]{} (1mat-3-6.south east); (1mat-4-6.north east) – node\[right=15pt\] [$n$]{} (1mat-4-6.south east); (1mat-5-6.north east) – node\[right=15pt\] [$n$]{} (1mat-5-6.south east); (1mat-6-6.north east) – node\[right=15pt\] [$n$]{} (1mat-6-6.south east); We shall also use $a$, $b$, $c$, $d$ and $e$ to refer to the sets of indices $\{1, \ldots, n^2\}$, $\{n^2 + 1, \ldots, n^2 + n\}$, $\{n^2 + n + 1, \ldots, n^2 + 2n\}$, $\{n^2 + 2n + 1, \ldots, n^2 + 3n\}$ and $\{n^2 + 3n + 1, \ldots, n^2 + 4n\}$ respectively, so that the columns indexed by these sets correspond to the blocks described above. We define $a_i = i, b_i = n^2 + i, c_i = n^2 + n + i, d_i = n^2 + 2n + i$ and $e_i = n^2 + 3n + i$ so that, for example, the $i$th column in block $c$ is column $c_i$. We first show that if $AG = B$ for some $G \in G_n$ then $C \in G_{n(n+4)}$, by showing how to write $C$ as a product of call matrices. Let (1mat) [ & & & & &\ & & & & &\ & & & & &\ & & & & &\ & & & & &\ & & & & &\ ]{}; (1mat-1-1.south west) – (1mat-1-6.south east); (1mat-3-1.south west) – (1mat-3-6.south east); (1mat-4-1.south west) – (1mat-4-6.south east); (1mat-5-1.south west) – (1mat-5-6.south east); (1mat-1-2.north east) – (1mat-6-2.south east); (1mat-1-3.north east) – (1mat-6-3.south east); (1mat-1-4.north east) – (1mat-6-4.south east); (1mat-1-5.north east) – (1mat-6-5.south east); at (1mat-1-2.west) [$1$]{}; at (1mat-1-3) [$A$]{}; at (1mat-1-4) [$0$]{}; at (1mat-1-5) [$0$]{}; at (1mat-1-6) [$0$]{}; at (1mat-3-2.north west) [$1$]{}; at (1mat-3-3.north) [$1$]{}; at (1mat-3-4.north) [$0$]{}; at (1mat-3-5.north) [$0$]{}; at (1mat-3-6.north) [$0$]{}; at (1mat-4-2.west) [$0$]{}; at (1mat-4-3) [$0$]{}; at (1mat-4-4) [$I_n$]{}; at (1mat-4-5) [$0$]{}; at (1mat-4-6) [$0$]{}; at (1mat-5-2.west) [$0$]{}; at (1mat-5-3) [$0$]{}; at (1mat-5-4) [$0$]{}; at (1mat-5-5) [$I_n$]{}; at (1mat-5-6) [$0$]{}; at (1mat-6-2.west) [$0$]{}; at (1mat-6-3) [$0$]{}; at (1mat-6-4) [$0$]{}; at (1mat-6-5) [$0$]{}; at (1mat-6-6) [$I_n$]{}; (1mat-1-1.north west) – node\[above=8pt\] [$a$]{} (1mat-1-2.north east); (1mat-1-3.north west) – node\[above=8pt\] [$b$]{} (1mat-1-3.north east); (1mat-1-4.north west) – node\[above=8pt\] [$c$]{} (1mat-1-4.north east); (1mat-1-5.north west) – node\[above=8pt\] [$d$]{} (1mat-1-5.north east); (1mat-1-6.north west) – node\[above=8pt\] [$e$]{} (1mat-1-6.north east); at (1mat-3-1.south west) [$Y_1 = $]{}; at (1mat-3-6.south east) [,]{}; (2mat) [ & & &\ & & &\ & & &\ & & &\ ]{}; (2mat-3-1.south west) – (2mat-3-4.south east); (2mat-1-3.north east) – (2mat-4-3.south east); at (2mat-2-2) [$I_{n(n+3)}$]{}; at (2mat-2-4) [$0$]{}; at (2mat-4-2) [$0$]{}; at (2mat-4-4) [$G$]{}; (2mat-1-1.north west) – node\[above=8pt\] [$a,b,c,d$]{} (2mat-1-3.north east); (2mat-1-4.north west) – node\[above=8pt\] [$e$]{} (2mat-1-4.north east); at (2mat-2-1.south west) [$Y_7 = $]{}; at (2mat-2-4.south east) [,]{}; $$\begin{aligned} & Y_2 = \prod_{i = 1}^n C[d_i, e_i], & Y_3 = \prod_{i = 1}^n C[c_i, d_i], \qquad\qquad & Y_4 = C[\{c_1, \ldots, c_n\}], \\ & Y_5 = \prod_{i = 1}^n C[b_i, e_i], & Y_6 = \prod_{i = 1}^n C[c_i, e_i]. \qquad\qquad &\end{aligned}$$ Since $n \geq 2$ it follows from Lemma \[boolingoss\] that $Y_1$ is a product of call matrices in $G_{n(n+4)}$. The assumption that $G \in G_n$ means that $Y_7$ can be written as a product of call matrices in $G_{n(n+4)}$. Matrices $Y_2$ through to $Y_6$ are all explicitly defined as products of call matrices, and we note that the products in $Y_2, Y_3, Y_5$ and $Y_6$ can be taken in any order. It is then straightforward to check that $C = Y_1 Y_2 Y_3 Y_4 Y_5 Y_6 Y_7$. Conversely, assume that $C \in G_{n(n+4)}$ and fix a sequence of call matrices $C_1, \dots, C_q$ such that $C = C_1 \cdots C_q$. We may clearly assume without loss of generality that the sequence contains no redundant factors, in the sense that $C_1\cdots C_t \neq C_1 \cdots C_t C_{t+1}$ (which since multiplication is monotonic means that $C_1\cdots C_t \prec C_1 \cdots C_t C_{t+1}$) for all $t$. By defining a particular scattered subsequence of these factors, we shall construct an element $G \in G_n$ satisfying $AG=B$. We start by observing that for any scattered subsequence $C_{t_1}, \ldots, C_{t_p}$ of $C_1, \ldots, C_q$ we must have $C_{t_1}\cdots C_{t_p} \preceq C$. We proceed by establishing a series of claims. For each $t$, if $C_t$ is of the form $C[i,j]$ with $i \in a \cup b$ and $j \in c \cup d \cup e$, then $C_t = C[b_k, e_k]$ for some $k \in \lbrace 1,\ldots, n \rbrace$. Moreover, for each $k$, the matrix $C[b_k, e_k]$ must appear exactly once in the sequence $C_1, \dots, C_q$. Suppose that $C_t = C[i,j]$, where $i \in a \cup b$ and $j \in c \cup d \cup e$. Since $C_t \preceq C$ we note that $i \notin a$, otherwise $C$ would contain a $1$ in block $a3$, $a4$ or $a5$. Similarly, we note that $j \notin c \cup d$, since otherwise $C$ would contain a 1 in block $b3$ or $d2$. Finally, let $i=b_k$ and note that if $j \neq e_k$ then $C$ would contain an off-diagonal 1 in block $b5$. It now follows that the only way that $C$ can contain the identity matrix in block $b5$ is for each of the matrices $C[b_k, e_k]$ to occur at least once in the sequence $C_1, \dots, C_q$. Assume for contradiction that $C[b_k, e_k]$ occurs more than once in the sequence, and suppose that $C_r$ and $C_s$ are the first two occurrences of this matrix. If $C_t$ is a factor occurring between $C_r$ and $C_s$ then $C_rC_tC_s \preceq C$. We cannot have $C_t$ of the form $C[b_k, j]$ with $j \in a \cup b$, as otherwise $C_rC_tC_s$ — and therefore also $C$ — would have a $1$ in block $a5$ or an off-diagonal $1$ in block $b5$. Similarly, it cannot be of the form $C[e_k, j]$ with $j \in c \cup d \cup e$, since then $C_rC_tC_s$ would have a $1$ in block $b3$ or $d2$ or an off-diagonal $1$ in block $b5$. By the first part of the claim, $C_t$ cannot be of the form $C[b_k, j]$ with $j \in c \cup d \cup e \setminus\{e_k\}$ or $C[e_k, j]$ with $j \in a \cup b \setminus\{b_k\}$. It also cannot be of the form $C[b_k, e_k]$ as it lies strictly between the first two factors of this form. Therefore $C_t$ is not of the form $C[b_k, j]$ or $C[e_k, j]$ for any $j$, so $C_r$ commutes with $C_t$. Since $C_t$ was an arbitrary factor between $C_r$ and $C_s$, and since $C_r = C_s$ is an idempotent, we have $C_r C_{r+1} \cdots C_{s-1} C_s = C_r C_s C_{r+1} \cdots C_{s-1} = C_r C_{r+1} \cdots C_{s-1}$, contradicting our assumption that the sequence $C_1, \dots, C_q$ contains no redundant factors and thus proving the claim. For $k \in \{1, \ldots n\}$ let $w_k$ be the unique index such that $C_{w_k} = C[b_k, e_k]$. For each $k \leq n$, column $k$ of block $e1$ in the product $C_1 \cdots C_{w_k -1}$ is equal to the zero vector, but the same column in $C_1 \cdots C_{w_k}$ is equal to column $k$ of $A$. We first show that in the product $C_1 \dots C_{w_k-1}$, the $k$th column of block $e1$ is equal to the zero vector. Assume that the vector is non-zero, so there is some $C_s$ with $s < w_k$ such that $C_1 \cdots C_{s-1}$ has the zero vector in this location, but $C_1 \cdots C_s$ does not. The factor $C_s$ must be of the form $C[e_k, j]$ and by Claim 1 and the fact that $s < w_k$, we have $j \in c \cup d \cup e$. We note that $j \notin c$, otherwise $C_s C_{w_k}$ — and hence also $C$ — would have a $1$ in block $b3$. Similarly we note that $j \notin e$, otherwise $C_s C_{w_k}$ would have an off-diagonal $1$ in block $b5$. This means $C_s$ must be of the form $C[e_k, d_i]$, but then columns $e_k$ and $d_i$ of $C_1 \cdots C_s$ are identical to each other, so this product has a $1$ in block $d1$, contradicting $C_1 \cdots C_s \preceq C$. Thus the $k$th column of block $e1$ in the product $C_1 \cdots C_{w_k-1}$ is equal to the zero vector. Since the last element of the product $C_1 \cdots C_{w_k}$ is $C_{w_k} = C[b_k, e_k]$, columns $b_k$ and $e_k$ of the product are identical and the second part of the claim will follow if the $k$th column of block $b1$ in this product is equal to the $k$th column of $A$. Suppose this does not hold; since it does hold in $C$, the sequence $C_{w_k +1}, \dots, C_q$ must include a subsequence which transforms the $k$th column of $b1$ into the $k$th column of $A$. Let $C[b_k, j]$ be one such call. Notice that $j \notin a$, since otherwise $C_{w_k}C[b_k, j] \preceq C$ would have a 1 in block $a5$. Similarly, $j \notin b$, since otherwise $C_{w_k}C[b_k, j]$ would have an off-diagonal 1 in block $b5$. The first part of Claim 1 now shows that the only remaining possibility is $j=e_k$. However, this would result in two factors of $C$ both equal to $C[b_k, e_k]$, contradicting the second part of Claim 1. Hence column $k$ of block $b1$ in the product $C_1 \cdots C_{w_k}$ is equal to the $k$th column of $A$, completing the proof of the claim. To provide a convenient starting point for induction, let $C_0$ be the $(n^2+4n)\times (n^2+4n)$ identity matrix, so that $C= C_0C_1 \cdots C_q$. For all $0\leq t\leq q$, every column in block $c1$, $d1$ or $e1$ of $C_0C_1 \cdots C_t$ is equal to either $\underline{0}$ or the maximum of some collection of columns of $A$. We prove this by induction on $t$. When $t = 0$ this is clearly true, as each of these columns is equal to $\underline{0}$. For $t \geq 1$ we shall assume that the condition holds for the columns of the appropriate blocks in $C_1 \cdots C_{t-1}$. By Claim 1, $C_t$ is of the form (i) $C[i,j]$ with $i,j \in a \cup b$, (ii) $C[i,j]$ with $i,j \in c \cup d \cup e$ or (iii) $C[b_k, e_k]$ for some $k$. In case (i) the columns of blocks $c1$, $d1$ and $e1$ are the same as the corresponding columns in the product $C_1 \cdots C_{t-1}$, which satisfy the condition by assumption. In case (ii) two of the columns of blocks $c1$, $d1$ and $e1$ are equal to the maximum of the corresponding columns from the product $C_1 \cdots C_{t-1}$, and so they satisfy the condition of the claim, and the rest of the columns are the same as the corresponding columns from $C_1 \cdots C_{t-1}$. In case (iii) by Claim 2, column $k$ of block $e1$ is equal to the $k$th column of $A$, so satisfies the claim, and all other columns of blocks $c1$, $d1$ and $e1$ are equal to the corresponding columns in $C_1 \cdots C_{t-1}$. The claim follows by induction. To prove the theorem we want to construct a matrix $G \in G_n$ such that $AG = B$. Consider those call matrices $C_t$ (in order) for which there exists $k \leq n$ such that $w_k<t$ and column $k$ of block $e1$ in the product $C_1 \cdots C_t$ is strictly larger than the same column in the product $C_1 \cdots C_{t-1}$. (Intuitively, these are the factors which modify some column $k$ of block $e1$ subsequent to the factor $C_{w_k} = C[b_k,e_k]$ which copies column $k$ of $A$ into this column.) Let $C_{t_1}, \ldots, C_{t_p}$ denote the subsequence of these matrices. We shall assume that this sequence is non-empty, since otherwise $B = A$ and $AG = B$ is trivially satisfied by $G = I_n$. Each element of the sequence $C_{t_1}, \ldots, C_{t_p}$ is of the form $C_{t_r} = C[e_j, e_k]$ for some $j,k \leq n$ such that $w_j < t_r$ and $w_k < t_r$. Let $C_{t_r}$ be an element of this subsequence. From the definition of the subsequence there is some $k \leq n$ such that $w_k < t_r$ and column $k$ of block $e1$ of $C_1 \cdots C_{t_r}$ is strictly larger than the same column in $C_1 \cdots C_{{t_r}-1}$. Thus $C_{t_r}$ must be of the form $C[i, e_k]$ for some $i \leq n^2 + 4n$. Claim 1 tells us that $i \notin a$. Since $t_r > w_k$, Claim 1 also tells us that $i \notin b$. By monotonicity, all columns in block $d1$ of $C_1 \cdots C_{t_r}$ are $\underline{0}$, so $i \notin d$. Assume for contradiction that $C_{t_r} = C[c_j, e_k]$ for some $j$. We observe several relations between columns of various matrices: - Column $j$ of $A$ is equal to column $j$ of block $c1$ in $C$. - By monotonicity, column $j$ of block $c1$ in $C$ is greater than or equal to the same column in $C_1 \cdots C_{t_r}$. - Since the last factor in the product $C_1 \cdots C_{t_r}$ is $C_{t_r} = C[c_j, e_k]$, columns $c_j$ and $e_k$ are equal in the resulting matrix. - By our choice of $k$, column $k$ of block $e1$ of $C_1 \cdots C_{t_r}$ is strictly greater than the same column in $C_1 \cdots C_{{t_r}-1}$. - Since $w_k < t_r$, by Claim 2 and monotonicity we know that column $k$ of block $e1$ in $C_1 \cdots C_{{t_r}-1}$ is greater than or equal to column $k$ of $A$. Putting these together, we find that column $j$ of $A$ is strictly greater than column $k$ of $A$, which contradicts the maximal column condition. Thus $C_{t_r}$ must be of the form $C[e_j, e_k]$ for some $j$. We know that $w_k < t_r$ by definition of the sequence, so for the second part of the claim it is enough to observe that $t_r \neq w_j$ since $C_{t_r} \neq C[b_j, e_k]$, and if $t_r < w_j$ then we would have $C[e_j, e_k]C[b_j, e_j] \preceq C$ and so $C$ would contain an off-diagonal $1$ in block $b5$, which is not the case. We see from this claim that there are only two types of call matrix in the sequence which can modify a column of block $e1$. The first matrix in the sequence which modifies column $i$ of block $e1$ is $C_{w_i}$, and all subsequent matrices (if any) which modify this column are elements of the subsequence $C_{t_1}, \ldots, C_{t_p}$. We now define a new sequence of call matrices in $G_n$. For $1 \leq r \leq p$, if $C_{t_r} = C[e_i, e_j]$ then let $D_r = C[i,j] \in G_n$. For each $r \leq p$ and $k \leq n$, column $k$ of block $e1$ of $C_1 \cdots C_{t_r}$ is equal to the zero vector if $t_r < w_k$, or equal to column $k$ of $A D_1 \cdots D_r$ otherwise. By Claim 2, for each $k \leq n$ the $k$th column of block $e1$ in the product $C_1 \cdots C_{w_k-1}$ is equal to the zero vector. By monotonicity, if $t_r < w_k$ then the same column is equal to the zero vector in the product $C_1 \cdots C_{t_r}$. We prove by induction on $r$ that for each $k$ such that $w_k < t_r$, column $k$ of block $e1$ of $C_1\cdots C_{t_r}$ is equal to column $k$ of $A D_1 \cdots D_r$. Let $t_0 = 0$. Then $t_0 < w_k$ for all $k$ so the claim holds for $r = 0$. Now assume the claim holds for $r-1$. That is, for each $k$ such that $w_k < t_{(r-1)}$, column $k$ of block $e1$ of $C_1\cdots C_{t_{(r-1)}}$ is equal to column $k$ of $A D_1 \cdots D_{r-1}$. Given $k \leq n$ such that $w_k < t_r$ we first consider the differences between column $k$ of block $e1$ in $C_1\cdots C_{t_{(r-1)}}$ and the same column in $C_1\cdots C_{t_r -1}$. No elements of the subsequence $C_{t_1}, \ldots, C_{t_p}$ occur among the matrices $C_{t_{(r-1)} + 1}, \ldots, C_{t_r -1}$ and we have noted that the only other factor of $C$ which can possibly modify column $k$ of block $e1$ is $C_{w_k}$. Thus the two columns under consideration are identical unless $t_{(r-1)} < w_k < t_r$, and in this case, by Claim 2, column $k$ of block $e1$ in $C_1\cdots C_{t_r -1}$ is equal to column $k$ of $A$. If $t_{(r-1)} < w_k < t_r$ then, since $C_{w_k}$ comes earlier in the sequence than any other matrix which can modify column $k$ of block $e1$, we see that $C_{t_r}$ is the first element of the sequence $C_{t_1}, \ldots, C_{t_p}$ which can possibly modify column $k$ of block $e1$, and thus $D_r$ is the first element of the sequence $D_1, \ldots, D_p$ which can possibly modify column $k$. Thus column $k$ of $AD_1 \cdots D_{r-1}$ is equal to column $k$ of $A$, and is therefore equal to column $k$ of $e1$ in $C_1\cdots C_{t_r -1}$. If $t_{(r-1)} < w_k < t_r$ does not hold then, since we are only considering $k$ such that $w_k < t_r$, we must have $w_k < t_{(r-1)}$. Thus by the inductive hypothesis we know that column $k$ of block $e1$ of $C_1\cdots C_{t_{(r-1)}}$ is equal to column $k$ of $A D_1 \cdots D_{r-1}$. Since none of the matrices $C_{t_{(r-1)} + 1}, \ldots, C_{t_r -1}$ modify this column, column $k$ of $e1$ in $C_1\cdots C_{t_r -1}$ is again equal to column $k$ of $A D_1 \cdots D_{r-1}$. Thus for all $k$ such that $w_k < t_r$ we know that column $k$ of $e1$ in $C_1\cdots C_{t_r -1}$ is equal to column $k$ of $A D_1 \cdots D_{r-1}$. By Claim 4, $C_{t_r} = C[e_i, e_j]$ for some $i,j \leq n$ such that $w_i, w_j < t_r$, and then $D_r$ is defined to be $C[i,j]$. Let $k < n$ be such that $w_k < t_r$. We first consider the case when $k \notin \{i,j\}$. Since $k \notin \{i,j\}$, column $k$ of block $e1$ of $C_1\cdots C_{t_r}$ is equal to the same column in $C_1\cdots C_{t_r -1}$. By the previous paragraph, this column is equal to column $k$ of $AD_1 \cdots D_{r-1}$, and again because $k \notin \{i,j\}$, this is equal to column $k$ of $AD_1 \cdots D_r$. In the case where $k \in \{i,j\}$, column $k$ of block $e1$ of $C_1\cdots C_{t_r}$ is equal to the maximum of columns $i$ and $j$ of block $e1$ of $C_1\cdots C_{t_r -1}$. Since $w_i, w_j < t_r$, by the previous paragraph this is equal to the maximum of columns $i$ and $j$ of $AD_1 \cdots D_{r-1}$. Since $D_r = C[i,j]$, column $k$ of $AD_1 \cdots D_r$ is also equal to the maximum of columns $i$ and $j$ of $AD_1 \cdots D_{r-1}$, which completes the proof of the claim. We now define $G = D_1 \cdots D_p$. For each $k \leq n$, the only factors of $C$ which can possibly modify column $k$ of block $e1$ are $C_{w_k}$ and elements of the subsequence $C_{t_1}, \ldots, C_{t_p}$. Therefore column $k$ of block $e1$ of $C$ is equal to the same column in $C_1 \cdots C_{t_p}$ if $w_k \leq t_p$, or it is equal to the same column in $C_1 \cdots C_{w_k}$ if $t_p < w_k$. If $w_k \leq t_p$ then by Claim 5 column $k$ of $e1$ in $C_1 \cdots C_{t_p}$ is equal to column $k$ of $AG = AD_1 \cdots D_p$, and therefore column $k$ of $e1$ in $C$ is also equal to column $k$ of $AG$. If $t_p < w_k$ then $C_{w_k}$ is the last factor of $C$ to modify column $k$ of block $e1$, so by Claim 2, column $k$ of $e1$ in $C$ is equal to column $k$ of $A$. We know from Claim 4 that if $C_{t_r} = C[e_j,e_k]$ for some $j$ and $r$ then $w_k < t_r$. However $t_r \leq w_k$ for all $r \leq p$, so no element of the sequence $C_{t_1}, \ldots, C_{t_p}$ is equal to $C[e_j,e_k]$ for any $j \leq n$. Therefore no element of the sequence $D_1, \ldots, D_p$ is equal to $C[j,k]$ for any $j \leq n$, and so column $k$ of $AG = AD_1\cdots D_p$ is also equal to column $k$ of $A$. In both cases, column $k$ of block $e1$ of $C$ is equal to column $k$ of $AG$. Therefore $AG$ is equal to block $e1$ of $C$, which is equal to $B$, and so $G \in G_n$ satisfies $AG = B$ as required to complete the proof. [99]{} B. Baker, R. Shostak, *Gossips and telephones*. Discrete Mathematics 2, 191–193, 1972. M. Brittenham, S. W. Margolis, J. Meakin, *Subgroups of the free idempotent generated semigroups need not be free*, J. Algebra 321 (10), 3026–-3042, 2009. A. E. Brouwer, J. Draisma, B. J. Frenk, *Lossy gossip and composition of metrics*, Discrete Comput. Geom. 53, 890–913, 2015. A. H. Clifford, G. B. Preston, The algebraic theory of semigroups. Vol. I., Mathematical Surveys, No. 7 American Mathematical Society, Providence, R.I., 1961. I. Dolinka, R. Gray, *Maximal subgroups of free idempotent generated semigroups over the full linear monoid*, Trans. Amer. Math. Soc. 366 (1), 419–-455, 2014. B. Frenk, *Tropical Varieties, Maps and Gossip*, PhD. thesis, Eindhoven University of Technology, 2013. M. R. Garey, D. S. Johnson, Computers and Intractability: A Guide to NP-Completeness. W.H. Freeman and Company, San Francisco, 1979. R. Gray, N. Ruskuc, *On maximal subgroups of free idempotent generated semigroups*, Israel J. Math. 189, 147–-176, 2012. R. Gray, N. Ruskuc, *Maximal subgroups of free idempotent-generated semigroups over the full transformation monoid*, Proc. Lond. Math. Soc. 104 (5), 997-–1018, 2012. J. A. Green, *On the structure of semigroups*, Ann. of Math. 54, 163–-172, 1951. A. Hajnal, E. C. Milner and E. Szemeredi, *A cure for the telephone disease*. Canad. Math. Bull. 15, 447–450, 1972. D. Jüngnickel, Graphs, Networks and Algorithms. Algorithms and Computation in Mathematics, 5. Springer, Heidelberg, 2013. V. Mazorchuk, B. Steinberg, *Double Catalan monoids*, J. Algebraic Combin. 36 (3), 333–-354, 2012. R. Tijdeman, *On a telephone problem*. Nieuw Arch. Wisk. 19 (3), 188–192, 1971. D. Yang, I. Dolinka, V. Gould, *Free idempotent generated semigroups and endomorphism monoids of free G-acts*, J. Algebra 429, 133–-176, 2015. [^1]: Email `[email protected]`. Peter Fenner’s research is supported by an EPSRC Doctoral Training Award. [^2]: Email `[email protected]`. [^3]: Email `[email protected]`. [^4]:
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we show that the derivative of the genus-1 Virasoro conjecture for Gromov-Witten invariants along the direction of quantum volume element holds for all smooth projective varieties. This result provides new evidence for the Virasoro conjecture.' author: - 'Xiaobo Liu [^1]' title: 'Genus-1 Virasoro conjecture along quantum volume direction' --- Introduction ============ The Virasoro conjecture predicts that the generating functions of the Gromov-Witten invariants of smooth projective varieties are annihilated by a sequence of differential operators which form a half branch of the Virasoro algebra. This conjecture was proposed by Eguchi-Hori-Xiong [@EHX] and modified by S. Katz [@CK]. In case the underlying manifold is a point, this conjecture is equivalent to Witten’s conjecture [@W], proved by Kontsevich [@K], that the generating function of intersection numbers on the moduli spaces of stable curves is a $\tau$-function of the KdV hierarchy. Together with Tian, we proved that the genus-0 part of the Virasoro conjecture holds for all compact symplectic manifolds (cf. [@LT]). For manifolds with semisimple quantum cohomology, the genus-1 part of this conjecture was proved by Dubrovin and Zhang [@DZ]. Without assuming semisimplicity, the genus-1 Virasoro conjecture was studied in [@L1] and [@L2]. Among other results, it was proved in [@L1] that the genus-1 Virasoro conjecture can be reduced to the the $L_1$-constraint. Using the genus-1 topological recursion relation, it was also proved that Virasoro constraints can be reduced to equations on the [*small phase space*]{}, i.e. the space of cohomology classes of the underlying manifold. Compatibility conditions for Virasoro conjectures were studied in [@L2]. Despite these efforts, the general case of the genus-1 Virasoro conjecture is still largely open. In this paper, we give more evidence to the genus-1 Virasoro conjecture without any assumption on the quantum cohomology of the underlying manifold. Let $M$ be a smooth projective variety. Choose a basis $\{ \gamma_{\alpha} \mid \alpha = 1, \ldots, N\}$ of the space of cohomology classes $H^{*}(M; \mathbb{C})$. For simplicity, we assume $H^{\rm odd}(M; \mathbb{C}) = 0$. We choose the basis in such a way that $\gamma_{1}$ is the identity of the cohomology ring and $\gamma_{\alpha} \in H^{p_{\alpha}, q_{\alpha}}(M)$ for some integers $p_{\alpha}$ and $q_{\alpha}$. Let $\{ t^{1}, \ldots, t^{N}\}$ be the coordinates on $H^{*}(M; \mathbb{C})$ with respect to this basis. We can identify each $\gamma_{\alpha}$ with the vector field $\frac{\partial}{\partial t^{\alpha}}$ and further identify each cohomology class with a constant vector field on $H^{*}(M; \mathbb{C})$. Let $$\label{eqn:balpha} b_{\alpha} = p_{\alpha} - \frac{1}{2}(d-1)$$ where $d$ is the complex dimension of $M$. Then the [*Euler vector field*]{} (on the small phase space) is defined to be $$E := c_{1}(M) + \sum_{\alpha} (b_{1} + 1 - b_{\alpha}) t^{\alpha} \gamma_{\alpha} .$$ We refer to [@LiT] [@RT] for definitions of Gromov-Witten invariants. In genus-$1$ case, it suffices to study only primary Gromov-Witten invariants since all genus-$1$ descendant invariants can be reduced to primary invariants due to the genus-$1$ topological recursion relation. Therefore we only consider primary Gromov-Witten invariants in this paper. Let $F_g$ be the generating function of genus-$g$ primary Gromov-Witten invariants of $M$. The [*$k$-point function*]{} is defined to be $${\left< \hspace{-2pt} \left< \, v_1 \cdots v_k \, \right> \hspace{-2pt} \right>_{g}} := \sum_{\alpha_{1}, \ldots, \alpha_{k}} f^{1}_{\alpha_{1}} \cdots f^{k}_{\alpha_{k}} \,\, \frac{\partial^k F_g}{\partial t^{\alpha_1} \cdots \partial t^{\alpha_k}},$$ for vector fields $v_{i} = \sum_{\alpha} f^{i}_{\alpha} \gamma_{\alpha}$ where $f^{i}_{\alpha}$ are functions on $H^*(M; \mathbb{C})$. Note that $F_g$ and ${\left< \hspace{-2pt} \left< \, \cdots \, \right> \hspace{-2pt} \right>_{g}}$ in this paper corresponds to $F_g^s$ and ${\left< \hspace{-2pt} \left< \, \cdots \, \right> \hspace{-2pt} \right>_{g, s}}$ in [@L1]. Let $\eta_{\alpha \beta} = \int_{M} \gamma_{\alpha} \cup \gamma_{\beta}$ be the intersection form on $H^{*}(M, {\Bbb C})$. We will use $\eta = (\eta_{\alpha \beta})$ and $\eta^{-1} = (\eta^{\alpha \beta})$ to lower and raise indices. For example ${\gamma^{\alpha}}:= \eta^{\alpha \beta} {\gamma_{\beta}}$ where repeated indices should be summed over entire range. We recall that the [*quantum product*]{} of two vector fields $v_1$ and $v_2$ is defined by $$v_1 {\circ}v_2 := {\left< \hspace{-2pt} \left< \, v_1 \, v_2 \, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \, {\gamma_{\alpha}}.$$ Define $$\Psi := {\left< \hspace{-2pt} \left< \, E^{2} \, \right> \hspace{-2pt} \right>_{1}} + \frac{1}{24} \sum_{\alpha} \left<\left< E E \gamma_{\alpha} \gamma^{\alpha} \right>\right>_{0} - \frac{1}{2}\sum_{\alpha} \left(b_{\alpha} (1-b_{\alpha}) - \frac{b_{1}+1}{6}\right) \left<\left< \gamma_{\alpha} \gamma^{\alpha} \right>\right>_{0}$$ where $E^2 = E {\circ}E$ is the quantum square of the Euler vector field. It was proved in [@L1] that, for any smooth projective variety $M$, the [*genus-1 Virasoro conjecture*]{} can be reduced to a single equation on $H^{*}(M; {\Bbb C})$: $$\label{eqn:VirSmall} \Psi = 0.$$ Moreover, since $E \Psi = \Psi$ (cf. [@L1 Lemma 6.3]), the genus-1 Virasoro conjecture holds if and only if $$E \Psi = 0.$$ Therefore, to prove the genus-1 Virasoro conjecture, it suffices to show that $v \Psi =0$ for all vector field $v$ on $H^{*}(M; {\Bbb C})$. It follows from the string equation that $\gamma_1 \Psi = 0$ where $\gamma_1 = E^0$ is the identity of the ordinary cohomology ring. In this paper we will give another vector field which always annihilates $\Psi$. Define the vector field $$\Delta := {\gamma^{\alpha}}{\circ}{\gamma_{\alpha}}.$$ If, in the definition of $\Delta$, we replace the quantum product “${\circ}$” by the ordinary cup product, we get a vector field proportional to the volume element. Therefore we call $\Delta$ the [*quantum volume element*]{}. The main result of this paper is the following \[thm:virD\] For all smooth projective varieties, $$\Delta \Psi = 0.$$ This result provides a new evidence for the genus-1 Virasoro conjecture. Properties of Euler vector fields {#sec:Euler} ================================= We first recall some basic properties of the Euler vector field $E$. We start with the [*quasi-homogeneity equation*]{} $$\left<\left< E \right>\right>_{g} = (3-d)(1-g) F_{g} + \frac{1}{2} \delta_{g, 0}\sum_{\alpha, \beta} {\cal C}_{\alpha \beta} t^{\alpha}_{0}t^{\beta}_{0} - \frac{1}{24} \delta_{g, 1} \int_{M} c_{1}(M) \cup c_{d-1}(M).$$ This equation is a consequence of the divisor equation. Define the grading operator $G$ by $$G(v) := \sum_{\alpha} b_{\alpha} f_{\alpha} {\gamma_{\alpha}}$$ for any vector field $v =\sum_{\alpha} f_{\alpha} {\gamma_{\alpha}}$. Derivatives of quasi-homogeneity equation has the form $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E \, v_{1} \, \cdots \, v_{k} \, \right> \hspace{-2pt} \right>_{g}} & = & \sum_{i=1}^{k} {\left< \hspace{-2pt} \left< \, v_{1} \, \cdots \, G(v_{i}) \, \cdots \, v_{k} \, \right> \hspace{-2pt} \right>_{g}} \nonumber \\ && - (2g+k-2)(b_{1}+1) {\left< \hspace{-2pt} \left< \, v_{1} \, \cdots \, v_{k} \, \right> \hspace{-2pt} \right>_{g}} \nonumber \\ && + \delta_{g, 0} \nabla^{k}_{v_{1}, \cdots, v_{k}} \left( \frac{1}{2} {\cal C}_{\alpha \beta} t_{0}^{\alpha} t_{0}^{\beta} \right) \label{eqn:dhomog}\end{aligned}$$ where ${\cal C}_{\alpha \beta}$ is defined by $c_1(M) \cup {\gamma_{\alpha}}= {\cal C}_{\alpha}^{ \beta} {\gamma_{\beta}}$, and $\nabla$ is the trivial connection on $H^*(M; \mathbb{C})$ defined by $\nabla {\gamma_{\alpha}}= 0$ for all $\alpha$. In particular, $$\label{eqn:Eg04pt} {\left< \hspace{-2pt} \left< \, E \, v_1 \, v_2 \, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \, {\gamma_{\alpha}}= G(v_1) {\circ}v_2 + v_1 {\circ}G(v_2) - G(v_1 {\circ}v_2) - b_1 v_1 {\circ}v_2.$$ Combining with [@L1 Lemma 4.2], we can obtain $$\begin{aligned} \nabla_{E^k} \Delta &=& {\left< \hspace{-2pt} \left< \, E^k \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \, {\gamma_{\beta}}\nonumber \\ &=& (k - b_1) E^{k-1} {\circ}\Delta - G(E^{k-1} {\circ}\Delta) - \sum_{i=1}^{k-1} \, \, \Delta {\circ}E^{i-1} {\circ}G(E^{k-i}) \nonumber \\ && - \sum_{i=1}^{k-1} G(\Delta {\circ}E^{i-1}) {\circ}E^{k-i} \label{eqn:dEkD}\end{aligned}$$ for $k \geq 1$. Covariant derivative of $E$ is given by $$\label{eqn:dEuler} \nabla_{v} E = - G(v) + (b_{1}+1)v.$$ Using the fact that $$\nabla_w (v_1 {\circ}v_2) = (\nabla_w v_1) {\circ}v_2 + v_1 {\circ}(\nabla_w v_2) + {\left< \hspace{-2pt} \left< \, w \, v_1 \, v_2 \, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \, {\gamma_{\alpha}},$$ we can also show that $$\label{eqn:dDE2} \nabla_{\Delta} E^2 = \Delta {\circ}G(E) - G(\Delta) {\circ}E - G(\Delta {\circ}E) + (b_1+2) \Delta {\circ}E.$$ Combining equations and , we have $$[E^2, \Delta] = - 2 b_1 E {\circ}\Delta - 2 G(E) {\circ}\Delta.$$ Proof of the main theorem ========================= For any vector fields $v_{1}, \ldots v_{4}$ on the small phase space, we define $$\begin{aligned} G_{0}(v_{1}, v_{2}, v_{3}, v_{4}) & = & \sum_{g \in S_{4}} \sum_{\alpha, \beta} \left\{ \frac{1}{6} \left<\left< v_{g(1)} v_{g(2)}v_{g(3)} \gamma^{\alpha} \right>\right>_{0} \left<\left< \gamma_{\alpha} v_{g(4)} \gamma_{\beta} \gamma^{\beta} \right>\right>_{0} \right. \\ && \hspace{40pt} + \frac{1}{24} \left<\left< v_{g(1)} v_{g(2)}v_{g(3)} v_{g(4)} \gamma^{\alpha} \right>\right>_{0} \left<\left< \gamma_{\alpha} \gamma_{\beta} \gamma^{\beta} \right>\right>_{0} \\ && \hspace{40pt} \left. - \frac{1}{4} \left<\left< v_{g(1)} v_{g(2)} \gamma^{\alpha} \gamma^{\beta} \right>\right>_{0} \left<\left< \gamma_{\alpha} \gamma_{\beta} v_{g(3)} v_{g(4)} \right>\right>_{0} \right\},\end{aligned}$$ and $$\begin{aligned} G_{1}(v_{1}, v_{2}, v_{3}, v_{4}) & = & \sum_{g \in S_{4}} 3 \left<\left< \{v_{g(1)} {\circ}v_{g(2)} \} \{v_{g(3)} {\circ}v_{g(4)} \} \right>\right>_{1} \\ && - \sum_{g \in S_{4}} 4 \left<\left< \{v_{g(1)} {\circ}v_{g(2)} {\circ}v_{g(3)} \} v_{g(4)} \right>\right>_{1} \\ && - \sum_{g \in S_{4}} \sum_{\alpha} \left<\left< \{ v_{g(1)} {\circ}v_{g(2)} \} v_{g(3)} v_{g(4)} \gamma^{\alpha} \right>\right>_{0} \left<\left< \gamma_{\alpha} \right>\right>_{1} \\ && + \sum_{g \in S_{4}} \sum_{\alpha} 2 \left<\left< v_{g(1)} v_{g(2)} v_{g(3)} \gamma^{\alpha} \right>\right>_{0} \left<\left< \{\gamma_{\alpha} {\circ}v_{g(4)} \} \right>\right>_{1}.\end{aligned}$$ Note that $G_{0}$ is completely determined by genus-0 data, while each term in $G_{1}$ contains genus-1 information. These two tensors are connected by Getzler’s equation (cf. [@Ge]): $$\label{eqn:Ge} G_{0} + G_{1} = 0.$$ Theorem \[thm:virD\] is obtained by applying this equation to $v_1 = v_2 = E$, $v_3 = {\gamma^{\alpha}}$, $v_4 = {\gamma_{\alpha}}$, and summing over $\alpha$. We first consider the genus-$1$ part of Equation . \[lem:G1EED\] $$\sum_{\alpha} G_1(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}}) = 24 \Delta {\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}}.$$ [**Proof**]{}: We will use the convention that repeated indices should be summed over their entire range. Therefore we will omit $\sum_{\alpha}$ in the left hand of this formula. To compute $G_1(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}})$, we notice that $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \{ E {\circ}{\gamma_{\alpha}}\} \, \{ {\gamma^{\alpha}}{\circ}E \} \, \right> \hspace{-2pt} \right>_{1}} &=& {\left< \hspace{-2pt} \left< \, E {\gamma_{\alpha}}{\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, E {\gamma^{\alpha}}{\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}{\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{1}} \\ &=& {\left< \hspace{-2pt} \left< \, E E {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma^{\alpha}}{\gamma^{\beta}}{\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}{\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{1}} \\ &=& {\left< \hspace{-2pt} \left< \, \{E^2 {\circ}{\gamma^{\mu}}\} \, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{1}} = {\left< \hspace{-2pt} \left< \, \{E^2 {\circ}{\gamma^{\alpha}}\} \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{1}} .\end{aligned}$$ In the second equality, we have used the associativity of the quantum product. This observation enables us to simplify the formula for $G_1(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}})$ and obtain $$\begin{aligned} && G_1(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}}) \nonumber \\ &=& 24 {\left< \hspace{-2pt} \left< \, E^2 \, \Delta \, \right> \hspace{-2pt} \right>_{1}} - 48 {\left< \hspace{-2pt} \left< \, \{ E {\circ}\Delta\} \, E \, \right> \hspace{-2pt} \right>_{1}} - 4 {\left< \hspace{-2pt} \left< \, E^2 \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} \nonumber \\ && - 16 {\left< \hspace{-2pt} \left< \, \{E {\circ}{\gamma^{\alpha}}\} \, {\gamma_{\alpha}}\, E \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} - 4 {\left< \hspace{-2pt} \left< \, \Delta \, E \, E \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} \nonumber \\ && + 24 {\left< \hspace{-2pt} \left< \, E \, E \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, \{ {\gamma_{\alpha}}{\circ}{\gamma_{\beta}}\} \, \right> \hspace{-2pt} \right>_{1}} + 24 {\left< \hspace{-2pt} \left< \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, \{ {\gamma_{\beta}}{\circ}E \} \, \right> \hspace{-2pt} \right>_{1}}. \label{eqn:G1EE}\end{aligned}$$ We now use formulas in Section \[sec:Euler\] to compute each term on the right hand side of this equation. Using equation , we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E^2 \, \Delta \, \right> \hspace{-2pt} \right>_{1}} &=& \Delta {\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}} - {\left< \hspace{-2pt} \left< \, \left\{ \nabla_{\Delta} E^2 \right\} \, \right> \hspace{-2pt} \right>_{1}} \\ &=& \Delta {\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}} - {\left< \hspace{-2pt} \left< \, \left\{ \Delta {\circ}G(E) - G(\Delta) {\circ}E - G(\Delta {\circ}E) + (b_1+2) \Delta {\circ}E \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ Since ${\left< \hspace{-2pt} \left< \, E \, \right> \hspace{-2pt} \right>_{1}}$ is a constant due to the quasi-homogeneity equation, by equation , we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \{ E {\circ}\Delta\} \, E \, \right> \hspace{-2pt} \right>_{1}} &=& \{ E {\circ}\Delta\} {\left< \hspace{-2pt} \left< \, E \, \right> \hspace{-2pt} \right>_{1}} - {\left< \hspace{-2pt} \left< \, \left\{ \nabla_{E {\circ}\Delta} E \right\} \, \right> \hspace{-2pt} \right>_{1}} \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G( E {\circ}\Delta) - (b_{1}+1) E {\circ}\Delta \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ By equation , we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E^2 \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} &=& {\left< \hspace{-2pt} \left< \, \left\{ (2-b_1) E {\circ}\Delta - G(E {\circ}\Delta) - G(E) {\circ}\Delta - E {\circ}G(\Delta) \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ By equation , we have $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, \{E {\circ}{\gamma^{\alpha}}\} \, {\gamma_{\alpha}}\, E \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G(E {\circ}{\gamma^{\alpha}}) {\circ}{\gamma_{\alpha}}+ E {\circ}{\gamma^{\alpha}}{\circ}G({\gamma_{\alpha}}) - G(E {\circ}\Delta) - b_1 E {\circ}\Delta \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ As a convention, we arrange the basis $\{\gamma_{1}, \ldots, \gamma_{N}\}$ of $H^{*}(M, \mathbb{C})$ in such a way that the degree $p_{\alpha} + q_{\alpha}$ of $\gamma_{\alpha} \in H^{p_{\alpha}, q_{\alpha}}$ is non-decreasing with respect to $\alpha$ and if two cohomology classes have the same dimension, we also require that the holomorphic dimension $p_{\alpha}$ is non-decreasing. Under this convention, we have $$G({\gamma^{\alpha}}) = (1- b_{\alpha}) {\gamma^{\alpha}}$$ for all $\alpha$, and $$G({\gamma^{\alpha}}) {\circ}{\gamma_{\alpha}}= \Delta - {\gamma^{\alpha}}{\circ}G({\gamma_{\alpha}}).$$ On the other hand, $$G({\gamma^{\alpha}}) {\circ}{\gamma_{\alpha}}= \eta^{\alpha \beta} G({\gamma_{\beta}}) {\circ}{\gamma_{\alpha}}= G({\gamma_{\beta}}) {\circ}{\gamma^{\beta}}= {\gamma^{\alpha}}{\circ}G({\gamma_{\alpha}}).$$ So we must have $$\label{eqn:Gaua} G({\gamma^{\alpha}}) {\circ}{\gamma_{\alpha}}= {\gamma^{\alpha}}{\circ}G({\gamma_{\alpha}}) = \frac{1}{2} \, \Delta.$$ Hence $$\begin{aligned} G(E {\circ}{\gamma^{\alpha}}) {\circ}{\gamma_{\alpha}}&=& {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} G({\gamma_{\beta}}) {\circ}{\gamma_{\alpha}}= G({\gamma_{\beta}}) {\circ}(E {\circ}{\gamma^{\beta}}) \nonumber \\ &=& \frac{1}{2} \, E {\circ}\Delta. \label{eqn:EGaua}\end{aligned}$$ Therefore we obtain $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \{E {\circ}{\gamma^{\alpha}}\} \, {\gamma_{\alpha}}\, E \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} &=& {\left< \hspace{-2pt} \left< \, \left\{ (1-b_1) E {\circ}\Delta - G(E {\circ}\Delta) \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ Similarly, $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \Delta \, E \, E \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, \right> \hspace{-2pt} \right>_{1}} &=& {\left< \hspace{-2pt} \left< \, \left\{ G(\Delta) {\circ}E + \Delta {\circ}G(E) - G(\Delta {\circ}E) - b_1 \Delta {\circ}E \right\} \, \right> \hspace{-2pt} \right>_{1}},\end{aligned}$$ and $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, E \, E \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, \{ {\gamma_{\alpha}}{\circ}{\gamma_{\beta}}\} \, \right> \hspace{-2pt} \right>_{1}} \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ {\gamma_{\alpha}}{\circ}\left( G(E) {\circ}{\gamma^{\alpha}}+ E {\circ}G({\gamma^{\alpha}}) - G(E {\circ}{\gamma^{\alpha}}) - b_1 E {\circ}{\gamma^{\alpha}}\right) \right\} \, \right> \hspace{-2pt} \right>_{1}} \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G(E) {\circ}\Delta - b_1 E {\circ}\Delta \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ To compute the last term in equation , we first compute $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\gamma_{\beta}}&=& G({\gamma_{\alpha}}) {\circ}{\gamma^{\alpha}}+ {\gamma_{\alpha}}{\circ}G({\gamma^{\alpha}}) - G(\Delta) - b_1 \Delta \nonumber \\ &=& (1-b_1) \Delta - G(\Delta) \label{eqn:Eaua}\end{aligned}$$ by equation . So the last term in equation is $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, \{ {\gamma_{\beta}}{\circ}E \} \, \right> \hspace{-2pt} \right>_{1}} &=& {\left< \hspace{-2pt} \left< \, \left\{ (1-b_1) E {\circ}\Delta - E {\circ}G(\Delta) \right\} \, \right> \hspace{-2pt} \right>_{1}}.\end{aligned}$$ After plugging the above formulas into equation , all terms on the right hand side cancel except the term $24 \Delta {\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}}$. The lemma is thus proved. $\Box$ Now we consider the genus-$0$ part of Equation . Let $$\label{eqn:PhiDef} \Phi := - \frac{1}{24} \sum_{\alpha} \left<\left< E E \gamma_{\alpha} \gamma^{\alpha} \right>\right>_{0} + \frac{1}{2}\sum_{\alpha} \left(b_{\alpha} (1-b_{\alpha}) - \frac{b_{1}+1}{6}\right) \left<\left< \gamma_{\alpha} \gamma^{\alpha} \right>\right>_{0}.$$ Then $$\Psi = {\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}} - \Phi$$ and the genus-1 Virasoro conjecture can be reduced to $${\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}} = \Phi.$$ \[lem:G0EED\] $$\sum_{\alpha} G_0(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}}) = -24 \Delta \Phi.$$ [**Proof**]{}: Again we will assume that repeated indices will be summed over their entire range. First, by definition of $G_0$, we have $$\begin{aligned} G_0(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}}) &=& 2{\left< \hspace{-2pt} \left< \, E \, E \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, {\gamma_{\alpha}}\, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} + 2 {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ && + {\left< \hspace{-2pt} \left< \, E \, E \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \Delta \, \right> \hspace{-2pt} \right>_{0}} - 2 {\left< \hspace{-2pt} \left< \, E \, E \, {\gamma^{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ && - 4 {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:G0EE}\end{aligned}$$ Note that the first and the fourth terms on the right hand side are canceled with each other. Applying equation to the second term, we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} &=& {\left< \hspace{-2pt} \left< \, \left\{ (1- b_1) \Delta - G(\Delta) \right\} \, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:EDED}\end{aligned}$$ Using equation again, we obtain $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \Delta \, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} &=& {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, \left\{(1-b_1) \Delta - G(\Delta) \right\} {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \nonumber\end{aligned}$$ Moreover, $$\begin{aligned} {\left< \hspace{-2pt} \left< \, G(\Delta) \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} &=& b_{\mu} {\left< \hspace{-2pt} \left< \, {\gamma^{\beta}}\, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \\ &=& {\left< \hspace{-2pt} \left< \, {\gamma^{\beta}}\, {\gamma_{\beta}}\, \left\{{\gamma^{\mu}}- G({\gamma^{\mu}}) \right\} \, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \\ &=& {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} - {\left< \hspace{-2pt} \left< \, {\gamma^{\beta}}\, {\gamma_{\beta}}\, G(\Delta) \, \right> \hspace{-2pt} \right>_{0}}.\end{aligned}$$ Moving the second term on the right hand to the left hand, we obtain $$\label{eqn:GDD} {\left< \hspace{-2pt} \left< \, G(\Delta) \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{2} \, {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}.$$ Hence, we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \Delta \, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} &=& \left(\frac{1}{2} - b_1 \right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:EDD}\end{aligned}$$ By equation , we have $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& (1- b_1) \left(\frac{1}{2} - b_1 \right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} - {\left< \hspace{-2pt} \left< \, G(\Delta) \, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:EDED2}\end{aligned}$$ To compute the last term on the right hand side of equation , we set $$f := {\left< \hspace{-2pt} \left< \, E \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}.$$ Applying equation , we have $$\begin{aligned} f &=& {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma^{\alpha}}) {\circ}{\gamma^{\mu}}+ {\gamma^{\alpha}}{\circ}G({\gamma^{\mu}}) - G({\gamma^{\alpha}}{\circ}{\gamma^{\mu}}) - b_1 {\gamma^{\alpha}}{\circ}{\gamma^{\mu}}\right\} \, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& (2-b_{\alpha}- b_{\mu}-b_1) {\left< \hspace{-2pt} \left< \, \left\{ {\gamma^{\alpha}}{\circ}{\gamma^{\mu}}\right\} \, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} - {\left< \hspace{-2pt} \left< \, G({\gamma^{\alpha}}{\circ}{\gamma^{\mu}}) \, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& (2-2b_{\alpha}-b_1) {\left< \hspace{-2pt} \left< \, {\gamma^{\alpha}}\, {\gamma^{\mu}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\beta}}\, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} - {\left< \hspace{-2pt} \left< \, {\gamma^{\alpha}}\, {\gamma^{\mu}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, G({\gamma_{\beta}}) \, {\gamma_{\mu}}\, E \, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \nonumber\end{aligned}$$ Switching $\alpha$ and $\beta$ in the last term, we have $$f = (2-b_1){\left< \hspace{-2pt} \left< \, E \, {\gamma_{\alpha}}\, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} -3{\left< \hspace{-2pt} \left< \, E \, G({\gamma_{\alpha}}) \, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:f3}$$ Applying equation again, we have $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, E \, {\gamma_{\alpha}}\, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}{\gamma_{\beta}}+ {\gamma_{\alpha}}{\circ}G({\gamma_{\beta}}) - G({\gamma_{\alpha}}{\circ}{\gamma_{\beta}}) - b_1 {\gamma_{\alpha}}{\circ}{\gamma_{\beta}}\right\} \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}. \nonumber\end{aligned}$$ By the associativity of the quantum product and equation , $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}{\gamma_{\beta}}\right \} \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} &=& {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}{\gamma^{\alpha}}\right \} \, {\gamma_{\beta}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \, = \, \frac{1}{2} \, {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \label{eqn:GD}\end{aligned}$$ and $$\begin{aligned} {\left< \hspace{-2pt} \left< \, G({\gamma_{\alpha}}{\circ}{\gamma_{\beta}}) \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} &=& {\left< \hspace{-2pt} \left< \, {\gamma_{\alpha}}\, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, G({\gamma_{\mu}}) \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, {\gamma_{\alpha}}\, \left\{ G({\gamma_{\mu}}) {\circ}{\gamma^{\alpha}}\right\} \, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{2} \, {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}.\end{aligned}$$ So we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E \, {\gamma_{\alpha}}\, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}}{\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} &=& \left(\frac{1}{2} - b_1 \right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:EDDD}\end{aligned}$$ Moreover $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, E \, G({\gamma_{\alpha}}) \, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G(G({\gamma_{\alpha}})) {\circ}{\gamma_{\beta}}+ G({\gamma_{\alpha}}) {\circ}G({\gamma_{\beta}}) - G(G({\gamma_{\alpha}}) {\circ}{\gamma_{\beta}}) - b_{1} G({\gamma_{\alpha}}) {\circ}{\gamma_{\beta}}\right\} \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}}. \nonumber\end{aligned}$$ Since $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}G({\gamma_{\beta}}) \right\} \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}{\gamma^{\alpha}}\right\} \, G({\gamma_{\beta}}) \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{2} \, {\left< \hspace{-2pt} \left< \, \left\{ {\gamma_{\alpha}}{\circ}{\gamma^{\alpha}}\right\} \, G({\gamma_{\beta}}) \, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& \frac{1}{2} \, {\left< \hspace{-2pt} \left< \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \left\{ G({\gamma_{\beta}}) {\circ}{\gamma^{\beta}}\right\} \, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{4} \, {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}\end{aligned}$$ and $$\begin{aligned} && {\left< \hspace{-2pt} \left< \, G( G({\gamma_{\alpha}}) {\circ}{\gamma_{\beta}}) \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, G({\gamma_{\alpha}}) \, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, G({\gamma_{\mu}}) \, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} = {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}{\gamma^{\mu}}\right\} \, G({\gamma_{\mu}}) \, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& {\left< \hspace{-2pt} \left< \, \left\{ G({\gamma_{\alpha}}) {\circ}{\gamma^{\alpha}}\right\} \, G({\gamma_{\mu}}) \, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{2} {\left< \hspace{-2pt} \left< \, \left\{ {\gamma_{\alpha}}{\circ}{\gamma^{\alpha}}\right\} \, G({\gamma_{\mu}}) \, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ &=& \frac{1}{2} {\left< \hspace{-2pt} \left< \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \left\{ G({\gamma_{\mu}}) {\circ}{\gamma^{\mu}}\right\} \, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{4} \, {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}},\end{aligned}$$ together with equation , we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, E \, G({\gamma_{\alpha}}) \, {\gamma_{\beta}}\, {\gamma^{\mu}}\, \right> \hspace{-2pt} \right>_{0}} {\left< \hspace{-2pt} \left< \, {\gamma_{\mu}}\, {\gamma^{\alpha}}\, {\gamma^{\beta}}\, \right> \hspace{-2pt} \right>_{0}} &=& \left(b_{\alpha}^2 - \frac{1}{2} b_{1} \right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:EGD}\end{aligned}$$ Combining results of equations , , and , we obtain that the last term on the right hand side of equation is $$f = (-3 b_{\alpha}^2 + b_{1}^2 - b_{1} + 1) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}.$$ Together with equation , we can simplify equation as $$\begin{aligned} G_0(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}}) &=& {\left< \hspace{-2pt} \left< \, E \, E \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \Delta \, \right> \hspace{-2pt} \right>_{0}} - 2 {\left< \hspace{-2pt} \left< \, G(\Delta) \, E \, {\gamma^{\mu}}\, {\gamma_{\mu}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ && + (12 b_{\alpha}^2 - 2 b_{1}^2 + b_{1} -3) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \label{eqn:G0EE2}\end{aligned}$$ On the other hand, by the definition of $\Phi$ in equation , we have $$\begin{aligned} 24 \Delta \Phi &=& - {\left< \hspace{-2pt} \left< \, \Delta \, E \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} - 2 {\left< \hspace{-2pt} \left< \, \left\{ \nabla_{\Delta} E \right\} \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ && + 12 \left(b_{\alpha} (1-b_{\alpha}) - \frac{b_{1}+1}{6}\right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \nonumber\end{aligned}$$ By equations and , we have $$\begin{aligned} {\left< \hspace{-2pt} \left< \, \left\{ \nabla_{\Delta} E \right\} \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} &=& {\left< \hspace{-2pt} \left< \, \left\{ - G(\Delta) + (b_{1}+1) \Delta \right\} \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \\ &=& - {\left< \hspace{-2pt} \left< \, G(\Delta) \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} + (b_{1}+1) \left(\frac{1}{2} - b_1 \right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}.\end{aligned}$$ Moreover, $$b_{\alpha}{\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} = {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, G({\gamma_{\alpha}}) \, \right> \hspace{-2pt} \right>_{0}} = \frac{1}{2} {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma^{\alpha}}\, {\gamma_{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}.$$ So we have $$\begin{aligned} 24 \Delta \Phi &=& - {\left< \hspace{-2pt} \left< \, \Delta \, E \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} + 2 {\left< \hspace{-2pt} \left< \, G(\Delta) \, E \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}} \nonumber \\ && + \left(- 12 b_{\alpha}^2 + 2 b_1^2 - b_1 + 3 \right) {\left< \hspace{-2pt} \left< \, \Delta \, {\gamma_{\alpha}}\, {\gamma^{\alpha}}\, \right> \hspace{-2pt} \right>_{0}}. \nonumber\end{aligned}$$ Comparing with equation , we obtain $$G_0(E, E, {\gamma^{\alpha}}, {\gamma_{\alpha}}) = - 24 \Delta \Phi.$$ The lemma is thus proved. $\Box$ [**Proof of Theorem \[thm:virD\]**]{}: Since $\Psi= {\left< \hspace{-2pt} \left< \, E^2 \, \right> \hspace{-2pt} \right>_{1}}- \Phi$, this theorem follows from Lemmas \[lem:G1EED\], \[lem:G0EED\] and Equation . $\Box$ [399]{} Cox, D., and Katz, S., [*Mirror symmetry and algebraic geometry*]{}. Providence, R.I. AMS, 1999. Dubrovin, B., Zhang, Y., [*Bihamiltonian hierarchies in 2D topological field theory at one-loop approximation*]{}, Comm. Math. Phys. 198 (1998) 311 - 361. Eguchi, T., Hori, K., and Xiong, C., [*Quantum Cohomology and Virasoro Algebra*]{}, Phys. Lett. B402 (1997) 71-80. Getzler, E., [*Intersection theory on $\bar{M}_{1,4}$ and elliptic Gromov-Witten Invariants*]{}, J. Amer. Math. Soc. 10 (1997) 973-998 Kontsevich, M., [*Intersection theory on the moduli space of curves and the matrix airy function*]{}, Comm. Math. Phys., 147 (1992). Li, J. and Tian, G., [*Virtual moduli cycles and Gromov-Witten invariants of general symplectic manifolds*]{}, Topics in symplectic 4-manifolds (Irvine, CA, 1996), 47-83. Liu, X., Elliptic Gromov-Witten invariants and Virasoro conjecture. [*Comm. Math. Phys.*]{} 216 (2001), 705-728. Liu, X., Genus-1 Virasoro conjecture on the small phase space, [*Symplectic Geometry and Mirror Symmetry*]{}, Preceedings of the 4th KIAS Annual International Conference, World Scientific, (2001), 265-279. Liu, X., and Tian, G., Virasoro constraints for quantum cohomology. [*J. Diff. Geom.*]{} 50 (1998), 537 - 591. Ruan, Y. and Tian, G., [*Higher genus symplectic invariants and sigma models coupled with gravity*]{}, Invent. Math. 130 (1997), 455-516. Witten, E., [*Two dimensional gravity and intersection theory on Moduli space*]{}, Surveys in Diff. Geom., 1 (1991), 243-310. Xiaobo Liu\ Beijing International Center for Mathematical Research,\ Beijing University, Beijing, China.\ &\ Department of Mathematics,\ University of Notre Dame,\ Notre Dame, IN 46556, USA\ Email: [*[email protected]*]{} [^1]: Research was partially supported by NSF grant DMS-0905227.
{ "pile_set_name": "ArXiv" }
--- abstract: 'It is known that, in continuous media, composite solitons with *hidden vorticity*, which are built of two mutually symmetric vortical components whose total angular momentum is zero, may be stable while their counterparts with explicit vorticity and nonzero total angular momentum are unstable. In this work, we demonstrate that the opposite occurs in discrete media: hidden vortex states in relatively small ring chains become unstable with the increase of the total power, while explicit vortices are stable, provided that the corresponding scalar vortex state is also stable. There are also stable mixed states, in which the components are vortices with different topological charges. Additionally, degeneracies in families of composite vortex modes lead to the existence of long-lived breather states which can exhibit vortex charge flipping in one or both components.' address: | $^1$Nonlinear Physics Centre, Research School of Physics and Engineering,\ The Australian National University, Canberra ACT 0200, Australia\ $^2$Department of Physical Electronics, School of Electrical Engineering,\ Faculty of Engineering, Tel Aviv University, Tel Aviv 69978, Israel author: - 'Daniel Leykam$^1$, Boris Malomed$^2$, and Anton S. Desyatnikov$^1$' title: Composite vortices in nonlinear circular waveguide arrays --- =1 Introduction ============ Optical vortex solitons [@desyatnikov_PROG], i.e., self-trapped beams containing phase singularities [nye1974,soskin,Pramana,review,review2,mark]{}, present an ideal setting for studying the relationship between topology and self-action effects, and may have applications to optical data transmission and processing [gibson2004, willner2012]{}. However, in local nonlinear media vortex solitons are often destroyed by the modulational instability, and their orbital angular momentum (OAM) is transferred to multiple filaments [firth1997,Torner,Torner2]{}. There are several approaches to suppressing the azimuthal instability of vortex solitons. Models with competing nonlinearities, such as cubic-quintic (CQ) [@Manuelo; @malomed2002], quadratic-cubic [@QC; @Herve] and nonlocal media [@NL1; @NL2] can support stable scalar vortex solitons. Two-component vortices may also be stabilized by the CQ nonlinearity [vectorCQ]{}. An alternative approach is to apply spatial modulation to the self-defocusing nonlinearity, making its local strength to grow, with radius $r$, faster than $r^{2}$ [@Barcelona], or (more often) to apply a spatially periodic modulation to the refractive index of a nonlinear medium [@Kart]. In the limit case of the deep periodic modulation, the medium reduces to an array of waveguides, the propagation through which is governed by discrete equations [@christo2003; @lederer2008]. When the nonlinear self-action suppresses the discrete diffraction in the arrays, discrete solitons emerge [@lederer2008]. The stability of discrete vortex solitons was predicted [@malomed2001; @pelinovsky2005] and subsequently observed in experiments [@neshev2004; @fleischer2004]. Interestingly, the stability hierarchy is inverted with respect to continuous media: higher-order vortices, including “supervortices" (ring patterns built of compact discrete vortex solitons) [@super], tend to be stable, while their lower order counterparts suffer instabilities [higherDiscrVort,law2009, terhalle2009, desyatnikov2011]{}. On the other hand, vortex solitons can also be stabilized by the action of the cross-phase modulation (XPM) between two or more mutually incoherent components of a composite beam. In addition to the above-mentioned two-component model with the CQ nonlinearity [@vectorCQ], examples include multipole solitons [@MP01; @MP02], two- and three-component necklace ring patterns [@desyatnikov2001; @boyd; @dip3_OL; @dip3_OC], and their counterparts in the Gross-Pitaevskii equation for Bose-Einstein condensates [@BEC; @Brtka]. Interactions between two vortical composite solitons were studied in the model with the saturable nonlinearity [Moti3]{}. Also studied were composite modes in which one component is vortical, while the other one is represented by a fundamental soliton [Moti]{}. In particular, solitons composed of vortex beams with oppositely rotating vortices, including symmetric counter-rotating pairs with the *hidden vorticity* (HV), whose OAM is exactly zero, were predicted to be much more robust than their counterparts with the explicit vorticity [desyatnikov2001,boyd,Brtka,desyatnikov2005,desyatnikov2007]{}. This feature can be demonstrated analytically in the framework of the one-dimensional two-component system, in which a counterpart of the HV states is represented by *hidden-momentum* counter-propagating wave pairs, with equal amplitudes and zero total momentum [@desyatnikov2007]. In the general case of an arbitrary number of symmetrically interacting components, the stability is determined by the total OAM of the composite beam [Moti2,math]{}. Vortex solitons of the HV type were recently observed in nematic liquid crystals [@izdebskaya2012]. A specific example of composite (*two-color*) solitary vortices is provided by those in media with the quadratic ($\chi ^{(2)}$) nonlinearity. While they are always unstable against splitting in the uniform media [firth1997,Torner,Torner2]{}, it was recently demonstrated that both single- and two-color vortices in $\chi ^{(2)}$ media can be stabilized by an external trapping potential [@HS]. HV modes can be defined in terms of $\chi ^{(2)}$ systems too, assuming that the fundamental-frequency beam is built of two different components, corresponding to orthogonal polarizations, which are parametrically coupled to a single component of the second harmonic. The HV modes are unstable in that system (without an external potential), but the addition of the competing self-defocusing cubic nonlinearity makes them almost completely stable [@HS]. This work aims to study combined effects of the above-mentioned approaches to the stabilization of vortices, *viz*., composite discrete vortex solitons in ring-shaped nonlinear lattices. We find that nonlinear modes of the HV type are subject to instabilities, which demonstrates that the above-mentioned *inverse* relation between the scalar and vortex stabilities and instabilities in discrete media, in comparison with continua, persists for composite modes as well. An additional feature of our discrete system is the existence of *mixed-charge* composite vortices. We conclude that their stability is tied to the topological charge of their brightest component. Further, an azimuthal modulation applied to discrete composite vortices may continuously deform them through a family of modes to *discrete necklace solitons*. When both components have equal total powers, this family is degenerate, and these deformations may be realized *dynamically* by perturbing a stationary mode. This can lead to simultaneous charge flipping of both components, similar to previous results in continua [@desyatnikov2007]. Additionally, due to the broken rotational symmetry of the discrete system, charge flipping of a single component only may occur too. We start by introducing the model, and obtaining analytically a class of “separable" nonlinear modes, in Section \[sec:model\]. Their stability is studied in Section \[sec:stability\]. The dynamics of perturbed modes and vortex-charge flipping are presented in Section \[sec:dynamics\]. Section \[sec:final\] concludes the paper and discusses possible experimental realizations. The model and vortex modes {#sec:model} ========================== We consider the discrete one-dimensional model governing the propagation of two incoherently coupled beams with amplitudes $A_{n}(z)$ and $B_{n}(z)$ through an array of nonlinear waveguides: $$\begin{aligned} &&i\partial _{z}A_{n}+A_{n-1}+A_{n+1}+(|A_{n}|^{2}+|B_{n}|^{2})A_{n}=0, \label{eqn:model} \\ &&i\partial _{z}B_{n}+B_{n-1}+B_{n+1}+(|A_{n}|^{2}+|B_{n}|^{2})B_{n}=0.\end{aligned}$$We apply periodic boundary conditions, $n+N\equiv n$, to define the ring-shaped arrays \[cf. Ref. [@Panos; @cuevas2009; @alvarez2011] which considered in detail similar equations on an infinite chain\]. The system conserves the powers, $P_{A}=\sum_{n}|A_{n}|^{2},P_{B}=\sum_{n}|B_{n}|^{2}$, and the Hamiltonian,$$H=\sum_{n}\left[ A_{n}A_{n+1}^{\ast }+A_{n}^{\ast }A_{n+1}+B_{n}B_{n+1}^{\ast }+B_{n}^{\ast }B_{n+1}+\frac{1}{2}\left( |A_{n}|^{2}+|B_{n}|^{2}\right) ^{2}\right] . \label{eqn:hamiltonian}$$The current flows in the two components between adjacent sites, $n$ and $n+1$, are $J_{n,n+1}^{(A)}=2\mathrm{Im}(A_{n}^{\ast }A_{n+1})$ and $J_{n,n+1}^{(B)}=2\mathrm{Im}(B_{n}^{\ast }B_{n+1})$. Discrete vortices correspond to circulation of the currents around the ring, and are characterized by the integer topological charges, $$m_{A}=\frac{1}{2\pi }\sum_{n}\mathrm{arg}(A_{n}^{\ast }A_{n+1}),\quad m_{B}=\frac{1}{2\pi }\sum_{n}\mathrm{arg}(B_{n}^{\ast }B_{n+1}),$$which take values within the range of $-N/2<m_{A,B}<N/2$ [@desyatnikov2011; @ferrando2005]. There is no vortex in a component if its charge vanishes. Because the Hamiltonian (\[eqn:hamiltonian\]) implies the Manakov-like nonlinearity, with equal coefficients of the XPM and SPM nonlinearities [Manakov]{}, it possesses an additional symmetry, associated with rotations that mix the two components, while preserving the power at each site: $$R_{1}(\varphi )=\left( \begin{array}{cc} \cos \varphi & \sin \varphi \\ -\sin \varphi & \cos \varphi \end{array}\right) ,R_{2}(\theta )=\left( \begin{array}{cc} \cos \theta & i\sin \theta \\ i\sin \theta & \cos \theta \end{array}\right) ,$$where the rotation matrices act on vector $(A_{n},B_{n})$. The corresponding conserved quantity is commonly called the isospin, $S=\sum_{n}A_{n}B_{n}^{\ast }$, which exists in the continuum limit too [@math]. Similar to the above definition of currents $J_{n,n+1}^{\left( A,B\right) }$, quantity $S-S^{\ast }=2\mathrm{Im}S$ is the total isospin-power flow. Since the coupling between components is incoherent, i.e., they cannot exchange powers, this flow must always be zero. Consequently, a global phase shift can always be applied to one of the components to set $\mathrm{Im}S=0$. Therefore in this case the isospin has no physical significance, but the fact that it is conserved during the propagation will have some consequences later. We look for nonlinear modes as $(A_{n},B_{n})=(U_{n}e^{i\beta _{A}z},V_{n}e^{i\beta _{B}z})$, where $\beta _{A,B}$ are the propagation constants and $U_{n},V_{n}$ are the site amplitudes. Here we consider a special class of solutions, similar to necklace-ring vector solitons in bulk nonlinear media [@desyatnikov2001], with constant total intensity on the ring, $$|U_{n}|^{2}+|V_{n}|^{2}=I\quad \mbox{for all}\quad n. \label{I}$$Under this constraint, the nonlinearity $\sim I$ is factorized, and the stationary equations for the amplitudes separate into two effectively decoupled linear equations, *viz*.,$$\begin{aligned} &(I-\beta _{A})U_{n}+U_{n-1}+U_{n+1}=&0, \label{stat} \\ &(I-\beta _{B})V_{n}+V_{n-1}+V_{n+1}=&0.\end{aligned}$$Exploiting the linearity of (6), we can apply the discrete Fourier transform, $(U_{n},V_{n})=\sum_{s}(a_{s},b_{s})e^{i\Theta _{s}n}$, where $\Theta _{s}=2\pi s/N$ is the phase winding of the Fourier mode with the discrete vortex of charge $s$. We thus obtain an analog of the dispersion relations [@desyatnikov2011], $\beta _{A,B}=I+2\cos \Theta _{p,q}$, where the mode indices, $p$ and $q$, may be different for two components with unequal propagation constants, $\beta _{A}\neq \beta _{B}$. Note that they are degenerate with respect to the sign of the mode indices $p$ and $q$, thus we should take the superpositions, $U_{n}=a_{+}e^{i\Theta _{p}n}+a_{-}e^{-i\Theta _{p}n}$ and $V_{n}=b_{+}e^{i\Theta _{p}n}+b_{-}e^{-i\Theta _{p}n}$, as a general solution. We can use the $U(1)$ invariance of each component to set $\mathrm{arg}a_{+}=-\mathrm{arg}a_{-}$, $\mathrm{arg}b_{+}=-\mathrm{arg}b_{-}$ without the loss of generality. Applying the change of variables, $a_{\pm }\equiv r_{\pm }e^{\pm i\chi _{1}}$, $b_{\pm }\equiv s_{\pm }e^{\pm i\chi _{2}}$, and $\Delta \equiv (1/2)(I-r_{+}^{2}-r_{-}^{2}-s_{+}^{2}-s_{-}^{2})$, the constraint equation (\[I\]) takes the form of $$\Delta =r_{+}r_{-}\cos [2(\Theta _{p}n+\chi _{1})]+s_{+}s_{-}\cos [2(\Theta _{q}n+\chi _{2})], \label{eqn:constraint}$$which must be satisfied for each value of $n=1,...,N$. This will fix some of the six parameters $r_{\pm },s_{\pm },\chi _{1,2}$. Solutions with symmetries will have redundancies in the $N$ constraint equations, leaving free parameters. In the continuum limit, there is an infinite number of constraints, hence only symmetric solutions survive in that limit. On the other hand, additional modes may exist in the discrete system with a small number of sites. We will focus on a particular family of solutions with three parameters $(P,\phi ,\varphi )$, $$\begin{aligned} U_{n} &=&\sqrt{P}\left( \cos \phi \right) [\left( \cos \varphi \right) e^{i\Theta _{p}n}+\left( \sin \varphi \right) e^{-i\Theta _{p}n}], \label{U} \\ V_{n} &=&\sqrt{P}\left( \sin \phi \right) [\left( \cos \gamma \right) e^{i\Theta _{q}n}-\left( \sin \gamma \right) e^{-i\Theta _{q}n}], \label{V} \\ \gamma &=&\frac{1}{2}\sin ^{-1}\left( \left( \cot ^{2}\phi \right) \sin \left( 2\varphi \right) \right) , \label{gam}\end{aligned}$$which can be seen as a discrete generalization of the necklace-ring vector solitons [@desyatnikov2001]. Here $P$ sets the total power of the mode, $\phi \in \lbrack 0,\pi /2]$ determines the relative power in its two components, and $\varphi \in \lbrack -\pi /4,+\pi /4]$ defines the azimuthal modulation. Notice that when the two components have different powers ($\phi \neq \pi /4$), they require different modulation strengths, $\gamma \neq \varphi $, to maintain the constant intensity constraint (\[I\]). Obviously, (\[gam\]) has a real solution only if $|\varphi |\leq (1/2)\sin ^{-1}(\tan ^{2}\phi )$. At the maximum allowed value of $|\varphi | $, the intensity of the second component is zero at some sites. By varying $\phi $ and $\varphi $, a scalar vortex ($\phi =0,\varphi =0$) can be continuously deformed into a *symmetric composite vortex* ($\phi =\pi /4,\varphi =0$) or a *discrete necklace soliton* ($\phi =\pi /4,\varphi =\pi /4$). Examples of these types of the modes are displayed in figure \[fig:system\]. ![Examples of different modes belonging to family (8)-(10) for the ring with $N=6$. (a) A scalar vortex ($\protect\phi =\protect\varphi =0$) with charge $p$. (b) A symmetric vector vortex ($\protect\phi =\frac{\protect\pi }{4},\protect\varphi =0$) of charge ($p,q$). (c) A discrete necklace ($\protect\phi =\frac{\protect\pi }{4},\protect\varphi =\frac{\protect\pi }{4}$). (d) Another discrete necklace, with $\protect\phi =\frac{\protect\pi }{4},\protect\varphi =-\frac{\protect\pi }{4}$.[]{data-label="fig:system"}](figure1){width="80mm"} In addition to solutions with the *explicit* ($p=q$) and *hidden* ($p=-q$) vorticities, for even $N$ there also exist solutions with mixed vorticity, $p=\pm (N/2-q)$. These appear because the constraint ([eqn:constraint]{}) only needs to be satisfied at a discrete set of points. The linear stability {#sec:stability} ==================== Linear stability is studied by introducing small perturbations of the form $A_{n}=(U_{n}+a_{n}e^{\lambda z}+b_{n}^{\ast }e^{\lambda ^{\ast }z})e^{i\beta _{A}z}$, $B_{n}=(V_{n}+c_{n}e^{\lambda z}+d_{n}^{\ast }e^{\lambda ^{\ast }z})e^{i\beta _{B}z}$ and linearizing (1) to derive an eigenvalue problem for $\lambda $. It is convenient to introduce vectors $\mathbf{v}_{n}=(a_{n},b_{n},c_{n},d_{n})$, $\mathbf{E}_{n}=(A_{n},A_{n}^{\ast },B_{n},B_{n}^{\ast })$, writing the eigenvalue problem as $$(H_{L}+H_{NL})\mathbf{v}_{n}=i\lambda \mathrm{diag}(1,-1,1,-1)\mathbf{v}_{n}, \label{eqn:stability}$$where $H_{L}$ is diagonal in the component space (i.e., it does not couple $a_{n}$ to $b_{n}$, etc.), but couples adjacent sites: $$H_{L}a_{n}=(I-\beta _{A})a_{n}+a_{n-1}+a_{n+1},$$and similarly for the other components (with $\beta _{A}$ replaced by $\beta _{B}$ for $c_{n},d_{n}$). On the other hand, $H_{NL}$ couples different components, rather than different sites: $$H_{NL}=\left( \begin{array}{cccc} |U_{n}|^{2} & U_{n}^{2} & U_{n}V_{n}^{\ast } & U_{n}V_{n} \\ U_{n}^{\ast 2} & |U_{n}|^{2} & U_{n}^{\ast }V_{n}^{\ast } & U_{n}^{\ast }V_{n} \\ U_{n}^{\ast }V_{n} & U_{n}V_{n} & |V_{n}|^{2} & V_{n}^{2} \\ U_{n}^{\ast }V_{n}^{\ast } & U_{n}V_{n}^{\ast } & V_{n}^{\ast 2} & |V_{n}|^{2}\end{array}\right) .$$This can be compactly written as an outer product, $H_{NL}=\mathbf{E}_{n}\otimes \mathbf{E}_{n}^{\dagger }$. Equations (\[eqn:stability\]) can be solved numerically, as an eigenvalue problem of dimension $4N$. For large $N$, it is relevant to consider some simple limits. For example, when $\varphi =0$ the problem simplifies significantly, the application of the discrete Fourier transform decoupling it into a set of “smaller" eigenvalue problems, each of dimension 4: $$\left( \begin{array}{cc} L_{A} & M \\ M & L_{B}\end{array}\right) \mathbf{v}_{s}=i\lambda \mathbf{v}_{s},$$where the matrices are defined as $$\begin{aligned} &&L_{A}=\left( \begin{array}{cc} P\cos ^{2}\phi +\kappa _{p,+} & P\cos ^{2}\phi \\ -P\cos ^{2}\phi & -P\cos ^{2}\phi -\kappa _{p,-}\end{array}\right) , \\ &&L_{B}=\left( \begin{array}{cc} P\sin ^{2}\phi +\kappa _{q,+} & P\sin ^{2}\phi \\ -P\sin ^{2}\phi & -P\sin ^{2}\phi -\kappa _{q,-}\end{array}\right) , \\ &&M=\frac{P}{2}\sin 2\phi \left( \begin{array}{cc} 1 & 1 \\ -1 & -1\end{array}\right) ,\end{aligned}$$with $\kappa _{m,\pm }\equiv 2(\cos \Theta _{m\pm s}-\cos \Theta _{m})$, and $|s|\leq N/2$. In general, the eigenvalues cannot be obtained in an explicit form, but simple expressions are available in some limits. Stability diagrams of family (\[U\])-(\[gam\]) in the limit of $\varphi =0$ for $N=6$ are shown in figure \[fig:stability\] for different combinations of topological charges. ![Stable (shaded) and unstable (white) regions in the $(P,\protect\phi )$ parameter plane for composite vortices with different topological charges, indicated in the brackets ($p,q$), in the ring built of $N=6$ sites. Families of hidden-vorticity modes are represented by the top and middle panels in the right column.[]{data-label="fig:stability"}](figure2){width="80mm"} When both components have the same charge (which corresponds to the explicit vorticity), the instability threshold is independent of $\phi $, and the stability problem reduces to that for the scalar case: lower charges are unstable, while higher ones are stable [@desyatnikov2011]. Another situation takes place for the HV states with hidden vorticity: the HV mode (1,-1) encounters instability at a higher power than its scalar counterpart; ultimately, the (1,-1) mode becomes unstable at large powers, in contrast to stable HV solitons in continua [@desyatnikov2007]. Further, the HV configuration of the (2,-2) type also becomes unstable at large powers, in contrast to its explicit-vorticity (2,2) counterpart, which is completely stable. Similar stability features are exhibited by the mixed vortex states of the (2,1), (2,-1) types: the state with the greater total charge $p+q$ has a larger stability region. The stability of these solitons depends on the topological charge of the brighter component, i.e., whether the brighter component is stable in the scalar case. A qualitatively similar behavior is observed for other values of $N$, with the stability dependent on whether the vortex charges are low (smaller than $N/4$) or high (larger than $N/4$). Vortices with the charges equal to $N/4$ represent a special case, as they form a one-parameter family of asymmetric vortex modes, introduced in [@alexander2004]. This fact complicates the stability analysis, as the stability also depends on the additional parameter. When $\phi =\pi /4$, an additional pair of zero eigenvalues appears for the hidden- and mixed vortex modes $(p,-p),(p,N/2-p),(p,p-N/2)$, and the linear stability analysis can no longer predict whether the modes are stable. The zero eigenvalues are associated with the degeneracy of family (\[U\])-([gam]{}) with respect to $\varphi $. Calculating the values of the conserved quantities for the family, we obtain $P_{A}=NP\cos ^{2}\phi $, $P_{B}=NP\sin ^{2}\phi $, $H=2NP(\cos ^{2}\phi \cos \Theta _{p}+\sin ^{2}\phi \cos \Theta _{q})+NP^{2}/2$ and $S=(1/2)NP\sin \left( 2\phi \right) [\cos (\varphi +\gamma )\delta _{p,q}+\sin (\varphi -\gamma )\delta _{p,-q}]$. We see that for the HV mode with ($p,-p$), when $\phi =\pi /4$, all quantities are independent of the azimuthal modulation $\varphi $, which means that, under small perturbations, the input with $\varphi =0$ can cycle through solutions with different values of $\varphi $, hence we must consider the stability of the family as $\varphi $ is varied. This family can be obtained by applying an isospin rotation to the $\varphi =0$ mode [@math]: $$\left( \begin{array}{c} U_{n} \\ V_{n}\end{array}\right) =R_{1}(\varphi )\sqrt{P}\left( \begin{array}{c} e^{i\Theta _{p}n} \\ e^{-i\Theta _{p}n}\end{array}\right) ,$$which follows from the fact that (\[gam\]) is solved by $\gamma =\varphi $ in this case. This rotation commutes with operator $H_{L}+H_{NL}$ in the linear stability problem, see (\[eqn:stability\]). Following the result of [@math], the stability is consequently independent of $\varphi $. This result is supported by direct numerical solutions of the linear stability problem. We stress that for other value of $\phi $, the stability does depend on $\varphi $. ![The stability of mixed vortex modes as a function of the azimuthal-modulation parameter ($\protect\varphi $) and total power $P$. Shaded: marginal stability. Unshaded: linear instability[]{data-label="fig:stability2"}](figure3){width="80mm"} The conserved quantities for the mixed vortex modes are independent of $\varphi $ for any value of $\phi $. However, the above reduction of the stability problem to the $\varphi =0$ case is not relevant, as the family cannot be generated by the isospin rotation. Therefore, the stability of the family depends on $\varphi $, as shown in figure \[fig:stability2\]. The azimuthal modulation lowers the instability threshold for the explicit vortex modes, and it can slightly increase the threshold for the HV mode. In practice, the family will become unstable when $P$ exceeds the lowest instability threshold that occurs as $\varphi $ is varied. Dynamics {#sec:dynamics} ======== The existence of zero eigenvalues in the linear stability problem means that higher-order terms will determine whether the degenerate families are stable. To check the stability, (1) was solved numerically using perturbed initial conditions. Figure \[fig:flip1\] shows the propagation of a perturbed charge-2 HV mode in the waveguide ring built of six sites ($N=6$). The high-frequency oscillations of site powers correspond to frequencies of stable eigenvalues, while the additional low-frequency oscillations correspond to the zero eigenvalue. During this oscillation, the power in each component acquires an azimuthal modulation, however the sum $I_{n}=|A_{n}|^{2}+|B_{n}|^{2}$ remains (on average) constant. ![Adiabatic dynamics of the degenerate hidden-vortex family, with parameters $P=1,\protect\phi =\protect\pi /4$, $\protect\varphi =\protect\pi /8$, topological charges (2,-2), and random perturbations added to the initial conditions at the 10% level. (a,c): On-site powers for the two components. (b,d): Vortex lines (red – positive charge, blue – negative charge).[]{data-label="fig:flip1"}](figure4){width="60mm"} When $z\approx 25$, the amplitudes of the $m=2$ and $m=-2$ Fourier modes are equal and the topological charges of both components vanish \[see figure \[fig:flip1\](b,d)\]. Each component has two pairs of sites with equal powers, and the power vanishes at a pair of sites in one component. The beam profiles at this point resembles those in figure [fig:system]{}(c), which corresponds to a discrete necklace beam with $\varphi =\pi /4$. Increasing $\varphi $ further, the charges of both components flip, and at $z\approx 100$ the azimuthal modulation vanishes, as $\varphi $ attains value $\pi /2$. There is a second charge flip at $z\approx 150$ ($\varphi =3\pi /4$), this time with the other component hosting the pair of vanishing site powers. Thus we see that the slow oscillations correspond to an adiabatic cycling through the degenerate family of solutions parametrized by $\varphi $. No member of the family is subject to linear instability, therefore the oscillations persist indefinitely long (in excess of $\Delta z=5000$, according to our numerical results). In contrast to the picture for the HV model outlined above, instability can occur during the adiabatic cycling of the mixed vortex modes, since their stability depends on $\varphi $. This is shown in figure \[fig:dynamics\], in which the initial condition is the perturbed $\varphi =0$ mode. Initially, it experiences large oscillations in terms of its azimuthal modulation, while preserving its vortex charges and the power sums $I_{n}$. At $z\approx 400$, an unstable value of $\varphi $ is reached and an instability emerges, leading to irregular dynamics in which neither the topological charges nor $I_{n}$ are conserved. If, instead, the $\varphi =\pi /4$ mode is used as the initial condition, this instability sets in immediately. ![The adiabatic dynamics of the degenerate mixed- vortex family, with $P=1,\protect\phi =\protect\pi /4$, $\protect\varphi =0$, topological charges (2,1) and random perturbations added to the initial conditions at the 5% level. (a) On-site powers of both components. (b) The total power at each site, $I_{n}=|A_{n}|^{2}+|B_{n}|^{2}$. (c) The topological charges.[]{data-label="fig:dynamics"}](figure5){width="60mm"} An additional feature demonstrated by the discrete system is the broken conservation of the OAM. Namely, the two components can exchange their angular momentum with the medium, as well as with each other. We show an example of this in figure \[fig:flip2\], in which one component exchanges the angular momentum with the medium, leading to the periodic reversal of its topological charge, while the charge of the other component remains conserved. This type of dynamics is generated by solutions (\[U\])-([gam]{}) with $\phi \neq \pi /4$ and $\varphi $ chosen so as to make the value of $\gamma $ in (\[gam\]) imaginary. The latter acts as a strong azimuthal perturbation, leading to the charge flipping of the second component. ![The charge- flipping dynamics of the single component, with $P=1/2,\protect\phi =\protect\pi /5$, $\protect\varphi =\protect\pi /10$, topological charges (2,-1), and random perturbations added to the initial conditions at the 2% level. (a,c): On-site powers for the two components. (b,d): Vortex lines (red – positive charge, blue – negative charge).[]{data-label="fig:flip2"}](figure6){width="60mm"} Conclusions {#sec:final} =========== We have studied the properties of composite vortex modes in circular arrays of nonlinear waveguides. The stability hierarchy of discrete composite vortices can be summarized by stating that the HV (hidden-vorticity) modes suffer instabilities above a critical power, while explicit vortex modes with high topological charges are stable. This hierarchy is opposite to that in continua. Additionally, mixed-vortex modes with different topological charges in the two components exist and can be stable. Degeneracies occur in these families of composite vortex modes, which results in long-lived breather states and persistent vortex-charge flipping. It should be stressed that the analysis was performed here for the relatively small ring chains, built of $N=6$ sites. For much larger rings, one may expect a change in the stability and dynamics, as, for a fixed diameter of the ring at $N\rightarrow \infty $, the system must go over into the continuum limit, with its reverse picture of the stability domains for the HV and explicit-vorticity modes. These effects are visible at high powers required for the self-localization in photonic lattices, therefore there is a possibility of observing them in experiments similar to those that were aimed at studying discrete vortices [@neshev2004; @fleischer2004] and multivortex solitons [@terhalle2008] in photorefractive crystals. A hexagonal lattice geometry corresponds to the ring built of $6$ sites in our model. We can obtain appropriate experimental parameters for the observation of composite vortices from Ref. [@terhalle2009], which studied double-charge discrete vortex solitons, corresponding to the scalar ($\phi = 0$) limit of our nonlinear modes. They used a 20mm long crystal with a bias voltage of 2.2kV/cm, a lattice wave beam with a total power of $75 \mu W$ and a $30\mu m$ period. Linear propagation was observed with a probe beam at $532nm$ with a total power of $P \approx 20nW$, while the nonlinear regime was reached at $P \approx 550 nW$. With these parameters, the propagation distance is long enough to observe the absence of discrete diffraction at high power (soliton formation) and the modulational instability of unstable modes. To observe composite vortex solitons, all that is required is to split the probe beam into two incoherent components, with the intensity and phase profiles generated using a spatial light modulator. Alternatively, our model can be realized directly in an integrated-optics setting, using a femtosecond-laser written ring of nonlinear waveguides and two incoherent beams [@femto], although because of the weaker nonlinearity, significantly higher beam powers would be required. Acknowledgment {#acknowledgment .unnumbered} ============== This work is supported by the Australian Research Council. References {#references .unnumbered} ========== [99]{} Desyatnikov A S, Kivshar Yu S and Torner L 2005 *Progress in Optics* **47** 291 Nye J F and Berry M V 1974 *Proc. R. Soc. London A* **336** 165 Soskin M S and Vasnetsov M V 2001 *Progress in Optics* **42** 219 Malomed B A, Peng G D, Chu P L, Towers I, Buryak A V and Sammut R A 2001 *Pramana* **57** 1061 Malomed B A, Mihalache D, Wise F and Torner L 2005 *J. Optics B: Quant. Semicl. Opt*. **7** R53 Kartashov Y V, Malomed B A and Torner L 2011 *Rev. Mod. Phys*. **83** 247 Dennis M R, O’Holleran K and Padgett M J 2009 *Progress in Optics* **52** 293 Gibson G, Courtial J, Padgett M, Vasnetsov M, Pasko V, Barnett S and Franke-Arnold S 2004 *Opt. Express* **12** 5448 Wang J, Yang J-Y, Fazal I M, Ahmed N, Yan Y, Huang H, Ren Y, Yue Y, Dolinar S, Tur M, Willner A E 2012 *Nat. Photonics* **6** 488 Firth W J and Skryabin D V 1997 *Phys. Rev. Lett.* **79** 2450 Torner L and Petrov D V 1997 *Electron. Lett*. **33** 608 Petrov D V, Torner L, Martorell J, Vilaseca R, Torres J P and Cojocaru C 1998 *Opt. Lett*. **23** 1444 Quiroga-Teixeiro M and Michinel H 1997 *J. Opt. Soc. Am. B* **14** 2004 Malomed B A, Crasovan L-C and Mihalache D 2002 *Physica D* **161** 187 Mihalache D, Mazilu D, Crasovan L-C, Towers I, Malomed B A, Buryak A V, Torner L and Lederer F 2002 *Phys. Rev. E* **66** 016613 Leblond H, Malomed B A and Mihalache D 2005 Phys. Rev. E **71** 036608 Briedis D, Petersen D, Edmundson D, Królikowski W and Bang O 2005 *Opt. Express* **13** 435443 Yakimenko A I, Zaliznyak Yu A and Kivshar Yu S 2005 *Phys. Rev. E* **71** 065603 Mihalache D, Mazilu D, Towers I, Malomed B A and Lederer F 2002 *J. Opt. A: Pure Appl. Opt*. **4**, 615 (2002) Borovkova O V, Kartashov Y V, Torner L and Malomed B A 2011 *Phys. Rev. E ***84** 035602(R) Kartashov Y V Vysloukh V A and Torner L 2009 *Progr. Opt*. **52** 63 Christodoulides D N, Lederer F and Silberberg Y 2003 *Nature* **424** 817 Lederer F, Stegeman G I, Christodoulides D N, Assanto G, Segev M and Silberberg Y 2008 *Phys. Reports* **463** 1 Malomed B A and Kevrekidis P G 2001 *Phys. Rev. E* **64** 026601 Pelinovsky D, Kevrekidis P and Frantzeskakis D 2005 *Physica D* **212** 20 Neshev D N, Alexander T J, Ostrovskaya E A, Kivshar Yu S, Martin H, Makasyuk I and Chen Z 2004 *Phys. Rev. Lett.* **92** 123903 Fleischer J W, Bartal G, Cohen O, Manela O, Segev M, Hudock J and Christodoulides D N 2004 *Phys. Rev. Lett* **92** 123904 Sakaguchi H and Malomed B A 2005 *Europhys. Lett*. **72** 698 Kevrekidis P G, Malomed B A, Chen Z and Frantzeskakis D J 2004 Phys. Rev. E **70** 056612 Law K J H, Song D, Kevrekidis P G, Xu J and Chen Z 2009 *Phys Rev A* **80** 063817 Terhalle B, Richter T, Law K J H, Göries D, Rose P, Alexander T J, Kevrekidis P G, Desyatnikov A S, Królikowski W, Kaiser F, Denz C and Kivshar Yu S 2009 *Phys. Rev A* **79** 043821 Desyatnikov A S, Dennis, M R and Ferrando A 2011 *Phys. Rev. A* **83** 063822 Desyatnikov A S, Neshev D, Ostrovskaya E A, Kivshar Yu S, Królikowski W, Luther-Davies B, Garcia-Ripoll J J and Perez-Garcia V M 2001 *Opt. Lett.* **26** 435 Desyatnikov A S, Neshev D, Ostrovskaya E A, Kivshar Yu S, McCarthy G, Królikowski W and Luther-Davies B 2002 *J. Opt. Soc. Am. B* **19** 586. Desyatnikov A S and Kivshar Yu S 2001 *Phys. Rev. Lett.* **87** 033901 Bigelow M S, Park Q-H and Boyd R W 2002 *Phys. Rev. E* **66** 046631 Desyatnikov A S, Kivshar Yu S, Motzek K, Kaiser F, Weilnau C and Denz C 2002 *Opt. Lett.* **27** 634 Motzek K, Kaiser F, Weilnau C, Denz C, McCarthy G, Królikowski W, Desyatnikov A and Kivshar Yu S 2002 *Opt. Commun.* **209** 501 Lashkin V M, Ostrovskaya E A, Desyatnikov A S and Kivshar Yu S 2009 *Phys. Rev. A* **80** 013615 Brtka M, Gammal A and Malomed B A 2010 *Phys. Rev. A* **82** 053610 Musslimani Z H, Soljačić M, Segev M and Christodoulides D N 2001 *Phys. Rev. E* **63** 066608 Musslimani Z H, Segev M, Christodoulides D N and Soljačić M 2000 *Phys. Rev. Lett*. **84** 1164 Desyatnikov A S, Mihalache D, Mazilu D, Malomed B A, Denz C and Lederer F 2005 *Phys. Rev. E* **71** 026615 Desyatnikov A S, Mihalache D, Mazilu D, Malomed B A and Lederer F 2007 *Phys. Lett. A* **364** 231 Musslimani Z H, Segev M and Christodoulides D N 2000 *Opt. Lett*. **25** 61 Desyatnikov A S, Pelinovsky D E and Yang J 2008 *J. Math. Sci.* **151** 3091 Izdebskaya Y V, Rebling J, Desyatnikov A S and Kivshar Y S 2012 *Opt. Lett.* **37** 767 Sakaguchi H and Malomed B A 2012 *J. Opt. Soc. Am. B* **29** 2741 Dong R, Rüter C E, Kip D, Cuevas J, Kevrekidis P G, Song D and Xu J 2011 *Phys. Rev. A* **83** 063816 Kevrekidis P G, Malomed B A, Frantzeskakis D J and Bishop A R 2003 *Phys. Rev. E* **67** 036614 Cuevas J, Hoq Q E, Susanto H and Kevrekidis P G 2009 *Physica D* **238** 2216 Alvarez A, Cuevas J, Romero F R and Kevrekidis P G 2011 *Physica D* **240** 767 Manakov S V 1973 Zh. Eksp. Teor. Fiz. **65** 505 Ferrando A, Zacarés M, García-March, M-A, Monsoriu J A and de Córdoba P F 2005 *Phys. Rev. Lett.* **95** 123901 Alexander T J, Sukhorukov A A and Kivshar Yu S 2004 *Phys. Rev. Lett.* **93** 063901 Terhalle B, Richter T, Desyatnikov A S, Neshev D N, Królikowski W, Kaiser F, Denz C and Kivshar Yu S 2008 *Phys. Rev. Lett* **101** 013903 Heinrich M, Keil R, Dreisow F, Tünnermann A, Szameit A, Nolte S 2011 *App. Phys. B* **104** 469
{ "pile_set_name": "ArXiv" }
--- abstract: 'For a massive multiple-input-multiple-output (MIMO) system using intelligent reflecting surface (IRS) equipped with radio frequency (RF) chains, the multi-channel RF chains are expensive compared to passive IRS, especially, when the high-resolution and high-speed analog to digital converters (ADC) are used in each RF channel. In this letter, a direction of angle (DOA) estimation problem is investigated with low-cost ADC in IRS, and we propose a deep neural network (DNN) as a recovery method for the low-resolution sampled signal. Different from the existing denoising convolutional neural network (DnCNN) for Gaussian noise, the proposed DNN with fully connected (FC) layers estimates the quantization noise caused by the ADC. Then, the denoised signal is subjected to the DOA estimation, and the recovery performance for the quantized signal is evaluated by DOA estimation. Simulation results show that under the same training conditions, the better reconstruction performance is achieved by the proposed network than state-of-the-art methods. The performance of the DOA estimation using 1-bit ADC is improved to exceed that using 2-bit ADC.' author: - 'Weifeng Han, Peng Chen, , Zhenxin Cao, [^1]' bibliography: - 'ref.bib' title: 'Deep Neural Network-Based Quantized Signal Reconstruction for DOA Estimation' --- reconstruction signal, deep neural network, quantization noise, DOA estimation, intelligent reflecting surface Introduction ============ a massive multiple-input-multiple-output (MIMO) system [@risi2014massive] to the millimeter wave (MMV) band will bring higher hardware costs and attenuate the signal significantly, so intelligent reflecting surface (IRS) [@8937491; @zheng2020intelligent] as a new method is introduced to enhance the wireless link and reduce costs of additional base stations. For the massive MIMO system, when the low-cost analog to digital converters (ADCs) are used in the IRS equipped with radio frequency (RF) chains, the ADC resolution is decreasing, and the signals will be deteriorated severely by the quantization noise. It is expected to ensure the quality of signal processing while decreasing costs. Direction of angle (DOA) estimation with 1-bit quantized measurements has been studied for many years. Ref. [@1039405] considers the influence of 1-bit measurements on the estimation accuracy in the scenario with additional white Gaussian noise (AWGN), and the estimation for the covariance matrix is also investigated. Similar to the DOA estimation, a line spectrum estimation with quantized signals is investigated in ref. [@824661], where design rules for antialiasing filter and sampling rate are proposed. They are applied to delta-sigma ADC to reduce the quantization noise, which has been widely used in the industry. Additionally, 1-bit quantized sparse arrays proposed in [@7952732] perform better than the unquantized uniform linear arrays (ULAs) in the DOA estimation problem. Deep learning has aroused great interest in the past years and has been combined with wireless communication and radar signal processing. In [@8400482], a channel and DOA estimation method based on deep neural network (DNN) is proposed in the massive MIMO system. Usually, for the DOA estimation using DNN, the output of DNN is the spatial angle, and this scheme can achieve high resolution. Moreover, there are some other kinds of DNN structures for the DOA estimation, such as the classifier-based network [@8485631], which divides the spatial angle into grids. The input of the network can also use a covariance matrix instead of the raw baseband signals. Denoising convolutional neural network (DnCNN) is first introduced in [@7839189] to remove Gaussian noise in image, and ref. [@8822737] applies DnCNN in the array signal processing with a better performance than the atomic norm minimization (ANM) denoising method. However, for the quantized signals, in addition to the reconstructed covariance matrix in [@1039405], there are few ways to deal with the reconstruction of quantized signals. In the hardware, the sigma-delta ADC can decrease the quantization noise [@824661], but it is expensive when the sampling rate increases, which is not applicable in the massive MIMO system and IRS who possess large-scale array. In this paper, we propose a reconstruction method for quantized signal based on DNN, which is composed of fully connected (FC) layers, batch normalization (BN) unit, residual block, and rectified linear unit (ReLU) activation function. Through offline training, ADC precision can be improved without additional performance loss. Then, the recovery performance for the quantized signal is evaluated by measuring the accuracy of the DOA estimation. Moreover, the reconstruction performance of the proposed network is also compared with state-of-the-art methods. The paper is structured as follows. The signal and quantization noise model are presented in section II. In Section III, we propose the scheme for signal reconstruction. Simulation results and discussion are illustrated in Section IV. Finally, in Section V, the paper is concluded. Signal and Quantization Noise Model =================================== Considering a ULA with $M$ sensors, $d$ is array spacing, and $\lambda$ is wavelength of signals. This ULA array is used to estimate the DOA of the $K$ far-field signals. For the $k$-th signal $\left(k=1,2,3,…,K\right)$, the complex signal and the DOA are denoted as $s_{k}$ and $\theta_{k}$, respectively. During the $n$-th sampling time, when the low-cost ADC is used, the received signal in ULA can be expressed as a baseband signal $$\begin{aligned} \mathbf{y}(n)=\sum_{k=1}^{K}s_{k}(n)\mathbf{a}(\theta_{k})+\mathbf{e}(n)+\mathbf{q}(n)=\mathbf{x}(n)+\mathbf{q}(n), \end{aligned}$$ where the received signal from $M$ sensors is $\mathbf{x}(n)=\sum_{k=1}^{K}s_{k}(n)\mathbf{a}(\theta_{k})+\mathbf{e}(n)\in{\mathbb{C}^{M\times1}}$, and the steering vector is $\mathbf{a}(\theta_{k})=\left[1,e^{j2\pi\frac{d}{\lambda}\sin(\theta_{k})},...,e^{j2\pi\frac{d(M-1)}{\lambda}\sin(\theta_{k})}\right]^{\text{T}}$. The quantized signal is $\mathbf{y}(n)\in{\mathbb{C}^{M\times1}}$, the complex additive Gaussian noise vector is $\mathbf{e}(n)=\left[e_{1}(n),...,e_{M}(n)\right]^{\text{T}}$, and $\mathbf{q}(n)=\left[q_{1}(n),...,q_{M}(n)\right]^{\text{T}}$ denotes the quantization noise. The $m$-th entry in $\mathbf{q}(n)$ is $$\begin{aligned} q_{m}(n)=\text{round}&\left(\frac{x_{m}(n)}{\Delta(B)}\right)\Delta(B)-x_{m}(n), \\&m=1,...,M, \end{aligned}$$ where round($\cdot$) is round operation, $\Delta(B)=2V/2^{B}$, and $V$ is maximum input voltage of ADC. The real part and imaginary part of $x_{m}(n)$ are distributed between $–V$ and $V$. $q_{m}$ is quantization noise produced by $B$-bit ADC, and follows an uniform distribution from $-0.5\Delta(B)$ to $0.5\Delta(B)$. We use rounding quantization, which resulting in that we have $2^{B}+1$ quantization level for $B$ bits quantization. All the quantizations used below are the same as those used in this section. In this paper, we will estimate the DOA $\theta_{k}$ from the quantized signals $\mathbf{y}(n)$.             Proposed Scheme For Signal Reconstruction {#sec:guidelines} ========================================= To recover the information from the quantized signals, we propose a DNN network with FC layers. With the quantized signal as input, the network is trained to output the original signal (unquantized signal, $\mathbf{x}(n)$) . As shown in Fig. 1, the DNN has $L$ hidden layers, we employ the ReLU [@nair2010rectified] as our activation function, which can be given by $$f_{\text{ReLU}}(x)=\max(0,x).$$ We also employ the residual block [@he2016deep] to prevent overfitting, gradient explosion and disappearance. The real and imaginary parts of $\mathbf{y}(n)$ are separated and reshaped into a long vector. For $M$ elements in the ULA, each signal being input to DNN is a tensor $\mathbf{Y}^{[0]}\in{\mathbb{R}^{1\times{2M}}}$. The first layer of this network is FC layer, and ReLU is applied to the output of first FC layer, so the output of first hidden layer can be expressed as $$\mathbf{Y}^{[1]}=f_{\text{ReLU}}\left(\mathbf{Y}^{[0]}\mathbf{W}^{[1]}\right),$$ where $\mathbf{W}^{[1]}\in{\mathbb{R}^{2M\times{N_{1}}}}$ is the weight of the first FC layer, $\mathbf{b}^{[1]}\in{\mathbb{R}^{N_{1}}}$ is its bias term, and $N_{l}$ is the amount of neuron in hidden layer $l$. Residual block is applied to layers after the hidden layer 1. For the first residual block, the following calculation is performed $$\begin{aligned} &\mathbf{Y}^{[2]}=f_{\text{ReLU}}\left(f_{\text{BN}}\left(\mathbf{Y}^{[1]}\mathbf{W}^{[2]}+\mathbf{b}^{[2]}\right)\right),\\ &\mathbf{Y}^{[3]}=f_{\text{ReLU}}\left(\mathbf{Y}^{[1]}+f_{\text{BN}}\left(\mathbf{Y}^{[2]}\mathbf{W}^{[3]}+\mathbf{b}^{[3]}\right)\right), \end{aligned}$$ where $\mathbf{W}^{[l]}\in{\mathbb{R}^{N_{l-1}\times{N_{l}}}}, \mathbf{b}^{[l]}\in{\mathbb{R}^{N_{l}}}$ for $l=2,3,...,L$. In residual block, we applied BN [@ioffe2015batch] to the output of each FC layer. BN uses the mean and standard deviation on small batches to continuously adjust the intermediate output of the neural network during training, thereby making the intermediate output of the entire neural network more stable. BN can be expressed as $$\begin{aligned} f_{\text{BN}}\left(\mathbf{Y}^{[l]}\right)=\mathbf{\gamma}&\odot\frac{\mathbf{Y}^{[l]}-E\left(\mathbf{Y}^{[l]}\right)}{\sqrt{var\left(\mathbf{Y}^{[l]}\right)+\epsilon}}+\mathbf{\beta}, \\&l=2,...,L-1, \end{aligned}$$ where $\epsilon$ is a small constant to ensure denominator is greater than 0. $\mathbb{E(\cdot)}, var(\cdot)$ calculate expectation and variance. $\mathbf{\gamma}, \mathbf{\beta}$ are scaling and shift parameters, respectively. $\odot$ denotes the element-wise product. To obtain the reconstruction of quantized signals, the activation function is not used in the last layer, and only linear operations are performed $$\mathbf{Y}^{[L]}=\mathbf{Y}^{[L-1]}\mathbf{W}^{[L]}+\mathbf{b}^{[L]}.$$ The loss function used for back propagation, reconstruction performance evaluation is as follows $$f_{\text{loss}}\left(\mathbf{Y}^{[L]}\right)=\frac{1}{2M}\parallel \mathbf{Y}^{[L]}-\mathbf{\psi}(n)\parallel^{2}_{2},$$ where $\mathbf{\psi}(n)=\left[\mathcal{R}\{\mathbf{y}(n)-\mathbf{q}(n)\}^{\text{T}},\mathcal{I}\{\mathbf{y}(n)-\mathbf{q}(n)\}^{\text{T}}\right]$, $\mathcal{R}\left\{\cdot\right\},\mathcal{I}\left\{\cdot\right\}$ take the real and imaginary parts of the complex vector, respectively. Under normal circumstances, multiple signals will be concatenated and input into the network, the loss value is the average for these signals.             Simulation Results And Discussion ================================= A 32 elements ULA with half wavelength spacing between sensors is used to reconstruct signals containing 3 different spatial angles, where the spatial scope is $\left[\ang{-30},\ang{30}\right]$. Pytorch framework and Adam optimizer are used to implement proposed scheme. Learning rate is set to $10^{-3}$ without decay. In the meantime, we set $L$ to 10, $N_{l}$ $(l = 1,2,3,…,L-1)$ to 1024, $N_{L}$ to 64 and size of mini-batch to 256, and the data set is shuffled randomly. The numerical results in this section are generated by inputting 1-bit quantized signal into the proposed DNN. The experimental platform is a PC with graphics card NVIDIA 2080Ti. Acquiring codes at <https://github.com/hwfhwf/DNN-For-Signal-Reconstruction>. First, by fine-tuning the proposed DNN to generate different networks, thus explaining why this network is used. Subsequently, comparing the DnCNN in [@8822737] and proposed DNN, multiple signal classification [@1143830] (MUSIC) algorithm is applied to evaluate neural network performance. The MUSIC algorithm uses 5 snapshots to compute the covariance matrix, and 1000 trials are performed. Finally, a method of compressing the model are examined, it can be deployed on mobile devices to reduce costs through numerical results. Assessment For Fine-Tuned Networks ----------------------------------- -- -- -- -- : Training time required for networks with different numbers of neurons (300 iterations)[]{data-label="table"} The performance of different networks based on proposed DNN are tested in 300 iterations, with 20000 training signals and 1000 test signals for SNR = 50 dB. As Fig. 2(a) shown, it is clear that adding layers will not bring large benefits by changing the number of layers of proposed DNN, and will lead to gradient explosion. As depicted in Fig. 2(b), increasing the number of neurons $N_{l}$ brings obvious benefits on test set, but DNN with 2048 neurons will cost much more time in training, as shown in Table . In Fig. 2(c), proposed DNN is adjusted on activation function, with/without residual block, with/without BN, respectively. It is worth noting that even in the case where the network is not that deep, the residual block can still suppress the gradient explosion. Although the curve gap is not obvious during training, on the test set, the loss of proposed method is significantly smaller than networks that has been changed, implying that this network performs better. Comparison Between Proposed DNN With DnCNN ------------------------------------------ Fig. 3(a) compares the proposed DNN with DnCNN in [@8822737]. For DnCNN, the number of filters is adjusted to 64 and 256 in each layer, which can be seen that over-fitting occurs in DnCNN. Increasing the training set size to $10^5$, test set size to 5000 with different $\text{SNR}\in{\{10\text{ dB},20\text{ dB},30\text{ dB},40\text{ dB},50\text{ dB}\}}$, the data set is evenly distributed in each SNR. Fig. 3(b) and Fig. 3(c) contain test loss of two network in different SNR. As the size of the training set increases, over-fitting no longer occurs for DnCNN, consequently more data is required. From two figures, the proposed DNN possesses better reconstruction capability, due to lower test loss in the whole SNR range. As shown in Fig. 3(d), the reconstructed signals output by the network corresponding to Fig.3(b) and Fig. 3(c) are used to estimate the DOA via MUSIC algorithm. As comparison, the DOA of 1$\thicksim$4 bits quantized signals also been estimated to evaluate signal reconstruction level. For estimated angle $\hat{\theta}_{k}$, MSE of angle can be expressed as: $\text{MSE}=\frac{1}{K}\sum_{k=1}^{K}\left(\hat{\theta}_{k}-\theta_{k}\right)^{2}$. With same quantized signal, reconstruction effect of DnCNN is very weak. Through proposed DNN, the performance of the DOA estimation using 1-bit ADC is improved to exceed that using 2-bit ADC. Further Training For Proposed DNN ---------------------------------   Increasing the size of training set to $10^7$ for each SNR to make the network more generalized. After 300 iterations, 1-bit quantized signal are input into the trained network to generate reconstructed signal, and the MUSIC algorithm is used to verify the recovery effect. By comparing with the MSE of 1$\thicksim$4-bit quantized signal, estimate the improvement effect obtained by this method. As is shown in Fig. 4(a), owing to proposed DNN, the performance of the DOA estimation using 1-bit ADC is improved to exceed that using 2-bit ADC on each SNR. In Fig. 4(b), the true angles of the signal are $\ang{-18.9346},\ang{8.6346},\ang{9.9462}$. In the scenario of SNR = 50 dB, the signal reconstructed by proposed network can achieve better resolution than 3-bit quantized signal around $\ang{9}$. Compressed Model ---------------- SNR (dB) ----------- ------- ------- ------- ------- ------- ------- -- -- -- -- Precision 16 FP 32 FP 16 FP 32 FP 16 FP 32 FP MSE 0.081 0.081 0.080 0.078 0.081 0.084 : MSE comparison between the simplified and the original model[]{data-label="table"} For field programmable gate array (FPGA) or embedded devices, the network model must be compressed as much as possible to ensure high processing speed, at the same time the accuracy of signal processing must be guaranteed. There are some methods to compress the model such as precision reduction and pruning for parameters in neural network. Our network parameters are initially single-precision floating-point (32 FP) numbers, turning them to half-precision floating-point (16 FP), then we can get 50% reduction in model size and faster data processing speed. From Table , precision from 32 FP to 16 FP hardly changes the MSE, it is feasible that continue to decrease the precision of parameters to achieve speed-performance trade-off. The operation we used above are for the trained network parameters. Greater benefits can be achieved by applying them to the training process, which is deeply discussed in [@han2015deep], including network quantization, pruning, weight sharing et al. Conclusion ========== This paper has proposed a DNN to decrease quantization noise and reconstruct the quantized signal. The numerical results have shown that under sufficient training, through proposed DNN, the performance of the DOA estimation using 1-bit ADC is improved to exceed that using 2-bit ADC, consequently ADCs with lower resolution can be employed without performance loss. This method can be applied to massive MIMO system, IRS with RF chains, thus greatly reducing costs. [^1]: The authors are with the State Key Laboratory of Millimeter Waves, Southeast University, email: {hanweifeng, chenpengseu, caozx}@seu.edu.cn.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider irreducible actions of locally compact product groups, and of higher rank semi-simple Lie groups. Using the intermediate factor theorems of Bader-Shalom and Nevo-Zimmer, we show that the action stabilizers, and all irreducible invariant random subgroups, are co-amenable in some normal subgroup. As a consequence, we derive rigidity results on irreducible actions that generalize and strengthen the results of Bader-Shalom and Stuck-Zimmer.' address: - 'Weizmann Institute of Science, Faculty of Mathematics and Computer Science, POB 26, 76100, Rehovot, Israel.' - 'Microsoft Research New England, One Memorial Drive, Cambridge MA, 02142, USA.' author: - Yair Hartman - Omer Tamuz bibliography: - 'irs\_rigidity.bib' title: Stabilizer Rigidity in Irreducible Group Actions --- [^1] Introduction ============ Let $G$ be a locally compact second countable (lcsc) group. An [ *invariant random subgroup*]{} (IRS) of $G$ is a random variable that takes values in ${{\mathrm{Sub}_{G}}}$, the space of closed subgroups of $G$, and whose distribution is invariant to conjugation by any element of $G$ [@abert2012kesten]. IRSs arise naturally as stabilizers of probability measure preserving (pmp) actions. Moreover, any IRS is the stabilizer of some pmp action (see [@abert2012growth Theorem 2.4] and also [@abert2012kesten; @creutz2013stabilizers-of-ergodic]). A subgroup $H \leq G$ is said to be [*co-amenable*]{} in $G$ if there exists a $G$-invariant mean on $G/H$ [@monod2003co]. A normal subgroup $N \lhd G$ is co-amenable in $G$ if and only if $G/N$ is amenable. We say that an IRS $K$ is co-amenable in $G$ if it is almost surely co-amenable in $G$. Likewise, if $K$ almost surely has some property (e.g., trivial, normal, co-finite), we say succinctly that $K$ has this property. Let $G = G_1\times G_2$ be a product of two lcsc groups. A pmp action $G \curvearrowright (X,{m})$ is [*irreducible*]{} (with respect to the decomposition $G=G_1 \times G_2)$ if the actions of both $G_1$ and $G_2$ are ergodic. Likewise, a pmp action of a semi-simple Lie group is said to be [*irreducible*]{} if the action of each simple factor is ergodic. An IRS $K$ is [*irreducible* ]{} if $G \curvearrowright ({{\mathrm{Sub}_{G}}}, \lambda)$ is irreducible, where $\lambda$ is the distribution of $K$ and $G$ acts on ${{\mathrm{Sub}_{G}}}$ by conjugation [@abert2012growth]. Note that if an action is ergodic or irreducible then so is the associated stabilizer IRS, while the opposite direction is not true in general. Our main results concern irreducible IRSs of product groups and of semi-simple Lie groups. These are generalizations of the theorems of Bader-Shalom [@bader2006factor Theorem 1.6] and of Stuck-Zimmer [@stuck1994stabilizers]; we do not require $G$ to have property (T). Nevertheless, in both cases, we rely on the corresponding Intermediate Factor Theorems: Bader-Shalom [@bader2006factor] for products and Nevo-Zimmer [@nevo2002generalization] for semi-simple Lie groups. \[thm:product-irs-rigidity\] Let $G=G_1 \times G_2$ be an lcsc group, and let $K$ be an irreducible IRS in $G$. Then there exist closed normal subgroups $N_1 \lhd G_1$ and $N_2 \lhd G_2$ such that $K$ is co-amenable in $N_1 \times N_2$. \[thm:lie-irs-rigidity\] Let $G$ be a connected semi-simple Lie group with finite center, no compact factors and ${\mathbb{R}}$-rank $\ge 2$. Let $K$ be an irreducible IRS in $G$. Then $K$ is either equal to a closed normal subgroup, or else $K$ is co-amenable in $G$. In particular, these theorems are rigidity results on the irreducible IRSs of groups such as $SL_2({\mathbb{R}}) \times SL_2({\mathbb{R}})$, to which the theorems of Bader-Shalom and Stuck-Zimmer do not apply. The fact that every pmp action gives rise to an IRS and vice versa allows us to derive a number of corollaries regarding irreducible actions. We assume throughout that $G$ acts measurably on standard measurable spaces. Given a pmp action $G \curvearrowright (X,{m})$ and a normal subgroup $N \lhd G$, denote by ${{N} \backslash\kern-.3em\backslash {(X,{m})}}$ the space of $N$-ergodic components of $(X,{m})$. \[thm:rigidity\] Let $G = G_1 \times G_2$ be an lcsc group, and let $G \curvearrowright (X,{m})$ be an irreducible pmp action. Then there exists a closed normal subgroup $N = N_1 \times N_2$, with $N_1 \triangleleft G_1$ and $N_2 \triangleleft G_2$, such that $G / N$ acts essentially freely on ${{N} \backslash\kern-.3em\backslash {(X,{m})}}$, and such that there exists an $N$-invariant mean on ${m}$-almost every $N$-orbit $Nx$. If furthermore $N$ has property (T), then the $N$-action on each $N$-ergodic component is essentially transitive. The conclusions of Theorem \[thm:product-irs-rigidity\] and Corollary \[thm:rigidity\] can be strengthened when more constraints are imposed on $G$. In particular, we consider the following notions: An lcsc group is said to be [*just non-compact*]{} if every closed normal subgroup is co-compact. An lcsc group is said to be [*just non-amenable*]{} if every closed normal subgroup is co-amenable. Note that if a group is just non-compact or simple then it is also just non-amenable. \[thm:just-non-amenable\] Let $G=G_1 \times G_2$ be an lcsc group, let both $G_1$ and $G_2$ be just non-amenable. Then every irreducible IRS is either co-amenable in $G$ or equal to a normal subgroup. In the following corollaries we show that when one of the factors has property (T), then all IRSs are either co-finite or equal to a normal subgroup. In fact, co-finite IRSs admit some more structure: any ergodic co-finite IRS is supported on a single orbit $\{H^g\}_{g\in G}$, for some co-finite $H \le G$ (see Corollary \[cor:tame\]). \[thm:rigidity-special\] Let $G = G_1 \times G_2$ be an lcsc group, and let $G_1$ be just non-amenable and have property (T). Let $G \curvearrowright (X,{m})$ be a faithful irreducible pmp action. Then the action $G \curvearrowright (X,{m})$ is either essentially free or essentially transitive. It follows that the associated stabilizer IRS is either trivial or co-finite in $G$. This constitutes a strengthening of the Essentially Free Actions Theorem of Bader-Shalom; they require that both $G_1$ and $G_2$ have property (T) and be just non-compact. A corollary of Theorem \[thm:lie-irs-rigidity\] is the following strengthening of the Stuck-Zimmer theorem [@stuck1994stabilizers]. \[thm:stuck-zimmer\] Let $G$ be a connected semi-simple Lie group with finite center, no compact factors and ${\mathbb{R}}$-rank $\ge 2$. Assume that one of the simple factors of $G$ has property (T). Then any faithful irreducible pmp $G$-action is either essentially free or essentially transitive, and its associated stabilizer is either trivial or a lattice in $G$. Stuck-Zimmer derive the same conclusions but require every simple factor to have property (T). Recently, Creutz [@creutz2013stabilizers] proved this result independently, using a different approach. In the same paper he also generalizes Bader-Shalom’s theorem to the case that $G_1$ has property (T) and both $G_1$ and $G_2$ are simple. Creutz and Peterson [@creutz2013stabilizers-of-ergodic] prove similar rigidity results for irreducible lattices and commensurators of lattices in semi-simple Lie groups, and also for product groups with the Howe-Moore property and property (T). In [@abert2012growth] it is shown that in the setting of Corollary \[thm:stuck-zimmer\], if $G$ has property (T) then every irreducible IRS is either equal to a normal subgroup or is a lattice. Intermediate factor theorems ---------------------------- Theorems \[thm:product-irs-rigidity\] and \[thm:lie-irs-rigidity\] are consequences of Theorem \[thm:ift-coamenability\] below, which is a result on any IRS that satisfies an [*intermediate factor theorem*]{} (IFT). The original proofs of Stuck-Zimmer and of Bader-Shalom are also each based on a corresponding IFT. In this work (in particular, in Theorem \[thm:ift-coamenability\]) we give a more general result of IRS rigidity, given an IFT. In particular, our proof does not require property (T). This gives a partial answer to a question asked in Stuck-Zimmer [@stuck1994stabilizers page 731]. Let ${\Pi({G,\mu})}$ be the Poisson boundary of a group $G$ with a measure $\mu$, and let $G \curvearrowright (X,{m})$ be a pmp action. A $G$-quasi-invariant probability space $(Y,\eta)$ is a $(G,\mu)$-[ *intermediate factor over $(X,{m})$*]{} if there exist $G$-factors $$\begin{aligned} {\Pi({G,\mu})} \times (X,{m}) \longrightarrow (Y,\eta) \longrightarrow (X,{m})\end{aligned}$$ such that the composition is the natural projection $$\begin{aligned} {\Pi({G,\mu})} \times (X,{m}) \longrightarrow (X,{m}).\end{aligned}$$ An example of an intermediate factor is $(Y,\eta) \cong (C,\xi) \times (X,{m})$, where $(C,\xi)$ is some $(G,\mu)$-boundary, or, equivalently, where $(C,\xi)$ is a $G$-factor of the Poisson boundary. In this case, the intermediate factor $(Y,\eta)$ is said to be [*standard*]{} [@furstenberg2009stationary]. Zimmer [@zimmer1982ergodic] proves an [*intermediate factor theorem*]{}, which was generalized (and had its proof corrected) by Nevo and Zimmer [@nevo2002generalization]: When $G$ is a connected semi-simple Lie group with finite center, no compact factors and higher rank, then there exists an admissible $\mu$ such that any $(G,\mu)$-intermediate factor over an irreducible pmp space is standard. Bader and Shalom [@bader2006factor] prove the same result for intermediate factors over irreducible actions of product groups. The following definition is inspired by these results. Let $G$ be an lcsc group. A pmp action $G\curvearrowright (X,{m})$ is said to be an [*IFT action*]{} if there exists an admissible $\mu$ on $G$ such that any $(G,\mu)$-intermediate factor over $(X,{m})$ is standard. Our main Theorems \[thm:product-irs-rigidity\] and \[thm:lie-irs-rigidity\] are consequences of the following theorem and the above mentioned intermediate factor theorems. \[thm:ift-coamenability\] Let $G$ be an lcsc group. Let $K \leq G$ be an IRS with distribution $\lambda$ such that $G\curvearrowright ({{\mathrm{Sub}_{G}}},\lambda)$ is an IFT action. Then there exists a closed normal subgroup $N \triangleleft G$ such that $K$ is co-amenable in $N$. The remainder of the paper is organized as follows. In Section \[sec:preliminaries\] we construct the [*normal closure of an IRS*]{} and the [*IRS coset space*]{}, and discuss random walks on groups and coset spaces, Poisson boundaries and Bowen spaces. In Section \[sec:proof3\] we prove Theorem \[thm:ift-coamenability\]. Finally, in Section \[sec:proof-of-corollaries\] we prove Theorems \[thm:product-irs-rigidity\] and \[thm:lie-irs-rigidity\] and their corollaries. Acknowledgments --------------- We would like to thank Uri Bader, Amos Nevo and Benjamin Weiss for useful discussions and motivating conversations. We would also like to thank Yehuda Shalom and Lewis Bowen for helpful comments on a first draft of this article. Preliminaries {#sec:preliminaries} ============= The normal closure of an IRS ---------------------------- Let $G$ be an lcsc group. Equip the space of closed subgroups of $G$, ${{\mathrm{Sub}_{G}}}$, with the Chabauty topology [@chabauty1950limite]. Then the $G$ action on ${{\mathrm{Sub}_{G}}}$ by conjugation is continuous. Denote by ${{\mathrm{IRS}(G)}}$ the set of conjugation invariant probability Borel measures on ${{\mathrm{Sub}_{G}}}$. These are the distributions of the IRSs of $G$. For $\lambda \in {{\mathrm{IRS}(G)}}$, let the [*normal closure*]{} ${\overline{{\left\langle {\lambda} \right\rangle}}}$ be the unique minimal subgroup that is the closure of a subgroup generated by a $\lambda$-full measure set of subgroups. ${\overline{{\left\langle {\lambda} \right\rangle}}}$ is well defined and is equal to the closure of the subgroup generated by all the subgroups in $\operatorname{supp}\lambda$. Clearly, ${\overline{{\left\langle {\lambda} \right\rangle}}}$ is a closed [*normal*]{} subgroup of $G$. Let $T$ be the closure of the subgroup that is generated by $\operatorname{supp}\lambda$. We prove the claim by showing that every closed subgroup $S$ that is the closure of a subgroup generated by a full measure set, contains $T$. Let $S$ be the closure of the subgroup generated by a set $A$ of subgroups, with $\lambda(A)=1$, and, without loss of generality, let $A \subseteq \operatorname{supp}\lambda$. Then the closure of $A$ equals $\operatorname{supp}\lambda$, and so for every $g \in T$ there exists a sequence $g_n \to g$ with $g_n \in S$. Since $S$ is closed, it follows that $g \in S$. IRS coset spaces ---------------- Given $\lambda \in {{\mathrm{IRS}(G)}}$, let ${{{{G} / {\lambda}}}}$ be the locally compact space of closed subsets of $G$ given by $$\begin{aligned} {{{{G} / {\lambda}}}}= \left\{g {H}\,:\,{H}\in \operatorname{supp}\lambda, g \in G\right\}, \end{aligned}$$ endowed with the Chabauty topology and equipped with the left and right $G$-actions on subsets of $G$. Note that $$\begin{aligned} {{{\lambda} \backslash {G}}}= \left\{{H}g\,:\,{H}\in \operatorname{supp}\lambda, g \in G\right\}\end{aligned}$$ is equal to ${{{{G} / {\lambda}}}}$ as a $G$-space of subsets of $G$, since $g{H}= g{H}g^{-1}g = {H}^gg$ and $\operatorname{supp}\lambda$ is conjugation invariant. Random walks on groups and Bowen spaces {#sec:bowen-spaces} --------------------------------------- Let $G$ be an lcsc group, and let $\mu$ be an [*admissible*]{} measure on $G$. Namely, let some convolution power of $\mu$ be absolutely continuous with respect to the Haar measure, and let $\operatorname{supp}\mu$ generate $G$ as a semi-group. A $\mu$-random walk on a group is a measure $\mathbb{P}_\mu$ on $G^{\mathbb{N}}$ given by the push-forward of $\mu^{\mathbb{N}}$ under the map $(g_1,g_2,g_3,\ldots) \mapsto (g_1,g_1g_2,g_1g_2g_3,\ldots)$. Equivalently, let $\{X_n\}_{n \in {\mathbb{N}}}$ be i.i.d. random variables with measure $\mu$, and let $Z_n = X_1 \cdots X_n$. Then a $\mu$-random walk is the distribution of $(Z_1,Z_2,\ldots)$. The [*(Furstenberg-)Poisson boundary*]{} [@furstenberg1971random; @furstenberg1974boundary] of a $\mu$-random walk, denoted by ${\Pi({G,\mu})}$, is Mackey’s point realization of the shift-invariant sigma-algebra of $(G^{\mathbb{N}},\mathbb{P}_\mu)$ (see, e.g., [@zimmer1978amenable]), also known as the space of shift-ergodic components. Let ${H}\leq G$ be a closed subgroup. The map $(g_1,g_2,\ldots) \mapsto ({H}g_1,{H}g_2,\ldots)$ pushes forward $\mathbb{P}_\mu$, the $\mu$-random walk on $G$, to $\mathbb{P}_\mu^{H}$, the $\mu$-random walk on ${H}\backslash G$. We denote by ${\Pi({G,\mu,{H}})}=(B_{H},\nu_{H})$ the Poisson boundary of this random walk, which is here equal to Mackey’s point realization of the shift-invariant sigma-algebra of $(({H}\backslash G)^{\mathbb{N}},\mathbb{P}_\mu^{H})$. Equivalently, ${\Pi({G,\mu,{H}})}$ is equal to ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$, the space of $K$-ergodic components of ${\Pi({G,\mu})}$. If ${H}$ is a [*normal*]{} subgroup, then the ${H}$-invariant sigma-algebra is $G$-invariant and hence ${\Pi({G,\mu,{H}})}$ is a $G$-space. In fact, it is a $(G,\mu)$-boundary, which is isomorphic to the Poisson boundary of the group $G/{H}$, equipped with the projection of $\mu$ (see [@bader2006factor Lemma 2.15]). L. Bowen [@bowen2010random] introduces what we shall call [ *Bowen spaces*]{} (see also earlier work by Kaimanovich [@kaimanovich2005amenability]). Let $\lambda \in {{\mathrm{IRS}(G)}}$. Consider the space $$\begin{aligned} \widetilde{ {{\mathrm{Sub}_{G}}}} = \{ \left({H}; {H}g_1, {H}g_2,\ldots \right) | {H}\in {{\mathrm{Sub}_{G}}}, g_n\in G \} \end{aligned}$$ with the topology induced from $(2^G)^{\mathbb{N}}$ and equipped with the measure $$\begin{aligned} d\mathbb{P}_\mu^\lambda \left({H}; {H}g_1, {H}g_2,\ldots \right) = d\mathbb{P}_\mu^{H}({H}g_1,{H}g_2,\ldots) d\lambda({H}). \end{aligned}$$ $G$ acts on $\widetilde{ {{\mathrm{Sub}_{G}}}}$ by $ g \left({H}; {H}g_1, {H}g_2,\ldots \right) = \left({H}^g; {H}^ggg_1, {H}^ggg_2,\ldots \right)$, and the natural shift maps $\left({H}; {H}g_1, {H}g_2,\ldots \right)$ to $\left({H}; {H}g_2,\ldots \right)$. Denote by ${\mathrm{B}_{\mu}({\lambda})}$ Mackey’s point realization of the shift-invariant sigma algebra. We shall refer to ${\mathrm{B}_{\mu}({\lambda})}$ as a [*Bowen space*]{}. Note that the $G$ action on $\widetilde{ {{\mathrm{Sub}_{G}}}}$ commutes with the shift, and so ${\mathrm{B}_{\mu}({\lambda})}$ is a $G$-space. It is useful to realize the Bowen space ${\mathrm{B}_{\mu}({\lambda})}$ as the space of pairs $$\begin{aligned} {\mathrm{B}_{\mu}({\lambda})} = \{({H},\xi)\,:\, {H}\in \operatorname{supp}\lambda, \xi \in {\Pi({G,\mu,{H}})}\},\end{aligned}$$ equipped with the measure $d\nu_\lambda({H},\xi) = d\lambda({H})d\nu_{H}(\xi)$. Therefore, ${\mathrm{B}_{\mu}({\lambda})}$ is the fiber bundle over $\operatorname{supp}\lambda$ in which the fiber over ${H}$ is ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$. Hence the Bowen space ${\mathrm{B}_{\mu}({\lambda})}$ is an intermediate factor over $({{\mathrm{Sub}_{G}}},\lambda)$: $$\begin{aligned} {\Pi({G,\mu})} \times ({{\mathrm{Sub}_{G}}},\lambda) \longrightarrow {\mathrm{B}_{\mu}({\lambda})} \longrightarrow ({{\mathrm{Sub}_{G}}},\lambda).\end{aligned}$$ We encourage the reader to study the details in Bowen’s original paper [@bowen2010random]. For further discussion and another application of Bowen spaces see [@hartman2012furstenberg]. Consider the process on ${{{\lambda} \backslash {G}}}$ that first chooses an element according to $\lambda$, and then applies a $\mu$-right random walk. Formally, this is the Markov chain $\{KZ_n \}_{n\in{\mathbb{N}}}$ where $K\sim \lambda$ and $Z_n \sim \mu^n$. It follows from the definitions that \[clm:bowen-is-poisson\] The Bowen space ${\mathrm{B}_{\mu}({\lambda})}$ is the Poisson boundary of the Markov chain $\{KZ_n\}_{n \in {\mathbb{N}}}$. The Poisson boundary here is the space of shift-ergodic components of the Markov chain measure on $({{{\lambda} \backslash {G}}})^{\mathbb{N}}$. Proof of the Co-amenable IRSs Theorem {#sec:proof3} ===================================== In this section we prove the following. \[thm:co-amenable-ift\] Let $K$ be an IRS in $G$ with distribution $\lambda \in {{\mathrm{IRS}(G)}}$ such that $G\curvearrowright ({{\mathrm{Sub}_{G}}},\lambda)$ is an IFT action. Then $K$ is co-amenable in ${\overline{{\left\langle {\lambda} \right\rangle}}}$. Note that Theorem \[thm:ift-coamenability\] is a direct consequence, since ${\overline{{\left\langle {\lambda} \right\rangle}}}$ is a closed normal subgroup of $G$. Let $\mu$ be an admissible measure on $G$ that is given by the IFT property, so that any $(G,\mu)$-intermediate factor over $({{\mathrm{Sub}_{G}}},\lambda)$ is standard. Fix $\lambda \in {{\mathrm{IRS}(G)}}$, and denote ${T}= {\overline{{\left\langle {\lambda} \right\rangle}}}$. Recall that ${\mathrm{B}_{\mu}({\lambda})}$ is the fiber bundle over $\operatorname{supp}\lambda$ in which the fiber over ${H}$ is ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$. It follows that when the Bowen space ${\mathrm{B}_{\mu}({\lambda})}$ is standard - that is, when it is equal to $(C,\xi) \times ({{\mathrm{Sub}_{G}}},\lambda)$ for some boundary $(C,\xi)$ - then ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}} \cong (C,\xi)$, for $\lambda$-almost every ${H}$. Note that while a priori ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$ is not a $G$-space, we can now conclude that it is. It follows that there exists a $\lambda$-full measure $A \subset {{\mathrm{Sub}_{G}}}$ such that $(C,\xi)$ is ${H}$-invariant for every ${H}\in A$. Without loss of generality we may take a compact model for ${\Pi({G,\mu})}$ such that $G\curvearrowright {\Pi({G,\mu})}$ is continuous. It follows that $(C,\xi)$ is ${T}$-invariant. We have therefore proven the following claim. \[clm:t-invariant\] Let ${\mathrm{B}_{\mu}({\lambda})}=(C,\xi)\times({{\mathrm{Sub}_{G}}},\lambda)$ be standard. Then ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}} ={{{T}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$ for $\lambda$-almost every ${H}$, and $(C,\xi)$ is ${T}$-invariant. Moreover, $(C,\xi)$ is the Poisson boundary of the quotient group ${{{G} / {{T}}}}$, equipped with the projection of $\mu$. The second part of the claim follows from the fact that ${{{T}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$ is the Poisson boundary of the projection of the $\mu$-random walk on $G$ on the quotient group ${{{G} / {{T}}}}$. Given the fact that ${{{H}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}} ={{{T}} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$, it should perhaps be not very surprising that ${H}$ is co-amenable in ${T}$, as Proposition \[thm:co-amenable-ift\] claims. Indeed, the fact that $G\curvearrowright {\Pi({G,\mu})}$ is an amenable action [@zimmer1978amenable] implies the following result, to which Kaimanovich [@kaimanovich2002thepoisson] gives a simple, elementary proof. \[clm:normal-co-amenable\] Let $N_1 \leq N_2$ be two normal subgroups of $G$ such that ${{N_1} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}} ={{N_2} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}$. Then $N_1$ is co-amenable in $N_2$. In this sense, we show that invariant random subgroups behave like normal subgroups, as has been observed in other contexts [@abert2012kesten]. ### Proof of Proposition \[thm:co-amenable-ift\] Let $\mu$ be a measure on $G$ such that $G\curvearrowright ({{\mathrm{Sub}_{G}}},\lambda)$ is an IFT action, so that the Bowen space ${\mathrm{B}_{\mu}({\lambda})}$ is standard. Let $K$ have distribution $\lambda$. To prove that $K$ is co-amenable in ${T}$, we first construct a sequence of asymptotically ${T}$-left invariant measures on ${{{\lambda} \backslash {G}}}$, in Lemma \[lem:t-invariant-on-glam\]. Then, we push this sequence to ${{{\lambda} \backslash {{T}}}}$ in Claim \[lem:t-invariant-on-tlam\]. Finally, in Claim \[clm:co-amenable-in-t\] we use this sequence to show the existence of a ${T}$-left invariant mean on ${{{{{T}} / {{H}}}}}$, for $\lambda$-almost every ${H}$. Denote by $\phi : {{{\lambda} \backslash {G}}}\to {{\mathrm{Sub}_{G}}}$ the map ${H}g =g{H}^{g^{-1}} \mapsto {H}^{g^{-1}}$. As a map ${{{{G} / {\lambda}}}}\to {{\mathrm{Sub}_{G}}}$, (recall that ${{{\lambda} \backslash {G}}}={{{{G} / {\lambda}}}}$), $\phi$ is the natural projection $g{H}\mapsto {H}$. Denote by $|\cdot|$ the total variation norm. \[lem:t-invariant-on-glam\] There exists a sequence of measures $\theta_n$ on ${{{\lambda} \backslash {G}}}$ such that $|t \theta_n - \theta_n| \to 0$ for all $t \in {T}$, and $\phi_*g\theta_n = \lambda$ for all $n\in {\mathbb{N}}$ and $g \in G$. Recall $\{KZ_n\}_{n \in {\mathbb{N}}}$, our Markov chain on ${{{\lambda} \backslash {G}}}$ given by $K \sim \lambda$ and $Z_n \sim \mu^n$. Denote by $\eta_n = \lambda * \mu^n$ the distribution of $KZ_n$. Let $\theta_n$ be given by $$\begin{aligned} \theta_n = \frac{1}{n}\sum_{k=0}^{n-1}\eta_k. \end{aligned}$$ Equip the space ${{{\lambda} \backslash {G}}}$ with a $\sigma$-finite measure $\alpha$ such that $\theta_n \prec \alpha$ for all $n$. Then $g \theta_n \prec \alpha$ for all $g \in G$ and $n$, since $\mu$ is generating. By [@kaimanovich1992measure Lemma 2.9], it follows that $|g \theta_n - \theta_n| \to |\nu_{g\lambda}-\nu_\lambda|$, where $\nu_{g\lambda}$ and $\nu_\lambda$ are the measures on the Poisson boundaries corresponding to the initial distributions $\lambda$ and $g\lambda$. Note that here, as $\lambda$ is a measure on ${{{\lambda} \backslash {G}}}$, the $G$-action is the left multiplication action rather than conjugation. Hence the initial distributions $g\lambda$ and $\lambda$ are in general different. By Claim \[clm:bowen-is-poisson\], $\nu_{\lambda}$ is the measure of the Bowen space ${\mathrm{B}_{\mu}({\lambda})}$. By the definition of the $G$ action on Bowen spaces, $\nu_{g\lambda} = g\nu_\lambda$. Hence we get that $$\begin{aligned} |g \theta_n - \theta_n| \to |g \nu_\lambda-\nu_\lambda|. \end{aligned}$$ Finally, by Claim \[clm:t-invariant\], $\nu_\lambda$ is ${T}$-invariant, and so $|t \theta_n - \theta_n| \to 0$ for all $t \in {T}$. We show that $\phi_* g\theta_n = \lambda$ by showing that $\phi_* g\eta_n = \lambda$. Let $KZ_n \sim \eta_n$, where $K \sim \lambda$ and $Z_n \sim \mu^n$. But $gKZ_n = gZ_nK^{Z_n^{-1}}$, and since $K^{Z_n^{-1}} \sim \lambda$ by the conjugation invariance of $\lambda$, we get that $\phi_*g\eta_n = \lambda$. Since ${T}$ includes every ${H}\in \operatorname{supp}\lambda$, $\lambda$ can be considered to be an element in ${\mathrm{IRS}({T})}$. In analogy to the definition of ${{{\lambda} \backslash {G}}}$, we can define ${{{\lambda} \backslash {{T}}}}$. We define $\phi :{{{\lambda} \backslash {{T}}}}\to {{\mathrm{Sub}_{G}}}$ likewise. Choose a ${T}$-equivariant measurable map $G \to {T}$ (recall that ${T}$ is normal in $G$). It induces a natural map ${{{\lambda} \backslash {G}}}\to {{{\lambda} \backslash {{T}}}}$. We can now push forward the measures $\theta_n$ from Lemma \[lem:t-invariant-on-glam\] above, to get the following. \[lem:t-invariant-on-tlam\] There exists a sequence of measures $\theta_n$ on ${{{\lambda} \backslash {{T}}}}$ such that $|t \theta_n - \theta_n| \to 0$ for all $t \in {T}$, and $\phi_*s\theta_n = \lambda$ for all $n\in{\mathbb{N}}$ and $s \in {T}$. We are now ready to take the last step in the proof of Proposition \[thm:co-amenable-ift\]. \[clm:co-amenable-in-t\] $K$ is co-amenable in ${T}$. Let $\theta_n$ be a sequence of probability measures on ${{{\lambda} \backslash {{T}}}}$ given by Claim \[lem:t-invariant-on-tlam\]. Thinking of $\theta_n$ as a measure on ${{{{{T}} / {\lambda}}}}$, and thinking of $\phi:{{{{{T}} / {\lambda}}}}\to {{\mathrm{Sub}_{G}}}$ as the natural projection $t{H}\mapsto {H}$, we disintegrate $\theta_n$ with respect to $\phi$: $$\begin{aligned} \theta_n = \int_{{{\mathrm{Sub}_{G}}}}\theta_n^{H}d\lambda({H}), \end{aligned}$$ so that $\theta_n^{H}$ is a measure on ${{{{{T}} / {{H}}}}}$. We likewise disintegrate $t\theta_n$ into the fiber measures $(t\theta_n)^{H}$. Note that $\phi(t s{H}) = \phi(s{H})$ for every $t,s \in {T}$. Hence both $\theta_n^{H}$ and $t \theta_n^{H}$ are supported on the same fiber, and we get that $(t\theta_n)^{H}= t\theta_n^{H}$. As both $t\theta_n$ and $\theta_n$ are projected by $\phi$ to $\lambda$, we can disintegrate $t\theta_n -\theta_n$ to get $$\begin{aligned} t\theta_n-\theta_n &= \int_{{{\mathrm{Sub}_{G}}}}\left((t\theta_n)^{H}-\theta_n^{H}\right)d\lambda({H}) \\ &= \int_{{{\mathrm{Sub}_{G}}}} \left(t\theta_n^{H}-\theta_n^{H}\right) d\lambda({H}),\end{aligned}$$ and in particular, $$\begin{aligned} |t\theta_n-\theta_n| = \int_{{{\mathrm{Sub}_{G}}}}\left|t\theta_n^{H}-\theta_n^{H}\right|d\lambda({H}).\end{aligned}$$ But for any $t\in{T}$ we have that $|t\theta_n-\theta_n| \to 0$, and therefore $\left|t\theta_n^{H}-\theta_n^{H}\right| \to 0$ for $\lambda$-almost every ${H}$. Finally, the existence of asymptotically invariant measures on ${{{{{T}} / {{H}}}}}$ implies the existence of a $T$-left invariant mean on ${{{{{T}} / {{H}}}}}$. This, in turn, implies that ${H}$ is co-amenable in $T$ [@greenleaf1969amenable]. This completes the proof of Proposition \[thm:co-amenable-ift\]. Proofs of main theorems and corollaries {#sec:proof-of-corollaries} ======================================= Let $\lambda \in {{\mathrm{IRS}(G)}}$ be the distribution of $K$, and denote ${T}= {\overline{{\left\langle {\lambda} \right\rangle}}}$. By the Bader-Shalom IFT, Proposition \[thm:co-amenable-ift\] implies that $K$ is co-amenable in ${T}$. Let $\mu = \mu_1 \times \mu_2$, where $\mu_i$ is an admissible measure on $G_i$. Then $$\begin{aligned} {\Pi({G,\mu})} = {\Pi({G_1,\mu_1})} \times {\Pi({G_2,\mu_2})}, \end{aligned}$$ and so, if we denote by $N_i \lhd G_i$ the projection of $T$ on $G_i$ we get that $$\begin{aligned} {{T} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}} = {{N_1} \backslash\kern-.3em\backslash {{\Pi({G_1,\mu_1})}}} \times {{N_2} \backslash\kern-.3em\backslash {{\Pi({G_2,\mu_2})}}} = {{(N_1 \times N_2)} \backslash\kern-.3em\backslash {{\Pi({G,\mu})}}}. \end{aligned}$$ By Theorem \[clm:normal-co-amenable\], it follows that $T$ is co-amenable in $N_1 \times N_2$. As $K$ is co-amenable in $T$, we conclude that $K$ is co-amenable in $N_1\times N_2$. By the Nevo-Zimmer IFT, Proposition \[thm:co-amenable-ift\] implies that $K$ is co-amenable in its normal closure $T$. If $T$ is central then $K$ is equal to a normal subgroup, and we are done. Otherwise, we show that $T=G$: Let $S$ be such that $G=TS$ and $T$ and $S$ commute [@ragozin1972normal]. It follows that $S$ acts trivially on $K$, and so, by the irreducibility of $K$, $S$ is trivial and $T=G$. Before proving Corollary \[thm:rigidity\] we make note of a few facts and prove a lemma. Let $G$ be an lcsc group, and let ${H}\le G$. By definition, if ${H}$ is co-amenable in $G$ then there exists a $G$-invariant mean on the homogeneous space $G/{H}$. If $G$ additionally has property (T), the existence of this invariant mean implies that there exists a $G$-invariant probability measure on $G/{H}$. That is, if $G$ has property (T) then the co-amenable subgroups are co-finite. We next show that it follows from Varadarajan’s ergodic decomposition theorem [@varadarajan1963groups] that the existence of invariant measures supported on orbits can be used to show that an action is essentially transitive. Let $G$ be an lcsc group acting on a standard measurable space $X$. Let $E(X)^G$ denote the space of $G$-invariant, ergodic probability measures on $X$. A [*decomposition map*]{} is a measurable map $\beta : X \to E(X)^G$ with the following properties. 1. $\beta$ is $G$-invariant. I.e., denoting $x \to \beta_x$, it holds that $\beta_{gx} = \beta_x$ for all $g \in G$ and $x \in X$. 2. For every $\eta \in E(X)^G$, it holds that $\eta(\beta^{-1}(\eta))=1$. 3. For every $G$-invariant measure $\theta$ it holds that $$\begin{aligned} \label{eq:decompo} \theta = \int_X\beta_xd\theta(x). \end{aligned}$$ \[thm:varadarajan\] For every action of an lcsc group $G$ on a standard measurable space $X$, there exists a decomposition map $\beta$. Furthermore, $\beta$ is essentially unique, in the sense that if $\beta$ and $\beta'$ are decomposition maps then $\theta (\{x \in X\,:\, \beta_x \neq \beta'_x\})=0$ for any $G$-invariant probability measure $\theta$. \[lem:co-finite-stabs\] Let $G$ be an lcsc group, and let $G \curvearrowright (X,{m})$ be a pmp action on a standard measurable space. Assume that there exists a $G$-invariant probability measure on ${m}$-almost every $G$-orbit. Then the action $G \curvearrowright (X,{m})$ is essentially transitive. Let $\beta : X \to E(X)^G$ be a decomposition map of $X$ with respect to the $G$-action. Let $x \in X$ be such that there exists a $G$-invariant and ergodic probability measure $\eta_x$ with $\eta_x(Gx)=1$. By the second property of decomposition maps, there exists an element $y \in Gx$ (in fact a $\eta_x$-full measure set of such elements) for which $\beta_y = \eta_x$. Since $\beta$ is $G$-invariant, we get that $\beta_x=\eta_x$ and in particular $\beta_x$ is supported on $Gx$. Let $A$ be an ${m}$-full measure set of $x \in X$ for which there exists a $G$-invariant, ergodic measure $\eta_x$ on $Gx$, and for which, by the above, $\beta_x=\eta_x$ is supported on a $G$-orbit. Then $$\begin{aligned} {m}= \int_X\beta_xd{m}(x) = \int_A\beta_xd{m}(x), \end{aligned}$$ and it follows by the ergodicity of ${m}$ that it is equal to some $\beta_x$. Hence ${m}$ is supported on a $G$-orbit, or, equivalently, the action is essentially transitive. The following is a corollary of Lemma \[lem:co-finite-stabs\]. \[cor:tame\] Let $G$ be an lcsc group, and let an ergodic IRS $K$ in $G$ be almost surely co-finite. Then its distribution is supported on a single orbit $\{{H}^g\,:\, g\in G\}$. In other words, the action of an lcsc group on its co-finite subgroups is [*tame*]{}. In Stuck-Zimmer this is proven for the case of Lie groups [@stuck1994stabilizers Corollary 3.2]. We are now ready to prove Corollary \[thm:rigidity\]. Let $K$ be the stabilizer IRS associated with the action $G \curvearrowright (X,{m})$. By Theorem \[thm:product-irs-rigidity\] we have that there exists a normal subgroup $N \lhd G$ such that $K$ is co-amenable in $N$. Since $N$ includes the stabilizers of ${m}$-almost every $x \in X$, it follows that $G/N$ acts essentially freely on ${{N} \backslash\kern-.3em\backslash {(X,{m})}}$. Let $(X,{m}_0)$ be an $N$-ergodic component of $m$. Since the stabilizer $N_x$ of $x$ is ${m}_0$-almost surely co-amenable in $N$, there exists an $N$-invariant mean on $N/N_x$, for ${m}_0$-almost every $x$. Since the orbit $Nx$ can be identified with $N/N_x$, it follows that there exists an $N$-invariant mean on the orbit $Nx$. If in addition $N$ has property (T) then, because $K$ is co-amenable in $N$, it is in fact co-finite in $N$. Hence, by again identifying each orbit $Nx$ with $N/N_x$, it follows that there exists an $N$-invariant measure on the orbit $Nx$ for ${m}_0$-almost every $x$. Therefore, by Lemma \[lem:co-finite-stabs\], $N$ acts essentially transitively on $(X,{m}_0)$. For discrete groups, it is straightforward to see that when an action’s stabilizers are co-amenable, then the action is weakly amenable (in the sense of Zimmer [@zimmer1978amenable]), or, equivalently, that the induced orbital equivalence relation is amenable [@connes1981amenable]. By Theorem \[thm:product-irs-rigidity\], there exist subgroups $N_1 \lhd G_1$ and $N_2 \lhd G_2$ such that $K$ is co-amenable in $N = N_1 \times N_2$. Since $K$ is irreducible, if either $N_1$ or $N_2$ is trivial then $K$ must equal $N_1 \times N_2$. Otherwise, because $G_1$ and $G_2$ are just non-amenable, $N_1$ is co-amenable in $G_1$ and $N_2$ is co-amenable in $G_2$. Hence $N$ is co-amenable in $G$, and so $K$ is co-amenable in $G$. Before proving Corollaries \[thm:rigidity-special\] and \[thm:stuck-zimmer\] we prove an analogue of Lemma \[lem:co-finite-stabs\] for almost direct products. $G$ is said to be an [*almost direct product*]{} of the subgroups $G_1$ and $G_2$ if the two groups commute and $G=G_1G_2$. \[lem:co-finite-stabs2\] Let $G=G_1G_2$ be an lcsc almost direct product, and let $G \curvearrowright (X,{m})$ be a pmp action on a standard measurable space that is $G_1$-ergodic. Assume that there exists a $G_1$-invariant probability measure on ${m}$-almost every $G$-orbit. Then the action $G \curvearrowright (X,{m})$ is essentially transitive. Let $\beta^1 : X \to E(X)^{G_1}$ be a decomposition map of the $G_1$-action on $X$. Note that, since $G_1$ and $G_2$ commute, it holds for every $\eta \in E(X)^{G_1}$ and $g \in G$ that $g\eta \in E(X)^{G_1}$, and so $G$ acts on $E(X)^{G_1}$. Hence one can consider the question of whether $\beta^1$ commutes with $G$. In fact, in order to follow the same arguments of Lemma \[lem:co-finite-stabs\] we will require that $\beta^1$ be essentially $G$-equivariant. That is, that there exists an $A \subseteq X$ with $\theta(A)=1$ for any $G_1$-invariant measure $\theta$, and such that $g\beta^1_x =\beta^1_{gx}$ for all $g\in G$ and $x\in A$. Assume first that $\beta^1$ satisfies this condition. Since $m$ is $G_1$-invariant, ${m}(A)=1$. Hence there exists an ${m}$-full measure set $A' \subseteq A$ such that there exists a $G_1$-invariant and ergodic measure $\eta_x$ on $Gx$ for all $x \in A'$. Fix some $x \in A'$. Since $\eta_x(Gx)=1$, we can find some $y = gx \in Gx$ such that $\beta^1_y=\eta_x$. Then $$\begin{aligned} g\beta^1_x = \beta^1_{gx} = \beta^1_y = \eta_x \end{aligned}$$ and so we conclude that $\beta^1_x$ is supported on $Gx$ for every $x \in A'$. Since ${m}$ is $G_1$-invariant we can write $$\begin{aligned} {m}= \int_X\beta^1_xd{m}(x) = \int_{A'}\beta^1_xd{m}(x) \end{aligned}$$ and, by the $G_1$-ergodicity of ${m}$, ${m}=\beta^1_x$ for some $x \in A'$ and in particular ${m}(Gx)=1$. Hence the action $G \curvearrowright (G,{m})$ is essentially transitive. Finally, we argue that $\beta^1$ is essentially $G$-equivariant. We first show the simple case that $G_2$ is countable, and then prove the more general case below. Let $\theta$ be a $G_1$-invariant measure, and let $g\in G$. Applying the decomposition  to $g\theta$ yields $$\begin{aligned} \theta = \int_Xg^{-1}\beta^1_{gx}d\theta(x). \end{aligned}$$ which, by the essential uniqueness of $\beta^1$, implies that there exists an $A_g \subset X$ such that $g\beta^1_x = \beta^1_{gx}$ for every $x\in A_g$, and such that $A_g$ has full measure with respect to any $G_1$-invariant measure. Hence, if $G_2$ is countable, we get that $\beta^1$ is equivariant on $A=\bigcap_{g\in G_2} A_g$, where $A$ also has full measure with respect to any $G_1$-invariant measure. We are left with the case that $G_2$, and so $G$, is not countable. Following Varadarajan, let $\mathcal{F}$ be the Banach space of all bounded measurable functions on $X$. Let $U^1:\mathcal{F}\to \mathcal{F}$ be the operator defined by $U^1(f)(x)= \int_X f(y) d\beta^1_x(y)$, for any $f\in \mathcal{F}$. Note that by definition, the equivariance of $\beta^1$ is equivalent to the equivariance of $U^1$. Now, for any $G_1$-invariant measure $\theta$, $U^1$ is the conditional expectation defined by the factor $(X,\theta)\to {{G_1} \backslash\kern-.3em\backslash {(X,\theta)}}$: $$\begin{aligned} U^1:L^\infty(X,\theta) \to L^\infty({{G_1} \backslash\kern-.3em\backslash {(X,\theta)}}). \end{aligned}$$ Since the actions of $G_1$ and $G_2$ on $X$ commute, $U^1$, as the conditional expectation map, is $G_2$-equivariant. Hence $U^1 : \mathcal{F} \to \mathcal{F}$ is $G$-equivariant on a $\theta$-full measure set. Since this holds for any $G_1$-invariant measure $\theta$, we conclude that $U^1$, and so the associated decomposition map $\beta^1$, are essentially $G$-equivariant, and the proof is complete. Let $K$ be the associated stabilizer IRS. By Theorem \[thm:product-irs-rigidity\] and Corollary \[thm:rigidity\], there exists a normal subgroup $N = N_1 \times N_2 \lhd G$ such that $K$ is co-amenable in $N$, $G / N$ acts essentially freely on ${{N} \backslash\kern-.3em\backslash {(X,{m})}}$, and ${m}$-almost every orbit $Nx$ admits an $N$-invariant mean. Since $G_1$ is just non-amenable, either $N_1$ is co-amenable in $G_1$ or else $N_1$ is trivial. In the latter case we have that $K$ is contained (in fact, co-amenable) in $\{e\} \times N_2$, and since $K$ is irreducible, it must almost surely equal $\{e\} \times N_2'$, for some $N_2' \lhd G_2$. Since the action $G \curvearrowright (X,{m})$ is faithful, $N_2' = \{e\}$ and the action is essentially free. We are therefore left with the case that $N_1$ is co-amenable in $G_1$. Then since $K$ is co-amenable in $N_1 \times N_2$ we get that $K$ is co-amenable in $G_1\times N_2$, and, identifying each $G_1 \times N_2$-orbit with $(G_1 \times N_2) / (G_1 \times N_2)_x$, we have that ${m}$-almost every $G_1 \times N_2$-orbit admits a $G_1 \times N_2$-invariant mean. In particular, it admits a $G_1$-invariant mean. Since $G_1$ has property (T), $m$-almost every $G_1 \times N_2$-orbit admits a $G_1$-invariant probability measure, and so each $G$-orbit admits a $G_1$-invariant measure. It follow from Lemma \[lem:co-finite-stabs2\] that ${m}$ is essentially transitive. Assume that $G_1$ is a simple factor of $G$ with property (T), and note that by Theorem \[thm:lie-irs-rigidity\] $K$ is either almost surely equal to a normal subgroup $N$, or $K$ is co-amenable in $G$. In the former case, since the action is faithful, the only possibility is $N=\{e\}$, and then the action is essentially free. In the latter case, as in the proof of Corollary \[thm:rigidity-special\], we can identify each orbit $Gx$ with $G/G_x$, and so, by the co-amenability of $G_x$, there exists a $G_1$-invariant mean on ${m}$-almost every orbit $Gx$. Since $G_1$ has property (T), ${m}$-almost every orbit $Gx$ admits a $G_1$-invariant measure. It then follows from Lemma \[lem:co-finite-stabs2\] that ${m}$ is essentially transitive. The identification of $G_x$ with $G/G_x$ now implies that $K$ is almost surely co-finite in $G$. Finally, it follows from the Borel Density Theorem that $K$ is almost surely a lattice in $G$. [^1]: Y. Hartman is supported by the European Research Council, grant 239885. O. Tamuz is supported by ISF grant 1300/08, and is a recipient of the Google Europe Fellowship in Social Computing. This research is supported in part by this Google Fellowship.
{ "pile_set_name": "ArXiv" }
--- abstract: | LISA Pathfinder (LPF) is a mission aiming to test the critical technology for the forthcoming space-based gravitational wave detectors. The main scientific objective of the LPF mission is to demonstrate test-masses free-falling with residual accelerations below $3 \times 10^{-14} \textnormal{ m} \textnormal{ s}^{-2} / \sqrt{\textnormal{Hz}}$ at $1$ mHz. Reaching such an ambitious target will require a significant amount of system optimisation and characterisation, which will in turn require accurate and quantitative noise analysis procedures. In this paper we discuss two main problems associated with the analysis of the data from LPF: Excess noise detection and Noise parameter identification. The mission is focused on the low frequency region ($\left[0.1, 10\right]$ mHz) of the available signal spectrum. In such a region the signal is dominated by the force noise acting on test masses. At the same time, the mission duration is limited to $90$ days and typical data segments will be $24$ hours in length. Considering those constraints, noise analysis is expected to deal with a limited amount of non-Gaussian data, since the spectrum statistics will be far from Gaussian and the lowest available frequency is limited by the data length. In this paper we analyze the details of the expected statistics for spectral data and develop two suitable excess noise estimators. One is based on the statistical properties of the integrated spectrum, the other is based on Kolmogorov-Smirnov test. The sensitivity of the estimators is discussed theoretically for independent data, then the algorithms are tested on LPF synthetic data. The test on realistic LPF data allows the effect of spectral data correlations on the efficiency of the different noise excess estimators to be highlighted. It also reveals the versatility of the Kolmogorov-Smirnov approach, which can be adapted to provide reasonable results on correlated data from a modified version of the standard equations for the inversion of the test statistic. Closely related to excess noise detection, the problem of noise parameter identification in non-Gaussian data is approached in two ways. One procedure is based on maximum likelihood estimator and another is based on the Kolmogorov-Smirnov goodness of fit estimator. Both approaches provide unbiased and accurate results for noise parameter estimation and demonstrate superior performance with respect to standard weighted least-squares and Huber’s norm. We also discuss the advantages of the Kolmogorov-Smirnov formalism for the estimation of confidence intervals of parameter values in correlated data. author: - Luigi Ferraioli - Martin Hewitson - Giuseppe Congedo - Miquel Nofrarias - Mauro Hueller - Michele Armano - Stefano Vitale title: Quantitative Analysis of LISA Pathfinder Test Mass Noise --- Introduction {#intro} ============ LISA Pathfinder (LPF), a European Space Agency mission, will be used to characterize and analyze all possible sources of disturbance which perturb free-falling test masses from their geodesic motion. The system is composed of a single spacecraft (SC) enclosing a scientific payload, the LISA Technology Package (LTP), which is composed of two test masses (TMs) whose position is sensed by an interferometer. The spacecraft cannot simultaneously follow both masses, therefore the trajectory of only one test mass is used as a drag-free reference along the measurement axis. In order to prevent the trajectories of the test-masses from diverging, the second test mass is capacitively actuated to follow the first (free-falling) TM. In the main science operating mode, the first interferometer channel measures the displacement of the SC relative to the free-falling TM. The second interferometer channel (the differential channel) measures the relative displacement between the two TMs. LPF is a controlled system which can only be fully assessed during flight operation, therefore a considerable number of experiments will be devoted to the identification of the details of the dynamics of the system. A dynamical model of LPF is built in advance on the basis of physical considerations and from the results of test campaigns. The dynamical model is parametric so that it can be updated on the basis of the experiments that will be conducted during mission operations. The overall aim of the process is to reach the best free-fall quality (below $3 \times 10^{-14} \textnormal{ m} \textnormal{ s}^{-2} / \sqrt{\textnormal{Hz}}$ at frequencies around $1$ mHz) in a step-by-step procedure in which the result of the previous experiment is used to adjust the detailed configuration of the following experiments [@LTPVitale2011; @LTPMartin2011; @LTPPaul2011; @ArmanoCQG2009]. Such a demanding program requires daily analysis of the instrument signals constrained by two major factors: the amount of available data is tightly constrained by LTP mission duration ($90$ days), the telemetry bandwidth, and the length of each data segment (typically $24$ h); the scientific interest is mainly focused on the analysis of noise sources which act directly on the TMs since that should provide a baseline reference for the forthcoming space-based gravitational waves observatories [@LISAYellowBook; @Cornish2005; @DECIGO; @ASTROD]. The direct forces on the TMs are expected to dominate the instrument output in the frequency range $\left[0.1, 10\right]$ mHz. Sample power spectra are typically calculated with Welch’s averaging periodogram method (WOSA) [@Percival]. In order to keep enough frequency resolution at low frequencies, the sample power spectra can not be averaged more than few times (we average $4$ times in the present paper), this results in highly non-Gaussian data for which we are developing dedicated techniques. In particular the paper aims to propose a solution for two major data analysis challenges encountered in LPF: Different measurements of the same physical quantity can exhibit different noise content if they are performed under slightly different environmental conditions. The objective of LISA Pathfinder data analysis during operations will be to discover such differences, understand their origin and adjust spacecraft physical parameters accordingly. Such a problem requires reliable excess noise detection procedures which have to be based on solid statistical considerations; Along with the demonstration of unprecedented test-mass free-fall, LPF will provide a model for the expected test mass force noise for future space-based gravitational wave detectors. In order to do this, we need to be able to match an analytical model to a noisy power spectral density measurement. The quality of the match must be statistically quantified. Both data analysis problems deal with sample spectra and the corresponding statistical properties. Section \[sect.noisexcess\] reports on the properties of different experimental procedures for the detection of noise excess. In particular we considered two cases where the noise excess is evaluated with respect to reference data or a reference model. The accuracy of the methods is theoretically analyzed for the case of broadband and band-limited excess noise. In Section \[sect.noisemodident\] the problem of noise parameter estimation for non-Gaussian data is explored and an algorithm based on maximum likelihood is derived. In parallel, the application of the Kolmogorov-Smirnov formalism to the construction of a goodness of fit estimator is discussed. Section \[sect.nonstatnoise\] reports briefly on the extension of the analysis procedures to the case of non-stationary noise and time-frequency investigations. In Section \[sect.testofperformance\] we provide an application of the developed algorithms to synthetic data with LPF-like qualities. This allows us to shed light on the effects of data correlations on the accuracy of the developed excess noise estimators. The analysis procedures are available as MATLAB tools in the framework of the LTPDA Toolbox [@HewitsonCQG2009; @matlab; @toolbox]; an object oriented MATLAB Toolbox for advanced data analysis. Quantitative detection of noise power variations {#sect.noisexcess} ================================================ The problem of the detection of noise power variation in consecutive measurements can be formulated in two different ways. Two different measurements of the power are compared; The different measurements of the power are compared with a reference model. The problem in the first case is of general character and can be applied to a wide range of experiments, the second case, on the other hand, assumes that a reference model for the noise power is known and data must be compared against the given model in order to understand if the system is performing under known conditions. The latter case is likely to be the scenario for LPF operations. The quantity typically used for the detection of noise power variations is the total energy content of the data series. It is defined as $\mathcal{E} = T \sum_i{\left|x_i\right|^2}$. Unfortunately, $\mathcal{E}$, provides a poor estimator for two reasons: As soon as the data series $x_0,\ldots,x_{N-1}$ departs from zero mean Gaussian white noise, the statistic of $\mathcal{E}$ becomes ill-defined and the definition of a confidence interval becomes cumbersome; $\mathcal{E}$ provides global information as it is not sensitive to noise changes in a given frequency band. While the first problem could be overcome (without little difficulty) by a numerical identification of the expected statistic, the second problem suggests that a spectral based estimator would provide supplementary information which could be fundamental in discriminating different noise sources. Detection of noise variations with a model {#subsect:noisexcsmodel} ------------------------------------------ In the case that a reference model is available, the detection of excess noise in the spectral domain can be effectively implemented with a test on the normalized Welch’s overlapped segment averaging (WOSA) spectrum $R_{\text{WOSA}}\left(f_k\right)$ defined in equation (\[eqn.normwosa\]). In particular, the test can pursue two different philosophies of which one aims to test a global scalar indicator of the properties of the data and another aims to test the details of the statistical distribution of the data. A sensible estimator for the first approach is provided by the integral of the normalized spectrum which, in the discrete case, can be written: $$\label{eqn:IRdef} \mathrm{IR} = \sum_{k=1}^{N_f}{R_{\text{WOSA}}\left(f_k\right)}.$$ In the simplifying assumption of independent spectral data, the statistic of each element of $R_{\text{WOSA}}\left(f_k\right)$ is described by a gamma distribution as defined by equation (\[eqn.probdistwosa\]) with $\delta = 1/N_s$ and $h = N_s$. The sum over the different values at the frequencies $f_k$ is still a gamma distribution with $\delta = 1/N_s$ and $h = N_s N_f$. The expectation value for $\mathrm{IR}$ is easily obtained as $E\left[\mathrm{IR}\right] = \delta h = N_f$. The natural estimator for the second approach is provided by the empirical cumulative distribution function (ECDF). The ECDF for a data series can be defined as $F_N\left(x\right) = z/N$, where $z$ is the number of observations reporting a value less than or equal to $x$. $x$ denotes the values taken by the data, in our case $R_{\text{WOSA}}\left(f_k\right)$ with $k = 1,\ldots,N_f$. The ECDF can be tested against the theoretical expectation provided by equation (\[eqn.probdistwosa\]). If the model well represents the given sample spectrum then $R_{\text{WOSA}}\left(f_k\right)$ is distributed according to the expected gamma distribution. Alternatively, if the sample spectrum contains a excess noise with respect to the reference model, then the distribution of the normalized WOSA spectrum will be different from the one reported in equation (\[eqn.probdistwosa\]). The hypothesis that the two distributions are equal can be tested if a ’distance’ between the ECDF and the theoretical reference, $F_{\Gamma}$, is defined as: $$\label{eqn.ksdistance} d_K\left(x\right) = \left|F_N\left(x\right) - F_{\Gamma}\left(x\right)\right|,$$ with $d_K\left(x\right)$ assuming values on the interval $\left[0,1\right]$ and $K = N_f$. Kolmogorov found that the statistical properties of $$\label{eqn.dKdef} d_K = \text{max}\left[d_K\left(x\right)\right]$$ are independent from the specific distributions under test [@Kolmogorov1933; @Feller1948]. This property qualifies $d_K$ as an excellent candidate for the construction of a general test for cumulative distribution functions, the limiting statistic for $d_K$ was identified by Kolmogorov himself and then inverted by Smirnov [@Smirnov1939; @Miller1956] who provided an analytical expression for the calculation of $d_K$ as a function of the significance level. General details about the Kolmogorov-Smirnov test (KS-test) are reported in appendix \[appendix.KStest\]. It is interesting to calculate the expected sensitivity for the two estimators. $\mathrm{IR}$ is expected to be distributed as a gamma distribution (at least when the model and the data are in agreement), the corresponding cumulative distribution function (CDF) is provided by the incomplete gamma function $\mathcal{P}\left(h,x/\delta\right)$ [@AbramoStegun]. The CDF assumes values in the interval $\left[0,1\right]$ as it defines the probability associated with the observation $x$. The inverse of the CDF provides the critical values, $x_{\rho}$, associated with a given probability, $\rho \in \left[0,1\right]$. The confidence range for a probability, $\rho$, for a gamma distributed variable can then be defined by the boundaries $x_{lw} = \mathcal{P}^{-1}\left(\alpha/2\right)$ and $x_{up} = \mathcal{P}^{-1}\left(1-\alpha/2\right)$, with $\alpha = 1 - \rho$. We assert that the measured sample spectrum is compatible with the reference model if $x_{lw} \leq \mathrm{IR} \leq x_{up}$ for the given probability, $\rho$, or significance level, $\alpha$. If the noise excess is provided by a scale factor, $\gamma$, which affects the noise on the complete band of frequencies, the expected value for $\mathrm{IR}$ changes to $E\left(\mathrm{IR}\right) = \gamma N_f$. Therefore the detection threshold for $\gamma$ is fixed by the interval $x_{lw}/N_f \leq \gamma \leq x_{up}/N_f$. In other words, the $\mathrm{IR}$ can detect a noise difference with respect to the reference model only if $\gamma < x_{lw}/N_f$ or if $\gamma > x_{up}/N_f$. If $\gamma$ is non-zero only in a restricted band of frequencies $\left[f_a,f_b\right]$, then the expectation value for $\mathrm{IR}$ changes to $E\left(\mathrm{IR}\right) = N_f - N_{ab} + N_{ab} \gamma$. In this case $\mathrm{IR}$ can detect the noise difference only if $\gamma < \left(x_{lw} - N_f + N_{ab}\right)/N_{ab}$ or $\gamma > \left(x_{up} - N_f + N_{ab}\right)/N_{ab}$. $N_{ab}$ is the number of frequency points in the interval $\left[f_a,f_b\right]$. In the case of the Kolmogorov-Smirnov (KS) estimator, the ECDF of the WOSA normalized spectrum is compared with the theoretical expectation $\mathcal{P}\left(N_s,x N_s\right)$. In the case that the noise excess is simply a constant scale factor, $\gamma$, over the whole frequency band, the expected CDF for the normalized WOSA spectrum is $\mathcal{P}\left(N_s,x N_s/\gamma\right)$. The expected value for the KS test variable is then written as: $$\label{eqn.ksvardetectth} d_K\left(\gamma\right) = \mathrm{max}\left|\mathcal{P}\left(N_s,\frac{x N_s}{\gamma}\right)-\mathcal{P}\left(N_s,x N_s\right)\right|.$$ Once a significance value $\alpha$ is provided, the corresponding critical value, $d_K\left(\alpha\right)$, can be calculated from the equations for the inversion of the limiting CDF for $d_K$ [@Smirnov1939; @Miller1956]. The two distributions in equation (\[eqn.ksvardetectth\]) are incompatible at the given significance level if $d_K\left(\gamma\right) > d_K\left(\alpha\right)$. This defines the detection threshold for $\gamma$. If $\gamma$ is non-zero only in a restricted band of frequencies $\left[f_a,f_b\right]$, then the expected distribution for the normalized WOSA estimator is difficult to calculate. Nevertheless the detection threshold for the KS estimator can be calculated numerically from synthetic data. ![Non-detection ranges for $\mathrm{IR}$ and KS estimators. Noise excess is assumed as a constant multiplicative factor $\gamma$ extending along the whole band of frequencies. Darkened intervals define the thresholds for $\gamma$ detection, if $\gamma$ is larger or smaller than the shaded values it can be detected with a confidence of $95 \%$. $N_s$ refers to WOSA averages. $N_f$ is the number of frequency points in the spectrum.[]{data-label="fig:fig1"}](./fig1.pdf){width="45.00000%"} ![Non-detection ranges for $\mathrm{IR}$ and KS estimators in the case that noise excess coefficient $\gamma$ is different from zero in a restricted band $\left[f_a,f_b\right]$. Data are presented as a function of the ratio $N_{ab} / N_f$ where $N_{ab}$ is the number of frequency points in the interval $\left[f_a,f_b\right]$ and $N_f$ is the total number of frequency points in the spectrum. $N_s$ refers to WOSA averages. Confidence level for excess detection is fixed to $95 \%$.[]{data-label="fig:fig2"}](./fig2.pdf){width="45.00000%"} In figure \[fig:fig1\], the ranges of non-detectability of the KS and $\mathrm{IR}$ estimators are reported as a function of $N_f$ for three different values of the WOSA averages $N_s$. Data refer to the case that the scale factor, $\gamma$, extends over the complete frequency band. As can be clearly seen, the $\mathrm{IR}$ estimator always has a better sensitivity than the KS estimator. The sensitivity for the case of a band limited excess noise is reported in figure \[fig:fig2\] as a function of the ratio $N_{ab}/N_f$. As in the previous case, the $\mathrm{IR}$ estimator provides a better sensitivity with respect to the KS estimator. The difference is particularly relevant for $N_{ab}/N_f < 0.07$, below such values the sensitivity of the KS estimator becomes poor. Both in figure \[fig:fig1\] and in figure \[fig:fig2\] the confidence level for $\gamma$ detection is fixed at $95 \%$. It is worth noting that KS algorithm can be used directly on time series to quantitatively gauging departures from a given distribution (e.g. Gaussian). Once the ECDF for the data is calculated, it can be compared with the expected distribution by calculating $d_K(x)$ from equation \[eqn.ksdistance\]. Since the distribution of $d_K(x)$ coefficients is known it is straightforward to set a confidence threshold. The procedure can be applied even in presence of correlations thanks to the generalizations discussed in section \[subsect.test.excessnoise\] and in appendix \[appendix.KStest\]. Detection of noise variations without a model --------------------------------------------- In the case that an excess of noise has to be detected by comparing different measurements the Parseval’s theorem [@Percival] ($\mathcal{E} = \sum_{k=0}^{N/2}{P\left(f_k\right)}$) suggests that the sum of the elements of a sample spectrum in a given frequency band $\mathcal{E}_{ab} = \sum_{k=a}^{b}{P\left(f_k\right)}$ could provide a sensitive estimator for noise power variations. Such an estimator would be loosely equivalent to the $\mathrm{IR}$ discussed above, except that its statistic is hard to determine in a typical experimental situation. The statistic of $P\left(f_k\right)$ at each frequency, $f_k$, is defined by equation (\[eqn.sampspprobdist\]). Therefore, in the case of non-white noise, its parameters depend on $S\left(f_k\right)$ and the statistic is different at different frequencies. Thus the statistic of $\mathcal{E}_{ab}$ is not easily known. An interesting alternative to $\mathcal{E}_{ab}$ is provided by the KS estimator. Given two data series $\left\{x_{n_1}\right\}$ and $\left\{x_{n_2}\right\}$ of length $N_1$ and $N_2$ respectively. The hypothesis that their ECDFs have the same limiting cumulative distribution function $F\left(x\right)$ can be tested if a distance in the ECDFs space is defined as: $$\label{eqn.ksdistanceemp} d_K\left(x\right) = \left|F_{n_1}\left(x\right) - F_{n_2}\left(x\right)\right|,$$ with $d_K\left(x\right)$ defined on the interval $\left[0,1\right]$ and $K = \left(N_1 N_2\right) /\left(N_1 + N_2\right)$ [@Feller1948]. Also in this case, the statistical properties of $$\label{eqn.dKdefemp} d_K = \text{max}\left[d_K\left(x\right)\right]$$ are independent from the distributions of $x_{n_1}$ and $x_{n_2}$. The same equations used in the case of the comparison with a given model can now be used for the inversion of the cumulative statistic of $d_K$ [^1]. Considering the simplifying assumption of independent spectral data, the sensitivity of the KS estimator at a given significance level can be calculated in analogy to what was discussed in the previous paragraph. ![Non detection ranges for the KS estimator for the comparison of two sample spectra. A) Noise excess coefficient $\gamma \neq 0$ on the full frequency band. $N_s$ is the number of WOSA averages. $N_f$ is the number of frequency points in the sample spectrum. B) $\gamma \neq 0$ only in a restricted interval of frequencies $\left[f_a,f_b\right]$. $N_{ab}$ is the number of points in the interval. []{data-label="fig:fig3"}](./fig3.pdf){width="45.00000%"} In figure \[fig:fig3\] the calculated interval for non-detection is reported for the case that the excess, $\gamma$, is extending over the whole frequency band and for the case that $\gamma \neq 0$ in $\left[f_a,f_b\right]$. Noise model identification {#sect.noisemodident} ========================== Closely related to the problem of excess noise detection, the problem of noise model identification is one of the principal scientific objectives of LPF mission. Of particular interest is the identification of a model for the force noise acting on the TMs, which can be used as a baseline for the forthcoming space-based gravitational wave observatories. The main constrains on the identification of force noise on the TMs in LPF are fixed by the limited data series length, typically $24$ h; the frequency range in which force noise on the TMs is dominating the signal is below $10$ mHz; force noise data are not directly accessible since the system measures and reports TM displacement, requiring that force noise on TMs be reconstructed by a numerical procedure [@LISA7Anneke; @LISA7Luigi]. The result of the combination of the first two constraints is that the number of segment averages in the WOSA procedure for sample spectrum estimation should be taken as low as possible so as to have a reasonable number of frequency points in the range $f \in \left[0.1,10\right]$ mHz. As a consequence, the distribution of the WOSA spectrum strongly departs from a Gaussian distribution [^2], meaning that the classical least-squares minimization procedure for parameter estimation is not well conditioned and a full maximum likelihood procedure is required. Maximum Likelihood Parameter Estimation {#subsect.mlikelihhod} --------------------------------------- If we replace $S\left(f_k\right)$ in equation (\[eqn.normwosa\]) with a parametric model for the spectrum, $S\left(f_k;\theta_1,\ldots,\theta_H\right)$, the normalized WOSA spectrum becomes parametric, $R_{\text{WOSA}}\left(f_k;\theta_1,\ldots,\theta_H\right)$, and can be used for the estimation of noise model parameters $\left\{\theta_1,\ldots,\theta_H\right\}$. In this section we develop the likelihood formalism for the simple case that the noise model is a function of a single parameter, $S\left(f_k;\theta\right) = \theta S_0\left(f_k\right)$. This allows us, not only to find a sensible goodness of fit estimator for $R_{\text{WOSA}}\left(f_k;\theta\right)$ that can be used also in the case of multiple parameters, but also to place the excess noise estimator, $\mathrm{IR}$, in a more solid theoretical framework. Indicating with $\theta_{\text{TRUE}}$ the ‘true’ value for the $\theta$ parameter, $R_{\text{WOSA}}\left(f_k;\theta\right)$ can be rewritten as: $$\begin{split} R_{\text{WOSA}}\left(f_k;\phi\right) & = \phi \frac{P_{\text{WOSA}}\left(f_k\right)}{\theta_{\mathrm{TRUE}} S_0\left(f_k\right)} \\ & = \phi R_{\text{WOSA}}^{\mathrm{true}}\left(f_k\right). \end{split}$$ Here $\phi = \theta_{\mathrm{true}} / \theta$. Assuming that $\theta_{\mathrm{TRUE}} S_0\left(f_k\right)$ correctly reproduces the expected value for the spectrum, the distribution of $R_{\text{WOSA}}^{\mathrm{true}}\left(f_k\right)$ is reported in equation (\[eqn.probdistwosa\]) with $\delta = 1/N_s$ and $h = N_s$. The distribution of the samples $R_{\text{WOSA}}\left(f_k;\phi\right)$ is then: $$f\left(y;h,\delta,\phi\right) = \frac{e^{-\frac{y}{\phi \delta}} y^{h-1} }{\left(\phi \delta\right)^h \Gamma\left(h\right) }.$$ Under the simplifying assumption that the values of $R_{\text{WOSA}}\left(f_k;\phi\right)$ are independent for different $f_k$, the likelihood function can be written as: $$\label{eqn.NormWOSALike} \mathcal{L}\left(h,\delta,\phi\right) = \prod_{k} f\left(y_k;h,\delta,\phi\right) \Delta y.$$ Here $\Delta y$ is a constant term required to have a finite probability from the probability distribution function $f\left(y_k;h,\delta,\phi\right)$. $y_k$ are observed samples corresponding to $R_{\text{WOSA}}\left(f_k;\phi\right)$. It is typically more convenient to work with the natural logarithm of the likelihood function $l\left(h,\delta,\phi\right) = \ln \mathcal{L}\left(h,\delta,\phi\right)$. $$l\left(h,\delta,\phi\right) \sim \left(h-1\right) \sum_k{\ln y_k} - \frac{1}{\phi \delta} \sum_k{y_k} - N_f h \ln \phi.$$ $N_f$ is the total number of frequency samples. Taking the first derivative with respect to $\phi$ and equating to $0$ we find the maximum likelihood estimator for the parameter $\phi$: $$\Lambda = \frac{1}{\phi} \sum_k{R_{\text{WOSA}}\left(f_k;\phi\right)} - N_f.$$ The value of $\phi$ at which $\Lambda = 0$ corresponds to the maximum likelihood estimation for the parameter. The $\Lambda$ estimator is unbiased, since, remembering that $\phi = \theta_{\mathrm{true}} / \theta$, it can be verified that $\lim_{\theta \rightarrow \theta_{\mathrm{true}}} E\left[\Lambda\right] = 0$. For the practical purpose we can find the zero crossing of the reduced estimator $$\tilde{\Lambda} = \sum_k{R_{\text{WOSA}}\left(f_k;\theta\right)} - N_f,$$ which is crossing zero at the same $\theta$ value of $\Lambda$, since $\phi \rightarrow 1$ when $\theta \rightarrow \theta_{\mathrm{true}}$. It is worth noting that $\tilde{\Lambda}$ is practically the $\mathrm{IR}$ excess noise estimator with the expectation value $N_f$ subtracted. It is then worth noting that $\tilde{\Lambda}$ can be used not only in the simple case of one parameter but it can also be applied in the general case of a model $S\left(f_k;\theta_1,\ldots,\theta_H\right)$ since the condition $\tilde{\Lambda} \rightarrow 0$ when $\left\{\theta_1,\ldots,\theta_H\right\} \rightarrow \left\{\theta_{1\mathrm{true}},\ldots,\theta_{H\mathrm{true}}\right\}$ is always satisfied. Kolmogorov-Smirnov Parameter estimation {#subsect.KSparamestimate} --------------------------------------- As discussed in section \[subsect:noisexcsmodel\], the KS estimator can be used as an alternative to a maximum likelihood procedure for parameter estimation. Thanks to the statistical properties of $R_{\text{WOSA}}\left(f_k;\theta\right)$, the closer $\theta$ is to $\theta_{\mathrm{true}}$, the better the distribution of $R_{\text{WOSA}}\left(f_k;\theta\right)$ is described by equation (\[eqn.probdistwosa\]) with $\delta = 1/N_s$ and $h = N_s$. Therefore, the Kolmogorov-Smirnov distance parameter provides an effective goodness-of-fit estimator: $$d_K\left(\theta\right) = \max\left|F_R\left(x;\theta\right)-\mathcal{P}\left(N_s,x N_s\right)\right|.$$ $F_R\left(\alpha\right)$ is the ECDF for the current $R_{\text{WOSA}}\left(f_k;\theta\right)$ estimate, $\mathcal{P}\left(N_s,x N_s\right)$ is the limiting distribution for $\theta \rightarrow \theta_{\mathrm{true}}$. KS estimation for $\theta$ is obtained by the minimization of $d_K\left(\theta\right)$ with respect to $\theta$. A confidence range for the parameter estimation can be readily defined from the non-rejection region of the KS-test at a given significance level. In practice, having defined a significance level, the corresponding critical values of the KS statistic, $d_K\left(\alpha\right)$, can be calculated with standard equations [@Miller1956] or Monte Carlo simulations in the case of correlated data. The values of $\bar{\theta}$ for which $d_K\left(\bar{\theta}\right) = d_K\left(\alpha\right)$ provide the boundary for the confidence range at the given significance. The KS statistic can also be successfully applied to multi-parameter identification, since the convergence of $F_R\left(\alpha\right)$ to $\mathcal{P}\left(N_s,x N_s\right)$ is always verified when $\left\{\theta_1,\ldots,\theta_H\right\} \rightarrow \left\{\theta_{1\mathrm{true}},\ldots,\theta_{H\mathrm{true}}\right\}$. Analysis of non-stationary noise {#sect.nonstatnoise} ================================ The implementation of noise analysis procedures was so far discussed in the context of stationary or [*pseudo-stationary*]{} noise [^3]. In the case of truly non-stationary noise the spectral content of a time series is investigated by time-frequency analysis techniques which include the spectrogram and the wavelet transform. The spectrogram is estimated by the square modulus of the short-time Fourier transform of the data [@Oppenheim1999]. It provides a direct extension of the PSD formalism to non-stationary time series. Given a data series of $N$ samples, a fraction of length $Q < N$ is windowed and the Fourier transform is applied. Then the data window is time shifted and the process is repeated. The calculation of the spectrogram is based on data windowing and the application of the Fourier transform, therefore the considerations noted in the previous sections for stationary noise can be applied directly to the spectrogram analysis of non-stationary noise. Since the short-time Fourier transform has the same resolution across the time-frequency plane, it is often preferable to resort to the wavelet transform. Wavelet transform is a decomposition of the time series over time-frequency elements that are obtained by scaling and translating a mother function $\psi \in \mathbf{L^2}\left(\mathbb{R}\right)$: $$\psi_{u,s} = \frac{1}{\sqrt{s}} \psi \left(\frac{t-u}{s}\right).$$ The function $\psi\left(t\right)$ has zero average and the wavelet elements $\psi_{u,s}$ are normalized to 1. The wavelet transform of a function $f(t)$ is then defined as: $$Wf(u,s) = \int_{-\infty}^{\infty} f(t)\frac{1}{\sqrt{s}} \psi^* \left(\frac{t-u}{s}\right) dt.$$ In the discrete case, the result of the wavelet transform on a time series is an array of coefficients $w_{u,s}$ where $u$ is the time index and $s$ is the scale index which is associated to a given frequency band [@MallatWavelet]. In the case of uncorrelated Gaussian noise, the distribution of the coefficients $w_{u,s}$ is still Gaussian with a certain amount of correlation introduced by the convolution-like transform operation [@MallatWavelet]. In such a favorable situation the extension of the method described in the stationary case it appears straightforward. In particular Kolmogorov-Smirnov procedures are excellent candidates since their robustness to correlation and the possibility of extending KS distance definition to a two-dimensional space [@Peacock1983; @Fasano1987]. Application to LPF synthetic data {#sect.testofperformance} ================================= In this section the different procedures for excess noise detection and parameter estimation are applied to synthetic LPF data. This provides not only an interesting framework for testing their accuracy and precision, but it also helps clarifying the role of correlations among spectral data. Synthetic data and noise projection ----------------------------------- LPF provides two output channels along the principal measurement axis which are sensing the displacement of the SC relative to the free falling TM and the relative displacement between the TMs. From the knowledge of the displacement signals and a linear model for the system dynamics, an effective force-per-unit-mass, $a_{\rm eff}$, acting on the TMs can be extracted. $a_{\rm eff}$ is the combination of the ‘true’ force-per-unit-mass acting on TMs and a projected interferometer readout noise. Details of the calculation are reported in appendix \[appendix.conv2acc\]. Following the same scheme, a given prediction for the different input noise sources can be projected to a prediction for the power spectrum of $a_{\rm eff}$, as reported in figure \[fig:fig4\]. Here the model for the spectrum of force noise acting on TMs and the model for the spectrum of readout noise are projected into a model for the power spectrum of $a_{\rm eff}$ (TMs + Readout in the figure notation), which represents this paper’s baseline for the spectrum of $a_{\rm eff}$. ![Projection of the spectrum of $a_{\rm eff}$. ‘Readout’ is the projection of the readout noise to $a_{\rm eff}$. ‘TMs’ is the projection of the force noise on the test masses to $a_{\rm eff}$. ‘TMs + Readout’ is the complete noise projection for $a_{\rm eff}$, it represents the baseline noise level assumed in the present paper. ‘LPF worst case’ refers to a worst case scenario for $a_{\rm eff}$ and ‘LPF Spec.’ corresponds to the mission specifications.[]{data-label="fig:fig4"}](./fig4.pdf){width="45.00000%"} In figure \[fig:fig4\] we also report the project specifications for LPF and the expected noise spectrum for $a_{\rm eff}$ in a worst case scenario. In our baseline we assumed a reduced force noise on the TMs compared to the worst case but choose to keep the worst case for the readout noise. This was done in order to represent one of the possible scenarios (not the best one) that can be experienced during the mission. The model assumed for the force noise on the TMs is characterized by a low frequency $1/f^2$ behavior and a flat part for $f > 1$ mHz. The model can be written as $S_{\rm TM}\left(f\right) = \theta S^0_{\rm TM}\left(f\right)$. $\theta$ is an adjustable parameter which assumes values $\theta = 1$ for the worst case scenario and $\theta = 0.1$ for our baseline model. It is worth noting that $S_{\rm TM}\left(f\right)$ is projected (together with the readout noise model) through LPF dynamics in order to obtain the expected noise spectrum for $a_{\rm eff}$ which we indicate with $S_a\left(f\right)$. $S_a\left(f\right)$ with $\theta = 0.1$ corresponds to ‘TMs + Readout’ in figure \[fig:fig4\], $S_a\left(f\right)$ with $\theta = 1$ corresponds, instead, to the ‘LPF worst case’. Excess noise detection {#subsect.test.excessnoise} ---------------------- A change of the noise level on $S_{\rm TM}\left(f\right)$ ($\theta \neq 0.1$) produces a variation of the energy content of $a_{\rm eff}$. Such variation, which may be ‘improperly’ identified as excess noise, can be detected with the procedures defined in section \[sect.noisexcess\]. In particular we tested the detection of excess noise between two data series and between a data series and a reference model. Synthetic data were produced according to the following procedure: 1. Different models for $S_a\left(f\right)$ are produced changing $\theta$ around the reference value $\theta = 0.1$. Readout noise level is kept fixed. 2. Corresponding noise time series for $a_{\rm eff}$ are generated using the procedure reported in [@FerraioliPRD2010]. The time series are $24$ h long and have a sampling frequency of $1$ Hz. 3. Sample spectra are calculated for each series with the WOSA algorithm. We chose a Blackman-Harris data window, $50 \%$ segment overlap and number of segments averages $N_s = 4$. 4. The analysis is restricted to the frequency interval $\left[0.1, 10\right]$ mHz since, as can be seen from figure \[fig:fig4\], it represents the region in which the force noise on the TMs dominates $S_a\left(f\right)$. Spectral data are tested for excess noise. We used the KS algorithm (equation (\[eqn.dKdefemp\])) in the case of the test of two data series. The data series for $\theta \neq 0.1$ are compared against the reference series with $\theta = 0.1$. In the case of the test of a data series against a model, both the KS algorithm (equation (\[eqn.dKdef\])) and the IR algorithm (equation (\[eqn:IRdef\])) are used. The reference model is the projected $S_a\left(f\right)$ for $\theta = 0.1$. The results of the tests are summarized in table \[tbl.testresults\]. KS critical values $d_K\left(\alpha\right)$ and IR confidence intervals are calculated for a significance level $\alpha = 0.05$ which corresponds to a $95 \%$ confidence. Each value of $\theta$ corresponds to a value of the in-band energy content $E\left(\theta\right)$ of $S_a\left(f\right)$ in the analyzed frequency band. We report in table \[tbl.testresults\] the relative change in energy $\Delta E / E$ corresponding to a relative change in $\theta$. The quantity $\Delta E / E$ plays the same role of the parameter $\gamma$ in figures \[fig:fig1\] and \[fig:fig3\], even though the two quantities are not completely comparable since $\gamma$ in figure \[fig:fig1\] assumes independence of the data. Spectral data are correlated among different frequency values because of two effects [@Percival; @Thomson1977]: Data windowing which corresponds to a convolution in the frequency domain of the window function with the sample spectrum; WOSA overlapped segment averaging. The first effect is unavoidable, the second effect, instead, can be attenuated by a proper choice of the segment overlap. It can be demonstrated [^4] that for a Blackmann-Harris window the effect is negligible with $50 \%$ overlap. ------------------------------- ---------------------------------------- --------------- -------------------------- ----------------------------------- --------------- -------------------------- ------------------------------------ ----------------------------------- ------------------------------------------- ------------------------------- ------------------------------- $d_K\left(\alpha\right)$ $d_K^{\rm MC}\left(\alpha\right)$ $d_K\left(\alpha\right)$ $d_K^{\rm eff}\left(\alpha\right)$ $d_K^{\rm MC}\left(\alpha\right)$ Conf. Int. MC Conf. Int. $\mathbf{\frac{\Delta E}{E}}$ $\mathbf{\frac{\Delta\theta}{\theta}}$ $d_K$ $0.1030$ $0.1006$ $d_K$ $0.0730$ $0.0982$ $0.0969$ $\sum_k{R_{\text{WOSA}}\left(f_k\right)}$ $\left[323.14, 359.33\right]$ $\left[311.33, 372.55\right]$ $-0.14$ $ -0.7 $ $ 0.2352 $ $ \times $ $ \times $ $ 0.2547 $ $ \times $ $ \times $ $ \times $ $ 250.74 $ $ \times $ $ \times $ $-0.08$ $ -0.4 $ $ 0.1424 $ $ \times $ $ \times $ $ 0.1103 $ $ \times $ $ \times $ $ \times $ $ 303.72 $ $ \times $ $ \times $ $-0.06$ $ -0.3 $ $ 0.1806 $ $ \times $ $ \times $ $ 0.1747 $ $ \times $ $ \times $ $ \times $ $ 274.54 $ $ \times $ $ \times $ $-0.05$ $ -0.25 $ $ 0.1740 $ $ \times $ $ \times $ $ 0.1685 $ $ \times $ $ \times $ $ \times $ $ 294.42 $ $ \times $ $ \times $ $-0.04$ $ -0.2 $ $ 0.0887 $ $ \surd $ $ \surd $ $ 0.0696 $ $ \surd $ $ \surd $ $ \surd $ $ 317.10 $ $ \times $ $ \surd $ $-0.03$ $ -0.15 $ $ 0.0906 $ $ \surd $ $ \surd $ $ 0.0433 $ $ \surd $ $ \surd $ $ \surd $ $ 324.87 $ $ \surd $ $ \surd $ $-0.01$ $ -0.05 $ $ 0.0771 $ $ \surd $ $ \surd $ $ 0.0302 $ $ \surd $ $ \surd $ $ \surd $ $ 339.47 $ $ \surd $ $ \surd $ $0.01$ $ 0.05 $ $ 0.0551 $ $ \surd $ $ \surd $ $ 0.0539 $ $ \surd $ $ \surd $ $ \surd $ $ 352.26 $ $ \surd $ $ \surd $ $0.03$ $ 0.15 $ $ 0.0629 $ $ \surd $ $ \surd $ $ 0.0370 $ $ \surd $ $ \surd $ $ \surd $ $ 345.38 $ $ \surd $ $ \surd $ $0.04$ $ 0.2 $ $ 0.0603 $ $ \surd $ $ \surd $ $ 0.0635 $ $ \surd $ $ \surd $ $ \surd $ $ 357.14 $ $ \surd $ $ \surd $ $0.05$ $ 0.25 $ $ 0.0519 $ $ \surd $ $ \surd $ $ 0.0818 $ $ \times $ $ \surd $ $ \surd $ $ 369.05 $ $ \times $ $ \surd $ $0.06$ $ 0.3 $ $ 0.0857 $ $ \surd $ $ \surd $ $ 0.1253 $ $ \times $ $ \times $ $ \times $ $ 388.01 $ $ \times $ $ \times $ $0.08$ $ 0.4 $ $ 0.1103 $ $ \times $ $ \times $ $ 0.1137 $ $ \times $ $ \times $ $ \times $ $ 402.14 $ $ \times $ $ \times $ $0.14$ $ 0.7 $ $ 0.1520 $ $ \times $ $ \times $ $ 0.2423 $ $ \times $ $ \times $ $ \times $ $ 459.64 $ $ \times $ $ \times $ ------------------------------- ---------------------------------------- --------------- -------------------------- ----------------------------------- --------------- -------------------------- ------------------------------------ ----------------------------------- ------------------------------------------- ------------------------------- ------------------------------- Since the standard statistics for the estimators ($d_K\left(\alpha\right)$ for KS and confidence interval for IR) are calculated in section \[sect.noisexcess\] with the assumption of data independence, the standard critical values, $d_K\left(\alpha\right)$, for the KS estimators and the confidence intervals for the IR estimator can be applied only if the correlations among data are negligible. If this is not the case, the effective statistic of the estimators can be numerically calculated with Monte Carlo simulations. The corresponding results of a Monte Carlo simulation with $N_{\rm MC} = 5000$ realizations of the reference data series are indicated in table \[tbl.testresults\] with the suffix MC. The symbol $\surd$ is used to indicate that the spectral data for the corresponding $\Delta E/ E$ is compatible with the reference. On the contrary the symbol $\times$ indicates a rejection. Observing the test results reported in columns $7$ and $9$ for the KS algorithm and the results in columns $11$ and $12$ for the IR algorithm it is readily seen that correlations among data play a role. In the case of the KS test, the comparison of the data with $d_K\left(\alpha\right)$ determines a rejection for $\Delta E/ E = 0.05$. On the other hand, the same value is accepted when Monte Carlo result $d_K^{\rm MC}\left(\alpha\right)$ is used. In the case of the IR estimator, the comparison with the standard confidence interval leads to the rejection in correspondence with $\Delta E / E = 0.05$ and $\Delta E / E = -0.04$. Such values are instead considered compatible by the Monte Carlo confidence interval. These results provide us with clear information that the presence of correlations among data has affected the tests statistics and therefore the standard equations, assuming independence among data, are not usable in this situation. Looking at the results for the KS test between two data series, we discover that $d_K\left(\alpha\right)$ and $d_K^{\rm MC}\left(\alpha\right)$ are practically equal. In fact the results for the two corresponding columns of table \[tbl.testresults\] (columns $4$ and $5$) are in perfect agreement. This is the practical result of one of the most interesting properties of the KS test. The test is based on equation (\[eqn.dKdefemp\]). It states that the two empirical cumulative distributions under test have the same limiting CDF. Since the sample spectra in our test are calculated following the same WOSA procedure, the degree of correlation among different frequency points is the same, therefore the test statistic is not spoiled. Comparing $d_K\left(\alpha\right)$ and $d_K^{\rm MC}\left(\alpha\right)$ of columns $7$ and $9$ we see that the effect of correlation is to increase the maximum expected spread between ECDF and limiting CDF. Therefore the effect of data correlation is to distort the expected statistic for $d_K$. The ‘distortion’ of $d_K$ statistic can be taken into account if an effective value for the parameter $K$ is introduced. In the case of the comparison of the ECDF for correlated data against a theoretical CDF the application of the standard values for $d_K$, where $K = N_f$, $N_f$ being the number of data elements, leads to a statistically unfair test. We then discovered that test fairness can be recovered if an effective value for $K$ is used rather than the standard $K = N_f$. In particular, for spectral data produced with the WOSA method, Blackmann-Harris window, $N_s = 4$ averages on $50 \%$ overlapped segments, we obtained $K_{\rm eff} = \beta N_f$ with $\beta = 0.55$ for a significance level $\alpha = 0.05$. It is worth noting that the value of $\beta$ is independent from the number of data points considered, and from the spectral shape, provided that the different shapes have reasonably comparable smoothness on a frequency interval comparable with the width of the first lobe of the data window. As an example, the value of $\beta = 0.55$ is valid for LPF-like data and for white-noise equivalently. The requirement on the smoothness of the spectrum is connected to the expression of the correlations introduced by data windowing. It can be demonstrated that if the spectrum can be assumed constant in a frequency range of the order of the width of the first lobe of data window [@Percival] then the correlations are independent from the particular shape of the spectrum and are determined only by the window function. For such class of spectra we expect the same value for $\beta$ once the required significance level is fixed. Noise model identification {#subsect.test.noisemodelident} -------------------------- The problem of noise parameter identification is strictly connected to the problem of excess noise detection by a comparison of a data series with a reference model. While in excess noise detection different data series are compared with a given reference, in parameter estimation different realizations of a parametric model are compared with a dataset in order to find the best fit. Due to this, the same algorithms (i.e. KS and IR) can be applied to the solution of the two problems. Precision and accuracy in the estimation of the parameter, $\theta$, controlling the excess force noise on the TMs is tested with a Monte Carlo simulation over $N_{\rm MC} = 5 \times 10^3$ realizations of the same process. The data series reproduce $a_{\rm eff}$ corresponding to the reference value $\theta = 0.1$. Sample spectra are calculated with the procedure described above and the analysis is restricted to the frequency range $\left[0.1, 10\right]$ mHz. For each realization, data are compared with a bank of models obtained by the projection of TMs force noise $S_{\rm TM}\left(f;\theta\right) = \theta S^0_{\rm TM}\left(f\right)$ and readout noise into $S_a\left(f;\theta\right)$ for different values of $\theta$ around the reference value. The KS and IR estimators are calculated for each element of the model bank, in particular it is simpler to analyze the results in terms of $\tilde{\mathrm{IR}} = \left|\mathrm{IR} - N_f\right|$. Both KS and $\tilde{\mathrm{IR}}$ are expected to have a minimum corresponding to the best estimate for the parameter $\theta$. The two methods proposed here are compared with the performance of a classical weighted least-squares method, which works by minimizing the mean squared error $\mathrm{MSE} = \sum_{f_k}{\left(\left(P_{WOSA}\left(f_k\right)-S_a\left(f_k;\theta\right)\right) / S_a\left(f_k;\theta\right)\right)^2 }$, and a Huber’s norm estimator (details are reported in appendix \[appendix.huber\]). The results of the analysis are reported in figure \[fig:fig5\], where the histograms of the best fit vales over $N_{\rm MC}$ realizations are reported for the four procedures. In the same figure we also show the evolution along the model grid of the four estimators for a particular set of data from the available $N_{\rm MC}$. ![image](./fig5.pdf){width="80.00000%"} The distributions for the best fit parameter are reasonably symmetric for all the estimators; mean values and sample standard deviations are respectively $\theta_{KS} = 0.100$, $\sigma_{KS} = 0.012$, $\theta_{IR} = 0.100$, $\sigma_{IR} = 0.012$, $\theta_{MSE} = 0.148$, $\sigma_{MSE} = 0.016$, $\theta_{Huber} = 0.130$ and $\sigma_{Huber} = 0.014$. From the analysis of the Monte Carlo results, it is readily seen that both the KS and the $\tilde{\mathrm{IR}}$ estimators provide equivalently precise and accurate results. On the other hand, the MSE algorithm provides a poor estimation, both from the accuracy and from the precision point of view. The best estimation for the parameter is $\theta_{MSE} = 0.148$, which is strongly biased with respect to the reference value of $\theta = 0.1$. Also, the distribution of the parameter values is wider than those obtained from the KS and $\tilde{\mathrm{IR}}$ estimators. Huber’s norm estimator, with the chosen parameter $c = 0.05$ \[appendix.huber\], performs better than MSE but the result $\theta_{Huber} = 0.130$ is still far from the true value. It is worth discussing the effect of correlations among spectral data on the estimation of the parameter $\theta$. As can be observed from the results of the MC simulation, the accuracy of the estimators is not affected by correlations; the results for KS and $\tilde{\mathrm{IR}}$ estimators are practically indistinguishable. Some problems with $\tilde{\mathrm{IR}}$ can arise from data correlations when the confidence interval for a single estimation is required. As discussed in the previous paragraph, correlations modify the statistic of the IR estimator making it impossible to easily calculate confidence intervals from the standard equations. The statistic of the KS estimator is affected by correlations too, but we have demonstrated that accurate critical values $d_K\left(\alpha\right)$ can be recovered if an effective value for $K$ is used, $K_{\rm eff} = \beta N_f$. $\beta$ is a shape parameter depending only on the correlations in the spectral data and on the required significance level. It can be calculated for a specific spectrum (e.g., white noise) and effectively extended to a wide family of spectral shapes. Using the value $\beta = 0.65$ reported in appendix \[appendix.KStest\], we obtain $d_K^{\rm eff}\left(\alpha\right) \approx 0.064$ for $\alpha = 1 - 0.68 \%$. The intersection of such a value with the curve of $d_K$ as a function of $\theta$ reported as black dots in figure \[fig:fig5\]A provides a $68 \%$ confidence interval for the single estimation. In this specific case such an interval is $\theta \in \left[0.086, 0.114\right]$. Conclusions =========== The problem of excess noise detection and noise parameter estimation for non-Gaussian data is analyzed in the framework of the LISA Pathfinder mission. Excess noise detection can be approached in two ways. In one way, the noise content of a data series is compared with a reference data series, in the other way the noise content of the data is compared with a reference model. In the first case, simple estimators like the total energy content in a data series are not suitable for formulating quantitative statements on a solid statistical basis. As an alternative a Kolmogorov-Smirnov (KS) estimator is proposed and successfully applied to LPF synthetic data. The KS estimator has the advantage of being independent of the statistical properties of the data under test. It is demonstrated in the paper that such a convenient property makes the estimator robust to correlations among spectral data. Two different estimators are investigated for the problem of comparing a data series to a model for excess noise detection. One estimator (IR) is based on the integral of the normalized WOSA spectrum, the other is a KS estimator for the comparison of an empirical cumulative distribution function with a limiting theoretical function. Despite the fact that IR estimator proves to have better sensitivity on independent data, the versatility of the KS estimator is highly advantageous on correlated data. Since the statistics of the estimators are distorted by the presence of correlations among test data, the standard and simple procedures for the determination of the confidence intervals, based on the inversion of the limiting distribution function, provide inaccurate results. While, in the case of the IR estimator, the problem can be overcome only with dedicated Monte Carlo simulations, we demonstrated that the introduction of a shape parameter allows us to use standard equations to calculate proper boundaries for KS confidence intervals. The shape parameter depends on the required significance level and on the data correlations. It does not depend on the number of data considered for the test. Since correlations among spectral data are mainly introduced by the windowing process, the shape parameter is fixed for a wide family of spectral shapes. For example, synthetic LPF data share the same shape parameter with white noise. Closely related to noise excess detection, the problem of noise parameter identification is analyzed with a maximum likelihood approach which, in the particular case of linear dependence on a single parameter, provides an algorithm analogous to the IR estimator used for excess noise detection. A KS algorithm was proposed as an alternative to the IR algorithm and the accuracy and precision of both were tested with a Monte Carlo simulation on LPF synthetic data. Both the IR and KS estimators were demonstrated to give equivalently good results, even though the capability of the KS to handle data correlation is a clear advantage for the definition of a confidence interval for the estimated noise parameter. Data analysis procedures introduced in this paper are easily extended to the vast context of time-frequency analysis of non-stationary noise. KS algorithm can be applied effectively both to spectrogram and wavelets coefficients provided that correlations among data are taken into account. KS algorithm has the advantage that can be generalized to two dimensions thus allowing to extend the analysis to the time-frequency plane. Statistic of the sample spectrum {#sec.spectrumstat} ================================ In the case of a discrete, real-valued stationary process $\left\{x_h\right\}$, the continuous spectral density function is defined as: $$S\left(f\right) = T \sum_{h=-\infty}^{\infty}{s_h e^{- \imath 2 \pi f h T}}.$$ Here $s_h$ is the autocovariance sequence of the process $\left\{x_h\right\}$, $T$ is the sampling period and $f$ is the frequency expressed in Hz. $f$ is defined in the range $\left|f\right| \leq f_{Nq} \equiv 1/2T$ and $f_{Nq}$ is known as the Nyquist frequency. In the case of a finite representation $x_0,\ldots,x_{N-1}$ of the discrete process $\left\{x_h\right\}$, the approximation to the spectral density function is provided by the sample spectrum $$\label{eqn.sampspect} \tilde{P}\left(f\right) = \frac{T}{N}\left|\sum_{h=0}^{N-1}{x_h e^{- \imath 2 \pi f h T}}\right|^2.$$ $f$ in this case is also defined on the interval $\left[-f_{Nq},f_{Nq}\right]$. If the sample spectrum is calculated on the grid of Fourier frequencies ($f_k = k/\left(N T\right)$, $\left|k\right| \leq N/2$) then it corresponds to the squared modulus of the discrete Fourier transform of the data sequence $x_0,\ldots,x_{N-1}$. In practical applications only the positive frequency part of the spectrum is considered, and the one sided sample spectrum is defined as $P\left(f_k\right) = 2 \tilde{P}\left(f_k\right)$ with $k = 0, 1, \ldots, N/2$. In the rest of the paper the one sided sample spectrum will be simply named the sample spectrum. If the data series $x_0,\ldots,x_{N-1}$ is Gaussian distributed and the elements $x_j$ are independent then the Fourier transform produces a complex series $X\left(f_k\right)$ whose elements are approximately independent and their real and imaginary parts are Gaussian distributed. The term $\left|X\left(f_k\right)\right|^2 = \left|\Re\left[X\left(f_k\right)\right]\right|^2 + \left|\Im\left[X\left(f_k\right)\right]\right|^2$ is then the sum of two independent variables distributed as a $\chi^2_{\nu}$ where $\nu$ is the number of degrees-of-freedom of the distribution ($\nu = 1$ in this case). If the correlations among the elements of the data series $x_0,\ldots,x_{N-1}$ are non-vanishing, the statistical properties of the sample spectrum, in the simplifying assumption of independent $P\left(f_k\right)$ elements, can be calculated from the case of a $\chi^2_{\nu}$ distributed variable $y$ multiplied by a constant $z = \lambda y$. The characteristic function for $z$ is $$\label{eqn.sampspcritical} \phi_z\left(t\right) = E\left[e^{\imath t \lambda y}\right] = \left(1 - 2 \imath t \lambda \right)^{-\frac{\nu}{2}}.$$ Here $E\left[\right]$ indicates the expected value. The inverse Fourier transform of $\phi_z\left(t\right)$ provides the probability density function for $z$: $$\label{eqn.sampspprobdist} \mathcal{F}^{-1}\left[\phi_z\left(t\right)\right] = \frac{e^{-\frac{z}{2\lambda}} z^{\left(\frac{\nu}{2}-1\right)}}{\left(2\lambda\right)^{\frac{\nu}{2}}\Gamma\left(\frac{\nu}{2}\right)}.$$ This is a gamma distribution, $f\left(z;k,\theta\right)$, with $k = \nu/2$ and $\theta = 2\lambda$. $\Gamma\left(k\right)$ is the gamma function. In the case of the sample spectrum at a given frequency, $z = P\left(f_k\right)$ and $\lambda = E\left[P\left(f_k\right)\right]/\nu \approx S\left(f\right)/\nu$ [^5]. It is useful, for the statistical analysis of the spectrum, to introduce the normalized sample spectrum $$R\left(f_k\right) = \nu \frac{P\left(f_k\right)}{S\left(f\right)},$$ which, at each frequency $f_k$, is distributed as $\chi^2_{\nu}$. The sample spectrum in equation (\[eqn.sampspect\]) can also be seen as a special case of the power spectral density, $S\left(f\right)$, when the infinite data series $\left\{x_h\right\}$ is chopped by a square data window of length $N$. This operation introduces a considerable amount of spectral leakage because of the convolution with the frequency response of the square data window [@Percival; @Harris1978]. Therefore it is common practice to multiply the time series $x_0,\ldots,x_{N-1}$ with a more performant data window, which increases the accuracy of the sample spectrum in the case of processes with a high dynamic range. The application of a data window introduces correlations among different elements of the sample spectrum. Such correlations affect the statistics of the spectrum, resulting in a change in the probability distribution of the sample spectrum. An analytical treatment of the spectrum statistics under such conditions is cumbersome, and it is easier to numerically evaluate (with a Monte Carlo simulation) the statistics of the sample spectrum for the case of interest. In order to improve the variance properties of the sample spectrum, Welch’s overlapped segment averaging (WOSA) method is applied [@Percival]. The data series $x_0,\ldots,x_{N-1}$ is divided in overlapping windowed segments. The estimates of the sample spectrum of each segment are then averaged. The practice of averaging overlapping segments can modify the expected statistics of the spectrum since the data in different segments can be correlated. Then, even in the simplifying assumption of vanishing spectral correlations from windowing and overlapping, the averaging process changes the statistics of the estimated sample spectrum. In the assumption of vanishing window and overlap correlations, the statistic of the WOSA spectrum corresponds to the average of $N_s$ gamma distributed variables $$P_{\text{WOSA}}\left(f_k\right) = \frac{1}{N_s}\sum_{j=1}^{N_s}{P_j\left(f_k\right)}.$$ The critical function for the sum is $\phi_{\text{P}_{\text{WOSA}}}\left(t\right) = \prod_j \phi_{j}\left(t\right)$, where $\phi_j\left(t\right)$ is the critical function for $\text{P}_j/N_s$. Thanks to equation (\[eqn.sampspcritical\]) $$\phi_{\text{P}_{\text{WOSA}}}\left(t\right) = \left(1 - \frac{\imath t \theta}{N_s}\right)^{-k N_s},$$ where $\theta = 2 S\left(f_k\right)/\nu$ and $k = \nu/2$. The inverse Fourier transform of $\phi_{\text{P}_{\text{WOSA}}}\left(t\right)$ provides the probability density function for the WOSA spectrum $$\label{eqn.probdistwosa} f_{\text{WOSA}}\left(z; h, \delta\right) = \frac{z^{\left(h-1\right)} e^{-\frac{z}{\delta}}}{\delta^h \Gamma\left(h\right)},$$ with $\delta = S\left(f_k\right)/N_s$ and $h = N_s$. Again it is useful to define a normalized WOSA spectrum as $$\label{eqn.normwosa} R_{\text{WOSA}}\left(f_k\right) = \frac{P_{\text{WOSA}}\left(f_k\right)}{S\left(f_k\right)},$$ which is gamma distributed (equation (\[eqn.probdistwosa\])) with $\delta = 1/N_s$ and $h = N_s$. Kolmogorov - Smirnov Test {#appendix.KStest} ========================= Kolmogorov - Smirnov is a well known test for distributions [@Kolmogorov1933; @Smirnov1939; @Feller1948; @Miller1956; @Fisher1983; @Wilk1968]. An empirical cumulative distribution (ECDF) is tested against a continuous theoretical model or, alternatively, two ECDFs are tested with the hypothesis that they share the same limiting cumulative distribution function. Indicating with $f\left(x\right)$ the probability density function associated with a given random process $X$, the corresponding cumulative distribution function (CDF) is defined as: $$F\left(x\right) = Prob\left[X \leq x \right] = \int_{-\infty}^x f\left(u\right) \, du.$$ Given a particular realization of the random process $X$: $$X_N = \left\{x_1, \ldots, x_N \right\},$$ ECDF is written as $F_N\left(x\right) = k/N$ where $k$ is the number of observations which is smaller or equal to $x$. Given two data series $X_N$ and $Y_M$, with $N$ and $M$ not necessarily equal, we can test if the two series are two particular realizations of the same random variable by analysis of their ECDFs. Under the hypothesis that the two data series comes from the same distribution function, Kolmogorov has demonstrated that the maximum distance between the two ECDFs has a limiting distribution which is independent from the statistical properties of the corresponding random variable. If the test is performed against a theoretical distribution, the distance is defined as: $$d_K = max\left|F_N\left(x\right) - F\left(x\right)\right|.$$ In such a case $K = N$. In alternative, If the test is performed between two ECDFs, $K = \left(N M\right)/ \left(N + M\right)$ and: $$d_K = max\left|F_N\left(x\right) - F_M\left(x\right)\right|.$$ The test is defined as follows: 1. In the case of the test on a single data series, the null hypothesis is that the data are realizations of a random variable which is distributed according to the given probability distribution. In the case of two data series, the null hypothesis is that the two data series are realizations of the same random variable, which means their ECDFs should tend to the same CDF. The test rejects or accepts the null hypothesis on the basis of the analysis of $d_{nm}$. 2. A significance level $\alpha$ is defined as the probability that the test rejects the null hypothesis when it is indeed true. 3. The test can be formulated in terms of critical values. The critical value for the test is the value of $d_K\left(\alpha\right)$ corresponding to the significance level. Then if $d_K > d_K\left(\alpha\right)$, the null hypothesis is rejected. KS theory was formulated for independent data sets and the available equations for critical values are valid only if this condition is satisfied [@Miller1956]. The test can be formulated also in the presence of data correlation but the distortion to $d_K$ statistic introduced should be taken into account. This is possible if an effective value for $K$ is introduced as $K_{\rm eff} = \beta K$, with $\beta$ a shape parameter depending only on the data correlations and the required significance level. Alternatively, realistic critical values can be calculated with dedicated Monte Carlo (MC) simulations [@Weiss1978]. The advantage of the shape parameter $\beta$ is that it depends only on correlations and the required significance level. Therefore it can be determined for a specific spectrum (e.g., white noise) and shared among a wide family of spectral shapes. Once $\beta$ is known it can be used to calculate critical values for correlated data using standard equations reported in the literature [@Miller1956]. Focusing on the particular problem, we performed a Monte Carlo estimation of $d_K\left(\alpha\right)$ for WOSA spectra representative of LPF. The number of frequency data considered is $N_f = 341$, corresponding $K$ values are $K = N_f$ in the case of the test against a theoretical distribution and $K = N_f/2$ in the case of the test between two ECDFs. The results are summarized in table \[tbl.kscritmc\]. In the same table we report the values of $\beta$ that are required to obtain proper critical values from the standard equations in the case of the test against a theoretical distribution. -------------------------------------------------------------------- -- -- -- -- **$\alpha$ & **$1-\alpha$ & **$K = N_f$ & $\beta$ & **$K = N_f/2$\ $0.32$ & $0.68$ & $0.0643$ & $0.65$ & $0.0723$\ $0.10$ & $0.90$ & $0.0863$ & $0.58$ & $0.0910$\ $0.05$ & $0.95$ & $0.0969$ & $0.55$ & $0.1006$\ $0.01$ & $0.99$ & $0.1214$ & $0.52$ & $0.1191$\ ******** -------------------------------------------------------------------- -- -- -- -- : Table of KS critical values for correlated spectral data. Critical values are calculated with a Monte Carlo simulation for different values of the significance level $\alpha$. Datasets used are representative of the data analyzed in the current paper. We reported the values for testing an ECDF against a theoretical CDF ($K = N_f$) and the values for testing two ECDFs for the same limiting CDF ($K = N_f/2$). We also reported the values of the shape parameter $beta$ that can be used to calculate proper critical values from standard equations in the case of correlated data. $\alpha$ refers to the significance level whereas $1-\alpha$ is the corresponding confidence level for the test. $N_f = 341$. \[tbl.kscritmc\] IR Test {#appendix.IRtest} ======= The statistics of the IR excess noise estimator are numerically estimated with a Monte Carlo simulation on the spectral data used for the present paper. Data series are $24$ h long synthetic reconstructions of the force-per-unit-mass expected on the TMs. WOSA spectra are calculated with $N_s = 4$ averages on $50 \%$ overlapped segments which were windowed with the Blackman-Harris window. Results are reported in table \[tbl.IRconf\]. ------------------------------------------------------ -- -- -- **$\alpha$ & **$1-\alpha$ & **$x_{lw}$ & **$x_{up}$\ $0.32$ & $0.68$ & $325.88$ & $357.41$\ $0.10$ & $0.90$ & $316.33$ & $367.28$\ $0.05$ & $0.95$ & $311.33$ & $372.55$\ $0.01$ & $0.99$ & $300.48$ & $380.69$\ ******** ------------------------------------------------------ -- -- -- : Table of confidence bounds for the IR estimator on correlated spectral data. The values are calculated with a Monte Carlo simulation for different values of the significance level $\alpha$. Data sets used are representative of the data analyzed in the current paper. The number of frequency points is $N_f = 341$, corresponding to the number of available spectral data in the range $\left[0.1, 10\right]$ mHz, for a time series $24$ h long and number of averages for the WOSA estimator $N_s = 4$. \[tbl.IRconf\] Conversion of displacement noise {#appendix.conv2acc} ================================ LPF can be considered as a three body controlled system composed of the two TMs and the spacecraft (SC). Its equations of motion along the measurement axis can be written as: $$\label{eqn:lpfeqmotion1} \begin{split} & m_1 \ddot{x_1} + m_1 \ddot{x_{sc}} + m_1 \omega_1^2 x_1 = f_1 \\ & m_2 \ddot{x_2} + m_2 \ddot{x_{sc}} + m_2 \omega_2^2 x_2 = f_2 + f_{c2} \\ & m_{sc} \ddot{x_{sc}} - m_1 \omega_1^2 x_1 - m_2 \omega_2^2 x_2 = f_{sc} - f_{c2} + f_{csc}\,. \\ \end{split}$$ Here: - $x_1$ and $x_2$ are TMs coordinates along the sensitive axis. They are relative coordinates in the SC reference frame. - $x_{sc}$ is the absolute SC coordinate along the sensitive axis. - $m_1$, $m_2$ and $m_{sc}$ are the masses of the two TMs and of the SC. - $\omega _1^2$ and $\omega _2^2$ are the parasitic stiffnesses coupling the TMs and the SC. The TMs are coupled to the spacecraft through the parasitic stiffness thus producing an oscillator like equation of motion. The spacecraft at the same time experiences reaction forces given by $- m_1 \omega _1^2 x_1$ and $ - m_2 \omega _2^2 x_2$. - $f_1$, $f_2$ and $f_{sc}$ are the forces acting on TMs and SC respectively. - $f_{c2}$ and $f_{csc}$ are control forces on the second TM and the SC respectively. Since $f_{c2}$ is an internal force to the system the SC experiences a reaction force $- f_{c2}$. - Dots over symbols represent time derivatives. In the main LPF science operation mode, one TM (indicated here as $\mathrm{TM}_1$) is in free-fall and provides the reference for the other TM ($\mathrm{TM}_2$) and the SC. In order to avoid unwanted drifting, both $\mathrm{TM}_2$ and the SC are controlled to follow $\mathrm{TM}_1$. It is worth noting that the system is, by construction, symmetric and the role of the two TMs can be inverted. In order to avoid confusion, we indicate with $\mathrm{TM}_1$ the free-fall reference and with $\mathrm{TM}_2$ the actuated TM. Moving to the Laplace domain, substituting for the SC dynamics and substituting for the differential coordinate $x_{\Delta}$ of $\mathrm{TM}_2$ with respect to $\mathrm{TM}_1$, the equation (\[eqn:lpfeqmotion1\]) can be rewritten as: $$\label{eqn:lpfeqmotion2} \begin{split} & s^2 x_1 \omega_1^2 \left(1+\frac{m_1}{m_{sc}}\right) x_1 + \omega_2^2 \frac{m_2}{m_{sc}} x_1 + \omega_2^2 \frac{m_2}{m_{sc}} x_{\Delta} = \\ & = \frac{f_1}{m_1} - \frac{f_{sc}}{m_{sc}} + \frac{1}{m_{sc}} H_2 o_{\Delta} - \frac{1}{m_{sc}} H_{sc} o_1 \\ & s^2 x_{\Delta} + \left(\omega_2^2 - \omega_1^2\right) x_1 + \omega_2^2 x_{\Delta} = \\ & = \frac{f_2}{m_2} - \frac{f_1}{m_1} + \frac{1}{m_2} H_2 o_{\Delta}. \end{split}$$ Here: - $o_1$ and $o_{\Delta}$ are output displacement signals as provided by the interferometer readout system. $o_1$ is the displacement between the SC and the $\mathrm{TM}_1$. $o_{\Delta}$ is the displacement of $\mathrm{TM}_2$ relative to $\mathrm{TM}_1$. - $H_2$ and $H_{sc}$ are transfer functions of the control systems on $\mathrm{TM}_2$ and SC. The force applied by the controllers is calculated on the basis of the output displacement, therefore $f_{c2} = H_2 o_{\Delta}$ and $f_{csc} = H_{sc} o_1$. Calculations can be more easily performed if we introduce a matrix notation: $$\label{eqn:position} \mathbf{x} = \left( \begin{array}{l} x_1 \\ x_{\Delta} \\ \end{array} \right)$$ $$\label{eqn:ifooutput} \mathbf{o} = \left( \begin{array}{l} o_1 \\ o_{\Delta} \\ \end{array} \right)$$ $$\label{eqn:forceinput} \mathbf{f} = \left( \begin{array}{l} f_1 \\ f_2 \\ f_{sc} \\ \end{array} \right)$$ $$\label{eqn:freedynamics} \mathbf{D} = \left( {\begin{array}{*{20}c} {s^2 + \omega _1^2 + \frac{m_1}{m_{sc}}\omega _1^2 + \frac{m_2}{m_{sc}}\omega _2^2 } & {\frac{m_2}{m_{sc}}\omega _2^2} \\ {\omega _2^2 - \omega _1^2 } & {s^2 + \omega _2^2} \\ \end{array}} \right)$$ $$\label{eqn:coupling} \mathbf{G} = \left( {\begin{array}{*{20}c} \frac{1}{m_1} & 0 & -\frac{1}{m_{sc}} \\ -\frac{1}{m_1} & \frac{1}{m_2} & 0 \\ \end{array}} \right)$$ $$\label{eqn:coupling} \mathbf{C} = \left( {\begin{array}{*{20}c} {-\frac{1}{m_{sc}} } & { \frac{1}{m_{SC}} } \\ 0 & {\frac{1}{m_2} } \\ \end{array}} \right)$$ $$\label{eqn:controllers} \mathbf{H} = \left( {\begin{array}{*{20}c} {H_{sc} } & 0 \\ 0 & {H_2 } \\ \end{array}} \right)$$ With the notation introduced, the equation of motion can be rewritten: $$\label{eqn:matrixdyn1} \mathbf{D} \cdot \mathbf{x} = \mathbf{G} \cdot \mathbf{f} + \mathbf{C} \cdot \mathbf{H} \cdot \mathbf{o}.$$ Then we should consider that the output displacement $\mathbf{o}$ corresponds to the measurement of $\mathbf{x}$ provided by the sensing system: $$\label{eqn:radout} \mathbf{o} = \mathbf{S} \cdot \mathbf{x} + \mathbf{o_{rn}}.$$ Here $\mathbf{S}$ is a $2 \times 2$ sensing matrix and $\mathbf{o_{rn}}$ is the readout noise. Substituting equation (\[eqn:radout\]) in equation (\[eqn:matrixdyn1\]), the dynamics can be rewritten in terms of the known output $\mathbf{o}$: $$\label{eqn:matrixdyn2} \mathbf{D} \cdot \mathbf{S}^{-1} \cdot \mathbf{o} - \mathbf{C} \cdot \mathbf{H} \cdot \mathbf{o} = \mathbf{G} \cdot \mathbf{f} + \mathbf{D} \cdot \mathbf{S}^{-1} \cdot \mathbf{o_{rn}}.$$ The quantity $\mathbf{G} \cdot \mathbf{f}$ represents the force-per-unit-mass acting on the test masses. Such a quantity is not directly known from the system output since the available signal is $\mathbf{o}$. From equation (\[eqn:matrixdyn2\]) it is readily seen that an effective force-per-unit-mass acting on the TMs can be reconstructed from the knowledge of $\mathbf{o}$ if the force applied by the control system ($\mathbf{C} \cdot \mathbf{H} \cdot \mathbf{o}$) is calculated and subtracted from the reconstructed dynamics $\mathbf{D} \cdot \mathbf{S}^{-1} \cdot \mathbf{o}$. Therefore the available quantity is $\mathbf{g} = \mathbf{G} \cdot \mathbf{f} + \mathbf{D} \cdot \mathbf{S}^{-1} \cdot \mathbf{o_{rn}}$ which contains the force-per-unit-mass acting on the TMs corrupted by the readout noise of the system. The expected values for $\mathbf{g}$ can be obtained thanks to the dynamics reported in equation (\[eqn:matrixdyn2\]) once the expected values for $\mathbf{f}$ and $\mathbf{o_{rn}}$ are known. The same applies to the projection of the noise spectra for $\mathbf{f}$ and $\mathbf{o_{rn}}$ to $\mathbf{g}$. Huber’s norm {#appendix.huber} ============ Huber’s norm [@Huber1973] is a way to construct a goodness of fit estimator which is more robust to outliers and non-Gausianity than the standard mean squared error (MSE). The norm is constructed as $\sum_i \rho(r_i)$, where $r_i$ are the residuals between a data series and a parametric model. The function $\rho(r_i)$ s defined as: $$\rho(r_i) = \begin{cases} \frac{1}{2} r_i^2 & \text{for } |r_i| < c \\ c |r_i| - \frac{1}{2} c^2 & \text{for } |r_i| \geqslant c. \end{cases}$$ The value of the threshold constant $c$ may depend on the given dataset, therefore some efforts must be spent to select the value of $c$ providing the most accurate results. Normalized residuals are defined as $r_i = \left(P_{WOSA}\left(f_i\right)-S_a\left(f_i;\theta\right)\right) / S_a\left(f_i;\theta\right)$ in accordance to the MSE definition in section \[subsect.test.noisemodelident\]. With such a definition, residuals are expected to be zero mean and unitary variance in correspondence of the ’true’ model $S_a\left(f_i;\theta\right)$. Different values of $c$ were tested with a Monte Carlo estimation on the first $1000$ data of the analysis reported in section \[subsect.test.noisemodelident\]. Huber’s norm is minimized for each MC iteration and the corresponding values of the parameter $\theta$ are stored. The histogram of $\theta$ shows a mean steadily fixed on $0.13$ for values of $c$ ranging from $0.001$ to $0.1$. Increasing the value of $c$, the distribution starts to shift toward the distribution obtained with MSE minimization. MSE and Huber distributions are practically undistinguishable for values $c > 2$. Since the ’true’ value for the parameter $\theta$ is set to $0.1$, Huber’s norm is performing better for values $c < 0.1$. It was then decided to use the value $c = 0.05$ for the final analysis reported in figure \[fig:fig5\]. References {#references .unnumbered} ========== [10]{} F. Antonucci et al., Class. Quantum Grav. [**28**]{}, 094002 (2011). F. Antonucci et al., Class. Quantum Grav. [**28**]{}, 094006 (2011). F. Antonucci et al., Class. Quantum Grav. [**28**]{}, 094001 (2011). M. Armano et al., Class. Quantum Grav. [**26**]{}, 094001 (2009). LISA International Science Team, LISA Unveiling a hidden Universe, ESA/SRE(2011)3, February 2011. J. Crowder and N. J. Cornish, Phys. Rev. D [**72**]{}, 083005 (2005). Kawamura S et al, Class. Quantum Grav. [**23**]{} S125 (2006). Ni W., Shiomi S. and Liao A., Class. Quantum Grav. [**21**]{} S641 (2004). D. B. Percival and A. T. Walden [*Spectral Analysis for Physical Applications*]{} (Cambridge University Press, Cambridge, 1993) p 290. M. Hewitson et al., Class. Quantum Grav. [**26**]{}, 094003 (2009). MATLAB http://www.mathworks.com LTPDA: a MATLAB toolbox for accountable and reproducible data analysis http://www.lisa.aei-hannover.de/ltpda. A. N. Kolmogorov, Ist. Ital. Attuari [**4**]{}, 83 (1933). W. Feller, Ann. Math. Stat. [**19**]{}, 177 (1948). N. Smirnov, Ann. Math. Stat. [**19**]{}, 279 (1948). L. H. Miller, J. Amer. Statistical Assoc. [**51**]{}, 273 (1956). M. Abramowitz and I. A. Stegun [*Handbook of mathematical functions*]{} (Dover Publications, New York, 1972) p 260. A. Monsky et al., Class. Quantum Grav. [**26**]{}, 094004 (2009). L. Ferraioli, M. Hueller and S. Vitale, Class. Quantum Grav. [**26**]{}, 094013 (2009). A. V. Oppenheim and R. W. Schafer [*Discrete-Time Signal Processing*]{} (Prentice-Hall, Upper Saddle River, New Jersey, USA) p 714. S. Mallat [*A Wavelet Tour of Signal Processing*]{} (Academic Press, San Diego, 1998). J. A. Peacock, Mon. Not. R. astr. Soc. [**202**]{}, 615 (1983). G. Fasano and A Franceschini, Mon. Not. R. astr. Soc. [**225**]{}, 155 (1987). L. Ferraioli et. al., Phys. Rev. D [**82**]{}, 042001 (2010). D. J. Thomson, Bell Sys. Tech. Journal [**56**]{}, 1769 (1977). F. J. Harris, Proc. IEEE [**66**]{}, 51 (1978). N. I. Fisher, Int. Stat. rev. [**51**]{}, 25 (1983). M. B. Wilk and R. Gnanadesikan, Biometrika [**55**]{}, 1 (1968). M. S. Weiss, J. Amer. Statistical Assoc. [**73**]{}, 872 (1978). P. J. Huber, Ann. Stat. [**1**]{}, 799 (1973). [^1]: As formulated by equation (\[eqn.dKdefemp\]), the test is searching for differences in the noise content of the data series. It does not provide information about which series has the largest noise content. Such information can be recovered graphically with a distribution plot (for example, a quantile-quantile plot). Alternatively, a pure excess noise test for a data series, $x_m$, with respect to a reference series, $x_n$, can be formulated by substituting $d_K$ with $d^{+}_K = max\left\{F_{n_2}\left(x\right) - F_{n_1}\left(x\right)\right\}$. The formalism used for the calculation of $d_K$ and $d^{+}_K$ is the same, therefore the procedures described in this paper can be applied in either case without any significant modification. [^2]: The distribution of the WOSA estimator for the sample spectrum is a gamma distribution, as reported in equation (\[eqn.probdistwosa\]). Such a distribution tends to a Gaussian distribution as the number of averages increases. The difference between the ‘true’ distribution and the corresponding Gaussian can be quantified by the maximum distance among cumulative distribution functions, $d$. As an example $d = 0.067$ for $N_s = 4$, $d = 0.03$ for $N_s = 16$, $d = 0.019$ for $N_s = 50$, $d = 0.013$ for $N_s = 100$ and $d = 0.0094$ for $N_s = 200$. It is worth noting that since $d$ is a difference in the space of the values of the CDF, it can assume values in $\left[0,1\right]$. As can be seen, the distribution of the WOSA spectrum reasonably approaches a Gaussian only for $N_s$ as large as $200$. [^3]: The term [*pseudo-stationary*]{} indicates a data series which is affected by slight non-stationarity of the kind that can be removed with mean subtraction or standard polynomial fit trend removal. [^4]: If $N_s$ is the number of averaging segments, $N_f$ the length of each segment and $h$ the shift factor, which is indicating the number of data points two consecutive data segments are shifted by, then the correction to the variance of the averaged process is proportional to [@Percival] $\Gamma\left(h\right) = \sum_{j=1}^{N_s-1}\sum_{t=1}^{N_f} h_t h_{t}^{j h}$. Here $h_{t}^{j h}$ are window elements shifted by a factor $j h$. With our choice of $N_s = 4$, a segment overlap of $50 \%$ and a total data length of $24$ h, we find $\Gamma\left(h\right) = 2.7 \times 10^{-4}$. [^5]: It is easy to verify that, in the case of $z = P\left(f_k\right)$ and $\lambda = E\left[P\left(f_k\right)\right]/\nu$, the probability density function $f\left(z;\nu,\lambda\right) = \frac{e^{-\frac{z}{2\lambda}} z^{\left(\frac{\nu}{2}-1\right)}}{\left(2\lambda\right)^{\frac{\nu}{2}}\Gamma\left(\frac{\nu}{2}\right)}$ is correctly represents the statistic of the sample spectrum. $E\left[z\right] = \frac{\nu}{2}2\lambda = E\left[P\left(f_k\right)\right]$ as expected for the spectrum [@Percival]. ${\rm var}\left[z\right] = \frac{\nu}{2} \left(2 \lambda\right)^2 = \frac{2}{\nu} E\left[P\left(f_k\right)\right] = E\left[P\left(f_k\right)\right]$ since $\nu = 2$ in the present case.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A spectral mixture (SM) kernel is a flexible kernel used to model any stationary covariance function. Although it is useful in modeling data, the learning of the SM kernel is generally difficult because optimizing a large number of parameters for the SM kernel typically induces an over-fitting, particularly when a gradient-based optimization is used. Also, a longer training time is required. To improve the training, we propose an approximate Bayesian inference for the SM kernel. Specifically, we employ the variational distribution of the spectral points to approximate SM kernel with a random Fourier feature. We optimize the variational parameters by applying a sampling-based variational inference to the derived evidence lower bound (ELBO) estimator constructed from the approximate kernel. To improve the inference, we further propose two additional strategies: (1) a sampling strategy of spectral points to estimate the ELBO estimator reliably and thus its associated gradient, and (2) an approximate natural gradient to accelerate the convergence of the parameters. The proposed inference combined with two strategies accelerates the convergence of the parameters and leads to better optimal parameters.' author: - | Yohan Jung, Kyungwoo Song, Jinkyoo Park\ Department of Industrial $\&$ Systems Engineering\ KAIST\ Daejeon, Republic of Korea\ ` \{\text{becre1776,gtshs2,jinkyoo.park}\}@kaist.ac.kr`\ bibliography: - 'neurips\_citation.bib' title: Approximate Inference for Spectral Mixture Kernel --- Introduction ============ In constructing a Gaussian process $(\mathcal{GP})$ model, selecting a proper kernel function is vital because the selected kernel determines the overall structure of the target function by specifying the covariance of $\mathcal{GP}$, a prior for the target function. Inspired by Bochner’s theorem [@bochner1959lectures] in which the spectral density has Fourier duality relationship with the stationary kernel, Wilson et al. [@wilson2013gaussian] modeled the spectral density of the kernel using a mixture of the Gaussian distribution and obtained a SM kernel by applying a Fourier transform to the modeled spectral density in an attempt to design a flexible kernel. This SM kernel is flexible enough to approximate any stationary kernel because the mixture of the Gaussian distribution can approximate well any spectral density of a stationary kernel [@kostantinos2000gaussian]. The SM kernel has motivated many researchers to devise more expressive kernels based on the spectral density, i.e., a Fourier duality of the kernel. Ulrich et al. [@ulrich2015gp] and Parra et al. [@parra2017spectral] proposed a cross-spectral mixture (CSM) kernel and a multi-output spectral mixture (MOSM) for a multi-output $\mathcal{GP}$ by modeling the cross-spectral density between the stochastic processes. Remes et al. [@remes2017non] propose a non-stationary spectral kernel by modeling the input-dependent spectral density. Despite the expressive power of the SM kernel and its variants, employing such kernels for a $\mathcal{GP}$ model is limited because they use numerous hyperparameters for the flexible spectral density modeling, increasing the difficulty of the training. Specifically, the training of many of the kernel hyperparameters is prone to an over-fitting, particularly when gradient-based optimization is employed [@warnes1987problems]. Moreover, computing the operations needed to update the hyperparameters requires additional time [@rasmussen2004gaussian], which prevents the SM kernel from being applied to the modeling of large-scale data. To tackle these issues, we propose an approximate Bayesian inference method for an SM kernel. Based on the intuition that learning of a SM kernel is equivalent to the learning of its spectral density distribution, we find the variational posterior distribution of the spectral density by employing the stochastic gradient variational bayes (SGVB) [@kingma2013auto], which is a sampling-based variational inference. To this end, we employ a variational distribution of the spectral points to approximate the spectral density of the SM kernel using sampled spectral points. To learn variational parameters of the spectral points, we first derive the regularized evidence lower bound (ELBO) estimator computed efficiently by the sampled spectral points. We then optimize the ELBO estimator with respect to the variational and other parameters by employing a sampling-based optimization through a reparameterization trick. To improve the inference, we propose two additional strategies: (1) a sampling strategy of spectral points used to reliably estimate the ELBO and its associated gradient, (2) a natural gradient that reflects the geometric information of the probability density to accelerate the convergence of the parameters. To validate the proposed methods, we run several experiments on kernel matrix approximation, an ablation study of the approximate inference, and a regression task using for large-scale datasets. Preliminaries ============= Spectral Mixture (SM) Kernel ---------------------------- We describe how the SM kernel is defined by the spectral density modeling. Bochner’s theorem states that stationary kernel $k(x_1-x_2)$ can be defined as the Fourier transform of spectral density $p(S)$ as $$\begin{aligned} & k(x_1-x_2) = \int e^{2\pi i S^{\mathrm{T}} {(x_1-x_2)} } p(S) dS \end{aligned}$$ for the inputs $x_{1},x_{2} \in R^{D}$. Wilson [@wilson2013gaussian] devised a SM kernel by representing the spectral density $p(S)$ using a $Q$ mixture of symmetric Gaussian distribution $p(S) = \sum_{q=1}^{Q} w_{q} \left( \frac{ N(s|\mu_{q},\sigma^{2}_{q}) + N(-s|\mu_{q},\sigma^{2}_{q}) }{2} \right)$ with the mean $\mu_{q} = [ \mu_{(q,1)},..,\mu_{(q,D)} ] \in R^{D} $ and the variance $\sigma^{2}_{q} = [ \sigma^{2}_{(q,1)},..,\sigma^{2}_{(q,D)} ] \in R^{D} $, and by then applying the Fourier Transform to $p(S)$ using Eq. (1). The derived SM kernel is expressed as $$\begin{aligned} & k_{SM}(x_1-x_2) = \sum_{q=1}^{Q} w_{q} \mathrm{exp} \left( -2{\pi}^2 \left( \sigma_{q}^{T} (x_1-x_2) \right)^2 \right) \mathrm{cos} \left(2\pi {\mu_{q}}^T (x_1-x_2) \right)\end{aligned}$$ Random Fourier Feature (RFF) ----------------------------- A random Fourier feature [@rahimi2008random] approximates the stationary kernel $k(x_{1}-x_{2})$ by applying a Monte Carlo integration to Eq. (1) with $M$ spectral points $\textbf{\em s} = \{ \textbf{\em s}_{i}\}_{i=1}^{M}$ sampled from $p(S)$ $$\begin{aligned} k(x_{1}-x_{2}) & \approx \frac{1}{M} \sum_{i=1}^{M} \cos({2\pi \textbf{\em s}_i}^{T}x_1) \cos({2\pi \textbf{\em s}_i}^{T}x_2) + \sin({2\pi \textbf{\em s}_i}^{T}x_1)\sin({2\pi \textbf{\em s}_i}^{T}x_2) \end{aligned}$$ If we let the feature map $\phi(x,{\textbf{\em s}}) = \frac{1}{\sqrt{M}} \left[\cos{(2\pi{\textbf{\em s}^{T}_{1}}x)},\sin{(2\pi{\textbf{\em s}^{T}_{1}}x)},..,\cos{(2\pi{\textbf{\em s}^{T}_{M}}x)},\sin{(2\pi{\textbf{\em s}^{T}_{M}}x)} \right] \in R^{1\times2M}$, Eq. (3) can be represented as $k(x_{1}-x_{2}) \approx \phi(x_{1},\textbf{\em s})\phi(x_{2},\textbf{\em s})^{T}$. Sparse Spectrum $\mathcal{GP}$ ------------------------------ For the scalable learning of a large dataset, sparse spectrum $\mathcal{GP}$ [@lazaro2010sparse] employs an approximate kernel obtained by the RFF. Let $\textbf{\em s} = \{ \textbf{\em s}_{i}\}_{i=1}^{M}$ be the spectral points used to approximate the stationary kernel by Eq. (3). In addition, let $f$ be a function, with $\mathcal{GP}$ prior, modeling the relation between $X=\{x_{1},..,x_{N}\}$ and $Y=\{y_{1},..,y_{N}\}$. Then, the prior distribution of $f(X) = [f(x_1),..,f(x_N)]$ using the approximate kernel can be defined as $$\begin{aligned} p\left(f(X)\right) = N\left( f(X);0,\Phi_{\textbf{\em s}}(X)\Phi_{\textbf{\em s}}(X)^{T} \right)\end{aligned}$$ where $\Phi_{\textbf{\em s}}(X) = \left[ \phi(x_{1},\textbf{\em s});...;\phi(x_{N},\textbf{\em s}) \right] \in R^{N \times 2M}$. If the likelihood is assumed as a Gaussian distribution i.e, $p(Y|f(X))=N \left( Y|f(X),{\sigma}^{2}_{\epsilon} I \right)$, the conditional marginal likelihood of the given spectral points $p(Y|X,\textbf{\em s})$ is computed as $N(Y;0,\Phi_{\textbf{\em s}}(X)\Phi_{\textbf{\em s}}(X)^{T} +{{\sigma}^{2}_{\epsilon}}I)$. To find the optimal spectral points $\textbf{\em s}$ that explain the data well, $-\log{p(Y|X,\textbf{\em s})}$ is minimized with respect to $\textbf{\em s}$. Evaluating $-\log{p(Y|X,\textbf{\em s})}$ uses the memories $O(NM)$ and takes computation time $O(NM^2)$ for computing the inversion and determinant by inversion lemma. When $M$ is much smaller than $N$, training $\mathcal{GP}$ model using $-\log{p(Y|X,\textbf{\em s})}$ takes less training time because the original $\mathcal{GP}$ model using the exact kernel uses the memories $O(N^{2})$ and takes the computation time $O(N^{3})$ [@rasmussen2004gaussian]. Natural Gradient Optimization (NGO) for the Probability Density Parameter ------------------------------------------------------------------------- It is known that natural gradient can be used in efficiently optimizing the probability density parameter [@amari1998natural]. We introduce the natural gradient to be used for our approximate inference. Mathematically, given the loss $\mathcal{L}(\theta)$ parameterized by the parameter $\theta$ of the probability density $p_{\theta}(z)$, for a small $\epsilon>0$, the natural gradient $\Tilde{\nabla}_{\theta} {\mathcal{L}}(\theta)$ can be defined as $$\begin{aligned} \Tilde{\nabla}_{\theta} {\mathcal{L}}(\theta) = \operatorname*{argmin}_{\substack{ \{\Delta\theta ;KL( p_{\theta} \Vert p_{\theta + \Delta\theta} ) = \epsilon \} }} \mathcal{L}(\theta + \Delta\theta)\end{aligned}$$ Based on the definition, the natural gradient is consistently defined regardless of the parameterization of the probability density. Because the natural gradient updates robustly the probability density parameters due to its consistent characteristics, it can accelerate the convergence for the inference. Approximate Inference for SM kernel =================================== Based on the intuition that the learning of a SM kernel is equivalent to the learning of its spectral density distribution, in this study, we seek to find the variational posterior distribution of the spectral density by employing the stochastic gradient variational bayes (SGVB) [@kingma2013auto]. Specifically, this section discusses 1) how to approximate the spectral density of the SM kernel by spectral points sampled from the variational distribution of the spectral points and how to construct the ELBO estimator using these sampled spectral points, 2) how to effectively sample the spectral points from the variational distribution to robustly compute the ELBO estimator, and 3) how to update the variational parameters by using the approximate natural gradient of the ELBO estimator. Regularized Lower Bound Estimator for Variational Sparse Spectrum Approximation ------------------------------------------------------------------------------- First, we assume the variational distribution of spectral points $S=\cup_{q=1}^{Q}\{s_{q,1},..,s_{q,m_{q}}\}$ as $q(S) = \prod_{q=1}^{Q}\prod_{i=1}^{m_q} N(s_{q,i} ;{\mu}_{q},{\sigma^{2}_{q}})$ where $m_q$ is the number of spectral points drawn from the $q$-th spectral density component such that the constructed random kernel can approximate the SM kernel with the hyperparameters $\{w_{q},\mu_{q},\sigma^{2}_{q}\}_{q=1}^{Q}$. If we sample the spectral points $\textit{\textbf{s}}_{q,i}$ from $N(s_{q,i} ;{\mu}_{q},{\sigma^{2}_{q}})$ as $\textit{\textbf{s}}_{q,i}=\mu_{q} + \sigma_{q} \circ \epsilon_i $ with $\epsilon_i \sim N(\epsilon;0,I)$ by the reparameterization trick, we can define the random feature map $\phi_{\mathrm{SM}}(x;\textit{\textbf{s}})$ with the sampled spectral points $\textit{\textbf{s}}=\cup_{q=1}^{Q}\{\textit{\textbf{s}}_{q,1},..,\textit{\textbf{s}}_{q,m_{q}}\}$ $$\begin{aligned} \phi_{\mathrm{SM}}(x;\textit{\textbf{s}}) = \Big[ \sqrt{w_{1}} \phi \left(x, \{\textit{\textbf{s}}_{1,i} \}_{i=1}^{m_1}\right), ... ,\sqrt{w_{Q}} \phi \left(x,\{\textit{\textbf{s}}_{Q,i}\}_{i=1}^{m_Q} \right) \Big] \in R^{1 \times 2M}\end{aligned}$$ where $\phi$ is the defined feature map used to calculate Eq. (3) and $M=\sum_{q=1}^{Q} m_q$. By employing this feature map $\phi_{\mathrm{SM}}(x;\textit{\textbf{s}})$ for the entire dataset $X=\{x_{1},..,x_{N}\}$, we can define the feature matrix as $\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)=[\phi_{\mathrm{SM}}(x_1;\textit{\textbf{s}});...;\phi_{\mathrm{SM}}(x_N;\textit{\textbf{s}})]\in R^{N\times 2M}$. The feature matrix can then be used to construct the unbiased estimator $\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T}$ to satisfy $\mathrm{E}\big[\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T} \big] = K_{\mathrm{SM}}(X,X)$. Proposition 1 states the error bound of this estimator. Let us denote $W_{0}={ \big( \sum_{q=1}^{Q} w^2_{q} \big) }^{1/2}$ and $m_0=\mathrm{min} \{m_1,..,m_Q\}$. Then, for a small $\epsilon >0$, the error bound of $\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T}$ using the matrix spectral norm ${\Vert \cdot \Vert}_{2}$ is obtained as $$\begin{aligned} \mathrm{Pr} \Big( {\big \Vert \Phi^{\mathrm{SM}}_{\textbf{s}}(X)\Phi^{\mathrm{SM}}_{\textbf{s}}(X)^{T} - K_{\mathrm{SM}}(X,X) \big \Vert }_{2} \geq \epsilon \Big) \leq N \exp{ \Big( \frac{ -3{\epsilon}^2 m_0}{ W_{0}N \big( 6{ \Vert K_{\mathrm{SM}}(X,X) \Vert}_{2} + 4\epsilon \big) } \Big) }\end{aligned}$$ Using the approximate SM kernel $\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T}$ with the equal number of spectral points $m_q=m$ for $q\in\{1,..,Q\}$, we can derive the regularized lower bound estimator $\hat{\mathcal{L}}_{N}$ as $$\begin{aligned} \log{p(Y|X)} &\geq \int \log{p(Y|X,S)}q(S) d S - KL(q(S)||p(S)) = \mathcal{L} \nonumber \\ &\approx \frac{1}{N}\sum_{n=1}^{N} \log {p(Y|X,\textbf{\em s}^{(n)})} - KL(q(S)||p(S)) = \hat{\mathcal{L}}_{N} \end{aligned}$$ where $\textbf{\em s}^{(n)}$ indicates the $n$-th set of the spectral points sampled from $q(S)$ and $\log {p(Y|X,\textbf{\em s}^{(n)})}$ is evaluated as $\log{N(Y;0,\Phi^{\mathrm{SM}}_{\textbf{\em s}^{(n)}}(X) \Phi^{\mathrm{SM}}_{ \textbf{\em s}^{(n)}}(X)^{T} + {\sigma}^{2}_{\epsilon}I)}$. $p(S)$ is a prior distribution of spectral density expressed as $\prod_{q=1}^{Q}\prod_{i=1}^{m_q} N(s_{q,i} ;\Tilde{{\mu}}_{q,i},{\Tilde{\sigma}^{2}_{q,i}})$ where $\Tilde{{\mu}}_{q,i}$ and ${\Tilde{\sigma}^{2}_{q,i}}$ are initialized based on the prior knowledge. The term $KL\left(q(S)||p(S)\right)$ prevents the model from being over-fitted to the training data. To update the parameters, we evaluate $\hat{\mathcal{L}}_{N}$ and compute its gradient with respect to the variational parameters of the spectral points $\{\mu_{q},\sigma^{2}_{q}\}_{q=1}^{Q}$, weight parameters $\{w_{q}\}_{q=1}^{Q}$, and noise parameter ${\sigma}^{2}_{\epsilon}$. We cast the problem of learning the SM kernel into the problem of estimating the variational parameters of the spectral points. The described procedure, denoted as SVSS, is our basic approximate inference method. In the following two subsections, we propose strategies to improve the SVSS. Efficient Sampling Strategy for Spectral Points ----------------------------------------------- As shown in Eq. (7), the value of $\hat{\mathcal{L}}_{N}$ might fluctuate depending on the sampled spectral point $\textit{\textbf{s}} \sim q(S)$ used to construct $\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T}$, especially when the number of total spectral points $M$ is small. Because the large volatility of $\hat{\mathcal{L}}_{N}$ will likely cause instability in learning, we propose a sampling strategy for spectral points to minimize such volatility. Specifically, we find the optimal ratios of the spectral points $\{m_q/\sum_{q=1}^{Q} m_q\}_{q=1}^{Q}$ for the spectral points $\textit{\textbf{s}}$ sampled from $q(S)$ such that the sampled spectral points minimize the sum of the variance for each element, i.e., $\sum_{i,j=1}^{N} \mathrm{Var} \big( [\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T}]_{i,j} \big)$. We consider only the upper off-diagonal entries of $\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)\Phi^{\mathrm{SM}}_{\textit{\textbf{s}}}(X)^{T}$ because this matrix is symmetric and diagonal terms are deterministic. Proposition 2 states the optimization problem obtaining the optimal ratios of the spectral points. Given the set of inputs $X=\{x_n\}_{n=1}^{N}$, let us define the set of pairwise distances $\{\tau_p\}_{p=1}^{P}$ where $\tau_p=|x_i-x_j|$ for some $i,j \in \{1,..,N\}$ and $i\neq j$. Let $m_{q}$ be the number of spectral points sampled from the variational distribution $N(\mu_q,\sigma^{2}_{q})$, and $M = \sum_{q=1}^{Q} m_{q} $ be the total number of spectral points. For the set of $\{\tau_{p}\}_{p=1}^{P}$, the optimal ratio ${p^{*}_q}=m^{*}_q/M$ to minimize $ \sum_{i<j} \mathrm{Var} \big( [\Phi^{\mathrm{SM}}_{\textbf{s}}(X)\Phi^{\mathrm{SM}}_{\textbf{s}}(X)^{T}]_{i,j} \big)$ is obtained as $$\begin{aligned} p^{*}_q = \frac{ w_{q} { \left[ \sum_{p=1}^{P} g_{q}(\tau_{p}) \right] }^{1/2} } { \sum_{q=1}^{Q} w_{q} {\left[ \sum_{p=1}^{P} g_{q}(\tau_{p}) \right] }^{1/2} }\end{aligned}$$ where $g_{q}(\tau) = 1 + k_{q}(2\tau) + k^{2}_{q}(\tau)$ and $k_{q}(\tau) = \exp{ \left ( -2{\pi}^2 (\tau^{T} \sigma_{q})^2 \right) } \cos{ \left( 2\pi \mu^{T}_{q}\tau \right) }$. The integer $m^{*}_{q}$ is determined as $\mathrm{max}\{1,{\lfloor Mp^{*}_q \rceil}\}$ where ${\lfloor m \rceil}$ denotes the integer closest to $m$. Proposition 2 states that the optimal numbers for the sampled spectral points $\textit{\textbf{s}}=\cup_{q=1}^{Q}\{\textit{\textbf{s}}_{q,1},..,\textit{\textbf{s}}_{q,m_{q}}\}$ are determined depending on the variational parameters, weight parameters, and inputs. When we apply this sampling strategy to the SVSS for a large dataset, the total number of the pairwise distances $P=N(N-1)/2$ is too large to be efficiently computed for every iteration. Specifically, we restrict $P$ by selecting a small subset of data. For every iteration, we randomly select the subset $\{\tau_{p_i}\}_{i=1}^{Pr} \subset \{\tau_p\}_{p=1}^{P}$ with the rate $r\in(0,1)$ used for calculating $p^{*}_{q}$. This random sampling will be validated in the experiment. Approximate Natural Gradient Update for Variational Parameters -------------------------------------------------------------- Once the ELBO estimator $\hat{\mathcal{L}}_{N}$ is computed, its stochastic gradients with respect to variational parameters of $q(S)$, as well as to other parameters, are computed and used to update the parameters. To accelerate the convergence of the parameters, we propose using the approximate natural gradient easily computed from the original gradient. We consider that variational parameters $\mu_q$ and $\sigma_q$ for $q\in\{1,..,Q\}$ are updated in the logarithm domain because these parameters should remain positive. Let $\mu^{(t)}_q$ and $\sigma^{(t)}_q$ be the $t$-th iterated parameters of $N(\mu_q,\sigma^2_q)$ which is $q$-th component distribution for $q(S)$. The natural gradient of $\hat{\mathcal{L}}_{N}$ w.r.t $\mu_q$ and $\sigma_q$ on log domain, i.e. $\Tilde{\nabla}_{\log{\mu_q}}$ and $\Tilde{\nabla}_{\log{\sigma_q}}$, can be approximated as $$\begin{aligned} \Tilde{\nabla}_{\log{\mu_q}} \hat{\mathcal{L}}_{N} &\approx \left( \frac{\sigma^{(t+1)}_q} {\mu^{(t)}_q} \right)^{2} \circ {\nabla}_{\log{\mu_q}} \hat{\mathcal{L}}_{N} \hspace{3 em} \Tilde{\nabla}_{\log{\sigma_q}} \hat{\mathcal{L}}_{N} \approx \frac{1}{2} \ \nabla_{\log{\sigma_q}} \hat{\mathcal{L}}_{N}\end{aligned}$$ under the condition $\Big| \left( \frac{\sigma^{(t+1)}_q} {\mu^{(t)}_q} \right)^{2} \circ {\nabla}_{\log{\mu_q}} \hat{\mathcal{L}}_{N} \Big| < 1$ and $ \Big| \nabla_{\log{\sigma_q}} \hat{\mathcal{L}}_{N} \Big| <1$ in element-wise sense. These constraints are satisfied by normalizing the revised gradient by its 2-norm ${\Vert \cdot \Vert}_{2}$. The derived gradients are used to update the parameters using the optimizer with the adaptive learning rate. Related Work ============ To train scalably the parameters of the kernel for large-scale data, variational inducing input method [@titsias2009variational; @hensman2013gaussian; @hensman2015scalable] and sparse spectrum method [@lazaro2010sparse; @gal2015improving; @hoang2017generalized] have been proposed. The variational inducing input (VFE) method introduces a small number of inducing inputs such that the variational distribution of the function-values on the inducing inputs best approximates the prior distribution of the function-values on all inputs. The sparse spectrum method employs the approximate kernel matrix obtained by a random Fourier feature (RFF) [@rahimi2008random] to represent a $\mathcal{GP}$ prior. Among the sparse spectrum methods, SSGP [@lazaro2010sparse] optimizes the spectral points used for constructing the approximate kernel. To relax the over-fitting issue of SSGP, VSSGP [@gal2015improving] applies a variational approximation to the spectral points. VSSGP independently estimates the hyperparameters of the kernel and variational parameters. However, the variational approximation of spectral points can directly approximate the spectral density of kernel through the designed feature map while relaxing the over-fitting. Being different from VSSGP, our approach train the parameters of the SM kernel by inferencing the variational distribution of spectral points. In addition, we consider the sampling strategy and approximate natural gradient to improve the performance of the approximate inference. Because of these differences, our method efficiently train the SM kernel while employing a smaller number of parameters. Experiments =========== In the first experiment, we show that our sampling strategy for spectral points helps approximate the SM kernel more accurately. In the second experiment, we validate that the proposed estimator, sampling strategy, and approximate natural gradient improve the inference for the SM kernel through an ablation study. In the last experiment, we conduct a regression task on a large-size UCI dataset [@Dua:2019] to evaluate the performance of the proposed approximate inference. We also include the results of additional experiments in the supplementary material. SM Kernel Gram Matrix Approximation ----------------------------------- We verify that the proposed sampling strategy enhances the SM kernel approximation over naive sampling approaches with equal sampling and naive weight sampling of the spectral points using $p_{q}=1/Q$ and $p_{q}=w_q / \sum_{q=1}^{Q} w_q$, respectively. Also, we show that even if a small number of inputs is used to calculate $\{p^{*}_{q}\}_{q=1}^{Q}$ by Eq. (8), our strategy can maintain the quality of the kernel approximation. To validate the sampling’s effect in a general setting, we consider the different cases of the number of total spectral points $M$ and the number of mixture components $Q$ for the SM kernel. First, we approximate the small-sized SM kernel $K\in R^{N \times N}$ with $N=100$. We use $X= \{0,.01,..,.99\}$ as the kernel inputs. For the hyperparameters, we initialize the weights of SM kernel in two ways by applying a uniform distribution $U$; $\{w_{q}\}_{q=1}^{Q} \sim U(0,20)$ and $\{w_{q}\}_{q=1}^{Q} \sim U(0.99,1.01)$. We consider these cases to show that the proposed sampling strategy approximates the SM kernel well regardless of its weight parameters. To investigate how the amount of the data used to compute $p^{*}_{q}$ affects the kernel approximation, we use only $r$ fraction among $P=N(N-1)/2$ pair-wise input distances with $r\in\{.05,.1,1\}$. For example, $r=1$ denotes that all data are used for calculating $p^{*}_{q}$. We evaluate the approximation quality based on the relative error ${\Vert K-\hat{K} \Vert}_{F} / {\Vert K \Vert}_{F}$ using the Frobenius norm $F$. Figures 2(a) and 2(b) describe the results of $\{w_{q}\}_{q=1}^{Q} \sim U(0,20)$ and $\{w_{q}\}_{q=1}^{Q} \sim U(0.99,1.01)$. In Figure 2(a), the proposed method approximates the kernel matrix most accurately, followed by naive weight sampling and equal sampling. By comparing the results of $r=1.0$ and $r=0.05$ in both graphs of (a), we can conclude that using a small fraction of the inputs for calculating $p^{*}_{q}$ does not degrade the performance under various conditions for $M$ and $Q$. In Figure 2(b), similarly, the proposed method outperforms both the naive weight sampling and equal sampling at different values of $M$ and $Q$. We believe that the improved performance is due to the optimal ratio $p^{*}_{q}$ which is computed by the variational parameters $\{\mu_q,\sigma^{2}_{q}\}_{q=1}^{Q}$ and inputs $\{\tau_p\}_{p=1}^{P}$ in addition to $\{w_{q}\}_{q=1}^{Q}$. Additionally, we apply our method on a large-scale high-dimensional real dataset: the Concrete dataset $(N=1,030 \ D=8)$ in Figure 3(a) and the Parkinsons dataset $(N=5,875 \ D=20)$ in Figure 3(b). Because $N$ is large, we consider the restricted value of $P=1,000,000$ and data rate of $r\in\{0.05,0.1\}$. We can see that our sampling strategy is the most effective at approximating the SM kernel in the real dataset. Approximate Inference for SM kernel ------------------------------------ We investigate the effectiveness of the proposed inference methods composed of the regularized lower bound estimator $\hat{\mathcal{L}}_{N}$, the sampling strategy of proposition 2, and the approximate natural gradient of proposition 3 by conducting an ablation study comparing the following inference methods: - SS: denotes the variation of sparse spectrum $\mathcal{GP}$ [@lazaro2010sparse] used to train the SM kernel approximated by Eq. (6). SS optimizes $\{w_{q},\mu_{q},\sigma^{2}_{q}\}_{q=1}^{Q}$ with the given $\{\epsilon_{i}\}_{i=1}^{QM} \sim N(0,I)$ without a reparameterization trick (RP). This approach is similar to the SM kernel inference in A la Carte [@yang2015carte] without applying the fastfood [@le2013fastfood] designed for a fast computation of a high dimensional dataset. - SVSS: optimizes the derived ELBO estimator Eq. (7) using the SGVB. We validate the $\hat{\mathcal{L}}_{N}$ with $N=1$. This is the basic inference method proposed in this study. - SVSS-Ws: denotes the SVSS combined with the weight sampling strategy of Proposition 2. - SVSS-Ng: denotes the SVSS combined with the natural gradient of Proposition 3. - SVSS-WsNg: denotes the SVSS combined with the weight sampling strategy of Proposition 2 and the natural gradient of Proposition 3. We use the UCI datasets in [@Dua:2019]: Skillcraft, Parkinsons, and Elevators. We run 5 repetitive experiments and obtain the statistical result. For each experiment, each data set is randomly separated into a 90$\%$ training set and 10$\%$ test set. In this experiment section, we report only the averaged value of the root mean square error (RMSE) of the prediction on the test data during training. Figure 4 describes the prediction results of the Skillcraft dataset with $Q\in \{4,8\}$ and $M\in \{3Q,5Q,15Q\}$ using the approximate kernel $\Phi^{\mathrm{SM}}_{\textbf{s}}(X)\Phi^{\mathrm{SM}}_{\textbf{s}}(X)^{T}$. By comparing SVSS with the SS and the SS with the reparameterization trick (SS + RP) for all cases, we can confirm that the regularizer KL term in $\hat{\mathcal{L}}_{N=1}$ helps relax the over-fitting and find better parameters. Notably, its effectiveness becomes clear as $Q$ increases, which implies that the SVSS approach is beneficial for learning numerous parameters. By comparing the results between SVSS, SVSS-Ng, SVSS-Ws, and SVSS-WsNg, we can see that the proposed sampling strategy and natural gradient accelerates the convergence of the variational parameters in all cases of $Q$ and $M$. Especially, the use of a sampling strategy and natural gradient creates the synergy to accelerate the parameter learning. Similar results are obtained for the other datasets. These results are provided in the supplementary material. UCI Dataset Regression Task --------------------------- To evaluate the approximate inference quality, we conduct a regression task specifically for the large-scale datasets that are difficult to be trained using a conventional inference method. For the comparison, we consider the following approximation methods in addition to the baseline inference methods compared in the previous experiments: - VFE: denotes the variational inducing input method [@titsias2009variational], which is a representative scalable inference method in $\mathcal{GP}$. We use not only the SM kernel but also the RBF (ARD) kernel because the RBF is the one of the most widely used kernels in the $\mathcal{GP}$ model. - VSS: denotes the variatioanl sparse spectrum approximation of $\mathcal{GP}$ [@gal2015improving] which improves the sparse spectrum $\mathcal{GP}$ [@lazaro2010sparse] by variational approximation of the spectral points and the inputs of the data through the VI [@jordan1999introduction; @hoffman2013stochastic]. This method assumes that the SM kernel is used. [![image](EXP1-3/ctslice_totalcomparison_partial4.pdf){width="95.00000%"}]{} [lrc ccccc ccc]{} &&&\ (r)[4-11]{} Dataset & $N$ & $d$ & VFE (RBF) & VFE (SM) & SS & VSS & **SVSS** & **SVSS Ws** & **SVSS WsNg** & **Exact (WsNg)**\ (r)[1-11]{} Concrete & 1,030 & 8 & 0.377 $\pm$ .005 & 0.402 $\pm$ .009 & 0.421 $\pm$ .041 & 0.618 $\pm$ .000 & 0.396 $\pm$ .033 & 0.347 $\pm$ .009 & **0.341** $\pm$ .009 & 0.833 $\pm$ .009\ Skillcraft & 3,325 & 18 & 0.300 $\pm$ .013 & 0.309 $\pm$ .016 & 0.320 $\pm$ .008 & 0.310 $\pm$ .015 & 0.296 $\pm$ .009 & 0.297 $\pm$ .013 & **0.295** $\pm$ .012 & 0.310 $\pm$ .007\ Parkinsons & 5,875 & 20 & 1.386 $\pm$ .096 & 1.670 $\pm$ .343 & 3.084 $\pm$ .250 & 6.618 $\pm$ .161 & 2.821 $\pm$ .100 & 1.505 $\pm$ .115 & 1.635 $\pm$ .148 & **0.528** $\pm$ .068\ Kin8nm & 8,192 & 8 & 0.145 $\pm$ .005 & 0.148 $\pm$ .010 & 0.167 $\pm$ .006 & 0.226 $\pm$ .045 & 0.177 $\pm$ .009 & 0.130 $\pm$ .005 & 0.128 $\pm$ .008 & **0.080** $\pm$ .002\ Elevators & 16,599 & 18 & 0.121 $\pm$ .003 & 0.103 $\pm$ .002 & 0.100 $\pm$ .005 & 0.124 $\pm$ .003 & 0.093 $\pm$ .001 & 0.095 $\pm$ .001 & 0.095 $\pm$ .001 & **0.089** $\pm$ .002\ ------------------------------------------------------------------------ $\text{Protein}^{*}$ & 45,730 & 9 & 0.613 $\pm$ .019 & 0.620 $\pm$ .019 & 0.631 $\pm$ .018 & 0.653 $\pm$ .016 & 0.621 $\pm$ .016 & 0.603 $\pm$ .018 & 0.601 $\pm$ .016 & **0.542** $\pm$ .023\ $\text{Blog}^{*}$ & 52,397 & 280 & 0.915 $\pm$ .029 & **0.771** $\pm$ .015 & 0.885 $\pm$ .040 & 0.919 $\pm$ .067 & 0.845 $\pm$ .013 & 0.784 $\pm$ .034 & 0.784 $\pm$ .030 & 0.846 $\pm$ .122\ $\text{CTsilce}^{*}$ & 53,500 & 385 & 6.522 $\pm$ .458 & 6.948 $\pm$ .329 & 8.122 $\pm$ .331 & 8.496 $\pm$ .106 & 10.988 $\pm$ .246 & 7.952 $\pm$ .169 & 7.867 $\pm$ .292 & **2.614** $\pm$ .175\ After 5 repetitive experiments, the statistical results are obtained. For each experiment, the training and test data are randomly selected with a ratio of 9:1. We use single GPU (V100-16GB). For the Protein, Blog, and CTslice datasets, VFE (SM) incurs a memory problem. For a fair comparison, we equally divide 5 partitions of the dataset and then obtain the averaged result as regression task in [@wilson2016deep]. Figure 5 compares the performance of the regression task for the CTslice datasets. We also present the prediction results obtained using the exact SM kernel with the parameters estimated by the SVSS and SVSS-WsNg. We can see that the parameters inferred by the proposed inference methods can be used for the exact SM kernel; The proposed inference method predicts the outputs more accurately using less computational time than other baseline inferences. Table 1 summarizes the RMSE for each dataset. Exact (WsNg) denotes the prediction results obtained by the exact SM kernel with the parameters estimated using SVSS-WsNg. We confirm that the proposed SVSS-WsNg and Exact (WsNg) achieve better prediction results for most of the dataset. Conclusion ========== In this research, we proposed a way to efficiently estimate the hyperparameters of an SM kernel by employing a sampling-based variational inference. Because we employ a regularized ELBO estimator as an objective function, we can relax the over-fitting issue in SM kernel training . In addition, we train the parameters of SM kernel in a scalable manner for large-scale data. To improve the inference quality, we propose a sampling strategy for spectral points to robustly compute the regularized ELBO estimator. We also propose an approximate natural gradient to optimize the variational parameters of the SM kernel. We validated that the combination of the sampling strategy and the approximate natural gradient used in the proposed approximate inference accelerates the convergence of the parameters and results in better parameters for the SM kernel. Broader Impact ============== In general, $\mathcal{GP}$ model is said to have the advantage of quantifying the uncertainty for the prediction of the model. This characteristic allows the $\mathcal{GP}$ model to be widely used for the decision-making because the quantified uncertainty of the prediction can be incorporated into decision-making. In particular, for the sensitive problems where the individual randomness is more reflected in the dataset and makes decision difficult, the quantified uncertainty can be a helpful factor to the decision. For example, when the newly developed medicine should be verified for its use, the deterministic prediction of clinical effect for the potential users would not be completely trustful because the clinical trial results used as a training dataset contain the individual error of the tester. Thus, the credibility of the prediction could be important in determining whether developed medicine is used. This example explains why the uncertainty about the prediction should be estimated accurately and explains why the elaborated $\mathcal{GP}$-based hybrid model have been proposed for the accurate uncertainty estimation and prediction. However, when the model becomes more complex, the learning of the model is likely to have problems as the case of the SM kernel. In this context, our approximate inference method can be used to train the complex $\mathcal{GP}$-based model using a large-scale dataset while alleviating the over-fitting issue. Furthermore, this approximate inference method can potentially help the process of making more reliable decisions.
{ "pile_set_name": "ArXiv" }
--- abstract: | We consider the classical theta operator $\theta$ on modular forms modulo $p^m$ and level $N$ prime to $p$ where $p$ is a prime greater than $3$. Our main result is that $\theta$ mod $p^m$ will map forms of weight $k$ to forms of weight $k+2+2p^{m-1}(p-1)$ and that this weight is optimal in certain cases when $m$ is at least $2$. Thus, the natural expectation that $\theta$ mod $p^m$ should map to weight $k+2+p^{m-1}(p-1)$ is shown to be false. The primary motivation for this study is that application of the $\theta$ operator on eigenforms mod $p^m$ corresponds to twisting the attached Galois representations with the cyclotomic character. Our construction of the $\theta$-operator mod $p^m$ gives an explicit weight bound on the twist of a modular mod $p^m$ Galois representation by the cyclotomic character. address: - 'Department of Mathematics, Simon Fraser University, 8888 University Drive, Burnaby, B.C., V5A 1S6, Canada' - 'Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark' author: - 'Imin Chen, Ian Kiming' title: On the theta operator for modular forms modulo prime powers --- Introduction {#intro} ============ Let $p$ be a prime number. We shall assume $p\ge 5$ throughout the paper in order to avoid certain technicalities when $p$ is $2$ or $3$. Further let $m \in {\mathbb N}$ and denote by $M_k(N,{\mathbb Z}_p)$ the ${\mathbb Z}_p$-module of modular forms of weight $k$ for $\Gamma_1(N)$ over ${\mathbb Z}_p$ and let $M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$ be the ${\mathbb Z}/p^m{\mathbb Z}$-module of modular forms for $\Gamma_1(N)$ over ${\mathbb Z}/p^m{\mathbb Z}$ as defined classically by $M_k(N,R) = M_k(N,{\mathbb Z}) \otimes R$. Note that this definition relies on the existence of an integral structure on $M_k(N,{\mathbb C})$ (see for instance [@ckw]). Let $k_1,\ldots,k_t$ be a collection of weights and let $f_i \in M_{k_i}(N,{\mathbb Z}_p)$. The $q$-expansion of an element in a direct sum of the $M_{k_i}(N,{\mathbb Z}_p)$’s or $M_{k_i}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$’s is defined by extending linearly on each component. When we write $f_1 + \ldots + f_t \equiv 0 \pmod{p^m}$, we shall mean that the $q$-expansion $f_1(q) + \ldots + f_t(q)$ lies in $p^m {\mathbb Z}_p[[q]]$. Similarly for $f_i \in M_{k_i}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$, we write $f_1 + \ldots + f_t \equiv 0 \pmod{p^m}$ if the $q$-expansion of $f_1 + \ldots + f_t$ equals $0$ in $({\mathbb Z}/p^m{\mathbb Z})[[q]]$. In such a case, we say that $f_1 + \ldots + f_t$ is congruent to $0$ modulo $p^m$. Let us recall the definition and basic properties of the standard Eisenstein series on $\operatorname{SL}_2({\mathbb Z})$, cf. §1 of [@serre_zeta], for instance: the series $$G_k := -\frac{B_k}{2k} + \sum_{n=1}^{\infty} \sigma_{k-1}(n) q^n$$ with $B_k$ the $k$-th Bernoulli number and $\sigma_t(n) := \sum_{d\mid n} d^t$ the usual divisor sum, is (with $q:=e^{2\pi iz})$ for $k$ an even integer $\ge 4$ a modular form on $\operatorname{SL}_2({\mathbb Z})$. Defining $E_k$ as the normalization $$E_k := -\frac{2k}{B_k} \cdot G_k$$ one has $E_k \equiv 1 \pmod{p^t}$ when (and only when) $k\equiv 0 \pmod{p^{t-1}(p-1)}$. There are natural inclusions (preserving $q$-expansions) $$\begin{aligned} M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}}) \hookrightarrow M_{k+p^{m-1}(p-1)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}}),\end{aligned}$$ induced by multiplication by $E_{p-1}^{p^{m-1}}$, using the fact that $E_{p-1}^{p^{m-1}} \equiv 1 \pmod {p^m}$. Note that $E_{p^{m-1}(p-1)} = E_{p-1}^{p^{m-1}}$ in $M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$, again by the $q$-expansion principle, so that the map can also be seen as induced by multiplication by $E_{p^{m-1}(p-1)}$. As is well-known, when we specialize the above series for $G_k$ to $k=2$ and define $$G_2 := -\frac{B_2}{4} + \sum_{n=1}^{\infty} \sigma_1(n) q^n = -\frac{1}{24} + \sum_{n=1}^{\infty} \sigma_1(n) q^n$$ then $G_2$ does not represent a modular form in the usual sense, but does so in the $p$-adic sense, cf. [@serre_zeta], §2. One defines $E_2$ as the normalization of $G_2$, i.e., $E_2 := -24 G_2$. Thus, $E_2$ is also a $p$-adic modular form. Consider the classical theta operator $\theta f = \frac{k}{12} E_2 + \frac{1}{12} \partial f$ of Ramanujan. Its effect on $q$-expansions is $\sum_n a_n q^n \mapsto \sum_n n a_n q^n$. Since $E_2$ is a $p$-adic modular form it is for any $m\in{\mathbb N}$ congruent modulo $p^m$ to a classical modular form of some weight. Thus we have $E_2 \equiv E_{p+1} \pmod{p}$, for example, and since we classically know that $\partial$ maps modular forms of weight $k$ to modular forms of weight $k+2$, one obtains the classical operator $\theta$ that maps $M_k(N,{\mathbb F}_p)$ to $M_{k+p+1}(N,{\mathbb F}_p)$. Studying this operator as well as its interaction with the ‘weight filtration’ (see below) is a key tool in the theory of modular forms modulo $p$; cf. for instance Jochnowitz’ proof of finiteness of systems of Hecke eigenvalues mod $p$ across all weights in [@jochnowitz], or Edixhoven’s results on the optimal weight in Serre’s conjectures [@edix]. As we have launched a framework for the study of modular forms mod $p^m$ in [@ckr] and [@ckw], it is natural to ask whether a $\theta$ operator with similar properties can be defined on such forms. We have discussed in [@ckw] how to attach Galois representations to eigenforms mod $p^m$, and it is clear from the properties of those attached representations that applying the $\theta$ operator corresponds on the Galois side to twisting by the cyclotomic character mod $p^m$. Hence, the construction of the theta operator mod $p^m$ yields an immediate application to mod $p^m$ Galois representations (see Corollary \[main-cor\]). Notice that Serre shows in [@serre_zeta Théorème 5] that there exists a $\theta$ operator on $p$-adic modular forms (of level $1$) whose effect on $q$-expansions is $\sum_n a_n q^n \mapsto \sum_n n a_n q^n$ and that sends a form of ($p$-adic) weight $k$ to a form of weight $k+2$. One can view our results as finding a partially explicit expression with explicit weights for the mod $p^m$ reduction of this operator. Hence, our results are then that on the one hand an extension of the $\theta$ operator from the mod $p$ to the mod $p^m$ situation is indeed possible, but that the interplay of the $\theta$ operator with the weights of the forms becomes much more complicated when $m>1$ and that, in fact, there are certain genuine qualitative differences between the case $m=1$ and the general cases $m>1$. Let us explain in detail. We show that a $\theta$ operator on modular forms mod $p^m$ can be defined such that $\theta$ maps $M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}}) \rightarrow M_{k+k(m)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$ with $k(m) := 2 + 2 p^{m-1}(p-1)$, such that the effect on $q$-expansions is $\sum_n a_n q^n \mapsto \sum_n n a_n q^n$, and such that $\theta$ satisfies simple commutation rules with Hecke operators $T_{\ell}$ for primes $\ell \neq Np$, cf. the first part of Theorem \[theta-single-wt\] below. The proofs use a number of results from [@serre_zeta] plus the observation that $f \mid V \equiv f^p \pmod p$ where $V$ is the classical $V$ operator. Define the weight $w_{p^m}(f)$ of an modular form $f$ mod $p^m$ with $f \not \equiv 0 \pmod p$ to be the smallest $k \in {\mathbb Z}$ such that $f$ is congruent modulo $p^m$ to an element of $M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. A classical fact, crucial for instance in the work [@jochnowitz], is that when $m=1$ we have $w_p(\theta f) \le w_p(f) + p + 1$ with equality if (and only if) $p \nmid w_p(f)$. One might expect the generalization of this to be that $w_{p^m}(\theta f) \le w_{p^m}(f) + 2 + p^{m-1}(p-1)$ (perhaps with equality in some cases). However, as the second part of Theorem \[theta-single-wt\] shows, this is false: \[theta-single-wt\] Let $p \ge 5$ be a prime. Put $$k(m) := 2 + 2 p^{m-1}(p-1).$$ \(i) The classical theta operator $\theta$ induces an operator $$\theta : M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}}) \rightarrow M_{k + k(m)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$$ whose effect on $q$-expansions is $\sum a_n q^n \mapsto \sum n a_n q^n$. \(ii) If $\ell \not= p$ is a prime and $T_{\ell}$ denotes the $\ell$-th Hecke operator, then $$T_{\ell} \theta = \ell \cdot \theta T_{\ell}$$ as linear maps $M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}}) \rightarrow M_{k + k(m)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. \(iii) Assume that $N$ is prime to $p$. Let $m\ge 2$ and $f \in M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$ with $f \not\equiv 0 \pmod p$. Suppose further that $p\nmid k$ and $w_p(f)= k$. Then $$w_{p^m}(\theta f) = k + 2 + 2 p^{m-1} (p-1).$$ The proof of the second part of the Theorem uses some results of [@katz_modp], which are applicable to general level $N$ prime to $p$. In particular, a main point is that if we consider the Eisenstein series $E_{p-1}$ and $E_{p+1}$ as modular forms modulo $p$ in the sense of Katz then $E_{p-1}$ (Hasse invariant) has only simple zeros and no zero common with $E_{p+1}$; cf. [@katz_modp Remark on p. 57]. This point allows one to compute the weight filtration of the last term of the expression below for the $\theta$ operator mod $p^m$, which is the controlling term. We would like to thank Nadim Rustom for a useful discussion about this point. We have in fact conducted an in-depth study of the relation between and $w_{p^m}(f)$ and $w_{p^m}(\theta f)$ in the case of level $N=1$ and for $m=2$. However, the results are rather complicated and will not be stated in this article. For simplicity, we have stated results in this paper for modular forms with coefficients in ${\mathbb Z}_p$ and hence reductions with coefficients in ${{\mathbb Z}/{p^m}{\mathbb Z}}$. The above theorem however is valid for coefficients in ${\overline{{\mathbb Z}/{p^m}{\mathbb Z}}}$ (see e.g. [@ckw], section 2.4 for a definition of this ring) using the same proofs. An immediate consequence of Theorem \[theta-single-wt\] to Galois representations is the following. We use the notation and terminology from [@ckw]. \[main-cor\] Let $p \ge 5$ be a prime, $\rho : G_{\mathbb Q}\rightarrow \operatorname{GL}_2({\overline{{\mathbb Z}/{p^m}{\mathbb Z}}})$ be a residually absolutely irreducible Galois representation, and $\chi :G_{\mathbb Q}\rightarrow ({{\mathbb Z}/{p^m}{\mathbb Z}})^\times$ be the reduction modulo $p^m$ of the $p$-adic cyclotomic character. Suppose $\rho \cong \rho_f$ for some weak eigenform $f \in S_k(N)({\overline{{\mathbb Z}/{p^m}{\mathbb Z}}})$. Then $\rho \otimes \chi \cong \rho_{g}$ for some weak eigenform $g \in S_{k + k(m)}(N)({\overline{{\mathbb Z}/{p^m}{\mathbb Z}}})$. Note first that $\theta$ maps cusp forms to cusp forms. Suppose then that $f \in S_k(N)({\overline{{\mathbb Z}/{p^m}{\mathbb Z}}})$ is a weak eigenform in the sense that $T_\ell f \equiv f(T_\ell) f \pmod{p^m}$ for all $\ell \nmid D$ for some integer $D$. Then $g = \theta f \in S_{k+k(m)}(N)({\overline{{\mathbb Z}/{p^m}{\mathbb Z}}})$ is a weak eigenform in the sense that $T_\ell g \equiv g(T_\ell) g \pmod{p^m}$ for all $\ell \nmid NDp$, and we have that $f(T_\ell) \chi(\ell) \equiv g(T_\ell) \pmod{p^m}$ for all $\ell \nmid Dp$. The hypothesis of residual absolute irreducibility then allows us to conclude that $\rho_f \otimes \chi \cong \rho_{\theta f}$. Regarding Corollary \[main-cor\], the main result of the paper [@cs] implies an analogous statement about twisting with the mod $p^m$ reduction of the Teichmüller character, with some differences: In [@cs], they consider the twist of $f$ by the mod $p^m$ reduction of the Teichmüller character where $f$ is a strong eigenform (again using the terminology of [@ckw]). One then finds a strong (and not merely weak) eigenform $g$ as in the Corollary, but apparently without control over the weight $k+k(m)$. The proof uses different methods (Coleman $p$-adic families of modular forms.) The theta operator modulo prime powers ====================================== Eisenstein series {#eisenstein_series} ----------------- We shall now develop an explicit expression for the truncation modulo $p^m$ of the $p$-adic Eisenstein series $G_2$. \[prop:G\_2\_modpm\] Let $m\in{\mathbb N}$. Define the positive even integers $k_0,\ldots,k_{m-1}$ as follows: If $m\ge 2$, define: $$k_j := 2 + p^{m-j-1}(p^{j+1}-1) \quad \mbox{for $j=0,\ldots,m-2$}$$ and $$k_{m-1} := p^{m-1}(p+1)$$ and define just $k_0 := p+1$ if $m=1$. Then $k_0 < \ldots < k_{m-1}$ and there are modular forms $f_0,\ldots,f_{m-1}$, depending only on $p$ and $m$, of level $1$ and of weights $k_0,\ldots,k_{m-1}$, respectively, that have rational $q$-expansions, satisfy $v_p(f_j) = 0$ for all $j$, and are such that $$G_2 \equiv \sum_{j=0}^{m-1} p^j f_j \pmod{p^m}$$ as a congruence between $q$-expansions. The form $f_{m-1}$ can be chosen to be $f_{m-1} = G_{p+1}^{p^{m-1}}$. When $m=2$ we can, and will, be a bit more explicit: \[prop:G\_2\_modp2\] We have: $$G_2 \equiv f_0 + p\cdot f_1 \pmod{p^2}$$ with modular forms $f_0$ and $f_1$ of weights $2+p(p-1)$ and $p(p+1)$, respectively, explicitly: $$G_2 \equiv G_{2+p(p-1)} + p\cdot G_{p+1}^p \pmod{p^2} .$$ It is amusing to note the following consequence of the Proposition: For $p \not=2, 3$, we have the following congruence of Bernoulli numbers, $$\frac{B_2}{2} \equiv \frac{B_{p(p-1)+2}}{p(p-1)+2} + p \frac{B_{p+1}}{p+1} \pmod {p^2}.$$ However, this can also be seen in terms of $p$-adic continuity of Bernoulli numbers (cf. [@washington], Cor. 5.14 on p. 61, for instance). Before the proofs of these propositions we need a couple of preparations. Let $k$ be an even integer $\ge 2$. Recall from [@serre_zeta] that if we choose a sequence of even integers $k_i$ such that $k_i \rightarrow \infty$ in the usual, real metric, but $k_i\rightarrow k$ in the $p$-adic metric, then the sequence $G_{k_i}$ has a $p$-adic limit denoted by $G_k^{\ast}$. This series $G_k^{\ast}$ is a $p$-adic modular form of weight $k$. It does not depend on the choice of the sequence $k_i$. In particular, we can, and will, choose $k_i := k + p^{i-1}(p-1)$, because if we chose another $k_i' = k + \lambda p^{i-1}(p-1)$, where $\lambda \ge 2$, then $G_{k_i'} \equiv G_{k_i} \pmod{p^i}$ so that $G_{k_i'} = G_{k_i} E_{p-1}^{p (\lambda-1)}$ in $M_{k_i'}({{\mathbb Z}/{p^i}{\mathbb Z}})$ is not essentially different. \[lemma:G\_k\^\*\] Let $k$ be an even integer $\ge 2$ and assume $(p-1)\nmid k$. Let $t\in{\mathbb N}$. Then $G_k^{\ast} \equiv G_{k+p^{t-1}(p-1)} \pmod{p^t}$. Let $u,v\ge t$. We claim that $G_{k_u} \equiv G_{k_v} \pmod{p^t}$. Since the series $G_{k_i}$ converges $p$-adically to $G_k^{\ast}$, the claim clearly implies the Lemma. If $u=v$ the claim is trivial, so suppose not, say $u<v$. Then $k_v-k_u = p^{v-1}(p-1) - p^{u-1}(p-1)$ is a multiple of $p^{t-1}(p-1)$, say $k_v - k_u = s\cdot p^{t-1}(p-1)$. We also have $k_v - k_u \ge 4$. Hence, we find that $G:=G_{k_u} \cdot E_{p^{t-1}(p-1)}^s$ is a modular form of weight $k_v$, and we have $G_{k_u} \equiv G \pmod{p^t}$. Now notice that, when $i\ge t$, we have: $$\sigma_{k_i-1}(n) = \sum_{d\mid n} d^{k-1+p^{i-1}(p-1)} \equiv \sum_{\substack{d\mid n \\ p\nmid d}} d^{k-1} \pmod{p^t}$$ as $d^{p^{i-1}(p-1)} \equiv 1 \pmod{p^t}$ when $p\nmid d$ and $i\ge t$, and as $d^{p^{i-1}(p-1)} \equiv 0 \pmod{p^t}$ when $p\mid d$ and $i\ge t$ (because $p^{t-1}(p-1) \ge t$ as long as $p \not= 2$). We conclude that the nonconstant terms of the series $G_{k_u}$ and $G_{k_v}$ are termwise congruent modulo $p^t$. The same is then true of the forms $G$ and $G_{k_v}$ that are both forms of weight $k_v$. Hence, the nonconstant terms of the form $(G-G_{k_v})/p^t$ are all $p$-integral. As $k_v \equiv k \not\equiv 0 \pmod{(p-1)}$, it follows from Théorème 8 of [@serre_modp] that the constant term of this form is in fact also $p$-integral. Hence, $$G_{k_u} \equiv G \equiv G_{k_v} \pmod{p^t}$$ as desired. Recall that the $V$ operator is defined on formal $q$-expansions as $$\left( \sum a_n q^n \right) \mid V := \sum a_n q^{np} .$$ \[cor:G\_2a\] We have $$G_2 \equiv \sum_{j=0}^{m-1} p^j \cdot \left( G_{2+p^{m-j-1}(p-1)}\mid V^j \right) \pmod{p^m}$$ as a congruence between formal $q$-expansions. Recall from [@serre_zeta], §2, the identity, valid for any even integer $k\ge 2$, that $$G_k = G_k^{\ast} + p^{k-1} \left( G_k^{\ast} \mid V \right) + \ldots + p^{t(k-1)} \left( G_k^{\ast} \mid V^t \right) + \ldots .$$ The identity is first an identity of formal $q$-expansions, but then shows that $G_k$ is a $p$-adic modular form as $V$ acts on $p$-adic modular forms, cf. [@serre_zeta], §$2$. If we specialize this identity to the case $k=2$, reduce modulo $p^m$, and note that the previous Lemma applies since $(p-1)\nmid 2$, the claim follows immediately. In the next paragraph and lemma, we use the notation $M_k(\Gamma,F)$ to mean the $F$-module of modular forms of weight $k$ over $F$, where $\Gamma$ is a congruence subgroup of $\operatorname{SL}_2({\mathbb Z})$ and $F$ is a subring of ${\mathbb C}$ or ${\mathbb C}_p$. We can also see the $V$ operator as an operator on modular forms: Suppose that $f\in M_k(\operatorname{SL}_2({\mathbb Z}),{\mathbb C})$. Then $(f\mid V)(z) = f(pz)$, and as is well-known $f\mid V \in M_k(\Gamma_0(p);{\mathbb C})$. The proof of the next lemma is a simple application of section 3.2 of [@serre_zeta]. Recall that if $f\in M_k(\operatorname{SL}_2({\mathbb Z}),{\mathbb Q}_p)$ is nonzero with $q$-expansion $\sum_n a_n q^n$ we define $v_p(f) := \min \{ v_p(a_n) \mid ~ n\in{\mathbb N}\}$ where $v_p(a_n)$ is the usual (normalized) $p$-adic valuation of $a_n$. \[lemma:V-operator\] Let $f\in M_k(\operatorname{SL}_2({\mathbb Z}),{\mathbb Q})$ and suppose $v_p(f) = 0$. Let $t\in{\mathbb N}$ and suppose that $s\in {\mathbb Z}_{\ge 0}$ is such that $$\inf (s+1,p^s+1-k) \ge t .$$ Then there is $h\in M_{k+p^s(p-1)}(\operatorname{SL}_2({\mathbb Z}),{\mathbb Q})$ with $v_p(h) = 0$, and such that $$f\mid V \equiv h \pmod{p^t}.$$ As we noted above, $f\mid V$ is a modular form of weight $k$ on $\Gamma_0(p)$. Since $f\mid V = \sum a_n q^{np}$ if $f=\sum a_n q^n$ we have $v_p(f\mid V) = 0$. Recall the Fricke involution for modular forms on $\Gamma_0(p)$ given by the action of the matrix $$W = {\left(\begin{smallmatrix}0&-1\\#3&0\end{smallmatrix}\right)} .$$ Since $$f\mid V = p^{-k/2} f\mid_k {\left(\begin{smallmatrix}p&0\\#3&1\end{smallmatrix}\right)} ,$$ (recall that the weight $k$ action is normalized so that diagonal matrices act trivially), since $${\left(\begin{smallmatrix}p&0\\#3&1\end{smallmatrix}\right)} {\left(\begin{smallmatrix}0&-1\\#3&0\end{smallmatrix}\right)} = {\left(\begin{smallmatrix}0&-1\\#3&0\end{smallmatrix}\right)} {\left(\begin{smallmatrix}p&0\\#3&p\end{smallmatrix}\right)} ,$$ and since $f$ is on $\operatorname{SL}_2({\mathbb Z})$ we see that $f\mid VW = p^{-k/2} f$ so that $v_p(f\mid VW) = -k/2$. Now let $E:=E_{p-1}$ and put $$g := E - p^{p-1} (E\mid V)$$ so that $g$ is a modular form of weight $p-1$ on $\Gamma_0(p)$. Then, if we put $$f_s := \operatorname{tr}((f\mid V)\cdot g^{p^s})$$ for $s\in{\mathbb Z}_{\ge 0}$ where $\operatorname{tr}$ denotes the trace from $\Gamma_0(p)$ to $\operatorname{SL}_2({\mathbb Z})$, it follows from section 3.2 of [@serre_zeta] that $f_s$ is a modular form on $\operatorname{SL}_2({\mathbb Z})$ of weight $k+p^s(p-1)$ and rational $q$-expansion. Furthermore, Lemme 9 of [@serre_zeta] implies that $$\begin{aligned} v_p(f_s-(f\mid V)) & \ge & \inf (s+1,p^s+1+v_p(f\mid VW) -k/2) \\ & = & \inf (s+1,p^s+1-k) \ge t .\end{aligned}$$ Hence, we can choose $h:=f_s$. As $f\mid V \equiv h \pmod{p^t}$ and $v_p(f\mid V) = 0$, we must have $v_p(h) = 0$. That the defined weights $k_0,\ldots,k_{m-1}$ satisfy $k_0 < \ldots < k_{m-1}$ is verified immediately. Thus, starting with Corollary \[cor:G\_2a\] we see that it suffices to show that for each $j\in \{ 0,\ldots ,m-1\}$ there is a modular form $f_j$ of weight $k_j$ with rational $q$-expansion and $v_p(f_j) = 0$, and such that $$G_{2+p^{m-j-1}(p-1)} \mid V^j \equiv f_j \pmod{p^{m-j}} .$$ If $m=1$, $j=0$ we just take $f_0 := G_{p+1}$, so assume $m\ge 2$. Then, if $j=m-1$, note that $$G_{p+1} \mid V^{m-1} \equiv G_{p+1}^{p^{m-1}} \pmod{p}$$ so that we can take $f_{m-1} := G_{p+1}^{p^{m-1}}$ that is of weight $k_{m-1} = p^{m-1}(p+1)$. So, suppose that $j\le m-2$. We claim that for $r=0,\ldots ,j$ there is a modular form $g_r$ of weight $2+p^{m-j-1}(p^{r+1}-1)$, rational $q$-expansion with $v_p(g_r)=0$, and such that $$G_{2+p^{m-j-1}(p-1)} \mid V^r \equiv g_r \pmod{p^{m-j}}$$ which is the desired when $r=j$. We prove the last claim by induction on $r$ noting that the case $r=0$ is trivial. So, suppose that $r<j$ and that we have already shown the existence of a modular form $g_r$ as above. Notice that $$p^{m-j+r} + 1 - (2 + p^{m-j-1}(p^{r+1}-1)) = p^{m-j-1} - 1 \ge m-j$$ holds because $m-j \ge 2$ (we used here that $p>2$). Thus we see that Lemma \[lemma:V-operator\] applies (taking $s = m-j+r$) and shows the existence of a modular form $g_{r+1}$ with rational $q$-expansion and $v_p(g_{r+1})=0$, such that $$g_r \mid V \equiv g_{r+1} \pmod{p^{m-j}},$$ and such that $g_{r+1}$ has weight $$2+p^{m-j-1}(p^{r+1}-1) + p^{m-j+r}(p-1) = 2+p^{m-j-1}(p^{r+2}-1),$$ and we are done. It is interesting to note that in the induction, the inequalities do not allow us to deal with the case $j = m-1$, but then we use the congruence $f \mid V \equiv f^p \pmod p$ to take care of the last term. Again by Corollary \[cor:G\_2a\] we have: $$G_2 \equiv G_{2+p(p-1)} + p\cdot \left( G_{p+1} \mid V \right) \pmod{p^2} .$$ Noting again that $G_{p+1} \mid V \equiv G_{p+1}^p \pmod{p}$ so that $$p\cdot \left( G_{p+1} \mid V \right) \equiv p\cdot G_{p+1}^p \pmod{p^2} ,$$ we are done. Definition and properties of the theta operator {#def_theta} ----------------------------------------------- Recall the classical $\theta$ operator acting on formal $q$-expansions as $q\frac{d}{dq}$, i.e., $$\theta \left( \sum a_n q^n \right) := \sum n a_n q^n,$$ and the operator $\partial$ defined by $$\frac{1}{12} \partial f := \theta f - \frac{k}{12} E_2 \cdot f = \theta f + 2k G_2\cdot f$$ when $f = \sum a_n q^n \in M_k(N,{\mathbb C})$ is a modular form of weight $k$ (we have $B_2 = \frac{1}{6}$ so that $E_2 = -24 G_2$). Then $\partial f$ is in $M_{k+2}(N,{\mathbb C})$, and $\partial$ defines a derivation on $M(N,{\mathbb C}) := \oplus_k M_k(N,{\mathbb C})$ (as follows by writing $\theta = \frac{1}{2\pi i} \cdot \frac{d}{dz}$ (as $q = e^{2\pi i z})$ and combining with the classical transformation properties of $E_2$ under the weight $2$ action of $\operatorname{SL}_2({\mathbb Z})$ given in ) The definition of $\partial$ implies that $\partial$ defines a derivation on $\oplus_k M_k(N,{\mathbb Z})$ and hence also on $M(N,{\mathbb Z}_p) := \oplus_k M_k(N,{\mathbb Z}_p)$. [*(i)*]{} Retain the notation of Proposition \[prop:G\_2\_modpm\] so that $$k_j := 2 + p^{m-j-1}(p^{j+1}-1) \quad \mbox{for $j=0,\ldots,m-2$} ,$$ and $$k_{m-1} := p^{m-1}(p+1) .$$ Then $k_0 < \ldots < k_{m-1}$ and by Proposition \[prop:G\_2\_modpm\] we have modular forms $f_0,\ldots,f_{m-1}$ (of level $1$ and) of weights $k_0,\ldots,k_{m-1}$, respectively, that have rational $q$-expansions, satisfy $v_p(f_j) = 0$ for all $j$, and are such that $$G_2 \equiv \sum_{j=0}^{m-1} p^j f_j \pmod{p^m} .$$ With $k(m) := 2 + 2 p^{m-1}(p-1)$ one checks that each number $k(m) - k_j$ is a multiple of $p^{m-j-1}(p-1)$, say $k(m) = k_j + t_j\cdot p^{m-j-1}(p-1)$ for $j=0,\ldots, m-1$. Since $E_{p-1}^{p^{m-j-1}} \equiv 1 \pmod{p^{m-j}}$ we find that $p^j E_{p-1}^{p^{m-j-1} t_j} \equiv p^j \pmod{p^m}$, and so the above congruence for $G_2$ can also be written as $$G_2 \equiv \sum_{j=0}^{m-1} p^j E_{p-1}^{p^{m-j-1} t_j} f_j \pmod{p^m}$$ where now each summand is a form of weight $k(m)$. Hence, for any an element $f\in M_k(N,{\mathbb Z}_p)$ we find that $$\begin{aligned} \label{theta-defn} \theta f & = \frac{1}{12} \partial f - 2k G_2\cdot f \\ & \equiv \frac{1}{12} E_{p-1}^{2p^{m-1}} \partial f - 2k f \sum_{j=0}^{m-1} p^j E_{p-1}^{p^{m-j-1} t_j} f_j \pmod{p^m} \\ & =: \theta_{p^m} f \in M_{k+k(m)}({\mathbb Z}_p)\end{aligned}$$ where now each summand on the right hand side is an element of $M_{k+k(m)}(N,{\mathbb Z}_p)$. Thus the classical theta operator induces a linear map $$\theta = \theta_{p^m} : M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}}) \rightarrow M_{k + k(m)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$$ the effect of which on $q$-expansions is $\sum a_n q^n \mapsto \sum n a_n q^n$. We still denote this operator (by abuse of notation) by $\theta$, but later when we need to distinguish from $\theta := \frac{1}{2\pi i} \cdot \frac{d}{dz}$, we will denote it by $\theta_{p^m}$. [*(ii)*]{} First assume for the prime $\ell$ that we have $\ell\nmid N$ (in addition to $\ell \neq p$). Recall that the diamond operator ${\left< \ell \right>}_k$ on a modular form $f$ of weight $k$ is defined by ${\left< \ell \right>}_k f = f \mid_k \gamma$ for any $\gamma = \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \operatorname{SL}_2({\mathbb Z})$, with $c \equiv 0 \pmod N$, $d \equiv \ell \pmod N$. (Note that we write the action of ${\left< \cdot \right>}_k$ from left, though it is given by the stroke operator which is from the right. This is fine because $({\mathbb Z}/N{\mathbb Z})^\times$ is abelian). As $\ell \neq p$ the operator ${\left< \ell \right>}_k$ induces a linear action on $M_k(N,{\mathbb Z}_p)$ (as follows from the well-known formula $\ell^{k-1} {\left< \ell \right>}_k = T_{\ell}^2 - T_{\ell^2}$ and the fact that the Hecke operators $T_n$ preserve $M_k(N,{\mathbb Z})$), and hence also on $M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. Using $\frac{1}{12} \partial f = \theta f -\frac{k}{12} E_2 f$ as well as the transformation property of $E_2$ given by $$\label{E-two} (E_2 \mid_2 \gamma)(z) = E_2(z) + \frac{12}{2\pi i} c j(\gamma,z)^{-1}$$ (see for instance [@ds]), a computation shows that we have $\frac{1}{12} \partial {\left< \ell \right>}_k f = {\left< \ell \right>}_{k+2} \frac{1}{12} \partial f$ for $f \in M_k(N,{\mathbb C})$. Now recall the above definition of $\theta_{p^m} : M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}}) \rightarrow M_{k + k(m)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$, as well as the fact that the forms $E_{p-1}$ and the $f_j$ occurring in the definition all have level $1$ and thus are fixed under the action of the operators ${\left< \ell \right>}_{p-1}$ and ${\left< \ell \right>}_{k_j}$, respectively. We deduce that: $$\theta_{p^m} {\left< \ell \right>}_k f = {\left< \ell \right>}_{k+k(m)} \theta_{p^m} f$$ for all $f \in M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. Let now $f = \sum a_n q^n \in M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. Then we have (with the usual convention that $a_{\frac{n}{\ell}} := 0$ if $\ell \nmid n$): $$\begin{aligned} {\left< \ell \right>}_k f &=& \sum b_n q^n \\ T_\ell f &=& \sum \left( a_{\ell n} + \ell^{k-1} \sum b_{\frac{n}{\ell}} \right) q^n \\ {\left< \ell \right>}_{k+k(m)} \theta_{p^m} f &=& \theta_{p^m} {\left< \ell \right>}_k f = \sum n b_n q^n \\ T_\ell \theta_{p^m} f &=& \sum \left( \ell n a_{\ell n} + \ell^{k+k(m)-1} \frac{n}{\ell} b_{\frac{n}{\ell}} \right) q^n \\ &=& \ell n \sum \left( a_{\ell n} + \ell^{(k-1) + k(m)-2} b_{\frac{n}{\ell}} \right) q^n \\ \ell \theta_{p^m} T_\ell f &=& \ell n \sum \left( a_{\ell n} + \ell^{k-1} b_{\frac{n}{\ell}} \right) q^n.\end{aligned}$$ For $\ell \mid N$, $\ell \not= p$, a similar calculation holds (the second term involving $b_{\frac{n}{\ell}}$ is omitted throughout). Thus, we have that $$T_\ell \theta_{p^m} f = \ell \theta_{p^m} T_\ell f$$ for all $f \in M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$ using the fact that $k(m)-2 = 2p^{m-1}(p-1)$ is divisible by $p^{m-1}(p-1)$ so that $$\ell^{k(m)-2} \equiv 1 \pmod{p^m}.$$ [*(iii)*]{} Now suppose that $m\ge 2$, that $f \in M_k(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$ with $f \not\equiv 0 \pmod p$, and suppose further that $p\nmid k$ and $w_p(f)= k$. Assume that we had $w_{p^m}(\theta f) < k + 2 + 2p^{m-1}(p-1) = k+k(m)$, i.e., that there exist a form $g\in M_{k'}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$ with $g=f$ as forms with coefficients in ${{\mathbb Z}/{p^m}{\mathbb Z}}$, and where $k'<k + k(m)$. We then know that $$k' \equiv k + k(m) \pmod{p^{m-1}(p-1)} ,$$ say $k+k(m) = k'+t\cdot p^{m-1}(p-1)$ with $t\ge 1$ (see [@ckw Corollary 22]; note we use the fact that modular forms in $M_k(N,{\mathbb Z}/p^m{\mathbb Z})$ can be lifted to classical modular forms over ${\mathbb Z}_p$ which is what is used in loc. cit. and that $N$ is prime to $p$). Putting $$h := E_{p-1}^{p^{m-1}(t-1)} g$$ we find that $$\theta f = E_{p-1}^{p^{m-1}} h$$ as an equality of forms in $M_{k+k(m)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. If we combine this with , we obtain: $$2kp^{m-1} E_{p-1}^{t_{m-1}} f_{m-1} f = -E_{p-1}^{p^{m-1}} h + \frac{1}{12} E_{p-1}^{2p^{m-1}} \partial f - 2k f \sum_{j=0}^{m-2} p^j E_{p-1}^{p^{m-j-1} t_j} f_j .$$ If we now use the fact that $p\nmid k$, that $p$ is odd, and that, as is easily checked, we have $$t_{m-1} = (k(m)-k_{m-1})/(p-1) = (p^m - 3 p^{m-1} + 2)/(p-1)< p^{m-1} ,$$ as well as $t_{m-1} < p^{m-j-1} t_j$ for $j=0,\ldots, m-2$, we deduce that $$p^{m-1} E_{p-1}^{t_{m-1}} f_{m-1} f = E_{p-1}^{t_{m-1}+1} h'$$ for some $h' \in M_{k+k(m)-(p-1)(t_{m-1}+1)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$. Hence we must have $h' \in p^{m-1} M_{k+k(m)-(p-1)(t_{m-1}+1)}(N,{{\mathbb Z}/{p^m}{\mathbb Z}})$, say $h'=p^{m-1}h''$, so that $$E_{p-1}^{t_{m-1}} f_{m-1} f \equiv E_{p-1}^{t_{m-1}+1} h'' \pmod{p} ,$$ and hence $$f_{m-1} f \equiv E_{p-1} h'' \pmod{p} .$$ It follows that $$w_p(f_{m-1} f) < k+k(m) - t_{m-1}(p-1) = k+k_{m-1} = k+p^{m-1}(p+1) .$$ Now recall (from Proposition \[prop:G\_2\_modpm\]) that $f_{m-1} = G_{p+1}^{p^{m-1}}$. As $G_{p+1} = -\frac{B_{p+1}}{2(p+1)} E_{p+1}$ with $\frac{B_{p+1}}{2(p+1)}$ invertible modulo $p$, we deduce $$w_p(E_{p+1}^{p^{m-1}} f) < k+p^{m-1}(p+1) .$$ However, as $w_p(f) = k$ by assumption (and as $p\nmid k$ so that, in particular, $k\neq p$), this conclusion contradicts Lemma \[lemma:filtration\_e\_p+1\_phi\] below. \[lemma:filtration\_e\_p+1\_phi\] Suppose that $p\neq \kappa\in{\mathbb N}$ and that $0\neq \phi \in M_\kappa(N,{\mathbb Z}/p{\mathbb Z})$ with $w_p(\phi) = \kappa$. Then, for $a\in {\mathbb N}$ we have: $$w_p(E_{p+1}^a \phi) = w_p(\phi) + a (p+1) .$$ By induction on $a$ it is clearly enough to prove the case $a=1$. Hence, let us assume $a=1$. Assume for a contradiction that we had $w_p(E_{p+1} \phi) < \kappa + p+1$. Then $$E_{p+1} \phi = E_{p-1} \psi$$ for some $\psi \in M_{2+\kappa}(N,{\mathbb Z}/p{\mathbb Z})$. Suppose first that $N \ge 5$. Let $\mathcal{M}_k(N,R)$ denote the space of modular forms of weight $k$ on $\Gamma_1(N)$ over $R$ as defined in [@di] using Katz’s definition. One has an injection $$M_k(N,R) \rightarrow \mathcal{M}_k(N,R)$$ sending classical modular forms over $R$ to Katz modular forms over $R$. Under the hypothesis $N \ge 5$, we have from [@di Theorem 12.3.7]) that: 1. $\mathcal{M}_k(N,{\mathbb Z}_p) \cong M_k(N,{\mathbb Z}_p)$ (as ${\mathbb Z}_p$ is flat over ${\mathbb Z}$) 2. $\mathcal{M}_k(N,{\mathbb Z}/p^m{\mathbb Z}) \cong M_k(N,{\mathbb Z}/p^m{\mathbb Z})$ if $k > 1$ and $N$ is prime to $p$. For $M \ge 3$, let $\mathcal{M}_k(\Gamma(M),R)$ denote the space of modular forms of weight $k$ on $\Gamma(M)$ over $R$ as defined by Katz in [@katz_padic], where $\Gamma(M)$ denotes naive full level $M$-structure. Regard the above identity as an identity of Katz modular forms on $\Gamma_1(N)$ over ${\mathbb Z}/p{\mathbb Z}$ and let us use some results from [@katz_modp]: By the remark after Lemma 1 of [@katz_modp], the forms $E_{p-1}$ and $E_{p+1}$ are without a common zero (the results in [@katz_modp] are for modular forms on $\Gamma(N)$ which implies the result for $\Gamma_1(N)$). But, by a theorem of Igusa [@igusa], $E_{p-1}$ vanishes to order $1$ at every supersingular point of $X_1(N)$ (see [@gross second paragraph following (4.6)]). Hence the equality $E_{p+1} \phi = E_{p-1} \psi$ means that $\phi$ vanishes at every zero of $E_{p-1}$. Thus, we must have ($\kappa > p-1$ and) $\phi = E_{p-1} \eta$ for some $\eta \in \mathcal{M}_{\kappa-(p-1)}(N,{\mathbb Z}/p{\mathbb Z})$. By hypothesis we have $\kappa-(p-1) \neq 1$ and so by (K2) we have that $\eta$ is classical in the sense that $\eta \in M_{\kappa-(p-1)}(N,{\mathbb Z}/p{\mathbb Z})$. But then $w_p(\phi) \le \kappa - (p-1)$, contrary to hypothesis. Suppose then that $N\le 4$. We can reduce to the previous case as follows. Choose an auxiliary prime $q\ge 5$ such that $q \nmid Np$ and such that $p \nmid t:= q^2-1$, and regard $\phi$ as a Katz modular form on $\Gamma_1(Nq)$. The above argument then shows that under the assumption $w_p(E_{p+1}\phi) < \kappa + p+1$ we will have $\phi = E_{p-1} \eta$ for some $\eta \in M_{\kappa-(p-1)}(Nq,{\mathbb Z}/p{\mathbb Z})$. Choose forms $\tilde\phi \in M_{\kappa}(N,{\mathbb Z}_p)$ and $\tilde\eta \in M_{\kappa-(p-1)}(N,{\mathbb Z}_p)$ that reduce to $\phi$ and $\eta$, respectively. Also, let $\zeta$ be a primitive $Nq$’th root of unity, and let $O$ be the ring of integers of ${\mathbb Q}(\zeta)$ completed at a prime ${\mathfrak p}$ above $p$. Let $\gamma_1,\ldots,\gamma_t$ be a set of left coset representatives for $\Gamma_1(Nq)$ in $\Gamma_1(N)$. We then have the trace map $$f \mapsto \operatorname{tr}f := \sum_{i=1}^t f\mid_k \gamma_i$$ as a linear map $M_k(Nq,{\mathbb C}) \rightarrow M_k(N,{\mathbb C})$ for any $k$. However, $\operatorname{tr}$ as defined above in fact also defines a a linear map $M_k(Nq,O) \rightarrow M_k(N,O)$: This follows as we know that if $f \in M_k(Nq,O)$ is a form for which the Fourier expansion (at $\infty$) has coefficients in $O$ then the same is true for $f\mid_k \gamma$ for any $\gamma\in\operatorname{SL}_2({\mathbb Z})$. To see this, use first that $O = O[1/Nq, \zeta]$ (we have $O[1/Nq] = O$ as $Nq$ is prime to $p$) and $M_k(Nq,O) = M_k(\Gamma_1(Nq),O) \subseteq M_k(\Gamma(Nq),O)$. By [@katz_padic Corollary 1.6.2] together with [@hida (1.4a)], we have that $$M_k(\Gamma(Nq),O) \cong \mathcal{M}_k(\Gamma(Nq),{\mathbb Z}_p).$$ The claim then follows from [@delrap Corollaire 3.13] or [@katz_padic A1.2]. Taking the trace we find the congruence $t\tilde \phi \equiv E_{p-1} (\operatorname{tr}\tilde \eta) \pmod{{\mathfrak p}}$ as a congruence between forms in $M_{\kappa}(N,O)$ (we used that $\tilde \phi$ and $E_{p-1}$ are on $\Gamma_1(N)$). Since $t = [\Gamma_1(N) : \Gamma_1(Nq)] = q^2-1$ is prime to $p$, the congruence shows that $\phi$ coincides with an element in $M_{\kappa-(p-1)}(N,O/{\mathfrak p})$. However, because $M_{\kappa-(p-1)}(N,{\mathbb Z}/p{\mathbb Z}) \otimes O/{\mathfrak p}\cong M_{\kappa-(p-1)}(N,O/{\mathfrak p})$ and $\phi \in M_{\kappa-(p-1)}(N,O/{\mathfrak p})$ has $q$-expansion lying in ${\mathbb Z}/p{\mathbb Z}[[q]]$ and $O/{\mathfrak p}$ is free over ${\mathbb Z}/p{\mathbb Z}$, it follows that $\phi$ in fact lies in $M_{\kappa-(p-1)}(N,{\mathbb Z}/p{\mathbb Z})$. But this contradicts $w_p(\phi) = \kappa$. I. Chen, I. Kiming, G. Wiese: ‘On modular Galois representations modulo prime powers’, Int. J. Number Theory **9** (2013), 91–113. I. Chen, I. Kiming, J. B. Rasmussen: ‘On congruences mod ${\mathfrak p}^m$ between eigenforms and their attached Galois representations’, J. Number Theory (3) **130** (2010), 608–619. R. Colemen and W. Stein: ‘Approximation of eigenforms of infinite slope by eigenforms of finite slope’, Geometric aspects of Dwork theory. Vol. I, II, 437–449, Walter de Gruyter GmbH & Co. KG, Berlin, 2004. P. Deligne, M. Rapoport: ‘Les schémas de modules de courbes elliptiques’, Modular functions of one variable, II (Proc. Internat. Summer School, Univ. Antwerp, Antwerp, 1972), pp. 143–316. Lecture Notes in Math., Vol. **349**, Springer, Berlin, 1973. F. Diamond and J. Im: ‘Modular forms and modular curves’, Seminar on Fermat’s Last Theorem (Toronto, ON, 1993–1994), CMS Conf. Proc. **17** (1995), 39–133. Amer. Math. Soc. 1995. F. Diamond and J. Shurman: ‘A first course in modular forms’, Grad. Texts in Math. **228**, Springer 2005. B. Edixhoven: ‘The weight in Serre’s conjectures on modular forms’, Invent. Math. (3) **109** (1992), 563–594. B. Gross: ‘A tameness criterion for Galois representations associated to modular forms mod $p$’, Duke Math. J. **61**, no.2 (1990), 445–517. H. Hida: ‘Iwasawa modules attached to congruences of cusp forms’, Ann. Sci. École Norm. Sup. (4) **19** (1986), 231–273. J.-I. Igusa: ‘Class number of a definite quaternion with prime discriminant’, Proc. Natl. Acad. Sci. USA **44** (1958), 312–314. N. Jochnowitz: ‘Congruences between systems of eigenvalues of modular forms’, Trans. Amer. Math. Soc. **270** (1982), 269–285. N. M. Katz: ‘$p$-adic properties of modular schemes and modular forms’, Modular functions of one variable, III (Proc. Internat. Summer School, Univ. Antwerp, Antwerp, 1972), pp. 69–190. Lecture Notes in Mathematics, Vol. **350**, Springer, Berlin, 1973. N. M. Katz: ‘A result on modular forms in characteristic $p$’, in J.-P. Serre, D. B. Zagier (eds.): Modular functions of one variable V, Lecture Notes in Math. **601** (1977), 53–61. J.-P. Serre: ‘Congruences et formes modulaires \[d’après H. P. F. Swinnerton-Dyer\]’, Séminaire Bourbaki, Lecture Notes in Math. **317**, 319–338, Springer 1973. J.-P. Serre: ‘Formes modulaires et fonctions zêta $p$-adiques, in W. Kuyk, J.-P. Serre: Modular functions of one variable III, Lecture Notes in Math. **350**, Springer 1973. L. C. Washington: ‘Introduction to cyclotomic fields’, Grad. Texts in Math. **83**, Springer 1982.
{ "pile_set_name": "ArXiv" }
--- abstract: | We simulate the inner 100pc of the Milky-Way Galaxy to study the formation and evolution of the population of star clusters and intermediate mass black holes. For this study we perform extensive direct $N$-body simulations of the star clusters which reside in the bulge, and of the inner few tenth of parsecs of the super massive black hole in the Galactic center. In our $N$-body simulations the dynamical friction of the star cluster in the tidal field of the bulge are taken into account via (semi)analytic soluations. The $N$-body calculations are used to calibrate a (semi)analytic model of the formation and evolution of the bulge. We find that $\sim 10$% of the clusters born within $\sim100$pc of the Galactic center undergo core collapse during their inward migration and form intermediate-mass black holes (IMBHs) via runaway stellar merging. After the clusters dissolve, these IMBHs continue their inward drift, carrying a few of the most massive stars with them. We predict that region within $\sim10$ parsec of the SMBH is populated by $\sim 50$ IMBHs of $\sim 1000$ . Several of these are expected to be accompanied still by some of the most massive stars from the star cluster. We also find that within a few milliparsec of the SMBH there is a steady population of several IMBHs. This population drives the merger rate between IMBHs and the SMBH at a rate of about one per 10Myr, sufficient to build the accumulate majority of mass of the SMBH. Mergers of IMBHs with SMBHs throughout the universe are detectable by LISA, at a rate of about two per week. author: - 'Simon F. Portegies Zwart, Holger Baumgardt, Stephen L. W. McMillan, Junichiro Makino, Piet Hut' - Toshi Ebisuzaki date: 'Received 2005 August 1; in original form 1687 October 3.6; Accepted xxxx xxx xx.' title: The ecology of star clusters and intermediate mass black holes in the Galactic bulge --- \#1[[**\[\#1 – Steve\]**]{}]{} \#1[[**\[\#1 – Steve\]**]{}]{} \#1[[**\[\#1 – Simon\]**]{}]{} \#1[[**\[\#1 – Simon\]**]{}]{} \#1[[**\[\#1 – Holger\]**]{}]{} \#1[[**\[\#1 – Holger\]**]{}]{} \#1[[**\[\#1 – Jun\]**]{}]{} \#1[[**\[\#1—Jun\]**]{}]{} \#1[[**\[\#1—Piet\]**]{}]{} \#1[[**\[\#1—Piet\]**]{}]{} \[firstpage\] Introduction ============ In recent years the Galactic center has been explored extensively over most of the electromagnetic spectrum, revealing complex structures and a multitude of intriguing physical phenomena. At the center lies a $\sim3.7\times10^6$ solar mass () black hole [@1997MNRAS.284..576E; @1998ApJ...509..678G; @2000Natur.407..349G]. The presence of a water-rich dust ring at about one parsec from SgrA\* . further underscores the complexity of this region, as does the presence within the central parsec of a few million year old population of very massive Ofpe/WN9 [@1993ApJ...414..573T] and luminous blue variable stars . These young stars may indicate recent star formation in the central region [@1993ApJ...408..496M; @2005astro.ph..7687N], or they may have migrated inward from larger distances to their current locations [@2001ApJ...546L..39G]. In addition, the [*Chandra*]{} X-ray Observatory has detected an unusually large number ($\apgt 2000$) of hard X-ray (2–10 keV) point sources within 23pc of the Galactic center [@2003ApJ...589..225M]. Seven of these sources are transients, and are conjectured to contain stellar-mass black holes [@2004astro.ph.12492M]; some may even harbor IMBHs [@2001ApJ...558..535M]. The Galactic center is a dynamic environment, where young stars and star clusters form in molecular clouds or thick dusty rings [@2004astro.ph..9541N; @2005astro.ph..7687N], and interact with their environment. Several star clusters are known to exist in this region [@1999ApJ...525..750F], and the star formation rate in the inner bulge is estimated to be comparable to that in the solar neighborhood [@2001ApJ...546L.101P], enough to grow the entire bulge over the age of the Galaxy. Of particular interest here are the several star clusters discovered within $\sim 100$pc of the Galactic center, 11 of which have reliable mass estimates . Most interesting of these are the two dense and young ($\aplt 10$Myr) star clusters Arches [@2002ApJ...581..258F] and the Quintuplet[@1999ApJ...514..202F], and the recently discovered groups IRS13E and IRS16SW [@2005astro.ph..4276L]. In this paper we study the relation between the star clusters in the inner $\sim 100$pc of the Galactic center and, to some extend, the partial formation of the central supermassive black hole. In particular we simulate the evolution of the star clusters born over a range of distances from the Galactic center. While we follow their internal dynamical evolution we allow the star clusters to spiral inwards towards the Galactic center until they dissolve in the background. During this process a runaway collision may have occurred in the cluster and we follow the continuing spiral-in of the resulting intermediate mass black hole. Our prescription for building an intermediate mass black hole has been well established in numerous papers concerning stellar collision runaways in dense star clusters . We just build on these earlier results for our description of the collision runaway and the way in which it leads to the formation of a black hole of intermediate mass. Eventually the IMBHs merge with the supermassive black hole, building the SMBH in the process. This model was initially proposed by [@2001ApJ...562L..19E], and here we validate the model by detailed simulations of the dynamical evolution of individual star clusters and the final spiral-in of the IMBH toward the SMBH. Using the results of the direct N-body simulations we calibrate a semi-analytic model to simulate a population of star clusters which are born within $\sim 100$pc over the age of the Galaxy. Collision Runaways and Cluster Inspiral {#Sect:picture} ======================================= A substantial fraction of stars are born in clusters and these have a power-law stellar mass functions fairly well described by a “Salpeter” exponent of -2.35, and with stellar masses ranging from the hydrogen burning limit ($\sim 0.08$) or a bit above [@2005ApJ...628L.113S] to an upper limit of $\sim100\,{\mbox{${\rm M}_\odot$}}$ or possibly as high as 150[@2005Natur.434..192F]. The massive stars start to sink to the cluster center immediately after birth, driving the cluster into a state of core collapse on a time scale ${{\mbox{$t_{\rm cc}$}}}\simeq 0.2{\mbox{${t_{\rm rh}}$}}$ [@2002ApJ...576..899P; @2004ApJ...604..632G], where [@1971ApJ...166..483S] $${\mbox{${t_{\rm rh}}$}}\simeq 2\,{\rm Myr} \left( {r \over [{\rm pc}]} \right)^{3/2} \left( {m \over [{\mbox{${\rm M}_\odot$}}]} \right)^{-1/2} {n \over {\mbox{${\log \lambda}$}}}.$$ Here $m$ is the cluster mass, $r$ is its half-mass radius, $n$ is the number of stars, and ${\mbox{${\log \lambda}$}}\simeq \log(0.1n)\sim10$. In sufficiently compact clusters the formation of a dense central subsystem of massive stars may lead to a “collision runaway,” where multiple stellar mergers result in the formation of an unusually massive object . If the mass of this runaway grows beyond $\sim300$ it collapses to an IMBH without losing significant mass in a supernova explosion [@2003ApJ...591..288H]. Recently, this model has been applied successfully to explain the ultraluminous X-ray source associated with the star cluster MGG-11 in the starburst galaxy M82 [@2004Natur.428..724P]. This model for creating an intermediate mass black hole in a dense star cluster was adopted by [@2005ApJ...628..236G], who continued by studying the evolution of massive $\apgt 10^6$ star clusters within about 60pc from the Galactic center. Their conclusions are consistent with the earlier $N$-body models [@2000ApJ...545..301K; @2003ApJ...593..352P; @2003ApJ...596..314M; @2004ApJ...607L.123K] and analytic calculations [@2001ApJ...546L..39G] in that massive clusters can reach the galactic center but in doing so they populate the inner few parsecs with a disproportionately large number of massive stars. The main requirement for a successful collision runaway is that the star cluster must experience core collapse (i) before the most massive stars explode as supernovae ($\sim3$Myr) and (ii) before the cluster dissolves in the Galactic tidal field. The collisional growth rate slows dramatically once the runaway collapses to an IMBH. We estimate the maximum runaway mass achievable by this process as follows. For compact clusters (${\mbox{${t_{\rm rh}}$}}\aplt100$ Myr), essentially all the massive stars reach the cluster core during the course of the runaway, and the runaway mass scales with the cluster mass: ${\mbox{${m_{\rm r}}$}}\simeq8\times10^{-4}m\,{\mbox{${\log \lambda}$}}$ [@2002ApJ...576..899P]. For systems with longer relaxation times, only a fraction of the massive stars reach the core in time and the runaway mass scales as $m{\mbox{${t_{\rm rh}}$}}^{-1/2}$ [@2004astro.ph.12622M] (see their Eq. 11). The relaxation based argument may result in higher mass runaways in star clusters with a very small relaxation time compared to the regime studied in Monte Carlo N-body simulations [@2004astro.ph.12622M]. A convenient fitting formula combining these scalings, calibrated by N-body simulations for Salpeter-like mass functions, is [@2002ApJ...576..899P; @2004astro.ph.12622M] $${\mbox{${m_{\rm r}}$}}\sim 0.01 m \left( 1 + \frac{{\mbox{${t_{\rm rh}}$}}}{\rm 100 Myr} \right)^{-1/2}\,.$$ Early dissolution of the cluster reduces the runaway mass by prematurely terminating the collision process. As core collapse proceeds, the orbit of the cluster decays by dynamical friction with the stars comprising the nuclear bulge. The decay of a circular cluster orbit of radius $R$ is described by (see \[Eq. 7-25\] in [@1987gady.book.....B], or [@2003ApJ...596..314M] for the more general case): $$\frac{dR}{dt} = -0.43 {Gm {\mbox{${\log \Lambda}$}}\over R^{(\alpha + 1)/2} v_c}\,, \label{Eq:df}$$ where $v_c^2 = GM(R)/R$, $\alpha = 1.2$, $M(R)$ is the mass within a distance $R$ from the Galactic center and we take ${\mbox{${\log \Lambda}$}}\sim8$ [@2003MNRAS.344...22S]. Numerical solution of this equation is required due to the complicating effects of stellar mass loss, which drives an adiabatic expansion of the cluster, and by tidal stripping, whereby the cluster mass tends to decrease with time according to $m(t) = m_0(1 - \tau/{\mbox{${t_{\rm dis}}$}})$, [@2002ApJ...576..899P]. Here $m_0$ is the initial mass of the cluster, $\tau$ is the cluster age in terms of the instantaneous relaxation time () within the Jacobi radius, and ${\mbox{${t_{\rm dis}}$}}$ is the time scale for cluster disruption: ${\mbox{${t_{\rm dis}}$}}\simeq 0.29{\mbox{${t_{\rm rJ}}$}}$ [^1]. Even after the bulk of the cluster has dissolved, a dense stellar cusp remains surrounding the newborn IMBH, and accompanies it on its descent toward the Galactic center. The total mass of stars in the cusp is typically comparable to that of the IMBH itself [@2004ApJ...613.1143B] and it is composed predominantly of massive stars, survivors of the population that initiated the core collapse during which the IMBH formed. Eventually even that cusp slowly decays by two-body relaxation [@2003ApJ...593L..77H], depositing a disproportionately large number of massive stars and the orphaned IMBH close to the Galactic center [@2005ApJ...628..236G]. Ultimately, the IMBH merges with the SMBH. Simulating star clusters within $\sim 100$pc from the Galactic center {#Sect:Nbody} ===================================================================== We have performed extensive direct N-body calculations to test the validity of the general scenario presented above, and to calibrate the semi-analytic model. Our analysis combines several complementary numerical, analytical and theoretical techniques in a qualitative model for the formation and evolution of the nuclear bulge of the Milky Way Galaxy. The semi-analytical model outlined in Sect.\[Sect:picture\], and which is based on equation \[Eq:df\] of [@2003ApJ...596..314M], is based on simple characterizations of physical processes, which we calibrate using large-scale N-body simulations. The initial conditions for these simulations are selected to test key areas in the parameter space for producing IMBHs in the inner $\sim 100$pc of the Galactic center. The N-body calculations employ direct integration of Newton’s equations of motion, while accounting for complications such as dynamical friction and tidal effects due to the Galactic field, stellar and binary evolution, physical stellar sizes and the possibility of collisions, and the presence of a supermassive black hole in the Galactic center. Two independent but conceptually similar programs are used: (1) the “kira” integrator, part of the Starlab software environment (see [http://www.manybody.org/$\sim$manybody/starlab.html]{}, [@2001MNRAS.321..199P]), and (2) NBODY4\ (see [http://www.sverre.org]{}) [@Aarseth2003]. Both codes achieve their greatest speed, as in the simulations reported here, when run in conjunction with the special-purpose GRAPE-6 (see [http://www.astrogrape.org]{}) hardware acceleration [@2003PASJ...55.1163M]. Both kira and NBODY4 incorporate parametrized treatments of stellar evolution and allow for the possibility of direct physical collisions between stars, thus including the two key physical elements in the runaway scenario described here (see also [[email protected]]). A collision occurs if the distance between two stars becomes smaller than the sum of the stellar radii, except that, for collisions involving black holes, we use the tidal radius instead. During a collision mass and momentum are conserved. These are reasonable assumptions since the relative velocity of any two colliding stars is typically much smaller than the escape speed from either stellar surface [@2003MNRAS.345..762L; @2005MNRAS.358.1133F]. We performed N-body simulations of star clusters containing up to 131,072 stars and starting at $R=1$pc, 2, 4, 10 and 100pc from the Galactic center, with various initial concentrations ($W_0=6$ and 9) and with lower limits to the initial mass function of 0.1 and 1. These simulations were carried out as part of the calibration of the semi-analytic model which we presented in Sect.\[Sect:analytic\]. One such comparison is presented in Figure\[fig:N-body\], which shows the orbital evolution and runaway growth in a star cluster born with 65536 stars in a circular orbit at a distance of 2pc from the Galactic center. The solid lines in the figure result from the semi-analytic model (based on equation \[Eq:df\] and [@2003ApJ...596..314M]), while the high precision N-body calculations are represented by dotted lines. They match quite well, indicating that the simple analytic model produces satisfactory results in reproducing the general features and physical scales of the evolution. As the cluster in Figure\[fig:N-body\] sinks toward the Galactic center, it produce one massive star through the collision runaway process. In Figure\[fig:image\] we show a snapshot of this simulations projected in three different planes at an age of 0.35Myr. By this time $\sim 30\%$ of the cluster has already dispersed and its stars have spread out into the shape of a disk spanning the inward-spiraling orbit. By the time of Figure \[fig:image\], a $\sim 1100$ collision runaway star has formed in the cluster center; this object subsequently continues to grow by repeated stellar collisions. The growth of the collision runaway is indicated by the dotted line in Figure\[fig:N-body\] running from bottom left to top right (scale on the right vertical axis). By an age of about 0.7Myr the cluster is almost completely disrupted and the runaway process terminates. After the cluster dissolves, the IMBH continues to sink toward the Galactic center, still accompanied by 10–100 stars which initially were among the most massive in the cluster. Near the end of its lifetime, the runaway star loses about 200 in a stellar wind and subsequently collapses to a $\sim 1000$ IMBH at about 2.4Myr. The IMBH and its remaining stellar companions continue to sink toward the Galactic center. The continuing “noise” in the dotted curve in Figure \[fig:N-body\] reflects the substantial eccentricity of the IMBH orbit. At an age of 2.5–3Myr, the remnant star cluster consisting of an IMBH orbited by a few of the most massive stars, quite similar to the observed star cluster IRS13, arrives in the inner 0.1 pc of the Galaxy (see sect.\[Sect:Observations\]). Merger with the central black hole {#Sect:finalparsec} ================================== When the IMBH arrives within about 0.1pc of the Galactic center the standard formula for dynamical friction[@1987gady.book.....B] is becoming unreliable, as the background velocity dispersion increases and the effects of individual encounters become more significant. It is important, however, to ascertain whether the IMBH spirals all the way into the SMBH, or if it stalls in the last tenth of a parsec, as higher-mass black holes may tend to do [@2005ApJ...621L.101M]. To determine the time required for the IMBH to reach the central SMBH, we have performed additional N-body calculations, beginning with a 1000 and a 3000 IMBH in circular orbits at a distance of 0.1pc from the Galactic center. Both IMBHs are assumed to have shed their parent cluster by the start of the simulation. The inner parsec of the Galaxy is modeled by 131,071 identical stars with a total stellar mass of $4\times 10^6$, distributed according to a $R^{-1.75}$ density profile; a black hole of $3\times 10^6$ resides at the center. The region within a milliparsec of the central SMBH is depleted of stars in our initial conditions. This is supported by the fact that the total Galactic mass inside that radius, excluding the central SMBH is probably less than $10^3\,{\mbox{${\rm M}_\odot$}}$ [@1998ApJ...509..678G; @2003ApJ...594..812G]. We stop the calculations as soon as the IMBH reaches this distance. Figure\[fig:final\_parsec\] (see also the dotted line in Figs.\[fig:N-body\]) shows the orbital evolution of the 1000 and 3000 IMBHs in our simulations. Although the black-hole orbits are initially circular, eccentricities on the order of $\aplt 0.6$ are induced quite quickly by close encounters with field stars. The rate of spiral-in near the SMBH is smaller than farther out, because the increasing velocity dispersion tends to reduce the effect of dynamical friction and because the IMBH reaches the inner depleted area. The central milliparsec was initially empty in our simulations, and there was insufficient time to replenish it during our calculations. It is unlikely that sufficient stellar mass exists within this region for dynamical friction to drive the IMBH much closer to the SMBH. (Interestingly, this distance is comparable to the orbital semi-major axis of the star S0-2, which is observed in a 15 year orbit around the Galactic center [@2003ApJ...586L.127G].) The time scale for a 1 mpc orbit to decay by gravitational radiation exceeds the age of the Galaxy for circular motion, so unless the IMBH orbit is already significantly eccentric, or is later perturbed to higher eccentricity ($\apgt 0.9$ to reduce the merger time to $\aplt 10^9$years) by encounters with field stars or another IMBH, the orbital decay effectively stops near the central SMBH. While the IMBH stalls, another star cluster may form, sink toward the Galactic center, and give rise to a new IMBH which subsequently arrives in the inner milliparsec (see Sec. 5). This process will be repeated for each new IMBH formed, until interactions become frequent enough to drive a flux of IMBHs into the loss cone where gravitational radiation can complete the merger process. We can estimate the number of IMBHs in a steady state in the inner few milliparsecs of the SMBH. The time scale for a close (90 degree deflection) encounter in a system of $n_{\rm IMBH}$ IMBHs is $$t_{\rm close} \sim \left(\frac{M_{\rm SMBH}}{m_{\rm IMBH}}\right)^2 \frac{t_{\rm orb}}{n_{\rm IMBH}}\,,$$ where $M_{\rm SMBH}$ and $m_{\rm IMBH}$ are the masses of the SMBH and the IMBH, respectively, and $t_{\rm orb} \sim 1-10$ years is the typical orbital period at a distance of 1 mpc from the SMBH. For $M_{\rm SMBH} \sim 10^6{\mbox{${\rm M}_\odot$}}$ and $m_{\rm IMBH} \sim 10^3{\mbox{${\rm M}_\odot$}}$, we find $t_{\rm close} \sim 1-10 \times 10^6/n_{\rm IMBH}$ years, comparable to the in-fall time scale unless $n_{\rm IMBH}$ is large. Close encounters are unlikely to eject IMBHs from the vicinity of the Galactic center, but they do drive the merger rate by replenishing the loss cone around the SMBH [@2004ApJ...606..788M; @2004ApJ...616..221G]. As IMBHs accumulate, the cusp around the SMBH eventually reaches a steady state in which the merger rate equals the rate of in-fall, with a roughly constant population of a few IMBHs within about a milliparsec of the SMBH. A comparable analysis was performed by [@2004ApJ...606L..21A] for stellar mass black holes around the SMBH, and if we scale their results to IMBHs we arrive at a similar steady state population. The evolution of a population of star clusters {#Sect:analytic} ============================================== We now turn to the overall evolution of the population of clusters which gave rise to the nuclear bulge. We have performed a Monte-Carlo study of the cluster population, adopting a star formation rate that declines as $1/t$ over the past 10 Gyr [@2004Natur.428..625H]. Cluster formation times are selected randomly following this star formation history, and masses are assigned as described below, until the total mass equals the current mass of the nuclear bulge within 100 pc of the Galactic center—about $10^9$. The total number of clusters thus formed is $\sim 10^5$ over the 10 Gyr period. For each cluster, we select a mass ($m$) randomly from a cluster initial mass function which is assumed to follow the mass distribution observed in starburst galaxies—a power-law of the form $N(m) \propto m^{-2}$ between $\sim10^3$ and $\sim10^7$ [@1999ApJ...527L..81Z]. The distance to the Galactic center ($R$) is again selected randomly, assuming that the radial distribution of clusters follows the current stellar density profile in the bulge between 1pc and 100pc[@1997MNRAS.284..576E; @1998ApJ...509..678G]. The current distribution of stars must reflect the formation distribution to a large extent, because most stars’ orbits don’t evolve significantly, but only the orbits of the more massive stellar clusters. The initial density profiles of the clusters are assumed to be $W_0=6$–9 King models. This choice of high-concentrated King models is supported by the recent theoretical understanding by [@2004ApJ...608L..25M] of the relation between age and core-radius for young star clusters in the large Magellanic cloud observed by [@2003MNRAS.338...85M]. We establish a cluster mass-radius relation by further assuming that clusters are born precisely filling their Jacobi surfaces in the Galactic tidal field. This provides a lower limit to the fraction of clusters that produce an IMBH and sink to the Galactic center. The evolution of each cluster, including specifically the moment at which it undergoes core collapse, the mass of the collision runaway (if any) produced, and the distance from the Galactic center at which the cluster dissolves, is then calculated deterministically using our semi-analytic model. After cluster disruption, the IMBH continues to sink by dynamical friction, eventually reaching the Galactic center. Results of the cluster population model {#Sect:Results} --------------------------------------- Figure\[fig:MC\] summarizes how the fates of the star clusters in our simulation depend on $m$ and $R$. Open and filled circles represent initial conditions that result in an IMBH reaching the central parsec by the present day. The various lines define the region of parameter space expected to contribute to the population of IMBHs within the central parsec, as described in the caption. Here we emphasize that our results depend linearly on the fraction of stars in the bulge that form in star clusters. The number of star clusters and IMBHs is proportional to this factor, which is not necessarily constant with time. Bear in mind also that, though theoretical uncertainties are about a factor of two, the systematic uncertainties can be much larger and depend critically on various assumptions in the models, like the amount of mass loss in the stellar wind of the collision product and the fate of the stellar remnant in the supernova explaion. The results of our calculations may be summarized as follows: 1. 5% – 10% of star clusters born within 100pc of the Galactic center produce an IMBH. 2. The mean mass of IMBHs now found in the inner 10pc is $\sim 1000$, whereas IMBHs between 90 and 100pc average $\sim 500$. 3. Over the age of the Galaxy ($\sim10^{10}$ years) a total of 1000–3000 IMBHs have reached the Galactic center, carrying a total mass of $\sim 1\times10^6{\mbox{${\rm M}_\odot$}}$. Here the range in masses stems from variations in the adopted stellar mass function. 4. At any instant, approximately $\sim 50$ IMBHs reside in the inner 10pc, about ten times that number lie within the nuclear star cluster (inner 30pc), and several lie within the innermost few tenths of a parsec. 5. One in every $\sim 30$ IMBHs is still accompanied by a remnant of its young (turn-off mass $\apgt 10$) star cluster when it reaches the inner parsec, resulting in a few IMBHs at any time in the inner few parsecs with young stars still bound to them, much like IRS13E or IRS16SW. On the basis of our N-body simulations of the central 0.1 pc in Sect.\[Sect:finalparsec\] we expect that the majority of IMBHs which arrive in the Galactic center eventually merge with the SMBH on a time scale of a few Myr, driven by the emission of gravitational radiation and interactions with local field stars and other IMBHs. In our simulations the in-fall rate has increased over the lifetime of the Galaxy (following our assumed star formation rate), from one arrival per $\sim 20$Myr to the current value of one every $\sim 5$Myr, with a time average IMBH in-fall rate of roughly one per $\sim 7$Myr. (A lower minimum mass in the initial mass function produces higher in-fall rates.) Some of the field stars near the SMBH may be ejected from the Galactic center with velocities of up to $\sim 2000$km/s following encounters with the hard binary system formed by the IMBH and the central SMBH [@1988Natur.331..687H; @2003ApJ...599.1129Y; @2005MNRAS.363..223G]. Support for this possibility comes from the recent discovery of SSDSJ090745.0+024507, a B9 star with a measured velocity of 709km/s directly away from the Galactic center [@2005ApJ...622L..33B]. IMBHs are potentially important sources of gravitational wave radiation. A merger between a 1000 IMBH and a $\sim 3 \times 10^6$ SMBH would be detectable by the LISA gravitational wave detector to a distance of several billion parsecs. Assuming that the processes just described operate in most spiral galaxies, which have a density of roughly 0.002 Mpc$^{-3}$ [@2004MNRAS.353..713K], we estimate a detectable IMBH merger rate of around two per week, with a signal to noise $\sim 10^3$. The current cluster population {#Sect:currentclusters} ------------------------------ Our semi-analytic model for the evolution of star clusters in the inner $\sim 100$pc of the Galaxy yields a steady-state distribution of cluster masses which we can compare with observed star clusters in the vicinity of the Galactic center. Figure\[fig:Borissova\] compares the observed mass distribution of young star clusters in the bulge with our steady-state solution. The data include the Arches cluster [@2002ApJ...581..258F], the Quintuplet [@1999ApJ...514..202F], IRS13E , IRS16SW [@2005astro.ph..4276L], and 7 recently discovered star clusters with reliable mass estimates . For comparison we show a realization of the present-day population of star cluster masses generated by our semi-analytic model. Using the adopted declining star-formation rate from Sect.\[Sect:analytic\] (see [@2004Natur.428..625H]), we find about $\sim 50$ star clusters within the central 100 pc at any given time, consistent with the earlier prediction of [@2001ApJ...546L.101P]. Assuming a flat (i.e. uniform) star formation rate, we predict $\sim 400$ clusters in the same region, about an order of magnitude more than currently observed. In our semi-analytic model, about 15% of all present-day star clusters host an IMBH or are in the process of producing one. Between 1% and 8% of star clusters with a present-day mass less than $10^4$ contain an IMBH, whereas more than 80% of clusters with masses between 30,000 and $2\times 10^5$ host an IMBH. For more massive clusters the probability of forming an IMBH drops sharply. Finally, we note that we are rather unlikely to find an orphaned very massive ($\apgt 200$) star. During the last 1 Gyr, only about 10–40 of such objects have formed in the inner 100pc of the Galaxy. Lower mass merger products, however, are quite common. The Pistol star[@1999ApJ...514..202F] may be one observational example. Discussion ========== Evolution of the merger product {#Sect:massloss} ------------------------------- One of the main uncertainties in our calculations is whether or not mass gain by stellar collisions exceeds mass loss by stellar winds. Although the accretion rate in our models is very high, mass loss rates in massive stars are uncertain, and it is conceivable that sufficiently high mass loss rates might prevent the merger product from achieving a mass of more than a few hundred . Mass loss in massive ($\apgt 100$) stars may be radiatively driven by optically thin lines. In this case it is possible to derive upper limits to the mass loss. Such calculations, including the von Zeipel [@1924MNRAS..84..665V] effect for stars close to the Eddington-Gamma limit, indicate that stellar wind mass loss rates may approach $10^{-3}$yr$^{-1}$ . If the star is rotating near the critical rate, the mass loss rate may be even larger . Outflow velocities, however, may be so small that part of the material falls back on the equatorial zone, where the mass loss is least . The calculations of match the observed mass loss rates for $\eta$ Carina, which has a peak of $1.6\pm 0.3\times10^{-3}$yr$^{-1}$ (assuming spherical symmetry) during normal outbursts, falling to $10^{-5}$yr$^{-1}$ during the intervening 5.5years . For young ($\aplt4$Myr) O stars in the small Magellanic cloud low ($\aplt 10^{-8}$/yr) mass loss rates were observed , indicating that massive stars may have much lower mass loss rates until they approach the end of their main-sequence lifetimes (see ). Thus it remains unclear whether the periods of high mass loss persist for long enough to seriously undermine the runaway scenario adopted here. We note that the collision runaways in our simulations are initiated by the arrival of a massive star in the cluster core . If such a star grows to exceed $\sim 300$, most collisions occur within the first 1.5Myr of the cluster evolution. The collision rate during the period of rapid growth typically exceeds one per $\sim 10^4$ years, sustained over about 1 Myr, resulting in an average mass accretion rate exceeding $10^{-3}$/yr, comparable to, and possibly exceeding, the maximum mass loss rates derived for massive stars. Furthermore, in our N-body simulations (and in the semi-analytic model), the stellar mass loss rate increases with time, with little mass loss at the zero-age main sequence and substantially more near the end of the main-sequence stellar lifetime ($\dot{m}_{\rm wind} \propto L^{2.2}$) . In other words, mass loss rates are relatively low while most of the accretion is occurring. This prescription for the mass loss rate matches that of evolutionary calculations for massive Wolf-Rayet stars . We also emphasize that a large mass loss rate in the merger product [*cannot*]{} prevent the basic mass segregation and collision process, even though it might significantly reduce the final growth rate . These findings are consistent with recent N-body simulations of small clusters in which the assumed mass loss rate from massive ($>120$) stars exceeded $10^{-3}$/yr [@2004astro.ph.10510B]. The stellar evolution of a runaway merger product has never been calculated in detail, and is poorly understood. However, it is worth mentioning that its thermal time scale significantly exceeds the mean time between collisions. Even if the star grows to $\apgt 10^3$, the thermal time scale will be 1–$4\times 10^4$years, still comparable to the collision rate. The accreting object will therefore be unable to reestablish thermal equilibrium before the next collision occurs. We note in passing that the supermassive star produced in the runaway collision may be hard to identify by photometry if the cluster containing it cannot be resolved: The runaway is mainly driven by collisions between massive stars, which themselves have luminosities close to the Eddington-Gamma limit. Since the Eddington luminosity scales linearly with mass, a collection of luminous blue variables at the Eddington luminosity are comparable in brightness to an equally massive single star. Spectroscopically, however, the collision runaway may be very different. Mass loss in the form of a dense stellar wind before the supernova can dramatically reduce the mass of the final black hole, or could even prevent black hole formation altogether [@2003ApJ...591..288H]. The runaway merger in fig.\[fig:N-body\] develops a strong stellar wind near the end of its lifetime before collapsing to a $\sim1000$ IMBH at $\sim2.4$Myr. It is difficult to quantify the effect of stellar winds on the final IMBH mass because the mass loss rate of such a massive star remains uncertain . However, it is important to underscore here the qualitative results that stellar winds are unable to prevent the occurrence of repeated collisions, and significantly limit the outcome only if the mass loss rate is very high—more than $\sim10^{-3}\,{\mbox{${\rm M}_\odot$}}/{\rm yr}$—and sustained over the lifetime of the star. The star clusters IRS13E and IRS16SW {#Sect:Observations} ==================================== The best IMBH candidate in the milky-way Galaxy was recently identified in the young association IRS13E in the Galactic center region. IRS13E is a small cluster of stars containing three spectral type O5I to O5III and four Wolf-Rayet stars, totaling at most $\sim300\,{\mbox{${\rm M}_\odot$}}$ . (The recently discovered cluster IRS16SW[@2005astro.ph..4276L] also lies near the Galactic center and reveals similarly interesting stellar properties.) Both clusters are part of the population of helium-rich bright stars in the inner parsec of the Galactic center . With a “normal” stellar mass function, as found elsewhere in the Galaxy, stars as massive as those in IRS13E are extremely rare, occurring only once in every $\sim 2000$ stars. However, in the Galactic center, a “top-heavy” mass function may be common [@2004astro.ph..9415F; @2005ApJ...628L.113S]. The mean proper motion of five stars in IRS13E is $\langle v\rangle_{\rm 2D} = 245$km/s;[@2005ApJ...625L.111S] an independent measurement of four of these stars yields $270$km/s . If IRS13E were part of the rotating central stellar disk [@2003ApJ...594..812G], this would place the cluster $\sim 0.12$pc behind the plane on the sky containing the SMBH, increasing its galactocentric distance to about 0.18pc, consistent with a circular orbit around the SMBH at the observed velocity. The five IRS16SW stars have $\langle v\rangle_{\rm 2D} \simeq 205$km/s [@2005astro.ph..4276L], corresponding to the circular orbit speed at a somewhat larger distance ($\sim 0.4$pc). The greatest distance between any two of the five stars in IRS13E with known velocities is $\sim 0.5$seconds of arc (0.02pc at 8.5kpc), providing a lower limit on the Jacobi radius: ${{r_{\rm J}}}\apgt 0.01$pc. It then follows from the Hills equation (${{r_{\rm J}}}^3 \simeq R^3 m/M$) that the minimum mass required to keep the stars in IRS13E bound is about 1300 (see also ). A more realistic estimate is obtained by using the measured velocities of the observed stars, using the expression: $m = \langle v^2 \rangle R/G$. The velocity dispersion of all stars, E1, E2, E3, E4, and E6, is about $\langle v \rangle \simeq 68$–84km/s, which results in a estimated mass mass of about 11000–16000. Such a high mass would be hard to explain with the collision runaway scenario. However, the stars E1 and E6 may not be members. The extinction of the latter star is smaller than that of the other stars, indicating that it may be closer to the sun than the rest of the cluster and therefore not a member [@2005ApJ...625L.111S]. One could also argue that star E1 should be excluded from the sample. With a high velocity in the opposite direction of the other stars it is equally curious as star E6 in both velocity space and the projected cluster image, where it is somewhat off from the main cluster position. Without star E1 the velocity dispersion of the cluster becomes $\langle v \rangle \simeq 47$–$50$km/s, which results in a estimated mass mass of about 5100–5800. These estimates for the total cluster mass are upper limits for the estimated mass of the dark point mass in the cluster center. If the cluster potential is dominated by a point mass object with a total mass exceeding the stellar mass by a seizable fraction, the stars are in orbit around this mass point. In that case some of the stars may be near the pericenter of their orbit. Since the velocity of a star at pericenter will be a factor of $\sim \sqrt{2}$ larger compared to the velocity in a circular orbit, the estimated black hole mass may therefore also be smaller by up to a factor of 2. We stress that the IMBH mass will be smaller than the total mass derived above since the cluster is made up out of the visible stars, unseen lower mass stars, possible stellar remnants and the potential IMBH. With 300 (seen) but possibly up to $\sim 1000$ of luminous material the mass for the IMBH is then reduced to 2000 – 5000. This is much more than the observed mass of the association, providing a lower limit on its dark mass component. Simulating IRS13E ----------------- With a present density of $\sim4 \times 10^8$[pc]{}$^{-3}$, a collision runaway in IRS13E is inevitable, regardless of the nature of the dark material in the cluster . Therefore, even if the cluster currently does not contain an IMBH, a collision runaway cannot be prevented if the stars are bound. We have tested this using N-body simulations of small clusters of 256 and 1024 stars, with masses drawn from a Salpeter mass function between 1 and 100. These clusters, with $W_0=6$–9 King model initial density profiles, exactly filled their Jacobi surfaces, and moved in circular orbits at 0.18pc from the Galactic center. We continued the calculations until the clusters dissolved. These simulations lost mass linearly in time, with a half-mass lifetime of a few 10,000 years, irrespective of the initial density profile. This is consistent with the results of independent symplectic N-body simulations [@2005astro.ph..2143L]. In each of these simulations a minor runaway merger occurred among roughly a dozen stars, creating runaways of $\aplt 250$. In another set of larger simulations with 1024–16386 stars, the runaway mergers were more extreme, with collision rates exceeding one per century! We draw two conclusions from these simulations. If the unseen material in IRS13E consists of normal stars, then (i) the cluster cannot survive for more than a few $\times10^4$ years, and (ii) runaway merging is overwhelmingly likely. If IRS13E is bound, a cluster of normal stars cannot be hidden within it, and the dark material must ultimately take the form of an IMBH of about 2000–5000 (see also ). Thus we argue that the properties of the dark-matter problem in IRS13E could be solved by the presence of a single IMBH of mass $\sim1000$–5000, consistent earlier discussions . The seven observed stars may in that case be the remnant of a larger star cluster which has undergone runaway merging, forming the IMBH during core collapse while sinking toward the Galactic center [@2001ApJ...562L..19E; @2003ApJ...593..352P; @2005ApJ...628..236G]. According to this scenario the stars we see are the survivors which have avoided collision and remained in tight orbits around the IMBH. Extensive position determinations with the National Radio Astronomy Observatory’s Very Long Baseline Array (VLBA) of Sgr A\* over an $\sim 8$ year baseline has revealed that the SMBH in the Galactic center (assuming 4 million Msun and a distance of 8.0kpc) is about $7.6\pm0.7$kms$^{-1}$ [@2004ApJ...616..872R]. An IMBH of 2000–5000 orbiting at a distance of $\sim0.18$pc would create the linear velocity of about 0.15–0.39kms$^{-1}$ for Sgr A\*, since the orbital velocity of IMBH is $\sim 310$km/s and its mass is $\aplt 1/800$ of the central BH, assuming a circular orbit. If observations with the VLBA continue with the same accuracy for the next decade, the IMBH in IRS13E can be detected by measuring the motion of Sgr A\*. X-ray and Radio observations of the Galactic center --------------------------------------------------- X-ray observations may offer a better chance of observing an individual IMBH near the Galactic center than the VLBA radio observations discussed in the previous section. Among the $\sim 2000$ X-ray point sources within 23pc of the Galactic center [@2003ApJ...589..225M], the source CXOGC J174540.0-290031 [@2004astro.ph.12492M], with $L_{2-8{\rm keV}} \simeq 8.5\times 10^{34}$ erg/s at a projected galactocentric distance of 0.11pc, is of particular interest. The peak radio intensity of this source is 0.1Jansky at 1GHz [@2005astro.ph..7221B], which corresponds to $L_r \sim 8\times 10^{30}$ erg/s at the distance of the Galactic center. Using the recently proposed empirical relation between X-ray luminosity, radio flux, and the mass of the accreting black hole [@2003MNRAS.345.1057M], $$\log L_r = 7.3 + 0.6 \log L_X + 0.8 \log M_{\rm bh}, \label{Eq:Merloni}$$ we derive a black hole mass of about 2000. Interestingly, this source has an 7.8 hour periodicity [@2004astro.ph.12492M], which, if it reflects the orbital period, would indicate a semi-major axis of $\sim 25$. The companion to the IMBH would then have a Roche radius of $\sim 1$, consistent with a 1 main-sequence star. Mass transfer in such a binary would be driven mainly by the emission of gravitational waves at a rate of $\sim 0.01$/Myr [@2004MNRAS.355..413P], which is sufficient to power an X-ray transient with the observed X-ray luminosity and a duty cycle on the order of a few percent [@2004MNRAS.355..413P]. It is a pleasure to thank drs. Clovis Hopman, Tom Maccarone and Mike Muno for interesting discussions, and Prof. Ninomiya for the kind hospitality at the Yukawa Institute at Kyoto University, through the Grants-in-Aid for Scientific Research on Priority Areas, number 763, “Dynamics of Strings and Fields,” from the Ministry of Education, Culture, Sports, Science and Technology, Japan. This work was made possible by financial support from the NASA Astrophysics Theory Program under grant NNG04GL50G, the Netherlands Organization for Scientific Research (NWO) under grant 630.000.001, The Royal Netherlands Academy of Arts and Sciences (KNAW), and the Netherlands Advanced School for Astronomy (NOVA). Part of the calculations in this paper were performed on the GRAPE-6 system in Tokyo and the MoDeStA platform in Amsterdam. , S. A. 2003, , Cambridge University press, 2003 , C., [Lamers]{}, H. J. G. L. M., [Molenberghs]{}, G. 2004, , 418, 639 , T., [Livio]{}, M. 2004, , 606, L21 , H., [Makino]{}, J. 2003, , 340, 227 , H., [Makino]{}, J., [Ebisuzaki]{}, T. 2004, , 613, 1143 , H., [Van Bever]{}, J., [Vanbeveren]{}, D. 2004, ArXiv Astrophysics e-prints , J., [Tremaine]{}, S. 1987, Galactic dynamics, Princeton, NJ, Princeton University Press, 1987, 747 p. , J., [Ivanov]{}, V. D., [Minniti]{}, D., [Geisler]{}, D., [Stephens]{}, A. W. 2005, , 435, 95 , G. C., [Roberts]{}, D. A., [Yusef-Zadeh]{}, F., [Backer]{}, D. C., [Cotton]{}, W. D., [Goss]{}, W. M., [Lang]{}, C. C., [Lithwick]{}, Y. 2005, ArXiv Astrophysics e-prints , W. R., [Geller]{}, M. J., [Kenyon]{}, S. J., [Kurtz]{}, M. J. 2005, , 622, L33 , T., [Makino]{}, J., [Tsuru]{}, T. G., [Funato]{}, Y., [Portegies Zwart]{}, S., [Hut]{}, P., [McMillan]{}, S., [Matsushita]{}, S., [Matsumoto]{}, H., [Kawabe]{}, R. 2001, , 562, L19 , A., [Genzel]{}, R. 1997, , 284, 576 , D. F. 2004, ArXiv Astrophysics e-prints , D. F. 2005, , 434, 192 , D. F., [Kim]{}, S. S. 2002, in ASP Conf. Ser. 263: Stellar Collisions, Mergers and their Consequences, p. 287 , D. F., [Kim]{}, S. S., [Morris]{}, M., [Serabyn]{}, E., [Rich]{}, R. M., [McLean]{}, I. S. 1999a, , 525, 750 , D. F., [McLean]{}, I. S., [Morris]{}, M. 1999b, , 514, 202 , D. F., [Najarro]{}, F., [Gilmore]{}, D., [Morris]{}, M., [Kim]{}, S. S., [Serabyn]{}, E., [McLean]{}, I. S., [Gilbert]{}, A. M., [Graham]{}, J. R., [Larkin]{}, J. E., [Levenson]{}, N. A., [Teplitz]{}, H. I. 2002, , 581, 258 , M., [Atakan G[" u]{}rkan]{}, M., [Rasio]{}, F. A. 2005a, ArXiv Astrophysics e-prints , M., [Benz]{}, W. 2005, , 358, 1133 , M., [Rasio]{}, F. A., [Baumgardt]{}, H. 2005b, ArXiv Astrophysics e-prints , K., [Miller]{}, M. C., [Hamilton]{}, D. P. 2004, , 616, 221 , M. A., [Freitag]{}, M., [Rasio]{}, F. A. 2004, , 604, 632 , R., [Sch[" o]{}del]{}, R., [Ott]{}, T., [Eisenhauer]{}, F., [Hofmann]{}, R., [Lehnert]{}, M., [Eckart]{}, A., [Alexander]{}, T., [Sternberg]{}, A., [Lenzen]{}, R., [Cl[' e]{}net]{}, Y., [Lacombe]{}, F., [Rouan]{}, D., [Renzini]{}, A., [Tacconi-Garman]{}, L. E. 2003, , 594, 812 , O. 2001, , 546, L39 , A. M., [Duch[\^ e]{}ne]{}, G., [Matthews]{}, K., [Hornstein]{}, S. D., [Tanner]{}, A., [Larkin]{}, J., [Morris]{}, M., [Becklin]{}, E. E., [Salim]{}, S., [Kremenek]{}, T., [Thompson]{}, D., [Soifer]{}, B. T., [Neugebauer]{}, G., [McLean]{}, I. 2003, , 586, L127 , A. M., [Klein]{}, B. L., [Morris]{}, M., [Becklin]{}, E. E. 1998, , 509, 678 , A. M., [Morris]{}, M., [Becklin]{}, E. E., [Tanner]{}, A., [Kremenek]{}, T. 2000, , 407, 349 , A., [Zwart]{}, S. P., [Sipior]{}, M. S. 2005, , 363, 223 , M. A., [Rasio]{}, F. A. 2005, , 628, 236 , B. M. S., [Milosavljevi[' c]{}]{}, M. 2003, , 593, L77 , A., [Panter]{}, B., [Jimenez]{}, R., [Dunlop]{}, J. 2004, , 428, 625 , A., [Fryer]{}, C. L., [Woosley]{}, S. E., [Langer]{}, N., [Hartmann]{}, D. H. 2003, , 591, 288 Hills, J. G. 1988, , 331, 687 , G., [White]{}, S. D. M., [Heckman]{}, T. M., [M[' e]{}nard]{}, B., [Brinchmann]{}, J., [Charlot]{}, S., [Tremonti]{}, C., [Brinkmann]{}, J. 2004, , 353, 713 , S. S., [Figer]{}, D. F., [Lee]{}, H. M., [Morris]{}, M. 2000, , 545, 301 , S. S., [Figer]{}, D. F., [Morris]{}, M. 2004, , 607, L123 , R. P. 2002, , 577, 389 , C. J., [Lada]{}, E. A. 2003, , 41, 57 , N., [Hamann]{}, W.-R., [Lennon]{}, M., [Najarro]{}, F., [Pauldrach]{}, A. W. A., [Puls]{}, J. 1994, , 290, 819 , Y., [Wu]{}, A. S. P., [Thommes]{}, E. W. 2005, ArXiv Astrophysics e-prints , J. C., [Thrall]{}, A. P., [Deneva]{}, J. S., [Fleming]{}, S. W., [Grabowski]{}, P. E. 2003, , 345, 762 , J. R., [Ghez]{}, A. M., [Hornstein]{}, S. D., [Morris]{}, M., [Becklin]{}, E. E. 2005, ArXiv Astrophysics e-prints , A. D., [Gilmore]{}, G. F. 2003, , 338, 85 , J. P., [Paumard]{}, T., [Stolovy]{}, S. R., [Rigaut]{}, F. 2004, , 423, 155 , J., [Fukushige]{}, T., [Koga]{}, M., [Namura]{}, K. 2003, , 55, 1163 , F., [Schaerer]{}, D., [Hillier]{}, D. J., [Heydari-Malayeri]{}, M. 2004, , 420, 1087 , S., [Portegies Zwart]{}, S. 2004, ArXiv Astrophysics e-prints astro-ph/0412622 , S. L. W., [Portegies Zwart]{}, S. F. 2003, , 596, 314 , K., [Haiman]{}, Z., [Narayanan]{}, V. K. 2001, , 558, 535 , A., [Heinz]{}, S., [di Matteo]{}, T. 2003, , 345, 1057 , D., [Piatek]{}, S., [Zwart]{}, S. P., [Hemsendorf]{}, M. 2004, , 608, L25 , D., [Poon]{}, M. Y. 2004, , 606, 788 , D., [Wang]{}, J. 2005, , 621, L101 , G., [Maeder]{}, A. 2003, , 404, 975 , G., [Maeder]{}, A. 2005, , 429, 581 , M. 1993, , 408, 496 , M. P., [Baganoff]{}, F. K., [Bautz]{}, M. W., [Brandt]{}, W. N., [Broos]{}, P. S., [Feigelson]{}, E. D., [Garmire]{}, G. P., [Morris]{}, M. R., [Ricker]{}, G. R., [Townsley]{}, L. K. 2003, , 589, 225 , M. P., [Pfahl]{}, E., [Baganoff]{}, F. K., [Brandt]{}, W. N., [Ghez]{}, A., [Lu]{}, J., [Morris]{}, M. R. 2004, ArXiv Astrophysics e-prints , F., [Krabbe]{}, A., [Genzel]{}, R., [Lutz]{}, D., [Kudritzki]{}, R. P., [Hillier]{}, D. J. 1997, , 325, 700 , S., [Cuadra]{}, J. 2004, ArXiv Astrophysics e-prints , S., [Sunyaev]{}, R. 2005, ArXiv Astrophysics e-prints , T., [Maillard]{}, J. P., [Morris]{}, M., [Rigaut]{}, F. 2001, , 366, 466 , S. F., [Baumgardt]{}, H., [Hut]{}, P., [Makino]{}, J., [McMillan]{}, S. L. W. 2004a, , 428, 724 , S. F., [Dewi]{}, J., [Maccarone]{}, T. 2004b, , 355, 413 Portegies Zwart, S. F., McMillan, S. L. W., & Gerhard, O. 2003, , 593, 352 , S. F., [Makino]{}, J., [McMillan]{}, S. L. W., [Hut]{}, P. 1999, , 348, 117 , S. F., [Makino]{}, J., [McMillan]{}, S. L. W., [Hut]{}, P. 2001a, , 546, L101 , S. F., [McMillan]{}, S. L. W. 2002, , 576, 899 , S. F., [McMillan]{}, S. L. W., [Hut]{}, P., [Makino]{}, J. 2001b, , 321, 199 , G. D., [Shapiro]{}, S. L. 1990, , 356, 483 , M. J., [Brunthaler]{}, A. 2004, , 616, 872 , A., [Bergman]{}, P., [Black]{}, J. H., [Booth]{}, R., [Buat]{}, V., [Curry]{}, C. L., [Encrenaz]{}, P., [Falgarone]{}, E., [Feldman]{}, P., [Fich]{}, M., [Floren]{}, H. G., [Frisk]{}, U., [Gerin]{}, M., [Gregersen]{}, E. M., [Harju]{}, J., [Hasegawa]{}, T., [Hjalmarson]{}, [Å]{}., [Johansson]{}, L. E. B., [Kwok]{}, S., [Larsson]{}, B., [Lecacheux]{}, A., [Liljestr[" o]{}m]{}, T., [Lindqvist]{}, M., [Liseau]{}, R., [Mattila]{}, K., [Mitchell]{}, G. F., [Nordh]{}, L., [Olberg]{}, M., [Olofsson]{}, A. O. H., [Olofsson]{}, G., [Pagani]{}, L., [Plume]{}, R., [Ristorcelli]{}, I., [Sch[' e]{}ele]{}, F. v., [Serra]{}, G., [Tothill]{}, N. F. H., [Volk]{}, K., [Wilson]{}, C. D., [Winnberg]{}, A. 2003, , 402, L63 , R., [Eckart]{}, A., [Iserlohe]{}, C., [Genzel]{}, R., [Ott]{}, T. 2005, , 625, L111 , P. F., [Fellhauer]{}, M., [Portegies Zwart]{}, S. F. 2003, , 344, 22 , L. J., [Hart]{}, M. H. 1971, , 166, 483 , A., [Brandner]{}, W., [Grebel]{}, E. K., [Lenzen]{}, R., [Lagrange]{}, A.-M. 2005, , 628, L113 , P., [Rieke]{}, G. H. 1993, , 414, 573 , R., [Kervella]{}, P., [Sch[" o]{}ller]{}, M., [Herbst]{}, T., [Brandner]{}, W., [de Koter]{}, A., [Waters]{}, L. B. F. M., [Hillier]{}, D. J., [Paresce]{}, F., [Lenzen]{}, R., [Lagrange]{}, A.-M. 2003, , 410, L37 , J. S., [de Koter]{}, A., [Lamers]{}, H. J. G. L. M. 2000, , 362, 295 , J. S., [de Koter]{}, A., [Lamers]{}, H. J. G. L. M. 2001, , 369, 574 , H. 1924, , 84, 665 , Q., [Tremaine]{}, S. 2003, , 599, 1129 , Q., [Fall]{}, S. M. 1999, , 527, L81 [^1]: Theoretical considerations suggest that the time scale for cluster dissolution has the form ${\mbox{${t_{\rm dis}}$}}= k {{\mbox{$t_{\rm hc}$}}}^{1/4} {\mbox{${t_{\rm rh}}$}}^{3/4}$, where is the cluster crossing time [@2003MNRAS.340..227B]. The constant $k$ may be obtained from direct N-body simulations of star clusters near the Galactic center [@2001ApJ...546L.101P], resulting in $k\simeq 7.5$, with and expressed in Myr.
{ "pile_set_name": "ArXiv" }
--- abstract: | The open problem of determining the exact value of the $n$-th linear polarization constant $c_n$ of ${\ensuremath{\mathbb{R}}}^n$ has received considerable attention over the past few years. This paper makes a contribution to the subject by providing a new lower bound on the value of $\sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_1,{\bf{y}} \> \cdots \< {\bf{x}}_n,{\bf{y}} \> |$, where ${\bf{x}}_1, \dots ,{\bf{x}}_n$ are unit vectors in ${\ensuremath{\mathbb{R}}}^n$. The new estimate is given in terms of the eigenvalues of the Gram matrix $[\< {\bf{x}}_i,{\bf{x}}_j\> ]$ and improves upon earlier estimates of this kind. However, the intriguing conjecture $c_n=n^{n/2}$ remains open. title: A geometric estimate on the norm of product of functionals --- Alfréd Rényi Institute of Mathematics, Hungarian Academy of Sciences, POB 127, H-1364, Budapest, Hungary. Tel: (+361) 483-8302, Fax: (+361) 483-8333, e-mail: [email protected] [**2000 Mathematics Subject Classification.**]{} Primary 46G25; Secondary 52A40, 46B07. [**Keywords and phrases.**]{} [*Polynomials over normed spaces, product of functionals, linear polarization constants, Gram matrices.*]{} Introduction ============ The present work contributes to study of the $n$-th linear polarization constant $c_n({\ensuremath{\mathbb{R}}}^n)$ of the $n$-dimensional real Euclidean space. We begin with introducing some (more general) standard terminology and giving a short account of some related results. Let $X$ denote a Banach space over the real or complex field ${\ensuremath{\mathbb{K}}}$. A function $P: \ X\to {\ensuremath{\mathbb{K}}}$ is a continuous [*n-homogeneous polynomial*]{} if there exists a symmetric, continuous $n$-linear form $L: \ X^n\to{\ensuremath{\mathbb{K}}}$ such that $P({\bf{x}})=L({\bf{x}},\dots ,{\bf{x}})$ for all ${\bf{x}}\in X$. We define $$\|P\|:=\sup \{ |P({\bf{x}})|: \ {\bf{x}}\in B\}$$ where $B$ denotes the unit ball of $X$. Considerable attention has been devoted to polynomials of the form $P({\bf{x}})=f_1({\bf{x}})f_2({\bf{x}})\dots f_n({\bf{x}})$, where $f_1, f_2,\dots ,f_n$ are bounded linear functionals on $X$. For any[*complex*]{} Banach space $X$ Benítez, Sarantopoulos and Tonge [@ben] have obtained $$\|f_1\| \ \|f_2\|\cdots \|f_n\|\le n^n\|f_1 f_2 \dots f_n\|,$$ and they also showed that, in general, the constant $n^n$ is best possible. For [*real*]{} Banach spaces, Ball’s solution [@ball2] of the famous plank problem of Tarski gives the same result. For specific spaces, however, the general constant $n^n$ can be lowered. This fact motivated the following (Benítez, Sarantopoulos, Tonge [@ben]) The n-th linear polarization constant of a Banach space $X$ is defined by $$\begin{aligned} \label{polconstdef} c_n(X)&:=& \inf \{M : \|f_{1}\| \cdots \|f_{n}\| \leq M \|f_{1}\cdots f_{n}\|\; (\forall f_{1}, \ldots ,f_{n} \in X^{\ast}) \} \\ & = & 1/ \inf_{f_{1}, \ldots ,f_{n} \in S_{X^{\ast}}} \sup_{\|{\bf{x}}\|=1}|f_{1}({\bf{x}}) \cdots f_{n}({\bf{x}})|.\end{aligned}$$ The linear polarization constant of $X$ is defined by $$\label{limitconstant} c(X):=\lim_{n \rightarrow \infty} c_{n} (X)^{ \frac{1}{n}}\;\;.$$ Let us recall that the above definition of $c(X)$ is justified since Révész and Sarantopoulos [@sar] showed that the limit does exist. Moreover, they also showed (both in the real and complex cases) that $c(X)=\infty$ if and only if $\dim X=\infty$. Note that it is easy to see that for any Banach space $X$ we have $$\label{finite} c_{n}(X)=\sup \left\{c_{n} (Y): Y {\mbox {is a subspace of}}\, X, \, \dim Y=n \right\}\,.$$ In particular, for a real or complex Hilbert space $H$ of dimension at least $n$, we always have $c_n(H)=c_n({\ensuremath{\mathbb{K}}}^n)$. Benítez, Sarantopoulos and Tonge [@ben] proved that for isomorphic Banach spaces $X$ and $Y$ we have $c_n(X)\le d^n(X,Y)c_n(Y)$, where $d(X,Y)$ denotes the Banach-Mazur distance of $X$ and $Y$. Note, that for any $n$-dimensional space $X$ a result of John [@john] states that $d(X,{\ensuremath{\mathbb{K}}}^n)\le \sqrt{n}$ (where ${\ensuremath{\mathbb{K}}}^n$ denotes the $n$-dimensional Hilbert space). The combination of these results mean that the determination of $c_n({\ensuremath{\mathbb{K}}}^n)$ gives information on the linear polarization constants of other spaces, too. In this paper we are going to focus our attention to Hilbert spaces. Pappas and Révész [@papp] showed that $c({\ensuremath{\mathbb{K}}}^n)= e^{-L(n,{\ensuremath{\mathbb{K}}})}$, where $$L(n,{\ensuremath{\mathbb{K}}}):= \int_{S}\textrm{log}|\< {\bf{x}},{\bf{e}}\> |d\sigma ({\bf{x}});$$ here $S$ and $\sigma$ denote the unit sphere and the normalized surface measure, respectively, and ${\bf{e}}\in S$ is an arbitrary unit vector. This result gives information on the asymptotic behaviour of $c_m({\ensuremath{\mathbb{K}}}^n)$ as $m\to \infty$. However, the exact values of $c_m({\ensuremath{\mathbb{K}}}^n)$ seem, in general, hopeless to determine. A remarkable result of Arias-de-Reyna [@rey] states that $c_n({\ensuremath{\mathbb{C}}}^n)=n^{n/2}$. Ball’s recent solution [@ball] of the complex plank problem also implies the same result. Compared to the complex case, the value of $c_n({\ensuremath{\mathbb{R}}}^n)$ seems harder to find. The determination of $c_n({\ensuremath{\mathbb{R}}}^n)$, by the definition and the Riesz representation theorem, boils down to determining $$I:= \inf_{{\bf{x}}_{1}, \dots ,{\bf{x}}_{n} \in S} \sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_{1},{\bf{y}}\> \cdots \< {\bf{x}}_{n},{\bf{y}}\> |$$ The estimate $I\le n^{-\frac{n}{2}}$ follows by considering an orthonormal system. The complex result of Arias-de-Reyna can be used to derive the following estimates (see [@sar], where the argument is based on an interesting complexification result of [@munoz]): $$\label{legjobb} n^{\frac{n}{2}}\le c_n({\ensuremath{\mathbb{R}}}^n)\le 2^{\frac{n}{2}-1}n^{\frac{n}{2}}.$$ A natural, intriguing conjecture, see [@ben], [@sar] is the following. [**Conjecture.**]{} $c_n({\ensuremath{\mathbb{R}}}^n)=n^{n/2}$. Marcus (communicated in [@marc], and elaborated later in [@sar]) gives the following estimate: If ${\bf{x}}_1, {\bf{x}}_2, \dots , {\bf{x}}_n$ are unit vectors in ${\ensuremath{\mathbb{R}}}^n$ then there exists a unit vector ${\bf{y}}$ such that $$\label{marcus} |\< {\bf{x}}_1,{\bf{y}}\> \cdots\< {\bf{x}}_n,{\bf{y}}\> |\ge (\l_1/n)^{n/2},$$ where $\l_1$ denotes the smallest eigenvalue of the Gram matrix $XX{^\ast}=[\< {\bf{x}}_i,{\bf{x}}_j\> ]$. Marcus also expressed the opinion that lower bounds on $\sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_1,{\bf{y}}\> \cdots \< {\bf{x}}_n,{\bf{y}}\> |$ should involve the eigenvalues $\l_1,\dots,\l_n$ of the Gram matrix $XX{^\ast}=[\< {\bf{x}}_i,{\bf{x}}_j\> ]$, i.e. we should look for estimates of the form $\sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_1,{\bf{y}}\> \cdots \< {\bf{x}}_n,{\bf{y}}\> |\ge f(\l_1, \dots , \l_n) n^{-n/2}$. Note that $\sum_j \l_j={\mathrm{Tr}\:}XX{^\ast}=n$. Therefore the above Conjecture can be formulated as $$\sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_1,{\bf{y}}\> \cdots \< {\bf{x}}_n,{\bf{y}}\> | \ge 1\cdot n^{-n/2}= \left(\frac{\l_1 +\dots +\l_n}{n}\right)^{n/2}n^{-n/2}.$$ In [@matpolar] the author proved that Marcus’ estimate can be improved to $$\sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_1,{\bf{y}}\> \cdots \< {\bf{x}}_n,{\bf{y}}\> | \ge \left( \frac{n}{\l_1^{-1}+ \cdots +\l_n^{-1}} \right)^{n/2}n^{-n/2}.$$ In the next section we will improve this result by replacing the harmonic mean of the numbers $\l_1, \dots , \l_n$ by the geometric mean. Also, in the course of the proof we use two ’geometrical’ lemmas which may be of independent interest. The original Conjecture (invoving the arithmethic mean of the numbers $\l_1, \dots , \l_n$), however, still remains open. A geometric lower bound ======================= For the sake of simplicity we introduce the following notations: Let $b_n$ denote the volume of the $n$-dimensional closed unit ball $B^{n}$ (we will not need the explicit value of $b_n$). Also, let $H_\alpha :=\{{\bf{z}}=(z_1, \dots , z_n)\in {\ensuremath{\mathbb{R}}}^n: \ |\prod_{j=1}^nz_j|\ge \alpha \cdot n^{-n/2}\}$. In order to prove our main result, Theorem \[theo1\], we will need the following two geometrical lemmas: \[lem1\] Let $E$ be an $n$-dimensional ellipsoid symmetric with respect to the origin (i.e. the image of the $n$-dimensional unit ball under a linear transformation of full rank) of volume $Vb_n$. Assume that the $n-1$-dimensional ’horizontal slice’ $E_0:=\{{\bf{z}}=(z_1, z_2, \dots z_n)\in E : \ z_n=0\}$ has $n-1$-dimensional volume $Sb_{n-1}$. Then the horizontal slice at height $h$, $E_h:=\{{\bf{z}}=(z_1, z_2, \dots z_n)\in E : \ z_n=h\}$ has $n-1$-dimensional volume $$f(V,S,h)=\left \{ \begin{array}{ll} (1-(\frac{S}{V}h)^2)^{\frac{n-1}{2}}b_{n-1} & \mathrm{if} \ |h|\le V/S\\ 0 & \mathrm{if} \ |h|> V/S \end{array}\right.$$ The essence of the lemma is that the function $f$ depends only on $V,S$ and $h$ and not on the actual ’shape’ of the ellipsoid. The statement of the lemma is clear if $E$ is a ’circular ellipsoid’ whose axes are the same as the coordinate axes, i.e. $E$ is the image of the unit ball $B^{n}$ under the diagonal transformation $$T:=\left ( \begin{array} {cccccc} S^{1/(n-1)}& \ & \ & \ & \ & 0 \\ \ & . & \ & \ & \ & \ \\ \ & \ & . & \ & \ & \ \\ \ & \ & \ & . & \ & \ \\ \ & \ & \ & \ & S^{1/(n-1)} & \ \\ 0 & \ & \ & \ & \ & \frac{V}{S} \end{array} \right)$$ In the general case, let $E=A[B^{n}]$ be the image of the unit ball $B^{n}$ under some transformation $A$, and assume that it posesses the prescribed parameters $V, \ S$, and let the height $h$ also be given. The natural idea of the proof is that we transform the ellipsoid $E$ to a circular ellipsoid whose axes are the coordinate axes and whose parameters are the same. Let ${\bf{r}}:=(r_1, \dots ,r_n)\in E$ denote the point of $E$ whose last coordinate $r_n$ is maximal among the points of $E$, and let ${\bf{q}}\in B^{n}$ be its inverse image, i.e. ${\bf{q}}=A^{-1}{\bf{r}}$. Let $L_0:=A^{-1}[E_0]$. Note that the $n-1$-dimensional $L_0$ is orthogonal to the vector ${\bf{q}}$, therefore there exists a unitary transformation $U$ which takes the horizontal slice $B_0^{n}$ of $B^{n}$ to $L_0$ and the vertical unit vector ${\bf{e}}_n$ to ${\bf{q}}$ (note that if $n-1>1$ then $U$ is not uniquely determined). Then we have $AU[B^{n}]=E$, $AU[B_0^{n}]=E_0$ and $AU({\bf{e}}_n)={\bf{r}}$. The transformation $AU$ maps the horizontal $n-1$-dimensional hyperplane $P_0:=\{{\bf{x}}=(x_1, \dots ,x_n)\in {\ensuremath{\mathbb{R}}}^n : \ x_n=0\}$ onto itself. Denote the restriction of $AU$ to $P_0$ by $C_0$. Now, take the $n-1$-dimensional transformation $\tilde{C}_1:=S^{1/(n-1)}C_0^{-1}$ of the horizontal hyperplane $P_0$. This preserves $n-1$-dimensional volume (i.e. it has determinant $\pm 1$) and takes the horizontal slice $E_0$ to $S^{1/(n-1)}B_0^{n}$. Consider now the $n$-dimensional transformation $$C_1:=\left ( \begin{array} {cc} \tilde{C}_1&{\bf{0}}\\ {\bf{0}}^T&1 \end{array} \right) .$$ It is clear that applying this transformation to $E$ the image ellipsoid $C_1[E]$ will posses the same parameters as $E$, i.e. the same volume $V$, the same $n-1$-dimensional volume $S$ of its horizontal slice $(C_1[E])_0=S^{1/n-1}B_0^{n}$, and the image $C_1[E_h]$ will still be at height $h$ and have the same $n-1$-dimensional volume as $E_h$. Let ${\bf{s}}:=(s_1, \dots ,s_n)=C_1{\bf{r}}$. Next we consider the transformation $$C_2:=\left ( \begin{array} {cccccc} 1 & \ & \ & \ & 0 & -s_1/s_n \\ \ & . & \ & \ & \ & . \\ \ & \ & . & \ & \ & . \\ \ & \ & \ & . & \ & . \\ \ & \ & \ & \ & 1 & -s_{n-1}/s_n \\ 0 & \ & \ & \ & 0 & 1 \end{array} \right)$$ Once again it is clear that the image $C_2C_1[E]$ has the same parameters $V, \ S$ as $E$, and $(C_2C_1[E])_h=C_2C_1[E_h]$ with equal $n-1$-dimensional volume. To finish the proof it is enough to observe that $C_2C_1[E]=C_2C_1AU[B^{n}]=T[B^{n}]$ with the diagonal transformation $T$ above. The next lemma establishes the connection between ellipsoids and products of functionals. \[lem2\] Assume $E$ is an $n$-dimensional ellipsoid of volume $Vb_n$ (not necessarily centered at the origin). Then $E\cap H_V\ne \emptyset$. The proof proceeds by induction with respect to $n$. For $n=1$ the statement is clear. For an arbitrary $n$ let ${\bf{c}}:=(c_1, \dots c_n)$ denote the centre of $E$, and assume, without loss of generality, that $c_n\ge 0$. Let $Sb_{n-1}$ be the $n-1$ dimensional volume of the horizontal slice $E\cap P_{c_n}$, where $P_{c_n}:=\{{\bf{x}}=(x_1, \dots ,x_n)\in {\ensuremath{\mathbb{R}}}^n : \ x_n=c_n\}$. Now, let $h:=\frac{V}{S\sqrt{n}}$ and consider the horizontal hyperplane $P:=P_{c_n+h}$. $P$ is an $n-1$-dimensional space, and $P\cap H_V=\{{\bf{z}}=(z_1, \dots z_{n-1}): \ |\prod_{j=1}^{n-1}z_j|\ge \frac{V}{c_n+h}n^{-n/2}\}\supset \{{\bf{z}}=(z_1, \dots z_{n-1}): \ |\prod_{j=1}^{n-1}z_j|\ge \frac{V}{h}n^{-n/2}\}$. Furthermore, the $n-1$-dimensional volume of $P\cap E$ is $(1-\frac{1}{n})^\frac{n-1}{2}Sb_{n-1}$ in view of Lemma \[lem1\] and the choice of $h$. Finally, observe that $\frac{V}{h}n^{-n/2}= (n-1)^\frac{-n+1}{2}\left( \frac{V}{h}(1-\frac{1}{n})^\frac{n-1}{2}\frac{1}{\sqrt{n}}\right )$ and $(1-\frac{1}{n})^\frac{n-1}{2}S= \frac{V}{h}(1-\frac{1}{n})^\frac{n-1}{2}\frac{1}{\sqrt{n}}$, therefore the inductive hypothesis applies. We are now in position to prove our new estimate on the norm of product of functionals. \[theo1\] Let unit vectors ${\bf{x}}_1,\dots , {\bf{x}}_n$ be given in ${\ensuremath{\mathbb{R}}}^n$, and let $\l_1, \dots ,\l_n$ denote the eigenvalues of the Gram matrix $XX{^\ast}=[\< {\bf{x}}_i,{\bf{x}}_j\> ]$ (the matrix $X$ is formed by the given vectors as rows). Then $$\label{geo} \sup_{\|{\bf{y}}\|=1}|\< {\bf{x}}_1,{\bf{y}}\> \cdots \< {\bf{x}}_n,{\bf{y}}\> | \ge \left (\prod_{j=1}^n\l_j\right )^{1/2}\cdot n^{-n/2}$$ We may assume that the vectors ${\bf{x}}_1, {\bf{x}}_2, \dots {\bf{x}}_n$ are linearly independent, otherwise the right hand side of the inequality is 0, and the estimate is meaningless. (We remark that other considerations , such as the ones in [@matpolar], also show that if we find a way to prove a good estimate in the case of linearly dependent vectors then we may get close to proving the original Conjecture. However, at present, there seems to be no better estimate than for the linearly dependent case.) The image $E$ of the unit ball $B^{n}$ under the transformation $X$ is an $n$-dimensional ellipsoid of volume $V=\left (\prod_{j=1}^n\l_j\right )^{1/2}b_n$, therefore Lemma \[lem2\] gives the required result. Finally, let us make the following remarks. An advantage of the proof applied above is that it is constructive in the sense that following the constructions of Lemma \[lem2\] we can actually find a vector ${\bf{y}}$ which satisfies . It is clear, however, that the estimate does not settle the original Conjecture. [1]{} J. Arias-de-Reyna: Gaussian variables, polynomials and permanents, [*Linear Algebra Appl.*]{} [**285**]{}(1998), 107-114. K.M. Ball: The plank problem for symmetric bodies, [*Invent. Math.*]{} [**104**]{}(1991), 535-543. K. M. Ball: The complex plank problem, [*Bull. London Math. Soc.*]{} [**33**]{}(2001), 433-442. T. Bang: A solution of the plank problem, [*Proc. Amer. Math. Soc.*]{} [**2**]{}(1951), 990-993. C. Benitez, Y. Sarantopoulos, A.M Tonge: Lower bounds for norms of products of polynomials, [*Math. Proc. Cambridge Philos. Soc.*]{} [**124**]{}(1998), 395-408. F. John: Extremum problems with inequalities as subsidiary conditions, [*Courant Aniversary Volume, Interscience, New York*]{} (1948), 187-204. M. Marcus: Letter to Y. Sarantopoulos, (1996). M. Matolcsi: On the $n$-th linear polarization constant of ${\ensuremath{\mathbb{R}}}^n$, [*Acta Math. Hung.*]{}, to appear. G. Mu$\tilde{n}$oz, Y. Sarantopoulos, A.M. Tonge: Complexifications of real Banach spaces, polynomials and multilinear maps, [*Studia Math.*]{} [**134**]{}(1999), 1-33. A. Pappas, Sz. Révész: Linear polarization constants of Hilbert spaces, [*J. Math. Anal. Appl.*]{} [**300/1**]{}(2004), 129-146. Sz. Révész, Y. Sarantopoulos: Plank problems, polarization and Chebyshev constants, [*J. Korean Math. Soc.*]{} [**41**]{}(2004), No. 1, 157-174.
{ "pile_set_name": "ArXiv" }
--- abstract: | We consider the kernel partial least squares algorithm for non-parametric regression with stationary dependent data. Probabilistic convergence rates of the kernel partial least squares estimator to the true regression function are established under a source and an effective dimensionality conditions. It is shown both theoretically and in simulations that long range dependence results in slower convergence rates. A protein dynamics example shows high predictive power of kernel partial least squares. [\ *Key words and phrases.*]{} Effective dimensionality, Long range dependence, Nonparametric regression, Source condition, Protein dynamics. author: - | Marco Singer$^a$, Tatyana Krivobokova$^a$, Axel Munk$^{a,b}$\ $^a$Institute for Mathematical Stochastics, Göttingen, Germany\ $^b$Max Planck Institute for Biophysical Chemistry, Göttingen, Germany bibliography: - 'bibliography.bib' title: Kernel partial least squares for stationary data --- =1 Introduction {#sec:Introduction} ============ Partial least squares (PLS) is a regularized regression technique developed by @Wol84 to deal with collinearities in the regressor matrix. It is an iterative algorithm where the covariance between response and regressor is maximized at each step, see @Hel88 for a detailed description. Regularization in the PLS algorithm is obtained by stopping the iteration process early. Several studies showed that partial least squares algorithm is competitive with other regression methods such as ridge regression and principal component regression and it needs generally fewer iterations than the latter to achieve comparable estimation and prediction, see, e.g., @FrankFriedman and @Krae07b. For an overview of further properties of PLS we refer to @Ros06. Reproducing kernel Hilbert spaces (RKHS) have a long history in probability and statistics [see e.g. @Berlinet]. Here we focus on the supervised kernel based learning approach for the solution of non-parametric regression problems. RKHS methods are both computationally and theoretically attractive, due to the kernel trick [@Sch98] and the representer theorem [@Wah99] as well as its generalization [@Sch01]. Within the reproducing kernel Hilbert space framework one can adapt linear regularized regression techniques like ridge regression and principal component regression to a non-parametric setting, see @Sau98 and @Ros00a, respectively. We refer to @bSch for more details on the kernel based learning approach. Kernel PLS was introduced in @Ros01 who reformulated the algorithm presented in @Lin93. The relationship to kernel conjugate gradient (KCG) methods was highlighted in @Blan10a. It can be seen in @Hanke that conjugate gradient methods are well suited for handling ill-posed problems, as they arise in kernel learning, see, e.g., @Vit06. @Ros03 investigated the performance of kernel partial least squares (KPLS) for non-linear discriminant analysis. @Blan10a proved the consistency of KPLS when the algorithm is stopped early without giving convergence rates. @Cap07 showed that kernel ridge regression (KRR) attains optimal probabilistic rates of convergence for independent and identically distributed data, using a source and a polynomial effective dimensionality condition. A generalization of these results to a wider class of effective dimensionality conditions and extension to kernel principal component regression can be found in @Dicker17. For a variant of KCG @Blan10b obtained probabilistic convergence rates for independent identically distributed data. The pointed explicitly out that their approach and results are not directly applicable to KPLS. We study of the convergence of the kernel partial least squares estimator to the true regression function when the algorithm is stopped early. Similar to @Blan10b we derive explicit probabilistic convergence rates. In contrast to previously cited works on kernel regression our input data are not independent and identically distributed but rather stationary time series. We derive probabilistic convergence results that can be applied for arbitrary temporal dependence structures, given that certain concentration inequalities for these data hold. The derived convergence rates depend not only on the complexity of the target function and of the data mapped into the kernel space, but also on the persistence of the dependence in the data. In the stationary setting we prove that the short range dependence still leads to optimal rates, but if the dependence is more persistent, the rates become slower. Kernel Partial Least Squares {#sec:problem} ============================ Consider the non-parametric regression problem $$\label{eq:model} y_t = f^\ast(X_t) + \varepsilon_t,~~t \in {{\mathbb Z}}.$$ Here $\{X_t\}_{t \in {{\mathbb Z}}}$ is a $d$-dimensional, $d \in {{\mathbb N}}$, stationary time series on a probability space $(\Omega,\calA,{\mathrm{P}})$ and $\{\varepsilon_t\}_{t \in {{\mathbb Z}}}$ is an independent and identically distributed sequence of real valued random variables with expectation zero and variance $\sigma^2 > 0$ that is independent of $\{X_t\}_{t \in {{\mathbb Z}}}$. Let $X$ be a random vector that is independent of $\{X_t\}_{t \in {{\mathbb Z}}}$ and $\{\varepsilon_t\}_{t \in {{\mathbb Z}}}$ with the same distribution as $X_0$. The target function we seek to estimate is $f^\ast \in {\calL^2\left({\mathrm{P}}^{X} \right)}$. For the purpose of supervised learning assume that we have a training sample $\{(X_t,y_t)\}_{t=1}^n$ for some $n \in {{\mathbb N}}$. In the following we introduce some basic notation for the kernel based learning approach. Define with $(\mathcal{H},\langle \cdot,\cdot\rangle_\calH)$ the RKHS of functions on ${{\mathbb R}}^d$ with reproducing kernel $k:{{\mathbb R}}^d \times {{\mathbb R}}^d \rightarrow {{\mathbb R}}$, i.e., it holds $$\label{eq:rep.property} g(x) = \langle g, k(\cdot,x) \rangle_\calH, ~~x \in {{\mathbb R}}^d, g\in \calH.$$ The corresponding inner product and norm in $\calH$ is denoted by $\langle \cdot,\cdot \rangle_\calH$ and $\|\cdot\|_\calH$, respectively. We refer to @Berlinet for examples of Hilbert spaces and their reproducing kernels. In the following we deal with reproducing kernel Hilbert spaces which fulfill the following, rather standard, conditions: 1. \[con:k1\] $\calH$ is separable, 2. \[con:k2\] There exists a $\kappa>0$ such that $|k(x,y)| \leq \kappa$ for all $x,y \in {{\mathbb R}}^d$ and $k$ is measurable. Under \[con:k1\] the Hilbert-Schmidt norm $\|\cdot\|_{{\mathrm{HS}}}$ for operators mapping from $\calH$ to $\calH$ is well defined. If condition \[con:k2\] holds, all functions in $\calH$ are bounded, see @Berlinet, chapter 2. The conditions are satisfied for a variety of popular kernels, e.g., Gaussian or triangular. The main principle of RKHS methods is the mapping of the data $X_t$ into $\calH$ via the feature maps $\phi_t = k(\cdot,X_t)$, $t=1,\dots,n$. This mapping can be done implicitly by using the kernel trick $\langle \phi_t,\phi_s \rangle_\calH = k(X_t,X_s)$ and thus only the $n \times n$ dimensional kernel matrix $K_n = n^{-1}[k(X_t,X_s)]_{t,s=1}^n$ is needed in the computations. Then the task for RKHS methods is to find coefficients $\alpha_1,\dots,\alpha_n$ such that $f_\alpha = \sum_{t=1}^n \alpha_t \phi_t$ is an adequate approximation of $f^\ast$ in $\calH$, measured in the ${\calL^2\left({\mathrm{P}}^{X} \right)}$ norm $\|\cdot\|_2$. There are a variety of different approaches to estimate the coefficients $\alpha_1,\dots,\alpha_n$, including kernel ridge regression, kernel principal component regression and, of course, kernel partial least squares. The latter method was introduced by @Ros01 and is the focus of the current work. It was shown by @Krae07b that the KPLS algorithm solves $$\label{eq:kpls.optim} {\widehat{\alpha}}_i = \arg\min\limits_{v \in \calK_i(K_n,y)} \|y - K_n v\|^2,~~i=1,\dots,n,$$ with $y=(y_1,\dots,y_n)^\T$. Here $\calK_i(K_n,y) = \mathrm{span}\left\{ y,K_n y, K_n^2y,\dots,K_n^{i-1}y \right\}$, $i=1,\dots,n$, is the $i$th order Krylov space with respect to $K_n$ and $y$ and $\| \cdot\|$ denotes the Euclidean norm rescaled by $n^{-1}$. The dimension $i$ of the Krylov space is the regularization parameter for KPLS. We will introduce several operators that will be crucial for our further analysis. Fist define two integral operators: the kernel integral operator $T^\ast:{\calL^2\left({\mathrm{P}}^{X} \right)}\rightarrow \calH, g \mapsto{\mathrm{E}}\{k(\cdot,X) g(X)\}$ and the change of space operator $T:\calH \rightarrow {\calL^2\left({\mathrm{P}}^{X} \right)}, g \mapsto g$, which is well defined if \[con:k2\] holds. It is easy to see that $T, T^\ast$ are adjoint, i.e., for $u \in \calH$ and $v \in {\calL^2\left({\mathrm{P}}^{X} \right)}$ it holds $\langle T^\ast v, u \rangle_\calH = \langle v, T u \rangle_2$ with $\langle\cdot,\cdot \rangle_2$ being the inner product in ${\calL^2\left({\mathrm{P}}^{X} \right)}$. The sample analogues of $T,T^\ast$ are $T_n:\calH \rightarrow {{\mathbb R}}^n, g \mapsto \{g(X_1),\dots,g(X_n)\}^\T$ and $T_n^\ast:{{\mathbb R}}^n \rightarrow \calH, (v_1,\dots,v_n)^\T \mapsto n^{-1} \sum_{t=1}^n v_t k(\cdot,X_t)$, respectively. Both operators are adjoint with respect to the rescaled Euclidean product $\langle u,v\rangle = n^{-1} u^\T v$, $u,v \in {{\mathbb R}}^d$ Finally, we define the sample kernel covariance operator ${S_n}= T^\ast_n T_n:\calH \rightarrow \calH$ and the population kernel covariance operator ${S}= T^\ast T:\calH \rightarrow \calH$. Note that it holds $K_n = T_n T_n^\ast$. Under \[con:k1\] and \[con:k2\] ${S}$ is a self-adjoint compact operator with operator norm $\|{S}\|_{\calL} \leq \kappa$, see @Cap07. With this notation we can restate (\[eq:kpls.optim\]) for the function $f_\alpha$ $$\label{eq:func.representation} f_{\widehat{\alpha}_i} = \arg \min_{g \in \calK_i({S_n},{T_n^\ast y})} \|y- \{g(X_1),\dots,g(X_n)\}^\T\|^2=\arg \min_{g \in \calK_i({S_n},{T_n^\ast y})}\|y-T_ng\|^2.$$ Hence, we are looking for functions that minimize the squared distance to $y$ constrained to a sequence of Krylov spaces. In the literature of ill-posed problems it is well known that without further conditions on the target function $f^\ast$ the convergence rate of the conjugate gradient algorithm can be arbitrarily slow, see @Hanke, chapter 3.2. One common a-priori assumption on the regression function $f^\ast$ is a source condition: 1. \[eq:source\] There exist $r \geq 0$, $R>0$ and $u \in {\calL^2\left({\mathrm{P}}^{X} \right)}$ such that $f^\ast = (T T^\ast)^{r} u$ and $\|u\|_2 \leq R$. If $r \geq 1/2$, then the target function $f^\ast \in {\calL^2\left({\mathrm{P}}^{X} \right)}$ coincides almost surely with a function $f \in \calH$ and we can write $f^\ast = T f$, see @Cuc02. With this the kernel partial least squares estimator $f_{\widehat{\alpha}_i}$ estimates the correct target function, not only its best approximation in $\mathcal{H}$. This case is known as the inner case. The situation with $r<1/2$ is referred to as the outer case. Under additional assumptions, e.g., the availability of additional unlabeled data, it is still possible that an estimator of $f^\ast$ converges to the true target function in ${\calL^2\left({\mathrm{P}}^{X} \right)}$ norm with optimal rates (with respect to the number $n$ of labeled data points). See @Vit06 for a detailed description of this semi-supervised approach for kernel ridge regression in the independent and identically distributed case. We do not treat the case $r<1/2$ in this work. A source conditions is often interpreted as an abstract smootheness condition. This can be seen as follows. Let $\eta_1 \geq \eta_2 \geq \dots$ be the eigenvalues and $\psi_1,\psi_2,\dots$ the corresponding eigenfunctions of the compact operator $S$. Then it is easy to see that the source condition \[eq:source\] is equivalent to $f = \sum_{j=1}^\infty b_j \psi_j \in {\calL^2\left({\mathrm{P}}^{X} \right)}$ with $b_j$ such that $\sum_{j=1}^\infty \eta_j^{-2(r+1/2)} b_j^2 < \infty$. Hence, the higher $r$ is chosen the faster the sequence $\{b_j\}_{j=1}^\infty$ must converge to zero. Therefore, the sets of functions for which source conditions hold are nested, i.e., the larger $r$ is the smaller the corresponding set will be. The set with $r=1/2$ is the largest one and corresponds to a zero smoothness condition, i.e., $\sum_{j=1}^\infty \eta_j^{-2} b_j^2 < \infty$, which is equivalent to $f \in \calH$. For more details we refer to @Dicker17. Consistency of Kernel Partial Least Squares {#sec:kpls.convergence} =========================================== The KCG algorithm as described by @Blan10b is consistent when stopped early and convergence rates can be obtained when a source condition \[eq:source\] holds. Here we will proof the same property for KPLS. Early stopping in this context means that we stop the algorithm at some $a=a(n) \leq n$ and consider the estimator $f_{{\widehat{\alpha}}_a}$ for $f^\ast$. The difference between KCG and KPLS is the norm which is optimized. The kernel conjugate gradient algorithm studied in @Blan10b estimates the coefficients $\alpha \in {{\mathbb R}}^n$ of $f_\alpha$ via ${\widehat{\alpha}}_i^{CG} = \arg\min_{v \in \calK_i(K_n,y)} \langle y- K_n v, K_n(y- K_n v)\rangle$. It is easy to see that this optimization problem can be rewritten for the function $f_\alpha$ as $$\min_{g \in \calK_i({S_n},{T_n^\ast y})} \|T_n^\ast y-S_ng\|_\calH^2= \min_{g \in \calK_i({S_n},{T_n^\ast y})} \|T_n^\ast\left(y-T_ng\right)\|_\calH^2,$$ compared to (\[eq:func.representation\]) for KPLS. Thus, KCG obtains the least squares approximation $g$ in the $\calH$-norm for the normal equation ${T_n^\ast y}= T^\ast _nT_n g$ and KPLS finds a function that minimizes the residual sum of squares. In both methods the solutions are restricted to functions $g \in \calK_i({S_n},{T_n^\ast y})$. An advantage of the kernel conjugate gradient estimator is that concentration inequalities can be established for both ${T_n^\ast y}$ and ${S_n}$ and applied directly as the optimization function contains both quantities. The stopping index for the regularization can be chosen by a discrepancy principle as ${{a^\ast}}= \min\{1\leq i \leq n: \|{S_n}f_{{\widehat{\alpha}}_i^{CG}} - {T_n^\ast y}\| \leq \Lambda_n\}$ with $\Lambda_n$ being a threshold sequence that goes to zero as $n$ increases. On the other hand, the function to be optimized for KPLS contains only $y$ and $T_n g = \{g(X_1),\dots,g(X_n)\}^\T$ for which statistical properties are not readily available. Thus, we need to find a way to apply the concentration inequalities for ${T_n^\ast y}$ and ${S_n}$ to this slightly different problem. This leads to complications in the proof of consistency and a rather different and more technical stopping rule for choosing the optimal regularization parameter $a^\ast$ is used, as can be seen in Theorem \[th:kpls\]. This stopping rule has its origin in @Hanke.\ In the following $\|\cdot\|_{\cal{L}}$ denotes the operator norm and $\|\cdot\|_{HS}$ is the Hilbert-Schmidt norm. \[th:kpls\] Assume that conditions \[con:k1\], \[con:k2\], \[eq:source\] hold with $r \geq 3/2$ and there are constants $C_\delta(\nu),C_\epsilon(\nu)>0$ and a sequence $\{\gamma_n\}_{n \in {{\mathbb N}}} \subset [0,\infty)$, $\gamma_n \rightarrow 0$, such that we have for $\nu \in (0,1]$ $$\begin{aligned} {\mathrm{P}}\left( \|{S_n}-{S}\|_{\calL} \leq C_\delta(\nu) \gamma_n \right) &\geq 1 - \nu/2,\\ {\mathrm{P}}\left( \|{T_n^\ast y}- {S}f\|_\calH \leq C_\epsilon(\nu) \gamma_n \right) &\geq 1- \nu/2.\end{aligned}$$ Define the stopping index $a^\ast$ by $$\label{eq:stopping} a^\ast = \min\left\{1 \leq a \leq n: \sum_{i=0}^a \|{S_n}f_{{\widehat{\alpha}}_i} - {T_n^\ast y}\|^{-2}_\calH \geq (C \gamma_n)^{-2} \right\},$$ with $C = C_\epsilon(\nu) + \kappa^{r-1/2}(r+1/2) R \{1 + C_\delta(\nu)\}$. Then it holds with probability at least $1-\nu$ that $$\begin{aligned} \|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 &= O\left\{\gamma_n^{2r/(2r+1)}\right\},\\ \|f_{{\widehat{\alpha}}_{a^\ast}} - f\|_\calH &= O\left\{\gamma_n^{(2r-1)/(2r+1)}\right\},\end{aligned}$$ with $f^\ast = T f$. It can be shown that the stopping rule (\[eq:stopping\]) always determines a finite index, i.e., the set the minimum is taken over is not empty, see @Hanke, chapter 4.3. The theorem yields two convergence results, one in the $\calH$-norm and one in the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-norm. It holds that $\|v\|_2 = \|S^{1/2}v\|_\calH$. These are the endpoints of a continuum of norms $\|v\|_\beta = \|S^\beta v\|_\calH$, $\beta \in [0,1/2]$ that were considered in @Nemirovskii86 for the derivation of convergence rates for KCG algorithms in a deterministic setting. The convergence rate of the kernel partial least squares estimator depends crucially on the sequence $\gamma_n$ and the source parameter $r$. If $\gamma_n = O(n^{-1/2})$, this yields the same convergence rate as Theorem 2.1 of @Blan10b for kernel conjugate gradient or @Vito05 for kernel ridge regression with independent and identically distributed data. For stationary Gaussian time series we will derive concentration inequalities in the next section and obtain convergence rates depending on the source parameter $r$ and the range of dependence. Note that Theorem \[th:kpls\] is rather general and it can be applied to any kind of dependence structure, as long as the necessary concentration inequalities can be established. The next theorem derives faster convergence rates under assumptions on the effective dimensionality of operator $S$, which is defined as ${d_\lambda}= \mathrm{tr}\{(S+\lambda)^{-1} S\}$. The concept of effective dimensionality was introduced in @Zho02 to get sharp error bounds for general learning problems considered there. If $\calH$ is a finite dimensional space it was shown in @Zho02 that $d_\lambda \leq \mathrm{dim}(\calH)$. For infinite dimensional spaces it describes the complexity of the interactions between data and reproducing kernel. If $d_\lambda = O(\lambda^{-s})$ for some $s \in (0,1]$, @Cap07 showed that the order optimal convergence rates $n^{-r/(2r+s)}$ are attained for KRR with independent and identically distributed data. The effective dimensionality clearly depends on the behaviour of eigenvalues of $S$. If these converge sufficiently fast to zero, nearly parametric rates of convergence can be achieved for reproducing kernel Hilbert space methods, see, e.g., @Dicker17. In particular, the behaviour of $d_\lambda$ around zero is of interest, since it determines how ill-conditioned the operator $(S+\lambda)^{-1}$ becomes. In the following theorem we set $\lambda = \lambda_n$ for a sequence $\{\lambda_n\}_{n\in {{\mathbb N}}} \subset (0,\infty)$ that converges to zero. \[th:kpls2\] Assume that conditions \[con:k1\], \[con:k2\], \[eq:source\] hold with $r \geq 1/2$ and that the effective dimensionality ${d_\lambda}$ is known. Additionally, there are constants $C_\delta(\nu),C_\epsilon(\nu),C_\psi >0$ and a sequence $\{\gamma_n\}_{n \in {{\mathbb N}}} \subset [0,\infty)$, $\gamma_n \rightarrow 0$, such that for $\nu \in (0,1]$ and $n$ sufficiently large $$\begin{aligned} {\mathrm{P}}\left\{ \|{S_n}-{S}\|_{\calL} \leq C_\delta(\nu) \gamma_n \right\} &\geq 1 - \nu/3,\\ {\mathrm{P}}\left\{ \|(S+\lambda_n)^{-1/2}({T_n^\ast y}- {S}f)\|_\calH \leq C_\epsilon(\nu) \sqrt{d_{\lambda_n}}\gamma_n \right\} &\geq 1- \nu/3,\\ {\mathrm{P}}\left\{ \|(S+\lambda_n)^{1/2}(S_n+\lambda_n)^{-1/2}\|_{\calL} \leq C_\psi \right\} &\geq 1 - \nu/3,\end{aligned}$$ Here $\{\lambda_n\}_{n \in {{\mathbb N}}} \subset (0,\infty)$ is a sequence converging to zero such that for $n$ large enough $$\label{eq:lambda.inequ} \gamma_n \leq \lambda_n^{r-1/2}.$$ Take $\zeta_n = \max\{\sqrt{\lambda_n d_{\lambda_n}} \gamma_n, \lambda_n^{r+1/2}\}$ Define the stopping index $a^\ast$ by $$\label{eq:stopping2} a^\ast = \min\left\{1 \leq a \leq n: \sum_{i=0}^a \|{S_n}f_{{\widehat{\alpha}}_i} - {T_n^\ast y}\|^{-2}_\calH \geq (C \zeta_n)^{-2}\right\},$$ with $C=4 R \max\{1, C_\psi^2,(r-1/2)\kappa^{r-3/2} C_\delta(\nu),2^{-1/2}R^{-1} C_\psi C_\epsilon(\nu)\}$. Then it holds with probability at least $1-\nu$ that $$\begin{aligned} \|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 &= O\left\{\lambda_n^{-1/2}\zeta_n\right\},\\ \|f_{{\widehat{\alpha}}_{a^\ast}} - f\|_\calH &= O\left\{\lambda_n^{-1}\zeta_n\right\},\end{aligned}$$ with $f^\ast = T f$. The condition (\[eq:lambda.inequ\]) holds trivially for $r=1/2$ as $\gamma_n$ converges to zero. For $r >1/2$ the sequence $\lambda_n$ must not converge to zero arbitrarily fast. In its general form Theorem \[th:kpls2\] does not give immediate insight in the probabilistic convergence rates of the kernel partial least squares estimator. Therefore, we state two corollaries, where the function $d_\lambda$ is specified. In both corollaries we explicitly state the choice of the sequence $\lambda_n$ that yield the corresponding rates. \[cor:pol.ed\] Assume that there exists $s \in (0,1]$ such that $ d_\lambda = O(\lambda^{-s}) $ for $\lambda \rightarrow 0$. Then under conditions of Theorem \[th:kpls2\] with $\lambda_n = \gamma_n^{2/(2r + s)}$ it holds with probability at least $1-\nu$ that $$\begin{aligned} \|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 &= O\left\{\gamma_n^{2r/(2r+s)}\right\}.\end{aligned}$$ Polynomial decay of the effective dimensionality $d_\lambda = \mathrm{tr}\{(S+\lambda)^{-1}S\}$ occurs if the eigenvalues of $S$ also decay polynomially fast, that is, $\mu_i = c_s i^{-1/s}$ for $s \in (0,1]$, since in this case $d_\lambda = \sum\limits_{i=1}^\infty \{1+\lambda/c_s i^{1/s}\}^{-1} = O(\lambda^{-s})$. This holds, for example, for the Sobolev kernel $k(x,y) = \min(x,y)$, $x,y \in [0,1]$ and data that are uniformly distributed on $[0,1]$, see @Ras14. If $\gamma_n = n^{-1/2}$, then the KPLS estimator converges in the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-norm with a rate of $n^{-r/(2r+s)}$. This rate is shown to be optimal in @Cap07 for KRR with independent identically distributed data. Note that the rate obtained in Theorem \[th:kpls\] corresponds to $\gamma_n^{-2r/(2r+s)}$ with $s=1$, i.e., the worst case rate with respect to the parameter $s \in (0,1]$. In the next corollary to Theorem \[th:kpls2\] we assume that the effective dimensionality behaves in a logarithmic fashion. \[cor:log.ed\] Let $d_\lambda = O\{\log(1+a/\lambda)\}$ for $\lambda \rightarrow 0$ and $a>0$. Then under the conditions of Theorem \[th:kpls2\] with $\lambda_n = \gamma_n^2 \log\{ \gamma_n^{-2}\}$ and $r=1/2$ it holds with probability at least $1-\nu$ that $$\begin{aligned} \|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 &= O\left\{\gamma_n \log(1/2 \gamma_n^{-2})\right\}.\end{aligned}$$ The effective dimensionality takes the special form considered in this corollary, for example, when the eigenvalues of $S$ decay exponentially fast. This holds, for example, if the data are Gaussian and the Gaussian kernel is used, see Section \[sec:gauss.kern\]. If $\gamma_n =O(n^{-1/2})$, then the convergence rate is of order $O\{n^{-1}\log(n)\}$, which are nearly parametric. It is noteworthy that the source condition only impacts the choice of the sequence $\lambda_n$, not the convergence rates of the estimator in the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-norm. Therefore, we stated the corollary for $r=1/2$, which is a minimal smoothness condition on $f^\ast$, i.e., that $f^\ast = T f$ almost surely for an $f \in \calH$. The rates obtained in Corollaries \[cor:pol.ed\] and \[cor:log.ed\] were derived in @Dicker17 for kernel ridge regression and kernel principal component regression under the assumption of independent and identically distributed data. Concentration Inequalities for Gaussian Time Series {#sec:concentration} =================================================== Crucial assumptions of Theorem \[th:kpls\] and \[th:kpls2\] are the concentration inequalities for ${S_n}$ and ${T_n^\ast y}$ and convergence of the sequence $\{\gamma_n\}_{n \in {{\mathbb N}}}$. Here we establish such inequalities in a Gaussian setting for stationary time series. At the end of this section we will state explicit convergence rates for $f_{\widehat{\alpha}_{{a^\ast}}}$ that depend not only on the source parameter $r \geq 1/2$ and the effective dimensionality ${d_\lambda}$, but also on the persistence of the dependence in the data. The Gaussian setting is summarized in the following assumptions 1. \[D1\] $(X_h,X_0)^\T \sim \mathcal{N}_{2d}(0,{\Sigma}_h)$, $h=1,\dots,n-1$, with $${\Sigma}_h = \left[ \begin{matrix} \tau_0 & \tau_h\\ \tau_h & \tau_0 \end{matrix} \right] \otimes {\Sigma}.$$ Here ${\Sigma}\in {{\mathbb R}}^{d \times d}$ and $V=[\tau_{|i-j|}]_{i,j=1}^{n} \in {{\mathbb R}}^{n \times n}$ are positive definite, symmetric matrices and $\otimes$ denotes the Kronecker product between matrices. Furthermore $X_0 \sim \mathcal{N}_d(0,\tau_0 \Sigma)$. 2. \[D2\] For the autocorrelation function $\rho_h = \tau^{-1}_0\tau_h$ there exists a $q>0$ such that $|\rho_h| \leq (h+1)^{-q}$ for $h =0,\dots,n-1$. Condition \[D1\] is a separability condition for the covariance matrices $\Sigma_h$, $h = 0,\dots,n-1$. Due to \[D1\] the effects (on the covariance) over time and between the different variables can be treated separately. Under condition \[D2\] it is easy to see that from $q>1$ follows the absolute summability of the autocorrelation function $\rho$ and thus $\{X_t\}_{t \in {{\mathbb Z}}}$ is a short memory process. Stationary short memory processes keep many of the properties of independent and identically distributed data, see, e.g., @bBrock. On the other hand $q \in (0,1]$ yields a long memory process, see, e.g., Definition 3.1.2 in @Giraitis. Examples of long memory processes are the fractional Gaussian noise with an autocorrelation function that behaves like $(h+1)^{-2(1-H)}$, with $H \in [0,1)$ being the Hurst coefficient. Stationary long memory processes exhibit dependencies between observations that are more persistent and many statistical results that hold for independent and identically distributed data turn out to be false, see @Samoro for details. The next theorem gives concentration inequalities for both estimators ${S_n}$ and ${T_n^\ast y}$ in a Gaussian setting with convergence rates depending on the parameter $q>0$. These inequalities are the ones needed in Theorem \[th:kpls\] and Theorem \[th:kpls2\]. Recall that ${d_\lambda}= \mathrm{tr}\{(S+\lambda)^{-1}S\}$ denotes the effective dimensionality of $S$. \[th:conc.equality\] (i) Define ${\mathrm{d}}\mu_h(x,y) = {\mathrm{d}}{\mathrm{P}}^{X_h,X_0}(x,y) - {\mathrm{d}}{\mathrm{P}}^{X_0}(x){\mathrm{d}}{\mathrm{P}}^{X_0}(y)$. Under Assumptions \[con:k1\] and \[con:k2\] it holds for $\nu \in (0,1]$ with probability at least $1-\nu$ that $$\begin{aligned} \|{S_n}- {S}\|^2_{\calL} &\leq \frac{2 \nu^{-1}}{n^2}\sum\limits_{h=1}^{n-1} (n-h) \int\limits_{{{\mathbb R}}^{2d}} k^2(x,y) {\mathrm{d}}\mu_h(x,y) + \frac{\nu^{-1}}{n} \left\{ {\mathrm{E}}k^2(X_0,X_0) - \|{S}\|^2_{{\mathrm{HS}}} \right\},\\ \|T_n^\ast y - {S}f\|^2_\calH &\leq \frac{2\nu^{-1}}{n^2}\sum\limits_{h=1}^{n-1} (n-h) \int\limits_{{{\mathbb R}}^{2d}} k(x,y)f(x)f(y){\mathrm{d}}\mu_h(x,y)\\ &+ \frac{\nu^{-1}}{n} \left[ {\mathrm{E}}\left\{k(X_0,X_0) f^2(X_0)\right\} - \| {S}f\|^2_\calH + \sigma^2 {\mathrm{E}}\{ k(X_0,X_0)\} \right].\end{aligned}$$ \(ii) Assume that additionally to \[con:k1\], \[con:k2\] also \[D1\], \[D2\] for $q >0$ are fulfilled. Denote $M=\sup_{x \in {{\mathbb R}}^d} |f(x)|$. Then there exists a constant $C(q)>0$ such that $$\begin{aligned} \|{S_n}- {S}\|_\calL &\leq \nu^{-1/2} \{\gamma_n^2(q) \kappa C_\gamma + n^{-1}(\kappa^2-\|S\|_{\mathrm{HS}}^2)\}^{1/2},\\ \|T_n^\ast y - {S}f\|_\calH &\leq \nu^{-1/2} \left[\gamma_n^2(q) M C_\gamma + n^{-1}\left\{ \kappa (M + \sigma^2) - \| {S}f\|^2_\calH \right\}\right]^{1/2} ,\end{aligned}$$ for $C_\gamma = C(q)\{(2\pi)^d \mathrm{det}(\Sigma)\}^{-1/2} \kappa d^{1/2}(1-4^{-q})^{-1/4(d+2)}$. The function $\gamma_n(q)$, $q>0$, is defined as $$\gamma_n(q) = \left\{ \begin{array}{clc} n^{-1/2} &,& q>1\\ n^{-1/2} \log(1/2n) &,& q=1\\ n^{-q/2} &,& q \in (0,1). \end{array} \right.$$ \(iii) Let \[con:k1\], \[con:k2\] and \[eq:source\] hold. Let $\gamma_n(q)$ be the function as defined in (ii). Then there exists a constant $\tilde{C}_\epsilon>0$ such that it holds with probability at least $1-\nu$ for $\lambda>0$ that $$\|(S+\lambda)^{-1/2}(T_n^\ast y - {S_n}f\|_\calH \leq \nu^{-1/2} \tilde{C}_\epsilon \sigma \sqrt{{d_\lambda}}\gamma_n(q).$$ \(iv) Let \[con:k1\], \[con:k2\], \[eq:source\], \[D1\] and \[D2\] hold. Let $\lambda_n^{-1/2}d_{\lambda_n}^{1/2} \gamma_n(q) \rightarrow 0$ for a sequence $\lambda_n \rightarrow 0$ and $\gamma_n(q)$ the function defined in (ii). Then there exists an $n_0 = n_0(\nu,q) \in {{\mathbb N}}$ such that with probability at least $1-\nu$ we have for all $n \geq n_0$ $$\|(S+\lambda_n)^{1/2}({S_n}+\lambda_n)^{-1/2}\|_\calL \leq \sqrt{2}.$$ The first part of the theorem is general and can be used to derive concentration inequalities not only in the Gaussian setting and is of interest in itself. The convergence rate is controlled by the sums appearing on the right hand side. If these sums are of $O(n)$ then the mean squared error of both ${S_n}$ and ${T_n^\ast y}$ will converge to zero with a rate of $n^{-1}$. On the other hand, if the sums are of order $O(n^{2-q})$ for some $q\in (0,1)$, the mean squared errors will converge with the reduced rate $n^{-q}$. The second part derives explicit concentration inequalities in the Gaussian setting described by \[D1\] and \[D2\] with rates depending on the range of the dependence measured by $q>0$. These inequalities appear in Theorem \[th:kpls\]. Parts (iii) and (iv) give the additional probabilistic bounds needed to apply Theorem \[th:kpls2\]. The condition $\lambda_n^{-1/2} d^{1/2}_{\lambda_n} \gamma_n(q) \rightarrow 0$ in Theorem \[th:conc.equality\] (iv) is fulfilled in the settings of Corollary \[cor:pol.ed\] and Corollary \[cor:log.ed\]. Theorem \[th:kpls\], Corollary \[cor:pol.ed\], Corollary \[cor:log.ed\] and Theorem \[th:conc.equality\] together imply \[cor:convergence\] Let the conditions of Theorem \[th:kpls2\] and \[D1\], \[D2\] hold. \(i) Assume that there exists $s\in (0,1]$ such that $d_\lambda = O(\lambda^{-s})$ for $\lambda \rightarrow 0$. Then with probability at least $1-\nu$ $$\|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 = \left\{ \begin{array}{cc} O\{n^{-r/(2r+s)}\}, & q>1,\\ O\{n^{-q r/(2r+s)}\}, & q \in (0,1). \end{array} \right.$$ If instead of conditions of Theorem \[th:kpls2\], conditions of Theorem \[th:kpls\] are assumed, then the convergence rates above have $s=1$. \(ii) Assume that there exists $a>0$ such that ${d_\lambda}= O\{\log(1+a/\lambda)\}$ for $\lambda \rightarrow 0$ and $r=1/2$. Then with probability at least $1-\nu$ $$\|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 = \left\{ \begin{array}{cc} O\{n^{-1/2}\log(1/2 n)\}, & q>1,\\ O\{n^{-q/2} \log(1/2 n^q)\}, & q \in (0,1). \end{array} \right.$$ Hence, for $q>1$ the kernel partial least squares algorithm achieves the same rates as if the data were independent and identically distributed. For $q \in (0,1)$ the convergence rates become substantially slower, highlighting that dependence structures that persist over a long time can influence the convergence rates of the algorithm. Source condition and effective dimensionality for Gaussian kernels {#sec:gauss.kern} ================================================================== The source condition \[eq:source\] and the effective dimensionality $d_\lambda$ are of great importance in the convergence rates derived in previous sections. Here we investigate these conditions for the reproducing kernel Hilbert space corresponding to the Gaussian kernel $k(x,y) = \exp(-l\|x-y\|^2)$, $x,y \in {{\mathbb R}}^d$, $l>0$, for $d=1$. Hence, the space $\calH$ is the space of all analytic functions that decay exponentially fast, see @Steinwart05anexplicit. We also impose the normality conditions \[D1\] and \[D2\] on $\{X_t\}_{t\in {{\mathbb Z}}}$, where now $\sigma^2_x =\Sigma \in {{\mathbb R}}$ due to $d=1$. The following proposition derives a more explicit representation for $f \in \calH$. \[prop:source\] Assume that \[con:k1\],\[con:k2\] and \[eq:source\] hold for $r \geq 1/2$. Let $d=1$, $X_0 \sim \mathcal{N}(0,\sigma^2_x), \sigma^2_x > 0$ and consider the Gaussian kernel $k(x,y) = \exp\{-l (x-y)^2\}$ for $x,y \in {{\mathbb R}}$, $l>0$. Then $f$ can be expressed for $\mu = r - 1/2 \in {{\mathbb N}}$ via $f(x) = \sum_{i=1}^\infty c_i L_\mu(x,z_i)$ for fixed $\{z_i\}_{i=1}^\infty,\{c_i\}_{i=1}^\infty \subset {{\mathbb R}}$ such that $\sum_{i,j=1}^\infty c_i c_j k(z_i,z_j) \leq R^2$, $R>0$. Here we have for $x,z \in {{\mathbb R}}$ $$\begin{aligned} L_\mu(x,z) &= \exp \left[ -1/2 \left\{ \frac{\det(\Lambda)(x^2+z^2)-2 l^{\mu+1} x z}{\det(\Lambda_{1:\mu})} \right\} \right], \end{aligned}$$ with $\Lambda \in {{\mathbb R}}^{(\mu+1)\times (\mu+1)}$ being a tridiagonal matrix with elements $$\Lambda_{i,j} = \left\{ \begin{array}{cll} \sigma^{-2}_x + 2 l &, & i=j<\mu+1\\ l &, & i=j=\mu+1\\ -l &, & |i-j| = 1\\ 0&, & else \end{array}\right.$$ for $i,j=1,\dots,\mu+1$ and $\Lambda_{1:\mu}$ is the $\mu \times \mu$-dimensional sub-matrix of $\Lambda$ including the fist $\mu$ columns and rows. Conversely any function $f^\ast = T f$ with $f$ of the above form fulfills a source condition $\ref{eq:source}$ with $r = \mu + 1/2$, $\mu \in {{\mathbb N}}$. Hence if we fix an $r \geq 1/2$ with $r-1/2 \in {{\mathbb N}}$ this theorem gives us a way to construct functions $f \in \calH$ with $f^\ast = Tf$ that fulfill \[eq:source\]. The next proposition derives the effective dimensionality ${d_\lambda}$ in this setting: \[prop:ed\] Let $d=1$, $X_0 \sim \mathcal{N}(0,\sigma^2_x)$ for some $\sigma^2_x>0$ and consider the Gaussian kernel $k(x,y) = \exp\{-l(x-y)^2\}$, $x,y \in {{\mathbb R}}$, $l>0$. Then there is a constant $D>0$ such that it holds for any $\lambda \in (0,1]$ $$d_\lambda = \mathrm{tr}\{(S+\lambda)^{-1} S\} \leq D\log(1+a/\lambda),$$ with $a =\sqrt{2}(1+\beta+\sqrt{1+\beta})^{-1/2} $, $\beta = 4 l \sigma^2_x$. With the latter result Corollary \[cor:log.ed\] is applicable and we expect convergence rates for the kernel partial least squares algorithm of order $O\{\gamma_n \log(1/2 \gamma_n^{-2})\}$ for a sequence $\{\gamma_n\}_n$ as in Theorem \[th:kpls2\]. Simulations {#sec:simulations} =========== To validate the theoretical results of the previous sections we conducted a simulation study. The reproducing kernel Hilbert space is chosen to correspond to the Gaussian kernel $k(x,y) = \exp(-l\|x-y\|^2)$, $x,y \in {{\mathbb R}}^d$, $l=2$, for $d=1$. The source parameter is taken $r=4.5$ and we consider the function $$f(x) = 4.37^{-1}\{3 {L}_4(x,-4) - 2 {L}_4(x,3)+ 1.5 {L}_4(x,9)\}, ~~ x \in {{\mathbb R}}.$$ The normalization constant is chosen such that $f$ takes values in $[-0.35,0.65]$ and $L_4$ is the exponential function given in Proposition \[prop:source\]. The function $f$ is shown in Figure \[fig:func\]. ![\[fig:func\] The function $f$ evaluated on $[-7.5,7.5]$ (black) and one realisation of the noisy data $y = f(x) + \varepsilon$ (grey). ](f_plot){width=".5\linewidth"} In condition \[D1\] we set $\sigma^2_x ={\Sigma}= 4$. For the matrix $V^2 = [\tau_{|i-j|}]_{i,j=1}^n \in {{\mathbb R}}^{n \times n}$ we choose three different structures for $n\in\{200,400,1000\}$. In the first setting $\tau_h = \mathbb{I}(h=0)$, which corresponds to independent data. The second setting with $\tau_h = 0.9^{-h}$ implies an autoregressive process of order one. Finally, the third setting with $\tau_h = (1+h)^{-q}$, $q=1/4$, $h =0,\dots,n-1$ leads to the long range dependent case. In a Monte Carlo simulation with $M=1000$ repetitions the time series $\{X_t^{(j)}\}_{t=1}^n$ are generated via $X^{(j)} = V N^{(j)}$ with $N^{(j)} \sim \mathcal{N}_n(0,\sigma^2 I_n)$, $j=1,\dots,M$, where $I_n$ is the $n \times n$-dimensional identity matrix. The residuals $\varepsilon_1^{(j)},\dots,\varepsilon_n^{(j)}$ are generated as independent standard normally distributed random variables and independent of $\{X_t^{(j)}\}_{t=1}^n$ . The response is defined as $y_t^{(j)} = f(X_t^{(j)}) + \eta\, \varepsilon_t^{(j)}$, $t=1,\dots,n$, $j=1,\dots,M$, with $\eta = 1/16$. The kernel partial least squares and kernel conjugate gradient algorithms are run for each sample $\{(X_t^{(j)},y_t^{(j)})^\T\}_{t=1}^n$, $j=1,\dots,M$, with a maximum of $40$ iteration steps. We denote the estimated coefficients with ${\widehat{\alpha}}_1^{(j,m)},\dots,{\widehat{\alpha}}_{40}^{(j,m)}$, $j=1,\dots,M$, with $m=CG$ meaning that the kernel conjugate gradient algorithm was employed and $m=PLS$ that kernel partial least squares was used to estimate $\alpha_1,\dots,\alpha_n$. The squared error in the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-norm is calculated via $$\widehat{e}_{n,\tau}^{(j,m)} = \min\limits_{a=1,\dots,40} \left[ \frac{1}{\sqrt{2\pi\sigma_x^2}}\int\limits_{-\infty}^\infty \left\{f_{\widehat{\alpha}_a^{(j,m)}}(x) - f(x) \right\}^2 \exp \left(-\frac{1}{2\sigma_x^2} x^2 \right) {\mathrm{d}}x \right],$$ for $j=1,\dots,M$, $n = 200,400,\dots,1000$ and $m \in \{CG, PLS\}$. The results of the Monte-Carlo simulations are depicted in the boxplots of Figure \[fig:box\]. ![\[fig:box\] Boxplots of the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ of kernel partial least squares (left side of each panel) and kernel conjugate gradient (right side of each panel) for different autocovariance functions $\tau$ and $n = 200,400,1000$. On the left is $\tau_h = \mathbb{I}(h=0)$, in the middle $\tau_h = 0.9^{-h}$ and on the right $\tau_h = (h+1)^{-1/4}$. ](MSE_box_1){width="\ssize"} ![\[fig:box\] Boxplots of the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ of kernel partial least squares (left side of each panel) and kernel conjugate gradient (right side of each panel) for different autocovariance functions $\tau$ and $n = 200,400,1000$. On the left is $\tau_h = \mathbb{I}(h=0)$, in the middle $\tau_h = 0.9^{-h}$ and on the right $\tau_h = (h+1)^{-1/4}$. ](MSE_box_2){width="\ssize"} ![\[fig:box\] Boxplots of the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ of kernel partial least squares (left side of each panel) and kernel conjugate gradient (right side of each panel) for different autocovariance functions $\tau$ and $n = 200,400,1000$. On the left is $\tau_h = \mathbb{I}(h=0)$, in the middle $\tau_h = 0.9^{-h}$ and on the right $\tau_h = (h+1)^{-1/4}$. ](MSE_box_3){width="\ssize"} For kernel partial least squares (left panels) one observes that independent and autoregressive dependent data have roughly the same convergence rates, although the latter have a somewhat higher error. In contrast, the long range dependent data show slower convergence with the larger interquartile range, supporting the theoretical results of Corollary \[cor:convergence\]. The ${\calL^2\left({\mathrm{P}}^{X} \right)}$-error of kernel conjugate gradient estimators is generally slightly higher than that of kernel partial least squares. Nonetheless, both of them have a similar behaviour. We also investigated the the stopping indices $a=1,\dots,40$ for which the errors $\widehat{e}_{n,\tau}^{(j,m)}$ were attained. These are shown in Figure \[fig:a\] for independent and identically distributed data. ![\[fig:a\] Boxplots of the optimal indices $a\in\{1,\dots,40\}$ for which the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ were attained. Kernel partial least squares is on the left of each panel and kernel conjugate gradient on the right. On the left is $n = 200$, on the right $n=1000$. The data were assumed to be independent and identically distributed. ](MSE_sim_a_200){width="\ssize"} ![\[fig:a\] Boxplots of the optimal indices $a\in\{1,\dots,40\}$ for which the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ were attained. Kernel partial least squares is on the left of each panel and kernel conjugate gradient on the right. On the left is $n = 200$, on the right $n=1000$. The data were assumed to be independent and identically distributed. ](MSE_sim_a_1000){width="\ssize"} It can be seen that the optimal indices for both algorithms have a rather similar behaviour. Kernel conjugate gradient stops slightly later, but overall the differences seem negligible. Figure \[fig:lines\] shows the mean (over $j$) of the estimated ${\calL^2\left({\mathrm{P}}^{X} \right)}$ errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ for different $n$, $\tau$ and $m \in \{CG,PLS\}$. The errors were multiplied by $n/\log(n)$ to illustrate the convergence rates. According to Proposition \[prop:ed\] and Corollary \[cor:convergence\] (ii) we expect the rates for the independent and autoregressive cases to be $n^{-1}\log(n)$, which is verified by the fact that the solid black and grey lines are roughly constant. For the long range dependent case we expect worse convergence rates which are also illustrated by the divergence of the dashed black line. ![\[fig:lines\] Mean of the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ of kernel partial least squares (left) and kernel conjugate gradient (right) for $n = 200,400,\dots,1000$ multiplied by $n/\log(n)$. The solid black line is for $\tau_h = \mathbb{I}(h=0)$, the grey line for $\tau_h = 0.9^{-h}$ and the dashed black line for $\tau_h = (h+1)^{-1/4}$. ](MSE_sim_pls_mean){width="\ssize"} ![\[fig:lines\] Mean of the ${\calL^2\left({\mathrm{P}}^{X} \right)}$-errors $\{\widehat{e}_{n,\tau}^{(j,m)}\}_{j=1}^{M}$ of kernel partial least squares (left) and kernel conjugate gradient (right) for $n = 200,400,\dots,1000$ multiplied by $n/\log(n)$. The solid black line is for $\tau_h = \mathbb{I}(h=0)$, the grey line for $\tau_h = 0.9^{-h}$ and the dashed black line for $\tau_h = (h+1)^{-1/4}$. ](MSE_sim_cg_mean){width="\ssize"} Application to Molecular Dynamics Simulations {#sec:protein} ============================================= The collective motions of protein atoms are responsible for its biological function and molecular dynamics simulations is a popular tool to explore this [@Henz07]. Typically, the $p \in {{\mathbb N}}$ backbone atoms of a protein are considered for the analysis with the relevant dynamics happening in time frames of nanoseconds. Although the dynamics are available exactly, the high dimensionality of the data and large number of observations can be cumbersome for regression analysis, e.g., due to the high collinearity in the columns of the covariates matrix. Many function-dynamic relationships are also non-linear [@Hub09]. A further complication is the fact that the motions of different backbone atoms are highly correlated, making additive non-parametric models for the target function $f^\ast$ less suitable. We consider T4 Lysozyme (T4L) of the bacteriophage T4, a protein responsible for the hydrolisis of 1,4-beta-linkages in peptidoglycans and chitodextrins from bacterial cell walls. The number of available observations is $n=4601$ and T4L consists of $p=486$ backbone atoms. Denote with $A_{t,i} \in {{\mathbb R}}^3$ the $i$-th atom, $i=1,\dots,p$, at time $t=1,\dots,n$ and $c_i \in {{\mathbb R}}^3$ the $i$-th atom in the (apo) crystal structure of T4L. A usual representation of the protein in a regression setting is the Cartesian one, i.e., we take as the covariate $X_t = (A_{1,t}^\T,\dots,A^\T_{p,t})^\T$, $t=1,\dots,n$, see @Bro83. The functional quantity to predict is the root mean square deviation of the protein configuration $X_t$ at time $t = 1,\dots,n$ from the (apo) crystal structure $C = (c_1^\T,\dots,c_d^\T)^\T$, i.e., $$y_t = \left\{ p^{-1} \sum\limits_{i=1}^{p} \|X_{i,t}-C_i\|^2 \right\}^{1/2}.$$ This nonlinear function was previously considered in @Hub09, where it was established that linear models are insufficient for the prediction. Figure \[fig:sample.series\] shows the time series corresponding to $X_{t,1}$ (i.e., the first coordinate of the first atom of T4L) on the left and the functional quantity $y_t$ on the right. These plots reveal certain persistent dependence over time. ![\[fig:sample.series\] Time series of $X_{t,1}$, i.e., the first coordinate of the first atom T4L consists of (left) and the root mean squared deviation $y_t$ between the protein configuration at time $t$ and the (apo) crystal structure. ](T4L_X1){width="\ssize"} ![\[fig:sample.series\] Time series of $X_{t,1}$, i.e., the first coordinate of the first atom T4L consists of (left) and the root mean squared deviation $y_t$ between the protein configuration at time $t$ and the (apo) crystal structure. ](T4L_Y){width="\ssize"} Fitting autoregressive moving average models of order $(3,2)$ ($ARMA(3,2)$) to $y_t$ and $ARMA(5,2)$ to $X_{t,1}$ shows that the smallest root of their respective characteristic polynomial is close to one ($1.009$ for $y_t$ and $1.003$ for $X_{t,1}$), highlighting that we are on the border of stationarity, see, e.g., @bBrock. Figure \[fig:acf\] depicts the autocorrelation functions of $X_{t,1}$ and $y_t$, the theoretical autocorrelation function of the corresponding autoregressive moving average process and $\rho_h \propto (h+1)^{-q}$ for $q=0.134$ for $X_{t,1}$ and $q=0.066$ for $y_t$. The latter, as highlighted in Section \[sec:concentration\], is an autocorrelation function for a stationary long range dependent process. ![\[fig:acf\] Autocorrelation plots of $X_{t,1}$ (left) and $y_t$ (right). The estimated autocorrelation function is grey, the theoretical one of a fitted $ARMA(3,2)$ process is solid black and $\rho_h \propto (h+1)^{-q}$ for a suitable choice of $q>0$ is dashed black. ](T4L_acf_X){width="\ssize"} ![\[fig:acf\] Autocorrelation plots of $X_{t,1}$ (left) and $y_t$ (right). The estimated autocorrelation function is grey, the theoretical one of a fitted $ARMA(3,2)$ process is solid black and $\rho_h \propto (h+1)^{-q}$ for a suitable choice of $q>0$ is dashed black. ](T4L_acf_y){width="\ssize"} These plots suggest that $X_{t,1}$ and $y_t$ follow some long-range stationary process. We apply kernel partial least squares to this data set with the Gaussian kernel $k(x,y) = \exp(-l \|x-y\|^2)$, $x,y \in {{\mathbb R}}^{3p}$, $l>0$. The function $f$ we aim to estimate is a distance between protein configurations, so using a distance based kernel seems reasonable. Moreover, we also investigated the impact of other bounded kernels such as triangular and Epanechnikov and obtained similar results. The first $25\%$ of the data form a training set to calculate the kernel partial least squares estimator and the remaining data are used for testing. The parameter $l>0$ is calculated via cross validation on the training set. In our evaluation we obtained $l = 0.0189$. Figure \[fig:prot\] compares the observed response in the test set with the prediction on the test set obtained by kernel partial least squares, kernel principal component regression and linear partial least squares. ![\[fig:prot\] Correlation (left) and residual sum of squares (right) between predicted values and the observed response on the test set depending on the number of used components for kernel partial least squares (solid black), partial least squares (grey) and kernel principal component regression (dashed black). ](T4L_pred_cor){width="\ssize"} ![\[fig:prot\] Correlation (left) and residual sum of squares (right) between predicted values and the observed response on the test set depending on the number of used components for kernel partial least squares (solid black), partial least squares (grey) and kernel principal component regression (dashed black). ](T4L_pred_rss){width="\ssize"} Apparently, kernel partial least squares show the best performance and the kernel principal components algorithm is able to achieve comparable prediction with more components only. Obviously, linear partial least squares can not cope with the non-linearity of the problem. This application highlights that kernel partial least squares still delivers a robust prediction even when the dependence in the data is more persistent, if enough observations are available. Proofs {#sec:proofs} ====== Proof of Theorem \[th:kpls\] {#sec:int.proof1} ---------------------------- The proof of Theorem \[th:kpls\] makes use of the connection between the partial least squares and the conjugate gradient algorithm. This section is structured as follows: First we will introduce the link between kernel partial least squares and kernel conjugate gradient. We will state some key facts about orthogonal polynomials and their relationship to the algorithm in Lemma \[lem:orth.pol\]. Then the consistency of kernel partial least squares is shown with the help of three error bounds that are obtained in Lemmas \[lem:error.difference\] – \[lem:derivative\]. With a slight abuse of notation we define $f_i = f_{{\widehat{\alpha}}_i}$ for $i=1,\dots,n$. We consider the kernel partial least squares algorithm as an optimization problem $$\label{ref:pls.opt} f_i = \arg\min\limits_{g \in \mathcal{K}_i({S_n},{T_n^\ast y})} \|y - T_n g\|^2, ~~ i=1,\dots,n.$$ This is the conjugate gradient algorithm CGNE discussed in chapter 2.2 of @Hanke. ### Orthogonal polynomials and some notation Denote with $\mathcal{P}_i$ the set of polynomials of degree at most $i = 0,\dots,n$. For functions $\psi,\phi:{{\mathbb R}}\rightarrow {{\mathbb R}}$ and $r \in {{\mathbb N}}_0$ define the inner products $[\psi,\phi]_r = \langle \psi({S_n}){T_n^\ast y}, {S_n}^r \phi({S_n}) {T_n^\ast y}\rangle_\calH$. From the definition of the Krylov space it is immediate that every element $v \in \calK_i({S_n},{T_n^\ast y})$, $i=1,\dots,n$, can be represented by a polynomial $q \in \mathcal{P}_{i-1}$ via $v = q({S_n}) {T_n^\ast y}$. The following discussion is based on @Hanke, chapter 2. There exist two sequences of polynomials $\{p_i\}_{i=0}^n, \{q_i\}_{i=0}^n \subset \mathcal{P}_i$, such that $f_i = q_{i-1}({S_n}) {T_n^\ast y}$ with $q_{-1} = 0$ and ${T_n^\ast y}- {S_n}f_i = p_i({S_n}){T_n^\ast y}$. Both sequences are connected by the equation $p_i(x) = 1 - x q_{i-1}(x)$, $x \in {{\mathbb R}}$, and the polynomials $\{p_i\}_{i=0}^n$ are orthogonal with respect to $[\cdot,\cdot]_0$. We will also consider other sequences of polynomials, namely $\{{q_{i}^{[r]}}\}_{i=0}^n,\{{p_{i}^{[r]}}\}_{i=0}^n \subset \calP_i$, ${q_{-1}^{[r]}} = 0$, such that ${p_{i}^{[r]}}(x) = 1 - x {q_{i-1}^{[r]}}(x)$, $x \in {{\mathbb R}}$, and the sequence $\{{p_{i}^{[r]}}\}_{i = 0}^n$ is orthogonal with respect to $[\cdot,\cdot]_{r}$. This yields for every $r \in {{\mathbb N}}_0$ a separate conjugate gradient algorithm with solution ${f_{i}^{[r]}} = {q_{i-1}^{[r]}}({S_n}) {T_n^\ast y}\in \calK_{i}({S_n},{T_n^\ast y})$ and residuals ${T_n^\ast y}- {S_n}{f_{i}^{[r]}} = {p_{i}^{[r]}}({S_n}){T_n^\ast y}$, $i=1,\dots,n$. The $p_i^{[r]}$, $i=0,\dots,n$, $r \in {{\mathbb N}}_0$, are called residual polynomials. As ${S_n}$ is self-adjoint, positive semi-definite and the kernel is bounded by $\kappa$ we know that its spectrum is a subset of $[0,\kappa]$, see @Cap07. This also implies that $\max\{\|{S}\|_{\calL},\|{S_n}\|_\calL\} \leq \kappa$, with $\|\cdot\|_\calL$ denoting the operator norm. The $i$ distinct roots of ${p_{i}^{[r]}}$ will be denoted by $0 < {x_{1,i}^{[r]}} < \dots {x_{i,i}^{[r]}} < \kappa$, $i=1,\dots,n$. We will summarize some key facts about the orthogonal polynomials in the next lemma. \[lem:orth.pol\] Let $r,s \in {{\mathbb N}}_0$ and $i=1,\dots,n$. Then we have: - The roots of consecutive orthogonal polynomials interlace, i.e., for $j=1,\dots,i$ it holds $$0 < {x_{j,i+1}^{[r]}} < {x_{j,i}^{[r]}}<{x_{j,i}^{[r+1]}}<{x_{j+1,i+1}^{[r]}}<{x_{j+1,i}^{[r]}} < \dots <{x_{i,i}^{[r+1]}} < {x_{i+1,i+1}^{[r]}} < \kappa,$$ - the optimality property $[{p_{i}^{[1]}},{p_{i}^{[1]}}]_{0}^{1/2} = \|{T_n^\ast y}- {S_n}{f_{i}^{[1]}}\|_\calH \leq \|{T_n^\ast y}- {S_n}h\|_\calH$ holds for all $h \in \calK_{i}({S_n},{T_n^\ast y})$, - on $x \in [0,{x_{1,i}^{[r]}}]$ it holds $0 \leq {p_{i}^{[r]}}(x) \leq 1$ and ${q_{i}^{[r]}}(x) \leq \left|\left({p_{i}^{[r]}}\right)'(0)\right| $, - $ {p_{n}^{[r]}} = {p_{n}^{[s]}}, $ - $\left({p_{i}^{[r]}}\right)'(0) = - \sum_{j=1}^i \left({x_{j,i}^{[r]}}\right)^{-1}$, - for $r \geq 1$ define $\phi_i(x) = {p_{i}^{[r]}}(x)\left({x_{1,i}^{[r]}}\right)^{1/2} \left( {x_{1,i}^{[r]}}- x\right)^{-1/2}$, $x \in [0,{x_{1,i}^{[r]}}]$, $i=1,\dots,n$. Then it holds for $u \geq 0$ that $x^u \phi^2_i(x) \leq u^u \left|\left({p_{i}^{[r]}}\right)'(0)\right|^{-u}$ with the convention $0^0 = 1$. [*Proof*]{}: (i) See @Hanke, Corollary 2.7. \(ii) See @Hanke, Proposition 2.1. \(iii) Due to part (i) we know that all $i$ roots of the polynomial ${p_{i}^{[r]}}$ are contained in $(0,\kappa)$. Furthermore ${p_{i}^{[r]}}(0) = 1 - 0{q_{i}^{[r]}} = 1$. Thus ${p_{i}^{[r]}}$ is convex and falling in $[0,{x_{1,i}^{[r]}}]$ and the first assertion follows. Because of the convexity of ${p_{i}^{[r]}}$ on $[0,{x_{1,i}^{[r]}}]$ we get ${q_{i}^{[r]}} (x) = x^{-1}\{1-{p_{i}^{[r]}}(x)\} \leq \left|\left({p_{i}^{[r]}}\right)'(0)\right|$. \(iv) See the discussion in @Hanke preceding Proposition 2.1 and use the facts that ${T_n^\ast y}\in \mathrm{range}({S_n})$ and ${S_n}$ is an operator of rank $n$. \(v) Write ${p_{i}^{[r]}}(x) = \prod_{j=1}^i(1 - x/ {x_{j,i}^{[r]}})$, $x \in [0,\kappa]$, and the result is immediate. \(vi) See equation (3.10) in @Hanke. $\square$ We denote for $x\geq 0$ by ${P}_x$ the orthogonal projection operator on the eigenspace corresponding to the eigenvalues of ${S_n}$ that are smaller or equal $x$ and ${P}_x^\perp = I_\calH - P_x$ with $I_\calH:\calH \rightarrow \calH$ being the identity operator. ### Preparation for the proof An important technical result that will be useful in the upcoming proof is \[lem:op.inequality\] Let $B,C:\calH \rightarrow \calH$ be two positive semi-definite, self-adjoint operators with $\max\{\|B\|_\calL,\|C\|_\calL\} \leq \kappa$. Then it holds for any $r \geq 0$ with $\zeta = \max\{r-1,0\}$ $$\|B^r - C^r\|_\calL \leq (\zeta+1) \kappa^\zeta \|B - C\|_\calL^{r-\zeta}.$$ [*Proof*]{}: See @Blan10b, Lemma A.6. $\square$ For the remainder of the proof we assume that we are on the set where it holds with probability at least $1-\nu$, $\nu \in (0,1]$, that $\|{S_n}- {{S}}\|_\mathcal{L} \leq C_\delta(\nu) \gamma_n$ and $\|{T_n^\ast y}- {{S}}f\| \leq C_\epsilon(\nu) \gamma_n$ for a sequence $\{\gamma_n\}_n$ converging to zero and constants $C_\delta=C_\delta(\nu),C_\epsilon=C_\epsilon(\nu)>0$. With Lemma 2.4 in @Hanke we see that the stopping iteration (\[eq:stopping\]) can also be expressed as $$\label{eq:alt.stop} {{a^\ast}}= \min\left\{1 \leq a \leq n: \|{S_n}{f_{a}^{[1]}}- {T_n^\ast y}\|_\calH \leq {C}\gamma_n \right\},$$ i.e., we stop the kernel partial least squares algorithm when a discrepancy principle for ${f_{a}^{[1]}}$ holds. It is easy to see that from \[eq:source\] it follows for $r \geq 1/2$ that 1. \[eq:source.H\] There exist $\mu \geq 0$, $R >0$ and $u \in \calH$ such that $f = {S}^{\mu} u$ and $\|u\|_\calH \leq R$. This condition is known as the Hölder source condition with $\mu = r-1/2$. Recall that $\calH \subseteq {\calL^2\left({\mathrm{P}}^{X} \right)}$ and $T:\calH \rightarrow {\calL^2\left({\mathrm{P}}^{X} \right)}$ is the change of space operator. Using the fact that $T$, $T^\ast$ are adjoint operators, $f_{{a^\ast}}= T f_{{a^\ast}}$ and $f^\ast = T f$ for $r \geq 1/2$ we see $$\begin{aligned} \|f_{{a^\ast}}- f^\ast\|_2 = \|T(f_{{a^\ast}}- f)\|_2 = \langle S(f_{{a^\ast}}- f), f_{{a^\ast}}- f\rangle_\calH = \|S^{1/2}(f_{{a^\ast}}- f)\|_\calH.\end{aligned}$$ An application of Lemma \[lem:op.inequality\] yields $$\begin{aligned} &\|f_{{a^\ast}}- f^\ast\|_2 =\|{{S}^{1/2}}(f_{{a^\ast}}- f)\|_\calH \leq \|{{S}^{1/2}}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH + \|{{S}^{1/2}}({f_{{{a^\ast}}}^{[1]}} - f)\|_\calH \notag \\ & \leq C_\delta^{1/2}\gamma_n^{1/2} \left( \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH + \|{f_{{{a^\ast}}}^{[1]}} - f\|_\calH \right) + \|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH + \|{S_n}^{1/2}({f_{{{a^\ast}}}^{[1]}}-f)\|_\calH. \label{eq:main.inequality}\end{aligned}$$ The following lemmas will deal with bounding the quantities in (\[eq:main.inequality\]). \[lem:error.difference\] Assume $C_x \in (0,1]$ such that $x_\ast = (C_x\gamma_n)^{1/(\mu+1)} < {x_{1,{{a^\ast}}-1}^{[1]}}$ and ${C}> C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R$. Under the conditions of the theorem it holds $\mu \geq 0$ $$\begin{aligned} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH & \leq \gamma^{\mu/(\mu+1)}_n \frac{{C}} {C_x^{1/(\mu+1)} \left[ 1- {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} \right]^2}\\ \|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH & \leq \gamma^{(2\mu+1)/(2\mu+2)}_n \frac{{C}} {C_x^{1/(2\mu+2)} \left[ 1- {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} \right]}.\end{aligned}$$ [*[Proof:]{}*]{} If the inner products $[\cdot,\cdot]_0$ and $[\cdot,\cdot]_1$ are the same the proof is done because both polynomial sequences are identical. We now observe that we have for ${{a^\ast}}= n$ due to Lemma \[lem:orth.pol\] (iv) $q_{n-1}(x) - {q_{n-1}^{[1]}}(x) = x^{-1}\{{p_{n}^{[1]}}(x)-p_{n}(x)\} = 0$, i.e., $\|f_{{{a^\ast}}}-{f_{{{a^\ast}}}^{[1]}}\|_\calH = 0$ and $\|{S_n}^{1/2}(f_{{{a^\ast}}}-{f_{{{a^\ast}}}^{[1]}})\|_\calH = 0$ and the proof is done. If the inner products differ and we have $0<{{a^\ast}}<n$ it holds $f_{{a^\ast}}\neq {f_{{{a^\ast}}}^{[1]}}$. Proposition 2.8 in @Hanke can now be applied for $0 < {{a^\ast}}< n$ and yields $q_{{{a^\ast}}-1}(x) - {q_{{{a^\ast}}-1}^{[1]}}(x) = x^{-1}\{{p_{{{a^\ast}}}^{[1]}}(x)-p_{{{a^\ast}}}(x)\} = \theta_{{{a^\ast}}} {p_{{{a^\ast}}-1}^{[2]}}(x)$, $x\geq 0$, with $\theta_{{{a^\ast}}} = ({p_{{{a^\ast}}}^{[1]}})'(0)-({p_{{{a^\ast}}}^{[0]}})'(0)>0$. We get $f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}} = q_{{{a^\ast}}-1}({S_n}){T_n^\ast y}- {q_{{{a^\ast}}-1}^{[1]}}({S_n}) {T_n^\ast y}= \theta_{{{a^\ast}}} {p_{{{a^\ast}}-1}^{[2]}}({S_n}) {T_n^\ast y}$. Proposition 2.9 in @Hanke yields $\theta_{{{a^\ast}}} = {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{-1} {{\left[{p_{{{a^\ast}}}^{[1]}},{p_{{{a^\ast}}}^{[1]}}\right]_{0}}}$. The optimality property of ${f_{{{a^\ast}}}^{[1]}}$ in Lemma \[lem:orth.pol\] (ii) shows that $$\label{eq:pol.opt} \|{T_n^\ast y}- {S_n}{f_{{{a^\ast}}}^{[1]}}\|_\calH = \|{p_{{{a^\ast}}}^{[1]}}({S_n}){T_n^\ast y}\|_\calH = {{\left[{p_{{{a^\ast}}}^{[1]}},{p_{{{a^\ast}}}^{[1]}}\right]_{0}}}^{1/2} \leq {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2}.$$ Combining these results yields $$\label{eq:est.difference} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH = \frac{{{\left[{p_{{{a^\ast}}}^{[1]}},{p_{{{a^\ast}}}^{[1]}}\right]_{0}}}}{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \leq \frac{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}}{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}}\|{p_{{{a^\ast}}}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH.$$ Recall that ${x_{1,{{a^\ast}}-1}^{[2]}}$ denotes the first root of ${p_{{{a^\ast}}-1}^{[2]}}$. It holds for any $0 \leq x \leq {x_{1,{{a^\ast}}-1}^{[2]}}$ that $0 \leq {p_{{{a^\ast}}-1}^{[2]}}(x) \leq 1$, see Lemma \[lem:orth.pol\] (iii), and thus $$\begin{aligned} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} &\leq \|{P}_x {p_{{{a^\ast}}-1}^{[2]}}({S_n}) \{{T_n^\ast y}- S f + Sf\}\|_\calH + \|{P}_x^\perp {p_{{{a^\ast}}-1}^{[2]}}({S_n}){T_n^\ast y}\|_\calH\\ &\leq C_\epsilon\gamma_n + \|{P}_x {p_{{{a^\ast}}-1}^{[2]}}({S_n}) {S}^{\mu+1} u\|_\calH + x^{-1/2} \|{P}_x^\perp {S_n}^{1/2} {p_{{{a^\ast}}-1}^{[2]}}({S_n}) {T_n^\ast y}\|_\calH\\ & \leq C_\epsilon\gamma_n + x^{\mu+1} R + \|{P}_x {p_{{{a^\ast}}-1}^{[2]}}({S}^{\mu+1}-{S_n}^{\mu+1})u \|_\calH + x^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}.\end{aligned}$$ In the second inequality \[eq:source.H\] with $\mu \geq 0$ was applied. By assumption $x_\ast = (C_x\gamma)^{1/(\mu+1)} \leq {x_{1,{{a^\ast}}-1}^{[1]}} < {x_{1,{{a^\ast}}-1}^{[2]}}$ due to the interlacing property of the roots of the polynomials ${p_{i}^{[r]}}$, $i=1,\dots,n$, $r \in {{\mathbb N}}_0$, see Lemma \[lem:orth.pol\] (i). Using Lemma \[lem:op.inequality\] we get $\|{S}^{\mu+1}-{S_n}^{\mu+1}\|_\calL \leq (\mu+1) \kappa^\mu C_\delta\gamma_n$ and setting $x=x_\ast$ we get $$\begin{aligned} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} &\leq C_\epsilon \gamma_n + x_\ast^{\mu+1} R + C_\delta \gamma_n(\mu+1) \kappa^\mu R + x_\ast^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2} \notag\\ & = \gamma_n \left\{ C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R \right\} + x_\ast^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}.\label{eq:gamma.bound}\end{aligned}$$ Due to (\[eq:alt.stop\]) and (\[eq:pol.opt\]) we have additionally $ {C}\gamma_n \leq \|{S_n}{f_{{{a^\ast}}-1}^{[1]}}- {T_n^\ast y}\|_\calH = \|{p_{{{a^\ast}}-1}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH \leq {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2}. $ Plugging this into (\[eq:gamma.bound\]) yields $${{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \leq {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} + x_\ast^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2},$$ or equivalently with $x_\ast = (C_x\gamma_n)^{1/(\mu+1)}$ $$\label{eq:twosidepol} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \leq \gamma^{-1/(2\mu+2)}_n \frac{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}} {C_x^{1/(2\mu+2)} \left[ 1- {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} \right] } ,$$ where by assumption ${C}> C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R$ and $x_\ast = (C_x\gamma)^{1/(\mu+1)}$. Combining (\[eq:est.difference\]), (\[eq:twosidepol\]) and $\|{p_{{{a^\ast}}}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH \leq {C}\gamma_n$ due to the stopping index (\[eq:alt.stop\]) yields $$\begin{aligned} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH &\leq \gamma^{-1/(\mu+1)}_n \frac{ \|{p_{{{a^\ast}}}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH } {C_x^{1/(\mu+1)} \left[ 1- {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} \right]^2 }\\ & \leq \gamma^{\mu/(\mu+1)}_n \frac{{C}} {C_x^{1/(\mu+1)} \left[ 1- {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} \right]^2}.\end{aligned}$$ For the second part of the proof we derive in the same way as (\[eq:est.difference\]) $$\|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH \leq \frac{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2}} {{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}} \|{p_{{{a^\ast}}}^{[1]}}({S_n}){T_n^\ast y}\|_\calH.$$ Using (\[eq:twosidepol\]) and $\|{p_{{{a^\ast}}}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH \leq {C}\gamma_n$ gives $$\|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH \leq \gamma^{(2\mu+1)/(2\mu+2)}_n \frac{{C}} {C_x^{1/(2\mu+2)} \left[ 1- {C}^{-1} \{C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R\} \right]},$$ finishing the proof. $\square$ \[lem:secondlemma\] For any $i =1,\dots,n$ and any $0 < x \leq {x_{1,i}^{[1]}}$ we have under the conditions of the theorem for $\mu \geq 1$ $$\begin{aligned} \|f - {f_{i}^{[1]}}\|_\calH &\leq R \left\{ x^\mu + C_\delta \mu \kappa^{\mu-1}\gamma_n \right\} + x^{-1} \left\{ \|{S_n}{f_{i}^{[1]}} - {T_n^\ast y}\|_\calH + (C_\epsilon + C_\delta \kappa^\mu R)\gamma_n \right\}\\ &+ (C_\epsilon + C_\delta \kappa^\mu R)\gamma_n | ({p_{i}^{[1]}})'(0)|, \\ \|{S_n}^{1/2}(f - {f_{i}^{[1]}})\|_\calH &\leq R \left\{ x^{\mu+1/2} + x^{1/2}C_\delta \mu \kappa^{\mu-1}\gamma_n \right\} \\ &+ x^{-1/2} \left\{ \|{S_n}{f_{i}^{[1]}} - {T_n^\ast y}\|_\calH + (C_\epsilon + C_\delta \kappa^\mu R)\gamma_n \right\}\\ &+ x^{1/2}(C_\epsilon + C_\delta \kappa^\mu R)\gamma_n | ({p_{i}^{[1]}})'(0)|.\end{aligned}$$ [*[Proof]{}:*]{} Denote $\bar{f}_i = {q_{i-1}^{[1]}}({S_n}) {S_n}f$ and consider $$\label{eq:eqsplit} \|f - {f_{i}^{[1]}}\|_\calH \leq \|{P}_x(f - \bar{f}_i)\|_\calH + \|{P}_x(\bar{f}_i - {f_{i}^{[1]}})\|_\calH + \|{P}_x^\perp ( f- {f_{i}^{[1]}})\|_\calH.$$ The first term of (\[eq:eqsplit\]) can be bound by an application of Lemma \[lem:op.inequality\] and \[eq:source.H\] with $\mu \geq 1$ $$\begin{aligned} \|{P}_x(f - \bar{f}_i)\|_\calH &= \|{P}_x\{I - {q_{i-1}^{[1]}}({S_n}){S_n}\}f \|_\calH = \|{P}_x{p_{i}^{[1]}}({S_n}) f \|_\calH = \|{P}_x{p_{i}^{[1]}}({S_n}) {S}^\mu u \|_\calH \\ &\leq \|{P}_x{p_{i}^{[1]}}({S_n}) {S_n}^\mu u \|_\calH + \|{P}_x{p_{i}^{[1]}}({S_n}) ({S}^\mu - {S_n}^\mu) u \|_\calH \\ & \leq R \left\{ x^\mu + C_\delta \mu \kappa^{\mu-1}\gamma_n \right\}.\end{aligned}$$ In the last inequality we used that on $0 \leq x \leq {x_{1,i}^{[1]}}$ we have $0 \leq {p_{i}^{[1]}}(x) \leq 1$. For the second term of (\[eq:eqsplit\]) we use Lemma \[lem:orth.pol\] (iii) ${q_{i}^{[1]}}(x) \leq |({p_{i}^{[1]}})'(0)|$ on $x \in [0,{x_{1,i}^{[1]}}]$. This yields $$\begin{aligned} \|{P}_x({f_{i}^{[1]}} - \bar{f}_i)\|_\calH & = \|{P}_x {q_{i}^{[1]}}({S_n}) ({S_n}f - {T_n^\ast y})\|_\calH\\ & \leq \|{P}_x {q_{i}^{[1]}}({S_n}) ({S}f - {T_n^\ast y})\|_\calH + \|{P}_x {q_{i}^{[1]}}({S_n})({S_n}- {S}) f\|_\calH\\ & \leq (C_\epsilon + C_\delta \kappa^\mu R)\gamma_n {\left|\left({p_{i}^{[1]}}\right)'(0)\right|}.\end{aligned}$$ Finally, we have $$\begin{aligned} \|{P}_x^\perp ( f- {f_{i}^{[1]}})\|_\calH &\leq x^{-1} \|{P}_x^\perp {S_n}( f- {f_{i}^{[1]}})\|_\calH\ \leq x^{-1} \left\{ \|{S_n}{f_{i}^{[1]}} - {T_n^\ast y}\|_\calH + \|{P}_x({T_n^\ast y}- {S_n}f)\|_\calH \right\} \\ &\leq x^{-1} \left\{ \|{S_n}{f_{i}^{[1]}} - {T_n^\ast y}\|_\calH + (C_\epsilon + C_\delta \kappa^\mu R) \gamma_n \right\}\end{aligned}$$ and thus the first inequality is proven. For the second inequality we use $$\|{S_n}^{1/2}(f - {f_{i}^{[1]}})\|_\calH \leq \|{P}_x {S_n}^{1/2}(f - \bar{f}_i)\|_\calH + \|{P}_x {S_n}^{1/2}(\bar{f}_i - {f_{i}^{[1]}})\|_\calH + \|{P}_x^\perp {S_n}^{1/2}( f- {f_{i}^{[1]}})\|_\calH.$$ In the same way as before we derive bounds for the three terms: $$\begin{aligned} \|{P}_x {S_n}^{1/2}(f - \bar{f}_i)\|_\calH & \leq x^{1/2} C_\delta \mu \kappa^{\mu-1} R\gamma_n + x^{\mu+1/2} R,\\ \|{P}_x {S_n}^{1/2}(\bar{f}_i - {f_{i}^{[1]}})\|_\calH &\leq x^{1/2} ( C_\epsilon + C_\delta R \kappa^\mu)\gamma_n {\left|\left({p_{i}^{[1]}}\right)'(0)\right|},\\ \|{P}_x^\perp {S_n}^{1/2}( f- {f_{i}^{[1]}})\|_\calH &\leq x^{-1/2} \{ \|{S_n}{f_{i}^{[1]}} - {T_n^\ast y}\|_\calH + (C_\epsilon + C_\delta \kappa^\mu R)\gamma_n\},\end{aligned}$$ completing the proof. $\square$ \[lem:derivative\] Assume that $C_x \in (0,1]$ is such that $x_\ast = (C_x\gamma)^{1/(\mu+1)} < {x_{1,{{a^\ast}}-1}^{[1]}}$ and ${C}> C_\epsilon + C_x R + C_\delta (\mu+1) \kappa^\mu R$. Under the conditions of the theorem it holds for $\mu \geq 0$ $$\begin{aligned} {\left|\left({p_{{{a^\ast}}}^{[1]}}\right)'(0)\right|} &\leq \gamma^{-1/(\mu+1)}_n \left[ C_x^{-1/(\mu+1)} \left\{ 1 - \frac{C_\epsilon + C_x R + C_\delta (\mu+1)\kappa^\mu R}{{C}} \right\}^{-2} \right.\\ &+ \left. \left\{ \frac{(2\mu+2)^{\mu+1} R} { {C}- C_\delta (\mu+1) \kappa^\mu R + C_\epsilon } \right\}^{1/(\mu+1)} \right]\end{aligned}$$ [*[Proof:]{}*]{} The proof is done in two steps by using the inequality ${\left|\left({p_{{{a^\ast}}}^{[1]}}\right)'(0)\right|} \leq {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} + \left| \left( {p_{{{a^\ast}}}^{[1]}} \right)'(0) - \left( {p_{{{a^\ast}}-1}^{[1]}} \right)'(0) \right|$. Consider first ${{a^\ast}}>1$. We will bound $\|{S_n}{f_{{{a^\ast}}-1}^{[1]}} - {T_n^\ast y}\|_\calH$ from above. Define $z ={x_{1,{{a^\ast}}-1}^{[1]}}$ and $\phi_{i}(x) = {p_{i}^{[1]}}(x) (z-x)^{-1/2} z^{1/2}$, $0 \leq x \leq z$. Due to Lemma \[lem:orth.pol\] (vi) it holds that $\sup_{0 \leq x \leq z} x^\nu \phi_{{{a^\ast}}-1}^2(x) \leq \nu^\nu |({p_{{{a^\ast}}-1}^{[1]}})'(0)|^{-\nu}$, $\nu \geq 0$. The proof of Lemma 3.7 in @Hanke shows that $${{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}}^{1/2} \leq \| {P}_{z} \phi_{{{a^\ast}}-1}({S_n}) {T_n^\ast y}\|_\calH.$$ This yields with \[eq:source.H\] $$\begin{aligned} &\|{S_n}{f_{{{a^\ast}}-1}^{[1]}} - {T_n^\ast y}\|_\calH = {{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}}^{1/2} \leq \| {P}_{z} \phi_{{{a^\ast}}-1}({S_n}) {T_n^\ast y}\|_\calH\\ & \leq \| {P}_{z} \phi_{{{a^\ast}}-1}({S_n}) {S}f\|_\calH + \| {P}_{z} \phi_{{{a^\ast}}-1}({S_n}) ({T_n^\ast y}-{S}f)\|_\calH\\ & \leq \| {P}_{z} \phi_{{{a^\ast}}-1}({S_n}) {S}f\|_\calH + C_\epsilon \gamma_n \left(\sup\limits_{0\leq x\leq z} \phi_{{{a^\ast}}-1}^2\right)^{1/2} \\ & \leq \|{P}_z \phi_{{{a^\ast}}-1}({S_n}){S_n}^{\mu+1} u \|_\calH + \|{P}_z \phi_{{{a^\ast}}-1}({S_n})({S_n}^{\mu+1}-{S}^{\mu+1}) u \|_\calH + C_\epsilon\gamma_n\\ & \leq R\left\{ \left(\sup\limits_{0\leq x\leq z} x^{2\mu+2} \phi_{{{a^\ast}}-1}^2\right)^{1/2} + C_\delta (\mu+1) \kappa^\mu \gamma_n \left(\sup\limits_{0\leq x\leq z} \phi_{{{a^\ast}}-1}^2\right)^{1/2} \right\} + C_\epsilon\gamma_n\\ & \leq {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-\mu-1} (2\mu+2)^{\mu+1} R + \{C_\delta (\mu+1) \kappa^\mu R + C_\epsilon\}\gamma_n.\end{aligned}$$ This gives together with ${C}\gamma_n \leq \| {S_n}{f_{{{a^\ast}}-1}^{[1]}} - {T_n^\ast y}\|_\calH$ $$\begin{aligned} {C}\gamma_n \leq {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-\mu-1} (2\mu+2)^{\mu+1} R + \left\{C_\delta (\mu+1) \kappa^\mu R + C_\epsilon\right\}\gamma_n.\end{aligned}$$ If ${C}> C_\delta (\mu+1) \kappa^\mu R + C_\epsilon$ we finally have $$\label{eq:diffm1bound} {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} \leq \gamma_n^{-1/(\mu+1)} \left\{ \frac{(2\mu+2)^{\mu+1} R} { {C}- C_\delta (\mu+1) \kappa^\mu R + C_\epsilon } \right\}^{1/(\mu+1)}.$$ If ${{a^\ast}}=1$ it holds ${p_{{{a^\ast}}-1}^{[1]}} = 1$ and thus ${\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} = 0$ and the inequality (\[eq:diffm1bound\]) is true as well. We will derive an upper bound on $\left| \left( {p_{{{a^\ast}}}^{[1]}} \right)'(0) - \left( {p_{{{a^\ast}}-1}^{[1]}} \right)'(0) \right|.$ Due to Corollary 2.6 of @Hanke we have $$\label{eq:deriv.diff} \left| \left( {p_{{{a^\ast}}-1}^{[1]}} \right)'(0) - \left( {p_{{{a^\ast}}}^{[1]}} \right)'(0) \right|\leq \frac{{{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}} } {{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}}.$$ We have $0 \leq x \leq {x_{1,{{a^\ast}}-1}^{[1]}} < {x_{1,{{a^\ast}}-1}^{[2]}}$ due to the interlacing property of the roots in Lemma \[lem:orth.pol\] (i) and thus $0 \leq {p_{{{a^\ast}}-1}^{[2]}}(x) \leq 1$ for $0 \leq x \leq {x_{1,{{a^\ast}}-1}^{[2]}}$. With that we get with \[eq:source.H\] $$\begin{aligned} &\|{p_{{{a^\ast}}-1}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH \leq {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \\ &\leq \|{P}_x {p_{{{a^\ast}}-1}^{[2]}}({S_n}) {T_n^\ast y}\|_\calH + x^{-1/2}\|{P}_x^\perp {S_n}^{1/2} {p_{{{a^\ast}}-1}^{[2]}}({S_n}) {T_n^\ast y}\|_\calH\\ &\leq \|{P}_x {p_{{{a^\ast}}-1}^{[2]}}({S_n})({T_n^\ast y}- {S}f)\|_\calH + \|{P}_x {p_{{{a^\ast}}-1}^{[2]}}({S_n}) {S}^{\mu+1} u\|_\calH + x^{-1/2}{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}\\ &\leq C_\epsilon\gamma_n + R\left\{ C_\delta(\mu+1)\kappa^\mu \gamma_n + x^{\mu+1} \right\} + x^{-1/2}{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}.\end{aligned}$$ For the choice $x_\ast = (C_x \gamma)^{1/(\mu+1)}$ we get $${{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}}^{1/2} \leq \gamma_n\left\{ C_\epsilon + C_\delta(\mu+1)\kappa^\mu R +C_x \right\} + x_\ast^{-1/2}{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}.$$ It holds ${{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}}^{1/2} = \| {S_n}{f_{{{a^\ast}}-1}^{[1]}} - {T_n^\ast y}\|_\calH \geq {C}\gamma_n$. This yields with ${C}> C_\epsilon + C_x R + C_\delta (\mu+1)\kappa^\mu R$ $${{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}} \leq \gamma_n^{-1/(\mu+1)} C_x^{-1/(\mu+1)} \left\{ 1 - \frac{C_\epsilon + C_x R + C_\delta (\mu+1)\kappa^\mu R}{{C}} \right\}^{-2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}.$$ Together with (\[eq:deriv.diff\]) we have $$\left| \left( {p_{{{a^\ast}}-1}^{[1]}} \right)'(0) - \left( {p_{{{a^\ast}}}^{[1]}} \right)'(0) \right|\leq \gamma^{-1/(\mu+1)}_n C_x^{-1/(\mu+1)} \left\{ 1 - \frac{C_\epsilon + C_x R + C_\delta (\mu+1)\kappa^\mu R}{{C}} \right\}^{-2}.$$ Combining this with (\[eq:diffm1bound\]) completes the proof. $\square$ ### Proof of Theorem \[th:kpls\] {#proof-of-theorem-thkpls} The proof is an application of Lemmas \[lem:error.difference\] - \[lem:derivative\] to (\[eq:main.inequality\]). First note that $r \geq 3/2$ implies $\mu \geq 1$ and thus this condition in Lemma \[lem:secondlemma\] holds. Let us choose $x_\ast = (C_x \gamma_n)^{1/(\mu+1)}$. Lemma \[lem:orth.pol\] (v) shows that ${\left|\left({p_{i}^{[r]}}\right)'(0)\right|} = \sum_{j=1}^i ({x_{j,i}^{[r]}})^{-1}$ for $i =1,\dots,n$, $r \in {{\mathbb N}}_0$. Thus it holds ${\left|\left({p_{i}^{[1]}}\right)'(0)\right|}^{-1} \leq {x_{1,i}^{[1]}}$. Equation (\[eq:diffm1bound\]) thus shows that $C_x$ can be chosen small enough such that $$x_\ast \leq {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1} \leq {x_{1,{{a^\ast}}-1}^{[1]}}$$ and $C_x <1$, which makes the first condition in Lemma \[lem:error.difference\] and \[lem:derivative\] hold true. The choice $C = C_\epsilon + (\mu+1)\kappa^\mu R (1 + C_\delta)$ gives the second condition. Now we need to check the remaining condition of Lemma \[lem:secondlemma\], namely that a $C_z$ can be chosen such that $(C_z \gamma_n)^{1/(\mu+1)} \leq {x_{1,{{a^\ast}}}^{[1]}}$ is true. Lemma \[lem:derivative\] yields a $C_z>0$ such that $C_z \gamma_n^{1/(\mu+1)} \leq {\left|\left({p_{{{a^\ast}}}^{[1]}}\right)'(0)\right|}^{-1} \leq {x_{1,{{a^\ast}}}^{[1]}}.$ Denote $z_\ast = (C_z \gamma_n)^{1/(\mu+1)}$ and Lemma \[lem:secondlemma\] can be applied. To ease notation we will denote everything in the derived bounds that does not depend on $\gamma_n$ as a constant $c_j$, $j \in {{\mathbb N}}$. Thus we get by combining Lemmas \[lem:secondlemma\] and \[lem:derivative\] that with probability at least $1-\nu$ $$\begin{aligned} \|f - {f_{{{a^\ast}}}^{[1]}}\|_\calH^2 &\leq c_1 \gamma_n^{\mu/(\mu+1)} + c_2 \gamma_n + c_3 \gamma_n^{1-1/(\mu+1)} + c_4 \gamma_n {\left|\left({p_{{{a^\ast}}}^{[1]}}\right)'(0)\right|}\\ & \leq c_1 \gamma_n^{\mu/(\mu+1)} + c_2 \gamma_n + c_3 \gamma_n^{\mu/(\mu+1)} + c_5 \gamma_n^{1-1/(\mu+1)} = O\{\gamma_n^{\mu/(\mu+1)}\}\end{aligned}$$ and $$\begin{aligned} &\|{S_n}^{1/2}(f - {f_{{{a^\ast}}}^{[1]}})\|_\calH^2\\ &\leq c_6 \gamma_n^{(\mu+1/2)/(\mu+1)} +c_7 \gamma_n^{1/(2\mu+2)}\gamma_n + c_8 \gamma_n^{-1/(2\mu+2)}\gamma_n + c_9\gamma_n^{1/(2\mu+1)}\gamma_n {\left|\left({p_{{{a^\ast}}}^{[1]}}\right)'(0)\right|}\\ & \leq c_6 \gamma_n^{(\mu+1/2)/(\mu+1)} +c_7 \gamma_n^{(2\mu+3)/(2\mu+2)} + c_8 \gamma_n^{(2\mu+1)/(2\mu+2)} + c_10\gamma_n^{1+1/(2\mu+2)-1/(\mu+1)}\\ & = O\{\gamma_n^{(2\mu+1)/(2\mu+2)}\}.\end{aligned}$$ Finally Lemma \[lem:error.difference\] gives $$\|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH^2 = O\{\gamma_n^{\mu/(\mu+1)}\}, ~~~~ \|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH =O\{\gamma_n^{(2\mu+1)/(2\mu+2)}\}.$$ Combining the above with (\[eq:main.inequality\]) yields $$\begin{aligned} \|f - f_{{a^\ast}}\|_\calH^2 &= O\{\gamma_n^{\mu/(\mu+1)}\},\\ \|f^\ast - f_{{a^\ast}}\|_2^2 &= O\{\gamma^{1/2} \gamma_n^{\mu/(\mu+1)}\} + O\{\gamma_n^{(2\mu+1)/(2\mu+2)}\} = O\{\gamma_n^{(2\mu+1)/(2\mu+2)}\},\end{aligned}$$ completing the proof with $\mu = r-1/2$. $\square$ Proof of Theorem \[th:kpls2\] ----------------------------- The overall design of this proof is similar to the one of Theorem \[th:kpls\] and makes heavy use of results obtained in @Blan10b. ### Preparation for the proof The stopping index \[eq:stopping2\] can be reformulated with $\mu = r-1/2$ as $$\label{eq:alt.stop2} {{a^\ast}}= \min\{1 \leq a \leq n: \|{S_n}{f_{a}^{[1]}}- {T_n^\ast y}\|_\calH \leq C \zeta_n\},$$ with $\zeta_n = \max\{\sqrt{\lambda_n {d_\lambda}}\gamma_n,\lambda_n^{\mu+1}\}$. We will derive the result in a similar way to Theorem \[th:kpls\]. First it holds due to (\[eq:main.inequality\]) $$\begin{aligned} &\|f_{{a^\ast}}- f^\ast\|_2 =\|{{S}^{1/2}}(f_{{a^\ast}}- f)\|_\calH \leq \|{{S}^{1/2}}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH + \|{{S}^{1/2}}({f_{{{a^\ast}}}^{[1]}} - f)\|_\calH \notag \\ & \leq C_\psi \lambda^{1/2} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH + C_\psi\|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH + \|{{S}^{1/2}}({f_{{{a^\ast}}}^{[1]}}-f)\|_\calH. \label{eq:main.inequality2}\end{aligned}$$ Now we prove the analogue versions of Lemma \[lem:error.difference\] – \[lem:derivative\]: \[lem:error.difference2\] Let $x = C_x \lambda_n$, $C_x>0$ such that $0< x < x_{1,{{a^\ast}}-1}^{[2]}$. Choose $C>\tilde{c}_2$, with $\tilde{c}_1 = R \max\{1,C_\psi^2,\mu \kappa^{\mu-1} C_\delta\}$ and $\tilde{c}_2 = 2\max\{C_\psi C_\epsilon \sqrt{C_x+1}, \tilde{c}_1 C_x (C_x^\mu+1)\}$. Then it holds $$\begin{aligned} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH &\leq \frac{C^3}{C_x(C-\tilde{c}_2)^2}\lambda^{-1}_n \zeta_n,\\ \|S_n^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH &\leq \frac{C^2}{C_x^{1/2}(C-\tilde{c}_2)} \lambda^{-1/2}_n \zeta_n.\end{aligned}$$ Proof: According to the proof of Lemma \[lem:error.difference2\] we can focus on the case $0< {{a^\ast}}< n$. Furthermore we have due to (\[eq:est.difference\]) $$\label{eq:est.difference2} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH \leq \frac{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}}{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}}\|{p_{{{a^\ast}}}^{[1]}}({S_n}) {T_n^\ast y}\|_\calH.$$ Using Lemma A.3 in @Blan10b (and the first line of its proof) we have for $0 < x < x_{1,{{a^\ast}}-1}^{[2]}$ with $\tilde{c}_1 = R \max\{1,C_\psi^2,\mu \kappa^{\mu-1} C_\delta\}$ $$\begin{aligned} \label{eq:norm.difference} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \leq C_\psi C_\epsilon \sqrt{x + \lambda_n}\sqrt{d_{\lambda_n}}\gamma_n + \tilde{c}_1 x \{x^\mu + Z_\mu(\lambda_n)\} + x^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}.\end{aligned}$$ Here we define $Z_\mu(\lambda) = \lambda^\mu \mathbb{I}(\mu \leq 1) + \gamma_n \mathbb{I}(\mu>1)$. Note that under the assumptions of the theorem it holds $Z_\mu(\lambda_n) \leq \lambda_n^\mu$. Choosing $x = C_x \lambda_n$ yields in (\[eq:norm.difference\]) $$\begin{aligned} &{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2}\\ &\leq C_\psi C_\epsilon \sqrt{C_x + 1} \sqrt{\lambda_n d_{\lambda_n}}\gamma_n + \tilde{c}_1 C_x(C_x^\mu+1)\lambda_n^{\mu+1} + C_x^{-1/2} \lambda_n^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}\\ &\leq \tilde{c}_2\max\{\sqrt{\lambda_n d_{\lambda_n}} \gamma_n,\lambda_n^{\mu+1}\} + C_x^{-1/2} \lambda_n^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}\\ & = \tilde{c}_2 \zeta_n + C_x^{-1/2} \lambda_n^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2} , \end{aligned}$$ with $\tilde{c}_2 = 2\max\{C_\psi C_\epsilon \sqrt{C_x+1}, \tilde{c}_1 C_x (C_x^\mu+1)\}$. Due to the stopping condition (\[eq:alt.stop2\]) we know that $${{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \geq {{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}}^{1/2} = \|{S_n}{f_{a}^{[1]}}- {T_n^\ast y}\|_\calH \geq C \zeta_n.$$ This gives $$\label{eq:norm.upperbound} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2} \leq \frac{C}{\sqrt{C_x}(C-\tilde{c}_2)} \lambda_n^{-1/2} {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}.$$ Plugging this into (\[eq:est.difference2\]) yields together with the definition of the stopping index ${{a^\ast}}$ $$\|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH \leq \frac{C^3}{C_x(C-\tilde{c}_2)^2}\lambda^{-1}_n \zeta_n.$$ In a similar way we derive for the second case $$\|S_n^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH = \frac{{{\left[{p_{{{a^\ast}}}^{[1]}},{p_{{{a^\ast}}}^{[1]}}\right]_{0}}}}{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}} \leq \frac{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}^{1/2}}{{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}^{1/2}}{{\left[{p_{{{a^\ast}}}^{[1]}},{p_{{{a^\ast}}}^{[1]}}\right]_{0}}}.$$ An application of (\[eq:norm.upperbound\]) yields $$\|S_n^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH \leq \frac{C^2}{\sqrt{C_x}(C-\tilde{c}_2)} \lambda_n^{-1/2}\zeta_n.$$ $\square$ \[lem:secondlemma2\] Denote $\tilde{c}_1 = R \max\{1,C_\psi^2,\mu \kappa^{\mu-1} C_\delta\}$. For any $i=1,\dots,n$ and $0 < x < x_{1,i}^{[1]}$ we have under the conditions of the theorem $$\begin{aligned} \|S^{1/2}(f_i^{[1]}-f)\|_\calH &\leq C_\psi\left[C_\delta+\sqrt{2}C_\epsilon + \lambda_n\left\{C_\delta {\left|\left({p_{i}^{[1]}}\right)'(0)\right|} + \sqrt{2}C_\epsilon x^{-1}\right\}\right]\sqrt{d_{\lambda_n}}\gamma_n\\ &+ \tilde{c}_1 (\sqrt{x}+\sqrt{\lambda_n})(x^\mu+\lambda_n^\mu) + (1+\sqrt{x^{-1}\lambda_n})x^{-1/2}\|{S_n}{f_{i}^{[1]}}-{T_n^\ast y}\|_\calH. \end{aligned}$$ Proof: Follow the proof of Lemma A.2 in @Blan10b. Note that $Z_\mu(\lambda_n) \leq \lambda_n^\mu$. $\square$ \[lem:derivative2\] Let $C>\max\{C_\psi C_\epsilon,\tilde{c}_2\}$, where $\tilde{c}_2$ is given in Lemma \[lem:error.difference2\]. Choose $x = C_x \lambda_n$ such that $0 < x \leq x_{1,{{a^\ast}}-1}^{[1]}$. Then there exists a constant $c^\ast>0$ such that $$\begin{aligned} {\left|\left({p_{a^\ast}^{[1]}}\right)'(0)\right|} \leq c^\ast \lambda^{-1}_n. \end{aligned}$$ Proof: In analogue to Lemma \[lem:derivative\] we will first derive an upper bound on ${\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}$. Lemma A.1 in @Blan10b yields $$\begin{aligned} \|{S_n}{f_{{{a^\ast}}-1}^{[1]}} - {T_n^\ast y}\|_\calH & \leq R (2\mu+2)^{\mu+1}\max\{1,C_\psi^{2\mu}\} {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-\mu-1}\\ &+ 2R \mu\kappa^{\mu-1} \max\{1,C_\delta,C_\psi^{2\mu},C_\psi^{2\mu} C_\delta\} {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1} Z_\mu(\lambda_n)\\ &+C_\epsilon C_\psi \left\{ {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1/2} + \sqrt{\lambda_n} \right\} \sqrt{d_{\lambda_n}}\gamma_n. \end{aligned}$$ Denote $\tilde{c}_3 = R (2\mu+2)^{\mu+1}\max\{1,C_\psi^{2\mu}\}$ and $\tilde{c}_4 = 2R \mu\kappa^{\mu-1} \max\{1,C_\delta,C_\psi^{2\mu},C_\psi^{2\mu} C_\delta\}$. The definition of ${{a^\ast}}$ gives $C \zeta_n \leq \|{S_n}{f_{{{a^\ast}}-1}^{[1]}} - {T_n^\ast y}\|_\calH$. Combining both inequalities, setting $x = C_x \lambda_n$ and keeping $\sqrt{\lambda_n d_{\lambda_n}} \gamma_n \leq \zeta_n$ in mind gives $$\begin{aligned} (C-C_\psi C_\epsilon C_\lambda) \zeta_n &\leq \tilde{c}_3 {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-\mu-1} + \tilde{c}_4 {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1} \lambda_n^\mu\\ &+ C_\epsilon C_\psi {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1/2} \sqrt{d_{\lambda_n}}\gamma_n\\ & \leq 3 \max\left\{ \tilde{c}_3 {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-\mu-1}, \tilde{c}_4 {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1}\lambda_n^\mu,\right.\\ &\left. C_\epsilon C_\psi {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|}^{-1/2} \sqrt{d_{\lambda_n}} \gamma_n \right\}.\end{aligned}$$ Now we assume that the maximum on the right hand side is attained in each of the three possible cases $$\begin{aligned} {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} & \leq \{3 (C- C_\epsilon C_\psi)^{-1} \tilde{c}_3 \}^{1/(\mu+1)} \zeta_n^{-1/(\mu+1)},\\ {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} & \leq 3 (C-C_\epsilon C_\psi)^{-1}\tilde{c}_4 \zeta_n^{-1} \lambda_n^\mu,\\ {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} & \leq 9 (C-C_\epsilon C_\psi)^{-2} C_\epsilon^2 C_\psi^2 \zeta_n^{-2} d_{\lambda_n} \gamma_n^2.\end{aligned}$$ Take $\tilde{c}_5 = \max[\{3 (C- C_\epsilon C_\psi)^{-1} \tilde{c}_3 \}^{1/(\mu+1)}, 3 (C-C_\epsilon C_\psi)^{-1}\tilde{c}_4, 9 (C-C_\epsilon C_\psi)^{-2} C_\epsilon^2 C_\psi^2]$. It is easy to see that $\zeta_n^{-1/(\mu+1)}, \zeta_n^{-1}\lambda_n^\mu$ and $\zeta_n^{-2} d_{\lambda _n}\gamma_n^2$ are all bound from above by $\lambda_n^{-1}$. Hence we get $$\label{eq:first.derivative.bound} {\left|\left({p_{{{a^\ast}}-1}^{[1]}}\right)'(0)\right|} \leq \tilde{c}_5 \lambda^{-1}_n.$$ For the final step in the proof we have due to (\[eq:deriv.diff\]) $$\left| \left( {p_{{{a^\ast}}-1}^{[1]}} \right)'(0) - \left( {p_{{{a^\ast}}}^{[1]}} \right)'(0) \right|\leq \frac{{{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}} } {{{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{1}}}}.$$ It holds ${{\left[{p_{{{a^\ast}}-1}^{[1]}},{p_{{{a^\ast}}-1}^{[1]}}\right]_{0}}} \leq {{\left[{p_{{{a^\ast}}-1}^{[2]}},{p_{{{a^\ast}}-1}^{[2]}}\right]_{0}}}$ and hence (\[eq:norm.upperbound\]) yields $$\label{eq:first.derivative.bound2} \left| \left( {p_{{{a^\ast}}-1}^{[1]}} \right)'(0) - \left( {p_{{{a^\ast}}}^{[1]}} \right)'(0) \right|\leq \frac{C^2}{C_x(C-\tilde{c}_2)^2} \lambda^{-1}_n.$$ The proof is complete by combining (\[eq:first.derivative.bound\]) and (\[eq:first.derivative.bound2\]). $\square$ ### Proof of Theorem \[th:kpls2\] We first restrict ourselves to the set where all concentration inequalities stated in the theorem hold simultaneously with probability at least $1-\nu$, $\nu \in (0,1]$. We only proof the convergence rates in the $\calL^2$-norm, the corresponding rates in the $\calH$-norm are done in the same way. The theorem is proven by an application of Lemmas \[lem:error.difference2\]–\[lem:derivative2\]. To that end we need to check the conditions of those. Equation (\[eq:first.derivative.bound\]) and the proof of Theorem \[th:kpls\] show that we can take $C_x = \min\{1/2,\tilde{c}_5\}$ to fulfill $0 < x \leq x_{1,{{a^\ast}}-1}^{[1]}$. Furthermore we can take $C=4 R \max\{1, C_\psi^2(\nu),(r-1/2)\kappa^{r-3/2} C_\delta(\nu),2^{-1/2}R^{-1} C_\psi(\nu) C_\epsilon(\nu)\}$ and the conditions of Lemma \[lem:error.difference2\] and \[lem:derivative2\] hold. Note that $x_{1,{{a^\ast}}-1}^{[1]} \leq x_{1,{{a^\ast}}-1}^{[2]}$ due to the interlacing property of the roots, see Lemma \[lem:orth.pol\] (i). For Lemma \[lem:secondlemma2\] we need to find a $0 < z < x_{1,{{a^\ast}}}^{[1]}$. By Lemma \[lem:derivative2\] there exists a constant $c^\ast>0$ such that $$(c_\ast)^{-1} \lambda_n \leq {\left|\left({p_{{{a^\ast}}}^{[1]}}\right)'(0)\right|}^{-1} \leq x_{1,{{a^\ast}}}^{[1]},$$ hence we choose $C_z = \min\{1/2,1/c^\ast\}$. Now, applying Lemmas \[lem:error.difference2\]–\[lem:derivative2\] to (\[eq:main.inequality2\]) gives the result (we again denote any constant that does not depend on $n$ with $C_i$, $i \in {{\mathbb N}}$) $$\begin{aligned} &\|f_{{a^\ast}}- f^\ast\|_2 \leq C_\psi \lambda_n^{1/2} \|f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}}\|_\calH + C_\psi\|{S_n}^{1/2}(f_{{a^\ast}}- {f_{{{a^\ast}}}^{[1]}})\|_\calH + \|{{S}^{1/2}}({f_{{{a^\ast}}}^{[1]}}-f)\|_\calH \\ \leq & C_1 \lambda_n^{-1/2} \zeta_n + C_2 \lambda_n {\left|\left({p_{i}^{[1]}}\right)'(0)\right|} \sqrt{d_{\lambda_n}}\gamma_n + C_3 \lambda_n^{\mu+1/2} + C_4 \lambda_n^{-1/2} \|{S_n}{f_{{{a^\ast}}}^{[1]}}-{T_n^\ast y}\|_\calH.\\ \leq& C_5 \lambda_n^{-1/2} \zeta_n + C_6 \sqrt{d_{\lambda_n}}\gamma_n + C_4 \lambda_n^{\mu+1/2} \leq \max\{C_4,C_5,C_6\} \lambda_n^{-1/2}\zeta_n.\end{aligned}$$ The error bound in the $\calH$-norm is proven in an analogue fashion. $\square$ Proof of Corollary \[cor:pol.ed\] --------------------------------- Take $\lambda_n = \gamma_n^{2/(2r+s)}$. It is immediate that $\lambda_n^{r-1/2} = \gamma_n^{(2r-1)/(2r+s)} \geq \gamma_n$ for $n$ sufficiently large, hence the inequality (\[eq:lambda.inequ\]) holds as soon as $\gamma_n \leq 1$. Then we have by Theorem \[th:kpls2\] that $$\|f_{{\widehat{\alpha}}_{a^\ast}} - f^\ast\|_2 = O\left\{\lambda_n^{-1/2}\zeta_n(\lambda_n)\right\} = O\left\{\gamma_n^{2r/(2r+s)}\right\}.$$ $\square$ Proof of Corollary \[cor:log.ed\] --------------------------------- Set $\lambda_n = \gamma_n^{1/r} \log\{1/(2r) \gamma_n^{-2}\}$. It is immediate that $\lambda_n \rightarrow 0$ as $\gamma_n$ converges to zero. For $r=1/2$ condition (\[eq:lambda.inequ\]) holds trivially. Let $r>1/2$, then we have $$\lambda_n^{r-1/2} = \gamma_n^{(r-1)/r} \log\{(r-1/2)/(2r) \gamma_n^{-2}\} \geq \gamma_n,$$ This is equivalent to $2r-1 \geq 2 \exp(\gamma_n) \gamma_n^2$, which holds for $n$ sufficiently large and $r > 1/2$. For the convergence rate we first show that $d_{\lambda_n} \gamma_n^2 \leq \lambda_n^{2r}$. We have $$d_{\lambda_n} \gamma_n^2 = \log\left\{ 1+ \frac{a}{\gamma_n^{1/r} \log^{1/r}(1/2 \gamma_n^{-1})} \right\} \gamma_n^2 \leq \log\left( \gamma_n^{-2} \right)\gamma_n^2.$$ Equivalently we need $a^r \leq \gamma_n (\gamma_n^{-2}-1)^r \log(1/2\gamma_n^{-2})$. As $\gamma_n$ converges to zero $\gamma_n(\gamma_n^{-2} -1)^{r}$ goes to infinity for any $r> 1/2$. Hence for suitably large $n$ it holds $\lambda_n^r \geq \sqrt{d_{\lambda_n}} \gamma_n$. Then the convergence rate is $\lambda_n^{-1/2} \zeta_n(\lambda_n) = \lambda_n^r = \gamma_n \log(1/2 \gamma^{-2})$. Because the convergence rate does not depend on $r \geq 1/2$ we can set $r = 1/2$. $\square$ Proof of Theorem \[th:conc.equality\] ------------------------------------- ### Preparation for the proof We denote with $\mathrm{tr} (A)$ the trace of a trace class operator $A: \calH \rightarrow \calH$ and the tensor product $(f_1 \otimes f_2) h = \langle f_1, h \rangle_\calH f_2$ for functions $f_1,f_2,h \in \calH$. We use the notation $k_t = k(\cdot, X_t)$. Note that it holds $\|A\|_{{\mathrm{HS}}}^2 = \mathrm{tr}(A^\ast A)$ for a Hilbert-Schmidt operator $A$. \[lem:sup1\] Under the assumptions \[con:k1\] and \[con:k2\] the following hold 1. ${\mathrm{tr}}\{(k_t \otimes k_t) (k_s \otimes k_s)\} = k^2(X_t,X_s)$, 2. $\|{S}\|_{{\mathrm{HS}}}^2 = \int_{{{\mathbb R}}^d}\int_{{{\mathbb R}}^d} k^2(x,y) {\mathrm{d}}{\mathrm{P}}^{X_0}(x) {\mathrm{d}}{\mathrm{P}}^{X_0}(y)$, 3. ${\mathrm{E}}[ {\mathrm{tr}}\{(k_0 \otimes k_0) {S}\}] = \|{S}\|_{{\mathrm{HS}}}^2$. 4. Let $X'$ and $X''$ be independent and identically distributed and denote $k' = k(\cdot,X')$, $k''=k(\cdot,X'')$. It holds for $\nu = 1,2$ and $\lambda>0$ $$\begin{aligned} \mathrm{E}\left[ \mathrm{tr}^\nu \left\{(S+\lambda)^{-1} k' \otimes k' k'' \otimes k''\right\} \right] = \mathrm{tr}^\nu\{(S+\lambda)^{-1} S^2\}.\end{aligned}$$ [*Proof:*]{} (i) Let $\{v_i\}_{i \in {{\mathbb N}}}$ denote an orthonormal base of $\calH$. Then it holds due to the reproducing property (\[eq:rep.property\]) $$\begin{aligned} \mathrm{tr}\left\{(k_t \otimes k_t)(k_s \otimes k_s)\right\} = \sum_{i=1}^\infty \langle v_i,k_t\rangle_\calH \langle v_i,k_s\rangle_\calH k(X_t,X_s) = \left\langle \sum_{i=1}^\infty \langle v_i, k_s \rangle_\calH v_i, k_t \right\rangle_\calH k(X_t,X_s).\end{aligned}$$ \(ii) $$\begin{aligned} \|{S}\|^2_{{\mathrm{HS}}} &= \sum_{i=1}^\infty \langle {S}v_i, {S}v_i\rangle_\calH = \sum_{i=1}^\infty \int\limits_{{{\mathbb R}}^d} \langle {S}v_i, k(\cdot,x) \rangle_\calH \langle v_i, k(\cdot,x)\rangle_\calH {\mathrm{d}}{\mathrm{P}}^{X}(x)\\ &= \int\limits_{{{\mathbb R}}^d} \int\limits_{{{\mathbb R}}^d} \left\langle \sum_{i=1}^\infty \langle v_i, k(\cdot,x) \rangle_\calH v_i, k(\cdot,y)\right\rangle_\calH k(x,y) {\mathrm{d}}{\mathrm{P}}^{X}(x) {\mathrm{d}}{\mathrm{P}}^{X}(y).\end{aligned}$$ The assertion follows because ${\mathrm{P}}^{X} = {\mathrm{P}}^{X_0}$. \(iii) $$\begin{aligned} {\mathrm{E}}[{\mathrm{tr}}\{(k_0 \otimes k_0) {S}\}] &= {\mathrm{E}}(\langle S k_0, k_0\rangle_\calH) = {\mathrm{E}}\left(\int\limits_{{{\mathbb R}}^d} \langle k_0, k(\cdot,x)\rangle^2_\calH {\mathrm{d}}{\mathrm{P}}^{X}(x)\right) \\ &= \int\limits_{{{\mathbb R}}^{d}}\int\limits_{{{\mathbb R}}^{d}} k^2(x,y) {\mathrm{d}}{\mathrm{P}}^{X}(x) {\mathrm{d}}{\mathrm{P}}^{X_0}(y) = \|S\|_{{\mathrm{HS}}}^2.\end{aligned}$$ \(iv) Because $S$ is a compact operator the spectral decomposition $S = \sum_{i=1}^\infty \mu_i \psi_i \otimes \psi_i$ holds (recall that $\{\mu_i$, $\psi_i\}_{i=1}^\infty$ is the eigensystem of $S$). Let $k(\cdot,x) = \sum_{i=1}^\infty \alpha_i(x) \psi_i$, $x \in {{\mathbb R}}^d$. For $\nu=1$ we have $$\begin{aligned} \mathrm{E}\left[\mathrm{tr}\{(S+\lambda)^{-1} k' \otimes k' k''\otimes k''\}\right]& = {\mathrm{E}}\left\{k(X',X'') \sum_{i=1}^\infty \langle \psi_i, k''\rangle_\calH \langle(S+\lambda)^{-1} k', \psi_i \rangle_\calH \right\}\\ & =\sum\limits_{i=1}^\infty \frac{1}{\mu_i+\lambda} {\mathrm{E}}\left\{k(X',X'') \langle \psi_i, k''\rangle_\calH \langle \psi_i,k' \rangle_\calH\right\}\\ & = \sum\limits_{i=1}^\infty \frac{1}{\mu_i+\lambda} \sum\limits_{j=1}^\infty {\mathrm{E}}\left\{\alpha_j(X')\alpha_i(X')\alpha_j(X'')\alpha_i(X'')\right\}\end{aligned}$$ On the other hand $$\begin{aligned} \mathrm{tr}\left\{ (S+\lambda)^{-1} S^2 \right\} & = \mathrm{tr}\left\{ (S+\lambda)^{-1} {\mathrm{E}}(k' \otimes k' k'' \otimes k'') \right\}\\ & = \sum\limits_{i=1}^\infty \langle (S+\lambda)^{-1} {\mathrm{E}}\{k(X',X'') \langle \psi_i,k''\rangle_\calH k'\} ,\psi_i \rangle_\calH\\ & = \sum\limits_{i=1}^\infty \frac{1}{\mu_i + \lambda } \langle {\mathrm{E}}\{k(X',X'') \langle \psi_i,k''\rangle_\calH k'\} ,\psi_i \rangle_\calH\\ & = \sum\limits_{i=1}^\infty \frac{1}{\mu_i + \lambda }\sum\limits_{j=1}^\infty {\mathrm{E}}\{\alpha_j(X')\alpha_i(X')\alpha_j(X'') \alpha_i(X'')\}\end{aligned}$$ and we are done. The proof for $\nu=2$ is along the same lines. $\square$ \[lem:acf.inequality\] Assume that condition \[D2\] holds. Then we have $$\begin{aligned} \label{eq:sum} n^{-2} \sum\limits_{h=1}^{n-1}(n-h) |\rho_h| \leq C(q) \left\{ \begin{array}{clc} n^{-1}&,& q>1\\ n^{-1} \log(n)&,& q=1\\ n^{-q} &,& q \in (0,1). \end{array} \right.,\end{aligned}$$ with $C(q) = \zeta(q) \mathbb{I}(q>1) + \{5-\log(4)\}\mathbb{I}(q=1) + \{2(1-q)^{-1}-(2-q)^{-1}+(2-q)^{-1} 2^{2-q}\}\mathbb{I}\{q\in(0,1)\}$. Here $\zeta$ denotes the Riemann zeta function. [*Proof*]{}: Recall that by condition \[D2\] we have $|\rho_h| \leq (h+1)^{-q}$, $h=0,\dots,n-1$ for some $q>0$. First assume $q \in (0,1]$. The integral test for series convergence gives lower and upper bounds for the hyperharmonic series as $$(1-q)^{-1}\{ (n+1)^{1-q}-2^{1-q} \} \leq \sum\limits_{h=2}^{n} h^{-q} \leq 2^{-q} + (1-q)^{-1} \{ n^{1-q}-2^{1-q} \}.$$ This yields $$\begin{aligned} &n^{-2} \sum\limits_{h=1}^{n-1} (n-h) (h+1)^{-q} = n^{-2} \sum\limits_{h=2}^{n} (n+1-h)h^{-q} = n^{-2} \left\{ (n+1)\sum\limits_{h=2}^n h^{-q} - \sum\limits_{h=2}^n h^{-(q-1)} \right\} \notag \\ &\leq n^{-2} \left[ (n+1) \left\{ 2^{-q} + (1-q)^{-1}(n^{1-q}-2^{1-q} \right\} - (2-q)^{-1} \left\{ (n+1)^{2-q} - 2^{2-q} \right\} \right].\label{eq:sum.inequality}\end{aligned}$$ Now let $q\in (0,1)$, then it holds from (\[eq:sum.inequality\]) and the fact that $n^{-2} \leq n^{-1} \leq n^{-q}$ $$\begin{aligned} &n^{-2} \sum\limits_{h=1}^{n-1} (n-h) (h+1)^{-q}\\ \leq & \frac{n+1}{n^2}\left\{ \frac{2^{-q}(1-q)-2^{1-q}}{1-q} \right\} + \frac{n+1}{n^{1+q}}(1-q)^{-1} - \frac{(n+1)^{2-q}}{n^2}(2-q)^{-1} +\frac{1}{n^2}\frac{2^{2-q}}{2-q}\\ \leq & n^{-q} [ \{ 2(1-q)^{-1} - (2-q)^{-1} \} + (2-q)^{-1}2^{2-q} ],\end{aligned}$$ due to $2^{-q} (1-q) - 2^{1-q}<0$. For $q=1$ we evaluate the limit $$\begin{aligned} &\lim\limits_{q \rightarrow 1\pm} n^{-2} \left[ (n+1) \left\{ 2^{-q} + (1-q)^{-1}(n^{1-q}-2^{1-q} \right\} - (2-q)^{-1} \left\{ (n+1)^{2-q} - 2^{2-q} \right\} \right]\\ & = (2 n^2)^{-1} [ 3-\log(4) - n \{1+\log(4)\} ] + n^{-2} (n+1) \log(n)\\ & \leq \frac{\log(n)}{n} \left[ 5 - \log(4) \right].\end{aligned}$$ The case $q>1$ is clear because the zeta-function $\zeta(q)$ is defined as the hyperharmonic series with coefficient $q$. $\square$ Denote with $g_{h}$ the common density of $(X_h,X_0)^\T$ and $g_0$ the density of $X_0$. The next lemma and the subsequent corollary will be used to show that the quantities appearing in the sums of Theorem \[th:conc.equality\] (i) can be linked to the autocorrelation function $\rho$: \[lem:normal.quantity\] Under the assumptions \[con:k1\], \[con:k2\] and \[D1\] it holds for $h>0$ with $\rho_h = \tau_0^{-1} \tau_h$ $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} k^2(x,y) \{g_h(x,y)-g_0(x)g_0(y)\}{\mathrm{d}}(x,y) &\leq \frac{\kappa^2} {\{(4 \pi \tau_0)^{d} \det({\Sigma})\}^{1/2}} \theta^{1/2}(\rho_h),\\ \int\limits_{{{\mathbb R}}^{2d}} k(x,y)f(x)f(y) \{g_h(x,y)-g_0(x)g_0(y)\} {\mathrm{d}}(x,y) &\leq \frac{\kappa M} {\{(4 \pi \tau_0)^{d} \det({\Sigma})\}^{1/2}} \theta^{1/2}(\rho_h),\end{aligned}$$ with $\theta(\rho) = 1 + (1- \rho^2)^{-d/2} - 2^{d+1} (4 - \rho^2)^{-d/2}$, $\rho \in [0,1)$. [*Proof*]{}: We will only proof the first inequality, the second one follows in the same way. By Jensen’s inequality and \[con:k2\] we know $$\begin{aligned} &\int\limits_{{{\mathbb R}}^{2d}} k^2(x,y) \{g_h(x,y)-g_0(x)g_0(y)\} {\mathrm{d}}(x,y) \\ &\leq \kappa^2 \left[ \int\limits_{{{\mathbb R}}^{2d}} \left\{ g_h^2(x,y) - 2 g_h(x,y) g_0(x) g_0(y) + g_0^2(x) g_0^2(y) \right\} {\mathrm{d}}(x,y) \right]^{1/2}.\end{aligned}$$ The first and third integral term can readily be calculated as $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} g_h^2(x,y){\mathrm{d}}(x,y) & = [(4 \pi)^d (\tau^2_0-\tau^2_h)^{d/2} \det(\Sigma)]^{-1}\\ \left\{\int\limits_{{{\mathbb R}}^{d}} g_0^2(x){\mathrm{d}}x\right\}^2 & = \{(4 \pi)^d \tau^{d}_0 \det(\Sigma)\}^{-1}. \end{aligned}$$ For the first equality we use $\det(A \otimes \Sigma) = \det(A)^d \det(\Sigma)^2$ for $A \in {{\mathbb R}}^{2 \times 2}$ and thus $$\label{eq:normalsquared} \int\limits_{{{\mathbb R}}^{2d}} g_h(x,y) g_0(x) g_0(y){\mathrm{d}}(x,y) = \frac{\int_{{{\mathbb R}}^{2d}} \exp\left( -1/2 z^\T G^{-1} z \right) {\mathrm{d}}z} {(2\pi)^{2d} \det(\Sigma)^2\tau^d_0 (\tau^2_0-\tau^2_h)^{d/2}},$$ with $$G^{-1} = \left\{\left( \begin{matrix} \tau_0 & \tau_h\\ \tau_h & \tau_0 \end{matrix} \right)^{-1} + \left( \begin{matrix} \tau^{-1}_0 & 0\\ 0 & \tau^{-1}_0 \end{matrix} \right) \right\}\otimes \Sigma^{-1}.$$ It holds $\det(G) = (4 \tau^2_0 - \tau^2_h)^{-d} (\tau^4_0 - \tau^2_0 \tau^2_h)^d \det(\Sigma)^2$. Thus we get with $(\ref{eq:normalsquared})$ $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} g_h(x,y) g_0(x) g_0(y){\mathrm{d}}(x,y) &= \frac{(2\pi)^d \tau^d_0 (\tau^2_0 - \tau^2_h)^{d/2} \det(\Sigma) } {(2\pi)^{2d} \det(\Sigma)^2 (4 \tau^2_0 - \tau^2_h)^{d/2}\tau^d_0 (\tau^2_0-\tau^2_h)^{d/2}}\\ & = \left\{ (2 \pi)^d ( 4 \tau^2_0 - \tau^2_h )^{d/2} \det(\Sigma) \right\}^{-1}, \end{aligned}$$ completing the proof by multiplying all terms with $\tau^{-d}_0 \tau^d_0$. $\square$ \[cor:rho.bound\] Under the assumptions \[con:k1\], \[con:k2\], \[D1\] and \[D2\] it holds for all $h>0$ and $q >0$ $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} k^2(x,y) \{g_h(x,y)-g_0(x)g_0(y)\}{\mathrm{d}}(x,y) &\leq \frac{ \kappa^2 d^{1/2} }{ \{(2 \pi)^d \det(\Sigma)\}^{1/2} } (1-4^{-q})^{-1/4(d-2)} |\rho_h|\\ \int\limits_{{{\mathbb R}}^{2d}} k(x,y)f(x)f(y) \{g_h(x,y)-g_0(x)g_0(y)\}{\mathrm{d}}(x,y) &\leq \frac{ \kappa M d^{1/2} }{ \{(2 \pi)^d \det(\Sigma)\}^{1/2} } (1-4^{-q})^{-1/4(d-2)} |\rho_h|.\end{aligned}$$ [*Proof*]{}: Recall that $\theta(\rho) = 1 + \{1-\rho^2\}^{-d/2} - 2^{d+1}\{4-\rho^2\}^{-d/2}$ for $\rho \in [0,1)$. We seek to find bounds on $\theta$ and the corollary can be proven by an application of Lemma \[lem:normal.quantity\]. By assumption \[D2\] we know there is a $\rho_\ast$ such that $\rho_h^2 \leq \rho_\ast^2<1$ for all $h > 0$. Thus consider $\rho \in [0,\rho_\ast]$. We start by finding a constant $C>0$ with $$\theta'(\rho) = \,\rho \left\{ (1-\rho^2)^{-d/2-1} - 2^{d+1} (4-\rho^2)^{-d/2-1} \right\} d \leq C \rho^2.$$ Thus $C$ can be taken as $C = d \left\{ (1-\rho_\ast^2)^{-d/2-1} - 2^{d+1} (4-\rho_\ast^2)^{-d/2-1} \right\}$. Thus we know that the slope of $\theta$ is always less than that of $C \rho^2$. Finally it holds that $\theta(0) = 0$ and thus $0 \leq \theta(\rho)\leq C \rho^2$, $\rho \in [0,\rho_\ast]$. Under condition \[D2\] it holds $\{1-\rho^2_\ast\}^{-d/2} \leq \{1-2^{-2 q}\}^{-d/2}$, completing the proof by using Lemma \[lem:normal.quantity\]. $\square$ The final preparatory result is used to derive the probablistic bound in Theorem \[th:conc.equality\] (iv) and is similar to Corollary \[cor:rho.bound\]: \[lem:normal.quantity2\] Under the assumptions \[con:k1\], \[con:k2\], \[D1\] and \[D2\] it holds for $\lambda>0$ $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} k(x,y) \langle (S+\lambda)^{-1} k(\cdot,x),k(\cdot,y)\rangle_\calH \left\{ f_h(x,y)-f_0(x)f_0(y) \right\}{\mathrm{d}}(x,y) \leq \tilde{c} |\rho_h|{d_\lambda},\end{aligned}$$ with $\tilde{c} = \sqrt{d \{1-4^{-q}\}^{-d-1}} \kappa$. Proof: Denote $\beta(x,y) = \langle (S+\lambda)^{-1} k(\cdot,x),k(\cdot,y)\rangle_\calH$. By the Cauchy-Schwarz inequality $$\begin{aligned} \notag \phi_h &= \int\limits_{{{\mathbb R}}^{2d}} k(x,y) \beta(x,y) \left\{ f_h(x,y)-f_0(x)f_0(y) \right\}{\mathrm{d}}(x,y)\\ \label{eq:two.integral.cauchy} & \leq \left[\int\limits_{{{\mathbb R}}^{2d}} k^2(x,y) \beta^2(x,y) f_0(x) f_0(y){\mathrm{d}}(x,y) \int\limits_{{{\mathbb R}}^{2d}}\left\{ \frac{f_h(x,y)}{\sqrt{f_0(x)f_0(y)}}-\sqrt{f_0(x)f_0(y)} \right\}^2{\mathrm{d}}(x,y)\right]^{1/2}.\end{aligned}$$ Denote by $X'$ and $X''$ two independent copies of $X_0$ and $k' = k(\cdot,X')$, $k''=k(\cdot,X'')$. We start by bounding the first integral term in the product: $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} k^2(x,y) \beta^2(x,y) f_0(x) f_0(y){\mathrm{d}}(x,y) & = {\mathrm{E}}\left\{ k^2(X',X'')\langle (S+\lambda)^{-1} k',k''\rangle_\calH^2\right\}\\ & = {\mathrm{E}}\left[ \mathrm{tr}^2\left\{(S+\lambda)^{-1} k' \otimes k' k'' \otimes k'' \right\} \right]\\ &\leq \kappa^2 \mathrm{tr}^2\{(S+\lambda)^{-1} S\} = \kappa^2 {d_\lambda}^2.\end{aligned}$$ In the second to last inequality we used Lemma \[lem:sup1\] (iv) and the definition of ${d_\lambda}$. The second integral in the product in (\[eq:two.integral.cauchy\]) is $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}}\left\{ \frac{f_h(x,y)}{\sqrt{f_0(x)f_0(y)}}-\sqrt{f_0(x)f_0(y)} \right\}^2{\mathrm{d}}(x,y) & = \int\limits_{{{\mathbb R}}^{2d}} \frac{f_h(x,y)}{\sqrt{f_0(x)f_0(y)}}{\mathrm{d}}(x,y) -1.\end{aligned}$$ We proceed in the same way as in the proof of Lemma \[lem:normal.quantity\] by using properties of the Gaussian distributions at hand. First we have $$F_h(x,y) = \frac{f_h^2(x,y)}{f_0(x)f_0(y)} = \frac{(2\pi)^d \tau_0^d \mathrm{det}(\Sigma)}{(2\pi)^{2d} \mathrm{det(\Sigma_h)}} \exp\left[ -\frac{1}{2} (x^\T,y^\T) \left\{ 2\Sigma_h^{-1} - \frac{1}{\tau_0} \left( \begin{matrix} \Sigma & 0\\ 0 & \Sigma \end{matrix} \right) \right\}^{-1} \left( \begin{matrix} x\\ y \end{matrix} \right) \right].$$ Denote $G^{-1} = 2\Sigma_h^{-1} - \tau_0^{-1} \left( \begin{matrix} \Sigma & 0\\ 0 & \Sigma \end{matrix} \right) $. It follows in a similar way as in the proof of Lemma \[lem:normal.quantity\] that $\mathrm{det}(G) = \tau_0^{2d} \mathrm{det}^2(\Sigma)$ and $\mathrm{det}(\Sigma_h) = (\tau_0^2 - \tau_h^2)^d \mathrm{det}^2(\Sigma)$. Hence we have $$\begin{aligned} \int\limits_{{{\mathbb R}}^{2d}} F_h(x,y) {\mathrm{d}}(x,y) &= \frac{(2\pi)^d \tau_0^d \mathrm{det}(\Sigma)}{(2\pi)^{2d}(\tau_0^2 - \tau_h^2)^d \mathrm{det}^2(\Sigma)} (2\pi)^d \tau_0^d \mathrm{det}(\Sigma)\\ & = \frac{\tau_0^{2d}}{(\tau_0^2-\tau_h^2)^d} = \frac{1}{(1-\rho_h^2)^d}.\end{aligned}$$ Under \[D2\] we have $\rho_h < 1$ for all $h>0$ and there is a $\tilde{\rho} = \max_h |\rho_h| \leq 2^{-q} < 1$ and hence it holds $(1-\rho_h^2)^{-d} \leq d (1-4^{-q})^{-d-1} \rho_h^2$ and we are done. $\square$ ### Proof of the theorem First note that the the operator norm is dominated by the Hilbert-Schmidt norm. By Markov’s inequality we have for $\nu \in (0,1]$ $$\begin{aligned} {\mathrm{P}}\left( \|{S_n}- {S}\|_{\mathrm{HS}}^2 \leq \nu^{-1} {\mathrm{E}}\|{S_n}- {S}\|_{\mathrm{HS}}^2 \right) &\geq 1 - \nu,\\ {\mathrm{P}}\left( \|{T_n^\ast y}- {S}f\|_\calH^2 \leq \nu^{-1} {\mathrm{E}}\|{T_n^\ast y}- {S}f\|_\calH^2 \right) &\geq 1 - \nu.\end{aligned}$$ (i) It holds due to ${S_n}= n^{-1} \sum_{t=1}^n k_t \otimes k_t$ $$\begin{aligned} {\mathrm{E}}\left(\|{S_n}- {S}\|^2_{{\mathrm{HS}}}\right) = \frac{1}{n^{2}}\sum\limits_{t,s=1}^n \left( {\mathrm{E}}[{\mathrm{tr}}\{(k_t \otimes k_t)(k_s \otimes k_s)\}] - 2 {\mathrm{E}}[{\mathrm{tr}}\{( k_0 \otimes k_0) {S}\}] + \|{S}\|^2_{{\mathrm{HS}}} \right).\end{aligned}$$ For the first summand we get ${\mathrm{E}}[{\mathrm{tr}}\{( k_t \otimes k_t)( k_s \otimes k_s)\}] = {\mathrm{E}}\{k^2(X_t,X_s)\}$, due to Lemma \[lem:sup1\] (i). Using the stationarity of $\{X_t\}_{t=1}^n$ and Lemma \[lem:sup1\] (iii) we get $$\begin{aligned} {\mathrm{E}}\left(\|{S_n}- {S}\|^2_{{\mathrm{HS}}}\right) = \frac{1}{n} \left\{{\mathrm{E}}\{k^2(X_0,X_0)\} - \|{S}\|^2_{{\mathrm{HS}}} \right\} + \frac{2}{n^{2}} \sum\limits_{h=1}^{n-1} (n-h) \left[ {\mathrm{E}}\{k^2(X_h,X_0)\} - \|{S}\|_{{\mathrm{HS}}}^2 \right],\end{aligned}$$ yielding the first result by an application of Lemma \[lem:sup1\] (ii). \ For the second equation we see due to the independence of $\{X_t\}_{t=1}^n$ and $\{\varepsilon_t\}_{t=1}^n$ that $$\|T^\ast_n y - {S}f\|^2_\calH = \sigma^2 n^{-1} {\mathrm{E}}\{k(X_0,X_0)\} + {\mathrm{E}}\left(\|{S_n}f - {S}f\|^2_\calH\right).$$ The rest follows along the same lines as the first part of the proof. \(ii) An application of part (i) of this theorem, Corollary \[cor:rho.bound\] and Lemma \[lem:acf.inequality\] yields this result. \(iii) Because the $\{\varepsilon_t\}_{t\in {{\mathbb Z}}}$ are independent and identically distributed and $\{X_t\}_{t \in {{\mathbb Z}}}$ is stationary it holds $$\begin{aligned} &\mathrm{E}\left\{\left\| (S+\lambda)^{-1/2}(S_n f - T_n^\ast y) \right\|^2_\calH\right\}\\ & = \mathrm{E}\left\{\left\| (S+\lambda)^{-1/2} T_n^\ast \varepsilon \right\|_\calH^2\right\} = n^{-2} \sum\limits_{t,s=1}^n \mathrm{E}\left\{ \left\langle \varepsilon_t (S+\lambda)^{-1} k_t, \varepsilon_s k_s \right\rangle_\calH\right\}\\ & = n^{-1}\sigma^2 \mathrm{E}\left\{ \left\langle (S+\lambda)^{-1} k_0, k_0 \right\rangle_\calH \right\} = n^{-1}\sigma^2 \mathrm{E}\left\{ \left\|(S+\lambda)^{-1/2} k_0\right\|^2_\calH\right\}.\end{aligned}$$ By the definition of ${d_\lambda}$ we get $$\mathrm{E}\{\|(S+\lambda)^{-1/2}k_0\|_\calH^2\} = \mathrm{E}[\mathrm{tr}\{ (S+\lambda)^{-1} k_0 \otimes k_0 \}] = \mathrm{tr}\{ (S+\lambda)^{-1} S \} = {d_\lambda}.$$ Using $n^{-1/2} \leq \gamma_n(q)$ proves the result. \(iv) Consider first $$\begin{aligned} {\mathrm{E}}\left\{ \|(S+\lambda)^{-1/2}({S_n}- S)\|_{\mathrm{HS}}^2\right\} &=n^{-2}\sum_{t,s=1}^n {\mathrm{E}}\left[ \mathrm{tr}\{ (S+\lambda)^{-1}(k_t \otimes k_t - S)(k_s \otimes k_s - S)\}\right]\\ &=n^{-1} {\mathrm{E}}\|(S+\lambda)^{-1/2}(k_0 \otimes k_0 - S)\|_{\mathrm{HS}}^2\\ & + \sum_{h=1}^{n-1} {\mathrm{E}}\left[ \mathrm{tr}\{(S+\lambda)^{-1}(k_0 \otimes k_0 - S)(k_h \otimes k_h - S)\} \right].\end{aligned}$$ Continuing with the expression inside the sums we expand $$\begin{aligned} \phi_h &= {\mathrm{E}}\left[ \mathrm{tr}\{(S+\lambda)^{-1}(k_0 \otimes k_0 - S)(k_h \otimes k_h - S)\} \right]\\ &= {\mathrm{E}}\left[ \mathrm{tr}\{(S+\lambda)^{-1} k_0 \otimes k_0k_h \otimes k_h\}\right] - \mathrm{tr}\{(S+\lambda)^{-1}S^2\}\\ &= {\mathrm{E}}\left\{ k(X_0,X_h) \langle (S+\lambda)^{-1} k_0, k_h\rangle_\calH \right\} - \mathrm{tr}\{(S+\lambda)^{-1}S^2\}.\end{aligned}$$ Using Lemma \[lem:sup1\] (iv) we see that $$\mathrm{tr}\{(S+\lambda)^{-1} S^2\} = \int\limits_{{{\mathbb R}}^{2d}} k(x,y) \langle (S+\lambda)^{-1} k(\cdot,x),k(\cdot,y)\rangle_\calH \mathrm{d}{\mathrm{P}}^X(x) \mathrm{d}{\mathrm{P}}^X(y).$$ Hence we have $$\phi_h = \int\limits_{{{\mathbb R}}^{2d}} k(x,y) \langle (S+\lambda)^{-1} k(\cdot,x),k(\cdot,y)\rangle_\calH \left\{ \mathrm{d}{\mathrm{P}}^{X_0,X_h}(x,y)- \mathrm{d}{\mathrm{P}}^{X_0}(x) \mathrm{d}{\mathrm{P}}^{X_0}(y)\right\}.$$ This can be bound by the results of Lemma \[lem:normal.quantity2\] and together with Lemma \[lem:acf.inequality\] there exists a constant $C(q)>0$ such that with probability at least $1-\nu$ $$\|(S+\lambda)^{-1/2}({S_n}- S)\|_\calL^2 \leq \nu^{-1} C^2(q) \gamma^2_n(q) {d_\lambda},$$ with $\gamma^2_n(q) = \left\{\begin{array}{cc} n^{-1},& q>1\\ n^{-1} \log(n), & q=1\\ n^{-q}, & q\in (0,1). \end{array} \right.$ This implies $\|(S+\lambda)^{-1/2}({S_n}- S)(S+\lambda)^{-1/2}\|_\calL \leq \nu^{-1/2} C(q) \lambda^{-1/2}\sqrt{{d_\lambda}}\gamma_n(q)$. Let $\lambda = \lambda_n$ be a sequence converging to zero such that $\lambda_n^{-1/2}\sqrt{d_{\lambda_n}} \gamma_n(q) \rightarrow 0$. Let $n$ be large enough such that $\nu^{-1/2} C(q)\lambda_n^{-1/2} \sqrt{d_{\lambda_n}} \gamma_n(q) < 1$. Using Lemma A.5 in @Blan10b we obtain $$\|(S+\lambda)^{1/2}({S_n}+ \lambda)^{-1/2}\| \leq [1-\nu^{-1/2} C(q) \lambda_n^{-1/2} \sqrt{d_{\lambda_n}} \gamma_n(q)]^{-1/2} \leq \sqrt{2}.$$ The latter inequality can be fulfilled for $n$ large enough such that $\nu^{-1/2} C(q) \lambda_n^{-1/2} \sqrt{d_{\lambda_n}} \gamma_n(q) \leq 1/2$. $\square$ Proof of Proposition \[prop:source\] ------------------------------------ Recall that ${S}u = {\mathrm{E}}\{u(X_0) k(\cdot,X_0)\}$ for $u \in \calH$. Define the independent random variables $Y_1,\dots,Y_\mu$ that are all distributed as $X_0$. First consider the following observation for $\mu \in {{\mathbb N}}$: $$\begin{aligned} S^\mu u &= S (S^{\mu-1}u) = {\mathrm{E}}_{Y_1} \{(S^{\mu-1}u)(Y_1) k(\cdot,Y_1)\} ={\mathrm{E}}_{Y_2}{\mathrm{E}}_{Y_1}\{(S^{\mu-2}u)(Y_2) k(Y_1,Y_2) k(\cdot,Y_1)\} = \dots \notag \\ & = {\mathrm{E}}_{Y_\mu} \cdots {\mathrm{E}}_{Y_1} \left\{\prod\limits_{\nu=1}^{\mu-1} k(Y_\nu,Y_{\nu+1}) u(Y_\mu) k(\cdot,Y_1)\right\}. \label{eq:iterated.expectation}\end{aligned}$$ We take $u = \sum_{i=1}^\infty c_i k(\cdot,z_i)$ for $\{z_i\}_{i \in {{\mathbb N}}}, \{c_i\}_{i \in {{\mathbb N}}} \subset {{\mathbb R}}$ such that $\|u\|_\calH^2 = \sum_{i,j=1}^\infty c_i c_j k(z_i,z_j) \leq R^2$. The fact that a function $u \in \calH$ can be represented as a linear combination of kernel functions is due to the Moore-Aronszajn Theorem, see @Berlinet. Define the matrix $\Gamma = [\Gamma_{i,j}]_{i,j=1}^{\mu+2} \in {{\mathbb R}}^{(\mu+2) \times (\mu+2)}$ via $$\Gamma_{i,j} = \left\{ \begin{array}{clc} \sigma_x^{-2} + 2 l &,& i=j = 2,\dots,\mu+1\\ l &, & i=j=1,\mu+2\\ -l &,& |i-j|=1\\ 0 &,& else \end{array}. \right.$$ Then we have via the integration of Gaussian functions and (\[eq:iterated.expectation\]) $$\begin{aligned} f(x) &= \frac{1}{(2\pi \sigma^2_x)^{\mu/2}} \sum\limits_{i=1}^\infty c_i \int\limits_{{{\mathbb R}}^\mu} \exp\left\{-1/2 (x,x_1,\dots,x_\mu,z_i) \Gamma (x,x_1,\dots,x_\mu,z_i)^\T \right\} {\mathrm{d}}(x_1,\dots,x_\mu)\\ & = \frac{1}{\sigma^\mu_x\det(\Gamma_{2:\mu+1})^{1/2}} \sum\limits_{i=1}^\infty c_i \exp\left[ - 1/2 \det(\Gamma_{2:\mu+1})^{-1}\left\{ \det(\Lambda_{1:\mu+1}) (x^2 + z_i^2) - 2 l^{\mu+1} x z_i \right\}\right].\end{aligned}$$ Here we used the symmetry property $\det(\Gamma_{2:\mu+2}) = \det(\Gamma_{1:\mu+1})$ as the first and last rows and columns of $\Gamma$ are identical. This concludes the proof. $\square$ Proof of Proposition \[prop:source\] ------------------------------------ In @Shi08 it was shown that the eigenvalues of $S$ have the form $\mu_i = a b^{i-1}$, $i=1,2,\dots$ with $$a = \sqrt{2}(1+\beta+\sqrt{1+\beta})^{-1/2},~~ b = (1+\beta+\sqrt{1+2\beta})^{-1} \beta.$$ and $\beta = 4 l \sigma^2_x$. It is clear that $0<b<1$ and hence $0 < \mu_i \leq a$. We have $d_\lambda = \sum_{i=0}^\infty \{1+ a^{-1} b^{-i}\lambda\}^{-1}$. Denote $f(x) =\{1+ a^{-1} b^{-x}\lambda\}^{-1}$. We want to apply the integral test to the sum. We have $\int_0^\infty f(x) \mathrm{d}x = \log^{-1}(b^{-1})\log(1+a \lambda^{-1})$. This yields the bounds $$\begin{aligned} \frac{\log(1+a/\lambda)}{\log(b^{-1})} \leq d_\lambda \leq \frac{1}{1+\lambda/a} + \frac{\log(1+a/\lambda)}{\log(b^{-1})}.\end{aligned}$$ On $\lambda \in (0,1]$ we get $d_\lambda \leq D \log(1+a/\lambda)$ for a constant $D>0$. This can be seen as follows: The function $g_1(\lambda)=(1+\lambda/a)^{-1}$ is bounded from above by $C_1 = 1$ and the function $g_2(\lambda) = (b^{-1})\log(1+a/\lambda)$ is lower bounded by $c_2 = \log(1+a)$ and has no upper bound. Hence on the set $I=\{\lambda \in (0,1]: g_2(\lambda) \geq C_1\}$ we can choose $C = 2$. On the set $I^c$ we have on the other hand $C g_2(x) \geq c_2 C \geq g_1(x) + g_2(x)$, hence we need $C = 2 c_2^{-1} C_1 = 2 \log^{-1}(1+a)$. The choice $D = 2 \log^{-1}(b^{-1}) \max\{1, \log^{-1}(1+a)\}$ is sufficient and we have $d_\lambda \leq D \log(1+a/\lambda)$, $\lambda \in (0,1]$. $\square$
{ "pile_set_name": "ArXiv" }
--- author: - Fabrice Vidal - Eric Gendron - Gérard Rousset - Tim Morris - Alastair Basden - Richard Myers - Matthieu Brangier - Fanny Chemla - Nigel Dipper - Damien Gratadour - David Henry - Zoltan Hubert - Andy Longmore - Olivier Martin - Gordon Talbot - Eddy Younger bibliography: - 'canary\_PhaseA\_AA\_FINAL.bib' title: 'Analysis of on-sky MOAO performance of CANARY using natural guide stars' --- Introduction ============ The understanding of the physics and of the formation of high redshift galaxies requires the multiplexed observations of very large number of targets with spatial and spectral resolution capabilities. Because of their faintness this science case also requires the photon-collecting area of an extremely large telescope (ELT) of 30-40 m class. During the E-ELT instrument phase A studies several MOS concepts were studied, amongst which EAGLE [@Cuby10], a near infrared multi-integral field spectrograph fed by Multi-Object Adaptive Optics (MOAO) [@Hammer02]. EAGLE is a near-infrared (IR) multi-object integral field spectrograph with 37.5 mas spatial sampling and a spectral resolution of 4000. It aims to simultaneously analyse 20 targets in a wide field of view (FOV) of about 7.5 arcmin in diameter. Thanks to the small angular size of the high-z galaxies, adaptive optics (AO) is required [@Puech2010]. Indeed, all the ground-based large telescopes suffer from spatial resolution degradations due to atmospheric turbulence, leading to an image full width at half maximum (FWHM) of the order of 1 arcsec. To overcome these effects, AO is required to compensate for the wavefront distortions in real time as demonstrated at the end of the 1980s in astronomy [@Rousset90]. However, conventional AO is only able to compensate for turbulence across a relatively small FOV of the order of a few tens of arc seconds in the near IR (H band for instance) due to the anisoplanatism effect. Therefore new AO technologies are required to overcome this strong limitation. Multiconjugate AO (MCAO) [@beckersMcao1988; @Marchetti08] and Ground Layer AO (GLAO) [@RigautVeniceGlao02; @Milton08] are novel AO system concepts that offer larger compensated FOV. However MCAO can deliver a good correction only in a FOV of the order of 1 or 2 arcmin while GLAO delivers a moderate correction, with a typical gain of a factor of 2 in FWHM with respect to the seeing [@bendekMMT2011; @Tokovinin2012], but in a much wider FOV up to the order of 10 arcmin. It is not possible to design and implement an MCAO system covering such a wide FOV because of the conservation of beam étendue. Using the specific case of a high redshift galaxy programme, only the galaxies themselves are of interest for the AO correction, not the continuous FOV. In addition, their angular extent is very limited to the order of 1 or 2 arcsec. The multi-object AO (MOAO) concept was initially proposed to tackle this problem [@Hammer02]. It aims to simultaneously compensate the turbulence for a large number of very faint small science objects distributed over a wide FOV. For that purpose one deformable mirror (DM) is implemented per target in a dedicated optical train feeding each integral field unit of a spectrograph. Moreover, the galaxies are too faint to be able to measure any wavefront distortion in real time. It is therefore necessary to find a number of bright guide stars (GS) within the wide instrumental FOV for that purpose. These guide stars can be natural GS (NGS), but for questions of sky coverage, laser GS (LGS) are needed. The light from each GS must be picked-off in the FOV and sent to a wavefront sensor (WFS). In such a configuration, the WFSs do not see any feedback from the correction applied on the DMs; they are thus working in an open loop mode. Finally all the WFS measurements made across the FOV must be mutually processed in order to reconstruct the 3D turbulent volume above the telescope. This is achieved using a tomographic approach [@Ragazzoni99]. Then a projection of the volume in the direction of each target [@Fusco01; @Vidal10JOSA] allows us to compute the correction to be applied to each DM in the system in open loop. A possible configuration for such a MOAO system is the one envisioned for the EAGLE instrument of the E-ELT [@Rousset10]. It makes use of 6 LGSs at the periphery of the FOV and between 4 to 6 additional NGSs selected in the field. The two main critical features of MOAO are tomographic reconstruction of the turbulence and the DM operation in open loop. In order to improve the technology readiness of the EAGLE instrument, a pathfinder was proposed to demonstrate the feasibility and the performance of MOAO [@Myers08]: this is the CANARY project. A first on-sky testing of on-axis open-loop command has been already reported [@Volt08] as well as first MOAO laboratory demonstrations [@Ammons10; @Vidal10AO4ELT]. The next step is to demonstrate on-sky the capability of the open loop tomography using a number of in-the-field GS, both laser and natural. This is the main goal of CANARY. Tomography has been successfully demonstrated in MCAO with MAD at VLT on NGS only [@Marchetti08] and more recently with a 5 LGS constellation on GeMS at Gemini Observatory [@Rigaut2013_Gems_1; @Neichel2014_Gems_2]. However, the MCAO approach uses multiple DMs conjugated at different altitude in the optical train of the instrument. This leads to a specificity of MCAO: the tomography problem is partially optically solved. This maybe directly taken into account by regularizing the pseudo-inverse of the interaction matrix between the WFSs and the DMs. In MCAO, Minimum Mean Square Error (MMSE) reconstructors without the turbulence profile knowledge still allows us to achieve a good level of performance [@Vidal2013]. In MOAO, with one DM per target the least-squares approach only leads to a partial correction of the wavefront, i.e. correcting only the ground layer. In fact, the knowledge of the turbulence profile is not crucial in MCAO but becomes a limiting factor in MOAO and is even more dramatic as the GS can be positioned farther off-axis from the target object. Considering the open loop operation, typical DM errors such as creep and hysteresis have been studied earlier [@Morzinski2008; @kellererDM2012]. For certain DM types including that in CANARY, they have been demonstrated to be small compared to the fitting error. In addition specific open loop procedures have to be developed as presented in section \[ControlAndCalibration\] for CANARY. In Sect. \[instrument\], we present the CANARY instrument with a hardware and software description and in section  \[ControlAndCalibration\] the control and calibration algorithms. The data reduction approach is described in Sect. \[reduc\]. In particular, the error budget computation is detailed. Section \[OnskyResults\] presents the on-sky results, the observing conditions and introduce the three selected NGS asterisms. The data are processed to retrieve a 15 layer $C_n^2(h)$ profiles in order to quantify the tomographic and open loop errors. Section \[details\] gives the detailed error budgets established for three sets of data during one night of observations. The results are compared to numerical simulations. We conclude in Sect. \[conclusion\]. The CANARY instrument {#instrument} ===================== Introduction ------------ The CANARY pathfinder implements a single target channel of a MOAO system and is deployed at one of the Nasmyth foci of the 4.2m William Herschel Telescope (WHT) at the Roque de Los Muchachos Observatory on the island of La Palma in Spain. CANARY is a fast track experiment. The design phase started beginning of 2008 while the first light happened in September 2010 [@Gendron11]. CANARY is a project planned in three phases (respectively called A, B, C), each phase of increasing system complexity [@Myers08]. It will lead to a comprehensive demonstration of the MOAO configuration as foreseen for EAGLE on the E-ELT [@Rousset10]. Phase A makes first use of three open-loop off-axis NGS WFSs and one on-axis open-loop-controlled DM of $8\times8$ actuator array. Phase B will use, in addition, four Rayleigh LGS to augment the turbulence tomography in open loop (see Fig. \[CANARY\]). Phase C will introduce the full configuration proposed for EAGLE. We will use a woofer DM in closed-loop (in fact the 8x8 actuator DM), as a first stage simulating M4 of the E-ELT [@VernetM4_2012] with an open-loop tweeter DM as a second stage simulating the MOAO channel. The tweeter DM in Phase C will be a higher-order DM from ALPAO with $17\times17$ actuators (241 useful). In order to reconfigure the system between the three phases, the optical design is build around a set of interchangeable optical modules. The Rayleigh LGSs used for these demonstrations have variable range-gate height and extension in order to simulate many of the LGS effects that will be encountered at the E-ELT. ![ CANARY MOAO demonstration configuration. Phase A of the system (left) uses three off-axis NGS to reconstruct the wavefront while one deformable mirror is driven in open loop on the on-axis star. Phase B (right) added 4 Rayleigh LGS in addition of the three NGS. []{data-label="CANARY"}](CANARY_noyear.pdf){width="8cm"} Phase A system -------------- ![The CANARY input focal plane is equipped with a Target Acquisition System (TAS) carrying 3 off-axis WFS working in open loop. Light from the central star is transported by the off-axis parabolic relay including the Tip-Tilt stage (TT) and the deformable mirror (DM) to the output corrected focus. At this focus, the light is reflected in the visible by a dichroic plate to a fourth WFS called the Truth Sensor. Finally, the IR light is transmitted to the IR camera observing at a central wavelength of $\lambda=1\,530$ nm.[]{data-label="CANARYBench"}](benchV2_BW.pdf){width="9cm"} ![Zoom of the CANARY focal plane. Upper part illustrates the on-sky configuration where the off-axis light is reflected to the off-axis WFS. The light from the central star is sent via the DM and TT to the IR camera and TS. Lower part illustrates the open-loop calibration configuration using the reverse path calibration source. The light is illuminating the DM and TT and is retroreflected at the focal plane on one off-axis WFS. This allows us to retrieve interaction matrices for the off-axis WFS and compute misalignment and scaling errors between all the WFS.[]{data-label="FocalPlaneZoom"}](FocalPlaneZoom_BW.pdf){width="9cm"} At the entrance of the bench, upstream of the Nasmyth focal plane, the field rotation is compensated using a K-mirror derotator. CANARY is equipped with three WFS that are able to patrol the derotated field ($2.5\arcmin$ in diameter, plate scale 0.22 mm arcsec$^{-1}$) and acquire the selected off-axis NGSs. The star that mimics the corrected science target is situated at the centre of the field. This on-axis central star is observed by an IR imaging camera and a fourth WFS called *Truth Sensor* (TS) which measures the DM-corrected wavefront (Figs  \[CANARYBench\] and  \[FocalPlaneZoom\]). The four WFS are identical, of Shack-Hartmann (SH) type, with $7\times7$ sub-apertures (only 36 illuminated). The total field of view per WFS sub-aperture is close to 4. They use Andor iXonEM 860 EMCCD cameras with $24~\mu$m pixel size, featuring $128\times128$ pixels. The measured read-out noise is 0.3, 0.5, 0.55 and 0.7 e$^-$ rms per pixel for the four cameras due to the Electron Multiplication gain. We use $16\times16$ pixels of $\approx 0.26\arcsec$ scale in each SH sub-aperture. The minimum distance between two off-axis WFSs is $20\arcsec$, limited by both hardware and software anti-collision systems. The central star beam is sent to the performance diagnostic IR camera and TS via a two off-axis parabolic mirror relay. The relay includes the CILAS $8\times8$ piezostack array DM conjugated to the WHT pupil. This DM was manufactured in 1991 and previously used at ESO in the ADONIS instrument [@beuzit94]. It is made in a very hard piezoelectric material and exhibits hysteresis and creep effects [@kellererDM2012]. Wavefront errors caused by “creeping” of the DM surface are observed but their effet on performance is minimal. The high-speed tip-tilt mirror is a copy of that used in the VLT instrument SPHERE [@Sphere2008]. The mirror pointing is controlled by a servo-loop based on internal position sensors to minimise open-loop error. At the output of the relay, the beam is split by a dichroic plate: the transmitted IR flux (1 to 2.5 $\mu$m) is sent to the IR camera. This is a Xeva-1.7-320 from Xenics, featuring $320\times256$ pixels of $30~\mu$m pitch, with a measured readout noise of $200\pm30~\mathrm{e}^-$ rms per pixel and a gain of $10~\mathrm{e}^- \ \mathrm{ADU}^{-1}$. The plate scale of the IR camera is $0.0371\arcsec$ per pixel. The combination of the H filter bandwidth, the QE efficiency of the camera and the atmosphere absorption gives an effective wavelength of the IR image centered at $\lambda=1\,530$ nm with a bandwidth of $\Delta \lambda = 160$ nm. The visible part of the spectrum (below 900 nm) is reflected by the dichroic and goes to the TS. The TS allows us to check the efficiency of the turbulence compensation in real-time, to perform system calibration tasks and to close the loop for performance comparison between conventional closed loop AO (hereafter called Single Conjugate AO: SCAO) and open loop MOAO. A telescope, turbulence and star simulator allow us to fully characterise the system when under testing in the laboratory or during daytime maintenance at the telescope. A deployable acquisition camera can be inserted to observe the whole field (2.5 arcmin diameter) for calibration of the telescope pointing and guide star acquisition. The entrance focal plane of CANARY is also equipped with deployable calibration sources of various diameters to emulate both seeing or diffraction-limited sources. In addition, a specific feature is the so-called reverse-path calibration source, discussed in Sect. \[interactionmatrix\]. A more detailed overview of the CANARY can be found in [@Gendron10]. Real-time computer {#wpurtc} ------------------ The real-time computer (RTC), called DARC (Durham Adaptive optics Real time Controller) is described in [@Dipper10] and [@Basden2010]. The system is driven at a selectable sampling frequency of up to 250 Hz, limited by the WFS camera readout rate. The four WFS cameras are synchronised and read through dedicated FPGA-based hardware. The AO controller is CPU based, and optimised for multi-threaded operation on multi-CPU architectures. It exhibits $0.7$ ms latency between the latest readout pixels of the WFSs and the time when the DM actuators are reaching half-stroke value. DARC also allows clients to obtain continuous or sub-sampled real-time data streams of SH images, slopes, and DM command voltages. DARC also features a number of different real-time algorithms. The SH image processing algorithm uses adaptive windowing ($12\times12 $ pixels) of the SH spots on a sub-aperture-by-sub-aperture basis, together with centroiding done on 10 brightest pixels [@basdenBrightestPix2012]. These features allowed us to cope with the dynamics required for open-loop wavefront sensing depending on the observing conditions and to reduce the effect of readout noise. Loop delay ---------- We measured the loop delay by introducing a known, white noise, voltage pattern on the DM command. By recording synchronised sets of TS slopes and DM voltages, we can deduce the delay from the WFS measurements. We measured a 1.5 and 1.9 frame delay at 150 and 250 Hz respectively. We define delay as the time taken from the beginning of the WFS integration to the application of the corresponding command on the DM. Control and calibrations algorithms {#ControlAndCalibration} =================================== Interaction matrices and deviations {#interactionmatrix} ----------------------------------- The interaction matrix between the DM and the truth sensor is an important calibration item in the control of CANARY. As in conventional AO, the interaction matrices reflect the optical relationship between the DM and the WFSs [@Boyer90]. They are measured experimentally by actuating the mirror and measuring the impact on a given wave-front sensor. They are even more important in our open-loop scheme, since we are now concerned with not only the relative amplitude between coefficients, but also with their [*absolute*]{} amplitude. Whereas a global multiplicative factor would just act as a loop gain when operating in closed loop, here it will operate as a scaling factor on the correction. In CANARY, we have chosen to simultaneously activate all actuators [@kellererDM2012; @Marchetti08] with a sinusoidal wave pattern using a specific temporal frequency for each and retrieve the matrix from the WFS measurements using a lock-in approach. This particular method has been demonstrated [@Vidal2009; @kellererDM2012] to be one of the best for identifying the DM model. The method has also demonstrated excellent on-sky capabilities for interaction matrix identification with the TS, the quality of the non-zero matrix coefficients only differs by 2% from the laboratory ones. We will call $M_i$ the interaction matrix between DM and TS, and $M_c$ its generalised inverse. The matrix $M_c$ is computed using a singular value decomposition of $M_i$. We observe a significant drop-off of the last eigenvalues and filtered the last 7 modes to keep a conditioning number of 50. In an open-loop scheme, the focal plane off-axis wave-front sensors have no optical feedback from the deformable mirror. This, in turns, means that no interaction matrix can be measured physically, since the WFS do not see the DM. This limitation, inherent to the MOAO scheme, is a severe drawback since the optical relation between wave-front sensors and DM does exist within the tomographic reconstructor. That is why CANARY is equipped with a reverse path calibration source, which is a seeing-limited source illuminating the DM from the output focal plane to one of the open-loop WFS located in the input focal plane. This is achieved using a retroreflector system, which preserves the pupil orientation (\[FocalPlaneZoom\]). Therefore we are able to record interaction matrices with each off-axis WFS, despite the open-loop scheme. Those matrices are the starting point for finding out all instrumental model parameters which are: pupil image translations and magnifications, rotations of the lenslet arrays and of the WFS cameras, and WFS sensitivities. This tool was first used for the purpose of fine alignment and also for final calibration of all the WFS parameters [@Vidal10JOSA; @Brangier2012]. These parameters could then be taken into account in the computation of the tomographic reconstructor. The Learn & Apply tomography algorithm {#LandA} -------------------------------------- The linear tomographic reconstruction matrix is derived using the *Learn & Apply* (L&A) algorithm from [@Vidal10JOSA], where the minimum mean square error (MMSE) reconstructor is computed from an atmospheric model directly identified on-sky from the WFS measurements. Engaging the open-loop compensation is preceded by the acquisition of an on-sky WFS data set, from which the optimized estimator is deduced. We detail hereafter the implementation. We define notations as follows. The set of local wavefront slopes along $x$ and $y$ directions, measured by a WFS $i$ with $p$ sub-apertures, is collected into a slope vector $\vec{S_i}$ $$\vec{S_i} = \left( \begin{array}{cc} sx_1 & \\ \vdots &\\ sx_p &\\ sy_1 & \\ \vdots &\\ sy_p &\\ \end{array} \right) - \vec{S}_{ref\ i}$$ After acquisition of number of frames *nframe* of an on-sky WFS data set we build a matrix $M_{offaxis}$ by concatenating horizontally the column-vectors formed by synchronous measurements $t=[1, ... , nframe]$ from the $n$ off-axis WFSs. Each WFS slope is subtracted from its own reference slope ($\vec{S}_{ref\ i}$) including all the static aberrations for the observing direction of the NGS $i$ (see Section \[RefslopesCal\]). $${M_{offaxis}} = \left( \begin{array}{ccc} \vec{S}_{1\ t=1} & \cdots & \vec{S}_{1\ t=nframe}\\ \vec{S}_{2\ t=1} & \cdots & \vec{S}_{2\ t=nframe}\\ \vdots & \vdots &\vdots\\ \vec{S}_{n\ t=1} & \cdots & \vec{S}_{n\ t=nframe} \end{array} \right).$$ Similarly, in the same time, synchronous slope vector measurements from the on-axis truth sensor are concatenated in the same way to form the matrix $M_{{central}}$ $${M_{central}} = \left( \begin{array}{cccccccc} \vec{S}_{central\ t=1} & \cdots & \vec{S}_{central\ t=nframe}\\ \end{array} \right).$$ [@Vidal10JOSA] infers that for this particular turbulence sequence the best tomographic estimator $M_t$ is the one that directly links inputs and outputs (i.e. $M_{offaxis}$ and $M_{central}$) as $$M_{central} = M_t.M_{offaxis} \ , \label{equationTomo}$$ and solving, in a least-squares sense, this equation for $M_t$ is given by [@Vidal10JOSA] $$M_t = (M_{central} {M^t_{offaxis}}) (M_{offaxis}{M^t_{offaxis}})^{-1}\ . \label{equationTomoSol}$$ It is noticeable that this estimator $M_t$ tends towards the MMSE estimator as the acquisition time tends towards infinity (and under hypothesis of stationarity), because both matrices $(M_{offaxis} {M^t_{offaxis}})$ and $(M_{central}{M^t_{offaxis}})$ tend towards the covariance matrix of the sensor measurements. We respectively call these matrices $C_{OffOff}$ and $C_{OnOff}$. Then we can write Eq. \[equationTomoSol\] as $$M_t = C_{OnOff} \, . \, C^{-1}_{OffOff} \ . \label{RetrieveTomo}$$ The covariance between two elementary slopes in $x$ of sub-apertures $\theta \in [1, ..., p] $ and $\nu~\in~[1,...,m]$ of respectively two WFS $i$ and $j$ is noted $<s_{x\theta} \, s_{x\nu}>_{ij}$. Here $p$ and $m$ are respectively the total number of subapertures for WFS $i$ and $j$. The covariance value can be theoretically computed from the Kolomogorov covariance maps : expressions in direct space have been given by [@butterleySLODAR] and some analytical approximations have recently been given by [@martinSpie2012]. We expressed covariance in the Fourier domain for a single turbulent layer as $$\begin{gathered} \mathcal{F} ( <s_{x\theta} \vec(k) s_{x\nu} \vec(k)>_{ij}) \propto k_x^2 \, \tilde{\Pi}_i(\vec{k}) \, \tilde{\Pi}^*_j(\vec{k}) \, r_0(h_l)^{-\frac{5}{3}} \, \| \vec{k} \|^{-\frac{11}{3}} \\ \times e^{ -2i\pi \left[ k_x (h_l(\alpha_i - \alpha_j) + x_{i\theta} - x_{j\nu} ) + k_y (h_l(\beta_i - \beta_j) + y_{i\theta} - y_{j\nu} ) \right] } \label{ExpressionCovariancegen}\end{gathered}$$ with $r_0 (h_l)$ the Fried parameter at altitude $h_l$, $\vec{k} = (k_x,k_y)$ the conjugate Fourier variable of the separation $\vec{r} = (x,y)$ between two sub-apertures $i$ and $j$, and $\tilde{\Pi}_i(\vec{k})$ the Fourier transform of the sub-aperture shape function $\Pi(\vec{r})$ (i.e. equal to 1 within the sub-aperture and 0 elsewhere). The separation between two WFS is characterised by the difference in their pointing directions ($\alpha_i,\beta_i$) and ($\alpha_j,\beta_j$), and the altitude $h_l$ of the considered layer. During a real scientific observation the number of time-independent realisations is limited, which leads to covariance matrices that have partly converged. Moreover, the general case is that it is impossible to acquire the data from the central direction with the truth sensor because of the faintness of the scientific target. We have therefore developed a method where we fit a model to the off-axis covariance matrices in order to extract the essential parameters of the turbulence and WFS configurations. Therefore, we are able to re-generate the modelled covariance matrices and compute the tomographic estimator from Eq. \[RetrieveTomo\]. We call $C_{OffOff\ raw}$ the real on-sky measured off-axis covariance matrix. In the CANARY case during this first run, we also use the TS data to fit the $C_{OnOff\ raw}$ matrix. We typically use sets from 10000 to 90000 synchronised slopes (i.e from $\approx$ 1 to 10 mn at 150 Hz) to compute the on-sky measured covariance matrices $C_{OffOff\ raw}$ and $C_{OnOff\ raw}$. In order to retrieve the parameters $h_l$, $r_0(h_l)$, ($\alpha_i$, $\beta_i$) and ($\alpha_j$, $\beta_j$), we minimize the distance $\epsilon$ between the covariance matrix model to the measured one: $$\begin{gathered} \epsilon = ||C_{OffOff\ raw} - \sum_l C_{OffOff}\left( h_l, r_0(h_l), (\alpha_i, \beta_i), (\alpha_j, \beta_j)\right)||^2 \\ + ||C_{OnOff\ raw} - \sum_l C_{OnOff}\left( h_l, r_0(h_l), (\alpha_i, \beta_i), (\alpha_j, \beta_j)\right)||^2\ . \label{critere}\end{gathered}$$ We use a Levenberg-Marquardt fitting algorithm to perform the minimisation of $\epsilon$. The number of layers is not retrieved by the algorithm but defined by the user before starting the fitting procedure. This method allowed us to retrieve on-sky the turbulence profile and geometric configuration from the recorded data in a few minutes. This step is called: *Learn*. In a sense, the *Learn* resembles a SLODAR analysis, since it allows us to retrieve the $C_n^2(h)$ profile, with the minimization process performing an inversion of the direct problem. However, the $C_n^2(h)$ profile is only a by-product of the [*Learn*]{}, the real output being the modeled covariance matrices. Using the determined turbulence and geometric parameters we compute the theoretical matrices $C_{OffOff}$ and $C_{OnOff}$, then using Eq. \[RetrieveTomo\] we compute the tomographic estimator $M_t$ which will be used in the RTC. We called this step *Apply*. To retrieve the parameters we used the following procedure: A first calibration is ran on the bench using one phase screen to simulate the Ground layer. We record a set of WFS slopes on all the WFS. The parameters fitted (using LM algorithm) are the pupil misalignments (x,y, theta and magnification G) in addition to the altitude of the layer (known with a limited optical precision on the bench). Once the pupil misalignments known we use them to align better the system and iterate until the x,y, theta and G are close to 0 (or 1 for G) and become negligible. Simultaneously we fit the strengths and altitudes of the layers together with the on-sky WFS positions (see section \[profile\]). Software implementation. {#implementation} ------------------------ In MOAO mode, $M_t$ is an output of the L&A tomographic algorithm. We emphasise that this matrix is able to reconstruct the slopes that the central sensor should see ($\vec{S_c}$) from the correlation with the off-axis slope measurements $\left( \vec{S_1}, ..., \vec{S_n} \right)$. Therefore, $M_t$ is a *slope tomographic estimator* for the central WFS: $$\vec{S_c} = M_t \left( \begin{array}{l} \vec{S}_1\\ \cdots\\ \vec{S}_{n} \end{array} \right)\ .$$ In the phase A configuration of CANARY we have 3 off-axis WFS and 1 central WFS (TS). For practical implementation we also consider the TS as a fourth off-axis WFS and we have $$\vec{S_c} = M_t \left( \begin{array}{l} \vec{S}_1\\ \vec{S}_2\\ \vec{S}_3\\ \vec{S}_{TS} \end{array} \right)\ .$$ The three left quarters of the matrix $M_t$, corresponding to off-axis parts, are computed according to Eq. \[equationTomoSol\], the right quarter is zero-padded in order to ignore the TS slopes in the command. We can therefore rewrite $M_t$ as $$M_t = \left( \begin{array}{ccccccccccccccccccccc} M_{t1} & M_{t2} & M_{t3} & 0\\ \end{array} \right)\ .$$ In GLAO mode, we average the slopes from the 3 off-axis WFSs leading to the matrix $M_t{_{GLAO}}$ defined as: $$M_t{_{GLAO}} = \frac{1}{3} \left( \begin{array}{ccccccccccccccccccccc } Id & Id & Id & 0\\ \end{array} \right)$$ with $Id$ being a square identity matrix. The GLAO scheme presented here is achieved with the DM operating in open-loop. This is slightly different to the usual GLAO definition that implies closed-loop operation. The term GLAO used in the rest of this paper describes our averaging of the open-loop off-axis WFS measurements, hence open-loop GLAO. Finally, as the TS is placed after the DM, we can use it to close the loop in SCAO mode. We disable the off-axis WFS slopes by filling the off-axis WFS part of the reconstructor with zeros. The SCAO reconstructor becomes $$\label{mctSCAO} M_t{_{SCAO}} = \left( \begin{array}{ccccccccccccccccccccc} 0 & 0 & 0 & Id \\ \end{array} \right)\ .$$ In CANARY, the TS command matrix $M_c$ (see also Sect. \[interactionmatrix\]) is a 72-by-54 array. This command matrix is used to link the slopes in the central direction ($\vec{S_c} $, for instance predicted by the $M_t$ matrix) to the voltage to apply on the DM. The final tomographic command matrix (the one loaded in the RTC), is noted $M_{ct}$ and is defined by $$M_{ct} = M_c M_t\ .$$ Depending on the mode we are running (MOAO, GLAO or SCAO), we use the corresponding $M_t$ matrix ($M_t$, $M_t{_{GLAO}}$ or $ M_t{_{SCAO}}$). AO controller {#Control} ------------- In close loop configuration (SCAO) we use a conventional integrator temporal controller which takes the form $$\label{closeloopEq} \vec{V}_{t} = \vec{V}_{t-1} + g.M_{ctSCAO}. \left( \begin{array}{l} \vec{S}_1(t)\\ \vec{S}_2(t)\\ \vec{S}_3(t)\\ \vec{S}_{TS}(t) \end{array} \right)$$ with $\vec{S}_{i}(t)$ the current slopes vector measurement and $\vec{S}_{ref\ i}$ the reference slopes vector of the $i^{\mathrm{th}}$ WFS. The parameter $g$ is the temporal loop gain of the integrator [@Gendron95]. In the RTC implementation, $g$ is a vector with a size of the number of actuators (54). It is possible in particular to filter the DM and the Tip-Tilt voltages differently. Notice in Eq. \[closeloopEq\] that the measurements from the off-axis WFSs are unused due to the zero terms in the identity matrix defining $M_{ctSCAO}$ (Eq. \[mctSCAO\]). Only the slopes from the TS are actually used to compute the DM voltages in this close loop configuration. The open-loop controller is a temporal filter of the form $$\label{openloopEq} \vec{V}_{t} = (1-g)\vec{V}_{t-1} + g.M_{ct} \left( \begin{array}{l} \vec{S}_1(t)\\ \vec{S}_2(t)\\ \vec{S}_3(t)\\ \vec{S}_{TS}(t) \end{array} \right)\ . $$ This controller is used for both GLAO and MOAO open loop modes. As the last quarter part of the $M_{ct}$ is filled with zeros, only the off-axis slopes are used to compute the vector $\vec{V}_{t}$. The final set of voltages to be applied to the DM by the RTC is in fact the vector $\vec{U}$: $$\label{staticOffset} \vec{U}(t) = \vec{V}(t) + \vec{V}_{offset}\ .$$ The offset vector $\vec{V}_{offset}$ is the static shape to be applied to the DM in order to produce the best static point spread function (PSF) on the IR camera. The computation of this vector is given in the next section. Reference slopes calibration and DM offset {#RefslopesCal} ------------------------------------------ The TS reference slopes $\vec{S}_{refTS}$ are deduced by recording the average position of the SH spot, $\vec{S}_{PlaneWF}$, with a reference source producing a flat wavefront placed at the entrance of the TS. We then close the loop on the TS in SCAO mode on the internal calibration source. We measure the Non Common Path Aberrations (NCPA) between the TS iteratively with a phase diversity algorithm [see @Sauvage2007]. We subtract their contribution, expressed in terms of slopes $\vec{S}_{NCPA}$, to the plane wavefront reference slopes to compute the final TS reference slopes: $\vec{S}_{refTS} = \vec{S}_{PlaneWF} - \vec{S}_{NCPA}$. The determination of $\vec{S}_{ref\ i}$ is different. In open loop, the off-axis WFSs measure their own aberrations, plus the telescope field aberrations and the derotator quasi-static aberrations in addition to the turbulence. These static aberrations are not compatible with the tomographic model underlying the estimator $M_t$, and they must be treated separately. The strategy is just to consider them as constant over the interval of observation, and subtract them from the measurements as reference slopes. We measure them *in-situ* by averaging a long time sequence of the turbulence (we use the same sequence that served to compute the tomographic estimator) measured by all the WFSs: $\vec{S}_{ref} = <\vec{S}_{turbu}>_t + \vec{S}_{static}$. As the turbulence has a zero average $<\vec{S}_{turbu}>_t= 0$, the average value of slopes measured by the off-axis WFSs are their corresponding static aberrations in the field and are set as reference slopes $\vec{S}_{ref\ i} = \vec{S}_{static\ i}$. DM offset voltages {#DmOffsetVoltage} ------------------ The on-axis telescope and derotator quasi-static aberrations $\vec{S}_{static\ TS}$ can also be measured using the same procedure. These measured on-axis aberrations are subtracted from the TS reference slopes in order to find the aberration slopes. To ensure their compensation on-axis, these slopes have to be converted into a static voltage vector $\vec{V}_{offset}$ (also taking into account the NCPA): $$\label{offset} \vec{V}_{offset} = - M_c (\vec{S}_{static\ TS} - \vec{S}_{PlaneWF} + \vec{S}_{NCPA})\ .$$ The vector $\vec{V}_{offset}$ remains applied on the DM (Eq. \[staticOffset\]) during the whole duration of the observation in open loop. This strategy works provided the condition $<\vec{S}_{turbu}>_t= 0$ is respected, otherwise a static error will propagate through the estimator $M_t$ due to incorrect reference slopes, and will add to the DM due to an incorrect $\vec{V}_{offset}$. This is exactly what is observed in reality, but we are fortunately able to demonstrate that those two errors terms opportunely cancel each other (or at least partly, under certain conditions). For the sake of simplicity in this explanation, we will lighten our notation. We call $s$ the vector of the static term for the off-axis WFS, and $t$ the turbulence. We write $s'$ and $t'$ the same quantities measured by the on-axis truth sensor. We also define $r$, the reference slope vector on the truth sensor. As said previously, the average of the quantities related to turbulence are non-zero, and we have $<t>=\epsilon_t \ne 0$ and $<t'>=\epsilon_t'\ne 0$. Our hypothesis is that, as the estimator $M_t$ is linear and designed to work on turbulence, we have $$\label{linearavg} \left\{ \begin{array}{l} M_t t = t' \\ M_t \epsilon_t = \epsilon_t'\ . \end{array} \right.$$ Now, at the end of the calibration of the reference slopes and the DM offsets, the off-axis WFS will be set with $(s+\epsilon_t)$ as their reference slopes, and the DM offset will be $-(s'+\epsilon_t'-r)$ (we skip $M_c$ for the sake of simplicity). Both terms contain errors, as they include partly-converged quantities $\epsilon$. However, now the measurements from the off-axis WFS for any turbulence $t$ are $(s + t - s - \epsilon_t)$. Thus, the total voltage applied on the DM will be $-M_t(t-\epsilon_t) -(s'+\epsilon_t'-r)$, and the wavefront on-axis will be $s'+t'-M_t(t-\epsilon_t) - (s'+\epsilon_t'-r)$. Taking into account relation \[linearavg\], it follows that all terms vanish and the expression is left with $r$. The TS will measure the wavefront with respect to its own reference slopes $r$, which gives $r-r=0$, i.e. a flat, corrected wavefront, in spite of the errors in both the reference slopes and the DM offset voltages. The success of the on-sky measurement method of all static offsets relies on the fact that $\vec{S}_{static\ i}$ and $\vec{V}_{offset}$ are estimated from the [*same*]{} set of data. In this case, the error we make is related only to partial convergence of the average of non-tomographic aberrations (unseen layers, drifts in the instrument or telescope, etc) and not to the partial convergence of the average of the turbulence itself. Data reduction {#reduc} ============== We have PSF images on the target direction (i.e. the central star) taken using the IR camera. In parallel to the IR images, the real-time slopes of all WFSs (off-axis and TS) plus DM voltage data are saved to determine the atmospheric parameters and evaluate the error budget. We record two types of synchronous data for all the WFSs: - *engaged slopes* recorded while the MOAO loop is engaged. It stands for the open loop sensing on the 3 off-axis WFS (used to compute the correction on-axis) and residual slopes seen by the TS. - *disengaged slopes* recorded while the MOAO loop is NOT running. DM is flattened allowing measurements of the turbulence in open loop on each of the 4 NGS WFS directions. We now present the procedure used to reconstruct different terms in the error budget of CANARY from these data. IR image performance estimation {#IRcalcul} ------------------------------- The background is subtracted from the IR images and dead pixels are removed using a pixel map previously established on a dark image. The IR images shown here are simply the average of 30 individual exposures of 1 second (not tip-tilt removed). The IR image was centered in a region of the detector with no dead pixel to avoid any bias in the estimation on the PSF. Strehl ratios (SR) have been computed on the IR images by normalizing their total energy to unity, and dividing their peak value by that of the diffraction-limited pattern sampled identically. This peak value is given by $a=\frac{\pi}{4}(D^2-o^2)p^2/\lambda^2$ with $o$ the central obscuration diameter, and $p$ the camera pixel scale in rd pixel$^{-1}$. First, we calibrate the pixel scale using an internal source at $\lambda=$1550 nm, and we ensure that the WFS pixel scale and IR camera pixel scale are consistent together. This is important to ensure that all absolute WFS-related values (r0, and any nm rms value measured from wavefront data) will be consistent with the IR camera pixel scale. Then a supplementary calibration has been done on-sky using a reference double star imaged on the IR camera to determine the final pixel scale. We close the loop using the TS and measure the Optical Transfert Function (OTF) of the IR image. From the OTF, we estimate the cut-off frequency at D/$\lambda$ and check the number of pixels in $\lambda/D$. The normalisation of the flux is dramatically sensitive to the estimation of background level to be subtracted. Special care has been taken with this operation, first narrowing the field to only 64$\times$64 pixels around the source and then estimating the residual background level using edge pixels. We estimate the uncertainty on the SR due to normalization errors in the background to be of the order of 0.02 (2%). Seeing estimation {#seeing} ----------------- The off-axis WFS continuously measure the open-loop turbulence, irrespectively of the loop being engaged, when operating in MOAO or SCAO modes. They can always be used to estimate the Fried parameter $r_0$ at any time. For each of the 3 off-axis WFS, $r_0$ is computed by fitting the theoretical variances of the Zernike decomposition of the Kolmogorov spectrum [@Noll76] to those of the experimentally reconstructed wave-front. Before fitting, the experimental variances are corrected from the wave-front sensing noise measured on the slopes as explained in Sect. \[Noise\], and propagated onto the Zernike coefficients. The wave-front is reconstructed on Zernike modes from $Z_2$ to $Z_{36}$, but the fit only takes into account modes 4 to 27. Tip-tilt is excluded, as it may be polluted by the telescope tracking or vibrations, and is definitely influenced by the outer scale $L_0$. We did not observed any significant vibration on the temporal spectra of the higher order modes. Modes 28 to 36 are excluded too, because as they are the last radial order they are more affected by aliasing effects. $r_0$ is given at $500$ nm and at the airmass of the observation (not rescaled to zenith). The final $r_0$ estimation seen by CANARY is computed by taking the mean of the estimated $r_0$ from the 3 off-axis WFS measurements. CANARY error budget {#CANARYErrBudget} ------------------- The estimated wavefront error $\sigma_{Err}$ can be translated in an expected SR using the formula SR = exp(-(2$\pi\sigma_{Err}/ \lambda)^2)$ and compared to the SR measured on the IR image. The overall error budget of CANARY is given for the IR on-axis channel where the images are recorded. The total error budget on the IR camera, denoted $\sigma_{ErrIR}$, can be expressed as $$\begin{gathered} \label{IRTotal} \sigma^2_{ErrIR} = \sigma^2_{NCPA} + \sigma^2_{FieldStat} + \sigma^2_{Fit} + \sigma^2_{TomoNoiseFilt} \\ + \sigma^2_{BW} + \sigma^2_{AliasGround} + \sigma^2_{AliasAlt} + \sigma^2_{Tomo} + \sigma^2_{OL}\ ,\end{gathered}$$ and we will assume in this paper that all these terms are independent, so that variances add up together. We define now each of the individual terms of the error budget and describe how we compute them. ### NCPA error $\sigma^2_{NCPA}$ {#NCPA} The best SR obtained on-bench, after the NCPA estimation by phase diversity and compensation by the DM, is 0.80 $\pm$ 0.02 ($\approx$ 115 nm rms ). It corresponds to residual NCPA and the high spatial frequencies non correctable by the DM. We call the both contributions as NCPA noted $\sigma_{NCPA}$. The best flat of the DM is 50 nm rms determined from interferometric measurements measured in May 2009. The poor error figure is explained by high spatial frequencies that developed between actuators due to the aging of the mirror. ### Field static aberrations $\sigma^2_{FieldStat}$ In MOAO, the off-axis WFSs and the TS measure, in addition to the atmospheric turbulence, static aberrations across the telescope field of view. For the TS the DM creep is also included in this term. These aberrations mainly come from the telescope and the derotator and may vary across the field. We consider them as quasi static aberrations as they slowly evolve during the night. We calibrate and subtract the field aberrations in the MOAO loop. However, a non-perfect determination of the off-axis static aberrations leads to a residual static error after the MOAO correction in the on-axis direction. Any additional DM creep due to the static term $V_{offset}$ applied to the DM during open-loop operation can also contribute to this error term. The total on-axis static error noted $\sigma_{FieldStat}$ can be measured by the TS by averaging its slope measurements while the loop is engaged. This error term is related to the inaccuracies in the calibration of the static offsets, that have already been described in Sect. \[DmOffsetVoltage\]. ### Fitting error $\sigma^2_{Fit}$ The estimation of the DM fitting error $\sigma_{Fit}$ (or undermodelling error) has been determined using a Monte-Carlo simulation. We have computed the residual phase variance of computer-generated Kolmogorov wave-fronts subtracted from their best least-squares fit on the DM modes. We have restricted the modes to the 47 actually controlled by the system, out of the 54 degrees of freedom (52 on the DM, 2 on the tip-tilt mirror). The 7 filtered modes correspond to a resulting conditioning number of 50. This fitting error was found to be $$\label{fitcanary} \sigma^2_{Fit} = 0.0122 \left( \frac{D}{r_0} \right) ^{5/3}.$$ Then, the fitting error term of our error budget is a number which will simply be derived from the knowledge of $r_0$, using the above equation. [@Roddier1999] gives the fitting error as function of the number of actuators $n_a$ in the pupil diameter: $\sigma^2_{Fit} = 0.335 n_{a}^{-5/3} ( {D}/{r_0} ) ^{5/3}$ and replacing $n_a$ with $2 \sqrt{N_{tot}/\pi}$ leads to $$\sigma^2_{Fit} = 0.274 \, N_{tot}^{-5/6} \left( \frac{D}{r_0} \right) ^{5/3},$$ which gives $N_{tot}=42$ in our case when identifying with Eq. \[fitcanary\]. This is extremely close to the number of actuators that are truly laying within the pupil, as depicted on Fig. \[pupil\]. This latter represents the configuration of the subapertures within the pupil and of the actuators of CANARY. One can see that a total of 12 actuators (represented with circles) are poorly seen by the WFS, leading to only 40 actuators in the pupil, a value compatible with $N_{tot}=42$. ![Geometric configuration of the 36 useful sub-apertures of a Shack-Hartmann WFS of CANARY across the WHT pupil (with a 0.286 normalised diameter central obscuration). Actuators within the pupil are represented with a cross. Actuators out of the pupil are also represented with circles.[]{data-label="pupil"}](pupil_BW.pdf){width="5cm"} Now, we can also compare Eq. \[fitcanary\] with this one derived by [@Conan94], $$\label{Noll} \sigma^2_{Fit} = 0.257 \, N_{Z}^{-5/6} \left( \frac{D}{r_0} \right) ^{5/3}$$ that gives the fitting error after the correction of the $N_z$ first Zernike polynomials. We find that our DM amounts to $N_Z \approx 38$ Zernike modes. In the following sections of this article, we express all error terms in nm rms. They are computed by using a Zernike decomposition of the WFS slopes. We have chosen to consider a wavefront reconstruction on $N_Z$= 36 Zernike polynomials corresponding to radial order $n=7$, as this is nearly equivalent to our DM. We define $M_{iz}$ the matrix containing the response of the SH to the Zernike modes 2 to 36, and $M_{rz}$ the Zernike reconstruction matrix (generalized inverse of the $M_{iz}$ matrix). ### Noise error $\sigma^2_{Noise}$ {#Noise} We compute the noise $\sigma^2_{Noise}$ on the WFS slope signals from their temporal autocorrelation. As noise and turbulence are two independent processes, the autocorrelation of their sum is the sum of autocorrelations. At $\Delta t$=0, in particular, the measured variance is the sum of variances of noise and turbulence. While the autocorrelation of noise is a Dirac function, that of turbulence is wide and locally smooth around $\Delta t =0$ : this allows us to separate them by fitting a parabola $(a . \Delta t^2+b)$ to the two points of the autocorrelation at $\Delta t$=1 and $\Delta t$=2 and extrapolating the purely turbulent variance at $\Delta t = 0$. This value is subtracted from the total variance at $\Delta t = 0$, to obtain the noise variance on each slope. The WFS slope noise is then propagated through the Zernike reconstruction matrix $M_{rz}$ on 36 modes in order to be converted into a wavefront error. We compute the propagated noise by making the assumption that there is no spatial correlation between noise on any of the slopes, i.e. the covariance matrix of the slope noise is a pure diagonal matrix, noted $\sigma_{slopesnoise}^2$. The propagated noise is then given by $$\label{noisePropagation} \sigma^2_{Noise} = \mathrm{Trace}(M_{rz} Diag(\sigma_{slopesnoise}^2) M_{rz}^t) \ .$$ This equation is used for the TS noise computation $\sigma^2_{NoiseTS}$. The diagonal terms on the noise matrix depends on the GS magnitude and considered sub-aperture. ### Noise through the tomographic estimator $\sigma^2_{TomoNoiseFilt}$ {#noiseTomo} This term is computed from noise on the off-axis WFS, computed as described in the previous paragraph, propagated through the tomographic slopes estimator matrix $M_t$, and propagated through the Zernike reconstruction matrix $M_{rz}$: $$\sigma^2_{TomoNoise} = \mathrm{Trace}\left( (M_{rz}M_t) Diag(\sigma^2_{Off-axis\ slopesnoise}) (M_{rz}M_t)^t \right)\ .$$ This white noise is filtered by the temporal controller (see Sect. \[Control\], Eq. \[openloopEq\]), and only a fraction of it is actually injected in the DM command. The reduction of the noise variance due to the low-pass filtering is given by a factor of $\frac{g}{2-g}$ (demonstrated in Appendix \[append\_g/2-g\]), where $g$ is the loop gain: $$\label{TomoNoisefilt} \sigma^2_{TomoNoiseFilt} = \left( \frac{g}{2-g} \right) \, \sigma^2_{TomoNoise}\ .$$ ### Bandwidth error $\sigma^2_{BW}$ {#bandwidthError} The bandwidth error is computed using a set of recorded engaged-loop WFS data using the off-axis WFS with the highest signal-to-noise ratio (in open loop). The principle is to numerically filter this real data set in order to reproduce frame-by-frame the loop behaviour in engaged mode, then derive the error from the simulated residuals. The method ensures the computation of a bandwidth error term that corresponds to the precise recorded turbulence conditions. A first step computes the open loop controller filtering (see Sect. \[Control\]) of the slope measurements: $$\label{filtFirstOrder} S'_{t} = (1 - g) \, S_{t-1} + g \, S_{t}\ .$$ Then a second step simulates the application of the DM command, which acts as a zero-order hold, and that occurs partly during the integration of the subsequent frame $t+1$, and partly during the frame after that at $t+2$. Writing the fractional delay, expressed in frames, as $(1+\alpha)$, with $0 \le \alpha < 1$, then the residual signal $R_t$ is $$\label{decalFraction} R_t = S_t - (\alpha \, S'_{t-2} + (1-\alpha) \, S'_{t-1}) \ ,$$ and the variance of this last signal $R_t$ characterizes the bandwidth error, except it just needs to be unbiased from noise. As the on-sky data $S_t$ unavoidably includes noise, we need to correct the variance of $R_t$ from the noise variance that has propagated through our filtering process, since we aim to compute here the bandwidth error associated only with turbulence (the impact of noise has been treated in  \[noiseTomo\]). It is demonstrated in Appendix \[append\_gaga\] that a pure white noise of unitary variance, filtered as described in Eqs. \[filtFirstOrder\] and \[decalFraction\] is output with a variance given by $(1 - 2 g \alpha (1-\alpha))g/(2-g)$. Finally, to express the bandwidth error as a wavefront error, we reconstruct the measurements $R_t $ in a Zernike basis using the matrix $M_{rz}$, and we compute the variance of each Zernike coefficient. The noise variance on each slope is also propagated on the Zernike coefficients as described in Eq. \[noisePropagation\], the factor $(1 - 2 g \alpha (1-\alpha))g/(2-g)$ is applied on each, and the result is subtracted from the variance of $R_t $. The BW error term also includes the vibration error term since the integrator controller does not efficiently reject the vibrations. The vibration term in CANARY has been studied in [@Kulcsar2012]. We measured Tip-Tilt vibration peaks between 20Hz and 50Hz and of the order of 100nm rms for the night of the 27th of September. ### Aliasing error $\sigma^2_{Alias}$ {#aliasing} Aliasing is due to the fact that the WFS spatially samples the wavefront across the pupil of the instrument. The high spatial frequencies of the wavefront disturbances will be undersampled if they are higher than the Nyquist frequency (i.e. higher than the half of the sampling frequency). Thereby, these high frequencies are mistaken for low spatial ones. Another way to say it is that the incoming wavefront $\phi$ does not entirely lie within the mirror subspace, it contains a component $\phi_\perp$ orthogonal to it (thus giving rise to the fitting error). This orthogonal wavefront is seen by the wavefront sensor but unfortunately produces non-zero measurements that will be reconstructed and mistaken by the system as mirror modes. [@quiros2010_genAliasing] has given an extensive analysis of aliasing in tomographic applications, and we do not aim to do the same here. We will use a simplified, approximate approach to disentangle the different aliasing effects on the system. As already stated in section \[CANARYErrBudget\] our main assumption is a negligible coupling between the terms of the error budget, in particular $\sigma^2_{Alias}$ with $\sigma^2_{BW}$ and $\sigma^2_{Tomo}$. First, we compute the covariance matrix of the turbulence on the Zernike basis $C_{zz}$ with 900 polynomials (which we assume to be large enough to represent an infinite number) normalized to $D/r_0=1$ using the formulae given by [@Noll76]. Then, we zero the lines and rows corresponding to the first modes up to $Z_{36}$ in order to obtain a covariance matrix of the turbulence $C_{zz\perp}$ only of high orders (of $\phi_\perp$, as named above). This matrix is supposed to mimic the statistics of the phase orthogonal to the DM space. We then compute the slopes covariance matrix of this high-order turbulence, noted $C_{ss}$, using a transformation matrix from Zernike to slopes $M_{iz\ 900}$, computed here for the first 900 Zernike polynomials: $$C_{ss} = M_{iz\ 900} \, C_{zz\perp} \, M_{iz\ 900}^t\ .$$ The matrix $C_{ss}$ represents the statistics of the aliased wavefront on the WFS. We finally use the Zernike reconstruction matrix $M_{rz}$ on the $C_{ss}$ slopes covariances matrix to compute the aliasing error, $$\label{aliasGendron} \sigma^2_{Alias} = \mathrm{Trace}\left( M_{rz} C_{ss} M_{rz}^t \right) \ \left( \frac{D}{r_0} \right)^{5/3} \, ,$$ and we should now evaluate how the aliasing will propagate through the control. For this, we consider the aliasing effect differently depending on whether layers are placed in altitude or at the ground. We assume that aliasing produced by layers located close to ground level is fully correlated between all the WFSs including the TS. This aliasing error is injected into the loop and is consequently fully applied to the wavefront by the DM. This introduces an error on the IR camera that is not seen on the TS slopes in first approximation. We compute its effects as a fraction $X_{Ground}$ of the total turbulence on the telescope: $$\sigma^2_{AliasGround} = X^{}_{Ground} \ \sigma^2_{Alias}\ .$$ Contrarily to ground, we assume the aliasing produced in altitude behaves as a white spatial noise, not correlated between the off-axis WFSs. We consider to be a noise contribution by averaging on the WFSs, $$\sigma^2_{AliasAlt} = X^{}_{Altitude} \ \frac{\sigma^2_{Alias}}{n_{WFS}}\ ,$$ where $ X_{Altitude}$ is the fraction of the turbulence which is not at the ground and $n_{WFS}=3$ the number of WFSs. Using the last two equations, we neglect the impact of the open loop filtering on the aliasing terms. Finally, the aliasing in altitude, produced only by the TS, is computed by $$\sigma^2_{AliasAltTS} = X^{}_{Altitude} \ \sigma^2_{Alias}\ .$$ ### Tomographic error $\sigma^2_{Tomo}$ {#tomo} We compute the tomographic error on a given data set by estimating the residuals between the non-engaged measurements of the TS and the synchronous tomographic prediction. We also have to unbias the tomographic error from several additional effects, that we detail in Eq. \[tomoeq\]. This latter is computed by multiplying the off-axis slopes $\vec{S}_{offAxis}$ of the set with the tomographic estimator $M_t$. When we are dealing with engaged-loop data, we do not have any non-engaged measurements of the TS. Our first attempt for retrieving non-engaged TS data was to subtract the contribution of the DM (with a proper multiplication with the interaction matrix $M_i$, Sect. \[interactionmatrix\]). This approach is incorrect, because it will include the DM open-loop error that cannot be disentangled from a tomographic error. We therefore decided to instead use a set of slopes taken with the loop *disengaged*, just before or after the engaged-loop set, to determine the tomographic error of the latter. This method works provided the value of $\sigma^2_{Tomo}$ is properly rescaled with respect to the $r_0$ value. A factor of $(r_0/r'_0)^{5/3}$ has to be applied to the variance. We then estimated the raw tomographic error by computing the difference between this tomographic prediction and the real measurements made by the TS without introducing any delay for each frame. We express the vector of the residuals onto a Zernike basis: $$\{ a_{i\ Tomoraw} \} = M_{rz}(\vec{S}_{TS} - M_t \vec{S}_{offaxis})\ .$$ The reference slopes are already subtracted from $\vec{S}_{offaxis}$ and $\vec{S}_{TS}$ as described in section \[LandA\]. Then the raw tomographic error $\sigma^2_{Tomoraw}$ is simply given by the sum of the variances of Zernike coefficients: $$\sigma^2_{Tomoraw} = \sum^{36}_{i=2} Var\left( \{ a_{i\ Tomoraw} \} \right)\ .$$ As we compare perfectly synchronised disengaged slope data, we have a direct access to the tomographic error with no temporal effect from the loop filter. However, it still needs to be corrected from the impact of noise and aliasing. We remove the TS noise $\sigma^2_{NoiseTS}$ and the propagated noise from the off-axis WFSs $\sigma^2_{TomoNoise}$. For aliasing, we remove the contribution of the aliasing effect in altitude on the TS $\sigma^2_{AliasAltTS} $ and propagated from the off-axis WFSs $\sigma^2_{AliasAlt}$. It follows that, provided *disengaged slopes*, the pure tomographic error $\sigma^2_{Tomo}$ can be estimated from $$\begin{gathered} \label{tomoeq} \sigma^2_{Tomo} = \sigma^2_{Tomoraw} - \sigma^2_{TomoNoise} - \sigma^2_{NoiseTS} \\ - \sigma^2_{AliasAltTS} -\sigma^2_{AliasAlt} \ .\end{gathered}$$ We emphasize that the on-sky measured term $\sigma^2_{Tomo}$ could itself be split into different terms, in particular the error term $\sigma^2_{Model}$ of the turbulence profile model, corresponding to an error made on the model used to build the reconstructor. We can have an estimation of this model error by comparing the $\sigma^2_{Tomo}$, evaluated on-sky with the tomographic error evaluated with a numerical simulation using a $C_n^2(h)$ profile strictly equal to the model that served to compute the estimator $M_t$. We give an example of such a comparison in the last section of this paper. ### Open loop error $\sigma^2_{OL}$ This term has also been named “go-to” error in the literature, and corresponds to the fact that the shape the mirror will take for a given set of voltages is not exactly the one that one would expect, due to hysteresis, drifts, non-linearities or any other effect that has not been taken into account in the mirror model. We also underline that in our analysis, the term $\sigma^2_{OL}$ only represents the dynamic wave-front error linked to the open loop behavior of CANARY DM. The static wave-front error term is given by $\sigma^2_{FieldStat}$. As stated in paragraph \[tomo\], the dynamic open loop error is difficult to disentangle from the tomographic error. We decide to estimate it from the *engaged slopes* where we can measure the residual wavefront error seen by the TS, denoted hereafter as $ \sigma_{ErrTS}$. The error seen by the TS can be computed as a sum of the individual terms: $$\begin{gathered} \label{ErrTS} \sigma^2_{ErrTS} = \sigma^2_{Tomo} + \sigma^2_{OL} + \sigma^2_{TomoNoiseFilt} + \sigma^2_{NoiseTS} \\ + \sigma^2_{AliasAltTS} + \sigma^2_{AliasAlt} + \sigma^2_{BW} + \sigma^2_{FieldStat}\ .\end{gathered}$$ Considering that we measure $\sigma^2_{ErrTS}$, $\sigma^2_{OL}$ is the last unknown in Eq. \[ErrTS\]. It can be estimated by $$\begin{gathered} \label{ErrOL} \sigma^2_{OL} = \sigma^2_{ErrTS} - \sigma^2_{Tomo} - \sigma^2_{TomoNoiseFilt} - \sigma^2_{NoiseTS} \\ - \sigma^2_{AliasAltTS} -\sigma^2_{AliasAlt} - \sigma^2_{BW} - \sigma^2_{FieldStat}\ .\end{gathered}$$ *Engaged slopes* are therefore required to compute an estimation of the open loop error and also *disengaged slopes* because an estimation of the $\sigma^2_{Tomo}$ term is needed. Because our estimation is computed by subtracting a large number of estimated terms from $\sigma^2_{ErrTS}$, $\sigma^2_{OL}$ may also include all the estimation errors (finite number of slope samples, approximations, bad calibration, etc). Therefore, it is only a crude estimation of $\sigma^2_{OL}$. On-sky Results {#OnskyResults} ============== We had 2 $\times$ 4 nights split between September (19th, 22nd, 26th and 27th) and November 2010 (from 23rd to the 26th). Unfortunately due to bad weather we lost most of the November nights. We focus in this paper on the results obtained on the fourth night of the September run (2010 Sep. 27th). We alternated the observations between the SCAO (closed loop on the TS), GLAO and MOAO modes as the turbulence profile evolved and as we changed asterisms. During the whole night, the sampling frequency of all WFSs was 150 Hz irrespective of the AO mode. Although the temporal controllers are slightly different between open and closed loop, the RTC latency is the same for all modes. Natural guide stars asterisms ----------------------------- CANARY makes use of star asterisms formed by four NGSs. The central one, placed on-axis and used for diagnostic purposes, mimics the science object that will benefit from the turbulence compensation. We selected asterisms with a distance between the central on-axis star and the three off-axis ones ranging from $15\arcsec$ to $65\arcsec$, while keeping all stars brighter than m$_V$=12. The three observed asterisms of the 54 identified asterisms for the September period are described in Table \[asterisms\] and Fig. \[Ast\]. Asterism \# 47 53 12 --------------- ----------- ----------- ----------- central $m_V$ 11 10.9 8.3 sep () 47.9 61.7 39.3 $m_V$ 9.9 11.2 11.2 sep () 40.6 49.1 31.4 $m_V$ 10.2 9.9 10.7 sep () 53 56.8 51.5 $m_V$ 8.7 9.8 10 airmass 1.02–1.55 1.11–1.50 1.05–1.09 : Three asterisms observed during the night of the 27th of September 2010. The columns indicate: the CANARY asterism reference number, the separation (in arcsec) of each off-axis star to the central one, the V band magnitudes of each and the range of airmass during observations.[]{data-label="asterisms"} ![Map of the 3 asterisms (DSS images) observed during the September run. The tomographic estimator was computed using the 3 off-axis stars and applied in the central direction (white arrows). The dashed circle represents a 2’ diameter field of view.[]{data-label="Ast"}](AsterismSky.pdf){width="9cm"} CANARY tomographic altitude resolution {#tomoreso} -------------------------------------- The maximum spatial frequency measured by a WFS is defined by $f_e = \frac{1}{2d_{ssp}}$ with $d_{ssp}$ the size of the sub-aperture projected on-sky ($d_{ssp}$ = 0.6m for CANARY). For that frequency, the tomographic altitude resolution $\Delta_h$, is computed with the relation $ \Delta_h = \frac{d_{ssp}}{\alpha} $, with $\alpha$ the separation angle between 2 WFSs. The maximum altitude accessible $h_{max}$ is given by the relation $h_{max} = \frac{D_{tel}}{\alpha}$. A large separation angle will allow us to take advantage of a good altitude resolution, while small separations give access to a higher sensing altitude. Because CANARY is equipped with 4 WFSs, each senses the wavefront in a different direction according to the NGS configuration (i.e. the observed asterism). This leads to different tomographic resolutions according to the considered pairs of WFSs. Table \[TomoResolution\] summarizes the altitude resolution for each baseline of each asterism. During the night, the altitude resolution was in the range 1000 to 4000 m. The best resolution of $ \Delta_h=$1083 m is achieved with the pairs of WFS 1 and 2 on asterism A53. It also means that we were more sensitive to a variation of a layer altitude on asterism 53 than asterisms A47 and A12. Because of the very large shift in altitude any layer above 7500 m cannot be seen by this particular pair of WFS but was completed by the measurement of narrower pairs like WFS 2 and 3 that allowed us to sense the turbulence profile up to 14000 m. ---------- -------- ---------- ------------ ----------- Asterism star $\alpha$ $\Delta_h$ $h_{max}$ couple () (m) (m) 1–C 47.9 2583 18085 2–C 40.6 3048 21337 A47 3–C 53 2335 16354 1–2 86.5 1430 10015 2–3 54.8 2258 15808 1–3 92.4 1339 9376 1–C 62.2 1990 13927 2–C 56.7 2183 15279 A53 3–C 48.7 2541 17789 1–2 114.2 1083 7586 2–3 60.26 2054 14376 1–3 105.27 1176 8229 1–C 38.1 3248 22738 2–C 31 3992 27946 A12 3–C 51.8 2389 16724 1–2 35.5 3486 24403 2–3 82.9 1493 10450 1–3 77.9 1589 11121 ---------- -------- ---------- ------------ ----------- : Vertical altitude resolution and maximum altitude in metres computed for each pairs of the 4 WFSs. There are 6 different combinations computed for each of the 3 observed asterisms. $\alpha$ is the angular separation between the 2 considered stars. []{data-label="TomoResolution"} The following sections present the evolution of the seeing conditions, the SR measured on the IR camera, the turbulence profile, the tomographic error, the open loop error and the residual field aberrations during the night of September 27th. Section \[details\] presents a detailed analysis of the error budget for three different cases during the night. Seeing conditions {#evolution} ----------------- Figure \[r0VStime\] plots the $r_0$ estimated from off-axis WFS data (unbiased from noise and aliasing, see Sect. \[seeing\]) versus the local time (in hours, negative before midnight). The $r_0$ value is not rescaled to zenith. Dashed lines represent the asterism change during the night. Asterism A47 was observed from the beginning of the night to [$02^\mathrm{h}00^\mathrm{m}$]{} local time, A53 from [$02^\mathrm{h}00^\mathrm{m}$]{} to [$05^\mathrm{h}00^\mathrm{m}$]{} and A12 from [$05^\mathrm{h}00^\mathrm{m}$]{} to the end of the night. The median value is $r_0=11.33$ cm giving a median seeing of $0.91\arcsec$ at $500$ nm. Worst and best seeing were respectively $1.23\arcsec$ and $0.63\arcsec$. ![$r_0$(cm) at 500 nm from WFS data versus local time (hour) during the night of the 27th September 2010. Dashed lines separate asterisms.[]{data-label="r0VStime"}](r0VStime.pdf){width="8cm"} IR camera images ---------------- Figure \[SR\_comparison3D\] shows an example of IR image comparison at $\lambda=1\,530$ nm for the three AO modes tested on CANARY plus a seeing limited image. The four images of 30 second exposure each were taken respectively at [$00^\mathrm{h}59^\mathrm{m}18^\mathrm{s}$]{} (Seeing), [$00^\mathrm{h}42^\mathrm{m}10^\mathrm{s}$]{} (GLAO), [$00^\mathrm{h}29^\mathrm{m}22^\mathrm{s}$]{} (MOAO) and [$00^\mathrm{h}32^\mathrm{m}28^\mathrm{s}$]{} (SCAO). Measured SR are respectively, 1%, 5.3%, 19.4% and 23.8%. As already underlined by [@Gendron11], the MOAO performance is close to the SCAO one. In particular, despite the strong ground layer observed during the night, we see that MOAO performs much better than GLAO. The following subsections will give more insights into this result. ![IR image comparison at $\lambda=1\,530$ nm. The four images of 30 seconds exposure each were taken at [$00^\mathrm{h}59^\mathrm{m}18^\mathrm{s}$]{} (Seeing), [$00^\mathrm{h}42^\mathrm{m}10^\mathrm{s}$]{} (GLAO), [$00^\mathrm{h}29^\mathrm{m}22^\mathrm{s}$]{} (MOAO) and [$00^\mathrm{h}32^\mathrm{m}28^\mathrm{s}$]{} (SCAO). Measured SR are respectively: 1%, 9%, 19.4% and 23.8%.[]{data-label="SR_comparison3D"}](perfMoao3Dtext.pdf){width="9cm"} Turbulence profile {#profile} ------------------ During our observations, the $C_n^2(h)$ retrieval step (i.e. the *Learn* phase of the L&A algorithm), was limited for practical software reasons (computation time) to only three fitted layers. Since September 2010, we have significantly improved the computation speed of the *Learn* step [see @martinSpie2012]. On-sky, we were jointly fitting, at the same time, the strengths and altitudes of 3 layers and the 4 WFS on-sky positions (observing directions). With the new fitting procedure we now employ a series of altitudes $h_i$, regularly spaced by $\Delta_h$, ranging from ground to $h_{max}$ (see Sect. \[tomoreso\]) and fit only their strength value $C_n^2(h_i)$. The positions of the WFS are measured by the Target Acquisition System encoders in the focal plane of the telescope. The fit is now more robust and thanks to an increase in speed, we are now able to fit up to 15-20 layers in a few tens of seconds. Post-processing the Phase A on-sky data allows us to retrieve detailed $C_n^2(h)$ profiles with up to 15 layers measured between 0 and 20 km (depending on the asterism geometry). We have selected one *disengaged slopes* data set per asterism to illustrate the identification of the turbulence profile and the estimation of the error budget. Table \[TomoSky\] summarises the parameters of these synchronised data sets taken at [$23^\mathrm{h}59^\mathrm{m}$]{} (using 28000 frames, acquisition time of [$3^\mathrm{h}06^\mathrm{m}$]{}), [$3^\mathrm{h}14^\mathrm{m}$]{} (58000 frames) and [$6^\mathrm{h}02^\mathrm{m}$]{} (28000 frames) on asterisms A\#47, A\#53 and A\#12, respectively. Figures \[23h59\], \[03h14\] and \[06h02\] present the $C_n^2(h)$ measured during the on-sky observations (left) and post-processed (right) using these sets of *disengaged slopes*. The turbulence profiles have not been rescaled by the airmass to maintain the profile as it was observed in the line of sight of CANARY. ![$C_n^2(h)$ retrieved on-sky at [$23^\mathrm{h}59^\mathrm{m}$]{} on Asterism $\#$47 (left). More detailed $C_n^2(h)$ with 15 layers (right) obtained after post-processing of the same data set.[]{data-label="23h59"}](23h59Full_BW.pdf){width="8cm"} ![$C_n^2(h)$ retrieved at [$3^\mathrm{h}14^\mathrm{m}$]{} on Asterism $\#$53 (left: on-sky; right: post-processed).[]{data-label="03h14"}](03h14Full_BW.pdf){width="8cm"} ![$C_n^2(h)$ retrieved at [$6^\mathrm{h}02^\mathrm{m}$]{} on Asterism $\#$12 (left: on-sky; right: post-processed).[]{data-label="06h02"}](06h02Full_BW.pdf){width="8cm"} At [$23^\mathrm{h}59^\mathrm{m}$]{} (Fig. \[23h59\]), we measured more than 95% of the turbulence below 1340 m (best resolution achievable on asterism A\#47, see also Table \[TomoResolution\]) during the on-sky operations. A weak layer of 5% was also detected at $\approx$ 3000 m. After post processing the data with a higher vertical resolution sampling, we measured the ground layer contribution at 91% and one high layer at $\approx$13500 m of 9% (right). On the [$3^\mathrm{h}14^\mathrm{m}$]{} dataset (Fig. \[03h14\]), we measured during the observations a ground layer contribution at 77% (74% post processed), a 14% layer contribution at 1080 m (11% post processed) and 1 layer at 10% at $\approx$4000 m (spread into 3 layers of 10% contribution with the new 15 layer profile). We additionally measured with the post processing 3 very weak layers around 11km (few percent) which was not measured and taken into account on-sky. At [$6^\mathrm{h}02^\mathrm{m}$]{} (Fig. \[06h02\]), we measured on-sky 85% in the ground layer contribution (82% post processed) and one layer at $\approx$2000 m of 15% (13% at 1490 m + 3 small layers at $\approx 15$ km of a total contribution of 5% with post-processing). We can see on these plots that even with only three fitted layers we were able to account for most of the turbulence distribution in our computed estimators. But it is also clear that we have a significant model error in our results because of the missing high altitude layers in the estimators. We will quantify this error in Sect. \[details\]. Tomographic error ----------------- Figure \[tomoVStime\] shows the tomographic error versus the local time. We used here sets of *disengaged slopes* to compute the tomographic error (see Eq. \[tomoeq\]). The graph plots the tomographic error for GLAO cases (crosses) and MOAO cases (L&A with 3 reconstructed layers: circles). At any time, the L&A reconstruction performs better than the GLAO reconstructor. The median tomographic error for all data taken over the night gives $\sigma_{tomoMOAO}=$216 nm rms for MOAO while the GLAO reconstruction gives $\sigma_{tomoGLAO}=$270 nm rms. On average throughout the night, MOAO reconstruction performed 160 nm rms better than the GLAO case. Most of the fluctuations in the tomographic error observed in Fig. \[tomoVStime\] are due to the large fluctuations in the seeing conditions during the night. ![Tomographic error (nm rms) measured using *disengaged slopes* versus local time (hour). Asterisms A47, A53 and A12 are respectively represented with crosses, circle and triangles. For each asterism we plot the MOAO and GLAO reconstruction respectively represented in plain and blank points.[]{data-label="tomoVStime"}](tomoVStime.pdf){width="8cm"} ![Tomographic error measured with MOAO (circles) and GLAO (crosses) estimator versus r$_0$(cm). Dashed lines represent a fit to each dataset by a law $\propto \left(D/r_0\right) ^{5/6.}$[]{data-label="tomoVSr0"}](tomoVSr0.pdf){width="8cm"} Figure \[tomoVSr0\] plots the tomographic error as function of $r_0$ value. We superimpose the expected $\left(D/r_0\right)^{5/6}$ law showing the overall behavior of this error term. This trend is globally verified by our estimates of this error. One can see the improvement of the L&A tomographic reconstruction versus the GLAO mode. The large scatter of the MOAO results is due to the large number of configurations tested during the night. Many MOAO datasets were taken when system parameters were not optimal and the observing conditions were also evolving. Figures \[03h14zer\] represent the statistics of Zernike expansion coefficients on the data set of *disengaged* slopes at 03h14. We compute the Zernike expansion of the uncorrected turbulence (black/diamonds), the GLAO reconstruction (black/square) MOAO correction with 3 reconstructed layers (black/circle) and 15 reconstructed layers (empty/cross), The uncorrected wavefront error is measured at a level of 1139 nm rms. As expected, we recognise the Zernike expansion sorted by radial orders (order 1 with tip and tilt, order 2 containing the defocus and the two astigmatisms, etc). The GLAO reconstruction gives a tomographic error of 314 nm rms while the optimized MOAO reconstruction with 3 layers gives 229 nm rms (with the MOAO estimator used on-sky). We also use the optimized 15 layers $C_n^2(h)$ measured a posteriori (see Fig. \[03h14\]) to compute a 15 layer tomographic estimator, apply it to the off-axis data and compare the prediction to the real measurements made on-axis by the TS. The post-processed estimator increases performance by reducing the model error, resulting in a wavefront error of 213 nm rms. Figure  \[03h14zerDiff\] illustrates the quadratic difference between the GLAO and MOAO reconstruction with 3 layers (black/triangles) as function of the Zernike number. One clearly sees the improvement for all modes using the MOAO reconstructor with 3 layers (tomographic error reduced by 215nm rms). Similarly Fig.  \[03h14zerDiff\] illustrates the quadratic difference between the MOAO reconstruction with 3 layers (black/circles) and MOAO reconstruction with 15 layers (black/triangles). For this dataset the total improvement by reconstructing 15 layers instead of 3 could be reduced by 84nm rms. ![Zernike expansion coefficient root mean square (log scale) of uncorrected wavefront (black/diamond), tomographic reconstruction with GLAO (black/square) MOAO correction with 3 reconstructed layers (black/circle) and 15 reconstructed layers (cross), for the *disengaged* slopes dataset of [$3^\mathrm{h}14^\mathrm{m}$]{}.[]{data-label="03h14zer"}](zer03h14.pdf){width="8cm"} ![We plot the quadratic difference between the GLAO and 3 layer MOAO reconstruction (triangles) as a function of the Zernike number using the same dataset as Fig \[03h14zer\]. The 3 layer reconstruction improves the tomographic reconstruction by 215nm rms in total. The quadratic difference between the 3 layer and 15 layer MOAO reconstruction is illustrated with circles. The 15 layers reconstruction allows us to reduce the tomographic error by 84nm rms.[]{data-label="03h14zerDiff"}](zer03h14_quadraticDiff.pdf){width="8cm"} Table \[TomoSky\] summarizes all the numbers for 3 data sets of *disengaged* slopes respectively at 23h59, 3h14, and 6h02 and highlights the quadratic differences between error terms. It shows that the 15 layer profile increases the performance of the tomographic reconstruction compared to the 3 layer reconstruction by an amount ranging between 80 and 60 nm rms. This comparison reveals the impact of the model error in the tomography error term, even in our observing conditions with a dominant ground layer. Correcting with the 3 layer estimator brings an improvement of 164, 215 and 173 nm rms over the GLAO estimator for the three data sets. Asterism $\#$A47 $\#$A53 $\#$A12 ----------------------------------- ---------------------------------- --------------------------------- --------------------------------- Disengaged slopes hour [$23^\mathrm{h}59^\mathrm{m}$]{} [$3^\mathrm{h}14^\mathrm{m}$]{} [$6^\mathrm{h}02^\mathrm{m}$]{} Recorded frame number 28000 58000 28000 $r_0$ (cm) 14.1 9.7 10.6 Total turbulent rms error 1052 1139 1133 $\sigma_{Tomo+Model}$ (GLAO) 249 314 286 $\sigma_{Tomo+Model}$ (3 layers) 187 229 228 $\sigma_{Tomo+Model}$ (15 layers) 175 213 219 quadratic diff. (3-15 layers) 66 84 63 quadratic diff. (GLAO-3 layers) 164 215 173 : Three disengaged data sets for the three asterisms and the corresponding tomographic error computed with a ground layer, 3 or 15 layer tomographic estimator (in nm rms).[]{data-label="TomoSky"} Open loop error --------------- Open loop error is represented as a function of $r_0$ in Fig. \[OLVSr0\]. Open loop error estimation requires *engaged slopes* to be processed with *disengaged slopes* taken together within a short time scale. Only 16 sets of data are available with engaged and disengaged slopes close enough (in time) to compute a valuable error budget during the night. We estimate that open loop error ranged between 160 and 50 nm rms during the night. Because of the method used to estimate this error (computed from the quadratic difference of numerous estimated error terms), we cannot reliably draw conclusions about the behavior with respect to $r_0$. The open-loop error (or go-to error) was measured on the bench as typically lower than 5% of the total corrected wavefront [@kellererDM2012]. For the three data sets of Table \[TomoSky\] this gives an error lower than 35nm rms. ![Openloop error versus $r_0$.[]{data-label="OLVSr0"}](OLVSr0.pdf){width="8cm"} Field static aberration error ----------------------------- The measured field (quasi-static) aberrations versus the local time is presented in Fig. \[staticVStime\]. Residual static aberrations are very small $\approx $10 nm rms in the SCAO case (triangles). This value can be explained by the larger number of freedom measured by the TS than the number really controlled by the DM. The measured field aberrations in MOAO (circles) is 110 nm rms averaged over all the night. This value is quite large compared to other error terms and should not be underestimated while designing an MOAO instrument. Within this value, we estimated the order of magnitude of the contribution of the DM creep to be around 60nm rms. It also shows that MOAO requires a regular update of these field aberrations either by calibrating them on-sky (CANARY scheme) or by computing them using a model of the telescope aberrations. ![Wavefront error (in nm rms) due to static aberrations when measured in SCAO (triangles) and MOAO (circles) configurations versus local time. []{data-label="staticVStime"}](staticVStime.pdf){width="8cm"} Error budget comparison {#details} ======================= In this section we present the error budget computed on-sky at 3 different times during the night, each made on a different asterism in order to characterize the performance on 3 different geometrical and atmospherical conditions. Each case requires both a disengaged and a sequential engaged data set, in order to be able to perform the full error budget. The 3 disengaged sets are the ones presented in Table \[TomoSky\] and described in the previous section. The hours for the different data sets are: Asterism Disengaged Engaged ---------- ---------------------------------- --------------------------------- A\#47 [$23^\mathrm{h}59^\mathrm{m}$]{} [$0^\mathrm{h}10^\mathrm{m}$]{} A\#53 [$3^\mathrm{h}14^\mathrm{m}$]{} [$3^\mathrm{h}24^\mathrm{m}$]{} A\#12 [$6^\mathrm{h}02^\mathrm{m}$]{} [$6^\mathrm{h}07^\mathrm{m}$]{} In order to compare to the on-sky results, we also perform three numerical simulations for the same conditions and we compute the corresponding error budget and IR images. We present the simulation parameters in the next section. Simulation parameters --------------------- We used the end to end adaptive optics simulation program called [*Yorick Adaptive Optics*]{} (YAO) written by [@YAO] to produce series of *engaged* and *disengaged* data sets of slopes for each of the 4 WFS (including TS). These data sets mimic the data that were acquired on-sky with CANARY. The YAO code was modified to perform a simulated MOAO correction, in particular using the tomographic estimator $M_{ct}$ (see Sect. \[implementation\]) and the open loop integrator scheme as described in Sect. \[Control\]. Finally, the 2 sets of data (*engaged* and *disengaged* slopes) produced by YAO are used as inputs to our on-sky data reduction software. We emphasize that we used the same data pipeline software to compute both the on-sky error budget and the simulated error budget in order to avoid any bias in the data reduction. YAO is able to simulate a multi-layer turbulence profile by generating independent phase screens for each layer. We have chosen the 15-layer profile deduced from the post-processed on-sky data (see Fig. \[23h59\], \[03h14\] and \[06h02\]) as the simulated $C_n^2(h)$. Simulated WFS positions in the field of view are the ones deduced from the measurement made during the on-sky observations by WFS position encoders within the Target Acquisition System. We simulate the on-sky WFS configuration with $7\times7$ sub-apertures and $16\times16$ pixels per sub-aperture. The centroids of spots are computed with the same centroiding method as that used on-sky, i.e. selecting the 10 brightest pixels in a sub-aperture and zeroing the others. The magnitudes of the guide stars are adjusted to match with the WFS noise variance measured directly from the on-sky data. The variance of the noise on the simulation dataset is therefore identical to the real noise encountered on-sky, with both measured as described in \[Noise\]. The sampling frequency is simulated at 150 Hz. In order to simulate a loop latency of 1.5 frames, we ran two numerical simulations to compute the bandwidth error at 1 and 2 frames delay (YAO only simulates latencies of an integer number of frames) and linearly interpolated the bandwidth error at 1.5 frames. We simulated 7000 iterations ($\approx$ 45 seconds at 150 Hz) to produce typical data sets on all the WFS. The estimation of the average wind speed was deduced from the on-sky datasets by measuring the FWHM of the temporal autocorrelation of disengaged slopes data, and we reproduced the same in the simulation by tuning the speed. The simulated wind speeds were set respectively to 6, 6 and 8 m s$^{-1}$ as measured from the datasets at [$0^\mathrm{h}10^\mathrm{m}$]{}, [$3^\mathrm{h}24^\mathrm{m}$]{} and [$6^\mathrm{h}07^\mathrm{m}$]{}. The loop gain is respectively set to 0.8, 0.6 and 0.6 corresponding to the values used in these 3 sets. The wind speed is kept the same for all the layers in the simulation. Other parameters of the simulation are identical to the CANARY configuration used on-sky (see Sect. \[instrument\]) and they are summarized in Table \[Simusparams\]. Asterism $\#$A47 $\#$A53 $\#$A12 --------------------------- ------------------------- ------------------------- ------------------------- frequency (Hz) 150 150 150 delay (frames) 1.5 1.5 1.5 simulated iterations 7000 7000 7000 total $r_0$ (cm) 16.3 10.0 13.0 $C_n^2(h)$ Fig. \[23h59\] right Fig. \[03h14\] right Fig. \[06h02\] right windspeed (m/s) 6 6 8 loop gain 0.8 0.6 0.8 nb pix/sub ap. $16\times16$ $16\times16$ $16\times16$ nb Brightest pix/sub ap. 10 10 10 WFS pixsize (/pix) 0.26 0.26 0.26 star magnitude (1,2,3,TS) \[10.2,11.7,11.4,12.5\] \[11.2,12.3,13.0,12.8\] \[13.2,13.8,11.9,10.8\] WFS noise (nm rms) \[70,127,108, 162\] \[99,142,187,176\] \[206,328,137,85\] RON (e$^-$) 0.5 0.5 0.5 On-sky error budget ------------------- In this section we discuss the on-sky error budget computed from *engaged slopes* listed at the beginning of Sect. \[details\]. We recall that the *engaged slopes* are used to compute, $r_0$, $\sigma_{ErrTS}$, $\sigma_{tomo\ Noise}$, $\sigma_{AliasAlt}$, $\sigma_{Alias_{Ground}}$, $\sigma_{AliasAltTS}$, $\sigma_{BW}$, $\sigma_{noiseTS}$, $\sigma_{Fit}$, and $\sigma_{Field\_stat}$. The important term $\sigma_{Tomo}$ is missing to this list: we need sets of *disengaged slopes* recorded a few minutes before the engaged ones to determine it (using Eq. \[tomoeq\]) and we then rescale it according to the $r_0$ value found on the disengaged set. Then we use Eq. \[ErrOL\] to estimate the open loop term $\sigma_{OL}$. Finally we estimate the overall error budget $\sigma_{ErrIR}$ on the IR camera using Eq. \[IRTotal\]. IR images (30 s exposure) were also recorded at a time very close to the engaged slopes. Results are presented in columns labelled *on-sky* in Table \[err\] and split in three parts, each for their respective couple of data sets of engaged and disengaged slopes. - The tomographic error (including the turbulence model error) is estimated on-sky at $\sigma_{Tomo}$ = 156, 219 and 188 nm rms respectively on asterisms A47, A53 and A12. Those 3 values are those already given in Table \[TomoSky\], but now rescaled with the $r_0^{-5/3}$ of the engaged sets. We recall they were computed using the 3-layer tomographic estimator used on-sky (Figs. \[23h59\], \[03h14\], and \[06h02\] on the left for the corresponding profiles). - The estimated open loop error ($\sigma_{OL}$) ranges between 55 and 140 nm rms. This error term also includes all the other error terms not identified in the error budget. - Noise propagated through the estimator ($\sigma_{TomoNoiseFilt}$, Eq. \[TomoNoisefilt\]) ranges between 48 and 97 nm rms. In spite of the large noise terms measured on the individual WFSs (up to 328 nm rms, see Table \[Simusparams\]), the noise actually injected in the loop is still one of the weakest error terms. This is explained firstly because we take advantage of multiple wavefront sensing directions allowing the averaging of the noise on the 3 off-axis directions. Secondly, the tomographic estimator takes into account the average noise variance in each sub-aperture (diagonal of the covariance matrix $C_{OffOff}$) and consequently deals optimally with a noisy WFS. Finally we use a temporal controller to filter part of the noise out. - We estimate the aliasing contribution at the ground ($\sigma_{AliasGround} $) at 83 to 132 nm rms, and in altitude ($\sigma_{AliasAlt}$) at 15 to 28 nm rms. Because of the predominance of the turbulence at ground level in the $C_n^2$ profile (never less than 75%), the ground layer contribution to the aliasing dominates. - The measured on-sky bandwidth error lies from 115 to 142 nm rms. The bandwidth error is the highest in A53 in particular because the loop gain was set only to 0.6 instead of 0.8 for the other cases, and the seeing was the worst. - We estimate the fitting error at 138, 206 and 165 nm rms which makes it the second largest contributor to the error on CANARY after the tomography. - We observe a non negligible contribution (between 77 and 106 nm rms) of the field static aberrations to the error budget even when using our dedicated calibration procedure to extract such a term from the measurements. This contribution may come from the DM creeping effect and from the evolution of the field aberrations during the observations through field derotation or telescope flexures. - Best SR errors were measured on the bench with the calibration sources at 115 nm rms and were considered as fixed during the night of observation (see Sect. \[NCPA\]). Finally, the estimated total error budget for the IR camera ($\sigma_{ErrIR} $) gives 297, 419 and 357 nm rms. Converting these to Strehl ratio using the approximate formula $$\mathrm{SR} \approx \exp(-(2\pi\sigma_{ErrIR}/ \lambda)^2)$$ at $\lambda = 1\,530$ nm, we find the estimated SR: 22.6%, 5.2% and 11.7%. Real IR images were recorded a few tens of seconds from the dataset of slopes. We measure on these images SR of 20.1%, 10.3% and 16.4% ($\pm$2%) respectively. The observed discrepancy is mainly due to the pessimistic estimation of the SR using the exponential formula from the error budget. This formula gives much better estimation when SR is higher, typically larger than 30%. We will come back to this point after the presentation of the simulated error budget. Asterism ------------------------------------------------- ----------- ---------- --------- --------- --------- --------- IR image hour Engaged slopes hour $r_0$ (cm) on-sky simul on-sky simul on-sky simul $\sigma_{Tomo}$ 156 161 219 220 188 201 $\sigma_{OL}$ 55 0 140 0 116 0 $\sigma_{TomoNoiseFilt}$ 48 78 56 79 97 156 $\sigma_{AliasGround} $ 95 95 132 132 83 83 $\sigma_{AliasAlt}$ 15 15 28 28 22 22 $\sigma_{BW}$ 115 101 142 162 128 145 $\sigma_{Fit} $ 138 138 206 206 165 165 $\sigma_{Field Stat} $ 77 0 106 0 72 0 $\sigma_{NCPA}$ 115 0 115 0 115 0 [ $\sigma_{ErrIR} $ ]{} [ 297 ]{} [ 265]{} [419]{} [376]{} [357]{} [347]{} $\exp(-(2\pi \sigma_{ErrIR}/\lambda)^2)$ ($\%$) 22.6 30.6 5.2 9.1 11.7 13.1 IR Image SR ($\%$) 20.1 33.3 10.3 16.5 16.4 18.4 [ Ratio]{} 1.12 0.92 0.50 0.55 0.71 0.71 Simulated error budget ---------------------- In this section we discuss the simulated error budget (columns *simul* in Table \[err\]). - We computed a 3-layer tomographic estimator, directly using the parameters of the simulation. The tomographic error found while using it is $\sigma_{Tomo} =$ 161, 220 and 201 nm rms. These values are in good agreement with the on-sky results 156, 219 and 188 nm rms. We recall that although the estimator is based on 3 layers, we introduced a 15 layers $C_n^2(h)$ profile in the simulated model of the atmosphere. Therefore in this particular case a representative tomographic model error was simulated. Table \[TomoSimu\] compares the performance of the 3-layer to the 15-layer estimator within the numerical simulation. The performance when correcting 15 layers gives 128, 161 and 157 nm rms, a significant reduction of the error. As the reconstructed profile perfectly matches the simulated one, it represents the pure tomographic error $\sigma_{PureTomo}$ without any turbulence profile modelling error contribution ($\sigma_{Model}$). The latter can be deduced by quadratically subtracting the tomographic performance of 15-layer estimator from the 3-layer one. This gives an estimation of the turbulence modeling error of $\sigma_{Model}$= 99, 150, 126 nm rms. These numbers can be directly compared with those in Table \[TomoSky\], where we performed the same analysis with two estimators based on 3 or 15 layers, but with the on-sky data. The error $\sigma_{Model}$ is roughly a factor of 2 higher in the simulation which means we are only able with the simulation to estimate the order of magnitude of the model error that impact our observational results. In any case, identifying only three layers to compute the estimator was clearly not sufficient and leads to around 120 nm of additional error. This significant contribution to the error budget underlines the requirement to increase the number of layers to be considered in the $C_n^2(h)$ profile. $\#$A47 $\#$A53 $\#$A12 --------------------------------- --------- --------- --------- $\sigma_{Tomo}$ (3 layers) 161 220 201 $\sigma_{PureTomo}$ (15 layers) 128 161 157 deduced $\sigma_{Model}$ 99 150 126 : Tomographic error measured from numerical simulations (in nm rms). The model error is deduced by subtracting the 3 layer performance to the 15 layer one.[]{data-label="TomoSimu"} - The simulated noise injected in the loop $\sigma_{TomoNoiseFilt}$ is respectively 78, 79 and 156 nm rms. Notice here that we reproduced the same amount of noise per WFS in the slope measurements with respect to the one measured on-sky. Nevertheless, the noise propagated by the tomographic estimator is higher in simulations than on-sky. This difference is explained by the way the tomographic estimator was computed in the simulations. We did not include the slopes noise variance in the diagonal of the $C_{OffOff}$ covariance matrix to compute the tomographic estimator. Therefore the simulated tomographic estimator fully propagates the noise, even though the noise is partially filtered by the integrator loop. - For the bandwidth error, we find $\sigma_{BW}=$101, 162 and 145 nm rms with the considered conditions. The $\sigma_{BW}$ computed by simulation is close to the on-sky values. - As we use an analytical expression based on the value of $r_0$ to derive both the aliasing and fitting error terms, we have reported the same number in the on-sky and simulations cases. - Open loop error, NCPA and field static aberrations errors were not simulated. The total error expected on the IR image $\sigma_{ErrIR}$ from the simulated set of slopes is 265, 376 and 347 nm rms. The expected SR from the simulated slopes is therefore 30.6%, 9.1% and 13.1% and must be compared to the SR measured on the simulated IR image which gives: 33.3%, 16.5% and 18.4%. Here the global performance is better because some of the on-sky contributors were not considered in the simulation. Relative error budget --------------------- Fig. \[RelErrBudget\] shows the on-sky relative error budget (in %) computed by the quadratic sum of the individual error terms presented in Table \[err\]. - The fitting error counts for less than a quarter of the total error budget. This error is irreducible in CANARY as it is fixed by the number of actuators on the DM. - The tomographic error is the dominant term and its relative value remains remarkably stable at $\approx$ 27% of the total error. This is the level of performance we expected from the initial simulations. Despite the change of guide star configuration and magnitudes, the estimator performs relatively well. $C_n^2(h)$ conditions were mainly dominated by a strong ground layer but MOAO provides better performance than GLAO. This demonstrates the robustness of our tomographic approach based on Learn and Apply. We emphasize again that because it is an on-sky measured term, it also includes the turbulence profile model error. - The next largest term in weight is the bandwidth error. It ranges between 11.5% and 15$\%$. It is significantly larger than the noise contributor, showing that a higher sampling frequency could have been used. - Noise injected in the loop contributes only a few percent of the total error ($\approx$ 2 to 7 $\%$) despite quite a large noise level in the individual WFS measurements (see Table \[Simusparams\]). - The open loop term, also containing other unknown errors, ranges from $\approx$ 3 to 11 $\%$ of the total error. This term is at the level of the other contributors listed below. This a satisfactory demonstration that open loop operation works. - The aliasing relative contribution ranges between 6 and 10 $\%$ of the total error. - The NCPA has a fixed value of 115 nm rms but its relative value evolves from 7.5$\%$ (strongest turbulence) to 14.7$\%$ (weakest turbulence). - The field static aberrations lead to a 4 to 6$\%$ relative error. This term was not expected at the beginning to have such a large contribution. It is clear that it will have to be properly mitigated in any MOAO system. ![Relative representation of the CANARY error budget at [$0^\mathrm{h}10^\mathrm{m}$]{}, [$3^\mathrm{h}24^\mathrm{m}$]{} and [$6^\mathrm{h}07^\mathrm{m}$]{}. Areas of the chart are proportional to variances.[]{data-label="RelErrBudget"}](piechartA47.pdf "fig:"){width="2.8cm"} ![Relative representation of the CANARY error budget at [$0^\mathrm{h}10^\mathrm{m}$]{}, [$3^\mathrm{h}24^\mathrm{m}$]{} and [$6^\mathrm{h}07^\mathrm{m}$]{}. Areas of the chart are proportional to variances.[]{data-label="RelErrBudget"}](piechartA53.pdf "fig:"){width="2.8cm"} ![Relative representation of the CANARY error budget at [$0^\mathrm{h}10^\mathrm{m}$]{}, [$3^\mathrm{h}24^\mathrm{m}$]{} and [$6^\mathrm{h}07^\mathrm{m}$]{}. Areas of the chart are proportional to variances.[]{data-label="RelErrBudget"}](piechartA12.pdf "fig:"){width="2.8cm"} Discussion ---------- We would like to emphasize that the tomography terms measured on-sky match remarkably well with their corresponding numerical simulations. This is probably one of the most important results of the on-sky demonstration made by CANARY at phase A. Such a comparison makes us confident in establishing the error budget of future instruments and also in the ability to process the instrument data to properly retrieve the $C_n^2(h)$ profile required for optimisation of the on-axis SR. The other main specific term of an MOAO instrument is the open loop error. From previous laboratory measurements a value of 3 to 5% of the total turbulence perturbation was expected [@kellererDM2012], leading here to 30 to 70 nm rms in our observing conditions. We measure on-sky values ranging from 50 to 160 nm rms. The difference may come from the way it has been evaluated. We measure it as the difference between the total error measured by the TS and the quadratic sum of all the error terms. Therefore the computed value can be influenced by any error in the estimation of the other terms. It contains the open loop error plus possibly all others unidentified errors such as drift in pupil alignment, badly seen turbulence on the bench etc... Only one term has been identified on-sky that was not included in the initial simulations of CANARY ; the quasi-static field aberrations residuals. They are of the order of 70 to 130 nm rms and are very difficult to calibrate properly. For us, this term includes the error linked to the mis-calibrated field dependent aberrations and the creep of the DM linked to the static voltage offsets applied to it. On the E-ELT, those static field aberrations will also not be fully static, but evolve slightly because of the field derotation and the residual errors in the active optics. This will bring a problem to all tomographic techniques and thus will have to be solved for the E-ELT instruments. Particular effort on new calibration schemes should be envisioned to reduce this term and we emphasize here the importance of the Truth Sensor for such calibrations in CANARY. The computed SR from the error budget shows an under-estimation of the measured SR on the IR camera. We explain the difference by several factors. Firstly, the $\exp(-(2\pi \sigma_{ErrIR}/\lambda)^2)$ formula tends to be pessimistic for SR below 30%. We also observe this effect in the simulations. The simulations allow us to quantify the mismatch between the measured image SR and the exponential formula. The ratio of the IR image SR and the exponential formula is very similar for the simulations and the on-sky data as given in Table \[err\]. Secondly, we estimate the error bars of the total error budget to be $\pm 50$ nm rms leading to $\pm$4.0% error on the predicted SR. Finally, we notice that $r_0$ fluctuated significantly (typically $\pm$2 cm in a few tens of seconds) between the recordings of the engaged slopes, the disengaged ones and the IR images. Unfortunately all the data were not perfectly synchronised. This may explain some discrepancies in Table \[err\]. Taking into account all these facts, we can say that the proposed error budget is validated. Conclusion ========== CANARY is a single channel MOAO demonstrator installed at the William Herschel Telescope. It obtained the first on-sky MOAO compensated images in September 2010. In this paper, we detail the calibration procedures used in CANARY, in particular to compute the optimized tomographic estimator required in MOAO and its software implementation. We present a method to process the slope data recorded on-sky and apply this to data taken during the night of the 27th of September 2010. The presence of the Truth Sensor allowed us to build a full error budget for the instrument. We evaluate 9 error terms including the tomographic and the open loop error, which are new terms introduced by the MOAO control scheme. The tomographic error is estimated from synchronous disengaged slopes while the open loop error also requires data taken when the MOAO loop is engaged. We have also computed other classical AO error terms like noise, aliasing, bandwidth, fitting and NCPA. For the error budget, we distinguish two terms in the aliasing due to both the ground and the high altitude layer contributions. We check and successfully compare the measured on-sky error budget with numerical simulations for 3 datasets differing in guide star configuration, magnitudes and atmospheric conditions. The conclusion is encouraging, with the open-loop term, although larger than expected, being small compared to the other terms, and the tomographic term behaving as expected from simulation. But we identify on-sky an additional term linked to the residual quasi-static field aberrations found to be of the order of the open-loop term. The turbulence profile derived from the instrument data is presented and used in the tomographic optimization. The optimized [*Learn & Apply*]{} tomographic estimator based on 3 layers computed during the on-sky operations performed at any time better than a basic ground layer reconstruction. The average performance of the 3 layer tomographic reconstruction was 216 nm rms while ground layer reconstruction performed at 270 nm rms representing an increase in the performance of 162 nm rms on average during the night. We are also able to evaluate the impact of the $C_n^2$ profile model error in the tomographic error by testing (in simulation) estimators computed with 3 or 15 layers. Although the tomographic error is the largest contributor to the error on CANARY, it remained stable contributing one quarter of the total error budget. The Strehl ratio on the IR camera ($\lambda=1\,530$ nm) was measured and compared to the estimated error budget. The agreement was good. In MOAO mode we measured SR in the range of 10% to 20% with a typical r$_0$ value between 10 to 16 cm. Comparison with the numerical simulation makes us confident in the on-sky error budget estimation. Therefore this paper brings some important insights for the establishment of the wavefront error budget of future MOAO instruments. CANARY represents a significant advance in the implementation of the future tomographic AO systems like MOAO, LTAO and even MCAO instruments. It has successfully demonstrated the tomographic optimization in a direction of interest and the accuracy of the novel open loop control scheme. The next phase of CANARY uses 4 Rayleigh Laser Guide stars in addition of the 3 natural guide stars. This phase also implements a modification of the Learn & Apply algorithm to perform with mixed measurements of natural and laser guide stars. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by Agence Nationale de la Recherche (ANR) program 06-BLAN-0191, CNRS / INSU, Observatoire de Paris, and Université Paris Diderot Paris 7 in France, Science and Technology Facilities Council, University of Durham and the Isaac Newton Group of telescopes in UK and European Commission Framework Programme 7 (E-ELT Preparation Infrastructure Grant 211257 and OPTICON Research Infrastructures Grant 226604). Appendix ======== White noise propagation through a first-order low-pass filter {#append_g/2-g} ------------------------------------------------------------- The expression of the first-order low-pass filter is $$\label{lowpass} y_n = (1-g)y_{n-1} + g x_n\ .$$ We want to determine the variance $\sigma_y^2$ of the output $y_n$, when $x_n$ is a white noise of known variance $\sigma_x^2$. We have $$y_n^2 = (1-g)^2 y_{n-1}^2 + g^2 x_n^2 + 2g(1-g)y_{n-1}x_n\ .$$ As $y_{n-1}$ and $x_n$ cannot be correlated, it follows that $$\sigma_y^2 = (1-g)^2 \sigma_y^2 + g^2 \sigma_x^2$$ or finally $$\label{cqfdg2g} \sigma_y^2 = \frac{g}{2-g} \sigma_x^2\ .$$ White noise propagation through a first-order low-pass filter with fractional shift {#append_gaga} ----------------------------------------------------------------------------------- We let $y_n$ be a signal resulting from the low-pass filtering described in Eq. \[lowpass\], of a white noise $x_n$ of variance $\sigma_x^2$. We consider the signal $s_n$ deduced from $y_n$ by a fractional delay: $$s_n = \alpha y_{n-2} + (1-\alpha) y_{n-1}\ .$$ We can replace $y_{n-1}$ in the above equation with its expression in terms of $y_{n-2}$ using Eq. \[lowpass\], and compute the square of $s_n$: $$\begin{aligned} s_n^2 & = & (\alpha y_{n-2} + (1-\alpha) y_{n-1})^2 \\ & = & ( \alpha y_{n-2} + (1-\alpha) (1-g) y_{n-2} + (1-\alpha) g x_{n-1})^2 \\ & = & ((1-g+\alpha g)y_{n-2} + (1-\alpha) g x_{n-1})^2 \\ & = & (1-g+\alpha g)^2 y_{n-2}^2 + (1-\alpha)^2 g^2 x_{n-1}^2 \\ & & + x_{n-1} y_{n-2} (1-g+\alpha g) (1-\alpha) g \ .\end{aligned}$$ Considering that $x_{n-1}$ and $y_{n-2}$ are not correlated, and using Eq. \[cqfdg2g\] for expressing the variance of $y_{n-2}$, it follows that $$\begin{aligned} \sigma_s^2 & = & \left( (1-g+\alpha g)^2 \frac{g}{2-g} + (1-\alpha)^2 g^2 \right) \sigma_x^2 \\ & = & \frac{g}{2-g} (1 - 2 g \alpha (1-\alpha)) \, \sigma_x^2\ . \label{cqfdag}\end{aligned}$$ The re-arrangement of terms as in Eq. \[cqfdag\] makes it clear that the result is symmetric between $\alpha$ and $(1-\alpha)$. Values of $\alpha$ of either 0 or 1 (integer shift) reduce the expression to the previous case, with a filtering coefficient of $g/(2-g)$. The particular case of $\alpha=0.5$ leads to $\sigma_s^2 = ( g/2) \, \sigma_x^2 $. Now, the variance of the difference between $x_n$ and $s_n$ can easily be deduced from the previous calculus, because $x_n$ and $s_n$ cannot be correlated. The variance of their difference is just equal to $$\sigma_{x-s}^2 = \left(1+\frac{g}{2-g} (1 - 2 g \alpha (1-\alpha)) \right) \, \sigma_x^2$$
{ "pile_set_name": "ArXiv" }
--- abstract: 'Transport properties provide important access to a solid’s quasiparticles, such as quasiparticle density, mobility, and scattering. The transport of heat can be particularly revealing because, in principle, all types of excitations in a solid may contribute. Heat transport is well understood for phonons and electrons, but relatively little is known about heat transported by magnetic excitations. However, during the last about two decades, the magnetic heat transport attracted increasing attention after the discovery of large and unusual signatures of it in low-dimensional quantum magnetic cuprate materials. Today it constitutes an important probe to otherwise often elusive, topological quasiparticles in a broader class of quantum magnets. This review summarizes the experimental foundation of this research, i.e. the state of the art for the magnetic heat transport in the mentioned cuprate materials which host prototypical low-dimensional antiferromagnetic $S=1/2$ Heisenberg models. These comprise, in particular, the two-dimensional square lattice, and one-dimensional spin chain and two-leg ladder spin models. It is shown, how studying the heat transport provides direct access to the thermal occupation and the scattering of the already quite exotic quasiparticles of these models which range from spin-1 spin wave and triplon excitations to fractionalized spin-1/2 spinons. Remarkable transport properties of these quasiparticles have been revealed: the spin-heat transport often is highly efficient and in some cases even ballistic, in agreement with theoretical predictions.' address: - 'IFW Dresden, 01069 Dresden, Germany' - 'Center for Transport and Devices, TU Dresden, 01069 Dresden, Germany' author: - Christian Hess title: 'Heat transport of cuprate-based low-dimensional quantum magnets with strong exchange coupling' --- heat transport,quantum magnetism ,low-dimensionality ,experiment ,cuprates Introduction {#sec:intro} ============ The heat transport properties of a solid provide important information about the generation, mobility, scattering, and dissipation of various kinds of excitations, such as electrons, phonons and magnons. Deep fundamental knowledge in this regard exists about the heat transport by phonons and electrons [@Berman], yet very little is known about heat transport by magnetic excitations. After the original prediction of magnetic heat transport in 1936 by Fröhlich and Heitler [@Froehlich36], it took almost 30 years until the first convincing experimental evidence for heat transport by classical spin waves was found in ferrimagnetic yttrium-iron-garnet (YIG) [@Luethi62; @Douglass63; @Rives69; @Walton73]. Analogous to the case of phononic and electronic heat conduction, the analysis of this magnon heat conductivity is intriguing because it should in principle yield valuable information about the excitation and scattering of magnetic excitations (e.g. off defects, phonons, and electrons). However, most of the pioneering experiments on YIG and following experiments on other materials [@Gorter69; @Lang77; @Coenen77] focused on the mere identification of the new heat transport phenomenon. Furthermore, these early experiments typically were restricted to spin waves emerging from magnetically ordered phases at very low temperature ($T<10~\mathrm{K}$). The first signature of magnetic heat transport at higher temperatures ($T>50~\mathrm{K}$) was observed for the one-dimensional quantum antiferromagnet $\mathrm{KCuF_3}$ [@Hirakawa75], remarkably *above* its Néel temperature of $T_N\approx38~\mathrm{K}$ [@Hutchings1969]. These very intriguing first results for the heat transport of a low-dimensional *quantum magnet* remained for quite some time largely unnoticed. However, the rigorous theoretical prediction of dissipationless (so-called *ballistic*) heat conduction in one-dimensional antiferromagnetic spin-1/2 chains [@Zotos97] and the discovery of huge magnetic contributions in quantum two-leg spin-1/2 ladder compounds [@Kudo1999; @Sologubenko00; @Hess01] triggered intense research on the heat transport of low-dimensional quantum spin systems both experimentally [@Kudo1999; @Sologubenko00; @Kudo01; @Hess01; @Hess02; @Hess03; @Hess04a; @Hess05; @Sologubenko00a; @Sologubenko01; @Kudo2003; @Ando1998; @Vasilev1998; @Hofmann2002; @Hofmann2003; @Sologubenko03; @Sologubenko03a; @Kordonis06; @Hlubek2010a; @Steckel2016; @Hess2007b; @Hess2006; @Otter2012; @Montagnese2013; @Ribeiro2005; @Hess2007; @Hlubek2010; @Hlubek2012; @Hlubek2011; @Mohan2014] and theoretically [@Heidrich02; @Heidrich03; @Orignac03; @Alvarez02; @Heidrich04; @Gros2004; @Saito2003; @Saito2003a; @Shimshoni03; @Kluemper2002; @Sakai2003; @Zotos04; @Karadamoglou04; @Louis2006; @Li2002; @Li2003; @Rozhkov05; @Chernyshev05; @Jung06; @Boulat2007; @Steinigeweg2016; @Chernyshev2016]. On the experimental side, more and more evidence for *unconventional* heat conduction in low-dimensional quantum magnets has been observed over the years in various materials. Today, the clearest and often surprising experimental findings are known for spin systems realized in copper oxides (cuprates). In these compounds, a substantial magnetic heat conduction is often observed, despite a pronounced quantum nature of their spin systems with inherent absence of long-range magnetic order or even short range spin-spin correlations. Their actual magnetic heat conductivity can be unambiguously detected, because if present, it is usually large even at room temperature and above, often dwarfing the phononic heat conductivity of the system and thereby sometimes reaches values which are comparable to the heat conductivity of a metal. The content of the review at hand is as follows: In the following paragraphs of this introductory Section \[sec:intro\], the essential basics of the spin models and the cuprate materials under scrutiny will be presented. Furthermore, the salient experimental signatures of magnetic heat conductivity in such systems will be presented, and the basics of analyzing the data will be introduced. This Section \[sec:intro\], i.e. the whole introductory part is based on [@Hess2007b]. Section  \[sec:spin\_planes\] and Section \[sec:ladders\] summarize the state of the art for the heat conductivity of the $S=1/2$ two-dimensional Heisenberg antiferromagnet on a square lattice (2D-HAF) and of $S=1/2$ two-leg spin ladders, respectively. Finally, Section \[sec:chains\] addresses the heat transport and related results on the spin dynamics of $S=1/2$ Heisenberg spin chains. Spin models ----------- We are considering $S=1/2$ models with Heisenberg interaction $$\label{Hamleit} {\cal H}=J_{i,j}\sum_{\langle i,j \rangle}{\bf S_i\cdot S_j}\,,$$ where the sum runs over all nearest neighbors in the system. Here we are investigating three different spin models: the 2D-HAF, the two-leg spin ladder, and spin chains. For the 2D-HAF and the spin chain $J_{i,j}=J$ and for the spin ladder $J_{i,j}=J_\Vert$ along the legs and $J_{i,j}=J_\bot$ along the rungs of the ladder. The corresponding low-dimensional quantum spin models are characterized by very peculiar ground states which are quantum disordered in the one-dimensional chain and ladder models. The elementary excitations which emerge from this ground states bear therefore a quite exotic character. For instance, the spin-spin correlations of $S=1/2$ Heisenberg spin chains, decay algebraically with distance between the spins [@Kluemper1993]. The elementary excitations are nevertheless well defined. They are fractionalized, i.e. $\Delta S=1$ spin-flip excitations of the system decay into so-called spinons which are gapless and carry a spin $S=1/2$ [@Faddeev1981]. On the other hand, the ground state of a two-leg ladder possesses more short-range spin-spin correlations which decay exponentially as a function of distance [@Dagotto96; @Dagotto99]. The elementary excitations are $S=1$ particles (usually called magnons or triplons) and are separated from the ground state by a spin gap $\Delta$ ($\Delta/k_B\approx400~\mathrm{K}$ in the case of the systems discussed here) [@Dagotto99]. Finally, the ground state of the 2D-HAF is a rather classical long-range ordered Néel state. However, also here due to incipient quantum disorder it only exists at temperature $T = 0$ and possesses a strongly reduced sublattice magnetization [@Manousakis91]. In this case the elementary excitations are well described using a spin wave framework where one should keep in mind that alternative descriptions have been discussed, too [@Coldea01; @Sandvik2001; @Ho2001]. In the case of hole doping, all these model systems yield interesting and exotic properties. A Luttinger liquid forms in hole-doped spin chains. Here the electronic excitations decay into collective excitations of holes (holons) and spins (spinons). This phenomenon is often referred to as spin-charge separation. Quite different properties have been predicted for two-leg spin ladders: superconductivity competing with a charge ordered ground state is expected in this case [@Dagotto92; @Dagotto96]. Finally, hole doping has great importance in the case of the 2D-HAF, because the interaction of the doped hole with the antiferromagnetic background forms a new quasiparticle, the spin polaron, which can be understood as a hole dressed with characteristic spin fluctuations [@Martinez1991]. The spin polaron is thought of playing a crucial role for the emergence of exotic ground states upon doping, including high-temperature superconductivity, which is observed in such systems [@Chernyshev1999; @Lee2006]. Note that the observation of spin-charge separation signatures upon controlled hole doping of $S=1/2$ Heisenberg chain materials has not yet been achieved experimentally, Nevertheless, signatures of it have been derived e.g. from angular resolved photo emission experiments on undoped chain compounds [@Koitzsch2006]. On the other hand, charge ordering and superconductivity are prominent experimental features of hole-doped spin ladder and 2D-HAF materials [@Vuletic2006; @Lee2006]. ![Illustration of low-dimensional spin structures: (a) a spin chain, (b) a two-leg spin ladder, and (c) a two-dimensional square lattice. Arrows represent localized electrons with spin $S=1/2$ spin and shaded bars symbolize strong antiferromagnetic exchange between them. (d) Schematic illustration of the underlying chemical building block giving rise to the localized spins and their interaction. Only the relevant [Cu]{}-$3d_{x^2-y^2}$ and [O]{}-$2p_x$ orbitals are indicated. Arrows represent the spins of the electrons involved. Figure reproduced from [@Hess2007b].[]{data-label="fig:1"}](fig1_neu){width="\textwidth"} Materials {#intromat} --------- The materials at focus of this review are cuprate compounds which indeed host spin arrangements in the geometrical form of chains, two-leg ladders, and square lattices with a strong antiferromagnetic Heisenberg exchange ($J/k_B\approx 1500-2000~\mathrm{K}$) between nearest neighbor spins. Sketches of such spin arrangements are shown in Fig. \[fig:1\]a-c. These low-dimensional arrangements of interacting spins arise from similarly low-dimensional structures composed of [Cu-O-Cu]{} bonds, within which a dominant antiferromagnetic exchange is present if these bonds are straight, i.e. a bonding angle of $180^\circ$ is realized as depicted in Fig. \[fig:1\]d. In all the considered cuprate systems the spins have $S=1/2$, resulting from the $3d^9$ configuration of $\mathrm{Cu}^{2+}$-ions and therefore possess generally a strong quantum nature. Good realizations of $S=1/2$ Heisenberg chains as depicted in Fig. \[fig:1\]a are found in the compounds [$\mathrm{CaCu_{2}O_{3}}$]{}, [$\mathrm{SrCuO_{2}}$]{}, and [$\mathrm{SrCuO_{2}}$]{}, where chains of straight [Cu-O-Cu]{} bonds and hence a strong antiferromagnetic exchange exists along one particular crystallographic direction only; the magnetic exchange perpendicular to this direction is much weaker [@Kiryukhin01; @Motoyama1996]. Two-leg spin ladders are realized (cf. Fig. \[fig:1\]b) in the $\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$ family of compounds, where parallel pairs of such [Cu-O-Cu]{} chains (the ladder legs) are coupled to each other via bridging O-ions, producing in straight [Cu-O-Cu]{} bonds perpendicular to the chains direction (the ladder rungs), where the interchain coupling (or rung coupling) perpendicular to the chain direction $J_\perp$ is of a similar magnitude to the intrachain coupling (or leg coupling), i.e. $J_\perp\approx J$ [@Dagotto99]. Ladder structures with more legs can in principle be created by coupling more chains to the structure; eventually, this would lead to a 2D-HAF in the infinite limit. A good realization of a 2D-HAF with $S=1/2$ is given by [$\mathrm{La_2CuO_4}$]{}and other related antiferromagnetic parent compounds of high-temperature superconductors. ![Anisotropic thermal conductivity of various low-dimensional spin materials as a function of temperature: a) the 2D-HAF as realized in [$\mathrm{La_2CuO_4}$]{}, b) the two-leg spin ladder material [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}, and c) the spin chain compound [$\mathrm{CaCu_{2}O_{3}}$]{}. Filled and open symbols represent [$\kappa_\Vert$]{}and [$\kappa_\bot$]{}of the materials. The solid line in a) represents a linear fit to the data in the range $T>100\mathrm~K$. The axis intercept of its extrapolation towards $T = 0$ is an approximation of [$\kappa_\mathrm{ph}$]{}in the fit range. Solid lines in b) and c) represent estimations for the phonon background. Figure adapted from [@Hess01; @Hess03; @Hess2007].[]{data-label="fig:signatures"}](fig_signatures){width="\textwidth"} Experimental signatures of magnetic heat transport {#sec:signatures} -------------------------------------------------- The typical experimental signature of low-dimensional magnetic heat transport in a given material is a very anisotropic heat conductivity tensor of the material which reflects the dimensionality of the spin system. The origin of the anisotropy is the magnetic heat conduction of the low-dimensional magnetic structures which add to the always present phononic heat conduction. In principle, electronic heat conduction could occur as well. However, for all materials discussed here, such a contribution is irrelevant because the electrical conductivity is very low. The experimental method which has been employed for obtaining the here discussed data is, in the largest portion of cases, the so-called standard steady-state method [@Berman] which is proven to be very reliable for obtaining the different components of the heat conductivity tensor. An exception concerns additional results on the two-leg spin ladder materials, where fluctuations in the magnetic signal provided the motivation for dynamic heat transport studies (Section \[sec:ladders\]). Fig. \[fig:signatures\] shows selected corresponding examples for experimental results of the heat conductivity $\kappa$ of [$\mathrm{La_2CuO_4}$]{}, [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}and [$\mathrm{CaCu_{2}O_{3}}$]{}, which are good representatives of the 2D-HAF, the two-leg spin ladder, and the isotropic antiferromagnetic Heisenberg chain [@Hess03; @Hess01; @Hess2007b; @Hess2007]. The data of the heat conductivity perpendicular to the low-dimensional magnetic structures ($\kappa_\bot$), i.e. along the directions with negligible magnetic exchange interaction, the temperature dependence of $\kappa$ is characteristic of conventional *phononic* heat conduction [$\kappa_\mathrm{ph}$]{}[@Berman]. More specifically, $\kappa(T)$ has a low-temperature peak centered at temperature $T\approx20\dots30~\mathrm{K}$, with a continuously decaying high-temperature edge. An exception is found for [$\mathrm{CaCu_{2}O_{3}}$]{}, where one component of [$\kappa_\bot$]{}increases monotonically. Such an increase is well known for systems with a strongly suppressed [$\kappa_\mathrm{ph}$]{}due to disorder [@Hess2007; @Hess04]. Completely different characteristics are found if $\kappa$ is measured parallel to the low-dimensional structures, i.e. along the directions with a large $J$ ($\kappa_\Vert$). Also in these cases, signatures of a phononic low-temperature peak are present. However, upon increasing the temperature further, [$\kappa_\Vert$]{}is very different from [$\kappa_\bot$]{}. In all three cases, [$\kappa_\Vert$]{}strongly increases for $T\gtrsim60~\mathrm{K}$ ($T\gtrsim90~\mathrm{K}$ in the case of [$\mathrm{La_2CuO_4}$]{}) with increasing temperature, and in the case of [$\mathrm{La_2CuO_4}$]{}and [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}, a high-temperature peak at $\mathrm 310$ K and $\mathrm 140$ K, respectively, is formed, while for [$\mathrm{CaCu_{2}O_{3}}$]{}the increase continues up to room temperature. This remarkably clear anisotropy of $\kappa$ is the clear qualitative evidence for large magnetic contributions to [$\kappa_\Vert$]{}of these three materials, i.e. a large [$\kappa_\mathrm{mag}$]{}of the low-dimensional spin structures. Generally, in the shown examples, [$\kappa_\mathrm{mag}$]{}can be well extracted from the measured data for [$\kappa_\Vert$]{}by subtracting the phononic contribution [$\kappa_\mathrm{ph}$]{}, which turns out to be reasonably approximated by the experimentally obtained [$\kappa_\bot$]{}. This is the essential basis for analyzing [$\kappa_\mathrm{mag}$]{}in order to yield information on the thermal generation and the scattering of the magnetic quasiparticles. Modeling {#kinmod} -------- In theoretical works, the attention often is focused on the possibility of ballistic magnetic heat transport in 1D-systems: in integrable models like the XXZ Heisenberg spin chain the Hamiltonian and the thermal current operator commute, i.e. once a thermal current is established in such a system, it will never decay [@Zotos97]. In other words, the thermal resistance vanishes and the magnetic thermal conductivity $\kappa_\mathrm{mag}$ diverges. While such surprising properties are well established for integrable spin models [@Zotos97], ballistic heat transport in non-integrable quasi 1D-systems (e.g. two-leg spin ladders) has been a subject of intense discussion [@Alvarez02; @Zotos04; @Heidrich04; @Boulat2007; @Steinigeweg2016]. However, in real materials scattering processes involving defects and other quasiparticles such as phonons and charge carriers must play an important role and render $\kappa_\mathrm{mag}$ finite in all cases [@Shimshoni03]. The analysis of $\kappa_\mathrm{mag}$ should hence provide further insight into the nature of these scattering processes and the dissipation of magnetic heat currents. ### Kinetic model {#kinetic-model .unnumbered} We set up a kinetic model which should be able to capture the most important features of the magnetic heat conductivity [@Hess2007b]. Apparently, the qualitative temperature dependence of $\kappa_\mathrm{mag}$ often is a simple peak structure ([$\mathrm{La_2CuO_4}$]{}and [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}) or a monotonic increase ([$\mathrm{CaCu_{2}O_{3}}$]{}) in the studied range $T=100\dots350~\mathrm K$, where the latter may be regarded as the low temperature edge of a peak. A peak structure is very common for the temperature dependence of the thermal conductivity $\kappa$ of any kind of heat carrying particle, such as phonons or electrons [@Berman]. In principle, one can expect that the same kinetic considerations which successfully describe the physics for phononic and electronic heat transport, can be applied to magnetic excitations as well. The basic physics which determines the $T$-dependence of $\kappa$ can be inferred from the kinetic estimate [@Ziman] $$\kappa=\frac{1}{d}\frac{1}{(2\pi)^d}\int c_{\bf k}v_{\bf k}l_{\bf k}d{\bf k}, \label{eq:kinetic_general}$$ with $d$ the dimensionality of the considered system, $c_{\bf k}=\frac{d}{dT}\epsilon_{\bf k}n_{\bf k}$ the specific heat, $v_{\bf k}$ the velocity and $l_{\bf k}$ the mean free path of a mode with wave vector ${\bf k}$. $\epsilon_{\bf k}$ and $n_{\bf k}$ are the energy and the statistical occupation function of the mode $\bf k$. Here, we are interested in magnetic excitations, i.e., a crucial parameter which determines the thermal occupation is the thermal energy $k_BT$ in relation to the magnetic exchange energy $J$. For all systems discussed here $J/k_B\approx 1500\dots2000\mathrm{~K}$, whereas the experimental data only extend over temperatures $T < 350~\mathrm{K}$. Thus, all considerations discussed in this review concern, from the viewpoint of the magnetic system, the situation of low temperature $k_BT\ll J$. The momentum of these particles naturally is confined to a small vicinity of the minima of the dispersion function. One can therefore safely ignore the momentum dependence of the mean free path, i.e., $l\approxeq l_{\bf k}$. Furthermore, at such low $T$, only a few magnetic modes are excited and contribute to the heat transport. If the temperature is also well below other relevant energy scales of the solid, such as the Debye temperature, scattering processes different from defect and boundary scattering which change the crystal momentum are rare. In this situation, the low-$T$ increase of $\kappa$ is (a) characteristic of the excitation of the heat carrying particle (reflecting the $T$-dependence of the specific heat if $v_{\bf k}$ is momentum independent) and (b) proportional to the mean free path $l\approxeq l_{\bf k}$. At higher $T$, the momentum-dependent scattering becomes increasingly important and eventually leads to a decrease of the mean free path and hence to a decrease of $\kappa$. This decrease is characteristic of the relevant scattering mechanisms and allows an advanced analysis which potentially provides crucial information about these mechanisms. The application of Eq. \[eq:kinetic\_general\] for the case of 1D and 2D magnetic systems considered here leads to the general result $$\kappa_\mathrm{mag}(T)\propto l_\mathrm{mag} f(T)~,\label{kmag_simple}$$ where $l_\mathrm{mag}(T)$ is a general magnetic mean free path based on the approximation $l_\mathrm{mag}\equiv l_{\bf k}$. As mentioned afore, for $k_BT\ll J$ this assumption is justified for the large-$J$ systems considered here because the heat carrying excitations exist in significant numbers only in the vicinity of the band minima, i.e. a very small fraction of the Brillouin zone. The function $f(T)$ depends on temperature in a manner which is characteristic of the considered spin system. Details will be discussed in the respective sections. Spin planes {#sec:spin_planes} =========== Qualitative proof of magnon heat transport ------------------------------------------ As already indicated in section \[sec:signatures\], the prominent signature of the magnetic heat conductivity in the 2D-HAF is a pronounced high-temperature peak in the in-plane thermal conductivity [$\kappa_{ab}$]{}, whereas the out-of-plane thermal conductivity [$\kappa_c$]{}is of purely phononic character (see Fig. \[fig:signatures\]). This strikingly anomalous heat conductivity was first observed in insulating parent compounds of cuprate high-temperature superconductors, in particular, [$\mathrm{La_2CuO_4}$]{}[@Morelli89; @Nakamura91] as well as in [$\mathrm{YBa_2Cu_3O_{7-\delta}}$]{}and [$\mathrm{PrBa_2Cu_3O_{7-\delta}}$]{}with $\delta \approx 1$ [@Cohn95]. After these pioneering experimental findings, the origin of the high-temperature anomaly in [$\kappa_{ab}$]{}remained elusive for several years. Nakamura et al. were the first to speculate that the high-temperature maximum could be related to heat carried by magnetic excitations [@Nakamura91]. However, several attempts to explain this peak by anomalous phononic heat transport involving scattering processes of acoustic phonons with soft optical phonons [@Cohn95] or magnons [@Morelli89] have been made. Eventually, inspired by the observation of an exceptionally large one-dimensional [$\kappa_\mathrm{mag}$]{}in two-leg spin ladder materials [@Sologubenko00; @Hess01] (see section \[sec:ladders\]), several groups reinvestigated the heat conductivity of [$\mathrm{La_2CuO_4}$]{}and related compounds [@Sun2003; @Sun2003a; @Berggold2006; @Hess04; @Hess03; @Hess03a; @Hess2007a; @Hofmann2003; @Yan2003; @Hess2007b]. Qualitatively, all these studies came to the conclusion that a conventional explanation for the high-temperature peak in [$\kappa_{ab}$]{}in the sense of electronic, phononic and even radiative heat transport can be excluded. Sizable electronic heat transport was readily ruled out due to the electronically insulating or only weakly conducting properties of [$\mathrm{La_2CuO_4}$]{}[@Hess03; @Hofmann2003] (and all other materials discussed below). The same holds for radiative heat transport [@Hess03] since the optical properties of [$\mathrm{La_2CuO_4}$]{}are almost isotropic in the relevant energy-range below $\mathrm h\nu\lesssim0.1$ eV [@Uchida91]. The possibility of anomalous phononic heat transport, however, deserves further attention here because it is difficult to exclude it based on the heat conductivity data on [$\mathrm{La_2CuO_4}$]{}without further information and accordingly has been specifically addressed in several of the mentioned studies. ![Thermal conductivity of [$\mathrm{La_2CuO_4}$]{}(bottom) and [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}(top). Full circles: $\kappa_c$. Open circles: $\kappa_{ab}$. Open squares: $\kappa_{\mathrm{mag}}$. Solid line: fit according to Eq. \[eq:fit2d\]. Dashed line: $\kappa_{ab,\mathrm {ph}}$. Reproduced from [@Hess03]. For similar measured data see also [@Nakamura91; @Sun2003a; @Yan2003].[]{data-label="fig:leucuo"}](fig_LaCuO_LEuCuO){width="70.00000%"} ### Anomalous phonon scattering versus magnetic heat transport {#anomalous-phonon-scattering-versus-magnetic-heat-transport .unnumbered} One thinkable hypothesis is that the heat conductivity of the 2D materials should be regarded as purely phononic, and thus the high-temperature anomaly in [$\kappa_{ab}$]{}should be regarded as not originating from an additional contribution but rather being the result of an unusual phonon damping by either magnetic excitations [@Morelli89] or soft phonon modes [@Cohn95]. Such phonon damping is known to be caused by *resonant* scattering of the heat-carrying phonons off such excitations. In such a case, a double-peak structure in $\kappa(T)$ results from a strong *reduction* of the heat conductivity in the temperature regime where the scattering is strongest. Well known examples where such a double peak is caused by local magnetic excitations and soft phonon modes are the heat conductivities of $\mathrm{SrCu_2(BO_3)_2}$ [@Hofmann2001], $\mathrm{SrTiO_3}$ [@Suemune1965; @Steigmeier1968], respectively. For [$\mathrm{La_2CuO_4}$]{}, one might consider such a scenario as being quite unlikely due to the complete absence of a double-peak in the heat conductivity perpendicular to the planes, [$\kappa_c$]{}. However, one could argue that both the magnetism and the crystal structure of the material are very anisotropic, and therefore an anisotropic phonon damping was thinkable. Hofmann et al. pointed out that in the 2D square-lattice cuprates, such as [$\mathrm{La_2CuO_4}$]{}, the dispersion of magnetic excitations which ranges from $\sim 0$ to $2J/k_B\sim2000~\mathrm{K}$ ($J$ is the in-plane exchange constant) is too stiff to cause phonon scattering on magnetic excitations being most pronounced in a narrow temperature interval around $100~\mathrm{K}$ [@Hofmann2003]. Damping of acoustic phonons due to soft optical phonon modes does, however, play an important role in the phonon heat conductivity of [$\mathrm{La_2CuO_4}$]{}, indeed. More specifically, soft optical phonon modes associated with the tilting of the $\mathrm{CuO_6}$ octahedra which are present in the entire so-called low-temperature orthorhombic (LTO) phase, i.e. at $T\leq T_\mathrm{HT}\approx500~\mathrm{K}$ [@Boeni88; @Boeni1989; @Birgeneau87], have been shown to cause a significant suppression in the phononic heat conductivity of [$\mathrm{La_2CuO_4}$]{}[@Hess03a]. Thus, in order to unambiguously rule out soft-phonon scattering as the origin of the anomalous [$\kappa_{ab}$]{}, the investigation of materials where the such soft phonons are not present is an obvious route. Hess et al. therefore studied the heat conductivity of [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}[@Hess03]. In this material the presence of Eu on the La-site has only little influence on the magnetism of the $\mathrm{CuO_2}$-planes [@Sun2003a] but has a strong impact on the structure, because it induces a new structural phase at low temperature, the so-called low-temperature tetragonal (LTT) phase, at $T\leq T_\mathrm{LT}\approx135~\mathrm{K}$ [@Buchner94; @Klauss00]. Fig. \[fig:leucuo\] (top panel) shows $\kappa_{ab}(T)$ and $\kappa_c(T)$ of a [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}single crystal. Both curves strongly resemble the findings for [$\mathrm{La_2CuO_4}$]{}(lower panel), yet exhibiting obvious differences: the low-temperature peaks of [$\kappa_{ab}$]{}and [$\kappa_c$]{}are slightly larger and more sharply shaped than in the case of [$\mathrm{La_2CuO_4}$]{}. Furthermore, a step-like anomaly is present at $T_\mathrm{LT}\approx135~\mathrm{K}$. Above $T_\mathrm{LT}$, $\kappa_c$ remains almost constant and stays below the value for the undoped case, while $\kappa_{ab}$ also exhibits a high-temperature maximum at $T\approx270~\mathrm{K}$ which is even larger than that of [$\mathrm{La_2CuO_4}$]{}. The difference between $\kappa_c$ of Eu-doped and of pure [$\mathrm{La_2CuO_4}$]{}can be attributed to a difference in phononic heat conduction: Upon doping [$\mathrm{La_2CuO_4}$]{}with Eu, enhanced scattering of phonons reduces $\kappa_c$ for $T>T_{LT}$, where both compounds have the same structure (LTO). The anomaly at $T_\mathrm{LT}$ ([$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}) signals the transition to a new structural phase for $T<T_\mathrm{LT}$ where $\kappa_{\mathrm {ph}}$ is enhanced. Indeed, soft phonon branches do exist in the LTO-phase of Rare Earth-doped [$\mathrm{La_2CuO_4}$]{}[@Keimer93], which naturally explains a suppression of $\kappa_{\mathrm{ph}}$ of [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}. The change of $\kappa_{\mathrm{ph}}$ at $T_\mathrm{LT}$ then follows from the discontinuous hardening of the soft phonon branch in the LTT-phase [@Martinez91; @Keimer93] and an associated reduced scattering rate of acoustic phonons.[^1] The more striking observation is, however, that the high-temperature peak in [$\kappa_{ab}$]{}of [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}is not affected by the transition. This unambiguously rules out that the double-peak structure in [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}(and thus also in [$\mathrm{La_2CuO_4}$]{}) is caused by soft-phonon scattering. Instead, the high-temperature peak has to be interpreted as stemming from another heat transport channel different from acoustic phonons which adds to the lattice thermal conductivity [$\kappa_\mathrm{ph}$]{}. This latter conclusion is corroborated by investigations on [$\mathrm{Sr_2CuO_2Cl_2}$]{}[@Hofmann2003] and on $\mathrm{R_2CuO_4}$ ($\mathrm{R = Pr}$, Nd, Sm, Eu, Gd) [@Berggold2006; @Jin2003], where the soft-phonon scattering as is present in [$\mathrm{La_2CuO_4}$]{}is rigorously excluded due to different structural phases, yet practically preserving the magnetism of the $\mathrm{CuO}_2$-planes. On the one hand, [$\mathrm{Sr_2CuO_2Cl_2}$]{}is practically isostructural to [$\mathrm{La_2CuO_4}$]{}with the small difference that (apart from having $\mathrm{Sr^{2+}}$-ions on the $\mathrm{La^{2+}}$-sites) $\mathrm{Cl^{1-}}$-ions occupy the positions of the apical $\mathrm{O^{2-}}$-ions of the $\mathrm{CuO}_6$ octahedra which in [$\mathrm{La_2CuO_4}$]{}form the two-dimensional spin-1/2 planes. Unlike [$\mathrm{La_2CuO_4}$]{}, however, [$\mathrm{Sr_2CuO_2Cl_2}$]{}remains in the so-called high-temperature tetragonal (HTT) phase down to lowest temperature and thus does not exhibit any lattice instability which could give rise to soft-phonon scattering as in [$\mathrm{La_2CuO_4}$]{}[@Hofmann2003]. On the other hand, the compounds $\mathrm{R_2CuO_4}$ ($\mathrm{R = Pr, Nd, Sm, Eu, Gd}$) crystallize in a structure quite different from that of [$\mathrm{La_2CuO_4}$]{}, viz. in the so-called tetragonal $T'$-phase without any apical oxygen but with very similar $\mathrm{CuO}_2$-planes as those in [$\mathrm{La_2CuO_4}$]{}. For $\mathrm{R = Pr, Nd, Sm}$ the $T'$-phase is stable down to lowest temperature, whereas for $\mathrm{R = Eu, Gd}$ the structure undergoes a phase transition towards an orthorhombic phase at $170~\mathrm{K}$ and $685~\mathrm{K}$, respectively [@Braden1994; @Vigoureux1997; @Berggold2006]. All three heat conductivity studies [@Hofmann2003; @Berggold2006; @Jin2003] reveal clearly a high-temperature hump in the in-plane thermal conductivity with a similar, albeit with a somewhat smaller magnitude than that of [$\mathrm{La_2CuO_4}$]{}, is also present in these layered cuprates despite having very different structural properties (see Fig. \[fig:SrCuOCl\_Hofmann2003\] for [$\kappa_{ab}$]{}of [$\mathrm{Sr_2CuO_2Cl_2}$]{}in comparison of that of [$\mathrm{La_2CuO_4}$]{}[@Hofmann2003; @Nakamura91]) . Thus the high-temperature peak turns out as a common feature in the layered cuprates which unambiguously is to be interpreted as a two-dimensional excess thermal conductivity within the magnetic $\mathrm{CuO}_2$-planes. Note that the high-temperature peaks in the thermal conductivity of [$\mathrm{YBa_2Cu_3O_{7-\delta}}$]{}, [$\mathrm{PrBa_2Cu_3O_{7-\delta}}$]{}with $\delta \approx 1$ [@Cohn95] should be interpreted in a similar way. ![In-plane thermal conductivity (here labeled $k(T)$) of [$\mathrm{Sr_2CuO_2Cl_2}$]{}(circles) and [$\mathrm{La_2CuO_4}$]{}(dotted line; data from [@Nakamura91]). Solid lines: fits to the phonon thermal conductivity using the Callaway model [@Callaway59]. Inset: The magnetic heat conductivity for [$\mathrm{Sr_2CuO_2Cl_2}$]{}(circles) and [$\mathrm{La_2CuO_4}$]{}(dotted line) (see text). Image taken from [@Hofmann2003].[]{data-label="fig:SrCuOCl_Hofmann2003"}](fig_SrCuOCl_Hofmann2003){width="70.00000%"} Thus, the bottom line of this section is that the high-temperature anomaly in the thermal conductivity of [$\mathrm{La_2CuO_4}$]{}and related layered insulating cuprates originates from a substantial two-dimensional excess thermal conduction which adds to the typical phononic thermal conductivity [$\kappa_\mathrm{ph}$]{}. In lack of other candidates it is straightforward to conclude that magnetic excitations (i.e. magnons) of the two-dimensional antiferromagnetic planes cause this unusual heat conduction.[^2] Hence, the thermal conductivity measured parallel to the $\mathrm{CuO}_2$-planes ([$\kappa_\Vert$]{}) can be considered as the sum of a conventional phonon heat conductivity [$\kappa_\mathrm{ph}$]{}and a magnetic contribution, i.e., $$\kappa_\Vert=\kappa_\mathrm{ph}+\kappa_\mathrm{mag} .$$ Extraction of the magnetic heat conductivity -------------------------------------------- Having qualitatively established that a sizable magnon contribution [$\kappa_\mathrm{mag}$]{}is present in the in-plane thermal conductivity of [$\mathrm{La_2CuO_4}$]{}, it is interesting to move one step further and to extract the temperature dependence of [$\kappa_\mathrm{mag}$]{}. In order to estimate this temperature dependence, it is of paramount importance to determine its phononic part as accurately as possible. The first step in this estimation is to exploit that [$\kappa_\mathrm{mag}$]{}is expected to roughly follow $\kappa_\mathrm{mag}\propto T^2$ (see below) at low temperature and therefore to become negligible in the temperature range of the phononic low-temperature peak. Thus one can estimate $\kappa_\Vert\approx\kappa_\mathrm{ph}$ at $T\lesssim40~\mathrm{K}$ and extrapolate $\kappa_\mathrm{ph}(T)$ towards higher temperature. One possible approach for this extrapolation is to use the Callaway model [@Callaway59] for fitting the low-temperature ($T\lesssim40~\mathrm{K}$) thermal conductivity and to use the thereby obtained fit parameters for the high-temperature extrapolation. This procedure has been applied to [$\mathrm{Sr_2CuO_2Cl_2}$]{}and $\mathrm{R_2CuO_4}$ ($\mathrm{R = La, Pr, Nd, Sm, Gd}$) [@Hofmann2003; @Berggold2006], see Fig. \[fig:SrCuOCl\_Hofmann2003\] and Fig. \[fig:kmag\_2D\_Berggold2006\] for the fits and the resulting [$\kappa_\mathrm{mag}$]{}. ![Magnetic contributions to the in-plane thermal conductivity, calculated via $\kappa_\mathrm{mag}=\kappa_\Vert-\kappa_\mathrm{ph}$, where [$\kappa_\mathrm{ph}$]{}is determined by a Callaway fit of the low-temperature maximum. Upper panel: Values calculated from measurements of [$\kappa_\Vert$]{}of $\mathrm{R_2CuO_4}$ [@Jin2003; @Berggold2006]. Lower panel: The same analysis for various data of [$\mathrm{La_2CuO_{4+\delta}}$]{}[@Nakamura91; @Yan2003; @Sun2003; @Berggold2006]. Inset: The maximum of the calculated [$\kappa_\mathrm{mag}$]{}vs. the Néel temperature of [$\mathrm{La_2CuO_{4+\delta}}$]{}. It is worth to note, that for high excess oxygen contents a spatial phase separation into hole-rich and hole-poor regions occurs in [$\mathrm{La_2CuO_{4+\delta}}$]{}at $T\lesssim290~\mathrm{K}$ [@Yu1996; @Zolliker90]. This explains the very small [$\kappa_\mathrm{mag}$]{}and the dip at $\sim 275~\mathrm{K}$ for the sample with $T_N=245~\mathrm{K}$ shown in Fig. \[fig:kmag\_2D\_Berggold2006\] [@Berggold2006]. Image adapted from [@Berggold2006].[]{data-label="fig:kmag_2D_Berggold2006"}](fig_kmag_2D_Berggold2006){width="70.00000%"} For [$\mathrm{La_2CuO_4}$]{}the application of the Callaway model for extrapolating [$\kappa_\mathrm{ph}$]{}is bound to be error-prone due to the dominant importance of soft-phonon scattering which is difficult to incorporate into the model. The step-like change of [$\kappa_c$]{}of [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}at $T_\mathrm{LT}$ (Fig. \[fig:leucuo\]) tellingly demonstrates this complication. One possibility to overcome this problem is to use a phenomenological approach the *measured* purely phononic [$\kappa_\bot$]{}perpendicular to the $\mathrm{CuO_2}$ planes and to exploit the negligibly small [$\kappa_\mathrm{mag}$]{}in the temperature range of the phononic low-temperature peak. Thus one can estimate $\kappa_\Vert\approx\kappa_\mathrm{ph}$ at $T\lesssim40~\mathrm{K}$ and extrapolate $\kappa_\mathrm{ph}=A\cdot\kappa_\bot$ for higher temperatures, where $A$ is a suitable scaling factor. Fig. \[fig:leucuo\] shows examples of in this way estimated [$\kappa_\mathrm{ph}$]{}for the in-plane heat conductivity [$\kappa_\Vert$]{}of [$\mathrm{La_2CuO_4}$]{}and [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}and the resulting [$\kappa_\mathrm{mag}$]{}[@Hess03]. As we shall see further below, the magnetic heat conductivity is well detectable also in polycrystalline samples. However, due to the polycrystalline nature of such samples, anisotropic information on $\kappa$ is averaged over. None of the above described approaches for estimating [$\kappa_\mathrm{ph}$]{}(and thus [$\kappa_\mathrm{mag}$]{}) can applied to such data in a simple way because on the one hand the application of the Callaway model to polycrystalline data requires further, non-clarified assumptions, and, on the other hand, for a polycrystal [$\kappa_\bot$]{}can obviously not be obtained independently. Hess et al. therefore estimated the phonon contributions $\kappa_{\mathrm {ph}}^{\mathrm {poly}}$ by fitting $\kappa$ at the high-temperature edge of its maximum (see Fig. \[fig:laznkap\]) by $\kappa_{\mathrm {ph}}=\alpha/T+\beta$ and by extrapolating this fit towards high temperature [@Hess03]. In turn, $\kappa_{\mathrm{mag}}^{\mathrm {poly}}$ on the polycrystals is obtained by $\kappa_{\mathrm{mag}}^{\mathrm {poly}}=\kappa-\kappa_{\mathrm {ph}}$. Note, that the measured $\kappa_{\mathrm{mag}}^{\mathrm {poly}}$ is smaller than the intrinsic $\kappa_{\mathrm{mag}}$ of these compounds by the factor of 2/3 due to averaging over all three components of the $\kappa$ tensor [@Hess03]. Analysis of the magnetic heat conductivity ------------------------------------------ ### Kinetic model {#kinetic-model-1 .unnumbered} The application [@Hess03; @Hess2007b] of the kinetic model (Eq. \[eq:kinetic\_general\]) to the 2D thermal conductivity of a single magnon dispersion branch (labeled by $i$) yields $$\tilde{\kappa}^i=\frac{1}{2}\frac{1}{(2\pi)^2}\int v_{\bf k}l_{\bf k} \frac{d}{dT}n_{\bf k}\epsilon_{\bf k}d{\bf k} ,$$ with $v_{\bf k}$, $l_{\bf k}$, $n_{\bf k}$ and $\epsilon_{\bf k}$ the velocity, mean free path, Bose-function and energy of a magnon, respectively. Note that $\kappa_{\mathrm{mag}}^i$ of a three-dimensional ensemble of planes, as realized in [$\mathrm{La_2CuO_4}$]{}, results from the multiplication of $\tilde{\kappa}^i$ with the number of planes per unit length, i.e., $\kappa_{\mathrm{mag}}^i=\frac{2}{c}\tilde{\kappa}^i$, where $c=13.2~\text{\rm\AA}$ is the lattice constant of [$\mathrm{La_2CuO_4}$]{}perpendicular to the planes. Then the total $\kappa_{\mathrm{mag}}$ is given by summing up $\kappa_{\mathrm{mag}}^i$ of each magnon branch. In order to calculate $\kappa_{\mathrm{mag}}^i$ one can approximate the magnon dispersion relation $\epsilon_{\bf k}$ of the two branches $i=1,2$ with the 2D-isotropic expression $$\epsilon_{\bf k}=\epsilon_k=\sqrt{\Delta^2_i+(\hbar v_0k)^2} ,$$ which describes the dispersion observed experimentally [@Keimer93; @Coldea01] for small values of $k$. Here, $v_0$ is the spin wave velocity while $\Delta_1$ and $\Delta_2$ denote the spin gaps of each magnon branch. Assuming a momentum independent mean free path, i.e., $l_{\bf k}\equiv l_{\mathrm{mag}}$ one finds for each magnon branch [@Hess03; @Hess2007b], $$\label{eq:fit2d} \kappa^i_\mathrm{mag}=\frac{{k_B^3T^2l_\mathrm{mag}}}{2\pi\hbar^2v_0c} \int_\frac{\Delta_i}{k_BT}^\infty x^2\sqrt{x^2-x_{\mathrm{0,i}}^2}\frac{e^x}{(e^x-1)^2} \, dx,$$ with the spin wave velocity $v_0\approx1.287\cdot10^5~\mathrm{m/s}$ [@Hayden91a]. The integral is temperature dependent via its lower boundary $x_{\mathrm{0,i}}=\Delta_i/(k_BT)$, where $\Delta_{1}/k_B\approx26~\mathrm{K}$ and $\Delta_{2}/k_B\approx58~\mathrm{K}$ [@Keimer93]. However, the temperature dependence is weak in the $T$-range where the experimental data are discussed and thus one roughly has $\kappa_{\mathrm{mag}}\propto T^2$, if one presumes [$l_\mathrm{mag}$]{}to be temperature independent. ### Low-temperature characteristics {#low-temperature-characteristics .unnumbered} Eq. \[eq:fit2d\] has been used to fit the [$\kappa_\mathrm{mag}$]{}data of [$\mathrm{La_2CuO_4}$]{}and [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}shown in Fig. \[fig:leucuo\] assuming a *temperature-independent* magnon mean free path $l_{\mathrm{mag}}$ [@Hess03]. In this procedure, an additive shift of the $\kappa_{\mathrm{mag}}$-curve was allowed for which yields a further free parameter apart from $l_{\mathrm{mag}}$ and accounts for the aforementioned uncertainties in the magnitude of $\kappa_{\mathrm{mag}}$. As can be seen in the figure, for both compounds satisfactory fits (solid lines) were obtained at intermediate temperature ranges (70-158 [K]{} for [$\mathrm{La_2CuO_4}$]{}and 54-131 [K]{} for [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}, [@Hess03]). While the slight deviations between the fitted and experimental data towards low $T$ are due to the uncertainties in $\kappa_{\bot,\mathrm{ph}}$ in this range, the deviations at high temperature can be understood in terms of $l_{\mathrm{mag}}$ becoming temperature-dependent due to enhanced magnon scattering off magnons or phonons. Notably, the data are consistent with a temperature-independent $l_{\mathrm{mag}}$ for $T$ both, within the fit interval and below, indicating that in this range the mentioned temperature-dependent scattering processes (magnon-magnon scattering, magnon-phonon scattering, or effects of a finite spin-correlation length at $T>T_N$) are frozen out and thus may be discarded. Therefore, relevant processes seem to be sample-boundary scattering or scattering off defects within the $\mathrm{CuO}_2$-planes. The analysis yields $l_{\mathrm{mag}}\approx1160~\text{\rm\AA}$ and $l_{\mathrm{mag}}\approx560~\text{\rm\AA}$ for [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}and [$\mathrm{La_2CuO_4}$]{}, respectively. Since these values are far too small to correspond to the crystal dimensions, which are of the order of millimeters, these values naturally can be interpreted as the size of two-dimensional defect-free grains. Thus, magnon-defect scattering is the most likely scattering process which dominates the low-temperature magnon transport. This conclusion is consistent with the fact that $\kappa_{\mathrm{mag}}$ and $l_{\mathrm{mag}}$ are quantitatively different for [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}and [$\mathrm{La_2CuO_4}$]{}: since the magnetic properties of both compounds are expected to be identical in essence (i.e., the same spin wave velocity $v_0$), unequal $\kappa_{\mathrm{mag}}$ can arise only due to a difference in densities of the magnetic defects that restrict $l_{\mathrm{mag}}$. ![Open circles: Thermal conductivity $\kappa$ of [$\mathrm{La_{2}Cu_{1-z}Zn_zO_{4}}$]{}polycrystals ($z=0$, 0.005, 0.01, 0.05) as a function of $T$. Solid lines: extrapolated phonon heat conduction [$\kappa_\mathrm{ph}$]{}assuming $\kappa_{\mathrm {ph}}=\alpha/T+\beta$. Similar thermal conductivity data have been obtained on single crystalline [$\mathrm{La_{2}Cu_{1-z}Zn_zO_{4}}$]{}by Sun et al. [@Sun2003] in qualitative agreement with the shown data. Image reproduced from [@Hess03].[]{data-label="fig:laznkap"}](fig_laznkap){width="70.00000%"} ### Magnon-defect scattering {#magnon-defect-scattering .unnumbered} There is no clear alternative way to measure the defect density of a given crystal directly, in order to allow a quantitative comparison of the magnon mean free path [$l_\mathrm{mag}$]{}with the defect density (or the defect distance). Hess et al. have therefore performed measurements of $\kappa$ on samples with a well-defined density of magnetic defects. Such defects can be induced in [$\mathrm{La_2CuO_4}$]{}, e.g., by substituting a small amount of non-magnetic Zn$^{2+}$-ions for the magnetic Cu$^{2+}$-ions. Representative results obtained for such samples of [$\mathrm{La_{2}Cu_{1-z}Zn_zO_{4}}$]{}are shown in Fig. \[fig:laznkap\]. The Zn-impurities represent both structural and magnetic impurities and therefore should affect the phononic as well as the magnetic peak in the heat conductivity. In fact, as can be inferred from the figure, the Zn-doping leads indeed to a gradual suppression of both, the phonon as well as the magnon contribution to $\kappa$. The analysis of the data for [$\mathrm{La_{2}Cu_{1-z}Zn_zO_{4}}$]{}using Eq. \[eq:fit2d\] in an analogous way as in the undoped case[^3] yields the interesting result for the low-temperature magnon mean free path that it scales linearly with the reciprocal Zn content [@Hess03]: $l_{\mathrm{mag}}\approx 0.74\cdot a/z$ (with the lattice constant $a$). Hence, these result suggests that [$l_\mathrm{mag}$]{}is about equal to the unidirectional distance between the Zn-ions within the $\mathrm{CuO_2}$ planes, i.e., [$\kappa_\mathrm{mag}$]{}can be directly used to measure these distances. It is important to note that the nature of the doped defects plays a decisive role in the effective strength of the magnon-defect scattering. In particular, *mobile* defects can be induced in [$\mathrm{La_2CuO_4}$]{}by hole-carrier doping. Such doping can be achieved either by substituting Sr for La or by enhancing the oxygen content of [$\mathrm{La_2CuO_4}$]{}, which at doping levels beyond about 5% leads to high-temperature superconductivity. Interestingly, in the case of Sr-doping, one observes a severe suppression of [$\kappa_\mathrm{mag}$]{}already at $\sim1\%$ doping level [@Sun2003; @Hess2007a], i.e. at much lower doping levels as compared to the afore-discussed Zn-doping. The different effect of these two doping schemes is consistent with the different impact of the defects on the antiferromagnetic correlations in the 2D-HAF. The doped charge strongly couples to the spin excitations resulting in a dressed quasiparticle (the spin polaron) which moves through the antiferromagnetic background. This motion of the spin polaron very effectively destroy the antiferromagnetic correlation of the spins in the plane (see Fig. \[fig:spinloch\] for an illustration) [@Martinez1991; @Chernyshev1999], whereas static non-magnetic defects (which can be viewed as immobile holes, see Fig. \[fig:spinloch\]a) only dilute the antiferromagnet and lead to a very moderate reduction of the spin correlation [@Brenig91; @Uchinokura95; @Hucker99a]. The strong suppression of [$\kappa_\mathrm{mag}$]{}already at $\sim1\%$ hole concentration suggests that the magnon mean free path amounts only a few lattice spacings, i.e. the effective ’size’ of disturbed antiferromagnet correlation due to the movement of a single hole is very large. ![Sketch of a hole moving in an 2D-antiferromagnet. a) A hole in an antiferromagnet. b) Antiferromagnet after hopping of the hole to a neighboring site. c) Antiferromagnet after two hopping processes. The gray shaded areas highlight the regions of destroyed antiferromagnetic correlation.[]{data-label="fig:spinloch"}](fig_spinloch){width="\textwidth"} Berggold et al. pointed out that the magnitude of the peak in [$\kappa_\mathrm{mag}$]{}of [$\mathrm{La_2CuO_4}$]{}significantly varies throughout the literature (see Fig. \[fig:kmag\_2D\_Berggold2006\], bottom panel) and suggested that the difference arises due to different levels of excess oxygen in the material [@Berggold2006]. More specifically, they pointed out that the reported maximum [$\kappa_\mathrm{mag}$]{}depends monotonically on the Néel temperature $T_N$ (inset of Fig. \[fig:kmag\_2D\_Berggold2006\]). This relation is plausible since at small hole doping levels $T_N$ is known to sensitively depend on the hole content not only for the case of Sr doping [@Hucker99a] but also for excess oxygen [@Chen1991]. However, one bear in mind that, as has been shown for the case of [$\mathrm{La_{2}Cu_{1-z}Zn_zO_{4}}$]{}, a reduced [$\kappa_\mathrm{mag}$]{}can also be caused by chemical impurities through limiting the mean free path, yet having little effect on $T_N$, in particular, in very clean crystals. Furthermore, it is natural to expect that structural defects apart from chemical impurities play a crucial role, i.e. the degree of crystal perfection should have an impact on the size of the two-dimensional grains, which limit the mean free path, as well. This notion is confirmed by recent data for the heat conductivity of ultrapure [$\mathrm{La_2CuO_4}$]{}with $T_N=325~\mathrm{K}$ and [$\kappa_\mathrm{mag}$]{}up to $\sim40$ [$\mathrm{Wm^{-1}K^{-1}}$]{}[@Mohanunpublished], i.e., a much higher value than those collected in Fig. \[fig:kmag\_2D\_Berggold2006\]. Two-leg spin ladders ==================== Materials aspects {#matpropladder} ----------------- Magnetic thermal transport in a cuprate-based two-leg spin ladder system up to present primarily has been observed in the material [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}. The crystal structure (see Fig. \[strukleit1\]) and also the connected physical properties of this material are significantly more complex as compared to all materials discussed in this review and thus deserve a special attention. [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}is composed of two main structural elements which define the physical properties. These are, on the one hand, Cu$_2$O$_3$ planes which realize $S=1/2$ two-leg spin ladders of the type as described in Section \[intromat\]. Each of these planes is formed by a parallel network of individual two-leg ladders in the $ac$-plane (see right panel of Fig. \[strukleit2\]). Each Cu spin (with $S=1/2$) interacts strongly antiferromagnetically via 180$^\circ$ Cu-O-Cu bonds with its two neighboring Cu spins along the ladder legs (leg interaction $J_\Vert$, parallel to the $c$-axis) and also antiferromagnetically with its one neighboring Cu spin on the same rung within one two-leg ladder (rung interaction $J_\bot$, parallel to the $a$-axis). These interactions result in an effective $S=1/2$ two-leg ladder model as described by Eq. \[Hamleit\] with $J_{i,j}=J_\Vert$ along the legs and $J_{i,j}=J_\bot$ along the rungs for each of the individual ladders. The interaction of the spins of one ladder to its neighboring ones is strongly frustrated and thus causes an effective decoupling of the individual ladders from each other [@Gopalan94]. The frustration arises since each Cu spin of one ladder interacts relatively weakly ferromagnetically via 90$^\circ$ Cu-O-Cu bonds with two strongly antiferromagnetically coupled Cu-spins (interaction $J_\Vert$) of the neighboring ladder. ![\[strukleit1\] Three-dimensional representation of the crystal structure of [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}with a view in the direction of the CuO$_3$-ladders and CuO$_2$-chains ($c$-axis). From [@Ammerahldiss].](strukleit1){width="\textwidth"} ![\[strukleit2\] Structural elements of [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}. Left: Plane of CuO$_2$ spin chains. Center: layer of (Sr, Ca, La). Right: Plane of Cu$_2$O$_3$ spin ladders with the structure of two separate two-leg ladders indicated in grey. The dashed line indicates the size of the in-plane average unit cell. Dark and light grey bullets represent the Cu and O sites, respectively. From [@Ammerahldiss].](strukleit2){width="\textwidth"} On the other hand, the structure contains planes (again in the $ac$-direction) of CuO$_2$ chains of edge-shared CuO$_4$ plaquettes with the main exchange path of two neighboring Cu spins within one chain via the approximately 90$^\circ$ Cu-O-Cu bonds, running along the $c$-axis (see left panel of Fig. \[strukleit2\]). The resulting magnetic interaction is about one order of magnitude smaller than those of the ladder and have been shown not to directly contribute to the magnetic heat transport [@Hess01]. The total crystal structure is composed of an alternating stacking of the ladder and chain planes along the $b$-axis where adjacent planes are separated by layers (Sr, Ca, La) ions (see Fig. \[strukleit1\] and central panel of Fig. \[strukleit2\]). The chains and the ladders possess along the $c$-axis different translational periods and thus form two incommensurate sublattices. Together with the (Sr, Ca, La) ions, the ladders constitute a sublattice unit cell with $a\approx 11.3\dots11.5$ Å, $b\approx 12.5\dots13.4$ Å, and $c_L\approx 3.9$ Å. The sublattice unit cell of the chains has the same $a$ and $b$ lattice constants but is somewhat shorter along the chain direction with $c_C\approx2.75$ Å. The almost commensurable average common unit cell is $c=7\times c_L\approx10\times c_C$. It should be noted, that the exact values of the lattice constants depend significantly on the relative ratio of the Sr, Ca, and La constituents [@Ammerahldiss]. Interestingly, the formal valencies of the stoichiometric compound are (Sr$^{2+}$)$_{14}$(Cu$^{2.25+})_{24}$(O$^{2-}$)$_{41}$. This means, the magnetically active ladder and chain structures are intrinsically hole-doped. These holes are not equally distributed among both structures. On the one hand, about one hole per formula unit is located in the ladders [@Osafune97; @Nucker00], whereas the amount in the chains is estimated to about 6 holes per formula unit. The intrinsically doped holes have been shown to undergo a charge ordered state in both the chains and in the ladders. In the former, the ordered holes are found to yield a state of non-interacting spin dimer singlets with an excitation gap of about 130 K [@Ammerahl00; @Matsuda99; @Regnault99; @Klingeler2005; @Klingeler2006], whereas the holes in the ladders form a long-range ordered hole crystal [@Abbamonte2004; @Rusydi2006]. We shall see further below that this charge ordering has profound impact on the magnetic heat transport properties. The isovalent substitution of Ca for Sr in [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}has strong consequences for the physical properties. The charge transport properties gradually develop from a semiconductor-like temperature dependence for [$\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$]{}at $x=0$ to an increasingly metallic characteristics at maximal $x\approx12$ [@Hess04a; @Vuletic2006]. Remarkably, at $x\gtrsim5$ the charge ordered state collapses [@Matsuda99; @Ammerahl00; @Kataev01] and at very high Ca doping levels, superconductivity occurs at high pressure [@Uehara96; @Isobe98]. The increasing metallic nature upon Ca doping commonly is interpreted as a result of a doping induced transfer of holes from the chains into the ladder substructures [@Osafune97; @Nucker00; @Dagotto99]. Alternatively, the divalent Sr and Ca ions can be substituted by trivalent La, which causes a drastic reduction of the total hole content in (Sr,Ca)$_{14-y}$La$_y$Cu$_{24}$O$_{41}$ with increasing $y$. Thus, at nominally $y=6$ the system reaches the realization of ’hole-free’ ladder and chain subsystems. It should be noted, however, that the highest La content at which phase pure single crystals can be achieved is $y\approx5.2$ [@Ammerahl99; @Ammerahldiss]. Quite importantly, already a moderate La-content of $y\approx2$ yields practically ’hole-free’ spin ladders [@Nucker00]. A further consequence of the La-doping is a rapid destruction of the dimer ground state in the chains and the formation of an antiferromagnetic ground state of the chains at $y\gtrsim4$ [@Ammerahldiss; @Matsuda96; @Matsuda98; @Matsuda00b; @Kumagai00; @Ammerahl00a]. The hole-free ladders in (Sr,Ca)$_{14-y}$La$_y$Cu$_{24}$O$_{41}$ at finite $y$ exhibit the theoretically expected properties of an $S=1/2$ two-leg quantum spin ladder in a remarkable way. Inelastic neutron scattering (INS) and optical spectroscopy experiments yield a large $J_\Vert/k_B\approx 1300\dots2200$ K and $J_\bot/k_B\approx1300\dots1440$ K with a gapped ground state and a large triplet excitation gap $\Delta\approx 310\dots410$ K [@Matsuda00b; @Windt01; @Notbohm2007]. For [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}, despite the finite amount of doped holes, the magnetism is barely different [@Eccleston98; @Deng2013], whereas for [$\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$]{}controversial report exist concerning the spin gap $\Delta$. INS results indicate that $\Delta$ remains practically unchanged up to the highest Ca-content [@Deng2013], whereas findings from nuclear magnetic resonance (NMR) experiments suggest a gradual decrease of $\Delta$ [@Kumagai97; @Takigawa98; @Imai98; @Magishi98]. It should be mentioned that the NMR findings suggest a somewhat larger $\Delta/k_B$ for $x=0$ (up to $\sim650$ K) than those from INS. Qualitative proof of magnon heat transport in the spin ladders {#sec:ladders} -------------------------------------------------------------- Kudo et al. were the first to report about unconventional magnon heat transport of the spin ladders in the material [$\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$]{}with the Ca content ranging from $x=0$ to $x=9$ [@Kudo1999]. The data presented in their pioneering study are qualitatively compatible with the data shown in Figures \[fig:signatures\], \[fig:La5\_kap\], and \[fig:SrCa\_kmag\] and also with data by Sologubenko et al. [@Sologubenko00]. In the following, instead of addressing these early works on first-generation single crystals of this material in detail, we focus on more recent work which allows better to carve out the relevant physics. The data shown in Fig. \[fig:signatures\] for the pristine material [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}are qualitatively similar to the afore-discussed case of [$\mathrm{La_2CuO_4}$]{}with the remarkable difference that the high-temperature anomaly in [$\kappa_\Vert$]{}, which is measured parallel to the spin ladders along the $c$-axis, is purely one-dimensional, since it is absent in both directions perpendicular to the ladders [@Hess01]. The magnetic nature of the anomaly in [$\kappa_\Vert$]{}is thus evident, the more so as no structural instabilities as is the case for the two-dimensional layered cuprates are known to be present in this material. However, for the sake of meticulous correctness, one should exclude other thinkable one-dimensional contributions to the heat transport parallel to the ladder structures in the material. In particular, one could imagine unusual optical phonon modes related to the quasi-one-dimensional structural elements present in the materials as well magnetic heat transport by the also present $\mathrm{CuO}_2$-chain structure in [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}. ![Thermal conductivity of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}as a function of temperature measured along the $a$ and $c$ axes, [$\kappa_\bot$]{}and [$\kappa_\Vert$]{}, respectively. The solid line represents an estimate of the phonon contribution to [$\kappa_c$]{}based on the Callaway model. Figure adapted from [@Hess01].[]{data-label="fig:La5_kap"}](fig_La5_kap){width="60.00000%"} Fig. \[fig:La5\_kap\] shows the thermal conductivity parallel ([$\kappa_\Vert$]{}) and perpendicular ([$\kappa_\bot$]{}) to the spin ladder structures in the compound [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}[@Hess01]. In this material, the $\mathrm{Sr^{2+}}$ site of the pristine compound [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}is substituted by two different ions, namely $\mathrm{Ca^{2+}}$ and $\mathrm{La^{3+}}$. The resulting structural disorder leads to a strong defect-scattering of the phonons in the system and, correspondingly, to a drastic suppression of its phonon heat conductivity [$\kappa_\mathrm{ph}$]{}, which is evident from a direct comparison of [$\kappa_\bot$]{}of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}and [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}in Fig. \[fig:La5\_kap\] and Fig. \[fig:signatures\], respectively. Interestingly, the high-temperature peak in [$\kappa_\Vert$]{}is not affected by this suppression. In contrast, in the direct comparison, this peak is even enhanced in [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. We shall see further below that this enhancement can be attributed to the vanishing hole content in the ladders of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. This observation clearly rules out that the high-temperature peak stems from unconventional phononic modes since the suppressed [$\kappa_\mathrm{ph}$]{}demonstrates clearly that these modes are heavily disturbed by the disorder in the material. Furthermore, also thinkable magnetic heat transport of the $\mathrm{CuO_2}$ chains can now safely be excluded because the magnetic state of these chains is fundamentally changed upon the substitution of La and Ca, for Sr [@Ammerahl00a; @Klingeler2006; @Klingeler2005; @Matsuda98]. Thus, the high-temperature peak observed in [$\kappa_\Vert$]{}of [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}unambiguously represents the magnon heat conductivity [$\kappa_\mathrm{mag}$]{}in the material. Furthermore, and remarkably, one finds the unusual materials property that $\kappa_\mathrm{ph}\ll\kappa_\mathrm{mag}$ at $T\gtrsim40~\mathrm{K}$ which is most evident for [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. This remarkable anisotropy of the heat conductivity tensor has recently been visualized by fluorescent microthermal imaging at room temperature, see Fig. \[fig:Otter\] [@Otter2009]. ![a) Fluorescent microthermal image of the $ac$-plane of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}generated by a localized heat pulse in the center of the image. It shows a highly anisotropic pattern due to the high thermal conductivity in the (diagonal) c-direction ([$\kappa_\Vert$]{}). b) Fluorescent microthermal image of the $ab$-plane of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}generated by a localized heat pulse in the center of the image shows an isotropic pattern. The distances on both x- and y-axis are in mm. Figure adapted from [@Otter2009].[]{data-label="fig:Otter"}](fig_Otter){width="\textwidth"} Extraction of the magnetic heat conductivity -------------------------------------------- Due to the presence of a sizable spin gap in [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}of the order of $\Delta/k_B\sim300\dots500~\mathrm{K}$ [@Kumagai97; @Eccleston98; @Imai98; @Matsuda00b; @Notbohm2007] one can safely assume that for low temperatures the magnon heat conductivity falls off as $\kappa_\mathrm{mag}\propto \exp[-\Delta/(k_BT)]$ and becomes negligible at $T\lesssim40~\mathrm{K}$ in comparison to the phonon heat conductivity [$\kappa_\mathrm{ph}$]{}. Thus, one can safely fit [$\kappa_\mathrm{ph}$]{}using the Callaway model [@Callaway59] in this temperature range and extrapolate this fit towards higher temperature in order to estimate the phononic contribution to [$\kappa_\Vert$]{}[@Sologubenko00; @Hess01]. The solid lines in Fig. \[fig:signatures\] and Fig. \[fig:La5\_kap\] represent corresponding results for [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}and [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}, respectively. Note, that the extreme suppression of [$\kappa_\mathrm{ph}$]{}in [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}renders the possible errors in the estimation of [$\kappa_\mathrm{ph}$]{}(and thus [$\kappa_\mathrm{mag}$]{}) very small, which is particularly important at low temperature where [$\kappa_\mathrm{mag}$]{}is still small. Assuming again $\kappa_\Vert=\kappa_\mathrm{ph}+\kappa_\mathrm{mag}$, the magnon heat conductivity data can now be computed from the measured data. The resulting temperature dependence of [$\kappa_\mathrm{mag}$]{}for [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}and [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}is depicted in Fig. \[fig:kmag\_59\_sr14\] [@Hess01]. ![Temperature dependence of [$\kappa_\mathrm{mag}$]{}of the ladder compounds [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}and [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. The solid line represent a fit to the low-temperature data of [$\kappa_\mathrm{mag}$]{}of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}on basis of the kinetic model (Eq. \[eq:kappaladder\]). Figure adapted from [@Hess01]. Similar data for [$\kappa_\mathrm{mag}$]{}of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}have been found by Sologubenko et al. [@Sologubenko00].[]{data-label="fig:kmag_59_sr14"}](fig_kmag_59_sr14){width="60.00000%"} Analysis of the magnon heat conductivity ---------------------------------------- The temperature dependence of the shown data is for both compounds very similar in the region of the low-temperature increase of [$\kappa_\mathrm{mag}$]{}. Differences become apparent only at $T\gtrsim 100~\mathrm{K}$: The increase of [$\kappa_\mathrm{mag}$]{}of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}weakens and after reaching a maximum ($\sim80$ [$\mathrm{Wm^{-1}K^{-1}}$]{}) at $\sim140~\mathrm{K}$, [$\kappa_\mathrm{mag}$]{}decreases strongly and saturates at $\sim10$ [$\mathrm{Wm^{-1}K^{-1}}$]{}for $T\gtrsim240~\mathrm{K}$. In contrast, [$\kappa_\mathrm{mag}$]{}of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}increases much stronger at $T\gtrsim 100~\mathrm{K}$. Also here a maximum value is reached ($\sim140$ [$\mathrm{Wm^{-1}K^{-1}}$]{}), but at clearly higher temperature $T\approx180~\mathrm{K}$, after which [$\kappa_\mathrm{mag}$]{}decreases only moderately and stays large at $\sim90$ [$\mathrm{Wm^{-1}K^{-1}}$]{}even at room temperature, i.e. it attains a value which is comparable to the heat conductivity of a metal. ### Kinetic model {#kinetic-model-2 .unnumbered} In the one-dimensional case, the kinetic model (Eq. \[eq:kinetic\_general\]) for the thermal conductivity of a single ladder yields [@Sologubenko00; @Sologubenko01; @Hess01] $$\tilde{\kappa}=\frac{1}{2\pi}\int v_{k}l_{k} \frac{d}{dT}n_{k}\epsilon_{k}dk , \label{eq:kinetic_1D}$$ again with $v_{k}$, $l_{k}$, and $\epsilon_{k}$ the velocity, mean free path, and energy of a magnon[^4]. The mathematical form of the occupation function $n_{k}$ is, however, unclear for the case of a two-leg spin ladder. Based on the fact that the elementary excitations of a two-leg ladder are triplet excitations, i.e. bosons, Sologubenko et al. approximated it by the Bose-function [@Sologubenko00]. However, Hess et al. argued that the Bose function leads to unphysically large triplet densities at higher temperatures and suggested an occupation function of the form $$\label{eq:triplets} n_k=\frac{3}{e^{\frac{\epsilon_k}{k_BT}}+3}$$ to account, on average, for the hard-core constraint of no on-site double occupancy for the triplet excitations [@Hess01]. The use of it and assuming $l_\mathrm{mag}\equiv l_{\bf k}$ leads then to the expression [@Hess01; @Hess2007b] $$\label{eq:kappaladder} \kappa_\mathrm{mag}=\frac{3 n_s {k_B}^2 }{\pi\hbar}l_\mathrm{mag}T \int_\frac{\Delta}{k_BT}^\frac{\epsilon_\mathrm{max}}{k_BT} x^2 \frac{e^x}{(e^x+3)^2}~dx,$$ where $n_s$ is now the number of ladders per unit area. Note, that Eq. \[eq:kappaladder\]) differs from the expression used by Sologubenko [*et al.*]{} [@Sologubenko00] for the heat-conductivity of one-dimensional bosons not only by the distribution function (\[eq:triplets\]) but also by an overall factor of three accounting for the triplet degeneracy. Thus, the following considerations and the discussion of analyzing the [$\kappa_\mathrm{mag}$]{}data for the two-leg ladder compounds will be based on Eq.\[eq:kappaladder\]. Note further, that this result for a one-dimensional system does not depend on the specific form of the dispersion function $\epsilon_k$ and the velocity $v_k$ [@Hess01]. The dispersion enters only through its band minimum, i.e., the spin gap $\Delta$ and its maximum $\epsilon_\mathrm{max}$. Experimentally, one finds a lower bound $\epsilon_\mathrm{max}\sim200$ meV in [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}[@Eccleston98; @Matsuda00b; @Notbohm2007]. At $T\lesssim300~\mathrm{K}$, the integral in Eq. \[eq:kappaladder\] therefore does only depend weakly on $\epsilon_\mathrm{max}$. Its exact value thus does not play an important role and one might even set $\epsilon_\mathrm{max}=\infty$ at very low temperature [@Hess2007b]. In any case, due the presence of a spin gap [$\kappa_\mathrm{mag}$]{}is expected to be exponentially suppressed at low temperature $k_BT\ll \Delta<J$ and one might approximate $\kappa_\mathrm{mag}\propto \exp[-\Delta/(k_BT)]$ presuming a temperature independent [$l_\mathrm{mag}$]{}. ### Low-temperature increase – thermal occupation of magnons {#low-temperature-increase-thermal-occupation-of-magnons .unnumbered} Apparently, for both materials [$\kappa_\mathrm{mag}$]{}follows such an exponential increase at low temperature (see dotted line in Fig. \[fig:kmag\_59\_sr14\]. An even more accurate description of the data is obtained if the data is described by Eq. \[eq:kappaladder\] [@Hess01]. The solid line in the figure shows a corresponding fit to the data of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}with a temperature independent [$l_\mathrm{mag}$]{}and the spin gap $\Delta$ as free parameters in the temperature range 54-102 K. The data for [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}can be fitted similarly well but in the somewhat reduced temperature interval 61-91 K (see [@Hess01], for the details). In both cases, this analysis yields very similar values for the magnon mean free path [$l_\mathrm{mag}$]{}and the spin gap. For the latter one finds $\Delta/k_B=418\pm15~\mathrm{K}$ and $\Delta/k_B=396\pm10~\mathrm{K}$ for [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}and [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}, respectively, which is in the same order of magnitude than results reported from neutron scattering and NMR measurements (see Section \[matpropladder\]) [@Kumagai97; @Eccleston98; @Imai98; @Matsuda00b; @Notbohm2007]. The results for the low-temperature mean free path $l_0$ are almost identical, specifically $l_0=2980\pm110$ [Å]{} and $l_0=2890\pm230\rm$ [Å]{} for [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}and [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}, respectively.[^5] This corresponds to about 750 lattice spacings along the ladder, which is surprisingly large in view of the rather complicated crystal structure of [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}[@Gopalan94], and, even more intriguing, in view of the spin-spin correlation length in a two-leg spin ladder which exponentially vanishes already after about 3-4 lattice spacings [@Dagotto99; @Dagotto96]. This means, the mean free path becomes orders of magnitude larger than the size of antiferromagnetically correlated spin configurations. This underpins the fundamentally different quantum nature of the magnons in the two-leg spin ladders as compared to antiferromagnetic spin wave type magnons which emerge from an antiferromagnetic ground state with infinite spin-spin correlation length, as is the case for the 2D-HAF. ### Ballistic heat transport? {#ballistic-heat-transport .unnumbered} The report of a large magnon thermal conductivity in [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}and [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}[@Sologubenko00; @Hess01] and pertinent large mean free paths soon after triggered a number of controversial theoretical works which addressed the conjecture that the heat transport of a two-leg spin ladder similar to that of a Heisenberg chain was *ballistic*, despite the non-integrability of the model (see also Section \[sec:chains\] for further elaboration on the connection between the integrability of the spin model and the thermal Drude weight) [@Alvarez02; @Gros2004; @Heidrich04; @Zotos04; @Jung06; @Boulat2007; @Steinigeweg2016]. One particular aspect of this context has been addressed by Alvarez and Gros, who claimed that the heat transport of a two-leg ladder indeed was ballistic [@Alvarez02]. More specifically, based on exact diagonalization results they found a finite thermal Drude weight and evaluated a much shorter magnon mean free path as compared to the above findings through using the kinetic model. Theoretically, these findings have been challenged by several groups: Possible problems due to finite size effects of the exact diagonalization results have been pointed out [@Heidrich04]. Furthermore, high-temperature approximations yielded a vanishing Drude weight [@Zotos04]. It has been argued, however, that the two-leg ladder model despite being non-integrable is still close enough to integrability to yield finite but large transport coefficients [@Boulat2007; @Jung06]. ![$\kappa_{\mathrm{mag}}$ of $\mathrm{Sr_{14}Cu_{24-x}Zn_xO_{41}}$ at $x=0.125, 0.25, 0.5, 0.75$. The gray shaded areas display the experimental uncertainty of $\kappa_{\mathrm{mag}}$ resulting from the uncertainties of $\kappa_{\mathrm{ph}}$ for the representative case $x=0.125$. Figure adapted from [@Hess2006].[]{data-label="fig:ladderZn_kmag"}](fig_LadderZn_kmag){width="60.00000%"} Experimentally, the magnitude of the mean free paths extracted by means of the kinetic model have been verified by a specific doping experiment. Similarly as in the 2D-case of [$\mathrm{La_2CuO_4}$]{}the heat conductivity of Zn-doped [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}has been studied, in order to render the determination of the mean free path model-independent [@Hess2006]. The non-magnetic Zn-ions act as scatterers for the magnons and cause a gradual suppression of [$\kappa_\mathrm{mag}$]{}as can be inferred from Fig. \[fig:ladderZn\_kmag\] which reproduces the obtained data for [$\mathrm{Sr_{14}Cu_{24-z}Zn_{z}O_{41}}$]{}. Since the mean distance of Zn-ions within a ladder, $d_\mathrm{Zn-Zn}$, can be computed from the Zn content $z$, it is instructive to compare this distance with the magnon mean free path which has been extracted from the [$\kappa_\mathrm{mag}$]{}data in Fig. \[fig:ladderZn\_kmag\] using the kinetic model (Eq. \[eq:kappaladder\]). The resulting [$l_\mathrm{mag}$]{}vs. $d_\mathrm{Zn-Zn}$ is shown in Fig. \[fig:dZn\_lmag\]. As is evident from the figure, $l_{\mathrm{mag}}$ roughly scales with $d_{\mathrm{Zn-Zn}}$, and a linear fit to the data points yields a slope of $1.17\pm 0.25$, which is close to unity. This confirms unambiguously that the values for $l_{\mathrm{mag}}$ obtained from the kinetic model are in fair agreement with real scattering lengths in the material [@Hess2006]. The above findings allow to stress several important points: Despite the strong quantum nature of the two-leg spin ladder (which manifests itself in a large spin gap and a short-range spin correlation) the experimentally observed magnetic heat conductivity can be successfully described by the simple Boltzmann-type kinetic model (Eq. \[eq:kappaladder\]). The model is non-integrable and thus ballistic heat transport is not expected from fundamental conservation laws [@Zotos2005; @Zotos97], yet the heat conductivity seems anomalously large because the experimental mean free paths are up to several orders of magnitude larger than the spin correlation lengths in the system. ![$l_{\mathrm{mag}}$ as a function of $d_{\mathrm{Zn-Zn}}$. Solid line: linear fit line through the origin. The error bars arise due to uncertainties in determining $\kappa_{\mathrm {ph}}$. Figure reproduced from [@Hess2006].[]{data-label="fig:dZn_lmag"}](fig_dZn_lmag){width="50.00000%"} ### Temperature dependent scattering processes {#sec:tdepmagnonscatt .unnumbered} We will now briefly address the characteristics of [$\kappa_\mathrm{mag}$]{}of the spin ladders at higher temperatures, in order to study the influence of temperature-dependent scattering processes on the magnon heat transport, i.e. scattering of magnons off other excitations in a solid such as phonons, charge carriers, and also other magnons. The spin ladder material [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}is an ideal ’playground’ in this regard because it can be doped in very different ways. Definitely, the possibility of doping the spin ladders with charge carriers is intriguing and unique among one-dimensional cuprate quantum magnets. ![$\kappa_{\mathrm{mag}}(T)$ of $\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$ ($x=0, 2, 3, 4, 5$). Top panel: Data for ($x=0, 2, 3$) in comparison with $\kappa_{\mathrm{mag}}$ of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. Lower panel: enlarged representation for $x=3, 4, 5$. Stars indicate the approximative $T^*$ for $x\leq3$. Similar results for $x=0, 2$ have been reported by [@Sologubenko00]. Figure adapted from [@Hess04a; @Hess2007b].[]{data-label="fig:SrCa_kmag"}](fig_kmag_SrCa){width="60.00000%"} ### Magnon-hole scattering {#magnon-hole-scattering .unnumbered} In the previous discussion of the low-temperature increase at $T\lesssim100$ K the fact that the hole content in the ladders of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}and [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}is very different did not play an important role. Apparently, at this low temperatures the presence of the hole in the ladders of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}is unimportant for the magnetic heat transport, as is reflected by the very similar values for [$\kappa_\mathrm{mag}$]{}(see Fig. \[fig:kmag\_59\_sr14\]) and thus the mean free path and the spin gap of both compounds. New light is shed on the relevance of the holes in the ladders from the direct comparison of the magnon heat conductivity [$\kappa_\mathrm{mag}$]{}of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}with that of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}in Fig. \[fig:kmag\_59\_sr14\] at $T\gtrsim100$ K. The impact of the higher hole content in the ladders of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}can directly be read off the figure: while [$\kappa_\mathrm{mag}$]{}is practically identical for both compounds at $T \lesssim100~\mathrm{K}$, one observes the already described differences at higher temperature. It is straightforward to explain the comparatively stronger high-temperature suppression of [$\kappa_\mathrm{mag}$]{}of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}by a significant magnon-hole scattering because the presence of holes in the ladders is the essential difference to the undoped ladders in [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. The equality of both [$\kappa_\mathrm{mag}$]{}data below the characteristic temperature $T_0\approx100~\mathrm{K}$ implies in addition that this scattering process, on the one hand, is completely unimportant at low temperature $T\lesssim T_0$, but on the other hand unfurls its full strength above another characteristic temperature $T^*\approx 240~\mathrm{K}$ where [$\kappa_\mathrm{mag}$]{}saturates [@Sologubenko00; @Hess01; @Hess02; @Hess04a]. This surprising temperature dependence of the magnon-hole scattering strength has been investigated towards its robustness against changes of the hole concentration [@Sologubenko00; @Hess04a]. [$\kappa_\mathrm{mag}$]{}was studied in a series of [$\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$]{}single crystals. Results for [$\kappa_\mathrm{mag}$]{}at low doping levels $x\leq5$ [@Hess04a] are shown in Fig. \[fig:SrCa\_kmag\]. One can infer quite clearly that $T^*$ is gradually shifted towards lower temperature with increasing $x$, i.e., the temperature interval in which [$\kappa_\mathrm{mag}$]{}is suppressed due to magnon-hole scattering is extended to lower temperatures. At $x=4$, 5 the magnon-hole scattering is apparently so dominant, that even the low-temperature peak is suppressed. The observed doping dependence of the magnon-hole scattering can clearly be related to the charge-ordered state [@Abbamonte2004; @Rusydi2006] of the holes in the ladder below $T^*$ [@Sologubenko00; @Hess01; @Hess02; @Hess04a]. More specifically, the charge order is accompanied by a drastic enhancement of the magnon mean free path [$l_\mathrm{mag}$]{}of the magnons where the scattering probability turns out to be practically one (inferred from comparing the [$l_\mathrm{mag}$]{}with the estimated distance of holes in the ladders, see below) in the case of completely mobile holes ($T>T^*$), and vanishes in the long-range ordered state ($T< T_0$) [@Sologubenko00; @Hess04a]. Note, that charge order is also present at $x=4$, 5 [@Vuletic2005; @Vuletic2003]. However, its onset temperature $T^*$ is already so low that [$\kappa_\mathrm{mag}$]{}is significantly suppressed. Note that the magnon heat conductivity in these samples is still well detectable, despite the substantial suppression, see Fig. \[fig:SrCa\_kmag\]. Sologubenko et al. have shown that a large [$\kappa_\mathrm{mag}$]{}exists even at a very high doping level $x=12$ [@Sologubenko00], where no charge order is present [@Rusydi2006; @Vuletic2005; @Vuletic2003]. The above qualitative analysis of the magnon-hole scattering in [$\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$]{}demonstrates, that [$\kappa_\mathrm{mag}$]{}of a quantum magnet beyond being intriguing in itself can also be used as a *probe* for the charge dynamics and order in the material. A very interesting quantitative result which can be inferred from the above discussion for the undoped [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}is that through the determination of the low-temperature value of the mean free path $l_0$, i.e. the mean free path deep in the charge ordered phase, the correlation length of the charge order $\xi$ can be estimated. Since $l_0\sim3000$ [Å]{} is much larger than the mean hole distance in the ladders and $l_0$ is practically identical for the hole-free [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}, the charge order should be perfect at least on this length scale, i.e. $\xi\gtrsim3000$ [Å]{} [@Hess04a]. It should be noted, that the magnon-hole scattering and the impact of the charge order on it can be well analyzed by studying the temperature dependence of the magnon mean free path $l_\mathrm{mag}(T)$ which can be extracted by comparing the measured [$\kappa_\mathrm{mag}$]{}data with Eq. \[eq:kappaladder\] if the previously extracted values for $\Delta$ are plugged into it. Hess et al. have decomposed $l_\mathrm{mag}(T)$ into a temperature independent magnon-defect part $l_0$ and a temperature-dependent part $l_h(T)$ which represents solely the magnon-hole scattering, using Matthiessen’s rule: $$l_{\mathrm{mag}}(T)^{-1}=l_0^{-1}+l_h(T)^{-1}. \label{eq:matthiesen_hole}$$ It could be shown that on the one hand $l_h$ at high temperatures $T>T^*$ is of the same order of magnitude than the mean distance of holes in the ladders. On the other hand, it was found that $l_h$ has a very similar temperature dependence as the electrical resistivity, which further confirms that [$\kappa_\mathrm{mag}$]{}depends sensitively on the charge dynamics in the material [@Hess04a]. ![$l_{\mathrm{mag}}$ of [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}as a function of temperature $T$. The solid and broken lines represent fits of $l_{\mathrm{mag}}$ accounting for magnon-phonon and magnon-magnon scattering respectively. Adapted from [@Hess01; @Hess05; @Hess2007b].[]{data-label="fig:La5_lmag_fits"}](fig_La5_lmag_fits){width="60.00000%"} ### Magnon-phonon scattering {#magnon-phonon-scattering .unnumbered} It is clear from the above discussion of magnon-hole scattering in [$\mathrm{Sr_{14-x}Ca_xCu_{24}O_{41}}$]{}, that it must be unimportant in [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. In spite of this, the measured [$\kappa_\mathrm{mag}$]{}substantially deviates for $T\gtrsim100~\mathrm{K}$ from the theoretically calculated [$\kappa_\mathrm{mag}$]{}based on a constant mean free path. Formally, this deviation can be captured by a temperature dependent mean free path $l_\mathrm{mag}(T)$, where the temperature dependence suggests the presence of a further temperature dependent scattering process. Hess et al. have analyzed this in more detail [@Hess01; @Hess05], the results of which have been summarized by [@Hess2007b]: Apart from magnon-hole scattering, which is unimportant in [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}, only magnon-magnon scattering or magnon-phonon scattering are thinkable processes which could cause this temperature dependence of [$l_\mathrm{mag}$]{}. As has been mentioned above, the $T$-dependence of $l_{\mathrm{mag}}$ in [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}can be calculated from the $\kappa_{\mathrm{mag}}$ data using Eq. \[eq:kappaladder\] by plugging in the previously extracted $\Delta/k_B=418~\mathrm{K}$ [@Hess01; @Hess05]. The resulting $l_{\mathrm{mag}}(T)$ as shown in Fig. \[fig:La5\_lmag\_fits\] reflects the different $T$-regimes which govern $\kappa_\mathrm{mag}$. For $T\lesssim110~\mathrm{K}$, $l_\mathrm{mag}$ is $T$-independent with a mean value $l_0=2980$ [Å]{} which reflects the scattering of magnons off static defects. In order to describe the $T$-dependent $l_{\mathrm{mag}}$ at higher $T$ it was assumed (as in the case of magnon-hole scattering, see Eq. \[eq:matthiesen\_hole\]) that all scattering mechanisms were independent of each other and Matthiessen’s rule applied $$l_{\mathrm{mag}}^{-1} = l_0^{-1}+\gamma_\mathrm{ph}d_\mathrm{ph}^{-1}+\gamma_\mathrm{mag}d_\mathrm{mag}^{-1}.$$ Here, $d_\mathrm{ph}$ and $d_\mathrm{mag}$ are the mean ’distances’ of phonons and magnons respectively, as calculated from the particle densities with $\gamma_\mathrm{ph}$ and $\gamma_\mathrm{mag}$ the corresponding scattering probabilities. Since it is unclear as to what extent the separate scattering mechanisms contribute to $l_{\mathrm{mag}}$, its behavior was analyzed based on the assumption that only one mechanism is active in addition to magnon-defect scattering. The case of dominant magnon-phonon scattering was modeled by three energy-degenerate non-dispersive optical branches along the ladder direction, yielding $$\frac{1}{d_\mathrm{ph}}=\frac{7.6\cdot10^{9}\rm m^{-1}}{\exp(\Delta_\mathrm{opt}/(k_BT))-1}$$ with $\Delta_\mathrm{opt}$ the optical gap (cf. [@Hess05] for details). The experimental $l_{\mathrm{mag}}$ was then fitted with $l_{\mathrm{mag}}^{-1} = l_0^{-1}+\gamma_\mathrm{ph}d_\mathrm{ph}^{-1}$ using $\gamma_\mathrm{ph}$ and $\Delta_\mathrm{opt}$ as free parameters. The fit (solid line in Fig. \[fig:La5\_lmag\_fits\]) describes the data fairly well. Remarkably, the value found for $\Delta_\mathrm{opt}/k_B=795~\mathrm{K}$ is of the same order of magnitude as the energy of the longitudinal [Cu-O]{} bond stretching mode which is involved in the two-magnon-plus-phonon absorption observed in optical spectroscopy [@Gruninger2000; @Windt01]. The scattering probability is obtained as $\gamma_\mathrm{ph}=3.2\cdot10^{-2}$, i.e. significantly smaller than that of magnon-hole scattering for mobile holes. For the assumption of dominant magnon-magnon scattering, a less satisfactory agreement was obtained with $l_{\mathrm{mag}}^{-1} = l_0^{-1}+\gamma_\mathrm{mag}d_\mathrm{mag}^{-1}$, where $$\frac{1}{d_\mathrm{mag}}=\frac{1}{\pi c_L}\int_0^\pi\frac{3}{3+\exp(\epsilon_k/k_BT)}dk$$ (broken line in Fig. \[fig:La5\_lmag\_fits\]). $c_L$ is the lattice constant along the ladders and $\epsilon_k$ was taken from Johnston et al. for the case of isotropic ladder coupling [@Johnston00], with $\epsilon_{k=\pi}/k_B=\Delta/k_B=418~\mathrm{K}$ employed. Note that $\gamma_\mathrm{mag}=0.05$ and thus of similar magnitude as $\gamma_\mathrm{ph}$. The comparison between both fits suggests that scattering off optical phonons is dominant in this compound, since an almost perfect description of [$l_\mathrm{mag}$]{}is obtained without the necessity to invoke magnon-magnon scattering as additional scattering mechanism. Nevertheless, a contribution from magnon-magnon scattering cannot be excluded on basis of this analysis. Scattering off optical phonons is a plausible scattering mechanism for magnons since the longitudinal [Cu-O]{} bond stretching mode directly affects the [Cu-O]{} distance and hence the magnetic exchange constant of the ladders $J$. It is worth mentioning that the scattering off acoustic phonons appears unlikely, since the magnon energies with a $\Delta/k_B\sim400~\mathrm{K}$ clearly are higher than that of acoustic phonons, unlike that of optical phonons. One should note, however, that a Debye temperature[^6] $\Theta_D~\sim300~\mathrm{K}$ in [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}is very close to the lower bound of the spread of measured values for the spin gap in two-leg spin ladder materials, see, e.g. [@Notbohm2007]. Thus, though optical phonons are a prominent candidate, acoustic phonons and also magnons cannot be excluded to play a role in the dissipation of magnon heat transport in the two-leg spin ladder compounds [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}. Time-dependent measurements --------------------------- All data discussed so far have been obtained with the use of the so-called steady-state method, see, e.g. [@Berman]. It should be noted, however, that concerning results for the two-keg spin ladders significant fluctuations of the absolute magnitude of the magnetic contribution to kappa of the very same sample have been observed by several groups. Details of this phenomenon are discussed in [@Hess_diss2002].[^7] There are a few attempts to overcome these experimental difficulties by dynamical heat transport studies focusing on a single material, namely [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. Interestingly, these studies consistently yield a lower value of [$\kappa_\Vert$]{}, which at room temperature amounts to about 60% of the data shown in Fig. \[fig:La5\_kap\]. This finding is also in reasonable agreement with more recent steady-state studies by Naruse et al. [@Naruse2013]. Apart from this quantitative result, these dynamic studies are however conflicting with respect to the extraction of the magnon-phonon relaxation time. Thus, the results of these studies [@Otter2012; @Montagnese2013; @Hohensee2014] are just briefly summarized here, while for an in-depth discussion the reader is referred to the original papers. Otter et al. have exploited the fluorescent microthermal imaging technique which has been mentioned already in the beginning of this section further to extract the time-dependence of the anisotropic heat spread after locally heating the surface of a [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}crystal in the $ac$-plane (see Fig. \[fig:Otter2\]) [@Otter2012]. By analyzing the time-dependent anisotropic heat spread with the three-dimensional heat diffusion equation, the obtained data were used to extract the room-temperature heat conductivity along the $a$- and $c$-directions, i.e. [$\kappa_\bot$]{}and [$\kappa_\Vert$]{}, respectively. Interestingly, while the extracted value for [$\kappa_\bot$]{}was found to be in good agreement with the steady-state results shown in Fig. \[fig:La5\_kap\], the value for [$\kappa_\Vert$]{}amounts, as already mentioned, only about 60% of the steady-state [$\kappa_\Vert$]{}. ![Time evolution of the heat diffusion from a hot spot in the spin ladder compound [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}. The left column shows the data, while the right column shows the best Gaussian fit to the data. Heating is done by a laser with a pulse duration of $\mathrm 20~\mu s$. The integration time for the probe UV-pulse is $\mathrm 20~\mu s$. The ladder direction is horizontal. The anisotropy of the diffusion process is clearly seen. Figure taken from [@Otter2012].[]{data-label="fig:Otter2"}](fig_Otter2){width="65.00000%"} Montagnese et al. employed a different but related dynamic method where they measured the time-of-flight of a heat pulse through a [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}sample for various sample thicknesses. This study allowed to extract the phonon-magnon equilibration time, yielding a very large $\tau_\mathrm{mp}\approx\mathrm 400~\mu s$ as compared to that for the spin chain compound [$\mathrm{SrCuO_{2}}$]{}where $\tau_\mathrm{mp}\approx\mathrm 10^{-12}~s$ [@Montagnese2013], and for which the mentioned fluctuations in [$\kappa_\Vert$]{}are absent. This finding is, however, in contrast with very recent measurement results using the time-domain thermoreflectance [@Hohensee2014] where a six orders of magnitude faster phonon-magnon equilibration time is found. At present, the origin of these strong discrepancies between both these pioneering studies remains unclear and is not discussed further at this point. Spin chains {#sec:chains} =========== Theoretical preliminaries ------------------------- ![a) Sketch of the real part of the thermal conductivity $\kappa(\omega)$ of the Heisenberg chain as a function of frequency $\omega$. The conductivity is given by $ \operatorname{Re}~\kappa(\omega) = K_\mathrm{th}(T)\delta(\omega)$ due to the exact conservation of the energy-current operator [@Niemeijer1971; @Zotos97]. Thus, the thermal Drude weight [$K_\mathrm{th}$]{}is nonzero at any finite temperature, and any contribution at finite frequencies vanishes. b) In a real experiment for spin chain materials, one may expect the Drude weight to be broadened into, e.g., a Lorentzian in frequency space by external scattering. $\tau$ is the inverse width of such a Lorentzian, and it is related to the life time of the current, or the inverse scattering rate, respectively. Reproduced from [@Heidrich_diss].[]{data-label="fig:kth_width"}](fig_Kth_tau){width="\textwidth"} The thermal conductivity of spin chains appears particularly interesting because unlike the situation in the previously discussed cases of spin planes and spin ladders, rigorous theoretical predictions exist, in particular for the spin-1/2 Heisenberg chain model which is assumed to be relevant for the cuprate spin chains discussed here. The most spectacular one is certainly the prediction of ballistic, i.e. dissipationless, heat transport which would result in a diverging heat conductivity [@Zotos97], which can be expressed in the form of a delta peak at zero frequency in the real part of the frequency-dependent thermal conductivity, i.e., $ \operatorname{Re}~\kappa(\omega) = K_\mathrm{th}(T)\delta(\omega)$ [@Heidrich_diss]. Moreover, the temperature dependence of the thermal Drude weight [$K_\mathrm{th}$]{}has been calculated *exactly* [@Kluemper2002]. Thus, in experimental investigations of the heat conductivity of spin-1/2 Heisenberg chain materials, one should expect to find signatures of the extraordinary properties of the model, though external scattering processes should complicate the picture. At low temperatures $T\lesssim 0.15 J/k_B$, i.e., $T\lesssim 300$ [K]{} for $J/k_B\sim2000~\mathrm{K}$, the thermal Drude weight depends linearly on temperature, where [@Kluemper2002; @Heidrich02; @Heidrich03; @Hess2007] $$K_\mathrm{th}=\frac{(\pi k_B)^2}{3\hbar}vT\,, \label{eq:kth}$$ with the velocity $v$ of the spinons at long wave lengths. In a real material, one might qualitatively expect that the external scattering processes cause the delta peak to broaden into, e.g., a Lorentzian in frequency space [@Heidrich_diss] with a width $1/\tau$, i.e. the scattering rate of the heat current (see Fig. \[fig:kth\_width\]). Hence, the heat conductivity of a single chain, $\tilde{\kappa}_\mathrm{mag}$, is rendered finite and may be approximated by $\tilde{\kappa}_\mathrm{mag}= K_\mathrm{th} \,\tau /\pi$ [@Hess2007]. If one naturally relates the experimental spinon mean free path [$l_\mathrm{mag}$]{}with the relaxation time $\tau$ via $l_\mathrm{mag}=v\tau$ and considers the number of spin chains per unit area $n_s$, one obtains [@Hess2007] $$l_\mathrm{mag}=\frac{3}{\pi}\frac{\hbar}{k_B^2n_s}\frac{\kappa_\mathrm{mag}}{T}\,,\label{eq:lmag_chain}$$ It is interesting and important to note, that the kinetic model in one dimension (Eq. \[eq:kinetic\_1D\]), if one uses a Fermi distribution in order to account for the fermionic character of the spinons, yields [@Hess2007b; @Sologubenko01] $$l_\mathrm{mag}=\frac{\pi}{2}\frac{\hbar}{k_B^2n_s}\frac{\kappa_\mathrm{mag}}{T}\left[\int_0^\frac{J\pi}{2k_BT}x^2\frac{\exp(x)}{(\exp(x)+1)^2}dx\right]^{-1}. \label{eq:lmag_chain_kinetic}$$ The integral in (\[eq:lmag\_chain\_kinetic\]) is only weakly temperature dependent and approaches $\pi^2/6$ for $T\rightarrow0$. Hence, at low temperatures $k_BT\ll J$, which with $J/k_B\sim2000~\mathrm{K}$ holds even at room temperature one obtains the same result as in the case of the Drude weight approach, i.e. Eq. \[eq:kth\] and \[eq:lmag\_chain\] [@Hess2007; @Hess2007b]. Spinon heat transport in ’dirty’ spin chains -------------------------------------------- ![$\kappa_a$ ($\square$), $\kappa_b$ ($\bigcirc$) and $\kappa_c$ ($\triangle$) of [$\mathrm{CaCu_{2}O_{3}}$]{}as a function of temperature. The dashed and solid lines represent a linear fit of the experimental data in the range $100\dots300$ [K]{} and the estimated $\kappa_\mathrm{mag}$ in this range. Extrapolations of $\kappa_\mathrm{mag}$ towards low temperatures (assuming a temperature independent magnetic mean free path as extracted for $T>100~\mathrm{K}$) corresponding to a finite ($\Delta=3$ [meV]{}) and a vanishing spin gap are represented by dotted and dashed-dotted lines, respectively. Figure reproduced from [@Hess2007].[]{data-label="fig:kappa_123"}](fig_kappa_123){width="60.00000%"} We start our discussion of the spinon heat conductivity in cuprate spin chain materials with the results for [$\mathrm{CaCu_{2}O_{3}}$]{}for which we have already briefly discussed its signatures of magnetic heat conductivity in Section \[sec:signatures\] (Fig. \[fig:signatures\]a, see [@Hess2007] for details on this study). The $\mathrm{Cu_2O_3}$ planes in this material actually form, like in [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}, a two-leg ladder structure. However, due to strongly bent [Cu-O-Cu]{} bonds along the rungs (bonding angle $\sim 123^\circ$) the rung exchange interaction is believed to be strongly reduced ($J_\bot/k_B\sim 100\dots300~\mathrm{K}$) as compared to the ladder legs ($J_\Vert\sim2000~\mathrm{K}$), and all other exchange interactions present in the compound are expected to be similarly small or even smaller [@Goiran06; @Kim03; @Kiryukhin01]. Thus, it is reasonable to consider this material rather as a chain compound with the chains running along the crystallographic $b$-axis. Indeed, INS confirms the absence of a spin gap for energies above $\sim3$ [meV]{} and excitation spectra which are compatible with weakly coupled spin-1/2 Heisenberg spin chains [@Lake2010]. Concerning its crystal structure, the material is rather disordered due to an inherent significant Ca and oxygen deficiency being balanced by excess Cu [@Ruck01; @Kim03]. We shall see that this disorder has a substantial impact on both the phononic as well the magnetic heat transport properties. Fig. \[fig:kappa\_123\] shows once more the heat conductivity of [$\mathrm{CaCu_{2}O_{3}}$]{}as a function of temperature, for all principal axes of a crystal (see [@Hess2007], for all details). $\kappa_a$ and [$\kappa_c$]{}were measured perpendicular to the chains (thus both are labeled [$\kappa_\bot$]{}in Fig. \[fig:signatures\]) and $\kappa_b$ was measured parallel to the chains. Since the material is electronically insulating, electronic heat conduction is negligible and we, therefore, expect these components to originate from phononic heat conduction plus a possible magnetic contribution. The thermal conductivity perpendicular to the chain direction is typical for a strongly suppressed phononic thermal conduction with a high phonon scattering rate [@Berman50; @Berman]: $\kappa_a$ and $\kappa_c$ exhibits only a weak $T$-dependence and possess absolute values ($\lesssim 4~\rm Wm^{-1}K^{-1}$) which are much smaller than the phonon heat conductivity [$\kappa_\mathrm{ph}$]{}of other chemically undoped chain or ladder cuprates such as [$\mathrm{SrCuO_{2}}$]{}, [$\mathrm{Sr_2CuO_3}$]{}(see further down), or [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}(Fig. \[fig:signatures\]b). The thermal conductivity perpendicular to the chains can, therefore, be considered to be purely phononic and the strong suppression is naturally explained as a direct consequence of the strong off-stoichiometry of [$\mathrm{CaCu_{2}O_{3}}$]{}. The completely different behavior of [$\kappa_b$]{}, which is measured *parallel* to the chains, in particular, the strong increase at $T\lesssim50~\mathrm{K}$ must arise from magnetic heat conduction in the chains. Such a strong increase of $\kappa$ with rising $T$ cannot be understood in terms of conventional phonon heat conduction by acoustic phonons, and also thinkable contributions from dispersive optical phonons which possibly play a role in the $T$ dependence of $\kappa_a$ can be excluded [@Hess2007]. Interestingly, the high-temperature increase of $\kappa_b$ turns into *linear* in $T$ for $T\gtrsim100~\mathrm{K}$. It is immediately clear that this linearity in temperature not only holds for the total thermal conductivity $\kappa_b=\kappa_\mathrm{ph}+\kappa_\mathrm{mag}$ but in particular also for the magnetic contribution [$\kappa_\mathrm{mag}$]{}because [$\kappa_\mathrm{ph}$]{}is only weakly temperature dependent as can be inferred from the phononic [$\kappa_a$]{}and [$\kappa_c$]{}. Therefore, the magnetic thermal conductivity of [$\mathrm{CaCu_{2}O_{3}}$]{}at $T\gtrsim100~\mathrm{K}$ apparently perfectly follows the expected temperature dependence of the thermal Drude weight [$K_\mathrm{th}$]{}as given in Eq. \[eq:kth\]. ![Temperature dependences of the thermal conductivities of [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}along the $a$, $b$, and $c$ axes. The dotted lines represent estimated limits of the thermal conductivity due to the finite size of the samples. The solid and dashed lines represent different evaluations of the phonon contribution, see [@Sologubenko01], where the figure has been adapted from, for details.[]{data-label="fig:kappa_SCO112_213_Solo"}](fig_kappa_SCO112_213_Solo){width="60.00000%"} Hess et al. have investigated this result further and extracted in a simple procedure the magnetic heat conductivity by presuming that the observed slope in the data is that of [$\kappa_\mathrm{mag}$]{}[@Hess2007]. As can be seen in Fig. \[fig:kappa\_123\], a linear fit in the range $100\dots300$ [K]{} (dashed line) describes the data almost perfectly and thus yields $\kappa_{\mathrm{ph},b}=1.2~\mathrm{Wm^{-1}K^{-1}}$ and $\kappa_\mathrm{mag}=0.055{\rm ~Wm^{-1}K^{-2}}\times T$ in this temperature regime. The extracted $\kappa_\mathrm{mag}$ for $T>100~\mathrm{K}$ is plotted in Fig. \[fig:kappa\_123\] as a solid line. $\kappa_\mathrm{mag}$ at lower $T$ cannot be inferred from the data. Obviously, the mean free path of such a linear [$\kappa_\mathrm{mag}$]{}should be temperature independent in the framework of the model described above. Comparison with Eq. \[eq:lmag\_chain\] yields $l_\mathrm{mag}=22\pm 3$ [Å]{} for the entire range $100\dots300$ [K]{} corresponding to about 6 lattice spacings [@Hess2007]. Such a mean free path appears to be is extremely short, in particular in view of the much larger mean free paths which were obtained on [$\mathrm{La_2CuO_4}$]{}and [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}(see the previous sections). The large impurity density which is implied by it is however qualitatively consistent with the fact that the material is intrinsically disordered. Furthermore, it is even quantitatively consistent with an independent measurement of the impurity density of the material, which in this special case is possible via the magnetic susceptibility (see [@Hess2007], for details). A temperature-independent [$l_\mathrm{mag}$]{}as is found here seems at first glance surprising. It may indicate, however, that spinon-impurity scattering arising from effective chain cuts through the impurities is dominating over all other scattering processes, in particular, spinon-phonon scattering. In the Boltzmann-type view that has been introduced above one may understand this, provided Matthiessen’s rule holds, by assuming $$l_\mathrm{mag}^{-1}=l_0^{-1}+l_\mathrm{sp}(T)^{-1}+l_\mathrm{ss}(T)^{-1}\, , \label{eq:matthiesen_allg}$$ where $l_{sp}$ and $l_\mathrm{ss}$ stand for *a priori* thinkable temperature-dependent spinon-phonon and spinon-spinon scattering lengths, respectively. A constant [$l_\mathrm{mag}$]{}may then naturally arise if $l_0 \ll l_{sp},\, l_\mathrm{ss}$, a situation which seems to be plausible in the situation of a ’dirty’ spin chain as obviously is realized in [$\mathrm{CaCu_{2}O_{3}}$]{}. In order to experimentally obtain access to these apparently masked scattering processes it seems straightforward to study [$\kappa_\mathrm{mag}$]{}of spin chain materials which are chemically cleaner than [$\mathrm{CaCu_{2}O_{3}}$]{}and where $l_0 \gg l_{sp},\, l_\mathrm{ss}$ can expected to be realized. In this situation, which shall be discussed in the next section, it should be possible to study the relevance of spinon-phonon and spinon-spinon scattering directly. ’Ballistic’ spinon heat transport in ’clean’ spin chains -------------------------------------------------------- The spin chain compounds [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}are both considered as being excellent realizations of the spin-1/2 Heisenberg model and are not prone to the stoichiometric problems that arise in [$\mathrm{CaCu_{2}O_{3}}$]{}. They are therefore ideal candidates to further studying relevant spinon scattering processes. Sologubenko et al. were the first to study the heat conductivity of these materials already in 2000 and 2001 [@Sologubenko00a; @Sologubenko01]. Fig. \[fig:kappa\_SCO112\_213\_Solo\] presents their data for the heat conductivity of these materials measured along all crystal directions. As can be seen in the figure, unlike the previously discussed cases, there is no clear double-peak structure or high-temperature increase visible in [$\kappa_\Vert$]{}. Yet, there obviously is an enhancement which develops in [$\kappa_\Vert$]{}at elevated temperature and causes a significant anisotropy between [$\kappa_\Vert$]{}and [$\kappa_\bot$]{}. Sologubenko already at that time concluded the presence of substantial spinon heat conduction in these materials based on these data. The solid and dashed lines in the Figure represent two different approaches to evaluate the phononic contribution to [$\kappa_\Vert$]{}. The detailed evaluation and analysis of the resulting spinon heat conductivity (see [@Sologubenko01]) is omitted here because more recent data allow a more precise view on the matter. These new data emerged recently by two groups who independently from each other discovered that the heat conductivity in [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}depends crucially on the chemical purity of sample, where in particular the spinon heat conductivity [$\kappa_\mathrm{mag}$]{}turned out to be extremely sensitive to chemical impurities [@Hlubek2010; @Hlubek2012; @Kawamata2010; @Kawamata2008]. ![[$\kappa_a$]{}and [$\kappa_c$]{}of [$\mathrm{SrCuO_{2}}$]{}for different purity values. Closed (open) symbols represent $c$ axis ($a$ axis) data, red circles (black diamonds) correspond to ’$4N$’ (’$2N$’) purity. Inset: The same data as in the main panel in double-logarithmic representation. Adapted from [@Hlubek2010].[]{data-label="fig:srcuo2_purity"}](fig_srcuo2_purity){width="60.00000%"} ### Ballistic spinon heat transport in SrCuO$\bf _2$ {#ballistic-spinon-heat-transport-in-srcuobf-_2 .unnumbered} Hlubek et al. [@Hlubek2010] and Kawamata et al. [@Kawamata2010] studied independently from each other the effect of chemical purity on the magnetic heat conductivity of double-chain compound [$\mathrm{SrCuO_{2}}$]{}, with very similar results. Fig. \[fig:srcuo2\_purity\] presents the thermal conductivities [$\kappa_\Vert$]{}and [$\kappa_\bot$]{}of [$\mathrm{SrCuO_{2}}$]{}which were obtained on single crystals with 99% (’$2N$’) and 99.99% (’$4N$’) chemical purity as obtained by Hlubek et al. [@Hlubek2010].[^8] Hlubek et al. report for both [$\kappa_\bot$]{}and [$\kappa_\Vert$]{}of the ’$2N$’ sample very similar results as previously found by Sologubenko et al. (see Fig. \[fig:kappa\_SCO112\_213\_Solo\]). However, they observed a drastically enhanced spinon heat conductivity of the samples with the higher purity (’$4N$’) [@Hlubek2010]: The thermal conductivity parallel to the spin chains of the ’$2N$’ sample, $\kappa_{\Vert,2N}$, exhibits a low-temperature peak at $\sim20~\mathrm{K}$ and a shoulder at $T\gtrsim40~\mathrm{K}$. This shoulder and the fact that $\kappa_{\Vert,2N}$ remains at all temperatures much larger than the thermal conductivity perpendicular to the chains $\kappa_{\bot,2N}$ (the figure shows data for the $a$-direction $\kappa_{a,2N}$ only, because data along the $b$ direction are practically identical [@Hlubek2010], see also Fig. \[fig:srcuo2\_purity\]) have been interpreted as the primary signatures of magnetic heat transport in the compound [@Sologubenko01]. Interestingly, for the ’$4N$’ sample, the heat transport perpendicular to the chains, $\kappa_{\bot,4N}$, is only slightly enhanced as compared to $\kappa_{\bot,2N}$ which is indicative of a weakly reduced phonon-defect scattering. In contrast, the heat transport parallel to the chains, $\kappa_{\Vert,4N}$ is drastically enhanced exhibiting a broad peak centered at $\sim28~\mathrm{K}$. At all temperatures $\kappa_{\Vert,4N}>\kappa_{\Vert,2N}$ up to room temperature. At low temperature, the enhancement is largest (exceeding a factor of 2 at $T\lesssim 70~\mathrm{K}$) where both curves approach each other at further increased temperature. Kawamata et al. pointed out that not only the chemical purity of the growth materials but also the oxygen stoichiometry, which can be influenced by annealing the samples, has a strong influence on the heat conductivity of [$\mathrm{SrCuO_{2}}$]{}parallel to the chains [@Kawamata2010]. Fig. \[fig:SrCuO2\_Kawamata\] shows corresponding data for samples with ’$3N$’ and ’$4N$’ chemical purity. As can be inferred from the figure, the heat conductivity parallel to the chains of as-grown samples is strongly enhanced by annealing the samples in $\mathrm O_2$ atmosphere. In contrast to this, the annealing procedure leads to a decrease of [$\kappa_a$]{}of the ’$4N$’ sample (see Figure). Note that the data for [$\kappa_c$]{}of the annealed ’$4N$’ sample are very similar to those of Hlubek et al. while only a very rough agreement is found for [$\kappa_a$]{}(see Fig. \[fig:srcuo2\_purity\]). The data by Hlubek et al. have been obtained on as-grown samples for which the annealing process lead to slight enhancement of the phonon heat conductivity [@Hlubekdiss]. Hlubek et al. drew several qualitative conclusions from their data [@Hlubek2010]: i) the increase of [$\kappa_\Vert$]{}upon enhancing the chemical purity in contrast to the negligible increase found in [$\kappa_\bot$]{}implies that the increase of [$\kappa_\Vert$]{}primarily concerns the magnetic heat conductivity [$\kappa_\mathrm{mag}$]{}. ii) the extreme low-temperature sensitivity to impurities of [$\kappa_\mathrm{mag}$]{}suggests that spinon-defect scattering is the dominating process for dissipating the heat current in this regime. iii) upon rising the temperature the spinon-defect scattering is increasingly masked by a further scattering process which leads to $\kappa_{\Vert,4N}$ and $\kappa_{\Vert,2N}$ approaching each other. Hlubek et al. suggested spinon-phonon scattering as the most reasonable candidate for the latter process, since phonons are inevitably present, and the only thinkable alternative, spinon-spinon scattering, is expected to be unimportant for the heat transport of the $S=1/2$ Heisenberg chain model [@Zotos97; @Kluemper2002; @Heidrich03; @Heidrich05]. ![Temperature dependence of the thermal conductivity along the $c$-axis (parallel to spin-chains, [$\kappa_c$]{}) and along the $a$-axis (perpendicular to spin-chains, [$\kappa_a$]{}) for as-grown and $\mathrm O_2$-annealed single-crystals of [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}grown from raw materials with ’$3N$’ and ’$4N$’ purity. Solid lines are fitting results for [$\kappa_a$]{}using the Callaway model. Reproduced from [@Kawamata2010][]{data-label="fig:SrCuO2_Kawamata"}](fig_SrCuO2_Kawamata){width="60.00000%"} Again assuming $ \kappa_\Vert=\kappa_\mathrm{ph}+\kappa_\mathrm{mag}$, the phonon heat conductivity parallel to the spin chains [$\kappa_\mathrm{ph}$]{}may be reasonably approximated from the measured [$\kappa_\bot$]{}, i.e., $\kappa_\mathrm{ph}\approx\kappa_\bot$ where a certain amount of anisotropy should be taken into account [@Sologubenko01; @Hlubek2010]. Thus, [$\kappa_\mathrm{mag}$]{}can again be estimated from the measured data by $\kappa_\mathrm{mag}=\kappa_\Vert-\kappa_\bot$. The relatively large anisotropy of the heat conductivity of the ’$4N$’ sample implies that any error in [$\kappa_\mathrm{ph}$]{}has only very little effect on [$\kappa_\mathrm{mag}$]{}, as long as $\kappa_\mathrm{mag}\gg\kappa_\bot$. ![[$\kappa_\mathrm{mag}$]{}of [$\mathrm{SrCuO_{2}}$]{}for different purities. Open symbols represent low-$T$ [$\kappa_\mathrm{mag}$]{}which is disregarded in the further analysis. The shaded areas show the uncertainty of the estimation of [$\kappa_\mathrm{mag}$]{}due to an assumed relative error of 30% for the phononic background. Inset: The same data as in the main panel in double-logarithmic representation. Adapted from [@Hlubek2010].[]{data-label="fig:srcuo2_kmag"}](fig_srcuo2_kmag){width="60.00000%"} Fig. \[fig:srcuo2\_kmag\] shows the thus extracted spinon heat conductivity [$\kappa_\mathrm{mag}$]{}of both the ’$2N$’ and the ’$4N$’ samples as a function of temperature $T$ [@Hlubek2010]. Both curves exhibit the characteristic peak structure which qualitatively reflects the competition between the increasing occupation of quasiparticle states and the growing importance of temperature dependent scattering processes with rising temperature [@Berman]. In the regime of the low-temperature increase, the latter are unimportant and only temperature independent boundary scattering dominates. The increase then is the result of the growing number of thermally excited spinons. The most reasonable candidate for the temperature dependent scattering is of course, as mentioned above, the spinon-phonon scattering. Within this scenario, the difference between both curves arises naturally from the different purity of both compounds which yields a corresponding different importance of boundary scattering. This corresponds very well with the concept that a single impurity within a chain serves as a boundary in one dimension. Note, that the peak value of $\sim660$ [$\mathrm{Wm^{-1}K^{-1}}$]{}of the ’$4N$’ is to the best of our knowledge a ‘record‘ as it exceeds the largest reported magnetic heat conductivities [@Sologubenko01; @Hess01] by more than a factor of 3. ![Magnetic mean-free paths of [$\mathrm{SrCuO_{2}}$]{}for different purities. The solid lines were calculated according to Eq. \[eq:matthiesen\_spez\]. The shaded area illustrates the uncertainty from the estimation of the phononic background. Adapted from [@Hlubek2010].[]{data-label="fig:srcuo2_lmag"}](fig_srcuo2_lmag){width="60.00000%"} A more direct view on the temperature dependence of spinon scattering processes is provided by the spinon mean free path that can computed according to Eq. \[eq:lmag\_chain\]. Resulting data are shown in Fig. \[fig:srcuo2\_lmag\]. Both data sets show a strong decrease with increasing temperature, consistent with the above expectation of a growing importance of spinon scattering with increasing temperature. At the same time, the different magnitude of the curves which is most pronounced at lowest temperature signals the different importance of impurity scattering. Following a conjecture of Sologubenko et al. [@Sologubenko01], the data in Fig. \[fig:srcuo2\_lmag\] have been modeled by Hlubek et al. using Matthiessen’s rule for the *extrinsic* spinon-defect and spinon-phonon scattering, i.e., employing a reduced version of Eq. \[eq:matthiesen\_allg\], where the *intrinsic* spinon-spinon scattering as expressed by $l_\mathrm{ss}$ has been discarded: $$l_\mathrm{mag}^{-1}=l_0^{-1}+l_\mathrm{sp}(T)^{-1}\, . \label{eq:matthiesen_spez}$$ An empirical expression has been used to model the spinon-phonon scattering, which represents a general umklapp process [@Sologubenko01; @Hlubek2010]: $$l_\mathrm{sp}(T)=\frac{\exp(T^*_u/T)}{A_sT}\, . \label{eq:spin-phonon_emp}$$ Here, $T^*_u$ represents the energy scale of the scattering phonons, and $A_s$ a scattering cross section. The solid lines in Fig. \[fig:srcuo2\_lmag\] represent fits to the data employing Eqs. \[eq:matthiesen\_spez\] and \[eq:spin-phonon\_emp\], where for both purity levels the same $T^*_u$ has been used [@Hlubek2010]. This reasonable assumption accounts for the expectation that the relevant energy scale for the scattering is the same for both purity levels. The other parameters $l_{0,2N}$, $l_{0,4N}$, $A_{s,2N}$, $A_{s,4N}$ were employed as free fit parameters. Ideally, the purity should be solely captured by the spinon-defect scattering lengths $l_{0,2N}$, $l_{0,4N}$, since one would expect also $A_{s,2N}\approx A_{s,4N}$, where deviations between these two values (up to 30%) should be allowed for in order to compensate geometrical errors of the individual measurements of the heat conductivity. As can be inferred from the figure, these constraints allow excellent fits. Hlubek et al. report the energy scale $T^*_u\sim200~\mathrm{K}$ which is of the same order of magnitude as the Debye temperature of the material, corroborating spinon-phonon scattering as being the dominant temperature dependent process. The analysis yields further the spinon-defect scattering lengths $l_{0,2N}\approx300$ [nm]{} and $l_{0,4N}\approx 1.6~\mu$[[m]{}]{}, which correspond to more than 750 and 4100 lattice spacings, respectively [@Hlubek2010]. This result is remarkable in two aspects: i) it suggests that indeed only the extrinsic spinon-defect and spinon-phonon scattering are relevant for relaxing the heat current. Thus the findings provide an experimental corroboration of the theoretical prediction of *ballistic* heat transport in the $S=1/2$ antiferromagnetic Heisenberg chain. ii) in the high-purity sample, an extraordinary scattering length of more than one micrometer is apparently present at low temperature, which apparently is only limited by the impurities in the chains. This suggests that much larger spinon mean free paths could be achievable in perfectly clean crystals of [$\mathrm{SrCuO_{2}}$]{}. ### Ballistic spinon heat transport in Sr$\bf _2$CuO$\bf _3$ {#ballistic-spinon-heat-transport-in-srbf-_2cuobf-_3 .unnumbered} After the pioneering experiments of Sologubenko et al. on the single chain material [$\mathrm{Sr_2CuO_3}$]{}[@Sologubenko00a; @Sologubenko01] with 99% purity (labeled ’$2N$’ hereafter), the spinon heat transport in this compound has been under scrutiny by two further studies. Kawamata et al. investigated the heat transport of the pure material on a crystal where chemicals with 99.9% purity (labeled ’$3N$’ hereafter) had been used for the crystal growth [@Kawamata2008]. Hlubek et al. investigated a crystal with even higher purity (99.99% purity of the starting chemicals, labeled ’$4N$’ hereafter), and performed a comparative study of the effect of the different purity levels [@Hlubek2012]. ![Thermal conductivity of [$\mathrm{Sr_2CuO_3}$]{}parallel to the spin chains ($\kappa_b$) for various purities. The dashed lines represent results from Sologubenko et al. with ’$2N$’ purity, reproduced from [@Sologubenko00a]. The dash-dotted line has been obtained by Kawamata et al for ’$3N$’ purity and is reproduced from [@Kawamata2008]. Inset: thermal conductivity of [$\mathrm{Sr_2CuO_3}$]{}perpendicular to the spin chains ($\kappa_c$) for ’$2N$’ (also reproduced from [@Sologubenko00a]) and ’$4N$’ purity. The solid lines are fits to the Callaway model [@Callaway59]. Adapted from [@Hlubek2012].[]{data-label="fig:kappa_213"}](Sr2CuO3-Purity-kappav1){width="60.00000%"} Fig. \[fig:kappa\_213\] depicts the corresponding data for the thermal conductivity parallel to the spin chains ($\kappa_\Vert$, main panel), and except for the $3N$ purity, perpendicular to them ($\kappa_\bot$, inset). $\kappa_\bot$, like in [$\mathrm{SrCuO_{2}}$]{}exhibits a temperature dependence that is characteristic for phononic heat conductivity. The height of the characteristic phononic peak at $T \approx 22~\mathrm{K}$ sensitively depends on the density of impurities in the system, which generate phonon-defect scattering. In fact, the data for both purities can be described well in the framework of a model by Callaway [@Callaway59], where the difference between both curves is largely captured by different point defect scattering strength (see Reference [@Hlubek2012] for details). Very similarly as described above for [$\mathrm{SrCuO_{2}}$]{}, the spinon heat conductivity has been obtained via $\kappa_\mathrm{mag} = \kappa_\Vert - \kappa_\bot$ for the ’$4N$’ sample, which is shown together with similarly obtained results from References [@Sologubenko00a; @Sologubenko01; @Kawamata2008] for ’$3N$’ and ’$2N$’ in Fig. \[fig:kappamag\_213\]. The very similar temperature dependence of [$\kappa_\mathrm{mag}$]{}as compared to that of [$\mathrm{SrCuO_{2}}$]{}, though with somewhat smaller values is evident. Hlubek et al. pointed out that the high-temperature decay of [$\kappa_\mathrm{mag}$]{}can be fit well with $\kappa_\mathrm{mag}\sim\exp(T^*/T)$, with $T^*$ a characteristic energy scale, which can be related to prevailing umklapp processes [@Hlubek2012]. ![Estimated magnetic thermal conductivity of [$\mathrm{Sr_2CuO_3}$]{}for ’$4N$’ (circles, squares), ’$3N$’ (dash-dotted line) [@Kawamata2008], and ’$2N$’ (dashed line) [@Sologubenko00a] purity. The shaded area illustrates the uncertainty from the estimation of the phononic background. The ’$4N$’ results shown in squares instead of circles have a large uncertainty. The thick solid line is a fit for $T > 80~\mathrm{K}$ with $\kappa_\mathrm{mag,fit1}\sim1/T^{2}$. The thin solid line is a fit with $\kappa_\mathrm{mag,fit2}\sim\exp(T^*/T)$. See Ref. [@Hlubek2012] where the figure has been reproduced from for a further discussion.[]{data-label="fig:kappamag_213"}](Sr2CuO3-Purity-kmagv3){width="60.00000%"} Hlubek et al. extracted from the data in Fig. \[fig:kappamag\_213\] the spinon mean free path following Eq. \[eq:lmag\_chain\], and analyzed the temperature dependence according to Eqs. \[eq:matthiesen\_spez\] and \[eq:spin-phonon\_emp\]. Remarkably, the analysis work equally well (see Fig. \[fig:lmag\_213\]) as before for [$\mathrm{SrCuO_{2}}$]{}with $l_0 = 0.54 \pm 0.05~\mu\mathrm{m}$ (corresponding to approximately 1370 lattice spacings), $T^*_u= 210 \pm 11~\mathrm{K}$ [@Hlubek2012]. It is important to point out that the energy scale for the phonon scattering $T^*_u$ is practically the same for both [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}. This is consistent with the fact that in both materials the spin chains are formed by practically the same $\mathrm{CuO_2}$ plaquettes, suggestive of a very similar local spin-phonon coupling. The low-temperature mean free path limit, $l_0$, is significantly smaller for [$\mathrm{Sr_2CuO_3}$]{}than that of [$\mathrm{SrCuO_{2}}$]{}. For a direct quantitative comparison, Fig. \[fig:lmag\_213\] shows [$l_\mathrm{mag}$]{}of both compounds in relation to the number of lattice spacings. Despite the same formal chemical purity, the double chain compound’s mean free path is about a factor of three larger, consistent with a relatively higher chemical stability of the compound [@Hlubek2012]. Hlubek et al. attempted for the first time a deeper insight into the spin-phonon interaction by calculating $l_\mathrm{sp}$ directly from a spin-phonon scattering theory for the XY-limit of the Heisenberg model [@Louis2006] with promising results at high temperature. However at lower temperature, the agreement between the experimental results and the theoretical $l_\mathrm{sp}$ becomes less satisfactory (see [@Hlubek2012] for details). Very recently, Chernyshev and Rozhkov reanalyzed the experimental data of Hlubek et al. for both [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}with an alternative theoretical model which specifies $l_\mathrm{sp}$ including realistic numerical values for the spin-phonon interaction [@Chernyshev2016]. Also in this case, the analysis relies for both compounds on just two types of scattering processes, viz. spinon-defect and spinon-phonon scattering. Thus the earlier statement for a strong experimental confirmation of ballistic heat transport holds for both materials. ![Magnetic mean free paths of [$\mathrm{SrCuO_{2}}$]{}(pentagon shape) and [$\mathrm{Sr_2CuO_3}$]{}(circles) for ’$4N$’ purity. The shaded area illustrates the uncertainty from the estimation of the phononic background. The lines are fits to the mean free paths. Figure taken from [@Hlubek2012].[]{data-label="fig:lmag_213"}](SrCuO2_Sr2CuO3_l_mag_Comparison_v5){width="60.00000%"} Spin chains with doping induced disorder ---------------------------------------- The ballistic transport in the isotropic Heisenberg chain implies the natural question about the impact of subtle disorder on the transport properties. Such disorder should break the integrability of the spin model and thus a very substantial impact on the spinon heat conductivity is expected. One possibility to approach this problem in [$\mathrm{SrCuO_{2}}$]{}or [$\mathrm{Sr_2CuO_3}$]{}is to substitute Ca for Sr in small amounts, thus creating bond disorder due to the different ionic radii of Sr$^{2+}$ and Ca$^{2+}$. It turns out that this disorder type has a strong impact not only on the spinon heat transport but also on the ground state of the spin chains as is evidenced by Nuclear Magnetic Resonance (NMR) measurements [@Ribeiro2005; @Hammerath2011; @Hammerath2014; @Hlubek2011; @Mohan2014]. Hammerath et al. performed NMR measurements at the $^{63}\mathrm{Cu}$ nucleus for [$\mathrm{Sr_{1-x}Ca_xCuO_{2}}$]{}at $x=0, 0.1$ and $\mathrm{Sr_{1.9}Ca_{0.1}CuO_3}$ determined the spin-lattice relaxation rate $1/T_1$ as a function of temperature [@Hammerath2011; @Hammerath2014]. The data very clearly reveal a doping induced spin gap which according to Hammerath et al. is solely induced by the bond disorder on the intrachain coupling [@Hammerath2014]. It should be noted, however, that a priori, it cannot be excluded that a finite amount of the doped Ca, instead of replacing the Sr in the structure, replaces Cu inside the $\mathrm{CuO_2}$-chain structures. In this case, a more severe site disorder would be the result which cuts the spin chains (which in the clean compounds possess a large average length of several thousand unit cells, according to the spinon mean free path, see Fig. \[fig:lmag\_213\]). Indeed, a specific investigation of such a site disorder has been addressed by substituting Ni for Cu *inside* the $\mathrm{CuO_2}$-chain structures of [$\mathrm{Sr_2CuO_3}$]{}as well as [$\mathrm{SrCuO_{2}}$]{}[@Utz2015; @Simutis2013] in NMR measurements and INS measurements. Both approaches led to the conclusion that finite size effects play an important role, i.e. the Ni effectively cuts the chains into finite segments which exhibit a gapped ground state. Remarkably, the Ca-induced disorder has a dramatic impact on the heat conductivity parallel to the spin chains [$\kappa_\Vert$]{}, while the phononic heat conductivity [$\kappa_\bot$]{}is only moderately affected. This implies a very strong suppression of the spinon heat conductivity [$\kappa_\mathrm{mag}$]{}upon increasing the disorder. This is illustrated in Fig. \[fig:kappa\_112\_Ca\]a for the Ca-doped double-chain compound, i.e. [$\mathrm{Sr_{1-x}Ca_xCuO_{2}}$]{}[@Hlubek2011]. A very similar result is obtained for Ca-doped [$\mathrm{Sr_2CuO_3}$]{}(not shown) [@Mohan2014]. ![a) Thermal conductivity $\kappa_c$ parallel to the spin chain ($\kappa_\Vert$) for [$\mathrm{Sr_{1-x}Ca_xCuO_{2}}$]{}at different doping levels. Inset: Thermal conductivity $\kappa_a$ for the same doping levels perpendicular to the spin chain ($\kappa_\bot$). b) [$l_\mathrm{mag}$]{}of [$\mathrm{Sr_{1-x}Ca_xCuO_{2}}$]{}for different levels of doping. Toward low temperatures, the error in the estimation of [$\kappa_\mathrm{mag}$]{}increases due to an increase in the phononic heat conductivity. Thus, the values of [$l_\mathrm{mag}$]{}are shown only at temperatures above which the error in [$l_\mathrm{mag}$]{}is reasonably small. The solid lines were calculated according to Eqs. \[eq:matthiesen\_spez\], \[eq:spin-phonon\_emp\]. Figures taken from [@Hlubek2011].[]{data-label="fig:kappa_112_Ca"}](Ca-doped){width="\textwidth"} The spinon heat conductivity can be calculated from the heat conductivity data in exactly the same way as it was already explained for the pure compounds above. Similarly, the temperature dependence of the spinon mean free path for various doping levels can be extracted via Eq. \[eq:lmag\_chain\], resulting in the data shown in Fig. \[fig:kappa\_112\_Ca\]b. Note, that a possible impact of the spin pseudogap can be excluded to have a significant impact on the spinon heat conductivity and the mean free path because at the size of the gap is too small ($\Delta/k_B\sim50~\mathrm{K}$ [@Hammerath2011; @Hammerath2014]) to play a significant role for the spinon heat conductivity which is studied at $T\gtrsim50$ [K]{} only [@Hlubek2011; @Mohan2014]. The data in Fig. \[fig:kappa\_112\_Ca\]b, and corresponding data for [$\mathrm{(Sr_{1-x}Ca_x)_2CuO_3}$]{}can be analyzed remarkably well as in the case of the clean compounds following Eqs. \[eq:matthiesen\_spez\], \[eq:spin-phonon\_emp\]. Thereby, essentially the spinon-defect scattering length $l_0$ turns out as the crucial parameter which is determined by the doping level whereas the phonon scattering term can be set identical for all doping levels (solid lines in Fig. \[fig:kappa\_112\_Ca\]b). It is very instructive to plot the obtained values of the spinon-defect scattering length $l_0$ as a function of the mean distance between two Ca atoms and the inverse of Ca concentration as is shown in Fig. \[fig:lmag\_Ca\]. For the single-chain material [$\mathrm{(Sr_{1-x}Ca_x)_2CuO_3}$]{}, $l_0$ apparently scales perfectly with the inverse of Ca concentration as $l_0 =\frac{2.91\cdot d_0}{x} $, where $d_0 = 3.91~\text{\rm\AA}$ is the lattice spacing between two Cu sites along the chain [@Mohan2014]. This implies that the Ca-induced disorder can perfectly be captured in terms of effective defects in the chain, where the scattering probability per defect is equally strong at all concentrations. ![Spinon-defect scattering length $l_0$ plotted against the inverse of Ca concentration ($x$) (lower abscissa) and the mean distance between two Ca atoms (upper abscissa) for [$\mathrm{SrCuO_{2}}$]{}(open symbols) and [$\mathrm{Sr_2CuO_3}$]{}(filled symbols); the solid lines are linear fits to the data. The ordinate on the right expresses the mean free path in terms of the number of lattice spacings between two Cu sites ($d_0$). Figure taken from [@Mohan2014].[]{data-label="fig:lmag_Ca"}](ca-doped-lmag){width="50.00000%"} Interestingly, the situation is somewhat different for [$\mathrm{Sr_{1-x}Ca_xCuO_{2}}$]{}(see Fig. \[fig:lmag\_Ca\]) [@Hlubek2011]. There, a linear scaling is present, too, however, with an offset: $l_0 =\frac{4.3\cdot d_0}{x} + l_\mathrm{lim}$, with $l_\mathrm{lim}\approx30\cdot d_0$ with $d_0=3.92~\text{\rm\AA}$. This indicates that already at intermediate concentrations ($x = 0.1$) the effect of Ca saturates and the mean free path of spinons is not reduced any further upon increasing the Ca concentration. The observed offset was originally interpreted to be due to a limit set by the disorder-induced long-distance decay of the spin-spin correlation [@Hlubek2011]. However, since the double-chain nature did not enter the latter interpretation, the fact, that the offset is absent in the single-chain material [$\mathrm{(Sr_{1-x}Ca_x)_2CuO_3}$]{}implies a different origin of the offset, which however remains open. Conclusion ========== We have seen that in the here discussed cuprate materials with large antiferromagnetic exchange interaction which realize low-dimensional $S=1/2$ Heisenberg systems in the form of chains, two-leg ladders, and planes, a sizeable magnetic heat conductivity arises. The possibility to extract the magnetic heat conductivity in a very clean manner in principle opens up a new approach for sensitively and comprehensively probing magnetic excitations in these systems. Indeed, for each these quantum systems, using the kinetic model, it is possible to formulate a relatively simple way towards rationalizing the observed temperature dependences of [$\kappa_\mathrm{mag}$]{}and to draw basic conclusions about the involved scattering processes of the heat-carrying magnetic excitations. So far, these considerations have been done individually for the various systems which are very different in terms of their elementary excitations, namely gapless spinons for the one-dimensional chains, triplon excitations with a large excitation gap for the ladders, and spin wave-like excitations in two-dimensions (with only small anisotropy gaps). Important conclusions could be drawn: Quite importantly, the kinetic model apparently yields realistic length scales of the magnetic mean free path [$l_\mathrm{mag}$]{}, despite the simplicity of the model. Based on this finding, the discussed data confirmed the ballistic nature of the heat transport of integrable spin chain systems which truly is a fundamental finding. On the other hand, we have seen that in the non-integrable two-leg spin ladders the magnetic heat transport is nevertheless substantial with [$l_\mathrm{mag}$]{}exceeding the spin-spin correlation length by about three orders of magnitude. ![Comparison of the magnetic heat transport of representative low-dimensional spin systems, namely for the spin chain compounds [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}[@Hlubek2010; @Hlubek2012], the two-leg spin ladder compound [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}[@Hess01], and the 2D-HAF material [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}[@Hess03]. a) Magnetic heat conductivity as a function of temperature, $\kappa_\mathrm{mag}(T)$. b) Magnetic mean free path as a function of temperature, $l_\mathrm{mag}(T)$, in units of lattice spacings.[]{data-label="fig:lmag_all"}](fig_kmag_lmag_comp.png){width="\textwidth"} Beyond these findings for the individual systems, it is interesting to directly compare the magnetic heat transport of these systems and to investigate whether the differences in the nature of the spin system also yield different transport and whether, nevertheless, similarities emerge. Fig. \[fig:lmag\_all\]a shows a collection of data of [$\kappa_\mathrm{mag}$]{}of the chain systems [$\mathrm{SrCuO_{2}}$]{}and [$\mathrm{Sr_2CuO_3}$]{}, of the two-leg ladder material [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}, and of the 2D-HAF [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}. These data can be considered in so far as being representative as they correspond to clean compounds where the up to present largest mean free paths have been extracted [@Hlubek2010; @Hlubek2012; @Hess01; @Hess03]. The data are the same as discussed in the previous sections, but this figure shows them for the first time together on the same scale and indeed reveals remarkable differences. Quite clearly, the chain compound [$\mathrm{SrCuO_{2}}$]{}exhibits a ’record’ value for [$\kappa_\mathrm{mag}$]{}of almost 700 [$\mathrm{Wm^{-1}K^{-1}}$]{}which is many times larger than that of (i) the single chain compound [$\mathrm{Sr_2CuO_3}$]{}and, furthermore, (ii) that of the 2-leg ladder and the 2D-HAF. The first observation (i) has been recognized earlier [@Hlubek2012] and already discussed in the previous section. A closer investigation of the mean free path of these two Heisenberg chain materials as shown in panel (b) of Fig. \[fig:lmag\_all\] reveals that [$l_\mathrm{mag}$]{}apparently can be separated in two regimes. On the one hand, this is at $T\gtrsim100$ K where there is virtually no difference between both curves. Here, according to the analysis, spinon-phonon scattering is dominant, where the data suggest that this type of scattering has no difference in both compounds. On the other hand, at $T\lesssim100$ K [$l_\mathrm{mag}$]{}of [$\mathrm{Sr_2CuO_3}$]{}approaches a much lower low-temperature limit than that of [$\mathrm{SrCuO_{2}}$]{}. Hence, the difference seems to imply a different inherent perfectness of the chains in the two compounds. Upon trying to obtain some further understanding of the more interesting observation (ii), it is important to note that the very large value of [$\kappa_\mathrm{mag}$]{}of [$\mathrm{SrCuO_{2}}$]{}occurs at relatively low temperature $T\lesssim100$ K where [$\kappa_\mathrm{mag}$]{}of both the ladder compound [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}and of the 2D-HAF material [$\mathrm{La_{1.8}Eu_{0.2}CuO_4}$]{}is very small. In both cases, this comparably small low-temperature value of [$\kappa_\mathrm{mag}$]{}can be straightforwardly be understood: for the two-leg ladder material [$\mathrm{Ca_{9}La_5Cu_{24}O_{41}}$]{}, the significant spin excitation gap $\Delta~\sim400$ K prohibits a significant population of triplons below about 100 K. Similarly, for the 2D-HAF, the low-temperature thermal population of magnetic excitations is much ’slower’ than that of the one-dimensional Heisenberg chain. One can convince oneself about this fact from considering the temperature dependence of the magnetic specific heat $c_\mathrm{mag}$ of both systems which is linear in temperature for the chain system whereas the two-dimensional spin plane possesses a quadratic low-temperature increase. Indeed, an estimation of the magnetic specific heat at $T=100$ K (i.e. $k_BT\ll J$) using the estimate $\kappa_\mathrm{mag}\sim c_\mathrm{mag} v_0 l_\mathrm{mag}$ and the respective kinetic expressions for [$\kappa_\mathrm{mag}$]{}in Eqs. \[eq:fit2d\] and \[eq:lmag\_chain\] yields an almost two orders of magnitude smaller $c_\mathrm{mag}$ for the 2D-HAF than that of the Heisenberg chain. Here, the magnetic velocity $v_0\sim Jd/\hbar$ with $d$ the distance between neighboring spins has been used. Thus, taking further into account the mean free path [$l_\mathrm{mag}$]{}at about 100 K being of a similar order of magnitude for all compounds (see further below), it is not surprising that the magnetic heat conductivity of the Heisenberg chain is by far the largest at low temperatures. Fig. \[fig:lmag\_all\]b compares the temperature dependence of the mean free paths [$l_\mathrm{mag}$]{}as extracted from the data in panel (a) upon using the kinetic model. Here one finds that [$l_\mathrm{mag}$]{}of the spin chain compounds additionally exceeds clearly that of the ladder and the plane compounds at $T\lesssim100$ K. These large low-temperature values of [$l_\mathrm{mag}$]{}additionally promote the extraordinarily large [$\kappa_\mathrm{mag}$]{}of the chains in this regime. Another interesting finding as revealed by Fig. \[fig:lmag\_all\]b is the fact that at $T\gtrsim100$ K the mean free path becomes significantly smaller than that of both the 2-leg ladder and of the plane compounds. Since the temperature-induced reduction of [$l_\mathrm{mag}$]{}of the spin chain materials has clearly been assigned to spinon-phonon scattering [@Hlubek2010; @Hlubek2012; @Chernyshev2016], this observation seems to imply that the spinon heat transport of the $S=1/2$ Heisenberg chain is much more prone to scattering of phonons than the two-leg Heisenberg ladder and the 2D-HAF counterparts. It remains to be clarified to what extent this notion is related to the integrability of the Heisenberg chain model, whose ballistic transport properties potentially could exhibit a particular sensitivity to distortions such as those induced by the phonons. Alternatively, differences in the magnetic and phononic spectrum could play an important role, too. Indeed such seems to be a crucial difference in the spin-phonon scattering of the Heisenberg chains where the gapless spinons can be expected to interact with acoustic phonons and that of the 2-leg spin ladders, where the energy of the acoustic phonons is too small to cause an important interaction with the gapped triplon excitations and instead scattering of optical phonons seems to dominate (see Section \[sec:tdepmagnonscatt\]). Finally, it is to be remarked that little is known from experiments about the temperature evolution of [$l_\mathrm{mag}$]{}of the 2D-HAF at elevated temperatures since apparently the data shown in Fig. \[fig:lmag\_all\]b as extracted from Ref. [@Hess03] barely touch this regime. An obvious aspect of future research is thus to extend the achievable temperature regime of reliable [$\kappa_\mathrm{mag}$]{}-data to much higher temperatures and to verify specific predictions about magnon scattering processes involving phonons and the correlation length [@Chernyshev2015]. As it has already been mentioned above, all the here discussed analysis is based on the kinetic model (see Section \[kinmod\]), the successful applicability of which is surprising in view of the strong quantum nature of the here studied spin models. Another possible direction of future research therefore could address ways to model the magnetic heat transport in the studied systems on a more microscopic level. It is worth to mention that the large magnetic heat conductivity values of the spin ladder and the chain systems may also open up new technological developments. The exploitation of the resulting anisotropic heat conductivity tensor for thermal management applications already has been investigated (see e.g. [@Otter2012]). Another, yet unexplored direction would be spin information transport experiments, where electrical pumping and detection of spin currents with the help of the spin-Hall and the inverse spin-Hall effect, as recently successfully used on classical spin systems [@Cornelissen2015]. Upon concluding, it is important to mention further active research directions which could not be touched in this review. One concerns research on $\kappa_{\mathrm{mag}}$ for systems where the magnetic exchange energy becomes comparable or much smaller than the Debye energy, and how it about evolves when the magnetic systems become less quantum in nature, i.e. when $S>1/2$. Initial work has already addressed these aspects. Intriguing heat conductivity results for one-dimensional spin systems with relatively small exchange interactions have been obtained for both organic and inorganic materials, see e.g. [@Sologubenko2007a; @Lue2008; @Sologubenko2009; @Jeon2016; @Steckel2016] for $S=1/2$ systems and [@Sologubenko03; @Sologubenko2008; @Kohama2011; @Sun2013; @Karadamoglou04; @Savin2005; @Kordonis06] for systems with large spin, see also the reviews [@Sologubenko2007; @Zhao2016]. Another dynamic field where magnetic heat transport is being investigated concerns frustrated spin systems, where it promises to become a good probe for accessing possible topological excitations. The class of spin-ice compounds of the type $R_2$Ti$_2$O$_7$ ($R$ a rare earth element) constitutes a subject on its own where the focus is on magnetic monopole-like excitations [@Klemke2011; @Kolland2012; @Toews2013; @Fan2013; @Kolland2013; @Scharffe2015; @Li2015; @Tokiwa2016; @Hirschberger2015; @Toews2018]. Another class concerns highly frustrated layered compounds [@Yamashita2008; @Yamashita2010] where very recently $J_\mathrm{eff}=1/2$-materials with Kitaev interactions came into focus [@Kitaev2006; @Hirobe2017; @Hentrich2018; @Hentrich2018a; @Kasahara2018; @Kasahara2018a] since such systems are conjectured as fascinating avenues for exploring the ’magnetic’ heat transport of topological fractionalized quasiparticles. Acknowledgments {#acknowledgments .unnumbered} =============== I am indebted to Neela Sekhar Beesetty, Wolfram Brenig, Bernd Büchner, David Cahill, Sang-Wook Cheong, Alexander Chernyshev, Stefan-Ludwig Drechsler, Hanan ElHaes, Jochen Geck, Ioannis Giapintzakis, Hans-Joachim Grafe, Franziska Hammerath, Fabian Heidrich-Meisner, Nikolai Hlubek, Vladislav Kataev, Rüdiger Klingeler, Andreas Klümper, Gernot Krabbes, Paul van Loosdrecht, Thomas Lorenz, Oleg Mityashkin, Ashwin Mohan, Satoshi Nishimoto, Peter Prelovšek, Pascal Reutler, Alexandre Revcolevschi, Patrick Ribeiro, Georg Roth, Romuald Saint-Martin, Chinnathambi Sekhar, Robin Steinigeweg, Yannic Utz, Anja Waske, Anja Wolter, Babak Zeini, Andrey Zheludev, and Xenophon Zotos for fruitful discussion or collaborations. Funding: This work was supported by the Deutsche Forschungsgemeinschaft \[grant numbers HE3439/7, HE3439/8, HE3439/9, HE3439/12, HE3439/13, SFB 1143 (project C07)\]; and the European Commission \[grant numbers FP6-032980, PITN-GA-2009-238475\]. References {#references .unnumbered} ========== [199]{} natexlab\#1[\#1]{}\[1\][`#1`]{} \[2\][\#2]{} \[1\][\#1]{} \[1\][[](http://dx.doi.org/#1)]{} \[1\][[](pmid:#1)]{} \[2\][\#2]{} , , , . , , , () . , , () . , , () . , , , , () . , , , , () . , , , , , () . , , , , () . , , , , , , () . , , , () . , , , , , () . , , , , () . , , , , , , , , , () . , , , , , , () . , , , , , , , , () . , , , , , , , , , () . , , , , , , () . , , , , , , , , () . , , , , , , , () . , , , , () . , , , , , , , () . , , , , , , () . , , , , , , () . , , , , , , , , () . , , , , , , , , () . , , , , , , , , , () . , , , , , , , , () . , , , , , , () . , , , , , () . , , , , , , () . , , , , , , , , , () . , , , , , , , () . , , () . , , , , , , , , () . , , , , , , , , , , , , , , () . , , , , , , , , , , , , () . , , , , , , () . , , , , , , , , , () . , , , , , , , , , () . , , , , , , , , () . , , , , , , , , , , , , , () . , , , , , , , , () . , , , , , () . , , , , , () . , , , , () . , , , () . , , , , , () . , , , () . , , () . , , () . , , , , () . , , , () . , , , () . , , () . , , , () . , , , , () . , , , () . , , , () . , , , () . , , , () . , , , , () . , , , , , , () . , , , , , () . , , , () . , , () . , , , () . , , , () . , , () . , , () . , , , , , , , , , () . , , , () . , , , , , () . , , , , () . , , , () . , , , () . , , , , () . , , , , , , , , , , () . , , , , , , , , () . , , , , , , , , , () . , , , , () . , , , () . , , , . , , , , , , , () . , , , , , , , , , () . , , , , () . , , , , , () . , , , , () . , , , , , , , , , , , , () . , , , , , () . , , , () . , , , , () . , , , , , , , () . , , , , , , , , , () . , , () . , , () . , , , , , , , , , , , () . , , , , , , , , , , , , () . , , , , , , , , , , , () . , , , , , () . , , , , , , , , , () . , , , , , , , , () . , , , , , () . , , , , , () . , , , , , , , () . , , , , , , , () . , , , , , , , , , () . , , , , , () . , , () . , , , , , , () . , , , , , , () . , , , , , , , , () . , , , () . , , , , , () . , , , , , , , , , () . , , , , , , () . , , , , , , , , , () . , , Ph.D. thesis, Universität zu Köln, . , , , , , () . , , , , , , , , , , , , , , , , , , , , , , , , () . , , , , , , () . , , , , , () . , , , , , , , , , () . , , , , , , , () . , , , , , , , , () . , , , , , , , , , , , () . , , , , , , , , () . , , , , , , , , () . , , , , , , , () . , , , , , , , , , , , () . , , , () . , , , , , , () . , , , , , , , , , () . , , , , , , () . , , , , () . , , , , , , () . , , , , , , , , , , , , () . , , , , , , , , , , , , , , , , () . , , , , , , , () . , , , , , , , , , , () . , , , , , () . , , , , , () . , , , , , , () . , , , , , , , , , () . , , , , , , , , () . , , () . , , , , , , , , , , () . , , , , , , , , , , , () . , , , , , , , () . , , , , , , , , , , , , , , , (). , , , , , () . , , , . , , , , , , , () . , , , , , () . , , , () . , , Ph.D. thesis, Technische Universität Carolo-Wilhelmina zu Braunschweig, . <http://www.digibib.tu-bs.de/?docid=00001712>. , , , , , , , , , , () . , , , , , , , , , , , () . , , , , , , , , , , () . , , , , , , , () . , , , , () . , , , , , , () . , , , , , , , , () . , , Ph.D. thesis, Technische Universität Dresden, . , , , , () . , , , , , , , , , , () . , , , , , , , , , , , () . , , , , , , , , , , () . , , , , , , , , , , , , , , , () . , , , () . , , , , , , () . , , , , , , , , () . , , , , , , , () . , , , , , , , , () . , , , , , , , , , , () . , , , , , , , () . , , , , , , , , , , , , , () . , , , , , , () . , , , , () . , , , , , () . , , , , , () . , , , , , , , () . , , , , , , , () . , , , , , , , () . , , , , , , , , , () . , , , , , , () . , , , , , , , () . . , , , , , , , , , , , , () . , , , , , , , , , , , , () . , , , , , () . , , , , , , , , , , () . , , , , , , , , , , () . , , , , , , , , , () . , , () . , , , , , , () . , , , , , , , , , , , , , , , , () . , , , , , , , , (). , , , , , , , , , , , , () . , , , , , , , , , , , , , () . [^1]: One should note that Sera et al. explain the change of $\kappa_{\mathrm{ph}}$ at $T_\mathrm{LT}$ by an observed change of the velocity of sound $v_s$ at $T_\mathrm{LT}$ [@Sera97]. The actual changes of $v_s$ are, however, far too small ($\sim1\%$) [@Yamada94] to account for the much larger changes of $\kappa_{\mathrm{ph}}$. These changes of $v_s$ at $T_\mathrm{LT}$ should thus be regarded as a further accompanying phenomenon of the structural phase transition. [^2]: Note that dispersing optical phonons can be ruled out to play an important role because their contribution to the thermal conductivity have been estimated to be about one order of magnitude smaller ($\sim1~\rm Wm^{-1}K^{-1}$) than that what is observed [@Hess04]. [^3]: It is worth mentioning that a slight reduction of the spin wave velocity [@Brenig91] and changes of the spin gaps induced by the Zn-ions can be safely ignored. These effects lead to corrections smaller than the experimental error. [^4]: It should be noted that the term ’magnon’ refers to the specific $\Delta S=1$ triplet excitations of the $S=1/2$ two-leg quantum spin ladder, which sometimes are also dubbed as ’triplon’ excitations. They are not be confused with the spin-wave type magnon excitations of the 2D-HAF. [^5]: Note, that $l_0$ specifically denotes the low-temperature value of [$l_\mathrm{mag}$]{}used for fitting the low-temperature increase of [$\kappa_\mathrm{mag}$]{}. As we shall see later, [$l_\mathrm{mag}$]{}becomes temperature dependent at higher temperature. Note further, that the slightly smaller values of $l_0$ and $\Delta$ for $\mathrm{Ca_9La_5Cu_{24}O_{41}}$ as compared to the values reported in the original paper by [@Hess01], are a consequence of the usage of more accurate lattice parameters and an optimized fit-interval. It is stressed that these small corrections have no further consequences on the conclusions drawn in the original paper. [^6]: For example, one finds $\Theta_D= 296~\mathrm{K}$ for [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}[@McElfresh89]. [^7]: A careful investigation of the impact of these fluctuations yields that all aspects of the aforementioned analysis of [$\kappa_\mathrm{mag}$]{}in [$\mathrm{(Sr,Ca,La)_{14}Cu_{24}O_{41}}$]{}remain valid. For example, for [$\mathrm{Sr_{14}Cu_{24}O_{41}}$]{}fluctuations of [$\kappa_\mathrm{mag}$]{}leave the extracted value of the spin gap practically unaffected while in the same measurement the fluctuations of the extracted mean free path value amount to not more than about 10% [@Hess_diss2002]. [^8]: The chemical purity refers to that of the primary chemicals [CuO]{} and $\mathrm{SrCO_3}$ used in for the crystal growth by the traveling solvent floating zone method [@Hlubek2010].
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a new Monte Carlo Markov Chain algorithm for CMB analysis in the low signal-to-noise regime. This method builds on and complements the previously described CMB Gibbs sampler, and effectively solves the low signal-to-noise inefficiency problem of the direct Gibbs sampler. The new algorithm is a simple Metropolis-Hastings sampler with a general proposal rule for the power spectrum, $C_{\ell}$, followed by a particular deterministic rescaling operation of the sky signal, $\mathbf{s}$. The acceptance probability for this joint move depends on the sky map only through the difference of $\chi^2$’s between the original and proposed sky sample, which is close to unity in the low signal-to-noise regime. The algorithm is completed by alternating this move with a standard Gibbs move. Together, these two proposals constitute a computationally efficient algorithm for mapping out the full joint CMB posterior, both in the high and low signal-to-noise regimes.' author: - 'J. B. Jewell, H. K. Eriksen, B. D. Wandelt, I. J. O’Dwyer, Greg Huey, and K. M. Górski' date: 'Received - / Accepted -' title: 'A Markov Chain Monte Carlo Algorithm for analysis of low signal-to-noise CMB data' --- Introduction ============ Since the detection of anisotropy in the Cosmic Microwave Background (CMB; Smoot et al. 1992), there has been an emphasis on likelihood or Bayesian methods for the inference of cosmological parameters and their error bars, or more generally, their confidence intervals. CMB analysis is most suitably addressed in a Bayesian, as opposed to frequentist, framework, simply because the observed microwave sky is interpreted as a single realization of a spatial random process. Early measurements of the CMB were limited to signal to noise ratios of order unity at relatively low angular scales, where direct evaluation of the likelihood for the power spectrum or cosmological parameters is possible. However, the ${\cal O}(N^{3})$ scaling of computational expense with pixel number $N$ prohibits direct likelihood evaluation for current and future CMB observations. Motivated by the scientific potential of CMB data with increasingly high spatial resolution, yet beset with systematics including partial sky coverage and foregrounds, an iterative method of sampling from the Bayes posterior, using a special case of Markov Chain Monte Carlo (MCMC) known as Gibbs sampling, was introduced by [@jewell:2002; @jewell:2004]. The method was later independently discovered and applied to COBE data [@wandelt:2004], numerically extended to high-resolution on the sphere [@eriksen:2004], applied to analysis of the WMAP [@bennett:2003; @hinshaw:2007; @page:2007] data [@odwyer:2004; @eriksen:2007a; @eriksen:2007b], as well as generalized to include inference of foreground model parameters [@eriksen:2008a; @eriksen:2008b]. While Gibbs sampling provably converges to the Bayes posterior over the entire range of angular scales probed by the data, the run-time required to generate enough independent samples at the low signal-to-noise, small angular scale regime was found to be prohibitive [@eriksen:2004]. The reason for this is that typical variations in the power spectrum from one sample to the next are determined by cosmic variance alone, whereas the posterior itself is given by both cosmic variance and noise. This results in a long correlation length in the sequence of spectra in the low signal to noise regime, thus requiring a very long run time to generate a sufficient number of independent samples. In this paper we generalize the original Gibbs sampling algorithm to include a new type of MCMC step alternating with standard Gibbs sampling, which solves this problem of slow probabilistic convergence in the low signal to noise regime. This method therefore makes possible an exact Bayesian approach to CMB analysis over the entire range of angular scales probed by current and future experiments. The paper is organized as follows. We first review the CMB Gibbs sampler, and describe the associated numerical difficulties in analysis at small angular scales. We then introduce the new MCMC step to the Markov chain, designed specifically to allow large variations in the high-$\ell$ CMB spectrum, precisely where the signal to noise is $\le 1$. We derive the required Metropolis-Hastings acceptance probability correctness in Appendix \[app:proof\], and numerically demonstrate the method in Section \[sec:simulations\], for both temperature and polarization. Finally, we summarize and conclude in Section \[sec:conclusions\]. Review of Gibbs Sampling ======================== The Joint Posterior ------------------- We begin by assuming that the observed data may be modelled by a signal and a noise term, $${\mathbf{d}}= {\mathbf{A}}{\mathbf{s}}+ {\mathbf{n}},$$ where ${\mathbf{d}}$ is a vector containing the data (at every pointing of the detectors), the matrix ${\mathbf{A}}$ involves both pointing and beam convolution (and where for this paper we will assume symmetric beams and neglect the details of this operation), and ${\mathbf{n}}$ is additive noise (here in the pixel domain). We assume both the CMB signal and noise to be Gaussian random fields with vanishing mean and covariance matrices ${\mathbf{S}}$ and ${\mathbf{N}}$, respectively. In harmonic space, where ${\mathbf{s}}= \sum_{\ell, m} a_{\ell m} Y_{\ell m}$, the CMB temperature covariance matrix is given by $\textrm{C}_{\ell m, \ell' m'} = \langle a_{\ell m}^* a_{\ell' m'}\rangle = C_{\ell} \delta_{\ell \ell'} \delta_{m m'}$, ${C_{\ell}}$ being the angular power spectrum. A generalization to polarization merely requires the replacement of the signal matrix diagonal elements with $3 \times 3$ matrices of the form $${\mathbf{C}}_{l} = \left[ \begin{array}{ccc} C_{l}^{TT} & C_{l}^{TE} & C_{l}^{TB} \\ C_{l}^{ET} & C_{l}^{EE} & C_{l}^{EB} \\ C_{l}^{BT} & C_{l}^{BE} & C_{l}^{BB} \end{array} \right]$$ For the discussion in this section, we focus on the temperature case, but note that the generalization to polarization is straightforward and discussed by @larson:2007. Given these asumptions, our goal is to quantify what has been learned about the underlying power spectrum of the CMB given the data, or how well the data constrain the cosmological parameters. One proceeds then, in a Bayesian framework, by writing down the posterior given the data, $$P(C_{\ell} | {\mathbf{d}}) \propto \mathcal{L}({\mathbf{d}}| C_{\ell} ) P(C_{\ell}).$$ Here $\mathcal{L}({\mathbf{d}}| C_{\ell})$ is the likelihood and $P(C_{\ell})$ is a prior on $C_{\ell}$. In order to derive the functional form of the likelihood, one imagines randomly choosing any relevant model \[here a power spectrum drawn from $P(C_{\ell})$\], and asks what sequence of effects needs to be modeled in order to simulate the data. Here, simulation is understood as conditioning on the chosen model, and leads to a joint density $$\begin{aligned} P({\mathbf{d}},{\mathbf{s}},C_{\ell}) & = & P({\mathbf{d}},{\mathbf{s}}| C_{\ell}) P(C_{\ell}) \nonumber \\ & = & P({\mathbf{d}}| {\mathbf{s}}) P({\mathbf{s}}| C_{\ell}) P(C_{\ell})\end{aligned}$$ where the last line follows directly from our data model through the assumption of additive noise. Specifically, the factors in the above are $$\begin{aligned} -2 \log P({\mathbf{s}}| C_{\ell} ) & = & {\mathbf{s}}^{t} {\mathbf{C}}^{-1} {\mathbf{s}}- \log |{\mathbf{C}}| \nonumber \\ -2 \log P({\mathbf{d}}| {\mathbf{s}}) & = & -({\mathbf{d}}-{\mathbf{s}})^{t} {\mathbf{N}}^{-1}({\mathbf{d}}-{\mathbf{s}}) - \log |{\mathbf{N}}|\end{aligned}$$ which follow from the assumption that both the signal and noise are independent Gaussian processes. The idea of a “simulation chain” provides a conceptually clear approach to constructing a joint density, from which we immediately have the Bayesian posterior $$P(C_{\ell} | {\mathbf{d}}) = \int d{\mathbf{s}}\ P(C_{\ell}, {\mathbf{s}}| {\mathbf{d}})$$ The relevance of the above for this paper lies in relating what we refer to as the [*joint posterior*]{}, $P(C_{\ell}, {\mathbf{s}}| {\mathbf{d}})$, and the more familiar likelihood $\mathcal{L}({\mathbf{d}}| C_{\ell}) \propto P(C_{\ell} | {\mathbf{d}}) / P(C_{\ell})$, Although we can analytically compute the integral of the joint posterior over the signal for the Gaussian signal and noise processes considered here, and therefore simply write down the functional form of the likelihood, it is too expensive to evaluate it for any specified $C_{l}$ given high-resolution data. Furthermore, for more complicated data models (i.e. including foreground model uncertainties) we will not be able to perform the integrals over the additional degrees of freedom. Both situations then instead motivate sampling from the joint posterior, and thereby generating samples from $P(C_{\ell} | {\mathbf{d}})$ without ever evaluating $P(C_{\ell} | {\mathbf{d}})$. We now discuss the original Gibbs sampling approach proposed and implemented by @jewell:2004, @wandelt:2004 and @eriksen:2004, and then introduce a new MCMC step which directly addresses the previously reported slow probabilistic convergence in the low signal to noise regime [@eriksen:2004]. The CMB Gibbs sampler {#sec:cmb_sampling} --------------------- As stated above, our goal is to sample from the joint posterior, $$- 2 \log P({\mathbf{s}}, C_{\ell}|{\mathbf{d}}) = \chi^{2}({\mathbf{d}}, {\mathbf{s}}) + {\mathbf{s}}^{t} {\mathbf{S}}^{-1} {\mathbf{s}}+ \log | {\mathbf{S}}| + \log P(C_{\ell}). \label{eq:cmb_posterior}$$ For notational convenience, we have here dropped constant factors of $2\pi$, and also defined $$\chi^{2}({\mathbf{s}}, {\mathbf{d}}) = ({\mathbf{d}}- {\mathbf{s}})^{t} {\mathbf{N}}^{-1}({\mathbf{d}}-{\mathbf{s}}).$$ One approach to sample from this posterior is to use an algorithm known as Gibbs sampling, where we can alternately sample from the respective conditional densities, $$\begin{aligned} {\mathbf{s}}^{i+1} &\leftarrow P({\mathbf{s}}| C_{\ell}^i, {\mathbf{d}}) \\ C_{\ell}^{i+1} &\leftarrow P(C_{\ell} | {\mathbf{s}}^{i+1}, {\mathbf{d}}).\end{aligned}$$ Here $\leftarrow$ indicates sampling from the distribution on the right-hand side. After some burn-in period, during which all samples must be discarded, the joint samples $({\mathbf{s}}^i, C_{\ell}^i)$ will be drawn from the desired density. Thus, the problem is reduced to that of sampling from the two *conditional* densities $P({\mathbf{s}}| C_{\ell}, {\mathbf{d}})$ and $P(C_{\ell} | {\mathbf{s}}, {\mathbf{d}})$. We now describe the sampling algorithms for each of these two conditional distributions, starting with $P(C_{\ell} | {\mathbf{s}}, {\mathbf{d}})$. First, note that $P(C_{\ell} | {\mathbf{s}}, {\mathbf{d}}) = P(C_{\ell} | {\mathbf{s}})$ which follows directly from the construction of the joint density of “everything” above. This is also intuitively easy to understand since if we already know the CMB sky signal, the data themselves tell us nothing new about the CMB power spectrum. Next, since the sky is assumed to be Gaussian and isotropic, the distribution reads $$P(C_{\ell} | {\mathbf{s}}) \propto P(C_{\ell}) \frac{e^{-\frac{1}{2} {\mathbf{s}}_{\ell}^{t}{\mathbf{S}}_{\ell}^{-1}{\mathbf{s}}_{\ell}}}{\sqrt{|{\mathbf{S}}_{\ell}|}} = P(C_{\ell}) \frac{e^{-\frac{2\ell+1}{2} \frac{\sigma_{\ell}}{C_{\ell}}}}{C_{\ell}^{\frac{2\ell+1}{2}}},$$ which, when interpreted as a function of $C_{\ell}$, is known as the inverse Gamma distribution. In this expression, $\sigma_{\ell} = \frac{1}{2\ell+1} \sum_{m} |a_{\ell m}|^2$ denotes the observed power spectrum of ${\mathbf{s}}$. Fortunately, there exists a simple textbook sampling algorithm for this distribution [e.g., @gupta:2000], and we refer the interested reader to the previous papers for details. For an alternative, and more flexible, sampling algorithm, see @wehus:2008. In order to describe the sky signal sampling step, we first define the mean-field map (or Wiener filtered data) to be $\hat{{\mathbf{s}}} = ({\mathbf{S}}^{-1} + {\mathbf{N}}^{-1})^{-1} {\mathbf{N}}^{-1} {\mathbf{d}}$, and note that the conditional sky signal density given the data and $C_{l}$ can be written as $$\begin{aligned} P({\mathbf{s}}| C_{\ell}, {\mathbf{d}}) &\propto e^{-\frac{1}{2} ({\mathbf{s}}-\hat{{\mathbf{s}}})^t ({\mathbf{S}}^{-1} + {\mathbf{N}}^{-1}) ({\mathbf{s}}-\hat{{\mathbf{s}}})}.\end{aligned}$$ Thus, $P({\mathbf{s}}| C_{\ell}, {\mathbf{d}})$ is a Gaussian distribution with mean equals to $\hat{{\mathbf{s}}}$ and a covariance matrix equals to $({\mathbf{S}}^{-1} + {\mathbf{N}}^{-1})^{-1}$. Sampling from this Gaussian distribution is straightforward, but computationally somewhat cumbersome. First, draw two random white noise maps $\omega_0$ and $\omega_1$ with zero mean and unit variance. Then solve the equation $$\left[{\mathbf{S}}^{-1} + {\mathbf{N}}^{-1}\right] {\mathbf{s}}= {\mathbf{N}}^{-1}{\mathbf{d}}+ {\mathbf{S}}^{-\frac{1}{2}} \omega_0 + {\mathbf{N}}^{-\frac{1}{2}} \omega_1. \label{eq:lin_sys}$$ for ${\mathbf{s}}$. Since the white noise maps have zero mean, one immediately sees that $\langle {\mathbf{s}}\rangle = \hat{{\mathbf{s}}}$, while a few more calculations show that $\langle {\mathbf{s}}{\mathbf{s}}^{t} \rangle = ({\mathbf{S}}^{-1} + {\mathbf{N}}^{-1})^{-1}$. The problematic part about this sampling step is the solution of the linear system in Equation \[eq:lin\_sys\]. Since this a $\sim10^6 \times 10^6$ system for current CMB data sets, it cannot be solved by brute force. Instead, one must use a method called Conjugate Gradients (CG), which only requires multiplication of the coefficient matrix on the left-hand side, not inversion. For details on these computations, together with some ideas on preconditioning, see @eriksen:2004. Convergence issues in the low signal-to-noise regime ---------------------------------------------------- As originally applied to high-resolution CMB data, the Gibbs sampling algorithm as described above has very slow convergence at the high-$\ell$, low signal-to-noise part of the spectrum. The reason for the slow convergence is easy to understand in light of the above: When sampling from $P(C_{\ell} | {\mathbf{s}})$, the typical step size is given by cosmic variance at all angular scales. In the high signal-to-noise regime, cosmic variance dominates the noise variance, and we are able to explore the full width of the posterior in only a few Gibbs iterations. However, in the low signal-to-noise end, cosmic variance is far smaller than the posterior variance, and it takes a prohibitively long time to converge probabilistically. This problem of “slow mixing” of the Gibbs sampler is illustrated in figures \[fig:TT\_trace\_plots\] and \[fig:TT\_correlation\_length\]. The long correlation length starting at signal-to-noise of unity leads to extremely long run times in order to produce a reasonable number of uncorrelated samples. A Low Signal-to-Noise MCMC Sampler ================================== When sampling from the true posterior, the goal is to produce as many independent samples from $P(C_{\ell}, {\mathbf{s}}| {\mathbf{d}})$ as possible. One might intuitively guess that it should be straightforward to establish good approximations to the posterior in the low signal-to-noise regime, since in the limit of vanishing signal to noise we simply recover the prior. This suggests that we look for a sampling scheme in which we first sample a new spectrum from some approximation to the true posterior independent on the current spectrum and CMB map, followed by sampling the CMB map from the conditional $P({\mathbf{s}}|C_{\ell}, {\mathbf{d}})$. The problem with such a direct scheme is that the accept probability will involve a ratio of determinants which are too expensive to compute. We are therefore motivated to look for a sampling scheme in which we can make a large variation in $C_{\ell}$ in the low signal-to-noise regime, and make an associated [*deterministic change*]{} in the CMB map, while still maintaining a reasonably high acceptance rate. The motivation for a deterministic change is that it will avoid introducing ratios of determinants which we cannot compute. Proposal rule and acceptance probability ---------------------------------------- Assume that we have defined a deterministic sampling scheme for ${\mathbf{s}}$, and that our new CMB map is given by some function $${\mathbf{s}}_{n+1} = F({\mathbf{s}}_{n}, C_{\ell}^{(n+1)}, C_{\ell}^{(n)} ).$$ Then the condition of detailed balance for our MCMC sampler requires that $$F^{-1}({\mathbf{s}}_{n+1}, C_{\ell}^{(n+1)}, C_{\ell}^{(n)}) = F({\mathbf{s}}_{n+1}, C_{\ell}^{(n)}, C_{\ell}^{(n+1)}),$$ or, in other words, that the inverse function is given by exchanging the order of the spectra in the function $F$. One simple function which has this property is $${\mathbf{s}}_{n+1} = \left(\frac{C_{\ell}^{(n+1)}}{C_{\ell}^{(n)}}\right)^{\frac{1}{2}} {\mathbf{s}}_{n}$$ The total proposal matrix is then $$\begin{aligned} w(C_{\ell}^{(n+1)}, {\mathbf{s}}_{n+1} | C_{\ell}^{(n)}, {\mathbf{s}}_{n}) & = & w(C_{\ell}^{(n+1)} | C_{\ell}^{(n)}, {\mathbf{d}}) \nonumber \\ & & \delta \left( {\mathbf{s}}_{n+1} - \left(\frac{C_{\ell}^{(n+1)}}{C_{\ell}^{(n)}}\right)^{-\frac{1}{2}} {\mathbf{s}}_{n} \right), \nonumber\end{aligned}$$ and the “reverse” proposal is $$\begin{aligned} w(C_{\ell}^{(n)}, {\mathbf{s}}_{n} | C_{\ell}^{(n+1)}, {\mathbf{s}}_{n+1}) & = & w(C_{\ell}^{(n)} | C_{\ell}^{(n+1)}, {\mathbf{d}}) \nonumber \\ & & \delta \left( {\mathbf{s}}_{n} - \left(\frac{C_{\ell}^{(n)}}{C_{\ell}^{(n+1)}}\right)^{-\frac{1}{2}} {\mathbf{s}}_{n+1} \right). \nonumber\end{aligned}$$ The condition of detailed balance including deterministic moves requires the consideration of some technical points which we leave to Appendix \[app:proof\]. There we show that the full Metropolis-Hastings accept probability reads $$\begin{aligned} A & = & \min \left[ 1, \frac{e^{- \chi^{2}({\mathbf{s}}_{n+1}, {\mathbf{d}})}}{e^{-\chi^{2}({\mathbf{s}}_{n}, {\mathbf{d}})}} \frac{w(C_{\ell}^{(n)} | C_{\ell}^{(n+1)}, {\mathbf{d}})}{w(C_{\ell}^{(n+1)} | C_{\ell}^{(n)}, {\mathbf{d}})} \right]\end{aligned}$$ The significance of the above is that we can make relatively large changes to the power spectrum in the low signal-to-noise regime, where ${\mathbf{N}}^{-1}$ is getting small, since the $\chi^2$ is affected only very mildly by changes in any low signal-to-noise mode. We note the interesting point (discussed more completely in Appendix \[app:cov\]) that if one changes variables in the joint posterior from CMB maps, ${\mathbf{s}}$, to whitened maps, ${\mathbf{x}}= {\mathbf{C}}_{\ell}^{-\frac{1}{2}} {\mathbf{s}}$, and then Gibbs sample in the new variables $(C_{\ell}, {\mathbf{x}})$, the resulting accept probability is numerically identitical to the above. However, we note the distinction here to emphasize the difference between MCMC algorithms implementing deterministic proposals of maps given $C_{\ell}$, and those sampling in a different set of variables, as there could be other deterministic proposal schemes or another change of variables which lead to improvements over the approach presented in this paper. For the numerical demonstration of the MCMC algorithm presented in this paper, we use a simple symmetric Gaussian proposal, truncated at $C_{\ell}>0$ (or, for polarization, the region where the resulting CMB covariance matrix is positive definite), for the power spectrum, $$w(C_{\ell}^{(n+1)} | C_{\ell}^{(n)}, {\mathbf{d}}) \propto e^{-\frac{1}{2} \left(\frac{C_{\ell}^{n+1}-C_{\ell}^{n}}{\tau_{\ell}}\right)^2} I(C_{\ell}>0),$$ where $\tau_{\ell}$ is a measure of the typical step size taken between two samples. Note that because this proposal density is symmetric, the ratio of $C_{\ell}$ proposals cancels, and the acceptance probability is entirely determined by the change in $\chi^{2}$. It should be noted that while the above MCMC step satisfies detailed balance, it is not [*irreducable*]{}, in the sense that there is not a non-vanishing probability in reaching any state from any other state in a finite number of MCMC steps; the phases are unchanged in each MCMC step. However, alternating these steps with a traditional Gibbs sampling step gives a combined “two-step” MCMC algorithm which indeed is irreducable, and therefore provably converges to the joint posterior. Once again, the details are left to the appendix for the interested reader. Optimization of the MCMC sampler -------------------------------- A general advantage of the Gibbs sampler is the fact that it is free of tunable efficiency parameters. The same is not true for the Metropolis-Hastings MCMC algorithm; for satisfactory sampling performance, it typically has to be tuned quite extensively. In this section, we describe three specific features that helps in this task, namely 1) step size tuning, 2) slice sampling and 3) binning. First, we have to ensure that the step size of our Gaussian proposal density roughly matches the width of the target distribution, in order to maintain both a reasonable acceptance rate and high mobility. We do this by performing an initial test run, producing typically a few hundreds $C_{\ell}$ samples, and compute the standard deviation of these samples for each $\ell$. These are then adopted as the proposal widths for the main run, scaled by some number less than unity, typically between 0.05 and 0.5. For the initial test run, we approximate the posterior width by the noise variance alone, $$\tau_{\ell}^{2} = \frac{2}{2\ell+1} \frac{N_{\ell}}{b_{\ell}^{2}},$$ because the MCMC sampler is used only in the low signal-to-noise regime. In this expression $N_{\ell}$ is the power spectrum of the instrumental noise alone, and $b_{\ell}$ is the product of the Legendre transform of the beam and the HEALPix window function. Next, Metropolis-Hastings MCMC is inefficient in spaces with too many free parameters. For this reason, we divide the power spectrum coefficients, $C_{\ell}$, into subsets, each containing typically only 10–20 multipoles. Then we propose changes to one subset at a time, while keeping all other multipoles fixed. Finally, we loop over subsets, and thus effectively implement a multipole slice Gibbs sampler for the full power spectrum. This is computationally feasible, because a single MCMC proposal only requires a single $\chi^{2}$ evaluation, which has a computational cost of a single spherical harmonic transform. Since drawing a full sky map from $P({\mathbf{s}}|C_{\ell}, {\mathbf{d}})$ in the classical Gibbs sampling step requires $\mathcal{O}(10^{2})$ spherical harmonic transforms, we can indeed afford to perform many MCMC proposals for each Gibbs step, without dominating the total cost. Nevertheless, for very high-resolution analysis it is often beneficial to bin several $C_{\ell}$’s together, both in order to increase the signal-to-noise of the joint coefficient, and to decrease the number of parameters that needs to be sampled by MCMC. We implement this by defining a new binned spectrum, weighted by $\ell(\ell+1)/2\pi$, as follows, $$C_{b} = \frac{1}{N_b}\sum_{\ell \in b} \frac{\ell(\ell+1)}{2\pi} C_{\ell}.$$ Here $b=[\ell_{\textrm{min}}, \ell_{\textrm{max}}]$ denotes the current bin, and $N_{b} = \ell_{\textrm{max}}-\ell_{\textrm{min}}+1$ is the number of multipoles within the bin. These new (and fewer) coefficients are then sampled with the above MCMC sampler, after which the original spectrum coefficients are given by $$C_{\ell} = \frac{2\pi}{\ell(\ell+1)} C_{b}.$$ Testing and Validation {#sec:simulations} ====================== We have implemented the new sampling step described above in the previously Gibbs sampling code called “Commander” [@eriksen:2004; @eriksen:2008a], and in this section we demonstrate its advantages compared to the old sampling algorithm. We consider two different cases, namely high-$\ell$ temperature and low-$\ell$ polarization analysis. In the former case, we also analyse two cases, with and without a sky cut. The former allows us to verify the results against an analytically known answer, while the second demonstrates that the sky cut does not degrade the sampling efficiency. Temperature analysis -------------------- The high-$\ell$ temperature simulation is designed to mimic the 5-year WMAP temperature data [@hinshaw:2008] with one exception, namely that the noise is assumed spatially uniform, in order to facilitate analytic comparison. Specifically, the CMB realization was drawn from the best-fit $\Lambda$CDM model derived from WMAP alone [@komatsu:2008], including multipoles up to $\ell_{\textrm{max}}=1000$, and then smoothed with the instrumental beam of the WMAP V1 differencing assembly, and pixelized at HEALPix[^1] resolution $N_{\textrm{side}}=512$. Finally, uniform noise of $\sigma_0 = 40\mu\textrm{K}$ RMS was added to each pixel. This corresponds to a signal-to-noise ratio of unity at $\ell \sim 550$, roughly similar to the 5-year WMAP data. We analyse this simulation both with and without the WMAP KQ85 sky cut [@gold:2008]. In both analyses, we adopted the Gaussian proposal density with tuned variances, as described above. We also bin the power spectrum in progressively wide bins, starting at $\ell = 600$, to maintain a reasonable signal-to-noise per sampled power spectrum parameter. Ten bins were sampled jointly per proposal, while all others were kept fixed. In the full-sky case, we produced a total of 31,800 samples over 60 chains, and in the cut sky case a total of 6800 samples. The cost for producing one sample in the latter, and by far most expensive, set was 2.5 CPU hours, for a total of 17000 CPU hours. The number of MCMC steps per Gibbs step was one in the former and 20 in the latter. (Since the the signal sampler dominates the cut sky Gibbs chain one can perform more low S/N steps without slowing down the overall code significantly.) In addition to these two main sample sets, we also produced two longer chains with each 3500 samples for the full-sky casee, both with and without the new MCMC step turned on, in order to compare the Markov chain correlation lengths before and after including the MCMC sampler. We first consider the full-sky data set, and in Figure \[fig:TT\_trace\_plots\] we show a segment of each of the two longer chains for three selected multipole bins. The top panel shows $\ell=600$, which is the first bin to be sampled by MCMC, the middle panel shows $\ell=732-742$, where there is still some signal in the data, and, finally, the bottom panel shows $\ell=855-1000$, which is strongly noise dominated. Starting with the top panel, we see that the red curve (Gibbs+MCMC) scatters significantly faster than the black curve (Gibbs only), implying more efficient sampling. This trend becomes even stronger with lower signal-to-noise, until the last case, where the Gibbs-only chain essentially does not move at all, while the MCMC sampler does probe the full range. Note, however, that even the MCMC sampler has a significant correlation length in this range, and this implies that there is still some room for improvement to be made in defining our proposals. Next, these considerations are quantified in Figure \[fig:TT\_correlation\_length\], where we plot the Markov chain correlation length as a function of distance in the chain, for six bins with and without the MCMC sampler. As first reported by @eriksen:2004, we see that the Gibbs-only correlation length increases dramatically with decreasing signal-to-noise, rendering the algorithm essentially useless in this regime. However, we also see that the new MCMC step effectively resolves this issue, as the correlation length (here defined by having a correlation less than 0.2) now is less than $\sim40$ steps. This is a dramatic improvement, and makes the algorithm useful even in this range. Nevertheless, we once again point out that it is possible to make further improvements by establishing better proposal densities. In Figure \[fig:TT\_gr\] we consider the convergence properties of the $\sim30$k samples set, by computing the Gelman-Rubin statistic $R$ [@gelman:1992] as a function of $\ell$. Typically, one recommends that $R$ should be less than, say, 1.2 in order to claim convergence. We see that this holds everywhere for this sample set, and typically it is even less than 1.05. Note also the step at $\ell=600$, showing clearly the beneficial effect of the MCMC sampler. Next, in Figure \[fig:TT\_posteriors\] we compare the marginal distributions derived from this sample set with the analytic result, $$P(C_{\ell}|\mathbf{d}) \propto \prod_{\ell \in b} \frac{e^{-\frac{2\ell+1}{2} \frac{\sigma_{\ell}^{\textrm{S+N}}}{b_{\ell}^2 C_{\ell}+N_{\ell}}}} {(b_{\ell}^2 C_{\ell}+N_{\ell})^{\frac{2\ell+1}{2}}}.$$ Here $b=[\ell_{\textrm{min}}, \ell_{\textrm{max}}]$ indicates a given multipole bin, $b_{\ell}$ denotes the product of the instrumental beam and the HEALPix pixel window, and $\sigma_{\ell}^{\textrm{S+N}}$ is the power spectrum of the noisy data map. We see that the new algorithm reproduces the analytic distributions very well, and this verifies the overall method. Finally, the cut-sky power spectrum with one-sigma confidence regions is shown in three panels in Figure \[fig:TT\_spectrum\], focusing on different $\ell$-ranges, namely all $\ell$’s, the $S/N \sim 1$ transition region, and the low $S/N$ region. This completes the high-$\ell$ temperature analysis validation. Polarization analysis --------------------- We now consider polarization analysis, and construct a new low-$\ell$ simulation for this purpose. This simulation does not mimic any planned experiment, but is rather designed to highlight the analysis method itself. Specifically, we drew a new CMB realization from the best-fit WMAP $\Lambda$CDM spectrum that includes a non-zero tensor contribution, including multipoles up to $\ell_{\textrm{max}}=150$, and convolved this with a $3^{\circ}$ FWHM Gaussian beam, and pixelized it at $N_{\textrm{side}} = 64$. Uniform noise of $5\mu\textrm{K}$ RMS was added to the temperature component, and $1\mu\textrm{K}$ RMS to the polarization components. The 5-year WMAP polarization sky mask was imposed on the data. We allowed for non-zero $C_{\ell}^{TT}$, $C_{\ell}^{TE}$, $C_{\ell}^{EE}$ and $C_{\ell}^{BB}$ spectra, but fixed $C_{\ell}^{TB} = C_{\ell}^{EB} = 0$. These spectra were then individually binned to maintain a reasonable signal-to-noise per bin. (Details on how to introduce individual binning of each power spectrum were recently described by Eriksen and Wehus, 2008.) Again, a tuned Gaussian proposal density was used in the MCMC step. A total of 12000 samples were produced over 12 chains, and the CPU time per sample was 55 seconds, for a total of $\sim200$ CPU hours. In Figure \[fig:pol\_trace\_plots\] we show one $C_{\ell}$ chain for each of the four sampled spectra, for the last (and therefore most difficult) bin in each case. Note that the $C_{\ell}^{EE}$ and $C_{\ell}^{BB}$ spectra have essentially vanishing signal-to-noise, and therefore these chains reach zero values. Clearly, we see that mixing properties of these chains are satisfactory, and the correlation lengths are quite short. In Figure \[fig:gr\_polarization\] we show the Gelman-Rubin statistics for each of the four power spectra, and with the single exception of the very last bin of $C_{\ell}^{EE}$, all $R$ values are well below 1.1. Thus, all spectra have converged well everywhere. Finally, in Figure \[fig:pol\_spectrum\] we show the reconstructed marginal power spectra for each polarization component, overplotted on the input spectrum. The agreement is very good. Note, however, that these spectra are direct marginals, and not a joint maximum likelihood estimate. They are therefore not individual unbiased estimators. In particular, the marginal $C_{\ell}^{EE}$ power spectrum is biased slightly high because of the combination of the $C_{\ell}^{TT}C_{\ell}^{EE} - (C_{\ell}^{TE})^2 > 0$ positivity constraint and relatively low signal-to-noise. Consideration of the joint polarization posterior, which *is* an unbiased estimator, is postponed to a future publication. Conclusions {#sec:conclusions} =========== We have presented a new MCMC algorithm for the high-L, low signal to noise limit of the joint posterior which solves the slow probabilistic convergence of the traditional Gibbs sampler in this regime. This in principle allows sampling over the joint posterior $p(C_{l}, {\mathbf{s}}| {\mathbf{d}})$ over the entire range of angular scales probed by current and future CMB experiments. The limiting computational burden is now entirely in the map-making step of Gibbs sampling, for which the cost per Gibbs iteration now scales with the expense of multiplication by the inverse noise matrix ${\mathbf{N}}^{-1}$. Assuming pixel uncorrelated (but scan weighted) noise as a good approximation at small angular scales, the cost of an ${\mathbf{N}}^{-1}$ multiplcation is that of a forward and inverse spherical harmonic transform, or ${\cal O}(\ell_{\textrm{max}}^{3})$. Future work will attempt to push the generalized Gibbs + MCMC sampling scheme presented here to smaller angular scales, ultimately limited by the degree to which we can compute harmonic transforms. We acknowledge use of the HEALPix[^2] software [@gorski:2005] and analysis package for deriving the results in this paper. HKE acknowledges financial support from the Research Council of Norway. Including Deterministic Proposals in MCMC {#app:proof} ========================================= Here we review the derivation of the accept probability in Markov Chain Monte Carlo when using deterministic proposals (or proposals where some of the degrees of freedom are specified as deterministic functions of the past state and/or proposed variations in some other degrees of freedom). We first briefly review the Metropolis-Hastings Markov Chain Monte Carlo algorithm and the proof of its convergence, and then turn to the special case involving deterministic proposals. Much of the review of the MCMC algorithm here follows [@Sokal:1989]. We also note that similar technical considerations including deterministic elements in proposals are presented in [@Green:1995] in the context of MCMC algorithms in which the dimension of the state space itself is included as a random variable to be sampled over. The goal is the construction of a transition matrix $T(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}})$ such that after initializing the Markov Chain with a sample from any probability density $p_{0}(C_{l}, {\mathbf{s}}| {\mathbf{d}})$, we generate samples from a sequence of probability densities $$p_{n+1}(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \equiv \int d(C_{l}', {\mathbf{s}}') \ T(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) \ p_{n}(C_{l}', {\mathbf{s}}' | {\mathbf{d}})$$ which eventually converge to an [*equilibrium density*]{} $\pi(C_{l}, {\mathbf{s}}| {\mathbf{d}})$ $$\pi (C_{l}, {\mathbf{s}}| {\mathbf{d}}) = \lim_{n \rightarrow \infty} p_{n}(C_{l}, {\mathbf{s}}| {\mathbf{d}})$$ We remind the reader of the sufficient conditions to establish convergence of an MCMC algorithm: [*stationarity*]{}, which means that the MCMC transition matrix satisfies $$\pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) = \int d(C_{l}', {\mathbf{s}}') \ T(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) \ \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}})$$ and [*irreducability*]{}, which means that for any two states, there is a finite number of iterations which give a non-vanishing probability to transition from one state to the other. It is well known that these two properties are sufficient to establish convergence, as can be seen simply from the triangle inequality $$\begin{aligned} \int d(C_{l}, {\mathbf{s}}) \ \left| \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) - p_{n}(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \right| & = & \int d(C_{l}, {\mathbf{s}}) \ \left| \int d(C_{l}', {\mathbf{s}}') \ T(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}') \left( \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) - p_{n-1}(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \right) \right| \nonumber \\ & \le & \int d(C_{l}, {\mathbf{s}}) \ \int d(C_{l}', {\mathbf{s}}') \ T(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}') \left| \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) - p_{n-1}(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \right| \nonumber \\ & = & \int d(C_{l}', {\mathbf{s}}') \ \left( \int d(C_{l}, {\mathbf{s}}) \ T(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}') \right) \left| \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) - p_{n-1}(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \right| \nonumber \\ & = & \int d(C_{l}', {\mathbf{s}}') \ \left| \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) - p_{n-1}(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \right| \nonumber\end{aligned}$$ The Metropolis-Hastings Markov Chain Monte Carlo algorithm is one method of constructing such a transition matrix. We choose [*any*]{} proposal matrix $w(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}})$ and then accept the proposed move with a probability $$0 \le A(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) \le 1$$ while rejecting the proposed move with probability $1 - A$ leads to a “null transition” where the next state in the Markov Chain remains the same. Application of this algorithm then leads to the sequence of probability densities which satisfy $$\begin{aligned} p_{n+1}(C_{l}, {\mathbf{s}}| {\mathbf{d}}) & = & \left( 1 - \int d(C_{l}', {\mathbf{s}}') \ A(C_{l}', {\mathbf{s}}' | C_{l} , {\mathbf{s}}, {\mathbf{d}}) w(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) \right) p_{n}(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \nonumber \\ & & + \int d(C_{l}', {\mathbf{s}}') \ A(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) w(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) p_{n}(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \end{aligned}$$ where the first term is the constribution to the probability density $p_{n+1}$ if we reject any proposed move, while the second term is the contribution from accepting the proposed move from any possible previous state. If we demand that, for a chosen proposal matrix, the accept probability satisfies $$\pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) w(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) A(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) = A(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) w(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}})$$ then we see that the MH MCMC algorithm satisfies stationarity, i.e. denoting by $T \circ \pi$ the density resulting from one application of the transition matrix to $\pi$, we have directly from detailed balance $$\begin{aligned} T \circ \pi & = & \left( 1 - \int d(C_{l}', {\mathbf{s}}') \ A(C_{l}', {\mathbf{s}}' | C_{l} , {\mathbf{s}}, {\mathbf{d}}) w(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) \right) \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \nonumber \\ & & + \int d(C_{l}', {\mathbf{s}}') \ A(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) w(C_{l}, {\mathbf{s}}| C_{l}', {\mathbf{s}}', {\mathbf{d}}) \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \nonumber \\ & = & \left( 1 - \int d(C_{l}', {\mathbf{s}}') \ A(C_{l}', {\mathbf{s}}' | C_{l} , {\mathbf{s}}, {\mathbf{d}}) w(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) \right) \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \nonumber \\ & & + \pi (C_{l}, {\mathbf{s}}| {\mathbf{d}}) \int d(C_{l}', {\mathbf{s}}') \ A(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) w(C_{l}', {\mathbf{s}}' | C_{l}, {\mathbf{s}}, {\mathbf{d}}) \nonumber \\ & = & \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}})\end{aligned}$$ We now turn to the case where our proposal is of the form $$w({\mathbf{s}}', C_{l}' | {\mathbf{s}}, C_{l}) = \delta \left[ {\mathbf{s}}' - F({\mathbf{s}}, C_{l}', C_{l}) \right] w(C_{l}' | C_{l}, {\mathbf{d}})$$ where we randomly propose a new power spectrum, posibly in a manner conditionally denpendent on the current spectrum and the data, and then deterministically compute a new CMB map with some function $${\mathbf{s}}' = F({\mathbf{s}}, C_{l}', C_{l})$$ To satisfy detailed balance with a non-vanishing accept probability our function must satisfy $$\begin{aligned} {\mathbf{s}}' & = & F({\mathbf{s}}, C_{l}', C_{l}) \nonumber \\ {\mathbf{s}}& = & F({\mathbf{s}}', C_{l}, C_{l}')\end{aligned}$$ or, that the inverse function is equivalent to interchanging the order of the power spectrum arguements $$F({\mathbf{s}}', C_{l}, C_{l}') = F^{-1}({\mathbf{s}}',C_{l}', C_{l})$$ In this paper, we have chosen one such function, given by $$F({\mathbf{s}}, C_{l}', C_{l}) = [{\mathbf{C}}']^{1/2} [{\mathbf{C}}]^{-1/2} {\mathbf{s}}$$ where interchanging the spectra in the function above does in fact give the inverse function itself. Our job now is to [*derive*]{} the accept probability such that we satisfy stationarity (as discussed above). For the proposal with deterministic changes to some of the degrees of freedom, stationarity is satisfied if $$\begin{aligned} (T \circ \pi)(C_{l}, {\mathbf{s}}| {\mathbf{d}}) & = & \left[ 1 - \int d(C_{l}'', {\mathbf{s}}'') \ A[C_{l}'', {\mathbf{s}}'' | C_{l}, {\mathbf{s}}] \delta[{\mathbf{s}}'' - F({\mathbf{s}}, C_{l}'', C_{l})] w(C_{l}'' | C_{l}, {\mathbf{d}}) \right] \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \nonumber \\ & & + \int d(C_{l}', {\mathbf{s}}') \ A[{\mathbf{s}}, C_{l} | {\mathbf{s}}', C_{l}'] \delta[{\mathbf{s}}- F({\mathbf{s}}', C_{l}, C_{l}')] w(C_{l} | C_{l}' , {\mathbf{d}}) \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \nonumber \end{aligned}$$ In order to determine the integral over the $\delta$-function in the accept term above, we recall the identity for $\delta[G({\mathbf{x}})]$, where $G({\mathbf{a}}) = 0$, $$\delta[G({\mathbf{x}})] = \frac{\delta({\mathbf{x}}- {\mathbf{a}})}{\left| \partial G / \partial {\mathbf{x}}\right|_{a} }$$ In our case, we can identify $$G({\mathbf{s}}') = {\mathbf{s}}- F({\mathbf{s}}', C_{l}, C_{l}')$$ which vanishes at $F^{-1}({\mathbf{s}},C_{l}, C_{l}') = F({\mathbf{s}}, C_{l}', C_{l})$. We also have the Jacobian $$\left| \frac{\partial G}{ \partial {\mathbf{s}}'} \right|_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} = \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')}$$ (i.e. $G({\mathbf{s}}')$ is considered a function of ${\mathbf{s}}'$ with the other CMB map ${\mathbf{s}}$ considered fixed) which therefore gives $$\delta[{\mathbf{s}}- F({\mathbf{s}}', C_{l}, C_{l}')] = \delta[{\mathbf{s}}' - F^{-1}({\mathbf{s}}, C_{l}, C_{l}')] \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')}$$ Inserting this into the condition for stationarity we have $$\begin{aligned} (T \circ \pi )(C_{l}, {\mathbf{s}}| {\mathbf{d}}) & = & \left[ 1 - \int d(C_{l}'', {\mathbf{s}}'') \ A[C_{l}'', {\mathbf{s}}'' | C_{l}, {\mathbf{s}}] \delta[{\mathbf{s}}'' - F({\mathbf{s}}, C_{l}'', C_{l})] w(C_{l}'' | C_{l}, {\mathbf{d}}) \right] \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \nonumber \\ & & + \int d(C_{l}', s') \ A[s, C_{l} | s', C_{l}'] \left( \delta[{\mathbf{s}}' - F^{-1}({\mathbf{s}}, C_{l}, C_{l}')] \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) w(C_{l} | C_{l}' , {\mathbf{d}}) \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \nonumber \\ & = & \left[ 1 - \int d(C_{l}'', {\mathbf{s}}'') \ A[C_{l}'', {\mathbf{s}}'' | C_{l}, {\mathbf{s}}] \delta[{\mathbf{s}}'' - F({\mathbf{s}}, C_{l}'', C_{l})] w(C_{l}'' | C_{l}, {\mathbf{d}}) \right] \pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) \nonumber \\ & & + \int d(C_{l}', {\mathbf{s}}') \ A[{\mathbf{s}}, C_{l} | {\mathbf{s}}', C_{l}'] \left( \delta[{\mathbf{s}}' - F({\mathbf{s}}, C_{l}', C_{l})] \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) w(C_{l} | C_{l}' , {\mathbf{d}}) \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}}) \nonumber \end{aligned}$$ where in the second line we again used the property that the inverse $F^{-1}$ is equivalent to $F$ with the spectra arguements interchanged. We see from the above that a sufficient condition for stationarity is $$\pi(C_{l}, {\mathbf{s}}| {\mathbf{d}}) w(C_{l}' | C_{l} , {\mathbf{d}}) A[{\mathbf{s}}', C_{l}' | {\mathbf{s}}, C_{l}] = A[{\mathbf{s}}, C_{l} | {\mathbf{s}}', C_{l}'] \left( \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) w(C_{l} | C_{l}' , {\mathbf{d}}) \pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}})$$ An accept probability which satisfies this condition therefore gives cancellation of the integrals over the $\delta$-functions for both the reject and accept contributions, leaving us exactly with $T \circ \pi = \pi$. We therefore have the accept probability $$A[{\mathbf{s}}', C_{l}' | {\mathbf{s}}, C_{l}] = \min \left[ 1, \frac{\pi(C_{l}', {\mathbf{s}}' | d)}{\pi(C_{l}, {\mathbf{s}}| {\mathbf{d}})} \frac{w(C_{l} | C_{l}', {\mathbf{d}})}{w(C_{l}' | C_{l}, {\mathbf{d}})} \left( \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) \right]$$ We give the expression above for the general case of any deterministic change in the CMB map with a function which satisfies $F({\mathbf{s}}, C_{l}, C_{l}') = F^{-1}({\mathbf{s}}, C_{l}', C_{l})$. We now explicitly evaluate this accept probability for the functional form chosen for this paper. Since we have $F({\mathbf{s}}', C_{l}, C_{l}') = [{\mathbf{C}}]^{1/2} [{\mathbf{C}}']^{-1/2} {\mathbf{s}}'$, we have $$\left( \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) = \frac{|{\mathbf{C}}|^{1/2}}{|{\mathbf{C}}'|^{1/2}}$$ Reminding the reader of the functional form of the joint posterior in eqn. \[eq:cmb\_posterior\], we have the accept probability given by $$\begin{aligned} A[{\mathbf{s}}', C_{l}' | {\mathbf{s}}, C_{l}] & = & \min \left[ 1, \frac{\pi(C_{l}', {\mathbf{s}}' | {\mathbf{d}})}{\pi(C_{l}, {\mathbf{s}}| {\mathbf{d}})} \frac{w(C_{l} | C_{l}', {\mathbf{d}})}{w(C_{l}' | C_{l}, {\mathbf{d}})} \left( \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) \right] \nonumber \\ & = & \min \left[ 1, \frac{e^{- \chi^{2}({\mathbf{s}}', {\mathbf{d}})}}{e^{- \chi^{2}({\mathbf{s}}, {\mathbf{d}})}} \frac{e^{- {\mathbf{s}}' [{\mathbf{C}}']^{-1} {\mathbf{s}}'}}{e^{{\mathbf{s}}{\mathbf{C}}^{-1} {\mathbf{s}}}} \frac{|{\mathbf{C}}|^{1/2}}{|{\mathbf{C}}'|^{1/2}} \frac{w(C_{l} | C_{l}', d)}{w(C_{l}' | C_{l}, d)} \left( \left| \frac{\partial F}{ \partial {\mathbf{s}}'} \right|^{-1}_{{\mathbf{s}}' = F^{-1}({\mathbf{s}}, C_{l}, C_{l}')} \right) \right] \nonumber \\ & = & \min \left[ 1, \frac{e^{- \chi^{2}({\mathbf{s}}', {\mathbf{d}})}}{e^{- \chi^{2}({\mathbf{s}}, {\mathbf{d}})}} \frac{e^{- {\mathbf{s}}' [{\mathbf{C}}']^{-1} {\mathbf{s}}'}}{e^{{\mathbf{s}}{\mathbf{C}}^{-1} {\mathbf{s}}}} \frac{|{\mathbf{C}}|^{1/2}}{|{\mathbf{C}}'|^{1/2}} \frac{w(C_{l} | C_{l}', {\mathbf{d}})}{w(C_{l}' | C_{l}, {\mathbf{d}})} \left( \frac{|{\mathbf{C}}'|^{1/2}}{|{\mathbf{C}}|^{1/2}} \right) \right] \nonumber \\ & = & \min \left[ 1, \frac{e^{- \chi^{2}({\mathbf{s}}', {\mathbf{d}})}}{e^{- \chi^{2}({\mathbf{s}}, {\mathbf{d}})}} \frac{w(C_{l} | C_{l}', {\mathbf{d}})}{w(C_{l}' | C_{l}, {\mathbf{d}})} \right] \label{eq:accept_prob}\end{aligned}$$ where the last line follows from the invariance of the quadratic form under the functional mapping ${\mathbf{s}}' [{\mathbf{C}}']^{-1} {\mathbf{s}}' = {\mathbf{s}}{\mathbf{C}}^{-1} {\mathbf{s}}$. Finally, we note that for the special case of a symmmetric proposal matrix where $w(C_{l}' | C_{l}, {\mathbf{d}}) = w(C_{l} | C_{l}', {\mathbf{d}})$, the accept probability is completely determined by the (exponeniated) change in $\chi^{2}$ $$A[{\mathbf{s}}', C_{l}' | {\mathbf{s}}, C_{l}] = \min \left[ 1, \frac{e^{- \chi^{2}({\mathbf{s}}', {\mathbf{d}})}}{e^{- \chi^{2}({\mathbf{s}}, {\mathbf{d}})}} \right]$$ As emphasized earlier in the main part of the text, the above allows large changes to the spectrum precisely where the signal to noise is getting small, as $\chi^{2}$ does not change much in this regime. Relation to Gibbs Sampling in a Change of Variables {#app:cov} =================================================== We note here another interesting approach to an MCMC algorithm in a [*different set of variables*]{} which in fact allows for large moves in the spectrum in the low signal to noise regime. We define the CMB map $${\mathbf{x}}= {\mathbf{C}}^{-1/2} {\mathbf{s}}$$ We therefore have the joint posterior [*in the new variables*]{} according to $$p(C_{l}, {\mathbf{s}}| {\mathbf{d}}) d(C_{l}, {\mathbf{s}}) = p(C_{l}, {\mathbf{x}}| {\mathbf{d}}) \left| \frac{\partial {\mathbf{s}}}{\partial x} \right| d(C_{l}, {\mathbf{x}})$$ which is explicitly, up to a normalization constant $$-2 \log p(C_{l}, {\mathbf{x}}| d) = ({\mathbf{d}}- {\mathbf{C}}^{1/2} {\mathbf{x}}) {\mathbf{N}}^{-1} ({\mathbf{d}}- {\mathbf{C}}^{1/2} {\mathbf{x}}) - \| {\mathbf{x}}\|^{2}$$ Then [*traditional Gibbs sampling in the new variables*]{} leads to an accept probability when changing the spectrum given the change of variable map $x$ as $$A(C_{l}', {\mathbf{x}}| C_{l}, {\mathbf{x}}) = \min \left[ 1, \frac{e^{-({\mathbf{d}}- [{\mathbf{C}}']^{1/2} {\mathbf{x}}) {\mathbf{N}}^{-1}({\mathbf{d}}- [{\mathbf{C}}']^{1/2} {\mathbf{x}})}} {e^{-({\mathbf{d}}- {\mathbf{C}}^{1/2} {\mathbf{x}}) {\mathbf{N}}^{-1}({\mathbf{d}}- {\mathbf{C}}^{1/2} {\mathbf{x}})}} \frac{w(C_{l} | {\mathbf{x}}, C_{l}', {\mathbf{d}})}{w(C_{l}' | {\mathbf{x}}, C_{l}, {\mathbf{d}})} \right]$$ where in the above the proposed variation in the spectrum can now be conditionally dependent on the current change of variable map ${\mathbf{x}}$. Assuming a symmetric proposal, or one conditionally independent of ${\mathbf{x}}$ leads to an accept probability which is [*numerically the same as \[eq:accept\_prob\]* ]{}, and also has the same property - large moves in the spectrum are possible in the low signal to noise regime. As a side note, we can see that $\log p(C_{l} | {\mathbf{x}}, {\mathbf{d}})$ is quadratic in $C_{l}^{1/2}$, and suggests a proposal given by a Gaussian in $C_{l}^{1/2}$. However there are two problems with this scheme - sampling in $C_{l}^{1/2}$ will result in re-introducing a Jacobian factor given by the ratio of $|C'|^{1/2} / |C|^{1/2}$ which results typically in low acceptance probabilities, and furthermore we cannot afford to exactly compute the local “Fisher” covariance matrix for each ${\mathbf{x}}$. Because of these difficulties, we in general need to produce a proposal for $C_{l}$ and then compute the accept probability above. We emphasize an important distinction between MCMC with deterministic steps [*in the original variables*]{} $(C_{l}, {\mathbf{s}})$ and Gibbs sampling in the change of variables $(C_{l}, {\mathbf{x}})$. It is only for the specific functional form that we have chosen for this paper that the numerical value of the accept probabilities for $A(C_{l}' , {\mathbf{s}}' | C_{l}, {\mathbf{s}})$ and $A(C_{l}', {\mathbf{x}}| C_{l}, {\mathbf{x}})$ are the same. At first glance, it might appear that a random variation in some of the variables followed by a deterministic change in the complementary set is always equivalent to random variation in a new set of variables. For notational convenience, we will assume the state space is separated into two sets of variables $({\mathbf{x}},{\mathbf{y}})$, i.e. for the CMB sampling context we have $({\mathbf{s}}, C_{l})$. Now, to make the distinction between a change of variables and deterministic steps in MCMC more precise, consider a “global” change of variables of the form $$\begin{aligned} {\mathbf{u}}& = & F({\mathbf{x}},{\mathbf{y}}) \nonumber \\ {\mathbf{v}}& = & {\mathbf{y}}\end{aligned}$$ with Jacobian $$\left| \begin{array}{cc} \frac{\partial {\mathbf{u}}}{\partial {\mathbf{x}}} & \frac{\partial {\mathbf{v}}}{\partial {\mathbf{x}}} \\ \frac{\partial {\mathbf{u}}}{\partial {\mathbf{y}}} & \frac{\partial {\mathbf{v}}}{\partial {\mathbf{y}}} \end{array} \right| = \left| \begin{array}{cc} \frac{\partial F}{\partial {\mathbf{x}}} & 0 \\ \frac{\partial F}{\partial {\mathbf{y}}} & {\mathbf{1}}\end{array} \right| = \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|$$ A Gibbs sampling step varying $v$ with $u$ fixed, has accept probability $$\begin{aligned} A({\mathbf{y}}_{n+1}, {\mathbf{u}}_{n} | {\mathbf{y}}_{n}, {\mathbf{u}}_{n}) & = & \min \left[ 1, \frac{\pi({\mathbf{y}}_{n+1} | {\mathbf{u}}_{n}, {\mathbf{d}})}{\pi({\mathbf{y}}_{n} | {\mathbf{u}}_{n}, {\mathbf{d}})} \frac{w({\mathbf{y}}_{n} | {\mathbf{u}}_{n}, {\mathbf{d}})}{w({\mathbf{y}}_{n+1} | {\mathbf{u}}_{n}, {\mathbf{d}})} \right] \nonumber \\ & = & \min \left[ 1, \frac{\pi({\mathbf{y}}_{n+1}, {\mathbf{x}}_{n+1} | {\mathbf{d}})}{\pi({\mathbf{y}}_{n} , {\mathbf{x}}_{n} | {\mathbf{d}})} \left( \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|_{{\mathbf{x}}_{n+1}, {\mathbf{y}}_{n+1}} \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|^{-1}_{{\mathbf{x}}_{n}, {\mathbf{y}}_{n}} \right) \frac{w({\mathbf{y}}_{n} | {\mathbf{u}}_{n}, {\mathbf{d}})}{w({\mathbf{y}}_{n+1} | {\mathbf{u}}_{n}, {\mathbf{d}})} \right] \nonumber \\\end{aligned}$$ where in the above we have the constraint $$\begin{aligned} {\mathbf{x}}_{n+1} & = & F^{-1}({\mathbf{u}}_{n}, {\mathbf{y}}_{n+1}) \nonumber \\ {\mathbf{x}}_{n} & = & F^{-1}({\mathbf{u}}_{n}, {\mathbf{y}}_{n}) \end{aligned}$$ Now consider an MCMC step in the original variables of the form $$w({\mathbf{x}}_{n+1}, {\mathbf{y}}_{n+1} | {\mathbf{x}}_{n}, {\mathbf{y}}_{n }) = w({\mathbf{y}}_{n+1} | {\mathbf{y}}_{n}, {\mathbf{x}}_{n}, {\mathbf{d}}) \delta \left( x_{n+1} - H(x_{n}, y_{n+1}, y_{n}) \right)$$ with general accept probability, according to the discussion above $$\begin{aligned} A({\mathbf{y}}_{n+1}, {\mathbf{x}}_{n+1} | {\mathbf{y}}_{n}, {\mathbf{x}}_{n}) & = & \min \left[ 1, \frac{\pi({\mathbf{y}}_{n+1}, {\mathbf{x}}_{n+1} | {\mathbf{d}})}{\pi({\mathbf{y}}_{n}, {\mathbf{x}}_{n} | {\mathbf{d}})} \frac{w({\mathbf{y}}_{n} | {\mathbf{x}}_{n}, {\mathbf{y}}_{n+1}, {\mathbf{d}})}{w({\mathbf{y}}_{n+1} | {\mathbf{x}}_{n+1}, {\mathbf{y}}_{n}, {\mathbf{d}})} \left( \left| \frac{\partial H}{\partial {\mathbf{x}}} \right|^{-1}_{{\mathbf{x}}_{n+1} = H^{-1}({\mathbf{x}}_{n},{\mathbf{y}}_{n}, {\mathbf{y}}_{n+1})} \right) \right]\end{aligned}$$ Interestingly enough this suggests that we can set $H$ to be the function $$H({\mathbf{x}}, {\mathbf{y}}_{n+1}, {\mathbf{y}}_{n}) = F^{-1} \left( F({\mathbf{x}}, {\mathbf{y}}_{n}), {\mathbf{y}}_{n+1} \right)$$ Does this function have the correct properties for its inverse? Assuming we have computed in the forward direction ${\mathbf{x}}' = H({\mathbf{x}}, {\mathbf{y}}_{n+1}, {\mathbf{y}}_{n})$, we can invert to find $x$ by computing sequentially $$\begin{aligned} F({\mathbf{x}}', {\mathbf{y}}_{n+1}) & = & F({\mathbf{x}}, {\mathbf{y}}_{n}) \nonumber \\ {\mathbf{x}}& = & F^{-1} \left( F({\mathbf{x}}', {\mathbf{y}}_{n+1}), {\mathbf{y}}_{n} \right) \nonumber \\ & \equiv & H({\mathbf{x}}', {\mathbf{y}}_{n}, {\mathbf{y}}_{n+1})\end{aligned}$$ where the last line follows from definition of the forward $H$. Since we have, by definition $$\begin{aligned} {\mathbf{x}}' & = & H({\mathbf{x}}, {\mathbf{y}}_{n+1}, {\mathbf{y}}_{n}) \nonumber \\ {\mathbf{x}}& \equiv & H^{-1}({\mathbf{x}}', {\mathbf{y}}_{n+1}, {\mathbf{y}}_{n}) \nonumber \end{aligned}$$ we therefore have shown that $$H^{-1}({\mathbf{x}}', {\mathbf{y}}_{n+1}, {\mathbf{y}}_{n}) = H({\mathbf{x}}', {\mathbf{y}}_{n}, {\mathbf{y}}_{n+1})$$ as required for a non-vanishing accept probability. The above as a function of $x$ has Jacobian $$\begin{aligned} \left| \frac{\partial H}{\partial {\mathbf{x}}} \right| & = & \left| \frac{\partial F^{-1}}{\partial {\mathbf{u}}} \right|_{( {\mathbf{u}}({\mathbf{x}}, {\mathbf{y}}_{n}), {\mathbf{y}}_{n+1})} \ \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|_{({\mathbf{x}},{\mathbf{y}}_{n})} \nonumber \\ & = & \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|^{-1}_{( {\mathbf{x}}, {\mathbf{y}}_{n+1})} \ \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|_{({\mathbf{x}},{\mathbf{y}}_{n})} \nonumber \\\end{aligned}$$ However, when evaluated at ${\mathbf{x}}_{n+1} = H^{-1}({\mathbf{x}}_{n}, {\mathbf{y}}_{n}, {\mathbf{y}}_{n+1})$, we will not in general satisfy the required equality required for numerical equivalence $$\left( \left| \frac{\partial H}{\partial {\mathbf{x}}} \right|^{-1}_{{\mathbf{x}}_{n+1} = H^{-1}({\mathbf{x}}_{n},{\mathbf{y}}_{n}, {\mathbf{y}}_{n+1})} \right) \neq \left( \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|_{{\mathbf{x}}_{n+1}, {\mathbf{y}}_{n+1}} \left| \frac{\partial F}{\partial {\mathbf{x}}} \right|^{-1}_{{\mathbf{x}}_{n}, {\mathbf{y}}_{n}} \right) \label{eq:jacobian_identity}$$ So in general, while we can use any function $F({\mathbf{x}},{\mathbf{y}})$ to generate deterministic moves in the original variables within MCMC, this is not equivalent to a Gibbs sampling step $p({\mathbf{y}}_{n+1} | {\mathbf{u}}_{n}, {\mathbf{d}})$ in the new variables using $(F({\mathbf{x}},{\mathbf{y}}), {\mathbf{y}})$ as a global change of variables. However, using the above construction for the CMB change of variables, we have explcitly $$\begin{aligned} F^{-1} \left( F({\mathbf{s}}, C_{l}), C_{l}' \right) & = & [{\mathbf{C}}']^{1/2} \left( {\mathbf{C}}^{-1/2} {\mathbf{s}}\right)\end{aligned}$$ which is exactly the functional form used for the deterministic MCMC steps. In this case, it is because the Jacobian of our deterministic change in the CMB map is independent of the current CMB map ${\mathbf{s}}$ (and only dependent on the proposed and current spectra) that we have numerical equivalence of the accept probabilities. So in summary, while we can use any mapping $F({\mathbf{x}},{\mathbf{y}})$ to generate deterministic steps for use in MCMC, the accept probability is not equivalent to a conditional step $p({\mathbf{y}}| {\mathbf{u}}, {\mathbf{d}})$ using $F({\mathbf{x}},{\mathbf{y}})$ in a change of variables due to the general “location” dependence of the Jacobian. Furthermore, setting $H({\mathbf{x}}, {\mathbf{y}}', {\mathbf{y}}) = F^{-1} ( F({\mathbf{x}},{\mathbf{y}}), {\mathbf{y}}')$ is not the most general form for a function that satisfies the detailed balance requirement $H^{-1}({\mathbf{x}}, {\mathbf{y}}', {\mathbf{y}}) = H({\mathbf{x}}, {\mathbf{y}}, {\mathbf{y}}')$. In this sense then, a change of variables as an approach to more efficiently generating samples from a probability density is distinct from a strategy of designing an MCMC algorithm (in any chosen representation of the variables) with deterministic changes of some of the degrees of freedom. Both approaches are interesting, and advances in either approach for Bayesian CMB analysis could lead to improvements over the approach presented in this paper. Abramowitz, M., & Stegun, I. A. 1972, Handbook of Mathematical Functions, New York: Dover, 1972, Bennett, C. L., et al.2003a, , 148, 1 Chu, M., Eriksen, H. K., Knox, L., G[ó]{}rski, K. M., Jewell, J. B., Larson, D. L., O’Dwyer, I. J., & Wandelt, B. D. 2005, , 71, 103002 Eriksen, H. K., et al. 2004, , 155, 227 Eriksen, H. K., et al.  2007a, , 656, 641 Eriksen, H. K., Huey, G., Banday, A. J., G[ó]{}rski, K. M., Jewell, J. B., O’Dwyer, I. J., & Wandelt, B. D. 2007b, , 665, L1 Eriksen, H. K., Jewell, J. B., Dickinson, C., Banday, A. J., G[ó]{}rski, K. M., & Lawrence, C. R. 2008a, , 676, 10 Eriksen, H. K., Dickinson, C., Jewell, J. B., Banday, A. J., G[ó]{}rski, K. M., & Lawrence, C. R. 2008b, , 672, L87 Eriksen, H. K. & Wehus, I. K. 2008a, , submitted, \[astro-ph/XXXXXX\] Gelman, A., & Rubin, D. 1992, Stat. Sci., 7, 457 Gold, B., et al. 2008, \[arXiv:0803.0715\] G[' o]{}rski, K. M., Hivon, E., Banday, A. J., Wandelt, B. D., Hansen, F.K., Reinecke, M., & Bartelmann, M. 2005, , 622, 759 Green, P.; 1995, Biometrika, 82: 711-732 Gupta, A. K. & Nagar, D. K. 2000, Matrix Variate Distributions Hinshaw, G., et al.  2007, , 170, 288 Hinshaw, G., et al.  2008, ApJ, submitted, \[arXiv:0803.0732\] Hivon, E., G[' o]{}rski, K. M., Netterfield, C. B., Crill, B. P., Prunet, S., & Hansen, F. 2002, , 567, 2 Jewell, J., Levin, S., & Anderson, C. H. 2002, astro-ph 0209560v1 Jewell, J., Levin, S., & Anderson, C. H. 2004, , 609, 1 Komatsu, E., et al.  2008, \[arXiv:0803.0547\] Larson, D. L., Eriksen, H. K., Wandelt, B. D., G[ó]{}rski, K. M., Huey, G., Jewell, J. B., & O’Dwyer, I. J. 2007, , 656, 653 Liu, J. S., Monte Carlo Strategies in Scientific Computing, Cambridge, USA: Springer, 2001, O’Dwyer, I. J., et al.  2004, , 617, L99 Page, L., et al. 2007, , 170, 335 Seljak, U., & Zaldarriaga, M. 1996, , 469, 437 Smoot, G. F., et al.  1992, , 396, L1 Sokal, A.D.; “Monte Carlo methods in statistical mechanics: foundations and new algorithms”, [*Cous de Troisième Cycle de la Physique en Suisse Romande*]{}, Lausanne. Taylor, J. F., Ashdown, M. A. J., & Hobson, M. P. 2007, MNRAS, submitted, \[arXiv:0708.2989\] Wandelt, B. D., Larson, D. L., & Lakshminarayanan, A. 2004, , 70, 083511 Zaldarriaga, M., & Seljak, U. 1997, , 55, 1830 [^1]: http://healpix.jpl.nasa.gov [^2]: http://healpix.jpl.nasa.gov
{ "pile_set_name": "ArXiv" }
--- title: Defect CFTs and holographic multiverse --- Introduction ============ In the quest to formulate a quantum theory of gravity, the holographic principle [@'tHooft:1993gx] is an appealing model-independent guiding principle, that has received a tremendous impetus after finding a concrete realization in the AdS/CFT correspondence [@Maldacena:1997re]. Given the wide and deep ramifications of the AdS/CFT correspondence, it is natural to try to realize the holographic principle as a way to formulate quantum gravity in spacetimes with different asymptotia, although there is no a priori guarantee that the problem is well defined or has a manageable answer. Since inflation plays a central role in modern cosmology, it has attracted a fair share of activity trying to provide a holographic dual for various inflationary models. One of the first ideas in this direction is due to Strominger, who suggested [@Strominger:2001gp] that a single inflating bubble is dual to a renormalization group flow between two Euclidean conformal fixed points. In this picture, the theory at the future infinity boundary would be the UV fixed point theory, so the time evolution of the universe would be dual to an upstream renormalization group flow. There have been many works trying to extend this proposal to the scenario of eternal inflation [@Freivogel:2006xu; @Freivogel:2009rf; @Bousso:2009dm], out of which we will focusing our attention in the recent proposal of the holographic multiverse [@Garriga:2008ks; @Garriga:2009hy]. The holographic multiverse proposal claims that there exists a holographic dual to eternal inflation, and that, as in [@Strominger:2001gp], it can be defined at the future infinity boundary of spacetime. Since now future infinity is boundary to regions in spacetime with different cosmological constants, it is expected that the number of degrees of freedom is not constant along the 3d boundary, but just constant on patches, with the particular value of a domain in the boundary tied to the Hubble expansion parameter of the corresponding region in the bulk of spacetime. Furthermore, the authors of [@Garriga:2008ks; @Garriga:2009hy] left open the possibility that additional degrees of freedom live at the 2d walls separating 3d domains of the boundary. In this note we want to focus precisely on these 2d walls on the hypothetical dual theory, and the implications that the dynamics of bubbles in the bulk of spacetime have for them. As we will argue, these issues lead naturally to consider defect conformal field theories, i.e. conformal field theories with a defect of lower dimension that preserves a subgroup of the original conformal symmetry group [@Cardy:1984bb]. Defect conformal field theories have been studied using the AdS/CFT correspondence [@Karch:2000gx; @DeWolfe:2001pq; @Aharony:2003qf; @Constable:2002xt], and we will also rely on this correspondence for our specific computations. To set the stage, consider the simplest scenario: two 4d deSitter solutions with different values of the cosmological constant, separated by a thin wall bubble. In the hypothetical dual theory defined at the future infinity boundary, this ought to correspond to two 3d regions with different UV CFTs (and different number of degrees of freedom), separated by a codimension one defect with some additional degrees of freedom living on it. As we review below, one can study the fluctuation modes of the bubble to try to narrow down the possible dual theory. The analysis of the bubble fluctuations is standard, but trying to interpret them as coming from a dual theory leads to some counterintuitive results [@Garriga:2009hy]; for instance, it was found that the quantity expected to correspond to the number of degrees of freedom with non-trivial boundary conditions at the wall can actually be smaller than the difference between the number of degrees of freedom on the two sides. This puzzling result implies that if the proposed duality does exist, we must sharpen our understanding on how it maps bubbles to defects in the dual theory, and what kind of interactions it requires among the defect degrees of freedom and the ambient ones in the dual theory. The purpose of this note is to improve our very limited casuistics, by working out a concrete example with a known 2+1 CFT. In setting up a field theory computation, we have to choose the three dimensional field theory and the two dimensional defect we are going to add to it. Let’s motivate our choices in turns. As for the choice of field theory, a possibility might be to start with a free field theory, but one should have reservations about a free field theory correctly capturing a regime where gravity is weakly coupled[^1]; we will not pursue this direction. If on the other hand, we consider a strongly coupled 3d CFT, our chances of performing a computation directly in the CFT are severely limited. As a way out of this conundrum, I will consider a 3d CFT that has a known AdS dual, namely, the IR limit of maximally supersymmetric 2+1 $SU(N)$ SYM theory, which is expected to be dual to M-theory on $AdS_4\times S^7$. In the large N limit, this reduces to something we can work with, 11D SUGRA on $AdS_4\times S^7$; having to take the large N limit is presumably not a limitation for the problem at hand, since we are interested in understanding the putative dual of large space-times, which should involve some sort of large N limit. A potential reservation is that while our aim is trying to make sense of some puzzling properties of a hypothetical Euclidean theory, which is presumably non-unitary [@Strominger:2001pn], we consider a Lorentzian unitary 2+1 field theory. While it is clear that the field theory we consider has to differ in some significant way from the theory we are after, the reason that our choice is eventually useful is that the computations we perform using AdS/CFT end up tracking closely the ones performed in [@Garriga:2009hy], so we can understand the different outcomes easily. As for the defect, we want a 1+1 wall in the 2+1 boundary. We can accomplish that by suitably embedding a probe brane in the background. Furthermore we want that the amount of flux on both sides of the brane is different, reflecting a difference in the value of the vacuum density energies. We accomplish this by adding a probe M5-brane with world-volume magnetic flux turned on. The M5-brane world-volume is $AdS_3\times S^3$, with the $S^3$ piece inside $S^7$, so from the four dimensional point of view it is a domain wall. We consider fluctuations about this probe brane embedding, and extract the contribution to the trace anomaly due to the 1+1 wall in the boundary. The results that we find for this specific example do not display any of the counterintuitive features encountered in [@Garriga:2009hy]. In the case at hand, the key difference appears to be that while in the $dS$ computations the curvature radius of the bubble is always equal or smaller than the Hubble expansion parameter, in $AdS$ the opposite is true. Although we have worked out a single example, this particular feature is common for $AdS_p$ branes in $AdS_q$ spaces, and we believe that this feature is the reflection on the gravity side that the behavior of the boundary degrees of freedom conforms to intuitive expectations. Given these results, there are various attitudes one can take. The first possibility is to take them as evidence that the duality proposed in [@Garriga:2008ks; @Garriga:2009hy] does not make sense. A less drastic possibility is that the type of unusual behavior encountered in [@Garriga:2009hy] doesn’t happen for Lorentzian CFTs which admit an AdS dual; it should be kept in mind that the Euclidean 3d CFT conjectured in [@Garriga:2008ks; @Garriga:2009hy] does not necessarily correspond to the Wick rotation of a sensible Lorentzian CFT. In the discussion section, we suggest that if the holographic multiverse proposal makes sense, the degrees of freedom in the defect might have to be non-unitarity. The plan of the paper is as follows. In section 2, we review the results of [@Garriga:2009hy] for bubble fluctuations and the difficulties they pose in trying to interpret them holographically. In section 3, we present the setup of the computation we are going to perform, by introducing the supergravity background dual to the CFT we consider, and the M5 probe brane that introduces the defect on the CFT. In section 4 we focus on a particular fluctuation mode around the solution, and compute its regularized action, which allows us to read off the contribution of this defect to the integrated trace anomaly of the CFT, and compare the result with those found in [@Garriga:2009hy]. We conclude with some discussion about possible implications for the hypothesized duality of [@Garriga:2008ks; @Garriga:2009hy]. Bubble fluctuations =================== The authors of [@Garriga:2009hy] proposed that eternal inflation is holographically dual to an Euclidean 3-dimensional field theory, controlled in the UV by a conformal fixed point. Their proposal extends ideas of Strominger about inflation as an upstream renormalization group flow [@Strominger:2001gp]. This CFT, and the renormalization group flow that describes the multiverse, if they exist at all, are probably bewilderingly complicated. As a way to extract concrete predictions out of their general proposal, the authors of [@Garriga:2009hy] considered a simplified scenario where a single spherical bubble separates two universes; the bubble has a thin wall with $dS_3$ metric separating two $dS_4$ spaces, that in general have different vacuum density energy. Analysis of fluctuations of this type of bubbles have been around for a while [@Garriga:1991ts]; the novel question addressed in [@Garriga:2009hy] is what can one deduce about the potential holographic dual from the knowledge of the bubble fluctuation spectrum. In this section we review their results, recasting them in a language more common in the AdS/CFT literature. The bulk spacetime is taken to be $dS_4$. It is best to consider $dS_4$ in flat coordinates $$ds^2_{dS_4}=-dt^2+e^{2Ht}\left(dr^2+r^2d\Omega_2\right)$$ since it makes manifest the identification of time evolution in the bulk with scale transformations in the boundary [@Strominger:2001gp], while for the induced metric on the $dS_3$ bubble we use conformal coordinates $$ds^2_{dS_3}=\left(\frac{R_0}{\cos \eta}\right)^2\left(-d\eta^2+d\Omega_2\right)\equiv \tilde a(\eta)^2 \left(-d\eta^2+d\Omega_2\right)$$ where $R_0$ is the curvature radius of the bubble. In the thin wall approximation one can express $R_0$ in terms of the tension $T$ of the bubble, the Hubble expansion parameter $H$, and the jump in the vacuum energy density between the interior and the exterior of the bubble $\Delta \rho_V$ [@Berezin:1987bc], $$R_0^2 \sim \frac{9T^2}{9H^2T^2+(\Delta \rho_V)^2} \label{thinwall}$$ As expected on geometric grounds, it follows from this formula that the curvature radius of the wall is always equal or smaller than that of the background, $R_0\leq H^{-1}$. In the thin wall approximation, the only fluctuations considered are the ones associated with the transverse position of the bubble wall. These can be captured by a scalar field $\phi$ living on the wall [@Garriga:1991ts]. For the specific case at hand, a thin $dS_3$ wall in $dS_4$, it turns out that this scalar field is tachyonic, with mass [@Garriga:1991ts] $$m^2_\phi=-\frac{3}{R_0^2} \label{tacmass}$$ This tachyonic mass can be understood as indicating that as the bubble expands, the fluctuations grow accordingly. In trying to pin down the possible dual field theory, the first thing that might come to mind is to determine the central charge of the UV CFT in terms of gravitational data, following the steps of [@Brown:1986nw; @Strominger:2001pn]. At first, this idea doesn’t look promising, as 3d CFTs defined on manifolds without boundaries have no trace anomaly. However, the presence of a domain wall (a 2d defect) can change that, if the CFT fields are required to satisfy non-trivial boundary conditions on it. A way to think of this is to consider that there are additional degrees of freedom living on the defect (placed at, say, $x=0$), so the full stress-energy tensor of the boundary theory is [@DeWolfe:2001pq; @Aharony:2003qf] $$T_{\mu \nu}^{TOTAL}=T_{\mu \nu}^{3d}+\delta(x)T^{2d}_{ij}\delta ^i_\mu \delta^j_\nu$$ and the defect stress-energy tensor $T^{2d}_{ij}$ can give rise to a non-zero integrated trace anomaly, even if $T^{3d}_{\mu \nu}$ cannot. Indeed, if $K_{ij}$ and $\hat R_{ij}$ are the extrinsic and intrinsic curvatures of the 2d defect, the integrated trace anomaly is given by an expression of the form [@Schwimmer:2008yh] $$a_{3/2}=\int d \Sigma_2 \sqrt{\hat g}\left[d_1(K_{ab}K^{ab}-\frac{1}{2}K^2)+d_2\hat R\right] \label{confan}$$ where we already assumed that the ambient 3d space is flat. The coefficients $d_1$ and $d_2$ depend on the specific theory. The term with $\hat R$ is the usual topological term in 2d, while the first term is only possible because the 2d defect is embedded into a 3d ambient space. This first term is sensitive to variations of the shape of the defect, and it was argued in [@Garriga:2009hy] that it is holographically dual to the action that governs the fluctuations of the bubble. This is in accordance with the standard AdS/dCFT dictionary [@Karch:2000gx; @DeWolfe:2001pq], where the probe brane degrees of freedom (bubble degrees of freedom in our case) are holographically dual to degrees of freedom on the defect (the 2d wall in our case). The arguments of [@Garriga:2009hy] allow to estimate the coefficient $d_1$ of the dCFT, through a computation of the action for the fluctuations of the bubble. A sketch of their computation goes as follows: while for an observer on the unperturbed brane the field that naturally captures fluctuations of the brane is the one measuring deviations normal to the brane - the field $\phi$ mentioned above - to make contact with a possible boundary theory, it is more convenient to trade it by a field, dubbed $\delta$ in [@Garriga:2009hy], that measures fluctuations for slices of constant $t$. An straightfoward evaluation of the action for these fluctuations yields a divergent result, as it is well-known to happen in AdS/CFT for the bulk holographic action, both for the full CFT and for submanifold observables. It is therefore necessary to introduce a cut-off value for the scale factor $\tilde a(\eta)$ and consider a regularized holographic action. It turns out that this regularized action contains a term logarithmic in the $\tilde a(\eta)$ cut-off , just as it is the case for even dimensional CFTs [@Henningson:1998ey] and even dimensional submanifold observables [@Berenstein:1998ij; @Graham:1999pm]. This logarithmic term induces a conformal anomaly for the finite piece of the regularized action. Thus, the anomaly of the renormalized action is fixed in terms of the coefficient of the logarithm term in the regularized action. In the case at hand, it is argued in [@Garriga:2009hy] that the log term of the regularized bubble action corresponds to the term with coefficient $d_1$ in the expression for the integrated trace anomaly (\[confan\]), allowing to determine it. Omitting numerical factors, this is given by $$d_1\sim \frac{TR_0}{H^2} \label{resultd}$$ The authors of [@Garriga:2009hy] further interpreted $d_1$ as roughly a measure of the number of degrees of freedom of the hypothetical dual field theory that must satisfy non-trivial boundary conditions at the defect (their reasoning being that if the wall was transparent to all the degrees of freedom, there would be no anomaly). However, this interpretation of $d_1$ and the estimate (\[resultd\]) are somewhat at odds. A first conterintuitive feature of this result is that for fixed $T$ and $H$, as $\Delta \rho_V$ increases, it follows from (\[thinwall\]) that $R_0$ decreases, and so does $d_1$. This would imply that as the two universes separated by the wall differ more, there are [*less*]{} degrees of freedom satisfying non-trivial boundary conditions at the wall. A second counterintuitive feature, emphasized already in [@Garriga:2009hy] is that, if we denote by $c_i$ the number of degrees of freedom on each side of the wall, $\Delta c$ should be a solid lower bound on the number of degrees of freedom with non-trivial boundary conditions, as it gives the difference of degrees on freedom between the two sides. Therefore, this interpretation of $d_1$ leads to the conclusion that $$d_1\geq \Delta c \label{puzzle}$$ However, it can be shown [@Garriga:2009hy] that this implies, within the thin wall approximation, that $R_0\sim H^{-1}$. One arrives then at the conclusion that the regime $R_0\ll H^{-1}$, while it can be reliably considered within the thin wall approximation, leads to a violation of (\[puzzle\]), so the intuition that led to the previous inequality must fail in this regime. To clarify the implications of these results, it seems appropriate to sharpen our intuition about the meaning and parametric dependence of $d_1$ in a situation when we do have a CFT realization. We turn now to that task. An $AdS_4$/dCFT computation: the setup ====================================== The 2+1 CFT we are going to consider is the IR fixed point of maximally supersymmetric 2+1 $SU(N)$ SYM. An explicit Lagrangian description of this CFT became available only recently [@Aharony:2008ug], in terms of a supersymmetric Chern-Simons theory with two gauge groups coupled to matter. This conformal theory does not have a marginal coupling that allows us to go to a free field theory limit: the theory is inherently strongly coupled, so direct computations on the CFT would be quite hard to perform. Furthermore, while there has been some work in providing a direct description of defects on this CFT, the results are not conclusive [@Ammon:2009wc]. Since we are not going to be using this Lagrangian description, we won’t present it here, referring the interested reader to [@Aharony:2008ug] for further details. For our purposes, what is important about this CFT is that it is the low energy limit of the world-volume theory of a stack of N M2-branes, and it is conjectured to be dual to M-theory on $AdS_4\times S^7$. In the large $N$ limit, this reduces to 11d SUGRA on the same background, becoming a tractable setup that allows explicit computations. Furthermore, it is straightforward to add a probe brane that reaches the boundary at a codimension one defect. To make contact with the discussion of the previous section, we are interested in having a defect that separates two CFTs with different numbers of degrees of freedom, and for that we will use the general idea of Karch and Randall [@Karch:2000gx] of having some of the M2-branes ending on the M5-brane probe. This means that the probe brane presents a bending, compensated by a non-trivial magnetic flux on the world-volume. A difference with the discussion in the previous section is that the defect we will consider here is a flat 1+1 wall, rather than a sphere. Before we embark in the actual computations, rethinking the arguments of [@Garriga:2009hy] in the light of AdS/dCFT leads to a refined interpretation of the coefficient $d_1$. When one considers defect CFTs in the AdS/CFT correspondence by adding a probe brane to an AdS background, there are two kinds of [*new*]{} degrees of freedom, those living on the brane and those living on the defect, and it is expected that holography acts a ’second time’ making them dual to each other in the appropriate limit. For instance, in the extensively studied case of a D5 probe in the background of D3 branes [@Karch:2000gx; @DeWolfe:2001pq], there are open strings with both ends on the D5 probe, that are additional degrees of freedom on the gravity side, and in the decoupling limit they are expected to be holographically dual to the degrees of freedom coming from open strings coming from 3-5 strings, that live on the defect, and enlarge the content of the boundary field theory. Since we are using brane fluctuations to determine the coefficient $d_1$, this coefficient must be sensitive to these [*new*]{} degrees of freedom on the defect. This interpretation of $d_1$ does not neccesarily clash with the previous one: in fact, ordinarily the degrees of freedom on the defect can be seen as enforcers of specific boundary conditions for the ambient fields at the defect, and can even be integrated out, leaving its mark as non-trivial boundary conditions for the ambient fields (see [@Gomis:2006sb] for a nice illustration of this). However, given the subtleties of the present problem, we prefer to keep these two points of view in sight. This interpretation already eases the tension in comparing $d_1$ with $\Delta c$, since $d_1$ is not immediately counting ambient degrees of freedom already present before adding the defect, and details about the couplings between defect and ambient fields in the boundary theory might alter the arguments leading to the inequality (\[puzzle\]). Nevertheless, it still leaves open the interpretation of the fact that in deSitter, $d_1$ decreases as the difference in vacuum energy on the two sides of the wall increases. We will return to this point in the discussion section. The background -------------- The bosonic content of 11d SUGRA is the 11d metric and a 3-form $C_{3}$. A maximally supersymmetric solution has metric of the form $AdS_4 (R/2) \times S^7(R)$ for any value of $R$. We write the $AdS_4\times S^7$ solution in a way that makes manifest a 2+1 Minkowski boundary, $$ds^2=\frac{r^4}{R^4}dx_{1,2}^2+\frac{R^2}{r^2}d\vec r^2$$ $$C_{(3)}=\frac{r^6}{R^6}dx^0\wedge dx^1\wedge dx^2$$ where $\vec r$ is an 8-dimensional vector in $R^8$. For what follows, it is convenient to split this 8d vector into two 4-d vectors, $\vec y$ and $\vec z$, with $\rho$ the norm of $\vec y$. The background metric reads then $$ds^2=\frac{(\rho^2+\vec z^2)^2}{R^4}dx^2_{1,2}+\frac{R^2}{\rho^2+\vec z^2}\left(d\rho^2+\rho^2 d\Omega_3^2+d\vec z^2\right) \label{pickst}$$ The reason to perform this split is to make manifest a choice of an $S^3$ inside $S^7$, that will become handy next, when we discuss the embedding of the probe brane. Adding a defect: the probe brane -------------------------------- The supergravity solution just presented is expected to be dual to a Lorentz invariant 2+1 CFT. We now want to add a codimension one defect on the CFT. These defects haven’t been much studied directly for the CFT at hand, but we can work around that by considering the dual description of adding a defect to the CFT. This is given by considering a probe brane in the previous geometry, such that it reaches the boundary of $AdS_4$ as a codimension one defect. The probe we consider is an M5-brane. The world-volume of a single M5-brane consists of a 6d $(2,0)$ tensor multiplet, whose bosonic content is given by five scalar fields and a self-dual three-form field strenght $F=dB$. The fact that the three-form on the world-volume of the M5-brane is self-dual makes it notoriously difficult to write down an action. A way to circumvect this problem is to introduce an extra non-dynamical scalar field $a(\xi)$, with a non-polynomial action; we will follow this route and as the world-volume action of the M5 brane we will take the PST action [@Pasti:1997gx], $$S_{PST}=T_{M5}\int d^6\xi \left[-\sqrt{-|g_{ij}+\check H_{ij}|}+\frac{\sqrt{-|g|}}{4\partial a \cdot \partial a} \partial _i a (*H)^{ijk}H_{jkl}\partial^l a\right] +T_{M5}\int \frac{1}{2}F\wedge P[C_{(3)}] \label{pstact}$$ where $H$ is a 3-form that combines $F$ and the pullback on the world-volume of the background 3-form potential $C_{(3)}$, $$H=F-P[C_{(3)}]$$ and $$\check H^{ij}=\frac{1}{3!\sqrt{-|g|}}\frac{1}{\sqrt{-(\partial a)^2}}\epsilon^{ijklmn}\partial _k a H_{lmn}$$ Now we look for a M5 embedding in the $AdS_4\times S^7$ geometry. Since we want a codimension one defect on the 2+1 boundary, we take $x^0,x^1$ as two of the world-volume coordinates. We also take $\rho$ as a world-volume coordinate, so the brane expands in the $AdS_4$ bulk as well. Finally, we identify the three remaining world-volume coordinates with the coordinates of the $S^3$ inside $S^7$ that we singled out in (\[pickst\]). With this identification of world-volume coordinates, a simple solution for the embedding equations has $x^2$ constant and the rest of the world-volume fields set to zero. This solution has world-volume metric $AdS_3\times S^3$ and reaches the boundary at a codimension one defect, but for our purposes it has the fatal drawback that the bulk on the two sides is the same. A way to get a jump in the number of degrees of freedom between the two sides of the brane was discussed by Karch and Randall [@Karch:2000gx]. In the brane configuration, before taking the near-horizon limit, it consists in having $k$ M2-branes ending on the M5. This causes a non-trivial bending on the M5, implying we must allow for a non-trivial $x^2(\rho)$, which is accompanied by a magnetic flux for three-form living on the world-volume on the brane, in the internal $S^3$. Armed with this insight, as ansatz for the remaining world-volume fields we take $$x^2=x^2(\rho),\hspace{1cm}\vec z=0, \hspace{1cm} F=q \hbox{ vol} (S^3)$$ and following [@Arean:2007nh] fix the gauge $a=x^1$. This ansatz is actually a particular case of the one considered in [@Arean:2007nh]. They consider a more general ansatz, with $|\vec z|=L$, that gives mass to some fields of the dual CFT, breaking conformal invariance. This more general ansatz might be relevant if we want to use this type of constructions to study the proposed duality beyond its UV fixed point, but in this work we will focus on this fixed point limit, and keep $\vec z=0$. The solution found in [@Pasti:1997gx; @Arean:2007nh] for this ansatz is $$x^2(\rho)=x^2(\infty)+\frac{q}{2\rho^2} \label{solprobe}$$ It can be checked that this is a $1/2$ BPS solution. The induced metric on the M5-brane world-volume is $$ds^2_{M5}=\frac{\rho^4}{R^4}dx^2_{1,1}+R^2 \left(1+\frac{q^2}{R^6}\right) \frac{d\rho^2}{\rho^2}+R^2d\Omega_3^2 \label{wvmetric}$$ so it is of the form $AdS_3(R_{eff}/2)\times S^3(R)$, with $$R_{eff}^2=R^2\left(1+\frac{q^2}{R^6}\right) \label{radeff}$$ We see that as consequence of a non-zero magnetic flux ($q\neq 0$), the curvature radius of the wall $R_{eff}$ is no longer equal to the Anti de Sitter radius $R$, but contrary to what happened for the thin wall in de Sitter, it’s always bigger, $R_{eff}\geq R$. This obvious difference between de Sitter and anti de Sitter embeddings appears to be at the root of the difficulties in giving an intuitive interpretation of (\[resultd\]). Computing the conformal anomaly =============================== Our next task is to compute the coefficient $d_1$ in the conformal anomaly (\[confan\]), for the defect we have presented in the previous section. As reviewed in section 2, in [@Garriga:2009hy] it was pointed out that a way to do so is to consider fluctuations of the brane inducing the defect, consider a regularized action with a cut-off in the holographic direction, and read off the coefficient from the term with the logarithmic dependence on the cut-off. Fluctuations ------------ To compute the renormalized action for fluctuation of the position of the brane, we need to know the Lagrangian controlling the fluctuations of the world-volume fields. This has been extensively studied in the case of D-branes, where the metric that appears is the open string metric, which needs not to coincide with the closed string metric (i.e. the induced world-volume metric). For M5-branes, it was argued in [@Gibbons:2000ck] that a close analog of the open string metric also controls the fluctuations, and this has been checked for a subset of the fluctuations [@Arean:2007nh]. A detailed analysis of the fluctuation spectrum, including fluctuations of all the world-volume fields, will be presented elsewhere [@inprep]. We are going to consider just fluctuations of the position of the brane, but since in general the equations of motion for the fluctuations are coupled, we need to argue that it is consistent to do so. First, it turns out [@inprep] that fluctuations of the scalars describing transverse directions to the $S^3$ in $S^7$ decouple from the rest, as it happens in analogous D-brane systems [@DeWolfe:2001pq], so it is consistent to set them to zero. Second, since the fluctuation of the position of the brane $x^2$ is a scalar in the $AdS_3$ part of the M5 world-volume, in the equations of motion of the fluctuations of the 2-form that have some $S^3$ index, it only appears with derivatives with respect to $S^3$ coordinates. The upshot is that, as long as we consider an $x^2$ fluctuation independent of $S^3$ coordinates, it decouples from the rest, and it is consistent to consider it on its own. This suits us well, since in the previous section the fluctuation of the $dS_3$ brane didn’t depend on any internal coordinates either. So we consider small fluctuations of $x^2$ around the solution (\[solprobe\]), $$x^2(\xi)=x^2_{sol}+h(\xi)$$ and expand the action (\[pstact\]) around the solution up to quadratic order in $h(\xi)$. Before expanding the action, it is convenient to rewrite the second term in (\[pstact\]) as $$\frac{\sqrt{-|g|}}{4\partial a \cdot \partial a}\partial _i a (*H)^{ijk}H_{jkl}\partial^l a= \frac{1}{4!}\epsilon^{1jklmn}H_{1jk}H_{lmn}$$ which upon inspection makes evident that the last two terms in (\[pstact\]) contribute only a linear term in $h(\xi)$, coming from the pullback of $C_{(3)}$. This linear term cancels a similar linear contribution from the first term in (\[pstact\]), and we conclude that all the quadratic fluctuations come from the square root in (\[pstact\]). While carrying out the expansion is quite tedious, the final result is reassuringly simple. The Lagrangian that controls the fluctuation of the transverse position is $${\cal L}_{fluc}=-\frac{1}{2}\left(\rho^3\frac{R_{eff}^2}{R^2}\sqrt{|g_{S^3}|}\right)\frac{\rho^4}{R^2R^2_{eff}} \tilde G^{ij}\partial _i h \partial _j h \label{fluclag}$$ The quantity in parentheses is the determinant of the square-root part of the action (\[pstact\]), evaluated at the solution (\[solprobe\]). The metric $\tilde G_{ij}$ is the following $$\tilde G_{ij}d\xi^id\xi^j=\frac{\rho^4}{R^4}dx^2_{1,1}+\frac{R^2_{eff}}{\rho^2}d\rho^2 +R_{eff}^2d\Omega_3^2 \label{openmet}$$ so it is $AdS_3(R_{eff}/2)\times S^3(R_{eff})$. This is the same metric that was found in [@Arean:2007nh] to control the fluctuations of the directions transverse to $S^3$ in $S^7$. As advertised, for $q\neq 0$ it differs from the world-volume induced metric (\[wvmetric\]), and indeed it resembles the open string metric that controls the kinetic term of the fluctuations of world-volume fields of D-branes. From the fluctuation Lagrangian (\[fluclag\]), we obtain the equation of motion for $h(\xi)$, and since we are requiring $h(\xi)$ to be independent of the $S^3$ coordinates, this equation of motion boils down to $$\partial_\rho(\rho^9\partial_\rho h)+R^4R_{eff}^2\rho^3 (-\partial_0^2+\partial_1^2)h=0$$ This equation is immediate to solve: we separate variables and look for a solution of the form $h(x,\rho)=e^{ikx}f(\rho)$. The resulting equation for $f(\rho)$ has two solutions; to pick the relevant one for our problem, we consider their behavior near the boundary, as $\rho \rightarrow \infty$. It turns out that one solution decays as $1/\rho^8$, while the other one becomes independent of $\rho$ as $\rho \rightarrow \infty$, becoming a plane wave at the boundary. This second solution captures variations of the shape of the boundary defect, so it is the one we pick. Explicitly, $$h(x,\rho)=C e^{ikx}\frac{k^2}{\rho^4}Y_2\left(\frac{R^2R_e|k|}{2\rho^2}\right) \label{flucsol}$$ To make contact with the discussion in the previous section, and the work of [@Garriga:2009hy], note that the fluctuation $h(\xi)$ is not a proper displacement, as seen by an observer in the unperturbed probe brane. To relate it to a proper distance, we compute the unit vector $n^\mu$ normal to the M5-brane profile (\[solprobe\]) $$(n^\rho,n^2)=\left(\frac{q\rho}{R^3R_3}, \frac{R^3}{\rho^2 R_e} \right)$$ so we can introduce a scalar $\phi(\xi)$ measuring proper displacement by defining $h=\phi n^2$ [@Garriga:1991ts]. For this scalar field, the equation of motion is $$\left(\Box_{AdS_3}-\frac{3}{R^2_{AdS_3}}\right)\phi=0$$ Notice that now this is a massive scalar field (as the fluctuation now has to climb the AdS well), while for the bubble in deSitter space, the scalar was tachyonic, see eq. (\[tacmass\]). The regularized action ---------------------- Having found the solution for the relevant fluctuation, our next task is to evaluate the action at this solution. As it is by now familiar, an straightforward evaluation yields a divergent result, so we regulate it by introducing a cut-off on $\rho$, $\rho_c$. The regulated action contains pieces analytic in $\rho_c$ that diverge in the limit $\rho\rightarrow \infty$, and a term logarithmic in $\rho_c$, which induces the anomaly in the finite piece. It is the coefficient of this $\log \rho_c$ we are after, since as we will show, it is $d_1/2$. The regularized action for the fluctuations is $$W[\rho_c]=-T_{M5}\int d^6 \xi \left(\frac{1}{2}\sqrt{g_{S^3}}\frac{\rho^7}{R^4}\tilde G^{ij}\partial _i h\partial _j h\right)$$ Since we are considering a fluctuation mode $h(\xi)$ independent of $S^3$ coordinates, the integration over $S^3$ is immediate, yielding $$W[\rho_c]=-\frac{2\pi^2}{2R^4}T_{M5}\int d^2x \int _0^{\rho_c} d\rho \rho^7 \tilde G^{ij}\partial _i h\partial _j h$$ Using the explicit form of the metric $\tilde G$, eq. (\[openmet\]), integration by parts, and the equations of motion, we obtain $$W[\rho_c]=-\frac{1}{2}\frac{2\pi^2\rho_c^9}{R^4R^2_{eff}}T_{M5}\int d^2x \; h(\rho_c)\partial _\rho h|_{\rho_c}$$ We now Fourier transform, evaluate the action with the solution (\[flucsol\]) and obtain $$W=-\frac{1}{2}\frac{2\pi^2}{R^4R_e^2}T_{M5}\int d^2k |h(k,\rho_c)|^2 \left[ -R^4R^2_{eff}\rho_c^4+\frac{(R^2R_{eff}|k|)^4}{2}\left(\gamma_E+\hbox{log}\frac{R^2R_e|k|^4}{2\rho^2_c}\right)+{\cal O}(1/\rho_c^2)\right]$$ The first term is analytic in $\rho_c$, and will be taken care of by a corresponding counterterm. The term relevant for us is the one with log $\rho_c^2$ since its coefficient is essentially $d_1$. To establish the precise relation, in the boundary we consider a fluctuation of the shape of the defect, as a plane wave $h(x)=Ae^{ikx}$. One easily sees that for small fluctuations $$K_{ij}K^{ij}-\frac{1}{2}K^2\simeq \frac{1}{2}k^4 h^2$$ Recalling the definition of the trace anomaly (\[confan\]), we arrive at $$\frac{d_1}{2} =\frac{\pi^2}{2}R^4R_{eff}^2T_{M5} \label{resultdd}$$ To translate this result into CFT variables, we use $$T_{M5}=\frac{1}{(2\pi)^5 \ell_P^6}\hspace{1cm} \frac{R}{\ell_P}=(32\pi^2 N)^{1/6}$$ where $\ell_P$ is the eleven-dimensional Planck length, and the quantization condition for the world-volume magnetic flux [@Camino:2001at] $$\frac{q}{\ell_P^3}=4\pi k$$ where $k$ is the number of M2-branes on one side of the M5 that recombine with it. We finally conclude that $$d_1=\frac{1}{\pi}\left(N+\frac{1}{2}k^2\right) \label{dincft}$$ Let’s pause to contemplate this result. As advertised, for fixed $N$, the minimum value of $d_1$ corresponds to the case when the vacuum density energy is the same on both sides of the defect (i.e. when $k=0$), which is the kind of behavior one would expect, in clear contrast to what happens in de Sitter, see the discussion below eq. (\[resultd\]). The key point to understand this difference is the relative plus sign in (\[dincft\]), which in the supergravity side is the relative plus sign in (\[radeff\]). We can also check if the inequality (\[puzzle\]) is always parametrically satisfied. To do so, we need the jump in the number of degrees of freedom between the two sides of the wall. In the large N limit, it is possible to compute its scaling from supergravity [@Klebanov:1996un], and one obtains $$c\sim N^{3/2} \Rightarrow \Delta c \sim (N+k)^{3/2}-N^{3/2}\sim N^{1/2}k$$ where in the last step we assumed that $k\ll N$. Comparing this with (\[dincft\]) we see that now the inequality is satisfied, with the particular parametric dependence $k\sim N^{1/2}$ being the only case when both quantities become comparable. Discussion ========== The main goal of this work was to continue the study of possible implications of the spectrum of bubble fluctuations for a potential holographic dual of an eternally inflating multiverse. The first results [@Garriga:2009hy] in this direction didn’t conform with intuitive expectations, since for instance, they appear to imply that as the difference in vacuum energies on both sides of the bubble grows, the number of degrees of freedom that ’feel’ the bubble decreases [^2]. In this work we have considered a similar setup, but with both de Sitter geometries (bulk and probe world-volume) traded with Anti de Sitter geometries. In this case we see that the coefficient $d_1$ behaves as expected. At the root of the difference lies the fact that the $dS_3$ world-volume in $dS_4$ has curvature radius smaller or equal to that of $dS_4$, while for the $AdS_3$ world-volume in $AdS_4$ just the opposite is true. Therefore, as long as the coefficient $d_1$ is proportional to this thin wall curvature radius, it behaves conforming the expectations of [@Garriga:2009hy] for $AdS$, but not for $dS$. If the hypothetical Euclidean QFT holographically dual to the multiverse does exist, a 2d defect should give a formula similar to (\[dincft\]), but with a minus relative sign, so in this case $d_1$ attains its maximum value when both sides of the wall have the same vacuum energy, as required by the arguments after eq. (\[resultd\]). Where could this relative minus sign come from, in the hypothetical dual theory? To reconcile the intuition that more ambient fields should feel the defect (or more fields live on the defect) as $\Delta \rho_V$ increases with the result that $d_1$ decreases, we would need degrees of freedom that contributely [*negatively*]{} to $d_1$. This suggests that these defect degrees of freedom might be non-unitary; we certainly know of non-unitary 2d CFTs with $c<0$, and while $d_1$ is not quite the $2d$ central charge (as it is the coefficient that accompanies the term with extrinsic rather than intrinsic curvature), this suggestion seems worth exploring. Put differently, the authors of [@Garriga:2009hy] speculated that a way out of their puzzles would be that the evaluation of $d_1$ involved unexpected cancellations on the field theory side. A simple example that might illustrate such cancellations is the worldsheet CFT in string theory, which has total vanishing central charge, due to the sum of two theories, one with $c>0$ and one with $c<0$[^3] . At a technical level, the main result in this paper is the computation of the coefficient $d_1$ in the conformal anomaly for a surface operator in a 2+1 CFT, following the prescription of [@Garriga:2009hy]. Surface operators in 3+1 dimensional theories have received a lot of attention, but there are very few explicit computations of their parameters [@Drukker:2008wr]. The kind of computation performed here might find applications in the study of surface operators, beyond our original motivation. Acknowledgements ================ This work grew out of many conversations with Jaume Garriga, and I would like to thank him for many explanations about [@Garriga:2008ks; @Garriga:2009hy]. I would also like to thank Ofer Aharony and Micha Berkooz for insightful discussions and Alfonso Ramallo for useful correspondence regarding [@Arean:2007nh]. I would like to thank the Department of Particle Physics at the Weizmann Institute of Science for hospitality during the last stages of this project. This research is supported by a Ramón y Cajal fellowship, and also by MEC FPA2007-66665C02-02, CPAN CSD2007-00042, within the Consolider-Ingenio2010 program, and AGAUR 2009SGR00168. [20]{} G. ’t Hooft, “Dimensional reduction in quantum gravity,” arXiv:gr-qc/9310026. L. Susskind, “The World As A Hologram,” J. Math. Phys.  [**36**]{}, 6377 (1995) \[arXiv:hep-th/9409089\]. J. M. Maldacena, “The large N limit of superconformal field theories and supergravity,” Adv. Theor. Math. Phys.  [**2**]{}, 231 (1998) \[Int. J. Theor. Phys.  [**38**]{}, 1113 (1999)\] \[arXiv:hep-th/9711200\]. A. Strominger, “Inflation and the dS/CFT correspondence,” JHEP [**0111**]{}, 049 (2001) \[arXiv:hep-th/0110087\]. B. Freivogel, Y. Sekino, L. Susskind and C. P. Yeh, “A holographic framework for eternal inflation,” Phys. Rev.  D [**74**]{}, 086003 (2006) \[arXiv:hep-th/0606204\]. B. Freivogel and M. Kleban, “A Conformal Field Theory for Eternal Inflation,” JHEP [**0912**]{}, 019 (2009) \[arXiv:0903.2048 \[hep-th\]\]. R. Bousso, “Complementarity in the Multiverse,” Phys. Rev.  D [**79**]{}, 123524 (2009) \[arXiv:0901.4806 \[hep-th\]\]. J. Garriga and A. Vilenkin, “Holographic Multiverse,” JCAP [**0901**]{}, 021 (2009) \[arXiv:0809.4257 \[hep-th\]\]. J. Garriga and A. Vilenkin, “Holographic multiverse and conformal invariance,” JCAP [**0911**]{}, 020 (2009) \[arXiv:0905.1509 \[hep-th\]\]. J. L. Cardy, “Conformal Invariance And Surface Critical Behavior,” Nucl. Phys.  B [**240**]{} (1984) 514. D. M. McAvity and H. Osborn, “Conformal Field Theories Near A Boundary In General Dimensions,” Nucl. Phys.  B [**455**]{}, 522 (1995) \[arXiv:cond-mat/9505127\]. J. Erdmenger, Z. Guralnik and I. Kirsch, “Four-Dimensional Superconformal Theories with Interacting Boundaries or Defects,” Phys. Rev.  D [**66**]{}, 025020 (2002) \[arXiv:hep-th/0203020\]. A. Karch and L. Randall, “Open and closed string interpretation of SUSY CFT’s on branes with boundaries,” JHEP [**0106**]{}, 063 (2001) \[arXiv:hep-th/0105132\]. O. DeWolfe, D. Z. Freedman and H. Ooguri, “Holography and defect conformal field theories,” Phys. Rev.  D [**66**]{}, 025009 (2002) \[arXiv:hep-th/0111135\]. O. Aharony, O. DeWolfe, D. Z. Freedman and A. Karch, “Defect conformal field theory and locally localized gravity,” JHEP [**0307**]{}, 030 (2003) \[arXiv:hep-th/0303249\]. N. R. Constable, J. Erdmenger, Z. Guralnik and I. Kirsch, “Intersecting D3-branes and holography,” Phys. Rev.  D [**68**]{}, 106007 (2003) \[arXiv:hep-th/0211222\]. P. McFadden and K. Skenderis, “The Holographic Universe,” arXiv:1001.2007 \[hep-th\]. P. McFadden and K. Skenderis, “Holography for Cosmology,” Phys. Rev.  D [**81**]{}, 021301 (2010) \[arXiv:0907.5542 \[hep-th\]\]. A. Strominger, “The dS/CFT correspondence,” JHEP [**0110**]{}, 034 (2001) \[arXiv:hep-th/0106113\]. J. Garriga and A. Vilenkin, “Perturbations on domain walls and strings: A Covariant theory,” Phys. Rev.  D [**44**]{}, 1007 (1991). J. Garriga and A. Vilenkin, “Quantum fluctuations on domain walls, strings and vacuum bubbles,” Phys. Rev.  D [**45**]{}, 3469 (1992). V. A. Berezin, V. A. Kuzmin and I. I. Tkachev, “Dynamics of Bubbles in General Relativity,” Phys. Rev.  D [**36**]{}, 2919 (1987). J. Garriga, “Nucleation rates in flat and curved space,” Phys. Rev.  D [**49**]{}, 6327 (1994) \[arXiv:hep-ph/9308280\]. J. D. Brown and M. Henneaux, “Central Charges in the Canonical Realization of Asymptotic Symmetries: An Example from Three-Dimensional Gravity,” Commun. Math. Phys.  [**104**]{}, 207 (1986). G. Kennedy, R. Critchley and J. S. Dowker, “Finite Temperature Field Theory With Boundaries: Stress Tensor And Surface Action Renormalization,” Annals Phys.  [**125**]{}, 346 (1980). A. Schwimmer and S. Theisen, “Entanglement Entropy, Trace Anomalies and Holography,” Nucl. Phys.  B [**801**]{}, 1 (2008) \[arXiv:0802.1017 \[hep-th\]\]. M. Henningson and K. Skenderis, “Holography and the Weyl anomaly,” Fortsch. Phys.  [**48**]{}, 125 (2000) \[arXiv:hep-th/9812032\]. M. Henningson and K. Skenderis, “The holographic Weyl anomaly,” JHEP [**9807**]{}, 023 (1998) \[arXiv:hep-th/9806087\]. D. E. Berenstein, R. Corrado, W. Fischler and J. M. Maldacena, “The operator product expansion for Wilson loops and surfaces in the large N limit,” Phys. Rev.  D [**59**]{}, 105023 (1999) \[arXiv:hep-th/9809188\]. C. R. Graham and E. Witten, “Conformal anomaly of submanifold observables in AdS/CFT correspondence,” Nucl. Phys.  B [**546**]{}, 52 (1999) \[arXiv:hep-th/9901021\]. O. Aharony, O. Bergman, D. L. Jafferis and J. Maldacena, “N=6 superconformal Chern-Simons-matter theories, M2-branes and their gravity duals,” JHEP [**0810**]{}, 091 (2008) \[arXiv:0806.1218 \[hep-th\]\]. M. Ammon, J. Erdmenger, R. Meyer, A. O’Bannon and T. Wrase, “Adding Flavor to AdS4/CFT3,” JHEP [**0911**]{}, 125 (2009) \[arXiv:0909.3845 \[hep-th\]\]. J. Gomis and F. Passerini, “Holographic Wilson loops,” JHEP [**0608**]{}, 074 (2006) \[arXiv:hep-th/0604007\]. P. Pasti, D. P. Sorokin and M. Tonin, “Covariant action for a D = 11 five-brane with the chiral field,” Phys. Lett.  B [**398**]{}, 41 (1997) \[arXiv:hep-th/9701037\]. D. Arean, A. V. Ramallo and D. Rodriguez-Gomez, “Holographic flavor on the Higgs branch,” JHEP [**0705**]{}, 044 (2007) \[arXiv:hep-th/0703094\]. G. W. Gibbons and P. C. West, “The metric and strong coupling limit of the M5-brane,” J. Math. Phys.  [**42**]{}, 3188 (2001) \[arXiv:hep-th/0011149\]. B. Fiol, “The spectrum of chiral operators of 2+1 dCFTs from M5-branes”, [*to appear*]{}. J. M. Camino, A. Paredes and A. V. Ramallo, “Stable wrapped branes,” JHEP [**0105**]{}, 011 (2001) \[arXiv:hep-th/0104082\]. I. R. Klebanov and A. A. Tseytlin, “Entropy of Near-Extremal Black p-branes,” Nucl. Phys.  B [**475**]{}, 164 (1996) \[arXiv:hep-th/9604089\]. T. Banks, B. Fiol and A. Morisse, “Towards a quantum theory of de Sitter space,” JHEP [**0612**]{}, 004 (2006) \[arXiv:hep-th/0609062\]. M. Henningson and K. Skenderis, “Weyl anomaly for Wilson surfaces,” JHEP [**9906**]{}, 012 (1999) \[arXiv:hep-th/9905163\]. N. Drukker, J. Gomis and S. Matsuura, “Probing N=4 SYM With Surface Operators,” JHEP [**0810**]{}, 048 (2008) \[arXiv:0805.4199 \[hep-th\]\]. [^1]: On the other hand, 3d super-renormalizable theories might be candidates for a holographic description of inflation in a regime where gravity is strongly coupled [@McFadden:2010na]. [^2]: This has a vague resemblance with another counterintuitive feature of deSitter space: for fixed cosmological constant, empty de Sitter has the maximal entropy possible, and adding a black hole decreases rather than increases the total entropy. See [@Banks:2006rx] for some ideas on the possible holographic realization of this feature. [^3]: I would like to thank Micha Berkooz for suggesting this example, and for related conversations.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Each topological group $G$ admits a unique universal minimal dynamical system $(M(G),G)$. When $G$ is a non-compact locally compact group the phase space $M(G)$ of this universal system is non-metrizable. There are however topological groups for which $M(G)$ is the trivial one point system (extremely amenable groups), as well as topological groups $G$ for which $M(G)$ is a metrizable space and for which there is an explicit description of the dynamical system $(M(G),G)$. One such group is the topological group $S_\infty$ of all permutations of the integers ${\mathbb Z}$, with the topology of pointwise convergence. We show that $(M(S_\infty),S_\infty)$ is a symbolic dynamical system (hence in particular $M(S_\infty)$ is a Cantor set), and give a full description of all its symbolic factors. Among other facts we show that $(M(G),G)$ (and hence also every minimal $S_\infty$) has the structure of a two-to-one group extension of proximal system and that it is uniquely ergodic.' address: | Department of Mathematics\ Tel Aviv University\ Ramat Aviv\ Israel author: - Eli Glasner title: 'The Cantor set of linear orders on ${\mathbb N}$ is the universal minimal $S_\infty$-system' --- (474,66)(0,0) (0,66)(1,0)[40]{}[(0,-1)[24]{}]{} (43,65)(1,-1)[24]{}[(0,-1)[40]{}]{} (1,39)(1,-1)[40]{}[(1,0)[24]{}]{} (70,2)(1,1)[24]{}[(0,1)[40]{}]{} (72,0)(1,1)[24]{}[(1,0)[40]{}]{} (97,66)(1,0)[40]{}[(0,-1)[40]{}]{} (143,66)[(0,0)\[tl\][Proceedings of the Ninth Prague Topological Symposium]{}]{} (143,50)[(0,0)\[tl\][Contributed papers from the symposium held in]{}]{} (143,34)[(0,0)\[tl\][Prague, Czech Republic, August 19–25, 2001]{}]{} [^1] [^2] [^3] This is a summary of a talk given at the Prague Topological Symposium of 2001 in which I described results obtained in a joint paper with B. Weiss. The paper is going to appear soon in GAFA [@GW]. Given a topological group $G$ and a compact Hausdorff space $X$, a dynamical system $(X,G)$ is a jointly continuous action of $G$ on $X$. If $(Y,G)$ is a second dynamical system then a continuous onto map $\pi:(X,G)\to (Y,G)$ which intertwines the $G$ actions is called [*a homomorphism*]{}. The dynamical system $(X,G)$ is [*point transitive*]{} if there exists a point $x_0\in X$ whose [*orbit*]{} $Gx_0$ is dense in $X$. $(X,G)$ is [*minimal*]{} if every orbit is dense. It can be easily shown that there exists a unique (up to isomorphism of dynamical $G$-systems) universal point transitive $G$-system $(\mathbf{L},G)$. One way of presenting this universal object is via the Gelfand space of the $C^*$-algebra $\mathcal{L}_l(G)$ of left uniformly $\mathbb{C}$-valued continuous functions on $G$. From the existence of $(\mathbf{L},G)$ one easily deduces the existence of a universal minimal dynamical system; i.e. a system $(M(G),G)$ such that for every minimal system $(X,G)$ there exists a homomorphism $\pi:(M(G),G)\to (X,G)$. Ellis’ theory shows that up to isomorphism this universal minimal dynamical system is unique, see e.g. [@E]. The existence of uncountably many characters of the discrete group ${{\mathbb Z}}$ already shows that the phase space $M({{\mathbb Z}})$ is non-metrizable. In fact one can show that $M(G)$ is non-metrizable whenever $G$ is non-compact locally compact group. A topological group $G$ has the [*fixed point on compacta property*]{} (f.p.c.) (or is [*extremely amenable*]{}) if whenever it acts continuously on a compact space, it has a fixed point. Thus the group $G$ has the f.p.c. property iff its universal minimal dynamical system is the trivial one point system. A triple $(X,d,\mu)$, where $(X,d)$ is a metric space and $\mu$ a probability measure on $X$, is called an $mm$[*-space*]{}. For $A\subseteq X$, $\mu(A)\geq 1/2$, and $\epsilon>0$ let $A_\epsilon$ be the set of all points whose distance from $A$ is at most $\epsilon$. A family of $mm$ spaces $(X_n,d_n,\mu_n)$ is called a [*Lévy family*]{} if for every $\epsilon$, $\alpha_n(\epsilon)\to 0$, where $\alpha(\epsilon)=1-\inf\{\mu(A_\epsilon) : A\subseteq X,\ \mu(A)\geq 1/2\}$. When a Polish group $(G,d)$ contains an increasing sequence of compact subgroups $\{G_n:n\in {{\mathbb N}}\}$ whose union is dense in $G$ and such that with respect to the corresponding sequence of Haar measures $\mu_n$, the family $(G_n,d,\mu_n)$ forms a Lévy family, then $G$ is called a [*Lévy group*]{}. In [@GM] Gromov and Milman prove that every Lévy group $G$ has the f.p.c. property. Many of the examples presently known of extremely amenable groups are obtained via this theorem. There are however other methods of obtaining such groups. Here is a partial list: 1. The unitary group $U(\infty)=\cup_{n=1}^\infty U(n)$ with the uniform operator topology (Gromov-Milman, [@GM]). 2. The monothetic Polish group $L_m(I,S^1)$, consisting of all (classes) of measurable maps from the unit interval $I$ into the circle group $S^1$ with the topology of convergence in measure induced by, say, Lebesgue measure on $I$ (Glasner, [@G]; Furstenberg-Weiss). More generally, $L_m(I,G)$, where $G$ is any locally compact amenable group (Pestov, [@P4]). 3. The group of measurable automorphisms $\operatorname{Aut}(X,\mu)$ of a standard sigma-finite measure space $(X,\mu)$, with respect to the weak topology (Giordano-Pestov [@GP]). 4. Using Ramsey’s theorem, Pestov has shown that the group $\operatorname{Aut}({{{\mathbb Q}},<})$, of order automorphism of the rational numbers with pointwise convergence topology, is extremely amenable, [@P2]. Thus, as we have seen, the universal minimal system $(M(G),G)$ corresponding to a non-compact $G$ is usually non-metrizable but can be, in some cases, trivial. Are there non-compact topological groups for which $M(G)$ is metrizable but non-trivial? The first such example was pointed out by Pestov [@P2] who used claim 4 above to show that the universal minimal dynamical system of the group $G$ of orientation-preserving homeomorphisms of the circle coincides with the natural action of $G$ on $S^1$. In [@U] V. Uspenskij shows that the action of a topological group $G$ on its universal minimal system $M(G)$ is never $3$-transitive. As a direct corollary he shows that for manifolds $X$ of dimension $>1$ (as well as for $X=Q$, the Hilbert cube) the corresponding group $G$ of orientation preserving homeomorphisms, $(M(G),G)$ does not coincide with the natural action of $G$ on $X$. Let $S_\infty$ be the group of all permutations of the integers ${{\mathbb Z}}$. With respect to the topology of pointwise convergence on ${{\mathbb Z}}$, $S_\infty$ is a Polish topological group. The subgroup $S_0\subset S_\infty$ consisting of the permutations which fix all but a finite set in ${{\mathbb Z}}$ is an amenable dense subgroup (being the union of an increasing sequence of finite groups) and therefore $S_\infty$ is amenable as well. In [@GM] Gromov and Milman conjectured, in view of the concentration of measure on $S_n$ with respect to Hamming distance, that $S_\infty$ has the f.p.c. property. In [@P2] and [@P3] V. Pestov has shown that, on the contrary, $S_\infty$ acts effectively on $M(S_\infty)$ and that, in fact, there is no Hausdorff topology making $S_0$ a topological group with the f.p.c. property. He as well as A. Kechris (in private communication) asked for explicit examples of $S_\infty$-minimal systems. The main result of our work [@GW] is the fact that the universal minimal system $(M(S_\infty),S_\infty)$ is a metrizable system, in fact a system whose phase space is the Cantor set. We also give in this work an explicit description of $(M(S_\infty),S_\infty)$ as a “symbolic" dynamical system and exhibit explicit formulas for all of its symbolic factors. Let me now describe these results in more details. For every integer $k\ge 2$ let $${{\mathbb Z}}^k_*=\{(i_1,i_2,\dots,i_k)\in {{\mathbb Z}}^k: i_1,i_2,\dots,i_k\ \text{ are distinct elements of ${{\mathbb Z}}$}\},$$ and set $\Omega^k=\{1,-1\}^{{{\mathbb Z}}^k_*}$. Consider the dynamical system $(\Omega^k,S_\infty)$, where for $\alpha\in S_\infty$ and $\omega\in \Omega^k$ we let $$(\alpha\omega)(i_1,i_2,\dots,i_k)= \omega(\alpha^{-1}i_1,\alpha^{-1}i_2,\dots,\alpha^{-1}i_k).$$ Let $\Omega^k_{alt}\subset\Omega^k$ consist of all the [*alternating*]{} configurations, that is those elements $\omega\in\Omega^k$ satisfying $$\omega(\sigma(i_1),\sigma(i_2),\dots,\sigma(i_k)) = \operatorname{sgn}(\sigma)\omega(i_1,i_2,\dots,i_k),$$ for all $\sigma\in S_k$ and $(i_1,i_2,\dots,i_k)\in {{\mathbb Z}}^k_*$. Clearly $\Omega^k_{alt}$is a closed and $S_\infty$-invariant subset of $\Omega^k$. A configuration $\omega\in\Omega^2$ [*determines a linear order*]{} on ${{\mathbb Z}}$ if it is alternating, and satisfies the conditions: $$\omega(m,n)=1\ \wedge\ \omega(n,l)=1 \quad \Rightarrow \quad \omega(m,l)=1.$$ Let $<_\omega$ be the corresponding linear order on ${{\mathbb Z}}$, where $m<_\omega n$ iff $\omega(m,n)=1$. Let $X=\Omega^2_{lo}$ be the subset of $\Omega^2$ consisting of all the configurations which determine a linear order. The correspondence $\omega \longleftrightarrow\ <_\omega$ is a surjective bijection between $\Omega^2_{lo}$ and the collection of linear orders on ${{\mathbb Z}}$. Clearly $X$ is a closed $S_\infty$-invariant set and using Ramsey’s theorem we shall show that $(X,S_\infty)$ is a minimal system. Say that a configuration $\omega\in \Omega^3$ is [*determined by a circular order*]{} if there exists a sequence $\{z_m: m\in {{\mathbb Z}}\} \subset S^1$ with $m\ne n \ \Rightarrow\ z_m\ne z_n$ such that: $\omega(l,m,n)=1$ for $(l,m,n)\in {{\mathbb Z}}^3_*$ iff the directed arc in $S^1$ defined by the ordered triple $(z_l,z_m,z_n)$ is oriented in the positive direction. Let $Y=\Omega^3_c\subset \Omega^3_{alt}$ denote the collection of all the configurations in $\Omega^3$ which are determined by a circular order. It follows that the set $Y=\Omega^3_c$ is closed and invariant and using Ramsey’s theorem one can show that it is minimal. If we go now to $\Omega^4_{alt}$, can one find a sequence of points $\{z_n\}$ on the sphere $S^2$ in general position such that the tetrahedron defined by any four points $z_{n_1},z_{n_2},z_{n_3},z_{n_4}$ has positive orientation when $n_1< n_2 < n_3 < n_4$? Starting with any sequence $\{z_n\}\subset S^2$ in general position one can use Ramsey’s theorem to find a subsequence with the required property. Another way to see this is to use the ‘moment curve’ $$t\mapsto (t,t^2,t^3).$$ Again it turns out that the orbit closure in $\Omega^4_{alt}$ which is determined by such a sequence forms a minimal dynamical system. It now seems as if going up to $\Omega^k_{alt}$ with larger and larger $k$’s we encounter more and more complicated minimal systems. However, as we show in [@GW], this is not the case and the entire story is already encoded in the simplest symbolic dynamical system $\Omega^2_{lo}$. $\Omega^2_{lo}$ is the universal minimal $S_\infty$-system. The fact that the topology on $S_\infty$ is zero-dimensional, and in fact given by a sequence of clopen subgroups, enables us to reduce this theorem to the following one. Every minimal subsystem $\Sigma$ of the system $(\Omega^k,S_\infty)$ is a factor of the minimal system $(\Omega^2_{lo},S_\infty)$. Finally let me mention two more facts concerning the system $(M(S_\infty),S_\infty)$. The universal minimal system $(\Omega^2_{lo},S_\infty)$ has the structure of a two-to-one group extension of a proximal system. The universal minimal system $(\Omega^2_{lo},S_\infty)$ is uniquely ergodic and therefore so is every minimal $S_\infty$-system. [1]{} Robert Ellis, *Lectures on topological dynamics*, W. A. Benjamin, Inc., New York, 1969. [MR ]{}[42 \#2463]{} Thierry. Giordano and Vladimir G. Pestov, *Some extremely amenable groups*, To appear in C. R. Acad. Sci. Paris Sér. I Math. arXiv:math.GR/0109138 [http://arxiv.org/abs/math.GR/0109138]{}, 2002. Eli Glasner, *On minimal actions of [P]{}olish groups*, Topology Appl. **85** (1998), no. 1-3, 119–125, 8th Prague Topological Symposium on General Topology and Its Relations to Modern Analysis and Algebra (1996). [MR ]{}[99c:54057]{} Eli Glasner and Benjamin Weiss, *Minimal actions of the group $\mathbb{S}(\mathbb{Z})$ of permutations of the integers*, To appear in Geom. Funct. Anal. (GAFA), 2002. M. Gromov and V. D. Milman, *A topological application of the isoperimetric inequality*, Amer. J. Math. **105** (1983), no. 4, 843–854. [MR ]{}[84k:28012]{} Vladimir G. Pestov, *On free actions, minimal flows, and a problem by [E]{}llis*, Trans. Amer. Math. Soc. **350** (1998), no. 10, 4149–4165. [MR ]{}[99b:54069]{} [to3em]{}, *Amenable representations and dynamics of the unit sphere in an infinite-dimensional [H]{}ilbert space*, Geom. Funct. Anal. **10** (2000), no. 5, 1171–1201. [MR ]{}[2001m:22012]{} [to3em]{}, *Ramsey-milman phenomenon, urysohn metric spaces, and extremely amenable groups*, Israel J. Math. **127** (2002), 317–358, arXiv:math.FA/0004010 [http://www.arxiv.org/abs/math.FA/0004010]{}. V. V. Uspenskij, *On universal minimal compact $g$-spaces*, Topology Proceedings **25** (2000), no. Spring, 301–308, arXiv:math.GN/0006081 [http://at.yorku.ca/b/a/a/k/56.htm]{}. [^1]: The author was an invited speaker at the Ninth Prague Topological Symposium. [^2]: This is a summary article. The results in this article will be treated fully in an article, written jointly with B. Weiss, to be published in Geometric and Functional Analysis (GAFA) [^3]: Eli Glasner, [*The Cantor set of linear orders on ${\mathbb N}$ is the universal minimal $S_\infty$-system*]{}, Proceedings of the Ninth Prague Topological Symposium, (Prague, 2001), pp. 119–123, Topology Atlas, Toronto, 2002
{ "pile_set_name": "ArXiv" }
--- abstract: 'Low decoding latency and complexity are two important requirements of channel codes used in many applications, like machine-to-machine communications. In this paper, we show how these requirements can be fulfilled by using some special quasi-cyclic low-density parity-check block codes and spatially coupled low-density parity-check convolutional codes that we denote as *compact*. They are defined by parity-check matrices designed according to a recent approach based on sequentially multiplied columns. This method allows obtaining codes with girth up to 12. Many numerical examples of practical codes are provided.' author: - title: 'Compact QC-LDPC Block and SC-LDPC Convolutional Codes for Low-Latency Communications' --- Introduction ============ Fast and reliable transmissions of short packets are a prerequisite for many modern applications, like machine-to-machine (M2M) communications [@Durisi2016]. Channel coding is commonly used for transmission reliability; however, constrained resources enforce the use of codes that can be encoded and decoded with low complexity [@BattaglioniPao]. At the same time, achieving low latency is crucial in this kind of applications, due to their real-time requirements. Quasi-cyclic low-density parity-check (QC-LDPC) block codes fit well this scenario, as they can be encoded and decoded with low-complexity, hardware-oriented techniques [@Li1; @Lan2007]. Iterative algorithms used for their decoding are adversely affected by the presence of short cycles in their associated Tanner graphs. Therefore, the minimum cycle length, also known as *girth* (and denoted by $g$ afterwards), should be kept as large as possible. QC-LDPC block codes are also the basis for the design of spatially coupled low-density parity-check convolutional codes (SC-LDPC-CCs). Due to their infinite length, SC-LDPC-CCs may seem unsuitable for resource-constrained contexts. However, sliding window (SW) decoding [@Felt1; @Lentmaier2005] of SC-LDPC-CCs with short constraint length can be performed over short windows, thus resulting in very good performance, often better than that of their block code counterparts. Motivated by these arguments, in this paper we provide design examples and assess the performance of either QC-LDPC block codes with smaller blocklength or SC-LDPC-CCs with smaller constraint length than those with comparable girth available in the literature. These codes are designed according to the approach we have recently introduced in [@Tadayon2018], and are denoted as *compact codes* to encompass block and convolutional LDPC codes with these features in one word. The method in [@Tadayon2018] is based on sequentially multiplied columns (SMCs) and, to the best of our knowledge, produces the most compact QC-LDPC block codes and SC-LDPC-CCs with $g=10, 12$ currently available in the literature. However, in [@Tadayon2018] a single design example is proposed. Here we design several codes with different rates and girth, and generalize the approach to the cases of $g=6,8$. Moreover, we relate the blocklength (for QC-LDPC block codes) and the constraint length (for SC-LDPC-CCs) of these codes to the latency and complexity of the decoding algorithms. Many theoretical lower bounds on the blocklength (constraint length) of QC-LDPC block codes (SC-LDPC-CCs) for several values of the girth have been proposed (see, for example, [@Hagi2009; @ATasdighi1; @Karimi2013; @Amirzadeh1] for QC-LDPC block codes and [@MBAT2017] for SC-LDPC-CCs). Compared with numerical results, these bounds are tight when $g=6,8$, but provide a loose indication when $g=10,12$. For this reason, we focus on the latter cases, quantifying the improvement achieved by the newly designed codes over previous solutions. The remainder of the paper is organized as follows. In Section \[Sec2\] we briefly remind the basic notions concerning QC-LDPC block codes and SC-LDPC-CCs. In Section \[Sec3\] we recall the SMC assumption. In Section \[Sec4\] we discuss the latency and complexity of the considered decoding algorithms. Section \[Sec5\] provides numerical results. Finally, Section \[Sec6\] concludes the paper. Notation {#Sec2} ======== CPM-based QC-LDPC block codes ----------------------------- We consider a special class of QC-LDPC block codes defined through a parity-check matrix formed by $m \times n$ circulant permutation matrices (CPMs) with size $N \times N$, where $N$ is known as the *lifting degree* of the code. Each CPM is denoted as $\mathbf{I}(p_{ij})$, $0 \leq i \leq m - 1$, $0 \leq j \leq n - 1$, and is obtained by cyclically shifting all the rows of the identity matrix by $p_{ij}$ positions, with $0 \leq p_{ij} \leq N - 1$. The code length is $L=nN$. QC-LDPC block codes can be equivalently represented through their [*exponent matrix*]{} $\mathbf{P}$, whose entries are the integer values $p_{ij}$. It is shown in [@MFossorier1] that a necessary and sufficient condition for the existence of a cycle with length $2k$ in the Tanner graph of a QC-LDPC block code is $$\sum_{i=0}^{k-1} \left( p_{m_{i}n_{i}} - p_{m_{i}n_{i+1}} \right) = 0 \mod N , \label{fore}$$ where $n_{k}=n_{0}$, $m_{i} \neq m_{i+1}$, $n_{i} \neq n_{i+1}$. Based on , let us introduce [*avoidable*]{} and [*strictly avoidable*]{} cycles for CPM-based QC-LDPC block codes. The former occur when $\sum_{i=0}^{k-1} \left( p_{m_{i}n_{i}} - p_{m_{i}n_{i+1}} \right) = \beta N$, $\beta > 0$. For the latter instead we have $\sum_{i=0}^{k-1} \left( p_{m_{i}n_{i}} - p_{m_{i}n_{i+1}} \right)=0$. SC-LDPC-CCs ----------- Besides QC-LDPC block codes, we consider time-invariant SC-LDPC-CCs, which are defined through a semi-infinite parity-check matrix in the form $$\mathbf{H} = \left[\begin{array}{cccccc} \arraycolsep=1.4pt\def\arraystretch{5pt} \mathbf{H}_0 & \mathbf{0} & \mathbf{0} & \ddots \\ \mathbf{H}_1 & \mathbf{H}_0 & \mathbf{0} & \ddots \\ \vdots & \mathbf{H}_1 & \mathbf{H}_0 & \ddots \\ \mathbf{H}_{m_h} & \vdots & \mathbf{H}_1 & \ddots \\ \mathbf{0} & \mathbf{H}_{m_h} & \vdots & \ddots \\ \mathbf{0} & \mathbf{0} & \mathbf{H}_{m_h} & \ddots \\ \vdots & \vdots & \vdots & \ddots \\ \end{array}\right], \label{eq:Hconv}$$ where each block $\mathbf{H}_i$, $i = 0, 1, 2, \ldots, m_h$, is a binary matrix with size $c \times a$. The syndrome former matrix is $\mathbf{H_s} = \left[ \mathbf{H}_0^T | \mathbf{H}_1^T | \mathbf{H}_2^T | \ldots | \mathbf{H}_{m_h}^T \right]$, where $^T$ denotes transposition; its size is $a\times (m_h+1)c$. According to , the code has asymptotic rate $R = \frac{a-c}{a}$. The height of the non-zero diagonal band in instead gives the syndrome former memory order $m_h$, and the code syndrome former constraint length is defined as $v_s = (m_h + 1) a$. Link between QC-LDPC block codes and SC-LDPC-CCs ------------------------------------------------ A common representation of the syndrome former matrix $\mathbf{H_s}$ of an SC-LDPC-CC has polynomials in $F_2[x]$ as its entries, where $F_2[x]$ is the ring of polynomials with coefficients in the Galois field $F_2$. In this case, the code is described by a $c \times a$ [*symbolic matrix*]{} $$\mathbf{H}(x)=\left[\begin{array}{ccc} h_{0,0}(x) & \ldots & h_{0,a-1}(x)\\ \vdots & \ddots &\vdots \\ h_{c-1,0}(x) & \ldots & h_{c-1,a-1}(x)\end{array}\right], \label{eq:Hx}$$ where each $h_{i,j}(x)$, $i = 0, 1, 2, \ldots, c-1$, $j = 0, 1, 2, \ldots, a-1$, is a polynomial in $F_2[x]$. The code representation based on $\mathbf{H_s}$ can be converted into that based on $\mathbf{H}(x)$ by using the following expression $$h_{i,j}(x)=\sum_{m=0}^{m_h} h_{m}^{(i,j)} x^{m}, \label{eq:bintopol}$$ where $h_{m}^{(i,j)}$ is the $(i, j)$-th entry of the matrix $\mathbf{H}_m$, the latter being the transpose of the $m$-th block of $\mathbf{H_s}$. We focus on codes described by a symbolic parity-check matrix containing only polynomials with unitary weight, also known as *monomial codes*. In this case, $\mathbf{H}(x)$ can be described through an exponent matrix in the form $$\mathbf{P}=\left[\begin{array}{llll} p_{0,0} & p_{0,1} & \ldots & p_{0,a-1}\\ p_{1,0} & p_{1,1} & \ldots & p_{1,a-1}\\ \vdots & \vdots & \ddots & \vdots\\ p_{c-1,0} & p_{c-1,1} & \ldots & p_{c-1,a-1}\end{array}\right], \label{eq:expomatrix}$$ where $p_{i,j}$ is the exponent of the (only) non-null term in $h_{i,j}(x)$. The syndrome former memory order $m_h$ is the largest difference, in absolute value, between any two elements of $\mathbf{P}$. Code design via SCM {#Sec3} =================== In this section we recall the basic assumptions of the design method proposed in [@Tadayon2018]. The design of the parity-check matrix of a QC-LDPC block code with lifting degree $N$ starts from an exponent matrix having the following form (SMC assumption) $$\mathbf{P}^{\mathrm{SMC}}_{m\times n}=\left[\begin{array}{c|c|c|c|c} \vec{0} & \vec{P}_{1} & \gamma_{2} \otimes \vec{P}_{1} & \ldots & \gamma_{n-1} \otimes \vec{P}_{1} \end{array}\right], \label{eq:SMCexpomatrix}$$ with $m,n,N\in \mathbb{N}$, $m<n\leq N$, and $\vec{0}$ and $\vec{P}_{1}$ being column vectors with $m$ entries in $ \{0,\ldots, N-1\}$. The vector $\vec{0}$ is filled with all zero entries, while the entries of the vector $\vec{P}_{1}$ are chosen as follows: the first entry is zero, the second entry is one and the other entries are chosen in $\{2,\ldots,N - 1\}$ in increasing order. Then, the subsequent vectors have the form $\gamma_{j} \otimes \vec{P}_{1}$ ($j = 2,\ldots, n - 1$), where $\otimes$ denotes multiplication mod $N$, and are computed from $\vec{P}_{1}$ through sequential multiplications by the coefficients $\gamma_{j}\in \{2,\ldots ,N - 1\}$ such that $\gamma_{j}<\gamma_{j + 1}$. The following proposition holds, which generalizes [@Tadayon2018 Proposition 1]. \[SMCproposition\] Let $\mathbf{P}^{\mathrm{SMC}}_{m\times n}$ be the exponent matrix of a QC-LDPC block code $C$ as defined in (\[eq:SMCexpomatrix\]). Suppose that the Tanner graph associated to the submatrix $\left[\begin{array}{@{}c@{}|@{}c@{}} \vec{0} & \vec{P}_{1} \end{array}\right]$ contains no strictly avoidable cycles of length up to $\lambda$, $\lambda \in \{4,\ldots,10\}$. Then, the Tanner graph of $C$ has no strictly avoidable cycle of length up to $\lambda$ for sufficiently large $N$ and a proper choice of $\gamma_j$’s. Similar to the proof of [@Tadayon2018 Proposition 1] and omitted here for saving space. The pseudocode for the algorithm that finds the smallest possible $\gamma_j$, $j=2,\ldots, n-1$, leading to the desired girth can be found in [@Tadayon2018 Algorithm 1]. Decoding latency and complexity {#Sec4} =============================== In this section we discuss the latency and complexity of the considered decoding algorithms. QC-LDPC block codes ------------------- QC-LDPC block codes can be efficiently decoded by means of belief propagation (BP) algorithms. These algorithms must be executed over the whole length of the codeword, that is, $L$. So, the decoding latency, expressed as the number of bits that must be awaited before the decoding process starts, is $$\Lambda_{\mathrm{BP}} = L = nN. \label{eq:lambdaBP}$$ The per-output-bit decoding complexity can be measured as the number of binary operations required per decoding instance per output bit. We refer to the implementation of the BP decoder proposed in [@Hu2001] and define the average per-output-bit decoding complexity as $$\begin{aligned} \Gamma_{\mathrm{BP}}=\frac{LI_{\mathrm{avg}}f(m,R)}{L}=I_{\mathrm{avg}}f(m,R), \end{aligned} \label{eq:complex}$$ where $I_{\mathrm{avg}}$ is the average number of decoding iterations and $f(x,R)=[8(8x+12R-11)+x]$. Notice that $\Lambda_{\mathrm{BP}}$ benefits from a reduction in the code blocklength, whereas the per-output-bit complexity does not depend on $L$. SC-LDPC convolutional codes {#subsec:4b} --------------------------- SW iterative algorithms perform BP over a window including $W$ blocks of $a$ bits each, and then let this window slide forward by $a$ bits before starting over again. For each decoding window position, the SW decoder gives the first $a$ decoded bits as output, before letting the window shift forward by $a$ bits. To ensure that the performance loss due to the non-infinite size of the sliding window is negligible, the number of blocks has to be $W=\alpha(m_h+1)$, with $\alpha \geq 5$. By using this value of $W$, we can express the decoding latency ($\Lambda_{\mathrm{SW}}$) and average per-output-bit complexity ($\Gamma_{\mathrm{SW}}$) of a SW decoder as $$\begin{cases} \Lambda_{\mathrm{SW}} = Wa = \alpha(m_h+1)a,\\ \begin{aligned} \Gamma_{\mathrm{SW}} &= \frac{WaI_{\mathrm{avg}} f(c,R)}{a}=\\ &=\alpha(m_h+1)I_{\mathrm{avg}} f(c,R). \end{aligned} \end{cases} \label{cas}$$ Note that SC-LDPC-CCs characterized by small values of $m_h$ can be decoded with small window sizes. According to , this results in a reduction of both the decoding latency and per-output-bit complexity. Numerical results {#Sec5} ================= By applying the method proposed in [@Tadayon2018], we have designed several codes with girth $g=10, 12$. The values of $N$ and $m_h$ obtained for these codes are often significantly smaller than those of other codes with the same rate and girth reported in the literature. In particular, we have considered $m = 3, 4$ and $n = 4, \ldots, 12$ for the QC-LDPC block codes, and $c = 3, 4$ and $a = 4, \ldots, 12$ for the SC-LDPC-CCs. This choice derives from the fact that codes with $m=1,2$ ($c=1,2$) entail undesirable properties which yield a very poor performance, whereas codes with $m>4$ ($c>4$) usually exhibit degraded waterfall performance and yield large decoding complexity. We have compared the obtained values of $N$ and $m_h$ with those available in the literature. To the best of our knowledge, the design approaches that have produced till now the codes with minimum values of $N$ and $m_h$ are those reported in [@sullivan1; @MBAT2017; @Bocharova1; @ATasdighi1; @Amirzadeh1]. The exponent matrices of the newly designed codes are reported in Tables \[Table:TTab1\] to \[Table:TTab8\]. The lifting degree (syndrome former memory order) of the most compact existing codes is given between square brackets. The first row and column of any QC-LDPC block code exponent matrix are filled with all-zero entries and omitted. Latency and complexity performance ---------------------------------- We denote the smallest lifting degree and syndrome former memory order found through the considered approach as $\tilde{N}$ and $\tilde{m}_h$, respectively. The corresponding minimum values obtained through previous approaches are instead denoted as $N^*$ and $m_h^*$, respectively. According to , we can compute the ratio of the decoding latency of the newly designed QC-LDPC block codes over that of previous QC-LDPC block codes as $$\Theta_N=\frac{\tilde{N}}{N^*} \label{eq:ratioQC}.$$ Similarly, starting from , we can compute the ratio of the decoding latency and per-output-bit complexity achieved by the newly designed SC-LDPC-CCs over classical SC-LDPC-CCs as $$\Theta_{m_h}=\frac{\tilde{m}_h+1}{m_h^*+1}. \label{eq:ratios}$$ The smaller the values of $\Theta_N$ and $\Theta_{m_h}$, the larger the improvement over classical codes. The smallest values of $\Theta_N$ and $\Theta_{m_h}$ we have obtained in our examples are $0.47$ and $0.23$, respectively. Error rate performance ---------------------- In this section we assess the performance of the new codes in terms of bit error rate (BER) and block error rate (BLER) through Monte Carlo simulations of binary phase shift keying modulated transmissions. We consider an SC-LDPC-CC designed according to [@Tadayon2018] (noted as $C_{1}$) under full-size BP decoding and SW decoding with different window sizes. Notice that the BLER refers to the block of $a$ target symbols decoded within each decoding window. The maximum number of decoding iterations is $100$, and the actual number of iterations is equal to the maximum for SW decoding (which does not use any stopping criteria based on parity-checks). We also consider an SC-LDPC-CC (noted as $C_{2}$) with the same code rate and girth as $C_{1}$, obtained by unwrapping a QC-LDPC block code designed following [@Bocharova1]. The parameters of the two codes and the value of $\Theta_{m_h}$ are shown in Table \[table:Tabparamsc\]. Their performance is shown in Figure \[fig:perf\]. We notice that $C_{\mathrm{1}}$ and $C_{\mathrm{2}}$ have almost coincident performance when $W\rightarrow \infty$. However, when the window size is relatively small, $C_{1}$ outperforms $C_{2}$. This happens because the small window sizes imply $\alpha<5$ for $C_{2}$, thus degrading performance, according to the discussion in Section \[subsec:4b\]. The code $C_{1}$, instead, has a value of $v_s$ which is about twice as small as that of $C_{\mathrm{2}}$, yielding values of $\alpha$ that are about twice as big as those of $C_{2}$, and this results in a better performance under SW decoding. For completeness, let us consider the codes in Table \[Table:TTab7\] with rate $R=\frac{a-3}{a}$, for $a=6,7,8$, denoted as $\bar{C}_a$, and assess the performance loss $\Delta_{\mathrm{dB}}$ in which they incur at $\mathrm{BER}=10^{-4}$, under full-size BP decoding, with respect to the most compact codes in the literature, noted as $C^*$. The results are reported in Table \[table:Tabloss\], from which we observe that the price paid in terms of $\Delta_{\mathrm{dB}}$ for the corresponding benefit in terms of $\Theta_{m_h}$ is very small. \[table:Tabparamsc\] Code $a$ $c$ $m_h$ $v_s$ $g$ $\Theta_{m_h}$ --------- ----- ----- ------- -------- ------ ---------------- -- -- $C_{1}$ $8$ $3$ $297$ $2384$ $12$ $C_{2}$ $8$ $3$ $652$ $5224$ $12$ : Parameters of the considered SC-LDPC-CCs with $R=\frac{5}{8}$ ![Simulated performance of SC-LDPC-CCs with $g=12$ as a function of the signal-to-noise ratio.[]{data-label="fig:perf"}](berbler.eps){width="78mm"} \[table:Tabloss\] $\bar{C}_6$ $C^*_{6}$ $C_7$ $C^*_{7}$ $C_8$ $C^*_{8}$ ------------------------ ------------- ----------- ------- ----------- ------- ----------- -- -- -- $\Delta_{\mathrm{dB}}$ $\Theta_{m_h}$ : $\Delta_{\mathrm{dB}}$ and $\Theta_{m_h}$ of SC-LDPC-CCs with $c=3$ and $g=12$ Conclusion {#Sec6} ========== Compact QC-LDPC block codes and SC-LDPC-CCs designed through a novel approach based on SMCs allow achieving reduced decoding latency and per-output-bit decoding complexity, while exhibiting comparable error rate performance with respect to previous solutions. [10]{} G. Durisi, T. Koch and P. Popovski, “Toward massive, ultrareliable, and low-latency wireless communication with short packets,” [* Proc. IEEE*]{}, vol. 104, no. 9, pp. 1711-1726, Sep. 2016. M. Battaglioni, M. Baldi and E. Paolini, “Complexity-constrained spatially coupled LDPC codes based on protographs,” [*Proc. [IEEE]{} ISWCS 2017*]{}, Bologna, Italy, pp. 49–53, Aug. 2017. Z. Li, L. Chen, L. Zeng, S. Lin, and W. Fong, “Efficient encoding of quasi-cyclic low-density parity-check codes,” [*IEEE Trans. Commun.*]{}, vol. 54, no. 1, pp. 71-81, Jan. 2006. L. Lan, L. Zeng, Y. Y. Tai, L. Chen, S. Lin, and K. Abdel-Ghaffar, “Construction of quasi-cyclic LDPC codes for AWGN and binary erasure channels: A finite field approach,” [*IEEE Trans. Inf. Theory*]{}, vol. 53, no. 7, pp. 2429-–2458, Jul. 2007. A. J. Felström and K. S. Zigangirov, “Time-varying periodic convolutional codes with low-density parity-check matrix,” [*IEEE Trans. Inf. Theory*]{}, vol. 45, no. 6, pp. 2181–2191, Sep. 1999. M. Lentmaier, A. Sridharan, K. S. Zigangirov, and D. J. Costello, “Terminated LDPC convolutional codes with thresholds close to capacity," [*Proc. [IEEE]{} ISIT 2005*]{}, Adelaide, Australia, pp. 1372–-1376, Sep. 2005. M. H. Tadayon, A. Tasdighi, M. Battaglioni, M. Baldi and F. Chiaraluce, “Efficient search of compact QC-LDPC and SC-LDPC convolutional codes with large girth," [*IEEE Commun. Lett.*]{}, vol. 22, no. 6, pp. 1156–1159, Jun. 2018. M. Hagiwara, M. P. C. Fossorier, T. Kitagawa, and H. Imai, “Smallest size of circulant matrix for regular (3, L) and (4, L) quasi-cyclic LDPC codes with girth 6,” [*IEICE Trans. Fundam. Electron., Commun. Comput. Sci.*]{}, vol. E92-A, no. 11, pp. 2891-–2894, Nov. 2009. A. Tasdighi, A. H. Banihashemi, and M. R. Sadeghi, “Efficient search of girth-optimal QC-LDPC codes,” [*IEEE Trans. Inf. Theory*]{}, vol. 62, no. 4, pp. 1552–1564, Apr. 2016. M. Karimi and A. H. Banihashemi, “On the girth of quasi cyclic protograph LDPC codes,” [*IEEE Trans. Inf. Theory*]{}, vol. 59, no. 7, pp. 4542-–4552, Jul. 2013. F. Amirzadeh and M. R. Sadeghi, “Lower bounds on the lifting degree of single-edge and multiple-edge QC-LDPC codes by difference matrices,” \[Online\], Available: https://arxiv.org/pdf/1709.00825.pdf M. Battaglioni, A. Tasdighi, G. Cancellieri, F. Chiaraluce, and M. Baldi, “Design and analysis of time-invariant SC-LDPC convolutional codes with small constraint length," [*IEEE Trans. Commun.*]{}, vol. 66, no. 3, pp. 918-931, Mar. 2018. M. P. C. Fossorier, “Quasi-cyclic low-density parity-check codes from circulant permutation matrices,” [*IEEE Trans. Inf. Theory*]{}, vol. 50, no. 8, pp. 1788-1793, Aug. 2004. X.-Y. Hu, E. Eleftheriou, E., D.-M. Arnold, and A. Dholakia, “Efficient implementations of the sum-product algorithm for decoding LDPC codes," [*Proc. [IEEE]{} GLOBECOM ’01*]{}, San Antonio, TX, pp. 1036–1036E, Nov. 2001. M. E. O’Sullivan, “Algebraic construction of sparse matrices with large girth,” [*IEEE Trans. Inf. Theory*]{}, vol. 52, no. 2, pp. 718–727, Feb. 2006. I. E. Bocharova, F. Hug, R. Johannesson, B. D. Kudryashov, and R. V. Satyukov, “Searching for voltage graph-based LDPC tailbiting codes with large girth,” [*IEEE Trans. Inf. Theory*]{}, vol. 58, no. 4, pp. 2265–2279, Apr. 2012.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Beijing Spectrometer (BES) at Beijing Electron Positron Collider (BEPC), IHEP has accumulated 2.5 $\times 10^7$ J/$\Psi$ and plans to increase the number to $10^8 \sim 10^9$ in the near future. In this paper I review and summarize the recent studies on the possibilities of probing for new physics at BES. This includes the theoretical works on and experimental attempt performed at BES in searching for flavor changing neutral current, CP violation and lepton flavor violation in the processes of J/$\Psi$ decays.' --- plus 1pt ‘@=12 -0.5in 0.0in 0.0in 8.5in -.375in =0 October 11, 2000 [**[Probing For New Physics In J/$\Psi$ Decays ]{}\ **]{} .3in [*Institute of High Energy Physics (IHEP), Academia Sinica\ *]{} [*Beijing 100039, P.R. China\ *]{} .8in **[Lepton flavor violating J/$\Psi$ decays]{}** =============================================== The J/$\Psi$ rare decay processes $J/ \Psi ~ \rightarrow {\bar l} l^\prime$ ( $l, ~ {l^\prime}$ = $\tau ,~ \mu , e$ ) conserve total lepton number, but violate the individual lepton numbers. In the standard model the lepton flavor symmetries are conserved, but speculated to be violated in many extensions of the standard model, such as supersymmetric standard models, left-right symmetric models and models where electroweak symmetry is broken dynamically. Recent Super-Kamiokande experiment results indicate that neutrinos have non-vanishing masses, mix with each others and consequently that lepton flavor symmetry and/or lepton number symmetry are broken symmetries. In cosmology mystery of matter and antimatter asymmetry might be understood in terms of the brokendown of the lepton number symmetry together with the non-perturbative effects (Sphaleron) of the standard electroweak theory. There has been a lot of studies both theoretically and experimentally on testing the lepton flavor conservation law. At present we have various bounds listed in the particle data book from $\mu$ decays, $\tau$ and Z gauge boson decays\[1\]. With a large sample of J/$\Psi$, BES will be able to make an additional experimental searching for lepton flavor violation. To estimate the branch ratio of lepton flavor violating J/$\Psi$ decays allowed by the current experimental data, Peccei, Wang and I\[2\] took a model-independent approach to new physics and introduced a four-fermion contact interaction $$\begin{aligned} \frac{4 \pi}{\Lambda^2} {\bar c} \gamma^\mu c ~ {\bar l} \gamma_\mu l^\prime , \ \end{aligned}$$ where $\Lambda$ is the new physics cutoff. This effective operator is forbidden in the standard model, however will be generated in theories where lepton flavor is not conserved, such as the minimal supersymmetric standard model with and/or without R parity, models with large extra dimension\[3\]. Therefore, any observed signal is a direct evidence for non- standard physics and will improve our understanding of flavor dynamics, especially in the lepton sector. There is no direct experimental limit on $\Lambda$ in (1). However, at one-loop, attaching the neutral gauge boson $Z$ to the charm quark loop generates an effective coupling of $Z$ to ${\bar l} l^\prime$ . From the limits given in the particle data book on $Br(Z \rightarrow {\bar l} l^\prime )$\[1\], we obtained the lower bounds on the branch ratio of the $J/ \Psi$ decay into leptons: $$\begin{aligned} &Br(J/\Psi \rightarrow \tau^+ e^- ) <2.7 \times 10^{-5};\\ &Br(J/\Psi \rightarrow \tau^+ \mu^-) <4.9 \times 10^{-5};\\ &Br(J/\Psi \rightarrow \mu^+ e^-) <8.3 \times 10^{-6}.\end{aligned}$$ Recently Nussinov, Peccei and I\[4\] have examined “unitarity inspired" relations between two- and three-body lepton flavor violating decays and found that the existing strong bounds on $\mu \to 3e$ and $ \mu\to e\gamma\gamma$ severly constrain two-body lepton flavor violating decays of vector bosons \[$J/\psi,~\Upsilon$, and $Z^o$\] or pseudoscalars \[$\pi^o,\eta$\] into $\mu^\pm e^\mp$ final states. However the bounds derived in Ref.\[4\] can be avoided if there is a kinematical suppression or as a result of some cancellations. Searching for lepton flavor violating decays of vector bosons such as $J/\Psi$ remains a worthwhile experimental challenge. BES started one year ago\[5\] and has been working on an experiment of searching for $J/\Psi \rightarrow e^\mp \mu^\pm$. They will publish their result officially in the near future. **[Single D meson production in $J/\Psi$ decays]{}** ==================================================== Kinematically $J/\Psi$ can not decay into D meson pairs, however it is able to decay into single D meson. In the standard model, these Cabbibo suppressed and/or favored weak decays have a typical branch ratio $\sim 10^{-8}$ or smaller, which is unobservable and because of which these processes serve as a probe of new physics. Recently Datta, O’Donnell, Pakvasa and I\[6\] have studied the possibility of searching for new flavor changing neutral current in the decay of $J/\Psi$. The purpose of the study is to answer whether new physics can enhance sufficiently for the processes to be observable in the near future experiments. We first perform a model independent analysis, then examine the predictions of the models, such as TopColor models, minimal sypersymmetric standard model with R-parity violation and the two Higgs doublet model. We found that the branch ratio of $J/\Psi \rightarrow D/ {\overline D} X_u$ could be as large as $10^{-5}$\[6\]. Experimentally with BES-II data $\sim 8 \times 10^{6} J/\Psi$, BES found no signal of single D production in $J/\Psi$ decays, and put limits on decay rates \[7\](at $90\%$ C.L. ): $Br ( J/\Psi \rightarrow {\overline D^0} \rho^0 ) \leq 1.25 \times 10^{-5}; ~~~~~ Br (J/\Psi \rightarrow D_s^+ K^- ) \leq 2.8 \times 10^{-5}$, and $Br( J/\Psi \rightarrow D / {\overline D} X_u ) \leq 1.7 \times 10^{-4}$. These are preliminary results. Next year BES will have collected $\sim 5 \times 10^7 J/\Psi$. The upper limit on $Br( J/\Psi \rightarrow D / {\overline D} X_u )$ is expected to be reduced to $\sim 3 \times 10^{-5}$, which is very close to the theoretical prediction. If there is still no signal found, it will put some constraints on models beyond the standard electroweak theory\[6\]. **[CP violation in $J/\Psi$ decays ]{}** ======================================== The origin of CP violation remains one of the outstanding problems in particle physics and cosmology. To pin down the sources and nature of CP violation in or beyond the Cabbibo-Kabayashi-Maskawa model, it would be necessary to consider different observations of CP violation in different channels from the K system, B sysytem, [*etc*]{}. The reaction of interest at BES is\[8\] $$e^+(p) + e^-(-p) \rightarrow J/\Psi \rightarrow A(q_-) + {\bar B(q_+)} + X ~~~~~, ~~~ (5)$$ where A (${\bar B}$) are charged particles, for instance $\pi^\pm$ in the processes of three or five pions decay of the $J/\Psi$, and $p, q_- $ and $q_+$ are corresponding momentum in the laboratory frame. Define CP/T-odd operator $$O_1 ~~ = ~~\frac{{\vec p} \cdot {\vec q_+} \times {\vec q_-}}{ | {\vec p} \cdot {\vec q_+} \times {\vec q_-} | }~~. ~~~ (6)$$ If there exists CP violating interaction in $J/\Psi$ decays, one would expect a non-vanishing expectation value of operator $O_1$. Theoretically there could be many sources responsible for the CP violation in the process. One of them is the Chromo-dipole moment of the charm quark $ i g_c \frac{d_g}{2} {\overline c}\sigma_{\mu\nu} \gamma_5 {\lambda^a} c G_a^{\mu\nu} $ where $d_g$ has ${\rm [mass]}^{-1}$ dimension and the $g_c$ the strong interaction coupling constant. Consider the process of three pion decay of the $J/\Psi$. Its branch ratio is around $1.5\%$. With $10^8 \sim 10^9 J/\Psi$ at BEPC II there will be around $10^6 \sim 10^7$ available for the analysis of CP violation. To estimate the experimental sensitivities to $d_g$, let us consider only the statistical uncertainties. Neglecting the systematical uncertainties one expects to be able at BES to probe for $d_g$ as small as $ ( 1/{\sqrt{10^6 \sim 10^7}}) \times ({1 / m_c}) \leq 10^{-17} {\it cm}$. One can easily construct different kind of CP/T-odd operators of three momentum products ($p, q_-, q_+ $) for the analysis of CP violation in $J/\Psi$ decays. For instance, $$O_2^{i j} \sim {( {\vec q_+ } - {\vec q_-} )}^i \cdot {( {\vec q_+} \times {\vec q_-} )}^j + (i <-> j). ~~~~ (7)$$ The initial electron and/or positron beams are not polarized at BEPC, otherwise one would be able to construct observables with the initial polarization vector $\vec \sigma$, $$O_3 \sim {\vec \sigma} \cdot {\vec q_-} \times {\vec q_+}.~~~~~~~~~~~ (8)$$ There has been proposal to measure CP violation in $e^+ e^- \rightarrow J/\Psi \rightarrow {\Lambda} {\bar\Lambda}$. With a large sample of $J/\Psi$, one expects to probe for and put a strong bound on the electric dipole moment of $\Lambda$\[9\]. Conclusion and comments ======================= In this paper I have concentrated on three kind of processes to probe for new physics in the decay of $J/\Psi$. There are some other rare decay modes which are interesting, but not reviewed. For instance, with a large sample of $J/\Psi$, it is possible and physically interesting to search for Goldstone or Pseudo-Goldstone J, such as Axion in the process $$J/\Psi \rightarrow \gamma + J ~~~~~~. ~~~~~~~~~~(9)$$ Another example is the invisible decay of $J/\Psi$ investigated by Chang, Lebedev and Ng\[10\] recently in models with extra Z-bosons, minimal sypersymmetric standard model with R-parity violation and decays into Goldstinos. The third example is provided by Bijnen and Maul\[11\], who recently have calculated in detail the branch ratio of $J/\Psi$ decay into photon + missing energy in the popular theory these days with large extra dimension. They found that the branch ratio could be as large as $10^{-5}$, which is measurable at BES. Before conclusion, we point out that even though we focus our discussions here on $J/\Psi$ decays, it is quite easy to apply the studies in this paper for $\Psi^\prime$ and $\Upsilon$ system\[12\]. [**ACKNOWLEDGMENTS**]{} I am grateful to my collaborators and colleagues for discussions. This work is supportted in part by the NSF of China. [**REFERENCES**]{} 1. Particle Data Group, C. Caso ${\it et al.}$, Europ. Phys. J, C3, 1 (1998). 2. “Probing for lepton flavor violation in decays of charmonium and bottomonium systems”, R.D. Peccei, Jian-Xiong Wang and Xinmin Zhang, May 1998 Note (unpublished); Xinmin Zhang, invited talked given at the national conference on high energy physics, Chengde, China, April (1998). 3. For examples, see, Z.K. Silagadze, hep-ph/9907328, July (1999); T. Huang, Z. Lin and X. Zhang, hep-ph/0009353 (2000); Chan Hong-Mo et al, hep-ph/0006338, hep-ph/0007004, hep-ph/0008313/0008324. 4. S. Nussinov, R.D. Peccei and X. Zhang, Hep-ph/0004153, April (2000). 5. G. Tong et al (Private communication). 6. A. Datta, P.J. O’Donnell, S. Pakvasa and X. Zhang, Phys. Rev. D60, 014011 (1999). 7. G. Rong et al (in preparation). 8. Xinmin Zhang, Jian-Xiong Wang, Jian-Pin Ma, Dongshen Du and Wu-Jun Huo (in preparation). 9. Xiao-Gang He, Jian-Pin Ma and McKellar, Phys. Rev. D49, 4548 (1994); Ye Yiun-Xiou and Ye Zheng-Yu (unpublished). 10. N. Chang, O. Lebedev and J.N. Ng, hep-ph/9806497, June (1998). 11. J. Bijnens and M. Maul, hep-ph/0006042, July (2000). 12. For example, it has been proposed to probe for CP violation in the process $\Psi^\prime \rightarrow J/\Psi + \pi^+ + \pi^- $, \[ Xinmin Zhang, Dongshen Du, Pin Wang, A. Datta, Jian-Xiong Wang and Jian-Pin Ma, May 1998 note, unpublished; The experimental analysis started already(Jin Li and Zhi-Jin Guo, private communication)\].
{ "pile_set_name": "ArXiv" }
--- abstract: | We present the first study of disordered jammed hard-sphere packings in four-, five- and six-dimensional Euclidean spaces. Using a collision-driven packing generation algorithm, we obtain the first estimates for the packing fractions of the maximally random jammed (MRJ) states for space dimensions $d=4$, $5$ and $6$ to be $\phi_{MRJ} \simeq 0.46$, $0.31$ and $0.20$, respectively. To a good approximation, the MRJ density obeys the scaling form $\phi_{MRJ}= c_1/2^d+(c_2 d)/2^d$, where $c_1=-2.72$ and $c_2=2.56$, which appears to be consistent with high-dimensional asymptotic limit, albeit with different coefficients. Calculations of the pair correlation function $g_{2}(r)$ and structure factor $S(k)$ for these states show that short-range ordering appreciably decreases with increasing dimension, consistent with a recently proposed “decorrelation principle,” which, among othe things, states that unconstrained correlations diminish as the dimension increases and vanish entirely in the limit $d \rightarrow \infty$. As in three dimensions (where $\phi_{MRJ} \simeq 0.64$), the packings show no signs of crystallization, are isostatic, and have a power-law divergence in $g_{2}(r)$ at contact with power-law exponent $\simeq 0.4$. Across dimensions, the cumulative number of neighbors equals the kissing number of the conjectured densest packing close to where $g_{2}(r)$ has its first minimum. Additionally, we obtain estimates for the freezing and melting packing fractions for the equilibrium hard-sphere fluid-solid transition, $\phi_F \simeq 0.32$ and $\phi_M \simeq 0.39$, respectively, for $d=4$, and $\phi_F \simeq 0.19$ and $\phi_M \simeq 0.24$, respectively, for $d=5$. Although our results indicate the stable phase at high density is a crystalline solid, nucleation appears to be strongly suppressed with increasing dimension. author: - Monica Skoge - Aleksandar Donev - 'Frank H. Stillinger' - Salvatore Torquato title: 'Packing Hyperspheres in High-Dimensional Euclidean Spaces' --- Introduction ============ Hard-sphere systems are model systems for understanding the equilibrium and dynamical properties of a variety of materials, including simple fluids, colloids, glasses, and granular media. The hard-sphere potential is purely repulsive; it is infinite when two spheres overlap, but otherwise zero. Despite the simplicity of the potential, hard-sphere systems exhibit rich behavior: they undergo a fluid-solid phase transition and can exhibit glassy behavior. Of particular recent interest are (nonequilibrium) disordered jammed packings of hard spheres and their statistical and mechanical properties, such as the maximally random jammed (MRJ) state [@rcp; @kansal], pair correlations [@aleks], isostaticity [@aleks], and density fluctuations [@alekshyper]. Such packings have been intensely studied computationally in two and three dimensions [@finney; @bennett; @tobochnik; @zinchenko; @rcp; @kansal; @rintoul; @aleks; @alekshyper; @ohern; @binarydisk] and in this paper we extend these studies to four, five and six dimensions. A hard-sphere packing in $d$-dimensional Euclidean space $\mathbb{R}^d$ is an arrangement of congruent spheres, no two of which overlap. As in a variety of interacting many-body systems [@chaikin], we expect studies of hard-sphere packings in high dimensions to yield great insight into the corresponding phenomena in lower dimensions. Analytical investigations of hard-spheres can be readily extended into arbitrary spatial dimension [@virial234; @virial234too; @EOSlubanmichels; @virial56; @Virial4_EvenDimensions; @frisch; @parisi; @finken; @philipse; @torquatobook; @hyperuniformity; @Pa06; @decorrelation; @torquato15; @To06] and high dimensions can therefore be used as a stringent testing ground for such theories. Along these lines and of particular interest to this paper, predictions have been made about long-wavelength density fluctuations [@hyperuniformity] and decorrelation [@decorrelation; @torquato15] in disordered hard-sphere packings in high dimensions. Additionally, the optimal packing of hard spheres in high dimensions is also of interest in error-correcting codes in communications theory [@conway]. Our focus in this paper will be the study of hard-sphere packings in four, five and six dimensions. Specifically, we consider both equilibrium packings for $d=4$ and $d=5$ and nonequilibrium packings representative of the maximally random jammed state for $d=4$, $d=5$ and $d=6$. Equilibrium thermodynamic properties of hard-sphere packings for $d=4$ and $d=5$ have been studied both computationally and with approximate theories [@EOSlubanmichels; @finken; @Lue]. For the low-density fluid, lower-order virial coefficients, $B_2$, $B_3$, and $B_4$, are known exactly for arbitrary dimensionality [@virial234; @virial234too; @Virial4_EvenDimensions]. Higher-order virial coefficients have been calculated by Monte Carlo simulation, $B_5$, $B_6$ and $B_7$ for both $d=4$ and $d=5$ [@virial56] and $B_8$ for $d=4$ [@virial56], and analytically [@Lyberg; @Virial89_HighDimensions]. The pair correlation function for equilibrium fluids has been studied and a decrease in ordering with increasing dimension was readily apparent [@bishop]. Hard-sphere systems have been shown to undergo a (first-order) fluid-solid phase transition by numerical simulations for $3 \leq d \leq 5$ [@michelstrapp] and with approximate theories for $d$ as high as $50$ [@finken]. The freezing points for $d=4$ and $d=5$ were estimated numerically to occur at packing fractions $\phi_F \approx 0.5 \phi_{max} = 0.31$ and $\phi_F \approx 0.4 \phi_{max} = 0.19$, respectively, and it was conjectured that freezing occurs at lower packing fractions relative to close packing as the dimension increases [@michelstrapp]. The [*packing fraction*]{} $\phi$ is the fraction of space $\mathbb{R}^d$ covered by the spheres, i.e., $$\phi=\rho v_1(R), \label{phi}$$ where $\rho$ is the number density, $$v_1(R) = \frac{\pi^{d/2}}{\Gamma(1+d/2)} R^d \label{v(R)}$$ is the volume of a $d$-dimensional sphere of radius $R$, and $\Gamma(x)$ is the gamma function [@torquatobook]. At sufficiently large densities, the packing of spheres with the highest jamming density has the greatest entropy because the free-volume entropy dominates over the degeneracy entropy. Therefore, the high-density equilibrium phase corresponds to the optimal packing, [*i.e.,*]{} maximal density. The densest packing for $d=3$ was recently proven by Hales [@hales] to be attained by the FCC lattice with packing fraction $\phi_{max} = \pi/\sqrt{18} = 0.7404 \dots$. The kissing number $Z$, the number of spheres in contact with any given sphere, for the FCC lattice corresponds to the maximal kissing number $Z_{max} = 12$ for $d=3$. One of the generalizations of the FCC lattice to higher dimensions is the $D_d$ checkerboard lattice, defined by taking a cubic lattice and placing spheres on every site at which the sum of the lattice indices is even ([*i.e.,*]{} every other site). The densest packing for $d=4$ is conjectured to be the $D_4$ lattice, with packing fraction $\phi_{max} = \pi^2/16 = 0.6168 \ldots$ and kissing number $Z = Z_{max} = 24$ [@conway], which is also the maximal kissing number in $d=4$ [@musin]. For $d=5$, the densest packing is conjectured to be the $D_5$ lattice, with packing fraction $\phi_{max} = 2\pi^2/(30\sqrt{2}) = 0.4652 \ldots$ and kissing number $Z = 40$ [@conway]. For $d=6$, the densest packing is conjectured to be the “root” lattice $E_6$, with density $\phi_{max} = 3\pi^3/(144\sqrt{3}) = 0.3729 \ldots$ and kissing number $Z = 72$ [@conway]. The maximal kissing numbers $Z_{max}$ for $d=5$ and $d=6$ are not known, but have the following bounds: $40 \leq Z_{max} \leq 46$ for $d=5$ and $72 \leq Z_{max} \leq 82$ for $d=6$ [@conway]. In very high dimensions, it has been suggested that random packings of spheres might have a higher density than ordered packings, enabling the intriguing possibility of disordered ground states and hence thermodynamic glass transitions [@decorrelation]; see also Ref. [@torquato15]. Equilibrium hard-sphere systems for $d=2$ and $d=3$ crystallize into ordered packings upon densification. However, for $d=3$, it has been found both experimentally [@Sc69] and computationally [@rintoul; @rcp; @aleks] that if the system is densified sufficiently rapidly, the system can be kept out-of-equilibrium and can jam in a disordered state. A jammed packing is one in which the particle positions are fixed by the impenetrability constraints and boundary conditions, despite thermal or mechanical agitation of the particles or imposed boundary deformations or loads. Depending on the boundary conditions, different jamming categories can be precisely defined, including local, collective and strict jamming [@linprogramming; @category1; @category2]. The density of disordered collectively jammed hard-sphere packings for $d=3$ is around $\phi \simeq 0.64$ for a variety of packing-generation protocols and has traditionally been called random close packing (RCP) [@torquatobook]. However, Ref. [@rcp] showed that RCP is ill-defined because “random” and “close packed” are at odds with one another and the precise proportion of each of these competing effects is arbitrary. Therefore, Ref. [@rcp] introduced the concept of the maximally random jammed (MRJ) state to be the most disordered jammed packing in the given jamming category. This definition presupposes an order metric $\psi$ can be defined such that $\psi = 1$ corresponds to the most ordered ([*i.e.,*]{} crystal) packing and $\psi = 0$ corresponds to the most disordered packing, in which there are no spatial correlations. Figure \[phipsi\] from Ref. [@rcp] shows where MRJ lies on a schematic diagram of the space of jammed packings in the density-disorder $\phi$-$\psi$ plane. ![A highly schematic plot of the subspace in the density-disorder $\phi-\psi$ plane, where strictly jammed three-dimensional packings exist, as adapted from Ref. [@rcp]. Point A corresponds to the lowest-density jammed packing, and it is intuitive to expect that a certain ordering will be needed to produce low-density jammed packings. Point B corresponds to the most dense jammed packing, which is also expected to be the most ordered. Point MRJ represents the maximally random jammed state. The jamming region in the $\phi$-$\psi$ plane will of course depend on the jamming category. The gray region is devoid of hard-sphere configurations. []{data-label="phipsi"}](phi_psi_jammed.eps){width="0.7\columnwidth"} In this paper, we numerically study MRJ packings of hard spheres for $d=4$, $5$ and $6$ that are at least collectively jammed and report the first estimates of the packing fractions of the MRJ states [@rcp] in these dimensions to be $\phi_{MRJ} \simeq 0.46$, $0.31$ and $0.20$, respectively. We find that short-range ordering exhibited by $g_{2}(r)$ and $S(k)$ appreciably diminishes with increasing dimension, consistent with a recently proposed “decorrelation principle” stating that unconstrained spatial correlations vanish asymptotically in high dimensions and that the $n$-particle correlation function $g_n$ for any $n \ge 3$ can be inferred entirely from a knowledge of the number density $\rho$ and the pair correlation function $g_2({\bf r})$  [@decorrelation; @footnote]. We also explore equilibrium properties, in particular the fluid-solid phase transition, for $d=4$ and $d=5$, and find a decreased tendency to crystallize with increasing dimension. This paper is organized as follows: Section II explains the simulation procedure, Section III gives equilibrium results for $d=4$ and $d=5$, Section IV gives results for disordered jammed packings for $d=4$, $5$ and $6$, and Section V summarizes and discusses our results. Simulation Procedure ==================== We use event-driven molecular dynamics and a modified Lubachevsky-Stillinger (LS) algorithm [@LS1], as in Ref. [@HSalg], to produce collectively-jammed hard-sphere packings. As in Ref. [@HSalg], our algorithm uses periodic boundary conditions applied to a hypercubic cell, in which a fundamental cell containing $N$ spheres is periodically replicated to fill all of Euclidean space. We also use the cell method, in which the computational domain is divided into cubic cells and only neighboring cells are checked when predicting collisions for a given sphere. Since the number of neighboring cells, as well as the number of spheres per cell, increases considerably with increasing dimension, working in high dimensions is computationally intensive. Additionally, eliminating excessive boundary effects requires on the order of ten sphere diameters per simulation box length, [*i.e.,*]{} on the order of $N=10^d$ spheres. Due to the increasing computational load with increasing dimension, we cannot yet study $d > 6$. Implementing the near-neighbor list (NNL) techniques from Ref. [@HSalg], as well as parallelization, are necessary in order to study higher dimensions. Dimension-independent C++ codes used to generate the data in this paper can be downloaded at [http://cherrypit.princeton.edu/Packing/C++/]{}. Starting from a Poisson distribution of points, the points grow into nonoverlapping spheres of diameter $D$ at an expansion rate $\gamma = dD/dt$, while the positions of the spheres evolve in time according to Newtonian mechanics, augmented with energy non-conserving collisions. Spheres receive an extra energy boost after the collision due to the positive expansion rate. In practice, the starting configurations for our packing algorithm are low density random-sequential-addition packings of spheres [@torquatobook]. As the density increases, statistics, such as pressure, are collected. In the limit $\gamma \rightarrow 0$, the system is in equilibrium; for small but nonzero $\gamma$, the system is in quasi-equilibrium; and for large $\gamma$, the system is out of equilibrium. Eventually, a jammed state with diverging collision rate is reached. For studies of amorphous jammed packings, the expansion must be initially fast to suppress crystallization and maximize disorder, but at sufficiently high pressure, the expansion rate must be slow enough to allow local particle rearrangements necessary to achieve jamming [@aleks]. Equilibrium and Metastable Properties ===================================== The temperature in equilibrium systems of hard spheres is a trivial variable; [*i.e.,*]{} it does not affect equilibrium configurational correlations, leaving only one independent thermodynamic state variable, which can be taken to be either the reduced pressure $p = PV/Nk_BT$ or the density $\phi$, related through the equation of state (EOS). Hard-sphere systems undergo a (first-order) fluid-solid phase transition, characterized by a melting point, [*i.e.,*]{} the density at which the crystal thermodynamically begins to melt, and a freezing point, [*i.e.,*]{} the density at which the fluid thermodynamically begins to freeze. Equilibrium properties, such as the melting and freezing points, are studied here using small expansion rates ($\gamma = 10^{-5}-10^{-9}$) and periodic rescaling of the average sphere velocity to one, such that the total change in kinetic energy of the system, due to the collisions between growing spheres, was kept small. Strictly speaking, a positive growth rate yields nonequilibrium packings but equilibrium packings result as the growth rate tends to zero. The packings were “equilibrated" by verifying that orders of magnitude of change in the expansion rate did not change the resulting equation of state. In this section we only consider four and five dimensions due to (presently) prohibitive computational costs for higher dimensions. Figure \[melt\] shows the reduced pressure $p$ as a function of density $\phi$ for (a) simulations of $d=4$ systems of spheres placed in a $D_4$ lattice with negative expansion rate $\gamma = -10^{-6}$ and (b) simulations of $d=5$ systems of spheres placed in a $D_5$ lattice with negative expansion rate $\gamma = -10^{-5}$. The pressure initially follows the (lower) crystal branch, until the system becomes mechanically unstable and jumps onto the (higher) fluid branch. Also plotted is the theoretical prediction of Luban and Michels (LM) for the equation of state [@EOSlubanmichels], which agrees well with our numerical results for the fluid branch for $d=4$, but less so for $d=5$. It is a computational observation that crystals become mechanically unstable, giving rise to a sudden jump in pressure, at a density close to the freezing point [@melting1; @melting2]. Such “superheating” (undercompression) is most likely due to the difficulty of achieving coexistence in finite systems, although we are not aware of a theoretical analysis. From the results in Fig. \[melt\], we estimate the freezing points for $d=4$ and $d=5$ to be $\phi_F \simeq 0.31-0.32$ and $\phi_F \simeq 0.19-0.20$, respectively. The melting points for $d=4$ and $d=5$ can also be estimated from the data in Fig. \[melt\]. Since throughout the coexistence region the fluid and solid have the same absolute pressure $P$, the melting density can be estimated as the density on the crystal branch with the same absolute pressure $P$ as that at the freezing point. The coexistence region is plotted in Fig. \[melt\] and the melting packing fractions for $d=4$ and $d=5$ are estimated to be $\phi_M \simeq 0.38-0.40$ and $\phi_M \simeq 0.24-0.26$, respectively. We also observe that the reduced pressure at the freezing point is $p_F \simeq 12$ in both $d=4$ and $d=5$, which agrees with the reduced pressure at the freezing point for $d=3$, $p_F \simeq 12.3$, obtained from free energy calculations [@frenkel]. ![(Color online) Reduced pressure $p$ as a function of density $\phi$, for a range of system sizes (see legend), for (a) $d=4$ systems of spheres, initially in a $D_4$ lattice, and negative expansion rate $\gamma = -10^{-6}$ and (b) $d=5$ systems of spheres, initially in a $D_5$ lattice, and negative expansion rate $\gamma = -10^{-5}$. $N$ was chosen to make a perfect $D_d$ lattice with periodic boundary conditions, [*i.e.,*]{} $N = (2n)^d/2$ for $n~ \epsilon ~\mathbb{Z}$. Also plotted is the theoretical prediction of Luban and Michels (LM) for the equation of state [@EOSlubanmichels]. Curves for larger system sizes lie farther to the right. []{data-label="melt"}](melt6.eps "fig:"){width="0.7\columnwidth"} ![(Color online) Reduced pressure $p$ as a function of density $\phi$, for a range of system sizes (see legend), for (a) $d=4$ systems of spheres, initially in a $D_4$ lattice, and negative expansion rate $\gamma = -10^{-6}$ and (b) $d=5$ systems of spheres, initially in a $D_5$ lattice, and negative expansion rate $\gamma = -10^{-5}$. $N$ was chosen to make a perfect $D_d$ lattice with periodic boundary conditions, [*i.e.,*]{} $N = (2n)^d/2$ for $n~ \epsilon ~\mathbb{Z}$. Also plotted is the theoretical prediction of Luban and Michels (LM) for the equation of state [@EOSlubanmichels]. Curves for larger system sizes lie farther to the right. []{data-label="melt"}](melt5D.eps "fig:"){width="0.7\columnwidth"} The melting point was also estimated for $d=4$ (higher dimensions are presently too computationally demanding) by slowly densifying a system of spheres, initially a fluid, and looking for the onset of partial crystallization, again by monitoring the reduced pressure $p$ as a function of density $\phi$. Due to the difficulty of observing coexistence in finite systems and the relatively high activation barrier, simulated hard-sphere systems become “supercooled” (overcompressed) and nucleation does not occur until the melting density is surpassed. Consequently, the density at which partial crystallization appears for sufficiently slow expansion provides a reasonable estimate for the melting density. Near jamming the reduced pressure is asymptotically given by the free-volume equation of state [@phiJ], $$p = \frac{PV}{Nk_BT} = \frac{1}{\delta} = \frac{d}{1-\phi/\phi_J},$$ which can be inverted to give an estimate $\tilde \phi_J$ of the jamming density, $$\tilde \phi_J = \frac{\phi}{1-d/p}.$$ Since the pressure increases very rapidly near jamming, it is more convenient to plot the estimated jamming density $\tilde \phi_J (\phi)$ instead of the pressure $p(\phi)$, as shown in Fig. \[freeze\] for a system of $648$ spheres in $d=4$. In such a plot, the onset of partial crystallization causes a dramatic jump in $\tilde \phi_J (\phi)$, as the jamming density of the crystal is much higher than the jamming density of a disordered packing. The intersection of the curves with the line $\tilde \phi_J (\phi) = \phi$ gives the final jamming density. Sufficiently fast expansion suppresses crystallization and leads to packing fractions around $0.45-0.47$. Slower expansion allows for partial crystallization, typically around $\phi_M \simeq 0.38-0.39$, which is our rough estimate of the melting point, in agreement with our estimate from the results in Fig. \[melt\]. More accurate estimates can only be obtained using free-energy calculations. Since crystallization is a nucleated process, it is not surprising that the same expansion rates $\gamma$ can crystallize at different packing fractions and onto different crystal branches, [*e.g.*]{} $\gamma = 10^{-8}$ (a) and (b) in Fig. \[freeze\]. ![(Color online) Left panel: The estimated jamming packing fraction $\tilde \phi_J$ as a function of density $\phi$ for systems of $648$ spheres for $d=4$ with various expansion rates (see legend and note that there are two samples labeled (a) and (b) for $\gamma=10^{-8}$). For the curves showing no partial crystallization ([*i.e.*]{}, $\gamma = 10^{-5}$, $10^{-6}$, and $10^{-7}$), curves with smaller expansion rates have larger peak heights. For the curves that show partial crystallization ([*i.e.*]{} $\gamma = 10^{-8}$ (a and b) and $10^{-9}$), curves with smaller expansion rate lie farther to the left. Right panel: The cumulative coordination $Z(r)$ ([*i.e.,*]{} the number of contacts) for the perfect $D_4$ lattice and for the partially crystallized packings at $p>10^{12}$ obtained for expansion rates $\gamma = 10^{-8}$ and $\gamma = 10^{-9}$. The jamming packing fraction for the $\gamma = 10^{-8}$ packing is $\phi = 0.511$, and the jamming packing fraction for the $\gamma = 10^{-9}$ packing agreed up to 12 significant figures with the density of the $D_4$ lattice, $\phi = \pi^2/16 \simeq 0.617$. []{data-label="freeze"}](648Jnsmall.eps){width="0.95\columnwidth"} To determine whether the crystallized packings were forming a $D_4$ lattice, the conjectured densest packing in four dimensions, we computed the average cumulative coordination number $Z(r)$, which is the average number of sphere centers within a distance $r$ from a given sphere center. The inset to Fig. \[freeze\] shows $Z(r)$ for a perfect $D_4$ lattice and for the crystallized packings with $\gamma = 10^{-8}$ and $\gamma = 10^{-9}$ (corresponding colors represent the same packing). The sharp plateaus for the $D_4$ lattice correspond to the coordination shells and the number of spheres in the first shell is the kissing number $Z_{max} = 24$. The packing shown with $\gamma = 10^{-9}$ formed a perfect $D_4$ lattice. The packing shown with $\gamma = 10^{-8}$ partially crystallized with a final density of $\phi \simeq 0.511$. ![The estimated jamming packing fraction $\tilde \phi_J$ as a function of density $\phi$ for a system of $10,000$ spheres for $d=4$ with various expansion rates. Curves with smaller expansion rates have larger peak heights. The curve labeled “mix” corresponds to the following sequence of expansion rates: $\gamma = 10^{-2}$ until $p=10$, $\gamma = 10^{-3}$ until $p=10^4$, $\gamma = 10^{-4}$ until $p=10^6$, and $\gamma = 10^{-5}$ until $p=10^{12}$. []{data-label="10kfreeze"}](10k.eps){width="0.95\columnwidth"} Figure \[10kfreeze\] shows the estimated jamming packing fraction $\tilde \phi_J$, as in Fig. \[freeze\], but for a system of $10,000$ spheres, instead of $648$ spheres, in four dimensions. In contrast to the $648$ sphere system, there is no sign of partial crystallization for the $10,000$-sphere system. In fact, molecular dynamics was performed at packing fractions of $\phi \simeq 0.38-0.42$ for 10 million collisions per sphere and there was no significant drop in pressure indicative of partial crystallization. The curves in Figs. \[freeze\] and \[10kfreeze\] exhibit a bump around $\phi_G \simeq 0.41$, suggesting a kinetic transition from the fluid branch to a glassy branch. ![(Color online) The estimated jamming packing fraction $\tilde \phi_J$ as a function of packing fraction $\phi$ for $d=3$. Shown are systems of $4096$ spheres with various expansion rates and systems of $10,976$ spheres placed in an FCC lattice with negative expansion rates $\gamma = -10^{-4}$, $-10^{-5}$, and $-10^{-6}$ (last three curves). Also plotted are approximations to the equilibrium EOS for the fluid phase, the coexistence region, and the crystal phase [@speedy], as well as the Percus-Yevick (PY) EOS for the fluid phase. Compare this figure to the curves shown in Figs. \[freeze\] and \[10kfreeze\]. For the curves showing no partial crystallization ([*i.e.*]{}, $\gamma = 32\times10^{-6}$, $64\times10^{-6}$, and $128\times10^{-6}$), curves with smaller expansion rates have larger peak heights. For the curves that show partial crystallization ([*i.e.*]{}, $\gamma = 10^{-6}$, $4\times10^{-6}$, and $16\times10^{-6}$), curves with smaller expanion rates lie farther to the left. For the melting curves ([*i.e.*]{}, $\gamma = -10^{-4}$, $-10^{-5}$, and $-10^{-6}$), curves with smaller compression rates lie farther to the right. []{data-label="3D"}](3Dequil2.eps){width="0.95\columnwidth"} Figure \[3D\] shows the estimated jamming packing fraction $\tilde \phi_J$ for systems of spheres for $d=3$ with various positive and negative expansion rates, for comparison with the results for $d=4$ and $d=5$ in Figs. \[melt\], \[freeze\] and \[10kfreeze\]. The locations of the freezing and melting points in $d=3$ have been determined from free-energy calculations [@frenkel] and good approximations to the EOS for both the fluid and crystal phases are known [@speedy]. Our estimates of the freezing and melting points as the densities at the onset of melting of a diluted crystal or of partial crystallization of a densified fluid, respectively, compare favorably to the true values computed from free-energy calculations in $d=3$. The bump around $\phi_G \simeq 0.59$, analogous to the bump in Fig. \[10kfreeze\] around $\phi_G \simeq 0.41$, is often cited as the approximate location of the “kinetic” glass transition [@chaikin2]. Comparing Figs. \[10kfreeze\] and \[3D\] reveals that the melting point and suggested kinetic glass transition are closer for $d=4$ than for $d=3$, which is a possible reason why there is a lower tendency to crystallize for $d=4$ than for $d=3$. Similar results have been observed for binary hard disks, a model glass former [@binarydisk]. Disordered Jammed Packings ========================== Packings representative of the maximally random jammed (MRJ) state are produced by a combination of expansion rates. The expansion rate must be initially high (compared to the average thermal velocity) to suppress crystallization and produce disordered configurations that are trapped in the neighborhood of a jammed packing. Near the jamming point, the expansion rate must be sufficiently slow to allow for particle readjustments necessary for collective jamming. Figure \[10kfreeze\] shows the final jamming packing fractions of packings created using a variety of expansion rates, as the packing fraction at which the curves intersect the line $\tilde \phi_J=\phi$. We see that by increasing the expansion rate, we attain packings with lower jamming packing fractions. By comparing Fig. \[10kfreeze\] and to the analogous plot for a $d=3$ system (Fig. \[3D\]), where it is widely accepted that $\phi_{MRJ} \simeq 0.64-0.65$ [@rcp; @kansal], we estimate the MRJ density for $d=4$ to be $\phi_{MRJ} \simeq 0.460 \pm 0.005$. A more accurate calculation of $\phi_{MRJ}$ demands a better theoretical understanding of order metrics and how the expansion rate in the algorithm affects the ordering in the produced packings; statistical errors are smaller than the effect of the packing-generation protocol. Systematic investigation of different protocol parameters, as done for $d=4$ in Fig. \[10kfreeze\], is currently too computationally intensive in higher dimensions. Reasonable estimates of $\phi_{MRJ}$ for both $d=5$ and $d=6$ are obtained using the following less computationally intensive procedure. First, the system of spheres is expanded, starting from zero initial kinetic energy ($T=0$), until it reached a high pressure (say, $p=100-1000$). Then the system is slowly expanded ($\gamma = 10^{-5}-10^{-3}$) and periodically cooled to $k_BT = 1$ until a very high pressure (say, $p=10^{12}$) is attained. The resulting packings are approximately collectively jammed, as demonstrated by very large relaxation times for the pressure during long molecular dynamics runs [@aleks]. Using this method we estimate the MRJ density for $d=5$ to be $\phi_{MRJ} \simeq 0.310 \pm 0.005$ and for $d=6$ to be $\phi_{MRJ} \simeq 0.200 \pm 0.01$. The MRJ packing fractions as well as important equilibrium packing fractions are summarized in Table \[table\]. It is useful to compare the MRJ packings fractions for $3 \le d \le 6$ to recent estimates of the [*saturation*]{} packing fraction $\phi_s$ for the random sequential addition (RSA) packing of hard spheres obtained by Torquato, Uche and Stillinger [@To06] in corresponding dimensions, which were shown to be nearly hyperuniform [@hyperuniformity]. These authors found that $\phi_s=0.38278 \pm 0.000046, 0.25454 \pm 0.000091, 0.16102 \pm 0.000036$ and $0.09394 \pm 0.000048$ for $d=3,4,5$ and $6$, respectively. The nonequilibrium RSA packing is produced by randomly, irreversibly, and sequentially placing nonoverlapping spheres into a volume. As the process continues, it becomes more difficult to find available regions into which the spheres can be added. Eventually, in the saturation (infinite-time) limit, no further additions are possible, and the maximal achievable packing fraction is the saturation value $\phi_s$ \[see Ref. [@torquatobook] and references therein\]. As expected, the RSA saturation packing fraction in dimension $d$ is substantially smaller than the corresponding MRJ value because, unlike the latter packing, the particles cannot rearrange. -------------- ----------------------------------- --------------------------- -------------------------- -------------------------- Packing fraction $d=3$ $d=4$ $d=5$ $d=6$ $\phi_F$ $0.494$ [@torquatobook; @frenkel] $0.32 \pm 0.01^*$ $0.19 \pm 0.01^*$ - $\phi_M$ $0.545$ [@torquatobook; @frenkel] $0.39 \pm 0.01^*$ $0.24 \pm 0.01^*$ - $\phi_{MRJ}$ $0.645 \pm 0.005$ [@kansal] $0.46 \pm 0.005^*$ $0.31 \pm 0.005^*$ $0.20 \pm 0.01^*$ $\phi_{max}$ $0.7405 \ldots$ [@hales] $0.6169 \ldots$ [@conway] $0.4652 \dots$ [@conway] $0.3729 \dots$ [@conway] -------------- ----------------------------------- --------------------------- -------------------------- -------------------------- Our estimates for the MRJ packing fraction are compared to a theoretical formula proposed by Philipse [@philipse] for the “random jamming density” $\phi_d$, $$\phi_d \simeq \frac{0.046 d^2 + 1.22d + 0.73}{2^d}, \label{phil}$$ which predicts $\phi_3 \simeq 0.601$, $\phi_4 \simeq 0.397$, $\phi_5 \simeq 0.249$, and $\phi_6 \simeq 0.152$. It is seen that Eq. (\[phil\]) underestimates MRJ density $\phi_{MRJ}$ in $d=3$ and becomes worse with increasing dimension. Following Ref. [@To06], we obtain a better scaling form by noting that the product $2^d \phi_{MRJ}$ for $3 \le d \le 6$ is well approximated by a function linear, rather than quadratic, in $d$ (see Fig. \[phi-fit\]), i.e., the scaling form for $\phi_{MRJ}$ is given by $$\phi_{MRJ}= \frac{c_1}{2^d}+\frac{c_2 d}{2^d}, \label{linear}$$ where $c_1=-2.72$ and $c_2=2.56$. Although the scaling form (\[linear\]) applies only in low dimensions such that $ d \ge 3$, theoretical arguments given by Torquato, Uche and Stillinger [@To06] suggest that the general scaling form (\[linear\]) persists in the high-dimensional asymptotic limit, albeit with different coefficients $c_1$ and $c_2$. In Ref. [@To06], the density lower bound $\phi_{MRJ} \ge (d+2)/2^d$ is derived for MRJ packings in any dimension. This MRJ density lower bound yields 0.3125, 0.1875, 0.109375, 0.0625 for $d=3,4,5$ and 6, respectively. We note that Parisi and Zamponi [@Pa06] suggest the MRJ density scaling $\phi_{MRJ} \sim (d \log d) / 2^d$. ![(Color online) Fit of the data for the product $2^d\phi_{MRJ} $ to the linear form (\[linear\]) for $3 \le d \le 6$ with $c_1=-2.72$ and $c_2=2.56$. []{data-label="phi-fit"}](density-2d.eps){width="0.55\columnwidth"} Pair Correlations ----------------- Our main interest is pair correlations in the jamming limit in four, five and six dimensions. We characterize jammed packings statistically using the pair correlation function $g_2(r)$ and structure factor $S(k)$. The pair correlation function measures the probability of finding a sphere center at a given distance from the center of another sphere, normalized by the average number density $\rho$ to go asymptotically to unity at large $r$; [*i.e.*]{} $$g_2(r) = \frac{\langle P(r) \rangle}{\rho s_1(r)},$$ where $P(r)$ is the probability density for finding a sphere center a distance $r$ from an arbitrary sphere center, $\langle \rangle$ denotes an ensemble average, and $s_1(r)$ is the surface area of a single hypersphere of radius $r$ [@torquatobook]: $s_1(r) = 2\pi^2 r^3$ in $d=4$, $s_1(r) = 8\pi^2 r^4/3$ in $d=5$ and $s_1(r) =\pi^3r^5$ in $d=6$. The structure factor $$S(k) = 1 + \rho \hat h(k) \label{skdef}$$ is related to the Fourier transform of the total correlation function $h(r) = g_2(r)-1$. It measures spatial correlations at wavenumber $k$ and in particular, large-scale density fluctuations at $k=0$ [@hyperuniformity]. The structure factor can be observed directly via scattering experiments [@chaikin]. In the jamming limit, the pair correlation function $g_2(r)$ consists of a $\delta$-function due to sphere contacts and a background part $g_2^b(r)$ due to spheres not in contact: $$g_2(r) = \frac{\bar Z \delta(r-D)}{\rho s_1(D)} + g_2^b(r), \label{delta}$$ where $\bar Z$ is the average kissing number. Figure \[pcf\] compares the pair correlation function for jammed packings of $10^5$ spheres in $d=3$, $4$, $5$ and $6$. Due to periodic boundary conditions, $g_2(r)$ can only be calculated up to half the length of the simulation box, which limits the calculation to $r/D \simeq 3$ for $d=6$. The well-known split second peak present in $d=3$ is strongly diminished as the dimension increases, [*i.e.,*]{} the amplitude of the split second peak decreases and the sharp cusps become rounded with increasing dimension. The split third peak present in $d=3$ with considerable structure and two shoulders vanishes almost completely in the higher dimensions. The oscillations are strongly damped with increasing dimension and the period of oscillations might also decrease slightly with increasing dimension; this latter possibility is revealed more vividly in the structure factor through the shift in the location of the maximum, as we will describe below. The inset to Fig. \[pcf\] shows the magnitude of the decaying oscillations in $h(r)$ on a semi-log scale. Though at the values of $r/D$ shown, up to about half the length of the simulation box, there is still structure in addition to the oscillations, especially apparent for $d=3$, it appears that the decay rate of the oscillations in $h(r)$ does not change significantly with dimension, whereas the amplitude of oscillations does. However, further studies with larger $r$ and therefore larger systems are needed to obtain more quantitative results. ![The pair correlation function $g_2(r)$ for MRJ packings of $10^4$ hard spheres for $d=3,~4,~5$ and $6$ at the respective densities reported in Table \[table\]. Pair separation is plotted in units of the sphere diameter $D$. \[For $d=6$, $g_2(r)$ was only calculated up to $r/D = 3$ due to the system size and periodic boundary conditions\]. The delta-function contribution \[[*cf.*]{} Eq. \[delta\]\] at contact, of course, is not shown. The inset shows $|h(r)| = |g_2(r)-1|$ on a logarithmic scale for $d=3,~4$ and $5$. Each curve for $g_2(r)$ is obtained from a single packing realization (not time-averaged). Curves for higher dimensions are increasingly diminished. []{data-label="pcf"}](pcf6D.eps){width="0.95\columnwidth"} We calculate the structure factor $S(k)$, defined in Eq. \[skdef\], for $d=4$ and $d=5$ by $$S(K) = 1 + 128\phi \int_0^{\infty} x^3 h(x) \frac{J_1(Kx)}{Kx}dx \label{s4D}$$ and $$S(K) = 1 + 480 \phi \int_0^{\infty} \frac{x^4 h(x)}{(Kx)^2} \left[\frac{\sin(Kx)}{Kx} - \cos(Kx) \right] dx, \label{s5D}$$ respectively, where $\phi = \pi^2 \rho D^4/32$ for $d=4$ and $\phi = \pi^2 \rho D^5/60$ for $d=5$, $x=r/D$ and $K=kD$ are the dimensionless radius and wave number, and $J_{\nu}(x)$ is the Bessel function of order $\nu$. We do not calculate the structure factor for $d=6$ because at present we do not have $g_2(r)$ over a sufficiently large range of $r$. Following Ref. [@alekshyper], rather than working directly with $g_2(x)$ as in Eq. (\[skdef\]), we consider the average cumulative coordination $Z(x)$, defined to be the following volume integral of $g_2(x)$: $$Z(x) = \rho \int_1^x s_1(x') g_2(x') dx'. \label{Zintegralg}$$ The excess coordination $\Delta Z(x)$, $$\begin{aligned} \Delta Z(x) & = & 1 + 64 \phi \int_0^x (x')^3 h(x') dx' \\ \Delta Z(x) & = & 1 + 160 \phi \int_0^x (x')^4 h(x') dx',\end{aligned}$$ for $d=4$ and $d=5$, respectively, is the average excess number of sphere centers inside a spherical window of radius $x$ centered at a sphere, compared to the ideal gas expectations, $16 \phi x^4$ for $d=4$ and $32 \phi x^5$ in $d=5$. We can rewrite Eq. (\[skdef\]) in terms of $\Delta Z(x)$ using integration by parts to get $$S(K) = - 2 \int_0^{\infty} \Delta Z(x) \frac{d}{dx}\frac{J_1(K x)}{K x} dx$$ and $$S(K) = -3\int_0^{\infty} \Delta Z(x) \frac{d}{dx}\left[\frac{\sin(K x)}{(K x)^3} - \frac{\cos(K x)}{(K x)^2} \right] dx,$$ for $d=4$ and $d=5$, respectively. Note that accurate evaluations of the integrals of $\Delta Z(x)$ require extrapolations of its large-$x$ tail behavior, for which we have used an exponentially-damped oscillating function [@footnoteZ]. Figure \[sk\] shows $S(k)$ for jammed packings of $10^5$ spheres in three, four and five dimensions. Qualitatively, $S(k)$ is somewhat similar for $d=3$, $4$, and $5$. However, with increasing dimension, the height of the first peak of $S(k)$ decreases, the location of the first peak moves to smaller wavelengths, and the oscillations become damped. The width of the first peak also increases with increasing dimension, which could indicate that the correlation length decreases with increasing dimension. The inset to Fig. \[sk\] shows $S(k)$ for a jammed packing and a fluid near the freezing point in four dimensions. The relation between the structure factor for the fluid and jammed packing is strikingly similar to what is found for $d=3$, except that the peaks of both curves for $d=4$ appear scaled down relative to $d=3$. Overall, our results for both $g_2(r)$ and $S(k)$ are consistent with a recently proposed “decorrelation” principle [@decorrelation]. We note that similar pair decorrelations are observed for RSA packings as the dimension increases up to $d=6$ [@To06]. ![The structure factor $S(k)$ for jammed packings of $10^5$ spheres for $d=3$, $4$ and $5$ at the respective densities reported in Table \[table\]. Inset: A comparison for $d=4$ of $S(k)$ for a jammed packing and for a fluid near the freezing point ($\phi\approx 0.31$). Each curve for $S(k)$ is obtained from a single packing realization (not time averaged). []{data-label="sk"}](sk.eps){width="0.95\columnwidth"} It is of interest to determine whether infinite-wavelength density fluctuations $S(k=0)$ vanish; systems with this property are called “hyperuniform” [@hyperuniformity]. For equilibrium fluids and crystals, $S(k = 0)$ is proportional to the isothermal compressibility and therefore must be positive. As for $d=3$, $S(k)$ for $d=4$ appears to go to zero faster near the origin for the jammed packing than for the fluid. However, we cannot reliably determine whether $S(k)$ vanishes at the origin because our calculation of $S(k)$ for small $k$ involved an extrapolation of the large-$x$ tail of $\Delta Z(x)$. Nevertheless, using larger system sizes of one million spheres, saturated [@saturated] MRJ packings for $d=3$ have been shown to be hyperuniform to a high accuracy [@alekshyper] and the comparison of $d=4$ and $d=5$ to $d=3$, shown in Fig. \[sk\], suggests that MRJ packings for $d=4$ and $d=5$ are also hyperuniform. Isostaticity ------------ We study the near-contact contribution to $g_2(r)$, [*i.e.,*]{} interparticle distances $r$ that are very close to the sphere diameter $D$, using the cumulative coordination number $Z(x)$, where as before $x=r/D$ is the dimensionless radius and $x-1$ is the dimensionless interparticle gap. Figure \[logzpower\] shows $Z(x)$ for jammed packings of $10,000$ spheres for $d=4$ and $d=5$ with rattlers removed [@no6Disostaticity]. The plateaus at $Z=8$ in Fig. \[logzpower\] (a) and $Z=10$ in Fig. \[logzpower\] (b) show that both packings are isostatic. Isostatic packings are jammed packings which have the minimal number of contacts necessary for collective jamming. For spheres, this occurs when the number of degrees of freedom equals the number of contacts (or constraints); each $d$-dimensional sphere has $d$ degrees of freedom, and hence the mean number of contacts experienced by a sphere necessary for jamming is $2d$, since each contact involves two spheres. Packings produced by the LS algorithm almost always contain a nonzero fraction of “rattlers”, which are spheres trapped in a cage of jammed neighbors, but free to move within the cage. We find approximately $\sim 1\%$ rattlers for $d=4$ and $\sim 0.6\%$ rattlers for $d=5$, as compared to $\sim 2-3\%$ rattlers for $d=3$ [@aleks]. Rattlers can be identified as having less than the required $d+1$ contacts necessary for local jamming and are removed to study the jammed backbone of the packing, which we focus on in this section. The insets to Fig. \[logzpower\] (a) and (b) show $Z(x) - 2d$, along with a power-law fit for intermediate interparticle gap $x-1$, $$Z(x) = \bar Z + Z_0 (x-1)^{\alpha},$$ where $\bar Z = 2d$. Since the packings are generally slightly subisostatic, we apply a small correction ($<0.1\%$) to the isostatic prediction of $2d$ by using the midpoint of the apparent plateau in $Z(x)$. The best-fit exponent is $\alpha \simeq 0.6$ in both $d=4$ and $d=5$, in agreement with that found for $d=3$ [@aleks]. The coefficients of the power law, $Z_0 \simeq 11$ in $d=3$, $Z_0 \simeq 24$ for $d=4$, and $Z_0 \simeq 40$ for $d=5$ are close to the corresponding kissing numbers of the densest packings, $Z = 12$ for $d=3$, $Z = 24$ for $d=4$, $40 \leq Z \leq 46$ for $d=5$ and $72 \leq Z \leq 80$ for $d=6$. Motivated by this observation, we measured the value of the gap $x-1$ at which the cumulative coordination $Z(x)$ equals the kissing number of the densest packing to be: $x-1 \simeq 0.35$, $0.34$, $0.31-0.36$ and $0.33-0.36$ in $d=3$, $4$, $5$ and $6$, respectively, which we can define to be the cutoff for the near-neighbor shell. This definition produces results similar to that of the more common definition of the cutoff for the near-neighbor shell as the value of the gap $x-1$ at the first minimum in $g_2$, which occurs at $x-1 \simeq 0.35$, $0.32$, $0.30$ and $0.28$ in $d=3$, $4$, $5$ and $6$, respectively. It is also interesting to observe that the power-law fit to $Z(x)$ is good over a rather wide range of gaps, almost up to the first minimum in $g_2$. We should, however, emphasize that the minimum of $g_2$ is not very precisely defined, especially due to decorrelation in high dimensions, and the choice of the gap at the minimum of $g_2$, or at which $Z(x)$ equals the kissing number of the densest packing, as a special point is somewhat arbitrary and not theoretically justified at present. ![The near-contact cumulative coordination $Z(x)$ \[[*c.f.*]{} Eq. \[Zintegralg\]\] for $10^4$-sphere MRJ packings for $d=4$ (a) and for $d=5$ (b), with rattlers removed. The inset shows $Z(x)$ on a log-log scale along with power-law fits for intermediate interparticle gap $x-1$ beyond contact. $10^5$-sphere MRJ packings in $d=5$ with final expansion rates of $\gamma = 10^{-4}$ give similar results; such packings with final expansion rates of $\gamma = 10^{-5}$ are (presently) too computationally expensive. Compare these plots to the equivalent results for $d=3$ in Ref. [@aleks] \[[*c.f.*]{} Fig. 8\]. []{data-label="logzpower"}](logzwor4D.eps "fig:"){width="0.8\columnwidth"} ![The near-contact cumulative coordination $Z(x)$ \[[*c.f.*]{} Eq. \[Zintegralg\]\] for $10^4$-sphere MRJ packings for $d=4$ (a) and for $d=5$ (b), with rattlers removed. The inset shows $Z(x)$ on a log-log scale along with power-law fits for intermediate interparticle gap $x-1$ beyond contact. $10^5$-sphere MRJ packings in $d=5$ with final expansion rates of $\gamma = 10^{-4}$ give similar results; such packings with final expansion rates of $\gamma = 10^{-5}$ are (presently) too computationally expensive. Compare these plots to the equivalent results for $d=3$ in Ref. [@aleks] \[[*c.f.*]{} Fig. 8\]. []{data-label="logzpower"}](logzpower5D.eps "fig:"){width="0.8\columnwidth"} Discussion ========== We have presented the first numerical results characterizing random jammed hard-sphere packings in four, five and six dimensions. We find disordered packings, representative of the maximally random jammed state, to be isostatic and have packing fractions $\phi_{MRJ} \simeq 0.46$, $\phi_{MRJ} \simeq 0.31$ and $\phi_{MRJ} \simeq 0.20$ for $d=4$, $5$ and $6$, respectively. For equilibrium sphere packings, we estimate the freezing and melting packing fractions for the fluid-solid transition in four dimensions to be $\phi_F \simeq 0.32$ and $\phi_M \simeq 0.39$, respectively, and in five dimensions to be $\phi_F \simeq 0.19$ and $\phi_M \simeq 0.24$, respectively. Additionally, a signature characteristic of the kinetic glass transition is observed around $\phi_G \simeq 0.41$ for $d=4$. We observe a significantly lower tendency to crystallize for $d=4$ than in $d=3$, which is likely due to the closer proximity of the melting and kinetic glass transition densities for $d=4$ [@binarydisk]. We find that in high dimensions the split-second peak in the pair correlation function $g_2$, present for $d=3$, gets dramatically diminished and oscillations in both $g_2$ and the structure factor $S(k)$ get significantly dampened. These findings are consistent with a recently proposed “decorrelation principle” [@decorrelation], stating that unconstrained spatial correlations vanish asymptotically in the high-dimensional limit and that the $n$-particle correlation function $g_n$ for any $n \ge 3$ can be inferred entirely from a knowledge of the number density $\rho$ and the pair correlation function $g_2({\bf r})$. Accordingly, in this limit the pair correlation function $g_2(r)$ would be expected to retain the delta-function contribution from nearest-neighbor contacts, but the extra structure representing unconstrained spatial correlations beyond a single sphere diameter would vanish. Figures \[pcf\] and \[sk\] show dramatically the decorrelation principle already taking effect in four, five and six dimensions. We note that decorrelation principle is also apparent in the same dimensions for RSA packings [@To06]. One should not be misled to believe that the decorrelation principle is an expected “mean-field" behavior. For example, it is well known that in some spin systems correlations vanish in the limit $d \rightarrow \infty$ and the system approaches the mean-field behavior. While this idea has meaning for spin systems with attractive interactions, hard-core systems, whose total potential energy is either zero or infinite, cannot be characterized by a mean field. Mean-field theories are limited to equilibrium considerations, and thus do not distinguish between “constrained" and “unconstrained" correlations because, unlike us, they are not concerned with non-equilibrium packings of which there an infinite number of distinct ensembles. The decorrelation principle is a statement about any disordered packing, equilibrium or not. For example, contact delta functions are an important attribute of non-equilibrium jammed disordered packings and have no analog in equilibrium lattice models of any dimension. The decorrelation principle is also justified on the basis of a rigorous upper bound on the maximal packing density in high dimensions [@decorrelation], which has no counterpart in mean-field theories. A particularly interesting property of jammed hard-sphere packings is hyperuniformity, the complete suppression of infinite wavelength density fluctuations, [*i.e.,*]{} the vanishing of the structure factor $S(k)$ as $k\rightarrow 0$. It has been recently conjectured that all saturated strictly-jammed packings are hyperuniform [@hyperuniformity] and calculations of the structure factor near $k = 0$ for $d=3$ using one million particle systems have strongly suggested that MRJ packings for $d=3$ are indeed hyperuniform [@alekshyper]. Though the system sizes used in this paper were too small to probe such large-scale density fluctuations without relying on dubious extrapolations, our numerical results for the structure factor for $d=4$ and $d=5$, as shown in Fig. \[sk\], are consistent with hyperuniformity. As in three dimensions, disordered jammed sphere packings show no signs of crystallization, are isostatic, and have a power-law divergence in $g_{2}(r)$ at contact. Interestingly, all three dimensions ($3$, $4$ and $5$) share the same power law exponent $1-\alpha \simeq 0.4$ when rattlers are removed, and show the first minimum of $g_{2}(r)$ close to where the cumulative coordination $Z(r)$ equals the kissing number of the densest lattice packing. Such a relation between the kissing numbers of the densest packings and MRJ packings for $d=3$, $4$, $5$ and $6$, if not coincidental, is very surprising and may be a consequence of the geometrical structure of MRJ packings. It suggests that disordered packings might be deformed crystal packings, in which the true contacts are deformed into near contacts, and only the minimal number of contacts necessary for jamming is preserved. This interpretation is to be contrasted with the usual interpretation of disordered packing in $d=3$ in terms of tetrahedral or icosahedral packings, without relation to the crystal (FCC) packing. The former interpretation is similar to the one of the MRJ state for binary hard disks as a random partitioning of the monodisperse triangular crystal into “small” and “large” disks, i.e., a deformed monodisperse triangular disk crystal in which a randomly chosen fraction of the particles have grown in size, as proposed in Ref. [@binarydisk]. It is important to point out that hard-sphere packings behave rather differently in two dimensions than in three and higher dimensions. For $d=2$, jammed hard-sphere systems are polycrystalline and there is a very weak, nearly continuous fluid-solid phase transition. Hence, there is no glassy behavior for $d=2$ and consequently no amorphous jammed packings. Glassy behavior, due to geometrical frustration arising from the inconsistency of local optimal packing rules and global packing constraints, first appears in three dimensions [@torquatobook]. It is likely that geometrical frustration generally increases with dimension, consistent with our observation that nucleation is suppressed with increasing dimension. Computational costs rise dramatically with increasing dimension and theoretical understanding based on observations in moderate dimensions is necessary. We believe that the numerical results presented in this work provide tests and motivations for such theories. M. S. was supported by the National Science Foundation and A. D. and S. T. were partially supported by the National Science Foundation under Grant No. DMS-0312067. [44]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , ****, (). , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ** (, , ). , ****, (). , ****, (). , ****, (). , , , ****, (); , , , ****, (); M. Bishop, A. Masters and A. Yu. Vlasov, J. Chem. Phys. [**122**]{}, 154502 (2005); M. Bishop and P.A. Whitlock, J. Chem. Phys. [**123**]{}, 014507 (2005). , ****, (); ibid., 1361 (2004); ibid., [**122**]{}, 15 (2006); N. Clisby and B.M. McCoy, J. Phys. [**64**]{}, 775 (2005). I. Lyberg, J. Stat. Phys. [**119**]{}, 747 (2005). , ****, (). , ****, (). , ****, (). , , , ****, (). , ** (, , ). , ****, (). , ****, (). , p. (). , (). , ****, (). S. Torquato, O. U. Uche and F. H. Stillinger, in preparation. , ** (, , ). L. Lue, J. Chem. Phys. [**122**]{}, 044513 (2005). , , , ****, (). , ****, (). , ****, (), , . G. D. Scott and D. M. Kilgour, [*Brit. J. Appl. Phys.*]{} [**2**]{} 863 (1969). , ****, (). , , , ****, (). , , , , ****, (). Ref. [@torquato15] presents the first exactly solvable disordered sphere-packing model (“ghost random sequential addition packing) in arbitrary space dimension. Specifically, it was shown that all of the $n$-particle correlation functions of this nonequilibrium model can be obtained analytically for all allowable densities and in any dimension. It provides an exact demonstration of the decorrelation principle; see also Ref. [@decorrelation] , ****, (). , , , ****, (). , ****, (). , , , ****, (). , ** (, ). , ****, (). , ****, (). , ** (, ), chap. , pp. . The specific extrapolation function we use for both $d=4$ and $d=5$ is of the form $$\Delta Z(x) = a_1 x e^{-a_2 x} \cos{(a_3 x + a_4)},$$ where $a_1, a_2, a_3,$ and $a_4$ are fitting parameters.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Motivated by specific connections to dark matter signatures, we study the prospects of observing the presence of a relatively light gluino whose mass is in the range $\sim (500-900)~\rm GeV$ with a wino-like lightest supersymmetric particle with mass in the range of $\sim (170-210)~\rm GeV$. The light gaugino spectra studied here is generally different from other models, and in particular those with a wino dominated LSP, in that here the gluinos can be significantly lighter. The positron excess reported by [the]{} PAMELA satellite data is accounted for by annihilations of the wino LSP and their relic abundance can generally be brought near the WMAP constraints due to the late decay of a modulus field re-populating the density of relic dark matter. We also mention the recent [FERMI photon]{} constraints on annihilating dark matter in this class of models and implications for direct detection experiments including CDMS and XENON. We study these signatures in models of supersymmetry with non-minimal soft breaking terms derived from both string compactifications and related supergravity models which generally lead to non-universal gaugino masses. At the LHC, large event rates from the three-body decays of the gluino in certain parts of the parameter space are found to give rise to early discovery prospects for the gaugino sector. Excess events at the 5 sigma level can arise with luminosity as low as $\mathcal{O}$(100) $\rm pb^{-1}$ at a center of mass energy of 10 TeV and $\lesssim \mathcal{O}$(1) $\rm fb^{-1}$ at $\sqrt{s} = 7 ~{\rm TeV}$.' address: | $^{a}$Michigan Center for Theoretical Physics,\ University of Michigan, Ann Arbor, MI 48109, USA\ $^{b}$Northeastern University, Dept. of Physics, Boston, MA, 02115, USA author: - 'Daniel Feldman$^{a}$, Gordon Kane$^{a}$, Ran Lu$^{a}$, and Brent D. Nelson$^{b}$' title: Dark Matter as a Guide Toward a Light Gluino at the LHC --- Introduction ============ Production of superpartners at the Large Hadron Collider (LHC), the event rates observed in the recent cosmic ray data [@PAM; @FERMI], and potentially in dark matter direct detection experiments [@DD], all may be linked to the composition of dark matter. A well motivated class of candidate models that can be tested on all of these fronts arises when the dark matter is the lightest R-parity odd supersymmetric particle (LSP) and has a substantial wino component. However, even if the LSP is produced at the LHC at a relatively light mass, a discoverable signal at the LHC may be difficult unless colored superpartners such as the gluino, are light. There are few top down models which generically imply an LSP that is dominantly wino with a light gluino. The breaking of supersymmetry (SUSY) through a pure anomaly mediated contribution does predict an LSP that is a wino, but needs an extension to provide a consistent model. String frameworks of interest which give rise to a light gluino and predict a wino LSP include those based on the fluxless sector of $G_2$ compactifications from which a realistic model of soft SUSY breaking has been constructed [@Acharya:2007rc]. Related models of soft breaking based on the heterotic string can also yield a light gluino with a wino-like LSP [@Kane:2002qp; @Gaillard:2007jr], and a light gluino/wino-like LSP system can also arise from gaugino mass non-universality in D-brane models of soft SUSY breaking [@Corsetti:2000yq]. That a wino-like LSP is consistent with the PAMELA satellite data with a mass of order 200 GeV [has been emphasized]{} [@Pierce; @Hisano:2008ti; @Kane:2009if; @Feldman:2009wv], and recent works have begun to study the implications of a light wino with a correspondingly light gluino [@Feldman:2009wv; @NUWISC]. Generally, for a pure wino, or wino-like dark matter candidate, the predictions on the relic abundance can be in the vicinity of the WMAP data [@WMAP]. Relatively light wino-like dark matter can produce the correct relic density and have a thermal history provided it contains non-negligible bino and Higgsino components in extended theories [@Feldman:2009wv], while the pure wino can do so in a non-thermal paradigm due to the decay of heavy moduli [@MoroiRandall; @Acharya:2008bk] (for recent related work see [@Endo; @GelminiGondolo; @Acharya:2009zt]). The heavy moduli can add large additional post freeze-out entropy to the primordial particle density from the moduli decay into the SUSY sector. This decay also leads to a release of winos which annihilate at a temperature much lower than freeze-out. For the class of models we are interested in here, the gluino mass is rather light, as dictated by the soft breaking of supersymmetry and electroweak symmetry breaking, leading to large gluino production cross sections with subsequent decays of the gluinos via three-body decay chains. Thus a prominent LHC signal arises from multijet production. Some early SUSY discovery prospects in multijets at the LHC over a broad class of models have been given in [@Hubisz:2008gg][@Randall:2008rw] (for reviews see [@Kane:2008zz],[@BSMLHC]). Distinctively, here we emphasize well motivated models that yield dark matter annihilation cross sections consistent with the recent PAMELA data, and also lead to a spectrum with a light gluino. The associated production of gluinos and a wino-like LSP lead to a simultaneous probe of supersymmetry at colliders and in present dark matter experiments, where the gluino is linked to the chargino and neutralino through its dominant three body decay channels. The analysis of electroweak gauginos at colliders with mass degeneracy between the LSP and chargino has been studied in great detail (for early work see [@Chen:1996ap] and for a recent analysis see [@Buckley:2009kv]) where the soft decays of the chargino can lead to a wino LSP and a charged pion, giving rise to a displaced vertex of a track length of a few centimeters. The organization of this paper is a follows: In Section \[g2\] we briefly review a soft breaking sector of interest which gives rise to a light wino and a light gluino and serves to illustrate the effects of the expected high jet multiplicity from the production of gluinos at the LHC. We then discuss the numerical simulations that allow us to make contact between the theory and the data and enable a connection between the predictions for the LHC and to possible signals of dark matter. Following this, in Section \[lhcmodels\] we analyze the early discovery prospects of such models at the LHC for benchmark models and also for a large collection of models. The above is all carried out in the framework of the $G_2$ models with a pure wino LSP and a light gluino. In Section \[GeneralPredictions\] we examine a larger class of models in the context of relic density and direct and indirect detection of dark matter. We include models which deviate from a pure wino, but still have a substantial wino component. Here, as before, the soft breaking of supersymmetry and radiative electroweak symmetry breaking dictate the mass of relatively light gluino in the models of interest. We optimistically conclude in Section \[conc\]. Soft Breaking with Tree and Anomalous Contributions \[g2\] ========================================================== General Framework ----------------- The underlying framework we work in is described by ${\cal N}=1$ supergravity. We first consider the $G_2-\rm MSSM$ [@Acharya:2007rc] which has a generalized sector of soft SUSY breaking derived from both [@Gaillard] a tree level supergravity contribution and an anomalous contribution.[^1] The soft parameters can be parametrized at the unification scale as $m_0 = s \cdot m_{3/2}$, $M_a= f_a \cdot m_{3/2}$, $A_{3} = a_{3}\cdot m_{3/2} $ where $m_{3/2} \sim {\cal O}(10 -100) ~{\rm TeV}$ is the gravitino mass, $m_0$ is a universal scalar mass, $M_a$ are the gaugino masses, and $A_{3}$ are the tri-linear couplings of the third generation. Here the parameters $(s,f_a,a_3)$ are functions of the microscopic theory which are determined entirely from the effective supergravity model. The soft parameters are well approximated by (for the complete analytical expressions see [@Acharya:2007rc]) $s \sim 1$, $f_a = f'_a \alpha_G -\epsilon ~\eta$, where $f'_{1,2,3} =(0.35, 0.58, 0.64)$, and $\eta = 1 -\alpha_G \delta$ parametrizes gauge coupling corrections in the tree level sector of the gaugino masses. The parameter $\epsilon \sim (0.02 -0.03) $ arises as a consequence of the hidden sector potential which is responsible for tuning the cosmological constant to zero. The terms entering for the tri-linears of the third generation are well approximated by $a_3 \sim (3/2) m_{3/2}$ up to small corrections in the normalized Yukawas and the normalized volume $V_7$ of the $G_2$ manifold, the latter of which enters in the determination of the gravitino mass. The ratio of the Higgs vacuum expectation values is generically in the range $\tan \beta \sim 1.5-2.0$ as $\mu$ and $B$ are both taken to arise from the quadratic term in the Kahler potential, and are similar in magnitude. Here the largeness of the gravitino mass decouples the scalars while the gaugino masses are suppressed relative to the gravitino mass, where the suppression enters via the volume of hidden sector three cycles. The physical values of the soft parameters are sensitive to the precise value of the unified gauge coupling and threshold corrections. The largeness of the gravitino mass generically drives the $\mu$ term to be order $m_{3/2}$ for electroweak symmetry breaking, which in turn induces a relatively large self energy correction [@PBMZ] to the electroweak gaugino masses [@Acharya:2007rc]. The models studied in reference [@Acharya:2007rc] did not have a solution of the $\mu$ problem, but simply assumed that $\mu$ and the associated soft breaking term $B \mu$ arose from the quadratic term in the Kahler potential. When the issues of embedding the Standard Model in the $G_2$ manifold are fully understood it could happen that $\mu$ is small by symmetry arguments, e.g. U(1) charges force the bilinear term to vanish. Then if $\mu$ was of the same order as $M_2$ there would be some higgsino mixture in the $G_2$ wino, so the theories would have a wino-like LSP instead of a pure wino one. In this paper we use “$G_2"$ to refer to the pure wino case of reference [@Acharya:2007rc]. If the theory developed to imply a wino-like LSP the results for describing the PAMELA data and LHC predictions would change very little; however predictions on dark matter direct detection would be considerably modified. Analysis \[sim\] ---------------- In this work we have implemented the complete analytical expressions for soft breaking terms of the $G_2$-MSSM into [SOFTSUSY]{}[@ben]. The analysis includes the gaugino mass threshold corrections [@PBMZ] with 2 loop scalar corrections, 2-loop RGEs for the Higgs and gaugino masses, $\mu$, and Yukawa and gauge couplings [@Martin:1993zk; @ben]. Branching fractions have been computed with [SUSYHIT]{} [@susyhit] and production of signal and backgrounds are generated with [PYTHIA ]{} [@PYTHIA] and [PGS]{} [@PGS] with the level 1 (L1) triggers designed to efficiently reproduce CMS specifications [@Ball:2007zza] (for detailed discussions see e.g. [@FLN]). Signal and background have been simulated at $\sqrt{s} = (7,10,14)~~ \rm TeV$ in order to generalize our predictions for preliminary LHC runs and future operational center of mass energies. Specifically, SM backgrounds have been generated with QCD multi-jet production due to light quark flavors, heavy flavor jets ($b \bar b$, $t \bar t$), Drell-Yan, single $Z/W$ production in association with quarks and gluons ($Z$+ jets / $W$+ jets), and $ZZ$, $WZ$, $WW$ pair production resulting in multi-leptonic backgrounds. Laboriously, samples were generated at $\sqrt{s} = (7,10,14) \rm TeV$ with up to 5 $\rm fb^{-1}$ of luminosity. In [PGS4]{} jets are defined through a cluster-based algorithm which has a heavy flavor tagging efficiency based on the parametrizations of the CDF Run II tight/loose (secondary) vertex b-tagging algorithm [@CDF2]. The standard criteria for the discovery limit of new signals is that the SUSY signals should exceed either $5\sqrt{N_{\rm SM}}$ or 10 whichever is larger, i.e., ${N^c_{\rm SUSY}}>{\rm Max}\left\{5\sqrt{N^c_{\rm SM}},10\right\}$, were $c$ indicates the channel of interest. The signature space of the models we probe has distinctive dark matter predictions. The models we consider are dominated by dark matter annihilations into $W^{+}W^{-}$ and can yield a significant flux of cosmic antimatter in the galactic halo (for early work see [@sa1; @sa2; @sa3; @sa4]). The annihilation cross section receives an enhancement relative to other SUSY modes since it is s-wave and has a relative strength dictated by the SU(2) gauge coupling and the wino component of the LSP. The models are made consistent with the relic density constraints as will be discussed. For the analysis of dark matter annihilation cross sections and their resultant fluxes we employ fragmentation functions from [DarkSUSY ]{}[@DS] using [PYTHIA.]{} In this work we also model cosmic fluxes with [GALPROP v50.1p]{} [@Strong:1998pw]. Early Discovery Prospects and Concrete Signatures \[lhcmodels\] =============================================================== In theories with wino LSPs, the dominant LHC production modes are not strictly those from strongly produced SUSY. The production modes of the wino ($\w$) and the lightest chargino ($\widetilde{C}_1$) are competitive with the gluino ($\g$) production and frequently are larger. However due to the small splittings (a fraction of a GeV) between the wino and chargino the decay products here are soft. [Except]{} for larger gluino masses, we find that most events that pass the triggers do indeed come from $\g \g$ production, though as much as 30% of the events come from electroweak production. Thus the dominant production modes are $pp \to [(\g \g), (\w \widetilde{C}_1), (\widetilde{C}^{\pm}_1,\widetilde{C}^{\mp}_1)]$. The decays modes lead to rich jet and missing [energy]{} signatures with a sizeable number of leptons in the final state. In particular the dominant decays are as follows: $\g \to [(\widetilde{N}_2 t \bar t),(\w b \bar b),(\w q \bar q), (\widetilde{C}^{-}_1 {\bar b} t + ~{\rm h.c.}), (\widetilde{C}^{-}_1 {\bar d} u + ~{\rm h.c.)} ]$ with secondary decays $ \widetilde{N}_2 \to \widetilde{C}_1 W^{*} \to (\widetilde{C}_1 l \nu_l), (\widetilde{C}_1 q \bar q')$ and $\widetilde{C}_1 \to \w W^{*} \to (\w l \nu_l),(\w q \bar q') $ with tertiary branchings of the produced standard model particles $t \to W b$ and $W \to [(q {\bar q'}),(l \nu_l)]$. The models are rather predictive as they typically require no more than 2-3 branchings to complete SUSY cascades resulting in lepton and jet signatures. While this is a typical signature of SUSY in a generic model, it is actually a prediction of the wino branch of the $G_2$ model as electroweak symmetry breaking corners the viable parameter space and thus the viable signature space. The decays of $\tilde{C}_1 \to \w$ and their jet and lepton by-products will be very soft yet there can be radiation of gluon from the initial or final state partons that can generate a relatively hard jet. Thus one can look for a hard monojet and n-jet events with large missing energy as an early indication of the production of supersymmetric events at the LHC. In Table (\[parbench\]) we illustrate some typical spectra found in the $G_2$ models for [$m_{\w} \sim (170-190)~\rm GeV$]{} (precisely in the mass range pointed to by the recent PAMELA data \[see Sec.(\[positrons\])\] along with the dominant branching ratio of the gluino [given]{} in [Table (\[branch\]).]{} $G^{m}_2$ $m_{3/2}~(\rm{TeV})$ $\delta$ $V_7$ $\tan\beta$ $m_{ \tilde g}$ $m_{ \tilde W}$ $m_{ \tilde{C}^{\pm}_1}$ $m_{ \tilde{N}_2}$ ----------- ---------------------- ---------- ------- ------------- ----------------- ----------------- -------------------------- -------------------- -- -- -- -- $G^1_2$ 38.950 -2.9 30.0 1.98 551 170.2 170.4 260 $G^2_2$ 21.186 -10.0 33.0 1.41 717 173.4 174.0 190 $G^3_2$ 20.700 -9.3 36.0 1.57 652 176.0 176.5 185 $G^4_2$ 20.618 -9.1 30.0 1.71 632 180.9 181.3 185 $G^5_2$ 35.492 -5.4 32.0 1.54 761 190.5 190.6 263 : Benchmark models predicting a light gluino and a LSP that is a wino with a degenerate chargino with a light second neutralino (which is mostly bino). The last four columns carry units of GeV. \[parbench\] For the $G_2$ models, a central prediction is a relatively light gluino over the range of wino mass that is capable of describing the PAMELA data [as is illustrated in Table (\[parbench\]).]{} ${\mathcal Br}({\tilde g} \to X)$ $G^1_2$ $G^2_2$ $G^3_2$ $G^4_2$ $G^5_2$ ------------------------------------------------------------ --------- --------- --------- --------- --------- -- -- -- -- -- -- -- ${\tilde g}\to b \overline b \w $ 14.2 6.4 10.2 11.3 19.5 ${\tilde g}\to q \overline q \w $ 21.0 7.4 12.6 14.6 10.0 ${\tilde g} \to t \overline t {\tilde N}_2$ - 47.6 21.8 14.5 14.6 $ {\tilde g} \to t \overline b \tilde{C}^{-} + h.c.$ 18.9 16.2 20.8 20.9 24.6 ${\tilde g}\to q_u {\overline q}_d \tilde{C}^{-} + h.c $ 41.5 14.6 25.2 29.0 24.9 : Dominant branching ratios of the gluinos. \[branch\] $G^{m}_2$ $\sigma(\g \g)$ (fb) $\sigma(\w \widetilde{C}_1)$ (fb) $\sigma(\widetilde{C}^{\pm}_1 \widetilde{C}^{\mp}_1)$ (fb) $\sigma_{\rm SUSY }$ (fb) $\sigma_{\rm eff} $ (fb) $N(4j)$ $\frac{N}{\sqrt{B}}|_{4j}$ $N(2b)$ $\frac{N}{\sqrt{B}}|_{2b}$ ----------- ---------------------- ----------------------------------- ------------------------------------------------------------ --------------------------- -------------------------- --------- ---------------------------- --------- ---------------------------- -- -- -- $G^1_1$ 1613 996 301 2910 1645 416 13.3 37 4.7 $G^2_2$ 236 970 277 1484 353 79 2.5 22 2.8 $G^2_3$ 481 903 280 1665 553 133 4.2 37 4.7 $G^2_4$ 648 877 246 1773 736 217 7.0 32 4.1 $G^2_5$ 182 696 208 1087 250 64 2.0 10 1.2 \[lhcdata\] In [Figure (\[fig1\])]{} (left upper panel) one observes that the models can produce detectable multi-jet signals even at $\sqrt{s}= 10$ TeV for $\mathcal L \sim 1 \rm fb^{-1}$ of integrated luminosity under the standard $5\sigma$ discovery reach criteria in the kinematic variable $M^{4jets}_{eff} = \sum_{J={1-4}} P^J_T(J) + P^{\rm miss}_T$. In [Figure (\[fig1\])]{} (right upper panel) we show the large number of multijet signals. The analysis shows that the model can produce a large excess in hadronic jets over the backgrounds. The large jet multiplicity arises from the three body decay of the gluinos and from jets arising from initial state radiation. We find the discovery limit is optimal for 4-5 jets with a lepton veto and large missing energy cut. The lower right panel exhibits $M^{2b}_{eff} = \sum_{J={1-2}} P^b_T(J) + P^{\rm miss}_T$ with larger luminosity. The lower left panel shows the discovery reach for the same model with $\sqrt{s} = (7,10,14) ~\rm TeV$ and $5\sigma$ can be reached with several hundred inverse picobarns of data. Global Analysis and Discovery Prospects of Early SUSY \[early\] --------------------------------------------------------------- Having established that the highly constrained, and therefore predictive $G_2$ model can give rise to detectable signals of SUSY with early LHC data (see also [@sekmen]), we now extend the analysis to a larger region of the $G_2$ parameter space rather than focusing on a benchmark model. We have performed a detailed scan of the parameter space of these models over the parameters discussed in Section \[g2\], consistent with radiative electroweak symmetry breaking subject to the constraint that the wino mass is in the range (170 - 210) GeV . We uncover a large parameter space where the gluino can be relatively light in the $G_2$ model. The majority of the models have a gluino in the mass range of 500 to 900 GeV (see Fig.(\[figrat\]) [for the corresponding gaugino mass ratios]{}). LHC predictions with light gluino have been studied recently [@Kuflik; @gnlsp; @Bhattacharya:2009ij; @Baer], but without considering the connection to the PAMELA data, which we pursue in the next section. In [Table (\[lhcdata\])]{} we display the relatively large total theoretical production cross section before cuts ($\sigma_{\rm SUSY}$ from gluino, neutralino, chargino production) and the effective SUSY cross section $\sigma_{\rm eff}$ (cross section after the L1 triggers have been passed). One observes that the L1 triggers are [well]{} optimized for these events as a large fraction of the SUSY cross section is maintained. The substantial missing energy arises in many of the models from the prompt branching of the gluino into 2 jets and the LSP wino. Event rates at the LHC are shown in the 4-jet channel and the 2b channel with just 1 fb$^{-1}$ of integrated luminosity at $\sqrt{s}= 10$ TeV along with the ratio of the signal to the square root of the background. These models can be discovered very early with the LHC and can begin to be probed at $\sqrt{s} = {\rm 7~TeV}.$ [ Figure ([\[fig3\]]{}) ]{} displays the effective SUSY production cross section after cuts ($\sigma_{\rm eff}$) as a function of gluino mass at various center of mass energies. The (shaded) colored regions are the necessary luminosity need for a $5 \sigma$ excess in steps of 200 pb$^{-1}$ of integrated luminosity where we require at least 5 jets and large missing energy $\geq 200$ GeV. Thus it is apparent from the analysis that nearly all the models can produce discoverable signals with low luminosity. Remarkably, we find that with only $\mathcal{O}(100\,{\rm pb}^{-1})$ of data at $\sqrt s= $ 10 TeV, the models will produce large jet-based signals which can be discovered over the SM backgrounds over a part of the parameter space, even for gluinos as light as 550 GeV with $\mathcal L \sim 500 ~\rm pb ^{-1}$ at $\sqrt{s}= 7$ TeV. Models with wino-like LSPs, and thus nearly degenerate charginos and neutralinos, are well known to be be difficult to study [@track]. The chargino lifetime can be of order a centimeter, and the second heavier neutralino can even have order tens of GeV splitting (see Table (\[parbench\]) for such theory motivated examples). Once a set of gluino candidates have been identified, an off-line analysis focused towards the study of the chargino and neutralino states in the gluino decay products will be necessary. General Implications of a Wino-Like LSP \[GeneralPredictions\] =============================================================== In this section we relax the tight constraints of the $G_2$ theory space and explore the possibility of an LSP which has a significant wino component (“wino-like"), but may also have non-negligible bino and Higgsino components. One natural class of models where such an LSP is achieved are in grand unified models such as $\rm SU(5)$, $\rm SO(10)$, and $\rm E_6$ where the GUT symmetry is broken by a non-singlet $F$ term leading to gaugino masses at the unification scale that are non-universal, i.e., $M_a = m_{1/2}(1+ \Delta_a)$, $a = 1,2,3$. Such soft breaking mass terms can give rise to a wino-like LSP with a light gluino if the high scale values of the gaugino masses, $M_2$ and $M_3$, are reduced relative to $M_1$. Relic Abundance of a Wino-Like LSP ---------------------------------- In a general setting, the relic density can be equal to the observed one with a wino-like or pure wino LSP due to the late decay of a modulus field. Such is possible in a universe that has a non-thermal cosmological history [@MoroiRandall]. Thus, for a single heavy modulus field $\Phi$, in the so-called instantaneous decay approximation one obtains a reheat temperature, $T_R$, due to the decay $\Gamma_{\Phi}$ by assuming all energy density of $\Phi$ is transferred into radiation. The modulus decays after freeze-out and the reheat temperature is $T_R = C^{1/4} \sqrt{{\overline M}_{\rm pl} \Gamma_{\Phi}}$, $C =90/(\pi^2 g_*(T_R))$. Here $\Gamma_{\Phi} = c_{\Phi} M^3_{\Phi}/{\Lambda^2}$ , $c_{\Phi} \sim 1$, where $ \Lambda \simeq {\overline M}_{\rm pl} \equiv {\overline M}_{\rm pl}/\alpha $, where $\alpha$ parametrizes deviations from the Planck scale (moduli couplings at (much lower) intermediate scales have been considered in [@Khalil:2002mu; @Conlon:2007gk]). For example, $\alpha = \sqrt{V_7}$ gives $\Lambda = {\overline M}_{\rm pl}/\sqrt{V_7} \sim (2-4) \times 10^{17} {\rm GeV}$ which may be interpreted as an effective string scale. Under this assumption of non-thermal (NT) production one has $\Omega_{\tilde W} \simeq \Omega_{\rm T}|_{T_R}$, where $\Omega_{\rm T} h^2$ can be computed in the usual manner (see i.e. [@Griest:1990kh]). For the s-wave dominated LSP [interaction]{}, we obtain $\Omega_{\tilde W} h^2 \simeq 0.32 \frac{1}{\alpha \sqrt{c_{\Phi}}} (\frac{3*10^{-7} {\rm GeV}^{-2}}{<\sigma v>})( \frac{m_{\tilde W}}{200 {\rm GeV}})(\frac{m_{3/2}}{100 {\rm TeV}})^{-3/2}$, where we have used $m_{\Phi} \lesssim 2 m_{3/2}$, and where $\w \equiv \widetilde W$. A saturation of the error corridor from the WMAP constraint on $\Omega h^2$ is then possible for a gravitino in the mass range $(40 -60)$ TeV. In the $G_2$ models specific calculations of the relic abundance from moduli decay have been carried out [@Acharya:2008bk] giving a relic density, from a string based construction, a few times larger than the experimental value unless the gravitino mass is order 100 TeV. In greater generality, the nature of soft breaking and the cosmological history of the universe may very well be closely tied together [@Acharya:2009zt]. On the other hand, in a non-thermal framework one can also approach the WMAP constraint so long as $T_R$ does not spoil BBN constraints [@GelminiGondolo]. In a thermal paradigm the relic abundance of a wino-like LSP can also be brought in accord with the WMAP data in the presence of residual Abelian gauge factors that survive down to the SUSY scale and mix weakly with the MSSM neutralinos leading to a co-annihilation enhancement [@Feldman:2009wv] in an otherwise depleted relic abundance from the large annihilations of the LSP. This is to be contrasted with enhancements in the halo cross section, i.e. through a Sommerfeld enhancement [@Hisano:2004ds] or through a Breit-Wigner enhancement [@Feldman:2008xs] or a boost in the flux via dark matter clumps [@Bergstrom:1998jj; @Hooper:2003ad; @Hooper:2008kv]. Thus predictions on the relic density consistent with the production of positrons in the halo are rather model dependant, but nevertheless can account for the proper relic abundance of dark matter in such models. Connection to the Positron Data \[positrons\] --------------------------------------------- The data released by the PAMELA collaboration indicates a large excess in positron flux in the halo. For the case of models with MSSM field content, annihilations of the LSP into $W$ bosons are dominant possible sources of positrons and indeed the $W^+W^-$ production provides the needed cross section in the halo to account for the PAMELA anomaly for a pure wino [@Kane:2009if][@Feldman:2009wv] without any boost factor in the positron flux ($\langle \sigma v \rangle \sim 2.5 \times 10^{-24} ~{\rm cm^3/s}$). The PAMELA data can also be fit when the LSP has a non-negligible Higgsino component [@Feldman:2009wv] with small boost (clump) factors in the positron flux $\sim 2-4$. Figure (\[figpam\]) illustrates fits to the data for [various neutralino masses]{} with no boost factor in the positron flux. The figure is meant to show that models with wino-like LSPs which describe the PAMELA positron ratio should have masses in the range near (170 - 200) GeV. Progress has been made towards a complete fit to both the PAMELA positron, antiproton data, and the FERMI $e^{+} +e^{-}$ flux data using [GALPROP]{} [@Kane:2009if] and more exhaustive analyses are currently under way. A lighter LSP could also produce the PAMELA signal with a different set of propagation parameters. If other effects, such as small density fluctuations are included, the LSP mass range could cover slightly heavier masses. Photon Line Spectrum and Recent Probes -------------------------------------- The relative strength of the photon line spectrum arising from dark matter annihilations in the galaxy [@Bergstrom:1997fh] is highly sensitive to the gaugino content of the LSP [@Ullio:2001qk][@Yaguna:2009cy]. Thus with an essentially pure wino, as in the G$_2$ models, $\langle \sigma v \rangle^{1-\rm loop}_{\gamma Z} \sim 10^{-26}~\rm cm^3/s$, for a wino mass corresponding to the line energy of Fig.(\[figdata2\]). Such models provide promising probes for dark matter candidates with the FERMI data [@Collaboration:2010ex] in the central galaxy and from dwarf galaxies. In Fig.(\[figdata2\]) we illustrate this effect for the recently released photon data [@Collaboration:2010ex]. The analysis shows that annihilations of a pure wino are not inconsistent with an isothermal profile (which may be favoured by recent simulations including baryons [@Governato:2009bg]). Such a constraint is highly dependent on the profile uncertainties. At present, the PAMELA data can be described consistently with the FERMI [photon data and]{} Fig.(\[figdata2\]) shows that FERMI is close to sensitivity needed to see a signal in the line source. Another probe of annihilating dark matter comes from the FERMI analysis on dwarf galaxies [@Farnier; @Murgia]. The recently reported results show the strongest constraints are from Ursa Minor and Draco implying a signal should be seen for wino masses below $\sim 300 ~\rm GeV$. This constraint assumes a NFW dwarf density [profile]{} [@Collaboration:2010ex] (see however [@Burkert:1995yz; @Governato:2009bg; @Pasetto:2010se]). There presently is a rather appreciable uncertainty in the predicted flux from the dwarf galaxies due in part to the integration over the density (squared) source of dark matter [@Bergstrom:2005qk],[@Strigari:2006rd]. For the case of the Draco dwarf galaxy Ref. [@Essig:2009jx] finds an uncertainty of a factor of 10 or more. A more detailed analysis will help shed light on these constraints. It would be premature to deduce that the constraints are ruling out models until the profile of the dwarf galaxies are better understood and the inclusion of more stars enters into the analyses. CDMS and XENON -------------- A related indication of wino-like dark matter (but not pure wino) is that of an enhanced spin independent (SI) cross section when the wino content is supplemented by non-negligible sources of Higgsino and bino content. The spin dependent cross section is also enhanced, and their contribution is not negligible, at least for Xenon based targets. For the SI interactions with admixtures of the above type one finds SI cross sections in the interesting region of $\sim {\mathcal O} (10^{-44}) ~ \rm cm^2$ [@Bertin:2002sq; @BirkedalHansen:2002am], (for recent related work see [@DDt; @nelson; @DDFLN; @bnelson; @Hisano:2009xv; @Cohen:2010gj]). For a pure wino, the tree level cross section involving the Higgs exchanges vanish and loop corrections [@Hisano:2004pv] are not large enough to bring the cross section up in the region that is presently testable. Thus observation of a signal in CDMS II, [XENON-100]{} (or [EDELWEISS]{} and other related experiments) would immediately exclude a pure wino LSP. Deviating from the pure wino by a few percent leads to a detectable (SI) cross section. For example, with soft breaking parameters $ (m_0,m_{1/2},A_0,\tan \beta, (\Delta_1, \Delta_2, \Delta_3)) =((3000,500, 0)~ \rm GeV,4,( 0, -.56, -0.80)) $, with ${\rm sign}(\mu)> 0$, the LSP forms a *wino-like eigenstate*: $ (N_{\tilde{B}}, N_{\widetilde{W}},N_{\tilde{H}_1 }, N_{\tilde{H}_2})=( 0.114, -0.983, 0.127, -0.061)$, *with both a large halo annihilation cross section, $\langle \sigma v \rangle_{\w \w \to W^{+} W^{-}}$, and detectable SI scattering cross sections, $\sigma_{\rm SI}(\w p)$.* Specifically, one obtains the [following:]{} $\sigma_{\rm SI}(\w p) = 1 \times 10^{-8} ~~ {\rm pb }$, and $\sigma_{\rm SD}(\w p)= 6 \times 10^{-6} ~~ { \rm pb }$ with $ \langle \sigma v \rangle_{\w \w \to W^{+} W^{-}} = 2 \times 10^{-24}~{\rm cm^3/s}$. Here the LSP mass is $m_{\w} = 181~ \rm GeV $ and the gluino mass is very light $m_{\tilde g} = 357 ~ \rm GeV$. Such a model would produce discoverable jet signatures immediately at the LHC. Thus, this class of model produces positrons in the halo which describe the PAMELA data, and produces a spin independent scattering cross section within reach of the CDMS and XENON experiments (see: [@wsFK] and [@bnelson] for a similar emphasis). On the other hand if XENON-100 sees no signal, and the PAMELA data turns over at higher energies, a pure wino remains a possible and well-motivated interpretation. Conclusion \[conc\] =================== In this letter we have studied collider and dark matter implications within the setting of soft supersymmerty breaking based on string compactifications and in related models with non minimal gaugino sectors. The implications of a pure wino and a wino-like LSP in association with the production of light gluinos at the LHC, along with a possible interpretation of dark matter annihilations as a cause for the rising positron ratio in PAMELA satellite data, all provide exciting possibilities for the early discovery of supersymmetry. Such a discovery will have strong implications for the underlying theory and for the nature of soft supersymmetry breaking, as well as for the cosmological history of the universe. An underlying theory which can accommodate the positron excess, can produce testable event rates in direct detection experiments, and lead to testable signatures at the LHC due to the presence of light gluinos, all can arise with an LSP that has a substantial wino component. In addition, the wino-like LSP can have a spin independent interaction cross section that can be rather large when a non-negligible Higgsino component is present. A theory of this kind provides a compelling candidate to explain the nature of dark matter, its relic density from re-heating, and its annihilations in the galaxy. Recent photon constraints from FERMI on the above class of models are also analyzed and we have shown that there is a large region of parameter space where a wino-like LSP is consistent with the constraints. The constraints are very sensitive to the gaugino content of the wavefunction of the LSP and to the assumed halo profile. This parameter space accommodates light gluinos and therefore jets and missing energy signals that can be tested with early data at the LHC. We have particularly emphasized, via specific models, a light gluino, and the importance of three-body decay chains which yield large jet multiplicities from the light gluino decays producing wino or wino-like LSPs. The resulting set of decays are strikingly simple and predictive with gaugino production controlling the event topologies. The nearly degenerate charginos and neutralinos arise from the three body decays of the gluino and could be identified with a careful analysis after collecting a sample of gluino events. Indeed the models discussed here are ripe for studies at the LHC with low luminosities and at start up center of mass energies due to their large multi-jet event rates. Acknowledgements {#acknowledgements .unnumbered} ================ We collectively would like to thank Bobby Acharya, Elliot Bloom, Katherine Freese, Simona Murgia, Aaron Pierce, Jing Shao, Scott Watson and Kathryn Zurek for a broad range of discussions. This work was supported by National Science Foundation Grant PHY-0653587, and support from the Michigan Center for Theoretical Physics (MCTP) and Department of Energy grant DE-FG02-95ER40899. References {#references .unnumbered} ========== [999]{} O. Adriani [*et al.*]{} \[PAMELA Collaboration\], Nature [**458**]{}, 607 (2009); Phys. Rev. Lett.  [**102**]{}, 051101 (2009), O. Adriani [*et al.*]{}, arXiv:1001.3522 \[astro-ph.HE\]. A. A. Abdo [*et al.*]{} \[The Fermi LAT Collaboration\], Phys. Rev. Lett.  [**102**]{} (2009) 181101 \[arXiv:0905.0025 \[astro-ph.HE\]\]. Z. Ahmed [*et al.*]{} \[CDMS Collaboration\], Phys. Rev. Lett.  [**102**]{}, 011301 (2009); J. Angle [*et al.*]{} \[XENON Collaboration\], Phys. Rev. Lett.  [**100**]{}, 021303 (2008) E. Armengaud [*et al.*]{}, \[The EDELWEISS Collaboration\] arXiv:0912.0805 \[astro-ph.CO\]. Z. Ahmed [*et al.*]{} \[The CDMS-II Collaboration\], arXiv:0912.3592 \[astro-ph.CO\]. B. S. Acharya, K. Bobkov, G. L. Kane, P. Kumar and J. Shao, Phys. Rev.  D [**76**]{}, 126010 (2007); Phys. Rev.  D [**78**]{}, 065038 (2008); B. S. Acharya and K. Bobkov, arXiv:0810.3285 \[hep-th\]. G. L. Kane, J. D. Lykken, S. Mrenna, B. D. Nelson, L. T. Wang and T. T. Wang, Phys. Rev.  D [**67**]{}, 045008 (2003); G. L. Kane, J. D. Lykken, B. D. Nelson and L. T. Wang, Phys. Lett.  B [**551**]{}, 146 (2003). M. K. Gaillard and B. D. Nelson, Int. J. Mod. Phys.  A [**22**]{}, 1451 (2007). A. Corsetti and P. Nath, Phys. Rev.  D [**64**]{}, 125010 (2001). P. Grajek, G. Kane, D. Phalen, A. Pierce and S. Watson, Phys. Rev.  D [**79**]{}, 043506 (2009). J. Hisano, M. Kawasaki, K. Kohri and K. Nakayama, Phys. Rev.  D [**79**]{}, 063514 (2009). G. Kane, R. Lu and S. Watson, Phys. Lett.  B [**681**]{}, 151 (2009). D. Feldman, Z. Liu, P. Nath, B. D. Nelson, Phys. Rev.  D [**80**]{}, 075001 (2009); D. Feldman, arXiv:0908.3727. B. Altunkaynak, L. L. Everett, I. W. Kim, B. D. Nelson and Y. Rao, arXiv:1001.5261 \[hep-ph\]. E. Komatsu [*et al.*]{} \[WMAP Collaboration\], Astrophys. J. Suppl.  [**180**]{}, 330 (2009); arXiv:1001.4538 \[astro-ph.CO\]. T. Moroi and L. Randall, Nucl. Phys.  B [**570**]{}, 455 (2000). B. S. Acharya, P. Kumar, K. Bobkov, G. Kane, J. Shao and S. Watson, JHEP [**0806**]{}, 064 (2008). M. Endo, K. Hamaguchi and F. Takahashi, Phys. Rev. Lett.  [**96**]{}, 211301 (2006); S. Nakamura and M. Yamaguchi, Phys. Lett.  B [**638**]{}, 389 (2006). G. B. Gelmini and P. Gondolo, Phys. Rev.  D [**74**]{}, 023510 (2006). B. S. Acharya, G. Kane, S. Watson and P. Kumar, Phys. Rev.  D [**80**]{}, 083529 (2009). J. Hubisz, J. Lykken, M. Pierini and M. Spiropulu, Phys. Rev.  D [**78**]{}, 075008 (2008). For recent work see: L. Randall and D. Tucker-Smith, Phys. Rev. Lett.  [**101**]{}, 221803 (2008). G. Kane and A. Pierce, “Perspectives on LHC physics,” [*World Scientific (2008) 337 p.*]{} P. Nath [*et al.*]{}, “The Hunt for New Physics at the Large Hadron Collider,” arXiv:1001.2693 \[hep-ph\]. C. H. Chen, M. Drees and J. F. Gunion, Phys. Rev.  D [**55**]{}, 330 (1997). M. R. Buckley, L. Randall and B. Shuve, arXiv:0909.4549 \[hep-ph\]. M. K. Gaillard, B. D. Nelson and Y. Y. Wu, Phys. Lett.  B [**459**]{}, 549 (1999); J. A. Bagger, T. Moroi and E. Poppitz, JHEP [**0004**]{}, 009 (2000). K. Griest and D. Seckel, Phys. Rev.  D [**43**]{}, 3191 (1991). S. P. Martin and M. T. Vaughn, Phys. Rev.  D [**50**]{}, 2282 (1994). D. M. Pierce, J. A. Bagger, K. T. Matchev and R. j. Zhang, Nucl. Phys.  B [**491**]{}, 3 (1997). B. Allanach, Comput. Phys. Commun.  [**143**]{}, 305 (2002). A. Djouadi, M. M. Muhlleitner and M. Spira, Acta Phys. Polon.  B [**38**]{}, 635 (2007). T. Sjostrand, S. Mrenna, P. Skands, JHEP [**0605**]{}, 026 (2006) J. Conway et. al, PGS-4. D. Feldman, Z. Liu and P. Nath, Phys. Rev. Lett.  [**99**]{}, 251802 (2007); JHEP [**0804**]{}, 054 (2008); arXiv:0911.0217 \[hep-ph\], PLB in Press. D. E. Acosta [*et al.*]{} \[CDF Collaboration\], Phys. Rev.  D [**71**]{}, 052003 (2005). G. L. Bayatian [*et al.*]{} \[CMS Collaboration\], J. Phys. G [**34**]{}, 995 (2007). P. Gondolo, J. Edsjo, P. Ullio, L. Bergstrom, M. Schelke and E. A. Baltz, JCAP [**0407**]{}, 008 (2004). A. W. Strong and I. V. Moskalenko, Astrophys. J.  [**509**]{}, 212 (1998). M. S. Turner and F. Wilczek, Phys. Rev.  D [**42**]{}, 1001 (1990); M. Kamionkowski and M. S. Turner, Phys. Rev.  D [**43**]{}, 1774 (1991); E. A. Baltz and J. Edsjo, Phys. Rev.  D [**59**]{}, 023511 (1998); A. Bottino, F. Donato, N. Fornengo and P. Salati, Phys. Rev.  D [**58**]{}, 123503 (1998); L. Bergstrom, J. Edsjo and P. Ullio, Astrophys. J.  [**526**]{}, 215 (1999). E. A. Baltz, J. Edsjo, K. Freese and P. Gondolo, Phys. Rev.  D [**65**]{}, 063511 (2002). G. L. Kane, L. T. Wang and J. D. Wells, Phys. Rev.  D [**65**]{}, 057701 (2002). J. Hisano, S. Matsumoto, O. Saito and M. Senami, Phys. Rev.  D [**73**]{}, 055004 (2006). S. Sekmen, Ph. D. thesis, CMS TS-2009/025. B. S. Acharya, P. Grajek, G. L. Kane, E. Kuflik, K. Suruliz and L. T. Wang, arXiv:0901.3367 \[hep-ph\]. D. Feldman, Z. Liu and P. Nath Phys. Rev.  D [**80**]{}, 015007 (2009). S. Bhattacharya, U. Chattopadhyay, D. Choudhury, D. Das and B. Mukhopadhyaya, arXiv:0907.3428 \[hep-ph\]. H. Baer, V. Barger, A. Lessa and X. Tata, JHEP [**0909**]{}, 063 (2009); H. Baer, S. Kraml, A. Lessa and S. Sekmen, arXiv:0911.4739 \[hep-ph\]. See Ref. [@Chen:1996ap] and J. L. Feng, T. Moroi, L. Randall, M. Strassler and S. f. Su, Phys. Rev. Lett.  [**83**]{}, 1731 (1999); T. Gherghetta, G. F. Giudice and J. D. Wells, Nucl. Phys.  B [**559**]{}, 27 (1999); A. J. Barr, C. G. Lester, M. A. Parker, B. C. Allanach and P. Richardson, JHEP [**0303**]{}, 045 (2003); U. Chattopadhyay, D. Das, P. Konar and D. P. Roy, Phys. Rev.  D [**75**]{}, 073014 (2007); M. Ibe, T. Moroi and T. T. Yanagida, Phys. Lett.  B [**644**]{}, 355 (2007); S. Asai, T. Moroi and T. T. Yanagida, Phys. Lett.  B [**664**]{}, 185 (2008). S. Khalil, C. Munoz and E. Torrente-Lujan, New J. Phys.  [**4**]{}, 27 (2002); M. Endo and F. Takahashi, Phys. Rev.  D [**74**]{}, 063502 (2006). J. P. Conlon and F. Quevedo, JCAP [**0708**]{}, 019 (2007). J. Hisano, S. Matsumoto, M. M. Nojiri and O. Saito, Phys. Rev.  D [**71**]{}, 063528 (2005); N. Arkani-Hamed, D. P. Finkbeiner, T. R. Slatyer and N. Weiner, Phys. Rev.  D [**79**]{}, 015014 (2009). D. Feldman, Z. Liu and P. Nath, Phys. Rev.  D [**79**]{}, 063509 (2009); M. Ibe, H. Murayama and T. T. Yanagida, Phys. Rev.  D [**79**]{}, 095009 (2009). L. Bergstrom, J. Edsjo, P. Gondolo and P. Ullio, Phys. Rev.  D [**59**]{}, 043506 (1999). D. Hooper, J. E. Taylor and J. Silk, Phys. Rev.  D [**69**]{}, 103509 (2004). D. Hooper, A. Stebbins and K. M. Zurek, Phys. Rev.  D [**79**]{}, 103513 (2009). M. A. DuVernois [*et al.*]{}, Astrophys. J.  [**559**]{}, 296 (2001); J. J. Beatty [*et al.*]{}, Phys. Rev. Lett.  [**93**]{}, 241102 (2004); M. Aguilar [*et al.*]{} \[AMS-01 Collaboration\], Phys. Lett.  B [**646**]{}, 145 (2007). L. Bergstrom and P. Ullio, Nucl. Phys.  B [**504**]{}, 27 (1997); Z. Bern, P. Gondolo and M. Perelstein, Phys. Lett.  B [**411**]{}, 86 (1997); P. Ullio and L. Bergstrom, Phys. Rev.  D [**57**]{}, 1962 (1998). P. Ullio, JHEP [**0106**]{}, 053 (2001). C. E. Yaguna, Phys. Rev.  D [**80**]{}, 115002 (2009). J. Hisano, S. Matsumoto, M. M. Nojiri and O. Saito, Phys. Rev.  D [**71**]{}, 015007 (2005). D. Feldman and G. Kane, “Perspectives on Supersymmetry II”; World Scientific, Dec. 2009. V. Bertin, E. Nezri and J. Orloff, JHEP [**0302**]{}, 046 (2003). A. Birkedal-Hansen and B. D. Nelson, Phys. Rev.  D [**67**]{}, 095006 (2003). D. Feldman, Z. Liu and P. Nath, Phys. Lett.  B [**662**]{}, 190 (2008); Phys. Rev.  D [**78**]{}, 083523 (2008). B. Altunkaynak, M. Holmes and B. D. Nelson, JHEP [**0810**]{}, 013 (2008). D. Feldman, Z. Liu and P. Nath, arXiv:0912.4217 \[hep-ph\]. M. Holmes and B. D. Nelson, arXiv:0912.4507 \[hep-ph\]. J. Hisano, K. Nakayama and M. Yamanaka, Phys. Lett.  B [**684**]{}, 246 (2010). T. Cohen, D. J. Phalen and A. Pierce, arXiv:1001.3408 \[hep-ph\]. A. A. Abdo, M. Ackermann and M. Ajello, arXiv:1001.4836 C. Farnier , \[FERMI LAT Collaboration\], RICAP09 (Roma International Conference on Astro-Particle Physics), May 13-15, 2009. S.  Murgia \[FERMI-LAT Collaboration\], 2009 Fermi Symposium 2-5 November 2009 - Washington DC. F. Collaboration, J. S. Bullock, M. Kaplinghat and G. D. Martinez, arXiv:1001.4531 A. Burkert, IAU Symp.  [**171**]{}, 175 (1996) \[Astrophys. J.  [**447**]{}, L25 (1995)\]. F. Governato [*et al.*]{}, Nature [**463**]{}, 607 (2009). S. Pasetto, E. K. Grebel, P. Berczik, R. Spurzem and W. Dehnen, arXiv:1002.1085 \[astro-ph.CO\]. L. Bergstrom and D. Hooper, Phys. Rev.  D [**73**]{}, 063510 (2006). L. E. Strigari, S. M. Koushiappas, J. S. Bullock and M. Kaplinghat, Phys. Rev.  D [**75**]{}, 083526 (2007). R. Essig, N. Sehgal and L. E. Strigari, Phys. Rev.  D [**80**]{}, 023506 (2009). [^1]: Models dominated by a tree level contribution to the soft terms are also considered in Section \[GeneralPredictions\].
{ "pile_set_name": "ArXiv" }
--- abstract: 'In the first section of the paper, we will give some basic definitions and properties about Crystalline Graded Rings. In the following section we will provide a general description of the center. Afterwards, the case where the grading group is Abelian finite will be handled. The center will have some properties of a crystalline graded ring, but not all. We will call this Arithmetically Crystalline Graded. The center is crystalline graded if the part of degree zero is a principal ideal domain. The last section deals with the case where the grading group is non-Abelian finite. Since this situation is much more complicated than the Abelian case, we primarily focus on the conditions to have a trivial center. The fact that the center is Arithmetically Crystalline Graded also holds in this case.' author: - | Tim Neijens\ University of Antwerp\ `[email protected]` - | Freddy Van Oystaeyen\ University of Antwerp\ `[email protected]` title: The Center of Crystalline Graded Rings --- Introduction {#introduction .unnumbered} ============ Crystalline graded rings have been introduced in [@NVO6] as a generalization of crossed products on one hand and of generalized Weyl algebras on the other, cf. [@B1], [@BVO3]. In this paper we focus on the situation where the subring of degree $e$, say $A_e$, of the crystalline graded ring $A$, is a Dedekind domain, $A_e = D$ with field of fractions $L$. Important examples in case the grading group is torsionfree Abelian, usually ${\mathbb{Z}}$ or ${\mathbb{Z}}^n$, include $D$-orders in well-known algebras over $L$ like the first Weyl algebra, the quantum plane, quantum $sl_2$, …, many algebras of quantum type appearing in : [@B], [@BVO3], [@NVO6], …However when attention is restricted to finite grading groups $G$, then crystalline graded rings over a commutative ring are necessarily PI-rings, then typical examples of quantum-type (quantized or $q$-deformed algebras) would correspond to the case where the deformation parameter $q$ is a root of unity. Hence the class of crystalline graded rings over $D$ consists of $D$-orders in classical crossed products over the field of fractions which are finite dimensional $L$-algebras. It is a natural problem to aim at an algebraic classification for crystalline graded rings over a Dedekind domain $D$, specializing to particular cases like discrete valuation rings, $k[T]$ or rings of integers in number fields for more concrete applications. On the other hand, most results extend to the situation where $A_e$ is the coordinate ring of a normal variety of dimension $d$, i.e. a Noetherian integrally closed domain (having global dimension $d$), but we do not consider this extension here. Applying the theory of maximal $D$-orders seems to be the obvious way to start analyzing the algebra structure of the noncommutative algebras under consideration but at an even more elementary level there is first the problem of determining the center of $A$. The first part of this paper deals with the determination of the center of a crystalline graded ring as before. Knowing that skew group rings and twisted group rings are special cases of our constructions, it is clear that rings of invariants for certain specific group actions and ray classes, as in the theory of projective representations of finite groups, will play an important part in this. Secondly, the particular case where the center $Z(A)$ is minimal (i.e. $Z(A)=D$ or $Z(A)= D^G$) should be the most easy to describe. In fact, the results concerning Clifford representations for Abelian groups, cf. [@CVO], suggest that such crystalline graded rings will be (maximal) $D$-orders in generalized Clifford algebras over the quotient field $L$ of $D$. Preliminaries ============= \[def1\]**Pre-Crystalline Graded Ring**\ Let $A$ be an associative ring with unit $1_A$. Let $G$ be an arbitrary group. Consider an injection $u: G \rightarrow A$ with $u_e = 1_A$, where $e$ is the neutral element of $G$ and $u_g \neq 0$, $\forall g \in G$. Let $R \subset A$ be an associative ring with $1_{R}=1_A$. We consider the following properties: (C1) : \[def2\] $A = \bigoplus_{g \in G} R u_g$. (C2) : \[def3\] $\forall g \in G$, $R u_g = u_g R$ and this is a free left $R$-module of rank $1$. (C3) : \[def4\] The direct sum $A = \bigoplus_{g \in G} R u_g$ turns $A$ into a $G$-graded ring with $R = A_e$. We call a ring $A$ fulfilling these properties a **Pre-Crystalline Graded Ring**. \ \[def5\] With conventions and notation as in Definition \[def1\]: 1. For every $g \in G$, there is a set map $\sigma_g : R \rightarrow R$ defined by: $u_g r = \sigma_g(r)u_g$ for $r \in R$. The map $\sigma_g$ is in fact a surjective ring morphism. Moreover, $\sigma_e = \textup{Id}_{R}$. 2. There is a set map $\alpha : G \times G \rightarrow R$ defined by $u_g u_h = \alpha(g,h)u_{gh}$ for $g,h \in G$. For any triple $g,h,t \in G$ the following equalities hold: $$\begin{aligned} \alpha(g,h)\alpha(gh,t)&=&\sigma_g(\alpha(h,t))\alpha(g,ht) \label{def6},\\ \sigma_g(\sigma_h(r))\alpha(g,h)&=& \alpha(g,h)\sigma_{gh}(r) \label{def7}. \end{aligned}$$ 3. $\forall g \in G$ we have the equalities $\alpha(g,e) = \alpha(e,g) = 1$ and $\alpha(g,g^{-1}) = \sigma_g(\alpha(g^{-1},g)).$ **Proof** See [@NVO6]. ${\hfill \Box}$\ Notation as above, the following are equivalent: 1. $R$ is $S(G)$-torsionfree. 2. $A$ is $S(G)$-torsionfree. 3. $\alpha(g,g^{-1})r=0$ for some $g \in G$ implies $r = 0$. 4. $\alpha(g,h)r=0$ for some $g,h \in G$ implies $r = 0$. 5. $R u_g = u_g R$ is also free as a right $R$-module with basis $u_g$ for every $g \in G$. 6. for every $g \in G$, $\sigma_g$ is bijective hence a ring automorphism of $R$. **Proof** See [@NVO6]. ${\hfill \Box}$\ [Any $G$-graded ring $A$ with properties **(C1),(C2),(C3)**, and which is $G(S)$-torsionfree is called a **crystalline graded ring**. In case $\alpha(g,h) \in Z(R)$, or equivalently $\sigma_{gh}=\sigma_g \sigma_h$, for all $g,h \in G$, then we say that $A$ is **centrally crystalline**.]{}\ We associate the following function to the $2$-cocycle $\alpha$: $$f_\alpha: G \times G \rightarrow K: (x,s)\mapsto \frac{\alpha(x,s)}{\alpha(xsx^{-1},x)}.$$ We will drop the subscript $\alpha$ if there is no confusion possible. For the remainder we also need the following definitions: [A **regular element of $G$ with respect to $\alpha$** is an $x \in G$ such that $\alpha(x,g)=\alpha(g,x)$ for every $g \in C(x):=\{g \in G| gx=xg\}$. The set of $\alpha$-regular elements in $G$ is denoted by $G_\textup{reg}$.]{}\ [An **$\alpha$-ray class** is defined to be the conjugation class of a regular element with respect to $\alpha$.]{}\ [\[cen1\] Given an $\alpha$-ray class, we define **the $\alpha$-ray class sum** as the sum of the base elements $u_g$ where $g$ is in the $\alpha$-ray class.]{}\ The Center ========== Setting ------- We will use the following notation and conventions. - $R$ is a commutative domain. - $K = {\mathbb{Q}}(R)$ is the field of fractions of $R$. - $G$ is a finite group. - $A {\mathop \diamondsuit \limits_{\sigma ,\alpha}}R$ is a crystalline graded ring. - $W = \textup{Ker} \sigma$. The Center ---------- For the case of a twisted group ring, see for example [@NVO5]. From now we consider $A= R {\mathop \diamondsuit \limits_{\sigma ,\alpha}}G$ where $R$ is a commutative domain and $G$ finite. Let $W = \textup{Ker}\sigma$. If we write an inverse, we mean the inverse in the field of fractions $K$ of $R$. \[cen2\] We have the following formulas $\forall x,g \in G$: 1. In $K {\mathop \diamondsuit \limits_{\sigma ,\alpha}}G : u_x^{-1} = u_{x^{-1}}\alpha^{-1}(x,x^{-1}) = \alpha^{-1}(x^{-1}, x)u_{x^{-1}}$. 2. $\alpha(gx, x)\sigma_{xgx^{-1}}(\alpha^{-1}(x, x^{-1}))=\alpha^{-1}(xgx^{-1}, x)$. **Proof** 1. $u_x u_{x^{-1}}\alpha^{-1}(x,x^{-1}) = \alpha(x,x^{-1})u_e \alpha^{-1}(x,x^{-1}) =1$ and $u_{x^{-1}}\alpha^{-1}(x,x^{-1}) u_x = \sigma_{x^{-1}}(\alpha^{-1}(x,x^{-1}))\alpha(x^{-1},x)=1$. 2. Use the $2$-cocycle relation (\[def6\]) for $(xg, x^{-1},x)$. ${\hfill \Box}$ [\[cen3\] With notation as above we have $$\sum\limits_{s \in G} {r_s u_s \in Z(A)} \Leftrightarrow \left\{ {\begin{array}{*{20}c} {r_s = 0} \hfill & {\forall s \notin W} \hfill \\ {\sigma _x (r_s )\alpha (x,s) = r_{xsx^{ - 1} }\alpha (xsx^{ - 1} ,x) } \hfill & {\forall x \in G, s\in W}. \hfill \\ \end{array} } \right.$$ \ ]{} **Proof** - Let $\sum r_s u_s \in Z(A)$, then $\forall r \in R$: $$\begin{aligned} \left(\sum_{s \in G}r_s u_s\right)r = r\left(\sum_{s \in G}r_s u_s\right)&\Leftrightarrow& \sum_{s \in G}r_s \sigma_s(r)u_s = \sum_{s \in G}r_s r u_s\\ \ &\Leftrightarrow& \sigma_s(r)=r \ \ \ \forall s\in G \textrm{ with } r_s \neq 0\\ \ &\Leftrightarrow& r_s = 0\ \ \ \forall s\notin W. \end{aligned}$$ - Let $x \in G$ and $s \in W$. We will use the formulas in Lemma \[cen2\]: $$\begin{aligned} u_x r_s u_s u_x^{-1} &=& \sigma_x(r_s)u_x u_s u_{x^{-1}}\alpha^{-1}(x, x^{-1})\\ \ &=& \sigma_x(r_s)\alpha(x,s)\alpha(xs, x^{-1})u_{xsx^{-1}}\alpha^{-1}(x, x^{-1})\\ \ &=& \sigma_x(r_s)\alpha(x,s)\alpha(xs, x^{-1})\sigma_{xsx^{-1}}(\alpha^{-1}(x, x^{-1}))u_{xsx^{-1}}\\ \ &=& \sigma_x(r_s)\alpha(x,s)\alpha(xs, x^{-1})\sigma_{xs}(\alpha( x^{-1},x))^{-1}u_{xsx^{-1}}\\ \ &=& \sigma_x(r_s)\alpha(x,s)\alpha(xsx^{- 1}, x)u_{xsx^{-1}}. \end{aligned}$$ And this finishes the proof. ${\hfill \Box}$\ If $g \in G$ is a degree of a nonzero element in $Z(A)$, say $0 \neq r_g u_g \in Z(A)$, then $\sigma_x(r_g)\alpha(x,g) = r_g \alpha(g,x)$ for $x \in G$ follows, or $\sigma_x(r_g) = r_g f(g,x)$. We now calculate in a straightforward way: $$\begin{aligned} r_g f(g,xy) &=& \sigma_{xy}(r_g)\\ \ &=&\sigma_x(\sigma_y(r_g))\\ \ &=& \sigma_x(r_g f(g,y))\\ \ &=& \sigma_x(r_g)\sigma_x(f(g,y))\\ \ &=& r_g f(g,x) \sigma_x(f(g,y)).\end{aligned}$$ Hence we arrive at (since $R$ is a domain): $$f(g,xy)=f(g,x)\sigma_x(f(g,y)).$$ A $g\in G$ corresponding to a nonzero $r_g u_g \in Z(A)$ corresponds therefore to a crossed homomorphism $f(g, -): G \rightarrow K$, i.e. to an element of $H^1(G, K)$ where $K$ is a $G$-module via the extension of the $G$-action on $R$ to the fraction field. The restriction of $f(g,-)$ to $W \subset G$ defines a multiplicative map $W \rightarrow K$. This situation appears often when $G$ is Abelian, the case we first treat in some detail hereafter. $G$ Abelian ----------- From now we consider $A= D {\mathop \diamondsuit \limits_{\sigma ,\alpha}}G$ where $D$ is a Dedekind domain and $G$ Abelian finite. Let $K$ be the field of fractions of $D$. In this section we will prove that the center of $A$ is crystalline graded itself in certain cases. It is easy to see that the center is graded since $G$ is Abelian. With $G$ Abelian, the second statement in Proposition \[cen3\] becomes $$\sigma_x(d_s)\alpha(x,s)= d_s \alpha(s,x) \ ,\ \ \forall x \in G, s \in W.\label{cen4}$$ And the function $f := f_\alpha$ becomes $$f: G \times G \rightarrow K : (g,h)\mapsto \frac{\alpha(g,h)}{\alpha(h,g)}.$$ So rewriting (\[cen4\]): $$\sigma_x(d_s)=d_s f(s,x) \ ,\ \ \forall x \in G, s \in W.\label{cen5}$$ Consider $B = D {\mathop \diamondsuit \limits_{\sigma ,\alpha}}W$. The first statement in (\[cen3\]) gives us $Z(A)\subset Z(B)$. Since the action on $Z(B)$ is trivial, we find that, using (\[cen5\]) $$Z(B)=\sum_{s \in W_{\textup{reg}}} D u_s,$$ where $W_{\textup{reg}} = \{s \in W | \alpha(s,x)=\alpha(x,s), \forall x \in W\}$. Fix a $d_s u_s \in Z(A)$, so $s \in K_{\textup{reg}}$ (looking at these elements is sufficient since the center is graded). We have two cases: either $s \in G_{\textup{reg}}$, or $s \notin G_{\textup{reg}}$.\ If $s \in G_{\textup{reg}}$, we have that $$\sigma_x(d_s) = d_s\ ,\ \ \forall x \in G \Rightarrow d_s \in D^G,$$ where we define $D^G = \{d \in D | \sigma_x(d)=d \ ,\forall x \in G\}$.\ Now consider the case that $s \notin G_{\textup{reg}}$ and define $\forall x$ with $\alpha(s,x)\neq\alpha(x,s)$ the set $I_{s,x}$: $$I_{s,x}=\{d\in D | \sigma_x(d) = d f(s,x)\},$$ and define $I_s$ as the intersection of all $I_{s,x}$, $x \in G$ with $\alpha(s,x)\neq \alpha(x,s)$. So now we have $$Z(A) = \sum_{s \in W \cap G_{\textup{reg}}}D^G u_s + \sum_{s \in W_{\textup{reg}}\backslash G_{\textup{reg}}}I_s u_s.$$ [\[cen6\]$I_s$ as defined above is a finitely generated $D^G$-bimodule of rank $1$. It is not multiplicatively closed.]{}\ **Proof**The fact that $I_s$ is a $D^G$-bimodule is easily checked. Not multiplicatively closed follows from $\sigma_x(mn)= \sigma_x(m)\sigma_x(n)=mnf(x,s)^2 \neq mnf(x,s)$ since $0$ and $1$ are the only idempotents.\ Since $D$ is finitely generated over $D^G$ and $D^G$ is a Dedekind domain ($G$ finite), we see that $I_s$ has finite rank. The rank is determined as follows. Take an $x$ such that $\alpha(s,x)\neq \alpha(x,s)$ and take $u,v \in I_s$: $$\frac{\sigma_x(u)}{\sigma_x(v)}=\frac{uf(s,x)}{vf(s,x)}= \frac{u}{v},$$ and this entails that $u/v \in K^G$ so the rank of $I_s$ is $1$. ${\hfill \Box}$\ [$Z(A)$ is crystalline graded over $D^G$ when $D$ is a principal ideal domain.]{}\ **Proof** This actually follows almost immediately from Proposition \[cen6\]. Since $I_s$ is a torsionfree module over a principal ideal domain, it is free. Since it has rank $1$ we find that $I_s = D^G d_s$ for some $d_s \in D$ and so our center becomes when we set $v_s = d_s u_s$ $$Z(A) = \sum_{s \in W \cap G_{\textup{reg}}}D^G u_s + \sum_{s \in W_{\textup{reg}}\backslash G_{\textup{reg}}}D^G v_s.$$ ${\hfill \Box}$\ We can now address to the question when the center is trivial, e.g. $Z(A)=D$ or $Z(A)=D^G$. This means that the only component that can appear is the component corresponding to $u_e=1$, and so our center will be $D^G$. There are a number of possibilities when this may happen: 1. $W_\textup{reg} = \{e\}$. 2. $W_\textup{reg} \cap G_\textup{reg} = \{e\}$ and $I_s = \{0\} \ \ \forall s \in W_\textup{reg} \backslash G_\textup{reg}$. We are not quite interested in the first condition, namely that $W$ has no regular elements, but we are interested in the second condition: - $\{e\}=W_\textup{reg}\cap G_\textup{reg} = G_\textup{reg} \cap W$. So no $G$-regular elements distinct from $e$ may be in $W$. - $I_s = \{0\} \ \ \forall s \in W_\textup{reg} \backslash G_\textup{reg}$. This means that there is no solution to $$\frac{\sigma_x(d)}{d} = f(x,s) \ \ \forall x \in G \textup{ with } \alpha(x,s)\neq\alpha(s,x).$$ $G$ Not Abelian --------------- When $G$ is not Abelian, we get a few extra problems. For one, the center will not be graded. Furthermore, we will need a lot more criteria to form the center. For this section, we will set $R=D$ to be a Dedekind domain. We set $$A = D {\mathop \diamondsuit \limits_{\sigma ,\alpha}}G,$$ $$B = D {\mathop \diamondsuit \limits_{\sigma ,\alpha}}W,$$ where $\sigma$ and $\alpha$ are as usual, $K = \textup{Ker}\sigma$. We have the following theorem from [@CVO]: [Let $R$ be a domain and let $G$ be a finite group. If a $2$-cocycle $\alpha \in Z^2(G, R)$ has the property that the corresponding function $f_\alpha(x,s)=1$, $\forall \alpha$-regular $s \in G$ and all $x \in G$, then the $\alpha$-ray class sums form an $R$-basis for the center $Z(R *_\alpha G)$ of the generalized crossed product $R *_\alpha G := R \mathop \diamondsuit \limits_{\textup{Id} ,\alpha }G$.]{}\ We can use this theorem on $B$ since the twist generated by $\sigma$ is trivial for $W$. We will call the property that $$f_\alpha(x,s)=1, \ \ \forall s \in G_{\textup{reg}},\ \ \forall x \in G,\label{cen7}$$ the universal regularity condition **URC** of $\alpha$ for $G$. (Remark: for now, $G_{\textup{reg}}$ has only meaning when the twist $\sigma$ is trivial.) Now denote $C_1, \ldots, C_t$ the ray classes of $W$ and $v_i:=v_{C_i}$ the ray class sum corresponding to $C_i$, $\forall i$. We find, if $\alpha$ satisfies the **URC** (\[cen7\]) for $W$: $$Z(B)= \sum_{i=1}^{t}D v_i.$$ [With notations as above, we have the following formula $\forall x \in G$, $\forall i = 1,\ldots, t$: $$u_x v_i u_x^{-1} = \sum_{g \in C_i} f(x,g)u_{xgx^{-1}}.$$]{} **Proof** We use the formulas in Lemma \[cen2\] $$\begin{aligned} u_x v_i u_x^{-1} &=& u_x v_i u_{x^{-1}}\alpha^{-1}(x,x^{-1})\\ \ &=& u_x \left(\sum_{g \in C_i}u_g\right)u_{x^{-1}}\alpha^{-1}(x,x^{-1})\\ \ &=& \left(\sum_{g \in C_i}\alpha(x,g)u_{xg}\right)u_{x^{-1}}\alpha^{-1}(x,x^{-1})\\ \ &=& \sum_{g \in C_i}\alpha(x,g)\alpha(xg,x^{-1})\sigma_{xgx^{-1}}(\alpha^{-1}(x,x^{-1}))u_{xgx^{-1}}\\ \ &=& \sum_{g \in C_i}\alpha(x,g)\alpha^{-1}(xgx^{-1}, x)u_{xgx^{-1}}\\ \ &=& \sum_{g \in C_i} f(x,g)u_{xgx^{-1}}. \end{aligned}$$ ${\hfill \Box}$\ For the remainder of this section, let us assume that $\alpha$ satisfies the **URC** (\[cen7\]) for $W$. We can now define $$W_{\textup{reg}}=\{g \in W | f(x,g)=1 \ \ \forall x \in W\},$$ $$G_{\textup{reg}}=\{g \in G | f(x,g)=1 \ \ \forall x \in G\}.$$ Like in the previous section, we can use Proposition \[cen3\] so we restrict to $Z(B)$.\ [\[cen8\] If the $2$-cocycle $\alpha$ satisfies the **URC** (\[cen7\]), we have that $\forall x \in G$, $\forall s \in W_\textup{reg}$, $xsx^{-1} \in W_\textup{reg}$.]{}\ **Proof** Since $\alpha$ satisfies the **URC** (\[cen7\]) we only have to check that for $g \in C_W(xsx^{-1})$ we have $u_g u_{xsx^{-1}} = u_{xsx^{-1}} u_g$. (One easily checks that this is equivalent with regularity.) We see that $x^{-1}gx \in C_W(s)$ and thus since $s \in W_\textup{reg}$ (we use the formulas in Lemma \[cen2\]) $$\begin{aligned} \ &\ & u_s u_{x^{-1}g x} = u_{x^{-1}g x}u_s\\ &\Rightarrow& u_s \alpha^{-1}(x^{-1}g,x)\alpha^{-1}(x^{-1},g)u_{x^{-1}}u_g u_x \\ &\ &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ = \alpha^{-1}(x^{-1}g,x)\alpha^{-1}(x^{-1},g)u_{x^{-1}}u_g u_x u_s \\ &\Rightarrow& u_s u_{x^{-1}}u_g u_x = u_{x^{-1}}u_g u_x u_s\\ &\Rightarrow& u_{x^{-1}}^{-1}u_s u_{x^{-1}}u_g= u_g u_x u_s u_x^{-1}\\ &\Rightarrow& \alpha^{-1}(x,x^{-1})u_x u_s u_{x^{-1}}u_g = u_g u_x u_s u_{x^{-1}}\alpha^{-1}(x,x^{-1})\\ &\Rightarrow& u_x u_s u_{x^{-1}}u_g = u_g u_x u_s u_{x^{-1}}\\ &\Rightarrow& u_{xsx^{-1}}u_g = u_g u_{xsx^{-1}}.\end{aligned}$$ And so $xsx^{-1} \in W_\textup{reg}$. ${\hfill \Box}$\ Take an element $y =\sum_{i = 1}^t d_i v_i \in Z(B)$, $d_i \in D$. We have to take a full sum now, since maybe the direct sum we have in $Z(B)$ does not hold in $Z(A)$. We calculate $$u_x \left(\sum_{i = 1}^t d_i v_i\right) u_x^{-1}= \sum_{i=1}^t \sigma_x(d_i)\sum_{g\in C_i}f(x,g)u_{xgx^{-1}}.\label{cen9}$$ So for $\sum_{i=1}^t d_i v_i$ to be an element of $Z(A)$ we must have that this equation (\[cen9\]) equals $\sum_{i=1}^t d_i v_i$, $\forall x \in G$. In other words, this means that for all appearing ray classes $C_i$, we have that $\forall g \in C_i$, $\forall x \in G$ that $xgx^{-1} \in W_\textup{reg}$. This is true according to Lemma \[cen8\]. We get the following condition on the coefficients $d_i$ of $y$ to be in the center of $A$, $\forall x \in G, \forall i = 1,\ldots, t,\forall g \in C_i$: $$\sigma_x(d_i)f(x,g) = d_j\ \ ,\ \ \textrm{ where } j \textrm{ is determined by } xgx^{-1} \in C_j. \label{cen10}$$ Let $d_i \neq 0$ for some $i$. We find, looking at equation (\[cen10\]), that all coefficients $d_j$ corresponding to the ray classes $C_j$ will be nonzero, where $j$ is determined by $xgx^{-1} \in C_j$ for some $x\in G$, some $g \in C_i$. Define for ease of notation the saturation class $\Gamma_i$, $\forall i = 1, \ldots, t$ as the set consisting of all ray classes $C$ (for $W$) such that if you take $x \in G$, $C$ in $\Gamma_i$ and $g \in C$, there is a $C' \in \Gamma_i$ with $xgx^{-1} \in C'$. So this means the general expression for an element $y \in Z(A)$ may be written as: $$y = \sum_{C \in \Gamma_i}d_C v_C\ \ \ \ \textrm{ for some }i.$$ Let $\bar x = \bar z \in G/W$ or equivalently $\exists k \in W: z =kx$.\ [Let $g,k \in W$, and $x \in G$. Then $$f(kx,g)=f(x,g)f(k, xgx^{-1}),$$ i.e. if $g \in W_\textup{reg}$ then $f(kx,g)=f(x,g)$.]{}\ **Proof** $$\begin{aligned} f(kx,g) &=& \frac{\alpha(kx,g)}{\alpha(kxgx^{-1}k^{-1}, kx)} = \frac{\alpha(kx,g)u_{kxg}}{\alpha(kxgx^{-1}k^{-1}, kx)u_{kxg}}=\frac{u_{kx}u_{g}}{u_{kxgx^{-1}k^{-1}}u_{kx}}\\ \ &=& \frac{\alpha^{-1}(k,x)u_k u_x u_g}{\sigma_{kxgx^{-1}k^{-1}}(\alpha^{-1}(k,x))u_{kxgx^{-1}k^{-1}}u_{k}u_{x}}\\ \ &=& \frac{\alpha(x,g)u_ku_{xg}}{\alpha(kxgx^{-1}k^{-1},k)\alpha^{-1}(k,xgx^{-1})u_k u_{xgx^{-1}}u_x}\\ \ &=& \frac{\alpha(x,g)u_k u_{xg}}{\alpha(xgx^{-1}, x)\alpha(kxgx^{-1}k^{-1},k)\alpha^{-1}(k, xgx^{-1})u_k u_{xg}}\\ \ &=& f(x,g)f(k,xgx^{-1}).\end{aligned}$$ ${\hfill \Box}$\ [If $\forall g \in C$, $C$ ray class in $W$, $\forall x,z \in G$ such that $\bar x = \bar z$ in $G/W$, then $$f(x,g) = f(z,g).$$]{}\ In particular, for $g \in C$, ray class in $W$ we have that $$F_g : G/W \rightarrow K : \bar x \mapsto f(x,g)$$ is well defined.\ Since the condition (\[cen10\]) is not as beautiful as the equation in the Abelian case (\[cen5\]), we will not describe the center fully. Instead we want to find conditions on the $2$-cocycle $\alpha$ and twist $\sigma$ to have a trivial center, e.g. $Z(A)=D$ or $Z(A)=D^G$. This for example happens if $W = \{e\}$. We can also describe the center when $W = G$, since there will be no twist and we can use the theory of twisted group rings. We can first assume that the center is of the following form: $$Z(A) = \bigoplus_{C \textup{ a certain ray class in }W}d_C v_C.$$This actually implies that a general element $y \in Z(A)$ is of the form $d_C v_C$ for a certain ray class $C$ in $W$. Not all ray classes $C$ will be present though. Looking at equation (\[cen9\]), for $y$ to be in $Z(A)$, we need to have if $d_C \neq 0$ then $\forall x \in G: xCx^{-1} = C$. In group theoretic words this means that the normalizer $N_G(C) = \{x \in G | xCx^{-1} = C\}=G$. So if $N_G(C)\neq G$, the component corresponding to $C$ will be $0$. So for now fix $C$ with $N_G(C)=G$ and set $u_x (d_C v_ C) u_x^{-1} = d_C v_C$ then we find $$\sigma_x(d_C)f(x,g) = d_C \ \ \ \forall g \in C.$$ If $d_C \neq 0$, then $f(x,g) = f(x,h)$ $\forall g,h \in C$, $\forall x \in G$. This in particular means that if $d_C \neq 0$, $f(x,g)$ with $g \in C$ is fully determined by one representative, say $s \in C$. We will call such $C$ constant under $f$. We define $$\Delta = \{C | C\textrm{ is a ray class in }W \textrm{ with } C \textrm{ constant under }f\}.$$ So from now on, fix $C \in \Delta$. We can now define $f(x, C)$ as $f(x, s)$ for some $s \in C$. Like in the Abelian case, we have two possibilities: $f(x, C)=1$, $\forall x \in G$ (we say $C \in G_{\textup{reg}}$) or $\exists x \in G: f(x,C)\neq 1$ (we say $C \notin G_{\textup{reg}}$). If $C \in G_{\textup{reg}}$ we see that $\sigma_x(d_C)=d_C$, $\forall x \in G$ or that $d_C \in D^G$. If $C \notin G_{\textup{reg}}$ we define for each $x \in G$ with $f(x,C)\neq 0$: $$I_{C,x}=\{d\in D | \sigma_x(d) = d \cdot f^{-1}(x,C)\},$$ and define $I_{C}$ to be the intersection of all $I_{C,x}$, $x\in G, f(x,C)\neq 0$. Similar to the Abelian case, our center now becomes: $$Z(D {\mathop \diamondsuit \limits_{\sigma ,\alpha}}G) = \sum_{C \in G_{\textup{reg}} \cap \Delta}D^G v_C + \sum_{C \in \Delta \backslash G_{\textup{reg}}}I_{C}v_C.$$ One can prove in exactly the same way as in the Abelian case:\ [$I_{C}$ as defined above is a finitely generated $D^G$-bimodule of rank $1$. It is not multiplicatively closed.]{}\ In order to have a trivial center, there are three possibilities: 1. $N_G(C)=G \Leftrightarrow C = \{e\}$. 2. $\Delta = \{C_e\}$. 3. $G_\textup{reg} = \{C_e\}$ and $I_C = \{0\}$, $\forall C \in \Delta \backslash G_{\textup{reg}}$. [Notice how similar these conditions are to having a trivial center in the Abelian case.]{}\ Now let us return to the bimodule considerations following Definition \[cen1\] now restricting to the situation where $R =D$ is a Dedekind domain. Since in a Dedekind domain any nonzero ideal is invertible, it follows that $I(g,g^{-1})$ as well as $I(g^{-1},g)$ are invertible and then $A_g A_{g^{-1}}=I(g,g^{-1})$ entails $A_g(A_{g^{-1}}I(g,g^{-1})^{-1}) = D$. One easily calculates: $$\begin{aligned} (A_{g^{-1}}I(g, g^{-1})^{-1})A_g &=& \sigma_{g^{-1}}(I(g,g^{-1})^{-1})A_{g^{-1}}A_g\\ \ &=& \sigma_{g^{-1}}(I(g, g^{-1})^{-1})I(g^{-1}, g).\end{aligned}$$ From $I(g, g^{-1})=\sigma_g(I(g^{-1},g))$ it then follows that the foregoing reduces to $$A_{g^{-1}}I(g,g^{-1})^{-1}A_g = I(g^{-1},g)^{-1}I(g^{-1},g) = D.$$ Consequently each $A_g$ is an invertible $D$-bimodule with inverse $A_{g^{-1}}I(g,g^{-1})^{-1}$. Clearly $A$ need not be strongly graded but the equation $$I(X,Y)I(XY,Z)= \sigma_X(I(Y,Z))I(X,YZ),\label{bla}$$ now gives rise to a map $$\phi:G \rightarrow \textup{Pic}(D): g \mapsto [A_g],$$ where $[A_g]$ denotes the class of the invertible $D$-bimodule $A_g$ in the class group (Picard group) $\textup{Pic}(D)$, together with a factor system $$I: G \times G \rightarrow \textup{Pic}(D):(g,h) \mapsto [I(g,h)].$$ The above equation (\[bla\]) $I$ defines a $2$-cocycle. Recall, $\textup{Pic}(D)$ is an Abelian group with respect to the operation induced by the tensor product, it may also be seen as the free Abelian group $\textup{Div}(D)$ generated by the prime ideals of $D$ modulo the subgroup generated by the principal ideals, in particular for a P.I.D. the Picard group is trivial. Hence $A$ may be viewed as a generalized crossed product $$A = \bigoplus_{g\in G} A_g,$$ defined by $\phi : G \rightarrow \textup{Pic}(D)$ and a $2$-cocycle $I$ in $\textup{Pic}(D)$, $I : G\times G \rightarrow \textup{Pic}(D)$. Let us call this type of crossed product a **class crossed product**. The notion of crystalline graded ring may be extended to include the cases where the $A_g$ are not necessarily free of rank $1$ but represent elements of the class group $\textup{Pic}(D)$ as above. We refer to such graded rings as being **arithmetically crystalline graded**. The following theorem is now just a rephrasing of the results we obtained: [\[cen11\]Let $A$ be crystalline graded by an Abelian group $G$ over a Dedekind domain $D$, then $Z(A)$ is arithmetically crystalline graded over a subgroup of $G$.]{}\ The statement generalizes to $A$ which are themselves arithmetically crystalline graded. A general definition, preceding those of generalized Weyl algebras and generalized crossed products was used by second author in some observations concerning strongly graded rings. The so-called $\delta$-strongly graded rings were characterized by $A_g A_{g^{-1}} = \delta_g$ being an invertible $R$-bimodule for each $g\in G$. Of course, if $R$ is not a Dedekind domain and in particular if it is not even commutative then $\textup{Pic}(R)$ is somewhat more complex to deal with, for example a $[M] \in \textup{Pic}(R)$ canonically defines an automorphism $\sigma_M$ of $Z(R)$ but not necessarily of $R$ (note that in the situation of Theorem \[cen11\] the $\sigma_{A_g}$ are exactly the $\sigma_g$ defined on $R$!). More detail about $\textup{Pic}(R)$ in the greater generality can be found in the book by H. Bass (K-theory), cf. [@B]. For us this will be useful in further work on the algebraic structure of general crystalline or arithmetically crystalline graded rings e.g. over Artinian algebras etc. [A]{} Bass, H., *Algebraic K-theory*, W. A. Benjamin, Inc. (1968), New York-Amsterdam. Bavula, V., *Generalized Weyl algebras and their representations*, Algebra i Analiz 4 (1992), no. 1, 75–97. English translation in St. Petersburg Mat. J. 4 (1993), no. 1, 71–92. Bavula, V.; Van Oystaeyen, F., *The simple modules of certain generalized crossed products*, J. of Algebra 194 (1997), no. 2, 521–566. Bavula, V.; Van Oystaeyen, F., *Simple holonomic modules over the second Weyl algebra*, Adv. Math. 150 (2000), no. 1, 80–116. Caenepeel, S.; Van Oystaeyen, F., *A note on generalized Clifford algebras and representations*, Comm. Algebra 17 (1989) no. 1, 93–102. Feit, W., *The representation theory of finite groups*, Dekker (1985), New York. McConnell, J.C.; Robson, J.C., *Noncommutative Noetherian rings*, John Wiley and Sons Ltd (1987), Brisbane. Nauwelaerts, E.; Van Oystaeyen, F., *The Brauer splitting theorem and projective representations of finite groups over rings*, J. Algebra 112 (1988), no. 1, 49–57. Nauwelaerts, E.; Van Oystaeyen, F., *Introducing crystalline graded algebras*, Algebras and Representation Theory vol 11(2008), no. 2, 133–148. Neijens, T.; Van Oystaeyen, F., *Centers of certain crystalline graded rings*, Algebras and Representation Theory, to appear. Nelis, P.; Van Oystaeyen, F., *The projective Schur subgroup of the Brauer group and root groups of finite groups*, J. of Algebra 137 (1991) no. 2, 501–518. Van Oystaeyen, F., *Azumaya strongly graded rings and ray classes*, J. of Algebra 103 (1986), no. 1, 228–240. Zmud, M., *Symplectic geometries and projective representations of finite Abelian groups*, (Russian) Mat. Sb. (N.S.) 87(129) (1972), 3–17.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A classical theorem by Ritt states that all the complete decomposition chains of a univariate polynomial satisfying a certain tameness condition have the same length. In this paper we present our conclusions about the generalization of these theorem in the case of finite coefficient fields when the tameness condition is dropped. *(Updated April 2008: see note at the beginning of the introduction.)*' author: - | Jaime Gutierrez[^1]\ Dpto. de Matemáticas, Estadística y Computación, Universidad de Cantabria\ E–39071 Santander, Spain\ [email protected] - | David Sevilla\ Department of Computer Science, Concordia University\ Montreal H3G 1M8, QC, Canada\ [email protected] title: 'On Ritt’s decomposition Theorem in the case of finite fields' --- Introduction ============ *(**Updated April 2008:** There exists a previously published article, of which the authors were not aware, where Ritt’s second theorem is shown to be true in all characteristics, provided we have the natural and necessary condition that no derivative vanishes identically. This goes much beyond the condition that $p$ does not divide the degrees (which is not a necessary condition). See additional references [@Zannier1993] and [@Schinzel2000]. We wish to thank Pr. Zannier for this information.)* Our starting point is the decomposition of polynomials and rational functions in one variable. First we define the basic concepts of this topic. Let ${\mathbb{K}}$ be any field, $x$ a transcendental over ${\mathbb{K}}$ and ${\mathbb{K}}(x)$ the field of rational functions in the variable $x$ with coefficients in ${\mathbb{K}}$. In the set $T={\mathbb{K}}(x)\setminus{\mathbb{K}}$ we define the binary operation of *composition* as $$g(x)\circ h(x)=g(h(x))=g(h).$$ We have that $(T,\circ)$ is a semigroup, the element $x$ being its neutral element. If $f=g\circ h$, we call this a *decomposition* of $f$ and say that $g$ is a *component on the left* of $f$ and $h$ is a *component on the right* of $f$. We call a decomposition *trivial* if any of the components is a unit with respect to decomposition. Given two decompositions $f=g_1\circ h_1=g_2\circ h_2$ of a rational function, we call them *equivalent* if there exists a unit $u$ such that $$h_1=u\circ h_2 \quad (\mbox{thus, } g_1=g_2\circ u^{-1}),$$ where the inverse is taken with respect to composition. Given $f\in T$, we say that it is *indecomposable* if it is not a unit and all its decompositions are trivial. We define a *complete* decomposition of $f\in{\mathbb{K}}(x)$ to be $f=g_1\circ\cdots\circ g_r$ where every $g_i$ is indecomposable. The notion of equivalent complete decompositions is straightforward from the previous concepts. Given a non–constant rational function $f(x)\in{\mathbb{K}}(x)$ where $f(x)=f_N(x)/f_D(x)$ with $f_N,f_D\in{\mathbb{K}}[x]$ and $(f_N,f_D)=1$, we define the *degree* of $f$ as $$\deg\,f=\max\{\deg\,f_N,\ \deg\,f_D\}.$$ We also define $\deg\,a=0$ when $a\in{\mathbb{K}}$. From now on, we will use the previous notation when we refer to the numerator and denominator of a rational function. Unless explicitly stated, we will take the numerator to be monic, even though multiplication by constants will not be relevant. Now we can properly state the problem of decomposition of univariate rational functions, although this will not be our main object of study. Given a univariate rational function, decide if it is decomposable, and in the affirmative case compute a non–trivial decomposition of the function. It is clear that the solution of this problem provides the computability of a complete decomposition of a function if it exists. Next, we introduce some basic results about univariate decomposition, see [@AGR95] for more details. \[prop-univ\] $ $ (i) : For every $f\in T$, $\deg\,f=[{\mathbb{K}}(x):{\mathbb{K}}(f)]$. (ii) : $\deg\,(g\circ h)=\deg\,g\cdot\deg\,h$. (iii) : $f(x)$ is a unit with respect to composition if and only if $\deg\,f=1$, that is, $f(x)=\displaystyle\frac{ax+b}{cx+d}$ with $a,b,c,d\in{\mathbb{K}}$ and $ad-bc\not=0$. (iv) : Every non–constant element of ${\mathbb{K}}(x)$ is cancelable on the right with respect to composition. In other words, if $f(x),h(x)\in T$ are such that $f(x)=g(h(x))$ then $g(x)$ is uniquely determined by $f(x)$ and $h(x)$. We can relate decomposition and Field Theory by means of the following classical result: \[th-luroth\] Let ${\mathbb{F}}$ be a field such that ${\mathbb{K}}\subset{\mathbb{F}}\subset{\mathbb{K}}(x)$. Then there exists $f\in{\mathbb{K}}(x)$ such that ${\mathbb{F}}={\mathbb{K}}(f)$. Also, if ${\mathbb{F}}$ contains a polynomial, $f$ can be chosen to be a polynomial. See for example [@Lur76] for a proof in the case ${\mathbb{K}}=\mathbb{C}$, [@Ste10] for one in the general case and [@Wae64] for an elementary one. Constructive proofs can be found in [@Net95], [@Sed86] and [@AGR95]. Now we state one of the classical Ritt’s theorems (see [@Rit22]) about the relations among the complete decompositions of a polynomial that satisfies a certain condition. First we have to define that condition. A polynomial $f\in{\mathbb{K}}[x]$ is *tame* when $\mathrm{char}\;{\mathbb{K}}$ does not divide $\deg\,f$. Ritt’s theorem essentially proves that all the decompositions have the same length and are related in a rather direct way. A *bidecomposition* is a 4-tuple of polynomials $f_1,g_1,f_2,g_2$ such that $f_1\circ g_1=f_2\circ g_2$, $(\deg\,f_1,\deg\,g_1)=1$ and $\deg\,f_1=\deg\,g_2$. \[ritt1\] Let $f\in{\mathbb{K}}[x]$ be tame and let $f=g_1\circ\cdots\circ g_r=h_1\circ\cdots\circ h_s$ be two complete decompositions of $f$. Then $r=s$, and the sequences $(\deg\,g_1,\ldots,\deg\,g_r)$, $(\deg\,h_1,\ldots,\deg\,h_s)$ are permutations of each other. Moreover, there exists a finite chain of complete decompositions $$f=f_1^{(j)}\circ\cdots\circ f_r^{(j)},\ j\in\{1,\ldots,k\},$$ such that $$f_i^{(1)}=g_i,\ f_i^{(k)}=h_i,\ i=1,\ldots,r,$$ and for each $j<k$, there exists $i_j$ such that the $j$-th and $(j+1)$-th decomposition differ only in one of these aspects: (i) : $f_{i_j}^{(j)}\circ f_{i_j+1}^{(j)}$ and $f_{i_j}^{(j+1)}\circ f_{i_j+1}^{(j+1)}$ are equivalent. (ii) : $f_{i_j}^{(j)}\circ f_{i_j+1}^{(j)}=f_{i_j}^{(j+1)}\circ f_{i_j+1}^{(j+1)}$ is a bidecomposition. See [@Rit22] for ${\mathbb{K}}=\mathbb{C}$, [@Eng41] for characteristic zero fields and [@FM69b] for the general case. In this paper we will study the generalization of this result to polynomials with coefficients in finite fields. To that end, we will also analyze the structure of intermediate fields between ${\mathbb{K}}(f)$ and ${\mathbb{K}}(x)$. It is already known that Ritt’s theorem is false when the tameness condition is dropped, see [@DW74] for a counterexample. Let $f=g(h)$. Then $f\in{\mathbb{K}}(h)$, thus ${\mathbb{K}}(f)\subset{\mathbb{K}}(h)$. Also, ${\mathbb{K}}(f)={\mathbb{K}}(h)$ if and only if $f=u\circ h$ for some unit $u$. This allows the following bijection among decompositions of a function $f$ and fields between ${\mathbb{K}}(f)$ and ${\mathbb{K}}(x)$: Let $f\in{\mathbb{K}}(x)$. In the set of decompositions of $f$ we have the equivalence relation given by the definition of equivalence of decompositions. If we denote as $[(g,h)]$ the class of the decomposition $f=g(h)$, the we have then the bijection: $$\begin{array}{ccc} \{[(g,h)]:f=g(h)\} & \longleftrightarrow & \{{\mathbb{F}}:{\mathbb{K}}(f)\subset{\mathbb{F}}\subset{\mathbb{K}}(x)\} \\ \left[(g,h)\right] & \longleftrightarrow & {\mathbb{F}}={\mathbb{K}}(h). \end{array}$$ Thanks to the Primitive Element Theorem (see for example [@Lan67]), we know that for each non–constant $f\in{\mathbb{K}}(x)$ there exist finitely many fields between ${\mathbb{K}}(f)$ and ${\mathbb{K}}(x)$. Due to the second part of Lüroth’s Theorem, every rational decomposition of a polynomial is equivalent to a decomposition whose components are polynomials. Therefore it suffices to care about polynomial decomposition in this case. In Section 2 we introduce several elementary results about univariate function fields that arise from Galois theory. In Section 3 we present a function that is fixed by all the automorphisms of a univariate function field over a finite field and several results related to it. In particular, we provide an essentially new counterexample of Ritt’s theorem for finite coefficient fields. The fixing group and the fixed field ==================================== In this section we introduce several simple notions from the classical Galois theory. Let $\Gamma({\mathbb{K}})=\mathrm{Aut}_{\mathbb{K}}{\mathbb{K}}(x)$ (we will write simply $\Gamma$ if there can be no confusion about the field). The elements of $\Gamma({\mathbb{K}})$ can be identified with the images of $x$ under the automorphisms, that is, with Möbius transformations (non–constant rational functions of the form $(ax+b)/(cx+d)\in{\mathbb{K}}(x)$), which are also the units of ${\mathbb{K}}(x)$ under composition. $ $ - Let $f\in{\mathbb{K}}(x)$. We define $G(f)=\{u\in\Gamma({\mathbb{K}}):\ f\circ u=f\}$. - Let $H<\Gamma({\mathbb{K}})$. We define ${\mathrm{Fix}}(H)=\{f\in{\mathbb{K}}(x):\ f\circ u=f\ \forall u\in H\}$. This definitions correspond to the classical Galois correspondences (not bijective in general) between the intermediate fields of an extension and the subgroups of its automorphism group, as the following diagram shows: $$\begin{array}{ccc} {\mathbb{K}}(x) & \longleftrightarrow & \{id\} \\ | & & | \\ {\mathbb{K}}(f) & \longrightarrow & G(f) \\ | & & | \\ Fix(H) & \longleftarrow & H \\ | & & | \\ {\mathbb{K}}& \longleftrightarrow & \Gamma \\ \end{array}$$ As ${\mathbb{K}}(f)={\mathbb{K}}(f')$ if and only if $f=u\circ f'$ for some unit $u$, we have that the application ${\mathbb{K}}(f)\mapsto G(f)$ is well–defined. We are interested in the computability of these elements, the following results solves one of the two parts of this question. \[gen-fixed-field\] Let $H=\{h_1,\ldots,h_m\}\subset{\mathbb{K}}(x)$ be a finite subgroup of $\Gamma$. Let $P(T)=\prod_1^m (T-h_i)\in{\mathbb{K}}(x)[T]$. Then any non–constant coefficient of $P(T)$ generates ${\mathrm{Fix}}(H)$. [Sketch of proof.]{} It can be shown that $P(T)$ is the minimal polynomial of $x$ over ${\mathrm{Fix}}(H)\subset{\mathbb{K}}(x)$. Then, a known proof of Lüroth’s theorem (see [@Net95]) gives the desired result. The previous theorem obviously provides an algorithm to compute the fixed field for a given finite subgroup of $\Gamma$: compute the symmetric elementary functions in $h_1,\ldots,h_m$ until a non–constant one is found. About the computation of the fixing group, an elementary but inefficient algorithm is given by the resolution of the equations given by $$f(x)-f\left(\frac{ax+b}{cx+d}\right)=0$$ in terms of $a,b,c,d$. Another algorithm (see [@Sev04]) combines this idea with certain normalization of the rational function, which simplifies the equations substantially. Next, we state several interesting properties of the fixed field and the fixing group, see [@Sev04] for details. \[H-inf-fin\] Let $H<\Gamma$. - $H$ is infinite $\Rightarrow {\mathrm{Fix}}(H)={\mathbb{K}}$. - $H$ is finite $\Rightarrow{\mathbb{K}}\varsubsetneq{\mathrm{Fix}}(H)$, ${\mathrm{Fix}}(H)\subset{\mathbb{K}}(x)$ is a normal extension, and in particular ${\mathrm{Fix}}(H)={\mathbb{K}}(f)$ with $\deg\,f=|H|$. \[props-fix\]$ $ (i) : Given a non–constant $f\in{\mathbb{K}}(x)$, $|G(f)|$ divides $\deg\,f$. Moreover, for any field ${\mathbb{K}}$ there is always a function $f\in{\mathbb{K}}(x)$ such that $1<|G(f)|<\deg\,f$. (ii) : $|G(f)|=\deg\,f\Rightarrow{\mathbb{K}}(f)\subseteq{\mathbb{K}}(x)$ is normal. Moreover, if the extension ${\mathbb{K}}(f)\subseteq{\mathbb{K}}(x)$ is separable, then $${\mathbb{K}}(f)\subseteq{\mathbb{K}}(x)\mathrm{\ is\ normal}\Rightarrow|G(f)|=\deg\,f.$$ (iii) : Given a finite subgroup $H$ of $\Gamma$, there is a bijection between the subgroups of $H$ and the fields between ${\mathrm{Fix}}(H)$ and ${\mathbb{K}}(x)$. Also, if ${\mathrm{Fix}}(H)={\mathbb{K}}(f)$, there is a bijection between the right components of $f$ (up to equivalence by units) and the subgroups of $H$. For the first item, we take $f=x^2\,(x-1)^2$ gives $G(f)$={x,1-x}. The other ones are straightforward. Finite fields ============= In this section, ${\mathbb{K}}={\mathbb{F}}_q$ where $q=p^m$ and $p=\mathrm{char}\;{\mathbb{F}}_q$, see [@LN97] for several useful results. As before, we will denote $\Gamma=\Gamma({\mathbb{F}}_q)$. For any ${\mathbb{K}}$, $\Gamma_0=\Gamma\cap{\mathbb{K}}[x]=\{ax+b:\ a\in{\mathbb{K}}^*,\ b\in{\mathbb{K}}\}$. ${\mathbb{K}}(x)$ is Galois over ${\mathbb{K}}$ (that is, the only functions fixed by $\Gamma({\mathbb{K}})$ are the constants) if and only if ${\mathbb{K}}$ is infinite. The “if” part is the first part of Theorem \[H-inf-fin\]. The “only if” part is a consequence of Theorem \[gen-fixed-field\], as $\Gamma({\mathbb{K}})$ is finite whenever ${\mathbb{K}}$ is finite. The interest of $\Gamma$ and $\Gamma_0$ in the case of finite fields lies in the fact that both groups provide non–trivial fixed fields. The fixed field for $\Gamma_0$ is generated by $(x^q-x)^{q-1}$. According to Theorem \[gen-fixed-field\] any non–constant coefficient of $Q(T)=\prod_{u\in\Gamma_0}(T-u)$ generates the field. But the constant term of $Q$ is precisely $\prod_{u\in\Gamma_0}u=(x^q-x)^{q-1}$. From now on, we will denote $P_q=(x^q-x)^{q-1}$. As $\Gamma_0\subset\Gamma$, if $f$ generates the fixed field for $\Gamma$ then $f=h(P_q)$ for some $h\in{\mathbb{K}}(x)$. Moreover, $h$ has degree $[\Gamma:\Gamma_0]=q+1$. Let $$h_q=(x^{q+1}+x+1)/x^q.$$ Then the rational function $f_q=h_q(P_q)$ generates ${\mathrm{Fix}}(\Gamma)$. It is easy to prove that $\Gamma_0\cup\{1/x\}$ generates $\Gamma$. As $f_q$ is a function of $P_q$ and its degree is equal to the order of the group, it suffices to show that $f_q(1/x)=f_q(x)$. A simple computation shows that this is indeed the case: let $y=x^{q-1}$. Then $P_q(x)=y(y-1)^{q-1}$ and $P_q(1/x)=(y-1)^{q-1}/y^q$. Thus, $$\begin{array}{cl} & f_q(1/x)-f_q(x)= \\ \\ = & \displaystyle\frac {\displaystyle\frac{(y-1)^{q^2-1}}{y^{q^2+q}}+\frac{(y-1)^{q-1}}{y^q}+1} {\displaystyle\frac{(y-1)^{q^2-q}}{y^{q^2}}} - \frac {y^{q+1}(y-1)^{q^2-1}+y(y-1)^{q-1}+1} {y^q(y-1)^{q^2-q}} = \\ = & \displaystyle\frac { (y-1)^{q^2-1} + y^{q^2}(y-1)^{q-1} + y^{q^2+q} - y^{q+1}(y-1)^{q^2-1} - y(y-1)^{q-1} - 1 } { y^q(y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ (y-1)^{q^2-1}(1-y^{q+1}) + (y-1)^{q-1}(y^{q^2}-y) + y^{q^2+q}-1 }{ y^q(y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ (y-1)^{q^2-1}(1-y^{q+1}) + (y-1)^{q-1}((y-1)^{q^2}-(y-1)) + y^{q^2+q}-1 }{ y^q(y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ (y-1)^{q^2-1}(1-y^{q+1}+(y-1)^q) - (y-1)^q + y^{q^2+q}-1 }{ y^q(y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ (y-1)^{q^2-1}(1-y^{q+1}+y^q-1) - (y-1)^q + (y^{q+1}-1)^q }{ y^q(y-1)^{q^2-q} } = \\ \end{array}$$ $$\begin{array}{cl} = & \displaystyle\frac{ -(y-1)^{q^2}y^q - (y-1)^q + (y-1)^q(1+y+\cdots+y^q)^q }{ y^q(y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ -(y-1)^{q^2}y^q + (y-1)^q(y+\cdots+y^q)^q }{ y^q(y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ -(y-1)^{q^2} + (y-1)^q(1+\cdots+y^{q-1})^q }{ (y-1)^{q^2-q} } = \\ = & \displaystyle\frac{ -(y-1)^{q^2} + (y^q-1)^q }{ (y-1)^{q^2-q} } = 0. \end{array}$$ Let $f\in{\mathbb{F}}_q(x)$. Let $\mathcal{C}=\{{\mathbb{K}}: {\mathbb{F}}_q\subseteq{\mathbb{K}}\subseteq{\mathbb{F}}_q(x)\}$ and $$\begin{array}{cccc} \phi: & \mathcal{C} & \longrightarrow & \mathcal{C} \\ & {\mathbb{F}}_q(f) & \rightarrow & {\mathrm{Fix}}(G(f))={\mathbb{F}}_q(f') \end{array}$$ which is a well–defined application. Then it is easy to check that $f'$ is a (not necessarily proper) right–component of $f$. Also, as $G(f)\subset\Gamma$, $f'$ is a right–component of $f_q$. Thus, ${\mathbb{F}}_q(f)\subseteq{\mathbb{F}}_q(f')$ and ${\mathbb{F}}_q(f_q)\subseteq{\mathbb{F}}_q(f')$, therefore ${\mathbb{F}}_q(f,f_q)\subseteq{\mathbb{F}}_q(f')$. ${\mathbb{F}}_q(f,f_q)={\mathbb{F}}_q(f')$. Let ${\mathbb{F}}_q(f,f_q)={\mathbb{F}}_q(m)$. Then there is a rational function $r(x,y)$ such that $r(f,f_q)=m$. For every $u\in G(f)$, $m\circ u=r(f\circ u,f_q\circ u)=r(f,f_q)=m$. Therefore, $m\in{\mathrm{Fix}}(G(f))={\mathbb{F}}_q(f')\ \Rightarrow\ {\mathbb{F}}_q(m)\subseteq{\mathbb{F}}_q(f')$. The other part is straightforward. The polynomial $P_q$ has at least two different decompositions: $$P_q=x^{q-1}\circ (x^q-x)=\left( x(x-1)^{q-1}\right) \circ x^{q-1}.$$ This gives at least two decompositions for $h_q$, both involving the component $\displaystyle\frac{x^{q+1}+x+1}{x^q}$. $ $ (i) : $\displaystyle\frac{x^{q+1}+x+1}{x^q}$ is indecomposable. (ii) : $x^q-x$ is decomposable iff q is composite, that is, $q=p^m$ with $m\geq 2$. (iii) : $x(x-1)^{q-1}$ is indecomposable. <!-- --> (i) : We will prove that for certain units $u,v\in{\mathbb{F}}_q(x)$, the function $$u\circ\frac{x^{q+1}+x+1}{x^q}\circ v$$ is indecomposable. In particular, let $u=x+1,v=1/(x-1)$. Then $$u\circ\frac{x^{q+1}+x+1}{x^q}\circ v=\frac{x^{q+1}}{x-1}.$$ As the degree is multiplicative with respect to composition, and so is the difference in the degrees of numerator and denominator (see [@Sev04 Theorem 1.14 and Corollary 1.15]), there is no possible decomposition for this function and the original function is also indecomposable. (ii) : As $G({x^q-x})=\{x-a:\ a\in{\mathbb{F}}_q\}$ and $|G(x^q-x)|=q=\deg\ x^q-x$, by Theorem \[props-fix\] there is a bijection between the decompositions of $x^q-x$ and the subgroups of its fixing group. But $G(x^q-x)$ has proper subgroups if and only if its order is composite. (iii) : Let $q=p^m$. Let $x(x-1)^{q-1}=g(h)$ with $g=x^{p^r}+g_0,\deg\,g_0\leq p^r-1$ and $h=x^{p^s}+h_0,\deg\,h_0\leq p^s-1$. Then $$g\circ h=h^{p^r}+g_0\circ h=(x^{p^s}+h_0)^{p^r}+g_0\circ h=x^q+{h_0}^{p^r}+g_0\circ h$$ with $\deg\,{h_0}^{p^r}\leq q-p^r$ and $\deg\,g_0\circ h\leq q-p^s$. But $$x(x-1)^{q-1}=x^q+x^{q-1}+\ldots+x^2+x,$$ thus either $r=0$ or $s=0$ and the decomposition is trivial. If $q$ is not prime, $P_q$ has two complete decomposition chains of different length. As there is a bijection between the subgroups of $\Gamma_0$ and the components of $(x^q-x)^{q-1}$ on the right, we will study those subgroups in order to determine whether this polynomial has complete decompositions of different length when $q$ is prime. $H_0=\{x+b:\ b\in{\mathbb{F}}_q\}$. $\Gamma_0$ is the semidirect product of $H_0$ and $\{ax:\ a\in{\mathbb{F}}_q^*\}$. Let $G$ be a subgroup of $\Gamma_0$. As $H_0$ has prime order, we have two cases: - $G\cap H_0=H_0$. Then $H_0\subseteq G$. If $ax+b\in G$, then for every $b'\in{\mathbb{F}}_q$ we have $ax+b'\in G$. In particular, $ax\in G$, and $G_0=\{a\in{\mathbb{F}}_q^*:\ ax\in G\}<{\mathbb{F}}_q^*$. But ${\mathbb{F}}_q^*$ is cyclic of order $q-1$, thus $G_0$ is cyclic of order $m\,|\,q-1$. In this case, $G=H_0\rtimes G_0\cong C_q\rtimes C_m$. - $G\cap H_0=\{x\}$. Then for every $a\in G_0$ there exists exactly one $b\in{\mathbb{F}}_q$ such that $ax+b\in G$, because $(ax+b)\circ(ax+b')^{-1}=x-b'+b$. As $G_0$ is cyclic, we have that $G$ is generated by some $a_0x+b_0$ where $a_0$ generates $G_0$ and $b_0\in{\mathbb{F}}_q$. This allows to prove the following theorem. If $q$ is prime, then all the maximal chains of subgroups of $\Gamma_0({\mathbb{F}}_q)$ have the same length. Let $G_0=\{x\}<G_1<\ldots<G_n=\Gamma_0({\mathbb{F}}_q)$ be a maximal chain. Let $i\in\{1,\ldots,n\}$ be such that $G_{i-1}\cap H_0=\{x\}$ and for all $j\geq i$, $H_0\subseteq G_j$. For each $j\geq i$ there exists a cyclic group $C_i$ of order $m_i$ with $m_i\,|\,q-1$ such that $G_i=H_0\rtimes C_i$. Thus, the numbers $m_i,m_{i+1},\ldots,m_n$ are a maximal chain of divisors of $q-1$ greater or equal than $m_i$. On the other hand, $G_{i-1}$ must be a cyclic group of order $m_i$, therefore the orders of $G_1,\ldots,G_{i-1}$ are a maximal chain of divisors of $m_i$. Therefore, the length of the chain $G_0,\ldots,G_n$ is equal to the number of prime factors in a complete factorization of $q-1$ plus two. The polynomial $(x^q-x)^{q-1}\in{\mathbb{F}}_q[x]$ has maximal decomposition chains of different lengths iff $q$ is not prime. It is possible to determine all the subgroups of $\Gamma({\mathbb{F}}_q)$ by finding all subgroups of $GL(2,q)$. Then all chains of subgroups can be computed, finding out whether the function $f$ has decompositions of different lengths. Conclusions =========== The results in the last section show some new information about the structure of decompositions of rational functions in the finite case; it is our hope that more can be said about possible versions of Ritt’s theorems for finite fields. Also, the algorithms presented here indicate that fast decomposition algorithms in the finite case can be achievable, by using this structure. [99]{} C. Alonso, J. Gutierrez, T. Recio, *A rational function decomposition algorithm by near-separated polynomials*. J. Symbolic Comput. 19 (1995), no. 6, 527–544. F. Binder, *Polynomial decomposition*. Master’s thesis, University of Linz, June 1995. C. Corrales-Rodrigáñez, *A note on Ritt’s theorem on decomposition of polynomials*. J. Pure Appl. Algebra 95 (1990), 293–296. F. Dorey, G. Whaples, *Prime and composite polynomials*. J. Algebra 28 (1974), 88–101. H. T. Engström, *Polynomial substitutions*. Amer. J. Math. 63 (1941), 249–255. M. Fried, R. Mac Rae, *On the invariance of chains of fields*. Illinois J. Math. 13 (1969), 165–171. S. Lang, *Algebra*. Addison–Wesley, Reading, Mass (1967). R. Lidl and H. Niederreiter, [*Finite fields*]{}, Cambridge University Press, Cambridge, 1997. P. Lüroth, *Beweis eines Satzes über rationale curven*. Math. Ann. 9 (1876), 163–165. E. Netto, *Ünber einen Lüroth-Gordaschen Satz*. Math. Ann. 9 (1895), 310–318. J. F. Ritt, *Prime and composite polynomials*. Trans. Amer. Math. Soc. 23 (1922), no. 1, 51–66. A. Schinzel, *Selected topics on polynomials*. Ann Arbor, University of Michigan Press, 1982. T. W. Sedeberg, *Improperly parametrized rational curves*. Computer Aided Geometric Design 3 (1986), 67–75. D. Sevilla, *Ritt’s Theorems and computation of unirational fields*. Ph. D. Thesis Dissertation. University of Cantabria, 2004. E. Steinitz, *Algebraische Theorie der Körper*. Journal für reine und angewante Mathematik, 137 (1910), 167–309. B. L. van der Waerden,*Modern Algebra*. Frederick Ungar Publishing Co., New York, 1964. R. Zippel, *Rational function decomposition*. Proceedings of ISSAC’91, ACM Press, 1991. U. Zannier, *Ritt’s second theorem in arbitrary characteristic*, Journal reine angew. Math., 445 (1993), 175–203. A. Schinzel, *Polynomials with Special Regard to Reducibility*. With an appendix by Umberto Zannier. Encyclopedia of Mathematics and its Applications, 77, Cambridge University Press, Cambridge, 2000. [^1]: Both authors are partially supported by Research Project MTM2004-07086 of the Spanish Ministry of Science and Technology
{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper formulates a novel problem on graphs: find the minimal subset of edges in a fully connected graph, such that the resulting graph contains all spanning trees for a set of specified subgraphs. This formulation is motivated by an unsupervised grammar induction problem from computational linguistics. We present a reduction to some known problems and algorithms from graph theory, provide computational complexity results, and describe an approximation algorithm.' author: - 'Nicholas Harvey$^1$ David Karger$^3$ Vahab Mirrokni$^4$ Virginia Savova$^2$ Leonid Peshkin$^{2,*}$' bibliography: - 'Notes.bib' title: Matroids Hitting Sets and Unsupervised Dependency Grammar Induction --- $^1$ - Dept of Computer Science, Univ of British Columbia, Vancouver, Canada\ $^2$ - Systems Biology Dept, Harvard Medical School, Boston, USA\ $^3$ - Dept of Computer Science, MIT, Boston, USA $^4$ - Google $^*$ - corresponding author [email protected] \ Introduction ============ Linguistic representations of natural language syntax arrange syntactic dependencies among the words in a sentence into a tree structure, of which the string is a one dimensional projection. We are concerned with the task of analyzing a set of several sentences, looking for the most parsimonious set of corresponding syntactic structures, solely on the basis of co-occurrence of words in sentences. We proceed by first presenting an example, then providing a general formulation of dependency structure and grammar induction. ![An illstration of a dependency structure.[]{data-label="fig:enc"}](figuredep1){width="9cm"} Consider a sentence [*"Her immediate predecessor suffered a nervous breakdown.*]{}" A dependency grammar representation of this sentence shown in Figure \[fig:enc\] captures dependency between the subject, the object and the verb, as well as dependency between the determiner and the adjectives and their respective nouns. In this sentence, the subject *predecessor* and the object *breakdown* are related to the verb *suffered*. The verb *suffered* is the root of the dependency structure, that is illustrated in the diagram by a link to the period. Figure \[fig:tree\] left represents the same dependency structure in a different way by ignoring the direction. Instead the dependence is related to the relative depth in the tree. In a dependency tree, each word is the mother of its dependents, otherwise known as their [head]{}. To linearize the dependency tree in Figure \[fig:tree\].left into a string, we introduce the dependents recursively next to their heads:\ iteration 1: suffered\ iteration 2: predecessor suffered breakdown\ iteration 3: her predecessor suffered a breakdown. Dependency and the related link grammars have received a lot of attention in the field of computational linguistics in recent years, since these grammars enable much easier parsing than alternatives that are more complex lexicalized parse structures. There are applications to such popular tasks as machine translation and information retrieval. However, all of the work is concerned with parsing, i.e. inducing a parse structure given a corpus and a grammar, rather than with grammar induction. Some work is concerned with inducing parameters of the grammar from annotated corpora, for example see work by Eisner on dependency parsing [@eisner96] or more recent work by McDonald et al. [@Pereira04] and references therein. It has been pointed out [@Pereira04] that parsing with dependency grammars is related to Minimal Spanning Tree algorithms in general and in particular Chu-Liu-Edmonds MST algorithm was applied to dependency parsing. An established computational linguistics textbook has the following to say on the subject [@manning99foundations]: [*“... doing grammar induction from scratch is still a difficult, largely unsolved problem, and hence much emphasis has been placed on learning from bracketed corpora.”*]{} If grammar is not provided to begin with, parsing has to be done concurrently with learning the grammar. In the presence of grammar, among all the possibilities one needs to pick a syntactic structure consistent with the grammar. In the absence of grammar, it makes sense to appeal to Occam’s razor principle and look for the minimal set of dependencies which are consistent among themselves. More formally, a dependency grammar consists of a lexicon of terminal symbols (words), and an inventory of dependency relations specifying inter-lexical requirements. A string is generated by a dependency grammar if and only if: - Every word but one (ROOT) is dependent on another word. - No word is dependent on itself either directly or indirectly. - No word is directly dependent on more than one word. - Dependencies do not cross. Unlike the first three constraints, the last constraint is a linearization constraint, usually introduced to simplify the structure and is empirically problematic. The structure in figure \[fig:tree\].left is an example of so-called projective parse, in which dependency links mapped onto the sentences word sequence do not cross. Figure \[fig:tree\].right illustrates an incorrect parse of the sentence with non-projective dependancies: “her”$\rightarrow$“suffered” is crossing “a”$\rightarrow$“predecessor”). While the vast majority of English sentences observe the projectivity constraint, other languages allow much more flexibility in word order. Non-projective structures include wh-relative clauses [@Pike43], parentheticals [@McCawley82], cross-serial constructions of the type found in Dutch and Swiss-German [@Ojeda88], as well as free or relaxed word order languages [@Pullum82]. Therefore, it is interesting whether grammar induction can be performed without regard to word order. A truly cross-linguistic formulation of dependency parsing corresponds to finding a spanning tree (parse) in a completely connected subgraph of word nodes and dependency edges. The grammar induction problem in the same setting corresponds to inducing the minimal fully-connected subgraph which contains spanning trees for all sentences in a given corpus. Consider three sentences: “Her immediate predecessor suffered a nervous breakdown.”, “Her predecessor suffered a stroke.”, “It is a nervous breakdown.” Intuitively, the repetition of a word cooccurrence informs us about grammatical co-dependence. Here is a formulation of the grammar induction problem as an optimization problem: Given a lexicon $V$ and a set of $k$ sentences $S_1,\ldots ,S_k$ s.t. $S_i \subset V$ (a.k.a. [*corpus*]{}) the objective is to find the most parsimonious combination of dependency structures. i.e. such set of spanning trees for all $S_i$ that has the minimal cardinality of a joint set of edges. In section \[sptree\] of this paper, we formally introduce the related graph-theoretic problem. In section \[lognhard\] we show that the problem is hard to approximate within a factor of $c \log n$ for weighted instances, and hard to approximate within some constant factor (APX-hard) for unweighed instances. In section \[matroids\], we generalize the problem to matroids. Here we prove that the problem is hard to approximate within a factor of $c \log n$, even for unweighed instances. We conclude with a positive result – an algorithm for the matroid problem which constructs a solution whose cardinality is within $O(\log n)$ of optimal. ![[**top:**]{} An instance of a problem for a graph $G$ consisting of two sub-graphs $G_1$ and $G_2$; [**bottom:**]{} examples of a correct solution on the left (green) and two incorrect solutions on the right (red). []{data-label="MSFexample"}](Fig3color.pdf){width="8.5cm"} The Problem for Spanning-Trees {#sptree} ============================== Let $G=(V,E)$ be a graph and let $S_1, \ldots, S_k$ be arbitrary subsets of $V$. Our objective is to find a set of edges $F \subseteq E$ such that - $F$ contains a spanning tree for each induced subgraph $G[S_i]$, and - $|F|$ is minimized. We call this the Min Spanning-Tree Hitting Set problem. Figure \[MSFexample\] illustrates one instance of this problem. A graph $G$ consist of two sub-graphs $G_1$ and $G_2$. We present one possible correct solution on the left ($|F|$ = 4) and two sample incorrect solutions ($|F|$ = 5) on the right. The Min Spanning-Tree Hitting Set problem may be generalized to include a weight function $w$ on the edges of $G$. The objective for the weighted problem is the same as before, except that we seek to minimize $w(F)$. Notice that the problem initially appears similar to the group Steiner problem [@Hauptmann2013ACO], since the objective is to connect certain subsets of the nodes. However, our condition on the subgraph is slightly different: we require that the given subsets of nodes are *internally* connected. To develop some intuition for this problem, let’s analyze a simple greedy ad-hoc solution: first, assign all the edges weight equivalent to the number of sub-graphs it is included into, i.e. count the frequency of node pairs in the input set; then fragment the graph into subgraphs, keeping the weights and run the standard MST algorithm, to find a spanning tree for each subgraph. Figure \[txt4dcmp\] presents a counterexample to simple heuristics approaches. The following sub-sets make up the input as indicated via edges of distinct color and pattern in the figure: $\{ 1,4,5 \}, \{ 2,4,5 \},$ $\{ 3,4,5 \}, \{ 1,4 \},$ $\{ 1,5 \}$, $\{ 2,4 \}, \{ 2,5 \},$ $\{ 3,4\}, \{ 3,5 \}$. The optimal solution to this instance does not contain the edge $\{ 4,5 \}$, yet this edge is a member of the most (namely three) sub-sets. ![A counterexample to simple heuristics approaches.[]{data-label="txt4dcmp"}](SpanFrstSmpl.pdf){width="6cm"} Hardness for Weighted Instances {#lognhard} ------------------------------- We now show that the weighted problem is NP-hard to approximate within a factor of $\log n$. To do so, we exhibit a reduction from Min Hitting Set, which is known to be hard to approximate within $\log n$. An instance of Min Hitting Set consists of a universe $U = \{u_1, \ldots, u_n\}$ and a collection of sets $\mathcal{T} = \{T_1, \ldots, T_m\}$, each of which is a subset of $U$. We construct a weighted instance of Min Spanning-Tree Hitting Set as follows. Let $r \not \in V$ be a new vertex. We set $$\begin{aligned} V =& U + r \\ E =& K_U \cup \left\{ \; \{r,u_i\} \, : \, \mbox{for all $u_i \in U$} \; \right\} \\ S_{\{i,j\}} =& \{u_i,u_j\} \, \, \, \, \mbox{for $1 \leq i < j \leq n$} \\ S'_i =& T_i + r \, \, \, \, \mbox{for $1 \leq i \leq m$},\end{aligned}$$ where $K_U$ denotes (the edges of) the complete graph on vertex set $U$. The edges belonging to $K_U$ have weight $1$ and the edges incident with $r$ have weight $n^3$. Let $h$ denote the minimum weight of a Spanning-Tree Hitting Set in $G$. Let $h'$ denote the minimum cardinality of a Hitting Set for $\mathcal{T}$. \[reduction\] $h = h' \cdot n^3 + \binom{n}{2}$. First we show that $h' \leq \big(h - \binom{n}{2} \big)/n^3$ Let $F$ be a spanning-tree hitting set. Clearly $K_U \subseteq F$, because of the sets $S_{\{i,j\}}$. So all edges in $F \setminus K_U$ are of the form $\{r,u_i\}$. Now define $C = \left\{\; u_i \,:\, \{r,u_i\} \in F \;\right\} $. We now show that $C$ is a hitting set. Consider a set $T_i$. Since $F$ contains a spanning tree for $S'_i$, it must contain some edge $\{r,u_i\}$. This shows that $C$ hits the set $T_i$. Now we show that $h \leq h' \cdot n^3 + \binom{n}{2}$. Let $C \subseteq U$ be a hitting set for $\mathcal{T}$. Let $F = K_U \cup \left\{\; \{r,u_i\} \,:\, \mbox{for all $u_i \in C$} \;\right\}$. We now show that $F$ is a spanning-tree hitting set. Each set $S_{\{i,j\}}$ is clearly hit by the set $K_U$. So consider a set $S'_i = T_i + r$. All edges $\{u_a,u_b\}$ with $a, b \in T_i$ are contained in $K_U$. Furthermore, since $C$ is a hitting set, there exists an element $u_a \in T_i \cap C$. This implies that $\{r,u_a\} \in F$, and hence $F$ contains a spanning tree for $G[S'_i]$. Given an instance $\mathcal{T}$ of Hitting Set, it is NP-hard to decide whether $OPT(\mathcal{T}) \leq f(n)$ or $OPT(\mathcal{T}) > \alpha \log n \cdot f(n)$ for some constant $\alpha > 0$ and some function $f$. To prove $\log n$-hardness of Min Spanning-Tree Hitting Set, we must similarly show that for any instance $y$, there exists a constant $\beta>1$ and a function $g$ such that it is NP-hard to decide whether $OPT(G) \leq g(y)$ or $OPT(G) > \beta \log n \cdot g(y)$. From our reduction, we know that it is NP-hard to distinguish between $$OPT(G) \leq f(n) \cdot n^3 + {\textstyle \binom{n}{2}}$$ $\text{\quad or \quad}$ $$OPT(G) > \alpha \log n \cdot f(n) \cdot n^3 + {\textstyle \binom{n}{2}}.$$ Now note that $$\begin{aligned} \frac{\alpha \log n \cdot f(n) \cdot n^3 + {\textstyle \binom{n}{2}}} {f(n) \cdot n^3 + {\textstyle \binom{n}{2}}} &= \\ \frac{\alpha \log n \cdot \big(f(n) \cdot n^3 + {\textstyle \binom{n}{2}}/(\alpha \log n) \big)} {f(n) \cdot n^3 + {\textstyle \binom{n}{2}}} &= \\ \alpha \log n \cdot \Bigg( 1 - \frac{{\textstyle \binom{n}{2}}\cdot \big(1- 1/ \alpha \log n) \big)} {f(n) \cdot n^3 + {\textstyle \binom{n}{2}}} \Bigg) & \geq \beta \log n\end{aligned}$$ for some constant $\beta > 0$. Letting $g(y) = f(n) \cdot n^3 + {\textstyle \binom{n}{2}}$, it follows that Min Spanning-Tree Hitting Set is NP-hard to approximate within $\log n$. Hardness for Unweighted Instances {#apxhard} --------------------------------- We show APX-hardness for the unweighted problem via a reduction from Vertex Cover. The approach is similar to the construction in Section \[lognhard\]. Suppose we have an instance $G'=(V',E')$ of the Vertex Cover problem. We use the fact that Vertex Cover is equivalent to Min Hitting Set where $U = E'$ and $\mathcal{T} = E'$. The construction differs only in that $E'$ is used in place of the edge set $K_U$; the sets $S_{\{i,j\}}$ are adjusted accordingly. Let $h$ denote the minimum cardinality of a Spanning-Tree Hitting Set in $G$. Let $c$ denote the minimum cardinality of a Vertex Cover in $G'$. A claim identical to Claim \[reduction\] shows that $h = c + \lvert E' \rvert$. Recall that Vertex Cover is APX-hard even for constant-degree instances; see, e.g., Vazirani [@Vazirani §29]. So we may assume that $ \lvert E' \rvert \leq \frac{d}{2} \lvert V' \rvert$. Given an instance $G'=(V',E')$ of Vertex Cover with degree at most some constant $d$, it is NP-hard to decide whether $OPT(G') \leq \alpha' {{\lvert V' \rvert}}$ or $OPT(G') > \beta' \lvert V' \rvert$ for some constant $\alpha' < \beta'$. To prove APX-hardness of Min Spanning-Tree Hitting Set, we must similarly show that for any instance $G$, there exists a constant $\gamma>1$ such that it is NP-hard to decide whether $OPT(G) \leq f(G)$ or $OPT(G) > \gamma f(G)$. From our reduction, we know that it is NP-hard to distinguish between $$OPT(G) \leq \alpha' (\lvert V \rvert -1) + (\lvert E \rvert - \lvert V \rvert + 1)$$ $\text{\quad or \quad}$ $$OPT(G) > \beta' (\lvert V \rvert -1) + (\lvert E \rvert - \lvert V \rvert + 1).$$ Now note that $$\begin{aligned} \frac{\beta' ({{\lvert V \rvert}}-1) + (\lvert E \rvert - \lvert V \rvert + 1)} {\alpha' (\lvert V \rvert -1) + (\lvert E \rvert - \lvert V \rvert + 1)} &= \\ 1 + \frac{(\beta' - \alpha') (\lvert V \rvert -1)} {\alpha' (\lvert V \rvert -1) + (\lvert E \rvert - {{\lvert V \rvert}} + 1)} &\geq \\ 1 + \frac{(\beta' - \alpha') ({{\lvert V \rvert}} - 1)} {(d/2-1 + \alpha') {{\lvert V \rvert}} + 1 - \alpha'} &= \\ 1 + \frac{\beta' - \alpha'}{d/2 + \alpha'} \cdot \frac{{{\lvert V \rvert}} - 1}{{{\lvert V \rvert}} - 1},\end{aligned}$$ which is a constant greater than $1$. Letting $\gamma$ be this constant, and letting $f(y) = \alpha' ({{\lvert V \rvert}}-1) + ({{\lvert E \rvert}} - {{\lvert V \rvert}} + 1)$, it follows that Min Spanning-Tree Hitting Set is APX-hard. The Problem for Matroids {#matroids} ======================== The Min Spanning-Tree Hitting Set can be rephrased as a question about matroids. Let $E$ be a ground set. Let $M_i = (E,{\mathcal{I}}_i)$ be a matroid for $1 \leq i \leq k$. Our objective is to find $F \subseteq E$ such that - $F$ contains a basis for each $M_i$, and - ${{\lvert F \rvert}}$ is minimized. We call this the Minimum Basis Hitting Set problem. Connection to Matroid Intersection ---------------------------------- Suppose we switch to the dual matroids. Note that $F$ contains a basis for $M_i$ if and only $E \setminus F \in {\mathcal{I}}_i^*$. Then our objective to find $F' \subseteq E$ such that - $F' \in {\mathcal{I}}_i^*$ for each $i$, and - ${{\lvert F' \rvert}}$ is maximized. Suppose that such a set is found, and let $F {\,:=\,}E \setminus F'$. The first property implies that $F$ contains a basis for each $M_i$. The second property implies that ${{\lvert F \rvert}}$ is minimized. Stated this way, it is precisely the Matroid $k$-Intersection problem. So, from the point of view of exact algorithms, Min Basis Hitting Set and Matroid k-Intersection problems are equivalent. However, this reduction is not approximation-preserving, and implies nothing about approximation algorithms. Hardness -------- Min Basis-Hitting Set is NP-hard. \[basishittinghard\] We do a reduction from the well-known problem Minimum Hitting Set. An instance of this problem consists of a family of sets ${\mathcal{C}}= \{ C_1, \ldots, C_k \}$. The objective is to find a set $F \subseteq E$ such that $F \cap C_i \neq \emptyset$ for each $i$. This problem is NP-complete. Now we reduce it to Minimum Basis Hitting Set. For each set $C_i$, set $M_i=(E,{\mathcal{I}}_i)$ be the matroid where $ {\mathcal{I}}_i = \left\{\; \{c\} \,:\, c \in C_i \;\right\} \cup \{ \emptyset \} $. That is, $M_i$ is the rank-1 uniform matroid on $C_i$. So a basis hitting set for these matroids corresponds precisely to a hitting set for the the sets ${\mathcal{C}}$. Min Basis Hitting Set is NP-hard to approximate with $c \log n$ for some positive constant $c$. It is well-known that Min Hitting Set is equivalent to Set Cover, and is therefore NP-hard to approximate within $c \log n$ for some positive constant $c$. Since reduction given in Theorem \[basishittinghard\] is approximation preserving, the same hardness applies to Min Basis Hitting Set. An Approximation Algorithm -------------------------- We consider the greedy algorithm for the Min Basis Hitting Set problem. Let ${\mathcal{O}}\subseteq E$ denote an optimum solution. Let ${\operatorname{rank}}_j$ denote the rank function for matroid $M_j$ and let $r_j$ be the rank of $M_j$, i.e., $r_j = {\operatorname{rank}}_j(E)$. Let $F_i$ denote the set that has been chosen after the $i{\ifmmode{^{\textrm{th}}}\else{\textsuperscript{th}\ }\fi}$ step of the algorithm. Initially, we have $F_0 = \emptyset$. For $S \subseteq E$, let $P(S,e) = \sum_{j=1}^k \big({\operatorname{rank}}_j(S+e) - {\operatorname{rank}}_j(S)\big)$; intuitively, this is the total “profit” obtained, or rank that is hit, by adding $e$ to $S$. Let $R_i$ denote $\sum_{j=1}^k \big(r_j - {\operatorname{rank}}_j(F_i)\big)$; intuitively, if the algorithm has chosen a set $F_i$, then $R_i$ is the total amount of “residual rank” that remains to be hit. Consider the $i{\ifmmode{^{\textrm{th}}}\else{\textsuperscript{th}\ }\fi}$ step of the algorithm. Let’s denote the profit obtained by choosing $e_i$ by $p_i = \max_{e \not \in F_{i-1}} P(F_{i-1},e)$. The greedy algorithm chooses an element $e_i \not \in F_{i-1}$ achieving the maximum profit. We now analyze the efficiency of this algorithm. Let ${\mathcal{O}}_i$ be a minimum-cardinality set that contains $F_i$ and is a basis hitting set. For any set $S \supseteq F_i$ and any $e \not \in S$, we have (by submodularity): $$\begin{aligned} {\operatorname{rank}}_j(S + e) + {\operatorname{rank}}_j(F_i) & \leq {\operatorname{rank}}_j(F_i + e) + {\operatorname{rank}}_j(S) \\ {\operatorname{rank}}_j(S + e) - {\operatorname{rank}}_j(S) &\leq {\operatorname{rank}}_j(F_i + e) - {\operatorname{rank}}_j(F_i) \\ P(S,e) &\leq P(F_i,e) \leq p_i\end{aligned}$$ This implies that each edge in ${\mathcal{O}}_i \setminus F_i$ has profit at most $p_i$. Since ${\mathcal{O}}_i$ must ultimately hit all of the residual rank, but each element hits at most $p_i$, we have $R_{i-1} \leq p_i \cdot {{\lvert {\mathcal{O}}_i \setminus F_i \rvert}}$. Now, note that ${{\lvert {\mathcal{O}}_i \setminus F_i \rvert}} \leq {{\lvert {\mathcal{O}}\rvert}}$. This is is because of the non-decreasing property of ${\operatorname{rank}}_j$: if ${\mathcal{O}}$ is a basis hitting set then so is ${\mathcal{O}}{\cup}F_i$. This observation yields the inequality $1 \leq p_i \cdot {{\lvert {\mathcal{O}}\rvert}} / R_{i-1}$. Suppose that the greedy algorithm halts with a solution of cardinality $s$. Then we have $$\begin{aligned} s &\leq \sum_{i=1}^s \frac{ p_i \cdot {{\lvert {\mathcal{O}}\rvert}} }{ R_{i-1} } \leq {{\lvert {\mathcal{O}}\rvert}} \cdot \sum_{i=1}^s \frac{ p_i }{ R_{i-1} } \\ &\leq {{\lvert {\mathcal{O}}\rvert}} \cdot \sum_{i=1}^s \: \sum_{0 \leq j < p_i} \frac{ 1 }{R_{i-1} - j} \leq {{\lvert {\mathcal{O}}\rvert}} \cdot \log R_0.\end{aligned}$$ Here, the last inequality follows from the fact that $R_i = R_{i-1} - p_i$ for $1 \leq i \leq s$. Note that $R_0 = \sum_{j=1}^k r_j$ is the total rank of the given matroids. ------------------------------------------------------------------------ 111111111111111= lexicon $V$, set of $k$ sentences $S_1,\ldots ,S_k$ s.t. $S_i \subset V$ ;\ [**Initialize:**]{} Assign each pair of words a count of sentences it appears in;\ Sort word pairs (edges) in decreasing order;\ through edges;\ Add top edge $e$ into the list of edges;\ Adjust edge weights in sentences which contained edge $e$;\ each sub-graph has a spanning tree (i.e. sentence has a parse);\ [**Output:**]{} a set of spanning trees for all $S_i$; ------------------------------------------------------------------------ \[alg\_txt\] The preceding argument shows that the greedy algorithm has approximation ratio $O(\log n)$, where $n$ is the length of the input. Table 1 presents description of the algorithm. Informally speaking, the algorithm could be explain as follows: [*Estimate potential number of sub-graphs each edge would contribute to if used. Loop through all edges, adding in (greedily) the edge which contributes to the most spanning trees, then re-calculate potential contributions.*]{} Contrast with Matroid Union --------------------------- Consider the matroid union problem for matroids $M_i^*$. The matroid union problem is: $$\max \left\{\; {{\lvert {\bigcup}_i S_i \rvert}} \,:\, S_i \in {\mathcal{I}}_i^* \;\right\}$$ But note that $S_i \in {\mathcal{I}}_i^*$ iff $r_i(V \setminus S_i) = r_i(V)$. In other words, $S_i \in {\mathcal{I}}_i^*$ iff $\bar{S_i}$ contains a basis for $M_i$. And maximizing the size of the union is the same as minimizing the size of the complement of the union. So an equivalent problem is: $$\min \left\{\; {{\lvert {\bigcap}_i \overline{S_i} \rvert}} \,:\, \mbox{$\overline{S_i}$ contains a basis for $M_i$} \;\right\}$$ The minimum does not change if we assume that $\overline{S_i}$ in fact *is* a basis. So, letting $T_i$ denote $\bar{S_i}$, we obtain the equivalent problem: $$\min \left\{\; {{\lvert {\bigcap}_i T_i \rvert}} \,:\, \text{$T_i$ is a basis for $M_i$} \;\right\}$$ This problem is solvable in polynomial time, because it is just matroid union in disguise. It is quite similar to the Minimum Basis Hitting Set problem, except that it has an “intersection” rather than an “union”. Empirical study --------------- We ran preliminary experiments with the approximation algorithm on adult child-directed speech from the CHILDES corpus [@CHILDES]. These experiments demonstrated that the algorithm performs better than the baseline adjacency heuristic because of its ability to pick out non-adjacent dependencies. For example, the sentence “Is that a woof?” is parsed into the following set of links: woof-is, that-is, a-woof. The links correspond to the correct parse tree of the sentence, In contrast, the baseline adjacency heuristic would parse the sentence into is-that; that-a; and a-woof, which fails to capture the dependence between the predicate noun “woof” and the verb, and postulates a non-existent dependency between the determiner “a” and the subject “that”. However, more work is needed to thoroughly assess the performance. In particular, one problem for direct application is the presence of repeated words in the sentence. The current implementation avoids the issue of repeated words in its entirety, by filtering the input text. An alternative approach is to erase the edges among repeated words from the original fully connected graph. This assumes that no word can be a dependent of itself, which might be a problem in some contexts (e.g. “I know that you know”). Related work which was not completed at the time of writing this manuscript seeks to incorporate adjacency as a soft linguistic constraint on the graph by increasing initial weight edges of adjacent words. Discussion ========== We presented some theoretical results for a problem on graphs which is inspired by the unsupervised link grammar induction problem from linguistics. Numerous possible directions for the future work would include searching for more efficient approximation algorithms under various additional constraints on admissible spanning trees, as well as characterizing instances of the problem which could be solved efficiently. Another possible direction is allowing “ungrammatical” corpus as input, e.g. searching efficiently for partial solutions, where several sentences remain unparsed or not fully parsed. Another direction is to look for a solution to a directed graph analog of the problem considered here, which would require finding minimal set of arborescences and relate to the directed dependency parsing. One other question which remains open is an edge weighing scheme which would reflect syntactic consideration and particular language-related constraints, as in the so-called Optimality Theory [@SavovaPhD]. Exploring relation of this problem to other application would be interesting. One such example could be an autonomous network design, where an objective is to efficiently design a network that must connect joint units of organizations which do not necessarily trust each other and want to maintain their own skeletal sub-network in case their partner’s links fail.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The Hodgkin and Huxley (H-H) model is a nonlinear system of four equations that describes how action potentials in neurons are initiated and propagated, and represents a major advance in the understanding of nerve cells. However, some of the parameters are obtained through a tedious combination of experiments and data tuning. In this paper, we propose the use of an iterative method (Landweber iteration) to estimate some of the parameters in the H-H model, given the membrane electric potential. We provide numerical results showing that the method is able to capture the correct parameters using the measured voltage as data, even in the presence of noise.' address: 'Departamento de Modelagem Computacional, Laboratório Nacional de Computação Científica, Av. Getúlio Vargas 333, 25651-070 Petrópolis, RJ, Brazil' author: - 'Jemy A. Mandujano Valle, Alexandre L. Madureira' bibliography: - 'HHinverse.bib' date: 'October 18, 2018' title: Parameter Identification Problem in the Hodgkin and Huxley Model --- [^1] Introduction. ============= In 1952 Hodgkin and Huxley [@H-H1952] used voltage-clamp technique to extract the parameters of the ionic channel model of the squid giant axon. In the space-clamped version of the H-H model, the membrane electrical potential $V:[0,T]\to{\mathbb R}$ solves $$\label{equation1} C_M\dot {V}(t)=I_{\ext}+I_\ion(t)\quad\text{in }(0,T],$$ where $C_M$ is the specific membrane capacitance, $V$ is the membrane potential, $\dot{V}$ is the rate of voltage change (dots denote time derivatives), $I_{\ext}$ is the specific external current applied on the membrane. The specific ionic current $I_\ion(t)$ is the sum of three currents $(I_\ion(t)=I_{\Na}(t)+I_\K(t)+I_{L}(t))$, potassium, sodium and leak currents, satisfying: $$\begin{aligned} I_{\Na}(t)&=&G_{\Na}\;m^a(V,t)\;h(V,t)^b\;(V(t)-E_{\Na});\label{equation2}\\ I_\K(t)&=&G_\K\;n^c(V,t)\;(V(t)-E_\K);\label{equation3}\\ I_{L}(t)&=&G_{L}\;(V(t)-E_{L}).\label{equation4}\end{aligned}$$ The constants $G_{\Na}$, $G_\K$ and $G_L$ are the maximal specific conductance for Na$^+$, K$^+$ and leakage channels, and $E_{\Na}$, $E_\K$, $E_{L}$ are the Nernst equilibrium potentials. The functions $m$ and $h$ are the activation and inactivation variables for $\Na^+$, and $n$ is the activation function for $K^+$. These functions are unitless gating variables that take values between $0$ and $1$. Also, the exponents $a$, $b$ and $c$ are positive numbers. The units of the other parameters are in Table \[ta1\]. **Parameters** **Units** **Units name** -------------------------- -------------- ------------------------------------ -- -- -- -- -- -- -- $C_M$ $\mu F/cm^2$ microfarad per square centimeter $V$ $m V$ millivolt $\dot V$ $V/s$ volts per second $I_{\ext}$, $I_{ion}$ $\mu A/cm^2$ microampere per square centimeter $G_{\Na}$, $G_\K$, $G_L$ $ mS/cm^2$ millisiemens per square centimeter $E_{\Na}$, $E_\K$, $E_L$ $mV$ millivolt : Units of the parameters; see [@H-H1952], Table 3. []{data-label="ta1"} The experiments performed by Hodgkin and Huxley [@H-H1952] suggest that $m$, $h$ and $n$ are functions that depend on time and the membrane potential. The exponent $c$ models the number of gating particles on the channel. In the case of active $\Na$ currents, experiments suggest that two types of independent gating particles are involved, $a$ activation gates $m$, and $b$ inactivation gates $h$ [@ermentrout2003]. In addiction, $m$ $n$ and $h$ satisfy the differential equations: $$\label{equation5} \dot{\mathcal{X}}(V,t)=\alpha_\mathcal{X}(V) (1-\mathcal{X}(V,t))-\beta_\mathcal{X}(V) \mathcal{X}(V,t) \quad\text{where }\mathcal{X}=m,n,h.$$ The functions $\alpha_\mathcal{X}$ and $\beta_\mathcal{X}$ depend on the membrane potential and are given by $$\label{e:alphasdef} \begin{array}{lll} \alpha_m=\frac{(25-V)/10}{\exp( (25-V)/10)-1},\;\;\; & \alpha_h=0.07\exp(-V/20),\;\;\; & \alpha_\mathsf {n}=\frac{(10-V)/100}{\exp((10-V)/10)-1 },\vspace*{0.2cm} \\ \beta_m=4\exp(-V/18), & \beta_h=\frac{1}{\exp((30-V)/10)+1},&\beta_\mathsf {n}=0.125\exp(-V/80). \end{array}$$ To equation we add the initial conditions $$\label{equation6} V(0)=V_0,\;\;\;m(0)=m_0,\;\;\;n(0)=n_0,\;\;\;h(0)=h_0.$$ Thus, (\[equation1\]-\[equation6\]) yield the following system of ordinary differential equation (ODE): $$\label{equation7} \left \{\begin{array}{l}\displaystyle C_M\dot {V}=I_{\ext}- G_{\Na}m^ah^b(V-E_{\Na})-G_\K n^c(V-E_\K)-G_{L}(V-E_L)\quad\text{for }t \in (0,T] \vspace*{0.15cm} \\ \displaystyle \dot{\mathcal{X}} =(1-\mathcal{X})\alpha_\mathcal{X} (V)-\mathcal{X}\beta_\mathcal{X}(V)\quad\text{where }\mathcal{X}=m,n,h\text{ and } t \in (0,T] \\ V(0)=V_0,\;\;\;\; m(0)=m_0,\;\;\;\; n(0)=n_0,\;\;\;\; h(0)=h_0, \end{array} \right.$$ and $C_M$, $I_{\ext}$, $E_{\Na}$, $E_\K$, $E_L$, $m_0$, $n_0$ and $h_0$ are known. Given all the parameters, it is possible to find a (theoretical or numerical) solution for . That is the *direct problem*. In *inverse problems*, one is given the voltage $V$ and has to compute one or more parameters. In this work, we consider two different *inverse* problems. The first one is to obtain the maximum conductances $G_{\Na}$, $G_\K$ and $G_{L}$ given the measurement of the membrane potential. For the second problem, the goal is to obtain the exponents $a$, $b$ and $c$, again given the measurement of the membrane potential. Using experimental data from the squid neuron, Hodgkin and Huxley obtained the parameters $a=3$, $b=4$ and $c=1$. Note, however, that other neurons may produce different parameters. Besides the Hodgkin and Huxley model, there are simplified models such as the cable equation, FitzHugh-Nagumo and Morris-Lecar models. Wilfrid Rall [@rall1977; @rall1992-1] developed the use of cable theory in computational neuroscience, as well as passive and active compartmental modeling of the neuron. In a previous paper [@mandujano2018], the authors determine conductances with nonuniform distribution in the equation of the cable with and without branches, using the Landweber iterative method. See also [@tadi2002; @bell2005; @avdonin2013; @avdonin2015], for identification of parameters in the cable equation, and [@cox2001-2; @cox2004; @pavel2012; @che2012; @pavel2013; @tuikina2017] for investigations on inverse problems in FitzHugh-Nagumo and Morris-Lecar models. In [@destexhe2007; @destexhe2004; @vich2017] the authors obtained approximately time-dependent but voltage-independent conductances, given the membrane potential, in a system of three ordinary differential equations (passive membrane equation). For the Hodgkin and Huxley model, the parameters of ionic channels are estimated in [@buhry2011; @buhry2012] using evolutionary algorithms. Inverse problems are said to be *ill-posed*. A problem is ill-posed in the sense of Hadamard [@hadamard2014] if any of the following conditions are not satisfied: there is a solution; the solution is unique; the solution has a continuous dependence on the input data (stability). Here we admit the existence of a single solution to the problem. However, stability is not guaranteed. Stability is necessary if we want to ensure that small variations in the data lead to small changes in the solution. Problems of instability can be controlled by regularization methods, in particular the Landweber iterative scheme [@binder1996; @chapko2004; @hanke1995; @neubauer2000]. This article is outlined as follows. Section \[section2\] presents our inverse problems for the H-H model along with some theoretical results, and in Section \[section3\] we show numerical results to describe the effectiveness of our strategy. Finally, we include in the Appendices some more technical arguments. Inverse Problem in the H-H model {#section2} ================================ In what follows, we describe an abstract formulation of the Landweber method or Landweber iteration [@kaltenbacher2008]. Consider  and let $x=(G_{\Na},G_\K,G_L)\in \mathbb{R}^3$ or $x=(a,b,c)\in \mathbb{R}^3$. Consider also the set of function $L^2(0,T)$, and the nonlinear operator $$\label{equa10} F: \mathbb{R}^3\rightarrow L^2(0,T),$$ defined by $F(x)=V$, where $V$ solves . In practical terms, the data $V$ are obtained by measurements. Therefore, we denote the measurements by $V^\delta$, of the which we assume to know the noise level $\delta$, satisfying $$\label{equa11} \|V-V^\delta\|_{L^2(0,T)}^2=\int_0^T|V(t)-V^\delta(t)|^2\,dt\le\delta.$$ To obtain an approximation of $x$, given $V^\delta$, we used the Landweber iteration $$\label{equation8} x^{ {k+1},\delta}=x^{ k,\delta}+w^{k,\delta}F'(x^{ k,\delta})^*(V^\delta-F(x^{ k,\delta})),$$ where $F'(x^{k,\delta})$ is the Gateaux-derivative of $F$ computed at $x^{k,\delta}$, and $F'(x^{k,\delta})^*$ is its adjoint. We also define $$w^{k,\delta}=\frac{{\|V^\delta-F(x^{k,\delta})\|}^2_{L^2(0,T)}}{{\| F'(x^{ k,\delta})^*(V^\delta-F(x^{ k,\delta}))\|}^2_{\mathbb{R}^3} }.$$ The iteration   begins with a guess $x^{1,\delta}$ and stops at the minimum $k_*=k(\delta,V^\delta)$, such that, for a given $\tau>2$ (see [@kaltenbacher2008], equation (2.14) ), $$\label{equation9} \|V^\delta-F(x^{ {k_*},\delta})\|_{L^2(0,T)}\le\tau\delta.$$ It is possible to show that, under certain conditions (we assume that is the case), $x^{ {k_*},\delta}$ converges to a solution of $F(x)=V$ as $\delta\to0$; see [@kaltenbacher2008] [Theorem 3.22]{}. Inverse Problem to obtain conductances in the H-H model {#subsection2.1} ------------------------------------------------------- The present goal is to estimate the maximum conductances $G_{\Na}$, $G_\K$ and $G_L$ while assuming that  holds. We assume that the exponents are $a=3$, $b=1$, and $c=4$. We denote our unknown parameters such as $x =\bG = (G_{\Na},G_\K, G_L)$, then from iteration we have $$\label{equation012} \bG^{ {k+1},\delta}=\bG^{ k,\delta}+w^{k,\delta}F'(\bG^{ k,\delta})^*(V^\delta-F(\bG^{ k,\delta})).$$ Given an initial approximation $\bG^{1,\delta}$ and $V^\delta$, we obtain a regularizing approximation $\bG^{k_{*},\delta}$ for $\bG$, from Landweber iteration . We denote $\bG^{k,\delta}=(G_{\Na}^{k,\delta},G_\K^{k,\delta},G_L^{k,\delta})$. In the next theorem, we compute the adjoint of the Gateaux derivative $F'(\bG^{k,\delta})^*$ to optimize from . \[Theorem1\] It follows from  that $$\label{equa26} \left( G_{\Na}^{k+1,\delta},G_\K ^{k+1,\delta},G_L ^{k+1,\delta}\right)= \left( G_{\Na}^{k,\delta},G_\K ^{k,\delta},G_L ^{k,\delta}\right)+w^{k,\delta}\left(X_{\Na}^{k,\delta},X_K^{k,\delta},X_L^{k,\delta}\right),$$ where $$w^{k,\delta}=\frac{{\|V^\delta-V^{k,\delta}\|}^2_{L^2(0,T)}}{ {\left\|\left(X_{\Na}^{k,\delta},X_K^{k,\delta},X_L^{k,\delta}\right) \right\|}^2_{\mathbb{R}^3 } },$$ and $$\begin{aligned} X_{\Na}^{k,\delta}&=&\int_0^T {\left(m^{k,\delta}\right)}^a {\left(h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})U^{k,\delta}\;dt,\label{equa14} \\ X_K^{k,\delta}&=&\int_0^T {\left(n^{k,\delta}\right) }^c(V^{k,\delta}-E_\K) U^{k,\delta}\;dt,\label{equa15} \\ X_{L}^{k,\delta}&=&\int_0^T {\left(n^{k,\delta}\right) }^c(V^{k,\delta}-E_\K) U^{k,\delta}\;dt.\label{equa16}\end{aligned}$$ The functions $m^{k,\delta}$, $n^{k,\delta}$, $h^{k,\delta}$ and $V^{k,\delta}$ solve, given $G_{\Na}^{k,\delta}$, $G_\K^{k,\delta}$ and $G_{L}^{k,\delta}$, $$\label{equati1} \left \{\begin{array}{l}\displaystyle C_M\dot{ V}^{k,\delta}=I_{\ext}- G_{\Na}^{k,\delta} {\left(m^{k,\delta}\right)}^a {\left(h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na}) -G_\K^{k,\delta}{\left(n^{k,\delta}\right) }^c(V^{k,\delta}-E_\K)\vspace*{0.1cm} \\ \hspace*{1.5cm} -G_{L}^{k,\delta}(V^{k,\delta}-E_L),\vspace*{0.2cm} \\ \dot{\mathcal{X}} =(1-\mathcal{X})\alpha_\mathcal{X} (V^{k,\delta})-\mathcal{X}\beta_\mathcal{X}(V^{k,\delta}) \quad\text{for }\mathcal{X}=m^{k,\delta},n^{k,\delta},h^{k,\delta}, \vspace*{0.2cm} \\ V^{k,\delta}(0)=V_0,\;\;\;m^{k,\delta}(0)=m_0,\;\;\;n^{k,\delta}(0)=n_0,\;\;\;h^{k,\delta}(0)=h_0, \end{array} \right.$$ and $\alpha_\mathcal{X}$, $\beta_\mathcal{X}$ are defined by . Finally, $U^{k,\delta}$ solve, given $m^{k,\delta}$, $n^{k,\delta}$, $h^{k,\delta}$ and $V^{k,\delta}$, $$\label{ab6.44} \left \{\begin{array}{l}\displaystyle C_M\dot {U}^{k,\delta}- \left( G_{\Na}^{k,\delta} {\left(m^{k,\delta}\right)}^a {\left(h^{k,\delta}\right)}^b+G_\K^{k,\delta} {\left(n^{k,\delta}\right)}^c+ G_L^{k,\delta}\right) U^{k,\delta}\vspace*{0.2cm} \\ -[(1-m^{k,\delta})\alpha'_{m^{k,\delta}}(V^{k,\delta})- m^{k,\delta}\beta'_ {m^{k,\delta}}(V^{k,\delta})]P^{k,\delta}\vspace*{0.2cm} \\ -[(1- n^{k,\delta})\alpha'_ {n^{k,\delta}}(V^{k,\delta})- n^{k,\delta}\beta_ {n^{k,\delta}}'(V^{k,\delta})]Q^{k,\delta}\vspace*{0.2cm} \\ -[(1-h^{k,\delta})\alpha'_{h^{k,\delta}}(V^{k,\delta})- h^{k,\delta}\beta'_{h^{k,\delta}}(V^{k,\delta})]R^{k,\delta}={V^\delta}-V^{k,\delta},\vspace*{0.3cm} \\ \displaystyle \dot P^{k,\delta}-[\alpha_{m^{k,\delta}}(V^{k,\delta})+\beta_{m^{k,\delta}}(V^{k,\delta})]P^{k,\delta} =-aG_{\Na}^{k,\delta}{\left(m^{k,\delta}\right)}^{a-1}{\left(h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})U^{k,\delta},\vspace*{0.3cm}\\ \dot Q^{k,\delta}-[\alpha_{n^{k,\delta}}(V^{k,\delta})+\beta_{n^{k,\delta}}(V^{k,\delta})]Q^{k,\delta} =-c G_\K^{k,\delta} {\left(n^{k,\delta}\right)}^{c-1}(V^{k,\delta}-E_\K)U^{k,\delta},\vspace*{0.3cm} \\ \dot R^{k,\delta}-[\alpha_{h^{k,\delta}}(V^{k,\delta})+\beta_{h^{k,\delta}}(V^{k,\delta})]R^{k,\delta} =-b G_{\Na}^{k,\delta} {\left(m^{k,\delta}\right)}^a {\left(h^{k,\delta}\right)}^{b-1}(V^{k,\delta}-E_{\Na})U^{k,\delta},\vspace*{0.3cm} \\ U^{k,\delta}(T)=0,\hspace*{0.5cm} P^{k,\delta}(T)=0,\hspace*{0.5cm} Q^{k,\delta}(T)=0,\hspace*{0.5cm} R^{k,\delta}(T)=0. \end{array} \right.$$ As previously mentioned, we assume that the constants $a$, $b$, $c$, $E_{\Na}$ ,$E_\K$, $E_{L}$, $C_M$, $I_{\ext}$, $m_0$, $n_0$ and $h_0$ are known data. See Appendix \[AppendixA\]. We next describe the computational scheme. \[Algorithm1\] Choose $\bG^{1,\delta}$ as an initial approximation for $\bG$ Compute $m^{1,\delta}$, $n^{1,\delta}$, $h^{1,\delta}$ and $V^{1,\delta}$ from  , replacing $\bG^{k,\delta}$ by $\bG^{1,\delta}$ k=1 Inverse Problem to obtain exponents in the H-H model {#subsection2.2} ---------------------------------------------------- Assume again that  holds and that $G_{\Na}$, $G_\K$ and $G_L$ are known. The goal of this subsection is to estimate the exponents $a$, $b$ and $c$. Denoting the unknown parameters by $x=\ba=(a,b,c)$ it follows from iteration  that $$\label{equation017} \ba^{ {k+1},\delta}=\ba^{ k,\delta}+w^{k,\delta}F'(\ba^{ k,\delta})^*(V^\delta-F(\ba^{ k,\delta})).$$ Given an initial approximation $\ba^{1,\delta}$ and the data $V^\delta$, we obtain a regularizing approximation $\ba^{k_{*},\delta}$ for $\ba$, from the Landweber iteration . Denote $\ba^{k,\delta}=(a^{k,\delta},b^{k,\delta},c^{k,\delta})$. In the next Theorem, we compute the adjoint of the Gateaux derivative $F'(\ba^{k,\delta})^*$ from . \[Theorem2\] Consider the iteration . It follows then that $$\label{equati.018} \left(a^{k+1,\delta},b^{k+1,\delta},c^{k+1,\delta}\right)=\left(a^{k,\delta},b^{k,\delta},c^{k,\delta}\right)+w^{k,\delta}\left(X_a^{k,\delta},X_b^{k,\delta},X_c^{k,\delta}\right),$$ where $w^{k,\delta}$ satisfies $$w^{k,\delta}=\frac{{\|V^\delta-V^{k,\delta}\|}^2_{ L^2(0,T)}}{ {\left\|\left(X_a^{k,\delta},X_b^{k,\delta},X_c^{k,\delta}\right) \right\|}^2_{\mathbb{R}^3 } },$$ and $$\begin{aligned} X_a^{k,\delta}&=&\int_0^T G_{\Na}(V^{k,\delta}-E_{\Na}) {\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}}{U^{k,\delta}}\ln(m^{k,\delta})\,dt,\label{equ16} \\ X_b^{k,\delta}&=&\int_0^T G_{\Na}(V^{k,\delta}-E_{\Na}){\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}}{U^{k,\delta}}\ln(h^{k,\delta})\,dt\label{equ17}, \\ X_c^{k,\delta}&=&\int_0^T G_\K (V^{k,\delta}-E_\K) {\left(n^{k,\delta}\right)}^{c^{k,\delta}}{U^{k,\delta}}\ln( {n^{k,\delta}})\,dt. $$ The functions $m^{k,\delta}$, $n^{k,\delta}$, $h^{k,\delta}$ and $V^{k,\delta}$ solve $$\label{equati344} \left \{\begin{array}{l}\displaystyle C_M\dot{ V}^{k,\delta}=I_{\ext}- G_{\Na} {\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}}(V^{k,\delta}-E_{\Na}) -G_\K{\left(n^{k,\delta}\right) }^{c^{k,\delta}}(V^{k,\delta}-E_\K)\vspace*{0.1cm}\\ \hspace*{1.5cm} -G_{L}(V^{k,\delta}-E_L),\vspace*{0.2cm}\\ \displaystyle \dot{\mathcal{X}} =(1-\mathcal{X})\alpha_\mathcal{X} (V^{k,\delta})-\mathcal{X}\beta_\mathcal{X}(V^{k,\delta}); \hspace*{0.4cm}\mathcal{X}=m^{k,\delta},n^{k,\delta},h^{k,\delta}, \vspace*{0.2cm}\\ V^{k,\delta}(0)=V_0;\;\;\;m^{k,\delta}(0)=m_0;\;\;\;n^{k,\delta}(0)=n_0;\;\;\;h^{k,\delta}(0)=h_0, \end{array} \right.$$ where $a^{k,\delta}$, $b^{k,\delta}$ and $c^{k,\delta}$ are given. Also, $U^{k,\delta}$ solve $$\label{equati377} \left \{\begin{array}{l}\displaystyle C_M\dot U^{k,\delta}-\left(G_{\Na} {\left(m^{k,\delta}\right)}^{a^{k,\delta}}{\left(h^{k,\delta}\right)}^{b^{k,\delta}}+G_\K {\left(n^{k,\delta}\right)}^{c^{k,\delta}}+G_L\right) {U^{k,\delta}}\\ \hspace*{1.cm} -[(1- {m^{k,\delta}})\alpha'_{m^{k,\delta}}(V^{k,\delta})- m^{k,\delta}\beta'_{m^{k,\delta}}(V^{k,\delta})]{P^{k,\delta}}\\ \hspace*{1.cm} -[(1- n^{k,\delta})\alpha'_{n^{k,\delta}}(V^{k,\delta})- n^{k,\delta}\beta_ {n^{k,\delta}}'(V^{k,\delta})]Q^{k,\delta}\\ \hspace*{1.cm} -[(1-h^{k,\delta})\alpha'_{h^{k,\delta}}(V^{k,\delta})- h^{k,\delta}\beta'_{h^{k,\delta}}(V^{k,\delta})]R^{k,\delta}=V^\delta-V^{k,\delta},\\ \displaystyle {\dot P}^{k,\delta}-[\alpha_{m^{k,\delta}}(V^{k,\delta})+\beta_{m^{k,\delta}}(V^{k,\delta})]P^{k,\delta} =\\\hspace*{1.cm}-a^{k,\delta}G_{\Na} {\left(m^{k,\delta}\right) }^{ a^{k,\delta}-1} {\left(h^{k,\delta}\right)}^{b^{k,\delta}}(V^{k,\delta}-E_{\Na})U^{k,\delta},\vspace*{0.2cm}\\ \displaystyle {\dot Q}^{k,\delta}-[\alpha_{n^{k,\delta}}(V^{k,\delta})+\beta_{n^{k,\delta}}(V^{k,\delta})]Q^{k,\delta} =\\\hspace*{1.cm}-c^{k,\delta}G_\K {\left(n^{k,\delta}\right)}^{c^{k,\delta}-1}(V^{k,\delta}-E_\K)U^{k,\delta},\vspace*{0.2cm}\\ \displaystyle {\dot R}^{k,\delta}-[\alpha_{h^{k,\delta}}(V^{k,\delta})+\beta_{h^{k,\delta}}(V^{k,\delta})]R^{k,\delta} =\\\hspace*{1.cm}-b^{k,\delta}G_{\Na} {\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}-1}(V^{k,\delta}-E_{\Na})U^{k,\delta},\vspace*{0.2cm}\\ U^{k,\delta}(T)=0;\hspace*{0.5cm} P^{k,\delta}(T)=0;\hspace*{0.5cm} R^{k,\delta}(T)=0;\hspace*{0.5cm} Q^{k,\delta}(T)=0, \end{array} \right.$$ given $m^{k,\delta}$, $n^{k,\delta}$, $h^{k,\delta}$ and $V^{k,\delta}$. The constants $ G_{\Na}$, $G_\K$, $E_{\Na}$ ,$E_\K$, $E_{L}$, $C_M$, $I_{\ext}$, $m_0$, $n_0$ and $h_0$ are given data. See Appendix . We next describe the computational scheme. \[Algorithm2\] Choose $\ba^{1,\delta}$ as an initial approximation for $\ba$ Compute $m^{1,\delta}$, $n^{1,\delta}$, $h^{1,\delta}$ and $V^{1,\delta}$ from  , replacing $\ba^{k,\delta}$ by $\ba^{1,\delta}$ k=1 Numerical simulation {#section3} ==================== To design our numerical experiments, we first choose $x$ ($x=\bG$ or $x=\ba$) and compute $V$ from . Of course, in practice, the values of $V$ are given by some experimental measurements, and thus subject to experimental/measurement errors. In our examples, for a given $\delta$, the noisy $V^\delta$ is obtained from $$\label{equ22} V^\delta(t)=V(t)+V(t) {\rand}_\varepsilon (t), \;\;\;\text{for all } t \in [0,T]$$ where $\rand_\varepsilon$ is a uniformly distributed random variable taking values in the range $[-\varepsilon,\varepsilon]$, and $\varepsilon=\delta/\|V\|_{L^2(0,T)}$. Next, given the initial guess $x^{1,\delta}$ and the data $V^\delta$ and $\delta$, we start to recover $x$ using Algorithm \[Algorithm1\] (for $x=\bG$) or Algorithm \[Algorithm2\] (for $x=\ba$). Note that we have the exact $x$, and we use that to gauge the algorithm performance. The absolute error of $V^\delta$ and its approximation $V^{k,\delta}$ defines the residual from $$\label{equti28} \Res_{k}= \| V^{\delta}-V^{k,\delta}\|_{L^2(0,T)}=\sqrt{\int_0^L \left( V^{\delta}(t)-V^{k,\delta}(t)\right)^2dt},\;\;\;k=1,2,\cdots,k_*.$$ The percent error of vector $x\in \mathbb{R}^3$ is defined by $$\label{equ23} \Error_{k}^x={ \frac{\|x-x^{k,\delta}\|_{\mathbb{R}^3}}{\|x\|_{ \mathbb{R}^3}}}\times 100\%,\;\;\;k=1,2,\cdots,k_*.$$ Each step of Algorithm \[Algorithm1\] and Algorithm \[Algorithm2\] involves solving two ODEs. Of course, there is no analytical solution for those equations, and the use of numerical methods is necessary. We use explicit Euler with a fixed time step $\Delta t$. In this section we will present two numerical simulations. In Example \[Exa2.1\] we estimate the conductances $G_{\Na}$, $G_\K$ and $G_L$, and in Example \[Exa3.1\] we estimate the exponents $a$, $b$ and $c$. Our simulation were computed with Matlab R2012b on a Dell PC, running on a Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz with 32 GB of RAM. See the code in the URL:<https://github.com/MandujanoValle/Conductances-HH>, to estimate the conductances $G_{\Na}$, $G_\K$ and $G_L$, and URL:<https://github.com/MandujanoValle/Exponents-HH>, to estimate the exponents $a$, $b$ and $c$. \[Exa2.1\] This example is a particular case from , with values (see [@cooley1966], page 586): $C_M=1\;[\mu F/cm^2]$, $E_{\Na}=115\;[mV]$, $E_\K=-12\;[mV]$, $E_L=10.598\;[mV]$, $G_{\Na}=120\;[mS/cm^2]$, $G_\K=36\;[mS/cm^2]$, $G_L=0.3\;[mS/cm^2]$, $I_{\ext}=0\;[\mu A/cm^2]$, $a=3$, $b=1$ and $c=4$. Let the initial conditions $V(0)=-25\;[mV]$, $m(0)=0.5$, $n(0)=0.4$ and $h(0)=0.4$. We consider $T=10 \;[mS]$ and $\Delta t=0.02$. Given $V^\delta$, the goal of this example is to approximate $\bG=(G_{\Na}, G_\K, G_L)\;[mS/cm^2]$. First, given $\bG = (120,36,0.3)\;[mS/cm^2]$, we compute $V$ from . Then, we calculate $V^\delta$ from given $\varepsilon$ (see table \[table1\]). Next, we consider $V$ and $\bG$ as unknowns. In this test we consider the initial guess $\bG^{1,\delta}=(0,0,0)\;[mS/cm^2]$ and $\tau=2.01$. Table \[table1\] presents the results for various levels of noise. When $\varepsilon$ decreases, the number of iterations grow resulting in a better approximation for $\bG=(G_{\Na},G_\K,G_L)\;[mS/cm^2]$ and smaller residuals. As expected, the result of the last column is close to $\tau\delta$, related to the stopping criteria . In Figures \[Figure1-1\], \[Figure1-2\] and \[Figure1-3\], we plot some results for $\varepsilon=5\%$ (Table 2, line 4). $\varepsilon $ $k_* $ $G_{\Na}^{k_*,\delta}$ $G_\K^{k_*,\delta}$ $G_L^{k_*,\delta}$ $Error_{k_*}^x$ $Res_{k_*}$ ---------------- --------- ------------------------ --------------------- -------------------- ----------------- ------------- -- -- -- $125\%$ $1$ $0$ $0 $ $0 $ $100 \;\%$ $161 $ $25\%$ $19303$ $114.08$ $28.49 $ $8.1727$ $9.9 \;\%$ $49$ $5\%$ $25012$ $115.07$ $30.59 $ $0.7938$ $5.8 \;\%$ $10 $ $1\%$ $33419$ $119.10$ $34.16 $ $0.3221$ $1.6 \;\%$ $2$ $0.2\%$ $48642$ $119.82$ $35.62 $ $0.3043$ $0.3 \;\%$ $0.4$ : Numerical results for Example \[Exa2.1\] for various values of $\varepsilon$, as in . The second column contains the number of iterations according to . The third, fourth and fifth columns are the approximations for $G_{\Na}$, $G_\K$ and $G_L$ respectively. The sixth column is the relative error of $\bG=(G_{\Na},G_\K,G_L)$ according to . The last column is the residue, see . []{data-label="table1"} ![For Example \[Exa2.1\]. The red line ($V$) is the exact membrane potential and blue line ($V^\delta$) is the membrane potential measurement; in this case $\varepsilon=5\%$. []{data-label="Figure1-1"}](Example1-1.pdf){height="7cm" width="12cm"} ![Figures for Example \[Exa2.1\] (estimation of the conductances) with $\varepsilon = 5\%$. The x-axis gives the number of iterations ($k$) and the y-axis gives the conductance. The red lines are the exact solutions and blue lines are the approximations. The figures \[Figure1-2\]-A, \[Figure1-2\]-B and \[Figure1-2\]-C display the estimates of the maximum conductances of sodium, potassium and leakage, respectively.[]{data-label="Figure1-2"}](Example1-2.pdf){height="8cm" width="13cm"} ![Example \[Exa2.1\] with $\varepsilon =5\%$. The x-axis indicates the number of iterations ($k$). The y-axis, in the figures $A$ and $B$ are the residual and error , respectively. . []{data-label="Figure1-3"}](Example1-3.pdf){height="8cm" width="13cm"} \[Exa3.1\] This example is another particular case from with values (see [@cooley1966], page 586): $C_M=1\;[\mu F/cm^2]$, $E_{\Na}=115\;[mV]$, $E_\K=-12\;[mV]$, $E_L=10.598\;[mV]$, $G_{\Na}=120\;[mS/cm^2]$, $G_\K=36\;[mS/cm^2]$, $G_L=0.3\;[mS/cm^2]$ , $I_{\ext}=0\;[\mu A/cm^2]$, $a=3$, $b=1$ and $c=4$. Let the initial conditions $V(0)=-25\;[mV]$, $m(0)=0.5$, $n(0)=0.4$ and $h(0)=0.4$. We consider the time $T=5\;[ms]$ with $\Delta t=0.02$. Given $V^\delta$, our goal is to approximate $\ba=(a, b, c)=(3,1,4)$. First we calculate $V$ from given $\ba=(3,1,4)$. Then, we calculate $V^\delta$ from given $\varepsilon$ (see table \[table1\]). We then consider $V$ and $\ba$ unknown. In this example we consider the initial guess $\ba^{1,\delta}=(0,0,0)$ and $\tau=2.01$. Table \[table3\] presents the results for various levels of noise. In figures \[Figure2-1\], \[Figure2-2\] and \[Figure2-3\], we plot some results for a level of noise $\varepsilon=1\%$. $\varepsilon$ $k_*$ $a^{ k_*,\delta}$ $b^{ k_*,\delta}$ $c^{ k_*,\delta}$ $Error_{k_*}^x$ $Res_{k_*}$ --------------- ---------- ------------------- ------------------- ------------------- ----------------- ------------- -- -- -- $125\;\%$ $1$ $0$ $0$ $0$ $ 100\;\%$ $170$ $25\;\% $ $11681$ $1.572$ $0.496$ $-0.300$ $ 89 \;\%$ $48$ $5\;\%$ $95605$ $2.970$ $0.807$ $2.626$ $ 27\;\%$ $9.7$ $1\;\%$ $188827$ $3.008$ $0.954$ $3.674$ $ 6\;\%$ $1.9$ $0.2\;\%$ $283487$ $3.002$ $0.990$ $3.930$ $ 1.4\;\%$ $0.4$ : Numerical results for Example \[Exa3.1\]. See Table \[table1\] for a description of the contents. []{data-label="table3"} ![For Example \[Exa3.1\] and $\varepsilon=1\%$. The red line ($V$) is the exact membrane potential and blue line ($V^\delta$) is the membrane potential measurement. []{data-label="Figure2-1"}](Example3-1.pdf){height="7cm" width="12cm"} ![For Example \[Exa3.1\] and $\varepsilon=1\%$. The x-axis is the number of iterations ($k$). In y-axis, the red lines are the exact solutions and blue lines are the approximations. The figures \[Figure2-2\]-A, \[Figure2-2\]-B and \[Figure2-2\]-C are the estimates of $a$, $b$ and $c$, respectively..[]{data-label="Figure2-2"}](Example3-2.pdf){height="8cm" width="13cm"} ![For Example \[Exa3.1\] and $\varepsilon=1\%$. The x-axis is the number of iterations ($k$). The y-axis, in the figures $A$ and $B$ are the residual and error , respectively.[]{data-label="Figure2-3"}](Example3-3.pdf){height="8cm" width="13cm"} Proof of Theorem \[Theorem1\] {#AppendixA} ============================= In this Appendix, we show Theorem \[Theorem1\]. Consider the operator F defined in . Evaluating $\bG^{k,\delta}$ in $F$ , we have $F(\bG^{k,\delta})=V^{k,\delta}$, where $V^{k,\delta}$, $m^{k,\delta}$, $n^{k,\delta}$ and $h^{k,\delta}$ solve the ODE . Let the vector $\boldsymbol{\theta}=(\theta_{\Na},\theta_\K,\theta_L)\in \mathbb{R}^3$ and $\lambda \in \mathbb{R}$, then evaluating $\bG^{k,\delta}+\lambda\boldsymbol{\theta}$ in the operator $F$, we have $F(\bG^{k,\delta}+\lambda\boldsymbol{\theta})=V^{k,\delta}_\lambda$, where $V^{k,\delta}_\lambda$, $m^{k,\delta}_\lambda$, $n^{k,\delta}_\lambda$ and $h^{k,\delta}_\lambda$ solve $$\label{ab6.2} \left \{\begin{array}{l}\displaystyle C_M\dot{V}^{k,\delta}_\lambda=I_{\ext} - \left(G_{\Na}^{k,\delta}+\lambda\theta_{\Na}\right){\left({m}_\lambda^{k,\delta}\right) }^a{\left({h}_\lambda^{k,\delta}\right)}^b\left(V^{k,\delta}_{\lambda}-E_{\Na}\right)\\\hspace*{0.5cm}-\left(G_\K^{k,\delta}+\lambda\theta_K\right){ \left(n_\lambda^{k,\delta}\right) }^c\left(V^{k,\delta}_{\lambda}-E_\K\right) -\left(G_{L}^{k,\delta}+\lambda\theta_L\right)\left(V^{k,\delta}_\lambda-E_L\right),\vspace*{0.2cm}\\ \displaystyle \dot{\mathcal{X}} =(1-\mathcal{X})\alpha_\mathcal{X} (V^{k,\delta})-\mathcal{X}\beta_\mathcal{X}(V^{k,\delta}); \hspace*{1.4cm}\mathcal{X}=m^{k,\delta}_\lambda,n^{k,\delta}_\lambda,h^{k,\delta}_\lambda, \vspace*{0.2cm}\\ V_\lambda^{k,\delta}(0)=V_0;\;\;\;\;m_\lambda^{k,\delta}(0)=m_0;\;\;\;\;n_\lambda^{k,\delta}(0)=n_0;\;\;\;\;n_\lambda^{k,\delta}(0)=n_0. \end{array} \right.$$ The Gateaux derivative of $F$ at $\bG^{k,\delta}$ in the direction $\boldsymbol{\theta}$ is given by $$\label{equ8} W^{k,\delta}=F'(\bG^{k,\delta})(\boldsymbol{\theta})=\lim_{\lambda\to0}\frac{F(\bG^{k,\delta}+\lambda\boldsymbol{\theta})-F(\bG^{k,\delta})}{\lambda}.$$ Also, we denote the following limits $$\label{ekl13} M^{k,\delta}=\lim_{\lambda \rightarrow 0}\frac{m^{k,\delta}_{\lambda}-m^{k,\delta}}{\lambda}, \hspace*{0.5cm}N^{k,\delta}=\lim_{\lambda \rightarrow 0}\frac{n^{k,\delta}_{\lambda}-n^{k,\delta}}{\lambda}, \hspace*{0.5cm}H^{k,\delta}=\lim_{\lambda \rightarrow 0}\frac{h^{k,\delta}_{\lambda}-h^{k,\delta}}{\lambda},$$ where $M^{k,\delta}$, $N^{k,\delta}$ and $H^{k,\delta}$ are the Gateaux derivatives of $m^{k,\delta}$, $n^{k,\delta}$ and $h^{k,\delta}$, respectively. Considering the difference between ODEs and , dividing by $\lambda$ and taking the limit $\lambda \rightarrow 0$, we have the following ODE $$\label{ab6.3} \left \{\begin{array}{l}\displaystyle C_M\dot {W} ^{k,\delta}+ \left(G_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a {\left( h^{k,\delta}\right)}^b+G_\K^{k,\delta} {\left( n^{k,\delta} \right)}^c+G_L^{k,\delta}\right)W^{k,\delta}=\vspace*{0.2cm}\\ \hspace*{0.8cm}-a G_{\Na}^{k,\delta} { \left(m^{k,\delta}\right) }^{a-1} M^{k,\delta} {\left(h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})\vspace*{0.2cm}\\ \hspace*{0.8cm}-bG_{\Na}^{k,\delta} {\left( m^{k,\delta}\right)}^a {\left(h^{k,\delta}\right)}^{b-1}H^{k,\delta}(V^{k,\delta}-E_{\Na})-cG_\K^{k,\delta} {\left(n^{k,\delta}\right)}^{c-1} N^{k,\delta}(V^{k,\delta}-E_\K)\vspace*{0.2cm}\\ \hspace*{0.8cm} -\theta_{\Na} {\left(m^{k,\delta}\right)}^a{\left(h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})-\theta_K {\left(n^{k,\delta}\right)}^c(V^{k,\delta}-E_\K) -\theta_L(V^{k,\delta}-E_L),\vspace*{0.3cm} \\ \displaystyle \dot{\mathcal{X}}+[\alpha_\mathcal{Y}(V^{k,\delta})+\beta_\mathcal{Y}(V^{k,\delta})]\mathcal{X} =[(1-\mathcal{Y})\alpha'_\mathcal{Y}(V^{k,\delta})-\mathcal{Y}\beta'_\mathcal{Y}(V^{k,\delta})]W^{k,\delta};\vspace*{0.2cm} \\ \hspace*{0.2cm} (\mathcal{X},\mathcal{Y})=( M^{k,\delta},m^{k,\delta}),(N^{k,\delta},n^{k,\delta}),(H^{k,\delta},h^{k,\delta}),\vspace*{0.3cm}\\ W^{k,\delta}(0)=0;\hspace*{0.5cm}M^{k,\delta}(0)=0;\hspace*{0.5cm}N^{k,\delta}(0)=0;\hspace*{0.5cm}H^{k,\delta}(0)=0. \end{array} \right.$$ This last equation is yet another system of coupled nonlinear differential equations, depending on the parameter $\boldsymbol{\theta}=(\theta_{\Na},\theta_K,\theta_L)$, representing an arbitrary point in $\mathbb{R}^3$. From Landweber iteration and $\boldsymbol{\theta} \in \mathbb{R}^3$ arbitrary, we have $$\begin{aligned} \langle \bG^{k+1,\delta}-\bG^{k,\delta},\boldsymbol{\theta} \;\rangle_{\mathbb{R}^3} &=&w^{k,\delta}\langle F'(\bG^{k,\delta})^*(V^\delta-F(\bG^{k,\delta})),\boldsymbol{\theta}\;\rangle_{\mathbb{R}^3},\\ &=&w^{k,\delta}\langle F'(\bG^{k,\delta})^*(V^\delta-V^{k,\delta}),\boldsymbol{\theta}\;\rangle_{\mathbb{R}^3}. \end{aligned}$$ By definition of adjoint operator $$\langle \bG^{k+1,\delta}-\bG^{k,\delta},\boldsymbol{\theta} \;\rangle_{\mathbb{R}^3}=w^{k,\delta}\langle V^\delta-V^{k,\delta},F'(x_k)(\boldsymbol{\theta})\;\rangle_{L^2[0,T]},$$ where the internal product in $L^2[0,T]$ is given by $\Phi=\int_0^T ( V^\delta-V^{k,\delta} )W^{k,\delta}\;dt$, and from  and the previous equation, $$\langle \bG^{k+1,\delta}-\bG^{k,\delta},\boldsymbol{\theta} \;\rangle_{\mathbb{R}^3}=w^{k,\delta}\langle V^\delta-V^{k,\delta},W^{k,\delta}\rangle_{L^2[0,T]}.$$ Denoting the last equality by $\Phi$, we gather that $$\label{ab6.5} \Phi= \frac{\langle \bG^{k+1,\delta}-\bG^{k,\delta},\boldsymbol{\theta} \;\rangle_{\mathbb{R}^3}}{w^{k,\delta}}=\langle V^\delta-V^{k,\delta},W^{k,\delta}\rangle_{L^2[0,T]}.$$ From the previous equation and the first equality from ODE , we obtain $$\begin{gathered} \label{ab6.6} \Phi =\int_0^T \left( C_M\dot {U}^{k,\delta}W^{k,\delta}-(G_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a {\left( h^{k,\delta}\right)}^b+G_\K^{k,\delta}{\left( n^{k,\delta}\right)}^c+G_L^{k,\delta})U^{k,\delta}W^{k,\delta}\right) \,dt\\ -\int_0^T \left[(1-{m^{k,\delta}})\alpha_{m^{k,\delta}}' (V^{k,\delta})-{m^{k,\delta}}\beta_{m^{k,\delta}}'(V^{k,\delta})\right]P^{k,\delta}W^{k,\delta}\,dt \\-\int_0^T\left[(1-{n^{k,\delta}})\alpha_{n^{k,\delta}}' (V^{k,\delta})-{n^{k,\delta}}\beta_{n^{k,\delta}}'(V^{k,\delta})\right]Q^{k,\delta}W^{k,\delta}\,dt\\ -\int_0^T\left[(1-{h^{k,\delta}})\alpha_{h^{k,\delta}}' (V^{k,\delta})-{h^{k,\delta}}\beta_{h^{k,\delta}}'(V^{k,\delta})\right]R^{k,\delta}W^{k,\delta}\,dt. \end{gathered}$$ Integrating the first term from by parts, and from the initial $(W^{k,\delta}(0)=0 )$ and final $(U^{k,\delta}(T)=0)$ conditions, we obtain $$\label{ab6.7} \int_0^T C_M\dot {U}^{k,\delta}W^{k,\delta}= \int_0^T C_MU^{k,\delta}\dot{W}^{k,\delta}.$$ Replacing equation in , we have $$\begin{aligned} \Phi &=&-\int_0^T \left( C_M\dot{W}^{k,\delta}+(G_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a{\left( h^{k,\delta}\right)}^b+G_\K^{k,\delta}{\left( n^{k,\delta}\right)}^c+G_L^{k,\delta})W^{k,\delta}\right) U^{k,\delta}\;dt\\ & &-\int_0^T \left[(1-{m^{k,\delta}})\alpha_{m^{k,\delta}}' (V^{k,\delta})-{m^{k,\delta}}\beta_{m^{k,\delta}}'(V^{k,\delta})\right]P^{k,\delta}W^{k,\delta}\;dt\\ & &-\int_0^T\left[(1-{n^{k,\delta}})\alpha_{n^{k,\delta}}' (V^{k,\delta})-{n^{k,\delta}}\beta_{n^{k,\delta}}'(V^{k,\delta})\right]{Q^{k,\delta}}W^{k,\delta}\;dt\\ & &-\int_0^T\left[(1-{h^{k,\delta}})\alpha_{h^{k,\delta}}' (V^{k,\delta})-{h^{k,\delta}}\beta_{h^{k,\delta}}'(V^{k,\delta})\right]R^{k,\delta}W^{k,\delta}\;dt.\end{aligned}$$ Replacing, the first equality from the ODE , in the first integral from the previous equation, we gather $$\begin{gathered} \label{ab6.8} \Phi=\int_0^T aG_{\Na}^{k,\delta}{m^{k,\delta}}^{a-1}{M^{k,\delta}}{\left( h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})U^{k,\delta}\;dt \\+\int_0^T bG_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a{h^{k,\delta}}^{b-1}\mathsf{H}(V^{k,\delta}-E_{\Na})U^{k,\delta}\;dt +\int_0^T cG_\K^{k,\delta}{n^{k,\delta}}^{c-1}\mathsf{N}(V^{k,\delta}-E_\K)U^{k,\delta}\;dt\\ +\int_0^T {\left( m^{k,\delta}\right)}^a{\left( h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})\alpha U^{k,\delta}\;dt +\int_0^T{\left( n^{k,\delta}\right)}^c(V^{k,\delta}-E_\K)\beta U^{k,\delta}\;dt \\ +\int_0^T (V^{k,\delta}-E_L)\gamma U^{k,\delta}\;dt -\int_0^T \left[(1-{m^{k,\delta}})\alpha_{m^{k,\delta}}' (V^{k,\delta})-{m^{k,\delta}}\beta_{m^{k,\delta}}'(V^{k,\delta})\right]P^{k,\delta}W^{k,\delta}\;dt\\ -\int_0^T\left[(1-{n^{k,\delta}})\alpha_{n^{k,\delta}}' (V^{k,\delta})-{n^{k,\delta}}\beta_{n^{k,\delta}}'(V^{k,\delta})\right]Q^{k,\delta}W^{k,\delta}\;dt \\-\int_0^T\left[(1-{h^{k,\delta}})\alpha_{h^{k,\delta}}' (V^{k,\delta})-{h^{k,\delta}}\beta_{h^{k,\delta}}'(V^{k,\delta})\right]R^{k,\delta}W^{k,\delta}\;dt. \end{gathered}$$ Multiplying the second equation from by ${M^{k,\delta}}$, and integrating in the interval $[0,T]$ it follows that $$\begin{gathered} \int_0^TP^{k,\delta}_t{M^{k,\delta}}-\left[ \alpha_{m^{k,\delta}} (V^{k,\delta})+\beta_{m^{k,\delta}}(V^{k,\delta})\right]P^{k,\delta}M^{k,\delta}\;dt=\\-\int_0^Ta G_{\Na}^{k,\delta}{\left(m^{k,\delta}\right)}^{a-1}{\left( h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})U^{k,\delta}{M^{k,\delta}}\;dt. \end{gathered}$$ Integrating by parts the first term from the previous equation, and using the initial conditions ${M^{k,\delta}}(0)=0$ and $P^{k,\delta}(0)=0$ we have $$\begin{gathered} \int_0^T\left(\dot M^{k,\delta}+\left[ \alpha_{m^{k,\delta}} (V^{k,\delta})+\beta_{m^{k,\delta}}(V^{k,\delta})\right]{M^{k,\delta}}\right)P^{k,\delta}\;dt=\\\int_0^Ta G_{\Na}^{k,\delta}{\left(m^{k,\delta}\right)}^{a-1}{\left( h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})U^{k,\delta}{M^{k,\delta}}\;dt. \end{gathered}$$ Then, from the previous equation and the second equation from ODE , for $(\mathcal{X},\mathcal{Y})=({M^{k,\delta}},{m^{k,\delta}})$, $$\begin{gathered} \label{ab6.9} \int_0^TaG_\K^{k,\delta}{\left(m^{k,\delta}\right)}^{a-1}{\left( h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})U^{k,\delta}{M^{k,\delta}}\;dt=\\ \int_0^T\left[(1-{m^{k,\delta}})\alpha_{m^{k,\delta}}' (V^{k,\delta})-{m^{k,\delta}}\beta_{m^{k,\delta}}'(V^{k,\delta})\right]W^{k,\delta} P^{k,\delta}\;dt.\end{gathered}$$ Multiplying the third equation from by $N^{k,\delta}$, and integrating in the interval $[0,T]$ we gather that $$\begin{gathered} \int_0^T\dot Q^{k,\delta}N^{k,\delta}-\left[ \alpha_{n^{k,\delta}} (V^{k,\delta})+\beta_{n^{k,\delta}}(V^{k,\delta})\right]Q^{k,\delta}N^{k,\delta}\;dt=\\-\int_0^TcG_\K^{k,\delta} {\left(n^{k,\delta}\right)}^{c-1}(V^{k,\delta}-E_\K)U^{k,\delta}\;dt. \end{gathered}$$ Integrating by parts the first term from previous equation, and using the initial conditions $N^{k,\delta}(0)=0$ and $Q^{k,\delta}(0)=0$ we have $$\begin{gathered} \int_0^T\left( \dot N^{k,\delta}+\left[ \alpha_{n^{k,\delta}} (V^{k,\delta})+\beta_{n^{k,\delta}}(V^{k,\delta})\right]{N^{k,\delta}}\right)Q^{k,\delta}\;dt=\\\int_0^TcG_\K^{k,\delta}{\left(n^{k,\delta}\right)}^{c-1}(V^{k,\delta}-E_\K)U^{k,\delta}\;dt. \end{gathered}$$ Then, from the previous equation and the second equation from ODE , for $(\mathcal{X},\mathcal{Y})=({N^{k,\delta}},{n^{k,\delta}})$, we have $$\begin{gathered} \label{ab6.10} \int_0^TcG_\K^{k,\delta} {\left(n^{k,\delta}\right)}^{c-1}(V^{k,\delta}-E_\K)U^{k,\delta}\;dt=\\ \int_0^T\left[(1-{n^{k,\delta}})\alpha_{n^{k,\delta}}' (V^{k,\delta})-{n^{k,\delta}}\beta_{n^{k,\delta}}'(V^{k,\delta})\right]W Q^{k,\delta}\;dt. \end{gathered}$$ Multiplying the fourth equation from by $H^{k,\delta}$, and integrating in the interval $[0,T]$ we gather that $$\begin{gathered} \int_0^T\dot R^{k,\delta}H^{k,\delta}-\left[ \alpha_{h^{k,\delta}} (V^{k,\delta})+\beta_{h^{k,\delta}}(V^{k,\delta})\right]R^{k,\delta}H^{k,\delta}\;dt=\\ -\int_0^Tb G_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a{\left(h^{k,\delta}\right)}^{b-1}(V^{k,\delta}-E_{\Na})U^{k,\delta}\;dt. \end{gathered}$$ Integrating by parts the first term from the previous equation, and using the initial conditions $H^{k,\delta}(0)=0$ and $R^{k,\delta}(0)=0$ we have, $$\begin{gathered} \int_0^T\left({\dot H^{k,\delta}}+\left[ \alpha_{h^{k,\delta}}(V^{k,\delta})+\beta_{h^{k,\delta}}(V^{k,\delta})\right]{H^{k,\delta}}\right)R^{k,\delta}\;dt=\\\int_0^Tb G_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a{\left(h^{k,\delta}\right)}^{b-1}(V^{k,\delta}-E_{\Na})U^{k,\delta}\;dt. \end{gathered}$$ Then, from the previous equation and the second equation from ODE , for $(\mathcal{X},\mathcal{Y})=({H^{k,\delta}},{h^{k,\delta}})$, we have $$\begin{gathered} \label{ab6.11} \int_0^Tb G_{\Na}^{k,\delta}{\left( m^{k,\delta}\right)}^a{\left(h^{k,\delta}\right)}^{b-1}(V^{k,\delta}-E_{\Na})U^{k,\delta}\;dt=\\ \int_0^T\left[(1-h^{k,\delta})\alpha_{h^{k,\delta}}' (V^{k,\delta})-h^{k,\delta}\beta_{h^{k,\delta}}'(V^{k,\delta})\right]W^{k,\delta} R^{k,\delta}\;dt. \end{gathered}$$ Substituting equations , , and in , we have $$\begin{gathered} \label{ab6.12} \Phi=\int_0^T {\left( m^{k,\delta}\right)}^a{\left( h^{k,\delta}\right)}^b(V^{k,\delta}-E_{\Na})\theta_{\Na} U^{k,\delta}\;dt +\int_0^T{\left( n^{k,\delta}\right)}^c(V^{k,\delta}-E_\K)\theta_\K U^{k,\delta}\;dt\\ +\int_0^T (V^{k,\delta}-E_L)\theta_{L} U^{k,\delta}\;dt.\end{gathered}$$ Substituting equations , and in equation we gather that $$\label{ab6.13} \Phi=X_{\Na}^{k,\delta}\;\theta_{\Na}+X_K^{k,\delta}\;\theta_\K+X_L^{k,\delta}\;\theta_{L}=\left\langle \left( X_{\Na}^{k,\delta},X_K^{k,\delta},X_L^{k,\delta}\right),\left(\theta_{\Na},\theta_\K,\theta_{L}\right)\right\rangle_{\mathbb{R}^3}.$$ From  and $$\frac{\langle \bG^{k+1,\delta}-\bG^{k,\delta},\boldsymbol{\theta} \;\rangle_{\mathbb{R}^3}}{w^{k,\delta}} =\left\langle \left( X_{\Na}^{k,\delta},X_K^{k,\delta},X_L^{k,\delta}\right),\boldsymbol{\theta}\right\rangle_{\mathbb{R}^3}.$$ Since $\boldsymbol{\theta} \in \mathbb{R}^3$ is arbitrary, we obtain . Proof of Theorem \[Theorem2\] {#AppendixB} ============================= In what follows we prove Theorem \[Theorem2\]. Consider the operator F defined in . Evaluating $\ba^{k,\delta}$ in $F$, we have $F(\ba^{k,\delta})=V^{k,\delta}$, where $V^{k,\delta}$, $m^{k,\delta}$, $n^{k,\delta}$ and $h^{k,\delta}$ solve ODE . Let the $\boldsymbol{\theta}=(\theta_{a},\theta_{b},\theta_{c})\in \mathbb{R}^3$ and $\lambda \in \mathbb{R}$, then $F(\ba^{k,\delta}+\lambda\boldsymbol{\theta})=V^{k,\delta}_\lambda$, where $V^{k,\delta}_\lambda$, $m^{k,\delta}_\lambda$, $n^{k,\delta}_\lambda$ and $h^{k,\delta}_\lambda$ solve $$\label{equati35} \left \{\begin{array}{l} C_M\dot{V}^{k,\delta}_\lambda=I_{\ext}-G_{\Na}{\left({m}_\lambda^{k,\delta}\right) }^{a^{k,\delta}+\lambda\theta_a}{\left({h}_\lambda^{k,\delta}\right)}^{b^{k,\delta}+\lambda\theta_b}\left(V^{k,\delta}_{\lambda}-E_{\Na}\right) \\ \hspace*{0.5cm}-G_\K^{k,\delta}{\left(n_\lambda^{k,\delta}\right)}^{c^{k,\delta}+\lambda\theta_c}\left(V^{k,\delta}_{\lambda}-E_\K\right)-G_{L}\left(V^{k,\delta}_\lambda-E_L\right),\vspace*{0.2cm} \\ \dot{\mathcal{X}} =(1-\mathcal{X})\alpha_\mathcal{X} (V^{k,\delta})-\mathcal{X}\beta_\mathcal{X}(V^{k,\delta}), \quad\text{for }\mathcal{X}=m^{k,\delta}_\lambda,n^{k,\delta}_\lambda,h^{k,\delta}_\lambda, \vspace*{0.2cm} \\ V_\lambda^{k,\delta}(0)=V_0,\;\;\;\;m_\lambda^{k,\delta}(0)=m_0,\;\;\;\;n_\lambda^{k,\delta}(0)=n_0,\;\;\;\;n_\lambda^{k,\delta}(0)=n_0. \end{array} \right.$$ Considering the difference between the ODEs  and , dividing by $\lambda$ and taking the limit $\lambda \rightarrow 0$, we have the ODE $$\label{equati36} \left \{\begin{array}{l}\displaystyle C_M{\dot {W}}^{k,\delta}+\left(G_{\Na}{\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}}+G_\K{\left(n^{k,\delta}\right)}^{c^{k,\delta}}+G_L\right)W^{k,\delta}=\\ \hspace*{0.5cm}-a^{k,\delta} G_{\Na} {\left(m^{k,\delta}\right)}^{a^{k,\delta}-1}M^{k,\delta}{\left(h^{k,\delta}\right)}^{b^{k,\delta}}(V^{k,\delta}-E_{\Na})\\ \hspace*{0.5cm}-\mathsf{b}G_{\Na} {\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}-1}H^{k,\delta}(V^{k,\delta}-E_{\Na})\\ \hspace*{0.5cm} -c^{k,\delta}G_\K {\left(n^{k,\delta}\right)}^{c^{k,\delta}-1}N^{k,\delta}(V^{k,\delta}-E_\K)\\ \hspace*{0.5cm} -G_{\Na} {\left(m^{k,\delta}\right)}^{a^{k,\delta}}\ln( m^{k,\delta}) {\left(h^{k,\delta}\right)}^{b^{k,\delta}}(V^{k,\delta}-E_{\Na})\theta_a\\ \hspace*{0.5cm}-G_{\Na}{\left(m^{k,\delta}\right)}^{a^{k,\delta}} {\left(h^{k,\delta}\right)}^{b^{k,\delta}}\ln( h^{k,\delta})(V^{k,\delta}-E_{\Na})\theta_b\\ \hspace*{0.5cm} -G_k{\left(n^{k,\delta}\right)}^\mathsf{c}\ln(n^{k,\delta})(V^{k,\delta}-E_\K)\theta_c,\vspace*{0.2cm} \\ \dot{\mathcal{X}}+[\alpha_\mathcal{Y}(V^{k,\delta})+\beta_\mathcal{Y}(V^{k,\delta})]\mathcal{X} =[(1-\mathcal{Y})\alpha'_\mathcal{Y}(V^{k,\delta})-\mathcal{Y}\beta'_\mathcal{Y}(V^{k,\delta})]{W^{k,\delta}}, \\ \hspace*{0.2cm}(\mathcal{X},\mathcal{Y})=({M^{k,\delta}},m^{k,\delta}),(N^{k,\delta},n^{k,\delta}),(H^{k,\delta},h^{k,\delta}),\vspace*{0.2cm}\\ W^{k,\delta}(0)=0,\;\;\;M^{k,\delta}(0)=0,\;\;\;N^{k,\delta}(0)=0,\;\;\;H^{k,\delta}(0)=0. \end{array} \right.$$ where $W^{k,\delta}$ is defined in equation by replacing $\bG^{k,\delta}$ by $\ba^{k,\delta}$. Also, $M^{k,\delta}$, $N^{k,\delta}$ and $H^{k,\delta}$ are defined in equation . This last equation is again a system of coupled nonlinear differential equations, parametrized by $\boldsymbol{\theta}=(\theta_a,\theta_b,\theta_c)$, where $\boldsymbol{\theta}\in\mathbb{R}^3$ is arbitrary. Considering , and proceeding as in Appendix \[AppendixA\], we gather . [^1]: The second author acknowledges the financial support of the Brazilian funding agency CNPq
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this note we propose to show how to find stable models of a one-parameter family of elliptic surfaces. The strategy we use was initiated by Abramovich-Vistoli in [@A-V]: that is to say to consider a fibred surface as a map from the base curve to the moduli of stable n-pointed curves, and to consider then the Kontsevitch space of such maps that are stable. In our case we will then be describing the moduli space of stable surfaces via the moduli space of Kontsevitch stable maps to $ {{\overline {\mathbf{M}}}}_{1,1},$ the moduli space of stable one-pointed elliptic curves. Then, following Kollár–Shepherd-Barron ([@K-S]) and V. Alexeev ([@A1]) we apply Mori’s Minimal Model Program in an explicit manner by means of toric geometry.' address: | Courant Institute of Mathematical Sciences\ New York University\ 251 Mercer Street\ New York, NY 10012\ USA author: - Gabriele La Nave title: Explicit stable models of elliptic surfaces with sections --- Introduction ============ Description of the problem -------------------------- When speaking of a [**surface**]{} we will always mean a reduced, integral, normal, projective noetherian scheme of dimension $2.$ For the purpose of studying the boundary of the moduli space, we need to introduce the following generalization of the well-known concept of [**elliptic surface**]{} (for convinience, we will mantain the name): An [**elliptic surface with zero section**]{} is the datum of a surface $X$ together with a proper map, $\pi : X \to C$ to a proper curve $C$ and a section $\sigma : C \to X,$ called the [**zero section**]{}, such that: 1. the generic fibre of $\pi$ is a stable complete curve of arithmetic genus $1;$ 2. the zero section is not contained in the singular locus of $\pi.$ Such an object is called [**relatively minimal**]{} if it is smooth and there is no $(-1)\text{-curve}$ in any fibre. Furthermore, we say it is [**minimal**]{} if it is smooth and contains no $(-1)\text{-curve}$ at all. Notice that if the base curve is not rational, the last two notions coincide. Note that our notion of elliptic surface differs from the usual one in that in point $1$ we ask the generic fibre to be stable, as opposed to the usual notion where the generic fibre is required to be smooth, and in that we do not ask the map $\pi$ to be flat. Whenever we are given such an object, we automatically get a rational map $C \dashrightarrow {{\overline {\mathbf{M}}}}_{1,1}$ to the moduli space of $1$-pointed elliptic curves: $ {{{\overline {\mathbf{M}}}}} _{1,1}$. If the base curve were smooth, we would then extend the map to a regular map on the whole $C\to {{\overline {\mathbf{M}}}}_{1,1}$. Composing this map with $\pi$ provides us with a regular map $X\to C \to {{{\overline {\mathbf{M}}}}}_{1,1} .$ Let $S$ be a fixed base scheme. Let $\pi: X\to C \to S$ be a proper $S$-map of relative dimension $2$ from a proper scheme $X$ to a proper scheme $C,$ with $X \to C$ generically fibred in stable curves of genus $1.$ Following Alexeev and Kollár–Shepherd-Barron (cf. [@A1], [@A2] and [@K-S]), we are naturally led to the following: \[logstabsurf\] A [*pair* ]{} $(\pi :{{\mathcal X}}\to {{\mathcal C}}\to S,{{\mathcal Q}})$ consisting of an $S$-morphism $X \to C$ fibred in generically stable curves of genus $1$ and a [*section* ]{} ${{\mathcal Q}}$ is called [**stable** ]{} if, for each geometric point $s$ in $S$: 1. ${{\mathcal X}}\to S$ and ${{\mathcal C}}\to S$ are flat morphisms of relative dimension $2$ and $1$ respectively; 2. the pair $({{\mathcal X}}_s , {{\mathcal Q}}_s )$ has semi-log-canonical singularities (see section \[prelim\] for a definition); 3. the relative log-canonical sheaf $ \omega _{X/S} (Q)$ is ${{\mathbb{Q}}}$-Cartier; 4. the relative log-canonical sheaf $ \omega _{X/S} (Q)$ is $S$-ample. If $S=Spec(k)$ for any field $k, $ we simply say that the pair is [**stable**]{} Similarly, a [*triple* ]{} $({{\mathcal X}}\to {{\mathcal C}}\to S,{{\mathcal Q}}, f: {{\mathcal X}}\stackrel{ \pi}{\to}{{\mathcal C}}\stackrel {j}{\to} {{{\overline {\mathbf{M}}}}}_{1,1})$ consisting of an $S$ morphism ${{\mathcal X}}\to {{\mathcal C}}$ fibred in generically smooth elliptic curves, a [*section* ]{} ${{\mathcal Q}}$ and a map $f$ as above will be called [**stable**]{} if conditions  1,  2 and  3 above hold and if $ \omega _{X/S} (Q) \otimes f^* {{\mathcal O}}(3)$ is [*ample*]{}. If $S=Spec(k)$ for any field $k, $ we simply say that the triple is [**stable**]{} Given a relatively minimal elliptic surface $X$ with section $Q,$ it is natural to consider its associated Weierstrass model $Y$: this is roughly obtained by contracting all the components of the fibres that do not meet $Q$ (see section \[wform\] for a more precise definition). This surface $Y$ has possibly a finite number of cuspidal fibres, with the property that near each such fibre $Y_0,$ the surface $Y$ has a local equation of the form $y^2 = x^3 + a x +b$ with $a,b \in {{\mathcal O}}_{C,0}$ with $min (\nu _0(a^3),\nu _0 (b^2)) <12 .$ Here $\nu _ 0 (g )$ denotes the order of vanishing of $g \in {{\mathcal O}}_{C,0}$ at $0.$ Such equations are called [**minimal Weierstrass equations**]{} and $Y$ is said to be in [**minimal Weiertrass form**]{}. Loosely speaking Weierstrass models, within the theory of elliptic surfaces, play the role that canonical models do for surfaces of general type. Indeed if the base curve $C$ is smooth and non-rational, they constitute the log-canonical models for elliptic surfaces with sections having $C$ as base curve (see Corollary \[semiampl\]). For definition \[logstabsurf\] to be of any use, we want the moduli functor it defines to include at least a large part of the locus of minimal Weierstrass equations. In other words, we want most pairs $(\pi:X \to C,Q)$ consisting of an ellitpic surface in [*minimal Weierstrass equation*]{} $X\to C$ with section $Q,$ to be stable according definition \[logstabsurf\]. Similarly, we want most triples $(\pi:S \to C,Q, X\stackrel{\pi}{\to} C \stackrel{j}{\to} {{\overline {\mathbf{M}}}}_{1,1})$ to be stable according to definition \[logstabsurf\]. In section \[wform\] we will show that a pair $(X\to C, Q)$ with $X\to C$ in minimal Weierstrass form is [ *stable* ]{} if $C$ is not rational, and that a triple $(\pi:S \to C,Q, X\stackrel{\pi}{\to} C \stackrel{j}{\to} {{\overline {\mathbf{M}}}}_{1,1})$ with $X\to C$ in minimal Weierstrass form is [*stable* ]{} either if $C$ is not rational, or, if it is rational, if $\pi: X \to {{\mathbb{P}}}^1$ is not an isotrivial family (i.e., mapped to a point in ${{\overline {\mathbf{M}}}}_{1,1}$ via $j$). This sets the course for us: the moduli problems we will be looking at are the ones defined by the functors: $$\begin{array}{cccc} {{\mathcal M}}_{pairs}: &{\mathcal Sch} &\longrightarrow & {\mathcal Sets} \\ & S & \to & {{\mathcal M}}_{pairs}(S) \end{array}$$ where: $${{\mathcal M}}_{pairs}(S):= \left\{ \begin{array}{ccc} &\text{ pairs over } S \ (\pi:{\mathcal X}\to {\mathcal C}, \mathcal Q,): \text{ which }\\ &\text{ are stable according to definition \ref{logstabsurf} } \end{array} \right\} / {\text{isomorphisms} }$$ and: $$\begin{array}{cccc} {{\mathcal M}}_{triples}: &{\mathcal Sch} &\longrightarrow & {\mathcal Sets} \\ &S & \to & {{\mathcal M}}_{triples}(S) \end{array}$$ where: $${{\mathcal M}}_{triples}(S):= \left\{ \begin{array}{ccc} &\text{ triples over } S \ (\pi:{\mathcal X}\to {\mathcal C}, \mathcal Q, f: {\mathcal X}\stackrel {\pi}{\to} {{\mathcal C}}\stackrel {j}{\to} { {{{\overline {\mathbf{M}}}}}}_{1,1}):\\ &\text{ which are stable according to definition \ref{logstabsurf} } \end{array} \right\} /{\text{isomorphisms} }$$ To render our functors of finite type, it is essential to fix some numerical invariants. The pair of rational numbers: $$A = c_1(\omega_X(Q))^2;\quad \chi = \chi ({{\mathcal O}}_X)$$ for the moduli of pairs and the triple of rational numbers: $$A = c_1(\omega_X(Q))^2;\quad B= c_1(\omega_X(D))\cdot c_1(f^*{{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}}(1)); \quad C = c_1(f^*{{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}}(1)))^2,$$ for the moduli of triples. Once a moduli problem is defined, there are two crucial questions one must pose and answer (preferably positively): 1)Is the associated functor proper? Is it a Deligne-Mumford stack? Here we will be addressing only the question of properness. Specifically we will prove that the functors ${{\mathcal M}}_{pairs}$ and ${{\mathcal M}}_{triples}$ satisfy the valutative criterion of properness. Such a criterion is usually called “stable reduction theorem” for moduli problems. In our case one has to prove that in a family of stable pairs (resp. triples) over the punctured disk, one can replace the central fibre of any compactification over the whole disk by a stable one, possibly after a base change. Previous work ------------- It must be said that the questions we are addressing here have been addressed and answered in much greater generality by J. Kollár and N. Shephard-Barron (cf. [@K-S]) and by V. Alexeev (cf. [@A1], [@A2]). In [@K-S] J. Kollár and N. Shephard-Barron prove a stable reduction theorem for the moduli of pairs $(X,D)$ consisting of a [*semi-log-canonical* ]{} surface $X$ and a ${{\mathbb{Q}}}$-Cartier divisor $D.$ In [@A1] and [@A2], V. Alexeev generalizes this to moduli of triples $(X,D, f:X \to M)$ where $X$ is a surface with at most semi-log-canonical singularities, $D \subset X$ is a ${{\mathbb{Q}}}\text{-Cartier}$ divisor, $f:X \to M$ is a proper morphism to a projective scheme $M$ and $\omega _{X} (Q) \otimes f^* A$ is ${{\mathbb{Q}}}$-Cartier and ample for every choice of a sufficiently ample divisor $A$ on $M.$ Moreover he proves that the corresponding moduli functor is bounded (in particular for $M={\operatorname{Spec}}(k)$ a point, one gets that the moduli space of pairs is bounded). Their proofs make use of the full strenght of Mori’s MMP (Minimal Model Program). Also, in the case of Weiestrass fibrations over ${{\mathbb{P}}}^1,$ in [@M], R. Miranda constructs a proper moduli space by identifying the GIT semi-stable points of the action of $k ^* \times SL(V_1 )$ on a suitable subset $T_N \subset V_{4N} \oplus V_{6N}.$ Here $V_1 = H^0({{\mathbb{P}}}^1, {{\mathcal O}}_{{{\mathbb{P}}}^1} (1)),$ $V_k = H^0({{\mathbb{P}}}^1, {{\mathcal O}}_{{{\mathbb{P}}}^1} (k))=Sym ^k (V_1 ),$ and the component $k^*$ of $k ^* \times SL(V_1 )$ acts on $T_N$ by $\lambda (A,B) = (\lambda ^{4N} , \lambda ^{6N})$ and $SL(V_1 )$ acts on $T_N$ by the action induced on it by the natural action of $SL(V_1 )$ on $Sym ^k(V_1 ).$ What we set out to do here, though, is to give an explicit description of the stable reduction process (in the sense of Alexeev–Kollar–Shephard-Barron, ie., in the sense of definition \[logstabsurf\]) in  1-parameter families and of the possible surfaces we may get at the boundary of the moduli spaces of elliptic pairs and of elliptic triples respectively. In fact, the work of Kollar–Shepherd-Barron and of Alexeev proves the valuative criterion for properness abstractly by means of the MMP. In order to explain better the significance of such an explicit description and how it compares with the construction of Kollar–Shepherd-Barron and of Alexeev, we may refer to the case of the space ${{\overline {\mathbf{M}}}}_{g,n}$ of [**Deligne- Mumford**]{} ([**DM**]{} for short) [**stable curves**]{} of a given genus g with sections $\sigma _i$ with $i=1,...,n,$ which compactifies the moduli space of [*smooth*]{} curves of genus $g$ with $n$ marked points ${{{\mathbf{M}}}}_{g,n}$ (cf. [@D-M] or [@H-M]). A (geometrically connected and proper) curve ${{\mathcal C}}\to S$ over a scheme $S$(resp. with sections $\sigma _i : S \to {{\mathcal C}}$) can be defined to be a D-M stable curve in two different and equivalent ways: [**the abstract description:**]{} 1. the singularities of ${{\mathcal C}}_s$ are at worst nodal for every closed point $s\in S;$ 2. the relative dualizing sheaf $\omega _{{{\mathcal C}}/S}$ (resp. $\omega _{{{\mathcal C}}/S}(\sum _{i=1} ^n \sigma _i)$ ) is ample. [**the combinatorial description:**]{} 1. as $1$ above, 2. if there is a rational irreducible component $R$ of ${{\mathcal C}}_s$ (for some closed point $s\in S$) then $R$ must meet the rest of ${{\mathcal C}}_s$ in at least $3$ points (resp. $R$ must contain at least three points that are either nodes of ${{\mathcal C}}_s$ or [*marked points*]{} coming from the sections $\sigma _i$); if there is a component $E$ of arithmetic genus $1,$ then $E$ must meet the rest of ${{\mathcal C}}_S$ in at least one point. An analogous picture holds in the case of [*Kontsevich stable maps*]{}. In this work we try to transpose the work of J. Kollár and N. Shepard-Barron and of V. Alexeev into an incarnation that would correspond to the combinatorial picture of the DM-stable curves given above. Up to now there are very few cases in which the degenerations in a  1-parameter family have been described explicitly. To name one of these, recently B. Hasset [@Ha] gave a description for the boundary of the moduli of pairs $({{\mathbb{P}}}^2, C)$ where $C$ is a plane quartic, and built an isomorphism between this space and ${{\overline {\mathbf{M}}}}_3,$ the moduli space of Deligne-Mumford stable genus $3$ curves (this is of importance also for the undersdanding of the locus of [*limiting plane curves*]{} in ${{\overline {\mathbf{M}}}}_ g,$ given that this locus coincides with the whole ${{\overline {\mathbf{M}}}}_3$ for degree $4$ curves). The present work adds to the list of cases that have been worked out. The strategy ------------ The strategy for the moduli of triples is the one initiated by Abramovich and Vistoli in [@A-V]: we consider an elliptic surface as a map from the base curve to ${{\overline {\mathbf{M}}}}_{1,1}.$ By marking the points on the base curve corresponding to the cuspidal fibres, thanks to the [**Purity Lemma**]{} of Abramovich and Vistoli (cf. [@A-V]), we can replace these fibres by finite cyclic quotients of stable curves (we will refer to such curves as [**twisted curves**]{}). The idea is to now make the map from the base curve to ${{\overline {\mathbf{M}}}}_{1,1}$ Kontsevich stable. The result is that at the limit there are surfaces $X$ that map to curves $C;$ these curves come endowed with a Kontsevich stable maps $C \to {{\overline {\mathbf{M}}}}_{1,1},$ wich correspond to the map to moduli; the general fibre of $X \to C$ is a stable curve and the componts of $X$ meet along fibres that are either stable or twisted. The problem is that one had to replace the cuspidal curves in our original elliptic surfaces: the new surfaces are not in Weierstrass form anymore. In order to deal with this problem, we prove an extension lemma (see lemma \[Extension Lemma\]), that allowes us to place back the cuspidal fibres, and so we can remove the extra marked points of the base. The double curves of $X$ are either stable or the [*twisted fibres*]{} of Abramovich–Vistoli (cf [@A-V2]). Looking at the irreducible components, we are therefore led to enlarging the class of [*minimal Weierstrass surfaces*]{} to what we call [**quasiminimal elliptic surfaces**]{}. Roughly speaking, these are elliptic surfaces $X\to C$ over smooth curves $C,$ which are in Weierstrass form away from a finite number of [*twisted fibres*]{} and such that the local Weierstrass equation away from the twisted fibres is minimal. The price we have to pay in removing the extra marked points, is that the map from the base curve to moduli might not be Kontsevich stable anymore: there may be isotrivial components mapping to a rational component of the base curve $C$, that meet the rest of $C$ in one or two points. It turns out (proposition \[extr ray\]) that the componets of $X$ dominating these isotrivial rational components are unstable. To be precise, the zero section of these components is an [*extremal ray* ]{} if the base curve meets the rest of $C$ in only one point, and it is contracted by the log-canonical map if it meets the rest of $C$ in two points. To deal with these components, we start by performing a few explicit birational transformations by means of local Weierstrass equations; in doing so we are led to enlarging the class of [*quasiminimal elliptic surfaces*]{} to what we call [**standard elliptic surfaces**]{}, which are, roughly speaking, elliptic surfaces with a finite number of twisted fibres, for which the local Weierstrass equations away from the twisted fibres are of the form $y^2 = x^3 + a x +b,$ for $a,b \in {{\mathcal O}}_{C,p},$ with the minimum of the order of vanishing of $a^3$ and $b^2$ not greater than $12$ at each point $p \in C.$ Note that if the above mensioned minimum does achieve $12$ at some point $p \in C,$ the surface has an [*elliptic singularity*]{} at the point $x=y=0$ over $p$ (see section \[LCWE\]). We continue by following the steps of the Minimal Model Program (MMP). By means of toric geometry we are able to explicitly perform the necessary log-flips and small log-contractions. The forgetfull functor sending, for each scheme $S,$ an $S$-triple $({{\mathcal X}}, {{\mathcal Q}},f:{{\mathcal X}}\to {{\mathcal C}}\to {{\overline {\mathbf{M}}}}_{1,1})$ to the $S$ pair $({{\mathcal X}}, {{\mathcal Q}}),$ is not well-defined at the level of moduli: it does not always produce a stable pair out of a stable triple. In fact, every component $X\to {{\mathbb{P}}}^1$ of a geometric fibre ${{\mathcal X}}_s \to {{\mathcal C}}_s ,$ for some geometric point $s$ in $S,$ that meets ${{\mathcal X}}_s$ in less than three fibres is [*unstable*]{}. Therefore, we need to perform more steps of the MMP. It will turn out that the same explicit operations described in the case of triples, go through for these more general settings in which the $j$-invariant associated to such components $X \to {{\mathbb{P}}}^1$ is not constant. The price we have to pay is that the map ${{\mathcal X}}\to {{\overline {\mathbf{M}}}}_{1,1} $ is no longer regular. When the general base curve ${{\mathcal C}}_{\eta}$ is isomorphic to ${{\mathbb{P}}}^1 ,$ the log-canonical bundle for the pair $({{\mathcal X}}, {{\mathcal Q}})$ is not nef: it is negative on ${{\mathcal Q}}_{\eta}.$ We therefore need to contract the zero-section in the central fibre, to obtain a [*stable pair*]{}. In order to do so we need to perform some extra birational operations on the total space of the family ${{\mathcal X}}.$ The result ---------- We need some definitions first. Loosely speaking a [**log-standard**]{} elliptic surface $X\to C$ (see definition \[logstand\]) is a certain explicit locally toric blow-up (the unique such [*semi-logcanonical*]{} blow-up) of a pair $(Y\to C, G+F+Q)$ consisting of a standard elliptic surface $Y\to C$ together with the marking of the zero section $Q$ and of some (possibly twisted) fibres $G= \bigcup _i G_i$ and $F= \bigcup _i F_i;$ the centers of the blow-ups are supported on points in $G_i \cap Q . $ We call [**splice**]{} the proper transform of a fibre $G_i$ of $Y \to C$ on which the point we blow-up is supported, and [**scion** ]{} the proper transform of each of the $F_i $’s. Such a [*log-standard elliptic surface*]{} is called [**strictly stable**]{}, if $Q ^2 < -1 .$ Our results are stated in Theorems \[stabredpairs\], \[stabred\], \[rationalbase\] and \[ellipticbase\]. Theorems \[stabredpairs\] and \[stabred\] roughly speaking say that, in the case in which the base curve has genus $g\geq 2,$ at the boundary we get a union of surfaces $X =\cup X_i$ attached one to the other along curves which are either stable or twisted fibers or along splices, mapping to a curve $\pi: X \to C ,$ with a section $Q .$ For the [**moduli of pairs**]{}, $C$ is a Deligne-Mumford stable curve and the components of $X$ are (possibly) of two kinds: 1. strictly stable log-standard components mappping dominantly onto irreducible components of $C;$ 2. components mapped to a point of $C$ via $\pi .$ The components mapped to a point, which we call [**pseudoelliptic**]{} (see definitions \[I\] and \[II\]), are further devided in two types. For the moduli of [*pairs*]{} $(X,Q),$ we have: 1. [**type I:**]{} A pair $(Y, G+F)$ consisting of a surface $Y$ endowed with a [*structure morphism*]{} $f:Y'\to Y ,$ which is regular and birational, from a [*log-standard*]{} surface $(Y' \to {{\mathbb{P}}}^1, Q+ G' + F'), $ with $1$ [ *scion*]{} $F'$ and a number of [*splices*]{} $G' = \bigcup _i G_i '$ and such that $F= f(F')$, $G_i = f({G _i}')$ and $G= \bigcup G_i.$ The surface $Y$ is attached to the rest of $X$ along $F$ and possibly along some (or all) of the $G_i$’s. Furthermore, the morphism $Y' \to Y$ is obtained by torically (and explicitly) blowing-down the zero section $Q$ of $Y' ;$ 2. [**type II:**]{} A pair $(Y ,G + F_1 +F_2 ) $ consisting of a surface $Y$ endowed with a birational morphism $f: Y' \to Y$ (the [*structure morphism*]{}) from a [*log-standard*]{} surface $(Y', Q+ G' +{F _1 } '+ {F_2 }')$ with $2$ [*scions*]{} ${F _1 } ' $ and $ {F_2 }' $ and a number of [*splices* ]{} $G' = \bigcup _i {G_i} ';$ also $ F_i = f({F _i}')$, $G_i = f({G _i}')$ and $G= \bigcup G_i.$ Moreover $Y$ is attached to the rest of $X$ along the $F_i$’s and possibly along some (or all) of the $G_i$’s. Furthermore, the morphism $Y' \to Y$ is obtained by torically (and explicitly) blowing-down the zero section $Q$ of $Y' ;$ We still call [*splice* ]{} and [*scion*]{} the images in a [ *pseudoellitpic surface*]{}, either of type I or of type II, of the corresponding curves via the structure morphism. The attaching of a scion to a splice is étale locally described by the fan given in Theorem \[log-flip\]. The attaching of a [*pseudoellitpic surface of type II*]{} along a marked fibre of $X$ is described étale locally by the cone given in Theorem \[logcancont\]. For the [**moduli of triples**]{} $(X,Q, f:X \to {{\overline {\mathbf{M}}}}_{1,1})$ we have the same situation except for having the [*isotrivial*]{} analogues of [**type I**]{} and [**type II**]{}, and we ask that the map $C\to {{\overline {\mathbf{M}}}}_{1,1}$ be Kontsevich stable. The [**type I**]{} surfaces arise from log-flips, and the [**type II**]{} ones from the log-canonical contraction of the zero section. (350,320)(0,0) (180,50)[(90,200) [![image](stable.eps)]{}]{} (200,10)[[fig.1]{}]{} In the rational base case (see Theorem \[rationalbase\]), there is no base curve at all. At the boundary here we get a union $X= \cup X_i$ where $X_i$ is a [**pseudoelliptic surface of type N**]{} with $N=0$ or $N=1.$ Furthermore the type $1$ pseudoelliptic surfaces come in pair, and they are attached one to the other (but not along as a scion). Here, by [**pseudoelliptic surface of type N**]{} we mean, loosely speaking, a surface $S$ and a map $g: S' \to S$ from a log-standard surface $(S\to {{\mathbb{P}}}^1, Q+\sum _{i=1} ^N F_i +\sum _{i=1} ^n G_i)$ with [**N** ]{} marked stable or twisted fibres $F_i$ and $n$ [*splices*]{} $G_i, $ furthermore, $g$ is obtained by an explicit toric blow-down of the zero section $Q$ of $Y.$ In the elliptic base case (see theorem \[ellipticbase\]) we have two more types of surfaces: type $E_0$ and type $E_{I_N}.$ Loosely speaking, a [**type $E_0$ pseudoelliptic surface**]{} (resp. [**type $E_{I_N}$ pseudoelliptic surface**]{}) is a surface $S$ endowed with a map $g: S' \to S$ from a log-standard elliptic surface $(S\to E, Q+F)$ with one marked fibre, and with base $E$ an elliptic curve (resp. a closed chain of ${{\mathbb{P}}}^1$’s), and such that the exceptional set of $g$ is the zero-section $Q.$ A type $E_0$ (resp. type $E_{I_N}$) pseudoelliptic surface has an elliptic (resp. degenerate cusp) singularity at $g(Q).$ Aknowledgements --------------- I would like to thank my advisor Dan Abramovich, whose invaluable and essential incouragement and teachings have been the sole reason for my accomplishing these results. The clarity of the exposition has benefited tremendously from his constant reading. Of course, any mistake of this work is to be blamed on me. Thanks are also due to Kenji Matsuki, for suggesting that the log-flips should be toric. On a personal level, I must express my gratitude to my parents, Angelo and Milva, and my wife Verity. Preliminaries {#prelim} ============= In this section we will give a brief list of results we need from toric geometry and the theory of log-canonical surfaces. Varieties are always integral, reduced schemes of finite type over an algebraically closed field $k$ of characteristic zero. Let $S$ be a scheme of finite type over $k .$ If $X$ is proper over $S$ we set: $$N_1 (X/S) = \{ \text{ ~1-cycles of } X/S \text{ modulo numerical equivalence} \} .$$ We then have a bilinear pairing: $$Pic(X) \times N_1 (X/S) \to {{\mathbb{Q}}}$$ defined by extending by linearity the map that associates $deg_C(\mathcal L)$ to the pair $({{\mathcal L}}, C),$ with ${{\mathcal L}}$ a line bundle on $X$ and $C$ an effective irreducible curve. We also set: $$NE(X/S)=\{ B \in N_1 (X/S) \text{: } Z \equiv \sum a_i C_i \text{, } a_1 \in {{\mathbb{Q}}}^+ \cup {0}\}$$ and we denote by ${\overline {NE}}(X/S)$ its closure with respect to the euclidean topology in $H^2(X,{{\mathbb{R}}}.)$ [**Kleiman’s Criterion of ampleness**]{} states that, if $X$ is $S \text{-proper} ,$ a divisor $A$ is $S$-ample on $X$ if and only if $A \cdot x > 0$ for each $x \in {\overline {NE}}(X/S) \setminus \{0 \}.$ An [**extremal ray**]{} is a ray $R \subset {\overline {NE}}(X/S)$ such that if $x_1 + x_2 \in R$ then $x_1 \text{, } x_2 \in R ,$ for each $x_1 , x_2 \in {\overline {NE}}(X/S).$ By the [**cone theorem**]{} (see [@Mo] for the smooth case and [@KMM] for the general case), given an extremal ray $R$ of $NE(X)$ there exists an [**extremal contraction**]{}, namely a morphism $\phi _R : X \to Y$ such that: 1. ${\phi _R} _* {{\mathcal O}}_X \simeq {{\mathcal O}}_Y ,$ i.e., $\phi _R$ has connected fibres; 2. a curve $C \subset X$ is contracted by $\phi _R$ if and only if its class $[C]$ in ${\overline {NE}}(X/S)$ is such that $[C] \in R .$ Let $\phi : X \to Y$ be an extremal contraction such that the codimension of the exceptional set $E \subset X$ is $\geq 2,$ and let $D$ be a ${{\mathbb{Q}}}-$Cartier divisor of $X.$ A variety $X ^+ ,$ together with a birational morphism: $$\phi ^+ : X^+ \to Y$$ is called a [**flip** ]{} (resp. [**log-flip**]{}) of $\phi$ if: 1. $X^+$ has only [*log-canonical singularities*]{} (see section \[slc\] below) 2. $K _{X^+}$ (resp. $K _{X^+} +D^+ $ where $D^+$ is the closure of $(f^+ \circ f^{-1})(D)$) is $f^+ $-ample. 3. the exceptional set of $\phi ^+$ has codimension $\geq 2$ in $ X^+ .$ Semi-log-canonical singularities {#slc} -------------------------------- By definition, a reduced scheme of finite type $X$ is said to be [**${{\mathbb{Q}}}$-Gorenstein**]{} if $\omega_X ^{[n]}$ is locally free for some $n$. Here $\omega_X ^{[n]} = ((\omega_X ^{\otimes n})^\vee )^\vee .$ For a ${{\mathbb{Q}}}$-Gorenstein variety, the smallest such $n$ is called the [*index*]{}. Following [@K-S] A surface $X$ is [**semi-smooth**]{} if it has only the following singularities: 1. $2$-fold normal crossings with local equation $x^2=y^2$ 2. pinch points with local equation $x^2=zy^2$ and A [**good semi-resolution**]{} of a surface $X$ is a proper map $g:Y \longrightarrow X$ satisfying the following properties 1. [ $Y$ is semi-smooth]{} 2. [ $g$ is an isomorphism in the complement of a codimension two subscheme of $Y$]{} 3. [ No component of the double curve $D$ of $Y$ is exceptional for $g$.]{} 4. [The components of $D$ and the exceptional locus of $X$ are smooth, and meet transversally.]{} Finally, given a birational map $g:Y \to X, $ we call [**discrepancies of**]{} $K_X + D$ [**associated to**]{} $g$ those integers $a_i$ such that: $\omega_Y^n ( \overline D ) = g^*\omega^{[n]}_X(D +ka_1 E_1 +...+ ka_n E_N),$ where $\overline D$ is the pushforward of $D$ via $g ^{-1} .$ When $D$ is the empty set, we call the $a_i$ just [**discrepancies**]{}. A surface $X$ is said to have [**semi-log-canonical**]{} singularities if 1. $X$ is Cohen-Macaulay and ${{\mathbb{Q}}}$-Gorenstein 2. $X$ is semi-smooth in codimension one 3. The discrepancies $a_i$ of a good semi-smooth resolution of $g:Y \to X$ are all greater than or equal to $-1 .$ It is not difficult to show that a surface $X$ is [**semi-log-canonical**]{} if and only if it is $S2$ and its normalization $\nu :X ^{\nu} \to X$ is such that the pair $(X, K_X + D)$ where $D \subset X ^{\nu}$ is the double curve of $\nu ,$ is log-canonical. Toric varieties --------------- Toric varities are obtained by suitably patching affine toric varieties, which are, roughly speaking, [*normal* ]{} zero sets of binomials. Given a lattice $N\cong\ {{\mathbb{Z}}}^n$ and a [*strictly convex rational polyhedral cone*]{} $\sigma\subset N_{{\mathbb{R}}}= N\otimes _{{\mathbb{Z}}}{{\mathbb{R}}}$ we will denote by $X_\sigma $ tghe affine toric variety associated with $\sigma$ (see [@F]) We denote by $\sigma^{(1)}$ the 1-dimensional edges of $\sigma$. The variety $X_\sigma$ is nonsingular if and only if the primitive points of $\sigma$ form a part of a basis of $N$. The toric variety $X_\sigma$ contains an n-dimensional algebraic torus $T={\mathbb G }^n_m$ as an open dense subset, and the action of $T$ on itself extends to a linear action on $X_\sigma$ ( hence the alternate name [*torus embedding*]{}). Thus, $X_\sigma$ is a disjoint union of orbits of this action. There is a one-to-one correspondence between the orbits and the faces of $\sigma$. In particular, 1-dimensional faces ${{\mathbb{R}}}_{+}v_i$ correspond to codimension 1 orbits ${\mathbb O}_{v_i}$. A [**fan**]{} $\Delta \subset N$ is a collection of strongly convex rational polyhedral cones such that:  1) each face of a cone in $\Delta $ is also a cone in $\Delta$;  2) the intersection of two cones in $\Delta$ is a common face of the two. To fans one associates [**toric varieties**]{}, that are obtained by suitably patching the affine toric varieties corresponding to the cones of the fan. A [**toric morphism**]{} $f:X_\sigma\to X_\tau$ is a dominant equivariant morphism of toric varieties corresponding to a linear map $f_\Delta: (N_\sigma,\sigma) \to (N_\tau,\tau)$. In this paper we will write $\langle f_1,...,f_2 \rangle $ for the cone in $N _{{{\mathbb{R}}}}$ generated by the lattice vectors $f_1,...,f_n \in N$ for some lattice $N.$ For a toric variety $ {NE}(X/S)$ is a closed cone, hence ${\overline {NE}}(X/S)= NE(X/S) .$ Indeed M. Reid in [@R1] proves: If $f: X \to S$ is a proper toric morphism, and if X is proper, then $$NE(X/S) = \sum {{\mathbb{Q}}}^+ {{\mathbb{O}}}_w$$ where ${{\mathbb{O}}}_w$ runs through the  1-dimensional strata of $X$ in fibre of $f.$ Futhermore, if $X$ is projective, $ NE(X/S) $ is spanned by a finite number of [*extremal rays*]{}. Set $\Delta ^{k} = \{ \text{ k-dimensional cones of }\Delta \} .$ If $\sigma _1 = \langle e_1, ...,e_{n-1}, e_n \rangle $ and $\sigma _2 = \langle e_1, ...,e_{n-1}, e_{n+1} \rangle $ and $w$ is the face $\langle e_1, ...,e_{n-1} \rangle ,$ we write $ \sigma (w) = \sigma _1 + \sigma _2$ for the cone: $\langle e_1, ...,e_{n-1},e_n , e_{n+1} \rangle .$ Since $\{ e_1, ...,e_{n-1}, e_n \}$ is a ${{\mathbb{Q}}}\text{-basis}$ for the lattice $N,$ there is a relation $\sum _{i=1} ^{n+1} a_i e_i =0 .$ Let $I_1 = \{ a_i \text{ such that } a_i <0 \}.$ If $R$ is an extremal ray (i.e., $R={{\mathbb{Q}}}^+ {{\mathbb{O}}}_w$ with ${{\mathbb{O}}}_w \in NE(X/S) ,$ an extremal ray) in a fan $F,$ write $F_R$ for the fan whose walls are $\Delta _R ^{n-1} = \Delta ^{n-1} \setminus R .$ The corresponding toric variety $Y= X(F_R)$ is the contraction of $R .$ Define a simplicial subdivision $\Delta ^+$ of $\Delta _R$ by defining $${\Delta ^+ }^n = \Delta _R ^n \setminus \big(\bigcup _{w \in R} \sigma(w) \cup \bigcup _{w\in R , i\in I_1} \sigma_i(w)\big),$$ where $\sigma _i(w):=\langle e_1,...,\hat {e_i}, ..., e_{n-1}, e_n, e_{n+1} \rangle$ M. Reid in [@R1] proves: The toric morphism $\phi _1 : X^+ =X(\Delta ^+) \to Y$ corresponding to the simplicial subdivision $\Delta ^+$ of $\Delta _R$ is projective and an isomorphism in codimension  1. $-R$ is identified with an extremal ray of $X^+$ and $\phi _1 = \phi _{-R}$ is the contraction of $-R .$ Toric  2-dimensional isolated singularities ------------------------------------------- Let $\sigma $ be the cone generated by the vectors $f_2$ and $v=kf_1 - nf_2$ in the lattice $N=f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}.$ Without loss of generality, we may assume that $k$ and $n$ are coprime. Then we can choose a unique $n' \in {{\mathbb{Z}}}$ such that $0 \leq n' <k$ and $n n' \equiv 1 (mod \text{ } k) .$ Therefore, if $nn'=1-kb$ we can map $N$ isomorphically into itself and $\sigma$ to the cone $\sigma ' $ generated by $f_2$ and $v' = kf_1 - n'f_2$ by means of the matrix: $$\bordermatrix {&& \cr & n &k \cr & b &-n' } \in SL_2 ({{\mathbb{Z}}}),$$ thus inducing an isomorphism of toric varieties: $X(\sigma) \simeq X(\sigma ').$ The surface $ X(\sigma ')$ is the normalization of: $$W =\{ (x,y,z) \in {{\mathbb{C}}}^3 \text{: } x^k=yz^{k-n'} \}$$ and it is isomorphic to the quotient of ${{\mathbb{C}}}^2$ by ${{\boldsymbol{\mu}}}_k$ via the action: $$\epsilon (x,y) = (\epsilon ^{n'} x, \epsilon y)$$ for a primitive k-th root of unity $\epsilon .$ Its minimal desingularization has as exceptional divisor a chain of rational curves $E_i$ with self intersections $E_i ^2 =-a_i\leq -2$ determined by the continued fraction: $$\frac{k}{n'}=a_1 - \frac{1}{ a_2 -\frac{1}{...-\frac{1}{a_r}} }.$$ Following [@B], we call such a singularity an $A_{n',k} \text{-singularity} $ or a $\frac{1}{k} (n',1)$ singularity. Some toric  3-dimensional isolated singularities ------------------------------------------------ Let $n_1,$ $n_2$ and $n_3$ be generators of a  3-dimensional lattice $N$ and let $\sigma$ be the cone $\langle n_1 , n_2 , an_1 + bn_2 +rn_3\rangle .$ Then the affine toric variety $X(\sigma)$ is isomorphic to the quotient of ${{\mathbb{C}}}^3$ by ${{\boldsymbol{\mu}}}_r$ acting as: $$\epsilon (x,y,z) = (\epsilon ^ {a} x, \epsilon ^ {b}y, \epsilon z) ,$$ where $\epsilon \in {{\boldsymbol{\mu}}}_r$ is a primitive  r-th root of unity. Following M. Reid ([@R3]), we shall refer to such a  3-dimenional isolated singularity as a $\frac{1}{r} (a,b,1)$ singularity, thus indicating the order of the cyclic group and the weights with which it acts. Weierstrass Forms {#wform} ================== Ampleness of the log-canonical divisor -------------------------------------- Let $\psi: Y \to C$ be a flat family of generically smooth stable curves of genus $1$ over a smooth curve $C,$ with zero section $Q,$ and let $\pi : X \to C$ be the surface obtained by contracting all the components of the possible singular fibres disjoint from $Q$. Then we can express $X$ in a Weierstrass form in the following way. Since $\pi$ is proper and flat and since $H^2(X_y, {{{\mathcal O}}}_{X,y})=0$, by the theorem of base change in cohomology ([@H] theorem 12.11) $R^1 \pi _* {{{\mathcal O}}}_X$ is locally free, and since its rank is one, it is an invertible sheaf on $C$. Let ${\mathcal L} := (R^1 \pi _* {{{\mathcal O}}}_X )^{\vee}$ its inverse, ${{\mathcal E}}:= {{\mathcal O}}_C \oplus {{\mathcal L}}^2 \oplus {{\mathcal L}}^3 $ and ${\mathcal P}:={\mathbb P}({\mathcal O}_C \oplus {\mathcal L}^2 \oplus { \mathcal L}^3)$ and let: $$x: \mathcal E \to {\mathcal L}^2,$$ $$y:\mathcal E \to {\mathcal L}^3$$ and $$z:\mathcal E \to {\mathcal O}_C$$ be the canonical projections onto the given factor. We have: \[weirstrassform\] There exist two sections: $g_2 \in H^0(C,{{\mathcal L}}^4)$ and $g_3 \in H^0(C,{{\mathcal L}}^6)$ such that $X$ is isomorphic to the Cartier divisor in $\mathcal P$ given by the equation: $$y^2z=x^3 -g_2 xz^2 -g_3 z^3 .$$ Moreover: 1. $\Delta =4 g_2 ^3 - 27 g_3 ^2 \in H^0(C, {\mathcal L}^{12})$ is non-zero; 2. the sections $g_2 ^3$ and $g_3 ^2$ of ${\mathcal L}^{12}$ do not vanish to order $\geq 12$ at any point of $C;$ 3. the zero section $Q$ of $X\to C$ corresponds to the section at infinity $(x,y,z)=(0,1,0) .$ The proof is the same as the one of Theorem 1’ in [@Mu], although they only state the theorem in the case the elliptic fibration has no singular fibres. Note that the equation makes sense since all the monomials that appear in it are sections of the vector bundle ${\mathcal Sym}^3 {\mathcal E}^* \otimes {\mathcal L}^6.$ The fact that we assumed that there is no component of the singular fibres that is disjoint from $Q$ implies that our $X$ is of this form. Notice that by the Leray spectral sequence we get: $$\chi ({\mathcal O}_S) = \chi (\pi _* {\mathcal O}_S) - \chi( R^1 \pi _* {\mathcal O}_S)= \chi ({{\mathcal O}}_C) -\chi ({{\mathcal L}}^{\vee}) =d,$$ where $d=c_1(\mathcal L)$. We have the important: \[canbundleform\] Let $\psi : Y \to C$ be a relatively minimal elliptic surface, with zero section $Q$ (in particular it has no multiple fibres). Then $$\omega _Y \simeq \pi ^* (\omega _C \otimes {\mathcal L} )$$ See [@B] theorem 12.1. Two important consequences are the following: \[weierstrasscase\] For a relatively minimal non-isotrivial elliptic surface $\pi: X \to C$ with smooth base curve $C$ and zero section $Q$ the divisor: $$L_X= K_X + Q + c_1(\pi^* j^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3))$$ is semi-ample and positive on every curve except for those ($-2$)-curves of the fibres that do not meet $Q.$ More generally, for the ${{\mathbb{Q}}}$-Cartier divisor: $$L:=a Q + \pi ^* (\beta +\lambda)$$ for $a\in {{\mathbb{Q}}}_+$ a number with $0<a \geq 1$ and $\beta $ a divisor on $C$, we have: 1. if $c_1 (\beta)>0,$ then $L$ is ample; 2. 1. if $0<a<1,$ $c_1 (\beta )=0$ and $c_1 (\lambda) >0,$ then $L$ is ample; 2. if $a=1,$ $c_1 (\beta ) = 0, $ and $c_1 (\lambda) >0,$ the line bundle $L$ is semiample, and for any irreducible curve $D\subset X$: $$L\cdot D=0 \text{ if and only if } D=Q;$$ 3. if $c_1 (\beta)<0$ and $c_1 (\beta +\lambda)>0$ then $Q$ is an extremal ray. Note that this includes $K_X +a Q+\pi^*(\alpha)$ for any $\alpha \in Div(C)$ with $c_1(\alpha )+2g-2 \geq 0 ,$ if $g$ is the genus of $C.$ In what follows we will use freely that $c_ (\lambda )= -Q ^2 .$ We will first show the general part and then we will reduce to it the statement involving the canonical bundle. First one checks easely that: $$\begin{aligned} (aQ +\pi ^* (\beta + \lambda))^2 &= 2a(\beta +\lambda)\cdot Q +a^2 Q^2\\ & = 2a[c_1(\beta) + c_1(\lambda)] -a^2 c_1(\lambda) \\ &=2a \big( 2c_1(\beta) + (2-a) c_1(\lambda)\big) \geq 2a( c_1(\beta) + c_1(\lambda)) .\end{aligned}$$ Note that this is positive in all three cases, showing that $L$ is big ($L$ is clearly effective). Also, if $F$ is a fibral divisor (i.e., if it is the class of a fibre), since $X$ is relatively minimal, we always have: $$L \cdot F= a >0 .$$ If $D \subset X$ is any irreducible curve, there are two possibilities: it either maps dominantly onto $C$ or it is a fibral divisor. Thus, we are left to deal with the former case. So we may assume that $D$ is an irreducible curve dominating $C .$ If $D \neq Q,$ since $Q$ is effective, we have: $D \cdot Q \geq 0 ,$ and then: $$L\cdot D \geq c_1 (\beta +\lambda) + aD\cdot Q \geq c_1 (\beta +\lambda)>0$$ in all the cases. Therefore, all the cases will be differentiated according to the behavior of $L$ on $Q.$ We have: $$L \cdot Q = a Q^2 + c_1 (\lambda )+ c_1 (\beta )= (1-a) c_1 (\lambda ) + c_1 (\beta )$$ [**Case (1) and Case (2) (i):**]{} In this case we have that: $L \cdot Q >0$ and we can conclude by means of the Nakai-Mosheizon criterion. [**Case (2) (ii):**]{} in this case $L\cdot D \geq c_1 (\lambda )>0 $ for any irreducible (non-fibral) curve $D \neq Q$ and $L \cdot Q = 0$ [**Case (3):**]{} in this case $Q$ is the only curve on which $L$ is negative. This entails that $Q$ generates an extremal ray. In fact, observe that if $c_1 + c_2 \in [Q] {{\mathbb{R}}}_+ \subset {\overline {NE}}(X)$ then $L \cdot (c_1 + c_2) <0.$ But if $c_1$ and $c_2$ are limits of sequences $C_{1i}$ and $C_{2i}$ of curves that differ from $Q,$ then $L \cdot C_{ji}>0$ for each $i$ and $j$, and thus $L\cdot c_j \geq 0$ for each $j,$ which is a contradiction, since it would imply that $L \cdot (c_1 + c_2) \geq 0.$ So at least one of the $c_i$ is in $[Q] {{\mathbb{R}}}_+ ,$ and again, since $Q$ is the only irreducible curve on which $L$ is negative, both $c_1$ and $c_2$ must be in $[Q] {{\mathbb{R}}}_+ .$ Now, for the statement about $L= K_X + Q + c_1(\pi^* j^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3))$, note that the by the canonical bundle formula (theorem \[canbundleform\]) we have that:$ K_X =\pi ^* ( K_C + \lambda) ,$ where $\lambda$ is a divisor class associated to $ \mathcal L .$ This concludes the proof. The following corollary is a special case of the previous one, but it is worth stating separately, for heuristic reasons. \[semiampl\] Same hypoteses on $\pi: X \to C$ and $C.$ If furthermore the curve $C$ is not rational, then the divisor: $$L_X= K_X + Q$$ is semi-ample and positive on every curve, except for those ($-2$)-curves of the fibres that do not meet $Q.$ What this means is that, if there are such $-2$-curves in the fibres, in order to make the log-canonical divisor ample we have to contract them. This explains the necessity of considering [*Weierstrass forms*]{} for the purpose of studying our moduli problems. Log-canonical singularities of Weierstrass equations {#LCWE} ---------------------------------------------------- In this section, given a Weierstrass equation $y^2= x^3 + a (t) x + b(t)$ with $a(t), b(t) \in k[[t]]$ we will only write the low degree terms of $a(t), b(t).$ After having dealt with the ampleness of the log-canonical divisor, the second question one needs to address in order to understand when an elliptic surface is stable is what kind of Weierstrass equations give rise to a log-canonical singularity. The following lemma answers this question: \[localcan\] Let $\pi:S \to {\operatorname{Spec}}k[[t]]$ be an elliptic surface with zero-section $Q$, given in Weierstrass form: $ y^2= x^3 + a t^ n x + b t^m,$ with $a,b$ units in $k[[t]]$. Assume furthermore that $j\neq \infty .$ Then the pair $(S,Q)$ is log-canonical if and only if $min( 3n, 2m) \leq 12 .$ Furthermore, if $F$ is a smooth fibre of $\pi$, the pair $(S,F+Q)$ is log-terminal if and only if S is, and if $F$ is a cuspidal fibre, $(S,Q+F)$ is never log-canonical. The proof of this lemma makes use of the list of Alexeev’s of dual graphs of log-canonical surface singularities (cf. [@K] ch.3) in the sufficient direction, and by calculating the discrepancies in the necessary direction. If $min (3n,2m) <12$ then the minimal resolution has as dual graph the Kodaira graphs: $I_0$, $I_1$, $I_n$, $II$, $III$, $IV$, $I_0 ^*$, $I_n ^*$, $II^*$, $III^*$, $IV^*$( see [@S], Ch. $IV$ pg.$354$). We will refer to table $(IV.3.1)$ pag.$41$ of [@M2] for the singularities corresponding to the various Kodaira types. In case $I_0,$ and $II$ the surface is smooth, and in case $I_n ^*$ it has a rational double point singularity, so there is nothing to prove. In case $III$ there is only one exceptional curve in the dual graph, and so we are in case $(1)$ of Alexeev’s list . In the case of Kodaira type $IV$, then graph of the resolution is of type $A_2$ and for the Kodaira type $I_n$ the dual graph is of type $A_n$ and again this is in case $(1)$ of Alexeev’s list. In the $*$ cases things become more interesting. $I_0 ^*$ corresponds to a $D_4$ diagram, and this is in case $(2)$ again with $(\Delta _1, \Delta _2, \Delta _3)=(2,2,2)$; $I_n ^*$ corresponds to a $D_{n+4}$ graph and this is still in case $(2)$ with $(\Delta _1, \Delta _2, \Delta _3)=(2,2,N)$; $IV^*$ to an $E_6$ and again this is in $(2)$ with $(\Delta _1, \Delta _2, \Delta _3)=(2,3,3)$; $III^*$ corresponds to an $E_7$ and we are again in case $(2)$ with $(\Delta _1, \Delta _2, \Delta _3)=(2,3,4)$; $II^*$ corresponds to $E_8$ and one gets case $(2)$ with $(\Delta _1, \Delta _2, \Delta _3)=(2,2,3)$. So these are all even canonical (they are all Du Val singularities). Now, if $min(3n,2m) \geq 12$, then one can write $(n,m)=(4k+n', 6k+m')$ with $n',m' \geq 0$ such that either $0 \leq n' \leq 3$ or $0 \leq m' \leq 5$. One can then consider the rational map: $ S' \to S$ given by $(x,y,t)= (x' t^{2k},y' t^{3k},t) $, where $S' \text{: } {y'} ^2 = {x'} ^3 + a t^{n'} x + b t^{m'}$. Now: $$\omega = \frac { dx \wedge dt} {2y} \in \Omega_{k(S)} ^2$$ is a basis for the $k(S)$-module $\Omega_{k(S)} ^2 $, and so is $$\omega ' = \frac { dx' \wedge dt} {2y'} \in \Omega_{k(S')} ^2$$ for the $k(S')$-module $\Omega_{k(S')} ^2 $, moreover: $$\pi ^* \omega = \frac { dx' t^{2k} \wedge dt} {2y' t^{3k}} = \frac { dx' \wedge dt} {2y' t^{k}}= \frac {\omega '} {t^{k}} \in \Omega_{k(S')} ^2 (kE) .$$ The surface $S'$ is canonical for what has just been showed above since $min (3n',2m') < 12.$ Therefore $S$ is not log-canonical if $k>1$, since the discrepancies would be at least $k$. One would be tempted to say that the $k=1$ case is settled and thereby claiming the log-canonicity in this case, but indeed the previous argument fails in this case, since the map $S' \to S$ exhibited above is not proper; there might still be exceptional divisors on a completion of $S' \to S$ with discrepancies $>1.$ Therefore a detailed analysis of the case $k=1$ is needed. In order to do that one has to find a partial resolution and compute discrepancies. To achieve this goal, since $y^2= x^3 + a t^{4} x + b t^{6} ,$ with $a$ and $b$ units in $k[t],$ can be thought of as the double cover of the affine plane ramified along the curve $R \text{: }x^3 + a t^{4} x + b t^{6}=0$, we can simply take an embedded resolution $D\subset M$ of this curve, blow-up again at the points in which components of $D$ with odd multiplicity meet to obtain $D' \subset M'$, and then take the double covering of the resulting surface along the total transform of $R$. This double covering has canonical singularities. The exceptional curve $E$ of a resolution of $y^2= x^3 + a t^4 x + b t^ 6 $ is a stable curve of arithemetic genus $1$ of self-intersection $-1$ attached to a rational curve of self-intersection $-1 .$ Hence the singularity is log-canonical (but not log-terminal). For the isotrivial $j= \infty $ case we have: \[infinity\] The local equation $y^2 = x^2 (x - \lambda t^k )$ with $\lambda$ a unit in $k[[t]],$ is semi-log-canonical if and only if $k\leq 2 .$ The proof is very similar to the one of lemma \[localcan\]. Let us denote by $X$ the surface defined by $y^2 = x^2 (x - \lambda t^k ).$ Let us show first that if $k \leq 2,$ then the singularity is semi-log-canonical. If $k=0$ then the equation is locally around the $(0,0,t)$ isomorphic to $u^2=w^2, $ and it is thus semi-smooth, hence semi-log-canonical. For $k=1,$ then seeting $z=x-\lambda t$ shows that this singularity is isomorphic to $y^2 = x^2 z$ which is again semi-smooth. For $k=2,$ after one blow-up we obtain a surface $X'$ and as exceptional divisor a nodal curve of the same type as the general fibre of $y^2 = x^2 (x-\lambda t^2).$ It is now easy to see that $X'$ is semismooth and that $X$ is semi-log-canonical. If $k >2 ,$ then we can conclude as in lemma \[localcan\] that $X$ is not semi-log-canonical. \[locstand\] We call [**standard Weierstrass equation**]{} an equation $ y^2= x^3 + a (t) x + b (t),$ which satisfies the condition $min( 3 n, 2m) \leq 12 ,$ where $a \text{, } b \in k[[t]]$ and $n$ and $m$ are respectively the order of vanishing of $a$ and $b$ at $t=0,$ or one of the form $y^2 = x^2 (x - \lambda t^k )$ with $k\leq 2 .$ So, we can directly infer from lemma \[localcan\] and \[infinity\] the important: A generically stable elliptic surface $X \to C$ mapping to a smooth curve $C$ is semi-log-canonical if and only if its local equation around each cusp is a standard Weierstrass equation. Types of cuspidal fibres {#types} ------------------------- In this section we will introduce a bit of terminology and notation we will be using in the discussion of the special cases in section \[special cases\]. Let $y^2 = x^3 + a x + b,$ with $a \text{, } b \in k[[t]] ,$ be a [*standard Weierstrass equation*]{} with $j \neq \infty$ (see below for the case $j=\infty$). Let us denote with $\nu _0 (f) $ the order of vanishing at $t=0$ of a power series $f \in k[[t]].$ In the following table we set $N := min (3 \nu _0 (a) , 2\nu _0 (b))$ and the first row will give a condition on $N,$ the second will give the Euler characteristic of the corresponding Kodaira fibre (see [@M2] table (IV.3.1) page 41) and the third will have a symbol we associate to the corresponding singularity. $N=$ $0$ $6$ $2$ $10$ $3$ $9$ $4$ $8$ $12$ -------- ----- ------- ------ --------- ------- --------- ------ -------- ------ -- $\chi$ $n$ $n+6$ $2$ $10$ $3$ $9$ $4$ $8$ type $I$ $I^*$ $II$ $II ^*$ $III$ $III^*$ $IV$ $IV^*$ $L$ here $n$ is related to the order of vanishing of $\Delta$ (see ramark below). Our notation differs from Kodaira’s in that we identify all the $I_n$ and $I_n ^*$ to two categories, namely: $I$ and $I^*.$ In Kodaira’s notation the former are characterized by the order of vanishing of the discriminant $\Delta = 4 a^3 + 27 b^2$ (in the $I_n$ case $\nu _0 (\Delta)=n$ and in the $I_n^*$ case $\nu _0 (\Delta)=n+6$). The reason for our notation stems from the fact that in our analysis we need not distinguish among them. Also the $L$ case does not appear in the classical litterature, because it is not a rational double point singularity. It is indeed elliptic, and hence it has moduli (for instance the $j$ invariant of the exceptional curve) as oppossed to the classical cases that do not. Note also that if a Weierstrass equation is isotrivial with $j=\infty$ then it is of the form $y^2 = x^3 -3 \lambda ^2 (t) x + 2 \lambda ^3(t) $ and it can be transformed to ${y'}^2 = {x'}^2 (x' + 2 \lambda (t))$ (and viceversa). This shows that the last equation is minimal if and only if $\nu _o (\lambda (t)) <2 ,$ which could also be argued by directly computing the discriminant. Abramovich-Vistoli’s fibred surfaces and prestable reduction ============================================================= Abramovich-Vistoli fibred surfaces {#abr} ---------------------------------- D. Abramovich and A. Vistoli in [@A-V] define [*famillies of fibered surfaces*]{} in order to compactify the moduli space of [*fibered surfaces*]{}, that is to say surfaces $X \to C$ mapping flatly and properly to a curve with stable fibers and with a number of sections $\sigma _1, ..., \sigma _n$ (or equivalently, to compactify the moduli space of Kontsevich stable curves into the Deligne-Mumford stack of stable $n$-pointed curves with fikxed genus). We will need to borrow the following definitions and results from their work: Let $\Gamma$ be a finite group acting on a family of Delgne-Mumford stable curves $Y\to V,$ over some scheme $V.$ Abramovich and Vistoli (in [@A-V]) give the following: This action is [**essential**]{} if each $\gamma \in Stab(v),$ for some geometric point $v\in V,$ acts nontrivially on the fibre $Y_s$ over $s.$ (see [**Def. 4.1**]{} in [@A-V]) Let $C\to S$ be a flat (not necessarily proper) family of nodal curves, $X\to C$ a proper morphism with one dimensional fibers, and $\sigma_1,\ldots,\sigma_\nu\colon C\to X$ sections of $\rho$. We will say that $X \to C \to S $ is a [**family of generically fibered surfaces**]{} if $X$ is flat over $S$, and the restriction of $\rho$ to $C_{sm}$ is a flat family of stable pointed curves. If $S$ is the spectrum of a field we will refer to $X\to C$ as a generically fibered surface. (see [**Def. 4.3**]{} in [@A-V]) A triple $(U,Y\to V \to S, \Gamma)$ is called a [**chart**]{} for a family of [*generically fibred surface*]{} $X\to C \to S$ if there is a diagram: $$\begin{array}{ccccc} Y&\to& X\times_{C} U & \to & X \\ {\downarrow}& & {\downarrow}& &{\downarrow}\\ V &\to& U & \to & C \\ {\downarrow}& & {\downarrow}& &{\downarrow}\\ S &=& S & = &S \end{array}$$ together with a group action $\Gamma\subset Aut_S(Y\to V)$ satisfying: 1. The morphism $U\to C$ is étale; 2. $V\to S$ is a flat (but not necessarily proper) family of nodal curves; 3. $\rho\colon Y\to V$ is a flat family of stable $\nu$-pointed curves of genus $\gamma$, 4. the action of $\Gamma$ on $\rho$ is essential; 5. we have isomorphisms of $S$-schemes $V/\Gamma\simeq U$ and $Y/\Gamma\simeq U\times_C X$ compatible with the projections $Y/\Gamma \to V/\Gamma$ and $U\times_C X\to U$, such that the sections $U\to U\times_C X$ induced by the $\sigma_i$ correspond to the sections $V/\Gamma\to Y/\Gamma$. The fibre above $p$ is called a [**twisted fibre**]{}. For our purposes, we do not need the chart $(U,Y\to V, \Gamma)$ to be [*minimal*]{}, i.e., we will not need it to satisfy property  4. Let $(U,Y\to V \to S, \Gamma)$ be a chart for $X\to C \to S,$ then: Let $\Gamma ' =Stab(v)$ be the stabilizer at the nodal point $v$ of a gemetric fibre $V _t$ of $V \to S$ and let $T_1$ and $T_2$ be the tangent spaces of each branch at the node. Then: 1. $\Gamma '$ is cyclic and it sends each branch of $V _t$ to itself; 2. the generator $\gamma$ of $\Gamma '$ acts on $T_1 $ and $T_2$ by multiplication with a primitive root of unity (of the order of $\Gamma '$). See [@A-V] Proposition 4.5. In the same situation, Abramovich and Vistoli (in [@A-V]) set the following: A chart $(U,Y\to V, \Gamma)$ is called balanced if for any nodal point of any gemetric fiber of $V,$ the action of the two roots of unity describing the action of a generator of the stabilizer on the tangent spaces $T_1$ and $T_2$ of the branches are inverse to each other. Let $X\to C\to S$ be a family of generically fibered surfaces, $\alpha _1 =(U_1,Y_1\to V_1,\Gamma_1)$, $\alpha _2 =(U_2,Y_2\to V_2,\Gamma_2)$ two charts; call $ p_i\colon V_1\times_C V_2\to V_i$ the $i^{\rm th}$ projection. Consider the scheme $$I = { \mathop{{\operatorname{Isom }}}\limits_{V_1\times_C V_2}}(p_1^*Y_1,p_2^*Y_2)$$ over $V_1\times_C V_2$ representing the functor of isomorphisms of the two families $p_1^*Y_1$ and $p_2^*Y_2.$ Abramovich and Vistoli call the [**transition scheme** ]{} from $\alpha _1$ to $\alpha _2$ the sheme theoretic closure of the section of $I$ over the inverse image $\widetilde V$ of $C_{sm}$ in $V_1\times_C V_2$ corresponding to the isomorphism $p_1^*Y_1 \mid _{\widetilde V}\simeq p_2^*Y_2\mid _{\widetilde V}$ Two charts $(U_1,Y_1\to V_1,\Gamma_1)$ and $(U_2,Y_2\to V_2,\Gamma_2)$ are [compatible]{} if their transition scheme $R$ is étale over $V_1$ and $ V_2$. A [**family of fibered surfaces**]{} $${{\mathcal X}}\to {{\mathcal C}}\to S$$ is a family of generically fibered surfaces $X\to C\to S$ such that $C\to S$ is proper, together with a collection $\{(U_\alpha, Y_\alpha\to V_\alpha,\Gamma_\alpha)\}$ of mutually compatible charts, such that the images of the $U_\alpha$ cover $C$. Such a collection of charts is called an [*atlas*]{}. A family of fibered surfaces is called [*balanced*]{} if each chart in its atlas is balanced. The family of generically fibered surfaces $X\to C\to S$ supporting the family of fibered surfaces ${{\mathcal X}}\to {{\mathcal C}}\to S$ will be called a family of [*coarse fibered surfaces*]{}. We can now state the theorem of theirs that we are going to be using here: \[fibredsurfaces\] Let ${{\mathcal X}}_\eta \stackrel{\pi}{\to} {{\mathcal C}}_\eta\to \eta$ be a [*balanced stable*]{} fibered surface, with induced map $f_\eta: {{\mathcal C}}_\eta \to {{\overline {\mathbf{M}}}}_{g,n},$ with sections ${\sigma _i }_{\eta}\subset {{\mathcal X}}_\eta$ and with sections ${D_i}_\eta \subset {{\mathcal C}}_\eta .$ Then there is a finite extension of discrete valuation rings $R\subset R_1$ and an extension $$\begin{array}{ccc} {{\mathcal X}}_\eta\times_\Delta \Delta_1 &\subset & {{\mathcal X}}_1 \\ {\downarrow}& & {\downarrow}\\ {{\mathcal C}}_\eta\times_\Delta \Delta_1 &\subset & {{\mathcal C}}_1 \\ {\downarrow}& & {\downarrow}\\ \{\eta_1\} & \subset & \Delta_1, \end{array}$$ with $\Delta ={\operatorname{Spec}}R$ and $\Delta _1 = {\operatorname{Spec}}R_1 ,$ such that: 1. ${{\mathcal X}}_1\to {{\mathcal C}}_1\to \Delta_1$ is a balanced stable family of fibered surfaces with sections $\sigma _i ;$ 2. there is a regular map $f_1 : {{\mathcal C}}_1 \to {{\overline {\mathbf{M}}}}_{g,n}$ extending $f_\eta \circ p_1,$ where $p_1 :{{\mathcal C}}_\eta\times_\Delta \Delta_1 \to {{\mathcal C}}_\eta$ is the natural projection; 3. $f_1$ is Kontsevich stable 4. $\omega _{{{\mathcal X}}_1 / {{\mathcal C}}_1} (\sum \sigma _i+ \sum \pi_1 ^* D_i) \otimes f_1 ^* {{\mathcal A}}$ is ample, for some ample line bundle ${{\mathcal A}}$ on ${{\overline {\mathbf{M}}}}_{g,n} .$ The extension is unique up to a unique isomorphism, and its formation commutes with further finite extensions of discrete valuation rings. See [@A-V] pg.20. prop. 2.1. and pg.28 prop. 8.13 The extension lemma and Prestable reduction {#extlemma} -------------------------------------------- Suppose one is given an elliptic surface ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta} \to \eta$ over the generic point $\eta$ of a DVR scheme $\Delta,$ and with induced $j-$map $j _{\eta}: {{\mathcal C}}_{\eta} \to {{\overline {\mathbf{M}}}}_{1,1} .$ If ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta}$ has no cuspidal fibres, then we can apply the theorem of Abramovich and Vistoli (cf. Theorem \[fibredsurfaces\] above), to extend ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta}$ and $j _{\eta}$ over the whole $\Delta .$ In case there are cuspidal fibres (and our surface is in Weierstrass form), the strategy we will adopt is to temporarely replace these with twisted fibres and mark the corresponding points on the base curve ${{\mathcal C}}_{\eta}.$ Once we do that, we want to be able to go back (at least generically) to Weiestrass forms. To this aim we prove the following: \[Extension Lemma\] Let $S$ be the spectrum of a two-dimensional complete regular local ring, $S= {\operatorname{Spec}}k[[s,t]],$ let $W= S \setminus p,$ with $p=V(s,t)$ the closed point, set $U= S \setminus V(t),$ $\eta$ the generic point of ${\operatorname{Spec}}k[[s]]$ and let $\pi :X _W \to W$ be a family of curves of genus $1$ with zero section, such that: 1. there is a map $j: S \to { {{\overline {\mathbf{M}}}}}_{1,1},$ the $j \text{-invariant};$ 2. the family $X_W \mid _U ,$ is a family of stable elliptic curves; 3. $X_W \mid _{W_\eta}$ is an elliptic surface in minimal Weierstrass form. then $ X _W$ extends over the whole $S,$ to a family $X\to S$ whose fibre over $s=0$ is an elliptic surface in minimal Weierstrass form. Let us consider the line bundle ${{{\mathcal L}}}'$ on $W$ whose dual is: $R^1 \pi _* {{{\mathcal O}}}_W$. Since $S$ is non-singular, this line bundle extends to a unique line bundle ${{\mathcal L}}$ over the whole $S$. By the Theorem of the base change in cohomology, ${{{\mathcal L}}}^* \otimes k (\eta) \simeq (R^1 \pi _* {{{\mathcal O}}}_{X}) \mid _{W_\eta}$, and thus the sections $g_2$ and $g_3$ of $H^0(W_\eta, ({{\mathcal L}}\otimes k (\eta))^4)$ and of $H^0(W_\eta, ({{\mathcal L}}\otimes k (\eta))^6)$ respectively, extend to two sections $a \text{ and } b$ of $H^0(S, {{\mathcal L}}^4)$ and of $H^0(S, {{\mathcal L}}^6)$ respectively. So we can consider: $$X \text{: } \left\{ y^2 = x^3 + a x + b \right\} \subset {{\mathbb{P}}}( {{{\mathcal O}}}_S \oplus {{\mathcal L}}^2 \oplus {{\mathcal L}}^3).$$ Note that since $a^3$ and $b^2$ are sections of the sixth power of the Hodge bundle $ {{\mathcal L}}^6$, $a$ and $b$ are defined up to the transformation: $(a,b) \to (\lambda ^4 a, \lambda ^6 b)$, where $\lambda \in {{\mathcal O}}_S ^* .$ \[Claim 1\] Let $V:=V(t)$, $A:= V(a)$ and $B:=V(b)$, and let $A= \bigcup A_i \cup h V$ and $B= \bigcup B_i \cup kV$ be respectively the decompositions in irreducible components of $A$ and $B$; then $V$ (if $h$ and $k$ are non-zero) is the only irreducible component that $A \text{ and } B$ can share. In fact, if they had another irreducible component in common, say $H,$ then , since in $ k[[s,t]] $ every prime ideal of height one is principal, there would be an $h \in k[[s,t]]$ such that $H=V(h)$ and so: $$X: y^2 = x^3 + h^{n} a' x + h^{m}b' ,$$ where $a' {\text{ and }} b'$ are non zero on $H.$ The assumption that the curves in the family are stable away from $V=V(t)$, is equivalent to the fact either one of $a' \text{ and } b'$ is nonzero on H and that $2n$ and $3m$ are divisible by $6$. In fact, if that were not the case, by pulling back our family $X$ via the map: $$k[[s,t]] \to k[[s,t]] \otimes \kappa (h)$$ where $\kappa (h)$ is the residue field at the prime $(h)$, we would produce a family of curves that is not stable and that cannot be reduced to a stable one by changing $a$ and $b$ via the transformation $(a,b) \to ( g^{-2k} a, g^{-3k} b)$ for some $g \in k[[s,t]]$ and $k \in {{\mathbb{Z}}}$. Therefore $2n$ and $3m$ are divisible by $6$ and we can apply the above mentioned transformation with $g=h,$ therefore $H$ does not exist. Furtermore we may assume, according to the [*purity lemma*]{} of Abramovich-Vistoli (see [**Lemma 2.1**]{}, pg.3 in [@A-V]), that the components $A_i$ do not intersect the $B_i$ away from $V.$ Indeed every point of $U$ satisfies the hypotheses of the purity lemma. If there were a point $p \in U$ such that $p \in A_i \cap B_i ,$ since $j_U : U \to {{\overline {\mathbf{M}}}}_{1,1}$ is well defined and extends the map $j_{U \setminus \{p\}} : U \setminus \{p\} \to {{\overline {\mathbf{M}}}}_{1,1}$ defined by the family over $U \setminus \{p\},$ according to the purity lemma there would exist a stable family ${{\mathcal X}}_U \to U$ extending ${{\mathcal X}}_{U \setminus \{p\}} \to {U \setminus \{p\}} ,$ and thus $p$ could not be in $A_i \cap B_i ,$ So, next step is to show: $A_i \cap B_i \cap V= \emptyset .$ Let us assume that some $A_i$ does meet some $B_j$ along $V .$ That means that there is some point $q \in V$ such that $a$ and $b$ are both zero at $q$. We can write $a'=t^n a''$ and $b'=t^m b''$ where $a'' \text{, } b'' \in k[[s,t]] \setminus (t)$ do not vanish identically along $V$, and the non negative integers $n \text{ and } m$ can be zero if, respectively $A$ and $B$ don’t contain $V$. Since we are assuming that some $A_i=V(a_i)$ does meet some $B_j=V(b_j)$ along $V$, there must be some point $q \in V$ such that $a''$ and $b''$ are both zero at $q$. The $j \text{-map}$ of the statement is now a map: $$j: {\operatorname{Spec}}k[[s,t]] \to {{{\overline {\mathbf{M}}}}}_{1,1}$$ that at this point has the form: $$j(s,t) = \frac {1728 a^3 t^{3n}}{4 a^3 t^{3n} + 27 b^2 t^{2m}}.$$ This restricts to a map: $$j_{G}: G \to {{{\overline {\mathbf{M}}}}}_{1,1}$$ of the same form for any divisor $G$ of ${\operatorname{Spec}}k[[s,t]]$. In particular we can restrict $j$ to $A_i$ and $B_j$, to get two morphisms: $$j_{A_i }: A_i \to {{{\overline {\mathbf{M}}}}}_{1,1}$$ and $$j_{B_j}: B_j \to {{{\overline {\mathbf{M}}}}}_{1,1}$$ Since these two maps are induced via restriction by $j$, they have to coincide at $p$, But from this we infer a contradiction, since $j_{A_i}\equiv 0$ and $j_{B_j} \equiv 1720 .$ Therefore they cannot meet along $V$, because by hypothesis $j: {\operatorname{Spec}}[[s,t]] \to {{{\overline {\mathbf{M}}}}}_{1,1}$ is well defined. Let $M \subset {{\mathbb{P}}}^r $ be a projective scheme. We call [**Kontsevich prestable**]{} a flat family of maps over a scheme $S$ $f: {{\mathcal C}}\to M $ if ${{\mathcal C}}\to S$ is a flat family of nodal curves. and We call [**quasiminimal**]{} an elliptic surface $X \to C$ such that $X \mid _{C_{sm}} \to C_{sm}$ is [*in minimal Weiestrass form*]{}, where $C_{sm} $ is the smooth locus of $C $ and such that, for each point $p \in C _{sing} $ there is a chart $(U,Y\to V, \Gamma)$ with $X \times _C V \simeq Y/\Gamma $ centered at $p.$ The first step towards proving the stable reduction theorem is to show the following theorem \[fict\] Let $\Delta $ be the spectrum of a DVR, $\eta \in \Delta$ the generic point and $({{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta} , {{\mathcal Q}}_{\eta} , f_{\eta} : {{\mathcal C}}_{\eta} \to {{\overline {\mathbf{M}}}}_{1,1}),$ be a triple consisting of a relatively minimal elliptic surface ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta}$ with section ${{\mathcal Q}}_{\eta}$ and Kontsevich-stable map to moduli $f_{\eta}.$ Then we can find, after possibly a finite base change $\Delta ' \to \Delta ,$ a map of $\Delta ' \text{-schemes}$ ${{\mathcal X}}' \to {{\mathcal C}}'$ such that: 1. the fibre over the special point $0 \in \Delta ' ,$ $ {{\mathcal X}}_0 \to {{\mathcal C}}_0$ is the semi-log-canonical union of [*relatively quasiminimal* ]{}elliptic surfaces with section ${{\mathcal Q}}_0;$ 2. the double curves of ${{\mathcal X}}_0$ are either stable or twisted fibres; 3. $f_{0} : {{\mathcal C}}_{0} \to {{\overline {\mathbf{M}}}}_{1,1}$ is [*Kontsevich prestable*]{}. We are given $({{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta} \to {\eta} ,f_{\eta} :{{\mathcal C}}_{\eta} \to {{\overline {\mathbf{M}}}}_{1,1}) .$ As mensioned in the introduction to this section, we can mark the points of ${{\mathcal C}}_{\eta}$ corresponding to cuspidal fibres of ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta}:$ let $\Sigma _{i \eta} $ be such divisor on ${{\mathcal C}}_{\eta}.$ We can then use the theorem of Abramovich-Vistoli (cf.[**Theorem**]{} \[fibredsurfaces\]), to get a triple $({{\mathcal X}}' \to {{\mathcal C}}' \to \Delta ', {{\mathcal Q}}' ,f': {{\mathcal C}}' \to {{\overline {\mathbf{M}}}}_{1,1} )$ consisting of a family of fibred surfaces ${{\mathcal X}}' \to {{\mathcal C}}' \to \Delta ',$ (with sections $s_i:\Delta ' \to {{\mathcal C}}'$ extending $\Sigma _{i \eta}$) a ${{\mathbb{Q}}}-\text{Cartier}$ divisor ${{\mathcal Q}}'$ and a regular map $f': {{\mathcal C}}' \to {{\overline {\mathbf{M}}}}_{1,1} $ such that condition  2 and the condition (stronger than  3 above) that $f' _0 $ together with the sections $s_j$ be Kontsevich stable hold. This family coincides with our family of elliptic curves in minimal Weierstrass form ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta} \to \eta $ over ${{\mathcal C}}' _{\eta} \setminus \bigcup _i {\Sigma _i} _\eta .$ Let $\Sigma _i $ = $\overline {{\Sigma _i} _{\eta}},$ the closure in ${{\mathcal C}}'$ of ${\Sigma _i} _\eta ,$ and $\Sigma = \bigcup \Sigma _i .$ Hence, we can apply lemma \[Extension Lemma\] to remove the twisted fibres lying above $\Sigma$ and to replace them with the cuspidal curves induced by the original ones lying above $\Sigma _{\eta},$ so that the generic fiber ${{\mathcal X}}' _{\eta '}$ is now isomorphic to the original surface ${{\mathcal X}}_{\eta} \to {\eta}.$ According to lemma \[Extension Lemma\] ${{\mathcal X}}' _0$ consists of [*quasiminimal*]{} elliptic surfaces. The map $f' _0$ may now be Kontsevich unstable (since we have removed the sections $s_i$), but it is clearly [ *prestable*]{}. This concludes the proof. We shall refer to theorem \[fict\] as the [**Prestable Reduction Theorem**]{}. Lemma \[Extension Lemma\] allowes us to remove the extra sections we added along the cuspidal fibres to use the stable reduction theorem of [@A-V] (see section $4$). But this comes with a price: when we do so, the family of maps $j: {{\mathcal C}}\to {{{\overline {\mathbf{M}}}}}_{1,1}$ may no longer be stable. In fact, if there is a rational component $C$ of the central fibre ${{\mathcal C}}_0$ on which $j$ is constant, and which meets the rest of the central fibre in only one point ( e.g., if the surface $X={{\mathcal X}}_0 \mid _C\to C$ contains two cuspidal curves or more and meets the rest of ${{\mathcal X}}_0$ transversally along one fibre), then $j$ is no longer stable. Indeed we will see in proposition \[extr ray\] that the components of ${{\mathcal X}}$ that map onto such curves are not stable in the sense of **Definition \[logstabsurf\].** The log-canonical divisor, Extremal rays and semiampleness {#iso} ----------------------------------------------------------- As mentioned in the previous subsection, we have to deal with components of the central fibre ${{\mathcal X}}_0$ that lie above a rational component of ${{\mathcal C}}_0$ contracted by the $j$-map. These are then [*isotrivial quasiminimal*]{} elliptic surfaces, attached to at least one component in two possible ways: either along a smooth or a twisted fibre. By the very definition, for such a surface the $j$-map is constant, therefore the log-canonical divisor (for triples) does not have the contribution coming from $j^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (1) .$ Hence the log-canoncial divisor for triples equals $ K _{{{\mathcal X}}} + {{\mathcal Q}}$ on such components. The zero section of such a component is always a log-flipping extremal ray, if this component is attached to the rest of the central fibre only along one fibre, and will be contracted by the log-canonical bundle if the isotrivial component is attached along two fibres. Indeed we have, even more generally, that this happens, even if the component $X \to C$ is [*not isotrivial*]{}, in case the log-canonical divior is $L= K _{{{\mathcal X}}} + {{\mathcal Q}}$ (the one for pairs). This last divisor coincides with the log-canonical divisor for triples on the isotrivial components, as we remarked above. This will turn out to be useful when dealing with pairs $(X,Q)$ only. \[extr ray\] Let ${\mathcal X}_0 = X \cup X'$ be a decomposition of the central fibre where $\pi:X \to C $ is a [*quasiminimal*]{} elliptic surface, where $C \simeq {{\mathbb{P}}}^1$ a smooth rational curve. Then: 1. If $X$ is attached to $X'$ only along one fibre $G$, then: $$L \cdot Q =-1$$ where $Q$ is the zero-section and $L= (K_{{{\mathcal X}}} + {{\mathcal Q}}) \mid _X=G+Q+K_X$ is the log-canonical divisor; 2. If $X$ is attached to $X'$ along two fibres $G_1 \text{ and } G_2$, then: $$L \cdot Q =0$$ where $Q$ is the zero-section and $L= (K_{{{\mathcal X}}} + {{\mathcal Q}}) \mid _ X=G_1 + G_2+Q+K_X$ is the log-canonical divisor. Let us assume first that the attaching fibres are stable. In this case, the zero section $Q$ goes entirely through the smooth locus of the morphism $\pi : {{\mathcal X}}\to {{\mathcal C}}$. Therefore, if we denote by $K_{X/C}$ the relative canonical divisor: $$K_{X/C} \cdot Q + Q^2 =c_1(\omega _Q \otimes \pi ^*\omega _C)=0$$ moreover the attaching fibres are reduced and therefore linearly equivalent to the general fibre $F$. Hence, since $\pi^* K_C= -2F$: $$L \cdot Q = (K_{X/C}+Q +G_1 + \pi ^* K_C)\cdot Q=(-2F +F)\cdot Q = -1$$ in the case of one attaching fibre, and: $$L \cdot Q = (K_{X/C}+Q +G_1+G_2 + \pi ^* K_C)\cdot Q=(-2F +F+F)\cdot Q = 0$$ in the case of two stable attaching fibres. If the attaching fibres are not stable, then there is a diagram: $$\begin{array}{ccc} Y &\stackrel{f} {\rightarrow} & X \\ {\downarrow}& &{\downarrow}\\ C' &\stackrel{\phi} {\rightarrow} &C \end{array}$$ such that $\phi :C' \to C$ is a finite morphism with Galois group $\Gamma$, $f: Y \to X $ is also finite and $Y \to C'$ is a relatively minimal elliptic surface. Moreover $\phi :C' \to C$ is ramified at two points, and: 1. $\Gamma \simeq {{\boldsymbol{\mu}}}_k$ if there is only one twisted attaching fibre with monodromy of order $h$, in this case the branch points are $0$ (where $(X_0)_{red} =G_1$) and infinity; 2. $\Gamma \simeq {{\boldsymbol{\mu}}}_k $ with $k = l.c.m.(k_1,k_2 )$ when there are two twisted attaching fibres of monodromy of orders $k_1$ and $k_2 ,$ respectively. Let $Q'=f^*Q$ and $F'=f^*G .$ Then, in case (1), by the projection formula we have: $$L \cdot Q = \frac {1}{k} f^* L \cdot Q' = \frac {1}{k} [(K_{X'} - (k-1)F' + F' + Q') \cdot Q']=-1$$ since $f^* G_1 = F'$ and by Riemann-Hurwitz: $$f^* K_X = K_{X'} - (k-1)F'.$$ In case (2), the Riemann-Hurwitz formula now reads: $$f^* K_X = K_{X'} - (k_1-1)F'-(k_2-1)F'$$ and ones again, by means of the projection formula, we can conclude: $$L \cdot Q = \frac {1}{k} f^* L \cdot Q' = \frac {1}{k} [(K_{X'} - (k_1 -1)F' +(k_2 -1) F'+k_1 F'+k_2 F' + Q') \cdot Q']=0.$$ It will turn out to be useful to generalize these computations for any [*quasiminal elliptic surface*]{} $X \to C$ with no limitation on the number of twisted fibres and the genus of the base curve. Let ${{\mathcal X}}_0 = X \cup X_R$ a decomposition of the central fibre such that $X$ meets $X_R$ in $r$ attaching fibres (stable and twisted) $\bigcup _{i=1} ^r G_i.$ Let $Q$ be the zero-section of $\pi :X\to C$ (where $C$ is the base curve of $X$). Let $k_i$ be order of the monodromy group around $G_i ,$ and $g$ the arithmetic genus of $C.$ We have: \[genlogcan\] With these hypotheses and notations, one has: $$L_X \cdot Q = 2g-2 + r .$$ The proof is very similar to the proof of the previous lemma. Define $\overline r $ to be the integer that equals $r$ if $r \equiv 0 \text{(mod 2)}$ and equals $r+1$ if $r \equiv 1 \text{(mod2)}.$ Let $p_i := \pi (G_i),$ $p_{r+1}\in C$ some extra point, and let $I$ to be the set that equals $\{ p_1, p_2 , ..., p_ {\overline r} \}.$ Now let $g:C' \to C$ be the finite ramified covering with Galois group ${{\boldsymbol{\mu}}}_k, $ with $k=lcm(k_i)$ totally ramified at the $p_i 's ,$ with $i \in I.$ Because of the underlying structure of quasiminimal elliptic surface,there is an atlas $(U_i,Y_i\to V_i,\Gamma _i \simeq {{\boldsymbol{\mu}}}_{k_i})$ such that $Y_i \to V_i$ is stable. In particular, the normalization of the pull-back $X' \to C'$ of $X \to C$ to $C'$ will be a family of stable curves, and therefore, if we denote by $Q'$ the pull-back of the zero section $Q \subset X,$ $Q$ must be entirely contained in the smooth locus of $X'.$ Let $f: X' \to X$ be the natural morphism. If we indicate by $g'$ the genus of $C',$ it must then be the case that $(K_{X'} +Q')\cdot Q' =2g'-2.$ But by Riemann-Hurwitz applied to $C' \to C$ we obtain: $$2g'-2 = k(2g-2) + {\overline r} (k-1)$$ The Riemann-Hurwitz formula applied to $f$ reads: $$K_{X'} = f^* K_X + \sum _{i=1} ^ {\overline r} (\frac{k}{k_i}-1) G_i '$$ where $G_i ' =(f^* G_i)_{red} .$ Therefore, if we let $Q' = f^* Q :$ $$L_X \cdot Q = \frac {1}{k} \big( K_{X'} - \sum _{i=1} ^ {\overline r} (\frac{k}{k_i} -1) G_i ' + \sum _{i=1} ^r \frac{k}{k_i} G_i' + Q' \big) \cdot Q' ,$$ where $ G_{\overline r}= \varnothing $ if $r \equiv 0 \text{ ( mod 2)} $ and $ G_{\overline r}$ is the fiber over the point $p_{r+1}$ otherwise. Thus, if we let $\epsilon (N)$ be the function over the integers that equals $0$ if $N \equiv 0 \text{ ( mod 2)} $ and $1$ otherwise we have: $$L_X \cdot Q =\frac {1}{k} (k(2g-2) + {\overline r} (k-1) -\epsilon (r)k +{\overline r})= 2g-2 + r.$$ \[self int\] Note that a similar computation can be carried out to measure the failure of $X$ to satisfy adjunction, i.e., to compute the [ *different*]{} (see [@K], page 175 for a definition). Another point worth observing is that if $X$ has only one attaching fibre, in $X \mid _U$: $$Q ^2 = \frac{1}{k} f^* Q \cdot Q' = \frac{1}{k} {Q'} ^2$$ so that this is the contribution a twisted fibre with monodromy $k$ gives to the self intersection of $Q .$ \[canint\] Note that we have showed, en passant, that: $$f^* L_{X}=L_{X'} + (1-k \epsilon (r)) G_{\overline r} .$$ \[twistedampleness\] Same hypotheses as in Proposition \[genlogcan\], then: 1. if $2g-2+r > 0,$ then $L_X$ is ample; 2. if $2g-2+r = 0,$ and $X\to C$ is non-isotrivial, then $L_X $ is semiample and for any irreducible curve $D$: $$L_X \cdot D =0 \text { if and only if } D=Q ;$$ 3. if $2g-2+r <0$ then $Q$ is an extremal ray. In fact, more generally the same holds for the ${{\mathbb{Q}}}$-Cartier divisor: $$M := K_X + \sum G_i + a Q$$ with $ 0<a<1$ any rational number. Recall from Remark \[canint\] that: $$f^* L_{X}=L_{X'} + (1-k \epsilon (r)) G '_{\overline r},$$ where $f:X' \to X$ is as constructed in the proof of Proposition \[genlogcan\]. Therefore: $$L_X ^2 = L_{X'}^2 + 2(1-k \epsilon (r)) L_{X'} \cdot G '_{\overline r}$$ since ${G '_{\overline r}} ^2=0.$ The surface $X'$ is now in minimal Weierstrass form, and therefore we can apply the [*canonical bundle formula*]{} (cf. theorem \[canbundleform\]), and write: $$f^* L_{X}= ({\pi '}^* (K_{C'} + \lambda '+ \sum _{i=1} ^r p_i ' +(1-k \epsilon (r)) p_ {\overline r} ')+Q',$$ where $p_i' \in C'$ denotes the point whose fibre via $\pi '$ is $G_i .$ Note that $c_1(K_{C'} + \sum _{i=1} ^r p_i ' +(1-k \epsilon (r)) p_ {\overline r} '=k (2g-2+r).$ We can then conclude (1), (2) and (3) by means of corollary \[weierstrasscase\]. We can now conclude also the more general statement about $M$ appealing to the same corollary, and by observing that if one defines the ${{\mathbb{Q}}}$-Cartier divisor on $X'$: $$M' := K_{X'} + \sum G_i ' + a Q'$$ one has that: $$f^* M= M' + (1- k \epsilon(r)) G_{\overline r} = ({\pi '}^* (K_{C'} + \lambda '+ \sum _{i=1} ^r p_i ' +(1-k \epsilon (r)) p_ {\overline r} ')+a Q' .$$ The special cases and standard elliptic surfaces {#special cases} ================================================ Special cases {#specas} ------------- Let ${{\mathcal Y}}\to {{\mathcal C}}\to S$ be an elliptic threefold in Weierstrass form $${{\mathcal Y}}= V(y^2 - x^3 - ax -b) \subset {{\mathbb{P}}}({{\mathcal O}}_{{{\mathcal C}}} \oplus {{\mathcal L}}^2 \oplus {{\mathcal L}}^3),$$ where ${{\mathcal L}}= \omega _{{{\mathcal Y}}\mid {{\mathcal C}}} .$ For any fibre ${{\mathcal C}}_t$ of the base ${{\mathcal C}}$ one can define an integer valued function on ${{\mathcal C}}_t$ as: $$N_{{{\mathcal C}}_t} (q)= min (\nu _q (a ^3 \mid _{{{\mathcal C}}_t}), \nu _q (b^2 \mid _{{{\mathcal C}}_t}) ),$$ where $\nu _q(g)$ is the order of vanishing of any section $g\in {{\mathcal O}}_{{{\mathcal C}}_t}$ at $q \in {{\mathcal C}}_t.$ In particular we will simply write $n(q)$ for this function defined for the central fibre ${{\mathcal C}}_0.$ If ${{\mathcal C}}$ contains a negative rational curve $E$ (not necessarely irreducible), one can contract it to obtain a morphism $\rho :{{\mathcal C}}\to {{\mathcal C}}' .$ On ${{\mathcal C}}'$ one can then construct an elliptic threefold ${{\mathcal Y}}' \to {{\mathcal C}}'$ in the following manner. Let $p'$ be the image of $E$ and $W'= {{\mathcal C}}' \setminus p' .$ Then, one naturally has a Weierstrass equation on $W'$ induced by ${{\mathcal Y}}\mid _{W'}.$ In fact $\rho$ induces an isomorphism $\rho _W : W \to W'$ where $W= {{\mathcal C}}\setminus E ,$ and, letting ${{\mathcal L}}'_{W'} = \rho _* \omega _{{{\mathcal Y}}_W / W'}$ and $a'_{W'}$ and $b'_{W'}$ respectively the push-forward of the section $a \mid _ W$ of ${{\mathcal L}}^2 \mid _W$ and $b\mid _W$ of ${{\mathcal L}}^3\mid _W,$ one can then define an ellitpic threefold in Weierstrass form: $${{{\mathcal Y}}}' = V(y^2 - x^3 - a'x -b') \subset {{\mathbb{P}}}({{\mathcal O}}_{{{\mathcal C}}'} \oplus {{{\mathcal L}}'} ^2 \oplus {{{\mathcal L}}'}^3).$$ Since ${{\mathcal C}}'$ is smooth at $p'$ one can then extend the line bundle ${{\mathcal L}}' _{W'}$ uniquely to a line bundle ${{\mathcal L}}'$ on ${{\mathcal C}}',$ and correspondingly uniquely extend the sections $a' _{W'}$ and $b' _{W'}$ to global sections $a'$ and $b'$ of ${{{\mathcal L}}' }^2$ and ${{{\mathcal L}}'}^3$ respectively. To put it differentely, we take the [*saturation* ]{} ${{\mathcal L}}' = (({ \rho _* {{\mathcal L}}}) ^{\vee})^{\vee}$ and the corresponding sections $a'$ and $b'$ to construct the Weiestrass equation presented above. In general, given an elliptic threefold ${{\mathcal X}}\to {{\mathcal C}}$ (not necessarely in Weierstrass form) with no multiple fibres, but possibly twisted fibres, and a contraction of a negative curve of the base $\rho : {{\mathcal C}}\to {{\mathcal C}}',$ one can define an elliptic threefold ${{\mathcal X}}'$ over ${{\mathcal C}}'$ as the elliptic threefold ${{\mathcal X}}'$ extension of a local Weierstrass model around the image point. That is to say we choose a Zariski neighborhood $Z$ of $E$ in ${{\mathcal C}}$ such that the only possible twisted fibres are above the nodes of $E.$ Let $W$ be a Zariski neighborhood of $E$ in $Z$ such that ${{\mathcal X}}_W$ has a Weierstrass representations away from the nodes of $E.$ $\rho :{{\mathcal C}}' \to {{\mathcal C}}$ induces a map $\rho : W \to W'$ to a Zariski neighborhood of $p' \in {{\mathcal C}}' .$ Now apply the construction above to $W'$ and $W.$ This construction may lead to a threefold which does not have log-canonical singularities, even if ${{\mathcal X}}$ had log-canonical singularities to begin with. Let ${{\mathcal X}}\to {{\mathcal C}}\to \Delta$ be a  1-parameter family of elliptic surfaces with section ${{\mathcal Q}},$ with general fibre ${{\mathcal X}}_{\eta}$ in minimal Weierstrass form. Assume furthermore that there is a component $X_1 \to C_1$ of the central fibre ${{\mathcal X}}_0$ that maps to a rational nodal not necessarely irreducible curve $C_1 \subset {{\mathcal C}},$ that $X_1$ is attached (transversally) to another component $X_2\to C_2$ of ${{\mathcal X}}_0$ along only one fibre $G$ (twisted or stable). One can then contract $C_1$ in ${{\mathcal C}}$ to a point $p\in {{\mathcal C}}'$ and therefore emulate the construction of ${{\mathcal X}}'$ above. The question one must ask oneself is: when does it happen that the surface $X_2 ' ,$ image of $X_2$ via the contraction of $X_1$ has log-canonical singularities? Proposition \[specialcas\] will answer just that. We call the cases when this occur, the [**special cases**]{}. In this chapter we will use (mostly implicitely) the following well known result: \[nonsing\] The contraction of $C_1$ in ${{\mathcal C}}$ produces a surface ${{\mathcal C}}'$ that is smooth in a Zariski neighborhood of the image point. The Idea -------- The idea of the construction of the log-canonical model, is based on the possibility (cf. section \[abr\]) of taking an étale neighborhood $V$ of $ p\in {{\mathcal C}}$ such that the pull-back of ${{\mathcal X}}$ to $V$ has a nodal fibre at $p,$ i.e., of finding a chart (not necessarely minimal) $(Y\to V, U, \Gamma )$ at least of a Zariski neighborhood $W$ of $p \in {{\mathcal C}}.$ Therefore $Y \to V$ has a Weiestrass model and we can naively contract the rational components of the central fibre of ${{\mathcal C}}\to S$ and write “explicit” equetions for the analityc singularity we thus obtain. Standard elliptic surfaces -------------------------- Let us assume that $C_1$ is irreducible and at set $p =C_1 \cap C_2 .$ Let moreover ${{\mathcal C}}_R= \overline {{{\mathcal C}}_0 \setminus C_1} .$ It might happen that the fibre of ${{\mathcal X}}$ over $p$ is [*twisted*]{}. But we can find a chart $(V,Y \to U, \Gamma )$ such that the fibre over the point lying above $p$ is stable. On $U$ we can define two divisors $A$ and $B$ as follows. Let $W$ and $W'$ as in section \[specas\], so that ${{\mathcal X}}\mid _ {W' \setminus p}$ is in Weierstrass form ${y^2= x^3 + ax +b} \subset {{\mathbb{P}}}({{\mathcal O}}_{W' \setminus p} \oplus {{\mathcal L}}_{W' \setminus p}^2 \oplus {{\mathcal L}}_{W' \setminus p})$ Since ${{\mathcal C}}$ is normal at $p,$ the two sections $a$ and $b$ extend to sections all over $W.$ Set $A= V(a^3)$ and $B=V(b^2).$ \[nonvan\] We have: $p \notin A \cap B.$ Let $U$ be a Zariski neighborhood of $p$ such that the only twisted fibre of ${{\mathcal X}}\mid _U$ is at $p$ and that there are no cuspidal fibres. Let $(V,Y \to V', \Gamma )$ be a chart of ${{\mathcal X}}_U \to U$ as in section \[abr\]. Then the thesis holds, because if both $A$ and $B$ passed through $p,$ the family $Y \to V'$ could not consist of stable curves. Let $c: {{\mathcal X}}\dashrightarrow {{\mathcal X}}' $ be the contraction constructed in the previous section and $ X_2 ' =c_* X_2 .$ Let $C_1$ be an irreducible $(-1)$-curve in ${{\mathcal C}}$ meeting the rest of ${{\mathcal C}}$ transversally in only one point $p.$ Let $C_2$ be the irreducible component of ${{\mathcal C}}$ meeting $C_1 .$ At this point we ask ourselves what happens to the self intersection of the zero section $Q_2$ of the component $X_2$ to which $X_1$ was attached. The answer is given by the following, where we let $Q_2 '$ be the zero-section of $X_2 ' \to C_2 ':$ \[self-int\] The self intersection ${Q_2 ' }^2$ of $Q_2 '$ in $X_2 '$ is: $${Q_2 ' }^2 = Q_2 ^2 +{Q_1 }^2$$ where $Q_2 ^2$ is taken in $X_2$ and ${Q_1 }^2$ is taken in $X_1 .$ Let us consider how the number $Q_2 ^2 + Q_1 ^2$ changes after the contraction of $X_1.$ Note that it is the same as $(Q_2+ Q_1)\cdot {{\mathcal Q}},$ where ${{\mathcal Q}}$ is the divisor of ${{\mathcal X}}$ swept out by the zero sections of the fibres of ${{\mathcal C}}\to S.$ On the other hand, $Q_1 + Q_2 \equiv {{\mathcal X}}_0 \cdot {{\mathcal Q}}- {{\mathcal X}}_R \cdot {{\mathcal Q}}$ where ${{\mathcal X}}_R = \overline {{{\mathcal X}}_0 -(X_1 +X_2)}$ is the rest of the central fibre ${{\mathcal X}}_0 ,$ and $\equiv$ denotes numerical equivalence. Hence $$Q_2 ^2 + Q_1 ^2 =({{\mathcal X}}_0 \cdot {{\mathcal Q}}- {{\mathcal X}}_R \cdot {{\mathcal Q}})\cdot {{\mathcal Q}}.$$ This number does not change after performing the contraction, since this operation does not touch ${{\mathcal X}}_R$ and ${{\mathcal X}}_t {{\mathcal Q}}$ is constant as $t$ varies among the geometric points of $\Delta$ because of the flatness of ${{\mathcal X}}\to \Delta .$ So $$(Q_2')^2 = Q_2^2 + Q_1^2$$ Our goal is to relate the invariant $\sum _{q \in {{\mathcal C}}\setminus \{p\}} N_C(q)$ with the self intersection of the zero section, therefore suggesting tha the latter should be treated as an [*invariant of the geometry near the zero section*]{}. This is taken care by: \[twelve\] Let $X \to {{\mathbb{P}}}^1$ an elliptic surface whose singular fibres are all standard Weierstrass equations, except for possibly having at most one twisted fibre over $0 \in {{\mathbb{P}}}^1$ whose monodromy group is isomorphic to ${{\boldsymbol{\mu}}}_h .$ Assume also that away from $0\in {{\mathbb{P}}}^1,$ the surface $X$ is in Weierstrass form (not necessarely minimal). Then $12 Q^2 = -[\sum _{q \in {{\mathbb{P}}}^1 \setminus {0}} n_C(q) + deg(j)]$ and in particular $\sum _{q \in {{\mathbb{P}}}^1 \setminus {0}} n_C(q) + deg(j) \leq 12$ if and only if $Q_1 ^2 \geq -1 .$ Let us assume first that $X\to {{\mathbb{P}}}^1$ is not isotrivial. Noether’s formula reads: $$\chi ({{\mathcal O}}_X) = \frac{ K_X ^2 + \chi (X)}{12}$$ and therefore, since $K_X ^2 =0$ and $\chi ({{\mathcal O}}_X) = c_1 ({{\mathcal L}})$ (see section \[weirstrassform\]) we obtain: $$\chi(X) = 12 c_1 ({{\mathcal L}})= -12 Q^2 .$$ But since $X\to {{\mathbb{P}}}^1$ is locally trivial in the euclidean (resp. étale) topology away from the cuspidal fibres, the (étale) Euler characteristic $\chi (X)$ must equal: $$\chi (X) = \chi (X_{\eta}) \chi (f^{-1} (C \setminus \delta) ) + \sum _{p\in {{\mathbb{P}}}^1} \chi (X _p),$$ where $X_{\eta}$ is the generic fibre and $\delta =\{ p \in {{\mathbb{P}}}^1 : X_p \text{ is singular } \} .$ Hence, since $\chi ( X_{\eta})=0, $ $$Q^2 =-\frac{1}{12} \sum_{p \in \delta} \chi (X_p) .$$ Performing a case by case analysis of the table in section \[types\] on sees that in all cases but for $I_n$ and $I_n ^*$ one has that $\chi (X_p) = N_p$ and in the latter cases $\chi (X_p)= N_p + n .$ Therefore, since by Corollary $IV$.4.2 of [@M2] one has that $deg(j) = \sum _{n\geq 1} n(i_n + i_n ^*)$ where $i_n$ and $i_n ^*$ indicate respectively the number of fibres of type $I_n$ and $I_n ^* ,$ we conclude that $$Q^2 =-\frac{1}{12} [\sum _{p \in \delta} N_p + deg(j)].$$ If there is a twisted fibre $G = (X \otimes k(p))_{red}$ (for some point $p \in {{\mathbb{P}}}^1$), there is a ramified morphism of order $h$ (the monodromy of $G$) of the base $s:C\simeq {{\mathbb{P}}}^1 \to {{\mathbb{P}}}^1,$ totally ramified at $0$ and at infinity, and a diagram: $$\begin{array}{ccc} X' & \to &X \\ \downarrow & &\downarrow \\ C & \stackrel{s}{\to} &{{\mathbb{P}}}^1 \end{array}$$ such that $X' \to C$ is stable. Thus we conclude by means of the same argument as above and the observations that $Q^2 =\frac { {Q' }^2}{h}$ (see remark \[self int\]) and that if $q' \in C$ is a point mapping to $q$ via $s,$ $N_{q'} = \frac {N_q}{h}.$ If $X\to {{\mathbb{P}}}^1$ is isotrivial, there is nothing to prove if $j\neq \infty,$ since in that case there are no $I$ or $I ^*$ fibres. \[specialcas\] The singularity of $X_2 '$ in a neighborhood of the fibre over the point $p'$ to which $C_1$ gets contracted is of type $y^2 = x^3 + a' x +b'$ with $min( \nu _{p'} ({a' \mid _{{C_2'}}}^3), \nu _{p'} ({b' \mid _{{C_2'}}}^2) ) = \sum _{q\in {{\mathcal C}}\setminus I} N_{C_1} (q)+deg(j \mid _{C_1}) .$ In particular $X_2 '$ is log-canonical if and only if $ \sum _{q\in {{\mathcal C}}\setminus I } N_{C_1} (q) +deg(j \mid _{C_1}) \leq 12 .$ We can construct ${{\mathcal X}}'$ as in section \[specas\]. Let $C_2 ' = \rho _* C_2$ and $p= C_1 \cap C_2 .$ According to proposition \[twelve\] we have that: $$12 {Q_2 '}^2 = -[\sum _{q \in C_2 '} N_{C_2 '}(q) + deg(j _{C_2 '})]$$ We remark that the fibre over $p'$ will no longer be twisted. Note also that even though $j'$ might only be a rational map on the whole ${{\mathcal C}}^c ,$ (the surface resulting from contracting $C_1$) it extends to a regular map when restricted to any smooth curve $B \subset {{\mathcal C}}^c .$ and: $$\begin{array}{cc} 12 (Q_1 ^2 +{Q_2 }^2) &= -[\sum _{q \in C_2 \setminus \{ p \}} N_{C_2 }(q) + deg(j_{C_2})] -[\sum _{q \in C_1 \setminus \{ p \}} N_{C_1 }(q) + deg(j_{C_1})]\\ &\\ &\\ \end{array}$$ On the other end, according to proposition \[self-int\] one has: $12 (Q_1 ^2 +{Q_2 }^2) =12 {Q_2 '}^2.$ Hence: $$\begin{array}{cc} \sum _{q \in C_2 '} N_{C_2 '}(q) &= \sum _{q \in C_2 \setminus \{ p \}} N_{C_2 }(q) +\sum _{q \in C_1 \setminus \{ p \}} N_{C_1 }(q) + deg(j_{C_1})]\\ &\\ &\\ \end{array}$$ since away from $p \in C_2$ we have not changed $X_2 \to C_2,$ and thus that $deg(j_{C_2})= deg(j _{C_2 '}).$ The proposition now follows from the observation that, since we have not changed $X_2 \to C_2$ away from $p,$ it must be that $N_{C_2 '}(q)= N_{C_2 }(q)$ for every $q\neq p .$ What this proposition says is that one can contract all these components to begin with, before proceding with the stable reduction. It is thus convenient to introduce the following generalization of the concept of minimal Weierstrass form: \[standard\] Let $X \to C$ a be an elliptic surface mapping to a nodal irreducible curve $C.$ Such a surface will be named [**standard elliptic surface**]{} if the local equation around each fibre above $C_{sm}$ is a [*standard Weierstrass equation*]{} (see definition \[locstand\]) and if, for each point $p \in C _{sing} $ there is a chart $(U,Y\to V, \Gamma)$ centered at $p$ with $X \times _C U \simeq Y/\Gamma .$ We have: \[divcont\] Same hypotheses as in theorem \[fict\]; then there is a finite base change of DVR schemes $\Delta ' \to \Delta$ and a $\Delta '$-family of elliptic surfaces $({{\mathcal X}}' \to {{\mathcal C}}', {{\mathcal Q}}' )$ with section ${{\mathcal Q}}' ,$ such that: 1. if $\eta '$ is the generic point of $\Delta ' ,$ ${{\mathcal X}}' _{\eta '} \simeq {{\mathcal X}}_{\eta}$; 2. the central fibre ${{\mathcal X}}' _0 $ is composed of [*standard elliptic surfaces*]{}; 3. if $X \to {{\mathbb{P}}}^1$ is a component of ${{\mathcal X}}' _0$ attached along only one fibre (twisted or stable), then ${{\mathcal Q}}\mid _X ^2 < -1$ in $X.$ 4. if we ask condition  3 only of isotrivial components, then there is a well defined regular map $j' : {{\mathcal C}}' \to {{\overline {\mathbf{M}}}}_{1,1} $ and if we set $f'= \pi ' \circ j',$ then $\omega _{{{\mathcal X}}' / \Delta '} (Q) \otimes {f'}^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3)$ is ample away from those isotrivial components of the central fibre ${{\mathcal X}}' _0$ that meet transversally the rest of ${{\mathcal X}}_0 '$ along one or two fibres of ${{\mathcal X}}_0 ' \to {{\mathcal C}}_0 '.$ According to theorem \[fict\] there exists a finite morphisms of DVR schemes $\overline {\Delta} \to \Delta $ and a triple $(\overline {{\mathcal X}}\to \overline {{\mathcal C}},\overline {{\mathcal Q}}, \overline f : \overline {{\mathcal X}}\to {{\overline {\mathbf{M}}}}_{1,1} )$ satisfying conditions  1,  2 and $\overline {{\mathcal X}}_0 \to \overline {{\mathcal C}}_0 $ is a union of [*realively quasiminimal*]{} ellitpic surfaces. Let $\overline B$ be a connected component of $\overline {{\mathcal C}}_0$ consisting of a tree of rational curves (we call such a component $\overline B$ a [**tree-like component**]{}), meeting the rest of $\overline {{\mathcal C}}_0$ trasversally in one point, and such that in $\overline X_ {\overline B} = \overline {{\mathcal X}}\mid _{\overline B} \to \overline B$ we have ${\overline {{\mathcal Q}}} _{\overline B} ^2 \geq -1.$ The proof will be by induction on the number $N$ of such components. Let us asume first that $N=1.$ Let $\overline C_1,$ be a “leaf”(i.e. an irreducible component furthest away from ${{\mathcal C}}_0 \setminus B$) of $\overline B$ Let $\rho :\overline {{\mathcal C}}\to \overline {{\mathcal C}}'$ be the contraction of $ \overline C_1.$ If $ \overline X_1 \to \overline C_1$ is isotrivial, the map $ \overline j : \overline {{\mathcal C}}\to {{\overline {\mathbf{M}}}}_{1,1}$ descends to a map $ \overline j ' : \overline {{\mathcal C}}'\to {{\overline {\mathbf{M}}}}_{1,1}.$ Let $\overline C_2 $ be the component of ${{\mathcal C}}_0$ meeting $\overline C_1,$ let $\overline C_2 '$ be its image in $\overline {{\mathcal C}}'$ via $\rho$ and $p' = \rho ( \overline C_1) .$ We can apply proposition \[specialcas\] to $\overline C_1$ and find a threefold $ \overline {{\mathcal X}}' \to \overline {{\mathcal C}}'$ that satisfies  1 and  2 (and with a regular map $ \overline j ' : \overline {{\mathcal C}}'\to {{\overline {\mathbf{M}}}}_{1,1}$ if $C_1$ was $j$-trivial) and such that the local equation of the fibre above $p'$ is in [*standard Weiestrass form*]{}. The image of the curve $\overline C_2 $ is again a leaf itself, because there are no other tree-like components. According to Lemma \[twelve\], if $\overline {{\mathcal Q}}'$ denotes the divisorial push-forward of $\overline {{\mathcal Q}}:$ $${\overline {{\mathcal Q}}' \mid _{\overline X_2 '}} ^2 ={\overline {{\mathcal Q}}\mid _{\overline X_2 }} ^2 + {\overline {{\mathcal Q}}\mid _{\overline X_1 }} ^2 .$$ If this number happens to be $\geq -1 ,$ then we can appy the procedure to $C_2 ' ,$ which is now attached to the rest of ${\overline {{\mathcal C}}}' _0$ only at one point, since we have contracted $\overline C_1 .$ We can inductively iterate this procedure untill we get to a component for which the self intersection of the zero-section is less than $-1 .$ Let us now do the general case: we assume we know the result for any number $k$ of tree-like components with $k<N .$ Let $\overline B$ be a tree-like component which has the property that it is attached to all the other tree-like components at one end only (such a component must exist, even though it may only consist of one edge). Then we can apply the previous argument to this tree-like component, and “prune” all its edges. Now the number of tree-like components is $N-1 ,$ and we can then conclude by induction. If we perform these operations only for tree-like components $\overline B$ for which $\overline X \to \overline B$ is isotrivial, then we get the desired triple $({{\mathcal X}}' \to {{\mathcal C}}', {{\mathcal Q}}', f:{{\mathcal X}}'\stackrel{\pi '}{\to }{{\mathcal C}}' \stackrel{j'}{\to } {{\overline {\mathbf{M}}}}_{1,1})$ on $\Delta ' = \overline \Delta .$ In this case, the claim about the ampleness of $\omega _{{{\mathcal X}}' / \Delta '} (Q) \otimes f^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3)$ away from those isotrivial components of the central ${{\mathcal X}}' _0$ that meet transversally the rest of ${{\mathcal X}}_0 '$ in one or two fibres, is a consequence of theorem \[fibredsurfaces\]. In fact we have only perfomed birational operations to some surfaces with constant $j$-invariant (namely, those surfaces $X$ for which ${{\mathcal Q}}\mid _ X ^2 \geq -1$) which met the central fibre ${\overline {{\mathcal X}}}_0$ along just one fibre, and so ${{\mathcal C}}' \to {{\overline {\mathbf{M}}}}_{1,1}$ is still Kontsevich prestable, i.e., $\omega _{{{\mathcal C}}'/ \Delta '} \otimes {j'}^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3)$ is semiample and ample away from those components $C \subset {{\mathcal C}}' _0 $ that meet the rest of ${{\mathcal C}}' _0 $ in one or two points and that are $j-$trivial; and since $\omega _{{{\mathcal X}}'/{{\mathcal C}}'}(Q)$ is relatively ample, we are done. A [*pair*]{} $(X \stackrel{\pi}{\to} C, Q)$ will be called [**strictly prestable**]{} if: 1. $X \to C$ is a [*standard elliptic surface*]{} and if for each rational component $B$ of $C$ that meets the rest of $C$ in only one point; 2. ${{{\mathcal Q}}\mid _ {X \mid _B}} ^2 < -1 .$ Similarly, a [*triple* ]{} $(X \stackrel{\pi}{\to} C, Q, f: X \stackrel{\pi}\to C \stackrel {j}{\to} {{\overline {\mathbf{M}}}}_{1,1})$ will be called [**stricly prestable**]{} if condition $(1)$ above hold and if $(2)$ holds only for those components $B$ for which $X \mid _ B \to B$ is [*isotrivial*]{}. Therefore, theorem \[divcont\] says that one can perform the [**strictly prestable reduction**]{} of a family of minimal elliptic surfaces, possinly after a finite base change. \[sum\] From the formula $ Q^2 =-\frac{1}{12} \sum_{p \in \delta} \chi (X_p) $ (see proposition \[twelve\]) and from table IV.3.1 of [@M2], we infer the following table for the contribution of a Kodaira fibre to $Q^2 :$ fibre type $I$ $I^*$ $II$ $II ^*$ $III$ $III^*$ $IV$ $IV^*$ L ---------------- ----- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---- -- cont. to $Q^2$ $0$ $-\frac{1}{2}$ $-\frac{1}{6}$ $-\frac{5}{6}$ $-\frac{1}{4}$ $-\frac{3}{4}$ $-\frac{1}{3}$ $-\frac{2}{3}$ -1 Note also that the contribution to the self intersection is exactly $\frac{-N}{12}$ (see the table in section \[types\]). The Toric Picture ================== One attaching fibre ------------------- In this section we will show that we can perform the necessary log-flips and log-canonical contractions torically. The first step towards understanding the toric picture in the case of one attaching fibre, is to understand what it looks like on the base curve ${{\mathcal C}},$ or equivalently on the zero-section ${{\mathcal Q}}.$ Let $R$ be a discrete valuation ring (DVR), $\Delta = {\operatorname{Spec}}(R) ,$ $\eta$ its generic point and $0$ its special point. For a toric variety $Z$ with torus $T$ we write $D_Z$ for the complement of the torus $D_Z = Z \setminus T .$ Let $ {{\mathcal C}}\to \Delta$ be a family of nodal curves ${{\mathcal C}}.$ Assume that the central fibre $ {{\mathcal C}}_0 $ has a rational component $C_1$ meeting the rest of ${{\mathcal C}}_0$ transversally in only one point $p , $ and assume that the singularity at $p$ is an $A_{k-1}\text{-singularity}. $ Let $C_2$ be the rest of ${{\mathcal C}}_0$, $S$ a divisor of ${{\mathcal C}}$ meeting $C_1$ transversally in only one point and ${{\mathcal I}}\subset {{\mathcal O}}_{{{\mathcal C}}}$ the ideal sheaf of $C_1 \cup C_2 \cup S .$ \[torbasonefib\] Let ${{\mathcal C}}\to \Delta$ as before. Then there is a Zariski open neighborood $U$ of $C_1$ in ${{\mathcal C}},$ a  2-dimensional toric variety $Z$ and an étale map $t:U \to Z$ such that: 1. the fan of $Z$ is $F = \langle f_1, f_1 + k f_2\rangle \cup \langle f_1 + k f_2 ,f_2 \rangle $ in the lattice $N = f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}};$ 2. the pull-back via $t$ of the ideal of $D_Z $ is the ideal ${{\mathcal I}}(U). $ Let $U \subset {{\mathcal C}}$ be a Zariski open neighborhood of $C_1$ such that $U \cap (C_1 \cap S) =p , $ and such that $S$ is not entirely contained in $U .$ We can now contract $C_1$ to a smooth point $q.$ Let $c : {{\mathcal C}}\to {{\mathcal C}}''$ be such contraction. Let $C_2 '' = c' _* C_2$ and $S'' = c'_* S$ and let ${{\mathcal I}}''$ be the ideal sheaf of $C_2'' \cup S'' .$ Therefore, we can find an étale neighborhood $U''$ of $q\in {{\mathcal C}}''$ and a map $t'' : U'' \to {{\mathbb{A}}}_{k} ^2$ to the toric variety ${{\mathbb{A}}}_{k}^2 .$ We shrink $U$ and $U''$ if necessary so that $c^{-1}(U'') =U .$ Let $Z'$ be the toric variety whose fan $F'$ is the union of the cones: $\sigma _1 =\langle f_1, f_1 + f_2 \rangle ,$ $\sigma _2 =\langle f_1 + f_2 , f_1 +2 f_2\rangle ,$ ....$\sigma _k = \langle f_1 + (k-1) f_2 , f_1 + k f_2\rangle , \sigma _{k+1}=\langle f_1 + k f_2 , f_2 \rangle $ in the lattice $N= f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}};$ and let $U'$ be normalization of $Z'$ in the function field $k(U)$ of $U.$ Thus, by definition we have morphisms $b: U' \to U $ and $t':U' \to Z'$ such that ${t'}^* {{\mathcal O}}(D_{Z'}) \simeq {{\mathcal O}}(C_1' + C_2 ' +S')$ where $C_i ' = b^* C_i$ and $S' =b^* S.$ The surface $U'$ is in fact the minimal resolution of $U.$ The morphism $t'$ induces a rational map $t :U \dashrightarrow Z .$ Let $W$ be its graph, then we have a commutative diagram: $$\begin{array}{cccccc} U'&&\stackrel{t'}{\longrightarrow} &&Z' \\ &\searrow \alpha&&&\\ b\downarrow &&W&&\downarrow \beta\\ & p_1 \swarrow &&\searrow p_2&&\\ U &&\stackrel{t}{\dashrightarrow} && Z \end{array}$$ Where $\alpha$ is the morphism whose existence is ensured by the minimality of $U' .$ If we show that $p_1$ and $p_2$ are finite, then $t$ is in fact a regular morphism and by construction it satisfies the properties of the thesis. But $\beta \circ t' = p_2 \circ \alpha ,$ and since ${t'} ^* F = E,$ if $E$ is the exceptional divisor of $b$ and $F$ the one of $\beta, $ we have that $p_2$ is finite ($t'$ is étale). Similarly one concludes that $p_1$ is finite and therefore $t$ is a morphism. Our goal is to look at a [*strictly prestable*]{} family ${{\mathcal X}}\to {{\mathcal C}}$ of elliptic surfaces, and in particular the base curve ${{\mathcal C}}$ is either a family of [*Kontsevich prestable*]{} curves (in the case of moduli of triples) or a surface obtained from a Kontsevich prestable family by contracting  [-1]{}-curves. So it may happen that ${{\mathcal C}}$ is singular, but the singularities are of type $A_{k-1}.$ In the particular context of lemma \[torbasonefib\], the total space of the family of base curves ${{\mathcal C}}$ might have a singularity of type $A_{k-1}$ at $p,$ in which case we want to be able to find a finite “toric” morphism from a smooth surface. We have: \[toricbasext\] In the hypothesis of lemma \[torbasonefib\], then there exist a ramified cyclic covering $f:V \to U$ of order $k$ ramified at $p$ and along some section $S$ and an étale map to a  2-dimensional toric variety $t' : V \to Z'$ such that: 1. the fan of $Z'$ is $F'=\langle f_1 ' , f_1 ' + f_2 '\rangle \cup \langle f_1 ' + f_2 ', f_2 '\rangle ;$ 2. the pull-back via $t'$ of the ideal of the toric divisor $D_{Z'} $ is the ideal of $C_1 ' \cup C_2 ' \cup S'$ where $C_2 ' = f^* C_2$ and $S' = f^* S ;$ 3. $f$ induces a toric morphism $f : Z' \to Z $ given by $f_1 '=f_1$ and $f_2 ' = k f_2$ Choose a divisor $S$ that meets $C_1$ transversally in only one point, by shrinking $U$ if necessary, we can apply lemma \[torbasonefib\] and find a toric variety $Z$ and an étale morphism $t:U\to Z$ such that $t^*{{\mathcal O}}_Z ({{\mathbb{O}}}_{f_1} + {{\mathbb{O}}}_{f_1 + kf_2} + {{\mathbb{O}}}_{f_2}) \simeq {{\mathcal O}}_U (C_1 + C_2 + S) .$ Since $t^* {{\mathcal O}}_Z (f_2) \simeq {{\mathcal O}}_U ( S)$ the cyclic covering corresponds to taking the sub-lattice $N' =f_1 ' {{\mathbb{Z}}}\oplus f_2 ' {{\mathbb{Z}}}$ of $N=f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}$ with $f_1'=f_1$ and $f_2'=kf_2 .$ $U'$ is the toric variety given by the fan $F' = \sigma _1 ' \cup \sigma _2 '$ union of the two cones generated by $\{ f_1', f_1 '+f_2 '\}$ and $\{ f_2', f_1' +f_2 '\}$ respectively in the lattice $N' .$ From the description of $Z' \to Z$ we can easely read off the ramification. Take $V$ to be the normalization of $U$ in $k(Z')$ with the induced morphisms $f: V \to U$ and $t': V \to Z'.$ By construction, they satisfy the hypotheses of the lemma. Let $({{\mathcal X}}\to {{\mathcal C}}\to \Delta, {{\mathcal Q}})$ be a family of [*strictly prestable*]{} elliptic surfaces with zero section ${{\mathcal Q}}.$ Let $X_1 \subset {\mathcal X}_0$ be component of the central fibre ${{\mathcal X}}_0$, mapping down to a rational curve $C_1 \subset {\mathcal C}_0;$ also, let $X_2 \subset {\mathcal X}_0$ be the rest of ${\mathcal X} _0$ to which $X_1$ is attached along one and only one fibre, which is either stable or twisted. We have: \[one fibre\] Let $Q_1 = {{\mathcal Q}}\mid _{C_1} $ and let ${{\mathcal S}}$ be a divisor of ${{\mathcal X}}$ meeting $X_1$ transversally in only one stable fibre. Then there is a Zariski open neighborhood ${{\mathcal U}}$ of $Q_1$ in ${{\mathcal X}}$ and a toric variety ${{\mathcal Y}}$ with an étale morphism $T:{{\mathcal U}}\to {{\mathcal Y}}$ such that: $T^*{{\mathcal O}}_{{{\mathcal Y}}} (D_{{{\mathcal Y}}}) \simeq {{\mathcal O}}_{{{\mathcal U}}} ( X_1 + X_2 + {{\mathcal S}})$ Let $U$ be a Zariski naighborhood of $C_1$ with an étale map from a toric variety $t:Z \to U$ as in lemma \[torbasonefib\]. Let $f:V \to U$ be as in lemma \[toricbasext\]. Since $V$ is smooth (at least in a neighborhood of $C_1 \cap C_2$), by the purity lemma of Abramovich and Vistoli (cf. [**Lemma 2.1**]{}, [@A-V] ) the attaching fibre $F := (X_1 \cap X_2)_{red}$ is a quotient of a stable curve by a cyclic group of order $h$ dividing $k$ and the normalization of the pull-back of ${{\mathcal X}}\mid _U$ to $V$ is a family of stable curves. Therefore we have a diagram: $$\begin{array}{ccl} {{{\mathcal V}}} & \stackrel{s}{\to} & {{\mathcal X}}\mid _U \\ \pi ' {{\downarrow}} & & {{\downarrow}}\pi\\ V &\stackrel{f}{\to} & U \end{array}$$ such that $ \pi' :{{\mathcal V}}\to V$ is a family of stable curves and the action of the Galois group $\Gamma$ of $f: V \to U$ extends to an action on ${{\mathcal V}}$ such that $ {{\mathcal V}}/\Gamma \simeq {{\mathcal X}}\mid _U.$ According to lemma and \[toricbasext\] we have Zariski neighborhoods $U$ of $C_1$ in ${{\mathcal C}}$ and $U'$ of $C_1 '$ in ${{\mathcal C}}'$ and toric varieties with étale maps $U \to Z$ and $U' \to Z'$ such that the induced map $Z' \to Z$ is toric. It is thus enough to show that the pull-back ${{\mathcal Y}}'$ of ${{\mathcal V}}$ to $Z'$ is toric and that the induced action of the Galois group $\Gamma$ on ${{\mathcal Y}}'$ is an action by a subgroup of the torus. In this case the quotient ${{\mathcal Y}}/ \Gamma \to {{\mathcal X}}\mid _U$ would be the étale map in the statement of the proposition. Indeed, the total space of the normal bundle ${{\mathcal N}}_{{{\mathcal Q}}/{{\mathcal X}}'} \mid _U$ of the zero section ${{\mathcal Q}}$ in ${{\mathcal X}}$ is such a space, since every line bundle on a toric variety is a toric bundle, according to the proposition of page  [63]{} in [@F]. Thus if we can show that the action of $\Gamma$ on ${{\mathcal C}}'$ lifts to an action of $\Gamma$ on ${{\mathcal Y}}$ and that $\Gamma$ acts as a subgroup of the torus on ${{\mathcal Y}}$. But $\Gamma$ is a finite group and for each point $(u,v) \in {{\mathcal Y}}$ it acts on the element $v$ of finite dimensional vector space ${{\mathcal N}}_{Q/{{\mathcal X}}} \otimes \kappa (u),$ where $\kappa (u)$ is the residue field of $u \in V.$ Therefore $\Gamma$ must act linearly on ${{\mathcal N}}_{Q/{{\mathcal X}}}$, i.e., via multiplication by a character, and since $\Gamma$ acts as a subgroup of ${{{\mathbb{C}}}^*} ^2$ on $U$ we are done. In the following lemma, $Q_1 ^2$ denotes the self-intersection as ${{\mathbb{Q}}}$-divisor of $Q_1$ in $X_1.$ \[toric one fibre\] In the same hypotheses as proposition \[one fibre\], the fan $F$ of the toric variety ${{\mathcal U}},$ is $F=\sigma _1= \langle e_1 , e_1 +e_2 , e_3 \rangle \cup \sigma _2 =\langle e_3 ,e_1 +e_2,w\rangle $ in the lattice $N := e_1 {{\mathbb{Z}}}\oplus w {{\mathbb{Z}}}\oplus e_3 {{\mathbb{Z}}}$ where $\{e_1 ,e_2 ,e_3 \}$ is the standard basis of ${{\mathbb{R}}}^3$ and $w =\frac{1}{k} ( e_2 + n e_3).$ Here $-n=k Q_1 ^2$ (see remark \[self int\]) and ${{\mathcal C}}$ has an $A_{k-1}$ singularity around $C_1 \cap C_2 .$ Let $$\begin{array}{ccc} {{{\mathcal X}}}' & \stackrel{f}{\to} & {{\mathcal X}}\\ \pi ' {{\downarrow}} & &{{\downarrow}}\pi\\ {\mathcal C}' &\stackrel{s}{\to} & \mathcal C \end{array}$$ be the diagram as in \[one fibre\]. In an étale neighborhood of $C_1$ the surface ${{\mathcal C}}$ is described by the quasi-projective toric surface defined by the fan $\Delta$ obtained by the two cones generated by $\{f_1 , v=f_1 + k f_2 \}$ and $\{f_2 , v \}$ respectively in the lattice $L= f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}+ (k f_2 ) {{\mathbb{Z}}},$ where $\{ f_1 , f_2 \}$ is the standard basis for ${{\mathbb{R}}}^2 .$ This is the surface we refer to in \[one fibre\]. The fan of $V$, i.e., the toric étale neighborhood of $C_1 '$ in ${{\mathcal C}}'$ as in \[one fibre\], is obtained by taking the the two cones generated by $\{ f_1 ' , f_1 ' + f_2 ' \}$ and $\{ f_2 ' , f_1 ' + f_2 ' \}$ in the sub-lattice $L' = L + \frac {1}{k} ( f_2 ') \subset L,$ where $f_1 '=f_1$ and $f_2 ' = k f_2.$ Set $\Lambda = Hom _{{{\mathbb{Z}}}} (L, {{\mathbb{Z}}})$ and $\Lambda ' = Hom _{{{\mathbb{Z}}}} (L', {{\mathbb{Z}}}),$ and let $\langle ,\rangle : \Lambda '/\Lambda \times L/L' \to {{\mathbb{Q}}}/{{\mathbb{Z}}}$ the natural pairing. The Galois group $\Gamma$ as in \[one fibre\] is isomorphic to $L /L' \simeq {{\mathbb{Z}}}/ k {{\mathbb{Z}}}$ and its action on the ring of functions of $V$ is given by: $$\gamma \cdot \chi ^{n'} = e^{2 \pi i \langle \gamma , n'\rangle } \cdot \chi ^{n'},$$ for $\gamma \in \Gamma$ and $n' \in L';$ i.e., by: $$\gamma u = u \text{ and } \gamma uv = \zeta uv,$$ where $u= \chi ^{f_1 '} \text{, } uv=\chi ^ {f_1 ' + f_2 '}$ and $\zeta = e^{2 \pi i \frac{1}{k}}$ is a primitiv $k\text{-th}$ root of unity. We refer to [@F] for the notation. Let $\{ e_1 ,e_2 , e_3 \}$ be the standard basis of ${{\mathbb{R}}}^3 .$ The fan of ${{\mathcal Y}}$ is the union of the two cones $\sigma _1 '$ and $\sigma _2 '$ generated by $\{ e_1 , e_1+ e_2 , e_3 \}$ and $\{ e_1+e_2 , w' \}$ respectively, in the lattice $N' = e_1 {{\mathbb{Z}}}\oplus e_2 {{\mathbb{Z}}}\oplus e_3 {{\mathbb{Z}}}+ w' {{\mathbb{Z}}}+ (e_1 + e_2 ) {{\mathbb{Z}}},$ for some vector $w'$. We want to find $w'.$ Well, we know that the projection onto $e_2 {{\mathbb{R}}}\oplus e_3 {{\mathbb{R}}}$ along $e_1 + e_2$ is the fan of a toric étale neighborhood of $Q_1 '$ in $X_1 ',$ where $X_1 ' = {{\mathcal X}}' \times _{C_1} C_1 '$ and $Q_1 ' ={{\mathcal Q}}' \cap X_1 '$ is the corresponding zero section. Let $\pi _{e_1 + e_2} : {{\mathbb{R}}}^3 \to e_2 {{\mathbb{R}}}\oplus e_3 {{\mathbb{R}}}\simeq {{\mathbb{R}}}^2$ be such projection. Then $\pi _{e_1 + e_2} (x,y,z) = (y-x,z) $ and $ \pi _{e_1 + e_2} (w') = (-1, n).$ But $w'$ must project onto the vector $e_2$ via the projection $\pi _{e_3} : {{\mathbb{R}}}^3 \to e_1 {{\mathbb{R}}}\oplus e_2 {{\mathbb{R}}}\simeq {{\mathbb{R}}}^2,$ since the latter maps the cone of ${{\mathcal Y}}$ onto the cone of $V,$ as the ray $e_1 {{\mathbb{R}}}_{+}$ represents the zero section of ${{\mathcal Y}}.$ Hence $w'= e_2 + n e_3.$ In order to find the fan of ${{\mathcal Y}}/ \Gamma$ as in \[one fibre\] we need to identify a lifting in $N'$ of the sublattice $L$ of $L'.$ So $N = e_1 {{\mathbb{Z}}}\oplus e_2 {{\mathbb{Z}}}\oplus e_3 {{\mathbb{Z}}}+ w' {{\mathbb{Z}}}+ (e_1 + e_2 ) {{\mathbb{Z}}},$ and we want to find $w.$ Let $ \pi _{e_3} : {{\mathbb{R}}}^3 \to e_1 {{\mathbb{R}}}\oplus e_2 {{\mathbb{R}}}$ be the projection. Since the divisor corresponding to $e_3$ is the zero section ${{\mathcal Q}}',$ $ \pi _{e_3} (w)= \frac{1}{k} e_2.$ Also, the action of ${{\mathbb{Z}}}/ k {{\mathbb{Z}}}$ on the divisor ${{\mathcal S}}$ corresponding to $w$ (i.e., the pull-back to ${{\mathcal Y}}$ of the divisor in ${{\mathcal X}}$ corresponding to $S \subset {{\mathcal C}},$ with the notation as in lemma \[one fibre\] ), is trivial, therefore it must be trivial on $X_1 ' \cap {{\mathcal Y}}.$ Thus $\pi _{e_1 + e_2} (w)$ must lie on $w'=e_2 + n e_3;$ from this and from $\pi _{e_3} (w)= \frac{1}{k} e_2,$ we infer that $w=\frac{1}{k}(e_2 + n e_3).$ Two attaching fibres -------------------- The analogous lemmas and proposition hold for the case of a chain of rational curves joining two curves in the central fibre: Let $ {{\mathcal C}}\to \Delta$ be a family of nodal curves ${{\mathcal C}}.$ Assume that the central fibre $ {{\mathcal C}}_0 $ has a chain of rational components $C=\bigcup _{i=1}^N C_i$ meeting the rest of ${{\mathcal C}}_0$ transversally in only two points $q_1 \text{ and } q_2 .$ The singularities of ${{\mathcal C}}$ around $p_i=C_i \cap C_{i+1}$ and the $q_i$ are at worst $A_{k_{i}-1}$ singularities, $i=1,...,N+1.$ We have: \[torbastwofib\] Let $B$ be the rest of ${{\mathcal C}}_0$ and ${{\mathcal I}}\subset {{\mathcal O}}_{{{\mathcal C}}}$ the ideal sheaf of $C \cup B .$ Then there is an étale neighborood $U$ of $C$ in ${{\mathcal C}},$ a  2-dimensional toric variety $Z$ and an étale map $t:U \to Z $ such that: 1. the fan of $Z$ is $F= \bigcup _i \sigma_i$ where $\sigma_1=\langle f_1, f_1 + k_{1} f_2\rangle,$ $\sigma_i=\langle f_1+\sum _{j=1} ^{i-1} k_{j}f_2, f_1 + \sum _{j=1} ^{i} k_{j} f_2\rangle $ and $\sigma_{r+1}= \langle f_1+ \sum _{j=1} ^{r} k_{j}f_2, f_2\rangle$ 2. the pull-back via $t$ of the ideal of $D_Z $ is the ideal ${{\mathcal I}}\otimes {{\mathcal O}}_U .$ The proof is analogous to the one of lemma \[torbasonefib\], except for a few variations. Let $W$ be a Zariski neighborhood of $C$ such that $W \cap (C \cap B) = \{p , q\}.$ We can now contract $C$ to a rational double point of type $A_{r}$ where $r+1=\sum _i k_i .$ Let $\rho :{{\mathcal C}}\to Y$ be the contraction map and let $W' = \rho (W)$ and $p' = \rho (C) .$ Choose $U'$ to be an étale neighborhood of $W'$ that splits the node. Hence, there is an isomorphism $\phi ^{\sharp}: R=k[[x,y,t]]/(xy-t^{r+1}) \to \hat {{{\mathcal O}}} _ {W' , p'},$ such that the pull-back of the maximal ideal $\mathfrak m _{W' , p'} \subset \hat {{{\mathcal O}}}_{W',p'} $ is the maximal ideal $(x,y,t) \subset R.$ The homomorphism of complete local rings $\phi ^{\sharp}$ produces an étale map $\phi :U' \to {\operatorname{Spec}}R$ for some étale neighborhood $U'$ of $p'$ in $Y .$ Note that the maximal ideal $\mathfrak m _{{{\mathcal C}}, p}$ is generated by local equations of the branches of $B\cap U.$ Let $Z'$ be the toric variety whose fan $F'$ is the union of the cones $\sigma _1,$ ..., $\sigma _r ,$ $\sigma _{r+1}$ respectively generated $\{f_1, f_1 + f_2 \} ,$ ..., $\{f_1, f_1 + (r+1) f_2 \} ,$ $\{f_1+(r+1) f_2, f_2 \} $ in the lattice $N= f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}.$ Let $U' \to Z'$ be the normalization of $U$ in the function field of $Z'.$ As in lemma \[torbasonefib\] we construct a commutative diagram: $$\begin{array}{cccccc} U'&&\stackrel{t'}{\longrightarrow} &&Z' \\ &\searrow \alpha&&&\\ b\downarrow &&W&&\downarrow \beta\\ & p_1 \swarrow &&\searrow p_2&&\\ U &&\stackrel{t}{\dashrightarrow} && Z \end{array}$$ where $Z$ is the toric variety whose fan is $F =\bigcup \sigma_i$ where $\sigma_i=\langle f_1+ \sum _{j=1} ^{i-1} k_{j} f_2, f_1 + \sum _{j=1} ^{i} k_{j} f_2\rangle $ in $f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}},$ for $i\neq 1, r+1$ and where $\sigma_1=\langle f_1, f_1 + k_{1} f_2\rangle,$ and $\sigma_{r+1}= \langle f_1+ \sum _{j=1} ^{r} k_{j}f_2, f_2\rangle .$ An argument similar to the one given there, shows that the rational map $U \dasharrow Z$ is in fact regular. Let $B '' = c' _* B$ and let ${{\mathcal I}}"$ be the ideal sheaf of $B .$ By construction the pull-back of the ideal sheaf of $D_Z$ is the ideal sheaf ${{\mathcal I}}\otimes {{\mathcal O}}_U,$ by construction. As in the case of one attaching fibre, we want to be able to find a toric finite covering that “untwists” all the fibres of ${{\mathcal X}}\to {{\mathcal C}}$ above the $p_i$’s. In the hypothesis of lemma \[torbastwofib\], let the possible $A$-singularity of ${{\mathcal C}}$ at the ponts $p_i$ be of type $A_{k_i-1} .$ We have: \[toricbasexttwo\] There is a ${{\boldsymbol{\mu}}}_n \times {{\boldsymbol{\mu}}}_n $-covering $f: V \to U$ ramified along $C$ and along the two branches of $B$ coming off $C$ and an étale map to a  2-dimensional toric variety $t' :V \to Z'$ such that: 1. the fan of $Z'$ is $F' = \sigma '_1 = \langle f_1', (f_1 '+ k_1 f_2)\rangle \cup ... \cup \sigma _i '= \langle f_1' + \sum _{j=1} ^{i-1} k_{j} f_2 ', f_1 ' +\sum _{j=1} ^{i} k_{j} f_2 ' \rangle \cup ...\cup \sigma _{r+1} '= \langle f_2', f_1 ' + \sum _{j=1} ^{r+1} k_{j} f_2 ', \rangle $ in the lattice $N' =f_1 ' {{\mathbb{Z}}}\oplus f_2 ' {{\mathbb{Z}}}.$ 2. the pull-back via $t'$ of the toric ideal sheaf $D_Z $ is the ideal sheaf ${{\mathcal I}}_{C ' \cup B ' \cup S'} \mid _V$ of $C ' \cup B ' \cup S'$ restricted on $V,$ where $B ' = f^* B;$ 3. $f$ induces a toric morphism $f : Z' \to Z $ given by $f_1 ' = n f_1$ and $f_2 ' = n f_2 .$ We can apply lemma \[torbasonefib\] and find a toric variety $Z$ and an étale morphism $t: Z \to U$ such that $t^*{{\mathcal O}}_Z ({{\mathbb{O}}}_{f_1} + \sum _{i=1} ^N {{\mathbb{O}}}_{f_1 + k_if_2} + {{\mathbb{O}}}_{f_2}) \simeq {{\mathcal O}}_U (B+C).$ Let $N'$ the sub-lattice of $N$ $N' =f_1 ' {{\mathbb{Z}}}\oplus f_2 ' {{\mathbb{Z}}}$ of $f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}$ with $f_1'=nf_1$ and $f_2'=nf_2 $ and let $Z'$ be the toric variety given by the fan $F' = \sigma _1 ' \cup \bigcup _i \sigma _i' \cup \sigma _{r+1} '$ union of the cones $\sigma '_1 = \langle f_1', (f_1 '+ k_1 f_2)\rangle , $ $\sigma _i '= \langle f_1' + \sum _{j=1} ^{i-1} k_{j}f _2 ', f_1 ' +\sum _{j=1} ^{i} k_{j} f_2 ' \rangle $ and $\sigma _{r+1} '= \langle f_2', f_1 ' + \sum _{j=1} ^{r+1} k_{j} f_2 ', \rangle $ in the lattice $N' .$ We obtain the desired ramified finite covering $V \to U$ by letting $V$ be the normalization of $U$ in the function field of $Z'.$ What we have in mind is to look at those families of elliptic surfaces with section $({{\mathcal X}}\to {{\mathcal C}}\to \Delta , {{\mathcal Q}}\to \Delta)$ such that the central fibre contains a chain of rational components meeting the rest of the central fibre in only two fibres, and show that the picture is toric in this case too. This will allow us to perform the small contractions torically. So, let $X=\bigcup _{i=1} ^n X_i \subset {\mathcal X}_0$ be a chain of components of the central fibre, mapping down to a chain of rational curves $C= \bigcup _{i=1} ^n C_i \subset {\mathcal C}_0.$ Assume that $C$ meets $B,$ the rest of ${{\mathcal C}}_0 ,$ transversally in only two points, and corrispondingly $X$ is attached to the rest of ${{\mathcal X}}_0$ along two fibres that we assume to be either stable or twisted. Let $Q=\bigcup _i Q_i= {{\mathcal Q}}\mid _C$ We have: \[many fibres\] There is an étale neighborhood ${{\mathcal U}}$ of $Q$ in ${{\mathcal X}}$ and an étale morphism to a toric variety $T:{{\mathcal U}}\to {{\mathcal Y}}$ such that $T^*{{\mathcal O}}_{{{\mathcal Y}}} (D_{{{\mathcal Y}}})\simeq {{\mathcal O}}_{{{\mathcal U}}}(X + {{\mathcal X}}\mid _B ).$ As in proposition \[one fibre\], we can find a diagram: $$\begin{array}{ccc} {{{\mathcal V}}} & \stackrel{s}{\to} & U \\ \pi ' {{\downarrow}} & &{{\downarrow}} \pi\\ V &\stackrel{f}{\to} & U \end{array}$$ where $ \pi ':{{\mathcal V}}\to V$ is a family of stable curves and the action of the Galois group $\Gamma \simeq {{\boldsymbol{\mu}}}_n \times {{\boldsymbol{\mu}}}_n$ of $f: V \to U$ extends to an action on ${{\mathcal V}}$ such that $\Gamma \backslash {{\mathcal V}}\simeq {{\mathcal X}}.$ Let $k_i$ with $i=2,...,N$ and $k_1$ and $k_{N+1}$ be the orders of singularities around $ p_i:=C_i \cap C_{i+1}$ and $p_j:=C \cap B_j$ respectively ($i=2,...,N$ and $j=1,2$); and let $h_i$ and $n_j$ be such that the group acting non-trivially on the fibre above $p_i$ is $G_i \simeq {{\mathbb{Z}}}/ h_i{{\mathbb{Z}}}$ and on the one above $q_j$ is $\Gamma _j \simeq {{\mathbb{Z}}}/n_j {{\mathbb{Z}}}.$ According to lemmas \[torbastwofib\] and \[toricbasexttwo\] we can find toric varieties and étale morphisms: $t:U \to Z $ and $t':U' \to Z'$ such that $f:Z' \to Z$ is toric. We can choose toric cyclic coverings: $$t_i :V_i \to U$$ with Galois group ${{\boldsymbol{\mu}}}_{k_i}$ and that desingularize the singularities around points $p_i .$ The covering $t:U' \to U $ factors through the normalization $W$ of the fibre product $V_1 \times _U ...\times _U V_{N} \times _U V_{N+1} .$ Let $\pi _i :W \to V_i$ be the standard projection. Since $W$ is smooth around those points $p_i '$ that map to $p_i $ via $t_i \circ \pi _i$ and the fibres of ${{\mathcal X}}_i = {t_i \circ \pi _i}^* {{\mathcal X}}$ above those points are stable. Hence the same is true for ${{\mathcal V}}.$ Therefore on ${{\mathcal V}}$ we have an étale neighborhood ${{\mathcal U}}'$ of the zero section $Q '$ in ${{\mathcal V}}$ and an étale map $T':{{\mathcal U}}' \to {{\mathcal Y}}'$ to a toric variety, when we take as ${{\mathcal Y}}'$ the total space of the normal bundle of ${{\mathcal Q}}'$ in ${{\mathcal X}}:$ ${{\mathcal N}}={{\mathcal N}}_{{{\mathcal Q}}'/ {{\mathcal X}}'} \mid _{U'}.$ Now the as in proposition \[one fibre\] the action of $\Gamma$ on ${{\mathcal Y}}'$ is linearizable, and therefore it acts as a subgroup of the torus. We take as $T: {{\mathcal U}}\to {{\mathcal Y}}$ the quotient of $T' .$ The statement that $T^*{{\mathcal O}}_{{{\mathcal Y}}} (D_{{{\mathcal Y}}})\simeq {{\mathcal O}}_{{{\mathcal U}}}(X + {{\mathcal X}}\mid _B )$ follows from the analogous statement for t in lemma \[toricbasexttwo\] , since we obviously have a commutative diagram: $$\begin{array}{ccc} {{{\mathcal Y}}} & \stackrel{T}{\to} & {{\mathcal U}}\\ {{\downarrow}} & &{{\downarrow}} \\ Z &\stackrel{}{\to} & U \end{array}$$ Let us keep the notation as in lemma \[many fibres\] and set $Q_i = {{\mathcal Q}}\mid _{X_i}.$ We have: \[toric two\] The fan of the toric variety ${{\mathcal U}}$ is the union of the cones $\sigma _1 \text{ ,} \sigma _2 \text{, } ... \sigma _N,$ respectively generated by: $\{ \frac{1}{n} e_1, w_1 =e_1+ k_1 e_2 ,e_3 \},$ ... $ \{ w_N = e_1 + (k_1 + ...+ k_N) e_2 + (n_1 k_2 +(n_1 +n_2) k_3+ ....+ (n_1 +...+n_{N-1}) k_N ) e_3 , w_{N+1} = e_1 + (k_1 + ...+ k_{N+1}) e_2 + (n_1 k_2 + (n_1 +n_2) k_3+....+(n_1+...+n_N) k_{N+1}) e_3, e_3 \}$ in the lattice $L = \frac{1} {n} e_1 {{\mathbb{Z}}}\oplus e_3 {{\mathbb{Z}}}\oplus (\frac {1}{n} e_2 + \frac{a} {h_1} e_3) {{\mathbb{Z}}},$ where the $n_i =Q_i ^2$ and $g.c.m(a,h_1)=1$ are the integers determining the actions on the first fibre (hence the action on all the other fibres is determined by this datum). Let $(x_1, y_1, t) \text{, }(x_2, y_2, t)$....$(x_{N+1} ,y_{N+1},t)$ be coordites around $q_1=B_1 \cap C_1$ $p_1= B_2 \cap C_2$ .... $q_2=C_N \cap B_2,$ so that these a neighborhood of $p_i$ (resp. $q_j$) in ${{\mathcal C}}$ has equation $x_i y_i =t^k_i$ (resp.$x_j y_j =t^k_j$). As in lemma \[many fibres\] we have a diagram: $$\begin{array}{ccc} {{{\mathcal Y}}} & \stackrel{f}{\to} & {{\mathcal Z}}\\ \pi ' {{\downarrow}} & &{{\downarrow}}\pi\\ V &\stackrel{p}{\to} & U \end{array}$$ where all the morphisms and varieties are toric and ${{{\mathcal Y}}} / \Gamma \simeq {{\mathcal Z}},$ where $\Gamma$ is the Galois group. The fan of $U$ is given by the cones generated by $\{f_1 , v_1=f_1 + k_1 f_2 \}$, $\{v_1, v_2=f_1 + (k_1+k_2) f_2 \}$,..., $\{v_N , v_{N+1}=f_1 + (k_1+...+k_{N+1}) f_2 \},$ in the lattice generated by $\{f_1, v_1, ...v_{N+1} \}$ where $\{ f_1, f_2 \}$ is the standard basis of ${{\mathbb{R}}}^2 .$ Hence the fan of $V$ is the union of the cones $\langle f_1 ' , v_1 '=f_1 '+ k_1 f_2 '\rangle $, $\langle v_1 ', v_2 '=f_1 '+ (k_1+k_2) f_2 '\rangle $,..., $\langle v_N ' , v_{N+1} '=f_1 ' + (k_1+...+k_{N+1}) f_2 '\rangle ,$ where $f_1 ' =n f_1 $ and $f_2 ' = n f_2.$ The fan of ${{\mathcal Y}}$ is the union $\sigma _1 ' \cup ... \cup \sigma _{N+1}'$ of the cones: $\sigma _1' =\langle e_1, (e_1 + k_1 e_2) ,e_3\rangle $, $\sigma _2 ' =\langle e_1 + k_1 e_2, w_1 ' ,e_3\rangle ,$ ...$,\sigma _{N+1} '=\langle w_N ', w_{N+1} ', e_3\rangle $ for some vectors $w_i ',$ in the lattice $L'$ generated by $\{e_1, e_1 + k_1 e_2, w_1 ' , ...,w_N ', w_{N+1} '\}.$ Our first goal is to find these vectors. Let $\pi _{e_3} : {{\mathbb{R}}}^3 \to {{\mathbb{R}}}^2$ be the projection along $e_3$ onto $e_1 {{\mathbb{R}}}\oplus e_2 {{\mathbb{R}}}$, and let $\pi _{w_i '} : {{\mathbb{R}}}^3 \to {{\mathbb{R}}}^2$ be the projection along $w_i '$ onto $e_2 {{\mathbb{R}}}\oplus e_3 {{\mathbb{R}}}.$ The zero section ${{\mathcal Q}}$ is the toric divisor of ${{\mathcal Z}}$ corresponding to the ray $e_3,$ hence all the $w_i \text{'s}$ must project to $e_1 +(k_1 +..+k_i) e_2,$ since $\pi \mid _{{{\mathcal Q}}} : {{\mathcal Q}}\to U$ is an isomorphism (strictly speaking we should write ${{\mathcal Q}}\mid _U$ here, but with abuse of notation we shall simply write ${{\mathcal Q}}$ for the rest of the discussion). Analogously the surface $X_1$ corresponds to the ray $e_1 +k_1 e_2,$ the image of the cones $\sigma _1$ and $\sigma _2$ in the lattice $L' / \langle e_1 +k_1 e_2\rangle $ should give the fan of a toric varity whose only complete toric curve (the zero section) has self intersection $-n_1 .$ Therefore, the vector $ \pi _{ e_1 +k_1 e_2}(w_1 ') =(k_2,z)$ must be proportional to $(1,n_1),$ since $\sigma _1$ maps to the cone given by the second quadrant, and so $w_1 ' =(1, k_1 + k_2 , n k_2) .$ Assuming by inductive hypothesis that $w_ N ' = (1, k_1 +..+k_N, n_1 k_2 +...+(n_1+...+n_{N-1}) k_N ), $ we want to show that $w_ {N +1} ' = (1, k_1 +..+k_{N+1}, n_1 k_2 +...+(n_1+...+n_{N}) k_{N+1} ).$ Since $\pi _{e_3} (w_N ')=(1,k_1 +..+k_{N+1}),$ we just have to show that the last coordinate is the one claimed. Note that $\pi _{w_N '} ( w_{N+1}')= (k_N,z- n_1k_1+...+(n_1 + ...+n_{N-1})k_N)$ and and $\pi _{w_N '} ( w_{N-1})= (-k_{N-1},- (n_1+...+n_{N-2})k_{N-1})$ so $z=n_1 k_2 +...+(n_1+...+n_{N-1}) k_N .$ All is left to do is to identify the action of $\Gamma.$ Obviously the vector $\frac {1} {n} e_1$ from the lattice of the base lifts to an element of the super-lattice $L$ of $L'.$ In fact the kernel of the map: $${{\mathbb{Z}}}/ n {{\mathbb{Z}}}\times {{\mathbb{Z}}}/ n {{\mathbb{Z}}}\to {{\mathbb{Z}}}/ k_1{{\mathbb{Z}}}$$ as in lemma \[many fibres\], must be contained in the cone generated by the two adjacent vectors $e_1$ and $e_1 + k_1 e_2$, since it acts trivially on the fibres above the corresponding point. Note that this kernel is generated by $\frac{1}{n} e_1$ and $\frac{k_1}{n} e_2.$ We want to know more, namely we want to determine the kernel of the map: $${{\mathbb{Z}}}/ n {{\mathbb{Z}}}\times {{\mathbb{Z}}}/ n {{\mathbb{Z}}}\to {{\mathbb{Z}}}/ h_1{{\mathbb{Z}}}.$$ and in doing so, lift the vector $\frac{1}{n} e_2$ from the lattice of the base curve $U.$ Let $\pi _{e_3}$ as before, then we are looking for a vector $w$ such that $\pi _{e_3} (w) = \frac{1}{n} e_2$ and such that it represents the action of ${{\mathbb{Z}}}/ h_1 {{\mathbb{Z}}}$ on the fibre. That is to say, we want: $w=(0, \frac{1}{n}, z)$ with $z $ such that $h_1 z \in {{\mathbb{Z}}}.$ Hence $w=(0 , \frac {1}{n}, \frac{a} {h_1}) $ for some $a$ with $g.c.d(a,h_1)=1;$ the integer $a$ is completely determined by the action of ${{\boldsymbol{\mu}}}_{h_1} \simeq {{\mathbb{Z}}}/ h_1 {{\mathbb{Z}}}$ on ${{\mathcal X}}' .$ This proves the lemma. Toric log flips and toric log-canonical contractions ==================================================== Log-flips --------- In this section we want to show that we can perform the log-flips torically. We will need the following: Given a pair $(X,D)$ consisting of a variety $X$ and a divisor $D=D_1 + ...+D_n$ (where the $D_i$ are irreducible and reduced), we say that a a pair $(T, D_T)$ consisting of a toric variety $T$ and its toric divisor, is a [**toric étale neighborhood**]{} of a subvariety $Y \subset X$ if there is an étale neighborhood $u:U \to X$ of $Y$ in $X$ such that: 1. there is an étale map $t: U \to T ;$ 2. $t^* {{\mathcal O}}_T (D_T) \simeq {{\mathcal O}}_U (D);$ where ${{\mathcal O}}_U (D ) := u^* {{\mathcal O}}_X (D).$ Let $({{\mathcal X}}\to {{\mathcal C}}\to \Delta, {{\mathcal Q}}, )$ be a family of [*strictly prestable*]{} elliptic surfaces over a DVR scheme $\Delta .$ Assume that the special fibre ${{\mathcal X}}_0 \to {{\mathcal C}}_0$ contains a surface $X_1 \to C_1$ over a rational curve $C_1$ and a surface $X_2 \to C_2$ along a fibre $G_1$ which is either stable or a twisted curve. Let $Q_i = X_i \cap {{\mathcal Q}}.$ In what follows, we will say thar a triple $({{\mathcal Y}}\to S \to \Delta , {{\mathcal D}}, {{\mathcal Y}}\to {{\overline {\mathbf{M}}}}_{1,1})$ is [*locally stable*]{} if the naturally induced triple $({{\mathcal Y}}\to S \to \Delta ,{{\mathcal D}}, {{\mathcal Y}}\to S \times {{\overline {\mathbf{M}}}}_{1,1})$ is stable. We have: \[log-flip\] The log-flip of $Q_1$ gives rise to a locally stable family $({{\mathcal X}}^+ ,{{\mathcal Q}}^+) \to {{\mathcal C}}^+ \to \Delta ,$ or, which is the same the triple $ ({{\mathcal X}}^+ ,{{\mathcal Q}}^+, {{\mathcal X}}^+ \to {{\mathcal C}}^+ \times {{\overline {\mathbf{M}}}}_{1,1})$ is stable. Moreover the central fibre decomposes as a union of surfaces $X_1 ^+ \cup X_2 ^+$ where $X_1 ^+ $ is obtained from $X_1$ by contracting $Q_1$ and log-flipping produces a new rational curve $Q_1^+$ which meets $Q_2 ^+$ (the zero section of $X_2^+$) within the smooth locus of $X_2 ^+ .$ Furthermore, there exists [*an étale toric neighborhood*]{} $(T,D_T)$ whose fan $F^+ \subset N$ is given by (keeping the notation as in lemma \[toric one fibre\]) the union of the two cones: $\sigma ^+ _1 + \langle e_1 , w , e_3 \rangle$ and $\sigma ^+ _2 = \langle e_1 , e_1 + e_2 , w \rangle .$ This singularity is canonical, and in particular ${{\mathcal X}}_0 ^+$ is [*semilog-canonical*]{}. First, we want to show that it is enough to perform the log-flip in an étale neighborhood of $Q $ in ${{\mathcal X}}.$ In fact, if $\mathcal U$ is an étale neighborhood of $Q$ in ${{\mathcal X}},$ let ${\mathcal U }^+$ be the log-flip of $Q.$ It is clear that ${{\mathcal X}}\setminus {\mathcal U}$ and patch together to form an Artin algebraic (or analytic) space ${{\mathcal X}}^+ ,$ so we just need to show that we can find an ample line bundle on it. Let $S \subset {{\mathcal C}}$ any effective horizontal divisor meeting $C_1$ transversally. After possibly an étale base change $\Delta ' \to \Delta$ we can make sure that $S$ is a section around $C_1 ,$ so we may assume it is to begin with. So the bundle $\omega _{{{\mathcal X}}/ \Delta} \otimes {{\mathcal O}}_{{{\mathcal X}}}({{\mathcal Q}}) \otimes {{\mathcal O}}_{{{\mathcal X}}}({{\mathcal X}}\mid _S)$ is ${{\mathbb{Q}}}\text{-Cartier}$ and contracts $Q,$ and is $\Delta \text{-ample}$ on ${{\mathcal X}}^+ ,$ since by hypothesis $Q$ is the only curve on which $\omega _{{{\mathcal X}}/ \Delta} \otimes {{\mathcal O}}_{{{\mathcal X}}} ({{\mathcal Q}})$ fails to be positive. It therefore suffices to construct ${{\mathcal X}}^+$ as an algebraic space, because the fact that it possesses an ample divisor makes it automatically a scheme. From lemmas \[one fibre\] and \[toric one fibre\] we know that there is an étale neighborhood of $Q_1$ in ${{\mathcal X}}$ that is toric and whose fan $F$ is the union of the two cones $\sigma _1=\langle e_1 e_1+e_2 , e_3 \rangle $ and $\sigma _2= \langle e_3 ,e_1 +e_2 , w \rangle $ in the lattice $N := \bigoplus _{i=1,3} e_i {{\mathbb{Z}}}\oplus w {{\mathbb{Z}}},$ where $\{e_1 ,e_2 ,e_3 \}$ is the standard basis of ${{\mathbb{R}}}^3$ and $w =\frac{1}{k} ( e_2 + n e_3).$ The log-flip is thus constructed by taking the fan $F^+=\sigma _1 ^+ \cup \sigma _2 ^+$ where $ \sigma _1 ^+ =\langle e_1 ,w , e_3\rangle $ and $\sigma _2 ^+ =\langle e_1, e_1 +e_2 , w \rangle $ (see [@R1] Theorem 2.4), and now the curve $Q^+$ corresponds to the face generated by $\{e_1 , w ,w- e_1 \}$ and $X_2 ^+$ to the vector $e_1 +e_2.$ The matrix $A$ associated to the vectors $\{ e_1 , e_3 , w \}$ of the lattice $N$ is : $$\bordermatrix {&&& \cr & 1 &0 &0 \cr & 0 &0 &1 \cr & 0 &\frac{1}{k} &\frac{n}{k}} .$$ Since $det(A)=- \frac{1}{k},$ (i.e., $A\in SL_3 ({\frac {1}{k}} {{\mathbb{Z}}})$) these vectors form a lattice basis for $N,$ hence the toric variety corresponding to the cone $\sigma _1 ^+$ is smooth (see [@F] page 29). The rest of the proposition is obvious. In particular, what this says is that if we have a chain of rational curves $C_1 \cup...\cup C_n \subset {{\mathcal C}}_0$ above which the zero-section is an extremal ray, we can contract them one by one and perform the log-flips above them inductively. We can “straighten up” the lattice $N$ in which the fans $F$ and $F^+$ live. Indeed, by applying the transformation: $$\bordermatrix {&&& \cr & 1 &0 &0 \cr & 0 &k &0 \cr & 0 &-n &1 } ,$$ we can send $N$ to the lattice $f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}\oplus f_3 {{\mathbb{Z}}}, $ and the fan $F$ becomes $\langle f_1,f_3 , (f_1+ kf_2 -nf_3)\rangle \cup \langle (f_1 + kf_2 -nf_3), f_2 , f_3 \rangle $ and the fan $F^+$ becomes $\langle f_1,f_2, f_3\rangle \cup \langle (f_1 + kf_2 -nf_3), f_1 , f_2 \rangle .$ In particular the singular point of the threefold ${{\mathcal X}}^+$ through which $Q_1 ^+$ goes is an $\frac{1}{n} (1,k, 1)$ threefold singularity. We need to understand what the different surfaces after the log-fip look like, in particular what their singularities are. We have: With the notation as in theorem \[log-flip\]. Let $n'$ and $a$ be the unique integer such that $nn' \equiv 1 \text{(mod k)}$ and $0 \leq ak- n' <n .$ Then the singularity of $X_2 ^+$ at $Q_1 ^+ \cap G^+$ is an $A_{k,m}$ singularity, where $m=ak-n' .$ Straightforward, using the description of the fan $F^+$ given in the remark above, since the surface $X_2 ^+$ corresponds to the ray $f_1 {{\mathbb{R}}}_{\geq 0} ,$ that is to say $F^+ = \langle e_1 , e_2 \rangle \cup \langle e_2 , -ke_1 + n e_2 \rangle .$ and: Let $k'$ be the unique integer such that $kk' \equiv 1 \text{ (mod n)}$ and $0 \leq an- k' <n .$ Then the singularity of $X_1 ^+$ at the point to which $Q_1$ gets contracted is an $A_{n,k'} .$ singularity For this is more convinient to look at the description of $F^+$ in theorem \[log-flip\]. The surface $X_1 ^+$ corresponds to the ray $\langle e_1 + e_2\rangle .$ Let $\pi : {{\mathbb{R}}}^3 \to {{\mathbb{R}}}^3$ be the projection along that ray onto $e_2 {{\mathbb{R}}}\oplus e_3 {{\mathbb{R}}}.$ Then $\pi (w) = (\frac{1}{k}, \frac{n}{k}) ,$ and $\pi (e_1)= (-1,0) .$ The cone $\langle \pi(w) , \pi(e_1)\rangle $ in the lattice $e_3 {{\mathbb{Z}}}\oplus \pi(w) {{\mathbb{Z}}}$ is equivalent, via the transformation $\pi(w) \to f_1$ and $e_3 \to f_2 ,$ to the cone $\langle f_1, (nf_2 -kf_1)\rangle ,$ in the fan $f_1 {{\mathbb{Z}}}\oplus f_2 {{\mathbb{Z}}}.$ The former gives rise to an affine toric variety isomorphic to the one in the statement (sending $f_1$ to $f_2$ and viceversa). We will also need: \[selfintinv\] Keeping the notations as above, one has that: $${Q_1 ^+ }^2 = \frac{1}{Q_1 ^2} .$$ The fan of a toric neighborhood of $Q_1 $ in $X_1 $ is given by: $$F = \langle e_1 , e_2 \rangle \cup \langle e_2 , -k e_1 + n e_2 \rangle$$ in the lattice $N = e_1 {{\mathbb{Z}}}\oplus e_2 {{\mathbb{Z}}};$ the one of a toric neighborhood of $Q_1 ^+$ in $X_2 ^+$ is given by: $$F ^+= \langle e_1 , e_2 \rangle \cup \langle e_2 , k e_1 - n e_2 \rangle .$$ Hence we have that $Q_1 ^2 = -\frac {n}{k}$ and ${Q_1 ^+} ^2 = -\frac {k}{n}.$ When performing a log-flip according to theorem \[log-flip\], even if the surface $X_2$ were [*standard* ]{} to begin with, the surface $X_2 ^+$ in general is not. In fact it is some toric blow-up of one. We are therefore naturally led to the following: \[logstand\] A [**log-standard elliptic surface**]{} $(Y \to C,Q,G+F)$ is a triple consisting of an elliptic surface $Y\to C, $ its zero section $Q$ and a marking of $s$ curves $G=\bigcup G_i$ and and $N$ of fibres $F=\cup F_j$ with a regular birational map $g:Y \to Y',$ called [**structure morphism**]{}, to a [*standard* ]{} elliptic surface $Y'\to C$ with zero section $Q'$ such that: 1. the exceptional divisor $E=\bigcup E_i$ is the disjoint union of smooth and irreducible rational curves $E_i$ meeting the $G_i$ transversally at one point, and $g(E_i)=p_i =g(G_i) \cap Q \in Y' ;$ 2. for each $i$ and $j$ the curves $G_i ':=g( G_i)$ and $F_j '= g( F_j)$ are either stable or a twisted fibres of $Y' $ and $G_i $ is the proper transform of $G_i '.$ 3. for each $i$ there exist étale neighborhoods $U \to Y$ of $E_i \subset Y$ and $U' \to Y'$ of $p_i \in Y'$ and morphisms to toric varieties $t:U\to T$ and $t' : U' \to T';$ 4. the fan of $T$ is the union of the two cones $\langle e_1 , e_2 \rangle \cup \langle e_1 , ke_1 -n e_2 \rangle $ in the lattice $N= e_1 {{\mathbb{Z}}}\oplus e_2 {{\mathbb{Z}}}, $ and $t^* {{\mathcal O}}(D_T ) \simeq {{\mathcal O}}_U ( E_i + G_i + Q );$ 5. the fan of $T'$ is the cone $\langle e_2 , ke_1 -n e_2 \rangle $ in $N,$ and ${t'}^* {{\mathcal O}}(D_T') \simeq {{\mathcal O}}_{U'} ( G_i' + Q ');$ 6. the morphims $U \to U'$ is induced by the toric blow-up $T \to T'$ induced by the subdivision $\langle e_1 , e_2 \rangle \cup \langle e_1 , ke_1 -n e_2 \rangle $ of $\langle e_2 , ke_1 -n e_2 \rangle. $ Here $D_Z$ is the toric divisor of a toric variety $Z.$ We shall call a divisor like $G_i$ a [**splice**]{} and one like $F_i$ a [**scion**]{}. (250,250)(0,0) (180,30)[(100,200) [![image](logstand.eps)]{}]{} (150,20)[[fig.6]{}]{} Something analogous happens to the surface $X_1 ^+ ,$ except that we now loose the fibration structure. Let us keep the same notation as above. We have: \[I\] A [** N-pseudoelliptic surface**]{} is a pair $(Y,G+F)$ consisting of a surface $Y,$ $N$ marked curves $F=\bigcup F_i$ and $s$ marked curves $G= \bigcup G_i$ with a regular birational map $g:Y' \to Y,$ called [**structure morphism**]{}, from a [*log-standard* ]{}elliptic surface $(\pi ' :Y' \to {{\mathbb{P}}}^1, Q' , G'+F')$ such that: 1. The proper transform of the $F_i$ are fibres of $Y' \to {{\mathbb{P}}}^1 ;$ 2. the exceptional divisor of $g$ is the zero-section $Q$ of $Y'$ and $g$ maps $Q$ to a point $p;$ 3. there exist a étale neighborhoods $V\to Y$ of $p \in Y$ and $V'\to Y'$ of $Q'\subset Y'$ and morphisms to toric varieties $\tau : V \to Z$ and $\tau ': V' \to Z'$ and a toric morphism $b: Z' \to Z$ such that $ b \circ \tau ' = \tau \circ g ;$ 4. the fan of $Z'$ is the union of the two cones $\langle e_1 , e_2 \rangle \cup \langle e_1 , ke_1 -n e_2 \rangle $ in the lattice $N$ and $\tau ^* {{\mathcal O}}(D_Z ) \simeq {{\mathcal O}}_U ( F_1' +S '+ Q ')$ for some fibre $S'$ of $Y' \to {{\mathbb{P}}}^1 ;$ 5. the fan of $Z$ is the cone $\langle e_2 , ke_1 -n e_2 \rangle $ in the lattice $N$ and $\tau ^* {{\mathcal O}}(D_{Z'} ) \simeq {{\mathcal O}}_U ( F_1 +S )$ where $S=g_* S' ;$ 6. the toric morphism $b: Z' \to Z$ is induced by the subdivision $\langle e_1 , e_2 \rangle \cup \langle e_1 , ke_1 -n e_2 \rangle $ of $\langle e_2 , ke_1 -n e_2 \rangle .$ We call such a $Y$ [**isotrivial** ]{}if $Y' \to {{\mathbb{P}}}^1$ is isotrivial. Furthermore, we call [**pseudolleptic surface of type I**]{} a [* 1-pseudoellitpic surface* ]{} $(Y,G+F_1);$ the component $G_i$ is still called a [**splice**]{} and $F_i$ is still called a [**scion**]{}. An $n$-pseudoelliptic surface is [*log-canonical*]{} if and only if $n \leq 2 .$ (250,250)(0,0) (180,30)[(100,200) [![image](2pseudoell.eps)]{}]{} (150,20)[[fig.7]{}]{} After performing any number of flips, we want to make sure that we know what happens to $X_1 ^+$ and $X_2 ^+$ in particular that we know when the restriction of the log-canonical divisor is nef and big on them. This is taken care by the following proposition: \[ample\] 1. Let $Y$ be a log-standard surface, $g: Y \to Y'$ its structure morphism with $Y'$ standard. Let $$L_{Y'} = K_{Y'} + \sum _{i=1} ^s G_i ' + \sum _{j=1}^r F_j ' +Q'$$ and $$L_Y = K_Y + \sum _{i=1} ^s G_i + \sum _{j=1}^r F_j+Q ;$$ then: $$g ^* L_{Y '} =L_{Y} + \sum _{i=1}^s E_i .$$ 2. Let $Y$ be an $n$-pseudoelliptic surface with structure morphism $g : Y' \to Y,$ with $Y'$ is log-standard. Let: $$L_{Y'} = K_{Y'} + \sum _{i=1}^s G_i ' + \sum _{j=1}^r F_j ' +Q$$ and $$L_Y = K_Y + \sum _{i=1}^s G_i + \sum _{j=1}^r F_j;$$ then: $$g ^* L_{Y } =L_{Y'} + \frac{2-n}{Q ^2} Q .$$ We start with proving part (1). We have that: $$g ^* L_{Y' } =L_{Y} + \sum _{i=1} ^s a_i E_i$$ given that $\sum _{i=1} ^s E_i$ is the exceptional divisor of $g.$ By an easy inductive argument (i.e., by performing the toric blow-ups one at the time), one can easily convince oneself that infact all the $a_i's$ must be equal to each other; let us indicate this rational number by $a.$ Let $\pi :Y\to C$ be the projection to the base curve of $Y,$ and let $g$ be the genus of $C.$ Furthermore, let $k$ be the least common multiple of all the orders of monodromy around the fibres $F_i .$ We can take a base change of oder $k$ as in proposition \[genlogcan\] to untwist the possible twisted fibres and get a diagram: $$\begin{array}{ccc} S &\stackrel{f} {\rightarrow} & Y \\ {\downarrow}& &{\downarrow}\\ B &\stackrel{\phi} {\rightarrow} &C \end{array}$$ Now the zero section $\overline Q$ of the surface $S$ is entirely contained in the smooth locus of $S,$ and therefore, if we indicate respectively by ${\overline F}_i$ and ${\overline G} _j $ the proper transforms of $F_i$ and $G_j ,$ by a computation analogous to the one in proposition \[genlogcan\], we get: $$L_Y \cdot Q= \frac {1}{k} (K_S +\sum _i {\overline F} _i + Q + k \sum _j {\overline G} _j)\cdot {\overline Q}= 2g-2+s .$$ Since according to proposition \[genlogcan\], we have: $$L_{Y'} \cdot Q'= 2g-2+s+r ,$$ we can infer that $a=1 .$ This completes the proof of part (1). As for part (2), we can again write: $$g^* L_{Y} = g ^* (K_{Y} +\sum _{i=1}^s G_i + \sum _ {j=1} ^r F_j ) =(K_{Y'} + \sum _{i=1}^s G_i '+ \sum _ {j=1} ^r F_j ' +Q ')+ a Q= =L_{Y '} + a Q .$$ But $ g ^* L_{Y} \cdot Q =0$ by the projection formula and $L_{Y '} \cdot Q =n-2 $ by lemma \[genlogcan\], therefore: $$n-2 + a Q ^2 = 0 .$$ This proves part (2). We need to say something about the positivity of the log-canonical bundle of a log-canonical surfaces and of n-pseudoelliptic ones. We have: \[flipextremalray\] 1. Let $Y\to C$ be log-standard with base curve $C$ of genus $g,$ with $r$ $F_j$ ’s and $n$ $G_i$’s. Moreover, assume that $ -1<E_i ^2<0 .$ Then the following hold: 1. if $2g-2+r >0,$ then $L_Y$ is ample; 2. if $2g-2+r=0 ,$ then $L_Y$ is semiample, and for any irreducible curve $D:$ $$L_Y \cdot D=0 \text{ iff } D=Q ;$$ 3. if $2g-2+r <0$ then $[Q] \in {\overline {NE}} (Y)$ is an extremal ray for $L_Y .$ 2. let $Y$ be $n-$pseudoelliptic, then: 1. if $n\geq 2$ then $L_Y$ is ample; 2. if $n=1$ and $Q ^2 < -1 ,$ then $L_Y$ is ample 3. if $n=0,$ then $L_Y$ is ample if and only if $Q^2 <-4 .$ We will start proving part (1). [**Proof of part (1)**]{} Recall from Proposition \[ample\], part (1), with the same notation as therein, that: $$g^*L_{Y'}=L_Y + \sum _{i=1}^s E_i .$$ Therefore, in order to show that $L_{Y}^2 >0,$ it suffices to show that $L_{Y'} ^2 >0$ and $2L_Y \cdot \sum _i E_i + \sum _i E_i ^2>0 .$ The first inequality is a consequence of corollary \[twistedampleness\], and: $$2L_{Y} \cdot \sum _{i=1}^s E_i + \sum _{i=1}^s E_i ^2 = - \sum _{i=1}^s E_i ^2 >0,$$ where we have used that $L_Y \cdot \sum _{i=1}^s E_i = g^*L_{Y'} \cdot \sum _{i=1}^s E_i + (\sum _{i=1}^s E_i)\cdot (\sum _{i=1}^s E_i) .$ Now,if $D$ is any irreducible curve on $Y$ other than one of the $G_i$’s or one of the $E_i$’s, then $L_Y \cdot D= L_{Y'} \cdot g(D)$ which is positive according to corollary \[twistedampleness\]. In case the irreducible curve is $E_i,$ then we have that: $$L_Y \cdot E_i = -E_i ^2 >0 ;$$ and in case it is $G_i,$ then, since $-G_i ^2 = G_i \cdot E_i,$ we have: $$L_Y \cdot G_i = L_{Y'} \cdot G_i ' - E_i \cdot G_i= Q' \cdot G_i ' - E_i \cdot G_i ,$$ and since $G_i ' \cdot Q' =\frac{1}{-E_i ^2} E_i \cdot G_i$ (this is obtained by writing $g^* Q ' = G_i + \sum a_i E_i$ and intersecting with $E_i$ to obtain $a_i$ and then intersect $g^*Q '$ with $G_i$), we have that: $$L_Y \cdot G_i= (\frac{1}{-E_i ^2}-1) E_i \cdot G_i >0$$ since by lemma \[selfintinv\]: $$-1 <E_i ^2 <0 .$$ We can then conclude part (1) with the aid of corollary \[twistedampleness\]. We now prove part (2). [**Proof of part (2)**]{} Recall from Proposition \[ample\], part (2) $$g^* L_Y =L_{Y'} + \frac {2-n}{Q^2} Q .$$ This implies that $L_Y \cdot G_i>0.$ If $n\geq 2$ $L_{Y'}$ is ample for the previous part, and $ \frac {2-n}{Q^2} Q$ is effective (or $0$ if $n=2$), so part (a) is proved. If $n=1,$ $$g^* L_{Y} = L _{Y'} +\frac {1}{Q^2} Q .$$ Let $g'':Y' \to Y''$ the structure morphism of the log-standard surface $Y' .$ Then, if we set: $$M:= K_{Y''} + \sum G_i '' +F'' +(1+ \frac{1}{Q^2}) Q'',$$ where $Q'' =g'' (Q')$, $F'' =g''(F')$ and ${G_i }'' = g'' ({G_i} '),$ then: $$g^* L_Y = {g''}^* M +\sum E_i$$ and again we can conclude by means of part (1) and corollary \[twistedampleness\], given that $0< a:= (1+\frac {1}{Q^2})<1 $ and that $\lambda >1 .$ Let us finally analyze the case in which $n=0.$ As before, we have: $$g^* L_{Y} ={g''}^* M +\sum E_i .$$ where now: $$M:= K_{Y''} + \sum G_i '' +(1+ \frac{2}{Q^2}) Q'',$$ Therefore one can once again conclude, with the aid of corollary \[twistedampleness\] (given that $Q^2 < -4 $) that $g^* L_{Y}$ is positive on every curve that is different from $Q$ and from one of the $E_i$’s. Indeed, every other curve, except for the $G_i$’s (and for these the computation is exactly like in part (1) of this proposition), will not meet the $E_i$’s. Also, since: $$L_{Y} ^2 = M^2 +\sum _{i=1} ^s E_i ^2$$ from the very same corollary, we get that $L_Y ^2$ is positive as soon as $Q^2 <-4 ,$ in fact, we have: $$L_Y ^2 = 2 a (s-2) + a (2-a) \lambda + \sum _{i=1} ^s E_i ^2 >0$$ with $a:= (1+ \frac{2} {Q^2})$ and $\lambda = - Q^2,$ and therefore, since $E_i ^2 >-1,$ as soon as: $$2 a (s-2) + a (2-a) \lambda -s >0$$ and this occurs as soon as $Q^2 < -4$ as claimed. So we only need to check that $L_Y \cdot E_i >0 ,$ but this is clearly true, due to the fact that $E_i ^2 <0$ and $M\cdot E_i =0 .$ This concludes the proof, again thanks to Nakai-Mosheizon. We also want to know how $X_2 ^+$ has changed. For one thing we know that now it must be a [*log-standard*]{} elliptic surface, even if it were only [*standard* ]{} to begin with. We want to know whether $X_2 ^+$ is [*strictly pre-stable*]{} when $X_2 $ is. We have: \[self-int2\] Mantaining the notation as in proposition \[flipextremalray\] part (1), the self intersection $Q^2$ of $Q$ in $Y$ is: $$Q ^2 = {Q'} ^2 + \frac{1}{E_i ^2}$$ where ${Q'} ^2$ is taken in $Y'.$ In particular, if $Y'$ is a strictly prestable standard surface (i.e., $ {Q'} ^2 <-1$) and $-1<E_i ^2 <0$ for each $i,$ then $Q^2 <-1$ as well. Write $g^* Q= Q' + \sum a_i E_i ,$ and intersect with $E_i ,$ to see that it must be that $a_i = \frac{1}{-E_i ^2}$ (since $Q\cdot E_i =1$). This yields: $${Q'} ^2 = Q \cdot b^* Q' = Q^2 + \frac{1}{-E_i ^2}$$ which concludes the proof. We can therefore make the following: We call a [*triple* ]{} $(X \to C, Q, G+F)$ consisting of a log-standard elliptic surface with a given marking [**strictly prestable**]{} if: 1. $(Q \mid _{X \mid _B})^2 <-1 $ for each rational component $B \subset C;$ 2. if $g: Y:= X \mid _B \to Y'$ is the structure morphism, then for each irredecubile component $E_i$ of the exceptional divisor $E$ of $g$ we have that $-1<E_i ^2 < 0 .$ Analogously we call a [*quadruple*]{} $(X \to C, Q+F, G, f:X \to C \stackrel {j}{\to} {{\overline {\mathbf{M}}}}_{1,1})$ consisting of a log-standard elliptic surface with marking and map to moduli [**strictly prestable**]{} if the same conditions are asked only of those rational components $B \subset C$ for which $j\mid _B : B \to {{\overline {\mathbf{M}}}}_{1,1}$ is constant. Small log-canonical contractions -------------------------------- Let $(\pi: {{\mathcal X}}\to {{\mathcal C}}\to \Delta , {{\mathcal Q}})$ be a family of [*strictly prestable*]{} log-standard elliptic surfaces with section over a DVR scheme (or a polydisk, if one favour the analytic flavor)) $\Delta,$ such that the special fibre ${{\mathcal X}}_0 $ contains a chain of surfaces $X:=X_1 \cup ...\cup X_N$ attached transvesally along stable or twisted fibres, with base curve a chain or rational curves $C:= C_1 \cup ...\cup C_N .$ Furthermore assume that $X$ is attached transversally to $\overline { {{\mathcal X}}_0 \setminus X}$ along one (twisted or stable) fibre each end. Let $Z_1 \to B_1$ and $Z_2 \to B_2$ be the adjacent surfaces (that is to say the irreducible components of $\overline { {{\mathcal X}}_0 \setminus X}$ that meet $X$ at each end). Let ${{\mathcal C}}' $ be obtained from ${{\mathcal C}}$ by contracting the curve $C,$ $\rho : {{\mathcal C}}\to {{\mathcal C}}'$ the contraction map and $p= \rho (C).$ Let also $w_1 =e_1+ k_1 e_2 \text{, ..., } w_i = e_1 + (k_1 + ...+ k_i) e_2 + (n_1+...+n_{i-1}) k_i e_3) \text{, ..., } w_{r+1} = e_1 + (k_1 + ...+ k_{r+1}) e_2 + (n_1+...+n_{r}) k_{r+1} e_3 ,$ and $L$ be the lattice $L = \frac{1} {n} e_1 {{\mathbb{Z}}}\oplus e_3 {{\mathbb{Z}}}\oplus (\frac {1}{n} e_2 + \frac{a} {h_1} e_3){{\mathbb{Z}}}.$ According to lemma \[extr ray\] ${{\mathcal Q}}\mid _C$ must be contracted by the log-canonical map, and in fact the contraction is described by the following: \[logcancont\] The stable model of ${{\mathcal X}}$ is a family ${{\mathcal X}}^c \stackrel {\pi '} {\to} {{\mathcal C}}' \to \Delta$ of surfaces such that the generic fibre and $\overline{{{\mathcal X}}^c _0 \setminus X}$ have not changed and $X^c = {\pi '} ^{-1} (p)$ is attached to the rest of the central fibre along marked curves. The singularity of this point in ${{\mathcal X}}^c$ is toric, and a toric neighborhood of $p$ in $({{\mathcal X}}^c , Z_1 ^c+ Z_2 ^c)$ (where $Z_i ^c:= {{\mathcal X}}^c \mid _ {\rho (B_i)}$) is given by the cone: $\sigma =\langle \frac{1}{n} e_1 , w_1 \text{, ... } w_r w_{N+1} , e_3 \rangle$ in the lattice $L .$ This singularity is canonical, and in particular ${{\mathcal X}}^c _0$ is [*semi-logcanonical*]{}. As in the proof of theorem \[log-flip\] we can reduce ourselves to finding the log-canonical contraction on an étale neighborhood, since by hypothesis $\omega _{{{\mathcal X}}/ \Delta} ({{\mathcal Q}})$ is $\Delta \text{-ample}$ except for contracting $Q .$ The rest is a direct consequence of [@R1], proposition \[many fibres\] and lemma \[toric two\]. The normalization ${X^c}^{\nu}$ of $X^c = X_i ^c$ consists of a union of normal surfaces $(Y_i ^c, {G_1^c} ^i , {G_2^c} ^i)$ with marked double curves ${G_1^c} ^i$ and ${G_2^c} ^i,$ and if $(Y_i ^c, {G_1^c} ^i , {G_2^c} ^i)$ denote components with double curves of the normalization of $X,$ we have a map $Y_i \to Y_i ^c$ which contracts the zero section to a point $p' \in {X^c}^{\nu}.$ \[fantypeII\] There is an étale neighborhood U of $p' \in Y_i ^c $ with a map to a toric variety $g:U \to Z$ such that: 1. the fan of $Z$ is the cone $\Delta = \langle e_1 , w =-e_1+ ne_2 \rangle$ in the lattice $N = ({\frac {1}{n}} e_1 + \frac {a}{h_1}e_2) {{\mathbb{Z}}}\oplus (\frac{1}{n}e_1 + \frac{n-a}{h_1}e_2){{\mathbb{Z}}},$ where $a$ and $h_1$ are completely determined by the monodromy of the action around one of the two “marked” fibres $G_1$ and $G_2 .$ 2. $g^* {{\mathcal O}}_Z (D_Z) \simeq {{\mathcal O}}_U (Q +G_1 +G_2),$ where $D_Z$ is the toric divisor. The lemma is a consequence of the following two observations. On the one hand that the projection along $e_3$ onto $e_1$ must determine the action on $Q\simeq {{\mathbb{P}}}^1,$ which is given by choosing the lattice $\frac{1}{n}e_1.$ On the other hand, let $h_1$ be the integer such that ${{\boldsymbol{\mu}}}_{h_1}$ is the subgroup of ${{\boldsymbol{\mu}}}_n$ acting nontrivially on $G_1.$ The kernel of the map $${{\boldsymbol{\mu}}}_n \times {{\boldsymbol{\mu}}}_n \to {{\boldsymbol{\mu}}}_{h_1}$$ is determined by a vector $v=(\frac{1}{n}, z)$ such that $h_1 z=a \in {{\mathbb{Z}}}.$ Thus the claim. What one has to prove now is that the log-canonical divisor on such components, which originally was only nef and big, has become ample after contracting the zero-section. In fact: \[ample2\] Let $X ^c $ be a $2-$pseudoelliptic surface, $\alpha : X \to X ^c ,$ its structure morphism, and assume that $X \to C$ is a [*strictly stable*]{} log-standard surface ($C$ rationl), with zero section $Q.$ Then: $\alpha ^* L_{X ^c } =L_{X} .$ In particular, $L_{X ^c }$ is ample. Write: $$\alpha ^* L_{X ^c } =L_{X} + a Q ,$$ and intersect with $Q.$ By the projection formula, $\alpha ^* L_{X ^c } \cdot Q = 0$ and $L_{X} \cdot Q =0$ according to lemma \[extr ray\], therefore the conclusion, on accounts of proposition \[flipextremalray\] part (2). \[II\] We call [**pseudoelliptic surface of type II**]{} a [*pair*]{} $(Y ^c, G^c +F_1 ^c +F_1 ^c )$ consisting of a [* 2-pseudoelliptic surface*]{} as in definition \[I\] with the extra conditions that $S$ in point $5)$ is $F_2 ^c ,$ and that we replace the fan of point $5)$ with the fan of lemma \[fantypeII\] Such a triple is called [**isotrivial**]{} if the surface $Y\to Y^c$ as in lemma \[fantypeII\] is isotrivial. The Stable reduction theorems ============================= Stable reduction of triples --------------------------- We are now ready to prove the stable reduction theorem in the relative case of elliptic surfaces with sections and endowed with a regular map to ${{\overline {\mathbf{M}}}}_{1,1}.$ \[stabred\] Let ${{{\mathcal X}}}_\eta\to {{{\mathcal C}}}_\eta\to \eta$ be a stable elliptic surface over a smooth curve ${{{\mathcal C}}}_\eta .$ Then there is a finite extension of discrete valuation rings $R\subset R'$ and a triple $( {{\mathcal X}}' \to {{\mathcal C}}' , {{\mathcal Q}}' , f':{{\mathcal X}}' \to {{\overline {\mathbf{M}}}}_{1,1})$ over $S'$ such that: 1. ${{\mathcal X}}' \to {{\mathcal C}}'$ gives rise to an extension: $$\begin{array}{ccc} {\mathcal X}_\eta\times_{\Delta} \Delta' &\subset & {\mathcal X}' \\ {\downarrow}& & {\downarrow}\\ {\mathcal C}_\eta \times_{\Delta} {\Delta}' &\subset & {\mathcal C}' \\ {\downarrow}& & {\downarrow}\\ \{\eta'\} & \subset & {\Delta}', \end{array}$$ compatible with the extension ${{\mathcal Q}}_{\eta} \times _S S' \subset {{\mathcal Q}}' ;$ 2. ${{\mathcal C}}'\to {{\overline {\mathbf{M}}}}_{1,1}$ is Konstevich stable. 3. the components of ${{\mathcal X}}'_0$ that dominate the components of ${{\mathcal C}}' _0$ are [*log-standard stricly prestable*]{} quadruples $(X,Q,G, f: X \to {{\overline {\mathbf{M}}}}_{1,1})$ where $G$ consists of either one or two fibres, which are either stable or twisted; if $X\to C$ in such a quadruple turns out to be isotrivial, then $C$ is not rational. 4. the components of ${{\mathcal X}}'_0$ that are mapped to a point of ${{\mathcal C}}' _0$ are [*isotrivial log-pseudoelliptic surfaces*]{} either of type I or II. In the former case they are attached to the rest of the central fibre according to lemma \[toric one fibre\] and in the latter according to lemma \[toric two\]. The extension is unique up to a unique isomorphism, and its formation commutes with further finite extensions of discrete valuation rings. By the [*strictly prestable reduction theorem*]{} (theorem \[divcont\]) we can find a finite base change $\Delta ' \to \Delta$ , $\Delta ' \text{-schemes} $ and $\Delta ' \text{-morphisms}$ unique up to unique isomorphisms $({{\mathcal X}}'\to {{\mathcal C}}' , {{\mathcal Q}}', f: {{\mathcal X}}'\stackrel{\pi '}{\to} {{\mathcal C}}' \stackrel{j '}{\to} {{\overline {\mathbf{M}}}}_{1,1})$ that extend $({{{\mathcal X}}}_\eta\to {{{\mathcal C}}}_\eta ,{{\mathcal Q}}_{\eta} , f_\eta ),$ which is [*strictly prestable*]{}. The extension commutes with further base changes and the log-canonical divisor $\omega _{{{\mathcal X}}' / \Delta '}({{\mathcal Q}}') \otimes {f'}^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3)$ is ample away from those isotrivial components of the central fibre ${{\mathcal X}}' _0$ the meet the rest of the central fibre in one or two fibres (twisted or stable). Let $X\to C$ such a component. Then ${j'}^* {{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} \simeq {{\mathcal O}}_C,$ and therefore $$\omega _{{{\mathcal X}}' / \Delta '} ({{\mathcal Q}}') \otimes f^*{{\mathcal O}}_{{{\overline {\mathbf{M}}}}_{1,1}} (3) \mid _X \simeq \omega _{{{\mathcal X}}/ \Delta '}({{\mathcal Q}}') \mid _X.$$ According to lemma \[extr ray\] we know that if an irreducible rational curve meets the rest of the central fibre in one point, we need to log flip the zero section above it. According to theorem \[log-flip\] we can then perform the log flip to get $({{\mathcal X}}^+ \to {{\mathcal C}}^+ , {{\mathcal Q}}^+ ).$ In doing so, one produces an isotrivial pseudoelliptic surface of type I $(X_1 ^+, G_1)$ attached to a log-standard elliptic surface $(X_2 ^+ ,Q _2 ^+, G_2^+)$ according to the fan in theorem \[log-flip\]. According to the same theorem, the zero section $Q_2 ^+$ of $X_2 ^+$ misses the singular point where it meets $X_1 ^+ ,$ so we can iterate the process, and prune the tree of all the $j \text{-trivial}$ rational curves meeting the rest of the central fibre in only one point. The log-canonical bundle is now nef and big after Proposition \[flipextremalray\], since the family ${{\mathcal X}}\to {{\mathcal C}}$ was [*strictly prestable*]{} to begin with (so it did not have any component whose zero section had self-intersection $\geq -1$), and according to proposition \[self-int2\] it stays such. Furthermore, according to theorem \[log-flip\], this way we only produce semi-logcanonical singularities. In order to make the log-canonical bundle [*ample*]{}, we need to contract all the chains of rational curves that meet the rest of the central fibre in two ends, according to lemma \[extr ray\]. But this is taken care by theorem \[logcancont\], and will produce [*isotrivial pseudoelliptic surfaces of type II*]{}. The log-canonical bundle is now ample on accounts of proposition \[ample2\]. On accounts of theorem \[logcancont\], the singularities we thus obtain are at most semi-loganonical. This ends the proof. Stable reduction for pairs {#stablepairs} -------------------------- Here we want to deal with the absolute case. As it has been mensioned earlier, the steps of the MMP in a one parameter family are going to be similar to the ones performed in the case of triples, except that we now need to perform the flips and the small contractions also in cases in which the $j$-map is not constant. \[stabredpairs\] Let ${{{\mathcal X}}}_\eta\to {{{\mathcal C}}}_\eta\to \eta$ be a stable elliptic surface over a smooth base curve ${{\mathcal C}}_\eta$ of genus $g\geq 2 .$ Then there is a finite extension of discrete valuation rings $R\subset R'$ and a pair $( {{\mathcal X}}' \to {{\mathcal C}}' , {{\mathcal Q}}' )$ over $S'$ such that: 1. ${{\mathcal X}}' \to {{\mathcal C}}'$ gives rise to an extension: $$\begin{array}{ccc} {\mathcal X}_\eta\times_SS' &\subset & {\mathcal X}' \\ {\downarrow}& & {\downarrow}\\ {\mathcal C}_\eta \times_S S' &\subset & {\mathcal C}' \\ {\downarrow}& & {\downarrow}\\ \{\eta'\} & \subset & S', \end{array}$$ compatible with the extension ${{\mathcal Q}}_{\eta} \times _S S' \subset {{\mathcal Q}}' ;$ 2. the components of ${{\mathcal X}}'_0$ that dominate the components of ${{\mathcal C}}' _0$ are [*log-standard strictly prestable* ]{} triples $(X,Q,G, )$ where $G$ consists of either one or two fibres, which are either stable or twisted. 3. the components of ${{\mathcal X}}'_0$ that are mapped to a point of ${{\mathcal C}}' _0$ are [*log-pseudoelliptic surfaces*]{} either of type I or II. In the former case they are attached to the rest of the central fibre according to lemma \[toric one fibre\] and in the latter according to lemma \[toric two\]. The extension is unique up to a unique isomorphism, and its formation commutes with further finite extensions of discrete valuation rings. By the [*strictly prestable reduction theorem*]{} (theorem \[divcont\]) we can find a finite base change $\Delta ' \to \Delta$ , $\Delta ' \text{-schemes} $ and $\Delta ' \text{-morphisms}$ unique up to unique isomorphisms $({{\mathcal X}}\to {{\mathcal C}}, {{\mathcal Q}}, f: {{\mathcal X}}\stackrel{\pi}{\to} {{\mathcal C}}\stackrel{j}{\to} {{\overline {\mathbf{M}}}}_{1,1})$ that extend $({{{\mathcal X}}}_\eta\to {{{\mathcal C}}}_\eta ,{{\mathcal Q}}_{\eta} , f_\eta ),$ which [*strictly prestable*]{}. The extension commutes with further base changes. The log-canonical divisor $\omega _{{{\mathcal X}}' / \Delta '}({{\mathcal Q}}')$ is ample away from those components fibred over a rational curve that meet the central fibre ${{\mathcal X}}' _0$ along one or two fibres (stable or twisted). In fact, let $r: Z \to B$ be a component of ${{\mathcal X}}' _0 .$ On the one hand, if $B$ is not rational, it is obvious that $\omega _{{{\mathcal X}}' / \Delta '}({{\mathcal Q}}')\otimes {{\mathcal O}}_Z \simeq \omega _X ({{\mathcal Q}}' \mid _Z ) \otimes {{\mathcal O}}_Z (D),$ where $D$ is the dual curve, is ample, since $\omega _{Z/B} ({{\mathcal Q}}' \mid _Z)$ is relatively ample (Kollar semipositivity theorem) and $\omega _B$ is ample. On the other hand, if $B$ is rational but meets the rest of ${{\mathcal C}}' _0$ in at least three points, then $r^*\omega _B (D)$ is ample. The rest of the proof can be translated word by word from theorem \[stabred\]. It is worth noting that if the base curve ${{\mathcal C}}_\eta$ is rational or elliptic, then the log-canonical bundle is not ample: one needs to contract all the base curves. In this sense it is probably more natural, in the rational and elliptic base curve case, to consider the moduli of triples (with map to ${{\overline {\mathbf{M}}}}_{1,1}$,) at least if one wants to preserve the fibration structure. The rational base case ---------------------- Here we deal with one of the two cases left out from section \[stablepairs\]: namely the case in which the base curve of the elliptic surface ${{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta} ,$ is rational. In this case, the log-canonical bundle is not ample even on the surface ${{\mathcal X}}_{\eta}$ itself: indeed we need to contract the zero section ${{\mathcal Q}}_{\eta}, $ to make the log-canonical bundle ample. We perform the stable reduction theorem for the triple $({{\mathcal X}}_{\eta} \to {{\mathcal C}}_{\eta}, {{\mathcal Q}}_{\eta}, j_{\eta} : {{\mathcal C}}_{\eta}\to {{\overline {\mathbf{M}}}}_{1,1}).$ We may then assume to despose of a stable $\Delta$-triple $({{\mathcal X}}\to {{\mathcal C}}\to \Delta, {{\mathcal Q}}\to \Delta , j: {{\mathcal C}}\to {{\overline {\mathbf{M}}}}_{1,1})$ over some DVR scheme $\Delta$ to begin with. We want to be able to contract the base curve in the general member. We can now state and prove: \[rationalbase\] Let $({{{\mathcal X}}}_\eta\to {{{\mathcal C}}}_\eta\to \eta , {{\mathcal Q}}_{\eta}\to \eta , {{{\mathcal X}}}_\eta \to {{\overline {\mathbf{M}}}}_{1,1} )$ be a stable triple over a [*rational*]{} smooth base curve ${{\mathcal C}}_\eta \simeq {{\mathbb{P}}}^1.$ Furthermore, let ${{{\mathcal Y}}}_\eta \to {{{\mathcal X}}}_{\eta}$ be the log-canonical contraction of ${{\mathcal Q}}_{\eta} .$ Then there is a finite extension of discrete valuation rings $R\subset R'$ and a pair $( {{\mathcal X}}' , {{\mathcal Q}}' )$ over $\Delta ' =Spec(R')$ such that: 1. ${{\mathcal X}}'$ gives rise to an extension: $$\begin{array}{ccc} {\mathcal X}_\eta\times_{\Delta } \Delta ' &\subset & {\mathcal X}' \\ {\downarrow}& & {\downarrow}\\ \{\eta'\} & \subset & \Delta ', \end{array}$$ compatible with the extension ${{\mathcal Q}}_{\eta} \times _{\Delta } \Delta ' \subset {{\mathcal Q}}' ;$ 2. the components of ${{\mathcal X}}'_0$ are [*n-pseudoelliptic surfaces*]{} with either $n=0$ or $n=1$. In the former case they are attached to the rest of the central fibre according to lemma \[toric one fibre\] and in the latter according to lemma \[toric two\]. The extension is unique up to a unique isomorphism, and its formation commutes with further finite extensions of discrete valuation rings. We can apply Theorem \[stabred\] to obtain a stable [*triple*]{} $({{{\mathcal Y}}}\to {{{\mathcal C}}} \to \Delta , {{\mathcal Q}}\to \Delta , {{{\mathcal X}}} \to {{\overline {\mathbf{M}}}}_{1,1} ), $ after a possible base change (unique up to a unique isomorphism) $\Delta ' \to \Delta,$ that satisfies the analogous of property  1. Either ${{\mathcal C}}_0$ consists of a simple tree (i.e., a chain of rational curves meeting transversally each only one consecutive curve at one point), or every [*chain-like*]{} component $C \subset {{\mathcal C}}_0$ (i.e., a component that consists of a chain) will meet another [*tree-like component*]{} in a leaf which is not an extremity (i.e., this leaf will meet two more leaves of tree-like component it belongs to). In the first case there are two possibilities: the tree (i.e., ${{\mathcal C}}_0$ has either an even or an odd number of leaves. If it has an odd number of components, there is a well-defined [*central leaf*]{}, that is to say the leaf which disconnets the tree in two components of equal lenght. In case there is an even number of components, then there is, analogously, a well defined concept of [*central pair of leaves*]{}. So in the case of a simple tree of odd lenght, let $B$ be the central curve. We can prune, starting from the two ends (by log-flipping according to theorem \[log-flip\]) all the leaves that belong to the two chain-like components that $B$ disconnets from ${{\mathcal C}}_0 .$ Let as call ${{\mathcal Y}}' \to {{\mathcal C}}' \to \Delta$ the family thus obtained from ${{\mathcal Y}}\to {{\mathcal C}}\to \Delta ,$ and ${{\mathcal Q}}'$ the new zero-section. At this point the zero section ${{\mathcal Q}}' \mid _B$ of the log-standard surface $Y:= {{\mathcal Y}}' \mid _B \to B$ has the property that $L _Y \cdot Q =-2, $ which is the same as in the general fibre. We can now divisorially contract the zero section ${{\mathcal Q}}'$ in ${{\mathcal Y}}$ (e.g., by means of the line bundle $\omega _{{{\mathcal Y}}'} ((1+a) {{\mathcal Q}}')$ where $a=\frac {2}{{{{\mathcal Q}}'_{\eta} }^2}$). The log-canonical bundle is now ample, according to proposition \[flipextremalray\]. Also, on accounts of theorems \[log-flip\] and \[logcancont\] the singularities thus produced are at most semi-logcanonical. The result in the special fibre is a configuration of two chains of $0$-pseudoelliptic surfaces attached to another $0$-pseudoelliptic surface (the surface obtained from $Y$ by contracting the zero-section). Analogously, if we a simple tree of even lenght, do the same operations word by word as above, barring that now $B$ is replaced by the [*central pair*]{} $B_1 \cup B_2$. The result, now consists of two chains of $0$-pseudoelliptic surfaces as above, only now attached to a union of two $1$-pseudoelliptic surfaces (the result of contracting the zero sections of ${{\mathcal Y}}' \mid _{B_1} \cup {{\mathcal Y}}' \mid _{B_2}$). In the event that there are two chain-like components , we can now individuate a [*spine*]{}, namely the one component that is attached to the other at an extremity. In this case we first prune (by means of theorem \[log-flip\]) the other component and reduce ourselves to considering only the spine, in other words reducing the problem to the previous case. We can now conclude by a simple induction argument. The elliptic base case ---------------------- Here, at last, we deal with the final case. We first need the following: A surface $Y$ with a structure morphism $g: Y'\to Y$ is said to be a [**pseudoelliptic surface of type $E_0$**]{} (resp. [**$E_{I_N}$**]{}) if the surface $(Y' \to E,Q,F) $ is a log-standard elliptic surface, with one marked fibre $F$ and a zero section $Q,$ mapping to an irreducible elliptic curve $E$ (resp. to a closed chain of rational curves $E$) as base curve, and if $g$ has the zero section $Q$ as exceptional curve. The singularity of $Y$ at $g(Q)$ is an [*elliptic*]{} (resp. [*degenerate cusp*]{}) singularity We have: \[ellipticbase\] Let $({{{\mathcal X}}}_\eta\to {{{\mathcal C}}}_\eta\to \eta , {{\mathcal Q}}_{\eta}\to \eta , {{{\mathcal X}}}_\eta \to {{\overline {\mathbf{M}}}}_{1,1} )$ be a stable triple over an [*elliptic*]{} smooth base curve ${{\mathcal C}}_\eta \simeq E.$ Furthermore, let ${{{\mathcal Y}}}_\eta \to {{{\mathcal X}}}_{\eta}$ be the [*pseudoelliptic surface of type $E_0$* ]{} obtained by contracting ${{\mathcal Q}}_{\eta} .$ Then there is a finite extension of discrete valuation rings $R\subset R'$ and a pair $( {{\mathcal X}}' , {{\mathcal Q}}' )$ over $\Delta ' =Spec(R')$ such that: 1. ${{\mathcal X}}'$ gives rise to an extension: $$\begin{array}{ccc} {\mathcal X}_\eta\times_{\Delta } \Delta ' &\subset & {\mathcal X}' \\ {\downarrow}& & {\downarrow}\\ \{\eta'\} & \subset & \Delta ', \end{array}$$ compatible with the extension ${{\mathcal Q}}_{\eta} \times _{\Delta } \Delta ' \subset {{\mathcal Q}}' ;$ 2. the components of the central fibre ${{\mathcal X}}'_0$ consist of a [*pseudoelliptic surface*]{} either of [*type $E_0$*]{} or of type [*$E_{I_N}$*]{}, attached to a configuration of $0$-pseudolliptic and $1$-pseudoelliptic surfaces as in theorem \[rationalbase\]. The extension is unique up to a unique isomorphism, and its formation commutes with further finite extensions of discrete valuation rings. We may assume, after theroem \[rationalbase\], that we have a family ${{\mathcal Y}}\to {{\mathcal C}}' \to \Delta$ whose generic fibre is isomorphic to ${{\mathcal X}}_{\eta} \to {{{\mathcal C}}'} _{\eta},$ and whose central fibre is a a surface ${{\mathcal Y}}_0$ whose components are a log-standard elliptic surface $(Y \to E , Q+F)$ with one marked fibre $F$ and having as a base curve $E$ either an elliptic curve or a closed chain of rational curves, attached along $F$ to a configuration of $0$-pseudolliptic and $1$-pseudoelliptic surfaces as in theorem \[rationalbase\]. We can now divisorially contract the zero section ${{\mathcal Q}}'$ of ${{\mathcal Y}}\to {{\mathcal C}}'$ to obtain the result stated. [MMMM]{} D. Abramovich and A. Vistoli, [*Compactifying the space of stable maps*]{}, preprint 1999. D. Abramovich and A. Vistoli, [*Complete moduli for fibered surfaces*]{}, preprint 1997. V.Alexeev,[*Moduli spaces $M \sb {g,n}(W)$ for surfaces,*]{} in [*Higher-dimensional complex varieties (Trento, 1994)*]{}, 1–2, de Gruyter, Berlin. V.Alexeev, [*Boundedness and $K^2$ for log surfaces,*]{} Int.J.Math. [*5*]{} (1994), no. 6. A.Grothendieck, [*“ Revetements etales et Groupe Fondamental”, Lectures Notes in Math. 224 ,*]{}, Springer Verlag, New York, [**1971**]{} W.Barth, C.Peters, A.Van de Ven [*Compact Complex Surfaces*]{} Springer-Verlag 1984 P. Deligne, D. Mumford, [*The irreducibility of the moduli space of curves of given genus*]{}, Publication I.H.E.S [**36**]{} (1969). A. J. de Jong and F. Oort, [*On extending families of curves*]{}, J. Alg. Geom. [**6**]{} no. 3 (1997), 545-562. W.Fulton, [*Introduction to Toric Varieties,*]{} Annals of Mathematics Studies N.$131$, Princeton University Press J. Harris, I. Morrison, [*Moduli of curves*]{}, GTM [**187**]{}, Springer-Verlag R. Hartshorne, [*Algebraic Geometry*]{}, GTM [**52**]{}, Springer-Verlag (1977) B. Hasset, [*Stable log surfaces and limits of quartic plane curves,*]{} Manuscripta Mathematica 100 (1999), 469-497 Y. Kawamata; K. Matsuda; K. Matsuki, [*Introduction to the minimal model program*]{}, Adv. Stud. Pure Mat. 10, 283-360 (1987) J.Kollár and N. Shepherd-Barron, [*Threefolds and deformations of surface singularities,*]{} Invent. Math. [*91*]{} (1988), 299-338. J. Kollár et al., [*Flips and abundance for algebraic threefolds*]{}, Asterisque, vol. 211, 1992. R. Miranda, [*The moduli of Weiertstrass Fibrations Over ${{\mathbb{P}}}^1$*]{}, Math. Ann. [*255*]{} (1981), 379-394 R. Miranda, [*The basic theory of elliptic surfaces* ]{} Università di Pisa, Dipartimento di Matematica, Pisa 1982 S. Mori, [*Threefolds whose canonical bundles are not numerically effective*]{} Ann. Math. [*116*]{} (1982), 133-176 D. Mumford and K. Suominen, [*Introduction to the theory of moduli* ]{}, 5-th Nordic Summer School in Mathematics, Oslo, 1970 M. Reid, [*Decomposition of Toric Morphisms*]{}, volume in honor of Shafarevich. M. Reid, [*Minimal Models of Canonical Threefolds*]{}, Advanced Studies in Pure Mathematics 1, edited by S.Iitaka. M. Reid, [*Young person’s guide to canonical singularities* ]{}, Algebraic Geometry Bowdoin 1985, Proceedings of Symposia in Pure Mathematics J. Silvermann, [*Advanced Topics in the Arithmetic of Elliptic Curves*]{}, Springer-Verlag
{ "pile_set_name": "ArXiv" }
--- abstract: 'On montre que sur toute hypersurface cubique complexe de dimension au moins 2, le groupe de Chow des cycles de dimension 1 est engendré par les droites. Le cas lisse est un théorème connu. La démonstration ici donnée repose sur un résultat sur les surfaces géométriquement rationnelles sur un corps quelconque (1983), obtenu via la K-théorie algébrique.' address: | Université Paris-Sud, CNRS, Paris-Saclay\ Mathématiques, Bâtiment 425\ 91405 Orsay Cedex\ France author: - 'J.-L. Colliot-Thélène' date: 28 mars 2017 title: Droites sur les hypersurfaces cubiques --- Over any complex cubic hypersurface of dimension at least 2, the Chow group of 1-dimensional cycles is spanned by the lines lying on the hypersurface. The smooth case has already been given several other proofs. Introduction {#introduction .unnumbered} ============ Soit $X$ une variété sur un corps quelconque. On note $CH_{i}(X)$ le groupe de Chow des cycles de dimension $i$ sur $X$ modulo l’équivalence rationnelle. Le théorème suivant est connu. \[lisse\] Soient $k$ un corps algébriquement clos de caractéristique zéro et $X \subset {{\mathbf P}}^n_{k}$, avec $n \geq 3$, une hypersurface cubique lisse. Le groupe de Chow $CH_{1}(X)$ des 1-cycles de $X$ modulo équivalence rationnelle est engendré par les droites contenues dans $X$. $\Box$ Pour $n=3$, c’est un résultat classique. Pour $n= 6$, c’est établi par Paranjape [@paranjape §4]. Celui-ci utilise l’existence d’un ${{\mathbf P}}^2$ contenu dans $X \subset {{\mathbf P}}^6$ pour fibrer $X \subset {{\mathbf P}}^6$ en quadriques de dimension 2 au-dessus de ${{\mathbf P}}^3$. Paranjape écrit qu’une méthode analogue vaut pour tout $n \geq 6$. Pour tout $n \geq 4$, le théorème est établi par M. Shen [@shen Thm. 1.1] par une méthode différente de celle de Paranjape. Dans la récente prépublication, M. Shen [@shen2 Thm. 4.1 ] étend ce résultat en un théorème pour toute hypersurface cubique lisse de dimension au moins 3 sur un corps de base non nécessairement algébriquement clos, lorsque l’hypersurface cubique contient une droite définie sur ce corps. Pour $n \geq 5$, le théorème est aussi un cas particulier d’un résultat de Tian et Zong sur les intersections complètes de Fano [@TZ Thm. 6.1] de dimension $m$ et de multidegré $(d_{1},\cdots,d_{c})$ avec $d_{1}+ \cdots d_{c} \leq m-1$ (résultat obtenu par encore une autre méthode). Comme le note déjà Paranjape [@paranjape], pour $n \geq 6$, l’énoncé implique que le groupe de Chow $CH_{1}(X)$ est égal à ${{\mathbb Z}}$. En effet le schéma de Fano des droites de $X$ est alors une variété de Fano (lisse, projective, faisceau anticanonique ample), et un théorème bien connu de Campana et de Kollár-Miyaoka-Mori dit que les variétés de Fano sont rationnellement connexes (par chaînes). Dans cette note, je donne une nouvelle démonstration du théorème \[lisse\], qui vaut pour les hypersurfaces cubiques quelconques. C’est le théorème \[sing\]. Il y a deux ingrédients. Le premier ingrédient est un résultat sur les surfaces projectives lisses géométriquement rationnelles sur les corps de dimension cohomologique 1 ([@CT], théorème \[K2\] ci-dessous), dont la démonstration utilise la K-théorie algébrique (théorème de Merkur’ev et Suslin). Le second ingrédient est classique : c’est la classification des types de surfaces cubiques singulières sur un corps algébriquement clos. La démonstration procède par sections hyperplanes et récurrence sur la dimension. Même pour une hypersurface cubique lisse donnée, elle impose de considérer toutes les hypersurfaces cubiques de dimension un de moins obtenues par section hyperplane, et celles-ci peuvent être singulières. Nous n’utiliserons que les propriétés les plus simples des groupes de Chow des variétés, telles qu’on les trouve dans le chapitre 1 du livre [@fulton], en particulier la suite de localisation [@fulton Prop. 1.8] et le comportement dans une fibration en droites affines [@fulton Prop. 1.9]. Étant donnée une variété $X$ projective sur un corps $K$, la R-équivalence sur l’ensemble $X(K)$ des points $K$-rationnels de $X$ est la relation d’équivalence engendrée par la relation élémentaire suivante : deux $K$-points $A$ et $B$ sont élémentairement liés s’il existe un $K$-morphisme $f : {{\mathbf P}}^1_{K} \to X$ tel que $A$ et $B$ soient dans $f({{\mathbf P}}^1(K))\subset X(K)$. Si deux $K$-points $A$ et $B$ sont R-équivalents, alors $A-B=0 \in CH_{0}(X)$. Groupe de Chow des zéro-cycles d’une hypersurface cubique sur un corps de fonctions d’une variable ================================================================================================== Le théorème suivant est   une conséquence immédiate de [@CT Prop. 4], puisque le groupe de cohomologie galoisienne $H^1(K,S)$ pour un $K$-tore $S$ sur un corps $K$ de dimension cohomologique 1 est nul. [@CT Theorem A (iv)] \[K2\] Soit $K$ un corps de caractéristique zéro et de dimension cohomologique 1. Soit $X$ une $K$-surface projective, lisse, géométriquement rationnelle. Le noyau de l’application degré $deg_{K} : CH_{0}(X) \to {{\mathbb Z}}$ est nul. Si $X$ possède un point rationnel, par exemple si $K$ est un corps $C_{1}$, alors l’application degré $$deg_{K} : CH_{0}(X) \to {{\mathbb Z}}$$ est un isomorphisme. $\Box$ Ce théorème s’applique en particulier aux surfaces cubiques lisses. Étudions maintenant le cas des surfaces cubiques quelconques. \[surfaces\] Soit $K$ un corps de caractéristique zéro et de dimension cohomologique 1. Soit $X \subset {{\mathbf P}}^3_{K}$ une surface cubique. Supposons $X(K) \neq \emptyset$, ce qui est le cas si $K$ est $C_{1}$, par exemple si $K$ est un corps de fonctions d’une variable sur un corps algébriquement clos. Alors l’application degré $$deg_{K}: CH_{0}(X) \to {{\mathbb Z}}$$ est un isomorphisme. Comme toute surface cubique lisse sur un corps algébriquement clos est rationnelle, le cas où $X$ est lisse est un cas particulier du théorème \[K2\]. Supposons $X$ singulière. Si $X \subset {{\mathbf P}}^3_{K}$ est un cône, tout point fermé de $X$ est rationnellement équivalent à un multiple d’un point $K$-rationnel du sommet du cône (cet argument vaut sur un corps quelconque). Si $X$ n’est pas un cône, mais n’est pas géométriquement intègre, alors c’est l’union d’un plan $P$ et d’une quadrique $Q$ géométriquement intègre, leur intersection est une conique $C$ dans ${{\mathbf P}}^2_{K}$. Toute telle conique possède un point $K$-rationnel, puisque $cd(K)\leq 1$, et $deg_{K} : CH_{0}(C) \to {{\mathbb Z}}$ est un isomorphisme. Fixons $m \in C(K)$. Tout point fermé du plan $P$ est rationnellement équivalent à un multiple de $m$. Si la quadrique $Q$ est un cône de sommet $q \in Q(K)$, tout point fermé de $Q$ est rationnellement équivalent à un multiple de $q$, et $m$ est rationnellement équivalent à $q$. Si la quadrique $Q$ est lisse, alors elle est $K$-rationnelle car elle possède un $K$-point, et $deg_{K} : CH_{0}(Q) \to {{\mathbb Z}}$ est un isomorphisme (en fait $Q(K)/R=\{*\}$). On conclut que $deg_{K} : CH_{0}(X) \to {{\mathbb Z}}$ est un isomorphisme. Supposons désormais que la surface cubique $X \subset {{\mathbf P}}^3_{K}$ n’est pas un cône et est géométriquement intègre. Elle est alors géométriquement rationnelle. Les diverses singularités possibles ont été analysées depuis longtemps (Schläffli, Cayley, B. Segre, Bruce–Wall [@brucewall], Demazure, Coray–Tsfasman [@coraytsfasman]). Si les points singuliers ne sont pas isolés, alors la surface cubique $X$ contient une droite double $D \subset X$, qui est définie sur $K$. Tout $K$-point de $X$ hors de $D$ est situé sur une droite définie sur $K$ rencontrant $D$, à savoir la droite résiduelle de l’intersection avec $X$ du plan défini par $D$ et le $K$-point. On a donc $X(K)/R=\{*\}$ et $deg_{K} : CH_{0}(X) \to {{\mathbb Z}}$ est un isomorphisme. Supposons désormais de plus que les points singuliers de $X$ sont isolés. Si $X$ possède un point singulier $K$-rationnel, alors $X(K)/R=\{*\}$ [@madore2008 Lemme 1.3], sous la simple hypothèse que toute conique sur $K$ possède un point rationnel. On a donc alors $X(L)/R= \{*\}$ pour toute extension finie de corps $L/K$. Ainsi $deg_{K} : CH_{0}(X) \to {{\mathbb Z}}$ est un isomorphisme. Supposons dorénavant de plus que l’on a $X_{sing}(K)=\emptyset$. Soit $f: Y \to X$ une résolution des singularités. Un argument simple (lemme de Nishimura) montre que l’application induite $Y(K) \to X(K)$ contient les $K$-points lisses de $X$ dans son image. Donc $Y(K) \to X(K)$ est surjectif. Par hypothèse, on a $X(K)\neq \emptyset$. Soient $P$ et $Q$ deux $K$-points de $X$. Soient $M$, resp. $N$, dans $Y(K)$ d’image $P$, resp. $Q$, dans $X(K)$. La $K$-surface $Y$ est projective, lisse, géométriquement rationnelle. Le théorème \[K2\] assure $M-N =0 \in CH_{0}(Y)$. Le morphisme propre $f$ induit $f_{*} : CH_{0}(Y) \to CH_{0}(X)$. On a donc $P-Q=0 \in CH_{0}(X)$. Si $R$ est un point fermé de $X$, de corps résiduel $L=K(R)$, suivant que $X_{L}$ possède un $L$-point singulier ou non, l’un des deux arguments ci-dessus garantit $R-M_{L} = 0 \in CH_{0}(X_{L})$, et donc $deg_{K} : CH_{0}(X) \to {{\mathbb Z}}$ est un isomorphisme. \[generique\] Soit $K$ un corps de caractéristique zéro et de dimension cohomologique 1. Soient $n \geq 3$ et $X \subset {{\mathbf P}}^n_{K} $, $n \geq 3$ une hypersurface cubique. Si l’on $X(K) \neq \emptyset$, par exemple si $K$ est un corps $C_{1}$, l’application degré $deg_{K}: CH_{0}(X) \to {{\mathbb Z}}$ est un isomorphisme. Soit $O$ un point $K$-rationnel et $P$ un point fermé de $X$, de corps résiduel $L=K(P)$. Sur $X_{L} \subset {{\mathbf P}}^n_{L} $, on dispose d’un point $L$-rationnel $p$ défini par $P$ et du $L$-point $q= O_{L}$. On choisit un espace linéaire $H \subset {{\mathbf P}}^n_{L}$ de dimension 3 qui contient $p$ et $q$. Soit $Y:= X_{L} \cap H$. Si $Y=H$, alors $p$ et $q$ sont $R$-équivalents sur $X_{L}$, donc $p-q = 0 \in CH_{0}(Y)$. Si $Y \subset H$ est une surface cubique, le théorème précédent assure aussi $p-q = 0 \in CH_{0}(Y)$ et donc $p-q=0 \in CH_{0}(X_{L})$. Ainsi $P- [L:K]O = 0 \in CH_{0}(X).$ Pour tout corps $K$ qui est $C_{1}$, et tout $n \geq 5$, un argument élémentaire [@madore2008 Prop. 1.4] montre que l’on a $X(K)/R=\{*\}$ pour toute hypersurface cubique (lisse ou non), d’où il résulte immédiatement que $deg_{K}: CH_{0}(X) \to {{\mathbb Z}}$ est un isomorphisme [@madore2008 Cor. 1.6]. C’est une question ouverte si sur un tel corps $K$, et déjà sur un corps $K$ de fonctions d’une variable sur le corps des complexes, on a $X(K)/R=1$ pour toute hypersurface cubique lisse $X \subset {{\mathbf P}}^n_{K}$ pour $n=3,4$. Groupe de Chow des 1-cycles d’une hypersurface cubique sur un corps algébriquement clos ======================================================================================= \[sing\] Soit $k$ un corps algébriquement clos de caractéristique zéro. Soit $X \subset {{\mathbf P}}^n_{k}$, avec $n \geq 3$ une hypersurface cubique. Le groupe de Chow $CH_{1}(X)$ est engendré par les droites contenues dans $X$. On va établir cet énoncé par récurrence sur $n \geq 3$. On commence par établir le cas $n=3$ par une discussion cas par cas. Dans un plan ${{\mathbf P}}^2$ tout 1-cycle est rationnellement équivalent à un multiple d’une droite. Pour une quadrique $Q \subset {{\mathbf P}}^3$ non singulière, le groupe de Picard de $Q$ est engendré par les deux classes de génératrices. Si $Y \subset {{\mathbf P}}^3$ de coordonnées $(x,y,z,t)$ est un cône défini par une équation $f(x,y,z)=0$, et de sommet $p$ de coordonnées $(0,0,0,1)$, $CH_{1}(Y)=CH_{1}(Y\setminus p)$ est engendré par les génératrices du cône. Ceci établit le résultat dans le cas où la surface cubique n’est pas intègre, et aussi dans le cas où c’est un cône. Supposons donc $X$ intègre et non conique. Si les singularités de $X$ ne sont pas isolées, alors $X$ possède une droite double. On peut alors [@brucewall §2, Case E] trouver des coordonnées homogènes $(x,y,z,t)$ de ${{\mathbf P}}^3$ telles que la surface soit donnée soit par l’équation $$x^2z+y^2t=0$$ soit par l’équation $$x^2z+xyt+y^3=0.$$ Dans le premier cas, le complémentaire des deux droites $x=y=0$ et $x=t=0$, découpées par $x=0$, est isomorphe au plan affine ${{\mathbf A}}^2$ de coordonnées $(y,t)$. Dans le second cas, le complémentaire de la droite $x=y=0$ découpée par $x=0$ est isomorphe au plan affine ${{\mathbf A}}^2$ de coordonnées $(y,t)$. Comme on a $CH_{1}({{\mathbf A}}^2)=0$, ceci établit que $CH_{1}(X)$ est engendré par des droites de $X$. Sinon, $X$ est normale, et si $f: X' \to X$ est sa désingularisation minimale, alors $X'$ est une surface de del Pezzo généralisée de degré 3, et les “droites” de $X'$ sont les transformées propres des vraies droites de $X$. Voir là-dessus [@coraytsfasman Exemple 0.5]. La projection $CH_{1}(X') \to CH_{1}(X)$ est clairement surjective, et le groupe $CH_{1}(X')=Pic(X')$ est engendré par les “droites” de $X'$ (courbes $D$ lisses de genre zéro avec $(D.D)=-1$ et les “racines irréductibles” (courbes lisses de genre zéro avec $(D.D)=-2$) qui sont des courbes contractées par $f$ sur les points singuliers de $X$. Donc $ CH_{1}(X)$ est engendré par les vraies droites de $X \subset {{\mathbf P}}^3$. Soit $n\geq 4$. Supposons le cas $n-1$ établi. Soit $X \subset {{\mathbf P}}^{n}$ une hypersurface cubique. On trouve dans $X$ une droite $D$ (il en existe sur toute surface cubique sur $k$ algébriquement clos) et on choisit $Q\simeq {{\mathbf P}}^{n-2} \subset {{\mathbf P}}^{n}$ un espace linéaire de dimension $n-2$ qui ne rencontre pas $D$ et qui n’est pas contenu dans $X$. On considère le pinceau des espaces linéaires ${{\mathbf P}}^{n-1} \subset {{\mathbf P}}^n$ qui contiennent $Q$. On trouve ainsi une variété $Y\subset X \times {{\mathbf P}}^1$ munie d’un morphisme propre $Y \to X$ et d’une fibration $Y \to {{\mathbf P}}^1$ dont les fibres au-dessus de $k$-points $s \in {{\mathbf P}}^1(k)$ sont des hypersurfaces cubiques $Y_{s} \subset {{\mathbf P}}^n_{k}$ sections hyperplanes de $X \subset {{\mathbf P}}^{n}_{k}$ (l’hypothèse que $X$ ne contient pas $Q$ garantit qu’aucun $Y_{s}$ n’est égal à ${{\mathbf P}}^n_{k}$) et dont la fibre générique est une hypersurface cubique $Y_{\eta} \subset {{\mathbf P}}^{n-1}_{K}$, avec $K=k({{\mathbf P}}^1)$. La droite $D$ définit une section de la fibration $Y \to {{\mathbf P}}^1$, soit une courbe $M \subset Y$, dont l’image se restreint en un $K$-point rationnel de $Y_{\eta}$. On dispose de la suite exacte $$\oplus_{s \in {{\mathbf P}}^1(k)} CH_{1}(Y_{s}) \to CH_{1}(Y) \to CH_{0}(Y_{\eta}) \to 0.$$ D’après le théorème \[generique\], la classe de $M$ dans $CH_{1}(Y)$ s’envoie sur un générateur de $CH_{0}(Y_{\eta})\simeq{{\mathbb Z}}$. L’application $CH_{1}(Y) \to CH_{1}(X)$ est surjective. En effet le morphisme $Y \to X$ induit un isomorphisme au-dessus du complémentaire du fermé propre $X \cap Q \subset Q$, et au-dessus de chaque point de $X \cap Q$, la fibre est une droite projective. L’image de $M$ est la droite $D$ de $X$, chaque groupe $CH_{1}(Y_{s})$ est par hypothèse de récurrence engendré par des droites de $Y_{s}$, dont les images dans $X$ sont des droites de $X$. [99]{} J. W. Bruce et C.T.C. Wall, On the classification of cubic surfaces, J. London Math. Soc. (2) [**19**]{} (1979) 245–256. J.-L. Colliot-Thélène, Hilbert’s Theorem 90 for $K_{2}$, with application to the Chow groups of rational surfaces, Invent. math. [**71**]{} (1983) 1–20. D. Coray et M. Tsfasman, Arithmetic on singular del Pezzo surfaces, Proc. Lond. Math. Soc. (3), [**57**]{}, 25–87 (1988). W. Fulton, Intersection Theory, Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, Bd. [**2**]{}, Springer-Verlag (1984). D. Madore, Équivalence rationnelle sur les hypersurfaces cubiques de mauvaise réduction, J. Number Theory [**128**]{} (2008), no. 4, 926–944. K. Paranjape, Cohomological and cycle-theoretic connectivity, Annals of Math. [**140**]{} (1994) 641-660. Mingmin Shen, On relations among 1-cycles on cubic hypersurfaces, J. Alg. Geometry [**23**]{} (2014) 539-569. Mingmin Shen, Rationality, universal generation and the integral Hodge conjecture, https://arxiv.org/abs/1602.07331v2 Zhiyu Tian et Hong R. Zong, One-cycles on rationally connected varieties. Compositio. Math. [**150**]{}, no. 3 (2014) 396–408.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The dependence of spin relaxation on the direction of the quantum wire under Rashba and Dresselhaus (linear and cubic) spin orbit coupling is studied. Comprising the dimensional reduction of the wire in the diffusive regime, the lowest spin relaxation and dephasing rates for (001) and (110) systems are found. The analysis of spin relaxation reduction is then extended to non-diffusive wires where it is shown that, in contrast to the theory of dimensional crossover from weak localization to weak antilocalization in diffusive wires, the relaxation due to cubic Dresselhaus spin orbit coupling is reduced and the linear part shifted with the number of transverse channels.' author: - 'P. Wenk' - 'S. Kettemann' bibliography: - 'WenkKettemann2010.bib' title: Direction Dependence of Spin Relaxation in Confined 2D Systems --- Introduction ============ Spin dynamics in semiconductors have been studied for decades, but still the prime condition for building spintronic devices, namely the understanding of spin relaxation, is not satisfactorily fulfilled. In the following we focus on materials where the dominant mechanism for spin relaxation is governed by the D’yakonov-Perel spin relaxation (DPR).[@perel] This mechanism results from lifting the spin degeneracy which is due to time inversion symmetry and spacial inversion symmetry and leads to the effect of slower spin dephasing the faster the momentum relaxes (motional narrowing[@Bloembergen1948; @Slichter1989; @wenkbook]). In one of the most studied systems GaAs/AlGaAs DPR is the most relevant mechanism in the metallic regime.[@dzhioev]\ Preserving time reversion symmetry, the spin splitting can be due to bulk inversion asymmetry (BIA)[@dresselhaus] and also due to the asymmetry arising from the structure of the quantum well (QW), the structure inversion asymmetry (SIA)[@rashba_1]. In Ref. it was shown how the spin relaxes in a quasi 1D electron system in a QW grown in the $[001]$ direction, depending on the width of the wire, where the normal of the boundary was pointing in the $[010]$ direction. It is already known that in a (001) 2D system with BIA and SIA we get an anisotropic spin-relaxation.[@Kainz2003; @Cheng2007; @Wu2010] This has also been studied numerically in quasi-1D GaAs wires[@Liu2009]. In this work, Sec.\[sec:001grothDir\], we present analytical results concerning this anisotropy for the 2D case as well as the case of QW with spin and charge conserving boundaries.\ We also extend our analysis to other growth directions, Sec.\[sec:110system\]: Searching for long spin decoherence times at room temperature, the (110) QW attracted attention.[@Adachi200136; @Dohrmann2004] The properties of spin relaxation in systems with this growth direction has also been related to weak localization (WL) measurements, Ref.. We present analytical explanations for dimensional spin relaxation reduction and discuss the crossover from WL to weak antilocalization (WAL), Sec.\[sec:WL\].\ As we will show in the following sections, the cubic Dresselhaus spin orbit coupling (SOC) gives always rise to a limitation of the spin relaxation time in the diffusive case, $W\gg l_e$, with the wire width W and the elastic mean free path $l_e$. However some of the experiments are done on ballistic wires and we need to modify the theory used in Ref. to enable us to study the crossover from diffusive to ballistic wires. In Sec.\[sec:DiffBallis\] we show how the spin relaxation which is due to cubic Dresselhaus SOC reduces with the number of channels in the QW.\ We consider the following Hamiltonian with SOC $$\label{hamiltonian} H = \frac{1}{2 m_e} ({\bf p} +e {\bf A} )^2 +V({\bf x}) -\frac{1}{2} \gamma\bsigma\left({\bf B}+{\bf B}_{\text{\tiny{SO}}}({\bf p})\right),$$ where $m_e$ is the effective electron mass. ${\bf A}$ is the vector potential due to the external magnetic field ${\bf B}$. ${\bf B}_{\text{\tiny{SO}}}^T = ({B_{\text{\tiny{SO}}}}_x, {B_{\text{\tiny{SO}}}}_y) $ is the momentum dependent SO field. $\bsigma$ is a vector, with components $\bsigma_i$, $i =x,y,z$, the Pauli matrices, $\gamma$ is the gyromagnetic ratio with $\gamma=g \mu_{\text{\tiny{B}}}$ with the effective g factor of the material, and $\mu_{\text{\tiny{B}}} = e/2m_e$ is the Bohr magneton constant. For example, III-V and II-VI semiconductors such as GaAs, InSb have zinc-blend structure. This BIA causes a SO interaction, which, to lowest order in the wave vector $\bf k$, is given by [@dresselhaus] $$-\frac{1}{2}\gamma {\bf B}_{\text{\tiny{SO,D}}}= \gamma_D\sum_i\hat{e}_i p_i(p_{i+1}^2-p_{i+2}^2)\label{Dresselhaus}$$ where the principal crystal axes are given by $i\in\{x,y,z\},i\rightarrow ((i-1) \mod 3)+1$ and the spin-orbit coefficient for the bulk semiconductor $\gamma_D$. We consider the standard white-noise model for the impurity potential, $V({\bf x})$, which vanishes on average $\langle V ({\bf x}) \rangle=0$, is uncorrelated, $\langle V({\bf x}) V({\bf x'})\rangle = \delta ({\bf x-x'})/2 \pi \nu \tau$, and weak, $\epsilon_F\tau\gg 1$. Here, $\nu = m_e/(2 \pi)$ is the average density of states per spin channel, $\epsilon_F$ is the Fermi energy and $\tau$ is the elastic scattering time. To address both, the WL corrections as well as the spin relaxation rates in the system, we analyze the Cooperon[@HIKAMI1980] $$\begin{aligned} \label{Cooperon0} {}&\hat{C}({\bf Q})^{-1}= \frac{1}{\tau} \left( 1 - \int \frac{d \varphi}{2\pi} \right. \nonumber\\ {}&\left.\frac{1}{1+{\mathit{i}}\tau(\mathbf{v}(\mathbf{Q}+2e\mathbf{A}+2m_e\hat a \mathbf{S}) + H_{\sigma'} + H_Z )}\right),\end{aligned}$$ where the integral is performed over all angles of velocity $\mathbf v$ on the Fermi surface, $ H_{\sigma'}=-(\mathbf{Q}+2 e\mathbf{A})\hat{a} \bsigma^\prime$ and the Zeeman coupling to the external magnetic field yields, $$H_Z=- \frac{1}{2} \gamma (\bsigma^\prime-\bsigma)\mathbf{B}.\label{zeemanTerm}$$ The coupling between the orbital motion and the spin ${\mathbf S=(\bsigma+\bsigma^\prime)/2}$ is described by the SOC operator $\hat a$. The spin quantum number is 1 instead of $1/2$ due to the electron-hole excitation. It follows that for weak disorder and without Zeeman coupling, the Cooperon depends only on the total momentum $\mathbf{Q}$ and the total spin $ {\bf S}$. Expanding the Cooperon to second order in $( {\bf Q} +2 e {\bf A} + 2m_e \hat{a} {\bf S} )$ and performing the angular integral which is for 2D diffusion (elastic mean free path $l_e$ smaller than wire width $W$) continuous from $0$ to $2 \pi$, yields: $$\hat{C} ({\bf Q})= \frac{1}{D_e( {\bf Q} + 2 e {\bf A} + 2 e {\bf A}_{\bf S})^2 + H_{\gamma_D} }.\label{Cooperon1}$$ The effective vector potential due to SO interaction is ${\bf A}_{\bf S} = m_e \hat{\alpha} {\bf S}/e$, where $\hat{\alpha} = \langle\hat{a}\rangle$ is averaged over angle. The SO term $H_{\gamma_D}$, which cannot be rewritten as a vector potential, is in our case due to the appearance of cubic Dresselhaus SOC. Example ------- To get an idea of the procedure we recall the situation presented in Ref.. Spin relaxation in a (001) quasi-1D wire in \[100\] direction:\ The Dresselhaus term, Eq.(\[Dresselhaus\]), for QWs grown in the $[001]$ direction is given by [@dresselhaus] $$-\frac{1}{2} \gamma{\bf B}_{\text{\tiny{SO,D}}} = \alpha_1 ( -\hat{e}_x p_x + \hat{e}_y p_y) + \gamma_D ( \hat{e}_x p_x p_y^2 - \hat{e}_y p_y p_x^2).$$ Here, $\alpha_1 = \gamma_D \langle p_z^2 \rangle$ is the linear Dresselhaus parameter, which measures the strength of the term linear in momenta $p_x, p_y$ in the plane of the 2D electron system (2DES). When $\langle p_z^2 \rangle \sim 1/a^2 \ge k_F^2$ ($a$ is the thickness of the 2DES, $k_F$, Fermi wave number), that term exceeds the cubic Dresselhaus terms which have coupling strength $\gamma_D$. Asymmetric confinement of the 2DES, a SIA, yields the Rashba term which does not depend on the growth direction $$-\frac{1}{2} \gamma{\bf B}_{\text{\tiny{SO,R}}} = \alpha_2 ( \hat{e}_x p_y - \hat{e}_y p_x),$$ with $\alpha_2$ the Rashba parameter.[@rashba_1; @rashba_2] Therefore the Cooperon Hamiltonian, in the case of Rashba and lin. and cubic Dresselhaus SOC is given by $$\label{H001} H_c:=\frac{\hat C^{-1}}{D_e}=({\bf Q} +2 e {\bf A}_{\mathbf{S}} )^2+(m_e^2\epsilon_F\gamma_D)^2(S_x^2+S_y^2),$$ with the effective vector potential $${\bf A}_{\mathbf{S}} = \frac{m_e}{e}\hat\alpha{\bf S}= \frac{m_e}{e}\left( \begin{array}{ccc} -\tilde\alpha_1 & -\alpha_2 & 0 \\ \alpha_2 & \tilde\alpha_1 & 0 \end{array} \right) \left(\begin{array}{c} S_x\\ S_y\\ S_z \end{array}\right),$$ with $\tilde\alpha_1=\alpha_1-m_e\gamma_D\epsilon_F/2$.\ It can be easily shown that the Hamiltonian Eq.(\[H001\]) has only non vanishing eigenvalues due to $(m_e^2\epsilon_F\gamma_D)^2$ in the 2D case.\ The term with $(S_x^2+S_y^2)$, which is due to cubic Dresselhaus SOC, is not reduced by reason of the boundary in the diffusive case. However two triplet eigenvalues of this term depend on the wire width, $$\begin{aligned} E_{QD1}={}&\frac{q_{s3}^2}{2},\\ E_{QD2,3}={}&\frac{q_{s3}^2}{2}\left(\frac{3}{2}\pm\frac{\sin(Q_{\text{\tiny{SO}}}W)}{2Q_{\text{\tiny{SO}}}W}\right),\end{aligned}$$ with $q_{s3}^2/2=(m_e^2\epsilon_F\gamma_D)^2$. In the following we are going to diagonalize the whole Hamiltonian and change the direction of the wire in the (001) plane. Spin Relaxation anisotropy in the (001) system {#sec:001grothDir} ============================================== 2D system {#sec:001grothDir2D} --------- We rotate the system in-plane through the angle $\theta$ (the angle $\theta=\pi/4$ is equivalent to \[110\]). This does not effect the Rashba term but changes the Dresselhaus one to[@Cheng2007; @Wu2010] $$\begin{aligned} {}&\frac{1}{\gamma_D}H_{D[001]}=\nonumber\\ {}&\sigma_y k_y\cos(2\theta)(\langle k_z^2 \rangle-k_x^2)-\sigma_x k_x\cos(2\theta)(\langle k_z^2 \rangle-k_y^2)\nonumber\\ {}&-\sigma_y k_x\frac{1}{2}\sin(2\theta)(k_x^2-k_y^2-2\langle k_z^2 \rangle)\nonumber\\ {}&+\sigma_x k_y\frac{1}{2}\sin(2\theta)(k_x^2-k_y^2+2\langle k_z^2 \rangle),\end{aligned}$$ with the wave vectors $k_i$. The resulting Cooperon Hamiltonian, including Rashba and Dresselhaus SOC, reads then $$\begin{aligned} \label{HrotatedWithDresselhaus} H_c={}&(Q_x+\alpha_{x1}S_x+(\alpha_{x2}-q_2)S_y)^2\nonumber\\ {}&+(Q_y+(\alpha_{x2}+q_2)S_x-\alpha_{x1}S_y)^2\nonumber\\ {}&+\frac{q_{s3}^2}{2}(S_x^2+S_y^2),\end{aligned}$$ where we set $$\begin{aligned} \frac{q_{s3}^2}{2}= {}&\left(m_e^2\epsilon_F\gamma_D\right)^2,\\ \alpha_{x1}={}&\frac{1}{2}m_e\gamma_D\cos(2\theta)((m_ev)^2-4\langle k_z^2 \rangle),\\ \alpha_{x2}={}&-\frac{1}{2}m_e\gamma_D\sin(2\theta)((m_ev)^2-4\langle k_z^2 \rangle)\\ ={}&\left(q_1-\sqrt{\frac{q_{s3}^2}{2}}\right)\sin(2\theta)\\ ={}&2m_e\tilde\alpha_1 \sin(2\theta),\end{aligned}$$ with $q_1=2m_e\alpha_1$, $q_2=2m_e\alpha_2$. We see that the part of the Hamiltonian which cannot be written as a vector field and is due to cubic Dresselhaus SOC does not depend on the wire direction in the (001) plane. ### Special case: Only lin. Dresselhaus SOC equal to Rashba SOC As a special example for the 2D case we set $q_{s3}=0$ and $q_{1}=q_2$. To simplify the search for vanishing spin relaxation we go to polar coordinates. Applying free wave functions (with $k_x,k_y$) to $H_c$, Eq.(\[HrotatedWithDresselhaus\]), we end up with (singlet part left out) $$\frac{H_c}{q_2^2}=\left( \begin{array}{ccc} 2+Q^2 & f_{\theta\phi} & -2{\mathit{i}}\exp(2{\mathit{i}}\theta) \\ & 4+Q^2 & f_{\theta\phi} \\ \text{c.c.} & & 2+Q^2 \end{array} \right)$$ with $k_x/q_2=Q\cos(\phi)$, $k_y/q_2=Q\sin(\phi)$ and $$f_{\theta\phi}=({\mathit{i}}-1)\sqrt{2}\exp({\mathit{i}}\theta)(\cos(\phi+\theta)-\sin(\phi+\theta))Q.$$ Vanishing spin relaxation is found at $Q=0$ for arbitrary values of $\theta$ (the spin with vanishing spin relaxation is pointing along the \[110\] direction[@Schliemann2003]). Another solution is found at $Q=2$ with the condition $\theta+\phi=3\pi/4$, which is equivalent to the $[\overline{1}10]$ crystallographic direction.[@Cheng2007] Quasi-1D wire {#sec:001grothDirWire} ------------- In the following we consider spin and charge conserving boundaries. Due to the SOC we have the following modified Neumann condition[@Wenk2010] $$\begin{aligned} \label{bc} \left(-\frac{\tau}{D_e}{\bf n}\cdot\langle{\bf v}_F [\gamma{\bf B}_{\text{\tiny{SO}}}({\bf k})\cdot{\bf S}]\rangle-{\mathit{i}}\partial_{\bf n}\right) C |_{\partial S} ={}& 0,\\ \intertext{where $\langle ... \rangle$ denotes the average over the direction of ${\bf v}_F$ and ${\bf k}$ which we rewrite for the rotated x-y system} (-{\mathit{i}}\partial_y + 2 e {\bf (A_{S})}_y ) C\left( x, y = \pm \frac{W}{2}\right) ={}& 0,\enspace\forall x,\quad\quad\end{aligned}$$ where ${\bf n}$ is the unit vector normal to the boundary $\partial S$ and x is the coordinate along the wire. In order to do a diagonalization taking only the zero-mode into account, we have to simplify the boundary condition. A transformation acting in the transverse direction is needed according to Eq.(\[HrotatedWithDresselhaus\]): $\hat{C}\rightarrow \tilde{\hat{C}} = U_A^{\phantom{\dagger}} \hat{C}U_A^\dagger$, by using the transformation $$\ U={\mathbbm{1}}_4-{\mathit{i}}\sin(q_s y)\frac{1}{q_s}A_y+(\cos(q_sy)-1)\frac{1}{q_s^2}A_y^2$$ with $A_y=(\alpha_{x2}+q_2)S_x-\alpha_{x1}S_y$ and$q_s=\sqrt{(\alpha_{x2}+q_2)^2+\alpha_{x1}^2}$. ### Spin relaxation We diagonalize the Hamiltonian, Eq.(\[HrotatedWithDresselhaus\]), after applying the transformation U, taking only the lowest mode into account. The spectrum of the Hamiltonian for small wire width, $W q_s<1$, is given by $$\begin{aligned} {}&E_{1/2}(k_x>0)=\nonumber\\ {}&k_x^2\pm k_x \left(2 q_{sm}-\frac{\left(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2\right)^2}{12 q_{sm}}W^2 \right)\nonumber\\ {}&+\frac{3}{2}\frac{q_{s3}^2}{2}+q_{sm}^2\mp \frac{q_{s3}^2}{2k_x}\frac{\left(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2\right)^2W^2}{96 q_{sm}}\nonumber\\ {}&-\frac{\left(\frac{q_{s3}^2}{2}+q_{sm}^2\right) \left(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2\right)^2}{24 q_{sm}^2}W^2\,\label{E1/2},\end{aligned}$$ $$\begin{aligned} {}&E_{1}(k_x=0)=q_{s3}^2+q_{sm}^2\nonumber\\ {}&-\frac{(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2)^2+\frac{q{s3}^2}{2}q_s^2}{12}W^2,\label{E1kx0}\\ {}&E_{2}(k_x=0)=\frac{q_{s3}^2}{2}+q_{sm}^2+\frac{q_{s3}^2}{2}\frac{q_2^2\alpha_{x1}^2}{3q_{sm}^2}W^2,\\ {}&E_3=k_x^2+\frac{q_{s3}^2}{2}+\frac{\left(\frac{q_{s3}^2}{2}+q_{sm}^2\right) \left(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2\right)^2}{12 q_{sm}^2}W^2,\end{aligned}$$ with $q_{sm}=\sqrt{(\alpha_{x2}-q_2)^2+\alpha_{x1}^2}$. First we notice that the only $\theta$ dependence is in the term $q_{sm}$, which disappears if the Dresselhaus SOC strength $\tilde\alpha_1$, which is shifted due to the cubic term, equals the Rashba SOC strength $\alpha_2$ and the angle of the boundary is $\theta=(1/4+n)\pi,\enspace n\in \mathbbm{Z}$. Assuming the term proportional to $W^2/k_x$ to be small, the absolute minimum can be found at $$\begin{aligned} {}&E_{1/2,min}=\frac{3}{2}\frac{q_{s3}^2}{2}+\frac{\left(q_{sm}^2-\frac{q_{s3}^2}{2}\right) \left(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2\right)^2}{24 q_{sm}^2}W^2\nonumber\\\end{aligned}$$ which is independent of the width $W$ if $\alpha_{x1}(\theta=0)=-q_2$ and/or the direction of the wire is pointing in $$\theta=\frac{1}{2}\arcsin\left(\frac{2\langle k_z^2\rangle (m_e\gamma_D)^2((m_ev)^2-2\langle k_z^2\rangle)-q_2^2}{\left(m_e^3v^2\gamma_D-4\langle k_z^2\rangle m_e\gamma_D\right)q_2}\right).$$ The second possible absolute minimum, which dominates for sufficient small width W and $q_{sm}\neq 0$ (compare with $E_2(k_x=0)$), is found at $$\begin{aligned} \label{E2min} {}&E_{3,min}=\frac{q_{s3}^2}{2}+\frac{\left(\frac{q_{s3}^2}{2}+q_{sm}^2\right) \left(\alpha_{x1}^2+\alpha_{x2}^2-q_2^2\right)^2}{12 q_{sm}^2}W^2.\nonumber\\\end{aligned}$$ The minimal spin-relaxation rate is found by analyzing the prefactor of $W^2$ in Eq.(\[E2min\]), Fig.(\[plot:theta\]). We see immediately that in the case of vanishing cubic Dresselhaus or in the case where $\alpha_{x1}(\theta=0)=-q_2$ we have no direction dependence of the minimal spin relaxation. Notice the shift of the absolute minimum away from $q_1=q_2$ due to $q_{s3}\neq 0$. In the case of $q_1<(q_{s3}/\sqrt{2})$ we find the minimum at $\theta=(1/4+n)\pi,\enspace n\in \mathbbm{Z}$, else at $\theta=(3/4+n)\pi,\enspace n\in \mathbbm{Z}$, which is indicated by the dashed line in Fig.(\[plot:theta\]). ![(Color online) Dependence of the $W^2$ coefficient in Eq.(\[E2min\]) on the lateral rotation $(\theta)$. The absolute minimum is found for $\alpha_{x1}(\theta=0)=-q_2$ (here: $q_1/q_2=1.63$) and for different SO strength we find the minimum at $\theta=(1/4+n)\pi,\enspace n\in \mathbbm{Z}$ if $q_1<(qs3/\sqrt{2})$ (dashed line: $q_1=(qs3/\sqrt{2})$) and at $\theta=(3/4+n)\pi,\enspace n\in \mathbbm{Z}$ else. Here we set $q_{s3}=0.9$. The scaling is arbitrary.[]{data-label="plot:theta"}](theta001.eps "fig:"){width="\columnwidth"}\ ### Spin dephasing Concerning spintronic devices it is interesting to know how an ensemble of spins initially oriented along the $[001]$ direction dephases in the wire of different orientation $\theta$. To do this analysis we only have to know that the eigenvector for the eigenvalue $E_1$ at $k_x=0$, Eq.(\[E1kx0\]), is the triplet state ${|S=1;m=0\rangle} = ( {|\uparrow \downarrow\rangle} + {|\downarrow \uparrow\rangle})/\sqrt{2} \equiv\linebreak {|\rightrightarrows\rangle}\hat=(0,1,0)^T$. This is equal to the z-component of the spin density whose evolution is described by the spin diffusion equation.[@Wenk2010] As an example we assume the case where cubic Dresselhaus term can be neglected and where the Rashba and lin. Dresselhaus SOC are equal. We notice that the dephasing is than width independent. At definite angles the dephasing time diverges - as for the in-plane polarized states with eigenvalue $E_2(k_x=0)$ - , $$\frac{1}{\tau_s(W)}=2D_eq_2^2(1-\sin(2\theta))$$ which is plotted in Fig.(\[plot:SpinDepWithMonteCarlo\]). We have longest spin dephasing time at $\theta=(1/4+n)\pi,\enspace n\in \mathbbm{Z}$. For $\theta=(3/4+n)\pi,\enspace n\in \mathbbm{Z}$ we get the 2D result $T_2=1/(4q_2^2D_e)$, which is given by the eigenvalue of the spin relaxation tensor [@dyakonov72_2; @dyakonov72_3; @Wenk2010], $$\label{dp} \frac{1}{\tau_{sij}} = \tau\gamma^2\left( \langle {\bf B}_{\rm SO}({\bf k})^2 \rangle \delta_{i j }-\langle { B}_{\rm SO}({\bf k})_i { B}_{\rm SO}({\bf k})_j\rangle\right)$$ to the triplet state ${|S=1;m=0\rangle}$.\ This gives an analytical description of numerical calculation done by J.Liu *et al.*, Ref..\ Switching on cubic Dresselhaus SOC leads to finite spin dephasing time for all angles $\theta$. In addition $T_2$ is than width dependent. In the case of strong cubic Dresselhaus SOC where $q_{s3}^2/2=q_1^2=q_2^2$, the dephasing time $T_2$ is angle independent and for $q_{s3}^2/2>q_1^2=q_2^2$ the minima in $T_2(\theta)$ change to maxima and vice versa. ![(Color online) The spin dephasing time $T_2$ of a spin initially oriented along the $[001]$ direction in units of $(D_eq_2^2)$ for the special case of equal Rashba and lin. Dresselhaus SOC. The different curves show different strength of cubic Dresselhaus in units of $q_{s3}/q_2$. In the case of finite cubic Dresselhaus SOC we set $W=0.4/q_2$. If $q_{s3}=0$: $T_2$ diverges at $\theta=(1/4+n)\pi,\enspace n\in \mathbbm{Z}$ (dashed vertical lines). The horizontal dashed line indicated the 2D spin dephasing time, $T_2=1/(4q_2^2D_e)$.[]{data-label="plot:SpinDepWithMonteCarlo"}](SpinDepWithMonteCarlo.eps "fig:"){width="\columnwidth"}\ ### Special case: $\theta=0$ In this case the longitudinal direction of the wire is \[100\].If we neglect the term proportional to $W^2/k_x$ in Eq.(\[E1/2\]) the lowest spin relaxation is found to be $$\frac{1}{D_e\tau_s}=\frac{q_{s3}^2}{2}+\frac{\left(\alpha_{x1}^2-q_{2}^2\right)^2 \left(q_s^2+\frac{q_{s3}^2}{2}\right) W^2}{12 q_s^2}$$ or $$\frac{1}{D_e\tau_s}=\frac{3 q_{s3}^2}{4}+\frac{\left(\alpha_{x1}^2-q_{2}^2\right)^2 \left(q_s^2-\frac{q_{s3}^2}{2}\right) W^2}{24 q_s^2},$$ depending whether $$-\frac{q_{s3}^2}{4}+\frac{\left(\alpha_{x1}^2-q_{2}^2\right)^2 \left(q_s^2+3 \frac{q_{s3}^2}{2}\right) W^2}{24 q_s^2}$$ is negative or positive. This shows that the cubic Dresselhaus term adds not only to the relaxation rate by a constant term but is also width dependent. However, this width dependence does not reduce the spin relaxation rate below $q_{s3}^2/2$. Spin relaxation in quasi-1D wire with \[110\] growth direction {#sec:110system} ============================================================== To get the spin-relaxation in a $[110]$ QW with Rashba and Dresselhaus SOC again we have to rotate the spacial coordinate system of the Dresselhaus Hamiltonian Eq.(\[Dresselhaus\]) but now with the rotation matrix $$R=\left( \begin{array}{lll} \frac{1}{\sqrt{2}} & 0 & \frac{1}{\sqrt{2}} \\ -\frac{1}{\sqrt{2}} & 0 & \frac{1}{\sqrt{2}} \\ 0 & 1 & 0 \end{array} \right).$$ We get $$\begin{aligned} \frac{H_{D[110]}}{\gamma_D} ={}&\sigma_x(-k_x^2k_z-2k_y^2k_z+k_z^3)\nonumber\\ {}&+\sigma_y(4k_xk_yk_z)\nonumber\\ {}&+\sigma_z(k_x^3-2k_xk_y^2-k_xk_z^2).\end{aligned}$$ The confinement in z-direction ($z\equiv$\[110\]) leads to $\langle k_z\rangle=\langle k_z^3\rangle=0$, and $\langle k_z^2\rangle=\int |\nabla\phi|^2dz$. The Hamiltonian for the QW in $[110]$ direction has then the following form[@Hassenkam1997] $$H_{[110]}=-\gamma_D\sigma_zk_x\left(\frac{1}{2}\langle k_z^2\rangle-\frac{1}{2}(k_x^2-2k_y^2)\right).$$ Including the Rashba SOC ($q_2$), noting that its Hamiltonian does not depend on the orientation of the wire,[@Hassenkam1997] we end up with the following Cooperon Hamiltonian $$\frac{C^{-1}}{D_e}=\left(Q_x-\tilde q_1 S_z-q_2S_y \right)^2+(Q_y+q_2S_x)^2+ \frac{\tilde q_{3}^2}{2} S_z^2.$$ with $\tilde q_1 = 2 m_e \frac{\gamma_D}{2} \langle k_z^2\rangle-\frac{\gamma_D}{2}\frac{m_e\epsilon_F}{2}$, $q_2 = 2 m_e \alpha_2$ and $\tilde q_3=(3m_e\epsilon_F^2(\gamma_D/2))$. We see immediately that in the 2D case states polarized in the z-direction have vanishing spin relaxation as long as we have no Rashba SOC. Compared with the (001) system the constant term due to cubic Dresselhaus does not mix spin directions. Here we set the appropriate Neumann boundary condition as follows: $$(-i\partial_y + 2 m_e \alpha_2 S_x) C\left(x,y = \pm\frac{W}{2}\right) =0,\enspace\forall x.\label{110boundaryCond}$$ The presence of Rashba SOC adds a vector potential proportional to $S_x$. Applying a non-abelian gauge transformation as before to simplify the boundary condition, we diagonalize the transformed Hamiltonian (App.(\[hamiltonian\])) up to second order in $q_2W$ in the 0-mode approximation. Special case: without cubic Dresselhaus SOC ------------------------------------------- The spectrum is found to be $$\begin{aligned} E_1={}&k_x^2+\frac{1}{12} \Delta^2 (q_2 W)^2,\label{110E1}\\ E_{2,3}={}&k_x^2+\frac{1}{24}\Delta^2\left(24-( q_2 W)^2\right)\nonumber\\ {}&\pm\frac{\Delta}{24}\sqrt{\Delta^2(q_2W)^4+4k_x^2(24-(q_2W)^2)^2},\end{aligned}$$ with the lowest spin relaxation rate found at finite wave vectors $k_{x_{\text{\tiny{min}}}}=\pm\frac{\Delta}{24}(24-(q_2 W)^2)$, $$\frac{1}{D_e\tau_s}=\frac{\Delta^2}{24}(q_2W)^2.$$ We set $\Delta=\sqrt{\tilde q_1^2+q_2^2}$. With cubic Dresselhaus SOC -------------------------- If cubic Dresselhaus SOC cannot be neglected, the absolute minimum of spin relaxation can also shift to $k_{x_{\text{\tiny{min}}}}=0$. This depends on the ratio of Rashba and lin. Dresselhaus SOC:\ If $q_2/q_1\ll1$, we find the absolute minimum at $k_{x_{\text{\tiny{min}}}}=0$, $$\begin{aligned} E_{min1}={}&\frac{\tilde q_3+\tilde q_1^2+ q_2^2}{2}-\Delta_c+\frac{1}{12}\Delta_c(q_2 W)^2,\end{aligned}$$ with $$\Delta_c=\frac{1}{2}\sqrt{(\tilde q_3+\tilde q_1^2)^2+2(\tilde q_1^2-\tilde q_3)q_2^2+q_2^4}.$$ If $q_2/q_1\gg 1$, we find the absolute minimum at $k_{x_{\text{\tiny{min}}}}\approx\pm\frac{\Delta}{24}(24-(q_2 W)^2)$, $$\begin{aligned} E_{min2}= {}&k_{x_{\text{\tiny{min}}}}^2-k_{x_{\text{\tiny{min}}}}q_2\left(\frac{\tilde q_1^2}{q_2^2}+2\right)-\frac{\tilde q_3^2}{16 k_{x_{\text{\tiny{min}}}} q_2}\nonumber\\ {}&+\tilde \Delta^2+\frac{\tilde q_3}{2}\left(\frac{\tilde q_1^2}{q_2^2}+1\right)\nonumber\\ {}&-\left(\frac{\tilde q_3 \tilde q_1^2}{12}-\frac{\tilde q_3^2 q_2}{3072 k_{x_{\text{\tiny{min}}}}^3}-\frac{q_2^2}{24}(\tilde q_3-\tilde q_1^2)\right.\nonumber\\ {}&+\frac{ q_2^4}{24}-\left(\frac{\tilde q_1^2}{24}+\frac{q_2^2}{12}\right)q_2 k_{x_{\text{\tiny{min}}}}\nonumber\\ {}&\left.-\frac{q_2}{k_{x_{\text{\tiny{min}}}}}\left(\left(\frac{\tilde q_3^2}{128}+\frac{\tilde q_3 \tilde q_1^2}{192}\right)-\frac{\tilde q_3 q_2^2}{96}\right)\right)W^2.\label{110Emin2}\end{aligned}$$ We can conclude that reducing wire width W will not cancel the contribution due to cubic Dresselhaus SOC to the spin relaxation rate. weak localization {#sec:WL} ================= In Ref. the crossover from WL to WAL due to change of wire width and SOC strength was explained in the case of a (001) system. Whether WL or WAL is present depends on the suppression of the triplet modes of the Cooperon. The suppression in turn is dominated by the absolute minimum of the spectrum of the Cooperon Hamiltonian $H_c$. The findings presented in Sec.\[sec:001grothDirWire\] therefore point out that e.g. the crossover width, at which the system changes from WL to WAL, can shift with the wire direction $\theta$. Recently experimental results on WL/WAL by J. Nitta *et al.*, Ref., seem to show a strong dependence on growth direction.\ In the (110) system the situation is different: In the 2D case it was shown by Pikus *et al.*, Ref., that in the absence of the Rashba terms the negative magnetoconductivity cannot be observed. In the case of a wire geometry we can conclude from Eqs.(\[110E1\]-\[110Emin2\]) that we have no width dependence if Rashba SOC vanishes. A change of the quantum correction to the static conductivity therefore cannot be achieved in this wire geometry by changing the wire width. The reason is the vector potential in the boundary condition, Eq.(\[110boundaryCond\]), which only depends on the Rashba SOC. Diffusive-Ballistic Crossover {#sec:DiffBallis} ============================= In the following we assume a (001) 2D system with both, Rashba and linear and cubic Dresselhaus SO coupling.\ Experiments measuring WL in diffusive QW with SOC[@hu05_2; @Schapers2009] are in great agreement with theoretical calculations by S. Kettemann, Ref.. But considering e.g. the works Ref., one realizes that the scope of application of the theory has to be extended to describe also the crossover to the ballistic regime, $l_e>W$. We have shown in Sec.\[sec:001grothDirWire\] that the presence of cubic Dresselhaus SOC in the sample leads to a finite spin relaxation even for wire widths $Q_{\text{\tiny{SO}}}W\ll 1$, regardless of the boundary direction in a (001) system. To account for the ballistic case we have to modify the derivation of the Cooperon Hamiltonian, Eq.(\[H001\]). In the case of a wire where the mean free path $l_e$ is comparable to the wire width W, we cannot integrate in Eq.(\[Cooperon0\]) over the Fermi surface in a continuous way. Instead, we assume $k_F/W$ to be finite and sum over the number of discrete channels $N=[k_F W/\pi]$, where $[\ldots]$ is the integer part. Because $H_{\gamma D}\sim \epsilon_F^2$ this constant term due to cubic Dresselhaus should reduce if we reduce the number of channels. If we expand the Cooperon to second order in $( {\bf Q} +2 e {\bf A} + 2m_e \hat{a} {\bf S} )$ before averaging over the Fermi surface, $\langle\ldots\rangle$, and use the Matsubara trick, we get $$\begin{aligned} \label{C1} \frac{C^{-1}}{D_e}={}&2f_1\left(Q_y+2\alpha_2S_x+2\left(\alpha_1-\gamma_D v^2\frac{f_3}{f_1}\right)S_y\right)^2\nonumber\\ {}& +2f_2\left(Q_x-2\alpha_2S_y-2\left(\alpha_1-\gamma_D v^2\frac{f_3}{f_2}\right)S_x\right)^2\nonumber\\ {}& +8\gamma_D^2v^4\left[\left(f_4-\frac{f_3^2}{f_2}\right)S_x^2+\left(f_5-\frac{f_3^2}{f_1}\right)S_y^2\right],\end{aligned}$$ with $m_e=1$ and functions $f_i(\varphi)$ (App.\[Appen2\]) which depend on the number of transverse modes N. In the diffusive case we can perform the continuous sum over the angle $\varphi$ in Eq.(\[f1\])-(\[f5\]), and we receive the old result with $f_1=f_2=1/2$, $f_3=1/8$ and $f_4=f_5=1/16$: $$\begin{aligned} H_c={}&(Q_y+2\alpha_2S_x+2\left(\alpha_1-\frac{1}{2}\gamma_D \epsilon_F\right)S_y)^2\nonumber\\ & + (Q_x-2\alpha_2S_y-2\left(\alpha_1-\frac{1}{2}\gamma_D \epsilon_F\right)S_x)^2\nonumber\\ & +(\gamma_D \epsilon_F)^2(S_x^2+S_y^2).\end{aligned}$$ Spin Relaxation at $Q_{\text{\tiny{SO}}}W\ll 1$ ----------------------------------------------- In the first section we analyzed the lowest spin relaxation in wires of different direction in a (001) system. We have shown, that for every direction there is still a finite spin relaxation at wire width which fulfill the condition $Q_{\text{\tiny{SO}}}W\ll 1$ due to cubic Dresselhaus SOC. It is clear that this finite spin relaxation vanishes when the width is equal to the Fermi wave length $\lambda_F$. In the following we show how this finite spin relaxation depends on the number of transverse channels $N$. We show in Ref. that the findings are consistent with calculations going beyond the perturbative ansatz. This is possible in a similar manner as has been done previously in Ref. for wires without SOC, which showed the crossover of the magnetic phase shifting rate, which had been known before in the diffusive and ballistic limit, only.\ To find the spectrum of the Cooperon Hamiltonian with boundary conditions as in Sec.(\[sec:001grothDirWire\]), we stay in the 0-mode approximation in the Q space and proceed as before: According to Eq.(\[C1\]), the non-Abelian gauge transformation for the transversal direction y is given by $$\label{trafo_ballistic} U=\exp\left(-{\mathit{i}}\left[2\alpha_2 S_x+2\left(\alpha_1-\gamma_D v^2\frac{f_3}{f_1}\right)S_y\right] y\right).$$ To concentrate on the constant width independent part of the spectrum we extract the absolute minimum at $Q=0$, Fig.(\[plot:modesCubicAlpha2\]) and Fig.(\[plot:modesCubicAlpha1\]). A clear reduction of the absolute minimum is visible. Due to the factor $f_3/f_1$ in the transformation U, the decrease of the minimal spin relaxation depends also on the ratio of Rashba and linear Dresselhaus SOC. ![(Color online) The lowest eigenvalues of the confined Cooperon Hamiltonian Eq.(\[C1\]), equivalent to the lowest spin relaxation rate, are shown for $Q=0$ for different number of modes $N=k_F W/\pi$. Different curves correspond to different values of $\alpha_2/q_s$.[]{data-label="plot:modesCubicAlpha2"}](modesCubicAlpha1.eps "fig:"){width="\columnwidth"}\ ![(Color online) The lowest eigenvalues of the confined Cooperon Hamiltonian Eq.(\[C1\]), equivalent to the lowest spin relaxation rate, are shown for $Q=0$ for different number of modes $N=k_F W/\pi$. Different curves correspond to different values of $\alpha_1/q_s$.[]{data-label="plot:modesCubicAlpha1"}](modesCubicAlpha2.eps "fig:"){width="\columnwidth"}\ From Eq.(\[C1\]) it is clear, that not only the $H_{\gamma_D}$ is affected by the reduction of the number of channels N but also the shift of the lin. Dresselhaus SOC, $\alpha_1$, in the orbital part. A model to extract the ratio of Rashba and lin. Dresselhaus SOC developed in Ref.[@epub7971] by Scheid *et al.* did not show much difference between the strict 1D case and the non-diffusive case with wire of finite width. The results presented here should allow for extending the model to finite cubic Dresselhaus SOC. Deducing from our theory, the direction of the SO field should change with the number of channels due to the mentioned N dependent shift. Conclusions =========== Summarizing the results, we have characterized the anisotropy and width dependence of spin relaxation in a (001) QW. There are special angles $\theta$ which are optimal for spin transport in quantum wires of finite width: The \[110\] and the $[\overline{1}10]$ direction. At \[110\] we find the the longest spin dephasing time $T_2$. If the absolute minimum of spin relaxation is found at \[110\] or $[\overline{1}10]$ direction depends on the strength of cubic Dresselhaus and wire width. The findings for the spin dephasing time are in agreement with numerical results. The analytical expression for $T_2$ allows to see directly the interplay between the cubic Dresselhaus SOC and the dimensional reduction, having effect on $T_2$. In addition we analyzed the special case of a (110) system and found the minimal spin relaxation rates depending on Rashba and lin. and cubic Dresselhaus SOC in the presence of boundaries. This results can be used to understand width and direction dependent WL measurements in QWs. Finally, we have shown how the reduction of channels in the wire reduces the finite spin relaxation rate which is due to cubic Dresselhaus SOC and does not reduce if the wire is small, $Wq_s\ll 1$, and diffusive, $W\gg l_e$. The change in channel number also changes the shift of lin. Dresselhaus SOC strength, $\tilde\alpha_1$. This has to be considered if extracting SOC strength from wires with only few transverse channels. P.W. thanks the Asia Pacific Center for Theoretical Physics for hospitality. This research was supported by DFG-SFB508 B9 and by WCU ( World Class University ) program through the Korea Science and Engineering Foundation funded by the Ministry of Education, Science and Technology (Project No. R31-2008-000-10059-0). Hamiltonian in \[110\] growth direction {#Appen1} ======================================= The Cooperon Hamiltonian in the 0-mode approximation is given as follows $$H_{c,0}= \left( \begin{array}{lll} A & B & C \\ B^* & D & E \\ C^* & E^* & F \end{array} \right) +M_{q3}, \label{hamiltonian}$$ with $$\begin{aligned} A={}&\frac{1}{4 q_2 W}(q_2 \left(4 k_x^2+3 \left(\tilde q_1^2+q_2^2\right)\right) W\nonumber\\ {}&-16 k_x \tilde q_1 \sin \left(\frac{q_2 W}{2}\right)+\left(\tilde q_1^2-q_2^2\right) \sin (q_2 W)),\\ B={}&\frac{{\mathit{i}}\left(4 k_x \sin \left(\frac{q_2 W}{2}\right)-\tilde q_1 \sin (q_2 W)\right)}{\sqrt{2} W},\\ C={}&-\frac{q_2 \left(\tilde q_1^2+q_2^2\right) W+\left(q_2^2-\tilde q_1^2\right) \sin (q_2 W)}{4 q_2 W},\\ D={}&\frac{q_2 \left(2 k_x^2+\tilde q_1^2+q_2^2\right) W+\left(q_2^2-\tilde q_1^2\right) \sin (q_2 W)}{2 q_2 W},\\ E={}&\frac{{\mathit{i}}\left(4 k_x \sin \left(\frac{q_2 W}{2}\right)+\tilde q_1 \sin (q_2 W)\right)}{\sqrt{2} W},\\ F={}&\frac{1}{4 q_2 W}(q_2 \left(4 k_x^2+3 \left(\tilde q_1^2+q_2^2\right)\right) W\nonumber\\ {}&+16 k_x \tilde q_1 \sin \left(\frac{q_2 W}{2}\right)+\left(\tilde q_1^2-q_2^2\right) \sin (q_2 W))\end{aligned}$$ and the term due to cubic Dresselhaus SOC $$\begin{aligned} {}& M_{q3}=\nonumber\\ {}& q_3\left( \begin{array}{lll} \frac{1}{4}{\sin\!\text{c}}(q_2 W)+\frac{3}{4} & 0 & \frac{1}{4}{\sin\!\text{c}}(q_2 W)-\frac{1}{4}\\ 0 & \frac{1}{2}-\frac{1}{2}{\sin\!\text{c}}(q_2 W) & 0 \\ \frac{1}{4}{\sin\!\text{c}}(q_2 W)-\frac{1}{4} & 0 & \frac{1}{4}{\sin\!\text{c}}(q_2 W)+\frac{3}{4} \end{array} \right).\end{aligned}$$ Summation over the Fermi Surface {#Appen2} ================================ The Cooperon Hamiltonian in the 2D case is given by $$\begin{aligned} \label{C0} H_c ={}& \tau v^2\{\langle \cos^2(\varphi)\rangle({\bf Q}+2m_e {\bf a.S})^2_x\nonumber\\ {}& +\langle \sin^2(\varphi)\rangle({\bf Q}+2m_e {\bf a.S})^2_y\nonumber\\ {}& +4m_e^2\gamma_D v^2\langle \cos^2(\varphi)\sin^2(\varphi)\rangle({\bf Q}+2m_e{\bf a.S})_x.S_x\nonumber\\ {}& -4m_e^2\gamma_D v^2\langle \sin^2(\varphi)\cos^2(\varphi)\rangle({\bf Q}+2m_e{\bf a.S})_y.S_y\nonumber\\ {}& +(2m_e^3\gamma_D v^2)^2(\langle \cos^2(\varphi)\sin^4(\varphi)\rangle S_x^2\nonumber\\ {}& +\langle \sin^2(\varphi)\cos^4(\varphi)\rangle S_y^2)\},\end{aligned}$$ with wave vector $\mathbf Q$. We set $$\begin{aligned} m_e \equiv{}& 1,\\ f_1 :={}&\langle\sin^2(\varphi)\rangle\label{f1},\\ f_2 :={}&\langle\cos^2(\varphi)\rangle,\\ f_3 :={}&\langle\sin^2(\varphi)\cos^2(\varphi)\rangle,\\ f_4 :={}&\langle\sin^4(\varphi)\cos^2(\varphi)\rangle,\\ f_5 :={}&\langle\sin^2(\varphi)\cos^4(\varphi)\rangle\label{f5}.\end{aligned}$$ Using the Matsubara trick we write $$\int_0^{2\pi}\frac{d\varphi}{2\pi}=\frac{2}{\pi N}\sum_{s=1}^N\frac{1}{\sqrt{1-\left(\frac{s}{N}\right)^2}}.$$ This gives us $$\begin{aligned} f_1 ={}&\frac{2}{\pi N}\sum_{s=1}^{N-1}\frac{s^2}{N^2 \sqrt{1-\left(\frac{s}{N}\right)^2}},\\ f_2 ={}&\frac{2}{\pi N}\sum_{s=1}^{N}\sqrt{1-\left(\frac{s}{N}\right)^2},\\ f_3 ={}&\frac{2}{\pi N}\sum_{s=1}^{N}\left(\frac{s}{N}\right)^2 \sqrt{1-\left(\frac{s}{N}\right)^2},\\ f_4 ={}&\frac{2}{\pi N}\sum_{s=1}^{N}\left(\frac{s}{N}\right)^4 \sqrt{1-\left(\frac{s}{N}\right)^2},\\ f_5 ={}&\frac{2}{\pi N}\sum_{s=1}^{N}\left(\frac{s}{N}\right)^2 \left(1-\left(\frac{s}{N}\right)^2\right)^{\frac{3}{2}}.\end{aligned}$$ Writing Eq.(\[C0\]) in a compact way gives us Eq.(\[C1\]).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Compactifications of IIB string theory with internal NS and RR three-form flux are computationally attractive in that the lifting of moduli is via a [*perturbative*]{} and often explicitly calculable (super)potential. We focus on the $T^6/Z_2$ orientifold, and provide an illustative ${\cal N} = 2$ example. For other choices of flux, the resulting equations of motion can be solved to yield ${\cal N}=0,1,2$ or 3 supersymmetry in four dimensions.(arXiv:0810.5197, CALT-68-2441)' author: - 'Michael B. Schulz' --- Moduli are a familiar by-product of string compactifications, but they do not exist in nature. Massless or nearly massless gravitationally coupled scalars generate long range interactions that have been excluded by fifth force experiments [@Smith]. They are also problematic in cosmology [@Linde].[^1] In traditional ${\cal N}=1$ heterotic compactifications, perturbatively massless moduli can be lifted by well-known nonperturbative effects like world-sheet instantons or Euclidean NS five-brane instantons. However, the computational difficulty of quantitatively understanding these effects has so far proven insurmountable. There does not seem to exist a single compact example in which anyone has computed the relevant instanton sums and explicitly found a supersymmetric minimum of the resulting potential. In contrast, when one turns on NS and RR three-form flux through nontrivial three-cycles in a four-dimensional compactification of IIB string theory, many of the moduli are lifted by a [*perturbative*]{} scalar potential [@GKP] $$\label{potential} {\cal V} \propto \int d^6y \sqrt{g_6}\, \bigl|G^{\rm ISD}\bigr|^2, \quad G = F^{\rm RR}_{(3)} - \varphi H^{\rm NS}_{(3)}.$$ Here, $\varphi$ is the dilaton-axion and $G^{\rm ISD}$ is the imaginary-anti-self-dual part of the complex flux $G$.[^2] This potential descends from the compact part of the flux kinetic terms in the ten-dimensional IIB supergravity action. For consistency of the compactification, the fluxes must satisfy a D3-brane charge tadpole cancellation condition, $$\label{tadpole} N_{\rm D3} + \frac{1}{2(2\pi)^4\alpha'^2} \int H^{\rm NS}_{(3)}\wedge F^{\rm RR}_{(3)} - \frac{1}{4} N_{\rm O3} = 0.$$ in which wedged fluxes contribute in exactly the same way as space-filling D3-branes and O3-planes.[^3] An obvious way to satisfy this condition is to start with a consistent string compactification involving space-filling D3-branes, and then construct new vacua by simply trading off D3-branes for fluxes. However, whereas space-filling D3-branes and O3-planes preserve ${\cal N}=4$ supersymmetry in four dimensions, the fluxes preserve less. To preserve at least ${\cal N}=1$ supersymmetry, we need vanishing dilatino variation, and vanishing gravitino variation for at least one gravitino. This implies that $G$ must be primitive ([*i.e.*]{}, $J\wedge G = 0$, with $J$ the Kähler form), and of type (2,1).[^4] The (2,1) condition can be imposed by a superpotential [@GVW] $$\label{superpot} W = \int G\wedge\Omega,$$ where $\Omega$ is the holomorphic (3,0) form. For proper Calabi-Yau compactification, the primitivity condition is trivial due to the absence of a fifth cohomology class. More generally, it is a linear constraint on the Kähler moduli which is easy to solve. On the other hand, the (2,1) condition is not so simple. When the periods of the holomorphic (3,0) form are known, we can compute the superpotential (\[superpot\]). But for Calabi-Yau orientifolds or F-theory compactifications, this generally involves complicated trancendental functions, for which it has not been possible to vary the superpotential and solve the resulting equations, except near singular (conifold) points in the moduli space of complex structure. Still, one might expect that the equations of motion [*are*]{} soluble when we choose a simple enough compactification manifold, and simplest choice is a torus. That is the choice we will make here, with one modification. Since we would like to turn on flux, Eq. (\[tadpole\]) requires that there also be orientifold planes. So, the compactification that we will consider is on the torus orientifold $T^6/Z_2$ [@FP; @KST]. To define the $T^6/Z_2$ orientifold, we first compactify on $T^6$, defined by $x^i\cong x^i+1$, $y^i\cong y^i+1$, $i=1,2,3$. Then, we mod out by the $Z_2$ parity operation $\Omega R_6(-1)^{F_L}.$ Here, $\Omega$ is worldsheet parity, $R_6$ is a reflection of all of the $T^6$ coordinates, and $(-1)^{F_L}$ is a parity operation that is required by supersymmetry.[^5] The massless states that survive the orientifold projection are the four-dimensional graviton $g_{\mu\nu}$, the scalars $g_{ab}$, $C_{abcd}$ and $\varphi$, and the twelve $U(1)$ gauge bosons $B_{a\mu}$ and $C_{a\mu}$. (Here $C$ denotes a RR potential, and $B$ the NS potential). It is also consistent with the orientifold projection to turn on internal NS and RR three-form fluxes. However, note that these fluxes are discrete by Dirac quantization, and non-dynamical in the massless sector, since the corresponding zero-modes are projected out. In the absence of flux, this orientifold describes the same theory as Type I on $T^6$, via T-duality in all six torus directions. The sixteen D9-branes of $SO(32)$ in Type I become sixteen D3-branes after T-duality. Also, the charge and tension of the D3-branes is cancelled by $2^6$ O3-planes located at the fixed points of the $Z_2$. So, the low energy effective field theory is the same ${\cal N}=4$ $SO(32)$ super-Yang-Mills, coupled to ${\cal N}=4$ supergravity. Once we replace some of the D3-branes with fluxes, this story is modified. The fluxes generate a potential for the scalars, and correspond to turning on charges that couple the scalars to the twelve $U(1)$ gauge fields. The result is a superhiggs mechanism in which many of the scalars get massive or are eaten by massive vectors, breaking ${\cal N}=4$ to ${\cal N}<4$ supersymmetry. As an example, consider the choice of flux [@KST] $$\begin{aligned} \label{fluxes} \frac{1}{(2\pi)^2\alpha'}F^{\rm RR}_{(3)} & = & 4dx^1\wedge dx^2\wedge dy^3 + 4dy^1\wedge dy^2\wedge dy^3,\\ \frac{1}{(2\pi)^2\alpha'}H^{\rm NS}_{(3)} & = & 4dx^1\wedge dx^2\wedge dx^3 + 4dy^1\wedge dy^2\wedge dx^3.\end{aligned}$$ Let us parametrize the complex structure as $dz^i = dx^i+\tau^i{}_jdy^j$, and normalize the holomorphic three-form so that $\Omega = dz^1\wedge dz^2\wedge dz^3$. Then, by wedging the appropriate three-forms together, it is easy to show that $$W = \int G\wedge\Omega \,\propto\, 1+ \bigl({\rm cof}\,\tau\bigr)_3{}^3 +\varphi\Bigl(\det\tau + \tau^3{}_3\Bigr).$$ For supersymmetric vacua, the equations of motion are that $D_I W = \partial_I W + (\partial_I{\cal K}) W = 0$, where ${\cal K}(\varphi^I,\bar{\varphi}^{\bar{I}})$ is the Kähler potential on moduli space. Since the superpotential is independent of Kähler moduli, this simplifies to $W= \partial_\varphi W = \partial_\tau W = 0$, which through a small amount of algebra can be shown to imply that $$\label{cpxstrmod} \varphi\,\tau^3{}_3 = -1, \quad \tau^1{}_1\,\tau^2{}_2-\tau^1{}_2\,\tau^2{}_1 = -1.$$ So, the moduli space of complex structure is complex four-dimensional, and can be parametrized by, say $\tau^1{}_1$, $\tau^2{}_2$, $\tau^3{}_3$, and $\tau^1{}_2$. In Eq. (\[fluxes\]), we expressed the flux as a linear combination of integral three-forms with integer coefficients, as required Dirac quantization. Using Eq. (\[cpxstrmod\]), we can also write the flux in terms of holomorphic and antiholomorphic forms. Restricting to $\tau^i{}_j$ diagonal and imaginary for simplicity, we find that $$\label{holoflux} G \propto dz^1\wedge d{\bar z}^{\bar2}\wedge dz^3 + d{\bar z}^{\bar1}\wedge dz^2\wedge dz^3.$$ This makes it clear that if Eq. (\[cpxstrmod\]) is satisfied then the complex flux is indeed of type (2,1). In addition, it is easy to show that the primitivity condition is satisfied on the appropriate subspace of Kähler moduli. As a final remark, note that if we replace $z^1$ and $z^2$ by their complex conjugates, then the flux (\[holoflux\]) is still (2,1) and primitive. In other words, there are two inequivalent complex structures in which the conditions for ${\cal N}=1$ supersymmetry are satisfied. This implies that the solution is actually ${\cal N}=2$ supersymmetric, and shows how to engineer solutions with anywhere from ${\cal N}=0$ supersymmetry (when there is [*no*]{} solution to $D_I W=0$) to ${\cal N} = 3$ supersymmetry (when the solution permits three independent complex structures). For a more complete discussion, including large classes of ${\cal N}=1$ solutions, we refer the reader to the work [@KST] on which this review is based, and the references contained therein. I would like to thank S. Kachru and S. P. Trivedi for the collaboration on which these proceedings are based. In addition, I would like to thank the organizers of the Cargèse 2002 ASI, as well as M. Berg, M. Haack, and A. Micu for useful discussions during the course of the school. This work was supported in part by the DOE under contracts DE-AC03-76SF00515 and DE-FG03-92-ER40701. [99]{} G. L. Smith, C. D. Hoyle, J. H. Gundlach, E. G. Adelberger, B. R. Heckel and H. E. Swanson, “Short Range Tests Of The Equivalence Principle,” Phys. Rev. D [**61**]{}, 022001 (2000). A. D. Linde, “Relaxing the Cosmological Moduli Problem,” Phys. Rev. D [**53**]{}, 4129 (1996) \[arXiv:hep-th/9601083\]. A. R. Frey and J. Polchinski, “${\cal N} = 3$ warped compactifications,” Phys. Rev. D [**65**]{}, 126009 (2002) \[arXiv:hep-th/0201029\]. S. B. Giddings, S. Kachru and J. Polchinski, “Hierarchies from fluxes in string compactifications,” Phys. Rev. D [**66**]{}, 106006 (2002) \[arXiv:hep-th/0105097\]. S. Gukov, C. Vafa and E. Witten, “CFT’s from Calabi-Yau four-folds,” Nucl. Phys. B [**584**]{}, 69 (2000) \[Erratum-ibid. B [**608**]{}, 477 (2001)\] \[arXiv:hep-th/9906070\]. S. Kachru, M. B. Schulz, P. K. Tripathy and S. P. Trivedi, “New supersymmetric string compactifications,” JHEP [**0303**]{}, 061 (2003) \[arXiv:hep-th/0211182\]. [^1]: As the early universe cools, approximate moduli can easily overshoot the minima of their potentials. When this happens, they contribute an energy density like that of matter rather than radiation, causing deviations from the successful predictions of big bang nucleosynthesis. [^2]: Since the compact manifold is six-dimensional, the hodge star operator squares to $-1$ and its eigenvalues are $\pm i$. The corresponding eigenfunctions are [*imaginary-self-dual*]{} (ISD) three-forms ($*a_{(3)} = ia_{(3)}$) and [*imaginary-anti-self-dual*]{} (IASD) three-forms ($*a_{(3)} = - ia_{(3)}$). [^3]: In F-theory compactifications, there would also be a contribution to Eq. (\[tadpole\]) from the Euler character of the fourfold, and from instantons on the compact part of D7-brane worldvolumes. [^4]: It can be shown that if $G$ is (2,1) and primitive then it is also ISD. So, if the supersymmetry conditions are satisfied, then the scalar potential (\[potential\]) is automaticaly minimized and equal to zero. [^5]: For massless modes, $(-1)^{F_L}$ acts as $-1$ on left-moving Ramond sector states and $+1$ otherwise. If this factor were not included, the resulting spectrum of states would not fill out supergravity multiplets.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Exoplanets orbiting M dwarfs present a valuable opportunity for their detection and atmospheric characterisation. This is evident from recent inferences of H$_2$O in such atmospheres, including that of the habitable-zone exoplanet K2-18b. With a bulk density between Earth and Neptune, K2-18b may be expected to possess a H/He envelope. However, the extent of such an envelope and the thermodynamic conditions of the interior remain unexplored. In the present work, we investigate the atmospheric and interior properties of K2-18b based on its bulk properties and its atmospheric transmission spectrum. We constrain the atmosphere to be H$_2$-rich with a H$_2$O volume mixing ratio of $0.02-14.8$%, consistent with previous studies, and find a depletion of CH$_4$ and NH$_3$, indicating chemical disequilibrium. We do not conclusively detect clouds/hazes in the observable atmosphere. We use the bulk parameters and retrieved atmospheric properties to constrain the internal structure and thermodynamic conditions in the planet. The constraints on the interior allow multiple scenarios between rocky worlds with massive H/He envelopes and water worlds with thin envelopes. We constrain the mass fraction of the H/He envelope to be $\lesssim 6$%; spanning $\lesssim 10^{-5}$ for a predominantly water world to $\sim6$% for a pure iron interior. The thermodynamic conditions at the surface of the H$_2$O layer range from the super-critical to liquid phases, with a range of solutions allowing for habitable conditions on K2-18b. Our results demonstrate that the potential for habitable conditions is not necessarily restricted to Earth-like rocky exoplanets.' author: - Nikku Madhusudhan - 'Matthew C. Nixon' - Luis Welbanks - 'Anjali A. A. Piette' - 'Richard A. Booth' title: 'The interior and atmosphere of the habitable-zone exoplanet K2-18b' --- Introduction {#sec:intro} ============ ![image](fig4.pdf){width="100.00000%"} Recent exoplanet detection surveys have revealed high occurrence rates of low-mass planets orbiting M Dwarfs [@dressing2015; @mulders2015]. The low masses, sizes and temperatures of M Dwarfs also mean that the planet-star contrast is favourable for planetary detection and characterisation. This ‘small-star opportunity’ has led to several detections of low-mass planets ($<10M_{\oplus}$) in the habitable-zones of M Dwarf hosts such as Trappist-1 [@gillon2017], Proxima Centauri [@anglada-escude2016], K2-18 [@montet2015; @foreman-mackey2015], and LHS 1140 [@dittmann2017]. The habitable-zone transiting exoplanet K2-18b is a particularly good example [@foreman-mackey2015; @montet2015]. The close proximity and small size of its host star make precise measurements of the planetary mass, radius, and atmospheric spectra viable [@benneke2017; @cloutier2019], as exemplified by the recent detection of H$_2$O in its atmosphere [@tsiaras2019; @benneke2019]. The habitable-zone temperature of K2-18b provides further impetus for detailed characterisation of its interior and atmosphere. Given its mass ($M_p$ = 8.63 $\pm$ 1.35 $M_{\oplus}$, @cloutier2019) and radius ($R_p$ = 2.610 $\pm$ 0.087 $R_{\oplus}$, @benneke2019), K2-18b has a bulk density (2.67$^{+0.52}_{-0.47}$ g/cm$^3$, @benneke2019). This density, between that of Earth and Neptune, may be thought to preclude a purely rocky or icy interior and require a hydrogen-rich outer envelope. However, the extent of such an envelope and the conditions at the interface between the envelope and the underlying interior have not been explored. We note that the mass and radius of the planet have recently been revised [@benneke2019], which may have impacted inferences made using previous values [@cloutier2017; @tsiaras2019]. Previous studies of planets with similar masses and radii, such as GJ 1214b, suggested envelope mass fractions $\lesssim$7% [@rogers2010; @nettelmann2011; @valencia2013]. GJ 1214b is expected to host super-critical H$_2$O below the envelope at pressures and temperatures too high to be conducive for life [@rogers2010]. However, while GJ 1214b has an equilibrium temperature ($T_{eq}$) of $\sim500$ K, K2-18b may be more favourable given its lower $T_{eq}\sim250-300$ K. In the present work, we conduct a systematic study to constrain both the atmospheric and interior composition of K2-18b based on extant data along with detailed atmospheric retrievals and internal structure models. [c|ccccc]{} Case 1: Full model, inhomogenous clouds and hazes & $-2.11 ^{+ 1.06 }_{- 1.19 }$ & $-8.20 ^{+ 2.53 }_{- 2.34 }$ & $-8.64 ^{+ 2.15 }_{- 2.06 }$ & $179.15$ & Reference\ No H$_2$O & N/A & $-1.11 ^{+ 0.53 }_{- 1.22 }$ & $-7.27 ^{+ 2.91 }_{- 2.92 }$ & $175.30$& $3.25$\ Case 2: Clear atmosphere & $-2.18 ^{+ 1.35 }_{- 1.44 }$ & $-8.27 ^{+ 2.59 }_{- 2.42 }$ & $-8.60 ^{+ 2.19 }_{- 2.16 }$ & $179.05$ & $1.20$\ Case 3: Opaque cloud deck & $-1.80 ^{+ 0.81 }_{- 1.22 }$ & $-8.13 ^{+ 2.64 }_{- 2.41 }$ & $-8.57 ^{+ 2.30 }_{- 2.17 }$ & $179.09$ & $1.06$\ Case 4: Inhomogenous clouds & $-2.10 ^{+ 1.07 }_{- 1.28 }$ & $-8.26 ^{+ 2.56 }_{- 2.34 }$ & $-8.61 ^{+ 2.18 }_{- 2.10 }$ & $179.41$ & N/A Atmospheric Properties {#sec:atmosphere} ====================== We retrieve the atmospheric properties of K2-18b using its broadband transmission spectrum reported by @benneke2019. The data include observations from the HST WFC3 G141 grism (1.1-1.7 $\mu$m), photometry in the Spitzer IRAC 3.6 $\mu$m and 4.5 $\mu$m bands, and optical photometry in the K2 band (0.4-1.0 $\mu$m). We perform the atmospheric retrieval using an adaptation of the AURA retrieval code [@pinhas2019; @welbanks2019]. Our model solves line-by-line radiative transfer in a plane-parallel atmosphere in transmission geometry. The model assumes hydrostatic equilibrium and considers prominent opacity sources in the observed spectral bands as well as homogeneous/inhomogeneous cloud/haze coverage. Clouds are included through a gray cloud deck with cloud-top pressure ($\text{P}_\text{c}$) as a free parameter. Hazes are included as a modification to Rayleigh-scattering through parameters for the scattering slope ($\gamma$) and a Rayleigh-enhancement factor ($a$). The opacity sources include H$_2$O [@rothman2010], CH$_4$ [@yurchenko2014], NH$_3$ [@yurchenko2011], CO$_2$ [@rothman2010], HCN [@barber2014], and collision-induced absorption due to H$_2$-H$_2$ and H$_2$-He [@richard2012]. The model comprises 16 free parameters: abundances of 5 molecules, 6 parameters for the pressure-temperature ($P$-$T$) profile, 4 cloud/haze parameters, and 1 parameter for the reference pressure $P_{\text{ref}}$ at $R_p$ [e.g., @welbanks2019b]. The Bayesian parameter estimation is conducted using the Nested Sampling algorithm MultiNest [@feroz2009] through PyMultiNest [@Buchner2014]. We conduct retrievals for four model configurations: (1) a full model including inhomogeneous clouds and hazes, (2) a clear atmosphere, (3) an atmosphere with an opaque cloud deck but no hazes, and (4) an atmosphere with inhomogeneous clouds but no hazes. The atmospheric constraints are shown in Figure \[fig:atmos\] and Table  \[table:atmos\]. We confirm the high-confidence detection of H$_2$O in a H$_2$-rich atmosphere as reported by @benneke2019 and @tsiaras2019. Our abundance estimates are consistent to within 1$\sigma$ between all four model configurations and with @benneke2019. The derived H$_2$O volume mixing ratio ranges between 0.02-14.80%, with median values of 0.7-1.6% between the 4 model cases, as shown in Table \[table:atmos\]. The case with an opaque cloud deck (a clear atmosphere) retrieves slightly higher (lower) H$_2$O abundances as expected [@welbanks2019]. Our derived H$_2$O abundance range corresponds to an O/H ratio of 0.2-176.8$\times$solar, assuming all the oxygen is in H$_2$O as expected in H$_2$-rich atmospheres at such low temperatures [@burrows1999]. The median H$_2$O abundance is 9.3$\times$solar for the full model, case 1. We cannot compare our results with @tsiaras2019 as their retrievals were based on only the HST WFC3 data and used older measurements of the planetary mass and radius which could have biased their inferences. We find a depletion of CH$_4$ and NH$_3$ in the atmosphere. For a H$_2$-rich atmosphere at $\sim$300 K, CH$_4$ and NH$_3$ are expected to be dominant carriers of carbon and nitrogen, respectively, in chemical equilibrium [@burrows1999], as also seen for the gas and ice giants in the solar system [@atreya2016]. Assuming solar elemental ratios (i.e., C/O = 0.55, N/O = 0.14), the CH$_4$/H$_2$O (NH$_3$/H$_2$O) ratio is expected to be $\sim$0.5 ($\sim$0.1). However, we do not detect CH$_4$ or NH$_3$ despite their strong absorption in the HST WFC3 and/or Spitzer 3.6 $\mu$m bands. As shown in Figure \[fig:atmos\], the retrieved posteriors of the CH$_4$ and NH$_3$ abundances are largely sub-solar, with 99% upper limits of 3.47$\times 10^{-2}$ and 5.75$\times 10^{-5}$, respectively. These sub-solar values are in contrast to the largely super-solar H$_2$O, arguing against chemical equilibrium at solar elemental ratios. We do not find strong evidence for clouds/hazes in the atmosphere. Our model preference for clouds/hazes, relative to the cloud-free case, is marginal (1.2$\sigma$) compared to @benneke2019 (2.6$\sigma$). Our retrieved cloud-top pressure ($P_c$) for the full case is weakly constrained to 0.1 mbar to 2 bar, close to the observable photosphere. Finally, we retrieve $P_{\text{ref}}$ for the full case to be $12 - 174$ mbar corresponding to $R_p$. The median value of 0.05 bar is used as the surface boundary condition, pressure $P_0$, for the internal structure models in section \[sec:interior\_model\]. ![Model mass-radius (M-R) relations for planets with different compositions. The mass fractions are shown in the legend. The solid magenta, teal, and orange curves show cases with three representative compositions, discussed in section \[sec:scenarios\], that all fit the mass and radius of K2-18b equally well. The dashed magenta line represents the same composition as the solid magenta line, but with a mixed H$_2$O-H/He envelope. Also shown are exoplanets whose masses and radii are known to $\geq 3\sigma$ with $T_{eq} <$ 1000 K, from TEPCat [@southworth2011].[]{data-label="fig:mr_diagram"}](MR_plot.pdf){width="\linewidth"} $ \begin{array}{cc} \includegraphics[width=0.5\textwidth]{k2_ternary.pdf} \includegraphics[width=0.5\textwidth]{core_gas.pdf} \end{array}$ Internal Structure and Composition {#sec:interior} ================================== In this section we use the observed bulk properties of K2-18b, namely the planetary mass ($M_p$), radius ($R_p$), and its atmospheric properties, to constrain its internal structure and thermodynamic conditions. Internal structure model {#sec:interior_model} ------------------------ We model the interior of the planet with a canonical four-layer structure. The model comprises a two-component Fe+rock core consisting of an inner Fe layer and an outer silicate layer, a layer of H$_2$O, and an outer H/He envelope. Such a model spans the possible internal structures and compositions of super-Earths and mini-Neptunes [e.g. @valencia2010; @valencia2013; @rogers2011; @lopez2014], as well as terrestrial planets and ice giants in the solar system [@guillot2014]. The mass fractions of the different components ($x_{\rm Fe}$, $x_{\rm rock}$, $x_{\rm H_2O}$, $x_{\rm env}$) are free parameters in the model and sum to unity. Our present model is adapted from a three-layer model for super-Earths from @madhu2012 comprising of Fe, rock, and H$_2$O, with the H/He envelope added in the present work. The model solves the standard internal structure equations of hydrostatic equilibrium and mass continuity assuming spherical symmetry. The equation of state (EOS) for each of the two inner layers is adopted from @seager2007 who use the Birch-Murnaghan EOS [@birch1952] for Fe [@ahrens2000] and MgSiO$_3$ perovskite [@karki2000]. For the H$_2$O layer we use the temperature-dependent H$_2$O EOS compiled by @thomas2016 from @french2009 [@sugimura2010; @fei1993; @seager2007] and @wagner2002. For the gaseous envelope we use the latest H/He EOS from @chabrier2019 for a solar helium mass fraction ($Y = 0.275$). The EOS in the H/He and H$_2$O layers can have a significant temperature dependence which we consider in our model. Past studies [@rogers2011; @valencia2013] considered analytic $P$-$T$ profiles for irradiated atmospheres derived using double gray approximations [@hansen2008; @guillot2010] with the internal and external fluxes and opacities as free parameters. We calculate self-consistent dayside $P$-$T$ profiles for K2-18b in the H/He envelope using the GENESIS code [@gandhi2017]. GENESIS solves line-by-line radiative transfer under assumptions of hydrostatic, radiative-convective and thermochemical equilibrium. We include opacity due to H$_2$O [@rothman2010], as detected in the transmission spectrum (section \[sec:atmosphere\]), H$_2$ Rayleigh scattering, clouds and H$_2$-H$_2$ and H$_2$-He collision-induced absorption. We use a H$_2$O abundance of 10$\times$solar (see section \[sec:atmosphere\]) and also use 10$\times$solar abundances for the cloud species. We include KCl, ZnS and Na$_2$S clouds [@morley2013], for which we obtain opacities from @pinhas2017. We further include water ice clouds using opacities from @Budaj2015. The $P$-$T$ profile also depends on the planetary internal flux, which is characterised by the internal temperature $T_\mathrm{int}$. We consider values of $T_\mathrm{int}$ which span the range expected for a planet with the mass and radius of K2-18b and an age of $1-10$ Gyr, with envelope compositions from solar to water-rich. We choose end-member cases of $T_\mathrm{int}=25\text{K}$ and $50\text{K}$, consistent with previous studies on planets of similar mass and radius, e.g., GJ 1214b [e.g., @valencia2013]. The GENESIS models are calculated between pressures of $10^{-5}-10^3$ bar, and assume full redistribution of the incident stellar irradiation. We explore a range of $P$-$T$ profiles and choose two representative cases, with different $T_\mathrm{int}$, discussed further in sections \[sec:interior\_constraints\] and \[sec:hhb\]. Where required by the internal structure model, the bottom of the $P$-$T$ profile of the H/He envelope is continued to deeper pressures using the adiabatic gradient from @chabrier2019. We also employ an adiabatic temperature profile in the H$_2$O layer, following @thomas2016. Constraints on interior composition {#sec:interior_constraints} ----------------------------------- Figure \[fig:mr\_diagram\] shows mass-radius relations for models with different interior compositions. We explore the full range of plausible interior compositions in three components: $x_{\rm core} = x_{\rm Fe} + x_{\rm rock}$, $x_{\rm H_2O}$, and $x_{\rm env}$, where $x_i$ = $M_i/M_p$ is the mass fraction of each component $i$. For each atmospheric $P$-$T$ profile considered, we explore two different core compositions: (1) an Earth-like core made of 33% Fe, 67% rock by mass, and (2) a pure Fe core, the densest possible composition. Here, we discuss results from two end-member cases: (1) a pure Fe core with $T_{\rm int} = 25$K, and (2) an Earth-like (33% Fe) core with $T_{\rm int} = 50$K. Solutions for all other cases lie between these two cases. As shown in Figure \[fig:ternary\], while a wide range of core and H$_2$O mass fractions are permitted, we place a stringent upper limit on the mass fraction of the H/He envelope: $x_{\rm env} = 6.2$%. This maximal $x_{\rm env}$ corresponds to the case of a pure Fe core, with $x_{\rm core}\sim94$%, underlying the H/He envelope with no $x_{\rm H_2O}$; here it is assumed that the atmospheric H$_2$O is not mixed in the envelope. However, if the retrieved atmospheric H$_2$O abundance is assumed to be well mixed in the envelope then the maximal $x_{\rm env} = 6$% with $x_{\rm H_2O}=0.4$% by mass; low, but still significantly higher than that of the Earth’s oceans ($\sim$0.02%). We find that a substantial gaseous H/He envelope is not necessary to explain the density of K2-18b. Figure \[fig:ternary\] shows the $x_{\rm env}$ required for different $x_{\rm core}$. At one extreme, a $\sim$100% H$_2$O interior with no rocky core can explain the data with an $x_{\rm env}$ of just $\sim$10$^{-6}$, comparable to the mass fraction of the Earth’s atmosphere. The presence of a rocky core would necessitate at least a thin H/He envelope. However, even considering a reasonable $x_{\rm core} = 10-50$% still requires $x_{\rm env}$ of only $\sim 10^{-5}-10^{-2}$, as shown in Figure \[fig:ternary\]. Model solutions with the hotter $P$-$T$ profile and/or lower Fe content in the core require smaller $x_{\rm env}$ for a given $x_{\rm core}$. We have also considered models with miscible H$_2$O and H/He envelopes. We follow the approach of @soubiran2015, using an additive volume law for mixtures. Assuming that the median H$_2$O mixing ratio in the atmosphere is representative of the mixed (H$_2$O-H/He) envelope, we find that the difference in radius between the mixed and non-mixed models is less than half of the measured uncertainty (see Figure \[fig:mr\_diagram\]). The constraint on the envelope mass fraction from this mixed case is $x_{\rm env} = 2.5-6.4$%, consistent with, and a subset of, the constraints discussed above. Note that in this case $x_{\rm env}$ includes both the H/He and H$_2$O mass fraction. $ \begin{array}{cc} \includegraphics[width=0.5\textwidth]{profile.pdf} \includegraphics[width=0.5\textwidth]{phase_diagram.pdf} \end{array}$ Atmosphere-Ocean Boundary {#sec:hhb} ------------------------- Our constraints on the interior compositions of K2-18b result in a wide range of thermodynamic conditions at the H$_2$O-H/He boundary (HHB). The pressure ($P_{\rm HHB}$) and Temperature ($T_{\rm HHB}$) at the HHB for the model solutions are shown in Figure \[fig:phase\_diagram\]. Each point on the HBB loci denotes the transition from the $P$-$T$ profile in the H/He envelope to the corresponding H$_2$O adiabat. The $P_{\rm HHB}$ and $T_{\rm HHB}$ depend on the H/He envelope mass fraction. For a given $P$-$T$ profile, larger envelopes result in higher $P_{\rm HHB}$ and $T_{\rm HHB}$. For example, solutions with $x_{\rm env} \gtrsim 1$% lead to $P_{\rm HHB}$ and $T_{\rm HHB}$ corresponding to the super-critical phase of H$_2$O. As shown in Figure \[fig:ternary\], solutions with higher $x_{\rm env}$ correspond to higher $x_{\rm core}$ and lower $x_{\rm H_2O}$. Conversely, solutions with lower $x_{\rm core}$, and hence lower $x_{\rm env}$ and higher $x_{\rm H_2O}$, lead to lower $P_{\rm HHB}$ and $T_{\rm HHB}$ with H$_2$O in vapour or liquid phases at the HHB. For example, an $x_{\rm core}\lesssim$ 30% leads to a $P_{\rm HHB}$ and $T_{\rm HHB}$ corresponding to the liquid phase of H$_2$O, for the cooler $P$-$T$ profile (with $T_{\rm int}=25$K). For $x_{\rm core}\sim$10% or less, the $P_{\rm HHB}$ and $T_{\rm HHB}$ approach STP conditions for liquid H$_2$O. Below the HHB, H$_2$O is found in increasingly dense phases spanning liquid, vapour, super-critical, and ice states depending on the location of the HHB and the extent of the H$_2$O layer, as shown in Figure \[fig:phase\_diagram\]. In the case of a mixed H$_2$O-H/He envelope, the HHB is undefined as it corresponds to an extreme case with no pure H$_2$O layer. Discussion {#sec:discussion} ========== Our constraints on the interior and atmospheric properties of K2-18b provide insights into its physical conditions, origins, and potential habitability. Possible Compositions and Origins {#sec:scenarios} --------------------------------- Here we discuss three representative classes that span the range of possible compositions, as indicated in Figures \[fig:mr\_diagram\], \[fig:ternary\] and \[fig:phase\_diagram\]. The specific cases chosen here fit the $M_{\rm p}$ and $R_{\rm p}$ exactly, as shown in figure \[fig:mr\_diagram\]. A wider range of solutions exist in each of these classes within the 1$\sigma$ uncertainties. [*Case 1: Rocky World.*]{} One possible scenario is a massive rocky interior overlaid by a H/He envelope. For example, a pure Fe core of 94.7% by mass with an almost maximal H/He envelope of 5% explains the data with minimal $x_{\rm H_2O}=0.3\%$, consistent with our retrieved H$_2$O abundance in the atmosphere. The HHB in this case is at $\sim 10^{6}$ bar, yielding supercritical H$_2$O close to the ice X phase. It is also possible in this case that the H$_2$O and H/He are mixed, meaning the HHB is not well-defined. Such a scenario is consistent with either H$_2$ outgassing from the interior [@elkins-tanton2008; @rogers2010] or accretion of an H$_2$-rich envelope during formation [@lee2016]. [*Case 2: Mini-Neptune.*]{} There are a range of plausible compositions consisting of a non-negligible H/He envelope in addition to significant H$_2$O and core mass fractions, akin to canonical models for Neptune and Uranus [@guillot2014]. One such example is a 45% Earth-like core with $x_{\rm env }=0.03\%$ and $x_{\rm H_2O}=54.97\%$. In this case the HHB is at $P_{\rm HHB} = 700$bar and $T_{\rm HHB}=1500$K, with H$_2$O in the supercritical phase. [*Case 3: Water World.*]{} A $\sim$100% water world with a minimal H$_2$-rich atmosphere ($x_{\rm env} \sim 10^{-6}$) is permissible by the data. However, such an extreme case is implausible from a planet formation perspective; some amount of rocky core is required to initiate further ice and gas accretion [@leger2004; @rogers2011; @lee2016]. For example, a planet with $x_{\rm core} = 10\%, x_{\rm H_2O} = 89.994\%$ and a thin H/He envelope ($x_{\rm env} = 0.006\%$) can explain the data. For this case, $P_{\rm HHB} = 130$bar and $T_{\rm HHB}=560$K, corresponding to liquid H$_2$O. For the same core fraction, solutions with even smaller H/He envelopes are admissible within the 1$\sigma$ uncertainties on $M_p$ and $R_p$, leading to $P_{\rm HHB}$ and $T_{\rm HHB}$ approaching habitable STP conditions. Potential Habitability ---------------------- A notional definition of habitability argues for a planetary surface with temperatures and pressures conducive to liquid H$_2$O [e.g., @kasting1993; @meadows2018]. Living organisms are known to thrive in Earth’s extreme environments (extremophiles). Their living conditions span the phase space of liquid H$_2$O up to $\sim$1000 bar pressures at the bottom of the Marianas Trench and $\sim$400 K temperatures near hydrothermal vents [e.g., @merino2019]. Whether or not habitable conditions prevail on K2-18b depends on the extent of the H/He envelope. The thermodynamic conditions at the surface of the H$_2$O layer span a wide range in the H$_2$O phase diagram. While most of these solutions lie in the super-critical phase, many others lie in the liquid and vapour phases. Model solutions with core mass fractions $<$15% and H/He envelopes $\lesssim10^{-3}$ allow for liquid H$_2$O at Earth-like habitable conditions discussed above. One plausible scenario is an ocean world, as discussed in section \[sec:scenarios\], with liquid water approaching STP conditions (e.g., 300 K, $\sim$1-10 bar) underneath a thin H/He atmosphere ($x_{\rm env}\lesssim$10$^{-5}$). A number of studies in the past have argued for potential habitability on planets with H/He-rich atmospheres orbiting M Dwarfs [e.g., @pierrehumbert2011; @seager2013; @koll2019]. Given our constraints above, we find that K2-18b has a realistic chance of being habitable. Furthermore, our constraints on CH$_4$ and NH$_3$ suggest chemical disequilibrium. Among other possibilities for chemical disequilibrium, e.g. photochemistry, the potential influence of biochemical processes may not be entirely ruled out. Future observations, e.g. with the *James Webb Space Telescope*, will have the potential to refine our findings. We argue that planets such as K2-18b can indeed have the potential to approach habitable conditions and searches for biosignatures should not necessarily be restricted to smaller rocky planets. N.M., M.N., A.P., R.B. acknowledge support from the UK Science and Technology Facilities Council (STFC). L.W. thanks the Gates Cambridge Trust for support toward his doctoral studies. We thank the anonymous reviewer for their helpful comments on the manuscript. This research is made open access thanks to the Bill & Melinda Gates foundation. natexlab\#1[\#1]{}\[1\][[\#1](#1)]{} \[1\][doi: [](http://doi.org/#1)]{} \[1\][[](http://ascl.net/#1)]{} \[1\][[](https://arxiv.org/abs/#1)]{} , T. J., ed. 1995, Mineral Physics [&]{} Crystallography: A Handbook of Physical Constants (American Geophysical Union), , G., [Amado]{}, P. J., [Barnes]{}, J., [et al.]{} 2016, , 536, 437, , S. K., [Crida]{}, A., [Guillot]{}, T., [et al.]{} 2018, [The Origin and Evolution of Saturn, with Exoplanet Perspective in, Saturn in the 21st Century]{}, ed. K. H. Baines [et al.]{} (Cambridge: Cambridge Univ. Press), 5–43, , R. J., [Strange]{}, J. K., [Hill]{}, C., [et al.]{} 2014, , 437, 1828, , B., [Werner]{}, M., [Petigura]{}, E., [et al.]{} 2017, , 834, 187, , B., [Wong]{}, I., [Piaulet]{}, C., [et al.]{} 2019, , 887, L14, , F. 1952, , 57, 227, , J., [Georgakakis]{}, A., [Nandra]{}, K., [et al.]{} 2014, , 564, A125, , J., [Kocifaj]{}, M., [Salmeron]{}, R., & [Hubeny]{}, I. 2015, , 454, 2, , A., & [Sharp]{}, C. M. 1999, , 512, 843, , G., [Mazevet]{}, S., & [Soubiran]{}, F. 2019, , 872, 51, , R., [Astudillo-Defru]{}, N., [Doyon]{}, R., [et al.]{} 2017, , 608, A35, —. 2019, , 621, A49, , J. A., [Irwin]{}, J. M., [Charbonneau]{}, D., [et al.]{} 2017, , 544, 333, , C. D., & [Charbonneau]{}, D. 2015, , 807, 45, , L. T., & [Seager]{}, S. 2008, , 685, 1237, Fei, Y., Mao, H., & Hemley, R. J. 1993, The Journal of Chemical Physics, 99, 5369, , F., [Hobson]{}, M. P., & [Bridges]{}, M. 2009, , 398, 1601, , D., [Montet]{}, B. T., [Hogg]{}, D. W., [et al.]{} 2015, , 806, 215, , M., [Mattsson]{}, T. R., [Nettelmann]{}, N., & [Redmer]{}, R. 2009, , 79, 054107, , S., & [Madhusudhan]{}, N. 2017, , 472, 2334, , M., [Triaud]{}, A. H. M. J., [Demory]{}, B.-O., [et al.]{} 2017, , 542, 456, , T. 2010, , 520, A27, , T., & [Gautier]{}, D. 2015, in Treatise on Geophysics, 2nd edn., ed. G. Schubert (Oxford: Elsevier), 529 – 557, , B. M. S. 2008, , 179, 484, Karki, B. B., Wentzcovitch, R. M., de Gironcoli, S., & Baroni, S. 2000, Phys. Rev. B, 62, 14750, , J. F., [Whitmire]{}, D. P., & [Reynolds]{}, R. T. 1993, , 101, 108, , D. D. B., & [Cronin]{}, T. W. 2019, , 881, 120, , E. J., & [Chiang]{}, E. 2016, , 817, 90, , A., [Selsis]{}, F., [Sotin]{}, C., [et al.]{} 2004, , 169, 499, , E. D., & [Fortney]{}, J. J. 2014, , 792, 1, , N., [Lee]{}, K. K. M., & [Mousis]{}, O. 2012, , 759, L40, , V. S., & [Barnes]{}, R. K. 2018, [Factors Affecting Exoplanet Habitability]{} (Springer International Publishing), 57, Merino, N., Aronson, H. S., Bojanova, D. P., [et al.]{} 2019, Frontiers in Microbiology, 10, 780, , B. T., [Morton]{}, T. D., [Foreman-Mackey]{}, D., [et al.]{} 2015, , 809, 25, , C. V., [Fortney]{}, J. J., [Kempton]{}, E. M. R., [et al.]{} 2013, , 775, 33, , G. D., [Pascucci]{}, I., & [Apai]{}, D. 2015, , 814, 130, , N., [Fortney]{}, J. J., [Kramm]{}, U., & [Redmer]{}, R. 2011, , 733, 2, , R., & [Gaidos]{}, E. 2011, , 734, L13, , A., & [Madhusudhan]{}, N. 2017, , 471, 4355, , A., [Madhusudhan]{}, N., [Gandhi]{}, S., & [MacDonald]{}, R. 2019, , 482, 1485, , C., [Gordon]{}, I. E., [Rothman]{}, L. S., [et al.]{} 2012, , 113, 1276, , L. A., [Bodenheimer]{}, P., [Lissauer]{}, J. J., & [Seager]{}, S. 2011, , 738, 59, , L. A., & [Seager]{}, S. 2010, , 716, 1208, , L. S., [Gordon]{}, I. E., [Barber]{}, R. J., [et al.]{} 2010, , 111, 2139, , S., [Bains]{}, W., & [Hu]{}, R. 2013, , 777, 95, , S., [Kuchner]{}, M., [Hier-Majumder]{}, C. A., & [Militzer]{}, B. 2007, , 669, 1279, , F., & [Militzer]{}, B. 2015, , 806, 228, , J. 2011, , 417, 2166, Sugimura, E., Komabayashi, T., Hirose, K., [et al.]{} 2010, Phys. Rev. B, 82, 134103, , S. W., & [Madhusudhan]{}, N. 2016, , 458, 1330, , R. 2008, Contemporary Physics, 49, 71, , A., [Waldmann]{}, I. P., [Tinetti]{}, G., [Tennyson]{}, J., & [Yurchenko]{}, S. N. 2019, NatAs, 3, 1086, , D., [Guillot]{}, T., [Parmentier]{}, V., & [Freedman]{}, R. S. 2013, , 775, 10, , D., [Ikoma]{}, M., [Guillot]{}, T., & [Nettelmann]{}, N. 2010, , 516, A20, , W., & [Pru[ß]{}]{}, A. 2002, Journal of Physical and Chemical Reference Data, 31, 387, , L., & [Madhusudhan]{}, N. 2019, , 157, 206, , L., [Madhusudhan]{}, N., [Allard]{}, N. F., [et al.]{} 2019, , 887, L20, , S. N., [Barber]{}, R. J., & [Tennyson]{}, J. 2011, , 413, 1828, , S. N., & [Tennyson]{}, J. 2014, , 440, 1649,
{ "pile_set_name": "ArXiv" }
--- abstract: 'We review the calculation of the prompt lepton flux, produced in the atmosphere by the semileptonic decay of charmed particles. We describe side by side the intermediary ingredients used by different authors, which include not only the charm production model, but also other atmospheric particle showering parameters. After evaluating separately the relevance of each single ingredient, we analyze the effect of different combinations over the final result. We highlight the impact of the prompt lepton flux calculation upon high-energy neutrino telescopes.' address: 'Service de Physique Théorique, CP 225, Université Libre de Bruxelles, Boulevard du Triomphe, 1050 Brussels, Belgium' author: - 'C.G.S. Costa' title: The Prompt Lepton Cookbook --- Cosmic Rays ,High Energy ,Charm ,Muon ,Neutrino 13.85.Tp ,96.40.Tv Introduction ============ Very-high-energy (above 1 TeV) neutrino astronomy is currently a subject of great interest, promising to expand our observational range of the Universe in an unique way[@reports]. Such energetic neutrinos may carry information from the sources of the highest energy phenomena ever observed in cosmic rays, possibly coming from active galactic nuclei (AGN) or gamma ray bursts (GRB). They may probe the early stages of the Universe and its farthest distances. In addition, they will contribute to the search of weakly interacting massive particles (WIMP), supernova explosions, monopoles, besides the discovery potential for new physics, which we can even not imagine yet. Neutrino telescopes under development, like the Antarctic Muon and Neutrino Detector Array (AMANDA) and the experiment at Lake Baikal, are already operational, and producing their first results[@amanda:ap00; @baikal:ap00]. In addition, great activity is planned for the near future[@spiering:nu00]. Aside from these perspectives, the operation of a neutrino detector at energies above 1 TeV poses challenging difficulties. One of the major limitations in the detection of a cosmic high-energy neutrino (from galactic or extra-galactic origin) is the background from atmospheric muons and neutrinos (produced by the interaction of high-energy cosmic rays in the atmosphere). The source of the atmospheric neutrino background changes with energy, in a way governed by the critical energy $\varepsilon_{critic}$ of the parent particle. This is the energy for which the decay and interaction lengths are equal. Above this energy the parent particle is likely to interact or be slowed down before decaying into a neutrino. The critical energy is calculated in terms of the particle rest energy $mc^2$, the mean life $\tau$ and, by adopting the isothermal atmosphere approximation, a scale constant $h_{o}$[@gaisser:92]: $$\varepsilon_{critic} \: = \: \frac{mc^2}{c\tau} \: h_{o}. \label{eq:epsilon}$$ Table \[table:pdg\] summarizes several particle properties, derived from the Review of Particle Physics [@pdg:00]. Comparing the critical energies we observe that muon decays contribute substantially to the atmospheric lepton flux only up to a few GeV’s, while the decays in flight of charged pions and kaons are still important up to about 1-10 TeV. They give rise to the “conventional” atmospheric lepton flux. Above this energy range, the semileptonic decay of very-short lived charmed particles (mainly $D$-mesons and $\Lambda_{c}^{+}$-hyperons) is the dominant source, despite their low production rate. The main contribution comes from the decay modes $$D \rightarrow K + \mu + \nu \:\: {\rm and} \:\: \Lambda_{c} \rightarrow \Lambda_{o} + \mu + \nu .$$ Muons and neutrinos thus generated are called “prompt leptons”, and they exhibit a flatter (and thus harder) energy spectrum. The lack of precise information on high-energy charm production in hadron-nucleus collisions leads to a great uncertainty in the estimate of the leptonic flux above 100 TeV. In addition, different authors do not use the same atmospheric particle showering routines, turning the comparison even more difficult. The predictions of resulting fluxes span up to four orders of magnitude! It is our purpose to bring some light to the forum of prompt lepton fluxes, describing side by side the many ingredients of the calculation. After evaluating separately the relevance of each single shower parameter, we analyze the effect of different combinations over the final result. We investigate the fluxes of $\mu$, $\nu_{\mu}$ and $\nu_{e}$, leaving the case of $\tau$ and $\nu_{\tau}$ for a further work. [lcrccr]{} Particle &Elementary &$mc^2$ &$c\tau$ &$\varepsilon_{critic}$$^{(1)}$ &Branching\ &contents &(MeV) &&(GeV) &ratio $B_{i}$$^{(2)}$\ $D^{+},D^{-}$ &$c\bar{d},\bar{c}d$ &1870 &317 $\mu$m &$3.8 \times 10^7$ &17.2 %\ $D^{o},\bar{D}^{o}$ &$c\bar{u},\bar{c}u$ &1865 &124 $\mu$m &$9.6 \times 10^7$ &6.8 %\ $D_{s}^{+},D_{s}^{-}$ &$c\bar{s},\bar{c}s$ &1969 &149 $\mu$m &$8.5 \times 10^7$ &5.2 %\ $\Lambda_{c}^{+}$ &$udc$ &2285 &62 $\mu$m &$2.4 \times 10^8$ & 4.5 %\ $\mu^{+},\mu^{-}$ &lepton &106 &659 m &1.0 &100 %\ $\pi^{+},\pi^{-}$ &$u\bar{d},\bar{u}d$ &140 &7.8 m &115 &100 %\ $K^{+},K^{-}$ &$u\bar{s},\bar{u}s$ &494 &3.7 m &855 &63.5 %\ $\Lambda^{o}$ &$uds$ &1116 &7.9 cm &$9.0 \times 10^4$ &0.1 %\ \ Calculation of the prompt lepton flux {#section:calculation} ===================================== The calculation of the prompt lepton flux has been carried out in the past (see, e.g. Ref.[@bugaev:98] and references therein), mainly with the purpose to investigate the effects of choosing a different charm production model. The outline of the calculation is basically the same in all of these works. We start from the primary cosmic ray flux at the top of the atmosphere, with a composition supposed to be dominated by protons, and evaluate the flux of nucleons at any atmospheric depth. Those nucleons interact with the nuclei of air to produce secondary particle showers. For energies above a few TeV, the secondary particles of interest to be followed are the charmed hadrons, for they will be the main source of high energy atmospheric leptons. We will, therefore, consider the contribution of mesons $D^{\pm}, D^{o}$ and $\bar{D}^{o}$ (comprising in the same notation the overall contribution of $D$ and $D^{*}$ mesons), of the strange $D_{s}^{\pm}$ and of the $\Lambda_{c}^{+}$-hyperon. Because the production rate of the mesons $D_{s}$ is relatively lower (about $20 \%$ of the $D$ production cross section), some authors neglect their contribution, although their critical energy and branching ratio are comparable to those of the other charmed particles (see Table \[table:pdg\]). When calculating the flux of $\tau$ and $\nu_{\tau}$, which we do not consider in this analysis, the role of $D_{s}$ turns out to be mostly important[@pr:99]. It is straightforward to calculate the flux of charm particles at any depth, and they will promptly decay yielding electrons, muons and neutrinos. We may integrate the flux for all possible charmed parent production and decay depths, and for all possible production and decay energies, leading to the flux of a chosen lepton at a given depth and energy. For detailed calculations we refer to Refs.[@bugaev:98; @volkova:87; @halzen:93], and we follow the notation of the latter, to present here only main results. Let’s write the primary cosmic-ray spectrum as a power law in energy: $$\Phi_{N} (E_{N},x=0) \: = \: N_{o} \: E_{N}^{-(\gamma+1)}, \label{eq:Fprime}$$ where $\Phi_{N}(E_{N},x)$, given in (GeV.cm$^2$.s.sr)$^{-1}$, is the differential flux of nucleons with energy $E_{N}$, in GeV, and $x$ is the slant depth penetrated by the cascade, measured in g/cm$^2$ from the top of the atmosphere ($x=0$) downward along the direction of the incident nucleon. The constant $N_{o}$ is the amplitude, or differential spectrum normalization; and $\gamma$ is the spectral index, or slope of the integral primary spectrum. After developing to a certain depth $x$, the nucleonic flux is given in terms of $\Lambda_{N}$, the nucleonic attenuation length[@gaisser:92]: $$\Phi_{N}(E_{N},x,\theta)\: = \: N_{o} \: E_{N}^{-(\gamma+1)} \: e^{-x/\Lambda_{N}}. \label{eq:Fnucx}$$ The resulting flux of secondary particles of type-$i$ ($i = D^{\pm}, D^{o}, \bar{D}^{o}, D_{s}^{\pm}, \Lambda_{c}^{+}$) is calculated by convolution of the nucleonic flux with the production spectrum of secondary particles: $$\Phi_{i}(E_{i},x,\theta) \: = \: K_{i} (E_{i},\gamma) \: \int_{0}^{x} dx' \: \left(\frac{x'}{x}\right)^\eta \: \exp\left\{-\frac{(x-x')}{\lambda_{i}} \:-\frac{x'}{\Lambda_{N}}\right\}, \label{eq:Fix}$$ defining $$\eta \: = \: \frac{\varepsilon_{critic}}{E_{i}\:\cos\!\theta},$$ where the dependence in the zenith angle holds for $\theta \leq 60^{o}$. For higher zenith angles the curvature of Earth must be taken into account. Both the nucleonic attenuation length $\Lambda_{N}$ and the charmed particle interaction length $\lambda_{i}$ are given in units g/cm$^2$. The production spectrum of charmed particles, weighted by the primary nucleonic spectrum, is written as: $$\begin{aligned} K_{i} (E_{i},\gamma) & = & \int_{E_{i}}^{\infty} dE_{N} \: \: \frac{N_{o}}{\lambda_{N}} \: E_{N}^{-(\gamma+1)} \: \frac{dW^{Ni}(E_{i},E_{N})}{dE_{i}}, \label{eq:Ki}\\ & = & \frac{N_{o}}{\lambda_{N}} \: E_{i}^{-(\gamma+1)} \: Z_{Ni}(\gamma). \label{eq:KiZi}\end{aligned}$$ In this notation, $dW^{Ni}/dE_{i}$ denotes the energy distribution of secondary particles, and represents the probability that a particle of type-$i$ with energy $E_{i}$ is created in the interaction of an incident nucleon $N$ of energy $E_{N}$ with an air nucleus. It is directly related to the inclusive cross section for secondary particle production. Eq. (\[eq:KiZi\]) is obtained assuming a mild energy dependence for the nucleonic interaction length $\lambda_{N}$, and defining $Z_{Ni}(\gamma)$, the particle production spectrum-weighted moment[@gaisser:92], also called production “Z-moment”: $$Z_{Ni}(\gamma) = \int_{0}^{1} x_{F}^{\! \gamma} \left( \frac{dW^{Ni}}{dx_{F}}\right) \: dx_{F} , \label{eq:Zch}$$ where $x_{F}$ is the Feynman variable $x_{F} \equiv p_{L}/p_{L}^{max}$, with $p_{L}$ as the produced particle longitudinal momentum. At the high-energy limit the Feynman-$x$ also represents the ratio of the final particle energy to the incident particle energy, $x_{F}=E_{i}/E_{N}$ (beware confusion with atmospheric depth $x$). In order to evaluate the flux $\Phi_{l}(E_{l},x,\theta)$ of leptons ($l=\mu$ or $\nu$), with energy $E_{l}$ and zenith angle $\theta$ at depth $x$, we need to fold the energy distribution $df^{l}/dE_{l}$ of the produced lepton with the spectrum $D_{i}(E_{i},x'',\theta)$ of decaying parents, for any decay depth $x''$ and any available parent energy $E_{i}$: $$\Phi_{l}(E_{l},x,\theta) \:=\: \int_{0}^{x} dx'' \:\: \int_{E_{i}^{max}}^{E_{i}^{min}} dE_{i} \:\: \frac{df^{l}}{dE_{l}} \:\: D_{i}(E_{i},x'',\theta), \label{eq:Fnil}$$ with $$\begin{aligned} D_{i}(E_{i},x'',\theta) &=& B_{i} \: \frac{1}{d_{i}} \: \Phi_{i}(E_{i},x'',\theta) , \label{eq:Di}\\ d_{i} &=& \frac{x''}{\eta} \:=\: \frac{x'' \cos\!\theta \:E_{i}}{\varepsilon_{critic}} \label{eq:decay}\end{aligned}$$ where $B_{i}$ is the branching ratio yielding leptons in the parent-$i$ decay (see Table \[table:pdg\]), and $d_{i}$ is the particle-$i$ decay length. The muon and neutrino production energy distributions $df^{l}/dE_{l}$ used in Eq. (\[eq:Fnil\]) are given by the semileptonic three-body decay phase space integrals, obtained from kinematics considerations[@gaisser:92; @hagedorn:63]. Some authors[@lipari:93; @tig:96; @prs:99] define a decay Z-moment in analogy to the production Z-moment, Eq. \[eq:Zch\], just replacing $dW^{Ni}/dE_{i}$ by $df^{l}/dE_{l}$. Making use of this definition, it is possible to write an approximate solution (valid for energies $E_{l} <\varepsilon_{critic}$) exploring the fact that the critical energy for charmed particles is very high. Nevertheless, in the present work we will use the complete solution for calculating the prompt lepton flux, given by the set of Equations (\[eq:Fix\]) to (\[eq:decay\]). Showering Process {#section:showering} ================= Once the calculation is established, the next step is to choose the ingredients that characterize the showering process in the atmosphere. We will browse through the literature to extract different parametrizations to be compared. The main parameters to define are: the primary spectrum normalization and slope, the nucleonic and charm interaction lengths, the nucleonic attenuation length and the charm production Z-moment. Primary spectrum {#section:primes} ---------------- The primary cosmic ray flux at the top of the atmosphere, Eq. (\[eq:Fprime\]), can be rewritten as to incorporate the effect of the change in slope (“knee”) observed in the energy spectrum, at energy $E_{knee}$: $$\begin{aligned} \Phi_{N} (E_{N}, x=0) & = & N_{1} \: E_{N}^{-(\gamma_{1}+1)}, \:\: E_{N} <E_{knee} \; \mbox{{\rm (in GeV)}}; \nonumber \\ & = & N_{2} \: E_{N}^{-(\gamma_{2}+1)}, \:\: E_{N} >E_{knee} \; \mbox{{\rm (in GeV)}}. \label{eq:primes}\end{aligned}$$ Table \[table:primes\] indicates some typical values for the parameters in Eq. (\[eq:primes\]), used by different authors. Lipari[@lipari:93] quotes a parametrization consistent with both the JACEE balloon borne experiments[@jacee:90] and the values given by Gaisser[@gaisser:92], adopting a single slope, because his analysis is mainly aimed at energies below the knee. The AKENO experiment[@akeno:84] obtained a description of the primary spectrum, covering the knee region, from data on size spectra of electrons and muons at high energy. Bugaev [*et al.*]{}[@bugaev:89] use a semiempirical model which takes into account detailed chemical composition of the primary spectrum, translated here in terms of Eq. (\[eq:primes\]). They propose two options (Model F and Model D), differing on the hypothesis responsible for the change in slope at the knee. Thunman, Ingelman and Gondolo (TIG) [@tig:96] follow the JACEE trend below the knee. Above the knee they adopt the same slope as Volkova [*et al.*]{}[@volkova:87], for which no normalization constant is reported, due to their interest limited to flux ratios. Figure \[fig:primes\] displays the energy spectra corresponding to those parametrizations. [lrrrrr]{} Label & &$E_{knee}$ &\ &$N_{1}$ &$\gamma_{1}$ &(GeV) &$N_{2}$ &$\gamma_{2}$\ Lipari &1.70 &1.70 & - & - & -\ Akeno &1.35 &1.62 &$4.67 \times 10^{6}$ &630 &2.02\ Bugaev (F) &1.02 &1.62 &$1.9 \times 10^{6}$ &323 &2.02\ Bugaev (D) &1.02 &1.62 &$5.2 \times 10^{5}$ &193 &2.02\ TIG &1.70 &1.70 &$5 \times 10^{6}$ &174 &2.00\ \ Interaction and attenuation lengths {#section:sigmas} ----------------------------------- The nucleonic interaction length, $\lambda_{N}$, represents the mean free path of nucleons in the atmosphere (given in g/cm$^2$). It is related to $\sigma_{in}^{\! N-air}$, the total inelastic cross section for collisions of nucleons $(N)$ with air nuclei, through the relation $$\lambda_{N} (E) = \frac{A. m_{p}}{\sigma_{in}^{\! N-air} (E)} = \frac{24100}{\sigma_{in}^{\! N-air} (E)}, \label{eq:lnuc}$$ where we used the average atomic number for air nuclei $A$=14.5, $m_{p}$ is the proton mass and $\sigma_{in}^{\! N-air}$ must be given in $mb$. There are different parametrizations to the inelastic $N-$air cross section as a function of energy. Some authors considered it to be constant[@volkova:87; @lipari:93]; others to be rising with energy: as a power-law[@akeno:84], as a logarithmic dependence[@bugaev:98; @bugaev:89], or as a log-squared dependence[@chs:95; @bhs:00]. The attenuation length, $\Lambda_{N}$, which governs the exponential decay of the nucleonic flux with increasing depth, see Eq. (\[eq:Fnucx\]), represents the net effect in the interplay between interaction losses and regeneration of the number of nucleons in the cascade development, and is given by $$\Lambda_{N} \: = \: \frac{\lambda_{N}}{\left( 1-Z_{NN}(\gamma) \right)}. \label{eq:Attnuc}$$ The interaction length $\lambda_{N}$ dictates the losses and the nucleon-to-nucleon spectrum-weighted Z-moment $Z_{NN}(\gamma)$ accounts for the survival rate of nucleons. $Z_{NN}$ is calculated in analogy to Eq. (\[eq:Zch\]), with the outgoing particles being the regenerated nucleons. Some authors use an approximately constant value[@bugaev:98; @lipari:93; @bhs:00], others adopt Feynman-scaling with a change in value at the knee energy[@gaisser:92; @chs:95], while in Ref. [@tig:96] it is assumed the violation of Feynman-scaling. The charm particle of type-$i$ may have an interaction length $\lambda_{i}$ in the atmosphere calculated analogously as for the nucleons, Eq. (\[eq:lnuc\]), substituting $\sigma_{in}^{\! N-air}$ by $\sigma_{in}^{\! i-air}$. As in the parametrization of the nucleonic inelastic cross section, we find authors adopting a charm cross section which is constant[@volkova:85], or which increases with energy either as a power-law[@mm:86], or logarithmically[@bugaev:98]. Charm production spectrum-weighted moments {#section:zch} ------------------------------------------ The key information for the evaluation of the prompt lepton flux is the behavior of the charm spectrum-weighted moments, given by a specific charm production model. Among these models, we select three different ones, to be compared in the present study. QGSM: : Quark Gluon String Model, a semiempirical model of charm production based on the non-perturbative QCD calculation by Kaidalov and Piskunova[@kaidalov:86], normalized to accelerator data, and applied to the prompt muon calculation by Volkova [*et al.*]{}[@volkova:87]. The Z-moments are calculated numerically from Eq. (\[eq:Zch\]), with $$\frac{dW^{Ni}}{dx_{F}} = \frac{\sigma_{NA}^{\! i}}{\sigma_{in}^{\! N-air}} \: \frac{df_{i}}{dx_{F}}, \label{eq:Zch1}$$ where the particle production spectrum is parametrized by: $$\begin{aligned} \frac{df_{i}}{dx_{F}} &=& \:\:\:\ \frac{1.08}{x_{F}} \:(1-x_{F})^{5} \:\:\:\mbox{{\rm for $D$ production}}, \nonumber \\ &=& \:\:\:\ 1.4 \:(1-x_{F})^{0.4} \:\:\: \mbox{{\rm for $\Lambda_{c}$ production}}. \end{aligned}$$ For the total inelastic cross section, $\sigma_{in}^{\! N-air}$, we used a parametrization with a $\log^{2}$ energy dependence[@chs:95]. Following Ref.[@volkova:87], a mild $\log$ energy dependence is assumed for the inclusive cross section of charm production, $\sigma_{NA}^{\! i }$, and the production of $D_{s}$ is neglected. Figure \[fig:qgsm\] displays the curves of Z-moments for the different charm components. RQPM: : Recombination Quark Parton Model, a phenomenological non-perturbative approach, taking into account the contribution of the intrinsic charm to the production process, in which a $c\bar{c}$ pair is coupled to more than one constituent of the projectile hadron, as described by Bugaev [*et al.*]{}[@bugaev:98; @bugaev:89]. Supposing Feynman-scaling to hold ([**RQPM-FS**]{}), the charm production Z-moments are given simply by: $$Z_{Ni}(\gamma) \:=\: Z_{\gamma} \:=\: \mbox{\rm constant}, \label{eq:Zch2}$$ with parameters for different particles shown in Table \[table:Zrqpm\]. Assuming the scaling violation ([**RQPM-SV**]{}), the parametrization turns out to be: $$Z_{Ni}(\gamma) \:=\: Z_{\gamma} \: \left( \frac{E_{N}}{E_{\gamma}} \right)^{\xi}, \label{eq:Zch3}$$ where $\xi= \: 0.177 -0.05 \,\gamma$. The parameters are also given in Table \[table:Zrqpm\]. We obtained the all-charged $D$-meson Z-moment ($i=D^{\pm}$) by averaging the individual conjugate moments, the same being done for neutral $D$-mesons, ($i=D^{o},\bar{D}^{o}$). Ref.[@bugaev:98] does not take into account the contribution of $D_{s}$ mesons. A comparison of the resulting Z-moments, with and without scaling, is provided in Figure \[fig:rqpm\]. [lcccrrr]{} Label & $\gamma$ & $\xi$ & $E_{\gamma}$ (GeV) & $Z_{\gamma}(D^{\pm})$ & $Z_{\gamma}(D^{o},\bar{D}^{o})$ & $Z_{\gamma}(\Lambda_{c}^{+})$\ RQPM-FS &1.62 & - & - &4.88 $\times 10^{-4}$ &4.73 $\times 10^{-4}$ &4.36 $\times 10^{-4}$\ &2.02 & - & - &3.14 $\times 10^{-4}$ &3.09 $\times 10^{-4}$ &2.95 $\times 10^{-4}$\ RQPM-SV &1.62 &0.096 &$10^{3}$ &5.55 $\times 10^{-4}$ &5.35 $\times 10^{-4}$ &4.9 $\times 10^{-4}$\ &2.02 &0.076 &$10^{6}$ &6.65 $\times 10^{-4}$ &6.55 $\times 10^{-4}$ &6.2 $\times 10^{-4}$\ pQCD: : Perturbative QCD, as calculated by TIG[@tig:96], using the Monte Carlo program PYTHIA[@pythia:94], explicitly evaluating the charm production, up to leading order (LO) in the coupling constant and including the next-to-leading order (NLO) distribution effects as an overall factor. Figure \[fig:pqcd\] presents the curves for pQCD calculations of the charm production Z-moments carried out supposing a primary spectrum either with, or without, the knee. The values of Z-moments for $D$-mesons were extracted directly from Ref.[@tig:96]. The Z-moments for $\Lambda_{c}^{+}$ and $D_{s}$ are derived by taking $Z (\Lambda_{c} ) \simeq \: 0.3 \: Z (D)$ and $Z (D_{s}) \simeq \: 0.2 \: Z (D)$, respectively, based on values assumed for the corresponding cross section ratios. More recently Gelmini, Gondolo and Varieschi (GGV)[@ggv:00] updated the calculation to include the full contribution of NLO predictions to the lepton fluxes. While TIG scales the LO cross sections by a constant factor of $K=2$ to obtain the NLO contribution, GGV evaluates explicitly the NLO component, as tailored by Mangano, Nason and Ridolfi[@mnr:93]. At the end, the net calculation corresponds as to multiply the LO term by an energy dependent factor $K$. In the $10^{2}$ to $10^{11}$ GeV energy interval, it starts at the lowest energies with $K=3$, decreases to around 2 for most of the intermediate energies, increasing slightly at the high energy extreme. However, the main difference between the two calculations comes from the extrapolation of the gluon parton distribution function, which produces higher charm cross sections even at LO. We do not duplicate here the particular effects implied over the Z-moments, leaving to consider the overall changes, resulting from GGV approach, at the evaluation of the lepton fluxes. Despite these differences, the atmospheric cascading routines are the same in both TIG and GGV calculations. A discussion of an alternative evaluation of the charm production based on pQCD, compared to TIG’s calculation, is presented by Pasquali [*et al.*]{}[@prs:99]. In Section \[section:results\] we compare the resulting prompt lepton fluxes, evaluated using each of these models. It is interesting to include in this comparison the results obtained by E. Zas [*et al.*]{}[@halzen:93]. They calculate extreme cases of charm production, at both low and high production rate limits. As for the high end, they assume a charm production cross section which is 10% of the total inelastic cross section (called Model-A), behaving as $\log^2 (s)$ at high energies, $\sqrt{s}$ being the center-of-mass system (c.m.s.) energy. At the low end lies a pQCD model at NLO, with structure functions given by Kwiecinski-Martin-Roberts-Stirling[@kmrs:93], adopting one choice of parameters that leads to relatively hard parton distribution functions (called Model-E). Results {#section:results} ======= High-energy atmospheric muon flux --------------------------------- A reproduction of calculated fluxes of high-energy muons reported by different authors is presented in Figure \[fig:reported\]. Comparing the conventional flux, from the decay of pions and kaons, obtained by Lipari and TIG, with the prompt flux, calculated by Volkova [*et al.*]{}, Bugaev [*et al.*]{}, TIG, GGV and Zas [*et al.*]{}, we note that the cross-over of the conventional and prompt components may occur at energies between $2 \times 10^3$ GeV and $2 \times 10^6$ GeV. The prompt flux intensity, itself, spans up to four orders of magnitude! We remark that the lowest curves, labeled TIG, GGV and Zas (E), are calculated assuming the pQCD charm production model. Although based on the same framework, these calculations of the charm cross section are subjected to theoretical uncertainties, which arises from the possible range of charm quark masses, as well as the factorization and renormalization scale dependence. Moreover, different assumptions are made for the parton distribution functions, needed at very small parton momentum fractions, not measured at accelerators. The curve labeled Bugaev is calculated within the RQPM. As an intrinsic-charm model, it predicts relatively hard inclusive spectra. In spite of that, the total inclusive cross section can be made rather large, since it depends strongly on the assumptions about the charm structure function of the projectile hadron. The curve labeled Volkova is obtained from a parametrization given in Ref.[@bugaev:98], because the original work[@volkova:87] quotes only flux ratios. The QGSM assumed in this calculation is considered to represent quite well open-charm production, and it is known to describe a wide variety of data on hadronic interactions. However, the model predicts a preferential production of certain secondary particle species which is not supported by experiment. In addition, predictions based on this model seem to exceed the experimental observations of horizontal air showers, measured by AKENO[@halzen:94]. Finally, the calculation labeled Zas (A) represents an extreme and crude upper bound for the prompt fluxes, based on the assumption that a charm is produced $10\%$ of the time in these high-energy interactions. Actually, it is ruled out by the bounds set to the charm cross section from the above mentioned analysis of AKENO data. One must be careful when analyzing differences in the curves of Figure \[fig:reported\]. We call attention to the fact that not only different charm models are being compared. The reported prompt fluxes are calculated adopting different atmospheric cascading routines. In order to better evaluate the observed discrepancies, we proceed the analysis by taking into account the separate effect of each single ingredient in the shower process. Later on, we analyze the effect of multiple ingredients, to evaluate how the choice of a different combination may affect the final result. Single ingredient effects ------------------------- If we calculate the fluxes of prompt neutrinos ($\nu_{\mu}$ and $\nu_{e}$), we obtain essentially the same values as the prompt muon flux (see for example Refs. [@bugaev:98; @tig:96; @ggv:00]). The reason is that both the parent ($D$ or $\Lambda_{c}$) and the daughter ($K$ or $\Lambda_{o}$) particles are massive compared to the leptons and the decay kinematics become blind to lepton family number or flavor. The prompt lepton flux is also essentially independent of $x$ (for depths greater than a few interaction lengths), due to the fact that the main contribution to the high-energy lepton flux must come from the first interactions of primary nucleons with air nuclei, while they are still energetic enough. In addition, for a fixed detection level, the flux is insensitive to the zenith angle $\theta$, up to the charm critical energy. This is also a consequence of the fact that the bulk of energetic leptons are produced high in the atmosphere and will reach the detector, regardless the amount of atmosphere traversed (less depth in the vertical direction, larger slant depths for showers close to the horizon). For energies above the critical energy, the charm particle decay length becomes comparable to the interaction length, and we feel the effects of its angular dependence, given in Eq. (\[eq:decay\]). Since the prompt lepton flux is almost independent of lepton flavor, detection depth and zenith angle, hereafter we will perform our comparisons using the muon-neutrino vertical flux ($\theta = 0^{o}$) at sea level ($x\approx 1000$ g/cm$^{2}$). First we investigate the effect of the primary spectrum at the top of the atmosphere (Section \[section:primes\]). To do so we calculate the prompt lepton flux, keeping all ingredients fixed, but the primary flux. Just for comparison purposes, we consider ingredients for the showering process used by Bugaev. The effect of the primary spectrum, alone, to the final result is considerable, as shown in Figure \[fig:result1\]. The spread on the resulting fluxes generally increases with energy. For example, at $10^{9}$GeV, the difference between using Lipari’s single slope and TIG’s primary is a factor 10 times, while they started together at $10^{3}$ GeV. Also a big shift is present for the curve with AKENO primary against all others with knee, above $10^{5}$ GeV. The next two ingredients considered are the nucleonic interaction length and the nucleonic Z-moment (Section \[section:sigmas\]). Apart from assuming a constant interaction length (as done by Lipari), for which the overestimated value at high energies pulls the neutrino flux down, the resulting fluxes are rather insensitive to the choice of $\lambda_{N}$. Similar situation occurs with the fluxes calculated changing only the nucleonic Z-moments, the difference being that the overestimated flux comes from TIG ($Z_{NN}(\gamma) \approx 0.5$, at energies below $10^{5}$ GeV), pushing the neutrino flux up, while other models ($Z_{NN}(\gamma) \approx 0.2 - 0.3$), result basically in the same final prompt flux. The fluxes are also rather insensitive to the charm interaction length up to $10^{7}$ GeV, as they should, since that is about the value of $\varepsilon_{critic}$ for charmed particles (see Table \[table:pdg\]). Above this range we discriminate the models up to a factor of two times, whether we use $\lambda_i$ constant or with a $\log(E)$ dependence. When evaluating the effects of the choice of charm production model (Section \[section:zch\]), the big uncertainties in the inclusive cross sections of charm production are transmitted to the calculated prompt lepton fluxes, as seen in Figure \[fig:result2\]. The spread between the prompt flux calculated with RQPM-SV and pQCD reaches a multiplicative factor of 20 at higher energies, solely due to the choice of $Z_{Ni}(\gamma)$. Extreme ingredient combinations ------------------------------- In the previous section the effect of each individual ingredient of the atmospheric showering process has been appreciated. Now, we investigate how the choice of multiple ingredients combined in different ways affects the resulting flux. Let’s consider an example. The choice of the production model QGSM makes the flux dominates over other choices of charm production only up to about $2 \times 10^5$ GeV, if all other parameters are fixed (Figure \[fig:result2\]). The combination of the QGSM with the AKENO primary flux, which solely contributes to larger prompt flux up to $10^{7}$ GeV (Figure \[fig:result1\]), makes this new set dominate upon a broader range of energy. Choosing a charm interaction length model that is responsible for increasing the prompt flux at higher energies, we build a set which dominates over yet a larger range. Such a combination of showering parameters leads to a maximum flux configuration. With a similar procedure we can build a minimum configuration. In this case it is interesting to consider the effect of Lipari’s primary up to the energy of the knee, above which Bugaev’s Model-D for the primary spectrum produces the minimum prompt flux. The other showering parameters may also be chosen as to stress the maximum or minimum possible behavior of a given flux calculation. Table \[table:combinations\] presents our suggestion of multiple ingredient combinations selected to obtain extreme outputs. The prompt lepton flux curve calculated for each charm production model (QGSM, RQPM or pQCD), can be shifted up or down, depending on the chosen combination of ingredients, an effect illustrated in Figures \[fig:result3\] to \[fig:result5\]. The band in each figure reflects the freedom to change the resulting flux between the maximum (MAX) and minimum (MIN) extreme combinations listed in Table \[table:combinations\]. Ingredient MAXIMUM Set MINIMUM Set ------------------------------------ ---------------------------- ----------------------------- $N_{0}, \gamma$ (for $E<E_{knee}$) AKENO Lipari $N_{0}, \gamma$ (for $E>E_{knee}$) AKENO Bugaev-D $\lambda_{N} (E)$ power-law[@akeno:84] constant[@lipari:93] $Z_{NN}(\gamma)$ scaling violation[@tig:96] constant with knee[@chs:95] $\lambda_{i} (E)$ constant[@volkova:85] $\log(E)$[@bugaev:98] : \[table:combinations\] Extreme ingredient combinations of atmospheric showering parameters, selected to obtain maximum or minimum behavior of prompt flux. Conclusion ========== The calculation of the prompt lepton flux produced in the atmosphere by the semileptonic decay of charmed particles is rather straightforward, but we cannot say the same for the analysis of the results. For instance, the lack of precise information on high-energy charm inclusive cross-section in hadron-nucleus collisions is accompanied by a variety of options for the particle showering process in the atmosphere. We described different ingredients of the calculation, comparing side by side several parametrizations for each one of them, and evaluated their relative importance to the final result. The major effects are due to the choice of the primary spectrum at the top of the atmosphere and, of course, to the choice of the charm particle production model. Only first nucleonic interactions play essential role in determining the prompt lepton flux down at sea level, therefore variations in the nucleonic attenuation lengths are not that relevant, while the charm interaction length have some influence, above the charm particle critical energy. We observed how different combinations of ingredients can shift the resulting flux curves up or down. Therefore, the comparison between different calculations must be carried out with great care. The bands displayed in Figures \[fig:result3\]-\[fig:result5\] correspond to the allowed regions for the prompt lepton flux calculated respectively with the charm production models QGSM, RQPM and pQCD, if the showering process parameters are mixed differently. The prompt lepton crossover energy, that is the energy above which the charm particle decay products dominate over the conventional pion and kaon decay induced fluxes, is yet an uncertain quantity. According to Figures \[fig:result3\]-\[fig:result5\], it may be anywhere between $10^{4}$ and $10^{6}$ GeV. If the cross-over is “low enough” (about $10^{4}$ GeV), then neutrino telescopes now operational can therefore take advantage of the isotropy of the prompt lepton flux, to search for an zenith angle independent component in their data. This can also be pursued by the analysis of the more copiously detected down-going muons. Exploiting the case of tau-neutrinos, which may produce a clear signature in high-energy neutrino detectors[@learned:95], will be addressed in a future analysis. There is also the need for a more comprehensive description of the available data on charm production cross section, and its extrapolation to higher energies. The prompt lepton flux is on the order-of-the-day of operating high-energy neutrino telescopes, because of the background it represents. Proposed experiments, like IceCube [@ice3:00], may turn the arguments the other way around, for their measurements with enhanced sensitivity may provide outstanding information on heavy quark interactions, just by discriminating atmospheric from cosmic neutrinos, at energies above tens of TeV. Acknowledgements {#acknowledgements .unnumbered} ================ The author would like to thank Freddy Binon, Jean-Marie Frère, Francis Halzen and C. Salles for many discussions and suggestions to the manuscript. This work was partially supported by the I.I.S.N. (Belgium) and by the Communauté Française de Belgique - Direction de la Recherche Scientifique, programme ARC. [99]{} T.K. Gaisser, F. Halzen and T. Stanev, Phys. Rep. [**258,**]{} 173 (1995); F. Halzen, Phys. Rep. [**333-334,**]{} 349 (2000). E. Andres [*et al.*]{} (The AMANDA Collaboration), Astropart. Phys. [**13,**]{} 1 (2000). V.A. Balkanov [*et al.*]{} (The BAIKAL Collaboration), Astropart. Phys. [**14,**]{} 61 (2000). C. Spiering, “Future High Energy Telescopes”, to appear in the [*Proceedings of the 19th International Conference On Neutrino Physics And Astrophysics - Neutrino 2000*]{}, June 16-21, 2000, Sudbury (Canada). T.K. Gaisser, [*Cosmic Rays and Particle Physics*]{}, (Cambridge University, Cambridge, 1992). D.E. Groom [*et al.*]{} (Particle Data Group), Eur. Phys. Jour. [**C15,**]{} 1 (2000). E.V. Bugaev, A. Misaki, V.A. Naumov, T.S. Sinegovskaya, S.I. Sinegovsky and N. Takahashi, Phys. Rev. D [**58,**]{} 054001 (1998). L. Pasquali and M.H. Reno, Phys. Rev. D [**59,**]{} 093003 (1999). L.V. Volkova, W. Fulgione, P. Galeotti and O. Saavedra, Nuovo Cimento C [**10,**]{} 465 (1987). E. Zas, F. Halzen and R.A. Vázquez, Astropart. Phys. [**1,**]{} 297 (1993). R. Hagedorn, in [*Relativistic Kinematics,*]{} (Benjamin, NY, 1963). P. Lipari, Astropart. Phys.[**1,**]{} 195 (1993). M. Thunman, G. Ingelman and P. Gondolo, Astropart. Phys. [**5,**]{} 309 (1996). L. Pasquali, M.H. Reno and I. Sarcevic, Phys. Rev. D [**59,**]{} 034020 (1999). T.H. Burnett [*et al.*]{} (The JACEE Collaboration), in [*Proceedings of the 21st International Cosmic Ray Conference*]{}, Adelaide, Australia, Vol.3, p. 101 (1990). T. Hara [*et al.*]{}, in [*Proceedings of the 18th International Cosmic Ray Conference*]{}, Bangalore, India, Vol.9, p. 198 (1983); M. Nagano [*et al.*]{}, J.Phys. G [**10,**]{} 1295 (1984). E.V. Bugaev, V.A. Naumov, S.I. Sinegovsky and E.S. Zaslavskaya, Nuovo Cimento C [**12,**]{} 41 (1989). C.G.S. Costa, F. Halzen and C. Salles, Phys. Rev. D [**52,**]{} 3890 (1995); Phys. Rev. D [**54,**]{} 5558 (1996). M.M. Block, F. Halzen and T. Stanev, Phys. Rev. Lett. [**83,**]{} 4926 (1999); Phys. Rev. D [**62,**]{} 077501 (2000). L.V. Volkova, Nuovo Cimento C [**8,**]{} 552 (1985); Y. Minorikawa and K. Mitsui, Lett. Nuovo Cimento [**44,**]{} 651 (1985). K. Mitsui, Y. Minorikawa and H. Komori, Nuovo Cimento C [**9,**]{} 995 (1986). See, e.g. A.B. Kaidalov and O.I. Piskunova, Z. Phys. C [**30,**]{} 145 (1986) and references therein. T. Sjostrand, Comput. Phys. Commun. [**82,**]{} 74 (1994). G. Gelmini, P. Gondolo and G. Varieschi, Phys. Rev. D [**61,**]{} 036005 (2000); Phys. Rev. D [**61,**]{} 056011 (2000). M.L. Mangano, P. Nason and G. Ridolfi, Nucl. Phys. [**B373,**]{} 295 (1992); Nucl. Phys. [**B405,**]{} 507 (1993). See, e.g. A.D. Martin, W.J. Stirling and R.G. Roberts, Phys. Rev. D [**47,**]{} 867 (1993). M.C. Gonzales-Garcia, F. Halzen, R.A. Vázquez and E. Zas, Phys. Rev. D [**49,**]{} 2310 (1994). J.G. Learned and S. Pakvasa, Astropart. Phys.[**3,**]{} 267 (1995). See, e.g. “The IceCube National Science Foundation Proposal,” available at the URL http://pheno.physics.wisc.edu/IceCube/ and also: F. Halzen [*et al.*]{}, “From the first neutrino telescope, the Antartic Muon and Neutrino Detector Array AMANDA, to the IceCube Observatory”, in [*Proceedings of the 26th International Cosmic Ray Conference,*]{} South Lake City (USA), edited by D. Kieda, M. Salamon and B. Dingus, Vol. 2, 428-431 (1999).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We construct a continuous family of exchangeable pairs by perturbing the random variable through diffusion processes on manifolds in order to apply Stein method to certain geometric settings. We compare our perturbation by diffusion method with other approaches of building exchangeable pairs and show that our perturbation scheme cooperates with the infinitesimal version of Stein’s method harmoniously. More precisely, our exchangeable pairs satisfy a key condition in the infinitesimal Stein’s method in general. Based on the exchangeable pairs, we are able to extend the approximate normality of eigenfunctions of Laplacian on a compact manifold to eigenfunctions of Witten Laplacian, which is of the form:$\Delta_w = \Delta - \nabla H$. We then apply our abstract theorem to recover a central limit result of linear statistics on sphere. Finally, we prove an an infinitesimal version of Stein’s method for exponential distribution and combine it with our continuous family of exchangeable pairs to extend an approximate exponentiality result of $|\operatorname{Tr}U|^2$, where $\operatorname{Tr}U$ is the trace of the first power of a matrix $U$ sampled from the Haar measure of unitary group, to arbitrary power and its analog for general circular ensemble.' author: - 'Weitao Du[^1]' title: Constructing exchangeable pairs by diffusion on manifolds and its application --- Introduction {#intro} ============ In [@cs72], Charles Stein introduced the celebrated Stein method along with the notion of exchangeable pairs for proving the rate of convergence for normal distribution. The main idea behind Stein’s method is to replace the characteristic function typically used to show distributional convergence with a characterizing operator $\mathcal{A}$. This characterizing operator of the target distribution defines a differential equation, the Stein equation, which can be solved and has a probabilistic representation. Then the problem of bounding the distance between two distributions is reduced to estimate the derivatives of Stein equation’s solution. The metric we use to measure the distance is usually of the form: $$d(\mu , \nu ) = \sup_{h\in \mathcal{H}} | \mathbb{E} h(X) - \mathbb{E}h(W)|\ ,$$ where $X$, $W$ are random variables with distribution $\mu$ and $\nu$ and $\mathcal{H}$ is a certain function class. We name a few distributions that have been tackled by Stein method: the Poisson distribution [@chen75; @cdm05; @hg], the exponential distribution [@cdm05; @cfr], the Beta distribution [@beta] and manifold-valued measures [@cdm05]. For a short survey, the reader can refer to [@cha]. In this article, we focus on building exchangeable pairs in order to apply Stein method efficiently. Recall that a pair $(X,X')$ of random elements on a common probability space is exchangeable if $$(X,X') \overset{d}{=} (X' , X)\ .$$ One standard way to construct an exchangeable pair is to go one step in a discrete time reversible Markov chain. This works well when the underlying problem is essentially discrete such as a random sequence with weak correlation. On the other hand, when there exists a continuous symmetry in the model, we can create a continuous family of exchangeable pairs by perturbing the random variable we want to study continuously In [@meckes-thesis], the author developed an infinitesimal version of Stein’s method which is coupled with a continuous family of exchangeable pairs. It was applies to several geometric settings like the sphere harmonics [@meckes-L] and the orthogonal group [@M07],[@wwr]. The procedure to produce the exchangeable pairs in [@M07; @meckes-L; @El2] was to perturb the original random variable $X$ along a deterministic flow such as geodesic flows. In [@Fj; @Fj10; @jn], the authors constructed the continuous family of exchangeable pairs by the heat kernel. Once we have the general set-up, the next step boils down to control the moments of certain functionals of the target random variable. This task is usually accomplished by combinatorial techniques.For instance, the so-called Weingarten calculus has been developed to calculate moments of a broad class of functions defined on the lie group. This kind of calculation also has a close relation with integrable probability, see [@al; @bs]. The main purpose of the present paper is to show how to create a continuous family of exchangeable pairs by diffusion process on manifold. When we set underlying process as Brownian motion, this method can be seen as a microscopic explanation of the heat kernel method in [@Fj10]. For example, the heat kernel perturbation was used in [@ja] to study linear eigenfunctions of Laplacian operator on unitary group. Besides covering the pure Laplacian case, our method also works for the Witten Laplacian by adding a drift generated by a gradient vector field $\nabla H$ to the Brownian motion, Where $H(x)$ is a weight function on the manifold. On the other hand, the diffusion perturbation scheme matches well with the infinitesimal Stein method. We will demonstrate this point by showing that a key condition we need to check to guarantee the approximate normality holds in general under our framework. We list the third condition of the infinitesimal version of Stein’s method (see [@El2]) as follows: $$\lim_{t \rightarrow 0} \frac{1}{t} {\mathsf E}[ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0 \ .$$ We prove this by utilizing a concentration property of Brownian motion as $t \rightarrow 0$ and Girsanov transform. Note that this condition is usually verified by calculating the fourth moment, which is a heavy computational task even if there is an explicit formula available. Moreover, this condition is universal in the sense that it also emerges in infinitesimal Stein method for other probabilistic distributions. To show how this universal condition appears, we convert the Stein method for exponential distribution in [@jn] into an infinitesimal version. Then we are able to study the distribution of $|\operatorname{Tr}(U^k)|^2$ for all $k > 0$, where $U$ is sampled by the Haar measure of unitary group. This is closely related with the central limit phenomenon of unitary matrices (see [@kg]). Properties of the norm of $\operatorname{Tr}(U^k)$ plays a role in comparing the circular unitary ensemble and the Riemann zeta function (see [@fm; @cd]), where a precise bound on the rate of convergence of $|\operatorname{Tr}(U^k)|$ is crucial. Furthermore, the analog of the trace power functions can be defined for general circular ensemble which is a key ingredient for studying linear statistics of the corresponding ensemble. Although the geometry of the underlying space has transformed from the compact group to an open simplex, we can still adapt our method to this case. Note that for general circular ensemble, the corresponding diffusion has a drift, so we need to harness the full power of the abstract theorem developed in section 3. Last but not least, our method may have the potential to generalize to metric measure space where the notion of diffusion can still be defined (see [@vjr; @ks]). The rest of the contents are organized as follows. We first introduce the notation in section 1.1. In section 2 and 3, the method of constructing exchangeable pairs by diffusion on manifold is given and then we prove a general theorem on the approximate normality of eigenfunctions of the Witten Laplacian, which is also known as the Bakery-Emery Laplacian (see [@xiang12; @dim]). The calculation in the process of proving the theorem will be encountered again in the following sections. Section 3 recovers a result in [@meckes-thesis] by implementing the Brownian motion perturbation. The next section treats the norm of the trace of the powers of unitary matrix. We first prove an infinitesimal approximation theorem on the basis of the estimates of the Stein equation built in [@jn]. Then we apply the theorem to prove the approximate exponentiality of $|\operatorname{Tr}(U^k)|^2$ for general $k > 0$, where $U$ is sampled by the Haar measure of unitary group. In the last section, we modify the general theorem in section 2 in order to handle Dyson Brownian motion, then we apply the theory to study the analog of $|\operatorname{Tr}(U^k)|^2$ for general circular ensemble. Notation -------- Let $(M,g)$ be an n-dimensional Riemannian manifold. We assume $M$ is compact if not particularly indicated. In local coordinates $(x^1, \dots, x^n)$ with tangent vectors $\{\frac{\partial}{\partial x_i} \}_{i=1}^{n}$, define $$(G(x))_{ij} = g_{ij} (x) = g( \frac{\partial}{\partial x_i} , \frac{\partial}{\partial x_j}),\ \ \ \ g^{ij}(x) = (G^{-1} (x))_{ij}\ .$$ Denote the covariant derivative on $M$ as $\nabla$. Then for a smooth function $H(x)$ (the potential function) on $M$, the gradient vector field is defined as $g^{-1} \nabla H$. We will abuse the notation and denote the gradient vector field as $\nabla H$ as well. In local coordinates, the Christoffel symbols $\Gamma^{m}{}_{ij}$ are given by $$\Gamma^{m}{}_{ij}={\frac{1}{2}}\,g^{mk}\left({\frac {\partial }{\partial x^{j}}}g_{ki}+{\frac {\partial }{\partial x^{i}}}g_{kj}-{\frac {\partial }{\partial x^{k}}}g_{ij}\right) \nonumber$$ and the components of the Ricci curvature tensor are given by $$Ric_{ij}={\frac {\partial \Gamma ^{\ell }{}_{ij}}{\partial x^{\ell }}}-{\frac {\partial \Gamma ^{\ell }{}_{i\ell }}{\partial x^{j}}}+\Gamma ^{m}{}_{ij}\Gamma ^{\ell }{}_{\ell m}-\Gamma ^{m}{}_{i\ell }\Gamma ^{\ell }{}_{jm}. \nonumber$$ For a general tensor $T$ on $M$ or a certain domain specified by the context, we denote the norm of $T$ induced by the metric $g$ as $\norm{T}$ and the maximum of the norm on the domain as $$\norm{T}_{max}\ .$$\ Let $ \Delta$ be the Laplacian operator with respect to g, in local coordinates, we have $$\Delta f = \operatorname{Tr}\nabla \nabla f = \frac{1}{\sqrt{det(G)}} \frac{\partial}{\partial x_j} (\sqrt{det(G)} \cdot g^{jk} \frac{\partial f}{\partial x_k}) \ .$$ The Witten Laplacian with respect to $H(x)$ is defined by $$\Delta_w : = \Delta - \nabla H\ .$$ We denote the orthonormal frame bundle of $M$ as $\mathcal O (M)$, which is the set of $(x, \phi^1,\dotsc,\phi^n)$, where $x \in M$ and $\phi^1,\dotsc,\phi^n$ forms an orthonormal basis for $T_x M$. Then the covariant derivative $ \nabla$ induces a connection 1-form $\omega$ on $\mathcal O (M)$. It’s well known that the tangent bundle of $\mathcal O (M)$ splits by the connection: $$T \mathcal O (M) = H T\mathcal O (M) + V T \mathcal O (M) \ ,$$ where we denote the vertical subbundle as $V T \mathcal O (M)$ and the horizontal subbundle, which is the kernel of $\omega$ as $H T\mathcal O (M)$. On the space of real $n\times n$ matrices, the Hilbert-Schmidt inner product and the corresponding norms are defined by $${\left\langle A, B \right\rangle}_{H.S.}=\operatorname{Tr}(AB^T),\ \ \ \ \norm{A}_{H.S}=\sqrt{\operatorname{Tr}(AA^T)}\ .$$ The operator norm of a matrix $A$ over ${\mathbb R}$ is defined by $$\norm{A}_{op}=\sup_{|v|=1,|w|=1}|{\left\langle Av, w \right\rangle}|.$$ Let X,Y be random variables, the Wasserstein distance $d_W (X,Y)$ is defined by $$d_W (X,Y) = \sup_{Lip(g) \leq 1} | \mathbb{E} g(X) - \mathbb{E} g(Y)|\ ,$$ where $Lip(g) = \sup_{x \neq y} \frac{|\mathbb{E} g(X) - \mathbb{E} g(Y)|}{|x-y|}$ is the Lipschitz constant of $g$. The total variation distance between X and Y is defined by $$d_{TV} (X,Y) = \frac{1}{2} \sup_{g} |\mathbb{E} g(X) - \mathbb{E} g(Y)|\ ,$$ where $g$ belongs to the class which consists of continuous functions which are bounded by 1. The Kolmogorov distance $d_K (X,Y) $ is defined by $$d_K (X,Y) = \sup_{s \in \mathbb{R}}| \mathbb{P}( X \leq s)- \mathbb{P}( Y \leq s)|\ .$$ Building the exchangeable pair ============================== Suppose $X$ is a random point in $M$. To create an exchangeable pair, we run a stochastic process starting at $X$. The specific stochastic process we will consider is defined by a stochastic differential equation on the orthonormal frame bundle $\mathcal O (M)$. We will briefly review the construction of stochastic differential equation on manifold, for the detail, the reader can see [@Hsu; @fyw]. Let $\pi_*$ be the be the isomorphism $\pi_* : H T\mathcal O (M) \rightarrow M$ induced by the canonical projection $\pi: \mathcal O (M) \rightarrow M$. Suppose we are given an element $\phi = (x, \phi^1,\dotsc,\phi^n)$ of $\mathcal O (M)$ and a vector $v \in T_x(M)$, then the horizontal lift of $v$ at $\phi$ is the unique tangent vector $\widetilde{v} \in H T\mathcal O (M)$ s.t $$\pi_* \widetilde{v} = v\ .$$ Let $ \mathcal H^i ( \phi) = \widetilde{\phi^i} $ be the horizontal lift of the orthonormal basis of $T_x M$, then we can define the horizontal Laplacian as $$\Delta_{\mathcal H} = \sum_{i=1}^n (\mathcal H^i)^2 \ .$$ We have the following relation with the Laplacian on $M$: $$\Delta_{ \mathcal H } ( f \circ \pi ) = \Delta (f) \circ \pi \ .$$ The stochastic differential equation is of the form: $$\left\{ \begin{array}{l} d U_t = \sum_{i=1}^d \mathcal H^i ( U_t ) \circ d B^i_t - \frac{1}{2} \widetilde{\nabla H} (U_t) dt , \\ \pi (U_0) = X. \end{array} \right.$$ The vector field $\widetilde{\nabla H}$ on orthonormal frame bundle is the horizontal lift of $\nabla H$ defined on $M$. We denote the projection of $U_t$ to $M$ as $X_t$, i.e $X_t = \pi (U_t)$. Since $M$ is compact, everything is smooth and bounded which is enough to insure the existence of a unique strong solution to (2.1). The stationary distribution of this process is $d\mu = \frac{1}{Z} e^{-H(x)} dx$, where Z is the normalization factor to make $d \mu $ a probabilistic measure. Since $X_t$ is reversible with stationary distribution $\mu$, we know that $(X_t , X)$ has the same law as $(X , X_t )$ If we sample $X$ from the stationary distribution $\mu$. In this way, we have created a family of exchangeable pairs $(X_t , X)$ indexed by the time t. There is a correspondence between the solution of the heat equation of Witten Laplacian and the underlying process $X_t$. Consider the following initial value problem for $\theta = \theta ( t,x)$: $$\left\{ \begin{array}{l} \frac{\partial \theta}{\partial t} = \frac{1}{2} \Delta_w \theta , \\ \theta (0,x) = f(x) , \ \ \ x \in M. \end{array} \right.$$ Then we have a Feynman-Kac representation $$\theta (t,x) = \mathbb{E} [ f ( X_t) | X_0 = x ] \ ,$$ for the detail, see Theorem 7.2.1 of [@Hsu]. On the other hand, from functional analysis, we know that the solution has a semigroup representation: $$\theta (t, x) = e^{t \Delta_w } f(x)\ ,$$ where $e^{t \Delta_w}$ is defined as $e^{t \Delta_w} : = I + t \Delta_w + t^2 \frac{\Delta_w^2}{2!} + \cdots$. We will need the asymptotics of $\theta (t, x) = \mathbb{E} [ f ( X_t) | X_0 = x ]$ with respect to the time $t$. In [@ja] and [@Fj10], this was down by means of the semigroup representation of heat equation. We give a probabilistic argument that also applies when the drift of the underlying process is singular. By Ito’s formula, $$\mathbb{E} [ f ( X_t) | X_0 = x ] = f(X_0) + \mathbb{E} [ \int_0^t \Delta_w f(X(s)) ds | X_0 = x ]\ . \eqno{(2.2)}$$ Since we assume $f$ is a smooth function on a compact manifold $M$, we can exchange the order of differentiation and expectation freely, $$\lim_{t \rightarrow 0} \mathbb{E} [ f ( X_t) - f(X_0) | X_0 = x ] = \Delta_w f( X_0)\ . \eqno{(2.3)}$$ 1. When we consider noncompact manifolds, we usual require a lower bound on the Bakry-Émery Ricci tensor associated with the Witten Laplacian: $$Ric_H : = Ric + \nabla^2 H\ ,$$ where $ \nabla^2 H$ is the Hessian two tensor of the potential function $H$. Otherwise, we may encounter analytical issues ( see[@fyw] ). 2. In [@meckes-L], the way to create the exchangeable pair is by perturbing $X$ by a geodesic flow. In that setting, we have $$\nonumber \begin{split}\label{taylor} f(X_\epsilon)-f(X)&=\epsilon\cdot\left.\frac{d(f\circ\gamma)}{dt}\right|_{t=0}+ \frac{\epsilon^2}{2}\cdot\left.\frac{d^2(f\circ\gamma)}{dt^2}\right|_{t=0} +O(\epsilon^3)\ . \end{split}$$ The first order term in $\epsilon$ is cancelled after taking expectation due to the symmetry. So the actual perturbing order in [@meckes-L] is $\epsilon^2$. However after adding the drift to the Laplacian, the symmetry is broken. So the original perturbation doesn’t work in our case. By Ito’s formula, $$df(X_t) = \sum_{i=1}^d ( \mathcal H^i f ) d B^i_t + [\frac{1}{2} \Delta f - \frac{1}{2} (\nabla H f ) ] dt\ , \eqno{(2.4)}$$ the Laplacian $\Delta$ and the drift $\nabla H$ scales in same order. As we will see in the next section, this is important if we want to apply the infinitesimal Stein method. Approximate normality of eigenfunctions of Witten Laplacian =========================================================== We need the following infinitesimal multivariate normal approximation theorem introduced in [@El2]: \[abscont\]Let $W$,$W_t$ be $ \mathbb{R}^d$ -valued random vector and $(W,W_t)$ forms a family of exchangeable pairs defined on a common probability space. Suppose there is an invertible matrix $\Lambda$, a positive definite symmetric matrix $\Sigma$, a random vector $E$, a random matrix $E'$ such that 1. $${\mathsf E}\left[W_t -W\big| W\right]=-t \Lambda W+t E + O(t^2),$$\[lindiff3\] 2. $${\mathsf E}\left[(W_t-W) (W_t - w)^T\big| W\right]=2 t \Lambda \Sigma + t E' + O(t^2),$$\[quaddiff3\] 3. For each $ \rho > 0$,$$\lim_{t \rightarrow 0} \frac{1}{t} {\mathsf E}[ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0$$\[cubediff3\] Then $$d_{W}(W,\Sigma^{\frac{1}{2}}Z)\le \norm{\Lambda^{-1}}_{op}\left[\frac{1}{2} \norm{\Sigma^{-\frac{1}{2}}}_{op} \mathbb{E}\norm{E'}_{H.S}+ \mathbb{E}\left|E\right|\right],$$ where $Z$ is a standard Gaussian random vector in $\mathbb{R}^d$. This approximation theorem is the version in [@El2], which is different from [@meckes-L] in the third condition. In [@meckes-L], the third condition is $$\frac{1}{t} \mathbb{E} |W_t-W|^3 = o(1)\ .$$ It’s hard to bound the third moment directly by our perturbation method. On the other hand, due to the concentration of the Gaussian measure, $\mathbb{P} ( | W_t - W |^2 > \rho ) $ goes to zero exponentially in t, the third condition in [@El2] holds as a trivial corollary. The function $f: M \rightarrow \mathbb{R}$ is an eigenfunction of $\Delta_w$ with eigenvalue $-\nu$ if $$\Delta_w f = \Delta f(x) - \frac{1}{2} \nabla H f(x) = - \nu f(x)\ .$$ We can always normalize $f(x)$ such that $\int_M f(x) d \mu (x) = 0$ and $\int_M f^2(x) d \mu (x) = 1$. Let $X$ be a random point of $M$ sampled from measure $d\mu$. Define the value distribution of $f$ with respect to $\mu$ as the distribution of the random variable $f(X)$. Consider a sequence of $L^2$-orthonormal (with respect to $d\mu (x)$) eigenfunctions of $\Delta_w $ with corresponding eigenvalues $- \nu_i$. Let $W$ be the random vector $(f_i(X))$. We construct $W_t$ as $(f_i(X_t))$, where $X_t$ is the stochastic process starting at $X$ defined previously. Then $( W , W_t )$ forms a family of exchangeable pairs parameterized by time t. Let $W_t$, $W$ be defined as above. Then $$\lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ W_t - W | W ] = - \nu_i f_i(X)\ . \eqno{(3.1)}$$   Applying (2.3), one has that $$\begin{aligned} \lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ W_t - W | W ] &= \frac{1}{t} \{ \mathbb{E} [ f_i( X_t ) | X ] - f_i(X) \} \\ & = \lim_{t \rightarrow 0} \frac{1}{t} [ f_i(X) + t \Delta_w f_i(X) + O(t^2) - f_i(X) ] \\ & = - \nu_i f_i(X)\end{aligned}$$ This implies that condition 1 of Theorem 3.1 is satisfied if we take $\Lambda = diag (\nu_1, \cdots , \nu_k ) $ and $E = 0$. Then the operator norm of $\norm{\Lambda^{-1}}$ satisfies $$\norm{\Lambda^{-1}}_{op} \leq \max_{1\le i\le k} (\mu_i^{-1})\ .$$ Let $\{ f_i(x) \}$ be any sequence of smooth functions on $M$ (not necessarily eigenfunctions). Let $W$,$W_t$ be the random vector $(f_i(X))$ and $(f_i(X_t))$ as above. Then $$\mathbb{E} [ (W_t - W)_i (W_t - W)_j | W ] = 2 t \langle \nabla f_i(X) , \nabla f_j(X) \rangle + O(t^2)\ , \eqno{(3.2)}$$ where $\langle \cdot , \cdot \rangle$ is the inner product induced by the metric g. $$\begin{aligned} \mathbb{E} [(W_t - W)_i (W_t - W)_j | W ] &= \mathbb{E} [ f_i(X_t)f_j(X_t) | W] - W_i \mathbb{E} [ f_j(X_t) | W] - W_j \mathbb{E} [ f_i(X_t) | W]+ f_i(X_t)f_j(X_t) \\ & =f_i(X)f_j(X) + t\Delta_w [f_i(X)f_j(X)] - 2f_i(X)f_j(X) - t f_i(X)\Delta_w f_j(X) \\ &\ \ \ - tf_j(X)\Delta_w f_i(X) + f_i(X_t)f_j(X_t) + O(t^2) \\ & = t\Delta_w [f_i(X)f_j(X)] - tf_i(X)\Delta_w f_j(X) - tf_j(X)\Delta_w f_i(X) + O(t^2) \ .\end{aligned}$$ Plugging in the identity $$\Delta_w [f_i(x)f_j(x)] = f_i(x)(\nabla H f_j(x) + \Delta f_j(x)) + f_j(x)(\nabla H f_i(x) + \Delta f_i(x)) + 2 \langle \nabla f_i(x) , \nabla f_j(x) \rangle\ ,$$ we get the final expression $$\mathbb{E} [(W_t - W)_i (W_t - W)_j | W ]= 2 t \langle \nabla f_i(X) , \nabla f_j(X) \rangle + O(t^2) \ .$$ So condition 2 of Theorem 3.1 is satisfied if we take $E' = 2 \langle \nabla f_i(X) , \nabla f_j(X) \rangle - 2\Lambda $ and $\Sigma$ to be the identity matrix. Although it’s hard to estimate $\langle \nabla f_i(X) , \nabla f_j(X) \rangle$ pointwise, we can calculate the expectation: $$\begin{aligned} \nonumber \mathbb{E} \langle \nabla f_i(X) , \nabla f_j(X) \rangle & = \sum_{k=1}^d \int_{M}\langle \nabla_k f_i(x) , \nabla_k f_j(x) \rangle d\mu \\ \nonumber & = \sum_{k=1}^d \int_{M} \langle \nabla_k f_i(x) , \nabla_k f_j(x) \rangle e^{-H(x)} dx\ . \nonumber\end{aligned}$$ Since $M$ has no boundary, after integrating by parts, we have $$\begin{aligned} \nonumber \mathbb{E} \langle \nabla f_i(x) , \nabla f_j(x) \rangle & = - \int_{M} \langle f_i(x) , \Delta f_j (x) \rangle d\mu + \int_{M} \langle f_i(x) , \nabla H f_j(x) \rangle d \mu \\ \nonumber & = - \int_{M} \langle f_i(x) , \Delta_w f_j(x) \rangle d \mu \\ \nonumber & = \delta_{ij} \nu_j\ , \nonumber\end{aligned}$$ thus, $$\mathbb{E} \norm{E'}_{H.S} = 2 \mathbb{E} \sqrt{\sum_{i,j=1}^k [ \langle \nabla f_i(X) , \nabla f_j(X) \rangle - \mathbb{E} \langle \nabla f_i(X) , \nabla f_j(X) \rangle ]}\ .$$ To verify condition 3, we need proposition 5.1.4 of [@Hsu] on the exit time of Brownian motion on manifold. Let $\tau_r$ denotes the first exiting time of the ball $B(x , r)$ of the Brownian motion starting at x. We have the following property: Let $i_M$ be the injectivity radius of M and $r < i_M$. Then $\exists$ a positive smooth function $C_r (x)$ s.t as $t \rightarrow 0$, $$\mathbb{P}_x \{ \tau_r \leq t \} \sim \frac{C_r (x)}{t^{(d-2)/2}} e^{-r^2 / 2t}\ ,$$ uniform in M. Once we have this key lemma on the exit time, we are ready to verify condition (3). For every $ \rho > 0$,$$\lim_{t \rightarrow 0} \frac{1}{t} {\mathsf E}[ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0\ .$$ For simplicity, we assume $k =1$. For general random vector $W = \{f_i(X)\}_{i=1}^k$, the proof is almost the same. We first prove the theorem when there is no drift and extend the argument by Girsanov transform. 1. $\nabla H = 0$, $\Delta_w = \Delta\ .$\ When $\nabla H = 0$, the underlying process is just Brownian motion. We can write the expectation in the following form: $$\lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = \lim_{t \rightarrow 0}\frac{1}{t} \mathbb{E} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) | W]\ .$$ Since M is compact, we can find a constant $C > 0$ such that the first derivative of f is bounded by C. Then $$| f(X_t) - f( X) | \leq C \cdot |X_t - X|\ .$$ Choose a radius $r$ such that $r \leq \frac{\rho}{C}$, then we have the following upper bound: $$\begin{aligned} \nonumber \lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] & \leq \lim_{t \rightarrow 0}\frac{1}{t} \mathbb{E} \mathbb{E} [ (f(X_t) - f(X))^2 \mathbb{I}( \tau_{\sqrt{\rho}} \leq t) | X] \\ \nonumber & \leq \lim_{t \rightarrow 0} 4 \cdot \norm{f}^2_{max} \frac{C_r (x)}{t^{d/2}} e^{-r^2 / 2t}\\ \nonumber & = 0\ , \nonumber\end{aligned}$$ where we use lemma 3.4 in the penultimate step. 2. $\nabla H = 0$, $\Delta_w = \Delta + \nabla H\ .$\ When $ H(x) \neq constant$, there is a nonzero drift term in the SDE: $$d U_t = \sum_{i=1}^d \mathcal H^i ( U_t ) \circ d B^i_t - \frac{1}{2} \widetilde{\nabla H} (U_t) dt \ .$$ By definition, the anti-development of $U_t$ can be written as $B_t + V_t$, where $V_t$ satisfies the following SDE: $$dU_t = \sum_{i=1}^d \mathcal H^i ( U_t ) \circ ( d B^i_t + \dot{V}(t) dt )\ .$$ Since the vector field $\widetilde{\nabla H(x)}$ is bounded, we know that the $|\dot{V(t)}|$ is also bounded and $V(t) = \int_0^t \dot{V_s} ds$ belongs to the Cameron-Martin space. Define $M_t$ as $$M_t = \exp( - \int_0^t \dot{V_s} dB_s - \frac{1}{2} \int_0^t |\dot{V_s}|^2 ds ) \ .$$ Since $\mathbb{E} [ e^{\frac{1}{2} \int_0^t |\dot{U_s}|^2 ds} ] < \infty$, by Novikov’s criterion, $M_t$ is a uniformly integrable martingale and Girsanov’s transform applies. We denote the original probabilistic measure as $P$ and define the new measure $Q$ as $$dQ = M_t dP\ ,$$ then the $\mathbb{R}^d$- valued process $ B_t + V_t$ is a Brownian motion under measure $Q$.\ This implies that the development of $B_t + V_t$ , which is exactly $U_t$, becomes a Brownian motion on $\mathcal O (M)$ under the new measure $Q$. This shows that $X_t = \pi(U_t)$ is a Brownian motion on $M$. Let $\tau^X_r$ be the first exit time of the ball $B(x , r)$ of $X_t$ under the original measure $P$, then $$\mathbb{E} \{ \mathbb{I}( \tau^X_r \leq t ) \} = \mathbb{E}_Q \{ \mathbb{I}( \tau^X_r \leq t ) \frac{dP}{dQ} \}\ .$$ lemma 3.4 implies: $$\begin{aligned} \nonumber \mathbb{E}_Q \{ \mathbb{I}( \tau^X_r \leq t ) \frac{dP}{dQ} \} &= \mathbb{E}_Q \{ \mathbb{I}( \tau_r \leq t ) \frac{dP}{dQ} \}\\ \nonumber & \leq \sqrt{\mathbb{E}_Q \{ \mathbb{I}( \tau_r \leq t ) \} \cdot \mathbb{E}_Q (\frac{dP}{dQ})^2}\ . \nonumber\end{aligned}$$ Since $\mathbb{E}_Q (\frac{dP}{dQ})^2$ is bounded, we can find a $ C' > 0$ such that $\mathbb{E}_Q (\frac{dP}{dQ})^2 \leq C'$. Therefore, $$\mathbb{E} \{ \mathbb{I}( \tau^X_r \leq t ) \} \leq C' \cdot \sqrt{\mathbb{P}( \tau_r \leq t)}\ .$$ This implies that $\mathbb{E} \{ \mathbb{I}( \tau^X_r \leq t ) \}$ also has the exponential decay as $t \rightarrow 0$ and the rest of the proof goes as in the Brownian motion case. Since we have checked all the conditions of Theorem 3.1, we get the following conclusion: \[eigenfunctions\] Let $\{ f_i (x) \}$ be an orthonormal (with respect to $d\mu (x)$) sequence of eigenfunctions of $\Delta_w $ with corresponding eigenvalues $- \nu_i$. Let $X$ be a random point of $M$ sampled from measure $d\mu$. Then if $W$ is the random vector: $(f_1(X), \ldots,f_k(X))$, $$d_W(W,Z)\le\left[\max_{1\le i\le k} (\mu_i^{-1})\right] \cdot \mathbb{E} \sqrt{\sum_{i,j=1}^k [ \langle \nabla f_i(X) , \nabla f_j(X) \rangle - \mathbb{E} \langle \nabla f_i(X) , \nabla f_j(X) \rangle ]}\ .$$ Linear spherical symmetric statistics ===================================== It requires further information to bound the $\mathbb{E} | \norm{\nabla f(X)}^2 - \mathbb{E} \norm{\nabla f(X)}^2 |$ term to get a central limit theorem out of Theorem 6. In this section, we turn from a general manifold to specific models where we can do precise calculation. As a warm-up example, we recover a corollary in [@meckes-thesis]. For one dimensional normal distribution, the stein equation is a first order differential equation: $$f'(x) - xf(x) = g(x) - \mathbb{E}g(Z)\ .$$ Since the Stein operator is first order for the univariate normal distribution, the corresponding metric is the total variation distance. In [@meckes-thesis], the author proved the following abstract approximation theorem: \[abscont\]Suppose that $(W,W_t)$ is a family of exchangeable pairs defined on a common probability space with ${\mathsf E}W=0$ and ${\mathsf E}W^2=1$. Suppose there is a random variables $E= E(W)$ and a constant $\lambda$ such that 1. $${\mathsf E}\left[W_t -W\big| W\right]=-\lambda t W+O(t^2),$$\[lindiff3\] 2. $${\mathsf E}\left[(W_t-W)^2\big| W\right]=2\lambda t +E + O(t^2),$$\[quaddiff3\] 3. For each $ \rho > 0$,$$\lim_{t \rightarrow 0} \frac{1}{t} {\mathsf E}[ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0 \ .$$\[cubediff3\] Then $$d_{TV}(W,Z)\le \frac{1}{\lambda} \mathbb{E} \left|E\right|\ ,$$ where $Z\sim N(0,1).$ We restate a corollary of Theorem 2.2 in [@meckes-thesis] with a slightly better constant: Let $X$ be a uniform random point on the scaled n dimensional sphere $\sqrt{n} S^{n-1}$, the sphere in $\mathbb{R}^n$ of radius $\sqrt{n}$. For $\theta \in S^{n-1}$, define $W = {\left\langle X, \theta \right\rangle}$. Then $$d_{TV}(W,Z)\le \frac{2\sqrt{2}}{\sqrt{(n-1)(n+2)}}\ ,$$ where Z is standard Gaussian $N(0,1)$. By spherical symmetry, it suffices to prove for the case $\theta = e_1$, where $e_1$ is the direction vector of the first coordinate. In the original paper [@meckes-thesis], the exchangeable pair was created by rotating $W$ in a random two-dimensional subspace through a deterministic angle $\sin^{-1} (\epsilon)$.\ Let $X_t$ be the Brownian motion on sphere starting at $X$. Since $X$ is sampled by the uniform measure on the sphere, which is the stationary distribution of the Brownian motion. Denote $W_t$, $W$ as the first coordinate component of $X_t$ and $X$. Then it’s easy to verify that $( W_t, W)$ forms a family of exchangeable pairs. The corresponding generator is $\Delta_{\sqrt{n} S^{n-1}}$ with no drift term. By lemma 3.2, we have $$\begin{aligned} \nonumber \mathbb{E} [ W_t - W | W] & = \Delta_{\sqrt{n} S^{n-1}} W + O( t^2)\\ \nonumber & = \frac{1}{n} \Delta_{S^{n-1}} W + O(t^2) \\ \nonumber & = \frac{1}{n} r^2 [ \Delta_{\mathbb{R}^n} W - \frac{\partial^2}{\partial^2 r } W - \frac{n - 1}{r} \frac{\partial}{\partial r} W ]\\ \nonumber & = - \frac{n-1}{n} W\ . \nonumber\end{aligned}$$ By lemma 3.3, we have $$\mathbb{E} [ (W_t - W)^2 | W ] = t \norm{\nabla X_1}^2 + O( t^2)\ .$$ Let $E = \norm{\nabla X_1}^2 - \mathbb{E} \norm{\nabla X_1}^2 $, then the only thing left is to bound the variance of $E$. We need the following proposition for integrating polynomials over spheres. Let $P(x)=|x_1|^{\alpha_1}$. Then if $X$ is uniformly distributed on $\sqrt{n} S^{n-1}$, $$\mathbb{E}\big[P(X)\big]=\frac{\Gamma(\beta_1)\Gamma(\frac{n}{2}) n^{\frac{1}{2} \alpha_1} (\Gamma(\frac{1}{2})^{n-1}}{\Gamma(\beta_1+\frac{n-1}{2})\pi^{n/2}},$$ where $\beta_1=\frac{1}{2}(\alpha_1+1)$ and $$\Gamma(t)=\int_0^\infty s^{t-1}e^{-s}ds=2\int_0^\infty r^{2t-1}e^{-r^2}dr.$$ When $\alpha_1 = 4$, we have $$\mathbb{E}\big[x_1^4] = \frac{3n}{n+2}\ .$$ Now we are ready to calculate the variance: Let $E$ be defined as above. Then $$Var [E]= \frac{2(n-1)}{n^2 (n+2)} \ .$$ $$\begin{aligned} \nonumber Var [E] & = \mathbb{E} || \nabla X_1 ||^4 - ( \mathbb{E} || \nabla X_1 ||^2 )^2\\ & = \mathbb{E} || \nabla X_1 ||^4 - \frac{(n-1)^2}{n^2}\ . \end{aligned}$$ Note that the covariant derivative on $\sqrt{n} S^{n-1}$ is the projection of the $\mathbb{R}^n$- gradient to the tangent space, which is the hyperplane orthogonal to the radial vector $\frac{\vec{x}}{||\vec{x}||}$. We have $$\begin{aligned} \nonumber || \nabla X_1 ||^2 &= < \nabla X_1, \nabla X_1> \\ \nonumber & = < \nabla_{\mathbb{R}^n} X_1 - (X \cdot\nabla_{\mathbb{R}^n} X_1)\frac{\vec{x}}{||\vec{x}||}, \nabla_{\mathbb{R}^n} X_1 - (X \cdot\nabla_{\mathbb{R}^n} X_1) \frac{\vec{x}}{||\vec{x}||}> \\ \nonumber & = ||\nabla_{\mathbb{R}^n} X_1 ||^2 - \frac{1}{n} (X \cdot\nabla_{\mathbb{R}^n} X_1)^2\\ \nonumber & = 1 - \frac{1}{n} X_1^2\ . \nonumber\end{aligned}$$ From proposition 8, $$\mathbb{E}[X_1^2] =1,\ \ \ \ \mathbb{E}[X_1^4] = \frac{3n}{n+2}$$ This implies $$\begin{aligned} \nonumber \mathbb{E}|| \nabla X_1 ||^4 & = 1 - \frac{2}{n} \mathbb{E} X_1^2 + \frac{1}{n^2} \mathbb{E} X_1^4 \\ \nonumber & = 1 - \frac{2}{n} + \frac{3}{n(n+2)}\ . \nonumber\end{aligned}$$ Plugging in (4.1), we get $$\begin{aligned} \nonumber Var [E] & = \mathbb{E} || \nabla X_1 ||^4 - \frac{(n-1)^2}{n^2}\\ \nonumber & = \frac{2(n-1)}{n^2 (n+2)}\ . \nonumber\end{aligned}$$ Since $X_1$ is an eigenfunction of $\Delta_{\sqrt{n} S^{n-1}}$ with eigenvalue $- \frac{n -1 }{n}$, we can apply Theorem 4.1: $$\begin{aligned} \nonumber d_{TV}(W,Z) & \leq \frac{2n}{n-1}\mathbb{E} |E| \\ \nonumber & \leq \frac{2n}{n-1} \sqrt{Var [E]}\\ \nonumber & \leq \frac{2n}{n-1} \sqrt{\frac{2(n-1)}{n^2 (n+2)}}\\ \nonumber & \leq \frac{2\sqrt{2}}{\sqrt{(n-1)(n+2)}}\ . \nonumber\end{aligned}$$ In [@meckes-thesis], $W_{\epsilon}$ was a rotation of $W$ by a deterministic angle, it’s obvious that the $\mathbb{E} | W_\epsilon - W|^3 $ is of order $O(\epsilon^3)$. On the other hand, our perturbation is by Brownian motion and the corresponding generator $\Delta$ is rotational invariant, so the spherical symmetry is preserved. Although we can calculate the third moment by proposition 4.3 in this case, it’s easier to verify the $\lim_{t \rightarrow 0} \frac{1}{t} {\mathsf E}[ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0 $ condition rather than bounding the higher moments in more complex cases. Approximate exponentiality of $|\operatorname{Tr}(U^k)|^2$ ========================================================== Let $U_n$ denotes the unitary group of dimension $n \times n$. In [@hr], the authors investigated the linear statistics of unitary group. We briefly review the relevant background. Let $g \in L^1 (T)$ be real valued function on the one dimensional torus. We can view $g$ as a $2\pi$-periodic function on $\mathbb{R}$ by identifying $g(e^{ix})$ with $g(x)$. Then $$\operatorname{Tr}g(U) := \sum_{j =1}^{j=n} g(\theta_j) \ ,$$ where $e^{i \theta_j},\ \ 1 \leq j \leq n$ are the eigenvalues of $U_n$. We can expand $g$ into Fourier series: $g(\theta) = \sum_{j \in \mathbb{Z} } g_k e^{ik\theta}$, then we have $$\operatorname{Tr}g(U) = \sum_{j \in \mathbb{Z}} g_k \operatorname{Tr}(U^k)\ .$$ In this way, the randomness is concentrated in the $\operatorname{Tr}(U^k )= \sum_{j =1}^{n} e^{ik\theta_j} $ for each frequency $k$, which can be seen as random Fourier basis functions. Thus it’s worth studying how the norm of those basis functions behave quantitatively as the dimension $n \rightarrow \infty$. In [@jn], the authors showed that as n tends to infinity, the distribution of $|\operatorname{Tr}(U_n)|^2$ tends to an exponential distribution with mean equals to 1. In this section, we will show that for a general power k, $|\operatorname{Tr}(U^k)|^2$ tends to an exponential distribution by an Infinitesimal version of stein method for exponential distribution. Moreover, the setting here can be generalized to general circular ensemble without any change, which will be the main topic of next section. Abstract approximation theorem for exponential distribution ----------------------------------------------------------- The first step towards establishing the stein method for exponential distribution is to find an operator $\mathcal{A}$, which acts on a large enough class of functions and ’characterizes’ the exponential distribution in the sense that $$\mathcal{A}f(Z)=0\ ,$$ for all $f$ in a large enough class of functions and $Z$ follows the exponential distribution of mean 1. It turns out the characterizing operator for the exponential distribution is of first order: $$\mathcal{A}f(x) = xf' (x) - (x - 1)f(x)\ .$$ It may be worth pointing out that the $-(x - 1)$ part of the Stein operator has a connection with the scale score function (see [@cl]) of the exponential distribution: $$\frac{d}{d \lambda} \log ( \lambda \exp{- \lambda x}) = \frac{1}{\lambda} - x\ .$$ For functions $h$ in the class of interest, define $f_h$ to be the solution of the Stein equation $$\mathcal{A}f_h(x)=h(x)- \mathbb{E}h(Z).$$ An iterative technique for bounding derivatives of solutions of Stein equations has been developed. However, for the infinitesimal version of exchangeable pair, we only need to bound the first derivative and the higher order term is automatically cut off. We restate the following fundamental lemma 2.2 in [@jn]: Let $Z$ be a mean one exponential random variable. If $h$ is a function such that the following integrals are well defined, then $$f(w) =f_h(w)=-\frac{e^w}{w}\int_w^\infty (h(x)-\mathbb{E} h(Z))e^{-x}dx \label{stnsol}$$ solves the differential equation $$wf'(w)-(w-1)f(w)=h(w)-\mathbb{E} h(Z). \label{stneq}$$ If $h$ is absolutely continuous with $\norm{h'}<\infty$, then $$\norm{f}\leq \left(1+\frac{2}{e}\right)\norm{h'}, \hspace{5mm} \norm{f'}\leq 2\norm{h'}. \label{abcobd}$$ Note that in our case, since we don’t have a bound of the form $\norm{f'} \leq C \norm{h}$, we won’t expect to have an approximation measured by the total variation distance. Instead, we need to introduce the following class of functions for $t,x \geq 0$, and $\delta \geq 0$, $$\label{htd} h_{t,\delta}(x) = \left\{ \begin{array}{ll} 1, & x \leq t-\delta,\\ 1-\frac{2(x-t+\delta)^2}{\delta^2}, &t-\delta<x\leq t-\delta/2 ,\\ \frac{2(x-t)^2}{\delta^2},&t-\delta/2<x\leq t ,\\ 0, & x > t . \end{array} \right.$$ We can derive the following bound on the Kolmogolov distance (see lemma 2.3 in [@jn]): \[lemhtd\] If $t\geq0$, $\delta>0$, and $h_{t,\delta}$ is defined by (5.1), then $$\norm{h_{t,\delta}}=1, \hspace{5mm} \norm{h_{t,\delta}'}=2/\delta, \hspace{5mm} \norm{h_{t,\delta}''}=4/\delta^2.$$ If $W\geq0$ is a random variable and $Z$ has the exponential distribution with mean one, then $$d_K(W,Z)\leq\sup_{t\geq0}\abs{\mathbb{E} h_{t,\delta}(W)-\mathbb{E} h_{t,\delta}(Z)}+\delta/2. \label{smo}$$ Now we are ready to prove the infinitesimal version of approximation theorem for exponential distribution with mean equals to 1: \[abscont\]Let Z be a mean 1 exponential distributed random variable. If $W \geq 0$ is random variable with finite second moment and $(W,W_t)$ is a family of exchangeable pairs defined on a common probability space. Let $F$ be a $\sigma -$algebra s.t $\sigma (W) \subseteq F$ and Suppose there are $F-$measurable random variables $E$ and $E'$ such that 1. $$\mathbb{E}\left[W_t -W\big| F\right]=\Lambda t(1-W)+tE + O(t^2),$$\[lindiff3\] 2. $$\mathbb{E}\left[(W_t-W)^2\big| F\right]=2\Lambda t W +tE'+ O(t^2),$$\[quaddiff3\] 3. For each $ \rho > 0$,$$\lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0 \ .$$\[cubediff3\] Then, for $\forall \delta \geq 0$, $$d_{K}(W,Z)\le \frac{1}{\Lambda \delta}\left[2 \mathbb{E}\left|E'\right|+(1+\frac{2}{e})\mathbb{E}\left|E\right|\right] + \frac{\delta}{2}\ . \eqno{(5.2)}$$ Since $(W_t, W)$ is an exchangeable pair, we have the identity $$\begin{aligned} \nonumber 0 &= \mathbb{E} [ (W_t - W) (f(W_t) + f(W))]\\ \nonumber & = \mathbb{E} [ (W_t - W) (f(W_t) - f(W)) + 2(W_t - W) f(W)] \ , \nonumber\end{aligned}$$ by Taylor expansion, $$0 = \mathbb{E} \{ \mathbb{E} [ (W_t - W)^2 | W] f' (W) + 2 \mathbb{E} [(W_t - W)| W] f(W)] + R\}\ .$$ For the reminder term $R$, there exists a real number $K$ depending on the function $f$, such that $$|R| \leq K |W_t - W|^3\ .$$ Fix $ \rho > 0$, decomposing the integrand into two sets $$\{|W_t - W| \leq \rho \} \ \ and \ \ \{|W_t - W| \ge \rho \},$$ we obtain $$\begin{aligned} \nonumber \lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} |R| &\leq \lim_{t \rightarrow 0} \frac{K}{t} \mathbb{E} [ |W_t - W|^3 \mathbb{I}( |W_t - W| \leq \rho ) + |W_t - W|^3 \mathbb{I}( |W_t - W| > \rho )]\\ \nonumber & \leq \lim_{t \rightarrow 0} K \rho \frac{\mathbb{E} |W_t - W|^2}{t} + \lim_{t \rightarrow 0} \frac{K}{t} \mathbb{E} [ |W_t - W|^2 \mathbb{I}( |W_t - W| > \rho )] \ , \nonumber\end{aligned}$$ by condition (3), the second term is zero and by condition (2), $$\lim_{t \rightarrow 0} \frac{K \rho}{t} \mathbb{E} | W_t - W|^2 \leq cK\rho$$ for some constant c that depends on the distribution of $W$. Let $\rho \rightarrow 0$, $$\lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E}|R| = 0\ .$$ Now we are left with two terms, dividing both side by t and take the limit, we have $$0 = \mathbb{E} [ 2 \Lambda W f'(W) + E' f'(W) + 2 \Lambda (1-W) f(W) + E f(W) ]\ .$$ Since $f$ is the solution of the stein equation with respect to $g(x)$, this implies $$0 = 2 \Lambda \mathbb{E} [ g(W) - \mathbb{E} g(Z) ] + \mathbb{E} [ E' f'(W) + Ef(W)]\ ,$$ by lemma 5.1, we have $$\mathbb{E} [ g(W) - \mathbb{E} g(Z) ] \leq \frac{1}{2 \Lambda} ( 1+ \frac{2}{e}) \norm{g'} \cdot \mathbb{E} |E| + \frac{1}{\Lambda} \norm{g'} \mathbb{E} |E'|\ .$$ Let $g(x)$ be in the function class of $h_{t,\delta}$, then (5.2) leads to $$\begin{aligned} \nonumber d_K(W,Z)&\leq\sup_{t\geq0}\abs{\mathbb{E} h_{t,\delta}(W)-\mathbb{E} h_{t,\delta}(Z)}+\delta/2 \\ \nonumber &\leq \frac{1}{\Lambda \delta} (1+ \frac{2}{e}) \mathbb{E} |E| + \frac{2}{\Lambda \delta} \mathbb{E} |E'| +\delta/2 \ . \nonumber\end{aligned}$$ Approximate normality of $|\operatorname{Tr}(U^k)|^2$ ----------------------------------------------------- To apply Theorem 5.3, the following lemma on the moment calculation of traces of unitary group is crucial (see [@hr; @ds]): Let $U$ be Haar distributed on $U(n,\mathbb{C})$. Let $(a_1,\cdots,a_k)$ and $(b_1,\cdots,b_k)$ be vectors of non-negative integers. Then one has that for all $n \geq \sum_{i=1}^k (a_i+b_i),$ $$\left[ \prod_{j=1}^k Tr(U^j)^{a_j} \overline{Tr(U^j)^{b_j}} \right] = \delta_{\vec{a} \vec{b}} \prod_{j=1}^k j^{a_j} a_j!.$$ Let $\lambda = (\lambda_1, \lambda_2, \dots )$ be a partition, that is, the sequence is in non-increasing order and only finite of $\lambda_i$ are nonzero. Let $m_j$ denotes the multiplicity of part j in $\lambda$ and $l(\lambda)$ the length of $\lambda$: $l(\lambda) = m_1(\lambda) + m_2(\lambda) + \dots,$ then $$p_{\lambda} (U) : = \prod_{i = 1}^{l(\lambda)} \operatorname{Tr}(U^i )^{m_i} \ .$$ We need a lemma from [@Lt] for the expression of the gradient and also the Laplacian of the trace powers of unitary group in terms of $p_{\lambda}$: Let $j > 0$ an integer, $p_j(U(n))$ be as above. Then $\overline{p_j} = p_{-j}$ and 1. $$\Delta_{U(n)} p_j = - nj p_j - j \sum_{1 \leq l <j} p_{l,j-l}.$$ 2. $$\Delta_{U(n)} p_{j,j} = - 2nj p_{j,j} - 2j^2 p_{2j} - 2j p_j \sum_{1 \leq l <j} p_{l,j-l} .$$ 3. $$\Delta_{U(n)} \left( p_j \overline{p_j} \right) = 2j^2n - 2njp_j \overline{p_j} - jp_j \sum_{1 \leq l < j} \overline{p_{l,j-l}} - j \overline{p_j} \sum_{1 \leq l < j} p_{l,j-l} .$$ Form this we can easily deduce the formula for the gradient of trace power functions: $$\begin{aligned} \nonumber (\nabla p_k(U)) \cdot (\nabla p_k(U)) &= \frac{1}{2}[ \Delta( p_k(U)^2) - 2 p_k(U) \Delta p_k(U)]\\ \nonumber & = -k^2 \cdot p_{2k}(U) \ . \nonumber\end{aligned}$$ Let $W(U) = |\operatorname{Tr}(U^k)|^2 / k = \operatorname{Tr}(U^k)\cdot \overline{\operatorname{Tr}(U^k)} / k$. We plan to prove that $W$ tends to the exponential distribution by checking the three conditions and identifying $\Lambda$, $E$ and $E'$ of Theorem 5.3. The first step is to perturb $W$ by running a Brownian motion starting at $U$ and we denote it by $U_t$. Let $W_t = |\operatorname{Tr}(U_t^k)|^2 / k$, then $(W_t , W)$ forms an exchangeable pair under the Haar measure of the unitary group. Since $ W_t = e^{\Delta t} W$, we have $$\begin{aligned} \nonumber \mathbb{E} [ W_t - W | W] & = t \Delta (p_k \overline{p_k} / k) + O(t^2)\\ \nonumber & = 2n( k - p_k \overline{p_k} ) - p_k \sum_{1 \leq l < k} \overline{p_{l,k-l}} - \overline{p_k} \sum_{1 \leq l < k} p_{l,k-l}\\ \nonumber &= 2nk( 1 -W) - p_k \sum_{1 \leq l < k} \overline{p_{l,k-l}} - \overline{p_k} \sum_{1 \leq l < k} p_{l,k-l} \ . \nonumber\end{aligned}$$ Unlike the previous case, the residue term $E : = p_k \sum_{1 \leq l < k} \overline{p_{l,k-l}} + \overline{p_k} \sum_{1 \leq l < k} p_{l,k-l}$ is not zero.\ By lemma 3.3, we have $$\begin{aligned} \nonumber \mathbb{E} [ (W_t - W)^2 | W] & = 2 t \nabla W \cdot \nabla W + O(t^2)\\ \nonumber &= 2t \nabla p_k \cdot \nabla p_k \cdot (\overline{p_k} \cdot \overline{p_k})/ k^2 \nonumber + 4t \nabla p_k \cdot \nabla \overline{p_k} \cdot ( p_k \cdot \overline{p_k}) / k^2 \\ \nonumber &\ \ + 2t \nabla \overline{p_k} \cdot \nabla \overline{p_k} \cdot (p_k \cdot p_k) / k^2 + O(t^2) \\ \nonumber & = -2 p_{2k} (\overline{p_k})^2 + 4 n p_k \overline{p_k} - 2 \overline{p_{2k}} (p_{2k})^2 + O(t^2)\\ \nonumber & = 4knW - 2p_{2k} (\overline{p_k})^2 - 2 \overline{p_{2k}} (p_k )^2 + O(t^2) \ . \nonumber\end{aligned}$$ We find out that $\Lambda = 2kn$, $E' = - 2p_{2k} (\overline{p_k})^2 - 2 \overline{p_{2k}} (p_k )^2$. The next task is to derive upper bounds for the norm of the two residue terms $E$ and $E'$. $$\begin{aligned} \nonumber \mathbb{E} [ E^2] &= \mathbb{E} [p_k^2 ( \sum_{1 \leq l < k} \overline{p_{l,k-l}})^2] + \mathbb{E}[\overline{p_k}^2 ( \sum_{1 \leq l < k} p_{l,k-l})^2] \\ \nonumber &\ \ \ \ + 2 p_k \overline{p_k} ( \sum_{1 \leq l < k} \overline{p_{l,k-l}} \cdot \sum_{1 \leq l < k} p_{l,k-l}) \ , \nonumber\end{aligned}$$ by lemma 5.5 and only keep terms with non-zero expectation, we have $$\begin{aligned} \nonumber \mathbb{E} [ E^2] &= \mathbb{E} \sum_{1 \leq l < k} (p_{l,k -l} \cdot p_k) \cdot \sum_{1 \leq l < k} (\overline{p_{l,k - l}} \cdot \overline{p_k}) \\ \nonumber & = \left\{ \begin{array}{lr} \sum_{1 \leq l < k} l \cdot (k - l) = \frac{k^3 - k}{6}, \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ if\ k\ is \ odd& \\ \sum_{1 \leq l < k} l \cdot (k - l) + \frac{k^2}{4} = \frac{2k^3 + 3k^2 - 2k}{12}. \ \ if\ k\ is \ even & \end{array} \right.\end{aligned}$$ It implies that for all $k > 0$, $$\mathbb{E} [ E^2] \leq \frac{2k^3 + 3k^2 - 2k}{12}\ ,$$ and $$\begin{aligned} \nonumber \mathbb{E} [ (E')^2] &= 4 \mathbb{E} [ (p_{2k})^2 (\overline{p_k})^4] + 4 \mathbb{E} [ (\overline{p_{2k}})^2 (p_k)^4]\\ \nonumber &\ \ \ + 8 \mathbb{E} [p_{2k} (p_k)^2 \cdot \overline{p_{2k}} (\overline{p_{k}})^2] \ . \nonumber\end{aligned}$$ By lemma 5.5, the first two terms vanish, we have $$\begin{aligned} \nonumber \mathbb{E} [ (E')^2] &=8 \mathbb{E} [p_{2k} \overline{p_{2k}} \cdot (p_k \overline{p_{k}})^2] \\ \nonumber & = 32 k^3\ . \nonumber\end{aligned}$$ Finally, by Theorem 5.3, we conclude that for $\forall \delta > 0$, $$\begin{aligned} \nonumber d_{K}(W,Z) &\le \frac{1}{\Lambda \delta}\left[2 \mathbb{E}\left|E'\right|+(1+\frac{2}{e})\mathbb{E}\left|E\right|\right] + \frac{\delta}{2} \\ \nonumber &\leq \frac{1}{\Lambda \delta}\left[2 \sqrt{\mathbb{E}\left|E'\right|^2}+(1+\frac{2}{e})\sqrt{\mathbb{E}\left|E\right|^2}\right] + \frac{\delta}{2}\\ \nonumber &\leq \frac{1}{2kn\delta}(1+\frac{2}{e})\sqrt{\frac{2k^3 + 3k^2 - 2k}{12}} + \frac{1}{kn\delta}\sqrt{32k^3} + \frac{\delta}{2}\\ \nonumber & \leq (\frac{1}{2} + 4\sqrt{2})\frac{\sqrt{k}}{n\delta} + \frac{\delta}{2}\ .\end{aligned}$$ Choose the optimal $\delta = \sqrt{(1+8\sqrt{2})\frac{\sqrt{k}}{n}}$, we have proved the following theorem: \[main2\] Let $W=|Tr(U^k)|^2$, where $U$ is from the Haar measure of unitary group and $k$ is a positive integer. Let $Z$ denotes the exponential random variable with mean one. Then we have $$d_{K}(W,Z) \le \sqrt{(1+8\sqrt{2})\frac{\sqrt{k}}{n}}\ .$$ General circular ensemble ========================= As we all know, the trace power functions on unitary matrix group are actually functions of the eigenvalues. From this point of view, it’s crucial to study the distribution of eigenvalues induced by the Haar measure of the Lie group. It turns out that the distribution of eigenvalues has the following form: $$p( \theta_1, \cdots , \theta_n ) = \frac{1}{Z_{n,\beta}} \prod_{1 \leq k < j \leq n} | e^{i \theta_k} - e^{i \theta_j }|^{\beta}\ ,$$ where $ ( \theta_1, \cdots , \theta_n ) \in \mathbb{R}^n_{[0,2\pi ]}$ and $Z_{n,\beta}$ is the normalization constant. In the unitary matrix case, $\beta = 2$. The circular orthogonal ensemble ($\beta = 1$) and the circular symplectic ensemble ($\beta = 4$) also belong to this circular ensemble family. In this section, we generalize the results of section 5 to the circular ensemble for all $\beta \ge 1$. Since there is no underlying Lie group for general $\beta$, we cannot use the Brownian motion on the Lie group to characterize it. Instead, we need to find a diffusion process that models the circular ensemble directly. The potential function of circular ensemble is $$H(\theta_1, \cdots , \theta_n ) = \beta \ln \prod_{1 \leq k < j \leq n} | e^{i \theta_k} - e^{i \theta_j }|\ .$$ Let’s order $( \theta_1, \cdots , \theta_n )$ such that $\theta_1 \leq \cdots \leq \theta_n$ in the compact set $\mathbb{R}^n_{[0,2\pi ]}$, then the corresponding drift vector field is $$\begin{aligned} \nonumber \nabla H &= \nabla_i [ \sum_{ 1 \leq k < j \leq n} \beta \ln | e^{i \theta_k} - e^{i \theta_j} | ]\frac{\partial}{\partial \theta_i} \\ & = \nabla_i [ \sum_{ 1 \leq k < i} \beta \ln \left (2 \sin \frac{\theta_k - \theta_i}{2}\right ) + \sum_{ i < k \leq n} \beta \ln \left (2 \sin \frac{\theta_i - \theta_k}{2}\right ) ]\frac{\partial}{\partial \theta_i}\\ \nonumber & = \frac{\beta}{2} \sum_{k \neq i} \cot \left ( \frac{\theta_i - \theta_k }{2} \right ) \ .\end{aligned}$$ So by our formal argument, the diffusion process for a general $\beta$ ensemble satisfies the following stochastic differential equation: $$d X^j_t = \sqrt{2} d B^j_t + \frac{\beta}{2} \sum_{ 1 \leq k \neq j \leq n} \cot \left ( \frac{\theta_j - \theta_k }{2} \right ) dt , \ j = 1,2, \cdots , n.$$ This exactly the circular Dyson Brownian motion (CDBM). The corresponding Witten Laplacian is $\Delta + \frac{\beta}{2} \sum_{ 1 \leq k \neq j \leq n} \cot \left ( \frac{\theta_j - \theta_k }{2} \right ) \frac{\partial}{\partial \theta_j}$, which is the so-called Dyson operator $D$. CDBM is well defined in the open simplex $$\Sigma_{n} := \{ X = (x_1, \dots , x_n ) \in \mathbb{R}^n_{[0,2\pi ]} \ : x_1 < \cdots < x_n \}$$ and $\partial \Sigma_{n}$ is the corresponding boundary set. We summarize the results we need in [@cl] (see also [@lh] ) as the following lemma: Let $\beta \ge 1$ and suppose that the initial data $X(0) \in \Sigma_{n}$. Then there exists a unique solution to (1) in the space of continuous functions $(X(t))_{t \ge 0} \in C( \mathbb{R}_{+} , \Sigma_{n})$. Moreover, the circular ensemble is the unique equilibrium of CDBM and CDBM is reversible with respect to this distribution. Now, let’s define the analogue of $\operatorname{Tr}(U^k)$ for general $\beta$ ensemble and $k \in \mathbb{Z}$: $$p_k (x) = \sum_{j =1}^n e^{ik x_j }, \ \ \ \ x = (x_1,\dots,x_n) \in \Sigma_{n}\ .$$ Note that this is a re-expression of $\operatorname{Tr}(U^k)$ if we set $(e^{ix_1}, \dots , e^{ix_n})$ as the ordered sequence of eigenvalues of unitary matrix and $\overline{ p_k} = p_{-k}$ as before. To state the main theorem of this section, we need to introduce some constants. Set $\alpha = 2/ \beta$, then $C_E$ ,$ C_{E'}$ is defined by $$C_E = \max \{ |A -1| , |B-1| \},\ \ \ \ C_{E'} = \max \{ |A' -1| , |B'-1| \},$$ where $$A = (1 - \frac{|\alpha - 1|}{n-2k + \alpha})^{2k},\ \ \ \ B = (1 + \frac{|\alpha - 1|}{n-2k + \alpha})^{2k},$$ and $$A' = (1 - \frac{|\alpha - 1|}{n-4k + \alpha})^{4k},\ \ \ \ B' = (1 + \frac{|\alpha - 1|}{n-4k + \alpha})^{4k}\ .$$ As in the last section, let Z be an exponential random variable with mean 1. Then we have the following theorem: Let $W= \frac{\beta}{2k} p_k (X) \overline{p_k}(X)$, where X is a random point sampled by the $\beta$ circular ensemble in $\Sigma_n$. Then there exists an integer $N(k,\beta)$ which depends on $k$ and $\beta$, such that for $n \ge N((k,\beta)$, the Kolmogorov distance between W and Z is bounded by: $$2 \sqrt{\frac{1}{\sqrt{\beta} n } \sqrt{80 C_{E'} k} + \frac{1}{ \beta n} (1+ \frac{2}{e}) \sqrt{2 C_{E'} k^3}}\ .$$ We need to let $n \ge N(k,\beta)$ to make the constants $A$ and $A'$ positive and less than one. This will be important in lemma 6.2 and lemma 6.3. We denote the CDBM starting from the random point X at time t as $X_t$ and let $W_t =\frac{\beta}{2k} p_k (X_t) \overline{p_k}(X_t)$. By the argument in section 2, we know that $(W,W_t)$ forms a continuous family of exchangeable pairs. To apply the Stein method, we need to generalize formulas in lemma 5.5 to general $\beta$ ensemble. Inspired by (1), we make an educated guess that $D p_j$ is a combination of $p_j$ and $\sum_{0 \leq l < j} p_{l}p_{j-l}$. We observe that $$\begin{aligned} \nonumber \sum_{l = 0}^{j} p_{l}p_{j-l} & = \sum_{l = 0}^{j} (\sum_{a=1}^{n} e^{ilx_a})\cdot (\sum_{b=1}^{n} e^{i(j-l)x_b}) \\ \nonumber & = \sum_{x_a \neq x_b} \frac{e^{i(j+1)x_a} - e^{i(j+1)x_b} }{e^{ix_a} - e^{ix_b}} + (j+1)p_j\\ \nonumber & = \sum_{x_a \neq x_b} \frac{e^{ijx_a} - e^{ijx_b} }{e^{ix_a} - e^{ix_b}}(e^{ix_a} + e^{ix_b}) - \sum_{x_a \neq x_b} \frac{e^{i(j-1)x_a} - e^{i(j-1)x_b} }{e^{ix_a} - e^{ix_b}} e^{i(x_a +x_b)}+ (j+1)p_j\ .\\ \nonumber\end{aligned}$$ Notice the symmetry between index a and b of the first term, we have $$\sum_{x_a \neq x_b} \frac{e^{ijx_a} - e^{ijx_b} }{e^{ix_a} - e^{ix_b}}(e^{ix_a} + e^{ix_b}) = 2 \sum_{x_a \neq x_b} \frac{e^{ijx_a} - e^{ijx_b} }{e^{ix_a} - e^{ix_b}} e^{ix_a}\ .$$ As for the second term, expand the $ e^{i(j-1)x_a} - e^{i(j-1)x_b}$ part, we have $$\begin{aligned} \nonumber \sum_{x_a \neq x_b}\frac{e^{i(j-1)x_a} - e^{i(j-1)x_b} }{e^{ix_a} - e^{ix_b}} e^{i(x_a +x_b)} &= \sum_{x_a \neq x_b} \{e^{i (l-1) x_a} e^{ix_b} + \cdots + e^{ix_a} e^{i(j - 1)x_b}\} \\\ \nonumber & = \sum_{x_a \neq x_b}\{\frac{e^{i(j+1)x_a} - e^{i(j+1)x_b} }{e^{ix_a} - e^{ix_b}}- e^{ijx_a} - e^{ijx_b}\} \\ \nonumber & = \sum_{l = 0}^{j} p_{l}p_{j-l} -2(n-1)p_j\ . \nonumber\end{aligned}$$ Combining (3) with (1) (2), we find that $$\begin{aligned} \nonumber \sum_{l = 0}^{j} p_{l}p_{j-l} & = \sum_{x_a \neq x_b} \frac{e^{ijx_a} - e^{ijx_b} }{e^{ix_a} - e^{ix_b}} e^{ix_a} + (n+j)p_j \\ \nonumber & = \sum_{x_a \neq x_b} \frac{e^{ix_a} + e^{ix_b} }{e^{ix_a} - e^{ix_b}} e^{ijx_a} + (n+j)p_j \\ \nonumber & = - i\sum_{x_a \neq x_b} \cot\left(\frac{x_a - x_b}{2}\right)e^{ijx_a} + (n+j)p_j \ .\\ \nonumber\end{aligned}$$ Now we are ready to calculate $D p_j$: $$\begin{aligned} \nonumber D p_j & = \frac{\beta}{2} \sum_{x_a \neq x_b}\cot\left(\frac{x_a - x_b}{2}\right) ij e^{ijx_a} - j^2 p_j\\ \nonumber & = -\frac{j \beta}{2} ( \sum_{l = 0}^{j} p_{l}p_{j-l} - (n+j)p_j) - j^2 p_j\\ & = -\frac{j \beta}{2} \sum_{l = 1}^{j-1} p_{l}p_{j-l} + (\frac{j^2 \beta}{2} - \frac{nj\beta}{2} - j^2)p_j \ .\end{aligned}$$ When $\beta = 2$, we recover formula (1) of lemma 5.5. To verify conditions of Theorem 5.3, we need to calculate $D (p_j \overline{p_j})$. From (6.2), it’s easy to derive the following formula: $$D (p_j \overline{p_j}) = -n\beta j p_j \overline{p_j} - (2- \beta)j^2 p_j \overline{p_j} + 2 j^2 n - \frac{\beta}{2} j \sum_{l = 1}^{j-1}p_{-j} p_{l-j} p_j - \frac{\beta}{2} j \sum_{l = 1}^{j-1}p_{l} p_{j-l}\overline{p_j}\ . \eqno{(6.3)}$$ By (1) and (6.3), we have $$\begin{aligned} \nonumber \mathbb{E} [ W_t - W | W] & = t D (\frac{\beta}{2k} p_k \overline{p_k}) + O(t^2)\\ \nonumber & = t[ \beta k n ( 1 - W) - \frac{\beta (2 - \beta)}{2}\cdot k p_k \overline{p_k} - \frac{\beta^2}{4} \sum_{l = 1}^{k-1}p_l p_{k-l} \cdot \overline{p_k} - \frac{\beta^2}{4} \sum_{l = 1}^{k-1} p_{-l}p_{l-k} \cdot p_k] + O(t^2) \ .\end{aligned}$$ Therefore, condition 1 of Theorem 5.3 is satisfied with $\Lambda = \beta k n$ and $$E = - \frac{\beta (2 - \beta)}{2}\cdot k p_k \overline{p_k} - \frac{\beta^2}{4} \sum_{l = 1}^{k-1}p_l p_{k-l} \cdot \overline{p_k} - \frac{\beta^2}{4} \sum_{l = 1}^{k-1} p_{-l}p_{l-k} \cdot p_k\ .$$ In order to identify $E'$ for the second condition of Theorem 5.3, by (2.3), we have $$\begin{aligned} \nonumber \mathbb{E} [ (W_t - W)_i (W_t - W)_j | W ] &= 2 t \langle \nabla W , \nabla W \rangle + O(t^2)\\ \nonumber & = t[ 2\beta k n W - \frac{\beta^2}{2} p_{2k}\cdot (p_{-k})^2 - \frac{\beta^2}{2} p_{-2k}\cdot (p_k)^2] + O(t^2) \ .\end{aligned}$$ This implies that condition 2 is satisfies with $$E' = - \frac{\beta^2}{2} p_{2k}\cdot (p_{-k})^2 - \frac{\beta^2}{2} p_{-2k}\cdot (p_k)^2\ .$$ To check condition 3 of Theorem 5.3, we must be careful when applying lemma 3.5 to CDBM. Since the drift of CDBM becomes singular when the random particle approaches the boundary of $\sum\nolimits_{n}$, the convergence rate of $\frac{1}{t} \mathbb{E}[ (W_t - W)^2 \mathbb{I}( (W_t - W)^2 > \rho | W ]$ is not uniform. However, from the proof of lemma 3.3, we can extract the following uniform bound: $$\begin{aligned} \nonumber \frac{1}{t} \mathbb{E}[ (W_t - W)^2 \mathbb{I}( (W_t - W)^2 > \rho | W ] & \leq \frac{1}{t} \mathbb{E}[ (W_t - W)^2 | W ] \\ \nonumber & \leq \norm{ \nabla W \cdot \nabla W}_{max} + 2 \norm{W}_{max} \cdot \norm{ D_c W }_{max} \ .\\ \nonumber\end{aligned}$$ From (6.3), $D_c W$ is a smooth function in the closure of $\sum\nolimits_{n}$, so $\norm{ D_c W }_{max}$ is well-defined. The dominated convergence theorem implies that $$\lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = \mathbb{E} \lim_{t \rightarrow 0}\frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) | W]\ .$$ To calculate $\lim_{t \rightarrow 0}\frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) | W]$, notice that we can find a $C > 0$ s.t $$| W_t - W | \leq C \cdot | X_t - X|\ .$$ Define the cut-off radius $ r_X = \min \{ \frac{\sqrt{\rho}}{C}, \frac{1}{2}d \left (X,\partial \Sigma_{n} \right) \}$, where $d \left (X,\partial \Sigma_{n} \right)$ is the distance between $X$ and $\Sigma_{n}$. As before, $\tau_{r_X}$ is the first exit time of ball $B(X ,r_X)$. Then it’s obvious that $$\mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) | W] \leq \mathbb{E} [ | W_t - W |^2 \mathbb{I}(\tau_{r_X} \leq t) | W]\ .$$ Furthermore, calculating the exit time from a compact set is a pure local thing. More precisely, we can introduce the stopped process $(U_n)_t = U_{t \wedge \tau_{r_X}}$ and denote the exit time for the stopped process as $\tau_{r_X}'$. Then $$\mathbb{P}( \tau_{r_X}' \leq t) = \mathbb{P}( \tau_{r_X}\leq t)\ .$$ Since the drift of the stopped process is bounded, the argument for proving lemma 3.5 applies. We conclude that $$\lim_{t \rightarrow 0}\frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) | W] = 0\ .$$ It follows that $$\lim_{t \rightarrow 0} \frac{1}{t} \mathbb{E} [ | W_t - W |^2 \mathbb{I}( | W_t - W |^2 > \rho ) ] = 0\ .$$ The remaining tasks are to estimate the second moment of $E$ and $E'$. Following the notations in [@ts], let $\rho = (\rho_1, \rho_2, \dots )$ be a partition, then the weight of $\rho$ is $|\rho| = \rho_1 + \rho_2 + \cdots$. For a partition $\rho$, $p_{\rho}$ is defined by (2.2) of [@ts]. To deal with the $E$ variable, We need to bound terms of the form $\mathbb{E} [ p_{\mu} \overline{p_{\nu}}]$ for partitions $\mu,\nu$ where $|\mu| = |\nu| =2$. Note that for general $\beta$, we don’t have precise formula like lemma 5.4. By Theorem 1 of [@ts] and some calculation, we get the following upper bound: Let $C_E$ be defined as in Theorem 6.1. Then for $0 < l,j \leq k -1 $, $$|\mathbb{E} [(p_k \overline{p_{k}})^2] \leq \frac{8 C_E}{\beta^2} k^2\ ;$$ $$|\mathbb{E} [p_l \cdot p_{k-l} \cdot p_j \cdot p_{k-j} \cdot \overline{p_k} \cdot \overline{p_k}]| \leq \frac{8 \sqrt{3} C_E}{\beta^3} k^3\ ;$$ $$|\mathbb{E} [p_{l} \cdot p_{k-l} \cdot p_{k} \cdot \overline{p_k} \cdot \overline{p_k}]| \leq \left( \frac{2}{\beta}\right)^{\frac{5}{2}} C_E \cdot k^2 \sqrt{k}\ ;$$ $$|\mathbb{E} [(p_{-l} \cdot p_{l-k} \cdot \overline{p_k} \cdot p_{j} \cdot p_{k-j} \cdot p_{k}] \leq \frac{4 C_E}{\beta^3} k^3\ .$$ From (10), $$\begin{aligned} \nonumber \mathbb{E}[E^2] &\leq\ \ \ \frac{\beta^2 (2 - \beta)^2}{4} k^2 (p_k \overline{p_k})^2 + \frac{\beta^4}{16} (\sum_{l=1}^{k-1} p_l p_{k-l} \overline{p_k})^2 + \frac{\beta^4}{16} (\sum_{l=1}^{k-1} p_{-l} p_{l-k} p_k)^2 \\ \nonumber & \ \ + \frac{\beta^3 (2 - \beta)}{4} \left(\sum_{l=1}^{k-1}k p_l p_{k-l} p_k (\overline{p_k})^2 + \sum_{l=1}^{k-1} k p_{-l} p_{l-k} \overline{p_k} (p_k)^2 \right) + \frac{\beta^4}{8} (\sum_{l=1}^{k-1} p_{-l} p_{l-k} p_k ) \cdot (\sum_{l=1}^{k-1} p_l p_{k-l} \overline{p_k}) \ .\\ \nonumber\end{aligned}$$ By the previous lemma, we have $$\begin{aligned} \nonumber \mathbb{E}[E^2] &\leq 2(2 - \beta)^2 C_E \cdot k^4 + \sqrt{3} \beta C_E k^3 (k-1)^2 + |4-2\beta|\sqrt{2\beta} C_E k^3 (k-1) \sqrt{k} + \frac{\beta C_E}{2} k^3 (k-1)^2 \\ & \leq 8 \beta^2 C_E \cdot k^5 \ . \tag{6.4} \end{aligned}$$ Now, for the $E'$ variable, we need to bound $\mathbb{E} [ p_{\mu} \overline{p_{\nu}}]$ for partitions $\mu,\nu$ where $|\mu| = |\nu| =4$. Applying Theorem 1 of [@ts] again, we have the following result. Let $C_E'$ be defined as in Theorem 6.1. Then for $0 < l,j \leq k -1 $, $$|\mathbb{E} [p_{2k} \cdot p_{2k} \cdot (p_{-k})^4 ]| \leq \frac{64 \sqrt{3} C_{E'}}{\beta^3} k^3\ ;$$ $$|\mathbb{E} [p_{-2k} \cdot p_{2k} \cdot (p_{-k})^2 \cdot (p_{k})^2]| \leq \frac{32 C_{E'}}{\beta^3} k^3\ .$$ It follows that $$\begin{aligned} \nonumber \mathbb{E}[E'^2] &= \frac{\beta^4}{4} ( p_{-k}^4 \cdot p_{2k} p_{2k} + p_k^4 \cdot p_{ -2k}p_{-2k} + 2 p_{-2k}p_{-k}^2 \cdot p_{2k} p_k^2) \\ \nonumber & \leq 32 \sqrt{3} \beta C_{E'} \cdot k^3 + 16 \beta C_{E'} \cdot k^3 \\ \nonumber & \leq 80 \beta C_{E'} \cdot k^3 \ . \tag{6.5}\end{aligned}$$ Combining (6.4) (6.5) with Theorem 5.3, we conclude that for $\forall \delta > 0$, $$d_K (W , Z) \leq \frac{1}{\beta k n \delta} (1+ \frac{2}{e}) \sqrt{8 \beta^2 C_E k^5} + \frac{2}{\beta k n \delta} \sqrt{80 \beta C_{E'} k^3} + \frac{\delta}{2}\ .$$ Let $\delta$ be the optimal value such that the right hand side achieves the minima, then we get $$d_K (W , Z) \leq 2 \sqrt{\frac{1}{\sqrt{\beta} n } \sqrt{80 C_{E'} k} + \frac{1}{ \beta n} (1+ \frac{2}{e}) \sqrt{2 C_{E'} k^3}}\ .$$ [**Acknowledgements.** ]{}The author is grateful to Prof. Elton Hsu for many helpful discussions on understanding stochastic calculus on manifolds and reading part of the manuscript. [20]{} Alexei Borodin and Leonid Petrov. (2014). *Integrable probability: From representation theory to Macdonald processes*. Probab. Surveys.Volume 11, 1-58. Benoıt Collins and Sho Matsumoto. (2017). *Weingarten calculus via orthogonality relations: new applications*. ALEA, Lat. Am. J. Probab. Math. Stat. . 14, 631–656. Cépa E. and Lépingle D. (2001). *Brownian particles with electrostatic repulsion on the circle: Dyson’s model for unitary random matrices revisited*. ESAIM Probab. Statist.5, 203–224. Chatterjee. S. (2014). *A short survey of Stein’s method*. arXiv:1404.1392. Chatterjee. S, Diaconis. P, Meckes. E. (2005). *Exchangeable pairs and Poisson approximation*. Probab. Surv.64–106. Chatterjee. S, Fulman. J and Rollin. A. (2011). *Exponential approximation by Stein’s method and spectral graph theory*. ALEA Lat. Am. J. Probab.Math. Stat. 8, 197-223. Chen, L. H. Y. (1975). *Poisson approximation for dependent trials*. Ann. Probab. No. 3, 534–545. Christian Döbler. (2015). *Stein’s method of exchangeable pairs for the Beta distribution and generalizations* . Electron. J. Probab. Volume 20. Christophe Ley. (2020). *Gauss and the identity function – a tale of characterizations of the normal distribution*. arXiv:2003.01827. Coram, M. and Diaconis, P. (2003). *New tests of the correspondence between unitary eigenvalues and the zeros of Riemann’s zeta function*. J. Phys. A. 36 , 2883-2906. D. Bakry, I. Gentil and M. Ledoux. (2014). *Analysis and geometry of Markov diffusion operators*. Grundlehren der Mathematischen Wissenschaften 348, Springer, Cham. Diaconis, P. and Shahshahani, M. (1994). *On the eigenvalues of random matrices. Studies in applied probability*. J. Appl. Probab. 31A, 49-62. Elton. Hsu.(2002). *Stochastic Analysis On Manifolds*. Graduate Studies in Mathematics.Volume 38. Feng-Yu Wang. (2013). *Analysis for Diffusion Processes on Riemannian Manifolds*. Advanced Series on Statistical Science & Applied Probability: Volume 18. Firk, F.W.K.; Miller, S.J. (2009). *Nuclei, Primes and the Random Matrix Connection*. Symmetry. 1, 64-105. Fulman, J. (2009). *Stein’s method and characters of compact Lie groups*. Comm. Math. Phys. 288, 1181-1201. Fulman. J, Adrian Röllin. (2011). *Stein’s method, heat kernel, and linear functions on the orthogonal groups*. arXiv:1109.2975. Fulman, J. (2012). *Stein’s method, heat kernel, and traces of powers of elements of compact Lie groups*. Electron. J. Probab. Volume 17. Fulman.J, Nathan Ross. (2012). *Exponential approximation and Stein’s method of exchangeable pairs* . Latin American journal of probability and mathematical statistics.Volume 10. H. L. Gan. (2015). *Conditional Poisson process approximation*. arXiv:1511.03251. Hughes, C. P. and Rudnick, Z. (2003). *Mock-Gaussian behaviour for linear statistics of classical compact groups*. Random matrix theory. J. Phys. A. 36, 2919-2932. Huiling Le, Alexander Lewis, Karthik Bharath, Christopher Fallaize. (2020). *A diffusion approach to Stein’s method on Riemannian manifolds*. arXiv:2003.11497. Kohei Suzuki. (2019). *Convergence of Brownian motions on metric measure spaces under Riemannian Curvature–Dimension*. Electron. J. Probab. No. 102, 1–36. Kurt Johansson, Gaultier Lambert. (2020). *Multivariate normal approximation for traces of random unitary matrices*. arXiv:2002.01879. László Erdős, Horng-Tzer Yau. (2017). *A Dynamical Approach to Random Matrix Theory*. Courant Lecture Notes. Volume 28. Levy, T.(2008). *Schur-Weyl duality and the heat kernel measure on the unitary group*. Adv. Math. 537-575. Meckes, E. (2006). *An infinitesimal version of Stein’s method of exchangeable pairs*. Doctoral dissertation, Stanford University. Meckes, E. (2007). *Linear functions on the classical matrix groups*. Trans. Amer. Math. Soc. 360, 5355-5366. Meckes, E. (2007). *On the approximate normality of eigenfunctions of the Laplacian*. Trans. Amer. Math. Soc. 361, 5377-5399. Meckes, E. (2009). *On Stein’s method for multivariate normal approximation*. High dimensional probability V: the Luminy volume. P.J. Forrester, T. Nagao. (1998). *Correlations for the Dyson Brownian motion model with Poisson initial conditions*. arXiv:cond-mat/9805330. Songzi Li, Xiang-Dong Li, Yong-Xiao Xie. (2013). *Generalized Dyson Brownian motion, McKean-Vlasov equation and eigenvalues of random matrices*. arXiv:1303.1240. Stein, Charles. (1972). *A bound for the error in the normal approximation to the distribution of a sum of dependent random variables*. Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability, Volume 2: Probability Theory, 583–602. Tiefeng Jiang and Sho Matsumoto. (2015). *Moments of traces of circular beta-ensembles*. Ann. Probab. Volume 43, Number 6. Vadim Kostrykin, Jürgen Potthoff and Robert Schrader. (2012). *Brownian motions on metric graphs*. Journal of Mathematical Physics 53, 095206. Wei Huang, Weitao Du, Richard Yi Da Xu. (2020). *On the Neural Tangent Kernel of Deep Networks with Orthogonal Initialization*. arXiv:2004.05867. Xiang-Dong Li. (2012). *Perelman’s entropy formula for the Witten Laplacian on Riemannian manifolds via Bakry–Emery Ricci curvature*. Math. Ann.353:403–437. [^1]: [ [email protected]]{}.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Diffractive lenses have recently been applied to the domain of multispectral imaging in the X-ray and UV regimes where they can achieve very high resolution as compared to reflective and refractive optics. Conventionally, spectral components are reconstructed by taking measurements at the focal planes. However, the reconstruction quality can be improved by optimizing the measurement configuration. In this work, we adapt a sequential backward selection algorithm to search for a configuration which minimizes expected reconstruction error. By approximating the forward system as a circular convolution and making assumptions on the source and noise, we greatly reduce the complexity of the algorithm. Numerical results show that the configuration found by the algorithm significantly improves the reconstruction performance compared to a standard configuration.' address: | Department of Electrical and Computer Engineering and Coordinated Science Laboratory,\ University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA bibliography: - 'bibliography.bib' title: | Optimal Measurement Configuration in Computational\ Diffractive Imaging --- Spectral imaging, diffractive optics, measurement configuration, subset selection, computational imaging Introduction {#sec:intro} ============ Spectral imaging is the formation of images at different wavelengths in the electromagnetic spectrum. With images usually taken in the visible, X-ray, ultraviolet (UV), or infrared bands, it has applications in medicine, geographic surveying, astronomy, and solar physics [@shaw2003spectral], [@garini2006spectral]. In spectral imaging, a polychromatic source must be first separated into its spectral components before being captured. There are a number of ways to achieve this, but a common method is to use a set of configurable optical filters. For example, the spectral imager on the Solar Dynamics Observatory uses a rotating drum of optical filters to selectively pass light of specific wavelengths of interest [@sdo]. A new approach is to use a diffractive lens to perform spectral imaging [@oktem2014icip]. Diffractive lenses are often preferred in the UV or X-ray regimes because manufacturing tolerances at these wavelengths can be more relaxed than reflective optics and still obtain a similar resolution [@davila2011high]. Since diffractive optics can be manufactured using a photolithographic process, they can be produced at a higher precision compared to the grinding process used to produce conventional reflective optics. Moreover, refractive optics are unsuitable for UV or X-ray imaging because glass is opaque at these wavelengths. Figures \[fig:diff\_lens\](b) and \[fig:diff\_lens\](c) are two examples of a pattern that can be etched into silicon wafer to produce a diffractive lens. Diffractive lenses have the property that the angle at which light exits the lens is determined by the light’s wavelength, which gives them a wavelength dependent focal length, as shown in Figure \[fig:diff\_lens\](a) [@attwood1999]. ![(a) diffraction of a polychromatic wave through a diffractive lens (b) Fresnel zone plate (c) photon sieve [@kipp2001sharper][]{data-label="fig:diff_lens"}](diffraction_ps_rgb){width="4.0cm"} (a) ![(a) diffraction of a polychromatic wave through a diffractive lens (b) Fresnel zone plate (c) photon sieve [@kipp2001sharper][]{data-label="fig:diff_lens"}](zoneplate.png){width="2.0cm"} (b) ![(a) diffraction of a polychromatic wave through a diffractive lens (b) Fresnel zone plate (c) photon sieve [@kipp2001sharper][]{data-label="fig:diff_lens"}](photonsieve){width="2.0cm"} (c) Measurements at the focal plane of each spectral component comprise of a sum of a focused image of one component and blurred images of all other components, as shown in Figure \[fig:pssi\_drawing\]. An inverse problem consisting of disentangling and deblurring of measurements must be solved in order to recover the original source components [@oktem2014icip]. However, this focal plane measurement configuration leads to suboptimal reconstructions, especially when spectral components are close in wavelength. Therefore, it is desired to determine the optimal measurement configuration before acquiring the data. ![Imaging a scene with emissions at wavelengths $\lambda_1$ and $\lambda_2$. Measurements $y_1$ and $y_2$ are taken at two positions where one wavelength is in focus and the other is out of focus.[]{data-label="fig:pssi_drawing"}](diffraction_system){width="8.5cm"} Finding the optimal measurement configuration can be seen as a *sensor placement problem*, which lies under the broader class of problems known as *subset selection*. Subset selection applies to many domains, such as array optimization for atmospheric imaging [@sharif], [@wang2017near], magnetic resonance imaging [@gao2000optimal], and detection problems [@yu1997sampling]. Methods like genetic algorithms, convex optimization [@joshi2008sensor], and hill climbing [@reeves1999sequential] with many selection criteria have been developed to solve such problems. However, most of these methods solve the problem of single-sensor/single-measurement systems where the placement of one sensor contributes a single row to the observation matrix. In contrast, many imaging systems are single-sensor/multiple-measurement (like our problem), where each sensor placed (measurement plane) contributes multiple rows to the observation matrix (one row per detector pixel). Single-sensor/single-measurement algorithms have been extended to the multiple measurement case, known as *clustering* algorithms. Examples include *clustered sequential backward selection* (CSBS) [@sharif], *clustered FrameSense* (CFS) [@ranieri2014near], *clustered maximum projection on minimum eigenspace* (CMPME) [@wang2018sampling]. In this paper we adapt CSBS to the diffractive imaging problem to automatically determine a measurement configuration from a set of candidate plane locations, which minimizes expected reconstruction error. Furthermore, we exploit structures in the imaging model to make the algorithm computationally feasible for large images. Forward Model and Statistical Formulation {#sec:format} ========================================= In this section, we mathematically model a diffractive imaging system and describe the process of recovering the spectral components. Consider a polychromatic source that has $S$ spectral components $\bm{x}_1, \dots, \bm{x}_S \in \mathbb R^{N_1\times N_1}$. Using a moving detector, we make $M$ measurements $\bm{y}_1, \dots, \bm{y}_M \in \mathbb R^{N_2\times N_2}$ at distances $d_1, \dots, d_M$ from the lens. We allow for repeated measurements at the same plane for a more flexible model that can take into account non equal exposure times. Due to linearity, each measurement is a superposition of blurred versions of the $S$ sources. More formally, $$\bm{y}_m = \sum_{s=1}^S \bm{a}_{m,s} \ast \bm{x}_s + \bm{n}_m \label{eq:fwd_model}$$ where $\bm{a}_{m,s} \in \mathbb R^{P\times P}$ is a blurring kernel known as a *point spread function* (PSF), $\ast$ is a 2D convolution, and $\bm{n}_m \in \mathbb R^{N_2 \times N_2}$ is additive measurement noise. Each PSF depends on the associated source wavelength and measurement location together with the diffractive lens parameters and can be computed efficiently [@ayazgok2020efficient]. Since convolution is a linear operation, we can rewrite the above equation as a linear system $$\underbrace{ \begin{bmatrix}\bm{y}_1 \\ \vdots \\ \bm{y}_M\end{bmatrix} }_{\bm{y}} = \underbrace{ \begin{bmatrix} \bm{A}_{1, 1} & \hdots & \bm{A}_{1, S} \\ \vdots & & \vdots \\ \bm{A}_{M, 1} & \hdots & \bm{A}_{M, S} \end{bmatrix} }_{\bm{A}_{\bm{d}}} \underbrace{ \begin{bmatrix}\bm{x}_1 \\ \vdots \\ \bm{x}_S\end{bmatrix} }_{\bm{x}} + \underbrace{ \begin{bmatrix}\bm{n}_1 \\ \vdots \\ \bm{n}_M\end{bmatrix} }_{\bm{n}} \label{eq:fourier_mtx}$$ where $\bm{y}_m \in \mathbb R^{N_2^2\times 1}$, $\bm{x}_s \in \mathbb R^{N_1^2\times 1}$ and $\bm{n}_m \in \mathbb R^{N_2^2\times 1}$ have been flattened from their original 2D shape, and each $\bm{A}_{m,s} \in \mathbb R^{N_2^2\times N_1^2}$ is a block-toeplitz matrix with toeplitz blocks formed from 2D convolution with PSF $\bm{a}_{m,s}$. We will refer to the matrix containing all $\bm{A}_{m,s}$ generated by measurements taken at $\bm{d} = \{d_1, \dots, d_M\}$ as $\bm{A}_{\bm{d}}$. The problem of where to take measurements $\bm{y}_1, \dots, \bm{y}_M$ has not been addressed and affects the reconstruction quality. In order to compare the impact of different measurement configurations on the reconstruction, it is necessary to define some cost for the measurement matrix $\bm{A}_{\bm{d}}$. A common cost metric is the expected reconstruction error, or expected *sum of squared errors* (SSE). However, we must have some strategy for the recovery of $\bm{x}$ to get reconstruction error and we must make statistical assumptions about $\bm{x}$. *Maximum a posteriori* (MAP) estimation is one such strategy. We assume the original spectral components and noise are distributed according to a normal distribution such that $\bm{x} \sim \mathcal{N}(\bm{x}_{0}, \bm{\Sigma}_{\bm{x}})$ and $\bm{n} \sim \mathcal{N}(0, \bm{\Sigma}_{\bm{n}})$. The MAP estimate is then $$\begin{aligned} \bm{x}_{MAP} &= \arg \max_{\bm{x} \in \mathbb{C}^n} p(\bm{x} | \bm{y}) = \arg \max_{\bm{x}} p(\bm{y}|\bm{x}) p(\bm{x})\\ &= \arg \min_{\bm{x}} \left[ - \log(p(\bm{y}|\bm{x})) - \log p(\bm{x})\right] \\ &= \bm{x}_0 + \left( \bm{A}_{\bm{d}}^H\bm{\Sigma}_{\bm{n}}^{-1} \bm{A}_{\bm{d}} + \bm{\Sigma}_{\bm{x}}^{-1}\right)^{-1} \cdot \bm{A}_{\bm{d}}^H \bm{\Sigma}_{\bm{n}}^{-1} (\bm{y} - \bm{A}_{\bm{d}} \bm{x}_0) \end{aligned}$$ The reconstruction error is defined as $\bm{e} = \bm{x} - \bm{x_{\text{MAP}}}$, and the expected sum of squared error cost is $E[{\left\lVert\bm{e}\right\rVert}_2^2]$. This expression can be rewritten in terms of the error covariance: $$\begin{aligned} \mathbb{E}[{\left\lVert\bm{e}\right\rVert}_2^2] = \mathbb{E}[\bm{e}^H\bm{e}] & = \mathbb{E}[\text{tr}(\bm{e}^H\bm{e})] = \mathbb{E}[\text{tr}(\bm{e}\bm{e}^H)] \\ & = \text{tr}(\mathbb{E}[\bm{e}\bm{e}^H]) = \text{tr}(\bm{\Sigma}_{\bm{e}})\end{aligned}$$ where the error covariance matrix is defined as $\bm{\Sigma}_{\bm{e}} = E[\bm{e}\bm{e}^H]$ and has the closed form expression: $$\bm{\Sigma}_{\bm{e}} = \left( \bm{A}_{\bm{d}}^H\bm{\Sigma}_{\bm{n}}^{-1} \bm{A}_{\bm{d}} + \bm{\Sigma}_{\bm{x}}^{-1}\right)^{-1}$$ Combining the above equations, we can now write a cost metric which lets us evalute the expected reconstruction error for a particular measurement configuration $\bm{d}$: $$\text{Cost}(\bm{d}) = \mathbb{E}[{\left\lVert\bm{e}\right\rVert}_2^2] = \text{tr}\left(\left( \bm{A}_{\bm{d}}^H\bm{\Sigma}_{\bm{n}}^{-1} \bm{A}_{\bm{d}} + \bm{\Sigma}_{\bm{x}}^{-1}\right)^{-1}\right)$$ Measurement Selection Algorithm =============================== With a method of evaluating the effect a particular configuration $\bm{d}$ has on reconstruction error, we can begin considering which configurations are best suited for minimizing error. For example, if we are provided with a set of $C$ candidate measurement locations, we may wish to find a subset of size $M$ which minimizes reconstruction error. This is known as a *subset selection* problem. One might think to simply search over all possible measurement configurations of size $M$, but this exhaustive search requires $\binom{C}{M}$ evalutions of cost, growing on the order of $O(C^M)$. CSBS is an alternative method which is more computationally feasible, where one measurement location is eliminated from $\bm{d}$ in each iteration until only $M$ locations remain. As reconstruction error generally increases as the number of measurements decreases, CSBS selects for elimination the measurement that incurs the smallest increase in cost in each iteration. $\bm{d} = \{d_1, \dots, d_C\}$ $d' = \arg \min_{d \in \bm{d}} \text{Cost}(\bm{d} \backslash \{d\}))$ $\bm{d} = \bm{d} \backslash \{d'\}$ Unlike an exhaustive search, the complexity of CSBS is not combinatorial. As the size of $\bm{d}$ shrinks with each iteration, the number of cost evaluations for each minimization step decreases. The total number of cost evaluations is $$\begin{aligned} \sum_{|\bm{d}| = M}^C |\bm{d}| &= O(C^2 - M^2) \end{aligned}$$ Fast Implementation =================== While the SSE cost applies to any general linear system, we can augment the complexity reduction achieved by CSBS by making assumptions about the structure of $\bm{A}_{\bm{d}}$, $\bm{\Sigma}_{\bm{n}}$ and $\bm{\Sigma}_{\bm{x}}$. Specifically, if we assume the blocks of these matrices are block-circulant with circulant blocks (BCCB), then they can be diagonalized by the 2D DFT matrix where operations involving multiplications and inversions are much faster. For $\bm{A}_{\bm{d}}$ this means that each block $\bm A_{m,s}$ is BCCB and corresponds to circular convolution with the kernel $\bm a_{m,s}$. For $\bm \Sigma_n$, we assume independent noise among image pixels and measurement planes with variance $(1/\lambda)$, so $\bm \Sigma_n=(1/\lambda) \bm I$ where $\bm I$ represents the identity matrix. For $\bm \Sigma_x$, each of its blocks being BCCB means that the covariance among image pixels are represented by 2D convolution kernels. Assuming $N \times N$ images, each $N^2 \times N^2$ block $\bm A_{m,s}$ of $\bm {A_d}$ can be decomposed as $\bm A_{m,s} = \bm F^{-1} \widetilde{\bm A}_{m,s} \bm F$ where $\widetilde{\bm A}_{m,s}$ is the diagonal matrix consisting of the 2D DFT of $\bm a_{m,s}$, and $\bm F$ is the 2D DFT matrix. This yields $$\bm {A_d}= \underbrace{ \begin{bmatrix} \bm F^{-1} \hspace{-0.2 in}& \hspace{-0.2 in} & \hspace{-0.20 in} \vspace{-0.1 in}\text{\Large 0} \\ \vspace{-0.05 in}&\hspace{-0.20 in} \ddots & \\ \text{\Large 0} \hspace{-0.17 in}& \hspace{-0.2 in} & \hspace{-0.15 in} \bm F^{-1} \end{bmatrix} }_{\widetilde{\bm F}^{-1}} \underbrace{ \begin{bmatrix} \widetilde{\bm A}_{1,1} \hspace{-0.1 in}&\hspace{-0.1 in} \hdots & \hspace{-0.1 in} \widetilde{\bm A}_{1,S} \vspace{-0.02 in}\\ \vspace{-0.02 in}\vdots &\hspace{-0.1 in} \ddots& \vdots \\ \widetilde{\bm A}_{M,1} \hspace{-0.1 in}&\hspace{-0.1 in} \hdots & \hspace{-0.1 in} \widetilde{\bm A}_{M,S} \\ \end{bmatrix} }_{\widetilde{\bm A}_{\bm d}} \underbrace{ \begin{bmatrix} \bm F \hspace{-0.1 in}& & \hspace{-0.18 in} \vspace{-0.1 in}\text{\Large 0} \\ \vspace{-0.05 in}& \hspace{-0.12 in}\ddots & \\ \hspace{0.05 in} \text{\Large 0} & & \hspace{-0.12 in} \bm F \end{bmatrix} }_{\widetilde{\bm F}} \label{eq:fourier_mtx}$$ so, we have $\bm{A_d} = \widetilde{\bm F}^{-1} \widetilde{\bm A}_{\bm d} \widetilde{\bm F}$, from which we get $\bm{A_d}^H \bm{A_d} = \widetilde{\bm F}^{-1} \widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d} \widetilde{\bm F}$. Applying the same procedure, we get $\bm \Sigma_x^{-1} = \widetilde{\bm F}^{-1} \widetilde{\bm \Sigma}_x^{-1} \widetilde{\bm F}$. The SSE cost for a measurement configuration $\bm d$ then becomes (scaling both terms with $\lambda$): $$\begin{aligned} \text{Cost}(\bm{d}) & = \text{tr}\left(\left( {\bm A}_{\bm d}^H {\bm A}_{\bm d} + \lambda \bm \Sigma_x^{-1} \right)^{-1} \right) \label{eq:naive_cost}\\ & = \text{tr}\left(\left(\widetilde{\bm F}^{-1} \left( \widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d} + \lambda \widetilde{\bm \Sigma}_x^{-1} \right) \widetilde{\bm F}\right)^{-1}\right) \nonumber \\ & = \text{tr}\left(\widetilde{\bm F}^{-1} \left( \widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d} + \lambda \widetilde{\bm \Sigma}_x^{-1} \right)^{-1} \widetilde{\bm F}\right) \nonumber \\ & = \text{tr}\left(\left( \widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d} + \lambda \widetilde{\bm \Sigma}_x^{-1} \right)^{-1} \right) \label{eq:fast_cost}\end{aligned}$$ where the computational complexity of evaluating is much less than due to the diagonalized blocks of $\widetilde{\bm A}_{\bm d}$ and $\widetilde{\bm \Sigma}_x^{-1}$. There are two contributors to the complexity of evaluating the cost at $\text{Cost}(\bm{d})$ for a particular configuration: The multiplication $\widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d}$, and the inversion $\left( \widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d} + \lambda \widetilde{\bm \Sigma}_x^{-1} \right)^{-1}$. In fact, the product $\widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d}$ only needs to be calculated once during the algorithm initialization, and it can be efficiently updated at each iteration by adding/subracting the contribution of the candidate plane that is iterated over, which can be precomputed. Thus, the complexity of overall CSBS algorithm is dominated by the inversion of $\left( \widetilde{\bm A}_{\bm d}^H \widetilde{\bm A}_{\bm d} + \lambda \widetilde{\bm \Sigma}_x^{-1} \right) \in \mathbb C ^{SN^2 \times SN^2}$ that is performed in each iteration. While the complexity of a standard inversion algorithm is $O((SN^2)^3)$, the diagonal structure of this matrix allows for a much faster inversion algorithm with complexity $O(S^3N^2)$, a speed-up of $N^4$ which is significant for large images. The total CSBS algorithm complexity is $$O_{sbs} = \sum_{|\bm{d}| = M}^C |\bm{d}| O(S^3N^2) = O(S^3N^2C^2)$$ Numerical Experiments ===================== In this section, we present numerical experiments that demonstrate that the measurement configuration selected by CSBS yields improved reconstructions over reconstructions obtained from measurements taken at focal planes. We use a photon sieve as the diffractive element in our simulations, which offers PSFs with sharper focus than Fresnel zone plates [@kipp2001sharper]. We begin by simulating a scenario with two spectral components that are close to each other in wavelength, shown as separate colors in Figure \[fig:results\](a). We use the MAP estimation framework given in Section 2 as the image reconstruction algorithm for both the focal plane and CSBS configurations. For a fair comparison between CSBS and focal plane reconstructions, we search over $\lambda$ to find the value which maximizes the focal plane reconstruction *structural similarity* (SSIM) [@ssim], then use this same $\lambda$ for the CSBS cost function and reconstruction. The final measurement configuration selected by CSBS is given in Figure \[fig:results\](d), where the two focal planes are marked with red and green bars. Figures \[fig:results\](b) and \[fig:results\](c) show the spectral component reconstructions for the focal plane configuration, and Figures \[fig:results\](e) and \[fig:results\](f) for CSBS configuration. The reconstruction SSIMs for the CSBS and focal plane reconstructions are 0.459 and 0.347, respectively. Our intuition on why CSBS chooses out of focus planes pertains to measurement variation of the PSF pairs for each candidate plane. The spectral components are very close together in wavelength, so the PSFs corresponding to in focus and out of focus components at the focal planes are very similar. This leads to poor measurement variation and makes disentangling the component contributions difficult. This is especially evident in Figure 3(c), where the features from one wavelength appear in the reconstruction of the other wavelength. Instead, CSBS chooses measurement locations where the PSF pairs have more variation at the expense of a less sharp in focus PSF, shown in Figure \[fig:psfs\_compare\]. ![(a) polychromatic source image with spectral components $\lambda_1$ and $\lambda_2$ (b) reconstruction of $\lambda_1$ from focal configuration (c) reconstruction of $\lambda_2$ from focal configuration (d) measurement locations selected by CSBS (e) reconstruction of $\lambda_1$ from CSBS configuration (f) reconstruction of $\lambda_2$ from CSBS configuration[]{data-label="fig:results"}](sources){width="2.5cm"} (a) ![(a) polychromatic source image with spectral components $\lambda_1$ and $\lambda_2$ (b) reconstruction of $\lambda_1$ from focal configuration (c) reconstruction of $\lambda_2$ from focal configuration (d) measurement locations selected by CSBS (e) reconstruction of $\lambda_1$ from CSBS configuration (f) reconstruction of $\lambda_2$ from CSBS configuration[]{data-label="fig:results"}](recon_focus1){width="2.5cm"} (b) ![(a) polychromatic source image with spectral components $\lambda_1$ and $\lambda_2$ (b) reconstruction of $\lambda_1$ from focal configuration (c) reconstruction of $\lambda_2$ from focal configuration (d) measurement locations selected by CSBS (e) reconstruction of $\lambda_1$ from CSBS configuration (f) reconstruction of $\lambda_2$ from CSBS configuration[]{data-label="fig:results"}](recon_focus2){width="2.5cm"} (c) ![(a) polychromatic source image with spectral components $\lambda_1$ and $\lambda_2$ (b) reconstruction of $\lambda_1$ from focal configuration (c) reconstruction of $\lambda_2$ from focal configuration (d) measurement locations selected by CSBS (e) reconstruction of $\lambda_1$ from CSBS configuration (f) reconstruction of $\lambda_2$ from CSBS configuration[]{data-label="fig:results"}](csbs_copies){width="3.6cm"} (d) ![(a) polychromatic source image with spectral components $\lambda_1$ and $\lambda_2$ (b) reconstruction of $\lambda_1$ from focal configuration (c) reconstruction of $\lambda_2$ from focal configuration (d) measurement locations selected by CSBS (e) reconstruction of $\lambda_1$ from CSBS configuration (f) reconstruction of $\lambda_2$ from CSBS configuration[]{data-label="fig:results"}](recon_csbs1){width="2.5cm"} (e) ![(a) polychromatic source image with spectral components $\lambda_1$ and $\lambda_2$ (b) reconstruction of $\lambda_1$ from focal configuration (c) reconstruction of $\lambda_2$ from focal configuration (d) measurement locations selected by CSBS (e) reconstruction of $\lambda_1$ from CSBS configuration (f) reconstruction of $\lambda_2$ from CSBS configuration[]{data-label="fig:results"}](recon_csbs2){width="2.5cm"} (f) ![(a) (b) PSFs at $\lambda_1$ focal plane. (c) (d) PSFs at a measurement location selected by CSBS (c) and (d) are less focused that (a) and (b), but have more measurement variation between them. []{data-label="fig:psfs_compare"}](psf_focus1){width="2cm"} (a) ![(a) (b) PSFs at $\lambda_1$ focal plane. (c) (d) PSFs at a measurement location selected by CSBS (c) and (d) are less focused that (a) and (b), but have more measurement variation between them. []{data-label="fig:psfs_compare"}](psf_focus2){width="2cm"} (b) ![(a) (b) PSFs at $\lambda_1$ focal plane. (c) (d) PSFs at a measurement location selected by CSBS (c) and (d) are less focused that (a) and (b), but have more measurement variation between them. []{data-label="fig:psfs_compare"}](psf_csbs1){width="2cm"} (c) ![(a) (b) PSFs at $\lambda_1$ focal plane. (c) (d) PSFs at a measurement location selected by CSBS (c) and (d) are less focused that (a) and (b), but have more measurement variation between them. []{data-label="fig:psfs_compare"}](psf_csbs2){width="2cm"} (d) ![Reconstruction SSIMs for varied number of spectral components, SNR (dB), and source separation (DOF). CSBS reconstruction SSIM and focal plane reconstruction SSIM are shown in orange and blue, respectively.[]{data-label="fig:ratio"}](ssims){width="8.5cm"} To show that this reconstruction improvement generalizes, we repeat the first experiment for $S = 2, 3, 4$ uniformly spaced spectral components under different noise levels and spectral component separations measured in depth of focus (DOF) [@davila2011high]. In Figure \[fig:ratio\], we plot the mean SSIM of the reconstructions obtained from measurements at focal planes (blue) and measurements at planes selected by CSBS (orange). The CSBS reconstructions generally have higher SSIM than the focal plane up until the spectral components are sufficiently separated (about 10 DOF), where reconstruction SSIM are about the same. Conclusion ========== We apply a variant of the sequential backward selection algorithm to the problem of acquisition in a diffractive spectral imaging system. The high dimensionality of large images makes a direct application of CSBS and SSE cost computationally intractable, so we have developed a more feasible implementation of this algorithm and perform an analysis of its complexity to show that it is significantly faster than the previous implementation for large images. Finally, we demonstrate CSBS on a simulated spectral imaging system and show that the optimized measurement configuration achieves equal or better reconstructions than a choice of measurements at the spectral component focal planes.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We study Maurer-Cartan moduli spaces of dg algebras and associated dg categories and show that, while not quasi-isomorphism invariants, they are invariants of strong homotopy type, a natural notion that has not been studied before. We prove, in several different contexts, Schlessinger-Stasheff type theorems comparing the notions of homotopy and gauge equivalence for Maurer-Cartan elements as well as their categorified versions. As an application, we re-prove and generalize Block-Smith’s higher Riemann-Hilbert correspondence, and develop its analogue for simplicial complexes and topological spaces.' author: - 'Joe Chuang, Julian Holstein, Andrey Lazarev' bibliography: - './biblibrary2.bib' title: 'Maurer-Cartan moduli and theorems of Riemann-Hilbert type' --- \[section\] \[thm\][Corollary]{} \[thm\][Lemma]{} \[thm\][Proposition]{} \[section\] \[section\] \[section\] [^1] Introduction ============ The simplest version of the Riemann-Hilbert correspondence is the statement, known for many decades, that the category of flat vector bundles on a smooth manifold $M$ is equivalent to the category of representations of its fundamental group $\pi_1(M)$. Recently a higher generalization of this statement was developed, [@Block09], where the category of representations of $\pi_1(M)$ was replaced by a differential graded category of infinity local systems on $M$ and the category of flat vector bundles by a differential graded (dg) category of certain modules, called *cohesive modules*, over $\Omega(M)$, the de Rham algebra of $M$. The correspondence is given by a certain $A_\infty$ functor. The proof in loc.cit. is technically complicated and our original motivation was to understand it in simple terms, particularly keeping in mind that one side of the equivalence – the category of infinity local systems – is essentially the same as the more classical notion of a cohomologically locally constant (clc) dg sheaf, i.e.  a dg sheaf whose cohomology forms an ordinary (graded) locally constant sheaf. An obvious approach to proving the desired result is based on the observation that $\Omega(M)$ is the global sections of the sheaf of de Rham algebras on $M$ and the latter is a soft resolution of the constant sheaf $\mathbb R$. Similarly, a dg module $N$ over $\Omega(M)$ could be sheafified and viewed as a module over the sheaf of de Rham algebras. Imposing suitable restrictions on $M$, one could hope that the resulting sheaf of modules would be quasi-isomorphic to a clc sheaf and that this procedure establishes an equivalence between the derived category of clc complexes of sheaves on $M$ and a suitable homotopy subcategory of dg $\Omega(M)$-modules (such as cohesive $\Omega(M)$-modules). Taking into account that the category of clc sheaves makes sense for spaces more general than manifolds, e.g. simplicial complexes, one could further ask whether this programme can be carried out in this more general context. Next, one could try to achieve a similar result working with the singular cochain complex of a topological space or a simplicial set, with values in rings other than $\mathbb R$, e.g. $\mathbb Z$. Finally, one should study the functorial properties of this construction, in particular its liftability to the suitable homotopy category of spaces that are being considered (manifolds, simplicial complexes, topological spaces or simplicial sets). Somewhat surprisingly, this naive approach does work and eventually produces pretty much all the results one would hope to obtain. The main difficulty in implementing the strategy outlined above is proving, in different contexts, that the associated dg sheaf of a dg $\Omega(M)$-module $N$ is a clc sheaf. To show this, one needs to work with the notion of a Maurer-Cartan (MC) element and the MC moduli set and establish various versions of the classical Schlessinger-Stasheff theorem [@Schlessinger12]. The latter states that, under appropriate conditions, homotopy equivalent MC elements must be gauge equivalent, and vice-versa. This result is usually formulated in the context of dg (pro)nilpotent Lie algebras but we need it for dg associative algebras. Schlessinger-Stasheff type results are established in this paper in two different contexts: analytical (for dg algebras such as the smooth de Rham algebra of a manifold) and algebraic (for dg algebras without any topology or with a pseudo-compact topology such as the singular cochain algebra of a topological space). The algebraic version of the Schlessinger-Stasheff theorem is particularly interesting and has ramifications far beyond higher Riemann-Hilbert correspondence; some of them have been explored in the present paper but others await further study. We associate to any dg algebra $A$ several dg categories, of which the most important is the category of twisted $A$-modules $\operatorname{Tw}(A)$. A version of this category (in the context where $A$ itself is a dg category) was first introduced in the seminal paper [@Bonda90] where it was called the category of (two-sided) twisted complexes and denoted by $\operatorname{Pre-Tr(A)}$ (in fact, $\operatorname{Tw}(A)$ is obtained from $\operatorname{Pre-Tr(A)}$ by adding infinite direct sums of objects). The homotopy category $\operatorname{H^{0}}(\operatorname{Tw}(A))$ is superficially similar to $D(A)$, the derived category of $A$, but is a finer invariant; in particular it is not, generally, a quasi-isomorphism invariant of $A$, unlike $D(A)$ (as pointed out by Drinfeld [@Drinfeld04 Remark 2.6]). It turns out that the correct notion to use in this context is that of *strong homotopy equivalence* of dg algebras. This is a chain homotopy equivalence that takes into account the multiplicative structure and it was not studied before, as far as we know. We show that two strongly homotopy equivalent dg algebras have quasi-equivalent categories of twisted modules. Furthermore, the notion of strong homotopy and strong homotopy equivalence exists also for dg coalgebras (equivalently, pseudo-compact dg algebras), such as the normalized chain complex of a simplicial set, and we show that two weakly equivalent Kan simplicial sets give rise to strongly homotopy equivalent dg coalgebras. This is an important ingredient in the proof of the singular version of the higher Riemann-Hilbert correspondence, but it also has philosophical significance as it shows that the singular chain coalgebra on a simplicial set that is not Kan (or fibrant) might have the wrong homotopy type. This phenomenon also showed up in the recent paper [@Zeinalia16] where a generalization of Adams’ cobar-construction to the non-simply connected case was established. Denoting by $C^*(X)$ the normalized cochain algebra of a Kan simplicial set $X$, we show that the homotopy category of twisted $C^*(X)$-modules is equivalent to the derived category of clc complexes of sheaves on $|X|$, the geometric realization of $X$. If $X$ is not Kan, the category $\operatorname{Tw}(C^*(X))$ has no homotopy invariant meaning, but one could speculate that it is related to the category of sheaves on $|X|$ that are constructible with respect to some stratification. A related idea is contained in Kontsevich’s preprint [@Kontsevich09 pp. 3-4]. The paper is organized as follows. Section \[basic\] introduces the notion of an MC element in a dg algebra as well as concomitant concepts: gauge equivalence, MC twisting and a new notion of *homotopy gauge equivalence* that is, as the name suggests, a relaxation of familiar gauge equivalence to an up to homotopy notion. Section \[twisted\] introduces twisted modules, and gives a comparison with Block’s cohesive modules [@Block10]. In Section \[smooth\] we study smooth homotopies of topological algebras and their MC elements, and prove an appropriate analogue of the Schlessinger-Stasheff theorem, its categorified version and show that homotopic maps of manifolds give rise to isomorphic functors between the corresponding categories of twisted modules over their de Rham algebras. In Section \[strong\] we introduce the notions of a strong homotopy of dg algebra morphisms and of a strong homotopy equivalence. A comparison is given with various weaker notions, of which the notion of *derivation homotopy* has been previously known, particularly in the context of rational homotopy theory. We obtain a suitable version of the Schlessinger-Stasheff theorem that implies that strongly homotopy equivalent dg algebras have quasi-equivalent dg categories of twisted modules and obtain a similar result for pseudo-compact dg algebras. In Section \[simplicial\] we apply our results to normalized cochain algebras of simplicial sets and show that weakly equivalent Kan simplicial sets give rise to quasi-equivalent categories of twisted modules. In Section \[sheaves\] we consider dg sheaves on a locally ringed space and, using our Schlessinger-Stasheff theorems, show that, under suitable assumptions, the homotopy category of perfect (i.e. finitely generated up to homotopy retractions) twisted modules over the dg algebra of global sections is equivalent to the derived category of perfect dg sheaves. This is applied in Section \[applications\] to produce versions of the higher Riemann-Hilbert correspondence for smooth, possibly non-compact, manifolds and finite-dimensional simplicial complexes, thus generalizing the results of [@Block09]. We also consider the case of the Dolbeault algebra and coherent sheaves on a complex manifold, slightly strengthening the result of [@Block10]. Finally, we treat the most interesting case, that of the singular cochain algebra on a topological space and the corresponding higher Riemann-Hilbert correspondence. The latter is obtained under very general assumptions, i.e. we consider any locally contractible topological space and its dg category of possibly infinitely generated and unbounded clc sheaves over any ring of finite homological dimension. The paper contains an appendix where relevant facts from the theory of nuclear spaces are collected. Notation and conventions {#sect-notations} ------------------------ We work in the category of $\mathbb Z$-graded dg modules over a fixed commutative ring ${k}$; an object in this category is a pair $(V,d_V)$ where $V$ is a graded ${k}$-module and $d_V$ is a differential on it; it will always be assumed to be of cohomological type (so it raises the degree of a homogeneous element). Unmarked tensor products and Homs will be understood to be taken over ${k}$. The *shift* of a graded ${k}$-module $V$ is the graded ${k}$-module $V[1]$ with $V[1]^i=V^{i+1}$. A *pseudo-compact* graded ${k}$-module is, by definition, an inverse limit of finitely generated free ${k}$-modules. Pseudo-compact ${k}$-modules are endowed with the topology of the inverse limit and form a category where maps are required to be continuous. The category of pseudo-compact ${k}$-modules is anti-equivalent to that of (discrete) free ${k}$-modules via ${k}$-linear duality. The category of pseudo-compact ${k}$-modules is monoidal: if $V=\varprojlim V_{\alpha}$ and $U=\varprojlim U_{\beta}$ are two pseudo-compact ${k}$-modules represented as inverse limits of finitely generated free ${k}$-modules, then $V\hat{\otimes}U:=\varprojlim_{\alpha,\beta}(V_{\alpha}\otimes U_\beta)$. Later on, the hat will always be omitted (but understood) for the tensor product of two pseudo-compact ${k}$-modules. We will also need to form the tensor product of a pseudo-compact ${k}$-module $V=\varprojlim V_{\alpha}$ and a discrete ${k}$-module $U$; such a tensor product will be defined as $V\hat{\otimes} U:=\varprojlim_\alpha(V_{\alpha}\otimes U)$ and, as before, the hat will be omitted but understood. Note that the tensor product of a pseudo-compact and discrete ${k}$-modules has a topology but is not, in general, pseudo-compact. A dg algebra is an associative monoid in the dg category of dg ${k}$-modules and in the examples we consider its underlying ${k}$-module is free. A (right) dg module over a dg algebra $A$ is a dg ${k}$-module $V$ together with a map $V\otimes A{\rightarrow}V$ of dg ${k}$-modules satisfying the usual conditions of associativity and unitality. Similarly a pseudo-compact dg algebra is a monoid in the monoidal category of pseudo-compact ${k}$-modules. Via continuous linear duality a pseudo-compact dg algebra becomes a dg coalgebra, and the two notions are therefore equivalent. We, however, will work consistently with pseudo-compact algebras rather than coalgebras. An important example of a pseudo-compact dg algebra over $\mathbb Z$ is the singular integer-valued cochain complex $C^*(X,\mathbb Z)$ of a topological space $X$ (or, more pertinently, its normalized version); it is pseudo-compact as dual to the dg coalgebra $C_*(X,\mathbb Z)$ of singular chains on $X$. We will consider dg *contramodules* over dg pseudo-compact algebras, cf. [@Positselski11; @Positselski15]; a (right) contramodule over a pseudo-compact algebra $A$ is a *discrete* ${k}$-module $V$ supplied with a ‘contra-action’ map $V\otimes A{\rightarrow}V$ satisfying the usual conditions of associativity and unitality. We reiterate that $V\otimes A$ is a completed tensor product so a contramodule is not merely an $A$-module where the topology on $A$ is disregarded. Prominent among contramodules are those of the form $V\otimes A$ with the $A$-(contra)action given by the right multiplication. These contramodules are free in the sense that if $U$ is another $A$-contramodule, then $\operatorname{Hom}_A(V\otimes A,U)\cong \operatorname{Hom}_{{k}}(V,U)$ just as it is in the case of usual free $A$-modules. Even though contramodules appear somewhat exotic objects, they are in fact encountered frequently. For example, if $X$ is a topological space and $V$ is a (possibly infinitely generated) free abelian group then $C^*(X,V)$, the singular cochain complex of $X$ with coefficients in $V$ is a free $C^*(X, \mathbb Z)$-contramodule. If $M$ is a dg object (such as a dg module, dg algebra etc), we will write $M^{\#}$ for its underlying graded object (i.e. graded module, graded algebra etc). A *dg category* in this paper will be understood to be a category enriched over dg ${k}$-modules. For example, if $A$ is a dg algebra then the category of dg $A$-modules is a dg category; similarly the category of contramodules over a pseudo-compact dg algebra is also a dg category. The dg space of homomorphisms in a dg category $C$ will be denoted by $\underline{\operatorname{Hom}}(-,-)$ and similarly for endomorphisms. The homotopy category $\operatorname{H^{0}}(C)$ of the dg category $C$ has the same objects as $C$ and for two objects $O_1, O_2$ in $C$ we have $\operatorname{Hom}_{\operatorname{H^{0}}(C)}(O_1,O_2):=H^0[\operatorname{\underline{Hom}}_C(O_1,O_2)]$. A dg functor $F:C{\rightarrow}C^\prime$ between two dg categories is *quasi-essentially surjective* if $\operatorname{H^{0}}(F):\operatorname{H^{0}}(C){\rightarrow}\operatorname{H^{0}}(C^\prime)$ is essentially surjective and *quasi-fully faithful* if $F$ induces quasi-isomorphisms on the $\operatorname{Hom}$-spaces; if both conditions are satisfied then $F$ is called a *quasi-equivalence*. A stronger notion is that of a *dg equivalence*: this is a dg functor $F:C{\rightarrow}C^\prime$ admitting a quasi-inverse dg functor $G:C^\prime{\rightarrow}C$, in the sense that there exist natural closed isomorphisms $F\circ G\cong {\mathbf 1}_{C^\prime}$ and $G\circ F\cong{\mathbf 1}_{C} $. A dg category is *strongly pre-triangulated* if it admits cones and shifts, and has a zero object (precise definitions can be found in e.g. [@Drinfeld04]), and *pre-triangulated* if it is quasi-equivalent to a strongly pre-triangulated category. A dg functor between pre-triangulated dg categories is a quasi-equivalence if and only if it induces an equivalence on their homotopy categories. A category dg-equivalent to a strongly pre-triangulated category is likewise strongly pre-triangulated. Examples of strongly pre-triangulated categories are provided by dg $A$-modules or dg $A$-contramodules where $A$ is a dg algebra or a dg pseudo-compact algebra respectively. If $X$ is a topological space, we denote by $C_*(X)$ its normalized singular chain dg coalgebra with coefficients in ${k}$ and by $C^*(X)$ its dual integer-valued normalized cochain (pseudo-compact) dg algebra; similarly if $X$ is a simplicial set, $C_*(X)$ and $C^*(X)$ will stand for its normalized chain dg coalgebra and normalized cochain (pseudo-compact) dg algebra. For a ${k}$-module $M$ we define by $\underline{M}$ the corresponding constant sheaf on a given topological space. For two dg sheaves $\mathcal {F,G}$ the corresponding dg sheaf of homomorphisms is denoted by ${\operatorname{{\mathcal{H}} \! \it{om}}}(\mathcal {F,G})$. We denote by $\Omega(M)$ the de Rham algebra of a smooth manifold $M$. If $K$ is a simplicial complex, then we write $\Omega(K)$ for its *piecewise smooth* de Rham algebra. Recall that a smooth form an $n$-simplex $\Delta^n$ is a smooth form on the interior of $\Delta^n$ such that it and all its derivatives extend continuously to the boundary of $\Delta^n$. It follows from Seeley’s extension theorem [@Seeley64] that such a form restricts to piecewise smooth forms on the faces of $\Delta^n$. The elements of $\Omega(K)$ are collections of smooth forms on the simplices of $K$ that are compatible with restriction maps. We define the sheaf $\Omega$ on $|K|$, the geometric realization of $K$, by setting $\Omega(U) = \lim_{\Delta^{n}\in K} {\mathcal{A}}^{*}(|\Delta^{n}|\cap U)$ for $U \subset |K|$. Then it is clear that $\Omega(K)$ coincides with the global sections of ${\Omega}$. When working with complete locally convex spaces $U$ and $V$, we will write $U\otimes V$ for the completed projective tensor product of $U$ and $V$; in the examples relevant to us, $U$ and $V$ will be nuclear, for which this choice or a tensor product is isomorphic to any other reasonable one. Acknowledgements ---------------- The authors would like to thank Jonathan Block, Chris Braun and Maxim Kontsevich for stimulating discussions. A substantial part of this paper was completed during the third author’s visit to IHES, and he wishes to thank this institution for excellent working conditions. Maurer-Cartan elements for algebras: basic notions, definitions and examples {#basic} ============================================================================ Let $A$ be a dg algebra. An element $x\in A^1$ is *Maurer-Cartan* or MC if it satisfies the equation $$\label{eq:MC} d(x) + x^{2}= 0.$$ The set of Maurer-Cartan elements in $A$ will be denoted by $\operatorname{MC}(A)$. The group $A^\times$ of invertible degree 0 elements in $A$ acts on $\operatorname{MC}(A)$ by *gauge equivalences*: for $g\in A^\times, x\in \operatorname{MC}(A)$ set $$g\cdot x=gxg^{-1}-d(g)g^{-1}$$ The *Maurer-Cartan moduli set* $\operatorname{\mathcal{MC}}(A)$ is the quotient of $\operatorname{MC}(A)$ modulo gauge equivalences. We now introduce the notion of MC *twisting*. For $x\in \operatorname{MC}(A)$ the dg $A$ module $A^{[x]}$ has $A$ as its underlying graded space and the differential $d^{[x]}:$ $$d^{[x]}(a):=d(a)+xa.$$ The right $A$-module structure on $A^{[x]}$ is the ordinary right multiplication. We will call $A^{[x]}$ the *module twisting* of $A$ by $x$. Similarly define the *algebra* twisting $A^x$ as the dg algebra having $A$ as an underlying graded algebra and the differential $d^x:$ $$d^x(a)=d(a)+[x,a].$$ Note that the MC condition (\[eq:MC\]) for $x$ implies (in fact, is equivalent to) $d^{[x]}$ squaring to zero in $A^{[x]}$. It also implies that $d^x$ squares to zero in $A^x$. With these definitions, $A^{[x]}$ becomes a dg $(A^x,A)$-bimodule. Let $X$ be a smooth manifold and $E{\rightarrow}X$ be a flat vector bundle on $X$. Consider $\operatorname{End}(E)$, the associated endomorphism bundle and set $A=\Omega(X,\operatorname{End}(E))$, the de Rham algebra of $X$ with values in $\operatorname{End}(E)$. The given flat structure determines a derivation $d$ on $A$ of square zero; if the bundle $E$ is topologically trivial then $d$ could be taken to be the ordinary de Rham differential. Then an MC element of $A$ is an $\operatorname{End}(E)$-valued 1-form $x$ on $X$ satisfying the MC equation (\[eq:MC\]). The set $\operatorname{MC}(A)$ is the set of all flat connections on the bundle $E$ and $\operatorname{\mathcal{MC}}(A)$ is the set of gauge equivalence classes of such flat connections. The complexes $A^{[x]}$ and $A^x$ are respectively one-sided and two-sided twisted de Rham complexes of $X$ with values in $\operatorname{End}(E)$. \[ex:universal\] Let $A:={k}[x], d(x)=-x^2$. Clearly $x$ is the unique non-zero MC element of $A$, and it is not gauge equivalent to 0. This algebra is universal in the sense that an MC element $y$ in a dg algebra $B$ is equivalent to a dg algebra map $A{\rightarrow}B$ with $x\mapsto y$. Note that $A$ is quasi-isomorphic to ${k}$, which implies that the MC moduli set is not quasi-isomorphism invariant. Recall that the category of (right) $A$-modules is enriched over dg modules: for any two right dg $A$-modules $M$ and $N$, we have the dg module of homomorphisms $\underline{\operatorname{Hom}}(M,N)$ from $M$ to $ N$; it is the graded vector space $\bigoplus_{n=-\infty}^\infty\operatorname{Hom}(M, N[n])$ with the differential $d(f)(m):= df(m)-(-1)^{|f|}f(dm)$. Then we have the following result whose proof is straightforward inspection. Let $x,y\in\operatorname{MC}(A)$. The dg space $A^{[x,y]}$ of right $A$-module homomorphisms $A^{[x]}{\rightarrow}A^{[y]}$ has $A$ as its underlying graded space and the differential $d^{[x,y]}:$ $$d^{[x,y]}(a):=d(a)+ya-(-1)^{|a|}ax.$$ The operations of left and right multiplications determine a dg $(A^y,A^x)$-bimodule structure on $A^{[x,y]}$. Note that for two right $A$-modules $M$ and $N$ a map $M{\rightarrow}N$ of right $A$-modules is precisely a zero-cocycle in $\underline{\operatorname{Hom}}(M,N)$. Then $M$ and $N$ are *homotopy equivalent* if there are maps of (right) $A$-modules $f:M{\rightarrow}N$ and $g:N{\rightarrow}M$ such that $f\circ g$ is cohomologous to $1\in\underline\operatorname{Hom}(N,N)$ and $g\circ f$ is cohomologous to $1\in \underline\operatorname{Hom}(M,M)$. The notion of a gauge equivalence of MC elements admits an important weakening to a *homotopy gauge equivalence*. Let $\operatorname{MC_{dg}}(A)$ be the dg category whose objects are MC elements of $A$ and for $x,y\in A$ the dg module of morphisms ${\operatorname{Hom}(x,y)_{\operatorname{MC_{dg}}(A)}:=\underline{\operatorname{Hom}}(A^{[x]},A^{[y]})}$. The correspondence $A\mapsto\operatorname{MC_{dg}}(A)$ is a functor from dg algebras to dg categories. Two MC elements $x,y\in A$ are called *homotopy gauge equivalent* if they are homotopy equivalent as objects in $\operatorname{MC_{dg}}(A)$. The *Maurer Cartan homotopy moduli set* $\operatorname{\mathcal{MC}_h}(A)$ is the set of isomorphism classes of objects in $\operatorname{H^{0}}(\operatorname{MC_{dg}}(A))$, i.e. the quotient of $\operatorname{MC}(A)$ modulo homotopy gauge equivalences. Thus, $x,y\in\operatorname{MC}(A)$ are homotopy gauge equivalent if there exist elements $g, h\in A^0$ such that 1. $dg+yg-gx=0$; 2. $dh+xh-hy=0$; 3. $hg$ is cohomologous to $1$ in $A^x$; 4. $gh$ is cohomologous to $1$ in $A^y$. Note that if $g\in A$ is invertible (i.e. $x$ and $y$ are *isomorphic*, as opposed to merely homotopy equivalent in $\operatorname{MC_{dg}}(A)$) then we could take $h=g^{-1}$ and conditions (2), (3) and (4) above are automatically implied by condition (1). In that case $x$ and $y$ are gauge equivalent. However, the following example shows that the relation of homotopy gauge equivalence is strictly weaker than that of gauge equivalence. Let $A:={k}\langle x,y,g,h,s,t\rangle$, the free algebra with two generators $x,y$ in degree 1, two generators $g,h$ in degree 0 and two generators $s,t$ in degree -1. The differential in $A$ is given by the formulae: $$\begin{aligned} d(x)&=-x^2,& d(y)&=-y^2,\\ d(g)&=gx-yg, & d(h)&=hy-xh ,\\ d(s)&=-xs+gh-1, & d(t)&=-yt+hg-1.\end{aligned}$$ It is clear that $x,y\in \operatorname{MC}(A)$ and that $g$ and $h$ provide maps of right dg $A$-modules $A^{[x]}{\rightarrow}A^{[y]}$ and $A^{[y]}{\rightarrow}A^{[x]}$ respectively that are homotopy equivalences with homotopies given by $s$ and $t$. As an aside, also note that $A$ is the universal dg algebra having two homotopy gauge equivalent MC elements in the sense that any other such dg algebra $B$ receives a unique map from $A$. Now $A$, being free, has no non-scalar invertible elements, and it follows that the MC elements $x$ and $y$ are not gauge equivalent, although they are homotopy gauge equivalent. Twisted modules and cohesive modules {#twisted} ==================================== We will now introduce the notion of a twisted module over a dg algebra $A$. A *twisted $A$-module* is a (right) dg $A$-module $M$ such that $M^{\#}$ is free as an $A^{\#}$-module. A twisted $A$-module is *finitely generated* if $M^{\#}$ is finitely generated. Finally, any twisted module that is a homotopy retract of a finitely generated twisted module is called a *perfect twisted module*. We will denote the dg category of twisted $A$-modules by $\operatorname{Tw}(A)$, and its full subcategories of finitely generated and perfect twisted $A$-modules by $\operatorname{Tw_{fg}}(A)$ and $\operatorname{Tw_{perf}}(A)$ respectively. If $A$ is a dg ring, then a dg $A$-module $M$ is sometimes called perfect if it represents a compact object in the derived category of $A$. This is *not* the same as a perfect twisted $A$-module, in particular the latter need not represent a compact object in a triangulated category. Later on, we will also use the notion of a perfect dg sheaf of modules. In all cases, our terminology will always be clear from the context and unambiguous. The data of a twisted $A$-module is clearly equivalent to that of an MC element $x\in\operatorname{End}(V)\otimes A$: for such an element, ${\mathbf 1}\otimes d_A+x$ gives a differential $D_V$ on $V\otimes A$ compatible with that of $A$ and any compatible differential on $V\otimes A$ must be of this form. We will often slightly abuse notation and write $V \otimes A$ for $(V \otimes A^{\#}, D_{V}).$ The correspondences $A\mapsto\operatorname{Tw}(A), A\mapsto\operatorname{Tw_{fg}}(A)$ and $A\mapsto\operatorname{Tw_{perf}}(A)$ are functors from dg algebras to dg categories; furthermore it is easy to see that $\operatorname{Tw}(A), \operatorname{Tw_{fg}}(A)$ and $\operatorname{Tw_{perf}}(A)$ are strongly pre-triangulated dg categories. Shifts are induced by the shift functor on $V$ and the cone on $f: (V \otimes A, D_V) {\rightarrow}(W \otimes A, D_W)$ is given by the complex $(V\oplus W[1])\otimes A$ with differential $\left(\begin{matrix} D_V & f \\ &D_W[1]\end{matrix}\right)$. The following result shows that the categories $\operatorname{Tw}(A)$ and $\operatorname{Tw_{perf}}(A)$ are closed with respect to taking retracts up to homotopy, i.e. their homotopy categories are idempotent complete. \[prop:idempotent\] Any idempotent morphism in $\operatorname{H^{0}}(\operatorname{Tw}(A))$ or $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A))$ is split. Since $\operatorname{H^{0}}(\operatorname{Tw}(A))$ is a triangulated category with direct sums, all idempotents in it split by [@Bokstedt93 Proposition 3.2]. The statement about $\operatorname{Tw_{perf}}(A)$ follows directly. \[remark:Morita\] We defined the category $\operatorname{Tw_{perf}}(A)$ as a certain subcategory of $\operatorname{Tw}(A)$. We see that $\operatorname{Tw}(A)$ is pre-triangulated, with $\operatorname{H^{0}}(\operatorname{Tw}(A))$ being idempotent complete; it is thus *Morita fibrant*, cf. [@Tabuada05] regarding this notion. Moreover, the inclusion of the category $\operatorname{Tw_{fg}}(A)$ of finitely generated twisted $A$-modules into $\operatorname{Tw_{perf}}(A)$ is a Morita morphism. Thus, $\operatorname{Tw_{perf}}(A)$ is a Morita fibrant replacement of $\operatorname{Tw_{fg}}(A)$ and could be defined, up to a quasi-equivalence, independently of the category $\operatorname{Tw}(A)$. The notion of a twisted $A$-module is closely related to that of a *cohesive* $A$-module cf. [@Block10]. Recall that a right dg $A$-module $M$ is cohesive if $M^{\#}$ is of the form $E \otimes_{A^{0}} A^{\#}$ for a graded right $A^{0}$-module $E$ that is projective, finitely generated in every degree and bounded. We will denote the dg category of cohesive $A$-modules by ${\mathcal{P}}_{A}$. The following result shows that any cohesive $A$-module is, up to a homotopy, a perfect twisted $A$-module. \[prop:retract\] Any cohesive $A$-module is a retract of a free cohesive $A$-module. The forgetful functor $A{\textrm{-Mod}}{\rightarrow}A^{\#}{\textrm{-Mod}}$ has a left adjoint sending a (right) $A^{\#}$-module $L$ to the $A$-module $G(L)$ consisting of formal symbols $x+dy$ for $x,y \in L$ with $A$-action given by $$(x+dy)a=xa+d(ya)-(-1)^{|y|}y\, da$$ and differential $d(x+dy) = dx$, see e.g. the proof of Theorem 3.6 in [@Positselski11]. The unit map $L {\rightarrow}G(L)$ is injective with cokernel isomorphic to $L[-1]$. In particular, if $L$ is projective, then $G(L)^{\#}$ is isomorphic to $L\oplus L[-1]$. Let $P$ be a dg $A$-module and assume that $P^{\#}$ is projective. Let $L$ be a (projective) $A^{\#}$-module such that $P^{\#}\oplus L$ is free. Then $P$ is a retract of $F \coloneqq P \oplus P[-1] \oplus G(L)$ and $F^{\#}$ is isomorphic to $P^{\#}\oplus P^{\#}[-1]\oplus L \oplus L[-1]$, which is a free $A^{\#}$-module. Note that if $P^{\#}$ is finitely generated then $F$ can be chosen so that $F^{\#}$ is of finite rank. In particular a cohesive module $M$ is a retract of a module $F$ such that $F^{\#}$ is a free $A^{\#}$-module of finite rank. But then we can write $F^{\#} = F'\otimes_{A^{0}}A^{\#}$ for some graded $A^{0}$-module $F'$ that is bounded and free of finite rank in every degree, i.e. $F$ is a free cohesive module. Under mild assumptions cohesive modules and perfect twisted modules agree. \[lemma-cohesiveretracts\] If $A$ is concentrated in non-negative degrees and $A$ is flat over $A^{0}$ then idempotents split in the homotopy category of cohesive modules. We call a bounded complex of finitely generated projective modules over $A^{0}$ *strictly perfect*; any dg-module $A^0$-module quasi-isomorphic to a strictly perfect will be called *perfect*. Let $h: E' \otimes_{A^0} A {\rightarrow}E' \otimes_{A^0} A$ be a homotopy idempotent. We can construct a splitting in the homotopy category of all $A$-modules by the well-known telescope trick. Writing $E = E' \otimes_{A^0} A$ we define a map $\sigma_{h} : \oplus_{{\mathbb{N}}} E \rightarrow \oplus_{{\mathbb{N}}} E$ defined by sending the $i$-th copy of $E$ to the $(i+1)$-st copy using $h$ and to the $i$-th copy using $1-h$. Then the cone of $\sigma_{h}$ splits $h$, i.e. there is an equivalence $E \simeq \operatorname{cone}(\sigma_{h})\oplus \operatorname{cone}(\sigma_{1-h})$. By construction $\operatorname{cone}(\sigma_{h})$ is of the form $(N' \otimes_{A^0} A, D_N)$ for some graded $A^0$-module $N'$. Moreover, inspecting the construction we see that $(N', D_N^{0})$ is equal to $\operatorname{cone}(\sigma_{h^{0}})$, which is the complex obtained by going through the above construction with $(E', h^{0})$ in place of $(E, h)$. To check this note that the underlying complex of $\operatorname{cone}(\sigma_{h})$ consists of a direct sum of copies of $E' \otimes A$ with some degree shifts. Writing the differential as a matrix each coefficient is given by ${\mathbf 1}$, $D_E$ or $h$. Dividing out by $A^{\geq 1}$ leaves a direct sum of shifted copies of $E'$ with differential given by a matrix of ${\mathbf 1}$, $D_E^{0}$ and $h^{0}$, which is exactly $\operatorname{cone}(\sigma_{h^{0}})$. The complex $\operatorname{cone}(\sigma_h)$ is a quasi-cohesive module in the sense of [@Block10] (i.e. a cohesive module without the assumption of finite generation) and we assumed that $A$ is in non-negative degrees and is flat over $A^{0}$. In this situation Theorem 3.2.7 of loc. cit. states that $\operatorname{cone}(\sigma_{h})$ is cohesive if $\operatorname{cone}(\sigma_{h^{0}})$ is perfect. But as $A$ is in non-negative degrees we can check that $h^{0}$ is a homotopy idempotent for $(E', D_E^{0})$ in $A^{0}{\textrm{-Mod}}$. In fact assuming $K$ is a homotopy from $h$ to $h^{2}$ then $K^{0}$ is a homotopy from $h^{0}$ to $(h^{0})^{2}$. Thus $\operatorname{cone}(\sigma_{h^{0}})$ is a direct summand of $E'$ in the homotopy category. We claim that this implies it is perfect. Following [@Quillen96] we say a map is algebraically nuclear if it factors through a strictly perfect complex. Then a complex is homotopy equivalent to a strictly perfect complex if and only if the identity is homotopy equivalent to a nuclear map, see [@Quillen96 Proposition 1.1]. Since the identity of $\operatorname{cone}(\sigma_{h^{0}})$ factors through $E'$ it is algebraically nuclear up to homotopy. This proves the claim and the lemma. \[cor-twistedcohesive\] If $A$ is concentrated in non-negative degrees and flat over $A^{0}$ then the dg categories $\operatorname{Tw_{perf}}(A)$ and ${\mathcal{P}}_{A}$ are quasi-equivalent. The inclusion of the dg category of finitely generated twisted modules $J:\operatorname{Tw_{fg}}(A) {\rightarrow}{\mathcal{P}}_{A}$ is quasi-fully faithful by construction; moreover it induces, by Proposition \[prop:retract\] and Lemma \[lemma-cohesiveretracts\] an equivalence on idempotent completions of its homotopy categories. It follows that $J$ is a Morita morphism and since by Lemma \[lemma-cohesiveretracts\] ${\mathcal{P}}_{A}$ is Morita fibrant, it could be viewed as a Morita fibrant replacement of $\operatorname{Tw_{fg}}(A)$. It is, thus, quasi-equivalent to $\operatorname{Tw_{perf}}(A)$, cf. Remark \[remark:Morita\]. Let $A=\Omega(X)$, the de Rham algebra of a smooth manifold $X$, and $E{\rightarrow}X$ be a flat vector bundle over $X$. Then $\Gamma(X,E)$, the sections of the bundle $E$, form a finitely generated projective module over $A^0$ and the given flat connection form on $E$ determines the structure of a cohesive $A$- module (and thus, of a perfect twisted $A$–module) on $\Gamma(X,E)\otimes_{A^0} A$. In good cases the homotopy category of twisted modules also agrees with Positselski’s derived category of the second kind [@Positselski11]. It follows from the proof of Proposition \[prop:retract\] that twisted $A$-modules agree up to homotopy with Positselski’s projective $A$-modules $A{\textrm{-Mod}}_{\text {proj}}$. Under certain conditions on $A^{\#}$, the underlying graded algebra of $A$, there is an equivalence $\operatorname{H^{0}}(A{\textrm{-Mod}}_{\text {proj}}) \cong D^{\text{ctr}}(A{\textrm{-Mod}})$. See Section 3.8 and 3.9 of [@Positselski11]. Note however, that the rings we are interested in in the second part of this paper do not satisfy Positselski’s conditions. The notions described in this and the previous section makes sense when $A$ is a pseudo-compact dg algebra. The definitions of $ \operatorname{\mathcal{MC}}(A), \operatorname{\mathcal{MC}_h}(A), \operatorname{MC_{dg}}(A), \operatorname{Tw}(A)$ and $\operatorname{Tw_{perf}}(A)$ are repeated verbatim. One slight subtlety is that the notions of twisted module over $A$ as a pseudo-compact dg algebra and as a discrete dg algebra (i.e. forgetting its pseudo-compact structure) are different, in general. This is because the tensor product of a pseudo-compact algebra and a (discrete) vector space is understood to be completed. Thus, for any dg algebra or pseudo-compact dg algebra $A$, we associated several invariants: the dg categories $\operatorname{MC_{dg}}(A), \operatorname{Tw}(A)$ and $\operatorname{Tw_{perf}}(A)$ as well as moduli sets $\operatorname{\mathcal{MC}}(A)$ and $\operatorname{\mathcal{MC}_h}(A)$. These are *not* quasi-isomorphism invariants of $A$ as, e.g. Example \[ex:universal\] demonstrates. Later on we will show that they are, nevertheless, homotopy invariants in two different contexts: analytic and algebraic. Smooth homotopies for dg algebras {#smooth} ================================= In this section we will consider dg Arens-Michael (AM) algebras. These are complete, Hausdorff, locally $m$-convex topological dg algebras over $\mathbb R$. For a detailed introduction see [@Pirkovskii08]. A special case of a dg AM algebra is a nuclear dg algebra, e.g the de Rham algebra $\Omega(X)$ where $X$ is a smooth manifold or a simplicial complex. For our purposes it is enough to know that any dg AM algebra is an inverse limit of dg Banach algebras. There is a natural notion of smooth homotopy between dg AM algebras. Let $f_0, f_1:A{\rightarrow}B$ be two continuous maps between dg AM algebras $A$ and $B$. A *smooth homotopy* between $f_0$ and $f_1$ is a map $F:A{\rightarrow}B\otimes\Omega[0,1]$ such that $({\mathbf 1}_A\otimes ev_{0})\circ F=f_0$ and $({\mathbf 1}_A\otimes ev_{1})\circ F=f_1$. Furthermore, $A$ and $B$ are called smooth homotopy equivalent if there are maps $f:A{\rightarrow}B$ and $g:B{\rightarrow}A$ such that $f\circ g$ and $g\circ f$ are smooth homotopic to ${\mathbf 1}_B$ and ${\mathbf 1}_A$ respectively. \[lem:diagonal1\] Any AM dg algebra $A$ is smooth homotopy equivalent to $A\otimes\Omega[0,1]$. It suffices to prove that $\Omega[0,1]$ is smooth homotopy equivalent to $\mathbb R$. This, in turn, would follow if we show that the map $1_{\Omega[0,1]}\circ ev_0:\Omega[0,1]{\rightarrow}\Omega[0,1]$ is smooth homotopic to the identity map on $\Omega[0,1]$. This last homotopy can be taken to be the diagonal map $\Delta:\Omega[0,1]{\rightarrow}\Omega[0,1]\otimes \Omega[0,1]$ induced by the multiplication $[0,1]\times[0,1]{\rightarrow}[0,1]$. \[prop:transitivity\] The relation of smooth homotopy on morphisms between AM algebras is an equivalence relation. Reflexivity is obvious and symmetry follows from the existence of a auto-diffeomorphism of $[0,1]$ swapping the endpoints. For transitivity consider a homotopy $F:A{\rightarrow}B\otimes\Omega[0,1]\cong B\otimes\Omega[0,\frac{1}{2}]$ such that $({\mathbf 1}_B\otimes ev_{0})\circ F=f_1$ and $({\mathbf 1}_B\otimes ev_{\frac{1}{2}})\circ F=f_2$, and another one $G:A{\rightarrow}B\otimes\Omega[0,1]\cong B\otimes\Omega[\frac{1}{2},1]$ such that $({\mathbf 1}_B\otimes ev_{\frac{1}{2}})\circ G=f_2$ and $({\mathbf 1}_B\otimes ev_{1})\circ G=f_3$, The homotopies $F$ and $G$ together constitute a map $$H:A{\rightarrow}B\otimes (\Omega[0,\frac{1}{2}]\times_{\mathbb{R}}\Omega[\frac{1}{2},1])$$ where the target of the last map could be viewed as $B$-valued forms on $[0,1]$ that are not necessarily smooth at $\frac{1}{2}$. To remedy the non-smoothness issue at $\frac{1}{2}$, let $h:[0,\frac{1}{2}]{\rightarrow}[0,\frac{1}{2}]$ be a diffeomorphism such that $h(0)=0,h(\frac{1}{2})=\frac{1}{2}$ and all higher derivatives of $h$ at $0$ and $\frac{1}{2}$ are zero. The correspondence $f{\rightarrow}F\circ h$ determines a homomorphism from the algebra of smooth functions on $[0,\frac{1}{2}]$ to the algebra of smooth functions whose derivatives at the endpoints are zero and similarly a homomorphism of the corresponding de Rham algebras. Note that this homomorphism preserves the values of the differential forms at the endpoints. Similarly, there is a homomorphism of the dg algebras $\Omega[\frac{1}{2},1]$ and $\Omega_0[\frac{1}{2},1]$. There is clearly a map of dg algebras $$\tilde{h}:\Omega_0[0,\frac{1}{2}]\times_{\mathbb{R}}\Omega_0[\frac{1}{2},1]{\rightarrow}\Omega[0,1]$$ whose image consists of the smooth forms on $[0,1]$ having zero derivatives at $\frac{1}{2}$. Then the map $$({\mathbf 1}_B\otimes\tilde{h})\circ H:A{\rightarrow}B\otimes\Omega[0,1]$$ is the desired homotopy between $f_1$ and $f_3$. There are also obvious notions of a polynomial or real analytic homotopy, both of which imply smooth homotopy. The relations of polynomial or analytic homotopy are not necessarily transitive. As in the discrete setting a MC element $x$ in a dg AM algebra $A$ is an element of degree $1$ such that $dx+x^2=0$ and we can define the gauge action etc. in the same way. Let $A$ be a dg AM algebra. Two MC elements $x_0, x_1\in A$ are called *smoothly homotopic* if there exists a MC element $X\in A\otimes\Omega[0,1]$ such that $({\mathbf 1}_A\otimes ev_{0})(X)=x_0$ and $({\mathbf 1}_A\otimes ev_{1})(X)=x_1$. We have the following result that is a direct consequence of Proposition \[prop:transitivity\]. The relation of smooth homotopy on MC elements of an AM algebra is an equivalence relation. \[lemma:transitive1\] Let $X=x(z)+y(z)dz$ be a smooth homotopy as above. Then it is equivalent to the system of equations $$\begin{aligned} \label{system} dx(z)+x(z)^2 &= 0, \\ \partial_zx(z)&= -dy(z)+[y(z),x(z)].\end{aligned}$$ \[thm-completess\] Two MC elements $x_0$ and $x_1$ are smoothly homotopic if and only if they are gauge equivalent via an element of $A^{\times}$ in the path component of $1$. Note first that we can, without loss of generality, assume that $A$ is a Banach space. Indeed, having a MC element in $A \coloneqq \varprojlim A_\alpha$ where $A_\alpha$ are Banach spaces, is the same as having a compatible collection of MC elements in every $A_\alpha$ (as MC elements are just maps from the algebra ${\mathbb{R}}[x \ | \ dx + x^{2} = 0]$). The same is true for gauge equivalences and also for homotopies since tensoring with the nuclear space $\Omega[0,1]$ commutes with inverse limits by Theorem \[thm:limit\]. The proof is similar to that in [@Chuang10 Theorem 4.4]. Suppose that two MC elements $x_0,x_1\in A$ are gauge equivalent; that means that there exists $g\in A^\times$ for which $g x_0g^{-1}-dg\cdot g^{-1} = x_1$. By assumption, there exists a smooth curve $g(z)$ with $g(0)=1$ and $g(1)=g$. Then define the homotopy $x(z)+y(z)$ in $A$ by $x(z)=g(z)x_0g^{-1}(z)-dg\cdot g^{-1}$ and $y(z)=\partial_z g(z)g^{-1}(z)$. Then a straightforward inspection shows that (\[system\]) holds. Conversely, suppose that there is a homotopy $x(z)+y(z)dz$ such that (\[system\]) holds. Consider the differential equation $$\label{difeq} \partial_z g(z)=y(z)g(z)$$ with the initial condition $g(0)=1$. (We note that this gives a compatible system of differential equations in Banach algebras.) If $g(z)$ is a solution of this differential equation and is invertible in $A$ then (\[system\]) would be satisfied with $g(z) x_0g^{-1}(z)-dg\cdot g^{-1}$ in place of $x(z)$ Since a solution of a linear differential equation in a Banach algebra is unique, the solution in the AM algebra $A$ is likewise unique and we conclude that, in fact, $x(z)=g(z)x_0g^{-1}(z)-dg\cdot g^{-1}$ and thus, $x_0$ and $x_1$ are gauge equivalent. But (\[difeq\]) does have the solution $g(z)=\operatorname{P}\exp\int_0^zy(t)dt$ where $\operatorname P\exp$ denotes the path ordered integral, defined by $$\operatorname{P}\exp\int_0^zy(t)dt \coloneqq 1 + \sum_{n=1}^{\infty} \int_{0\leq t_{1}\leq \dots \leq t_{n}\leq z} y(t_{n}) \cdots y(t_{1}) \ dt_{1}\cdots dt_{n}.$$ By [@Araki73 Propositions 3 and 4], $g(z)$ is invertible. \[cor-completehomotopygauge\] A smooth homotopy equivalence between two AM dg algebras $A$ and $B$ induces bijections $\operatorname{\mathcal{MC}}(A)\cong\operatorname{\mathcal{MC}}(B)$ and $\operatorname{\mathcal{MC}_h}(A)\cong\operatorname{\mathcal{MC}_h}(B)$. Let $f: A {\rightarrow}B$ and $g: B {\rightarrow}A$ be homomorphisms such that $g \circ f$ and $f\circ g$ are smoothly homotopic to the identity. Then $f$ and $g$ induce functions between $\operatorname{\mathcal{MC}}(A)$ and $\operatorname{\mathcal{MC}}(A)$. To see they are inverse note that, for each $a \in A$, $g \circ f(a)$ is smoothly homotopic to $a$ and thus by Theorem \[thm-completess\] it is gauge equivalent to $a$, similarly $f \circ g(b)$ is gauge equivalent to $b$. The bijection $\operatorname{\mathcal{MC}_h}(A)\cong\operatorname{\mathcal{MC}_h}(B)$ is proved in the same way. We would like to consider twisted modules over AM algebras. Since an endomorphism algebra of an infinite-dimensional space is, in general, not AM, it is not clear whether an arbitrary (infinitely generated) twisted module is a reasonable notion. For an AM algebra $A$, we consider the dg category $\operatorname{Tw_{fg}}(A)$ of dg $A$-modules of the form $V\otimes A$ where $V$ is a finite-dimensional $\mathbb R$-space and denote $\operatorname{Tw_{perf}}(A)$, its Morita fibrant replacement. The latter can be obtained, e.g. by taking the closure of the Yoneda embedding of $\operatorname{Tw_{fg}}(A)$ with respect to homotopy idempotents. \[rk:twfgam\] Note that the definition of $\operatorname{Tw_{perf}}(A)$ depends, strictly speaking, on whether $A$ is viewed as an AM algebra or a discrete one since in the latter case $\operatorname{Tw_{perf}}(A)$ is defined in terms of $\operatorname{Tw}(A)$ which is not considered for an AM algebra $A$. Nevertheless, this is only an ambiguity up to quasi-equivalence since the notion of a Morita fibrant replacement is well-defined up to a quasi-equivalence of dg categories, cf. Remark \[remark:Morita\]. Corollary \[cor-twistedcohesive\] continues to hold for an AM algebra $A$ with the same proof. A map $A{\rightarrow}B$ of AM algebras induces functors $\operatorname{MC_{dg}}(A){\rightarrow}\operatorname{MC_{dg}}(B)$ and $\operatorname{Tw_{perf}}(A){\rightarrow}\operatorname{Tw_{perf}}(B)$. It is natural to ask how these induced functors differ for smoothly homotopic maps. The following result answers this question. \[prop:homotopic\] Let $f,g:A{\rightarrow}B$ be two smoothly homotopic maps of AM algebras. Then the induced functors on *homotopy categories* $\operatorname{H^{0}}\operatorname{MC_{dg}}(A){\rightarrow}\operatorname{H^{0}}\operatorname{MC_{dg}}(B)$ and $\operatorname{H^{0}}\operatorname{Tw_{perf}}(A){\rightarrow}\operatorname{H^{0}}\operatorname{Tw_{perf}}(B)$ are isomorphic. \[lem:homotopic\] For an AM algebra $A$ consider the two natural maps $${\mathbf 1}_A\otimes ev_{0,1}:A\otimes\Omega[0,1]\rightrightarrows A.$$ Then the induced functors $$\operatorname{H^{0}}\operatorname{MC_{dg}}(A\otimes\Omega[0,1])\rightrightarrows \operatorname{H^{0}}\operatorname{MC_{dg}}(A)$$ are isomorphic. The map $i:A{\rightarrow}A\otimes \Omega[0,1];a\mapsto a\otimes 1$ induces a quasi-equivalence $$\operatorname{MC_{dg}}(A)(i):\operatorname{MC_{dg}}(A){\rightarrow}\operatorname{MC_{dg}}(A\otimes\Omega[0,1]).$$ Indeed, $i$ is a smooth homotopy equivalence by Lemma \[lem:diagonal1\], and Corollary \[cor-completehomotopygauge\] then implies that $\operatorname{MC_{dg}}(A)(i)$ is quasi-essentially surjective (even essentially surjective). The quasi-fully faithfulness of $\operatorname{MC_{dg}}(A)(i)$ follows from acyclicity of $\Omega[0,1]$. The composition $({\mathbf 1}_A\otimes ev_{0})\circ i:A{\rightarrow}A$ is clearly the identity map on $A$ and it follows that the map ${\mathbf 1}_A\otimes ev_{0}$ induces the functor $\operatorname{H^{0}}\operatorname{MC_{dg}}(A\otimes\Omega[0,1])\rightarrow \operatorname{H^{0}}\operatorname{MC_{dg}}(A)$ that is quasi-inverse to the one induced by $i$. The same can be said about the functor induced by ${\mathbf 1}_A\otimes ev_{1}$. Since quasi-inverse functors are determined uniquely up to an isomorphism, the desired claim follows. . Let $h:A{\rightarrow}B\otimes\Omega[0,1]$ be a smooth homotopy between $f$ and $g$. Then $({\mathbf 1}_B\otimes ev_1)\circ h=f$ and $({\mathbf 1}_B\otimes ev_2)\circ h=g$, and applying Lemma \[lem:homotopic\] we obtain the desired result. \[thm:homotopyequivalence1\] Let $A$ and $B$ be two dg AM algebras that are smoothly homotopy equivalent. Then there is a quasi-equivalence between the dg categories 1. $\operatorname{MC_{dg}}(A)$ and $\operatorname{MC_{dg}}(B)$, 2. $\operatorname{Tw_{perf}}(A)$ and $\operatorname{Tw_{perf}}(B)$. Let $f:A{\rightarrow}B$ and $g:B{\rightarrow}A$ be the dg algebra maps such that $f\circ g$ is smoothly homotopic to ${\mathbf 1}_B$ and $g\circ f$ is smoothly homotopic to ${\mathbf 1}_A$. These maps induce functors $\operatorname{MC_{dg}}(A)(f):\operatorname{MC_{dg}}(A){\rightarrow}\operatorname{MC_{dg}}(B)$ and $\operatorname{MC_{dg}}(A)(g):\operatorname{MC_{dg}}(B){\rightarrow}\operatorname{MC_{dg}}(A)$. Corollary \[cor-completehomotopygauge\] implies that $\operatorname{MC_{dg}}(A)(f)$ is quasi-essentially surjective and Proposition \[prop:homotopic\] – that $\operatorname{MC_{dg}}(A)(f)$ and $\operatorname{MC_{dg}}(A)(g)$ induce an equivalence $\operatorname{H^{0}}\operatorname{MC_{dg}}(A){\rightarrow}\operatorname{H^{0}}\operatorname{MC_{dg}}(B)$. It follows that $\operatorname{MC_{dg}}(A)$ and $\operatorname{MC_{dg}}(B)$ are quasi-equivalent. The same argument establishes a quasi-equivalence between $\operatorname{Tw_{fg}}(A)$ and $\operatorname{Tw_{fg}}(B)$ after one observes that a finitely generated twisted $A$-module is the same as an MC element in the dg algebra $A\otimes\operatorname{End}(V)$ where $V$ is a graded finitely generated free ${k}$-module and similarly for $B$. It follows that $\operatorname{Tw_{perf}}(A)$ and $\operatorname{Tw_{perf}}(B)$ (as Morita fibrant replacements of $\operatorname{Tw_{fg}}(A)$ and $\operatorname{Tw_{fg}}(B)$) are quasi-equivalent. If $X$ is a smooth manifold or a simplicial complex we write $\operatorname{MC}(X)$ and $\operatorname{Tw}(X)$ for the dg categories $\operatorname{MC}(\Omega(X))$ and $\operatorname{Tw}(\Omega(X))$ respectively. For a smooth (piecewise smooth in the case of simplicial complexes) homotopy $X\times[0,1]{\rightarrow}Y$ of maps between $X$ and $Y$, the associated map $\Omega(Y){\rightarrow}\Omega(Y\times[0,1])\cong\Omega{Y}\otimes\Omega[0,1]$ (see Corollary \[cor:product\] regarding the last isomorphism) is a smooth homotopy of the corresponding dg AM algebras. Therefore, we have the following result. \[cor:deRham\] Let $f,g:X{\rightarrow}Y$ be (piecewise) smooth homotopic maps between $M$ and $N$. Then the induced functors $$\operatorname{MC_{dg}}(A)(f), \operatorname{MC_{dg}}(A)(g):\operatorname{H^{0}}\operatorname{MC}(X)\rightrightarrows\operatorname{H^{0}}\operatorname{MC}(Y);$$ $$\operatorname{MC_{dg}}(A)(f), \operatorname{MC_{dg}}(A)(g):\operatorname{H^{0}}\operatorname{Tw}(X)\rightrightarrows\operatorname{H^{0}}\operatorname{Tw}(Y)$$ are isomorphic. If $X$ and $Y$ are (piecewise) smooth homotopy equivalent smooth manifolds or simplicial complexes then the following dg categories are quasi-equivalent. $$\operatorname{MC_{dg}}(X) \text{~and~} \operatorname{MC_{dg}}(Y),$$ $$\pushQED{\qed} \operatorname{Tw_{perf}}(X) \text{~and~} \operatorname{Tw_{perf}}(Y). \qedhere \popQED$$ Strong homotopies for dg algebras {#strong} ================================= In this section we introduce the notion of strong homotopy between maps of dg algebras and the concomitant notion of strong homotopy equivalence of dg algebras. All definitions, results and proofs are applicable verbatim to dg pseudo-compact algebras as long as we keep in mind our conventions that homomorphisms of pseudo-compact algebras are assumed to be continuous, unmarked tensor products are automatically completed etc. Let $I$ be the singular simplicial set of the unit interval $[0,1]$; recall that the set $I_n$ of $n$-simplices of $I$ is the set of singular $n$-simplices of $[0,1]$, i.e. the set of continuous maps $\Delta^n{\rightarrow}[0,1]$ where $\Delta^n$ is the standard topological $n$-simplex. We will consider a collection of simplicial subsets of $I$ defined as follows. 1. The simplicial set $K_0$ is generated by two nondegenerate simplices $a_0,b_0$ in degree zero corresponding to the endpoints of $[0,1]$ and one nondegenerate simplex $a_1$ in degree one corresponding to the linear path from $0$ to $1$ in $[0,1]$, viewed as a $1$-simplex in $[0,1]$. 2. The simplicial set $K_1$ contains all the simplices of $K_0$ and has, additionally, one other nondegenerate $1$-simplex $b_1$ corresponding to the linear path from $1$ to $0$ in $[0,1]$. 3. Assuming that for $n\geq 1$ the simplicial set $K_n$ has been defined, we let $K_{n+1}$ contain all the simplices of $K_n$ plus two additional nondegenerate simplices $a_n,b_n$ in degree $n$ defined as follows. Writing $\Delta^n$ as the convex hull of its vertices $x_0,\ldots, x_n$, we let $a_n:\Delta^n{\rightarrow}[0,1]$ and $b_n:\Delta^n{\rightarrow}[0,1]$ be the affine maps for which $$a(x_i)=\begin{cases}0, \text{if $i$ is even,}\\1,\text{if $i$ is odd}\end{cases} \text{and~} b(x_i)=\begin{cases}1, \text{if $i$ is even,}\\0,\text{if $i$ is odd.}\end{cases}$$ 4. The simplicial set $K_\infty$ is the union of the nested sequence of simplicial sets $K_1\subset K_2\subset\ldots$. We have the following inclusions of the simplicial sets introduced above: $$K_0\subset K_1\subset\ldots \subset K_\infty\subset I$$ as well as their geometric realizations $|K_n|$. It is clear that $|K_0|$ is a cell decomposition of $[0,1]$ with two 0-cells and one 1-cell. Furthermore, for $n=1,\ldots,\infty$ the cell complex $|K_n|$ is homeomorphic to the $n$-sphere $S^n$ with two cells in each dimension. \[lemma:retract\] The simplicial set $K_\infty$ is a retract of $I$. Consider the category $\mathcal K$ with two objects and two mutually inverse morphisms between them. The simplicial set $K_\infty$ is, by definition, the nerve of $\mathcal K$. Since $\mathcal K$ is a groupoid, its nerve is a Kan simplicial set (cf. [@Goerss99 Lemma 3.5]). Clearly, the inclusion $K_\infty{\rightarrow}I$ is an acyclic cofibration and it follows that it admits a splitting, exhibiting $K_\infty$ as a retract of $I$. Since the simplicial sets $K_n, 1\leq n<\infty$ are not contractible, they are not retracts of $I$. The simplicial set $K_0$, while contractible, is still not a retract of $I$ since it is not Kan. We denote by $K_n^*, n=0,1,\ldots, \infty$ and $I^*$ the complexes of normalized cochains on the corresponding simplicial sets with values in ${k}$. Endowed with the Alexander-Whitney product, these become dg algebras, in fact pseudo-compact dg algebras (as duals to dg coalgebras). We re-iterate that, even though $K_\infty^*$ is a degree-wise finitely generated free ${k}$-module, it will be regarded as pseudo-compact, in particular tensor products with it will always be understood in the completed sense, as per our convention. Note that this subtlety is vacuous for $K_n^*, n<\infty$ as these free ${k}$-modules have totally finite rank. We have the following tower of surjective maps of dg pseudo-compact algebras: $$K_0^*\leftarrow K_1^*\leftarrow\ldots\leftarrow K_\infty^*\leftarrow I^*.$$ Note that any pseudo-compact dg algebra in this tower admits two maps $ev_1$ and $ev_2$ to ${k}$ corresponding to the inclusion of the two endpoints of $[0,1]$ into the corresponding simplicial set. We can define the notion of a $K$-multiplicative homotopy of dg algebra maps where $K$ is any simplicial subset of $I$ containing the 0-simplices corresponding to the endpoints of $[0,1]$. In the following definition $K$ is $K_n, n=0, 1,\ldots,\infty$, or $I$. Let $f,g:A{\rightarrow}B$ be two dg algebra maps. An *elementary $K$-homotopy* between them is a map $H:A{\rightarrow}B\otimes K^*$ such that $({\mathbf 1}_B \otimes ev_1)(H)=f$ and $({\mathbf 1}_B\otimes ev_2)(H)=g$. We say that $f$ and $g$ are $K$-homotopic, if they are related by the equivalence relation generated by elementary $K$-homotopy. If $K=I$, we will refer to $K$-homotopy as *strong homotopy*. Furthermore, $A$ and $B$ are called $K$ homotopy equivalent if there are maps $f:A{\rightarrow}B$ and $g:B{\rightarrow}A$ such that $f\circ g$ and $g\circ f$ are $K$-homotopic to ${\mathbf 1}_B$ and ${\mathbf 1}_A$ respectively. If $K=I$ we will refer to a $K$ homotopy equivalence as a *strong homotopy equivalence*. It is easy to see that for $n>0$ the relation of elementary homotopy is symmetric but not transitive and for $n=0$ it is not even symmetric. Furthermore, using *normalized* cochains is essential: almost all of our results will fail for un-normalized cochains. For example, the un-normalized singular cochain algebra of the one-point topological space is the dg algebra of Example \[ex:universal\] having non-trivial dg categories of MC elements and twisted modules. Since $K_\infty$ is a retract of $I$, the notions of strong homotopy and strong homotopy equivalence are equivalent to those of a $K_\infty$ homotopy and $K_\infty$ homotopy equivalence respectively. It is this notion of multiplicative homotopy that is of chief relevance to this paper. Also of interest is the notion of $K_0$ homotopy (sometimes called *derivation homotopy*); it has been used in rational homotopy theory, cf. for example [@Anick89]. \[lem:diagonal2\] Any dg algebra $A$ is strongly homotopy equivalent to $A\otimes I^*$ (and thus, also to $A\otimes K_\infty^*$). The multiplication map $[0,1]\times[0,1]{\rightarrow}[0,1]$ makes $I^*$ into a bialgebra and the coproduct map $I^*{\rightarrow}I^*\otimes I^*$ could be viewed as a strong homotopy between the identity map on $I^*$ and a projection onto ${k}$. It follows that $I^*$ (and thus, $K_\infty^*$) is strongly homotopy equivalent to ${k}$ and the desired statement is an immediate consequence. Similarly, $K_0^*$ is $K_0$-homotopy equivalent to ${k}$ and so $A$ is $K_0$ homotopy equivalent to $A\otimes K_0$; we will not use this result. Since for $0<n<\infty$ the algebra $K_n^*$ is not acyclic, it is not $K_n^*$ homotopy equivalent to ${k}$. We will see later on (Example \[eg:K\_0\]) that $K_0^*$ is not $K_2$ homotopy equivalent to ${k}$. \[prop:comp\] Let $f,g:A{\rightarrow}B$ be two $K$-homotopic dg algebra maps. If $C$ is a third dg algebra then for any dg algebra map $h:C{\rightarrow}A$ the maps $f\circ h,g\circ h:C{\rightarrow}B$ are $K$-homotopic. Similarly for any dg algebra map $k:B{\rightarrow}C$ the maps $k\circ f, k\circ g:A{\rightarrow}C$ are $K$-homotopic. It suffices to treat the case of an elementary homotopy. If $H:A{\rightarrow}B\otimes K^*$ is an elementary homotopy between $f$ and $g$ then $H\circ h$ is an (elementary) $K$-homotopy between $f\circ h$ and $g\circ h$. Similarly, $(k\otimes {\mathbf 1}_{K^*})\circ H$ is an (elementary) homotopy between $k\circ f$ and $ k\circ g$. Using Proposition \[prop:comp\] we can define the $K$-homotopy category of dg algebras as having dg algebras as objects and $K$-homotopy classes of maps as morphisms. Of most interest is the case $K=K_\infty$ and $K=K_0$ as $K_\infty^*$ and $K_0^*$ are acyclic dg algebras. As was mentioned earlier, $K_0^*$ is not $K_\infty$ contractible and so, the relation of $K_\infty$ homotopy equivalence is strictly finer than that of $K_0$-equivalence. Moreover, the existence of a $K$-homotopy category of dg algebras suggest the existence of a closed model category structure underpinning it. The standard closed model structures on dg algebras having quasi-isomorphisms as weak equivalences, should then be localizations of the $K_\infty$ closed model structure. The main advantage of the dg algebra $K^*_\infty$ over $I^*$ is that the former is much smaller, and $K_{{\infty}}$, as well as its quotients $K_n,n<\infty$, admits an explicit description. 1. The dg algebra $K_\infty^*$ is generated by two elements $e,f$ in degree zero and two elements $s,t$ in degree one, subject to the relations $$\begin{aligned} e^2 &=e, &f^2 &=f, & ef &=fe=0;\\ fs&=s,&se&=s, &sf &=es=0;\\ tf&=t, &et&=t, &ft &=te=0;\\ t^2 &=s^2=0\end{aligned}$$ with the differential specified by the formulae $$\begin{aligned} d(e)&=t-s, &d(t)&=s-t; \\ d(s)&=ts+st, &d(t)&=st+ts.\end{aligned}$$ 2. The algebra $K_n^*,0<n<\infty$ is the quotient of $K_\infty^*$ by the dg ideal generated by monomials in $s$ and $t$ of length $>n$. 3. The algebra $K_0^*$ is the quotient of $K_\infty^*$ by the dg ideal generated by $t$ and polynomials in $s$ of degree $>1$. Statements (2) and (3) clearly follow from (1). To prove (1), we use the interpretation of $K^*_\infty$ as the normalized cochain algebra of the nerve of the category with two objects and two mutually inverse morphisms between them as in the proof of Lemma \[lemma:retract\]. It follows that $K^*_\infty$ is the path algebra of the graded quiver with arrows $s$ and $t$ placed in degree 1. The stated relations in $K^*$ are precisely the relations in this path algebra, with the elements $e$ and $f$ corresponding to the length zero paths at the vertices of the above quiver. The formula for the differential in $K^*_\infty$ is straightforward to obtain. Strong homotopies for MC elements --------------------------------- There is a corresponding notion of K-homotopy for MC elements. Two MC elements $x_0, x_1$ in a dg algebra $A$ are called K-homotopic if there exists an MC element $X\in A \otimes K^*$ such that $({\mathbf 1}_A \otimes ev_0)(X)=x_0$ and $({\mathbf 1}_A \otimes ev_1)(X)=x_1$. If $K=I$, this will be referred to as *strong homotopy* of MC elements. It turns out that the notions of $K_2$ homotopy and homotopy gauge equivalence for MC elements are equivalent. \[lemma:K2gauge\] Let $x,x^\prime$ be two MC elements in a dg algebra $A$. Then $x$ and $x^\prime$ are homotopy gauge equivalent if and only if they are $K_2$-homotopic. Let $X\in A\otimes K^*_2$ be a $K_2$-homotopy between $x$ and $x^\prime$. We could write $$X=xe+x^\prime f+ys+y^\prime t+zts+z^\prime st,$$ where $y,y^\prime$ and $z,z^\prime$ are elements of $A$ of degrees $0$ and $1$ respectively. Writing down the MC equation for $X$ and equating to zero the coefficients at $e,f,s,t, st$ and $ts$ we obtain: $$\begin{aligned} x^2+dx=0; (x^\prime)^2+dx^\prime &=0\\ dy+x^\prime(y+1)-(y+1)x &=0\\ dy^\prime+x(y^\prime+1)-(y^\prime+1)x^\prime &=0\\ (y+1)(y^\prime+1)-1+dz+[x,z] &=0\\ (y^\prime+1)(y+1)-1+dz^\prime+[x^\prime,z^\prime] &=0.\end{aligned}$$ The first line above is the statement that $x$ and $x^\prime$ are MC elements in $A$, the second and third – that the elements $y+1$ and $y^\prime+1$ determine right $A$-module maps $A^{[x]}{\rightarrow}A^{[x^\prime]}$ and $A^{[x^\prime]}{\rightarrow}A^{[x]}$ respectively and the last two lines – that the elements $(y+1)(y^\prime+1)$ and $(y^\prime+1)(y+1)$ are cohomologous to $1$ in $A^x$ and $A^{x^\prime}$ respectively. It follows that $x$ and $x^\prime$ are homotopy gauge equivalent. Conversely, if $x$ and $x^\prime$ are homotopy gauge equivalent, then performing the above calculations in the reverse order, we find a $K_2$-homotopy between $x$ and $x^\prime$. Rather surprisingly, the notions of $K_2$ and $K_\infty$ homotopy for MC elements are equivalent. This could be interpreted as a strong homotopy analogue of the Schlessinger-Stasheff theorem. Strikingly, it holds with no assumptions on the dg algebra in question. To show this, we need a few preliminary results. Recall that we introduced a category $\mathcal K$ having two objects $O_1$ and $O_2$ and two mutually inverse morphisms between them. Let $\mathcal K_\infty$ be the dg category with the same set of objects $O_1$ and $O_2$ and a set of free generators: $$x_n:O_1{\rightarrow}O_2; y_n:O_2{\rightarrow}O_1 \text{ for } n=0,1,\ldots$$ with $|x_n|=|y_n|=n$. The differential $d$ is given on the generators as follows: $$\begin{aligned} d(x_0)&=0, &d(y_0)&=0;\\ d(x_1)&=y_0x_0-1, &d(y_1)&=x_0y_0-1\end{aligned}$$ and for $n>0$: $$\begin{aligned} d(x_{2n})=& \sum_{i=0}^{n-1}(x_{2i}x_{2(m-i)-1}-y_{2(m-i)-1}y_{2i});\\ d(y_{2n})=& \sum_{i=0}^{n-1}(y_{2i}y_{2(n-i)-1}-x_{2(n-i)-1}y_{2i});\\ d(x_{2n+1})=&\sum_{i=0}^ny_{2i}x_{2(n-i)}-\sum_{i=0}^{n-1}x_{2i-1}x_{2(n-i)-1};\\ d(y_{2n+1})=&\sum_{i=0}^nx_{2i}y_{2(n-i)}-\sum_{i=0}^{n-1}y_{2i+1}y_{2(n-i)-1}.\end{aligned}$$ Note that $\mathcal K_\infty$ is a cofibrant dg category. Clearly there is a surjection $\mathcal K_\infty{\rightarrow}\mathcal K$ whose kernel is generated by all $x_n, y_n,n>0$. Then we have the following result. \[lemma:resolution\] The map $\mathcal K_\infty{\rightarrow}\mathcal K$ is a quasi-isomorphism, i.e. $\mathcal K_\infty$ is a cofibrant resolution of $\mathcal K$. This is proved in [@Markl01 Theorem 9,]; note that $\mathcal K$ and $\mathcal K_\infty$ are called ‘coloured operads’ in the cited reference but these are really dg categories as they do not support operations of higher arities. - The proof of Lemma \[lemma:resolution\] in [@Markl01] is computational. In fact, the resolution $\mathcal K_\infty{\rightarrow}\mathcal K$ is the standard reduced bar-cobar resolution of the category $\mathcal K$. The existence of such a resolution seems to be a well-known fact and is mentioned, in, e.g. [@Drinfeld04; @Keller06]. We are, however, unaware of any reference where this general fact has been given a full proof. - A different (smaller) resolution of the category $\mathcal K$ was described in [@Drinfeld04 Corollary 3.7.3]. - A one object analogue of the dg category $\mathcal K$ is the algebra ${k}[s,s^{-1}]$ with $|s|=0$. A cofibrant resolution of this algebra was constructed in [@Braun15]; the formulae are essentially the same as for $\mathcal K_\infty$. \[lemma:category-algebra\] Let $A$ be a dg algebra and $x, x^\prime\in \operatorname{MC}(A)$. Then there is a 1-1 correspondence between strong homotopies from $x$ to $x^\prime$ and dg functors $\mathcal K_\infty{\rightarrow}\operatorname{MC_{dg}}(A)$ sending $O_1$ and $O_2$ to $x$ and $x^\prime$ respectively. Let $X\in \operatorname{MC}(A\otimes K^*_\infty)$ be a strong homotopy from $x$ to $x^\prime$. We could write $$X=xe+x^\prime f +u_0s+v_0t+u_1st+v_1ts+\ldots.$$ In other words the coefficient of $X$ at the monomial $st\ldots t$ or $st\ldots s$ of length $n$ is $u_n$ and the coefficient at the monomial $ts\ldots t$ or $ts\ldots s$ of length $n$ is $v_n$. Note that the $u_n,v_n$ are elements of $A$ of degree $n$. Similarly, a dg functor $F:\mathcal C{\rightarrow}\operatorname{MC_{dg}}(A)$ such that $F(O_1)=x$ and $F(O_2)=x^\prime$ is determined (since $\mathcal C$ is a free category) by a collection of elements $$\begin{aligned} F(x_{2n}) &\in \operatorname{Hom}_{\operatorname{MC_{dg}}(A)}(x,x^\prime),\\ F(y_{2n}) &\in \operatorname{Hom}_{\operatorname{MC_{dg}}(A)}(x^\prime,x),\\ F(x_{2n+1})&\in \operatorname{Hom}_{\operatorname{MC_{dg}}(A)}(x,x) ,\\ F(y_{2n+1})&\in \operatorname{Hom}_{\operatorname{MC_{dg}}(A)}(x^{\prime},x^\prime) \end{aligned}$$ where $n=0,1,\ldots$. The correspondence between these two sets of data is given by $$F(x_0)=u_0+1, F(y_0)=v_0+1$$ and, for $n>0$: $$F(x_n)=u_n, F(y_n)=v_n.$$ Finally, a somewhat tedious but straightforward calculation, similar to that of Lemma \[lemma:K2gauge\] shows that the MC equation $d(X)+X^2=0$ translates into the condition that $F$ is a dg functor (i.e. determines a dg map on $\operatorname{Hom}$-complexes). \[thm-singularss\] Let $A$ be a dg algebra. Then two MC elements in $A$ are strongly homotopic if and only if they are homotopy gauge equivalent. If two MC elements $x, x^\prime$ in $A$ are strongly (or $K_\infty$) homotopic then they are $K_2$ homotopic since $K_2^*$ is a quotient of $K_\infty^*$ and thus by Lemma \[lemma:K2gauge\] they are homotopy gauge equivalent. Conversely, let $x,x^\prime\in \operatorname{MC}(A)$ be homotopy gauge equivalent and consider a map $f:x{\rightarrow}x^\prime$ inducing an isomorphism in $\operatorname{H^{0}}(\operatorname{MC_{dg}}(A))$. Let $\mathcal K_0$ be the ${k}$-linear category generated by two objects $O$ and $O^\prime$ and a single morphism $i:O{\rightarrow}O^\prime$. Then there is a unique dg functor $F:\mathcal K_0{\rightarrow}\operatorname{MC_{dg}}(A)$ mapping $i$ to $f$. Since $f$ represents an isomorphism in $\operatorname{H^{0}}(\operatorname{MC_{dg}}(A))$, the functor $F$ factors through $L_i(\mathcal K_0)$, the derived localisation of $\mathcal K_0$, cf. [@Toen07b]. On the other hand, it follows from the proof of [@Toen07b Corollary 9.7] that $L_i(\mathcal K_0)$ is quasi-equivalent to the category $\mathcal K$ consisting of two mutually inverse isomorphisms between two objects $O_1$ and $O_2$. Since $\mathcal K_\infty$ is a cofibrant replacement of $\mathcal K$, we obtain a dg functor $\mathcal K_\infty{\rightarrow}\operatorname{MC_{dg}}(A)$ taking $O_1$ and $O_1$ to $x$ and $x^\prime$ respectively. By Lemma \[lemma:category-algebra\] this implies that $x$ and $x^\prime$ are strongly homotopic. For $n=2,3\ldots,\infty$ the relation of $K_n$-homotopy on MC elements of a dg algebra is an equivalence relation. Indeed, by Theorem \[thm-singularss\] two MC elements in a dg algebra $A$ are $K_n$-homotopic if and only if they are homotopy equivalent as objects in $\operatorname{MC_{dg}}(A)$. The latter relation is obviously an equivalence relation. \[cor:strongh\] A strong homotopy equivalence between two dg algebras $A$ and $B$ induces a bijection $\operatorname{\mathcal{MC}_h}(A)\cong\operatorname{\mathcal{MC}_h}(B)$. The given strong homotopy equivalence between A and B clearly induces a bijection of MC elements up to strong homotopy. By Theorem \[thm-singularss\] this becomes a bijection of MC elements up to homotopy gauge equivalence, i.e. a bijection $\operatorname{\mathcal{MC}_h}(A)\cong\operatorname{\mathcal{MC}_h}(B)$. The following result is a strong homotopy analogue of Proposition \[prop:homotopic\] and its proof is completely analogous, after replacing $\Omega[0,1]$ with $I^*$ and Corollary \[cor-completehomotopygauge\] with Corollary \[cor:strongh\]. \[prop:homotopic1\] Let $f,g : A{\rightarrow}B$ be two strongly homotopic maps of dg algebras. Then the induced functors on homotopy categories: 1. $\operatorname{H^{0}}\operatorname{MC_{dg}}(f), \operatorname{H^{0}}\operatorname{MC_{dg}}(g): \operatorname{H^{0}}\operatorname{MC_{dg}}(A){\rightarrow}\operatorname{H^{0}}\operatorname{MC_{dg}}(B)$; 2. $\operatorname{H^{0}}\operatorname{Tw}(f), \operatorname{H^{0}}\operatorname{Tw}(g): \operatorname{H^{0}}\operatorname{Tw}(A){\rightarrow}\operatorname{H^{0}}\operatorname{Tw}(B)$; 3. $\operatorname{H^{0}}\operatorname{Tw_{perf}}(f), \operatorname{H^{0}}\operatorname{Tw_{perf}}(g): \operatorname{H^{0}}\operatorname{Tw_{perf}}(A){\rightarrow}\operatorname{H^{0}}\operatorname{Tw_{perf}}(B)$ are isomorphic. The following result is a strong homotopy analogue of Theorem \[thm:homotopyequivalence1\] and the proof is completely analogous, after replacing $\Omega[0,1]$ with $I^*$, Corollary \[cor-completehomotopygauge\] with Corollary \[cor:strongh\] and Proposition \[prop:homotopic\] with Proposition \[prop:homotopic1\]. \[thm:homotopyequivalence\] Let $A$ and $B$ be two dg algebras that are strongly homotopy equivalent. Then there are quasi-equivalences of dg categories between 1. $\operatorname{MC_{dg}}(A)$ and $\operatorname{MC_{dg}}(B)$, 2. $\operatorname{Tw}(A)$ and $\operatorname{Tw}(B)$, 3. $\operatorname{Tw_{perf}}(A)$ and $\operatorname{Tw_{perf}}(B)$. \[eg:K\_0\] Assume that $2$ is not invertible in ${k}$ and consider the dg algebra $K_0^*$; recall that it is the path algebra of the quiver with $|s|=1$ with the differential being $\operatorname{ad}(s)$. It is clear that $s\in K^*_0$ is an MC element. It is easy to see that the $K^*_0$-module $K^{*[x]}_0$ is not isomorphic to $K^*_0$ as its first homology group is ${k}/2\neq 0$. It follows that $s$ is not homotopy gauge equivalent to zero and therefore by Theorem \[thm:homotopyequivalence\], $K^*_0$ is not $K_2$ homotopy equivalent to ${k}$. \[eg-polyderham\] Now let ${k}$ be a field of characteristic zero and consider $A:={k}[z,dz]$, the polynomial de Rham algebra of the line. This is quasi-isomorphic to ${k}$, and even polynomially homotopy equivalent to ${k}$, however $\operatorname{MC_{dg}}(A)$ is not quasi-equivalent to $\operatorname{MC_{dg}}({k})$ (and so, $A$ is not $K_2$ homotopy equivalent to ${k}$). To see this note that MC elements are just polynomial 1-forms and a map in $\operatorname{MC_{dg}}({k}[z,dz])$ between two such elements $x$ and $y$ is a polynomial $f$ satisfying $df + fx + yf = 0$. This differential equation will not usually have polynomial solutions, so different choices of $x$ and $y$ give a large number of MC elements in $A$ which do not map to one another (and thus represent non-isomorphic objects in $\operatorname{H^{0}}\operatorname{MC_{dg}}(A)$). The finitely generated twisted modules represented by these $\operatorname{MC}$ elements are examples of ${\mathcal{O}}$-coherent ${\mathcal{D}}$-modules with irregular singularities at infinity. Categories of twisted modules associated with simplicial sets {#simplicial} ============================================================= In this section we consider twisted modules over the dg pseudo-compact algebra $C^*(X)$, the normalized cochain complex of a simplicial set $X$. We have the dg categories $\operatorname{MC_{dg}}(C^*(X)), \operatorname{Tw}(C^*(X))$ and $\operatorname{Tw_{perf}}(C^*(X))$ that we will abbreviate to $\operatorname{MC_{dg}}(X), \operatorname{Tw}(X)$ and $\operatorname{Tw_{perf}}(X)$ respectively. These dg categories are not (up to quasi-equivalence) invariants of the weak homotopy type of $X$, however they are homotopy invariants of $X$ in a sense that we will now make precise. Recall cf. [@Goerss99] that two maps of simplicial sets $f,g:X{\rightarrow}Y$ are called homotopic if they can be extended to a map $X\times K_0{\rightarrow}Y$; recall that $K_0$ stands for the standard simplicial interval having two nondegenerate 0-simplices and one nondegenerate 1-simplex. This notion of homotopy is completely adequate only in the case where $Y$ is a Kan complex (in which case it is an equivalence relation). We will now introduce the notion of a *strong homotopy* of maps between simplicial sets and the concomitant notion of strong homotopy equivalence. Let $C$ be a fibrant cylinder object for the simplicial point. For example, we can take $C=I$ or $C=K_\infty$. Then $X\times C$ is a cylinder object for any simplicial set $X$; moreover it is *very good* in the sense that the natural projection $X\times C{\rightarrow}X$ is a fibration of simplicial sets. We will denote by $i_0, i_1:X{\rightarrow}X\times C$ the two natural inclusions of $X$ into $X\times C$. An elementary strong homotopy of maps of simplicial sets $f,g:X{\rightarrow}Y$ is a map $h:X\times C{\rightarrow}Y$ such that $h\circ i_0 =f$ and $h \circ i_1=g$. the maps $f$ and $g$ are called strongly homotopic if there is a chain of elementary homotopies connecting $f$ and $g$. Furthermore, $X$ and $Y$ are called strongly homotopy equivalent if there are maps $f:X{\rightarrow}Y$ and $g:Y{\rightarrow}X$ such that $f\circ g$ and $g\circ f$ are strongly homotopic to ${\mathbf 1}_Y$ and ${\mathbf 1}_X$ respectively. The relation of strong homotopy does not depend on the choice of a very good cylinder object $C$. Any such very good cylinder is strongly homotopy equivalent to the point. Let $P$ stand for the simplicial point. For any two very good cylinder objects $C$ and $C^{\prime}$ of $P$ consider the diagram $$\xymatrix{ P\coprod P\ar^{i_0\coprod i_1}[d]\ar^{i_0\coprod i_1}[r]&C^{\prime}\ar[d]\\ C\ar[r]&P. }$$ Since the left downward arrow is a monomorphism and thus a cofibration of simplicial sets, and the right downward arrow is a fibration ($C^{\prime}$ being fibrant), there exists a filler $C{\rightarrow}C^{\prime}$. It follows that any strong homotopy based on $C^{\prime}$ gives rise to a strong homotopy based on $C$. Symmetrically, any strong homotopy based on $C$ gives rise to a strong homotopy based on $C^{\prime}$; this proves the first claim of the proposition. The second claim follows from general theory of closed model categories: any very good cylinder object is weakly equivalent to the point; then, being a fibrant-cofibrant object it is homotopy equivalent to the point through any fixed good cylinder object, i.e. it is strongly homotopy equivalent to the point. Two natural candidates for $C$ are $K_\infty$ and $I$, the singular simplicial set of the unit interval $[0,1]$. The multiplication on $[0,1]$ turns $I$ into a simplicial monoid, and the multiplication map $I\times I{\rightarrow}I$ could be viewed as a homotopy between the identity map on $I$ and the map to the point. This is an explicit strong homotopy equivalence between $I$ and the point. \[prop:strongsimplicial\] Let $f,g:X{\rightarrow}Y$ be two maps of simplicial sets. If $f$ and $g$ are strongly homotopic, then the induced maps of pseudo-compact dg algebras $f^*,g^*:C^*(Y){\rightarrow}C^*(X)$ are strongly homotopic. If two simplicial sets $X$ and $Y$ are strongly homotopy equivalent, then the pseudo-compact dg algebras $C^*(X)$ and $C^*(Y)$ are strongly homotopy equivalent. The second statement of the proposition follows from the first. For the first, choosing $X\times I$ as a very good cylinder object for $X$, consider a homotopy $h:X\times I{\rightarrow}Y$ such that $h\circ i_0=f$ and $h\circ i_0=g$. This gives rise to a map of dg pseudo-compact algebras $C^*(Y){\rightarrow}C^*(X\times I)$ and, composing the latter with the Eilenberg-Zilber map $C^*(X\times I){\rightarrow}C^*(X)\otimes I^*$ (which is known to be a dg algebra map) we obtain the desired strong homotopy between $f^*$ and $g^*$. \[cor:equivalence\] Let $X$ and $Y$ be two strongly homotopy equivalent simplicial sets. Then the following dg categories are quasi-equivalent: 1. $\operatorname{MC_{dg}}(X)$ and $\operatorname{MC_{dg}}(Y)$, 2. $\operatorname{Tw}(X)$ and $\operatorname{Tw}(Y)$, 3. $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw_{perf}}(Y)$. This is a direct consequence of Theorem \[thm:homotopyequivalence\] and Proposition \[prop:strongsimplicial\]. \[cor:weaklyKan\] Let $X$ and $Y$ be two weakly equivalent Kan simplicial sets. Then the following dg categories are quasi-equivalent: 1. $\operatorname{MC_{dg}}(X)$ and $\operatorname{MC_{dg}}(Y)$, 2. $\operatorname{Tw}(X)$ and $\operatorname{Tw}(Y)$, 3. $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw_{perf}}(Y)$. Two weakly equivalent Kan simplicial sets are homotopy equivalent through any given very good cylinder object. In particular, they are strongly homotopy equivalent. The conclusion then follows from Corollary \[cor:equivalence\]. This also has a consequence for singular cochain algebras of topological spaces. \[cor:mctopologicalequivalence\] Let $X$ and $Y$ be weakly equivalent topological spaces. Then there are quasi-equivalences of dg categories between 1. $\operatorname{MC_{dg}}(X)$ and $\operatorname{MC_{dg}}(Y)$, 2. $\operatorname{Tw}(X)$ and $\operatorname{Tw}(Y)$, 3. $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw_{perf}}(Y)$. In particular, if $X$ is a contractible topological space, then the dg categories $\operatorname{MC_{dg}}(X)$, $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw}(X)$ are quasi-equivalent to the category of free ${k}$-modules of rank 1, the category of finitely generated free dg ${k}$-modules and the category of all free dg ${k}$-modules, respectively. Since the topological spaces $X$ and $Y$ are weakly equivalent, so are their singular simplicial sets. Since the latter are Kan complexes, the claim follows from Corollary \[cor:weaklyKan\]. Reduced and minimal twisted modules ----------------------------------- Let $A$ be a non-negatively graded pseudo-compact algebra, such as $C^*(X)$ for a simplicial set $X$, and $M:=V\otimes A$ be a twisted $A$-module. The differential $D_M$ on $M$ is determined by its restriction on $V$; furthermore we have: $D_M|_{V\otimes 1}=d^0+d^1+\ldots$ where $d^n:V{\rightarrow}V\otimes A^n$. In particular, $d^0:V{\rightarrow}V\otimes A^0$ determines an $A^0$-linear differential on $V\otimes A^0$. A twisted $A$-module $M$ as above is called *reduced* if $d_0$ factors through $V\hookrightarrow V\otimes A^{0}: v\mapsto v\otimes 1$, i.e. if it is induced by a differential in the graded ${k}$-module $V$. If, further, $d_0=0$, we will call the twisted $A$-module $M$ *minimal*. We will denote by $\operatorname{Tw^{red}}(A)$, $\operatorname{{Tw}^m}(A)$, $\operatorname{Tw_{perf}^m}(A)$, and $\operatorname{Tw_{perf}^{red}}(A)$ the categories of reduced, minimal twisted $A$-modules and their perfect versions respectively. If $A=C^*(X)$ for a simplicial set $X$, we will denote these categories by $\operatorname{Tw^{red}}(X)$, $\operatorname{{Tw}^m}(X)$, $\operatorname{Tw_{perf}^m}(X)$ and $\operatorname{Tw_{perf}^{red}}(X)$ respectively. If $A$, in addition to being non-negatively graded, is *connected* i.e. $A^0={k}$, then clearly any twisted $A$-module is reduced. Such is the case, when $A=C^*(X)$ for a reduced simplicial set $X$. In this situation, twisted $A$-modules could be viewed as cofibrant objects in a certain closed model category of $A$-contramodules, cf. [@Positselski11 Theorem 8.2(b)]; then $\operatorname{H^{0}}(\operatorname{Tw}(A))$ is the corresponding derived category of 2nd kind. The notion of a minimal twisted module is similar to that of a minimal $A_\infty$-module, [@Keller01]; indeed in the case when $A$ is a completed tensor algebra representing an $A_\infty$ algebra, then a minimal twisted $A$-module is a contramodule corresponding to a minimal $A_\infty$-module under the comodule-contramodule correspondence, cf. [@Positselski11 Theorem 5.2]. A homotopy equivalence between two minimal twisted modules is necessarily an isomorphism. It suffices to show that any endomorphism of a minimal twisted module that is homotopic to the identity is invertible. Let $V\otimes A$ be such a minimal $A$-module; then its dg algebra of endomorphisms is $A\otimes\operatorname{End}(V)$; by minimality the differential in it has the form $D_A=d_A^1+d_A^2+\ldots$ where $$d_A^n|_{V\otimes 1}: V{\rightarrow}V\otimes A^n.$$ Let $f\in A\otimes\operatorname{End}(V)$ be a closed endomorphism homotopic to the identity; thus $f=1+D_A(g)$ for some $g\in A\otimes\operatorname{End}(V)$. Then $D_A(g)$ must have the form $D_A(g)=d_A^1(g)+d_A^2(g)+\ldots$ with $d_A^n(g)\in A^n\otimes\operatorname{End}(V)$ and therefore $f$ is invertible: $f^{-1}=1+\sum_{i=1}^{\infty}(-1)^i(\sum_{n=1}^\infty d_A^i(g))$. The following result is analogous to the well-known theorem on minimal $A_\infty$ modules [@Keller01]. \[prop:minimal\] If ${k}$ is a field then any reduced twisted $A$-module is homotopy equivalent to a minimal one. Let $A\otimes V$ be a reduced twisted $A$-module; it has differential $d^0+d^\prime:=d^0+d^1+\ldots$ as described above and $d^0$ makes $V$ into a dg ${k}$-vector space. Since ${k}$ is a field, $V$ admits a decomposition $V\cong H(V)\oplus d_0(V)\oplus U$ with $d^0$ mapping $U$ isomorphically onto $V$. Denote by $t:V{\rightarrow}V$ the projection onto $H(V)$ and by $s:V{\rightarrow}V$ the operator that is inverse to $d^0$ on $d^0(V)$ (viewed as an operator $U{\rightarrow}d^0(V)$) and whose restriction on $H(V)$ and $U$ is zero. The pair of operators $(s,t)$ determines an abstract Hodge decomposition on $V$ (cf. for example [@Chuang17] concerning this notion) and we can apply the Perturbation Lemma as formulated in [@Chuang17 Corollary 3.17]. Namely, the twisted module $A\otimes V$ is isomorphic to the direct sum of $M_1:=A\otimes(d^0(V)\oplus U)$ and $M_2:=A\otimes H(V)$ where $M_1$ is supplied with the differential ${\mathbf 1}_A\otimes d^0$ and $M_2$ with the differential $td^\prime(1+sd^\prime)^{-1}t$. Since $M_1$ is clearly homotopy equivalent to zero and $M_2$ is minimal, the claim follows. Then we have the following result. \[cor:minimal\] If $A$ and $B$ are two non-negatively graded pseudo-compact dg algebras that are strongly homotopy equivalent. Then the following categories are quasi-equivalent: 1. $\operatorname{{Tw}^m}(A)$ and $\operatorname{{Tw}^m}(B)$, 2. $\operatorname{Tw_{perf}^m}(A)$ and $\operatorname{Tw_{perf}^m}(B)$. Let $A$ be a connected pseudo-compact dg algebra over a field. Then the following dg categories are quasi-equivalent: 1. $\operatorname{Tw}(A)$ and $\operatorname{{Tw}^m}(A)$, 2. $\operatorname{Tw_{perf}}(A)$ and $\operatorname{Tw_{perf}^m}(A)$. Let $f:A{\rightarrow}B$ and $g:B{\rightarrow}A$ be two maps that are inverse up to $K_2$-homotopy. These maps induce dg functors $\operatorname{Tw}(f):\operatorname{Tw}(A){\rightarrow}\operatorname{Tw}(B)$ and $\operatorname{Tw}(g):\operatorname{Tw}(B){\rightarrow}\operatorname{Tw}(A)$ inducing quasi-equivalence of the corresponding dg categories by (the pseudo-compact analogue of) Theorem \[thm:homotopyequivalence\]. These functors restrict to the categories of minimal twisted modules and, using Proposition \[prop:minimal\] we see, that these restrictions give mutually inverse quasi-equivalences. The statement about perfect minimal twisted modules is proved similarly. Finally, if $A$ is connected, any twisted $A$-module is automatically reduced, and the proof is finished by appealing to Proposition \[prop:minimal\]. \[thm:connected\] Let $X$ be a connected Kan simplicial set. Then the pseudo-compact dg algebra $C^*(X)$ is strongly homotopy equivalent to a connected one. Choosing a vertex of $X$ amounts to constructing a map $P{\rightarrow}X$ from the one-point simplicial set $P$ to $X$. Let $X^\prime$ be the simplicial set defined by the pullback diagram $$\xymatrix{ X^\prime\ar^f[r]\ar[d]&X\ar[d]\\ P\ar[r]&\operatorname{cosk_0}(X). }$$ Here $\operatorname{cosk_0}(X')$ is the 0-coskeleton (the zeroth stage of the Moore-Postnikov tower of $X$). The simplicial set $X^\prime$ has a single vertex corresponding to the map $P{\rightarrow}X$ and so $C^*(X)$ is a connected pseudo-compact dg algebra. It is well-known (e.g. [@May67 Proposition 8.2, Theorem 8.4]) that $\operatorname{cosk_0}(X)$ is a weakly contractible Kan simplicial set, and it follows that $X^\prime$ is likewise Kan. Then $f:X^\prime{\rightarrow}X$ is a strong homotopy equivalence with a strong homotopy inverse $g:X{\rightarrow}X^\prime$ (in fact it is clear that $X^\prime$ is a deformation retract of $X$ so that $g\circ f={\mathbf 1}_{X^\prime}$). By Proposition \[prop:strongsimplicial\] $C^*(X)$ and $C^*(X^\prime)$ are strongly homotopy equivalent. Combining Corollary \[cor:minimal\] and Theorem \[thm:connected\] we obtain the following result. \[cor-twistedreduced\] 1. Let $X$ be a Kan simplicial set. Then there is a quasi-equivalence between the following dg categories: 1. $\operatorname{Tw}(X)$ and $\operatorname{Tw^{red}}(X)$, 2. $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw_{perf}^{red}}(X)$. 2. If ${k}$ is a field then, additionally, the following dg categories are quasi-equivalent: 1. $\operatorname{Tw}(X)$ and $\operatorname{{Tw}^m}(X)$, 2. $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw_{perf}^m}(X)$. 3. If $X,X^\prime$ are weakly equivalent Kan simplicial sets, then the following dg categories are quasi-equivalent: 1. $\operatorname{{Tw}^m}(X)$ and $\operatorname{{Tw}^m}(X^\prime)$, 2. $\operatorname{Tw_{perf}}(X)$ and $\operatorname{Tw_{perf}^m}(X^\prime)$. If $X$ is a reduced simplicial set and ${k}$ is a field then $C^*(X)$ is a local pseudo-compact dg algebra (which is the dual to a conilpotent dg coalgebra $C_*(X)$). The category of local pseudo-compact dg algebras admits the structure of a closed model category, see [@Positselski11] where this result is formulated in the dual language of coalgebras. It makes sense to ask whether weakly (or strongly) homotopy equivalent reduced simplicial sets give rise to weakly equivalent (in the sense of the aforementioned closed model category) local pseudo-compact dg algebras. A partial answer to this question could be extracted from the recent paper [@Zeinalia16] where it is proved that if $X$ is a singular simplicial set of a topological space (or, more generally, a Kan simplicial set) then the cobar-construction of $C_*(X)$ is quasi-isomorphic to the dg algebra of chains on the loop space of $X$. Note that this generalizes the classical result of Adams on the cobar-construction [@Adams56] in that simple connectivity of $X$ is not assumed. This result implies that for two weakly equivalent *Kan* simplicial sets $X$ and $X^\prime$ the pseudo-compact local dg algebras $C^*(X)$ and $C^*(X^\prime)$ are indeed weakly equivalent. The Kan condition is essential; e.g. taking for $X$ a simplicial circle with one non-degenerate simplex in degree zero and another in degree one (which is not a Kan simplicial set), a straightforward inspection shows that the cobar-construction of $C_*(X)$ is isomorphic to ${k}[x]$ with $|x|=0$ whereas the singular chain algebra on $\Omega(S^1)=\mathbb Z$ is ${k}[x,x^{-1}]\neq {k}[x]$. Twisted modules and sheaves {#sheaves} =========================== Generalities on dg sheaves -------------------------- In this section we will present the local to global arguments needed to apply Schlessinger-Stasheff type results to infinity local systems. The results obtained here are directly used in Section \[sect-derham\] and Section \[sect-dolbeault\] and the methods of proof are used in Section \[sect-singular\]. Let $X$ be a topological space, always assumed paracompact and Hausdorff in this section. Let ${\mathcal{R}}$ be a sheaf of $k$-algebras on $X$ and let ${\mathcal{A}}= ({\mathcal{A}}^{\bullet}, d)$ be a sheaf of non-negatively graded dg ${\mathcal{R}}$-algebras. Write ${\mathcal{A}}^{\#}$ for $({\mathcal{A}}^{\bullet}, 0)$. We will consider the dg algebra $A \coloneqq {\mathcal{A}}(X)$. Write $A{\textrm{-Mod}}$ for the dg category of dg $A$-modules and ${\mathcal{A}}{\textrm{-Mod}}$ for the dg category of sheaves of dg modules over ${\mathcal{A}}$. Write ${\mathcal{R}}{\textrm{-Mod}}$ for the dg category of sheaves of dg ${\mathcal{R}}$-modules. There is an adjunction $p^{*}\dashv p_{*}$ between dg modules over ${\mathcal{A}}$ and dg modules over $A$ which is induced by the map $p: (X,{\mathcal{A}}) {\rightarrow}(*, A)$ of dg ringed spaces. For free modules we recall the following straightforward result: \[lemma-protoswan\] The adjunction $p^{*} \dashv p_{*}$ induces a dg equivalence between dg $A$-modules and dg ${\mathcal{A}}$-modules whose underlying $A^{\#}$-modules, respectively ${\mathcal{A}}^{\#}$-modules, are free. We first forget the differential and the grading and consider a ringed space $(X, {\mathcal{R}})$ and let $R = {\mathcal{R}}(X)$. Let $q$ be the map $(X, {\mathcal{R}}) {\rightarrow}(*, R)$. Then $q^{*}$ induces an equivalence between free $R$-modules and free ${\mathcal{R}}$-modules: Any free ${\mathcal{R}}$-module is of the form $q^{*}V$ for a free $R$-module $V$ and $q_{*}q^{*}(W) = (q^{*}W)(X) = W$, which gives the isomorphism $\operatorname{Hom}_{{\mathcal{R}}}(q^{*}V, q^{*}W) \cong \operatorname{Hom}_{R}(V, W)$. Thus the unit and counit of the adjunction are isomorphisms. They are also compatible with the grading and the differential, hence they are isomorphisms of dg modules, resp. dg sheaves, proving the lemma. For a fine sheaf ${\mathcal{R}}$ we can compare categories of locally free sheaves and projective modules over the ring of global sections. Recall that a sheaf ${\mathcal{F}}$ is *fine* if for any locally finite open cover $\{U_{i}\}$ of $X$ there is family of morphisms $\phi_{i}: {\mathcal{F}} {\rightarrow}{\mathcal{F}}$ such that $\sum \phi_{i} = {\mathbf 1}_{{\mathcal{F}}}$ and such that $\phi_{i}$ has support contained in $U_{i}$. On a paracompact Hausdorff space fine sheaves are always soft and thus ${\Gamma}$-acyclic. A module over a fine sheaf of rings is automatically fine. For more details see e.g. [@Wells07 Section II.3]. The following is Swan’s theorem as stated in [@Morye09 Corollary 3.2]. \[thm-swan\] Assume $(X, {\mathcal{R}})$ is a locally ringed space of finite covering dimension with ${\mathcal{R}}$ a fine sheaf of commutative algebras. Then the category of finitely generated projective ${\mathcal{R}}(X)$-modules is equivalent to the category of locally free ${\mathcal{R}}$-modules of bounded rank. We note that Lemma \[lemma-protoswan\] gives an equivalence of dg categories, but the two sides have a priori very different homotopy theories: For dg sheaves the natural class of weak equivalences is given by local quasi-isomorphisms, i.e. morphisms which restrict to quasi-isomorphisms on all stalks. To make this more precise we recall that the categories ${\mathcal{R}}{\textrm{-Mod}}^{psh}$ and ${\mathcal{A}}{\textrm{-Mod}}^{psh}$ of presheaves of dg ${\mathcal{R}}$-modules, respectively presheaves of dg ${\mathcal{A}}$-modules, have model structures. To define this, first recall the definition of a *hypercover*. A hypercover of a presheaf $P$ on a topological space $X$ is an augmented simplicial presheaf $C_{\bullet} {\rightarrow}P$ such that all $C_{n}$ are coproducts of representable presheaves and for all $n$ the map $C_{n} {\rightarrow}\operatorname{cosk}_{n-1}\operatorname{sk}_{n-1}C$ is a cover. In particular a hypercover of $X$ is defined to be a hypercover of the presheaf $h_{X}$ that $X$ represents, and it may be represented by a cover $\mathfrak U_{0} {\rightarrow}X$ together with covers $\mathfrak U_{n} {\rightarrow}\lim_{k < n} \mathfrak U_{k}$. The fundamental example of a hypercover is the nerve of a Čech cover. In this case all the covers (except for $\mathfrak U_{0} {\rightarrow}X$) are isomorphisms. The *projective model structure* on presheaves of dg ${\mathcal{R}}$-modules has fibrations and weak equivalences defined object-wise. The *local model structure* on presheaves of dg ${\mathcal{R}}$-modules is the localization of the projective model structure at all hypercovers. Then weak equivalences are given by maps inducing weak equivalences on stalks. We will abuse notations and refer to them as quasi-isomorphisms. An object $P$ is fibrant if it is a *hypersheaf*, i.e. for any open subset $U \subset X$ and hypercover $\mathfrak U_{\bullet} {\rightarrow}U$ there is a quasi-isomorphism $P(U) \simeq \check C(\mathfrak U_{\bullet}, P)$. Here the right hand side is the Čech complex of a hypercover, defined exactly like the Čech complex for a cover. We say a hypercover is *contractible* if every $\mathfrak U_{n}$ is a coproduct of contractible open sets. Any locally contractible topological space admits a contractible hypercover. We will in the following sometimes compute Čech cohomology with respect to a hypercover, but not much generality is lost if the reader wants to mentally replace them by Čech covers. The local model structure on presheaves of dg ${\mathcal{A}}$-modules is defined in the same way (or it can be transferred via the forgetful functor, see e.g. [@Goerss06]). The homotopy category of this model category is the usual derived category of ${\mathcal{R}}{\textrm{-Mod}}$, respectively ${\mathcal{A}}{\textrm{-Mod}}$. The adjunction $p_{*} \dashv p^{*}$ is Quillen. For more details on the local model structure see [@Dugger04a] (in the case of simplicial presheaves) and [@Choudhury15] (for chain complexes). We now assume ${\mathcal{R}} {\rightarrow}{\mathcal{A}}$ is an object-wise quasi-isomorphism, i.e. in particular a local weak equivalence. Writing $J$ for the forgetful functor we have a dg adjunction $-\otimes_{{\mathcal{R}}}{\mathcal{A}}\dashv J: {\mathcal{R}}{\textrm{-Mod}}\rightleftarrows {\mathcal{A}}{\textrm{-Mod}}$. We consider the derived categories of ${\mathcal{R}}{\textrm{-Mod}}$ and ${\mathcal{A}}{\textrm{-Mod}}$ and write $RJ$ for the total derived functor of $J$, i.e. the lift of $J$ to the derived categories. As ${\mathcal{A}}$ and ${\mathcal{R}}$ are quasi-isomorphic $RU$ is an equivalence. \[defn-F\] Let $F = J \circ p^{*}: A{\textrm{-Mod}}{\rightarrow}{\mathcal{R}}{\textrm{-Mod}}$ be the dg functor given by composition of the two functors defined above. We will use $F$ to map different categories of twisted modules to dg sheaves. In the remainder of this paper we will abuse notation and write $F$ for different choices of $A$ as well as for the restriction of $F$ to $\operatorname{Tw}(A)$, $\operatorname{Tw_{perf}}(A)$ and $\operatorname{Tw_{fg}}(A)$. Twisted modules and perfect complexes {#sect-cohesivesheaves} ------------------------------------- We now consider the functor $F: A{\textrm{-Mod}}{\rightarrow}{\mathcal{R}}{\textrm{-Mod}}$ in more detail. We let $D(X, {\mathcal{R}})$ or simply $D({\mathcal{R}})$ be the derived category of ${\mathcal{R}}{\textrm{-Mod}}$. In this section we will assume $X$ is *locally good*, which is defined as follows. We say a ringed space $(U, {\mathcal{R}})$ is *good* if the natural map $\mathcal{R}(U){\rightarrow}Ru_{*}{\mathcal{R}}|_{U}$ is a quasi-isomorphism (here $u:U{\rightarrow}*$ is the map to the one-point space). Then $X$ is locally good if its topology has a basis of good open sets. Most spaces of interest are locally good, for example algebraic schemes, analytic spaces and locally contractible topological spaces with the constant structure sheaf. Good neighbourhoods are given by affine subvarieties, Stein subspaces and contractible subsets respectively. A dg ${\mathcal{R}}$-module on $X$ is *strictly perfect* if it is bounded and a direct summand of a free sheaf of finite rank in each degree. A dg ${\mathcal{R}}$-module $G$ is *perfect* if for every $x \in X$ there is a neighbourhood $U$ such that $G|_{U}$ is quasi-isomorphic to a strictly perfect dg sheaf. We denote by ${D_{\operatorname{perf}}}(X, {\mathcal{R}})$ or ${D_{\operatorname{perf}}}({\mathcal{R}})$ the subcategory of $D(X, {\mathcal{R}})$ consisting of perfect dg sheaves of ${\mathcal{R}}$-modules. We will say a perfect dg sheaf of ${\mathcal{R}}$-modules is *globally bounded* if there are integers $a, b$ and $N$ such that there is a cover $\{U_{i}\}$ such that each $G|_{U_{i}}$ is quasi-isomorphic to a strictly perfect dg sheaf $G^{U}$ which is concentrated in degrees $[a, b]$ and has at most $N$ generators. We let ${D_{\operatorname{perf}}}^{B}({\mathcal{R}})$ denote the subcategory of globally bounded perfect dg sheaves. In many cases of interest all perfect dg sheaves are globally bounded. An example of a non-globally bounded one is given by the following construction. Consider ${\mathbb{C}}$ equipped with the holomorphic (or smooth) structure sheaf. Then the skyscraper sheaf ${\mathbb{C}}_{n}$ at $n \in {\mathbb{C}}$ is perfect and so is $\oplus_{n \in {\mathbb{N}}} {\mathbb{C}}_{n}^{\oplus n}$. But this sum is clearly not globally bounded. For later use we also define ${D_{\operatorname{lf}}}(X, {\mathcal{R}})$ or ${D_{\operatorname{lf}}}({\mathcal{R}})$ to be the subcategory of $D({\mathcal{R}})$ consisting of locally free dg sheaves of ${\mathcal{R}}$-modules, i.e. those which are locally quasi-isomorphic to free ${\mathcal{R}}$-modules without any finiteness assumptions. In the case ${\mathcal{R}}=\underline{{k}}$, the locally free dg sheaves of ${\mathcal{R}}$-modules will be referred to as cohomologically locally constant (clc) sheaves. We will need the following: \[lemma-perfectkaroubi\] ${D_{\operatorname{lf}}}({\mathcal{R}})$, ${D_{\operatorname{perf}}}({\mathcal{R}})$ and ${D_{\operatorname{perf}}}^{B}({\mathcal{R}})$ are idempotent complete. The result for ${D_{\operatorname{lf}}}({\mathcal{R}})$ follows from [@Bokstedt93]. Next, recall that for any ring perfect dg modules are exactly compact objects in the derived category, and since compact objects are closed under direct summands so are perfect dg sheaves, see e.g.[@Bokstedt93 Proposition 6.4]. We consider a perfect dg sheaf of the form $G \simeq M \oplus N$ and will now show $M$ is perfect. It follows from the definition that any point has a neighbourhood $U$ on which we may assume $G$ is strictly perfect. Then the restriction $G|_{U}$ is isomorphic to the sheaf associated to $G(U)$ (apply Lemma \[lemma-protoswan\] restricted to finitely generated modules and extended to the idempotent completion). We write $u: U {\rightarrow}*$. Then $G|_{U} \cong u^{*}G(U)\simeq Lu^{*}G(U)$ as $G(U)$ is cofibrant. We may assume $U$ is good and then $G(U) \simeq Ru_{*}G \simeq Ru_{*}M\oplus Ru_{*}N$. $G(U)$ is perfect and thus so is $Ru_{*}M$. As $G \simeq Lu^{*}G(U)$ it follows that $M \oplus N \simeq Lu^{*}Ru_{*}M \oplus Lu^{*}Ru_{*}N$. Since the map $M {\rightarrow}Lu^{*}Ru_{*}N$ in the derived category corresponds to the zero map $Ru_{*}M {\rightarrow}Ru_{*}N$ under an adjunction it is zero and $M \simeq Lu^{*}Ru_{*}M$. Thus $M$ is perfect. The same argument applies to globally bounded dg sheaves. \[prop-cohesivesheaves\] Let $(X, {\mathcal{R}})$ be a ringed space, and let ${\mathcal{A}}$ be a fine sheaf of dg algebras on $X$ such that there is a quasi-isomorphism ${\mathcal{R}} {\rightarrow}{\mathcal{A}}$ satisfying condition (\*) below. Then the associated sheaf functor $p^{*}$ gives a quasi-fully faithful functor $\operatorname{Tw_{perf}}(A) {\rightarrow}{\mathcal{A}}{\textrm{-Mod}}$ and $F = J\circ p^{*}$ induces an embedding of triangulated categories $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A)) {\rightarrow}{D_{\operatorname{perf}}}^{B}(X, {\mathcal{R}})$. Proposition \[prop-cohesivesheaves\] only depends on the construction of $\operatorname{Tw_{perf}}(A)$ up to quasi-equivalence and thus holds equally if we consider $A$ as a dg AM algebra, see Remark \[rk:twfgam\]. The crucial assumption is the following - For every free graded $k$-module $G$, every $x \in X$ with a neighbourhood $U'$ and any MC element $\xi \in \operatorname{MC}({\mathcal{A}}(U')\otimes \operatorname{\underline{End}}(G))$ there is a neighbourhood $x\in U \subset U'$ such that $\xi|_{U}$ is homotopy gauge equivalent to an element in the image of the Maurer Cartan set of ${\mathcal{R}}(U)\otimes \operatorname{\underline{End}}(G)$. We will be particularly interested in cases where (\*) is the consequence of the following stronger condition: - For every $x \in X$ and every free graded $k$-module $G$ there is a neighbourhood $U$ such that ${\mathcal{R}} {\rightarrow}{\mathcal{A}}$ induces a quasi-equivalence $\operatorname{MC_{dg}}({\mathcal{R}}(U)\otimes \operatorname{\underline{End}}(G)) \simeq \operatorname{MC_{dg}}({\mathcal{A}}(U)\otimes \operatorname{\underline{End}}(G))$. By Lemma \[lemma-protoswan\] the restriction $p^{*}: {\mathcal{\operatorname{Tw_{fg}}}}(A) {\rightarrow}{\mathcal{A}}{\textrm{-Mod}}$ is quasi-fully faithful. Then $p^{*}$ on $\operatorname{Tw_{perf}}(A)$ is just the extension of $p^{*}|_{\operatorname{Tw_{fg}}(A)}$ to homotopy idempotents and it follows that $p^{*}$ is also quasi-fully faithful. It remains to prove the statement on homotopy categories. We have the following composition $$\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A)) \stackrel {p^{*}}\longrightarrow \operatorname{H^{0}}({\mathcal{A}}{\textrm{-Mod}}) \stackrel {q_{{\mathcal{A}}}} \longrightarrow D({\mathcal{A}}) \stackrel {RJ} \longrightarrow D({\mathcal{R}})$$ where $q_{{\mathcal{A}}}$ is the quotient by quasi-isomorphisms. In Lemma \[lemma-fullyfaithful\] we will show that $q_{{\mathcal{A}}}$ is fully faithful on the image of $p^{*}$. It is well-known that $RJ$ is fully faithful. Thus $RJ \circ q_{{\mathcal{A}}} \circ \operatorname{H^{0}}(p^{*})$ is fully faithful. It is clearly compatible with shifts and cones. The fact that $RJ$ lands in ${D_{\operatorname{perf}}}({\mathcal{R}})$ is Lemma \[lemma-mc\]. \[lemma-mc\] The dg functor $F$ sends perfect twisted $A$-modules to globally bounded perfect sheaves of ${\mathcal{R}}$-modules. By Lemma \[lemma-perfectkaroubi\] it suffices to show that a finitely generated twisted $A$-module $E$ is sent to a globally bounded perfect dg sheaf. We may write $E = (G\otimes A, D)$ where $G$ is a free graded module over $k$. It suffices to show that $F(E) \cong (\underline G \otimes {\mathcal{A}}, D)$ is perfect locally. On any $U$ we know that $D|_{U} - {\mathbf 1}\otimes d_{A(U)}$ can be represented by a MC element $\xi$ in $ {\mathcal{A}}(U) \otimes \operatorname{\underline{End}}(G)$. Fix some $x$. For a suitably small neighbourhood we may assume that $\xi$ is as in in condition $(*)$. Thus there is a homotopy gauge equivalence $g \in {\mathcal{A}}(U)\otimes \operatorname{\underline{End}}(G)$ between $\xi$ and some element $\eta$ in the image of ${\mathcal{R}}(U)\otimes\operatorname{\underline{End}}(G)$. It follows that $g$ gives a homotopy equivalence from $(G \otimes {\mathcal{A}}(U), D_{U})$ to $(G \otimes {\mathcal{A}}(U), d_{G}\otimes {\mathbf 1}+ {\mathbf 1}\otimes d_{{\mathcal{A}}})$ where $d_{G}$ is some differential on $G \otimes {\mathcal{R}}(U)$. Thus we obtain a quasi-isomorphism from $F(E)|_{U}$ to the perfect dg sheaf $(\underline G\otimes {\mathcal{R}}|_{U}, d_{G})$ of ${\mathcal{R}}|_{U}$-modules. Boundedness follows immediately from finite generation of $E$. \[lemma-fullyfaithful\] The natural functor $q_{{\mathcal{A}}}: \operatorname{H^{0}}({\mathcal{A}}{\textrm{-Mod}}){\rightarrow}D({\mathcal{A}})$ is fully faithful when restricted to the image of perfect twisted modules. It suffices to consider finitely generated twisted modules, so we fix $(V \otimes A, D_{V})$, $ (W \otimes A, D_{W}) \in {\mathcal{\operatorname{Tw_{fg}}}}(A)$ and compute $R\operatorname{\underline{Hom}}_{D({\mathcal{A}})}(\underline V\otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}})$. The derived Hom can be computed as derived global sections of the sheaf Hom $U \mapsto R\operatorname{\underline{Hom}}_{{\mathcal{A}}|_{U}}((\underline V\otimes {\mathcal{A}})|_{U}, (\underline W \otimes {\mathcal{A}})|_{U})$. We first compute locally. We write ${\mathcal{V}} = \underline V\otimes {\mathcal{R}}$ and ${\mathcal{W}} = \underline W \otimes {\mathcal{R}}$. Then let $U$ be any good open set as in condition (\*), such that $(\underline V\otimes {\mathcal{A}})|_{U}$ is homotopy equivalent to $({\mathcal{V}}|_{U}, d_V)\otimes_{{\mathcal{R}}|_{U}} {\mathcal{A}}|_{U}$, say. Then we can compute: $$\begin{aligned} R\operatorname{\underline{Hom}}_{{\mathcal{A}}|_{U}}((\underline V\otimes {\mathcal{A}})|_{U}, (\underline W \otimes {\mathcal{A}})|_{U}) & \simeq R\operatorname{\underline{Hom}}_{{\mathcal{R}}|_{U}}(({\mathcal{V}}|_{U}, d_V), (\underline W\otimes {\mathcal{A}})|_{U}) \\ & \simeq \operatorname{\underline{Hom}}_{{\mathcal{R}}|_{U}}(({\mathcal{V}}|_{U}, d_V), (\underline W\otimes {\mathcal{A}})|_{U}).\end{aligned}$$ As $({\mathcal{V}}|_{U}, d_{V})$ is free it is a cofibrant dg sheaf over ${\mathcal{R}}$ and the Hom space is underived. To compute global sections we pick a hypercover $\mathfrak U$ consisting of good open sets $U$ satisfying condition (\*). By the above the Hom presheaf on $\mathfrak U$ may be written as $U \mapsto \operatorname{\underline{Hom}}_{{\mathcal{R}}|_{U}}({\mathcal{V}}, {\mathcal{W}})\otimes_{{\mathcal{R}}(U)} {\mathcal{{\mathcal{A}}}}(U)$ with a suitable differential. We compute Čech cohomology. Since ${\mathcal{A}}^{\#}$ is fine each $(\operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{R}}}({\mathcal{V}}, {\mathcal{W}}) \otimes_{{\mathcal{R}}} {\mathcal{A}})^{i}$ has no higher cohomology. We filter the map $\epsilon: \check C^{*}(\mathfrak U, \operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{R}}}({\mathcal{V}}, {\mathcal{W}}) \otimes_{{\mathcal{R}}} {\mathcal{A}}) {\rightarrow}\check C^{0}(\mathfrak U, \operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{R}}}({\mathcal{V}},{\mathcal{W}})\otimes_{{\mathcal{R}}} {\mathcal{A}})$ by the degree of coefficients. The associated map $\operatorname{Gr}(\epsilon)$ is a quasi-isomorphism and, since the filtration is exhaustive and Hausdorff, $\epsilon$ is a quasi-isomorphism, too. Putting all of this together we have: $$\begin{aligned} R\operatorname{\underline{Hom}}_{D({\mathcal{A}})}(\underline V \otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}}) &\simeq \check C^{*}(\mathfrak U, R\operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{A}}}(\underline V\otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}})) \\ &\simeq \check C^{*}(\mathfrak U, \operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{R}}}({\mathcal{V}}, \underline W\otimes {\mathcal{A}})) \\ &\simeq \check C^{*}(\mathfrak U, \operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{R}}}({\mathcal{V}}, {\mathcal{W}}) \otimes_{{\mathcal{R}}} {\mathcal{A}}) \\ &\simeq \check C^{0}(\mathfrak U, \operatorname{{\mathcal{H}} \! \it{om}}_{{\mathcal{R}}}({\mathcal{V}}, {\mathcal{W}}) \otimes_{{\mathcal{R}}} {\mathcal{A}}) \\ &\simeq \operatorname{\underline{Hom}}_{{\mathcal{A}}{\textrm{-Mod}}}(\underline V\otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}}). \qedhere\end{aligned}$$ \[cor-quasifullyfaithful\] If ${\mathcal{A}}$ is flat over ${\mathcal{R}}$ then $F$ is quasi-fully faithful. We use the notation from the proof of Lemma \[lemma-fullyfaithful\]. It suffices to compare $\operatorname{\underline{Hom}}_{{\mathcal{R}}}(\underline V\otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}})$ and $\operatorname{\underline{Hom}}_{{\mathcal{A}}}(\underline V\otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}})$. Locally on $U$ the terms are quasi-isomorphic to $\operatorname{\underline{Hom}}_{{\mathcal{R}}}(({\mathcal{V}}, d_{V})\otimes {\mathcal{A}}, \underline W\otimes {\mathcal{A}})$ and $\operatorname{\underline{Hom}}_{{\mathcal{R}}}(({\mathcal{V}}, d_{V}), \underline W\otimes {\mathcal{A}})$, respectively. If ${\mathcal{A}}$ is flat over ${\mathcal{R}}$ they are quasi-isomorphic. The local-to-global argument remains unchanged. \[lemma-surjective\] Let $(X, {\mathcal{R}})$ and ${\mathcal{A}}$ be as in Proposition \[prop-cohesivesheaves\]. If moreover $(X, {\mathcal{A}}^{0})$ is locally ringed, ${\mathcal{A}}^{0}$ is commutative, $A$ is flat over $A^{0}$ and $X$ has finite covering dimension then ${D_{\operatorname{perf}}}(X, {\mathcal{R}})$ lies in the image of $\operatorname{H^{0}}(F)$. Consider a globally bounded perfect dg sheaf ${\mathcal{V}}$ of ${\mathcal{R}}$-modules on $X$. Then ${\mathcal{V}}\otimes_{{\mathcal{R}}} {\mathcal{A}}^{0}$ is a perfect dg sheaf of ${\mathcal{A}}^{0}$-modules. Let ${\Gamma}_{{\mathcal{V}}} \coloneqq ({\Gamma}(X, {{\mathcal{V}}}\otimes_{{\mathcal{R}}} {\mathcal{A}}), D_{{{\mathcal{V}}}}\otimes {\mathbf 1}+ {\mathbf 1}\otimes d_{A})$. This is a dg sheaf of $A$-modules which is not necessarily a perfect twisted module. However, it is a dg $A$-module of the form $Q\otimes_{A^{0}} A$ where $Q$ is some dg $A^{0}$-module. Such objects are called *quasi-cohesive modules* in [@Block10]. By Lemma 2.3 and [@Morye09 Proposition 2.5] the associated sheaf functor will send ${\Gamma}_{{\mathcal{V}}}$ to ${{\mathcal{V}}}$. We will show that ${\Gamma}_{{\mathcal{V}}}$ is homotopy equivalent to a cohesive module, which by \[prop:retract\] is in turn homotopy equivalent to a perfect twisted module ${\Gamma}_{{{\mathcal{V}}}}'$. Then it is clear that $F({\Gamma}_{{{\mathcal{V}}}}')$ is homotopy equivalent to ${{\mathcal{V}}}$. As ${\Gamma}_{{\mathcal{V}}}$ is a quasi-cohesive modules we may apply [@Block10 Theorem 3.2.7]. It suffices to show that ${\Gamma}(X, {{\mathcal{V}}}\otimes_{{\mathcal{R}}} {\mathcal{A}}^{0})$ is a perfect dg sheaf of $A^{0}$-modules to deduce that there is a cohesive module homotopy equivalent to ${\Gamma}_{{\mathcal{V}}}$. By assumption $E \coloneqq {{\mathcal{V}}} \otimes_{{\mathcal{R}}} {\mathcal{A}}^{0}$ is a perfect dg sheaf. Thus there is a cover $\{U_{i}\}_{i \in I}$ of $X$ such that each $E|_{U_{i}}$ is quasi-isomorphic to a strictly perfect dg sheaf $E'_{i}$. Following the argument in in [@Wells07 Proposition III.4.1] we may actually assume $I$ is finite as long as we only demand $E'_{i}$ to be strictly perfect on each connected component of $U_{i}$. Of course $U_{i}$ can have infinitely many components, but since $E$ is globally bounded we may still choose the $E'_{i}$ to be strictly perfect. As ${\mathcal{A}}^{0}$ is fine we may apply Theorem \[thm-swan\] degree by degree and see that each $ E'_{i}(U_{i})$ is a finitely generated bounded complex of projective ${\mathcal{A}}^{0}(U_{i})$-modules. We use the fact that ${\mathcal{A}}^{0}$ is fine to write the identity on $E(X)$ as a finite sum $\sum_{i} \phi_{i}$ where $\operatorname{supp} \phi_{i} \subset U_{i}$. Thus every $\phi_{i}: E(X) {\rightarrow}E(X)$ factors (up to homotopy) through $E(U_{i})$. By the above $E(U_{i})$ is strictly perfect and thus $\phi_{i}$ is algebraically nuclear (up to homotopy), cf. the proof of Lemma \[lemma-cohesiveretracts\]. Then [@Quillen96 Proposition 1.1] shows that $\phi_{i}$ is homotopy equivalent to a map $x \mapsto \sum_{j} f_{ij}(x)e_{ij}$ for some finite family of functions $f_{ij}: E(X) {\rightarrow}A^{0}(X)$ and objects $e_{ij} \in E(X)$. Thus ${\mathbf 1}\simeq \sum_{i} \sum_{j} f_{ij}(x)e_{ij}$. So the identity on $E(X)$ is algebraically nuclear and $E(X) ={\Gamma}(X, {{\mathcal{V}}}\otimes_{{\mathcal{R}}} {\mathcal{A}}^{0})$ is perfect. We can now compare perfect twisted modules with perfect dg sheaves. The following two results are needed in the next section. \[thm-twistedsheaves\] Let $(X, {\mathcal{R}})$ and ${\mathcal{A}}$ be as in Proposition \[prop-cohesivesheaves\]. If moreover $(X, {\mathcal{A}}^{0})$ is locally ringed, ${\mathcal{A}}^{0}$ is commutative, $A$ is flat over $A^{0}$ and $X$ has finite covering dimension then $F$ induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A)) {\rightarrow}{D_{\operatorname{perf}}}^{B}(X, {\mathcal{R}})$. This is Proposition \[prop-cohesivesheaves\] together with Lemma \[lemma-surjective\], which says that the functor $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A)) {\rightarrow}{D_{\operatorname{perf}}}^{B}(X, {\mathcal{R}})$ is essentially surjective. \[thm-twistedsheaves2\] Let $(X, {\mathcal{R}})$ and ${\mathcal{A}}$ be as in Proposition \[prop-cohesivesheaves\]. If moreover $(X, {\mathcal{A}}^{0})$ is locally ringed, ${\mathcal{A}}^{0}$ is commutative and ${\mathcal{R}}$ is the constant sheaf $\underline k$ then $F$ induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A)) {\rightarrow}{D_{\operatorname{perf}}}(X, \underline k)$. Again this follows by Proposition \[prop-cohesivesheaves\] together with essential surjectivity. By Lemma \[lemma-perfectconstant\] below we may identify perfect dg sheaves with dg sheaves with locally constant cohomology. Then we note that any locally constant sheaf ${\mathcal{M}}$ on $X$ is in the essential image of $F$. ${\mathcal{M}}\otimes {\mathcal{A}}^{0}$ is locally free and thus ${\Gamma}(X, {\mathcal{A}} \otimes {\mathcal{M}})$ is a cohesive module by Theorem \[thm-swan\] and thus homotopy equivalent to a perfect twisted module by Proposition \[prop:retract\]. Clearly ${\mathcal{M}}$ is quasi-isomorphic to $F({\Gamma}(X, {\mathcal{A}} \otimes {\mathcal{M}}))$. As $F$ is quasi-fully faithful and $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A))$ is triangulated this shows that any subcategory of ${D_{\operatorname{perf}}}(X, \underline k)$ containing locally constant sheaves and closed under triangles is in the essential image of $\operatorname{H^{0}}(F)$. But any perfect complex over $\underline k$ is a finite extension of its cohomology sheaves, thus contained in the image of $\operatorname{H^{0}}(F)$. We note that the global boundedness condition on perfect dg sheaves is automatic for clc sheaves. \[lemma-perfectconstant\] Let $X$ be locally contractible. Then ${D_{\operatorname{lf}}}(X, \underline{{k}})$ is equivalent to the derived category of clc sheaves. Moreover, ${D_{\operatorname{perf}}}(X, \underline k)$ is equivalent to the category of sheaves with locally constant cohomology sheaves whose fibres are perfect when considered as dg modules over $k$. The cohomology of ${\mathcal{M}} \in {D_{\operatorname{lf}}}(X, \underline k)$ is locally given as the cohomology of a complex of ${k}$-modules, and thus constant. Conversely consider a dg sheaf ${\mathcal{M}}$ and some contractible open set $U$ on which its cohomology is a constant $\underline k$-module. As $U$ has no cohomology ${\mathcal{M}}|_{U}$ is quasi-isomorphic to a direct sum of its cohomology sheaves. Using free resolutions of the cohomology sheaves shows that ${\mathcal{M}}$ is locally quasi-isomorphic to a free dg sheaf. The statement for perfect dg sheaves follows similarly. \[cor-cohesiveperfect\] In the setting of Theorem \[thm-twistedsheaves\] and Theorem \[thm-twistedsheaves2\] we also have $\operatorname{H^{0}}({\mathcal{P}}_{A}) \cong {D_{\operatorname{perf}}}^{B}(X, {\mathcal{R}})$. There is an embedding $\operatorname{H^{0}}({\mathcal{P}}_{A}) {\rightarrow}\operatorname{H^{0}}(\operatorname{Tw_{perf}}(A))$ so it remains to show essential surjectivity, which follows by inspecting the proofs. Corollary \[cor-quasifullyfaithful\] shows that if we assume ${\mathcal{A}}$ is flat over ${\mathcal{R}}$ then moreover $F$ is quasi-fully faithful in Theorems \[thm-twistedsheaves\] and \[thm-twistedsheaves2\] and gives a quasi-equivalence with a dg category of perfect complexes. In general, the equivalences of homotopy categories may be be enhanced to quasi-equivalences of dg categories between $\operatorname{Tw_{perf}}(A)$ and the dg-category of fibrant cofibrant dg ${\mathcal{R}}$-modules which are perfect dg sheaves. As presheaves in the image of $F$ are fibrant it suffices to compose $F$ with functorial cofibrant replacement. Applications ============ The de Rham algebra {#sect-derham} ------------------- In this section the ground ring $k$ is ${\mathbb{R}}$ and $X$ is a connected smooth manifold. We consider perfect twisted modules over the de Rham algebra ${\Omega}(X)$. We denote by ${\Omega}$ the dg sheaf of de Rham algebras. Recall that we consider ${\Omega}(X)$ as a dg AM algebra and that all tensor products are understood to be completed. Using what we have done so far we can recover and generalise the main result of [@Block09], up to replacing infinity local systems by clc sheaves with cohomology sheaves of finite rank. Note that one may consider cohesive modules (or equivalently perfect twisted modules) over the de Rham algebra $A$ as *${\mathbb{Z}}$-graded connections*. By Theorem \[thm-swan\] we may consider a a complex of finitely generated projective ${\Omega}^{0}(X)$-modules $E$ as a dg vector bundle ${\mathcal{E}}$, and the differential becomes a ${\mathbb{Z}}$-graded connection ${\mathbb{E}}: {\mathcal{E}} {\rightarrow}{\mathcal{E}} \otimes_{{\Omega}^{0}} {\Omega}$ satisfying $d{\mathbb{E}} + {\mathbb{E}}^{2} = 0$. This is the natural derived analogue of a vector bundle with a flat connection. \[thm-derham\] Let $X$ be a connected manifold (not necessarily compact). Then the dg functor $F:\operatorname{Tw_{perf}}(X) {\rightarrow}\underline{{\mathbb{R}}}{\textrm{-Mod}}$ sending $E$ to $U \mapsto E \otimes_{{\Omega}(X)} {\Omega}(U)$ is quasi-fully faithful and induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(X)) \cong {D_{\operatorname{perf}}}(X, \underline{{\mathbb{R}}})$. The theorem follows from Theorem \[thm-twistedsheaves\] or Theorem \[thm-twistedsheaves2\] together with Corollary \[cor-quasifullyfaithful\], applied to ${\mathcal{R}} = \underline {{\mathbb{R}}}$ and ${\mathcal{A}} = {\Omega}$. To check the conditions fix some point $x \in X$ and some perfect dg $\underline {\mathbb{R}}$-module $G$. We consider the smooth homotopy equivalence ${\mathbb{R}}\otimes \operatorname{\underline{End}}(G) {\rightarrow}{\Omega}(U)\otimes \operatorname{\underline{End}}(G)$ given by inclusion and evaluation at $x$. Then we apply Corollary \[cor:deRham\](i) to verify that the de Rham algebra satisfies (\*\*) and the assumptions of Proposition \[prop-cohesivesheaves\]. \[rk-infinitylocal\] To recover the results of [@Block09] we use Corollary \[cor-cohesiveperfect\] and then recall that perfect dg sheaves over $\underline{{\mathbb{R}}}$ are clc sheaves by Lemma \[lemma-perfectconstant\], which are in turn equivalent to various other notions of *infinity local systems*. In fact, under mild assumptions, the following are all quasi-equivalent dg categories: 1. perfect clc sheaves, sometimes called homotopy locally constant sheaves, i.e. fibrant cofibrant dg sheaves whose cohomology sheaves are locally constant of finite rank, 2. perfect dg modules over the dg algebra of chains on the Moore loop space of $X$, 3. the dg category obtained from the cotensor action of singular simplices on $X$ on the dg category of perfect chain complexes, see [@Holstein3], 4. (combinatorial) infinity local systems on a simplicial set as explicitly described in terms of a Maurer-Cartan condition in [@Block09]. One can extend all these notions by dropping the assumption of perfectness and the quasi-equivalences still hold. The equivalence of (1) and (2) follows from [@Holstein1] and [@Holstein2], (2) and (3) are identified in [@Holstein1]. The correspondence of (3) and (4) follows from [@Holstein3]; note that there is a difference of definition between the objects considered in (3) and (4) for an arbitrary simplicial set, but on fibrant simplicial sets the definitions agree. In [@Holstein1] it is shown that all of these can be interpreted as categorified cohomology of $X$, i.e. cohomology of $X$ with coefficients in the dg category of perfect complexes. Keeping with this viewpoint one could consider the dg category of cohesive modules over ${\Omega}(X)$ as categorified de Rham cohomology. Unravelling definitions we may also see that the category (4) for a reduced simplicial set $K$ agrees precisely with our definition of $\operatorname{Tw}(K)$. One may deduce that the two notions agree for arbitrary Kan complexes from Corollary \[cor-twistedreduced\] and homotopy invariance of infinity local systems. The main result of [@Block09] shows that if $X$ is a compact manifold and $k = {\mathbb{R}}$ then the dg category of infinity local systems as in (4) is $A_{{\infty}}$-quasi-equivalent to the dg category of ${\mathbb{Z}}$-graded connections, using computations with iterated integrals. In Theorem \[thm-derham\] we directly establish a quasi-equivalence of cohesive modules for the de Rham algebra with (1). There is, incidentally, also a direct proof comparing ${\mathbb{Z}}$-graded connections to (2), also using iterated integrals [@Arias16]. We now extend this result to the case where we replace the manifold $X$ by a simplicial complex $K$. We write ${\Omega}(K)$ for de Rham algebra of piecewise smooth differential forms on $K$. Piecewise smooth differential forms define a sheaf on the underlying topological space $|K|$ of $K$ that we also denote by ${\Omega}$. \[thm-derhampiecewise\] Let $K$ be a connected finite dimensional simplicial complex. Then the functor $F: \operatorname{Tw_{perf}}(K){\rightarrow}\underline{{\mathbb{R}}}{\textrm{-Mod}}$ sending $E$ to $U \mapsto E \otimes_{{\Omega}(X)} {\Omega}(U)$ is quasi-fully faithful and induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(K)) \cong {D_{\operatorname{perf}}}(|K|, \underline{{\mathbb{R}}})$. First we show that piecewise smooth functions (and thus piecewise smooth forms) form a fine sheaf on $K$. It is enough to construct, given two closed subsets $A$ and $B$ of $K$, a section $s$ of ${\Omega}^0$ that is equal to 1 on $A$ and $0$ on $B$. We proceed by induction on the dimension of the simplex. So assume we have constructed the restriction of $s$ to $k$-simplices and denote it by $s'$. Consider a $(k+1)$-simplex $L$. We have to check that we can separate $L\cap A$ and $L\cap B$ by a function that restricts to $s'$ on the boundary. As ${\Omega}^0$ is fine on $L$, we may choose a section $t$ that is equal to 1 on $L\cap A$ and 0 on $L\cap B$. Then on the boundary of $L$ we observe that the function $t - s'$ is 0 on the intersections of $A$ and $B$ with the boundary of $K$. We can easily find a smooth function $t'$ on $K$ that restricts to $t-s'$ and which has support disjoint from $A$ and $B$. Then we let $s = t - t'$. We can clearly do this for all $(k+1)$-simplices simultaneously as we did not change $s'$. Now we need to check that ${\Omega}$ satisfies condition (\*) to deduce the theorem from Theorem \[thm-twistedsheaves2\]. The other conditions on $(|K|,{\Omega})$ are immediate. Note that we cannot use Theorem \[thm-twistedsheaves\] as ${\Omega}(K)$ is not flat over ${\Omega}^0(K)$. Let $x \in |K|$. There is a neighbourhood $U$ of $x$ and a piecewise linear contracting homotopy $H: U \times [0,1] {\rightarrow}U$. This induces a map $H^{*}: {\Omega}(U) = \lim_{\Delta \in K} \Omega(U \cap \Delta) {\rightarrow}\lim_{\Delta \in K} \Omega((U \cap \Delta) \times [0,1]) \cong \Omega(U) \otimes \Omega[0,1]$. Here for the last equivalence we use Corollary \[cor:product\]. Thus the map $H^{*}$ gives a smooth homotopy equivalence between ${\mathbb{R}}$ and $\Omega(U)$. Now we apply Theorem \[thm:homotopyequivalence1\] to deduce condition (\*\*) and apply Theorem \[thm-twistedsheaves2\]. Note that Theorem \[thm-derhampiecewise\] would be false for the polynomial de Rham algebra that is used for example in rational homotopy theory, cf. Example \[eg-polyderham\]. The Dolbeault algebra {#sect-dolbeault} --------------------- In this subsection $k$ is ${\mathbb{C}}$ and $X$ is a (not necessarily compact) complex manifold equipped with its sheaf of holomorphic functions ${\mathcal{O}}_{X}$. We revisit Block’s proof [@Block10] that the derived category of perfect dg coherent sheaves on a complex manifold $X$ is equivalent to the homotopy category of cohesive modules over the Dolbeault algebra $({\mathcal{A}}^{0*}(X), \bar\partial)$. Note that the main result in the previous section draws from the methods in [@Block10]; we have generalised the setting and added some details regarding faithfulness of the functor from twisted modules to perfect dg sheaves. Thanks to Lemma \[lemma-surjective\] we may answer the implicit question in [@Block10 Remark 4.1.4]. Let $X$ be a complex manifold. The functor $F: \operatorname{Tw_{perf}}({\mathcal{A}}^{0*}(X), \bar \partial) {\rightarrow}{\mathcal{O}}_{X}{\textrm{-Mod}}$ sending $V \otimes {\mathcal{A}}^{0*}(X)$ to its dg associated sheaf is quasi-fully faithful and induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw_{perf}}{({\mathcal{A}}^{0*}(X), \bar\partial)}) \cong {D_{\operatorname{perf}}}^{B}(X, {\mathcal{O}}_{X})$. The equivalence of homotopy categories follows from Theorem \[thm-twistedsheaves\]. Condition (\*) is exactly the content of [@Block10 Lemma 4.1.5]. Quasi-full faithfulness follows from Corollary \[cor-quasifullyfaithful\], which applies since smooth functions are flat over holomorphic functions. To see this, note real analytic functions are flat over holomorphic functions and smooth functions are flat over analytic functions, see [@Malgrange66 Corollary V.1.2]. One might try to also view this result through a suitable Schlessinger-Stasheff theorem. There are, however, considerable conceptual obstacles to implementing this. Note that the inclusion ${\mathcal{O}}(U) {\rightarrow}({\mathcal{A}}^{0*}(U),\bar \partial)$ does not have a section as a function of topological vector spaces for any open set $U$, see [@Mityagin71 Proposition 5.4]. On a closed poly-disk $D$ there is a section (not compatible with restrictions), but it is not clear how to construct a homotopy equivalence between ${\mathcal{O}}(D)$ and $({\mathcal{A}}^{0*}(D),\partial)$, or even what the correct notion of homotopy equivalence would be. The singular cochain algebra {#sect-singular} ---------------------------- In this subsection $X$ is a topological space and $C^{*}(X)$ the pseudo-compact dg algebra of its normalized singular cochains. We will assume that $X$ is connected and locally contractible, and that $k$ has finite homological dimension. We will consider infinitely generated modules, so recall from Section \[sect-notations\] that whenever we consider $M\otimes C^*(X)$ for some graded $k$-module $M$ we will understand it as the completed tensor product. To define a functor from $\operatorname{Tw}(C^*(X))$ to $\underline k{\textrm{-Mod}}$ we recall that the presheaf of singular cochains with coefficients in any abelian group $L$, given for an open set $U\subset X$ by $U \mapsto C_{\text {sing}}^*(U, L)$, has a sheafification given by $U \mapsto C^{*}_{\text{sing}}(U,L)/C_0^{*}(U,L)$. Here $C_{0}^{*}(U,L)$ consists of those singular cochains on $U$ such that there is an open cover of $U$ on which they vanish. See [@Sella16] for details. We write ${\mathcal{C}}^{*}(L) = ({\mathcal{C}}^{*}(L), d_{{\mathcal{C}}})$ for the normalization of $C_{\text{sing}}^{*}(U,L)/C_{0}^{*}(U,L)$. This is a flabby sheaf if $X$ is semi-locally contractible and there is a quasi-isomorphism $\underline L {\rightarrow}{\mathcal{C}}^{*}(L)$. When $L = k$ we drop it from the notation, and we note that $\underline L \otimes {\mathcal{C}}^{*} \cong {\mathcal{C}}^{*}(L)$. Let us consider the dg functor $F: \operatorname{Tw}(X) {\rightarrow}\underline k{\textrm{-Mod}}$ defined by $$\label{eq:def-F}{F(M)(U)=M\otimes_{C^{*}(X)}{\mathcal{C}}^*(U)},$$ where $U$ is an open subset of $X$. Note that as $C^{*}(X)$ is different from ${\mathcal{C}}^{*}(X)$ this differs from Definition \[defn-F\]. Then we have the following result. \[thm-singularcochains\] The dg functor $F: \operatorname{Tw}(X) {\rightarrow}\underline k{\textrm{-Mod}}$ defined above is quasi-fully faithful, and induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw}(X)) \cong {D_{\operatorname{lf}}}(X, \underline k)$. The proof is somewhat long and technical and will occupy the rest of the paper. Many of the technical complications of the proof disappear under the assumption that ${k}$ is a field. Given a simplex $\sigma$ we will denote its vertices by $\sigma_{0}, \dots, \sigma_{n}$ and write $\sigma_{i_{0}\dots i_{k}}$ for the subsimplex spanned by $\sigma_{i_{0}}, \dots, \sigma_{i_{k}}$. Recall that we may write objects of $\operatorname{Tw}(X)$ as $(V \otimes C^{*}(X), D_{V})$ where $V$ is some free graded ${k}$-module. \[lemma-locallyconstant\] Let $V$ be a ${k}$-module considered as a dg module concentrated in degree 0. Then there is a bijective correspondence between $C^*(X)$-modules of the form $(V \otimes C^*(X), D_V)$ and functors from the fundamental groupoid $\Pi(X)$ of $X$ to $\operatorname{End}(V)$, where the latter is viewed as a linear category with one object. It suffices to identify $\operatorname{MC}(C^{*}(X, \operatorname{End}(V)))$ with functors from $\Pi(X)$ to $\operatorname{End}(V)$. Consider an element $f\in \operatorname{MC}(C^{*}(X, \operatorname{End}(V)))$, which is by definition a 1-cochain. To define the functor $\Phi(f): \Pi(X) {\rightarrow}\operatorname{End}(V)$ it suffices to specify it on the morphisms of $\Pi(x)$. For a singular 1-simplex $\sigma$ of $X$, viewed as a morphism of $\Pi(X)$, set $\Phi(f)([\sigma]) = 1+f(\sigma)$. Assuming $f$ is MC we obtain, for any singular 2-simplex $\tau$, $$0 = (df + f^{2})(\tau) = f(\tau_{01}) - f(\tau_{02})+ f(\tau_{12}) + f(\tau_{01})f(\tau_{12}).$$ We deduce that $1+f(\tau_{02}) = (1+f(\tau_{01}))(1+f(\tau_{12}))$. This shows that homotopic paths have the same image and concatenation of paths is sent to multiplication, so $\Phi(f)$ is a well-defined functor. Conversely, given a functor $F: \Pi(X) {\rightarrow}\operatorname{End}(V)$, define a cochain $\Psi(F)$ by $\Psi(F)(\sigma) = F([\sigma]) - 1$. Given any two-simplex $\tau$ we know $F(\tau_{01}) \circ F(\tau_{12}) = F(\tau_{02})$, and the same computation as above shows $\Psi(F)$ is MC. The maps $\Psi$ and $\Phi$ are inverse to each other. Lemma \[lemma-locallyconstant\] is compatible with the correspondence between locally constant sheaves and representations of the fundamental groupoid. With the notation of the proof, for every MC element $x \in \operatorname{MC}(\operatorname{End}(V) \otimes C^{*}(X))$ we have that $(\underline V \otimes {\mathcal{C}}^{*}, d + x)$ is a soft resolution of the locally constant sheaf ${\mathcal{V}}$ associated to $\Phi(g)$. To check the monodromy we may observe that if $f: C_{*}(X) {\rightarrow}V$ represents a section and $\sigma$ is a 1-simplex connecting two points $\sigma_{1}$ and $\sigma_{0}$ then the cocycle condition $(d_{C}+g)(f)(\sigma) = 0$ gives $f(\sigma_{0}) = \Phi(g)([\sigma])(f(\sigma_{1}))$. \[lemma-mcsheafcohomology\] Let $A, B$ be $k$-algebras concentrated in degree zero and $x, y$ be MC elements in $A \otimes C^{*}(X)$ and $B \otimes C^{*}(X)$ respectively. Let $M$ be a $(B,A)$-bimodule and consider the dg module $(M \otimes C^{*}(X), D_{M})$ where $D_{M}(f) = df + yf - (-1)^{|f|}fx$. Then the natural quotient map $q_{M}:M \otimes C^{*}(X) {\rightarrow}M\otimes {\mathcal{C}}^{*}(X)$ is a quasi-isomorphism. We proceed exactly in the same way as to establish the quasi-isomorphism $C^{*}(X) {\rightarrow}{\mathcal{C}}^{*}(X)$. For the reader’s convenience we provide some details. We first observe that ${\mathcal{C}}^{*}(X) = \varinjlim C^{*}_{\mathfrak U}(X)$ where the limit is over covers of $X$ and $C^{*}_{\mathfrak U}(X)$ are those singular cochains which vanish on $\mathfrak U$. So it suffices to show $q_{M}^{\mathfrak U}: M \otimes C^{*}(X){\rightarrow}M\otimes C^{*}_{\mathfrak U}(X)$ is a quasi-isomorphism for every cover $\mathfrak U$. The natural quotient map $q^{\mathfrak U}: C^{*}(X){\rightarrow}C^{*}_{\mathfrak U}(X)$ has a homotopy inverse $P$ induced by iterated barycentric subdivision of simplices. Inspection of the proof e.g. in [@Hatcher02 Proposition 2.21] shows that this homotopy equivalence is entirely formal, depending only on the boundary operator $d$ and the operator $b$ induced by taking the cone over the barycentre of a simplex. As long as they satisfy $db + bd = {\mathbf 1}$ one may define a subdivision chain map $S$ and the homotopy $T$ from $S$ to the identity, and use these to define the desired map $P$, see below. Thus we may repeat the whole construction with twisted coefficients. Write $X = x+1$ and $Y = y+1$. We may write the differential on $M \otimes C^{*}(X)$ as $$(D_{M}f)(\sigma) = Y(\sigma_{01})f(\partial_{0}\sigma) + \sum_{i=1}^{n-1}(-1)^{i} f(\partial_{i}\sigma) + (-1)^{n}f(\partial_{n})X(\sigma_{n-1,n}).$$ We observe that this is just a two-sided version of the usual singular cochain complex with local coefficients. Let $\beta(\sigma)$ denote the cone over the barycentre of $\sigma$. We then define $b_{M}f(\sigma)$ as $Y(\sigma_{0b})f(\beta\sigma)$ where $\sigma_{0b}$ denotes the 1-simplex connecting $\sigma_{0}$ to the barycentre of $\sigma$. A straightforward computation, using the fact that $Y(\sigma_{0b})Y(\sigma_{b0}) = {\mathbf 1}$ by Lemma \[lemma-locallyconstant\], shows $b_{M}D_{M}+D_{M}b_{M} = {\mathbf 1}$. We inductively define a twisted subdivision ${S_M}(f) = D_{M}S_{M}b_{M}(f)$, with $S_{M}(f) = f$ on a $0$-cochain, and a chain homotopy $T_{M}(f) = ({\mathbf 1}- D_{M}T_{M})b_{M}$, with $T_{M}(f)= 0$ on a $0$-cochain. Then $D_{M}T_{M} + T_{M}D_{M} = {\mathbf 1}- S_{M}$. For $m \geq 0$ let $H_{m}= \sum_{0\leq i < m} S^{i}_{M}T_{M}$, this is a chain homotopy from ${\mathbf 1}$ to $S_{M}^m$. For every simplex $\sigma$ there is a minimum $m(\sigma)$ such that $\beta^{m(\sigma)}(\sigma)$ is contained in $\mathfrak U$. We then define the map $H$ by $H(f)(\sigma) = H_{m(\sigma)}(f)(\sigma)$ and the map $P_{M} = S_{M}^{m(\sigma)} + D_{M}H_{m(\sigma)} - D_{M}H$. One can check that $H$ is a chain homotopy between the identity and $P_{M}\circ q^{\mathfrak U}_{M}$. Moreover $P_{M}$ is a right inverse of $q^{\mathfrak U}_{M}$. This establishes the desired homotopy equivalence. Details are as for the untwisted dual case, which may be found in [@Hatcher02]. Associated to a representation of the fundamental groupoid $R:\Pi(X){\rightarrow}\operatorname{End}(V)$ is, according to Lemma \[lemma-locallyconstant\], an MC element $\Psi(R)$. The corresponding twisted module $(V\otimes C^*(X),D_V)\cong (V\otimes C^*(X))^{[\Psi(R)]}$ coincides with the singular complex of $X$ with local coefficients corresponding to the representation $R$. To a $\Pi(X)$-bimodule, i.e. a representation $P:\Pi\times\Pi^{\text{op}}{\rightarrow}\operatorname{End}(M)$ one similarly associates a pair $\Psi_1(P),\Psi_2(P)$ of MC elements and a two-sided twisted complex $(M\otimes C^*(X))^{[\Psi_1(P),\Psi_2(P)]}$; this complex was used in the proof of Lemma \[lemma-mcsheafcohomology\]. Any $\Pi(X)$-bimodule determines, via the canonical functor $\Pi(X){\rightarrow}\Pi(X)\times\Pi(X)^{\text{op}}$, a left $\Pi(X)$-module. It is easy to see that for a singular $n$-cochain $f$ with values in $M$, the map $$f{\rightarrow}(f)\cdot(\sigma_{n, n-1}\cdot \sigma_{n-1,n-2}\cdot\ldots\cdot\sigma_{1,0})$$ determines an isomorphism from the two-sided complex with local coefficients to the one-sided complex. This is analogous to the well-known isomorphism between the two-sided Hochschild complex of a group and a one-sided complex, cf. [@MacLane67 Chapter 6, p. 293]. The following lemma is only needed if ${k}$ is not a field. In that case not all locally constant sheaves are locally free, but the underlying graded module of a twisted module needs to be free. \[lemma-freeresolution\] Any locally constant sheaf ${\mathcal{V}}$ on $X$ is the image under $F$ of a bounded twisted $C^*(X)$-module $W\otimes C^*(X)$. By Lemma \[lemma-locallyconstant\] we know there is a $C^*(X)$-module $(V\otimes C^*(X), D_V)$ mapping to ${\mathcal{V}}$. The only problem is that $V$ might not be free over $k$. We pick a finite free resolution $q: (W, d_{W}) {\rightarrow}V$. Now we need to construct a differential $D_W$ on $W \otimes C^*(X)$ that maps to $D_V$. $D_V$ is determined by the map $D^{1}_{V}|_{V}: V {\rightarrow}V \otimes C^{1}(X)$. For degree reasons all the maps $D^{i}_{V}: V {\rightarrow}V \otimes C^{i}(X)$ for $i \neq 1$ are zero. As $W$ is free we can lift $D_{V}^{1}$ to a chain map $w_{1}: W {\rightarrow}W \otimes C^{1}(X)$. As $D_{V}^{2}=0$ we know $w_{1}^{2} = d_{W}(w_{2})$ for some $w_{2}: W {\rightarrow}W[-1] \otimes C^{2}$. We let $w_{0}= d_{W}$. Then this is the beginning of an inductive construction of a homomorphism $\sum_{i \geq 0} w_{i}$ that will define a differential on $W \otimes C^*(X)$. Assume we are given $w_{i}$ for $i \leq k$ satisfying $\sum_{i=0}^{n} w_{i}w_{n-i} = 0$ for every $n \leq k$. Then $\sum_{i=1}^{k} w_{i}w_{k+1-i}$ is an object of $\operatorname{End}(W)\otimes C^{k+1}$. We now compute $[w_{0}, \sum_{i=1}^{k} w_{i}w_{k+1-i}]$ to check that $\sum w_{i}w_{k+1-i}$ is a $d_{W}$-cocycle. We observe that $$\sum_{0 \leq m, i, j \leq k; m+i+j=k+1} [w_{m}, w_{i}w_{j}] = 0$$ by symmetry. Then we split the sum as $$\left[w_{0}, \sum_{i=1}^{k} w_{i}w_{k+1-i}\right] + \sum_{m = 1}^{k} \left[w_{m}, \sum_{i=0}^{k+1-m} w_{i} w_{k+1-m-i}\right] = 0$$ But for $m \geq 1$ all $\sum_{i} w_{i} w_{k+1-m-i}$ are $0$ by induction. Thus the first term in the sum is $0$, which is what we had to show. As $H^{-k}(\operatorname{\underline{End}}((W, d_{W}))) = \operatorname{Ext}^{-k}(V, V) = 0$ we see that the cocycle $\sum_{i=1}^{k} w_{i}w_{k+1-i}$ is a boundary and we can define $w_{k+1}$ such that $\sum_{i=0}^{k+1} w_{i}w_{k+1-i} = 0$. As $W$ is finite this process terminates. $D_{W}|_{W} \coloneqq \sum w_{i}$ defines a differential on $W\otimes C^*(X)$ that is compatible with $D_{V}$. Now we filter $q: (W \otimes C^*(X), D_W) {\rightarrow}(V \otimes C^*(X), D_V)$ by the singular cochain degree. This is a complete exhaustive filtration and the associated graded map consists of quasi-isomorphisms $(W, d_{W})\otimes C^{p}(X) \simeq V \otimes C^{p}(X)$, thus $q$ is a quasi-isomorphism. In fact, $q$ is a quasi-isomorphism if we replace $X$ by any open subset $U$ and thus we have constructed $(W \otimes C^*(X), D_{W})$ whose image under $F$ is quasi-isomorphic to ${\mathcal{V}}$. \[lemma-singularclc\] For any twisted module $E$ the sheaf $F(E)$ is clc. Consider $P = (V \otimes C^*(X), D_V)$ in $\operatorname{Tw}(C^*(X))$ and restrict $P$ to a contractible subset $U \subset X$. We use the weak equivalence between $U$ and a point and apply Corollary \[cor:mctopologicalequivalence\] to show that $P|_{U}$ is weakly equivalent to a constant sheaf on $U$ with fibre $(V, d_{V})$. \[lemma-singularff\] The natural functor $\operatorname{Tw}(C^{*}(X)) {\rightarrow}{\mathcal{C}}^{*}{\textrm{-Mod}}$ is quasi-fully faithful. Given twisted modules $(V \otimes C^*(X), D_V)$ and $(W \otimes C^*(X), D_W)$ over $C^*(X)$ we know that $\operatorname{\underline{Hom}}_{\operatorname{Tw}({C^*(X)})}(V \otimes C^{*}(X), W \otimes C^{*}(X))$ is given by $(\operatorname{\underline{Hom}}(V,W)\otimes C^*(X))$ with a differential defined by $f \mapsto D_W \circ f - (-1)^{|f|}f \circ D_V$ on $\operatorname{\underline{Hom}}(V,W)$. We then compute $\operatorname{\underline{Hom}}_{{\mathcal{C}}^{*}}(\underline V\otimes {\mathcal{C}}^*, \underline W \otimes {\mathcal{C}}^*)$. By Lemma \[lemma-protoswan\] it is homotopy equivalent to module homomorphisms $\operatorname{\underline{Hom}}_{{\mathcal{C}}^{*}(X)}(V\otimes {\mathcal{C}}^*(X), W \otimes {\mathcal{C}}^*(X))$. This space in turn can be computed as $\operatorname{\underline{Hom}}(V, W) \otimes {\mathcal{C}}^{*}(X)$ with differential induced by $f \mapsto D_{W}\circ f -(-1)^{|f|}f \circ D_{V}$. Writing $M$ for $\operatorname{\underline{Hom}}(V,W)$ we now need to show that there is a quasi-isomorphism $M \otimes C^{*}(X) \simeq M \otimes {\mathcal{C}}^{*}(X)$. Note that $M \otimes C^{*}(X)$ with its differential $D_{M}$ is not a $C^{*}(X)$-module, and in particular not a twisted module. We may still consider its sheafification. By Lemma \[lemma-singularclc\] we know $\underline V\otimes {\mathcal{C}}^{*}$ and $\underline W \otimes {\mathcal{C}}^{*}$ are clc and thus so is $\operatorname{{\mathcal{H}} \! \it{om}}(V, W) \otimes {\mathcal{C}}^{*}$. Moreover, by Corollary \[cor-twistedreduced\] we may assume that $V \otimes C^{*}$ and $W \otimes C^{*}$ are reduced, so we may assume that $D^{0}_{M}$ induces a differential on $M$. We consider the natural map induced by the quotient $C^{*}(X) {\rightarrow}{\mathcal{C}}^{*}(X)$ and filter both sides by the singular degree. We claim the associated spectral sequences agree on the second sheet, showing the map is a quasi-isomorphism as the filtration is complete exhaustive. For the first spectral sequence we have ${}^{I}E_{1}^{pq} = H^{q}(M\otimes C^{0}(X))\otimes_{C^{0}(X)} C^{p}(X)$, which we may rewrite as $H^{q}(M)\otimes C^{p}(X)$. The second sheet computes cohomology of a dg module $(H^{q}(M)\otimes C^*(X), d_{1})$, which satisfies the conditions of Lemma \[lemma-mcsheafcohomology\] for $A = \operatorname{End}(V)$ and $B = \operatorname{End}(W)$. For the second spectral sequence one has ${}^{II}E_{1}^{pq} = H^{q}(M)\otimes {\mathcal{C}}^{p}(X)$, and by Lemma \[lemma-mcsheafcohomology\] the $E_{2}$-terms agree. Here the first spectral sequence computes the Ext groups between clc sheaves using the singular cochain complex, and the second spectral sequence computes the Ext groups using a soft resolution. Recall that to any dg ${k}$-module $C^*$ one can associate its canonical truncation ${\tau_{\leq i}C^*}$ obtained by replacing $C^n$ with zero for $n>i$ and with ${\ker(C^i{\rightarrow}C^{i+1})}$ for $n=i$. Then $\tau_{\leq i}C^*$ is a dg submodule in $C^*$ and we set $\tau_{\geq i+1}C^*:=C^*/(\tau_{\leq i}C^*)$. This construction works for complexes over any abelian category, in particular one can define canonical truncations for dg sheaves of ${k}$-modules. The following result shows that there are corresponding truncation functors for twisted $C^*(X)$-modules. \[lem:truncation\] For every twisted $C^*(X)$-module $M$ there is a twisted module $\tau_{\leq i}M$ and a map $\tau_{\leq i}M{\rightarrow}M$ such that $F(\tau_{\leq i}M){\rightarrow}F(M)$ is isomorphic in the derived category of dg ${k}$-sheaves on $X$ to the canonical map $\tau_{\leq i}F(M){\rightarrow}F(M)$. Similarly there is twisted module $\tau_{\geq i}M$ and a map $M{\rightarrow}\tau_{\geq i}M$ such that $F(M){\rightarrow}F(\tau_{\geq i}M)$ is isomorphic in the derived category of dg ${k}$-sheaves on $X$ to the canonical map $F(M){\rightarrow}\tau_{\geq i}F(M)$. We will prove the statement for the truncation $\tau_{\leq i}$; the other claim for $\tau_{\geq i}$ will follow by taking $\tau_{\geq i}M$ to be the cone of the map $\tau_{\geq i-1}M\rightarrow M$. Let $(V \otimes C^*(X), D_{V})$ be a twisted $C^*(X)$-module that will be assumed to be reduced (or we replace it by a reduced one by Corollary \[cor-twistedreduced\]). Note that $D_V$ restricts to $(\tau_{\leq i}(V) \otimes [C^*(X)])$ and so $ (\tau_{\leq i}(V) \otimes [C^*(X)], D_{V})$ is well-defined as a dg $C^*(X)$-module. This may not be a twisted $C^*(X)$-module since $\tau_{\leq i}(V)$ may not be free over ${k}$. We pick a ${k}$-free resolution $(W,d_W)$ of $\tau_{\leq i}(N)$ and, arguing as in the proof of Lemma \[lemma-freeresolution\], construct a differential $D_W$ on $W\otimes C^*(X)$ together with a filtered quasi-isomorphism $(W\otimes C^*(X), D_W){\rightarrow}(\tau_{\leq i}(V) \otimes [C^*(X)], D_{V})$. Let us set $\tau_{\leq i}(V \otimes C^*(X), D_{V}):=(W\otimes C^*(X), D_W)$. We need to show that the truncation so obtained agrees with the truncation of dg sheaves upon applying the functor $F$. This is a local statement, and so it suffices to prove it with $X$ replaced by a small contractible neighbourhood $U\subset X$. This is, however, obvious since the twisted $C^*(U)$-module $(W\otimes C^*(U), D_W)$ is homotopy equivalent to the (untwisted) tensor product of complexes $(W,d_W)$ and $C^*(U)$ by Corollary \[cor:mctopologicalequivalence\]. We use Lemma \[lemma-perfectconstant\] to identify ${D_{\operatorname{lf}}}(X, \underline k)$ with cohomologically constant sheaves. Then by Lemma \[lemma-singularclc\] the image of $F$ consists of locally free dg sheaves. Next we show that the functor $\operatorname{H^{0}}(\operatorname{Tw}(X)) {\rightarrow}{D_{\operatorname{lf}}}(X, \underline k)$ induced by $F$ is fully faithful. To this end note that this functor can be represented as the following composition: $$\operatorname{H^{0}}(\operatorname{Tw}(X)) {\rightarrow}\operatorname{H^{0}}({\mathcal{C}}^{*}{\textrm{-Mod}}) {\rightarrow}{D_{\operatorname{lf}}}(X, {{\mathcal{C}}^{*}}) {\rightarrow}{D_{\operatorname{lf}}}(X, \underline k)$$ By Lemma \[lemma-singularff\] the first functor is fully faithful. To show the second functor is fully faithful on the image of $\operatorname{H^{0}}(\operatorname{Tw}(X))$ we claim $\operatorname{\underline{Hom}}(\underline V \otimes {\mathcal{C}}^{*}, \underline W \otimes {\mathcal{C}}^{*}) \simeq R\operatorname{\underline{Hom}}(\underline V \otimes {\mathcal{C}}^{*}, \underline W \otimes {\mathcal{C}}^{*})$. We deduce this claim by following verbatim the proof of Lemma \[lemma-fullyfaithful\]. By Corollary \[cor-twistedreduced\] (1) we have a homotopy equivalence $(V \otimes C^{*}(U), D_{W}) \simeq (V,d_{V})\otimes C^{*}(U)$ on any contractible set $U$. This takes the place of condition (\*). We allow for unbounded dg sheaves, but this does not affect the proof as the filtration by degree of $\operatorname{{\mathcal{H}} \! \it{om}}(V,W)\otimes {\mathcal{C}}^{*}$ is still exhaustive and Hausdorff. Note that the dg $k$-module $(V, d_{V})$ is cofibrant as it is free in each degree and $k$ has finite homological dimension. Hence the associated sheaf $\underline V \otimes {\mathcal{C}}^{*}$ is also cofibrant. Since $\underline {k}\simeq {\mathcal{C}}^{*}$ we have $D(X, {\mathcal{C}}^{*}) \cong D(X, \underline {k})$ and $\operatorname{H^{0}}(\operatorname{Tw}(X)) {\rightarrow}D_{lf}(X, \underline {k})$ is fully faithful. Moreover, as ${\mathcal{C}}^{*}$ is flat over $k$ we may refine the argument and show, as in the proof of Corollary \[cor-quasifullyfaithful\], that the functor $F: \operatorname{Tw}(X) {\rightarrow}\underline {k}{\textrm{-Mod}}$ is quasi-fully faithful. Now we determine the quasi-essential image of $F$. The subcategory of ${D_{\operatorname{lf}}}(X, \underline k)$ given by bounded dg sheaves is the smallest triangulated subcategory inside the derived category of dg ${k}$-module sheaves on $X$ containing all locally constant sheaves. This follows since any bounded element in ${D_{\operatorname{lf}}}(X, \underline k)$ is an iterated extension of its cohomology sheaves. The image of $F$ contains all locally constant sheaves by Lemma \[lemma-freeresolution\]. Thus, since $F$ is compatible with cones, the quasi-essential image of $F$ contains all *bounded* clc sheaves. Observe that every bounded below clc sheaf ${\mathcal{M}}$ in ${D_{\operatorname{lf}}}(\underline k)$ is a homotopy colimit (in the sense of [@Bokstedt93]) of its truncations, $\operatorname*{hocolim}_i \tau_{\leq i} {\mathcal{M}} \simeq {\mathcal{M}}$. By Lemma \[lemma-freeresolution\] we may lift the diagram $\tau_{\leq i}{\mathcal{M}}$ to a diagram $\{P_i\}$ in $\operatorname{H^{0}}(\operatorname{Tw}(X))$. As $\operatorname{H^{0}}(\operatorname{Tw}(X))$ has arbitrary direct sums we may define $P = \operatorname*{hocolim}_i P_{i}$ and there is a natural map ${\mathcal{M}} {\rightarrow}F(P)$ which is an isomorphism on cohomology (as we can check on truncations using $\tau_{\leq i} F(P) \simeq F(\tau_{\leq i} (P)) \simeq \tau_{\leq i} {\mathcal{M}}$ by Lemma \[lem:truncation\]). Thus, all *bounded below* clc sheaves are in the quasi-essential image of $F$. Finally we write a bounded above clc sheaf ${\mathcal{M}}$ as a limit of bounded dg sheaves; ${\mathcal{M}} \cong \lim \tau_{\geq i} {\mathcal{M}}$. We will explicitly construct a twisted $C^*(X)$-module $Q$ with a map $F(Q) {\rightarrow}{\mathcal{M}}$ such that $\tau_{\geq i}F(Q) \simeq \tau_{\geq i}{\mathcal{M}}$, showing $F(Q) \simeq {\mathcal{M}}$. To find $Q$ we proceed as follows. We fix some $Q_{0} = Q_{0}' \otimes C^{*}(X)$ mapping to $\tau_{\geq 0}{\mathcal{M}}$ and then construct twisted modules $Q_{i} = Q_{i}'\otimes C^{*}(X)$, $i < 0,$ inductively. We may write $\tau_{\geq i} {\mathcal{M}}$ as an extension of $\tau_{\geq i+1}{\mathcal{M}}$ by $H^{i}({\mathcal{M}})[-i]$. Using Lemma \[lemma-freeresolution\] we obtain $W_{i} \otimes C^*(X)$ mapping to $H^{i}({\mathcal{M}})[-i]$ under $F$ where $W_{i}$ is a finite complex of free ${k}$-modules; moreover, because ${k}$ is of finite homological dimension $\operatorname{gd}({k})$, the length of $W_i$ is bounded independently of $i$. As $F$ is quasi-full we may lift the extension map $\tau_{\geq i+1}{\mathcal{M}} {\rightarrow}H^{i}({\mathcal{M}})[-i+1]$ to $\eta_{i}: Q_{i+1} {\rightarrow}W_{i}\otimes C^*(X)[1]$. Now the cone on $\eta_{i}$ is defined as the twisted module of the form $Q'_{i} \otimes C^{*}(X)$ where $Q'_{i} = Q'_{i+1} \oplus W_{i}[1]$ and the differential is given by $D_{Q_{i}} = (D_{Q_{i+1}}, D_{W} + \eta_{i})$, see Section \[twisted\]. Thus we let $Q_{i}$ be the cone of the map $\eta_{i}$. By construction there is a quasi-isomorphism $F(Q_{i}) \simeq \tau_{\geq i}{\mathcal{M}}$. By construction $Q_{i}'$ is eventually constant, to be precise the degree $m$ part $(Q_{i}')^{m}$ is independent of $i$ if $i < m - \operatorname{gd}({k}) -1$. We define a graded ${k}$-module $Q'$ by $(Q')^{m} \coloneqq (Q_{i}')^{m}$ for sufficiently small $i$. Similarly, the differential $D_{Q_{i}}$ restricted to $Q_{i}'$ is eventually constant and we define $D_{Q}$ on $(Q')^{m}$ to be $D_{Q_{i}}$ (for sufficiently small $i$) and extend by the Leibniz rule. Then $Q = (Q' \otimes C^{*}(X), D_{Q})$ is the desired twisted module. There is a natural map $Q {\rightarrow}Q_{i}$ and the maps $F(Q_{i}) {\rightarrow}\tau_{\geq i}{\mathcal{M}}$ induce a map $F(Q) {\rightarrow}{\mathcal{M}}$. We need to check that $\tau_{\geq i}F(Q) \simeq F(\tau_{\geq i}Q)$ is equivalent to $\tau_{\geq i}{\mathcal{M}}$. By construction $\tau_{\geq i+1} Q' = Q'_{i+1} \oplus \tau_{\geq i+1}W_{i}[1] \oplus \tau_{\geq i+1}W_{i-1}[2] \oplus \dots$. All summands but the first are acyclic for $D^{0}$, thus after applying $F$ we can show that $F(\tau_{\geq i} Q) \simeq \tau_{\geq i}{\mathcal{M}}$. This shows that that every *bounded above* clc sheaf is in the quasi-essential image of $F$. As every dg sheaf $\mathcal F$ is an extension of a bounded above sheaf $\tau_{\geq 0}\mathcal F$ by a bounded below dg sheaf $\tau_{\leq 0}\mathcal F$, it follows that $F$ is quasi-essentially surjective. With $X$ as above $F$ induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(X) ) {\rightarrow}{D_{\operatorname{perf}}}(X, \underline k)$. We follow the proof of Theorem \[thm-singularcochains\]. In particular this means we define the functor $F$ on $\operatorname{Tw_{fg}}(X)$, the dg category of finitely generated twisted modules, and obtain an embedding $\operatorname{H^{0}}(\operatorname{Tw_{fg}}(X)) {\rightarrow}{D_{\operatorname{perf}}}(X, \underline k)$. As the right hand side is idempotent complete this extends to an embedding $\operatorname{H^{0}}(\operatorname{Tw_{perf}}(X)) {\rightarrow}{D_{\operatorname{perf}}}(X, \underline k)$. Essential surjectivity needs a little extra care. Considering any perfect $\underline{k}$-module as a finite extension of its cohomology sheaves it suffices to find a preimage for a locally constant sheaf ${\mathcal{M}}$. The fiber of ${\mathcal{M}}$ may not be free, but by Lemma \[lemma-perfectconstant\] it is quasi-isomorphic to a strictly perfect dg module $P$ over $k$. Next we find a dg module $Q$ over $k$ such that $P \oplus Q$ is free and of finite rank in each degree, and $Q$ has cohomology concentrated in degree $0$. To do this let us write $P$ as $P^{n}{\rightarrow}\dots {\rightarrow}P^{0}$. We pick for each $P^{i}$ a $k$-module $R^{i}$ such that $P^{i}\oplus R^{i}$ is free of finite rank. Then let $Q^{i} = R^{i} \oplus \bigoplus_{j = i+1}^{n} R^{j}\oplus P^{j}$ and define differentials inductively. The map $d^{n}: Q^{n}{\rightarrow}Q^{n-1}$ is just the inclusion of $R^{n}$, and $d^{i}$ is defined as the natural inclusion into $Q^{i-1}$ of the cokernel of $d^{i-1}$. With this definition the cohomology of $Q$ is a $k$-module $N$ concentrated in degree 0. We now consider the locally constant sheaf ${\mathcal{M}}' = {\mathcal{M}} \oplus \underline N$. By construction its fiber has a finite free resolution of finite rank. We use Lemma \[lemma-freeresolution\] to lift ${\mathcal{M}}'$ to a finitely generated twisted module, using the fact that we may choose $W$ in the proof of Lemma \[lemma-freeresolution\] to be of finite rank. But ${\mathcal{M}}$ is a summand of ${\mathcal{M}}'$. Thus, as $\operatorname{Tw_{perf}}(X)$ is equivalent to an idempotent complete subcategory of ${D_{\operatorname{perf}}}(X, \underline k)$, it follows that ${\mathcal{M}}$ is in the essential image of $\operatorname{Tw_{perf}}(X)$. For an early incarnation of MC elements on singular cochains see [@BrownJr59]. There twisting cochains are used to express singular chains on a fibre space in terms of singular chains on base and fibre. One may interpret this as higher MC elements on $C^{*}(X)$ representing certain infinity local systems. There is a version of Theorem \[thm-singularcochains\] for simplicial sets. Let $X$ be a connected Kan complex. Then there is a quasi-fully faithful functor from $\operatorname{Tw}(X)$ to the category of dg sheaves of ${k}$-modules on $|X|$, the geometric realization of $X$, which induces an equivalence $\operatorname{H^{0}}(\operatorname{Tw}(X)) \cong {D_{\operatorname{lf}}}(|X|)$. The singular simplicial set of $|X|$, is weakly equivalent to $X$. Since both are Kan simplicial sets, by Corollary \[cor:weaklyKan\] their categories of twisted modules are quasi-equivalent. Now the result follows from \[thm-singularcochains\] since $|X|$ is locally contractible. Nuclear spaces {#app-nuclear} ============== In this Appendix we collect some facts about Grothendieck’s nuclear spaces used in the main text, for the reader’s convenience. Our main sources are [@Jarchow12] and [@Treves67]. We will consider complete locally convex Hausdorff topological vector spaces over $\mathbb{R}$ which will be referred to below as simply ‘vector spaces’. If we have a linear continuous injection $U\hookrightarrow V$ that is a homeomorphism of $U$ on its image, we will refer to $U$ as a subspace of $V$. Let $U$ and $V$ be vector spaces. Their projective tensor product ${U\otimes_\pi V}$ is a vector space having a universal property with respect to continuous bilinear maps out of $U\times V$, i.e. for any vector space $W$ there is a natural isomorphism between the set of continuous bilinear maps $U\otimes_\pi V{\rightarrow}W$ and the space of bilinear continuous maps $U\times V{\rightarrow}W$. It is clear that the above defines $U\otimes_\pi V$ up to a unique isomorphism, and there is an explicit construction (that we will not need) showing that the vector space with the required universal property exists. There are other natural notions of a tensor product of vector spaces, of which the most important is that of an *injective* tensor product denoted by $U\otimes_\epsilon V$, [@Jarchow12 Chapter 16]. There is a canonical continuous map $U\otimes_\pi V{\rightarrow}U\otimes _\epsilon V$. A vector space $U$ is called *nuclear* if for any vector space $V$ the canonical map $U\otimes_\pi V{\rightarrow}U\otimes_\epsilon V$ is a topological isomorphism. From now on we will refer to projective tensor products as simply tensor products and omit the corresponding subscript. The category of nuclear spaces and continuous linear maps is closed with respect to various natural operations. \[thm:stable\] The collection of nuclear spaces is stable with respect to forming arbitrary direct products, tensor products and passage to subspaces. See [@Jarchow12 Corollary 21.2.3]. \[cor:stable\] The category of nuclear spaces contains arbitrary limits. This follows immediately from Theorem \[thm:stable\] since any limit can be constructed using direct products and passing to subspaces. It turns out that the operation of tensor product with a nuclear space commutes with arbitrary limits: \[thm:limit\] Let $U_\alpha$ be a diagram of vector spaces and continuous linear maps and $V$ be a nuclear space. Then there is a natural topological isomorphism $$(\underleftarrow{\lim}_\alpha U_\alpha)\otimes V\cong \underleftarrow{\lim}_\alpha( U_\alpha\otimes V).$$ It suffices to show that tensor products commutes with direct products and passing to kernels. This follows from [@Jarchow12 Proposition 16.2.5 and Theorem 16.3.1], taking into account that injective and projective tensor products with a nuclear space are isomorphic. A lot of vector spaces one encounters in analysis are nuclear. In particular: \[thm:smooth\] Let $W$ be an open subset of some topological simplex $\Delta^{n}$. Then the algebra $C^\infty(W)$ of smooth functions on $W$ is nuclear. Let $I^n_\epsilon\subset\mathbb{R}^n$ denote the $n$-dimensional cube in $\mathbb{R}^n$ with side of length $\epsilon>0$. Using Seeley’s extension theorem, more specifically its version for domains with corners [@Kriegl97 Proposition 24.10], we conclude that the restriction map ${C^\infty(\mathbb{R}^n){\rightarrow}C^\infty(I^n_\epsilon)}$ has a continuous splitting and, since $C^\infty(\mathbb{R}^n)$ is nuclear, [@Treves67 Corollary to Theorem 51.5], its retract $C^\infty(I^n_\epsilon)$ is likewise nuclear. Moreover, clearly the algebra of smooth functions on any closed subset in $\mathbb{R}^n$ diffeomorphic to $I^n_\epsilon$ also forms a nuclear space as it is isomorphic to $C^\infty(I^n_\epsilon)$. We then deduce nuclearity of $W$ by representing it as a union of a collection of subsets diffeomorphic to $I^n_\epsilon$ and using Theorem \[thm:limit\]. Given a smooth manifold $X$ we consider its de Rham algebra ${\Omega}(X)$ and for a simplicial complex $K$ we consider its piecewise smooth de Rham algebra ${\Omega}(K)$. We also consider the piecewise smooth de Rham algebra on any open subset $U$ of the underlying space $|K|$ of $K$. Theorem \[thm:smooth\] implies the following. \[cor:smooth\] If $W$ be an open subset of $\mathbb{R}^n$ or of some standard simplex $\Delta^{n}$ then the dg algebra $\Omega(W)$ is nuclear. \[thm:product\] Let $U, W$ be open subsets of topological simplices $\Delta^{n}$ and $\Delta^m$ respectively for some $n,m>0$. Then there is a natural topological isomorphism $\Omega(U\times W)\cong \Omega(U)\otimes \Omega(W)$. It suffices to prove the isomorphism $C^\infty(U\times W)\cong C^\infty(U)\otimes C^\infty(W)$. Arguing as in the proof of Theorem \[thm:smooth\], we represent $U$ and $V$ as unions of subsets diffeomorphic to cubes $I^n_\epsilon$ and $I^m_{\epsilon^\prime}$; it will be sufficient to prove the desired isomorphism for $U=C^\infty(I^n_\epsilon)$, $V=C^\infty(I^m_{\epsilon^\prime})$. Since $C^\infty(I^n_\epsilon)$ and $C^\infty(I^m_{\epsilon^{\prime}})$ are retracts of $C^\infty(\mathbb{R}^n)$ and $C^\infty(\mathbb{R}^m)$ respectively, the natural map ${C^\infty(I^n_\epsilon)\otimes C^\infty(I^m_{\epsilon^\prime}){\rightarrow}C^\infty(I^n_\epsilon\times I^m_{\epsilon^\prime})}$ is a retract of the map $C^\infty(\mathbb{R}^n)\otimes C^\infty(\mathbb{R}^m){\rightarrow}C^\infty(\mathbb{R}^n\times \mathbb{R}^m)$ and so is an isomorphism since the latter map is, [@Treves67 Theorem 51.6]. \[cor:product\] Let $U$ and $W$ be open subsets of the underlying spaces of simplicial complexes $K$ and $L$. Then ${\Omega}(U \times W) \cong {\Omega}(U)\otimes {\Omega}(W)$. By definition ${\Omega}(U) = \lim_{\Delta \in K} {\Omega}(U \cap |\Delta|)$. As the tensor product commutes past the limits by Theorems \[thm:smooth\] and \[thm:limit\] it suffices to check the result for open subsets of the standard simplex, which is is the content of Theorem \[thm:product\]. [^1]: This work was partially supported by EPSRC grants EP/N015452/1 and EP/N016505/1
{ "pile_set_name": "ArXiv" }
--- abstract: 'We define an quantum entropy conditioned on post-selection which has the von Neumann entropy of pure states as a special case. This conditional entropy can take negative values which is consistent with part of a quantum system containing less information than the whole which can be in a pure state. The definition is based on generalised density operators for post-selected ensembles. The corresponding density operators are consistent with the quantum generalisation of classical conditional probabilities following Dirac’s formalism of quasi-probability distributions.' author: - Sina Salek - Roman Schubert - Karoline Wiesner bibliography: - 'e012119.bib' title: 'Negative Conditional Entropy of Post-Selected States' --- Introduction \[sec.prelim\] =========================== Post-selection refers to keeping the record of the outcome of some ensemble quantum measurement only for those parts of the ensemble which at a later point in time are in a desired, so-called postselected state and discarding the remaining results. Ensembles prepared in a state $\ket{\psi}$ and post-selected in a state $\ket{\phi}$ are described by a generalised density operator as [@Reznik1995; @Hosoya2010], This generalisation is appropriate when a weak or no measurement has been performed between pre- and post- selection. For a proposal that incorporates strong measurements see [@Silva2014]. Nevertheless, we restrict our attention to the case where no strong measurement has been performed. These generalised density operators for post-selected ensembles are used to obtain the so-called weak values $\Pi_w$ [@Aharonov1988] of an operator $\Pi$ as Experimentally, a weak value is obtained by weakly coupling an ensemble of states to a measuring apparatus, and post-selecting at a later time. To have an intuition about weak measurement on pre- and post- selected ensembles, take the 3-box problem [@Aharonov1999]. Here, at time $t=0$ a state $\ket{\psi}$ is prepared in a superposition of states $\ket{A}$, $\ket{B}$ and $\ket{C}$ (the three boxes), *e.g.* $\ket{\psi}=\frac{1}{\sqrt{3}}(\ket{A}+\ket{B}+\ket{C})$. At a later time $t=1$ the system is weakly measured in the basis $\{\ket{A},\ket{B},\ket{C}\}$ and then post-selected in some other state $\ket{\phi}$ which is not orthogonal to $\ket{\psi}$, *e.g.* $\ket{\phi}=\frac{1}{\sqrt{3}}(\ket{A}+\ket{B}-\ket{C})$. The resulting weak values of the operator projecting into the three boxes at time $t=1$ with post-selection in state $\ket{\phi}$ at time $t=2$, calculated from Eq. (\[weakval\]), are, respectively, These results can also be mathematically studied in the framework of Dirac quasi-probabilities. In 1945, Paul Dirac introduced a complex phase-space distribution to make an “Analogy Between Classical and Quantum Mechanics"[@Dirac1945], given by where the $a_m$ and $b_n$ are the eigenvalues of the operators $\Pi_A$ and $\Pi_B$, and $\Pi_A^m$ and $\Pi_B^n$ are the projectors onto the corresponding eigenstates. The Dirac distribution (\[dirac\]) satisfies all the conditions of classical Kolmogorov probabilities, except that it is not a positive real function. It was shown that the negativity and complexness of this function is due to the non-commutativity of the quantum mechanical observables [@Johansen2007]. The Dirac distribution is normalised and gives correct marginals, and it obeys the sum rule and the product rule, and it is compatible with Bayes’ law. Note, that Dirac distributions are not limited to phase space. In fact any two operators with non-vanishing overlap between each of their eigenstates can be used to construct a Dirac decomposition, as long as the operators have the same Hilbert space dimension as the state $\rho$, and their eigenvectors are mutually non-orthogonal and none of those eigenvectors are orthogonal to the state $\rho$. Hence a space spanned by any two such observables would be sufficient to describe all the information available from the state $\rho$. This is due to the fact that one can describe any quantum state of $d$-dimensional Hilbert Space with $d^2-1$ elements. For theoretical considerations on the Dirac distribution see the work by Johansen [@Johansen2007] and Hofmann [@Hofmann2012]. An experimental procedure for measuring the Dirac distribution of a general quantum state has been given by Lundeen and Bamber [@Lundeen2012]. We will now describe how weak values can be understood in the framework of Dirac distributions. Note that we can rewrite the weak value, Eq. (\[weakval\]), as Now we can interpret $\Pi_w$ as a conditional Dirac distribution of an ensemble which is pre- and post-selected in states $\ket{\psi}$ and $\ket{\phi}$, respectively. Using Bayes’ law and defining $\rho_\psi:=\ket{\Psi}\bra{\Psi}$ and $ \rho_\phi:=\ket{\Phi}\bra{\Phi}$, the weak value can be written as In this last equation the interpretation as a conditional quasi-probability follows from the use of Bayes’ law. This close connection between weak values and Dirac quasi-probabilities gives an operational meaning to the complex values of the Dirac distribution as a result of weak measurements. In the case where the measurement was performed by coupling the momentum of the measurement pointer to the quantum system, the real part of the weak value refers to the shift in the position of the measurement pointer, while the imaginary part refers to the shift in the momentum of the measurement pointer [@Jozsa2007]. The same interpretation can be given to the real and complex part of the Dirac distribution. Given this operational meaning of the Dirac distribution, we proceed to use this formalism to make an analogy between classical conditional probabilities and quantum conditional states. In the following, we establish first that Eq. (1) is actually a form of a conditional state by using the framework of Dirac distributions. This allows us then to define a corresponding conditional entropy of post-selected quantum states. Quantum *conditional* states and *conditional* Entropy ====================================================== Quantum states, conditioned on post-selection --------------------------------------------- Using the conditional Dirac distribution obtained in Eq. (\[conditional\]), we now construct the corresponding conditional Dirac decomposition $\rho_{\psi|\phi}$ in analogy to Eq. (\[rhorec\]), given post-selection in some state $\ket{\phi}$. The summation here runs over the eigenstates of a projection operator $\Pi$ measured weakly in between the times of pre-selection and post-selection. We define $\Pi:=\ket{h}\bra{h}$, where $\ket{h}$ is one out of a complete set of basis states $\{\ket{h}\}$ which are mutually non-orthogonal to the state $\ket{\phi}$.[^1] The conditional Dirac decomposition $\rho_{\psi|\phi}$ we obtain is $\rho_{\psi|\phi}$ is the density operator of all paths leading from state $\ket{\psi}$ to post-selected state $\ket{\phi}$. One may note that the operator $\rho_{\psi|\phi}$ is not Hermitian. It has been argued in the weak measurement literature why this should not be a cause of concern [@Lundeen2012]. $\rho_{\psi|\phi}$ thus defined is a trace-one operator and can indeed, by construction, be determined by weak measurements. In the case of the three-box problem, the eigenstates $\ket{h}$ of the projector $\Pi$ would be $\ket{A}$, $\ket{B}$, and $\ket{C}$, representing the system in being in one of the three boxes $A$, $B$, and $C$ at time $t=1$. By writing and inserting it into Eq. (\[condi\]) we obtain which is indeed the generalised density operator for post-selected ensembles, defined in Eq. (\[W\]). This links the interpretation Eq. (\[condi\]) as conditional quantum states to the generalised density operator for post-selected ensembles as in Eq. (\[W\]). We observe that the application of Bayes’ law in the definition (\[condi\]) results in a density operator which is the extension of classical conditional probabilities. This analogy can be further clarified by multiplying $\rho_{\psi|\phi}$ by probability $\Pr(\phi)$ of the system ending in state $\ket{\phi}$ and sum Eq. (\[condi\]) over a complete basis $\{\ket{\phi}\}$. Thus, we retrieve the density operator of Eq. (\[rhorec\]) which contains the full information about the system, with $\Pr(\phi)=Tr[\rho \ket{\phi} \bra{\phi}]$ being the probability of the state being in the state $\ket{\phi}$ at the time of post-selection. This is what is expected from the analogy with classical probabilities. Entropies, conditioned on post-selection ----------------------------------------- In classical information theory, the entropy of random variable $X$ conditioned on selection of a particular instance $y$ of random variable $Y$ is $H(X|Y=y)=-\sum_x \Pr(x|y) \log \Pr(x|y)$, where $\Pr(x|y)$ is the conditional probability of $x$ given the occurrence of $y$. Note that this function is the average $\Big\langle - \log \Pr(x|y)\Big\rangle$. The classical conditional entropy is, again, an average, i.e. $H(X|Y)=\sum_y \Pr(y)H(X|Y=y)$. In analogy to the classical conditional entropy, we define the conditional entropy of a state $\ket{\psi}$, given post-selection in one out of a possible set of states $\ket{\phi}$, to be where $\rho^\dagger_{\psi|\phi}$ is the conjugate transpose of $\rho_{\psi|\phi}$. We note that this particular form of the entropy is an average of the logarithm of the density operator of the system: $$S_c(\psi|\Phi=\phi) = \Big\langle - \log \Big((\rho^{ }_{\psi|\phi}\rho^\dagger_{\psi|\phi})^{1/2}\Big)\Big\rangle~.$$ One could choose other functional forms for the conditional entropy here. For instance, one could choose the function $-\frac{1}{2}\mbox{Tr}[\rho^{ }_{\psi|\phi}\rho^\dagger_{\psi|\phi}{\log}(\rho^{ }_{\psi|\phi}\rho^\dagger_{\psi|\phi})]$ instead. However, this would not correspond to an average of the logarithm. This makes our choice for the conditional entropy function a natural one. Note that the argument of the logarithm gives the Singular values of the operator $\rho_{\psi|\phi}$. The Singular Value Decomposition is a generalisation of diagonalisation needed here due to the particular choice of basis in the definition of the two-state density operator (Eq. \[condi\]). For Hermitian operators, such as ordinary density operators, the singular values and the eigenvalues are the same. Finally, we define the general quantum conditional entropy as an average of the entropy conditioned on a particular choice of post-selection as Eqs. (\[CS\] – \[CS2\]) can be simplified. By inserting Eq. (\[W\]) into Eq. (\[CS\]) we obtain Inserting this into Eq. (\[CS2\]) we obtain the simplified expression This new expression is very useful for finding upper and lower bounds of the conditional entropy. We can see that $S_C$ has a negative lower bound of $\frac{1}{d}\log\frac{1}{\sqrt{d}}$, where $d$ is the dimension of the Hilbert Space. This lower bound is reached when all states $\ket{\phi}$ have the same magnitude in overlap with the state $\ket{\psi}$. $S_C$ is bounded from above by the von Neumann entropy of the system without post-selection. The conditional entropy $S_C$ can be understood as the amount of information contained in a system conditioned on a particular post-selection. The case discussed here is the special case of pre-selection in a pure state which, by definition, has zero von Neumann entropy. The upper bound of the conditional entropy is zero in this case and thus the conditional entropy of a pure pre-selected state $\ket{\psi}$ can be negative. This is a necessary consequence from choosing a subset of the entire (pure) system for post-selection which necessarily decreases the entropy below zero. Thus, a negative conditional entropy of post-selected ensembles becomes intuitive as a subset contains less information than the whole. The concept of negative conditional entropy appears in other contexts in quantum information. For instance, for a bipartite system $\rho_{AB}$ in a pure, i.e. maximally entangled state, the joint entropy is zero. On the other hand, the locally the subsystems $\rho_A$ and $\rho_B$ are in maximally mixed states. Hence, the entropy of one subsystem, conditioned on the measurement of the other, $S(\rho_A|\rho_B)=S(\rho_{AB})-S(\rho_{B})<0$, takes on a negative value. This conditional entropy was also calculated in the form of $S(\rho_A|\rho_B)=-\mbox{Tr}[\rho_{AB} {\log}\rho_{A|B}]$, where $\rho_{A|B}=\lim_{n \to +\infty}[\rho_{AB}^{1/n}(\mathds{1}_A \otimes \rho_B)^{-1/n}]^n$ [@Cerf1996] which is closer in outlook to the construction introduced in Eq. (\[CS2\]). #### Example: 3-box post-selection We will now revisit the 3-box problem [@Aharonov1999] from the beginning and calculate the conditional entropies just defined. Choosing pre- and post-selection as before, i.e. $\ket{\psi}=1/\sqrt{3}(\ket{A}+\ket{B}+\ket{C})$ and $\ket{\phi}=1/\sqrt{3}(\ket{A}+\ket{B}-\ket{C})$, one obtains for the conditional entropy $S_c$ of the system, Eq. (\[CS\]), $S_c(\psi|\Phi=\phi)=-\ln 3$. In order to calculate the conditional entropy $S_C$, Eq. (\[CS2\]), given that the Hilbert space dimension of the system is three, one needs to perform the same calculation as above for two other post-selected states, $\ket{\phi'}$ and $\ket{\phi''}$. The only restrictions on the choice of these two post-selected states are that they need to be mutually non-orthogonal to the state of each box, i.e. to $\ket{A}$, $\ket{B}$ and $\ket{C}$. And in addition they need, together with our original state $\ket{\phi}$, to span the Hilbert space of the state $\ket{\psi}$. Take these states to be and Calculating $S_c$ of Eq. (\[CS\]) for for $\ket{\phi'}$ and $\ket{\phi''}$ gives $S_c(\psi|\Phi=\phi')=-\log_3 4.10$ and $S_c(\psi|\Phi=\phi)=-\log_3 1.10$. With probabilities $\Pr(\phi)=0.11$, $\Pr(\phi')=0.06$ and $\Pr(\phi'')=0.83$ (given by $|\bra{\phi}\psi\rangle|^2$ and correspondingly for $\phi^\prime$ and $\phi^{\prime\prime}$), the conditional entropy Eq. (\[CS2\]) becomes $S_C(\psi|\Phi)=-0.26$, with the logarithm being calculated in base 3 for convenience. The 3-box problem illustrates the negativity of the conditional entropy where each path through a box contains partial information of the system as a whole. #### Conclusion In this work we have defined an entropy for post-selected ensembles. To this end, we have adapted the formalism of Dirac quasi-probabilities, and showed that the two-state density operator of post-selected ensembles is a generalisation of the classical conditional probability, given that a particular final state is selected. We have found upper and lower bounds on such entropies and interpreted the values as the amount of information contained in a system conditioned on a particular post-selection. Furthermore, we showed that these states have properties beyond their classical counterparts. Most notably, conditional quantum entropies as defined in this work can have negative values. This is consistent with our interpretation. This new quantum entropy should open up avenues for studying the properties of weak measurement of more general states than discussed in this paper, such as mixed post-selected states. [**Acknowledgements**]{} We thank Jeff Lundeen, Holger Hofmann and Sandu Popescu for their helpful comments and stimulating discussions. K.W. thanks EPSRC for financial support. [^1]: We use the letter $h$ for the basis states of the projection operator to honour the closeness of these concepts to the consistent histories approach of quantum measurement, see [@Hartle2008].
{ "pile_set_name": "ArXiv" }
--- abstract: | Frohlich, Morchio and Strocchi long ago proved that Lorentz invariance is spontaneously broken in QED because of infrared effects. We develop a simple model where the consequences of this breakdown can be explicitly and easily calculated. For this purpose, the superselected $U(1)$ charge group of QED is extended to a superselected “Sky" group containing direction-dependent gauge transformations at infinity. It is the analog of the Spi group of gravity. As Lorentz transformations do not commute with Sky, they are spontaneously broken. These abelian considerations and model are extended to non-Abelian gauge symmetries. Basic issues regarding the observability of twisted non-Abelian gauge symmetries and of the asymptotic ADM symmetries of quantum gravity are raised. title: Spontaneous Lorentz Violation in Gauge Theories --- Introduction ============ Some time ago, Frohlich, Morchio and Strocchi [@Frohlich:1978bf; @Frohlich:1979uu] and Buchholz [@Buchholz:1986uj] proved that Lorentz invariance is spontaneously broken in QED because of infrared effects. It is broken for the same reason that the Higgs field spontaneously breaks gauge symmetry: the Lorentz group cannot be implemented on the representation space of local observables. In this paper, we develop a model for the calculation of this violation in scattering processes. It is similar to the one we developed earlier for the induced electric dipole moment from QCD $\theta$-angle [@Balachandran:2012bn]. A $(2+1)$-dimensional version of this approach was developed recently in [@Balachandran:2012md]. It also overlaps with the work of Buchholz [@Buchholz:1986uj] and Buchholz and Fredenhagen [@Buchholz:1981fj]. The calculation itself will be presented elsewhere in order to keep the length of the present paper under control. An additional contribution in this work is the extension of the above considerations to non-Abelian gauge theories like QCD where too spontaneous Lorentz violation is seen. In Section 2, we review the theory of gauge invariance. It emerges from the basic contributions of Hanson, Regge and Teitelboim [@Hanson:1976cn] and has been explained before. There are different kinds of transformation groups in gauge theories. There are those implementing the Gauss law and acting as identity on quantum states. The QED and QCD (global) charges also arise from Gauss law, but they may not vanish on quantum state vectors. Rather they serve to define superselection sectors or inequivalent representations of the local algebra of observables. We recall this difference, as it plays a role in our further considerations. In QED, charged sources emit infrared photons. They create a cloud of electromagnetic field at “far” distances, in the “sky” of the source. There are clear similarities between this cloud and the CMB radiation of cosmology. The charge operator does not reflect the possibility of this cloud which can carry net zero electric charge. The proper description of this cloud involves direction-dependent gauge transformations at large distances. The enlarged gauge group which includes such transformations is our Sky group. It is similar to the spatial infinity or Spi group in gravity literature [@Ashtekar:1978zz]. The Sky group of QED serves to define superselection sectors. This abelian Sky group is introduced in Section 3. Section 4 examines Lorentz symmetries in the presence of the Sky group. Unlike $U(1)$, generic elements of the Sky group are not Lorentz invariant, and therefore, we argue that in this sector there is spontaneous Lorentz symmetry breaking. Summing up, our understanding is this: the treatment of infrared photons requires angle-dependent gauge transformations at infinity and the latter break Lorentz symmetry. A concrete and simple model for the calculation of this breakdown in scattering processes does not exist in the literature. For this purpose in section 5 we introduce an algebra generated by two unitaries $U(\chi)$ and $V(\omega)$. The $\chi$ is a function on $\mathbb{R}^3$ with a limit $\chi_\infty$ as the radius $r \rightarrow \infty$: $$\lim_{r \rightarrow \infty} \chi(r, \hat{n}) = \chi_\infty (\hat{n}), \quad \hat{n} \cdot \hat{n} = 1.$$ The $\omega$ is a closed two-form, also with a limit $\omega_\infty$ as $r \rightarrow \infty$: $$\lim_{r \rightarrow \infty} \omega (r,\hat{n}) = \omega_\infty(\hat{n}). \label{omegalimit}$$ They generate a Weyl-like algebra $$U(\chi) V(\omega) = C(\chi_\infty, \omega_\infty) V(\omega) U(\chi), \quad C(\chi_\infty,\omega_\infty) = {\rm central\,\,elements}. \label{sky_algebra}$$ The meaning of these operators is this: $U(\chi)$ represent the elements of the Sky group while $V(\omega)$ intertwines its different representations. The algebra (\[sky\_algebra\]) is only a modification of a Buchholz-Fredenhagen algebra [@Buchholz:1981fj] introduced by them also to discuss gauge theories. Local observables cannot affect $\chi_\infty$ or $C$ as they depend on data at infinity. We see in section 5 that the Lorentz group does change it, and so gets spontaneously broken. If $|\cdot\rangle_{C=0}$ is a vector state on which $U(\chi)$ is $\mathbb{1}$, then $$V(\omega) |\cdot \rangle_0 = |\cdot \rangle_C \label{vertex}$$ is a vector state on which $U(\chi)$ has value $C$. We show in Section 5 how to see the effect of $C$ in scattering processes by a simple modification of the fermion mass. We note that we do not prescribe a choice for $C$ here. It is determined by the details of the scattering process and perhaps can be modeled from earlier work on the infrared problem. We hope to address this issue in our follow-up paper [@bqv]. There is an alternative choice for $V(\omega)$, call it $W(\theta)$, which commutes with the Gauss law and intertwines different eigenspaces of the (global) charge operator. It is $$W(\theta) = e^{i \theta S_{CS}(A)}$$ where $S_{CS}(A)$ is the abelian Chern-Simons action for the spatial slice of QED. For the fermion mass term modified as previously, the vector states $$|\cdot \rangle'_\theta = W(\theta) |\cdot \rangle'_{\theta=0}$$ also lead to Lorentz symmetry violation (The prime is to distinguish these vectors from those in (\[vertex\])). But the algebra with relation (\[sky\_algebra\]) is modified. The twist $V(\omega)$ reverses to $V(-\omega) = V(\omega)^{-1}$ under ${\cal C}$ and ${\cal P}$ and is ${\cal T}$-invariant, while $W(\theta)$ is ${\cal C}$-invariant while reversing under ${\cal P}$ and ${\cal T}$. Thus $|\cdot \rangle_\chi$ for $\chi \neq 1$ breaks ${\cal C}$ and ${\cal P}$ while $|\cdot \rangle_\theta$ for $\theta \neq 0$ breaks ${\cal P}$ and ${\cal T}$. Both keep ${\cal CPT}$ intact. Section 6 takes up non-Abelian gauge groups. The analog of $U(\chi)$ exists here as well and serves to define the non-Abelian Sky group. But a $V(\omega)$ commuting with Gauss law does not exist. Instead, we must perforce use the non-Abelian $W(\theta)$ where $S_{CS}(A)$ is the non-Abelian Chern-Simons action for the spatial three-manifold. We get a new algebra from $U(\chi)$ and $W(\theta)$ and a new model for Lorentz violation. With Section 7, we close with qualitative remarks on [*non-Abelian*]{} gauge invariance. Thus for example observables commute with [*all*]{} gauge transformations. For this reason, if the gauge bundle is twisted such as in QCD, it immediately confronts us with issues like the “problem of color” [@Nelson:1983bu; @Balachandran:1983xz; @Balachandran:1983fg; @Nelson:1983fn] and mixed states [@Balachandran:2011bv]. In quantum gravity where diffeos assume the role of the gauge group, and observables commute with them, asymptotic symmetries, including the Poincaré group cease to be observable.Only their center, like mass and total spin can be measured. Thus for a quantum black hole, it is impossible to measure the probability of finding the component of angular momentum in a generic direction in any exact theory of quantum gravity. We explain these points in conclusion. QED: The Gauss Law and the Charge ================================= In the canonical approach to QED, the equal-time vector potentials $A_i$ and the electric fields $E_i$ are canonically conjugate: $$[A_i(\vec{x},t), E_j(\vec{y},t)] = i \delta_{ij} \delta^3(x-y).$$ In elementary treatments, the classical Gauss law is imposed as a constraint on quantum vector states $|\cdot\rangle$: $$(\partial_i E_i + j_0) |\cdot \rangle =0, \quad j_0 = {\rm charge\,\,density}.$$ It has been understood for a long time that this formulation of quantum Gauss law is unsound: $A_i$ and $E_j$ are operator-valued distributions so that their derivatives must be reformulated in terms of test functions, their duals. For test functions, we can pick $C_0^\infty(\mathbb{R}^3)$, which is the space of infinitely differentiable functions on $\mathbb{R}^3$ vanishing at $\infty$. Note that we do not require the derivatives of functions to vanish at infinity. Then the quantum Gauss law is $$G(\Lambda) |\cdot \rangle = \int d^3 x [-(\partial_i \Lambda) E_i + \Lambda j_0](\vec{x},t) |\cdot \rangle =0, \quad \forall \Lambda \in C_0^\infty(\mathbb{R}^3). \label{qGausslaw}$$ For classical fields, we can here partially integrate without generating boundary terms, getting $$\int d^3 x \Lambda(x)(\partial_i E_i + j_0)(\vec{x},t) = 0, \quad \forall \Lambda \in C_0^\infty(\mathbb{R}^3),$$ which can be interpreted as the classical Gauss law. Let us now enlarge $C_0^\infty (\mathbb{R}^3)$ to a larger space $C^\infty(\mathbb{R}^3;S^2_\infty)$. It consists of $C^\infty$-functions which have an angle-dependent limit $\chi_\infty$ as the radius variable $r$ goes to infinity: $$\chi \in C^\infty(\mathbb{R}^3;S^2_\infty) \implies \lim_{r \rightarrow \infty} \chi(\vec{x}) = \chi_\infty (\hat{n}), \quad \hat{n} \cdot \hat{n} =1.$$ It is possible to interpret this space as the space of functions after blowing up infinity, as in the treatment of Spi group. We are in effect attaching a sphere $S^2_\infty$ at infinity, with its points $\hat{n}$. Clearly, $$C^\infty_0(\mathbb{R}^3) \subset C^\infty(\mathbb{R}^3;S^2_\infty)$$ We can now also consider the operators $$Q(\chi) = \int d^3x \big(-\partial_i \chi E_i + \chi j_0 \big)(\vec{x},t ) \label{charge}$$ There is no reason for $Q(\chi)$ to vanish on physical states unless $\chi \in C^\infty_0(\mathbb{R}^3)$. In that case $Q(\chi)$ becomes $G(\chi)$. In quantum theory, it is $Q(\chi)$ which gives the charge operator $Q_0$ after the condition $\chi_\infty (\hat{n}) =1$: $$Q_0 = Q(\chi)|_{\chi_\infty(\hat{n})=1}.$$ The global $U(1)$ group is generated by $Q_0$. The Sky Group ============= We generalize the global $U(1)$ group to a bigger “Sky” group, call it Sky, with generators $Q(\chi)$. Of course only its quotient by Gauss law group can act nontrivially on the quantum states. So what we mean by Sky is the quotient of the group generated by $Q(\chi)$ by the Gauss law group.\ [*The Need for Sky*]{}\ In QED, infrared photons of arbitrarily small wavelengths are emitted by the sources. Even if at some time $t=-T$, we assume that there are no such photons for $r>R$, and that charge densities are compactly supported, the photons they emit will reach the $t=0$ surface for $r>R$ if $T$ is large enough (see Figure 1). So we cannot rule out radiation field at arbitrarily large distances.\ ![Photons from the distant past arrive at the $t=0$ surface at large distances.](lightcone.eps){height="9cm"} But then, for their proper description using $A_i$ and to treat gauge invariance, we should allow gauge transformations at arbitrarily large distances. Neither the Gauss law $G(\Lambda)$ nor the charge $Q_0$ generate these transformations. So we expand global $U(1)$ to Sky. We work in the $A_0=0$ gauge. For this reason, we need not introduce time dependence in the elements of the Sky group.\ [*Sky is superselected*]{}\ [*All*]{} observables must preserve the physical state condition (\[qGausslaw\]). If ${\cal A}$ is the algebra of observables and $\alpha \in {\cal A}$, we thus have $$[G(\Lambda), \alpha]=0.$$ We require that ${\cal A}$ is local as well. A good way to explain locality in this context is as follows. In a Lagrangian quantum field theory, fields $\phi$ and their conjugate momenta $\pi$ all commute at distinct spatial points at equal times: $$[\phi(\vec{x},t),\pi(\vec{y},t)]=0.$$ We can define local operators as their spatial integrals at a fixed time with test functions of compact spatial support. Locality implies that Sky commutes with ${\cal A}$: it lies in the commutant ${\cal A}'$ of ${\cal A}$: $${\rm Sky} \subseteq {\cal A'}$$ To see this, let $\alpha \in {\cal A}$ be localized on a compact set $K$. Now for every $\chi$, we can find a $\Lambda$ such that $\chi$ and $\Lambda$ coincide on $K$: $$\chi \restriction K = \Lambda \restriction K$$ which means that $$Q(\chi \restriction K) = G(\Lambda \restriction K). \label{QisG}$$ But since $\alpha$ is localized in $K$, by definition, $$\begin{aligned} [Q(\chi), \alpha] &=& [Q(\chi \restriction K), \alpha], \\ \protect{[}G(\Lambda), \alpha] &=& [G(\Lambda \restriction K), \alpha].\end{aligned}$$ Using (\[QisG\]), we then get $$[Q(\chi), \alpha] = 0.$$ This means that $Q(\chi)$ are superselected. They play the role of Casimir operators for Lie group representations. Their distinct eigenvalues label inequivalent representations of local algebras of observables. The Poincaré Group as Global Symmetry ===================================== The generators of Poincaré group, just as those of any global symmetry, involve integrations of local fields over all space. They do not commute with [*all*]{} local fields localized in any compact spatial region. But as the group they generate preserves the algebra of local observables, they form an automorphism of this algebra. But such an automorphism may or may not be implementable in a particular representation of this algebra. A simple example can illustrate this point. The group $SU(3)$ and its group algebra $\mathbb{C}SU(3)$ admit complex conjugation or charge conjugation as its automorphism. Now $SU(3)$ (and therefore $\mathbb{C} SU(3)$) has two inequivalent three-dimensional irreducible representations (IRRs), the [**3**]{} and the ${\bf \bar{3}}$. The above automorphism exchanges them. If we consider just the IRR [**3**]{}, then the above automorphism cannot be implemented on the representation space. Now going back to the charge operator $Q_0$, it is Poincaré invariant because current is conserved. The proof relies on the fact that if $g$ is a Poincaré transformation and $g^*\chi$ is the Lorentz transform of the test function $\chi$ with constant value $\tilde{\chi}$ at infinity, then $g^* \chi$ has the same value $\tilde{\chi}$ at infinity. We conclude that the Poincaré group respects charge superselection rule. The latter places no obstruction to its unitary implementation. For theories with mass gap, there is no further obstruction. The Poincaré group in such theories is not spontaneously broken. But for the full Sky group, the conclusions are different. If $\chi_\infty$ has non-trivial $\hat{n}$ dependence, and $R$ is a rotation, then $(R^* \chi)_\infty (\hat{n}) = \chi_\infty(R^{-1} \hat{n}) \neq \chi_\infty(\hat{n})$. Hence rotations, and therefore Lorentz transformations do not commute with those of Sky. Spacetime translations in contrast are Sky-invariant. For example for spatial translations, this follows from the fact that $$\frac{(\vec{x}+\vec{a})}{|\vec{x} + \vec{a}|} \stackrel{r \rightarrow \infty}{\longrightarrow} \frac{\vec{x}}{|\vec{x}|} = \hat{n}.$$ As for time translations, since the Hamiltonian density $${\cal H}(x) = \frac{1}{2} \Big( F_{0i}^2 + \frac{1}{2} F_{ij}^2 \Big) + \cdots$$ itself is gauge invariant and commutes with $Q(\chi)$, elements of the Sky group also are time-translation invariant. We conclude that unless $$Q(\chi) |\cdot \rangle =0 \label{g_invariant}$$ for $\chi_\infty$ not constant on $S^2_\infty$, the Lorentz group is spontaneously broken. The Sky Changes the $S$-matrix ============================== Let us start with quantum states $|\cdot \rangle_0$ carrying the trivial representation of Sky where (\[g\_invariant\]) is fulfilled. We will now construct the operator $V(\omega)$ of Section 1 which acting on $|\cdot\rangle_0$ creates the vector state $$|\cdot \rangle_\omega = V(\omega) |\cdot \rangle_0$$ which carries a non-trivial Sky-representation, while at the same time it commutes with $G(\Lambda)$ and respects Gauss law. With $V(\omega)$ at our disposal, there is a simple modification of the charged fermion-mass term which picks up the effect of $\omega$ and gives a Lorentz-violating $S$-matrix. Gauss law and charge conservation are maintained by the new interaction, while if $\omega=0$ and there is empty Sky, it does not affect scattering either. Let $\omega$ be a closed two-form, $$d \omega = 0,$$ and consider $$V(\omega) = \exp i \left(Q_0 \int \omega \wedge A \right). \label{Vdef}$$ In the above, $Q_0$ commutes with $\int \omega \wedge A$. That is because on ${\mathbb R}^N$, any closed form is exact, and hence the integral of $\omega$ over any two-sphere is 0. So there is no ordering ambiguity in (\[Vdef\]). The presence of $Q_0$ in $V(\omega)$ ensures that in any sector with net zero charge, $V(\omega)$ is identity and hence there is no Lorentz violation [@Frohlich:1978bf]. $G(\Lambda)$ commutes with $V(\omega)$ since $$\int \omega \wedge d \Lambda = -\int d \omega \wedge \Lambda = 0$$ for $\Lambda \in C_0^\infty$. Hence if $|\cdot \rangle$ is annihilated by $G(\Lambda)$, then so is $V(\omega |\cdot \rangle$: $$G(\Lambda) V(\omega) |\cdot \rangle = 0.$$ This means that $V(\omega)$ maps admissible vectors compatible with Gauss law to other such admissible vectors. But if $$U(\chi) = e^{i Q(\chi)} \in {\rm Sky}$$ we see that $$U(\chi) V(\omega) = e^{-i Q_0 \int d \chi \wedge \omega} V(\omega) U(\chi) = C(\chi_\infty, \omega_\infty) V(\omega) U(\chi),$$ where $C$ depends only on $\chi_\infty$ and $\omega_\infty$ since the integral $\int \omega \wedge d \chi$ depends only on them. This generalized “Weyl" algebra is very similar to the Buchholz-Fredenhagen algebra [@Buchholz:1981fj]. Hence $V(\omega)$ maps an empty Sky $|\cdot\rangle_0$ to a non-empty one $|\cdot \rangle_\omega$. The operator $V(\omega)$ is not a local observable, nor can it be identified as an observable in some sense, as it does not commute with Sky. It is like a charged field, but it commutes with $Q_0$.\ [*Remark*]{}\ $V(\omega)$ does not create charge as follows from (\[omegalimit\],\[Vdef\]) $$Q_0 V(\omega) |\cdot \rangle = V(\omega) Q_0 |\cdot \rangle$$ We are after a simple model to calculate the effects of Lorentz breaking. There is in fact a simple approach to such a model. It was devised earlier [@Balachandran:2012bn] to calculate electric dipole moment from QCD $\theta$ and can be adapted for the present purpose. Such a model is one where the fermion mass term ${\cal L}^0_m:= m \bar{\Psi}\Psi$ is modified to $${\cal L}^\chi_m = \frac{m}{2} \big(U(\chi) + U(\chi)^\dagger\big) \bar{\Psi}\Psi.$$ Since $$_\omega \langle \cdot | U(\chi) |\cdot \rangle_\omega = C(\chi_\infty, \omega_\infty) _0\langle \cdot | U(\chi) |\cdot \rangle_0 = C(\chi_\infty, \omega_\infty)\,\, _0\langle \cdot |\cdot \rangle_0,$$ the amplitudes for the vectors $|\cdot \rangle_\omega$ are obtained from the amplitudes for the vectors $|\cdot \rangle_0$ if we replace $m$ by $$M(C(\chi_\infty,\omega_\infty)) \equiv m \Re C(\chi_\infty, \omega_\infty) = m \cos \Big( Q_0 \int d \chi \wedge \omega\Big).$$ We can then set $U(\chi)=1$, which is its value on $|\cdot \rangle_0$. The Sky-deformed mass $M(C(\chi_\infty, \omega_\infty))$ depends in general on $\hat{n}$. Because of this dependence, it can give scattering amplitudes that violate Lorentz invariance. Its choice has to be dictated by the infrared cloud of a specific process. For instance, $\chi$ could be chosen to be the profile of the coherent state of soft radiation that accompanies incoming or outgoing charged particles. The explicit form of such is profile has been derived in [@Eriksson:1970dc].\ [*An Alternative Choice for $V(\omega)$*]{}\ Instead of $V(\omega)$, we could have chosen any other unitary operator commuting with $G(\Lambda)$, but not commuting with a general $Q(\chi)$. There is one particularly interesting choice of this sort. It is the one which generalizes to the non-Abelian case. It is $$W(\theta) = e^{i \theta S_{CS}(A)}$$ where $$S_{CS} = \frac{1}{8\pi^2}\int A \wedge d A$$ is the Chern-Simons term. Then $$U(\chi) W(\theta) = e^{-\frac{i \theta}{8 \pi^2} \int d \chi \wedge dA} W(\theta) U(\chi) := C(\chi_\infty, \theta) W(\theta) U(\chi).$$ As before we can consider $$\begin{aligned} W(\theta) |\cdot \rangle_0 &=& |\cdot \rangle_\theta, \\ U(\chi) |\cdot \rangle_0 &=& |\cdot \rangle_0.\end{aligned}$$ The mass term ${\cal L}^0_m$ between the $\theta$-sectors is equivalent to the mass term $${\cal L}^\theta_m = M(\theta) \bar{\Psi} \Psi, \quad M(\theta) = m \cos \Big(\frac{\theta}{8\pi^2} \int d \chi \wedge dA \Big). \label{csmass}$$ in the $\theta=0$ sector. Since $dF=0$, this too depends only on asymptotic data. For the $A$ which occurs in (\[csmass\]) as well, such data must come from the state of the infrared photons. Non-Abelian Generalization ========================== The Gauss law and $Q(\chi)$ easily generalize to non-Abelian semi-simple compact Lie groups $H$. It is enough to replace $\Lambda$ and $\chi$ by functions valued in its Lie algebra $\underline{H}$ and take traces at appropriate spots. Thus for instance $$Q(\chi) = \int d^3 x \Big(- {\rm Tr} (D_i \chi) E_i + \chi J_0 \Big), \quad \chi = \chi^\alpha \lambda_\alpha, \quad \chi^\alpha \in C^\infty(\mathbb{R}^3;S^2_\infty).$$ Here $\lambda_\alpha$ is a basis for say the adjoint representation of $\underline{H}$. When $\chi^\alpha = \Lambda^\alpha \in C^\infty_0 (\mathbb{R}^3)$, we get the Gauss law operators $G(\Lambda)$. They vanish on physical quantum states. The argument that local observables commute with $Q(\chi)$ is still valid so that the non-Abelian Sky group they generate is superselected. If $Q(\chi)|\cdot \rangle \neq 0$ when $\chi_\infty$ has a non-trivial angular dependence on $\hat{n}$, then this superselection sector is not Lorentz-invariant. Lorentz group is spontaneously broken. All this is as before. But the operator $V(\omega)$ does not generalize to the non-Abelian case as its naive generalization does not commute with $G(\Lambda)$. Only $W(\theta)$ does so, with $S_{CS}(A)$ being the non-Abelian Chern-Simons term: $$\begin{aligned} W(\theta) &=& e^{i \theta S_{CS}(A)}, \\ S_{CS}(A)&=& \frac{1}{8\pi^2}{\rm Tr} \Big( A \wedge d A + \frac{2}{3} A \wedge A \wedge A \Big).\end{aligned}$$ This operator commutes with $G(\Lambda)$, but not with $Q(\chi)$ for a generic $\chi$: $$U(\chi) W(\theta) = C(\chi_\infty, \theta,A) W(\theta) U(\chi), \quad U(\chi) = e^{i Q(\chi)}, $$ where the explicit expression for $C(\chi_\infty, \theta,A)$ can be read off from Eq. (13.59) of [@Nair:2005iw]. It simplifies if $\chi^\alpha$ is non-zero for only one value $\alpha_0$ of $\alpha$. Any gauge transformation induced by $e^{i Q(\chi)}$ is a product of such transformations. For the mass-deformed model, the deformed mass is a generalization of (\[csmass\]) $$M(\theta) = m \cos \Big(\int \Re C(\chi_\infty, \theta, A)\Big). \label{nabmass}$$ It comes from absorbing the twist $W(\theta)$ of the twisted quantum state $$|\cdot \rangle_\theta = W(\theta) |\cdot \rangle_0$$ in the mass term. The twisted sector of QCD with twist $\theta$ is precisely the QCD $\theta$-sector.It is the one where quarks and baryons acquire electric dipole moment. In the absence of Sky, there is $2\pi$-periodicity on $\theta$. But that is absent in $M(\theta)$: Lorentz violation is not periodic in $\theta$. In addition to modifying $m \bar{\Psi}\Psi$ by $U(\chi)$, we can also modify it by another operator discussed in [@Balachandran:2012bn]. That will let us calculate electric dipole moment as well. This operator too comes from a gauge principle and has shared properties with $U(\chi)$. They occur together in the twisted mass and mutually affect each other. Final Remarks ============= There are particular features of [*non-Abelian*]{} gauge groups which merit more study. Let us focus on the subgroup of the non-Abelian Sky where $\chi_\infty$ is a constant function. We can write $$\chi_\infty(\hat{n}) = \chi_\infty^\alpha \lambda_\alpha, \quad \chi_\infty^\alpha \,\,{\rm constants}.$$ Correspondingly, $$Q(\chi) = \chi^\alpha_\infty Q(\lambda_\alpha) := \chi^\alpha_\infty Q_\alpha$$ where $Q_\alpha$ generate the non-Abelian global group $H$. (Only the asymptotic value of $\chi$ matters because of the Gauss law.) In QCD, $H$ is the color group $SU(3)_c$. It is what is normally studied in literature. So we can concentrate on $H$ and assume without loss of generality that $H$ is a simple compact Lie group. But $H$ is superselected. So are $Q_\alpha$. $H$ is also by assumption non-Abelian. Therefore not all elements $h = e^{i \phi_\alpha Q_\alpha}$ of $H$ can characterize a superselection sector. That is because local observables commute with $H$. So the vector states with the same value of a fixed $h_0 = e^{i \phi^0_\alpha Q_\alpha}$ are invariant under local observables. This value is part of the characterization of a superselection sector. But a generic $h \in H$, $H$ being non-Abelian, changes this value and hence the superselection sector. Such $h$ changing this value are “spontaneously broken”. The choice of a particular value of $h_0$ is like choosing a direction for the Higgs field. Thus unlike the abelian $U(1)$ of electric charge or its group algebra $\mathbb{C}U(1)$, the full color group $SU(3)$ or its group algebra $\mathbb{C}SU(3)$ for instance cannot be used or implemented in a superselection sector. Global color cannot participate as a symmetry group in QCD.\ [*Remarks*]{}\ The above features of non-Abelian gauge groups were first encountered in the problem of color [@Balachandran:1983xz; @Balachandran:1983fg]. It was subsequently understood as a generic feature of non-Abelian twisted bundles (see [@Balachandran:1991zj; @Balachandran:1991ea]). It is encountered in numerous physical systems: molecules, non-Abelian monopoles, QCD, ... including when $H$ is a discrete group as for molecules [@Balachandran:1991ea]. But we can use a maximal abelian subalgebra ${\cal A}_c$ of $\mathbb{C}H$ ($c$ for commutative) for labeling the superselection sectors.. If $H$ is a simple Lie group and $T^k$ is the maximal torus generated by its Cartan subalgebra, then it can be the algebra generated by the center $\mathcal{C}(\mathbb{C}H)$ of $H$ and the algebra $\mathbb{C}T^k$ of $T^k$. Thus ${\cal A}_c$ contains the Casimir invariant. Now ${\cal A}_c$ is [*commutative*]{}, so its irreducible representations are one-dimensional. Elsewhere [@Balachandran:2013kia] we have discussed this issue further. Here we summarize the main points. If the $H$-bundle is trivial, we can show that pure states exist on the local algebra ${\cal A}$, but they are obtained by tracing out the $H$-degrees of freedom as ${\cal A}$ is $H$-blind. But if the $H$-bundle is twisted, the states obtained from a superselection sector and a GNS construction are all mixed. We will also show that the QCD bundle is twisted. Similar reasoning can be applied to quantum gravity. Previous work [@Balachandran:2011gj] treated mapping class groups partially from this perspective. Also in asymptotically flat spacetimes, the Poincaré group emerges as the analogue of the global gauge groups above in the ADM formalism. It is [*non-Abelian*]{} so that it is [*spontaneously broken*]{} to the analogue of ${\cal A}_c$. What this means requires more work to understand adequately. But it does indicate that a quantum black hole has features different from its classical description. For example, suppose we decide on $J_3$, the third component of angular momentum, as one unbroken generator creating an unbroken circle group in ${\cal A}_c$. But then all the $\hat{n} \cdot \vec{J}, \hat{n} \neq (0,0,1)$ are spontaneously broken, and $J_3$ has a fixed numerical value in a superselection sector. No observable internal to the black hole can change it. It is not then clear if it makes quantum sense to interrogate if the black hole is spinning in a direction distinct from the $3$-direction, and with what angular velocity. The role of cluster decomposition, so important to analyze such issues for electric charge [@Haag:1992hx] has not been analyzed in quantum gravity. We remark that in QED, electric charge density $j_0$ is gauge invariant (commutes with $Q(\chi)$) and can be turned into a local observable. So there is quantum meaning for claims of measurement of charge in a finite volume $V$: it is the measurement of $$\int_V d^3 x \,j_0,$$ a local observable. But for non-Abelian $H$, $Q_\alpha$ are [*not*]{} integrals of fully gauge invariant local densities $q_\alpha$ ($q_\alpha$ do not commute with all $Q(\chi)$). The same goes for say $J_3$ in quantum gravity: such observables are not integrals of local diffeomorphism-invariant densities. Thus while local measurements of electric charge density are possible, that is not so for non-Abelian $H$-charge densities or their diffeomorphism analogues in gravity. That already poses problems for us to adapt Haag’s ideas (“positron behind the moon") to ${\cal A}_c$ or its gravity analogue, even though this algebra is abelian.\ [*Remark on Covariantization*]{}\ In the text, we have used the fixed-time formalism all the way. That leads to a “blow-up" of spatial infinity where a sphere $S^2_\infty$ gets attached. This is enough to treat all but Lorentz boosts. But Lorentz boosts cannot act on such an infinity. We can accommodate the latter by attaching the $2+1$ de Sitter space, which carries the action of the Lorentz group, at spatial infinity. That this must be so can be inferred from the analysis of the Spi group [@Ashtekar:1978zz]. But for a satisfactory treatment, we need the covariant Peierls form of commutation relations [@Peierls:1952cb; @DeWitt:2003pm; @Bimonte:2003cx]. For free abelian electromagnetism, this can be written down compatibly with gauge invariance. For non-Abelian gauge theories including QCD, gauge invariance requires us to deal with an interacting non-linear theory. But since the connected component of the Lorentz group ${\cal L}_+^\uparrow$ is simple, if rotations are spontaneously broken, so is ${\cal L}_+^\uparrow$. Thus indirectly, we can infer the fate of ${\cal L}_+^\uparrow$. We may finally note that the analysis in this paper can be readily adapted to all spacetimes of dimension $\geq2+1$. The $2+1$ case has in fact already been treated in [@Balachandran:2012md]. Acknowledgments =============== APB was supported by the Institute of Mathematical Sciences, Chennai and by Instituto Internacional de Fisica, Natal, Brasil. [99]{} J. Frohlich, G. Morchio and F. Strocchi, Annals Phys.  [**119**]{}, 241 (1979). J. Frohlich, G. Morchio and F. Strocchi, Phys. Lett. B [**89**]{}, 61 (1979). D. Buchholz, Phys. Lett. B [**174**]{}, 331 (1986). A. P. Balachandran, T. R. Govindarajan and A. R. de Queiroz, Eur. Phys. J. Plus [**127**]{}, 118 (2012) \[arXiv:1204.6609 \[hep-th\]\]. A. P. Balachandran, S. Kurkcuoglu and A. R. de Queiroz, Mod. Phys. Lett. A [**28**]{}, 1350028 (2013) \[arXiv:1208.3175 \[hep-th\]\]. D. Buchholz and K. Fredenhagen, Commun. Math. Phys.  [**84**]{}, 1 (1982). See Section 9. A. J. Hanson, T. Regge and C. Teitelboim, “Constrained Hamiltonian Systems,” RX-748. A. Ashtekar and R. O. Hansen, J. Math. Phys.  [**19**]{}, 1542 (1978). A. P. Balachandran, A. R. de Queiroz and S. Vaidya (in preparation). P. C. Nelson and A. Manohar, Phys. Rev. Lett.  [**50**]{}, 943 (1983). A. P. Balachandran, G. Marmo, N. Mukunda, J. S. Nilsson, E. C. G. Sudarshan and F. Zaccaria, Phys. Rev. D [**29**]{}, 2919 (1984). A. P. Balachandran, G. Marmo, N. Mukunda, J. S. Nilsson, E. C. G. Sudarshan and F. Zaccaria, Phys. Rev. D [**29**]{}, 2936 (1984). P. C. Nelson and S. R. Coleman, Nucl. Phys. B [**237**]{}, 1 (1984). A. P. Balachandran and A. R. de Queiroz, Phys. Rev. D [**85**]{}, 025017 (2012) \[arXiv:1108.3898 \[hep-th\]\]. K. E. Eriksson, Phys. Scripta [**1**]{}, 3 (1970). A. P. Balachandran, G. Marmo, B. S. Skagerstam and A. Stern, [*Classical Topology and Quantum States*]{}, Singapore, Singapore: World Scientific (1991) A. P. Balachandran, A. Simoni and D. M. Witt, Int. J. Mod. Phys. A [**7**]{}, 2087 (1992). A. P. Balachandran, A. R. de Queiroz and S. Vaidya, Phys. Rev. D [**88**]{}, 025001 (2013) \[arXiv:1302.4924 \[hep-th\]\]. A. P. Balachandran and A. R. de Queiroz, JHEP [**1111**]{}, 126 (2011) \[arXiv:1109.5290 \[hep-th\]\]. V. P. Nair, [*Quantum field theory: A Modern Perspective*]{} New York, USA: Springer (2005) R. Haag, [*Local Quantum Physics: Fields, Particles, Algebras*]{}, Berlin, Germany: Springer (1992) R. E. Peierls, Proc. Roy. Soc. Lond. A [**214**]{}, 143 (1952). B. S. DeWitt, [*The Global Approach to Quantum Field Theory, Vols 1 and 2*]{}, Oxford, UK: Clarendon (2003). G. Bimonte, G. Esposito, G. Marmo and C. Stornaiolo, Int. J. Mod. Phys. A [**18**]{}, 2033 (2003) \[hep-th/0301113\], and references therein.
{ "pile_set_name": "ArXiv" }
--- abstract: 'As part of a project to map nearby Seyfert galaxies in the near-IR with adaptive optics, we present high spatial resolution, near-IR images in J,H, and K of the nuclei of NGC 3227 and NGC 2992, obtained with the Adaptive Optics Bonnette (AOB) on CFHT. The $\sim0.15"$ resolution allows us to probe structures in the core region at unprecedented scales. With NGC 3227, we are able to identify an inwards spiraling starburst in all three near-IR bands. NGC 2992 shows evidence for emission along a radio loop. Compared with HST optical images, dust obscuration becomes significantly less pronounced at longer wavelengths, revealing the true geometry of the core regions. The observed structures may help to elucidate fueling mechanisms for the central engine, as well as providing insight into the unification paradigms. The results are tempered with the discovery of AOB-related artifacts in the central arcsec of observed AGN galaxies which take on the appearance of spiral/disk structures.' author: - 'Scott C. Chapman, Gordon A. H. Walker' - 'Simon L. Morris' title: | The core structure of AGN:\ perils of adaptive optics artifacts --- INTRODUCTION {#intro} ============ It is generally accepted that pronounced activity in galaxies hosting Active Galactic Nuclei (AGN) results from accretion onto a supermassive black hole. This paradigm has led to a plethora of research into AGN, of which the problems of overcoming the angular momentum barrier to fuel the nucleus and unification of the AGN types have risen to the forefront as especially vexing and controversial. Near-IR imaging has proven to be a powerful means to study these AGN problems since the dust extinction is reduced, and the contrast between the central AGN and the underlying stellar population is improved. A fresh perspective can be gained on the core regions of AGNs through the high resolution images now possible thanks to adaptive optics (AO). We have begun a project to map nearby Seyfert galaxies in the near-IR with adaptive optics in order to study core morphologies and address possible fueling mechanisms. Here, we present observations of NGC 3227 and NGC 2992 obtained with the AO system on the 3.6 m Canada-France-Hawaii Telescope. Both galaxies live in disturbed environments:\ $\bullet$ NGC 3227 is an SABa galaxy classified as Sy1.5, interacting with its dwarf elliptical neighbor, NGC 3226. It has been much studied in recent years as it contains many of the elements thought to be related to the formation and evolution of active nuclei: emission line regions excited by both starburst and AGN continuum, strong interaction, and a stellar bar (Gonzales Delgado & Perez 1997, Arribas & Mediavilla 1994).\ $\bullet$ NGC 2992 is an Sa galaxy seen almost edge-on and interacting with NGC 2993. It possesses an active Seyfert nucleus classified as Sy1.9. A large and prominant dust lane runs through the center of the galaxy roughly north to south, splitting the nuclear region in two. Ulvestad and Wilson (1984) found that the radio structure of the nucleus of NGC 2992 has the shape of a “figure-8”, with a maximum extent of about 2000 pc, oriented out of the plane of the galactic disk. Most of of the 6cm radio emission from the center of the galaxy arises in the loops of the figure-8 rather than in the nucleus. OBSERVATIONS {#sec2} ============ The imaging observations were obtained at the CFHT in March, 1997, using the MONICA near-IR camera (Nadeau et al. 1994) mounted at the f/20 focus of the Adaptive Optics Bonnette (AOB). The detector is a Rockwell NICMOS3 array with 256x256 pixels and a 0.034“/pix scale. The CFHT AOB is based on curvature wavefront sensing (Roddier 1991), and uses a 19 zone bimorph mirror to correct the wavefront distortions. As the field size is small (9”x9“), blank sky images were taken intermitantly between science frames. On-source images were taken in a mosaic of 4 positions, alternately putting the galaxy core in each of the four quadrants of the array. Flux and PSF calibrations were performed using the UKIRT standard stars fs13 and fs25. Flat-field images were taken on the dome with the lamps turned on and off to account for the thermal glow of the telescope. The nuclei of the galaxies themselves were used as the guiding source for the AO system, a clear point source for NGC 3227 and a more extended core for NGC 2992. The natural seeing averaged 0.6”-0.8“ throughout the observations resulting in relatively high strehl ratios in all bands, and FWHM of 0.14”, 0.17“, 0.22” at K, H, J bands respectively. At a distance of 15Mpc for NGC3227, 1“=76pc using $H_0=50$. The estimated distance of NGC2992 at 20Mpc implies an angular scale of 100pc / 1”. THE COMPLETE SAMPLE {#sec3} =================== Candidate galaxies for our survey were chosen based primarily on core guidability with AOB. Our sample galaxies therefore tend to have the brightest cores amongst nearby Seyferts (though not necessarily unresolved) and may not be representative of the lower luminosity active nuclei. However, our Seyferts typically have bright X-ray flux and are perhaps the best sample with which to study the AGN phenomenon, since they are uncontaminated with borderline AGN galaxies. There are roughly equal numbers of Seyfert type 1 and 2 galaxies, since a bright extended core seems to provide a good guiding source with curvature sensing adaptive optics. We note however that the types of artifacts around the central AGN spike vary with the core morphology, and identifying physical structures is a nontrivial matter within the central arcsecond. The complete sample is presented in Table 1, along with alternate designations, redshift, filters obtained, Seyfert class (S=saturated core, U=unresolved core, R=resolved core; all Sy2 are resolved; (S) = HST saturated core); galaxy classification, and camera used: KIR or MONICA. AOB artifacts - the difficulty of identifying physical structure ---------------------------------------------------------------- The exciting results on AGN cores emerging from AOB studies are tempered with the discovery of achromatic, AOB-related artifacts in the central arcsec surrounding the bright nuclei of observed AGN galaxies. These artifacts masquerade convincingly as the expected types of morphologies in the centers of Seyfert galaxies: small-scale spiral arms, edge-on disks or tori, double nuclei, and outflows. Figure 1 shows a representative sample of AGN cores. Elliptical isophotes were fit to the galaxy core region. A smooth model was then built from the resulting fit, and subtracted from the raw image. Alternatively, median filtered images were subtracted from the core with the same result. The most pronounced effect in all the raw images is an extended core to the North. Figure 1 clearly depicts that these structures exist at the same scale with similar shape and orientation relative to North, regardless of Seyfert type or distance to the galaxy. In particular NGC3998, NGC3516, and Mkn348 all show a cross-like pattern suggestive of small-scale spiral arms plus a bar or edge-on torus. The K-band image of NGC3998 is displayed in Figure \[artfig2\], and shows the effect of this cross pattern artifact is significant even in the raw image. The northern component of the cross encompasses more than 10% of the peak of the central core. In the case of NGC3227, the AOB core was compared with HST-NICMOS data (Alonso-Herrero, private communication) and the core cross pattern was found not to exist, even at very low signal levels. Note that a similar feature may have been mistakenly identified as a physical structure in NGC1068 ([@rouan98]). We stress, however, that the AOB system is perfectly reliable in identifying systems such as binary stars or galaxies with true double nuclei, if the relative flux is less than a factor of 10, or if the artifact is carefully taken into account (see [@knap97] and [@lai98] for resolved double nuclei in Mkn273 and Mkn231 respectively which appear to be real). A technique has been developed to reconstruct the AOB PSF from the modal control used during the actual observations ([@veran98]). The algorithm requires a fairly bright guiding source to accurately reproduce the PSF, and not all our galaxy cores satisfy this criterion. Reconstructed PSFs for 4 of the galaxies depicted in Figure \[artfig1\] (all with bright cores), are shown in figure \[artfig3\] with a smooth model subtracted. Structures are similar in scale to the actual galaxy data, showing some kind of “lobe" configuration. The actual galaxy counterparts to these reconstructions display a broad range in core morphology from the extremely bright pointsource in NGC4151 to the bright but extended Seyfert 2 Mkn620. The reconstructed PSF has essentially the same morphology in all cases, and certainly does not reproduce the actual artifact morphology in the galaxy cores. However, the reconstruction does not take into account the extended nature of the underlying galaxy and the nuclear core, and would not be expected to reproduce the artifact if it was related to this. Non-common light path to the wavefront sensor and IR array may also contribute to the discrepancy. The structures are most prominent when a bright point source dominates the core, and is less apparent in the weaker or more extended cores, possibly indicating that the artifact is strongest when the AOB correction is at its best. The morphology is also not dependent on the camera used with AOB (both the 1k$^2$ pixel KIR and 256$^2$ pixel MONICA show identical artifacts, once the images have been rotated to the same orientation on the sky). Although some stars imaged with AOB were reported not to show extended PSF cross-patterns, we find that the structures appear to contaminate our calibration stellar images corrected with AOB in a similar manner (Chapman et al. 1999a in preparation). Thus the extended underlying galaxy is likely not the primary cause of the cross-shaped and double/triple-lobed artifacts. Further modeling of the response of the curvature based wavefront sensing AO system is required to ascertain the root of such structures. Deconvolution techniques may be able to account for these artifacts sufficiently, once their true nature is uncovered. This will allow the central arcsecond of active galaxies to be probed with more confidence using AOB. [lcccccccl]{} Galaxy & Alternate & z & filters & sy class & gal type & instrument ic 4329a & eso 445-g50 & 0.016 & J,H,K,CO & S1 & Sa,S0 & KIR\ mrk 1066 & UGC 2456 & 0.012 & J,H,K & 2 & Sc,SB0 & KIR\ mrk 1330 & NGC 4593 & 0.009 & J,H,K & U?1 & Sb/c, SBb & KIR\ mrk 3 & UGC 3426 & 0.014 & J,H,K & 2 & S0 & KIR\ mrk 620 & NGC 2273 & 0.006 & J,H,K & 2 & SBb,SBa & KIR\ mrk 744 & NGC 3786 & 0.010 & J,H,K & U1.8(S) & Sb,Sa &KIR\ mrk 766 & NGC 4253 & 0.012 & J,H,K & U1.5(S) & SBc,Sa &KIR\ ngc 2639 & NGC 4593 & 0.011 & J,H,K & R1 & Sb,Sa & KIR\ ngc 2992 & & 0.007 & J,H,K,CO & 2 & Sa & KIR/MON\ ngc 3516 & UGC 6153 & 0.009 & J,H,K & S1.5 & S0 & KIR\ ngc 3998 & & 0.008 & J,H,K & U1.5 & E & KIR\ ngc 4051 & & 0.002 & J,H,K & S1 & Sb,SBbc &KIR\ ngc 4151 & & 0.002 & J,H,K & S1 & Sa/ & KIR\ ngc 4968 & ESO 508-g6 & 0.009 & H & 2 & Sa/SB0a &KIR\ ngc 5033 & & 0.002 & J,H,K & 2 & & KIR/MON ngc 5135 & ESO 444-g32 & 0.013 & K & 2 & Sc,SBaba & KIR\ ngc 5273 & & 0.004 & H & R1.8 & Sa/ & KIR ngc 3081 & & 0.007 & K & 2 & SBca & KIR ngc 3393 & & 0.012 & K & 2 & Sa/SBa & KIR\ mrk 1376 & NGC 5506 & 0.007 & J & R1.9 & Sa/edge & KIR\ ngc 3227 & UGC 5620 & 0.003 & J,H,K & U1.5(S) & SBa & MONICA\ ngc 5548 & MRK 1509 &0.017 & J,H,K & U1.5(S) & Sa & MONICA\ mrk 348 & NGC 262 & 0.014 & H,K & 2 & Sa,S0a & MONICA\ ngc 1068 & & 0.003 & H & 2 & Sb/ & MONICA ngc 7469 & & 0.016 & H & U1(S) & Sb/c & MONICA ngc 1241 & & 0.013 & H & 2 & Sb/c, SBb & MONICA ngc 1275 & & 0.013 & K & 1.5 & Sa/ & MONICA ngc 1386 & & 0.002 & H,K & 2 & Sb/c & MONICA\ ngc 5728 & & 0.005 & K & 2 & Sb/ & MONICA ngc 5929 & UGC 9851 & 0.008 & J,H,K & 2 & S0,Sab & MONICA ngc 5953 & & 0.007 & H & 2 & Sb/ & MONICA ngc 6814 & & 0.003 & J,H,K,CO & U1 & Sa/ & MONICA ngc 7465 & MRK 313 & 0.006 & J,H,K & 2 & Sa/ & MONICA ngc 7582 & ESO 291-g16 & 0.005 & H,K & 2 & Sa/ & MONICA ngc 7590 & & 0.005 & H & 2 & Sb/ & MONICA ngc 7743 & & 0.007 & H,K,CO & 2 & Sa/S0a & MONICA ngc 5005 & & 0.002 & H & 2 & S0/ & MONICA NGC3227 ======= The CFHT J,H,K- and HST V-band images are presented in Figure \[n3fig1\] on a log scale. A diffuse, ellongated structure containing wispy spiral bands is seen surrounding the nucleus in all wavelengths. Subtraction of a smooth model reveals that this region is punctuated with bright knotty structures tracing out a mini-spiral pattern within a region 3“x2” (Figure \[n3fig3\]). The colours of these knots are consistent with a red supergiant population, with a scattered AGN light contribution near the core. However, the presence of the AOB artifacts in the very core region make the identification of physical knots difficult. We explored several methods of removing the low frequency galactic component, including various smoothing filters, a one-dimensional elliptical isophote model, and a multi-component (bulge+disk+point source) elliptical isophote model. All methods consistently unveil the knotty spiral structure. However, in the central arcsec of the galaxy, subtracting isophotal fitting models results in prominent artifacts which obscure structural details as described in section 3.1. Color maps are formed by convolving the images to the worse resolution of a given pair and taking the flux ratio (Figure \[n3fig2\]). Any color gradients in these images can result from several different processes: 1) change in dust 2) change in stellar population 3) change in gas. The most prominent feature is an irregular-shaped patch to the southwest. The fact that this region appears clearly as a deficit in the V-band image, and takes on a patchy morphology is strong evidence for dust obscuration as the source of the color gradient. The region is therefore most pronounced the V-K color map, since the K image is least affected by dust. The J-K image indicates that substantial dust still affects this region in the J-band. The nucleus is also very red relative to V and J, possibly as a result of thermal dust emission in the K band. The red colors of the knotty spiral starburst stand out from a region slightly bluer than the larger scale bulge of the galaxy. The images are distorted by PSF artifacts in the central 0.5", with strong diffraction spikes in the HST image. However, there appears to be a disk-like feature most clearly visible in the J-K and V-K color maps, as it seems to be bluer than the rest of the galaxy at K. The position angle of 43$^\circ$ indicates that this elongation is unrelated to the AOB artifacts seen in figure 1. The 1D profiles of the galaxies are similar at J, H and K, displaying bumps in ellipticity at 1.5 and 0.5 arcsec radius, confirming the presence and position angle of the above disk and the enhanced region coincident with the spiral starburst. The isophotes are twisted of order 10$^\circ$ in both cases. The images are also compared to the 6cm and 18cm MERLIN radio continuum emission, both of which align with the axis of the nuclear spiral as seen in Figure \[n3fig3\]. Previous explanations for the radio structure (Mundell et al 1995), invoked the standard unified AGN model to explain this emission as collimated outflow. However, there is an offset in orientation of the \[OIII\] “cone" and the small-scale radio features. A projection effect would be possible, but this would necessitate that the NE side of the disc is closer to us than the SW side. This could only occur if the spiral arms were leading rather than trailing (Mundell et al 1995). As figure \[n3fig3\] shows, the radio aligns well with both the orientation and some of the knots of the starburst spiral, and the obvious interpretation is that we are seeing synchrotron emission from SNe remnants. Discussion ---------- Table 2 The structural components of NGC3227 [llllll]{} & [Scale]{} & [PA ($^\circ$)]{} & [Ellipticity]{} & [Observed with]{} & [Function in galaxy]{} Large Scale galaxy & 1-10 kpc & -25 & 0.5 & V-band & Large Scale bar & 1-5 kpc & -45 & 0.8 & galaxy subtracted & funnel material to ILR at 7“Extended \[OIII\] & 1kpc & 35 & & OIII filter/ OASIS \[SIII\] & collimated emission? Circum-nuclear ring & 1 kpc & & & H$_\alpha$ & ILRMedium Scale bar & 100-1000pc & ? & & submm CO & funnel material to ILR at 2” K-band ellipse & 200 pc & -10 & 0.2 & model subtract/ color maps& spiral starburst, ILRRadio jets/blob & 100 pc & -10 & & MERLIN 8/16cm& SN/outflow?K-V blob & 100 pc & -10 & 0.1 & K-V map, raw Vband & bluer than galaxy K-J annulus & 100 pc & 40 & 0.3 & K-J, K-V, J maps & twisted disk/bars, scattered AGN light Several possible scenerios emerge from these results. We tabulate the observed structures in this galaxy from the largest to the smallest in Table 2. On the largest scales Gonzalez Delgado et al. (1997) noted that a large-scale bar appears to transport material towards an inner radius which corresponds to the calculated inner Linblad resonance (ILR) at roughly 7“. At this point, prominent dust and HII regions indicate substantial star formation. Within this region, a molecular bar of length $\sim1$kpc is observed in CO with and ILR of 2” (Schinnerer 1998). This radius corresponds with the outer extent of the spiral starburst rings in our images. With such nested bar structure repeating itself at these two larger scales, it is natural to speculate that the small scale elongation seen in the color maps and profile analysis may be yet another bar potential funneling material down to the scales where viscous forces may take over to fuel the AGN. However, a larger scale extended \[OIII\] region (Arribas et al. 1995) lies to the northeast and has been interpretted as a narrow-line region ionized by the AGN, collimated into a bi-cone by a small-scale ($\sim pc$) dusty torus. The fact that this collimation axis roughly aligns with our “bar" may be an indication that this elongation actually represents scattered AGN light. This is made all the more convincing by the blue colour of the elongation. On the other hand, if our observed small-scale elongation is some sort of twisted disk as found in CentaurusA by Schreier et al. (1998), its plane lies roughly perpendicular to the axis defined by the radio “jet" observed at 6 and 18cm, and would be consistent with a collimated radio jet normal to an accretion disk plane. For the radio emission to be interpretted as an outflow, the collimated \[OIII\] ionization picture would then have to be abandoned. A more detailed analysis can be found in Chapman et al. (1999b). NGC2992 ======= In figure \[n2fig\], we present an H-band AOB image of the central 7" of NGC2992. Of obvious note are the jet-like extent to the NW, and the elongated isophotes to the SW along the galaxy disk. The V-band HST image is also depicted with 8.4GHz radio contours overlaid. Although it is still clear that the isophotes are elongated to the SW along the galaxy disk, the galaxy morphology is much more distorted due to the effects of dust and there is no indication of the extension to the NW out of the galaxy plane. The V-band image shows that the radio emission lying along the galaxy disk has no obvious opical emission associated and appears to lie well within the dust lane. When we form an H-V color map, we now observe highly reddened emission lying along the radio contours in the disk of the galaxy, however with what appears to be a loop to the north, associated with an inner loop of radio emission. By subtracting a model image consisting of either elliptical fitted isophotes, or a smooth median filtered image, we are able to discern the spiral arms along the disk, as well as an extension to the West (figure \[n2fig\]), also noted in [@alonso98]. There is clearly some radio emission coincident with the southern spiral arm, which breaks up into a similar knotty morphology to the H-band model-subtracted image. At faint levels, it is also possible to discern knotty features along the northern red H-V loop, the colors of which are consistent with star formation. The extended feature aligns with the beginning of the H-V loop noted above, but continues outward to fill the radio loop. There is, however, little sign of optical or near-IR counterparts to the radio loops out past the disk of the galaxy, even at K-band where the ability to see through the dust lane is greatest. This is evidence for the actual figure-8 loops lying out of the galactic disk plane, superimposed over fairly strong disk emission related to star formation. If the extended near-IR emission to the West is related to the NW radio loop, the dust lane must extend out past the loop in order to heavily obscure the optical/near-IR emission. This sort of near-IR “jet" may exist towards the SE radio loop as well, although largely obscured by the galaxy disk. There are several favored models for such figure-8 radio emission. The most convincing in light of our new near-IR imaging is that the loops result from some sort of expanding gas bubbles which are seen preferentially as limb-brightened loops (Wehrle 1988). Such outflows may be associated with the AGN core, which is consistent with the orientation of the proposed ionization cones observed at larger scales Allen et al. 1998. Superwinds in starbursts would blow preferentially out of the galaxy plane, such as in NGC253 (Unger et al. 1987), providing a similar mechanism even without a strong AGN driving the outflow. Here, however, the \[OIII\] emission is likely to be associated, but not continuum emission thus our data likely rule out this latter model. These possibilities are further explored in Chapman et al. (1998c) [dum]{} Allen et al. 1998, astro/ph9809123 Alonso-Herrero A., Simpson C., Ward M.J., Wilson A.S., 1998, ApJ, 495, 196 Antonucci R., 1993, 31, 473 Chapman S. & Morris S.L., in preparation 1999a. Chapman S. & Morris S.L., 1999b, submitted to MNRAS. Chapman S., Morris S.L., Alonso-Herrero A., Falcke H., in preparation 1999c. Falcke, H. et al., 1988, AJ, 95, 1689 Arribas S., Mediavilla E. 1994, , 437, 149 Gonzalez Delgado R.M., Perez E. 1997, , 284, 931 Knapen J., et al, 1997 , 500, 500 Lai O. et al., 1998, A&A, in press. Mundell C.G., et al. 1995, , 275, 67 Nadeau D., Murphy D.C., Doyon R., Rowlands N., 1994, 106, 909 Roddier F.J., Graves J.E., McKenna D. , Northcott M.J. , 1991, Vol. 1524, p. 248 Rouan D. et al., 1998, A&A, in press. Schreier E.J. et al. 1998, , in press Schinnerer E. et al. 1998, AGM 14, H08 Ulvstad & Wilson A. 1984, UNGER, S. W., PEDLAR, A., AXON, D. J., WHITTLE, M., MEURS, E. J. A., WARD, M. J., 1987 MNRAS, 228, 671 Veran J.P., Rigaut F., Maître H., Rouan D., 1997, Journ. of the Optic. Soc. of America A 14, 11 Wehrle A.E., Morris M., 1988, AJ, 95, 1689
{ "pile_set_name": "ArXiv" }
--- abstract: | We discuss the discrete data assimilation problem for the 3D viscous primitive equations arising in the modeling of large scale phenomena in oceanic dynamics. Our main result states possibility of asymptotically reliable prognosis based on a discrete sequence of finite number of scalar observations. Our method is quite general and can be applied to a wide class of dissipative systems. [**Keywords:** ]{} 3D viscous primitive equations, data assimilation, determining functionals. [**2010 MSC:**]{} 35R60, 37N10, 76F25, 76F55 author: - 'Igor Chueshov[^1],' title: | Discrete data assimilation via\ Ladyzhenskaya squeezing property in\ the 3D viscous primitive equations --- Introduction {#introduction .unnumbered} ============ Data assimilation problem is a question how to incorporate available observation data in computational schemes to improve quality of predicting of the future evolution of the corresponding dynamical system. This problem has a long history and was studied by many authors at different levels (see, e.g., the monographs [@data-book; @kal2003-data] and the references therein). In this paper we consider the case when observations of the system are making in some sequence $\{t_n\}$ of moments of time and use the same formulation of the data assimilation problem as in [@HO-Titi-data-as2011]. Our main goal is to demonstrate the role of the so-called Ladyzhenskaya squeezing property (which is valid for a wide class parabolic type PDEs, see [@lad1; @lad2]) in the solving of data assimilation problems. As in [@HO-Titi-data-as2011] we also involve the notion of determining modes or, more generally, determining functionals. However our method is different from the approach developed in [@HO-Titi-data-as2011] for the 2D Navier–Stokes equations. Moreover, the method based on the squeezing property looks more general and can be applied to a wide class of dissipative systems admitting the Ladyzhenskaya property. In this paper as a model we choose the system of the 3D viscous primitive equations which arise in geophysical fluid dynamics for modeling large scale phenomena in oceanic motions. In this case data assimilation problem is related with reliability of weather predictions. The paper is organized as follows. In Section \[sec:3PE\] we describe the model and quote its properties which need for data assimilation. The main technical tools are the Ladyzhenskaya squeezing property and the theory of determining functionals which we discuss shortly in Section \[sec:def-f\]. In Section \[sec:data\] formulate the data assimilation problem, introduce the notion of asymptotically reliable prognosis and prove our main result concerning a finite number of scalar observations in a discrete sequence of times. 3D Primitive equations {#sec:3PE} ====================== The primitive equations are based on the so-called hydrostatic approximation of the 3D Navier-Stokes equations for velocity field $u$ and coupled to thermo- and salinity equations which are taken into account via small variation of density (or equivalently via buoyancy $b$), see, e.g., the survey [@temam2008-srw] and the literature cited there. Below to simplify the presentation we consider periodic boundary conditions of the same type as in [@petcu] and [@temam2008-srw] (see also [@Chueshov-squeszing2012]). However, it should be noted that our results remains valid in the case of free type boundary conditions like in [@CaoTiti]. The case of mixed free-Dirichlet boundary conditions (see [@kuka]) is more complicated and requires a separate consideration. Let $${{\cal O}}= (0,L_1)\times (0,L_2)\times (-L_3/2,L_3/2) \subset {{\mathbb R}}^3.$$ We denote by $\bar{x}=(x,z)=(x_1,x_2,z)$ the spatial variable in ${{\cal O}}$ and suppose that ${{\nabla}}$, ${{\rm div\, }}$ and $\Delta$ are the gradient, divergence and Laplace operators in the (horizontal) variable $x=(x_1,x_2)$. After the reduction based on the hydrostatic relation we arrive (see, e.g., [@temam2008-srw]) at the following equations for the horizontal fluid velocity $$v(\bar{x},t)=(v^1(\bar{x},t);v^2(\bar{x},t)),~~ \bar{x}=(x,z),$$ for the (surface) pressure $p=p(x,t)$ and for the buoyancy $b= b(\bar{x},t)$: $$\begin{aligned} \label{fl.1a} v_t +(v,{{\nabla}})v & -\left[\int_0^{z} {{\rm div\, }}vd\xi \right]{\partial}_z v -\nu [\Delta v +{\partial}_{zz} v] +f v^{\perp} \notag \\ & =-\nabla \left[ p(x,t)+\int_0^{z} b d\xi\right] +G_f\quad {\rm in\quad} {{\cal O}}\times(0,+\infty),\end{aligned}$$ $$\label{fl.4a} b_t +(v,{{\nabla}}) b - \left[\int_0^{z} {{\rm div\, }}vd\xi \right] {\partial}_zb-\nu [\Delta b +{\partial}_{zz} b] =G_b\quad {\rm in\quad} {{\cal O}}\times(0,+\infty),$$ where $\nu>0$ is the dynamical viscosity, $f$ is the Coriolis parameter and $v^{\perp}=(-v^2;v^1)$. The functions represents $G_f$ and $G_b$ are volume sources related to the fluid field and the buoyancy. As in [@petcu; @temam2008-srw] the equations in (\[fl.1a\]) and (\[fl.4a\]) supplied with the conditions: $$\label{fl.4a-bc1} {\rm div}\,\int_{-L_3/2}^{L_3/2} v dz=0;~~ v ~\mbox{is periodic in $\bar{x}$ and even in $z$},~~\int_{{{\cal O}}} v d\bar{x}=0;$$ $$\label{fl.4a-bc2} b ~\mbox{is periodic in $\bar{x}$ and odd in $z$}.$$ We also impose initial data: $$\label{fl.4a-id} v(0)=v_0,~~~b(0)=b_0.$$ We note that the vertical component of the velocity field has the form $$w(\bar x,t)=-\int_0^{z} {{\rm div\, }}v(x,\xi,t) d\xi~~\mbox{for every $\bar{x}=(x,z)\in {{\cal O}}$}$$ and thus the full velocity field $(v^1,v^2,w)$ satisfies incompressibility condition. We emphasize that the (surface) pressure $p$ in (\[fl.1a\]) depends on 2D (horizontal) variable $x$ only. Basing on this observation a new effective approach [@CaoTiti] for proving of the global well-posedness for problems like (\[fl.1a\]) and (\[fl.4a\]) have been implemented, see [@CaoTiti] and the discussion therein. The system of the viscous 3D primitive equations was intensively studied for the different types of boundary conditions (see the literature cited in the survey [@temam2008-srw]). The existence of week solutions was established in [@LTW-ocean]; for global well-posedness of strong solutions we refer to [@CaoTiti] and also to [@kobelkov; @kuka; @petcu]. The uniqueness of weak solutions is still unknown. The question on a global attractor for the viscous 3D primitive equations was considered in [@ju] (see also the papers [@petcu] and [@Chueshov-squeszing2012] devoted to the periodic case). We denote by $\dot{H}_{per}^s({{\cal O}})$ the Sobolev space of order $s$ consisting of periodic functions such that $\int_{{{\cal O}}} f d\bar{x}=0$ and introduce the following spaces: $$V_s=\left\{ v=(v^1;v^2)\in [\dot{H}_{per}^s({{\cal O}})]^2 :\; v^i~\mbox{is even in}~z,\; {\rm div}\,\int_{-L_3/2}^{L_3/2} v dz=0 \right\}$$ for $s\ge 0$. We equip $H\equiv V_0$ with $L_2$-norm $\|\cdot\|$ and denote by $(\cdot,\cdot)$ the corresponding inner product. It is convenient to endow the spaces $V_1$ and $V_2$ with the norms $\|\cdot\|_{V_1}= \|\nabla_{x,z}\cdot\|$ and $\|\cdot\|_{V_2}= \|\Delta_{x,z}\cdot\|$. Here and below we use the notations ${{\nabla}}_{x,z}$ and $\Delta_{x,z}$ for gradient and Laplace operation in the 3D variable $(x,z)$. We also introduce state spaces for the buoyancy variable by the formulas $$E_s=\left\{ b\in \dot{H}_{per}^s({{\cal O}}) :\; b~\mbox{is odd in}~z \right\},~~~s\ge0.$$ We equip them with the standard Sobolev norms. We suppose $W_s=V_s\times E_s$ with the corresponding (Hilbert) product norms. As it was already mentioned, starting with [@CaoTiti] the global well-posedness of the equations in and was studied by many authors [@kobelkov; @kuka; @petcu; @temam2008-srw]. The following result on well-posedness of strong solutions in the case of periodic boundary conditions was basically proved in [@petcu] (see also [@temam2008-srw] and Remark 2.2 in [@Chueshov-squeszing2012]). \[pr:titi\] Let $G_f\in V_0$, $G_b\in E_0$, and $U_0=(v_0;b_0)\in W_1$. Then problem (\[fl.1a\])–(\[fl.4a-id\]) has a unique strong solution $(v(t);b(t))$: $$U(t;U_0)\equiv (v(t);b(t))\in C({{\mathbb R}}_+; W_1)\cap L_2(0,T; W_2),~~~\forall T>0.$$ This solution generates a dynamical system $(S_t, W_1)$ with the evolution operator $S_t$ defined by the relation $S_tU_0=U(t;U_0)$. The operator $S_t$ satisfies the Lipschitz property: $$\|S_tU-S_tU_*\|_{W_1}\le C_{T,{{\varrho}}}\|U-U_*\|_{W_1},~~ t\in [0,T],$$ for every $T>0$ and $U,U_*\in {\mathscr{B}}_1({{\varrho}})\equiv\{U : \|U\|_{W_1}\le {{\varrho}}\}$. If $G_f\in V_{m-1}$, $G_b\in E_{m-1}$ and $U_0=(v_0;b_0)\in W_m$ for some $m\ge 2$ then (see [@petcu; @temam2008-srw]) the solution $U$ lies in the class $C({{\mathbb R}}_+; W_m)\cap L^{loc}_2({{\mathbb R}}_+; W_{m+1})$. This observation makes it possible to use smooth approximations of solutions in the calculations with multipliers (see, e.g., [@Chueshov-squeszing2012]). For our goal the following assertion is important. \[pr:frame-for-kicks\] Let the hypotheses of Proposition \[pr:titi\] be in force. Then - There exist positive constants $a_0$ and $a_1$ such that $$\label{S-sq1} \|S_tU\|\le e^{-a_0t}\|U\| +a_1 K_G,~~~ t\ge 0, ~~ U\in W_1,$$ where $K_G^2=\|G_f\|^2+\|G_b\|^2$. - If we assume in addition that $$\label{GG-cond} G_f\in V_1, ~G_b\in E_1~~ and~ also~~ ({\partial}_zG_f;{\partial}_zG_b)\in \big[L_6({{\cal O}})\big]^3,$$ then for every ${{\varrho}}>0$ and $0<{{\alpha}}\le\beta<+\infty$ there exists the constant $C({{\alpha}},\beta, {{\varrho}})>0$ such that $$\label{S-sq2} \|S_tU\|_{W_2}\le C({{\alpha}},\beta, {{\varrho}})~~\mbox{for every}~ t\in [{{\alpha}},\beta], ~~\|U\|\le {{\varrho}}.$$ The first statement is achieved by the standard multipliers $v$ and $b$ applied to and , see [@petcu; @temam2008-srw], for instance. The second statement is a more complicated and based mainly on the calculations given in [@CaoTiti] and [@petcu]. The corresponding argument involves the splitting of the system into 2D Navier-Stokes type equations coupled with 3D Burgers type model (see [@CaoTiti] and also [@petcu; @temam2008-srw]) and consists of several steps based on the application of the same multipliers as in [@CaoTiti; @petcu; @temam2008-srw]. The spatial periodicity of the system allows us to use freely higher order multipliers like $\Delta_{x,z}^2 v$ and $\Delta_{x,z}^2 b$. For some related details we refer to the paper [@Chueshov-squeszing2012] which contains a very similar argument in the proof of Theorem 3.1 on the existence of a smooth absorbing set. Proposition \[pr:frame-for-kicks\] implies that the system $(S_t,W_1)$ possesses an absorbing set which is bounded in $W_2$. More precisely we have the following assertion. \[co:absorbing\] Let be in force. Then there exists $K>0$ such that the ball $${\mathscr{B}}\equiv {\mathscr{B}}_{2}(K)=\{ U\in W_2: \|U\|_{W_2}\le K\}$$ is absorbing for the dynamical system $(S_t, W_1)$ generated by problem (\[fl.1a\])–(\[fl.4a-id\]), i.e., for any bounded set $B$ in $W_1$ there is $t_B$ such that $$S_tB\subset {\mathscr{B}}~~~\mbox{for all}~ t\ge t_B.$$ It follows from that $$\|S_tU\|\le 1+ a_1 K_G~~\mbox{for all}~~ t\ge t_B,$$ and thus by we have that $$\|S_{t+1}U\|_{W_2} \le K\equiv C(1,1, 1 +a_1 K_G)~~\mbox{for all}~~ t\ge t_B,$$ i.e. the ball ${\mathscr{B}}$ possesses the desired property. We can also prove the Lipschitz property in $H$ provided one of two solutions belongs to $W_2$. \[pr:Lip-H\] Let $U_1(t)$ and $U_2(t)$ be two strong solutions to (\[fl.1a\])–(\[fl.4a-id\]). Assume that $\|U_1(t)\|_{W_2}\le R$ for all $t\in [0,T]$ for some $T>0$. Then $$\label{LipH-1} \| U_1(t)-U_1(t)\|\le C_T(R) \| U_1(0)-U_1(0)\|,~~~\forall\, t\in [0,T].$$ We note (see, e.g., [@temam2008-srw]) that problem (\[fl.1a\])–(\[fl.4a-id\]) can be written in the form $${\partial}_t U+\nu AU+B(U,U)+CU=G,~~~ U(0)=U_0,$$ where $A$ is a positive self-adjoint operator in $H$ generated by the bilinear form $$\label{a-form} a(U,U_*)=\int_{{\cal O}}\left[ {{\nabla}}_{x,z} v\cdot{{\nabla}}_{x,z} v_* + {{\nabla}}_{x,z} b\cdot{{\nabla}}_{x,z} b_*\right] dxdz,$$ where $U=(v;b)$ and $U_*=(v_*;b_*)$ are from $W_1$, $C$ is a bounded skew-symmetric operator, and $B(U,U)$ is a quadratic operator possessing the properties $$B(U^*,U),U)= 0, ~~~ B(U,U^*),U)\le C\|U\|^{3/2}_{W_1} \|U\|^{1/2} \|U^*\|_{W_2} \label{LipH-2}$$ for every $U^*\in W_2$ and $U\in W_1$. Thus for the difference $V=U_1(t)-U_2(t)$ we have that $${{\frac12}}{\partial}_t\|V(t)\|^2 +\nu a(V(t),V(t))+ B(V,U_1),V)=0$$ which, via and Gronwall’s lemma, implies the relation in . We note that the operator $A$ generated by form has a discrete spectrum. This means that there exists an orthonormal basis $\{ e_k\}$ in $H$ such that $$\label{a-form2} A e_k={\lambda}_k e_k,~~~0<{\lambda}_1\le {\lambda}_2\le\ldots,~~~\lim_{k\to+\infty}{\lambda}_k=+\infty.$$ We denote by $P_N$ the orthoprojector onto Span$\{e_1,\ldots,e_N\}$ and $Q_N=I-P_N$. The following Ladyzhenskaya squeezing property (see [@lad1; @lad2]) of the evolution operator $S_t$ is the main ingredient of our further data assimilation considerations. \[pr:frame-for-kicks2\] Let be in force. Then for every $q<1$, $0<{{\alpha}}\le \beta<+\infty$ and $L$ there exists $N_*=N({{\alpha}},\beta,L,q)$ such that $$\|Q_N[S_t U-S_tU_*]\|_{W_1}\le q \|U-U_*\|_{W_1},~~\forall\, t\in [{{\alpha}},\beta], ,~~\forall\, N\ge N_*,$$ for any $U$ and $U_*$ from the set $${\mathscr{D}}=\left\{ U\in W_2\,: \; \|S_t U\|_{W_2}\le L~~ for~all~t\in [0,\beta]\right\}.$$ The same type argument as in the proof of Theorem 3.5 in [@Chueshov-squeszing2012] leads to the desired result. Observation/measurement functionals {#sec:def-f} =================================== To describe observation/measurement procedure we use a finite family ${{\cal L}}$ of linear continuous functionals $\{l_j: j=1,\ldots,N\}$ on the phase space. If $U$ is a phase vector which corresponds to some state of the system, then, similar to [@HO-Titi-data-as2011], we can treat the values $\{l_j(U): j=1,\ldots,N\}$ as a set of observation data. Our task is now to determinate the state $U$ with the help of observation functionals $\{l_j\}$. Therefore to describe admissible observations it is natural to involve well-developed theory of determining functionals. This theory starts with the pioneering paper [@FP-det] on determining modes and was developed by many authors for different classes of PDE systems and different families of functionals (see the recent discussion in [@FJKT-det]). For a general theory of the determining functionals we refer to [@CT-det], see also [@Chueshov; @cl-mem; @cl-book] for a development of this theory based on the notion of the completeness defect. The concept of completeness defect which was introduced in [@Chu97; @Chu98] seems a convenient tool in characterization of observation functionals. \[de7.8.21\] [Let $V$ and $H$ be reflexive Banach spaces and $V$ is continuously and densely embedded into $H$. [*The completeness defect*]{} of a set ${\cal L}$ of linear functionals on $V$ with respect to $H$ is the value $$\label{7.8.21} \epsilon_{\cal L}(V,H) =\sup\{ \parallel w \parallel_{H} : w\in V,\, l (w)=0,\, l\in {\cal L},\, \parallel w \parallel _{V} \le 1 \}\;.$$ ]{} It is obvious that $\epsilon_{{\cal L}_1} (V,H)\ge\epsilon_{{\cal L}_2} (V,H)$ provided Span${\cal L}_1\subset{\rm Span}{\cal L}_2$. In addition, $\epsilon_{\cal L} (V,H)=0$ if and only if the class of functionals ${\cal L}$ is complete in $V$; this means that the property $l(w)=0$ for all $l\in {\cal L} $ implies $w=0$. We can also generalize the notion of the completeness defect by considering some seminorms $\mu_V$ in instead of the norm $\|\cdot\|_H$ (see, e.g., [@cl-book]). Below we use the so-called interpolation operators which are related with the set of functionals given. To describe their properties we need the following notion. \[de7.8.22\] [Let $V\subset H$ be separable Hilbert spaces and $R$ be a linear operator from $V$ into $H$. As in [@Au72] the value $$e^H_V (R) =\sup\{\Vert u-Ru\Vert_H\; :\; \Vert u\Vert_V\le 1\}\equiv \|I-R\|_{V\mapsto H}$$ is said to be the [*global approximation error*]{} in $H$ arising in the approximation of elements $v\in V$ by elements $Rv$. Here and below $\| \cdot \|_{V\mapsto H}$ denotes the operator norm for linear mappings from $V$ into $H$. ]{} The following assertion (see [@Chu98; @Chueshov] for the proof) shows that the completeness defect provides us with a bound from below for the best possible global approximation error. \[th7.8.22\] Let $V$ and $H$ be the separable Hilbert spaces such that $V$ is compactly and densely embedded into $H$. Let ${\cal L}$ be a set of linear functionals on $V$. Then we have the following relations $$\epsilon_{\cal L}(V,H)=\min\{ e^H_V (R)\; :\; R\in {\cal R}_{\cal L}\},$$ where ${\cal R}_{\cal L}$ is the family of linear bounded operators $R: V\mapsto H$ and such that $Rv=0$ for all $v\in {\cal L}^{\perp}=\{ v\in V\; :\; l(v)=0, \; l\in {\cal L}\}$. Moreover, we have that $$\label{eL-opt} \epsilon_{\cal L}(V,H)= e^H_V (I-Q_{{\cal L}})=\sup\{ \| Q_{{\cal L}}u\|_H\; :\;\Vert u\Vert_V\le 1\},$$ where $Q_{{\cal L}}$ is the orthoprojector in $V$ onto ${\cal L}^{\perp}$. One can show (see [@Chueshov]) that any operator $R\in {\cal R}_{\cal L}$ has the form $$\label{R-op} Rv=\sum_{j=1}^N l_j(v) \psi_j, ~~~ \forall\, v\in V,$$ where $\{\psi_j\}$ is an arbitrary finite set of elements from $V$. This why ${\cal R}_{\cal L}$ is called the set of interpolation operators corresponding to the set ${{\cal L}}$. An operator $R\in {{\cal R}}_{{\cal L}}$ is called [*Lagrange* ]{} interpolation operator, if it has form with $\{\psi_j\}$ such that $l_k(\psi_j)=\delta_{kj}$. In the case of Lagrange operators we have that $R^2=R$, i.e., $R$ is a projector. We also note that the operator $Q_{{\cal L}}$ in has the following structure $$Q_{{\cal L}}=I-P_{{\cal L}}~~~\mbox{with}~~P_{{\cal L}}v=\sum_{j=1}^N (\xi_j,v)_V \xi_j, ~~~ \forall\, v\in V,$$ where $\{\xi_j\}$ is the orthonormal basis in the orthogonal supplement ${{\cal M}}_{{\cal L}}$ to the annulator ${\cal L}^{\perp}$ in $V$. We call $P_{{\cal L}}$ the [*optimal*]{} interpolation operator corresponding to the set ${{\cal L}}$. Our main example is related with the eigen-basis of the operator $A$ defined by the form . \[modes\] [Denote by ${\cal L}$ the set of functionals ${\cal L}=\{ l_j(u)=(u,e_j)\; :\; j=1,2,\ldots,N\}$, where $ \{ e_k \}$ are eigenfunctions of the operator $A$ given by the form , see . The optimal interpolation operator $P_{{\cal L}}$ is Lagrange in this case and has the form $$\label{RL-opt} P_{{\cal L}}v=\sum_{j=1}^N (e_j,v) e_j, ~~~ \forall\, v\in W_1.$$ Moreover, $\epsilon_{\cal L}(W_i,H)=e^H_{W_i}(P_{{\cal L}})={\lambda}_{N+1}^{-i/2}$, $i=1,2$. Thus the completeness defect and the global approximation error $ \|I-P_{{\cal L}}\|_{W_i\mapsto H}$ can be made small after an appropriate choice of $N$. ]{} Discrete data assimilation {#sec:data} ========================== We consider the discrete data assimilation problem in the sense due to [@HO-Titi-data-as2011]. The paper [@HO-Titi-data-as2011] is focused on the case where the measurement data is taken at a sequence of discrete times $t_n$ in contrast with the papers [@Olson-titi2013; @Olson-titi2003] which consider *continuous* data assimilation. All these papers deal with for the incompressible two-dimensional Navier–Stokes equations. Following the idea presented in [@HO-Titi-data-as2011] we accept the following definition. \[de:data-assim\] [ Let $U(t)=S_tU_0$ be a solution to (\[fl.1a\])–(\[fl.4a-id\]) with initial data $U_0$ at time $t_0$. Let ${{\cal L}}=\{l_j\}$ be a finite family of functionals on $H$ (each functional $l_j$ is interpreted as a single observational measurement). Let $R_{{\cal L}}$ be some Lagrange interpolation operator related with ${{\cal L}}$ such that the sequence $\{r_{{\cal L}}^n\equiv R_{{\cal L}}U(t_n)\}$ represents the (joint) observational measurements of the reference solution $U(t)$ at a sequence $\{t_n\}$ of times, we call the sequence $\{r_{{\cal L}}^n\}$ [*observation values*]{}. Now we can construct [*prognostic values*]{} at time $t_n$ by the formula $$\label{progn1} u_n= (1-R_{{\cal L}}) S_{t_n-t_{n-1}} u_{n-1}+r_{{\cal L}}^n ,~~~ n=1,2,\ldots,$$ where $u_0$ is (unknown) vector which, according to [@HO-Titi-data-as2011], corresponds to an initial guess of the reference solution $U(t_0)$. We can also define the prognostic (piecewise continuous) trajectory as $$\label{progn-traj} u(t) = S_{t- t_n} u_n~~\mbox{for}~~ t \in [t_n, t_{n+1}), ~~~ n=0,1,2,\ldots$$ We say that the prognosis is [*asymptotically reliable*]{} at a sequence of times $t_n$ if $$\|U(t_n)-u_n\|_{W_1} \to0 ~~~\mbox{as}~~n\to +\infty.$$ ]{} Our goal is to find conditions on $R_{{\cal L}}$, $t_n$ and $\eta$ which guarantee that the prognosis based on a finite number of single observations is asymptotically reliable. We assume that $0<\alpha\le t_{n+1}-t_n\le \beta<+\infty$ for some positive ${{\alpha}}$ and $\beta$. The following assertion gives us a dissipativity property for prognostic values which is important for our application of the Ladyzhenskaya squeezing property. \[le:dis-prognosis\] Assume that $\|U(t)\|_{W_2}\le K$ for all $t\ge t_0$. Let $$\| R_{{\cal L}}\|_{W_2\mapsto H}\le c_0~~\mbox{and}~~ \| 1-R_{{\cal L}}\|_{H\mapsto H}\le c_1~ with~ c_1<e^{a_0{{\alpha}}},$$ where $a_0$ is the constant in . Then there exists $n_*>0$ such that $$\label{un-dis1} \|u_n\|\le 1+\varrho_*~~~ \mbox{for all}~~ n\ge n_*,$$ where $\varrho_*=(a_1K_G+c_0K)(1-c_1e^{-a_0{{\alpha}}})^{-1}$. If we assume in addition that $\| 1-R_{{\cal L}}\|_{W_2\mapsto W_2}\le c_2$, then $$\label{un-dis2} \|u_n\|_{W_2}\le \varrho \equiv c_2C({{\alpha}},\beta, 1+{{\varrho}}_*)+ (1+c_2) K~~~ \mbox{for all}~~ n\ge m_*\equiv 1+n_*,$$ where $C({{\alpha}},\beta, {{\varrho}})$ is the constant from . One can see from Proposition \[pr:frame-for-kicks\] that $$\|u_n\|\le c_1e^{-a_0{{\alpha}}} \|u_{n-1}\|+a_1K_G+c_0K,~~n=1,2,\ldots$$ This implies that $$\|u_n\|\le q_*^n \|u_0\|+\varrho_*,~~n=1,2,\ldots$$ where $q_*= c_1e^{-a_0{{\alpha}}}$. This yields . To prove we note that $$\|u_n\|_{W_2}\le c_2 \|S_{t_n-t_{n-1}}u_{n-1}\|_{W_2}+ (1+c_2)K,~~n=1,2,\ldots$$ Hence follows from and . In the case of (spectral) modes we have the following assertion. \[co:diss\] Let ${{\cal L}}$ be the same as in Example \[modes\]. If we take in $R_{{\cal L}}$ to be the interpolation operator given by , then there exist positive constants $C(K_G, K,{{\alpha}},\beta)$ and $m_*$ such that $$\|u_n\|_{W_2}\le \varrho \equiv C(K_G, K,{{\alpha}},\beta)~~~ \mbox{for all}~~ n\ge m_*.$$ In this case $c_0={\lambda}_1^{-1}$ and $c_1=c_2=1$. Now we are in position to obtain the main result. \[th:prognosis\] Assume that ${{\cal L}}$ is a finite family of functionals on $H$ and there is a Lagrange interpolation operator $R_{{\cal L}}$ possessing the properties: $$\label{cond-WH} \| 1-R_{{\cal L}}\|_{H\mapsto H}\le c_1~~\mbox{and}~~\| 1-R_{{\cal L}}\|_{W_2\mapsto W_2}\le c_2$$ with the constants $c_1$ and $c_2$ independent of ${{\cal L}}$ such that $c_1<e^{a_0{{\alpha}}}$, where $a_0$ is the constant in . Then there exists $\epsilon_*>0$ such that under the condition $\epsilon(W_1,H)\le \epsilon_*$ the prognosis in is asymptotically reliable for every $u_0\in H$. In the case of the modes described in Example \[modes\] there exists $N_*$ such that the prognosis is asymptotically reliable with $R_{{\cal L}}=P_{{\cal L}}$, where $P_{{\cal L}}$ is given by with some $N\ge N_*$. We obviously have that $$U(t_n)-u_n= (1-R_{{\cal L}})[S_{t_n-t_{n-1}} U(t_{n-1})- S_{t_n-t_{n-1}} u_{n-1}] ,~~~ n\ge m_*.$$ In the case of modes we have that $I-R_{{\cal L}}=Q_N$. Therefore using Corollary \[co:diss\] by Proposition \[pr:frame-for-kicks2\] we can choose $N_*$ such that and thus $$\|U(t_n)-u_n\|_{W_1}= q \|U(t_{n-1}) - u_{n-1}\|_{W_1},~~~ n\ge m_*,$$ with $q<1$. This implies $$\|U(t_n)-u_n\|_{W_1}\to0~~\mbox{as}~~ n\to+\infty$$ with exponential speed. Therefore the statement of the theorem is valid in the case of modes. It is obvious that under conditions the hypotheses of Lemma \[le:dis-prognosis\] are in force. Therefore in the general case Proposition \[pr:frame-for-kicks2\] implies that $$\begin{aligned} \|S_{\Delta_n} & U(t_{n-1})- S_{\Delta_n} u_{n-1}\|_{W_1} \notag \\ &\le q_N \|U(t_{n-1}) - u_{n-1}\|_{W_1} +{\lambda}_N^{1/2} \|S_{\Delta_n} U(t_{n-1})- S_{\Delta_n} u_{n-1}\| \label{progn4}\end{aligned}$$ for $n\ge m_*$ with $\Delta_n =t_n-t_{n-1}$, where $q_N<1$ can be chosen as small as we need at the expense of $N$. By Proposition \[pr:Lip-H\] we have that $$\|S_{\Delta_n} U(t_{n-1})- S_{\Delta_n} u_{n-1}\|\le C_\beta({{\varrho}}) \| U(t_{n-1})- u_{n-1}\|, ~~ n\ge m_*.$$ Since $l_j(U(t_{n-1}))=l_j(u_{n-1})$, this gives $$\|S_{\Delta_n} U(t_{n-1})- S_{\Delta_n} u_{n-1}\|\le \epsilon(W_1,H)C_\beta({{\varrho}}) \| U(t_{n-1})- u_{n-1}\|_{W_1}, ~~ n\ge m_*.$$ Thus yields $$\|U(t_{n})- u_{n}\|_{W_1} \le \tilde{q} \|U(t_{n-1}) - u_{n-1}\|_{W_1}$$ for $n\ge m_*$, where $$\tilde{q}=\|I-R_{{\cal L}}\|_{W_1\mapsto W_1}\left[ q_N+{\lambda}^{1/2}\epsilon(W_1,H) C_\beta({{\varrho}})\right].$$ By the operators interpolation from condition we have that $$\|I-R_{{\cal L}}\|_{W_1\mapsto W_1}\le \sqrt{c_1c_2}.$$ Hence we can choose $N$ and $\epsilon(W_1,H)$ such that $\tilde{q}<1$. Therefore the prognosis is asymptotically reliable with exponential speed. We conclude our considerations with several remarks. \[re:conclusion1\] [ As an example of set ${{\cal L}}$ functionals $\{l_j\}$ satisfying we can consider [*generalized modes*]{} which are defined by the formulas: $$l_j(u) =(Ke_j,u),~~~ j=1,\ldots, N,$$ where $\{e_j\}$ is the eigen-basis of the operator $A$ and $K$ is a linear invertible self-adjoint operator in $H$ with maps $W_2$ into itself and is bounded in both spaces $H$ and $W_2$. In this case the operator $R_{{\cal L}}$ has the form with $\psi_j=K^{-1}e_j$. One can see that we can apply Theorem \[th:prognosis\] with ${{\alpha}}$ greater than $\frac1{a_0}\ln (1+\|K\|_{H\mapsto H})$. ]{} \[re:conclusion2\] [ Under the conditions of Theorem \[th:prognosis\] we also have that $$\lim_{t\to+\infty}\|U(t)-u(t)\|_{W_1}=0$$ for the prognostic trajectory given by . Thus the prognosis is also reliable in the sense used in [@HO-Titi-data-as2011]. ]{} \[re:conclusion3\] The number of functionals which provides an asymptotically reliable prognosis according to Theorem \[th:prognosis\] is finite. However the estimates for this number which follows from the statement of theorem are not optimal and even not constructive. The derivation of optimal bounds for the numbers requires more careful analysis of constants related to dissipativity and squeezing properties of individual trajectories. We refer to [@HO-Titi-data-as2011] for more constructive approach based on the multipliers technique and developed in the case of the 2D Navier–Stokes equations for the reference solution from the global attractor. [99]{} J.-P. Aubin, [*Approximation of Elliptic Boundary-Value Problems*]{}, Wiley, New York, 1972. A. Azouani, E. Olson, E. S. Titi, Continuous data assimilation using general interpolant observables, Preprint arXiv:1304.0997 (2013). C. Cao, E.S. Titi, Global well-posedness of the three-dimensional viscous primitive equations of large scale ocean and atmosphere dynamics, Annals of Math., 166 (2007), 245–267. B. Cockburn, D.A. Jones, E. S. Titi, Estimating the number of asymptotic degrees of freedom for nonlinear dissipative systems. Math. Comp., 66 (1997), 1073–1087. I. Chueshov, On the finiteness of the number of determining elements for von Karman evolution equations, [*Math. Meth. Appl. Sci.*]{}, 20 (1997), 855–865. I. Chueshov, Theory of functionals that uniquely determine asymptotic dynamics of infinite-dimensional dissipative systems, [*Russian Math. Surv.*]{}, 53 (1998), 731–776. I. Chueshov, *Introduction to the Theory of Infinite-Dimensional Dissipative Systems*. University Lectures in Contemporary Mathematics, Acta, Kharkov, 2002 (from the Russian edition (Acta, 1999)); see also $http://www.emis.de/monographs/Chueshov/$. I. Chueshov, A squeezing property and its applications to a description of long time behaviour in the 3D viscous primitive equations. Preprint arXiv:1211.4408 (2012); to be published in Proc. Royal Soc. Edinburgh, Ser.A. I. Chueshov, I. Lasiecka, *Long-Time Behavior of Second Order Evolution Equations with Nonlinear Damping*. Memoirs of AMS, vol.195, no. 912, AMS, Providence, RI, 2008. I. Chueshov, I. Lasiecka, [*Von Karman Evolution Equations*]{}. Springer, New York, 2010. *Data Assimilation. Making Sense of Observations*, (Eds: W.Lahoz, B. Khattatov, R. Ménard), Springer, New York, 2010. C. Foias, M. S. Jolly, R. Kravchenko, E. S. Titi, Navier-Stokes equations, determining forms, determining modes, inertial manifolds, dissipative dynamical systems, ArXiv:1208.5134v1 (2012). C. Foias, G. Prodi, Sur le comportement global des solutions non-stationnaires des équations de Navier-Stokes en dimension 2. Rend. Sem. Mat. Univ. Padova, 39 (1967), 1–34. K. Hayden, E. Olson and E.S. Titi, Discrete data assimilation in the Lorenz and 2D Navier–Stokes equations, Physica D, 240 (2011), 1416–1425. N. Ju, The global attractor for the solutions to the 3d viscous primitive equations, Disc. Cont. Dyn. Sys., 17 (2007), 159–179. E.Kalnay, *Atmospheric modeling, data assimilation and predictability*. Cambridge University Press, 2003. G. M. Kobelkov, Existence of a solution “in the large" for the 3D large-scale ocean dynamics equations. C. R. Acad. Sci. Paris, Ser.I, (2006) 343, 283–286. I. Kukavica, M. Ziane, On the regularity of the primitive equations of the ocean, Nonlinearity, 20 (2007), 2739–2753 O.A.Ladyzhenskaya, Finding minimal global attractors for the Navier-Stokes equations and other partial differential equations, Russian Math. Surveys, 42 (1987), 27–73. O.A.Ladyzhenskaya, [*Attractors for Semigroups and Evolution Equations*]{}. Cambridge University Press, 1991. J. L. Lions, R. Temam, S. Wang, On the equations of the large scale ocean, Nonlinearity, 5 (1992), 1007–1053. E. Olson and E.S. Titi, Determining modes for continuous data assimilation in 2-D turbulence, Journal of Statistical Physics, 113 (2003), 799–840. M. Petcu, On the three dimensional primitive equations, Adv. Dif. Eq., 11 (2006), 1201–1226. M. Petcu, R. Temam, M. Ziane, Some mathematical problems in geophysical fluid dynamics, in: Special Volume on Computational Methods for the Atmosphere and the Oceans, in: [*Handbook of Numerical Analysis, vol. 14*]{}, Elsevier, 2008, pp. 577–750. [^1]: Department of Mechanics and Mathematics, Karazin Kharkov National University, Kharkov, 61022, Ukraine, e-mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present an angle-resolved photoemission (ARPES) study on the layered transition-metal dichalcogenide 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ in the metallic commensurate charge-density-wave (CDW) phase. A model calculation of the spectral function captures the main features of the ARPES spectra well qualitatively, that is, the gross splits of unreconstructed band structure in the absence of the CDW superlattice. The observed enhancement of the size of the gap between the lower and middle fragments of the Ta [*5d*]{} band along the $\Gamma$M line by cooling is interpreted in terms of the increase in the CDW-related potential.' author: - 'Y. Aiura' - 'I. Hase' - 'H. Bando' - 'K. Yagi-Watanabe' - 'K. Tanaka' - 'K. Ozawa' - 'T. Iwase' - 'Y. Nishihara' - 'O. Shiino' - 'M. Oshima' - 'M. Kubota' - 'K. Ono' title: 'Spectroscopic Signatures on Increase in Charge-Density-Wave Potential of 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$' --- Quasi-two-dimensional (quasi-2D) materials, like the layered transition metal dichalcogenides, 1[*T*]{}-TaS$_{x}$Se$_{2-x}$, have attracted much attention because of their various unique physical properties combined with the formation of charge-density wave (CDW) [@Wilson75]. 1[*T*]{}-TaS$_{2}$ shows a metal-to-insulator (MI) transition at 180 K which occurs followed by nearly commensurate (NC) to commensurate (C) CDW transition [@Kim94a]. The MI transition is now understood in terms of a Mott localization, triggered by the NC-C transition [@Fazekas80]. On the other hand, isostructural and isoelectronic 1[*T*]{}-TaSe$_{2}$ remains metallic to very low temperature, suggesting that the Ta 5d electrons in Se compound are less susceptible to the Mott localization than those in S one [@Perfetti03]. The MI transition of 1[*T*]{}-TaS$_{x}$Se$_{2-x}$ occurs between x=1.2 and x=1.5 [@Shiino96; @Shiino00; @Horiba02a]. Spectroscopic signatures of an energy gap at the Fermi energy (E$_{F}$) were observed in the C phase in the sample with x=1.5 by scanning tunneling spectroscopy (STS), but not in the sample with x=1.2 [@Shiino96]. 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ exhibits only one (metallic) C phase below the incommensurate to C transition temperature, but the resistivity in the C phase is complicated [@Shiino96; @Horiba02a]. Therefore, 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ is suitable system for studying the anomalous metallic behavior near the Mott localization, because it is possible to control the potential of the superlattice [*in the same CDW phase*]{}. Angle-resolved photoemission spectroscopy (ARPES) and STS, which probes the single-particle spectral function, can provide a direct view of changes in the dramatic rearrangement of the electronic structure near the MI transition. The importance of the electron correlation effects in 1[*T*]{}-TaS$_{2}$ and 1[*T*]{}-TaSe$_{2}$ has been recognized enough by virtue of various STS [@Kim94a; @Kim94b] and ARPES [@Smith85; @Manzke89; @Claessen90; @Dardel92; @Zwick98; @Pillo99; @Pillo00; @Horiba02b; @Aiura03a; @Perfetti03]. In this paper, we report ARPES results on 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ in the metallic C phase taken at 300 K and 90 K. It was shown that the size of the gap, which was observed between the lower and middle fragments of the Ta [*5d*]{} band at the 0.65$\Gamma$M point, grows by cooling. From model calculations, it can be understood that the gap size is closely connected with the CDW-related potential. The width of the peak near E$_{F}$ in the calculated density of states (DOS) narrows with the increase in the CDW-related potential, which lends support to the notation of the anomalous metallic behavior near the Mott transition. In previous empirical tight-binding calculation of 1[*T*]{}-TaS$_{2}$ in the presence of commensurate CDW superlattice by Smith [*et al.*]{}, it has been shown that the Ta-derived [*d*]{} band collapses into three sub-manifolds separated by gaps [@Smith85]. Signatures of those manifolds and/or gaps were partially confirmed by many ARPES experiments [@Smith85; @Claessen90; @Manzke89; @Dardel92; @Zwick98]. However, it remains to be explained why the experimental band dispersive behavior of those manifolds except those gaps resemble so closely the theoretical band calculations in the normal state in the absence of the CDW superlattice, and why the three bands in the lower and middle manifolds could not be observed clearly even using the high-resolution ARPES [@Zwick98]. From a recent ARPES study, moreover, it was shown that the remnant Fermi surface of 1T-TaS$_{2}$ exhibits the symmetry of the one-particle normal state one [@Pillo99]. Those ARPES results mean that the overall shape of the electron wave functions is governed by the Fourier components of the crystal potential for the unreconstructed lattice, rather than the CDW-related contribution [@Claessen90; @Aiura03a]. In such weak competing potentials, recently, Voit [*et al.*]{} proposed model calculations on a quasi-one-dimensional (quasi-1D) material [@Voit00]. The band structure and spectral weight distribution obtained by diagonalizing the Hamiltonian truncated [*at the first order*]{} reproduced ARPES spectra very well. To interpret the ARPES spectra of quasi-2D material 1[*T*]{}-TaS$_{x}$Se$_{2-x}$ we developed the model calculations proposed for a quasi-1D material. First, we calculated band structure in absence of CDW superlattice using empirical tight-binding calculations [@Smith85]. Since the d$_{yz}$ and d$_{zx}$ orbitals reside at energies well above E$_{F}$, we can drop them from the basis set. Therefore, three bands that consist of the d$_{3z^{2}-r^{2}}$, d$_{x^{2}-y^{2}}$ and d$_{xy}$ orbitals are shown in the calculation. Since only the lowest band among them crosses E$_{F}$ and is occupied, we treated it mainly in the model calculations. The Hamiltonian is $$H_{{\bf k+g_{i}},{\bf k+g_{j}}}=\varepsilon({\bf k}+{\bf g_{i}})\delta_{{\bf g_{i}},{\bf g_{j}}}+[V_{{\bf g_{i}},{\bf g_{j}}}c^{\dag}_{{\bf k+g_{i}}}c_{{\bf k+g_{j}}}+H.c.].$$ $\varepsilon({\bf k+g_{j}})$ is the kinetic energy of the lowest band, the potential is $V_{{\bf g_{i}},{\bf g_{j}}}=V\delta_{|{\bf g_{i}}-{\bf g_{j}}|,g}$, ${\bf g_{i}}$ is one of the reciprocal lattice points of the $\sqrt{13}\times\sqrt{13}$ superstructure in the unreconstructed $1\times1$ BZ ($i$=0, . . ., 12), and [*g*]{} describes the size of the smallest reciprocal lattice vector ($\frac{4}{\sqrt{13}}\frac{\pi}{a}$, where $a$ is lattice constant). The single-particle spectral function $\rho({\bf k},\varepsilon)$ is given by $$\rho({\bf k}+{\bf g_{i}},\varepsilon)=\sum_{j=0}^{12} [a_{j}({\bf k})]_{i}^{2}\delta[\varepsilon-E_{j}({\bf k})],$$ in terms of the components of the eigenvectors $a_{j}$ associated with eigenvalues $E_{j}$ of the Hamiltonian (1). ![(a) Unreconstructed band of 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ along the $\Gamma$M line without the CDW superlattice (thick curve) and the Umklapp bands (thin and dotted curves) in the case for V=0 eV. The single-particle spectral function in the case for (b) V=0.05 and (c) 0.10 eV. The size of the symbols is proportional to the spectral weight. Calculated Ta [*d*]{} band density of states in the case for (d) V=0, (e) V=0.05, and (f) V=0.10 eV, respectively.](figure1) In the case for V=0 eV, the band of 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ along the $\Gamma$M line in absence of the CDW superlattice (thick curve) is shown in Fig. 1 (a). No spectral weight is expected for any of the additional bands derived from the Umklapp process (thin and dotted curves), i.e., weak Bragg diffraction through reciprocal-lattice vectors of the CDW superlattice. Figure 1 (b) and (c) show the single-particle spectral function in the case for V=0.05 and 0.10 eV, respectively. Figures 1 (d), (e), and (f) shows the calculated DOS for V=0, 0.05 and 0.10 eV, respectively. To compare with the ARPES spectra, the single-particle spectral function was multiplied by the Fermi-Dirac distribution function at 300K or 90K, and then convoluted with Gaussian functions of 150 meV and $\pm0.1\AA^{-1}$ FWHM which represent the instrumental energy and momentum resolution, respectively. Figure 2 (a) and (c) show the intensity plot and the energy distribution curves (EDCs) from the calculated single-particle spectral weight function in the case for V=0.05 eV and T=300K. Figure 2 (b) and (d) show the calculated intensity plot and the EDCs in the case for V=0.10 eV and T=90K. ![(Color online) (a) Intensity plot and (c) EDCs from the calculated single-particle spectral weight function in the case for V=0.05 eV and T=300K. (b) The intensity plot and (d) the EDCs in the case for V=0.10 eV and T=90K. Bright and dark parts correspond to high and low spectral weight, respectively.](figure2) We performed ARPES measurements at BL-1C of the Photon Factory (KEK, Tsukuba) using an electron spectrometer mounted on a two-axis goniometer (VG ARUPS10) [@Ono01]. The sample goniometer used here provides independent polar, azimuth and tilt rotation of the sample (R-Dec Co. Ltd., [*i*]{} GONIO LT) [@Aiura03b]. All ARPES spectra were taken at the photon energy (h$\nu$) of 40 eV. The samples were mounted vertically and only photoelectrons emitted from the plane defined by the light beam and the surface normal were observed. The emission angle of the photoelectron measured from the surface normal was varied by rotating the energy analyzer horizontally, whereas the angle of incidence of the light was fixed to $45^{\circ}$. The azimuth angle was varied by rotating the samples to the surface normal. Single crystals of 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ were grown by the iodine transport technique and characterized by resistivity measurement [@Shiino96; @Horiba02a]. The samples were cleaved [*in situ*]{} at a base pressure of $3\times10^{-10}$ Torr. We measured ARPES spectra at 300 K and at 90 K. The energy and spatial resolution were 0.15 eV and $\pm2^{\circ}$, respectively. ![(Color online) ARPES intensity plots at (a) 300K and (b) 90K and EDCs at (c) 300 K and (d) 90K along the high symmetry $\Gamma$M line (h$\nu$=40eV).](figure3) Figure 3 shows ARPES intensity plots at (a) 300K and (b) 90K and EDCs at (c) 300K and (d) 90K along the high symmetry $\Gamma$M line. At first sight, it is shown in Fig. 3 (a) that a Ta [*5d*]{} band disperses downward in energy along the high symmetry line from the $\Gamma$ point to the M point. Contrary to the prediction of previous band calculations of 1[*T*]{}-TaS$_{2}$ [@Mattheiss73; @Myron75; @Woolley77] and 1[*T*]{}-TaSe$_{2}$ [@Myron75; @Woolley77], however, the observed Ta [*5d*]{} band never crosses E$_{F}$ and lie around the $\Gamma$ point at a binding energy of about 0.2 eV, which is consistent with the previous ARPES spectra in 1[*T*]{}-TaS$_{2}$ in the metallic NC phase [@Pillo99; @Pillo00] and 1[*T*]{}-TaSe$_{2}$ in the metallic C phase [@Aiura03a]. Furthermore, the peak width of the Ta [*5d*]{} band is broadened at the 0.65$\Gamma$M point (thick curve in Fig. 3 (c)). This broad peak is divided into two structures by cooling (thick curve in Fig. 3 (d)). As the result, the intensity plot at 90K in Fig. 3 (b) shows the split of the Ta [*5d*]{} band at the 0.65$\Gamma$M point clearly. In addition, a weak hump appears at a binding energy of about 0.4eV around the $\Gamma$ point, as shown by ticks around the $\Gamma$ point in Fig. 3 (d). It is obvious to reproduce those spectral features of the Ta [*5d*]{} band by the model calculations well, except the absence of crossing over E$_{F}$. The split of the Ta [*5d*]{} band at the 0.65$\Gamma$ point and the appearance of the weak hump at a binding energy of about 0.4eV around the $\Gamma$ point is ascribed to the increase in V as shown in Fig. 2. Judging from comparison of the ARPES spectra at the 0.65$\Gamma$M point in Figs. 3 (c) and (d) (thick curves) and the corresponding spectral weight distributions in Figs. 2 (c) and (d) (thick curves), the potentials for the ARPES spectra at 300 K and 90 K were estimated at 0.05 eV and 0.1 eV, respectively. Although the split of the Ta [*5d*]{} band in 1[*T*]{}-TaS$_{2}$ and 1[*T*]{}-TaS$_{1.5}$Se$_{0.5}$ in the metallic NC phase is almost equal to that in 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ shown here, it is dramatically enhanced in the insulating C phase, meaning that the influence of the CDW-related potential became dramatically large in the insulating C phase [@Zwick98; @Horiba03]. With increasing V, the calculated spectral weight distribution shifts to a low binding energy side (not shown), which is consistent with the observed shift of the ARPES spectra in 1[*T*]{}-TaS$_{1.5}$Se$_{0.5}$ due to the NC-C phase transition [@Horiba02a; @Horiba03]. The calculated intensity plot and the EDCs for V=0.10 eV show another gap near E$_{F}$ at about 0.4$\Gamma$M due to the split of the Ta [*5d*]{} bandas as shown in Figs. 2(b) and (d). That is, the Ta [*5d*]{} band splits into three fragments under the influence of the CDW-related potential. The gap near E$_{F}$ at the 0.4$\Gamma$M point was well observed for 1[*T*]{}-TaS$_{2}$ in the insulating C phase [@Zwick98]. Also for 1[*T*]{}-TaS$_{2}$ in the metallic NC phase [@Pillo00] and for 1[*T*]{}-TaSe$_{2}$ in the metallic C phase [@Aiura03a], the gap was not clear because the upper fragment in the metallic phase is very broad compared with that in the insulating phase. No crossing of the Fermi level is supposed to be induced by the electron correlation effects in the Ta [*5d*]{} band. According to the model of Tosatti and Fazekas [@Fazekas80], the (partially filled) upper fragment straddling E$_{F}$ split into the lower Hubbard band (LHB) and the upper Hubbard band (UHB) by the electron correlation effects. For 1[*T*]{}-TaS$_{2}$ in the insulating C phase the flat LHB was observed around the $\Gamma$ at a binding energy of 0.19 eV [@Zwick98] and in the metallic NC phase [@Pillo00], and at about 0.2 eV for 1[*T*]{}-TaSe$_{2}$ in the metallic C phase [@Aiura03a]. For the metallic phase, E$_{F}$ lies in a pseudogap created by the tails of two overlapping Hubbard subbands. Based on our model calculation, the width of the upper fragment straddling E$_{F}$, i.e., the width of the Hubbard bands, depends on the strength of the potential, V. This is supported by the calculated DOS spectra in Figs. 1 (d), (e) and (f). No peak in DOS for V=0 eV is shown near E$_{F}$ in Fig. 1(d). A broad peak appears under V=0.05 eV (Fig. 1 (e)) and becomes sharp with increasing the strength of V to 0.1 eV (Fig. 1 (f)). ![(Color online) Experimental band dispersion at 90 K along the high symmetry lines $\Gamma$M, MK and K$\Gamma$. (b) Corresponding intensity plot of the calculated spectral weight distribution in the case for V=0.1eV and T=90 K.](figure4) Figure 4 shows (a) the experimental band dispersion at 90 K along the high symmetry lines $\Gamma$M, MK and K$\Gamma$ and (b) the corresponding intensity plots of the calculated spectral weight distribution in the case for V=0.1eV and T=90 K. It is shown that the model calculations reproduce the main features of the ARPES results very well, e.g., the dominant spectral feature from the lower subband along the MK line and the appearance of the weak spectral weight around the middle of the $\Gamma$K like. This means that the model calculations presented here, which are originally proposed for a quasi-1D material [@Voit00], are also useful for a quasi-2D material, 1[*T*]{}-TaS$_{x}$Se$_{2-x}$. It is reasonable to suppose that the slight difference in the energy position of the spectral features near E$_{F}$ is based on the electron correlation effects, which are not included in the model calculation. Since V in Eq. (1) represents the coupling strength between the unreconstructed band and the CDW-derived bands, the spectral changes by temperature may depend on the size of the atomic displacement and/or the domain size of the CDW superlattice. To investigate the electronic behavior due to the MI transition by a Mott transition, 1[*T*]{}-TaS$_{2}$ is not suitable because the MI transition is triggered by the structural NC-C phase transition. Recently, Perfetti [*et al.*]{} reported that a MI transition without accompanying such structural phase transition occurs at the surface of 1[*T*]{}-TaSe$_{2}$, which is an ideal material to tune the crucial parameter (W/U), where U is the on-site Coulomb correlation energy and W is the bandwidth [@Perfetti03]. However, it is hard to know the role of the CDW-related potential in the electronic behavior of the Ta 5d in a situation close to the Mott transition since the ARPES spectra were shown only in the limited area of the Brillouin zone. To study this, the elucidation of the entire electronic structure of the Ta [*5d*]{} band is desired strongly. In summary, we presented ARPES spectra of 1[*T*]{}-TaS$_{1.2}$Se$_{0.8}$ in the metallic phase taken at 300 K and 90 K and compared them with the calculated spectra based on the model proposed by Voit [*et al.*]{} [@Voit00]. It was shown that the calculated single-particle spectral function capture the main features of the ARPES spectra well qualitatively. We can be fairly certain that the model calculations shown here play an important role to understand the physical properties of quasi-two-dimensional transition-metal dichalcogenides.\ This work was partly done under Project No. 2002G174 at the Institute of Material Structure Science in KEK. [99]{} J. A. Wilson, F. J. DiSalvo, and S. Mahajan, Adv. Phys. [**24**]{}, 117 (1975). J.-J. Kim, W. Yamaguchi, T. Hasegawa, and K. Kitazawa, Phys. Rev. Lett. [**73**]{}, 2103 (1994). P. Fazekas and E. Tosatti, Physica B [**99**]{}, 183 (1980). A recent ARPES work on 1[*T*]{}-TaSe$_{2}$ reported that a metal-insulator Mott transition occurs at the surface, which is driven by the localization of the Ta [*5d*]{} band at the surface; L. Perfetti, A. Georges, S. Florens, S. Biermann, S. Mitrovic, H. Berger, Y. Tomm, H. Höchst, and M. Grioni, Phys. Rev. Lett. [**90**]{}, 166401 (2003). O. Shiino, T. Endo, W. Yamaguchi, H. Sugawara, T. Hasegawa, and K. Kitazawa, Czech. J. Phys. [**46**]{}, 2621 Suppl. S5 (1996). O. Shiino, T. Watanabe, T. Endo, T. Hanaguri, K. Kitazawa, M. Nohara, H. Takagi, C. Murayama, N. Takeshita, N. Môri, T. Hasegawa, and W. Yamaguchi, Physica B [**284-288**]{}, 1673 (1996). K. Horiba, K. Ono, M. Oshima, Y. Aiura, and O. Shiino, Surf. Rev. Lett. [**9**]{}, 1085 (2002). J. -J. Kim, C. Park, W. Yamaguchi, O. Shiino, K. Kitazawa and T. Hasegawa, Phys. Rev. B [**56**]{}, 15 573 (1997). N. V. Smith, S. D. Kevan and F. J. DiSalvo, J. Phys. C [**18**]{} ,3175 (1985). R. Manzke, T. Buslaps, B. Pfalzgraf, M. Skibowski, and O. Anderson, Europhys. Lett. [**8**]{}, 195 (1989). R. Claessen, B. Burandt, H. Carstensen, and M. Skibowski, Phys. Rev. B [**41**]{}, 8270 (1990). B. Dardel, M. Grioni, D. Malterre, P. Weibel, Y. Baer, and F Lévy, Phys. Rev. B [**45**]{}, 1462 (1992); [*ibid.*]{} B [**46**]{}, 7407 (1992). F. Zwick, H. Berger, I. Vobornik, G. Margaritondo, L. Forró, C. Beeli, M. Onellion, G. Panaccione, A. Taleb-Ibrahimi, and M. Grioni, Phys. Rev. Lett. [**81**]{}, 1058 (1998). Th. Pillo, J. Hayoz, H. Berger, M. Grioni, L. Schlapbach, and P. Aebi, Phys. Rev. Lett. [**83**]{}, 3494 (1999). Th. Pillo, J. Hayoz, H. Berger, R. Fasel, L. Schlapbach, and P. Aebi, Phys. Rev. B [**62**]{}, 4277 (2000); P. Aebi, Th. Pillo, H. Berger, and F. Lévy, J. Electron Spectrosc. Relat. Phenom. [**117-118**]{}, 433 (2001). K. Horiba, K. Ono, J. H. Oh, T. Kihara, S. Nakazono, M. Oshima, O. Shiino, H. W. Yeom, A. Kakizaki, and Y. Aiura, Phys. Rev. B [**66**]{}, 073106 (2002). Y. Aiura, H. Bando, R. Kitagawa, S. Maruyama, Y. Nishihara, K. Horiba, M. Oshima, O. Shiino, M. Nakatake, Phys. Rev. B [**68**]{}, 073408(2003). J. Voit, L. Perfetti, F. Zwick, H. Berger, G. Margaritondo, G. Grüner, H. Höchst, M. Grioni, Science [**290**]{}, 501 (2000). K. Ono, J. H. Oh, K. Horiba, M. Mizuguchi, M. Oshima, T. Kiyokura, F. Maeda, Y. Watanabe, A. Kakizaki, T. Kikuchi, A. Yagishita, and H. Kato, Nucl. Instrum. Meth. A [**467**]{}, 573 (2001); K. Ono, H. W. Yeom, K. Horiba, J. H. Oh, S. Nakazono, T. Kihara, K. Nakamura, T. Mano, M. Mizuguchi, M. Oshima, Y. Aiura, and A. Kakizaki, [*ibid.*]{} [**467**]{}, 1497 (2001). Y. Aiura, H. Bando, T. Miyamoto, A. Chiba, R. Kitagawa, S. Maruyama, Y. Nishihara, Rev. Sci. Instrum. [**74**]{}, 3177 (2003). L. F. Mattheiss, Phys. Rev. B [**8**]{}, 3719 (1973). H. W. Myron and A. J. Freeman, Phys. Rev. B [**11**]{}, 2735 (1975). A. M. Woolley and G. Wexler, J. Phys. C [**10**]{}, 2601 (1977) Horiba, Ph. D. thesis, Tokyo University, 2003.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We describe a Quillen equivalence between *quasi-categories* and *relative categories* which is surprisingly similar to Thomason’s Quillen equivalences between *simplicial sets* and *categories*.' address: - 'Department of Mathematics, Massachusetts Institute of Technology, Cambridge, MA 02139' - 'Department of Mathematics, Massachusetts Institute of Technology, Cambridge, MA 02139' author: - 'C. Barwick' - 'D.M. Kan' title: 'A Thomason-like Quillen equivalence between quasi-categories and relative categories' --- Introduction {#sec:Intro} ============ In [@JT] Joyal and Tierney constructed a Quillen equivalence $${\boldsymbol{S}} {\mathchoice{\longleftrightarrow} {\leftrightarrow} {\leftrightarrow} {\leftrightarrow}}{\mathrm{s}}{\boldsymbol{S}}$$ between the Joyal structure on the category ${\boldsymbol{S}}$ of small simplicial sets and the Rezk structure on the category ${\mathrm{s}}{\boldsymbol{S}}$ of simplicial spaces (i.e. bi-simplicial sets) and in [@BK] we described a Quillen equivalence $${\mathrm{s}}{\boldsymbol{S}} {\mathchoice{\longleftrightarrow} {\leftrightarrow} {\leftrightarrow} {\leftrightarrow}}{\mathbf{RelCat}}$$ between the Rezk structure on ${\mathrm{s}}{\boldsymbol{S}}$ and the induced Rezk structure on the category ${\mathbf{RelCat}}$ of relative categories. In this note we observe that the resulting composite Quillen equivalence $${\boldsymbol{S}} {\mathchoice{\longleftrightarrow} {\leftrightarrow} {\leftrightarrow} {\leftrightarrow}}{\mathbf{RelCat}}$$ admits a description which is almost identical to that of Thomason’s [@T] Quillen equivalence $${\boldsymbol{S}} {\mathchoice{\longleftrightarrow} {\leftrightarrow} {\leftrightarrow} {\leftrightarrow}}{\mathbf{Cat}}$$ between the classical structure on ${\boldsymbol{S}}$ and the induced on on the category ${\mathbf{Cat}}$ of small categories, as reformulated in [@BK]\*[6.7]{}. [ 6 pt plus 9 pt ]{}To do this we recall from [@BK]\*[4.2 and 4.5]{} the notion of The two-fold subdivision of a relative poset {#sec:2foldsbdv} ============================================ For every $n \ge 0$, let $$\check{{\boldsymbol{n}}} \qquad\text{(resp. $\hat{{\boldsymbol{n}}}$)}$$ denote the relative poset which has as underlying category the category $$0 \longrightarrow {{\cdot\;\cdot\;\cdot}}\longrightarrow n$$ and in which the weak equivalences are only the *identity maps* (resp. *all maps*). Given a relative poset ${\boldsymbol{P}}$, its **terminal** (resp.**initial**) **subdivision** then is the relative poset $\xi_{t}{\boldsymbol{P}}$ (resp. $\xi_{i}{\boldsymbol{P}}$) which has 1. as *objects* the *monomorphisms* $$\check{{\boldsymbol{n}}} \longrightarrow {\boldsymbol{P}} \qquad\text{($n \ge 0$)}$$ 2. as *maps* $$\begin{aligned} (x_{1}\colon \check{{\boldsymbol{n}}}_{1} \to {\boldsymbol{P}})&\longrightarrow (x_{2}\colon \check{{\boldsymbol{n}}}_{2} \to {\boldsymbol{P}})\\ \text{(resp. }(x_{2}\colon \check{{\boldsymbol{n}}}_{2} \to {\boldsymbol{P}}) &\longrightarrow (x_{1}\colon \check{{\boldsymbol{n}}}_{1} \to {\boldsymbol{P}})\text{)} \end{aligned}$$ the commutative diagrams of the form $$\xymatrix{ {\check{{\boldsymbol{n}}}_{1}} \ar[rr] \ar[dr]_{x_{1}} && {\check{{\boldsymbol{n}}}_{2}} \ar[dl]^{x_{2}}\\ & {{\boldsymbol{P}}} }$$ and [3]{} as *weak equivalences* those of the above diagrams in which the induced map $$x_{1}n_{1} \longrightarrow x_{2}n_{2} \qquad\text{(resp.\ $x_{2}0 \longrightarrow x_{1}0$)}$$ is a weak equivalence in ${\boldsymbol{P}}$. The **two-fold subdivision** of ${\boldsymbol{P}}$ then is the relative poset $$\xi{\boldsymbol{P}} = \xi_{t}\xi_{i}{\boldsymbol{P}} {\rlap{\enspace .}}$$ Conclusion {#sec:conc} ========== In view of [@JT]\*[4.1]{} and [@BK]\*[5.2]{} we now can state: 1. *the left adjoint in the above composite Quillen equivalence* $${\boldsymbol{S}} {\mathchoice{\longleftrightarrow} {\leftrightarrow} {\leftrightarrow} {\leftrightarrow}}{\mathbf{RelCat}}$$ *is the colimit preserving functor which for every integer $n \ge 0$ sends* $$\Delta[n] \in {\boldsymbol{S}} \text{\qquad to\qquad} \xi\check{{\boldsymbol{n}}} \in {\mathbf{RelCat}}$$ *and the right adjoint sends an object $X \in {\mathbf{RelCat}}$ to the simplicial set which in dimension $n$ ($n \ge 0$) consists of the maps $\xi\check{{\boldsymbol{n}}} \to X \in {\mathbf{RelCat}}$.* while, in view of the fact that ${\mathbf{Cat}}$ is canonically isomorphic to the full subcategory $$\widehat{\mathbf{Cat}}\in{\mathbf{RelCat}}$$ spanned by the relative categories in which *every map is a weak equivalence* and [@BK]\*[6.7]{}, [2]{} *the left adjoint in Thomason’s Quillen equivalence* $${\boldsymbol{S}} {\mathchoice{\longleftrightarrow} {\leftrightarrow} {\leftrightarrow} {\leftrightarrow}}\widehat{\mathbf{Cat}}$$ *is the colimit preserving functor which, for every integer $n \ge 0$, sends* $$\text{$\Delta[n]\in{\boldsymbol{S}}$ \emph{ to } $\xi\hat{{\boldsymbol{n}}} \in\widehat{\mathbf{Cat}}$} \qquad \text{($n \ge 0$)}$$ *while the right adjoint sends an object $X \in \widehat{\mathbf{Cat}}$ to the simplicial set which in dimension $n$ ($n \ge 0$) consists of the maps* $$\xi\hat{{\boldsymbol{n}}} \longrightarrow X \in \widehat{\mathbf{Cat}}$$
{ "pile_set_name": "ArXiv" }
--- abstract: '[The linearized Davey-Stewartson equation with varing coefficients is solved by Fourier method. The approach uses the inverse scattering transform for the Davey-Stewartson equation.]{}' --- =msbm10 [**The Fourier method for the linearized Davey-Stewartson I equation\ O. M. Kiselev**]{}\ Ufa Institute of Mathematics, Russian Academy of Sciences,\ 112 Chernyshevsky str., Ufa, 450000, Russia\ E-mail: [email protected] Introduction ============ The Davey-Stewartson equation (DS) is the well-known subject for investigations because of two causes. First, the equation models a nonlinear interaction between a long surface wave and a short surface wave [@D-S]. Second, DS is integrable by an inverse scattering transform (IST) [@F-A] and then one can investigate a solution structure in details. In this work we consider a linear system of equations with varing coefficients, which come out when one linearizes the DS at a nonzero solution as a background. We develope the Fourier method for the above linear system. This method is based on IST results for DS [@F-A; @F-S; @N]. The basic functions, which are used for the Fourier expansion, are assosiated with a scattering problem for a Dirac system. A realized approach one can see as generalizing on (2+1)-dimensions (two spatial variables and time) system of pioneer works of D.Kaup [@K1; @K2], in which the Fourier method was formulated for linearized (1+1)-dimensional integrable equations. The obtained results make possible to study perturbations of the DS, in general putting out the equation from the class of integrable equations. The Dirac equation and the basic functions ========================================== Here we consider a Goursat problem for the Dirac equation [@F-A]-[@N]: ( [cc]{} \_& 0\ 0 & \_ ) =-[12]{}( [cc]{} 0 & q\_1\ q\_2 & 0 )\[de\] Define a solutions of (\[de\]) as follows [@F-A]: $$\begin{array}{cc} \psi^+_{11}|_{\xi\to-\infty}=\exp(ik\eta),&\psi^+_{12}|_{\xi\to-\infty}=0,\\ \psi^+_{21}|_{\eta\to\infty}=0,&\psi^+_{22}|_{\eta\to-\infty}=\exp(-ik\xi);\\ \psi^-_{11}|_{\xi\to-\infty}=\exp(ik\eta),&\psi^-_{12}|_{\xi\to+\infty}=0,\\ \psi^-_{21}|_{\eta\to-\infty}=0,&\psi^-_{22}|_{\eta\to-\infty}=\exp(-ik\xi).\\ \end{array}$$ Denote by $(\chi,\mu)_{q}$ a bilinear form: (,)\_q=\_[-]{}\^\_[-]{}\^dd(\_1\_1q\_2+\_2\_2q\_1), \[f1\] where $\chi_i$ and $\mu_i$ are elements of the columns $\chi$ and $\mu$. Denote by $\phi^{(1)}$ and $\phi^{(2)}$ the solutions which are dual to $\psi_{(1)}$ (first column of the matrix $\psi^+$) and $\psi_{(2)}$ (second column of the matrix $\psi^-$) with respect to the bilinear form (\[f1\]). It is known, that the solution of (\[de\]) satisfies a nonlocal Riemann-Hilbert problem [@F-S]. Formulate this problem for $\psi^-_{11}$ and $\psi^+_{12}$. Denote by $\psi^{(1)}$ a row $\{\psi^-_{11},\psi^+_{12}\}$, then [@F-S]: $$\psi^{(1)}=E^{(1)}(ik\eta)+S[s]\psi^{(1)}.$$ Here $E^{(1)}(z)$ is the first row of a matrix $E(z)=diag(\exp(z),\exp(-z))$, the operator $S[s]$ is defined by formula: $$S[s]\psi^{(1)}=\big[\exp(ik\eta)\bigg(\exp(-ik\eta)\int_{-\infty}^{\infty}dl s_1(k,l)\psi^+_{12}(\xi,\eta,l)\bigg)^-,$$ $$\exp(ik\xi)\bigg(\exp(-ik\xi)\int_{-\infty}^{\infty}dl s_1(k,l)\psi^-_{11}(\xi,\eta,l)\bigg)^+ \big],$$ where $$\bigg(f(k)\bigg)^{\pm}={1\over2i\pi}\int_{-\infty}^{\infty} {dk'\,f(k')\over k'-(k\pm i0)}.$$ Denote by $<\chi,\mu>_s$ a bilinear form: \_s=\_[-]{}\^\_[-]{}\^dk dl (\^1(l)\^1(k)s\_2(k,l)+\^2(l)\^2(k)s\_1(k,l)), \[f2\] where $\chi^{j}$ is the element of the row $\chi$. Let $\varphi^{(j)},\, j=1,2$ be rows which are solution of the equations conjugated to the equation for the rows $\psi^{(j)}=\{\psi^-_{j1}, \psi^+_{j2}\}$ with respect to the bilinear form (\[f2\]). Let $q_1$ and $q_2$ satisfy following conditions: $\partial^{\alpha} q_{1,2}\in L_1\cap C$ at $|\alpha|\le3$. If $h_1(\xi,\,\eta)$ and $h_2(\xi,\,\eta)$ satisfy the conditions $\partial^{\alpha}h_{1,2}\in L_1\cap C$ at $|\alpha|\le4$, then $h_1$ and $h_2$ may be represented in the form: [c]{} h\_1=[-1]{}&lt;\^[(1)]{}(,,l),\^[(1)]{}(,,k)&gt;\_[h]{},\ h\_2=[1]{}&lt;\^[(2)]{}(,,l),\^[(2)]{}(,,k)&gt;\_[h]{}, \[inverse\] where [c]{} h\_1=[14]{}(\^+\_[(1)]{}(,,k),\_[(1)]{}(,,l))\_h,\ h\_2=[14]{}(\^-\_[(2)]{}(,,k),\_[(2)]{}(,,l))\_h. \[direct\] Of course, if $q_1=q_2=0$, then the formulae (\[direct\]) and (\[inverse\]) are ordinary Fourier transform with respect to two variables $\xi,\eta\in\Real$. The Fourier method for the linearized Davey-Stewartson I equation. ================================================================== We shall consider the Davey-Stewartson I equation: [cc]{} i\_t Q+(\_\^2+\_\^2)Q+(g\_1+g\_2)Q=0,\ \_g\_1=-[2]{}\_|Q|\^2,\_g\_2=-[2]{}\_|Q|\^2,=1. \[ds1\] Linearization of this equation on $Q,\,g$ as a background gives: [cc]{} i\_t U+(\^2\_+\_\^2)U+(g\_1+g\_2)U+(V\_1+V\_2)Q=0,\ \_V\_1=-[2]{}\_(Q|U+|Q U),\_V\_2=-[2]{}\_(Q|U+|Q U). \[lds1\] The equation (\[ds1\]) is a compatible condition for (\[de\]) at $q_1=Q$, $q_2=\varepsilon\bar Q$ and for a following system [@A-S]: $$\partial_t \psi= i\Big(\begin{array}{cc} 1&0\\0&-1\end{array}\Big) (\partial_\xi-\partial_\eta)^2\psi+ i\Big(\begin{array}{cc} 0&q_1\\q_2&0\end{array}\Big) (\partial_\xi-\partial_\eta)\psi+$$ +( [cc]{} ig\_1&-i\_q\_1\ i\_q\_2&-ig\_2 ) . \[dtpsi\] Using the systems (\[de\]) and (\[dtpsi\]) one can prove the following statement: Let $U$ be the solution of (\[lds1\]) and $Q$ be the solution of the Davey-Stevartson I equation with boundary conditions: $g_1|_{\xi\to\infty}=0$ and $g_2|_{\eta\to\infty}=0$, and $U$ and $Q$ satisfy the conditions of the theorem 1 for the functions $h_1$ and $q_1$ at $\forall t\in[0,T]$ respectively, then: \_t U\_1=i(k\^2+l\^2)U\_1,\_t U\_2=-i(k\^2+l\^2)U\_2, \[efc\] where $\hat U_1={1\over4\pi}(\psi^{-}_{(1)},\phi_{(1)})_U$, $\hat U_2={1\over4\pi}(\psi^{+}_{(2)},\phi_{(2)})_U$. [**Inverse statement.**]{} Let $\hat U_{1}(t,k,l)$ and $\hat U_{2}(t,k,l)$ be integrable with respect to $k,\,l$ with factor $(1+k^2)(1+l^2)$ and satisfy equations (\[efc\]), then the function $$U(\xi,\eta,t)={-1\over\pi} <\psi^{(1)}(\xi,\eta,l),\varphi^{(1)}(\xi,\eta,k)>_{\hat U}$$ satisfies linearized Davey-Stewartson I equation (\[lds1\]) Solving of the system (\[lds1\]) is reduced to solving of the trivial equations (\[efc\]) for the Fourier coefficients $\hat U_{1,2}$. The theorems 1 and 2 give solving (\[lds1\]) by Fourier method. [99]{} A. Davey, K. Stewartson. On three-dimensional packets of surface waves. Proc. Royal Soc. London, 1973, v.338, p.101-110. A.S. Fokas, M.J. Ablowitz. On the inverse scattering transform of multidimensional nonlinear equations related to first-order systems in the plane. J. Math. Phys. 25(8), 1984, p.2494-2505. A.S. Fokas, P.M.Santini. Dromions and a boundary value problem for the Davey-Stewartson I equation. Physica D, 44(1990), p.99-130. L.P.Niznik. Inverse scattering problems for hyperbolic equations. Naukova Dumka, Kiev, 1991, 232 pp. (in russian) D.J.Kaup. Closure of the squared Zakharov-Shabat eigenstates. J.of Math.Anal.and Appl., v.54, p.849-864, (1976). D.J.Kaup. A perturbation expansion for the Zakharov-Shabat inverse scaterring transform. SIAM J.on Appl.Math., v.31, n1, p.121-133, (1976). M.J.Ablowitz, H.Segur. Solitons and the Inverse Scattering Transform. SIAM, Philadelphia., 1981.
{ "pile_set_name": "ArXiv" }
--- abstract: 'A multi-channel algebraic scattering theory, to find solutions of coupled-channel scattering problems with interactions determined by collective models, has been structured to ensure that the Pauli principle is not violated. By tracking the results in the zero coupling limit, a correct interpretation of the sub-threshold and resonant spectra of the compound system can be made. As an example, the neutron-${}^{12}$C system is studied defining properties of ${}^{13}$C to 10 MeV excitation. Accounting for the Pauli principle in collective coupled-channels models is crucial to the outcome.' author: - 'L. Canton$^{(1)}$' - 'G. Pisent$^{(1)}$' - 'J. P. Svenne$^{(2)}$' - 'D. van der Knijff$^{(3)}$' - 'K. Amos$^{(4)}$' - 'S. Karataglidis$^{(4)}$' bibliography: - 'MCAS-PRL.bib' title: 'Role of the Pauli principle in collective-model coupled-channels calculations' --- At energies above 25 MeV, by using optical potentials formed by full folding effective two-nucleon interactions with microscopic (nucleon based) descriptions of the target structure, the importance of treating the Pauli principle has been well established [@Am00]. However, in the domain of low-energy nucleon scattering for which an explicit coupled-channels theory of scattering is essential, the significance of Pauli exclusion effects has not been well defined. Many coupled-channels codes are available, some of which perform phenomenological collective-model calculations searching on parameter values of the chosen function forms to find a best fit to experimental data. But while it has long been known that any such models violate the Pauli principle [@Ma69; @Gr96], quantification of that violation is lacking. To study the effects of the Pauli principle in a macroscopic (collective model) approach is not a trivial task. In a recent publication [@Am03], the orthogonalizing pseudo-potential (OPP) method [@Ku78; @Sa69] was generalized to treat this problem. That was a small though important part of the full theoretical framework of the multi-channel algebraic scattering (MCAS) theory of scattering [@Am03]. Therein the OPP was used in finding the spectra, bound and resonance properties, of ${}^{13}$C. However, implications of the role of the Pauli principle in collective model coupled-channel calculations arising from the use of the OPP was not discussed. Such is a purpose of this letter. Another is that the method could be pertinent for any study requiring coupled channel solutions of quantal systems involving fermions. As the example, we study the effects introduced by the Pauli principle in collective, geometrical-type, models for low-energy nucleon-nucleus processes that can be characterized from the spectrum of the compound nucleus. That spectrum includes the states that lie below the nucleon-nucleus threshold and in the continuum as revealed by the narrow and broad resonances that lie upon a smooth but energy dependent background of the elastic scattering cross section. This can be done in a systematic and self-consistent way since the MCAS approach facilitates such a determination of the sub-threshold bound states and resonances of the compound nucleus. This theory, with which one solves the coupled-channel Lippmann-Schwinger (LS) equations for the nucleon-nucleus system considered, is built upon sturmian expansions of an interaction matrix of potential functions. The MCAS method has been developed in momentum space and the starting matrix of potentials may be formed by folding effective two-nucleon interactions with one-body density matrices of the target (studies in progress) or, as is more common, from a collective model description of the target states and excitations. As in that recent publication [@Am03], we have used a rotational collective-model representation with deformation taken to second order. We chose Woods-Saxon functions and their various derivatives to be the form factors for all components each with characteristic operators of diverse type. The interactions were allowed to depend on parity as well. With such a characterization, we were able to describe all important aspects, at positive and negative energies, in the neutron-$^{12}$C system. With the MCAS approach and a collective model prescription for the starting matrix of potentials, the OPP is used in the process by which the sturmians are specified. The OPP inclusion ensures that all sturmians in the (finite) set selected as the basis of expansion of the matrix of potentials contain few or no components equivalent to the external nucleon being placed in an already densely occupied orbit. That scheme is an approximation as we discuss later by assessing the spectra of $^{12,13}$C and the single neutron spectroscopic amplitudes that link them using information obtained from large space no-core shell model calculations. But it is a good approximation. The role of the Pauli principle is studied by comparing results found with and without using the OPP scheme to select the sturmians that form the expansion set. Note that the actual matrix of potentials is the same throughout though extra information on single nucleon plus a core nucleus state underlying each sub-threshold bound and resonance in the compound system has been obtained by taking the zero deformation limit. Full details of the MCAS scheme have been published [@Am03] and the reader is referred there for those, as well as for specifics of the notation we use herein. In momentum space for potential matrices $V_{cc'}(p,q)$, one seeks the solution of coupled LS equations \[see Eq. (1) in Ref. [@Am03]\], which involve both open and closed channel contributions. With incident energy $E$, the channel wave numbers for the open and closed channels are $k_c$ and $h_c$ respectively. Solutions of those LS equations are sought using expansions of the potential matrix elements in (finite) sums of energy-independent separable terms, $$V_{cc'}(p,q) \sim \sum^N_{n = 1} \hat{\chi}_{cn}(p) \eta^{-1}_n \hat{\chi}_{c'n}(q)\; , \label{finiteS}$$ where $\hat{\chi}_{c'n}(q)$ are the Fourier-Bessel transforms of the selected sturmians whose eigenvalues are $\eta_n$. To predict observables one requires the multichannel $S$-matrix. In terms of the multi-channel $T$-matrix, that has closed algebraic form $$\begin{aligned} S_{cc'} & = & \delta_{cc'} -i \pi \mu \sqrt{k_c k_{c'}}\; T_{cc'} \nonumber \\ T_{cc'} & = & \sum_{n,n' = 1}^N \hat{\chi}_{cn}(k_c) \left([\mbox{\boldmath $\eta$} - \mathbf{G}_0]^{-1} \right)_{nn'}\ \hat{\chi}_{c'n'}(k_{c'})\ , \label{multiS} \end{aligned}$$ where now $c,c'$ refer to open channels only. In this representation, $\mathbf{G}_0$ and have matrix elements $$\begin{aligned} \left[ \mathbf{G}_0 \right]_{nn'} & = \mu \left[ \sum^{\text{open}}_c \int^{\infty}_0 \hat{\chi}_{cn}(x) \frac{x^2}{k^2_c - x^2 +i\varepsilon} \hat{\chi}_{cn'} \, dx \right. \nonumber \\ & \phantom{= \mu} - \left. \sum^{\text{closed}}_c \int^{\infty}_0 \hat{\chi}_{cn}(x) \frac{x^2}{h^2_c + x^2} \hat{\chi}_{cn'}(x) \, dx \right] \; , \nonumber \\ \left[ \mbox{\boldmath $\eta$} \right]_{nn'} & = \eta_n \delta_{nn'} \; . \label{xiGels} \end{aligned}$$ The bound states of the compound system are defined by the zeros of the matrix determinant when the energy $E$ is negative, and so link to the zeros of $\{ \left| \mbox{\boldmath $\eta$}-\mathbf{G}_0\right| \}$ when all channels in Eq. (\[xiGels\]) are closed. As noted above the sturmians are solutions of homogeneous Schrödinger equations for the matrix of potentials. In coordinate space if those potentials are designated by local forms $V_{cc'}(r) \delta(r-r')$, the OPP method is to use sturmians that are solutions for nonlocal potentials $$\mathcal{V}_{cc'}(r,r') = V_{cc'}(r)\delta(r-r') + \lambda A_c(r) A_c(r')\delta_{cc'} ,$$ where $A(r)$ is the radial part of the single particle bound state wave function in channel $c$ spanning the phase space excluded by the Pauli principle. The OPP method holds in the limit $\lambda \to \infty$, but use of $\lambda = 100$ MeV suffices. The spectrum of $^{12}$C also was calculated in the shell model using the program OXBASH [@Ox86] and with the MK3W interaction [@Wa89]. The positive parity states of $^{12}$C were calculated in a complete $(0+2)\hbar\omega$ space using this interaction, while the negative parity states were calculated in a restricted $(1+3)\hbar\omega$ space. In both calculations the same single particle basis of $0s$ up to and including the $0f1p$ shell was used. Hence the restriction from a full $(1+3)\hbar\omega$ study is that we have not included the $0g1d2s$ shell. With exceptions, most notably the super-deformed $0^+_2$ state at 7.654 MeV and the known collective $3^-$ state at 9.64 MeV, the calculated spectrum to 20 MeV excitation agrees well with observation [@Am00]. So also do results of calculations [@Am00] of elastic and inelastic scattering data (form factors from electron scattering and differential cross sections and analyzing powers from proton scattering) without the need for any *a posteriori* core polarization corrections. Of interest here are the details of the low lying spectrum. First, in Table \[C12occs\], we list the nucleon shell occupancies in the three lowest states of $^{12}$C. Clearly the $0s$- and $0p$- shells have dense occupancy: essentially 4 nucleons filling the $0s$ shell while there are almost 8 nucleons in the $0p$ shell. Those eight nucleons are distributed between the sub-shells, so blocking the $0p_{\frac{3}{2}}$ orbit as we do in using the OPP method is an approximation. Note also that the second excited state of $^{12}$C, the $0^+_2$(7.654 MeV) is well known to be a super-deformed $3\alpha$ chain and, as such, a much larger space is needed for a good description. orbit $0^+_1$ $2^+_1$ $0^+_2$ -------------------- --------- --------- --------- $0s_{\frac{1}{2}}$ 1.963 1.962 1.968 $0p_{\frac{3}{2}}$ 3.054 2.858 3.075 $0p_{\frac{1}{2}}$ 0.842 1.028 0.804 higher orbits 0.124 0.129 0.120 : \[C12occs\] Shell occupancies of protons (or neutrons) in states of $^{12}$C. The lowest three $0^+$ states in our shell model are the ground, at 12.25, and at 23.03 MeV. They are much more spread than measured energies and have structure $$\begin{aligned} \left| {}^{12}\text{C} \left( 0^+_1 \right) \right\rangle & = 80.525\% \left| 0\hbar\omega \right\rangle + 19.475\% \left| 2\hbar\omega \right\rangle \\ \left| {}^{12}\text{C} \left( 0^+_2 \right) \right\rangle & = 78.213\% \left| 0\hbar\omega \right\rangle + 21.786\% \left| 2\hbar\omega \right\rangle \\ \left| {}^{12}\text{C} \left( 0^+_3 \right) \right\rangle & = \; \: 9.066\% \left| 0\hbar\omega \right\rangle + 90.934\% \left| 2\hbar\omega \right\rangle \; . \end{aligned}$$ Notice that the first dominantly $2\hbar \omega$ state lies at 23.03 MeV excitation; a calculated energy that can be expected to fall with the addition of higher $\hbar \omega$ components. That has not been seen sufficiently at least to the $4\hbar \omega$ level with an *ab initio* shell model [@Na00]. So while the $0^+_3$ state may be the one that is super-deformed, a convergence in energy will require a greatly increased space. But as the $0^+_2$ state is not very important in the formation of resonances and bound states [@Am03], use of our shell model should suffice for the comparisons we make. This may have more bearing on future results as we move to use a microscopic MCAS in which the matrices of interaction potentials will be formed using nucleon-based structure. Next we consider how each state in the low excitation spectrum of $^{13}$C maps onto a single neutron added to any of the three selected states of $^{12}$C. The relevant one-body spectroscopic amplitudes for $I^\pi \to J^{\pi'}$, $$S_{j \frac{1}{2}} = \frac{1}{\sqrt{2(2J+1)}} \left\langle \left( {}^{13}\text{C} \right) J^{\pi'} \left|\left|\left| a^{\dag}_{j,\frac{1}{2}} \right|\right|\right| \left( {}^{12}\text{C} \right) I^{\pi}_i \right\rangle \; ,$$ are listed in Table \[C13-C12\]. ----------------- -------- -------------------- ----------- -------------------- ----------- -------------------- ----------- $\frac{1}{2}^-$ (g.s.) $0p_{\frac{1}{2}}$ $-$0.7285 $0p_{\frac{3}{2}}$ $-$1.0040 $0p_{\frac{1}{2}}$ $-$0.4738 $\frac{1}{2}^+$ (3.09) $1s_{\frac{1}{2}}$ $-$0.9088 $0d_{\frac{5}{2}}$ $-$0.3162 $1s_{\frac{1}{2}}$ $-$0.0605 $\frac{3}{2}^-$ (3.68) $0p_{\frac{3}{2}}$ 0.4504 $0p_{\frac{3}{2}}$ $-$1.0040 $0p_{\frac{3}{2}}$ $-$0.3284 $0f_{\frac{5}{2}}$ $-$0.8342 $\frac{5}{2}^+$ (3.85) $0d_{\frac{5}{2}}$ 0.8129 $0d_{\frac{5}{2}}$ 0.4799 $0d_{\frac{5}{2}}$ 0.0096 $0d_{\frac{3}{2}}$ $-$0.1361 $1s_{\frac{1}{2}}$ 0.0840 $\frac{5}{2}^+$ (6.86) $0d_{\frac{5}{2}}$ $-$0.2147 $0d_{\frac{5}{2}}$ 0.5372 $0d_{\frac{5}{2}}$ $-$0.0102 $0d_{\frac{3}{2}}$ $-$0.0907 $1s_{\frac{1}{2}}$ $-$0.7714 $\frac{5}{2}^+$ (8.88) $0d_{\frac{5}{2}}$ $-$0.0349 $0d_{\frac{5}{2}}$ $-$0.2568 $0d_{\frac{5}{2}}$ 0.2829 $0d_{\frac{3}{2}}$ $-$0.2694 $1s_{\frac{1}{2}}$ $-$0.2391 ----------------- -------- -------------------- ----------- -------------------- ----------- -------------------- ----------- : \[C13-C12\] Dominant components of shell model spectroscopic amplitudes. Energies in brackets are in MeV. The shell model calculations gave more values for addition of that neutron in higher shell states, but those spectroscopic amplitudes (not listed) all have magnitudes less than 0.1. Results of calculations of the neutron-$^{12}$C system reported previously [@Am03], used the parameter values that are specified in Table \[OMparams\]. ------------------------------------------------------------------------------------------- $V_0(\pi)$ $V_{\ell \ell}(\pi)$ $V_{\ell s}(\pi)$ $V_{Is}(\pi)$ ----------- ----------------- ---------------------- ------------------- --------------- -- $\pi = -$ $-$49.144 4.559 7.384 $-$4.770 $\pi = +$ $-$47.563 0.610 9.176 $-$0.052 Geometry $R_0 = 3.09$ fm $a = 0.65$ fm $\beta_2 = -0.52$ ------------------------------------------------------------------------------------------- : \[OMparams\] Parameter values of the base potential (in MeV). In Fig. \[n-12C-fig\], the results are compared with data, both elastic scattering cross sections as well as the spectrum of $^{13}$C. Therein it is clear that the three states of $^{12}$C suffice to deal with information to $\sim 10$ MeV excitation in the compound with corroboration in the scattering of up to 5 MeV. Spin-parity assignments, bound state energies and resonance centroids, widths of the resonances, and the background scattering all are very well matched by the calculations. -------------------------------------------------------------------------------------------------------------- $J^\pi$ With Pauli No Pauli n+${}^{12}$C -------------------------------------------------------------- ------------ ---------- -------------------- -- $\frac{1}{2}^+$ - $-$26.57 $0s_{1/2} + 0^+_1$ $\frac{3}{2}^+, \frac{5}{2}^+$ - $-$22.13 $0s_{1/2} + 2^+_1$ $\frac{1}{2}^+$ - $-$18.91 $0s_{1/2} + 0^+_2$ $\frac{3}{2}^-$ - $-$8.849 $0p_{3/2} + 0^+_1$ $\frac{1}{2}^-$ $-$4.685 $-$4.685 $0p_{1/2} + 0^+_1$ $\frac{1}{2}^-, \frac{3}{2}^-, \frac{5}{2}^-, \frac{7}{2}^-$ - $-$4.410 $0p_{3/2} + 2^+_1$ $\frac{3}{2}^-$ - $-$1.195 $0p_{1/2} + 0^+_2$ $\frac{1}{2}^+$ $-$0.837 $-$0.837 $1s_{1/2} + 0^+_1$ $\frac{3}{2}^-, \frac{5}{2}^-$ $-$0.246 $-$0.246 $0p_{1/2} + 2^+_1$ $\frac{5}{2}^+$ $-$0.171 $-$0.171 $0d_{5/2} + 0^+_1$ $\frac{1}{2}^-$ 2.969 2.969 $0p_{1/2} + 0^+_2$ $\frac{3}{2}^+, \frac{5}{2}^+$ 3.601 3.601 $1s_{1/2} + 2^+_1$ $\frac{1}{2}^+, \frac{3}{2}^+, \frac{5}{2}^+, \frac{7}{2}^+, 4.267 4.267 $0d_{5/2} + 2^+_1$ \frac{9}{2}^+$ -------------------------------------------------------------------------------------------------------------- : \[Zero-b+Vss\] Pauli effects on sub-threshold and bound states in the continuum in the limit $\beta_2 \to 0$ (with $V_{ss} = 0$). A most interesting feature is what occurs as the coupling tends to zero. In that limit, all of the compound resonances shrink to be bound states in the continuum. In this limit, calculations were made with the spin-spin interaction strengths set to zero, and so offsetting a splitting that is most evident with the two odd parity states built from coupling with a $0p_{\frac{1}{2}}$ neutron. The results of these limit calculations are collected in Table \[Zero-b+Vss\]. Therein the states are listed in the order from most bound to largest continuum energy whether they are real or spurious. The energies listed in columns 2 and 3 respectively were found in the zero deformation limit with and without the OPP treatment of Pauli blocking. In the last column we display what dominant character (neutron orbit coupled to state in $^{12}$C) is found for each state in $^{13}$C. Disregarding the Pauli principle clearly gives many spurious states. However, notice that there are matching entries for every resonance state whether the Pauli principle is taken into account or not. That has led to the erroneous belief that a simple adjustment of parameter values is all that is needed to define scattering cross sections and that the Pauli principle effects are unimportant for scattering. Not only is that phenomenology not guaranteed to work in other cases but also the mixing of components caused by finite deformation is quite different when the Pauli principle is or is not satisfied. A calculation made ignoring the Pauli principle gives an incorrect description of all states. The resonance centroids tend to three limits. The highest is at 4.267 MeV with five entries from $\frac{1}{2}^+$ to $\frac{9}{2}^+$ as is formed by attaching a $0d_{5/2}$ neutron to the $2^+$ state in $^{12}$C. The second is at 3.601 MeV having two entries which equate to a $1s_{1/2}$ neutron coupled to the $2^+$ state of the target. The third, the only odd parity resonance ascertained from these calculations within the range of energies to 5 MeV, is identified as a $\frac{1}{2}^-$ resonance. It lies 7.65 MeV above the calculated value for the $^{13}$C ground state and can then be associated with binding a $0p_{1/2}$ neutron to the second $0^+$ state of $^{12}$C. The bound states are less clear with regard to dominant particle coupling character. From shell model calculations, the $\frac{1}{2}^-$ (ground state) and the $\frac{3}{2}^-$ state are sizable mixtures of $p$-shell nucleon coupling to both the ground and $2^+$ states in $^{12}$C. But the $\frac{1}{2}^+$ and $\frac{5}{2}^+$ bound states in $^{13}$C are dominantly formed respectively by a $1s_{1/2}$ and a $0d_{5/2}$ neutron coupled to the ground state of $^{12}$C. The energies found in the zero coupling limit support the inferences made above. Notably, the bound $\frac{5}{2}^+$ tends to $-0.171$ MeV in that limit, as $4.267 - (-0.171) = 4.438$ MeV; the excitation energy of the $2^+$ state in ${}^{12}$C. Likewise the doublet $\frac{3}{2}^+, \frac{5}{2}^+$ tends to $3.601$ MeV and as the bound state $\frac{1}{2}^+$ tends to $-0.837$ MeV in the zero coupling limit, $3.601 - (-0.837) = 4.438$ MeV, the excitation energy of the $2^+$ state in $^{12}$C once more. Finally the $\frac{3}{2}^-$ and $\frac{5}{2}^-$ states both are bound by $-0.246$ MeV and so the energy gap of that pair from the $\frac{1}{2}^-$ state ($-4.685$ MeV), is 4.439 MeV. But deformation makes significant admixing of these nucleon plus core nucleus elements. Indeed it is this spurious mixing that is the most serious concern about calculations that do not take the Pauli principle into account. To reveal that, we have repeated the limit calculations excising the OPP effects and thereby solving the problem in a way equivalent to coordinate-space solutions of such coupled channels equations. That calculation clearly has given spurious states. Summarizing, the MCAS approach has been used to evaluate (low-energy) n-$^{12}$C elastic scattering and to characterize sub-threshold states of $^{13}$C. A collective model prescription with the three lowest states in the $^{12}$C spectrum was used. The results well match observed data but only when allowance for the influence of the Pauli principle was made. Without such allowance, many spurious states result. Most strikingly, the ground state of $^{13}$C then has the wrong spin-parity and a binding far in excess of the known value. But more disturbing is that when states may be matched (in energy and spin-parity) their underlying nucleon plus $^{12}$C compositions are wrong. By tracking results to the $\beta_2 \to 0$ limit, the dominant parentage of each sub-threshold and resonant state in this system has been identified.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Standard photographs of streamer discharges show a two-dimensional projection. We here present stereo-photographic images that resolve their three-dimensional structure. We describe the stereoscopic set-up and evaluation, and we present results for positive streamer discharges in air at 0.2 - 1 bar in a point plane geometry with a gap distance of 14 cm and a voltage pulse of 47 kV. In this case an approximately Gaussian distribution of branching angles of $43\textdegree\pm12\textdegree$ is found; these angles do not significantly depend on the distance from the needle or on the gas pressure.' author: - 'S Nijdam[^1], J S Moerman, T M P Briels, E M van Veldhuizen' - U Ebert title: 'Stereo-photography of point-plane streamers in air' --- A streamer is a rapidly extending discharge channel that can appear when a high voltage is applied to any ionizable medium; most studies are done in air. Streamers precede phenomena like sparks, leaders and lightning. The main difference is that streamers do not significantly increase the gas temperature; they are rather governed by impact ionization and space charge effects [@Ebert2006]. Streamers are directly observed in nature in the form of sprites [@Pasko2007], that are enormous atmospheric discharges above active thunderstorms at about 40 - 90 km altitude. Streamers also have many technical applications, in ozone generation and consecutive disinfection, in bio fuel processing, plasma assisted combustion and aviation; for a short review with references, we refer to [@Ebert2006]. A largely unexplored issue in streamer research is the breakup of single channels into many. Such branching events are commonly seen in experiments [@Veldhuizen2002; @Briels2006; @Briels2008a]; multiple branching actually determines the gas volume that is crossed by streamers and consecutively chemically activated for plasma processing purposes. However, up to now, only the conditions of the first branching event have been resolved in microscopic models [@Liu2004; @Arrayas2002; @Montijn2006; @Luque2007; @Pancheshnyi2005]. On the other hand, the distribution of branching lengths and angles is an ingredient of models for the complete branching tree on larger scales [@Niemeyer1984; @Akyuz2003; @Pasko2001]. In the present paper, we resolve these lengths and, in particular, the angles in experiments. Imaging of streamer discharges is usually done with conventional or digital cameras [@Pancheshnyi2005; @Briels2006; @Winands2006]. This leads to two-dimensional (2D) representations of what is essentially a three-dimensional (3D) phenomenon. These 2D representations can cause problems of interpretation. For example, it is impossible to see whether an apparent loop or reconnection is really what it seems to be. It is also impossible to get a complete picture of the 3D spatial structure and to measure branching angles. For this purpose, we have implemented a stereo-photography method which makes it possible to image streamer discharges in 3D. In this way, we resolve the imaging ambiguities in the fundamental physical phenomena, help understanding which gas volumes are actually treated by the discharge, and supply experimental data for larger scale models. The stereoscopic technique that we use has been around for a very long time [@Brewster1856; @Faugeras1993] and has been used for a large variety of topics. Some phenomena similar to streamers that have been studied with stereo-photography are sparks [@MacAlpine1999], flames [@Ng2003] and dusty plasmas [@Jr2004]. To generate streamers, we use the experimental set-up that is discussed thoroughly in [@Briels2006], and we use the electric circuit called C-supply in [@Briels2006]. In this set-up a capacitor is charged negatively with a DC power supply. This capacitor is then discharged by means of a spark-gap switch. This results in a positive voltage peak on the needle inside the vacuum vessel. A positive corona discharge then propagates from the needle to the grounded plate. Both needle and plate are highlighted in figure \[fig:Setup\]. In the present measurements, a positive voltage of 47 kV with a rise-time of about 30 ns was applied to the point, 14 cm above the plate. The atmosphere in the vacuum vessel consisted of ambient air at different pressures (200, 565 and 1000 mbar). ![\[fig:Setup\]Overview of the stereoscopic measurement set-up with a schematic drawing of the 2 image paths.](Stereo.pdf){width="9cm"} MacAlpine et al. [@MacAlpine1999] have studied sparks with a camera and a prism. In this study two images were taken using a prism to form an image at a right angle to the directly-observed one. In this way the complete 3D-structure of the spark path can be reconstructed with great accuracy. Similar work was reported by M. Makarov [@Makarov]. However, this method only works well for structures that have very few channels (e.g. the one spark of [@MacAlpine1999]). When there are many channels, it is very difficult to correlate them pairwise from two images taken at an angle of 90$\textdegree$. ![\[fig:Stereo-image\]Stereo image as recorded by camera. Settings: positive voltage on tip, *$U=47\mathrm{kV}$*, $p=200\mathrm{mbar}$, $\alpha=13\textdegree$, $d=14\mathrm{cm}$. The intensity has been scaled so that the structure in the bottom part can be clearly seen. One streamer section has been marked with a white line in both images.](StereoOriginalLines.pdf){width="9cm"} In our case, we want to study streamer discharges that contain many (10-100) streamers. For this purpose a similar method can be used, but with a much smaller angle between the two image paths so that the two images of one streamer can be recognized. To achieve a smaller angle, one camera has been used in combination with two prisms and two flat mirrors as shown in figure \[fig:Setup\]. With this set-up two images (from different viewing angles) are captured on one camera frame; therefore they are temporarily perfectly synchronized. An example of such a camera frame is shown in figure \[fig:Stereo-image\]. ![\[fig:3D-reconstruction\]Orthogonal views of the 3D reconstruction of streamer structure shown in figure \[fig:Stereo-image\]. The section originally marked with the white line is now marked with an arrow in both views.](Figure3.png){width="9cm"} From the two 2D-images the 3D structure of the streamer channels can be reconstructed in the following manner: a straight section of a streamer channel is selected in both images. The end points of these two lines are now translated from 2D (xy) to 3D (xyz). In principle, an exact trigonometric evaluation would supply absolute locations in space. However, as we are only interested in local observables (branching angles and lengths), we have used a simplified approach assuming that the cameras are far from the system and have a very large focal length. Indeed, the distance between camera and streamers is about 1 m, while distances between recently splitted streamer branches never exceed 2 cm. The two images give the 2D coordinates $(x_{l},y_{l})$ and $(x_{r},y_{r})$ of identical streamer parts within the left or right image respectively, where the origins of the respective coordinate systems are chosen in the electrode tip. The depth coordinate $z$ is then approximated as $z=\left(x_{r}-x_{l}\right)/\left(2\cdot\sin\left(\alpha/2\right)\right)$, where $\alpha$ is the full angle between the two optical paths (as indicated in figure \[fig:Setup\], in the present measurements $\alpha=13\textdegree$). The 3D $x$ and $y$-coordinates are calculated as $x=\left(x_{r}+x_{l}\right)/2$ and $y=\left(y_{r}+y_{l}\right)/2$. The error in streamer distances after splitting that results from this simplification is less than 0.2 mm. The dominant error comes from the visual determination of the locations of streamer section end points on the stereoscopic images. In many situations it is difficult to locate the exact point of branching, especially where two streamers are very close to each other. The total error is approximately 1 mm for local observables and 5 mm for absolute locations. The two 2D lines have now been translated into one 3D streamer section. This can be done for all suitable streamer sections in the image. When all these 3D streamer sections are now plotted in 3D-space, we get some insight in the real structure of the streamer discharge. The 3D reconstruction of the example from figure \[fig:Stereo-image\] is shown in figure \[fig:3D-reconstruction\]. Here it can clearly be seen that the streamer section marked with the white lines in figure \[fig:Stereo-image\] is not part of a loop. This information can not be derived from just one of the original 2D images. One of the measurements that can be performed now is measuring branching angles. The measured angles are the inner angles between two 3D streamer sections, represented as vectors. The technique described here also has some limitations, the most important one is that it is not possible to process discharge images that contain more than about 50 streamer channels. ![image](Combination.pdf) Figures \[fig:histograms\]a-c show histograms of the measured branching angles for 200, 565 and 1000 mbar and figure \[fig:histograms\]d combines the results for all pressures into one histogram. As can be seen, the distribution is roughly Gaussian, with average values between 39$\textdegree$ and 46$\textdegree$ and standard deviations of 11$\textdegree$ to 13$\textdegree$. The average branching angle shows a slight decrease as a function of pressure. However, it is not clear whether this is statistically significant due to the limited amount of data points (about 35 points per pressure setting). ![\[fig:anglevsposition\]Measured branching angle as function of $p\cdot d$.](pd.pdf) The length scales of streamers are expected and observed to scale quite well with pressure. However, density fluctuations do not scale with density [@Ebert2006; @Briels2008]; if they play a significant role in streamer branching, one would expect the branching distribution to depend on pressure. Therefore, in figure \[fig:anglevsposition\] the branching angle is plotted as function of $p\cdot d$ where $p$ is the pressure and $d$ the vertical distance from the tip (the $y$-coordinate) at the point of branching. If the branching behaviour would differ for streamer sections close to the tip from sections close to the cathode plane, this would be visible in this plot. Also a pressure dependence would be visible. However, only a small dependence on $p\cdot d$ can be observed. This dependence is statistically not significant given the large spread and measurement error in the data set (correlation coefficient $R^{2}=0.15$). The ratio of streamer length between branching events over streamer width has also been measured. This ratio is about 15 for all pressures. This is a bit higher than the ratio of 12 found by Briels *et al.* [@Briels2008]. In conclusion, we have built a stereographic set-up that is able to reconstruct 3D spatial structures of streamer discharges. This enables us to get more insight into what really happens in such a discharge. For example, we are now able to see if something that looks like a streamer reconnecting to another streamer is indeed what it seems. Up to now, such statements relied on multiple observations from 2D images [@Briels2006]. We are also able to measure branching angles of streamers. We have found that the branching angle for streamers in an overvolted gap of 16 cm does not significantly depend on pressure and $p\cdot d$ and is distributed normally with an average of 43$\textdegree$ and a standard deviation of 12$\textdegree$. S.N. and T.B. acknowledge support by STW-project 06501, part of the Netherlands Organisation for Scientific Research NWO. [10]{} U Ebert, C Montijn, T M P Briels, W Hundsdorfer, B Meulenbroek, A Rocco, E M van Veldhuizen [*Plasma Source Sci Technol* ]{}**15** s118 (2006) V P Pasko [*Plasma Source Sci Technol* ]{}**16** s13 (2007) E M van Veldhuizen, W R Rutgers [*J Phys D Appl Phys*]{} **35** 2169 (2002) T M P Briels, J Kos, E M van Veldhuizen, U Ebert [*J Phys D Appl Phys*]{} **39** 5201–5210 (2006) T M P Briels, E M van Veldhuizen, U Ebert [\[]{}unpublished\] Arrayás M, Ebert U, Hundsdorfer W [*Phys Rev Lett*]{} **88** 174502 (2002) N Liu, V P Pasko [*J Geophys Res*]{} **109** 1 (2004) S Pancheshnyi [*Plasma Sources Sci T*]{} **14** 645 (2005) C Montijn, U Ebert, W Hundsdorfer [*Phys Rev E*]{} **73** 065401 (2006) A Luque, U Ebert, C Montijn, W Hundsdorfer [*Appl Phys Lett*]{} **90** 081501 (2007) L Niemeyer, L Pietronero, H J Wiesmann [*Phys Rev Lett*]{} **52** 1033 (1984) M Akyuz, A Larsson, V Cooray, G Strandberg [*J Electrostat*]{} **59** 115 (2003) V P Pasko, U S Inan, T F Bell [*Geophys Res Lett*]{} **28** 3821 (2001) G J J Winands, Z Liu, A J M Pemen, E J M van Heesch, K Yan, E M van Veldhuizen [*J Phys D Appl Phys*]{} **39** 3010–3017 (2006) D Brewster [*The stereoscope: its history, theory and construction*]{} (John Murray, London 1856) O D Faugeras [*Three-Dimensional Computer Vision: A Geometric Viewpoint*]{} (MIT Press, Cambridge, Mass., 1993) J M K MacAlpine, D H Qiu, Z Y Li [*IEEE T Dielect El In*]{} **6** 331–336 (1999) W B Ng, Y Zhang [*Exp Fluids*]{} **34** 484–493 (2003) E Thomas Jr, J D Williams, J Silver [*Phys Plasmas*]{} **11** L37 (2004) M Makarov, Renault Technocentre, Bryancourt, France [\[]{}unpublished\] T M P Briels, E M van Veldhuizen, U Ebert [\[]{}unpublished\] [^1]: Corresponding e-mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a scheme in which an ion trap quantum computer can be used to make arbitrarily accurate measurements of the quadrature phase variables for the collective vibrational motion of the ion. The electronic states of the ion become the ‘apparatus’, and the method is based on regarding the ‘apparatus’ as a quantum computer register which can be prepared in appropriate states by running a Fourier transform algorithm on the data stored within it. The resolution of the measurement rises exponentially with the number of ions used.' address: 'Department of Physics, The University of Queensland, QLD 4072, Australia' author: - 'C. D’Helon and G.J.Milburn' title: Quantum measurements with a quantum computer --- Quantum computing offers promise for greatly enhanced efficiency of implementation for difficult computational problems[@Bennett95; @DiVincenzo95; @Ekert96]. The most important example to date is the quantum factoring algorithm of Shor[@Shor94]. Recent advances in quantum error correction[@Calderbank96; @Steane96; @Laflamme96] and fault tolerant computation[@Shor96; @DiVinShor96] indicate that there is nothing in principle which will prevent the practical realisation of a quantum computer. The most promising suggestion currently is the ion trap realisation of Cirac and Zoller[@CirZol95]. In this scheme, the internal electronic states of a string of ions becomes entangled with the lowest centre-of-mass vibrational mode of the ion trap. An experimental demonstration of a fundamental quantum logic gate in a trapped ion realisation was provided by Monroe et al[@Monroe96]. In this paper we propose that an ion trap quantum computer be viewed as a means of making accurate measurements on the centre-of-mass mode of the ions. The electronic states of the ions forms an ‘apparatus’ coupled to the ‘system’, the vibrational mode. However using the kind of discrete unitary transformations which realise a quantum computing circuit, we can prepare the apparatus in a variety of states to facilitate new quantum limited measurements on the system. This is in the spirit of the suggestion of Wineland and coworkers[@Wineland94] that highly entangled states may be used to improve frequency measurements. In recent experiments on trapped ions[@Wineland96] it was possible to subject the vibrational motion of the ion to a variety of unitary transformations by using sequences of Raman pulses. By selecting the Raman detuning carefully, a variety of oscillator states may be excited. In some case the unitary transformation of the vibrational motion is conditioned on the electronic state of the ion, and may be different for the ground and excited states. This conditional dependence was recently used to generate a Schrödinger cat state for the vibrational mode. In other cases the unitary transformation of the oscillator is independent of the electronic state. Consider a single ion laser cooled to the ground state of a ion trap in the Lamb-Dicke limit. Assume that the ion is illuminated by laser fields which are well detuned from resonance. Let the interaction Hamiltonian coupling the vibrational and electronic states of a single ion be given by $$\hat{H_I}^{(i)}=\hbar(\alpha_i a^\dagger +\alpha_i^*a)|e\rangle_i\langle e|$$ where $\alpha_i$ is proportional to a classical laser field acting on the ith-ion in a dispersive regime, $a,a^\dagger$ are the creation and annihilation operators for the vibrational mode of the ion, and $|e\rangle_i$ is the excited electronic state of the ith ion If we now generalise to the case of N ions in a linear trap, the electronic state of the system may be described by a binary string $S=(S_i=1,0; i=1,N)$ , assuming some ordering of the ions, and where a $1,0$ represent the excited state and ground state respectively. Alternatively the state may be described by an integer $k$ for which the binary string is the binary code $k=\sum_{i=1}^NS_i 2^i$. There are $2^N$ possible states for the electronic system so that $k=0,1,\ldots, K=2^N-1$. Assume now that the coupling strengths $\alpha_i$ are adjusted from ion to ion so that $\alpha_i=\alpha 2^i$. The coupling Hamiltonian between the electronic states and the lowest collective vibrational mode of the ions may then be written $$\hat{H_I}=\hbar(\alpha a^\dagger +\alpha^* a)\hat{\Upsilon}$$ where $$\hat{\Upsilon}=\sum_{k=0}^K k|k\rangle\langle k|$$ The phase of $\alpha$ can be adjusted, so for simplicity let us take it to be real. The unitary operator describing the coupling of the vibrational and electronic states at the end of a sequence of pulses is then given by $$U=e^{ir\hat{x}\hat{\Upsilon}}$$ where $\hat{x}=a+a^\dagger$ is the operator describing the in-phase component of the collective vibrational amplitude in an interaction picture rotating at the ion trap frequency, and $r$ is a real parameter. It is clear that the interaction can realise a measurement of the vibrational quantity $\hat{x}$. In this interpretation we regard the N ion register as an ‘apparatus’ to measure the ‘system’, the vibrational mode. However if all the ions are initially in the ground state, they will remain in the ground state under this unitary interaction as $\hat{\Upsilon}$ commutes with the projection operators for each electronic energy eigenstate. In order for the interaction to change the state of the ions we must first pre-process the electronic register to a state which can be displaced by the unitary interaction. The required states $|\bar{k}\rangle$ are eigenstates of an operator $\hat{\Phi}$ which is canonically conjugate to $\hat{\Upsilon}$. That is to say $\hat{\Upsilon}$ must act as a pure differential operator in the $\{|\bar{k}\rangle\ k=0,\ldots,K\}$ basis. This basis is simply a discrete Fourier transform of the original electronic basis $\{|k\rangle\; k=0,\ldots,K\}$. Thus $$|\bar{l}\rangle=\frac{1}{\sqrt{K}}\sum_{k=0}^K\exp\left (\frac{2\pi ikl}{K+1}\right) |k\rangle \label{FT}$$ It is at this point that we remember the quantum computer interpretation of this system. The transformation of Eq (\[FT\]) may be realised by running a Fourier transform algorithm on the electronic register ions[@Beckman96; @Dhelon96]. Let the initial state of the system be $$|\Psi\rangle =|\psi\rangle_v\otimes|0\rangle$$ where $|\psi\rangle_v$ is an arbitrary vibrational state and $|0\rangle$ indicates all the ions are in the ground state. The [*first step*]{} is to apply a sequence of $\pi/2-$pulses to the ions to place them in a symmetric superposition of the ground and excited states. The resulting electronic state is precisely the state $|\bar{0}\rangle$, the 0-state in the Fourier transform basis. This state is equivalently a uniform superposition over all possible electronic energy eigenstates. In the [*second step*]{}, the unitary interaction is then implemented to couple the vibrational and electronic states. In the [*third step*]{} an inverse Fourier transform is run on the electronic register. At the end of these three steps the state of the system is $$|\Psi^\prime\rangle=\frac{1}{K+1}\sum_{k,l=0}^K\int_{-\infty}^\infty dp\phi(p) e^{-\frac{2\pi i kl}{K+1}}|p+rk\rangle_v\otimes|l\rangle$$ where $\phi(p)$ is the momentum probability amplitude for the initial vibrational state. In the final step we readout the state of the electronic register. This is very much like the readout of the output register in the Shor algorithm. For trapped ions this can be done with very high quantum efficiency using quantum jump techniques[@Wineland96]. The result of this readout is a binary string describing which ions are in the ground state and which are in the excited state. Equivalently the result is the integer $l$ encoded by this binary string. The (unnormalised) conditional vibrational state of the system after a readout result, $l$, is $$|\tilde{\psi}\rangle_v=\frac{1}{K+1}\sum_{k=0}^K\int_{-\infty}^\infty dp \phi(p) e^{-\frac{2\pi i kl}{K+1}}|p+rk\rangle_v$$ The probability for this result is found to be $$P(l)=\frac{\sqrt{2\pi}}{(K+1)^2}\sum_{k,k^\prime=0}^Ke^{-\frac{2\pi i (k-k^\prime)l}{K+1}}\chi(r(k-k^\prime))$$ where $$\chi(k)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty dx e^{ikx} P(x)$$ with $P(x)$ being the initial distribution of the in-phase quadrature variable $\hat{x}$, of the vibrational state. Clearly $\chi(k)$ is just the characteristic function for the ‘position’ distribution for the initial vibrational state. In what follows we will assume that $r=1$. The probability distribution for the output $l$ is thus seen to a kind of discrete approximation to the probability distribution $P(x)$. In the case of a minimum uncertainty Gaussian state with position variance $\Delta$, and zero mean, the readout distribution takes the form $$P(l) =\frac{1}{K+1}\left (1+\sum_{m=1}^K(K+1-m)\cos\left(\frac{2\pi ml}{K+1}\right )e^{-m^2\Delta/2}\right ).$$ In the limit $\Delta\rightarrow\infty$ we have an effective uniform position distribution. In this limit it is easy to see that $P(l)=1$ for all $l$. In the opposite limit $\Delta\rightarrow 0$ we have a very well defined position at $x=0$, and $P(l)=\delta_{l,0}$. In general the sum can be truncated at $m_{trunc}=\frac{2\sqrt{2}}{\Delta}$ when subsequent terms contribute less than $e^{-4}$, which we refer to as the tolerance $\epsilon$. For a large number of ions, the truncation becomes a good approximation. In figure \[fig1\] we illustrate the intermediate behaviour for $N= 9$ ions and various values of $\Delta$,(the tolerance used is $\epsilon=0.01$. In these plots we have reflected the values at $l\ >\ N/2$ to negative values so as to ensure the plot is symmetric around $0$. We see that as the number of ions increases we get a better and better approximation to the true ‘position’ distribution $P(x)$. The dimensionless position $x$ is related to the index $l$ by $x=\frac{2\pi}{2^N}l$, thus the corresponding position distribution $P^\prime(x)$ obtained from $P(l)$ is defined over the domain $[0,2\pi)$. Thus for a fixed number of ions $N$, the range of the uncertainty $\Delta$ in the initial position distribution $P(x)$ is restricted, if the measured position distribution $P^\prime(x)$ is to be essentially the same as $P(x)$. The lower limit imposed on $\Delta$ is determined by the number of ions in the register, whereas the upper limit on $\Delta$ is independent of the size of the register and is given by $\Delta_{max}\approx10$. As the uncertainty approaches this upper limit, the measured electronic distribution $P(l)$ becomes flat, so that it does not provide any more information about the actual position distribution $P(x)$. However, we are assuming that the collective motion of the ions is in the Lamb-Dicke regime, and hence we require [*a priori*]{} that the uncertainty in the position is bounded by $\Delta x \ll \sqrt{N}/\eta$, where $\eta$ is the Lamb-Dicke parameter of the centre-of-mass vibrational mode. Alternatively, if we are given a lower limit for values of $\Delta$, then we can determine the minimum number of ions $N_{min}$ required to be able to measure the initial position distribution $P(x)$ accurately, and this is given by $$N_{min}\approx 8.14-\frac{1}{2}\log_2 \Delta$$ for a tolerance of $\epsilon =0.01$ in the summation series i.e., this corresponds to $N_{min}=25$ for $\Delta=10^{-10}$, and $N_{min}=10$ ions for $\Delta=0.1$. The graphs plotted in Figures \[fig2\](a)-(b) demonstrate that the value of the uncertainty $\Delta x$ in the measured position distribution $P^\prime(x)$ settles to the value of the initial uncertainty $\Delta$ as the number of ions increases past the appropriate minimum number $N_{min}$, thus indicating that the measured position distribution $P^\prime(x)$ is approximately equal to $P(x)$. We have illustrated a way in which an ion trap quantum computer can be used to measure the distribution of the in-phase and out-of-phase quadrature variables for the collective vibrational motion of the ion. The method is based on regarding the ions as a quantum computer register which can be prepared in appropriate states by running a Fourier transform algorithm on the data stored in the ionic register. The first point to note is that we are not restricted to any particular quadrature phase variable. As the phase of the laser pulses coupling the system and apparatus may be varied we can in principle measure any rotated quadrature phase. Given a sufficient set of such distributions a quantum tomographic data inversion could be done to reconstruct the initial vibrational state. The second point to note is that as the number of ions increases the states available to the computer rises exponentially, providing an increasingly accurate readout of a system variable with a continuous spectrum. This generalises previous schemes to readout vibrational states using only a single ion; a two state apparatus. In that case many repetitions of the measurement must be preformed to get an average over particular vibrational variables. In our scheme we can do better (albeit at the considerable expense of running a quantum algorithm), as we have an apparatus with many more states, and thus one which is better adapted to a vibrational degree of freedom. Finally we note that the scheme is capable of considerable generalisation. For example by choosing different couplings between the ion and vibrational modes we can get access to many other vibrational variables. For example if we choose a sequence of Raman pulses which provided a ‘squeezing’ interaction for the vibrational motion, we can measure the distribution of eigenstates of the squeeze operator[@Chen94]. Furthermore, there may be other quantum algorithms that realise different measurement schemes. GJM would like to thank The Institute for Theoretical Physics, University of California, Santa Barbara, for support during a visit where this work was completed. C.H.Bennett, Physics Today,[**48**]{} 24,(1995). D.P. DiVincenzo, Science, [**269**]{}, 255 (1995). A.Ekert and R. Jozsa, Rev. Mod. Phys. [**68**]{},733 (1996). P. Shor, Proc. 35th. Ann. Symp. on Foundations of Computer Science, IEEE Press, Los Alamitos, 124, (1994). A.R.Calderbank and P. Shor, Phys. Rev A [ **54**]{}, 1098, (1996). A.M.Steane, Phys. Rev. Letts. [**77**]{}, 793, (1996). R. Laflamme, C. Miquel, J.-P. Paz and W.H. Zurek, Phys. Rev. Letts, [**77**]{}, 198, (1996). P.W. Shor in Proc 37th Symp. on Foundations of Computer Science, to appear. D.P. DiVincenzo and P.W. Shor, Phys. Rev. Lett. [**77**]{}, 3260, (1996). J.I. Cirac and P. Zoller, Phys. Rev. Lett. [**74**]{}, 4714, (1996). C. Monroe, D. Meekhof, B.E King, W.M. Itano, and D. J. Wineland, Phys. Rev. Lett, [**75**]{}, 4714 (1995). D.J.Wineland, J.J. Bollinger, W.M. Itano, D.J. Heinzen, Phys. Rev A [**50**]{} 67, (1994). D.Meekhof, C.Monroe, B.E.King, W.M.Itano and D.J.Wineland, Phys. Rev. Lett. [**76**]{}, 1796 (1996). D. Beckman, A. Chari, S. Devabhaktuni and J. Preskill, quant-ph/9602016. C. Dhelon and G.J. Milburn, to appear Phys. Rev A (1996). G.J.Milburn, Wenyu Chen and K.R. Jones, Phys. Rev A. [**50**]{}, 801 (1994)
{ "pile_set_name": "ArXiv" }
--- abstract: 'Consistent reconstruction is a method for producing an estimate $\widetilde{x} \in \R^d$ of a signal $x\in \R^d$ if one is given a collection of $N$ noisy linear measurements $q_n = \langle x, \varphi_n \rangle + \epsilon_n$, $1 \leq n \leq N$, that have been corrupted by i.i.d. uniform noise $\{\epsilon_n\}_{n=1}^N$. We prove mean squared error bounds for consistent reconstruction when the measurement vectors $\{\varphi_n\}_{n=1}^N\subset \R^d$ are drawn independently at random from a suitable distribution on the unit-sphere $\mathbb{S}^{d-1}$. Our main results prove that the mean squared error (MSE) for consistent reconstruction is of the optimal order $\mathbb{E}\|x - \widetilde{x}\|^2 \leq K\delta^2/N^2$ under general conditions on the measurement vectors. We also prove refined MSE bounds when the measurement vectors are i.i.d. uniformly distributed on the unit-sphere $\mathbb{S}^{d-1}$ and, in particular, show that in this case the constant $K$ is dominated by $d^3$, the cube of the ambient dimension. The proofs involve an analysis of random polytopes using coverage processes on the sphere.' address: 'Vanderbilt University, Department of Mathematics, Nashville, TN 37240, USA' author: - 'Alexander M. Powell' - 'J. Tyler Whitehouse' date: 'April 9, 2013' title: | Error bounds for consistent reconstruction:\ random polytopes and coverage processes --- Introduction ============ We consider the problem of estimating an unknown signal $x\in \R^d$ from a collection of $N \geq d$ noisy linear measurements $$\label{qn-def} q_n = \langle x, \varphi_n \rangle + \epsilon_n, \ \ \ 1 \leq n \leq N,$$ where $\{\varphi_n\}_{n=1}^N \subset \R^d$ is a known spanning set for $\R^d$, and where the unknown noise $\{\epsilon_n\}_{n=1}^N$ has been independently drawn according to the uniform distribution on a known interval $[-\delta, \delta]$. [*Consistent reconstruction*]{} is a method for producing an estimate $\widetilde{x} \in \R^d$ of $x$ from the noisy measurements . Consistent reconstruction selects $\widetilde{x}$ as any solution to the linear feasibility problem $$\label{cr-def} \forall \ 1 \leq n \leq N, \ \ \ \ | \langle \widetilde{x}, \varphi_n \rangle - q_n | \leq \delta.$$ In other words, simply seeks an estimate $\widetilde{x}$ that is consistent with the knowledge that the noise is bounded in $[-\delta,\delta]$. Our main contribution in this work is to provide sharp bounds on the mean squared estimation error associated with consistent reconstruction and to quantify how accurately recovers $x$ from the measurements as a function of the number of measurements $N$ and the dimension $d$. In our analysis, $\{\varphi_n\}_{n=1}^N\subset \R^d$ will be i.i.d. random vectors drawn from a suitable distribution on the unit-sphere $\mathbb{S}^{d-1}$, and special attention will be given to the case when each $\varphi_n$ is uniformly distributed on $\mathbb{S}^{d-1}$. Consistent reconstruction has received particular attention in the signal processing literature as a method for recovering signals from quantized samples. Deterministic round-off errors that arise in quantization are frequently modeled using uniform noise. For example, the use of uniform noise models in quantization is typically justified by dithering, [@RG], or with high resolution asymptotics as the quantizer step-size approaches zero, [@JWW]. Consistent reconstruction and its variants have been shown to be an effective method for signal recovery in memoryless scalar quantization (MSQ), [@GVT; @VT94b; @RG; @Z01; @Z03], Sigma-Delta ($\Sigma\Delta$) quantization, [@VT94a], compressed sensing, [@JHF], and finite rate of innovation sampling, [@JB06]. A key point often observed in practice is that when compared with linear reconstruction, consistent reconstruction can reduce the mean squared reconstruction error by an extra multiplicative factor that scales inversely with the sampling rate. The structure of uniformly distributed noise plays an essential role in our analysis of consistent reconstruction . It is useful to note that estimation with uniform noise in falls outside of several classical approaches to estimation theory. For example, asymptotic normality theorems in maximum likelihood estimation (MLE) typically require suitable smoothness assumptions on the underlying noise distribution which uniform noise does not satisfy, e.g., [@F]. Similarly, the multiparameter Cramer-Rao bound gives lower bounds for minimum variance unbiased estimation, but also requires suitable regularity on the noise distribution. Finally, recall that linear estimation commonly yields mean squared error (MSE) bounds of order $1/N$, and this is, for example, optimal for Gaussian noise with respect to the Cramer-Rao bound. However, when dealing with uniform noise, linear estimation is typically sub-optimal and it is possible to provide more accurate recovery than MSE of order $1/N$. We provide a rigorous analysis of the mean squared error in consistent reconstruction for the general estimation problem with suitable random measurement vectors $\{\varphi_n\}_{n=1}^N \subset \mathbb{R}^d$, and we prove that the mean squared error is of the optimal order $K/N^2$ with precise control on the constant $K$. A sample consequence of this is that $N = \mathcal{O}(d^{3/2})$ random measurements will suffice to achieve highly accurate mean squared reconstruction error (compared to $\mathcal{O}(d^2)$ measurements with linear reconstruction). Consistent reconstruction has a simple geometrical interpretation. Since the bounded noise satisfies $|\epsilon_n| \leq \delta$, each noisy measurement $q_n$ gives the information that the unknown true signal $x\in\R^d$ lies in the $2\delta$-thick slab $$\label{Sn-def} S_n = S_n(q_n, \varphi_n)= \{ u \in \R^d : | \langle u, \varphi_n \rangle - q_n | \leq \delta \}.$$ Consequently, the consistency equations are equivalent to requiring that $\widetilde{x}\in\R^d$ lies in the [*consistent reconstruction polytope*]{} defined by $$\label{QN-def} Q_N = \bigcap_{n=1}^N S_n.$$ The assumption that $\{\varphi_n\}_{n=1}^N$ spans $\R^d$ ensures that $Q_N$ is a compact set. Moreover, $Q_N$ is almost surely a nondegenerate polytope with nonempty interior. Since the true signal $x$ is always contained in the polytope $Q_N$, it is clear that the system is feasible. The main object of interest in this article will be the worst case error associated with consistent reconstruction. Recalling that generally has infinitely many solutions, the worst case error may be defined as follows. If $\widetilde{x} \in Q_N$ is any solution to the consistent reconstruction system , then the error $(x - \widetilde{x})$ lies in the following [*error polytope*]{} $$\label{PN-def} P_N = \bigcap_{n=1}^N E_n$$ where $$\label{En-def} E_n = \{ u \in \R^d : | \langle u, \varphi_n \rangle - \epsilon_n | \leq \delta \}.$$ The polytope $P_N$ is obtained by translating $Q_N$ to the origin by the vector $x$. In terms of the polytopes $P_N$ and $Q_N$, the worst case error associated with consistent reconstruction can be defined as $$\label{WN-def} W_N = \max \{ \| u \| : u \in P_N \} = \max \{ \| u - x \| : u \in Q_N \}.$$ Viewed geometrically, the worst case error $W_N$ is precisely the radius of the smallest closed ball centered at 0 that contains the error polytope $P_N$. Overview and main results {#overview-and-main-results .unnumbered} ------------------------- The main contribution of this article is to prove that the expected worst case error squared for consistent reconstruction is of the optimal order $$\label{overview-eq} \mathbb{E}|W_N|^2 \leq \frac{K\delta^2}{N^2},$$ where the constant $K>0$ depends on the distribution of the random vectors $\{\varphi_n\}_{n=1}^N$. Our first main result, Theorem \[main-thm1\], proves the mean squared error bound under general assumptions on the i.i.d. unit-norm random vectors $\{\varphi_n\}_{n=1}^N \subset \mathbb{S}^{d-1}$. Our second main result, Theorem \[mainthm2\], proves a refined version of when $\{\varphi_n\}_{n=1}^N \subset \mathbb{S}^{d-1}$ are i.i.d. random vectors that are uniformly distributed on $\mathbb{S}^{d-1}$ and, in particular, shows that the constant $K$ in is dominated by $d^3$, the cube of the ambient dimension. The paper is organized as follows. Section \[background:sec\] provides background on estimation with uniform noise, and also discusses preliminaries and notation such as surface measure on the sphere and geodesic $\epsilon$-nets. Section \[unidirection-sec\] analyzes the size of the error polytope $P_N$ in a fixed direction. Section \[cover-sec\] provides necessary background and results on coverage processes on the sphere which will be used in the proofs of our main theorems. Section \[genthm-sec\] states and proves our first main theorem, Theorem \[main-thm1\], which shows that consistent reconstruction satisfies under general assumptions. Section \[unif-sec\] states and proves our second main theorem, Theorem \[mainthm2\], which shows that for random vectors that are uniformly distributed on the unit-sphere consistent reconstruction satisfies with a constant $K$ that is dominated by $d^3$. Background and notation {#background:sec} ======================= Estimation with uniform noise and consistent reconstruction {#est-background:sec} ----------------------------------------------------------- In this section we briefly recall some background on estimation with uniform noise and consistent reconstruction. This will help provide perspective on our main results. Begin by recalling linear reconstruction. Let $N \geq d$. Suppose that $\{\varphi_n\}_{n=1}^N \subset \R^d$ spans $\R^d$ and that $\{f_n\}_{n=1}^N \subset \R^d$ is any dual frame satisfying $$\forall x \in \R^d, \ \ \ x = \sum_{n=1}^N \langle x, \varphi_n \rangle f_n.$$ Suppose for the moment that $\{\epsilon_n\}_{n=1}^N$ are simply independent zero mean random variables with variance $\mathbb{E}|\epsilon_n|^2 = \sigma^2$, and that one wishes to estimate $x\in \R^d$ from the noisy measurements $q_n = \langle x, \varphi_n \rangle + \epsilon_n$, $1 \leq n \leq N$. Using the dual frame $\{f_n\}_{n=1}^N$ to linearly reconstruct $\widetilde{x}\in\R^d$ by $$\widetilde{x} = \sum_{n=1}^N q_n f_n$$ yields the estimation error $$\label{linrec-mse} \forall x \in \R^d, \ \ \ \mathbb{E}\| x- \widetilde{x}\|^2 = \sigma^2 \sum_{n=1}^N \| f_n\|^2.$$ If each $\varphi_n$ is assumed to be unit-norm, $\|\varphi_n\|=1$, then the MSE in is bounded below by $$\label{utf-mse} \forall x \in \R^d, \ \ \ \mathbb{E}\|x-\widetilde{x}\|^2 \geq \frac{d^2 \sigma^2}{N},$$ for example, see [@GKK]. Moreover, if each $\| \varphi_n\| =1$ then equality holds in precisely when $\{\varphi_n\}_{n=1}^N$ is a unit-norm tight frame for $\R^d$ and when $f_n = \frac{d}{N} \varphi_n$ is taken as the associated canonical dual frame, e.g., see [@GKK]. The case of unit-norm tight frames yields the mean squared error $\mathbb{E}\| x - \widetilde{x}\|^2 = {d^2 \sigma^2}/{N}$. It is natural to ask how much one can improve on the standard $1/N$ mean squared accuracy if one uses nonlinear reconstruction. We are specifically interested in the case of uniform noise, and assume henceforth that $\{ \epsilon_n\}_{n=1}^N$ are i.i.d. uniform random variables on $[-\delta, \delta]$. We begin by mentioning a Bayesian lower bound due to Rangan and Goyal, [@RG]. Suppose that the vectors $\{ \varphi_n \}_{n=1}^{\infty} \subset \R^d$ are unit-norm and that $x\in \R^d$ is an absolutely continuous random vector and that $q_n = \langle x, \varphi_n \rangle + \epsilon_n$. Let $\widetilde{x}_N = \widetilde{x}_N( \{q_n\}_{n=1}^N, \{ \varphi_n \}_{n=1}^N)$ be any estimator which maps each observed input $(\{q_n\}_{n=1}^N, \{ \varphi_n \}_{n=1}^N)$ to an estimate $\widetilde{x}_N\in\R^d$ of the signal $x\in \R^d$. It was shown in [@RG] that the mean squared error is bounded below in the following manner: $$\label{rg-lower-bnd} {\rm lim \thinspace inf}_{N \to \infty} \ N^2 \thinspace \mathbb{E} \| x - \widetilde{x}_N\|^2 >0.$$ Unlike the expectation in is taken over both a random signal $x$ and the noise $\{\epsilon_n\}_{n=1}^N$. The lower bound shows that for estimation in the setting of uniform noise one cannot expect MSE that is more accurate than $1/N^2$. Related lower bounds for quantization problems can be found in [@GVT]. In [@RG], Rangan and Goyal proposed an estimation algorithm for that achieves the optimal $1/N^2$ error rate. Their algorithm starts with an arbitrary $\widetilde{x}_0 \in \R^d$, and iteratively produces estimates $\widetilde{x}_n \in \R^d$ with the following soft-thresholding algorithm $$\label{rg-alg} \widetilde{x}_n = \widetilde{x}_{n-1} + \frac{\varphi_n \thinspace T_{\delta} (q_n - \langle \widetilde{x}_{n-1}, \varphi_n \rangle)}{\|\varphi_n\|^2},$$ where the soft-thresholding function $T_\delta:\mathbb{R} \to \mathbb{R}$ is defined by $$\label{soft-thresh} T_{\delta}(t) = \begin{cases} t - \delta, & \hbox{ if } t > \delta,\\ 0, & \hbox{ if } |t| \leq \delta,\\ t + \delta, & \hbox{ if } t < -\delta. \end{cases}$$ The error analysis of the algorithm in [@RG] assumed that $\{ \varphi_n\}_{n=1}^N$ are independent identically distributed versions of a random vector $\varphi$ satisfying the following condition $$\label{rg-exp-cond} \exists \alpha>0, \ \forall x \in \mathbb{S}^{d-1} , \ \ \ \mathbb{E}|\langle x, \varphi \rangle| \geq \alpha>0.$$ It was proven in [@RG] that if the i.i.d. random vectors $\{ \varphi_n \}_{n=1}^N$ satisfy then $$\forall s<1, \ \forall x \in \R^d, \ \ \ \lim_{N\to \infty} N^{2s} \| x - \widetilde{x}_N\|^2 =0, \ \ \ \hbox{ almost surely.}$$ Moreover, the mean squared error was later proven to satisfy $\mathbb{E}\|x- \widetilde{x}_N\|^2 \leq C/N^2$ for a suitable constant $C>0$ in [@AP-RG]. The algorithm need not produce globally consistent estimates but instead employs local updates that can be sensitive to ordering issues. Consistent reconstruction provides a maximum likelihood estimate (MLE) for . Let $q=q(x)$ be the $N \times 1$ random vector that is parametrized by $x$ and whose $n$th entry is $q_n = \langle x, \varphi_n\rangle + \epsilon_n$. The associated likelihood function is given by $\mathcal{L}(x | q) = \chi_{Q_N}(x)$, where $\chi_{Q_N}$ is the indicator function of the consistent reconstruction polytope $Q_N$ in . Thus, the likelihood function $\mathcal{L}(x|q)$ is maximized precisely when $x$ is a consistent estimate satisfying . Moreover, asymptotic normality results from MLE do not apply here since $\mathcal{L}(x|q)$ does not satisfy the smoothness assumptions that are typically needed, [@F]. We conclude this section with the following simple one-dimensional example to provide intuition into the desired $1/N^2$ error rate for consistent reconstruction. \[cr-1dim-ex\] Let $x \in \R$ and $q_n = x + \epsilon_n$, where $\{\epsilon_n \}_{n=1}^N$ are i.i.d. uniform random variables on $[-\delta,\delta]$. Consider the problem of estimating $x$ from the noisy observations $\{q_n\}_{n=1}^N$. In this one-dimensional example, consistent reconstruction simply selects any estimate $\widetilde{x}_N$ that lies in the interval $I_N=[A_N,B_N]$ where $$A_N = \max \{ q_n - \delta : 1 \leq n \leq N\} \ \ \ \ \ \hbox{ and } \ \ \ \ \ B_N = \min \{ q_n + \delta: 1 \leq n \leq N\},$$ and the associated worst case estimation error is given by $$w_N = \max \{ |x - A_N|, |x - B_N| \}.$$ Elementary order statistics computations show that $$\mathbb{E} |x-A_N|^2 = \mathbb{E} |x-B_N|^2 =\frac{8\delta^2}{(N+1)(N+2)}$$ and that the worst error $w_N$ satisfies the following mean squared error bound $$\mathbb{E} |w_N|^2 = \frac{14\delta^2}{(N+1)(N+2)}.$$ An important technical issue for analyzing in $\mathbb{R}^d$ will be that the geometry of the error polytope becomes non-trivial in higher dimensions. Preliminaries and notation -------------------------- In this section we collect some necessary notation and background results concerning measure on the sphere and epsilon-nets. We shall denote an open spherical cap on the unit-sphere $\mathbb{S}^{d-1}\subset \R^d$ with center $\varphi \in \mathbb{S}^{d-1}$ and angular radius $0<\theta<\pi$ by $${\rm Cap}(\varphi, \theta) = \{ u \in \mathbb{S}^{d-1} : \langle u, \varphi \rangle > \cos \theta \}.$$ If $0<\theta<\pi/2$ then the relative measure (normalized with respect to $\mathbb{S}^{d-1}$) of ${\rm Cap}(\varphi, \theta)$ is given by, e.g., [@BCL10], $$\label{rel-cap-meas} r_{d-1}(\theta) = \frac{\Gamma(\frac{d}{2})}{\sqrt{\pi}\Gamma(\frac{d-1}{2})} \int_0^{\theta} (\sin u)^{d-2} d u.$$ It is useful to note, e.g., [@W46], that when $d\geq 2$ the constant $$\label{CdDef} C_d = \frac{\Gamma(\frac{d}{2})}{\sqrt{\pi}\Gamma(\frac{d-1}{2})}$$ satisfies $$\label{C-d-estimate} \sqrt{1 - \frac{1}{d} \thinspace} \thinspace \sqrt{\frac{d-1}{2\pi}} \leq C_d \leq \sqrt{\frac{d-1}{2\pi}} \ \ \ \hbox{ and } \ \ \ \lim_{d\to\infty}\frac{C_d}{\sqrt{d}}=\frac{1}{\sqrt{2\pi}}.$$ Fix any $x_0 \in \mathbb{S}^{d-1}$ and let the random vector $\varphi\in \mathbb{R}^{d-1}$ be uniformly distributed on the unit-sphere $\mathbb{S}^{d-1}$. Note that, by rotation invariance, the distribution of the random variable $Z= |\langle x_0, \varphi \rangle|$ does not depend on $x_0$. The pdf of the random variable $Z$ is given by, e.g., [@Stam82], $$\label{Zpdf} f_Z(z) = \begin{cases} 2C_d (1-z^2)^{\frac{d-3}{2}}, & \hbox{ if } z \in [0,1],\\ 0, & \hbox{ if } z \not\in [0,1]. \end{cases}$$ The geodesic distance between between two points $x,y \in \mathbb{S}^{d-1}$ on the sphere will be denoted by $$d(x,y) = \arccos( \langle x, y \rangle).$$ Given $\epsilon>0$, we say that a set $\mathcal{N}_{\epsilon} \subset \mathbb{S}^{d-1}$ is a geodesic $\epsilon$-net of $\mathbb{S}^{d-1}$ if $$\forall x \in \mathbb{S}^{d-1}, \ \exists z \in \mathcal{N}_{\epsilon}, \ \ \hbox{ such that } \ \ d(x,z) \leq \epsilon.$$ A standard argument shows that if $d\geq 2$ then there exist geodesic $\epsilon$-nets $\mathcal{N}_{\epsilon}$ of $\mathbb{S}^{d-1}$ with cardinality satisfying $$\begin{aligned} \label{eps-net-card-eq} \# \left( \mathcal{N}_{\epsilon} \right) &\leq \frac{1}{r_{d-1}(\epsilon/2)} \leq \frac{\sqrt{\pi} \thinspace \Gamma(\frac{d-1}{2})}{\Gamma(\frac{d}{2})} \left( \frac{\pi}{2}\right)^{d-2} \frac{(d-1)}{(\epsilon/2)^{d-1}} \leq \left( \frac{8}{\epsilon}\right)^{d-1}.\end{aligned}$$ For example, if $\mathcal{N}_{\epsilon}$ is a maximal $\epsilon$-separated (with respect to geodesic distance) subset of $\mathbb{S}^{d-1}$ then $\mathcal{N}_{\epsilon}$ is a geodesic $\epsilon$-net for $\mathbb{S}^{d-1}$ and satisfies . Error polytope size in a fixed direction {#unidirection-sec} ======================================== In this section we study the radial size of the error polytope $P_N$ in a [*fixed*]{} direction. Given a unit-vector $\psi \in \mathbb{S}^{d-1}$, define the radial size of the error polytope in the direction $\psi$ by $$R_N(\psi) = \max \{ r \geq 0 : r\psi \in P_N \}.$$ Note that $W_N \geq R_N$ since the worst case error satisfies $$W_N = \sup \{R_N(\psi): \psi \in \mathbb{S}^{d-1} \}.$$ The next result follows immediately from Example \[cr-1dim-ex\] and provides a simple lower bound on $\mathbb{E}|R_N(\psi)|^2$ for a general choice of $\{\varphi_n\}_{n=1}^N\subset \mathbb{S}^{d-1}$. \[GenRbndProp\] Let $\{\varphi_n \}_{n=1}^N \subset \mathbb{S}^{d-1}$ be arbitrary and let $\psi \in \mathbb{S}^{d-1}$. Then the worst case error in the direction $\psi$ for consistent reconstruction satisfies $$\forall N, \ \ \ \mathbb{E}|R_N(\psi)|^2 \geq \frac{8\delta^2}{(N+1)(N+2)}.$$ Note that $$R_N(\psi) = \min \{ x_n : 1 \leq n \leq N \},$$ where $$x_n = \begin{cases} {(\epsilon_n+\delta)}/{|\langle \varphi_n, \psi \rangle| }, & \hbox{ if } \langle \varphi_n, \psi \rangle \geq 0, \\ {(\delta - \epsilon_n)}/{|\langle \varphi_n, \psi \rangle|}, & \hbox{ if } \langle \varphi_n, \psi \rangle < 0. \\ \end{cases}$$ Since $|\langle \varphi_n, \psi \rangle| \leq 1$ and since $(\epsilon_n+\delta)$ and $(\delta -\epsilon_n)$ are both uniformly distributed on $[0, 2\delta]$ it follows that $x_n \geq \xi_n$, where $\xi_n$ is uniformly distributed on $[0,2\delta]$. The proof now follows from Example \[cr-1dim-ex\]. In the remainder of this section, we study $R_N(\psi)$ in the case when the vectors $\{ \varphi_n \}_{n=1}^N \subset \R^d$ used to define the error polytope $P_N$ are i.i.d. uniformly distributed random vectors on the unit-sphere $\mathbb{S}^{d-1}$. In this case, rotation invariance implies that the distribution of $R_N=R_N(\psi)$ is independent of $\psi$. We will make use of the following lemma whose proof follows from similar steps as in Proposition \[GenRbndProp\]. \[YN-lemma\] Let $\xi$ be a uniform random variable on $[0, 2\delta]$, and define the random variable $Z=|\langle e_0, \varphi \rangle|$, where $e_0 = (1, 0, 0, \cdots, 0) \in \R^d$ and the random vector $\varphi$ is uniformly distributed on unit-sphere $\mathbb{S}^{d-1}$. Let $\xi$ and $Z$ be independent and define the random variable $X= {\xi}/{Z}.$ Let $\{X_n\}_{n=1}^N$ be $N$ independent versions of the random variable $X$ and consider the associated order statistic $$Y_N = \min \{ X_n : 1 \leq n \leq N \}$$ The random variable $R_N$ has the same distribution as the random variable $Y_N$. \[RadialThmDimd\] Let $d \geq 2$. Suppose that $\{\varphi_n\}_{n=1}^N\subset \mathbb{S}^{d-1}$ are i.i.d. uniformly distributed random vectors on $\mathbb{S}^{d-1}$. Then for $N \geq 3$ the worst case error in the direction $\psi \in \mathbb{S}^{d-1}$ for consistent reconstruction satisfies $$\label{RadThmDimdBnd} \mathbb{E}|R_N(\psi)|^2 = \frac{2\delta^2(d-1)^2}{(C_d)^2(N+1)(N+2)} + 2\delta^2\alpha_{d,N},$$ where $C_d$ is as in and $$- \left( \frac{2C_d}{(d-1)} \right) \left( 1 - \frac{C_d}{(d-1)} \right)^{N+1} \leq \alpha_{d,N} \leq 54 (C_d)^2 \left( 1 - \frac{2 C_d}{(d-1)} \right)^N.$$ Note that by , $0< (1 - \frac{C_d}{d-1})<1$ and $0 < (1 - \frac{2C_d}{d-1}) <1$. By Lemma \[YN-lemma\] we need to estimate the following integral $$\label{RadialThmDimdEq1} \mathbb{E}|R_N(\psi)|^2 = \mathbb{E}|Y_N|^2 = 2 \int_0^{\infty} \lambda \Pr[Y_N >\lambda] d\lambda = 2 \int_0^{\infty} \lambda \left( \Pr[X >\lambda] \right)^N d\lambda.$$ [*Step I.*]{} If $0 \leq \lambda \leq 2\delta$ then using gives $$\begin{aligned} \Pr [ X > \lambda] &= \Pr [ \xi > \lambda Z] = \int_0^1\Pr[\xi >\lambda z] f_Z(z) dz \notag \\ & = \int_0^1 \left( \frac{2\delta - \lambda z}{2\delta} \right) 2 C_d(1-z^2)^{\frac{d-3}{2}} dz \notag \\ &= \left( 1 - \frac{\lambda C_d}{\delta(d-1)} \right). \label{stepI-eq1}\end{aligned}$$ A computation shows that $$\begin{aligned} \int_0^{2\delta} & \lambda \left( \Pr[X >\lambda] \right)^N d\lambda = \int_0^{2\delta} \lambda \left( 1 - \frac{\lambda C_d}{\delta(d-1)} \right)^N d\lambda \notag \\ &= \frac{\delta^2(d-1)^2}{(C_d)^2} \left( \frac{1}{(N+1)(N+2)} - \left( 1 - \frac{2C_d}{(d-1)} \right)^{N+1} \left( \frac{1}{N+1} - \frac{(1 - \frac{2C_d}{(d-1)})}{N+2} \right) \right). \label{step1-eq2}\end{aligned}$$ [*Step II.*]{} If $\lambda \geq 2\delta$ then then gives $$\Pr[X>\lambda] \leq \Pr[X>2\delta] \leq \left( 1 - \frac{2 C_d}{(d-1)} \right).$$ Thus $$\begin{aligned} \int_{2\delta}^{6\pi\delta C_d} \lambda \left( \Pr[X >\lambda] \right)^N d\lambda &\leq \int_{2\delta}^{6\pi\delta C_d} \lambda \left( 1 - \frac{2 C_d}{(d-1)} \right)^N d\lambda \notag \\ & \leq 18\pi^2\delta^2 (C_d)^2 \left( 1 - \frac{2 C_d}{(d-1)} \right)^N. \label{stepII-eq1}\end{aligned}$$ [*Step III.*]{} If $\lambda \geq 2 \delta$ then a computation using and $d\geq2$ shows that $$\begin{aligned} \Pr[X>\lambda] &= 2 C_d \int_0^{2\delta/\lambda} \left( \frac{2\delta - \lambda z}{2\delta}\right) (1-z^2)^{\frac{d-3}{2}} dz \notag \\ & \leq 2 C_d \int_0^{2\delta/\lambda} (1-z^2)^{-\frac{1}{2}} dz \notag \\ & \leq\frac{2\pi \delta C_d}{\lambda}. \label{stepIII-eq1}\end{aligned}$$ If $d\geq 2$ then along with $\Gamma(1/2)=1/\sqrt{\pi}$ and $\Gamma(3/2) = \sqrt{\pi}/2$ implies that $\frac{1}{3} \leq (1 - \frac{2C_d}{d-1})$. Equation implies that $$\begin{aligned} \int_{6\pi\delta C_d}^{\infty} \lambda \left( \Pr[X>\lambda] \right)^N d \lambda & \leq \int_{6\pi\delta C_d}^{\infty} \lambda \left( \frac{2 \pi \delta C_d}{\lambda} \right)^N d\lambda \notag \\ & = \left( \frac{1}{3} \right)^N \frac{36\pi^2 \delta^2 (C_d)^2}{(N-2)}\notag \\ & \leq 36 \pi^2\delta^2 (C_d)^2 \left( 1 - \frac{2 C_d}{(d-1)} \right)^N. \label{stepIII-eq2}\end{aligned}$$ Combining , , and now yields the desired conclusion . \[lower-bnd-cor\] Suppose that $\{\varphi_n\}_{n=1}^N\subset \mathbb{S}^{d-1}$ are i.i.d. uniformly distributed random vectors on $\mathbb{S}^{d-1}$. Then $${\rm lim \thinspace inf}_{N \to \infty} N^2 \thinspace \mathbb{E}|W_N|^2 \geq \lim_{N\to \infty} N^2 \thinspace \EE |R_N|^2 = 2 \delta^2\left(\frac{d-1}{2\,C_d}\right)^2 \geq \pi \delta^2 (d-1).$$ Consistent reconstruction and coverage processes {#cover-sec} ================================================ In this section we describe a useful connection between the worst case error $W_N$ and coverage processes on the sphere $\mathbb{S}^{d-1}$. This relationship will play a central role in the proofs of our upper bounds on $\mathbb{E}|W_N|^2$ that appear in the subsequent sections. Worst case error and coverage processes --------------------------------------- The expected worst case error squared $\mathbb{E}|W_N|^2$ can be represented as $$\label{MSE-intbyparts} \mathbb{E}|W_N|^2 = 2 \int_0^{\infty} \lambda \ \Pr [ W_N > \lambda ] \thinspace d \lambda.$$ So, a main step towards bounding $\mathbb{E}|W_N|^2$ is to bound the probability $\Pr [ W_N > \lambda ]$. The next lemma shows that bounding $\Pr [ W_N > \lambda ]$ can be reformulated as a coverage problem. \[cover-lem\] For each $\lambda>0$ define the set $$\label{Bn-def} B_n=B_n(\lambda) = \left\{ u \in \mathbb{S}^{d-1} : \langle u , \varphi_n \rangle > \frac{\epsilon_n+\delta}{\lambda} \ \hbox{ or } \ \langle u , \varphi_n \rangle < \frac{\epsilon_n - \delta}{\lambda} \right\}.$$ Then $$\label{coverage-eq2} \forall \thinspace \lambda>0, \ \ \ \Pr [ W_N \geq \lambda ] = \Pr \left( \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda) \right).$$ In particular, $$\label{coverage-eq} \forall \thinspace \lambda>0, \ \ \ \Pr [ W_N > \lambda ] \leq \Pr \left( \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda) \right).$$ Let $E_n$ be as in and define the set $$\begin{aligned} A_n =A_n(\lambda)= ({\lambda \mathbb{S}^{d-1}) \backslash E_n} = \{ u \in \lambda \mathbb{S}^{d-1} : \langle u , \varphi_n \rangle > \epsilon_n + \delta \ \hbox{ or } \ \langle u , \varphi_n \rangle < \epsilon_n - \delta \}.\end{aligned}$$ Observe that $$\label{cover-lambda} W_N \geq \lambda \ \ \ \hbox{if and only if} \ \ \ \lambda \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N A_n(\lambda).$$ It only remains to rescale to the unit-sphere. Using the map $u \mapsto u /\lambda$ and the set $B_n(\lambda)$ defined by one has that $$\label{coverage-equiv} \lambda \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N A_n(\lambda) \ \ \ \hbox{if and only if} \ \ \ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda).$$ The proof of now follows by combining and . We shall refer to the set $B_n(\lambda)$ as a bi-cap since it can be expressed as the union of two antipodal (possibly empty) open spherical caps $$\label{bicap-eq} B_n(\lambda) = {\rm Cap}(\varphi_n, \theta_n^{+})\cup {\rm Cap}(-\varphi_n, \theta_n^{-}),$$ where the angular radii $\theta_n^{+}$ and $\theta_n^{-}$ are given by $$\label{thetaplus} \theta^{+}_n = \begin{cases} \arccos \left( \frac{\delta+\epsilon_n}{\lambda} \right), & \hbox{ if } \delta + \epsilon_n < \lambda,\\ 0, & \hbox{ otherwise,} \end{cases}$$ and $$\label{thetaminus} \theta^{-}_n = \begin{cases} \arccos \left( \frac{\delta- \epsilon_n}{\lambda} \right), & \hbox{ if } \delta - \epsilon_n < \lambda,\\ 0, & \hbox{ otherwise.} \end{cases}$$ In particular, depending on the size of the parameters $\epsilon_n$ and $\lambda$, each set $B_n(\lambda)$ is either: (i) a union of two disjoint spherical caps with antipodal centers, or (ii) a single spherical cap, or (iii) the empty set. Background on coverage processes -------------------------------- In our general analysis of consistent reconstruction, the coverage problem in Lemma \[cover-lem\] involves spherical caps with both random angular radii and random centers. Random coverage problems have a long and technical history, e.g., see [@S; @BCL10], but in high dimensions the literature is still limited when considering caps with both random center and random size. In fact, even in the case of constant sized caps with random centers on $\mathbb{S}^{d-1}$, bounds on coverage probabilities were only recently obtained in [@BCL10]. Some noteworthy results for randomly sized caps include [@SH] which contains exact results in dimension $d=2$ with general distributions on the random arclengths, and [@J86] which contains asymptotic results (as the random cap size becomes small) on general manifolds. In this section, we shall provide some necessary background on coverage processes in the case of spherical caps with angular radii of fixed size and random centers $\{\varphi_n\}_{n=1}^N$ that are uniformly distributed on the unit-sphere. The techniques and results that we will use later are especially influenced by [@BCL10] and [@FN77]. For the remainder of this section let $\{\varphi_n\}_{n=1}^N \subset \mathbb{S}^{d-1}\subset \R^{d}$ be i.i.d. random vectors that are uniformly distributed on $\mathbb{S}^{d-1}$, and let $0<\theta< \pi/2$ be fixed. Consider the following non-coverage probability $$\label{p-def} p(N,d-1,\theta) = \Pr \left( \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N {\rm Cap}(\varphi_n,\theta) \right).$$ The following theorem contains the best known bounds on $p(N,d-1,\theta)$ when $0<\theta<\pi/2$, see Theorem 1.1 and Proposition 5.5 in [@BCL10]. The work in [@BCL10] is stated for coverage by closed caps but the following result remains true for open caps. \[thmBCL\] If $0 < \theta < \pi/2$ and $N \geq d \geq 2$ then $$\begin{aligned} p(N,d-1,\theta) \leq 2^{1-N} \sum_{k=0}^{d-1} \binom{N-1}{k} + \binom{N}{d} \ \left( \frac{d\sqrt{d-1}}{2^{d-1}} \right) \ F_{N,d-1}(\theta), \label{BCL-bnd}\end{aligned}$$ where $$\begin{aligned} F_{N,d-1}(\theta)= \int_0^{\cos \theta} (1-t^2)^{((d-1)^2-2)/2}(1 - r_{d-1}(\arccos t))^{N-d-2} dt\end{aligned}$$ and $r_{d-1}( \arccos t)$ is defined using . We briefly comment on why Theorem \[thmBCL\] holds when $p(N,d-1,\theta)$ is defined as in using open spherical caps instead of closed spherical caps as in [@BCL10]. For this it suffices to note that if $0<\alpha <\theta<\pi/2$ then $$\Pr \left( \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N {\rm Cap}(\varphi_n,\theta) \right) \leq \Pr \left( \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N \overline{{\rm Cap}(\varphi_n,\alpha)} \right),$$ and that $$\lim_{\alpha \to \theta} F_{N,d-1}(\alpha) = F_{N,d-1}(\theta).$$ We shall later need bounds on $p(N,d-1, \theta)$ when $\arccos(1/\sqrt{d}) \leq \theta < \pi/2$, i.e., when the cap height is less than $1/\sqrt{d}$. \[pbound-lem\] If $N \geq \frac{2d}{\ln(12/11)}\approx (22.99)d$ and $\arccos(1/\sqrt{d}) \leq \theta < \pi/2$ then $$\begin{aligned} p(N,(d-1), \theta) & \leq 2 \sqrt{d} \ \left( 13 \right)^d \left( \frac{11}{12} \right)^{N/2}.\end{aligned}$$ In particular, if $s> \frac{2 \ln(13)}{\ln(12/11)}\approx 58.96$ then $$\lim_{d \to \infty} p(sd,(d-1), \theta)=0. \label{pds-lim}$$ Hoeffding’s inequality shows that if $(N-1) \geq 4(d-1)$ then $$\begin{aligned} 2^{1-N} \sum_{k=0}^{d-1} \binom{N-1}{k} & \leq e^{1/8} e^{-N/8}. \label{HemiBnd}\end{aligned}$$ By Lemma 2.1 of [@BGKKLS] we have that $$\arccos(1/\sqrt{d}) \leq \theta < \pi/2 \implies \frac{1}{12} \leq r_{d-1}(\theta) \leq \frac{1}{2}.$$ Thus, for $\arccos(1/\sqrt{d}) \leq \theta < \pi/2$ we have $$\begin{aligned} F_{N,d-1}(\theta) &\leq \int_0^{1/\sqrt{d}} (1-t^2)^{((d-1)^2-2)/2}(1 - r_{d-1}(t))^{N-(d-1)-1} dt \notag \\ &\leq \frac{1}{\sqrt{d}} \left( \frac{11}{12} \right)^{N-d}. \label{F-bnd}\end{aligned}$$ Combining , , gives $$\begin{aligned} p(N,d-1, \theta) &\leq e^{1/8}e^{-N/8} + \binom{N}{d} \frac{d\sqrt{d-1}}{2^{d-1}} \frac{1}{\sqrt{d}} \left( \frac{11}{12} \right)^{N-d} \notag \\ & \leq e^{1/8}e^{-N/8} + 2d \left( \frac{12}{22} \right)^d \left( \frac{11}{12} \right)^{N} \binom{N}{d}. \label{p-eq-unsimp}\end{aligned}$$ We shall use the following bounds to further simplify . First, note that by Stirling’s approximation $$\label{binom-stirling} \binom{N}{d} \leq \frac{N^d}{d!} \leq \frac{1}{\sqrt{2\pi d}} \left( \frac{eN}{d} \right)^d.$$ Also, it follows from $\ln(x) \leq (x/c) + (\ln(c) -1)$ that one has $$\label{dd-bnd} \forall c>0, \ \ \ \left( \frac{eN}{d} \right)^d \leq c^d e^{eN/c} e^{-d}.$$ Applying and with $c = 2e/ \ln(12/11)$ to yields that if $N \geq \frac{2d}{\ln(12/11)}$ then $$\begin{aligned} p(N,d-1, \theta) & \leq e^{1/8}e^{-N/8} + 2d \left( \frac{12}{22} \right)^d \left( \frac{11}{12} \right)^{N} \binom{N}{d}\\ & \leq e^{1/8}e^{-N/8} + \frac{2\sqrt{d}}{\sqrt{2\pi}} \left( \frac{12}{22} \right)^d \left( \frac{2}{\ln(12/11)} \right)^d \left(\frac{11}{12} \right)^{N/2} \\ & \leq e^{1/8}e^{-N/8} + \sqrt{d} \thinspace \left( 13 \right)^d \left( \frac{11}{12} \right)^{N/2} \\ & \leq 2 \sqrt{d} \thinspace \left( 13 \right)^d \left( \frac{11}{12} \right)^{N/2}.\end{aligned}$$ This completes the proof. Upper bounds for general distributions {#genthm-sec} ====================================== In this section we prove that consistent reconstruction achieves MSE of the optimal order $\mathbb{E}|W_N|^2 \lesssim 1/N^2$ under rather general conditions on the i.i.d. random measurement vectors $\{\varphi_n\}_{n=1}^N \subset \R^d$. Our error bounds use the following admissibility condition. We assume that $\{\varphi_n\}_{n=1}^N\subset \R^d$ are independent identically distributed versions of a unit-norm random vector $\varphi\in \mathbb{S}^{d-1}$, and we further assume that there exist constants $\alpha \geq 1, s>0$ such that $$\label{general-cond} \forall \thinspace 0\leq t \leq 1, \ \forall x \in \mathbb{S}^{d-1}, \ \ \ \Pr \left( |\langle x, \varphi \rangle | \leq t \right) \leq \alpha \thinspace t^s.$$ Roughly speaking, the admissibility condition ensures that the random vector $\varphi$ cannot be too concentrated on any subspace of $\R^d$ with positive codimension. \[unif-gencond-ex\] If $d\geq 3$ and if the random vector $\varphi$ is uniformly distributed on the unit-sphere $\mathbb{S}^{d-1}$ then $\varphi$ satisfies with $s=1$ and $\alpha = 2C_d$, where $C_d$ is as in . This follows since $0 \leq f_Z(z)\leq 2C_d$ in . Similarly, when $d=2$, a direct computation shows that if $\varphi$ is uniformly distributed on the unit-circle $\mathbb{S}^{1}$ then $\varphi$ satisfies with $s=1$ and $\alpha = 1$. This follows using $2C_2 = 2/\pi$ and $\arcsin(t) \leq (\pi/2)t$. \[beta-dominate-ex\] Suppose that the unit-norm random vector $\varphi_1\in \mathbb{S}^{d-1}$ satisfies with $\alpha=\alpha_1$ and $s=s_1$. Suppose that the unit-norm random vector $\varphi_2 \in \mathbb{S}^{d-1}$ has the property that there exists $\beta>0$ such that for every Borel subset $B\subset \mathbb{S}^{d-1}$ there holds $\Pr [ \varphi_2\in B] \leq \beta \thinspace \Pr [ \varphi_1\in B].$ Then $\varphi_2$ satisfies with $\alpha=\beta \alpha_1$ and $s=s_1$. If the random vector $\varphi$ is uniformly distributed on an open subset of $\mathbb{S}^{d-1}$ then $\varphi$ satisfies . This follows by combining Examples \[unif-gencond-ex\] and \[beta-dominate-ex\]. If the random vector $\varphi$ has a point mass at $v\in \R^d$ with $\Pr [ \varphi = v]>0$ then $\varphi$ does not satisfy . To see this, let $x_v\in \mathbb{S}^{d-1}$ be any vector that is orthogonal to $v$. Then $\Pr [ \langle \varphi, x_v \rangle =0] \geq \Pr [ \varphi =v] >0$ shows that does not hold. We are now ready to state and prove our first main theorem. \[main-thm1\] Suppose that $\{\varphi_n\}_{n=1}^N \subset \mathbb{S}^{d-1}$ are i.i.d. versions of a unit-norm random vector $\varphi$ that satisfies . For all $N \geq (d+2)/s$ the expected worst case error squared satisfies $$\label{thm1bnd} \mathbb{E} |W_N|^2 \leq \frac{10^5 \delta^2d^2(2 \alpha)^{2/s}\ln^2(16 (2\alpha)^{1/s})}{(N+1)(N+2)} + \delta^2 32^{d+1}(2\alpha)^{(d+1)/s}\left(\frac{1}{2} \right)^N.$$ The proof is divided into several steps. [*Step I.*]{} To use the error represention we need to bound $\Pr [ W_N > \lambda]$. By Lemma \[cover-lem\] this will be done by bounding the coverage probability . We begin by discretizing the coverage problem with an $\epsilon$-net argument developed in [@FN77]. Given any $\epsilon>0$, let $\mathcal{N}_{\epsilon} = \{z_m\}_{m=1}^M \subset \mathbb{S}^{d-1}$ be a geodesic $\epsilon$-net for $\mathbb{S}^{d-1}$ with cardinality satisfying $$M=\#(\mathcal{N}_{\epsilon}) \leq \left( \frac{8}{\epsilon} \right)^{d-1}.$$ Recall that the bi-cap $B_n(\lambda)$ is defined by and as $$B_n(\lambda) = {\rm Cap}(\varphi_n, \theta_n^{+})\cup {\rm Cap}(-\varphi_n, \theta_n^{-}).$$ Next define the shrunken bi-cap $$T_{\epsilon}(B_n(\lambda)) = {\rm Cap}(\varphi_n, T_{\epsilon}(\theta_n^{+}))\cup {\rm Cap}(-\varphi_n, T_{\epsilon}(\theta_n^{-})),$$ where $T_{\epsilon}(\theta_n^{+})$ and $T_{\epsilon}(\theta_n^{-})$ are defined by , and . The key discretization step is to proceed as in [@FN77] and note that $$\label{cover-discrete} \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda) \ \ \ \implies \mathcal{N}_{\epsilon} \not\subset \bigcup_{n=1}^N T_{\epsilon}(B_n(\lambda)).$$ Since the shrunken bi-caps $\{T_{\epsilon}(B_n(\lambda))\}_{n=1}^N$ are independent and identically distributed, implies $$\begin{aligned} \Pr ( W_N > \lambda ) &\leq \Pr \left( \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda) \right) \notag \\ &\leq \Pr \left( \mathcal{N}_{\epsilon} \not\subset \bigcup_{n=1}^N T_{\epsilon}(B_n(\lambda)) \right) \notag\\ & \leq \sum_{m=1}^{M} \Pr \left( z_m \notin \bigcup_{n=1}^N T_{\epsilon}(B_n(\lambda)) \right) \notag \\ & = \sum_{m=1}^{M} \large\left( \Pr \left[ z_m \notin T_{\epsilon}(B_1(\lambda)) \right] \large\right)^N \notag\\ & \leq \left( \frac{8}{\epsilon} \right)^{d-1} \sup_{z\in \mathbb{S}^{d-1}} \left( \Pr [ z \notin T_{\epsilon}(B_1(\lambda)) ] \right)^N.\label{cover-discrete-bnd}\\ \notag\end{aligned}$$ [*Step II.*]{} We now use to bound $\Pr[W_N>\lambda]$ in the case when $\lambda \geq 4 \delta$. In this case note that each $B_n(\lambda)$ is a genuine bi-cap that consists of two antipodal non-empty spherical caps. Since $\lambda >2 \delta$ it is straightforward from that $$B_n(\lambda) \supset \left\{ u \in \mathbb{S}^{d-1} : \langle u , \varphi_n \rangle > \frac{2\delta}{\lambda} \ \hbox{ or } \ \langle u , \varphi_n \rangle < \frac{-2\delta}{\lambda} \right\},$$ and it follows that the shrunken bi-cap $T_{\epsilon}(B_n(\lambda))$ satisfies $$\begin{aligned} T_{\epsilon}(B_n(\lambda)) &\supset \left\{ u \in \mathbb{S}^{d-1} : |\langle u , \varphi_n \rangle| > \cos \left( \arccos \left(\frac{2\delta}{\lambda}\right) -\epsilon \right) \right\} \notag \\ & \supset \left\{ u \in \mathbb{S}^{d-1} : |\langle u , \varphi_n \rangle| > \frac{2\delta}{\lambda} +\epsilon \right\}. \label{bicap-subset}\end{aligned}$$ For the remainder of this step we fix $\epsilon = \frac{2\delta}{\lambda}$. The assumption along with shows that if $\lambda \geq 4\delta$ then $$\begin{aligned} \Pr [ z \notin T_{\epsilon}(B_1(\lambda))] &\leq \Pr \left( |\langle z, \varphi \rangle| \leq\frac{2\delta}{\lambda} + \epsilon \right) \label{eq-useinthm2}\\ & = \Pr \left( |\langle z, \varphi \rangle| \leq\frac{4\delta}{\lambda} \right) \notag\\ & \leq \alpha \left( \frac{ 4\delta }{\lambda} \right)^s. \label{step2bnd}\end{aligned}$$ Thus and imply that if $\lambda \geq 4\delta$ then $$\label{step2result} \Pr [ W_N > \lambda ] \leq \left( \frac{4\lambda}{\delta} \right)^{d-1} \left( \alpha \left( \frac{ 4\delta }{\lambda} \right)^s \right)^N = 4^{d-1} (4^s \alpha)^N \left( \frac{\delta}{\lambda}\right)^{sN-d+1}.\\$$ Since $8\delta(2\alpha)^{1/s} \geq 4\delta$ note that if $N \geq (d+2)/s$ then implies $$\begin{aligned} \int_{8\delta(2\alpha)^{1/s}}^{\infty} \lambda \Pr [ W_N > \lambda] d \lambda & \leq 4^{d-1} (4^s \alpha)^N \int_{8\delta(2\alpha)^{1/s}}^{\infty} \lambda \left( \frac{\delta}{\lambda}\right)^{sN-d+1} d \lambda \notag\\ & = 4^{d-1} \left(\frac{1}{2(2^s)} \right)^N \delta^2 \frac{(8(2\alpha)^{1/s})^{d+1}}{(sN-d-1)} \notag \\ & \leq \delta^2 (32(2\alpha)^{1/s})^{d+1} \left(\frac{1}{2} \right)^N. \label{int2}\end{aligned}$$ [*Step III.*]{} Next, we bound $\Pr[W_N>\lambda]$ in the case when $0< \lambda \leq {4(2\alpha)^{1/s}\delta}$. It will be useful to begin with the following symmetrization argument. Let $\{b_n\}_{n=1}^N$ be i.i.d. Bernoulli random variables satisfying $\Pr[b_n=1] = \Pr[b_n=-1]=1/2$, and additionally suppose that $\{b_n\}_{n=1}^N$ is independent of $\{\varphi_n\}_{n=1}^N$ and $\{\epsilon_n\}_{n=1}^N$. Note that the i.i.d. bi-caps $\{B_n(\lambda)\}_{n=1}^N$ have the same distribution as the i.i.d. bi-caps $\{B_n^{\prime}(\lambda)\}_{n=1}^N$ defined by $$B_n^{\prime}(\lambda) = {\rm Cap}(b_n \varphi_n, \theta^+_n(\lambda))\cup {\rm Cap}(-b_n \varphi_n, \theta^-_n(\lambda)).$$ This follows from , , and the fact that $\epsilon_n$ is uniformly distributed on $[-\delta,\delta]$. Consequently, $$\Pr[\mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda)] = \Pr[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B^{\prime}_n(\lambda)].$$ Let $I(\lambda)$ denote the number of $\{\epsilon_n + \delta\}_{n=1}^N$ that lie in the interval $\left[0,\frac{\lambda}{2(2\alpha)^{1/s}} \right] $, namely: $$I(\lambda) = \# \left( \{\epsilon_n + \delta\}_{n=1}^N \cap \left[0,\frac{\lambda}{2(2\alpha)^{1/s}} \right] \right).$$ Let $$\label{qdef} q(k, d-1, \alpha,s)=\Pr \left[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^k {\rm Cap} \left(b_n \varphi_n, \arccos\left(\frac{1}{2(2\alpha)^{1/s}}\right) \right)\right].$$ Thus $$\begin{aligned} \Pr [ W_N>\lambda] &\leq \Pr[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda)] \notag \\ & = \Pr[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B^{\prime}_n(\lambda)] \notag \\ & \leq \Pr[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N{\rm Cap}(b_n \varphi_n, \theta^+_n(\lambda))] \notag \\ & = \sum_{k=0}^N \Pr \left[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N {\rm Cap}(b_n \varphi_n, \theta^+_n(\lambda) ) \ \bigg\vert \ I(\lambda) = k\right] \ \Pr [ I(\lambda)=k] \notag \\ & \leq \sum_{k=0}^N \Pr \left[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^k {\rm Cap}(b_n \varphi_n, \arccos(1/2(2\alpha)^{1/s}) )\right] \ \Pr [ I(\lambda)=k] \notag \\ & = \sum_{k=0}^N q(k, d-1, \alpha,s) \ \Pr[ I(\lambda) = k] \notag \\ & = \sum_{k=0}^N q(k, d-1, \alpha,s) \binom{N}{k} \left( 1 - \frac{\lambda}{4\delta(2\alpha)^{1/s}}\right)^{N-k} \left( \frac{\lambda}{4\delta(2\alpha)^{1/s}} \right)^k. \label{thm1step3eq}\end{aligned}$$ [*Step IV.*]{} In this step we bound the integral $\int_0^{4\delta (2\alpha)^{1/s}} \lambda \Pr[W_N > \lambda] d\lambda$. Equation and properties of the beta function imply $$\begin{aligned} &\int_0^{4\delta(2\alpha)^{1/s}} \lambda \Pr[W_N > \lambda] d\lambda \notag \\ & \leq \sum_{k=0}^N q(k, d-1, \alpha,s) \binom{N}{k} \int_0^{4\delta (2\alpha)^{1/s}} \lambda \left( 1 - \frac{\lambda}{4\delta(2\alpha)^{1/s}} \right)^{N-k} \left( \frac{\lambda}{4\delta(2\alpha)^{1/s}} \right)^k d \lambda \notag\\ &= \sum_{k=0}^N q(k, d-1, \alpha,s) \binom{N}{k} {16\delta^2(2\alpha)^{2/s}} \int_0^1 u^{k+1} (1- u)^{N-k} d \lambda \notag \\ &=\sum_{k=0}^N q(k, d-1, \alpha,s) \binom{N}{k} {16\delta^2(2\alpha)^{2/s}} \frac{(k+1)! (N-k)!}{(N+2)!} \notag \\ &=\frac{16\delta^2(2\alpha)^{2/s}}{(N+1)(N+2)} \sum_{k=0}^N q(k, d-1, \alpha,s) \thinspace (k+1) \notag \\ &\leq \frac{16\delta^2(2\alpha)^{2/s}}{(N+1)(N+2)} \left( 1 + 2 \sum_{k=1}^N k \thinspace q(k, d-1,\alpha,s) \right). \label{thm1stepivqsum}\end{aligned}$$ [*Step V.*]{} In this step we bound the quantity $q(k, d-1, \alpha,s)$ appearing in and defined in . We shall again employ an $\epsilon$-net argument as in Step I. For the remainder of this step fix $$\label{thm1stepvepsdef} \epsilon = \frac{1}{2(2 \alpha)^{1/s}} \left( 1 - \frac{1}{4(2\alpha)^{2/s}}\right)^{-1/2},$$ and let $\mathcal{N}_{\epsilon}$ be a geodesic $\epsilon$-net for $\mathbb{S}^{d-1}$ with cardinality $\#(\mathcal{N}_{\epsilon}) \leq \left( \frac{8}{\epsilon} \right)^{d-1}$. A similar argument as used to obtain yields the following $$\begin{aligned} q(k, d-1,\alpha,s) & = \Pr \left[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^k {\rm Cap}(b_n \varphi_n, \arccos(1/2(2\alpha)^{1/s}) ) \right] \notag \\ & \leq \Pr [ \mathcal{N}_{\epsilon} \not\subset \bigcup_{n=1}^k {\rm Cap}(b_n \varphi_n, \arccos(1/2(2\alpha)^{1/s}) - \epsilon )] \notag \\ &\leq \left( \frac{8}{\epsilon} \right)^{d-1} \left( \sup_{z\in \mathbb{S}^{d-1}} \Pr \left[ z \notin {\rm Cap}(b_1 \varphi_1, \arccos(1/2(2\alpha)^{1/s}) - \epsilon ) \right] \right)^k. \label{thm1stepvqepsnetbnd}\end{aligned}$$ By and , for an arbitrary $z\in \mathbb{S}^{d-1}$ we may compute as follows $$\begin{aligned} \Pr &\left[ z \notin {\rm Cap}(b_1 \varphi_1, \arccos(1/2(2\alpha)^{1/s})- \epsilon ) \right] \notag \\ &= \frac{1}{2} \Pr \left[ \langle z, \varphi_1 \rangle \leq \cos(\arccos(1/ 2(2\alpha)^{1/s}) - \epsilon) \right] + \frac{1}{2} \Pr \left[ -\langle z, \varphi_1 \rangle \leq \cos(\arccos(1/ 2(2\alpha)^{1/s}) - \epsilon) \right] \notag \\ &= \frac{1}{2} \left( 1 + \Pr\left[ |\langle z, \varphi_1 \rangle| \leq \cos(\arccos(1/ 2(2\alpha)^{1/s}) - \epsilon) \right] \right) \notag\\ &= \frac{1}{2} \left( 1 + \Pr\left[ |\langle z, \varphi_1 \rangle| \leq \frac{\cos \epsilon}{2(2\alpha)^{1/s}} + \left(1 - \frac{1}{4(2\alpha)^{2/s}}\right)^{1/2} \sin \epsilon \right] \right) \notag \\ &\leq \frac{1}{2} \left( 1 + \Pr\left[ |\langle z, \varphi_1 \rangle| \leq \frac{1}{2(2\alpha)^{1/s}} + \left(1 - \frac{1}{4(2\alpha)^{2/s}}\right)^{1/2} \epsilon \right] \right) \notag \\ &= \frac{1}{2} \left( 1 + \Pr\left[ |\langle z, \varphi_1 \rangle| \leq \frac{1}{(2\alpha)^{1/s}} \right] \right) \notag \\ & \leq \frac{1}{2} \left( 1 + \alpha \left|\frac{1}{(2\alpha)^{1/s}}\right|^s \right) \notag \\ &= \left(\frac{3}{4}\right). \label{thm1stepv34bnd}\end{aligned}$$ Combining , , and gives $$\begin{aligned} q(k,d-1,\alpha,s) &\leq \left( \frac{8}{\epsilon} \right)^{d-1} \left( \frac{3}{4} \right)^k \notag \\ & = \left( 8 \sqrt{4(2\alpha)^{2/s}-1} \right)^{d-1} \left( \frac{3}{4} \right)^k \notag\\ & \leq \left( 16 (2\alpha)^{1/s} \right)^{d} \left( \frac{3}{4} \right)^k. \label{StepVeq}\end{aligned}$$ [*Step VI.*]{} In this step we bound the sum $\sum_{k=1}^N k \thinspace q(k, d-1, \alpha,s)$ appearing in . For the remainder of this step let $$\label{stepviKdef} K = \frac{2d \ln(16 (2\alpha)^{1/s})}{\ln(4/3)},$$ and note that $K>1$. If $k \geq K$ then by $$q(k,d-1,\alpha,s) \leq \left(\frac{3}{4}\right)^{k/2}.$$ Thus $$\begin{aligned} \sum_{k=K+1}^N k \thinspace q(k, d-1,\alpha,s) &\leq \sum_{k=K+1}^N k \left(\frac{3}{4}\right)^{k/2} \notag \\ &\leq \sum_{k=1}^{\infty} (k+K) \left(\sqrt{\frac{3}{4}}\right)^{k+K}\notag \\ &\leq 2K \sum_{k=1}^{\infty} k \left(\sqrt{\frac{3}{4}}\right)^{k}\notag \\ &\leq 120K. \label{Kupsum}\end{aligned}$$ Here we have used that $$\label{geosum} \forall \thinspace 0< r< 1, \ \ \ \sum_{k=1}^{\infty} k r^k = \frac{1}{(1-r)^2}.$$ Since $0 \leq q(k, d-1,\alpha,s) \leq 1$ we also have $$\begin{aligned} \sum_{k=1}^{K} k \thinspace q(k, d-1, \alpha,s) & \leq \sum_{k=1}^{K} k \leq K^2. \label{Klowsum}\end{aligned}$$ Since $K>1$, combining and gives $$\label{stepviqsumeq} 1+2\sum_{k=1}^N k q(k, d-1,\alpha,s) \leq 1+2(120K+K^2) \leq 250K^2.$$ By , , and we conclude that $$\begin{aligned} \int_0^{4\delta(2\alpha)^{1/s}} \lambda \Pr[W_N> \lambda] d \lambda &\leq \frac{16 \delta^2(2 \alpha)^{2/s} (250K^2)}{(N+1)(N+2)} \notag \\ &\leq\frac{(16)(250) \delta^2(2 \alpha)^{2/s}}{(N+1)(N+2)} \left(\frac{2d \ln(16 (2\alpha)^{1/s})}{\ln(4/3)}\right)^2 \notag \\ & \leq\frac{10^5 \delta^2d^2(2 \alpha)^{2/s}\ln^2(16 (2\alpha)^{1/s})}{(N+1)(N+2)}. \label{thm1stepvieq}\end{aligned}$$ [*Step VII.*]{} It remains to bound the integral $\int_{4\delta(2\alpha)^{1/s}}^{8\delta(2\alpha)^{1/s}} \lambda \Pr[W_N>\lambda] d \lambda$. Note that $$4\delta(2\alpha)^{1/s} \leq \lambda \implies \Pr[W_N>\lambda] \leq \Pr[W_N>4\delta(2\alpha)^{1/s}].$$ Since $4\delta(2\alpha)^{1/s}\geq 4\delta$ it follows from that $$\begin{aligned} \Pr[W_N>4\delta(2\alpha)^{1/s}] & \leq 4^{d-1}(4(2\alpha)^{1/s})^{d-1} \left( \frac{1}{2}\right)^N.\end{aligned}$$ Consequently $$\begin{aligned} \int_{4\delta(2\alpha)^{1/s}}^{8\delta(2\alpha)^{1/s}} \lambda \Pr[W_N>\lambda] d \lambda &\leq 4^{d-1}(4(2\alpha)^{1/s})^{d-1} \left( \frac{1}{2}\right)^N 24 \delta^2 (2\alpha)^{2/s} \notag \\ & \leq 2\delta^2 16^d (2\alpha)^{(d+1)/s} \left( \frac{1}{2}\right)^N. \label{thm1stepviieq}\end{aligned}$$ The bound now follows from , , and . This completes the proof. Suppose that $\{\varphi_n\}_{n=1}^N \subset \mathbb{S}^{d-1}$ are i.i.d. versions of a unit-norm random vector $\varphi$ that satisfies . There exist absolute constants $c_1, c_2>0$ such that if $$N \geq c_2 d \ln (32(2\alpha)^{1/s})$$ then the expected worst case error squared satisfies $$\label{corthm1bnd} \mathbb{E}|W_N|^2 \leq \frac{c_1 \delta^2 d^2(2\alpha)^{2/s} \ln^2(16(2\alpha)^{1/s}) }{(N+1)(N+2)}.$$ If $d\geq3$ and $\{\varphi_n\}_{n=1}^N\subset \mathbb{S}^{d-1}$ are i.i.d. uniformly distributed random vectors on $\mathbb{S}^{d-1}$ then Example \[unif-gencond-ex\] and show that one can take $s=1$ and $\alpha = 2C_d \leq 2\sqrt{\frac{d-1}{2\pi}}\leq \sqrt{d}$. In this case the error bound shows that there exist constants $a,c$ such that if $N \geq a d \ln d$ then $$\mathbb{E}|W_N|^2 \leq \frac{c \thinspace \delta^2 d^3 \ln^2 d}{(N+1)(N+2)}. \label{d3logd}$$ In the next section (see Theorem \[mainthm2\]) we show that the logarithmic term in can in fact be removed. For perspective, note that the admissibilty condition is stronger than the condition that was used to analyze the Rangan-Goyal algorithm. In particular, it is straightforward to show that if $\varphi \in \mathbb{S}^{d-1}$ is a random vector that satisfies , then $$\forall x \in \mathbb{S}^{d-1}, \ \mathbb{E} |\langle x, \varphi \rangle| \geq \left(\frac{1}{\alpha}\right)^s \left( 1- \frac{1}{s+1} \right)>0.$$ We conclude this section by noting that Theorem \[main-thm1\] does not generally hold under the weaker condition . For this we first show if $\varphi$ has a point mass then the conclusion does not hold. \[atom-counter-examp\] Suppose $\{ \varphi_n\}_{n=1}^N \subset \R^d$ are i.i.d. versions of a random vector $\varphi$ that has a point mass at $v \in \R^d$ that occurs with positive probability $c=\Pr [ \varphi = v]>0$. Then $$\forall \lambda >0, \ \ \ \Pr [W_N > \lambda ] \geq \Pr [W_N = \infty] \geq \Pr [ \varphi_1 = \varphi_2 = \cdots = \varphi_N = v] = c^N.$$ Thus $$\mathbb{E}|W_N|^2 = 2 \int_0^{\infty} \lambda \Pr [W_N > \lambda ] d \lambda \geq 2 \int_0^{\infty} \lambda c^N d \lambda = \infty.$$ In particular, the conclusion of Theorem \[main-thm1\] fails if $\varphi$ has a point mass. Let $\{e_n\}_{n=1}^d \subset \R^d$ be an orthonormal basis. Let $\varphi$ be the discrete random vector defined by $\Pr [ \varphi = e_n] = 1/d$ for each $1\leq n \leq d$. Then $$\forall x \in \R^d, \ \ \ \mathbb{E} | \langle x, \varphi \rangle | = (1/d)\sum_{n=1}^d | \langle x, e_n \rangle | \geq (1/d) \left( \sum_{n=1}^d | \langle x, e_n \rangle |^2 \right)^{1/2} = (1/d)\|x\|.$$ So, $\varphi$ satisfies , but by Example \[atom-counter-examp\], the conclusion of Theorem \[main-thm1\] does not hold. Upper bounds for uniformly distributed measurements {#unif-sec} =================================================== In this section we prove refined bounds for consistent reconstruction when $\{\varphi_n\}_{n=1}^N\subset \mathbb{S}^{d-1}$ are i.i.d. random vectors that are uniformly distributed on the unit-sphere. In this case, our next main result shows that the dimension dependent constant $K$ in is essentially dominated by the cube of the ambient dimension. \[mainthm2\] If the random vectors $\{\varphi_n\}_{n=1}^N \subset \R^d$ are i.i.d. uniformly distributed on $\mathbb{S}^{d-1}$ and if $N \geq d+2$ then $$\mathbb{E}|W_N|^2 \leq \frac{2e^{12}\delta^2 d^3}{(N+1)(N+2)} + 26\delta^2 d^{3/2}\left( \frac{11}{12} \right)^{N/2} e^{\frac{d \ln(1024d)}{2}}.$$ In particular, $$\limsup_{d\to \infty} \mathbb{E}|W_{\lceil d^{3/2} \rceil}|^2 \leq 2e^{12}\delta^2 \ \ \ \hbox{ and } \ \ \ \lim_{d\to \infty} \mathbb{E}|W_{ \lceil d^{3/2} \ln d \thinspace \rceil}|^2 = 0.$$ The proof is divided into several steps. We need to bound the integral $$\label{thm2-eq1} \mathbb{E}|W_N|^2 = 2 \int_0^{\infty} \lambda \Pr [ W_N > \lambda ] d \lambda.$$ [*Step I.*]{} In this step we provide preliminary bounds on $\Pr [ W_N > \lambda ]$ when $0 \leq \lambda \leq 2\delta\sqrt{d}$. Note that $\{\epsilon_n + \delta\}_{n=1}^N$ are i.i.d. uniform random variables on $[0, 2\delta]$ and define the random variable $$J(\lambda) = \# \left( \{ \epsilon_n+\delta \}_{n=1}^N \cap \left[0, \frac{\lambda}{\sqrt{d}}\right] \right).$$ Lemma \[cover-lem\], , imply that $$\begin{aligned} \Pr [ W_N>\lambda] &\leq \Pr[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda)] \notag \\ & \leq \Pr[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N{\rm Cap}(\varphi_n, \theta^+_n(\lambda))] \notag \\ & = \sum_{k=0}^N \Pr \left[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N {\rm Cap}(\varphi_n, \theta^+_n(\lambda) ) \ \bigg\vert \ J(\lambda) = k\right] \ \Pr [ J(\lambda)=k] \notag \\ & \leq \sum_{k=0}^N \Pr \left[ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^k {\rm Cap}(\varphi_n, \arccos(1/\sqrt{d}) )\right] \ \Pr [ J(\lambda)=k] \notag \\ & = \sum_{k=0}^N p(k, d-1, \arccos(1/\sqrt{d})) \ \Pr[ J(\lambda) = k] \notag \\ & = \sum_{k=0}^N p(k, d-1, \arccos(1/\sqrt{d})) \binom{N}{k} \left( 1 - \frac{\lambda}{2\delta\sqrt{d}}\right)^{N-k} \left( \frac{\lambda}{2\delta\sqrt{d}} \right)^k \label{thm2-eq2}\end{aligned}$$ The above computations made use of definition and the independence of the $\{\epsilon_n\}_{n=1}^N$ and $\{\varphi_n\}_{n=1}^N$.\ [*Step II.*]{} In this step we bound the integral $\int_0^{2\delta \sqrt{d}} \lambda \Pr[W_N > \lambda] d\lambda$. Equation and properties of the beta function imply $$\begin{aligned} \int_0^{2\delta\sqrt{d}} & \lambda \Pr[W_N > \lambda] d\lambda \notag \\ & \leq \sum_{k=0}^N p(k, d-1, \arccos(1/\sqrt{d})) \binom{N}{k} \int_0^{2\delta\sqrt{d}} \lambda \left( 1 - \frac{\lambda}{2\delta\sqrt{d}}\right)^{N-k} \left( \frac{\lambda}{2\delta\sqrt{d}} \right)^k d \lambda \notag\\ &= \sum_{k=0}^N p(k, d-1, \arccos(1/\sqrt{d})) \binom{N}{k} (4\delta^2 d) \int_0^1 u^{k+1} (1- u)^{N-k} d \lambda \notag \\ &=\sum_{k=0}^N p(k, d-1, \arccos(1/\sqrt{d})) \binom{N}{k} (4\delta^2 d) \frac{(k+1)! (N-k)!}{(N+2)!} \notag \\ &=\frac{4\delta^2 d}{(N+1)(N+2)} \sum_{k=0}^N p(k, d-1, \arccos(1/\sqrt{d})) \thinspace (k+1) \notag \\ &\leq \frac{4\delta^2 d}{(N+1)(N+2)} \left( 1 + 2 \sum_{k=1}^N k \thinspace p(k, d-1, \arccos(1/\sqrt{d})) \right). \label{intsumsqrtd}\end{aligned}$$ [*Step III.*]{} In this step we bound the sum $\sum_{k=1}^N k \thinspace p(k, d-1, \arccos(1/\sqrt{d}))$ appearing in . First note that $$\sum_{k=1}^{60d} k \thinspace p(k, d-1, \arccos(1/\sqrt{d})) \leq \sum_{k=1}^{60d} k = \frac{60d(60d+1)}{2} \leq 3600d^2. \label{thm2-eq3}$$ Next by Lemma \[pbound-lem\] and $$\begin{aligned} \sum_{k=60d+1}^N k \thinspace p(k, d-1, \arccos(1/\sqrt{d})) & \leq \sum_{k=60d+1}^N k \left( 2 \sqrt{d} (13)^d \left( \frac{11}{12} \right)^{k/2} \right)\notag \\ & = 2 \sqrt{d} (13)^d \sum_{k=60d+1}^N k\left( \frac{11}{12} \right)^{k/2}\notag \\ & \leq 2 \sqrt{d} (13)^d \sum_{k=1}^{\infty} (k+60d) \left( \sqrt{\frac{11}{12}} \right)^{k+60d}\notag \\ & \leq 2 \sqrt{d} (120d) \sum_{k=1}^{\infty} k \left( \sqrt{\frac{11}{12}} \right)^{k} \notag \\ & = 2 \sqrt{d} (120d) \left( \frac{\sqrt{132}}{(\sqrt{12} - \sqrt{11})^2} \right) \notag \\ & \leq 12700d^{3/2}. \label{thm2-eq4}\end{aligned}$$ Combining and shows that $$\label{thm2-eq5} \sum_{k=1}^N k \thinspace p(k, d-1, 1/\sqrt{d}) \leq 3600d^2 + 12700d^{3/2}.$$ Now and show that $$\int_0^{2\delta \sqrt{d}} \lambda \Pr[W_N > \lambda] d\lambda \leq \frac{e^{12}\delta^2d^3}{(N+1)(N+2)}. \label{thm2-eq5b}\\$$ [*Step IV.*]{} In this step we bound the integral $\int_{2\delta\sqrt{d}}^{4\delta\sqrt{d}} \lambda \Pr[W_N>\lambda] d \lambda$. Note that if $\lambda \geq 2\delta \sqrt{d}$ then implies that $\theta^+_n(\lambda) \geq \arccos(1/\sqrt{d})$ which, in turn, implies that $$\label{thm2-eq6} {\rm Cap}(\varphi_n, \arccos(1/\sqrt{d})) \subset {\rm Cap}(\varphi_n, \theta^+_n(\lambda))$$ Lemma \[cover-lem\] and equations and imply that $$\begin{aligned} \Pr [W_N>\lambda] &= \Pr [ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N B_n(\lambda)] \\ &\leq \Pr [ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N {\rm Cap}(\varphi_n, \theta^+_n(\lambda))]\\ &\leq \Pr [ \mathbb{S}^{d-1} \not\subset \bigcup_{n=1}^N {\rm Cap}(\varphi_n, \arccos(1/\sqrt{d}))]\\ &= p(N, d-1, \arccos(1/\sqrt{d})).\end{aligned}$$ Thus by Lemma \[pbound-lem\] $$\begin{aligned} \int_{2\delta\sqrt{d}}^{4\delta\sqrt{d}} \lambda \Pr [ W_N > \lambda] d\lambda &\leq p(N, d-1, \arccos(1/\sqrt{d})) \int_{2\delta\sqrt{d}}^{4\delta\sqrt{d}} \lambda d \lambda\notag\\ &\leq 6 \delta^2d \left( 2 \sqrt{d} (13)^d \left(\frac{11}{12} \right)^{N/2} \right)\notag\\ &= 12 \delta^2 d^{3/2} (13)^d \left(\frac{11}{12} \right)^{N/2}.\label{thm2-eq7}\end{aligned}$$ [*Step V.*]{} In this step we bound the integral $\int_{4\delta\sqrt{d}}^{\infty} \lambda \Pr[W_N>\lambda] d\lambda$. For this we specialize the $\epsilon$-net argument used in the proof of Theorem \[main-thm1\] to the case when $\{\varphi_n\}_{n=1}^N$ are i.i.d. uniformly distributed on $\mathbb{S}^{d-1}$. Using and gives $$\begin{aligned} \Pr [W_N>\lambda] &\leq \left( \frac{8}{\epsilon} \right)^{d-1} \sup_{z \in \mathbb{S}^{d-1}} \left( \Pr [ z \notin T_{\epsilon}(B_1(\lambda))] \right)^N\notag\\ & \leq \left( \frac{8}{\epsilon} \right)^{d-1} \sup_{z \in \mathbb{S}^{d-1}} \left( \Pr \left[ |\langle z, \varphi\rangle| \leq \frac{2\delta}{\lambda} + \epsilon \right] \right)^N. \label{thm2-eq8}\end{aligned}$$ Since $\varphi$ is uniformly distributed on $\mathbb{S}^{d-1}$ it follows from Example \[unif-gencond-ex\] and that $$\label{thm2-eq9} \forall z \in \mathbb{S}^{d-1}, \ \ \ \Pr \left[ | \langle z, \varphi \rangle | \leq \frac{3\delta}{\lambda} \right] \leq \frac{3 \delta \sqrt{d}}{\lambda}.$$ Taking $\epsilon = {\delta}/{\lambda}$ in and using shows that $$\begin{aligned} \Pr [W_N>\lambda] &\leq \left( \frac{8\lambda}{\delta} \right)^{d-1} \sup_{z\in\mathbb{S}^{d-1}} \left( \Pr \left[ | \langle z, \varphi \rangle | \leq \frac{3\delta}{\lambda}\right] \right)^N\notag\\ &\leq \left( \frac{8\lambda}{\delta} \right)^{d-1} \left( \frac{3\delta \sqrt{d}}{\lambda} \right)^N. \label{thm2-eq10}\end{aligned}$$ It now follows from that if $N \geq d+2$ then $$\begin{aligned} \int_{4\delta\sqrt{d}}^{\infty} \lambda \Pr [ W_N > \lambda] d\lambda &\leq \int_{4\delta\sqrt{d}}^{\infty} \lambda \left( \frac{8\lambda}{\delta} \right)^{d-1} \left( \frac{3\delta \sqrt{d}}{\lambda} \right)^N d\lambda \notag\\ &= \delta^2 (3\sqrt{d})^N 8^{d-1} \frac{(4\sqrt{d})^{d-N+1}}{N-d-1}\notag\\ &\leq \delta^2\sqrt{d} \left( \frac{3}{4} \right)^N \left( 32\sqrt{d} \right)^d. \label{thm2-eq11}\end{aligned}$$ [*Step VI.*]{} Combining , , and yields $$\begin{aligned} \frac{1}{2}\mathbb{E}|W_N|^2 &\leq \frac{e^{12} \delta^2 d^3}{(N+1)(N+2)} +12 \delta^2 d^{3/2} (13)^d \left(\frac{11}{12} \right)^{N/2} +\delta^2\sqrt{d} \left( \frac{3}{4} \right)^N \left( 32\sqrt{d} \right)^d\\ & \leq \frac{e^{12} \delta^2 d^3}{(N+1)(N+2)} + 13\delta^2 d^{3/2}\left( \frac{11}{12} \right)^{N/2} \left( 32\sqrt{d} \right)^d\\ &=\frac{e^{12} \delta^2 d^3}{(N+1)(N+2)} + 13\delta^2 d^{3/2}\left( \frac{11}{12} \right)^{N/2} e^{\frac{d \ln(1024d)}{2}}.\end{aligned}$$ This completes the proof.\ There exist absolute constants $a,c>0$ such that if $N \geq a d \ln d$ and $d \geq 2$ and if the random vectors $\{\varphi_n\}_{n=1}^N \subset \R^d$ are i.i.d. uniformly distributed on $\mathbb{S}^{d-1}$ then $$\label{cor-bnd-unif} \mathbb{E}|W_N|^2 \leq \frac{c \thinspace d^3 \delta^2}{N^2}.$$ It is instructive to compare the dimension dependence of consistent reconstruction with linear reconstruction. Fix an error tolerance $\eta>0$. Suppose that the random vectors $\{\varphi_n\}_{n=1}^N \subset \mathbb{S}^{d-1}$ are uniformly distributed on $\mathbb{S}^{d-1}$, that $\{\epsilon_n\}_{n=1}^N$ are uniformly distributed on $[-\delta, \delta]$, and that the $\{\varphi_n\}_{n=1}^N$ and $\{\epsilon_n\}_{n=1}^N$ are mutually independent. The bound shows that if $N \geq \frac{d^{3/2}\delta \sqrt{c}}{\eta}$ then the mean squared error for a consistent estimate $\widetilde{x}_{\rm cr}$ satisfies $$\mathbb{E}\|x - \widetilde{x}_{\rm cr} \|^2 \leq \mathbb{E}|W_N|^2 \leq \eta^2.$$ Thus $N=\mathcal{O}(d^{3/2})$ measurements are sufficient for consistent reconstruction to achieve $\eta^2$-precise mean squared error. For linear reconstruction with an arbitrary dual frame, applying with $\sigma = \delta^2/3$ shows that linearly reconstructed $\widetilde{x}_{\rm lin}$ satisfies $$\mathbb{E}\| x - \widetilde{x}_{\rm lin}\|^2 \geq \frac{d^2\delta^2}{3N}.$$ Thus at least $N \geq \frac{d^2\delta^2}{3\eta^2}$ measurements are necessary for linear reconstruction to achieve $\eta^2$-precise mean squared error. Acknowledgments =============== A.M. Powell was supported in part by NSF DMS 1211687 and NSF DMS 0811086, and also gratefully acknowledges the Academia Sinica Institute of Mathematics (Taipei, Taiwan) for its hospitality and support. The authors thank Yaniv Plan, Mark Rudelson, Roman Vershynin, and Elena Yudovina for helpful comments. The authors especially thank Elena Yudovina for a suggestion which led to an improved proof of Theorem \[main-thm1\] that is more precise than an earlier version and that is also more consistent with the proof of Theorem \[mainthm2\]. [1]{} A. Brieden, P. Gritzmann, R. Kannan, V. Klee, L. Lovász and M. Simonivits. Deterministic and randomized polynomial-time approximation of radii. Mathematika, 48 (2001), 63–105. P. Bürgisser, F. Cucker and M. Lotz. Coverage processes on spheres and condition numbers for linear programming. The Annals of Probability, [38]{} (2010), 570–604. Z. Cvetkovi[ć]{} and M. Vetterli. On simple oversampled [A/D]{} conversion in $L^2(\mathbb{R})$. IEEE Transactions on Information Theory, [47]{} (2001), no. 1, 146–154. Z. Cvetkovi[ć]{}. Resilience properties of redundant expansions under additive noise and quantization. IEEE Transactions on Information Theory, [49]{} (2003), no. 3, 644–656. T.S. Ferguson. A Course in Large Sample Theory. Chapman & Hall, London, 1996. L. Flatto and D.J. Newman. Random coverings. Acta Mathematica, 138 (1977), 241–264. V.K. Goyal, J. Kova[č]{}evi[ć]{} and J.A. Kelner. Quantized frame expansions with erasures. Applied and Computational Harmonic Analysis, 10 (2001), no. 3., 203–233. V.K. Goyal, M. Vetterli and N.T. Thao. Quantized overcomplete expansions in $\RR^n$: analysis, synthesis, and algorithms. IEEE Transactions on Information Theory, 44 (1998), no.1, 16–30. L. Jacques, D.K. Hammond and J.M. Fadili. Dequantizing compressed sensing: when oversampling and non-[G]{}aussian constraints combine. IEEE Transactions on Information Theory, [57]{} (2011), no. 1, 560–571. S. Janson. Random coverings in several dimensions. Acta Mathematica, 156 (1986), 83-118. D. Jimenez, L. Wang and Y. Wang. White noise hypothesis for uniform quantization errors. SIAM Journal on Mathematical Analysis, 38 (2007), no. 6, 2042–2056. I. Jovanovi[ć]{} and B. Beferull-Lozano. Oversampled [A/D]{} conversion and error-rate dependence of nonbandlimited signals with finite rate of innovation. IEEE Transactions on Signal Processing, [54]{} (2006), no. 6, 2140–2154. A.M. Powell. Mean squared error bounds for the Rangan-Goyal soft thresholding algorithm. Applied and Computational Harmonic Analysis, 29 (2010), no. 3, 251–271. S. Rangan and V.K. Goyal. Recursive consistent estimation with bounded noise. IEEE Transactions on Information Theory, 47 (2001), no. 1, 457–464. A.F. Siegel and L. Holst. Covering the circle with random arcs of random sizes. Journal of Applied Probability, 19 (1982), no. 2, 373–381. H. Solomon. Geometric Probability. Society for Industrial and Applied Mathematics, Philadelphia, 1978. A.J. Stam. Limit theorems for uniform distributions on spheres in high-dimensional Euclidean spaces. Journal of Applied Probability, 19 (1982), no. 1, 221-228. N.T. Thao and M. Vetterli. Deterministic analysis of oversampled [A/D]{} conversion and decoding improvement based on consistent estimates. IEEE Transactions on Signal Processing, 42 (1994), no. 3, 519–531. N.T. Thao and M. Vetterli. Reduction of the MSE in $R$-times oversampled A/D Conversion from $\mathcal{O}(R^{-1})$ to $\mathcal{O}(R^{-2})$. IEEE Transactions on Signal Processing, 42 (1994), 200–203. N.T. Thao and M. Vetterli. Lower bound on the MSE in over sampled quantization of periodic signals using vector quantization analysis. IEEE Transactions on Information Theory, 42 (1996), 469–479. J.G. Wendel. Note on the gamma function. American Mathematical Monthly, 55 (1946), no. 9, 563–564.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Traditional optimal design of experiment theory is developed on Euclidean space. In this paper, new theoretical results of optimal design of experiments on Riemannian manifolds are provided. In particular, it is shown that D-optimal and G-optimal designs are equivalent on manifolds and provide a lower bound for the maximum prediction variance. In addition, a converging algorithm that finds the optimal experimental design on manifold data is proposed. Numerical experiments demonstrate the competitive performance of the new algorithm.' author: - | Hang Li[^1]\ Department of Industrial and Manufacturing Engineering,\ Pennsylvania State University\ and\ Enrique Del Castillo\ Department of Industrial and Manufacturing Engineering,\ Pennsylvania State University\ and\ George Runger\ Department of Biomedical Informatics,\ Arizona State University bibliography: - 'DOE.bib' title: '**Optimal Design of Experiments on Riemannian Manifolds**' --- \#1 1 [1]{} 0 [1]{} [**Optimal Design of Experiments on Riemannian Manifolds**]{} [*Keywords:*]{} Manifold learning, active learning, high-dimensional data analysis, regularization. Introduction {#sec:intro} ============ Supervised learning models typically need to be trained on large amounts of labeled instances to perform well. While many modern systems can easily produce a large number of unlabeled instances at low cost, the labeling process can be very difficult, expensive or time-consuming. However, different labeled instances contain various information and contribute to the learning process in different ways. Therefore, an interesting question arises: how to choose the most informative instances to label so that one can improve learning rate of the model and reduce the labeling cost at the same time? In statistics, the problem of selecting which instances to label is referred as *Design of Experiments* (**DOE**) ([@Federov; @box2005statistics; @EDC_RSMbook]). DOE is a systematic method to explore the relationship between process input (or experimental “factors") and output responses under limited resources for conducting experiments. Traditional DOE was developed for physical experiments in agricultural and industrial applications where the goal is to optimize some continuous function of the covariates. A classic theory of experimental design exists for linear statistical models that assume the response is a function of a small number of covariates (for a summary of this theory, see, e.g., [@Federov]). In these problems, the number of covariates or “factors" of interest in an experiment is relatively small and so is the number of tests or size of the experiment, given the high experimental cost. However, with the development of modern technology, scientists and engineers frequently face different challenges arising not only from higher dimensional data but also from more complex data structures. Furthermore, in experiments where the response is an image that needs to be classified or text that needs to be recognized and categorized, the dimension of each data instance is much higher than those typically dealt with in the classical DOE literature. A particular type of data complexity specially important in experiments with image and text data occurs when the problem data actually lies on an unknown manifold of much smaller dimension of the space in which it appears to reside. The goal of this paper is to discuss a new methodology for designing optimal experimental designs that minimize the number of experimental runs for high-dimensional manifold data and at the same time acquire as much useful information about the response as possible. As far as we know, no existing work has provided a theoretical justification of optimal experimental design methods for high-dimensional manifold data. This paper contributes to the theoretical development of DOE methods on manifolds. In particular, we prove a new Equivalence Theorem for a continuous optimal design on a Riemannian manifolds, and also provide a converging algorithm for finding the optimal design. The rest of this paper is organized as follows. In Section \[sec:ODOE\], we briefly introduce the traditional *Optimal Design of Experiment* (**ODOE**) on Euclidean space, and then explain the idea of manifold learning, in particular the *manifold regularization* model from [@BNS06JMLR]. More importantly, a manifold-based ODOE scheme is discussed. In Section \[sec:theory\], we provide the theoretical justification of optimal experimental design and present a new equivalence theorem of ODOE on Riemannian manifolds. In Section \[sec:alg\], we illustrate our proposed algorithm and provide a convergence analysis. In Section \[sec:numerical\], several numerical experiments are conducted to demonstrate the effectiveness of the proposed algorithm for finding optimal designs on manifold data. In Section \[sec:conc\], we conclude this paper and discuss some possible future directions. Optimal Design of Experiments on Manifolds {#sec:ODOE} ========================================== Traditional ODOE on Euclidean Space ----------------------------------- Consider initially a classical linear regression model $$\label{eqn:lr} y = f({x},\beta) + {\varepsilon}\; = \; \beta^\top g({x}) + {\varepsilon},$$ where $g: \mathbb{R}^d \rightarrow \mathbb{R}^p$ is some nonlinear mapping function that maps from the input space ${x}\in \mathbb{R}^d$ to the feature space $\mathbb{R}^p$, $\beta \in \mathbb{R}^p$ is a column vector of unknown parameters, and ${\varepsilon}$ is assumed to have a $N(0,\sigma^2)$ distribution . Given a sample of $n$ design points $\{{x}_i\}_{i=1}^n$, if the corresponding response values $\{y_i\}_{i=1}^n$ are available, the well-known ordinary least squares estimates of the $\beta$ parameters are given by: $$\label{OLS_est} \hat \beta= \operatornamewithlimits{argmin}_{\beta\in \mathbb{R}^p} \left\{ \sum_{i=1}^n ( y_i - \beta^\top g({x}_i))^2 \right\} \; = \; (X^\top X)^{-1} X^\top Y$$ where $X$ is a $n \times p$ design matrix with $i$-th row defined as $g({x}_i)^\top$, and $Y$ is a $n \times 1$ response vector. As a result, the corresponding fitted function is $\hat f({x}) = \hat \beta^\top g({x})$. Classical work on *Optimal Design of Experiments* (**ODOE**) was developed by [@kiefer_wolfowitz_1960] and summarized by [@F1972TOE] . The goal of ODOE is to find an experimental design that is optimal with respect to some statistical criterion related either to the model parameter estimates or to the model predictions. For example, given the linear regression model (\[eqn:lr\]), the *D-optimality* criterion minimizes the determinant of the covariance matrix of the parameter estimates $\operatorname{Var}(\hat \beta)= \sigma^2 (X^\top X)^{-1}$, while the *G-optimality* criterion minimizes the maximum prediction variance $\operatornamewithlimits{max}_{i=1,...,n}\big\{\operatorname{Var}(\hat{y}_i) \big\}$. These and similar criteria are called “*alphabetic optimality*" design criteria by [@box2007response]. Traditional ODOE methods assume both the covariate and response data lie on an Euclidean space under the classical linear regression model (\[OLS\_est\]). However, in applications to high-dimensional image and text datasets, traditional ODOE is not applicable as it does not consider the intrinsic manifold structure these type of datasets usually have ([@TSL00S; @RS00S]). While there have been recent attempts at applying alphabetic optimality criteria to manifold learning models ([@H10IEEETIP; @CCBWZZ2010; @Alaeddini2019IISE]), no theoretical justification exists, as far as we know, to these methods, and no guarantees can be given for their success other than empirical experiments. A new theory for the optimal experimental designs is therefore needed that explicitly considers high-dimensional manifold data, justify existing methods if possible, and shows a principled way to develop new methods. Before we discuss the design of experiments on manifolds, first we need to introduce a manifold learning model by [@BNS06JMLR] that will be used in the sequel. Manifold Regularization Model ----------------------------- In the well-known paradigm of machine learning, the process of learning is seen as using the training data $\{{x}_i\}_{i=1}^{n}$ to construct a function $f: \mathcal{X} \rightarrow \mathbb{R}$ that maps a data instance ${x}$ to a label variable $y$. Let $P$ be the joint distribution that generates labeled data $\{({x}_i,y_i)\}_{i=1}^{l} \subset \mathcal{X}\times \mathbb{R}$ and $P_\mathcal{X}$ be the marginal distribution that generates unlabeled data [$\{{x}_i\}_{i=l+1}^{n} \subset \mathcal{X} \subset \mathbb{R}^d$]{}. In order to extend the learning of functions to nonlinear manifolds, [@BNS06JMLR] assume that the conditional distribution $P(y|{x})$ varies smoothly as ${x}$ moves along a manifold that supports $P_\mathcal{X}$. In other words, if two data points ${x}_1, {x}_2 \in \mathcal{X}$ are close as measured by an intrinsic (or geodesic) distance on this manifold, then the two probabilities of the labels, $P(y|{x}_1)$ and $P(y|{x}_2)$, will be similar. These authors developed a semi-supervised learning framework that involves solving the following double regularized objective function: $$\label{eqn:obj1} \hat{f}=\operatornamewithlimits{argmin}\limits_{f\in \mathcal{H}_\mathcal{K}} \left\{ \sum_{i=1}^{l} V({x}_i,y_i,f)+\lambda_A \|f\|_{\mathcal{H}_\mathcal{K}}^2+\lambda_I \|f\|_I^2 \right\}$$ where $V$ is a given loss function (such as squared loss $(y_i-f({x}_i))^2$), $\mathcal{H}_\mathcal{K}$ is a *Reproducing Kernel Hilbert Space* (**RKHS**) ([@A50TAMS]) with associated Mercer kernel $\mathcal{K}$, $ \| f \|_{\mathcal{H}_\mathcal{K}}^2$ is a penalty term with the norm equipped in $\mathcal{H}_\mathcal{K}$ that imposes smoothness conditions in the ambient space ([@W90SM]), and $\|f\|_I^2$ is a penalty term for non-smoothness along geodesics on the intrinsic manifold structure of $P_\mathcal{X}$. Moreover, $\lambda_A$ and $\lambda_I$ are two regularization parameters that control the amount of penalization in the ambient space and in the intrinsic manifold that supports $P_\mathcal{X}$, respectively. Some other work in recent literature can be explained as a particular case of this general framework. For example, the spatial regression model proposed by [@ESP2016Bio] can be seen as the manifold regularization model (\[eqn:obj1\]) without the ambient space regularization. There are also different work about nonparametric regression models on manifolds ([@CWJASA2013; @MPT2014JASA; @LTZD2017JASA]), but in this paper we focus on the manifold regularization model from [@BNS06JMLR]. Intuitively, the choice of $\|f\|^2_I$ should be a smoothness penalty corresponding to the probability distribution $P_\mathcal{X}$. However, in most real-world applications $P_\mathcal{X}$ is not known, and therefore empirical estimates of the marginal distribution must be used. Considerable research has been devoted to the case when $P_\mathcal{X}$ is supported on a compact manifold $\mathcal{M} \subset \mathbb{R}^d$ ([@RS00S; @TSL00S; @BN03NC; @DG03PNAS; @CLLMNWZ05PNAS]). Under this assumption, it can be shown that problem (\[eqn:obj1\]) can be reduced to $$\begin{aligned} \label{eqn:mrobj2} \hat f = \operatornamewithlimits{argmin}\limits_{f\in \mathcal{H}_\mathcal{K}} \left\{ \sum_{i=1}^{l} V({x}_i,y_i,f)+\lambda_A \|f\|_{\mathcal{H}_\mathcal{K}}^2+\lambda_I \mathbf{f}^\top L \mathbf{f} \right\}\end{aligned}$$ where $\mathbf{f}=[f({x}_1),...,f({x}_{n})]^\top$ and $L$ is the Laplacian matrix associated with the data adjacency graph $\mathcal{G}$ that is constructed on all the labeled and the unlabeled data points $\{{x}_i\}_{i=1}^{n}$. In particular, the graph Laplacian $L$ approximates the *Laplace-Beltrami* operator acting on the continuous Riemannian manifold $\mathcal{M}$ (see [@B03UCTHESIS; @L04YALETHESIS; @BN05COLT; @CLLMNWZ05PNAS; @HAL05COLT]). In this way, Belkin et al. provide a theoretical justification to the common trick in manifold learning of using a graph and geodesic distances on the graph as an approximate representation of the manifold $\mathcal{M}$, providing a precise sense in which the graph approaches $\mathcal{M}$ as the number of data points gets denser. This way, the term $\mathbf{f}^\top L \mathbf{f}$ serves as an approximation for $\|f\|_I^2 $, and enforces the penalization on the lack of smoothness of $f$ as it varies between adjacent points in the graph $\mathcal{G}$. In addition, [@BNS06JMLR] proceed to prove a representer theorem (similar to the theory of splines in [@W90SM]), which shows how the solution of the infinite dimensional problem (\[eqn:mrobj2\]) can be represented in terms of a finite sum over the labeled and unlabeled points: $$\begin{aligned} \label{representer} f({x}) = \sum_{i=1}^{n} \alpha_i \mathcal{K}({x}_i,{x})\end{aligned}$$ where $\mathcal{K}(\cdot,\cdot)$ is the Mercer kernel associated with the ambient space $\mathcal{H}_\mathcal{K}$. More details about the manifold regularized model can be found in [@BNS06JMLR]. Regularized ODOE on Manifolds ----------------------------- As far as we know, the first discussion of regularized ODOE comes from [@Vuchkov1977], who proposed a ridge-type procedure for ODOE based on the ridge regression estimator: $$\label{ridge} \hat \beta_{\text{ridge}} = \operatornamewithlimits{argmin}_{\beta\in \mathbb{R}^p} \left\{ \sum_{i=1}^l ( y_i - \beta^\top g({x}_i))^2 + \lambda_{\text{ridge}} \| \beta\|^2 \right\}$$ Vuchkov’s motivation was to use the ridge estimator to solve the singular or ill-conditional problems that might exist in the sequential D-optimal design algorithm when the number of design points is smaller than the number of parameters to estimate. The ridge solution (\[ridge\]) can be seen as a particular case of the more general learning problem (\[eqn:mrobj2\]) where $V$ is a squared-loss function, and the RKHS $\mathcal{H}_K$ is equipped with a $L^2$-norm and the manifold regularization parameter $\lambda_I=0$. To discuss the optimal experimental design for the general manifold regularization model (\[eqn:mrobj2\]), we first need to clarify some notation. Without loss of generality, assume a sequential experimental design generation, starting with no labeled data at the beginning of the sequence. Let $\{{z}_i\}_{i=1}^k \subset \{{x}_i\}_{i=1}^n$ be the set of points that has been labeled at the $k$-th iteration, and $\textbf{y}=(y_1,...,y_k)^\top$ be the corresponding label vector. Given a square loss function, the manifold regularization model (\[eqn:mrobj2\]) becomes the *Laplacian Regularized Least Squares* (**LapRLS**) problem: $$\label{eqn:LapRLS} \hat f = \operatornamewithlimits{argmin}\limits_{f\in \mathcal{H}_K} \left\{ \sum_{i=1}^{k}(y_i-f({z}_i))^2+\lambda_A \|f\|_{\mathcal{H}_\mathcal{K}}^2+\lambda_I \textbf{f}^\top L \textbf{f} \right\}.$$ Substituting the representer theorem (\[representer\]) to (\[eqn:LapRLS\]), we get a convex differentiable objective function with respect to $\alpha$: $$\label{eqn:obj_alpha} \hat{\alpha} = \operatornamewithlimits{argmin}\limits_{\alpha\in \mathbb{R}^n} \left\{ ({{\bf y}}-K_{XZ}^\top \alpha)^\top ({{\bf y}}-K_{XZ}^\top \alpha) +\lambda_A \alpha^\top K \alpha +\lambda_I \alpha^\top KLK \alpha \right\},$$ where $K_{XZ}$ and $K$ are the Gram matrix defined by $$\begin{aligned} K_{XZ}= \begin{bmatrix} \mathcal{K}({x}_1,{z}_1) & ... & \mathcal{K}({x}_1,{z}_k)\\ \vdots & \ddots & \vdots \\ \mathcal{K}({x}_n,{z}_1) & ... & \mathcal{K}({x}_n,{z}_k) \end{bmatrix}_{n \times k}, \; K_{XX}= \begin{bmatrix} \mathcal{K}({x}_1,{x}_1) & ... & \mathcal{K}({x}_1,{x}_n)\\ \vdots & \ddots & \vdots \\ \mathcal{K}({x}_n,{x}_1) & ... & \mathcal{K}({x}_n,{x}_n) \end{bmatrix}_{n \times n},\end{aligned}$$ and $\mathcal{K}$ is the kernel embedded in the RKHS $\mathcal{H}_\mathcal{K}$. Taking a derivative of (\[eqn:obj\_alpha\]) with respect to $\alpha$ and making it equal to 0, we arrive at the following expression: $$\begin{aligned} \hat{\alpha} = (K_{XZ}K_{XZ}^\top +\lambda_A K + \lambda_I KLK)^{-1} K_{XZ}\mathbf{y}\end{aligned}$$ Consider a linear regression model form (\[eqn:lr\]) and a linear kernel for $\mathcal{H}_K$, the regression parameters $\beta$ can be estimated by $$\begin{aligned} \label{beta_est1} \hat{\beta}=X^\top \hat{\alpha} = X^\top (XZ_k^\top Z_k X^\top+\lambda_A XX^\top + \lambda_I XX^\top L XX^\top)^{-1}XZ^\top \mathbf{y}\end{aligned}$$ where $$\begin{aligned} Z_k=\left [ \begin{array}{c} g({z}_1)^\top\\ \vdots \\ g({z}_k)^\top \end{array} \right], \; X=\left [ \begin{array}{c} g({x}_1)^\top\\ \vdots \\ g({x}_n)^\top \end{array} \right], \; \mathbf{y}=\left [ \begin{array}{c} y_1\\ \vdots \\ y_k \end{array} \right].\end{aligned}$$ By some simple linear algebra (a formal proof is provided in the Appendix), the estimated parameters $\hat{\beta}$ (\[beta\_est1\]) can be simplified to $$\begin{aligned} \label{beta_est2} \hat{\beta}=(Z_k^\top Z_k + \lambda_A I_p + \lambda_I X^\top LX)^{-1} Z^\top \mathbf{y}\end{aligned}$$ [@H10IEEETIP] demonstrated that the covariance matrix of (\[beta\_est2\]) can be approximated as: $$\label{eqn:cov_appr} \mathrm{Cov} (\mathbf{\hat{\beta}}) \approx \sigma^2 (Z_k^\top Z_k +\lambda_A I_p + \lambda_I X^\top L X)^{-1}.$$ The determinant of covariance matrix (\[eqn:cov\_appr\]) is the statistical criterion we will minimize to obtain a D-optimal design. Before we discuss the optimal design algorithm, first we will provide some theoretical justification of ODOE on Riemannian manifolds in the following section. Theoretical Results {#sec:theory} =================== When the determinant of $Z_k^\top Z_k + \lambda_A I_p + \lambda_I X^\top L X$ is maximized, one obtains a so-called [*D-optimal*]{} experimental design. In Euclidean space, “continuous" or “exact" optimal design theory (which considers the proportion of experimental tests allocated to different locations over the space) indicates the equivalence between the D-optimality criteria and the so-called *G-optimality* criteria, where the maximum prediction variance is minimized, as stated by the celebrated *Kiefer-Wolfowitz* (**KW**) theorem ([@kiefer_wolfowitz_1960; @Kiefer1974]). In analogy with the KW theorem, we aim to develop a new equivalence result for optimal experimental design based on the manifold regularization model (\[eqn:mrobj2\]), which can then be used for designing an experiment on a Riemannian manifold. Assume there is an infinite number of points $x$ that are uniformly distributed on a Riemannian manifold $\mathcal{M}$. Let $\epsilon$ be a continuous design on $\mathcal{M}$. For any continuous design $\epsilon$, based on the *Carathéodory Theorem*, it’s known that $\epsilon$ can be represented as $$\begin{aligned} \epsilon =\left \{ \begin{array}{cccc} z_1, z_2, ..., z_{n_0}\\ q_1,q_2,...,q_{n_0} \end{array} \right\}, \; \mathrm{where} \; \sum_{i=1}^{n_0} q_i=1.\end{aligned}$$ For any $\epsilon$, the corresponding information matrix of LapRLS model is defined as $$\label{Mcont} M_{Lap}(\epsilon)=\int_{z \in \mathcal{X}} \xi(z)g(z)g(z)^\top dz + \lambda_A I_p + \lambda_I \int_{x \in \mathcal{M}} g(x) \Delta_{ \mathcal{M}} g(x)^\top d\mu,$$ where $\xi$ is a probability measure of design $\epsilon$ on the experimental region $\mathcal{X}\subseteq \mathcal{M} \subset \mathbb{R}^p$, $\Delta_{ \mathcal{M}}$ is the *Laplace-Beltrami* operator on $\mathcal{M}$, and $\mu$ is the uniform measure on $\mathcal{M}$. Note that the last two terms in (\[Mcont\]) are independent of the design $\epsilon$, thus for simplicity, define $$C = \lambda_A I_p + \lambda_I \int_{x \in \mathcal{M}} g(x) \Delta_{ \mathcal{M}} g(x)^\top d\mu.$$ Then (\[Mcont\]) can be written as $$M_{Lap}(\epsilon)=\int_{z \in \mathcal{X}} \xi(z)g(z)g(z)^\top dz + C.$$ Based on the parameters estimates (\[beta\_est2\]), for a given continuous design $\epsilon$, the prediction variance at a test point $z$ is $$\begin{aligned} d(z,\epsilon)=\mathrm{Var}\Big[\mathbf{\hat{\beta}}^\top g(z) \Big]=g(z)^\top \mathrm{Cov}(\mathbf{\hat{\beta}}) g(z) = \sigma^2 g(z)^\top M^{-1}_{Lap}(\epsilon) g(z)\end{aligned}$$ As it can be seen, under the LapRLS model one can obtain a D-optimal design by maximizing the determinant of $M_{Lap}(\epsilon)$ and a G-optimal design by minimizing $\operatornamewithlimits{max}\limits_{z \in \mathcal{X}} d(z,\epsilon)$. Similarly to the optimal design of experiments in Euclidean space, we prove next an equivalence theorem on Riemannian manifolds that shows how the D and G optimality criteria lead to the same optimal design. Before the equivalence theorem is discussed, we need to prove some auxiliary results. The proofs of these proposition are provided in the Appendix. \[prop1\] Let $\epsilon_1$ and $\epsilon_2$ be two designs with the corresponding information matrices $M_{Lap}(\epsilon_1)$ and $M_{Lap}(\epsilon_2)$. Then $$M_{Lap}(\epsilon_{3})=(1-\alpha) M_{Lap} (\epsilon_1) + \alpha M_{Lap} (\epsilon_2),$$ where $M_{Lap}(\epsilon_3)$ is the information matrix of the design $$\epsilon_{3}=(1-\alpha)\epsilon_{1}+\alpha \epsilon_2.$$ \[prop2\] Let $\epsilon_1$ and $\epsilon_2$ be two designs with the corresponding information matrices $M_{Lap}(\epsilon_1)$ and $M_{Lap}(\epsilon_2)$. Then $$\frac{d \log |M_{Lap}(\epsilon_3)|}{d \alpha}=\operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon_3)\big [M_{Lap}(\epsilon_2)-M_{Lap}(\epsilon_1)\big] \Big\},$$ where $M_{Lap}(\epsilon_3)$ is the information matrix of the design $$\epsilon_3=(1-\alpha)\epsilon_1+\alpha\epsilon_2.$$ \[prop3\] For any continuous design $\epsilon$, 1. $$\label{dint} \int_{z\in \mathcal{X}} d (z,\epsilon) \xi(z) dz=p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \}$$ 2. $$\label{maxd_ineq} \max_{z \in \mathcal{X}} d(z,\epsilon) \geq p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \}$$ \[prop4\] The function $\log |M_{Lap}(\epsilon)|$ is a strictly concave function. Based on Propositions \[prop1\]-\[prop4\], we can now prove the equivalence theorem for the LapRLS model. In summary, the following theorem demonstrates that the D-optimal design and G-optimal design are equivalent on the Riemannian manifold $\mathcal{M}$. It also provides the theoretical value of maximum prediction variance of the LapRLS model when the D/G optimal design is achieved. \[Thm1\] The following statements are equivalent: 1. the design $\epsilon^*$ maximizes $\det(M_{Lap}(\epsilon))$ 2. the design $\epsilon^*$ minimizes $\operatornamewithlimits{max}\limits_{z \in \mathcal{X}} d(z,\epsilon)$ 3. $\operatornamewithlimits{max}\limits_{z \in \mathcal{X}} d(z,\epsilon^*)=p-\operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon^*) C \Big \}$ **Proof** **(1)** **1 $\Rightarrow$ 2** Let $\epsilon^*$ be the design that maximizes $| M_{Lap}(\epsilon)|$ and define $\tilde{\epsilon}=(1-\alpha)\epsilon^*+\alpha \epsilon$, where $\epsilon$ is some arbitrary design. According to Proposition \[prop2\], we have that $$\begin{aligned} \frac{d \log |M_{Lap}(\tilde{\epsilon})|}{d \alpha} &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\tilde{\epsilon})\big [M_{Lap}(\epsilon)-M_{Lap}(\epsilon^*)\big] \Big\}\end{aligned}$$ When $\alpha=0$, we have $\tilde{\epsilon}=\epsilon^*$. Thus $$\begin{aligned} \frac{d \log |M_{Lap}(\tilde{\epsilon})|}{d \alpha} \bigg |_{\alpha=0} &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*)\big [M_{Lap}(\epsilon)-M_{Lap}(\epsilon^*)\big] \Big\} \\ &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) M_{Lap}(\epsilon) \Big\}-\operatorname{Tr}(I_p) \\ &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) M_{Lap}(\epsilon) \Big\}- p\end{aligned}$$ Since $\epsilon^*$ is the maximal solution, then $$\begin{aligned} \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) M_{Lap}(\epsilon) \Big\}- p \leq 0.\end{aligned}$$ Without loss of generality, assume the design $\epsilon$ has only one instance $z \in \mathcal{X}$. Then we have $$\begin{aligned} M_{Lap}(\epsilon)&=&g(z)g(z)^\top+C \end{aligned}$$ and $$\begin{aligned} \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) \big [g(z)g(z)^\top+C \big] \Big\}- p &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) g(z)g(z)^\top \Big\}+ \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} - p\\ &=& d (z,\epsilon^*) + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} - p \\ &\leq& 0 \label{Mderivative}\end{aligned}$$ Thus $$\label{dleq} d (z,\epsilon^*) \leq p - \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}$$ In addition, based on Proposition \[prop3\], we have $$\label{dgeq} \max_{z \in \mathcal{X}} d(z,\epsilon^*) \geq p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon^*) C \Big \}$$ Combining (\[dleq\]) and (\[dgeq\]), we can conclude that the D-optimal design $\epsilon^*$ minimizes $\operatornamewithlimits{max}\limits_{z \in \mathcal{X}} d(z,\epsilon)$. **(2)** **2 $\Rightarrow$ 1** Let $\epsilon^*$ be the design that minimizes $\operatornamewithlimits{max}\limits_{z \in \mathcal{X}} d(z,\epsilon)$, but assume it is not D-optimal. Based on Proposition \[prop4\], we know there must exist a design $\epsilon$ such that: $$\begin{aligned} \label{increaseD} \frac{d \log |(1-\alpha) M_{Lap}(\epsilon^*) + \alpha M_{Lap}(\epsilon)|}{d \alpha} \bigg |_{\alpha=0} &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) M_{Lap}(\epsilon) \Big\}- p >0\end{aligned}$$ where $$\begin{aligned} M_{Lap}(\epsilon) = \int_{z \in \mathcal{X}} \xi(z) g(z)g(z)^\top dz +C.\end{aligned}$$ Then $$\begin{aligned} &&\operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) M_{Lap}(\epsilon) \Big\}- p \\ &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) \big[\int_{z \in \mathcal{X}} \xi(z) g(z)g(z)^\top dz +C \big] \Big\}- p \\ &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) \int_{z \in \mathcal{X}} \xi(z) g(z)g(z)^\top dz \Big\} + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \\ &=& \operatorname{Tr}\Big\{ \int_{z \in \mathcal{X}} M_{Lap}^{-1}(\epsilon^*) \xi(z) g(z)g(z)^\top dz \Big\} + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \\ &=& \int_{z \in \mathcal{X}} \xi(z) \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) g(z_i)g(z_i)^\top \Big\} dz + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \\ &=& \int_{z \in \mathcal{X}} \xi(z) \operatorname{Tr}\Big\{ g(z)^\top M_{Lap}^{-1}(\epsilon^*) g(z) \Big\} dz + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \\ &=& \int_{z \in \mathcal{X}} \xi(z) d(z,\epsilon^*) dz + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \end{aligned}$$ Since $\epsilon^*$ is the design that minimizes $\operatornamewithlimits{max}\limits_{z \in \mathcal{X}}d(z,\epsilon)$, by Proposition \[prop3\], we have $$\begin{aligned} \max_{z \in \mathcal{X}} d(z,\epsilon^*)= p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} \end{aligned}$$ Thus, for any $z \in \mathcal{X}$, $$\begin{aligned} d(z,\epsilon^*) &\leq& p - \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} \end{aligned}$$ Then $$\begin{aligned} \int_{z \in \mathcal{X}} \xi(z) d(z,\epsilon^*) dz &\leq& \int_{z \in \mathcal{X}} \xi(z) \Bigg(p-\operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} \Bigg) dz \\ &=&\Bigg(p-\operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} \Bigg) \int_{z \in \mathcal{X}}\xi(z) dz \\ &=& p-\operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}\end{aligned}$$ Therefore, we have $$\begin{aligned} \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) M_{Lap}(\epsilon) \Big\}- p &=& \int_{z \in \mathcal{X}} \xi(z) d(z,\epsilon^*) dz + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \\ &\leq& p - \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\}- p \\ &=& 0.\end{aligned}$$ This contradicts with (\[increaseD\]). Therefore, the design $\epsilon^*$ is also D-optimal. **(3) 1 $\Rightarrow$ 3** Let $\epsilon^*$ be the D-optimal design. From the previous proof, in particular Equation (\[dleq\]), we know that $$\begin{aligned} \max_{z\in \mathcal{X}} d(z,\epsilon^*) = p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \}. \end{aligned}$$ **(4) 3 $\Rightarrow$ 1** Let $\epsilon^*$ be the design such that $$\begin{aligned} \max_{z\in \mathcal{X}} d(z,\epsilon^*) = p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \}. \end{aligned}$$ Then, for any $z \in \mathcal{X}$, $$\label{eq65} d (z,\epsilon^*) + \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon^*) C \Big\} - p \leq 0.$$ Based on the previous proof, we know that equation (\[eq65\]) implies that there is no improving direction for the D-optimal criteria. Thus $\epsilon^*$ is the D-optimal design.\ **(5)** Since 1 $\Leftrightarrow$ 2, 1 $\Leftrightarrow$ 3, then **2 $\Leftrightarrow$ 3** and the equivalence theoreme is proved. Different from the classical equivalence theorem on Euclidean space, Theorem \[Thm1\] demonstrates the equivalence of D-optimal design and G-optimal design on the Riemannian manifold. In addition, for any given design $\epsilon$, Equation (\[maxd\_ineq\]) provides a new lower bound for the maximum prediction variance. Theorem \[Thm1\] shows that this lower bound (\[maxd\_ineq\]) can be achieved at the D/G optimal design $\epsilon^*$. Therefore, Theorem \[Thm1\] also provides a theoretical justification that the optimal D/G design $\epsilon^*$ would minimize the maximum prediction variance of the model. Proposed Algorithm and Convergence Analysis {#sec:alg} =========================================== Before we discuss our proposed algorithm for finding optimal experimental design on manifolds, some auxiliary results need to be derived. \[prop5\] Let $M_{Lap}(\epsilon_k)$ be the information matrix of the design $\epsilon_k$ at $k$-th iteration. Let $M_{Lap}(\epsilon(z))$ be the information matrix of the design concentrated at one single point $z$. Given $\epsilon_{k+1}=(1-\alpha)\epsilon_{k}+\alpha \epsilon(z)$, then $$\begin{aligned} \label{eqprop4} |M_{Lap}(\epsilon_{k+1})|=(1-\alpha)^p \Big|M_{Lap}(\epsilon_k) \Big| \Big[1+\frac{\alpha}{1-\alpha} d(z,\epsilon_k)+\frac{\alpha}{1-\alpha} \operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C) \Big] \end{aligned}$$ \[prop6\] Let $M_{Lap}(\epsilon_k)$ be the information matrix of the design $\epsilon_k$ at $k$-th iteration. Construct the design $\epsilon_{k+1}$ at $(k+1)$-th iteration as $$\begin{aligned} \epsilon_{k+1}=(1-\alpha_k)\epsilon_k+\alpha_k \epsilon(z_{k+1})\end{aligned}$$ where $$\begin{aligned} 0 \leq \alpha_k \leq \frac{d(z_{k+1},\epsilon_k)-(p-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C))}{p[d(z_{k+1},\epsilon_k)-(1-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C))]}, \; \; z_{k+1}=\operatornamewithlimits{argmax}\limits_{z \in \mathcal{X}} d (z, \epsilon_k).\end{aligned}$$ Then the resulting sequence $\Big\{|M_{Lap}(\epsilon_k)|\Big\}_k$ is non-decreasing. Based on Proposition \[prop5\] and \[prop6\], we propose a new algorithm to find the D-G optimal experimental design on a manifold, as shown in Algorithm 1. Note how after obtaining an optimal design for the data to be labeled, and obtaining the corresponding labels, we use both labeled [*and*]{} unlabeled instances to train the manifold regularized model. We next provide a convergence analysis of the proposed algorithm. **Input:** Some initial design $\epsilon_k$, $$\begin{aligned} \epsilon_k=\left \{ \begin{array}{cccc} z_1, z_2, ..., z_k\\ q_1,q_2,...,q_k \end{array} \right\}, \; \mathrm{where} \; \sum_{i=1}^k q_i=1\end{aligned}$$ Compute the information matrix $$\begin{aligned} M_{Lap}(\epsilon_k)=\sum_{i=1}^k q_i g(z_i)g(z_i)^\top +C\end{aligned}$$ 1. Find $z_{k+1}$ s.t. $$\begin{aligned} z_{k+1}=\operatornamewithlimits{argmax}\limits_{z \in \mathcal{X}} d(z,\epsilon_k) \end{aligned}$$ 2. Update the design $$\begin{aligned} \epsilon_{k+1}=(1-\alpha_k)\epsilon_k+\alpha_k \epsilon(z_{k+1})\end{aligned}$$ where $\alpha_k$ is a user choice that satisfies $$\begin{aligned} 0 \leq \alpha_k \leq \frac{d(z_{k+1},\epsilon_k)-[p-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)]}{p\{d(z_{k+1},\epsilon_k)-[1-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)]\}} \end{aligned}$$ 3. Compute the information matrix $M_{Lap}(\epsilon_{k+1})$, set $k=k+1$ and repeat step 1-3. **Output:** Optimal Design on manifolds. The iterative procedure in Algorithm 1 converges to the D-optimal design $\epsilon^*$, $$\begin{aligned} \lim_{k \rightarrow \infty} |M_{Lap}(\epsilon_k)|=|M_{Lap}(\epsilon^*)| \end{aligned}$$ **Proof** Let the design $\epsilon_0$ not be D-optimal. Based on Proposition \[prop6\], we have $$\begin{aligned} |M_{Lap}(\epsilon_0)| < |M_{Lap}(\epsilon_1)| \leq \cdots \leq |M_{Lap}(\epsilon_k)| \leq |M_{Lap}(\epsilon^*)|\end{aligned}$$ It is known that any bounded monotone sequence converges. Thus the sequence $|M_{Lap}(\epsilon_0)|$, $|M_{Lap}(\epsilon_1)|$, ..., $|M_{Lap}(\epsilon_k)|$ converges to some limit $|M_{Lap}(\hat\epsilon)|$. Next we need to show $$\begin{aligned} |M_{Lap}(\hat\epsilon)|=|M_{Lap}(\epsilon^*)| \end{aligned}$$ The proof proceeds by contradiction. Assume $$\begin{aligned} \label{eqassumption} |M_{Lap}(\hat\epsilon)|<|M_{Lap}(\epsilon^*)| \end{aligned}$$ By the convergence of the sequence $|M_{Lap}(\epsilon_0)|$, $|M_{Lap}(\epsilon_1)|$, ..., $|M_{Lap}(\epsilon_k)|$, we know that, for $\forall \eta >0$, there $\exists k_0 \in \mathbb{N}$ s.t. $$\begin{aligned} |M_{Lap}(\epsilon_{k+1})|-|M_{Lap}(\epsilon_k)| < \eta \; \mathrm{for} \; \forall k > k_0 \end{aligned}$$ Based on Proposition \[prop5\], we have $$\begin{aligned} (1-\alpha_k)^p\Big(1+\frac{\alpha_k}{1-\alpha_k} d(z_{k+1},\epsilon_k)+\frac{\alpha_k}{1-\alpha_k} \operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C) \Big) |M_{Lap}(\epsilon_k)| -|M_{Lap}(\epsilon_k)| &<& \eta \end{aligned}$$ Then, $$\begin{aligned} \label{eq93} (1-\alpha_k)^p\Big(1+\frac{\alpha_k}{1-\alpha_k} [d(z_{k+1},\epsilon_k)+\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C) ] \Big) & <& 1+ \eta |M_{Lap}(\epsilon_k)|^{-1}\end{aligned}$$ Define $\tau_k=d(z_{k+1},\epsilon_k)-[p-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)]$, then we can rewrite (\[eq93\]) as $$\begin{aligned} \label{eq94} (1-\alpha_k)^p\Big(1+\frac{\alpha_k}{1-\alpha_k} [\tau_k+p] \Big) < 1+ \eta |M_{Lap}(\epsilon_k)|^{-1}\end{aligned}$$ Define a function $\mathrm{T}(\tau_k, \alpha_k)$ as $$\begin{aligned} \label{eq:tau:def} \mathrm{T}(\tau_k, \alpha_k)=(1-\alpha_k)^p\Big(1+\frac{\alpha_k}{1-\alpha_k} [\tau_k+p] \Big) \end{aligned}$$ Then $$\begin{aligned} \frac{\partial \mathrm{T}}{\partial \tau_k}&=&(1-\alpha_k)^p \frac{\alpha_k}{1-\alpha_k} \end{aligned}$$ Clearly, $\frac{\partial \mathrm{T}}{\partial\tau_k}>0$ for $0 \leq \alpha_k<1$. Thus, for a given $\alpha_k$, $\mathrm{T}(\tau_k, \alpha_k)$ is an increasing function with respect to $\tau_k$. On the other hand, $$\begin{aligned} \frac{\partial \mathrm{T}}{\partial\alpha_k}=-p(1-\alpha_k)^{p-1}\Big(1+\frac{\alpha_k}{1-\alpha_k} [\tau_k+p] \Big) +(1-\alpha_k)^p [\tau_k+p] \frac{1}{(1-\alpha_k)^2} \end{aligned}$$ Let $\frac{\partial \mathrm{T}}{\partial\alpha_k} \geq 0$, we have $$\begin{aligned} (1-\alpha_k)^{p-2} [\tau_k+p] &\geq& p(1-\alpha_k)^{p-1}\Big(1+\frac{\alpha_k}{1-\alpha_k} [\tau_k+p] \Big) \\ \tau_k+p &\geq& p (1-\alpha_k)+p (\tau_k+p) \alpha_k \\ \tau_k+p &\geq& p - p \alpha_k +p\tau_k\alpha_k +p^2\alpha_k\\ \tau_k &\geq& \alpha_k ( p^2 + p\tau_k-p) \\ \alpha_k &\leq&\frac{\tau_k}{p(p + \tau_k-1)}\end{aligned}$$ Thus, for $0\leq \alpha_k \leq \frac{\tau_k}{p(p + \tau_k-1)}$ and a given $\tau_k$, $\mathrm{T}(\tau_k,\alpha_k)$ is an increasing function. In particular, plug in the formula of $\tau_k$, we have $$\begin{aligned} \frac{\tau_k}{p(p + \tau_k-1)}=\frac{d(z_{k+1},\epsilon_k)-[p-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)]}{p(d(z_{k+1},\epsilon_k)-[1-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)])}. \end{aligned}$$ Notice that $0\leq \alpha_k \leq \frac{\tau_k}{p(p + \tau_k-1)}$ is the same choice of $\alpha$ in the proposed algorithm. In addition, based on Equation (\[eq:tau:def\]), it can be seen that for any $0 < \alpha_k<\frac{\tau_k}{p(p + \tau_k-1)}$ and $\tau_k >0$, we have $\mathrm{T}(\tau_k,\alpha_k) > 1$. Note that $\eta$ is an arbitrary positive number in equation (\[eq94\]), which implies $\tau_k$ need to be an infinitely small positive number to satisfy equation (\[eq94\]), i.e. given $\forall\zeta>0$, there $\exists \tilde{k} (\zeta) \in \mathbb{N}$ s.t. $$\begin{aligned} \tau_k=d(z_{k+1},\epsilon_k)-[p-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)] < \zeta \; \mathrm{for} \; k>\tilde{k}(\zeta) \end{aligned}$$ However, based on the assumption (\[eqassumption\]) and the Theorem \[Thm1\], we have that $$\begin{aligned} d(z_{k+1},\epsilon_k)-[p-\operatorname{Tr}(M^{-1}_{Lap}(\epsilon_k) C)] \geq \delta_k >0 \; \mathrm{for} \; \forall k.\end{aligned}$$ Choosing $\zeta < \delta_k$, we have a contradiction, and therefore, the convergence theorem is proved. From the derivation of Algorithm 1, it is not difficult to notice that ODOEM is a model-dependent design. The corresponding manifold regularization model (\[eqn:LapRLS\]) need to be trained after a desired number of instances is labeled. As it is shown before, Algorithm 1 is a converging algorithm on a continuous design space. However, sometimes the experimental design space is not continuous and only a set of candidate points is available. For a discrete design space with a set of candidate points, one can evaluate each candidate point and choose the point with maximum prediction variance. The resulting sequence of $|M_{Lap}(\epsilon_k)|$ is still non-decreasing, since $$\begin{aligned} |M_{Lap}(\epsilon_k)+g(z_{k+1})g(z_{k+1})^\top|&=&|M_{Lap}(\epsilon_k)|[1+g(z_{k+1})^\top M^{-1}_{Lap}(\epsilon_k) g(z_{k+1})] \\ &\geq& |M_{Lap}(\epsilon_k)|\end{aligned}$$ where $z_{k+1}=\operatornamewithlimits{argmax}\limits_{z \in \mathcal{X}\setminus Z_k} d(z,\epsilon_k)=\operatornamewithlimits{argmax}\limits_{z \in \mathcal{X}\setminus Z_k} g(z)^\top M^{-1}_{Lap}(\epsilon_k) g(z)$. Numerical Results {#sec:numerical} ================= To illustrate the empirical performance of the proposed ODOEM algorithm in practice, we consider its application to both synthetic datasets, low dimensional manifold datasets that permit straightforward visualization of the resulting designs, and also its application to the high dimensional real-world image datasets. Synthetic Datasets ------------------ In this section, we generated four different two-dimensional manifold datasets: data on a Torus, on a Möbius Strip, on a figure “8” immersion of a Klein bottle and on a classic Klein bottle. Each of the first three datasets contains 400 instances and the last dataset contains 1600 instances. For all four datasets, we plot these two-dimensional manifolds in a three-dimensional Euclidean space, as shown in Figure 1-8. The colors on these manifolds represent the corresponding response values $\{y_i\}_{i=1}^n$ (or their estimates based on different experimental designs), which were defined by $$y=sin(u)+sin^2(u)+cos^2(v)$$ where $u \in [0,2\pi)$ and $v \in [0,2\pi)$. The red numbers on the manifolds represent the sequence of labeled instances by different design algorithms. In order to show the improvement provided by using the ODOEM algorithm, we compare it with a classical D-optimal design algorithm on a kernel regression model, which does not consider the manifold structure. For both of the learning models, we choose a RBF kernel and set the range parameter to be 0.01. In addition, we choose $\lambda_A=0.01$ in both models for numerical stability, and $\lambda_I=-\ln(k/n)$ in ODOEM (a discussion of the choice of $\lambda_I$ is provided in the Appendix). For some real-world applications, the data may not strictly lie on a given manifold due to noise. In order to explore the robustness of the ODOEM algorithm to noise, we also let the four synthetic datasets fluctuate around their manifolds by adding noise to $\{{x}_i\}_{i=1}^n$. In other words, for each of the four manifolds, we investigate both of the case when the data $\{{x}_i\}_{i=1}^n$ lie exactly on the given manifold and the case when $\{{x}_i\}_{i=1}^n$ are not exactly on the manifold. The results are shown in Figure 1-4. As it can be seen, on all four synthetic datasets, ODOEM performs much better than kernel regression D-optimal Design in terms of instance selection and function fitting, under both of the case with noises and the case without noises. [0.31]{} ![Torus example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Torus. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Torus. (a) The colors represent the true response values defined on the Torus. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM. As it can be seen here and in Figures 2-4, the predictions (c) approximate the true function on the manifold (a) well even if the data are observed with noise large enough that it distorts the observed manifold with respect to the original one.](t1.jpg "fig:"){width="\textwidth"} [0.31]{} ![Torus example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Torus. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Torus. (a) The colors represent the true response values defined on the Torus. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM. As it can be seen here and in Figures 2-4, the predictions (c) approximate the true function on the manifold (a) well even if the data are observed with noise large enough that it distorts the observed manifold with respect to the original one.](t2.jpg "fig:"){width="\textwidth"} [0.31]{} ![Torus example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Torus. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Torus. (a) The colors represent the true response values defined on the Torus. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM. As it can be seen here and in Figures 2-4, the predictions (c) approximate the true function on the manifold (a) well even if the data are observed with noise large enough that it distorts the observed manifold with respect to the original one.](t3.jpg "fig:"){width="\textwidth"} [0.31]{} ![Torus example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Torus. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Torus. (a) The colors represent the true response values defined on the Torus. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM. As it can be seen here and in Figures 2-4, the predictions (c) approximate the true function on the manifold (a) well even if the data are observed with noise large enough that it distorts the observed manifold with respect to the original one.](t4.jpg "fig:"){width="\textwidth"} [0.31]{} ![Torus example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Torus. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Torus. (a) The colors represent the true response values defined on the Torus. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM. As it can be seen here and in Figures 2-4, the predictions (c) approximate the true function on the manifold (a) well even if the data are observed with noise large enough that it distorts the observed manifold with respect to the original one.](t5.jpg "fig:"){width="\textwidth"} [0.31]{} ![Torus example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Torus. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Torus. (a) The colors represent the true response values defined on the Torus. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM. As it can be seen here and in Figures 2-4, the predictions (c) approximate the true function on the manifold (a) well even if the data are observed with noise large enough that it distorts the observed manifold with respect to the original one.](t6.jpg "fig:"){width="\textwidth"} [0.31]{} ![Möbius Strip. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Möbius Strip. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Möbius Strip. (a) The colors represent the true response values defined on the Möbius Strip. (b) 100 labeled instances (red numbers) and fitted response values by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](m1.jpg "fig:"){width="\textwidth"} [0.31]{} ![Möbius Strip. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Möbius Strip. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Möbius Strip. (a) The colors represent the true response values defined on the Möbius Strip. (b) 100 labeled instances (red numbers) and fitted response values by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](m2.jpg "fig:"){width="\textwidth"} [0.31]{} ![Möbius Strip. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Möbius Strip. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Möbius Strip. (a) The colors represent the true response values defined on the Möbius Strip. (b) 100 labeled instances (red numbers) and fitted response values by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](m3.jpg "fig:"){width="\textwidth"} [0.31]{} ![Möbius Strip. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Möbius Strip. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Möbius Strip. (a) The colors represent the true response values defined on the Möbius Strip. (b) 100 labeled instances (red numbers) and fitted response values by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](m4.jpg "fig:"){width="\textwidth"} [0.31]{} ![Möbius Strip. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Möbius Strip. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Möbius Strip. (a) The colors represent the true response values defined on the Möbius Strip. (b) 100 labeled instances (red numbers) and fitted response values by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](m5.jpg "fig:"){width="\textwidth"} [0.31]{} ![Möbius Strip. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Möbius Strip. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Möbius Strip. (a) The colors represent the true response values defined on the Möbius Strip. (b) 100 labeled instances (red numbers) and fitted response values by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](m6.jpg "fig:"){width="\textwidth"} [0.31]{} ![Figure “8” Immersion of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Figure “8” Immersion. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Figure “8” Immersion. (a) The colors represent the true response values defined on the Figure “8” Immersion. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](f1.jpg "fig:"){width="\textwidth"} [0.31]{} ![Figure “8” Immersion of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Figure “8” Immersion. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Figure “8” Immersion. (a) The colors represent the true response values defined on the Figure “8” Immersion. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](f2.jpg "fig:"){width="\textwidth"} [0.31]{} ![Figure “8” Immersion of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Figure “8” Immersion. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Figure “8” Immersion. (a) The colors represent the true response values defined on the Figure “8” Immersion. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](f3.jpg "fig:"){width="\textwidth"} [0.31]{} ![Figure “8” Immersion of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Figure “8” Immersion. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Figure “8” Immersion. (a) The colors represent the true response values defined on the Figure “8” Immersion. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](f4.jpg "fig:"){width="\textwidth"} [0.31]{} ![Figure “8” Immersion of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Figure “8” Immersion. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Figure “8” Immersion. (a) The colors represent the true response values defined on the Figure “8” Immersion. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](f5.jpg "fig:"){width="\textwidth"} [0.31]{} ![Figure “8” Immersion of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Figure “8” Immersion. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Figure “8” Immersion. (a) The colors represent the true response values defined on the Figure “8” Immersion. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](f6.jpg "fig:"){width="\textwidth"} [0.31]{} ![Bottle Shape of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Bottle. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Bottle. (a) The colors represent the true response values defined on the Bottle. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](k1.jpg "fig:"){width="\textwidth"} [0.31]{} ![Bottle Shape of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Bottle. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Bottle. (a) The colors represent the true response values defined on the Bottle. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](k2.jpg "fig:"){width="\textwidth"} [0.31]{} ![Bottle Shape of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Bottle. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Bottle. (a) The colors represent the true response values defined on the Bottle. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](k3.jpg "fig:"){width="\textwidth"} [0.31]{} ![Bottle Shape of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Bottle. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Bottle. (a) The colors represent the true response values defined on the Bottle. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](k4.jpg "fig:"){width="\textwidth"} [0.31]{} ![Bottle Shape of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Bottle. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Bottle. (a) The colors represent the true response values defined on the Bottle. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](k5.jpg "fig:"){width="\textwidth"} [0.31]{} ![Bottle Shape of Klein Bottle example. Top: When $\{{x}_i\}_{i=1}^n$ lie on a Bottle. Bottom: When $\{{x}_i\}_{i=1}^n$ are not exactly on a Bottle. (a) The colors represent the true response values defined on the Bottle. (b) 100 labeled instances (red numbers) and fitted response values (colors) by kernel regression D-optimal Design. (c) 100 labeled instances (red numbers) and fitted response values (colors) by ODOEM.](k6.jpg "fig:"){width="\textwidth"} Columbia Object Image Library ----------------------------- To demonstrate application to real-world datasets, we tested our ODOEM algorithm on the Columbia Object Image Library (**COIL-20**). COIL-20 is a database of grey-scale images of 20 different objects and these images were taken at pose intervals of 5 degrees for each object. There are two versions of this database. In this paper, we choose the processed database that contains 1440 $32 \times 32$ normalized images. In this set of experiments, the input data $\{{x}_i\}_{i=1}^n$ are the object images and the response values $\{y_i\}_{i=1}^n$ are the corresponding angles of these images. Given an object image, our goal is to estimate the angle of this object in the image. Among 20 different objects, we choose four different objects as illustration: a Rubber Duck, a Cannon, a Toy Car and a Piggy Bank. For each object, we apply the ODOEM algorithm to decide which instances to label and then train the LapRLS model (\[eqn:LapRLS\]) to predict the angles of the images using the labeled and unlabeled instances. Comparisons were made with the following alternative algorithms: - Kernel regression model with a classical D-optimal Design; - Kernel regression model with a random sampling scheme; - Kernel regression model with a $L_2$-discrepancy uniform design; - Kernel regression model with a minimax uniform design; - Kernel regression model with a maximin uniform design; - SVM model with MAED ([*Manifold Adaptive Experimental Design*]{} [@CH12IEEETKDE]); - SVM model with TED ([*Transductive Experimental Design*]{}, [@YZXG2008]). For both kernel regression and SVM, we used a RBF kernel and fixed the range parameter at 0.01. The results are shown in Figure \[coil20\_fig1\]-\[coil20\_fig3\]. Figure \[coil20\_fig1\]-\[coil20\_fig2\] illustrate the first four images selected by classical D-optimal design and ODOEM for training the model. Figure \[coil20\_fig3\] demonstrates the fitting performance (in terms of MSE) of different algorithms. Based on the results, the following comments can be made: (a) Compared to the classical D-optimal design, there is a greater dispersion (in terms of angles) within the first four images selected by ODOEM, which improves the learning curve in Figure \[coil20\_fig3\]; (b) For some uniform design criteria, the corresponding optimization is not convex. Since the images are labeled sequentially, there is no guarantee that the global optimal can be achieved. This explains why some uniform designs do not work very well in these experiments. (c) MAED also benefits from incorporating the manifold structure into the design process. It leads to better fitting performance than other algorithms, except for ODOEM. (d) ODOEM outperforms all the other algorithms on all four object images. [0.47]{} ![Top: The first four Rubber Duck images selected by classical D-optimal design and ODOEM. Bottom: The first four Cannon images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig1"}](duck1.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Rubber Duck images selected by classical D-optimal design and ODOEM. Bottom: The first four Cannon images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig1"}](duck2.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Rubber Duck images selected by classical D-optimal design and ODOEM. Bottom: The first four Cannon images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig1"}](cannon1.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Rubber Duck images selected by classical D-optimal design and ODOEM. Bottom: The first four Cannon images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig1"}](cannon2.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Toy Car images selected by classical D-optimal design and ODOEM.. Bottom: The first four Piggy Bank images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig2"}](car1.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Toy Car images selected by classical D-optimal design and ODOEM.. Bottom: The first four Piggy Bank images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig2"}](car2.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Toy Car images selected by classical D-optimal design and ODOEM.. Bottom: The first four Piggy Bank images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig2"}](pig1.jpg "fig:"){width="\textwidth"} [0.47]{} ![Top: The first four Toy Car images selected by classical D-optimal design and ODOEM.. Bottom: The first four Piggy Bank images selected by classical D-optimal design and ODOEM. The true angle is labeled on top of each image.[]{data-label="coil20_fig2"}](pig2.jpg "fig:"){width="\textwidth"} [0.475]{} ![MSE comparison among different algorithms on all four objects.[]{data-label="coil20_fig3"}](duck.jpg "fig:"){width="\textwidth"} [0.475]{} ![MSE comparison among different algorithms on all four objects.[]{data-label="coil20_fig3"}](cannon.jpg "fig:"){width="\textwidth"} [0.475]{} ![MSE comparison among different algorithms on all four objects.[]{data-label="coil20_fig3"}](car.jpg "fig:"){width="\textwidth"} [0.475]{} ![MSE comparison among different algorithms on all four objects.[]{data-label="coil20_fig3"}](pig.jpg "fig:"){width="\textwidth"} Conclusions {#sec:conc} =========== In this paper, we have developed a theoretical framework of optimal experimental designs on Riemannian manifolds. In particular, we have shown that D-optimal design and G-optimal design are equivalent on manifolds and have provided a new lower bound for the maximum prediction variance, demonstrating that this lower bound can be achieved at the D/G optimal design. In addition, we proposed a converging algorithm for the optimal design of experiments on manifolds. Finally we compared our proposing algorithm with other popular designs and models on several synthetic datasets and real-world image problems, and illustrated the competitive performance of our algorithm. There are several directions of future research in this work. First, further research can be done to develop a systematic procedure for choosing the regularization parameters $\lambda_A$ and $\lambda_I$. As discussed before, cross-validation is not a feasible strategy in a sequential learning problem since there are few or none labeled instances available at the beginning. Instead of using fixed values for regularization parameters, model selection criterion with theoretical justification might provide better learning performance. Similar work has been discussed by [@LCR2019TEST], where they maximize the likelihood function to choose the values of $\lambda_A$ and $\lambda_I$ in a Gaussian Process model. Secondly, there are other optimality criterion than the D/G “alphabetic" criteria in the field of optimal design of experiments. Under different optimal design criteria, new theoretical results of experimental design on manifolds can be explored. Thirdly, for very large scale problems with billions of discrete candidate points, evaluating each point with the corresponding design criteria is exhausting. Some modifications of our algorithm can be investigated, such as applying unsupervised clustering techniques first and then evaluate a representative point from each cluster. Appendix {#sec:appendix .unnumbered} ======== **Equation (\[beta\_est2\]) Proof**\ Let $A=(Z_k^\top Z_k + \lambda_A I_p + \lambda_I X^\top LX)^{-1}$. Then: $$\begin{aligned} AX^\top(XAX^\top)^{-1}XZ^\top \mathbf{y}&=&(Z^\top X)^{-1} Z^\top X AX^\top(XAX^\top)^{-1}XZ^\top \mathbf{y}\\ &=& (Z^\top X)^{-1}Z^\top XZ^\top \mathbf{y}\\ &=&Z^\top \mathbf{y}\end{aligned}$$ Thus, we have $$\begin{aligned} X^\top(XAX^\top)^{-1}XZ^\top \mathbf{y}&=& A^{-1} Z^\top \mathbf{y}\end{aligned}$$ and therefore equation (\[beta\_est1\]) can be reduced to equation (\[beta\_est2\]).\ **Proposition 1 Proof** $$\begin{aligned} M_{Lap}(\epsilon_{3})&=&\int_{z \in \mathcal{X}} \xi_{3}(z)g(z)g(z)^\top dz +C \\ &=& \int_{z \in \mathcal{X}} \big [(1-\alpha)\xi_{1}(z)+\alpha \xi_{2}(z)\big] g(z)g(z)^\top dz +C \\ &=& (1-\alpha) \int_{z \in \mathcal{X}} \xi_1(z)g(z)g(z)^\top dz+(1-\alpha) C +\alpha \int_{z \in \mathcal{X}} \xi_{2}(z)g(z)g(z)^\top dz + \alpha C \\ &=& (1-\alpha) M_{Lap} (\epsilon_1) + \alpha M_{Lap} (\epsilon_2) \end{aligned}$$ **Proposition \[prop2\] Proof** Let $M_{ij}$ be the $(i,j)$ cofactor of the matrix $M_{Lap}(\epsilon_3)$ and let $m_{ij}$ be the $(i,j)$ element of the matrix $M_{Lap}(\epsilon_3)$. Then: $$\begin{aligned} \frac{d \log |M_{Lap}(\epsilon_3)|}{d \alpha} &=&|M_{Lap}(\epsilon_3)|^{-1} \frac{d |M_{Lap}(\epsilon_3) |}{d \alpha} \\ &=& |M_{Lap}(\epsilon_3)|^{-1} \sum_{i=1}^p \sum_{j=1}^p M_{ij} \frac{d m_{ij}(\alpha)}{d \alpha} \\ &=& \sum_{i=1}^p \sum_{j=1}^p \Big(M^{-1}_{Lap}(\epsilon_3)\Big)_{{ji}} \Big(\frac{d M_{Lap}(\alpha)}{d \alpha}\Big)_{{ij}} \\ &=& \operatorname{Tr}\Big(M^{-1}_{Lap}(\epsilon_3) \frac{d M_{Lap}(\alpha)}{d \alpha} \Big) \\ &=& \operatorname{Tr}\Big\{ M^{-1}_{Lap}(\epsilon_3) \frac{d \big[(1-\alpha) M_{Lap} (\epsilon_1) + \alpha M_{Lap} (\epsilon_2)\big]}{d \alpha} \Big\} \\ &=& \operatorname{Tr}\Big\{ M_{Lap}^{-1}(\epsilon_3)\big [M_{Lap}(\epsilon_2)-M_{Lap}(\epsilon_1)\big] \Big\}\end{aligned}$$ **Proposition \[prop3\] Proof** 1\. $$\begin{aligned} && \int_{z\in \mathcal{X}} d (z,\epsilon) \xi(z) dz \\ &=& \int_{z\in \mathcal{X}} g(z)^\top M^{-1}_{Lap}(\epsilon) g(z) \xi(z) dz \\ &=& \int_{z\in \mathcal{X}} \operatorname{Tr}\Big \{g(z)^\top M^{-1}_{Lap}(\epsilon) g(z) \Big \} \xi(z) dz \\ &=& p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} \\ &=& \int_{z\in \mathcal{X}} \operatorname{Tr}\Big \{M^{-1}_{Lap}(\epsilon) \big[g(z) g(z)^\top+C-C \big] \Big \} \xi(z) dz\end{aligned}$$ $$\begin{aligned} &=& \int_{z\in \mathcal{X}} \operatorname{Tr}\Big \{M^{-1}_{Lap}(\epsilon) \big[g(z) g(z)^\top+C \big ] - M^{-1}_{Lap}(\epsilon) C \Big \} \xi(z) dz \\ &=& \int_{z\in \mathcal{X}} \Bigg (\operatorname{Tr}\Big \{M^{-1}_{Lap}(\epsilon) \big[g(z) g(z) ^\top +C \big ] \Big \} - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} \Bigg) \xi(z) dz \\ &=& \operatorname{Tr}\Big \{M^{-1}_{Lap}(\epsilon) \Big [\int_{z\in \mathcal{X}} g(z) g(z)^\top\xi(z) dz +C \Big] \Big \} - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \int_{z\in \mathcal{X}} \xi(z) dz\Big \}\\ &=& \operatorname{Tr}\Big \{M^{-1}_{Lap}(\epsilon) \Big [\int_{z\in \mathcal{X}} g(z) g(z)^\top \xi(z) dz +C \Big] \Big \} - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} \\ &=& \operatorname{Tr}\Big \{M^{-1}_{Lap}(\epsilon) M_{Lap}(\epsilon)\Big \} - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} \\ &=& p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} \end{aligned}$$ 2\. $\int_{z\in \mathcal{X}} d (z,\epsilon) \xi(z) dz = p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \} $ implies that $p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \}$ is the mean value of $d (z,\epsilon)$ for given design $\epsilon$. Thus, we have $$\max_{z \in \mathcal{X}} d(z,\epsilon) \geq p - \operatorname{Tr}\Big \{ M^{-1}_{Lap}(\epsilon) C \Big \}$$ **Proposition \[prop4\] Proof** Let $\epsilon_1$ and $\epsilon_2$ be two arbitrary designs on the experimental region $\mathcal{X}$ and let $M_{Lap}(\epsilon_1)$ and $M_{Lap}(\epsilon_1)$ be the corresponding information matrices. Define the set of information matrices on $\mathcal{X}$ as $$M_{Lap}(\mathcal{X}):= \{ M_{Lap}(\epsilon)| \xi \in \Xi \}$$ where $\Xi$ is the set of all probability measure on $\mathcal{X}$. Clearly, $M_{Lap}(\epsilon_1), M_{Lap}(\epsilon_2) \in M_{Lap}(\mathcal{X})$. Based on Proposition \[prop1\], we have that $$M_{Lap}(\epsilon_3)=(1-\alpha)M_{Lap}(\epsilon_1) + \alpha M_{Lap}(\epsilon_2) \in M_{Lap}(\mathcal{X})$$ where $M_{Lap}(\epsilon_3)$ is the information matrix for the design $\epsilon_3=(1-\alpha)\epsilon_1+\alpha \epsilon_2$. This implies that $M_{Lap}(\mathcal{X})$ is a convex set. In addition, in order to prove $\log |M_{Lap}(\epsilon)|$ is strictly concave, we also need to show that $$\log |(1-\alpha) M_{Lap}(\epsilon_1)+\alpha M_{Lap}(\epsilon_2)| > (1-\alpha) \log |M_{Lap}(\epsilon_1)|+ \alpha \log |M_{Lap}(\epsilon_2)|$$ for $\forall \; M_{Lap}(\epsilon_1) \neq M_{Lap}(\epsilon_2)$ and $\forall \alpha \in (0,1)$. It is known that, for any positive-definite matrices $A$ and $B$, $$|(1-\alpha)A+\alpha B| \geq |A|^{1-\alpha} |B|^\alpha, \; \mathrm{where} \;\alpha \in (0,1),$$ where the equality holds only if $A=B$. Since $M_{Lap}(\epsilon)$ is positive-definite, we have that $$|(1-\alpha) M_{Lap}(\epsilon_1)+\alpha M_{Lap}(\epsilon_2)| > |M_{Lap}(\epsilon_1)|^{1-\alpha} |M_{Lap}(\epsilon_2)|^\alpha.$$ Therefore, $$\log |(1-\alpha) M_{Lap}(\epsilon_1)+\alpha M_{Lap}(\epsilon_2)| > (1-\alpha) \log |M_{Lap}(\epsilon_1)|+ \alpha \log |M_{Lap}(\epsilon_2)|$$ **Proposition \[prop5\] Proof** Based on Proposition \[prop1\], we have $$\begin{aligned} M_{Lap}(\epsilon_{k+1})&=&(1-\alpha) M_{Lap}(\epsilon_k)+\alpha M_{Lap}(\epsilon(z)) \\ &=& (1-\alpha) M_{Lap}(\epsilon_k)+\alpha (g(z)g(z)^\top+C) \\ &=& (1-\alpha) \Big [ M_{Lap}(\epsilon_k)+\frac{\alpha}{1-\alpha} g(z)g(z)^\top+ \frac{\alpha}{1-\alpha} C \Big]\end{aligned}$$ Then $$\begin{aligned} &&|M_{Lap}(\epsilon_{k+1})|\\ &=& (1-\alpha)^p \Big|M_{Lap}(\epsilon_k)(I_p+\frac{\alpha}{1-\alpha}M_{Lap}(\epsilon_k)^{-1}g(z)g(z)^\top+ \frac{\alpha}{1-\alpha} M_{Lap}(\epsilon_k)^{-1} C) \Big|\\ &=& (1-\alpha)^p \Big|M_{Lap}(\epsilon_k)\Big| \Big|I_p+\frac{\alpha}{1-\alpha}M_{Lap}(\epsilon_k)^{-1}g(z)g(z)^\top+ \frac{\alpha}{1-\alpha} M_{Lap}(\epsilon_k)^{-1} C \Big|\end{aligned}$$ Assume $\alpha$ is an infinitesimal number. It’s known that $$\begin{aligned} &&\Big|I_p+\frac{\alpha}{1-\alpha}M_{Lap}(\epsilon_k)^{-1}g(z)g(z)^\top+ \frac{\alpha}{1-\alpha} M_{Lap}(\epsilon_k)^{-1} C \Big|\\ &=&1+\frac{\alpha}{1-\alpha}\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1}g(z)g(z)^\top+ M_{Lap}(\epsilon_k)^{-1} C)\\ &=& 1+\frac{\alpha}{1-\alpha} d(z,\epsilon_k)+\frac{\alpha}{1-\alpha} \operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C)\end{aligned}$$ Therefore, we have $$\begin{aligned} |M_{Lap}(\epsilon_{k+1})|=(1-\alpha)^p \Big|M_{Lap}(\epsilon_k) \Big| \Big[1+\frac{\alpha}{1-\alpha} d(z,\epsilon_k)+\frac{\alpha}{1-\alpha} \operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C) \Big] \end{aligned}$$ **Proposition \[prop6\] Proof** Based on Equation (\[eqprop4\]), $|M_{Lap}(\epsilon_{k+1})|$ is clearly an increasing function with respect to $d(z,\epsilon_k)$. In order to maximize the value of $ \log |M_{Lap}(\epsilon_{k+1})|$, we choose $z_{k+1}=\operatornamewithlimits{argmax}\limits_{z \in \mathcal{X}} d (z, \epsilon_k)$. Thus, we have that $$\begin{aligned} \log |M_{Lap}(\epsilon_{k+1})| &=& p \log (1-\alpha)+\log |M_{Lap}(\epsilon_k)| \\ &&+ \log \Big [1+\frac{\alpha}{1-\alpha} d(z_{k+1},\epsilon_k)+\frac{\alpha}{1-\alpha} \operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C) \Big ].\end{aligned}$$ It can be shown that $$\begin{aligned} & &\frac{\partial \log |M_{Lap}(\epsilon_{k+1})| }{\partial \alpha} \\ &=& \frac{d(z_{k+1},\epsilon_k)-(p-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))+p\alpha(1-d(z_{k+1},\epsilon_k)-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))}{(1-\alpha)[(1-\alpha)+\alpha d(z_{k+1},\epsilon_k)+\alpha \operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C)]}.\end{aligned}$$ Since $$\frac{\partial \log |M_{Lap}(\epsilon_{k+1})| }{\partial \alpha} \geq 0,$$ then $$\begin{aligned} d(z_{k+1},\epsilon_k)-(p-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))+p\alpha(1-d(z_{k+1},\epsilon_k)-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C)) \geq 0.\end{aligned}$$ After simplification, we have $$\begin{aligned} \alpha \leq \frac{d(z_{k+1},\epsilon_k)-(p-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))}{p[d(z_{k+1},\epsilon_k)-(1-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))]}\end{aligned}$$ Clearly, $$\begin{aligned} \label{eq:alpha:range} 0 \leq \alpha \leq \frac{d(z_{k+1},\epsilon_k)-(p-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))}{p[d(z_{k+1},\epsilon_k)-(1-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))]}\end{aligned}$$ is the non-decreasing direction for the value of $ \log |M_{Lap}(\epsilon_{k+1})|$. In addition, based on the Proposition \[prop3\] Equation (\[maxd\_ineq\]), it is clear that $$\begin{aligned} \frac{d(z_{k+1},\epsilon_k)-(p-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))}{p[d(z_{k+1},\epsilon_k)-(1-\operatorname{Tr}(M_{Lap}(\epsilon_k)^{-1} C))]} \geq 0\end{aligned}$$ which guarantees the existence of $\alpha$ in Equation (\[eq:alpha:range\]). Therefore, $\Big\{|M_{Lap}(\epsilon_k)|\Big\}_k$ is a non-decreasing sequence. **Choice of** $\lambda_I$ The regularization parameters $\lambda_A$ and $\lambda_I$ are usually selected by cross-validation. However, ODOEM is a sequential design algorithm and the order of labeled instance is important. The cross-validation idea of randomly dividing the labeled instances into training set and validation set does not work here. Thus, one can set fixed values for $\lambda_A$ and $\lambda_I$. In our experiments, we set $\lambda_A=0.01$ for numerical stability and generate a decreasing sequence of $\lambda_I$ by setting $\lambda_I=-\ln(k/n)$, where $k$ is the number of labeled instance at $k$-th iteration and $n$ is the total number of instances. The reason we choose a decreasing sequence of $\lambda_I$ comes from the penalized loss function (\[eqn:LapRLS\]) and the performance evaluation criterion MSE$=\sum_{i=1}^n (y_i -\hat{f}(z_i))^2$. For manifold regularization model, the estimated learning function $\hat{f}$ is achieved by minimizing the objective function (\[eqn:LapRLS\]). At early iterations, there are only few labeled instances, and $\hat{f}$ would benefit more from penalizing the learning function along the manifold structure (second regularization term). As the number of labeled instances increase, larger $\lambda_I$ might not lead to smaller MSE. For example, let’s consider the extreme scenario when all the instances have been labeled, i.e. $k=n$. If one want to achieve smaller MSE$=\sum_{i=1}^n (y_i -\hat{f}(z_i))^2$, it is better to estimate $\hat{f}$ by $$\hat{f}=\operatornamewithlimits{argmin}\limits_{f\in \mathcal{H}_K} \sum_{i=1}^n (y_i -f(z_i))^2,$$ instead of using $$\hat{f}=\operatornamewithlimits{argmin}\limits_{f\in \mathcal{H}_K} \sum_{i=1}^{n}(y_i-f(z_i))^2+\lambda_A \|f\|_{\mathcal{H}_\mathcal{K}}^2+\lambda_I \textbf{f}^\top L \textbf{f}.$$ In summary, for a learning problem with a fixed number of labeled instances, $\lambda_I$ can be chosen using cross-validation. For a learning problem with sequentially labeled instances, we set $\lambda_I=-\ln(k/n)$ so that we can get a decreasing sequence of $\lambda_I$ as $k$ increases and $\lambda_I=0$ when all the instances have been labeled. [^1]: The authors gratefully acknowledge *NSF grant CMMI 1537898*
{ "pile_set_name": "ArXiv" }
--- abstract: 'This article presents an overview of applications of logic programming, classifying them based on the abstractions and implementations of logic languages that support the applications. The three key abstractions are join, recursion, and constraint. Their essential implementations are for-loops, fixed points, and backtracking, respectively. The corresponding kinds of applications are database queries, inductive analysis, and combinatorial search, respectively. We also discuss language extensions and programming paradigms, summarize example application problems by application areas, and touch on example systems that support variants of the abstractions with different implementations.' author: - | Yanhong A. Liu\ [ Computer Science Department, Stony Brook University]{}\ [[email protected]]{} bibliography: - '\\bibdir/strings.bib' - '\\bibdir/liu.bib' - '\\bibdir/IC.bib' - '\\bibdir/PT.bib' - '\\bibdir/PA.bib' - '\\bibdir/Veri.bib' - '\\bibdir/Lang.bib' - '\\bibdir/Algo.bib' - '\\bibdir/Perform.bib' - '\\bibdir/DB.bib' - '\\bibdir/SE.bib' - '\\bibdir/Sys.bib' - '\\bibdir/Sec.bib' - '\\bibdir/misc.bib' - '\\bibdir/crossref.bib' date: 'December 24, 2017' title: | Logic programming applications:\ What are the abstractions and implementations?[^1] --- Introduction ============ Common reasoning with logic is the root of logic programming, which allows logic rules and facts to be expressed formally and used precisely for inference, querying, and analysis in general. Logic formalisms, or languages, allow complex application problems to be expressed declaratively with high-level abstractions and allow desired solutions to be found automatically with potentially efficient low-level implementations. The biggest challenge in logic programming has been the need for efficient implementations. Much progress has been made, with efficient implementations in some cases beating manually written low-level code. However, inadequate performance in many cases has led to the introduction of non-declarative features in logic languages and resulted in the writing of obscure logic programs. Despite the challenges, the most exciting aspect of logic programming is its vast areas of applications. They range from database queries to program analysis, from text processing to decision making, from security to knowledge engineering, and more. These vast, complex, and interrelated areas make it challenging but necessary to provide a deeper understanding of the various kinds of applications in order to help advance the state of the art of logic programming and realize its benefits. This article presents an overview of applications of logic programming based on a study of the abstractions and implementations of logic languages. The rationale is that abstractions and implementations are the enabling technologies of the applications. The abstractions are essential for determining what kinds of application problems can be expressed and how they can be expressed, for ease of understanding, reuse, and maintenance. The underlying implementations are essential for high-level declarative languages to be sufficiently efficient for substantial applications. We discuss the following essential abstractions, where data abstractions are for expressing the data, and control abstractions are for expressing computations over the data: 1. data abstractions: objects and relationships; 2. control abstractions: (1) join, (2) recursion, and (3) constraint, which capture bounded, cyclic, and general computations, respectively. In logic languages, the data abstractions as objects and relationships are essential for all three control abstractions. The essential techniques for implementing the three control abstractions listed are (1) for-loops, (2) fixed points, and (3) backtracking, respectively. The corresponding kinds of applications are   : \(1) database-style queries, e.g., for ontology management, business intelligence, and access control;   : \(2) inductive analysis, e.g., for text processing, program analysis, network traversal, and trust management;   : \(3) combinatorial search, e.g., for decision making, resource allocation, games and puzzles, and administrative policy analysis. We categorize application problems using these three control abstractions because they capture conceptually different kinds of problems, with inherently different implementation techniques, and at the same time correspond to very different classes of applications. Note that the same application domain may use different abstractions and implementations for different problems. For example, enterprise software may use all three of traditional database queries, inductive analysis, and combinatorial search, for business intelligence and decision making; and security policy analysis and enforcement may use database-style queries for access control, inductive analysis for trust management, and combinatorial search for administrative policy analysis. We also discuss additional extensions, especially regular-expression paths for higher-level queries and updates for modeling actions; additional applications; and abstractions used in main programming paradigms. We also touch on several well-known systems while discussing the applications. There is a large body of prior work, including surveys of logic programming in general and applications in particular, as discussed in Section \[sec-relate\]. This article distinguishes itself from past work by analyzing classes of applications based on the language abstractions and implementations used. The rest of the article is organized as follows. Section \[sec-lang\] presents essential abstractions in logic languages. Sections \[sec-join\], \[sec-recursion\], and \[sec-constraint\] describe abstractions, implementations, and applications centered around join, recursion, and constraint. Section \[sec-ext\] discusses additional language extensions, applications, and programming paradigms. Section \[sec-relate\] discusses related literature and future directions. Logic language abstractions {#sec-lang} =========================== Logic languages provide very high-level data and control abstractions, using mostly very simple language constructs. We describe these abstractions and their meanings intuitively. Data abstractions ----------------- All data in logic languages are abstracted, essentially, as objects and relationships. [[[**Objects.**]{}]{}]{} Objects are primitive values, such as numbers and strings, or structured values whose components are objects. Examples of primitive values are integer number 3 and string ’Amy’. We enclose a string value in single quotes; if a string starts with a lower-case letter, such as ’amy’, the quotes can be omitted, as has been conventional in logic languages. Examples of structured values are şucc(3), father(amy), andçert(’Amy’,birth(’2000-02-28’,’Rome’)), denoting the successor integer of 3, the father of amy, and the certificate that ’Amy’ was born on ’2000-02-28’ in ’Rome’, respectively. The names of structures, such as şucc, father, çert, and birth above, are called function symbols. They correspond to object constructors in object-oriented languages. [[[**Relationships.**]{}]{}]{} Relationships are predicates, or properties, that hold among objects. In particular, (,...,), i.e., predicate over objects ,..., being true, is equivalent to (,...,) in , i.e., tuple (,...,) belonging to relation —a table that holds the set of tuples of objects over which is true. Examples of relationships are male(bob), is\_parent(bob,amy), andissue(mario,’Amy’,birth(’2000-02-28’,’Rome’)), denoting that bob is male, bob is a parent of amy, and mario issued a certificate that ’Amy’ was born on ’2000-02-28’ in ’Rome’, respectively. Structured values can be easily captured using relationships, but not vice versa. For example, f being the structured value father(c) can be captured using relationship is\_father(f,c), but relationship is\_parent(p,c) cannot simply be captured as p being the structured value parent(c) when ç has two parents. Such high-level data abstraction allows real-world objects or their lower-level representations, from bits and characters to lists to sets, to be captured easily without low-level implementation details. For example, - bits and characters are special cases of integers and strings, respectively. - lists are a special case of linearly nested structured values, and - sets are a special case of relations consisting of tuples of one component. Objects and relationships can be implemented using well-known data structures, including linked list, array, hash table, B-tree, and trie, usually taking Ø[1]{} or Ø[n]{} time to access an object, where is the size of the data. Control abstractions -------------------- \[sec-control\] Control in logic languages is abstracted at a high level, as logical inference or logic queries over asserted relationships among objects: - asserted relationships can be connected by logical connectives: conjunction (read “and”), disjunction (read “or”), negation (read “not”), implication (read “then”), backward implication (read “if”), and equivalence (read “if and only if”); - variables can be used in place of objects and be quantified over with universal quantifier (read “all”) and existential quantifier (read “some”); and - one can either infer all relationships that hold or query about certain relationships, among all objects or among certain objects. Rules and facts are the most commonly supported forms in existing logic languages: [[[**Rules.**]{}]{}]{} A rule is of the following form, where is called the conclusion, and other assertions are called the hypotheses. Each assertion is a predicate over certain objects, where variables may be used in place of objects. Intuitively, left arrow ([&lt;]{}) indicates backward implication, comma (,) denotes conjunction, and all variables in a rule are implicitly universally quantified, i.e., the rule holds for all values of the variables. [&lt;]{}, ..., . For example, the second rule below says: X is a grandfather of Y if X is the father of Z and Z is a parent of Y, and this holds for all values of variables X, Y, and Z; the other rules can be read similarly. Following logic language conventions, names starting with an upper-case letter are variables. is\_parent(X,Y) [&lt;]{}is\_father(X,Y). is\_grandfather(X,Y) [&lt;]{}is\_father(X,Z), is\_parent(Z,Y). is\_ancestor(X,Y) [&lt;]{}is\_parent(X,Z), is\_ancestor(Z,Y). is\_positive(succ(N)) [&lt;]{}is\_positive(N). The second rule is a join query—its two hypotheses have a shared variable, and it concludes a new predicate. The third and fourth rules are recursive—the predicate in the conclusion depends on itself in a hypothesis, or in general possibly indirectly through another predicate. Note that disjunction of a set of hypotheses can be expressed using a set of rules with the same conclusion. [[[**Facts.**]{}]{}]{} A fact is a rule that has no hypotheses and is denoted simply as assertion0. For example, is\_father(bob,amy). says that bob is the father of amy, and is\_positive(1). says that 1 is positive. The meaning of a set of rules and facts is the least set of facts that contains all the given facts and all the facts that can be inferred, directly or indirectly, using the rules. This set can be computed by starting with the given facts and repeatedly applying the rules to conclude new facts—i.e., matching hypotheses of rules against facts, instantiating variables in rules with values in matched facts, and adding instantiated conclusions of rules as new facts. However, - repeated application of rules might not terminate if function symbols are used in the rules, because facts about infinitely many new objects may be concluded, e.g., the fourth example rule above may infer is\_positive(succ(1)), is\_positive(succ(succ(1))), and so on. - when only certain relationships about certain objects are queried, application of rules may stop as soon as the query can be answered, e.g., if only is\_positive(succ(1)) is queried, application of rules can stop after one use of the given rule and the given fact. Rules that do not contain function symbols are called Datalog rules. For example, the first three example rules given earlier in this section are Datalog rules. General logic forms have also been increasingly supported, typically by extending the rule form above: [[[**Negation in the hypotheses.**]{}]{}]{} A hypothesis in a rule may be prefixed with not, denoting negation of the asserted relationship. For example, the following rule says: for all values of X and Y, X is the mother of Y if X is a parent of Y and X is not male. is\_mother(X,Y) [&lt;]{}is\_parent(X,Y), not male(X). Difficulties arise when negation is used with recursion. For example, what can be inferred from the following rule? Is good(zak) true or false? good(zak) [&lt;]{}not good(zak). [[[**More general forms.**]{}]{}]{} More general forms include disjunction and negation in the conclusion and, most generally, quantifiers all and şome in any scope, not only the outermost scope. For example, the first rule below says: X is male or female if X is a person. The second rule says: X is not a winning position if, for all Y, there is no move from X to Y or else Y is a winning position. male(X) or female(X) [&lt;]{}person(X). not win(X) [&lt;]{}all Y: not move(X,Y) or win(Y). The meaning of recursive rules with negation is not universally agreed upon. The two dominant semantics are well-founded semantics (WFS) [@van+91well; @van93alt] and stable model semantics (SMS) [@GelLif88]. Both WFS and SMS use the closed-world assumption, i.e., they assume that what cannot be inferred to be true from the given facts and rules, is false. - WFS gives a single 3-valued model, with the additional truth value undefined besides ţrue and false. - SMS gives zero or more 2-valued models, using only ţrue and false. Other formalisms and semantics include partial stable models, also called stationary models [@prz94well]; first-order logic with inductive and fixed-point definitions, called FO(ID) and FO(FD) [@denecker2008logic; @hou2010fo]; and the newly proposed founded semantics and constraint semantics [@LiuSto18Founded-LFCS]. The first two are both aimed at unifying WFS and SMS. The last unifies and cleanly relates WFS, SMS, and other major semantics by allowing the assumptions about the predicates and rules to be specified explicitly. For practical applications, logic languages often also support predefined relationships among objects, including equality, inequality, and general comparisons. Cardinality and other aggregates over relationships are often also supported. Combinations of control abstractions ------------------------------------ There are many possible combinations of the language constructs. We focus on the following three combinations of constructs as essential control abstractions. We identify them by join, recursion, and constraint. They capture bounded, cyclic, and general computations, respectively. : \(1) [[**Join**]{}]{}—with join queries, no recursive rules, and restricted negation and other constructs; the restriction is that, for each rule, each variable in the conclusion must also appear in a hypothesis that is a predicate over arguments. Implementing this requires that common objects for the shared variables be found for the two hypotheses of a join query to be true at the same time; the number of objects considered are bounded, by the predicates in the hypotheses, following a bounded number of dependencies. : \(2) [[**Recursion**]{}]{}—with join queries, recursive rules, and restricted negation and other constructs; the restriction is as for join above plus that a predicate in the conclusion of a rule does not depend on the negation of the predicate itself in a hypothesis. Implementing this requires repeatedly applying the recursive rules following cyclic dependencies, potentially an unbounded number of times if new objects are in some conclusions. : \(3) [[**Constraint**]{}]{}—with join queries, recursive rules, and unrestricted negation and other constructs; unrestricted negation and other constructs can be viewed as constraints to be satisfied. Implementing this could require, in general, trying different combinations of variable values, as in general constraint solving. Table \[tab-abs\] summarizes these three essential control abstractions and the corresponding kinds of computations and applications. [@l@[ ]{}|l@[ ]{}||l@[ ]{}|l@[ ]{}|l@[ ]{}||l@[ ]{}||@[ ]{}l@]{} & Essential & ---------- Has join queries ---------- : Essential control abstractions of logic languages.[]{data-label="tab-abs"} & ---------- Has rec. rules ---------- : Essential control abstractions of logic languages.[]{data-label="tab-abs"} & ------------ Has neg. and others ------------ : Essential control abstractions of logic languages.[]{data-label="tab-abs"} & Computations & Application kinds\ (1)& Join & yes & no & restricted & bounded & database-style queries\ (2)& Recursion & yes & yes & restricted & cyclic & inductive analysis\ (3)& Constraint & yes & yes & unrestricted & general & combinatorial search\ Join and database-style queries {#sec-join} =============================== Join queries are the most basic and most commonly used queries in relating different objects. They underlie essentially all nontrivial queries in database applications and many other applications. Join queries ------------ A join query is a conjunction of two hypotheses that have shared variables, concluding possible values of variables that satisfy both hypotheses. A conjunction of two hypotheses that have no shared variables, i.e., a Cartesian product, or a single hypothesis can be considered a trivial join query. A join query corresponds to a rule whose predicate in the conclusion is different from predicates in the hypothesis, so the rule is not recursive. A non-recursive rule with more than two hypotheses corresponds to multiple join queries, as a nesting or chain of join queries starting with joining any two hypotheses first. For example, the first rule below, as seen before, is a join query. So is the second rule; it defines şibling over X and Y if X and Y have a same parent. The third rule defines a chain of red, green, and blue links from X to Y through U and V; it can be viewed as two join queries—join any two hypotheses first, and then join the result with the third hypothesis. is\_grandfather(X,Y) [&lt;]{}is\_father(X,Z), is\_parent(Z,Y). sibling(X,Y) [&lt;]{}is\_parent(Z,X), is\_parent(Z,Y). chain(X,Y) [&lt;]{}link(X,U,red), link(U,V,green), link(V,Y,blue). In general, the asserted predicates can be about relationships among any kinds of objects—whether people, things, events, or anything else, e.g., students, employees, patients, doctors, products, courses, hospitals, flights, interviews, and hangouts; and the join queries can be among any kinds of relationships—whether family, friend, owning, participating, thinking, or any other relation in the real world or conceptual world. Join queries expressed using rules correspond to set queries. For example, in a language that supports set comprehensions with tuple patterns [@RotLiu07Retrieval-PEPM; @Liu+16IncOQ-PPDP] the is\_grandfather query corresponds to is\_grandfather = {(X,Y): (X,Z) in is\_father, (Z,Y) in is\_parent} Without recursion, join queries can be easily supported together with the following extensions, with the restriction that, for each rule, each variable in the conclusion must also appear in a hypothesis that is a predicate over arguments, so the domain of the variable is bounded by the predicate; queries using these extensions can be arbitrarily nested: - unrestricted negation, other connectives, and predefined relationships in additional conditions, - aggregates, such as count and max, about the relationships, and - general universal and existential quantifiers in any scope. These subsume all constructs in the şelect statement for SQL queries. Essentially, join queries, with no recursion, relate objects in different relationships within a bounded number of steps. Implementation of join queries ------------------------------ \[sec-join-impl\] A join query can be implemented straightforwardly using nested for-loops and if-statements, where shared variables in different hypotheses correspond to equality tests between the corresponding variables. For example, the is\_grandfather query earlier in this section can be implemented as is\_grandfather = {} for (X,Z1) in is\_father: – time factor: number of is\_father pairs for (Z2,Y) in is\_parent: – time factor: number of is\_parent pairs if Z1 == Z2: is\_grandfather.add(X,Y) In a language that supports set comprehensions, such as Python, the above implementation can be expressed as is\_grandfather = {(X,Y) for (X,Z1) in is\_father for (Z2,Y) in is\_parent if Z1 == Z2} For efficient implementations, several key implementation and optimization techniques are needed, described below; additional optimizations are also needed, e.g., for handling streaming data or distributed data. [[[**Indexing.**]{}]{}]{} This creates an index for fast lookup based on values of the indexed arguments of a relation; the index is on the shared arguments of the two hypotheses. For example, for any fact is\_father(X,Z), to find the matching is\_parent(Z,Y), an index called, say, çhildren{Z}—mapping the value of Z, the first argument of is\_parent, to the set of corresponding values of second argument of is\_parent—significantly speeds up the lookup, improving the time factor for the inner loop to the number of children of Z: is\_grandfather = {(X,Y) for (X,Z) in is\_father for Y in children{Z}} [[[**Join ordering.**]{}]{}]{} This optimizes the order of joins when there are multiple joins, e.g., in a rule with more than two hypotheses. For example, for the rule for çhain, starting by joining the first and third hypotheses is never more efficient than starting by joining either of these hypotheses with the second hypothesis, because the former yields all pairs of red and blue links, even if there are no green links in the middle. [[[**Tabling.**]{}]{}]{} This stores the result of common sub-joins so they are not repeatedly computed. Common sub-joins may arise when there are nested or chained join queries. For example, for the rule for çhain earlier in this section, consider joining the first two hypotheses first: if there are many red and green link pairs from a value of X to a value of V, then storing the result of this sub-join avoids recomputing it when joining with blue links to find each target Y. [[[**Demand-driven computation.**]{}]{}]{} This computes only those parts of relationships that affect a particular query. For example, a query may only check whether is\_father(dan,bob) holds, or find all values of X for is\_father(dan, X), or find all is\_father pairs, as opposed to finding all relationships that can be inferred. Basic ideas for implementing the extensions negation, aggregates, etc. are as follows, where nested queries using these extensions are computed following their order of dependencies: - negation, etc. in additional conditions: test them after the variables in them become bound by the joins. - aggregates: apply the aggregate operation while collecting the query result of its argument. - quantifiers: transform them into for-loops, or into aggregates, e.g., an existential quantification is equivalent to a count being positive. Efficient implementation techniques for join queries and extensions have been studied in a large literature, e.g., [@ioa1996query]. Some methods also provide precise complexity guarantees, e.g., [@wil02jcss; @Liu+16IncOQ-PPDP]. Applications of join queries ---------------------------- Join queries are fundamental in querying complex relationships among objects. They are the core of database applications [@KifBL06], from enterprise management to ontology management, from accounting systems to airline reservation systems, and from electronic health records management to social media management. Database and logic programming are so closely related that one of the most important computer science bibliographies is called DBLP, and it was named after Database and Logic Programming [@ley2002dblp]. Join queries also underlie applications that do not fit in traditional database applications, such as complex access control policy frameworks [@ansi04role]. We describe three example applications below, in the domains of ontology management, enterprise management, and security policy frameworks. They all heavily rely on the use of join queries and optimizations, especially indexing. We give specific examples of facts, rules, and indexing for the first application. [[[**Ontology management—Coherent definition framework (CDF).**]{}]{}]{} CDF is a system for ontology management that has been used in numerous commercial projects [@Gomes2010mknf], for organizing information about, e.g., aircraft parts, medical supplies, commercial processes, and materials. It was originally developed by XSB, Inc. Significant portions have been released in the XSB packages [@xsb14]. The data in CDF are classes and objects. For example, XSB, Inc. has a part taxonomy, combining UNSPSC (United Nations Standard Products and Services Code) and Federal INC (Item Name Code) taxonomies, with a total of over 87,000 classes of parts. The main relationships are variants of isa, ḩasAttr, and allAttr. Joins are used extensively to answer queries about closely related classes, objects, and attributes. Indexing and tabling are heavily used for efficiency. Appropriate join order and demand-driven computation are also important. An example fact is as follows, indicating that specification ’A-A-1035’ in ontology şpecs has attribute ’MATERIAL’ whose value is ’ALUMINUM ALLOY UNS A91035’ in material\_taxonomy. Terms çid(Identifier, Namespace) represent primitive classes in CDF. hasAttr\_MATERIAL(cid(’A-A-1035’,specs), cid(’ALUMINUM ALLOY UNS A91035’,material\_taxonomy)). An example rule is as follows, meaning that a part PartNode has attribute’PART-PROCESS-MATERIAL’ whose value is process-material pair (Process,Material) in’ODE Ontology’ if PartNode has attribute ’PROCESS’ whose value is Process, and Process has attribute ’PROCESS-MATERIAL’ whose value is Material. hasAttr\_PART-PROCESS-MATERIAL(PartNode, cid(’process-material’(Process,Material),’ODE Ontology’)) [&lt;]{} hasAttr\_PROCESS(PartNode, Process), hasAttr\_PROCESS-MATERIAL(Process, Material). An example of indexing is for ḩasAttr\_ATTR, for any ATTR, shown below, in XSB notation, meaning: use as index all symbols of the first argument if it is bound, or else do so for the second argument. \[\*(1), \*(2)\] XSB, Inc. has five major ontologies represented in CDF, for parts, materials, etc., with a total of over one million facts and five meta rules. The rules are represented using a Description Logic form—an ontology representation language. The example rule above is an instance of such a rule when interpreted. The indexing used supports different appropriate indices for different join queries. CDF is used in XSB, Inc.’s ontology-directed classifier (ODC) and extractor (ODE) [@swift2012xsb]. ODC uses a modified Bayes classifier to classify item descriptions. For example, it is used quarterly by the U.S. Department of Defense to classify over 80 million part descriptions. ODE extracts attribute-value pairs from classified descriptions to build structured knowledge about items. ODC uses aggregates extensively, and ODE uses string pattern rules. [[[**Enterprise management—Business intelligence (BI).**]{}]{}]{} BI is a central component of enterprise software. It tracks the performance of an enterprise over time by storing and analyzing historical information recorded through online transaction processing (OLTP), and is then used to help plan future actions of the enterprise. LogicBlox simplifies the hairball of enterprise software technologies by using a Datalog-based language [@green12logicblox; @aref2015design]. All data are captured as logic relations. This includes not only data as in conventional databases, e.g., sale items, price, and so on for a retail application, but also data not in conventional databases, e.g., sale forms, display texts, and submit buttons in a user interface. Joins are used for easily querying interrelated data, as well as for generating user interfaces. Many extensions such as aggregates are also used. For efficiency, exploiting the rich literature of automatic optimizations, especially join processing strategies and incremental maintenance, is of paramount importance. Using the same Datalog-based language, LogicBlox supports not only BI but also OLTP and prescriptive and predictive analytics. “Today, the LogicBlox platform has matured to the point that it is being used daily in dozens of mission-critical applications in some of the largest enterprises in the world, whose aggregate revenues exceed \$300B” [@aref2015design]. [[[**Security policy frameworks—Core role-based access control (RBAC).**]{}]{}]{} RBAC is a framework for controlling user access to resources based on roles. It became an ANSI standard [@ansi04role] building on much research during the preceding decade and earlier, e.g., [@Landwehr:Heitmeyer:McLean:84; @Ferraiolo+92; @Gavrila:Barkley:98; @ferraiolo01proposed]. Core RBAC defines users, roles, objects, operations, permissions, sessions and a number of relations among these sets; the rest of RBAC adds a hierarchical relation over roles, in hierarchical RBAC, and restricts the number of roles of a user and of a session, in constrained RBAC. Join queries are used for all main system functions, especially the ÇheckAccess function, review functions, and advanced review functions on the sets and relations. They are easily expressed using logic rules [@barker+04effi; @barker+06term]. Efficient implementations rely on all main optimizations discussed, especially auxiliary maps for indexing and tabling [@Liu+06ImplCRBAC-PEPM]. Although the queries are like relational database queries, existing database implementations would be too slow for functions like ÇheckAccess. Unexpectedly, uniform use of relations and join queries also led to a simplified specification, with unnecessary mappings removed, undesired omissions fixed, and constrained RBAC drastically simplified [@LiuSto07RBAC-ONR]. Recursion and inductive analysis {#sec-recursion} ================================ Recursive rules are most basic and essential in relating objects that are an unknown number of relationships apart. They are especially important for problems that may require performing the inference or queries a non-predetermined number of steps, depending on the data. Recursive rules and queries --------------------------- Given a set of rules, a predicate depends on a predicate if is in the conclusion of a rule, and either is in a hypothesis of the rule or some predicate is in a hypothesis of the rule and depends on . A given set of rules is recursive if a predicate in the conclusion of a rule depends on itself. For example, the second rule below, as seen in Section \[sec-control\], is recursive; the first rule is not recursive; the set of these two rules is recursive, where the first rule is the base case, and the second rule is the recursive case. is\_ancestor(X,Y) [&lt;]{}is\_parent(X,Y). is\_ancestor(X,Y) [&lt;]{}is\_parent(X,Z), is\_ancestor(Z,Y). In general, recursively asserted relationships can be between objects of any kind, e.g., relatives and friends that are an unknown number of connections apart in social networks, direct and indirect prerequisites of courses in universities, routing paths in computer networks, nesting of parts in products, supply chains in supply and demand networks, transitive role hierarchy relation in RBAC, and repeated delegations in trust management systems. Recursive queries with restricted negation correspond to least fixed-point computations. For example, in a language that supports least fixed points, the is\_ancestor query corresponds to the minimum is\_ancestor set below, where, for any sets Ş and Ţ, Ş subset T holds iff every element of Ş is an element of Ţ: min is\_ancestor: is\_parent subset is\_ancestor, {(X,Y): (X,Z) in is\_parent, (Z,Y) in is\_ancestor} subset is\_ancestor With cyclic predicate dependencies, recursion allows the following restricted extensions to be supported while still providing a unique semantics; there is also the restriction that, for each rule, each variable in the conclusion must also appear in a hypothesis that is a predicate over arguments, as in extensions to join queries: - stratified negation, where negation and recursion are separable, i.e., there is no predicate that depends on the negation of itself, and - other connectives and predefined relationships in additional conditions, aggregates, and general quantifiers, as in extensions for join queries, when they do not affect the stratification. Essentially, recursive rules capture an unbounded number of joins, and allow inference and queries by repeatedly applying the rules. Implementation of recursive rules and queries --------------------------------------------- Inference and queries using recursive rules can be implemented using while-loops; for-loops with predetermined number of iterations do not suffice, because the number of iterations depends on the rules and facts. Each iteration applies the rules in one step, so to speak, until no more relevant facts can be concluded. For example, the is\_ancestor query earlier in this section can be implemented as is\_ancestor = is\_parent while exists (X,Y): (X,Z) in is\_parent, (Z,Y) in is\_ancestor, (X,Y) not in is\_ancestor: is\_ancestor.add((X,Y)) Each iteration computes the existential quantification in the condition of the while-loop, and picks any witness (X,Y) to add to the result set. It can be extremely inefficient to recompute the condition in each iteration after a new pair is added. For efficient implementations, all techniques for joins are needed but are also more critical and more complex. In particular, to ensure termination, - tabling is critical if relationships form cycles, and - demand-driven computation is critical if new objects are created in the cycles. For the is\_ancestor query, each iteration computes the following set, which is a join, plus the last test to ensure that only a new fact is added: {(X,Y): (X,Z) in is\_parent, (Z,Y) in is\_ancestor, (X,Y) not in is\_ancestor} Two general principles underlying the optimizations for efficient implementations are: 1. incremental computation for expensive relational join operations, with respect to facts that are added in each iteration. 2. data structure design for the relations, for efficient retrievals and tests of relevant facts. For the restricted extensions, iterative computation follows the order of dependencies determined by stratification; additional aggregates, etc.that do not affect the stratification can be handled as described in Section \[sec-join-impl\] for computing the join in each iteration. Efficient implementation techniques for recursive queries and extensions have been studied extensively, e.g., [@AbiHulVia95]. Some methods also provide precise complexity guarantees, e.g., [@mcallester99; @Ganzinger:McAllester:01; @LiuSto09Rules-TOPLAS; @TekLiu11RuleQueryBeat-SIGMOD]. Applications of recursive rules and queries ------------------------------------------- Recursive rules and queries can capture any complex reachability problem in recursive structures, graphs, and hyper-graphs. Examples are social network analysis based on all kinds of social graphs; program analysis over many kinds of flow and dependence graphs about program control and data values; model checking over labeled transition systems and state machines; routing in electronic data networks, telephone networks, or transportation networks; and security policy analysis and enforcement over trust or delegation relationships. We describe three example applications below, in the domains of text and natural language processing, program analysis, and distributed security policy frameworks. They all critically depend on the use of recursive rules and efficient implementation techniques, especially tabling and indexing. [[[**Text processing—Super-tokenizer.**]{}]{}]{} Super-tokenizer is an infrastructure tool for text processing that has been used by XSB, Inc.’s ontology-directed classifier (ODC) and extractor (ODE) for complex commercial applications [@swift2012xsb]. It was also developed originally at XSB, Inc. Super-tokenizer supports the declaration of complex rewriting rules for token lists. For example, over 65,000 of these rules implement abbreviations and token corrections in ODC and complex pattern-matching rules in ODE for classification and extraction based on combined UNSPSC and Federal INC taxonomies at XSB, Inc. Recursion is used extensively in the super-tokenizer, for text parsing and processing. The implementation uses tabled grammars and trie-based indexing in fundamental ways. Super-tokenizer is just one particular application that relies on recursive rules for text processing and, more generally, language processing. Indeed, the original application of Prolog, the first and main logic programming language, was natural language processing (NLP) [@pereira2002prolog], and a more recent application in NLP helped the IBM Watson question answering system win the Jeopardy Man vs. Machine Challenge by defeating two former grand champions in 2011 [@lally2011natural; @lally2012question]. [[[**Program analysis—Pointer analysis.**]{}]{}]{} Pointer analysis statically determines the set of objects that a pointer variable or expression in a program can refer to. It is a fundamental program analysis with wide applications and has been studied extensively, e.g., [@hind01; @sri2013alias]. The studies especially include significantly simplified specifications using Datalog in more recent years, e.g., [@smara15pointer], and powerful systems such as bddbddb [@whaley2005using] and Doop [@bravenboer2009strictly], the latter built using LogicBlox [@green12logicblox; @aref2015design]. Different kinds of program constructs and analysis results relevant to pointers are relations. Datalog rules capture the analysis directly as recursively defined relations. For example, the well-known Andersen’s pointer analysis for C programs defines a points-to relation based on four kinds of assignment statements [@and94thesis], leading directly to four Datalog rules [@SahaRam05]. Efficient implementation critically depends on tabling, indexing, and demand-driven computation [@SahaRam05; @TekLiu11RuleQueryBeat-SIGMOD]. Such techniques were in fact followed by hand to arrive at the first ultra fast analysis [@HeiTar01ultra; @HeiTar01demand]. Indeed, efficient implementations can be generated from Datalog rules giving much better, more precise complexity guarantees [@LiuSto09Rules-TOPLAS; @TekLiu11RuleQueryBeat-SIGMOD] than the worst-case complexities, e.g., the well-known cubic time for Andersen’s analysis. Such efficient implementation with complexity guarantees can be obtained for program analysis in general [@mcallester99]. Commercial tools for general program analysis based on Datalog have also been built, e.g., by Semmle based on CodeQuest [@haj+2006codequest]. [[[**Security policy frameworks—Trust management (TM).**]{}]{}]{} TM is a unified approach to specifying and enforcing security policies in distributed systems [@BlaFeiLac96; @GraSlo00; @ruohomaa2005trust]. It has become increasingly important as systems become increasingly interconnected, and logic-based languages have been used increasingly for expressing TM policies [@bonatti2010datalog], e.g., SD3 [@Jim01], RT [@LiMitWin02], Binder [@DeT02], Cassandra [@BecSew04], and many extensions, e.g., [@becker2012foundations; @sultana2013selective]. Certification, delegation, authorization, etc. among users, roles, permissions, etc. are relations. Policy rules correspond directly to logic rules. The relations can be transitively defined, yielding recursive rules. For example, one of the earliest TM frameworks, SPKI/SDSI [@Ell+99], for which various sophisticated methods have been studied, corresponds directly to a few recursive rules [@Hri+07SPKI-PPDP], and efficient implementations with necessary indexing and tabling were generated automatically. TM studies have used many variants of Datalog with restricted constraints [@LiMit03], not unrestricted negation. A unified framework with efficient implementations is still lacking. For example, based on the requirements of the U.K. National Health Service, a formal electronic health records (EHR) policy was written, as 375 rules in Cassandra [@Bec05ehr], heavily recursive. As the largest case study in the TM literature, its implementation was inefficient and incomplete—techniques like indexing were deemed needed but missing [@Bec05thesis]. Constraint and combinatorial search {#sec-constraint} =================================== Constraints are the most general form of logic specifications, which easily captures the most challenging problem-solving activities such as planning and resource allocation. Constraint satisfaction ----------------------- A constraint is, in general, a relationship among objects but especially refers to cases when it can be satisfied with different choices of objects and the right choice is not obvious. For example, the rule below says that X is a winning position if there is a move from X to Y and Y is not a winning position. It states a relationship among objects, but its meaning is not obvious, because the concluding predicates are recursively defined using a negation of the predicate itself. win(X) [&lt;]{}move(X,Y), not win(Y). In general, constraints can capture any real-world or conceptual-world problems, e.g., rules for moves in any game—whether recreational, educational, or otherwise; actions with conditions and effects for any planning activities; participants and resource constraints in scheduling—whether for university courses or manufacturer goods production or hospital surgeries; real-world constraints in engineering design; as well as knowledge and rules for puzzles and brain teasers. Given constraints may have implications that are not completely explicit. For example, the win rule implies not just the first constraint below, but also the second, by negating the conclusion and hypotheses in the given rule, following the closed-world assumption; the second constraint makes the constraint about not win explicit: win(X) if some Y: move(X,Y) and not win(Y) not win(X) if all Y: not move(X,Y) or win(Y) Indeed, with general constraints, objects can be related in all ways using all constructs together with join and recursion: unrestricted negation, other connectives, predefined relationships, aggregates, and general quantifiers in any scope. However, due to negation in dependency cycles, the meaning of the rules and constraints is not universally agreed on anymore. - Well-founded semantics (WFS) gives a single, 3-valued model, where relationships that are true or false are intended to be supported from given facts, i.e., well-founded, and the remaining ones are undefined. - Stable model semantics (SMS) gives zero or more 2-valued models, where each model stays the same, i.e., is stable, when it is used to instantiate all the rules; in other words, applying the rules to each model yields the same model. For example, for the win example, - if there is only one move, move(a,b), not forming a cycle, then WFS and SMS both give that win(b) is false and win(a) is true; - if there is only one move, move(a,a), forming a self cycle, then WFS gives that win(a) is undefined, and SMS gives that there is no model; - if there are only two moves, move(a,b) and move(b,a), forming a two-move cycle, then WFS gives that win(a) and win(b) are both undefined, and SMS gives two models: one with win(a) true and win(b) false, and one with the opposite results. Despite the differences, WFS and SMS can be computed using some shared techniques. Implementation of constraint satisfaction ----------------------------------------- Constraint solving could in general use straightforward generate-and-test—generate each possible combination of objects for solutions and test whether they satisfy the constraints—but backtracking is generally used, as it is much more efficient. [[[**Backtracking.**]{}]{}]{} Backtracking incrementally builds variable assignments for the solutions, and abandons each partial assignment as soon as it determines that the partial assignment cannot be completed to a satisfying solution, going back to try a different value for the last variable assigned; this avoids trying all possible ways of completing those partial assignments or naively enumerating all complete assignments. For example, the win(X) query can basically try a move at each next choice of moves and backtrack to try a different move as soon as the current move fails. Expressed using recursive functions, this corresponds basically to the following: def win(X): return (some Y: move(X,Y) and not\_win(Y)) def not\_win(X): return (all Y: not move(X,Y) or win(Y)) This backtracking answers the query correctly when the moves do not form a cycle. However, it might not terminate when the moves form a cycle, and the implementation depends on the semantics used. Both WFS and SMS can be computed by using and extending the basic backtracking: - WFS computation could track cycles, where executing a call requires recursively making the same call, and infer undefined for those queries that have no execution paths to infer the query result to be true or false. - SMS computation could generate possible partial or complete variable assignments, called grounding, and check them, possibly with the help of an external solver like Boolean satisfiability (SAT) solvers or satisfiability modulo theories (SMT) solvers. For efficient implementations, techniques for join and recursion are critical as before, especially tabling to avoid repeated states in the search space. Additionally, good heuristics for pruning the search space can make drastic performance difference in computing SMS, e.g., as implemented in answer set programming (ASP) solvers. [[[**Backjumping.**]{}]{}]{} One particular optimization of backtracking in SMS computation is backjumping. Backtracking always goes back one level in the search tree when all values for a variable have been tested. Backjumping may go back more levels, by realizing that a prefix of the partial assignment can lead to all values for the current variable to fail. This helps prune the search space. For extensions that include additional constraints, such as integer constraints, as well as aggregates and quantifiers, an efficient solver such as one that supports mixed integer programming (MIP) can be used. Efficient implementation techniques for constraint solving have been studied extensively, e.g., for ASP solvers [@leone2006dlv; @GebKKS12book]. Applications of constraint satisfaction --------------------------------------- The generality and power of constraints allow them to be used for all applications described previously, but constraints are particularly important for applications beyond those and that require combinatorial search. Common kinds of search problems include planning and scheduling, resource allocation, games and puzzles, and well-known NP-complete problems such as graph coloring, k-clique, set cover, Hamiltonian cycle, and SAT. We describe three example applications, in the domains of decision making, resource allocation, and games and puzzles. They all require substantial use of general constraints and efficient constraint solvers exploiting backtracking, backjumping, and other optimizations. [[[**Enterprise decision making—Prescriptive analysis.**]{}]{}]{} Prescriptive analysis suggests decision options that lead to optimized future actions. It is an advanced component of enterprise software. For example, for planning purposes, LogicBlox supports prescriptive analysis using the same Datalog-based language as for BI and OLTP [@green12logicblox; @aref2015design]. The data are objects and relations, same as used for BI, but may include, in particular, costs and other objective measures. Constraints capture restrictions among the objects and relations. When all data values are provided, constraints can simply be checked. When some data values are not provided, different choices for those values can be explored, and values that lead to certain maximum or minimum objective measures may be prescribed for deciding future actions. Efficient implementations can utilize the best constraint solvers based on the kinds of constraints used. LogicBlox’s integrated solution to decision making based on BI and OLTP has led to significant success. For example, for a Fortune 50 retailer with over \$70 billion in revenue and with products available through over 2,000 stores and digital channels, the solution processes 3 terabytes of data on daily, weekly, and monthly cycles, deciding exactly what products to sell in what stores in what time frames; this reduces a multi-year cycle of a challenging task for a large team of merchants and planners to an automatic process and significantly increases profit margins [@logicblox15plan]. [[[**Resource allocation—Workforce management (WFM) in Port of Gioia Tauro.**]{}]{}]{} WFM handles activities needed to maintain a productive workforce. The WFM system for automobile logistics in the Port of Gioia Tauro, the largest transshipment terminal in the Mediterranean, allocates available personnel of the seaport such that cargo ships mooring in the port are properly handled [@ricca2012team; @LeoneRicca2015asp]. It was developed using the DLV system [@leone2006dlv]. The data include employees of different skills, cargo ships of different sizes and loads, teams and roles to be allocated, and many other objects to be constrained, e.g., workload of employees, heaviness of roles, and contract rules. Constraints include matching of available and required skills, roles, hours, etc., fair distribution of workload, turnover of heavy or dangerous roles, and so on. The constraints are expressed using rules with disjunction in the conclusion, general negation, and aggregates. The DLV system uses backtracking and a suite of efficient implementation techniques. This WFM system was developed by Exeura s.r.l. and has been adopted by the company ICO BLG operating automobile logistics in the Port of Gioia Tauro [@LeoneRicca2015asp], handling every day several ships of different sizes that moor in the port [@ricca2012team]. [[[**Games and puzzles—N-queens.**]{}]{}]{} We use a small example in a large class of problems. The n-queens puzzle is the problem of placing n queens on a chessboard of n-by-n squares so that no two queens threaten each other, i.e., no two queens share the same row, column, or diagonal. The problem is old, well-studied, and can be computationally quite expensive [@bell2009survey]. The allowed placements of queens can be specified as logic rules with constraints. Naively enumerating all possible combination of positions and checking the constraints is prohibitively expensive. More efficient solutions use backtracking, and furthermore backjumping, to avoid impossible placement of each next queen as soon as possible. Stronger forms of constraints may also be specified to help prune the search space further [@GebKKS12book]. For example, backtracking can solve for one or two scores of queens in an hour, but backjumping and additional constraints help an ASP system like Clingo solve for 5000 queens in 3758.320 seconds of CPU time [@schaub14slides]. Many other games and puzzles can be specified and solved in a similar fashion. Examples are all kinds of crossword puzzles, Sudoku, Knight’s tour, nonograms, magic squares, dominos, coin puzzles, graph coloring, palindromes, among many others, e.g., [@demoen05first; @edmunds15puzzles; @hett15prolog; @malita15logic; @kje15lpicat]. Further extensions, applications, and discussion {#sec-ext} ================================================ We discuss additional language extensions and applications, summarize applications based on the key abstractions used, touching on example logic programming systems, and finally put the abstractions into the perspective of programming paradigms. Extensions ---------- Many additional extensions to logic languages have been studied. Most of them can be viewed as abstractions that capture common patterns in classes of applications, to allow applications to be expressed more easily. Important extensions include: - regular-expression paths, a higher-level abstraction for commonly-used linear recursion; - updates, for real-world applications that must handle changes; - time, for expressing changes over time, as an alternative to supporting updates directly; - probability, to capture uncertainty in many challenging applications; and - higher-order logic, to support applications that require meta-level reasoning. We discuss two of the most important extensions below: [[[**Regular-expression paths.**]{}]{}]{} A regular-expression path relates two objects using regular expressions and extensions. It allows repeated joins of a binary relation to be expressed more easily and clearly than using recursion; such joins capture reachability and are commonly used. For example, is\_ancestor(X,Y), defined in Section \[sec-recursion\] using two rules including a recursive rule, can now be defined simply as below; it indicates that there are one or more is\_parent relationships in a path from X to Y: is\_ancestor(X,Y) [&lt;]{}is\_parent+(X,Y). This is also higher-level than using recursion, because the recursive rule has to pick one of three possible forms below: with is\_parent on the left, as seen before; with is\_parent on the right; and with both conjuncts using is\_ancestor. is\_ancestor(X,Y) [&lt;]{}is\_parent(X,Z), is\_ancestor(Z,Y). is\_ancestor(X,Y) [&lt;]{}is\_ancestor(X,Z), is\_parent(Z,Y). is\_ancestor(X,Y) [&lt;]{}is\_ancestor(X,Z), is\_ancestor(Z,Y). Depending on the data, the performance of these forms can be asymptotically different in most implementations. Regular-expression paths have many important applications including all those in Section \[sec-recursion\], especially graph queries, with also parametric extensions for more general relations, not just binary relations [@deM+03; @Liu+04PRPQ-PLDI; @LiuSto06GraphQL-PADL; @Tek+10ImplGraphQL-PPDP]. [[[**Updates.**]{}]{}]{} An update, or action, can be expressed as a predicate that captures the update, e.g., by relating the values before and after the update and the change in value. The effect of the update could be taken immediately after the predicate is evaluated, similar to updates in common imperative languages, but this leads to lower-level control flows that are harder to reason about. Instead, it is better for the update to take effect as part of a transaction of multiple updates that together satisfy high-level logic constraints. For example, with this approach, the following rule means that adopted\_by\_from holds if the updates add\_child and del\_child and the check adoption\_check happen as a transaction. adopted\_by\_from(C,X,Y) [&lt;]{}add\_parent(X,C), del\_parent(Y,C), adoption\_check(C,X,Y). It ensures at a high-level that certain bad things won’t happen, e.g., no child would end with one fewer parent or one more parent than expected. Transaction logic is an extension of logic rules for reasoning about and executing transactional state changes [@BonKif94]. LPS, a Logic-based approach to Production Systems, captures state changes by associating timestamps with facts and events, and this is shown to correspond to updating facts directly [@kow15reactive]. Logic languages with updates have important applications in enterprise software [@green12logicblox; @aref2015design]. Transaction logic can also help in planning [@basseda2014planning]. Additional implementation support can help enhance applications and enable additional applications. A particular helpful feature is to record justification or provenance information during program execution [@roychoudhury2000justif; @damasio2013justif], providing explanations for how a result was obtained. The recorded information can be queried to improve understanding and help debugging. Additional applications ----------------------- Many additional applications have been developed using logic programming, especially including challenging applications that need recursion and those that furthermore need constraint. Table \[tab-appl\] lists example application areas with example application problems organized based on the main abstractions used. Note that application problems can often be reduced to each other, and many other problems can be reduced to the problems in the table. For example, model checking a property of a system [@Clarke+99; @Clarke:Grumberg:Long:94] can be reduced to planning, where the goal state is a state violating the property specified, so a plan found by a planner corresponds to an error trace found by a model checker [@MCAP14; @APMC14]. Administrative policy analysis also has correspondences to planning, by finding a sequence of actions to achieve the effect of a security breach [@stoller2011symbolic]. [@[ ]{}m[10.9ex]{}@[ ]{}||@[ ]{}m[21.5ex]{}@[ ]{}|@[ ]{}m[26.5ex]{}@[ ]{}|@[ ]{}m[24ex]{}@[ ]{}]{} Area & Using join & Using recursion & Using constraint\ Data management & business intelligence,\* many databasejoin queries & route queries, many databaserecursive queries & data cleaning, data repair\ Knowledge management & ontologymanagement\* & ontology analysis & reasoning withknowledge\ Decision support & & supply-chain management, market analysis & prescriptive analysis,\* planning, scheduling, resource allocation\*\ Linguistics & & text processing,\* context-free parsing,semantic analysis & context-sensitiveanalysis, deep semantics analysis\ Program analysis & type checking,many local analyses & pointer analysis,\*type analysis, many dependency analyses & type inference ,many constraint-basedanalyses\ Security & role-basedaccess control\* & trust management,\* hierarchical role-basedaccess control & administrative policyanalysis,cryptanalysis\ Games and puzzles & & Hanoi tower, many recursion problems & n-queens,\*Sudoku,many constraint puzzles\ Teaching & course management & course analysis & question analysis,problem diagnosis,test generation\ Table \[tab-appl\] is only a small sample of the application areas, with example application problems or kinds of application problems in those areas. Many more applications have been developed, in many more areas, using systems that support variants of the abstractions with different implementations. Some examples are: - XSB has also been used to develop applications for immunization survey [@burton12], standardizing data, spend analysis, etc. [@xsb15case], and it is discussed in many publications[^2]. - LogicBlox has also been used to create solutions for predicting consumer demand, optimizing supply chain, etc. [@logicblox15sol] and more [@green12logicblox]. - ASP systems have been used in bioinformatics, hardware design, music composition, robot control, tourism, and many other application areas [@wasp05; @schaub11asp], including part of a decision support system for the Space Shuttle flight controller [@nogueira2001prolog; @balduccini2005model]. - Logic systems have been developed for additional applications, e.g., PRISM [@sato1997prism] and ProbLog [@de2007problog] for probabilistic models; XMC [@rama2000xmc] and ProB [@leuschel2008prob] for verification; and NDlog [@loo09decl], Meld [@ashley-rollman-iclp09], Overlog [@alvaro2010declare], and Bloom [@bloom] for network and distributed algorithms. Languages and systems with more powerful features such as constraints for general applications are often also used in less challenging application areas such as those that need only join queries. For example, DLV has also been used in ontology management [@ricca2009ontodlv]. Additional discussion on abstractions ------------------------------------- We give an overview of the main abstractions in the larger picture of programming paradigms, to help put the kinds of applications supported into broader perspective. The three main abstractions—join, recursion, and constraint—correspond generally to more declarative programming paradigms. Each is best known in its corresponding main programming community: - Join in database programming. Database systems have join at the core but support restricted recursion and constraints in practice. - Recursion in functional programming. Functional languages have recursion at the core but do not support high-level join or constraints. - Constraint in logic programming. Logic engines support both join and recursion at the core, and have increasingly supported constraints at the core as well. The additional extensions help further raise the level of abstraction and broaden the programming paradigms supported: - Regular-expression paths raise the level of abstraction over lower-level linear recursion. - Updates, or actions, are the core of imperative programming; they help capture real-world operations even when not used in low-level algorithmic steps. - Time, probability, higher-order logic, and many other features correspond to additional arguments, attributes, or abstractions about objects and relationships. One main paradigm not yet discussed is object-oriented programming. Orthogonal to data and control abstractions, objects in common object-oriented languages provide a kind of module abstraction, encapsulating both data structures and control structures in objects and classes. Similar abstractions have indeed been added to logic languages as well. For example, F-logic extends traditional logic programming with objects [@KifLW95] and is supported in Flora-2 [@flora14]; it was also the basis of a highly scalable commercial system, Ontobroker [@ontobroker], and a recent industry suite, Ergo [@Grosof15ruleml]. For another example, ASP has been extended with object constructs in OntoDLV [@ricca2009ontodlv]. Finally, building practical applications requires powerful libraries and interfaces for many standard functionalities. Many logic programming systems provide various such libraries. For example, SWI-Prolog has libraries for constraint logic programming, multithreading, interface to databases, GUI, a web server, etc, as well as development tools and extensive documentation. Related literature and future work {#sec-relate} ================================== There are many overview books and articles about logic programming in general and applications of logic programming in particular. This article differs from prior works by studying the key abstractions and their implementations as the driving force underlying vastly different application problems and application areas. Kowalski [@kow14lp] provides an extensive overview of the development of logic programming. It describes the historical root of logic programming, starting from resolution theorem-proving; the procedural interpretation and semantics of rules with no negated hypotheses, called Horn clause programs; negation as failure, including completion semantics, stratification, well-founded semantics, stable model semantics, and ASP; as well as logic programming involving abduction, constraints, and argumentation. It focuses on three important issues: logic programming as theorem proving vs. model generation, with declarative vs. procedural semantics, and using top-down vs. bottom-up computation. Our description of abstractions and implementations aims to separate declarative semantics from procedural implementations. Other overviews and surveys about logic programming in general include some that cover a collection of topics together and some that survey different topics separately. Example collections discuss the first 25 years of logic programming from 1974 [@apt1999lp25] and the first 25 years of the Italian Association of Logic Programming from 1985 [@dovier2010lp25]. Example topics surveyed separately include logic programming semantics [@fitting2002fixpoint], complexity and expressive power [@dantsin2001complexity], constraints [@jaf+1994constr], ASP and DLV [@grasso2013asp], deductive databases [@Ceri:Gottlob:Tanca:90; @AbiHulVia95; @RamUll95survey; @Min+14history], and many more. Our description of abstractions and implementations is only a highly distilled overview of the core topics. Overviews and surveys about logic programming applications in particular are spread across many forums. Example survey articles include an early article on Prolog applications [@roth1993practical], DLV applications [@grasso2009some; @grasso2011asp; @LeoneRicca2015asp], applications in Italy [@dal2010lp25], emerging applications [@huang2011datalog], and a dedicated workshop AppLP—Applications of Logic Programming [@WarLiu17AppLP-arxiv]. For example, the early article [@roth1993practical] describes six striking practical applications of Prolog that replaced and drastically improved over systems written previously using Fortran, C++, and Lisp. Example collections of applications on the Web include one at TU Wien [@wasp05], one by Schaub [@schaub11asp], and some of the problems in various competitions, e.g., as described by Gebser et al. [@gebser17sixth]. We try to view the applications by the abstractions and implementations used, so as to not be distracted by specific details of very different applications. There are also many articles on specific applications or specific classes of applications. Examples of the former include team building [@ricca2012team], program pointer analysis [@smara15pointer], and others discussed in this article. Examples of the latter include applications in software engineering [@cian1995report], DLV applications in knowledge management [@grasso2009some], and IDP applications in data mining and machine learning [@bruynooghe2014predicate]. We used a number of such specific applications as examples and described some of them in slightly more detail to illustrate the common technical core in addition to the applications per se. #### Directions for future work. There are several main areas for future study: (1) more high-level abstractions that are completely declarative, (2) more efficient implementations with complexity guarantees, and (3) more unified and standardized languages and frameworks with rich libraries. These will help many more applications to be created in increasingly complex problem domains. Acknowledgment {#acknowledgment .unnumbered} -------------- I would like to thank David S. Warren for his encouragement over the years at Stony Brook, and his patient and stimulating explanations about logic programming in general and XSB implementation in particular. I am grateful to Molham Aref, Francesco Ricca, and David Warren for helpful suggestions and additional information about applications using LogicBlox, DLV, and XSB, respectively. I thank Molham Aref and others at LogicBlox, Jon Brandvein, Christopher Kane, Michael Kifer, Bob Kowalski, Bo Lin, Francesco Ricca, Scott Stoller, Tuncay Tekle, David Warren, Neng-Fa Zhou, and anonymous reviewers for helpful comments on drafts of this article. [^1]: This work was supported in part by NSF under grants CCF-0964196, CCF-1248184, CCF-1414078, and IIS-1447549; and ONR under grants N00014-15-1-2208. [^2]: A Google Scholar search with +XSB +”logic programming” returns over 2300 results, July 2, 2017.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Computing a basis for the exponent lattice of algebraic numbers is a basic problem in the field of computational number theory with applications to many other areas. The main cost of a well-known algorithm [@ge1993algorithms; @kauers2005algorithms] solving the problem is on computing the primitive element of the extended field generated by the given algebraic numbers. When the extended field is of large degree, the problem seems intractable by the tool implementing the algorithm. In this paper, a special kind of exponent lattice basis is introduced. An important feature of the basis is that it can be inductively constructed, which allows us to deal with the given algebraic numbers one by one when computing the basis. Based on this, an effective framework for constructing exponent lattice basis is proposed. Through computing a so-called pre-basis first and then solving some linear Diophantine equations, the basis can be efficiently constructed. A new certificate for multiplicative independence and some techniques for decreasing degrees of algebraic numbers are provided to speed up the computation. The new algorithm has been implemented with Mathematica and its effectiveness is verified by testing various examples. Moreover, the algorithm is applied to program verification for finding invariants of linear loops.' author: - | Tao Zheng and Bican Xia\ \ title: An Effective Framework for Constructing Exponent Lattice Basis of Nonzero Algebraic Numbers --- =10000 =10000 \[Algebraic Algorithms\] Introduction ============ For $x=(x_1,\ldots,x_n)^T$$\in(\overline{\mathbb{Q}}^*)^n$, where $\overline{\mathbb{Q}}$ denotes the set of algebraic numbers, vectors $\alpha=(k_1,\ldots,k_n)^T$$\in\mathbb{Z}^n$ satisfying $x^\alpha$$= x_1^{k_1}x_2^{k_2}\cdots x_n^{k_n}=1$ form an *exponent lattice*, which accepts a basis since $\mathbb{Z}$ is Noetherian. Computing that basis is a significant problem from both practical and theoretical points of view. Exponent lattice has various applications. For example, based on computing exponent lattice, an algorithm was proposed to compute the Zariski closure of a finitely generated group of invertible matrices in [@derksen2005quantum], the growth behavior when $k$$\rightarrow $$\infty$ of rational linear recurrence sequence was studied in [@recurrence], the problem of finding integers $k_1,\ldots,k_n$ *s.t.* $\mathcal{G}_1^{k_1}\cdots\mathcal{G}_n^{k_n}$ is a rational function for $G-$solutions $\mathcal{G}_j$ in [@chen2011structure] could be solved, also an algorithm was provided to compute the ideal of algebraic relations among C-finite sequences in [@kauers2008computing]. A class of loop invariants called L-invariants introduced in [@lvov2010polynomial] are closely related to exponent lattice: each vector in the lattice corresponds to an L-invariant. Moreover Theorems 3 and 5 in [@structure] show that a part of the invariant ideal of a linear loop is exactly the lattice ideal defined by the exponent lattice. The lattice ideal accepts a set of finite generators corresponding to a Markov basis of the lattice which can be computed from the usual basis by the method in [@markov]. The first result leading to the computability of the exponent lattice basis was presented in [@masser1988linear], which bounds a basis of the lattice inside a box that contains finitely many vectors, allowing one to do exhaustive search inside the box to obtain a basis of the lattice. A more efficient algorithm to solve this problem was proposed in [@ge1993algorithms]. It is redescribed in [@kauers2005algorithms] $\S7.3$ and implemented as a Mathematica package **FindRelations**. According to [@kauers2005algorithms], “the runtime is usually negligible compared to the time needed for computing the primitive element”. That means **FindRelations** usually takes much time to compute a primitive element of the extended field generated by the given algebraic numbers. When degree of the extended field is slightly large, the function tends to fail to return an answer within one hour. In this paper, a special kind of exponent lattice basis is introduced. An important feature of the basis is that it can be inductively constructed, which allows us to deal with the given algebraic numbers one by one when computing the basis. Based on this, an effective framework for constructing exponent lattice basis is proposed. Through computing a so-called pre-basis first and then solving some linear Diophantine equations, the basis can be efficiently constructed. A new certificate for multiplicative independence and some techniques for decreasing degrees of algebraic numbers are provided to speed up the computation. The paper is organized as follows. Section \[sec:def\] introduces basic conceptions used throughout the paper. The main result on constructing a basis of the lattice (Theorem \[basisthm\]) and a certificate for multiplicative independence (Theorem \[suff\]) are given in Section \[sec:main\]. Main algorithms computing the pre-basis and the basis are then presented in Section \[sec:alg\]. Section \[sec:red\] devotes to design algorithms decreasing degrees of algebraic numbers. Experiments are carried out and an application to compute invariant ideal of linear loops is illustrated in Section \[sec:example\]. Finally, the paper is concluded in Section \[sec:conclusions\]. Definitions and conceptions {#sec:def} =========================== A sequence $\{x_i\}_{i=1}^n\subset{\overline{\mathbb{Q}}}^*$ is *multiplicatively dependent* *(dependent* for short*)* if there are integers $k_1,\ldots,k_n$ not all zero such that $ x_1^{k_1}\cdots x_n^{k_n}=1$, otherwise it is *multiplicatively independent* *(independent* for short*)*. *Note:* An empty sequence $\epsilon$ containing no algebraic numbers is multiplicatively independent with **length**($\epsilon)=0$, and is a subsequence of any sequence of nonzero algebraic numbers. \[mis\] For sequence $\{x_i\}_{i=1}^n\subset{\overline{\mathbb{Q}}}^*$, its subsequence $S$ is a *maximal independent sequence* if : (i) $S$ is multiplicatively independent and (ii) ***length***$(S)<n$ implies any supersequence $T$ of $S$ with ***length***$(T)>$***length***$(S)$ is multiplicatively dependent. *Note:* $\epsilon$ is a maximal independent sequence of $x_1,\ldots,x_n$ iff every $x_j$ is a root of unity. A number $\mathfrak{a}\in\overline{\mathbb{Q}}^*$ can be *pseudo*-*multiplicatively represented* by $x_1,\ldots,x_n\in\overline{\mathbb{Q}}^*$ if there are integers $k_1,\ldots,k_n$ and $k\neq0$ such that $\mathfrak{a}^k=x_1^{k_1}\cdots x_n^{k_n}$. If $n=0$, $\mathfrak{a}$ can be *pseudo*-*multiplica*-*tively* *represented* by $\epsilon$ means $\mathfrak{a}$ is a root of unity. The *order* of a root of unity $\mathfrak{a}$, denoted by $\textbf{\emph{Order}}(\mathfrak{a})$, is the least positive integer so that $\mathfrak{a}^k=1$. \[pmr\] Suppose $S$ is a maximal independent sequence of $\{x_i\}_{i=1}^n\subset{\overline{\mathbb{Q}}}^*$. Then $x_i$ can be pseudo-multiplicatively represented by $S$ for each $i=1,...,n$. \[rank\] Let $S=\{y_i\}_{i=1}^m\subset\overline{\mathbb{Q}}^*$ and $T= \{x_j\}_{j=1}^n\subset\overline{\mathbb{Q}}^*$. If every $y_j$ can be pseudo-multiplicatively represented by $T$ and $m>n$, then $S$ is multiplicatively dependent. Let $S$ and $T$ be two maximal independent sequences of $x_1,\ldots,x_n$, then $\emph{\textbf{length}}(S)=\emph{\textbf{length}}(T)$. \[algerank\] The length of any maximal independent sequence of $\{x_i\}_{i=1}^n\subset\overline{\mathbb{Q}}^*$ is called the *rank* of $x_1,\ldots,x_n$, denoted by $\textbf{\emph{rank}}(x)$. \[rx\] Define $\mathcal{R}_x=\{v\in\mathbb{Z}^n|x^v=1\}$ with $x=(x_1,\ldots,x_n)^T$ in $({\overline{\mathbb{Q}}^*})^n$ and $x^v = x_1^{k_1}\cdots x_n^{k_n}$ if $v = (k_1,\ldots,k_n)^T$. $\mathcal{R}_x$ is called *the exponent lattice* of $x$. The elements of $\mathcal{R}_x\backslash\{\mathbf{0}\}$ are called *dependent vectors* of $x$. A $\mathbb{Z}$-independent finite set $\{\alpha_1,\ldots,\alpha_r\}\subset\mathcal{R}_x$ is called *a basis* of $\mathcal{R}_x$ if $\forall v\in\mathcal{R}_x, \exists k_1,\ldots,k_r\in\mathbb{Z}$ s.t. $v=\sum_{i=1}^r k_i\alpha_i$, where $\mathbb{Z}$-independent means $\ell_1\alpha_1+\cdots+\ell_r\alpha_r=\mathbf{0}$ implies $\ell_1=\cdots=\ell_r=0$ for integers $\ell_1,\ldots,\ell_r$. Main results {#sec:main} ============ Main Theorems for Constructing a Basis -------------------------------------- For a vector $v=(z_1,\ldots,z_n)^T\in \mathbb{C}^n$, if $1\leq k\leq n$, denote $v|k= (z_1,\ldots,z_k)^T$ and $v(k)$ the coordinate of $v$ indexed by $k$. Denote $[j]=\{1,\ldots,j\}$ for integer $j\geq 1$. Set $I\subset[n]$, denote by $\mathbb{Z}^I$ the lattice consists of all integer vectors whose coordinates are indexed by $I$. In addition, we simplify the notation $\mathbb{Z}^{[j]}$ to $\mathbb{Z}^j$. For any two subsets $I\subset J\subset [n]$, each vector in $\mathbb{Z}^I$ is considered to be a vector in $\mathbb{Z}^J$ with extra coordinates, indexed by $J\backslash I$, equal to $0$. This ambiguity does not lead to any troubles but brings us convenience. For $x=(x_1,\ldots,x_n)^T$$\in(\overline{\mathbb{Q}}^*)^n$ and $v\in\mathbb{Z}^{I}$, define $x^v=\prod_{i\in I}x_i^{v(i)}$. This value remains the same if one considers $v$ to be a vector in $\mathbb{Z}^J$, for any $J\supset I$, and computes as $x^v=\prod_{j\in J}x_j^{v(j)}$. \[bv\] For $x\in(\overline{\mathbb{Q}}^*)^n$ and each $j\in[n]$, define [(i)]{} $V_j=\{v\in\mathbb{Z}^j| v(j)>0,\; x^v=1\}$, [(ii)]{} $J=\{j\in [n]|V_j\neq\emptyset\}$, [(iii)]{} $\textbf{\emph{min}}_j=\min_{v\in V_j} v(j)$ and [(iv)]{} $\tilde{V}_j=\{v\in V_j|v(j)=\textbf{\emph{min}}_j\}$ for each $j\in J$. Then $\tilde{V}_j\neq\emptyset$ for any $j\in J$. A set of vectors of the form $ \{u_j\}_{j\in J}$, where $u_j\in\tilde{V}_j$ for each $j\in J$, is called a set of *basis vectors*. The name is justified by Theorem \[basisthm\] that is to come. \[divide\] $\forall j\in J$, $\forall v \in V_j$, ***min***$_j=u_j(j)$ divides $v(j)$. Write $v_j(j)= q\cdot \textbf{min}_j+r$, where $0\leq r<\textbf{min}_j$. Assume that $0<r<\textbf{min}_j$, then $(v_j-qu_j)(j)=r>0$ and $(v_j-qu_j)\in V_j$. Then $r<\textbf{min}_j$ contradicts Definition \[bv\] (iii). In the following, we fix a certain set of basis vectors $\{u_j\}_{j\in J}$ defined in Definition \[bv\]. Define inductively $n+1$ sets of integer vectors as follows: $$\begin{array}{l} {B_0} = \emptyset ,\\ {B_j} = \left\{ {\begin{array}{*{20}{c}} {{B_{j - 1}},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{\rm{if}}\;{V_j} = \emptyset},\\ {{B_{j - 1}} \cup \{ {u_j}\} ,\;\;\;\;{\rm{if}}\;{V_j} \ne \emptyset }, \end{array}} \right. \end{array} \label{Basis}$$ for each $j\in[n]$. We have the following main theorem: \[basisthm\] Consider $B_n$ to be a subset of$\;\mathbb{Z}^n$, then $B_n=\{u_j\}_{j\in J}$ is a basis of the exponent lattice $\mathcal{R}_x$. If $B_n=\emptyset$, then $\mathbb{R}_x=\{\mathbf{0}\}$ and the conclusion holds. Define $\textbf{Tail}(v)=\max\{i\in[n]|v(i)\neq0\}$. We use induction on $\textbf{Tail}(v)$ to prove that $\forall v\in\mathcal{R}_x\backslash\{\mathbf{0}\}, v\in\textbf{span}(B_n)$. If $\textbf{Tail}(v)=1$, $x_1$ is a root of unity and $\textbf{Order}(x_1)|v(1)$. Thus $v(1)\in\textbf{span}\{u_1\}\subset B_n$. Suppose $\textbf{Tail}(v)=k\geq2$ and $\forall v\;(v\in\mathcal{R}_x\wedge\textbf{Tail}(v)<k)\Rightarrow v\in\textbf{span}(B_n)$. Then by Proposition \[divide\], $u_k(k)$ divides $|v(k)|$. Let $\tilde{v}= v-\frac{v(k)}{u_k(k)}\cdot u_k\in\mathcal{R}_x$, then $\textbf{Tail}(\tilde{v})<k$. Therefore $\tilde{v}\in\textbf{span}(B_n)$ and $v=\tilde{v}+v(k)/u_k(k)\cdot u_k\in\textbf{span}(B_n)$. Moreover, the vectors in $B_n$ are obviously $\mathbb{Z}$-independent. Theorem \[basisthm\] indicates how one constructs inductively a basis of $\mathcal{R}_x$. The point is to compute $u_j$. Before that, we digress a little to see the main idea to obtain a maximal independent sequence of the numbers $x_1,x_2,\ldots,x_n$. Set: $$\begin{array}{l} {S_0} = \epsilon ,\\ {S_j} = \left\{ {\begin{array}{*{20}{l}} {{S_{j - 1}},\;\;\;\;\;\;\;\;{\rm{if}}\;{S_{j - 1}},{x_j}\;{\rm{is\; dependent,}}}\;\\ {{S_{j - 1}},{x_j},\;\;{\rm{if}}\;{S_{j - 1}},{x_j}\;{\rm{is \;independent,}}} \end{array}} \right. \end{array}\label{Sj}$$ for each $j\in[n]$. Here $S_{j - 1},{x_j}$ is the sequence obtained by attatching $x_j$ to the tail of the sequence $S_{j-1}$. The sequence $\epsilon,x_j$, for instance, means $x_j$. Also we define for $j\in[n]$: $$\begin{array}{l} {I_0} = \emptyset,\\ {I_j} = \left\{ {\begin{array}{*{20}{l}} {{I_{j - 1}},\;\;\;\;\;\;\;\;\;\;\;\;\;{\rm{if}}\;{S_{j - 1}},{x_j}\;{\rm{is\; dependent,}}}\\ {{I_{j - 1}}\cup\{j\},\;\;{\rm{if}}\;{S_{j - 1}},{x_j}\;{\rm{is \;independent.}}} \end{array}} \right. \end{array} \label{Ij}$$ \[misthm\] $S_n$ is a maximal independent sequence of $x_1,\ldots,x_n$. The following proposition connects Theorem \[basisthm\] and \[misthm\]. \[connection\] $V_j\neq\emptyset\Leftrightarrow$ the sequence $S_{j-1},x_{j}$ is dependent. “$\Rightarrow$ Set $v=(-k_1,-k_2,\ldots,-k_{j-1},k_j)^T$$\in V_j$, *s.t.* $k_j>0$ and $x^v=1$, thus $x_j^{k_j}=x_1^{k_1}x_2^{k_2}\cdots x_{j-1}^{k_{j-1}}$. Set $ {S_{j - 1}}\text{to be }{x_{{i_1}}},{x_{{i_2}}}, \cdots ,{x_{{i_m}}}, $ as defined in (\[Sj\]). By Theorem \[misthm\] and Proposition \[pmr\], each of $x_1,\ldots,x_{j-1}$ can be pseudo-multiplicatively represented by $S_{j-1}$: $$\left\{ {\begin{array}{l} \;\;\;{x_1^{{r_1}} = x_{i_1}^{{a_{11}}} \cdots x_{i_m}^{{a_{m1}}}},\\ \;\;\;\;\;\;\;\;\;\; \vdots \\ {x_{j-1}^{{r_{j-1}}} = x_{i_1}^{{a_{1,{j-1}}}} \cdots x_{i_m}^{{a_{m,{j-1}}}}}, \end{array}} \right.$$ where $r_1,r_2,\ldots,r_{j-1}\in\mathbb{Z}^*$, $a_{st}\in\mathbb{Z}$, for $s\in[m],\;t\in[j-1]$. Define $\Pi=r_1 r_2\cdots r_{j-1}$ and $\pi_{t}=\Pi/r_{t}$, then $$\begin{array}{l} x_j^{\Pi \cdot {k_j}} = x_1^{\Pi \cdot {k_1}}x_2^{\Pi \cdot {k_2}} \cdots x_{j - 1}^{\Pi \cdot {k_{j - 1}}}\\ \;\;\;\;\;\;\;\;\;\;{\rm{ = }}\;x_1^{{r_1} \cdot {k_1} \cdot {\pi _1}}x_2^{{r_2} \cdot {k_2} \cdot {\pi _2}} \cdots x_{j - 1}^{{r_{j - 1}} \cdot {k_{j - 1}} \cdot {\pi _{j - 1}}}\\ \;\;\;\;\;\;\;\;\;\; = {(x_{{i_1}}^{{a_{11}}} \cdots x_{{i_m}}^{{a_{m1}}})^{{k_1} \cdot {\pi _1}}} \cdots {(x_{{i_1}}^{{a_{1,j - 1}}} \cdots x_{{i_m}}^{{a_{m,j - 1}}})^{{k_{j - 1}} \cdot {\pi _{j - 1}}}} \end{array}$$ with $\Pi\cdot k_j\neq0$, thus the sequence $S_{j-1},x_j$ is dependent. “$\Leftarrow$”: Since $S_{j-1}$ is multiplicatively independent, $S_{j-1}$ is a maximal independent sequence of $S_{j-1},x_j$. According to Proposition \[pmr\], $x_j$ can be pseudo-multiplicative represented by $S_{j-1}$: $x_j^{l} = x_{i_1}^{l_1} \cdots x_{i_m}^{{l_m}}$ with $0\neq l, l_1,\ldots,l_m\in\mathbb{Z}$. Define $v\in \mathbb{Z}^{I_{j-1}\cup\{j\}}$ by $v(i_s)=l_s,\;$for $ s\in[m]$ and $v(j)=-l$. Then $v$ or $-v\in V_j$, hence $V_j\neq\emptyset$. Proposition \[connection\] implies that $$(|B_j|-|B_{j-1}|)+(\textbf{length}(S_j)-\textbf{length}(S_{j-1}))=1$$ for any $j$, thus $|B_n|-|B_0|+\textbf{length}(S_n)-\textbf{length}(S_{0})=n.$ Now $|B_n|=\textbf{rank}(\mathcal{R}_x)$ by Theorem \[basisthm\], $\textbf{length}(S_n)=\textbf{rank}(x)$ by Theorem \[misthm\] and $|B_0|=\textbf{length}(S_{0})=0$, thus $\textbf{rank}(\mathcal{R}_x)+\textbf{rank}(x)=n$. We obtain the following theorem. For $x\in(\overline{\mathbb{Q}}^*)^n$, $\emph{\textbf{rank}}(\mathcal{R}_x)+\emph{\textbf{rank}}(x)=n$. Note that $I_{j-1}\cup\{j\}\subset [j]$, thus $\mathbb{Z}^{I_{j-1}\cup\{j\}}\subset\mathbb{Z}^j$ by convention. Define a subset of $V_j$ : $$U_j=V_j\cap\mathbb{Z}^{I_{j-1}\cup\{j\}}$$ for $j\in[n]$. It is clear, from the “$\Leftarrow$” part of the proof of Proposition \[connection\], that $U_j\neq\emptyset$ whenever the sequence $S_{j-1},x_j$ is dependent. Indeed, the vector $v$ or $-v$ constructed there is in $U_j$. Conversely, if $U_j\neq\emptyset$, any vector in $U_j$ provides a dependent vector for the sequence $S_{j-1},x_j$. Hence $U_j\neq\emptyset$ iff the sequence $S_{j-1},x_j$ is dependent, which is equivalent to $V_j\neq\emptyset$. One sees $\{j\in[n]\;|\;U_j\neq\emptyset\}=J$ (Definition \[bv\] (ii)). \[prebv\] A set of vectors of the form $ \{w_j\}_{j\in J}$, where $w_j\in U_j$ for each $j\in J$, is called a set of *pre-basis vectors*. In the following, we fix a certain pre-basis $\{w_j\}_{j\in J}$. Since $w_j\in U_j\subset V_j$, $u_j(j)$ divides $w_j(j)$ by Proposition \[divide\]. A Certificate for Multiplicative Independence --------------------------------------------- The degree of an algebraic number $\mathfrak{a}$ is denoted by $\textbf{deg}(\mathfrak{a})$. \[nondegen\] A sequence $\{x_i\}_{i=1}^n\subset\overline{\mathbb{Q}}^*$ is called *degenerate* if $[\mathbb{Q}[x_1,x_2,\ldots,x_n]:\mathbb{Q}]<\prod_{i=1}^n\textbf{\emph{deg}}(x_i)$, otherwise *non-degenerate*. \[suff\] If $\{x_i\}_{i=1}^n$$\subset\overline{\mathbb{Q}}^*$ is non-degenerate and $ \prod_{i\in[n]}x_i^{k_i}=1$ for integers $k_1,k_2,\cdots,k_n$, then $\forall i\in [n],\; x_i^{k_i}\in \mathbb{Q}.$ Suppose $\textbf{deg}(x_i)=d_i$ for $i\in[n]$. The non-degenerate condition implies that the set $ \{x_1^{l_1}x_2^{l_2}\cdots x_n^{l_n}|\forall i \in [n],\;0\leq l_i\leq d_i-1\} $ is a basis of the $\mathbb{Q}$-linear space $\mathbb{Q}[x_1,x_2,\ldots,x_n]$. For each $i\in[n]$, a basis of the $\mathbb{Q}$-linear space $\mathbb{Q}[x_i]$ is $ \{1,x_i,x_i^2,\ldots,x_i^{d_i-1}\}. $ Expanding $x_i^{k_i}$ along that basis we obtain: $$x_i^{k_i}= a_{i0}+a_{i1}x_i+a_{i2}x_i^2+\cdots+a_{i,d_i-1}x_i^{d_i-1}, \label{span}$$ where $a_{ij}\in \mathbb{Q}$, $i\in[n]$, $j+1\in[d_i]$. Then $\prod_{i\in[n]}x_i^{k_i}=1$ means: $$\sum_{\forall i\in [n],\;0\leq l_i \leq d_i-1}a_{1l_1}a_{2l_2}\cdots a_{nl_n}x_1^{l_1}x_2^{l_2}\cdots x_n^{l_n}=1. \label{compare}$$ Since $\{x_1^{l_1}x_2^{l_2}\cdots x_n^{l_n}|\forall i \in [n], l_i+1\in[d_i]\}$ is a basis, two sides of (\[compare\]) have exactly the same coefficients, [*i.e.*]{} $$\left\{ {\begin{array}{*{20}{c}} a_{10}a_{20}\cdots a_{n0}=1,\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\\ a_{1l_1}a_{2l_2}\cdots a_{nl_n}=0,\;\text{if}\; \exists i\in[n],\;l_i>0, \end{array}} \right.\label{coe}$$ Consider $a_{ij}\in\mathbb{Q}$ with $i\in[n]$, $j \in[d_i-1]$, then $a_{10}a_{20}\cdots a_{i-1,0}a_{ij}\cdot$\ $a_{i+1,0}\cdots a_{n0}=0$ by the $2^{\text{nd}}$ equality of (\[coe\]). Thus $a_{ij}=0$ follows from the $1^{\text{st}}$ equality of (\[coe\]). Then (\[span\]) is reduced to $x_i^{k_i}=a_{i0}\in \mathbb{Q}$. \[rorn\] Nonzero algebraic number $\mathfrak{a}$ is a *root of rational* if $\exists k\in\mathbb{Z},k>0,\mathfrak{a}^k\in \mathbb{Q}$. The smallest such $k$ is *the rational order* of $\mathfrak{a}$, denoted by $\textbf{\emph{Rorder}}(\mathfrak{a})$. For $\mathfrak{a}$ not a root of rational, define $\textbf{\emph{Rorder}}(\mathfrak{a})=0$. For convenience, define $\textbf{\emph{R}}(\mathfrak{a})= \mathfrak{a}^{\textbf{\emph{Rorder}}(\mathfrak{a})}$. [*Note:*]{} If $\mathfrak{a}^\ell\in\mathbb{Q}$ for integer $\ell$, then $\textbf{Rorder}(\mathfrak{a})$ divides $\ell$. \[suffcor\] If $x_1,\ldots,x_n$ satisfy non-degenerate condition in Definition \[nondegen\] and none of them is a root of rational, then $x_1,\ldots,x_n$ are multiplicatively independent. Set $x_1^{k_1}\cdots x_n^{k_n}=1$ for integers $k_j$. By Theorem \[suff\], $\forall i\in[n]$, $x_i^{k_i}\in \mathbb{Q}$. Since $x_i$ is not a root of rational, $k_i=0,\forall i\in[n]$. More generally, set $$P=\{i\in[n]\;|\;x_i \text{ is a root of rational}\}=\{i_1,i_2,\ldots,i_s\}.\label{reducedequation}$$ Suppose the non-degenerate condition holds, then $\prod_{i\in[n]}x_i^{k_i}=1$ implies $k_i=0$ for $i\notin P$. The problem is reduced to solving the equation $x_{i_1}^{k_{i_1}}x_{i_2}^{k_{i_2}}\cdots x_{i_s}^{k_{i_s}}=1$. In fact, $x_{i_1},x_{i_2},\ldots, x_{i_s}$ are dependent iff $\textbf{R}(x_{i_1}),\textbf{R}(x_{i_2}),\ldots,\textbf{R}(x_{i_s})$ (Definition \[rorn\]) are. More generally, for each $i\in [n]$, let $x_i^{{1}/{q_i}}$ be a certain $q_i$-th root of $x_i$. Formally define $x_i^{{p_i}/{q_i}}= (x_i^{{1}/{q_i}})^{p_i}$, for integers $q_i>0,p_i\neq 0$, then \[scale\] $x_1^{{p_1}/{q_1}},x_2^{{p_2}/{q_2}},\ldots, x_n^{{p_n}/{q_n}}$ are multiplicatively dependent $\Leftrightarrow x_1,x_2,\ldots,x_n$ are multiplicatively dependent. *“$\Leftarrow$”:* Set $(k_1,k_2,\ldots,k_n)^T\in \mathbb{Z}^n\backslash\{\mathbf{0}\}$ and $x_1^{k_1}x_2^{k_2}\cdots x_n^{k_n}=1$. Define $p=p_1p_2\cdots p_n$ and $\tilde{p}_i=p/p_i$, then $p=p_i\tilde{p}_i$ and $$(x_1^{k_1}x_2^{k_2}\cdots x_n^{k_n})^p=x_1^{p_1\tilde{p}_1k_1}x_2^{p_2\tilde{p}_2k_2}\cdots x_n^{p_n\tilde{p}_nk_n}=1.$$ Thus $(x_1^{{p_1}/{q_1}})^{q_1\tilde{p}_1k_1}(x_2^{{p_2}/{q_2}})^{q_2\tilde{p}_2k_2}\cdots(x_n^{{p_n}/{q_n}})^{q_n\tilde{p}_nk_n}=1.$ Since $\exists i\in[n], k_{i}\neq0, q_{i}>0$ and $\tilde{p}_{i}\neq 0$, $q_{i}\tilde{p}_{i}k_i\neq0$. We find a dependent vector for $x_1^{{p_1}/{q_1}},x_2^{{p_2}/{q_2}},\ldots, x_n^{{p_n}/{q_n}}$. *“$\Rightarrow$”:* Set $(k_1,k_2,\ldots,k_n)^T\in \mathbb{Z}^n\backslash{\{\mathbf{0}\}}$, such that $$(x_1^{p_1/q_1})^{k_1}(x_2^{p_2/q_2})^{k_2}\cdots (x_n^{p_n/q_n})^{k_n}=1.\label{relation}$$ Define $q= q_1q_2\cdots q_n$ and $\tilde{q}_i=q/q_i$, then $q=q_i\tilde{q}_i$. Taking power to $q$ for both sides of (\[relation\]), we have $$(x_1^{1/q_1})^{q_1\tilde{q}_1p_1k_1}(x_2^{1/q_2})^{q_2\tilde{q}_2p_2k_2}\cdots (x_n^{1/q_n})^{q_n\tilde{q}_np_nk_n}=1.$$ Thus $x_1^{\tilde{q}_1p_1k_1}x_2^{\tilde{q}_2p_2k_2}\cdots x_n^{\tilde{q}_np_nk_n}=1$. Suppose $k_i\neq0$ for an $i\in[n]$. Note that $\tilde{q}_{i}>0$ and $p_{i}\neq0$, thus $\tilde{q}_{i}p_{i}k_{i}\neq0$. We obtain a dependent vector for $x_1,\ldots,x_n$. In the proof of Proposition \[scale\], from a dependent vector for any $x_1^{{p_1}/{q_1}},x_2^{{p_2}/{q_2}},\ldots, x_n^{{p_n}/{q_n}}$, one recovers a dependent vector for the original numbers $x_1,x_2,\ldots,x_n$. In particular, one recovers a dependent vector for $\{x_i\}_{i\in P}$ (defined in (\[reducedequation\])) from a dependent vector for $\textbf{R}(x_{i_1}),\textbf{R}(x_{i_2}),\ldots,\textbf{R}(x_{i_s})$, which are all rational numbers. Computing multiplicatively dependent vectors for rational numbers is equivalent to solving linear Diophantine equations. We show this in the following by an example. \[rationalcase\] Set $x_1=\frac{21}{4},x_2=\frac{27}{50},x_3=\frac{245}{32},x_4=\frac{16}{7}$. We solve the equation $x_1^{k_1}x_2^{k_2}x_3^{k_3}x_4^{k_4}=1$ with unknown integers $k_i$. One factors those rationals and collects factors sharing a same base: $$\begin{array}{l} \quad1=x_1^{k_1}x_2^{k_2}x_3^{k_3}x_4^{k_4}\\ \\ \quad\;\;\;=\Big(\frac{21}{4}\Big)^{k_1}\cdot\Big(\frac{27}{50}\Big)^{k_2}\cdot\Big(\frac{245}{32}\Big)^{k_3}\cdot\Big(\frac{16}{7}\Big)^{k_4}\\ \\ \quad\;\;\;=\Big(\frac{3\cdot7}{2^2}\Big)^{k_1}\cdot\Big(\frac{3^3}{2\cdot5^2}\Big)^{k_2}\cdot\Big(\frac{5\cdot7^2}{2^5}\Big)^{k_3}\cdot\Big(\frac{2^4}{7}\Big)^{k_4}\\ \\ \quad\;\;\;=2^{-2k_1-k_2-5k_3+4k_4}\cdot3^{k_1+3k_2}\cdot5^{-2k_2+k_3}\cdot7^{k_1+2k_3-k_4}. \end{array}\label{above}$$ By the fundamental theorem of arithmetic, any vector $(l_1,\ldots,l_4)^T$ in $\mathbb{Z}^4\backslash\{\mathbf{0}\}$ satisfies $2^{l_1}3^{l_2}5^{l_3}7^{l_4}\neq1$. Hence (\[above\]) is equivalent to $$\left\{ {\begin{array}{*{20}{c}} -2k_1&-&k_2&-&5k_3&+&4k_4&=&0,\\ k_1&+&3k_2&&&&&=&0,\\ &-&2k_2&+&k_3&&&=&0,\\ k_1&&&+&2k_3&-&k_4&=&0, \end{array}} \right.$$ which has a unique solution $\mathbf{0}$. Hence $x_1,x_2,x_3,x_4$ are multiplicatively independent. Main algorithms {#sec:alg} =============== Preprocess for Algebraic Numbers -------------------------------- There are three ways to reduce the degrees of the algebraic numbers that we are going to deal with. First, we use a function **RootOfUnityTest** to decide whether a nonzero algebraic number is a root of unity and return its order if it is. Second, **RootOfRationalTest** in Algorithm \[algror\] is a function to decide whether a nonzero algebraic number is a root of rational and return its rational order. Finally, the function **DegreeReduction** in Algorithm \[algdegred\] is developed in order that for a given $\mathfrak{a}\in\overline{\mathbb{Q}}^*$, one finds an integer $q \geq 1$ *s.t.* $\textbf{deg}(\mathfrak{a}^q)=\min_{k\in \mathbb{Z},k\geq1}\textbf{deg}(\mathfrak{a}^k)$ and the minimal polynomial of $\mathfrak{a}^q$. \[reddegree\] The *reduced degree* of $\mathfrak{a}\in\overline{\mathbb{Q}}^*$ is defined by $ \textbf{\emph{rdeg}}(\mathfrak{a})$\ $=\min_{k\in \mathbb{Z},k\geq1}\textbf{\emph{deg}}(\mathfrak{a}^k). $ *The set of reducing exponents* of $\mathfrak{a}$ is given by $ \textbf{\emph{Rexp}}(\mathfrak{a})=\{q\in\mathbb{Z}^*|\textbf{\emph{deg}}(\mathfrak{a}^q)=\textbf{\emph{rdeg}}(\mathfrak{a})\} $. We say $\mathfrak{a}$ is *degree reducible* if $\textbf{\emph{rdeg}}(\mathfrak{a})<\textbf{\emph{deg}}(\mathfrak{a})$, otherwise *degree irreducible*. Set $\mathfrak{a}=(\sqrt{5}-2)e^{\frac{\pi \sqrt{-1}}{3}}$ with minimal polynomial $f(t)=t^4 + 4 t^3 + 17 t^2 - 4 t + 1$. Then ***deg***$(\mathfrak{a}^3)=2<4=\textbf{\emph {deg}}(\mathfrak{a})$. Hence $\mathfrak{a}$ is degree reducible. We permute algebraic numbers $x_1,x_2,\ldots,x_n$ as follows $$\mathfrak{x}^T=(\mathfrak{x}_1,\ldots,\mathfrak{x}_n)=(\alpha_1,\ldots,\alpha_{r},\beta_1,\ldots,\beta_s,\gamma_1,\ldots,\gamma_m) \label{rearrange}$$ *s.t.* $\{\alpha\}_{i=1}^r$ are roots of unity, $\{\beta\}_{i=1}^s$ are roots of rational but none of which is a root of unity and none of $\{\gamma\}_{i=1}^m$ is a root of rational. Suppose we find $p_i\in\textbf{Rexp}(\gamma_i)$, $p_i\geq1$ for $i\in[m]$. Then we deal with (reduced) algebraic numbers: $$(y_1,\ldots,y_n)=(1_1,\ldots,1_r,\textbf{R}(\beta_1),\ldots,\textbf{R}(\beta_s),\gamma_1^{p_1}\cdots,\gamma_m^{p_m}). \label{reduced}$$ Set $\tilde{\gamma}_i=\gamma_i^{p_i}$, then none of $\{\tilde{\gamma}_i\}_{i=1}^m$ is a root of rational. And assume that $\{\tilde{\gamma}_i\}_{i=1}^t$, for some $t\leq m$, satisfy the non-degenerate condition in Definition \[nondegen\]. Here we ensure the non-degenerate condition by checking the equality $ \textbf{deg}(\tilde{\gamma_1}+\tilde{\gamma_2}+\cdots+\tilde{\gamma_t})=\prod_{i=1}^{t}\textbf{deg}(\tilde{\gamma_i}). $ By Corollary \[suffcor\], $\{\tilde{\gamma}\}_{i=1}^t$ are multiplicatively independent. More generally, we have the following observation: \[ind\] Suppose $\{\delta_i\}_{i=1}^l\subset\{\textbf{\emph{R}}(\beta_i)\}_{i=1}^s$ is a multiplicatively independent sequence. Algebraic numbers $\tilde{\gamma}_1,\ldots,\tilde{\gamma}_t$ satisfy the non-degenerate condition, none of which is a root of rational. Then the sequence $\delta_1,\ldots,\delta_l,\tilde{\gamma}_1,\ldots,\tilde{\gamma}_t$ is multiplicatively independent. Since $\delta_i$ are rational, $\delta_1,\ldots,\delta_l,\tilde{\gamma}_1,\ldots,\tilde{\gamma}_t$ satisfy the non-degenerate condition. Set $ \delta_1^{k_1}\cdots \delta_l^{k_l}\tilde{\gamma}_1^{m_1}\cdots\tilde{\gamma}_t^{m_t}=1 $ with integer exponents. Then $\tilde{\gamma}_i^{m_i}\in\mathbb{Q}$ by Theorem \[suff\]. Since $\tilde{\gamma}_i$ is not a root of rational, $m_i=0$ for $i\in[t]$. Then $ \delta_1^{k_1}\cdots \delta_l^{k_l}=1 $ with $\delta_1,\ldots,\delta_l$ multiplicatively independent, so $k_i=0$ for $i\in[l]$. Algorithm \[getbasis\] is the main algorithm of this paper, its key parts are Algorithm \[alggetpre-basis\] (**GetPreBasis**) and Algorithm \[pre-basis2basis\] (**PreBasis2Basis**). The function **Isomorphism** in Step 7 is given by Algorithm \[iso\]. Preprocess $(x_1,x_2,\ldots,x_n)$ to obtain: $\mathfrak{x}$ in (\[rearrange\])$, (y_i)_{i=1}^n$ in (\[reduced\]), {**Order**$(\alpha_i)\big\}_{i=1}^r$, $\big\{\textbf{Rorder}(\beta_i)\big\}_{i=1}^s$ and $\big\{p_i\in\textbf{Rexp}(\gamma_i)\big\}_{i=1}^m$ ;\ \[getbasis\] $\{PreBasis,I\}=\textbf{ GetPreBasis }\;\;\;\;\;\;\;\;\;\;\;\;\;\;$ $\big((y_i)_{i=1}^n,\begin{small}\{\textbf{Order}(\alpha_i)\}_{i=1}^r,\{\textbf{Rorder}(\beta_i)\}_{i=1}^s,\end{small}\{p_i\}\begin{small}_{i=1}^m\end{small}\big);$\ $J=[n]\backslash I$;\ **if** $(J==\emptyset)$ **{Return** $\{\emptyset,[n]\}$;**} endif**\ Set $J=\{j_1,j_2,\cdots\}$, $j_1<j_2<\cdots$, $PreBasis=\{w_{j_1},w_{j_2},\cdots\};$\ $g=\text{GCD}(w_{j_1}(1),w_{j_1}(2),\ldots,w_{j_1}(j_1))$;\ $a=\textbf{Isomorphism}(\mathfrak{x},w_{j_1}/g,g);$ $u_{j_1}=w_{j_1}/\text{GCD}(a,g);\;$(\[ini\])\ $Basis=\{u_{j_1}\}$;\ Since algebraic numbers are permuted in (\[rearrange\]), re-express each $u_j$ and $I$ in the original indices;\ **Return** $\{Basis, I\}$ Constructing a Set of Pre-Basis Vectors --------------------------------------- In the following, all of $B_j,S_j,I_j,V_j,U_j,J,\tilde{V}_j,\textbf{min}_j,u_j,w_j$ are defined for $\mathfrak{x}$ in (\[rearrange\]) instead of $x=(x_1,\ldots,x_n)^T$. We introduce Algorithm \[alggetpre-basis\] to compute $J$ (Definition \[bv\] (ii)) and extract pre-basis vectors $w_j\in U_j$ (Definition \[prebv\]) for each $j\in J$, as preparation for computing a basis $\{u_j\}_{j\in J}$. Moreover, the set $I$ in Algorithm \[alggetpre-basis\] is successively equal to $I_0,I_1,\ldots,I_{n}$ defined in (\[Ij\]) as the algorithm runs. Algorithm \[alggetpre-basis\] deals with algebraic numbers in (\[reduced\]) while returning dependent vectors $w_j$ for $\mathfrak{x}$ in (\[rearrange\]). This is by recovering technique mentioned in the proof of Proposition \[scale\]. Algorithm \[alggetpre-basis\] works in the spirit of formula (\[Ij\]). Moreover, a dependent vector $w_j$ is obtained if the multiplicative dependence condition holds in (\[Ij\]). Steps 2-5 deal with those roots of unity, while Steps 6-15 process roots of rational. Step 16 adds those numbers of non-degenerate property to the set $I$ by Proposition \[ind\]. Finally, Steps 17-26 cope with the rest numbers that are not root of unity and failed in the non-degenerate condition test. For designing the function **DecideDependence**, we need results in [@van1977multiplicative; @loxton1983multiplicative; @masser1988linear]. Theorem 1 of [@van1977multiplicative], Theorem 3 of [@loxton1983multiplicative] and Theorem $G_m$ of [@masser1988linear] are of the same form as follows: \[bnd\] Set $x\in(\overline{\mathbb{Q}}^*)^n$ to be multiplicatively dependent, then there is an efficiently computable number $\emph{\textbf{bnd}}$ s.t. $\exists v\in \mathcal{R}_x\backslash\{\mathbf{0}\}$, $\|v\|_\infty\leq \emph{\textbf{bnd}}$. The number **bnd** is given by corresponding theorems in [@van1977multiplicative; @loxton1983multiplicative; @masser1988linear]. Other similar theorems can be found in [@loher2004uniformly; @tori; @mat]. By definition $S_{j-1}$ is multiplicatively independent. If the sequence $S_{j-1},\mathfrak{x}_j$ is multiplicatively dependent, by Theorem \[bnd\] it follows that $\exists v\in\mathbb{Z}^{I_{j-1}\cup\{j\}}\backslash\{\mathbf{0}\},$ *s.t.* $\mathfrak{x}^v=1,\|v\|_\infty<\textbf{bnd} $ and $v(j)>0$. Function **DecideDependence** accepts a sequence of nonzero algebraic numbers $b_1,b_2,\ldots,b_\iota,b_{\iota+1}$, with$\{b_j\}_{j=1}^\iota$ multiplicatively independent, as its input. By exhaustive search in the box defined by **bnd**, **DecideDependence** returns $\{\textbf{True}, v\}$ if it finds a dependent vector $v$, $\{\textbf{False},\mathbf{0}\}$ otherwise. **DecideDependence** can be replaced by the function **FindRelations** in [@kauers2005algorithms] $\S7.3$ which accepts an $x\in(\overline{\mathbb{Q}}^*)^n$ as its input and returns a basis of $\mathcal{R}_x$. In this case, it returns one basis vector or $\emptyset$. $J=\emptyset$; $I=\emptyset$; $j=1$; $\text{$PreBasis$}=\emptyset$;\ **if** [$(s>0)$]{} **{**[$I=I\cup\{r+1\};$**} endif**\ ]{} $I=I\cup\{r+s+1,r+s+2,\ldots,r+s+t\}$;\ **Return** $\{PreBasis,I\}$ Recovering Basis from Pre-Basis ------------------------------- We recover a basis $\{u_j\}_{j\in J}$ (Definition \[bv\]) from the pre-basis $\{w_j\}_{j\in J}$ (Definition \[prebv\]). Algorithm \[alggetpre-basis\] obtains $J=\{j\in[n]\;|\;V_j\neq\emptyset\}=\{j\in[n]\;|\;U_j\neq\emptyset\}$ on which $u_j$ and $w_j$ are defined. We compute $u_j$ inductively. That is, if $J=\{j_1,j_2,\ldots,j_r\}$ then we compute $u_{j_2}$ based on $u_{j_1}$, compute $u_{j_3}$ based on $u_{j_1}$ and $u_{j_2}$, and so forth. ### Recovering a Basis Vector To begin with, we need the first basis vector $u_{j_1}$. By Theorem \[basisthm\], $w_{j_1}=\tau u_{j_1}$ for an integer $\tau>0$. Set $g= \text{GCD}(w_{j_1}(1),w_{j_1}(2),$\ $\ldots,w_{j_1}(j_1))$. Consider $\tilde{w}= w_{j_1}/g$, then $u_{j_1}=k \tilde{w}$ for an integer $k>0$. By the definition of $u_j$, $k= \min K$, where $ K=\{q\in\mathbb{Z}\;|\;q>0, \mathfrak{x}^{q\cdot\tilde{w}}=1\}. $ Noting that $(\mathfrak{x}^{\tilde{w}})^g=\mathfrak{x}^{w_{j_1}}=1$, we compute by Algorithm \[iso\] an integer $0\leq a<g$ *s.t.* $\mathfrak{x}^{\tilde{w}}=e^{2a\pi\sqrt{-1}/g}$. Then $\mathfrak{x}^{q\cdot\tilde{w}}=1\Leftrightarrow g|q\cdot a. $ Thus $$\begin{array}{rcl} K&=&\{q\in\mathbb{Z}\;|\;q>0,qa\text{ is a multiple of }g\}\\ &=&\{q\in\mathbb{Z}\;|\;q>0,qa\text{ is a common multiple of }a\text{ and }g\}. \end{array}$$ Hence if $a>0$, $\min K={\text{LCM}(a,g)}/{a}={g}/{\text{GCD}(a,g)}$. If $a=0$ then $\mathfrak{x}^{\tilde{w}}=1$ and $\min K=1=g/\text{GCD}(a,g)$. In a nutshell $$u_{j_1}={g}/{\text{GCD}(a,g)}\cdot \tilde{w}={w_{j_1}}/{\text{GCD}(a,g)}. \label{ini}$$ Suppose $u_{j_1},\ldots,u_{j_k}$ are obtained, we compute a vector in $\tilde{V}_{j_{k+1}}$ (Definition \[bv\] (iv)) to be $u_{j_{k+1}}$. Set $\tau= {w_{j_{k+1}}(j_{k+1})}/{\textbf{min}_{j_{k+1}}}$$\in \mathbb{Z}$ and $\Lambda=\{\lambda\in[w_{j_{k+1}}(j_{k+1})]|\lambda$ divides $w_{j_{k+1}}(j_{k+1})\}.$ For each $\lambda\in\Lambda$, define an equation $$\textbf{E}_\lambda:\left\{ {\begin{array}{*{20}{c}} v({j_{k + 1}}) &=& \frac{{{w_{{j_{k + 1}}}}({j_{k + 1}})}}{\lambda }&\;\;\text{(a)}\\ {{\bar w}_{{j_{k + 1}}}} - \lambda \bar v &=& \sum\limits_{\iota = 1}^k {{q_\iota }} {u_{{j_\iota }}}&\;\;\text{(b)}\\ {\mathfrak{x}^v} &=& 1&\;\;\text{(c)} \end{array}} \right.\label{E}$$ with $v\in\mathbb{Z}^{j_{k+1}}$ unknown vector and $q_\iota$ unknown integers. Here $\overline{v}= v|(j_{k+1}-1)$ and $\overline{w}_{j_{k+1}}= w_{j_{k+1}}|(j_{k+1}-1)$ are defined at the beginning of $\S$3.1. \[equi\] For $\lambda\in\Lambda$, the following conditions are equivalent : *(i)* there are integers $\{q_\iota\}_{\iota=1}^k$, s.t. $\{v,q_\iota\}$ is a solution to **E**$_\lambda$, *(ii)* $v\in V_{j_{k+1}}$and $v(j_{k+1})=w_{j_{k+1}}$$(j_{k+1})/\lambda$. \(i) $\Rightarrow$ (ii) is trivial. Suppose (ii) holds, then so do (\[E\]) and (\[E\]). Since $(w_{j_{k+1}}-\lambda v)(j_{k+1})=0, (\overline{w}_{j_{k+1}}-\lambda \overline{v})\in \mathcal{R}_{\mathfrak{x}|(j_{k+1}-1)}=\textbf{span}\{u_{j_1},u_{j_2},\ldots,u_{j_k}\}$ by Theorem \[basisthm\]. So (\[E\]) holds for some $q_1,q_2,\ldots,q_k\in\mathbb{Z}$, thus (i) follows. Consider Proposition \[equi\] when $\lambda=\tau\in\Lambda$. Since $\tilde{V}_{j_{k+1}}\neq \emptyset$, any $v\in\tilde{V}_{j_{k+1}}$ satisfies condition (ii). Thus $v$ satisfies condition (i) as well. So $\textbf{E}_\tau$ has a solution. Define $\tilde{\Lambda}=\{\lambda\in\Lambda\;|\;\textbf{E}_\lambda \text{ has a solution}\}$ and $\lambda_m=\max \tilde{\Lambda}$, we indicate that $\tau=\lambda_m$. \[maxlambda\] Set $\lambda_m=\max \tilde{\Lambda}$, then $\lambda_m=\tau=\frac{w_{j_{k+1}}(j_{k+1})}{\textbf{\emph{min}}_{j_{k+1}}}$. Each solution to **E**$_{\lambda_m}$ can be projected to a vector in $\tilde{V}_{j_{k+1}}$. Since $\textbf{E}_{\lambda_m}$ has a solution, by Proposition \[equi\], $\exists v\in V_{j_{k+1}}$ *s.t.* $v(j_{k+1})=w_{j_{k+1}}(j_{k+1})/\lambda_m$. Then $\textbf{min}_{j_{k+1}}\leq w_{j_{k+1}}(j_{k+1})/\lambda_m$ by Definition \[bv\] (iii). This is $\tau\geq\lambda_m$. Since $\textbf{E}_\tau$ has a solution, $\tau\in\tilde{\Lambda}$. Hence $\tau\leq\lambda_m$. Suppose $\{v,q_\iota\}$ is a solution to $\textbf{E}_{\lambda_m}$. By Proposition \[equi\] its projection $v\in V_{j_{k+1}}$ and $v(j_{k+1})=w_{j_{k+1}}(j_{k+1})/\lambda_m=w_{j_{k+1}}(j_{k+1})/\tau=\textbf{min}_{j_{k+1}}$. Hence $v\in\tilde{V}_{j_{k+1}}$ by definition. In the spirit of Proposition \[maxlambda\], we arrange the numbers in $\Lambda$ decreasingly as: $\lambda^{(1)}>\lambda^{(2)}>\cdots.$ Then we solve a sequence of equations in the order: $\textbf{E}_{\lambda^{(1)}},\textbf{E}_{\lambda^{(2)}},\cdots.$ We stop as soon as some $\textbf{E}_{\lambda^{(i)}}$ has a solution (otherwise we move to the next one). Since $\textbf{E}_{\lambda^{(i)}}$ is the first set of equations to have a solution, $\lambda^{(i)}=\lambda_m$. Any solution to $\textbf{E}_{\lambda^{(i)}}$ can be projected to a vector in $\tilde{V}_{j_{k+1}}$, which we take as the value of $u_{j_{k+1}}$. ### Solving the Equation $\textbf{E}_\lambda$ First, solve the linear Diophantine equation (\[E\]). If (\[E\]) has no solutions, then neither has $\textbf{E}_\lambda$. Suppose (\[E\]) has general solution: $$\left( {\begin{array}{*{20}{c}} L\\ Q \end{array}} \right) = \left( {\begin{array}{*{20}{c}} {L_0}\\ {Q_0} \end{array}} \right) + {z_1}\left( {\begin{array}{*{20}{c}} {L_1}\\ {Q_1} \end{array}} \right) + \cdots + {z_s}\left( {\begin{array}{*{20}{c}} {L_s}\\ {Q_s} \end{array}} \right)\label{generalsolution}$$ where $L_i\in \mathbb{Z}^{j_{k+1}-1}$, $Q_i\in\mathbb{Z}^{k}$ and $z_i$ are any integers. $L$ gives the value of $\overline{v}$ while $Q$ gives the value of the vector $(q_1,q_2,\ldots,q_k)^T$. Moreover, $$\left\{ {\begin{array}{*{20}{c}} {{{\overline w}_{{j_{k + 1}}}} - \lambda {L_0} = \sum\limits_{\iota = 1}^k {{Q_0}(\iota )} {u_{{j_\iota }}}},\\ { - \lambda {L_i} = \sum\limits_{\iota = 1}^k {{Q_i}(\iota )} {u_{{j_\iota }}},\;{\rm{for}}\;i \in[s].} \end{array}} \label{solu}\right.$$ That is, $(L_0^T,Q_0^T)^T$ is a solution to (\[E\]) and $(L_i^T,M_i^T)^T$ form a basis of the solution lattice to the homogeneous version of (\[E\]). We concern ourselves with the problem whether there are integers $z_1,z_2,\ldots,z_s$ *s.t.* while taking $v=(L^T,{w_{j_{k+1}}(j_{k+1})}/{\lambda})^T$ as in (\[generalsolution\]), (\[E\]) holds. If there are, then $\textbf{E}_\lambda$ has a solution, otherwise it has not. In fact $$\begin{array}{rcl} \mathfrak{x}^v&=&{\mathfrak{x}^{{L_0} + {z_1}{L_1} + \cdots + {z_s}{L_s}}} \cdot \mathfrak{x}_{j_{k + 1}}^{{{{w_{{j_{k + 1}}}}({j_{k + 1}})}}/{\lambda }}\\ &=& \mathfrak{x}^{z_1L_1} \cdots \mathfrak{x}^{z_sL_s}\cdot (\mathfrak{x}^{L_0}\mathfrak{x}_{j_{k + 1}}^{ w_{j_{k + 1}}(j_{k + 1})/\lambda}). \end{array}$$ Since $\mathfrak{x}^{u_{j_\iota}}=1$, it follows from the 1$^\text{st}$ equation of (\[solu\]) that $\mathfrak{x}^{\lambda {L_0}}=\mathfrak{x}^{{{\overline w}_{{j_{k + 1}}}}}$. Moreover, $\mathfrak{x}^{{{\overline w}_{{j_{k + 1}}}}}\cdot \mathfrak{x}_{j_{k+1}}^{w_{j_{k+1}}(j_{k+1})}=1$ by definition, one obtains $$\begin{array}{rl} &{\big({{\mathfrak{x}^{{L_0}}}\mathfrak{x}_{j_{k + 1}}^{{w_{{j_{k + 1}}}}({j_{k + 1}})/\lambda }}\big)^\lambda }\\ = &{\mathfrak{x}^{\lambda {L_0}}}\mathfrak{x}_{j_{k + 1}}^{{w_{{j_{k + 1}}}}({j_{k + 1}})}\\ =& {\mathfrak{x}^{{{\overline w}_{{j_{k + 1}}}}}}\mathfrak{x}_{j_{k + 1}}^{{w_{{j_{k + 1}}}}({j_{k + 1}})}\\ = &1. \end{array}$$ For $L_i,i\in[s]$, using the 2$^\text{nd}$ equation of (\[solu\]), we have ${(\mathfrak{x}^{L_i})}^{\lambda}=\mathfrak{x}^{\lambda L_i}=1$. Now we observe that $$\Gamma_0=\mathfrak{x}^{L_0}\mathfrak{x}^{ w_{j_{k + 1}}(j_{k + 1})/\lambda}_{j_{k + 1}}\;\text{ and }\;\Gamma_i= \mathfrak{x}^{L_i}, i\in[s]\label{Gammai}$$ are roots of the equation $\Gamma^\lambda-1=0$. By using Algorithm \[iso\], for each $\Gamma_i$, $i\in\{0\}$$\cup[s]$, one obtains an integer $0\leq a_i<\lambda$ *s.t.* $\Gamma_i=e^{{2a_i\pi\sqrt{-1}}/{\lambda}}$. The set of all roots of the equation $\Gamma^\lambda-1=0$ is isomorphic to the group $\mathbb{Z}/\langle\lambda\rangle=\{0,1,\ldots,\lambda-1\}$ with addition modulo $\lambda$. The problem is reduced to whether there are integers $z_1,\ldots,z_s$ *s.t.* $\Gamma_1^{z_1}\Gamma_2^{z_2}\cdots \Gamma_s^{z_s}\Gamma_0=1$, which is equivalent to $\exists p\in\mathbb{Z}$, $$a_1z_1+\cdots +a_sz_s+a_0=p\lambda.\label{single}$$ This equation with $z_1,\ldots,z_s, p$ unknown integers can be efficiently solved. If (\[single\]) has no solutions, neither has $\textbf{E}_\lambda$. Otherwise we get the values of $z_i$ and obtain $L$ and $Q$ from (\[generalsolution\]). Then $\{v=(L^T,{w_{j_{k+1}}(j_{k+1})}/{\lambda})^T, (q_1,\ldots,q_k)=Q^T\}$ is a solution to $\textbf{E}_\lambda$. Finally, assign $u_{j_{k+1}}$$=(L^T,{w_{j_{k+1}}(j_{k+1})}/{\lambda})^T$. We summarize all these by Algorithm \[pre-basis2basis\]. Let $\lambda^{(1)}=w_{j_{k+1}}(j_{k+1})>\lambda^{(2)}>\cdots>\lambda^{(\gamma)}=1$ be all positive numbers that divide $w_{j_{k+1}}(j_{k+1})$;\ Denote a rectangle by $ [a,b;c,d]=\{z\in\mathbb{C}\;|\;a\leq\Re(z)\leq b,c\leq\Im(z)\leq d\}, $ call $a,b,c,d$ the coordinates of the rectangle $[a,b;c,d]$. By complex root isolation in [@isolation], we isolate each $x_j$ by small rectangles $R_j$ on the complex plane with rational coordinates. Assume in addition that $\forall j,0\not\in R_j$. For each $R_j$, efficiently chose an interval $\theta_j=[\underline\theta_j,\overline\theta_j]$ with $\underline\theta_j\leq\overline\theta_j$ rational numbers, *s.t.* $R_j\subset\{\rho e^{\pi\eta\sqrt{-1}}\in\mathbb{C}|\rho>0,\;\underline\theta_j\leq\eta\leq\overline\theta_j\}$. That can be done by standard method such as series expansion. Using interval arithmetic, we provide Algorithm \[iso\] for implementing the group isomorphism $\{\Gamma\;|\;\Gamma^\lambda=1\}\rightarrow\mathbb{Z}/\langle\lambda\rangle$. In Step 2, the condition “$\exists a,b\in\mathbb{Z}\cap[0,\lambda), a\neq b$ *s.t.* $\{2a/\lambda,2b/\lambda\}\subset\Theta$” is equivalent to “the length of $\Theta\geq 2/\lambda$”. The number “50%” makes sure that the length of each $\theta_j$ converges to $0$ (hence so does the length of $\Theta$) so that the algorithm terminates. $\Theta=\theta_1=\cdots=\theta_n=[0,2]$;\ **Return** the only integer $a$ so that $2a/\lambda\in\Theta$; Degree Reduction Functions in Preprocessing {#sec:red} =========================================== Recognizing Roots of Rational ----------------------------- \[necess\] For $p(t)=t^d+a_{d-1}t^{d-1}+\cdots+a_0\in\mathbb{R}[t]$ whose roots $z_1,\ldots,z_d$ are all of modulus $1$, $p(t)=\pm t^dp(1/t)$. Define $\tilde{p}(t)= t^dp(1/t)$. Roots of $\tilde{p}(t)$ are $(z_1^{-1},\ldots,z_d^{-1})=$\ $(\overline{z}_1,\ldots,\overline{z}_d)$. This is a re-permutation of $(z_1,\ldots,z_d)$. Thus $p(t)=$\ $c\cdot\tilde{p}(t), c\in\mathbb{C}\backslash\{0\}$. In fact $c=1/a_0$, and $|a_0|=\prod_{i\in[d]}|z_i|=1.$ \[samerorder\] For irreducible $p(t)\in\mathbb{Q}[t]$ with roots $z_1,\ldots,z_d$ $\in\mathbb{C}$, exactly one of the following conditions holds: (i) none of $z_1,\ldots,z_d$ is a root of rational; (ii) all of $z_1,\ldots,z_d$ are roots of rational, and $\textbf{\emph{R}}(z_1)=\textbf{\emph{R}}(z_j)$, $\textbf{\emph{Rorder}}(z_1)=\textbf{\emph{Rorder}}(z_j)$, $2\leq j\leq d$. Suppose $z_1$ is a root of rational of rational order $k$ and $\textbf{R}(z_1)$\ $=r$. Then $z_1^k-r=0$ and hence $p(t)|t^k-r$. Thus $p(z_j)=0$ implies $z_j^k-r=0$, $j=2,\ldots,d$. It follows that $\textbf{Rorder}(z_j)|k$. Exchanging the roles of $z_1$ and $z_j$, one sees that $k|\textbf{Rorder}(z_j)$. Hence $k=\textbf{Rorder}(z_j)$ and $\textbf{R}(z_j)=r$. Set $\mathfrak{a}$ to be a root of rational, $p(t)$ its monic minimal polynomial with $p(0)=a_0$. All complex roots of $p(t)$ are of the same modulus, say $\ell$, by Proposition \[samerorder\], then $|a_0|=\ell^d$. Define monic polynomial $ \overline{p}(t)= p(\sqrt[d]{|a_0|}\cdot t)/|a_0| \in \mathbb{R}[t]$ whose roots are all of modulus $1$. Then if the conclusion of Proposition \[necess\] fails to hold for $\overline{p}(t)$, one concludes that $\mathfrak{a}$ is not a root of rational. \[rorpro\] (Lemma 3.5 in [@lemma]) Let $F$ be a field, $E=F(\alpha)$, $[E:F]=d$ and $\alpha^m\in F$. If $f(t)$ is the monic irreducible polynomial of $\alpha$ over $F$, then $\zeta \alpha^d=(-1)^d f(0)\in F$ for some $\zeta $ s.t. $ \zeta^m=1$. Set $\alpha=\mathfrak{a}, F=\mathbb{Q}, E=\mathbb{Q}(\mathfrak{a})$ and $m=\textbf{Rorder}(\mathfrak{a})$, then $\zeta =(-1)^da_0/\mathfrak{a}^d$. By Proposition \[rorpro\], $\zeta^{\textbf{Rorder}(\mathfrak{a})}=1.$ Thus $\exists k_1\in\mathbb{Z}$ $k_1\cdot\textbf{Order}(\zeta)=\textbf{Rorder}(\mathfrak{a}).$ Note that $\zeta^{\textbf{Order}(\zeta)}=1$ means $\mathfrak{a}^{d\cdot\textbf{Order}(\zeta)}=((-1)^da_0)^{\textbf{Order}(\zeta)}\in\mathbb{Q}$. This implies $\exists k_2\in \mathbb{Z}$ *s.t.* (ii) $ k_2\cdot\textbf{Rorder}(\mathfrak{a})=d\cdot\textbf{Order}(\zeta). $ Combining (i) and (ii) we have $k_1\cdot k_2=d.$ These lead to Algorithm \[algror\] deciding whether an algebraic number, given its minimal polynomial, is a root of rational. Compute $\overline{p}(t)=p(\sqrt[d]{|a_0|}\cdot t)/|a_0|$;\ **if** $(\overline{p}(t)\neq \textbf{sgn}(a_0)\cdot t^d\overline{p}(1/t))$ **{Return** $\{0,1\}$;**} endif**\ $ f(t)=\textbf{MinimalPolynomial}(\mathfrak{a}^d/((-1)^da_0))$;\ $\{RootOfUnity, order\}=\textbf{RootOfUnityTest}(f(t))$;\ **if** (RootOfUnity$==\textbf{False}$) **{Return** $\{0,1\}$;**} endif**\ \[algror\] Arrange positive divisors of $d$ increasingly: $\lambda_1<\lambda_2<\cdots$;\ Computing Reduced Degree ------------------------ Set $\mathfrak{a}$ to be an algebraic number with minimal polynomial $p(t)$ whose complex roots are $\mathfrak{a}=z_1,z_2,\ldots,z_d$. For integer $m\geq 1$, if the minimal polynomial of $\mathfrak{a}^m$ is $f(t)$, then $\{z\in\mathbb{C}|f(z)=0\}=\{z_i^m\;|\;i\in[d]\}$. Define $\mathbb{E}=\mathbb{Q}(\mathfrak{a},z_2,\ldots,z_d)$, $G= \text{Gal}(\mathbb{E}/\mathbb{Q})$ and $G(\mathfrak{a}^m)=\{\sigma(\mathfrak{a}^m)|\sigma\in G\}=\{\beta_1=\mathfrak{a}^m,\beta_2,\ldots,\beta_r\}$. Then it is not hard to see that $f(t)=c\cdot\prod_{i=1}^{r}(t-\beta_i)$ for some $c\in\mathbb{Q}^*$. Hence $\{z\in\mathbb{C}|f(z)=0\}=\{\sigma(\mathfrak{a}^m)|\sigma\in G\}=\{\big(\sigma(\mathfrak{a})\big)^m|\sigma\in G\}=\{z_i^m\;|\;i\in[d]\}$. The last equality holds since the Galois group of irreducible polynomial $p(t)$ operates transitively on its roots. Set $ R_m=\{z_i^m\;|\;i\in[d]\}$ to be complex roots of the minimal polynomial of $\mathfrak{a}^m$, then $|R_m|=\textbf{deg}(\mathfrak{a}^m)$. By Definition \[reddegree\], $\textbf{rdeg}(\mathfrak{a})$\ $=\min_{m\in\mathbb{Z},m\geq1}|R_m|$. If $z_i/z_j$ is a root of unity for some $1\leq i\neq j\leq d$, then $\mathfrak{a}$ is degree reducible. Set $m=\mathbf{Order}(z_i/z_j)$, then $\mathbf{deg}(\alpha^m)=|R_m|<|R_1|=d=\mathbf{deg}(\alpha)$. Define an equivalent relation $\sim$ on $\mathbb{C}$ by: $z\sim z'$ iff $z/z'$ is a root of unity. Because $z_i^m= z_j^m$ implies $z_i\sim z_j$, $|R_1/$$\sim$$|\leq|R_m|$ for any integer $m\geq1$. Denote by $M$ the unitary order of $p(t)$ ([@yokoyama1995finding] Definition 2.3), then $\forall i,j\in[d], (z_i\sim z_j \Leftrightarrow z_i^M=z_j^M)$. Thus $|R_M|=|R_1/$$\sim$$|$. Combine this with (i) and (ii), it follows that $|R_1/$$\sim$$|=\min_{m\in\mathbb{Z},m\geq1}|R_m|=\textbf{rdeg}(\mathfrak{a}).$ Since $z_i\sim z_j$ iff $z_i^m\sim z_j^m$, we conclude that $|R_1/$$\sim$$|=|R_m/$$\sim$$|$. \[correct\] $|R_m/$$\sim$$|=|R_m|\Rightarrow \textbf{\emph{rdeg}}(\mathfrak{a})=|R_m|=\textbf{\emph{deg}}(\mathfrak{a}^m)$. This is obvious since $|R_m|=|R_m/$$\sim$$|=|R_1/$$\sim$$|$ and (iii) holds. [@yokoyama1995finding] provides an efficient algorithm **Unitary-Test** so that for $p(x)\in \mathbb{Q}[x]$, if the quotient of a pair of roots of $p$ is a root of unity, it returns $\{\textbf{True},k\}$, where $k$ is the order of the quotient. Otherwise it returns $\{\textbf{False}, 0\}$. Based on this, we develop Algorithm \[algdegred\] to compute a reducing exponent and the reduced degree of a given algebraic number. The algorithm terminates since the number of roots of $f(t)$ decreases in Step 5. Its correctness is justified by Proposition \[correct\]. $prod=1;$ $f(t)=p(t)$;\ Suppose $f(t)$ has a root $r$;\ $\{Reducible,k\}=\textbf{Unitary-Test}(f(t))$;\ **if** ($Reducible==\textbf{False}$) **{Return** $\{prod,f(t)\}$;**} endif**\ $prod=prod\cdot k;$ $f(t)=\textbf{MinimalPolynomial }(r^k)$;\ **Goto** Step 2;\[algdegred\] Experiments and Application {#sec:example} =========================== We show experimental results verifying the effectiveness of the framework to tackle problems larger than those that **FindRelations** can handle. The Mathematica package **FindRelations** is available at$$\text{\textcolor{blue}{https://www3.risc.jku.at/research/combinat/software/}.}$$We implemented our algorithms with Mathematica. All results are obtained on a laptop of WINDOWS 7 SYSTEM with 4GB RAM and a 2.53GHz Intel Core i3 processor with 4 cores. [c|c|c|c|c]{} NO.&TD&RTD&R/B&Time (s) FD/FF/GE\ 1&9.3$\times10^6$&1.5$\times10^6$&4/3&26/2196/&gt;3600\ 2&1.2$\times10^5$&336&5/2&114.8/114.3/&gt;3600\ 3a&30&30&3/0&0.035/0.032/8.9\ 3b&120&120&4/0&2.69/2.68/&gt;3600\ 3c&600&600&5/0&360/361/&gt;3600\ 4a&108&108&4/1&11.18/0.79/0.11\ 4b&256&256&4/0&&gt;3600/188.4/78.6\ 4c&729&729&4/2&&gt;3600/2049/&gt;3600\ 5a&$4.7\times10^4$&1&1/5&0.433/0.433/2.227\ 5b&$6.6\times10^7$&1&4/2&0.712/0.720/&gt;3600\ 5c&$3.8\times10^{27}$&1&11/9&203.98/203.77/&gt;3600\ In the table TD stands for the product of degrees of given algebraic numbers, RTD the product of reduced degrees, R the rank (Definition \[algerank\]) and B cardinality of the basis. FD and FF are respectively the main algorithm and the main algorithm with **DecideDependence** replaced by **FindRelations**, while GE means applying **FindRelations** directly to the problem. TD (RTD) is the upper bound of the degree of the extended field generated by (reduced) given algebraic numbers. It indicates the difficulty for algorithm GE (FD, FF) to compute in the extended field. For algorithm FD (FF), the number of input algebraic numbers for function **DecideDependence** (**FindRelations**) is at most R$+1$ throughout the computation, thus R also characterizes the computational difficulty. These three algorithms (FD, FF, GE) are all very efficient for small inputs (*e.g.*, roots of an irreducible polynomial of degree 3) which we leave out in the table. The runtimes of these algorithms depend intensively on the input. We see that these algorithms perform differently from each other in the examples. In Example 1 both TD and RTD are large. GE fails to give an answer within an hour, while FD (FF) deals with partial input by **DecideDependence** (**FindRelations**) and is faster. In Example 2, RTD is much smaller than TD, FD (FF) is more efficient than GE which does not contain degree reduction. The non-degenerate condition holds for Examples 3a-c. The certificate for multiplicative independence is effective in this case. In Examples 4a-b GE does better than FD (FF) since the degree reduction does not help, R$/$(R$+$B) is close to $1$ and the non-degenerate condition fails to hold. However FF still gives correct answer. FF outperforms GE in Example 4c since R/(R$+$B) is relatively small. Examples 5a-c deal with the case where all input numbers are roots of rationals. FD (FF) can handle very large problem of this type. To conclude, (i) RTD $<<$ TD, (ii) R/(R$+$B) $<<$ 1 and (iii) the non-degenerate condition holds for more input numbers are good for FD (FF) to tackle problems larger than those that GE can handle. An interesting application is to compute the invariant polynomial ideal of linear loops considered in [@lvov2010polynomial; @structure] of the form: $$X= b; {\rm While} \textbf{ True }{\rm do} \;X= AX;$$ Here $b\in\mathbb{Q}^m$, $A\in\mathbb{Q}^{m\times m}$ is diagonalizable with nonzero eigenvalues. Let $X= (X_1,\ldots,X_m)^T$ be the vector of indeterminates and $I(A,b)= \{f(X)\in\mathbb{C}[X]\;|\;f(A^kb)=0,\forall k\geq0\}$ the *invariant polynomial ideal*. Suppose $A^T$$=$$ PDP^{-1}$, $D=\textbf{diag}\{x_1,x_2,\ldots,x_m\}$, $x= (x_1,x_2,\ldots,x_m)^T$ and $\tilde{b}= P^Tb\in(\mathbb{C}^*)^m$. For $v\in\mathcal{R}_x$, define $v_+=(\max\{v(1),0\},\ldots,\max\{v(m),0\})^T$ and $v_-= (-v)_+$. Then by Theorems 3 and 2 in [@structure] one concludes $$I(A,b)=\langle\{({\tilde{b}}^{v_-})(P^TX)^{v_+}-({\tilde{b}}^{v_+})(P^TX)^{v_-}|v\in\mathcal{R}_x\}\rangle. \label{thm3}$$ Define $\mathfrak{g}_v(X)=({\tilde{b}}^{v_-})(P^TX)^{v_+}-({\tilde{b}}^{v_+})(P^TX)^{v_-}$, we observe that $\forall k\in\mathbb{Z}$, $\mathfrak{g}_v(X)|\mathfrak{g}_{kv}(X)$. Hence if $\textbf{rank}(\mathcal{R}_x)=1$ with only one basis vector $u$, (\[thm3\]) becomes $ I(A,b)=\langle\mathfrak{g}_u(X)\rangle. $ Set $p_i(X) \in \mathbb{C}[X],i\in[s]$, consider the loop: $$\begin{array}{l} X= b;\\ While\;(p_1(X)==0\land\cdots\land p_s(X)==0)\\ do\;(X= AX;) \end{array}\label{loop}$$ which does not terminate iff $\mathfrak{g}_u(X)|p_i(X),i\in[s]$. Set $$A = \left( {\begin{array}{*{20}{c}} 4&{226}&2&1&{ - 117}\\ 1&{126}&1&0&{ - 64}\\ 0&{ - 91}&0&{ - 1}&{46}\\ 0&{80}&1&0&{ - 40}\\ 4&{232}&2&1&{ - 120} \end{array}} \right),b = \left( {\begin{array}{*{20}{c}} 2\\ { - 1}\\ 0\\ 3\\ { - 4} \end{array}} \right).$$ The only basis vector of the exponent lattice defined by eigenvalues of matrix $A$ is $u=(0, 1, 1, 0, 1)^T$. Computing by definition shows\ $ \\\mathfrak{g}_u(X)= 7674169 X_1^3 - 31858655 X_1^2 X_2 + 22396826 X_1 X_2^2 - 165997 X_4^3- $\ $ 255380 X_2 X_3^2 + 12769 X_3^3 +153228 X_4 X_5^2 + 1442897 X_5^3+459684 X_1 X_3^2 - $\ $ 3051791 X_1 X_2 X_4 + 3639165 X_1^2 X_3 +8504154 X_2 X_3 X_5- 472453 X_2^2 X_3 -$\ $ 5694974 X_1 X_4 X_5 +1391821 X_2 X_3 X_4- 127690 X_1 X_4^2 + 2106885 X_2 X_4^2-$\ $ 2311189 X_1 X_3 X_5-1442897 X_3 X_4 X_5 - 906599 X_4^2 X_5 -3639165 X_1 X_5^2 - $\ $ 204304 X_3 X_4^2 - 5465132 X_1^2 X_5 + 35817045 X_1 X_2 X_5 + 3971159 X_1^2 X_4 - $\ $ 2387803 X_2 X_5^2- 995982 X_3 X_5^2 -28347180 X_2^2 X_5 +9947051 X_2 X_4 X_5 -$\ $ 12769 X_3^2 X_4 + 5528977 X_2^3 - 7380482 X_2^2 X_4 - 293687 X_3^2 X_5 -$\ $ 8899993 X_1 X_2 X_3 + 855523 X_1 X_3 X_4- 858983399. $ \ The ideal in (\[thm3\]) is a lattice ideal up to an invertible linear transformation in the coordinates. When there are at least two basis vectors, we may compute according to [@markov], from the basis vectors, a Markov basis $\{\eta_j\}_{j=1}^r$, *s.t.* $ I(A,b)=\langle\{\mathfrak{g}_{\eta_j}(X)\}_{j=1}^r\rangle $ by Lemma A.1 of [@markov]. Then (\[loop\]) does not terminate iff $p_i(X)\in\langle\{\mathfrak{g}_{\eta_j}(X)\}_{j=1}^r\rangle,i\in[s]$. Conclusions {#sec:conclusions} =========== This paper provided an effective framework to construct exponent lattice basis in an inductive way. In many situations the framework can handle problems larger than those that **FindRelations** can do. It is efficient especially when degrees of algebraic numbers to deal with can be intensively reduced, when the rank of algebraic numbers is small and when non-degenerate condition holds for many of the input algebraic numbers. The non-degenerate condition provides a relatively cheaper certificate for multiplicative independence. It usually holds for randomly picked algebraic numbers, thus it can be useful for proving their multiplicative independence. However, it often fails when numbers are algebraically dependent. Though the problem of computing exponent lattice basis for nonzero algebraic numbers is still difficult (the exponent lattice basis for roots of a general rational polynomial of degree $5$ cannot be computed fast), our framework still casts light on how we may handle larger problems. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== This work was supported partly by NSFC under grants 61732001 and 61532019. The authors thank Professor Shaoshi Chen for providing helpful references and Haokun Li for useful advice on programming. [99]{} S. Almagor, B. Chapman, M. Hosseini, et al. *Effective Divergence Analysis for Linear Recurrence Sequences.* arXiv preprint arXiv:1806.07740, 2018. F. Barrera Mora. *On subf$\grave{\text{I}}$ields of radical extensions*. Communications in Algebra, 1999, 27(10): 4641-4649. S. Chen, R. Feng, G. Fu, et al. *On the structure of compatible rational functions*. Proceedings of the 36th international symposium on Symbolic and algebraic computation. ACM, 2011: 91-98. H. Derksen, E. Jeandel, P. Koiran. *Quantum automata and algebraic groups*. Journal of Symbolic Computation, 2005, 39(3-4): 357-371. G. Ge. *Algorithms related to multiplicative representations*. PhD thesis, University of California, Berkeley, 1993. R. Hemmecke, P. N. Malkin. *Computing generating sets of lattice ideals and Markov bases of lattices*. Journal of Symbolic Computation, 2009, 44(10): 1463-1476. J. L. G. Hewage, J. Lanel. *Complex root isolation*. PhD thesis, Oakland University, 2009. M. Kauers. *Algorithms for nonlinear higher order difference equations*. PhD thesis, RISC-Linz, Linz, Austria, 2005. M. Kauers, B. Zimmermann. *Computing the algebraic relations of C-finite sequences and multisequences*. Journal of Symbolic Computation, 2008, 43(11): 787-803. T. Loher, D. Masser. *Uniformly counting points of bounded height*. Acta Arithmetica, 2004, 3(111): 277-297. J. Loxton, A. van der Poorten. *Multiplicative dependence in number fields*. Acta Arithmetica, 1983, 42(3): 291-302. M. S. Lvov. *Polynomial invariants for linear loops*. Cybernetics and Systems Analysis, 2010, 46(4): 660-668. M. S. Lvov. *The Structure of Polynomial Invariants of Linear Loops*. Cybernetics and Systems Analysis, 2015, 51(3): 448-460. D. W. Masser. *Linear relations on algebraic groups*. New Advances in Transcendence Theory, 1988: 248-262. A. van der Poorten , J. Loxton. *Multiplicative relations in number fields*. Bulletin of the Australian Mathematical Society, 1977, 16(1): 83-98. K. Yokoyama, Z. Li, I. Nemes. *Finding roots of unity among quotients of the roots of an integral polynomial*. Proceedings of the 1995 international symposium on Symbolic and algebraic computation. ACM, 1995: 85-89. D. Bertrand. *Duality on tori and multiplicative dependence relations*. Journal of the Australian Mathematical Society, 1997, 62(2): 198-216. E. M. Matveev. *On linear and multiplicative relations*. Sbornik: Mathematics, 1994, 78(2): 411-425.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Luo and Tan gave a new identity for hyperbolic surfaces with/without geodesic boundary in terms of dilogarithms of the lengths of simple closed geodesics on embedded three-holed spheres or one-holed tori in [@LuoTan11]. However, the identity was trivial for a hyperbolic one-holed torus with geodesic boundary. In this paper we adapt the argument from Luo and Tan to give an identity for hyperbolic tori with one geodesic boundary or cusp in terms of dilogarithm functions on the set of lengths of simple closed geodesics on the torus. As a corollary, we are also able to express the Luo-Tan identity as a sum over all immersed three-holed spheres $P$ which are embeddings when restricted to the interior of $P$.' address: | Department of Mathematics\ National University of Singapore\ Singapore 119076 author: - Hengnan Hu and Ser Peow Tan title: New identities for small hyperbolic surfaces --- [^1] Introduction ============ In the last couple of decades, several interesting identities had been derived by various authors for hyperbolic manifolds and surfaces, notably, those of Basmajian [@Basmajian93], McShane [@McShane98; @McShane04; @McShane06], Bridgeman [@Bridgeman11], Bridgeman and Kahn [@BridgemanKahn10] and Luo and Tan [@LuoTan11]. For a survey of the identities, generalizations and their connections, see Bridgeman and Tan [@BridgemanTan13b] and also [@BridgemanTan13a]. For hyperbolic surfaces, the identities derived were for surfaces with geodesic boundary and/or cusps, except for the Luo-Tan identity which works for surfaces with or without boundary. However, the identity derived by Luo-Tan was trivial in the case of hyperbolic one-holed tori as in this case, there was only one embedded simple surface, namely the surface itself. This contrasts for example with McShane’s identity which was first proved for hyperbolic tori with one cusp by McShane in his thesis [@McShane91] and involves a non-trivial sum over the lengths of all simple closed geodesics, and subsequently generalized by Mirzakhani for hyperbolic tori $T$ with one geodesic boundary (more generally hyperbolic surfaces with geodesic boundary) in [@Mirzakhani07], and Tan, Wong and Zhang for hyperbolic tori with one cone singularity (more generally hyperbolic surfaces with cone singularities of cone angle $\le \pi$) in [@TanWongZhang06], and more generally for ${\rm{SL}}(2, {\mathbb{C}})$ characters of $\pi_1(T)$ satisfying Bowditch’s Q-conditions in [@TanWongZhang08; @TanWongZhang06b]. The main aim of this paper is to derive a non-trivial version of the Luo-Tan identity for hyperbolic tori with one geodesic boundary or cusp, and as a corollary, to express the Luo-Tan identity for closed hyperbolic surfaces as a sum over all immersed geometric three-holed spheres $P$ in the surface which are embeddings when restricted to the interior of $P$. The basic idea is to exploit the close connection between four-holed spheres and one-holed tori, and to use the non-trivial Luo-Tan identity for hyperbolic four-holed spheres with four geodesic boundaries of equal lengths to obtain a corresponding identity for the one-holed tori. An earlier version of the results here form part of the thesis of the first author [@Hu13]. Let $T$ be a hyperbolic torus with one geodesic boundary or cusp. Let $$SG_T=\{\gamma ~:~ \gamma ~~\hbox{is a non-peripheral simple closed geodesic on $T$}\}$$ and let $SS_T$ be the set of lengths (counted with multiplicity) of the elements of $SG_T$, we call $SS_T$ the [*simple length spectrum*]{} of $T$. The main result of this paper is the following identity for $T$: \[thm:oneholed\] For a hyperbolic one-holed torus $T$ with boundary geodesic $K$ of length $k>0$, we have $$\label{eq:main} \sum_{b}\Bigg\{\mathcal{L}\left(\frac{\cosh(\frac{k}{2})+1}{\cosh(\frac{k}{2})+\cosh(b)}\right)+ 2 \mathcal{L}\left(\frac{\cosh(\frac{k}{4}+\frac{b}{2})}{\cosh(\frac{k}{4}) e^{\frac{b}{2}}}\right) -2 \mathcal{L}\left(\frac{2\sinh(\frac{b}{2})}{(1+e^{-\frac{k}{2}}) e^{\frac{b}{2}}}\right)\Bigg\}=\frac{\pi^2}{2}$$ where the sum is over the simple length spectrum $SS_T$ and $\mathcal{L}(z)$ is the Roger’s dilogarithm function. By taking the limit as $k \rightarrow 0$ in the proof of Theorem \[thm:oneholed\], we obtain the following identity for hyperbolic tori with one cusp. \[thm:oncepuncturetorus\] For a once-punctured torus $T$ equipped with a complete hyperbolic structure of finite area, $$\label{newpunctureid} \sum_{b}\left\{\mathcal{L}\left(\operatorname{sech}^2(\frac{b}{2})\right) + 2\mathcal{L}\left(\frac{1 + e^{-b}}{2}\right)-2 \mathcal{L}\left(\frac{1 - e^{-b}}{2}\right)\right\}=\frac{\pi^2}{2}$$ where the sum is over the simple length spectrum $SS_T$. Compare the above result with McShane’s identity, which states that, with the same conditions, $$\label{eq:McShane} \sum_{b}\frac{1}{1+e^{b}}=\frac{1}{2}.$$ Theorem \[thm:oncepuncturetorus\] has an equivalent form in terms of traces and representations into ${\rm{PSL}}(2,\mathbb{R})$ (see Theorem \[thm:puncturenew\]). Recall that an embedded, [*geometric*]{} three-holed sphere/one-holed torus (collectively, simple surface) in a hyperbolic surface $S$ is an embedding from a three-holed sphere $P$/one-holed torus $T$ into $S$ such that the boundaries are mapped to geodesics. The Luo-Tan identity for closed hyperbolic surfaces in [@LuoTan11] is given by: \[thm:LuoTan\] Let $S$ be a closed, orientable hyperbolic surface of genus $\textsf{g} \ge 2$. There exist functions $f$ and $g$ involving the Roger’s dilogarithm function of the lengths of the simple closed geodesics on a three-holed sphere or a one-holed torus, such that $$\label{eq:originalLuoTan} \sum_{P} f(P) + \sum_{T} g(T) = 8 \pi^2 (\textsf{g}-1)$$ where the first sum is over all embedded geometric three-holed spheres $P \subset S$ and the second sum extends over all embedded geometric one-holed tori $T \subset S$. As a corollary of Theorem \[thm:oneholed\] we can express the Luo-Tan identity as a sum over all immersed geometric three-holed spheres $P$ of $S$ such that the restriction to $int(P)$, the interior of $P$ is an embedding into $S$. We define a [*quasi-embedded geometric*]{} three-holed sphere in $S$ (orientable) to be an immersion of a three-holed sphere $P$ into $S$ which is injective on the interior $int(P)$ of $P$ such that the boundaries are mapped to geodesics, but two of the boundaries are mapped to the same geodesic. Thus, a quasi-embedded geometric three-holed sphere is contained in a unique embedded geometric one-holed torus, and conversely, every embedded geometric one-holed torus together with a non-peripheral simple closed geodesic on the torus determines a quasi-embedded geometric three-holed sphere. We have: \[cor:newform\] Let $S$ be a closed, orientable hyperbolic surface of genus $\textsf{g} \ge 2$. There exist functions $f_1$ and $f_2$ involving the dilogarithms of the lengths of the simple closed geodesics in a hyperbolic three-holed sphere such that $$\label{eq:newformLuoTan} \sum_{P_1} f_1(P_1) + \sum_{P_2} f_2(P_2) = 8 \pi^2 (\textsf{g}-1)$$ where the first sum is over all embedded geometric three-holed spheres $P_1 \subset S$ and the second sum is over all quasi-embedded geometric three-holed spheres $P_2$ of $S$. The explicit formulae for $f_1$ and $f_2$ are given in §\[ss:pseudo\], $f_1$ is in fact just the function $f$ in Theorem \[thm:LuoTan\]. Note that the main advantage of this reformulation is that in Theorem \[thm:LuoTan\], the function $g$ in was expressed in terms of an infinite sum and so the identity really involves double sums, in this formulation, the identity is expressed as a simple sum over all geometric three-holed spheres which are embeddings when restricted to the interior. [**Remark.**]{}We can also easily get the reformulation of the Luo-Tan identity for non-closed hyperbolic surfaces whose Euler characteristics are strictly less than $-1$ as in Theorem 1.2 of [@LuoTan11] (Arxiv version). The rest of this paper is organized as follows. In §\[ss:notion\], we set the notation and apply the main theorem in [@LuoTan11] to a hyperbolic four-holed sphere $Q$ with four geodesic boundaries all of length $c$, to obtain an identity for $Q$. In §\[ss:proofs\], we describe the relation between hyperbolic four-holed spheres $Q$ with geodesic boundaries all of length $c$, and hyperbolic one-holed tori $T$ with one geodesic boundary of length $k=2c$. We then prove Theorems \[thm:oneholed\], and \[thm:oncepuncturetorus\] by transcribing the identity for $Q$ to the one for $T$. Finally, in §\[ss:pseudo\], we prove Corollary \[cor:newform\] giving the explicit formulas for $f_1$ and $f_2$. [*Acknowledgements.*]{} We would like to thank Martin Bridgeman, Feng Luo and Hugo Parlier for helpful discussions and comments. Notation and results for four-holed spheres {#ss:notion} =========================================== We establish the Luo-Tan identity for hyperbolic four-holed spheres with boundaries of equal lengths in this section. [**The Roger’s Dilogarithm.** ]{} Recall that the Roger’s Dilogarithm function $\mathcal{L}(z)$ for $z\le 1$ is defined by $$\mathcal{L}(z):=-\frac{1}{2} \int_0^z(\frac{\log u }{1-u}+\frac{\log (1-u)}{u}) \, du$$ see [@Roger1907]. $\mathcal{L}(z)$ can also be thought of as a normalization of the regular dilogarithm function $$Li_2(z) = \sum_{n=1}^\infty \frac{z^n}{n^2}\qquad \mbox{ for } |z| < 1,$$ where $${\mathcal L}(z) = Li_2(z) + \frac{1}{2}\log|z|\log(1-z).$$ ${\mathcal L}(z)$ is an increasing function on $z \le 1$ and it enjoys many remarkable properties, see for example [@Kirillov89; @Leonard91; @GeRiRa99] for a detailed discussion. Here we mention some of the more useful properties. We have $\mathcal{L}(0)=0$, $\mathcal{L}(\frac{1}{2})=\frac{\pi^2}{12}$, $ \mathcal{L}(1) =\frac{\pi^2}{6}$ and $\mathcal{L}(-1) =-\frac{\pi^2}{12}$. The Euler relations are given by: $$\mathcal{L}(x)+\mathcal{L}(1-x) = \mathcal{L}(1)=\frac{\pi^2}{6}$$ for $0 \le x \le 1$ and $$\mathcal{L}(-x)+\mathcal{L}(-x^{-1}) = 2\mathcal{L}(-1)=-\frac{\pi^2}{6}$$ for $x>0$. The Landen’s identity is $$\mathcal{L}(\frac{-x}{1-x})=-\mathcal{L}(x)$$ for $0\le x \le 1$. Note $\displaystyle \lim_{x \to -\infty}\mathcal{L}(x) = -\mathcal{L}(1) = -\frac{\pi^2}{6}$. The pentagon relation is as follows: for $x, y \in [0,1]$ and $xy \neq 1$, $$\mathcal{L}(x)+\mathcal{L}(y)+\mathcal{L}(\frac{1-x}{1-xy}) + \mathcal{L}(\frac{1-y}{1-xy}) =\mathcal{L}(xy)+\frac{\pi^2}{3}.$$ [**The Lasso function**]{}. The Lasso function $La(x,y)$ in [@LuoTan11], which corresponds to the measure of a certain subset of the unit tangent bundle of a three-holed sphere is defined in terms of the Roger’s Dilogarithm function as follows: $$\label{eq:lassofunction} La(x,y):=\mathcal{L}(y)+\mathcal{L}(\frac{1-y}{1-xy})-\mathcal{L}(\frac{1-x}{1-xy})$$ for $0 < x < y < 1$. [**Remark.**]{} In fact, the function $La(x,y)$ is well-defined on the square domain $0 \le x,y \le 1$. Here it is enough to focus on the upper triangle region. [**Orthogeodesics**]{}. Given a (cone)-hyperbolic surface $S$, an [*orthogeodesic*]{} from $A$ to $B$ is a closed geodesic arc $\alpha$ from the set $A$ to the set $B$ which is orthogonal to $A$ and $B$. Here we usually consider $A$ and $B$ to be geodesics, however, they may be points on $S$ or cone singularities on $S$ (where the orthogonality condition there is vacuous), or they may be horocycles around cusps. The orthogeodesic is [*simple*]{} if the restriction to the interior is injective. [**The Luo-Tan identity for four-holed spheres**]{}. Let $Q$ be a hyperbolic four-holed sphere with boundary geodesics $C_1,C_2,C_3,C_4$ all of length $c > 0$ and let $A$ be a non-peripheral simple closed geodesic on $Q$, see Figure \[fig:4sphere\]. We obtain two isometric three-holed spheres $P_1$, $P_2$ by cutting the four-holed sphere $Q$ along $A$. Suppose that $P_1$ has boundaries $C_i$, $C_j$ and $A$, where $\{i,j\} \subset \{1,2,3,4\}$. Let $m_A$ be the length of the simple orthogeodesic from $C_i$ to $A$ in $P_1$, $p_A$ be the length of the simple orthogeodesic from $A$ to itself in $P_1$ and $q_A$ the length of the simple orthogeodesic from $C_i$ to $C_j$ in $P_1$, see Figure \[fig:4sphere\]. Note that by symmetry, $m_A$ is also the length of the simple orthogeodesic from $C_j$ to $A$ and that the lengths of the corresponding simple orthogeodesics on $P_2$ are given by the same quantities. Let $U(Q)$ be the set of unit tangent vectors over $Q$ and $\mu$ be the measure on $U(Q)$ which is invariant under the geodesic flow. We have $\mu(U(Q))=-4 \pi^2 \chi(Q) =8 \pi^2$ and applying the Luo-Tan identity to $Q$ gives: \[le:4boundary\] For a hyperbolic four-holed sphere $Q$ with boundary geodesics all of length $c >0$, $$\label{eq:fournew} \sum_{A} \Bigg\{\mathcal{L}\left(\tanh^2 (\frac{p_A}{2})\right) + 2 \mathcal{L}\left(\tanh^2 (\frac{m_A}{2})\right) - 2 La\left(e^{-c}, \tanh^2(\frac{m_A}{2})\right) \Bigg\} =\frac{\pi^2}{2}$$ where the sum is over all non-peripheral simple closed geodesics $A$ on $Q$ and for every non-peripheral simple closed geodesic $A$, $m_A$ and $p_A$ are the lengths of the simple orthogeodesics as described above (see Figure \[fig:4sphere\]). Note that $e^{-c} < \tanh^2(\frac{m_A}{2})$ in the above since for a fixed $c>0$, there is a unique real $m'$ such that $ \sinh(m')\sinh(\frac{c}{2}) = 1$ and for any hyperbolic structure on $Q$ and every non-peripheral simple closed geodesic $A$ on $Q$, $m' < m_A $, from which the inequality follows. Furthermore, we can use hyperbolic identities and the properties of the Roger’s dilogarithm function $\mathcal{L}(z)$ given at the beginning of the section to obtain an equivalent form of the above identity in terms of just the quantities $\ell(A)$ and $c$. We omit the elementary but somewhat tedious details. We have the following reformulation of identity (\[eq:fournew\]) for $Q$: $$\label{eq:foursimpleid} \sum_{A} \Bigg\{ \mathcal{L}\left(\frac{\cosh(c)+1}{\cosh(c)+ \cosh(\frac{\ell(A)}{2})}\right)+ 2\mathcal{L}\left(\frac{\cosh(\frac{c}{2}+\frac{\ell(A)}{4})}{\cosh(\frac{c}{2}) e^{\frac{\ell(A)}{4}} }\right) -2 \mathcal{L}\left(\frac{2\sinh(\frac{\ell(A)}{4})}{(1+e^{-c})e^{\frac{\ell(A)}{4}}}\right)\Bigg\} =\frac{\pi^2}{2}$$ where the sum extends over all non-peripheral simple closed geodesics $A$ on $Q$ and $\ell(A)$ denotes the length of $A$. Taking the limit as $c$ tend to $0$, we obtain: \[le:4cusped\] For a hyperbolic quadruply-punctured sphere $Q$, $$\label{eq:ideal4sphere} \sum_{A} \left\{\mathcal{L}\left(\operatorname{sech}^2(\frac{\ell(A)}{4})\right)+2\mathcal{L}\left(\frac{1+e^{-\frac{\ell(A)}{2}}}{2}\right) -2 \mathcal{L}\left(\frac{1-e^{-\frac{\ell(A)}{2}}}{2}\right) \right\} = \frac{\pi^2}{2}$$ where the sum is over all simple closed geodesics $A$ on $Q$ and $\ell(A)$ denotes the length of $A$. Proofs of Theorems \[thm:oneholed\] and \[thm:oncepuncturetorus\] {#ss:proofs} ================================================================= In the following we use the widely known relationship between hyperbolic structures on one-holed tori and four-holed spheres. We claim that a hyperbolic one-holed torus with geodesic boundary of length $2c$ is paired with a hyperbolic four-holed sphere with geodesic boundaries all of length $c$, via a (two-to-one and four-to-one respectively) branched covering over a hyperbolic orbifold of genus $0$ with three cone points of angle $\pi$ and one geodesic boundary of length $c$. We provide some details below for the convenience of the reader. [**One-holed torus and the elliptic involution**]{}. Let $T$ be a hyperbolic one-holed torus with boundary geodesic $K$ of length $k$ and let $B$ be a non-peripheral simple closed geodesic on $T$ with length $b$. Let $P_B$ be the simple orthogeodesic from $K$ to itself which is disjoint from $B$ and let $p_B$ be its length. Let $Q_B$ be the simple orthogeodesic from $B$ to itself which intersects $B$ at only its end points, and let $q_B$ be its length. Finally, let $M_B$ and $M_B'$ be the simple orthogeodesics from $K$ to $B$ which intersect $B$ only at the endpoint, they both have the same length which we denote by $m_B$. These geodesics and orthogeodesics on $T$, as well as on the three-holed sphere obtained by cutting $T$ along $B$ are shown in Figure \[fig:1holedtorus\]. Let $p_1$ and $p_2$ be the (not necessarily distinct) points on $B$ which are the endpoints of $Q_B$. Note that the two points coincide if and only if the twist parameter of the Fenchel-Nielsen coordinates with respect to the curve $B$ is an integer multiple of $b$. Generically, $p_1$ and $p_2$ separate $B$ into two arcs, we denote the midpoints of these two arcs by $\omega_1$ and $\omega_2$. Note that $\omega_1$ and $\omega_2$ are diametrically opposite on $B$. Let $C$ be any other simple closed geodesic on $T$ which intersects $B$ exactly once. Then $C$ intersects $P_B$ exactly once at the midpoint of $P_B$ which we denote by $\omega_3$ and it intersects $B$ at either $\omega_1$ or $\omega_2$. Furthermore, these points of intersection of $C$ with $P_B$ and $B$ are diametrically opposite on $C$, and the point missed by $C$ is the midpoint of $P_C$ which is the simple orthogeodesic from $K$ to itself which is disjoint from $C$. As every simple closed geodesic on $T$ can be obtained from a basic triple of simple closed geodesics which pair-wise intersect once, by moving along a binary tree, we see that every simple closed geodesic passes through exactly two of the three points $\omega_1, \omega_2, \omega_3$, which are diametrically opposite on the geodesic. The three points $\omega_1, \omega_2, \omega_3$ are the Weierstrass points of $T$ and rotation by $\pi$ at any one of $\omega_i$ is an isometry $\iota$ of $T$ which fixes all three points. $\iota$ is the elliptic involution of $T$, it fixes every non-peripheral simple closed geodesic, but reverses the direction. We have $T/\langle \iota\rangle $ is a hyperbolic orbifold $\mathcal{O}$ of genus $0$ with three cone points $c_1, c_2, c_3$ of cone angle $\pi$ and one boundary $\overline K$ with length $\frac{k}{2}$. $T$ is a double branched cover of $\mathcal{O}$ and the Weierstrass points $\omega_1, \omega_2, \omega_3$ cover the cone points $c_1, c_2, c_3$ respectively, see Figure \[fig:torusorbifold\]. The boundary $K$ double covers the boundary $\overline K$ of ${\mathcal O}$, the geodesic $B$ double covers a simple geodesic arc $\overline B$ on $\mathcal{O}$ joining $c_1$ and $c_2$, so $b=2\ell(\overline{B})$. $P_B$ double covers a simple orthogeodesic $\overline{P_B}$ from $c_3$ to the boundary disjoint from $\overline B$, so $p_B=2\ell(\overline{P_B})$. $M_B$ and $M_B'$ (as in Figure \[fig:1holedtorus\]) cover the simple orthogeodesic $\overline{M_B}$ on ${\mathcal O}$ from $\overline K$ to $\overline B$, whose interior is disjoint from $\overline B$, so $m_B=\ell(\overline{M_B})$. $Q_B$ double covers the simple orthogeodesic $\overline{Q_B}$ from $c_3$ to $\overline B$, whose interior is disjoint from $\overline B$, so $q_B=2\ell(\overline{Q_B})$. Reversing the construction, and keeping track of the covering data, we see every such hyperbolic orbifold ${\mathcal O}$ with boundary geodesic of length $k/2$ is double branch covered by a unique hyperbolic torus $T$ with geodesic boundary of length $k$. [**Four-holed spheres and the ${\mathbb{Z}}_2 \times {\mathbb{Z}}_2$ action**]{}. The situation for a hyperbolic four-holed sphere $Q$ with four boundary geodesics $C_1,C_2,C_3,C_4$ of equal length $c$ is similar but slightly more complicated. Let $A$ be a non-peripheral simple closed geodesic on $Q$ with length $a$. Cutting along $A$ gives two isometric three-holed spheres $P$ and $P'$, without loss of generality, assume that $C_1, C_2 \subset \partial P$ and $C_3, C_4 \subset \partial P'$. Let $P_A$ (resp. $ P_A'$) be the simple orthogeodesic on $P$ (resp. $P'$) from $A$ to itself with length $p_A$ and let $x,y$ (resp. $x',y'$) be the endpoints of $P_A$ (resp. $P_A'$) on $A$, so that $x,x',y,y'$ are in cyclic order on $A$. The midpoints of the four intervals obtained divide $A$ into four segments of equal length, denote these points by $\omega_1, \omega_2, \omega_1', \omega_2'$ respectively, where $\omega_i$ and $\omega_i'$ are diametrically opposite on $A$.. Let $M_A^i$, $i=1,2,3,4$ be the simple orthogeodesic from $C_i$ to $A$ in $P$ and $P'$, they have equal length, denoted by $m_A$ and let $Q_A$ (resp. $Q_A'$) be the simple orthogeodesic from $C_1$ to $C_2$ (resp. $C_3$ to $C_4$) in $P$ (resp. $P'$), with length $q_A$. See Figure \[fig:4holedorbifold\]. Now let $C$ be any simple closed geodesic which intersects $A$ exactly twice. Then, $C$ intersects $Q_A$ and $Q_A'$ exactly once at their midpoints, which we denote by $\omega_3$, $\omega_3'$ respectively. Furthermore, $C$ intersects $A$ in either the pair of points $\omega_1, \omega_1'$ or the pair $\omega_2,\omega_2'$, and these together with $\omega_3, \omega_3'$ are equidistributed on $C$. As in the one-holed torus case, we can in fact show that every non-peripheral simple closed geodesic on $Q$ passes through exactly two of the pairs $\{\omega_i, \omega_i\}$, which are equally distributed on the geodesic. These six points are the Weierstrass points of $Q$. A $\pi$ rotation about say $\omega_1$ is an isometry of $Q$ taking $P$ to $P'$, rotation about $\omega_1'$ induces the same involution on $Q$, which we denote by $\iota_1$. We define $\iota_2, \iota_3$ similarly, together, they generate a ${\mathbb{Z}}_2 \times {\mathbb{Z}}_2$ action on $Q$ since $\iota_1\iota_2=\iota_3$ etc., so $\langle \iota_1, \iota_2\rangle \cong Z_2 \times {\mathbb{Z}}_2$. $Q /\langle \iota_1, \iota_2 \rangle$ is a genus $0$ hyperbolic orbifold ${\mathcal O}$ with three cone points $c_1, c_2, c_3$ of cone angle $\pi$ and one boundary $\overline{C}$ with length $c$. $Q$ is a quadruply branched cover of ${\mathcal O}$ with $\omega_i$ and $\omega_i'$ projecting to the same cone singularity $c_i$ ($i=1,2,3$) on ${\mathcal O}$. The four peripheral geodesics $C_i$ on $Q$ cover the peripheral geodesic $\overline{C}$ on ${\mathcal O}$ (so they have the same length), and $A$ is a quadruple cover of a simple geodesic $\overline A$ from $c_1$ to $c_2$ in ${\mathcal O}$, so $a=4 \ell({\overline A})$. $M_A^i$ $i=1,2,3,4$ cover the same simple orthogeodesic $\overline{M_A}$ on ${\mathcal O}$ from ${\overline C}$ to ${\overline A}$, so $m_A=\ell(\overline{M_A})$. $P_A$ and $P_A'$ each double covers the same simple orthogeodesic $\overline{P_A}$ on ${\mathcal O}$ from $c_3$ to $\overline A$, so $p_A=2\ell(\overline{P_A})$, and $Q_A$ and $Q_A'$ double covers the same simple orthogeodesic ${\overline Q_A}$ on ${\mathcal O}$ from $c_3$ to ${\overline C}$ which is disjoint from $\overline A$, so that $q_A=2 \ell(\overline{Q_A})$. Again, by reversing the construction, and keeping track of the covering data, we see every hyperbolic orbifold ${\mathcal O}$ with three cone points $c_1,c_2,c_3$ of angle $\pi$ and one geodesic boundary of length $c$ is quadruply branch covered by a unique hyperbolic four-holed sphere $Q$ with geodesic boundaries all of length $c$. [**Tori and four-holed spheres**]{}. To summarize, every hyperbolic one-holed torus $T$ with geodesic boundary of length $2c$ is a double branched cover over a unique hyperbolic orbifold ${\mathcal O}$ with three cone points of angle $\pi$ and one geodesic boundary of length $c$, which in turn is quadruply branched covered by a unique hyperbolic four-holed sphere $Q$ with four geodesic boundaries all of length $c$. Furthermore, if $T$ and $Q$ are the hyperbolic one-holed torus and four-holed sphere covering the same orbifold ${\mathcal O}$, a non-peripheral simple closed geodesic $B$ on $T$ double covers a simple geodesic arc $\gamma$ from $c_i$ to $c_j$ on $\mathcal{O}$, which in turn is quadruply covered by a unique non-peripheral simple closed geodesic $A$ on $Q$. Moreover, the simple orthogeodesic $M_B$ (or $M_B'$) on $T$ covers the simple orthogeodesic on $\mathcal{O}$ from $\overline{K}$ to $\gamma$, which in turn is covered by the simple orthogeodesic $M_A^1$ (or $M_A^2$, $M_A^3$, $M_A^4$) on $Q$; the simple orthogeodesic $Q_B$ on $T$ double covers the simple orthogeodesic on $\mathcal{O}$ from $c_l$ to $\gamma$, which in turn doubly covered by the simple orthogeodesic $P_A$ (or $P_A'$) on $Q$; and the simple orthogeodesic $P_B$ on $T$ double covers the simple orthogeodesic on $\mathcal{O}$ from $c_l$ to the boundary which is disjoint from $\gamma$, which is in turn doubly covered by the simple orthogeodesic $P_A$ (or $P_A'$) on $Q$. Therefore there is a one to one correspondence between the set of non-peripheral simple closed geodesics $\{B\}$ on $T$ and the set of non-peripheral simple closed geodesics $\{A\}$ on $Q$, and if the simple closed geodesic $B$ on $T$ corresponds to the simple closed geodesic $A$ on $Q$, we have $\ell(A)=2\ell(B)$, $m_A=m_B$, $p_A=q_B$ and $q_A=p_B$. [**Identities for one-holed tori**]{}. From the discussion above, we see that the identity in Lemma \[le:4boundary\] induces the following identity on the hyperbolic one-holed torus: \[thm:oneholednew\] For a hyperbolic one-holed torus $T$ with boundary geodesic $K$ of length $k>0$, we have $$\label{eq:1holednon0} \sum_{B} \left\{\mathcal{L}\left(\tanh^2 (\frac{q_B}{2})\right)+2\mathcal{L}\left(\tanh^2(\frac{m_B}{2})\right) - 2 La\left(e^{-\frac{k}{2}}, \tanh^2(\frac{m_B}{2})\right) \right\}=\frac{\pi^2}{2}$$ where $B$ is over all non-peripheral simple closed geodesics on $T$. For every non-peripheral simple closed geodesic $B$ on $T$, $q_B$ and $m_B$ are the lengths of the simple orthogeodesics shown in Figure \[fig:1holedtorus\]. Note for a fixed $k>0$, there is a unique real $m''$ such that $\sinh(m'') \sinh(\frac{k}{4}) = 1$. For any hyperbolic structures on $T$ and every non-peripheral simple closed geodesics $B$ on $T$, $m'' < m_B$, thus we have $ e^{-\frac{k}{2}} < \tanh^2(\frac{m_B}{2})$. Theorem \[thm:oneholednew\] in turn is equivalent to Theorem \[thm:oneholed\] by using the reformulation of . Let $k \to 0$, we get a once-punctured torus T. The relation between once-punctured tori and quadruply-punctured spheres applies, thus from the identity for a hyperbolic quadruply-punctured sphere, we have Theorem \[thm:oncepuncturetorus\]. There is an equivalent interpretation of Theorem \[thm:oncepuncturetorus\], which is motivated by Bowditch [@Bowditch96]. Let $\Gamma$ be the fundamental group of T which a free group of rank two i.e. $\Gamma=\langle \alpha, \beta \rangle$. Any complete hyperbolic structure on a once-punctured torus T is determined by a discrete faithful and type-preserving (trace of commutator of a generating pair for $\pi_1(T)$ is $-2$) representation $\rho:\Gamma \to {\rm{PSL}}(2,\mathbb{R})$. Define $g' \sim g$ in $\Gamma$ if and only if $g'$ is conjugate to $g$ or $g^{-1}$, thus $\Gamma/\!\!\sim$ represents the set of free homotopy classes of (unoriented) closed curves on T. Let $\Omega \subset \Gamma /\!\!\sim$ be the set of free homotopy classes of essential (i.e. non-trivial and non-peripheral) simple closed curves on T. For each free homotopy class $X \in \Omega$, there exists a unique non-peripheral simple closed geodesic $B$ on T representing $X$ and we have $$\label{eq:basichyprep} {\rm{tr}}^2(\rho(g))=4\cosh^2(\frac{\ell(B)}{2})$$ where $g \in \Gamma$ is an arbitrary element representing the homotopy class $X$ and $\ell(B)$ denotes the length of $B$. We define a function $\phi$ on $\Omega$ by $\phi(X):=x={\rm{tr}}^2(\rho(g))$ where $g \in \Gamma$ represents the homotopy class $X$. Then we have \[thm:puncturenew\] For any discrete faithful, type-preserving representation $\rho:\Gamma \to {\rm{PSL}}(2,\mathbb{R})$, $$\label{eq:newoncepuncture} \sum_{X \in \Omega} \left\{ \mathcal{L}(\frac{4}{x}) + 2\mathcal{L}(\frac{x}{x+\sqrt{x^2-4x}}) -2\mathcal{L}(\frac{\sqrt{x^2-4x}}{x+\sqrt{x^2-4x}})\right\} = \frac{\pi^2}{2}$$ where the sum is over all free homotopy classes in $\Omega$. Reformulation of the Luo-Tan identity {#ss:pseudo} ===================================== In the following, we reformulate the function $g$ of Theorem 1.1 in [@LuoTan11], which is $$\label{eq:originalG} g(T)=4 \pi^2-\sum_B 8\left[2 La\left(e^{-\ell(B)}, \tanh^2(\frac{m_B}{2})\right) +\mathcal{L}\left(\operatorname{sech}^2 (\frac{p_B}{2})\right)\right]$$ where the sum is over all non-peripheral simple closed geodesics $B$ on $T$, and for each non-peripheral simple closed geodesic $B$, $\ell(B)$ is the length of $B$, $m_B$ and $p_B$ are the lengths of the simple orthogeodesics as in Figure \[fig:1holedtorus\]. Here the inequality $e^{-\ell(B)}<\tanh^2(\frac{m_B}{2})$ holds for every non-peripheral simple closed geodesic $B$ on $T$. By using to write $4 \pi^2$ as a sum over $B \in SG_T$, we have: $$\begin{aligned} g(T)=\sum_B 8\Bigg\{ & \mathcal{L}\left(\tanh^2 (\frac{q_B}{2})\right) + 2\mathcal{L}\left(\tanh^2 (\frac{m_B}{2})\right) -\mathcal{L}\left(\operatorname{sech}^2(\frac{p_B}{2})\right)\\ & - 2 La\left(e^{-\ell(B)}, \tanh^2(\frac{m_B}{2})\right) - 2La\left(e^{-\frac{k}{2}}, \tanh^2(\frac{m_B}{2})\right)\Bigg\} {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:newgfun}\end{aligned}$$ where the sum is over all non-peripheral simple closed geodesics $B\in SG_T$, and for each non-peripheral simple closed geodesic $B$, $\ell(B)$ is the length of $B$, $m_B$, $p_B$ and $q_B$ are the lengths of the simple orthogeodesics $M_B, P_B$ and $Q_B$ respectively as in Figure \[fig:1holedtorus\]. For any hyperbolic surface $S$ and an embedded geometric 3-holed sphere $P \subset S$, we denote the boundary geodesics by $L_1, L_2, L_3$ with lengths $l_1,l_2,l_3$, respectively. For $\{i, j, k\} = \{1,2,3\}$, let $M_i$ be the shortest geodesic arc between $L_j$ and $L_k$ with its length $m_i$ and $B_i$ be the shortest non-trivial geodesics arc from $L_i$ to itself with its length $p_i$. We define $$\begin{aligned} f_1(P)&:=4 \pi^2 - 8\left[\sum_{i=1}^3 \left(\mathcal{L}\left(\operatorname{sech}^2(\frac{m_i}{2})\right) + \mathcal{L}\left(\operatorname{sech}^2(\frac{p_i}{2})\right)\right) + \sum_{i \neq j} La\left(e^{-l_i},\tanh^2(\frac{m_j}{2})\right) \right]\\ &:= 8\left[\sum_{i=1}^3 \left(\mathcal{L}\left(\tanh^2(\frac{m_i}{2})\right) - \mathcal{L}\left(\operatorname{sech}^2(\frac{p_i}{2})\right)\right) -\sum_{i \neq j} La\left(e^{-l_i},\tanh^2(\frac{m_j}{2})\right) \right]. {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:newf1fun}\end{aligned}$$ For a quasi-embedded geometric 3-holed sphere $P$ of $S$, let $B \subset S$ be the simple closed geodesic which is the image of two of the boundaries of $P$, and $T$ the one-holed torus which is the image of $P$ in $S$. We denote the length of $B$ by $\ell(B)$, and let $m_B$, $p_B$ and $q_B$ the lengths of the simple orthogeodesics $M_B, P_B$ and $Q_B$ as in Figure \[fig:1holedtorus\]. We define $$\begin{aligned} f_2(P):=8\Bigg[&\mathcal{L} \left(\tanh^2 (\frac{q_B}{2})\right) + 2\mathcal{L}\left(\tanh^2 (\frac{m_B}{2})\right) - \mathcal{L}\left(\operatorname{sech}^2(\frac{p_B}{2})\right)\\ &- 2 La\left(e^{-\ell(B)}, \tanh^2(\frac{m_B}{2})\right) - 2La\left(e^{-\frac{k}{2}},\tanh^2(\frac{m_B}{2})\right)\Bigg]. {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:newf2fun}\end{aligned}$$ Thus becomes $$g(T)= \sum_{P} f_2(P)$$ where the sum is over all quasi-properly immersed geometric 3-holed spheres $P$ of $T$. Therefore we get Corollary \[cor:newform\] simply by reformulating the function $g$ in . [99]{} George E Andrews, Richard Askey and Ranjan Roy, *Special Functions*, Volume 71 of the Encyclopedia of mathematics and its applications, Cambridge University Press, Cambridge, UK; New York, 1999. Ara Basmajian, *The orthogonal spectrum of a hyperbolic manifold*, Amer. J. Math. 115 (1993), no. 5, 1139–1159. B. H. Bowditch, *A proof of [M]{}c[S]{}hane’s identity via [M]{}arkoff triples*, Bull. London Math. Soc. 28 (1996), no. 1, 73–78. Martin Bridgeman, *Orthospectra of geodesic laminations and dilogarithm identities on moduli space*, Geom. Topol. 15 (2011), no. 2, 707–733. Martin Bridgeman and Jeremy Kahn, *Hyperbolic volume of manifolds with geodesic boundary and orthospectra*, Geom. Funct. Anal. 20 (2010), no. 5, 1210–1230. Martin Bridgeman and Ser Peow Tan, *Moments of the boundary hitting function for the geodesic flow on a hyperbolic manifold.*, Geom. Topol. 18 (2014), no. 1, 491-520. Martin Bridgeman and Ser Peow Tan, *Identities on Hyperbolic Manifolds*, arXiv:1309.3578\[math.GT\]. Hengnan Hu, *Identities on hyperbolic surfaces, Group actions and the Markoff-Hurwitz equations*, Ph.D. Thesis, National University of Singapore, 2013. A. N. Kirillov, *Identities for the Rogers dilogarithmic function connected with simple Lie algebras*, Journal of Soviet Mathematics 47(1989), 2450–2459. Leonard Lewin, *Structural Properties of Polylogarithms*, of the Mathematical Surveys and Monographs, American Mathematical Society, Providence, Rhode Island, 1991. Feng Luo and Ser Peow Tan, *A dilogarithm identity on moduli spaces of curves*, arXiv:1102.2133\[math.GT\], to appear, J. Differential Geom. Greg McShane, *A remarkable identity for lengths of curves*, Ph.D. Thesis, University of Warwick, 1991. Greg McShane, *Simple geodesics and a series constant over [T]{}eichmuller space*, Invent. Math. 132 (1998), no. 3, 607–632. Greg McShane, *Weierstrass points and simple geodesics*, Bull. London Math. Soc. 36 (2004), no. 2, 181–187. Greg McShane, *Simple geodesics on surfaces of genus 2*, Ann. Acad. Sci. Fenn. Math. 31 (2006), no. 1, 31–38. Maryam Mirzakhani, *Simple geodesics and Weil-Petersson volumes of moduli spaces of bordered Riemann surfaces*, Invent. Math. 167 (2007), no. 2, 179–222. L. J. Roger, *On [F]{}unction [S]{}um [T]{}heorems [C]{}onnected with the [S]{}eries $\sum_{1}^{\infty} \frac{x^n}{n^2}$*, Proc. London Math. Soc. S2-4 (1907), no. 1, 169–189. Ser Peow Tan, Yan Loi Wong and Ying Zhang, *Generalizations of McShane’s identity to hyperbolic cone-surfaces*, J. Differential Geom. [**72**]{} (2006), no. 1, 73–111. Ser Peow Tan, Yan Loi Wong and Ying Zhang, *Necessary and sufficient conditions for McShane’s identity and variations*, Geometriae Dedicata, 119 No. 1 (2006), 119-217. Ser Peow Tan, Yan Loi Wong and Ying Zhang, *Generalized Markoff maps and McShane’s identity*, Adv. Math. 217 (2008), no. 2, 761–813. [^1]: The second author is partially supported by the National University of Singapore academic research grant R-146-000-186-112.
{ "pile_set_name": "ArXiv" }
--- bibliography: - 'mybib.bib' --- Introduction ============ The unsolved problem of the origin of cosmic rays is one of the biggest challenges in high-energy astrophysics. In hadronic interactions of cosmic rays with matter, high-energy neutrinos are produced in the direct environment of cosmic ray sources. Possible candidates for these sources are, for example, Gamma Ray Bursts (GRBs) [@Mesz_GRB; @Waxman_Nu] or Active galactic Nuclei (AGNs) [@HE_AGNs; @WaxBa_nu]. From Fermi Acceleration, the resulting energy-dependent differential flux ($\frac{\mathrm{d}\phi}{\mathrm{d}E_\nu}$) of astrophysical neutrinos is expected to follow a power law of $E_\nu^{-\gamma}$, where $E_\nu$ is the neutrino energy and $\gamma = 2+\epsilon$ is close to the spectral index of the cosmic ray production process [@FrancisAGNs]. The so-called inefficiency ($\epsilon$) is often assumed to be $0\leq \epsilon \leq 1$, while $\gamma \approx 2$ is favored by Fermi Acceleration [@FrancisAGNs].\ Neutrinos are also expected in coincidence with high-energy photons. Thus, the measurement of thousands of gamma-ray sources, like AGNs [@AGNGammaToNu; @FermiAGNs], provides a large number of potential neutrino sources.\ Since neutrinos do not experience deflections or scattering, they are ideal messengers for observing and tracing the hadronic interactions described above. The detection of high-energy neutrinos of cosmic origin would give important insights for identifying the sources of cosmic rays [@Tjus_sources; @WaxBah].\ The IceCube neutrino telescope [@icecube] recently detected a diffuse cosmic flux of high-energy neutrino events in two years of IceCube data by searching for neutrino-induced events with an interaction vertex within the detector [@HESE_2year]. This permits one to reduce the dominant atmospheric muon background. In a follow-up analysis, this diffuse flux was investigated in more detail using three years of IceCube data [@HESE]. This high energy starting events analysis (HESE) yielded 37 events compared to an expected background of 15.0 events from atmospheric muons and neutrinos. These events could provide information about potential cosmic ray astrophysical sources and motivate additional searches. IceCube, with its large field-of-view, offers a unique opportunity to study the production and interaction of high-energy cosmic rays using neutrinos. The detector, which is located at the geographical South Pole, has a detection volume of $\sim1\,\mathrm{km}^3$ deep in the Antarctic glacier and an additional $\sim1\,\mathrm{km}^2$ surface air shower detector, called IceTop. The IceCube detector consists of digital optical modules (DOMs) [@DAQ], placed on strings deployed vertically at depths between 1450m and 2450m. The strings hold 60 DOMs each equipped with a photomultiplier tube and digitizing electronics to detect neutrinos by measuring Cherenkov radiation of their secondary particles [@PMTs]. A detailed description of the data acquisition system can be found in [@DAQ]. The detector was built in several stages between 2005 and 2010, such that each year several strings were added, until reaching the final configuration of 86-strings, containing more than 5000 DOMs. IceCube has completed different point source searches, including an energy-dependent likelihood point source search scanning the full sky [@all_sky], as well as searches for flaring and periodic neutrino emission [@flare]. Additionally there are searches for diffuse neutrino emission looking for deviations in the two dimensional distribution of energy and zenith angle [@diffuse]. Point source searches are most sensitive for finding individual sources of astrophysical neutrinos among the background of atmospheric events (neutrinos and muons from cosmic ray interactions at Earth). Diffuse searches, on the other hand, are most sensitive for detecting within this background the presence of high-energy astrophysical neutrinos throughout the sky, without identifying individual sources. In between these two scenarios is the possibility that many weak sources exist. These could contribute to the detected diffuse signal and create a small number of events clustering on the background event distribution, while the individual clusters remain too weak to be detected by the point source searches. In this paper we present two searches for such small-scale clustering. The first is an autocorrelation test using a two-point (2-pt) correlation function performed in the northern and southern hemisphere. Since most of the signal-like events are at high energies, we extended the autocorrelation test to include the most energetic events in an additional test (2-pt HE). The second test is a multipole expansion of the skymap of neutrino arrival directions. The goal of both methods is to gain sensitivity to faint sources at unknown positions in the sky with unidentified energy spectra using a three year data-set of the partially completed detector. Both methods are complementary to previous searches and therefore an important addition to IceCube searches for a cosmic neutrino flux. A similar search for clustering was recently performed by ANTARES, details can be found in [@Antares_2pt]. The paper is organized as follows: In section \[datasample\], the data sample used and the generation of pseudo experiments is described. The 2-pt correlation test is explained in section \[sec:method\_2pt\], and the multipole analysis in section \[sec:method\_mp\]. The performance of both analyses, decribed by the discovery potential, is given in section \[sec:DataAnalysis\]. In section \[sec:Result\], the experimental result is presented and exclusion limits are calculated. Systematic uncertainties are discussed in section \[sec:systematics\]. In section \[sec:conclusion\] a conclusion is drawn. Conclusions {#sec:conclusion} =========== Two methods to search for a small-scale anisotropy with IceCube were presented. The results of both searches are consistent with background expectations with small underfluctuation. Depending on the number of assumed sources, the resulting upper limits range from $10^{-8}$ GeV/$\mathrm{cm}^{2}\mathrm{s}^{-1}$ for one source to $10^{-9}$ GeV/$\mathrm{cm}^{2}\mathrm{s}^{-1}$ for 3500 $E^{-2}$ neutrino sources in the northern hemisphere. Limits were also set for other assumed energy spectra, including $E^{-3}$ and $E^{-2.25}$ in the northern hemisphere. Since both analyses use a data-driven background estimation they are more robust against systematic uncertainties than estimations from MC simulations.\ Considering the astrophysical flux previously observed in IceCube [@HESE], a small number ($\leq 10$) of isotropically distributed sources in the northern hemisphere of very hard energy spectra, like $E^{-2}$, is excluded as it was by former IceCube analyses [@all_sky]. For softer energy spectra, the analyses presented here disfavor the observed flux to come from less than $\sim20$ sources for $E^{-2.25}$ and from less than $\sim 5000$ sources for $E^{-3}$. Additionally, for sources distributed along the galactic plane in the northern hemisphere the autocorrelation limit is close to the flux predicted by HESE. In the southern hemisphere, the data sample contains predominantly atmospheric muons from cosmic ray air showers above the detector. Due to this background the autocorrelation analysis is not sensitive to a population of sources at the HESE flux level. For all these tests, the sources are assumed to have the same flux at Earth, since the true spatial flux distribution is not known for the observed astrophysical flux.\ For hard energy spectra, the 2-pt correlation analysis is more sensitive than the multipole analysis since it uses the energy information as an additional variable. For soft energy spectra, the multipole analysis becomes slightly more competitive.\ **Acknowledgements** We acknowledge the support from the following agencies: U.S. National Science Foundation-Office of Polar Programs, U.S. National Science Foundation-Physics Division, University of Wisconsin Alumni Research Foundation, the Grid Laboratory Of Wisconsin (GLOW) grid infrastructure at the University of Wisconsin - Madison, the Open Science Grid (OSG) grid infrastructure; U.S. Department of Energy, and National Energy Research Scientific Computing Center, the Louisiana Optical Network Initiative (LONI) grid computing resources; Natural Sciences and Engineering Research Council of Canada, WestGrid and Compute/Calcul Canada; Swedish Research Council, Swedish Polar Research Secretariat, Swedish National Infrastructure for Computing (SNIC), and Knut and Alice Wallenberg Foundation, Sweden; German Ministry for Education and Research (BMBF), Deutsche Forschungsgemeinschaft (DFG), Helmholtz Alliance for Astroparticle Physics (HAP), Research Department of Plasmas with Complex Interactions (Bochum), Germany; Fund for Scientific Research (FNRS-FWO), FWO Odysseus programme, Flanders Institute to encourage scientific and technological research in industry (IWT), Belgian Federal Science Policy Office (Belspo); University of Oxford, United Kingdom; Marsden Fund, New Zealand; Australian Research Council; Japan Society for Promotion of Science (JSPS); the Swiss National Science Foundation (SNSF), Switzerland; National Research Foundation of Korea (NRF); Danish National Research Foundation, Denmark (DNRF) \
{ "pile_set_name": "ArXiv" }
--- abstract: 'Person re-identification is to retrieval pedestrian images from no-overlap camera views detected by pedestrian detectors. Most existing person re-identification (re-ID) models often fail to generalize well from the source domain where the models are trained to a new target domain without labels, because of the bias between the source and target domain. This issue significantly limits the scalability and usability of the models in the real world. Providing a labeled source training set and an unlabeled target training set, the aim of this paper is to improve the generalization ability of re-ID models to the target domain. To this end, we propose an image generative network named identity preserving generative adversarial network (IPGAN). The proposed method has two excellent properties: 1) only a single model is employed to translate the labeled images from the source domain to the target camera domains in an unsupervised manner; 2) The identity information of images from the source domain is preserved before and after translation. Furthermore, we propose IBN-reID model for the person re-identification task. It has better generalization ability than baseline models, especially in the cases without any domain adaptation. The IBN-reID model is trained on the translated images by supervised methods. Experimental results on Market-1501 and DukeMTMC-reID show that the images generated by IPGAN are more suitable for cross-domain person re-identification. Very competitive re-ID accuracy is achieved by our method.' author: - | Jialun Liu, Wenhui Li, Hongbin Pei\ College of Computer Science and Technology, Jilin University, China\ bibliography: - 'egbib.bib' title: 'Identity Preserving Generative Adversarial Network for Cross-Domain Person Re-identification' --- ![Example images from 2,4,6 cameras of Market-1501 and 6,7,8 cameras of DukeMTMC-reID, respectively. The images with a same identity have different appearances in different camera views. []{data-label="fig:carton"}](fig1.png){width="1.0\linewidth"} Introduction {#Introduction} ============ As one of the most challenging problems in the field of computer vision, the re-ID task aims at searching for the relevant images from a gallery according to a query. The traditional person re-identification approaches [@koestinger2012large; @liao2015person; @zhang2016learning] are based on hand-crafted features. With the recent development of deep learning, many excellent deep learning-based supervised methods [@zheng2016person; @sun2017svdnet; @hermans2017defense; @sun2017beyond; @zhong2018camera] are proposed for the person re-identification task. Although those approaches has obtained a dramatic performance improvement, they have many limitations in the real-world mainly including 1) the image labeling process is very expensive and impractical for supervised learning; 2) when re-ID models trained on a source domain and used on a target domain, the bias between the source and the target domain leads to performance degradation notably. Thus, it is critical to improve the generalization capacity of the supervised re-ID methods. A common strategy to solve this problem is unsupervised domain adaptation [@hoffman2017cycada; @long2015learning; @tzeng2017adversarial; @ghifary2014domain], which assume that the source and the target domain contain a same set of classes. However, the assumption is not appropriate for the person re-ID task in which the source and the target domain have completely different persons (classes). Very recently, few person re-ID methods[@deng2018image; @wei2017person] based on unsupervised domain adaptation are proposed, which use a generative adversarial network (GAN) to translate images from source domain to target domain. However, these methods only take into consideration the general gap between source domain and target domain but ignore the bias between source domain and target camera domains. Actually, images captured by different cameras often have many obviously different styles. As Figure \[fig:carton\] shows, the images with a same identity have different appearances in different camera views. Images captured by one camera can be regarded as a subdomain of target domain. Thereby, DukeMTMC-reID and Market1501 has 8 and 6 subdomains respectively because of their camera number. In the real world, the distribution of one subdomain may greatly differ from the distribution of the other ones, because the types of camera and scenes for image acquisition are different. In this case, it is improper to take the target domain as a whole. The better solution of domain adaptation is to reduce the *bias* between source domain and each subdomain (camera domain) in target domain. In this paper, we propose a novel and efficient unsupervised domain adaptation approach named Identity Preserving Generative Adversarial Network (IPGAN), which can generate images for target camera domains learning. The design of IPGAN is motivated by three aspects. Firstly, A significant motivation is to reduce the *bias* between source and each target camera domains. Secondly, although the style of images in source domain is changed, the translated image should preserve the same identity information with its corresponding original image. Thirdly, the computational cost of cross domain person re-ID should not be very expensive because the dataset usually very large. To achieve the above three objectives, we design IPGAN which is inspired by StarGAN [@choi2017stargan]. In IPGAN, we design a novel semantic discriminator to implement the semantic constraint that the identity information of images from source domain keeps invariance before and after translation. Such a semantic discriminator brings more challenges to a generator in GAN framework. Using IPGAN, we can create a new dataset in an unsupervised manner, which inherits the labels from source domain and has the style of target camera domains. Then, we train the reID model on the new dataset in a supervised way. For the person re-ID task, we further propose IBN-reID model inspired by [@pan2018two]. The model can eliminate appearance variance in shallow layers, and holds discrimination of the learned features in deep layers. In the model, instance normalization and batch normalization are integrated. Instance normalization provides visual and appearance invariance, while batch normalization accelerates training and preserves discriminative features. The IBN-reID model has better generalization ability than the baseline model [@deng2018image], especially in the case of deploying the model trained on source domain to target domain without any domain adaptation. In summary, this paper has the following contributions: - To solve domain adaptation in person re-ID task, we propose IPGAN, a novel and efficient unsupervised learning methods. It works by mapping the images from source domain to target camera domains with one single model meanwhile, preserving the identity information of images from source domain. - we present IBN-reID model which intuitively induces appearance invariance into re-ID model. It has better generalization ability than baseline model. - Many experimental results show that the proposed methods achieve very competitive re-ID accuracy and it is efficient and applied. ![image](fig2.png){width="1\linewidth"} Related Work {#Related Works} ============ **Generative Adversarial Networks.** Generative adversarial networks (GANs) [@Goodfellow2014Generative] has shown remarkable performance improvement in various computer vision tasks, especially image-to-image translation, in recent years. For image-to-image translation task, pix2pix [@isola2017image] uses a conditional GANs to learn mappings from input to output images by combining adversarial loss and $L_{1}$ loss. However, this method needs paired data to train its model. For unpaired image-to-image translation, several methods are proposed [@zhu2017unpaired; @kim2017learning; @yi2017dualgan; @zheng2017unlabeled]. UNIT [@liu2017unsupervised] combines variational autoencoders [@kingma2013auto] and CoGAN [@liu2016coupled], in which the two generators share same weights. CycleGAN [@zhu2017unpaired] and DiscoGAN [@kim2017learning] use a cycle consistency to preserve key attributes. However, all the aforementioned frameworks only consider the mapping from source domain to target domain. Different form them, we propose a new framework which can translate images from source domain to each target camera domains using only a single model and be used to improve the performance of cross domain person re-ID. **Unsupervised domain adaptation.** Our work is one of unsupervised domain adaptation method where the labels of target images are unavailable [@long2015learning; @Chen2018Person; @Ajakan2014Domain]. In existing methods, a popular idea is to reduce the divergence between source domain and target domain [@sun2016return; @gretton2012kernel; @Ganin2015Unsupervised]. CORAL [@sun2016return] gets good performance in various computer vision tasks by aligning the mean and covariance of two data distributions of source and target domain. By introducing the Maximum Mean Discrepancy (MMD), [@gretton2012kernel] try to reduce the MMD distance between the two domains. DANN [@Ganin2015Unsupervised] integrates the gradient reversal layer (GRL) into the standard architecture to ensure the similar distribution of features on both domains. There are many methods which attempt to produce fake-labels for the unlabeled samples [@Rohrbach2013Transfer; @Saito2017Asymmetric; @sener2016learning; @zhu2006semi]. For instance, [@zhu2006semi] trains a classifier on labeled and fake-labeled samples to the predict labels of unlabeled samples. In [@Saito2017Asymmetric], three classifiers are modelled to generate fake-labels for samples in target domain. Recently, many domain adaptation approaches [@hoffman2017cycada; @bousmalis2017unsupervised] based on Generative Adversarial Networks [@Goodfellow2014Generative] are proposed. CyCADA [@hoffman2017cycada] achieves domain adaptation at both the pixel-level and feature-level by utilizing pixel cycle consistency and semantic losses. However, the above domain adaption approaches can not be used for the cross domain person re-ID task, because they assume that the source and target domain have same class labels. Actually, in the community of person re-ID, identities(classes) of different datasets have no overlap. **Unsupervised person re-ID.** Most existing re-ID models are based on supervised learning [@koestinger2012large; @Xiong2014Person; @Wang2014Person; @zhang2016learning; @Wang2016Joint; @Xiao2016Learning; @Chen2016A; @subramaniam2016deep]. These models suffer from poor scalability in the real-world environment. To solve this scalability issue and improve the generalization ability, unsupervised methods based on hand-crafted features [@Farenzena2010Person; @Kodirov2016Person; @Rui2017Person; @Gray2008Viewpoint; @liao2015person; @cheng2011custom; @kodirov2015dictionary; @Wang2014Person; @bazzani2013symmetry] can be chosen and applied. These methods aim to design or learn robust feature for person re-ID. But, they ignore the distribution of samples in the dataset and yield much weaker performance on large-scale dataset than supervised learning methods. ![image](fig3.PNG){width="1\linewidth"} Recent works [@Liu2017Stepwise; @fan2018unsupervised] employ deep learning technology to estimate the labels of unlabeled target dataset. [@fan2018unsupervised] proposes an unsupervised approach which utilizes K-means to offer fake-labels for the unlabeled samples and train the re-ID model with those samples iteratively. [@Liu2017Stepwise] uses K-reciprocal nearest neighbors to get proximate label information for unsupervised video re-ID. A few unsupervised domain adaptation for person re-ID methods has proposed [@wang2018transferable; @wei2017person; @deng2018image; @peng2016unsupervised; @zhong2018generalizing]. Based on CycleGAN [@zhu2017unpaired], SPGAN translates images from source domain to target domain via self-similarity [@deng2018image] which works by keeping the underlying identity information during translation. Then, the translated images are utilized to train re-ID model in a supervised way. Similar with [@deng2018image], [@wei2017person] keep same contents during transferring. The above methods attempt to reduce the bias between source and target domain on image space and feature space, however they all ignore the divergence of image style caused by target camera domains. [@zhong2018camstyle] considers the intra-domain image variations caused by target cameras and learn discriminative representations of target domain. But this method cannot keep the same identity information between original images and translated images because identify semantic constraint is not considered. Furthermore, the above approaches have limited scalability in handling multiple domains since different models need to be trained on each pair of domains. Unlike them, our method explicitly considers the difference between source domain and target camera domains, and we can learn the relations among multiple domains using only a single model. Proposed Method {#Proposed Method} =============== Baseline Re-ID Model {#sec:baseline} -------------------- The person re-ID task can be regarded as a classification problem. Thus, we use a classification model ResNet-50 [@he2016deep] as a backbone network for person re-ID. We use two FC layers to replace the final 1,000-dim fully connected(FC) layer after the Pooling-5 layer. The dimensions of the two FC layer are 1,024 and $N$, where $N$ is the number of classes in the dataset. The cross-entropy loss function is used to optimizing the model parameters. Our training process follows the ID discriminative embedding (IDE) [@zheng2016person]. ![image](fig4.png){width="1.0\linewidth"} StarGAN Revisit {#StarGAN} --------------- StarGAN [@choi2017stargan] is a novel and efficient generative adversarial network that learns the mapping relations among multiple domains using only a single model. To make the generated images indistinguishable from real images, the adversarial loss is defined as: $$\begin{split} \mathcal{L}_{adv}(x,c) = & \thinspace {\mathbb{E}}_{x} \left[ \log{{D}_\mathcal{T}(x)} \right] \> \> + \\ & \thinspace {\mathbb{E}}_{x, c}[\log{(1 - {D}_\mathcal{T}(G(x, c)))}], \end{split} \label{eq1}$$ where $G$ generates an image $G(x, c)$ to fake $D$. The term ${D}_\mathcal{T}(x)$ is a probability distribution over sources. The goal of StarGAN is to translate x to an output image y which is classified as the target domain $c$. For this goal, a classifier is added on the top of $D$. The domain classification loss of real/fake images is defined as: $$\mathcal{L}_{dom}^{r}(x,c') = {\mathbb{E}}_{x, c'}[-\log{{L}_{dom}(c'|x)}], \label{eq2}$$ $$\mathcal{L}_{dom}^{f}(x,c) ={\mathbb{E}}_{x, c}[-\log{{D}_{dom}(c|G(x, c))}]. \label{eq3}$$ where the term ${D}_{dom}(c'|x)$ represents a probability distribution over domain labels computed by $D$. By minimizing Eq.\[eq2\] and Eq.\[eq3\], $D$ learns to classify a real image $x$ to its original domain $c'$ and fake images to target domain $c$. To guarantee that translated images preserve the content of their original images, StarGAN introduces a cycle-consistent loss [@zhu2017unpaired; @kim2017learning] which is defined as: $$\mathcal{L}_{rec}(x,c,c') = {\mathbb{E}}_{x, c, c'} [{||x - G(G(x, c), c')||}_{1} ], \label{eq4}$$ IPGAN {#IPGAN} ----- One pedestrian has different appearances in different camera views and the bias is shown in Figure \[fig:carton\]. In this paper, each camera domain in target domain is defined as a subdomain. In the real-world, it is entirely possible that the distribution of one subdomain may greatly differ from the distribution of the other ones, it is improper to take the target domain as a whole. A better way to smooth the bias between source and target domain is to learn image-image translation models that translate images from source domain to target camera domains rather than the whole target domain. To this end, we propose the Identity Preserving Generative Adversarial Network (IPGAN). Our method can ensure the transferred image has a similar style as the style in target camera domain. The method is also able to keep the identity information of images from source domain during the translation. IPGAN consists of a style transfer model $G(x, c)$, a domain discriminator ${D}_{dom}$, and a semantic discriminator ${D}_{sem}$, as illustrated in Figure \[fig:IPGAN-framework\]. The construction of IPGAN requires a source training set , the identity labels of source training set , a target training set, and the camera labels of target training set. Compared to identity labels, camera labels can be obtained along with surveillance videos without tedious and expensive manual labeling. There is no identity information of target set during image-image translation. Thus, IPGAN is an unsupervised learning method for person re-ID. **Source to target camera domains Image-Image Translation.** In this work, we employ StarGAN [@choi2017stargan] to generate new training dataset without identity semantic constraint. Our goal is to train a single generator model that can translate images among the $L+1$ domains which consist of $L$ camera domains from target dataset and one source domain. The objective functions with respect to G and D are given, respectively, as $$\mathcal{L}_{\mathcal{D}_{StarGAN}} = - \mathcal {L}_{adv} + {\lambda}_{dom}\thinspace\mathcal{L}_{dom}^{r}, \label{eq5}$$ $$\mathcal{L}_{\mathcal{G}_{StarGAN}} = \mathcal {L}_{adv} + {\lambda}_{dom}\thinspace\mathcal{L}_{dom}^{f} + {\lambda}_{rec}\thinspace\mathcal{L}_{rec}, \label{eq6}$$ Specifically, following the training strategy in [@choi2017stargan], the generator G contains 2 convolutional layers, 6 residual blocks and 2 transposed convolution layers. The discriminator $D_{dom}$ has the same structure as PatchGANs [@isola2017image]. ---------------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- rank-1 rank-5 rank-10 mAP rank-1 rank-5 rank-10 mAP Supervised Learning 84.8 93.7 96.2 65.3 72.6 84.5 87.6 52.0 Baseline+Direct Transfer 44.3 61.9 69.7 18.4 30.2 45.1 51.5 16.1 Baseline+CycleGAN [@deng2018image] 49.9 67.1 74.2 22.6 39.2 54.8 60.5 20.1 Baseline+CamStyle [@zhong2018camstyle] 58.8 78.2 84.3 27.4 48.4 62.5 68.9 25.1 Baseline+StarGAN **54.3** **74.7** **81.5** **24.4** **44.1** **60.0** **65.8** **21.9** [Baseline+IPGAN]{} **56.4** **76.0** **82.5** **25.6** **46.8** **62.0** **67.9** **26.7** ---------------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- \[table:1\] **Identity semantic constraint loss function.** As analyzed in the above, we aim to preserve the identity information of images from source domain during image-image translation. To fulfill this goal, we design a semantic discriminator $D_{sem}$, which is used for identity preserving. The consistency on person identity is important for person re-ID training. The structure of $D_{sem}$ is similar to the baseline(Section \[sec:baseline\]). The identity semantic constraint loss function is: $$\mathcal{L}_{sem}(x_{s},y) ={\mathbb{E}}_{x_{s}, y}[-\log{{D}_{sem}(y|G(x_{s}, c))}], \label{eq7}$$where $G$ generates an image $G(x_{s}, c)$ conditioned on both the input source domain images $x_{s}$ and the target camera domain label $c$. A $x_{s}$ corresponds to a identity label $y$. In a min-batch, ${D}_{sem}$ aims to classify a generated $G(x_{s}, c)$ to its corresponding original identity label $y$. The above two parts are integrated in a framework IPGAN. Its structure is $\big\{G, ({D}_{dom} ,{D}_{sem})\big\}$ as illustrated in Figure \[fig:IPGAN-framework\]. The generator $G$ maps source domain images to the styles of target camera domains. The discriminator ${D}_{dom}$ is used to distinguish real and fake images and judge the domain that the translated images belong to. The ${D}_{sem}$ enforces the images from source domain keeping identity information after translation. Note that ${D}_{sem}$ is a pretrained classifier with source domain training set and only used to optimize $G$. Only the identity information from source domain is applied. When we train the IPGAN, the parameters of ${D}_{sem}$ is fixed. Finally, the overall IPGAN objective function can be written as: $$\mathcal{L}_{\mathcal{D}_{IPGAN}} = - \mathcal {L}_{adv} + {\lambda}_{dom}\thinspace\mathcal{L}_{dom}^{r}, \label{eq8}$$ $$\begin{split} \mathcal{L}_{\mathcal{G}_{IPGAN}} = & \thinspace \mathcal {L}_{adv} + {\lambda}_{dom}\thinspace\mathcal{L}_{dom}^{f}\> \> + \\ & \thinspace {\lambda}_{rec}\thinspace\mathcal{L}_{rec} + {\lambda}_{sem}\thinspace\mathcal{L}_{sem}, \end{split} \label{eq9}$$ where ${\lambda}_{dom}$, ${\lambda}_{rec}$, and ${\lambda}_{sem}$ are three hyper-parameters, that tradeoff the importance of domain classification, reconstruction loss, and identity semantic loss. We use ${\lambda}_{dom}$ = 1, ${\lambda}_{rec}$ =10, and ${\lambda}_{sem}$ = 1 in our experiments. With the learned IPGAN model, for an image in source domain we generate $L+1$ fake samples whose styles are similar to $L+1$ domains, and meanwhile the identity information is keeping during the Image-Image translation. Note that the $L+1$ samples contain a fake image whose style is same as its original style and the fake image should be ignored. Finally, we train re-ID model with the style transferred images in a supervised way. IBN-re-ID model {#IBN-re-ID model} --------------- To address the problem of the appearance gap between source and target camera domains, a intuitive way is to introduce appearance invariance into re-ID models. Inspired by [@pan2018two], we propose a novel deep network named IBN-reID for person re-ID. In IBN-reID, the appearance variance is mainly reflected in shallow features, and the change of content information is reflected in deep features. The characteristic of IBN-reID is that batch normalization and instance normalization are utilized together in the framework. The advantage of batch normalization is preserving discrimination between individual samples by deep features, but the drawback is that it makes CNNs vulnerable to appearance transforms. On the contrary, instance normalization eliminates individual contrast, but diminishes useful information at the same time. Instance normalization and batch normalization are integrated as IBN-block which learns to capture and eliminate appearance variance, while maintains discrimination of the learned features. To our best knowledge, it’s the first attempt to introduce IBN-block into person re-ID. IBN-reID has better generalization ability than baseline models, in the case of deploying the model trained on source domain to target domain without any domain adaptation. We use ResNet-50 [@he2016deep] as base model that consists of four groups of residual blocks. We add three IN layers after the first convolution layer (conv1) and the first two convolution groups (conv2\_x, conv3\_x), respectively. The latter layers are the same as baseline model (Section \[sec:baseline\]). Experiment ========== Datasets -------- Two widely used benchmark datasets are chosen for experiments, Market-1501 [@zheng2015scalable] and DukeMTMC-reID [@zheng2017unlabeled; @ristani2016performance] because both datasets 1) are large-scale and 2) camera labels for each image is available. **The Market-1501**[@zheng2015scalable] dataset contains 32,668 images from 1501 identities collected from 6 cameras. All of the images are produced by deformable part mode (DPM) [@felzenszwalb2010object]. The dataset is split into two non-over-lapping parts: 12,936 images from 751 identities for training and 19,732 images from 750 for testing. Moreover, 3,368 query images from 750 identities are used to retrieve the matching persons in the gallery. **The DukeMTMC-reID** [@zheng2017unlabeled; @ristani2016performance] is also a large-scale re-ID dataset which is collected from 8 cameras. It contains 16,522 training images from 702 identities, 2,228 query images from another 702 identities and 17,661 gallery images, 36411 images belonging to 1404 identities in total. During the course of the experiment, training sets of Market-1501 and DukeMTMC-reID are used for style transfer. We use rank-1 accuracy and mean average precision (mAP) for evaluation on both datasets. Implementation Details ---------------------- We use Pytorch to train IPGAN on NVIDIA GeForce GTX Titan xp GPU using the training set of Market-1501 and DukeMTMC-re-ID. We use a single model to learn the mapping between source domain and target camera domains. For the generator network, we use instance normalization in all layers except the last output layer. For the discriminator $D_{dom}$ network, we use Leaky ReLU with a negative slope of 0.01. The structure of discriminator $D_{sem}$ is simailr to baseline model. The input images are resized to $128 \times 64$. The learning rate is 0.0001 at the first 100 epochs and linearly reduces to zero for the last 100 epochs. for each image in the source training set, we generate $L$ style-transferred images (the number of target cameras). These fake images are regarded as a new training set which is used to train re-ID model. ----------- --------------------- --------------------- --------------------- --------------------- Real Image Accuracy Fake Image Accuracy Real Image Accuracy Fake Image Accuracy StarGAN 1.0 0.22 1.0 0.28 [IPGAN]{} 1.0 **0.99** 1.0 **0.97** ----------- --------------------- --------------------- --------------------- --------------------- \[table:2\] -------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- rank-1 rank-5 rank-10 mAP rank-1 rank-5 rank-10 mAP Baseline+Direct Transfer 44.3 61.9 69.7 18.4 30.2 45.1 51.5 16.1 Baseline+StarGAN 54.3 74.7 81.5 24.4 44.1 60.0 65.8 21.9 Baseline+IPGAN 56.4 75.6 82.5 25.6 46.8 62.0 67.9 25.7 IBN-reID+Direct Transfer **45.7** **62.8** **70.8** **19.8** **32.4** **47.4** **54.6** **17.3** IBN-reID+StarGAN **56.0** [74.6]{} **81.5** **25.5** **44.6** **60.0** **66.0** **22.3** [IBN-reID+IPGAN]{} **57.2** **76.0** **82.7** **28.0** **47.0** **63.0** **68.1** **27.0** -------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- \[table:3\] Evaluation ---------- **Baseline accuracy.** When the re-ID model is baseline model. The supervised learning method and the direct transfer method are specified in Table \[table:1\]. When trained and tested both on the target set, excellent accuracy can be achieved. However, when trained on source dataset and directly tested on target dataset, the performance drops significantly. For instance, the baseline model trained and tested on Market-1501 achieves $84.8\%$ in item of rank-1 accuracy, but drops to $44.3\%$ when trained on DukeMTMC-reID training set and tested on Market-1501 testing set. The reason is the *bias* of data distributions between different domains. --------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- rank-1 rank-5 rank-10 mAP rank-1 rank-5 rank-10 mAP LOMO [@liao2015person] 27.2 41.6 49.1 8.0 12.3 21.3 26.6 4.8 BOW [@zheng2015scalable] 35.8 52.6 60.3 14.8 17.1 28.8 34.9 8.3 UMDL [@peng2016unsupervised] 34.5 52.6 59.6 12.4 18.5 31.4 37.6 7.3 PUL [@fan2018unsupervised] 45.5 60.7 66.7 20.5 30.0 43.4 48.5 16.4 CAMEL [@yu2017cross] 54.5 - - 26.3 - - - - PTGAN [@wei2017person] 38.6 - 66.1 - 27.4 - 50.7 - SPGAN [@deng2018image] 51.5 70.1 76.8 22.8 41.1 56.6 63.0 22.3 TJ-AIDL [@wang2018transferable] 58.2 74.8 81.1 26.5 44.3 59.6 65.0 23.0 CamStyle [@zhong2018camstyle] **58.8** **78.2** **84.3** 27.4 **48.4** 62.5 **68.9** 25.1 IPGAN 56.4 75.6 82.5 25.6 46.8 62.0 67.9 25.7 [IPGAN+IBN-reID]{} 57.2 76.0 82.7 **28.0** 47.0 **62.8** 68.1 **27.0** --------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- \[table:4\] **The impact of ¡°source domain to target camera domains Image-Image Translation¡±.** Firstly, following the work of [@deng2018image], we use CycleGAN [@zhu2017unpaired] to translate the labeled images from the source domain to the target domain and then train the baseline re-ID model with the translated images in a supervised way. As show in Table \[table:1\]. When trained on DukeMTMC-reID training set and tested on Market-1501 testing set, rank-1 accuracy improves from $44.3\%$ to $49.9\%$ and mAP accuracy improves from 18.4 to 22.6. when trained on Market-1501 training set and tested on DukeMTMC-reID testing set, rank-1 accuracy improves from $30.2\%$ to $39.2\%$ and mAP accuracy improves from $16.1$ to $20.1$. Secondly, we consider the bias between source domain and target camera domains. Very resently, CamStyle [@zhong2018camstyle] is a similar work with ours and implement the translation from source domain to target camera domains with CycleGAN [@zhu2017unpaired], while we use StarGAN [@choi2017stargan] to implement it. As show in Table \[table:1\], compared to CamStyle[@zhong2018camstyle], despite the results of “Baseline+StarGAN” are inferior to it, “Baseline+StarGAN” is more efficient. As show in Figure \[fig:fig5\], when translate from Market-1501(source domain) to DukeMTMC-reID(8 camera domains), CamStyle [@zhong2018camstyle] needs 16 pairs of {G, D} (CycleGAN has two generator-discriminator pairs). When translation from DukeMTMC-reID(source domain) to Market-1501(6 camera domains), it needs 12 pairs of {G, D} on Market-1501. StarGAN [@choi2017stargan] only uses two pairs of $\big\{G, {D}_{dom}\big\}$ to complete it. Results are showed in Table \[table:1\]. Compared to the “Baseline+CycleGAN”, the “Baseline+StarGAN” gains +$4.4\%$ improvements in item rank-1 accuracy and +$1.8$ in item of mAP when tested on Market-1501. When tested on DukeMTMC-reID, the performance gains +$4.9\%$ in item of rank-1 accuracy and +$1.8$ in item of mAP. Through such source to target camera domains adaptation method, effective improvement can be achieved. **The impact of IPGAN.** Compared to StarGAN [@choi2017stargan] , IPGAN is same efficiency as StarGAN [@choi2017stargan] but owns identity semantic constraint. We conduct experiment to verify the influence of the identity semantic constraint. As show in Table \[table:2\], we analyze the difference of images translated by StarGAN and IPGAN. We use DukeMTMC-reID/Market-1501 training set to train two classifiers which are used to predict the identities of translated images. The classification accuracy of DukeMTMC-reID/Market-1501 training set after translated by StarGAN is only $28.0\%/22.0\%$. But, when DukeMTMC-reID/Market-1501 training set is translated by IPGAN, the classification accuracy is $97.0\%/99.0\%$. The fake images generated by IPGAN keep almost the same identity information as the original real images. However, most of the images generated by StarGAN lose identity semantic information. The performance of re-ID is showed in Table \[table:1\]. “Baseline+IPGAN” gains a better performance than “Baseline+StarGAN” in item of rank-k and mAP accuracy. **The impact of ¡°IBN-block¡±.** To further improve re-ID performance on target dataset, we propose IBN-re-ID model (Figure \[fig:IBN-framework\]). As show in Table \[table:3\], compared to methods with baseline re-ID model, when we use IBN-reID model as re-ID model, the rank-k and mAP accuracy are improved in varying degree. More obviously “IBN-reID+Direct Transfer” gains rank-1 accuracy in $32.4\%$ and mAP in 17.3 when tested on DukeMTMC-reID, surpassing the “Baseline+Direct Transfer” by +$2.2\%$ and +$1.2$, respectively. The similar improvement is obtained when tested on Market-1501. However, compared with “Baseline+StaGAN/IPGAN”, the models with IBN-reID, i.e., “IBN-reID+StarGAN/IPGAN”, get slight improvements in term of rank-1 accuracy. The main reason is that the *bias* between source and target camera domains is significantly reduced by StarGAN and IPGAN. Thus, in IBN-reID, the instance normalization only provide limited helps which eliminate appearance variance in shallow layers. This weaken the generalization capacity of IBN-reID. Even so, the mAP values have been significantly improved and the models get the best results. **Comparison with State-of-the-art Methods.** We compare the proposed method with the state-of-the-art unsupervised learning methods. Table \[table:4\] presents the comparison when tested on Market-1501 and DukeMTMC-reID. Firstly, we compare with two hand-crafted features: Bag-of-Words(BoW) [@zheng2015scalable] and local maximal occurrence (LOMO)[@liao2015person]. Their inferiority is obvious. Secondly, we compare the proposed methods with three unsupervised methods, including CAMEL [@yu2017cross], PUL [@fan2018unsupervised], and UMDL [@peng2016unsupervised]. Finally we also compare four unsupervised domain adaptation approaches, including PTGAN [@wei2017person], SPGAN(+LMP) [@deng2018image], TJ-AIDL [@wang2018transferable] and CamStyle [@zhong2018camstyle]. Comparing with those methods, when tested on Market-1501, The proposed method achieves rank-1 accuracy = $57.2\%$ and **the best mAP = 28.0**. When tested on DukeMTMC-reID, The result achieves rank-1 accuracy = $47.0\%$ and **the best mAP = 27.0**. Compared with CamStyle [@zhong2018camstyle], the proposed method achieves a similar rank-k accuracy and slight improvement in item of mAP. Especially when tested on DukeMTMC-reID, we gain $+1.9$ higher in mAP. The proposed method achieves a very competitive re-ID accuracy on “Market-1501 to Duke” and “DukeMTMC-reID to Market-1501”. Conclusion ========== In this paper, we propose the identity preserving generative adversarial network (IPGAN), a novel and efficient unsupervised domain adaptation method for person re-identification. IPGAN is used to translate images from source domain to target camera domains with identities preserving and thus we have a new training set. Then, we use the new training set to train re-ID model. Moreover, to obtain better performance, we propose IBN-reID model, which has better generalization ability than baseline model. Experiments on the Market-1501 and DukeMTMC-reID datasets show that we can achieve more competitive performance. In the feature, we will propose a new method to solve identity semantic constraint.
{ "pile_set_name": "ArXiv" }
--- abstract: 'M82 X-1 is one of the brightest ultraluminous X-ray sources (ULXs) known, which, assuming Eddington-limited accretion and other considerations, makes it one of the best intermediate-mass black hole (IMBH) candidates. However, the ULX may still be explained by super-Eddington accretion onto a stellar-remnant black hole. We present simultaneous [[*NuSTAR*]{}]{}, [[*Chandra*]{}]{} and [[*Swift*]{}/XRT]{} observations during the peak of a flaring episode with the aim of modeling the emission of M82 X-1 and yielding insights into its nature. We find that thin-accretion disk models all require accretion rates at or above the Eddington limit in order to reproduce the spectral shape, given a range of black hole masses and spins. Since at these high Eddington ratios the thin-disk model breaks down due to radial advection in the disk, we discard the results of the thin-disk models as unphysical. We find that the temperature profile as a function of disk radius ($T(r)\propto r^{-p}$) is significantly flatter ($p=0.55^{+ 0.07}_{- 0.04}$) than expected for a standard thin disk ($p=0.75$). A flatter profile is instead characteristic of a slim disk which is highly suggestive of super-Eddington accretion. Furthermore, radiation hydrodynamical simulations of super-Eddington accretion have shown that the predicted spectra of these systems are very similar to what we observe for M82 X-1. We therefore conclude that M82 X-1 is a super-Eddington accretor. Our mass estimates inferred from the inner disk radius imply a stellar-remnant black hole ([$M_{\rm BH}$]{}=$26^{+9}_{-6}$ [$M_{\odot}$]{}) when assuming zero spin, or an IMBH ([$M_{\rm BH}$]{}=$125^{+45}_{-30}$ [$M_{\odot}$]{}) when assuming maximal spin.' author: - 'Murray Brightman$^{1}$, Fiona A. Harrison$^{1}$, Didier Barret$^{2,3}$, Shane W. Davis$^{4}$, Felix F[ü]{}rst$^{1}$, Kristin K. Madsen$^{1}$, Matthew Middleton$^{5}$, Jon M. Miller$^{6}$, Daniel Stern$^{7}$, Lian Tao$^{1}$, Dominic J. Walton$^{7,1}$' bibliography: - 'bibdesk.bib' title: 'A broadband X-ray spectral study of the intermediate-mass black hole candidate M82 X-1 with NuSTAR, Chandra and Swift' --- Introduction ============ The ultraluminous X-ray source M82 X-1 is one of the best candidates for an intermediate-mass black hole ($100<M_{\rm BH}<10000$ [$M_{\odot}$]{}) based on several indirect factors. These include the source’s high luminosity, which can reach $\sim10^{41}$  [e.g. @ptak99b; @rephaeli02; @kaaret06], far greater than the Eddington limit of a stellar-remnant black hole of mass $\sim$10 [$M_{\odot}$]{} that is typical of X-ray binaries in our own Galaxy ($\sim10^{39}$ ); detection of low-frequency quasi-periodic oscillations (QPOs) in the power spectrum [54 mHz, @strohmayer03; @dewangan06; @mucciarelli06], indicative of a compact, unbeamed source; as well as twin-peaked QPOs at 3.3 and 5.1 Hz, which lead to a mass estimate using scaling laws between the QPOs frequencies and mass [@pasham14]. The mass estimates for X-1 vary considerably, however, and have large uncertainties. This makes its status as an IMBH is not yet firmly established. The most recent estimate came from the twin peak QPOs, which give a mass of 428$\pm$105 [$M_{\odot}$]{} [@pasham14]. On the other hand, modeling of the accretion disk emission by [@okajima06] instead found that the source can be explained by a $\sim30$ [$M_{\odot}$]{} stellar remnant black hole radiating at several times its Eddington limit. At moderate Eddington ratios ([$\lambda_{\rm Edd}$]{}$\equiv L/L_{\rm Edd}\ll1$), the accretion on to a black hole can be described by the standard “thin” accretion disk model [@shakura73 SS73]. For the standard disk model, the accretion disk is geometrically thin and optically thick where viscous heating in the disk is balanced by radiative cooling and the local temperature of the disk, $T$, decreases with radius, $r$ as $T(r)\propto r^{-0.75}$. Under the assumption that the disk extends down to the innermost stable circular orbit [ISCO, e.g. @steiner10], spectral modeling yields the temperature of the disk at the ISCO, which in turn yields the inner radius. The inner radius is directly proportional to the mass of the black hole, albeit with a large degeneracy with the black hole’s spin, which can be used for mass estimates. However, as the mass accretion rate increases, advective cooling dominates over radiative cooling and the thin-disk model breaks down. The scale height of the disk increases and thus is referred to as the “slim” disk model [@abramowicz88]. For a slim disk, the local temperature of the disk has a flatter temperature profile as a function of radius with $T(r)\propto r^{-0.5}$ [@watarai00]. Slim disks have been proposed as mechanisms to explain ULXs as super-Eddington stellar remnant black hole accretors rather than IMBHs [e.g. @kato98; @poutanen07]. In addition to the modified disk spectrum, the emission from super-Eddington accretion is expected to produce winds/outflows [@king03] which may also modify the emission spectrum [e.g. via Compton scattering, @kawashima09]. Indeed, high-velocity, ionized outflows have recently been detected in the high-resolution X-ray grating spectra of NGC 1313 X-1 and NGC 5408 X-1 [@pinto16] and confirmed in a follow-up study with CCD resolution data for NGC 1313 X-1 [@walton16]. Therefore, spectral modeling of the emission from ULXs and testing for a departure from the thin-disk model can yield important information regarding their nature. For M82 X-1, however, modeling of the disk emission has yielded conflicting results. [@feng10] observed the source with [[*XMM-Newton*]{}]{} and [[*Chandra*]{}]{} over the course of a flaring episode and fitted the spectra with the standard thin-disk model. They found that the luminosity of the disk, $L$, scaled with inner temperature as $L\propto T^4$ which is expected from a thin accretion disk with a constant inner radius. From this they inferred a black hole mass in the range $300-810$ [$M_{\odot}$]{}, assuming that the black hole is rapidly spinning in order to avoid extreme violations of the Eddington limit. However, using a different [[*XMM-Newton*]{}]{} dataset, [@okajima06] modeled the emission of the accretion disk from M82 X-1 instead finding that the temperature profile of the disk was too flat ($T(r)\propto r^{0.61}$) to be consistent with the standard thin-accretion disk and concluded that it was in the slim-disk condition. Applying a theoretical slim-disk model their estimate for the mass of the black hole was [$M_{\rm BH}$]{}$\approx19-32$[$M_{\odot}$]{}. Considering broader band data afforded by [[*Suzaku*]{}]{}/XIS and HXD-PIN allowed [@miyawaki09] (M09) to better distinguish between thin and slim-disk models. While M09 also consider a slim-disk conclusion based on a high inferred Eddington ratio, they instead prefer the power-law state interpretation, finding the spectrum to be too hard to be explained by emission from an optically thick accretion disk. These conflicting results may stem from the fact that spectral studies of X-1 are complicated by the presence of another ultraluminous X-ray source only 5 from X-1, which was recently identified as an ultraluminous X-ray pulsar [@bachetti14]. Since this source can reach luminosities of $10^{40}$  [@feng07; @kong07; @brightman16], and is only resolvable from X-1 with [[*Chandra*]{}]{}, its contribution to the X-ray spectrum of M82 must be taken into account when modeling the spectrum of X-1 with other X-ray instruments. Furthermore, X-1 and X-2 are embedded in bright diffuse emission [@ranalli08], which further complicates analysis. X-1 is also bright enough to cause pile-up effects on the [[*Chandra*]{}]{} detectors, which can severely distort the spectrum. In this paper we report on simultaneous observations of M82 with [[*NuSTAR*]{}]{}, [[*Chandra*]{}]{} and [[*Swift*]{}/XRT]{} during an episode of flaring activity from X-1. We aim to improve upon previous works with the combination of [[*Chandra*]{}]{} to spatially resolve X-1 from X-2 and the diffuse emission below 8 keV, and [[*NuSTAR*]{}]{} to gain broadband sensitive spectral coverage, especially above 10 keV. Our goal is to determine if the emission from the disk is indeed consistent with a standard thin-accretion disk, which would support the IMBH scenario, or if it shows a significant departure from this model that would indicate a super-Eddington accretor of lower mass. In section \[sec\_obs\] we describe our observations, including details of the [[*Swift*]{}/XRT]{} monitoring that showed the increased flux from M82, triggering our [[*NuSTAR*]{}]{} and [[*Chandra*]{}]{} Director’s Discretionary Time (DDT) requests, including details of the data reduction, while in section \[sec\_spec\] we describe our spectral analysis where we test various emission models for X-1. In section \[sec\_x1\] we describe the results from the disk models and the mass estimation of X-1. In section \[sec\_comp\] we discuss our results with respect to previous analyses and we finish with a discussion of alternative interpretations of the high-energy spectrum section \[sec\_alt\]. We conclude and summarize in section \[sec\_conc\]. A distance of 3.3 Mpc to M82 is assumed throughout [@foley14]. Observations and Data reduction {#sec_obs} =============================== The [[*Chandra*]{}]{} and [[*NuSTAR*]{}]{} observations were taken simultaneously, along with a [[*Swift*]{}]{}/XRT observation between 2015 June 20$-$21. Table \[table\_obsdat\] provides a description of the observational data. The following sections describe the individual observations and data reduction. Spectral fitting was carried out using [xspec]{} v12.8.2 [@arnaud96] and all uncertainties quoted are at the 90% level. -------------------- ------------- --------------------- --------------------- ------------ ---------- ------------- ------------------- Observatory ObsID Start date (UT) End date (UT) Instrument Exposure Energy band Count rate (ks) (keV) (counts s$^{-1}$) [[*NuSTAR*]{}]{}  90101005002 2015-06-20T14:21:07 2015-06-21T06:21:07 FPMA 37.4 3$-$30 0.88 FPMB 37.4 3$-$30 0.87 [[*Swift*]{}]{}  00081460001 2015-06-20T17:33:18 2015-06-20T19:21:55 XRT 1.6 0.5$-$10 0.77 [[*Chandra*]{}]{}  17678 2015-06-21T02:46:16 2015-06-21T06:17:39 ACIS-I 9.3 0.5$-$8 2.45 -------------------- ------------- --------------------- --------------------- ------------ ---------- ------------- ------------------- Swift/XRT --------- M82 has been monitored by [[*Swift*]{}/XRT]{} [@burrows05] every few days since 2012 (albeit with several gaps), including a period of intense monitoring of SN2014J at the beginning of 2014. Figure \[fig\_swift\_ltcrv\] shows a light curve of the 2$-$10 keV emission from the galaxy, which was produced from all 238 [[*Swift*]{}]{}/XRT observations of the galaxy taken since the beginning of 2012 until the beginning of 2016. We used the [heasoft]{} tool [xselect]{} to filter events from a 49 radius region centered on the peak of the emission and to extract the spectrum. Background events were extracted from a nearby circular region of the same size. The spectra were grouped with a minimum of 1 count per bin and the Cash statistic was used for spectral fitting. The spectra were fitted in the range 0.2$-$10 keV with an absorbed power-law. The luminosity in Figure \[fig\_swift\_ltcrv\] is the observed luminosity between 2$-$10 keV given a redshift of 0.00067. The observed flux in the 2$-$10 keV band is shown on the right axis. ![image](swift_M82_ltcrv.pdf){width="180mm"} As shown in Figure \[fig\_swift\_ltcrv\], between 2012 and early 2015 the observed 2$-$10 keV flux from the galaxy varied between 1$-$2$\times10^{-11}$ , implying a 2$-$10 keV luminosity of 1$-$2$\times10^{40}$ . On 2015 March 20 the flux from the galaxy increased by several factors to $>4\times10^{-11}$ , and continued to flare up to this level for several months. The [[*Swift*]{}/XRT]{} images pointed towards one of the ULXs as the origin of the brightening, however its spatial resolution and pointing accuracy did not allow us to pinpoint the emission. Chandra ------- Since the angular separation of X-1 and X-2 is only 5, only [[*Chandra*]{}]{} [@weisskopf99] can spatially resolve the emission from these two sources. We were therefore granted a DDT [[*Chandra*]{}]{} observation to ascertain the origin of the increased X-ray flux from M82. This observation was taken with ACIS-I at the optical axis with only a 1/8th sub-array of pixels on chip I3 turned on. M82 was placed 4 off-axis to smear out the PSF in order to mitigate the effect of pile-up from the bright point-sources. The sub-array of pixels was used to decrease the readout time of the detector, further mitigating the effect of pile-up. Figure \[fig\_chandra\_img\] shows the [[*Chandra*]{}]{} image of the central region of M82, which shows that the cause of the increased X-ray emission from M82 was X-1. We also indicate the other point-sources in this figure, including the ultraluminous pulsar (X-2). ![[[*Chandra*]{}]{} ACIS-I 0.5$-$8 keV image of M82 taken on 2015 July 21. The large green circle shows the [[*NuSTAR*]{}]{} and [[*Swift*]{}]{}/XRT extraction region which has a radius of 49. The brightest point-sources within this region are labelled. North is up, east is left, indicated by the arrows in the upper right corner, which are 10 long. []{data-label="fig_chandra_img"}](17678_img_labels.pdf){width="80mm"} We proceeded to extract the [[*Chandra*]{}]{} spectra of the point-sources using the [ciao]{} (v4.7, CALDB v4.6.5) tool [specextract]{}, from elliptical regions to encompass the shape of the off-axis [[*Chandra*]{}]{} PSF. For X-1 we used a semi-major axis of 3 and a semi-minor axis of 2. For X-2 we used 2 and 1 and for the other sources between 1$-$2. A small region close to the center of the image was used for background subtraction for X-1, X-2 and X-3. For the other sources, a background region outside the galaxy was used. Figure \[fig\_chandra\_spec\] shows the [[*Chandra*]{}]{} spectra of all these sources, including the spectrum of the diffuse emission within the central 49 of the galaxy. X-1 dominates the emission from the galaxy in the [[*Chandra*]{}]{} band with 1.29 , whereas the diffuse emission contributes 1.03 , X-2 contributes 0.11  and the other point sources contribute $\leq$0.01  individually. ![[[*Chandra*]{}]{} ACIS-I spectra of the seven brightest point-sources within the [[*NuSTAR*]{}]{} and [[*Swift*]{}]{}/XRT extraction regions (top) and the spectrum of the diffuse emission in M82 within the same region compared to the brightest source, X-1 (bottom). []{data-label="fig_chandra_spec"}](chandra_X7_data.pdf){width="90mm"} NuSTAR ------ In addition to the [[*Chandra*]{}]{} observation, we were granted a [[*NuSTAR*]{}]{} [@harrison13] DDT observation to obtain a source-dominated spectrum of M82 above 10 keV during a flaring event of X-1. The raw [[*NuSTAR*]{}]{} data were reduced using the [nustardas]{} software package version 1.4.1. The events were cleaned and filtered using the [nupipeline]{} script with standard parameters. The [nuproducts]{} task was used to generate the spectra and the corresponding response files. Spectra were extracted from a circular aperture of radius 49 centered on the peak of the emission shown in Figure \[fig\_chandra\_img\]. The background spectra were extracted from a region encompassing the same detector chip as the source, excluding the source extraction region and avoiding the wings of the PSF as much as possible. Data from both focal plane modules (FPMA and FPMB) are used for simultaneous fitting, without co-adding. Figure \[fig\_nustar\_spec\] shows the [[*NuSTAR*]{}]{} spectrum of M82 from this observation. The background is also plotted showing that the spectrum remains source-dominated up to $\sim$30 keV. For this reason, we limit our spectral analysis to below 30 keV. The [[*NuSTAR*]{}]{} spectrum can be described with an absorbed cutoff power-law ([zwabs\*cutoffpl]{}) where [$N_{\rm{H}}$]{}$=3.7\pm0.7\times10^{22}$ [cm$^{-2}$]{}, $\Gamma=1.74^{+0.15}_{-0.16}$ and $E_{\rm C}=8.89^{+1.30}_{-1.05}$ keV with a normalization of $2.894^{+0.005}_{-0.004}\times10^{-2}$, which is shown in Figure \[fig\_nustar\_spec\]. The absorption model [zwabs]{} uses Wisconsin [@morrison83] cross-sections and [@anders82] abundances. We leave the cross-normalization between the two FPMs to float, where $C_{\rm B}/C_{\rm A}=1.05\pm0.01$, which is close to typical values [@madsen15]. The flux in the $3-79$ keV band is 5.50$\times10^{-11}$ . ![[[*NuSTAR*]{}]{} 3$-$79 keV FPMA (red) and FPMB (green) spectra of M82 taken on 2015 July 20-21 (top) which are fitted with an absorbed cutoff power-law model. The backgrounds are plotted as black and grey histograms respectively. The data are source-dominated between 3$-$30 keV, above which the data are dominated by the background. The bottom panel shows the data-to-model ratio of the fit with the absorbed cutoff power-law model.[]{data-label="fig_nustar_spec"}](fit_nustar_zwabscutoffpl.pdf){width="90mm"} Spectral analysis and results {#sec_spec} ============================= Constraining the diffuse emission {#sec_chan_spec} --------------------------------- While X-1 dominates the X-ray emission from M82 during our observations, the diffuse emission and X-2 contribute significantly to the flux in the [[*Chandra*]{}]{} band within the [[*NuSTAR*]{}]{} and [[*Swift*]{}]{} extraction regions. Therefore in order to properly model the broadband emission from X-1, we must account for these other significant sources of emission. For our spectral analysis we model the fainter point sources as part of the diffuse emission. The diffuse emission from M82 was studied in depth by [@ranalli08] using deep [[*XMM-Newton*]{}]{} EPIC and RGS data. They find that this emission is best described by thermal plasma emission from hot gas with a double-peaked temperature, with peaks at $\sim0.5$ and $\sim7$ keV, for which they use the [apec]{} model in [xspec]{}. Following this we fit the diffuse emission with a combination of [apec]{} models, subjected to photo-electric absorption. We leave the abundances to vary for each component as well as the [$N_{\rm{H}}$]{}. We find that the $E<1$ keV spectrum requires two [apec]{} models with temperatures of $\sim0.3$ and $\sim1$ keV. We note that the temperature distribution of the lower energy peak in [@ranalli08] is broad, which is most likely why we find the need for two components rather than one. For the third high-energy component, we find that the temperature is not well constrained, so we fix this value at 7 keV. The spectral fit of the diffuse emission is shown in Figure \[fig\_chandra\_diffuse\] and the spectral parameters are given in Table \[table\_diffpar\]. The [$\chi^2$]{}/degrees of freedom (DOF) of this fit is 172.64/204=0.846. The 0.5$-$8 keV flux (luminosity) of the diffuse emission is 1.47$\times10^{-11}$  (1.91$\times10^{40}$ ) during this observation. ![[[*Chandra*]{}]{} 0.5$-$8 keV ACIS-I spectrum of the diffuse emission from M82 extracted from the same 49 region used for the [[*NuSTAR*]{}]{} and [[*Swift*]{}]{}/XRT spectral extraction, but with the emission from X-1 and X-2 masked out. The top panel shows the unfolded spectrum fitted with a combination of three [apec]{} models subjected to line-of-sight photo-electric absorption. The bottom panel shows the data-to-model ratio.[]{data-label="fig_chandra_diffuse"}](fit_diffuse_3apec.pdf){width="90mm"} [apec 1]{} [apec 2]{} [apec 3]{} ------------------------------------------- --------------------------- -------------------------- ------------------------ [$N_{\rm{H}}$]{} ($10^{22}$[cm$^{-2}$]{}) 1.45$^{+0.17}_{-0.10}$ 0.74$^{+0.13}_{-0.17}$ 9.28$^{+2.24}_{-1.76}$ $kT$ (keV) 1.01$\pm0.05$ 0.28$\pm0.03$ 7 (fixed) Adund. 0.78$^{+0.26}_{-0.57}$ 4.96$^{+u}_{-4.27}$ 0.04$^{+0.13}_{-0.04}$ norm . 0.024$^{+0.007}_{-0.006}$ 0.006$^{+0.04}_{-0.001}$ 0.013$\pm0.002$ : Spectral parameters of the diffuse emission[]{data-label="table_diffpar"} Accounting for emission from X-2 -------------------------------- A detailed spectral analysis of the [[*Chandra*]{}]{} spectra of X-2 was presented in [@brightman16] which found that these could be equally well described by absorbed power-law, cutoff power-law or disk blackbody models. While only [[*Chandra*]{}]{} can spatially resolve X-2 from nearby point-sources, its limited bandpass does not allow it to constrain broad band emission models for X-2. [[*NuSTAR*]{}]{} on the other hand cannot spatially resolve X-2. However, due to the coherent pulsations from the source and the timing capabilities of [[*NuSTAR*]{}]{}, the pulsed emission in the 3$-$79 keV band was measured. This pulsed spectrum is well described by an absorbed ([$N_{\rm{H}}$]{} fixed to $3\times10^{22}$ [cm$^{-2}$]{}) cutoff power-law ([cutoffpl]{}), where $\Gamma=0.6\pm0.3$ and $E_{\rm C}=14^{+5}_{-3}$ keV. The pulsed spectrum compares well with the phase-averaged [[*Chandra*]{}]{} spectra using the same model, which yields $\Gamma=0.70^{+0.68}_{-0.65}$ and $E_{\rm C}=6.19^{+50.9}_{-2.9}$ keV. Therefore in our present analysis, we model the emission from X-2 with the [cutoffpl]{} model. From our new [[*Chandra*]{}]{} data on X-2, we measure [$N_{\rm{H}}$]{}=2.92$^{+0.73}_{-0.85}\times10^{22}$ [cm$^{-2}$]{}, $\Gamma=1.16^{+0.44}_{-1.70}$ and an unconstrained cutoff energy. The 0.5$-$8 keV flux (luminosity) from X-2 was 5.94$\times10^{-12}$  (7.73$\times10^{39}$ ) during our observation. Combined spectral analysis of X-1 --------------------------------- Finally, for X-1, we find that the [[*Chandra*]{}]{} spectra are heavily piled-up with a pile-up fraction of $>10$%, despite our efforts to mitigate this with a sub-array of pixels and by locating the source off axis. For this reason we opt not to use the [[*Chandra*]{}]{} data on X-1 and instead infer the spectral properties of X-1 from the [[*Swift*]{}]{}/XRT and [[*NuSTAR*]{}]{} data, having accounted for the diffuse emission and X-2 with modeling of the [[*Chandra*]{}]{} data. For our broadband spectral study of X-1 we conduct simultaneous fitting of the combined [[*Chandra*]{}]{}, [[*NuSTAR*]{}]{} and [[*Swift*]{}/XRT]{} datasets. In Section \[sec\_chan\_spec\] we described our spectral fitting of the diffuse emission within the [[*NuSTAR*]{}]{} and [[*Swift*]{}/XRT]{} extraction regions. We fix all the spectral parameters of this component in our joint fit. For X-2, modeled with a cutoff power-law, we allow the spectral parameters to vary in the fit, since X-2 is expected to contribute significantly above 8 keV and be constrained by the [[*NuSTAR*]{}]{} data. For X-1 we test a variety of emission models. We use five spectral datasets in our joint fit, the [[*Chandra*]{}]{} 0.5$-$8 keV spectrum of the diffuse emission, the [[*Chandra*]{}]{} 0.5$-$8 keV spectrum of X-2, the [[*Swift*]{}/XRT]{} 0.5$-$10 keV spectrum of the integrated emission from M82 and the [[*NuSTAR*]{}]{} FPMA and FPMB 3$-$79 keV spectrum of the integrated emission from M82. For each dataset we allow a cross-normalization constant to vary to allow for instrument cross-calibration uncertainties that cannot vary beyond $\pm10$% of the FPMA normalization [@madsen15]. Results from the disk models for X-1 {#sec_x1} ==================================== The main goal of this paper is to test various disk emission models for X-1 and to assess if the emission is consistent with the thin-disk model associated with sub-Eddington accreting black holes, or whether a significant departure from this is seen which could imply super-Eddington accretion. We begin by testing the simplest thin-disk model [diskbb]{} (Figure \[fig\_comb\_spec2\], top left panel), which describes the emission from an accretion disk with multiple blackbody components with a temperature, $T$, that varies with the radius, $r$, as $T(r)\propto r^{-0.75}$. The parameters of the model are the temperature at the inner disk radius and the normalization, which is related to the inner disk radius as $((R_{\rm in}/$km$)/(D/10 $kpc$))^2 \times $cos$\theta$, where $D$ is the distance to the source and $\theta$ is the inclination angle of the disk. The fit with this model yields a statistically acceptable fit with [$\chi^2$]{}/DOF=972.8/964, where the inner temperature of the disk is constrained to be $T_{\rm in}=1.89^{+ 0.10}_{- 0.09}$ keV. As noted by several previous authors, this temperature is hotter than expected for an accretion disk around an IMBH [e.g. @strohmayer03]. Following this, we progress to more sophisticated models, specifically those that include the black hole mass and spin as parameters and compute the Eddington ratio in a self-consistent manner. The first of these is [kerrbb]{} [@li05], which specifically takes into account general relativistic effects due to a spinning Kerr black hole (Figure \[fig\_comb\_spec2\], bottom left panel). The model accounts for self-irradiation of the disk and limb darkening, which can be switched on or off. The torque at the inner boundary of the disk is also allowed to be zero. We switch on self-irradiation and limb darkening and set the torque at the inner boundary to zero. The color correction factor, $\kappa$, which accounts for the deviation of the local disk spectrum from a black body due to electron scattering, is set to 1.7 [@davis11]. The fit with this model provides an improved fit over [diskbb]{} with [$\chi^2$]{}/DOF=961.1/962 and provides a constraint on the black hole mass of 80$^{+72}_{-60}$ [$M_{\odot}$]{} (where the spin is unconstrained between $a_{*}=0.99$ and $a_{*}=-0.99$) assuming that the disk extends to the ISCO. However, the model implies that M82 X-1 is shining at a super-Eddington rate, [$\lambda_{\rm Edd}$]{}=17$^{+106}_{-1}$. [bhspec]{} [@davis05; @davis06] is similar to [kerrbb]{} in that it is a fully relativistic accretion disk model, but differs from [kerbb]{} by the treatment of the emission at the surface of the disk, where [bhspec]{} does not assume the spectrum at the surface to be black body, but instead uses stellar-like atmosphere calculations to model the vertical structure of the disk (Figure \[fig\_comb\_spec2\], bottom right panel). We use a version of this model with the parameter ranges [$M_{\rm BH}$]{}$=100-10^4$ [$M_{\odot}$]{}, [$\lambda_{\rm Edd}$]{}$=0.03-1$ and $a_{*}=0-0.99$. Since [bhspec]{}, which is a grid model, does not allow for [$\lambda_{\rm Edd}$]{}$>1$, the results from this model are different from [kerrbb]{}. A fit with this model gives [$\chi^2$]{}/DOF=1002.6/962 and implies [$M_{\rm BH}$]{}$=950^{+ 210}_{- 120}$ [$M_{\odot}$]{}, where the spin and Eddington ratios hit their upper limits, $a_{*}=0.9900^{+u}_{-0.0004}$ and [$\lambda_{\rm Edd}$]{}=1.00$^{+ u}_{- 0.01}$. By effectively assuming Eddington-limited accretion, the model is forced to higher black hole masses in order increase the flux, while simultaneously being forced to high spins to keep the temperature of the disk high. By being forced to its parameter limits, the results from [bhspec]{} point towards super-Eddington accretion as does [kerrbb]{}. However, at close-to-Eddington ratios and above, the standard thin disk that these models assume does not hold, thus we discard them as unphysical in this regime and do not rely on their black hole mass estimates. Finally, the thermal emission from X-1 has previously been shown to be best described by a slim disk, rather than a thin disk [e.g. @okajima06]. At high Eddington ratios, the standard thin disk is expected to transition to a slim disk which is dominated by advection [@abramowicz88] and has a different temperature profile as a function of radius than a thin disk does, where the temperature of the disk, $T$, at a given radius, $r$, is described as $T(r)\propto r^{-0.5}$ [@watarai00]. We test if M82 X-1 is consistent with the slim disk prediction using the simple [diskpbb]{} model, which is similar to [diskbb]{}, but with a variable exponent to the temperature profile, $p$ (i.e. $T(r)\propto r^{-p}$, Figure \[fig\_comb\_spec2\], top right panel). This model provides the best fit of all the disk models we have tested with [$\chi^2$]{}/DOF=956.6/963. The derived parameters are $T_{\rm in}=2.36^{+ 0.29}_{- 0.30}$ keV and $p=0.55^{+ 0.07}_{- 0.04}$. Allowing a variable $p$ leads to a significant improvement in the fit statistic over a fixed $p$ of 0.75 ([$\Delta\chi^2$]{}$=-16$). The derived $p$ value is significantly lower than for a thin disk, and consistent with a slim disk, in confirmation of previous results (e.g. M09). While the slim disk state indicated by the data imply a high-Eddington ratio, the [diskpbb]{} model we use to model the spectrum is in fact a simplification since at high Eddington rates, outflows, geometric collimation and other effects are also expected to influence the observed emission. Nonetheless, radiation hydrodynamics simulations of super-Eddington accretion have been carried out [e.g. @kawashima12; @jiang14] and produce spectra that are roughly consistent with the spectral shape of slim-disk models. For example, [@kawashima12] compare their predicted spectra with the shape of the observed spectrum of several ULXs. They find that for face-on geometries, their results compare very well to the spectrum of NGC 1313 X-2, which is also well fitted with a [diskpbb]{} model with $p\approx$0.5 [@gladstone09; @bachetti13]. A black hole mass estimate can also be made from slim-disk models, which is inferred from the inner radius measured by the model. To estimate the mass this way, we follow the prescription of [@soria15], who estimated the mass of M83 ULX-1, which was also found to be in the slim-disk regime. The inner radius of the disk is related to the normalization of the model: $R_{\rm in}=\xi\kappa^2N^{1/2}($cos$\theta)^{-1/2}(D/10~{\rm kpc})$ km, where $\xi$ is a geometric factor and depends on how close to the ISCO the disk reaches its maximum temperature, $\kappa$ is the color correction factor, $N$ is the normalization of the model, and $\theta$ is the inclination of the disk. The mass of the black hole is related to $R_{\rm in}$ by [$M_{\rm BH}$]{}$=1.2R_{\rm in}c^2/G\alpha$, where $\alpha=1.24$ for a maximally spinning black hole in an astrophysical context [$a_*=0.998$, @thorne74] or $\alpha=6$ for a black hole with no spin. The factor of 1.2 results from the fact that the inner radius of a slim disk extends within the ISCO [@vierdayanti08]. While for standard thin-disks, $\xi=0.412$ [@kubota98] and $\kappa=1.7$ [@shimura95; @davis05], [@soria15] note that at high Eddington rates, $\kappa$ increases to 3 [e.g. @watarai03] and $\xi=0.353$ [@vierdayanti08], which we adopt here. Given the normalization of the [diskpbb]{} model, $N=0.033^{+0.028}_{-0.014}$, the distance of 3.3 Mpc to M82 and assuming a face-on inclination of the disk ($\theta=0$[[$^\circ$]{}]{}), the formula above yields [$M_{\rm BH}$]{}=$26^{+9}_{-6}$ [$M_{\odot}$]{} for a non-spinning black hole, or [$M_{\rm BH}$]{}=$125^{+45}_{-30}$ [$M_{\odot}$]{} for a maximally spinning black hole. The luminosity of the disk from the [diskpbb]{} model is 5.11$^{+ 0.42}_{- 0.30}\times10^{40}$ , which corresponds to [$\lambda_{\rm Edd}$]{}=14$^{+5}_{-3}$ for a non-spinning black hole, or [$\lambda_{\rm Edd}$]{}=3$\pm1$ for a maximally spinning black hole. The black hole mass estimates are also degenerate on the inclination angle assumed. Given an extreme inclination of 85[[$^\circ$]{}]{}, the formula above yields [$M_{\rm BH}$]{}=$88^{+30}_{-22}$ [$M_{\odot}$]{} for a non-spinning black hole, or [$M_{\rm BH}$]{}=$424^{+149}_{-104}$ [$M_{\odot}$]{} for a maximally spinning black hole. which corresponds to [$\lambda_{\rm Edd}$]{}=0.9$^{+0.3}_{-0.3}$ for a non-spinning black hole, or [$\lambda_{\rm Edd}$]{}=3$\pm1$ for a maximally spinning black hole. We list the best fit spectral parameters of the models above in Table \[table\_specpar3\] and show the fitted spectra and data-to-model ratios in Figure \[fig\_comb\_spec2\]. ![image](combined_spectra3.pdf){width="160mm"} ![image](combined_ratios3.pdf){width="160mm"} [l l l l l]{} Comparison with previous results {#sec_comp} ================================ As described above, we found that the spectrum of M82 X-1 requires a departure from the standard disk model, with a $p$-value of 0.55$^{+0.07}_{-0.04}$, which is significantly less than the $p=0.75$ that reproduces the standard disk, indicating that advection is significant in the disk. This was found similarly by M09, although with larger $p$-values (0.61-0.65). The best-fit temperature of the [diskpbb]{} model is 2.36$^{+0.29}_{-0.30}$ keV, which is not as hot as what M09 find (3.4-3.6 keV). It should be noted that the luminosity of X-1 during our observations was a factor of $2-4$ higher than when it was observed by [[*Suzaku*]{}]{}, thus our results may not be strictly comparable. Interestingly, an increase in the disk temperature with decreasing luminosity is predicted by the model of supercritically accreting black holes of [@poutanen07]. However, M09 did not benefit from simultaneous spatially resolved [[*Chandra*]{}]{} data and they estimate the contribution from X-2 from previous [[*Chandra*]{}]{} observations, leading to systematic uncertainties in their best-fit parameters. Similarly, [@okajima06] fitted a long [[*XMM-Newton*]{}]{} exposure of M82 with the [diskpbb]{} model, which gave $p=0.61^{+0.03}_{-0.02}$ and a temperature of 3.73$^{+0.58}_{-0.40}$ keV. However, these authors neglected contributions to the [[*XMM-Newton*]{}]{} spectrum by X-2 and the diffuse emission. In addition to the [diskpbb]{} model, [@okajima06] fit their [[*XMM-Newton*]{}]{} spectrum of M82 with their own slim-disk model, deriving a black hole mass between 19$-$32 [$M_{\odot}$]{}, depending on the physical processes assumed (e.g. blackbody emission, Comptonisation, gravitational redshift, relativistic effects). This is consistent with our mass estimate of 26 [$M_{\odot}$]{} for a non-spinning black hole. [@okajima06] do not consider a spinning black hole. [@feng10] also observed X-1 during a period of high flux with [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{}. They fit the spectrum with the [diskbb]{} model, obtaining disk temperatures from 1.10 keV to 1.52 keV, depending on the luminosity which ranged from $1.8-7.9\times10^{40}$ . Our fit with this model yielded a temperature of 1.89$^{+0.10}_{-0.09}$ keV and a disk luminosity of 4.78$^{+0.26}_{-0.12}\times10^{40}$ , a somewhat higher temperature than the range found by [@feng10], despite a comparable luminosity. [@feng10] also applied the [kerrbb]{} model to their combined [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{} dataset in order to estimate the black hole mass. They test two different scenarios for the spin of the black hole, $a_{*}$=0 and $a_{*}$=0.9986. They rule out a non-rotating black hole due to the model yielding an Eddington ratio exceeding the limit by a factor of 160. Their maximally spinning scenario yields a black hole mass in the range 300$-$810 [$M_{\odot}$]{} with a disk inclination of 59$-$79 degrees (90% confidence) and an Eddington ratio of 2.5. These authors did not however consider a slim disk. Furthermore, they relied on a spectral model to account for pile-up, which introduces uncertainties into the spectral parameters and luminosity estimations, and performed their fits in a relatively narrow energy band with no coverage above 10 keV. Pile-up does not affect the [[*NuSTAR*]{}]{} data at the observed count rates. Our results put the mass of M82 X-1 at the lower end of previous estimates, and are consistent with a stellar remnant black hole without spin, or a borderline IMBH with maximal spin. Interestingly the mass range is consistent with the masses of the merging black holes discovered through the detection of their gravitational wave signal [@abbott16]. Our mass estimates are considerably lower than those from thin-disk modeling [@feng10] or QPO analysis [@pasham14]. A subsequent analysis of the QPOs by [@stuchlik15] found that the black hole mass estimate from these are strongly model dependent, and given this put a range on the mass as $140<$[$M_{\rm BH}$]{}$<660$ [$M_{\odot}$]{}, where the lower masses correspond to a non-spinning black hole, and as such are at odds with our result. In our estimation, we have assumed a face-on inclination of the disk. Larger inclinations would increase the mass estimate by (cos$\theta$)$^{-1/2}$, but even an extreme inclination value, such as $\theta=85$[[$^\circ$]{}]{}, only increases the non-spinning estimate to [$M_{\rm BH}$]{}=$62^{+22}_{-15}$ [$M_{\odot}$]{}. Alternative interpretations of the high-energy spectrum {#sec_alt} ======================================================= For all of the continuum models we have fitted for X-1, the spectrum turns down above 10 keV, however a significant signal remains, which we have modeled with a cutoff power-law representing emission from X-2. However, since the spatial resolution of [[*NuSTAR*]{}]{} does not allow us to determine directly that X-2 is indeed the source of the emission, we consider other scenarios for its origin. For other ULXs, an additional hard component has been found in excess the disk component. This has been interpreted as Compton-scattered disk emission by a corona [e.g. @gladstone09; @walton15]. We therefore test if the spectral shape above 10 keV, can be explained by such a component for X-1. For this we add the [simpl]{} model [@steiner09], which takes a fraction of the seed disk model photons (we test with our best-fitting [diskpbb]{} model) and up-scatters them into a power-law component. The parameters of this model are the power-law index ($\Gamma$) and the scattered fraction ($f_{\rm scatt}$). We find that the addition of this component does not change the fit since the hard emission is already being accounted for by the [cutoffpl]{} model for X-2. To test to what extent the [simpl]{} model can account for this instead, we fix the parameters of the [cutoffpl]{} model to the best-fit parameters found by the long [[*Chandra*]{}]{} exposure in [@brightman16] ($\Gamma=0.70$ and $E_{\rm C}=6.19$ keV), since these parameters lead to a softer spectrum for X-2. In this case the best fit [simpl]{} parameters are $\Gamma=2.37^{+ 0.44}_{- 0.43}$ and $f_{\rm scatt}=0.15^{+ 0.09}_{- 0.04}$ with [$\chi^2$]{}/DOF=974.7/963. In comparison, for Holmberg II X-1 [@walton15] found $\Gamma=3.1^{+ 0.3}_{- 1.2}$ and $f_{\rm scatt}=0.4^{+ 0.5}_{- 0.3}$, which are similar to what we find for M82 X-1 considering the uncertainties. The addition of the [simpl]{} model also leads to changes in the [diskpbb]{} model parameters which become $T_{\rm in}=2.01^{+ 0.13}_{- 0.19}$ keV, $p=0.59^{+ 0.06}_{- 0.04}$ and $N=0.075^{+0.028}_{-0.014}$. The temperature and $p$-value are within the statistical uncertainties of the previous fit with [diskpbb]{}, however the normalization has increased. This in turn leads to an increase in the black hole mass estimate, which given the same assumptions as above becomes [$M_{\rm BH}$]{}=$39^{+14}_{-8}$ [$M_{\odot}$]{} for a non-spinning black hole, or [$M_{\rm BH}$]{}=$188^{+70}_{-38}$ [$M_{\odot}$]{} for a maximally spinning black hole. These estimates are not significantly different from our initial estimates. Another alternative source of the high-energy signal is the hot diffuse gas in the center of M82 in which the ULXs are embedded. We have modeled this with three [apec]{} models with temperatures of 0.28, 1.01 and 7 keV. For the first two components, the temperatures were constrained by the [[*Chandra*]{}]{} data, however the highest temperature component could not be constrained, and was fixed at 7 keV based on previous results by [@ranalli08]. We now explore if this hot [apec]{} component can account for the high-energy signal by freeing the temperature. For this we keep the spectrum of X-2 fixed as described above. We find that the signal can indeed be described by a hot [apec]{} model where $kT=13.2^{+2.4}_{-1.0}$ keV with [$\chi^2$]{}/DOF=963.31/962. While [@ranalli08] find that their temperature profile peaks at $\sim7$ keV, their data and/or model do not allow for exploration of temperatures $>10$ keV, thus the temperature that we found may be consistent with their results. We note that [@cappi99] also found evidence for a hot thermal component in M82 from [*BeppoSAX*]{} data, with a temperature of $\sim5-8$ keV. The origin of this emission was discussed by [@ranalli08], including non-thermal bremsstrahlung emission from star formation and unresolved point sources, however these were ruled out and no clear conclusion was reached. For this scenario, the parameters of the [diskpbb]{} model for X-1 become $T_{\rm in}=2.21^{+ 0.20}_{- 0.15}$ keV, $p=0.55^{+ 0.06}_{- 0.04}$ and $N=0.042^{+0.032}_{-0.017}$. These new parameters imply [$M_{\rm BH}$]{}=$29^{+10}_{-7}$ [$M_{\odot}$]{} for a non-spinning black hole, or [$M_{\rm BH}$]{}=$140^{+46}_{-32}$ [$M_{\odot}$]{} for a maximally spinning black hole. Again these parameters are very similar to those assuming that the high-energy emission originates from X-2, and therefore the assumption regarding the origin of this emission does not significantly change our results regarding X-1. Summary and Conclusions {#sec_conc} ======================= In this paper we have presented analysis of simultaneous [[*NuSTAR*]{}]{}, [[*Chandra*]{}]{} and [[*Swift*]{}/XRT]{} observations of the ultraluminous X-ray source M82 X-1 during a period of flaring activity. The [[*Chandra*]{}]{} data have allowed us to spatially resolve the source from the other bright sources of X-rays in the galaxy, specifically that of the nearby ultraluminous pulsar, X-2, and the bright diffuse emission. Combined with [[*NuSTAR*]{}]{} and [[*Swift*]{}/XRT]{} data, this provides a sensitive measurement of the 0.5$-$30 keV spectrum of the source. We have fitted standard thin accretion disk models for sub-Eddington accretion to the spectrum finding that they require super-Eddington accretion rates in order to reproduce the observed spectrum. Since the thin accretion disk models do not hold at high Eddington ratios, we discard the thin-disk models as unphysical. We directly test for the departure from the thin-disk model using a disk model that allows for a variable temperature as a function of radius of the disk ([diskpbb]{}), finding that the temperature profile is ($T(r)\propto r^{-0.55}$), which is significantly flatter than expected for a thin disk, and is instead characteristic of a slim disk, which is expected at high Eddington ratios. While at high Eddington rates, outflows and geometric collimation are also expected to influence the observed emission, which our simple model does not account for, radiation hydrodynamics simulations of super-Eddington accretion have shown that the predicted spectra are very similar to what we observe for M82 X-1. We therefore conclude that the ULX is a super-Eddington accretor. Our mass estimates inferred from the inner disk radius imply a stellar-remnant black hole ([$M_{\rm BH}$]{}=$26^{+9}_{-6}$ [$M_{\odot}$]{}) when assuming zero spin, or an IMBH ([$M_{\rm BH}$]{}=$125^{+45}_{-30}$ [$M_{\odot}$]{}) when assuming maximal spin. Acknowledgements ================ This work made use of Director’s Discretionary Time on [[*Chandra*]{}]{}, for which we thank Belinda Wilkes for approving and the [[*Chandra*]{}]{} X-ray Center for implementing. We also use Director’s Discretionary Time on [[*NuSTAR*]{}]{}, for which we thank Fiona Harrison for approving and the [[*NuSTAR*]{}]{} SOC for co-ordinating with [[*Chandra*]{}]{}. The [[*NuSTAR*]{}]{} mission is a project led by the California Institute of Technology, managed by the Jet Propulsion Laboratory, and funded by NASA. We thank the [[*NuSTAR*]{}]{} Operations, Software and Calibration teams for support with the execution and analysis of these observations. This research has made use of the [[*NuSTAR*]{}]{} Data Analysis Software (NuSTARDAS) jointly developed by the ASI Science Data Center (ASDC, Italy) and the California Institute of Technology (USA). AZ acknowledges funding from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013)/ERC Grant Agreement n. 617001. [*Facilities:*]{}
{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the problem of finding an $\epsilon$-approximate Nash equilibrium in an [anonymous]{} game with seven pure strategies is complete in PPAD, when the approximation parameter $\epsilon$ is exponentially small in the number of players.' author: - | Xi Chen\ Columbia University - | David Durfee\ Georgia Institute of Technology - | Anthi Orfanou\ Columbia University bibliography: - 'citations.bib' - 'Reference.bib' title: On the Complexity of Nash Equilibria in Anonymous Games --- Introduction ============ The celebrated theorem of Nash [@NASH50; @NASH51] states that every game has an equilibrium point. The concept of Nash equilibrium has been tremendously influential in economics and social sciences ever since (e.g., see [@NashSocial]), and its computation has been one the most well-studied problems in the area of Algorithmic Game Theory. For [normal form games]{} with a bounded number of players, much progress has been made during the past decade in understanding both the complexity of Nash equilibrium [@AbbottKaneValiant; @ChenDengTengSparse; @ChenTengValiant; @DGPJournal; @2Nash; @Mehta14] as well as its efficient approximation . In this paper we study a large and important class of *succinct multiplayer* games called *anonymous games* (see [@sch73; @Mil96; @Blo99; @Blo05; @Kal05] for studies of such games in the economics literature). These are special multiplayer games in that the payoff of each player depends only on (1) the pure strategy of the player herself, and (2) the *number* of other players playing each pure strategy, instead of the full pure strategy profile. In such a game, the (expected) payoff of a player is *highly symmetric* over (pure or mixed) strategies of other players. For instance, two players switching their strategies would not affect the payoff of any other player. A consequence of this very special payoff structure is that $O(\alpha n^{\alpha-1})$ numbers suffice to completely describe the payoff function of a player, when there are $\alpha$ pure strategies shared by $n$ players. Notably this is polynomial in the number of players when $\alpha$ is bounded, and hence the game is [succinctly representable]{}. *Throughout the paper, we focus on succinct anonymous games with a bounded number of pure strategies.* Other well-studied multiplayer games with a succinct representation include graphical, symmetric, and congestion games (for more details see [@PapRough08]). While graphical and congestion games are both known to be hard to solve [@fabrikant04; @AckermannPLS; @Skopalik], there is indeed a polynomial-time algorithm for computing an exact Nash equilibrium in a symmetric game [@PapRough08]. Because anonymous games generalize symmetric games by allowing player-dependent payoff functions, it is a natural question to ask whether there is an efficient algorithm for finding an (exact or approximate) Nash equilibrium in an anonymous game as well. Culminating in a sequence of beautiful papers [@PTAS07; @2PTAS08; @cPTAS08; @obliv; @dask14] Daskalakis and Papadimitriou obtained a polynomial-time approximation scheme (PTAS) for $\epsilon$-approximate Nash equilibria in anonymous games with a bounded number of strategies (see more discussion on related work in Section \[sec:work\]). However, the complexity of finding an exact Nash equilibrium in such games remains open, and was conjectured to be hard for PPAD in [@obliv; @dask14].[^1] In this paper we give an affirmative answer to the conjecture of Daskalakis and Papadimitriou, by showing that it is PPAD-complete to find an $\epsilon$-approximate Nash equilibrium in an anonymous game, when the approximation parameter $\epsilon$ is exponentially small in $n$. To formally state our main result, let $(\alpha,c)$- denote the problem of finding a $(2^{-n^c})$-approximate Nash equilibrium in an anonymous game with $\alpha$ pure strategies and payoffs from $[0,1]$.[^2] Here is our main theorem: \[main-theorem\] For any $\alpha\ge 7$ and $c>0$, the problem $(\alpha,c)$- is PPAD-complete. *The greatest challenge to establishing the PPAD-completeness result stated above is posed by the rather complex but also highly symmetric payoff structure of anonymous games.* Before discussing our approach and techniques in Section \[sec:ours\], we first review related work in Section \[sec:work\], then define anonymous games formally and introduce some useful notation in Section \[sec:pre\]. Related Work {#sec:work} ------------ Anonymous games have been studied extensively in the economics literature [@sch73; @Rashid83; @Mil96; @Blo99; @Blo05; @Kal05; @Sandholm05], where the game being considered is usually nonatomic and consists of a continuum of players but a finite number of strategies. For the discrete setting, two special families of anonymous games are symmetric games [@PapRough08; @Bra09] and congestion games [@ros]. [@PapRough08] gave a polynomial-time for finding an exact Nash equilibrium in a symmetric game. For congestion games, PLS-completeness of pure equilibria was established in [@fabrikant04; @AckermannPLS; @Skopalik][^3], and efficient approximation algorithms for various latency functions were obtained in [@CFGS1; @CFGS2; @Chien07]. While an anonymous game does not possess a pure Nash equilibrium in general, it was shown in [@PTAS07; @Azrieli; @dask14] that when the payoff functions are $\lambda$-Lipschitz, there exists an $\epsilon$-approximate pure Nash equilibrium and it can be found in polynomial time, where $\epsilon$ has a linear dependency on $\lambda$. Furthermore, in [@BRE12] Babichenko presented a best-reply dynamic for $\lambda$-Lipschitz anonymous games with two strategies which reaches an approximate pure equilibrium in $O(n\log n)$ steps. Regarding our specific point of interest, i.e., (mixed) Nash equilibria in anonymous games with a scaling number of players but a non-scaling number of strategies, there have been a sequence of positive and negative results obtained by Daskalakis and Papadimitriou [@PTAS07; @cPTAS08; @2PTAS08; @obliv] (summarized in the journal version [@dask14]). We briefly review these results below. In [@PTAS07], Daskalakis and Papadimitriou presented a PTAS for finding an $\epsilon$-approximate Nash equilibrium in an anonymous game with two pure strategies, with running time $n^{O(1/\epsilon^2)}\cdot U$, where $U$ denotes the number of bits required to describe the payoffs. The running time was subsequently improved in [@2PTAS08] to $\text{poly}(n)\cdot (1/\epsilon)^{O(1/\epsilon^2)}\cdot U$. The first PTAS in [@PTAS07] is based on the existence of an $\epsilon$-approximate Nash equilibrium consisting of integer multiples of $\epsilon^2$, while the second PTAS in [@2PTAS08] is based on the existence of an $\epsilon$-approximate Nash equilibrium satisfying the following special property: either at most $O(1/\epsilon^3)$ players play mixed strategies, or all players who mix play the same mixed strategy. Later [@cPTAS08] extended the result of [@PTAS07], giving the only known PTAS for anonymous games with any bounded number of pure strategies with time $n^{g(\alpha,1/\epsilon)}\cdot U$ for some function $g$ of $\alpha$, number of pure strategies, and $1/\epsilon$. All three PTAS obtained in [@PTAS07; @2PTAS08; @cPTAS08] are so-called *oblivious* algorithms [@obliv], i.e., algorithms that enumerate a set of mixed strategy profiles that is independent of the input game as candidates for approximate Nash equilibria (hence, the game is used only to verify if a given mixed strategy profile is an $\epsilon$-approximate Nash equilibrium). In [@obliv], Daskalakis and Papadimitriou showed that any oblivious algorithm for anonymous games must have running time exponential in $1/\epsilon$. In contrast to this negative result, they also presented a [*non-oblivious*]{} PTAS for two-strategy anonymous games with running time $\text{poly}(n)\cdot (1/\epsilon)^{O(\log^2 (1/\epsilon))} \cdot U$. Anonymous Games and Polymatrix Games {#sec:pre} ------------------------------------ Before giving a high-level description of our approach and techniques in Section \[sec:ours\], we first give a formal definition of anonymous games and introduce some useful notation. Consider a multiplayer game with $n$ players $[n]=\{1,\ldots,n\}$ and $\alpha$ pure strategies $[\alpha]=\{1,\ldots,\alpha\}$ with $\alpha$ being a constant. For each pure strategy $b\in [\alpha]$, let $\psi_b(\tt)$ denote the number of $b$’s in a tuple $\tt\in [\alpha]^{n-1}$, and define $\Psi(\tt) = (\psi_{1}(\tt),\ldots, \psi_{\alpha}(\tt)),$ which we will refer to as the *histogram* of pure strategies in $\tt$. In an anonymous game, the payoff of each player $p\in [n]$ depends only on $\Psi(\ss_{-p})$ and her own strategy $s_p$, given a pure strategy profile $\ss\in [\alpha]^n$. (We follow the convention and use $\ss_{-p}\in [\alpha]^{n-1}$ to denote the pure strategy profile of the $n-1$ players other than player $p$ in $\ss$.) Informally, $\Psi(\ss_{-p})$ can be described as what player $p$ “sees” in the game when $\ss$ is played. We now formally define [anonymous games]{}. An [anonymous game]{} $\calG = (n,\alpha,\{\textsf{\emph{payoff}}_p\})$ consists of a set $[n]$ of $n$ players, a set $[\alpha]$ of $\alpha$ pure strategies, and a payoff function $\emph{\pay}_p : [\alpha]\times \k \rightarrow \mathbb{R}$ for each player $p\in [n]$, where $$\begin{aligned} &\k = \big\{(k_1,\ldots,k_\alpha): k_j \in \mathbb{Z}_{\ge 0}\text{ for all $j$ and }\sum_{j=1}^\alpha k_j = n-1 \big\}&\end{aligned}$$ is the set of all histograms of pure strategies played by $n-1$ players. Specifically, when $\ss\in [\alpha]^{n}$ is played, the payoff of player $p$ is given by $\textsf{\emph{payoff}}_p(s_p,\Psi(\ss_{-p}))$. As usual, a [mixed strategy]{} is a probability distribution $\xx=(x_{1},\ldots,x_{\alpha})$, and a [mixed strategy profile]{} $\bxx$ is an ordered tuple of $n$ mixed strategies $(\xx_p:p\in [n])$, one for each player $p$. Given $\bxx$, let $u_p(b,\bxx)$ denote the *expected payoff* of $p$ playing $b\in [\alpha]$, which has the following explicit expression: $$u_p(b,\bxx) = \sum_{\kk \in \k}{\pay}_p(b,\kk) \cdot \Pr_{\bxx}[p,\kk],$$ where $\Pr_{\bxx}[p,\kk]$ denotes the probability of player $p$ seeing histogram $\kk$ under $\bxx$: $$\Pr_{\bxx}[p,\kk] = \sum_{\ss_{-p} \in \Psi^{-1}(\kk)} \left( \hspace{0.05cm}\prod_{q\ne p} x_{q,s_q}\right). $$ Note that $s_q$ denotes the pure strategy of player $q$ from a profile $\ss_{-p}\in \Psi^{-1}(\kk)$. We also use $u_p(\bxx)$ to denote the expected payoff of player $p$ from playing $\xx_p$: $$u_p(\bxx)=\sum_{b\in [\alpha]} x_{p,b}\cdot u_p(b,\calX).$$ It is worth pointing out that, while $u_p(b,\bxx)$ contains exponentially many terms, it can be computed in polynomial time using dynamic programming [@PTAS07; @dask14] when $\alpha$ is a constant. For a detailed presentation of the algorithm for $2$-strategy anonymous games, see [@dask14]. This then implies that checking whether a given profile $\calX$ is a (approximate) Nash equilibrium is in polynomial time. Next we define (approximate) Nash equilibria of an anonymous game. Given an anonymous game $\calG=(n,\alpha,\{\textsf{\emph{payoff}}_p\})$, we say a mixed strategy profile $\calX$ is a *Nash equilibrium* of $\calG$ if $u_p(\calX)\ge u_p(b,\calX)$ for all players $p\in [n]$ and strategies $b\in [\alpha]$. For $\epsilon\ge 0$, we say $\calX$ is an *$\epsilon$-approximate Nash equilibrium* if $u_p(\calX)+\epsilon\ge u_p(b,\calX)$ for all $p\in [n]$ and $b\in [\alpha]$. For $\epsilon\ge 0$, we say $\calX$ is an *$\epsilon$-well-supported Nash equilibrium* if $u_p(a,\calX)+\epsilon<u_p(b,\calX)$ implies that $x_{p,a}=0$, for all $p\in [n]$ and $a,b\in [\alpha]$. As discussed in Section \[sec:ours\], the hardness part of Theorem \[main-theorem\] is proved using a polynomial-time reduction from the problem of finding a well-supported Nash equilibrium in a *polymatrix game* (e.g. see [@Cai11]). For our purposes, such a game (with $n$ players and two strategies each player) can be described by a payoff matrix $\AA \in [0,1]^{2n\times 2n}$ with $A_{k,\ell}=0$ for all $k ,\ell \in \{2i-1,2i\}$ and $i \in [n]$. Each player $i\in [n]$ has two pure strategies that correspond to rows $2i-1$ and $2i$ of $\AA$. Let $\AA_{j}$ denote the $j$th row of $\AA$. Given a vector $\yy\in \mathbb{R}_{\ge 0}^{2n}$, where $(y_{2i-1},y_{2i})$ is the mixed strategy of player $i$, expected payoffs of player $i$ for playing rows $2i-1$ and $2i$ are $\AA_{2i-1}\cdot \yy$ and $\AA_{2i}\cdot \yy$ respectively. An *$\epsilon$-well-supported Nash equilibrium* of $\AA$ is a vector $\yy\in \mathbb{R}_{\ge 0}^{2n}$ such that $y_{2i-1}+y_{2i}=1$ and $$\begin{aligned} \AA_{2i-1} \cdot \yy > \AA_{2i} \cdot {\yy} +\epsilon \hspace{0.06cm} \Rightarrow\hspace{0.06cm} y_{2i}=0 \ \ \ \text{and}\ \ \ \AA_{2i}\cdot \yy > \AA_{2i-1}\cdot \yy +\epsilon \hspace{0.06cm} \Rightarrow\hspace{0.06cm} y_{2i-1}=0,\end{aligned}$$ for all players $i\in [n]$. We need the following result on such games: The problem of computing a $(1/n)$-well-supported Nash equilibrium in a\ polymatrix game is PPAD-complete. Our Approach and Techniques {#sec:ours} --------------------------- A commonly used approach to establishing the PPAD-hardness of approximate equilibria is to design gadget games that can perform certain arithmetic operations on entries of mixed strategies of players (e.g. see [@DGPJournal; @2Nash]). Such gadgets would then yield a reduction from the problem of solving a generalized circuit [@DGPJournal; @2Nash], a problem complete in PPAD. However, we realized that this approach may not work well with anonymous games; we found that it was impossible to design an anonymous game $G_{=}$ that enforces equality constraints.[^4] Instead we show the PPAD-hardness of anonymous games via a reduction from the problem of finding a $(1/n)$-well-supported equilibrium in a two-strategy polymatrix game (see Section \[sec:pre\]). Given a $2n\times 2n$ polymatrix game $\AA$, our reduction constructs an anonymous game $\calG_\AA$ with $n$ “main” players $\{P_1,\ldots,P_n\}$ (and two auxiliary players). We have each main player $P_i$ simulate in a way a player $i$ in the polymatrix game, as discussed below, such that any $\epsilon$-well-supported Nash equilibrium of $\calG_\AA$ with an exponentially small $\epsilon$ can be used to recover a $(1/n)$-well-supported Nash equilibrium of the polymatrix game $\AA$ efficiently. We then prove a connection between approximate Nash equilibria and well-supported Nash equilibria of anonymous games to finish the proof of Theorem \[main-theorem\]. *The greatest challenge to establishing such a reduction is posed by the complex but highly structured, symmetric expression of expected payoffs in an anonymous game.* As discussed previously in Section \[sec:pre\], the expected payoff $u_p(b,\calX)$ of player $p$ is a linear form of probabilities $\Pr_{\cal{X}}[p,\kk]$, each of which is function over mixed strategies of all players other than $p$. This rather complex function makes it difficult to reason about the set of well-supported Nash equilibria of an anonymous game, not to mention our goal is to embed a polymatrix game in it. To overcome this obstacle, we need to find a special (but hard enough) family of anonymous games with certain payoff structures which allow us to perform a careful analysis and understand their well-supported equilibria. The bigger obstacle for our reduction, however, is to in some sense *remove the anonymity of the players and break the inherent symmetry underlying an anonymous game*. To see this, a natural approach to obtain a reduction from polymatrix games is to directly encode the $2n$ variables of $\yy$ in mixed strategies of the $n$ “main” players $\{P_1,\ldots,P_n\}$. More specifically, let $\{s_1, s_2\}$ denote two special pure strategies of $\calG_\AA$, and we attempt to encode $(y_{2i-1}, y_{2i})$ in $(x_{i,s_1} , x_{i,s_2} )$, probabilities of $P_i$ playing $s_1,s_2$, respectively. The reduction would work if expected payoffs of $P_i$ from $s_1$ and $s_2$ in $\calG_\AA$ can always match closely expected payoffs of player $i$ from rows $2i - 1$ and $2i$ in $\AA$, given by two linear forms $\AA_{2i-1}\cdot \yy$ and $\AA_{2i}\cdot \yy$ of $\yy$. However, it seems difficult, if not impossible, to construct $\calG_\AA$ with this property, since anonymous games are highly symmetric: the expected payoff of $P_i$ is a symmetric function over mixed strategies of all other players. This is not the case for polymatrix games: a linear form such as $\AA_{2i}\cdot \yy$ in general has different coefficients for different variables, so different players contribute with different weights to the expected payoff of a player (and the problem of finding a well-supported equilibrium in $\AA$ clearly becomes trivial if we require that every row of $\AA$ has the same entry). An alternative approach is to encode the $2n$ variables of $\yy$ in probabilities $\Pr_{\cal{X}}[p,\kk]$. This may look appealing because expected payoffs $u_p(b,\calX)$ are linear forms of these probabilities so one can set the coefficients $\pay_p(b,\kk)$ to match them easily with those linear forms $\AA_{j}\cdot \yy$ that appear in the polymatrix game $\AA$. However, the histogram $\kk$ seen by a player $p$ (as a vector-valued random variable) is the sum of $n-1$ vector-valued random variables, each distributed according to the mixed strategy of a player other than $p$. The way these probabilities $\Pr_{\cal{X}}[p,\kk]$ are derived in turn imposes strong restrictions on them,[^5] which makes it a difficult task to obtain a correspondence between the $2n$ free variables in $\yy$ and the probabilities $\Pr_{\cal{X}}[p,\kk]$. Our reduction indeed follows the first approach of encoding $(y_{2i-1},y_{2i})$ in $(x_{i,s_1},x_{i,s_2})$ of player $P_i$. More exactly, the former is the normalization of the latter into a probability distribution. Now to overcome the difficulty posed by symmetry, we *enforce* the following *“scaling” property* in every well-supported Nash equilibrium $\calX$ of $\calG_\AA$: probabilities of $P_i$ playing $\{s_1,s_2\}$ satisfy $$\label{scales} x_{i,s_1}+x_{i,s_2}\approx 1/N^i,$$ where $N$ is exponentially large in $n$. This property is established by designing an anonymous game called *generalized radix game* $\calG_{n,N}^*$, and then using it as the base game in the construction of $\calG_\AA$. We show that (\[scales\]) holds approximately for every anonymous game that is payoff-wise *close* to $\calG_{n,N}^*$. In particular, (\[scales\]) holds for any well-supported equilibrium of $\calG_\AA$, as long as we make sure $\calG_\AA$ is close to $\calG_{n,N}^*$. The “scaling” property plays a crucial role in our reduction because, as the base game for $\calG_\AA$, it helps us reason about well-supported Nash equilibria of $\calG_\AA$; it also removes anonymity of the $n$ “main” players $P_i$ (since they must play the two special pure strategies $\{s_1,s_2\}$ with probabilities of different scales) and overcome the symmetry barrier. Equipped with the “scaling” property (\[scales\]), we prove a key technical lemma called the *estimation lemma*. It shows that one can compute efficiently coefficients of a linear form over probabilities of histograms $\Pr_\calX[P_i,\kk]$ seen by player $P_i$, which guarantees to approximate additively $x_{j,s_1}$ (or $x_{j,s_2}$) i.e. probability of another player $P_j$ plays $s_1$ (or $s_2$), whenever the profile $\calX$ satisfies the “scaling” property (this holds when $\calG_\AA$ is close to $\calG_{n,N}^*$ and $\calX$ is a well-supported equilibrium of $\calG_\AA$). As $$(y_{2j-1},y_{2j})\approx N^j(x_{j,s_1},x_{j,s_2})$$ given (\[scales\]), these linear forms for $x_{j,s_1},x_{j,s_2}$ can be combined to derive a linear form of $\Pr_\calX[P_i,\kk]$ to approximate additively any linear form of $\yy$, particularly $\AA_{2i-1}\cdot \yy$ or $\AA_{2i}\cdot \yy$ that appear as expected payoffs of player $i$ in the polymatrix game $\AA$. The proof of the estimation lemma is the technically most involved part of the paper. We indeed derive explicit expressions for coefficients of the desired linear form where substantial cancellations yield an additive approximation of $x_{j,s_1}$ or $x_{j,s_2}$. Finally we combine all ingredients highlighted above to construct an anonymous game $\calG_\AA$ from polymatrix game $\AA$. This is done by first using the estimation lemma to compute, for each main $P_i$ coefficients of linear forms of probabilities $\Pr_\calX[P_i,\kk]$ seen by $P_i$ that yield additive approximations of $x_{j,s_1}$ and $x_{j,s_2}$. We then perturb payoff functions of players $P_i$ in the generalized radix game $\calG_{n,N}^*$ using these coefficients so that 1) the resulting game $\calG_\AA$ is close to $\calG_{n,N}^*$ and thus, any well-supported equilibrium $\calX$ of $\calG_\AA$ automatically satisfies the “scaling” property; 2) expected payoffs of $P_i$ playing $s_1,s_2$ in a well-supported equilibrium $\calX$ of $\calG_\AA$ match additively expected payoffs of player $i$ playing rows $2i-1,2i$ in $\AA$, given $\yy$ derived from $\calX$ by normalizing $(x_{j,s_1},x_{j,s_2})$ for each $j$. The correctness of the reduction, i.e., $\yy$ is a $(1/n)$-well-supported equilibrium of $\AA$ whenever $\calX$ is an $\epsilon$-well-supported equilibrium of $\calG_\AA$ with an exponentially small $\epsilon$, follows from these properties of $\calG_\AA$. Organization ------------ In Section 2, we define the radix game, and show that it has a unique Nash equilibrium as a warm-up. We also use it to define the generalized radix game which serves as the base of our reduction. In section 3, we characterize well-supported Nash equilibria of anonymous games that are close to the generalized radix game (i.e., those that can be obtained by adding small perturbations to payoffs of the generalized radix game). In section 4, we prove the PPAD-hardness part of the main theorem. Our reduction relies on a crucial technical lemma, called the estimation lemma, which we prove in Section 5. We prove the membership in Section 6, and conclude with open problems in Section 7. Warm-up: Radix Game =================== In this section, we first define a $(n+2)$-player anonymous game $\calG_{n,N}$, called the *radix game*. As a warmup for the next section, we show that it has a unique Nash equilibrium. We then use the radix game to define the *generalized radix game* $\calG_{n,N}^*$, by making a duplicate of a pure strategy in $\calG_{n,N}$. The latter will serve as the base game for our polynomial-time reduction from polymatrix games. Radix Game ---------- The radix game $\calG_{n,N}$ to be defined has a unique Nash equilibrium of a specific form: given $N\ge 2$ as an integer parameter of the game, each of the $n$ “main” players mixes over the first two strategies with probabilities $1/N^i$ and $1-1/N^i$, respectively, for each $i\in [n]$, in the unique Nash equilibrium. The remaining two “special” players are created to achieve the aforementioned property. \[gadget1\] Let $n\ge 1$ and $N\ge2$ denote two integer parameters. Let $\delta=1/N$. Let $\calG_{n,N}$ denote the following anonymous game with $n+2$ players $\{\hspace{-0.03cm}P_1,\ldots,P_n,Q,R\}$ and $6$ pure strategies $\{s,t,q_1,q_2,r_1,r_2\}$. We refer to $\{P_1,\ldots,P_n\}$ as the *main* players. Each main player $P_i$ is only interested in strategies $s$ and $t$ *(*e.g., by setting her payoff of playing any other four actions to be $-1$ no matter what other players play*)*. Player $Q$ is only interested in strategies $\{q_1,q_2\}$, and player $R$ is only interested in strategies $\{r_1,r_2\}$. Next we define the payoff function of each player. When describing the payoff of a player below we always use $\kk=(k_s,k_t,k_{q_1},k_{q_2},k_{r_1},k_{r_2})$ to denote the histogram of strategies this player sees. 1. For each $i\in [n]$, the payoff of player $P_i$ when she plays $s$ only depends on $k_s$: $$\text{\emph{\textsf{payoff}}}_{P_i}(s,\kk)= \begin{cases} \hspace{0.06cm} \delta^i+ \prod_{j\in [n]} \delta^j& \text{if $k_s=n-1$}\\[0.9ex] \hspace{0.06cm} \prod_{j\in [n]} \delta^j & \text{otherwise.} \end{cases}\vspace{0.04cm}$$ The payoff of player $P_i$ when she plays $t$ only depends on $k_{r_1}$: $$\text{\emph{\textsf{payoff}}}_{P_i}(t,\kk)=\begin{cases} \hspace{0.06cm}2& \text{if $k_{r_1}= 1$}\\[0.4ex] \hspace{0.06cm}0 & \text{otherwise.} \end{cases}$$ 2. The payoff of player $Q$ when she plays $q_1$ or $q_2$ is given by $$\text{\emph{\textsf{payoff}}}_Q(q_1,\kk)=\begin{cases} \hspace{0.06cm}1& \text{if $k_{s}= n$}\\[0.4ex] \hspace{0.06cm}0 & \text{otherwise} \end{cases} \ \ \ \text{and}\ \ \ \ \text{\emph{\textsf{payoff}}}_Q(q_2,\kk)=\begin{cases} \hspace{0.06cm}1& \text{if $k_{r_1}= 1$}\\[0.4ex] \hspace{0.06cm}0 & \text{otherwise.} \end{cases}$$ 3. The payoff of player $R$ when she plays $r_1$ or $r_2$ is given by $$\text{\emph{\textsf{payoff}}}_R(r_1,\kk)=\begin{cases} \hspace{0.06cm}1& \text{if $k_{q_1}=1$}\\[0.4ex] \hspace{0.06cm}0 & \text{otherwise} \end{cases} \ \ \ \text{and}\ \ \ \ \text{\emph{\textsf{payoff}}}_R(r_2,\kk)=\begin{cases} \hspace{0.06cm}1& \text{if $k_{q_2}=1$}\\[0.4ex] \hspace{0.06cm}0 & \text{otherwise.} \end{cases}$$ This finishes the definition of the radix game $\calG_{n,N}$. $\calG_{n,N}$ is an anonymous game with payoff functions taking values from $[-1,2]$. Since the main players $P_i$ are only interested in $\{s,t\}$, $Q$ is only interested in $\{q_1,q_2\}$, and $R$ is only interested in $\{r_1,r_2\}$, each Nash equilibrium $\calX$ of $\calG_{n,N}$ can be fully specified by a $(n+2)$-tuple $\calX=(x_1,\ldots,x_n,y,z)\in [0,1]^{n+2}$, where $x_i$ denotes the probability of $P_i$ playing strategy $s$ for each $i\in [n]$, $y$ denotes the probability of $Q$ playing $q_1$, and $z$ denotes the probability of $R$ playing $r_1$. Given $\calX=(x_1,\ldots,x_n,y,z)$ we calculate the expected payoff of each player as follows (we skip $\calX$ in the expected payoffs $u_p(b,\calX)$, when $\calX$ is clear from the context, and we use $u_i$ to denote the expected payoff of $P_i$ instead of $u_{P_i}$ for convenience): \[fact1\] Given $\calX=(x_1,\ldots,x_n,y,z)$, the expected payoff of player $P_i$ for playing $s$ is $$u_{i}(s)=\delta^i\cdot \emph{\Pr}\big[k_s=n-1\big] + \prod_{j \in [n]} \delta^j = \delta^i \prod_{j\ne i \in [n]} x_j + \prod_{j \in [n]} \delta^j.$$ The expected payoff of $P_i$ for playing $t$ is $u_{i}(t)=2z$. The expected payoff of player $Q$ for playing $q_1$ is $$u_{Q}(q_1)= \emph{\Pr}\big[k_s=n\big]=\prod_{i\in [n]} x_i.\vspace{-0.03cm}$$ The expected payoff of $Q$ for playing $q_2$ is $u_Q(q_2)=z$. The expected payoff of $R$ for playing $r_1$ is $u_R(r_1)=y$ and that for $r_2$ is $u_R(r_2)=1-y$. We show that $x_i=\delta^i$ in a Nash equilibrium $\calX$ of $\calG_{n,N}$. We start with the following lemma. \[notuseful1\] In a Nash equilibrium $\calX=(x_1,\ldots,x_n,y,z)$ of $\calG_{n,N}$, we have that $z=\prod_{i\in [n]}x_i$. \[util\_s5\] Assume for contradiction that $z>\prod_{i} x_i$. As $u_{Q}(q_2) > u_{Q}(q_1)$ and $\calX$ is a Nash equilibrium, player $Q$ never plays $q_1$ and thus, $y=0$. This in turn implies $u_{R}(r_2)=1 >0= u_{R}(r_1)$ and $z=0$, which contradicts with the assumption that $z>\prod_i x_i\ge 0$. Next, assume for contradiction that $z<\prod_i x_i$, giving us that $u_{Q}(q_2) < u_{Q}(q_1)$. Player $Q$ never plays $q_2$ and $y=1$. This implies that $u_{R}(r_1)>u_{R}(r_2)$ and thus $z=1$, which contradicts with the assumption that $z<\prod_i x_i\le 1$ (as $x_i\in [0,1]$). This finishes the proof of the lemma. We now show that the radix game $\calG_{n,N}$ has a unique Nash equilibrium $\calX$ with $x_i=\delta^i$. \[notuseful2\] In a Nash equilibrium $\calX=(x_1,\ldots,x_n,y,z)$ of $\calG_{n,N}$, we have $x_i = \delta^i$ for all $i\in [n]$. \[powers\] First we show that $\prod_{i\in [n]} x_i=\prod_{i\in [n]} \delta^i$. Consider for contradiction the following two cases: 1. [Case 1]{}: $\prod_{i\in [n]} x_i<\prod_{i\in [n]} \delta^i$. Then there is an $i\in [n]$ such that $x_i<\delta^i$. For $P_i$, we have $$\label{eq:ha1} u_i(s)=\delta^i\prod_{j\ne i} x_j + \prod_{j\in [n]} \delta^j > \prod_{j\in [n]} x_j + \prod_{j\in [n]} x_j= 2\prod_{j\in [n]} x_j=2z=u_i(t).$$ This implies that $x_i=1$, contradicting with the assumption that $x_i<\delta^i<1$ as $N\ge 2$. 2. [Case 2]{}: $\prod_{i\in [n]} x_i>\prod_{i\in [n]} \delta^i$. Then there is an $i\in [n]$ such that $x_i>\delta^i$. For $P_i$, we have $$\label{eq:ha2} u_i(s)=\delta^i \prod_{j\ne i} x_j + \prod_{j\in [n]} \delta^j < \prod_{j\in [n]} x_j + \prod_{j\in [n]} x_j= 2\prod_{j\in [n]} x_j=2z=u_i(t).$$ This implies that $x_i=0$, contradicting with the assumption that $x_i>\delta^i>0$. As a result, we must have $\prod_i x_i=\prod_i \delta^i$, which also implies that $x_i>0$ for all $i\in [n]$. Now we show that $x_i=\delta^i$ for all $i$. Assume for contradiction that $x_i\ne \delta^i$ for some $i\in [n]$. 1. [Case 1]{}: $x_i<\delta^i$. Then the same strict inequality (\[eq:ha1\]) holds for $P_i$, which implies that\ $x_i=1$, contradicting with the assumption that $x_i<\delta^i<1$ as $N\ge 2$. 2. [Case 2]{}: $x_i>\delta^i$. Then the same strict inequality (\[eq:ha2\]) holds for $P_i$, which implies that\ $x_i=0$, contradicting with the assumption that $x_i>\delta^i>0$. This finishes the proof of the lemma. Notice that Lemma \[notuseful1\] and \[notuseful2\] together imply that $\calG_{n,N}$ has a unique Nash equilibrium because of Lemma \[notuseful1\] as well as the fact that $0<z<1$ implies $u_R(r_1)=y=1-y=u_R(r_2)$ and thus $y=1/2$. Generalized Radix Game ---------------------- We use $\calG_{n,N}$ to define an anonymous game $\calG^*_{n,N}$, called the *generalized radix game*, with the same set of $n+2$ players $\{P_1,\ldots,P_n,Q,R\}$ but seven strategies $\{s_1,s_2,t,q_1,q_2,r_1,r_2\}$. To this end, we replace strategy $s$ in $\calG_{n,N}$ with two of its duplicate strategies $s_1,s_2$ in $\calG^*_{n,N}$ and make sure that the players in $\calG^*_{n,N}$ treat both $s_1$ and $s_2$ the same as the old strategy $s$, and have their payoff functions derived from those of players in $\calG_{n,N}$ in this fashion. We will show in the next section that in any Nash equilibrium of $\calG^*_{n,N}$, player $P_i$ must have probability exactly $\delta^i$ distributed among $s_1,s_2$. For readers who are familiar with previous PPAD-hardness results of Nash equilibria in normal form games [@DGPJournal; @2Nash], this is the same trick used to derive the game *generalized matching pennies* from *matching pennies*. We define $\calG^*_{n,N}$ formally as follows. Let $n\ge 1$ and $N\ge 2$ be two parameters. Let $\delta=1/N$. We use $\calG^*_{n,N}$ to denote an anonymous game with the same $n+2$ players $\{P_1,\ldots,P_n,Q,R\}\hspace{-0.03cm}$ as $\calG_{n,N}$ but now $7$ pure strategies $\{s_1,s_2,t,q_1,q_2,r_1,r_2\}$. The payoff function $\emph{\pay}_T^*$ of a player $T$ in $\calG^*_{n.N}$ is defined using $\emph{\pay}_T$ of the same player $T$ in $\calG_{n,N}$ as follows: $$\emph{\pay}_T^*\Big(b,\big(k_{s_1},k_{s_2},k_t, k_{q_1},k_{q_2},k_{r_1},k_{r_2}\big)\Big)= \emph{\pay}_T\Big(\phi(b),\big(k_{s_1}+k_{s_2},k_t, k_{q_1},k_{q_2},k_{r_1},k_{r_2}\big)\Big),$$ where $\phi(s_1)=\phi(s_2)=s$ and $\phi(b)=b$ for every other pure strategy. Since the payoff of player $P_i$ is always $-1$ when playing $q_1,q_2,r_1$ or $r_2$, she is only interested in $s_1,s_2$ and $t$. Similarly $Q$ is only interested in $q_1,q_2$ and $R$ is only interested in $r_1,r_2$. As a result, a Nash equilibrium $\calX$ of $\calG^*_{n,N}$ can be fully specified by $2n+2$ numbers $(x_{i,1},x_{i,2},y,z:i\in [n])$, where $x_{i,1}$ (or $x_{i,2}$) denotes the probability of $P_i$ playing strategy $s_{1}$ (or strategy $s_2$, respectively), so the probability of $P_i$ playing $t$ is $1-x_{i,1}-x_{i,2}$. We also let $x_i=x_{i,1}+x_{i,2}$ for each $i\in [n]$. Given the definition of $\calG^*_{n,N}$ from $\calG_{n,N}$, Lemma \[notuseful2\] suggests $x_i=x_{i,1}+x_{i,2}=\delta^i$, for all $i\in[n]$, in every Nash equilibrium $\calX$ of $\calG^*_{n,N}$. This indeed follows from the main lemma of the next section concerning $\epsilon$-well-supported Nash equilibria of not only the generalized radix game $\calG_{n,N}^*$ itself, but also anonymous games obtained by perturbing payoff functions of $\calG_{n,N}^*$. Generalized Radix Game after Perturbation ========================================= In this section, we analyze $\epsilon$-well-supported Nash equilibria of anonymous games obtained by perturbing payoff functions of the generalized radix game $\calG^*_{n,N}$. Recall that $n\ge 1$ and $N\ge 2$, and we use $\pay_T^*$ to denote the payoff function of a player $T$ in $\calG^*_{n,N}$. Given $x,y\in \mathbb{R}$ and $\xi\ge 0$, we write $x=y\pm \xi$ to denote $|x-y|\le \xi$. We first define anonymous games that are close to $\calG_{n,N}^*$. For $\xi\ge 0$, we say an anonymous game $\calG$ is *$\xi$-close* to $\calG_{n,N}^*$ if 1. $\calG$ has the same set $\{P_1,\ldots,P_n,Q,R\}$ of players and same set of $7$ strategies as $\calG_{n,N}^*$. 2. For each player $T\in \{P_1,\ldots,P_n,Q,R\}$, her payoff function $\emph{\pay}_T$ in $\calG$ satisfies $$\textsf{\emph{payoff}}_T(b,\kk)=\textsf{\emph{payoff}}_T^*(b,\kk)\pm \xi,$$ for all $b\in \{s_1,s_2,t,q_1,q_2,r_1,r_2\}$ and all histograms $\kk$ of strategies played by $n+1$ players. To characterize $\epsilon$-well-supported Nash equilibria of a game $\calG$ $\xi$-close to $\calG_{n,N}^*$ we first show that when $\epsilon$, $\xi$ are small enough, each player in $\calG$ remains only interested in a subset of strategies, i.e., $\{s_1,s_2,t\}$ for $P_i$, $\{q_1,q_2\}$ for $Q$, and $\{r_1,r_2\}$ for $R$, in any $\epsilon$-well-supported Nash equilibrium of $\calG$. \[lem:interested\] Let $\calG$ be an anonymous game $\xi$-close to $\calG_{n,N}^*$ for some $\xi\ge 0$. When $2\hspace{0.02cm}\xi+\epsilon<1$, every $\epsilon$-well-supported Nash equilibrium of $\calG$ satisfies: player $P_i$ only plays $\{s_1,s_2,t\}$; player $Q$ only plays $\{q_1,q_2\}$; player $R$ only plays $\{r_1,r_2\}$. We only prove (1) since the proof of (2) and (3) is similar. Given an $\epsilon$-well-supported Nash equilibrium $\calX$, as the payoff of $P_i$ when playing $b\notin \{s_1,s_2,t\}$ is always $-1$ in $\calG_{n,N}^*$, her expected payoff when playing $b$ in $\calG$ is at most $-1+\xi$; as the payoff of $P_i$ when playing $b\in \{s_1,s_2,t\}$ is always nonnegative in $\calG_{n,N}^*$, her expected payoff in $\calG$ is at least $-\xi$. It follows from $2\hspace{0.02cm}\xi+\epsilon<1$ and the assumption of $\calX$ being an $\epsilon$-well-supported equilibrium that $P_i$ only plays strategies in $\{s_1,s_2,t\}$ with positive probability. It follows from Lemma \[lem:interested\] that an $\epsilon$-well-supported Nash equilibrium of $\calG$ can be fully described by a tuple of $2n+2$ numbers $(x_{i,1},x_{i,2},y,z:i\in [n])$, when $\xi,\epsilon$ satisfy $2\hspace{0.02cm}\xi+\epsilon<1$: $x_{i,1}$ denotes the probability of $P_i$ playing $s_1$, $x_{i,2}$ denotes the probability of $P_i$ playing $s_2$, $y$ denotes the probability of $Q$ playing $q_1$, and $z$ denotes the probability of $R$ playing $r_1$. Recall that $\delta=1/N\le 1/2$. Let $\kappa=\prod_{i\in [n]} \delta^i.$ We prove the main lemma of this section. \[lem:perturb\] Let $\calG$ denote an anonymous game that is $\xi$-close to $\calG_{n,N}^*$. Suppose that $\xi,\epsilon\ge 0$ satisfy $$\label{condition} \tau=\frac{36\hspace{0.02cm}\xi+18\hspace{0.02cm}\epsilon}{\kappa}\le 1/2.$$ Then every $\epsilon$-well-supported Nash equilibrium of $\calG$ satisfies $x_{i,1}+x_{i,2}=\delta^i\pm \tau\delta^i$ for all $i\in [n]$. Let $\hspace{-0.03cm}\calX\hspace{-0.03cm}=(x_{i,1},x_{i,2},y,z:i\in [n])$ be an $\epsilon$-well-supported Nash equilibrium of $\calG$. For each $i\in $ $[n]$ we let $x_i=x_{i,1}+x_{i,2}$. Since $\calG$ is $\xi$-close to $\calG_{n,N}^*$, we have the following estimates: 1. The expected payoff of $P_i$ for playing strategy $s_1$ or $s_2$ is $$\begin{aligned} &u_i(s_1),\hspace{0.04cm}u_{i}(s_2)=\left(\delta^i\cdot \Pr\big[k_{s_1}+k_{s_2}=n-1\big] + \prod_{j \in [n]} \delta^j\right)\pm\xi = \left(\delta^i \prod_{j\ne i} x_j +\kappa\right) \pm\xi,&\end{aligned}$$ where we write $k_{s_1},k_{s_2}$ to denote the numbers of players that play $s_1,s_2$ respectively, as seen by player $P_i$ (same below). The expected payoff of $P_i$ for playing $t$ is $u_{i}(t)=2z\pm \xi$. 2. The expected payoff of $Q$ for playing $q_1$ is $$\begin{aligned} &u_{Q}(q_1)= \Pr\big[k_{s_1}+k_{s_2}=n\big]\pm \xi=\prod_{j\in [n]} x_j\pm \xi.&\end{aligned}$$ The expected payoff of $Q$ for playing $q_2$ is $u_Q(q_2)=z\pm \xi$. 3. The expected payoff of $R$ for playing $r_1$ is $u_R(r_1)=y\pm \xi$ and for $r_2$ is $u_R(r_2)=(1-y)\pm \xi$. To rest of the proof follows those of Lemma \[util\_s5\] and Lemma \[powers\]. First we show that $z$ must satisfy $$\label{haha1} z=\prod_{j\in [n]} x_j\pm (2\xi+\epsilon).$$ The proof is the same as that of Lemma \[util\_s5\], using the assumption that $\calX$ is $\epsilon$-well-supported. Given (\[haha1\]), next we show that the $x_i$’s satisfy $$\label{haha2} \prod_{i\in [n]}x_i=\prod_{i\in [n]} \delta^i\pm (6 \xi+3\epsilon)=\kappa\pm (6 \xi+3\epsilon).$$ To this end we follow the proof of the first part of Lemma \[powers\] and consider the following two cases: 1. [Case 1]{}: $\prod_{i\in [n]} x_i<\kappa-(6\xi+3\epsilon)$. Then there exists an $i\in [n]$ such that $x_i<\delta^i$. For $P_i$: $$\hspace{-0.1cm}u_i(s_1)\ge \delta^i\prod_{j\ne i} x_j + \kappa-\xi > 2\prod_{j\in [n]} x_j +5\xi+3\epsilon\ \ \ \text{and}\ \ \ u_i(t)\le 2z+\xi\le 2\prod_{j\in [n]}x_j+5\xi+2\epsilon.$$ This implies that $P_i$ does not play $t$ in $\calX$, an $\epsilon$-well-supported Nash equilibrium of $\calG$, and thus, $x_i=x_{i,1}+x_{i,2}=1$, contradicting with $x_i<\delta^i<1$ as $N\ge 2$. 2. [Case 2]{}: $\prod_{i\in [n]} x_i>\kappa+(6\xi+3\epsilon)$. Then there exists an $i\in [n]$ such that $x_i>\delta^i$. For $P_i$: $$\hspace{-0.1cm}u_i(s_1),\hspace{0.03cm}u_i(s_2)\le \delta^i \prod_{j\ne i} x_j + \kappa+\xi < 2\prod_{j\in [n]} x_j -5\xi-3\epsilon\ \ \ \text{and}\ \ \ u_i(t)\ge 2\prod_{j\in [n]} x_j-5\xi-2\epsilon.$$ This implies that $P_i$ plays neither $s_1$ nor $s_2$ and thus, we have $x_{i,1}=x_{i,2}=0$ and $x_i=0$ as well, contradicting with $x_i>\delta^i>0$. By (\[haha1\]) and (\[haha2\]), $z=\kappa\pm (8\xi+4\epsilon)$. (\[haha2\]) also implies that $x_i>0$ since $\kappa>0$ and $\kappa\ge 72\xi+36\epsilon$ by (\[condition\]). Finally, assume for contradiction that either $x_i<(1-\tau)\delta^i$ or $x_i>(1+\tau)\delta^i$ for some $i\in [n]$. 1. Case 1: [$x_i<(1-\tau)\delta^i$]{}. Then using $\tau\le 1/2$ and $1\le 1/(1-\tau)\le 2$, we have $$\begin{aligned} u_i(s_1)-u_i(t)\ge \delta^i\prod_{j\ne i} x_j+ \kappa-2z-2\xi> \frac{\kappa-6\xi-3\epsilon}{1-\tau} +\kappa -2z-2\xi\ge\tau\kappa-30\xi-14\epsilon.\end{aligned}$$ Plugging in the definition of $\tau$ in (\[condition\]), we have $u_i(s_1)-u_i(t)>\epsilon$ and thus, $x_i=1$, which contradicts with the assumption that $x_i<(1-\tau)\delta^i<1$. 2. Case 2: $x_i>(1+\tau)\delta^i$. Then using $\tau\le 1/2$ and $2/3\le 1/(1+\tau)\le 1$, we have $$\begin{aligned} u_i(s_1)-u_i(t)\le \delta^i\prod_{j\ne i} x_j+\kappa-2z+2\xi < \frac{\kappa+6\xi+3\epsilon}{1+\tau} +\kappa -2z+2\xi\le -\frac{2\tau\kappa}{3}+24\xi+11\epsilon.\end{aligned}$$ The same inequality holds for $u_i(s_2)-u_i(t)$. Plugging in (\[condition\]), we have $u_i(s_1)-u_i(t)<-\epsilon$ as well as $u_i(s_2)-u_i(t)<-\epsilon$. This in turn implies that $x_{i,1}=x_{i,2}=0$ and thus, $x_i=0$, which contradicts with the assumption that $x_i>(1+\tau)\delta^i>0$. This finishes the proof of the lemma. Reduction from Polymatrix Games to Anonymous Games {#main reduction} ================================================== In this section we prove the hardness part of Theorem \[main-theorem\]. For this purpose we present a polynomial time reduction from the problem of finding a $1/n$-well-supported Nash equilibrium in a polymatrix game to the problem of finding an $\epsilon$-well-supported Nash equilibrium in an anonymous game with $7$ strategies, for some exponentially small $\epsilon$. We first give some intuition behind this quite involved reduction in Section \[hahasec\]. Details of the reduction and the proof of its correctness are then presented in Section \[sec:reduction\] and \[sec:correctness\], respectively, with a key technical lemma proved in Section \[sec:estimation\]. We finish the proof of the hardness part in Section \[sec:hardness\] by showing that any approximate Nash equilibrium of an anonymous game can be converted into a well-supported equilibrium efficiently (since Theorem \[main-theorem\] is concerned with approximate Nash equilibria). Overview of the Reduction {#hahasec} ------------------------- Given as input a polymatrix game specified by a matrix $\AA \in [0,1]^{2n\times 2n}$, our goal is to construct in polynomial time an anonymous game $\calG_\AA$, and show that every $\epsilon$-well-supported Nash equilibrium of $\calG_\AA$, where $\epsilon=1/2^{n^6}$, can be used to recover a $(1/n)$-well-supported equilibrium of $\AA$ in polynomial time. Note that this is not exactly the PPAD-hardness result as claimed in Theorem \[main-theorem\] but we will fill in the gap in Section \[sec:hardness\] with some standard arguments. Given $\AA$, we construct $\calG_\AA$ by perturbing payoff functions of the Generalized Radix game ${\cal{G}}^*_{n,N}$ with $N=2^n$, so that $\calG_\AA$ is $\xi$-close to $\calG^*_{n,N}$ for some exponentially small $\xi>0$ to be specified later. (Thus, $\calG_\AA$ has the same set of $n+2$ players $\{P_1,\ldots,P_n,Q,R\}$ as well as the same set of $7$ strategies $\{s_1,s_2,t,q_1,q_2,r_1,r_2\}$ as $\calG^*_{n,N}$.) By Lemma \[lem:interested\] and Lemma \[lem:perturb\] we know that every $\epsilon$-well-supported equilibrium of $\calG_\AA$ can be fully described by a tuple $\calX=(x_{i,1},x_{i,2},y,z:i\in [n])$ that satisfies $$\label{eq:approximate} x_{i,1}+x_{i,2}\approx \delta^i$$ for each $i\in [n]$, where $\delta=1/N=1/2^n$. Our construction of $\calG_\AA$ has player $P_\ell$ simulate row $2\ell-1$ and $2\ell$ of the polymatrix game $\AA$ for each $\ell\in [n]$. The goal is to show at the end that, after normalizing $(x_{\ell,1},x_{\ell,2})$, i.e., probabilities of $P_\ell$ playing $s_1,s_2$ in an $\epsilon$-well-supported equilibrium $\calX$ of $\calG_\AA$, into a distribution $(y_{2\ell-1},y_{2\ell})$: $$\label{eq:back} y_{2\ell-1}=\frac{x_{\ell,1}}{x_{\ell,1}+x_{\ell,2}}\ \ \ \text{and}\ \ \ y_{2\ell}=\frac{x_{\ell,2}}{x_{\ell,1}+x_{\ell,2}},$$ we get a $(1/n)$-well-supported Nash equilibrium $\yy=(y_1,\ldots,y_{2n})$ of $\AA$. By (\[eq:approximate\]) we have $$y_{2\ell-1}\approx N^\ell\cdot x_{\ell,1}\ \ \ \text{and}\ \ \ y_{2\ell}\approx N^{\ell}\cdot x_{\ell,2}.$$ For player $P_\ell$ to simulate row $2\ell-1$ and $2\ell$ of the polymatrix game $\AA$, we perturb the original payoff function $\pay^*_\ell$ of $P_\ell$ in ${\cal{G}}^*_{n,N}$ in a way such that the following two linear forms of $\yy$: $$\AA_{2\ell-1}\cdot \yy=\sum_{j\notin \{2\ell-1,2\ell\}}A_{2\ell-1,j}\cdot y_j\ \ \ \text{and} \ \ \ \AA_{2\ell}\cdot \yy=\sum_{j\notin\{2\ell-1,2\ell\}} A_{2\ell,j}\cdot y_j$$ appear as additive terms in the expected payoffs $u_\ell(s_1,\calX)$ and $u_\ell(s_2,\calX)$ of $P_\ell$ obtained from $s_1,s_2$, respectively. Let $u^*_\ell(\sigma, {\bxx})$ denote the expected payoff of player $P_\ell$ in the original generalized radix game ${\cal{G}}^*_{n,N}$ for strategies $\sigma \in \{s_1,s_2\}$. Then more specifically, we would like to perturb carefully the payoff functions of ${\cal{G}}^*_{n,N}$ such that for every $\ell\in [n]$, the expected payoffs of player $P_\ell$ in an $\epsilon$-well-supported Nash equilibrium $\bxx$ of $\cal{G}_\AA$ satisfy $$\begin{aligned} u_\ell(s_1,\bxx)&\approx u^*_\ell(s_1,\bxx) + \xi^*\cdot \AA_{2\ell-1}\cdot \yy \nonumber\\ \label{reduct} &\approx u^*_\ell(s_1,\bxx)+\xi^* \sum_{j\ne \ell} N^j \big(A_{2\ell-1,2j-1}\cdot x_{j,1}+A_{2\ell-1,2j}\cdot x_{j,2}\big)\\[0.5ex] u_\ell(s_2, {\bxx})&\approx u^*_\ell(s_2, {\bxx}) + \xi^* \cdot \AA_{2\ell}\cdot \yy \nonumber\\ \label{reduct2} &\approx u^*_\ell(s_2,\bxx)+\xi^* \sum_{j\ne \ell} N^j \big(A_{2\ell,2j-1}\cdot x_{j,1}+A_{2\ell,2j}\cdot x_{j,2}\big)\end{aligned}$$ where $\xi^*$ is a parameter small enough to make sure that the resulting game is $\xi$-close to $\calG_{n,N}^*$. If one can perturb the payoff functions of players $P_\ell$ in $\calG^*_{n,N}$ so that (\[reduct\]) and (\[reduct2\]) hold for every $\epsilon$-well-supported Nash equilibrium $\calX$ of $\calG_\AA$, then the vector $\yy$ obtained from $\calX$ using (\[eq:back\]) must be a $(1/n)$-well-supported equilibrium of $\AA$. To see this, assume for contradiction that $$\label{eq:contradiction} \AA_{2\ell-1}\cdot \yy>\AA_{2\ell}\cdot \yy+1/n$$ but $y_{2\ell}>0$. Using (\[eq:contradiction\]), (\[reduct\]), and (\[reduct2\]), we have $u_\ell(s_1,\bxx)$ is bigger than $u_\ell(s_2,\bxx)$ by $\xi^*/n$ (assuming that errors hidden in both (\[reduct\]) and (\[reduct2\]) are negligible). As long as our choice of $\xi^*$ satisfies $\xi^*/n>\epsilon$ we must have $x_{\ell,2}=0$ and thus, $y_{2\ell}=0$ from (\[eq:back\]). However, perturbing the generalized radix game so that (\[reduct\]) and (\[reduct2\]) hold is challenging. While $$\label{eq:desired} \sum_{j\ne \ell} N^j\big(A_{2\ell-1,2j-1}\cdot x_{j,1}+A_{2\ell-1,2j}\cdot x_{j,2}\big)\ \ \ \text{and}\ \ \ \sum_{j\ne \ell} N^j\big(A_{2\ell,2j-1}\cdot x_{j,1}+A_{2\ell,2j}\cdot x_{j,2}\big)$$ are merely two linear forms of $(x_{j,1},x_{j,2}:j\ne \ell)$ from $\calX$, they are extremely difficult to obtain due to the nature of anonymous games: the expected payoff of player $P_\ell$ is $$\label{eq:expected} u_\ell(\sigma,\calX)=\sum_{\kk \in K} \pay_\ell(\sigma,\kk) \cdot \Pr_{\cal{X}}[P_\ell,\kk] ,$$ a linear form of $\Pr_{\cal{X}}[P_\ell,\kk]$, the probability of $P_\ell$ seeing histogram $\kk$ given $\calX$. As each $\Pr_{\cal{X}}[P_\ell,\kk]$ is a highly complex and symmetric expression of variables in $\calX$, it is not clear how one can extract from (\[eq:expected\]) the desired linear forms of (\[eq:desired\]). This is where the fact that $x_{i,1}+x_{i,2}\approx \delta^i$ helps us tremendously. (Recall that this holds as long as the generalized radix game $\calG^*_{n,N}$ and $\calG_\AA$ are $\xi$-close.) The core of the construction of $\calG_\AA$ uses the following key technical lemma which we refer to as the *estimation lemma*. It shows that under any mixed strategy profile $\calX=(x_{i,1},x_{i,2},y,z:i\in [n])$ such that $x_{i,1}+x_{i,2}\approx \delta^i$, there is indeed a linear form of $\Pr_{\cal{X}}[P_\ell,\kk]$ that gives us a close approximation of $x_{j,1}$ (or $x_{j,2}$), $j\ne \ell$, and its coefficients can be computed in polynomial time in $n$. We delay its proof to Section \[sec:estimation\]. \[lem:estimation-lemma\] Let $N=2^n$ and ${\lambda= 2^{-n^3}}$. Given $\ell\in [n]$ and $j\ne \ell\in [n]$ one can compute in polynomial time in $n$ vectors $\BB^{[\ell,j]},\CC^{[\ell,j]}$ of length $|K|$ *(*indexed by $\kk\in K$*)* such that every mixed strategy profile $\calX=(x_{i,1},x_{i,2},y,z:i\in [n])$ with $x_{i,1}+x_{i,2}=\delta^i\pm \lambda$ for all $i$ satisfies $$\sum_{\kk \in K} B^{[\ell,j]}_\kk \cdot \emph{\Pr}_{\calX}[P_\ell,\kk] = x_{j,1} \pm O\hspace{-0.05cm}\left( j^2 \delta^{j+1}\right)\ \ \ \text{and}\ \ \ \sum_{\kk \in K} C^{[\ell,j]}_\kk \cdot \emph{\Pr}_{\calX}[P_\ell,\kk] = x_{j,2} \pm O\hspace{-0.05cm}\left( j^2 \delta^{j+1}\right).$$ Moreover, the absolute value of each entry of $\BB^{[\ell,j]}$ and $\CC^{[\ell,j]}$ is at most $N^{n^2}$. With the estimation lemma in hand we can derive linear forms of $\Pr_{\cal{X}}[P_\ell,\kk]$ that are close approximations of the two linear forms of $(x_{j,1},x_{j,2}:j\ne \ell)$ in (\[eq:desired\]). We then use the coefficients of these linear forms of $\Pr_{\cal{X}}[P_\ell,\kk]$ to perturb $\calG_{n,N}^*$ and wrap up the construction of $\calG_\AA$. Construction of Anonymous Game $\calG_\AA$ {#sec:reduction} ------------------------------------------ Let $\AA\in [0,1]^{2n\times 2n}$ denote the input polymatrix game. We need the following parameters: $$N=2^n,\ \ \delta={1}/{N}= {2^{-n}},\ \ \lambda={2^{-n^3}},\ \ \xi={2^{-n^4}},\ \ \xi^*={2^{-n^5}}\ \ \text{and}\ \ \epsilon={2^{-n^6}}.$$ We remark that we do not attempt to optimize the parameters here but rather set them in different scales to facilitate the analysis later. We use the polynomial-time algorithm promised in the Estimation Lemma to compute $\BB^{[\ell,j]}$ and $\CC^{[\ell,j]}$, for all $\ell\in [n]$ and $j\ne \ell\in [n]$. Starting with the generalized radix game ${\cal{G}}^*_{n,N}$, we modify payoff functions of players $P_1,\ldots,P_n$ as follows *(*payoff functions of $Q$ and $R$ remain unchanged*)*. Let $\emph{\pay}^*_\ell$ denote the payoff function of $P_\ell$ in $\calG_{n,N}^*$. Then for each player $P_{\ell}$ and each histogram $\kk \in K$, we set $$\begin{aligned} \emph{\pay}_{{\ell}} (s_1,\kk) &= \emph{\pay}_{{\ell}}^* (s_1,\kk) + \xi^*\sum_{j\ne \ell} N^j\left(A_{2\ell-1,2j-1}\cdot B_\kk^{[\ell,j]}+ A_{2\ell-1,2j}\cdot C_{\kk}^{[\ell,j]}\right)\\[0.5ex] \emph{\pay}_{\ell} (s_2,\kk) &= \emph{\pay}^*_{{\ell}} (s_2,\kk) + \xi^* \sum_{j\neq \ell} N^j\left(A_{2\ell, 2j - 1}\cdot B_\kk^{[\ell,j]} + A_{2\ell, 2j}\cdot C_\kk^{[\ell,j]}\right), \end{aligned}$$ and keep all other payoffs of $P_\ell$ the same *(*i.e., $\emph{\pay}_\ell(\sigma,\kk)= \emph{\pay}_\ell^*(\sigma,\kk)$ for all $\sigma\notin\{s_1,s_2\}$*)*. A few properties of $\calG_\AA$ then follow directly from its construction. First, observe that entries of $\AA$ lie in $[0,1]$ and entries of $\BB^{[\ell,j]}$ and $\CC^{[\ell,j]}$ have absolute values at most $N^{n^2}=2^{n^3}$. We have Given $\AA\in [0,1]^{2n\times 2n}$, $\cal{G}_\AA$ is an anonymous game $\xi$-close to ${\cal{G}}^*_{n,N}$ where $\xi ={2^{-n^4}}$. By Lemma \[lem:interested\], an $\epsilon$-well-supported Nash equilibrium of $\calG_\AA$ is fully described by a $(2n+2)$-tuple $\calX=(x_{i,1},x_{i,2},y,z:i\in [n])$, where $P_i$ plays strategies $s_1,s_2$ and $t$ with probabilities $x_{i,1},x_{i,2}$ and $1-x_{i,1}-x_{i,2}$, respectively. We also get the following corollary from Lemma \[lem:perturb\]. Every $\epsilon$-well-supported equilibrium $\calX =(x_{i,1},x_{i,2},y,z:i\in [n])$ of $\calG_\AA$ satisfies $$x_{i,1}+x_{i,2}=\delta^i\pm \lambda,\ \ \ \text{for all $i\in [n]$.}$$ Therefore, the conditions of the estimation lemma are met. It follows that \[simpleproperty\] Given an $\epsilon$-well-supported equilibrium $\calX$ of $\calG_\AA$, the expected payoffs of $P_\ell$ satisfy $$\begin{aligned} u_\ell(s_1,\calX)&=u_\ell^*(s_1,\calX)+\xi^*\sum_{j\ne \ell} N^j\big(A_{2\ell-1,2j-1}\cdot x_{j,1} +A_{2\ell-1,2j}\cdot x_{j,2}\big)\pm O(n^3\xi^*\delta)\ \ \ \ \text{and}\\[0.6ex] u_\ell(s_2,\calX)&=u_\ell^*(s_2,\calX)+\xi^*\sum_{j\ne \ell} N^j\big(A_{2\ell,2j-1}\cdot x_{j,1} +A_{2\ell,2j}\cdot x_{j,2}\big)\pm O(n^3\xi^*\delta).\end{aligned}$$ Correctness of the Reduction {#sec:correctness} ---------------------------- We are now ready to show that, given an $\epsilon$-well-supported Nash equilibrium $\calX$ of $\calG_\AA$, the vector $\yy$ derived from $\calX$ using (\[eq:back\]) is a $(1/n)$-well-supported Nash equilibrium of the polymatrix game $\AA$. \[reduct\_1\] Let $\calX=(x_{i,1}, x_{i,2},y,z:i\in [n])$ be an $\epsilon$-well supported Nash equilibrium of ${\cal{G}}_\AA$. Then the vector $\yy\in [0,1]^{2n}$ derived from $\calX$ using *(\[eq:back\])* is a $(1/n)$-well-supported Nash equilibrium of $\AA$. Firstly, note that $x_{i,1}+x_{i,2}>0$ so $\yy$ is well defined and satisfies $y_{2i-1}+y_{2i}=1$ for all $i$. Assume towards a contradiction that $\yy$ derived from $\calX$ using (\[eq:back\]) is not a $(1/n)$-well-supported Nash equilibrium of $\AA$, i.e., there is a player $\ell\in [n]$ such that, without loss of generality, $$\label{final11} \AA_{2\ell -1}\cdot \yy > \AA_{2\ell}\cdot \yy + 1/n$$ but $y_{2\ell}>0$, which in turn implies that $x_{\ell,2}>0$. Since $x_{j,1}+x_{j,2}=\delta^j\pm \lambda$, we have $$y_{2j-1}=\frac{ x_{j,1}}{x_{j,1}+x_{j,2}} = N^jx_{j,1} \pm\hspace{0.04cm} O(N^{2j}\lambda)=N^jx_{j,1}\pm O\hspace{0.04cm} (N^{2n}\lambda).$$ Similarly we also have $y_{2j}=N^jx_{j,2}\pm O\hspace{0.04cm}(N^{2n}\lambda)$. Combining these with Property \[simpleproperty\], we have $$\begin{aligned} \nonumber u_\ell(s_1,\calX)&=u_\ell^*(s_1,\calX)+\xi^*\cdot \AA_{2\ell-1}\cdot \yy\pm \left(O(n^3\xi^*\delta)+O(n\xi^*N^{2n}\lambda)\right)\ \ \ \ \text{and}\\[0.8ex] u_\ell(s_2,\calX)&=u_\ell^*(s_2,\calX)+\xi^*\cdot \AA_{2\ell}\cdot \yy\pm \left(O(n^3\xi^*\delta)+O(n\xi^*N^{2n}\lambda)\right).\label{final22} \\[-2.6ex] \nonumber\end{aligned}$$ By our choices of parameters, $n\xi^*N^{2n}\lambda\ll n^3\xi^*\delta$ so the former can be absorbed into the latter. Combining (\[final11\]) and (\[final22\]) (as well as the fact that $u_\ell^*(s_1,\calX)=u_\ell^*(s_2,\calX)$ because the payoffs of $s_1$ and $s_2$ are exactly the same in the generalized radix game $\calG^*_{n,N}$), we have $$u_\ell(s_1,\calX)-u_\ell(s_2,\calX)\ge \xi^*\left(\AA_{2\ell-1}\cdot \yy- \AA_{2\ell}\cdot \yy\right)-O\hspace{0.04cm}(n^3\xi^*\delta)\ge \xi^*/n-O\hspace{0.04cm}(n^3\xi^*\delta)>\epsilon,$$ for sufficiently large $n$, by our choices of parameters $\delta$, $\xi^*$ and $\epsilon$. It then follows that $x_{\ell,2}=0$, since $\calX$ is assumed to be an $\epsilon$-well-supported Nash equilibrium of $\calG_\AA$, contradicting with $y_{2\ell}>0$. This finishes the proof. Proof of the Hardness Part of Theorem \[main-theorem\] {#sec:hardness} ------------------------------------------------------ From our definitions of $\calG_{n,N}^*$ and $\calG_\AA$, it is clear that all payoffs of $\calG_\AA$ are in $[-1,3]$. Using standard arguments (invariance of Nash equilibria under shifting and scaling), we can easily see that given an anonymous game $\calG = (n,\alpha,\{\textsf{payoff}_p\})$ such that all payoffs are in the interval $[a,b]$, where $a,b \in \mathbb{R}$ and $a < b$, a mixed strategy profile $\bxx$ is an $(b-a)\epsilon$-well-supported equilibrium of $\calG$ if and only if $\bxx$ is an $\epsilon$-well-supported equilibrium of $\calG' = (n,\alpha,\{{\pay}'_p\})$, where $$\label{shifting_scaling} \pay'_p(\sigma,\kk) = \frac{\pay_p(\sigma,\kk) - a}{b-a}.$$ The new game $\calG'$ now has all payoffs from in $[0,1]$. As a result, we can construct $\calG'_\AA$ from $\calG_\AA$ in polynomial time such that all payoffs of $\calG_\AA'$ lie in $[0,1]$, and Lemma \[reduct\_1\] holds for all $(\epsilon/4)$-well-supported Nash equilibria of $\calG'_\AA$. It follows that \[before\_padding\] Fix any $\alpha\ge 7$. The problem of finding a $2^{-(n^6+2)}$-well-supported Nash equilibrium of an anonymous game with $\alpha$ actions and $[0,1]$ payoffs is [PPAD-hard]{}. This can be further strengthened using a standard padding argument. \[padding\] Fix any $\alpha\in \mathbb{N}$ and $a>b>0$. There is a polynomial-time reduction from the problem of finding a $(2^{-n^a})$-well-supported equilibrium to that of finding a $(2^{-n^b})$-well-supported equilibrium, in an anonymous game with $\alpha$ actions and $[0,1]$ payoffs. For convenience, we will refer to the problem of finding a $(2^{-n^a})$-well-supported equilibrium as problem $\textsc{A}$ and the other as problem $\textsc{B}$. Let $\calG\hspace{-0.01cm}=\hspace{-0.01cm}(n,\alpha,\{\pay_p\})$ denote an input anonymous game of problem <span style="font-variant:small-caps;">A</span>. We define a new game $\padG=(n^{t},\alpha,\{\pay_p'\})$ as follows, where $t=a/b>1$ and thus, $n^t>n$. To this end, define a map $\phi: \mathbb{Z}^{\alpha} \rightarrow \mathbb{Z}^{\alpha}$ such that $\phi(k_1,\ldots,k_{\alpha}) = (k_1 - (n^t - n), k_2,\ldots,k_{\alpha})$. We then define payoff functions of players $\{1,\ldots,n^t\}$ in $\padG$ as follows: - For each $i > n$, the payoff function of player $i$ is given by $$\text{{\textsf{payoff}}}'_{i}(\sigma,\kk)=\begin{cases} \hspace{0.06cm}1& \text{if $\sigma = 1$}\\[0.4ex] \hspace{0.06cm}0 & \text{otherwise} \end{cases}$$ So player $i$ always plays strategy $1$ in any $\epsilon$-well-supported equilibrium with $\epsilon<1$. - The payoff of each player $i\in [n]$ is given by $$\text{{\textsf{payoff}}}'_{i}(\sigma,\kk)= \begin{cases} \hspace{0.06cm} \text{{\textsf{payoff}}}_{i}(\sigma,\phi(\kk))& \text{if $k_1\ge n^t-n$}\\[0.6ex] \hspace{0.06cm} 0 & \text{otherwise} \end{cases}\vspace{0.04cm}$$ Note that in any $\epsilon$-well-supported equilibrium with $\epsilon<1$, the latter case never occurs. By the definition of $\padG$, it is easy to show that $\calX$ is an $\epsilon$-well-supported equilibrium in $\padG$, for some $\epsilon<1$, iff 1) each player $i>n$ plays strategy $1$ with probability $1$ and 2) the mixed strategy profile of the first $n$ players in $\calX$ is an $\epsilon$-well-supported equilibrium of $\calG$. As a result, a solution to $\padG$ as an input of problem <span style="font-variant:small-caps;">B</span> must be an $\epsilon$-approximate equilibrium of $\calG$ with $\epsilon=2^{-(n^t)^b}=2^{-n^a}.$ As $\padG$ can be constructed from $\calG$ in polynomial time, this finishes the proof of the lemma. Combining Corollary \[before\_padding\] and Lemma \[padding\], we have \[after\_padding\] Fix any $\alpha\ge 7$ and $c>0$. The problem of finding a $(2^{-n^c})$-well-supported Nash equilibrium in an anonymous game with $\alpha$ actions and $[0,1]$ payoffs is PPAD-hard. To prove the hardness part of Theorem \[main-theorem\], we next give a polynomial-time algorithm to compute a well-supported equilibrium from an approximate equilibrium. \[poly:equiv\] Let $\calG\hspace{-0.01cm} =\hspace{-0.01cm} (n,\alpha, \{\textsf{\emph{payoff}}_p\}\hspace{-0.01cm})$ be an anonymous game with payoffs from $[0,1]$. Given an $\epsilon^2/(16\alpha n)$-approximate Nash equilibrium $\bxx$ of $\calG$, one can compute in polynomial time an $\epsilon$-well-supported Nash equilibrium $\byy$ of $\calG$. Let $\bxx=(\xx_i:i\in [n])$ be an $\epsilon'$-approximate Nash equilibrium of $\calG$, with $\epsilon'=\epsilon^2/(16\alpha n)$. For each player $i \in [n]$, we have for any mixed strategy $\xx_i'$, $$\label{nash_def} u_i(\xx_i',\bxx_{-i}) \leq u_i(\bxx) +\epsilon',$$where we let $u_i(\xx_i',\bxx_{-i})$ denote the expected payoff of player $i$ when she plays $\xx_i'$ and other players play $\bxx_{-i}$. Let $\sigma_i$ be a strategy with the highest expected payoff for player $i$ (with respect to $\calX_{-i}$): $$u_i(\sigma_i,\bxx)=\max_{k \in[ \alpha]} \hspace{0.06cm}u_i(k,\bxx),$$ and let $J_{i} = \{j:u_i(\sigma_i,\bxx) \geq u_i(j,\bxx) + \epsilon/2\}$. We then define a mixed strategy $\yy_i$ for player $i$ using $\xx_i$, $\sigma_i$ and $J_i$ as follows: Set $y_{i,j}=0$ for all $j \in J_{i}$, and set $$y_{i,\sigma_i}=x_{i,\sigma_i} + \sum_{j \in J_{i}} x_{i,j}.$$ All other entries of $\yy_i$ are the same as $\xx_i$. As $\yy_i$ increases the expected payoff of player $i$ by at least $$({\epsilon}/{2})\cdot \sum_{j \in J_{i}} x_{i,j},$$ we have from (\[nash\_def\]) that $\sum_{j \in J_{i}} x_{i,j} \leq 2\epsilon'/\epsilon$. Repeating this for every player $i\in [n]$, we obtain a new mixed strategy profile $\byy$ (clearly $\byy$ can be computed in polynomial time given $\bxx$). We finish the proof of the lemma by showing that $\byy$ is indeed an $\epsilon$-well-supported Nash equilibrium of $\calG$. Below we write $\zeta=2\epsilon'/\epsilon$. First, by the definition of $\byy$, $|x_{i,j} - y_{i,j}| \leq \zeta$ for all $i,j$. Thus, for any pure strategy profile $\ss_{-i}$, $$\begin{aligned} &\prod_{q \neq i} y_{q,s_q} \ge \prod_{q \neq i} \max\big\{0, x_{q,s_q} - \zeta\big\} \ge \prod_{q \neq i} x_{q,s_q} -\zeta \cdot \sum_{q\ne i} \prod_{p\notin \{i,q\}} x_{p,s_p}\ \ \ \ \text{and}\\[0.5ex] &\prod_{q \neq i} x_{q,s_q} \ge \prod_{q \neq i} \max\big\{0, y_{q,s_q} - \zeta \big\} \ge \prod_{q \neq i} y_{q,s_q} -\zeta \cdot \sum_{q\ne i} \prod_{p\notin \{i,q\}} y_{p,s_p}.\end{aligned}$$ Since all payoffs are in $[0,1]$, we have for any player $i\in [n]$ and pure strategy $j\in [\alpha]$ that $$\begin{aligned} \big|u_i(j,\byy) - u_i(j,\bxx)\big| &\leq \sum_{\ss_{-i}\in [\alpha]^{n-1}}\Bigg| \hspace{0.04cm}\prod_{q \neq i} y_{q,s_q}-\prod_{q \neq i} x_{q,s_q}\hspace{0.02cm}\Bigg| \\[1ex] &\le \zeta \sum_{\ss_{-i}} \hspace{0.04cm}\sum_{q\ne i} \prod_{p\notin \{i,q\}} x_{p,s_p} +\zeta \sum_{\ss_{-i}} \hspace{0.04cm}\sum_{q\ne i} \prod_{p\notin \{i,q\}} y_{p,s_p} \\[1ex]&= \zeta \sum_{q\ne i} \hspace{0.04cm}\sum_{\ss_{-i}} \prod_{p\notin \{i,q\}} x_{p,s_p} +\zeta \sum_{q\ne i} \hspace{0.04cm}\sum_{\ss_{-i}} \prod_{p\notin \{i,q\}} y_{p,s_p} \\[0.8ex] &\le \alpha\hspace{0.02cm}\zeta \sum_{q \neq i} \left( \sum_{s_r:r\notin \{i,q\}} \hspace{0.04cm}\prod_{p \notin \{i,q\} } x_{p,s_p} \right) + \alpha\hspace{0.02cm}\zeta \sum_{q \neq i} \left(\sum_{s_r:r\notin \{i,q\}} \hspace{0.04cm}\prod_{p \notin \{i,q\} } y_{p,s_p} \right) \\[0.7ex] &=2(n-1)\alpha\hspace{0.02cm}\zeta.\end{aligned}$$ This implies that for any pure strategies $j,k \in [\alpha]$ we have $$\big|(u_i(j,\bxx) - u_i(k,\bxx)) - (u_i(j,\byy) - u_i(k,\byy))\big|< \epsilon/2.$$ Therefore, the new mixed strategy profile $\byy=(\yy_i:i\in [n])$ satisfies $$u_i(j,\byy) < u_i(k,\byy) + \epsilon\ \Rightarrow \ u_i(j,\bxx) < u_i(k,\bxx) + \epsilon/2 \ \Rightarrow\ y_{i,j} = 0$$ for all $i,j$ and $k$. This finishes the proof of the lemma. Fix any $\alpha\ge 7$ and $c>0$. It then follows from Lemma \[poly:equiv\] that the problem of finding a $(2^{-n^{c/2}})$ well-supported equilibrium in an anonymous game with $\alpha$ actions and $[0,1]$ payoffs is polynomial-time reducible to problem $(\alpha,c)$-. As the former problem is PPAD-hard by Corollary \[after\_padding\], $(\alpha,c)$- is PPAD-hard. The finishes the proof of the hardness part of Theorem \[main-theorem\]. Proof of the Estimation Lemma {#sec:estimation} ============================= We prove the estimation lemma (Lemma \[lem:estimation-lemma\]) in this section. Recall that there are $n$ main players $P_1,\ldots,P_n$, and they are only interested in three strategies $\{s_1,s_2,t\}$. For convenience we will refer to $s_1$ as strategy $1$, $s_2$ as strategy $2$, and $t$ as strategy $3$ in this section. Player $P_i$ plays strategy $b\in [3]$ with probability $x_{i,b}$, and $\sum_{b} x_{i,b}=1$. While $x_{i,b}$’s are unknown variables, by the assumption of the lemma we are guaranteed that $$\label{eq:setup} x_{i,1}+x_{i,2}=\delta^i\pm \lambda,\ \ \ \text{where $\lambda=\delta^{n^2}$.}$$ Throughout this section we will *fix* two distinct integers $r,\ell\in [n]$, and the goal will be to derive an approximation of the unknown $x_{r,1}$ for $P_\ell$ using a linear form of the following probabilities: $$\label{linearform} \Big\{\Pr\big[k_{1} = i, k_{2} = j\big]:i,j \in [0:n-1]\Big\},\ \text{ where $\Pr\big[k_{1} = i, k_{2} = j\big] = \sum_{\substack{\kk \in K \\ k_1 = i, k_2 = j}} \Pr_{\bxx}[P_\ell,\kk]$},$$ and $k_{b}$ denotes the random variable that counts players playing $b\in [3]$ other than player $P_\ell$ herself. We will show that coefficients in the desired linear form can be computed in polynomial time in $n$. First we would like to give the reader some intuition for the rest of the section, by showing how one can get a good estimate of $x_{1,1}$ and $x_{2,1}$, assuming $\ell > 2$. We believe this to be useful for more easily understanding the rest of the section, but the reader should feel free to skip it, if desired. As $N=2^n$ is large, we have $x_{i,3} \approx 1$ for each $i$. This gives $$\emph{\Pr}\big[k_{1} = 1, k_{2} = 0\big] \approx x_{1,1} + x_{2,1} + \cdots + x_{n,1} = x_{1,1}\pm O(\delta^2)$$ as $x_{i,1} \leq \delta^i+\lambda$. Similarly, $\emph{\Pr}\big[k_{1} = 2, k_{2} = 0\big] \approx x_{1,1}x_{2,1} \pm O(\delta^4)$. Using $x_{i,1}+x_{i,2}\approx \delta^i$, we have $$\emph{\Pr}\big[k_{1} = k_{2} = 1\big] \approx x_{1,1}(\delta^2 - x_{2,1}) + x_{2,1}(\delta - x_{1,1}) \pm O(\delta^4) = \delta^2x_{1,1} + \delta x_{2,1} - 2x_{1,1}x_{2,1} \pm O(\delta^4).$$ Combining all three estimates, we have $$N\Big(\emph{\Pr}\big[k_{1} =k_{2} = 1\big] + 2\cdot \emph{\Pr} \big[k_{1} = 2, k_{2} = 0\big]\Big) - \delta\cdot \emph{\Pr}\big[k_{1} = 1, k_{2} = 0\big]\approx x_{2,1} \pm O(\delta^3).$$ Since $x_{2,1}\le \delta^2+\lambda$, the linear form on the LHS gives us an additive approximation of $x_{2,1}$. We need some notation in order to generalize and formalize this. Let $\calS=[n]\setminus \{\ell\}$, the set of players observed by player $P_\ell$. Let $k_b$, $b\in [3]$, denote the random variable that counts players from $\calS$ that play strategy $b$. We write $\calL=\{i\in \calS:i\le r\}$ and $m=|\calL|$, i.e., $\calL = [r]$ and $m=r$ if $\ell > r$, and $\calL = [r] \setminus \{\ell\}$ and $m=r-1$ if $\ell < r$. We start by understanding the following probabilities $$\Big\{\Pr\big[k_{1} = m -j, k_{2} = j\big] : j \in [0:m]\Big\}.$$ It will become clear that players from $\calS\hspace{-0.03cm}\setminus\hspace{-0.03cm}\calL$ have probabilities too small to significantly affect these probabilities (so their contribution will just be absorbed into the error term). For $j\in [0:m]$, let $\Delta_j$ denote the set of partitions of $\calS$ into sets of size $m-j,j$ and $n-1-m$: $$\Delta_{j} = \Big\{(\calS_1,\calS_2,\calS_3):\text{$\calS_1,\calS_2,\calS_3$ are pairwise disjoint,}\ \calS_1 \cup \calS_2 \cup \calS_3 = \calS,\hspace{0.04cm} |\calS_1| = m - j,\hspace{0.04cm} |\calS_2| = j\Big\}.$$ So, by definition, we have $$\Pr\big[k_{1} = m -j, k_{2} = j\big] = \sum_{(\calS_1,\calS_2,\calS_3) \in \Delta_{j}} \left(\prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calS_2} x_{i,2} \prod_{i \in \calS_3} x_{i,3}\right).$$ By (\[eq:setup\]) we can write $x_{i,1}+x_{i,2}=\delta^i+\lambda_i$ for some $\lambda_i$ with $|\lambda_i|\le \lambda$. We can substitute to get $$\label{eq:11} \Pr\big[k_{1} = m -j, k_{2} = j\big] = \sum_{(\calS_1,\calS_2,\calS_3) \in \Delta_{j}}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calS_2} (\delta^i + \lambda_i - x_{i,1}) \prod_{i \in \calS_3} (1 - \delta^i - \lambda_i)\right).\vspace{0.06cm}$$ Next, we split $\Delta_j$ into two sets $\Delta_{j}^*$ and $\Delta_j'$: $(\calS_1,\calS_2,\calS_3)\in \Delta_j$ is in $\Delta_j^*$ if $\calS_1\cup\calS_2 =\calL$; otherwise, it is in $\Delta_j'$. This splits the sum in (\[eq:11\]) into two sums accordingly, one over $\Delta_j^*$ and one over $\Delta_j'$. We show in the following lemma that the contribution from the second sum is negligible. \[error1\] Given the parameters in *(\[eq:setup\])*, we have $$\sum_{(\calS_1,\calS_2,\calS_3) \in \Delta_{j}'}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calS_2} (\delta^i+\lambda_i- x_{i,1}) \prod_{i \in \calS_3} (1 - \delta^i-\lambda_i)\right) = O\bigg(\delta\prod_{i\in \calL} \delta^i\bigg).$$ Since all terms in the sum are nonnegative, it suffices to show that $$\label{eq:22} \sum_{(\calS_1,\calS_2,\calS_3) \in \Delta_{j}'}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calS_2} (\delta^i+\lambda_i- x_{i,1})\right) = O\bigg(\delta\prod_{i\in \calL} \delta^i\bigg).$$ Fix a set $\calT \subseteq \calS$ such that $|\calT| = m$ but $\calT \neq \calL$. We have $$\prod_{i\in \calT} (\delta^i+\lambda_i)= \prod_{i\in \calT} \big(x_{i,1} + (\delta^i+\lambda_i - x_{i,1})\big) = \sum_ {\calS_1 \subseteq \calT}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calT \setminus \calS_1} (\delta^i+\lambda_i - x_{i,1})\right).$$ Since every term on the RHS is nonnegative, we have $$\sum_{\substack{\calS_1 \subseteq \calT\\ |\calS_1| = m - j}}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calT \setminus \calS_1} (\delta^i+\lambda_i - x_{i,1}) \right)\leq \prod_{i \in \calT} (\delta^i+\lambda_i) =(1+o(1))\cdot \prod_{i\in \calT} \delta^i,$$ given that $\lambda_i=\delta^{n^2}$ in (\[eq:setup\]). Let $h(\calT)=\prod_{i\in \calT} \delta^i$. To prove (\[eq:22\]), it now suffices to show that $$\sum_{\substack{\calT\subseteq \calS\\ |\calT|=m,\hspace{0.03cm} \calT\ne \calL}} h(\calT)= O\big(\delta\cdot h(\calL)\big)= O\bigg(\delta\prod_{i\in \calL}\delta^i\bigg).$$ For this purpose, notice that $h(\calT)\le \delta\cdot h(\calL)$ for any $\calT$ such that $\calT\subseteq \calS$, $|\calT|=m$, but $\calT\ne \calL$. It is also easy to see that there is at most one $\calT$ such that $h(\calT)=\delta\cdot h(\calL)$. Because every other $\calT$ has $h(\calT)\le \delta^2\cdot h(\calL)$ and the total number of $\calT$’s is at most $2^{n-1}=N/2$, we have $$\sum_{\calT} h(\calT)\le \delta\cdot h(\calL)+(N/2)\cdot \delta^2\cdot h(\calL) =O(\delta\cdot h(\calL)),$$ as $\delta=1/N$. This finishes the proof of the lemma. Combining (\[eq:11\]) and Lemma \[error1\], we have $${\Pr}\big[k_{1} = m -j, k_{2} = j\big] =\hspace{-0.1cm} \sum_{\substack{\calS_1 \subseteq \calL\\ |\calS_1| = m-j}} \hspace{-0.1cm} \left(\prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calL \setminus \calS_1} (\delta^i+\lambda_i - x_{i,1}) \prod_{i \notin \calL}\hspace{0.08cm}(1 - \delta^i-\lambda_i)\right)\hspace{-0.04cm}\pm O(\delta\cdot h(\calL)).$$ The next lemma further simplifies this estimate by absorbing all the $\lambda_i$’s into the error term. \[lem:error2\] Given the parameters in *(\[eq:setup\])*, we have $$\emph{\Pr}\big[k_{1} = m -j, k_{2} = j\big] = \sum_{\substack{\calS_1\in \calL\\ |\calS_1|=m-j}}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calL\setminus \calS_1} (\delta^i- x_{i,1}) \prod_{i \notin \calL} (1 - \delta^i)\right) \pm O(\delta\cdot h(\calL)).$$ First the number of $\calS_1$’s is at most $2^{n-1}<N$. Further, fixing an $\calS_1$ and multiplying out $$\prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calL\setminus\calS_1} (\delta^i + \lambda_i - x_{i,1})\prod_{i \notin\calL} (1 - \delta^i - \lambda_i)$$ will yield $3^j\cdot 3^{n-1-m} \le 3^{n-1}< N^2$ many terms. The absolute value of each term with at least one $\lambda_i$ must be less than or equal to $\lambda$ because all factors are less than or equal to $1$. There are at most $N^2$ many such terms, for each $\calS_1$, and there are at most $N$ different $\calS_1$’s. Using $N^3\lambda\ll \delta h(\calL)$ by (\[eq:setup\]), we can absorb all terms with at least one $\lambda_i$ into the error term $O\hspace{0.02cm}(\delta\cdot h(\calL))$. Using Lemma \[lem:error2\] and the fact that $\prod_{i\notin \calL}(1-\delta^i)>1/2$ as $\delta=1/2^n$, we have $$\left(\prod_{i \notin \calL} (1 - \delta^i)\right)^{-1} \Pr\big[k_{1} = m - j, k_{2} = j\big] = \sum_{\substack{\calS_1 \subseteq \calL\\ |\calS_1| = m-j}} \left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calL \setminus \calS_1} (\delta^i - x_{i,1})\right)\pm O(\delta\cdot h(\calL)).$$ To understand the RHS better, we define a polynomial $P_d^m$ for each $d\in [0:m]$ to be $$P_d^{m} = \sum_{\calT \subseteq \calL,\hspace{0.04cm} |\calT| = d} \left(\prod_{i \in \calT} x_{i,1} \prod_{i \in \calL \setminus \calT} \delta^i\right),$$ and prove the following lemma that establishes a connection between them. \[lem:lin\_eq\] Given $P_d^m$ defined above, we have $$\label{compare} \sum_{\substack{\calS_1 \subseteq \calL\\ |\calS_1| = m-j}}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calL \setminus \calS_1} (\delta^i - x_{i,1})\right) = \sum_{i = 0}^j \hspace{0.05cm}(-1)^i \cdot {m - j + i \choose m - j} \cdot P_{m - j + i}^{m}$$ Note that every monomial that appears on the two sides of (\[compare\]) has the form $\prod_{i\in \calT} x_{i,1}$ for some $\calT \subseteq \calL$ with $|\calT| =d\ge m - j$. Fix such a $\calT$. The coefficient of $\prod_{i\in \calT} x_{i,1}$ on RHS of (\[compare\]) is $$(-1)^{d-m+j}\cdot {d\choose m-j}\cdot \prod_{i\in \calL\setminus \calT}\delta^i.$$ On the other hand, for an $\calS_1\subseteq \calL$ with $|\calS_1|=m-j$, we have $$\prod_{i \in \calS_1} x_{i,1} \prod_{i \in\calL \setminus \calS_1} (\delta^i - x_{i,1}) = \sum_{\calS' \subseteq \calL \setminus \calS_1}\left( \prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calS'} (-x_{i,1}) \prod_{i \in \calL \setminus \{\calS_1 \cup \calS'\}} \delta^i \right).$$ Hence, $\prod_{i \in \calT} x_{i,1} $ occurs exactly once in this sum if and only if $\calS_1 \subseteq \calT$, and will take the form $$\prod_{i \in \calS_1} x_{i,1} \prod_{i \in \calT \setminus \calS_1} (-x_{i,1})\prod_{i \in \calL \setminus \calT} \delta^i = (-1)^{d-m+j}\prod_{i \in \calT} x_{i,1} \prod_{i \in \calL\setminus \calT}\delta^i.$$ Further, there are ${d \choose m - j}$ many $\calS_1$ such that $\calS_1 \subseteq \calT$ and $|\calS_1|=m-j$. The lemma is proven. Combining Lemma \[lem:error2\] and \[lem:lin\_eq\], we immediately get the following corollary: For any $j\in [0:m]$, we have \[coro:lin\_eq\] $$\left(\prod_{i \notin \calL} (1 - \delta^i)\right)^{-1} \emph{\Pr}\big[k_{1} = m - j, k_{2} = j\big] = \sum_{i = 0}^j \hspace{0.05cm}(-1)^i \cdot {m - j + i \choose m - j} \cdot P_{m- j + i}^{m} \pm O\big(\delta\cdot h(\calL)\big).$$ Taking a step back, we have derived a set of linear equations that hold with high precision over $\Pr[k_{1} = m, k_{2} = 0],\ldots,\Pr[k_{1} = 0, k_{2} = m]$ and $P_m^{m},\ldots,P_0^{m}$. This then allows us to attain a close approximation for $P_1^{m}$, using a linear form of the $m$ probabilities. Note that $$\label{final2} P_1^{m} = \sum_{i \in \calL}\hspace{0.03cm} x_{i,1} \prod_{j \in \calL \setminus \{i\}} \delta^j =h(\calL)\cdot \sum_{i\in \calL} N^i\cdot x_{i,1}$$ is a linear form of the $x_{i,1}$’s, $i\in \calL$, including $x_{r,1}$ (recall that $r$ is the largest integer in $\calL$). So from here, it will be straightforward to get an approximation of $x_{r,1}$. The next lemma gives us a linear form to approximate $P_1^m$. \[lem:inverse\] The $m$ probabilities and $P_1^m$ satisfy $$\left(\prod_{i \notin \calL} (1-\delta^i)\right)^{-1} \hspace{0.1cm} \sum_{j=1}^{m} j\cdot {\emph{\Pr}}\big[k_{1} = j, k_{2} = m-j\big] = P_1^{m} \pm O\big(m^2\delta\cdot h(\calL)\big).$$ By Corollary \[coro:lin\_eq\] (and replacing $j$ init by $m-j$), we see that it suffices to show that $$\label{eq:uu} P_1^{m} = \sum_{j=1}^{m} j\cdot \left(\sum_{i=0}^{m-j} (-1)^i\cdot {j + i \choose j}\cdot P_{j+i}^{m}\right).$$ Consider $P_d^{m}$ for some $d \in [m]$. $P_d^{m}$ appears in the $j$th term on the RHS of (\[eq:uu\]) if and only if $d \geq j$, and when this is the case, the coefficient of $P_d^m$ is $${ j}\cdot (-1)^{d - j} \cdot {d \choose j}.\vspace{-0.15cm}$$ So the RHS of (\[eq:uu\]) is $$\sum_{d=1}^m P_d^{m}\cdot \left( \sum_{j=1}^d (-1)^{d-j}\cdot j\cdot { d \choose j}\right).$$ For $d = 1$, the coefficient of $P_1^m$ is clearly $1$. For $d>1$, using $j {d \choose j } = d {d - 1 \choose j - 1}$ we have $$\sum_{j=1}^d (-1)^{d-j}\cdot { j }\cdot { d \choose j} = d\cdot \sum_{j=1}^d (-1)^{d-j}\cdot {d - 1 \choose j - 1} = d\cdot \sum_{j=0}^{d-1} (-1)^{d-j- 1} { d - 1 \choose j} = 0. \vspace{0.06cm}$$ This finishes the proof of the lemma. Lemma \[lem:inverse\] gives us a linear form to approximate $P_1^m$. Denote this linear form by $Y_m$. Then for the special case when $\calL=\{r\}$ (so $r$ is the only integer in $\calL$), we are done since $P_1^m$ is exactly $x_{r,1}$, and we have attained a linear form that approximates $x_{r,1}$ with error $O(m^2\delta\cdot h(\calL))$. Otherwise suppose $|\calL|>1$. We use $r'$ to denote the largest integer in $\calL$ other than $r$ and write $\calL'=\{i\in \calS:i\le r'\}$ ($|\calL'|=m-1$). Repeating the same line of proof so far over $\calL'$ and $m-1$, we obtain a linear form of $\Pr [k_1=m-1-j,k_2=j]$, $j\in [0:m-1]$, denoted by $Y_{m-1}$, to approximate $$\label{final1} P_1^{m-1}=\sum_{i\in \calL'} x_{i,1}\prod_{j\in \calL'\setminus \{i\}} \delta^j=h(\calL')\cdot \sum_{i\in \calL'} N^i\cdot x_{i,1}$$ with error $O(m^2\delta\cdot h(\calL'))$. By the definition of $P_1^m$ and $P_1^{m-1}$ in (\[final2\]) and (\[final1\]), we have $$x_{r,1}=\delta^r\left(\frac{P_1^m}{h(\calL)}-\frac{P_1^{m-1}}{h(\calL')}\right).$$ As a result, we have obtained a linear form $$\label{fffff} \delta^r\left(\frac{Y_m}{h(\calL)}-\frac{Y_{m-1}}{h(\calL')}\right) =x_{r,1}\pm O\hspace{0.03cm}(m^2\delta^{r+1})$$ over $\Pr[k_{1} = m - j, k_{2} = j]$, $j \in [0:m]$ and $\Pr[k_{,1} = m - 1 - j, k_{2} = j]$, $j \in [0:m-1]$. Finally, it follows easily from our derivation of $Y_m$ and $Y_{m-1}$ that coefficients of this linear form can be computed in polynomial time in $n$, and every coefficient has absolute value at most $N^{m^2}$. å Membership in PPAD ================== In this section we show that $(\alpha,c)$- is in PPAD for any constants $\alpha\in \mathbb{N}$ and $c>0$, i.e. the problem of finding an $\epsilon$-approximate equilibrium in an anonymous game $\calG = (n,\alpha,\{\pay_p\})$ with payoffs from $[0,1]$ is in PPAD, where $\epsilon=1/2^{n^c}$. Below we use $\size(\calG)$ to denote the input size of an anonymous game $\calG$, i.e., length of the binary representation of $\calG$. We write $\size(a)$ to denote the length of the binary representation of a rational number $a$, and let $\size(\aa)=\sum_{i} \size(a_i)$ for a rational vector $\aa$ (e.g., a rational mixed strategy profile). Fix constants $\alpha\in \mathbb{N}$ and $c>0$. We show the membership of $(\alpha,c)$- by reducing it to a “weak-approximation” fixed point problem [@FIXP] (see [@FIXP] for the difference between *weak* and *strong* approximations). Given $\calG=(n,\alpha,\{\pay_p\})$, we define a map $F:\Delta\rightarrow \Delta$ (this is the map commonly used to prove the existence of Nash equilibria, e.g., see [@NASH51]), where $$\Delta=\Big\{(\xx_i:i\in [n]): \xx_i\in \mathbb{R}_+^\alpha\ \text{is a mixed strategy of player $i\in [n]$}\Big\}$$ is the set of all mixed strategy profiles. For each $i \in [n]$ and $j \in [\alpha]$, the $(i,j)$th component of $F$ $$\label{defF} F_{i,j} (\bxx) = \frac{x_{i,j} + \max{(0, u_i(j,\bxx) - u_i(\bxx))}}{1 + \sum_{k \in [\alpha]}\max{(0, u_i(k,\bxx) - u_i(\bxx))}},$$ where $\calX=(\xx_i:i\in [n])\in \Delta$ and $\xx_i=(x_{i,1},\ldots,x_{i,\alpha})$ for each $i\in [n]$. Observe that $F$ is continuous and maps $\Delta$ to itself. We also have \[poly-comp\] The map $F$ defined above is *polynomial-time computable*: Given a rational $\bxx\in \Delta$, $F(\bxx)$ is rational and can be computed in polynomial time in $\size(\calG)$ and $\size(\bxx)$. This follows from the fact that there is a polynomial-time dynamic programming algorithm (see [@dask14]) that computes $u_i(j,\bxx)$, given $\calG$ and $\calX$. We say $\bxx\in \Delta$ is an *$\epsilon$-approximate fixed point* of $F$ if $\|F(\bxx)-\bxx\|_{\infty}\le \epsilon$. We prove Lemma \[lem:fixedpoint\] in Section \[sec:fixedpoint\], showing that approximate fixed points of $F$ are approximate Nash equilibria of $\calG$. \[lem:fixedpoint\] Given $\bxx \in \Delta$ and $0 \leq \epsilon \le 1$, if $\|F(\bxx) - \bxx\|_\infty \leq \epsilon$, then we have $u_i(j,\bxx) \leq u_i(\bxx) + \epsilon'$ for all players $i \in [n]$ and pure strategies $j \in [\alpha]$, where $\epsilon' = \alpha^2\epsilon^{1/3}$. So to find an $\epsilon$-approximate Nash equilibrium $\calX$ of $\calG$, it suffices to find an $(\epsilon^3/\alpha^6)$-approximate fixed point of $F$. Moreover, we show in Section \[sec:Lipschitz\] that $F$ is *polynomially Lipschitz continuous*: \[poly-cont\] For all $\bxx , \byy\in \Delta$, we have $$\|F(\bxx) - F(\byy)\|_\infty \leq 10\hspace{0.03cm}n \hspace{0.03cm}\alpha^{n+2}\cdot \| \bxx - \byy \|_\infty.$$ Combining Property \[poly-comp\] and Lemma \[poly-cont\], it follows from Proposition 2.2 (Part 2) of [@FIXP] that given $\calG$ and $\epsilon$ (in binary), the problem of finding an $\epsilon$-approximate fixed point $\bxx$ of $F$ is in PPAD. The PPAD membership of $(\alpha,c)$-  then follows from Lemma \[lem:fixedpoint\]. Proof of Lemma \[lem:fixedpoint\] {#sec:fixedpoint} --------------------------------- For convenience, we write $\max_{i,k}(\bxx) = \max{(0, u_i(k,\bxx) - u_i(\bxx))}$ for $i\in [n]$ and $k\in [\alpha]$. In the pursuit of a contradiction, assume that there exist a player $i\in [n]$ and an action $\ell\in [\alpha]$ such that $u_i(\ell,\bxx)> u_i(\bxx) + \epsilon'$. This, along with the fact that $\max_{i,k}(\bxx) \in [0,1]$, implies that, $$\begin{aligned} \label{sum:ineq} &\epsilon' < \sum_{k \in [\alpha]} \max_{i,k}(\bxx) \leq \alpha-1. &\end{aligned}$$ We will show that cases $x_{i,\ell} \leq \alpha\hspace{0.02cm}\epsilon^{1/3}$ and $x_{i,\ell} >\alpha\hspace{0.02cm}\epsilon^{1/3}$ both result in the existence of a strategy $j \in [\alpha]$ such that $|F_{i,j}(\bxx) - x_{i,j}| > \epsilon$, contradicting our initial assumption. 1. [Case 1]{}: $x_{i,\ell} \leq \alpha\hspace{0.02cm}\epsilon^{1/3}$. Apply (\[defF\]), (\[sum:ineq\]) and $\epsilon'=\alpha^2\epsilon^{1/3}$ to get $$F_{i,\ell}(\bxx) > \frac{x_{i,\ell} + \epsilon'}{\alpha}\ \Rightarrow\ F_{i,\ell}(\bxx) - x_{i,\ell} > \frac{\epsilon' - (\alpha - 1) x_{i,\ell}}{\alpha} \geq \frac{\epsilon' - (\alpha - 1) \alpha\hspace{0.02cm}\epsilon^{1/3}}{\alpha} = \epsilon^{1/3}\ge \epsilon.$$ 2. [Case 2]{}: $x_{i,\ell} > \alpha\hspace{0.02cm}\epsilon^{1/3}$. Let $J = \{j\in [\alpha]:u_i(j,\bxx) \leq u_i(\bxx)\}$. We must have $$\begin{aligned} &\sum_{j \in J} x_{i,j}\big(u_i(\bxx) - u_i(j,\bxx)\big) \geq x_{i,\ell}\big(u_i(\ell,\bxx) - u_i(\bxx)\big),&\end{aligned}$$ where $u_i(\bxx) - u_i(j,\bxx) \leq 1 - \epsilon'$, $u_i(\ell,\bxx) - u_i(\bxx) \geq \epsilon'$, and $x_{i,\ell} > \alpha\hspace{0.02cm}\epsilon^{1/3}$. Therefore, $$\begin{aligned} &\sum_{j \in J} x_{i,j}&\hspace{-0.22cm} \geq \frac{\alpha\hspace{0.02cm}\epsilon'\epsilon^{1/3}}{1 - \epsilon'},\end{aligned}$$ which implies that there exists some strategy $j \in J$ such that $x_{i,j} \geq \epsilon'\epsilon^{1/3}/(1 - \epsilon')$. Apply (\[defF\]) and (\[sum:ineq\]) to get $F_{i,j}(\bxx) < x_{i,j}/(1 + \epsilon')$, which implies that $$\big|F_{i,j}(\bxx) - x_{i,j}\big| > \frac{\epsilon' x_{i,j}}{1 + \epsilon'} \geq \frac{(\epsilon')^2\epsilon^{1/3}}{ (1 - \epsilon')(1 + \epsilon')} \geq \alpha^4 \epsilon\geq \epsilon.$$ This finishes the proof of Lemma \[lem:fixedpoint\]. Proof of Lemma \[poly-cont\] {#sec:Lipschitz} ---------------------------- As $\bxx - \byy$ is of length $n\hspace{0.02cm}\alpha$, we have $\|\bxx-\byy\|_1\le n\hspace{0.02cm}\alpha\cdot \|\bxx-\byy\|_\infty$. Thus, it suffices to show that $$\| F(\bxx) - F(\byy)\|_\infty \leq 16\hspace{0.02cm}\alpha^{n+1}\| \bxx - \byy\|_1.$$ Fix $i\in [n]$ and $j\in [\alpha]$. We have $$\big|\hspace{0.02cm} F_{i,j}(\bxx) - F_{i,j}(\byy) \hspace{0.02cm}\big| = \left|\hspace{0.06cm} \frac{x_{i,j} + \max_{i,j}(\bxx)}{1 + \sum_{k \in [\alpha]}\max_{i,k}(\bxx)} - \frac{y_{i,j} + \max_{i,j}(\byy)}{1 + \sum_{k \in [\alpha]}\max_{i,k}(\byy)} \hspace{0.06cm}\right|. \vspace{0.04cm}$$ Multiplying the terms in the RHS to get a common denominator, which is clearly $ \geq 1$, we get $$\begin{aligned} \label{main-bound} \big|\hspace{0.02cm} F_{i,j}(\bxx) - F_{i,j}(\byy) \hspace{0.02cm}\big| &\leq \big|x_{i,j} - y_{i,j}\big| + \Bigg|\hspace{0.04cm} x_{i,j} \sum_{k \in [\alpha]}\max_{i,k}(\byy) - y_{i,j} \sum_{k \in [\alpha]}\max_{i,k}(\bxx)\hspace{0.04cm} \Bigg| \\[0.5ex] & \ \ \ + \Big| \max_{i,j}(\bxx) - \max_{i,j}(\byy)\hspace{0.02cm} \Big| + \Bigg|\hspace{0.01cm} \max_{i,j}(\bxx) \sum_{k \in [\alpha]}\max_{i,k}(\byy) - \max_{i,j}(\byy) \sum_{k \in [\alpha]}\max_{i,k}(\bxx) \hspace{0.04cm}\Bigg|. \nonumber\end{aligned}$$ To bound $| F_{i,j}(\bxx) - F_{i,j}(\byy)|$, we shall use the following simple trick several times in the rest of the proof. If $a_1,a_2,b_1,b_2 \in [0,1]$, then we have $$|\hspace{0.02cm}a_1a_2 - b_1b_2\hspace{0.02cm}| = |\hspace{0.02cm}(a_1 - b_1)a_2 + b_1(a_2 - b_2)\hspace{0.02cm}| \leq |\hspace{0.02cm}a_1 - b_1\hspace{0.02cm}| + |\hspace{0.02cm}a_2 - b_2\hspace{0.02cm}|,$$ which easily extends to $$|\hspace{0.02cm}a_1 \cdots a_n - b_1 \cdots b_n\hspace{0.02cm}| \leq |\hspace{0.02cm}a_1 - b_1\hspace{0.02cm}| + \cdots + |\hspace{0.02cm}a_n - b_n\hspace{0.02cm}|,$$ when all the $a_i$’s and $b_i$’s are in $[0,1]$. Now we come back to (\[main-bound\]). By the definition of $\max_{i,j}(\bxx)$, we have $$\begin{aligned} \Big| \max_{i,j}(\bxx) - \max_{i,j}(\byy)\hspace{0.02cm} \Big| &\leq \big|(u_i(j,\bxx) - u_i(\bxx)) - (u_i(j,\byy) - u_i(\byy))\big|\\ &\leq \big|u_i(j,\bxx) - u_i(j,\byy)\big| + \big|u_i(\bxx) - u_i(\byy)\big|.\end{aligned}$$ As $\bxx,\byy\in \Delta$ we have $x_{i,j},y_{i,j} \in [0,1]$. Since all payoffs of $\calG$ are in $[0,1]$, we have $u_i(j,\bxx),u_i(j,\byy)$, $u_i(\bxx),u_i(\byy) \in [0,1]$ for all $i,j$, which in turn implies that $\max_{i,j}(\bxx),\max_{i,j}(\byy) \in [0,1]$. Using these properties above, along with the trick, we can conclude $$\begin{aligned} \Bigg|\hspace{0.04cm} x_{i,j} \sum_{k \in [\alpha]}\max_{i,k}(\byy) &- y_{i,j} \sum_{k \in [\alpha]}\max_{i,k}(\bxx)\hspace{0.04cm} \Bigg| \leq \sum_{k \in [\alpha]}\Big|\hspace{0.03cm}x_{i,j}\cdot \max_{i,k}(\byy) - y_{i,j}\cdot \max_{i,k}(\bxx)\hspace{0.03cm}\Big| \\[0.8ex] &\leq \sum_{k\in [\alpha]}\Big(|x_{i,j} - y_{i,j}| + |u_i(k,\bxx) - u_i(k,\byy)| +|u_i(\bxx) - u_i(\byy)|\Big).\end{aligned}$$ Similarly, we also have $$\begin{aligned} &\Bigg|\max_{i,j}(\bxx) \sum_{k \in [\alpha]}\max_{i,k}(\byy) - \max_{i,j}(\byy) \sum_{k \in [\alpha]}\max_{i,k}(\bxx)\hspace{0.04cm} \Bigg|\\[0.8ex] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \leq \sum_{k\in [\alpha]}\Big(|u_i(j,\bxx) - u_i(j,\byy)| + 2|u_i(\bxx) - u_i(\byy)| + |u_i(k,\bxx) - u_i(k,\byy)| \Big).\end{aligned}$$ Plugging all these back into (\[main-bound\]), we have $$\begin{aligned} \big| F_{i,j}(\bxx) - F_{i,j}(\byy) \big|&\leq (1 + \alpha)\cdot |\hspace{0.02cm}x_{i,j}-y_{i,j}\hspace{0.02cm}| + (1+3\alpha)\cdot |\hspace{0.02cm}u_i(\bxx) - u_i(\byy)\hspace{0.02cm}|\\&\ \ \ \ \ \ + (1+\alpha)\cdot|\hspace{0.02cm}u_i(j,\bxx) - u_i(j,\byy)\hspace{0.02cm}|+ 2\cdot \sum_{k \in [\alpha]}|\hspace{0.02cm}u_i(k,\bxx) - u_i(k,\byy)\hspace{0.02cm}|.\end{aligned}$$ Finally, we bound $|u_i(k,\bxx) - u_i(k,\byy)|$ in terms of $\|\bxx - \byy\|_1$. Let $S$ be the set of pure strategy profiles. Then, by applying the trick and the fact that all payoffs are in $[0,1]$, it follows that $$\begin{aligned} |\hspace{0.02cm}u_i(k,\bxx) - u_i(k,\byy)\hspace{0.02cm}| &\leq \sum_{\ss \in S_{-i}} \Bigg|\prod_{q \neq i} x_{q,s_q} - \prod_{q \neq i} y_{q,s_q}\Bigg| \leq \sum_{\ss \in S_{-i}} \sum_{q \neq i} |\hspace{0.02cm}x_{q,s_q} - y_{q,s_q}\hspace{0.02cm}| \leq \alpha^{n-1}\| \bxx - \byy\|_1 \\[1ex] |\hspace{0.02cm} u_i(\bxx) - u_i(\byy)\hspace{0.02cm} | &\leq \sum_{\ss \in S} \Bigg|\prod_{q\in [n]} x_{q,s_q} - \prod_{q \in [n]} y_{q,s_q}\Bigg| \leq \sum_{\ss \in S} \sum_{q \in [n]} |\hspace{0.02cm}x_{q,s_q} - y_{q,s_q}\hspace{0.02cm}| \leq \alpha^n \| \bxx - \byy \|_1.\\[-2.7ex]\end{aligned}$$ Applying these inequalities, along with $|\hspace{0.02cm}x_{i,j} - y_{i,j}\hspace{0.02cm}| \leq \|\bxx - \byy\|_1$, we get $$\big|F_{i,j}(\bxx) - F_{i,j}(\byy)\big| \leq 10\hspace{0.02cm} \alpha^{n+1}\cdot \| \bxx - \byy \|_1.$$ This finishes the proof Lemma \[poly-cont\]. Open Problems ============= Can the number of strategies be further reduced from seven in our PPAD-hardness result? Specifically, could we construct an anonymous game similar to the radix game $\calG_{n,N}$, particularly its set of approximate Nash equilibria after perturbation, but without the four special (auxiliary) pure strategies $\{q_1,q_2,r_1,r_2\}$? While we believe this to be possible, constructing such a game can be highly non-trivial and would require specifying different payoffs for many of the possible outcomes seen by each player. Accordingly, proving a result similar to Lemma \[lem:perturb\] after duplicating the first strategy would be even more difficult. However, even the construction of such a game would only reduce the number of strategies used in the hardness proof down to three (due to the strategy duplication in the generalized radix game later), leading to the next open question: Is there an FPTAS for two-strategy anonymous games? As was posited by Daskalakis and Papadimitriou, it remains unclear whether a rational two-strategy anonymous game always has a rational Nash equilibrium. Additionally, in their sequence of paper’s proving a PTAS for a bounded number of strategies, Daskalakis and Papadimitriou found that the form of the $\Pr_\bxx[p,\kk]$ is significantly simpler for two-strategy anonymous games. Correspondingly, we found that constructing useful gadgets for reductions with just two strategies to be very difficult, suggesting that an FPTAS for two-strategy anonymous games is certainly a possibility. Moreover, could there be an FPTAS for anonymous games with any bounded number of pure strategies? There is no clear way to strengthen our current construction to obtain a PPAD-hardness result for ${1}/{\text{poly}(n)}$-approximate Nash equilibrium. In order for the estimation lemma to hold, we need $x_{i,1} + x_{i,2} \approx \delta^i$ for all $i$. So even if we set $N = 2$, ensuring that $x_{i,1} + x_{i,2} = \delta^i \pm O(1/\text{poly}(n))$ would still not be sufficient for the estimation lemma to hold. Accordingly, in order to modify our construction to get such a hardness result, we would need to construct an anonymous game, which contains $n$ players with the same properties as the main players in the generalized radix game, but with the additional property that $O(1/\text{poly}(n))$ shifts in the payoffs would only cause $O(1/2^{\text{poly}(n)})$ shifts in $x_{i,1} + x_{i,2}$, which seems incredibly unlikely. [^1]: When the number of pure strategies is a sufficiently large constant, an anonymous game with rational payoffs may not have any rational equilibrium (e.g., by embedding in it a rational three-player game with no rational equilibrium). But for the case of two strategies, it remains unclear as whether every rational anonymous game has a rational Nash equilibrium, which was posed as an open problem in [@dask14]. [^2]: Since we are interested in the additive approximation, all payoffs are normalized to take values in $[0,1]$. [^3]: These PLS-hardness results have no implication to the setup of this paper since the number of pure strategies in the congestion games considered there are unbounded. [^4]: For example, we can rule out the existence of an anonymous game $G_{=}$ with 4 players and 2 pure strategies such that $\xx$ is a Nash equilibrium of $G_{=}$ if and only if $x_{1}=x_{2} \in [\mu,\nu] \subseteq [0,1]$ and $x_{3}=x_{4} \in [\mu',\nu'] \subseteq [0,1]$, where we use $x_i$ to denote the probability that player $i$ plays the first pure strategy. [^5]: For example, as it is pointed out in [@PTAS07; @cPTAS08] for anonymous games with two strategies, players can always be partitioned into a few sets such that the probabilities $\Pr_{\cal{X}}[p,\kk]$ over $\kk$ must follow approximately a Poisson or a discretized Normal distribution on each set respectively.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We identify a new four-wave mixing process in which two nearly collinear pump beams produce phase-dependent gain into a weak bisector signal beam in a self-defocusing Kerr medium. Phase matching is achieved by weak-wave advancement caused by cross-phase modulation between the pump and signal beams. We relate this process to the inverse of spatial modulational instability and suggest a time-domain analog.' author: - 'C. F. McCormick' - 'R. Y. Chiao' - 'J. M. Hickmann' title: 'Weak-wave advancement in nearly collinear four-wave mixing' --- Modulational instability (MI) is a common and important nonlinear effect in which a continuous wave breaks up into periodic, localized wave packets. Its long-term dynamics can lead to Fermi-Pasta-Ulam (FPU) recurrence, in which a nonlinear system returns to its initial energy configuration rather than thermalizing[@Fermi1965]. Interest in MI has increased since the recent experimental observation of FPU recurrence of nonlinear optical pulses in optical fibers[@VanSimaeys2001; @VanSimaeys2002; @Akhmediev2001]. MI in the spatial domain is known as beam filamentation and is a four-wave mixing process in which the phase matching is provided by cross-phase modulation[@Agrawal1990; @hickmann1992; @Boyd1999]. Recent experimental work has demonstrated the existence of a connected spatial effect, photon-photon scattering in the presence of a Kerr nonlinearity[@Mitchell2000]. In this experiment, two counter-propagating beams “collide” in a rubidium vapor cell and photons scattered at $\pm 90^{\circ}$ from the beam axis are observed to be correlated in time, while photons at other angles are not. These results suggest a momentum-conserving photon-photon scattering process mediated by the Kerr nonlinearity. In this paper we discuss a related process, in which the two “colliding” beams are in a frequency-degenerate, nearly-collinear configuration. Normally this configuration generates high-order diffracted beams, used for example in dephasing time measurements. We predict that a different kind of four-wave mixing process should be possible with this geometry, in which two photons of the noncollinear pump beams are absorbed and two collinear signal photons are generated in the direction of a bisector line between the pump beams. We demonstrate that through a phenomenon similar to *weak-wave retardation* [@Chiao1966], phase matching is possible even if all four waves have the same frequency. We conclude by identifying this process as the inverse of the process that generates standard beam filamentation. We begin by calculating the nonlinear polarization component in the direction of the signal beam, $P_{4}$. For simplicity we assume that this process will be stimulated, introducing a small seed field $E_{3}$ in the direction of the bisector line (see Fig. \[configuration\]). The general expression for the polarization of the signal field is $$\begin{aligned} P_{4} &=& 3\chi_{xxxx}^{(3)} \{ |E_{4}|^{2}E_{4} + 2(|E_{1}|^{2} + |E_{2}|^{2} + |E_{3}|^{2})E_{4} \nonumber \\ && + E_{1}E_{2}E_{4}^{\ast } \exp (i(\vec{k}_{1}+\vec{k}_{2}-\vec{k}_{3}-\vec{k}_{4}) \cdot \vec{z} - i(\omega_{1}+\omega_{2}-\omega_{3}-\omega_{4})t) \} \label{polarization}\end{aligned}$$ where $\chi _{xxxx}^{(3)}$ is a component of the third-order nonlinear susceptibility tensor and $E_{j}$ are the (complex) fields associated with the pump ($j=1,2$) and signal ($j=3,4$) beams. The first term in this equation represents self-phase modulation (SPM), the second cross-phase modulation (XPM) and the third four-wave mixing. For a weak signal beam, SPM is negligible. The XPM terms from the pump beams are large and play an important part in the four-wave mixing process, while the XPM term from the signal beam may be neglected. The exponential four-wave mixing term indicates that in order for this polarization to efficiently transfer energy into the signal field $E_{4}$, there must be energy conservation and phase matching between the different wave vectors associated with each field. ![Wave vector configuration for weak-wave advancement.[]{data-label="configuration"}](configuration.eps) In the degenerate case of equal frequencies, this phase matching appears difficult to achieve. However, the signal wave experiences a decrease in the refractive index generated by the XPM term associated with the pump beams, assuming a negative nonlinear Kerr coefficient (the self-defocusing case). This difference in the refractive index is given by $$\Delta n=-\frac{3}{4n_{0}}|Re[\chi_{xxxx}^{(3)}]|(|E_{1}|^{2}+|E_{2}|^{2})$$ where $n_{0}$ is the linear index of refraction. Consequently, there exists a shortening of the weak-wave momentum vector by $\Delta k=-|\Delta n|\omega_{0}/c$, an effect which we call *weak-wave advancement* in analogy with “weak-wave retardation” in the case of the self-focusing sign of the Kerr nonlinearity[@Chiao1966]. For a certain parameter range, weak-wave advancement allows phase matching between the four wave vectors. In this range, this four-wave mixing process should transfer energy to the weak signal beam as efficiently as its counterpart. We now take the special case of degenerate signal fields $E_{3}$ and $E_{4}$. In this case, we use Eq. \[polarization\] to calculate the evolution equation for $E_{4}$ in the paraxial and plane wave limits, finding $$\frac{\partial E_{4}}{\partial z} = i4\gamma P (E_{4} + 0.5 E^{*}_{4} e^{i(6\gamma P - \Delta k)z}e^{i(\phi_{1} + \phi_{2})}) \label{evolution}$$ where $\gamma \equiv 6\pi |\vec{k}_1|\chi_{xxxx}^{(3)}/n_{0}^{2}$, $P$ is the pump beam power, $\Delta k \equiv |\vec{k}_{1} + \vec{k}_{2} - 2\vec{k}_{4}|$ is the momentum mismatch and $\phi_{1}, \phi_{2}$ are the phases of the pump beams. Taking the self-defocusing case ($\gamma P < 0$) we guess a solution to Eq. \[evolution\] given by $$E_{4} = A e^{gz} e^{i(3\gamma P - \Delta k/2)z} e^{i \phi_{4}} \label{trialsolution}$$ where $A$ is the signal field amplitude, $g$ is a purely real exponential gain coefficient and $\phi_{4}$ is the signal beam phase (for a similar treatment, see reference [@Agrawal2001], p. 392). Since the phase $\phi_{4}$ is a free parameter, Eq. \[trialsolution\] is a solution of Eq. \[evolution\] when the momentum mismatch is in the range $$0 \leq \Delta k \leq 6 |\gamma P| \label{range}$$ with an exponential gain $g$ given by $$g = \sqrt{3(\gamma P)^{2} - (\gamma P)\Delta k - 0.25(\Delta k)^{2}} \ .$$ The signal beam gain is plotted against the collision half-angle in Fig. \[gain\] for several values of $\gamma P/k$. The maximum gain $g = 2|\gamma P|$ occurs when $\Delta k = 2|\gamma P|$. ![Signal beam gain as a function of collision half-angle $\alpha$ for $\lambda = 700$nm. Note that $\Delta k = 2|\vec{k}_{1}| (1 - \cos \alpha)$.[]{data-label="gain"}](gain.eps) The geometry considered here has an important difference from the geometry of spatial modulational instability. Unlike the case of beam filamentation[@Chiao1966B; @Campillo1973], the relative phase of the signal and pump beams is under experimental control. For the optimal choice of $\Delta k$ as above, the signal beam gain is maximized for the phase relation $$\Delta \phi \equiv \phi_{1} + \phi_{2} - 2 \phi_{4} = \pi/2 \label{phase}$$ For other relative phases the signal beam initially experiences either gain ($0 < \Delta \phi < \pi$) or loss ($\pi < \Delta \phi < 2\pi$). However, we find from numerical solutions to Eq. \[evolution\] that as the signal beam propagates it accumulates phase so as to satisfy Eq. \[phase\]. The signal beam thus eventually experiences gain even if it was initially lossy. The distance over which this rephasing occurs increases for phases far from $\Delta \phi = \pi/2$, becoming infinite for $\Delta \phi = 3\pi/2$. It should be noted that the exact phase relation $\Delta \phi = 3\pi/2$ is unstable, and any phase noise will cause $\phi_{4}$ to evolve towards satisfying Eq. \[phase\]. Numerical solutions for the log of the signal beam amplitude are plotted against distance in Fig. \[gainvphase\] for several pump-signal phase relations. ![Logarithm of signal beam amplitude as a function of propagation distance.[]{data-label="gainvphase"}](gainvphase.eps) We believe that this gain should be experimentally observable with current materials and laser powers. We anticipate that the principal difficulty will lie in distinguishing the signal beam from the pump beams, since their only distinguishing feature is the propagation direction. Given a minimum angular resolution between the beams of $\alpha \approx 1^{\circ}$, satisfying the relation in Eq. \[range\] implies $$|\gamma P| / k \geq \frac{\Delta k}{6k} = \frac{1}{3} (1 - \cos \alpha) \approx 5 \times 10^{-5}. \label{nonlinearity}$$ For a material with $|\chi^{(3)}_{xxxx}| = 10^{-6} \mathrm{cm}^{2}/ \mathrm{W}$, Eq. \[nonlinearity\] is satisfied for $P \geq 3 \mathrm{W/cm}^{2}$. Nonlinearities of this order are available in current materials including atomic vapors[@Wang2002]. We recall that weak-wave retardation in degenerate four-wave mixing is responsible for transverse modulational instability and resulting laser beam filamentation in the spatial domain[@Boyd1999]. This behavior agrees with the idea that time-domain modulational instability may be thought of as a four-wave mixing process where the fiber nonlinearity provides phase matching between the different fields[@Agrawal2001]. In modulational instability in the time domain, a CW beam breaks up into a periodic pulse train with the simultaneous appearance of associated spectral side bands[@Tai1986]. We conjecture that there should exist a time-domain version of the weak-wave advancement Kerr-mediated four-wave mixing effect described here. This effect should occur if we propagate light at two different frequencies $\omega _{0}+\Omega $ and $\omega_{0}-\Omega $ through an optical fiber in the normal dispersion regime. These two optical frequency components correspond to the noncollinear pump beams in the spatial case discussed here. As a result, a stationary signal at the frequency $\omega _{0}$ should be generated, corresponding to the bisecting signal beam in the spatial case. In conclusion, we have identified a new four-wave mixing process mediated by the Kerr nonlinearity, and related this process to a different, inverse kind of modulational instability. We believe that this nonlinear scattering process could be at the heart of an experimental realization of a photonic Bose-Einstein condensate, which would require that photons with high transverse momentum (“hot” photons) be scattered into low transverse momentum modes (become “cool” photons)[@Chiao2000]. We thank M. Trassinelli and J. Garrison for helpful discussions. This work was supported in part by the ONR. JMH thanks the support from Instituto do Milênio de Informação Quântica, CAPES, CNPq, FAPEAL, PRONEX-NEON, ANP-CTPETRO. [99]{} E. Fermi, J. Pasta, and H. C. Ulam, in *Collected Papers of Enrico Fermi*, edited by E. Segrè (The University of Chicago, Chicago, 1965), vol. 2 pp. 977-988. G. Van Simaeys, Ph. Emplit, and M. Haelterman, “Experimental demonstration of the Fermi-Pasta-Ulam recurrence in a modulationally unstable optical wave,” Phys. Rev. Lett. **87**, 033902 (2001). G. Van Simaeys, Ph. Emplit, and M. Haelterman, “Experimental study of the reversible behavior of modulational instability in optical fibers,” J. Opt. Soc. Am. B **19**, 477 (2002). N. N. Akhmediev, “Nonlinear physics - Deja vu in optics,” Nature **413**, 267 (2001). G. P. Agrawal, “Transverse modulation instability of copropagating optical beams in nonlinear Kerr Media,” J. Opt. Soc. Am. B **7**, 1072 (1990). J. M. Hickmann, A. S. L. Gomes, and C. B. de Araújo, “Observation of spatial cross-phase modulation effects in a self-defocusing nonlinear medium,” Phys. Rev. Lett. **68**, 3547 (1992). R. W. Boyd, and G. S. Agarwal, “Preventing laser beam filamentation through use of the squeezed vacuum,” Phys. Rev. A **59**, R2587 (1999). M. W. Mitchell, C. J. Hancox, and R. Y. Chiao, “Dynamics of atom-mediated photon-photon scattering,” Phys. Rev. A **62**, 043819 (2000). R. Y. Chiao, P. L. Kelley, and E. Garmire, “Stimulated four-photon interaction and its influence on stimulated Rayleigh-wing scattering,” Phys. Rev. Lett. **17**, 1158 (1966). G. P. Agrawal, *Nonlinear Fiber Optics*, 3$^{rd}$ed. (Academic, San Diego, 2001). R. Y. Chiao, M. A. Johnson, S. Krinsky, H. A. Smith, C. H. Townes, and E. Garmire, “A new class of trapped light filaments,” IEEE J. of Quant. Elec. **QE-2** 467 (1966). A. J. Campillo, S. L Shapiro, and B. R. Suydam, “Periodic breakup of optical beams due to self-focusing,” Appl. Phys. Lett. **23**, 628 (1973). H. Wang, D. Goorskey, and M. Xiao, “Dependence of enhanced Kerr nonlinearity on coupling power in a three-level atomic system,” Opt. Lett. **27**, 258 (2002). K. Tai, A. Hasegawa, and A. Tomita, “Observation of modulational instability in optical fibers,” Phys. Rev. Lett. **56**, 135 (1986). R. Y. Chiao, “Bogoliubov dispersion relation for a ’photon fluid’: Is this a superfluid?,” Opt. Comm. **179**, 157 (2000).
{ "pile_set_name": "ArXiv" }
--- abstract: 'The determinantal point process (DPP) is an elegant probabilistic model of repulsion with applications in various machine learning tasks including summarization and search. However, the maximum a posteriori (MAP) inference for DPP which plays an important role in many applications is NP-hard, and even the popular greedy algorithm can still be too computationally expensive to be used in large-scale real-time scenarios. To overcome the computational challenge, in this paper, we propose a novel algorithm to greatly accelerate the greedy MAP inference for DPP. In addition, our algorithm also adapts to scenarios where the repulsion is only required among nearby few items in the result sequence. We apply the proposed algorithm to generate relevant and diverse recommendations. Experimental results show that our proposed algorithm is significantly faster than state-of-the-art competitors, and provides a better relevance-diversity trade-off on several public datasets, which is also confirmed in an online A/B test.' author: - | Laming Chen\ Hulu LLC\ Beijing, China\ `[email protected]`\ Guoxin Zhang[^1]\ Kuaishou\ Beijing, China\ `[email protected]`\ Hanning Zhou\ Hulu LLC\ Beijing, China\ `[email protected]`\ bibliography: - 'reference.bib' title: Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommendation Diversity --- [^1]: This work was conducted while the author was with Hulu.
{ "pile_set_name": "ArXiv" }
--- author: - 'C. Maitra[^1]' - 'F. Acero' - 'C. Venter' bibliography: - 'j0855b.bib' date: 'Received ... / Accepted ...' title: 'Constraining the geometry of [[PSRJ0855$-$4644]{}]{}: A nearby Pulsar Wind Nebula with Double Torus/Jet Morphology' --- =1 Introduction {#sec:intro} ============ Pulsar wind nebulae (PWNe) are powered by young energetic pulsars and are excellent sites to study the energetics of and particle spectrum injected by young rotation-powered pulsars. The unprecedented spatial resolution of [[*Chandra *]{}]{}has unravelled highly structured compact PWNe around many of these sources, including equatorial and polar outflows [@weisskopf2000; @gaensler2006; @pwnchandra]. An accurate knowledge of the PWN morphology is crucial to constrain the geometry of the pulsar, especially the pulsar’s line of sight $\zeta_{\rm PSR}$ defined as the angle between the direction of the observer and the pulsar spin axis, and the magnetic obliquity $\alpha$ defined as the angle between the magnetic axis and the spin axis of the pulsar. Such constraints can in turn provide useful insights for models of high-energy emission from pulsar magnetospheres, such as the polar-cap (PC) model [@1996ApJ...458..278D], the outer gap (OG) model [@CHR86a; @1995ApJ...438..314R; @Romani96], and the slot gap or two-pole caustic (TPC) model [@Arons83; @Dyks03]. [[PSRJ0855$-$4644]{}]{} was discovered by the Parkes multibeam radio survey lying on the south-eastern rim of the RX J0852.0$-$4622 supernova remnant [SNR; @2003MNRAS.342.1299K]. The measured spin characteristics, such as the spin period $P=65$ ms and its first derivative $\dot{P} = 7.3 \times 10^{-15} ~s ~s^{-1}$ gives a spin-down luminosity $\dot{E} = 1.1 \times 10^{36}$ erg/s (assuming a moment of inertia $I = 10^{45} g ~cm^{2}$ for standard neutron star parameters), and characteristic age $\tau_{\rm c} \equiv {P}/(2\dot{P}) = 140$ kyrs. A dedicated [[*XMM-Newton *]{}]{}[[observation ]{}]{}revealed that the X-ray counterpart of the pulsar is surrounded by extended non-thermal emission, the associated PWN [@facero2013]. The emission was 150 in extent, including two large-scale structures with an angular separation of $\sim$180$\dg$, resembling possible jets. In addition, comparison of column densities provided an upper limit to the distance of the pulsar [[PSRJ0855$-$4644]{}]{} of $d\leq 900$ pc. With this revised distance, [[PSRJ0855$-$4644]{}]{} is the second most energetic pulsar, after the Vela pulsar, within a distance of 1 kpc from Earth, and could therefore contribute to the local cosmic-ray e$^{-}$/e$^{+}$ spectrum [@facero2013]. This furthermore makes this pulsar the highest $\dot{E}$/$d^{2}$ system not detected at $\gamma$-ray energies with the *Fermi*-LAT instrument [@abdo13-2PC]. As mentioned earlier, determination of the pulsar/PWN geometry holds the key to understand this paradox. We present results of a 38 ks on-axis [[*Chandra *]{}]{}observation to study the small-scale morphology of the nebula surrounding PSR J0855-4644 and its geometry. With the arcsecond spatial resolution of [[*Chandra *]{}]{}we have resolved the compact [[*XMM-Newton *]{}]{}source that was assumed to be the X-ray counterpart of the pulsar into a $\sim$ 10 extended emission with jets and a possible double torus. We performed a morphological and spectral analysis, and constrained the observer angle of the system. From geometric light curve modelling, we could also independently constrain the ($\alpha$, $\zeta_{\rm PSR}$) space by comparing our predictions to the radio pulse profile of the pulsar (and taking the non-detection of pulsed $\gamma$ rays into account). The [[observations ]{}]{}and analysis are described in Section 2. Section 3 presents the spatial analysis including the imaging and morphological fitting of the source. Section 4 presents the detailed spectral analysis of the pulsar, the surrounding nebula, and its decomposed regions. Section 5 details our geometric light curve modelling and comparison with observations. Section 6 is the discussion, and Section 7 the summary and future work. Observations and analysis {#sec:obs} ========================= The Advanced CCD Imaging Spectrometer (ACIS) onboard [[*Chandra *]{}]{}offers simultaneous high resolution images and moderate-resolution spectra. ACIS consists of 10 planar CCDs, 4 front-illuminated ones (ACIS-I) are arranged in $2 \times 2$ array, and 6 back-illuminated ones (ACIS-S) in a $1 \times 6$ array. The [[*Chandra *]{}]{}[[observation ]{}]{}(ObsID 13780) was carried out with the ACIS-S as the primary instrument, and was performed in full-frame timed exposure (TE) mode which uses the standard 3.2 s frame time and the FAINT telementry format. The [[observation ]{}]{}was performed on 3 September 2012 with [[PSRJ0855$-$4644]{}]{} positioned at the aimpoint of the S3 CCD with an exposure of 38.02 ks. Data reduction was performed using CIAO 4.7, CALDB 4.6.7, and the standard prescribed analysis procedure[^2]. The event file was reprocessed with *chandra$\_$repro*, which incorporates the subpixel repositioning algorithm EDSER as a default to obtain a better angular resolution for sources near the centre of the field of view. The effective exposure of the observation after filtering was 37.98 ks. We also checked for the possible presence of pileup in the data. Many times, multiple photons impinge on a single pixel during readout, leading to it being read out as one single photon with a large pulse height, and causing photon pileup. Using the tool [*pileup$\_$map*]{}, the estimated fraction of pileup in the centremost pixel was $<5\%$ and so the effect was not important for our observation. Images were created using the task *dmcopy*. Spatial analysis was performed using the *Sherpa* analysis package 4.4[^3]. The task [*specextract*]{} was used for extracting source and background spectra and response files from regions of interest. The selection of regions used for spectral extraction is described in Section 4. Spectra were fitted using *XSPEC v12.8.2*[^4]. Figure \[images\]a shows a full resolution [[*Chandra *]{}]{}image in the energy range 0.5–8 keV centred on [[PSRJ0855$-$4644]{}]{}. The point source seen with [[*XMM *]{}]{}that was previously thought to be the pulsar X-ray counterpart is in fact composed of an extended structure elongated in the south-east/north-west direction along a symmetry axis (See Figure \[images\]a). This component of the PWN is more compact ($\sim$ 10) compared to the much larger diffuse nebula seen with [[*XMM *]{}]{}measuring 150 [@facero2013]. Signatures of the diffuse nebula can be seen in the [[*Chandra *]{}]{}image in the high-energy 2–8 keV band (free of emission from the Vela Senior SNR) after smoothing it at the scale of the compact PWN, 10 (See Figure \[images\]d). The fine structures of the compact PWN can be seen more clearly from its subpixel images of the same region after dividing them into soft [*i.e.*]{} 0.5–2 keV and hard, [*i.e.*]{} 2–8 keV energy bands (See Figures \[images\]b and c). The elongated PWN structure is composed of two lobes lying on the symmetry axis, more or less equidistant from the pulsar in the middle. The pulsar spectrum is soft and almost disappears in the 2–8 keV energy band. The nebular structure on the other hand is much harder and extends more to the east in the form of a tail-like elongation. The structure is reminiscent of axisymmetric features commonly seen in young PWN systems like the equatorial outflows (tori), and collimated polar outflows (jets) [@gaensler2006; @pwnchandra]. The two lobes symmetric about the pulsar can be explained by the three possible scenarios: a) double-torus b) double-torus and jets, and c) jets-only morphology. In the first scenario the lobes could be the two rings of a double torus structure as it would appear in sky projection, spaced more or less equidistant from the pulsar centered along the torus axis (which is coincident with the symmetry axis, a.k.a. the spin axis). The brightened central parts of the lobes are due to the doppler boosting of the torus. In the second scenario, the morphology of the PWN around [[PSRJ0855$-$4644]{}]{}  is that of a double torus-jet structure, with the presence of both the equatorial torus and polar jet features contributing to the lobes. This is similar to the PWNe systems seen in Vela and PSR J2021+3651 [@helfand2001; @hessels2004]. In the third case, the lobes could be the jets along the pulsar spin axis (the symmetry axis in Figure \[images\]a). In all three cases, however the eastern part of the nebula shows a further protrusion in the form of an outer jet. While it is hard to distinguish between these scenarios, and disentangle the dominance of either the jets or torus in the present observation, the presence of torus/jets are strong indicators of $\zeta_{\rm PSR}$ and can constrain the geometry of the system as shown in Section 3. To accurately determine the position of the point source (pulsar), we created a subpixel image of the same (at one-fifth of the ACIS pixel resolution) and applied the source detection algorithm [*celldetect*]{} in the soft energy band of 0.5–2 keV where the pulsar is visible. The coordinates of the point source are RA(J2000)$=08^h55^m36.138^s$ Dec$=-46\deg44\arcm13.57$ considering an error of 0.6 at 90 $\%$ confidence level in absolute [[*Chandra *]{}]{}astrometry. This is consistent with the radio position of the pulsar. The net count rate from the point source after subtracting the nebular component is 0.0005 counts/s, and from the entire nebula is 0.02 counts/s. The details of background subtraction are discussed in the spectral analysis Section \[sec-spec\]. X-ray spatial analysis {#sec:xspec} ====================== We performed detailed spatial analysis of the compact PWN surrounding [[PSRJ0855$-$4644]{}]{}. To probe the structures of the nebula, especially the axisymmetric features, we created a counts profile in the box region shown in Figure \[images\]a. We also performed morphological fitting of the PWN using the torus fitting model of [@ngromani2004; @ngromani2008], and determined its geometrical parameters like the torus radii, the position angle, and $\zeta_{\rm TORUS}$. Figure \[radial\] shows the 0.5–8 keV counts profile decomposed into two energy bands. The profile was created in the box region of length 10 shown in Figure \[images\]a, along the symmetry axis, averaged over a region (4.5 in width) perpendicular to it. The box was centred with respect to the X-ray counterpart of the pulsar, and we overlaid the point-spread function (PSF) in the figure. The PSF of the [[observation ]{}]{}was simulated using the [[*Chandra *]{}]{}ray tracer chaRT[^5] which simulates the High Resolution Mirror Assembly based on the energy spectrum of the source and the [[observation ]{}]{}exposure. The output of chaRT was modelled with the software MARX[^6] taking the instrumental effects and the EDSER subpixel algorithm into account to be consistent with the observational data. The best-fit spectrum of the point source (see Section \[sec-spec\]) was used. The profile clearly shows all the structures of the nebula including the two lobes symmetric about the pulsar. The western lobe has higher counts than the eastern one, although consistent within errors. The region west of the pulsar declines more steeply with radius, almost immediately after the lobed feature. The eastern part on the other hand has two additional bumps, the farthest one corresponding to a tail-like protrusion seen in the images, which might be the outer jet. The decomposed counts profile highlights the dominance of some features of the PWN in certain energy ranges. As seen from the images, and also from spectral analysis (later in Section \[sec-spec\]), the X-ray counterpart of the pulsar is clearly seen in the energy band of 0.5–2 keV, while it is less dominant above 2 keV. The east lobe of the nebula shows an indication of being harder, being more dominant in the energy range of 2–8 keV. The eastern outer jet has equal dominance in the soft and hard energy bands. ![The top panel shows the counts profile of the source extracted from the box region in Figure 1a (0.5–8 keV) compared with the 2-D PSF (0.5–8 keV, red dashed line). The bottom panel portrays the same profile in the two energy bands 0.5–2 keV and 2–8 keV. []{data-label="radial"}](count_profile_decomposed.pdf) Section 2 discusses three possible scenarios for the morphology of the PWN around [[PSRJ0855$-$4644]{}]{}. In two of the three scenarios where the lobes either correspond entirely to the arcs of the torus, or to one of the torus components along with the jets, it is possible to provide quantitative measurements of the PWN geometry by modelling the lobes as a double torus as in the pulsar wind torus fitting scheme of [@ngromani2008]. In the third case, identifying the lobes entirely as jets instead does not alter the geometry of the system as the torus axis is coincident with the spin axis of the pulsar along which jets are formed [@ngromani2008]. The parameters of the torus model are the torus axis position angle (PA) $\Psi$ (N-E)[^7] Spatial fitting was performed on the full resolution image, in a region of size 47 $\times$ 45 centred on the pulsar to constrain the background better. The PSF described in the previous section was loaded as a table model in [*Sherpa*]{} to model the pulsar emission. The remaining excess consisted of the background and the emission from the nebula. The background was modelled with a constant. The lobes of the nebula were modelled with the relativistic torus model of [@ngromani2008] as a double torus seperated by a distance *d*. The background and the double-torus model was then convolved with the PSF. The $\delta$, if left free, has a tendency to grow and absorb the unmodelled larger-scale PWN components. As this parameter does not provide any physical information on the shock or the geometry of the system, it was frozen to 0.5 provided by an initial visual estimate. Figure \[corr\] shows the best-fit model along with the data used for fitting and the residuals. The best-fit parameters were determined using the C-statistic [@cash1979][^8]. The best-fit parameters are listed in Table \[tabsherpa\]. The 1$\sigma$ confidence intervals are quoted here. The statistical errors were estimated by performing Monte Carlo Poisson realizations of the best-fit model as recommended in [@ngromani2008]. To characterize the [**uncertainty on the estimates associated with the unmodelled components in the fit**]{}, especially the eastern tail region, we blanked out the region in the fit and noted the change in parameters. Considering that the [**two**]{} errors add in quadrature, we find that the pulsar line of sight $\zeta_{\rm PSR}$ is inclined at $32.5\dg\pm 4.3\dg$ with an upper limit of $\zeta_{\rm PSR}$ &lt; $37\dg$. Parameter Value Units --------------------- -------------------------- -------------- $\Psi^{a}$ $114.4\pm 2.3 \pm 1.1$ $\dg$ $\zeta_{\rm PSR}$ $32.5 \pm 4.0 \pm 1.7$ $\dg$ R $2.1\pm 0.06 \pm 0.08$   $\beta_{\rm shock}$ $0.41\pm 0.06 \pm 0.06$ – *d* $3.6\pm0.70 \pm 0.30$   Background $0.46\pm0.02$ counts/pixel : Final morphological best fit with a double-torus model. The model was convolved with the PSF as described in the text. \ $^{a}$ For all the parameters except the background counts/pixel, the first error denotes the statistical error, and the second the [**uncertainty on the estimates associated with the unmodelled components in the fit**]{}.\ \[tabsherpa\] X-ray spectral analysis {#sec-spec} ======================= We have performed a detailed spectral analysis of the X-ray counterpart of [[PSRJ0855$-$4644]{}]{} and the surrounding compact PWN. To look for changes in the spectral parameters in different parts of the nebula we also extracted spectra from different regions by dividing the PWN into an inner and outer annular part, and have examined the spectra of the east and west lobes separately. The analysis was performed in the energy range of 0.5–8 keV. The C-statistic was used for spectral fitting and errors were estimated at a 90$\%$ confidence interval. The regions used for spectral extraction are shown in Fig. \[spec-1\]. For the pulsar, a circular region of radius 0.7 was extracted, centred on the best-fit coordinates of the source. In the case of the compact PWN, the outer radius was optimized from the spatial fitting, with the region corresponding to the pulsar excised. For all the regions, the background spectrum was extracted from an annular region with inner and outer radii of 75 and 100 (extent of the diffuse nebula seen with [*XMM*]{}: see Fig. \[images\]d). This ensured the background was free from the fainter diffuse nebular emission that was seen with the [[*XMM-Newton *]{}]{}observations. In the case of the pulsar, apart from the annular background spectrum, the astrophysical background due to the nebula was considered, and was modelled with the best-fit spectrum of the entire PWN, scaled with the ratio of the two extraction regions. We also checked for an additional astrophysical background contribution from the pulsar in the nebular region, and found it to be a negligible fraction of the total background. We used an absorbed power law as the spectral model for the compact PWN and all the decomposed regions marked in Fig. \[spec-1\]. In order to account for the photoelectric absorption by the interstellar gas along the line of sight, a free absorption ([*XSPEC*]{} model [*tbabs*]{}) component was used. This component was determined from the fit of the entire nebula and frozen henceforth for all other regions as it is not expected to vary locally. The value obtained with [[*Chandra *]{}]{}is $N_{\rm H} = 0.70^{0.25}_{-0.20}\times10^{21}$ cm$^{-2}$, which is consistent with that obtained using [*XMM-Newton*]{}, derived from a 15  radius region ($N_{\rm H}$ = 0.64$^{0.13}_{-0.11}\times10^{21}$ cm$^{-2}$). The spectrum of the PWN is hard with a power-law index $\Gamma$ of $1.12\pm0.25$. The unabsorbed bolometric luminosity L$_{\rm eff}$ (0.5–8 keV) is 3.3$\times10^{31}\ergsec$ assuming a distance of $d\sim900$ pc. Comparing the spectra of the inner and outer nebula, we did not detect any steepening of the power-law spectral index $\Gamma$ which could have been indicative of synchrotron cooling. In the case of the east and west lobe, while there is an indication from the counts profile (Fig. \[radial\]) that the eastern lobe has a harder spectrum than the western region (higher counts in 2–8 keV band in the east lobe compared to 0.5–2 keV), we do not find a significant difference in the spectral indices, given the current level of statistics. The pulsar spectrum is faint and soft with the statistics deteriorating drastically beyond 5 keV. The spectrum was fitted with an absorbed blackbody as well as a neutron star hydrogen atmosphere model [[*nsa*]{}; @2002yCat..33861001G; @zalvin2009]. Both models provide acceptable fits to the data with comparable C-stat values. The fit with the [*nsa*]{} model shows an effective temperature T$_{\rm eff}$ of $2.2 \times 10^{6}$ K and an unabsorbed bolometric luminosity $L_{\rm eff}$ (0.5–8 keV) of $1.4 \times10^{30}\ergsec$ assuming a distance of $d=900$ pc. The blackbody fit of the thermal component gives a similar temperature $T_{\rm BB}$ of $2.0 \times10^{6}$ K and $L_{\rm eff}$ of $1.3 \times10^{30}\ergsec$. Both indicate an effective radius $R_{\rm eff}$ of $\sim 1.5$ km suggesting that the emission originates from hot spots on the poles of the neutron star, rather than being an emission from the entire surface. An additional power-law component (related to potential non-thermal emission) was not required for the fit. The pulsar and the compact PWN spectrum along with its best-fit model and residuals are shown in Fig. \[spec-2\]a and  \[spec-2\]b, and the best-fit parameters for all the regions along with the C-statistic values are given in Table \[table-specfit\]. ![image](aug25_model.pdf) ![image](chandra_reg_mod_rgb.pdf) ------------------- ------------------------ ------------------------ ---------------------- ----------------------- --------------------------------- ---------------------------------- ---------------------- Region Absorption Photon index $kT_{\rm BB}^{\inf}$ $kT_{\rm eff}^{\inf}$ $F_{x}$$^{a}$ $F_{x}$$^{a}$ C-stat/$N_{\rm dof}$ – $10^{21}$cm$^{-2}$ $\Gamma$ keV keV 0.5–8 keV 2–10 keV Pulsar 0.70(f) – $0.20\pm0.05$ $0.22\pm0.05$ ($2.48\pm1.04$)$\times10^{-15}$ – 248.48/442 Compact nebula $0.70^{+0.25}_{-0.20}$ $1.12\pm0.25$ – – ($2.80\pm0.24$)$\times10^{-13}$ ($3.2\pm0.42$)$\times10^{-13}$ 466.50/440 Inner nebula 0.70 (f) $1.20\pm0.16$ – – ($1.94\pm0.40$)$\times10^{-13}$ ($2.17\pm0.32$)$\times10^{-13}$  434.86/440 Outer nebula 0.70 (f) $0.90^{+0.28}_{-0.24}$ – – ($8.87\pm1.3$)$\times10^{-14}$ ($1.07\pm0.21$)$\times10^{-13}$ 369.53/440 West lobe 0.70 (f) $1.30\pm0.33$ – – ($3.24\pm0.72$)$\times10^{-14}$ ($3.55\pm1.02$)$\times10^{-14}$ 255.71/442 East lobe 0.70 (f) $0.93\pm0.38$ – – ($3.15\pm0.78$)$\times10^{-14}$ ($3.86\pm1.20$)$\times10^{-14}$ 234.26/442 \[table-specfit\] ------------------- ------------------------ ------------------------ ---------------------- ----------------------- --------------------------------- ---------------------------------- ---------------------- \ $^{a}$Observed flux in units of $\ergcms$\ (f) Parameter fixed for consistency between fit regions.\ Instrumental line at $1.49 \pm 0.11$ keV sigma fixed. Constraints from geometric light curve modelling ================================================ The fact that [[PSRJ0855$-$4644]{}]{} is radio-loud but $\gamma$-quiet (despite its large pseudo luminosity $\dot{E}/d^2$) may be used to derive constraints on the pulsar geometry. The radio peak multiplicity and width may also aid in obtaining geometric constraints. One can compare the predictions of geometric light curve models with the observed radio profile to constrain both $\alpha$ and $\zeta_{\rm PSR}$. We used the TPC model [@Dyks03] and a geometric version of the OG model [@CHR86a; @Romani96], in conjuction with a semi-empirical hollow-cone radio model [@Gonthier04; @Harding07_Geminga]. The latter model is based on work by @Rankin93 [@ACC02], with the radio emission altitude expression as in @Kijak03. For more details, see @Venter09. We generated an ‘atlas’ of $\gamma$-ray and radio light curves on an ($\alpha$,$\zeta_{\rm PSR}$) grid, at a 5$^{\circ}$ resolution. In the next two sections, we attempt to provide conservative limits on $\alpha$ and $\zeta_{\rm PSR}$ based on the qualitative features of the observed radio profile (peak multiplicity, duty cycle) and the non-detection of $\gamma$-ray pulsations. Additionally, we perform a $\chi^2$ fit using the observed radio profile to further constrain $\alpha$ and $\zeta_{\rm PSR}$, in a subsequent section. Radio visibility and peak multiplicity -------------------------------------- For the radio models, we used period of $P = 65$ ms. This important parameter sets the radio beam width (assuming emission from a single altitude for simplicity), which scales as $P^{-1/2}$ (as does the polar cap opening angle $\theta_{\rm PC}$, with which it is closely associated). The beam width, in conjunction with the observer angle $\zeta_{\rm PSR}$, determine the radio visibility of the pulsar. Conversely, the $\gamma$-ray emission originates from a range of altitudes and undergo beaming effects, so the caustic $\gamma$-ray beam shape in this case is relatively independent of the period (unless the observer’s line of sight crosses the non-emitting polar cap on the stellar surface when $\alpha\approx\zeta_{\rm PSR}$, which will cause a dip in the $\gamma$-ray light curves). We set the radio frequency to $\nu=1400$ MHz (same as for the observed profile). The observer orientation with respect to the radio beam provides three regimes in terms of radio pulse shape predictions. For an impact angle $\beta \equiv \zeta_{\rm PSR}-\alpha\sim 0\dg$, a double-peaked structure will be seen. For larger values of $|\beta|$, a single peak will be visible, since the observer skims the edge of the hollow cone. For even larger values of $|\beta|$, the cone will be completely missed, and the pulsar will be radio-quiet. Additionally, for small $\alpha$, the duty cycle is large, and for large $\alpha$ and small enough $|\beta|$, one starts to see radio beams from both magnetic poles (leading to two profiles with about an 0.5 difference in normalized phase). Therefore, the fact that the observed radio peak of [[PSRJ0855$-$4644]{}]{} is single provides a lower limit $|\beta|\gtrsim10\dg-15\dg$ (the lower value is for low values of $\alpha$). On the other hand, radio visibility provides an upper limit on $|\beta|$: from the radio model we can derive the constraint $|\beta|\lesssim 30\dg$. This result is somewhat dependent on the assumed emission profile of the cone. Currently, we use an offset-Gaussian flux profile that peaks at a co-latitude $\bar{\theta}$. Since we assume emission from a single height, this maximum relative flux is independent of $\alpha$ and $\zeta_{\rm PSR}$. The Gaussian profile has tails reaching beyond the last open $B$-field lines (which are tangent to the light cylinder at a cylindrical radius $r=cP/2\pi$, where the corotation speed equals the speed of light), into the corotating region. If one wants to restrict the emission by not allowing any radio flux from this region (i.e., terminating the tails at $\theta=\theta_{\rm PC} \approx \sqrt{2\pi R/Pc}$), this would lead to smaller radio beams (by a few degrees) and therefore to tighter constraints on $|\beta|$. However, given the inherent uncertainties in the beam geometry, it might not be wise to pursue such stricter constraints on $|\beta|$ that depend on the detailed beam structure. We can therefore summarize our constraint from this section as $10\dg\lesssim|\beta|\lesssim 30\dg$. This constraint is also borne out by the $\chi^2$ fit to the data performed below. $\gamma$-ray invisibility (non-detection) ----------------------------------------- The pulsar geometry is much better constrained if radio and $\gamma$-ray data (or a non-detection in the latter case) are used simultaneously [@Pierbattista15]. In our case, the lack of $\gamma$-ray visibility leads to further constraints on $\alpha$ and $\zeta_{\rm PSR}$, within the context of the TPC and OG models. In Figure \[fig:TPC\], we have plotted $\gamma$-ray (black) and radio (magenta) light curves on an $(\alpha,\zeta_{\rm PSR})$ grid at a resolution of 10$^{\circ}$. We have normalized the $\gamma$-ray intensity so that the maximum (typically at $\alpha=\zeta_{\rm PSR}=90\dg$) is unity and the flux of the other combinations of $\alpha$ and $\zeta_{\rm PSR}$ relative to this maximum is clear[^9]. The geometric models cannot predict absolute[^10] $\gamma$-ray fluxes, neither can they give an indication of the $\gamma$-ray flux relative to the radio flux at a particular viewing geometry, so one cannot use a sensitivity-limit argument to constrain $\alpha$ and $\zeta_{\rm PSR}$. Rather, one can only use visibility (and radio pulse shape) arguments to constrain these angles. We note that we have followed @Venter09 and assumed a gap width of 0.05$\theta_{\rm PC}$ for both the TPC and OG models. It is true that in physical pulsar models, the gap width sensitively depends on the pulsar period and magnetic field. For example, @Muslimov2003 give expressions for the slot gap width as a function of $P$ and $B$, and the width scales linearly with $P$. When using geometrical models, however, the usual approach is to assume a reasonable value for this width which results in good light curve fits. Our choice of 5% is also supported by the results of @Johnson2014 who fit the light curves of 40 $\gamma$-ray millisecond pulsars, finding that their best-fit gap width never exceeded 10%, but is usually smaller. Further justification comes from Figure 2 of @Muslimov2003. While it is true that the gap width influences the shape of the resulting $\gamma$-ray light curves, the actual $\gamma$-ray pulse shape is of less importance in our case due to the following reasons: (i) Observational data imply that the gaps must be narrow in order to reproduce the observed shapes; (ii) We are only using the geometrical gap models to decide for which parameter ranges ($\alpha$ and $\zeta_{\rm PSR}$) the $\gamma$-ray pulse should be visible. Therefore, the width of the pulse is less important than the relative intensity at different $(\alpha,\zeta_{\rm PSR})$, and different gap width choices should not alter our main conclusions. As for the magnetic field structure, we have assumed the vacuum retarded dipole [@Deutsch55; @Cheng00; @Dyks04], which has been standard in recent years, allowing other authors to compare their results to ours. For the TPC model, one can see that the $\gamma$-ray light curves are visible at almost all angles (except $\alpha=\zeta_{\rm PSR}\sim10\dg$). The low-level $\gamma$-ray emission reflects the emission outside of the main caustic beam. However, the relative flux increases dramatically when the observer samples the bright caustic (e.g., compare $(\alpha,\zeta_{\rm PSR})=(10\dg,60\dg)$ and $(\alpha,\zeta_{\rm PSR})=(10\dg,70\dg)$). Given the $\gamma$-quiet and single visible radio pulse constraints, this would imply both $\alpha$ and $\zeta_{\rm PSR} \lesssim 50\dg$, assuming that the low level of $\gamma$-ray flux (and broad $\gamma$-ray peaks) would not be detectable. The case for the OG model is perhaps a bit clearer. Since no $\gamma$-ray emission is generated below the null charge surface where the Goldreich-Julian charge density is zero [@GJ69], $\gamma$ rays are only visible for $\zeta_{\rm PSR} \gtrsim50\dg$, and for large $\alpha$. For example, at $(\alpha,\zeta_{\rm PSR}) \sim (50\dg,50\dg)$, one starts to sample the main caustic beam. (For larger $\zeta_{\rm PSR}$, one samples the caustic twice during one rotation, leading to double peaks.) Since the radio model is the same for both the TPC and OG cases, one derives the same constraints on $\alpha$ and $\zeta_{\rm PSR}$ for both $\gamma$-ray models. At a finer resolution, one finds that the constraint may be written as $\alpha \lesssim 55\dg$ and $\zeta_{\rm PSR} \lesssim 55\dg$. It would be difficult to pick a best model between the OG and TPC in this case, since this would depend on whether the low-level TPC emission is deemed detectable or not. This is difficult to decide given the lack of absolute flux predictions. Radio profile fit ----------------- In this section, we constrain $\alpha$ and $\zeta_{\rm PSR}$ by fitting the observed radio profile from [@2003MNRAS.342.1299K] using our radio model and a $\chi^2$ minimization procedure. We are only interested in reproducing the pulse shape, so we normalize the flux maximum to unity. We estimate the error on the data by calculating the standard deviation of the flux for the phase range $\phi<0.4$ (this range was chosen as the ‘off-peak region’), yielding $\sigma_{\rm data} \sim 0.06$. The number of degrees of freedom is $N_{\rm dof} = 255$. We used $\chi^2 = \sum_i[(Y_{\rm model}-Y_{\rm data})/\sigma_{\rm data}]^2$ for $i=1,\dots,N_{\rm bins}$. The radio model was run for $\alpha \in [0\dg,90\dg]$ and $\zeta_{\rm PSR} \in [0\dg,90\dg]$ at a $1\dg$ resolution, for $P=65$ ms and $\nu=1400$ MHz, and initially for 720 phase bins. In order to perform a $\chi^2$ fit, we rebinned (smoothed) the model to have the same number of bins as that of the data ($N_{\rm bins} = 256$) using a Gaussian Kernel Density Estimator (KDE) with step size $h=1/N_{\rm bins}$ [@Rosenblatt56; @Parzen62]. We furthermore minimized $\chi^2(\alpha,\zeta_{\rm PSR})$ for each fixed value of $\alpha$ and $\zeta_{\rm PSR}$ over a phase shift parameter $d\phi \in [0,0.99]$ with a resolution of $0.01$. This is to allow alignment of the model and observed light curves, circumventing any issues that may stem from the uncertain definition of $\phi=0$ [cf. @Johnson2014]. The background level was set to zero, and the model curve was treated as being cyclical. Figure \[fig:chi2\] shows the $\log_{10}[\chi^2(\alpha,\zeta_{\rm PSR})]$ map. The dark red colour indicates where the radio is invisible ($|\beta|\gtrsim30\dg$ as mentioned above). As one moves inward from larger to smaller $|\beta|$ values, the edge of the cone is clipped and the pulse is too narrow (cyan colour, $|\beta|\sim25\dg$); at $|\beta|\sim10\dg-15\dg$ a wider single peak is obtained that reasonably fits the data (blue colour), and at even smaller $|\beta|\lesssim10\dg$, a double peak is obtained (greenish colour). At small $\alpha$ and $|\beta|$ (e.g., $\alpha\sim15\dg$ and $\beta\sim5\dg$), the predicted single peak is very wide (yellow colour), and at $\alpha\gtrsim0\dg$ and $|\beta|\sim5\dg$, the edge of the cone is clipped for most of the rotation, so that the light curve is nearly flat, or very broad, with a very large duty cycle (light red colour). The best fit is found at $(\alpha,\zeta_{\rm PSR}) = (22\dg, 8\dg)$ for $d\phi = 0.73$, giving $\chi^2/N_{\rm dof} = 285/255$. The sigma contours (yellow lines) are calculated using $\log_{10}[\chi^2_{\rm min}+\sigma_i]$, with $\sigma_1 = 2.3$, $\sigma_2 = 6.18$, and $\sigma_3 = 11.83$ [@Lampton76]. They are quite small (and barely visible on the $\chi^2$ contour plot) and the $1\sigma$ contours imply strong constraints on the best-fit values of $\alpha$ and $\zeta_{\rm PSR}$. However, we note that there is an alternative fit at $(\alpha,\zeta_{\rm PSR}) = (9\dg, 25\dg)$ for $d\phi = 0.74$, giving $\chi^2/N_{\rm dof} = 294/255$, which lies within $3\sigma$ of the best fit. Other fits within $3\sigma$ are $(\alpha,\zeta_{\rm PSR}) = (23\dg, 10\dg)$ and $(\alpha,\zeta_{\rm PSR}) = (24\dg, 9\dg)$. These best-fit values satisfy the constraints derived in the two preceding sections. Figure \[fig:bestfitLC\] shows the best-fit radio light curve overplotted on the data. ![image](TPC1_phiglobal2.pdf){width="50.00000%"} ![image](OG1_phiglobal2.pdf){width="50.00000%"} ![The $\log_{10}\chi^2/N_{\rm dof}(\alpha,\zeta_{\rm PSR})$ map with the best fit indicated by a yellow cross. The $1\sigma$, $2\sigma$, and $3\sigma$ contours are very small and are indicated by yellow lines.\[fig:chi2\]](log10_chi2_finest2_fin_chi_nu.pdf) ![The best-fit radio light curves overplotted on the data.\[fig:bestfitLC\]](LCs_finest2_fin.pdf) Discussion ========== Thanks to a dedicated [[*Chandra *]{}]{}observation, the compact PWN around the fast-spinning energetic pulsar [[PSRJ0855$-$4644]{}]{} has been resolved. The presence of two lobes, axisymmetric about a fainter pulsar counterpart, can have contributions from either the torus or jet structures or both of them, with the dominance of one component over the other indistinguishable by the present observation. Using the double torus model of @ngromani2008 to spatially fit the lobes captures the characteristic features of the PWN. Most importantly, it provides a robust estimate of $\zeta_{\rm PSR}$ agreeable with constraints derived from geometric modelling of the radio pulse profile and non-detection of $\gamma$-ray pulsations from the pulsar. X-ray pulsar ------------ The X-ray counterpart of [[PSRJ0855$-$4644]{}]{} is faint and soft with a blackbody temperature of $0.2\pm0.05$ keV and $L_{\rm eff}$ of $1.3\times10^{30}\ergsec$. The blackbody is emitted from a region with $R_{\rm eff}\sim1.5$ km, which indicates a polar cap nature of this emission rather than from the surface of the neutron star. Assuming co-located non-thermal X-ray and $\gamma$-ray emission regions, detection of thermal emission from the polar cap and the lack of non-thermal X-rays indicate that we are viewing the source close to the pole of the neutron star, with small $|\beta|$. Given the absence of detection of non-thermal emission from the pulsar, we cannot determine the parameters of the system like the X-ray efficiency of the pulsar ($\eta_{\rm psr} \equiv L_{\rm psr}/\dot{E}$) or the ratio of the efficiency of the PWN to its pulsar ($\eta_{\rm pwn}/\eta_{\rm psr}$). Given the spin period and its first derivative ($P$ and $\dot{P}$ ) as determined from radio observations [@2003MNRAS.342.1299K] , the equatorial surface magnetic field can be estimated to be $B \equiv 3.2 \times 10^{19} (P\dot{P})^{\frac{1}{2}} \equiv$ 0.7$\times10^{12}$ G (assuming a dipole magnetic field). This is in the range expected for a neutron star powering a PWN. Morphology and properties of the compact PWN -------------------------------------------- The spectrum of the compact nebula is hard, compatible with a power law of $\Gamma=1.12\pm0.25$. This is compatible with that observed in other PWNe [@pwnchandra] and is also compatible with predictions of particle wind models assuming relativistic Fermi acceleration [@2001MNRAS.328..393A]. At a distance of $d=900$ pc, the compact PWN radius $\sim$ 10 corresponds to a physical size of 0.06 pc. This is similar to the inner PWN of Vela [$\sim$ 0.1 pc; @pavlov2001]. The fainter and diffuse PWN extends much farther as hinted from the [[*XMM *]{}]{}[[observation ]{}]{}[@facero2013] corresponding to a physical size of $\sim$ 0.6 pc. An $L_{\rm eff}$ (0.5–8 keV) of $3.3\times10^{31}\ergsec$ indicates a nebular efficiency ($\eta_{\rm pwn} \equiv L_{\rm psr}/\dot{E}$) of $\sim$ 10$^{-5}$, which is similar to what is observed for Vela and some Vela-like pulsars and their PWNe (which have their $\dot{E}$ in the same ballpark range). This is, however, only a lower limit as the contribution from the diffuse nebula has not been taken into account in this case. Morphological modelling of the inner PWN with a double-torus model provides a reasonable fit, resulting in an observer angle $\zeta_{\rm PSR} < 37\dg$. The eastern outer tail-like feature which is not accounted for in the spatial model could represent the outer jet of the nebula. In that case the system would have dominance of one of the jet components, which also supports a small value of $\zeta_{\rm PSR}$. Pulsar geometry & its implications ---------------------------------- Constraints obtained from spatial modelling indicate $\zeta_{\rm PSR}\lesssim 37\dg$ (although with some uncertainty, since we cannot rule out a higher value of $\zeta_{\rm PSR}$ solely from the spatial modelling). Constraints from geometric[^11] light curve modelling indicate $\alpha\lesssim55\dg$ and $\zeta\lesssim55\dg$, and $10\dg\lesssim|\beta|\lesssim30\dg$ (due to a single visible radio peak, and from the width of the radio pulse). A $\chi^2$ fit to the radio light curve yields a best fit at $(\alpha,\zeta_{\rm PSR}) = (22\dg, 8\dg)$, with a few alternative fits within $3\sigma$, including one at $(\alpha,\zeta_{\rm PSR}) = (9\dg, 25\dg)$. Assuming (nearly) co-located regions of non-thermal X-ray and $\gamma$-ray emission, this is also consistent with the lack of (pulsed) non-thermal X-rays from the pulsar. To further break the degeneracy of $\alpha$-$\zeta_{\rm PSR}$ parameter space and improve constraints on the individual values of $\alpha$ and $\zeta_{\rm PSR}$, one could pursue radio polarization modelling, as well as X-ray light curve modelling, should radio polarization data become available and X-ray pulsations be detected[^12]. The obtained range of ($\alpha$, $\zeta_{\rm PSR}$) implies we are viewing the system relatively close to the spin axis of the pulsar. Limits on $\beta$ further implies we are also viewing the system relatively close to the magnetic poles of the pulsar. Interestingly, @R16 recently found that radio profile widths of a sample of 35 young, energetic radio pulsars that are not detected in $\gamma$ rays are typically broader than those of another sample of 35 pulsars which are detected in both radio and $\gamma$ rays, and that the boundary between these two populations increases with $\dot{E}$. The most plausible scenario to qualitatively explain this is the geometry: one needs a small $\alpha$ (to have wide radio pulses) and a small $|\beta|$ (to cross the radio beam and thus detect the radio pulsar), implying a small $\zeta$ (which also, crucially, ensures missing the $\gamma$-ray emission concentrated at the equator, making the pulsar invisible in $\gamma$ rays). This corresponds exactly to our independent arguments used above to constrain the geometry of PSR J0855-4644. @G14 have furthermore argued for small $\zeta_{\rm PSR}$ in the case of nearby, energetic millisecond pulsars that are undetected in $\gamma$ rays, suggesting that the geometric arguments made above hold for the entire pulsar population. Summary & Future work {#sec:conc} ===================== We have carried out a detailed spatial and spectral study of the compact PWN and X-ray counterpart of the [fast-spinning]{} and energetic pulsar [[PSRJ0855$-$4644]{}]{}. The pulsar is faint and soft in the X-rays (accounts for $\sim$ 1% of the total emission) surrounded by a brighter, hard and compact PWN of about 0.06 pc in extent. The PWN is elongated along a symmetry axis composed of two lobes more or less equidistant from the pulsar in the middle. The morphology is compatible with either a double-torus morphology although other models (double-torus and jets or jets only) can not be ruled out in the present observation. Spatial modelling of the two lobes of the PWN with a double-torus model captures the gross structure of the nebula and provides an estimate of the observer angle of the pulsar $\zeta_{\rm PSR}$ &lt; $37\dg$. Spectral analysis confirms the hard nature of the PWN spectrum ($\Gamma = 1.12\pm0.25$) with no changes in spectral parameters detected between the outer and inner PWN, or the east and west lobe in the present observation. The pulsar spectrum is compatible with that of a blackbody of $T_{\rm BB} = 2.0 \times 10^{6}$ K and effective radius $R_{\rm eff}\sim$ 1.5 km, suggesting that the emission originates from the hot spots on the poles of the neutron star. We find a consistent scenario emerging from constraints on the geometry of the system, derived independently from the X-ray morphological modelling of the nebula and from the radio / $\gamma$-ray light curve modelling. These constraints are compatible with values of $\alpha\lesssim55\dg$ and $\zeta_{\rm PSR}\lesssim55\dg$, and $10\dg\lesssim|\beta|\lesssim30\dg$, (additionally, we find a best-fit radio profile at $(\alpha,\zeta_{\rm PSR}) = (22\dg, 8\dg)$), which implies that we are viewing the pulsar relatively close to one of its magnetic poles, and relatively close to the spin axis of the pulsar. In this scenario, we do not expect to see the $\gamma$-ray pulsations which is supposed to originate from the outer magnetosphere [e.g., @Grenier15] from caustics mostly located near the equatorial regions at large $\zeta_{\rm PSR}$. This proposed scenario (which we do not claim to be the only plausible one) would explain why, despite the high $\dot{E}$/$d^{2}$ value for this system, no $\gamma$-ray pulsations have been detected so far with the *Fermi*-LAT telescope. For co-located regions of non-thermal X-ray and $\gamma$-ray emission, this is also consistent with the lack of non-thermal X-rays from the pulsar. [[PSRJ0855$-$4644]{}]{} is the most energetic pulsar after Vela in the nearby environment (distance $<$ 1 kpc), and therefore has a great appeal in studying the morphology in detail and enhancing our understanding of the PWNe. A future [[*Chandra *]{}]{} [[observation ]{}]{}will be used to further test for changes in position and brightness of the lobes with respect to the pulsar symmetry axis to disentangle the dominance of either (torus vs. jet) components. Obtaining radio polarization data to look for sweeps of the position angle will also be useful to put further constraints on $\beta$. Finally, X-ray pulsations and the light curve shape may lead to independent constraints on the pulsar geometry and mass-to-radius ratio [e.g., @Bogdanov13_J0437]. We gratefully acknowledge fruitful discussions with Alice Harding. This work is based on the research supported wholly / in part by the National Research Foundation of South Africa (Grant Numbers 87613, 90822, 92860, 93278, and 99072). The Grantholder acknowledges that opinions, findings and conclusions or recommendations expressed in any publication generated by the NRF supported research is that of the author(s), and that the NRF accepts no liability whatsoever in this regard. [^1]: Present address: Max-Planck-Institut f[ür]{} extraterrestrische Physik, Giessenbachstra[ß]{}e, 85748 Garching, Germany [^2]: http://cxc.harvard.edu/ciao/ [^3]: http://cxc.harvard.edu/sherpa4.4/ [^4]: http://heasarc.gsfc.nasa.gov/xanadu/xspec/ [^5]: http://cxc.harvard.edu/chart/ [^6]: http://cxc.harvard.edu/chart/threads/marx/ [^7]: Position angle 0 corresponds to North and 90 to East, observer angle $\zeta_{\rm TORUS}$ between the torus axis and the observer line of sight, radius $R$ and a finite thickness or “blur” $\delta$ of the cross section of the torus, and the post-shock flow velocity $\beta_{\rm shock}$. For the double torus there is an additional parameter *d* for the separation between the two torii. It is symmetrically offset along the torus axis with the pulsar positioned in the middle. The torus axis is coincident with the pulsar spin axis, $\zeta_{\rm TORUS}=\zeta_{\rm PSR}$ and so $\zeta_{\rm TORUS}$ will be referred as $\zeta_{\rm PSR}$ in the rest of the paper. [^8]: The Cash statistic is a likelihood function defined using Poisson statistics. Probability P$_{x,y}$ of observing d$_{x,y}$ counts out of the expected c$_{x,y}$ counts in pixel (x,y) is $P_{x,y}=\frac{c_{x,y}e^{-c_{x,y}}}{d_{x,y}!}$ . It may be used regardless of the number of counts in each pixel and passes to $\chi^{2}$ statistics at c$_{x,y} \ge 20$. The best-fit parameters are determined by the negative logarithm of the likelihood function summed over all image pixels. [^9]: This relative flux is not seen for the radio pulses, since the radio code predicts the same maximum conal flux from a single altitude, based on $P$, $\dot{P}$, and $\nu$ at $\theta = \bar{\theta}$, independent of $\alpha$ and $\zeta_{\rm PSR}$, as mentioned earlier. [^10]: Absolute predictions of flux, and therefore relative $\gamma$-ray vs. radio flux, are only possible when using full radiative models, which is outside the scope of the current paper. [^11]: We note that while the geometric models assume constant emissivity in the corotating frame, it is very likely that there will be an azimuthal dependence of this quantity in more detailed models, given such a dependence of the electric field. This should terminate some of the low-level (off-peak) TPC radiation at small $\alpha$ and $\zeta_{\rm PSR}$. More physical models which self-consistently modulates the emissivity via an electric field and associated particle transport and radiation calculations is beyond the scope of the current geometric models. [^12]: It is also true that should (weak) $\gamma$-ray pulsations be detected in future, this will imply quite large values for $\alpha$ and $\zeta_{\rm PSR}$, while the value of $|\beta|$ should remain small.
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this short survey article, we aim to provide an up to date information on the progress made towards Schurs exponent conjecture and related conjectures. We also mention the connection between Schurs exponent conjecture and Noether’s Rationality problem.' address: | School of Mathematics, Indian Institute of Science Education and Research Thiruvananthapuram,\ 695551 Kerala, India. author: - 'V.Z. Thomas' bibliography: - 'Bibliography.bib' title: ' On Schurs exponent conjecture and its relation to Noether’s Rationality problem' --- Schur Multiplier ,group actions. 20B05 ,20D10 ,20D15 ,20F05 ,20F14 ,20F18 ,20G10 ,20J05 ,20J06 Introduction ============ Let $\operatorname{\mathit{exp}}(G)$ denote the exponent of the group $G$, which is the smallest positive integer $n$ such that $g^n=1$ for all $g\in G$, and let $H_2(G, \mathbb{Z})$ denote the second homology group with coefficients in $\mathbb{Z}$. Schur’s exponent conjecture can now be stated as \[C1\] if $G$ is a finite group, then $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z})) \mid \operatorname{\mathit{exp}}(G)$. In [@MRJ], the author mentions the above as a conjecture. So this conjecture is at least 46 years old. In [@BKW], the authors found a remarkable counterexample to this conjecture. Their counterexample involved a $2$-group of order $2^{68}$ with $\operatorname{\mathit{exp}}(G)=4$ and $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z}))=8$. In [@PM1], the author mentions another counterexample to Schur’s conjecture. His counterexample involves a $2$-group of order $2^{11}$ with $\operatorname{\mathit{exp}}(G)=4$ and $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z}))=8$. All the counterexamples produced so far involve $2$ groups. During the write up of this survey, the author was working on groups of exponent $5$. The author could manage to prove the conjecture for $R(2,5)$ and $R(3,5)$, where $R(d,5)$ denotes the largest finite quotient of the $d$ generator Burnside group $B(d,5)$ of exponent $5$. For $R(4,5)$, it is known that the nilpotency class is at most 24 (cf. [@HNV]). I felt, I could prove Schur’s conjecture if the nilpotency class of $R(4,5)$ was less than 24, but I only managed to prove that $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z})) \mid (\operatorname{\mathit{exp}}(G))^2$. By this point, I had started to feel that the conjecture may not be true for $R(d,5)$, where $d>4$. To decide whether to continue my approach further or not, I contacted M. R. Vaughan-Lee asking him if the nilpotency class of $R(4,5)$ is less than 24. He replied saying that it is an interesting question, but it is not known and perhaps very difficult to decide, but most probably it is less than 24. I explained to him why I was interested in the nilpotency class of $R(4,5)$ and its connection to Schur’s conjecture and mentioned that one could expect a counterexample for groups of exponent 5 and shared a preliminary version of this survey article with him. He got interested in the problem and found a remarkable counterexample to Schur’s conjecture. His counterexample involves a $4$ generator group $G$ of exponent $5$, nilpotency class 9, but the $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z}))=25$. The order of this group is $5^{4122}$. The group $G$ is defined as follows. He first defines a four generator group $H$ with the presentation $<a,b,c,d|[b,a]=[d,c]>$. Then he defines $G$ to be the largest quotient of $H$ with exponent $5$ and nilpotency class $9$. He also obtained another counterexample of a $4$ generator group of order $3^{11983}$, nilpotency class $9$ and exponent $9$, with $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z}))=27$. Much before these counterexamples were produced by Vaughan-Lee, the author of [@PM5] conjectures the following \[C2\] if G is a finite group, then $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z})) \mid (\operatorname{\mathit{exp}}(G))^2$. The authors of [@APT] conjecture that, \[C3\] if G is a finite p-group, then $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z})) \mid p\ \operatorname{\mathit{exp}}(G)$. The authors of [@APT] conjecture this at least 6 months prior to the counterexample produced by Vaughan-Lee. Clearly the counterexamples for Schur’s conjecture given by the authors of [@BKW] and [@PM1] are not counterexamples for Conjecture \[C3\]. The counterexamples produced by Vaughan-Lee to Schur’s conjecture are not counterexample to Conjecture \[C3\]. If Conjecture \[C3\] holds, then using a standard argument given in Theorem 4, Chapter IX of [@JPS], it will follow that $\operatorname{\mathit{exp}}(H_2(G,\mathbb{Z})) \mid (\operatorname{\mathit{exp}}(G))^2$ for any finite group $G$.\ For the benefit of the readers, without further ado, we collect the classes of groups for which the above conjectures have been proved. After that we will give a brief outline of the strategy used to prove the above conjectures. - Groups for which the Bogomolov multiplier is trivial. ( cf. [@KK] for $\Sh$-rigid groups and triviality of Bogomolov multiplier for such groups.) - If the commuting probability of a finite group $G$ is greater than 0.25. (cf. [@JM]) - Central extensions of most finite simple groups. (cf. [@DLT1]) - $p$-groups of nilpotency class less than or equal to $3$ for all primes $p$. (cf. [@PM2]) - Powerful $p$-groups. (cf. [@LM], [@APT]) - $p$-groups of maximal class. (cf. [@PM6]) - Potent $p$-groups (cf. [@PM4], [@APT]) - $p$-groups of nilpotency class less than or equal to $5$ for an odd prime $p$. (cf. [@APT]) - $p$-groups of nilpotency class less than or equal to $p$ for an odd prime $p$. (cf. [@APT]) - Metabelian $p$-groups of exponent $p$. (cf. [@PM1], [@APT]) - $p$-central metabelian $p$ groups. (cf. [@APT]) - An odd $p$-group with an abelian Frattini subgroup. (cf. [@AT]) - A $p$-group such that the commutator subgroup of $G$ is cyclic. (cf. [@AT]) - Finite metacyclic groups. (cf. [@AT]) - Abelian by cyclic groups. (cf. [@AT]) - $3$-Engel groups. (cf. [@PM7]) - $4$-Engel groups of exponent $e$, where $e$ is not divisible by $2$ or $5$. (cf. [@PM7]) - A group $G$ with $\operatorname{\mathit{exp}}(G/Z(G)) = p$ and $p \in \{2,3\}$. (cf. [@AT]) - A finitely generated group $G$ such that $\operatorname{\mathit{exp}}(G/Z(G)) = 6$. (cf. [@AT]) <!-- --> - A group of exponent $e$, where $e\leq 7$. (cf. [@NS2]) - An odd solvable group of exponent $p$ and derived length $3$. (cf. [@AT]) - A $p$-group having an abelian normal subgroup $N$ of index $p^l$, $p \neq 2$ and $l$ is less than max $\{7,p+2\}$. (cf. [@AT]) - An odd $p$-group whose commutator subgroup is powerful. (cf. [@AT]) - An odd $p$-group $G$ such that $\gamma_{p+1}(G)$ is powerful. (cf. [@AT]) <!-- --> - An odd $p$-group of nilpotency class less than or equal to $7$. (cf. [@AT]) - A $p$-group $G$ such that $\operatorname{\mathit{exp}}(Z(G))=p$ and nilpotency class of $G$ is at most $p+1$. (cf. [@AT]) - A group having an abelian normal subgroup $N$ of index $p^2$, and $p \neq 2$. (cf. [@AT]) - A $2$-group such that the frattini subgroup of $G$ is abelian. (cf. [@AT]) - An odd $p$-group whose frattini subgroup is powerful. (cf. [@AT]) - A $p$-group $G$ with $\operatorname{\mathit{exp}}(G/Z(G)) = p$ and $p \leq 7$. (cf. [@AT]) Common Strategies ================= In this section, we briefly explain the common strategies used to prove the Theorems in the previous section. One of the tools used is a construction introduced by R. Brown and J.-L. Loday in [@BL1] and [@BL2], called the nonabelian tensor product $G\otimes H$. The nonabelian tensor product of groups is defined for a pair of groups that act on each other provided the actions satisfy the compatibility conditions of Definition \[D:1.1\] below. Note that we write conjugation on the left, so $^gg'=gg'g^{-1}$ for $g,g'\in G$ and $^gg'\cdot g'^{-1}=[g,g']$ is the commutator of $g$ and $g'$. \[D:1.1\] Let $G$ and $H$ be groups that act on themselves by conjugation and each of which acts on the other. The mutual actions are said to be compatible if $$^{(^h g)}h_1=\; (^h(^g(^{h^{-1}}h_1))) \;and\; ^{(^g h)}g_1=\ (^g(^h(^{g^{-1}}g_1))) \;\mbox{for \;all}\; g,g_1\in G, h,h_1\in H.$$ \[D:1.2\] If $G$ and $H$ are groups that act compatibly on each other, then the **nonabelian tensor product** $G\otimes H$ is the group generated by the symbols $g\otimes h$ for $g\in G$ and $h\in H$ with relations $$\label{E:1.1.1} gg'\otimes h=(^gg'\otimes \;^gh)(g\otimes h),$$ $$\label{E:1.1.2} g\otimes hh'=(g\otimes h)(^hg\otimes \;^hh'),$$ for all $g,g'\in G$ and $h,h'\in H$. The nonabelian tensor square, $G\otimes G$, of a group $G$ is a special case of the nonabelian tensor product of a pair of groups $G$ and $H$, where $G=H$, and all actions are given by conjugation. There exists a homomorphism $\kappa : G\otimes G \rightarrow G^{\prime}$ sending $g\otimes h$ to $[g,h]$. Let $\nabla (G)$ denote the subgroup of $G\otimes G$ generated by the elements $x\otimes x$ for $x\in G$. The exterior square of $G$ is defined as $G\wedge G= (G\otimes G)/\nabla (G)$ and denote the induced homomorphism again by $\kappa : G\wedge G \rightarrow G^{\prime}$. Let $M(G):=\ker ( G\wedge G \rightarrow G' )$. It has been shown in [@CM] that $M(G)\cong H_{2}(G, \mathbb{Z})$. Now we outline most common methods to prove the previous Theorems stated in the Introduction. - [**Strategy 1**]{} Try to find the exponent of $G\wedge G$. If the exponent of $G\wedge G$ divides the exponent of $G$, then Schurs Conjecture holds true, since $M(G)$ is a subgroup of $G\wedge G$. - [**Strategy 2**]{} The second strategy is almost similar to the first one. Instead of estimating the exponent of $G\wedge G$, one estimates the exponent of the commutator subgroup of a covering group of $G$, thanks to Schur (cf. [@GK]), the commutator subgroups of any two covering groups are isomorphic. - [**Strategy 3**]{} One can also estimate the exponent of the Bogomolov multiplier of the group $G$. Noether’s Rationality Problem and Schur’s Exponent Conjecture ============================================================= Let $k$ be any field and $G$ be a finite group. Let $G$ act on the rational function field $k(x_g, g\in G)$ by $k$ automorphisms defined by $g.x_h:=x_{gh}$ for all $g,h\in G$. Denote by $F(G)$ the fixed field $k(x_g, g\in G)^{G}$. Noether’s problem asks whether $F(G)$ is purely transcendental over $k$. Bogomolov multiplier is the group $B_0(G)=\ker\{H^2(G, \mathbb{Q}/{\mathbb{Z}})\to \bigoplus_A H^2(A,\mathbb{Q}/{\mathbb{Z}})\}$, where $A$ runs over all Abelian subgroups of $G$. Let $M_0(G)$ be the subgroup of $G\wedge G$ generated by $x\wedge y$ such that $[x,y]=1$. It is shown in [@PM8] that for a finite group $G$, $B_0(G)$ is non-canonically isomorphic to $M(G)/M_0(G)$. More on Bogomolov multiplier and its relation to Noether’s Rationality problem can be found in [@B1], [@B2] and [@S]. Bogomolov multiplier $B_0(G)$ can be seen as an obstruction to Noether’s rationality problem. In particular, if $B_0(G)\neq 0$, then Noether’s Rationality problem has a negative answer for $G$. In this section, we show that if the Bogomolov multiplier $B_0(G)=0$, then Schur’s Conjecture \[C1\] holds. This statement gives us a connection between Noether’s Rationality problem and Schur’s exponent conjecture. Sometime in 2015, in an evening walk, Guram Donadze mentioned to the author that triviality of the Bogomolov multiplier implies Schur’s conjecture. Since the author could not find a reference to this statement, we produce it here for the benefit of the reader and also because the proof is short. \[L:3.1\] Let $G$ be a finite group. If $B_0(G)=0$, then Conjecture $\ref{C1}$ holds true. Note that triviality of $B_0(G)$ implies that $M(G)=M_0(G)$. Since $[x,y]=1$, we have that $x^n\wedge y=(x\wedge y)^n$. Taking $n$ to be the exponent of $G$ and noting that each $x\wedge y$ in $M_0(G)$ is contained in the center of $G\wedge G$, finishes the proof. Thus all counterexamples to Schur’s conjecture will provide a negative answer to Noether’s Rationality problem. Triviality of Bogomolov multiplier is not a necessary condition for Conjecture \[C1\] to hold. Most $p$ groups of maximal class have non-trivial Bogomolov multiplier (cf. [@AJ]), but Conjecture \[C1\] holds for $p$ groups of maximal class. On the other hand, the Bogomolov multiplier is trivial for all cyclic groups, but Noether’s Rationality problem does not hold for all cyclic groups. (cf. [@RGS], [@BP]) Bounds depending on nilpotency class, derived length and the rank. ================================================================== Several authors have given bounds for the exponent of the Schur Multiplier in terms of the exponent of the group, the nilpotency class, derived length and the rank of the group. The authors of [@SN] improved the bounds given by the authors of [@LM] depending on the exponent and the rank of the group. Below we list some of the bounds given in terms of the nilpotency class and the derived length of a group. - $exp(M(G))\mid (exp(G))^{\ceil{\frac{c}{2}}}$. (cf. [@GE2]) - $exp(M(G))\mid (exp(G))^{2(\floor{\log_2 c})}$. (cf. [@PM1]) - $exp(M(G))\mid (exp(G))^m$, where $m=\floor{\log_{p-1} c}+1$. (cf. [@NS2]) - $exp(M(G))\mid (exp(G))^n$, where $n = 1+\ceil{\log_{p-1} (\frac{c+1}{p+1})}$. (cf. [@APT]) Note that the last bound above is an improvement of all the previous bounds. - $exp(M(G))\mid (exp(G))^d$, if $p$ is odd. (cf. [@NS1], [@APT]) - $exp(M(G))\mid 2^{d-1}(exp(G))^{d}$, if $p=2$. (cf. [@NS1], [@APT]) - If the exponent of $G$ is $p$, then $exp(G\wedge G) \mid (exp(G))^{d-1}$, if $p$ is odd. (cf. [@AT]) - If the exponent of $G$ is $2$, then $exp(G\wedge G) \mid 2^{d-2}(exp(G))^{d-1}$. (cf. [@AT]) Many classes of groups for which Conjecture $\ref{C1}$ is true falls under the class of Regular groups. The authors of [@NS2] and [@APT] proved the following: The following statements are equivalent: - $exp(M(G))\mid exp(G)$ for all regular $p$-groups $G$. - $exp(M(G))\mid exp(G)$ for all groups $G$ of exponent $p$. By the counterexample given by Vaughan-Lee, it follows that Schur’s conjecture is not true for regular $p$-groups. Interesting Questions ===================== The following are the questions that the author finds most interesting with regards to Schurs exponent conjecture.\ [**Question 1**]{}: Is Conjecture $\ref{C1}$ true for metabelian groups?\ We know that Conjecture $\ref{C1}$ is false in general. But we do not have any counterexample of groups of nilpotency class less than $9$ to Conjecture $\ref{C1}$. So the following question is interesting.\ [**Question 2**]{}: Let $p$ be an odd prime. Is Conjecture $\ref{C1}$ true for $p$ groups of class less than or equal to $8$?
{ "pile_set_name": "ArXiv" }
--- author: - | Shenghua Liu,$^{1}$ Houdong Zheng,$^{2}$ Huawei Shen,$^{1}$ Xiangwen Liao,$^{2}$ Xueqi Cheng$^{1}$\ \ \ \ bibliography: - 'social.bib' title: 'Learning Sentimental Influences from Users’ Behaviors' ---
{ "pile_set_name": "ArXiv" }
--- author: - Rajdeep Mukherjee - Saurabh Joshi - 'John O’Leary' - Daniel Kroening - Tom Melham bibliography: - 'paper.bib' title: 'Hardware/Software Co-verification Using Path-based Symbolic Execution' ---
{ "pile_set_name": "ArXiv" }
--- abstract: 'The famous twistor construction tells us that harmonic surfaces in $S^4$ can be obtained as projections of complex legendrian curves in ${{\mathbb C}}P^3$. A mostly computational attempt to study the tropicalizations of the latter curves is presented. Several combinatorial properties of these tropical curves are proven and conjectured. The code in Macaulay2 to compute examples is supplied. The journey starts with drawing a legendrian rational cubics through three generic points and classification of legendrian curves on a quadric surface and ends with a proof of one combinatorial property of legendrian tropical curves and heuristical arguments using tropical differential forms for the second property.' address: 'CINVESTAV, Departmento de Matemáticas, Av. Instituto Politécnico Nacional, 2508, Col. San Pedro Zacatenco, México, D.F. CP 07360, MEXICO, [email protected]' author: - Nikita Kalinin bibliography: - 'legendrian.bib' title: Tropical approach to legendrian curves in CP 3 --- Tropical geometry, legendrian curves, macaulay2 Introduction ============ Inspired by Gromov-Witten invariants, one can try to count holomorphic curves under some additional restrictions. We will see what can be said about counting of legendrian curves in ${{\mathbb C}}P^3$, passing through prescribed number of generic points and one generic line. This problem goes back to I. Vainsencher; his student, Éden Amorim [@contato] also studied it, but with rather different methods (localizations) and a bit different setup (he was looking for the number of rational legendrian curves through $2d+1$ generic lines). In fact, my goal was to enlarge tropical gear, conquest new territories such as tropical differential forms, tropical distributions, etc. Finally, some combinatorial properties of tropical contact curves have been found, but the mystery remains. This paper contains the following material. In Section \[sec\_contactomorphisms\] we compute the element of $Sp(4,{{\mathbb C}})$ which brings 3 generic points to three given generic points. In Section \[sec\_quadric\] we classify all algebraic legendrian curves on a quadric surface. In Section \[sec\_macocubics\] we predict, using Macaulay2, that the set of rational legendrian cubics through three generic points is a line in the space of coefficients, and we prove this in Section \[sec\_cubicsproof\]. In Section \[sec\_codetropical\], we show how to draw the tropicalization of all legendrian cubics through three points and how to draw a particular curve in this family. In Lemma \[lem\_locally\] we prove local necessary conditions (tropical legendrian tangency property, Figure \[fig\_tangency\]) for a curve to be a tropicalization of an algebraic legendrian curve. We conjecture that the divisibility property (Definition \[def\_divisibility\], Figure \[fig\_line\]) holds for all legendrian tropical curves, demonstrate this in some particular cases and give in Section \[sec\_applicationrefined\] a sketch of a proof of this conjecture, using tropical differential forms. Complete proof will appear in [@tropforms]. Below we summarize what is known about complex legendrian curves in ${{\mathbb C}}P^3$, and mention some related questions. However, in what follows we do not use these facts, our approach is completely elementary, if not only computational. We use Macaulay2 and Mathematica. The code is incorporated into this text, so the reader can verify the results. If a proof is omitted, it means that it is a straightforward computation. The recent study of the [*complex*]{} legendrian curves (see also [@Alarcon:2014mz]), by itself, is motivated by minimal surfaces in four dimensional sphere (read [@meeks2012survey] and [@brendle2013minimal] about minimal surfaces in ${{\mathbb R}}^3$ and $S^3$, respectively). The map $(z_1,z_2,z_3,z_4)\to(z_1+jz_2,z_3+jz_4)$ from ${{\mathbb C}}^4$ to $\mathbb H^2$ yields so-called twistor (or Penrose) map $\phi:{{\mathbb C}}P^3\to\mathbb H P^1=S^4$, and Bryant has shown [@bryant] that the images of the legendrian curves in ${{\mathbb C}}P^3$ under $\phi$ are superminimal surfaces in $S^4$. Furthermore, each minimal immersion $S^2\to S^4$ can be obtained as $\phi(C)$ where $C$ is a rational legendrian curve in ${{\mathbb C}}P^3$. Then, each Riemann surface $M$ can be mapped to a legendrian curve in ${{\mathbb C}}P^3$, using two meromorphic functions $(f,g)$ on $M$. This proves that for each Riemann surface $M^2$ there exists a conformal minimal immersion $M^2\to S^4$, and such immersions are nowadays constructed mostly by this approach (see also [@vlachos2012isometric] about isometric deformation of minimal surfaces in $S^4$). The area of the image of a harmonic map $f:S^2\to S^4$ equals $4\pi d$ if $f(S^2)$ comes as $\phi(C)$, where $C$ is a legendrian rational curve in ${{\mathbb C}}P^3$ of degree $d$. The dimension of the space $\mathfrak M_{d,0}$ of legendrian maps ${{\mathbb C}}P^1\to {{\mathbb C}}P^3$ of degree $d$ is proven to be $2d+4$ [@harmonic],[@0686.58009],[@verdier1984applications], [@verdier1988applications] (and see [@0914.58005] for the legendrian maps ${{\mathbb C}}P^1\to{{\mathbb C}}P^{2n+1}$). This is done via studying the pairs $(f,g)$ of meromorphic functions $\big($secretly $(f,g)=(\frac{z_1}{z_2},\frac{z_3}{z_4})$$\big)$ of degree $d$ with the same ramification divisor. Up to degree six the space $\mathfrak M_{d,0}$ is a smooth complex manifold [@1255.58004]. See also a survey [@contactSurvey] and references therein about minimal immersions $S^2\to S^{2n}$. In [@complexContact](unpublished, see also [@MR1608659]) it is proven that if $d\geq g+3$, then the part of the space $\mathfrak M_{d,g}$, which consists of smooth contact curves in ${{\mathbb C}}P^3$ of degree $d$ and genus $g$, is smooth. Besides, it is proven there, that a complete intersection cannot be a contact curve. That complicates the study of the contact curves of higher genus, which was approached in [@0997.53043], [@0902.53043] [@0902.53043]. In [@0943.58007] it is proven that the dimension of $\mathfrak M_{d,g}$ is $2d-g+4$ for $d\geq \max(2g,g+2)$, and in [@0997.53043] it is proven that the dimension of each irreducible component of $\mathfrak M_{d,g}$ is between $2d-4g+4$ and $2d-g+4$, where upper bound is always attained by the totally geodesic immersions (whose image belongs to a line) and the lower bound is obtained on $\mathfrak M_{6,1}$ and $\mathfrak M_{8g+1+3k,g}$. See [@0902.53043], for further details about other possible pairs $(d,g)$ with non-trivial contact curve. All this means that for $g\geq 1$ we need to take the degree $d$ of the curve at least $6$ what is now beyond our abilities to compute. For a general overview of complex contact varieties and deformations of contact curves see [@buczinsky], [@complexContact], [@0801.14014]. Real algebraic contact structures are numerous, the questions about polynomial distributions went back to [@MR0016748; @MR0217061], see Example \[ex\_realform\]. For the works of the same spirit we mention the study of legendrian curves of minimal degree through two points with prescribed tangency [@giluch2005real] and contact curves in $PSL(2,{{\mathbb C}})$ [@contactsl2]. For an introduction to tropical geometry see [@BIMS] and [@2013arXiv1311.2360B] and references therein. Acknowledgement {#acknowledgement .unnumbered} --------------- Research is supported by the grant 168647 (Early PostDoc.Mobility) of the Swiss National Science Foundation. I would like to thank M. Shkolnikov and G. Mikhalkin for their useful suggestions. The contact structure on CP\^3 ============================== A section $\omega$ of $P(\Omega^1({{\mathbb C}}P^3))$ is said to be a [*contact holomorphic form*]{} on ${{\mathbb C}}P^3$ if $\omega\wedge d\omega$ is nowhere zero. The form $\omega =ydx-xdy+wdz-zdw$ is contact. Indeed, consider the restriction of $\omega$ to the chart $w=1$. We have $$\omega_{w=1}=u=dz+ydx-xdy, u\wedge du = -2dx\wedge dy\wedge dz\ne 0,$$ similar formulae hold in other charts. Each contact holomorphic form $\omega$ on ${{\mathbb C}}P^3$ is of the type $$(py-qz+aw)dx+(-px+rz+bw)dy+(qx-ry+cw)dz+(-ax-by-cz)dw \label{eq_contactform}$$ where $a,b,c,p,q,r$ are constants and $pc+qb+ra\ne 0$. Furthermore, all such forms are equivalent under the $GL(4,{{\mathbb C}})$ action. We only sketch the proof. Let $\alpha$ be a holomorphic contact form in ${{\mathbb C}}P^3$. The form $\alpha\wedge d\alpha$ gives a section of the canonical bundle. Considering transition function for $\alpha$ we conclude that $c_1({{\mathbb C}}P^3)=2c_1(\alpha)$ (the first Chern class of the linear bundle generated by $\alpha$). It means that if $Pdx+Qdy+Rdz$ is a contact form in the chart $w=1$, then it extends to the whole ${{\mathbb C}}P^3$ only if the transition function to another charts have $w$ in denominator in degree at most two. Therefore $P,Q,R$ are polynomials of degree one. The explicit form of all such polynomials that $\omega$ is contact follows from the direct computation. On the other hand, there are many algebraic contact structures on ${{\mathbb R}}P^3$. \[ex\_realform\] Consider $$\omega' =(yz^2+yw^2)dx+(-xz^2-xw^2)dy+(x^2w+y^2w+w)dz+(-x^2z-y^2z-z)dw, \text{ or}$$ $$\begin{aligned} \omega'=& (x^2y+y^3+yz^2+yw^2)dx-(x^3+xy^2+xz^2+xw^2)dy\\ &+(x^2w+y^2w+z^2w+w^3)dz-(x^2z+y^2z+z^3+zw^2)dw.\end{aligned}$$ Then, a small perturbation of the coefficients of a real contact structure doesn’t affect the fact $\omega\wedge d\omega \ne 0$, hence we can arbitrarily perturb the coefficients of $\omega'$. At the same time, it seems to be hard to enumerate real algebraic curves which are contact with respect to these contact structures. Any irreducible algebraic curve $C\in {{\mathbb C}}P^3$ of degree at least three is legendrian with at most one holomorphic contact structure. Indeed, when we intersect the distribution given by with the distribution given by $\omega=yxd-xdy+wdz-zdw$, we obtain a vector field $v$ almost everywhere (except finite collection of points as the code below shows). On the other hand, we know that there is a line, tangent to the obtained distribution, through each point in ${{\mathbb C}}P^3$. Therefore the integral curves for $v$ are lines almost everywhere. Hence, the only locus where a curve, tangent to both distribution, can leave, is the set where these two contact forms coincide, i.e. a finite collection of points. use QQ[p,q,r,a,b,c,x,y,z,w] a1=p*y-q*z+a*w a2=-p*x+r*z+b*w a3=q*x-r*y+c*w a4=-a*x-b*y-c*z I=ideal(a1*x-a2*y, a2*w-a3*x,a3*z-a4*w) C=minimalPrimes I J=C_8 --all the other ideals C_0,C_1,... give lines if we fix a,b,c,p,q,r dim J --=7, 7-6=1, because we have 6 parameters p,q,r,a,b,c -- so it is just a collection of points. The global Reeb vector field for the contact structure $\omega=ydx-xdy+wdz-zdw$ is given by $y\frac{\partial}{\partial x} -x\frac{\partial}{\partial y}+w\frac{\partial}{\partial z} - z\frac{\partial}{\partial w}$. Its trajectories (i.e. the fibers of the Penrose map) are $$\varphi(t)=\left(A\frac{(t^2-1)}{(t^2+1)},2A\frac{t}{(t^2+1)},B\frac{((t+k)^2-1)}{((t+k)^2+1)},2B\frac{(t+k)}{((t+k)^2+1)}\right)$$ and $(\frac{t^2-1}{t^2+1})'=\frac{4t}{t^2+1},(\frac{2t}{t^2+1})'=\frac{t^2-1}{t^2+1}$. So, the Reeb vector field just rotates in $xy$ plane and $zw$ plane on the same angle. For each fixed angle this gives a linear transformation. Contact form automorphisms -------------------------- It is known that the group of automorphisms of ${{\mathbb C}}P^3$ which preserve the form $\omega = ydx-xdy+wdz-zdw$ is[^1] the symplectic group $Sp(4,{{\mathbb C}})$. The dimension count gives $\dim Sp(4,{{\mathbb C}})=10$ and $\dim PGL(4,{{\mathbb C}})=15$, what agrees with the fact the set of all contact structures Eq.  is five-dimensional. \[prop\_generators\] We list the set of generators of this group $Sp(4,{{\mathbb C}})$. 1\) $x\to x+\lambda y, \begin{pmatrix} 1 & \lambda & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ \end{pmatrix}$ 2\) $x\to y, y\to -x, \begin{pmatrix} 0 & 1 & 0 & 0 \\ -1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ \end{pmatrix}$ 3\) $x\to z, y\to w, \begin{pmatrix} 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ \end{pmatrix}$ 4\) $x\to x+\lambda w, z\to z+\lambda y, \begin{pmatrix} 1 & 0 & 0 & \lambda \\ 0 & 1 & 0 & 0 \\ 0 & \lambda & 1 & 0 \\ 0 & 0 & 0 & 1 \\ \end{pmatrix} $ 5\) $x\to\lambda x,y\to y/\lambda, \begin{pmatrix} \lambda & 0 & 0 & 0 \\ 0 & 1/\lambda & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ \end{pmatrix}$ The restriction of a contact structure on a plane $z=w=0$ is $pydx-pxdy=0$ by an easy computations. Therefore the vector field, generated by the contact form, at a point $(x,y)$ equals to a vector $(x,y)$, so the only integral curves are the lines passing through the origin. Since all the planes are equivalent under the action of $GL(4,{{\mathbb C}})$, all the planar contact curves are collections of lines. Let us choose an arbitrary plane $L$. Each contact curve in $L$ is a collection of lines through a point $p\in L$. Moreover, $L$ is the contact plane at $p$, i.e. $L$ is the zero set of $\omega$ computed at this point $p$. Macaulay2 code for the action of contactomorphism group on triplets of points {#sec_contactomorphisms} ----------------------------------------------------------------------------- All the elements of $Sp(4,{{\mathbb C}})$ which preserve $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$ are of the form $$\mathrm{Stab^3_\mu}:x\to x, y \to y+\mu(z-x), z\to z, w\to w-\mu(z-x), \begin{pmatrix} 1 & 0 & 0 & 0 \\ -\mu & 1 & \mu & 0 \\ 0 & 0 & 1 & 0 \\ \mu & 0 & \mu & 1 \\ \end{pmatrix} \label{eq_fixator}$$ Direct calculation. It is easy to send any point of ${{\mathbb C}}P^3$ to $(0,0,0,1)$ by an element of $\mathrm{Sp}(4)$. Then, points of ${{\mathbb C}}P^3$ can be divided in two classes: those, lying on the plane $L$ through $(0,0,0,1)$ such that $\omega((0,0,0,1))|_{L}=0$ and all the others. The subgroup of $\mathrm{Sp}(4)$, stabilizing $(0,0,0,1)$ acts on both these classes transitively. Now, consider a point $p$ which is not on the contact planes through $(0,0,0,1)$ and $(1,1,1,1)$. We will prove that there exists an element in the subgroup of $\mathrm{Sp}(4)$ stabilizing $(0,0,0,1)$ and $(1,1,1,1)$ that sends $p$ to $(-1,1,-1,1)$. \[lem\_transitive\] The group $\mathrm{Sp}(4)$ is [*generically 3-transitive*]{}, i.e. every three points $p_1,p_2,p_3\in {{\mathbb C}}P^3$ in general position can be sent to every three points $q_1,q_2,q_3\in {{\mathbb C}}P^3$ in general position by an element $a\in \mathrm{Sp}(4,{{\mathbb C}})$. In general, the set $\{a\in \mathrm{Sp}(4, {{\mathbb C}})|a(p_i)=q_i,i=1,2,3\}$ is of dimension one. The following code in Macaulay2 produces a matrix preserving the contact form $\omega$, which brings generic points (a1,b1,c1,1),(a2,b2,c2,1),(a3,b3,c3,1) to (0,0,0,1),(1,1,1,1) and (-1,1,-1,1). The group of contactomorphisms is generated by the matrices A,B,C,D (see Proposition \[prop\_generators\] and below formulae) and by straightforward combination of them we arrive to the answer. We present here the whole list function that we need further. newPackage("Contact") export isContact export cubicSurface export transformation export inverseMatrix --check if the functions x(s),y(s),...give a contact curve isContact = (x,y,z,w,s)-> (a:= y*diff(s,x)-x*diff(s,y)+w*diff(s,z)-z*diff(s,w)) -- surface through all legendrian cubics cubicSurface=(x,y,z,w)->(-2*x^3-21*x^2*z+27*y^2*z-60*x*z^2-25*z^3+54*y*z*w+27*z*w^2) --generators of the contactomorphic group,their action on a matrix M -- x->x+lambda y A=(lambda,M)->(matrix{{1,lambda,0,0},{0,1,0,0},{0,0,1,0},{0,0,0,1}}*M) --x->y,y->-x B=M->(matrix{{0,-1,0,0},{1,0,0,0},{0,0,1,0},{0,0,0,1}}*M) --x<->z,y<->w C=M->(matrix{{0,0,1,0},{0,0,0,1},{1,0,0,0},{0,1,0,0}}*M) --x->x+ lambda w, z-> z+lambda y D=(lambda,M)->(matrix{{1,0,0,lambda},{0,1,0,0},{0,lambda,1,0},{0,0,0,1}}*M) --special contactomorphism, which preserve (0,0,0,1),(1,1,1,1) and --brings (x,y,z,w) to (-1,1,-1,1) Q=QQ[x,y,z,w]; c:= (x-y+z-w)/(4*z); a:=-(y+z)/(2*z); b:=(-x+y+z-w)/(2*(y-z)); M:=matrix{{a+1,b+c-a+a*b-1,-b*(a+1),0},{a,c+a*(b-1),-b*a,0},{0,0,c,0},{0,b,c-1-b,1}}; M0:=inverse(M); den:= denominator M0_(0,0); M0=den*M0; M0=lift(M0,Q); --this function gives the matrix for the contactomorphism which --brings (a1,b1,c1,1)(a2,b2,c2,1)(a3,b3,c3,1) to -- (0,0,0,1),(1,1,1,1) and (-1,1,-1,1) transformation = (a1,b1,c1,a2,b2,c2,a3,b3,c3,R)->( use R; T1:=matrix{{a1,a2,a3},{b1,b2,b3},{c1,c2,c3},{1,1,1}}; T2:=C(T1); lambda1 := -(entries(T2)_0)_0; T3:=C(A(lambda1,T2)); lambda2:=-(entries(T3)_0)_0/(entries(T3)_0)_1; T4:=B(A(lambda2,T3)); lambda3:=-(entries(T4)_0)_0/(entries(T4)_0)_3; T5:=D(lambda3,T4); T6:=B(C(T5)); lambda4 :=-((entries(T6)_1)_0 + (entries(T6)_1)_1)/(entries(T6)_1)_1; T7:=C(B(A(lambda4,T6))); lambda5 := ((entries(T7)_1)_2-(entries(T7)_1)_0)/(entries(T7)_1)_1; T8:=B(A(lambda5,T7)); lambda6 := ((entries(T8)_1)_2-(entries(T8)_1)_0)/(entries(T8)_1)_1; T9:=A(lambda6,T8); T9=lift(liftit(T9,R),R); (p1,p2,p3,p4):=((entries(T9)_2)_0,(entries(T9)_2)_1,(entries(T9)_2)_2,(entries(T9)_2)_3); M1:=sub(M0,{x=>p1,y=>p2,z=>p3,w=>p4}); revers:=XX->( YY:=A(-lambda4,B(B(B(C(A(-lambda5,B(B(B(A(-lambda6,(M1*XX))))))))))); C(A(-lambda1,C(A(-lambda2,B(B(B(D(-lambda3,C(B(B(B(YY)))))))))))) ); SR:=revers(matrix{{1,0,0,0},{0,1,0,0},{0,0,1,0},{0,0,0,1}}); SR=lift(liftit(SR,R),R); SR) inverseMatrix=(SR,S)->(t:=promote(SR,frac S); t=inverse(t); lift(liftit(t,S),S)) liftit = (SR2,R)->( SR:=SR2; li:=(a,b)->( c:=b; try lift(a,R) then c=c else (den:=denominator a; c=den*b); c); SR=li((SR_(0,0)),SR); SR=li((SR_(0,1)),SR); SR=li((SR_(0,2)),SR); SR=li((SR_(1,0)),SR); SR=li((SR_(1,1)),SR); SR=li((SR_(1,2)),SR); SR=li((SR_(2,0)),SR); SR=li((SR_(2,1)),SR); SR=li((SR_(2,2)),SR); SR=li((SR_(3,0)),SR); SR=li((SR_(3,1)),SR); SR=li((SR_(3,2)),SR); if (numgens source SR)==4 then( SR=li((SR_(0,3)),SR); SR=li((SR_(1,3)),SR); SR=li((SR_(2,3)),SR); SR=li((SR_(3,3)),SR);); SR) end Note that we are going to apply this for the computations with Puiseux series. The direct approach will consume too much memory, that is why in the code above we do everything by small steps. Curves on a hypersurface of degree two {#sec_quadric} -------------------------------------- Consider a contact form $\omega$ as in . We are going to find the restriction of $\omega$ on the surface $X$, given by $$\label{eq_quadric} \{xy-zw=0\}=Im(f\colon{{\mathbb C}}P^1\times {{\mathbb C}}P^1\to {{\mathbb C}}P^3), f\colon(\mu:\mu'),(\nu:\nu')\to (\mu\nu',\mu'\nu,\mu\nu,\mu'\nu').$$ Note, that any irreducible hypersurface $X'$ of degree 2 in ${{\mathbb C}}P^3$ is projectively equivalent to $X$, therefore in this way we will describe all the legendrian curves on all the non-degenerate hypersurfaces $X'$ of degree $2$. Computing in the affine chart $(\mu,\nu)\to (\mu,\nu,\mu\nu,1)\in X$, we obtain $$f_*:\frac{\partial}{\partial \mu}\to \frac{\partial}{\partial x} +y \frac{\partial}{\partial z}, \frac{\partial}{\partial \nu}\to \frac{\partial}{\partial y} +x \frac{\partial}{\partial z}.$$ The fact that $\omega (f_*(M\frac{\partial}{\partial \mu}+N\frac{\partial}{\partial \nu})=0$ at $ (\mu,\nu,\mu\nu,1)$ is equivalent to $$(p\nu-q\mu\nu+a)M+(-p\mu+r\mu\nu+b)N+(q\mu-r\nu+c)(M\nu+N\mu)=0, \text{\ i.e.}$$ $$M(p\nu +a-r\nu^2+c\nu)+N(-p\mu+b+q\mu^2+c\mu)=0.$$ If a curve is locally of type $(\mu(t),\nu(t))$, then its tangent vector is given by the formula $\mu'\frac{\partial}{\partial \mu}+\nu'\frac{\partial}{\partial\nu}$. But this, after reparametrization, means that $$\label{diffur} \frac{d\mu}{dt} = (c-p)\mu+b+q\mu^2, \frac{d\nu}{dt} = -((p+c)\nu +a-r\nu^2).$$ We are looking for the algebraic leafs of this foliation. See [@zbMATH05995244; @zbMATH01602924] for details about space of foliations with algebraic leafs, [@zbMATH05555269] for the classification of the quadratic systems with the first integral. Consider the curve $(t,t^2,t^3,1)$ which lies on the hypersurface $\{xy-zw=0\}$. It is legendrian with respect to the form $3dx-3dy+wdz-zdw=0$, so we put $p=3,c=1,q=a=r=b=0$ and becomes $(\mu',\nu') = (-2\mu,-4\nu)=(\mu,2\nu)$, hence $\mu = e^t,\nu=e^{2t}$ which is the same as $(\mu,\nu)=(t,t^2)$, and subsequently $\mu\nu=t^3$. Depending on the coefficients, each equation $\frac{dx}{dt}=c_0+c_1x+c_2x^2$ after a linear change of the coordinates (over complex numbers) becomes one in the following list: - $\frac{dx}{dt}=c,$ - $\frac{dx}{dt}=cx,$ - $\frac{dx}{dt}=cx^2,$ - $\frac{dx}{dt}=c(x^2-1)$. If $\frac{d\mu}{dt}=c_0(\mu^2-1),\frac{d\nu}{dt}=c_1(\nu^2-1)$, then $\frac{d\mu}{\mu^2-1}=c_3\frac{d\nu}{\nu^2-1}$. That implies $\log(\frac{\mu-1}{\mu+1})=c_4\log(\frac{\nu-1}{\nu+1})+c_5$, and finally $c_6(\frac{\nu-1}{\nu+1})=c_7(\frac{\mu-1}{\mu+1})^{d_1}$ which is algebraic if $d_1\in{{\mathbb Q}}$. To the contrary, the case $\frac{d\mu}{dt}=c_0(\mu^2-1), \frac{d\nu}{dt}=c_1\nu^2$ always gives a non-algebraic curve if $c_0c_1\ne 0$ because this gives an equation of the type $\frac{\mu-1}{\mu+1}=e^{1/\nu}$. \[th\_quadricsurface\] After a linear change of coordinates $\tilde\mu=c_0+c_1\mu,\tilde\nu=c_2+c_3\nu$ any legendrian curve on the quadric $xy-zw=0$ with parametrization can be written in one of the following standard forms : - $c_0(\frac{\nu-1}{\nu+1})^{d_1}=c_1(\frac{\mu-1}{\mu+1})^{d_2}$, - $c_0\nu^{d_1}=c_1\mu^{d_2}$, - $c_0(\frac{\nu-1}{\nu+1})^{d_1}=c_1\mu^{d_2}$ , - $c_0(\frac{\mu-1}{\mu+1})^{d_1}=c_1\nu^{d_2}$, - $c_0\mu\nu+c_1\mu+c_2\nu=0$, - $c_0\mu+c_1\nu+c_2=0$, - $c_0\mu\nu+c_1\mu+c_2=0$, - $c_0\mu\nu+c_1\nu+c_2=0$, - $\mu=c_0$, - $\nu=c_0$, where $c_i\in{{\mathbb C}},d_i\in {{\mathbb N}}_0$ are some constants. \[pr\_quadrics\] Given this classification one might count the legendrian curves of given degree and genus lying in a quadric. For example, all rational quartics lie on a quadric surface. Legendrian curves of degrees one and two ---------------------------------------- A map $f:M\to {{\mathbb C}}P^3$ is [*totally geodesic*]{} if $f(M)$ is a legendrian line. Let us study the rational legendrian curves of degrees one and two. In the case $\deg x,y,z=1$ or $2$, it happens that such curves are parametrized by $(f,p + qf, r + pf)$, where $f$ is a polynomial of degree 1 or 2. Consider a general line $l = (a_0+b_0s,a_1+b_1s,a_2+b_2s,a_3+b_3s)$ in ${{\mathbb C}}P^3$. Putting it into the contact form we conclude that the line $l$ is legendrian iff $a_1b_0-a_0b_1+a_3b_2-a_2b_3=0$. It means that for a point $A$ we have 1-dimensional family of legendrian lines through $A$, this family is just the contact plane through $A$. Therefore the number of legendrian lines through 1 point and 1 line equals 1. Let us observe one important property of legendrian lines. One can think about a line $l$ in ${{\mathbb C}}P^3$ as four section $x,y,z,w$ of $\mathcal O(1)$ on ${{\mathbb C}}P^1$. Let $X,Y,Z,W$ be the roots of $x,y,z,w$, $x=a_0+b_0s, X=-\frac{a_0}{b_0}$, $y=a_1+b_1s, Y = -\frac{a_1}{b_1}$, etc. The following three conditions are equivalent: \[prop\_balance\] - the line $l$ is legendrian,\ - $y(X)/z(X) = w(Z)/x(Z)$,\ - $x(Y)/w(Y)=z(W)/y(W)$.\ Look at table with values of $x,y,z,w$ in $X,Y,Z,W$. $$\begin{pmatrix} X=&(0,&\frac{a_1b_0-b_1a_0}{b_0},&\frac{a_2b_0-b_2a_0}{b_0},&\frac{a_3b_0-b_3a_0}{b_0})\\ Y=&(\frac{a_0b_1-a_1b_0}{b_1}, &0,&\frac{a_2b_1-a_1b_2}{b_1},&\frac{a_3b_1-a_1b_3}{b_1})\\ Z=&(\frac{a_0b_2-a_2b_0}{b_2},&\frac{a_1b_2-a_2b_1}{b_2},&0,&\frac{a_3b_2-a_2b_3}{b_2})\\ W=&(\frac{a_0b_3-a_3b_0}{b_3},&\frac{a_1b_3-a_3b_1}{b_3},&\frac{a_2b_3-a_3b_2}{b_3},&0)\\ \end{pmatrix}$$ \[pr\_weil\] Is it possible to generalize this theorem for the curves of higher degree? As we see in Example \[ex\_tropicalcubic4\], the straightforward generalization does not take place. Legendrian cubics via Macaulay2 {#sec_macocubics} ------------------------------- Let us find all the legendrian cubics passing through three generic points in ${{\mathbb C}}P^3$. We parametrize our curve and suppose that is passes through chosen points at $t=-1,0,1$. clearAll --coefficients of the parametrization of the cubic mainvar=(a0,a1,a2,a3,b0,b1,b2,b3,c0,c1,c2,c3,d0,d1,d2,d3) R=QQ[mainvar]; P=R[s]; --polynomials for each coordinate x=a0+a1*s+a2*s*s+a3*s*s*s; y=b0+b1*s+b2*s*s+b3*s*s*s; z=c0+c1*s+c2*s*s+c3*s*s*s; t=d0+d1*s+d2*s*s+d3*s*s*s; cont = y*diff(s,x)-x*diff(s,y)+t*diff(s,z)-z*diff(s,t) --in M we have our relation for variables since in in the variable cont --(as a polynomial in z) all the coef. should be zeroes (C,M) = coefficients cont (A,B,C)=(0,1,-1) xA=sub(x,{s=>A}); xB=sub(x,{s=>B}); xC=sub(x,{s=>C}); yA=sub(y,{s=>A}); yB=sub(y,{s=>B}); yC=sub(y,{s=>C}); zA=sub(z,{s=>A}); zB=sub(z,{s=>B}); zC=sub(z,{s=>C}); tA=sub(t,{s=>A}); tB=sub(t,{s=>B}); tC=sub(t,{s=>C}); --choose random points (p11,p12,p13,p14)=(29,-6,13,11) (p21,p22,p23,p24)=(-3,-17,7,-5) (p31,p32,p33,p34)=(16,-5,6,23) --conditions that our curve passes through chosen points (i1,i2,i3)=(p14*xA-p11*tA,p14*yA-p12*tA,p14*zA-p13*tA) (j1,j2,j3)=(p24*xB-p21*tB,p24*yB-p22*tB,p24*zB-p23*tB) (k1,k2,k3)=(p34*xC-p31*tC,p34*yC-p32*tC,p34*zC-p33*tC) use R; N= M_0; l=i->lift(i,R); J = ideal(i1,i2,i3,l(N_0),l(N_1),l(N_2),l(N_3),l(N_4)) S = minimalPrimes J J0 = S_0; J1 = S_1; J2 = S_2; --S_3 does not exist di=i->dim variety i; use P; Null = ideal(x,y,z,t) --if Null is a subset of our ideal, -- it means that x,y,z,t are all zeroes at some point, -- so we are not interested in such coefficients a0,a1, ... di J0 -- 7 di J1 -- 8 that raise our suspicions... di J2 -- 7 --ideal(s-A) means evaluation at A isSubset(Null, promote(J0,P)+ideal(s-A)) -- false, isSubset(Null, promote(J1,P)+ideal(s-A)) --true, eliminate! isSubset(Null, promote(J2,P)+ideal(s-A)) --false use R; S0 = minimalPrimes (J0+ideal(j1,j2,j3)); J00=S0_0; J01=S0_1; --S0_2 do not exist use P isSubset(Null, promote(J00,P)+ideal(s-B)) --false isSubset(Null, promote(J01,P)+ideal(s-B))--true! eliminate use R; S01 = minimalPrimes (J00+ideal(k1,k2,k3)) J000=S01_0; J001=S01_1; use P; isSubset(Null, promote(J000,P)+ideal(s-C)) --false isSubset(Null, promote(J001,P)+ideal(s-C)) --true eliminate di J000--1 degree J000--1, so it is linear! --------- S2 = minimalPrimes (J2 + ideal(j1,j2,j3)) J20=S2_0 --S2_1 does not exist isSubset(Null, promote(J20,P)+ideal(s-B)) --true, eliminate! As we see, we impose all the constrains on the coefficients $a_i,b_i,c_i,d_i$ and found that the subvariety of the coefficients of legendrian cubics through three generic points is of dimension one (as expected) and of degree one (it was not expected). Legendrian cubics {#sec_cubicsproof} ----------------- Any rational non-planar cubic is equivalent to $(t,t^2,t^3,1)$. We can choose a contact form $\omega_1$ such that $(t,t^2,t^3,1)$ was legendrian with respect to it. The cubic $(t,t^2,t^3,1)$ is legendrian with respect to only one contact structure $\omega_1=3ydx-3xdy+wdz-zdw$. Direct calculation, using . We fix the contact form $w_1$, then by a contactomorphism we can bring any three generic points to the points $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$. The main result of this section is the following theorem (in the previous section we already predicted that the family of such curves should be parametrized by a line). \[th\_contact\] All the rational cubics passing through $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$ and tangent to $\omega_1=3ydx-3xdy+wdz-zdw$ are of the form $$\label{eq_contactfamily} l(t,\mu)=(t,t^2+\mu(t-t^3),t^3, 1-3\mu(t-t^3)).$$ The result of the theorem is not surprising. We have the orbit of the action of $\mathrm{Stab^3_\mu}$ (see Eq. ) on $(t,t^2,t^3,1)$. Therefore, the only problem is to show that there are no other solutions. \[cor\_answer\] For each complex contact form on ${{\mathbb C}}P^3$ the number of rational contact cubics through three generic points and a line in general position is equal to three. We intersect the family with a generic line $L$ of the type $(t', p_1 + q_1 t', p_2 + q_2t', p_3+q_3t')$. Because of the genericity, $L$ does not pass through $(0,0,0,1)=l(0,\mu)$, therefore we may suppose that at any intersection of $L$ and $l(t,\mu)$ we have $t\ne 0$. Therefore, at a point of intersection we have $t'=ct$ for some $c$, and then $$\label{eq_lineintersection} p_1 + q_1 t'=c(t^2+\mu(t-t^3)),p_2 + q_2t'=ct^3,p_3+q_3t'=c(1-3\mu(t-t^3)).$$ We have $3(p_1+q_1t')+p_3+q_3t'=c(3t^2+1)$, therefore, substituting $t'=ct$ we obtain $c=\frac{3p_1+p_3}{3t^2-3q_1t-q_3t+1}$. Then, using the first equality in , we get $\mu = \frac{p_1+q_1ct-ct^2}{c(t-t^3)}$. Then, since $c(t^3-q_2t)=p_2$, we have $t^3-q_2t=\frac{p_2}{3p_1+p_3}(3t^2-3q_1t-q_3t+1)$. Choosing $p_2,q_2$ appropriately, we see that the last equation usually has three roots. Each rational cubic curve has a parametrization of the form $$\begin{aligned} (a_0+a_1t+a_2t^2+a_3t^3,b_0+b_1t+b_2t^2+b_3t^3,c_0+c_1t+c_2t^2+c_3t^3,d_0+d_1t+d_2t^2+d_3t^3).\end{aligned}$$ We supposed that our cubic passes through points $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$ at $t=0,1,-1$ respectively. Substituting $t=0$ in the parametrization, we obtain $a_0=b_0=c_0=0,d_0=1$. Substitutions $t=\pm 1$ give us $$\begin{aligned} a_1+a_2+a_3&=b_1+b_2+b_3=c_1+c_2+c_3=1+d_1+d_2+d_3, \\ a_1-a_2+a_3&=-b_1+b_2-b_3=c_1-c_2+c_3=1-d_1+d_2-d_3.\end{aligned}$$ Therefore, $a_2=b_1+b_3=c_2=d_1+d_3, a_1+a_3=b_2=c_1+c_3=1+d_2$. Substituting indeterminates with bigger indices as functions of the indeterminates with smaller indices we obtain that our curve is parametrized by $$\begin{aligned} (a_1t+a_2t^2+&(b_2-a_1)t^3,b_1t+b_2t^2+(a_2-b_1)t^3,\\ &c_1t+a_2t^2+(b_2-c_1)t^3,1+d_1t+(b_2-1)t^2+(a_2-d_1)t^3).\end{aligned}$$ Evaluating the form $3ydx-3xdy+wdz-zdw$ on the curve we obtain $$\begin{aligned} 3(b_1t+b_2t^2+&(a_2-b_1)t^3)(a_1+2a_2t+3(b_2-a_1)t^2)\\ &-3(a_1t+a_2t^2+(b_2-a_1)t^3)(b_1+2b_2t+3(a_2-b_1)t^2)\\ &+(1+d_1t+(b_2-1)t^2+(a_2-d_1)t^3)(c_1+2a_2t+3(b_2-c_1)t^2)\\ &-(c_1t+a_2t^2+(b_2-c_1)t^3)(d_1+2(b_2-1)t+3(a_2-d_1)t^2) =0.\end{aligned}$$ The coefficient before $t^0$ should be equal to $0$, so $c_1=0$. The parametrization rewrites as $$\begin{aligned} 3(b_1t&+b_2t^2+(a_2-b_1)t^3)(a_1+2a_2t+3(b_2-a_1)t^2)-3(a_1t+a_2t^2+(b_2-a_1)t^3)(b_1+2b_2t+3(a_2-b_1)t^2)\\ &+(1+d_1t+(b_2-1)t^2+(a_2-d_1)t^3)(2a_2t+3b_2t^2) -(a_2t^2+b_2t^3)(d_1+2(b_2-1)t+3(a_2-d_1)t^2) = 0.\end{aligned}$$ Coefficient before $t^1$ equals $2a_2$, so $a_2=0$. $$\begin{aligned} 3(b_1t+b_2t^2-&b_1t^3)(a_1+3(b_2-a_1)t^2)-3(a_1t+(b_2-a_1)t^3)(b_1+2b_2t-3b_1t^2)\\ &+(1+d_1t+(b_2-1)t^2-d_1t^3)(3b_2t^2) -(b_2t^3)(d_1+2(b_2-1)t-3d_1t^2) = \\ 3(b_1t+b_2t^2-&b_1t^3)(a_1+3(b_2-a_1)t^2)-3(a_1t+(b_2-a_1)t^3)(b_1+2b_2t-3b_1t^2)\\ &+b_2t^2(3+3d_1t+3(b_2-1)t^2-3d_1t^3 -d_1t-2(b_2-1)t^2+3d_1t^3)=\\ 3(b_1t-b_1t^3)(a_1+3&(b_2-a_1)t^2)-3(a_1t+(b_2-a_1)t^3)(b_1-3b_1t^2)\\ &+b_2t^2(3(b_2-a_1)t^2-3a_1+3+2d_1t+(b_2-1)t^2) =\\ b_1t^3(-3a_1-&9(b_2-a_1)t^2+9(b_2-a_1)+9a_1-3(b_2-a_1)+9(b_2-a_1)t^2)+b_1t(3a_1-3a_1)\\ &+b_2t^2(3(b_2-a_1)t^2-3a_1+3+2d_1t+(b_2-1)t^2)=\\ 6b_1b_2&t^3+b_2t^2(3(b_2-a_1)t^2-3a_1+3+2d_1t+(b_2-1)t^2)=\\ b_2t^2(6b_1t&+3(b_2-a_1)t^2-3a_1+3+2d_1t+(b_2-1)t^2) =\\ b_2t^2(t(6&b_1+2d_1)+t^2(4b_2-3a_1-1)-3a_1+3) =0\\\end{aligned}$$ Therefore, either $b_2=0$ or $a_1=1, b_2=1,d_1=-3b_1$. In the first case the curve is going to be as follows: $$(a_1t-a_1t^3,b_1t-b_1t^3,0,1+d_1t-t^2-d_1t^3) = (a_1t,b_1t,0,1+d_1t).$$ what is not really a cubic, but in the second case we have $$(t,b_1t+t^2-b_1t^3,t^3,1-3b_1t+3b_1t^3) = (t,t^2+\mu(t-t^3),t^3, 1-3\mu(t-t^3)).$$ As we predicted in the previous Section, we obtained a linear family of cubics. One can look at what happens in the limiting case $\mu=\infty$. The family of curves converges (if we look at the parametrizations) to a point $(0,-1/3,0)$. On the other hand their tangent vectors at $t=0,1,-1$ converge to $(0,1,0),(-3,-4,-3),(3,-4,3)$ respectively. Then, contact lines from $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$ with these tangent vectors all intersect in $(0,-1/3,0)$. So, the family $l(t,\mu)$ converges to these three lines as $\mu\to\infty$, these three lines with the embedded point $(0,-1/3,0)$ is a point on the boundary of the Hilbert scheme of rational cubics in ${{\mathbb C}}P^3$ (see [@cubicsComp] for more details about the compactification of the space of rational cubics). \[pr\_legendrian\] Is it true for higher degrees? The hypothesis is that there always exist at least $d$ legendrian rational curves of degree $d$ passing through $d$ generic points and a line. An heuristic argument is as follows. We take the one dimensional family (because $\mathrm{Stab^3_\mu}$ acts on these curves) of the degree $d$ legendrian curves through $d$ points which all belong to a plane $L$, and write the equation of the surface that they sweep. Then we intersect this surface with $L$. We obtain a collection of $d$ lines in the intersection, therefore the degree of the surface is at least $d$, therefore there is al least $d$ legendrian curves through $d$ generic points and one generic line. Also, this approach might work for any genus, as long as the set of the curves is not empty. Cubic surface containing the family of legendrian cubics -------------------------------------------------------- Here we find the equation of the surface containing all the cubics in the previous subsection. The equation of this surface is $F(x,y,z,w)=2x^3+21x^2z-27y^2z-54yzw-27zw^2+60xz^2+25z^3$. Such a surface intersects a generic line in three points, this gives another proof of Corollary \[cor\_answer\]. We found the family of legendrian cubics through $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$ and tangent to $\omega_1=3ydx-3xdy+wdz-zdw$. Therefore the family $(3t,t^2+\mu(t-t^3),t^3, 1-3\mu(t-t^3))$ is tangent to $\omega=ydx-xdy+wdz-zdw$ but passes through $(0,0,0,1),(3,1,1,1),(-3,1,-1,1)$. Therefore we apply the contactomorphism $ \psi= \begin{pmatrix} 1/2 & 0 & -1/2 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 1/2 & 0 & 1/2 \\ \end{pmatrix},$ which brings these three points to the standard ones. After the action of $\psi$, the parametrization of the family of legendrian rational cubics through the points $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$ is $$\label{eq_allcubics} (3t-t^3, 2t^2+2\mu(t-t^3), 2t^3,1+t^2-2\mu(t-t^3)).$$ In the following computation we choose the affine coordinates $(x/z,y/z,1,w/z)$. The following code produce the equation $F=0$ mentioned in the beginning of this section. clearAll path = append(path, "~/Dropbox/legendrian") loadPackage "Contact" R=QQ[x,y,w] T=QQ[t,m] cubic ={(3*t^2-1)/(2),(2*t^1+2*m*(t^2-1))/(2),(t^3+t^1-2*m*(t^2-1))/(2)} --sanity check isContact(-t^3+3*t,2*t^2+2*m*(t-t^3),2*t^3,t^2+1 - 2*m*(t-t^3),t) f=map(T,R,cubic) I=ker f l=mingens I g=l_0_0 -- 2x^3+21x^2-27y^2-54yw-27w^2+60x+25 R=QQ[x,y,z,w] g1=2*x^3+21*x^2*z-27*y^2*z-54*y*z*w-27*z*w^2+60*x*z^2+25*z^3 --g1 coincides with g at z=1 sub(g1,{x=>3*t-t^3,y=>2*t^2+2*m*(t-t^3),z=>2*t^3,w=>1+t^2-2*m*(t-t^3)}) sub(g1,{x=>0,y=>0,z=>0,w=>1})--=0 sub(g1,{x=>1,y=>1,z=>1,w=>1})--=0 sub(g1,{x=>-1,y=>1,z=>-1,w=>1})--=0 -- surface containing all legendrian cubics through (0,0,0,1),(1,1,1,1),(-1,1,-1,1) -- cubicSurface(x,y,z,w)-> == assigned in the main package. Tropicalization of legendrian curves {#sec_tropicallegendrian} ==================================== In what follows we suppose that small letters stand for complex numbers of complex-valued functions of $t$. By the same big letters we denote the corresponding limit of $\log_t$. For example, $\lim_{t\to\infty}\log_t|a(t)|=A$. We do not specify each time whether a small letter stands for a function or is a complex number, because it is clear from the context. \[def\_tropicallegendrian1\] A tropical curve $C\subset {{\mathbb T}}P^3$ is a tropical legendrian curve if there exists a family $C_t\subset {{\mathbb C}}P^3$ of complex legendrian curves, such that $\lim_{t\to\infty}\log_t(C_t)=C$ in the Hausdorff sense. Now we state the problem: given three points $P_1,P_2,P_3\in{{\mathbb T}}P^3$ in general position, we want to describe the family $S(P_1,P_2,P_3)$ of tropic legendrian cubics through them. The problem is correct in the sense that the set of the tropicalizations of the legendrian curves through generic liftings of $P_1,P_2,P_3$ to the families of points $p^t_1,p^t_2,p^t_3$ such that $\lim_{t\to\infty} \log_t(p^t_i)=P_i$ coincide. Indeed, a small change of points $p_1,p_2,p_3$ to $p'_1,p'_2,p'_3$ affects $S_t(p_1,p_2,p_3)$ only a few, because we can apply the contactomorphism which brings $p_1,p_2,p_3$ to $p'_1,p'_2,p'_3$, and since we found its form, we see that its matrix is close to the identity matrix. That finishes the proof. Similar statements are valid in general situation. If we consider some family of complex curves, then the dimension of the deformation space of a generic curve in this family coincides with the dimension of the deformation space of a generic tropical curve in the corresponding family of tropical curves. On the other hand, if a generic complex curve is superabundant, the same is true for the corresponding tropical curves. Another argument could be that the family of legendrian curves through three points sweeps a cubic surfaces, and instead of above perturbations we can just look at the tropicalization of this surface. Tropical legendrian lines ------------------------- Consider a tropical legendrian line $L$, i.e. $L$ is the limit of legendrian complex lines $l_t$. Equations in Proposition \[prop\_balance\] survive under tropical limit, and in the same notation but in the tropical situation we have $y(X)+x(Z) + x(Y)+y(W)=z(X)+w(Z)+z(W)+w(Y)$. But $(x(Y),y(X),z(X),w(Y))$ is the coordinates of the first vertex and $(x(Z),y(W),z(W),w(Z))$ is the coordinates of the second. So, sum of two vertices lies on the plane $X+Y=Z+W$, we call this condition [*tropical legendrian divisibility condition*]{} and it completely defines tropical legendrian lines. Consider a tropical curve $C\subset {{\mathbb T}}^3$, pick an edge $E$ of $C$ with two vertices $A,B$. \[def\_divisibility\] We say that a tropical curve [*looks like a legendrian line*]{} along the edge $E$ if - the edge $E$ is parallel to the vector $(1,1,0)$, - $A$ belongs to the part of ${{\mathbb T}}^3$ where $X+Y<Z$, valency of the vertex $A$ is three, and two other edges (others than $E$) from $A$ are parallel to $(1,0,0)$ and $(0,1,0)$, - $B$ belongs to the part of ${{\mathbb T}}^3$ where $X+Y>Z$, valency of the vertex $B$ is three, and two other edges (others than $E$) from $B$ are parallel to $(0,0,1)$ and $(1,1,1)$. We say that $C$ satisfies the tropical legendrian divisibility property (Figure \[fig\_line\]) along $E$ if $C$ looks like a legendrian line along $E$ and the middle point of $E$ belongs to the plane $X+Y=Z$. We say that $C$ satisfies the tropical legendrian divisibility property if for each edge $E$ of $C$, such that $C$ looks like a legendrian line along $E$, this edge $C$ satisfies the tropical legendrian divisibility property along $E$. \[y= [(0.9cm,0.4cm)]{}, z=[(0cm,0.6cm)]{}, x=[(0.5cm,-0.4cm)]{},scale=0.7\] (3,8,11)–(7,4,11)–(3,0,3)–(-1,4,3)–cycle; (1,-2,7)–(1,2,7)–(5,6,7)–(5,6,3); (-3,2,7)–(1,2,7); (5,6,7)–(7,8,9); (3,4,7) node [$\bullet$]{}; (3,4,7) node\[below\] [$R$]{}; (1,2,7) node\[above\] [$P$]{}; (5,6,7) node\[above\][$Q$]{}; (5,5,10) node [$X+Y=Z$]{}; \[prop\_lines\] All the tropical lines satisfying the tropical legendrian divisibility property can be obtained as tropical limits of complex legendrian curves. If a tropical line has vertices of the type $(A,B,C),(A',B',C')$ with $A+A'+B+B'=C+C'$ then these two points are of the following types. - $(A,B,A+B+X),(A+X,B+X,A+B+X)$ and the corresponding family of legendrian curves are $(t^A+t^As,2t^B+t^Bs,t^{A+B+X}+t^{A+B}s)$, - $(A,B,A+B),(A+X,B,A+B+X)$ and $(t^A+t^As,t^B+t^{B-X}s,t^{A+B}-(t^{A+B}-t^{A+B-X})s)$, - $(A,B,A+B),(A,B+X,A+B+X)$, similar to the previous case. Tropical legendrian rational cubics via Macaulay2 ------------------------------------------------- We found all the legendrian cubics through $(0,0,0,1),(1,1,1,1),(-1,1,-1,1)$, but if we tropicalize these points, they go to $(-\infty,-\infty,-\infty),(0,0,0),(0,0,0)$ in the affine coordinates, which seems to be not very interesting. We want to draw all tropical tropical legendrian curves through three points $(X_1,Y_1,Z_1),(X_2,Y_2,Z_2),(X_3,Y_3,Z_3)\in{{\mathbb T}}^3$ with different $Z_1,Z_2,Z_3$ and $X_i+Y_i\leq Z_i$. We start with an example. \[ex\_tropicalcubic\] Let $(X_1,Y_1,Z_1)=(9,2,14), (X_2,Y_2,Z_2)=(-16,-23,-2)$, and $(X_3,Y_3,Z_3)=(-31,-22,-12)$. Then, by direct calculation we can find the tropical cubic sweeping all tropical legendrian cubics through these three points, it is given by $\max(3x + 321, 2x + y + 328, 2x + z + 327, 2x + 325, x + y + z + 334, x + y + 332, x + z + 347, x + 2y + 335, x + 2z + 349, x + 337, 3y + 342, 2y + z + 340, 2y + 328, y + 2z + 356, y + 342, y + z + 354, 3z + 347, 2z + 361, z + 359, 347)$ Then we can directly find one of the legendrian curves, it is drawn in the surface. We give the following code in Mathematica [@draw]. fs = {3*x + 321, 2*x + y + 328, 2*x + z + 327, 2*x + 325, x + y + z + 334, x + y + 332, x + z + 347, x + 2*y + 335, x + 2*z + 349, x + 337, 3*y + 342, 2*y + z + 340, 2*y + 328, y + 2*z + 356, y + 342, y + z + 354, 3*z + 347, 2*z + 361, z + 359, 347}; conds = Table[{Equal @@ fs1, And @@ Table[First@fs1 >= f, {f, Complement[fs, fs1]}]}, {fs1, Subsets[fs, {2}]}]; aa = ContourPlot3D[ x - y == 15, {x, -50, 50}, {y, -50, 50}, {z, -50, 50}, ContourStyle -> {Yellow, Opacity[0.5]}, AxesLabel -> {"x", "y", "z"}]; cc = Graphics3D[{PointSize[Large], Green, Point[{9, 2, 14}], Green, Point[{-16, -23, -2}], Green, Point[{-31, -22, -12}]}]; bb = Table[ ContourPlot3D[Evaluate@First@c, {x, -32, 50}, {y, -32, 50}, {z, -32, 50}, RegionFunction -> Function[{x, y, z} , Last@c], AxesLabel -> {x, y, z}, ContourStyle -> Directive[Orange, Opacity[0.5]]], {c, conds}]; s1 = Graphics3D[{AbsoluteThickness[10], Blue, Line[{{26/3, 26/3 - 15, -50}, {26/3, 26/3 - 15, -12}}]}]; s2 = Graphics3D[{AbsoluteThickness[10], Blue, Line[{{26/3, 26/3 - 15, -12}, {12, -3, 2}}]}]; ... (*s1,s2,... stands for parts of the blue tropical curve *) Show[bb, cc, s1, s2, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14, s15, PlotRange -> {{-20, 20}, {-20, 20}, {-20, 20}}] ![The tropical cubic containing all the legendrian cubics through $(9,2,14), (-16,-23,-2),(-31,-22,-12)$. []{data-label="fig_tropicalcubic1"}](cubicSurface1 "fig:"){width="7cm"} ![The tropical cubic containing all the legendrian cubics through $(9,2,14), (-16,-23,-2),(-31,-22,-12)$. []{data-label="fig_tropicalcubic1"}](cubicSurface2 "fig:"){width="7cm"} On Figure \[fig\_tropicalcubic1\] we see that even though the three chosen points are in the half-space $X+Y<Z$, we cannot guarantee that the part of the curve in the half-space $X+Y>Z$ will be in one plane of the type $X-Y=c$. ![The tropical cubic containing all the legendrian cubics through $(20,2,17), (20,18,25),(12,20,5)$. []{data-label="fig_tropicalcubic2"}](cubicSurface3 "fig:"){width="8cm"} ![The tropical cubic containing all the legendrian cubics through $(20,2,17), (20,18,25),(12,20,5)$. []{data-label="fig_tropicalcubic2"}](cubicSurface4 "fig:"){width="8cm"} If we choose our three fixed points not in the part $X+Y<Z$, then the picture becomes even more complicated (see Figure \[fig\_tropicalcubic2\]) and lack any visible distinction from a random tropical cubic surface. Code for producing tropical legendrian cubics and their spanning surface {#sec_codetropical} ------------------------------------------------------------------------ We want to draw the tropical cubic surface which contains the tropical legendrian cubics through three fixed points. So, we choose three points, which have desired tropical limits as $t$ tends to infinity. Then, using code in Section \[sec\_contactomorphisms\] we find the contactomorphic matrix $M$ which brings our points to the standard ones, and then we know the explicit Eq.  for the cubic surface, containing all the legendrian rational cubics through these three standard points. Using this matrix $M$, we obtain the equation of such a surface for our points, then we take the valuation, and this gives the tropical polynomial defining the tropical cubic surface. clearAll path = append(path, "~/Dropbox/legendrian/") loadPackage "Contact" S=QQ[p] a1=2*p^(13); b1=2*p^(20); c1=p^(33); a2=2*p^(11); b2=2*p^(5); c2=p^(31); a3=2*p^(4); b3=2*p^(13); c3=2*p^(27); M1=transformation(a1,b1,c1,a2,b2,c2,a3,b3,c3,S);--see Section 1.2 R=S[symbol x, symbol y,symbol z,symbol w,s,m, Degrees=> {{1,0,0},{0,1,0},{0,0,1},{0,0,0},{0,0,0},{0,0,0}}] M3=matrix{{x},{y},{z},{w}} mma=(inverseMatrix(M1,S)*M3)_0 (x1,y1,z1,w1)=(mma_0,mma_1,mma_2,mma_3) g=cubicSurface(x1,y1,z1,w1) leadterms2:=(ww)->( (M,C):=coefficients(ww, Variables=>{x,y,z,w}); C1:=matrix(for i in entries(C_0) list (degree lift(leadTerm(i),S))); (M*C1)_(0,0)) gg=leadterms2(g) mon = {x^3,x^2*y,x^2*z,x^2*w,x*y*z,x*y*w,x*z*w,x*y^2, x*z^2,x*w^2,y^3,y^2*z,y^2*w,y*z^2,y*w^2,y*z*w,z^3,z^2*w,z*w^2,w^3} ggg=for i in mon list ((matrix{degree(i)}* matrix{{x},{y},{z},{0}})+coefficient(i,gg)) ans= html(for i in ggg list i_0_0) The variable [*ans*]{} gives us {3*x+228, 2*x+y+217, 2*x+z+208, 2*x+236, x+y+z+203, x+y+230, x+z+223, x+2*y+209, x+2*z+192, x+250, 3*y+200, 2*y+z+196, 2*y+223, y+2*z+185, y+243, y+z+216, 3*z+172,2*z+205, z+236, 263} which we put into the code in Example \[ex\_tropicalcubic\] as [*fs*]{}. This gives the following pictures, see Figure \[fig\_tropicalcubics3\]. ![The tropical cubic containing all the legendrian cubics through $(13,20,33), (11,5,31),(4,13,27)$. []{data-label="fig_tropicalcubics3"}](cubicSurface5 "fig:"){width="8cm"} ![The tropical cubic containing all the legendrian cubics through $(13,20,33), (11,5,31),(4,13,27)$. []{data-label="fig_tropicalcubics3"}](cubicSurface6 "fig:"){width="8cm"} \[ex\_tropicalcubic4\] Now we are ready to find some concrete examples of tropical legendrian rational cubic curves. A priori, it is not simple, because all what we have now is only the parametrization of a curve. If we try to directly take the valuation of the parametrization, we obtain only a part of our curve. So, we will use another approach. A rational curve can by determined by the roots of its coordinates. So, if a curve has a parametrization of the type $(x(s),y(s),z(s),w(s))$, we find $x_0,x_1,x_2$ which approximate the roots of $x(s)$, i.e. ${\mathrm{val}}(x(x_i))<<0$, then we do the same for other coordinate functions. Then we can calculate $({\mathrm{val}}(x(x_0)),{\mathrm{val}}(y(x_0)),{\mathrm{val}}(z(x_0)),{\mathrm{val}}(w(x_0)))$, and then the same for $x_i,y_i,z_i,w_i, i=1,2,3$. While trying to find $x_0$ we sometimes experience the problem that we need to take square roots or cubic roots. The way to avoid such problems is to choose an appropriate prime number such that the Newton method for finding roots of polynomials works without passing to the extensions of the base field. clearAll; path = append(path, "~/Dropbox/legendrian/"); loadPackage "Contact" S=QQ[p]; N=2897; S=ZZ/N[p]; a1=1/p^4; b1=1/p^4; c1=p^4; a2=3*p^12; b2=4/p^8; c2=5*p^(12); a3=6*p^16; b3=7*p^(8); c3=2*p^(32); R=QQ[s,m,t,Inverses => true,MonomialOrder => Lex] if char(S)>0 then R=ZZ/N[s,m,t,Inverses => true,MonomialOrder => Lex] var={s,m}; mapR=map(R,S,{t}); M1=transformation(a1,b1,c1,a2,b2,c2,a3,b3,c3,S); T1=mapR(M1); leadterms:=(ww)->( (M,C):=coefficients(ww, Variables=>var); C1:=transpose matrix{for i in entries(C_0) list (degree leadTerm(i))_0}; (M*C1)_(0,0)) Tcoeff:=(w)->(coefficients(w,Variables=>{s})) -- we are going to sample points on the curve setpoint:=(s1,m1,curve)->( point=sub(curve,{s=>s1,m=>m1}); ll=degree(point_0_3); (degree(point_0_0) -ll,degree(point_0_1) -ll,degree(point_0_2) -ll) ) --newton method to find roots newton:=(f,s0,k)->( s1:=s0; f1:=sub(f,{s=>s+s1}); (a,b):=coefficients(f1,Variables=>{s}); d1:=degree leadTerm(b_0_3); d2:=degree leadTerm(b_0_2); d:=d1-d2; den:=0; cc:=0; for i from 1 to k do( f1=sub(f,{s=>s+s1}); (a,b)=coefficients(f1,Variables=>{s}); d1=degree leadTerm(b_0_3); d2=degree leadTerm(b_0_2); d=d1-d2; cc=(coefficient(t^(d2_0),leadTerm(b_0_2))); if char(R)==0 then den=1/cc else ( for i from 1 to N do (if mod(i*cc-1,N)==0 then den=i); ); s1=s1-t^(d_0)*(coefficient(t^(d1_0),leadTerm(b_0_3)))*den; ); s1 ) --all contact cubic curves curve=matrix{{-s^3+3*s},{2*s^2+m*(s-s^3)},{2*s^3},{s^2+1 - m*(s-s^3)}} result1=T1*curve; result=result1; --we have result1, which is the curve over S, need to draw, --but for that we need to sample points on the curve x:=result_0_0; y:=result_0_1; z:=result_0_2; w:=result_0_3; for i from 0 to 3 do (print leadterms(result_0_i)) m0=t^2 for i from -30 to 5 do print setpoint(t^(i),m0,result) f=sub(z,{m=>m0,s=>s}); leadterms(x) f=sub(x,{m=>m0}); use R; rx1=newton(f,t^(-18),10) rx2=newton(f,1,10) rx3=newton(f,-1,10) setpoint(rx3,m0,result) leadterms(y) f=sub(y,{m=>m0}) ry1=newton(f,t^(-4),20) ry2=newton(f,1,10) ry3=newton(f,-1,10) setpoint(ry2,m0,result) leadterms(z) f=sub(z,{m=>m0}); rz1=newton(f,t^(-10),10) rz2=newton(f,94*t^(-1),10) rz3=newton(f,-94*t^(-1),10) setpoint(rz3,m0,result) leadterms(w) f=sub(w,{m=>m0}); rw1=newton(f,-1+t^(-10),10) rw2=newton(f,-1+132*t^(-5),10) rw3=newton(f,-1-132*t^(-5),10) setpoint(rw1,m0,result) As always, the first few terms of the roots we need to find by hands, see the part with $w$-coordinate in the above example. The numbers $132$ and $2897$ are chosen in such a way that the series in $t$ for roots of $w$ may be written using integer (finally, sure, from a finite field) coefficients. I do not know a general algorithm which works in any case and does not require a personal tuning. Calculating [*setpoint(..., m0,results)*]{} for [*...=rx1,rx2,...rw3*]{} we obtain the following points on our curve: $$\begin{aligned} (-4,8,32),(22,-12,32),&(-8,-2,12),(12,-28,12),(-24,-4,4),(10,-44,6),(13,-1,-1),\\ &(13,-1,-3),(8,-4,-16),(44,30,52),(37,23,40),(37,23,40).\end{aligned}$$ So, we find the vertices of our curve: $(22,8,32),(12,-2,12)$ — these are the vertices lying in the planes $Z=32$ and $Z=12$. As it should be, $X-Y=14$ for both vertices as the rest of the curve in the part $X+Y>Z$ should be in a plane $X-Y=c$. But at the bottom we see a different picture. The leg going to $-\infty$ by $X$ coordinate comes to the vertex $(8,-4,4)$ on $X+Y=Z$, where the leg going to $-\infty$ by $Z$-coordinate branches. Then the curve goes to $(10,-4,6)$, where the leg, going to $-\infty$ by $Y$-coordinate, branches. Also, the curve came to the plane $X-Y=14$ in order to unify with the other parts of the curve. The rest of the curve is also uniquely determined, see Figure \[fig\_tropicalcubics4\]. ![On the left, the tropical cubic surface and the plane $X-Y=14$, on the right we see the part of our curve, in $X-Y=14$ with $X+Y\geq Z$. []{data-label="fig_tropicalcubics4"}](cubicSurface7 "fig:"){width="8cm"} ![On the left, the tropical cubic surface and the plane $X-Y=14$, on the right we see the part of our curve, in $X-Y=14$ with $X+Y\geq Z$. []{data-label="fig_tropicalcubics4"}](cubicSurface8 "fig:"){width="8cm"} As we see in this example, we have a part of tropical cubic curve which looks like a tropical line. Also, it has vertices $(22,8,32)$ and $(24,10,32)$. As we see, $(22+8)+(24+10)=32+32$, i.e. the middle point of the interval between the vertices is on the plane $X+Y=Z$. We list here the values of the coordinates at the roots of the other coordinates. $\begin{pmatrix} -\infty & 8 & 32 & 0 \\ -\infty & -2 & 12 & 0 \\ -\infty & -4 & 4 & 0 \\ \end{pmatrix}$ $\begin{pmatrix} 10 & -\infty & 6 & 0 \\ 12 & -\infty & 12 & 0 \\ 22 & -\infty & 32 & 0 \\ \end{pmatrix}$ $\begin{pmatrix} 13 & -1 & -\infty & 0 \\ 13 & -1 & -\infty & 0 \\ 8 & -4 & -\infty & 0 \\ \end{pmatrix}$ $\begin{pmatrix} 14 & 0 & 22 & -\infty \\ 14 & 0 & 17 & -\infty \\ 14 & 0 & 17 & -\infty \\ \end{pmatrix}$ This example is related to Remark \[pr\_weil\]. It is not clear what and how should something be added in order to obtain a similar identity. Direct attempt would be something like the following. Denote by $Y(X)$ the sum of $Y$-coordinate at the roots of the $X$-coordinate, etc. If we try to verify the identity $Y(X)-Z(X)=W(Z)-X(Z)$, we see that it does not hold. Tropical differential forms =========================== Infinitesimal considerations via the logarithmic Gauss derivative ----------------------------------------------------------------- We define logarithmic Gauss derivative, cf. [@MR3119697; @martin2013discriminant; @MR2931884]. For $s\in{{\mathbb K}}$ and $f:{{\mathbb K}}\to{{\mathbb K}}$ we define the [*logarithmic Gauss derivative*]{} $f_\gamma$ at $s$ by $(f_\gamma) (s)=s\frac{f'(s)}{f(s)}$. So, $f_\gamma$ is a function ${{\mathbb K}}\to{{\mathbb K}}$. We are going to study the valuation of logarithmic Gauss derivates of tropical polynomials. We start with a polynomial $f:{{\mathbb K}}\to{{\mathbb K}}$ where ${{\mathbb K}}$ is a valuation field, $f(s)=\sum_{i=0}^n a_is^i.$ The tropicalization of $f$ is ${\mathrm{Trop}}(f)(S)=\max_{i=0}^n({\mathrm{val}}(a_i)+iS)$, where $S={\mathrm{val}}(s)$ is the tropical parameter. \[lemma\_derivative\] Consider $s\in{{\mathbb K}}, S={\mathrm{val}}(s)$ and $k$ such that ${\mathrm{val}}(a_k)+kS>{\mathrm{val}}(a_i)+iS$ for all $i\in 0,\dots, n, i\ne k$. In this case, - if $k\ne 0$, then ${\mathrm{val}}(f'(s))={\mathrm{val}}(f(s))-S$. Also, $f_\gamma(s)=k+$ higher order terms, ${\mathrm{val}}(f_\gamma(s))=0$; - if $k=0$, then ${\mathrm{val}}(f'(s))<{\mathrm{val}}(f(s))-{\mathrm{val}}(s)$, and ${\mathrm{val}}(f_\gamma(s))<0$. We calculate $f'(s)=\sum_{i=1}^n ia_is^{i-1}$. Note that ${\mathrm{val}}(a_is^i)={\mathrm{val}}(ia_is^{i-1})+S$, except the case $i=0$. Therefore $\max_{i=0}^n({\mathrm{val}}(a_i)+iS)=\max_{i=1}^n({\mathrm{val}}(ia_i)+(i-1)S)+S$ everywhere except the place where the maximum is attained for $i=0$. To resolve the case $i=0$ we consider $g=\sum_{i=1}^n a_is^i$. Clearly, $f'=g'$, ${\mathrm{val}}(g(s))<{\mathrm{val}}(f(s))$ for $s$ such that the monomial $a_0$ strictly dominates all the $a_is^i$ in $f$. Now, for the function $g$ we have the equality ${\mathrm{val}}(g'(s))={\mathrm{val}}(g(s))-S$, which concludes the proof. Therefore the logarithmic Gauss derivative (if it has non-negative valuation) tells us the slope of a tropical edge. Tropical legendrian tangency property ------------------------------------- \[lem\_locally\] If $C\subset {{\mathbb K}}^3$ is legendrian curve, a point $p=(x, y, z)\in C$ and ${\mathrm{val}}(xy)>{\mathrm{val}}(z)$, then the difference between the $X$-coordinate and $Y$ coordinate of ${\mathrm{Trop}}(C)$ near ${\mathrm{Val}}(p)$ is locally constant. We substitute local parametrization into the form $\omega=ydx+xdy+wdz+zdw$. Let us choose local parametrization such that $w=1$. The condition that our curve is legendrian reads as $$\label{eq_toriccontact} x(s)y(s)\left(\frac{x'(s)}{x(s)}-\frac{y'(s)}{y(s)}\right)+z(s)\left(\frac{z'(s)}{z(s)}\right)=0.$$ Locally, ${\mathrm{val}}(x(s)),{\mathrm{val}}(y(s)),{\mathrm{val}}(z(s))$ are tropical polynomials, so we can apply the reasoning form Lemma \[lemma\_derivative\]. Since $z_\gamma(s)={\mathrm{val}}\left(\frac{z'}{z}(s)\right)+{\mathrm{val}}(s)$ is at most zero, then in order to have zero in we need that ${\mathrm{val}}(s)\left(\frac{x'(s)}{x(s)}-\frac{y'(s)}{y(s)}\right)<0$ by Lemma \[lemma\_derivative\], i.e. $x_\gamma(s)=y_\gamma(s)$ locally. Therefore either the leading term is constant in both $x(s),y(s)$ (and so changing $s$ does not move the considered point), or $x_\gamma(s)=k+$ higher order terms, and $y_\gamma(s)=l+$ higher order terms, and so we have to have $k=l$. \[prop\_locally2\] Analogously, we can prove that in the part where ${\mathrm{val}}(xy)<{\mathrm{val}}(z)$ a tropical legendrian curve has locally constant $Z$-coordinate. \[def\_distribution\] A tropical curve $C\subset {{\mathbb T}}^3=\{(X,Y,Z)|X,Y,Z\in{{\mathbb T}}\}$ satisfies the [*tropical legendrian tangency property*]{}, if in the half-space $X+Y>Z$ this curve $C$ is tangent to the distribution $dX-dY$, and in the half-space $X+Y<Z$ this curve $C$ is tangent to the distribution $dZ$. In other words, if $Z>X+Y$ (resp. $Z<X+Y$), then the curve locally lies in a plane of the type $Z=\mathrm{const}$ (resp. $X-Y=\mathrm{const}$). \[prop\_tandency\] A tropical legendrian curve $C\subset {{\mathbb T}}P^3$ (Definition \[def\_tropicallegendrian1\]) satisfies the tropical legendrian tangency property. In particular, all the edges $E$ of $C$ with endpoints on different sides of the plane $X+Y=Z$ are parallel to the vector $(1,1,0)$. Indeed, this proposition follows from Lemma \[lem\_locally\] and Proposition \[prop\_locally2\]. A tropical line $L\subset{{\mathbb T}}^3$ is a tropical legendrian line (Definition \[def\_tropicallegendrian1\]) if and only if $L$ satisfies the tropical legendrian divisibility property (Definition \[def\_divisibility\]) and the tropical legendrian tangency property (Definition \[def\_distribution\]). We just verify that all the tropical legendrian lines (we have the full description in Proposition \[prop\_lines\]) satisfy the hypothesis of our proposition. \[y= [(0.9cm,0.4cm)]{}, z=[(0cm,0.6cm)]{}, x=[(0.5cm,-0.2cm)]{},scale=0.7\] (7,5,6)–(4,2,6)–(8,6,14)–(11,9,14)–cycle; (5,2,7)–(4,3,7)–(2,1,7)–(3,0,7)–cycle; (6,3,9)–(5,4,9)–(3,2,9)–(4,1,9)–cycle; (8,5,13)–(7,6,13)–(5,4,13)–(6,3,13)–cycle; (1,2,7) node [$\bullet$]{}; (2,3,9) node [$\bullet$]{}; (5,1,13) node [$\bullet$]{}; (-1,2,7)–(4,2,7)–(5.5,3.5,7); (4,-2,7)–(4,2,7); (0,3,9)–(5,3,9)–(6.5,4.5,9); (5,1,9)–(5,3,9); (5,0.5,13)–(5,3,13)–(8.5,6.5,13); (2,3,13)–(5,3,13); (9,5,11) node [$X-Y=14$]{}; (3,1,7.3) node [$Z=7$]{}; (4,2,9.3) node [$Z=9$]{}; (5,5,13.3) node [$Z=13$]{}; \[ex\_tangency\] Let us draw a tropical legendrian curve through points $(1,2,7), (2,3,9), (5,1,13)$. Suppose also that it has a point in the half-space $X+Y>Z$ with coordinates satisfying $X-Y=2$. Thanks to tropical legendrian tangency condition, the picture should be like in Figure \[fig\_tangency\]. We do not know how this curve looks like in the half-space $X+Y>Z$, except that it contains the points $(4.5,2.5,7),(5.5,3.5,9),(7.5,5.5,13)$. Tropical divisibility property for curves ----------------------------------------- Here we prove a version of the divisibility property (Definition \[def\_divisibility\]) for the rational curves which have only one intersection with the plane $w=0$, such a curve always has a parametrization of the type $$\label{eq_div1} (a_0+a_1s+a_2s^2+\ldots: b_0+b_1s+b_2s^2+\ldots: c_0+c_1s+c_2s^2+\ldots:1) \subset {{\mathbb K}}P^3.$$ Suppose that somewhere our parametrization looks like $$\label{eq_div2} (x(s),y(s),z(s))= (\ldots+ a_ls^l+\ldots+a_ks^k+\ldots,\ldots+ b_ks^k+\ldots, c_0+\ldots + c_ps^p+\ldots)$$ in the affine coordinates. This notation means that on the interval for $s$ that we consider, at the beginning, the monomials $a_ls^l,b_ks^k,c_0$ are the dominating monomials in the $x,y,z$ coordinate respectively. Then, at some moment, $s=s_1$ in the $x$-coordinate the monomial $a_ks^k$ becomes dominant. And then, at some moment $s=s_2$ the monomial $c_ps^p$ becomes dominant in the $z$-coordinate. We suppose that nothing more happens during this period of time. That means that the tropicalization $C$ of our curve has the following behavior in between on ${\mathrm{val}}(s_1)\leq S\leq {\mathrm{val}}(s_2)$. The curve starts in the part $Z\geq X+Y$ of the space and has a vertex $P$ at $S={\mathrm{val}}(s_1)$, because ${\mathrm{val}}(a_ls_1^l)={\mathrm{val}}(a_ks_1^k)$ and $C$ changes its direction. Then $C$ goes to the part $Z<X+Y$ and has a vertex $Q$ there, when ${\mathrm{val}}(c_0)={\mathrm{val}}(c_ps_2^p)$. Also, these conditions mean that the leg of $C$ which goes from $Q$ to $-\infty$ by $Z$-coordinate does not branch. We can reformulate the conditions Eq. , as follows. We consider a part of a tropical curve $C$, namely, an edge $PQ$ of $C$. The direction of the edge $PQ$ is $(k,k,0)$, the direction of edges from $P$ are $(l,k,0)$ and $(k-l,0,0)$, these edges are the infinite legs of $C$. Also $Q$ has a vertical infinite leg in the direction $(0,0,p)$. \[th\_affinedivisibility\] In the above hypothesis, $(k-l)P+pQ$ belongs to the plane $X+Y=Z$. Let us find the valuations of the roots $s_1,s_2$ of the first and the third coordinates, respectively. We have ${\mathrm{val}}(s_1)= \frac{{\mathrm{val}}(a_l)-{\mathrm{val}}(a_k)}{k-l},{\mathrm{val}}(s_2)=\frac{{\mathrm{val}}(c_0)-{\mathrm{val}}(c_p)}{p}$. We denote $A_k={\mathrm{val}}(a_k),B_k={\mathrm{val}}(b_k)$, etc. Therefore the coordinates of these two vertices $P,Q$ of $C$ are given by $$\begin{split} P=(A_k+k\frac{A_l-A_k}{k-l}, B_k+k\frac{A_l-A_k}{k-l},C_0), \\ Q= (A_k+k\frac{C_0-C_p}{p},B_k+k\frac{C_0-C_p}{p},C_0). \end{split}$$ The condition that $(k-l)P+pQ$ belongs to the plane $X+Y=Z$ is the following: $$\begin{split} (k-l)A_k+k(A_l-A_k)+(k-l)B_k +k(A_l-A_k)+pA_k+ \\ k(C_0-C_p)+pB_k+k(C_0-C_p)=(k-l+p)C_0. \end{split}$$ So we need to verify that $$\label{eq_needprove} A_k(k-l-k-k+p)+A_l(k+k)+B_k(k-l+p)=2kC_p + C_0(k-l+p-2k).$$ Since $ydx-xdy+dz=0$, we have $$(b_ks^k)(la_ls^{l-1}+ka_ks^{k-1})-(a_ls^l+a_ks^k)kb_ks^{k-1} = pc_ps^{p-1}.$$ This implies $p-1=k+l-1$ and $(k-l)a_lb_k=pc_p$ i.e. $A_l+B_k=C_p$. Finally, it is easy to verify that Eq.  follows from the equalities $p-k-l=0$ and $A_l+B_k=C_p$. Divisibility conditions for “line”-similar parts of the tropical rational legendrian cubics ------------------------------------------------------------------------------------------- We will draw a tropical legendrian cubic $C$ through three generic points $P_1,P_2,P_3$. We will choose these three points $P_1,P_2,P_3$ in the half-space $X+Y<Z$ with different $Z$-coordinates $Z_1<Z_2<Z_3$ respectively. By Proposition \[prop\_tandency\] any such curve will satisfy tropical legendrian tangency property (Definition \[def\_distribution\]). Therefore, there exists a plane with equation $X-Y=c$ such that $C\cap \{(X,Y,Z)|Z<X+Y\}$ belongs to this plane (see Example \[ex\_tangency\]). Suppose also that $C$ has no edges inside the plane $X+Y=Z$ (cf. with Figure \[fig\_tropicalcubic1\] where we have such an edge). \[th\_cubicdivisibility\] Suppose that a tropical rational legendrian cubic as above looks like a tropical legendrian line (Definition \[def\_divisibility\], Figure \[fig\_line\]) along an edge $PQ$. Then the middle point $(P+Q)/2$ of this edge $PQ$ belongs to the plane $X+Y=Z$. For such a curve, we always can choose a local parameter $s$ such that a neighborhood of $PQ$ is given by $$(a_0+a_1s+\dots, b_0+b_1s+\dots, c_0+c_1s+\dots, d_0+d_1s+\dots),$$ if $PQ$ is on the level $Z_1$ or $Z_3$. If $PQ$ is in the middle, i.e. on the level of $Z_2$, then we can choose the following parametrization which also looks like a line: $$(a_1s+a_2s^2+\dots, b_0+b_1s+b_2s^2+\dots, c_0+c_1s+c_2s^2+\dots, d_0+d_1s+\dots),$$ and for some parameter $s_0$ we have $${\mathrm{val}}(a_1s_0)={\mathrm{val}}(a_2s_0^2), {\mathrm{val}}(b_1s_0)={\mathrm{val}}(b_2s_0^2), {\mathrm{val}}(c_1s_0)={\mathrm{val}}(c_2s_0^2),{\mathrm{val}}(d_0)={\mathrm{val}}(d_1s_0).$$ We use the same notation as in the previous section. In the first case the proof is the same as for a line, because we do not see the higher degree terms. For the second case, it is a computation. We evaluate the contact form on the curve and obtain $b_0a_1+d_0c_1-c_0d_1+s(2a_2b_0+2d_0c_2)+s^2(\dots)$. This means that $A_2+B_0=D_0+C_2$ and $B_0+A_1=C_0+D_1$ because $C_1-C_0<D_1-D_0$. Now we compute the coordinates of the vertices $P,Q$, they happen to be $$(A_1+B_0-B_1,B_0,C_0,D_0) \text{ and\ } (A_2+2(C_1-C_2),B_2+2(C_1-C_2),2C_1-C_2,D_0),$$ we need to verify that $$A_1+2B_0-B_1+A_2+B_2=C_0+D_0+2C_2+D_1-C_1.$$ This follows from $B_1-B_2=C_1-C_0=A_1-A_2=D_0-D_1, B_0+A_2=D_0+C_2$. Note that we heavily use a parametrization of a curve, including information about how many roots of each coordinate we have near an edge. So we use information, which can not be deduced from the local picture, unless we already know very well how the curve looks like combinatorially. Integration on a tropical line {#sec_integration} ------------------------------ We start with an example; then we briefly explain how this example proves the tropical legendrian divisibility property. A formal proof with complete definitions will appear in [@tropforms]. Consider the line $L=\{(x,y)\in({{\mathbb C}}^*)^2|x+y-1=0\}$ and the form $\Omega=\frac{dy}{y}$ on it. When $t\to\infty$, the image of $L$ under the map ${{\text{Log}}}_t$ tends to the locus ${\mathrm{Trop}}(L)$ where the function $\max(X,Y,0)$ is not smooth. Consider the interval $AB\subset {\mathrm{Trop}}(L)$ where $A=(-a,0), B=(-b,0), b>a>0$. It is natural to consider points $A_t=(t^{-a},1-t^{-a}), B_t=(t^{-b}, 1-t^{-b})\in L$ as “preimages” of $A,B$ with respect to the map ${{\text{Log}}}_t$. Let us compute the integral $\int_{A_t}^{B_t}\frac{dy}{y} = \ln (1-t^{-a})-\ln(1-t^{-b})\sim t^{-a}$. Therefore it is natural to say that “the asymptotic” of $\Omega$ near the point $(0,a)$ is $t^{-a}$. After the toric change $x\to x, y\to x/y$ of coordinates, the form $\Omega$ becomes $\frac{dx}{x}-\frac{dy}{y}$. Consequently, it is natural to say that “asymptotic” of $\frac{dx}{x}-\frac{dy}{y}$ near the point $(a,a)$ is $t^{-a}$. Application for tropical legendrian curves {#sec_applicationrefined} ------------------------------------------ Suppose that we have a tropical legendrian curve $C$ which looks like a tropical legendrian line (Definition \[def\_divisibility\]) along an edge $e\subset C$. Then we consider the form $(\frac{dx}{x}-\frac{dy}{y})$ on the left part of $e$. Take the point $P$ at the intersection of $C$ with the plane $X+Y=Z+W$. As we showed in the previous section, “the asymptotic” of $(\frac{dx}{x}-\frac{dy}{y})$ at $P$ is $t^{-l}$ where $l$ is the distance from $P$ to the left vertex of $e$. By the same reasoning, for the right hand side of $e$ we get that “the asymptotic” of $(\frac{dz}{z}-\frac{dw}{w})$ at $P$ is $t^{-l'}$ where $l'$ is the distance from $P$ to the right vertex of $e$. Since $xy/zw$ is asymptotically a constant near $P$ and $(\frac{dx}{x}-\frac{dy}{y}) = \frac{zw}{xy}(\frac{dz}{z}-\frac{dw}{w})$ we conclude that $l=l'$ which suggests that the tropical legendrian divisibility property holds in the full generality. References {#references .unnumbered} ========== [^1]: We have 6 conditions on the coefficients of a matrix $A\in GL(4,{{\mathbb C}})$, since $A$ preserves $\omega$, and the condition $\det A\ne 0$, but one can check (by Macaulay2 for example), that the set of such $A\subset {{\mathbb C}}^{16}$ is a quasiprojective variety of dimension 10.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Structural properties of a finite number ($N = 2 - 20$) of point charges (classical electrons) confined laterally in a two-dimensional two-minima potential are calculated as a function of the distance ($d$) between the minima. The particles are confined by identical parabolic potentials and repel each other through a Coulomb potential. Both ground state and metastable electron configurations are discussed. At zero distance previous results of other calculations and experiments are reproduced. Discontinuous transitions from one configuration to another as a function of $d$ are observed for $N = 6, 8, 11, 16, 17, 18, 19$.' author: - 'M. Marlo' - 'M. Alatalo' - 'A. Harju' - 'R. M. Nieminen' date: - - title: Classical electrons in laterally coupled diatomic 2D artificial molecule --- INTRODUCTION {#sec:intro} ============ Quantum dots (sometimes called artificial atoms) are nanoscale semiconductor structures where a small number of electrons are confined into a small spatial region.[@Ashoori; @Kastner] The electron motion is usually further restricted to two dimensions. There is strong theoretical evidence for the existence of a limit where the electron system crystallises to Wigner molecules, which is seen as the localisation of the electron density around positions that minimise the Coulomb repulsion. [@Egger; @Reimann; @Creffield; @Reusch; @Arin_paperi; @Koskinen; @Maksym_molaspects] In the limit of weak confinement (low density) or a very strong magnetic field the quantum effects are quenched or obscured and the classical electron correlations start to dominate the properties of the system. The ultimate limit is a purely classical system where only the Coulomb repulsion between the electrons defines the ground state. The problem reduces to finding the classical positions of electrons (which depend on the forms of the confining and the interaction potentials) that minimise the total energy of the system. There is growing interest in calculating[@Imamura; @Nagaraja; @Partoens; @Veikon_paperi; @Wensauer; @Kolehmainen; @Jannouleas] and measuring[@Livermore; @Oosterkamp; @Brodsky] the properties of coupled quantum dots. Due to the 2D nature of quantum dots the two-atom system is different whether the quantum dots are coupled in the plane in which the electrons are confined ([*laterally*]{} coupled) or in the perpendicular direction ([*vertically*]{} coupled). Especially for [*laterally*]{} coupled quantum dots only a limited number of studies have appeared.[@Veikon_paperi; @Wensauer; @Kolehmainen; @Jannouleas] Classical studies serve as a good starting point for more demanding quantum mechanical calculations. Moreover, the study of classical electrons in [*vertically*]{} coupled artificial atoms has revealed interesting structural transitions in the ground state electron configurations as a function of the distance between the atoms.[@PartoensClassicalPos] Apart from quantum dots in the classical limit the point charges in 2D can be used to model also other physical systems. Examples include vortex lines in superconductors and superfluids and electrons on the surface of liquid He (see Ref. and references therein). In the theoretical field, the ground state configurations of a confined classical 2D electron system have been studied in the case of a single artificial atom in Refs. and for the vertically coupled artificial atom molecule as a function of the inter-atom distance in Ref. . Recently, also some experimental studies of 2D confined charged classical particle systems have appeared to reflect the classical cluster patterns in 2D.[@dustparticles; @exp_classical] Classical point charges in a two-dimensional infinite plane crystallise into a hexagonal lattice at low temperatures. Parabolic confinement in the artificial atom, on the other hand, favours circularly symmetric solutions. The ground state configuration is thus determined by two competing effects, circular symmetry and hexagonal coordination, thus resulting in non-trivial particle configurations. The reported configurations of the electron clusters in a single artificial atom do not all agree between different studies. The differences can be partly explained by the different forms of confinement and interaction potentials. However, when the number of particles, $N$, confined in the atom is one of the following $N = 2-5,7,10,12,14,19$ all results are in agreement while differences appear for $N = 6,8,9,11,13,15-18,20$ (for $N \leq 20$). In this paper we consider two laterally coupled artificial atoms and classical electrons in the molecule. The changes in the ground state electron configurations are studied for $N = 2 - 20 $ electrons in the molecule as the inter-atom distance is changed. The energies of the metastable states are also calculated at different distances and their role in the structural transitions in the ground state electron configurations is discussed. We also reproduce electron configurations of the single parabolic artificial atom. The differences between different calculations and experimental results are discussed in the limit of single atom. Monte Carlo simulation ====================== The classical electrons in the artificial atom molecule are modelled with the Hamiltonian $$\begin{aligned} H &=& \frac{1}{2} \ m^* \omega_0^2 \ \sum_{i=1}^{N} \min \left[ ({\vec r_i} - d/2 )^2, ({\vec r_i}+d/2 )^2 \right] \nonumber \\ &+& \frac{e^2}{4 \pi \epsilon_0 \epsilon} \ \sum_{i<j}\frac{1}{|{\vec r_i} - {\vec r_j}|}.\end{aligned}$$ Each of the $N$ electrons is described with coordinates $\vec{r_i} = (x_i, y_i)$ in two-dimensional space. The harmonic confinements are positioned symmetrically around the origin with distance $d$ between the minima. $m^*$ is the electron effective mass, $\omega_0$ the confinement strength and $\epsilon$ the dielectric constant. We measure the energy in meV and distance in Å. The confinement strength was set to $\hbar \omega_0 = 3$ meV and typical GaAs parameters were chosen to the effective mass and the dielectric constant: $m^* = 0.067 \ m_e$ and $\epsilon = 13$. The calculated energy values and distances could be scaled to correspond to different values of $\omega_0, \ m^*, \ \epsilon$, but changing the parameters also changes the effective distance between the minima, $d$, and then the minimum energy configuration may not be the same anymore. Therefore we have have one significant parameter in the system, $d$, which scales as $ \propto (m^* \omega_0^2 \epsilon)^{-1/3}$. The minimum energy as a function of the positions of the particles, $ E_{tot} = \min \ E({\vec r_1},...,{\vec r_N})$, is solved with a standard Metropolis Monte Carlo method [@Metropolis] starting from a randomly chosen initial electron configuration ${\vec r_1},...,{\vec r_N}$. The accuracy and simulation time needed with the Metropolis algorithm was found to be well sufficient for the current problem. We compared the calculated energies in the limit of a single artificial atom to those given in Ref. , and the results were found to be in complete agreement within the given accuracy. In the simulations we choose four different distances between the atoms and perform 300 test runs at each particle number ($N = 2 - 20 $) and distance ($d = 0, 200, 600$ and $1000$ Å). In addition to minimum energy configurations we also obtain metastable states that are higher in energy compared to the ground state. When the ground and metastable states are obtained at $d = 0, 200, 600, 1000$ Å  we study the structural transitions between ground state electron configurations at the intermediate distances. The electron configurations obtained from the fixed $d$ calculations are taken as an input to Monte Carlo minimisations where the attempt step is set so small that the electron configuration cannot change to another. Then the distance is slightly altered ($d \rightarrow d \pm 1$ Å) and a new energy with slightly modified positions is calculated for the configuration defined by the input. The calculated new configuration is taken as an input to the next calculation with a new distance between the atoms, and so all distances between $d = 0,200,600,1000$ Å are well sampled. However, it may happen that a configuration becomes unstable as the distance is changed. In that case the simulation converges to some other stable configuration, which can be seen as a sudden jump to a new energy value in the $E(d)$-plots. The energies of all states are studied as a function of the distance and structural transitions in the ground state configurations are examined. Results ======= The results for the ground and metastable states are summarised in Table \[Tab:N\]. The electron configurations are given at four different distances between the artificial atoms. The ground state energy and the corresponding configuration at the four distances is represented in the row following the particle number $N$. If there exist metastable states at the given $N$ and $d$ the energy difference $\Delta E/N$ to the ground state and the electron configuration for the metastable state is also reported. However, not all metastable configurations are marked in Table \[Tab:N\], since when starting the simulation from random positions more electrons can be trapped in one artificial atom than in the other. Only metastable states with either the same number of electrons per atom (for even $N$) or only one more at one than the other atom (for odd $N$) are reported. The notation for the configurations in a single artificial atom is chosen so that electrons are thought to be organised in (nearly) concentric shells around the potential minimum: (N$_1$,N$_2$,N$_3$), where N$_1$ denotes the number of electrons in the innermost shell, N$_2$ the next shell and N$_3$ the number of electrons in the outermost shell. (For $N \leq 20$ only three shells are occupied). For laterally coupled two-atom artificial molecule we have chosen the following notation for configurations: At $d = 200$ Å the configuration is marked as if it would still be on a single atom centred around the midpoint connecting the two atoms. At $d = 600$ and $1000$ Å  the configurations are given as configurations of two separated atoms. For example, as Table \[Tab:N\] and Fig \[Fig:NN\] (a) indicate, with eight particles in the single artificial atom ($d = 0$ Å) the ground state is (1,7), one electron in the centre and seven electrons in the circular shell, and there exist no metastable states. At $d = 200$ Å a new ground state has appeared with configuration (2,6) (Fig. \[Fig:NN\] (b)) while (1,7) has changed to a metastable state (see also Table \[Tab:N\]). At distances $d = 600 $ Å  and $d = 1000 $ Åthe notation is changed to two-atom configurations and for $N = 8$ the ground state is marked with (4),(4), see Figs. \[Fig:NN\] (c) and (e). The notation for configurations is not always exhaustive. The relative orientation of different shells and the relative orientations of the configurations of the two atoms at $d = 600, 1000$ Å do not always become clear from Table \[Tab:N\]. For example, when either or both atoms are left with four electrons ($N = 7,8,9$), the orientation of the square(s) relative to the other atom changes as the distance is increased. At smaller distances the position of the square of four electrons is such that the tips of the squares are in the same line with the positions of the minima (see Fig. \[Fig:NN\] (c) for $N = 8$ at $d = 600$ Å). As the distance is increased the square (or two squares with $N = 8$) turns onto its side (see Fig. \[Fig:NN\] (e) for $N = 8$ at $d = 1000$ Å). For $N = 8$ at $d = 600$ Å there also exists a metastable state where one of the squares is lying on its side and the other on the tip (Fig. \[Fig:NN\] (d)). Even though we divide electrons into shells in our notation it does not mean that the shells are strictly circular even for $d = 0$. This can be seen clearly for $N = 12$ in Fig \[Fig:NN\] (f), where the outer shell resembles more like an incomplete triangle with the tips missing. The configuration marked with (1,5)$^*$ in the two-atom configurations in Table \[Tab:N\] cannot be identified strictly to (6) but neither to (1,5). Therefore we choose the notation (1,5)$^*$ to describe the configuration. The difference between (1,5)$^*$ and (1,5) can be seen with the (1,5)$^*$,(1,5) metastable state in Fig. \[Fig:NN\] (g). The configurations of the ground and metastable states for $N = 13$ at $d = 200$ Å are marked in the same way in the Table, but are different as can be seen in Figs. \[Fig:NN\] (h) and (i). For the highest energy metastable state for $N = 17, d = 200$ Å the two-atom notation would have described the configuration better, Fig. \[Fig:NN\] (l). One metastable state for $N = 19$ and the ground state and one metastable state for $N = 20$ at $d = 200$ Å could not be described with the shell structure notation. The configurations are depicted in Figs. \[Fig:NN\] (m),(n) and (o), respectively. $\begin{array}{|c|c|c|} \hline \includegraphics*[height=25mm]{Fig_1a.ps} & \includegraphics*[height=25mm]{Fig_1b.ps} & \includegraphics*[height=25mm]{Fig_1c.ps} \\ \hline \includegraphics*[height=25mm]{Fig_1d.ps} & \includegraphics*[height=25mm]{Fig_1e.ps} & \includegraphics*[height=25mm]{Fig_1f.ps} \\ \hline \includegraphics*[height=25mm]{Fig_1g.ps} & \includegraphics*[height=25mm]{Fig_1h.ps} & \includegraphics*[height=25mm]{Fig_1i.ps} \\ \hline \includegraphics*[height=25mm]{Fig_1j.ps} & \includegraphics*[height=25mm]{Fig_1k.ps} & \includegraphics*[height=25mm]{Fig_1l.ps} \\ \hline \includegraphics*[height=25mm]{Fig_1m.ps} & \includegraphics*[height=25mm]{Fig_1n.ps} & \includegraphics*[height=25mm]{Fig_1o.ps} \\ \hline \end{array}$ ---- --------- -------------- ---------- --------- -------------- ---------- --------- -------------- --------------------- --------- -------------- ------------- N $E/N$ $\Delta E/N$ config. $E/N$ $\Delta E/N$ config. $E/N$ $\Delta E/N$ config. $E/N$ $\Delta E/N$ config. \[meV\] \[meV\] \[meV\] \[meV\] \[meV\] \[meV\] \[meV\] \[meV\] 2 2.736 (2) 1.777 (2) 0.875 (1),(1) 0.546 (1),(1) 3 4.780 (3) 3.894 (3) 2.940 (1),(2) 2.541 (1),(2) 4 6.696 (4) 5.588 (4) 4.351 (2),(2) 3.795 (2),(2) 5 8.531 (5) 7.340 (5) 5.915 (2),(3) 5.244 (2),(3) +0.099 (1,4) 6 10.231 (1,5) 8.939 (6) 7.234 (3),(3) 6.395 (3),(3) +0.074 (6) +0.033 (1,5) 7 11.816 (1,6) 10.459 (1,6) 8.664 (3),(4) 7.720 (3),(4) 8 13.384 (1,7) 11.933 (2,6) 9.932 (4),(4) 8.850 (4),(4) +0.016 (1,7) +0.003 (4),(4) 9 14.913 (2,7) 13.335 (2,7) 11.246 (4),(5) 10.097 (4),(5) +0.022 (1,8) 10 16.361 (2,8) 14.680 (2,8) 12.441 (5),(5) 11.195 (5),(5) +0.012 (3,7) 11 17.746 (3,8) 16.053 (3,8) 13.701 (5),(1,5)$^*$ 12.361 (5),(1,5) +0.003 (2,9) 12 19.111 (3,9) 17.354 (3,9) 14.873 (1,5)$^*$,(1,5)$^*$ 13.434 (1,5),(1,5) +0.011 (4,8) +0.004 (4,8) +0.011 (1,5)$^*$,(1,5) 13 20.433 (4,9) 18.624 (4,9) 16.048 (1,5)$^*$,(1,6) 14.511 (1,5),(1,6) +0.024 (4,9) 14 21.738 (4,10) 19.854 (4,10) 17.168 (1,6),(1,6) 15.518 (1,6),(1,6) +0.014 (5,9) 15 23.010 (5,10) 21.072 (5,10) 18.302 (1,6),(1,7) 16.587 (1,6),(1,7) +0.029 (1,5,9) +0.035 (5,10) +0.234 (1,6),(2,6) 16 24.259 (1,5,10) 22.271 (6,10) 19.373 (1,7),(1,7) 17.583 (1,7),(1,7) +0.009 (5,11) +0.006 (5,11) +0.024 (1,7),(2,6) 17 25.473 (1,6,10) 23.448 (6,11) 20.468 (1,7),(2,7) 18.611 (1,7),(2,7) +0.005 (1,5,11) +0.010 (1,6,10) +0.006 (1,7),(1,8) +0.018 (1,7),(1,8) +0.016 (1,5,11) +0.023 (2,6),(2,7) +0.018 (6,11) +0.041 (2,6),(1,8) 18 26.660 (1,6,11) 24.597 (1,6,11) 21.522 (1,8),(2,7) 19.579 (2,7),(2,7) +0.026 (1,7,10) +0.006 (6,12) +0.001 (1,8),(2,7) +0.017 (2,7),(1,8) 19 27.841 (1,6,12) 25.728 (1,6,12) 22.572 (2,7),(2,8) 20.569 (2,7),(2,8) +0.003 (1,7,11) +0.004 (1,7,11) +0.001 (1,8),(2,8) +0.016 (1,8),(2,8) +0.016 - 20 29.000 (1,7,12) 26.843 - 23.583 (2,8),(2,8) 21.585 (2,8),(2,8) +0.024 (1,6,13) +0.001 - +0.003 (2,7,11) ---- --------- -------------- ---------- --------- -------------- ---------- --------- -------------- --------------------- --------- -------------- ------------- As the distance between the atoms is increased it is not always clear whether the electrons just follow the two atoms drawn apart and continuously change to two separated atoms. Sometimes metastable states change to a ground state and the ground state to a metastable state as the distance between the atoms is increased. The clearest example can be seen in Table \[Tab:N\] for six electrons between $d = 0$ and $d = 200 $ Å. At $d = 0$ the (1,5) configuration is the ground state and (6) the metastable state. At $d = 200$ Å it is the other way around: (6) is the ground and (1,5) a metastable state. The energy as a function of distance for two alternative configurations is shown in Fig.\[Fig:N6N8N16N19\] (a). The transition point, marked with a small circle, is at $d = 111.6 $ Å. The transition is continuous with respect to energy as a function of distance, but the curvature of the $E(d)$-curve is different and therefore the first derivative of energy with respect to $d$ is discontinuous. This is a first-order discontinuous structural transition in the electron configuration. Hereafter, by discontinuous structural transitions we mean the qualitative change in the ground state electron configuration which is discontinuous with respect to $\partial E/\partial d$ at the transition point. ![image](Fig_2a.ps){height="60mm"} ![image](Fig_2b.ps){height="60mm"} ![image](Fig_2c.ps){height="60mm"} ![image](Fig_2d.ps){height="60mm"} ![image](Fig_2e.ps){height="60mm"} In addition to $N = 6$, for $N = 8, 11, 16$ and $19$ one qualitative change in the electron configuration is observed as a function of $d$. For $N = 8$ at $d = 135.9$ Å the electron configuration changes from (1,7) to (2,6), see Fig \[Fig:N6N8N16N19\] (b). Notice that the (2,6) configuration is not stable at the limit of one atom and becomes unstable approximately below $d = 17$ Å. For $N = 11$ there exists one metastable state, (2,9), at $d = 200$ Å, which at $214.2$ Å changes to a ground state as is depicted in Fig. \[Fig:N6N8N16N19\] (c). With $N=16$ the configuration changes from (1,5,10) to (6,10) at $d = 102.4 $ Å. The energy differences between ground and metastable states are small and metastable states exist only in the proximity of the transition point (Fig. \[Fig:N6N8N16N19\] (d)). At $d = 80 $ Å  the (1,5,10) configuration which is plotted below the energy curve is still the ground state and the (6,10) configuration plotted above the energy curve has appeared as a metastable state. For $N = 19$ Fig. \[Fig:N6N8N16N19\] (e) shows how the (1,6,12) configuration, which is the ground state at $d = 200 $ Å, has changed to a rather unsymmetric configuration at $d = 250$ Å. This unsymmetric configuration changes continuously to (2,8),(1,8) (lower middle plot in (e)) which at $586.0 $ Å  changes discontinuously to (2,8),(2,7) configuration. The (2,8),(2,7) configuration appears approximately at $415$ Å as a metastable state. The (1,8),(2,8) persists as a metastable state to the greatest studied distance of $1000 $ Å as is also indicated in Table \[Tab:N\]. The other metastable states at $d = 0$ and $200$ Å do not change to a ground state. ![\[Fig:N17N18\] Energy as a function of distance for $N = 17$ and $N = 18$ with the ground state configurations along the energy curve. The small circles indicate the discontinuous structural transition points.](Fig_3a.ps "fig:"){height="60mm"} ![\[Fig:N17N18\] Energy as a function of distance for $N = 17$ and $N = 18$ with the ground state configurations along the energy curve. The small circles indicate the discontinuous structural transition points.](Fig_3b.ps "fig:"){height="60mm"} More than one discontinuous transformation in the electron configurations is found for $N = 17$ and $18$, see Figs. \[Fig:N17N18\] (a) and (b). For $N = 17$ the (1,6,10) changes to (6,11) at $d = 145.0$ Å. The (6,11) configuration parts to (1,7),(1,8) two-atom configuration, which is the ground state up to $d = 501.1$ Å, where the (2,7) configuration of the other atom becomes more stable than (1,8), thus (1,7),(1,8)$\rightarrow$(1,7),(2,7). The (1,7),(2,7) configuration persists as a ground state to greater distances. Qualitatively similar transformations are seen for $N = 18$ as for $N = 17$. First the centred cluster (1,6,11) changes to an open configuration (6,12) at $d = 233.9$ Å. The open configuration follows the separation of atoms adopting the configuration (1,8),(1,8), where as in $N = 17$, the (2,7) becomes more stable than (1,8). However, we now see two discontinuous transformations. The first at $d = 522.8$ Åwhen (1,8),(1,8)$\rightarrow$(2,7),(1,8) and the second at $d = 593.4$ Å when (2,7),(1,8)$\rightarrow$(2,7),(2,7). ![\[Fig:N12N14N20\] [**(a)**]{} - [**(c)**]{} Ground state electron configurations along $E(d)$ curve for $N = 12, 14$ and $20$.](Fig_4a.ps "fig:"){height="60mm"} ![\[Fig:N12N14N20\] [**(a)**]{} - [**(c)**]{} Ground state electron configurations along $E(d)$ curve for $N = 12, 14$ and $20$.](Fig_4b.ps "fig:"){height="60mm"} ![\[Fig:N12N14N20\] [**(a)**]{} - [**(c)**]{} Ground state electron configurations along $E(d)$ curve for $N = 12, 14$ and $20$.](Fig_4c.ps "fig:"){height="60mm"} For other electron numbers besides the reported $N = 6,8,16,17,18,19$ we do not observe discontinuous structural transitions in the electron configuration as the distance between the two atoms is increased. A few examples of continuous electron configuration changes are shown in Fig. \[Fig:N12N14N20\]. For $N = 12$ the (3,9) configuration transforms continuously to resemble the (6),(1,5)$^*$ two-atom configuration. Between $200$ Å and $450$ Å the row of electrons pushes itself forward when the atoms move apart, resulting in the symmetric configuration (1,5)$^*$,(1,5)$^*$. For $N = 14$ the electron configuration follows the separation of atoms in a symmetrical form, but after $d = 250$ Å both sides start to twist towards the (1,6),(1,6) two-atom configuration. For $N = 20$ the transformation is hard to describe, but it is continuous. The three distinct metastable states at $d = 0$ and $200$ Å, seen in Table \[Tab:N\], never change to a ground state and vanish at other distances. ![\[Fig:dmyy\] Change in the chemical potential ($E_{N+1} / (N+1) - E_N / N$) at $d = 0, 200, 600$ and $1000$ Å.](Fig_5.ps){height="60mm"} The changes in energy per particle of the ground states at the four studied distances as a function of $N$ are shown in Fig. \[Fig:dmyy\]. At $d =$ 0 there are small troughs at $N = 3, 6, 10$ and $17$, at adding the fourth, seventh, eleventh and eighteenth particle. Moving to greater distances between the atoms, the change in the chemical potential is clearly peaked. Going to an odd number of particles increases the chemical potential much more than going to an even number of particles. Interesting is the intermediate distance of $d =$ 200 Å where this trend is observed for $N = 2,3,4$ and $9,10,11$, but otherwise the curve shows no clear structure and does not strictly follow the shape of the $d = 0$ Å curve either. \[sec:discussion\]Discussion ============================ At $d = 0$ (single artificial atom) our results are in agreement with other Monte Carlo (MC) and molecular dynamics (MD) studies with parabolic confinement and pure Coulomb interaction. [@Bolton_Superlatt; @BedanovPeeters; @Schweigert; @Ying-Ju; @Date; @Kong] However, for $N = 17$ Bolton [[*et al. *]{}]{} [@Bolton_Superlatt] obtain the (1,5,11) configuration instead of (1,6,10) which our and other calculations[@BedanovPeeters; @Schweigert; @Ying-Ju; @Kong] predict. Ref. may contain an error since in a later work by Bolton[@Bolton's_thesis] the configuration for $N = 17$ was reported to be (1,6,10). There is also a difference for $N = 21$ in Ref. which was later corrected.[@Bolton's_thesis] Besides calculating the ground state configurations Kong [[*et al. *]{}]{} [@Kong] also examined metastable states for $N = 1 - 40$. Our results are in agreement (we calculated configurations only for $N \leq 20$) both in ground and metastable states except that for $N = 18$ Kong [[*et al. *]{}]{}found two metastable states whereas we see only one. In addition to the (1,6,11) and (1,7,10) they also obtain (6,12) as a metastable state. We repeated the simulation with 3000 independent test runs, but were still unable to find the (6,12) configuration. We can conclude that different calculations for the $r^2$ confinement and $1/r$ interaction potential are in good general agreement. The few experiments on charged particles trapped in 2D as well as calculations with different forms of interaction and confinement potentials reveal also different configurations for the cluster patterns. The interaction between the particles could be logarithmic, which is the case with infinite charged lines moving in 2D (vortex lines etc.) or perhaps Yukawa type with a strong but short-range repulsion (screened Coulomb interaction). The form of the confinement is usually chosen to be parabolic (Lai and I [@Ying-Ju] tested also a steeper confinement with $r^4$ contribution). However, for the question whether the potential in experiments with clusters is parabolic there is no clear answer. Therefore it is not surprising that the experiments and also calculations with different functional forms of interaction and confinement result in different cluster patterns. Lai and I [@Ying-Ju] calculated and summarised the configuration patterns with different interactions and tested also $r^4$ contribution to the confinement and compared the results to dust particle experiments. [@dustparticles] Saint Jean [[*et al. *]{}]{} [@exp_classical] measured the configurations with electrostatically interacting charged balls of millimeter size moving on a plane conductor. They made a comparison with simulations and quite surprisingly found the best agreement with a relatively old simulation with vortex lines in a superfluid[@Campbell] with logarithmic interaction, which again was not in agreement with the dust particle experiments[@dustparticles] nor with the purely logarithmic interaction used by Lai and I. [@Ying-Ju] Despite the differences there are some particle numbers where the configuration seems to be the same regardless of the experiment or functional form of the interaction or confinement. These particle numbers are $N = 3,4,5,7,10,12,14,19$. Partoens [[*et al. *]{}]{}[@PartoensClassicalPos] examined the ground states of even number of classical electrons evenly distributed in two [*vertically*]{} coupled artificial atoms as a function of the distance between the atoms. As in our study of [*laterally*]{} coupled atoms discontinuous transitions between configurations occur as a function of the distance $d$. The difference is that for vertically coupled atoms one can see intuitively that transitions should occur between $d = 0$ and $d \rightarrow \infty$ whereas in laterally coupled atoms the configurations can be pulled apart with some $N$ without qualitative (discontinuous) changes in the electron configurations. For vertically coupled atoms discontinuous transitions (first order with respect to energy) in electron configurations are observed for all even $N \leq 20$ whereas for laterally coupled atoms we see also purely continuous changes as $d \rightarrow \infty$. To summarise, we have calculated ground and metastable state configurations of classical point charges confined in two dimensions with two laterally coupled parabolic potential wells. Ground and metastable electron configurations were studied as a function of the distance between the atoms and discontinuous (in $\partial E / \partial d$) transitions in the ground state configurations were observed for particle numbers $N = 6,8,11,16,17,18,19$. The configurations of purely classical electrons in laterally coupled two-minima potential have an interesting and complex spectrum as the distance between the minima is changed. This work has been supported by the Academy of Finland through its Centers of Excellence Program (2000-2005). [31]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , , , , ****, (). , , , ****, (). , , , , ****, (). , , , ****, (). , , , . , , , ****, (). , , , , ****, (). , , , ****, (). , , , ****, (). , ****, (). , , , ****, (). , , , , ****, (). , , , , ****, (). , **** (). , , , , , ****, (). , , , , , , , ****, (). , , , , , ****, (). , , , ****, (). , , , (). , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , , , , , ****, (). , , , ****, (). , , , ****, (). , Ph.D. thesis, ().
{ "pile_set_name": "ArXiv" }
--- abstract: | Reservoir Computing (RC) is a well-known strategy for designing Recurrent Neural Networks featured by striking efficiency of training. The crucial aspect of RC is to properly instantiate the hidden recurrent layer that serves as dynamical memory to the system. In this respect, the common recipe is to create a pool of randomly and *sparsely* connected recurrent neurons. While the aspect of sparsity in the design of RC systems has been debated in the literature, it is nowadays understood mainly as a way to enhance the efficiency of computation, exploiting sparse matrix operations.\ In this paper, we empirically investigate the role of sparsity in RC network design under the perspective of the richness of the developed temporal representations. We analyze both sparsity in the recurrent connections, and in the connections from the input to the reservoir. Our results point out that sparsity, in particular in input-reservoir connections, has a major role in developing internal temporal representations that have a longer short-term memory of past inputs and a higher dimension. author: - bibliography: - 'references.bib' title: 'Sparsity in Reservoir Computing Neural Networks [^1] ' --- Reservoir Computing, Echo State Networks, Short-term Memory, Sparse Recurrent Neural Networks Introduction ============ Recurrent Neural Networks (RNNs) [@kolen2001field] are a fundamental tool for adaptive processing of dynamically evolving information, with excellent performance in fields such as time-series forecasting [@laptev2017time], machine translation [@sutskever2014sequence], speech and text processing [@graves2013speech; @nallapati2016abstractive], just to mention a few. An increasing number of works are analyzing the role of sparsity in the design of trained (dynamical) neural networks systems, for example through pruning [@narang2017exploring] or re-wiring [@bellec2017deep] connections. The characterization emerging from these studies is that having sparse connections between neurons is not only advantageous in computational terms - as it enables fast sparse matrix computations - but can also be beneficial to obtain a better performance in practice. Moreover, in the context of neurobiologically-inspired information processing systems, a sparse degree of connectivity between neurons has been shown to improve the quality of the developed internal representations [@litwin2017optimal]. Interestingly, the optimal amount sparsity in the numerical simulations matched observed properties of cerebellum-like circuits. Reservoir Computing (RC) neural networks [@lukovsevivcius2009reservoir; @schrauwen2007overview; @jaeger2004harnessing] represent an intriguing development in the field of RNNs. In RC, the recurrent hidden layer of a RNN is left untrained after initialization subject to asymptotic stability conditions of the corresponding dynamical system. As a result, learning is applied only to a simple readout component with striking advantages in terms of required training times compared to fully trained RNNs. Pushing the involved algorithms towards extreme simplicity and efficiency makes the RC approach very well suited for real-world application scenarios featured by (possibly severe) resource constraints, such as neuromorphic hardware implementations [@larger2012photonic] or cyber-physical systems where the learning modules are embedded at the edge [@bacciu2014experimental]. A typical strategy in the design of RC networks is to setup the recurrent layer in a sparse way. The initial intuition was that sparsity in the recurrent untrained layer could enable a decoupling of state variables and hence richer representations [@jaeger2001echo]. Successively, several authors pointed out empirical evidences contrary to the initial intuition (see, e.g., [@schrauwen2007overview; @xue2007decoupled; @gallicchio2011architectural]). Currently, the sparse design of reservoirs is commonly understood mainly as a way to speedup state computations, without a practical effect on the resulting performance. However, the impact of sparsity on the performance of RC neural networks has been typically studied limited to the recurrent connections only. In this paper, we intend to shed more light on the role of sparsity in RC by extending the analysis to both recurrent and input connections. Specifically, we empirically show the effect of recurrent and input sparsity in reservoirs, evaluated by means of short-term memory capacity and effective dimension of the resulting state trajectories. The rest of this paper is structured as follows. We introduce the basics of RC methodology in Section \[sec.RC\], discussing initialization and sparsity of reservoirs. Then, in Section \[sec.richness\] we present the concepts of short-term memory capacity and effective reservoir dimension. Our experimental analysis is described in Section \[sec.experiments\]. Finally, in Section \[sec.conclusions\] we draw our conclusions and sketch possible developments. Reservoir Computing Neural Networks {#sec.RC} =================================== Here we give a brief description of the RC design methodology for RNNs, focusing on the Echo State Network (ESN) [@jaeger2004harnessing; @jaeger2001echo] model. An RC network is a neural information processing system that treats data in the form of (temporal) sequences. Architecturally, the neural network is composed by a hidden recurrent layer called *reservoir*, and an output layer called *readout*. Fig. \[fig.esn\] illustrates the building blocks of a typical RC network. In what follows, we denote the number of reservoir neurons, i.e., the reservoir dimension, by $N$, and the state of the reservoir system at time $t$ by ${\mathbf{h}}(t) \in {\mathbb{R}}^{N}$. This state is evolved by following a state update equation: $$\label{eq.reservoir} {\mathbf{h}}(t) = \tanh({\mathbf{U}}{\mathbf{x}}(t) + {\mathbf{W}}{\mathbf{h}}(t-1)),$$ where ${\mathbf{x}}(t) \in {\mathbb{R}}^{M}$ is the M-dimensional input at time $t$, ${\mathbf{U}}\in {\mathbb{R}}^{N \times M}$ is the input weight matrix, modulating the influence of the external input on the current state, and ${\mathbf{W}}\in {\mathbb{R}}^{N \times N}$ is the recurrent weight matrix, which controls the impact of previous state on the current state. The state is typically set to a zero vector as initial condition, i.e. ${\mathbf{h}}(0) = \mathbf{0} \in {\mathbb{R}}^{N}.$ Note that here we dropped from the reference to bias terms to focus the analysis on the external stimulating input signal alone. Both weight matrices ${\mathbf{U}}$ and ${\mathbf{W}}$ remain untrained after initialization (see Section \[sec.initialization\]). The reservoir system is coupled with a linear readout layer that computes an $L$-dimensional output at each time step, i.e. ${\mathbf{y}}(t) \in {\mathbb{R}}^{L}$, as an affine transformation of the reservoir state: $$\label{eq.readout} {\mathbf{y}}(t) = {\mathbf{V}}{\mathbf{h}}(t) + \mathbf{b},$$ where ${\mathbf{V}}\in {\mathbb{R}}^{L \times N}$ is a readout weight matrix and $\mathbf{b} \in {\mathbb{R}}^{L}$ is a bias vector (that assumes a constant unitary input bias for the readout). The readout parameters are the only ones that undergo a training process, typically in closed-form fashion by using pseudo-inversion [@lukovsevivcius2009reservoir]. Initialization of Reservoirs {#sec.initialization} ---------------------------- The fundamental characterization of RC neural networks is that all the reservoir parameters remain untrained after initialization. Such initialization is performed in agreement to asymptotic stability conditions expressed by the Echo State Property (ESP) [@yildiz2012re; @jaeger2001echo; @gallicchio2019chasing], which essentially require to control the magnitude of the weights in ${\mathbf{U}}$ and ${\mathbf{W}}$. Usually, both the input weights in ${\mathbf{U}}$ and the recurrent weights in ${\mathbf{W}}$ are randomly drawn from a uniform distribution in $[-1,1]$. After that, the elements in ${\mathbf{U}}$ are re-scaled by a factor $\omega_{in}$, which takes the role of input scaling. The weights in ${\mathbf{W}}$ are re-scaled to control the largest absolute eigenvalue, i.e., the spectral radius $\rho$, typically to a value smaller than 1 [@jaeger2001echo]. The design strategy of the reservoir topology (i.e., the way in which the reservoir neurons are connected among each other) has been subject of several studies in literature (see, e.g., [@strauss2012design; @rodan2010minimum]). While some of the proposed reservoir organizations can be beneficial in specific application circumstances, a random and sparse topological organization of the reservoir is the architecture of choice in general cases. This is the focus of our analysis in this paper. Making the connections among reservoir neurons sparse has the fundamental practical advantage to reduce the cost of state update operations in . Actually, for densely connected reservoirs (and assuming $N >> M$) the cost of state updating scales as $\mathcal{O}(N^2)$, i.e. quadratically with the reservoir size. A first approach to make the reservoir sparsely connected would be to impose a (small) fixed percentage, say $C$, of non-zero weights in the involved weight matrices matrices. Although reducing the running times in practice for smaller reservoirs, this approach would asymptotically scale as $\mathcal{O}(N^2 \, C/100)$, hence still quadratically with the reservoir size. A more effective approach, which is adopted in this paper, is to fix the number, say $\chi_R$, of incoming recurrent connections for each reservoir unit. This indeed makes the state update cost as small as $\mathcal{O}(N \, \chi_R)$, i.e. scaling only linearly with the number of neurons in the reservoir. A similar strategy can be adapted for the setup of the input connections. In this case, to ensure that each input dimension is actually forwarded to the reservoir, we fix the number of outgoing connections from each input units, denoted as $\chi_I$. The sparse architectural reservoir setup used in this paper is exemplified in Figure \[fig.sparsity\]. Notice that in this case, every row of ${\mathbf{W}}$ has exactly $\chi_R$ non-zero values, and every column of ${\mathbf{U}}$ has exactly $\chi_I$ non-zero elements, with both $\chi_R$ and $\chi_I$ being not greater than $N$. Short-term Memory and Effective Reservoir Space Dimension {#sec.richness} ========================================================= The role of the recurrent reservoir system is to embed the input time-series into an internal “state” representation, given by the activation of the reservoir neurons over time. Here we analyze the quality of such internal reservoir representation by quantifying its short-term memory and effective dimension.\ *Short-term Memory Capacity (MC)* [@jaeger2001short] tests the ability of a recurrent neural system to reconstruct its driving input time-series from the transient state dynamics. More in detail, the reservoir is driven by a uni-dimensional time-series, $x(t), t = 1, 2, \ldots$, and different readout units are trained to recall progressively delayed versions of the input. I.e., the i-th readout unit $y_i(t)$ should approximate $x(t-i)$. The MC of an RC network is then quantified as follows: $$\label{eq.mc} MC = \sum_{i = 1}^\infty \frac{cov^2(x(t-i),y_i(t))}{\sigma^2(x(t-i)) \sigma^2(y_i(t))},$$ i.e., as the sum of squared correlation coefficients of the delayed input and reconstructed signals.\ *Effective Dimension ($N_{eff}$)*[@abbott2011interactions; @litwin2017optimal] is a measure of the number of orthogonal directions in the neuronal system’s state trajectory over time. While the evolution of the reservoir system in is described by an $N$-dimensional state vector ${\mathbf{h}}(t)$, the actual reservoir trajectory lies into a lower-dimensional manifold whose dimension can be quantified as follows: $$\label{eq.dimension} N_{eff} = \frac{(\sum_{i = 1}^N \lambda_i)^2}{\sum_{i=1}^N\lambda_i^2},$$ where $\lambda_i$, $i = 1, 2, \ldots, N$, denote the eigenvalues of the covariance matrix of the reservoir state activation over time. When measured for a reservoir under the driving influence of an external time-series, gives an estimate of the number of directions of reservoir state variability that are (linearly) uncorrelated along the observed trajectory. Experimental Analysis {#sec.experiments} ===================== We measured the short-term memory (MC) and the effective reservoir dimension ($N_{eff}$) introduced in Section \[sec.richness\] for RC networks varying the amount of recurrent and input connections. Our experimental settings are described in Section \[sec.settings\], while the results are reported in Section \[sec.results\]. Settings {#sec.settings} -------- We used a uni-dimensional signal as driving input for the reservoir (i.e., $M = 1$). To maximally test the intrinsic quality of reservoir representations, we used iid randomly sampled inputs $x(t)$ from a uniform distribution (in $[-0.8, 0.8]$). The length of the generated input time-series was $6000$, and the number of reservoir neurons was fixed to $N = 100$. To compute MC, we used the first $5000$ time-steps as training set[^2], using the remaining $1000$ time-steps to assess the MC score. The total number of delays used for the computation of was 200, which is in practice sufficient to account for all the non-negligible contributions for 100-dimensional reservoirs. The last $1000$ time-steps of the dataset were also used to compute the effective reservoir dimension $N_{eff}$ (see ). In our experiments, we used RC networks with spectral radius $\rho = 0.9$ and input scaling $\omega_{in} = 1$. While this setup is of common use in RC practice, we also ran preliminary experiments with other choices of these hyper-parameters, finding that the outcomes are not qualitatively different. We varied both the number of recurrent connections ($\chi_R$) and of input connections ($\chi_I$) from 1 to 100 (with step of 1). For each configuration we averaged the results over 50 reservoir realizations. Results {#sec.results} ------- [0.48]{} \[fig.MC\] [0.48]{} \[fig.dim\] The achieved values of MC and $N_{eff}$ in correspondence of the possible sparsity settings (values of $\chi_R$ and $\chi_I$) are shown in Fig. \[fig.results\]. We can draw two major observations from the results. First, the number of input connections has a decisive impact on both the short-term memory and the effective reservoir dimension of the networks. Indeed, maximally sparse input connections, with $\chi_I = 1$, achieved the highest performances. Interestingly, simply propagating the input to all the reservoir neurons degrades the performance sensibly. Second, the role of sparsity in recurrent connections seems to be much less important. In fact, the trend in Fig. \[fig.results\] indicates that for a given input connectivity, the achieved results are not much sensible to the exact number of recurrent connections (after a minimum number has been exceeded). The results are further detailed in Fig. \[fig.results2\], which shows the best result for each choice of input (resp. recurrent) connectivity in Fig. \[fig.results2\](a) (resp. Fig. \[fig.results2\](b)), as well as the results achieved for maximally sparse input connectivity, i.e. for $\chi_I = 1$, in Fig. \[fig.results2\](c). Figs. \[fig.results2\](a)-(b) confirm the already observed trends. On the one hand the performance of the RC networks tends to deteriorate for less sparse input weight matrices. On the other hand, a modest number of recurrent connections is already sufficient to achieve a performance not far from the highest possible one. For RC networks with $\chi_I = 1$ (Fig. \[fig.results2\](c)), both MC and $N_{eff}$ saturate for fairly small values of $\chi_R$, without appreciable differences for settings with more than $20$ recurrent connections per reservoir neuron. [0.32]{} \[fig.dim\] [0.32]{} \[fig.dim\] [0.32]{} \[fig.MC\] Conclusions {#sec.conclusions} =========== We have empirically analyzed the performance of RC neural networks in relation to sparsity of input and recurrent connections. Our results indicate that under commonly used reservoir configurations, the number of non-zero connections can play a decisive role in determining the richness of the developed representations. In particular, while a modest number of recurrent connections is already sufficient to achieve good performance, we found that maximally sparse input to reservoir connections lead to the best results both in terms of short-term memory and in terms of effective dimension of the state manifold. Overall, our analysis points out a simple rule of thumb for shaping reservoir weight matrices in case of uni-dimensional driving time-series: (i) connect the input to just one reservoir neuron, and (ii) set a small number of incoming recurrent connections ($\approx 20\%$) for each reservoir neuron. The study presented in this paper can be seen as preparatory to opening further and deeper lines of research. First of all, the role of sparsity can be investigated in synergy with *structured* (rather than random) recurrent reservoir topologies, such as those based on cyclic [@rodan2010minimum] or small-world [@kawai2019small] connections. Similarly, the study can be extended towards *deep* RC neural networks [@gallicchio2017deep; @gallicchio2018design], where multiple reservoir layers are connected in a pipeline. In this case, the sparsity of input connections for higher layers has the even more intriguing role of modulating the extent of signal propagation between consecutive internal representations. Neuromorphic hardware implementations [@moughames2019three; @freiberger2019towards; @partzsch2011analyzing; @larger2012photonic] of deep recurrent neural systems are an important example of a domain where such insights can be capitalized in practice. Under a broader perspective, and outside the RC world, the analysis presented here pointed out that a sparse setting of RNN connections brings advantages even before learning of the non-zero connections. How these architectural advantages can be further exploited by (supervised or unsupervised) *training* is another exciting open research question. [^1]: **This paper is currently under review.** This work has been partially supported by the European Union’s Horizon 2020 Research and Innovation program, under project TEACHING (Grant agreement ID: 871385). https://www.teaching-h2020.eu [^2]: We used pseudo-inversion to train the readout, discarding the first 1000 time-steps as initial transient.
{ "pile_set_name": "ArXiv" }
--- abstract: | The general form of the anisotropy parameter of the expansion for Bianchi type-III metric is obtained in the presence of a single diagonal imperfect fluid with a dynamically anisotropic equation of state parameter and a dynamical energy density in general relativity. A special law is assumed for the anisotropy of the fluid which reduces the anisotropy parameter of the expansion to a simple form ($\Delta\propto H^{-2}V^{-2}$, where $\Delta$ is the anisotropy parameter, $H$ is the mean Hubble parameter and $V$ is the volume of the universe). The exact solutions of the Einstein field equations, under the assumption on the anisotropy of the fluid, are obtained for exponential and power-law volumetric expansions. The isotropy of the fluid, space and expansion are examined. It is observed that the universe can approach to isotropy monotonically even in the presence of an anisotropic fluid. The anisotropy of the fluid also isotropizes at later times for accelerating models and evolves into the well-known cosmological constant in the model for exponential volumetric expansion. **Keywords** Bianchi type III $\cdot$ Anisotropic fluid $\cdot$ Dark energy $\cdot$ Isotropization author: - 'Özgür Akarsu[^1]' - 'Can Battal K[i]{}l[i]{}nç[^2]' title: Bianchi type III models with anisotropic dark energy --- -1cm *Ege University, Faculty of Science, Dept. of Astronomy and Space Sciences, 35100 Bornova, [İ]{}zmir/Turkey.* Introduction {#intro} ============ Some large-angle anomalies, which appear to indicate violation of the statistical isotropy, have been found in the cosmic microwave background (CMB) radiation (see for the anomalies). Non-trivial geometry, i.e., broken spherical symmetry, of the universe seems as the most promising explanation of these anomalies (see ). For instance, Jaffe et al. showed that removing a Bianchi component from the WMAP initial data release can account for several large-angle anomalies and leave a statistically isotropic sky. The low value of the quadrupole moment, which has been known since the first COBE results, is the most apparent of these anomalies and still exist in the high resolution WMAP data . Campanelli et al. showed that, in Bianchi type-I framework, allowing the universe to be plane-symmetric with eccentricity (regardless of the origin) at decoupling of order $10^{-2}$ can resolve the quadrupole problem without effecting higher multipoles. They also concluded that WMAP data require an addition to the standart cosmological model that resembles the Bianchi morphology, which are homogeneous but not necessarily isotropic (see for review on Bianchi metrics). Thus somehow in the cosmological models, the universe should have achieved a slightly anisotropic geometry inspite of the inflation. One may classify the models according to whether this occurs at an early time or at late times of the universe. In the context of the former class, the generic inflationary models can be modified in a way to end inflation with slightly anisotropic geometry, e.g. . In the context of the latter class, the isotropy of the space that achieved during inflation can be distorted in the late time acceleration of the universe by modifiying the dark energy (DE), e.g. . In generic inflationary models characterized by an accelerated expansion in the early universe, space is assumed to be homegenous and isotropic from the begining and inflation is driven by a scalar field, which is isotropic. Some authors also studied scalar field models in Bianchi type space-times (e.g. ). However, any possible anisotropy of the Bianchi metrics should have been died away during the inflation . On the other hand, scalar fields may be replaced by vector fields, which give rise to anisotropic EoS parameter, once the metric is generalized to Bianchi type metrics. Inflation by using vector field first introduced by Ford [@Ford], nevertheless it was suffering from fine tuning problem. Recently, Koivisto and Mota have considered several new classes of viable vector field alternatives to the inflaton within in the Bianchi type-I framework. Golovnev et al. [@Golovnev] has constructed a succesful model, which either could give completely isotropic universe or slightly anisotropic universe at the end of inflation. Such models would also provide opportunity to construct more realistic inflationary models than the Bianchi type inflationary models which are driven by a scalar field and isotropize as evolved. Because in such models one generalizes not only the metric, but also the EoS parameter of the inflaton in accordance with the metric, and afterwards it can be showed whether the metric and/or the energy source evolves towards the isotropy. On the other hand, these lowest multipoles represent the scale of the horizon at approximately the DE domination begins, since then it is natural to associate these anomalies with the present acceleration of the universe and the intrinsic nature of the DE rather than inflation (thus, inflaton). An anisotropic DE energy can derive an anisotropic late time acceleration and can break the isotropy that had been achieved during inflation. The paramount characteristic of the DE is a constant or slightly changing energy density as the universe expands, but we do not know the nature of the DE very well (see for reviews on the DE). DE has conventionally been characterized by the equation of state (EoS) parameter $w=p/\rho$ which is not necessarily constant, where $\rho$ is the energy density and $p$ is the pressure. The simplest DE candidate is the vacuum energy ($w=-1$), which is mathematically equivalent to the cosmological constant ($\rm{\Lambda}$). The other conventional alternatives, which can be described by minimally coupled scalar fields, are quintessence ($w\geq-1$), phantom energy ($w\leq-1$) and quintom (that can cross from phantom region to quintessence region as evolved) and have time dependent EoS parameters. In all these models, DE is handled as an isotropic fluid. However there is no a priori reason to assume the DE is isotropic in nature. In principle, the EoS parameter of DE may be generalized by determining the EoS parameter separately on each spatial axis in a consistent way with the considered metric, since the energy density is a scalar quantity but the pressure is vectorial. Such DE candidates can also be studied in the context of vectorial fields and such candidates have been proposed by several authors (see ). Unlike Robertson-Walker (RW) metric Bianchi type metrics can admit a DE that wields an anisotropic EoS parameter according to their characteristics. The cosmological data -from the large-scale structures [@Tegmark] and Type Ia supernovae observations- do not rule out the possibility of an anisotropic DE either . In last years cosmological models in the presence of an anisotropic DE within the Bianchi type-I framework have been studied by several authors. Rodrigues has proposed a $\rm{\Lambda}$CDM cosmological model extension whose DE component preserves its nondynamical character but wields anisotropic vacuum pressure. Koivisto and Mota have presented a two-fluid model in the presence of an anisotropic DE and perfect fluid which are interacting, and presented a vector field action for DE as an example of the possibility of an anisotropic DE. They have shown that such models are cosmologically viable and can explain the large-angle anomalies in the CMB. Koivisto and Mota have investigated cosmologies where the accelerated expansion of the universe is driven by a field with an anisotropic equation of state by introducing two skewness parameters to quantify the deviation of pressure from isotropy. They have studied the dynamics of the background expansion and analyzed a special case of an anisotropic cosmological constant in detail. Akarsu and Kilinc [@Akarsu] have proposed a two-fluid model in the presence of a perfect fluid and dynamical DE which wields dynamical and anisotropic EoS parameter. All of the above studies are based on the idea that an anisotropic fluid gives rise to an anisotropy in the expansion in Bianchi type-I space-time. However, an anisotropic fluid must not necessarily promote the anisotropy in the expansion. Candidates of such energy sources may also act so as to support isotropization of the expansion as has mentioned by Akarsu and Kilinc [@Akarsu] in relatively earlier times and as has shown in this study, within the Bianchi type-III framework, in the entire history of the universe. Thus, even if we observe an isotropic expansion in the present universe we still cannot rule out possibility of DE with an anisotropic EoS. Bianchi type-III cosmological models in the presence of DE have been studied in general relativity in the last thirty years. Moussiaux et al. [@Moussiaux] have given an exact particular solution of the Einstein field equations for vacuum with a cosmological constant. Lorenz [@Lorenz1] has presented a model with dust and a cosmological constant. Chakraborty and Chakraborty have given a bulk viscous cosmological model with variable gravitational constant ($G$) and $\rm{\Lambda}$ in [@Chakraborty]. Singh et al. [@SinghIII] have investigated a model with variable $G$ and $\rm{\Lambda}$ in the presence of perfect fluid by assuming a conservation law for the energy-momentum tensor. Recently, Tiwari [@Tiwari] has studied a model in the presence of perfect fluid and a time dependent $\rm{\Lambda}$ with constant deceleration parameter. Bali & Tinker [@Bali] have investigated a model in the presence of bulk viscous barotropic fluid with variable $G$ and $\rm{\Lambda}$. Letelier [@Letelier] has examined some two-fluid cosmological models, which have similar symmetries to those Bianchi type-III models, where the distinct four-velocity vectors of the two non-interacting perfect fluids generate an axially symmetric anisotropic pressure. In this study we have first obtained the general form of the anisotropy parameter of the expansion for Bianchi type-III metric in the presence of a single diagonal imperfect fluid with a dynamically anisotropic EoS parameter and a dynamical energy density in general relativity. Then we have made an assumption on the anisotropy of the fluid in a way to reduce the anisotropy parameter of the expansion to a simple form and obtained a hypothetical fluid that obeys to a special form of an anisotropic EoS parameter. The exact solutions of the Einstein field equations have been obtained by assuming two different volumetric expansion laws in a way to cover all possible expansions: namely, exponential expansion and power-law expansion. Some features of the evolution of the metric and the dynamics of the anisotropic DE fluid have been examined. It has been shown that, in the Bianchi type-III framework, there can be solutions in which anisotropic fluid does not promote anisotropic expansion. Field equations {#sec:1} =============== We consider the homegenous and anisotropic space-time described by Bianchi type-III metric in the form $$ds^{2}=dt^{2}-A(t)^{2}dx^{2}-B(t)^{2}e^{-2\alpha x}dy^{2}-C(t)^{2}dz^{2}$$ where $A(t)$, $B(t)$ and $C(t)$ are the scale factors (metric tensors) and functions of the cosmic time $t$, and $\alpha\neq0$ is a constant. (Bianchi type-I metric can be recovered by choosing $\alpha=0$, but the underlying Lie algebra of the isometry group of the Bianchi type-I and type-III metrics are completely different [@Ellis2]. This metric does not cover Robertson-Walker metric, but gets its closest form to RW metric when $A(t)=B(t)=C(t)$, thus we may talk about its approaching to isotropy, but not a total isotropization of this metric.) The simplest generalization of the EoS parameter of a perfect fluid may be to determine the EoS parameter separately on each spatial axis by preserving the diagonal form of the energy-momentum tensor in a consistent way with the considered metric as disccused in the introduction. (In fact, within the Bianchi type-III framework we would allow the off-diagonal terms with ${T}_{1}^{2}={T}_{2}^{1}$ to be non-null. However, in this study, we are dealing only with an anisotropic fluid whose energy-momentum tensor is in diagonal form.) Thus, we may write down the energy-momentum tensor of the fluid in the following form, $${{T}_{\nu}}^{\mu}=\mathrm{diag}[{{T}_{0}}^{0},{{T}_{1}}^{1},{{T}_{2}}^{2},{{T}_{3}}^{3}].$$ Then we may parametrize it as follows, $$\begin{aligned} {{T}_{\nu}}^{\mu}=\mathrm{diag}[\rho,-{p_{x}},-{p_{y}},-{p_{z}}]=\mathrm{diag}[1,-{w_{x}},-{w_{y}},-{w_{z}}]\rho\\ \nonumber =\mathrm{diag}[1,-w,-(w+\gamma),-(w+\delta)]\rho,\end{aligned}$$ where $\rho$ is the energy density of the fluid; ${p_{x}}$, ${p_{y}}$ and ${p_{z}}$ are the pressures and ${w_{x}}$, ${w_{y}}$ and ${w_{z}}$ are the directional EoS parameters on the $x$, $y$ and $z$ axes respectively; $w$ is the deviation-free EoS parameter of the fluid. We have parametrized the deviation from isotropy by setting ${w_{x}}=w$ and then introducing skewness parameters $\delta$ and $\gamma$ that are the deviations from $w$ respectively on the $y$ and $z$ axes. $w$, $\delta$ and $\gamma$ are not necesarilly constants and can be functions of the cosmic time $t$. The Einstein field equations, in natural units ($8\pi G=1$ and $c=1$), are $$G_{\mu\nu}=R_{\mu\nu}-\frac{1}{2}Rg_{\mu\nu}=-{T}_{\mu\nu}$$ where $g_{\mu\nu}u^{\mu}u^{\nu}=1$; $u^{\mu}=(1,0,0,0)$ is the four-velocity vector; $R_{\mu\nu}$ is the Ricci tensor; $R$ is the Ricci scalar, ${T}_{\mu\nu}$ is the energy-momentum tensor. In a comoving coordinate system, Einstein’s field equations (4), for the anisotropic Bianchi-III space-time (1), in case of (3), read as $$\frac{\dot{A}}{A}\frac{\dot{B}}{B}+\frac{\dot{A}}{A}\frac{\dot{C}}{C}+\frac{\dot{B}}{B}\frac{\dot{C}}{C}-\frac{\alpha^2}{A^{2}}=\rho,$$ $$\frac{\ddot{B}}{B}+\frac{\ddot{C}}{C}+\frac{\dot{B}}{B}\frac{\dot{C}}{C}=-w\rho,$$ $$\frac{\ddot{A}}{A}+\frac{\ddot{C}}{C}+\frac{\dot{A}}{A}\frac{\dot{C}}{C}=-(w+\delta)\rho,$$ $$\frac{\ddot{A}}{A}+\frac{\ddot{B}}{B}+\frac{\dot{A}}{A}\frac{\dot{B}}{B}-\frac{\alpha^2}{A^{2}}=-(w+\gamma)\rho,$$ $$\alpha\left(\frac{\dot{A}}{A}-\frac{\dot{B}}{B}\right)=0$$ where the over dot denotes derivation with respect to the cosmic time $t$. General discussion on isotropization and the solution {#sec:2} ===================================================== The anisotropy of the expansion can be parametrized after defining the directional Hubble parameters and the mean Hubble parameter of the expansion. The directional Hubble parameters in the directions of $x$, $y$ and $z$ for the Bianchi type-III metric defined in (1) may be defined as follows, $$H_{x}\equiv\frac{\dot{A}}{A}\textnormal{,}\qquad H_{y}\equiv\frac{\dot{B}}{B}\qquad \textnormal{and}\qquad H_{z}\equiv\frac{\dot{C}}{C}$$ and the mean Hubble parameter is given as $$H=\frac{1}{3}\frac{\dot{V}}{V}=\frac{1}{3}\left(\frac{\dot{A}}{A}+\frac{\dot{B}}{B}+\frac{\dot{C}}{C}\right),$$ where $V=ABC$ is the volume of the universe. The anisotropy parameter of the expansion is defined as $$\Delta\equiv\frac{1}{3}\sum_{i=1}^{3}\left(\frac{H_{i}-H}{H}\right)^{2},$$ where $H_{i}$ (i=1,2,3) represent the directional Hubble parameters in the directions of $x$, $y$ and $z$ respectively. $\Delta=0$ corresponds to isotropic expansion. The space approaches isotropy, in case of diagonal energy-momentum tensor (${T}^{0i}=0$, where $i=1,\,2,\,3$) if $\Delta\rightarrow 0$, $V\rightarrow +\infty$ and ${T}^{00}>0$ ($\rho>0$) as $t\rightarrow+\infty$ (see [@Collins] for details). After giving the above definition (12), if we use only the Einstein field equations (5-9) without introducing any constraint, we can obtain the most general form of the anisotropy parameter of the expansion for Bianchi type-III model in the presence of an anisotropic fluid with a diagonal energy-momentum tensor in general relativity. Solution of the Eq. (9) gives $$B=c_{1}A,$$ where $c_{1}$ is the positive constant of integration. We substitute (13) into (7), subtract the result from (6), and obtain that the skewness parameter on the $y$ axis is null, $$\delta=0,$$ which means that the directional EoS parameters, hence the pressures, on the $x$ and $y$ axes are equal. On the other hand, the deviation of the directional EoS parameter from $w$ on the $z$ axis, $\gamma$, is not constrained to be null by the Einstein field equations. Now, using (10), (11), and (13), anisotropy parameter of the expansion (12) can be reduced to $$\Delta=\frac{2}{9}\frac{1}{H^{2}}\left(H_{x}-H_{z}\right)^{2}.$$ $H_{x}-H_{z}$, the difference between the expansion rates on $x$ and $z$ axes, can be obtained by using the field equations. The field equations are reduced to the equations below, when Eqs. (13) and (14) are substituted into (5-9), $$\frac{{\dot{A}}^{2}}{A^{2}}+2\frac{\dot{A}}{A}\frac{\dot{C}}{C}-\frac{\alpha^2}{A^{2}}=\rho,$$ $$\frac{\ddot{A}}{A}+\frac{\ddot{C}}{C}+\frac{\dot{A}}{A}\frac{\dot{C}}{C}=-w\rho,$$ $$2\frac{\ddot{A}}{A}+{\frac{\dot{A}}{A}}^{2}-\frac{\alpha^2}{A^{2}}=-(w+\gamma)\rho.$$ On solving the equation which is obtained by subtracting (17) from (18) we obtain $$H_{x}-H_{z}=\frac{\dot{A}}{A}-\frac{\dot{C}}{C}=\frac{\lambda}{V}+\frac{1}{V}\int\left(\frac{\alpha^{2}}{A^{2}}-\gamma\rho\right)Vdt,$$ where $\lambda$ is the real constant of integration and the term with $\gamma$ is the term that arises due to the possible intrinsic anisotropy of the fluid. Finally using (19) in (15) we obtain the anisotropy parameter of the expansion, $$\Delta=\frac{2}{9}\frac{1}{H^{2}}\left[\lambda+\int \left(\frac{\alpha^{2}}{A^{2}}-\gamma\rho\right)Vdt\right]^{2}V^{-2}.$$ The anisotropy parameter of the expansion can be reduced to the equation below (21) for a Bianchi type-III cosmological model in the presence of a perfect (thus isotropic) fluid by choosing $\gamma=0$, $$\Delta=\frac{2}{9}\frac{1}{H^{2}}\left[\lambda+\alpha^{2}\int\frac{V}{A^{2}}dt\right]^{2}V^{-2}.$$ The integral term in (20) vanishes for $$\gamma=\frac{\alpha^{2}}{\rho A^{2}},$$ which also leads to the following energy-momentum tensor $$\begin{aligned} {{T}_{\nu}}^{\mu}=\mathrm{diag}\left[1,-w,-w,-w-\frac{\alpha^{2}}{\rho A^{2}}\right]\rho,\end{aligned}$$ and reduces the anisotropy parameter of the expansion to the following form $$\Delta=\frac{2}{9}\frac{\lambda^{2}}{H^{2}}V^{-2}.$$ One can check that this behaviour of the $\Delta$ (24) we obtained by using an *anisotropic fluid* (23) in Bianchi type-III space-time is equivalent to the ones that can be obtained similarly for Bianchi type-I and Bianchi type-V space-times by using any *isotropic fluid*. Then one would see that the results we obtain for $\Delta$ in the models given below are equivalent to the ones obtained in for Bianchi type-I and in for Bianchi type-V space-time models in case of *isotropic fluid*. The vanishing of the integral term also reduces the difference between the expansion rates on $x$ and $z$ to the following form, $$H_{x}-H_{z}=\frac{\lambda}{ABC}.$$ We can also obtain the most general form of the energy density in Bianchi type-III framework by using the first field equation (5) and the definition of the anisotropy parameter of the expansion (12), $$\rho =3H^{2}\left(1-\frac{\Delta}{2}\right)-\frac{\alpha^{2}}{A^{2}}.$$ Below we present the exact solutions of the model in the presence of an anisotropic fluid described by the energy-momentum tensor given in (23), i.e., we use (22) in the Einstein field equations (16-18); $$\frac{{\dot{A}}^{2}}{A^{2}}+2\frac{\dot{A}}{A}\frac{\dot{C}}{C}=\rho+\frac{\alpha^{2}}{A^{2}}=(1+\gamma)\rho,$$ $$\frac{\ddot{A}}{A}+\frac{\ddot{C}}{C}+\frac{\dot{A}}{A}\frac{\dot{C}}{C}=-w\rho,$$ $$2\frac{\ddot{A}}{A}+\frac{{\dot{A}}^2}{A^{2}}=-w\rho.$$ Now we have three linearly independent equations (27-29) and four unknown functions ($A$, $C$, $w$ and $\rho$), thus an extra equation is needed to solve the system completely. To do that we have used two different volumetric expansion laws, $$V=c_{2}e^{3kt}$$ and $$V=c_{2}t^{3m},$$ where $c_{2}$, $k$ and $m$ are positive constants. In this way, all possible expansion histories, the exponential expansion (30) and the power-law expansion (31), have been covered. The models with the exponential expansion and power-law for $m>1$ exhibit accelerating volumetric expansion. On the other hand while model for $m=1$ exhibits volumetric expansion with constant velocity, the models for $m<1$ exhibit decelerating volumetric expansion. Thus, phenomenologically, the anisotropic fluid we dealed here can be considered in the context of DE in the models with exponential expansion and power-law expansion for $m>1$. Model for exponential expansion {#sec:3} =============================== After solving the field equations (27-29) for the exponential volumetric expansion (30) by considering (13) and (25), we obtain the scale factors as follows $$A=\left(\frac{c_{2}}{c_{1}c_{3}}\right)^{\frac{1}{2}}e^{kt-\frac{1}{9}\frac{\lambda}{kc_{2}}e^{-3kt}},$$ $$B=\left(\frac{c_{1}c_{2}}{c_{3}}\right)^{\frac{1}{2}}e^{kt-\frac{1}{9}\frac{\lambda}{kc_{2}}e^{-3kt}},$$ $$C={c_{3}}e^{kt+\frac{2}{9}\frac{\lambda}{kc_{2}}e^{-3kt}},$$ where ${c_{3}}$ is a positive constant of integration. The mean Hubble parameter is, $$H=k$$ and the directional Hubble parameters on the $x$, $y$ and $z$ axes are, respectively, $$H_{x}=H_{y}=k+\frac{1}{3}\frac{\lambda}{c_{2}}e^{-3kt}\qquad\textnormal{and}\qquad H_{z}=k-\frac{2}{3}\frac{\lambda}{c_{2}}e^{-3kt}.$$ Using the directional and mean Hubble parameters in (15) we obtain $$\Delta=\frac{2}{9}\frac{\lambda^2}{{c_{2}}^{2}k^{2}}{e^{-6kt}}.$$ One can check that this behaviour of the $\Delta$ is equivalent to the ones obtained for exponential expansion in Bianchi type-I [@Kumar] and Bianchi type-V cosmological models with isotropic fluid. We can obtain the energy density of the fluid by using the scale factors in (27); $$\rho =3k^{2}-\frac{1}{3}\frac{\lambda^2}{{c_{2}}^2}e^{-6kt}-\alpha^2\frac{{c_{1}}{c_{3}}}{c_{2}}e^{-2kt+\frac{2}{9}\frac{\lambda}{kc_{2}}e^{-3kt}}.$$ The deviation-free part of the anisotropic EoS parameter may be obtained by using (32) and (38) in (29); $$w=\frac{\lambda^2+9{{c_{2}}^2}k^2e^{6kt}}{\lambda^2-9{{c_{2}}^2}k^2e^{6kt}+3\alpha^2{c_{1}}{c_{2}}{c_{3}}e^{4kt+\frac{2}{9}\frac{\lambda}{kc_{2}}e^{-3kt}}}.$$ While the $\delta$ is already found to be zero, we can obtain $\gamma$ by using (32) and (38) in (22) as follows $$\gamma =-\frac{3\alpha^2{c_{1}}{c_{2}}{c_{3}} e^{4kt+\frac{2}{9}\frac{\lambda}{kc_{1}}e^{-3kt}}}{\lambda^2-9{{c_{2}}^2}k^2e^{6kt}+3\alpha^2{c_{1}}{c_{2}} {c_{3}}e^{4kt+\frac{2}{9}\frac{\lambda}{kc_{2}}e^{-3kt}}}.$$ The anisotropy of the expansion ($\Delta$) is not promoted by the anisotropy of the fluid and decreases to null exponentially as $t$ increases. The space approaches to isotropy in this model, since $\Delta\rightarrow 0$, $V\rightarrow\infty$ and $\rho>0$ as $t\rightarrow\infty$. Both terms with $\lambda$ and $\alpha$ contribute to the energy density of the fluid $\rho$ negatively. The energy density ($\rho$), the deviation-free EoS parameter ($w$) and the skewness parameter ($\gamma$) are dynamical. As $t\rightarrow\infty$, the anisotropic fluid isotropizes and mimics the vacuum energy, which is mathematically equivalent to the cosmological constant ($\rm{\Lambda}$), i.e., $\gamma\rightarrow 0$, $w\rightarrow -1$ and $\rho\rightarrow 3k^{2}$. One can observe that the universe approaches to isotropy monotonically even in the presence of the anisotropic fluid, and the anisotropic fluid isotropizes and evolves to the cosmological constant in case of exponential volumetric expansion. These observations are worth to pay attention, since we are inclined to think that anisotropy in an energy source gives rise to increase anisotropy in the expansion. Model for power-law expansion {#sec:4} ============================= After solving the field equations (27-29) for the power-law volumetric expansion (31) by considering (13) and (25), we obtain the scale factors as follows $$A=\left(\frac{c_{2}}{c_{1}c_{3}}\right)^{\frac{1}{2}}t{^m} e^{-\frac{1}{3}\frac{\lambda}{c_{1}}\frac{t^{1-3m}}{3m-1}},$$ $$B=\left(\frac{c_{1}c_{2}}{c_{3}}\right)^{\frac{1}{2}}t^{m} e^{-\frac{1}{3}\frac{\lambda}{c_{1}}\frac{t^{1-3m}}{3m-1}},$$ $$C={c_{3}}t^{m} e^{\frac{2}{3}\frac{\lambda}{c_{1}}\frac{t^{1-3m}}{3m-1}},$$ where ${c_{3}}$ is a positive constant of integration. The mean Hubble parameter is $$H=\frac{m}{t}$$ and the directional Hubble parameters on the $x$, $y$ and $z$ axes are, respectively, $$H_{x}=H_{y}=\frac{m}{t}+\frac{1}{3}\frac{\lambda}{c_{2}}t^{-3m}\qquad\textnormal{and}\qquad H_{z}=\frac{m}{t}-\frac{2}{3}\frac{\lambda}{c_{2}}t^{-3m}.$$ Using the directional and mean Hubble parameters in (15) we obtain $$\Delta=\frac{2}{9}\frac{\lambda^2}{{c_{2}}^2}\frac{t^{2-6m}}{m^2}.$$ One can check that this behaviour of the $\Delta$ is equivalent to the ones obtained for the models that correspond to the power-law expansion in Bianchi type-I [@Kumar] and Bianchi type-V cosmological models with isotropic fluid. We can obtain the energy density of the fluid by using the scale factors in (27); $$\rho =3m^2t^{-2}-\frac{1}{3}\frac{\lambda^2}{{c_{2}}^2}t^{-6m}-\alpha^{2}\frac{c_{1}c_{3}}{c_{2}}t^{-2m}e^{\frac{2}{3}\frac{\lambda}{c_{2}}\frac{t^{1-3m}}{3m-1}}.$$ The deviation-free part of the anisotropic EoS parameter may be obtained by using (41) and (47) in (29); $$w=\frac{\lambda^{2}t^{2}+3m{c_{2}}^2(3m-2)t^{6m}}{\lambda^{2}t^{2}+3\alpha^{2}c_{1}c_{2}c_{3}t^{4m+2}e^{\frac{2}{3}\frac{\lambda}{c_{2}}\frac{t^{1-3m}}{3m-1}}-9m^{2}{c_{2}}^{2}t^{6m}}.$$ While the $\delta$ is already found to be zero, we obtain $\gamma$ by using (41) and (47) in (22) as follows $$\gamma=-\frac{3\alpha^{2}c_{1}c_{2}c_{3}t^{4m+2}e^{\frac{2}{3}\frac{\lambda}{c_{2}}\frac{t^{1-3m}}{3m-1}}}{\lambda^{2}t^{2}+3\alpha^{2}c_{1}c_{2}c_{3}t^{4m+2} e^{\frac{2}{3}\frac{\lambda}{c_{2}}\frac{t^{1-3m}}{3m-1}}-9m^{2}{c_{2}}^{2}t^{6m}}.$$ The volume of the universe expands indefinitely for all values of $m$. Anisotropy of the expansion ($\Delta$) is not promoted by the anisotropy of the fluid. It behaves monotonically, decays to zero for $m>1/3$ and diverges for $m<1/3$ as $t\rightarrow\infty$, and is constant for $m=1/3$. One can see that the terms with $\lambda$ and $\alpha$ contribute the energy density of the fluid ($\rho$) negatively, and thus we can determine which values of $m$ are convenient for which times of the universe by applying the condition $\rho>0$. If $\lambda$ is null, according to the condition on $\rho$, the models for $m<1$ may represent the relatively earlier times of the universe, the models for $m>1$ may represent the relatively later times of the universe and the model for $m=1$ may represent the entire universe provided that $3c_{2}>{\alpha}^{2}c_{1}c_{3}$. The cases for non-zero values of $\lambda$ should be examined seperately. The models for $0<m\leq1/3$ may represent the relatively earlier times of the universe, $1/3<m<1$ may represent the intermediate times of the universe and $m>1$ may represent the relatively later times of the universe. Finally the model for $m=1$ may represent the relatively later times of the universe if $3c_{2}>{\alpha}^{2}c_{1}c_{3}$, otherwise the intermediate times. Thus, we may examine the behaviours of $\Delta$, $w$ and $\gamma$ as $t\rightarrow\infty$ only for $m\geq 1$. For $m>1$, $\Delta\rightarrow 0$ and $V\rightarrow\infty$ as $t\rightarrow\infty$, thus universe approaches to isotropy. $w\rightarrow-1+\frac{2}{3m}$ and $\gamma\rightarrow0$ as $t\rightarrow\infty$, which means that the EoS parameter of the fluid isotropizes and approaches a value in quintessence region with regard the value of $m$ at the later times of the universe for accelerating models. For the model $m=1$, under the condition mentioned in the previous paragraph, $w\rightarrow -(3-{\alpha}^{2}c_{1}c_{3}/c_{2})^{-1}$ and $\gamma\rightarrow ({\alpha}^{2}c_{1}c_{3})/(3c_{2}-{\alpha}^{2}c_{1}c_{3})$ as $t\rightarrow\infty$. Thus, the space approaches to isotropy since $\Delta\rightarrow 0$ and $V\rightarrow\infty$ as $t\rightarrow\infty$, but the fluid does not for the model $m=1$. Similarly to the model with exponential expansion, the universe approaches to isotropy monotonically even in the presence of the anisotropic fluid for $m>1$ and for $m=1$ with appropriate values of the constants. However, the anisotropic fluid isotropizes only in the accelerating models ($m>1$) at later times of the universe and its EoS parameter evolves into the quintessence region. Conclusion ========== We have first obtained the general form of the anisotropy parameter for the expansion of Bianchi type-III metric in the presence of a single imperfect fluid with a dynamically anisotropic equation of state (EoS) parameter and a dynamical energy density in general relativity. Then we have made an assumption on the anisotropy of the fluid in a way to reduce the anisotropy parameter of the expansion to a simple form and obtained a hypothetical fluid with an special anisotropic EoS parameter. The exact solutions of the Einstein field equations have been obtained by assuming two different volumetric expansion laws in a way to cover all possible expansions: namely, exponential expansion and power-law expansion. The anisotropy of the fluid, expansion and space have been examined. It is observed that eventhough the fluid we used wields an anisotropic EoS parameter, its anisotropy does not promote anisotropy in the expansion. The expansion anisotropy decays to zero monotonically in the models with the exponential expansion and in the power-law expansion when $m>1/3$. The universe approaches to isotropy in the accelerating models (exponential expansion and the power-law models with $m>1$). The anisotropy of the fluid isotropizes at later times of the universe in the accelating models. The fluid evolves into the vacuum energy with $w=-1$, which is mathematically equivalent to the cosmological constant ($\rm{\Lambda}$) at the later times of the universe in the model for exponential expansion. The EoS parameter of the fluid evolves into the quintessence region at later times of the accelerating ($m>1$) universes in the power-law models. The anisotropic fluid we used here can be considered in the context of dark energy (DE), at least phenomenologically, in the accelerating models. This model is of interest because it shows that even in the presence of an anisotropic fluid, the universe can approach to isotropy monotonically and also the anisotropy of the fluid can isotropize in the accelerating models. In short, an accelerated expansion period isotropizes both the expansion anisotropy and the anisotropy of the fluid in our study. Thus, even if we observe an isotropic expansion in the present universe we still cannot rule out possibility of DE with an anisotropic EoS. Additionaly we can conclude that an anisotropic DE does not necessarilly distort the symmetry of the space, and consequently even if it turns out that spherical symmetry of the universe that achieved during inflation has not distorted in the later times of the universe, we can not rule out the possibility of an anisotropic nature of the DE at least in Bianchi type-III framework. *Acknowledgments* Özgür Akarsu was supported in part by The Scientific and Technological Research Council of Turkey (TÜB[İ]{}TAK). Smoot, G.F., et al.: Astrophys. J. **396**, L1 (1992) Bennett, C.L., et al.: Astrophys. J. **464**, L1-L4 (1996) Hinshaw, G., et al.: Astrophys. J. Suppl. **148**, 135 (2003) Hinshaw, G., et al.: Astrophys. J. Suppl. **170**, 288 (2007) Hinshaw, G., et al.: Astrophys. J. Suppl. **180**, 225-245 (2009) Eriksen, H.K., et al.: Astrophys. J. **605**, 1420 (2004) de Oliveira-Costa, A., et al.: Phys. Rev. D **69**, 063516 (2004) Cruz, M., et al.: Mon. Not. R. Astron. Soc. **356**, 29 (2004) Hansen, F. K., et al.: Mon. Not. R. Astron. Soc. **354**, 641 (2004) Land, K., Magueijo, J.: Phys. Rev. Lett. **95**, 071301 (2005) Copi, C.J., et al.: Mon. Not. R. Astron. Soc. **367**, 79-102 (2006) Copi, C.J., et al.: arXiv:astro-ph/0808.3767v1 Jaffe, T.R. et al.: Astrophys. J. **629**, L1-L4 (2005) Jaffe, T.R. et al.: Astrophys. J. **643**, 616-629 (2006) Jaffe, T.R., et al.: Astron. Astrophys. **460**, 393 (2006) Campanelli, L. et al.: Phys. Rev. Lett. **97**, 131302 (2006) Campanelli, L. et al.: Phys. Rev. D **76**, 063007 (2007) Ellis, G.F.R.: Gen. Relativ. Gravit. **38(6)**, 1003-1015 (2006) Ellis, G.F.R.: Cosmological Models. In: Bonometto, S., et al. (ed.) Modern Cosmology, pp. 108-158. Institute of Physics Publishing, Bristol and Philadelphia (2002) Koivisto T., Mota, D.F.: Phys. Rev. D **73**, 083502 (2006) Koivisto T., Mota, D.F.: arXiv:astro-ph/0707.0279 (2007) Rodrigues, D.C.: Phys. Rev. D **77**, 023534 (2008) Guth, A.H.: Phys. Rev. D **23**, 347 (1981) Sato, K.: Mon. Not. R. Astron. Soc. **195**, 467 (1981) Linde, A.D.: Phys. Lett. B **108**, 389 (1982) Albrecht, A., Steinhardt, P.J.: Phys. Rev. Lett. **48**, 1220 (1982) Feinstein, A., Ibanez, J.: Class. Quant. Gravit. **10**, 93-98 (1993) Aguirregabiria, J.M., et al.: Phys. Rev. D **48**, 4662-4668 (1993) Ford, L.H.: Phys. Rev. D **40**, 967 (1989) Koivisto T., Mota, D.F.: J. Cosmol. Astropart. Phys. **08**, 021 (2008) Golovnev, A., et al.: J. Cosmol. Astropart. Phys. **06**, 009 (2008) Sahni, V., Starobinsky A.A.: Int. J. Mod. Phys. D **9**, 373-444 (2000) Sahni, V.: Lect. Notes Phys. **653**, 141-180 (2004) Alam, U., et al.: J. Cosmol. Astropart. Phys. **0406**, 008 (2004) Sahni, V., Starobinsky A.A.: Int. J. Mod. Phys. D **15**, 2105-2132 (2006) Copeland, E.J., et al.: Int. J. Mod. Phys. D **15**, 1753-1936 (2006) Padmanabhan, T.: Gen. Relativ. Gravit. **40**, 529-564 (2008) Turner M.S., Huterer D.: J. Phys. Soc. Jap. **76**, 111015 (2007) Carroll, S.M., Hoffman, M.: Phys. Rev. D **68**, 023509 (2003) Armendariz-Picon, C.: J. Cosmol. Astropart. Phys. **07**, 007 (2004) Kiselev, V.V.: Class. Quant. Gravit. **21**, 3323 (2004) Zimdahl, W., et al.: Phys. Rev. D **64**, 063501 (2001) Novello, M., et al.: Phys.Rev. D **69**, 127301 (2004) Wei, H., Cai, R.G.: Phys. Rev. D **73**, 083002 (2006) Tegmark, M., et al.: (SDSS Collaboration). Phys. Rev. D **69**, 103501 (2004) Riess, A.G., et al.: Astrophys. J. **607**, 665-687 (2004) Astier, P., et al.: Astron. Astrophys. **447**, 31-48 (2006) Mota, D.F., et al.: Mon. Not. R. Astron. Soc. **382**, 793-800 (2007) Koivisto T., Mota, D.F.: J. Cosmol. Astropart. Phys. **018** (2008) Akarsu, O., Kilinc, C.B.: Gen. Relativ. Gravit. (2009). doi:10.1007/s10714-009-0821-y Moussiaux, A. et al.: J. Phys. A **14**, L277-L280 (1981) Lorenz, D.: J. Phys. A **15**, 2997 - 2999 (1982) Chakraborty, N.C., Chakraborty, S.: Il Nuovo Cimento B **116**, 191-198 (2001) Singh, J.P., et al.: Chin. Phys. Lett. **24**, 3325 (2007) Tiwari, R.K.: Astrophys. Space Sci. **319**, 85-87 (2009) Bali, R., Tinker, S.: Chin. Phys. Lett. **26**, 029802 (2009) Letelier, P.S.: Phys. Rev. D **22**, 807 (1980) Collins, C.B., Hawking, S.W.: Astrophys. J. **180**, 317-334 (1973) Kumar, S., Singh, C.P.: Astrophys. Space. Sci. **312**, 57-62 (2007) Singh, C.P., et al.: Astrophys. Space. Sci. **315**, 181-189 (2008) Singh, J.P., Baghel, P.S.: Int. J. Theor. Phys. **48**, 449-462 (2009) [^1]: E-Mail: [email protected] [^2]: E-Mail: [email protected]
{ "pile_set_name": "ArXiv" }
--- abstract: 'Vascular graphs can embed a number of high-level features, from morphological parameters, to functional biomarkers, and represent an invaluable tool for longitudinal and cross-sectional clinical inference. This, however, is only feasible when graphs are co-registered together, allowing coherent multiple comparisons. The robust registration of vascular topologies stands therefore as key enabling technology for group-wise analyses. In this work, we present an end-to-end vascular graph registration approach, that aligns networks with non-linear geometries and topological deformations, by introducing a novel over-connected geodesic vascular graph formulation, and without enforcing any anatomical prior constraint. The 3D elastic graph registration is then performed with state-of-the-art graph matching methods used in computer vision. Promising results of vascular matching are found using graphs from synthetic and real angiographies. Observations and future designs are discussed towards potential clinical applications.' author: - 'Stefano Moriconi^()^' - 'Maria A. Zuluaga' - 'H. Rolf J[ä]{}ger' - Parashkev Nachev - Sébastien Ourselin - 'M. Jorge Cardoso' bibliography: - './Moriconi2018Elastic\_MICCAI18.bib' title: Elastic Registration of Geodesic Vascular Graphs --- Introduction ============ Vascular graphs can be obtained from angiographies using connectivity paradigms and network extraction algorithms by embedding high-level features, such as spatial location, direction, scale, and bifurcations. However, the correct extraction of subject-specific vascular topologies, in complex (cerebro)vascular networks, can be challenging when rather tortuous and tangled structures are present. In other cases, anatomical cycles and their variants (i.e. the circle of Willis, anastomoses and fenestrations) [@jinkins2000atlas], the presence of pathology (e.g. tangled arterio-venous malformations, neoplastic and embryologic plexiforms), and image-related limitations (e.g. unresolved kissing vessels) dramatically increase the network complexity, and sometimes impede the extraction of the vascular topology as a tree. A viable approach is to consider a data-driven vectorial prior from an early group-wise vascular graph registration. Defining a group-wise vectorial prior first embeds the likelihood of connectivity patterns from a population, and subsequently injects a probabilistic prior towards the inference of the most meaningful subject-specific vascular topology. The same vectorial prior could also embed morphometric parameters, functional and hemodynamic descriptors and surrogate biomarkers, constituting thus a labelled multi-spectral vascular atlas. By registering the obtained vectorial atlas over a set of similar vascular graphs, a number of group-level clinical analyses would be allowed, from inter-subject comparisons of the underlying vascular morphology, to longitudinal studies of vascular pathologies, on which clinical prediction and therapeutic inference ultimately depend. The robust alignment of multiple topologies is of critical relevance and represents a methodological bottleneck for population-level analyses. The alignment of networks and vectorial graphs raised increasing interest among the scientific community in the last decade. Motivated by registering acyclically connected structures from biomedical imaging, (e.g. vascular and respiratory trees), [@feragen2012hierarchical; @charnoz2005design; @xue2006automatic; @wang2017automatic; @petersen2013airways; @serradell2015nonrigid] introduced different registration techniques, which mostly rely on pairwise matching distances between junction nodes and connecting edges. Following an initial alignment, these methods usually minimise a similarity cost function or maximise a probabilistic likelihood between pairs of nodes/edges or sub-trees and graph kernels, and hierarchically evaluate the correspondences at different levels of tree-depth. Whilst only few formulations would register generic spatial graphs [@serradell2015nonrigid], in all cases the considered topologies were either hierarchically pre-defined as trees, or determined beforehand on a specific anatomical compartment. Also, since these methods exploit node locations, branches geometry, arborescence depth, or the parent-child relation of a rooted tree, they require the explicit tree topology to accurately capture the underlying vasculature, where each bifurcation is correctly annotated as its connectivity pattern. The registration of noisy topologies, (i.e. mis-connections, missing branches and short-cuts), and non-linearly deformed geometries remains a challenging and open problem. \[Fig01\] ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![Geodesic Vascular Graph and GM problem of non-linearly deformed topologies. Extraction of a fully-connected topology from an initial set of nodes (left). Associated graph representations and minimum spanning trees for two topologically different instances ($\mathcal{G}^{A}$ and $\mathcal{G}^{B}$) of the same underlying vascular anatomy (center). Graphs alignment and nodes matching for the generalised GM problem (right).](Fig1_1 "fig:"){width="16.70000%"} ![Geodesic Vascular Graph and GM problem of non-linearly deformed topologies. Extraction of a fully-connected topology from an initial set of nodes (left). Associated graph representations and minimum spanning trees for two topologically different instances ($\mathcal{G}^{A}$ and $\mathcal{G}^{B}$) of the same underlying vascular anatomy (center). Graphs alignment and nodes matching for the generalised GM problem (right).](Fig1_2 "fig:"){width="16.70000%"} ![Geodesic Vascular Graph and GM problem of non-linearly deformed topologies. Extraction of a fully-connected topology from an initial set of nodes (left). Associated graph representations and minimum spanning trees for two topologically different instances ($\mathcal{G}^{A}$ and $\mathcal{G}^{B}$) of the same underlying vascular anatomy (center). Graphs alignment and nodes matching for the generalised GM problem (right).](Fig1_3 "fig:"){width="10.50000%"} ![Geodesic Vascular Graph and GM problem of non-linearly deformed topologies. Extraction of a fully-connected topology from an initial set of nodes (left). Associated graph representations and minimum spanning trees for two topologically different instances ($\mathcal{G}^{A}$ and $\mathcal{G}^{B}$) of the same underlying vascular anatomy (center). Graphs alignment and nodes matching for the generalised GM problem (right).](Fig1_4 "fig:"){width="10.50000%"} ![Geodesic Vascular Graph and GM problem of non-linearly deformed topologies. Extraction of a fully-connected topology from an initial set of nodes (left). Associated graph representations and minimum spanning trees for two topologically different instances ($\mathcal{G}^{A}$ and $\mathcal{G}^{B}$) of the same underlying vascular anatomy (center). Graphs alignment and nodes matching for the generalised GM problem (right).](Fig1_5 "fig:"){width="37.50000%"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -6pt In this preliminary work, we address vascular graph matching (GM) by relaxing assumptions on the acyclic (un)directed graph structure and the anatomical hierarchical prior from any vascular compartment. The idea is to consider and register the vasculature as an over-connected graph: a redundant topology encoding the likelihood of connections between neighbouring nodes with minimal paths. This enhanced connectivity pattern would compensate for topological inaccuracies, for non-linear deformations of branches, and would enrich the registration space-search with distinctive features. The pairwise graph registration problem can be subsequently solved using generic GM algorithms. In the following sections, the proposed approach is first described, then, an experimental set-up is presented, comprising graphs from synthetic and real angiographies. The accuracy of different GM algorithms is evaluated on correct nodes correspondences. Observations and conclusions are discussed, focusing on future developments and potential applications. Methods {#section_Methods} ======= Aiming at the pairwise alignment of vascular topologies within a deformable and anatomical prior-free framework, we first introduce a novel over-connected geodesic vascular graph (GVG), then the generic GM problem is presented together with the proposed affinity metrics based on vessels geometry and their redundant geodesic connectivity. The two-steps registration pipeline is described also listing the considered GM algorithms. ### Geodesic Vascular Graph. {#subsection_GVG} We define the undirected geodesic vascular graph $\mathcal{G} = (N,E)$ in $\mathbb{R}^3$, as the set of nodes $\mathbf{n}_i \in N$, and the associated set of connecting geodesic edges $\mathbf{e}_v \in E$, encoding the graph adjacency list. Each geodesic edge $\mathbf{e}_v$ is defined as the 3D shortest path joining a generic pair of nodes, by solving the Eikonal equation [@kimmel1998computing] over a vascular smoothly connected manifold as in [@moriconi2017vtrails]. However, an exhaustive search is here performed by connecting all pairs of nodes independently, or up to a pre-defined spatial neighborhood $\nu$. This determines an over-connected vascular graph of minimal paths, which fully captures the underlying vasculature with enhanced geodesic redundancy . Together with the formulation of the over-connected $\mathcal{G}$, we also introduce a set of edge- and node-attributes. The edge-attributes $\mathbf{e}_v=\{\mathbf{p}_v , l_v , u_v\}$ comprise the dense sampling $\mathbf{p}_v$ of each shortest path in 3D (i.e. the point coordinates sequence as in Fig. 2,3), its associated euclidean length $l_v$ and the geodesic integral energy $u_v$ integrated along the path, as in [@moriconi2017vtrails]. The node-attributes $\mathbf{n}_i = \{\mathbf{c}_i , d_i\}$ include the spatial location $\mathbf{c}_i$ as coordinates in $\mathbb{R}^3$, and the geodesic node degree $\mathsmaller{d_i = \frac{1}{| \tilde{\mathbf{e}}_v |} \sum_{\tilde{\mathbf{e}}_v}{u_v} }$, with $\tilde{\mathbf{e}}_v$ the set of incident edges of cardinality $| \tilde{\mathbf{e}}_v |$. ### Graph Matching Problem and Affinity Metrics. {#subsection_GMp_AffMetrics} As presented in [@zhou2016factorized], the problem of matching a pair of graphs $\mathcal{G}^{A}$ and $\mathcal{G}^{B}$ requires the definition of an affinity matrix $\mathbf{K}$ to measure the similarity between each pair of nodes and edges. Given the node cardinality $i = |\mathbf{n}^{A}_{i}|$, and $j = |\mathbf{n}^{B}_{j}|$, the symmetric affinity matrix $\mathbf{K} \in \mathbb{R}^{ij \times ij}$ encodes the similarity between nodes along its diagonal elements, whereas the edges similarity is encoded in the off-diagonal ones. Given $\mathbf{K}$, the problem of graph matching consists in finding the optimal correspondence $\mathbf{X}$ between all the nodes , so that a compatibility functional $J(\mathbf{X})$ is maximised with a quadratic assignment problem (QAP) [@loiola2007survey], $$\label{eq1} \max{~J(\mathbf{X})} = \text{vec}(\mathbf{X})^{t} ~ \mathbf{K} ~ \text{vec}(\mathbf{X}), \vspace{-2pt}$$ where $\mathbf{X}$ is constrained to be a one-to-one mapping between the sets of nodes $\mathbf{n}^{A}_{i}$ and $\mathbf{n}^{B}_{i}$, and $\text{vec}(\mathbf{X})$ denotes the vectorisation of the correspondence matrix. We formulate both node- and edge-similarity metrics for the definition of the affinity matrix $\mathbf{K}$, by adopting the matrix factorisation as in [@zhou2016factorized], i.e. $\mathbf{K}_{\mathbf{n}^{\text{\textit{AB}}}}$ and $\mathbf{K}_{\mathbf{e}^{\text{\textit{AB}}}}$ respectively. In detail, we define $$\begin{gathered} \label{eq2-3} \vspace{-6pt} \mathbf{K}_{\mathbf{n}^{\text{\textit{AB}}}} = e^{-\left( \alpha_{1} \frac{\mathbf{C}^{\text{\textit{AB}}}}{\sigma_{\mathbf{C}}} ~+~ \alpha_{2} \frac{\mathbf{D}^{\text{\textit{AB}}}}{\sigma_{\mathbf{D}}} \right)} ~~~~ \text{with} ~~ \alpha_{1} + \alpha_{2} = 1,~~~ \text{and}\\ \mathbf{K}_{\mathbf{e}^{\text{\textit{AB}}}} = e^{-\left( \beta_{1} \frac{\mathbf{P}^{\text{\textit{AB}}}}{\sigma_{\mathbf{P}}} ~+~ \beta_{2} \frac{\mathbf{L}^{\text{\textit{AB}}}}{\sigma_{\mathbf{L}}} ~+~ \beta_{3} \frac{\mathbf{U}^{\text{\textit{AB}}}}{\sigma_{\mathbf{U}}} \right)} ~~~~ \text{with} ~~ \beta_{1} + \beta_{2} + \beta_{3} = 1, \vspace{-6pt}\end{gathered}$$ where $\mathbf{C}^{\text{\textit{AB}}}$ and $\mathbf{D}^{\text{\textit{AB}}}$ are the pairwise $\ell^{2}$-norm matrices between the two sets of node coordinates $\{\mathbf{c}^{A}_i,\mathbf{c}^{B}_j\}$, and geodesic degrees $\{d^{A}_i,d^{B}_j\}$, as well as $\mathbf{P}^{\text{\textit{AB}}}$, $\mathbf{L}^{\text{\textit{AB}}}$ and $\mathbf{U}^{\text{\textit{AB}}}$ are the pairwise average symmetric distance matrices of the connecting minimal paths $\{\mathbf{p}^{A}_v,\mathbf{p}^{B}_w\}$, and the pairwise $\ell^{2}$-norm matrices between the sets of the euclidean lengths $\{l^{A}_v,l^{B}_w\}$ and geodesic integral energies $\{u^{A}_v,u^{B}_w\}$, respectively. The normalisation factors $\sigma_{\mathbf{C},\mathbf{D},\mathbf{P},\mathbf{L},\mathbf{U}}$ are the standard deviations estimated from the off-diagonal elements of the associated distance matrices over the considered population of graphs. Lastly, $\alpha_{1}$, $\beta_{1}$ and $\beta_{2}$ weight the geometrical similarities among nodes and edges, whereas $\alpha_{2}$ and $\beta_{3}$ represent the respective geodesic trade-off. We refer to [@zhou2016factorized] for the composition of $\mathbf{K}$ from the factorised components $\mathbf{K}_{\mathbf{n}^{\text{\textit{AB}}}}$ and $\mathbf{K}_{\mathbf{e}^{\text{\textit{AB}}}}$, and for the QAP solver implementation. ### Graph Registration. {#subsection_GraphRegPipeline} Although some GM algorithms do not require any spatial initialisation of the graphs, we present a two-steps approach by combining an early coarse alignment strategy to facilitate the further registration by reducing biases due to pure rigid mis-alignment. #### Rigid Alignment. The globally-optimal iterative closest point [@yang2016go] is run on $\mathcal{G}^{A}$ and $\mathcal{G}^{B}$ as coarse geometrical initialisation. Here, the dense cloud of samples, i.e. the nodes coordinates $\{\mathbf{c}^{A}_{i},\mathbf{c}^{B}_{j}\}$ and the sequences of edge points $\{\mathbf{p}^{A}_{v},\mathbf{p}^{B}_{w}\}$, is retrieved for the spatial rigid pre-alignment. searches the entire 3D motion space, and, under the minimisation of an $L_{2}$ error metric based on a branch-and-bound scheme, guarantees the global optimality of the rigid mapping, even in presence of noisy data, outliers, and partial samples overlap. #### Fine Graph Matching. Classic GM algorithms employed in computer vision, are considered for the fine registration. We account for Graduated Assignment (GA) [@gold1996graduated], Spectral Matching (SM) [@leordeanu2005spectral], Spectral Matching with Affine Constraints (SMAC) [@cour2007balanced], Probabilistic Matching (PM) [@zass2008probabilistic], Integer Projected Fixed Point (IPFP-U/SM) [@leordeanu2009integer], Re-weighted Random Walk Matching (RRWM) [@cho2010reweighted], and the current state-of-the-art, the non-rigid Factorized Graph Matching (FGM) [@zhou2016factorized]. The deformable graph matching problem, detailed in [@zhou2016factorized], formulates the unknown graph correspondence being constrained with a geometric transformation $T$. A composition of transformations (i.e. similar, affine, and non-rigid) are incorporated into the compatibility function (\[eq1\]), and subsequently estimated by optimising jointly the correspondence matrix $\mathbf{X}$ and the composite transformation $T$ itself. We employed the undirected-graph versions of the listed algorithms. Implementations and configurations are available from authors’ websites. Experiments and Results {#section_ExpRes} ======================= \[Fig02\] ------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------- -- -- $\mathbf{\mathcal{G}^{A}}$ $\mathbf{\mathcal{G}^{A}_\mathsmaller{{\mathcal{D}_{40\%}\mathcal{T}_{30\%}}}}$ **Rigid Alignment** ![Example of sGVG, simulated deformations, rigid alignment and resulting GM.](Fig2_1a "fig:"){width="19.50000%"} ![Example of sGVG, simulated deformations, rigid alignment and resulting GM.](Fig2_1b "fig:"){width="19.50000%"} ![Example of sGVG, simulated deformations, rigid alignment and resulting GM.](Fig2_2_sq "fig:"){width="19.50000%"} ------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------- -- -- -6pt #### Dataset. A set of 10 synthetic over-connected geodesic vascular graphs (sGVG) and associated minimum spanning trees (sGVT) are obtained from 3D vascular tree images [@vascusynth] (isotropic 100$\times$100$\times$100 voxels), as in \[subsection\_GVG\]. Each graph comprises 80 nodes, i.e. the vascular junction and end-points, over-connected within a neighbourhood of radius $\nu = 35$ . A total of 10 fully over-connected geodesic vascular graphs (aGVG) as well as the respective minimum spanning trees (aGVT) of the basilar artery are derived as in \[subsection\_GVG\] from Time-of-Flight MRI angiographies (0.35$\times$0.35$\times$0.5 mm), where anatomical vascular junctions and endpoints were manually labelled following [@jinkins2000atlas]. #### Synthetic Graphs. We randomly deform the synthetic datasets sGVG and sGVT with a non-linear geometrical displacement field (i.e. max magnitude $\mathcal{D}_{30\%}$, $\mathcal{D}_{40\%}$, $\mathcal{D}_{50\%}$ of the graph spatial embedding), a topological pruning (i.e. reducing by $\mathcal{T}_{30\%}$, $\mathcal{T}_{40\%}$, $\mathcal{T}_{50\%}$ the original connectivity), and a combination of both, for a representative set of alterations (Fig. 2). The deformed graphs were then registered with the respective unaltered topologies. The accuracy of the GM is given by the percentage of correct correspondences, and differences of registration performances between sGVG and sGVT are evaluated with a paired Wilcoxon signed rank test. #### Angiographic Graphs. Both aGVG and aGVT are pairwise aligned, covering all possible inter-subject combinations within the same dataset. The matching accuracy is given by the percentage of correct correspondence among the labelled nodes. Differences between aGVG and aGVT are evaluated with a paired Wilcoxon signed rank test. ### Synthetic Graph Matching. {#Res_SynthGM} In Fig. 4 (charts), the GM accuracy is reported for the synthetic datasets, for each algorithm and for the simulated levels of deformation. The affinity metrics trade-offs are arbitrarily defined as $\mathbf{\alpha} = \left[0.5,0.5\right]$, and $\mathbf{\beta} = \left[0.25,0.25,0.5\right]$ in all cases, to balance the similarity features. Similar trends of performances are observed for the considered GM algorithms across different levels of increasing deformation. Overall, FGM reported the best matching accuracy together with RRWM in both sGVG and sGVT, whereas the other algorithms showed globally varying performances. Purely geometrical displacements did not affect the registration, whereas more severe topological pruning showed a visible drop of accuracy in both sGVG and sGVT, as well as the combination of joint deformations at different degrees. Overall, better matching is found for sGVG compared to sGVT at the same level of alteration. A significant accuracy drop ($p<0.05$) is found for the registration of tree-like structures, proportional to the combined deformation. This suggests that the proposed registration pipeline would benefit from both geometrical and geodesic information arising from a more dense and redundant over-connected pattern, rather than an explicit vascular tree hierarchy, in presence of non-linear deformations. ### Angiographic Graph Matching. The accuracy of the pairwise registration for both aGVG and aGVT datasets is reported in Fig. 4 (table). The affinity metrics trade-offs adopted here are the same as those for the synthetic experiments. Overall, discrete matching is obtained for the state-of-the-art FGM ($\text{61.26}\,\pm\,\text{21.91} \%$), as well as for GA ($\text{65.16}\,\pm\,\text{20.39}\%$) and SM ($\text{62.83}\,\pm\,\text{22.96}\%$). The considered angiographic dataset presented large deformations and anatomically different variants (Fig. 3). In line with results of \[Res\_SynthGM\], the registration of over-connected topologies (aGVG) showed significantly higher accuracy ($p<0.05$), compared to the respective hierarchical minimum spanning trees (aGVT). Globally, nodes mismatch occurred in correspondence of nodes with lower degree and centrality, where higher confusion is found for spatially close vascular end-points and neighbouring branches. Conversely, the correspondence of superior/inferior and left/right branches was correctly preserved in the majority of cases. \[Fig03\] ----------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------- **aGVG $\&$ Label set** **Rigid Alignment** **Fine GM (FGM) - acc: 84.21$\%$** ![aGVG label set [@jinkins2000atlas] and pairwise registration of anatomical topologies.](Fig3_1 "fig:"){width="24.90000%"} ![aGVG label set [@jinkins2000atlas] and pairwise registration of anatomical topologies.](Fig3_2 "fig:"){width="26.50000%"} ![aGVG label set [@jinkins2000atlas] and pairwise registration of anatomical topologies.](Fig3_3 "fig:"){width="43.00000%"} ----------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------- -6pt \[Fig04\] -- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ![Accuracy of GM: synthetic datasets sGVG vs. sGVT (charts), and angiographic datasets aGVG vs. aGVT (table). Values are mean $\pm$ SD, (median), $~\mathbf{\ast} = p<0.05$.](Fig4_1 "fig:"){width="57.00000%"} -- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -6pt Discussion and Conclusions {#section_DiscConc} ========================== We presented a vascular graph matching approach to pairwise and elastically register similar topologies, in presence of non-linear deformations. A novel formulation of the vascular network is first introduced using an over-connected geodesic vascular graph. Then, the non-rigid nodes correspondence assignment is solved with a two-steps alignment comprising an optimal rigid registration of the network geometrical embedding, and a set of graph matching algorithms employed in computer vision. For the first time, a general registration of vascular graphs, accounting for noisy over-connected topologies with possible cycles, could be performed by relaxing the explicit hierarchical vessel-tree structure or connectivity patterns specific of a vascular compartment. The use of multiple GM strategies, on the one hand, is motivated by the unconstrained formulation of the GVG, on the other hand, it is justified by the different connectivity lattice of the introduced GVG. The latter can dramatically differ from the connectivity patterns found in computer vision applications (i.e. 3D polygonal subdivision and/or triangulations in 2D), therefore, established GM algorithms may show rather different performances. Early results show, however, good matching from synthetic vascular graphs even in presence of mild-to-moderate non-linear deformations. With the same registration pipeline, we aligned over-connected and redundant topologies, as well as hierarchical undirected tree-structures. Despite these share the same similarity features, the graph matching reported significantly different accuracies, where better nodes correspondences are found for the over-connected topologies. This suggests that the overhead information from the redundant connectivity may enrich the registration space-search with more distinctive cues. Similarly, the registration of geodesic vascular graphs from angiographic datasets reported appreciable matching, even in cases of large spatial deformations and anatomically different topologies, whereas the registration of the associated tree-like structures showed significantly lower accuracies, in line with the synthetic experiments. On the basis of this early evidence, we assume the problem of vascular tree- and graph-registration could be generalised with a multi-spectral network alignment, where further developments towards a more robust design for vascular applications may better incorporate both geometrical and geodesic vascular features. Although most of the GM algorithms considered in this work are used for 2D applications in computer vision, their general formulation allows the alignment of any generic network, regardless the dimensional embedding, and offer a rich ground for ad-hoc methodological developments. From a clinical perspective, the successful vascular graph alignment would lead to the definition of a co-registered group-wise prior to improve the inference of patient specific anatomical topologies. In last instance, the co-registration of a vascular vectorial prior would pave the way for group-wise analyses with potential applications in neurovascular cross-sectional and longitudinal studies. ### Acknowledgements: {#acknowledgements .unnumbered} The study is co-funded from the Wellcome Trust, the EPSRC grant EP/H046410/1, and the National Institute for Health Research, University College London Hospitals, Biomedical Research Centre.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We analyse strangeness production in proton-proton (pp) collisions at SPS and RHIC energies, using the recently advanced NeXus approach. After having verified that the model reproduces well the existing data, we interpret the results: strangeness is suppressed in proton-proton collisions at SPS energy as compared to electron-positron ($ \mathrm{e}^{+}\mathrm{e}^{-} $) annihilation due to the limited masses of the strings produced in the reaction, whereas high energy pp and $ \mathrm{e}^{+}\mathrm{e}^{-} $ collisions agree quantitatively. Thus strangeness suppression at SPS energies is a consequence of the limited phase-space available in string fragmentation.' author: - | Hans-Joachim Drescher$ ^{\$}\protect $, Jörg Aichelin$ ^{\dagger }\protect $, Klaus Werner$ ^{\dagger }\protect $\ $ ^{\$}\protect $New York University, 4 Washington Place, New York NY 10003\ $ ^{\dagger }\protect $SUBATECH, La Chantrerie 4, rue Alfred Kastler BP 20722 - 44307 Nantes-cedex 3 title: 'Strangeness Suppression in Proton-Proton Collisions' --- Introduction ============ Strangeness enhancement in ultra-relativistic nucleus-nucleus collisions has been proposed as a signal for the formation of a quark-gluon plasma[@rafelski]. At the SPS it has been observed for example by the WA97 [@WA97] collaboration, and the experiments of RHIC will certainly analyse strangeness is detail. When one talks about strangeness enhancement one has first to specify the point of reference which are usually proton-proton reactions. How meaningful is this point of reference? In order to answer this question, we analyse the strangeness production of proton-proton collisions at different energies as compared to $ \mathrm{e}^{+}\mathrm{e}^{-} $annihilation which is rather energy independent. This choice is justified by the fact that particle production seems to be universal in all kinds of elementary high energy reactions. The [<span style="font-variant:small-caps;">neXus</span>]{} Model and the Role of String Fragmentation ====================================================================================================== Before we start with an analysis of the physics of multiplicities of different hadrons we explain the basic features of our approach ([<span style="font-variant:small-caps;">neXus</span>]{}) which describes simultaneously high energy electron-positron annihilation and hadron-hadron scattering. The details may be found in reference [@nexus]. The common feature between hadron-hadron collisions and electron-positron annihilation is the creation of strings which finally produce observable hadrons. In the former case the exchange of a Pomeron leads to the formation of two strings, in the latter a string is spanned between the quark-antiquark created by the decay of a virtual photon or a Z-boson. At low energies the string just consists of two partons at the end-points, at higher energies perturbative gluons appear in initial or final state radiation which are mapped onto the string as the so-called kinks. Once a string is created it evolves according to the Nambu-Goto Lagrangian for a classical relativistic string. In order to produce hadrons we use the area-law of Artru-Menessier. Here, the probability of the string to break is proportional to the invariant area swept over in Minkowski space. The breaking is then determined by one parameter, the break probability. If it is small, the string breaks at later times, producing less but heavier fragments and vice-verse. Flavor production is governed by two additional parameters, the probability to create a strange quark-antiquark pair (otherwise up or down pairs are created in equal fractions) and the probability to create a diquark-antidiquark pair. The former therefore governs strangeness production, the latter baryon production and the combination of both rules the creation of hyperons. The decay of strings can be seen as a longitudinal (one dimensional) microscopic phase space decay, therefore is is more difficult to produce heavier particles than lighter ones. In figure \[fig:1a\] we show particle yields of $ \mathrm{e}^{+}\mathrm{e}^{-} $ annihilation from our model compared with data from the Opal collaboration [@opal]. The two parameters for strangeness and diquarks have been adjusted to fit these data, and the model is capable to describe a multitude of data. One can convince oneself in reference [@nexus] that also event-shapes and differential spectra are reproduced nicely. The same model applied to hadron-hadron collisions gives the results shown in figure \[fig:1bc\]. Here we compare two energies which are close to the ones we are going to use for our analysis. Furthermore we consider only negatives or anti-baryons as produced particles, results for the other particles agree in a similar way with data. We can conclude that for $ e^{+}e^{-} $annihilation as well as for pp and anti-proton-proton ($ \mathrm{p}\bar{\mathrm{p}} $) collisions NeXus agrees with the experimentally observed particle yields. Interpretation of Results ========================= We are now going to interpret the above-mentioned results based on [<span style="font-variant:small-caps;">neXu</span>]{}s calculations. In figure \[fig:2\] we show multiplicities of particles produced in pp collisions at 17.3 GeV (SPS) and at 200 GeV (RHIC) as compared with $ \mathrm{e}^{+}\mathrm{e}^{-} $ at 91.2 GeV. To account for spin-degeneracy we divide the obtained multiplicity by the factor $ 2j+1 $. First of all one sees that the particle yields fall roughly exponentially with the particle mass. This is a simple phase-space effect: heavier particles are more difficult to produce. Striking is the unexpected similarity of pp at 200 GeV with $ \mathrm{e}^{+}\mathrm{e}^{-} $ at 91.2 GeV. As described above, the formation of strings is quite different in pp as compared to $ \mathrm{e}^{+}\mathrm{e}^{-} $ reactions. The spectra obtained for 17.3 GeV is considerably steeper. We see as well very little difference between strange and non strange hadrons, all fall on a common curve. This effect can be seen clearer in figure \[fig:3\], where the multiplicities are plotted normalized to the ones of $ \mathrm{e}^{+}\mathrm{e}^{-} $. The ratio for pp interactions with respect to $ \mathrm{e}^{+}\mathrm{e}^{-} $ at RHIC energies is close to one. Only the heaviest particle - the Omega - is slightly suppressed in pp. At SPS energies the yields for pp collisions show a completely different behavior: the ratio with respect to $ \mathrm{e}^{+}\mathrm{e}^{-} $ falls off strongly as a function of the mass. Where does this effect come from? No new physics enters between the two energies, with exception of the minijets which are more abundant at higher energies. But this influences only differential spectra like that of transverse momenta and not the relative abundance of particles. The answer becomes quite clear when we look at the masses of the strings which finally produce the particles. Figure \[fig:4\] shows the distribution $ \frac{dn}{dm} $ of string masses produced at the two different energies. We leave out the case of $ \mathrm{e}^{+}\mathrm{e}^{-} $ since here we have in most cases one string of mass 91.2 GeV. Only if a quark-anti-quark pair is produced during the final state radiation, we end up with more than one string. This process is however much less important than gluon radiation. In pp interactions most of the strings have still low masses, which is a direct consequence of parton distribution functions peaking at low x. But the evolution of the tails is quite different. Whereas at 17.3 GeV the distribution is steeply falling with almost no strings at all above 10 GeV, the strings for pp at 200 GeV reach much higher masses. More conclusive is figure \[fig:5\] where we see the corresponding cumulative distributions, i.e. the fraction $$F(m)=\frac{\int _{0}^{m}\frac{dn}{dm'}dm'}{\int _{0}^{\infty }\frac{dn}{dm'}dm'}$$ of strings with masses below $ m $. At 17.3 GeV 50% of strings are lighter than 1 GeV, at 200 GeV the fraction is only 10%. Strings below 1 GeV cannot produce any baryons. If we want to create a $ \Omega $ given a $ \mathrm{s}-\bar{\mathrm{d}} $ string, we will find in addition a $ \bar{\Xi } $, since we have to break the string with the creation of a $ \mathrm{ss}-\bar{\mathrm{s}}\bar{\mathrm{s}} $ pair. Therefore the minimum mass is above 3 GeV in the best case scenario, where one strange quark is already given by the initial string. Consequently, it is hard to create $ \Omega $’s at low beam energies since only 10% of the strings have the necessary mass, whereas at RHIC-energies 70% of the strings could kinematically produce $ \Omega $’s. Conclusions =========== We can conclude that the different masses for the strings at the different beam energies are responsible for a possible suppression of heavy hadrons in pp collisions as compared to $ \mathrm{e}^{+}\mathrm{e}^{-} $annihilation. If the hadron mass is small as compared to the typical string energy the hadron multiplicity ratios reach asymptotic values. A further increase of the string energy leads only to an overall increase of the produced hadron multiplicity leaving their relative ratio unchanged. If the string mass becomes comparable to the hadron masses, the production of these hadrons are suppressed due to the very limited phase space available. This explains why in thermal fits [@becattini] a common constant temperature has been found for high energy $ \mathrm{p}\bar{\mathrm{p}} $ collisions and $ \mathrm{e}^{+}\mathrm{e}^{-} $annihilation. The physical origin of this phenomenon has in our model nothing to do with the formation of a thermal system. Therefore it is premature to identify this fit parameter with a true temperature. Employing a string fragmentation model which describes the kinematical variables as well the multiplicities of particle species in $ \mathrm{e}^{+}\mathrm{e}^{-} $, $ \mathrm{pp} $ and $ \mathrm{p}\bar{\mathrm{p}} $ collisions allows to interpret the results of [@becattini] in physical terms without touching the claim that data can be well fitted using the functional forms of a grand canonical description. [1]{} J. Rafelski, Phys.Rept.88:331,1982 , Phys. Rev. Lett. 48:1066 (1982) F. Antinori et al., Nucl.Phys.A661:130 (1999) H.J. Drescher et al., hep-ph/0007198 accepted by Phys. Rep., Phys.Rev.Lett.86:3506 (2001) Opal collaboration: R. Ackers et al., Zeit.Phys. C63:181 (1994), G. Alexander et al., Zeit.Phys. C73:569 (1997), Zeit.Phys. C73:587 (1997) M. Gazdzicki, Nucl.Phys. A528:754 (1991) UA5 Collaboration, R.E. Ansorge et al., Nucl.Phys. B328:36 (1989) F. Beccattini, Z.Phys.C76:269 (1997), Z.Phys.C72:491 (1996)
{ "pile_set_name": "ArXiv" }
--- abstract: 'A continuous-variable measurement-device-independent quantum key distribution (CV-MDI QKD) protocol using squeezed states is proposed where the two legitimate partners send Gaussian-modulated squeezed states to an untrusted third party to realize the measurement. Security analysis shows that the protocol can not only defend all detector side channels, but also attain higher secret key rates than the coherent-state-based protocol. We also present a method to improve the squeezed-state CV-MDI QKD protocol by adding proper Gaussian noise to the reconciliation side. It is found that there is an optimal added noise to optimize the performance of the protocol in terms of both key rates and maximal transmission distances. The resulting protocol shows the potential of long-distance secure communication using the CV-MDI QKD protocol.' author: - 'Yi-Chen Zhang$^1$' - 'Zhengyu Li$^{2,3}$' - 'Song Yu$^{1,}$' - Wanyi Gu$^1$ - Xiang Peng$^2$ - 'Hong Guo$^{1,2,3}$' title: | Continuous-variable measurement-device-independent quantum key distribution\ using squeezed states --- Introduction ============ Quantum key distribution (QKD) [@Gisin_RevModPhys_2002; @Scarani_RevModPhys_2009] is the most prominent application of quantum information science, which accomplishes the secure key distribution phase of an encrypted communication between two legitimate partners, i.e., Alice and Bob. The continuous-variable approach of QKD (CV-QKD) [@Braunstein_RevModPhys_2005; @Xiang-Bin_PhysReport_2007; @Weedbrook_RevModPhys_2012], based on the Gaussian modulation of Gaussian states, has attracted much attention in the past few years [@Weedbrook_RevModPhys_2012; @Madsen_nature_2012; @Jouguet_nature_2013; @Weedbrook_PhysRevA_2013] mainly because its associated homodyne or heterodyne detection offers the prospect of high detection efficiency and high repetition rate. Generally speaking, there are eight types of one-way Gaussian CV-QKD protocols, which are classified according to Alice’s sending states (squeezed or coherent states), Bob’s measurement methods (homodyne or heterodyne measurement), and reconciliation methods (direct or reverse reconciliation ) [@Ralph_PhysRevA_1999; @Hillery_PhysRevA_2000; @Cerf_PhysRevA_2001; @Gottesman_PhysRevA_2001; @Grosshans_PhysRevLett_2002; @Grosshans_Nature_2003; @Weedbrook_PhysRevLett_2004; @Patron_PhysRevLett_2009]. A CV-QKD protocol using squeezed states, heterodyne detection, and reverse reconciliation [@Patron_PhysRevLett_2009] outperforms these eight Gaussian protocols [@Patron_PhysRevLett_2009; @G_Patron_PhD_2007], which can be treated as the protocol using squeezed states and homodyne detection [@Ralph_PhysRevA_1999; @Hillery_PhysRevA_2000; @Cerf_PhysRevA_2001; @Gottesman_PhysRevA_2001] followed by a Gaussian added noise. Furthermore, a trusted added noise model is introduced into the squeezed state and homodyne detection protocol, and it performs a longer transmission distance with carefully chosen noise parameters [@Pirandola_PhysRevLett_2009; @Patron_PhysRevLett_2009; @G_Patron_PhD_2007]. The increased performance can be understood as that adding trusted noise to the receiver will cause the mutual information between the eavesdropper and the receiver to decrease more than that between the two legitimate partners. Comparing to the generation of a coherent state, generating a squeezed state is much more difficult, which becomes the most difficult part of implementing the CV-QKD protocol using squeezed states. Recent research shows that the largest achievable two-mode squeezing in a stable optical configuration has already been reached at about $10$ dB [@Eberle_OptExpr_2013], and an experiment using the CV-QKD protocol with squeezed states and homodyne detection has been successfully demonstrated [@Madsen_nature_2012; @Peuntinger_arXiv_2014], which shows the potential for implementing the squeezed-state CV-QKD in real life. In this paper, we introduce the use of squeezed states into a recently proposed protocol, the continuous-variable measurement-device-independent QKD (CV-MDI QKD) protocol [@Zhengyu_PhysRevA_2013; @Pirandola_arXiv_2013]. The MDI-QKD protocol [@Pirandola_PhysRevLett_2012_MDI; @Lo_PhysRevLett_2012_MDI; @XFMa_PhysRevA2_2012] was first proposed to defend detector side channels. Then many methods were introduced to improve the secret key rate and transmission distance of the protocol, such as using different kinds of sources [@XBWang_PhysRevA_2013; @Zhou_PhysRevA_2013; @Li_OpticsLetters_2014], explicitly utilizing the decoy states to dramatically increase the secret key rates [@XBWang_PhysRevA2_2013; @XBWang_PhysRevA_2014], enhancing the practical security by tight finite-size analysis [@XFMa_PhysRevA_2012; @TTSOng_PhysRevA_2012; @Curty_NatCommun_2014], etc. All the efforts were aimed at improving the performance of the protocol, which are also our pursuits on the CV-MDI QKD protocol. Here we first present the equivalent entanglement-based (EB) scheme and the prepare-and-measure (PM) scheme of the squeezed-state CV-MDI QKD. We find that the transmission distance of the squeezed-state CV-MDI QKD protocol is longer than the coherent-state-based protocol. In addition, we introduce the trusted added noise model to the receiver, which can further improve the transmission distance. Furthermore, in the most asymmetric case, even if the variance in the EPR is as small as $5.04$ (referring to $10$ dB squeezing [@Text]), the transmission distance can still reach $100.5$ km, which shows the potential for actually implementing the CV-MDI QKD protocol using squeezed states. The rest of this paper is organized as follows. In Sec. \[sec:2\], we propose a CV-MDI QKD protocol using squeezed states. We optimize the CV-MDI QKD protocol using squeezed states by adding optimal Gaussian noise to the reconciliation side in Sec. \[sec:3\]. In Sec. \[sec:4\], we show the numerical simulation results of the secret key rate and give the optimal value of the added noise under different situations. Our conclusions are drawn in Sec. \[sec:5\]. \[sec:2\] SQUEEZED-STATE CV-MDI QKD ==================================== In this section, we first present the idea and basic notions of the CV-MDI QKD protocol using squeezed states and then derive the secure bound of the protocol. The standard PM description of the CV-MDI QKD protocol using squeezed states is shown in Fig. 1 and is described as follows: ![ (Color online) The prepare-and-measurement scheme of the CV-MDI QKD protocol using squeezed states where quantum channels and Charlie are fully controlled by Eve. Practical detectors on Charlie’s side have the same quantum efficiency and electronic noise.](fig1.eps){width="8.0cm"} *Step 1*. Alice and Bob randomly draw values $x_A (p_A)$ and $x_B (p_B)$ from two Gaussian distributed set with $0$ mean and variance $V_A-1$ and $V_B-1$, respectively, and use these numbers to modulate *x*-quadrature (*p*-quadrature) of the squeezed state. The modulation processing can be achieved by mixing the x-quadrature (p-quadrature) squeezed states on a beamsplitter of high transmissivity ($T \sim 99 \%$) with a coherent state of intensity $\frac{{x_A^2\left( {p_A^2} \right)}}{{1 - T}}$ and $\frac{{x_B^2\left( {p_B^2} \right)}}{{1 - T}}$ [@Patron_PhysRevLett_2009]. Then they send these states to the untrusted third party (Charlie) through two different quantum channels. *Step 2*.Charlie combines two received modes $A'$ and $B'$ with a beam splitter (50:50), and the output modes of the beam splitter are $C$ and $D$. Then he measures *x*-quadrature of mode $C$ and *p*-quadrature of mode $D$ by homodyne detectors and publicly announces the measurement results $x_{C}, p_{D}$ to Alice and Bob through classical channels. *Step 3*. After receiving Charlie’s measurement results $x_{C}, p_{D}$, Bob modifies his data to ${x'_B}$ (${p'_B}$), while Alice keeps her data $x_A (p_A)$ unchanged. *Step 4*. Once Alice and Bob have collected a sufficiently large amount of correlated data, they first perform a parameter estimation from a randomly chosen sample of final data ${x_A}$ (${p_A}$) and ${x'_B}$ (${p'_B}$). Then Alice and Bob proceed with classical data postprocessing namely, information reconciliation and privacy amplification using an authenticated public channel. The reconciliation can be performed in two ways: either direct reconciliation (DR) or reverse reconciliation (RR). ![ (Color online) The entanglement-based scheme of the CV-MDI QKD protocol using squeezed states where all detectors represent homodyne detection and EPR states are two-mode squeezed states. Two quantum channels and Charlie are fully controlled by Eve, but Eve has no access to the apparatuses in Alice’s and Bob’s stations. The imperfection of the detectors is characterized by quantum efficiency $\eta$ and electronic noise $\upsilon_1 = \upsilon_2 = \upsilon_{el}$.[]{data-label="fig2"}](fig2.eps){width="7cm"} The PM description presented above is equivalent to the EB scheme shown in Fig. \[fig2\]. Although the EB version does not correspond to the actual implementation, it is fully equivalent to the PM version from a secure point of view, and it provides a powerful description for establishing security proof [@Grosshans_QIC_2003]. The EB scheme of the proposed CV-MDI QKD using squeezed states can be described as follows: *Step 1*. Alice and Bob respectively generate an Einstein-Podolsky-Rosen (EPR) state EPR$_1$ and EPR$_2$ with variance $V_A$ and $V_B$ and keep mode $A_3$ and $B_3$ in each side. Then they send the other mode $A_2$ and $B_2$ to the untrusted third party (Charlie) through two different quantum channels with length $L_{AC}$ and $L_{BC}$. *Step 2*.Charlie combines two received modes $A_1$ and $B_1$ with a beam splitter (50:50) and the output modes of the beam splitter are $C$ and $D$. Then he measures the *x*-quadrature of mode $C$ and *p*-quadrature of mode $D$ with homodyne detectors and publicly announces the measurement results $x_{C}, p_{D}$ to Alice and Bob through classical channels. *Step 3*.Bob displaces the mode $B_3$ to $B_4$ by operation $D\left(x_{C}', p_{D}'\right)$ which depends on Charlie’s announced results $\left\{x_C, p_D\right\}$. Then Bob measures the mode $B_4$ to get the final data $\left\{x_B \left(p_B\right)\right\}$ using homodyne detection which randomly detects the *x*-quadrature or *p*-quadrature. Alice also measures the mode $A_3$ to get the final data $\left\{x_A \left(p_A\right)\right\}$ using homodyne detection. *Step 4*. Once Alice and Bob have collected a sufficiently large amount of correlated data, they use an authenticated public channel to perform parameter estimation from a randomly chosen sample of final data $\left\{x_A, p_A\right\}$ and $\left\{x_B, p_B\right\}$. Then Alice and Bob proceed with classical data postprocessing namely information reconciliation and privacy amplification to distill a secret key. The reconciliation can be performed in two ways: either direct reconciliation (DR) or reverse reconciliation (RR). The detector’s imperfection is characterized by quantum efficiency $\eta$ and electronic noise $\upsilon_{el}$, which is shown in Fig. \[fig2\]. The variance $\upsilon_{1,2}$ of the thermal state ${\rho _{{F_0}}}$ and ${\rho _{{I_0}}}$ is chosen to obtain the appropriate expression for practical homodyne detection $\upsilon_{1,2} = 1 + {{{\upsilon _{el}}} \mathord{\left/ {\vphantom {{{\upsilon _{el}}} {\left( {1 - \eta } \right)}}} \right. \kern-\nulldelimiterspace} {\left( {1 - \eta } \right)}}$ [@Lodewyck_PhysRevA_2007]. From the analysis above, one can find that the EB scheme proposed here shares the same demonstration with the one in Ref. [@Zhengyu_PhysRevA_2013] except for that what measurements Alice and Bob use are substituted by homodyne detection. Thus, the EB scheme discussed here is equivalent to the conventional CV-QKD with squeezed states and homodyne detection. The (asymptotical) secret key rate ${K}$ against collective attacks for reverse reconciliation is given by [@Devetak_ProcRSoc_2005] $$\begin{aligned} {K} = \beta I\left( {A:B} \right) - \chi \left( {B:E} \right),\end{aligned}$$ where $\beta \in[0,1]$ is the reconciliation efficiency, $I(A:B)$ is the classical mutual information between Alice and Bob, $\chi(B:E)$ is the Holevo quantity [@Nielsen_QCQI] $$\begin{aligned} \chi \left( {B:E} \right) = S\left( {{\rho _E}} \right) - \sum\nolimits_{x_B} {p\left( x_B \right)S\left( {{\rho _{E|x_B}}} \right)},\end{aligned}$$ where $S(\rho)$ is the von Neumann entropy of the quantum state $\rho$, $x_B$ is Bob’s measurement result obtained with the probability $p\left( x_B \right)$, $\rho _{E|x_B}$ is the corresponding state of Eve’s ancillary, and ${\rho _E} = \sum\nolimits_{x_B} {p\left( x_B \right){\rho _{E|x_B}}}$ are Eve’s partial states. Firstly, Eve is able to purify the whole system $\rho_{A_3 B_4}$ to maximize her information, we have $S\left( {{\rho _E}} \right) = S\left( {{\rho _{{A_3}{B_4}}}} \right)$. Secondly, after Bob’s projective measurement resulting in $x_B$, the system $\rho_{A_3 E}$ is pure, so that ${S\left( {E|x_B} \right) = S\left( {{A_3}|x_B} \right)}$. In practical experiment, we calculate the covariance matrix ${{\gamma _{{A_3}{B_4}}}}$ of correlated variables from a randomly chosen sample of measurement data. According to the Gaussian optimality theorem, we assume the final state ${{\rho _{{A_3}{B_4}}}}$ shared by Alice and Bob is Gaussian to maximize the quantum information available to Eve. Thus, the entropies $S({\rho _{{A_3}{B_4}}})$ and $\sum\nolimits_{x_B} {p\left( x_B \right)S\left( {{\rho _{{A_3}|x_B}}} \right)}$ can be calculated using the covariance matrices ${\gamma _{{A_3}{B_4}}}$ characterizing the state ${\rho _{{A_3}{B_4}}}$ and ${{\gamma _{{A_3}|x_B}}}$ characterizing the state ${{\rho _{{A_3}|x_B}}}$. So the expression for ${\chi _{BE}}$ can be further simplified as follows $$\begin{aligned} \label{eq} \chi \left( {B:E} \right) = \sum\limits_{i = 1}^2 {G\left( {\frac{{{\lambda _i} - 1}}{2}} \right) - G\left( {\frac{{{\lambda _3} - 1}}{2}} \right)},\end{aligned}$$ where $G(x) = (x + 1)\log_2 (x + 1) - x\log_2 x$, ${\lambda _{1 \left(2\right)}}$ are the symplectic eigenvalues of the covariance matrix ${{\gamma _{{A_3}{B_4}}}}$ and ${\lambda _3}$ is the symplectic eigenvalue of the covariance matrix ${\gamma _{{A_3}|x_B}}$. As discussed above, in experiment Alice and Bob can get the covariance matrix ${\gamma _{{A_3}{B_4}}}$ from parameter estimation step. The covariance matrix ${\gamma _{{A_3}{B_4}}}$ depends on the system and the gain of the displacement, which is written as $$\label{eq2} {\gamma _{{A_3}{B_4}}} = \left[ {\begin{array}{*{20}{c}} {{\gamma _{{A_3}}}} & {\sigma _{{A_3}{B_4}}^T} \\ {{\sigma _{{A_3}{B_4}}}} & {{\gamma _{{B_4}}}} \\ \end{array}} \right] = \left[ {\begin{array}{*{20}{c}} {a{I_2}} & {c{\sigma _z}} \\ {c{\sigma _z}} & {b{I_2}} \\ \end{array}} \right],$$ where ${{{\rm I}_n}}$ is the $n \times n$ identity matrix and ${\sigma _z}$ = diag (1, -1). The symplectic eigenvalues ${\lambda _{1 - 2}}$ of the above matrix are given by $$\label{eq12} \lambda _{1,2}^2 = \frac{1}{2}\left[ {{\rm A} \pm \sqrt {{{\rm A}^2} - 4{{\rm B}^2}} } \right],$$ where we have used the notations $$\left\{ \begin{array}{l} {\rm A} = {a^2} + {b^2} - 2{c^2} \\ {\rm B} = ab - {c^2} \\ \end{array} \right..$$ The symplectic eigenvalues ${\lambda _3}$ of the matrix ${\gamma _{{A_3}|{x_B}}} = {\gamma _{{A_3}}} - \sigma _{{A_3}{B_4}}^T{\left( {X{\gamma _{{B_4}}}X} \right)^{-1}}{\sigma _{{A_3}{B_4}}}$ ($X$ = diag (1, 0)), after Bob’s homodyne measurement, is given by $${\lambda _3^2} = a\left( {a - {{{c^2}} \mathord{\left/ {\vphantom {{{c^2}} b}} \right. \kern-\nulldelimiterspace} b}} \right).$$ ![ (Color online) The entanglement-based scheme of the modified squeezed-state CV-MDI QKD protocol with general Gaussian added noise on the reconciliation side. Two quantum channels and Charlie are fully controlled by Eve, but Eve has no access to the apparatuses in Alice’s and Bob’s stations. The imperfection of the detectors is characterized by quantum efficiency $\eta$ and electronic noise $\upsilon_1 = \upsilon_2 = \upsilon_{el}$.[]{data-label="fig3"}](fig3.eps){width="7cm"} \[sec:3\] MODIFIED SQUEEZED-STATE CV-MDI QKD ============================================ In this section, we propose a modified CV-MDI QKD protocol using squeezed states by adding proper classical Gaussian noise to the reconciliation side (Alice’s side for the DR protocol or Bob’s side for the RR protocol). This method has practical benefits because there exists certain preparation noise [@Filip_PhysRevA_2008; @Shen_JPhysB_2009; @Shen_PhysRevA_2011; @Weedbrook_PhysRevLett_2010; @Weedbrook_PhysRevA_2012; @Weedbrook_PhysRevA_2014] and detection noise [@Jouguet_nature_2013; @Lodewyck_PhysRevA_2007] in a practical system. If we optimize such noise in the way we discussed below, the performance of the protocol will be improved. ![(Color online) Secret key rates of the coherent-state (black), squeezed-state (blue) CV-MDI QKD protocol, and the modified squeezed-state CV-MDI QKD protocol (red) in the symmetric case ($L_{AC} = L_{BC}$) with perfect homodyne detectors ($\eta = 1$, ${\upsilon _{el}} = 0$) and imperfect homodyne detectors ($\eta = 0.9$, ${\upsilon _{el}} = 0.015$). The dot-dashed line and solid line represent the situation for using perfect and imperfect detectors, respectively. Here we use the ideal reconciliation efficiency $\beta = 1$, large variance $V_A = V_B = 10^5$, and $\varepsilon = 0.002$. []{data-label="fig4"}](fig4.eps){width="8.5cm"} The EB scheme of the modified protocol is illustrated in Fig. 3, where Alice and Bob implement the original squeezed-state CV-MDI QKD protocol as we proposed in the last section but Bob adds some proper noise before his homodyne detection. In the EB scheme, the added Gaussian phase-insensitive noise is modeled by mixing the original mode $B_4$ with a thermal state (half of an EPR) of variance $N_R$ by a beam splitter of transmissivity $T_R$, thus ${\chi _N} = {{\left( {1 - {T_R}} \right){N_R}} \mathord{\left/ {\vphantom {{\left( {1 - {T_R}} \right){N_R}} {{T_R}}}} \right. \kern-\nulldelimiterspace} {{T_R}}}$. In the corresponding PM scheme, the added noise means that Bob adding proper classical Gaussian noise of variance ${\chi _N} = {{\left( {1 - {T_R}} \right){N_R}} \mathord{\left/ {\vphantom {{\left( {1 - {T_R}} \right){N_R}} {{T_R}}}} \right. \kern-\nulldelimiterspace} {{T_R}}}$ before sending the modulated squeezed states to Charlie. Then we follow the analysis of Sec. \[sec:2\] for collective attacks, but it is clear from Fig. 3 that two additional modes $N_1$ and $N_3$ need to be included in the calculation of the reverse reconciliation protocol. Here we only derive the expression of the reverse reconciliation protocol and for the direct reconciliation protocol we can use the similar method to calculate. By replacing the Eq. \[eq\], $\chi \left( {B:E} \right)$ is calculated from the following equation: $$\begin{aligned} \chi \left( {B:E} \right) = \sum\limits_{i = 1}^2 {G\left( {\frac{{{\lambda _i} - 1}}{2}} \right) - \sum\limits_{i = 3}^5 {G\left( {\frac{{{\lambda _i} - 1}}{2}} \right)} } ,\end{aligned}$$ where ${\lambda _{1 \left(2\right)}}$ remains the same as the Eq. \[eq12\] while ${\lambda _{3 \left(4, 5\right)}}$ represents the symplectic eigenvalues of the covariance matrix ${\gamma _{{A_3 N_1 N_3}|x_B}}$, which is given by $${\gamma _{{A_3}{N_1}{N_3}|{x_B}}} = {\gamma _{{A_3}{N_1}{N_3}}} - \sigma _{{A_3}{N_1}{N_3}{B_5}}^T{\left( {X{\gamma _{{B_5}}}X} \right)^{ - 1}}{\sigma _{{A_3}{N_1}{N_3}{B_5}}},$$ where the matrices ${\gamma _{{A_3}{N_1}{N_3}}}$, ${\gamma _{{B_5}}}$ and $\sigma _{{A_3}{N_1}{N_3}{B_5}}$ can all be derived from the decomposition of the matrix $${\gamma _{{A_3}{N_1}{N_3}{B_5}}} = \left[ {\begin{array}{*{20}{c}} {{\gamma _{{A_3}{N_1}{N_3}}}} & {\sigma _{{A_3}{N_1}{N_3}{B_5}}^T} \\ {{\sigma _{{A_3}{N_1}{N_3}{B_5}}}} & {{\gamma _{{B_5}}}} \\ \end{array}} \right].$$ The above matrix can be derived with an appropriate rearrangement of lines and columns from the matrix describing the system in Fig. 3 $${\gamma _{{A_3}{B_5}{N_3}{N_1}}} = {Y_{B_4 N_2}}\left( {{\gamma _{{A_3}{B_4}}} \oplus {\gamma _{{N_2}{N_1}}}} \right)Y_{B_4 N_2}^T,$$ where ${\gamma _{{A_3}{B_4}}}$ is the same as expressed by Eq. \[eq2\], ${\gamma _{{N_2}{N_1}}}$ is the standard covariance matrix of an EPR state with variance $N_R$ and ${Y_{{B_4}{N_2}}} = {I_2} \oplus {Y^{BS}} \oplus {I_2}$ where ${Y^{BS}}$ can be written by $${Y^{BS}} = \left[ {\begin{array}{*{20}{c}} {\sqrt {{T_R}} \cdot{I_2}} & {\sqrt {1 - {T_R}} \cdot{I_2}} \\ { - \sqrt {1 - {T_R}} \cdot{I_2}} & {\sqrt {{T_R}} \cdot{I_2}} \\ \end{array}} \right].$$ \[sec:4\] NUMERICAL SIMULATION AND DISCUSSION ============================================= In this section, the performance of the proposed and the modified squeezed-state CV-MDI QKD protocol is illustrated and compared with the coherent-state based protocol [@Zhengyu_PhysRevA_2013; @Pirandola_arXiv_2013]. ![(Color online) Optimal added noise for the modified squeezed-state CV-MDI QKD protocol in symmetric case ($L_{AC} = L_{BC}$) using perfect homodyne detectors ($\eta = 1$, ${\upsilon _{el}} = 0$) and imperfect homodyne detectors ($\eta = 0.9$, ${\upsilon _{el}} = 0.015$). Here we use the ideal reconciliation efficiency $\beta = 1$, large variance $V_A = V_B = 10^5$, and $\varepsilon = 0.002$. []{data-label="fig5"}](fig5.eps){width="8.5cm"} As discussed above, in a practical experiment, Alice and Bob can get the covariance matrix ${\gamma _{{A_3}{B_4}}}$ from parameter estimation step. However, in a numerical simulation, a model of simulating what the channels are and what Charlie does is needed to get ${\gamma _{{A_3}{B_4}}}$. To compare with the performance of the coherent-state CV-MDI QKD protocol [@Zhengyu_PhysRevA_2013], we use the same method to simulate the channels’ environment (two independent entangling cloner attacks) and Charlie’s measurement (standard Bell-state measurement), which is illustrated in Fig. \[fig2\]. The relationships are as follows: $$\begin{aligned} \left\{ \begin{array}{l} {{\hat A}_1} = \sqrt {{T_1}} {{\hat A}_2} + \sqrt {1 - {T_1}} {{\hat E}_1} \\ {{\hat B}_1} = \sqrt {{T_2}} {{\hat B}_2} + \sqrt {1 - {T_2}} {{\hat E}_2} \\ {{\hat C}_1} = {{{{\hat A}_1}} \mathord{\left/ {\vphantom {{{{\hat A}_1}} {\sqrt 2 - }}} \right. \kern-\nulldelimiterspace} {\sqrt 2 - }}{{{{\hat B}_1}} \mathord{\left/ {\vphantom {{{{\hat B}_1}} {\sqrt 2 }}} \right. \kern-\nulldelimiterspace} {\sqrt 2 }} \\ {{\hat D}_1} = {{{{\hat A}_1}} \mathord{\left/ {\vphantom {{{{\hat A}_1}} {\sqrt 2 + }}} \right. \kern-\nulldelimiterspace} {\sqrt 2 + }}{{{{\hat B}_1}} \mathord{\left/ {\vphantom {{{{\hat B}_1}} {\sqrt 2 }}} \right. \kern-\nulldelimiterspace} {\sqrt 2 }} \\ {{\hat C}_2} = \sqrt \eta {{\hat C}_1} + \sqrt {1 - \eta } {{\hat F}_0} \\ {{\hat D}_2} = \sqrt \eta {{\hat D}_1} + \sqrt {1 - \eta } {{\hat I}_0} \\ {{\hat B}_{4x}} = {{\hat B}_{3x}} + g{{\hat C}_{2x}} \\ {{\hat B}_{4p}} = {{\hat B}_{3p}} - g{{\hat D}_{2p}} \\ \end{array} \right..\end{aligned}$$ ![(Color online) A comparison among the maximal transmission distance for the coherent-state, squeezed-state CV-MDI QKD protocol, and the modified squeezed-state CV-MDI QKD protocol with perfect homodyne detectors ($\eta = 1$, ${\upsilon _{el}} = 0$) and imperfect homodyne detectors ($\eta = 0.9$, ${\upsilon _{el}} = 0.015$), within which the key rate $K$ is positive. Here we use the ideal reconciliation efficiency $\beta = 1$, large variance $V_A = V_B = 10^5$, and $\varepsilon = 0.002$. []{data-label="fig6"}](fig6.eps){width="8.8cm"} The parameters that will affect the secret key rate are the reconciliation efficiency $\beta$, the variance of Alice’s and Bob’s modulation $V_A-1$, $V_B-1$, the transmission efficiency $T_1$, $T_2$, excess noise $\varepsilon_1$, $\varepsilon_2$ of two quantum channels, the inefficiency $\eta$ and the electronic noise ${\upsilon _{el}}$ of the practical homodyne detector. Here we first choose large variance $V_A = V_B = 10^5$ to see the performance of ideal modulation, then we will use the practical variance $V_A = V_B = 5.04$ to see the realistic performance. Excess noises are $\varepsilon_1 = \varepsilon_2 = \varepsilon = 0.002$ and transmittances are $T_1 = 10^{-\alpha L_{AC}/10}$, $T_2 = 10^{-\alpha L_{BC}/10}$ ($\alpha = 0.2$ dB/km) for simulation, which are standard parameters in one-way CV-QKD experiment [@Jouguet_nature_2013]. Furthermore, we use $\eta = 1$, ${\upsilon _{el}} = 0$ representing for the perfect homodyne detector and $\eta = 0.9$, ${\upsilon _{el}} = 0.015$ for the imperfect detector. ![(Color online) A comparison among the maximal transmission distance for the coherent-state, squeezed-state CV-MDI QKD protocol, and the modified squeezed-state CV-MDI QKD protocol with perfect homodyne detectors ($\eta = 1$, ${\upsilon _{el}} = 0$) and imperfect homodyne detectors ($\eta = 0.9$, ${\upsilon _{el}} = 0.015$), within which the key rate $K$ is positive. Here we use the realistic variance $V_A = V_B = 5.04$ and $\varepsilon = 0.002$. []{data-label="fig7"}](fig7.eps){width="8.8cm"} ![(Color online) Secret key rates of the coherent-state (black), squeezed-state (blue) CV-MDI QKD protocol, and the modified squeezed-state CV-MDI QKD protocol (red) in the most asymmetric case ($L_{BC} = 0 km$) with perfect homodyne detectors ($\eta = 1$, ${\upsilon _{el}} = 0$) and imperfect homodyne detectors ($\eta = 0.9$, ${\upsilon _{el}} = 0.015$). The dot-dash line and solid line represent the situation of using perfect and imperfect detectors, respectively. Here we use the realistic parameters: $V_A = V_B = 5.04$ and $\varepsilon = 0.002$. []{data-label="fig8"}](fig8.eps){width="8.5cm"} First, we consider the performance of the symmetric case where the length of two quantum channels is equal ($L_{AC} = L_{BC}$). We calculate the secret key rate $K$ as a function of transmission distance $d = L_{AC} = L_{BC}$ with perfect detectors or imperfect detectors. The simulation results are shown in Fig. \[fig4\], where we also calculate the CV-MDI-QKD protocol using coherent states [@Zhengyu_PhysRevA_2013] for comparison. We find that the secret key rate of the squeezed-state CV-MDI QKD protocol is always larger than coherent-state based protocol. Explicitly, the total maximal transmission distance ($L_{AB} = L_{AC}+L_{BC}$) of our squeezed-state CV-MDI QKD protocol increases both $6.1$ km using perfect detectors and imperfect detectors than that of the coherent-state based protocol. Furthermore, by adding proper Gaussian noise in Bob’s side, the performance of the modified squeezed-state CV-MDI QKD protocol improves. The total maximal transmission distance increases both $1.3$ km and the optimal added noise is illustrated in Fig. \[fig5\]. We also observe that the imperfections of the homodyne detectors decrease $L_{AB}$, i. e. , using $\eta = 0.9$, ${\upsilon _{el}} = 0.015$ homodyne detectors decrease $L_{AB}$ $7.2$ km for the modified squeezed-state CV-MDI QKD protocol. This is because the imperfections of the homodyne detectors dramatically increase the equivalent excess noise of the whole system [@Zhengyu_PhysRevA_2013]. Second, we also find that when Charlie’s position is close to Bob, the total maximal transmission distance $L_{AB}$ will increase to a relatively longer distance. To understand this phenomenon intuitively, we can treat the EB scheme of the CV-MDI QKD protocol as a CV teleportation process, in which Bob and Charlie preshare an EPR source and then teleport mode $A¡¯$ from Alice to Bob. Any loss and noise in Bob¡¯s channel will decrease the quality of the EPR source and will decrease the fidelity of the teleportation result. Thus, here we also consider the performance of the asymmetric case where $L_{AC} \neq L_{BC}$. As illustrated in Fig. \[fig6\], $L_{AB}$ increases when $L_{BC}$ decreases. In the asymmetric case, the performance of the modified squeezed-state CV-MDI QKD protocol is also improved by adding the proper noise to Bob’s side. The optimal input noise can enhance $L_{AB}$ to $91.2$ km when using perfect detectors and $20.0$ km when using imperfect detectors. Then we change the variance from ideal ($V_A = V_B = 10^5$) to a practical one ($V_A = V_B = 5.04$), which is shown in Fig. \[fig7\]. The total maximal transmission distance $L_{AB}$ can also reach $84.4$ km using perfect detectors and $13.7$ km using imperfect detectors, which allows one to directly use the EPR state as the source in a practical experiment. And if Alice and Bob use EPR sources, the squeezed-state CV-MDI QKD protocol will be more secure because they could completely outplay side-channel attacks in their private spaces [@Pirandola_PhysRevLett_2012_MDI]. Compared with the performance of using perfect and imperfect detectors, we find that the inefficiency and electronic noise of the practical homodyne detector have a significant influence on the transmission distance. It is possible that these imperfections can be compensated by optical phase-sensitive-amplifiers [@Fossier_JPhysB_2009; @My_JPhysB_2014]. ![(Color online) Optimal input noise for the modified squeezed-state CV-MDI QKD protocol in the most asymmetric case ($L_{BC} = 0 km$) using perfect homodyne detectors ($\eta = 1$, ${\upsilon _{el}} = 0$) and imperfect homodyne detectors ($\eta = 0.9$, ${\upsilon _{el}} = 0.015$). Here we use the realistic parameters: $V_A = V_B = 5.04$ and $\varepsilon = 0.002$. []{data-label="fig9"}](fig9.eps){width="8.5cm"} Finally, we consider the performance of the most asymmetric case where we put Charlie on Bob’s side ($L_{BC} = 0 km$). As illustrated in Fig. \[fig8\], $L_{AB}$ increases dramatically compared with the symmetric case. In the most asymmetric case, the secret key rate of the squeezed-state CV-MDI QKD protocol is always larger than that of the coherent-state based protocol. The transmission distance of the coherent-state protocol is zero. Thus, there is no line representing the coherent-state CV-MDI QKD protocol using imperfect detectors in Fig. \[fig8\]. Thus, the total maximal transmission distance increases $15.7$ km and $10.5$ km when using perfect detectors and imperfect detectors. Furthermore, the modified squeezed-state CV-MDI QKD protocol achieves optimal performance than the original protocol by adding proper noise in Bob’s side, i. e. , $L_{AB}$ increases more $4.9$ km and $3.3$ km. The optimal input noise is illustrated in Fig.  \[fig9\]. The reason of the improvement is that the added noise not only lowers the mutual information between Alice and Bob, but also lowers that between Eve and Bob. When the influence on Eve and Bob is stronger than that on Alice and Bob, the secret key rate is enhanced and the performance is improved. \[sec:5\]CONCLUSION =================== In this paper, we proposed a continuous-variable measurement-device-independent QKD protocol using squeezed states, which outperforms the coherent-state-based protocol in terms of the secret key rate and maximal transmission distance. Security analysis shows that the protocol is immune to attacks against the detector and secure against collective attacks in the asymptotical limit. Furthermore, we also presented a method to optimize the performance of the reverse reconciliation squeezed-state continuous-variable measurement-device-independent QKD protocol by adding proper Gaussian noise to Bob’s side. It is found that there is an optimal noise Bob need to add to maximize the secret key rate and total transmission distance for the reverse reconciliation protocol. The resulting protocol exhibits the optimal performance and shows the potential of long-distance secure communication using the continuous-variable measurement-device-independent QKD protocol. We restricted our discussion to collective-attack cases in this paper. Since the protocol using squeezed states and homodyne detection is also shown to be secure against coherent attacks in the finite-size regime [@Furrer_PhysRevLett_2012; @Furrer_arXiv_2014], an interesting extension to this paper would be to further derive the security bound against coherent attacks. This work was supported in part by the National Basic Research Program of China (973 Pro-gram) under Grant 2012CB315605 and 2014CB340102, in part by the National Science Fund for Distinguished Young Scholars of China (Grant No. 61225003), in part by the National Natural Science Foundation under Grant 61101081, 61271191 and 61271193, in part by the Fund of State Key Laboratory of Information Photonics and Optical Communications, and in part by the Fundamental Research Funds for the Central Universities. [99]{} N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. **74**, 145 (2002). V. Scarani, H. Bechmann-Pasquinucci, N. J. Cerf, M. Dušek, N. Lütkenhaus, and M. Peev, Rev. Mod. Phys. **81**, 1301 (2009). S. L. Braunstein and P. van Loock, Rev. Mod. Phys. **77**, 513 (2005). X.-B. Wang, T. Hiroshima, A. Tomita, M. Hayashi, Phys. Rep. **448**, 1 (2007). C. Weedbrook, S. Pirandola, R. García-Patrón, N. J. Cerf, T. C. Ralph, J. H. Shapiro, and S. Lloyd, Rev. Mod. Phys. **84**, 621 (2012). L. S. Madsen, V. C. Usenko, M. Lassen, R. Filip and U. L. Andersen, Nat. Commun. **3**, 1083 (2012). P. Jouguet, S. Kunz-Jacques, A. Leverrier, P. Grangier, and E. Diamanti, Nat. Photon. **7**, 378 (2013). C. Weedbrook, Phys. Rev. A **87**, 022308 (2013). T. C. Ralph, Phys. Rev. A **61**, 010303(R) (1999). M. Hillery, Phys. Rev. A **61**, 022309 (2000). N. J. Cerf, M. Lévy, and G. Van Assche, Phys. Rev. A **63**, 052311 (2001). D. Gottesman and J. Preskill, Phys. Rev. A **63**, 022309 (2001). F. Grosshans and P. Grangier, Phys. Rev. Lett. **88**, 057902 (2002). F. Grosshans, G. Van Assche, J. Wenger, R. Brouri, N. J. Cerf, and P. Grangier, Nature (London) **421**, 238 (2003). C. Weedbrook, A. M. Lance, W. P. Bowen, T. Symul, T. C. Ralph, and P. K. Lam, Phys. Rev. Lett. **93**, 170504 (2004). R. García-Patrón and N. J. Cerf, Phys. Rev. Lett. **102**, 130501 (2009). R. García-Patrón, Ph.D. thesis, Université Libre de Bruxelles, Bruxelles (2007). S. Pirandola, R. García-Patrón, S. L. Braunstein, and Seth Lloyd, Phys. Rev. Lett. **102**, 050503 (2009). T. Eberle, V. H[ä]{}ndchen, and R. Schnabel, Opt. Expr. **21** 11546 (2013). C. Peuntinger, B. Heim, C. R. M[ü]{}ller, C. Gabriel, C. Marquardt, and G. Leuchs, Phys. Rev. Lett. **113**, 060502 (2014). Z. Li, Y.-C. Zhang, F. Xu, X. Peng, and H. Guo, Phys. Rev. A **89**, 052301 (2014). S. Pirandola, C. Ottaviani, G. Spedalieri, C. Weedbrook, S. L. Braunstein, arXiv:1312.4104 (2013). S. L. Braunstein and S. Pirandola, Phys. Rev. Lett. **108**, 130502 (2012). H.-K. Lo, M. Curty, and B. Qi, Phys. Rev. Lett. **108**, 130503 (2012). X. Ma and M. Razavi, Phys. Rev. A **86**, 062319 (2012). Q. Wang and X.-B. Wang, Phys. Rev. A **88**, 052332 (2013). C. Zhou, W.-S. Bao, W. Chen, H.-W. Li, Z.-Q. Yin, Y. Wang, and Z.-F. Han, Phys. Rev. A **88**, 052333 (2013). M. Li, C.-M. Zhang, Z.-Q. Yin, W. Chen, S. Wang, G.-C. Guo, and Z.-F. Han, Opt. Lett. **39**, 4 (2014). X.-B. Wang, Phys. Rev. A **87**, 012320 (2013). Y.-H. Zhou, Z.-W. Yu, and X.-B. Wang, Phys. Rev. A **89**, 052325 (2014). X. Ma, C.-H. F. Fung, and M. Razavi, Phys. Rev. A **86**, 052305 (2012). T.-T. Song, Q.-Y. Wen, F.-Z. Guo, and X.-Q. Tan, Phys. Rev. A **86**, 022332 (2012). M. Curty, F. Xu, W. Cui, C. C. W. Lim, K. Tamaki, H.-K. Lo, Nat. Commun. **5**, 3732 (2014). In experimental papers, squeezing is often measured in decibels, defined so that a squeezing degree $r$ corresponds to $10{\log _{10}}\left( {{e^{2r}}} \right)dB$. Thus, 10 dB corresponds to $r = 1.15$, i.e., to an EPR variance $V = \cosh \left( {2r} \right) = 5.04$. F. Grosshans, N. J. Cerf, J. Wenger, R. Tualle-Brouri, and P. Grangier, Quantum Inf. Comput. **3**, 535 (2003). J. Lodewyck, M. Bloch, R. García-Patrón, S. Fossier, E. Karpov, E. Diamanti, T. Debuisschert, N. J. Cerf, R. Tualle-Brouri, S. W. McLaughlin, and P. Grangier, Phys. Rev. A. **76**, 042305 (2007). I. Devetak and A. Winter, Proc. R. Soc. London Ser. A **461**, 207 (2005). M. A. Nielsen and I. L. Chuang, *Quantum Computation and Quantum Communication* (Cambridge University Press, Cambridge, 2000). R. Filip, Phys. Rev. A **77**, 022310 (2008). Y. Shen, J. Yang, and H. Guo, J. Phys. B **42**, 235506 (2009). Y. Shen, X. Peng, J. Yang, and H. Guo, Phys. Rev. A **83**, 052304 (2011). C. Weedbrook, S. Pirandola, S. Lloyd, and T. C. Ralph, Phys. Rev. Lett. **105**, 110501 (2010). C. Weedbrook, S. Pirandola, and T. C. Ralph, Phys. Rev. A **86**, 022318 (2012). C. Weedbrook, C. Ottaviani, and S. Pirandola, Phys. Rev. A **89**, 012309 (2014). S. Fossier, E. Diamanti, T. Debuisschert, R. Tualle-Brouri, and P. Grangier, J. Phys. B **42**, 114014 (2009). Y.-C. Zhang, Z. Li, C. Weedbrook, S. Yu, W. Gu, M. Sun, X. Peng, and H. Guo, J. Phys. B **47**, 035501 (2014). F. Furrer, T. Franz, M. Berta, A. Leverrier, V. B. Scholz, M. Tomamichel and R. F. Werner, Phys. Rev. Lett. **109**, 100502 (2012). F. Furrer, Phys. Rev. A **90**, 042325 (2014).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We use equations for Faddeev amplitudes to solve the shell-model problem for four nucleons in the model space that includes up to $14 \hbar\Omega$ harmonic-oscillator excitations above the unperturbed ground state. Two- and three-body effective interactions derived from the Reid93 and Argonne V8’ nucleon-nucleon potentials are used in the calculations. Binding energies, excitations energies, point-nucleon radii and electromagnetic and strangeness charge form factors for $^4$He are studied. The structure of the Faddeev-like equations is discussed and a formula for matrix elements of the permutation operators in a harmonic-oscillator basis is given. The dependence on harmonic-oscillator excitations allowed in the model space and on the harmonic-oscillator frequency is investigated. It is demonstrated that the use of the three-body effective interactions improves the convergence of the results.' address: ' Department of Physics, University of Arizona, Tucson, Arizona 85721 ' author: - 'P. Navrátil[^1] and B. R. Barrett' title: ' Four-nucleon shell-model calculations in a Faddeev-like approach ' --- Introduction {#sec1} ============ Many different methods have been used to solve the few-body problem in the past. One of the most viable approaches appears to be the Faddeev method [@Fad60]. It has been successfully applied to solve the three-nucleon bound-state problem for various nucleon-nucleon potentials [@PFGA80; @PFG80; @CPFG85; @FPSS93; @NHKG97]. For solution of the four-nucleon problem one can employ Yakubovsky’s generalization of the Faddeev formalism [@Ya67] as done, e.g., in Ref. [@GH93]. Alternatively, other methods have also been succesfully used in the past, such as, the Green’s function Monte Carlo method [@GFMC] or the correlated hyperspherical harmonics expansion method [@VKR95]. On the other hand, when studying the properties of more complex nuclei one typically resorts to the shell model. In that approach, the harmonic-oscillator basis is used in a truncated model space. Instead of the free nucleon-nucleon potential, one utilizes effective interactions appropriate for the truncated model space. Examples of such calculations are the large-basis no-core shell-model calculations that have recently been performed [@ZBVHS; @NB96]. In these calculations all nucleons are active, which simplifies the effective interaction as no hole states are present. The effective interaction is determined for a system of two nucleons in a harmonic-oscillator well interacting by the nucleon-nucleon potential and is subsequently used in the many-particle calculations. In a recent paper we combined the shell-model approach to the three-nucleon problem with the Faddeev formalism [@NB98]. That allowed us to extend the shell-model calculations to a model space of excitations of $32\hbar\Omega$ above the unperturbed ground state and to study the convergence with respect to the size of the model space. In the present paper we generalize these earlier calculations to the four-nucleon problem. We introduce equations for Faddeev amplitudes that are fully antisymmetrized for three nucleons. As the center-of-mass term is removed, we are able to work in a model space up to an excitation of $14\hbar\Omega$ above the unperturbed ground state. For comparison the largest shell-model calculations so far for $^4$He are those performed by R. Ceuleneer et al., in which a $10\hbar\Omega$ model space was utilized [@CVS88]. The main motivation for the present work is to test the shell-model approach and the effective interactions derived from realistic nucleon-nucleon (NN) potentials that are used in the conventional shell-model applications for more complex systems. As the equations that we employ can be conveniently used with three-body interactions or three-body effective interactions, we investigate, in addition to two-body effective interactions, also three-body effective interactions in the present formalism. Such effective interactions are not typically used in the traditional applications. We show that the inclusion of three-body effective interactions improves the overall convergence of the results. At the same time our work serves as an alternative method to solving the four-nucleon problem. We can study the convergence properties of the results with the increasing size of the model space. If convergence is achieved, our results will approach the exact solutions obtained by other methods. In our formalism we seek simultaneously solutions for both the ground-state and the excited states. In the past, the variational Monte Carlo method was used to investigate the excited states of $^4$He using realistic NN potentials [@CPW84]. In most four-nucleon calculations with realistic NN potentials, however, only the ground-state properties were evaluated [@GH93; @VKR95]. On the other hand, earlier studies that investigated the excited-state properties usually did not employ realistic NN potentials [@CVS88; @CH97]. Recently, the four-nucleon resonant and scattering states were investigated using realistic NN potentials in the framework of the resonating group method [@HH97], the correlated-hyperspherical-harmonics method [@VRK98] as well as in the solution of the Faddeev-Yakubovsky equations in configuration space [@CC98]. The present calculation is simplified by using a compact formula for the matrix elements of the permutation operators in the harmonic-oscillator (HO) basis. Also, because of the way we do the model-space truncation, we keep equivalence of the Faddeev-like and Schrödinger equations throughout the calculation. In addition to calculation of ground-state and excited state energies and point-nucleon rms radii, we also evaluate electromagnetic (EM) and strangeness form factors in the impulse approximation. In section \[sec2\] we first discuss the Faddeev equations for the shell-model problem of three nucleons. Then a generalization to the four-nucleon system is introduced. In section \[sec3\] we present the energy, radii and form factor results for $^4$He. Conclusions are given in section \[sec4\]. Shell model and Faddeev-like formalism {#sec2} ====================================== In shell-model studies the one- plus two-body Hamiltonian for the A-nucleon system, i.e., $$\label{ham} H=\sum_{i=1}^A \frac{\vec{p}_i^2}{2m}+\sum_{i<j}^A V_{\rm N}(\vec{r}_i-\vec{r}_j) \; ,$$ where $m$ is the nucleon mass and $V_{\rm N}(\vec{r}_i-\vec{r}_j)$, the NN interaction, is usually modified by adding the center-of-mass HO potential $\frac{1}{2}Am\Omega^2 \vec{R}^2$, $\vec{R}=\frac{1}{A}\sum_{i=1}^{A}\vec{r}_i$. This potential does not influence intrinsic properties of the many-body system. It provides, however, a mean field felt by each nucleon and allows us to work with a convenient HO basis. The modified Hamiltonian, depending on the HO frequency $\Omega$, can be cast into the form $$\label{hamomega} H^\Omega=\sum_{i=1}^A \left[ \frac{\vec{p}_i^2}{2m} +\frac{1}{2}m\Omega^2 \vec{r}^2_i \right] + \sum_{i<j}^A \left[ V_{\rm N}(\vec{r}_i-\vec{r}_j) -\frac{m\Omega^2}{2A} (\vec{r}_i-\vec{r}_j)^2 \right] \; .$$ The one-body term of the Hamiltonian (\[hamomega\]) can be re-written as a sum of the center-of-mass term, $H^\Omega_{\rm cm}=\frac{\vec{P}_{\rm cm}^2}{2Am} +\frac{1}{2}Am\Omega^2 \vec{R}^2$, where $\vec{P}_{\rm cm}=\sum_{i=1}^A \vec{p}_i$, and a term depending only on the relative coordinates. In the present application we use a basis, which explicitly separates center-of-mass and relative-coordinate wave functions. Therefore, the contribution of the center-of-mass term is trivial and will be omitted from now on. The shell-model calculations are performed in a finite model space. Therefore, the interaction term in Eq. (\[hamomega\]) must be replaced by an effective interaction. In general, for an $A$-nucleon system, an $A$-body effective interaction is needed. In practice, the effective interaction is usually approximated by a two-body effective interaction. In the present study we will also employ a three-body effective interaction. As approximations are involved in the effective interaction treatment, large model spaces are desirable. In that case, the calculation should be less affected by any imprecision of the effective interaction. The same is true for the evaluation of any observable characterized by an operator. In the model space, renormalized effective operators are required. The larger the model space, the less renormalization is needed. We may take advantage of the present approach to perform shell-model calculations in significantly larger model spaces than are possible in the conventional shell-model approach. At the same time we can investigate convergence properties of effective interactions. Three-nucleon system -------------------- In this subsection we repeat the steps discussed in Ref. [@NB98] that are needed to solve the three-nucleon shell-model problem in the Faddeev formalism. For a three-nucleon system, i.e., $A=3$, the following transformation of the coordinates \[rtrans\] $$\begin{aligned} \vec{r}&=&\sqrt{\textstyle{\frac{1}{2}}}(\vec{r}_1-\vec{r}_2) \; , \\ \vec{y}&=&\sqrt{\textstyle{\frac{2}{3}}}[\textstyle{\frac{1}{2}}(\vec{r}_1 +\vec{r}_2)-\vec{r}_3] \; ,\end{aligned}$$ and, similarly, of the momenta, can be introduced that brings the relative-coordinate part of the one-body HO Hamiltonian into the form $$\label{H0} H_0 = \frac{\vec{p}^2}{2m} + \frac{1}{2}m\Omega^2 \vec{r}^2 + \frac{\vec{q}^2}{2m} + \frac{1}{2}m\Omega^2 \vec{y}^2 \; .$$ Eigenstates of this Hamiltonian, $$\label{hobas} |n l s j t, {\cal N} {\cal L} {\cal J}, J T \rangle \; ,$$ are then used as the basis for the three-nucleon calculation. Here $n, l$ and ${\cal N}, {\cal L}$ are the HO quantum numbers corresponding to the harmonic oscillators associated with the coordinates and momenta $\vec{r}, \vec{p}$ and $\vec{y}, \vec{q}$, respectively. The quantum numbers $s,t,j$ describe the spin, isospin and angular momentum of the relative-coordinate partial channel of particles 1 and 2, while ${\cal J}$ is the angular momentum of the third particle relative to the center of mass of particles 1 and 2. The $J$ and $T$ are the total angular momentum and the total isospin, respectively. The Faddeev equation for the bound system can be written in the form $$\label{Fadeq} \tilde{H}|\phi\rangle = E|\phi\rangle \; ,$$ with $$\label{Fadham} \tilde{H}= H_0 + V(\vec{r}) {\cal T} \; .$$ Here, $V(\vec{r})=V_{\rm N}(\sqrt{2}\vec{r})-\textstyle{\frac{1}{A}}m \Omega^2 \vec{r}^2$ is the potential and ${\cal T}$ is given by $$\label{metric} {\cal T}=1+{\cal T}^{(-)}+{\cal T}^{(+)} \; ,$$ with ${\cal T}^{(+)}$ and ${\cal T}^{(-)}$ the cyclic and the anticyclic permutation operators, respectively. Previously [@NB98], we derived a simple formula for the matrix elements of ${\cal T}^{(-)}+{\cal T}^{(+)}$ in the basis (\[hobas\]), namely $$\begin{aligned} \label{t13t23} &&\langle n_1 l_1 s_1 j_1 t_1, {\cal N}_1 {\cal L}_1 {\cal J}_1, J T | {\cal T}^{(-)}+{\cal T}^{(+)} | n_2 l_2 s_2 j_2 t_2, {\cal N}_2 {\cal L}_2 {\cal J}_2, J T\rangle = - \delta_{N_1,N_2} \nonumber \\ && \times \sum_{LS} \hat{L}^2 \hat{S}^2 \hat{j}_1 \hat{j}_2 \hat{\cal J}_1 \hat{\cal J}_2 \hat{s}_1 \hat{s}_2 \hat{t}_1 \hat{t}_2 (-1)^L \left\{ \begin{array}{ccc} l_1 & s_1 & j_1 \\ {\cal L}_1 & \textstyle{\frac{1}{2}} & {\cal J}_1 \\ L & S & J \end{array}\right\} \left\{ \begin{array}{ccc} l_2 & s_2 & j_2 \\ {\cal L}_2 & \textstyle{\frac{1}{2}} & {\cal J}_2 \\ L & S & J \end{array}\right\} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & \textstyle{\frac{1}{2}} & s_1 \\ \textstyle{\frac{1}{2}} & S & s_2 \end{array}\right\} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & \textstyle{\frac{1}{2}} & t_1 \\ \textstyle{\frac{1}{2}} & T & t_2 \end{array}\right\} \nonumber \\ &&\times \left[(-1)^{s_1+s_2+t_1+t_2-{\cal L}_1-l_1} \langle {\cal N}_1 {\cal L}_1 n_1 l_1 L | n_2 l_2 {\cal N}_2 {\cal L}_2 L \rangle_{\rm 3} +\langle n_1 l_1 {\cal N}_1 {\cal L}_1 L | {\cal N}_2 {\cal L}_2 n_2 l_2 L \rangle_{\rm 3} \right] \; ,\end{aligned}$$ where $N_i=2n_i+l_i+2{\cal N}_i+{\cal L}_i, i\equiv 1,2$, $\hat{j}=\sqrt{2j+1}$ and $\langle {\cal N}_1 {\cal L}_1 n_1 l_1 L | n_2 l_2 {\cal N}_2 {\cal L}_2 L \rangle_{\rm 3}$ is the general HO bracket for two particles with mass ratio 3 as defined, e.g., in Ref. [@Tr72]. The expression (\[t13t23\]) can be derived by examining the action of ${\cal T}^{(+)}$ and ${\cal T}^{(-)}$ on the basis states (\[hobas\]). A similar derivation for a different basis is described, e.g., in Refs. [@HKT72; @Glo83]. Let us note that it follows from the antisymmetry of the two-nucleon states and from the symmetry properties of the HO brackets that the contributions of ${\cal T}^{(-)}$ and ${\cal T}^{(+)}$ in (\[t13t23\]) are identical. The eigensystem of the operator ${\cal T}$ (\[metric\]) consists of two subspaces. The first subspace has eigenstates with the eigenvalue 3, which form totally antisymmetric physical states. The second subspace has eigenstates with the eigenvalue 0, which form a not completely antisymmetric, unphysical subspace of states. We found these properties of ${\cal T}$ by direct calculation using the relation (\[t13t23\]). It is, however, a general result. The same structure of eigenstates was also obtained in Ref. [@RY95] using a different basis. The eigenvalue structure follows from the fact that $\frac{1}{3}{\cal T}$ has the properies of a projection operator. It is possible to hermitize the Hamiltonian (\[Fadham\]) on the physical subspace, where it is quasi-Hermitian. The Hermitized Hamiltonian takes the form $$\label{Fadhamh} \bar{H}= H_0 + \bar{{\cal T}}^{1/2}V(\vec{r})\bar{{\cal T}}^{1/2} \; ,$$ where $\bar{{\cal T}}$ operates on the physical subspace only. The operator ${\cal T}$ (\[metric\]) is diagonal in $N=2n+l+2{\cal N}+{\cal L}$. Note that any basis truncation other than one of the type $N\le N_{\rm max}$ violates, in general, the Pauli principle and mixes physical and unphysical states. Here, $N_{\rm max}$ characterizes the maximum of total allowed HO quanta in the model space and is an input parameter of the calculation. The truncation into totally allowed oscillator quanta $N\le N_{\rm max}$, however, preserves the equivalence of the Hamiltonians (\[Fadham\]) and (\[Fadhamh\]) on the physical subspace. Four-nucleon system ------------------- By relying on the results obtained for the three-nucleon system, as described in the previous subsection, we can extend the formalism to the four-nucleon system. We use the Hamiltonian (\[hamomega\]) with $A=4$. By introducing the coordinate (and momentum) transformations, \[fourtran\] $$\begin{aligned} \vec{r}&=&\sqrt{\textstyle{\frac{1}{2}}}(\vec{r}_1-\vec{r}_2) \;, %\nonumber \\ \vec{y}&=&\sqrt{\textstyle{\frac{2}{3}}}[\textstyle{\frac{1}{2}}(\vec{r}_1 +\vec{r}_2)-\vec{r}_3] \;, %\nonumber \\ \vec{z}&=&\textstyle{\frac{\sqrt{3}}{2}}[\textstyle{\frac{1}{3}}(\vec{r}_1 +\vec{r}_2+\vec{r}_3)-\vec{r}_4] \;, %\nonumber\end{aligned}$$ we obtain the one-body part of the Hamiltonian (\[hamomega\]) in the form $$H_0 = \frac{\vec{p}^2}{2m} + \frac{1}{2}m\Omega^2 \vec{r}^2 + \frac{\vec{q}^2}{2m} + \frac{1}{2}m\Omega^2 \vec{y}^2 + \frac{\vec{o}^2}{2m} + \frac{1}{2}m\Omega^2 \vec{z}^2 \; ,$$ with the center-of-mass term omitted. A possible generalization of the Faddeev equation (\[Fadeq\]) for four identical particles can be written in the form $$\label{fadfour} \tilde{H}|\psi_{(123)4}\rangle = E|\psi_{(123)4}\rangle \; ,$$ with $$\tilde{H}|\psi_{(123)4}\rangle \equiv H_0 |\psi_{(123)4}\rangle + \textstyle{\frac{1}{2}}(V_{12}+V_{13}+V_{23}) (|\psi_{(123)4}\rangle +|\psi_{(432)1}\rangle +|\psi_{(134)2}\rangle +|\psi_{(142)3}\rangle )\; ,$$ and $$\label{metric4} (|\psi_{(123)4}\rangle +|\psi_{(432)1}\rangle +|\psi_{(134)2}\rangle +|\psi_{(142)3}\rangle ) = (1-{\cal T}_{14}-{\cal T}_{24}-{\cal T}_{34})|\psi_{(123)4}\rangle \equiv {\cal T}_4 |\psi_{(123)4}\rangle \; .$$ Here, $|\psi_{(123)4}\rangle $ is a four-fermion Faddeev amplitude completely antisymmetrized for particles 1,2, and 3. There are three other equations that can be obtained from Eq. (\[fadfour\]) by permuting particle 4 with particles 1, 2, and 3. Their sum then leads to the Schrödinger equation. We note that the present equations are different from the traditional Faddeev-Yakubovsky equations [@Ya67], which combine Faddeev amplitudes depending on two sets of relative coordinates. We are working with a complete orthonormal basis. It is, therefore, sufficient and convenient to use a single set of coordinates defined by the relations (\[fourtran\]). Unlike the Faddeev amplitudes used typically in the Faddeev-Yakubovsky equations, the amplitudes appearing in Eq. (\[fadfour\]) are antisymmetrized with respect to the first three particles. Those amplitudes are obtained, as described below, in a straightforward manner with the help of our three-nucleon HO formalism introduced earlier. The present equations allow us to employ easily real three-body interactions or three-body effective interactions. The latter property makes them particularly useful for the present extension of shell-model calculations for four nucleons. At the same time, the use of Faddeev amplitudes antisymmetrized for particles 1, 2 and 3 allows us to reduce the dimmension of the basis significantly. We start the four-nucleon calculation using the basis $$\label{fourbas} |N_1 i J_1 T_1, n_z l_z {\cal J}_4, J T \rangle \; .$$ with the three-fermion part given by the antisymmetrized eigenstates of ${\cal T}$ (\[metric\]) corresponding to eigenvalue 3, e.g., $$\label{threebant} |N_1 i J_1 T_1\rangle = \sum c_{n l s j t {\cal N} {\cal L} {\cal J}_{3}}^{N_{1 } i J_{ 1} T_{ 1}} |n l s j t, {\cal N} {\cal L} {\cal J}_3, J_1 T_1\rangle \; ,$$ where $N_1=2n+l+2{\cal N}+{\cal L}$ and $i$ counts the eigenstates of ${\cal T}$ with the eigenvalue 3 for given $N_1$ and $J_1,T_1$. Further, $n_z, l_z$ are the HO quantum numbers corresponding to the harmonic oscillator associated with the coordinate $\vec{z}$ and the momentum $\vec{o}$ and ${\cal J}_4$ is the angular momentum of the fourth particle relative to the center of mass of particles 1, 2 and 3. As in the case of the three-particle transposition operators (\[t13t23\]), a compact formula can be derived for the matrix elements of the four-particle transposition operators in the basis (\[fourbas\]), e.g., $$\begin{aligned} \label{t4} &&\langle N_{\rm 1L} i_{\rm L} J_{\rm 1L} T_{\rm 1L}, n_{z \rm L} l_{z \rm L} {\cal J}_{\rm 4L}, JT | {\cal T}_{14} +{\cal T}_{24}+{\cal T}_{34} |N_{\rm 1R} i_{\rm R} J_{\rm 1R} T_{\rm 1R}, n_{z \rm R} l_{z \rm R} {\cal J}_{\rm 4R}, JT \rangle \nonumber \\ &&=\delta_{N_{\rm L},N_{\rm R}} \sum c_{n_{\rm L} l_{\rm L} s_{\rm L} j_{\rm L} t_{\rm L} {\cal N}_{\rm L} {\cal L}_{\rm L} {\cal J}_{3\rm L}}^{N_{1 \rm L} i_{\rm L} J_{\rm 1L} T_{\rm 1L}} c_{n_{\rm R} l_{\rm R} s_{\rm R} j_{\rm R} t_{\rm R} {\cal N}_{\rm R} {\cal L}_{\rm R} {\cal J}_{3\rm R}}^{N_{1 \rm R} i_{\rm R} J_{\rm 1R} T_{\rm 1R}} \hat{L}_{\rm 1L}^2 \hat{L}_{\rm 1R}^2 \hat{S}_{\rm 1L}^2 \hat{S}_{\rm 1R}^2 \hat{L}_2^2 \hat{S}_2^2 \nonumber \\ &&\times \hat{j}_{\rm L} \hat{j}_{\rm R} \hat{\cal J}_{\rm 3L} \hat{\cal J}_{\rm 3R} \hat{\cal J}_{\rm 4L} \hat{\cal J}_{\rm 4R}\hat{J}_{\rm 1L}\hat{J}_{\rm 1R} \hat{T}_{\rm 1L}\hat{T}_{\rm 1R} (-1)^{T_{\rm 1L}-T_{\rm 1R}+S_{\rm 1L}+S_{\rm 1R}} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & s_{\rm R} & S_{\rm 1R} \\ \textstyle{\frac{1}{2}} & S_2 & S_{\rm 1L} \end{array}\right\} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & t_{\rm R} & T_{\rm 1R} \\ \textstyle{\frac{1}{2}} & T & T_{\rm 1L} \end{array}\right\} \nonumber \\ &&\times \left\{ \begin{array}{ccc} l_{\rm L} & s_{\rm L} & j_{\rm L} \\ {\cal L}_{\rm L} & \textstyle{\frac{1}{2}} & {\cal J}_{\rm 3L} \\ L_{\rm 1L} & S_{\rm 1L} & J_{\rm 1L} \end{array}\right\} \left\{ \begin{array}{ccc} l_{\rm R} & s_{\rm R} & j_{\rm R} \\ {\cal L}_{\rm R} & \textstyle{\frac{1}{2}} & {\cal J}_{\rm 3R} \\ L_{\rm 1R} & S_{\rm 1R} & J_{\rm 1R} \end{array}\right\} \left\{ \begin{array}{ccc} L_{\rm 1L} & S_{\rm 1L} & J_{\rm 1L} \\ l_{z\rm L} & \textstyle{\frac{1}{2}} & {\cal J}_{\rm 4L} \\ L_2 & S_2 & J \end{array}\right\} \left\{ \begin{array}{ccc} L_{\rm 1R} & S_{\rm 1R} & J_{\rm 1R} \\ l_{z\rm R} & \textstyle{\frac{1}{2}} & {\cal J}_{\rm 4R} \\ L_2 & S_2 & J \end{array}\right\} \nonumber \\ &&\times \hat{L}'^2 (-1)^{L'} \left\{ \begin{array}{ccc} l_{\rm R} & L_2 & L' \\ l_{z\rm R} & {\cal L}_{\rm R} & L_{\rm 1R} \end{array}\right\} \left\{ \begin{array}{ccc} l_{\rm R} & L_2 & L' \\ l_{z\rm L} & l' & L_{\rm 1L} \end{array}\right\} \left[ \hat{s}_{\rm L} \hat{s}_{\rm R}\hat{t}_{\rm L} \hat{t}_{\rm R} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & \textstyle{\frac{1}{2}} & s_{\rm R} \\ \textstyle{\frac{1}{2}} & S_{\rm 1L} & s_{\rm L} \end{array}\right\} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & \textstyle{\frac{1}{2}} & t_{\rm R} \\ \textstyle{\frac{1}{2}} & T_{\rm 1L} & t_{\rm L} \end{array}\right\} \right. \nonumber \\ &&\times (-1)^{l_{z\rm R}+L_{\rm 1L}} \left( (-1)^{l_{z\rm L}} \langle n' l' n_{z\rm L} l_{z\rm L} L' | n_{z\rm R} l_{z\rm R} {\cal N}_{\rm R} {\cal L}_{\rm R} L' \rangle_{\rm 8} \langle n_{\rm L} l_{\rm L} {\cal N}_{\rm L} {\cal L}_{\rm L} L_{\rm 1L} | n' l' n_{\rm R} l_{\rm R} L_{\rm 1L} \rangle_{\rm 3} \right. \nonumber \\ &&\left. +(-1)^{t_{\rm R}-t_{\rm L}+s_{\rm R}-s_{\rm L} +{\cal L}_{\rm R} -l_{\rm L}-{\cal L}_{\rm L}} \langle n_{z\rm L} l_{z\rm L} n' l' L' | {\cal N}_{\rm R} {\cal L}_{\rm R} n_{z\rm R} l_{z\rm R} L' \rangle_{\rm 8} \langle {\cal N}_{\rm L} {\cal L}_{\rm L} n_{\rm L} l_{\rm L} L_{\rm 1L} | n_{\rm R} l_{\rm R} n' l' L_{\rm 1L} \rangle_{\rm 3}\right) \nonumber \\ &&\left. - \delta_{l_{\rm L},l_{\rm R}}\delta_{s_{\rm L},s_{\rm R}} \delta_{t_{\rm L},t_{\rm R}}\delta_{{\cal N}_{\rm L},n'} \delta_{{\cal L}_{\rm L},l'} (-1)^{{\cal L}_{\rm R}+l_{z\rm R}} \langle n_{z\rm L} l_{z\rm L} {\cal N}_{\rm L} {\cal L}_{\rm L} L' | {\cal N}_{\rm R} {\cal L}_{\rm R} n_{z\rm R} l_{z\rm R} L' \rangle_{\rm 8} \right] \; ,\end{aligned}$$ where $N_{\rm X}=2n_{\rm X}+l_{\rm X}+2{\cal N}_{\rm X}+{\cal L}_{\rm X} +2n_{z\rm X}+l_{z\rm X}, {\rm X}\equiv {\rm L \; or \; R}$, and, e.g., the expression $\langle n_{z\rm L} l_{z\rm L} {\cal N}_{\rm L} {\cal L}_{\rm L} L' | {\cal N}_{\rm R} {\cal L}_{\rm R} n_{z\rm R} l_{z\rm R} L' \rangle_{\rm 8}$ denotes a general HO bracket for two particles with mass ratio 8, as defined in Ref. [@Tr72]. Similarly, as in Eq. (\[t13t23\]) the brackets for two particles with mass ratio 3 also appear in the relation (\[t4\]). In the derivation of the expression (\[t4\]) we relied on the antisymmetry of the basis states with respect to particles 1,2 and 3. The calculation was facilitated by application of the operators $-{\cal T}_{13}$ and $-{\cal T}_{23}$. The relation (\[t4\]) appears to be non-symmetric. However, its numerical evaluation leads to a symmetric matrix. It may also appear that the angular momentum sums in (\[t4\]) can be summed up. In fact, it is possible to simplify the expression by introducing a $15j$-coefficient of the fifth kind as defined, e.g., in Ref. [@YLV60], but, as such coefficients are seldomly used, we prefer to keep the summations in the explicit form. On the other hand, a significant simplification of the expression (\[t4\]) can be obtained, when the symmetry relations of different terms are exploited. First, it follows from the properties of the HO brackets and from the antisymmetry of the two-nucleon states that the contributions of ${\cal T}_{14}$ and ${\cal T}_{24}$ are identical. Second, using the fact that the states (\[fourbas\]) are antisymmetrized for the particles 1, 2 and 3 it follows that all three permutation operators appearing in (\[t4\]) give identical contributions to the expression (\[t4\]). The computation of ${\cal T}_{34}$ is the simplest. In that case a partial summation of the angular momentum coefficients can be performed, yielding a compact expression $$\begin{aligned} \label{t34} &&\langle N_{\rm 1L} i_{\rm L} J_{\rm 1L} T_{\rm 1L}, n_{z \rm L} l_{z \rm L} {\cal J}_{\rm 4L}, JT | {\cal T}_{34} |N_{\rm 1R} i_{\rm R} J_{\rm 1R} T_{\rm 1R}, n_{z \rm R} l_{z \rm R} {\cal J}_{\rm 4R}, JT \rangle \nonumber \\ &&=\delta_{N_{\rm L},N_{\rm R}} \sum c_{n l s j t {\cal N}_{\rm L} {\cal L}_{\rm L} {\cal J}_{3\rm L}}^{N_{1 \rm L} i_{\rm L} J_{\rm 1L} T_{\rm 1L}} c_{n l s j t {\cal N}_{\rm R} {\cal L}_{\rm R} {\cal J}_{3\rm R}}^{N_{1 \rm R} i_{\rm R} J_{\rm 1R} T_{\rm 1R}} \nonumber \\ &&\times \hat{\cal J}_{\rm 3L} \hat{\cal J}_{\rm 3R} \hat{\cal J}_{\rm 4L} \hat{\cal J}_{\rm 4R}\hat{J}_{\rm 1L}\hat{J}_{\rm 1R} \hat{T}_{\rm 1L}\hat{T}_{\rm 1R} (-1)^{T_{\rm 1L}+T_{\rm 1R}+{\cal J}_{\rm 3L}+{\cal J}_{\rm 3R}} \left\{ \begin{array}{ccc} \textstyle{\frac{1}{2}} & t & T_{\rm 1R} \\ \textstyle{\frac{1}{2}} & T & T_{\rm 1L} \end{array}\right\} \nonumber \\ &&\times \hat{K}^2 \left\{ \begin{array}{ccc} j & {\cal J}_{\rm 3L} & J_{\rm 1L} \\ {\cal J}_{\rm 3R} & K & {\cal J}_{\rm 4L} \\ J_{\rm 1R} & {\cal J}_{\rm 4R} & J \end{array}\right\} \left\{ \begin{array}{ccc} {\cal L}_{\rm L} & l_{\rm zR} & K \\ {\cal J}_{\rm 4R} & {\cal J}_{\rm 3L}& \textstyle{\frac{1}{2}} \end{array}\right\} \left\{ \begin{array}{ccc} {\cal L}_{\rm R} & l_{\rm zL} & K \\ {\cal J}_{\rm 4L} & {\cal J}_{\rm 3R}& \textstyle{\frac{1}{2}} \end{array}\right\} \left\{ \begin{array}{ccc} l_{\rm zL} & {\cal L}_{\rm R} & K \\ l_{\rm zR} & {\cal L}_{\rm L} & L' \end{array}\right\} \nonumber \\ &&\times \hat{L}'^2 (-1)^{{\cal L}_{\rm R}+l_{z\rm L}+L'} \langle n_{z\rm L} l_{z\rm L} {\cal N}_{\rm L} {\cal L}_{\rm L} L' | {\cal N}_{\rm R} {\cal L}_{\rm R} n_{z\rm R} l_{z\rm R} L' \rangle_{\rm 8} \; .\end{aligned}$$ Thus, by multiplying the expression (\[t34\]) by three we obtain the same matrix element as from (\[t4\]). We note that a generalization of the evaluation of the permutation operator matrix element (\[t34\]) to a more complex system, than the presently studied $A=4$, is straightforward. Its simplicity suggests that the present formalism can be extended to systems with $A>4$. Similarly, as for the operator ${\cal T}$ (\[metric\]), eigenstates of the operator ${\cal T}_4$ defined by the relation (\[metric4\]) can be subdivided into two subspaces. A physical subspace is spanned by totally antisymmetric states, in this case corresponding to the eigenvalue 4, and a spurious subspace is spanned by eigenvectors corresponding to the eigenvalue 0. It is possible to symmetrize the Hamiltonian $\tilde{H}$ appearing in Eq. (\[fadfour\]) on the physical subspace. The symmetrized Hamiltonian then takes the form $$\label{foureq} \bar{H}= H_0 + \bar{{\cal T}_4}^{1/2}\textstyle{\frac{1}{2}} (V_{12}+V_{13}+V_{23})\bar{{\cal T}_4}^{1/2} \; ,$$ where $\bar{\cal T}_4$ operates only on the physical subspace. In our calculations, described later, we diagonalize the symmetrized Hamiltonian (\[foureq\]) in the physical basis formed by the eigenstates of $\bar{\cal T}_4$. The operator ${\cal T}_4$ (\[metric4\]) is diagonal in $N=2n+l+2{\cal N}+{\cal L}+2n_z+l_z$. A basis truncation defined by a restriction on the totally allowed oscillator quanta $N\le N_{\rm max}$ preserves the equivalence of the Hamiltonians (\[Fadham\]) and (\[Fadhamh\]) on the physical subspace. Effective interactions ---------------------- From solving two-nucleon systems in a HO well, interacting by soft-core potentials, one learns that excitations up to about $300\hbar\Omega$ ($N_{\rm max}=300$) are required to get almost exact solutions. We anticipate, therefore, that at least the same number of excitations should be allowed to solve the many-nucleon system. The Faddeev formulation has the obvious advantage compared with the traditional shell-model approach that the center-of-mass coordinate is explicitly removed. Even then, it is not feasible to solve the eigenvalue problem either for (\[Fadhamh\]) or for (\[foureq\]) in such a large space. On the other hand, shell-model calculations are always performed by employing effective interactions tailored to a specific model space. In practice, these effective interactions can never be calculated exactly, because, in general, for an A-body effective interaction is required for an A-nucleon system. We may, however, take advantage of the present approach to perform shell-model calculations in significantly larger model spaces than are possible in conventional shell-model approach. At the same time we can investigate convergence properties of effective interactions. If convergence is achieved, we should obtain the exact solution, since by construction the effective interactions that we employ satisfy the condition $V_{\rm eff}\rightarrow V$ for $N_{\rm max}\rightarrow\infty$. Usually, the effective interaction is approximated by a two-body effective interaction determined from a two-nucleon system. In the present calculations we replace matrix elements of the potential $V(\vec{r})$ by matrix elements of an effective two-body interaction, derived in a straightforward manner for each relative-coordinate partial channel. The relevant two-nucleon Hamiltonian is then $$\label{hamomega2} H_2\equiv H_{02}+V= \frac{\vec{p}^2}{2m} +\frac{1}{2}m\Omega^2 \vec{r}^2 + V_{\rm N}(\sqrt{2}\vec{r})-\frac{m\Omega^2}{A}\vec{r}^2 \; ,$$ which can be solved as a differential equation or, alternatively, can be diagonalized in a sufficiently large harmonic oscillator basis. For a four-nucleon system we set $A=4$ in Eq. (\[hamomega2\]), which implies that we are dealing with a bound-state problem. To construct the two-body effective interaction we employ the Lee-Suzuki [@LS80] similarity transformation method, which gives the effective interaction in the form $P V_{\rm eff}P = P V P + PV Q\omega P$, with $\omega$ the transformation operator satisfying $\omega=Q \omega P$, and $P$ and $Q=1-P$, the projectors on the model and the complementary spaces, respectively. Our calculations start with exact solutions of the Hamiltonian (\[hamomega2\]) and, consequently, we construct the operator $\omega$ and, then, the effective interaction directly from these solutions. Let us denote the relative-coordinate two-nucleon HO states, which form the model space, as $|\alpha_P\rangle$, and those which belong to the Q-space, as $|\alpha_Q\rangle$. Then the Q-space components of the eigenvector $|k\rangle$ of the Hamiltonian (\[hamomega2\]) can be expressed as a combination of the P-space components with the help of the operator $\omega$ $$\label{eigomega} \langle\alpha_Q|k\rangle=\sum_{\alpha_P} \langle\alpha_Q|\omega|\alpha_P\rangle \langle\alpha_P|k\rangle \; .$$ If the dimension of the model space is $d_P$, we may choose a set ${\cal K}$ of $d_P$ eigenevectors, for which the relation (\[eigomega\]) will be satisfied. Under the condition that the $d_P\times d_P$ matrix $\langle\alpha_P|k\rangle$ for $|k\rangle\in{\cal K}$ is invertible, the operator $\omega$ can be determined from (\[eigomega\]). In the present application we select the lowest states obtained in each channel. Their number is given by the number of basis states satisfying $2n+l\leq N_{\rm max}$. Once the operator $\omega$ is determined, the effective hamiltonian can be constructed as follows $$\label{effomega} \langle \gamma_P|H_{2\rm eff}|\alpha_P\rangle =\sum_{k\in{\cal K}} \left[ \langle\gamma_P|k\rangle E_k\langle k|\alpha_P\rangle +\sum_{\alpha_Q}\langle\gamma_P|k\rangle E_k\langle k|\alpha_Q\rangle \langle\alpha_Q |\omega|\alpha_P\rangle\right] \; .$$ It should be noted that $P|k\rangle=\sum_{\alpha_P}|\alpha_P\rangle\langle\alpha_P|k\rangle$ for $|k\rangle\in{\cal K}$ is a right eigenvector of (\[effomega\]) with the eigenvalue $E_k$. This Hamiltonian, when diagonalized in a model-space basis, reproduces exactly the set ${\cal K}$ of $d_P$ eigenvalues $E_k$. Note that the effective Hamiltonian is, in general, quasi-Hermitian. It can be hermitized by a similarity transformation determined from the metric operator $P(1+\omega^\dagger\omega)P$. The Hermitian Hamiltonian is then given by [@S82SO83] $$\label{hermeffomega} \bar{H}_{\rm 2eff} =\left[P(1+\omega^\dagger\omega)P\right]^{1/2} H_{\rm 2eff}\left[P(1+\omega^\dagger\omega) P\right]^{-1/2} \; .$$ Finally, the two-body effective interaction used in the present calculations is determined from the two-nucleon effective Hamiltonian (\[hermeffomega\]) as $V_{\rm 2eff}=\bar{H}_{\rm 2eff}-H_{02}$. We note that the interaction $V_{12}+V_{13}+V_{23}$ in Eq. (\[foureq\]) is then replaced by ${\cal T}^{1/2}V_{\rm 2eff}{\cal T}^{1/2}$, which is evaluated in a straightforward way in the basis (\[fourbas\]). As pointed out before, the structure of the Hamiltonian (\[foureq\]) allows us to employ easily three-body effective interactions in addition to the above discussed two-body effective interactions. We can replace $V_{12}+V_{13}+V_{23}$ in Eq. (\[foureq\]) by $V_{\rm 3eff}$ that can be derived from the three-nucleon solutions in a similar manner as the two-body effective interaction is derived from the two-nucleon solutions. To find $V_{\rm 3eff}$ we solve the three-nucleon system described by the Hamiltonian (\[Fadhamh\]) with $V(\vec{r})=V_{\rm N}(\sqrt{2}\vec{r})-\textstyle{\frac{1}{A}}m \Omega^2 \vec{r}^2$. As $A=4$ we are dealing with a bound three-nucleon problem. It can be solved in a three-nucleon model space characterized by $N_{\rm 3max}\approx 30$ [@NB98]. First, we compute the two-body effective interaction appropriate for the model space defined by $N_{\rm 3max}$, as discussed earlier in this subsection. Then the three-nucleon system is solved in the same space. Afterwards we construct the three-body effective interaction for a model space defined by $N_{\rm max}<N_{\rm 3max}$. In the present paper we use model spaces up to $N_{\rm max}=14$. The effective interaction is constructed exactly, as described above, using Eqs. (\[eigomega\],\[effomega\],\[hermeffomega\]) with $H_{\rm 2eff}$ replaced by $H_{\rm 3eff}$. The energies $E_k$ and the states $|k\rangle$ correspond to the three-nucleon system eigenstates, however, and the states $|\alpha_P\rangle$ and $|\alpha_Q\rangle$ are three-nucleon basis states (\[threebant\]) with the model-space condition $N_1\equiv 2n+l+2{\cal N}+{\cal L}\leq N_{\rm max}$ and the $Q$-space condition $N_{\rm max}<N_1\leq N_{\rm 3max}$. The three-body effective interaction is computed for different three-nucleon channels characterized by $J_1, T_1$ and parity and is obtained from the hermitized effective Hamiltonian as $V_{\rm 3eff}=\bar{H}_{\rm 3eff}-H_{0}$, where $H_0$ is given by Eq. (\[H0\]). The interaction $V_{\rm 3eff}$ then replaces $V_{12}+V_{13}+V_{23}$ in Eq. (\[foureq\]). We note that by construction in the limit $N_{\rm max}\rightarrow N_{\rm 3max}$ the three-body effective interaction approaches the two-body effective interaction $V_{\rm 3eff}\rightarrow {\cal T}^{1/2}V_{\rm 2eff}{\cal T}^{1/2}$ and with $N_{\rm max}\rightarrow \infty$ the effective interaction approaches the bare interaction $V_{\rm 2eff}\rightarrow V$. Application to $^4$He {#sec3} ===================== In the present paper we use the Reid93 NN potential [@SKTS] and the Argonne V8’ NN potential, introduced in Ref. [@GFMC]. We work in the isospin formalism; the charge invariant potential $V_{\rm N} =\frac{1}{3} V_{pp}+\frac{1}{3} V_{nn}+\frac{1}{3} V_{np}$ is used for each $T=1$ wave in the calculations with the Reid93 potential. The Coulomb potential is added to $V_{pp}$ in this case. On the other hand, the calculations with the Argonne V8’ potential, which is isopin invariant, do not include the Coulomb potential. Our calculation progresses in several steps. The model space is characterized by the condition $N\le N_{\rm max}$, $N=2n+l+2{\cal N}+{\cal L}+2n_z+l_z$. First, the three-nucleon antisymmetrized basis is constructed by diagonalizing ${\cal T}$ (\[metric\]) in the basis (\[hobas\]) for all $N_1\equiv 2n+l+2{\cal N}+{\cal L}\le N_{\rm max}$ and all $J_1, T_1$. Then the four-nucleon antisymmetrized basis is calculated by diagonalizing ${\cal T}_4$ (\[metric4\]) in the basis (\[fourbas\]) for $N=N_1+2n_z+l_z\le N_{\rm max}$ with $N$ even for positive parity states and $N$ odd for negative parity states. We present results for $J=0$ and $T=0$ only, but for both parities. We note that the four-nucleon basis computation is independent of $\Omega$ and is done only once. The next step is the effective interaction calculation. The two-body effective interaction is derived from the Eqs. (\[eigomega\])-(\[hermeffomega\]). The condition for the relative-coordinate two-body effective-interaction model space is then $2n+l\le N_{\rm max}$. When solving the two-nucleon relative-coordinate Hamiltonian (\[hamomega2\]) in the full space, we truncate the HO basis by keeping the states typically up to $n = 152$. The two-body effective interaction is constructed for all partial-wave channels up to $j=6$. The resulting effective interaction is finally used as input for the four-nucleon calculation, where the Hamiltonian (\[foureq\]) is diagonalized. Instead of a two-body effective interaction, we may use a three-body effective interaction, as discussed in the previous section. The three-body effective is computed only for the most important three-nucleon channels $J_1 T_1$. In particular, we evaluated the three-body effective interaction for $J_1=1/2, 3/2$, $T_1=1/2$ and for both positive and negative parity. For the channels with higher $J_1$ the two-body effective interaction corresponding to $N_{\rm max}$ is used instead. For the parameter $N_{\rm 3max}$ characterizing the three-nucleon full space, we used $N_{\rm 3max}=28$ for $J_1=1/2$ and $N_{\rm 3max}=24$ for $J_1=3/2$. We also performed calculations with the inclusion of the three-body effective interaction for $J_1=5/2$ and found it to have little effect. Let us remark that the present method for solving the four-nucleon shell-model problem is fully equivalent to the standard shell-model approach. In particular, it is straightforward to transform the relative-coordinate two-body effective interaction used in the present calculations to the two-particle basis used for the shell-model input by the standard transformation [@HG83]. We used the transformed interactions for the model spaces up to an 8$\hbar\Omega$ space to test our results. The shell-model diagonalization was then performed by employing the Many-Fermion-Dynamics Shell-Model Code [@VZ94], which can be utilized for calculations with model spaces comprising up to 9 major HO shells, i.e., $N_{\rm max}=8$ for $^4$He. We obtained the same results from both the Faddev-like calculation and the standard shell-model calculation. The Faddeev-like calculation has, obviously, much smaller dimension and can be extended to larger model spaces. We also note that we applied the discussed formalism to four-electron system in a related study recently [@NBG98]. Our results compared well with those obtained by the Stochastic Variational Method [@VOS]. Energies and point-nucleon rms radii ------------------------------------ Our results for the ground-state and excited-state energies and point-nucleon rms radii are presented in Figs. \[figr93e\]-\[figv8p23\], where the dependencies on the model-space size and the HO energy are shown. A summary of the largest model-space ($N_{\rm max}=14$ for the positive-parity states and $N_{\rm max}=13$ for the negative-parity states) results is given in Table \[tab1\]. Let us mention an unusual feature of the present calculations, namely, the convergence from below for the ground-state energy. It is caused by the asymmetric treatment of the HO terms that are added and subtracted to the Hamiltonian in the process of evaluating the effective interaction. Our effective interaction is computed for a two- or three-nucleon system bound in an HO potential. Therefore, artificial binding from this potential is included in the effective interaction and the four-body effects coming from the entire four-nucleon calculation may not completely compensate for this spurious binding in a particular model space. We note that this type of over-binding in the no-core shell-model calculations was noticed in previous studies [@ZBVC; @ZBVM; @NB96]. This effect decreases as the model-space size increases, as is demonstrated in our earlier three-nucleon shell-model calculations [@NB98]. In Fig. \[figr93e\] we present the calculated dependence of the ground-state energy and the first-excited $0^+ 0$ state energy on the model-space size, characterized by $N_{\rm max}$. The two-body effective interaction employed was derived from the Reid93 NN potential. Results for $\hbar\Omega=14, 17, 19, 22$ MeV are shown. The corresponding dependence of the point-nucleon rms radius is presented in Fig. \[figr93r\]. A slow convergence with the increasing model-space size can be observed for the energies with a significantly faster rate for the ground state compared to the first excited $0^+ 0$ state. Also, much stronger dependence of the excited state on the HO energy $\hbar\Omega$ is apparent. The results of the point-nucleon rms-radius calculation demonstrate even more the differences between the ground state and the first excited $0^+ 0$ state. While the ground-state radius has almost converged and shows little $\hbar\Omega$ dependence, the first excited $0^+ 0$ state displays a strong dependence of its energy on $\hbar\Omega$ and a steady increase of its radius with increasing model-space size. Let us remark that in our approach we obtain the ground state as well as the excited states by diagonalizing the Hamiltonian. This implies that the excited states are expanded in the same harmonic-oscillator basis used for the ground state. While such an approach has technical advantages, it might not be physically sound. Cautious interpretation of the excited-state results is, therefore, necessary. The significantly different convergence rate of the ground state and of the first excited $0^+ 0$ state manifests the different nature of the two states. Let us note that if the model-space size was increased up to the point at which total convergence of the excited state was achieved, our procedure would yield isolated three- and one-body clusters with an infinite rms radius and a total energy of the three-nucleon system. It is possible, though, that we could observe a meta-stability prior to the onset of the cluster separation, as the resonance is sharp and low-lying. The present model-space sizes, however, are not yet sufficient to arrive at that point. That we have not reached this point can be seen from the lack of convergence and, in particular, from the rather small rms radius, which shows a significant increase with $N_{\rm max}$ and a strong dependence on $\Omega$. The importance of the three-body effective interaction can be judged from the results shown in Fig. \[figr93et\]. The ground-state and excited $0^+ 0$ state energies obtained in a calculation that employs the three-body effective interaction is compared to a calculation performed by using only the two-body effective interaction. Results for two different values of the HO energy, $\hbar\Omega=17$ and 19 MeV, are presented. The dashed lines connect the two-body effective interaction calculation results that are identical to those in Fig. \[figr93e\] that correspond to HO energies of $\hbar\Omega=17$ and 19 MeV. The full lines connect the results obtained in calculations with the three-body effective interaction. It is apparent that the three-body effective interaction improves the converegence considerably. It is especially true for the ground state. The difference between the $N_{\rm max}=6$ and $N_{\rm max}=14$ energies is significantly smaller in the calculation that employs the three-body effective interaction. It can also be seen that the two-body effective interaction results approach the three-body effective interaction results in the largest spaces used in our calculations. In addition, the dependence on the HO energy decreases in the three-body effective interaction calculation compared to the two-body effective interaction calculation. This holds for both the ground state and the first excited $0^+ 0$ state. However, the inclusion of the three-body effective interaction clearly has a larger overall impact on the ground-state results. The influence of the three-body effective interaction on the point-nucleon rms radius is depicted in Fig. \[figr93rt\]. Again we observe a better stability of the radii computed using the three-body effective interaction. In particular, the ground-state point-nucleon rms radius shows convergence in both the model-space-size dependence and the HO-frequency dependence. On the other hand, the three-body effective interaction does not improve the convergence of the excited state in any significant way in the model spaces that we employed. In Fig. \[figr93etpm\] we show the calculated energies of the first $0^- 0$ state obtained using two-body effective interactions in model spaces up to $N_{\rm max}=13$. For a comparison, the results for the ground state and the first excited $0^+ 0$ state from Fig. \[figr93et\] are also presented. It is interesting to note that the $0^- 0$ state shows a better convergence and stability with respect to the $N_{\rm max}$ change as well as a weaker dependence on $\hbar\Omega$ than the first excited $0^+ 0$ state. This observation is confirmed also in the point-nucleon rms radius calculation as can be seen in Fig. \[figr93rtpm\]. In the experiment, the $0^- 0$ excitation energy, 21.01 MeV, is higher than the excitation energy of the first $0^+ 0$ state, 20.21 MeV. Though in our calculations their positions are reversed, it is visible from Fig. \[figr93etpm\] that the extrapolation to larger $N_{\rm max}$ leads to correct ordering of the two states. A possible interpretation of this observation is that the excited $0^+ 0$ state is associated with a radial excitation and, thus, it is more sensitive to the HO basis used in our calculations. The energy and radius results, obtained using the Argonne V8’ NN potential, are presented in Figs. \[figv8pe\] and \[figv8pr\], respectively. The three-body effective interaction was used in calculating these results, for three different HO energies, $\hbar\Omega=16, 19$ and 22 MeV. The dotted line represents the value -25.92 MeV obtained for the ground state, using the GFMC [@Pipcom]. Similarly, as in the calculations with the Reid93 NN potential, we get the best convergence for the ground state for the highest value of $\hbar\Omega$, while for the excited state the best results are obtained for the lowest $\hbar\Omega$. The same discussion, given earlier, for the excited-state convergence using the Reid93 NN potential, is also valid for the calculations using the V8’ NN potential. The energy convergence is very slow and there is no sign of convergence of the point-nucleon rms radius of the excited $0^+ 0$ state. A significant dependence on $\hbar\Omega$ prevails for all the model-spaces studied. On the other hand, the ground-state energy shows good convergence and approaches the GFMC result, in particular for the $\hbar\Omega=22$ MeV calculation. The ground-state point-nucleon rms radius is almost $\hbar\Omega$ independent and converged. It agrees with the GFMC value of 1.485 fm. We note that results on the first excited $0^+ 0$ state obtained using the resonating group method were reported in Ref. [@HH97]. The Bonn potential employed in that work gives very similar results for the ground state as those obtained using the Argonne V8’. It is, therefore, reasonable to make a comparison for the excited state results. The first excited $0^+ 0$ state energy reported in Ref. [@HH97] was -6.42 MeV, which is about 10% below our result of $N_{\rm max}=14$ and $\hbar\Omega=16$ MeV calculation. The reported rms radius, 3.02 fm, is slightly above our calculation. In order to further compare the convergence and the $\Omega$ dependence of the results obtained with two- and three-body effective interactions, we present a similar calculation as that of Fig. \[figr93et\], obtained using the Argonne V8’ NN potential and a larger $\Omega$ difference, in Fig. \[figv8p23\]. The full lines correspond to the three-body effective interaction calculations, also shown in Fig. \[figv8pe\], while the dashed lines connect the two-body effective interaction results. Two HO energies of $\hbar\Omega=$16 and 22 MeV were used. The dotted line represents the GFMC result. Again we observe a better stability of the three-body effective-interaction results with respect to the model-space size changes, a smaller $\Omega$ dependence, and a faster convergence, in paticular for the ground-state. In Table \[tab0\] we show the absolute value of the ground-state energy differences obtained in the calculations with HO energies of $\hbar\Omega=16$ MeV and $\hbar\Omega=22$ MeV for both the two-body and the three-body effective-interaction calculations in different model spaces. We can see that the differences obtained with the three-body effective interaction are almost two times smaller in model spaces with $N_{\rm max}=6-10$. The differences decrease with the enlargement of the model space for $N_{\rm max}\geq 8$. We note that by construction the present two- and three-body effective-interaction calculations would become identical for $N_{\rm max}=28$. In Table \[tab1\] we present a summary of our results obtained in the largest model spaces used in the present study, e.g., $N_{\rm max}=14$ for the positive-parity states and $N_{\rm max}=13$ for the negative-parity states. The positive-parity state results were obtained using the three-body effective interaction. For the Argonne V8’ NN potential calculations we also include the GFMC ground-state results [@Pipcom] for a comparison. We note that the Faddeev-Yakubovski equation solution gives -25.03 MeV [@GH93] for the Nijmegen NN potential [@NRS78], which gives comparable results to the Reid93 NN potential for the three-nucleon problem. The experimental binding energy of $^4$He is 28.296 MeV. The discrepancy between the experimental and calculated values are usually attributed to the real three-nucleon forces that were not taken into account either in our calculation or in the other calculations, which we discussed. We note that the difference in the binding energies obtained using the V8’ and the Reid93 NN potentials is mainly due to the Coulomb interaction included in an isospin-invariant manner only in the calculations with the Reid93 NN potential. Charge form factors ------------------- A sensitive test of the wave-functions obtained in our calculations is the evaluation of charge form factors. Using the formalism of Ref. [@MSD94], we calculated the charge EM and strangeness form factors in the impulse approximation. The one-body contribution to the charge operator is given by Eq. (15) in Ref. [@MSD94], e.g., $$\label{chargeop} \hat{M}_{00}^{(a)}(q)^{[1]}=\frac{1}{2\sqrt{\pi}}\sum_{k=1}^A \left\{\frac{G_E^{(a)}(\tau)}{\sqrt{1+\tau}}j_0(q r_k)+ \left[ G_E^{(a)}(\tau) - 2 G_M^{(a)}(\tau) \right] 2\tau \frac{j_1(q r_k)}{q r_k} {\bf \sigma}_k \cdot {\bf L}_k \right\} \; ,$$ where $\tau=\frac{q^2}{4 m_N ^2}$, ${\bf L}_k$ is the $k$-th nucleon orbital momentum, $G_E^{(a)}(\tau)$ and $G_M^{(a)}(\tau)$ are the one-body electric and magnetic form factors, respectively. The superscript $(a)$ refers to $(T=0)$ for isoscalar EM form factor or to $(s)$ for the strangeness form factor. We use the parametrization of the one-body form factors as discussed in Ref. [@MSD94] \[obform\] $$\begin{aligned} G_E^{(p)}(\tau) &=& G_V^D(\tau) \; , \\ G_M^{(p)}(\tau) &=& \mu_p G_V^D(\tau) \; , \\ G_E^{(n)}(\tau) &=& -\mu_n \tau G_V^D(\tau)\xi_n(\tau) \; , \\ G_M^{(n)}(\tau) &=& \mu_n G_V^D(\tau) \; , \\ G_E^{(s)}(\tau) &=& \rho_s \tau G_V^D(\tau)\xi_s(\tau) \; , \\ G_M^{(s)}(\tau) &=& \mu_s G_V^D(\tau) \; , \end{aligned}$$ with \[obformpar\] $$\begin{aligned} G_V^D(\tau) &=& (1+\lambda^D_V \tau)^{-2} \; , \\ \xi_n &=& (1+\lambda_n \tau)^{-1} \; , \\ \xi_s &=& (1+\lambda^{(s)}_E \tau)^{-1} \; . \end{aligned}$$ The isoscalar EM form factor is given by $G_{E,M}^{(T=0)}=\frac{1}{2}[G_{E,M}^{(p)}+G_{E,M}^{(n)}]$, and for the parameters appearing in Eqs. (\[obform\],\[obformpar\]), one has numerically $\mu_p=2.79, \mu_n=-1.91, \lambda_V^D=4.97$, and $\lambda_n = 5.6$. Following Ref. [@MSD94], we also set the strangeness radius $\rho_s =-2.0$ and $\lambda^{(s)}_E = \lambda_n$. Limits on these parameters are to be determined in the experiments at the Thomas Jefferson Accelerator Facility (TJNAF). The first strangeness magnetic-moment measurement was reported recently [@M97] and an experimental value $\mu_s=+0.23$, obtained with a large error. We use this value in our calculations. Our charge form factor calculations are presented in Figs. \[figEM22\] -\[figratEMS22\]. The charge form factors given in the figures were calculated using the one-body operator (\[chargeop\]) as $F_C^{(a)}(q)=2\sqrt{\pi}\langle f, 0^+ 0|\hat{M}_{00}^{(a)}(q)^{[1]} |i,0^+ 0\rangle$. We show only results obtained with the Argonne V8’ NN potential; the Reid93 NN potential gives almost identical results for the charge form factors, when the same HO energy $\hbar\Omega$ is employed. Our calculated elastic EM charge form factor is given in Fig. \[figEM22\] together with the inelastic EM charge form factor corresponding to the transition to the first excited $0^+ 0$ state. These results were obtained using the HO energy $\hbar\Omega=22$ MeV and the three-body effective interaction in the $N_{\rm max}=14$ model space. In this calculation we obtained the best description of the ground state. The calculation of the elastic charge form factor in the impulse approximation can be directly compared to that presented in Fig. 2 of Ref. [@MSD94], performed using Variational Monte Carlo (VMC) wave-functions and the Argonne V14 NN potential. There, the minimum was obtained at $q\approx 3.55$ fm$^{-1}$, while the experimental minimum is at $q\approx 3.2$ fm$^{-1}$. The difference can be explained with the help of meson-exchange-current contributions. The elastic charge EM form factor obtained in our calculation compares well with that obtained by the VMC wave-functions. It is shifted further to higher $q$, namely, we get the minimum at $q\approx 3.75$ fm$^{-1}$. We note that a second minimum appears in our calculated elastic charge form factor at $q\approx 7.25$ fm$^{-1}$. The second minimum at a similar position was found in the VMC calculations presented in Ref. [@SPR90]. To examine the form factor dependence on $\hbar\Omega$, we repeated these calculations for different choices of the HO energy. In Fig. \[figEM19\] we show the result obtained with $\hbar\Omega=19$ MeV. All other characteristics are the same as in the calculation of Fig. \[figEM22\]. The minimum here is shifted further to higher $q$, we have it at $q\approx 3.85$ fm$^{-1}$. The difference between the two results is rather small but still it shows that our calculation is not completely converged and, in particular, the description of the high transferred-momentum part of the form factors requires the use of even larger model spaces than we employed. We note that the inelastic form factor has a stronger $\Omega$ dependence than the elastic form factor. As discussed in the previous subsection, convergence of the excited state has not been achieved in our calculations within the model spaces employed. Therefore, our calculated inelastic form factors must be taken with some degree of caution. Let us remark that, in addition to the transition form factor, we also computed the form factor of the first excited $0^+ 0$ state. That form factor was also evaluated in Ref. [@HH97] in the resonating group method approach using the Bonn potential. Similarly as in that work, our calculated $0^+_2$ form factor is almost an order of magnitude smaller than the ground-state form factor for a wide range of $q$. The first minimum is shifted in our calculation to larger $q$, more or less to the position of the transition form factor minimum, and the second minimum is shifted to smaller $q$ compared to our ground-state form factor. Our calculated elastic strangeness form factor together with the inelastic EM charge form factor corresponding to the transition to the first excited $0^+ 0$ state are shown in Fig. \[figS22\]. These results were obtained using the same wave-functions as those used for calculations presented in Fig. \[figEM22\], namely we had $\hbar\Omega=22$ MeV, the model-space size characterized by $N_{\rm max}=14$ and the three-body effective interaction was employed. The elastic form factor can be compared with the impulse approximation VMC result of Fig. 3 in Ref. [@MSD94]. Similarly, as for the EM elastic form factor, our calculation compares well with VMC result. We note, however, the different value of strangeness magnetic moment used in Ref. [@MSD94] ($\mu_s=-0.2$). Finally, in Fig. \[figratEMS22\] we present the ratio of the EM and strangeness form factors from Figs. \[figEM22\] and \[figS22\]. The ratio of the elastic charge form factors is particularly interesting, as it can be experimentaly obtained from the measurement of the parity-violating left-right asymmetry for scattering of polarized electrons from a $^4$He target. Experiments of this type are now under preparation at TJNAF. Conclusions {#sec4} =========== In the present study we used equations for Faddeev amplitudes, antisymmetrized for three nucleons, to solve the shell-model problem for the four-nucleon system. We performed calculations in larger model spaces, up to an HO excitation of $14\hbar\Omega$ above the unperturbed ground-state, than in any other shell-model study so far. The main motivation for the present work was to test the shell-model approach and the effective interactions that we want to apply to more complex systems, e.g., $p$-shell nuclei, in particular. The effective interactions that we employed were derived from realistic NN potentials, i.e., the Reid93 and the Argonne V8’. In addition to the two-body effective interactions, we also computed the three-body effective interactions and demonstrated that their use significantly improves the convergence of the results. Our calculations depend on the model-space size and on the HO frequency $\Omega$. The effective interactions were constructed in such a way that in the large model-space limit the effective interactions approach the bare NN interaction. Thus our results should converge to the exact solutions. The dependence on the model-space size and $\Omega$ was investigated. We found quite different behavior of the ground state and the first excited $0^+ 0$ state. Our ground-state energy and point-nucleon radius results begin to converge and are close to or in agreement with those obtained by the GFMC method. For the first excited $0^+ 0$ state, our results, the point-nucleon radius, in particular, show large model-space and $\Omega$ dependence. This implies that significantly larger model spaces would still be needed in order to obtain the exact solutions. The nature of the $0_2^+$ state is discussed in the literature [@CVS88; @CH97]. The Coulomb interaction plays an important role in the description of this state. In the present calculations we did not include the isospin breaking. Our formalism is quite general, however, and allows the use of interactions that break the isospin symmetry. On the other hand, the calculated properties of the $0^-_1$ state show better convergence behavior. In the model spaces studied, we obtained lower excitation energy of the $0^-_1$ state than of the $0^+_2$ state, contrary to experiment. The extrapolation of the model-space dependence of these two energies to larger model spaces shows, however, that the correct ordering of the states will be obtained. Apparently, the $0^+_2$ state is associated with a radial excitation and, thus, it is more sensitive to the HO basis used in the expansion. A sensitive test of our calculated wave functions is the computation of the charge EM and strangeness form factors. Our impulse-approximation results show little dependence on the NN potential and our best results are close to the corresponding form factors obtained using the VMC wave functions and the Argonne V14 NN potential. In particular, we observe both the first and the second minima in the elastic charge form factor in positions close to those obtained using the VMC calculations. In addition to the elastic charge form factors, we also evaluated the form factors for the transition to the $0^+_2$ state in the impulse approximation. In general, the energy scales of the bound $0s$ nucleons are significantly different from the scattering energies of the resonances. This difference can only be accounted for with a large $N_{\rm max}$ in our approach. Consequently, the results for the excited states and the transition form factors obtained within the limited model spaces of the present work should be taken with some caution. In the future we would like to apply the formalism, discussed in the present paper, to a more extensive study of the negative-parity states of $^4$He. In particular, it is desirable to use still larger model spaces to investigate the excited-state convergence properties. The most important result of the present work is, however, the successful use of the three-body effective interaction. This three-body effective interaction can be computed for more complex nuclei as well and, in principle, used, after a transformation to an appropriate three-nucleon basis, in standard shell-model calculations. A more practical approach, however, is to make use of the three-body effective interaction knowledge for the renormalization of the two-body effective interaction. Work in this direction is under way. In addition, the present formalism may be used to compute the four-body effective interaction for nuclei with $A>4$. We plan to extend the shell-model Faddeev-like approach that we have successfully applied to three- and four-nucleon systems to systems with more than four nucleons, using also a formalism of equations for components with lower degree of antisymmetry than the full wave-function developed in Ref. [@GPK]. L.D. Faddeev, Zh. Eksp. Teor. Fiz. [**39**]{}, 1459 (1960) \[Sov. Phys.-JETP [**12**]{}, 1014 (1961)\]. G.L. Payne, J.L. Friar, B.F. Gibson, and I.R. Afnan, Phys. Rev. [**C 22**]{}, 823 (1980). G.L. Payne, J.L. Friar, and B.F. Gibson, Phys. Rev. [**C 22**]{}, 832 (1980). C.R. Chen, G.L. Payne, J.L. Friar, and B.F. Gibson, Phys. Rev. [**C 31**]{}, 2266 (1985). J.L. Friar, G.L. Payne, V.G.J. Stoks, and J.J. de Swart, Phys. Lett. [**B 311**]{}, 4 (1993). A. Nogga, D. Hüber, H. Kamada, and W. Glöckle, Phys. Lett. B [**409**]{}, 19 (1997). O. A. Yakubovsky, Sov. J. Nucl. Phys. [**5**]{}, 937 (1967). W. Glöckle and H. Kamada, Phys. Rev. Lett. [**71**]{}, 971 (1993). B. S. Pudliner, V. R. Pandharipande, J. Carlson, S. C. Pieper and R. B. Wiringa, Phys. Rev. [**C 56**]{} 1720, (1997); R. B. Wiringa, Nucl. Phys. [**A 631**]{}, 70c (1998). M. Viviani, A. Kievsky, and S. Rosati, Few-Body Systems [**18**]{}, 25 (1995). D. C. Zheng, J. P. Vary, and B. R. Barrett, Phys. Rev. [**C 50**]{}, 2841 (1994); D. C. Zheng, B. R. Barrett, J. P. Vary, W. C. Haxton, and C. L. Song, Phys. Rev. [**C 52**]{}, 2488 (1995). P. Navrátil and B. R. Barrett, Phys. Rev. [**C 54**]{}, 2986 (1996); Phys. Rev. [**C 57**]{}, 3119 (1998). P. Navrátil and B. R. Barrett, Phys. Rev. [**C 57**]{}, 562 (1998). R. Ceuleneer, P. Vandepeutte, and C. Semay, Phys. Rev [**C 38**]{}, 2335 (1988). J. Carlson, V. R. Pandharipande, and R. B. Wiringa, Nucl. Phys. [**A424**]{}, 47 (1984). A. Csòtò and G. M. Hale, Phys. Rev. [**C 55**]{}, 2366 (1997). H. M. Hofmann and G. M. Hale, Nucl. Phys. [**A613**]{}, 69 (1997). M. Viviani, S. Rosati, and A. Kievsky, Phys. Rev. Lett. [**81**]{}, 1580 (1998). F. Ciesielski and J. Carbonell, Phys. Rev. [**C 58**]{}, 58 (1998); F. Ciesielski, J. Carbonell, and C. Gignoux, Nucl. Phys. [**A631**]{}, 653c (1998). L. Trlifaj, Phys. Rev. [**C 5**]{}, 1534 (1972). E. P. Harper, Y. E. Kim, and A. Tubis, Phys. Rev. [**C 6**]{}, 126 (1972). W. Glöckle, [*The Quantum Mechanical Few-Body Problem*]{} (Springer-Verlag, New York, 1983). V. A. Rudnev and S. L. Yakovlev, Phys. of At. Nuclei [**58**]{}, 1662 (1995). A. P. Yutsis, I. B. Levinson, and V. V. Vanagas, Mathematical Apparatus of the Theory of Angular Momentum, (Israel Program for scientific Translations, Jerusalem 1962). K. Suzuki and S.Y. Lee, Prog. Theor. Phys. [**64**]{}, 2091 (1980). K. Suzuki, Prog. Theor. Phys. [**68**]{}, 246 (1982); K. Suzuki and R. Okamoto, Prog. Theor. Phys. [**70**]{}, 439 (1983). V. G. J. Stoks, R. A. M. Klomp, C. P. F. Terheggen, and J. J. de Swart, Phys. Rev. [**C 49**]{} 2950 (1994). A. G. M. van Hees and P. W. M. Glaudemans, Z. f. Phys. [**A314**]{}, 323 (1983). J. P. Vary and D. C. Zheng, “The Many-Fermion-Dynamics Shell-Model Code”, Iowa State University (1994) (unpublished). P. Navrátil, B. R. Barrett, and W. Glöckle, Phys. Rev. [**C 59**]{}, (February 1999). K. Varga and Y. Suzuki, Phys. Rev C [**52**]{}, 2885 (1995); K. Varga, Y. Ohbayasi and Y. Suzuki, Phys. Lett. B [**396**]{}, 1 (1997). D.C. Zheng, B.R. Barrett, J.P. Vary and R.J. McCarthy, Phys. Rev. [**C 49**]{}, 1999 (1994). D.C. Zheng, B.R. Barrett, J.P. Vary and H. Müther, Phys. Rev. [**C 51**]{}, 2471 (1995). S. C. Pieper, private communication. M. M. Nagels, T. A. Rijken, and J. J. de Swart, Phys. Rev. [**D 17**]{}, 768 (1978). M. J. Musolf, R. Schiavilla, and T. W. Donnelly, Phys. Rev. [**C 50**]{}, 2173 (1994). B. Mueller et al., Phys. Rev. Lett. [**78**]{}, 3824 (1997). R. Schiavilla, V. R. Pandharipande, and D. O. Riska, Phys. Rev. [**C 41**]{}, 309 (1990). G. P. Kamuntavičius, Few-Body Systems [**1**]{}, 91 (1986); Sov. J. Part. Nuclei, [**20**]{}, 109 (1988). $N_{\rm max}$ 6 8 10 12 14 ------------------------- ------- ------- ------- ------- ------- $|\Delta E_{\rm 2eff}|$ 1.311 1.466 1.265 1.037 0.834 $|\Delta E_{\rm 3eff}|$ 0.778 0.782 0.676 0.601 0.550 : Absolute value of the ground-state energy differences obtained in the calculations with HO energies of $\hbar\Omega=16$ MeV and $\hbar\Omega=22$ MeV with the two-body (second row) and the three-body (third row) effective interactions in different model spaces. The effective interactions were derived from the Argonne V8’ NN potential. The corresponding energy dependence is shown in Fig. . []{data-label="tab0"} Argonne V8’ NN potential -------------------------- ------------------------------------- ---------------------- ---------------------- ---------------------- ----------- State Variable $\hbar\Omega=16$ MeV $\hbar\Omega=19$ MeV $\hbar\Omega=22$ MeV GFMC $0^+_1$ $E$ \[MeV\] -26.62 -26.30 -26.07 -25.92(8) $\sqrt{\langle r^2 \rangle}$ \[fm\] 1.481 1.485 1.485 1.485(10) $0^+_2$ $E$ \[MeV\] -5.77 -4.89 -3.93 $E_x$ \[MeV\] 20.86 21.42 22.14 $\sqrt{\langle r^2 \rangle}$ \[fm\] 2.906 2.777 2.658 $0^-_1$ $E$ \[MeV\] -6.70 -6.17 -5.59 $E_x$ \[MeV\] 19.93 20.14 20.48 $\sqrt{\langle r^2 \rangle}$ \[fm\] 2.349 2.263 2.186 Reid93 NN potential State Variable $\hbar\Omega=17$ MeV $\hbar\Omega=19$ MeV $0^+_1$ $E$ \[MeV\] -25.69 -25.47 $\sqrt{\langle r^2 \rangle}$ \[fm\] 1.487 1.489 $0^+_2$ $E$ \[MeV\] -5.00 -4.39 $E_x$ \[MeV\] 20.69 21.08 $\sqrt{\langle r^2 \rangle}$ \[fm\] 2.873 2.787 $0^-_1$ $E$ \[MeV\] -5.91 -5.54 $E_x$ \[MeV\] 19.78 19.93 $\sqrt{\langle r^2 \rangle}$ \[fm\] 2.339 2.281 : Results for the ground-state and excited state energies and point-nucleon rms radii, as well as the excitation energies ($E_x$), obtained in the largest model spaces used in the present study, $N_{\rm max}=14$, (13) for the positive- (negative-)parity states, respectively, are presented. All the states have isospin $T=0$. The positive-parity-state calculations were performed using the three-body effective interaction. Results for different HO energies are given in separate columns. The GFMC ground-state results [@Pipcom] are shown for comparison.[]{data-label="tab1"} [^1]: On leave of absence from the Institute of Nuclear Physics, Academy of Sciences of the Czech Republic, 250 68 Řež near Prague, Czech Republic.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We calculate the amount of primordial matter density contrast and the size of the very early universe in the recent Quantum Big Bang theory \[Arxiv:0705.4549\[gr-qc\](2007)\] of the cosmological constant. We obtain $(\delta\rho/\rho)_M = 1.75 \times 10^{-5}$, [*without*]{} the introduction of an adjustable free parameter. Harrison-Zel’dovich $k$-dependence with $A = 64/9\pi^2 = 0.72$ and $n = 1$ in $|\delta_k|^2 = Ak^n$ arises inherently. The size of the universe with which it enters the classical Friedmann-Robertson-Walker (FRW) phase comes out to be 0.2 cm. We conclude that the hypothesis of classical inflation at an early stage of cosmic evolution is [**not**]{} needed.' author: - | Budh Ram\ Physics Department\ New Mexico State University\ Las Cruces, New Mexico 88003, USA\ \ and\ \ Prabhu-Umrao Institute of Fundamental Research\ A2/214 Janak Puri, New Delhi 110058, India\ title: '**Primordial matter density contrast and the size of the very early universe in the Quantum Big Bang theory of the cosmological constant**' --- In a recent paper \[1\] we constructed the Quantum Big Bang theory of the cosmological constant in which its presently observed very small value $\sqrt{\Lambda/3} = 1.00 \times 10^{-61}$ – appears naturally. (Throughout this paper we use units in which $G = c = \hbar = 1$). The Quantum Big Bang theory \[1\] is constructed by applying the method \[2\] earlier given by Ram, Ram and Ram to the classical de Sitter line element. In the resulting quantum theory there appear discrete de Sitter zero-point quantum states with zero-point energies $\d{3\over2} \omega_0, \d{3\over2} \omega_1, \d{3\over2} \omega_2, \cdots; \omega_0, \omega_1, \omega_2, \cdots$ being the frequencies of [*independent*]{} oscillators. As a result of the first quantum bang (QB) the zero-point energy $\d{3\over2} \omega_0$ is reduced to $\d{3\over2} \omega_1$ and the difference $\d{3\over2} (\omega_0-\omega_1) = 2(1)\omega_1$ is transformed into one pair of mass quanta, each quantum of frequency $\omega_1$. As a result of the 2nd QB, the zero-point energy $\d{3\over2} \omega_1$ is further reduced to $\d{3\over2}\omega_2$, the difference $\d{3\over2} (\omega_1 - \omega_2) = 2(2)\omega_2$ being converted to 2 pairs of mass quanta, each mass quantum of frequency $\omega_2$. And so on, till as a result of the last 43-rd QB, 43 pairs of mass quanta, each of frequency $\omega_{43}$ are created. Everyone of these mass quanta contributes to the matter density $\rho_M$, the contribution of each quantum being proportional to the [*square*]{} of its frequency $\omega$ (see \[1\]). First we wish to calculate the matter density contrast $(\delta\rho/\rho)_M$ given in terms of the frequency contrasts $\d{\Delta \omega_n \over \omega_n} \ (n = 1,2,\cdots,43)$ by the relation ()\_M = \^[43]{}\_[n=1]{} 2([\_n \_n]{}) \[one\] since the frequencies $\omega_1,\omega_2,\cdots,\omega_{43}$ contribute [*independently*]{} to the matter density contrast. In order to do that, let us look closely at the conceptually mathematical nature of a QB. After the application in \[1\] of the method \[2\] to the de Sitter line element the resulting quantum equation $(\omega = \sqrt{\Lambda/3})$ (-[12]{} [d\^2 dr\^2]{} + [12]{} \^2 r\^2)U = [1 8]{} U \[two\] has only one eigenvalue $\epsilon = \d{1 \over 8\pi}$ in = (2n + [32]{}) ,  n = 0,1,2,. \[three\] Assigning $\epsilon = \d{1 \over 8\pi}$ to the $n = 0$ ground state results in an oscillator of frequency $\omega_0 = \d{1 \over 12\pi}$. On the other hand, assigning this eigenvalue $\d{1 \over 8\pi}$ to the $n = 1$ state results in an oscillator of a different frequency $\omega_1 = \d{1 \over 28\pi}$, the relation between $\omega_0$ and $\omega_1$ being \_0 = (2(1) + [32]{}) \_1. \[four\] Eq.(\[four\]) is interpreted as: the first QB transforms the energy difference $\d{3\over2} (\omega_0 - \omega_1) = 2(1) \omega_1$ to one pair of mass quanta, each mass quantum of frequency $\omega_1$. In quantum language, the first QB is transition of the isotropic oscillator of frequency $\omega_1$, from its $|n=1\rangle$ state to its $|n=0\rangle$ ground state and spontaneous emission (a la Einstein) of one $p$-[*quantum*]{} (paired-quantum) of frequency $\omega_1$. This is equivalent to a [*Planck*]{} oscillator spontaneously emitting one quantum of frequency $\omega_1$, and is thus analogous to an atomic transition from one quantum state to another with spontaneous emission of a light quantum of frequency $\omega_1$ (see \[3\] and references cited therein). Hence one can calculate the spread $\Delta \omega$ in the frequency $\omega$ of the spontaneously emitted mass-quantum in the same manner one calculates the spread in the frequency of the light-quantum in the atomic case, namely, by using the famous [*Golden Rule*]{} of time-dependent perturbation theory \[4\]. Thus = \_[AB]{} = 2\_|B|H\_I|A|\^2 d. \[five\] In Eq.(\[five\]), $\omega_{AB}$ is the total transition probability per unit time for spontaneous emission of a mass quantum of frequency $\omega$ for the isotropic oscillator transition $|A\rangle \rightarrow |B\rangle$, \_= [\^2 (2)\^3]{} \[six\] is the density of states in the $\omega$ continuum with the emitted quantum of momentum $\vec k$ ($k = \omega$) directed within the solid angle $d\Omega$, and $H_I$ is the self-interaction of the oscillator which we take to be H\_I = p\_r = [dr d]{}  ( [is  the proper  time,  see \[1\]]{}) \[seven\] analogous to the atomic case. For the interaction (\[seven\]), the squared matrix element |B|H\_I|A|\^2 = \^2 |B|r|A|\^2. \[eight\] Then the contribution $\left(\d{\delta\rho \over \rho}\right)_n$ from the $n$-th QB to the total matter density contrast is simply given by ()\_n = ([2\_n \_n]{}) = [2\^3\_n ]{} |0|r|n|\^2. \[nine\] For a given $n$ the matrix element $|\langle 0|r|n\rangle|^2$ can be easily calculated using the isotropic oscillator wave functions given in \[5\] (we explicitly reproduce the first three wave functions in the Appendix). For the first QB |0|r|1|\^2 = [2 3\_1]{} \[ten\] giving ()\_1 = [4\^2\_1 3\^2]{} = 1.75 10\^[-5]{} \[eleven\] with $\omega_1 = \d{1 \over 28\pi}$ (see \[1\]). For the second QB |0|r|2|\^2 = [1 30\_2]{} \[twelve\] giving ()\_2 = [\^2\_2 15\^2]{} = 7 10\^[-8]{} \[thirteen\] with $\omega_2 = 3.1 \times 10^{-3}$ (see \[1\]). Contribution from the 3rd QB is of order $10^{-11}$, i.e., negligible, and it is expected to be negligible for the remainder of the QB’s in view of how the $\omega_n$’s decrease with increasing $n$. Thus the major contribution to the density contrast comes from the first QB and $(\delta\rho/\rho)_M$ remains almost constant at the value $\simeq 1.75 \times 10^{-5}$ during the successive 41 QB’s. And it enters the classical FRW phase with this value. With $\omega_1 = k_1 \rightarrow k$, Eq.(\[eleven\]) is rewritten as ()\^2 = [16 9\^4]{} k\^3 k. \[fourteen\] Note the Harrison-Zel’dovich $k$-dependence in (\[fourteen\]). Equating (\[fourteen\]) with Eq. (7.80) of \[6\] (see also \[7\]), namely ([M M]{})\^2\_R= [A 2]{} ([k\^3 k 2\^2]{})\_[k=R\^[-1]{}]{}, \[fifteen\] one gets for the Zel’dovich amplitude $A$ in $|\delta_k|^2 = Ak^n \ (n=1)$ the value \[8\] A = [64 9\^2]{} = 0.72. \[sixteen\] In essence the matter density contrast $\left(\d{\delta\rho \over \rho}\right)_M \simeq 1.75 \times 10^{-5}$, with its Harrison-Zel’dovich $k$-dependence \[9\] and $A = 0.72$, results from the Quantum Big Bang. And then the classical FRW description takes over. Next we answer the question: what is the size of the universe when it enters the classical FRW phase? The universe enters the classical FRW description after the [**last**]{} 43-rd QB with zero-point energy $\d{3\over2} \omega_{43} = 1.5 \times 10^{-61}$ (see \[1\]). The state of the universe with this zero-point energy is described by the radial quantum equation \[1\] \_[43]{} = (2(43) + [32]{})\_[43]{} \_[43]{} \[seventeen\] with a completely spherically symmetric state function $\psi_{43}$. Thus the size of the universe with which it enters the classical FRW phase, $a_{ECP}$, is simply $\sqrt{4\pi}$ times \[1,10\] the square root of the expectation value of $r^2$ for the state $\psi_{43}$, viz., a\_[ECP]{} = . \[eighteen\] The expectation value of $r^2$ for the state $\psi_{43}$ is most easily calculable by using the well-known result \[11\] that for a given quantum state of the isotropic oscillator the expectation value of the potential energy equals half the total energy: \^2\_[43]{} r\^2 \_[43]{} = [12]{} (2(43) + [32]{})\_[43]{} \[nineteen\] which, with $\omega_{43} = 10^{-61}$ (see \[1\]), gives = 3 10\^[31]{} (= 4.8 10\^[-2]{}  [cm]{}). \[twenty\] Or a\_[ECP]{} = 0.2  [cm]{}. \[twentyone\] Thus the universe enters the classical FRW phase with a size equal to 0.2 cm. With the present size of the universe $\simeq 10^{28}$ cm and the present cosmic microwave background (CMB) radiation temperature $\simeq 2.4 \times 10^{-4}$ eV (2.73 K), the temperature and entropy of the very early universe come out to be $10^{16}$ GeV and $10^{87}$, respectively. It is clear from the results obtained in this paper that the hypothesis of classical inflation \[12\] at an early stage of cosmic evolution is [**not**]{} needed. [**Acknowledgments**]{} The author thanks Rajeev Bhalerao, Arun Ram, Nilam Ram, Narayan Banerjee, Subha Majumdar, and Anatoly Klypin for useful conversations, the Tata Institute of Fundamental Research, Mumbai for a pleasant stay, and the Mathematics Department, Melbourne University, Australia for a delightful visit. He is grateful to Rajeev Bhalerao for his continued help and support. [**Appendix**]{} Here we reproduce the first three radial wave functions $R_{n,\ell=0} (r), \ n = 0,1,2$, for the isotropic oscillator from \[5\]. $$\begin{aligned} R_0 (r) &=& \pi^{-1/4} \alpha^{3/2} 2 e^{-{1\over2}(\alpha r)^2}, \ \alpha = \omega^{1/2}, \nonumber \\[2mm] R_1 (r) &=& \pi^{-1/4} \alpha^{3/2} {2^{3/2} \over 3^{1/2}} \left({3\over2} - (\alpha r)^2\right) e^{-{1\over2}(\alpha r)^2}, \nonumber \\[2mm] R_2(r) &=& \pi^{-1/4} \alpha^{3/2} {2^{3/2} \over (15)^{1/2}} \left[{15 \over 4} - 5(\alpha r)^2 + (\alpha r)^4\right] e^{-{1\over2}(\alpha r)^2}. \nonumber\end{aligned}$$ [**References and Notes**]{} 1. Budh Ram, ArXiv:0705.4549 \[gr-qc\] (2007). 2. B. Ram, A. Ram, and N. Ram, ArXiv:gr-qc/0504030 (2005); B. Ram, Phys. Lett. , 1 (2000). See also, B. Ram and J. Shirley, ArXiv:gr-qc/0604074 (2006). 3. B. Ram, ArXiv:quant-ph/0503109 (2005). 4. W. Heitler, [*The Quantum Theory of Radiation*]{}, third edition (Dover Publications, New York 1984), §17 and 18. 5. M.A. Preston and R.K. Bhaduri, [*Structure of the Nucleus*]{} (Addison-Wesley Publishing Co. 1975), p.120. 6. J.V. Narlikar, [*An Introduction to Cosmology*]{}, 3rd edition (Cambridge Univ. Press 2002), p.258. 7. T. Padmanabhan, [*Structure Formation in the Universe*]{} (Cambridge Univ. Press 1993), Chap.5. 8. The best fit for the power-law LCDM model for WMAP + 2d FGRS + Ly$\alpha$ data gives $A \simeq 0.75$; see D.N. Spergel [*et al.*]{}, Ap J. Supplement Series , 175 (2003). 9. The Harrison-Zel’dovich model is consistent with the WMAP three-year data; see W.H. Kinney, E.W. Kolb, A. Melchiorri, and A. Riotto, ArXiv:astro-ph/0605338 (2006). 10. H.P. Robertson, Phil. Mag. [**5**]{}, 835 (1928). 11. L.I. Schiff, [*Quantum Mechanics*]{}, 3rd edition (McGraw-Hill, New York, 1955), p.180. 12. A.H. Guth, Phys. Rev. [**D23**]{}, 347 (1981).
{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $\mathbb{F}_{\Theta }=G/P_{\Theta }$ be a generalized flag manifold, where $G$ is a real noncompact semi-simple Lie group and $P_{\Theta }$ a parabolic subgroup. A classical result says the Schubert cells, which are the closure of the Bruhat cells, endow $\mathbb{F}_{\Theta }$ with a cellular CW structure. In this paper we exhibit explicit parametrizations of the Schubert cells by closed balls (cubes) in $\mathbb{R}^{n}$ and use them to compute the boundary operator $\partial $ for the cellular homology. We recover the result obtained by Kocherlakota \[1995\], in the setting of Morse Homology, that the coefficients of $\partial $ are $0$ or $\pm 2$ (so that $\mathbb{Z}_{2}$-homology is freely generated by the cells). In particular, the formula given here is more refined in the sense that the ambiguity of signals in the Morse-Witten complex is solved.' address: - 'Department of Mathematics, Federal University of Juiz de Fora, Juiz de Fora 36036-900, Minas Gerais, Brazil' - 'Imecc - Unicamp, Departamento de Matemática. Rua Sérgio Buarque de Holanda, 651, Cidade Universitária Zeferino Vaz. 13083-859 Campinas São Paulo, Brazil' author: - Lonardo Rabelo - 'Luiz A. B. San Martin' bibliography: - 'biblio.bib' title: Cellular Homology of Real Flag Manifolds --- [^1] [^2] *AMS 2010 subject classification:* 57T15, 14M15. *Key words and phrases:* Flag manifolds, cellular homology, Schubert cells. Introduction {#introduction .unnumbered} ============ Let $\mathbb{F}_{\Theta }=G/P_{\Theta }$ be a flag manifold of the non-compact semi-simple Lie group $G$ where $P_{\Theta }$ is a parabolic subgroup. A classical result says that a cellular structure of $\mathbb{F}_{\Theta }$ is given by the Schubert cells $\mathcal{S}_{w}^{\Theta }$ which are the closure of the Bruhat cells, that is, the components of the Bruhat decomposition $$\mathbb{F}_{\Theta }=\coprod_{w\in \mathcal{W}/\mathcal{W}_{\Theta }}N\cdot wb_{\Theta },$$where $N$ is the nilpotent component of the Iwasawa decomposition, $\mathcal{W}$ is the Weyl group of the corresponding Lie algebra and $\mathcal{W}_{\Theta}$ is the subgroup of $\mathcal{W}$ associated with $\Theta$. In order to compute the cellular homology of $\mathbb{F}_{\Theta}$, our first task in this paper is to provide explicit parametrizations of the Schubert cells $\mathcal{S}_{w}^{\Theta }$ by cubes $\left[ 0,\pi \right] ^{d}\subset \mathbb{R}^{d}$ which are defined in terms of the reduced decompositions of $w$. This description turns out to be useful to get algebraic formulas for the boundary operator $\partial $ of the cellular homology. Our strategy consists by working firstly in the maximal flag manifolds, denoted by $\mathbb{F}$, and then by projecting down the Schubert cells via the canonical map $\pi _{\Theta }: \mathbb{F}\rightarrow \mathbb{F}_{\Theta }$. To parametrize a Schubert cell $\mathcal{S}_{w}$, $w\in \mathcal{W}$, in the maximal flag manifold $\mathbb{F}$, we start with a minimal decomposition $w=r_{1}\cdots r_{n}$ of $w$ as a product of reflections $r_{i}=r_{\alpha _{i}}$ with respect to the simple roots. Then, similar to the construction of Bott-Samelson dessingularization, we see $\mathcal{S}_{w}$ as a product $K_{1}\cdots K_{n}\cdot b_{0}$, where $b_{0}=P$ is the origin of $\mathbb{F}$ and $K_{i}$ are maximal compact subgroups of rank one Lie groups $G_{i}$ (see Section \[secschubertmax\]). This presents $\mathcal{S}_{w}$ as successive fibrations by spheres $S^{d_{i}}$, where $d_{i}$ are the multiplicities of the roots $\alpha _{i}$ - which may be not equal to $1$. Thus a parametrization $\Phi _{w}:B^{d}\rightarrow \mathcal{S}_{w}$ of a cell of dimension $d=d_{1}+\cdots d_{n}$ is obtained by viewing $S^{d_{i}}$ as the ball $B^{d_{i}}$ whose boundary is collapsed to a point. The case of interest for homology are the roots $\alpha _{i}$ with multiplicity $d_{i}=1$. This is because the boundary operator $\partial $ for the cellular homology takes the form $\partial \mathcal{S}_{w}=\sum c\left( w,w^{\prime }\right) \mathcal{S}_{w^{\prime }}$ with $w^{\prime }=r_{1}\cdots \widehat{r_{i}}\cdots r_{n}$ and the index $i$ is such that $d_{i}=1$. In this case, the characteristic map $\Phi _{w}$ is defined in $B^{d-1}\times \left[ 0,\pi \right] $ and the coefficient $c\left( w,w^{\prime }\right) $ is the sum of the degrees of the attaching maps, that is, the restrictions of $\Phi _{w}$ to $B^{d-1}\times \{0\}$ and $B^{d-1}\times \{\pi \}$ (see Section \[sechommax\], in particular the example of $\mathrm{Sl}\left( 3,\mathbb{R}\right) $ in Subsection \[subsecsl3\]). This way we get that any coefficient $c\left( w,w^{\prime }\right) $ is $0$ or $\pm 2$. In particular, the $\mathbb{Z}_{2}$-homology is the vector space with basis $\mathcal{S}_{w}$, $w\in \mathcal{W}$. Once the maximal flag manifold is worked out, we get the boundary operator $\partial ^{\Theta }$ in a general flag manifold $\mathbb{F}_{\Theta }$. Actually we can prove that for a cell $\mathcal{S}_{w}^{\Theta }$ in $\mathbb{F}_{\Theta }$ there exists a unique (minimal) cell $\mathcal{S}_{w}$ in $\mathbb{F}$ with $\pi _{\Theta }\left( \mathcal{S}_{w}\right) =\mathcal{S}_{w}^{\Theta }$. Then $\partial ^{\Theta }$ is obtained directly from the $\partial $ applied to the minimal cells. These results were already obtained by Kocherlakota [@Koc95] in the realm of Morse homology. In [@Koc95], Theorem 1.1.4, it is proved that the boundary operator for the Morse-Witten complex has coefficients $0$ or $\pm 2 $ as well. Clearly the cellular and the Morse-Witten complexes are intimately related since the Bruhat cells are the unstable manifolds of the gradient flow of a Morse function (see Duistermat-Kolk-Varadarajan [@DKV83] ). Nevertheless the cellular point of view has the advantage of showing the geometry in a more evident way. For instance, in the Subsection \[secgrad\], we provide a description of the flow lines of the gradient flow inside a Bruhat cell in terms of characteristic maps of the cellular decomposition. Also, the choice of minimal decompositions for the elements of $\mathcal{W}$ fix certain signs that are left ambiguous in the Morse-Witten complex. The construction of cellular decompositions of group manifolds and homogeneous spaces is an old theme. For the classical compact Lie groups one can build cells using products of reflections via a method that goes back to Whitehead [@Whi44] and was later developed by Yokota [@Yok55], [@Yok56]. By projection the decomposition on group level induces decompositions on the Stiefel manifolds $V_{n,k}$, that were exploited by Miller [@Mil53] to get several homological properties of these manifolds. On the contrary the cellular decompositions of the group manifolds do not project, in general, to cells in the flag manifolds. Hence that method does not yield cellular decomposition of the flag manifolds. On the other hand the Schubert cells are central objects in the study of (co) homological properties of the flag manifolds (see e.g. Bernstein-Gelfand-Gelfand [@BGG73] and references therein). In the complex case the cellular homology is computed trivially since the cells are all even dimensional hence boundary operator $\partial =0$ and the homology groups are freely generated. We refer also to Casian-Stanton [@CS99] for an approach through representation theory of algebraic reductive groups. For the real flag manifolds $\partial $ is not, in general, trivial and its computation requires explicit expressions for the gluing maps between the cells as we provide in this paper. To the best of our knowledge there is no systematic construction of the cellular decomposition of the flag manifolds (of arbitrary semi-simple Lie groups) through the Bruhat cells and their closures the generalized Schubert cells. The cells constructed here appeared before (up to cells of dimension two) in Wiggerman [@Wig98], that uses them to get generators and relations for the fundamental groups of the flag manifolds. Also in Rabelo [@Rab16] and Rabelo-Silva [@RL18] the method of this paper is used to compute the integral homology of the Real isotropic Grassmannians (those of type B,C and D). The article is organized as follows: In Section \[secschubertmax\] we construct the parametrizations of the Schubert cells on the maximal flag manifolds and analyze the attaching (gluing) maps. In particular, in the subsection \[secgrad\] we look at some aspects of the gradient flow yielding Morse homology. Section \[sechommax\] is devoted to the boundary operator $\partial $ on the maximal flag manifold. The partial flag manifolds are treated in Section \[secpartial\]. In this point we would like to thank Lucas Seco for his comments on some proofs and for his interest in the problem suggesting interesting references related to this question. Notation {#secpre .unnumbered} -------- Flag manifolds are defined as homogeneous spaces $G/P$ where $G$ is a noncompact semi-simple Lie group and $P$ is a parabolic subgroup of $G$. Let $\mathfrak{g}$ be a noncompact real semi-simple Lie algebra. The flag manifolds for the several groups $G$ with Lie algebra $\mathfrak{g}$ are the same. With this in mind we take always $G$ to be the identity component of the automorphism group of $\mathfrak{g}$, which is centerless. Take a Cartan decomposition $\mathfrak{g} = \mathfrak{k} \oplus \mathfrak{s}$ with $\mathfrak{k}$ the compactly embedded subalgebra and denote by $\theta$ the corresponding Cartan involution. Let $\mathfrak{a}$ be a maximal abelian subalgebra contained in $\mathfrak{s}$ and denote by $\Pi$ the set of roots of the pair $(\mathfrak{g}, \mathfrak{a})$. Fix a simple system of roots $\Sigma \subset \Pi$. Denote by $\Pi^{\pm}$ the set of positive and negative roots respectively and by $\mathfrak{a}^+$ the Weyl chamber $$\mathfrak{a}^+ = \{ H \in \mathfrak{a} : \alpha (H) > 0 \mbox{ for all } \alpha \in \Sigma \}.$$ Let $\mathfrak{n} = \displaystyle \sum_{\alpha \in \Pi^+} \mathfrak{g}_\alpha $ be the direct sum of root spaces corresponding to the positive roots. The Iwasawa decomposition of $\mathfrak{g}$ is given by $\mathfrak{g} = \mathfrak{k} \oplus \mathfrak{a} \oplus \mathfrak{n}$. The notations $K, A$ and $N$ are used to indicate the connected subgroups whose Lie algebras are $\mathfrak{k}, \mathfrak{a}$ and $\mathfrak{n}$ respectively. A sub-algebra $\mathfrak{h} \subset \mathfrak{g}$ is said to be a Cartan sub-algebra if $\mathfrak{h}_{\mathbb{C}}$ is a Cartan sub-algebra of $\mathfrak{g}_{\mathbb{C}}$. If $\mathfrak{h} = \mathfrak{a}$ is a Cartan sub-algebra of $\mathfrak{g}$ we say that $\mathfrak{g}$ is a split real form of $\mathfrak{g}_{\mathbb{C}}$. A minimal parabolic subalgebra of $\mathfrak{g}$ is given by $\mathfrak{g} = \mathfrak{m} \oplus \mathfrak{a} \oplus \mathfrak{n}$ where $\mathfrak{m}$ is the centralizer of $\mathfrak{a}$ in $\mathfrak{k}$. Let $P$ be the minimal parabolic subgroup with Lie algebra $\mathfrak{p}$ which is the normalizer of $\mathfrak{p}$ in $G$. We call $\mathbb{F} = G/P$ the maximal flag manifold of $G$ and denote by $b_0$ the base point $1 \cdot P$ in $G/P$. Associated to a subset of simple roots $\Theta\subset \Sigma$ there are several Lie algebras and groups. We write $\mathfrak{g}(\Theta)$ for the semi-simple Lie algebra generated by $\mathfrak{g}_{\pm \alpha}$, $\alpha \in \Theta$. Let $G(\Theta)$ be the connected group with Lie algebra $\mathfrak{g}(\Theta)$. Moreover, let $\mathfrak{n}_\Theta$ be the subalgebra generated by the roots spaces $\mathfrak{g}_{-\alpha}$, $\alpha \in \Theta$ and put $$\mathfrak{p}_\Theta = \mathfrak{n}_\Theta \oplus \mathfrak{p}.$$ The normalizer $P_{\Theta }$ of $\mathfrak{p}_{\Theta }$ in $G$ is a standard parabolic subgroup which contains $P$. The corresponding flag manifold $\mathbb{F}_{\Theta }$ is called a partial flag manifold of $G$ or flag manifold of type $\Theta $. We denote by $b_{\Theta }$ the base point $1\cdot P_{\Theta }$ in $G/P_{\Theta }$. Such a flag manifold can also be written as $\mathbb{F}_{\Theta }=K/K_{\Theta }$ where $K_{\Theta }=P_{\Theta }\cap K$. The Weyl group $\mathcal{W}$ associated to $\mathfrak{a}$ is the finite group generated by the reflections over the root hyperplanes $\alpha = 0$ contained in $\mathfrak{a}$, $\alpha \in \Sigma $, and can be alternatively given as the quotient $M^* / M$ where $M^*$ and $M$ are respectively the normalizer and the centralizer of $\mathfrak{a}$ in $K$ (the Lie algebra of $M$ is $\mathfrak{m}$). We use the same letter to denote a representative of $w$ in $M^*$. For the subset $\Theta \subset \Sigma$, there exists the subgroup $\mathcal{W}_\Theta$ which acts trivially on $\mathfrak{a}_\Theta = \{ H \in \mathfrak{a} : \alpha (H) = 0 , \alpha \in \Theta\}$. Alternatively, $\mathcal{W}_\Theta $ may be seen as the subgroup of the Weyl group generated by the reflections with respect to the roots $\alpha \in \Theta$. Viewing the elements of $\mathcal{W}$ as product of simple reflections, the length $\ell(w)$ of $w \in \mathcal{W}$, is the number of simple reflections in any reduced expression of $w$ which is equal to the cardinality of $\Pi_w = \Pi^+ \cap w \Pi^-$, the set of positive roots sent to negative roots by $w^{-1}$. If $w = r_1 \cdots r_n$ is a reduced expression of $w$ then $$\Pi_w = \{ \alpha_{1}, r_1 \alpha_2, \ldots, r_1 \cdots r_{n-1}\alpha_n\}.$$ There are two equivalent definitions of order between elements in the Weyl group (see Humphreys [@Hum90]). 1. First, two elements are connected, denoted $w_{1}\rightarrow w_{2}$, if $\ell(w_{1})<\ell(w_{2})$ and there is a root $\alpha $ (not necessarily simple) such that $w_{1}r_{\alpha }=w_{2}$. Now that $w_{1}<w_{2}$ if there are $u_{1},\ldots ,u_{k}\in \mathcal{W}$ with $$w_{1}\rightarrow u_{1}\rightarrow \cdots u_{k}\rightarrow w_{2}.$$ It may happen that $w_{1}\rightarrow w_{2}$ with $\ell(w_{1})+1=\ell(w_{2})$ but there is no simple root with $w_{1}r_{\alpha }=w_{2}$. The definion may be changed by multiplication in the left $r_{\alpha }w_{1}=w_{2}$ because $r_{\alpha }w_{1}=w_{1}(w_{1}^{-1}r_{\alpha }w_{1})=w_{1}r_{\beta }$ with $\beta =w^{-1}\alpha $. 2. $w_{1}\leq w_{2}$ if given a reduced expression $w_{2}=r_{1}\cdots r_{\ell(w_{2})}$ then $w_{1}=r_{i_{1}}\cdots r_{i_{k}}$ for some indices $i_{1}<\cdots <i_{k}$. There is a unique $w_{0}\in \mathcal{W}$ such that $w_{0}\Pi ^{+}=\Pi ^{-}$ which we call the principal involution and is the maximal element in the Bruhat-Chevalley order. A partial flag manifold is the base space for the natural equivariant fibration $\pi _{\Theta }:\mathbb{F}\rightarrow \mathbb{F}_{\Theta }$ whose fiber is $P_{\Theta }/P$. This fiber is a flag manifold of a semi-simple Lie group $M_{\Theta }\subset G$ whose rank is the order of $\Theta $. The Weyl group of $M_{\Theta }$ is the subgroup $\mathcal{W}_{\Theta }$. Its orbit through $b_{0}$ is contained in the fiber $\pi _{\Theta }^{-1}\pi _{\Theta }(b_{0})$. In particular, the group $M_{\Theta }$ is of rank one if $\Theta $ is a singleton. For example, if $\alpha $ is a simple root, the fiber of $\mathbb{F}\rightarrow \mathbb{F}_{\alpha }=G/P_{\alpha }$ which is $P_{\alpha }/P$, coincides with the (unique) flag manifold of the group $G(\alpha )$ whose Lie algebra is $\mathfrak{g}(\alpha )$, generated by $\mathfrak{g}_{-\alpha } $ and $\mathfrak{g}_{\alpha }$. These rank one flag manifolds are spheres $S^{m}$, where $m=\dim (\mathfrak{g}_{\alpha }+\mathfrak{g}_{2\alpha })$. The Bruhat decomposition presents the flag manifolds as a union of $N$-orbits (or one of its conjugates). It says that the $N$-orbits on a flag manifold $\mathbb{F}_{\Theta}$ is finite and coincide with the orbits that goes through the $A$-fixed points. Let $b_{\Theta }$ be the origin of $\mathbb{F}_{\Theta }$. Then the set $A$-fixed points coincides with the orbit $M^{\ast }b_{\Theta }$. This set is finite and is in bijection with $\mathcal{W}/\mathcal{W}_{\Theta }$. Thus the Bruhat decomposition reads $$\mathbb{F}_{\Theta }=\coprod_{w\in \mathcal{W}/\mathcal{W}_{\Theta }}N\cdot wb_{\Theta }\quad, \quad w\in M^{\ast },$$where $N\cdot w_{1}b_{\Theta }=N\cdot w_{2}b_{\Theta }$ if $w_{2}\mathcal{W}_{\Theta }=w_{1}\mathcal{W}_{\Theta }$. When there is an equivariant fibration $\mathbb{F}_{\Theta _{1}}\rightarrow \mathbb{F}_{\Theta _{2}}$ (in particular when $\mathbb{F}_{\Theta _{1}}=\mathbb{F}$) the $N$-orbits project onto $N$-orbits by equivariance, hence the fibration respects the Bruhat decompositions. Each $N$-orbit through $w$ is diffeormophic to an Euclidean space. Such an orbit $N\cdot wb_{\Theta }$ is called a Bruhat cell. Its dimension is given by the formula $$\dim \left( N\cdot wb_{\Theta }\right) =\displaystyle\sum_{\alpha \in \,\Pi _{w}\,\setminus \,\langle \Theta \rangle }m_{\alpha }$$ where $m_{\alpha }$ is the multiplicity of the root space $\mathfrak{g}_{\alpha }$ and $\langle \Theta \rangle $ denotes the roots in $\Pi $ generated by $\Theta $ (see the Lemma \[lemproperties\] for the maximal flag case and Lemma \[minimal\_element\] for the partial flag case). In particular, the Bruhat cell $N\cdot w_{0}b_{\Theta }$ is an open and dense orbit. The closure of the Bruhat cells are called (generalized) Schubert cells. A Schubert cell is the closure of a Bruhat Cell: $$\mathcal{S}_{w}^{\Theta }=\mathrm{cl}(N\cdot wb_{\Theta }).$$ The Schubert cells endow the flag manifolds with a cellular decomposition. For a maximal flag manifold we avoid the superscript $\Theta $ and write simply $$\mathcal{S}_{w}=\mathrm{cl}(N\cdot wb_{0})$$ We recall the following well known facts (see [@DKV83] or Warner [@War72]). \[BC-order\] $\mathcal{S}_{w_{1}}^{\Theta }\subset \mathcal{S}_{w_{2}}^{\Theta }$ if and only if $w_{1}\leq w_{2}$. \[BC-order-1\] $\mathcal{S}_{w}^{\Theta }=\displaystyle{\cup _{u\leq w}N\cdot ub_{\Theta }}$. In the forthcoming sections we will look carefully at the cellular decompositions of the flag manifolds given by the Schubert cells. Before going into them we present examples showing that classical cell decompositions of compact groups are not well behaved with respect to projections to flag manifolds. **Example:** In the cellular decomposition of $\mathrm{SO}\left( 3\right) $ of [@Whi44] and [@Mil53] there are $4$ cells of dimensions $0$, $1$, $2$ and $3$. The $2$-dimensional cell is given by the map $f:\mathbb{RP}^2 \rightarrow \mathrm{SO}\left( 3\right) $ given by $f\left( \left[ x\right] \right) =r_{x}d$, $x\in \mathbb{R}^{3}\setminus \{0\}$, where $r_{x}$ is the reflection in $\mathbb{R}^{3}$ with respect to the plane orthogonal to $x$ and $d=\mathrm{diag}\{1,1,-1\}$ needed to correct the determinant. This map is viewed as a two-cell $B_{2}\rightarrow \mathrm{SO}\left( 3\right) $ by taking the interior of the $2$-ball $B_{2}$ as the set $\{\left[ x\right] \in \mathbb{RP}^{2}:x_{3}\neq 0\}$ where $x=\left( x_{1},x_{2},x_{3}\right) $. The boundary of $B_{2}$ is mapped to the $1$-dimensional cell which is the image under $f$ of $\mathbb{RP}^{1}=\{\left[ \left( x_{1},x_{2},0\right) \right] \in \mathbb{RP}^{2}\}$. If $\{e_{1},e_{2},e_{3}\}$ is the standard basis of $\mathbb{R}^{3}$ then $f\left( \left[ e_{1}\right] \right) =\mathrm{diag}\{-1,1,-1\}$, $f\left( \left[ e_{2}\right] \right) =\mathrm{diag}\{1,-1,-1\}$ and $f\left( \left[ e_{3}\right] \right) =\mathrm{id}$. These three elements belong to the group $M$ where $\mathbb{F}=\mathrm{SO}\left( 3\right) /M$ is the maximal flag manifold of $\mathrm{Sl}\left( 3,\mathbb{R}\right) $. Hence the projection to $\mathbb{F}$ of the $2$-cell in $\mathrm{SO}\left( 3\right) $ is not a cell in $\mathbb{F}$ because $f\left( \left[ e_{i}\right] \right) $, $i=1,2,3$ are projected to the same point, namely the origin of $\mathbb{F}$. For other examples we recall the cellular decomposition of $\mathrm{SU}\left( n\right) $ given in [@Yok56], Theorem 7.2, where the positive dimensional cells have dimension $\geq 3$. Hence this construction does not yield, by projection $\mathrm{SU}\left( n\right) \rightarrow \mathrm{SU}\left( n\right) /H$, a cellular decomposition of $\mathrm{SU}\left( n\right) /H$ if this homogeneous space has non trivial homology at the levels $1$ or $2$. This happens, for instance, with the flag manifolds of $\mathrm{Sl}\left( n,\mathbb{C}\right) $, that have nontrivial $H_{2}$. Also, the maximal compact subalgebra of the split real form of the exceptional type $E_{7}$ is $\mathfrak{su}\left( 8\right) $. However the maximal flag manifold of a split real form has nontrivial fundamental group (and hence $H_{1}$) as follows by Johnson [@Joh04] and [@Wig98] . Schubert cells in maximal flag manifolds\[secschubertmax\] ========================================================== In this section we give a detailed description of the Schubert cells in the maximal flag manifolds. This description includes a parametrization by compact groups (subsets of them) which allows explicit expressions for the gluing maps between the cells. The partial flag manifolds will be trated in the Section \[secpartial\]. Schubert cells and product of compact subgroups ----------------------------------------------- The main result here is a suitable parametrization for the Schubert cells which is the basis for the computation of the boundary operator for the cellular homology. As before, $\mathbb{F}=G/P$ is the maximal flag manifold. We denote by $\mathbb{F}_{i}=G/P_{i}$ the partial flag manifolds where $P_{i}=P_{\{\alpha _{i}\}}$, with $\alpha _{i}$ a simple root. The canonical fibration is $\pi _{i}:\mathbb{F}\rightarrow \mathbb{F}_{i}$. The Schubert cells are firstly described by the fiber-exhausting map $\gamma _{i}$ defined by $$\gamma _{i}(X)=\pi _{i}^{-1}\pi _{i}(X)\quad,\quad X\subset \mathbb{F},$$that is, $\gamma _{i}\left( X\right) $ is the union of the fibers of $\pi _{i}:\mathbb{F}\rightarrow \mathbb{F}_{i}$ crossing $X\subset \mathbb{F}$. Notice that each $\gamma _{i}$ is an equivariant map, i.e., $g\gamma _{i}\left( X\right) =\gamma _{i}\left( gX\right) $, for all $g\in G$ and $X\subset \mathbb{F}$, since the projections $\pi _{i}$ are equivariant maps. For $w\in \mathcal{W}$, put $N^{w}=wNw^{-1}$. Every Schubert cell is the image of some $g\in G$ of $\mathrm{cl}\left( N^{w}wb_{0}\right) $. The following result was proved in [@San98]. Let $w=r_{1}\cdots r_{n}$ be a reduced expression of $w\in \mathcal{W}$ as a product of reflections with respect to the simple roots. Then, for any $k=1,\ldots ,n$, we have $$\mathrm{cl}(N^{w}b_{0})=\gamma _{1}\cdots \gamma _{k}\left( \mathrm{cl}\left( N^{w}r_{1}\cdots r_{k}b_{0}\right) \right) .$$ In particular, for $k=n$ we have $$\mathrm{cl}(N^{w}b_{0})=\gamma _{1}\cdots \gamma _{n}\left( \mathrm{cl}\left( wNw^{-1}wb_{0}\right) \right) =\gamma _{1}\cdots \gamma _{n}\{wb_{0}\} \label{forSchubert0}$$because $Nb_{0}=b_{0}$. From this equality we get the following expression for the Schubert cell $\mathcal{S}_{w}$. \[corschubexaust\]Let $w=r_{1}\ldots r_{n}$ be a reduced expression as a product of reflections with respect to the simple roots in $\Sigma $. Then, $$\mathcal{S}_{w}=\gamma _{n}\cdots \gamma _{1}\{b_{0}\}$$(Note that the order of the indexes is reversed.) We have $\mathrm{cl}(Nw\cdot b_{0})=w\left( \mathrm{cl}(N^{w^{-1}}b_{0})\right) $, hence by (\[forSchubert0\]) with $w^{-1}=r_{n}\cdots r_{1}$ instead of $w$ we have $$\mathcal{S}_{w}=w\gamma _{n}\cdots \gamma _{1}(w^{-1}b_{0})=\gamma _{n}\cdots \gamma _{1}\{b_{0}\},$$where the last equality follows by equivariance. Now we change slightly the above expression in terms of exhausting-fiber maps to get the Schubert cells as unions of successive orbits of the parabolic subgroups $P_{i}$. This construction is in the same spirit as the Bott-Samelson dessingularization (see [@DKV83]). It starts with the remark that the fiber $\gamma _{i}\{b_{0}\}$ of $\pi _{i}:\mathbb{F}\rightarrow \mathbb{F}_{i}$ through the origin is the orbit $P_{i}\cdot b_{0}$. In general, the fiber through $g\cdot b_{0}\in \mathbb{F}$ is given by $g\cdot \gamma _{i}\{b_{0}\}$ by equivariance of $\gamma _{i} $. Now, if we have two iterations $\gamma _{2}\gamma _{1}$, then by equivariance we get $$\begin{aligned} \gamma _{2}\gamma _{1}\{b_{0}\} &=&\gamma _{2}\left( \bigcup_{g\in P_{1}}g\cdot b_{0}\right) \label{union} \\ &=&\left( \bigcup_{g\in P_{1}}g\cdot \gamma _{2}(b_{0})\right) \notag \\ &=&\left( \bigcup_{g\in P_{1}}g\cdot \left( P_{2}b_{0}\right) \right) \notag \\ &=&P_{1}P_{2}\cdot b_{0}. \notag\end{aligned}$$ Proceeding successively by induction, we obtain $$\mathcal{S}_{w}=\gamma _{n}\cdots \gamma _{1}\{b_{0}\}=P_{1}\cdots P_{n}\cdot b_{0},$$where the indexes of $P_{1}\cdots P_{n}$ is the same as those of minimal decomposition $w=r_{i}\ldots r_{n}\in \mathcal{W}$. The same expression still holds with the compact $K_{i}=K\cap P_{i}$ instead of $P_{i}$. In fact, $K_{i}\cdot b_{0}=P_{i}\cdot b_{0}$ by the Langlands decomposition $P_{i}=K_{i}AN$ and $AN\cdot b_{0}$. Hence the same arguments yield the following description of the Schubert cells. \[Schubert\_2\]Let $w=r_{1}\cdots r_{n}$ be a reduced expression as a product of reflections with respect to the simple roots in $\Sigma $. Then, $$\mathcal{S}_{w}=K_{1}\cdots K_{n}\cdot b_{0}.$$(Here, different from Corollary \[corschubexaust\], the indexes of the $r_{i}$’s and $K_{i}$’s are in the same order). **Remark:** In general, there is more than one reduced expression for $w\in \mathcal{W}$, which provides distinct compact subgroups $K_{i}$ and distinct parametrizations. **Example:** Let $G=\mathrm{Sl}(n,\mathbb{R})$ with $\mathfrak{g}=\mathfrak{s}\mathfrak{l}(n,\mathbb{R})$. The simple roots are given by $\alpha _{i,i+1}=\lambda _{i}-\lambda _{i+1}$. The compact group $K_{i}$ associated to the simple root $\alpha _{i,i+1}$ is given by the rotations $$R_{i}^{t}=\exp (tA_{i,i+1})=\left( \begin{array}{cccccc} 1 & & & & & \\ & \ddots & & & & \\ & & \cos t & \sin t & & \\ & & -\sin t & \cos t & & \\ & & & & \ddots & \\ & & & & & 1 \\ & & & & & \end{array}\right)$$ where $A_{i,i+1}=E_{i,i+1}-E_{i+1,i}$. In this case, a Schubert cell has the form $$\mathcal{S}_{w}=R_{i_{1}}^{t_{1}}\cdots R_{i_{m}}^{t_{m}}\cdot b_{0},$$that is, is the image of the map $(t_{1},\ldots ,t_{m})\mapsto R_{i_{1}}^{t_{1}}\cdots R_{i_{m}}^{t_{m}}\cdot b_{0}\in \mathbb{F}$. Continuing with the example, let $n=3$, with $\mathcal{W}$ the permutation group in three letters. The Schubert cell $\mathcal{S}_{(13)}$ is the whole flag $\mathbb{F}_{1,2}^{3}$ since $(13)$ is the principal involution. If we decompose $(13)=(12)(23)(12)$, $\mathcal{S}_{(13)}$ may be parametrized as: $$\left( \begin{array}{ccc} \cos t_{1} & \sin t_{1} & \\ -\sin t_{1} & \cos t_{1} & \\ & & 1 \\ & & \end{array}\right) \left( \begin{array}{ccc} 1 & & \\ & \cos t_{2} & \sin t_{2} \\ & -\sin t_{2} & \cos t_{2} \\ & & \end{array}\right) \left( \begin{array}{ccc} \cos t_{3} & \sin t_{3} & \\ -\sin t_{3} & \cos t_{3} & \\ & & 1 \\ & & \end{array}\right) \cdot b_{0}.$$If we choose to write $(13)=(23)(12)(23)$ we parametrize $\mathcal{S}_{(13)}$ as: $$\left( \begin{array}{ccc} 1 & & \\ & \cos t_{1} & \sin t_{1} \\ & -\sin t_{1} & \cos t_{1} \\ & & \end{array}\right) \left( \begin{array}{ccc} \cos t_{2} & \sin t_{2} & \\ -\sin t_{2} & \cos t_{2} & \\ & & 1 \\ & & \end{array}\right) \left( \begin{array}{ccc} 1 & & \\ & \cos t_{3} & \sin t_{3} \\ & -\sin t_{3} & \cos t_{3} \\ & & \end{array}\right) \cdot b_{0}.$$ In these examples the parameter $t_{i}$ range in the interval $[0,\pi ]$ because $R_{i}^{\pi }\cdot b_{0}=b_{0}$ for any $i$ ($b_{0}=\left( V_{1}\subset V_{2}\right) $ where $V_{1}$ is the one dimensional subspace of $\mathbb{R}^{3}$ spanned by the first basic vector and $V_{2}$ is spanned by the first two basic vectors). This is a general feature since our cell maps will be defined in cubes $\left[ 0,\pi \right] ^{m}$. Bruhat cells inside the Schubert cell ------------------------------------- The next results determine the points of a Schubert cell $\mathcal{S}_{w}=K_{1}\cdots K_{n}\cdot b_{0}$ which are in the corresponding Bruhat cell $N\cdot wb_{0}$. \[Schubert\_lema\] Let $w=r_{1}\cdots r_{n-1}r_{n}$ a minimal decomposition. Define $v=wr_{n}=r_{1}\cdots r_{n-1}$. Let the parabolic subgroup $P_{n}=P_{\{\alpha _{n}\}}$ with $r_{n}$ the reflection with respect to $\alpha _{n}$ and $\mathbb{F}_{n}=G/P_{n}$. Let $\pi _{n}:\mathbb{F}\rightarrow \mathbb{F}_{n}$ be the canonical projection and denote by $b_{n}$ the origin of $G/P_{n}$. Then we have the disjoint union$$\label{main_fiber_equation} \pi _{n}^{-1}(N\cdot wb_{n})=(N\cdot wb_{0})\,\dot{\cup}\,(N\cdot vb_{0})$$ The fiber $\pi _{n}^{-1}(wb_{n})$ is the flag manifold of a rank one group. Its Bruhat decomposition reads $$\pi _{n}^{-1}(wb_{n})=\{vb_{0}\}\,\dot{\cup}\,\left( \pi _{n}^{-1}(wb_{n})\cap (N\cdot wb_{0})\right) .$$Indeed, in $\pi _{n}^{-1}(wb_{n})$ there are a $0$-cell which is $\{vb_{0}\}$ and an open cell. This latter one is $\pi _{n}^{-1}(wb_{n})\cap (N\cdot wb_{0})$ because it is contained in the Bruhat cell $N\cdot wb_{0}$ and $vb_{0}\notin N\cdot wb_{0}$. The result follows by acting $N$. In fact, $wb_0 \in \pi_n^{-1}(wb_n) \cap (N \cdot wb_0)$, hence $N \cdot wb_0 = N \left( \pi_n^{-1}(wb_n) \cap (N \cdot wb_0) \right)$. Also, by equivariance of $\pi_n$ we get $N \pi_n^{-1}(wb_n) = \pi_n^{-1}(N\cdot wb_n)$. Then, $$\pi_n^{-1}(N\cdot wb_n) = N \left( \{ vb_0 \} \, \dot{\cup} \, ( \pi_n^{-1}(wb_n) \cap (N \cdot wb_0) ) \right) = ( N \cdot vb_0 ) \, \dot{\cup} \, (N \cdot wb_0). \qedhere$$ We notice that Equation (\[main\_fiber\_equation\]) is equivalent to $$\begin{aligned} \pi_n^{-1}(N \cdot wb_n) &=& \mathcal{S}_v \, \dot{\cup} \, (N \cdot wb_0).\end{aligned}$$ because $\pi_n^{-1}(N \cdot wb_n) \cap \mathcal{S}_v = N\cdot vb_0$ and $\pi_n^{-1}(N \cdot wb_n) = \pi_n^{-1}(N \cdot vb_n)$ since $wb_n = vb_n$ in $\mathbb{F}_n$. \[propSchubert\_3\]Write $\mathcal{S}_{w}=K_{1}\cdots K_{n}\cdot b_{0}$. Take $b=u_{1}\cdots u_{n}\cdot b_{0}$, with $u_{i}\in K_{i}$. Then $b\in \mathcal{S}_{w}\setminus N\cdot wb_{0}$ if and only if $u_{i}\in M$ for some $i=1,\ldots ,n$. In other words, an element $b \in \mathcal{S}_{w}$ is inside the Bruhat cell $N\cdot wb_{0}$ if and only if there is no $u_{i}\in M$. Suppose that $u=u_{i}\in M$ for some $i$. Then $u\in K_{j}$ for all $j$, since $M\subset K_{j}$, so that $v_{j}=uu_{j}u^{-1}\in K_{j}$. Hence $b$ can be rewritten as $b=u_{1}\cdots u_{i-1}v_{i+1}\cdots v_{n}u\cdot b_{0}$. Since $ub_{0}=b_{0}$, it follows that $b\in \mathcal{S}_{v}$, with $v=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$, which implies that $b\notin N\cdot wb_{0}$ since $v<w$ and $\mathcal{S}_{w}\setminus N\cdot wb_{0}=\cup _{u<w}\mathcal{S}_{u}$. For the converse we use induction on the length of $w$. If $w = r_1$ has length one, then the Schubert cell is $\mathcal{S}_{r_1} = {b_0} \cup (N\cdot u_1b_0)$. So if $u_1\notin M $, then $u_1b_0 \neq b_0$ and hence $u_1\cdot b_0 \in N\cdot u_1b_0$. For $n>1$, let $b=u_{1}\cdots u_{n}\cdot b_{0}$ with $u_{i}\notin M$. We must show that $b\in N\cdot wb_{0}$. Put $x=u_{1}\cdots u_{n-1}\cdot b_{0}$. Note that $b\neq x$ for otherwise $u_{n}b_{0}=b_{0}$ which gives $u_{n}\in M$, contradicting the assumption. The induction hypothesis says that $x \in N \cdot vb_0$, $v=r_1 \cdots r_{n-1}$. Moreover, $\pi_n(b_0) = \pi_n(u_nb_0)$ which implies that $\pi_n(x) = \pi_n(b)$, that is, $x$ and $b$ are in the same fiber of $\pi_n$. Hence $\pi_n(b) \in \pi_n(N\cdot wb_n)$, so that by Lemma \[Schubert\_lema\], $b \in (N\cdot vb_0) \cup (N\cdot wb_0)$. Now $b \notin N\cdot vb_0$ for otherwise $b = x$. In fact, as $\pi_n(b) = \pi_n(x) = zb_n $, for some $z \in N$, we have $b \in \pi_n^{-1}(zb_n) \cap N \cdot vb_0 $. Since this intersection reduces to ${\ zb_0 }$ we have $x = zb_0$, because $x \in N \cdot vb_0$. Hence $b \in N\cdot wb_0$, concluding the proof. Parametrization of subsets of compact subgroups ----------------------------------------------- The next step is to find subsets of the subgroups $K_{i}$ that cover $\mathcal{S}_{w}=K_{1}\cdots K_{n}\cdot b_{0}$ and thus find parametrizations of the cells. The fiber $P_{i}/P$ of the projection $\mathbb{F}\rightarrow \mathbb{F}_{i}$ is the flag of the rank one Lie group $G(\alpha )$ whose Lie algebra is $\mathfrak{g}(\alpha )$, generated by $\mathfrak{g}_{\pm \alpha }$. The flag manifold $\mathbb{F}_{\alpha }$ of $G(\alpha )$ is a sphere $S^{m}$ with dimension $m=\dim \mathfrak{s}_{\alpha }-1$ where $\mathfrak{s}_{\alpha }$ is the symmetric part of the Cartan decomposition of $\mathfrak{g}\left( \alpha \right) $. If $\{1,w\}$ is the Weyl group of $G\left( \alpha \right) $ and $b_{0}$ is the origin of $\mathbb{F}$ then $b_{0}$ and $wb_{0}$ are antipodal points in $S^{m}$. The parametrization we seek is provided by the following lemma whose general proof is only sketched below. In the sequel we write down the details for the case when $\dim \mathfrak{g}_{\alpha }=1$ and $\mathfrak{g}_{2\alpha }=\{0\}$ so that $\mathfrak{g}\left( \alpha \right) \approx \mathfrak{sl}\left( 2,\mathbb{R}\right) $. \[parametrization\]Let $G=G(\alpha )$ be a real one rank group of rank with maximal compact subgroup $K=K_{\alpha }$ and the corresponding flag manifold $\mathbb{F}=S^{m}$ with origin $b_{0}$. Let $B^{m}$ be the closed ball in $\mathbb{R}^{m}$. Then, there exists a continuous map $\psi :B^{m}\rightarrow K$ such that - $\psi (S^{m-1}) \subset M$ and hence $\psi(S^{m-1}) \cdot b_0 = b_0$. - If $x\in B^{m}\setminus S^{m-1}$, then $\psi (x)\cdot wb_{0}$ is a diffeomorphism onto the Bruhat cell which is the complement of $b_0$. For the proof of  the lemma recall that the following list exhaust the rank one Lie algebras (see [@War72], pages 30-32). In the list $d_{\alpha }=\dim \mathfrak{g}_{\alpha }$ and $d_{2\alpha }=\dim \mathfrak{g}_{2\alpha } $. - $\mathfrak{so}\left( 1,n\right) $; $d_{\alpha }=n-1$, $d_{2\alpha }=0$; $\dim \mathfrak{s}=n$. (This class includes $\mathfrak{sl}\left( 2,\mathbb{R}\right) \approx \mathfrak{sp}\left( 1,\mathbb{R}\right) \approx \mathfrak{so}\left( 1,2\right) $, $\mathfrak{sl}\left( 2,\mathbb{C}\right) \approx \mathfrak{so}\left( 1,3\right) $ and $\mathfrak{su}^{\ast }\left( 4\right) \approx \mathfrak{so}\left( 1,5\right) $.) - $\mathfrak{su}\left( 1,n\right) $; $d_{\alpha }=2\left( n-1\right) $, $d_{2\alpha }=1$; $\dim \mathfrak{s}=2n$. (This class includes $\mathfrak{so}^{\ast }\left( 6\right) \approx \mathfrak{su}\left( 1,3\right) $.) - $\mathfrak{sp}\left( 1,n\right) $; $d_{\alpha }=4\left( n-1\right) $, $d_{2\alpha }=3$; $\dim \mathfrak{s}=4n$. - A real form of the exceptional Lie algebra $F_{4}$; $d_{\alpha }=8$, $d_{2\alpha }=7$; $\dim \mathfrak{s}=16$. The exceptional algebra $F_{4}$ does not appear as a $\mathfrak{g}\left( \alpha \right) $ in any Lie algebra different from itself because apart from $F_{4}$ the multiplicities $d_{2\alpha }$ are at most $3$ (see [@War72], pages 30-32). Hence, we can discard it. On the other hand the classical groups $\mathrm{SO}(1,n)$, $\mathrm{SU}(1,n)$ and $\mathrm{Sp}(1,n)$ contain the compact subgroups $\mathrm{SO}(n)$, $\mathrm{SU}(n)$ and $\mathrm{Sp}(n)$ whose actions on the respective flag manifolds $S^{n-1}$, $S^{2n-1}$ and $S^{4n-2}$ are the standard ones coming from the linear actions in $\mathbb{R}^{n}$, $\mathbb{C}^{n}$ and $\mathbb{H}^{n}$, respectively. In each case the origin $b_{0}$ of the flag is the first basic vector $e_{1}$ while $wb_{0}=-e_{1}$. Now, take matrices $$A_{\gamma }=\left( \begin{array}{cc} 0 & -\overline{\gamma }^{T} \\ \gamma & 0\end{array}\right)$$with $\gamma $ in $\mathbb{R}^{n-1}$, $\mathbb{C}^{n-1}$ and $\mathbb{H}^{n-1}$ respectively, such that $\left\Vert \gamma \right\Vert =1$. If $m=n-1 $, $2n-1$ or $4n-1$ then $U$ is one of the groups $\mathrm{SO}(n)$, $\mathrm{SU}(n)$ and $\mathrm{Sp}(n)$. If $U = \mathrm{SO}(n)$ then the map $\psi :S^{m-1}\times \left[ 0,2\pi \right] \rightarrow U$ given by $\psi \left( \gamma ,t\right) =e^{tA_{\gamma }}$ satisfies the requirements of Lemma \[parametrization\], because $e^{tA_{\gamma }}\cdot e_{1}=\cos te_{1}+\sin t\widetilde{\gamma }$ where $\widetilde{\gamma }=\left( 0,\gamma \right) $. The complex and quaternionic cases are made similarly with slight modifications. If $U= \mathrm{SU}(n)$, let $B^{2n-2}=\{t\gamma :\left\Vert \gamma \right\Vert =1$, $t\in \left[ 0,\pi \right] \}$ and define the map $\psi :B^{2n-2}\times \left[ -\pi ,\pi \right] \rightarrow \mathrm{SU}\left( n\right) \subset K $ by $\psi \left( t\gamma ,\theta \right) =e^{tA_{\gamma }}e^{D_{\theta }}$ where $$D_{\theta }=\mathrm{diag} \left\{ i\theta ,-\frac{i}{n-1}\theta ,\ldots ,-\frac{i}{n-1}\theta \right\}.$$It follows that $\psi$ is the desired parametrization. If $U= \mathrm{Sp}(n)$, the map $\psi :B^{4n-4}\times B^{3}\rightarrow \mathrm{Sp}\left( n\right) \subset K $ that realizes the parametrization is defined by $\psi \left( t\gamma ,q\right) =e^{tA_{\gamma }}e^{D_{q}}$, where $B^{4n-4}=\{t\gamma \in \mathbb{H}^{n-1}:\left\Vert \gamma \right\Vert =1$, $t\in \left[ 0,\pi \right] \}$, $B^{3}=\{q\in i\mathbb{H}:\left\Vert q\right\Vert \leq \pi \}$ and $$D_{q}=\mathrm{diag}\left\{q,-\frac{1}{n-1}q,\ldots ,-\frac{1}{n-1}q\right\}.$$ From now on we consider the case when $\dim \mathfrak{g}_{\alpha }=1$ and $\mathfrak{g}_{2\alpha }=\{0\}$, so that $\mathfrak{g}(\alpha )\approx \mathfrak{sl}(2,\mathbb{R})$ and compact Lie algebra of $K_{\alpha }$ is $\mathfrak{so}\left( 2\right) $. This is the only relevant case to the computation of homology of the flag manifolds (c.f. Proposition [propappendix]{}). Let $\theta $ be the Cartan involution. Take $0\neq X_{\alpha }\in \mathfrak{g}_{\alpha }$ and $Y_{\alpha }=\theta (X_{\alpha })\in \mathfrak{g}_{-\alpha }$ such that $\langle X_{\alpha },Y_{\alpha }\rangle =\frac{2}{\langle \alpha ,\alpha \rangle }$. Hence, $[X_{\alpha },Y_{\alpha }]=H_{\alpha }^{\vee }=\frac{2H_{\alpha }}{\langle \alpha ,\alpha \rangle }$. Denote by $A_{\alpha }=X_{\alpha }+Y_{\alpha }\in \mathfrak{k}$. The Lie algebra $\mathfrak{g}(\alpha )=\mathfrak{g}_{-\alpha }\oplus \langle H_{\alpha }^{\vee }\rangle \oplus \mathfrak{g}_{\alpha }$ is isomorphic to $\mathfrak{s}\mathfrak{l}(2,\mathbb{R})$. Explicitly, write $\rho :\mathfrak{s}\mathfrak{l}(2,\mathbb{R})\rightarrow \mathfrak{g}(\alpha )$, with $\rho (H)=H_{\alpha }^{\vee }$, $\rho (X)=X_{\alpha }$ and $\rho (Y)=Y_{\alpha }$ where$$\begin{aligned} H=\left( \begin{array}{cc} 1 & 0 \\ 0 & -1\end{array}\right) ,\,\hspace{1cm}X=\left( \begin{array}{cc} 0 & -1 \\ 0 & 0\end{array}\right) ,\,\hspace{1cm}Y=\left( \begin{array}{cc} 0 & 0 \\ 1 & 0\end{array}\right) .\end{aligned}$$ This homomorphism extends to a homomorphism $\phi :\mathfrak{s}\mathfrak{l}(2,\mathbb{C})\rightarrow \mathfrak{g}_{\mathbb{C}}(\alpha )$. Note that $\mathrm{ad}\circ \phi $ is a representation of $\mathfrak{s}\mathfrak{l}(2,\mathbb{C})$ in $\mathfrak{g}_{\mathbb{C}}$. As $\mathrm{Sl}(2,\mathbb{C})$ is simply connected, this representation extends to a representation $\Phi $ of $\mathrm{Sl}(2,\mathbb{C})$ in $\mathfrak{g}_{\mathbb{C}}$ and they are related by $e^{\mathrm{ad}\circ \phi (X)}=\Phi (\exp (X))$ for any $X\in \mathfrak{s}\mathfrak{l}(2,\mathbb{C})$. We have $$\begin{aligned} e^{\mathrm{ad}(\pi A_{\alpha })}=e^{\mathrm{ad}\circ \phi (A)}=\Phi (\exp (\pi A)),\end{aligned}$$where $A=X+Y$. But in $\mathrm{Sl}(2,\mathbb{C})$ we have $$\begin{aligned} \exp (\pi A_{\alpha }^{\prime }) = \exp \left( \begin{array}{cc} 0 & -\pi \\ \pi & 0\end{array}\right) &=& \left( \begin{array}{cc} 1 & 0 \\ 0 & -1\end{array}\right) \\ &=& \exp \left( \begin{array}{cc} i\pi & 0 \\ 0 & -i\pi\end{array}\right) = \exp (i\pi H).\end{aligned}$$Therefore, $$\begin{aligned} e^{\mathrm{ad}(\pi A_{\alpha })} &=&\Phi (\exp ({i\pi H})) = e^{\mathrm{ad}\circ \phi (i\pi H)} \notag \\ &=&e^{\mathrm{ad}(i\pi H_{\alpha }^{\vee })}.\end{aligned}$$ Put $$\begin{aligned} m_{\alpha }=\exp (\pi iH_{\alpha }^{\vee })=\exp (\pi A_{\alpha }).\end{aligned}$$ Then $m_{\alpha }$ centralizes $A$ ($m_{\alpha }=\exp (\pi iH_{\alpha }^{\vee })$) and belongs to $K$ ($m_{\alpha }=\exp (\pi A_{\alpha })$). Hence $m_{\alpha }\in M=Z_{K}(\mathfrak{a})$. Now consider the curve $\gamma (t)=\exp (tA_{\alpha })\cdot b_{0}$ in the fiber of $\mathbb{F}\rightarrow G/P_{\alpha }$ through the origin. Since $m_{\alpha }\in M$, $\gamma (\pi )=m_{\alpha }b_{0}=b_{0}$. Actually $\gamma (t)$ covers the fiber in the interval $[0,\pi ]$. In $\mathfrak{s}\mathfrak{l}(2,\mathbb{R})$ we have that$$\begin{aligned} \mathrm{Ad}(e^{tA})H &=&\left( \begin{array}{cc} \cos t & -\sin t \\ \sin t & \cos t\end{array}\right) \left( \begin{array}{cc} 1 & 0 \\ 0 & -1\end{array}\right) \left( \begin{array}{cc} \cos t & \sin t \\ -\sin t & \cos t\end{array}\right) \\ &=&\left( \begin{array}{cc} \cos 2t & \sin 2t \\ -\sin 2t & \cos 2t\end{array}\right) .\end{aligned}$$That is, $\mathrm{Ad}(e^{tA})H=-\sin 2tX+\cos 2tH+\sin 2tY$. Applying the formula $$\rho \left( \mathrm{Ad}(e^{tA})H\right) =\mathrm{Ad}(e^{tA})H_{\alpha }^{\vee }$$ we get $\mathrm{Ad}(e^{tA})H_{\alpha }^{\vee }=-\sin 2tX_{\alpha }+\cos 2tH_{\alpha }^{\vee }+\sin 2tY_{\alpha }.$ This shows that $e^{tA}$ centralizes $H_{\alpha }^{\vee }$ if and only if $t=n\pi $. In particular, $e^{tA}\in M$ if and only if $t=n\pi $. Hence, the period of $\gamma $ is exactly $\pi $. Summarizing, \[compact\_parametrization\_1\] The one-dimensional version of the Lemma [parametrization]{} is realized by $$\begin{aligned} \psi :[0,\pi ]\rightarrow K_{\alpha }\,,\,t\mapsto \exp (tA_{\alpha }).\end{aligned}$$In particular, $\psi (0)=1$ and $\psi (\pi )=m_{\alpha }=\exp (\pi A_{\alpha })$. Moreover, if $X\in \mathfrak{g}_{\beta }$, then: $$\begin{aligned} \mathrm{Ad}(m_{\alpha })(X)=\mathrm{Ad}(\exp (\pi iH_{\alpha }^{\vee }))(X)=e^{\mathrm{ad}(\pi iH_{\alpha }^{\vee })}(X)=e^{\pi i\epsilon (\alpha ,\beta )}(X),\end{aligned}$$where $\epsilon (\alpha ,\beta )=\frac{2\langle \alpha ,\beta \rangle }{\langle \alpha ,\alpha \rangle }$ is the Killing number. This implies that \[compact\_parametrization\_2\] The root spaces $\mathfrak{g}_{\beta }$ are invariant by the action of $\mathrm{Ad}(m_{\alpha })$ and $$\begin{aligned} \mathrm{Ad}(m_{\alpha })_{\left\vert \mathfrak{g}_{\beta }\right. }=(-1)^{\epsilon (\alpha ,\beta )}\mathrm{id}.\end{aligned}$$ Gluing cells ------------ A Schubert cell $\mathcal{S}_{w}$ is obtained from smaller cells $\mathcal{S}_{v}$, $v<w$, by gluing a cell of dimension $\dim (N\cdot wb_{0})$. Once this proccess is done for each $w\in \mathcal{W}$, we get a cellular decomposition for $\mathbb{F}$ which is explictly given by characterisc maps and attaching maps. (We follow the terminology of Hatcher [@Hat02]: the *characteristic map* is defined in a closed ball while the *attaching map* is the restriction characteristic map to the boundary of the ball.) In order to define a characteristic map for $\mathcal{S}_{w}$, $w\in \mathcal{W}$, we must choose a reduced expression $$w=r_{1}\cdots r_{n}$$as product of simple reflections $r_{i}=r_{\alpha _{i}}$. We know that $\mathcal{S}_{w}=K_{1}\cdots K_{n}\cdot b_{0}$. By Lemma \[parametrization\], for each $i$, there exists $\psi _{i}:B^{d_{i}}\rightarrow K_{i}$, where $d_{i}$ is the dimension of the fiber of $\mathbb{F}\rightarrow \mathbb{F}_{i} $, that is, the dimension of the flag of $G(\alpha _{i})$. Let $B_{w}=B^{d_{1}}\times \cdots \times B^{d_{n}}$ be the ball with dimension $d=d_{1}+\cdots +d_{n}$. Then the characteristic map $\Phi _{w}:B_{w}\rightarrow \mathbb{F}$ is defined by $$\Phi _{w}(t_{1},\ldots ,t_{n})=\psi _{1}(t_{1})\cdots \psi _{n}(t_{n})\cdot b_{0}.$$ **Remark:** Distinct decompositions of $w$ yields different characteristic maps, so the notation $\Phi _{w}$ should include the minimal decomposition of $w$ (for example, $\Phi _{r_{1}\cdots r_{n}}$). To keep this simpler notation we shall fix later a choice of minimal expressions for each $w\in \mathcal{W}$. \[characteristic\_map\] Let $w=r_{1}\cdots r_{n}$ be a minimal decomposition. Let $\Phi _{w}:B_{w}\rightarrow \mathbb{F}$ be the map defined above and take $\mathbf{t}=(t_{1},\ldots ,t_{n})\in B_{w}$. Then, $\Phi _{w}$ is a characteristic map for $\mathcal{S}_{w}$, that is, 1. $\Phi_w (B_w) \subset \mathcal{S}_w$. 2. $\Phi _{w}(\mathbf{t})\in \mathcal{S}_{w}\setminus N\cdot wb_{0}$ if and only if $\mathbf{t}\in \partial B_{w}=S^{d-1}$. 3. $\Phi |_{B_{w}^{\circ }}:B_{w}^{\circ }\rightarrow N\cdot wb_{0}$ is a diffeormorphism ($B_{w}^{\circ }$ is the interior of $B_{w}$). The first condition holds by construction since $\psi_i(t_i) \in K_i$ and hence $\Phi _{w}(t_{1},\ldots ,t_{n})=\psi _{1}(t_{1})\cdots \psi _{n}(t_{n})\cdot b_{0}\in K_{1}\cdots K_{n}\cdot b_{0}=\mathcal{S}_{w}$. The second statement follows as a consequence of the Proposition \[propSchubert\_3\] by which we have that $\psi _{1}(t_{1})\cdots \psi _{n}(t_{n})$ is not in $N\cdot wb_{0}$ if and only if some $\psi _{i}(t_{i})\in M$. By Lemma [compact\_parametrization\_1]{}, this implies that $t_{i}\in \{0,\pi \}=\partial S^{d_{i}-1}$, that is, $\mathbf{t}\in \partial S^{d-1}$. Finally, we already have that $\Phi |_{B_{w}^{\circ }}$ is a surjective map. Let us prove its injectivity by induction on the lenght $l\left( w\right) $ of $w$. If $l\left( w\right) =1$, this is Lemma [compact\_parametrization\_1]{}. For $l\left( w\right) >1$, suppose that $\Phi _{w}\left( \mathbf{t}\right) =\Phi _{w}\left( \mathbf{s}\right) $ with $\mathbf{t}=\left( t_{1},\ldots ,t_{n}\right) $ and $\mathbf{s}=\left( s_{1},\ldots ,s_{n}\right) $ in $B_{w}^{\circ }$. Then we claim that $x=y$ where $$\begin{aligned} x &=&\psi _{1}(t_{1})\cdots \psi _{n-1}(t_{n-1})\cdot b_{0} \\ y &=&\psi _{1}(s_{1})\cdots \psi _{n-1}(s_{n-1})\cdot b_{0}.\end{aligned}$$In fact, the elements $\psi _{i}(t_{i})$ and $\psi _{i}(s_{i})$ are not in $M $, hence by Proposition \[propSchubert\_3\] both $x,y\in N\cdot vb_{0}$, $v=r_{1}\cdots r_{n-1}$. Also, $\pi _{n}\left( x\right) =\pi _{n}\left( \Phi _{w}\left( \mathbf{t}\right) \right) =\pi _{n}\left( \Phi _{w}\left( \mathbf{s}\right) \right)= \pi _{n}\left( x\right) $, that is, $x$ and $y$ belong to the same fiber of $\pi _{n}:\mathbb{F}\rightarrow \mathbb{F}_{n}$. It follows by Lemma \[Schubert\_lema\] that $x=y$ since $N\cdot vb_{0}$ meets each fiber of $\pi _{n}$ in a unique point. By the induction hypothesis $(t_{1},\ldots ,t_{n-1})=(s_{1},\ldots ,s_{n-1})$, so that $\psi _{1}(t_{1})\cdots \psi _{n-1}(t_{n-1})=\psi _{1}(s_{1})\cdots \psi _{n-1}(s_{n-1})$. Applying this to the equality $\Phi _{w}\left( \mathbf{t}\right) =\Phi _{w}\left( \mathbf{s}\right) $ we conclude that $\psi _{n}(t_{n})\cdot b_{0}=\psi _{n}(s_{n})\cdot b_{0}$, which in turn implies that $t_{n}=s_{n}$, $l\left( r_{n}\right) =1$. Therefore, $\Phi _{w}$ is a closed continuous and bijective map, hence it is a homeomorphism (differentiability comes from the construction of the maps $\psi _{i}$). As a consequence of the last item of the above proposition, we have the following construction. Let $d=\dim \mathcal{S}_{w}=\dim N\cdot wb_{0}$. The sphere $S^{d}$ is the quotient $B_{w}/\partial (B_{w})$ where the boundary is collapsed to a point. We can do the same with the Schubert cell $\mathcal{S}_{w}$. Define $\sigma _{w}=S_{w}/(S_{w}\setminus N\cdot wb_{0})$, i.e., the space obtained by identifying the complement of the Bruhat cell $\mathcal{S}_{w}\setminus N\cdot wb_{0}$ in $\mathcal{S}_{w}$ to a point. As $\Phi _{w}(\partial (B_{w}))\subset S_{w}\setminus N\cdot wb_{0}$, it follows that $\Phi _{w}$ induces a map $S^{d}\rightarrow \sigma _{w}$ which is a homeomorphism. The inverse of this homeomorphism will be denoted by $$\Phi _{w}^{-1}:\sigma _{w}\rightarrow S^{d} \label{Sphere}$$(although this is not the same as the inverse of $\Phi _{w}$). A very useful data is the determination of pairs $w,w^{\prime }\in \mathcal{W}$ for which $w^{\prime} \leq w$ and $\dim \mathcal{S}_{w}-\dim \mathcal{S}_{w^{\prime }}=1$. \[propappendix\] Let $w,w^{\prime }\in \mathcal{W}$. The following statements are equivalent. 1. $\mathcal{S}_{w^{\prime }}\subset \mathcal{S}_{w}$ and $\dim \mathcal{S}_{w}-\dim \mathcal{S}_{w^{\prime }}=1$. 2. If $w=r_{1}\cdots r_{n}$ is a reduced expression of $w\in \mathcal{W}$ as a product of simple reflections, then \(i) $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ is a reduced expression. \(ii) If $r_{i}=r_{\alpha _{i}}$ then $\mathfrak{g}(\alpha _{i})\cong \mathfrak{s}\mathfrak{l}(2,\mathbb{R})$. This is the same as saying the fiber of $\mathbb{F}\rightarrow \mathbb{F}_{i}$ has dimension $1$. In fact, $\mathcal{S}_{w^{\prime }}\subset \mathcal{S}_{w}$ if and only if $w^{\prime }<w$ in the Bruhat-Chavalley order. In this case if $w=r_{1}\cdots r_{n}$ and $w^{\prime }=r_{i_{1}}\cdots r_{i_{j}}$ are reduced expressions then $\dim \mathcal{S}_{w}$ is $\dim \mathcal{S}_{w^{\prime }}$ plus the sum of the multiplicities of the roots missing in the reduced expression for $w^{\prime }$. Hence $\dim \mathcal{S}_{w}-\dim \mathcal{S}_{w^{\prime }}=1$ if and only if $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ and $\alpha _{i}$ has multiplicity $1$, that is, $\mathfrak{g}(\alpha _{i})\cong \mathfrak{s}\mathfrak{l}(2,\mathbb{R})$. **Remark:** Given $w^{\prime }$ as above, the decomposition $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ is unique. In fact, if $w=r_{1}\cdots r_{i}\cdots r_{j}\cdots r_{n}$ and $w^{\prime }=r_{1}\cdots r_{i}\cdots \hat{r_{j}}\cdots r_{n}$ then $r_{i+1}\cdots r_{j}=r_{i}\cdots r_{j-1}$ which cannot happen (see [@San10], Chapter 9). Gradient flow\[secgrad\] ------------------------ It is known that the vector field $\widetilde{H}$ with flow $\exp tH$ induced on a flag manifold $\mathbb{F}_{\Theta }$ by a regular element $H\in \mathfrak{a}^{+}$ is the gradient of a Morse function. For this flow, the singularities are $wb_{0}$, $w\in \mathcal{W}$, whose unstable and stable manifolds are Bruhat cells $W^{u}(wb_{0})=N\cdot wb_{0}$ and $W^{s}(ub_{0})=N^{-}\cdot ub_{0}$ (see [@DKV83], for details). Below we describe these orbits in terms of characteristic maps of the cellular decomposition constructed above. Take $w=r_{1}\cdots r_{n}\in \mathcal{W}$ with the characteristic map $\Phi _{w}$ and let $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ such that $\mathcal{S}_{w^{\prime}}\subset \mathcal{S}_{w}$ and $\dim \mathcal{S}_{w^{\prime }}=\dim \mathcal{S}_{w}-1$, by Proposition \[propappendix\]. Note that by construction $wb_{0}=\Phi _{w}(\pi/2,\ldots ,\pi/2,\ldots ,\pi/2)$, that is, $wb_0$ is the image of the center of the cube. Now, consider the path $$\phi _{i}(t)=\Phi _{w}\left(\pi/2,\ldots ,t,\ldots ,\pi/2 \right)\qquad \, , \, t\in \lbrack 0,\pi ],$$where $t$ is in the $i$-th position. Then $\phi (\pi/2)=w\cdot b_{0} $, $\phi (0)=w^{\prime }\cdot b_{0}$ comes from the $0$-face and $\phi (\pi )=w^{\prime }\cdot b_{0}$ comes from the $\pi $-face. Below we prove that the two pieces of $\phi _{i}\left( t\right) $, from $\pi /2$ to $\pi $ and from $0$ to $\pi /2$ (in reversed direction) are the two gradient lines joining the singularities $w\cdot b_{0}$ and $w^{\prime }\cdot b_{0}$. In what follows, we write $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}=u\cdot v$, i.e., $u=r_{1}\cdots r_{i-1}$ and $v=r_{i+1}\cdots r_{n}$. Put $X_{\beta }=\mathrm{Ad}(u)X_{\alpha _{i}}$, $Y_{\beta }=\theta (X_{\beta })$ and $A_{\beta }=X_{\beta }+Y_{\beta }$. \[negativo\] With the above notation, we have that $$\phi _{i}(t)=\exp (sA_{\beta })wb_{0}$$ where $s=t+\pi/2\in \lbrack -\pi/2,\pi/2]$ if $t\in \lbrack 0,\pi]$. The result is a consequence of the following computation. $$\begin{aligned} \ \phi _{i}(t) &=&u\exp (tA_{i})vb_{0} = u\exp (tA_{i})r_{i}r_{i}vb_{0} \\ &=&u\exp (tA_{i})\cdot \exp ((\pi/2)A_{i})r_{i}vb_{0} \\ &=&\exp \left( \left(t+\pi/2 \right)\mathrm{Ad}(u)A_{i}\right) wb_{0} \\ &=&\exp (sA_{\beta })wb_{0}\end{aligned}$$ where $\beta =u\alpha _{i}$ implies that $A_{\beta }=\mathrm{Ad}(u)A_{i}$. Let us consider $\phi _{i}(s)=\exp (sA_{\beta })wb_{0}$. It follows that $\phi _{i}(0)=wb_{0}$, $\phi _{i}(\pm \pi/2)=w^{\prime }b_{0}$. $\exp(t Y_{\beta})wb_0 = wb_0$. The main idea is translate to the origin. That is $$\exp (tX_{-\beta })wb_{0}=w(w^{-1}\exp (tX_{-\beta }))wb_{0}=w\exp (tAd(w^{-1}X_{-\beta }))b_{0}.$$The root $\beta $ is positive while $w^{-1}\beta $ is negative. As $u\left( \mathfrak{g}_{\alpha }\right) =\mathfrak{g}_{u\alpha }$ we have that $ \mathrm{Ad}(w^{-1}X_{-\beta })\in \mathfrak{g}_{-w^{-1}\beta }$. But since $ w^{-1}\beta $ is a negative root, it follows that $-w^{-1}\beta $ is positive. Hence, it belongs to $\mathfrak{n}^{+}\subset \mathfrak{p}$ and therefore $\exp (tAd(w^{-1}X_{-\beta }))b_{0}=b_{0}$ for all $t\in \mathbb{R}$. Let $t=\tan (s),r=-\sin (s)\cos (s),\lambda ={\cos (s)}^{-1}$. Hence $$\phi _{i}(s)=e^{tX_{\beta }}e^{rY_{\beta }}e^{\log (\lambda )H_{\beta }^{\vee }}wb_{0}.$$ Note however that both matrices $e^{\log (\lambda )H_{\beta }^{\vee }}$ and $e^{rY{\beta }}$ fix $wb_{0}$. This gives that $$\phi _{i}(s)=e^{\tan (s)X_{\beta }}wb_{0}\,,\,s\in \left( -\pi/2,\pi/2\right). \label{fluxo}$$Finally, we have that $$\lim_{t\rightarrow \pm \infty }\exp ^{tX_{\beta }}wb_{0}=\phi _{i}(\pm \pi/2)=w^{\prime }b_{0}.$$ From (\[fluxo\]) we get the behaviour of the gradient flow $h^{t}=\exp tH$ with $H\in \mathfrak{a}^{+}$. Let $s \neq 0$. It is easy to see that $h^t (\phi_i(s)) = \exp (\tan(s) e^{t \beta(H)}X_\beta) \cdot wb_0$. This may be written as $h^t (\phi_i(s)) = \phi_i( s^{\prime })$ with $s^{\prime} = \arctan (\tan(s) e^{t \beta(H)})$. Hence, we conclude that the gradient flow leaves the path $\phi_i$ invariant. Observe that $\beta$ is a positive root. So $e^{t \beta(H)}X_\beta \to 0$ as $t \to -\infty$ and hence $s^{\prime }\to 0$, i.e., $\lim_{t \to - \infty} h^t \phi_i (s) = \phi_i(0) = wb_0 .$ When $t\rightarrow +\infty $ it follows that $\tan (s)e^{t\beta (H)}X_{\beta }\rightarrow \pm \infty $ depending only in the sign of $\tan (s)$. Hence $s^{\prime }\rightarrow \pm \pi/2$, i.e., $\lim_{t \to + \infty} h^t \phi_i (s) = \phi_i(\pm \pi/2) = w^{\prime }b_0.$ Thus we get the desired result. $\phi_i(s)$ give the two gradient flow lines between $wb_0$ and $w^{\prime }b_0$. One of them belongs to the interval $s \in (-\pi/2, 0)$ while the other belongs to the interval $s \in (0, \pi/2)$. Homology of maximal flag manifolds\[sechommax\] =============================================== The cellular homology of a CW complex is defined from a cellular decomposition of the complex and is isomorphic to the singular homology of the space. It means that the homology group does not depend on the choice of the cellular decomposition, although the boundary operator may change according to the choice of the cellular decomposition, i.e., the way the cells are glued. In view of that, we **fix** once and for all reduced expressions $$w=r_{1}\cdots r_{n}$$as a product of simple reflections, for each $w\in \mathcal{W}$. After making these choices we define as before the characteristic map $\Phi _{w} $, $w\in \mathcal{W}$, that glues the ball $B_{w}$ in the union of Schubert cells $\mathcal{S}_{u}$ with $u<w$. The boundary map ---------------- We recall (in our context) the definition of the cellular boundary maps giving the homology with coefficients in a ring $R$ (see [Hat02]{}). Let $\mathcal{C}$ be the $R$-module freely generated by $\mathcal{S}_{w}$, $w\in \mathcal{W}$. The boundary maps $\partial :\mathcal{C}\rightarrow \mathcal{C}$ are defined by $$\partial \mathcal{S}_{w}=\sum_{w^{\prime }}c(w,w^{\prime })\mathcal{S}_{w^{\prime }}$$where the coefficients $c(w,w^{\prime })\in R$  satisfy the properties: 1. $c(w,w^{\prime })=0$ in case $\dim \mathcal{S}_{w}-\dim \mathcal{S}_{w^{\prime }}\neq 1$. 2. If $\dim \mathcal{S}_{w}-\dim \mathcal{S}_{w^{\prime }}=1$ then $c(w,w^{\prime })=\deg \left( \phi _{w,w^{\prime }}:S_{w}^{d-1}\rightarrow S_{w^{\prime }}^{d-1}\right) $, where $\phi _{w,w^{\prime }}$ is the composition of the following maps: 1. The attaching map: $\Phi _{w}|_{\partial }(B_{w}^{d}):S_{w}^{d-1}=\partial (B_{w}^{d})\rightarrow \mathcal{S}_{w}\setminus N\cdot wb_{0}=\cup _{u<w}\mathcal{S}_{u}=X^{d-1}$, where $X^{d-1}$ denotes the $(d-1)$-squeleton of $\mathcal{S}_{w}$. 2. The quotient map $X^{d-1}\rightarrow X^{d-1}/(X^{d-1}\setminus \mathcal{S}_{w^{\prime }})$ where we take the cell $\mathcal{S}_{w^{\prime }} $ inside $X^{d-1}$ and identify its complement in $\mathcal{S}_{w^{\prime }}$ to a point. 3. The identification: $X^{d-1}/(X^{d-1}\setminus \mathcal{S}_{w^{\prime }})\cong \mathcal{S}_{w^{\prime }}/(\mathcal{S}_{w^{\prime }}\setminus N\cdot w^{\prime }b_{0})$ which are in the same space. This last one is $\mathcal{S}_{w^{\prime }}/(\mathcal{S}_{w^{\prime }}\setminus N\cdot w^{\prime }b_{0})=\sigma _{w^{\prime }}$ by definition. 4. $\Phi _{w^{\prime }}^{-1}:\sigma _{w^{\prime }}\rightarrow S_{w^{\prime }}^{d-1}$. This is the map defined in (\[Sphere\]). **Remark:** There is a subtlety which must be emphasized: $\phi _{w,w^{\prime }}$ is a map $S^{d-1}\rightarrow S^{d-1}$ whose domain is the boundary of a ball in some $\mathbb{R}^{N}$ (the ball $B_{w}$) and, hence, it is a canonically defined sphere. However, the codomain is the space $\sigma _{w^{\prime }}$ which is homeomorphic to $S^{d-1}$. To get the boundary map a homeomorphism $\sigma _{w^{\prime }}\rightarrow S^{d-1}$ must be fixed beforehand, since distinct homeomorphisms may yield maps with distinct degrees. Here is where it is needed to choose in advance the reduced expressions of $w\in \mathcal{W}$. To compute the the degree $c(w,w^{\prime })=\deg \left( \phi _{w,w^{\prime }}:S_{w}^{d-1}\rightarrow S_{w^{\prime }}^{d-1}\right) $ when $w=r_{1}\cdots r_{n}$ and $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ are minimal decompositions we proceed with the following steps. ### Step 1: Domain and codomain spheres {#step-1-domain-and-codomain-spheres .unnumbered} First we identify the spheres $S_{w}^{d-1}$ in the domain and $S_{w^{\prime }}^{d-1}$ in the codomain. Remember that $B_{w}=B^{d_{1}}\times \cdots \times B^{d_{n}}$ where $B^{d_{i}}$ is the $1$-dimensional choosen to be the interval $[0,\pi ]$, as in the construction of Lemma \[compact\_parametrization\_1\]. The dimension of $B_{w}$ is $d=d_{1}+\cdots +d_{n}$ and the domain of $\phi _{w,w^{\prime }}$ is $$S_{w}^{d-1}=\partial (B_{w})=\{(t_{1},\ldots ,t_{n}):\exists j,t_{j}\in \partial B^{d_{j}}\}$$the union of faces" of $B_{w}$. On the other hand let $B_{w^{\prime }}=B^{d_{1}}\times \cdots \times \hat{B^{d_{i}}}\times \cdots \times B^{d_{n}}$. Then codomain is the sphere $S_{w^{\prime }}^{d-1}$ obtained by collapsing to a point the boundary of $B_{w^{\prime }}$. This is seen by items (c) and (d) in the above definition of $\partial $. ### Step 2: $\protect\sigma _{w^{\prime }}$ in the image $\Phi _{w}(S_{w}^{d-1})$. {#step-2-protectsigma-_wprime-in-the-image-phi_ws_wd-1. .unnumbered} The second step is to see how $\sigma _{w^{\prime }}$ sits inside the image $\Phi _{w}(S_{w}^{d-1})$. The following lemma says how is the pre-image of $N\cdot w^{\prime }b_{0}$ under $\Phi _{w}$. $\Phi _{w}(t_{1},\ldots ,t_{n})\in N\cdot w^{\prime }b_{0}$ if and only if $t_{j}\in (B^{d_{j}})^{\circ }$, $j\neq i$ and $t_{i}\in \partial B^{d_{i}}$, that is, $t_{i}=0$ or $\pi $. If $t_{i}\in \partial B^{d_{i}}$ then $\psi _{i}(t_{i})\in M$ by Lemma [compact\_parametrization\_1]{}. This implies that $$\Phi _{w}(t_{1},\ldots ,t_{n})=\psi _{1}(t_{1})\cdots \psi _{n}(t_{n})\cdot b_{0}\in K_{1}\cdots \hat{K_{i}}\cdots K_{n}=\mathcal{S}_{w^{\prime }}$$ since $M\subset K_{s}$ for all sub-index $s$. By Proposition [propSchubert\_3]{}, we see that $\Phi _{w}(t_{1},\ldots ,t_{i},\ldots ,t_{n})\in N\cdot w^{\prime }b_{0}$ if and only if $\psi _{j}(t_{j})\notin M$ for $j\neq i$, which in turn is equivalent to $t_{j}\in (B^{d_{j}})^{\circ } $, $i\neq j$, by Lemma \[compact\_parametrization\_1\]. In other words the pre-image $\Phi _{w}^{-1}\left( N\cdot w^{\prime }b_{0}\right) \subset B_{w}$ is the union of the interior of the two faces corresponding to the $i$-th coordinate, that is, the faces where $t_{i}=0$ and $t_{i}=\pi $, respectively. In the quotient $\sigma _{w^{\prime }}=\mathcal{S}_{w^{\prime }}/(\mathcal{S}_{w^{\prime }}\setminus N\cdot w^{\prime }b_{0})$ the faces of $\partial B_{w}$ corresponding to the $j$-th coordinates, $j\neq i$ are collapsed to a point. ### Step 3: Degrees {#step-3-degrees .unnumbered} The degree of $\phi _{w,w^{\prime }}$ is the sum of the degree of two maps, namely the maps obtaining by restricting to the faces$$\mathcal{F}_{0}^{i}=\{(t_{1},\ldots ,0,\ldots ,t_{n})\}\quad \mathrm{and}\quad \mathcal{F}_{\pi }^{i}=\{(t_{1},\ldots ,\pi ,\ldots ,t_{n})\}.$$The values of $\phi _{w,w^{\prime }}$ in these faces are given by $$\begin{aligned} f_{i}^{0}\left( \mathbf{t}\right) &=&\Phi _{w^{\prime }}^{-1}\left( \psi _{1}(t_{1})\cdots \psi _{i}(0)\cdots \psi _{n}(t_{n})\cdot b_{0}\right) \\ &=&\Phi _{w^{\prime }}^{-1}\left( \psi _{1}(t_{1})\cdots 1\cdots \psi _{n}(t_{n})\cdot b_{0}\right) . \\ f_{i}^{\pi }\left( \mathbf{t}\right) &=&\Phi _{w^{\prime }}^{-1}\left( \psi _{1}(t_{1})\cdots \psi _{i}(\pi )\cdots \psi _{n}(t_{n})\cdot b_{0}\right) \\ &=&\Phi _{w^{\prime }}^{-1}\left( \psi _{1}(t_{1})\cdots m_{\alpha _{i}}\cdots \psi _{n}(t_{n})\cdot b_{0}\right) .\end{aligned}$$where $\mathbf{t}=\left( t_{1},\ldots ,\widehat{t_{i}},\ldots ,t_{n}\right) $ and $\Phi _{w^{\prime }}$ is given by a choice of a reduced expression $w^{\prime }=s_{1}\cdots s_{m}$ (choosed in advance) which may be different from the reduced expression $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$. The degree of $\phi _{w,w^{\prime }}$ is the sum of the degrees of $f_{i}^{0} $ and $f_{i}^{\pi }$ which may be considered as maps $S^{d-1}\rightarrow S^{d-1}$ by collapsing the boundary to points of the faces. Now, the degree of a map $\varphi $ can be computed as the sum of the local degrees in the inverse image of $\varphi ^{-1}(\xi )$ which has a finite number of points (see [@Hat02], Proposition 2.30). In the case of our map $\phi _{w,w^{\prime }}$, the maps $f_{i}^{0}$ and $f_{i}^{\pi }$ are homeomorphisms so that pre-image $\phi _{w,w^{\prime }}^{-1}(\xi )$ of a generic point has two points. Namely a point $x_{1}$ in the the face $\mathcal{F}_{0}^{i}$ and another one $x_{2}$ in the face $\mathcal{F}_{\pi }^{i}$. The local degree at $x_{1}$ is the degree of $f_{i}^{0}$ since $f_{i}^{0}$ is a homeomorphism. The same the local degree at $x_{2}$ is the degree of $f_{i}^{\pi }$. Finally the degrees of $f_{i}^{0}$ and $f_{i}^{\pi }$ are $\pm 1$ since they are homeomorphisms. **Summarizing:** To get the degree of $\phi _{w,w^{\prime }}$ we must restrict $\Phi _{w^{\prime }}^{-1}\circ \Phi _{w}$ to the faces $\mathcal{F}_{0}^{i}$ and $\mathcal{F}_{\pi }^{i}$ and view these faces as spheres (with the boundaries collapsed to points). The sum of the degrees of these two restrictions is the degree of $\phi _{w,w^{\prime }}$. The restrictions of $\Phi _{w^{\prime }}^{-1}\circ \Phi _{w}$ to the faces $\mathcal{F}_{0}^{i}$ and $\mathcal{F}_{\pi }^{i}$ are homeomorphisms and hence have degree $\pm 1$. It follows that the total degre of $\phi _{w,w^{\prime }}$ is $0$ or $\pm 2$. This is one of the main results on the homology of flag manifolds. The coefficient $c(w,w^{\prime })=\deg (f_{i}^{0})+\deg (f_{i}^{\pi })=0$ or $\pm 2$, for any $w,w^{\prime }\in \mathcal{W}$. In particular, in the case of $\mathbb{Z}_{2}$ coefficients all boundary maps vanish. The homology of $\mathbb{F}$ over $\mathbb{Z}_{2}$ is a vector space of dimension $|\mathcal{W}|$. **Remark:** The above computations are particularly interesting when the simple root $\alpha _{i}$ has multiplicity $\dim \mathfrak{g}_{\alpha _{i}}=1$. If all the simple roots have multiplicity $\geq 2$ then the boundary operator $\partial $ is identically zero and homology is freely generated by the Schubert cells. This happens in the classical case of the complex Lie algebras, where any root has (real) multiplicity two. An example of a real Lie algebra where the simple roots have multiplicity $\geq 2$ is the real form of $\mathfrak{sl}\left( n,\mathbb{C}\right) $ whose Satake diagram is (170,25) (0,0) (0,15) (3,15)[(1,0)[20]{}]{} (26,15) (29,15)[(1,0)[20]{}]{} (52,15) (55,15)[(1,0)[17]{}]{} (72,15) (24,3) (7,0)[(1,0)[2]{}]{} (11,0)[(1,0)[2]{}]{} (14,0)[(1,0)[2]{}]{} (96,15)[(1,0)[17]{}]{} (116,15) (119,15)[(1,0)[20]{}]{} (142,15) (145,15)[(1,0)[20]{}]{} (168,15) In this case the simple roots are complex and hence their multiplicities are $\geq 2$. Illustration\[subsecsl3\] ------------------------- In order to illustrate the above description of the boundary operator $\partial $ we consider here the maximal flag manifold $\mathbb{F}$ of the split real form of $\mathfrak{sl}\left( 3,\mathbb{R}\right)$. In this case, the Weyl group is $S_{3}$, the permutation group in three elements. The simple reflections are $(12)=r_{\alpha _{1,2}}=r_{1}$ and $(23)=r_{\alpha _{2,3}}=r_{2}$. Only $(13)$ has two reduced expressions: $(13)=(12)(23)(12)$ and $(13)=(23)(12)(23)$. We fix the following minimal decompositions $$1,(12),(23),(123)=(12)(23),(132)=(23)(12),(13)=(12)(23)(12).$$Let $A=E_{1,2}-E_{2,1}$ and $B=E_{2,3}-E_{3,2}$ be the matrices whose exponentials provide parametrizations for the compact groups $K_{1}$ and $K_{2}$ respectively. With these choices, the characteristic maps are 1. $\Phi_1 (0) = b_0$. 2. $\Phi_{(12)}(t)= e^{tA} \cdot b_0$, $t \in [0,\pi]$. 3. $\Phi_{(23)}(t)= e^{tB} \cdot b_0$, $t \in [0,\pi]$. 4. $\Phi_{(123)}(t,s)=e^{tA}e^{sB}\cdot b_0$, $(t,s) \in [0,\pi]^2$. 5. $\Phi_{(132)}(t,s)=e^{tB}e^{sA}\cdot b_0$, $(t,s) \in [0,\pi]^2$. 6. $\Phi_{(13)}(t,s,z)=e^{tA}e^{sB}e^{zA}\cdot b_0$, $(t,s,z) \in [0,\pi]^3$. Then we obtain expressions for $c(w,w^{\prime })$. 1. $c((12),1)=0$ and $c((23),1)=0$ since there is a unique $0$-cell. 2. $c((123),(12))=0$. Note that $(12)=(12)\widehat{(23)}$. So we need to consider the degree of the two maps $f_{2}^{0},f_{2}^{\pi }:S^{1}\rightarrow S^{1}$ defined by $f_{2}^{0}(t,0)=e^{tA}e^{0B}\cdot b_{0}=e^{tA}\cdot b_{0}$ and $f_{2}^{\pi }(t,\pi )=e^{tA}e^{\pi B}\cdot b_{0}=e^{tA}\cdot b_{0}$. Their degrees are obtained by comparing the orientation in the respective face of the boundary of the cube $[0,\pi ]^{2}$, which is $S^{1}$ oriented counter-clockwise, with the orientation given by the attaching map $e^{tA}\cdot b_{0}$. Hence, the degree of $f_{2}^{0}$ is $1$ since as $t$ increases the curve $(t,0)$ and $e^{tA}\cdot b_{0}$ go in the same direction as $\Phi _{(12)}$. On the other hand the degree of $f_{2}^{\pi }$ is $-1$ since as $t$ increases, the curve $(t,0)$ and the image $e^{tA}\cdot b_{0}$ are in opposite directions. Hence $c((123),(12))=+1+(-1)=0$. 3. $c((123),(23))=-2$. Here $(23)=\widehat{(12)}(23)$. So we consider the degree of the two maps $f_{1}^{0},f_{1}^{\pi }:S^{1}\rightarrow S^{1}$ given by $f_{1}^{0}(0,s)=e^{0A}e^{sB}\cdot b_{0}=e^{sB}\cdot b_{0}$ and $f_{1}^{\pi }(\pi ,s)=e^{\pi A}e^{sB}\cdot b_{0}=\exp (s\mathrm{Ad}(e^{\pi A})B)\cdot b_{0}=e^{-sB}\cdot b_{0}$ because $\mathrm{Ad}(e^{\pi A})B=-B$. Since $e^{-sB}\cdot b_{0}=e^{(\pi -s)B}e^{\pi B}\cdot b_{0}=e^{(\pi -s)B}\cdot b_{0}$, the function $f_{1}^{\pi }$ defined in $[0,\pi ]$ is given by $f_{1}^{\pi }(0,s)=e^{(\pi -s)B}\cdot b_{0}$. Hence the degree of $f_{1}^{0}$ is $-1$ as above. The degree of $f_{1}^{\pi } $ is also $-1$ since it is the degree of the function $s\mapsto \pi -s$. Hence $c((123),(12))=(-1)+(-1)=-2$. 4. $c((132),(121))=-2$ and $c((132),(23))=0$, which can be seen the same as above. 5. $c((13),(123))=0$. Note that $(123)=(12)(23)\widehat{(12)}$. So we consider the maps $f_{3}^{0},f_{3}^{\pi }:S^{2}\rightarrow S^{2}$. 1. $f_{3}^{0}(t,s,0)=e^{tA}e^{sB}e^{0A}\cdot b_{0}=e^{tA}e^{sB}\cdot b_{0} $, and we have $$\deg f_{3}^{0}=-1.$$In fact, the boundary of the cube $[0,\pi ]^{3}$ is $S^{2}$ oriented with the normal vector pointing outwards. The face $(t,s,0)$ (in this order) when viewed in the domain is negatively orientated while in the codomain the orientation agrees with the $S^{2}$ orientation. Hence the degree is $-1$. 2. $f_{3}^{\pi }(t,s,\pi )=e^{tA}e^{sB}e^{\pi A}\cdot b_{0}=e^{tA}e^{sB}\cdot b_{0}$ with $$\deg f_{3}^{\pi }=1.$$In this case the face $(t,s,\pi )$ agrees with the positive orientation. 6. $c((13),(132))=0$. Note that $(132)=\widehat{(12)}(23)(12)$. So we consider the maps $f_{1}^{0},f_{1}^{\pi }:S^{2}\rightarrow S^{2}$. 1. $f_{1}^{0}(0,t,s)=e^{0A}e^{tB}e^{sA}\cdot b_{0}=e^{tB}e^{sA}\cdot b_{0} $ with $$\deg f_{3}^{0}=-1.$$In this case the face $(0,t,s)$ (in this order) in the domain hasa negative orientation while in the codomain the orientation agrees with the positive one. Hence the degree is $-1$. 2. $f_{1}^{\pi }(\pi ,t,s)=e^{\pi A}e^{tB}e^{sA}\cdot b_{0}=\exp (-tB)e^{sA}\cdot b_{0}$ since $Ad(e^{\pi A}B)=-B$ and $Ad(e^{\pi A}A)=A$. We want to describe this map with a domain in $[0,\pi ]^{2}$. So, first $\exp (-tB)e^{sA}\cdot b_{0}=\exp ((\pi -t)B)e^{\pi B}e^{sA}\cdot b_{0}$. Finally, since $Ad(e^{\pi B}A)=-A$ we get $$\exp (-sB)e^{sA}\cdot b_{0}=e^{(\pi -s)B}e^{(\pi -s)A}\cdot b_{0}.$$Hence the degree of $f_{1}^{\pi }$ is the degree of $(t,s)\mapsto (\pi -t,\pi -s)$. This degree is $+1$ since it preserves the orientation. Summarizing, the boundary operator is given by - $\partial_3 \mathcal{S}_{(13)}=0$; - $\partial_2 \mathcal{S}_{(123)} = -2 \mathcal{S}_{(23)} $ and $\partial_2 \mathcal{S}_{(132)}= -2\mathcal{S}_{(12)}$; - $\partial_1 \mathcal{S}_{(12)} = \partial \mathcal{S}_{(23)}=0$. Hence the integer homology groups are - $H_3(\mathbb{F},\mathbb{Z}) = \mathbb{Z}$ generated by $\mathcal{S}_{(13)}$. - $H_2(\mathbb{F},\mathbb{Z}) = 0$ ($\ker \partial_2 =0$). - $H_1(\mathbb{F},\mathbb{Z}) = \mathbb{Z}_2 \oplus \mathbb{Z}_2$ ($\ker \partial_1$ is $\mathbb{Z}\oplus\mathbb{Z}$ and the image of $\partial_2$ is $2\mathbb{Z}\cdot \mathcal{S}_{(12)} \oplus 2\mathbb{Z}\cdot \mathcal{S}_{(23)}$). Algebraic expressions for the degrees ------------------------------------- Here we compute the coefficients $c(w,w^{\prime })$ in terms of the roots by finding the degrees of the maps involved. For a diffeomorphism $\varphi $ of the sphere its degree is local degree at a point $x$ which in turn is the sign of the determinant $\det (d\varphi _{x})$ with respect to a volume form of $S^{d}$. Let us apply this in our context. We let $w=r_{1}\cdots r_{n}$ and $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ be reduced expressions, with $r_{i}=r_{\alpha _{i}}$ and assume throughout that the simple root $\alpha _{i}$ has multiplicity $d_{i}=d_{\alpha _{i}}=1$. We must find the degrees of $f_{i}^{0}$ and $f_{i}^{\pi }$ defined by 1. $f_i^0 (t_1, \ldots, 0, \ldots, t_n) = \Phi_{w^{\prime }}^{-1}\left( \psi_1(t_1) \cdots 1 \cdots \psi_n(t_n) \cdot b_0 \right).$ 2. $f_i^\pi (t_1, \ldots, \pi , \ldots, t_n) = \Phi_{w^{\prime }}^{-1}\left( \psi_1(t_1) \cdots m_{\alpha_i} \cdots \psi_n(t_n) \cdot b_0 \right).$ In these expressions $\Phi _{w^{\prime }}^{-1}$ is defined by a previously chosen reduced expression $w^{\prime }=s_{1}\cdots s_{m}$ of $w^{\prime }$ which may be distinct of $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$. On the other hand $w^{\prime }=r_{1}\cdots \hat{r_{i}}\cdots r_{n}$ can be used to define another characteristic map, which will be denoted by $\Psi _{w^{\prime }}$. This new characteristic map define new functions 1. $p_i^0 (t_1, \ldots, 0, \ldots, t_n) = \Psi_{w^{\prime }}^{-1}\left( \psi_1(t_1) \cdots 1 \cdots \psi_n(t_n) \cdot b_0 \right)$ and 2. $p_{i}^{\pi }(t_{1},\ldots ,\pi ,\ldots ,t_{n})=\Psi _{w^{\prime }}^{-1}\left( \psi _{1}(t_{1})\cdots m_{\alpha _{i}}\cdots \psi _{n}(t_{n})\cdot b_{0}\right) $. The two pair of functions are related by $$f_{i}^{\epsilon }=\left( \Phi _{w^{\prime }}^{-1}\circ \Psi _{w^{\prime }}\right) \circ p_{i}^{\epsilon }\hspace{1cm}\epsilon =0,\pi .$$ The composition $\Phi _{w^{\prime }}^{-1}\circ \Psi _{w^{\prime }}$ (also understood as a map between spheres in which the boundary are collapsed to points) are homeomorphisms of spheres and, hence, have degree $\pm 1$. Hence we can concentrate on the computation of degrees of the ${p_{i}^{\epsilon }} $’s since the total degree will be multiplied by $\pm 1$. Before getting these degrees we make the following discussion on the orientation of the faces of the cube $\left[ -1,1\right] ^{d}$, centered at the origin of $\mathbb{R}^{d}$, which is given with the basis $\{e_{1},\ldots ,e_{d}\}$. Starting with the $(d-1)$-dimensional sphere $S^{d-1}$ we orient the tangent space at $x\in S^{d-1}$ by a basis $\{f_{2},\ldots ,f_{d}\}$ such that $\{x,f_{2},\ldots ,f_{d}\}$ is positively oriented. The faces of $[-1,1]^{d}$ are oriented accordingly: Given a base vector $e_{j}$, we let $F_{j}^{-}$ be the face perpendicular to $e_{j}$ that contains $-e_{j}$ and $F_{j}^{+}$ the one that contains $e_{j}$. Then $F_{j}^{-}$ has the same orientation as the basis $e_{1},\ldots ,\hat{e_{j}},\ldots ,e_{d}$ if $j$ is even ($-e_{j},e_{1},\ldots ,\hat{e_{j}},\ldots ,e_{d}$ is positively oriented in $\mathbb{R}^{d}$) and opposite orientation if $j$ is odd. Therefore the orientation of $F_{j}^{-}$ is $\left( -1\right) ^{j}$ the orientation of $e_{1},\ldots ,\hat{e_{j}},\ldots ,e_{d}$. Analogously, the orientation of $F_{j}^{+}$ is $\left( -1\right) ^{j+1}$ the orientation of $e_{1},\ldots ,\hat{e_{j}},\ldots ,e_{d}$. The following facts about the action of an element $m\in M$ will be used below in the computation of the degrees. \[lemproperties\]For a root $\alpha $ consider the action on $\mathbb{F}$ of $m=m_{\alpha }=\exp (\pi A_{\alpha })\in M$. Then 1. $mwb_{0}=wb_{0}$ and $mNm^{-1}=N$. Therefore $m$ leaves invariant any Bruhat cell and hence any Schubert cell $\mathcal{S}_{w}$. 2. The restriction of $m$ to $N\cdot wb_{0}$ is a diffeomorphism. 3. The differential $dm_{wb_{0}}$ identifies to $\mathrm{Ad}\left( m\right) $ restricted to the subspace$$\sum_{\beta \in \Pi_w}\mathfrak{g}_{\beta }.$$ Since $\mathrm{Ad}(m_{\alpha })\mathfrak{g}_{\beta }=\mathfrak{g}_{\beta }$, $\beta \in \Pi $ (cf. Lemma \[compact\_parametrization\_2\]), the first and second statements follow easily. For the third statement we use the notation $X\cdot x=d/dt\left( \exp tX\right) _{t=0}$, $x\in \mathbb{F}$ and $X\in \mathfrak{g}$. Also, for $A\subset \mathfrak{g}$ write $A\cdot x=\{X\cdot x:X\in A\}$. Note that $N\cdot wb_{0}=w(w^{-1}Nw)\cdot b_{0}$, and the tangent space to $(w^{-1}Nw)\cdot b_{0}$ at $b_{0}$ is spanned by $\mathfrak{g}_{\alpha }\cdot b_{0}$ with $\alpha <0$ such that $\alpha =w^{-1}\beta $ and $\beta >0 $, that is, $w\cdot \alpha >0$. Since $\left( dw\right) (\mathfrak{g}_{\alpha }\cdot b_{0})=\mathfrak{g}_{w\cdot \alpha }\cdot b_{0}$, it follows that $T_{wb_{0}}\left( N\cdot wb_{0}\right) $ is spanned by $\mathfrak{g}_{\beta }\cdot b_{0}$ with $\beta =w\cdot \alpha >0$ such that $w^{-1}\cdot \beta =\alpha <0$. Hence the result. The next statement computes the degree of ${p_{i}^{\epsilon }}$’s in terms of Killing numbers. \[propdegresig\]$\deg (p_{i}^{0})=(-1)^{I}$ and $\deg (p_{i}^{\pi })=(-1)^{I+1+\sigma }$, where $$\sigma =\sigma \left( w,w^{\prime }\right) =\sum_{\beta \in \Pi _{u}}\frac{2\langle \alpha _{i},\beta \rangle }{\langle \alpha _{i},\alpha _{i}\rangle }\dim \mathfrak{g}_{\beta },\hspace{0.5cm}\Pi _{u}=\Pi ^{+}\cap u\Pi ^{-},\hspace{0.5cm}u=r_{i+1}\cdots r_{n}, \label{forsigmasoma}$$and $I$ is the sum of the multiplicities of the roots $\alpha _{j}$ with $j\leq i$. The map $p_{i}^{0}$ is the projection of the face of a $d$-dimensional cube onto the face of a $(d-1)$-dimensional cube, i.e., in coordinates $$(t_{1},\ldots ,0,\ldots ,t_{n}) \mapsto (t_{1},\ldots ,\hat{t_{i}},\ldots ,t_{n}).$$ Note that with respect to the basis $e_{1},\ldots ,e_{d}$ the $t_{i}$-coordinate appears in the $I$-th position. Hence, by the orientation of the cube, discussed above, the projection preserves or reverses orientation if $I$ is even or odd, respectively. Therefore, $\deg (p_{i}^{0})=(-1)^{I}$. To get $\deg (p_{i}^{\pi })$ write $m_{i}=m_{\alpha _{i}}$ for the element of $M$ appearing in the expression of $p_{i}^{\pi }$. Its action on $\mathbb{F}$ leaves invariant any Bruhat cell $N\cdot wb_{0}$ (because $m_{i}Nm_{i}^{-1}=N$ and $m_{i}wb_{0}=wb_{0}$), and hence any Schubert cell. Moreover, the restriction of $m_{i}$ to $N\cdot wb_{0}$ is a diffeomorphism (given by the conjugation $y\in N\mapsto m_{i}ym_{i}^{-1}$). In particular, we restrict the action of $m_{i}$ to the cell $\mathcal{S}_{u} $, $u=r_{i+1}\cdots r_{n}$. Using the parametrization of this cell by the cube $B_{u}$ we get $$m_{i}\psi _{i+1}(t_{i+1})\cdots \psi _{n}(t_{n})\cdot b_{0}=\psi _{i+1}(s_{i+1})\cdots \psi _{n}(s_{n})\cdot b_{0},$$with $(s_{i+1},\ldots ,s_{n})=\overline{m}_{i}(t_{i+1},\ldots ,t_{n})$ with $\overline{m}_{i}:B_{u}\rightarrow B_{u}$ continuous and a diffeomorphism of the interior of $B_{u}$. Hence, $p_{i}^{\pi }(t_{1},\ldots ,\pi ,\ldots ,t_{n})$ becomes the projection of the $i-1$ first coordinates and the composition of $\overline{m}_{i}$ with the projection of the last $j$-coordinates, $j=i+1,\ldots ,n$. From the choice of the orientation of $B_{w}=\left[ 0,\pi \right] ^{d}$, the face $(t_{1},\ldots ,\pi ,\ldots ,t_{n})$ of $B_{w}$ has orientation $(-1)^{I+1}$ with respect to the orientation of the coordinates $(t_{1},\ldots ,\widehat{t_{i}},\ldots ,t_{n})$. Hence, after collapsing the boundary to a point, we get the degree $$\deg p_{i}^{\pi }=(-1)^{I+1}\deg \overline{m}_{i}.$$The degree of $\overline{m}_{i}$ equals its local degree at one point which in turn is sign of the determinant of the differential $d({m}_{i})_{ub_{0}}$ restriced to the tangent space to Bruhat cell $N\cdot ub_{0}$ at $ub_{0}$: $$\deg (p_{i}^{\pi })=(-1)^{I+1}\mathrm{sgn}[\det \left( d({m}_{i})_{ub_{0}}|T_{ub_{0}}(N\cdot ub_{0})\right) ].$$By the third statement in the Lemma \[lemproperties\], $T_{ub_{0}}(N\cdot ub_{0})$ identifies to $\sum_{\beta \in \Pi_w}\mathfrak{g}_{\beta }$. Once we have the generators $\mathfrak{g}_{\beta }\cdot ub_{0}$, $\beta \in \Pi_w$ for $T_{ub_{0}}(N\cdot ub_{0})$ together with the action of $\mathrm{Ad}(m_{i})$ over $\mathfrak{g}_{\beta }$ given by the Lemma [compact\_parametrization\_2]{}, $\mathrm{Ad}(m_{\alpha })_{\left\vert \mathfrak{g}_{\beta }\right. }=(-1)^{\epsilon (\alpha ,\beta )}\mathrm{id}$ we conclude that the signal of $\det \left( d({m}_{i})_{ub_{0}}|T_{ub_{0}}(N\cdot ub_{0})\right) =(-1)^{\sigma }$ where $$\sigma =\sum_{\beta \in \Pi _{u}}\frac{2\langle \alpha _{i},\beta \rangle }{\langle \alpha _{i},\alpha _{i}\rangle }\dim \mathfrak{g}_{\beta }. \qedhere$$ Summarizing, we have the following algebraic expression for the coefficient $c\left( w,w^{\prime }\right) $. \[teoforcw\]Let be $\sigma \left( w,w^{\prime }\right) $ be defined as in (\[forsigmasoma\]). Then $$c(w,w^{\prime })=\deg \left( \Phi _{w^{\prime }}^{-1}\circ \Psi _{w^{\prime }}\right) (-1)^{I}(1-(-1)^{\sigma \left( w,w^{\prime }\right) }).$$ We will now derive another formula for $\sigma \left( w,w^{\prime }\right) $ that does not depend on the reduced expressions of $w$ and $w^{\prime }$. This formula is the same one given by Theorem A of [@Koc95]. For $w\in \mathcal{W}$, let $$\phi (w)=\sum_{\beta \in \Pi _{w}}\dim \mathfrak{g}_{\beta }\cdot \beta$$be the sum of roots in $\Pi _{w}=\Pi ^{+}\cap w\Pi ^{-}$ counted with their multiplicity. As before let $w=r_{1}\cdots r_{n}$ and $w^{\prime }=r_{1}\cdots \widehat{r_{i}}\cdots r_{n}$ be reduced expressions. Let $\beta $ be the unique root (not necessarily simple) such that $w=r_{\beta }w^{\prime }$, that is, $\beta =r_{1}\cdots r_{i-1}\alpha _{i}$. Then $$\phi (w)-\phi (w^{\prime })=(1-\sigma )\beta$$where $\sigma =\sigma (w,w^{\prime })$ is the sum (\[forsigmasoma\]). By the reduced expressions $w^{-1}=r_{n}\cdots r_{1}$ and $w^{\prime -1}=r_{n}\cdots \hat{r_{i}}\cdots r_{1}$ and $u^{-1}=r_{n}\cdots r_{i+1}$ we obtain the sets 1. $\Pi _{w}=\{\alpha _{1},r_{1}\alpha _{2},\ldots ,r_{1}\cdots r_{i-1}\alpha _{i},r_{1}\cdots r_{i}\alpha _{i+1},\ldots ,r_{1}\cdots r_{n-1}\alpha _{n}\}$. 2. $\Pi _{w^{\prime }}=\{\alpha _{1},r_{1}\alpha _{2},\ldots , r_{1}\cdots r_{i-1}\alpha _{i+1},\ldots ,r_{1}\cdots \hat{r_{i}}\cdots r_{n-1}\alpha _{n}\}$. 3. $\Pi _{u}=\{\alpha _{i+1},r_{i+1}\alpha _{i+2},\ldots ,r_{i+1}\cdots r_{n-1}\alpha _{n}\}$. The first $(i-1)$ roots of $\Pi _{w}$ and $\Pi _{w^{\prime }}$ coincide. The remaining ones are related by the equalities $$(r_{1}\cdots r_{i-1})r_{i}\cdots r_{j}\alpha _{j+1}=r_{\beta }(r_{1}\cdots r_{i-1})r_{i+1}\cdots r_{j}\alpha _{j+1} \quad,\quad j=i,\ldots ,n-1,$$because $(r_{1}\cdots r_{i-1})r_{i}(r_{1}\cdots r_{i-1})^{-1}=r_{r_{1}\cdots r_{i-1}\alpha _{i}}=r_{\beta }$. It follows that the remaining roots $r_{1}\cdots r_{j}\alpha _{j+1}$ and the roots $r_{1}\cdots \hat{r_{i}}\cdots r_{j}\alpha _{j+1}$ have the same multiplicity $d_{j}$, $j=i,\ldots ,n-1$. Write $\gamma _{j}=r_{i+1}\cdots r_{j}\alpha _{j+1}$, so that $\Pi _{u}=\{\gamma _{i},\gamma _{i+1},\ldots ,\gamma _{n-1}\}$. Then$$\phi (w)-\phi (w^{\prime })=\beta +\sum_{j=i}^{n-1}d_{j}(r_{1}\cdots r_{i-1})\left( r_{i}(\gamma _{j})-\gamma _{j}\right) \label{root_equation}$$because $\beta =r_{1}\cdots r_{i-1}\alpha _{i}$ has multiplicity $1$ as $\alpha _{i}$. Since $r_{i}(\gamma _{j})-\gamma _{j}=-\displaystyle\frac{2\langle \alpha _{i},\gamma _{j}\rangle }{\langle \alpha _{i},\alpha _{i}\rangle }\alpha _{i} $ we rewrite (\[root\_equation\]) as $$\begin{aligned} \phi (w)-\phi (w^{\prime }) &=&\left( 1-\sum_{j=i}^{n-1}d_{j}\frac{2\langle \alpha _{i},\gamma _{j}\rangle }{\langle \alpha _{i},\alpha _{i}\rangle }\right) \beta \label{root_equation_2} \\ &=&(1-\sigma )\beta \notag\end{aligned}$$ concluding the proof. Combining the above proposition with Theorem \[teoforcw\] we get immediately the following formula for $c\left( w,w^{\prime }\right) $ (cf. [@Koc95], Theorem A). \[teoforcw1\]$$c(w,w^{\prime })=\deg \left( \Phi _{w^{\prime }}^{-1}\circ \Psi _{w^{\prime }}\right) (-1)^{I}(1+(-1)^{\kappa(w,w')}) \label{forcewseg}$$where $\kappa(w,w')$ is the integer defined by $\phi (w)-\phi (w^{\prime })=\kappa(w,w')\cdot \beta $ and $\beta $ is the unique root such that $w=r_{\beta }w^{\prime }$. **Remark:** If $w=r_{1}\cdots r_{n}$ and $w^{\prime }=r_{1}\cdots r_{i-1}$ then $c(w,w^{\prime })=0$ because $m_{\alpha _{n}}$ does not affect the computations of the degrees (see Proposition \[propdegresig\]). **Example of $\mathrm{Sl}(3,\mathbb{R})$:** Let us use Formula (\[forcewseg\]) to rederive the homology of the maximal flag of $\mathrm{Sl}(3,\mathbb{R})$, the split real form of the algebra whose Dynkin diagram is $A_2$. Let fix the same reduced expressions for elements in $\mathcal{W}$ and note that the unique element which has more than one reduced expression is $(13)$ which implies that the factor $\left( \Phi _{w^{\prime }}^{-1}\circ \Psi _{w^{\prime }}\right) $ is $1$ in all cases. In this case, we have the following table \[table1\] which determines completely the coefficients $c(w,w^{\prime })$, as in Subsection \[subsecsl3\]. \[table1\] [0.4 ]{}[||@c|c|c||]{}\ ${\mathcal{W}}$ & ${\Pi _{w}}$ & ${\phi (w)}$\  $1$ & $\emptyset $ & $0$\  $(12)$ & $\alpha _{1}$ & $\alpha _{1}$\  $(23)$ & $\alpha _{2}$ & $\alpha _{2}$\  $(123)$ & $\alpha _{1},\alpha _{1}+\alpha _{2}$ & $2\alpha _{1}+\alpha _{2}$\  $(132)$ & $\alpha _{2},\alpha _{1}+\alpha _{2}$ & $\alpha _{1}+2\alpha _{2}$\  $(13)$ & $\Pi ^{+}$ & $2\alpha _{1}+2\alpha _{2}$\ For instance, let us compute $\partial_3 \mathcal{S}_{(13)}=0$. According to the table \[table1\], $\sigma((13),(123))=2(\alpha_1+\alpha_2)-(2\alpha_1+\alpha_2)= \alpha_2$ and $\sigma((13),(132))= 2(\alpha_1+\alpha_2)-(\alpha_1+2\alpha_2)=\alpha_1$. It implies that $\kappa((13),(123))=\kappa((13),(132))=1$ by which we conclude that $c((13),(123))=c((13),(132))=0$. **Example of $G_2$:** Let us apply the results above for the two groups with Dynkin diagram $G_{2}$, namely the complex group and the split real form. In the complex case, we already have $\partial =0$. Now we proceed to the real case. Let $\Sigma =\{\alpha _{1},\alpha _{2}\}$ be the simple roots. The set $\Pi ^{+}\setminus \Sigma =\{\alpha _{3}=\alpha _{2}+\alpha _{1},\alpha _{4}=\alpha _{1}+2\alpha _{2},\alpha _{5}=\alpha _{1}+3\alpha _{2},\alpha _{6}=2\alpha _{1}+3\alpha _{2}\}$ contains the remaining positive roots. The Weyl group with the respective fixed reduced expressions is $\mathcal{W}=\{1,r_{1},r_{2},s_{1}=r_{1}r_{2},s_{2}=r_{2}r_{1},r_{1}s_{2},r_{2}s_{1},s_{1}^{2},s_{2}^{2},r_{1}s_{2}^{2},r_{2}s_{1}^{2},s_{1}^{3}\}$, where $r_{i}=r_{\alpha _{i}}$ are the simple reflections and $s_{1}^{3}=s_{2}^{3}$ is the unique element with two different minimal decompositions. The next table \[table2\] presents the data useful to compute the homology coefficients. \[table2\] [0.45 ]{}[||@c|c|c||]{}\ ${\mathcal{W}}$ & ${{\Pi _{w}}}$ & ${\phi (w)}$\  $1$ & $\emptyset $ & $0$\  $r_{1}$ & $\alpha _{1}$ & $\alpha _{1}$\  $r_{2}$ & $\alpha _{2}$ & $\alpha _{2}$\  $s_{1}$ & $\alpha _{1},\alpha _{3}$ & $2\alpha _{1}+\alpha _{2}$\  $s_{2}$ & $\alpha _{2},\alpha _{5}$ & $\alpha _{1}+4\alpha _{2}$\  $r_{1}s_{2}$ & $\alpha _{1},\alpha _{3},\alpha _{6}$ & $4\alpha _{1}+4\alpha_{2}$\  $r_{2}s_{1}$ & $\alpha _{2},\alpha _{5},\alpha _{4}$ & $2\alpha _{1}+6\alpha_{2}$\  $s_{1}^{2}$ & $\alpha _{1},\alpha _{3},\alpha _{6},\alpha _{4}$ & $5\alpha_{1}+6\alpha _{2}$\  $s_{2}^{2}$ & $\alpha _{2},\alpha _{5},\alpha _{4},\alpha _{6}$ & $4\alpha_{1}+9\alpha _{2}$\  $r_{1}s_{2}^{2}$ & $\alpha _{1},\alpha _{3},\alpha _{6},\alpha _{4},\alpha_{5}$ & $6\alpha _{1}+9\alpha _{2}$\  $r_{2}s_{1}^{2}$ & $\alpha _{2},\alpha _{5},\alpha _{4},\alpha _{6},\alpha_{3}$ & $5\alpha _{1}+10\alpha _{2}$\  $s_{1}^{3}$ & $\Pi ^{+}$ & $6\alpha _{1}+10\alpha _{2}$\ By (\[forcewseg\]) the boundary operator is given as - $\partial_6 \mathcal{S}_{s_{1}^{3}}=0$; - $\partial_5 \mathcal{S}_{r_{1}s_{2}^{2}}=-2\mathcal{S}_{s_{2}^{2}}$ and $\partial_5 \mathcal{S}_{r_{2}s_{1}^{2}}=-2\mathcal{S}_{s_{1}^{2}}$; - $\partial_4 \mathcal{S}_{s_{1}^{2}}=\partial_4 \mathcal{S}_{s_{2}^{2}}=0$; - $\partial_3\mathcal{S}_{r_{1}s_{2}}=\partial_3 \mathcal{S}_{r_{2}s_{1}}=0$; - $\partial_2 \mathcal{S}_{s_{1}}=-2\mathcal{S}_{r_{2}}$ and $\partial_2 \mathcal{S}_{s_{2}}=-2\mathcal{S}_{r_{1}}$; - $\partial_1 \mathcal{S}_{r_{1}}=\partial_1 \mathcal{S}_{r_{2}}=0$. Hence - $H_6(\mathbb{F},\mathbb{Z})= \mathbb{Z}$. - $H_5(\mathbb{F},\mathbb{Z})= 0$. - $H_4(\mathbb{F},\mathbb{Z}) = \mathbb{Z}_2 \oplus \mathbb{Z}_2$. - $H_3(\mathbb{F},\mathbb{Z}) = \mathbb{Z} \oplus \mathbb{Z}$. - $H_2(\mathbb{F},\mathbb{Z}) = 0$. - $H_1(\mathbb{F},\mathbb{Z}) = \mathbb{Z}_2 \oplus \mathbb{Z}_2$. Partial flag manifolds\[secpartial\] ==================================== In this section we project down the constructions made for the maximal flag manifolds, via the canonical map $\pi _{\Theta }:\mathbb{F}\rightarrow \mathbb{F}_{\Theta }$, to obtain analogous results for the homology of a partial flag manifold. In $\mathbb{F}_{\Theta }$ the Schubert cells are $\mathcal{S}_{w}^{\Theta }$, $w\in \mathcal{W}/\mathcal{W}_{\Theta }$, with $\mathcal{S}_{w}^{\Theta }=\mathcal{S}_{w_{1}}^{\Theta }$ if $w\mathcal{W}_{\Theta }={w_{1}}\mathcal{W}_{\Theta }$. The next lemma chooses a special representative in $w\mathcal{W}_{\Theta }$ for $\mathcal{S}_{w}^{\Theta}$. \[minimal\_element\] There exists an element $w_{1}=wu$ of the coset $w\mathcal{W}_{\Theta }$ such that $$\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{w_{1}}.$$This element is unique and minimal with respect to the Bruhat-Chevalley order. By Proposition 1.1.2.13 of [@War72] any $v\in \mathcal{W}$ can be written uniquely as $$v=v_{s}v_{u}$$with $v_{s}\in \mathcal{W}_{\Theta }$ and $v_{u}$ satisfying $\Pi ^{+}\cap v_{u}\Pi ^{-}\cap \langle \Theta \rangle =\emptyset $, that is, no positive root in $\langle \Theta \rangle $ is mapped to a negative root by $v_{u}^{-1} $. Note that the condition for $v_{u}$ is equivalent to $\Pi ^{-}\cap v_{u}\Pi ^{+}\cap \langle \Theta \rangle =\emptyset $, since a root $\alpha >0$ belongs to $\Pi ^{+}\cap v_{u}\Pi ^{-}\cap \langle \Theta \rangle $ if and only if $-\alpha \in \Pi ^{-}\cap v_{u}\Pi ^{+}\cap \langle \Theta \rangle $. Let $w^{-1}=v_{s}v_{u}$ be the decomposition for $w^{-1}$ so that $w=v_{u}^{-1}v_{s}^{-1}$. Then $w_{1}=v_{u}^{-1}\in w\mathcal{W}_{\Theta }$ is the required element. In fact $\Pi ^{-}\cap w_{1}^{-1}\Pi ^{+}\cap \langle \Theta \rangle =\emptyset $, and hence $\Pi ^{+}\cap w_{1}\Pi ^{-}\cap w_{1}\langle \Theta \rangle =\emptyset $. Now the tangent space $T_{w_{1}b_{0}}(N\cdot w_{1}b_{0})$ is $$\langle \mathfrak{g}_{\beta }\cdot w_{1}b_{0}\,:\beta \in \Pi ^{+}\cap w_{1}\Pi ^{-}\rangle$$(cf. Lemma \[lemproperties\]). On the other hand the tangent space to the fiber $\pi _{\Theta }^{-1}(w_{1}b_{\Theta })$ is the translation under $w_{1} $ of the tangent space to fiber at origin. Hence, $T_{w_{1}b_{\Theta }}\pi _{\Theta }^{-1}(w_{1}b_{\Theta })$ is spanned by $w_{1}(\mathfrak{g}_{\alpha }\cdot b_{0})$, with $\alpha \in \langle \Theta \rangle $ and $\alpha <0$. By the translation formula, we have $w_{1}(\mathfrak{g}_{\alpha }\cdot b_{0})=\mathfrak{g}_{w_{1}\alpha }\cdot w_{1}b_{0}$. Therefore, by writting $\gamma =w_{1}\alpha $ we conclude that $T_{w_{1}b_{\Theta }}\pi _{\Theta }^{-1}(w_{1}b_{\Theta })$ is spanned by $\mathfrak{g}_{\gamma }\cdot w_{1}b_{0}$ with $w_{1}^{-1}\gamma \in \langle \Theta \rangle $ and $w^{-1}\gamma <0$, that is, with $w_{1}^{-1}\gamma \in \Pi ^{-}\cap \langle \Theta \rangle $. So that $$T_{w_{1}b_{0}}(\pi _{\Theta }^{-1}\left( w_{1}b_{\Theta }\right) )=\langle \mathfrak{g}_{\gamma }\cdot w_{1}b_{0}\,:\gamma \in w_{1}\Pi ^{-}\cap w_{1}\langle \Theta \rangle \rangle .$$Since $\Pi ^{+}\cap w_{1}\Pi ^{-}\cap w_{1}\langle \Theta \rangle =\emptyset $, it follows that none of roots $\gamma $ spanning $T_{w_{1}b_{0}}(\pi _{\Theta }^{-1}\left( w_{1}b_{\Theta }\right) )$ can be positive. Therefore, $T_{w_{1}b_{0}}(N\cdot w_{1}b_{0})\cap T_{w_{1}b_{0}}(\pi _{\Theta }^{-1}\left( w_{1}b_{\Theta }\right) )=\{0\}$. This implies that the differential of $\pi _{\Theta }$ maps $T_{w_{1}b_{0}}(N\cdot w_{1}b_{0})$ isomorphically to the tangent space of $\pi _{\Theta }\left( N\cdot w_{1}b_{0}\right) =N\cdot w_{1}b_{\Theta }$. Hence the two Bruhat cells have the same dimension. Finally, $N\cdot w_{1}b_{\Theta }$ has the minimum possible dimension among the cells $N\cdot wb_{0}$, $w\in w_{1}\mathcal{W}$, because all of them project onto $N\cdot w_{1}b_{\Theta }$. Hence $w_{1}$ has minimal length in $w_{1}\mathcal{W}$ which is known to be unique and minimal with respect to the Bruhat-Chevalley order as well (see Deodhar [@Deo77]). We will denote by $\mathcal{W}_{\Theta }^{\min }$ the set of minimal elements of the cosets in $\mathcal{W}/\mathcal{W}_{\Theta }$. Now we contruct a cellular decomposition for $\mathbb{F}_{\Theta }$ with the aid of the minimal elements $w\in \mathcal{W}_{\Theta }^{\min }$ in their cosets $w\mathcal{W}_{\Theta }$, satisfying $\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{w}$. Using a reduced decomposition of such minimal element $w$ we have new functions $\Phi _{w}^{\Theta }$ defined in the same way, but replacing the origin $b_{0}$ of $\mathbb{F}$ by the origin $b_{\Theta }$ of $\mathbb{F}_{\Theta }$, that is, $$\Phi _{w}^{\Theta }(t_{1},\ldots ,t_{n})=\psi _{1}(t_{1})\cdots \psi _{n}(t_{n})\cdot b_{\Theta }.$$By equivariance, $\Phi _{w}^{\Theta }=\pi _{\Theta }\circ \Phi _{w}$. This function satisfies the required properties to be a characteristic map for the Schubert cells $\mathcal{S}_{w}^{\Theta }$. Take $w\in \mathcal{W}_{\Theta }^{\min }$ so that $\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{w}$ and let $w=r_{1}\cdots r_{n}$ be a reduced expression as a product of simple reflections. Let $\Phi _{w}^{\Theta }:B_{w}\rightarrow \mathbb{F}_{\Theta }$ be defined by $\Phi _{w}^{\Theta }=\pi _{\Theta }\circ \Phi _{w}$ and take $\mathbf{t}=(t_{1},\ldots ,t_{n})\in B_{w}$. Then $\Phi _{w}^{\Theta }$ is a characteristic map for $\mathcal{S}_{w}^{\Theta }$, that is, satisfies the following properties: 1. $\Phi _{w}^{\Theta }(B_{w})=\mathcal{S}_{w}^{\Theta }$. 2. $\Phi _{w}^{\Theta }(\mathbf{t})\in \mathcal{S}_{w}^{\Theta }\setminus N\cdot wb_{\Theta }$ if and only if $\mathbf{t}\in \partial B_{w}=S^{d-1}$. 3. $\Phi |_{w}^{\Theta }{B_{w}^{\circ }}:B_{w}^{\circ }\rightarrow N\cdot wb_{\Theta }$ is a homeomorphism, where $B_{w}^{\circ }$ is the interior of $B_{w}$. This is the Proposition \[characteristic\_map\] in this generalized flag context. The first item follows by equivariance of $\pi _{\Theta }$. The second assertion is true because $\pi _{\Theta }(\mathcal{S}_{w}\setminus N\cdot wb_{0})=\mathcal{S}_{w}^{\Theta }\setminus N\cdot wb_{\Theta }$ and $\Phi _{w}^{\Theta }=\pi _{\Theta }\circ \Phi _{w}$. The last item is a consequence of the equality of the dimensionsx of Bruhat cells $N\cdot wb_{0} $ and $N\cdot wb_{\Theta }$. Now we can find out the boundary maps $\partial ^{\Theta }$ with coefficients $c^{\Theta }([w],[w^{\prime }])$, where $[w]$ denotes the the coset $w\mathcal{W}_{\Theta }$. We have $c^{\Theta }([w],[w^{\prime }])=0$, unless 1. $\dim \mathcal{S}_w^\Theta = \dim \mathcal{S}_{w^{\prime }}^\Theta + 1$ and 2. $\mathcal{S}_{w^{\prime }}^{\Theta }\subset \mathcal{S}_{w}^{\Theta }$. Here the inclusions among the Schubert cells are also given by the Bruhat-Chevalley order (cf. Proposition \[BC-order\]), namely $\mathcal{S}_{w^{\prime }}^{\Theta }\subset \mathcal{S}_{w}^{\Theta }$ if and only if there is $u\in w^{\prime }\mathcal{W}_{\Theta }$ such that $u<w$. (This follows immediately from the projections $\pi _{\Theta }\mathcal{S}_{w}=\mathcal{S}_{w}^{\Theta }$.) Actually, we have the following complement to Lemma \[minimal\_element\]. \[lemumenos\] Let $w\in \mathcal{W}_{\Theta }^{\min }$ minimal in its coset and suppose that there exists $u\in w^{\prime }\mathcal{W}_{\Theta }$ with $u<w$ and $\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{w^{\prime }}^{\Theta }+1$. Then $u$ is minimal in $w^{\prime }\mathcal{W}_{\Theta }$. We have $\dim \mathcal{S}_{w}=\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{u}^{\Theta }+1\leq \dim \mathcal{S}_{u}+1$. But if $u<w$ then $\dim \mathcal{S}_{u}\leq \dim \mathcal{S}_{w}-1$, so that $\dim \mathcal{S}_{w}\leq \dim \mathcal{S}_{u}+1\leq \dim \mathcal{S}_{w}$, implying that $$\dim \mathcal{S}_{u}=\dim \mathcal{S}_{w}^{\Theta }-1=\dim \mathcal{S}_{u}^{\Theta }.$$Hence $u$ is minimal in its coset. **Remark:** The assumption $\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{w^{\prime }}^{\Theta }+1$ in Lemma \[lemumenos\] is essential. Without it there may be $u\in w^{\prime }\mathcal{W}_{\Theta }$ which is not minimal although $u<w$ and $w$ is minimal. Geometrically this happens when $\dim \mathcal{S}_{w}=\dim \mathcal{S}_{u}^{\Theta }+1$ but $\dim \mathcal{S}_{w}^{\Theta }>\dim \mathcal{S}_{w^{\prime }}^{\Theta }+1$, which may give $c\left( w,u\right) \neq 0$ and $c^{\Theta }\left( [w],[u]\right) =0$. The following example illustrates this situation. **Example:** In the Weyl group $S_{4}$ of $A_{3}$ take $w=(12)(23)(34)$ and $\Theta =\{\alpha _{23}\}$. Then $w$ is minimal in the coset $w\mathcal{W}_{\{\alpha _{23}\}}$. The roots $\alpha _{12}$, $(12)\alpha _{23}=\alpha _{12}+\alpha _{23}$ and $(12)(23)\alpha _{34}=\alpha _{12}+\alpha _{23}+\alpha _{34}$ are positive roots mapped to negative roots by $w^{-1}$ and none of these roots lie in $\langle \Theta \rangle $. However, $w^{\prime }=(12)(23)=(123)$ is not minimal in its coset since $(12)<(12)(23)$ and and both belong to the same coset. Now $\dim \mathcal{S}_{(12)(23)}=\dim \mathcal{S}_{w}-1$ but $\mathcal{S}_{(12)(23)}^{\Theta }=\dim \mathcal{S}_{(12)}^{\Theta }=\dim \mathcal{S}_{w}^{\Theta }-2$. Now if $w$ and $w^{\prime }$ belong to $\mathcal{W}_{\Theta }^{\min }$ and $\dim \mathcal{S}_{w}^{\Theta }=\dim \mathcal{S}_{w^{\prime }}^{\Theta }+1$ then there are the homeomorphisms $\pi _{\Theta }:N\cdot wb_{0}\rightarrow N\cdot wb_{\Theta }$ and $\pi _{\Theta }:N\cdot w^{\prime }b_{0}\rightarrow N\cdot w^{\prime }b_{\Theta }$. This implies that the attaching map between $\mathcal{S}_{w}^{\Theta }$ and $\mathcal{S}_{w^{\prime }}^{\Theta }$ defined by $\Phi _{w}^{\Theta }=\pi _{\Theta }\circ \Phi _{w}$ and $\Phi _{w^{\prime }}^{\Theta }\circ \Phi _{w^{\prime }}$ is the same as the attaching map between $\mathcal{S}_{w}$ and $\mathcal{S}_{w^{\prime }}$. Hence the coefficients for $\partial ^{\Theta }$ and $\partial $ are the same, that is, $$c^{\Theta }([w],[w^{\prime }])=c(w,w^{\prime }).$$ Hence the computation of $c^{\Theta }([w],[w^{\prime }])$ reduces to a computation on $\mathbb{F}$. The cellular homology of $\mathbb{F}_{\Theta }$ is isomorphic to the homology of $\partial _{\min }^{\Theta }$ which is the boundary map of the free module $\mathcal{A}_{\Theta }^{\min }$ generated by $\mathcal{S}_{w}$, $w\in \mathcal{W}_{\Theta }^{\min }$, obtained by restricting $\partial $ and projecting onto $\mathcal{A}_{\Theta }^{\min }$. $c^{\Theta }([w],[w^{\prime }])=0$ or $\pm 2$. In particular taking coefficients in $\mathbb{Z}_{2}$, $\partial ^{\Theta }=0$ and the $\mathbb{Z}_{2}$-homology of $\mathbb{F}_{\Theta }$ is freely generated by $\mathcal{S}_{[w]}^{\Theta }$, $[w]\in \mathcal{W}/\mathcal{W}_{\Theta }$. **Remark:** Let $w$ be minimal in its coset $w\mathcal{W}_{\Theta }$ and suppose that $u<w$ is of the form $w=ur_{\beta }$, with $\beta $ a simple root and $\Theta =\{\beta \}$. Hence $u$ is minimal in its coset. In fact, this conditions imply that $w\beta <0$. In fact, $w\beta =ur_{\beta }(\beta )=-u\beta $ and $u\beta \in \Pi _{w}$. So, if $u$ is not minimal in its coset, there is $\gamma >0$ such that $u^{-1}\gamma <0$ and $\gamma \in \langle \Theta \rangle $. As $w$ is minimal, by the same fact we know that $w^{-1}\gamma >0$ (otherwise we would have $\Pi ^{+}\cap w\Pi ^{-}\cap \langle \Theta \rangle \neq \emptyset $). This implies that $r_{\beta }(u^{-1}\gamma )>0$ and hence $u^{-1}\gamma =-\beta $. Hence $\gamma =-u\beta =ur_{\beta }(\beta )=w\beta <0 $. This is a contradiction because $\gamma >0$. **Example:** Let us consider the example of $G_{2}$ with $\Theta =\{\alpha _{1}\}$. We have the following cosets $$\mathcal{W}=\{1,r_{1}\},\{r_{2},s_{2}\},\{s_{1},r_{1}s_{2}\},\{r_{2}s_{1},s_{2}^{2}\},\{s_{1}^{2},r_{1}s_{2}^{2}\},\{r_{2}s_{1}^{2},s_{1}^{3}\}.$$The boundary maps for the minimal element in each coset is computed using the table \[table2\]. - $\partial_5 \mathcal{S}_{r_{2}s_{1}^{2}}=-2\mathcal{S}_{s_{1}^{2}}$; - $\partial_4 \mathcal{S}_{s_{1}^{2}}=0$; - $\partial_3 \mathcal{S}_{r_{2}s_{1}}=0$; - $\partial_2 \mathcal{S}_{s_{1}}=-2\mathcal{S}_{r_{2}}$; - $\partial_1 \mathcal{S}_{r_{2}}=0$. Hence - $H_{5}(\mathbb{F}_{\alpha _{1}},\mathbb{Z})=0$ (in particular $\mathbb{F}_{\{\alpha _{1}\}}$ is not orientable). - $H_4(\mathbb{F}_{\alpha_1},\mathbb{Z})= \mathbb{Z}_2$. - $H_3(\mathbb{F}_{\alpha_1},\mathbb{Z}) = \mathbb{Z} $. - $H_2(\mathbb{F}_{\alpha_1},\mathbb{Z}) = 0$. - $H_1(\mathbb{F}_{\alpha_1},\mathbb{Z}) = \mathbb{Z}_2$. As another source of examples, we refer to the papers [@Rab16] and [@RL182] which computes the coefficients of the isotropic and orthogonal Grassmannians. [^1]: Supported by FAPESP grant number 08/04628-6 [^2]: Supported by CNPq grant n$^{\mathrm{o}}$ 305513/2003-6 and FAPESP grant n$^{\mathrm{o}}$ 07/06896-5
{ "pile_set_name": "ArXiv" }
--- abstract: 'Quantitative understanding of human behaviors provides elementary comprehension of the complexity of many human-initiated systems. In this paper, we investigate the behavior of people on the $BBS$ forum by the statistical analysis of the amounts of view and reply of posts. According to our statistics, we find that the amounts of view and reply of posts follow the power law distributions with different power exponent. Furthermore, we discover that the amounts of view and reply of posts have nonlinear relationship. They are related by power function and show us straight line in log-log plot. Based on the estimation of slope and intercept of the line, we can characterize the behaviors quantitatively and know that people of Chinese forum and those of foreign forum have different preference towards replying to and viewing the posts. At last, we analyze the burstiness and memory in replying time series. They show some universal properties among different forum. All of them locate themselves in the high-$B$, low-$M$ region.' author: - | Jiefei Yu$^{1}$, Yanqing Hu$^{1}$, Min Yu$^{2}$, Zengru Di$^{1}$[^1]\ \ *1. Department of Systems Science, School of Management,*\ *Beijing Normal University, Beijing 100875, P.R.China*\ *2. Department of Information Technology and Management,*\ *Beijing Normal University, Beijing 100875, P.R.China* title: '**Analyzing Netizen’s View and Reply Behaviors on the Forum**' --- [**[Keyword]{}**]{}: Human Dynamics, $BBS$ forum, Burstiness and Memory [**[PACS]{}**]{}: 89.75.Da - Systems obeying scaling laws; 89.20.Hh - World Wide Web, Internet; 89.65.-s - Social and economic systems Introduction ============ Human behavior, as an academic issue in science, has a history about one century from Watson [@1]. As a joint interest of sociology, psychology and economics, human behavior has been extensively investigated during the last decades. A basic assumption embedded in the previous analysis on human dynamics is that its temporal statistics are uniform and stationary, which can be properly described by a Poisson process. Accordingly, the interevent time distribution should have an exponential tail [@2]. However, through studying the distributions of interevent time of the human behavior such as the people to send out or reply E-mail and surface mail, Albert Barabási discovered that these human behaviors present obvious deviation to the Poisson process: long time silent and short time high frequency eruption simultaneously present in these human behaviors, and their distributions of interevent time have power law fat tail [@3]. From E-mail communications [@3; @4] to surface mail communications [@5; @6; @7] and short message communications [@8], from financial activities to library loans [@6], from web browsing [@9] to on-line movie watchings [@10], and more and more examples show that the distributions of interevent time of the human behavior can not be described by a Poisson process but may use power function to fit well. According to the previous studies about human dynamics, we find that its methodology, excavating statistical laws from the historical records of human activities, can apply to analyze other issues. For instance, we can also use some of those datasets to quantify the herd behavior of an individual, that is to say, following the opinions of the majority of people in his/her social surrounding in an irrational way. With the development of information technology, web surfing has become a part of people’s daily life, including E-mail communication, dealing with financial matters, reading news online, browsing web, downloading software and so on. In the cyber world, $BBS$ forum provides people a platform for mutual communication. Those behaviors like “lurking, spamming, sofast”, etc. are the ways in which Netizens publicly express their ideas. The means of expression is so unique for its dual role of “openness” and “concealment”. Therefore the $BBS$ forum is loved by the majority of Internet users. On the $BBS$ forum, the new threads and Internet users are increasing, and the Internet users can browse information and express their own points. In the process of browsing information and expressing their own points, the most common group is formed on the $BBS$ forum, that is to say, browsing the posts of the same subject will form a provisional group, different themes has different groups. However, in the provisional group, some people simply browse information, and others will publish their own opinions. Then there are the amounts of view and reply on the $BBS$ forum. So what is the difference between the different groups, and how are these differences reflected are the interesting issues to be researched. In our paper, we focus on the actions of Internet users through the $BBS$ forum to solve these problems. The dates come from the $Sina$ Forum and $Chat Avenue$ Forum, the first one is the largest Chinese community and the other one is the current leader in free online chat rooms. Our paper is organized as follows. In Sec.\[sect2\], we present that the amounts of view and reply of posts which follow the power law distributions and have different power exponents. In Sec.\[sect3\], we address the nonlinearity relationship of the amounts of view and reply of posts and we also calculate the fitted line to characterize the forum. According to the fitted line, we find that the differences between the different forums. In Sec.\[sect4\], we analyze the burstiness and memory in replying time series. The amounts of view and reply {#sect2} ============================= Under computer program’s help, we can carry on the statistics to the amounts of view and reply about posts of different themes. Here we collect the dates from $Sina$ Community Forum and $Chat Avenue$ Forum. $Sina$ Community Forum is the world’s largest Chinese community and the Internet’s most comprehensive well-known $BBS$, has a huge core group of users, the theme of the plate covers culture, life, society, current affairs, sports, entertainment and other areas [@11]. We got the amounts of view and reply about posts of the entertainment, sports, life, economics themes from $Sina$ Community Forum. Economic Commentary: From April 17, 2001 to November 11, 2008, there were 32,480 posts. Pet Club: From December 1, 2005 to November 17, 2008, there were 39,950 posts. China Football: From June 14, 2001 to July 13, 2008, there were 95,891 posts. Pastime Puzzles: From November 19, 2004 to July 8, 2008, there were 49,346 posts. ![ The cumulative distribution functions $P(x)$ and their maximum likelihood power-law fits, for the posts of the amounts of view and reply from $Sina$ Community Forum.The red lines stand for the reply of cumulative distribution and the blue lines stand for the view of cumulative distribution. []{data-label="power"}](chinapost.eps){height="10cm" width="95.00000%"} $Chat Avenue$ is the current leader in free online chat rooms. Most of users are from the USA, Canada, Australia and the United Kingdom. And it offers a very active message forum with over 120,000 members, an arcade, and photo gallery [@12]. We got the amounts of view and reply about posts of World Events-Politics and News, Jokes, General, and Graphics Forum. World Events-Politics and News: From June 10, 2006 to February 23, 2009, there were 1,558 posts. Jokes: From June 10, 2006 to February 25, 2009, there were 1,583 posts. General: From June 10, 2006 to March 3, 2009, there were 1,250 posts Graphics Forum: From June 10, 2006 to March 13, 2009, there were 1,131 posts. According to the number of posts from $Sina$ Forum and $Chat Avenue$ Forum, we can see that the Chinese Internet Forums are more active than foreign ones. ![ The cumulative distribution functions $P(x)$ and their maximum likelihood power-law fits, for the posts of the amounts of view and reply from $Chat Avenue$ Forum. The red lines stand for the reply of cumulative distribution and the blue lines stand for the view of cumulative distribution. []{data-label="fp"}](foreign.eps){height="10cm" width="95.00000%"} A power law distribution is observed when we inspect the total number of views and replies a post receives (see FIG.\[power\] and FIG.\[fp\]), indicating that the vast majority of posts generates little interest, while a few posts are highly popular. At the same time, the distribution of the amounts of view and reply has the different power exponents. Mostly the power exponents of reply are larger than that of view, although the amounts of reply are smaller than that of view (see Table \[valuepower\]). The power exponents $\alpha$ range from $1.6$ to $3.5$. The values of power exponent and $x_{min}$ are based on the method of Ref.[@13]. From Table \[valuepower\], we also can discover that the exponents of view from $Chat Avenue$ Forum are bigger than those of view from $Sina$ Community Forum. But the exponent of reply is no much difference. This shows that the viewers of posts from $Sina$ Community Forum have greater range, this may be related to the forum rules — more attended posts will be topped and forum also can recommend some wonderful posts to Internet users. \[t\][|c|c|c|c|c|c|]{} Forum &Theme&&\ & &$\alpha$& $x_{min}$&$\alpha$& $x_{min}$\ &Economic Commentary&2.05& 256&2.91& 57\ $Sina$&Pet Club &1.94 &207&2.66& 13\ &China Football&1.87& 119&3.06 &60\ &Pastime Puzzles&1.62 &92&2.51& 46\ &World News &2.57&382&2.81&26\ $Chat Avenue$&Jokes &3.5&213&3.31&7\ &General &2.52&570&2.5&31\ &Graphics &1.98&175&3.22&49\ Besides, we also study other themes from $Sina$ Forum, all of them have the similar broad distribution. In the Table \[others\], we list the values of power exponent and $x_{min}$ of other themes from $Sina$ Forum. And they have the similar characteristic, namely the values of power exponent range from $1.6$ to $3.5$, and the power exponents of view are smaller than those of reply. \[t\][|c|c|c|c|c|c|]{} Forum &Theme&&\ & &$\alpha$& $x_{min}$&$\alpha$& $x_{min}$\ &Arrange Finance&2.07& 232&2.12& 9\ &Education Viewpoint &1.61 &1121&1.92& 3\ &Postgraduate Communication&2.04& 2372&2.13 &5\ &Diet Forum&1.92 &493&3.28& 17\ $Sina$ &08European Cup &2.05&162&2.55&35\ &Film Television World &1.81&1144&2.74&94\ &People Gossip &2.12&896&2.38&38\ &The ebb and flow of Overseas &1.93&394&2.69&47\ &Words on World&2.09&13689&2.18&8\ Some questions should be put forward, for example, why the distribution of the amounts of view and reply follow the power but have the different power exponents, and what relationship exists between the views and replies. We will solve these questions in the next section. The relationship of views and replies {#sect3} ===================================== The relationship between the amounts of view and those of reply can be observed through the FIG.\[scatterplot\], and it shows non-linearity which could be described by Eq.(\[eq1\]). $$\log y=A+B\log x\label{eq1}$$ ![ The relationship between the amounts of view and those of reply is approximately linear in the log-log coordinate.[]{data-label="scatterplot"}](cfrelationship.eps){height="10cm" width="95.00000%"} where $y$ stands for the amounts of reply, $x$ stands for the amounts of view. From analysis of Sect.\[sect2\], we know that the amounts of view and those of reply can be described by power law distribution. Therefore, here we suppose that Probability Density Function of views is $$f(x)=c_1 x^{-r_1}\label{eq2}$$ The Probability Density Function of replies is $$f(y)=c_2 y^{-r_2}\label{eq3}$$ From Eq.(\[eq1\]) and Eq.(\[eq3\]), we can get that $$f(y)=\frac{c_1}{B}e^{\frac{A(r_1-1)}{B}} y^{-(1+\frac{r_1-1}{B})}\label{eq4}$$ Contrasting to Eq.(\[eq3\]), we can find that $$c_2=\frac{c_1}{B}e^{\frac{A(r_1-1)}{B}}\label{eq5}$$ $$r_2=1+\frac{r_1-1}{B}\label{eq6}$$ From Eq.(\[eq5\]) and Eq.(\[eq6\]) we can evaluate that $$B=\frac{r_1-1}{r_2-1}$$ $$A=\frac{1}{r_2-1}\left(\log B+\log c_2 -\log c_1 \right)$$ Here $r_1$ and $r_2$ are known quantities, so we must evaluate unknown quantities $c_1$ and $c_2$. We know the Probability Density Function is $$f(z)=\left\{\begin{array}{ll} cz^{-r}&\textrm{$0<z<z_{max}$}\\ 0 &\textrm{others} \end{array}\right.$$ According to the characteristic of the Probability Density Function $$\sum_{z_{min}}^{z_{max}}cz^{-r}=\frac{n}{N}$$ We can evaluate the value of $c$. $$c=\frac{n}{N}\frac{1}{\sum_{z_{min}}^{z_{max}} z^{-r}}$$ where $n$ stands for the number of $z_{min}<z<z_{max}$, $z_{min}$ can be used from Table \[valuepower\], $z_{max}$ is the largest number which can easily get from the series of views or replies. $N$ stands for the all number of $z$. Therefore, we can easily evaluate the value of $c_1$ and $c_2$ $$c_1=\frac{n_1}{N}\frac{1}{\sum_{x_{min}}^{x_{max}} x^{-r_1}}$$ $$c_2=\frac{n_2}{N}\frac{1}{\sum_{y_{min}}^{y_{max}} y^{-r_2}}$$ So we can get the value of $A$ $$A=\frac{1}{r_2-1}\left(\log \frac{r_1-1}{r_2-1}+ \log \frac{n_2}{n_1}+\log \frac{\sum_{x_{min}}^{x_{max}} x^{-r_1}}{\sum_{y_{min}}^{y_{max}}y^{-r_2}} \right)$$ According to Eq.(\[eq1\]), we can know that the fitted line is determined by the parameters $A$ and $B$. For the same amounts of view, when the value of $B$ is larger, there would be more people to reply the post. The value of $e^A$ gives the probability of the first viewer to reply the post. The value of $e^{-A/B}$ gives the number of people who had viewed the post before there is the first reply to the post. In terms of the above analysis, from Table \[value\], we can know that when there is the same amounts of view, the groups in the $Chat Avenue$ Forum reply more than the groups in the $Sina$ Forum. Most people in the $Sina$ Forum prefer viewing to replying the posts. The differences between $Chat Avenue$ Forum and $Sina$ Forum come from the different culture and social environment possibly. \[t\][|c|c|c|c|c|]{} Forum&Theme&$A$&$B$&$-A/B$\ &Economic Commentary &-2.74 &0.550 &4.98\ $Sina$&Pet Club &-3.35 &0.566&5.92\ & China Football&-3.44 &0.422&8.15\ &Pastime Puzzles&-3.70 &0.411&8.39\ &World News &-5.17&0.967&5.35\ $Chat Avenue$& Jokes &-7.83&1.08&7.25\ & General &-3.53&1.01&3.50\ &Graphics &-0.456&0.441&1.03\ Burstiness and memory {#sect4} ===================== The dynamics of a wide range of real systems display a bursty, characterized by short timeframes of intense activity followed by long times of no or reduced activity. In Ref.[@13] K.-I. Goh and A.-L. Barabási propose to characterize the bursty nature of real signals using orthogonal measures quantifying two distinct mechanisms leading to burstiness: the interevent time distribution (see Eq.\[eq12\]) and the memory(see Eq.\[eq13\]). $$B=\frac{\sigma_{\tau}-m_{\tau}}{\sigma_{\tau}-m_{\tau}} \label{eq12}$$ where $m_{\tau}$ and $\sigma_{\tau}$ are the mean and the standard deviation of $P(\tau)$, respectively. $$M=\frac{1}{n_{\tau}}\sum_{i=1}^{n_{\tau-1}}\frac{(\tau_{i}-m_{1})(\tau_{i+1}-m_2)}{\sigma_1\sigma_2} \label{eq13}$$ where $n_{\tau}$ is the number of interevent times measured from the signal and $m_1(m_2)$ and $\sigma_1(\sigma_2)$ are sample mean and sample standard deviation of $\tau_i$’s ($\tau_{i+1}$’s), respectively($i$=$1,..., n_{\tau} -1$). The time of posting and replying can be as a time signal, then we can easily calculate the interevent time distribution and the memeory (see Table \[tab3\]). \[t\][|c|c|c|c|]{} Forum &Theme&$B$&$M$\ &People Gossip &0.8243 &0.0378\ $Sina$& The ebb and flow of Overseas &0.7856&0.0457\ &Words on World&0.7296&0.01108\ &World News&0.8897 &0.0057\ $ChatAvenue$ &Jokes &0.8337&0.0057\ &Graphics &0.8258&0.0348\ In Table \[tab3\], the themes of People Gossip, The ebb and flow of Overseas and Words on World belong to $Sina$ Community Forum, and the themes of World News, Jokes and Graphics come from $Chat Avenue$ Forum. From the Table \[tab3\], we can discover that all the themes of the values $B$ and $M$ have no much differences which will further explain that the replying behaviors of $Sina$ Community Forum and $Chat Avenue$ Forum are similarity. All of them can illustrate that the replying time series are bursty signal, in other words, the front people who reply the post can not affect the behind ones, they would reply a post depending on their interests. Conclusion and discussion ========================= In this paper we discovered that the amounts of view and reply on $BBS$ forum appear the power law distributions, the power exponents range from $1.6$ to $3.5$. Moreover, mostly the power exponents of reply are bigger than that of view. Furthermore, we found that the amounts of view and those of reply has the nonlinear relationship. According to the relationship, we presented two parameters — the slope and intercept of the log-log straight line which can evaluate human’s interests on replying or viewing. With the above measurements and statistics, we find some differences between $Sina$ Community Forum and $Chat Avenue$ Forum, which may be caused by cultural, social environments, and forum rules. At last, we analyze the burstiness and memory in replying time series. It could be found that all the time series have similar qualitative properties. The $Sina$ Community Forum and $Chat Avenue$ Forum have no much differences in replying time series. Acknowledgement {#acknowledgement .unnumbered} =============== This paper is supported by NSFC under the grant Nos. 70771011,70471080. [99]{} J. B. Watson, Psychology as the Behaviorist Views It, Psychological Review **20**, 158 - 177(1913) F. A. Haight, Handbook of the Poisson Distribution, Wiley, New York, (1967) A. -L. Barabási, The origin of burst and heavy tails in human dynamics, Nature **435**, 207 - 211(2005) A. -L. Barabási, K. -I. Goh, A. Vázquez, Reply to Comment on ‘The origin of bursts and heavy tails in human dynamics’, arXiv: physics/0511186 J. G. Oliveira, A. -L. Barabási, Darwin and Einstein correspondence patterns, Nature **435**, 1251(2005) A. Vázquez, J. G. Oliveira, Z. Dezs$\ddot{o}$, K. -I. Goh, I. Kondor, A. -L. Barabási, Modeling burst and heavy tails in human dynamics, Phys. Rev. E **73**, 036127(2006) N. N. Li, Z. J. Deng, N. Zhang, J. M. Li, Y. Qin, T. Zhou, Empirical analysis on temporal statistics of human correspondence patterns, Physica A **387**, 6391(2008) W. Hong, X. P. Han, T. Zhou, B. -H. Wang, Heavy-Tailed Statistics in Short-Message Communication, CHIN. PHYS. LETT **26**, 028902(2008) Z. Dezsö, E. Almaas, A. Lukács, B. Rácz, I. Szakadát, A. -L. Barabási, Dynamics of information access on the web, Phys. Rev. E **73**, 066132(2006) T. Zhou, H. A. -T. Kiet, B. J. Kim, B. -H. Wang, P. Holme, Role of activity in human dynamics, EPL **82**, 28002(2008) http://people.sina.com.cn/ http://www.chat-avenue.com/about.html Aaron Clauset, Cosma Rohilla Shalizi, M. E. J. Newman, Power-law distributions in empirical data, arXiv: 0706.1062vl K. -I. Goh and A. -L. Barabási, Burstiness and memory in complex systems, EPL **81**, 48002(2008) [^1]: Author for correspondence: [email protected]
{ "pile_set_name": "ArXiv" }
--- author: - | [**J.M. Gordillo**]{}, Departamento de Ingenería Aeroespacial y Mecánica de Fluidos, Universidad de Sevilla, Spain\ [**J. Rodríguez-Rodríguez**]{}, Grupo de Mecánica de Fluidos, Universidad Carlos III de Madrid, 28911, Leganés, Spain title: 'Comment on “Revision of Bubble Bursting: Universal Scaling Laws of Top Jet Drop Size and Speed”' --- In a recent Letter [@PRL2017] Gañán-Calvo presents scalings, in very good agreement with experiments, for the velocities and diameters of the first jet drops produced after bubble bursting. It is the purpose of this comment to show that the physical arguments given to explain such scalings are not consistent with theoretical, experimental and numerical evidences reported in the literature, and confirmed by our own simulations. Indeed, the low-Oh limit of Eq. (7) in [@PRL2017] expresses that the jet velocity is $V\propto V_c\,Oh^{1/2}$, with $V_c\propto \sqrt{\sigma/(\rho\,R_0)}$ the capillary velocity and $Oh=\mu/\sqrt{\rho\,R_0\,\sigma}$ the Ohnesorge number. This result, combined with Eq. (4) in [@PRL2017] yields $V'=V_c$. In addition, the justification of the scalings in [@PRL2017] rests on the assumption that all terms in the momentum equation are of the same order of magnitude at the instant of jet ejection, $\rho\,V^2/L\sim \mu V'/L^2\Rightarrow \rho\,V^2\sim \mu V'/L$ from which, since $V\propto V_c\,Oh^{1/2}$ and $V'=V_c$, it is deduced that the unknown length $L$ coincides with the initial radius of the bubble, namely, $L\propto R_0$. In [@PRL2017], the jet velocity $V$ is deduced from the mass balance $V\,R^2\propto V'\,L\,R$, a fact meaning that the proposed physical mechanism assumes that the fluid entering the jet comes from a region of width $L\propto R_0$ of velocity $V'\propto V_c$. Thus, it is hypothesized in [@PRL2017] that viscosity sets in motion a region of width comparable to the initial radius of the bubble with a velocity comparable with the capillary velocity $V_c$, a fact which is in contradiction with the well known results by [@Moore], where it is shown that the width of the region where vorticity produced by viscous effects to comply with the shear-free condition at the interface is concentrated in a boundary layer region of width much smaller than $R_0$. Indeed, the numerical results in figure 1 illustrate that the thickness of the boundary layer formed in the high curvature region located at the base of the ejected jet, which depends on the Ohnesorge number, is far smaller than $R_0$. ![\[fig:vorticity\]Vorticity contours for a capillary wave generated upon the recoil of the bubble cap and that focuses at the apex of the cavity leading to the formation of a Worthington-like jet. It is worth remarking that the thickness of the boundary layer induced by the wave propagation is of the order of the wavelength or amplitude, and much smaller than the cavity size. This case corresponds to $Bo = 0.05$ and $Oh = 10^{-2}$.](figure_vorticity_Oh0008.jpg){height="4.5cm"} \ Contrarily to the physical mechanism presented in [@PRL2017], which attributes to viscosity the origin of the jet generated after the bursting of a bubble resting on a free interface, in [@PRLBursting] we reveal that such jet emerges as a consequence of a purely inertial mechanism, triggered by the presence of the fastest capillary wave which, once generated during the rim retraction process, propagates towards the apex of the air cavity [@Duchemin2002], breaking the self similarity of the inertio-capillary collapse of the void [@Zeff; @Sierou; @Deike; @Thoroddsen_2018]. However, viscosity plays a role in the selection of the wavelength breaking the self-similarity, and thus in the modulation of the jet’s initial velocity.\ Funding from MINECO under Projects DPI2017-88201-C3-1-R and DPI2017-88201-C3-3-R is acknowledged. We thank our colleagues and friends for useful suggestions. [6]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , , , , ****, (). , , , , ****, (). , ****, (). , , , , , , , ****, (). , , , , ****, ().
{ "pile_set_name": "ArXiv" }
--- abstract: 'The H.E.S.S. experiment in Namibia, Africa, is a high energy gamma ray telescope sensitive in the energy range from $\sim$100Gev to a few tens of TeV, via the use of the atmospheric Cherenkov technique. To minimize the systematic errors on the derived fluxes of the measured sources, one has to calculate the impact of the atmospheric properties, in particular the extinction parameter of the Cherenkov light ($\sim$300-650nm) exploited to observe and reconstruct atmospheric particle showers initiated by gamma-ray photons. A lidar can provide this kind of information for some given wavelengths within this range. In this paper we report on the hardware components, operation and data acquisition of such a system installed at the H.E.S.S. site.' address: 'LUPM, IN2P3/CNRS and Un.Montpellier II' author: - 'J. Bregeon, M. Compin, S. Rivoire, M. Sanguillon, G. Vasileiadis [^1]' title: 'An elastic lidar system for the H.E.S.S. Experiment' --- Lidar; Atmospheric monitoring; Gamma Ray Astronomy Introduction ============ The H.E.S.S. (High Energy Stereoscopic System) experiment consists of 5 imaging Cherenkov telescopes situated in the Namibia Khomas Highland desert (1800m asl). Its main objective is the study of galactic and extragalactic sources in the energy range of $\sim$100GeV to a few tens of TeV coupled to a substantial flux sensitivity (1$\%$ Crab units). The combination of the 5 telescopes data analysis provides good background rejection and angular resolution. The detection technique used by H.E.S.S., namely the measurement of Cherenkov light emitted from secondary particles created 5 to 10 km above ground level, demonstrates by itself the importance of knowing any variation of the atmospheric quality [@H.E.S.S.]. Precise flux and energy spectra calculations for the observed sources could suffer if the level of extinction of cherenkov-generated photons varies due to aerosols or thin particles present in the atmosphere during the gamma shower development. Similar experiments use by default the uniform rate of background cosmic-ray initiated showers as a measurement of atmospheric clarity[@CTC]. To obtain more precise estimations additional instruments such as radiometers, transmissometers or lidars can be implemented. Operating a lidar during the data-taking phase permits us to model the atmospheric transmission above the site, which in sequence could be used to simulate different background conditions by varying the aerosol density, much like the real background data. Once this is achieved, these transmission tables are used to produce corrected tables for energy and effective area by means of gamma-ray simulations. ![Aerial view of the H.E.S.S. site, the lidar hut is visible at the lower right part of the photo. The distance between the telescopes and the lidar been 850m to avoid any interference due to the laser shots.[]{data-label="fig1"}](hess.jpg){width="8.0cm"} This paper describes the design, operation and performance of a lidar system installed at the H.E.S.S. experiment site, fig.1. It is organized as follows. We give a summary of the relevant atmospheric parameters in section 2. Section 3 describes the current lidar hardware. In section 4, the signal detection and processing is presented. First results of the analysis of the atmospheric parameters are summarized in section 5. A summary concludes the paper. ![Characteristics of a Cherenkov shower development of an incoming $\gamma$ ray particle in the atmosphere.[]{data-label="fig2"}](cher.png){width="5.cm"} Atmospheric Monitoring ====================== In order to study very high energy cosmic rays, the atmosphere can be used as the Cherenkov radiator, while the radiation is imaged onto single photon sensitive detector planes of telescopes with large parabolic mirrors. A $\gamma$ ray interacting in the atmosphere will produce an air shower of secondary particles at an elevation of 10-15km. The shower particles produce Cherenkov light and the resulting light cone will cover a disk of a radius of 100-120m at the ground level with an intensity of 10-100 photons m$^{-2}$, fig.2. The telescopes looking up the night sky will be able to detect the Cherenkov light and measure the intensity, orientation and shape of the air shower, which is related to the primary energy and direction of the $\gamma$ ray. The loss of Cherenkov light from a shower (as viewed by the H.E.S.S. telescopes) is mainly due to molecular and Mie scattering. The presence of aerosols can affect the data two-fold. If close to ground level the recorded photon yield can be lower affecting the telescope trigger thus affecting the reconstructed shower energy. If close to the shower maximum, at around 10km height , the shape and brightness of the camera images would be affected. The introduction of lidar measurements could permit the calculation of the extinction parameter of the atmosphere $\alpha$ thus reducing the systematic error on the final energy spectrum calculation. The back-scattered signal emitted by a lidar, after being back-scattered by the atmosphere and collected by a telescope, is measured by a photomultiplier tube (PMT). The light pulse height from a distance R is given by the lidar equation $$P(R)=K\frac{\beta(R)}{R^{2}}e^{-2\tau(R)}$$ where $\beta(R)$ is the back-scattering coefficient and $\tau(R)$ is the integral of the extinction coefficient $\alpha(r)$ along the path. Both these quantities are sums of the aerosol and molecular contributions: $\alpha(R)=\alpha_{mol}(R)+\alpha_{aer}(R), \beta(R)=\beta_{mol}(R)+\beta_{aer}(R)$. The molecular part can be evaluated either using standard models, knowing the pressure and temperature vs. height from atmospheric measurements or radiosonde data from nearby meteorological sites. On the other hand the extraction of the aerosol coefficients, based on the lidar equation, demands algorithms based on either the Klett [@klett] or Fernald [@fern] methods. Since both methods need a priori assumptions about the molecular vs. aerosol parameter, the calculated value of $\alpha$ comes with a relatively important systematic error. For the purposes of this paper we have used the Fernald approach to present a preliminary analysis of our data while to evaluate the molecular contribution we used also radiosonde data available for our site. ![Schematic diagram of the H.E.S.S. lidar System. The Quantel laser is mounted on the right-hand side of the telescope, while a pair of photomultipliers are situated behind the primary mirror and at a distance of few cm further away than the optimal focal point of thje telescope to achieve a more uniform illumination of the photomultplier’s cathode.[]{data-label="fig3"}](lidarm.png){width="7cm"} Lidar Hardware and Data Acquisition =================================== The lidar telescope was conceived in 1997 and construction was completed in 1999. Until 2001 it was installed at the Themis site (Pyrenees,France) for the purpose of the Celeste experiment [@celeste]. A mechanical aspect of the telescope structure and light collection is sown in figure 3. Due to poor weather conditions on site it was used only for short intervals which led to frequent mechanical failures. During summer 2004 it was transferred and installed at the H.E.S.S. site, in Namibia (Africa). It was installed in a dedicated hut, situated at a distance of 850m from the H.E.S.S. facilities, minimizing the possibility that the cameras of the H.E.S.S. telescopes sees the lidar laser beam. After a complete rebuild, involving replacement of obsolete parts and software upgrades, it started routine data taking operations during the summer of 2006. Mirror and Mount ---------------- For the collection of the backscattered light, the lidar uses a Cassegrain type telescope, fig. 3. The primary parabolic mirror is of $\Phi=60$cm diameter with a focal length of $\phi=102 cm$, while the secondary has a diameter of 8cm and $\phi=10$cm. The mirrors were produced by Compact using BK7 glass coated with aluminum and a reflectivity of $80\%$ in the range 300-600nm. The average spot size in the focus is 1.5mm FWHM. It is mounted in a fully steerable alt-azimuth frame equipped with DC servomotors with a maximum speed of $5^{o}$/s. The absolute pointing direction is close to $0.7^{o}$ accuracy. A three point mounting system allows for alignment and collinearity of the mirror and laser beam. The whole apparatus is installed in a 5$\times$5m hut equipped with a motorized roof, protecting it from rain and harsh conditions when the lidar is not in operation. All motors can be controlled either locally or remotely via Ethernet connections. Laser ----- The choice of the laser for our purpose is dictated by the following requirements: the wavelength of the laser has to match as much as possible the observed Cherenkov photons energy spectrum (300-650nm). It follows that a double wavelength laser is needed, giving us two reference points to compare with this spectrum. Repetition rate should be high enough to reduce collection time, while the laser power should be adjustable to avoid interference with the H.E.S.S. telescope optical systems. To meet these requirements a Quantel Brilliant 20 Nd:YAG laser was used. It is equipped with two cavities generating the 2nd and 3rd harmonics at 532 and 355nm. The repetition rate is 10Hz while the per-pulse energy is 180mJ and 65mJ respectively. The laser is mounted aside the telescope structure. It is guided in a bi-axial configuration at a distance of 43cm from the optical axis of the telescope. Signal Detection, Trigger and Digitization ------------------------------------------ A pair of Photonis XP2012B photomultipliers are used for backscatter light detection. The return signal is split by means of a dichroic filter, mounted at the focal point, separating the 532nm from the 355nm component. We use the photon counting method as a measurement which results in a saturation of the signal for the first few hundred meters. The photomultiplier gain reaches the value of $3\times10^{6}$ at 1450V. The whole 51mm-diameter photocathode window is exposed to maximize light detection. The output signals are fed through a pair of 1.5m long cables to a two-channel 12bit Compuscope Octopus CAGE 8265 Digitization board that runs at 65Mhz. This gives us an altitude resolution of 2.5m. A series of PCI boards command the servomotors and hydraulic brakes. Finally a Labview interface supervises the whole sequence of steering and data-taking of the lidar system on a Windows XP based industrial CPU. The laser system is being driven by means of a client-server protocol. The server is running on the H.E.S.S. DAQ system while the dedicated lidar CPU runs the client part. Upon reception of an acquisition demand, accompanied by several parameters such as pointing direction, laser power and run duration, the client executes the required configuration , points the telescope and triggers the laser system up to the nominal laser repetition rate. The DAQ is triggered by the laser Q-switch synchronization pulse generated at every successful laser shot. Operation Mode and Maintenance ------------------------------ The lidar system can be operated in two ways. Either in stand-alone mode or in slave mode where it is being driven by the H.E.S.S. central data acquisition system. The idea was that during the first months of operations we used the standalone mode for a complete debugging and data analysis optimization while later we switched to slave mode for normal data taking. During normal H.E.S.S. operation the dedicated lidar runs are executed within the time interval that separates the consecutive physics runs, a 180 sec time interval. A 1200 laser shots profile is executed every time and stored in the H.E.S.S. database in form of a ROOT file [@root]. During the commissioning phase of the lidar, it was pointed out by the H.E.S.S. experiment that the option to steer the laser beam of the lidar could cause some issues with the operation of the Cherenkov telescopes. The danger being that the laser beam crosses the field of view of the telescopes while in operation, resulting in damage to the telescope’s camera readout system. It was decided then to point the lidar telescope and laser beam in a fixed direction ($75^{o}$ azimuth and $25^{o}$ WE direction) to avoid any overlap between the two experiments. Maintenance was centered around the re-alignement of the optical system. This operation was repeated once a year to assure optimal operation of the lidar. At the same occasion the laser flash lamp was also being replaced even if the typical lifetime ($>50$ millions shots) was not reached. The entire de-ionized water cooling system was also being serviced and filters were replaced on a yearly basis as well. Since during winter time the temperature could drop below zero during the night, the laser temperature head was kept between 5 and 30 degrees by means of a temperature regulated air flow system. ![Backscatter signal obtained from a distance up to 3000m using horizontal laser acquisitions. Solid line correspond to simulated performance using the OSLO optical design software for our setup [@oslo].[]{data-label="fig4"}](ov1.png){width="11cm"} Signal Treatment ---------------- The recorded signal is described by the so-called lidar equation : $$P(r)=P_{0}\frac{ct_{0}}{2}\beta(r)\frac{A}{r^{2}}e^{-2\tau(r)}=P_{0}\frac{ct_{0}}{2}\beta(r)\frac{A}{r^{2}}e^{-2\int^{r}_{0}a(r')dr'}$$ where $P(r)$ is the signal received at time $t$ from photons scattered at a distance $r$ from the lidar, $P_{0}$ is the transmitted laser power, $t_{0}$ is the laser pulse duration, $\beta(r)$ is the backscattering coefficient, $\tau(r)$ is the optical depth, $\alpha(r)$ is the extinction coefficient, and $A$ is the effective area of the detector. The latter is proportional to the overlap function of the telescope. Using the experimental method described in [@overp], we calculated the overlap function using horizontal scans. The results (shown in fig.4) are compatible with an analytical calculation regarding the minimum useful observation altitude, calculated at around 2000m. For analysis purposes we used an initial altitude of $r_{0}$=800m above site due to signal saturation due to stray light entering the telescope tube, still we have applied the overlap correction factor shown above for the rest of the data points. Data after background subtraction, calculated in the range of 20-25km, where we expect no laser signal, were smoothed using the typical Savinsky-Golay algorithm. Resolution was also reduced from the initial 2.3m to 92m using the gliding window technique. This was necessary to improve the statistics per altitude unit but also to be coherent with the expected Monte Carlo resolution used with the H.E.S.S. energy spectra reconstruction algorithms. On the latter the atmosphere is divided in layers of a given thickness (200m) and all effects due to aerosol or molecule presence are bein g studied using this minimum path length. --------------------------- ----------- -- -- -- Marine particles 20-35 sr Sahara dust 50-80 sr Urban particles 35-70 sr Biomass burning particles 70-100 sr --------------------------- ----------- -- -- -- : Typical Lidar ratio values at 532nm and different aerosol types Analysis method and data sets ============================= Backscattering and extinction are both caused by particles and molecules. The first contribution is related to aerosol and cloud presence, while the second is due to Raleigh scattering. It follows that we can write from equation (2) : $$\begin{aligned} \beta(r)=\beta_{aer}(r)+\beta_{mol}(r) \\ \alpha(r)=\alpha_{aer}(r)+\alpha_{mol}(r)\end{aligned}$$ The molecular scattering part can be determined using pressure and temperature profiles from meteorological measurements or approximated from appropriate atmospheres, in our case desert type. So the only part that needs to be determined is the aerosol scattering part. It is common to introduce two ratios in this case, the particle extinction-to-backscatter ratio (lidar ratio) $$L_{aer}(r)=\frac{\alpha_{aer}(r)}{\beta_{aer}(r)}$$ and in similar manners the molecular lidar ratio $$L_{mol}(r)=\frac{\alpha_{mol}(r)}{\beta_{mol}(r)}=\frac{8\pi}{3}sr$$ Typical values for the lidar ratio $L_{aer}$ ( shown in Table 1) have been discussed in the literature [@profi]. Taking into account these considerations and after some integrations, equation 2 takes the following Bernoulli type form : $$\begin{aligned} \frac{d ln(S(r)L_{aer}(r) exp\left \{ -2 \int_0^r (L_{aer}(r')-L_{mol}) \beta(r')_{aer} dr' \right \})}{dr} \\ =\frac{1}{Y(r)}\frac{dY(r)}{dr}-2Y(r)\end{aligned}$$ where S(r) corresponds to the the range-corrected signal $S(r)=R^2P(r)$ and $Y(r)=L_{aer}(r) \left \{\beta(r)_{aer}(r)+\beta(r)_{mol}(r)\right \}$. To resolve this equation we follow the Fernal method, assuming a reference range where no aerosol contribution to the backscattered signal is expected and integrating from then on (backward integration). For the last two and a half years we have operated our lidar in a regular base following the observation schedule of the H.E.S.S. experiment. A total of 2650 profiles, both at 355nm and 532nm, were accumulated covering a wide variety of weather conditions. Most of the datasets concern the Namimbian winter periods, June to October, where the H.E.S.S. experiment is most active and naturally only during night time. The results presented here are based on a subset of these data after omitting nights where weather conditions were bad enough that H.E.S.S. observations weren’t possible, namely rain, heavy winds or complete overcast skys. On the other hand our lidar was used intensively when aerosol levels were elevated or thin clouds were present. A particular case where we have focused our efforts concerned the period where fires and other biomass burning activities are very active in the region. Such events are expected to influence the total trigger rate of the H.E.S.S. telescope so a dedicated campaign was conducted. ![Normalized range corrected return signal (blue line). Green line correspons to the molecular corss section calculated profile.[]{data-label="fig5"}](li.jpg){width="12cm"} Results ======= ![Typical extinction profiles at 355 (left) and 532 (right) nm obtained during summer 2013 on site. Solid lines corresponds to Rayleigh scattering after pressure and temperature from radio sonde profiles taken at a nearby station.[]{data-label="fig6"}](prof.png){width="11cm"} Atmospheric transparency ------------------------ The lidar range-corrected signals were first compared and normalized to the molecular scattering cross section profile computed from measured pressure and temperature on site. A suitable region where no aerosol or cloud presence was expected was chosen for this purpose, typically between 6 and 8km above site. A typical case is shown in fig.5. Following this, a reference range-bin has been set as the calibration point required for the Klett-based analysis. The altitudinal distribution of $\alpha_{obs}$ was then calculated. Examples of the extinction coefficient obtained from these measurements and for various conditions are shown in fig.6. We use data points up to 11km since above this altitude the return signal was too weak. The solid line on these plots represents the $\alpha_{Ray}$ calculated for pure Raleigh scattering. We used a lidar ratio of 70 for all our analyses, a value that corresponds to desert conditions, see table 1. Error bars indicate one standard deviation caused by statistical uncertainties and have been calculated from the law of error propagation by assuming a Poisson noise distribution on lidar signals. ![Timeline of the extinction coefficient measured during summer 2013. A biomass type burning effect is visible towards the end of this period. Extensive presence of dust particles is clearly seen up to altitudes of 5km above sea level. []{data-label="fig7"}](biob.png){width="12cm"} As it was mentioned before, an interesting effect during the winter period in Africa is the biomass burning effect. This is mainly due either to fires been initiated by farmers to enrich their fields or by accidental or weather-related reasons. High concentrations of smoke and dust produce relatively significant concentrations of aerosols at low altitudes. The timeline plot shown in fig.7 represents extinction coefficients calculated for the period of 6 June to 24 November 2013. Data were taken every second night, accumulating on an average 8-10 profiles per night, evenly distributed throughout the night period. Extinction was measured on both wavelengths 355nm and 532nm (only 532nm shown). A strong concentration of low altitude aerosols and dust, up to 3000m above the lidar horizon, has being observed between the 5th of September and 16th of November, a period that coincides with the above described local activities ![Variations of the optical depth based on datasets of winter 2013. Red line is based on MODTRAN5 typical desert aerosol profile on the right side the Ansgtrom exponent 532nm/355nm for the same period. Particle size corresponds to that due to biomass burning effect.[]{data-label="fig8"}](oda.png){height="8cm" width="11cm"} Vertical Aerosol Optical Depth and Angstrom exponent ---------------------------------------------------- In this section we evaluate the opacity of the atmosphere in terms of the Aerosol Optical Depth (AOD). The AOD for a given height r, $\tau_{AOD}$, is defined as the integral of the extinction coefficient $\alpha_{obs}$ between the Lidar ground level and r as follows : $$\tau_{AOD}(r)=\int_0^r\alpha_{obs} (r')dr'$$ The median of $\tau_{AOD}$(r) for the 355nm and 532nm line as a function of altitude is shown on fig.8. This covers the time period mentioned above. Superimposed, red line, is a typical profile expected using the a MODTRAN5 [@mod5] simulation package. We used a mixture of desert climat with aerosol presence parametrization, while the atmospheric conditions corresponded to the mean vlaues ones registered on site for this specific period. Overall accordance is satisfactory, while the variations of the measured AODs are expected since cloud and dust presence on site is know to vary substantially with a single night. Using the above obtained AODs for these two wavelengths we were able to calculate the Angstrom exponent 532nm/355nm shown on fig.8, right side plot. Values vary between 1.3 and 2.2, corresponding to biomass burning types of particles, in accordance with the observed period and related activities in the region. Conclusions =========== We’ve described the H.E.S.S. elastic lidar, and explain operation and maintenance matters. We’ve shown that the lidar works well, provides good quality data and that we’re able to analyse them to show the presence of biomass burning aerosols on site. Next steps are to fine tune the analysis, show the impact of aerosols on the instrument response, correct the instrument response using atmospheric transmission from lidar data in simulations, and show we can retrieve Cherenkov data for not perfectly clear nights, and demonstrate the validity of the procedure on one or more bright sources, see for example [@magic]. All this we’ll be published in a near future. Main issue of the elastic lidar is that you have to assume an aerosol type, what leads to quite large uncertainties, that’s why we’re working on a Raman lidar for CTA. Konrad Bernl[" o]{}hr, AP, Vol. 12, Issue 4, pp. 255-268 (2000) J.Hahn et al., EPJ Web of Conferences 89, 02002 (2015) M.Naurois, Gamma 2012, AIP Conf.Proc. Vol. 558, pp. 540-543 (2001) R.Brun et al., ROOT, NIM. in Phys. Res. A 389 p 81-86 (1997) http://www.lambdares.com/oslo J. Klett, Applied Optics, Vol. 24, Issue 11, pp. 1638-1643 (1985) D.M[" u]{}ller et al., J. Geoph. Res, Vol. 112, D16202 (2007) F.G Fernald, Applied Optics, Vol. 23, pp. 652 (1984) Y. Sasano et al, Applied Optics, Vol. 18, pp. 3908 (1979) MODTRAN5 (Berk et al., 2006) C.Fruck et al., EPJ Web of Conferences 89, 02003 (2015) [^1]: Corresponding author: [email protected]
{ "pile_set_name": "ArXiv" }
=15.5pt ${}$ [0.05cm]{}[0.05cm]{} [**Unruh detectors and quantum chaos in JT gravity** ]{} \ **Abstract** We identify the spectral properties of Hawking-Unruh radiation in the eternal black hole at ultra low energies as a probe for the chaotic level statistics of quantum black holes. Level repulsion implies that there are barely Hawking particles with an energy smaller than the level separation. This effect is experimentally accessible by probing the Unruh heat bath with a linear detector. We provide evidence for this effect via explicit and exact calculations in JT gravity building on a radar definition of bulk observables in the model. Similar results are observed for the bath energy density. This universal feature of eternal Hawking radiation should resonate into the evaporating setup. Introduction {#sect:1} ============ One of the main features of finite volume AdS/CFT is that the CFT in question is a unitary and discrete system. The latter follows from the fact that the boundary on which the CFT lives is compact. This implies that quantum gravity in AdS must be a unitary and discrete quantum system. There is however at least one further constraint on the spectrum of quantum gravity in AdS on which we want to focus in this work. It is a consequence of the fact that black holes are chaotic quantum systems [@bhrm]. Roughly speaking there are two hallmarks of chaotic quantum systems. The first is exponential dependence on changes in initial conditions. In a quantum theory this translates into the exponential growth of operators with time [@operatorgrowth]. One popular way to probe this effect is by computing out-of-time-ordered correlators [@SS; @SSmultiple; @Shenker:2014cwa; @bound]. The gravitational translation of this fast scrambling is the exponential redshift close to gravitational horizons, resulting in gravitational shockwaves [@SS; @SSmultiple; @Shenker:2014cwa]. The second hallmark of quantum chaotic systems concerns their level statistics. In particular quantum chaotic systems are very well characterized by the property of level repulsion. Roughly speaking two subsequent energy levels of a chaotic system are rarely close together. A more precise version of this is that the local spectral statistics of any quantum chaotic system can be described using random matrix statistics.[^1] Black holes are fast scramblers [@sekinosusskind] and shockwave interactions are a universal feature of quantum gravity. Therefore quantum black holes are expected to be chaotic quantum systems [@bhrm]. But if quantum gravity is quantum chaotic, then what is the bulk gravitational interpretation of level repulsion and of random matrix level statistics? Recently this question has been addressed within Jackiw-Teitelboim (JT) gravity [@sss2].[^2] The answer in this theory is that Euclidean wormhole contributions to the gravitational path integral are responsible for level repulsion and random matrix level statistics in the spectrum of the gravitational theory. We expect this conclusion to be true universally.\  \ We conclude that any acceptable theory of quantum gravity in AdS is a discrete and unitary quantum system with random matrix level statistics. One prime consequence of this spectral behavior is the specific late time behavior of boundary correlators. Consider for example the two-point function of a discrete quantum chaotic system with an $L$ dimensional Hilbert space with levels $\lambda_1\dots \lambda_L$: $$\begin{aligned} {\left\langle {\mathcal{O}}(0){\mathcal{O}}(t) \right\rangle}_\beta &= \int_{\mathcal{C}}d E_1\,e^{-\beta E_1}\,\int_{\mathcal{C}}d E_2\,e^{it(E_1-E_2)}\,\rho(E_1,E_2)\,\rvert{\mathcal{O}}_{E_1 E_2}\rvert^2\,,\label{erraticbdy}\end{aligned}$$ where $$\rho(E_1,E_2)=\sum_{i=1}^L \delta(E_1-\lambda_i)\sum_{j=1}^L \delta(E_2-\lambda_j)\,.\label{spikes}$$ For a quantum chaotic system the eigenvalue thermalization hypothesis [@eth1; @eth2; @bhrm; @phil] states that $\rvert{\mathcal{O}}_{E_1 E_2}\rvert^2$ is a smooth function of $E_1$ and $E_2$. At early times the Fourier transform in cannot distinguish the delta functions from a coarse-grained version of this spectrum. The result is that at early times this correlator decays exponentially with time. This is the quasinormal mode decay known from quantum fields on a black hole background. At exponentially late times however the correlator essentially oscillates erratically around an in general nonzero average [@maldainfo]. The nonzero averaged value is explained by random matrix theory [@bhrm]. The erratic oscillations are testimony to the fundamental discreteness of the theory. In fact this erratically oscillating behavior can itself be viewed as distinguishing a chaotic from a regular quantum system [@haake]. In gravity, the bulk explanation for the late time behavior of this correlator involves rather exotic gravitational effects. In particular in order to explain the nonzero average one needs to sum over an infinite number of Euclidean wormhole amplitudes in the Euclidean path integral [@sss; @sss2; @Blommaert:2019hjr; @phil]. To explain the erratic oscillations in JT gravity from the Euclidean path integral one furthermore needs to realize that a single discrete quantum chaotic system can be described by a version of JT gravity which includes a specific set of branes in the gravitational path integral [@paper6; @maxfieldmarolf; @wophilbert].\  \ In this work we would like to introduce a structurally different probe of the chaotic level statistics of quantum black holes. We propose to investigate the low-energy spectral properties of the eternal Hawking-Unruh radiation as detected by a linear Unruh-DeWitt detector. There are two main reasons why we believe this to be an interesting observable. 1. It has recently been advocated [@rw1; @rw2] that the same Euclidean wormholes which play a crucial role in explaining the late time behavior of correlators [@sss; @sss2; @Blommaert:2019hjr; @phil] are key to understanding unitary black hole evaporation from the bulk point of view. Ultimately we want to understand black hole evaporation by tracking what happens to all Hawking quanta emitted from the horizon during the evaporation process, which is more fine-grained bulk information than the early-late entanglement entropy considered in [@rw1; @rw2]. We expect to already see traces of the mechanism in the eternal Hawking-Unruh radiation. 2. Ultimately we want to have a pure bulk gravitational intuition of quantum gravity. In this sense our observable stands out as compared to for example late time boundary correlators. It is inherently a bulk observable. One further incentive to investigate bulk probes in quantum gravity is that in de Sitter or in flat space it is not so natural to formulate questions in the dual theory and so to make progress there we will need to strengthen our understanding of bulk observables in quantum gravity.[^3] This work is organized as follows. In **section \[sect:1\]** we argue on general grounds for a modification of the semiclassical formula for the Unruh-Hawking emission probability which accounts for the fact that black holes have chaotic level statistics. We then introduce the setup of JT gravity which we will use to gather concrete evidence for these ideas. In **section \[s:udw\]** we explicitly verify these expectations. In particular, we couple a massless scalar field to JT gravity and compute the response rate of an Unruh-DeWitt detector which couples linearly to this scalar field. We analyze the detector response in three layers of approximation. Firstly we use the semiclassical approximation. Secondly we discuss the effects of coupling to the Schwarzian reparameterization mode. Finally we include the effects of Euclidean wormholes which give rise to random matrix level statistics. We observe a depletion of the detector response rate at extremely low energies as testimony to the chaotic level statistics of quantum black holes. We generalize to bulk fermionic matter and to other detector couplings. In **section \[s:hb\]** we compute the spectral energy density in the Unruh heat bath. For the Schwarzian system (disk topology) this was investigated in [@Mertens:2019bvy]. Perhaps surprisingly this spectral energy density is not precisely identical to the detector response. We explain that this is due to ordering ambiguities that arise when promoting classical expressions to gravitational operators. Despite these subtle differences with the detector setup, we find a similar depletion at ultra low energies due to level repulsion. In **section \[s:concl\]** we comment on gravitational dressings, higher genus contributions to bulk correlators and evaporation. Certain more technical aspects of the story have been relegated to appendices. Unruh detectors and level repulsion {#s:exp} ----------------------------------- The main conceptual point in this work is that the semiclassical Planckian black body law for Hawking radiation does not take into account the chaotic level statistics of the quantum black holes which emit these quanta.\  \ We first briefly review the expected level statistics of black holes. For an arbitrary quantum mechanical system we denote the probability to find an energy level between $E_1$ and $E_1+ d E_1$ and a second energy level between $E_2$ and $E_2+d E_2$ as $$\rho(E_1,E_2)\, d E_1\,d E_2.$$ For a system of which the level spacings are Poisson distributed, the probability to find a level between $E_1$ and $E_1+d E_1$ is independent of the probability to find a second level between $E_2$ and $E_2+d E_2$ $$\rho(E_1)\rho(E_2)\, d E_1\, d E_2.$$ In other words different levels are uncorrelated. An example of a system with implicit Poisson level statistics is a particle in a very large box, typically used to derive the Planckian black body law. For chaotic quantum systems, the story is quite different. The local level statistics of a quantum chaotic system are described by random matrix theory [@haake; @mehta]. Level statistics in random matrix theory is universal: the multi-density correlators of essentially any chaotic quantum systems (without time-reversal symmetry) are those of the Gaussian unitary ensemble (GUE).[^4] The GUE two-level correlator is[^5] $$\rho(E_1,E_2) = \rho(E_1)\rho(E_2)-\rho(E_1)\rho(E_2)\,{\text{sinc}}^2\, \pi \rho(E_1)(E_1-E_2).\label{re1e2}$$ This leads to the normalized two-density correlator $$\frac{\rho(E_1,E_2)}{\rho(E_1)\rho(E_2)}=\quad \raisebox{-15mm}{\includegraphics[width=50mm]{rep.pdf}}\quad.\label{8}$$ This depletion around $E_1=E_2$ is the characteristic level repulsion of quantum chaotic systems. In particular for the GUE we have quadratic level repulsion meaning there is a quadratic zero in level correlators when any two eigenvalues approach.[^6]\  \ Now let’s think about a process where a system with such chaotic level statistics emits an energy quantum. In particular let’s consider the probability for a black hole with energy $E$ to emit a massless scalar particle with energy $\omega$ which we then detect in our detector. In $d$ spacetime dimensions this probability is proportional to the usual Planckian black body law[^7] $$R_{\text{BE}}(\omega) \sim \frac{\omega^{d-1}}{e^{\beta \omega}-1}.\label{be}$$ The Hawking modes in the Unruh heat bath originate from a decay process between two black hole energy levels. So every energy level $\omega$ of the Hawking radiation corresponds to the difference of two energy levels of the quantum chaotic black hole system $$\omega= E_1-E_2.\label{difference}$$ Therefore the level density for the Unruh modes in the heat bath is by definition sensitive and proportional to the two-level spectral density $\rho(E_1,E_2)$ of the underlying quantum black hole. For this reason we expect that the intrinsically chaotic level statistics of quantum black holes modifies the detection formula for the probability to detect a massless scalar Hawking particle with energy $\omega$ that has been emitted by a black hole with energy $E$ as $$\boxed{ R(\omega) \, \sim\, R_{\text{BE}}(\omega)(1-{\text{sinc}}^2 \pi \rho(E)\,\omega\,) \, \sim\, R_{\text{BE}}(\omega) \frac{\rho(E,E-\omega)}{\rho(E)\rho(E-\omega)}} \,. \label{expectationunruh}$$ This appropriately takes into account level repulsion.[^8] Of course this level repulsion is only visible at ultra low energies $\omega$. At such low energies the semiclassical Planckian black body low goes like $\sim \omega^{d-2}$. Level repulsion modifies this behavior at ultra low energies to $\sim \omega^d$. The effect is most clearly visible in 2d where the total detection probability looks like at very low energies $\omega$. Before proceeding, let us note that in quantum gravity there will be further modifications to the detector response for highly energetic Hawking modes $\omega\sim E$ for which we may no longer approximate the setup as a light probe particle travelling on a heavy black hole background $E\gg 1$. Related quantum gravitational effects will kick in for Planck size black holes $E\sim 1$. These effects are not governed by random matrix theory. The model {#s:setup} --------- In the remainder of this work we will gather concrete evidence in favor of via exact calculations in JT gravity. The action of JT gravity is [@Jackiw:1984je; @Teitelboim:1983ux][^9] $$\label{JTaction} S_{\text{JT}}[g,\Phi] = S_0\,\chi + \frac{1}{16\pi G}\int d x\, \sqrt{g}\,\Phi \left(R + 2\right) + \frac{1}{8\pi G}\int_\partial d\tau\, \sqrt{h}\,\Phi_{\partial} \left(K- 1 \right)\,.$$ Here the Euler character $\chi$ comes from the usual Einstein-Hilbert action in 2d. Its only effect is a weighting of different topologies. The extremal entropy is $S_0 = \Phi_0/4 G$.[^10] Integrating over the bulk values of the Lagrange multiplier dilaton field $\Phi$ localizes the gravitational path integral on hyperbolic Riemann surfaces $R+2=0$. Any hyperbolic Riemann surface can be built be gluing together different subregions of the [Poincaré ]{}upper half plane. In Lorentzian signature $$ds^2=\frac{d Z^2-d F^2}{Z^2} = -\frac{4d U d V}{(U-V)^2}\quad, \quad Z>0,\label{poinc}$$ where we introduced the Poincaré lightcone coordinates $U=F+Z$ and $V=F-Z$. There is an asymptotic boundary at $Z=0$ where we will have to impose interesting boundary conditions. Focusing on one such asymptotic boundary, there are several ways of discussing the reduction of the dynamics of this model into a purely boundary degree of freedom [@Almheiri:2014cka; @jensen; @malstanyang; @ads2]. Here we briefly review an intrinsically real-time approach [@ads2]. In order to treat this model within the holographic paradigm, we envision a dynamical boundary curve $(F(t),Z(t))$ as UV-cutoff as $Z\to 0$. This curve is specified by off-shell boundary conditions on the metric and dilaton field. The geometry is taken in Fefferman-Graham gauge, which entails partially fixing the diff-group near the boundary such that the leading part of the geometry is the Poincaré metric and the interesting dynamics is in the subleading pieces as $Z \to 0$. Physically this simply means one can only compare different spacetimes if they share the same asymptotics. This leads to the constraint $Z(t) = \varepsilon\, F'(t)$, determining the wiggly boundary in terms of a single reparameterization $F(t)$ mapping the [Poincaré ]{}time $F$ to the proper time $t$ of an observer following the boundary trajectory. This single function $F(t)$ generates the 1d conformal group (as the residual diff’s that preserve Fefferman-Graham gauge) and this is the usual endpoint in AdS/CFT. However, it can be shown that this system is special in the sense that also the dilaton field $\Phi$ blows up near the boundary $Z=0$. This means it has to be treated on the same footing as the metric, since once again one cannot compare spacetimes with different (dilaton) asymptotics. So, as a second constraint, we choose the boundary curve to satisfy $\Phi_\partial = a/\varepsilon$ in terms of a dimensionful quantity $a$ that determines the theory of interest.[^11] Combining the ingredients, one finds that the Lorentzian version of the action reduces to a Schwarzian derivative action [@Almheiri:2014cka; @jensen; @malstanyang; @ads2] $$\label{SSchL} S_L[F] = C\int dt \, \text{Sch}(F,t)\quad, \quad \text{Sch}(F,t) =\frac{F'''}{F'} - \frac{3}{2}\left(\frac{F''}{F'}\right)^2.$$ The coupling constant $C=a/8\pi G$ has units of length and controls the gravitational fluctuations of the wiggly boundary of the disk.[^12] Since path-integral computations are always performed in Euclidean signature, we Wick-rotate $t \to -i\tau$ and $F \to -i F$ to get the Euclidean Schwarzian action $$\label{SSch} S[F] =- C\int d\tau \, \text{Sch}(F,\tau)$$ With multiple boundaries one has to glue several such Schwarzian systems together [@sss2; @Blommaert:2018iqz]. Higher genus contributions can be included as explained in [@sss2]. In the lowest genus (disk) case it is more convenient to describe the model using the time reparametrization $f(\tau)$ defined as $$F(\tau) = \tan \frac{\pi}{\beta} f(\tau)\quad, \quad f(\tau+\beta) = f(\tau) + \beta\quad, \quad \dot{f}(\tau) \geq 0 \, .$$ The last two equations characterize $f(\tau)$ as reparametrizing a circle that is the boundary of the 2d Euclidean disk. The quantity $\beta$ is the boundary length and is interpretable as the inverse temperature. We will adhere to this notation of $F(\tau)$ and $f(\tau)$ in the remainder of this work.\  \ Using a myriad of techniques, JT gravity has been exactly solved for an entire class of boundary correlation functions. Gravitational contributions to correlation functions in JT gravity come in several flavors in terms of $G$. 1. There are perturbative $G$ corrections. These can be viewed as boundary graviton interactions and can be obtained via Schwarzian perturbation theory [@malstanyang; @Stanford:2017thb; @Qi:2019gny]. 2. There are nonperturbative $G$ corrections associated with an exact solution of Schwarzian correlation functions [@altland; @altland2; @schwarzian; @Mertens:2018fds; @paper3; @Blommaert:2018iqz; @kitaevsuh; @zhenbin; @Iliesiu:2019xuh; @suh]. 3. There are furthermore nonperturbative $G$ corrections associated with Euclidean wormhole contributions to the Euclidean path integral [@sss; @sss2; @Blommaert:2018iqz; @phil; @wophilbert]. These represent higher genus Riemann surfaces ending on the wiggly boundary. Indeed via we see that such contributions are perturbative in $e^{S_0}$ and hence nonperturbative in $G$. 4. Finally there are nonperturbative $e^{S_0}$ contributions which are hence doubly nonperturbative in $G$. In the gravitational language these are due to brane effects. JT gravity has a dual formulation as a double-scaled random matrix model [@sss2]. The doubly nonperturbative effects can be considered as hallmarks of this dual matrix integral description. One of the properties which makes JT gravity so interesting is that we have analytic control over all these types of corrections. The centerpiece formulas and take all such corrections into account. Detectors in the Unruh heat bath {#s:udw} ================================ In this sector we will consider a massless scalar field minimally coupled to JT gravity:[^13] $$\label{fieldact} S= \frac{1}{2} \int d x\,\sqrt{-g}\, g^{\mu\nu}\,\partial_\mu \phi\, \partial_\nu \phi.$$ In particular, we aim to probe the emission spectrum of massless scalar Hawking-Unruh particles by a 2d quantum black hole. To do so, we imagine an experiment where we probe the heat bath using a linear Unruh-DeWitt detector. We will isolate the effects of different types of gravitational interactions by working in three improving layers of approximation. Semiclassical analysis ---------------------- Let us first consider physics on the gravitational saddle, which is a black hole with inverse Hawking temperature $\beta$ $$ds^2 = \frac{\pi^2}{\beta^2}\frac{ dz^2-dt^2}{\sinh^2 \frac{2\pi}{\beta} z}\quad, \quad z>0\,.\label{thermal}$$ The semiclassical relation between the ADM mass $M$ of the black hole and the Hawking temperature is $\sqrt{M}=2\pi C/\beta$. We will henceforth set $C=1/2$. The asymptotic boundary is at $z=0$ and the semiclassical horizon is at $z=\infty$. This is just a conformal rescaling of flat space. The massless field $\phi$ is insensitive to this conformal rescaling of the metric and hence the solutions to the equations of motion of are left-and right-moving plane waves. Introducing lightcone coordinates $u=t+z$ and $v=t-z$ we have the mode expansion $$\label{modex} \phi(u,v) = \int_0^\infty \frac{d\omega}{\sqrt{4\pi \omega}}\,\left( a_\omega e^{-i\omega u } + a^\dagger_\omega e^{i\omega u } - a_\omega e^{-i\omega v } - a^\dagger_\omega e^{i\omega v }\right),$$ implementing Dirichlet boundary conditions on the boundary $z=0$ of the half plane . The modes are orthogonal with respect to the usual Klein-Gordon inner product, and the modes satisfy $$[a_\omega, a^{\dagger}_{\omega'}] = \delta(\omega-\omega').\label{24}$$ The Wightman two-point function in the thermal state of the CFT is [@spradlin] $${\left\langle \phi(u_1,v_1)\phi(u_2,v_2) \right\rangle}_{{\scriptscriptstyle \text{CFT}}} = - \frac{1}{4\pi} \ln \abs{\frac{\sinh \frac{\pi}{\beta}(u_1-u_2)\sinh \frac{\pi}{\beta}(v_1-v_2)}{\sinh \frac{\pi}{\beta}(u_1-v_2)\sinh \frac{\pi}{\beta}(v_1-u_2)}}.\label{25}$$ All matter correlators will be denoted by ${\left\langle \dots \right\rangle}_{{\scriptscriptstyle \text{CFT}}}$. This expression can be equivalently read as evaluated in the Poincaré vacuum state, defined by taking positive-frequency modes with respect to the [Poincaré ]{}time $F$. As is well known the correlator looks thermal in $t$ coordinates because of the thermal coordinate transformation $F = \tanh \frac{\pi}{\beta} t$ relating the two [@spradlin]. We will be interested in understanding the frequency content of this correlator, and in particular on what it has to say about the underlying black hole.\  \ The Unruh-DeWitt detector is a simple quantum mechanical detector model [@Unruh:1976db; @DeWitt:1980hx]. It linearly couples a quantum mechanical system with degree of freedom $\mu(t)$ to the scalar quantum field $\phi(u(t),v(t))$ via the local interaction Hamiltonian: $$H_\text{int}(t) = g \, \mu(t)\, \phi(u(t),v(t)).\label{26}$$ Typically one models the detector system to be a two-level system described by its “monopole moment” $\mu$. Here $g \ll 1$ is a tiny coupling and $(u(t),v(t))$ is the worldline of the detector. We assume the detector is initially in its ground state $\ket{0}$. We want to compute the probability $P(\omega)$ for the final state of the detector to be the energy eigenstate $\ket{\omega}$ in first order perturbation theory in the detector coupling $g$. Within perturbation theory, the Hilbert space factorizes as $\mathcal{H}_{\text{det}} \otimes \mathcal{H}_{\text{matter}}$. Since the detector is ignorant of the final state $\ket{\psi}$ of the matter sector, one finds: $$\label{27} P(\omega) =\sum_{\psi} \left|\bra{0,M} -i\int_{-\infty}^{+\infty} d t H_\text{int}(t)\ket{\omega,\psi}\right|^2.$$ Here $\ket{M}$ denotes the thermal state of the matter sector. Summing over $\psi$, one finds the response rate $R(\omega)$ which is defined as the probability per unit time to see the detector transition:[^14] $$\begin{aligned} \label{29} R(\omega)=& \, g^2\,\abs{\bra{\omega}\mu(0)\ket{0}}^2\\&\lim_{{{\small \text{T}}} \to +\infty}\frac{1}{{\small \text{T}}} \int_{-{\small \text{T}}}^{+{\small \text{T}}}dt_1\int_{-{\small \text{T}}}^{+{\small \text{T}}}dt_2\, e^{-i\omega (t_1-t_2)}{\left\langle \phi(u(t_1),v(t_1))\phi(u(t_2),v(t_2)) \right\rangle}_{{\scriptscriptstyle \text{CFT}}}.\nonumber\end{aligned}$$ This quantity represents the probability rate for the detector to get excited to energy level $\omega$. By energy conservation the matter state gets depleted by a similar energy $\omega$. In the vacuum associated to the time coordinate $t$ the result would be zero but this is not so in the Poincaré vacuum. This is the essence of the Unruh effect. The analogous situation in flat space is by interpreting the black hole time coordinate $t$ as Rindler (or Schwarzschild) time, and the Poincaré time $T$ as Minkowski time. For a thorough recent review see [@Crispino:2007eb]. In the [Poincaré ]{}vacuum one finds the semi-classical thermal Unruh population by computing the integral on the second line of using and assuming a stationary detector worldline. The answer is $$\label{210} R(\omega) = g^2 \abs{\bra{\omega}\mu(0)\ket{0}}^2 \, 2\, \frac{\sin^2 \omega z}{\omega^2}\, \frac{\omega}{e^{\beta \omega} -1}\,.$$ One indeed recognizes the Planckian black body law in two dimensions . Coupling to Schwarzian reparameterizations ------------------------------------------ Our goal for this subsection and the following is to compute the Fourier transform of the bulk two-point function on the second line of in two different levels of approximation. In this subsection we will ignore Euclidean wormhole (or higher genus) contributions to the Euclidean JT gravity path integral. This means we incorporate the gravitational corrections of only items 1 and 2 of the list in section \[s:setup\].\  \ Within quantum gravity, physical bulk locations and bulk observables must be defined in a diff-invariant manner. To do so in a holographic context we are led to define a point in the bulk by anchoring the bulk point to the asymptotic boundary [@Donnelly:2015hta; @gid3; @Almheiri:2017fbd; @ref3; @Lewkowycz:2016ukf; @ref4; @ref5; @ref6; @Chen:2017dnl; @Chen:2018qzm; @Engelhardt:2016wgb]. One particularly natural way to do this in this context is by geodesic localizing using lightrays [@Engelhardt:2016wgb; @Blommaert:2019hjr; @Mertens:2019bvy]. In JT gravity the boundary is one-dimensional. Therefore the physical coordinates used to define bulk points are boundary time coordinates. In particular we need two such time coordinates $t_1$ and $t_2$. We associate these to the lightcone coordinates $u$ and $v$ of a point in the bulk. Here $v$ is the physical boundary time at which an observer sends a signal to a given bulk point and $u$ is the physical boundary time at which the observer receives back the signal after reflecting off some fictitious mirror. The boundary curve $(F(t),Z(t))$ to which we anchor a bulk point is described by a single function $F(t)$, where $Z$ is determined in terms of $F$ by the boundary conditions. The actual wiggling of the boundary as explained above , is negligible for $\varepsilon\ll 1$. The field $F(t)$ represents the map between [Poincaré ]{}coordinates and the physical boundary time coordinate $t$ [@Almheiri:2014cka; @jensen; @malstanyang; @ads2]. Consequently in terms of the two boundary times $u$ and $v$, the location of the bulk point in [Poincaré ]{}coordinates is defined as $U= F(u)$ and $V =F(v)$. For a more detailed explanation of this construction, see [@Blommaert:2019hjr; @Mertens:2019bvy; @wopjordan]. The bulk metric constructed in this fashion is $$ds^2=-\frac{F'(u)F'(v)}{(F(u)-F(v))^2}\,du\,dv.\label{189}$$ Following this same logic, we are led to define massless scalar bulk observables $\Phi$ which implicitly depend on the Schwarzian reparameterization $f$ as: $$\label{dress} \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u,v] = \phi(f(u),f(v)).$$ This is just a regular massless scalar field but the location of the insertion of this operator in [Poincaré ]{}coordinates $(Z,F)$ depends on the details of the wiggly boundary $F(t)$ $$\Phi[f_1{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u,v]= \raisebox{-17mm}{\includegraphics[width=21mm]{points1.pdf}}\quad,\quad \Phi[f_2{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u,v]= \raisebox{-17mm}{\includegraphics[width=22.7mm]{points2.pdf}}\quad.$$ This implicit dependence on $F(t)$ of bulk operators couples bulk matter to the Schwarzian reparameterization mode. We note that the definition of bulk operators in quantum gravity corresponds in the language of [@Donnelly:2015hta; @gid1; @gid2; @gid3; @gid4; @gid5; @gid6; @gid7] to specifying a particular gravitational dressing of a bare bulk matter operator such that the dressed operator is diff-invariant.\  \ In every fixed metric , the two-point function of two operators of the type is then by definition of just the reparameterization of $$\label{twoqft} {\left\langle \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1] \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}} = - \frac{1}{4\pi}\left|\ln \frac{\sinh \frac{\pi}{\beta}(f(u_1)-f(u_2))\sinh \frac{\pi}{\beta}(f(v_1)-f(v_2))}{\sinh \frac{\pi}{\beta}(f(u_1)-f(v_2))\sinh \frac{\pi}{\beta}(f(v_1)-f(u_2))}\right|.$$ In this same way we define the entire trajectory of the Unruh-DeWitt detector:[^15] $$\includegraphics[width=0.2\textwidth]{bulkframeUDW.pdf}$$ This shows the constant $z$ worldlines of the Unruh-DeWitt detector for two different clock ticking patterns $F(t)$ (blue and red) at the same radial coordinate $z$. Close to the horizon and semi-classically, these become constant accelerated worldlines. Close to the boundary at $z=\epsilon$, this becomes the wiggly boundary curve itself. The interaction Hamiltonian now features the diff-invariant dressed field $\Phi$ $$H_\text{int}(t) = g\, \mu(t)\,\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u(t),v(t)].\label{216}$$ Including gravitational degrees of freedom and perturbatively in $g$ the Hilbert space may be considered to factorize as $\mathcal{H}_\text{det}\otimes \mathcal{H}_\text{matter+gravity}$. The second factor is the whole coupled system of matter and gravity. Computing the probability for the detector to evolve from an initial state $\ket{0}$ to a final state $\ket{\omega}$ one finds an identical formula but with the interaction replaced by and with $\psi$ summed over $\mathcal{H}_\text{matter+gravity}$. Doing the sum over the final states $\psi$ results in $$\begin{aligned} \bra{\mM}\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1]\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2]\ket{\mM} =\int_{\text{micro M}} [{\mathcal{D}}f]\, {\left\langle \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1] \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}}\,e^{-S[f]}.\label{215}\end{aligned}$$ This implements a matter-coupled quantum gravity path integral in the microcanonical ensemble of fixed energy $M$, defined by inverse Laplace transforming the canonical ensemble path integral. Here the Schwarzian action is .\  \ We can now proceed with the actual computation, for which we use a technical trick [@Blommaert:2019hjr]: $${\left\langle \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1] \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}}=\int_{v_1}^{u_1} d t_1\int_{v_2}^{u_2}dt_2\,{\left\langle {\mathcal{O}}[f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}t_1]{\mathcal{O}}[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}t_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}}.\label{hkll}$$ Here the matrix element on the right hand side is a thermal boundary two-point function of a massless scalar: $${\left\langle {\mathcal{O}}[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}t_1]{\mathcal{O}}[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}t_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}}=-\frac{1}{4\pi}\,\frac{ f'(t_1)f'(t_2)}{\frac{\beta^2}{\pi^2}\sinh^2 \frac{\pi}{\beta}(f(t_1)-f(t_2))}.\label{218}$$ We notice that is essentially implementing the HKLL bulk reconstruction in each of the metrics [@hkll1; @hkll2; @kll; @kl; @Lowe:2008ra]. This reverse-engineered version of bulk reconstruction is pivotal since the Schwarzian path integral in over the right hand side of can be easily computed [@Blommaert:2019hjr], as the disk boundary-to-boundary propagator in JT gravity: $$\int [{\mathcal{D}}f]\, {\left\langle {\mathcal{O}}[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}t_1]{\mathcal{O}}[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}t_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}}\, e^{-S[f]}\,=\quad\raisebox{-10mm}{\includegraphics[width=45mm]{1lbiloc.pdf}}\quad.\label{40}$$ The answer is [@altland; @altland2; @schwarzian; @Mertens:2018fds; @paper3; @Blommaert:2018iqz; @kitaevsuh; @zhenbin; @Iliesiu:2019xuh]:[^16] $$\frac{1}{Z(\beta)}\int_0^\infty d M\,e^{-\beta M}\int_{0}^{+\infty} dE \, e^{-i(t_1-t_2)(E-M)}\, \rho_0(M)\rho_0(E)\,\rvert {\mathcal{O}}_{M E}^1\rvert^2. \label{41}$$ Here we introduced the notation:[^17] $$\label{221} \rho_0(E)=\frac{e^{S_0}}{2\pi^2}\sinh 2\pi \sqrt{E}\quad,\quad \rvert{\mathcal{O}}_{E_1E_2}^\ell\rvert^2 = e^{-S_0}\frac{\Gamma(\ell\pm i\sqrt{E_1}\pm i\sqrt{E_2})}{\Gamma(2\ell)}\, .$$ Within the microcanonical ensemble of fixed energy $M$, the boundary-to-boundary correlator becomes: $$\int_{0}^{+\infty} d E \, e^{-i(t_1-t_2)(E-M)}\, \rho_0(E)\,\rvert {\mathcal{O}}_{M E}^1\rvert^2 \, .$$ The response rate for a detector at $z_1=z_2=z$ is now readily found by performing the elementary integrations in and . One obtains: $$\begin{aligned} \label{response} R(\omega) &= g^2\,\abs{\bra{\omega}\mu(0)\ket{0}}^2\, 2 \, \frac{\sin^2 \omega z}{\omega^2}\, \rho_0(M-\omega)\, \rvert {\mathcal{O}}_{M M-\omega}^1\rvert^2 \,.\end{aligned}$$ Notice that by it is independent of the ground-state degeneracy $e^{S_0}$. The “greybody” interference factor $2 \sin^2 \omega z/\omega^2$ is determined in part by choosing Dirichlet boundary conditions for the massless scalar field in .[^18] For other boundary conditions on the scalar field, the interference pattern changes, see appendix \[app:obc\]. In the semiclassical regime $M \gg 1$ and $\omega\ll M$, reproduces the classical answer . For more generic values of $\omega$ however there are gravitational backreaction effects. The most prominent such backreaction effect is that the response function abruptly stops at $\omega = M$ with a square root edge via .\  \ In order to better understand from a physical point of view why vanishes at $\omega=M$, it is convenient to go back to the mode expansions . We can use it to define raising and lowering operators in the matter Hilbert space: $$\begin{aligned} \label{creaan} a_\omega &= \frac{i}{\sqrt{\pi \omega}} \int_{-\infty}^{+\infty} du \,\partial_u \phi(u) \, e^{i\omega u}, \qquad a^{\dagger}_\omega = -\frac{i}{\sqrt{\pi \omega}} \int_{-\infty}^{+\infty} du \, \partial_u \phi(u) \, e^{-i\omega u} \, .\end{aligned}$$ Quite analogously, we could define raising and lowering operators of gravitationally dressed matter fields in the Hilbert space of the coupled system of matter and gravity $$\begin{aligned} \label{creabn} A_\omega[f] &= \frac{i}{\sqrt{\pi \omega}} \int_{-\infty}^{+\infty} du \,\partial_u \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u] \, e^{i\omega u}, \qquad A^{\dagger}_\omega[f] = -\frac{i}{\sqrt{\pi \omega}} \int_{-\infty}^{+\infty} du \, \partial_u \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u] \, e^{-i\omega u}\, .\end{aligned}$$ Working with these dressed (or diff-invariant) raising and lowering operators, we find that the response rate is proportional to a number operator expectation value[^19] $$\begin{aligned} \label{num1} \bra{\mM}A^\dagger_\omega[f] A_\omega[f] \ket{\mM} &= \frac{V}{2\pi\omega} \,\rho_0(M-\omega)\, \rvert\mathcal{O}_{M M-\omega}^1\rvert^2\,.\end{aligned}$$ This formula explains the step function in a quite natural manner: the dressed operator $A_\omega$ extracts an energy $\omega$ from the gravity system. Of course it is impossible to extract more energy from this system than the finite energy $M$ which it had to begin with and so we have $$A_\omega[f] \ket{\mM}= 0\quad,\quad \omega>M.\label{229}$$ Therefore and should be expected to vanish for $\omega >M$. Using similar techniques, one could compute more involved matrix elements. For example $$\bra{\mM}A^\dagger_{\omega_1}[f]\dots A^\dagger_{\omega_n}[f]A_{\omega_n}[f]\dots A_{\omega_1}[f]\ket{\mM}.$$ One finds, in accordance with the fact that operators such as $A_\omega$ deplete the system of an energy $\omega$, that this amplitude vanishes if $\omega_1+\dots \omega_n>M$. Level repulsion {#s:levrep} --------------- We would now like to include Euclidean wormhole contributions to the JT gravity path integral which computes the massless bulk two point function. This includes all 4 items of the list of gravitational corrections in section \[s:setup\].\  \ It is not a priori obvious how to dress the massless scalar bulk two-point function, whose radar construction was inherently Lorentzian, with higher genus contributions to the Euclidean path integral. We here propose a very natural way of doing so. We start with the disk (genus zero) contribution first. One may then use the bulk reconstruction formula to write the bulk matter two-point function in terms of a Euclidean disk JT gravity path integral with a boundary-to-boundary matter propagator $$\bra{\mM}\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1]\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2]\ket{\mM}\,=\int_{v_1}^{u_1}dt_1\int_{v_2}^{u_2} dt_2 \quad\raisebox{-10mm}{\includegraphics[width=45mm]{1lbiloc.pdf}}\quad.\label{231}$$ Including Euclidean wormhole connections for the boundary two-point function on the right hand side has recently been understood [@phil; @Blommaert:2019hjr; @wophilbert]. One just sums over all higher genus Riemann surfaces which end on the union of the boundary circle and the boundary-to-boundary bilocal line. We now define bulk correlators by applying the bulk reconstruction formula to these higher-topology boundary correlators. So our definition of bulk operators is really a “bulk reconstruction first” approach. This includes for example a contribution of the type $$\bra{\mM}\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1]\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2]\ket{\mM}\,\supset\int_{v_1}^{u_1}dt_1\int_{v_1}^{u_2} dt_2 \quad\raisebox{-10mm}{\includegraphics[width=45mm]{1lbilochandle.pdf}}\quad.\label{232}$$ This definition ends up reproducing our generic intuition which we view as an argument that this definition makes sense. Similarly in [@wophilbert] we will find that this definition ends up reproducing other intuitive expectations about bulk observables. As it turns out, it is quite feasible to sum over all such amplitudes via the definition of JT gravity as a double-scaled matrix integral [@sss2]. In the end, the effect of including such contributions is quite elegant. In terms of the formulas, one ends up effectively replacing $\rho_0(M)\rho_0(E)$ in the boundary two-point function by $\rho(M,E)$ which is the universal answer from random matrix theory[^20] $$\frac{1}{Z(\beta)}\int_{\mathcal{C}}d M\,e^{-\beta M}\int_{\mathcal{C}}dE \, e^{-i(t_1-t_2)(E-M)}\, \rho(M,E)\,\rvert {\mathcal{O}}_{M E}^1\rvert^2. \label{233}$$ Here we have $$\rho(M,E) = \rho(M)\rho(E)-\rho(M)\rho(E)\,{\text{sinc}}^2\, \pi \rho(M)(M-E) + \rho(M)\delta(M-E). \label{49}$$ This formula holds if both $M$ and $E$ are far enough from the spectral edge $E\gg e^{-2S_0/3}$. Closer to the spectral edge, we can resort to similarly universal formulas for the two-level spectral density in the Airy model [@sss2; @paper6]. The first term corresponds geometrically to the disconnected contribution of adding higher topology to each of the two sides of the bilocal line in . The second term is due to Riemann surfaces connecting both sides of the bilocal such as the annulus in . Each of these terms furthermore includes nonperturbative contributions in $e^{-S_0}$ due to brane effects [@sss2; @phil; @paper6]. These are oscillatory and hence not necessarily small. The final term represents a contact term that may or may not have a geometric interpretation.[^21] We can now immediately write down the analogue of which takes into account these Euclidean wormholes via the substitution $$\boxed{R(\omega) = g^2\,\abs{\bra{\omega}\mu(0)\ket{0}}^2\,2\,\frac{\sin ^2\omega z}{\omega^2}\, \frac{\rho(M,M-\omega)}{\rho(M)}\, \rvert {\mathcal{O}}_{M M-\omega}^1\rvert^2.}\label{51}$$ It is straightforward to similarly write down the modifications to the expectation values of the dressed number operator . Notice that now there is no step function. This is because in the matrix integral description there is a tiny but nonzero probability for eigenvalues to be in the forbidden region $E<0$ of the energy contour ${\mathcal{C}}$ [@sss2]. We note that that the prefactor $g^2\,\abs{\bra{\omega}\mu(0)\ket{0}}^2$ is intrinsic to the detector, which we (the observer) can determine within the free detector theory. Dividing out this known prefactor, and computing in the probe approximation $M\gg 1$ and $\omega\ll M$ we find $$\frac{R(\omega)}{2 \, \sin^2 \omega z/\omega^2} = \frac{\omega}{e^{\beta\omega}-1} (1-{\text{sinc}}^2 \pi \rho(M)\,\omega\,) \, . \label{237}$$ This explicitly confirms our general expectation about the effects of level repulsion on the detection rate, via a bulk JT gravity calculation. Schematically the detector finds the following result $$R(\omega) = 2 \, \frac{\sin^2 \omega z}{\omega^2}\quad \raisebox{-15mm}{\includegraphics[width=65mm]{depl.pdf}}\quad\label{234}$$ The red curve denotes the semiclassical Planckian black body law with initial linear decay. At extremely low energies with $\omega$ of order $e^{-S_0}$ we explicitly see the depletion of the Hawking-Unruh spectrum due to level repulsion in quantum gravity.\  \ To get a better grip on the physics we are probing we can briefly consider the bulk gravitational dual to one single quantum chaotic system with Hamiltonian $H$. In JT gravity, we can effectively achieve this by considering a new definition of JT gravity for which the path integral itself includes a bunch of additional branes [@sss2; @paper6; @maxfieldmarolf; @wophilbert]. The data of these branes are one-to-one with the spectral data of the dual quantum mechanical system. For this setup in particular, we want to consider a version of JT gravity that includes branes which fix the eigenvalues $\lambda_1\dots \lambda_L$ of the Hamiltonians [@paper6]. Essentially we replace in and $$\rho(E,M)=\sum_{i=1}^L \delta(E-\lambda_i)\sum_{j=1}^L \delta(M-\lambda_j).\label{239}$$ We imagine probing in a state of the coupled matter and gravity system with energy $M$ identical to one of the eigenvalues $\lambda$. We find via and $$R(\omega)= 2\,\frac{\sin^2 \omega z}{\omega^2} \, \sum_{i=1}^L \delta(\omega-\lambda+\lambda_i)\, \rvert {\mathcal{O}}_{\lambda \lambda_i}^1\rvert^2.$$ In the probe approximation $\lambda\gg 1$ and $\omega\ll \lambda$ this approximates to[^22] $$R(\omega)_{\lambda} = 2\,\frac{\sin^2 \omega z}{\omega^2}\, \frac{\omega}{e^{\beta \omega}-1}\sum_{i=1}^L \frac{1}{\rho(\lambda_i)}\delta(\omega-\lambda+\lambda_i) .\label{55}$$ This makes explicit formula which states that the energy levels $\omega$ of the Unruh-Hawking modes must match to energy differences of the quantum chaotic black hole system: the gravitationally dressed raising and lowering operators generate level transitions within the quantum gravitational system $$\raisebox{-15mm}{\includegraphics[width=48mm]{trans.pdf}}\quad\label{242}$$ We remove (orange) an energy difference $\omega_{ij}$ from an initial state (red) with energy $\lambda_i$ and are left with a final state (blue) with energy $\lambda_j$.\  \ The result carries an interesting interpretation. It is commonly assumed that the detailed information about the microstate of a gravitational system is somehow located behind the horizon. In however we have an explicit experiment which can be performed by an observer hovering outside of the semiclassical horizon. From the resulting set of delta spikes as observed by his detector, he is in principle able to determine the energy spectrum of all the gravitational microstates. Here we take the details of one gravitational microstate to mean knowledge of all the levels $\lambda_1\dots \lambda_L$. Fermionic matter ---------------- The result is expected to hold quite universally. We may provide evidence for this by testing it in more general situations. Therefore in this section we consider an Unruh-DeWitt detector coupled to massless fermionic bulk matter in JT gravity. In the next section \[s:proper\] we consider other detector couplings. We note that massless bulk fermions coupled to JT gravity are expected to play quite an important role in the elusive 2d gravity theory dual to the SYK model. Furthermore a large number of such massless bulk fermionic fields drastically enhance the evaporation rate for 4d magnetically charged black holes (which have a near horizon description as JT gravity) [@Maldacena:2018gjk; @Maldacena:2020skw].\  \ We will consider as detector interaction the simplest possible coupling of a bosonic detector $\tilde{\mu}(t)$ to a bulk Dirac fermion $\psi(u,v)$ [@Takagi:1986tf] $$\begin{aligned} H_\text{int} &= g \, \tilde{\mu}(t) \, (\bar{\psi} \psi)(u(t),v(t)).\label{ferco}\end{aligned}$$ See [@Hummer:2015xaa] for a comprehensive study of different Unruh-DeWitt detector types and their effects and [@Gray:2018ifq; @Louko:2016ptn] for recent studies. The normal ordering defined by $(\bar{\psi} \psi) = \bar{\psi}\psi - \left\langle \bar{\psi}\psi\right\rangle_{{\scriptscriptstyle \text{CFT}}}$ is required in order to sensibly define this detector coupling. The detector variable is chosen as $$\label{rel} \tilde{\mu}(t) = \frac{L_{\text{det}}}{\Omega(u(t),v(t))} \,\mu(t) \, .$$ We recognize the same detector variable $\mu(t)$ from the bosonic detector . Furthermore we introduced the Weyl factor $\Omega(u(t),v(t))$ of the metric $$ds^2=-\frac{1}{\Omega(u,v)^{2}}\,du\, dv = -\frac{F'(u)F'(v)}{(F(u)-F(v))^2} du\, dv \, .$$ The relation can be motivated by dimensional analysis. The coupling $\mu(t)$ in the bosonic detector has units of inverse length whereas $\tilde{\mu}(t)$ is dimensionless. The Weyl factor $\Omega(u(t),v(t))$ transforms between the detector length scale $L_\text{det}$ and the local length scale. We will suppress the detector length scale $L_{\text{det}}$. It is however not difficult to consider a detector with coupling that does not include this Weyl factor. We comment on this in section \[s:proper\]. Notice that because of the coupling to a composite operator in , we we are not probing for a fermionic emission spectrum, but instead are measuring the emission of a fermion-antifermion pair. This jives with the interpretation that we are probing with the Unruh-DeWitt detector the probability of level transitions within an underlying bosonic black hole system, for which level transitions are only possible upon emission of bosonic quanta.\  \ Let us develop this example in more detail. The 2d massless curved spacetime Dirac equation has the following Weyl rescaling property: if $\psi$ is a solution in the metric $g_{\mu\nu}$, then $\Omega^{1/2}\psi$ is a solution in the metric $\Omega^{-2}g_{\mu\nu}$. We can use this to write the mode expansion of the Dirac field in AdS$_2$ in [Poincaré ]{}coordinates ($U,V$) as:[^23] $$\begin{aligned} \label{modefermi} \frac{\psi(U,V)}{(U-V)^\frac{1}{2}}&=\psi(U)+\psi(V)\\&= \,\,\frac{1}{\sqrt{4\pi}}\,\left(\begin{array}{c} 1 \\ i \end{array}\right)\,\sum_{\omega>0} \left(e^{- i \omega U} a_{\omega} +e^{ i \omega U} b_{\omega}^{\dagger}\right) +\,\frac{1}{\sqrt{4\pi}}\,\left(\begin{array}{c} i \\ 1 \end{array}\right)\,\sum_{\omega>0} \left(e^{- i \omega V} a_{\omega} +e^{ i \omega V} b_{\omega}^{\dagger}\right) \, .\nonumber\end{aligned}$$ We have implicitly adopted Dirichlet boundary conditions here.[^24] Using this mode expansion, it is easy to obtain the Wightman bulk two point function in the [Poincaré ]{}vacuum: $$\frac{{\left\langle \psi_\alpha(U_1,V_1)\bar{\psi}_\beta(U_2,V_2) \right\rangle}_{{\scriptscriptstyle \text{CFT}}}}{ (U_1-V_1)^\frac{1}{2}\,(U_2-V_2)^\frac{1}{2}} = S_{\alpha\beta}(U_1,V_1,U_2,V_2) \, ,$$ where $$S(U_1,V_1,U_2,V_2)=\frac{1}{2}\left[ \begin{array}{cc} \frac{1}{U_1-U_2} - \frac{1}{U_1-V_2} - \frac{1}{V_1-V_2} + \frac{1}{V_1-U_2} & \frac{i}{U_1-U_2} + \frac{i}{U_1-V_2} + \frac{i}{V_1-V_2} + \frac{i}{V_1-U_2} \\ \frac{i}{U_1-U_2} - \frac{i}{U_1-V_2} + \frac{i}{V_1-V_2} - \frac{i}{V_1-U_2} & -\frac{1}{U_1-U_2} - \frac{1}{U_1-V_2} + \frac{1}{V_1-V_2} + \frac{1}{V_1-U_2} \end{array}\right] \nonumber \, .$$ In terms of the detector we are led to compute the bulk two-point function of fermion-antifermion pairs. By taking Wick contractions, one finds $$\begin{aligned} \label{contr} \frac{\left\langle (\bar{\psi} \psi) (U_1,V_1) (\bar{\psi} \psi) (U_2,V_2)\right\rangle_{{\scriptscriptstyle \text{CFT}}}}{(U_1-V_1)\,(U_2-V_2)}&= \sum_{\alpha\,\beta}\contraction{}{\overline{\psi}_\alpha}{\psi_\alpha \overline{\psi}_\beta}{\psi_\beta} \contraction{\overline{\psi}_\alpha}{\psi_\alpha}{}{\overline{\psi}_\beta} \overline{\psi}_\alpha \psi_\alpha \overline{\psi}_\beta \psi_\beta = - \Tr S^2(U_1,V_1,U_2,V_2)\\&= -\frac{1}{(U_1-V_2)^2} - \frac{1}{(V_1-U_2)^2} + \frac{2}{(U_1-U_2)(V_1-V_2)}\nonumber \,.\end{aligned}$$ Notice that by definition of normal-ordering, there are no contractions to be considered within each composite operator $(\bar{\psi}\psi)(U,V)$.\  \ As in for the bosonic case, we can couple this observable to the Schwarzian by applying a coordinate reparameterization and then computing the path integral. Before doing so, let us note that the Weyl factor in the detector couplings cancels with the Weyl factors on the left hand side of in the detector transition rate . So from hereon let us drop all such factors. In reparameterized bulk metrics and ignoring the Weyl factors, the fermion pair two-point function becomes: $$\begin{aligned} \label{contr2} \nonumber &{\left\langle (\bar{\Psi} \Psi) [f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1,v_1] (\bar{\Psi} \Psi) [f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2,v_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}} \\&= -\frac{F'(u_1)F'(v_2)}{(F(u_1)-F(v_2))^2} - \frac{F'(v_1)F'(u_2)}{(F(v_1)-F(u_2))^2} + 2\, \frac{F'(u_1)F'(u_2)}{(F(u_1)-F(u_2)}\,\frac{F'(v_1)F'(v_2)}{(F(v_1)-F(v_2)}\,,\end{aligned}$$ in terms of the gravitationally dressed field $\Psi$. The conformal scaling factors $F'$ are explained because the holomorphic and antiholomorphic components $\psi(u)$ and $\psi(v)$ in are $\ell=1/2$ conformal primaries. Following and normalizing by the operator intrinsic prefactor on the first line of , one now computes the response rate as: $$\begin{aligned} R(\omega)=&\lim_{{\small \text{T}}\to \infty}\frac{1}{{\small \text{T}}}\int_{-{\small \text{T}}}^{+{\small \text{T}}}dt_1\int_{-{\small \text{T}}}^{+{\small \text{T}}}dt_2\,e^{-i\omega(t_1-t_2)}\nonumber\\&\qquad\qquad\qquad\qquad\bra{\mM}(\bar{\Psi} \Psi) [f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_1(t_1),v_1(t_2)] (\bar{\Psi} \Psi) [f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u_2(t_1),v_2(t_2)]\ket{\mM}\, .\label{254}\end{aligned}$$ As in the bosonic case we will consider a detector following a trajectory at fixed $z$. In implementing the Schwarzian path integral of there are a few subtleties related with operator ordering ambiguities which we must first address. 1. All three terms in correspond to the product of two Schwarzian $\ell=1/2$ bilocals. In our case the quantum mechanical operator ordering is fixed by the nested Wick contractions in , leading to three nested Schwarzian four-point functions. 2. The first two terms of correspond to two Schwarzian bilocals with the same start and end points. The Schwarzian correlator can be simplified using the following property: $$\quad\raisebox{-10mm}{\includegraphics[width=45mm]{l1.pdf}}\quad=\quad\raisebox{-10mm}{\includegraphics[width=45mm]{l1plus.pdf}}\quad.\label{256}$$ Both these JT gravity amplitudes can be immediately evaluated using the result of [@schwarzian; @Mertens:2018fds; @paper3; @Blommaert:2018iqz]. Labeling the energies as $E_1$, $E$ and $E_2$ from left to right of the diagram , we have the result:[^25] $$\begin{aligned} \int_0^\infty d E\, \rho_0(E)\, \rvert{\mathcal{O}}^{\ell_1}_{E_1 E}\rvert^2\,\rvert{\mathcal{O}}^{\ell_2}_{E E_1}\rvert^2 = \rvert{\mathcal{O}}^{\ell_1+\ell_2}_{E_1 E_2}\rvert^2\, .\label{257} \end{aligned}$$ This means we can use the classical identity $$\left(\frac{F'_1 F'_2}{(F_1-F_2)^2} \right)^{\ell_1 } \left(\frac{F'_1 F'_2}{(F_1-F_2)^2} \right)^{\ell_2 } = \left(\frac{F'_1 F'_2}{(F_1-F_2)^2} \right)^{\ell_1+\ell_2 } \, ,$$ also at the quantum level. We also have the more general identity: $$\quad\raisebox{-10mm}{\includegraphics[width=45mm]{l1it.pdf}}\quad=\quad\raisebox{-10mm}{\includegraphics[width=45mm]{l1plus.pdf}}\quad.\label{259}$$ This property also reduces the Schwarzian path integral of the last term in (for which $c=2i z$) to a single bilocal computation. Proceeding by the calculation of we find that this boils down to computing three Schwarzian diagrams of the same type. Their prefactors combine as an interference factor as $ - \frac{1}{2}e^{i\omega 2z} - \frac{1}{2}e^{-i\omega 2z} + 1 = 2 \sin^2 \omega z$, leading to the result $$R(\omega) = 2\,\sin^2 \omega z\, \rho_0(M-\omega)\, \rvert{\mathcal{O}}_{M M-\omega}^1 \rvert^2\,.\label{260}$$ Finally, we should include Euclidean wormholes to this computation. The calculation is essentially identical to the one which led to . We note that does not technically hold when including such Euclidean wormholes because we could have Euclidean wormholes connecting the middle region with the two other regions. Such corrections are important in the intermediate $E$ integral whenever $E \approx M$ and or $E\approx M-\omega$ where the contribution to the $E$ integral is pushed to zero by quadratic level repulsion. This region has energy width $\sim e^{-S_0}$ and order $1$ height. Doing the $E$ integral results in an order $e^{-S_0}$ correction to the response rate. This is negligible. In other words we may safely still use even when including Euclidean wormhole contributions to both diagrams. The final result is hence essentially identical to the bosonic answer and provides a second example of our generic expectation in the probe approximation $M\gg 1$ and $\omega \ll M$. The difference sits only in the interference factor which differs by an overall $\omega^2$. This comes purely from dimensional reasons since there is also a length scale $L_{\text{det}}$ in the coupling . The relative prefactor between and is the dimensionless combination $L_{\text{det}}^2 \, \omega^2$. More general detector couplings {#s:proper} ------------------------------- The detector couplings $\mu(t)$ and $\tilde{\mu}(t)$ we have defined previously, transform as scalar densities under coordinate transformations, and correspond to time measurements on the boundary clock. It is straightforward to consider coupling to the proper bulk time of the worldline, by changing the interaction terms into $$\begin{aligned} \label{propcou} S_\text{int} &= g\int d\tau_p \, \mu(\tau_p)\, \phi(u(\tau_p),v(\tau_p)) = g\int dt \, \Omega(u(t),v(t))^{-1} \, \mu(t)\, \phi(u(t),v(t)) \, \\ S_\text{int} &= g\int d\tau_p \, \tilde{\mu}(\tau_p) \, (\bar{\psi} \psi)(u(\tau_p),v(\tau_p)) = g \int dt \, \Omega(u(t),v(t))^{-1} \, \tilde{\mu}(t) \, (\bar{\psi} \psi)(u(t),v(t)) \, . \nonumber \end{aligned}$$ Here $dt \, \Omega(u(t),v(t))^{-1} = d\tau_p$ is the proper time along the bulk worldline. At the quantum gravity level, the above coupling is replaced by a Hermitian coupling including the dressed fields $$\begin{aligned} S_\text{int} = \frac{g}{2}\int dt \, \mu(t) \Bigl( \Omega(u(t),v(t))^{-1} \,\Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u(t),v(t))] + \Phi[f{\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u(t),v(t))]\Omega(u(t),v(t))^{-1} \Bigr) \nonumber \, .\end{aligned}$$ Notably also $\Omega^{-1}$ depends on the gravity variable $f$. An analogous formula holds for the fermion detector. The explicit Weyl factors $\Omega^{-1}$ in this expression yield additional Schwarzian bilocal lines with length $2z$ between their endpoints. Diagrammatically we have: $$\label{weyl} \frac{1}{2} \quad\raisebox{-10mm}{\includegraphics[width=45mm]{ProperFig.pdf}}\quad+ \frac{1}{2}\quad\raisebox{-10mm}{\includegraphics[width=45mm]{ProperFig2.pdf}}\quad.$$ The bilocal lines coming from the explicit Weyl factors are represented by the red lines. The important point is that since this additional structure only depends on $z$, it does not participate in the Fourier transform over $t$, and factorizes from the amplitudes. Furthermore Euclidean wormholes connecting to these new regions in only give subdominant contributions because we are integrating over the corresponding energy labels of these regions. Hence the only effect of including these Weyl factors in is that and receive an additional overall prefactor which depends on $z$ but which is crucially independent of $\omega$. By consequence one still finds the same physics as in for example . The fermionic coupling requires exactly the same treatment with Weyl factors on either side of the $\ell=1/2$ pair of lines in . We note to conclude that one could consider more generic detector couplings, of higher order in the fields, such as: $$H_\text{int} = g\,\mu(t)\,f(\phi(u(t),v(t))\quad,\quad f(\phi(u(t),v(t)) = \sum_{n} c_n (\phi^n)(u(t),v(t)) \, .$$ The precise Schwarzian and Euclidean wormhole computations are more involved. The resulting detector response would be obtained as the same Taylor series expansion: $$R_g(\omega)=\sum_n c_n R_n(\omega) \, .$$ Intuitively, we expect to see level repulsion in any such response rate given its universal role in random matrix correlation functions. Energy density in the Unruh heat bath {#s:hb} ===================================== In this section we investigate the spectral energy density in the Unruh heat bath. Semi-classically this is given by $\bra{\mM}\omega\,a_\omega^\dagger a_\omega \ket{\mM}$, so we might expect the gravitational corrections to be accounted for by considering instead $\bra{\mM}\omega\,A_\omega^\dagger[f] A_\omega[f] \ket{\mM}$. However, this is not true. This traces back to operator ordering ambiguities when promoting quantum field observables in a fixed metric to operators in a theory of matter-coupled quantum gravity.\  \ Let us first focus on coupling to the Schwarzian reparametrization. The local energy density in the Unruh heat bath can be computed independently using the coincident limit of the bulk two-point function (regularized via point splitting) [@spradlin]. The result is: $$\label{Unruhflux} \left\langle :T_{uu}[f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u]:\right\rangle_{{\scriptscriptstyle \text{CFT}}} = -\frac{1}{24\pi}\text{Sch}\left(\tanh\frac{\pi}{\beta}f,u \right), \quad \left\langle :T_{vv}[f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}v]:\right\rangle_{{\scriptscriptstyle \text{CFT}}} = -\frac{1}{24\pi}\text{Sch}\left(\tanh\frac{\pi}{\beta}f,v\right) \, .$$ Doing the gravitational path integral one finds [@Mertens:2019bvy] $$\begin{aligned} \label{Unruhflux2} \bra{\mM}:T_{uu}[f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u]:\ket{\mM} &= \bra{\mM}:T_{vv}[f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}v]:\ket{\mM} =\frac{M}{12 \pi}\, .\end{aligned}$$ The total energy in the heat bath is then[^26] $$E_{\text{bath}} = \int_{-\infty}^{+\infty} du\, \bra{\mM}:T_{uu}[f {\,\rule[-1.2pt]{.2pt}{1.7ex}\,}u]:\ket{\mM}=\frac{V M}{12\pi}\, .\label{33}$$ The bath energy $E_{\text{bath}}$ is defined operationally by summing the bath energy densities $T_{uu}$ across an entire spatial slice. However, this is not the total matter energy that backreacts on the geometry: including the vacuum energy gives a net zero result. In flat spacetime this famously leads to the statement that the energy in the Unruh heat bath does not deform the flat geometry one started with.\  \ Our goal here is to find a spectral quantity that reproduces this bath energy $E_{\text{bath}}$, and the first candidate is $\bra{\mM}\omega\,A_\omega[f]^\dagger A_\omega[f] \ket{\mM}$. However, one finds very explicitly from that: $$\int_{0}^{+\infty}d\omega \bra{\mM}\omega\,A_\omega^\dagger[f] A_\omega[f] \ket{\mM}\, \neq \, E_{\text{bath}}\, .\label{34}$$ The potential confusion is closely related to the fact that the operators $A_\omega$ constructed in are not quite the same as raising and lowering operators acting on the matter Hilbert space (nor where they constructed to be). For example, we may compute: $$\begin{aligned} \bra{\mM}[A_{\omega_1},A^\dagger_{\omega_2}]\ket{\mM} = \frac{\delta(\omega_1-\omega_2)}{\omega_1}\,&\rho_0(M+\omega_1)\,\rvert{\mathcal{O}}_{M \, M+\omega_1}^1\rvert^2\nonumber \\ &-\frac{\delta(\omega_1-\omega_2)}{\omega_1}\,\rho_0(M-\omega_1)\,\rvert {\mathcal{O}}_{M\, M-\omega_1}^1\rvert^2\, .\label{35}\end{aligned}$$ In the probe approximation $M\gg 1$ and $\omega\ll M$ this does reduce to $\delta(\omega_1-\omega_2)$. However for more generic values of $M$ and $\omega$, it does not. We will below define the spectral energy density $\omega N_\omega[f]$ that does satisfy this relation: $$\int_{0}^{+\infty}d\omega \bra{\mM}\omega N_\omega[f]\ket{\mM} = E_{\text{bath}} \, ,$$ and find that it corresponds to a symmetrized version of $\omega A^{\dagger}_\omega A_\omega$. Symmetric number operator ------------------------- The goal here is to introduce a symmetric number operator $N_\omega[f]$, which does compute the bath spectral energy density. Consider first semi-classical matter. Following one defines a number operator $${\left\langle a^\dagger_\omega a_\omega \right\rangle}_{{\scriptscriptstyle \text{CFT}}} = \frac{1}{\pi\omega}\int_{-\infty}^{+\infty} du_1\int_{-\infty}^{+\infty}du_2\, e^{-i\omega(u_1-u_2)}\, {\left\langle \partial\phi(u_1) \partial\phi(u_2) \right\rangle}_{{\scriptscriptstyle \text{CFT}}}\, .\label{38}$$ This features the Wightman bulk two-point function. Using the canonical commutation relation $$a^\dagger_\omega a_\omega =\frac{1}{2}a^\dagger_\omega a_\omega+\frac{1}{2} a_\omega a^\dagger_\omega -\frac{1}{2}\delta(0)\, ,\label{39}$$ we can write this in a more symmetric way as: $$\begin{aligned} \label{symoper} {\left\langle a^\dagger_\omega a_\omega \right\rangle}_{{\scriptscriptstyle \text{CFT}}} &= \frac{1}{\pi\omega}\int_{-\infty}^{+\infty} du_1\int_{-\infty}^{+\infty}du_2\, e^{-i\omega(u_1-u_2)}\, {\left\langle \partial\phi(u_1) \partial\phi(u_2) \right\rangle}_{{\scriptscriptstyle \text{CFT}}}\nonumber \\&\qquad+ \frac{1}{\pi\omega}\int_{-\infty}^{+\infty} du_1\int_{-\infty}^{+\infty}du_2\, e^{-i\omega(u_1-u_2)}\, {\left\langle \partial\phi(u_2) \partial\phi(u_1) \right\rangle}_{{\scriptscriptstyle \text{CFT}}} - \frac{1}{2} \delta(0)\,.\end{aligned}$$ Semiclassically, this manipulation is completely harmless. The point is that and turn out not to be equivalent when promoting the Wightman two-point functions to operators in quantum gravity. This has to do with operator orderings. The two Wightman bulk two-point functions are not precisely identical. Including an appropriate $i\varepsilon$ regulator we have $$\begin{aligned} \label{wight} {\left\langle \partial\phi(u_1) \partial\phi(u_2) \right\rangle}_{{\scriptscriptstyle \text{CFT}}}&=-\frac{1}{4\pi}\,\frac{1}{\frac{\beta^2}{\pi^2}\sinh^2 \frac{\pi}{\beta}(u_1-u_2+i\epsilon)}\nonumber\\ {\left\langle \partial\phi(u_2) \partial\phi(u_1) \right\rangle}_{{\scriptscriptstyle \text{CFT}}}&=-\frac{1}{4\pi}\,\frac{1}{\frac{\beta^2}{\pi^2}\sinh^2\frac{\pi}{\beta}(u_1-u_2-i\epsilon)}\,.\end{aligned}$$ In Schwarzian quantum gravity, we can then likewise define the symmetrized version of the dressed operators $A^{\dagger}_\omega A_\omega$ and $A_\omega A^{\dagger}_\omega$ as: $$\begin{aligned} B^\dagger_\omega[f] B_\omega[f] &= \frac{1}{2}A^\dagger_\omega[f] A_\omega[f] + \frac{1}{2}A_\omega[f] A^\dagger_\omega[f] -\frac{1}{2}\delta(0)\, \\ B_\omega[f] B^\dagger_\omega[f] &= \frac{1}{2}A^\dagger_\omega[f] A_\omega[f] + \frac{1}{2}A_\omega[f] A^\dagger_\omega[f] + \frac{1}{2}\delta(0)\, .\end{aligned}$$ By construction these modes $B_\omega$ automatically satisfy the commutator relation: $$[B_{\omega_1}[f],B_{\omega_2}^\dagger[f]]=\delta(\omega_1-\omega_2)\, .\label{313}$$ Dressing the bulk Wightman two-point functions in to include for Schwarzian gravitational interactions as in , we write for $N_\omega[f] = B^\dagger_\omega B_\omega$ [@Mertens:2019bvy]: $$\begin{aligned} \label{planckp} &{\left\langle N_\omega[f] \right\rangle}_{{\scriptscriptstyle \text{CFT}}} \\\nonumber &= -\frac{1}{8\pi^2\omega}\int_{-\infty}^{+\infty} d u_1 \int_{-\infty}^{+\infty} du _2\, e^{-i\omega (u_1-u_2)}\,\frac{f'(u_1)f'(u_2)}{\frac{\beta^2}{\pi^2}\sinh^2 \frac{\pi}{\beta}(f(u_1)-f(u_2)+i\varepsilon)} - \frac{1}{(u_1-u_2+i\varepsilon)^2}\\\nonumber &\quad -\frac{1}{8\pi^2\omega}\int_{-\infty}^{+\infty} d u_1 \int_{-\infty}^{+\infty} du _2\, e^{-i\omega (u_1-u_2)}\,\frac{f'(u_1)f'(u_2)}{\frac{\beta^2}{\pi^2}\sinh^2 \frac{\pi}{\beta}(f(u_1)-f(u_2)-i\varepsilon)} - \frac{1}{(u_1-u_2-i\varepsilon)^2}\,.\end{aligned}$$ Here we have rewritten the delta-function in as a term which explicitly subtracts the poles in the Wightman two-point functions.[^27] With this operator, we first compute the energy: $$\int_0^\infty d\omega\,{\left\langle \omega N_\omega[f] \right\rangle}_{{\scriptscriptstyle \text{CFT}}} \, .$$ By swapping the integration variables $u_1$ and $u_2$ in the integrals of , we see that the terms on the middle and last line are mapped into one another if we change the sign of $\omega$. Therefore, including a $1/2$ factor to enlarge the integration over $\omega$ along the entire real axis, we obtain a factor $$\int_{-\infty}^{+\infty} d\omega\,e^{-i\omega(u_1-u_2)}=2\pi \delta(u_1-u_2) \, .$$ One then finds directly from Taylor expanding [^28] $$\begin{aligned} - \frac{1}{8\pi} \int_{-\infty}^{+\infty} du_1\int_{-\infty}^{+\infty} d u_2 \,\delta(u_1-u_2)\,&\frac{f'(u_1)f'(u_2)}{\frac{\beta^2}{\pi^2}\sinh^2 \frac{\pi}{\beta}(f_1-f_2-i\epsilon)} - \frac{1}{(u_{1}-u_{2}-i\epsilon)^2} + (\epsilon \to - \epsilon) \nonumber \\&\qquad\qquad= - \frac{1}{24\pi} \int d y\, \text{Sch}\left(\tanh \frac{\pi}{\beta}f,y\right) \, .\end{aligned}$$ As classical functions this is true in any case. However, in the Schwarzian path integral this is only true when we work with the symmetrically dressed operators $B_\omega$. The reason is that the two Wightman two-point functions are not equal to one another after the Schwarzian path integral.[^29] The above trick $\omega \to - \omega$ maps one into the other, but since we started with a symmetric combination this has no impact. This confirms that we may view the Schwarzian path integral of as the spectral energy density $\bra{\mM}\omega N_\omega[f] \ket{\mM}$ in the Unruh heat bath. This agrees with the formulas presented in [@Mertens:2019bvy]. Analytical analysis ------------------- Doing the Schwarzian path integrals in , we find $$\begin{aligned} \bra{\mM}\omega N_\omega[f]\ket{\mM} = \frac{V}{4\pi}\,\rho_0(M+\omega)\,\rvert {\mathcal{O}}_{M M+\omega}^1\rvert^2+\frac{V}{4\pi}\,\rho_0(M-\omega)\,\rvert {\mathcal{O}}_{M M-\omega}^1\rvert^2-\frac{\omega}{2}\,\delta(0) \, .\label{319}\end{aligned}$$ To deal with the delta-function, we will choose to calibrate our measurement to the zero energy [Poincaré ]{}state[^30] $$\bra{0}\omega N_\omega[f]\ket{0}=\frac{V}{4\pi}\,\rho_0(\omega)\,\rvert {\mathcal{O}}_{0\, 0+\omega}^1\rvert^2+\frac{V}{4\pi}\,\rho_0(-\omega)\,\rvert {\mathcal{O}}_{0\,0-\omega}^1\rvert^2-\frac{\omega}{2}\,\delta(0)\, .\label{320}$$ We will henceforth only discuss this relative spectral energy density $$\begin{aligned} \bra{\mM}\omega N_\omega[f]\ket{\mM}=&\frac{V}{4\pi}\,\rho_0(M+\omega)\,\rvert {\mathcal{O}}_{M M+\omega}^1\rvert^2-\frac{V}{4\pi}\,\rho_0(\omega)\,\rvert {\mathcal{O}}_{0\, 0+\omega}^1\rvert^2\nonumber\\&+\frac{V}{4\pi}\,\rho_0(M-\omega)\,\rvert {\mathcal{O}}_{M M-\omega}^1\rvert^2-\frac{V}{4\pi}\,\rho_0(-\omega)\,\rvert {\mathcal{O}}_{0\,0-\omega}^1\rvert^2 \, .\label{321}\end{aligned}$$ In the semi-classical probe approximation $M\gg 1$ and $\omega\ll M$ one recovers the classical answer for the spectral energy density $$\bra{\mM}\omega N_\omega[f]\ket{\mM}=\frac{V}{2\pi}\frac{\omega}{e^{\beta\omega}-1}\, .\label{322}$$ The Schwarzian answer which combines and is accurate for any $M\gg e^{-2S_0/3}$ and for any $\omega\gg e^{-S_0}$. This includes Planck sized black holes where $M\sim 1$. For such tiny black holes and with $\omega < M$ one finds slightly lower spectral energy density in as compared to . See figure \[Urandomz2\]. On the other hand for $\omega>M$ the Schwarzian result gives a slightly higher occupation as compared to the classical answer . Including Euclidean wormhole corrections to the Schwarzian correlations is done analogously as in , and leads to: $$\begin{aligned} \nonumber \bra{\mM}\omega N_\omega[f]\ket{\mM}=&\frac{V}{4\pi}\,\frac{\rho(M,M+\omega)}{\rho(M)}\,\rvert {\mathcal{O}}_{M M+\omega}^1\rvert^2-\frac{V}{4\pi}\,\frac{\rho(0,0+\omega)}{\rho(0)}\,\rvert {\mathcal{O}}_{0\,0+\omega}^1\rvert^2\nonumber\\&+\frac{V}{4\pi}\,\frac{\rho(M,M-\omega)}{\rho(M)}\,\rvert {\mathcal{O}}_{M M-\omega}^1\rvert^2-\frac{V}{4\pi}\,\frac{\rho(0,0-\omega)}{\rho(0)}\,\rvert {\mathcal{O}}_{0\,0-\omega}^1\rvert^2\, .\label{323}\end{aligned}$$ The behavior of this function is in fact quite similar to . In particular we see quadratic level repulsion for $\omega \sim e^{-S_0}$. One notable difference with is the effect of the zero energy subtractions. As we review in appendix \[s:zeroref\] these come which much slower wiggles as compared to the oscillations in . The last term in is in the forbidden region and is suppressed as $\sim e^{-S_0}$ in any case, making it negligible in practice. Supersymmetric JT gravity and the resulting random matrix completion from the Altland-Zirnbauer ensembles have a hard spectral edge at $\omega=0$ [@Stanford:2019vob] removing the forbidden region alltogether. We have checked numerically for that this indeed computes the spectral energy density , by satisfying $$\label{toten} \int_0^{+\infty} d\omega \bra{\mM}\omega N_\omega[f]\ket{\mM} = \frac{V M}{12\pi} \,.$$ Upon including higher topology in , there are several tiny corrections. This is perfectly fine since neither nor is precise when including Euclidean wormhole corrections. We note that via one finds that such corrections are at least suppressed by a factor $e^{-S_0}$. Therefore they can be neglected.[^31] For completeness, we note that the contact term contribution in implies a further contribution to of the form: $$\begin{aligned} \bra{\mM}\omega N_\omega[f]\ket{\mM} \, \supset \, \delta(\omega)\,\frac{V}{2\pi}\, \bigl(\rvert {\mathcal{O}}_{M M}^1\rvert^2-\,\rvert{\mathcal{O}}_{0\,0}^1\rvert^2 \, \bigr) \,.\end{aligned}$$ It is impossible to measure precisely zero energy so this contribution seems less interesting. On the other hand, there are several known examples of zero energy modes being important in a black hole context [@Donnelly:2014fua; @Donnelly:2015hxa; @Blommaert:2018rsf; @Blommaert:2018oue]. Numerical analysis ------------------ It is clarifying to explicitly plot the spectral energy density as computed in the three levels of improving approximation in , and . We note that the zero-energy kernels in are different from the universal answer , which only holds far enough from the spectral edge. For the zero-energy kernels we may utilize instead results from the exactly solvable Airy model.[^32] We present some details in appendix \[s:zeroref\].\  \ The effects which we aim to see in the plots are 1. The main effect of the Schwarzian corrections is that the Schwarzian answer (red) lies below the semiclassical answer (green) for small $\omega$ and above the semiclassical answer for large $\omega$. 2. Secondly and most importantly, we can clearly see the signs of level repulsion. There is a depletion in the spectral density of the Unruh heat bath for $\omega\ll e^{-S_0}$ and furthermore we see high-frequency wiggles in the regime where $\omega$ is order $e^{-S_0}$. 3. Finally, there are similar wiggles associated to the zero-energy subtraction in . These will play a role when $\omega$ is order $e^{-2S_0/3}$. For $M\gg 1$ these Airy wiggles effectively become invisible as all contributions to grow exponentially with $M$. We will consider a parametric regime where all these effects are clearly visible. Therefore we take $M=2$ and $S_0=10$ in both figures \[Urandomz2\] and \[Urandomfull\]. ![(left) Zoom-in on the region with $\omega$ or order $e^{-S_0}$ where we see level repulsion and high-frequency wiggles in the exact (blue) result. (right) Zoom-in on the region $\omega$ of order $e^{-2S_0/3}$ where we see the slower Airy wiggles from the zero energy reference . Furthermore one sees that the Schwarzian answer (red) is lower than the semi-classical answer (green).[]{data-label="Urandomz2"}](Urandomzv2.pdf){width="95.00000%"} ![The Schwarzian effects (right) and level repulsion effects (left) are simultaneously visible in this log plot of $\bra{\mM}\omega N_\omega[f]\ket{\mM}$. One may compare the exact answer (blue) with the Schwarzian answer (red) and the semiclassical answer (green).[]{data-label="Urandomfull"}](Urandomfull.pdf){width="95.00000%"} Concluding remarks {#s:concl} ================== The main goal of this work was to advocate and provide evidence that Hawking-Unruh radiation is highly sensitive (at ultra low energies) to level repulsion in the chaotic spectrum of the underlying quantum black hole. We made this explicit by probing a massless scalar field coupled to JT gravity using an Unruh-DeWitt detector. The calculation involves including Euclidean wormhole corrections to a massless scalar bulk two-point function in JT gravity. Due to random matrix universality for quantum black holes, we expect our conclusion to be quite universal and to qualitatively hold in any number of dimensions. One immediate way to test the universality of these ideas is to consider an Unruh-DeWitt detector in super JT gravity or to consider charged versions of JT gravity. Results in this direction are forthcoming.\  \ We end this work by emphasizing three features of our setup that deserve more work.\  \ ***Gravitational dressings***\  \ One important aspect of this work is navigating through different types of operator dressings and operator ordering ambiguities. In particular, for the gravitationally dressed matter modes $A_\omega[f]$ in , we obtained the expectation value of their correlator . Schematically, we way write this in a canonical language as $$[A_{\omega_1}^\dagger[f], A_{\omega_2}[f]]=\delta(\omega_1-\omega_2)+ \text{gravitational corrections}.$$ The first contribution is due to the canonical oscillator algebra of the undressed matter modes $a_\omega$. The second contribution is due to canonical commutators of the gravitational variables. Indeed, the matter modes $A_\omega[f]$ include a dressing with gravitational variables. See for example [@Donnelly:2015hta; @gid3] and section 5 of [@ads2]. This should be contrasted with the behavior of the dressed matter modes $B_\omega[f]$ in for which we write by construction $$[B_{\omega_1}^\dagger[f], B_{\omega_2}[f]]=\delta(\omega_1-\omega_2).$$ In defining $B_\omega[f]^\dagger B_\omega[f]$ via we have specified a symmetric gravitational dressing for this combination of modes, aimed in a precise way such that the canonical algebra of the modes would receive no gravitational corrections. This means a single mode $B_\omega[f]$ is by itself not a diff-invariant observable but the gravitational dressed composite operator $N_\omega[f] = B_\omega[f]^\dagger B_\omega[f]$ is. On the other hand, the modes $A_\omega[f]$ are well defined diff-invariant operators. The definition of the Unruh-DeWitt detector requires we use such individually well-defined diff-invariant modes.\  \ More broadly speaking there are numerous operator ordering ambiguities whenever we are promoting a correlator or a field in semiclassical physics to an operator in quantum gravity. One might refer to all of these as choices associated with gravitational dressings [@Donnelly:2015hta; @gid3]. However in the real world we are all gravitationally dressed composite objects with an implicit choice of dressing. This begs the questions: which dressings are actually natural in quantum gravity and which are just mathematical curiosities? Given a particular physical context, what is the appropriate dressing? In our particular case of Hawking-Unruh radiation, we were able to pinpoint a natural type of dressing for two distinct experiments. For the Unruh-DeWitt detector experiment we were led to work with the dressed matter modes $A_\omega[f]$ whereas in an experiment which measures the spectral energy density in the heat bath the modes $B_\omega[f]$ turned out to be relevant. Nevertheless, these dressing ambiguities remain largely elusive.\  \ ***Higher genus bulk correlators***\  \ Another important aspect of our story is a prescription for how to include Euclidean wormhole corrections to inherently Lorentzian bulk observables in JT gravity. Let us summarize the general idea by an algorithm 1. Find an expression for the bulk matter correlator in the reparameterized metrics . 2. Attempt to rewrite that expression as a combination of Schwarzian bilocal operators. An algorithmic way of doing so is to reverse engineer bulk reconstruction in each of the metrics . Write each term as a Euclidean JT gravity path integral on the disk including corresponding boundary to boundary matter propagators. 3. Include Euclidean wormhole contributions to each such individual JT gravity boundary correlator as done in e.g. [@phil]. The result is a correlation function in a double-scaled matrix integral. This is identical to summing over all Riemann surfaces which end on the union of the bilocal lines and the boundary.[^33] Effectively this leads to replacing $\rho_0(E_1)\dots \rho_0(E_n)$ with $\rho(E_1\dots E_n)$ in the Schwarzian correlators. The correlators $\rho(E_1\dots E_n)$ are multi-level correlators of a double-scaled matrix integral and may be obtained via universal random matrix cluster functions [@mehta; @paper6]. 4. Sum over all these boundary correlators in order to obtain the bulk correlator. One pragmatic way to argue for this prescription is that the resulting bulk correlators end up showcasing certain generic physical principles. One example is level repulsion in Hawking-Unruh radiation as discussed in this work. A second example is the behavior of bulk matter correlators at large distances in a finite entropy system as will be discussed in [@wophilbert].\  \ ***Implications for evaporation?***\  \ Our discussion on the Unruh-DeWitt detector experiment is an idealization of a more realistic experiment where the measurement takes place over an infinite amount of time. In a more realistic experiment, we would measure for a finite time ${\small \text{T}}$. This can be implemented in the formulas by introducing in the coupling an additional switching function $\chi(t)$ which has a width of order ${\small \text{T}}$. This introduces a factor $\chi(t_1)\chi(t_2)$ in the integrand on the second line of . If we denote the Fourier transform of $\chi(t)$ by $\hat{\chi}(\omega)$ then is replaced by a convolution of the previous answer with the frequency content of the switching function $$R(\omega) \sim \frac{1}{{\small \text{T}}} \int_{\mathcal{C}} d\tilde{\omega} \, \hat{\chi}^2(\tilde{\omega}-\omega) \, \frac{\rho(M,M - \tilde{\omega})}{\rho(M)} \left|\mathcal{O}^1_{M \, M - \tilde{\omega}}\right|^2\, \frac{\sin^2 z \tilde{\omega}}{\tilde{\omega}^2} \, .$$ This convolution replaces but also for example with a version that is smooth on frequency scale of order $1/{{\small \text{T}}}$. This implies we must measure for a time ${{\small \text{T}}}\gg 1/\beta$ in order to resolve the semiclassical Planckian black body law. The level repulsion in and the delta spikes in can only be resolved if we measure for a time ${{\small \text{T}}} \gg e^{S_0}$. In a non-evaporating setup this remains a sensible experiment. However for an evaporating black hole on these time scales we are in the regime where the Page curve is decreasing [@rw1] and any eternal approximation no longer applies. Nevertheless it is quite natural that the information about the microstates in our eternal model of quantum gravity can only be accessed from a measurement that takes longer than what would be the Page time in an evaporating setup. At the very least this work emphasizes that we should expect extraordinary nonperturbative effects in gravity to highly affect Hawking radiation at long time scales ${{\small \text{T}}}\gg e^{S_0}$. These effects can be probed directly in the quantum gravity bulk using an Unruh-DeWitt detector. Acknowledgements {#acknowledgements .unnumbered} ================ We thank Jordan Cotler, Julius Engelsoy and Joaquin Turiaci for useful discussions. AB and TM gratefully acknowledge financial support from FWO Vlaanderen. Detectors with other boundary conditions {#app:obc} ======================================== In section \[s:udw\] we studied an Unruh-DeWitt detector that couples to a massless scalar field with Dirichlet boundary conditions . One might be interested in bosonic bulk matter with other boundary conditions. The result generalizes as follows. For example, imposing Neumann boundary conditions one finds $$\begin{aligned} &{\left\langle \Phi[f\rvert u_1,v_1]\Phi[f\rvert u_2,v_2] \right\rangle}_{{\scriptscriptstyle \text{CFT}}}\nonumber\\&\quad=-\frac{1}{4\pi}\ln (F(u_1)-F(u_2))(F(v_1)-F(v_2))(F(v_1)-F(u_2))(F(u_2)-F(v_2))\, .\label{a1}\end{aligned}$$ As a Schwarzian insertion, notice that this operator is not SL$(2,\mathbb{R})$ invariant. For the Fourier transform, we can integrate by parts twice: $$\begin{aligned} \nonumber &\lim_{{{\small \text{T}}} \to +\infty}\frac{1}{{\small \text{T}}} \int_{-{\small \text{T}}}^{+{\small \text{T}}}dt_1\int_{-{\small \text{T}}}^{+{\small \text{T}}} dt_2\, e^{-i\omega (t_1-t_2)}\,\ln (F(u_1)-F(u_2))\\&\qquad =\frac{1}{\omega^2}\lim_{{{\small \text{T}}} \to +\infty}\frac{1}{{\small \text{T}}} \int_{-{\small \text{T}}}^{+{\small \text{T}}}dt_1\int_{-{\small \text{T}}}^{+{\small \text{T}}} dt_2\, e^{-i\omega (t_1-t_2)}\,\frac{F'(u_1)F'(u_2)}{(F(u_1)-F(u_2))^2} \, .\end{aligned}$$ Notice that hence by integrating the above operator , the result *is* SL$(2,\mathbb{R})$ invariant. The full answer is a sum over four such terms, each with a generically different phase factor. Summing those phase factors given a greybody factor $\cos^2 z\omega / \omega^2$ replacing the $\sin^2 z\omega / \omega^2$ in the Dirichlet case. Including Euclidean wormhole corrections, one obtains $$R(\omega) = 2\frac{\cos^2 z\omega}{\omega^2}\, \frac{\rho(M,M-\omega)}{\rho(M)}\, \rvert {\mathcal{O}}_{M M-\omega}^1\rvert^2\, .\label{a4}$$ In terms of an experimental response the effect in this case is quite different to . The semiclassical answer now blows up for $\omega\ll 1$ due to the double pole in the interference factor. Level repulsion regulates this pole and results in a finite answer for $\omega\ll e^{-S_0}$. For generic conformally invariant boundary conditions, the greybody factor is $$\frac{1-a \cos 2\omega z}{\omega^2}\,.$$ Hera $a=1$ is Dirichlet, $a=-1$ is Neumann, and $a=0$ corresponds to transparant boundary conditions, where one only takes the first two factors in the logarithm of . So we see that the behavior of for $\omega\ll e^{-S_0}$ is an exception to the general rule where the semiclassical answer has a double pole for $\omega\ll 1$ which is regulated by level repulsion. Airy model and zero energy reference term {#s:zeroref} ========================================= Let us briefly address the zero energy [Poincaré ]{}reference contribution to the spectral energy density $$\bra{0}\omega N[f|\omega]\ket{0}=\frac{V}{4\pi}\,\frac{\rho(0,\omega)}{\rho(0)}\Gamma(1 \pm i\sqrt{\omega})^2+\frac{V}{4\pi}\,\frac{\rho(0,-\omega)}{\rho(0)}\Gamma(1 \pm i\sqrt{\omega})^2 \, .\label{b1}$$ We note that the contribution from the second term in is essentially negligible for every $\omega$ as it is evaluated in the forbidden region where there is barely any density.\ Consider first $\rho(0,\omega)$ for $\omega\ll 1$. Both energies are close to the spectral edge and therefore we must use the two-level spectral density of the Airy model instead of : $$\frac{\rho(E_1,E_2)}{\rho(E_2)}=\rho(E_1)+\delta(E_1-E_2)-\frac{K(E_1,E_2)^2}{\rho(E_2)} \, .$$ Here the spectral density is $$\rho(E) = e^{\frac{2S_0}{3}}\,{\text{Ai}}'(- e^{\frac{2S_0}{3}} E)^2-e^{\frac{2S_0}{3}}\xi {\text{Ai}}(- e^{\frac{2S_0}{3}} E)^2\,.\label{b2}$$ Furthermore the Airy kernel is[^34] $$\label{airykernel} K(E_1,E_2) = \frac{{\text{Ai}}'(-e^{\frac{2S_0}{3}}E_1){\text{Ai}}(-e^{\frac{2S_0}{3}}E_2) - {\text{Ai}}'(-e^{\frac{2S_0}{3}}E_2){\text{Ai}}(-e^{\frac{2S_0}{3}}E_1)}{E_1-E_2}\, .$$ This kernel is only relevant for $E_1-E_2$ of order $e^{-S_0}$. Otherwise one finds: $$\frac{\rho(0,\omega)}{\rho(0)}=\rho(\omega)\quad,\quad \omega \gg e^{-S_0}\, .$$ Using this one finds: $$\bra{0}N_\omega[f]\ket{0}=\frac{V}{4\pi}\quad,\quad \omega\gg 1 \, .\label{b6}$$ This zero-energy subtraction ensures that the relative spectral energy density $\bra{M}\omega N_\omega[f]\ket{M}$ in goes to zero smoothly for $\omega \gg M$. We furthermore note that is positive definite for any $M>0$. This is explicit in figure \[Urandomz2\]. For $M\gg 1$ is is also quite easy to check this explicitly. This whole zero energy contribution is strongly suppressed by exponentials of $M$ at any $\omega$ scale.\  \ In order to construct a plot of we use the Airy result for $\omega\ll 1$, and the smooth JT gravity spectral density $\rho(\omega)$ as determined in [@sss2] $$\rho(\omega)=\rho_0(\omega)-\frac{e^{-S_0}}{4\pi\omega}\,\cos(2\pi e^{S_0}\int_0^\omega d E\,\rho_0(E)) \, .\label{b7}$$ This can be trusted for $\omega\gg e^{-2S_0/3}$. Since $S_0 \gg 1$, these regions overlap so we can construct the whole plot. Both $\rho(\omega)$ and $\bra{0}N_\omega[f]\ket{0}$ are plotted for any positive $\omega$ in figure \[Urandomref\] which uses $S_0=10$. Notice the wiggles in the latter which shine through in figure \[Urandomz2\] (right). ![ (left) Spectral density $\rho(\omega)$ in JT gravity for any positive energy $\omega$. The Airy answer is accurate for $\omega$ much smaller then the right dotted line. The JT gravity answer of [@sss2] is accurate for $\omega$ much greater then the first dotted line. (right) $\bra{0}N_\omega[f]\ket{0}$ for any positive $\omega$. The asymptotic limit is shown (green). []{data-label="Urandomref"}](airyLRv2.pdf){width="95.00000%"} [99]{} J. S. Cotler [*et al.*]{}, “Black Holes and Random Matrices,” JHEP [**1705**]{} (2017) 118 Erratum: \[JHEP [**1809**]{} (2018) 002\] [[arXiv:1611.04650 \[hep-th\]]{}](https://arxiv.org/abs/1611.04650). D. A. Roberts, D. Stanford and A. Streicher, “Operator growth in the SYK model,” JHEP **06**, 122 (2018) [[arXiv:1802.02633 \[hep-th\]]{}](http://arxiv.org/abs/1802.02633) S. H. Shenker and D. Stanford, “Black holes and the butterfly effect,” JHEP [**1403**]{}, 067 (2014) [[arXiv:1306.0622 \[hep-th\]]{}](http://arxiv.org/abs/1306.0622) S. H. Shenker and D. Stanford, “Multiple Shocks,” JHEP **12**, 046 (2014) [[arXiv:1312.3296 \[hep-th\]]{}](http://arxiv.org/abs/1312.3296) S. H. Shenker and D. Stanford, “Stringy effects in scrambling,” JHEP [**1505**]{}, 132 (2015) [](https://arxiv.org/abs/1412.6087) J. Maldacena, S. H. Shenker and D. Stanford, “A bound on chaos,” JHEP **08**, 106 (2016) [[arXiv:1503.01409 \[hep-th\]]{}](http://arxiv.org/abs/1503.01409) M. Mehta, Random Matrices. Pure and Applied Mathematics. Elsevier Science, 2004. F. Haake, Quantum Signatures of Chaos. Springer, 2010. D. Kapec, R. Mahajan and D. Stanford, “Matrix ensembles with global symmetries and ’t Hooft anomalies from 2d gauge theory,” JHEP [**2004**]{} (2020) 186 [[arXiv:1912.12285 \[hep-th\]]{}](http://arxiv.org/abs/1912.12285) Y. Sekino and L. Susskind, “Fast Scramblers,” JHEP **10**, 065 (2008) [[arXiv:0808.2096 \[hep-th\]]{}](http://arxiv.org/abs/0808.2096) P. Saad, S. H. Shenker and D. Stanford, “JT gravity as a matrix integral,” arXiv:1903.11115 \[hep-th\]. [[arXiv:0106112 \[hep-th\]]{}](http://arxiv.org/abs/0106112) J. Maldacena, D. Stanford and Z. Yang, “Conformal symmetry and its breaking in two dimensional Nearly Anti-de-Sitter space,” PTEP [**2016**]{}, no. 12, 12C104 (2016) [[arXiv:1606.01857 \[hep-th\]]{}](http://arxiv.org/abs/1606.01857). H. T. Lam, T. G. Mertens, G. J. Turiaci and H. Verlinde, “Shockwave S-matrix from Schwarzian Quantum Mechanics,” JHEP [**1811**]{} (2018) 182 [[arXiv:1804.09834 \[hep-th\]]{}](http://arxiv.org/abs/1804.09834) M. Srednicki, “Chaos and quantum thermalization,” Physical Review E [**[50]{}**]{} no. 2, (1994) 888. J. Deutsch, “Quantum statistical mechanics in a closed system,” Physical Review A [**[43]{}**]{} no. 4, (1991) 2046. P. Saad, “Late Time Correlation Functions, Baby Universes, and ETH in JT Gravity,” [[arXiv:1910.10311 \[hep-th\]]{}](http://arxiv.org/abs/1910.10311) J. M. Maldacena, “Eternal black holes in anti-de Sitter,” JHEP [**0304**]{} (2003) 021 [[arXiv:0106112 \[hep-th\]]{}](http://arxiv.org/abs/0106112) P. Saad, S. H. Shenker and D. Stanford, “A semiclassical ramp in SYK and in gravity,” arXiv:1806.06840 \[hep-th\]. [[arXiv:1806.06840 \[hep-th\]]{}](http://arxiv.org/abs/1806.06840). A. Blommaert, T. G. Mertens and H. Verschelde, “Clocks and Rods in Jackiw-Teitelboim Quantum Gravity,” JHEP [**1909**]{} (2019) 060 [[arXiv:1902.11194 \[hep-th\]]{}](http://arxiv.org/abs/1902.11194) A. Blommaert, T. G. Mertens and H. Verschelde, “Eigenbranes in Jackiw-Teitelboim gravity,” [[arXiv:1911.11603 \[hep-th\]]{}](http://arxiv.org/abs/1911.11603) D. Marolf and H. Maxfield, “Transcending the ensemble: baby universes, spacetime wormholes, and the order and disorder of black hole information,” [[arXiv:2002.08950 \[hep-th\]]{}](http://arxiv.org/abs/2002.08950) A. Blommaert, “To appear” G. Penington, S. H. Shenker, D. Stanford and Z. Yang, “Replica wormholes and the black hole interior,” [[arXiv:1911.11977 \[hep-th\]]{}](http://arxiv.org/abs/1911.11977) A. Almheiri, T. Hartman, J. Maldacena, E. Shaghoulian and A. Tajdini, “Replica Wormholes and the Entropy of Hawking Radiation,” [[arXiv:1911.12333 \[hep-th\]]{}](http://arxiv.org/abs/1911.12333) L. Aalsma and G. Shiu, “Chaos and complementarity in de Sitter space,” [[arXiv:2002.01326 \[hep-th\]]{}](http://arxiv.org/abs/2002.01326) A. Blommaert and J. Cotler, “To appear” J. Cotler and K. Jensen, “Emergent unitarity in de Sitter from matrix integrals,” [[arXiv:1911.12358 \[hep-th\]]{}](http://arxiv.org/abs/1911.12358) T. G. Mertens, “Towards Black Hole Evaporation in Jackiw-Teitelboim Gravity,” JHEP [**1907**]{} (2019) 097 [[arXiv:1903.10485 \[hep-th\]]{}](http://arxiv.org/abs/1903.10485) D. Stanford and E. Witten, “JT Gravity and the Ensembles of Random Matrix Theory,” [[arXiv:1907.03363 \[hep-th\]]{}](http://arxiv.org/abs/1907.03363) R. Jackiw, “Lower Dimensional Gravity,” Nucl. Phys. B [**252**]{}, 343 (1985) C. Teitelboim, “Gravitation and Hamiltonian Structure in Two Space-Time Dimensions,” Phys. Lett.  [**126B**]{} (1983) 41. A. Almheiri and J. Polchinski, “Models of AdS$_{2}$ backreaction and holography,” JHEP [**1511**]{} (2015) 014 [[arXiv:1402.6334 \[hep-th\]]{}](http://arxiv.org/abs/1402.6334) K. Jensen, “Chaos in AdS$_2$ Holography,” Phys. Rev. Lett.  [**117**]{} (2016) no.11, 111601 [[arXiv:1605.06098 \[hep-th\]]{}](http://arxiv.org/abs/1605.06098) J. Engelsoy, T. G. Mertens and H. Verlinde, “An investigation of AdS$_{2}$ backreaction and holography,” JHEP [**1607**]{}, 139 (2016) [[arXiv:1606.03438 \[hep-th\]]{}](http://arxiv.org/abs/1606.03438). A. Blommaert, T. G. Mertens and H. Verschelde, “Fine Structure of Jackiw-Teitelboim Quantum Gravity,” [[arXiv:1812.00918 \[hep-th\]]{}](http://arxiv.org/abs/1812.00918). D. Stanford and E. Witten, “Fermionic Localization of the Schwarzian Theory,” JHEP [**1710**]{} (2017) 008 [[arXiv:1703.04612 \[hep-th\]]{}](http://arxiv.org/abs/1703.04612) Y. H. Qi, S. J. Sin and J. Yoon, “Quantum Correction to Chaos in Schwarzian Theory,” JHEP [**1911**]{} (2019) 035 [[arXiv:1906.00996 \[hep-th\]]{}](http://arxiv.org/abs/1906.00996) D. Bagrets, A. Altland and A. Kamenev, “Sachdev-Ye-Kitaev model as Liouville quantum mechanics,” Nucl. Phys. B [**911**]{}, 191 (2016) [[arXiv:1607.00694 \[cond-mat.str-el\]]{}](http://arxiv.org/abs/1607.00694); D. Bagrets, A. Altland and A. Kamenev, “Power-law out of time order correlation functions in the SYK model,” Nucl. Phys. B [**921**]{} (2017) 727 [[arXiv:1702.08902 \[cond-mat.str-el\]]{}](http://arxiv.org/abs/1702.08902). T. G. Mertens, G. J. Turiaci and H. L. Verlinde, “Solving the Schwarzian via the Conformal Bootstrap,” JHEP [**1708**]{} (2017) 136 [[arXiv:1705.08408 \[hep-th\]]{}](http://arxiv.org/abs/1705.08408). T. G. Mertens, “The Schwarzian Theory - Origins,” JHEP [**1805**]{} (2018) 036 [[arXiv:1801.09605 \[hep-th\]]{}](http://arxiv.org/abs/1801.09605). A. Blommaert, T. G. Mertens and H. Verschelde, “The Schwarzian Theory - A Wilson Line Perspective,” JHEP [**1812**]{} (2018) 022 [[arXiv:1806.07765 \[hep-th\]]{}](http://arxiv.org/abs/1806.07765). A. Kitaev and S. J. Suh, “Statistical mechanics of a two-dimensional black hole,” [[arXiv:1808.07032 \[hep-th\]]{}](http://arxiv.org/abs/1808.07032). Z. Yang, “The Quantum Gravity Dynamics of Near Extremal Black Holes,” [[arXiv:1809.08647 \[hep-th\]]{}](http://arxiv.org/abs/1809.08647). L. V. Iliesiu, S. S. Pufu, H. Verlinde and Y. Wang, “An exact quantization of Jackiw-Teitelboim gravity,” JHEP [**1911**]{} (2019) 091 [[arXiv:1905.02726 \[hep-th\]]{}](http://arxiv.org/abs/1905.02726) S. J. Suh, “Dynamics of black holes in Jackiw-Teitelboim gravity,” JHEP **03**, 093 (2020) [[arXiv:1912.00861 \[hep-th\]]{}](http://arxiv.org/abs/1912.00861) M. Spradlin and A. Strominger, “Vacuum states for AdS(2) black holes,” JHEP **11**, 021 (1999) [[arXiv:9904143 \[hep-th\]]{}](http://arxiv.org/abs/9904143) W. G. Unruh, “Notes on black hole evaporation,” Phys. Rev. D [**14**]{} (1976) 870. B. S. DeWitt, “Quantum Gravity: The New Synthesis,” L. Crispino, C.B., A. Higuchi and G. E. Matsas, “The Unruh effect and its applications,” Rev. Mod. Phys. **80** (2008), 787-838 [[arXiv:0710.5373 \[gr-qc\]]{}](http://arxiv.org/abs/0710.5373) W. Donnelly and S. B. Giddings, “Diffeomorphism-invariant observables and their nonlocal algebra,” Phys. Rev. D [**93**]{} (2016) no.2, 024030 Erratum: \[Phys. Rev. D [**94**]{} (2016) no.2, 029903\] [[arXiv:1507.07921 \[hep-th\]]{}](http://arxiv.org/abs/1507.07921) S. B. Giddings and A. Kinsella, “Gauge-invariant observables, gravitational dressings, and holography in AdS,” JHEP [**1811**]{}, 074 (2018) [[arXiv:1802.01602 \[hep-th\]]{}](http://arxiv.org/abs/1802.01602) A. Almheiri, T. Anous and A. Lewkowycz, “Inside out: meet the operators inside the horizon. On bulk reconstruction behind causal horizons,” JHEP [**1801**]{}, 028 (2018) [[arXiv:1707.06622 \[hep-th\]]{}](http://arxiv.org/abs/1707.06622) H. Verlinde, “Poking Holes in AdS/CFT: Bulk Fields from Boundary States,” [[arXiv:1505.05069 \[hep-th\]]{}](http://arxiv.org/abs/1505.05069) A. Lewkowycz, G. J. Turiaci and H. Verlinde, “A CFT Perspective on Gravitational Dressing and Bulk Locality,” JHEP [**1701**]{}, 004 (2017) [[arXiv:1608.08977 \[hep-th\]]{}](http://arxiv.org/abs/1608.08977) M. Miyaji, T. Numasawa, N. Shiba, T. Takayanagi and K. Watanabe, “Continuous Multiscale Entanglement Renormalization Ansatz as Holographic Surface-State Correspondence,” Phys. Rev. Lett.  [**115**]{}, no. 17, 171602 (2015) [[arXiv:1506.01353 \[hep-th\]]{}](http://arxiv.org/abs/1506.01353) Y. Nakayama and H. Ooguri, “Bulk Locality and Boundary Creating Operators,” JHEP [**1510**]{}, 114 (2015) [[arXiv:1507.04130 \[hep-th\]]{}](http://arxiv.org/abs/1507.04130) K. Goto, M. Miyaji and T. Takayanagi, “Causal Evolutions of Bulk Local Excitations from CFT,” JHEP [**1609**]{}, 130 (2016) [[arXiv:1605.02835 \[hep-th\]]{}](http://arxiv.org/abs/1605.02835) H. Chen, A. L. Fitzpatrick, J. Kaplan and D. Li, “The AdS$_{3}$ propagator and the fate of locality,” JHEP [**1804**]{}, 075 (2018) [[arXiv:1712.02351 \[hep-th\]]{}](http://arxiv.org/abs/1712.02351) H. Chen, A. L. Fitzpatrick, J. Kaplan and D. Li, “The Bulk-to-Boundary Propagator in Black Hole Microstate Backgrounds,” JHEP [**1906**]{}, 107 (2019) [[arXiv:1810.02436 \[hep-th\]]{}](http://arxiv.org/abs/1810.02436) N. Engelhardt and G. T. Horowitz, “Towards a Reconstruction of General Bulk Metrics,” Class. Quant. Grav.  [**34**]{}, no. 1, 015004 (2017) [[arXiv:1605.01070 \[hep-th\]]{}](http://arxiv.org/abs/1605.01070) W. Donnelly and S. B. Giddings, “Observables, gravitational dressing, and obstructions to locality and subsystems,” Phys. Rev. D [**94**]{}, no. 10, 104038 (2016) [[arXiv:1607.01025 \[hep-th\]]{}](http://arxiv.org/abs/1607.01025) W. Donnelly and S. B. Giddings, “How is quantum information localized in gravity?,” Phys. Rev. D [**96**]{}, no. 8, 086013 (2017) [[arXiv:1706.03104 \[hep-th\]]{}](http://arxiv.org/abs/1706.03104) S. B. Giddings, “Quantum gravity: a quantum-first approach,” LHEP [**1**]{}, no. 3, 1 (2018) [[arXiv:1805.06900 \[hep-th\]]{}](http://arxiv.org/abs/1805.06900) W. Donnelly and S. B. Giddings, “Gravitational splitting at first order: Quantum information localization in gravity,” Phys. Rev. D [**98**]{}, no. 8, 086006 (2018) [[arXiv:1805.11095 \[hep-th\]]{}](http://arxiv.org/abs/1805.11095) S. B. Giddings, “Gravitational dressing, soft charges, and perturbative gravitational splitting,” Phys. Rev. D [**100**]{}, no. 12, 126001 (2019) [[arXiv:1903.06160 \[hep-th\]]{}](http://arxiv.org/abs/1903.06160) S. Giddings and S. Weinberg, “Gauge-invariant observables in gravity and electromagnetism: black hole backgrounds and null dressings,” [[arXiv:1911.09115 \[hep-th\]]{}](http://arxiv.org/abs/1911.09115) A. Hamilton, D. N. Kabat, G. Lifschytz and D. A. Lowe, “Local bulk operators in AdS/CFT: A Boundary view of horizons and locality,” Phys. Rev. D [**73**]{} (2006) 086003 [[arXiv:0506118 \[hep-th\]]{}](http://arxiv.org/abs/0506118) A. Hamilton, D. N. Kabat, G. Lifschytz and D. A. Lowe, “Holographic representation of local bulk operators,” Phys. Rev. D [**74**]{} (2006) 066009 [[arXiv:0606141 \[hep-th\]]{}](http://arxiv.org/abs/0606141) D. Kabat, G. Lifschytz and D. A. Lowe, “Constructing local bulk observables in interacting AdS/CFT,” Phys. Rev. D [**83**]{} (2011) 106009 [[arXiv:1102.2910 \[hep-th\]]{}](http://arxiv.org/abs/1102.2910) D. Kabat and G. Lifschytz, “Local bulk physics from intersecting modular Hamiltonians,” JHEP [**1706**]{} (2017) 120 [[arXiv:1703.06523 \[hep-th\]]{}](http://arxiv.org/abs/1703.06523) D. A. Lowe and S. Roy, “Holographic description of asymptotically AdS(2) collapse geometries,” Phys. Rev. D [**78**]{} (2008) 124017 [[arXiv:0810.1750 \[hep-th\]]{}](http://arxiv.org/abs/0810.1750) J. Maldacena, A. Milekhin and F. Popov, ‘Traversable wormholes in four dimensions,’’ [[arXiv:1807.04726 \[hep-th\]]{}](http://arxiv.org/abs/1807.04726) J. Maldacena, “Comments on magnetic black holes,” [[arXiv:2004.06084 \[hep-th\]]{}](http://arxiv.org/abs/2004.06084) S. Takagi, “On The Response Of A Rindler Particle Detector. 3.,” Prog. Theor. Phys.  [**74**]{} (1985) 501. doi:10.1143/PTP.74.501 D. Hümmer, E. Martin-Martinez and A. Kempf, “Renormalized Unruh-DeWitt Particle Detector Models for Boson and Fermion Fields,” Phys. Rev. D **93** (2016) no.2, 024019 [[arXiv:1506.02046 \[quant-ph\]]{}](http://arxiv.org/abs/1506.02046) F. Gray and R. B. Mann, “Scalar and Fermionic Unruh Otto engines,” JHEP **11** (2018), 174 [[arXiv:1808.01068 \[hep-th\]]{}](http://arxiv.org/abs/1808.01068) J. Louko and V. Toussaint, “Unruh-DeWitt detector’s response to fermions in flat spacetimes,” Phys. Rev. D [**94**]{} (2016) no.6, 064027 [[arXiv:1608.01002 \[hep-th\]]{}](http://arxiv.org/abs/1608.01002) W. Groenevelt, “The Wilson function transform," [[arXiv:0306424 \[math.CA\]]{}](https://arxiv.org/abs/math/0306424); “Wilson function transforms related to Racah coefficients," [[arXiv:math/0501511 \[math.CA\]]{}](https://arxiv.org/abs/math/0501511). A. Fabbri, J. Navarro-Salas and G. J. Olmo, “Particles and energy fluxes from a CFT perspective,” Phys. Rev. D [**70**]{} (2004) 064022 [[arXiv:0403021 \[hep-th\]]{}](https://arxiv.org/abs/hep-th/0403021). T. G. Mertens and G. J. Turiaci, “Defects in Jackiw-Teitelboim Quantum Gravity,” JHEP [**1908**]{} (2019) 127 [[arXiv:1904.05228 \[hep-th\]]{}](http://arxiv.org/abs/1904.05228) T. G. Mertens, “To appear” W. Donnelly and A. C. Wall, “Entanglement entropy of electromagnetic edge modes,” Phys. Rev. Lett.  [**114**]{} (2015) no.11, 111603 [[arXiv:1412.1895 \[hep-th\]]{}](http://arxiv.org/abs/1412.1895) W. Donnelly and A. C. Wall, “Geometric entropy and edge modes of the electromagnetic field,” Phys. Rev. D [**94**]{} (2016) no.10, 104053 [[arXiv:1506.05792 \[hep-th\]]{}](http://arxiv.org/abs/1506.05792) A. Blommaert, T. G. Mertens, H. Verschelde and V. I. Zakharov, “Edge State Quantization: Vector Fields in Rindler,” JHEP [**1808**]{} (2018) 196 [[arXiv:1801.09910 \[hep-th\]]{}](http://arxiv.org/abs/1801.09910) A. Blommaert, T. G. Mertens and H. Verschelde, “Edge dynamics from the path integral - Maxwell and Yang-Mills,” JHEP [**1811**]{} (2018) 080 [[arXiv:1804.07585 \[hep-th\]]{}](http://arxiv.org/abs/1804.07585) [^1]: For a system with internal symmetries level repulsion or random matrix statistics emerges within subsets of energy levels where each subset has fixed values of quantum numbers related to these internal symmetries. See for example [@mehta; @haake; @Kapec:2019ecr]. [^2]: Gravitational shockwaves in this model have been studied in [@malstanyang; @shockwaves]. [^3]: For example in order to probe scrambling in de Sitter one is led to investigate shockwaves and out-of-time-ordered correlators in the bulk [@aalsmashiu; @wopjordan]. In this setup we do not have access to intuition from a unitary dual theory [@cotleremergent]. [^4]: We will focus on such systems here. The discussion is immediately modified to other ensembles [@Stanford:2019vob]. [^5]: There is furthermore a contact term contribution $\rho(E_1)\delta(E_1-E_2)$ which is largely irrelevant to our story here. We define $${\text{sinc}}\, x =\frac{\sin x}{x}.$$ [^6]: The Gaussian orthogonal ensemble (GOE) and Gaussian symplectic ensemble (GSE) have linear respectively quartic level repulsion. [^7]: One recognizes the Bose-Einstein distribution. The factor $\omega^{d-1}$ is proportional to the level density for a particle in a $d-1$ dimensional box. This formula should furthermore be augmented with suitable greybody factors associated with the centrifugal barrier and with interference terms associated with reflections off of asymptotic boundaries. We leave them implicit here. [^8]: The middle formula is the unique rate which probes the two level spectral density for transitions of energy $\omega$ close to a heavy black hole of mass $E$ and which matches with the semiclassical answer for $\omega\gg 1/\rho(E)$. [^9]: We work in units where the AdS length $L=1$. [^10]: For example if we want to interpret JT gravity as the low-energy limit of a 4d near-extremal black hole, we have $S_0 \sim \frac{Q}{G}$ where $Q$ is the charge of the extremal black hole. [^11]: On-shell, the equations of motion link the dilaton field to the matter content in the theory, and these constraints are then sufficient to find the Schwarzian equation of motion [@ads2]. [^12]: It is conventional to choose units such that $C=1/2$. [^13]: For earlier discussion in this context see for example [@Blommaert:2019hjr; @Mertens:2019bvy]. [^14]: The Fourier transform appears due to $$\bra{\omega}\mu(t)\ket{0}=e^{i\omega t}\bra{\omega}\mu(0)\ket{0}.$$ [^15]: Infinitesimally separated bulk points along such a worldline are separated by $$ds^2 = -\frac{F'(t+z)F'(t-z)}{(F(t+z)-F(t-z))^2}dt^2$$ and since $F' \geq 0$ are hence for any off-shell $F$ time-like separated, proving that the resulting trajectory is always timelike. It is possible for the trajectory to become lightlike at points where time stops flowing $F'=0$. [^16]: There is an implicit $i\varepsilon$ in the exponential $t_1-t_2-i\varepsilon$ as a Euclidean damping factor [@schwarzian]. This is related to the particular ordering of both operators in the Wightman two-point function. We will be more explicit about this in section \[s:hb\] where both operator orderings are relevant. [^17]: One should multiply all signs. [^18]: Note that for $z\ll 1$ this quantity goes to zero like $z^2$ as demanded by the extrapolate dictionary for a massless scalar field. [^19]: Here the factor $V$ is the total volume outside the semiclassical black hole horizon and is an artifact of the matter theory. [^20]: The integration contour is chosen as in [@sss2]. It follows the positive real axis. However in the classically forbidden region $E<0$ one needs to choose an appropriate contour for convergence. The contributions to observables from the forbidden part of the integration contour is highly suppressed in all cases by powers of $e^{-S_0}$. Therefor we can essentially neglect these contributions. [^21]: We note [@phil] that is conform the idea that this expression represents the ensemble average over different Hamiltonians $H$ of the two point function in a discrete quantum chaotic system. The idea is to take such a set of discrete quantum chaotic systems and to first ensemble average over unitaries $U$ which diagonalize $H$. One then invokes a version of the eigenvalue thermalization hypothesis $$\sum_{a,b}{\mathcal{O}}_{a}{\mathcal{O}}_{b}\int d U\,U_{i a}\,U_{j a}^*\,U_{j b}\,U_{i b}^* = \rvert{\mathcal{O}}_{E_i E_j}\rvert^2.$$ The assumption is that one point functions in the averaged theory vanish (which they do in JT gravity). Furthermore $\rvert{\mathcal{O}}_{E_i E_j}\rvert^2$ are smooth functions on energy scales of order the typical level spacing $e^{-S_0}$. The result is . Furthermore ensemble averaging over the eigenvalues of the Hamiltonians (with a well chosen potential $V(H)$ for Hamiltonians $H$ in the ensemble average) one then indeed finds . [^22]: Here $\rho(\lambda_i)$ is to be understood as a function that varies only on scales much larger than $e^{-S_0}$. [^23]: We use the Dirac algebra convention where $\gamma_0= i \sigma_1$ and $\gamma_1 = \sigma_3$, in terms of the Pauli $\sigma$-matrices. [^24]: Variation of the massless Dirac action gives a boundary condition: $$\left. \bar{\psi} \gamma^1 \psi \right\rvert_\partial = 0 \, .$$ This equation holds also in curved spacetime as one checks that all Weyl scaling factors end up cancelling. In terms of spinor components $\psi_\alpha, \, \alpha=1,2$, this becomes: $$\psi_1^*\, \psi_2\rvert_\partial = \psi_1 \,\psi_2^*\rvert_\partial \, .$$ The Dirichlet boundary discussed above corresponds to: $$\psi_1\rvert_\partial =\psi_2\rvert_\partial \, .$$ This results eventually in the $4\sin^2\,\omega z$ greybody interference factor in . Other possibilities include for example setting $\psi_1=0$ or $\psi_2=0$. This results instead in a $4 \cos^2\, \omega z$ interference factor. See also appendix \[app:obc\]. [^25]: This follows from a generalization of the Barnes identities and is identical to the orthonormality relation of the zeroth Wilson polynomial [@groenevelt]. In particular the identity is: $$\begin{aligned} \frac{1}{4\pi i}\int_{-\infty}^{+\infty}ds \frac{\Gamma(a \pm s) \Gamma(b \pm s) \Gamma( c \pm s) \Gamma (d\pm s)}{\Gamma(\pm 2s)} = \frac{\Gamma(a+b)\Gamma(a+c)\Gamma(a+d)\Gamma(b+c)\Gamma(b+d)\Gamma(c+d)}{\Gamma(a+b+c+d)}. \end{aligned}$$ This identity holds when $\Re a,b,c,d >0$. We checked it numerically as well. [^26]: We have included the contribution from the $v$-lightcone component by using the doubling trick. [^27]: This is directly found by using the integral representation of the step function. See also [@Fabbri:2004yy] for the semi-classical versions of these equations. [^28]: The validity of such a series expansion within Schwarzian correlators is not entirely straightforward. We refer to [@schwarzian; @Mertens:2019tcm] for some comments and to [@toth] for a thorough analysis. [^29]: Unlike the semi-classical answer , the Schwarzian answers and its complex conjugate differ by more than an infinitesimal term. [^30]: Other options exist, and will be explored elsewhere. [^31]: Actually, for all intents and purposes this suppression is exact. The sine kernel and contact term contributions in cancel perfectly upon integration (by construction). Furthermore there are only quick wiggles but their integral gives a suppressed effect as well. [^32]: The kernel is accurate as long as both $E, M\gg e^{-2S_0/3}$. [^33]: This is true in a precise sense for the two-point function considered in this work. However it is a slight oversimplification for higher point functions. For example in an out-of-time-ordered disk four-point functions, two bilocals cross. This crossing can be avoided on higher genus surfaces if one bilocal traverses a handle whilst the other travels underneath it. In the end however such contributions can also be written in terms of multi-level correlators $\rho(E_1\dots E_n)$ [@wophilbert]. [^34]: For a recent derivation of this kernel from a brane correlator computation, see appendix A.4 of [@paper6].
{ "pile_set_name": "ArXiv" }
--- abstract: 'The understanding of the basic properties of the ultra - high energy extensive air showers is strongly dependent on the description of the hadronic interactions in a energy range beyond that probed by the LHC. One of the uncertainties present in the modeling of the air showers is the treatment of diffractive interactions, which are dominated by non - perturbative physics and usually described by phenomenological models. These interactions are expect to affect the development of the air showers, since they provide a way of transporting substantial amounts of energy deep in the atmosphere, modifying the global characteristics of the shower profile. In this paper we investigate the impact of the diffractive interactions in the observables that can be measured in hadronic collisions at high energies and ultra - high energy cosmic ray interactions. We consider three distinct phenomenological models for the treatment of diffractive physics and estimate the influence of these interactions on the elasticity, number of secondaries, longitudinal air shower profiles and muon densities for proton - air and iron - air collisions at different primary energies. Our results demonstrate that the diffractive events has a non - negligible effect on the observables and that the distinct approaches for these interactions, present in the phenomenological models, are an important source of theoretical uncertainty for the description of the extensive air showers.' author: - 'L.B. Arbeletche, V.P. Gonçalves and M.A. Müller' title: 'Investigating the influence of diffractive interactions on ultra - high energy extensive air showers' --- Introduction {#intro} ============ Understanding the behaviour of high energy hadron reactions from a fundamental perspective within Quantum Chromodynamics (QCD) is an important goal of particle physics. One of the main open questions is the treatment of the diffractive processes, which are characterized by the presence of an intact hadron and large rapidity gaps in the final state (For a recent review see Ref. [@forward]). These processes are in general described in terms of a color singlet object: the Pomeron ($\pom$). This object, with the vacuum quantum numbers, was introduced phenomenologically in the Regge theory as a simple moving pole in the complex angular momentum plane, to describe the high-energy behaviour of the total and elastic cross-sections of the hadronic reactions [@collins]. The diffractive events are dominated by low transverse momentum processes, i.e. processes in which the strong running coupling constant is large and the useful perturbative methods are not valid. These processes are in general described by phenomenological models based on first principles of the Quantum Field Theory and basic properties of QCD [@kaidalov]. It implies a large theoretical uncertainty, with a strong impact on the predictions for the magnitude and energy dependence of the diffractive cross section. For example, while some models [@qgsjet] predict that $\sigma_{diff} \propto \ln^2 s$ at asymptotic energies, other models [@sibyll] predict $\sigma_{diff} \propto \ln \, s$. As a consequence, the contribution of the diffractive events for the total cross sections at very high energies still is an open question. Recent experimental results from the Run 1 of the LHC have shed some light on the energy behaviour of the single and double diffractive cross sections [@lhcdata] and more precise results are expected in the Run 2 [@forward]. The expectation is that these data could be used to constrain the basic assumptions present in the phenomenological models, decreasing the uncertainty in its predictions at larger energies. The understanding of the hadronic interactions also is fundamental for the description of the ultra - high energy cosmic ray (UHECR) air showers [@annual; @lipari], with the reconstruction of the primary UHECR properties being strongly dependent on the treatment of the diffractive and non-diffractive events present in the hadron – air interactions. Due to the dominance of soft physics, the models of hadronic interactions in the generator models used in the simulation of extensive air showers (EAS) are largely phenomenological and have to be extrapolated from accelerator energies, where they are calibrated, to the UHECR energies. In recent years, the underlying theoretical framework present in these generators have been improved and the LHC data have been used to tuning the basic cross sections [@qgsjet; @sibyll23; @epos]. However, the available collider data do not cover the full kinematic region of interest in UHECR interactions [@ralph_epj]. In particular, experimental data for the particle production at very forward rapidities still are scarce, with its theoretical description still being an open question (For a recent study see, e.g. Ref. [@werner]). As a consequence of the theoretical uncertainty present in the description of the hadronic interactions present in the EAS, the air shower simulations still are one of the main source of systematic uncertainty in the interpretation of cosmic ray data [@david; @parsons; @auger]. Our main goal in this paper is to give a quantitative estimate of the uncertainty associated to the treatment of the diffractive interactions on the shower observables. These interactions are expect to affect the development of the air showers, since diffractive interactions have direct impact on the inelasticity – the relative energy loss of the leading secondary particle – providing a way of transporting substantial amounts of energy deep in the atmosphere. Consequently, they modify the global characteristics of the shower profile. In particular, a higher diffraction rate implies a slower EAS development, modifying the position of the shower maximum $\langle X_{max}\rangle$, and a smaller number of secondaries in each interaction. In what follows we will study the impact of the diffractive interactions on ultra - high energy extensive air showers through the comparative analysis of the predictions for the EAS observables from the standard versions of the Sibyll [@sibyll], QGSJET [@qgsjet] and EPOS [@epos] hadronic models, which are available in the framework of the CORSIKA air shower simulation package [@corsika]. Our study is strongly motivated by the analysis performed thirteen years ago in Ref. [@garcia1], where the authors have studied the same problem using the pre - LHC hadronic models and the AIRES program [@aires] for the simulation of the EAS development. In what follows we will update that analysis by considering the current hadronic models and include new results for Fe-Air collisions which were not considered in that reference. In order to investigate the contribution of these interactions, we will present, for each hadronic model, a comparison between predictions obtained using the full simulations, i.e. including the non - diffractive and diffractive events, with those derived excluding the diffractive events of the EAS development. In particular, in Section \[hadron\] we will compare the predictions of the different hadronic models for the distribution of number of secondaries, fraction of diffractive events, average fraction of pions and elasticity distributions, considering individual $p$-Air and $Fe$-Air collisions at different values for the primary energy. The impact on the EAS observables is studied in Section \[shower\], where we will analyse the impact of the diffractive interactions on the longitudinal profiles of charged particles and muons as well as on the position of the shower maximum. Finally, in Section \[conc\] we summarize our main conclusions. Impact of the diffractive interactions in hadronic collisions {#hadron} ============================================================= As discussed in the Introduction, the treatment of the hadronic interactions in the air shower simulation codes is based on phenomenological models. In what follows we will consider the Sibyll 2.1, QGSJET – II 04 and EPOS LHC models, present in the CORSIKA package, and compare its predictions for distinct observables in $p$-Air and $Fe$-Air collisions at different center-of-mass energies. Before to present our results, lets present a brief review of the main assumptions of the different phenomenological models (For a more detailed review see, e.g., Ref. [@david]). Most current hadronic interaction models are based on basic quantum field theory principles, such as unitarity and analyticity of scattering amplitudes, and use Gribov - Regge theory [@gribov] of multi - Pomeron exchange between nucleons as the basis for the treatment of high energy, soft interactions. Perturbative QCD is considered to describe hard interactions with high transverse momentum, which becomes important at high energies. In general, simplifications are made in the implementation of the hard processes if they are not directly relevant to the production of high energy secondaries. QGSJET and Sibyll consider the eikonal model and assume unitarized cross sections with the real eikonal function being given by a sum of soft and hard contributions. EPOS also is based on Gribov - Regge theory and provides a energy - conserving quantum mechanical multiple scattering approach in terms of parton ladders. At high energies, nonlinear effects associated to the high partonic density becomes important and should be taken into account [@glr; @hdqcd]. QGSJET provides a microscopic treatment of nonlinear interaction effects in hadronic and nuclear collisions in terms of Pomeron - Pomeron interaction diagrams. On the other hand, in Sibyll these effects are modelled by means of an energy - dependent cutoff for the minijet production. In contrast, EPOS employs a phenomenological approach for nonlinear interaction effects and address the energy - momentum correlations between multiple scattering processes at the amplitude level. The generalization from $pp$ to $pA$ and $AA$ collisions is usually performed via Glauber - Gribov formalism [@glauber; @gribov2], taking into account inelastic screening and low mass diffraction effects. Finally, diffractive interactions are treated differently in the distinct models. In Sibyll and QGSJET models, the diffraction dissociation is described in terms of the Good - Walker formalism [@gw], where the colliding hadrons are represented by superpositions of elastic scattering eigenstates which undergo different absorption during the collision. In Sibyll, high mass diffraction is described in terms of a 2 - channel eikonal approach. On the other hand, the description of this process in QGSJET is based on all - order resummation of cut enhanced $\pom$ - diagrams. In contrast, a particular kind of Pomeron is used to define a diffractive event in EPOS. Depending of each event configuration it can be classified as non - diffractive, low mass diffraction without central particle production, or high mass diffraction. A basic feature of the diffractive interactions is that there is a leading particle whose energy is much larger than the energies of the other particles and the total number of secondaries is generally small. The resulting final states have a high elasticity. In previous studies, these characteristics were used to tag the diffractive events [@garcia1; @garcia2]. In contrast, in our study the classification of the diffractive and non - diffractive events was made using internal variables of the respective hadronic generators. In particular, in the EPOS LHC and QGSJET-II 04 generators, the events were classified according to the variable [typevt]{}, which define the collision type: if [typevt = 1]{}, the event is classified as non-diffractive; otherwise, the event is considered a diffractive one. As this variable is read after the event is generated, it allows to reject diffractive events when generating samples of non-diffractive events. In contrast, in Sibyll 2.1, a hadron-air interaction is classified as diffractive requiring that there is only one wounded nucleon in the target ([NW = 1]{}) and that the interaction with this nucleon is diffractive ([JDIFF(0) &gt; 0]{}). Again, variables are read after the event have been generated. For nucleus-nucleus interactions, we implemented a distinct scheme for this model. As Sibyll 2.1 treats nucleus-nucleus collisions in a semisuperposition model, where the *A* projectile nucleons are considered independent particles, it generates *A* superposed interactions. We found reasonable to classify as diffractive the collisions whose all superposed interactions were diffractive. Otherwise, if at least one interaction is non-diffractive, the event is labelled as non-diffractive. ![Predictions of the different hadronic generators for diffractive and non - diffractive events in the elasticity – multiplicity ($f_L \times N_{sec}$) plane considering $10^{4}$ $p$ - Air collisions at $\sqrt{s} = 10$ TeV. Diffractive events are represented by red squares and the non-diffractive one by light blue circles.[]{data-label="tagging"}](qgs_0.eps "fig:"){width="5"} ![Predictions of the different hadronic generators for diffractive and non - diffractive events in the elasticity – multiplicity ($f_L \times N_{sec}$) plane considering $10^{4}$ $p$ - Air collisions at $\sqrt{s} = 10$ TeV. Diffractive events are represented by red squares and the non-diffractive one by light blue circles.[]{data-label="tagging"}](epos_0.eps "fig:"){width="5"} ![Predictions of the different hadronic generators for diffractive and non - diffractive events in the elasticity – multiplicity ($f_L \times N_{sec}$) plane considering $10^{4}$ $p$ - Air collisions at $\sqrt{s} = 10$ TeV. Diffractive events are represented by red squares and the non-diffractive one by light blue circles.[]{data-label="tagging"}](sib_0.eps "fig:"){width="5"} In order to illustrate the classification between diffractive and non - diffractive events, in Fig. \[tagging\] we present our results for the elasticity-multiplicity plane considering $10^{4}$ $p$ - Air collisions at $\sqrt{s} = 10$ TeV. The elasticity is characterized by the leading energy fraction $f_L$, which is defined by the ratio between the energy of the secondary with maximum energy $E_{lead}$ (leading particle) and the primary energy $E_{prim}$, i. e. $f_L \equiv E_{lead}/E_{prim}$. The predictions of the three different hadronic generators are presented separately, with the diffractive events being represented by red squares and the non-diffractive one by light blue circles. We have that the most of the diffractive events appear on the region of low multiplicity and elasticity close to 1, in agreement with the theoretical expectation. However, the distribution is different for the distinct hadronic models. In what follows, we will analyse the phenomenological implications of these differences. We have generated samples of $10^{4}$ $p$ - Air and $Fe$ - Air collisions at different center of mass energies for each event generator model: QGSJET-II 04, EPOS LHC and Sibyll 2.1. We assume that the target is a mixture of nitrogen, oxygen and argon, exactly as it is implemented in the CORSIKA package [@corsika]. After the events have been generated, we register all secondaries, imposing a cutoff on the kinetic energy of the secondary particle. Basically, we excluded from the secondary list all particles with kinetic energy smaller than 40 MeV. It is assumed to eliminate from the secondary list, the fragments of the target nucleus, which are included in the EPOS LHC case, but not in the others models. Such cutoff is expected to have no influence in the air shower development [@garcia2], since particles with such low values of kinetic energy should not be able to propagate in the atmosphere. [ As]{} we are interested in the particular case of diffractive collisions, we generated separately samples without diffractive events and full samples including non - diffractive and diffractive events. In what follows we will compare the predictions from these two configurations, which allow us to estimate the impact of diffraction on the average properties of the hadronic collisions as well as to investigate the theoretical uncertainty associated to the distinct treatment of diffraction present in the hadronic generators considered in our study. ![Comparison between the Sybill, QGSJET and EPOS LHC predictions for the number of secondaries considering $p$-Air (upper panels) and $Fe$-Air (lower panels) collisions at three different values for the primary energy.[]{data-label="p_nsec_dist"}](p_nsec_dist.eps "fig:"){width="\textwidth"}\ ![Comparison between the Sybill, QGSJET and EPOS LHC predictions for the number of secondaries considering $p$-Air (upper panels) and $Fe$-Air (lower panels) collisions at three different values for the primary energy.[]{data-label="p_nsec_dist"}](Fe_nsec_dist.eps "fig:"){width="\textwidth"} Initially, in Fig. \[p\_nsec\_dist\] (upper panels) we present the predictions for the distributions of number of secondaries $N_{sec}$ produced in $p$-Air collisions considering three different values for the primary energy: $E_{prim} =$ 100 GeV, 1 TeV and 100 PeV. We have that for $E_{prim} =$ 100 GeV, the distinct predictions are similar. One the other hand, for $E_{prim} =$ 1 TeV, these distributions are different, with the presence of two peaks: one at small values of $N_{sec}$, associated to diffractive events and other at larger values of $N_{sec}$, related to non - diffractive events. Finally, for $E_{prim} =$ 100 PeV ($\sqrt{s_{NN}} \approx 14$ TeV) the distributions extend up to thousands of secondary particles, and the peaks become more evident. In the case of Sibyll 2.1, the peak at low $N_{sec}$ is suppressed. In Fig. \[p\_nsec\_dist\] (lower panels) we shown the $N_{sec}$ distributions for [ $Fe$-Air]{} collisions. In this case we have that the predictions of the different models for the peak at low - $N_{sec}$ are very distinct, being enhanced in the Sibyll case and suppressed for both QGSJET-II 04 and EPOS LHC with the increasing of the primary energy. ![Predictions of the different hadronic generators for the primary energy dependence of the average number of secondaries produced in [ $p$-Air]{} (left panel) and [ $Fe$-Air]{} (right panel) collisions. The solid (open) symbols represent simulations including diffractive interactions (non-diffractive events only).[]{data-label="nsec_media"}](p_nsec_media.eps "fig:"){width="49.00000%"} ![Predictions of the different hadronic generators for the primary energy dependence of the average number of secondaries produced in [ $p$-Air]{} (left panel) and [ $Fe$-Air]{} (right panel) collisions. The solid (open) symbols represent simulations including diffractive interactions (non-diffractive events only).[]{data-label="nsec_media"}](Fe_nsec_media.eps "fig:"){width="49.00000%"} The predictions of the different hadronic generators for the dependence of the average number of secondaries produced in [ $p$-Air]{} collisions on the primary energy are presented in Fig. \[nsec\_media\] (left panel). The full simulations, including diffractive and non - diffractive events, are represented by solid symbols. On the other hand, the simulations including only non - diffractive events are represented by open symbols. Initially, lets compare the predictions of the different hadronic generators for the full simulations. In this case we have that for $E_{prim} \approx 10$ PeV, the EPOS LHC and QGSJET-II 04 generators predict, on average, very similar values for the number of secondaries. As the energy increases, differences become significant, with the QGSJET-II 04 predicting the largest number of secondaries at the highest energies, while Sibyll 2.1 predicts the smaller value. One have that the relative difference between the hadronic generators is smaller after the LHC tuning. In particular, the QGSJET prediction for the number of secondaries produced at ultra - high energies in [ $p$-Air]{} collisions was substantially reduced after the LHC tune. Our predictions for $Fe$ - Air collisions are presented in Fig. \[nsec\_media\] (right panel). In this case, Sibyll 2.1 predicts the largest number of secondaries for primary energies smaller than 10 EeV. At higher energies, the QGSJET-II 04 predictions are slightly larger than Sibyll 2.1, with the EPOS LHC predicting the lowest number of secondaries in the whole considered energy range. As already observed in $p$-Air collisions, the QGSJET-II 04 and EPOS LHC results are very similar at energies lower than 10 PeV and become very distinct at higher energies. Lets now analyse the impact of the diffractive events on the average number of secondaries. As can be observed in Fig. \[nsec\_media\] (left panel) by the comparison between the solid and dashed lines, one have that for [ $p$-Air]{} collisions the contribution of the diffractive events is small for the Sibyll 2.1. One the other hand, these events have a non - negligible impact on the QGSJET-II 04 and EPOS LHC predictions, with the contribution of diffraction increasing with the primary energy. In contrast, in the case of the predictions for [ $Fe$-Air]{} collisions, presented in Fig. \[nsec\_media\] (right panel), Sibyll 2.1 receives the most significant influence of diffractive interactions on the mean secondary multiplicity, while QGSJET-II 04 one is almost not influenced by diffraction. One have that in $p$-Air and $Fe$-Air collisions, the presence of the diffractive interactions reduces the average number of secondaries produced in the collisions. It is expected, since diffractive interactions produce less secondaries than non-diffractive collisions. Consequently, it is expected that the sample of collisions without diffraction has more secondaries on the average. ![Fraction of diffractive events in [ $p$-Air]{} (left panel) and [ $Fe$-Air]{} (right panel) collisions.[]{data-label="fracdiff"}](p_fracdiff.eps "fig:"){width="49.00000%"} ![Fraction of diffractive events in [ $p$-Air]{} (left panel) and [ $Fe$-Air]{} (right panel) collisions.[]{data-label="fracdiff"}](Fe_fracdiff.eps "fig:"){width="49.00000%"} In order to estimate how the contribution of the diffractive events change with the energy, lets consider the relative probability of diffraction, which is related to the ratio of the diffractive to the total cross sections and determines how many diffractive events are expected in a sample of collisions. The relative probabilities of diffraction predicted by the different hadronic generators for [$p$-Air]{} and [ $Fe$-Air]{} collisions at different primary energies are presented in Fig. \[fracdiff\]. For [ $p$-Air]{} collisions (left panel), the EPOS LHC predicts the largest fraction of diffractive events in the whole energy range, being $\approx$ 40% at 100 GeV and 18% for 100 EeV. On the other hand, Sibyll 2.1 predicts that the contribution of the diffractive events is of the order of 20% at 100 GeV, decreasing at larger energies. In particular, Sibyll 2.1 predicts the smaller contribution of these events at ultra - high energies. Finally, QGSJET-II 04 shows a different behaviour, increasing from 10% at 100 GeV to 17% at 100 TeV and then decreasing for 12% at 100 EeV. The results for [ $Fe$-Air]{} collisions are presented in Fig. \[fracdiff\] (right panel). As observed in $p$-Air collisions, EPOS LHC predicts the largest fraction of diffractive events at all primary energies. However, the magnitude of this contribution is smaller, with values around 13% in the energy range considered. The Sibyll 2.1 and QGSJET-II 04 predictions also are smaller in comparison to the $p$-Air collisions, with the Sibyll 2.1 one increasing with the energy and predicting similar values to the EPOS LHC at ultra - high energies. ![Energy dependence of the average fraction of pions produced in [ $p$-Air]{} (left panel) and [$Fe$-Air]{} (right panel) collisions.[]{data-label="fracpi"}](p_fracpi.eps "fig:"){width="49.00000%"} ![Energy dependence of the average fraction of pions produced in [ $p$-Air]{} (left panel) and [$Fe$-Air]{} (right panel) collisions.[]{data-label="fracpi"}](Fe_fracpi.eps "fig:"){width="49.00000%"} In Fig. \[fracpi\] we present the predictions of the different hadronic generators for the energy dependence of the fraction of pions produced in [$p$-Air]{} (left panel) and [$Fe$-Air]{} (right panel) collisions, which is given by the ratio between the number of pions produced and the total number of secondaries. We have that EPOS LHC predicts the largest fraction of pions for all primary energies, with the diffractive events decreasing the fraction. On the other hand, as expected from the Fig. \[fracdiff\], the fraction of pions predicted by the Sibyll 2.1 is almost not influenced by the diffractive events. The predictions for [$Fe$-Air]{} collisions are presented in Fig. \[fracpi\] (right panel). In this case, we have that the fraction of pions predicted by the three hadronic generators is smaller than in [ $p$-Air]{} collisions and the QGSJET-II 04 predicts the largest fraction in the energy range considered. Moreover, we have that the Sibyll 2.1 prediction at large energies is strongly influenced by the diffractive events. ![Predictions of the hadronic generators for the elasticity distribution considering [$p$-Air]{} collisions at three representative primary energies.[]{data-label="p_elast_dist"}](p_elast_dist.eps){width="59.00000%"} ![Energy dependence of the average elasticity considering [ $p$-Air]{} collisions.[]{data-label="elast_media"}](p_elast_media.eps){width="49.00000%"} In Fig. \[p\_elast\_dist\] we present the predictions for the elasticity distribution considering [ $p$-Air]{} collisions and three representative primary energies: 1 TeV, 100 PeV and 100 EeV. The large $f_L$ region is detailed in the small plots on the right. We can verify the presence of [a]{} peak for $f_L \approx 1$, which is related to the diffractive interactions, where the leading particle carry most of the primary energy. At small energies ($E_{prim} = 1$ TeV), the predictions of the different models are similar, except by the fact that the diffractive peak predicted by the QGSJET - II 04 is smaller in comparison to the other models. As the energy increases, the distributions starts to be different, with the QGSJET - II 04 predicting the largest fraction of non - diffractive events at small - $f_L$, which is associated to the fact that this model predicts the largest number of secondaries at high energies \[See Fig. \[nsec\_media\] (left panel)\]. We have that the EPOS LHC and QGSJET - II 04 predict the increasing of the peak for $f_L \approx 1$. In contrast, Sibyll 2.1 predicts that it becomes smaller at large energies, in agreement with the results presented in Fig. \[fracdiff\]. Finally, in order to investigate the influence of the diffractive interactions in the average elasticity $\langle f_L \rangle$ [in $p$-Air collisions]{}, in Fig. \[elast\_media\] we present the predictions for the energy dependence of $\langle f_L \rangle$. We have that the presence of the diffractive events implies a higher average elasticity, since these events populate the region of large $\langle f_L \rangle$. Moreover, the largest impact is in the EPOS LHC predictions, which is associated to the fact that this model predicts the largest peak for $f_L \approx 1$. Our results for $p$-Air and $Fe$-Air collisions demonstrated that the diffractive interactions modify the magnitude of the number of secondaries and the inelasticy of these collisions. Moreover, our results indicated that the distinct treatment of these interactions, present in the different hadronic generators, implies a non - negligible theoretical uncertainty in the predictions. In the next Section, we will expand our analysis for air shower observables. Impact of the diffractive interactions on Shower Observables {#shower} ============================================================ In order to investigate the impact of the diffractive interactions on the air shower observables we will use the software CORSIKA (Version 7.4.005) to simulate air showers generated by primary protons and iron nuclei with energies of $10^{17}$ and $10^{20}$ eV, reaching the atmosphere with a zenith angle of 60º. The simulations have been performed considering two distinct configurations: (a) full simulations, which include diffractive and non - diffractive interactions in the shower development, and (b) non - diffractive (ND) simulations generated removing the diffractive interactions of the shower development. For the description of the hadronic interactions we will consider the same models discussed in the previous Section. ![Average longitudinal profiles of charged particles for showers generated by protons and iron nuclei with primary energies of $10^{17}$ eV (upper panels) and $10^{20}$ eV (lower panels). Solid (dashed) lines represent the full (non - diffractive) simulations.[]{data-label="long_ch_17"}](long_ch_17.eps "fig:"){width="\textwidth"}\ ![Average longitudinal profiles of charged particles for showers generated by protons and iron nuclei with primary energies of $10^{17}$ eV (upper panels) and $10^{20}$ eV (lower panels). Solid (dashed) lines represent the full (non - diffractive) simulations.[]{data-label="long_ch_17"}](long_ch_20.eps "fig:"){width="\textwidth"} In Fig. \[long\_ch\_17\] we present the predictions for the mean longitudinal profiles of charged particles for showers generated by protons and iron nuclei at $10^{17}$ eV (upper panels) and $10^{20}$ eV (lower panels). The solid lines represent showers generated including the non - diffractive and diffractive interactions (full simulations), while the dashed lines represent showers where the diffractive interactions were removed of the shower development (non - diffractive simulations). The impact of the diffractive interactions is very clear: the presence of the diffractive events implies that the average number of particles is smaller and the maximum is shifted to higher atmospheric depths, i.e. the air showers develop slower in the atmosphere, in agreement with the results obtained in Ref. [@garcia1]. Moreover, we have that the Sibyll 2.1 predictions are almost insensitive to the inclusion of the diffractive interactions, which agrees with the fact that the contribution of these interactions at high energies is small in this model. $E_{prim}$ Primary QGSJET-II 04 EPOS LHC Sibyll 2.1 ------------------- --------- --------------- --------------- --------------- [ $10^{17}$ eV]{} p 24.5 $g/cm^2$ 23.3 $g/cm^2$ 8.3 $g/cm^2$ Fe 17.0 $g/cm^2$ 20.0 $g/cm^2$ 11.2 $g/cm^2$ [ $10^{20}$ eV]{} p 19.6 $g/cm^2$ 26.0 $g/cm^2$ 10.9 $g/cm^2$ Fe 16.5 $g/cm^2$ 22.5 $g/cm^2$ 8.3 $g/cm^2$ : Predictions for the shift in $\langle X_{max} \rangle$ due to the presence of the diffractive events.[]{data-label="xmax_shift_17"} ![Depth of maximum $\langle X_{max} \rangle$ as a function of the primary cosmic ray energy. Solid (dashed) are from simulations including (removing) the diffractive interactions of the air shower development. []{data-label="xmax"}](xmax_qgsii.eps "fig:"){width="30.00000%"} ![Depth of maximum $\langle X_{max} \rangle$ as a function of the primary cosmic ray energy. Solid (dashed) are from simulations including (removing) the diffractive interactions of the air shower development. []{data-label="xmax"}](xmax_epos.eps "fig:"){width="30.00000%"} ![Depth of maximum $\langle X_{max} \rangle$ as a function of the primary cosmic ray energy. Solid (dashed) are from simulations including (removing) the diffractive interactions of the air shower development. []{data-label="xmax"}](xmax_sib21.eps "fig:"){width="30.00000%"} The influence of the diffractive interactions also can be estimated by the analysis of the shift in the depth of maximum $\langle X_{max} \rangle$, given by $\langle X_{max}^{(full)} \rangle - \langle X_{max}^{(ND)} \rangle$. In Table \[xmax\_shift\_17\] we present our results for this quantity. We can see that the influence of the diffractive interactions on the profiles is smaller in the case of Sibyll 2.1, for both primaries. This is related to the fact that Sibyll 2.1 usually produces less diffractive events (See Fig. \[fracdiff\]) in hadron-nucleus collisions than the other models. Additionally, we have that the showers generated using the EPOS LHC are more influenced by diffraction than those generated by the other models. This also can be verified in Fig. \[xmax\], where the energy dependence of $\langle X_{max} \rangle$ is presented. Moreover, we have that the impact of the diffractive interactions on this observable is almost energy-independent for the three considered models of hadronic interactions. We have that the impact of the diffractive interactions in the shower maximum position depends of the treatment of the diffractive physics and it is of the order of the typical experimental precision for the $\langle X_{max} \rangle$ measurements. Such uncertainty has as consequence the degradation of the accuracy of other quantities, as for example, the composition of the UHECR [@osta]. ![Predictions for the average longitudinal profiles of muons for air showers generated by protons and iron nuclei with primary energies of $10^{17}$ eV (upper panels) and $10^{20}$ eV (lower panels) . Solid (dashed) lines represent the full (non - diffractive) simulations.[]{data-label="long_mu_17"}](long_mu_17.eps "fig:"){width="\textwidth"}\ ![Predictions for the average longitudinal profiles of muons for air showers generated by protons and iron nuclei with primary energies of $10^{17}$ eV (upper panels) and $10^{20}$ eV (lower panels) . Solid (dashed) lines represent the full (non - diffractive) simulations.[]{data-label="long_mu_17"}](long_mu_20.eps "fig:"){width="\textwidth"} Lets now analyse the impact on the muonic content of the air showers, which is known to be related to the primary composition and also to the properties of the hadronic interactions on the shower development. In particular, Auger collaboration has recently measured the depth of maximum production of muons $X_{max}^{\mu}$ in high energy air showers and has verified that QGSJET-II 04 bracketed the data with simulations of proton and iron induced showers, while EPOS LHC underestimates values of $X_{max}^{\mu}$ [@Aab:2014dua]. Auger collaboration also showed that both models underestimate the muon content that reach the ground [@Aab:2014pza], with the discrepancies being larger for the QGSJET-II 04 predictions than those from the EPOS LHC model. In Fig. \[long\_mu\_17\] we present the predictions of the different hadronic generators for the longitudinal profiles of the muonic component for air showers generated at $10^{17}$ eV (upper panels) and $10^{20}$ eV (lower panels). Qualitatively, the influence of diffraction is the same observed for the profiles of charged particles: simulations without diffractive processes produce more muons and the profiles are shifted towards lower atmospheric depths. It is worth note that EPOS LHC predicts more muons than the other models and that its predictions for the profiles are more dependent on the diffractive interactions. ![Ratio between the predictions for muon densities obtained considering only non - diffractive interactions and those derived including diffractive and non - diffractive interactions as a function of the distance to the shower core. Showers initiated by protons (iron nuclei) are represented by blue circles (red triangles). Solid (open) symbols denote showers initiated by primary reaching the atmosphere with a zenithal angle of 60º (30º). []{data-label="late_mu_17"}](late_razao_mu_17.eps){width="\textwidth"} Finally, lets analyse the influence of the diffractive interactions on the density of muons that hit the ground ($\rho_{\mu}$). In particular, we estimate the ratio between the predictions for $\rho_{\mu}$ obtained considering only non - diffractive interactions and those derived including diffractive and non - diffractive interactions. In Fig. \[late\_mu\_17\] we present our results for the ratio $\rho_{\mu}^{(ND)}/\rho_{\mu}^{(full)}$ as a function of distance to the core shower, assuming a primary energy of $10^{17}$ eV and an observation level of 1400 m (Pierre Auger Observatory), which corresponds to an atmospheric depth of 1760 $g/cm^2$ [ for a zenith angle of 60º. For showers generated with a zenith angle of 30º degrees, such altitude corresponds to an atmospheric depth of $\approx$ 1000 $g/cm^2$.]{} We present results assuming [ both values]{} for the zenith angle. We have that the ratio increases with distance to the shower core and that the influence of the diffractive interactions becomes non - negligible at large distances, specially for small values of the zenith angle. A final comment is in order. After the completion of our study, a new version of the event generator Sibyll have been released [@sibyll23]. One of the implications of the modifications implemented in Sibyll 2.3 is the enhancement of the diffractive interactions. As a consequence, the new version predicts, on average, less secondaries in the whole energy range considered for both primaries. One have verified that our results are not strongly modified. The main modification is that the differences between non - diffractive and full simulations of the distinct observables are enhanced by a factor smaller than 1.25 ($\lesssim$ 25%) in comparison with the former version Sibyll 2.1, for all energies and primaries. Therefore, our main conclusions about the impact of the diffractive interactions remain valid using the Sibyll 2.3. Summary {#conc} ======= In this paper we invertigated the impact of the diffractive interactions on distinct observables for $p$-Air and $Fe$-Air collisions as well as in ultra - high energy cosmic ray interactions. Our results demonstrated that the distinct phenomenological models, present in the CORSIKA package, predict a different magnitude for the fraction of diffractive events and for its energy dependence. As a consequence, the influence of these interactions on the number of secondaries and fraction of pions is strongly model dependent. We demonstrated that the predictions of these models for the elasticiy are distinct, which directly modifies the air shower development. Our results for $\langle X_{max} \rangle$ indicated that the impact of the diffractive interactions on this observable is almost energy-independent for the three considered models of hadronic interactions and it is of the order of the typical experimental precision for the $\langle X_{max} \rangle$ measurements. Moreover, we shown that the average longitudinal profile of muons is sensitive to the diffractive interactions, in particular for small zenithal angles. Our results indicated that the diffractive interactions has a non - negligible influence on the observables and that the treatment of the diffractive physics is an important source of uncertainty in the description of the extensive air showers. This work was financed by the Brazilian funding agencies CAPES, CNPq and FAPERGS. [99]{} K. Akiba [*et al.*]{} \[LHC Forward Physics Working Group Collaboration\], J. Phys. G [**43**]{}, 110201 (2016) P. D. B. Collins, [*An Introduction to Regge theory and high energy physics*]{} (Cambridge University Press, Cambridge, England, 1977). A. B. Kaidalov, Phys. Rept.  [**50**]{}, 157 (1979); M. G. Ryskin, A. D. Martin, V. A. Khoze and A. G. Shuvaev, J. Phys. G [**36**]{}, 093001 (2009) S. Ostapchenko, Phys. Rev. D [**83**]{}, 014018 (2011); EPJ Web Conf.  [**52**]{}, 02001 (2013). E. J. Ahn, R. Engel, T. K. Gaisser, P. Lipari and T. Stanev, Phys. Rev. D [**80**]{}, 094003 (2009) G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**72**]{}, 1926 (2012); B. Abelev [*et al.*]{} \[ALICE Collaboration\], Eur. Phys. J. C [**73**]{}, no. 6, 2456 (2013); G. Antchev [*et al.*]{} \[TOTEM Collaboration\], Phys. Rev. Lett.  [**111**]{}, no. 26, 262001 (2013); V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Rev. D [**92**]{}, no. 1, 012003 (2015) R. Engel, D. Heck and T. Pierog, Ann. Rev. Nucl. Part. Sci.  [**61**]{}, 467 (2011). P. Lipari, EPJ Web Conf.  [**53**]{}, 07004 (2013). F. Riehn, R. Engel, A. Fedynitch, T. K. Gaisser and T. Stanev, arXiv:1510.00568 \[hep-ph\]. T. Pierog, I. Karpenko, J. M. Katzy, E. Yatsenko and K. Werner, Phys. Rev. C [**92**]{}, no. 3, 034906 (2015) R. Ulrich, C. Baus and R. Engel, EPJ Web Conf.  [**99**]{}, 11001 (2015). S. Ostapchenko, M. Bleicher, T. Pierog and K. Werner, Phys. Rev. D [**94**]{}, no. 11, 114026 (2016) D. d’Enterria, R. Engel, T. Pierog, S. Ostapchenko and K. Werner, Astropart. Phys.  [**35**]{} (2011) 98 R. D. Parsons, C. Bleve, S. S. Ostapchenko and J. Knapp, Astropart. Phys.  [**34**]{}, 832 (2011) A. Aab [*et al.*]{} \[Pierre Auger Collaboration\], Phys. Rev. Lett.  [**117**]{}, no. 19, 192001 (2016) D. Heck, J. Knapp, J.N. Capdevielle, G. Schatz, and T. Thouw, Report [**FZKA 6019**]{} (1998), Forschungszentrum Karlsruhe; available from\ http://www-ik.fzk.de/corsika/physics\_description/corsika\_phys.html R. Luna, A. Zepeda, C. A. Garcia Canal and S. J. Sciutto, Phys. Rev. D [**70**]{} (2004) 114034 S. J. Sciutto, astro-ph/0106044. V. N. Gribov, Sov. Phys. JETP [**26**]{}, 414 (1968) \[Zh. Eksp. Teor. Fiz.  [**53**]{}, 654 (1967)\]. L. V. Gribov, E. M. Levin and M. G. Ryskin, Phys. Rept.  [**100**]{}, 1 (1983). J. L. Albacete and C. Marquet, Prog. Part. Nucl. Phys.  [**76**]{}, 1 (2014). R. J. Glauber and G. Matthiae, Nucl. Phys. B [**21**]{}, 135 (1970). V. N. Gribov, Sov. Phys. JETP [**29**]{}, 483 (1969) \[Zh. Eksp. Teor. Fiz.  [**56**]{}, 892 (1969)\]. M. L. Good and W. D. Walker, Phys. Rev.  [**120**]{}, 1857 (1960). C. A. Garcia Canal, S. J. Sciutto and T. Tarutina, Phys. Rev. D [**79**]{} (2009) 054006 S. Ostapchenko, Phys. Rev. D [**89**]{}, no. 7, 074009 (2014) A. Aab [*et al.*]{} \[Pierre Auger Collaboration\], Phys. Rev. D [**90**]{} (2014) no.1, 012012 Addendum: \[Phys. Rev. D [**90**]{} (2014) no.3, 039904\] Erratum: \[Phys. Rev. D [**92**]{} (2015) no.1, 019903\] A. Aab [*et al.*]{} \[Pierre Auger Collaboration\], Phys. Rev. D [**91**]{} (2015) no.3, 032003 Erratum: \[Phys. Rev. D [**91**]{} (2015) no.5, 059901\]
{ "pile_set_name": "ArXiv" }
--- abstract: | Recent developments in the field of Networking have provided opportunities for networks to efficiently cater application specific needs of a user. In this context, a routing path is not only dependent upon the network states but also is calculated in the best interest of an application using the network. These advanced routing algorithms can exploit application state data to enhance advanced network services such as anycast, edge cloud computing and cyber physical systems (CPS). In this work, we aim to design such a routing algorithm where the router decisions are based upon convex optimization techniques.\ author: - | Sumit Maheshwari\ *WINLAB/ECE, Rutgers University*\ `[email protected]` title: | Joint Optimization of Application Specific Routing in an Anycast Network\ [^1] --- Convex optimization; Application Specific Routing; [99]{} Michael, Nithin, et al. “Analysis of application-aware on-chip routing under traffic uncertainty.” Networks on Chip (NoCS), 2011 Fifth IEEE/ACM International Symposium on. IEEE, 2011. Shareef, Ali, Aliha Shareef, and Yifeng Zhu. “Optrix: Energy aware cross layer routing using convex optimization in wireless sensor networks.” Networking, Architecture and Storage (NAS), 2015 IEEE International Conference on. IEEE, 2015. Maheshwari, Sumit, et al. “A joint parametric prediction model for wireless internet traffic using Hidden Markov Model.” Wireless networks 19.6 (2013): 1171-1185. www.orbit-lab.org. Kohler, Eddie, et al. “The Click modular router.” ACM Transactions on Computer Systems (TOCS) 18.3 (2000): 263-297. Sun, Yan, et al. “ASAR: An ant-based service-aware routing algorithm for multimedia sensor networks.”Frontiers of Electrical and Electronic Engineering in China3.1 (2008): 25-33. Chen, Lei, and Wendi B. Heinzelman. “QoS-aware routing based on bandwidth estimation for mobile ad hoc networks.”IEEE Journal on selected areas in communications23.3 (2005): 561-572. [^1]:
{ "pile_set_name": "ArXiv" }
--- author: - 'K.K. Knudsen' - 'P.P. van der Werf' - 'W. Jaffe' date: 'Received / Accepted ' title: 'A Submillimetre Selected Quasar in the Field of [^1]' --- Introduction ============ Major advances in submillimetre (submm) continuum observations came with the Submillimetre Common-User Bolometer Array (SCUBA, Holland et al. [@Holland]), which is mounted at the 15m James Clerk Maxwell Telescope (JCMT) at Mauna Kea, Hawaii. This dual-channel instrument for the first time allowed sensitive mapping, making it possible to survey larger areas of the sky to greater depths than previously possible at submm wavelengths. This development led to the discovery of a new class of objects of high infrared (IR) luminosity, located at cosmological distances (e.g., Smail et al. [@Smail]). Even though these objects are less common than Lyman-break galaxies at similar redshifts, they would dominate the cosmic star formation rate density at these redshifts, if star formation is indeed the source of their high luminosity (e.g., Blain et al. [@Blainetal99]). Since the importance of these objects was realized, a number of submm surveys have been performed or are in progress (e.g. Smail et al. [@SIBK] ; Eales et al. [@Eales]; Scott et al. [@scott]; Chapman et al. [@chapman_lens]; Knudsen et al. in prep.). One of the biggest challenges for those surveys has turned out to be the follow-up observations and the identification of the counterparts causing the submm emission. As a result, the determination of the nature and redshift of these objects has been significantly hampered. While currently more than a hundred submm sources have been detected, less than 20 of these have reliably been identified with sources at other wavelengths, in spite of painstaking attempts. The number of published spectroscopic redshifts is even significantly smaller. The majority of the reliably identified counterparts are very or extremely red objects (e.g., Frayer et al. [@smmj00266]; Smail et al.  [@smail_smmero]); several are also exhibiting active galactic nuclei (AGN) features in their spectra (Ivison et al. [@smmj02399]). Because of the small number of secure identifications, any new identification adds important information to our understanding of the submm population. A survey with different selection criteria, radio-preselected and submm-detected, has produced 10 spectroscopic redshifts in a sample of 34 sources (Chapman et al. [@chapman_nat]). We are carrying out an extensive SCUBA survey of a number of galaxy clusters fields, aimed at detecting gravitationally amplified background galaxies: the Leiden-SCUBA Lensed Survey (Knudsen et al. [*in prep.*]{}). In the course of doing the optical identifications and follow-up of this survey we discovered one of our submm sources to be a previously unknown type-1 quasar (previously reported in Knudsen et al. [@knudsen00]). While submm surveys of optically selected quasars have been quite succesful (Isaak et al. [@Isaak]), this object is the first type-1 quasar first discovered by its submm emission. In contrast, type-2 quasars have been detected in small numbers in other submillimetre surveys (e.g. , Vernet & Cimatti [@vernet]), and IRAS-radio-optical quasars have been selected before at a wide range of redshifts (e.g.  in Irwin et al.  [@irwin]). In this paper we present the observations of the quasar. We discuss unusual properties of the object, its optical spectrum, and its IR spectral energy distribution, and compare the results to optically selected quasars. We adopt an $\Omega_0 = 0.3$ and $\Lambda = 0.7$ cosmology with $H_0 = 70\ {\rm km\,s}^{-1}{\rm Mpc}^{-1}$. Observations and results {#sect:obs} ======================== Submillimetre data ------------------ The SCUBA data of the $z=0.088$ galaxy cluster have been obtained during five nights in September and December 1997, March 1998 and December 1999. The first data were obtained in a program to study the cooling flow in the cluster itself. In these data a bright point source was detected. Consequently, extra data was obtained to study this object better. The total integration time was 6.6 hours (excluding overheads), recording data at both 850$\mu {\rm m}$ and 450$\mu {\rm m}$ simultaneously in jiggle-map mode. The data were obtained mostly under good conditions with 850$\mu {\rm m}$ zenith atmospheric opacity typically around 0.2. The pointing was checked regularly and was found to be stable. Calibration maps of CRL618 were also obtained. The data were reduced using the [SURF]{} (SCUBA User Reduction Facility) and [KAPPA]{} software packages (Jenness & Lightfoot [@surf]). The resulting images have an angular resolution of $15''$ at $850\,\mu$m and $8''$ at $450\,\mu$m. Source extraction and estimation of the uncertainties were carried out using a method based on Mexican Hat wavelets (Cayón et al. [@mexhat]; Barnard et al. in prep; Knudsen et al. in prep), which was adopted for the entire Leiden-SCUBA Lensed Survey, and which will be described in a forthcoming publication (Knudsen et al. in prep.), where the full survey will be presented. This method was adopted because it is mathematically rigorous and its performance on SCUBA jiggle maps can be fully characterized. Monte Carlo simulations have been performed to determine the noise and uncertainties of the derived parameters. The area-weighted noiselevels of the maps are 2 mJy at $850\ \mu{\rm m}$ and 14 mJy at $450\ \mu{\rm m}$. In the $850\,\mu{\rm m}$ map four sources were detected of which the brightest has a flux of $S_{850} = 25\pm 2.8\ {\rm mJy}$. This is the only source in the map with detected $450\,\mu{\rm m}$ emission, $S_{450} = 55\pm 17\ {\rm mJy}$. It was detected with a signal-to-noise of 15, for which the formal positional uncertainty including the pointing uncertainty of the JCMT is $3.2''$. This is the object for which we are here presenting the follow-up observations. Fluxes and positions are presented in Table \[tab:flux\_pos\]. Optical identification ---------------------- For identification and redshift determination of the SCUBA source(s) optical imaging and spectroscopy was obtained with FORS1 at VLT-UT1 (Antu) in Chile, in September 1999. Four 15 min exposures in $I-$band were aquired in photometric conditions. The frames were bias-subtracted, flatfielded and stacked. The resulting image is shown in Fig. \[fig:FORS1SCUBA\] with the SCUBA $850\ \mu{\rm m}$ contours overlayed. The seeing measured in the final image is $0\farcs 9$. The standard star field (Landolt [@landolt]) was used for the calibration. The source detection and photometry was performed using SExtractor (Bertin & Arnouts [@SEx]). The center of is coincident with an $I=20.5\pm 0.1\ {\rm mag}$ point source at $\alpha = 04^{\rm h}13^{\rm m}27\fs28$, $\delta = 10\degr 27\arcmin40\farcs4$ (J2000). There are no other apparent candidate counterparts. The optical position is within the error circle of the submm observation. One of the other SCUBA sources () coincides with a galaxy, which, given its size and magnitude, is a probable cluster member. There are no obvious candidate counterparts for the two other SCUBA sources. Using the DIRBE/FIRAS maps (Schlegel et al. [@dustmaps]), a Galactic reddening $E(B-V) \approx 0.52\ {\rm mag}$ is derived — a substantial reddening. The corrected $I$ magnitude of the optical counterpart for SMMJ04135+10277 is thus $19.4\pm 0.1\ {\rm mag}$. Optical spectroscopy -------------------- FORS1 spectroscopy of SCUBA sources in the A478 field was also obtained in September 1999. We used FORS1 in Multi-Object Spectroscopy (MOS) mode to obtain spectra of a number of targets, using grism 150I+17, without order sorting filter. This setup gives a spectral resolution of 260 at 720nm with the $1''$ slit which we employed. Overlap of the second spectral order may affect the wavelength region longwards of 650nm, but was in the present case found not to affect the spectra. Two exposures of 1800sec were obtained in a seeing of $1.3''$. The spectra were bias-subtracted and flatfielded. Wavelength calibration was achieved using exposures of He and Ar lamps. Correction for telluric absorption and flux calibration was carried out using observations of the white dwarf EG274 ($V=11.03$), which we corrected for photospheric absorption features. The multislit mask included both the bright SCUBA source SMMJ04135+10277 and the fainter SCUBA source SMMJ04134+10270. The extracted spectra were corrected for galactic foreground absorption using the DIRBE/FIRAS maps. The optical spectrum of (Fig. \[fig:spectrum\]) shows broad emission lines, of which the four most prominent can be identified with Ly$\alpha$+N[v]{}, Si[iv]{}+O[iv]{}\], C[iv]{} and C[iii]{}. In addition the spectrum shows a power-law continuum. All of these features are characteristic of quasars. Bluewards of the Ly$\alpha$ emission line, Ly$\alpha$ forest absorption is seen. We use the C[iii]{}, C[iv]{} and the Si[iv]{}+O[iv]{}\] lines, with the largest weight on the symmetric C[iii]{} line, to determine the redshift. We find the value $z = 2.837 \pm 0.003$, consistently for the peak value of all three profiles. The spectrum of confirms its membership of the A478 cluster. It shows the characteristic spectrum of a quiescent elliptical galaxy with no evidence for nuclear activity. This source will be discussed together with the rest of the survey in Knudsen et al. (in prep.). Near-infrared spectroscopy -------------------------- Since restframe ultraviolet emission lines of quasars can be significantly blueshifted with respect to the systemic velocity (e.g., Carswell et al. [@Carswell]), we also attempted to obtain additional redshift information using restframe optical lines. Unfortunately, at $z\approx2.84$, the brightest lines (H$\alpha$, H$\beta$, \[O[iii]{}\] 5007Å) are all in wavelength regions where the earth atmosphere is opaque. We therefore concentrated on the \[O[ii]{}\] 3727Å line which is redshifted to the blue edge of the H-band window, a region strongly affected by atmospheric absorption lines. We used ISAAC on VLT-UT1 (Antu) in February 2002, to take H-band spectra of . We used the medium resolution grating with a $1''$ slit to obtain an $R=3000$ spectrum between 1.41 and $1.49\,\mu$m, which should contain the \[O[ii]{}\] line for redshifts between 2.78 and 3.00. In addition, we obtained a low-resolution spectrum ($R=500$) of the entire H-band, in an attempt to detect H$\gamma$, which although intrinsically faint, should at least lie in a clear part of the spectrum. Both spectra were obtained in photometric conditions and in an optical seeing of $0.8''$, by nodding the object along the slit. Total integration time was 44minutes in both spectra. The individual frames were pairwise subtracted in order to remove the bright OH nightsky lines, flatfielded and coadded. Wavelength calibration was derived from the OH nightsky lines. Correction for telluric absorption and flux calibration were achieved using the B5V star Hip25499 (H=5.62) and the B2V star Hip28142 (H=7.497), corrected for photospheric absorption. Unfortunately, while the continuum of the quasar is clearly detected in both spectra, no emission features are seen. Undoubtedly, this is due to atmospheric absorptions in the region of the redshifted \[O[ii]{}\] line, and the faintness of the relevant features in the rest of the H-band spectrum. passband RA(J2000) Dec(J2000) $f_{\nu}$ -------------------- ------------- ------------- --------------------------- $850\mu{\rm m}$ 04:13:27.2 +10:27:43 $25\pm2.8\, {\rm mJy}$ $450\mu{\rm m}$ 04:13:27.2 +10:27:42 $55\pm17\, {\rm mJy}$ $180\mu{\rm m}$ $< 620\, {\rm mJy}$ $14.3\mu{\rm m}$ 04:13:27.24 +10:27:44.5 $470\pm80\, {\rm mJy}$ $6.7\mu{\rm m}$ 04:13:27.88 +10:27:43 $200\pm30\, {\rm mJy}$ $I$ 04:13:27.28 +10:27:41.4 $19.4\pm 0.1\, {\rm mag}$ $4.86\, {\rm GHz}$ 04:13:27.26 +10:27:40.5 $220\pm35\, \mu{\rm Jy}$ $1.4\, {\rm GHz}$ $< 750\, \mu{\rm Jy}$ : Coordinates and fluxes at different wavelengths for SMMJ04135+10277. The fluxes as they are listed here have not been corrected for the gravitational lensing. \[tab:flux\_pos\] ISO data -------- We also inspected the archive of the Infrared Space Observatory (ISO, Kessler et al. [@ISO]) and extracted observations of A478 using both the mid-infrared camera (CAM, Cesarsky et al. [@CAM]) and the mid/far-infrared photometer (PHOT, Lemke et al.[@PHOT]). The PHOT data were obtained using the P22 raster mode at $180\,\mu$m, with $92''$ pixels, on February 21, 1998. The data were reduced using the PHOT Interactive Analysis (Gabriel et al. [@PIA]). Initial data reduction steps included discarding of corrupted data, non-linearity correction, and deglitching of individual ramps. After fitting all integration ramps with a first order polynomial, further deglitching and data editing, and dark current subtraction, the data were corrected for detector drifts and for vignetting, and calibrated using the internal Fine Calibration Sources. The resulting image shows a $0.62\pm0.19\,$Jy source, the centroid of which is however displaced by $46''$ from SMMJ04135+10277. This displacement is less than the ISO angular resolution at $180\,\mu$m, but much more than the nominal ISO pointing uncertainty of $2''$. While the centroid of a faint source can be displaced somewhat when placed on top of a highly structured background, in the present case the offset is so large that the $180\,\mu$m detection cannot reliably be associated with the quasar. Hence in the following we label this as an upper limit. The CAM data were obtained using the LW3 filter (effective wavelength $14.3\,\mu$m) on February 21, 1998, and using the LW2 filter (effective wavelength $6.7\,\mu$m) on March 21, 1998 using the CAM01 raster observing mode with $6''$ pixels. The data were reduced using the CAM Interactive Analysis package (Ott et al. [@CIA]). Processing steps consisted of dark current subtraction, deglitching and correction of transients using the PRETI method (Aussel et al. [@Aussel]), which is particularly suited for the detection of faint sources, flatfielding using a flatfield derived from the stacked dataframes, and mosaic contruction taking into account the image distortion. This resulted in clear detections of SMMJ04135+10277 at $14.3\,\mu$m with a flux density of $0.47\pm0.08\,$Jy, and at $6.7\,\mu$m with a flux density of $0.20\pm0.03\,$Jy. Positions of these sources are listed in Table \[tab:flux\_pos\]. CO $J=3{\to}2$ emission ----------------------- In a recent commisioning project of the new COBRA spectrometer on the Owens Vally Radio Observatory, the CO $J=3{\to}2$ emission line has been detected from the quasar. This detection confirms the nature of as a hyperluminous IR quasar. The redshift implied by the CO line is $z=2.84$, which is indeed somewhat higher than the optically determined redshift. This result will be discussed in detail in Hainline et al. (in prep). Discussion {#sect:res_disc} ========== SMMJ04135+10277 and the optical quasar population ------------------------------------------------- Only little is known about the importance of AGNs in the submm population. Most studies comparing X-ray and submm observations conclude that the submm population is powered by star formation rather than AGNs and especially quasars (e.g., Almaini et al. [@almaini]), based on the lack of overlap of X-ray and submm sources in deep studies. This has been confirmed in a study combining very deep [*Chandra*]{} observations with SCUBA observations of the HDF-N, where Alexander et al. ([@alexander]) found that a significant fraction of bright submm sources ($f_{850\mu{\rm m}}>5\ {\rm mJy}$) harbour an AGN, however, the AGN is not powerful enough to power the submm emission. This makes SMMJ04135+10277 a particularly interesting object, since here we have a bright submm source that is unequivocally identified with a type-1 quasar. Given what is known about the abundance of type-1 quasars, is this an ordinary object that we should have expected to find in our survey, or are we dealing with an exceptional case? We here estimate the probability of finding a high redshift submm emitting quasar in our survey. The total area of our survey is $65\ {\rm arcmin}^2$ (Knudsen et al., in prep.). Using the optical spectrum we estimate that the quasar has a $B$ magnitude $B \sim 21.0-21.5\ {\rm mag}$. Based on the counts of Kennefick et al. ([@Kennefick]), we find that there is only a 20% probability of finding a quasar with $z > 2.3$ and $16.5\,{\rm mag} < B < 22\,{\rm mag}$ in our survey. Furthermore, the probability that such a quasar is a bright submm source is also less than unity, as shown by Priddey et al. ([@Priddey]), who did a submm study of optically selected quasars at $1.5 < z < 3$. For the sub-sample of quasars with $z>2.3$, only 30% of these had detectable submm emission down to 6.8 mJy and all of those are fainter than . Combining the numbers we estimate only a 6% chance of detecting a submm bright quasar at $z>2.3$ in our survey, if that quasar was drawn from the population of optically selected quasars. We also estimate the expected number of bright submm sources in the surveyed area, regardless of their physical nature. According to the number counts from Smail et al.  ([@SIBK]) we should expect to find two sources with $850\,\mu$m fluxes between 20 and 25mJy. Our observations (Knudsen et al., in prep.) are in agreement with that number. Comparing this to the small chance of finding a high redshift submm emitting quasar in our survey, this result suggests that the bright part of the submm population does not originate from dusty quasars, and that is an unusual object. Optical spectrum {#subsect:opt_spec} ---------------- Since is the first quasar selected based on its submm emission, it is of interest to compare its properties to those of optically selected quasars. Turning first to the optical spectrum, the shapes of the C[iii]{}, C[iv]{} and the Si[iv]{}+O[iv]{}\] lines appear as expected. However, the Ly$\alpha$+N[v]{} emission line has a more unexpected shape. The peak and blue wing appear to be absorbed. Furthermore, the strength of the line relative to the other emission lines is unusually low for a quasar. Since dust is present in this quasar, it is natural to assume that atomic hydrogen will also be present, so that associated absorption may play a role in suppressing the Ly$\alpha$ emission. However, for a more detailed assessment of this effect, a higher resolution spectrum is needed. Comparing the optical spectrum of SMMJ04135+10277 to that of quasars selected at other wavelengths (see e.g. Francis et al. ([@Fran92]) for a composite spectrum), there are no significant differences except for the suppressed Ly$\alpha$ emission. We note that SMMJ04135+10277 belongs to the optically fainter part of the quasar population. Can this be the effect of strong absorption by dust, which would then simultaneously account for the luminous dust emission from SMMJ04135+10277? We obtain a measure of the isotropic luminosity of the quasar nucleus using the C[iv]{} emission line; obviously, the Ly$\alpha$ line cannot be used since it appears to be absorbed, and the continuum cannot be used because of the effects of relativistic beaming, which cannot reliably be quantified. The observed flux of the C[iv]{} line is $1.0\cdot10^{-14}\,$ergs$^{-1}$cm$^{-2}$. A comparison sample can be constructed from the optically selected submm emitting quasars studied by Priddey et al. ([@Priddey]), using the spectra from Hagen et al. ([@Hagen]). This comparison sample covers redshifts from 2.60 to 2.79 and can therefore be compared directly to SMMJ04135+10277. The comparison sample has C[iv]{} fluxes from 1.7 to $4.1\cdot10^{-14}\,$ergs$^{-1}$cm$^{-2}$, roughly a factor of 3 higher than SMMJ04135+10277. The rest-frame [*equivalent width*]{} of C[iv]{} on the other hand shows the opposite trend: while SMMJ04135+10277 has a C[iv]{} restframe equivalent width of $\sim 170\,$Å, values in the comparison sample are approximately a factor of 10 lower, ranging from 13 to $25\,$Å. In other words, the quasar continuum is fainter by about a factor of 30 than would be expected for its C[iv]{} flux. It is highly unlikely that extinction could account for this, since the quasar continuum and the broad line region should be viewed through approximately the same obscuring column. Furthermore, as Fig. \[fig:spectrum\] shows, the quasar continuum is characterized by a blue power law. The slope of this continuum does not indicate the presence of abnormally large absorption. Therefore a more likely explanation of the optical faintness of this quasar is a large viewing angle away from the direction of relativistic beaming. The beamed flux density is proportional to $\delta^p$ with $p\sim4$, where the Doppler factor $\delta=[\gamma(1-\beta\cos\theta)]^{-1}$, where $\beta$ is the bulk velocity in units of the speed of light, and $\gamma=(1-\beta)^{-1/2}$ is the corresponding Lorentz factor, and $\theta$ is the angle away from the beam (Urry & Padovani [@UrryPadovani]). Therefore a decrease in $\delta$ of a factor 2.3 would be sufficient to produce a factor 30 decrease in the beamed continuum with respect to the lines. The required angle away from the beam cannot be calculated since $\beta$ is not known. However, as shown by Urry & Padovani ([@UrryPadovani]), variations in $\delta$ of this magnitude are entirely reasonable for angles $\theta<20\degr$, provided $\gamma>2$. This estimate confirms the viability of our suggestion that the optical faintness of the quasar is due to a large viewing angle away from the direction of relativistic beaming, and not to abnormally large extinction. If in fact the optical spectrum is still dominated by the doppler boosted jet then our detection of this one object suggests that a much larger number of yet unidentified sources are similar AGNs viewed from a larger angle to the jet axis. We finally note that it would be interesting to make the same comparison with low-$z$ far-IR detected quasars, addressing also the properties of the dust emssion spectrum. This comparison would require spectrophotometry of quasars in the vacuum ultraviolet. Going further, we can investigate whether the observed submm emission from SMMJ04135+10277 is likely powered by the AGN or whether the presence of an additional power source is indicated. In the comparison sample, the observed $850\,\mu$m fluxes range from 6.8 to 10.0mJy, increasing monotonically with C[iv]{} flux. The three times fainter C[iv]{} flux of SMM04135+10277 thus would suggest an AGN-powered $850\,\mu$m flux of approximately 3mJy. The observed flux is almost a factor of 10 higher. This result suggests that the bulk of the submm emission from SMM04135+10277 is not powered by the AGN but by an additional source of energy, most likely vigorous star formation in the host galaxy. If this interpretation is correct, high resolution imaging of CO lines and dust emission with ALMA should reveal an extended source. Gravitational magnification --------------------------- The low redshift ($z=0.088$) of A478 is far from the optimal lensing redshift ($z \sim 0.2$), and no arcs are detected in the vicinity of the quasar. This suggests that the gravitational magnification is small and that the quasar is not subject to differential lensing, which otherwise would influence the shape of the spectral energy distribution (SED). We calculate the magnification of the quasar using LENSTOOL (Kneib et al.[@lenstool]). The cluster is modelled using two components: the overall cluster potential with a core radius $r_{\rm c} = 250\ {\rm kpc}$ and a velocity dispersion $\sigma = 905\ {\rm km\,s}^{-1}$ and the potential of the cD galaxy with $r_{\rm c} = 2\ {\rm kpc}$ and $\sigma = 350\ {\rm km\,s}^{-1}$ (Allen et al. [@A478numbers1]; Zabludoff et al. [@A478numbers2]). We find that the quasar is magnified by a factor of 1.3. Hence, all fluxes should be corrected for this value. In all calculations in the following sections of this paper the fluxes have been corrected for the gravitational lensing. Spectral energy distribution ---------------------------- From archival data from the NRAO Very Large Array a radio source near the position of the SMMJ04135+10277 was found. The fluxes measured are $S(4.86{\rm GHz})= 220\pm35\,\mu{\rm Jy}$ and $S(1.4{\rm GHz}) < 750\,\mu{\rm Jy}$ ($3\sigma$) (M. Yun, private comm.). Based on this low radio flux, it is concluded that the quasar is radio-quiet (according to the radio-power criterion given by Stocke et al. [@stocke] to divide quasars into radio-loud and radio-quiet types). Hence, the non-thermal contribution to the submm flux is expected to be small and is here neglected. Combined, we have then six points on the SED and two upper limits, ranging from the radio to the optical regime (see Table \[tab:flux\_pos\]). The SCUBA points and the two ISO points at $14.3\,\mu{\rm m}$ and $170\,\mu{\rm m}$ are in the rest frame all at wavelengths typical for thermal emission by dust. The ISO $6.7\,\mu{\rm m}$ point in rest frame is $1.7\,\mu{\rm m}$, which together with the optical point is expected to originate from stellar light, possibly contaminated with non-thermal emission from the AGN. The radio emission is attributed to non-thermal synchrotron emission. In the following we will focus on the thermal dust emission. Of the thermal emission, the two SCUBA points are most likely due to the cool dust typically described by a modified blackbody, whereas the $14.3\,\mu{\rm m}$ point arises from a hot component. The shape of the SED between the cold and hot component seems to be different for different quasars (see e.g. Haas et al. [@haas_pg]). We have no measurements between $450\ \mu{\rm m}$ and $14.3\ \mu{\rm m}$, which makes an analysis of the IR SED difficult. We do, however, attempt to make a tentative analysis in which we compare with known objects and also estimate parameters like temperature and luminosity. We first compare the quasar with other known quasars. Comparing to high-$z$ quasars is not trivial, since the high-$z$ quasars which have well-sampled IR SEDs, are often strongly lensed and their observed SEDs may have suffered differential lensing. We therefore first focus on low-$z$ quasars. Haas et al. ([@haas_pg]) have made a detailed study of the IR SED of Palomar-Green (PG) quasars. The majority of these quasars are at fairly low redshift. We compare to three low-$z$ PG quasars with well-sampled SEDs, , and (all shown in Fig. \[fig:SED\]). All three SEDs are redshifted to $z=2.837$. If the SEDs are scaled to the quasar $850\,\mu{\rm m}$ point, the comparison gives the impression of a deficit in the mid/near-IR emission of SMMJ04135+10277. Alternatively, inspired by the findings of Archibald et al. (2001) and Page et al. (2001), that the star formation rate observed in AGNs is higher at higher redshift, leading to enhanced long-wavelength emission at higher redshift, we may choose instead to scale the low-$z$ SEDs to the observed $14.3\,\mu{\rm m}$ point, i.e., the hot dust emission associated with the AGN. This, not unexpectedly, then suggests an excess in the far-IR-submm emission from SMMJ04135+10277. This result corroborates our earlier conclusion that a significant portion of the observed $850\,\mu$m emission of SMMJ04135+10277 results from extended star formation, and is not powered directly by the AGN. The SED of the strongly lensed $z=3.87$ quasar APM08279+5255 (Lewis et al. [@lewis] and references therein) is also shown in Fig. \[fig:SED\]. It has also been appropriately shifted and scaled to the $14.3\,\mu{\rm m}$ point. In this case the submm/FIR deficit relative to the is even more pronounced, corroborating the discussion above. Given this result, it is also of interest to compare the SED of SMMJ04135+10277 to the SEDs of well-studied starburst galaxies. We use the SEDs of the starburst galaxy and the ultraluminous infrared galaxy (ULIRG) (extracted from the NASA Extragalactic Database), redshifted to $z=2.837$, for comparison. Scaled to the $850\,\mu{\rm m}$ point, the far-IR/submm range matches quite well, whereas the mid/near-IR emission is much brighter for SMMJ04135+10277. This result is expected, as starbursts are known not to have the hot dust component that is characteristic of AGNs, especially quasars (Sanders et al.  [@Sanders]; Barvainis [@Barvainis]). Considering that the SED of the quasar is not well-sampled, and that we are looking at only one quasar, no definite conclusions can be drawn about the precise shape of the IR SED and the power source of the dust emission of SMM04135+10277 in particular, or of the submm-selected quasar population in general. Tentatively, though, these observations support the suggestion that quasars at higher redshift have a high submm and far-IR flux, suggesting a higher star formation rate. To study this in detail, however, observations filling the big gaps in the IR SED are needed. Such data can possibly be obtained with SIRTF. Finally, we calculate the dust temperature, dust mass, and total luminosity in dust emission of SMMJ04135+10277. For the dust emission, we use blackbody emission, modified by the frequency-dependent mass absorption coefficient $$k_{\rm d}(\nu_{\rm rest}) = 1.5\,{\rm cm}^2\,{\rm g}^{-1} \left( \frac{\nu_{\rm rest}}{375\ {\rm GHz}} \right)^\beta,$$ using the average value from the literature at $800\,\mu$m (see Hughes et al. 1997 for a discussion of the assumptions and uncertainties in this parameter) and assume $\beta = 1.5$. Since the IR SED cannot be fit by a single modified blackbody, we first fit the cold dust component sampled by the SCUBA points. Using only the SCUBA points, we find a temperature $T_{\rm cold} = 29\pm 2\ {\rm K}$ and a dust mass of $M_{\rm d} = (1.8\pm0.3)\times 10^{9}\ {\rm M_{\odot}}$ for the cold dust component. The total luminosity of a modified blackbody spectrum can be calculated using the following analytical expression: $$L_{\rm tot} = 4\pi D_{\rm L}^2 \frac{S(\nu_{\rm obs})}{\nu_0^{\beta}B(\nu_0,T)} \frac{2h}{c^2} \left ( \frac{kT}{h} \right ) ^{4+\beta} \Gamma (4+\beta) \zeta(4+\beta),$$ where $h$ and $k$ are Planck’s respectively Boltzmann’s constants, $D_{\rm L}$ is the luminosity distance, and the two last factors are the Gamma function and Riemann’s zeta function. $S(\nu_{\rm obs})$ is the flux density at the observing frequency, and $\nu_0$ is the corresponding rest frequency. The total luminosity of the cold component is found to be $(2.4\pm0.5) \cdot 10^{13}\,{\rm L_\odot}$. To estimate the total IR luminosity, we assume a powerlaw between the peak of the modified blackbody curve and the observed $14.3\,\mu{\rm m}$ point. The powerlaw is integrated from $200\,\mu{\rm m}$ (observed frame), where the modified blackbody and the powerlaw balance eachother, to $14.3\,\mu{\rm m}$ with the result of $(5.4\pm 1)\cdot 10^{12}\,{\rm L_\odot}$. In total the IR luminosity (corrected for gravitational amplification) is then $L_{\rm IR} = (2.9\pm 0.5)\cdot 10^{13}\,{\rm L_\odot}$, dominated by the cold dust component. Using a different method based on the analysis in Blain et al.  ([@blain03]), where the whole IR SED is fit with a single temperature modified blackbody with a powerlaw on the Wien side ranging all the way into the mid-IR, a temperature of 38K is found and a total IR luminosity of $1.8\times 10^{13}\,{\rm L_{\odot}}$. This gives a higher temperature, though a slightly lower luminosity, compared to the fit above where a cold component was fitted to the two SCUBA points. The temperature as we find is lower than that found in other high-$z$ quasars such as APM08279+5255, which has a temperature of $120-220\,$K determined for a pure blackbody (Lewis et al. [@lewis]), or BR1202$-$0725, which has a dust temperature of $50-68\,$K (Leech et al. [@leech]). Both quasars have luminosities in order of $10^{14-15}\,{\rm L_{\odot}}$, thus brighter than SMMJ04135+10277, so that higher dust temperatures might be expected. On the other hand, these two quasars are strongly lensed and it is possible that differential lensing distorts the integrated SED and overemphasizes warm dust components. We finally attempt to compare the radio-submm flux density ratio with the relevant simulations performed by Blain ([@blain]), which are based on the IR-radio correlation observed at low redshift. As we do not have a 1.4GHz flux density measurement, we estimate it by assuming that the radio SED is a power law, $f\propto \nu^{-\alpha}$, with slope $\alpha = -0.8$ and scale it to the observed flux density at 4.86GHz. We find $f_{1.4\,{\rm GHz}} = 595\,\mu{\rm Jy}$ (not corrected for the gravitational lensing). Still assuming $\beta = 1.5$, we use Fig. 4 in Blain ([@blain]) by interpolating between his two models with $T=20\,{\rm K}$ respectively $T=40\,{\rm K}$. For $z=2.837$ this gives a flux density ratio of between $1.4\,$GHz and $850\,\mu$m of $\sim3.5\cdot10^{-3}$. Therefore, the observed $850\,\mu$m flux would imply $f_{1.4\,{\rm GHz}} \sim 88\,\mu{\rm Jy}$ if the quasar strictly followed the local IR-radio correlation. This number is however a factor $6-7$ lower than what we had just estimated based above. This indicates that SMMJ04135+10277 has more radio emission (for its IR emission) than e.g., the ULIRG , which was used for the template SED in Blain (1999). This result is not surprising, as quasar radio emission is powered by both the synschrotron emission from stellar remnants and the synchrotron emission from the central black hole. Conclusions {#sect:concl} =========== We have discovered a type-1 quasar behind the cluster of galaxies A478. The quasar, , was discovered by it submm emission. The quasar has a redshift of $z=2.837$ and is radio quiet. The quasar is optically faint, but has a large submm flux. Using number counts of quasars and of submm sources, we argue that SMM04135+10277 is an unusual object. It is in any case a remarkable object since there is little overlap between deep submm and X-ray samples, suggesting that the incidence of powerful AGNs among submm galaxies is low. The slope of the rest-frame UV continuum is similar to that of optically selected quasars (Francis et al. [@Fran92]) and does not exhibit any signs of extraordinary dust extinction. This leads us to suggest that the line of sight to the quasar nucleus is not abnormally obscured. Analysis of optical continuum, spectral lines and submm emission leads us to conclude that the optical faintness of the quasar results from a large viewing angle from the direction of relativistic beaming, and that a significant amount of the submm flux is not powered by the active nucleus. More likely the cold dust is heated by a high rate of star formation in the environment surrounding the quasar. Comparison of the sparsely sampled IR SED to that of other objects tentatively supports this conclusion. The total IR luminosity is found to be $(2.9\pm0.5)\cdot 10^{13}\,{\rm L_{\odot}}$ and is dominated by the emission from cool dust. We thank Remo Tilanus for taking most of the SCUBA data presented in this paper, Min Yun for providing us with the VLA archive data, and Jean-Paul Kneib for making his LENSTOOL program available for us. We also thank the referee, Andrew Blain, for useful comments. KKK is supported by the Netherlands Organization for Scientific Research (NWO). The JCMT is operated by the Joint Astronomy Centre on behalf of the United Kingdom Particle Physics and Astronomy Research Council (PPARC), the Netherlands Organization for Scientific Research, and the National Research Council of Canada. The National Radio Astronomy Observatory is a facility of the National Science Foundation, operated under cooperative agreement by Associated Universities, Inc. 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. The ISOCAM data presented in this paper was analysed using “CIA”, a joint development by the ESA Astrophysics Division and the ISOCAM Consortium. The ISOCAM Consortium is led by the ISOCAM PI, C. Cesarsky. The ISOPHOT data presented in this paper was reduced using PIA, which is a joint development by the ESA Astrophysics Division and the ISOPHOT consortium, with the collaboration of the Infrared Analysis and Processing Center (IPAC) and the Instituto de Astrofísica de Canarias (IAC). Alexander, D.M., Bauer, F.E., Brandt, W.N., et al., 2003, AJ, 125, 383 Allen, S.W., Fabian, A.C., Johnstone, R.M., et al., 1993, MNRAS, 262, 910 Almaini, O., Scott, S.E., Dunlop, J.S., et al., 2003, MNRAS, 338, 303 Archibald, E.N., Dunlop, J.S., Hughes, D.H., et al., 2001, MNRAS, 323, 417 Aussel, H., Cesarsky, C.J., Elbaz, D., & Starck, J.L., 1999, A&A, 342, 313 Barvainis, R., 1987, ApJ, 320, 537 Bertin, E. & Arnouts, S., 1996, A&A, 117, 193 Blain, A.W., 1999, MNRAS, 309, 955 Blain, A.W., Smail, I., Ivion, R.J., & Kneib, J.-P., 1999, MNRAS, 302, 632 Blain, A.W., Barnard, V.E. & Chapman, S.C., 2003, MNRAS, 338, 733 Carswell, R.F., Mountain, C.M., Robertson, D.J., et al., 1991, ApJ, 381, L5 Cayón, L.,Sanz, J.L., Barreiro, R.B., et al., 2000, MNRAS 315, 757 Cesarsky, C., Abergel, A., Agnese, P., et al., 1996, A&A, 315, L32 Chapman, S.C., Scott, D., Borys, C. & Fahlman, G.G., 2002, MNRAS, 330, 92 Chapman, S.C., Blain, A.W., Ivison, R.J. & Smail, I.R., 2003, Nature, 422, 695 Eales, S., Lilly, S., Gear, W., et al., 1999, ApJ 515, 518 Francis, P.J., Hewett, P.C., Foltz, C.B. & Chaffee, F.H., 1992, ApJ 398, 476 Frayer, D.T., Smail I., Ivison R.J. & Scoville N.Z., 2000, AJ, 120, 1668 Gabriel, C., Acosta-Pulido, J., Heinrichsen, I., Morris, H., & Tai, W.-M., 1997, [*Astronomical Data Analysis Software and Systems VI*]{}, A.S.P. Conference Series, Vol. 125, 1997, G. Hunt and H.E. Payne, eds., p.108 Hagen, H.-J., Engels, D. & Reimers, D., 1999, A&AS, 134, 483 Haas, M., Müller, S.A.H., Chini, R., et al., 2000, A&A, 354, 453 Holland, W.S., Robson, E.I., Gear, W.K., et al., 1999, MNRAS, 303, 659 Hughes, D.H., Dunlop, J.S. & Rawlings, S., 1997, MNRAS, 289, 766 Irwin, M.J., Ibata, R.A., Lewis, G.F. & Totten, E.J., 1998, ApJ, 505, 529 Isaak, K.G., Priddey, R.S., McMahon, R.G., et al., 2002, MNRAS, 329, 149 Ivison, R., Smail, I., Blain, A., Kneib, J.-P. & Frayer, D., 1999, ApSS, 266, 285 Jenness, T. & Lightfoot, J.F., 1998, [*Astronomical Data Analysis Software and Systems VII*]{}, A.S.P. Conference Series, Vol. 145, 1998, R. Albrecht, R.N. Hook and H.A. Bushouse, eds., p.216 Kennefick, J.D., Osmer, P.S., Hall, P.B. & Green, R.F., 1997, AJ, 114, 2269 Kessler, M.F., Steinz, J.A., Anderegg, M.E., et al., 1996, A&A, 315, L27 Kneib, J.-P., Mellier, Y., Fort, B. & Mathez, G., 1993, A&A, 273, 367 Knudsen, K.K., van der Werf, P.P. & Jaffe W., 2001, in “Deep Millimeter Surveys: Implications for Galaxy Formation and Evolution”, eds. J.D. Lowenthal and D.H. Hughes, p168 (astro-ph/0009024) Landolt, A.U., 1992, AJ 104, 340 Leech, K.J., Metcalfe, L. & Altieri, B., 2001, MNRAS, 328, 1125 Lemke, D., Klaas, U., Abolins, J., et al., 1996, A&A, 315, L64. Lewis, G.F., Chapman, S.C., Ibata, R,A., Irwin, M.J. & Totten, E.J., 1998, ApJL, 505, 1 Ott, S., Abergel, A., Altieri, B., et al., 1997, [*Astronomical Data Analysis Software and Systems VI*]{}, A.S.P. Conference Series, Vol. 125, 1997, G. Hunt and H.E. Payne, eds., p.34 Page, M.J., Stevens, J.A., Mittaz, J.P.D. & Carrera, F.J., 2001, Science, 294, 2516 Priddey, R.S., Isaak, K.G., McMahon, R.G. & Omont, A., 2002, astro-ph/0211646 Sanders, D.B., Phinney, E.S., Neugebauer, G., Soifer, B.T. & Matthews, K., 1989, ApJ, 347, 29 Schlegel, D.J., Finkbeiner, D.P. & Davis, M., 1998, ApJ 500, 525 Scott, S.E., Fox, M.J. & Dunlop, J. S., 2002, MNRAS, 331, 817 Smail, I., Ivison, R.J. & Blain, A.W., 1997, ApJ, 490, L5 Smail, I., Ivison, R.J., Kneib, J.-P., et al., 1999, MNRAS, 308, 1061 Smail, I., Ivison, R.H., Blain, A.W. & Kneib, J.-P., 2002, MNRAS, 331, 495 Stocke, J.T., Morris, S.L., Weymann R.J. & Foltz, C.B., 1992, ApJ, 396, 487 Urry, C.M., & Padovani, P., 1995, PASP, 107, 803 Vernet, J. & Cimatti, A., 2001, A&A, 380, 409 Zabludoff, A.I., Huchra, J.P. & Geller, M.J., 1990, ApJS, 74, 1 [^1]: Based on observations made with ESO Telescopes at the Paranal Observatory under programme IDs 63.O-0087 and 68.A-0111. Also based on observations with ISO, an ESA project with instruments funded by ESA member states (especially the PI countries: the Netherlands, the United Kingdom, Germany, and France) with the participation of ISAS and NASA.
{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a multi-mode model to describe an arbitrary N-photon state with a wide spectral range and some arbitrary temporal distribution. In general, some of the $N$ photons are spread out in time while other may overlap and become indistinguishable. From this model, we find that the temporal (in)distinguishability of photons is related to the exchange symmetry of the multi-photon wave function. We find that simple multi-photon detection scheme gives rise to a more general photon bunching effect with the famous two-photon effect as a special case. We then send this N-photon state into a recently discovered multi-photon interference scheme. We calculate the visibility of the multi-photon interference scheme and find that it is related to the temporal distinguishability of the $N$ photons. Maximum visibility of one is achieved for the indistinguishable N-photon state whereas the visibility degrades when some of the photons are separated and become distinguishable. Thus we can identify an experimentally measurable quantity that may quantitatively define the degree of indistinguishability of an N-photon state. This presents a quantitative demonstration of the complementary principle of quantum interference.' author: - 'Z. Y. Ou' title: | Temporal Distinguishability of an N-Photon State\ and Its Characterization by Quantum Interference --- \[sec:level1\]Introduction ========================== The coherence properties of an optical field are best described by the field correlation function in space and time [@mw]. Most commonly used quantity to characterize the coherence property of an optical field is the coherence time or coherence length for temporal coherence. Roughly speaking, the coherence length of an optical field is the distance within which the field can be described as a single uninterrupted wave train. In other words, any two points within the coherence length will have a fixed phase relationship. However, this description is primarily concerned with wave aspect of an optical field and is based on the interference effect observed in intensity or single photon interference effect. More specifically in terms of the quantum coherence theory [@glau], it is related to the field correlation function of $$\begin{aligned} \Gamma(\tau) = \langle \hat E^{(-)}(t+\tau) \hat E^{(+)}(t) \rangle,\label{1}\end{aligned}$$ where $$\begin{aligned} \Big[\hat E^{(-)}\Big]^{\dag} = \hat E^{(+)}(t) ={1\over\sqrt{2\pi}}\int d\omega\hat a(\omega) e^{-i\omega t}\label{E}\end{aligned}$$ for a quasi-monochromatic field [@man66] and the average is over the quantum state of the field. The visibility of the single-photon interference fringes is simply the absolute value of the normalized field correlation function [@bw]: $$\begin{aligned} \gamma(\tau) = \Gamma(\tau)/\Gamma(0). \label{2}\end{aligned}$$ However, this description becomes rudimentary when we start to deal with the cases involving more than one photon in quantum information. One may use a higher order correlation function such as the intensity correlation function [@glau] $$\begin{aligned} &&\Gamma^{(N)}(t_1,t_2,...,t_N) \cr &&\hskip 0.2in = \langle \hat E^{(-)}(t_1)...\hat E^{(-)}(t_N) \hat E^{(+)}(t_N)...\hat E^{(+)}(t_1) \rangle.~~~~~\label{Gma-N}\end{aligned}$$ which is related to an N-photon coincidence measurement. However, this function does not provide any information about photon entanglement, i.e., quantum superposition of different states. The realization of multi-particle entanglement is paramount in achieving most of the tasks in quantum computing and quantum information processing [@sho; @gro]. While there are many ways to create entangled multi-particle state, the straightforward method is to start from independent single photons [@hof]. Knill, Laflamme, and Milburn [@klm] have shown that quantum computing can be realized with single photons and some linear optical elements via multi-photon interference. N-photon entanglement is thus produced from single-photon states. This is one of the primary reasons behind the big rush in creating light sources with single-photon on demand [@mic2; @san]. While most of the analysis are based on the single mode model, i.e., all the photons in one single temporal mode, this is, on the other hand, impossible to achieve in experiment. The multi-mode nature of light inevitably reduces the effect of photon interference and leads to degradation in information processing. One often uses the fidelity quantity of quantum states to characterize the degradation. But this description has emphasis only on the end result of the process and spares the true culprit of the process, that is, the multi-mode nature of light. For monochromatic field of only one frequency component, the field can be represented by an infinite wave train. Photons can appear anywhere in this wave train and are indistinguishable from each other. They will produce maximum effect of entanglement. However, when many frequency components are excited, an optical field is no longer monochromatic and the wave train becomes finite with a length of the order of the coherence length of the field. With multiple photons, we generally cannot use a single wave packet to describe them. We cannot assign separate wave packets to describe each photon, either. This is because of the possibility of multi-photon entanglement. Thus, an issue is raised about how to describe the different situations of temporal distribution of photons and distinguish these situations experimentally. Recently, this issue was addressed in the four-photon case [@tsu; @ou2] for distinguishing a genuine four-photon polarization entangled state from a state made of two well separated pairs of photons. The difference lies in the multi-photon interference: an entangled four-photon state will give rise to the strongest multi-photon interference effect whereas two well separate pairs produce less interference effect. This is in consistence with the complementary principle of quantum mechanics which states that quantum interference is a result of indistinguishability of the paths but if the the paths are distinguishable, the interference effect will be gone. Partial distinguishability will lead to reduced interference effect, as described by Eq.(\[2\]) in the coherence theory for the single-photon interference. Four-photon interference experiments were performed to distinguish an entangled four-photon state from two independent pairs of photons [@tsu; @ou2; @rhe1; @rhe2] However, the above mentioned interference scheme on the four-photon state cannot be generalized to arbitrary photon number. More recently, Sun [*et al*]{} [@sun1; @sun2] and Resch [*et al*]{} [@res] independently constructed a quantum state projection measurement scheme and applied it to a maximally entangled N-photon state (the so-called NOON state) for the demonstration of multi-photon de Broglie wavelength without a NOON state. It turns out that this new projection measurement scheme is based on a multi-photon interference effect that depends on the temporal distribution of the photons involved. Since the new scheme can be easily generalized to arbitrary photon number, it can be used to study the relation between the multi-photon interference effect and the temporal distinguishability of an N-photon state. We will show that the various scenarios of temporal distribution of photons give rise to different visibility in the multi-photon interference, which provides a direct measure of the degree of temporal distinguishability of a multi-photon state in a similar fashion to the coherence theory \[Eq.(\[2\])\]. This is a quantitative investigation into the complementary principle of quantum interference. In the following, we will first review the two-photon and four-photon cases to look for the relation between temporal distinguishability and multi-photon interference. We then will generalize to an arbitrary N-photon state and present the criteria for photon indistinguishability and distinguishability. In Sect.IV, we use quantum coherence theory to calculate the result from a direct N-photon coincidence measurement and discuss the generalized photon bunching effect. This measurement process is not sensitive to the different temporal distribution of the photons. In Sect.V, we introduce the newly constructed NOON state projection measurement and demonstrate how it can be used to characterize the degree of temporal indistinguishability for the simple three-photon case. We will generalize the discussion for three-photon case to the general N+1-photon case. In Sect.VI, we will discuss an even more general case and present the numerical results for a few special cases. We conclude with a discussion. Temporal distinguishability for the case of two photons and for the case of two-pairs of photons ================================================================================================ The first discussion about the temporal distinguishability was by Grice and Walmsley [@wam], who investigated the visibility in a Hong-Ou-Mandel interferometer [@hom] with a two-photon state input from type-II parametric down-conversion. Later on, Atatüre et al [@ser] performed experiment and confirmed the degradation of the two-photon interference visibility predicted in Ref.[@wam] due to temporal distinguishability. In the discussion of Ref.[@wam], the multi-mode description of the two-photon state is given by $$\begin{aligned} |\Phi_{2}\rangle = \int d\omega_1d\omega_2\Phi(\omega_1,\omega_2)\hat a_s^{\dag}(\omega_1)\hat a_i^{\dag}(\omega_2)|0\rangle,\label{phi-st}\end{aligned}$$ where $s, i$ denote the two correlated signal and idler photons from parametric down-conversion. For type-II process, we have $\Phi(\omega_1,\omega_2) \ne \Phi(\omega_2,\omega_1)$ due the birefringent effect of the nonlinear crystal on the ordinary and extra-ordinary rays. The maximum visibility in the two-photon Hong-Ou-Mandel interferometer has the form of $$\begin{aligned} {\cal V}_2 = M_2 \equiv {\int d\omega_1 d\omega_2 \Phi^*(\omega_1, \omega_2)\Phi(\omega_2, \omega_1)\over \int d\omega_1 d\omega_2 |\Phi(\omega_1, \omega_2)|^2}.\label{9-2}\end{aligned}$$ $M_2$ is defined as a degree of permutation symmetry. Note that $M_2 =M_2^*$ and $0\le |M_2|\le 1$. The visibility or the degree of permutation symmetry is one if and only if $\Phi(\omega_1,\omega_2)$ satisfies the permutation symmetry relation: $$\begin{aligned} \Phi(\omega_1, \omega_2)=\Phi(\omega_2, \omega_1).\label{9-3}\end{aligned}$$ As stated in Ref.[@wam], this permutation relation is a signature of spectral indistinguishability of the two photons, that is, we cannot tell the difference between the two photons through their spectra. This in turn gives temporal indistinguishability if we consider the Fourier transformation: $$\begin{aligned} G(t_1, t_2) = {1\over 2\pi}\int d\omega_1d\omega_2 \Phi(\omega_1, \omega_2) e^{-i(\omega_1t_1+\omega_2t_2)}.\label{9-4}\end{aligned}$$ Combination of Eqs.(\[9-3\], \[9-4\]) gives directly the symmetric relation: $$\begin{aligned} G(t_1, t_2)=G(t_2, t_1),\label{9-5}\end{aligned}$$ for all times of $t_1, t_2$. On the other hand, the visibility is zero if $\Phi(\omega_1, \omega_2)$ does not have any overlap with $\Phi(\omega_2, \omega_1)$, which is characterized by the orthogonal relation: $$\begin{aligned} \int d\omega_1 d\omega_2 \Phi^*(\omega_1, \omega_2)\Phi(\omega_2, \omega_1) = 0\label{9-6}\end{aligned}$$ or in time $$\begin{aligned} \int dt_1dt_2 G^*(t_1, t_2)G(t_2, t_1) = 0.\label{9-7}\end{aligned}$$ This orthogonal relation indicates that the two functions $G(t_1, t_2), G(t_2, t_1)$ have no overlap. At this point, it is not easy to see what is the physical meaning of Eq.(\[9-7\]). However, if we go back to Eq.(\[9-3\]) and introduce a non-symmetric factor of $e^{i\omega_2T}$, we find that the equivalent $\Phi(\omega_1, \omega_2)$ in Eq.(\[phi-st\]) in this case will be $\Phi^{\prime}(\omega_1, \omega_2)\equiv \Phi(\omega_1, \omega_2)e^{i\omega_2T}$, which is not symmetric with respect to $\omega_1,\omega_2$ even if $\Phi(\omega_1, \omega_2)$ is symmetric. This extra phase can be introduced by acting the evolution operator $\hat U(T) = \exp(-i\omega_2\hat a_i^{\dag}\hat a_i T)$ on the state in Eq.(\[phi-st\]) for an extra free propagation time $T$ of the idler photon. This then creates a time delay $T$ between the two photons. Then the visibility in Eq.(\[9-2\]) becomes $$\begin{aligned} {\cal V}_2(T) = {\int d\omega_1 d\omega_2 \Phi^*(\omega_1, \omega_2)\Phi(\omega_2, \omega_1)e^{i(\omega_1-\omega_2)T}\over \int d\omega_1 d\omega_2 |\Phi(\omega_1, \omega_2)|^2}.\label{v-p}\end{aligned}$$ Notice that if the delay is large enough \[$T >> T_c\sim 1/\Delta\omega_{PDC}$ with $\Delta\omega_{PDC}$ as the range of $\Phi(\omega_1, \omega_2)$\], we will have ${\cal V}_2(T) = 0$ or $\Phi^{\prime}(\omega_1, \omega_2)$ satisfies Eq.(\[9-6\]). Since $T$ is the relative delay between the two photons before they meet at the beam splitter of the Hong-Ou-Mandel interferometer, we may believe that there is a large enough delay between the two photons so that the two photons become distinguishable in time when they arrive at the beam splitter. So the orthogonal relation in Eq.(\[9-6\]) or Eq.(\[9-7\]) corresponds to the situation when the two photons are well separated in time and form two non-overlapping and distinguishable wave packets. Therefore, the visibility in the Hong-Ou-Mandel interferometer in Eq.(\[9-2\]) is a direct measure of temporal distinguishability of the two photons. This is very much similar to the role of the field correlation function $\gamma$ of Eq.(\[2\]) in defining optical coherence of a field. For the four-photon case, temporal distinguishability between two pairs of photons was first studied by Ou, Rhee and Wang [@rhe1; @rhe2] in a similar scheme as the Hong-Ou-Mandel interferometer but with four photons. It was found that the visibility in four-photon interference is directly related to a quantity ${\cal E/A}$, which is a measure of the temporal distinguishability of photon pairs from parametric down-conversion: when ${\cal E/A} << 1$, the pairs are well separated from each other corresponding to the so-called $2\times 2$ case but when ${\cal E/A}=1$, the two pairs are overlap in time and form an indistinguishable four-photon state corresponding to the $4\times 1$ case. From the definition of the quantities ${\cal E}$ and ${\cal A}$ in Ref.[@rhe2], we rewrite them as $$\begin{aligned} &&{\cal E} = \int d\omega_1d\omega_2d\omega_1^{\prime}d \omega_2^{\prime}\Phi^*(\omega_1, \omega_2)\Phi^* (\omega_1^{\prime}, \omega_2^{\prime})\cr &&\hskip 1.2in \times \Phi(\omega_1^{\prime}, \omega_2)\Phi(\omega_1, \omega_2^{\prime}), \label{8-181}\end{aligned}$$ and $$\begin{aligned} {\cal A} = \int d\omega_1d\omega_2d\omega_1^{\prime}d \omega_2^{\prime}|\Phi(\omega_1, \omega_2)\Phi (\omega_1^{\prime}, \omega_2^{\prime})|^2, \label{8-182}\end{aligned}$$ where $\Phi(\omega_1, \omega_2)$ is the two-photon wave function in Eq.(\[phi-st\]). On the other hand, the four-photon state from Ref.[@rhe2] has the form of $$\begin{aligned} |\Phi_{4}\rangle &=& \int d\omega_1d\omega_2d\omega_1^{\prime}d \omega_2^{\prime}\Phi_4(\omega_1,\omega_2;\omega_1^{\prime}, \omega_2^{\prime})\cr &&\hskip 0.4in \times\hat a_s^{\dag}(\omega_1)\hat a_i^{\dag}(\omega_2)\hat a_s^{\dag}(\omega_1^{\prime})\hat a_i^{\dag}(\omega_2^{\prime})|0\rangle,\label{phi4-st}\end{aligned}$$ where $\Phi_4(\omega_1,\omega_2;\omega_1^{\prime}, \omega_2^{\prime})\equiv \Phi(\omega_1,\omega_2)\Phi(\omega_1^{\prime}, \omega_2^{\prime})$. Then we can rewrite the expression for ${\cal E}$ and ${\cal A}$ in Eqs.(\[8-181\], \[8-182\]) and obtain the quantity ${\cal E/A}$ as $$\begin{aligned} {\cal E \over A} = {\int d\omega_1d\omega_2d\omega_1^{\prime}d \omega_2^{\prime}\Phi_4^*(\omega_1, \omega_2,\omega_1^{\prime}, \omega_2^{\prime}) \Phi_4(\omega_1^{\prime}, \omega_2;\omega_1, \omega_2^{\prime})\over \int d\omega_1d\omega_2d\omega_1^{\prime}d \omega_2^{\prime}|\Phi_4(\omega_1, \omega_2;\omega_1^{\prime}, \omega_2^{\prime})|^2}.\label{e-a}\end{aligned}$$ Recall that this quantity is a measure of the temporal distinguishability of two pairs of photons. But from Eq.(\[e-a\]), we find that this quantity is again dependent on the permutation of the wave function $\Phi_4(\omega_1,\omega_2;\omega_1^{\prime}, \omega_2^{\prime})$ similar to that in Eq.(\[9-2\]) and it is one if and only if we have the permutation symmetry of $$\begin{aligned} \Phi_4(\omega_1, \omega_2,\omega_1^{\prime}, \omega_2^{\prime}) =\Phi_4(\omega_1^{\prime}, \omega_2;\omega_1, \omega_2^{\prime}).\label{sym4}\end{aligned}$$ Therefore from the discussion on the meaning of the quantity ${\cal E/A}$, we find that the symmetry relation in Eq.(\[sym4\]) corresponds to the case when the two pairs are completely overlap in time and become temporally indistinguishable (the $4\times 1$ case) whereas the orthogonal relation $$\begin{aligned} &&\int d\omega_1d\omega_2d\omega_1^{\prime}d \omega_2^{\prime}\Phi_4^*(\omega_1, \omega_2,\omega_1^{\prime}, \omega_2^{\prime})\cr && \hskip 1in \times \Phi_4(\omega_1^{\prime}, \omega_2;\omega_1, \omega_2^{\prime})=0\label{orth4}\end{aligned}$$ leads to the case of completely separated pairs of photons (the $2\times 2$ case). From the experiments and analysis on four-photon interference with two pairs of photons by parametric down-conversion [@rhe1; @rhe2; @sun1], we find that the visibility is not zero even for ${\cal E/A} =0$. This can be attributed to the existence of two-photon interference since we usually have two-photon indistinguishability with exchange symmetry in Eq.(\[9-3\]). Note that ${\cal E/A}$ concerns the permutation symmetry between two different pairs, i.e., exchange between the group of $\{\omega_1,\omega_2\}$ and the group of $\{\omega_1^{\prime},\omega_2^{\prime}\}$. The exchange within each group is symmetric due to Eq.(\[9-3\]). Next we will generalize Eqs.(\[9-3\], \[sym4\]) and Eqs.(\[9-6\], \[orth4\]) to an arbitrary $N$-photon case and relate them to the visibility of some $N$-photon interference experiment. Description of a temporally distributed N-photon state ====================================================== Now we can generalize Eqs.(\[9-3\], \[9-6\]) of the two-photon case and Eqs.(\[sym4\], \[orth4\]) of the two-pair case to arbitrary $N$ case. An arbitrary N-photon state of wide spectral range can be generally described by $$\begin{aligned} &&|\Phi_N\rangle = {\cal N}^{-1/2}\int d\omega_1d\omega_2...d\omega_N \Phi(\omega_1, ..., \omega_N)\times\cr &&\hskip 1.0in \times\hat a^{\dag}(\omega_1)\hat a^{\dag}(\omega_2)...\hat a^{\dag}(\omega_N)|0\rangle,\label{Phi-state}\end{aligned}$$ where the normalization factor ${\cal N}$ is given by $$\begin{aligned} &&{\cal N} = \int d\omega_1d\omega_2...d\omega_N \Phi^*(\omega_1, ..., \omega_N)\times\cr &&\hskip 1.0in\times\sum_P \Phi(P\{\omega_1, ..., \omega_N\}),\label{N}\end{aligned}$$ where $P$ is the permutation operator on the indices of 1,2,..., $N$. and the sum is over all possible permutation. There are totally $N!$ terms. So the value of ${\cal N}$ ranges from $I$ to $N!I$ with $I = \int d\omega_1d\omega_2...d\omega_N |\Phi(\omega_1, ..., \omega_N)|^2$. The maximum value of $N!I$ is reached when $$\begin{aligned} \Phi(\omega_1, ..., \omega_N) = \Phi(P\{\omega_1, ..., \omega_N\})\end{aligned}$$ for all $P$. Similar to Eqs.(\[9-3\], \[sym4\]), this corresponds to a case when the $N$ photons are indistinguishable in time. We refer to this case as the $N\times 1$ case, meaning that all $N$ photons are in one indistinguishable temporal mode. This single-mode description of an N-photon state is more vivid in the special case when $\Phi(\omega_1, ..., \omega_N)$ is factorized as $\phi(\omega_1)\phi(\omega_2)...\phi(\omega_N)$ and the N-photon state simply becomes $$\begin{aligned} &&|\Phi_N\rangle = {1\over N!} \hat A(\phi)^{\dagger N}|0\rangle = |N\rangle_{\phi}\end{aligned}$$ with $$\begin{aligned} \hat A(\phi) = \int d\omega \phi(\omega)\hat a (\omega) ~~~~(\int d\omega |\phi(\omega)|^2 = 1).\end{aligned}$$ Note that $\hat A(\phi)$ satisfies $[\hat A, \hat A^{\dag}]=1$ and represents the annihilation operator for a single temporal mode characterized by $\phi(\omega)$. The single-photon state $|1\rangle_{\phi}$ has a single-photon detection probability of $|g(\tau)|^2$ with a temporal shape of $$\begin{aligned} g(\tau) = {1\over \sqrt{2\pi}}\int d\omega \phi(\omega)e^{-i\omega t}\label{g}\end{aligned}$$ and normalization relation $$\begin{aligned} \int d\tau|g(\tau) |^2 = 1.\label{g-norm}\end{aligned}$$ The other extreme case of ${\cal N} = I$ requires $\Phi(\omega_1,...,\omega_N)$ be orthogonal to all the permuted functions $\Phi(P\{\omega_1,...,\omega_N\})$ in the similar ways in Eqs.(\[9-6\], \[orth4\]) and thus corresponds to the situation when all photons are well separated in time. We refer to this case as the $1\times N$ case, meaning that each photon is in its separate temporal mode and there are totally $N$ independent modes. For the situations in between the two extreme cases, the value of ${\cal N}$ is between $I$ and $N!I$. For example, assume that the spectral amplitude $\Phi(\{\omega\})$ have partial permutation symmetry, that is, $$\begin{aligned} \Phi(\omega_1, ..., \omega_N) = \Phi(P_{\{n_i\}}\{\omega_1, ..., \omega_N\}),\label{sym}\end{aligned}$$ where the permutation $P_{\{n_i\}}$ only applies to a subgroup of $\{\omega_1, \omega_2,...,\omega_N\}$. In the meantime, it also satisfies the orthogonal relations: $$\begin{aligned} \int d\omega_1...d\omega_N\Phi^*(\omega_1, ..., \omega_N) \Phi(P_{ij}\{\omega_1, ..., \omega_N\})=0\label{orth}\end{aligned}$$ for permutation $P_{ij}$ between different subgroups ($\{n_i\}$ and $\{n_j\}$, $i\ne j$) defined in Eq.(\[sym\]). Then it can be easily shown that ${\cal N} = n_1!n_2!...n_k!I$. In the simple case when $\Phi(\omega)$ can be factorized as $$\begin{aligned} &&\Phi(\omega_1, ..., \omega_N) = \phi_1(\omega_1)...\phi_1(\omega_{n_1}) \phi_2(\omega_{n_1+1})\times \cr &&\hskip 1.2in \times...\phi_2(\omega_{n_1+n_2}) ...\phi_k(\omega_N)~~~~~~\end{aligned}$$ with the orthogonal relations $$\begin{aligned} \int d\omega_1d\omega_2\phi_i^*(\omega_1)\phi_j^*(\omega_2) \phi_i(\omega_2)\phi_j(\omega_1)=0~~(i\ne j),\label{orth2}\end{aligned}$$ the N-photon state in Eq.(\[Phi-state\]) becomes $$\begin{aligned} &&|\Phi_N\rangle = {1\over n_1!}|n_1\rangle_{\phi_1} {1\over n_2!}|n_2\rangle_{\phi_2}... {1\over n_k!}|n_k\rangle_{\phi_k}.\label{Phi-state2}\end{aligned}$$ This is the situation when the $N$ photons are divided into $k$ subgroups with $n_i (i=1,2,...,k)$ photons in each group in a single temporal mode characterized by $\phi_i$. This situation is denoted as $n_1+...+n_k$ case. For simplicity of later argument, let us consider another special kind of N-photon state with $$\begin{aligned} \Phi(\omega_1, ..., \omega_N) = \phi(\omega_1)e^{i\omega_1T_1}...\phi(\omega_N)e^{i\omega_NT_N}.\label{Phi-T}\end{aligned}$$ With this $\Phi$, the N-photon state can be viewed as direct product of $N$ identical single photon wave packets: $$\begin{aligned} |N\rangle_T = |T_1\rangle\otimes|T_2\rangle\otimes...\otimes|T_N\rangle,\label{Nprod}\end{aligned}$$ with $$\begin{aligned} |T_i\rangle = \int d\omega \phi(\omega)e^{i\omega T_i}\hat a^{\dag}(\omega)|0\rangle.\end{aligned}$$ This state can be viewed as from single-photon sources such as quantum dots (see below for details). However, the quantum state in Eq.(\[Nprod\]) is not normalized. Substituting Eq.(\[Phi-T\]) into Eq.(\[N\]), we have the normalization factor as $$\begin{aligned} &&{\cal N} = \int d\omega_1d\omega_2...d\omega_N \Big[\prod_k|\phi(\omega_k)|^2e^{-i\omega_kT_k}\Big]\times\cr &&\hskip 0.7in\times\sum_P P\Big[\exp\big\{\sum_m i\omega_mT_m\big\}\Big].\label{NT}\end{aligned}$$ When $T_1=T_2=...=T_N$, we recover the case when all $N$ photons are in one single temporal mode with ${\cal N} =N!$ ($N\times 1$). On the other hand, if $|T_i-T_j|>> 1/\Delta\omega (i\ne j)$ with $\Delta\omega$ as the bandwidth of $\phi(\omega)$, we have ${\cal N} =1$. This is the case when all the photons are well separated from each other ($1\times N$). The N-photon state in Eq.(\[Phi-state\]) describes a state when all photons are in one spatial and polarization mode. They only differ in spectral mode. In practice, although $N=2$ case can be easily obtained from degenerate parametric down-conversion, such a state with $N >2$ is not easy to produce directly. It can be produced indirectly from single-photon states with a set of beam splitters as shown in Fig.1, where the single-photon sources are, for example, quantum dots. The quantum state for the input fields has the general form of $$\begin{aligned} &&|\Psi_N\rangle_{in} = \int d\omega_1d\omega_2...d\omega_N \Phi(\omega_1, ..., \omega_N)\times\cr &&\hskip 1.0in \times\hat a_1^{\dag}(\omega_1)\hat a_2^{\dag}(\omega_2)...\hat a_N^{\dag}(\omega_N)|0\rangle ~~~~~~~\label{Phi-state-ND}\end{aligned}$$ with the normalization relation: $$\begin{aligned} \int d\omega_1d\omega_2...d\omega_N \Phi^*(\omega_1, ..., \omega_N)\Phi(\omega_1, ..., \omega_N) =1. \label{norm-ND}\end{aligned}$$ Here $\hat a_j^{\dag} (j=1,...,N)$ is the creation operator for each input mode. Photons are possible to exit at any of the $N$ output ports. To produce a state of the form in Eq.(\[Phi-state\]), however, we only consider the possibility when all $N$ photon exit at one port, say, $b_N$ port. It is straightforward using the beam splitter theory to show that the projected state is $$\begin{aligned} &&\mathbb{P}|\Psi_N\rangle_{out} = {1\over N^{N/2}}\int d\omega_1d\omega_2...d\omega_N \Phi(\omega_1, ..., \omega_N)\times\cr &&\hskip 1.1in \times\hat b_N^{\dag}(\omega_1)\hat b_N^{\dag}(\omega_2)...\hat b_N^{\dag}(\omega_N)|0\rangle, ~~~~~~~\label{Phi-state-ND}\end{aligned}$$ which is in the form of Eq.(\[Phi-state\]). This state is not normalized because it is a projected state with the probability of projection as $P(|\Phi_N\rangle) =||\mathbb{P}|\Psi_N\rangle_{out}||^2 = {\cal N}/N^N$. The delay factors $\{e^{i\omega_jT_j}\}$ in Eq.(\[Phi-T\]) can be easily introduced on individual mode $\hat a_j$ before the beam splitters via the free-field evolution operator $\hat U_j(T_j) = \exp(-i\omega_j\hat a^{\dag}_j\hat a_j T_j)$. More generally, to include different spatial and polarization modes, the N-photon state has the following shape $$\begin{aligned} |\Phi_N\rangle = {\cal N}_k^{-1/2}\int d\omega_1^{(1)}...d\omega_{n_1}^{(1)}...d\omega_1^{(k)}...d\omega_{n_k}^{(k)} \Phi(\{\omega^{(1)}\}, ..., \{\omega^{(k)}\})\hat a^{\dag}_1(\omega_1^{(1)})...\hat a^{\dag}_1(\omega_{n_1}^{(1)})...\hat a^{\dag}_k(\omega_{n_k}^{(k)})|0\rangle,~~~~~~~~~\label{Phi-state-SP}\end{aligned}$$ where $\{\omega^{(1)}\} = \omega^{(1)}_1,...,\omega^{(1)}_{n_1}$, etc. The normalization factor ${\cal N}_k$ takes the form of $$\begin{aligned} {\cal N}_k = \int d\{\omega^{(1)}\}...d\{\omega^{(k)}\} \Phi^*(\{\omega^{(1)}\},...,\{\omega^{(k)}\})\sum_{P_1,...,P_k} \Phi(P_1\{\omega^{(1)}\}, ..., P_k\{\omega^{(k)}\}).\label{Nk}\end{aligned}$$ ${\cal N}_k$ now ranges from $I$ to $n_1!...n_k!I$. The special case when $\Phi(\{\omega^{(1)}\}, ..., \{\omega^{(k)}\})$ factorizes is similar as before. Direct N-Photon Measurement: Photon Bunching Effect for N Photons ================================================================= Next, we consider an N-photon joint measurement with the joint probability density given from the quantum coherence theory in Eq.(\[Gma-N\]). The average is over the quantum state of the system given in Eq.(\[Phi-state\]) for an arbitrary N-photon state. For simplicity, we first apply it to the state in Eq.(\[Nprod\]). To carry out the quantum average, it is easier to first find the N-photon detection probability amplitude: $$\begin{aligned} &&C^{(N)}(t_1,t_2,...,t_N) = \langle 0| \hat E^{(+)}(t_N)...\hat E^{(+)}(t_1)|\Phi_N \rangle.~~~~~\label{CN}\end{aligned}$$ Then $\Gamma^{(N)}(t_1,t_2,...,t_N) = |C^{(N)}(t_1,t_2,...,t_N)|^2$. From Eq.(\[E\]) for the field operator and Eq.(\[Phi-T\]) for $\Phi$, it is straightforward to obtain $$\begin{aligned} &&C^{(N)}(t_1,t_2,...,t_N) \cr &&\hskip 0.5in = \sum_P P[g(t_1-T_1)...g(t_N-T_N)],\end{aligned}$$ where the permutation operation $P$ is on $t_1t_2...t_N$ and there are $N!$ terms in the sum. The overall probability of detecting $N$ photons together (N-photon coincidence) is an integral of $\Gamma^{(N)}(t_1,t_2,...,t_N)$ over all times $t_1,...,t_N$: $$\begin{aligned} &&P_N = \int dt_1...dt_N \Gamma^{(N)}(t_1,t_2,...,t_N) \cr &&\hskip 0.23in = \int dt_1...dt_N \bigg|\sum_P P[g(t_1-T_1)...g(t_N-T_N)]\bigg|^2.\cr &&\label{PN}\end{aligned}$$ In the extreme case when $T_1 = T_2 = ... = T_N$, we obtain $P_N{(N\times 1)} = (N!)^2 I$ while in the other extreme case when $|T_i-T_j|>>1/\Delta\Omega$, we have $P_N{(1\times N)}= N!I$. Therefore, we seem to have $$\begin{aligned} P_N{(N\times 1)} = N! P_N(1\times N),\end{aligned}$$ or $$\begin{aligned} P_N{(N\times 1)}/ P_N(1\times N) = N!,\end{aligned}$$ that is, the N-photon detection probability is $N!$ larger in the case of $N$ identical photons than in the case of $N$ separated photons. This can be thought of as the Bosonic photon bunching effect for $N$ photons. The case of $N=2$ gives the familiar photon bunching factor of 2. However, as we know, the N-photon state in Eq.(\[Nprod\]) is not normalized. With the normalization factor considered, we have instead $$\begin{aligned} P_N(N\times 1) = P_N(1\times N) = N!.\label{PN-N}\end{aligned}$$ For the case in between the two extreme cases, we may evaluate Eq.(\[PN\]) as $$\begin{aligned} P_N = \int dt_1...dt_N \sum_{P^{\prime}} P^{\prime}[g^*(t_1-T_1)...g^*(t_N-T_N)] \sum_P P[g(t_1-T_1)...g(t_N-T_N)]. \label{PN2}\end{aligned}$$ Since the sum is over all permutations, the integral does not change if we make the variable change: $\{t_1...t_N\}\rightarrow P^{\prime}\{t_1...t_N\}$, i.e., $$\begin{aligned} &&P_N = \sum_{P^{\prime}} \int dt_1...dt_N g^*(t_1-T_1)...g^*(t_N-T_N) \sum_P P[g(t_1-T_1)...g(t_N-T_N)]\cr &&\hskip 0.23in = N! \int dt_1...dt_N g^*(t_1-T_1)...g^*(t_N-T_N) \sum_P P[g(t_1-T_1)...g(t_N-T_N)]. \label{PN3}\end{aligned}$$ It can be further shown that $$\begin{aligned} \int dt_1...dt_N g^*(t_1-T_1)...g^*(t_N-T_N) \sum_P P[g(t_1-T_1)...g(t_N-T_N)] = {\cal N},\label{PN4a}\end{aligned}$$ where $\cal N$ is given in Eq.(\[NT\]). Thus we have $$\begin{aligned} P_N = N!\cal N.\label{PN4}\end{aligned}$$ For a normalized N-photon state, we have $P_N = N!$ in any case as in Eq.(\[PN-N\]). For the multi-spatial and polarization state in Eq.(\[Phi-state-SP\]), we may find $P_N$ after some lengthy manipulation as that leads to Eq.(\[PN4\]): $$\begin{aligned} P_N = n_1!...n_k!{\cal N}_k\label{PN5}\end{aligned}$$ for the un-normalized state and $P_4= n_1!...n_k!$ for the normalized state. Hence, it is impossible to characterize different cases of temporal entanglement with just simple direct multi-photon detection for the normalized state. Furthermore, even for the un-normalized state, we cannot explore the temporal indistinguishability among different spatial and polarization modes with multi-photon detection, for ${\cal N}_k$ depends only on the permutation symmetry within photons in one spatial and polarization mode. Before we proceed further, it is interesting to evaluate the multi-photon detection rates in some special cases. For example, for the single-photon detection rate $P_1$, we have $$\begin{aligned} &&P_1 = \int dt \langle\Phi_N|\hat E^{\dagger}(t) \hat E(t)|\Phi_N\rangle.\end{aligned}$$ With some manipulation, it can be shown that $P_1 = N{\cal N}$ for an un-normalized N-photon state and $P_1 = N$ for a normalized N-photon state. The reason that we still discuss the un-normalized case of an N-photon state is because we encounter this kind of state in practice when a projection measurement is involved such as that in Fig.1. Consider, for example, a multi-photon state from degenerate parametric down-conversion, which, for small $\eta$, has the form of [@sun1] $$\begin{aligned} &&|\Phi_{PDC}\rangle = C\Big(|0\rangle + \eta |\Phi_{2D}\rangle +{\eta^2\over 2}|\Phi_{4D}\rangle +...\Big),\label{PDC}\end{aligned}$$ with $$\begin{aligned} |\Phi_{2D}\rangle = \int d\omega_1d\omega_2\Phi(\omega_1,\omega_2)\hat a^{\dag}(\omega_1)\hat a^{\dag}(\omega_2)|0\rangle\end{aligned}$$ and $$\begin{aligned} &&|\Phi_{4D}\rangle = \int d\omega_1d\omega_2d\omega_1'd\omega_2' \Phi(\omega_1,\omega_2)\Phi(\omega_1',\omega_2')\times\cr&&\hskip 0.8in\times \hat a^{\dag}(\omega_1)\hat a^{\dag}(\omega_2)\hat a^{\dag}(\omega_1')\hat a^{\dag}(\omega_2')|0\rangle.~~~~~~\end{aligned}$$ Here $C$ in Eq.(\[PDC\]) is a normalization factor but because $|\eta|<< 1$, $|C|\approx 1$ no matter what function $\Phi(\omega_1,\omega_2)$ is. Two-photon and four-photon detections project the state to $\eta|\Phi_{2D}\rangle$ and $\eta^2|\Phi_{4D}\rangle/2$, respectively, which are not normalized. From Eqs.(\[N\],\[PN4\]), we then have $$\begin{aligned} &&P_2 = 2|\eta|^2\int d\omega_1d\omega_2[|\Phi(\omega_1,\omega_2)|^2 +\cr&&\hskip 1in+\Phi^*(\omega_1,\omega_2)\Phi(\omega_2,\omega_1)].~~~~\label{P2}\end{aligned}$$ The last term is related to the permutation symmetry or the degree of two-photon temporal distinguishability and can be viewed as a two-photon bunching effect. For a state from parametric down-conversion in the degenerate case as in Eq.(\[PDC\]), we usually have the symmetry $\Phi(\omega_1,\omega_2) = \Phi(\omega_2,\omega_1)$ so that $$\begin{aligned} &&P_{2D} = 4|\eta|^2\int d\omega_1d\omega_2|\Phi(\omega_1,\omega_2)|^2 .~~~~\label{P2D}\end{aligned}$$ Similarly for four-photon case, we have $$\begin{aligned} &&P_{4D} = 48|\eta|^2({\cal A}+2{\cal E}) = 3 P_2^2 (1+2{\cal E/A}),~~~~\label{P4D}\end{aligned}$$ where ${\cal E}, {\cal A}$ are given in Eqs.(\[8-181\], \[8-182\]), respectively. The dependence on ${\cal E/A}$ indicates that the extra term in Eq.(\[P4D\]) is a pair bunching effect – a generalized photon bunching effect for a multi-photon state. Direct measurement by Sun [*et al*]{} [@sun2] confirmed the four-photon bunching effect in Eq.(\[P4D\]). Another example is from non-degenerate parametric down-conversion in type-II $\chi^{(2)}$ medium. The quantum state is similar to that in Eq.(\[PDC\]) [@rhe2]: $$\begin{aligned} &&|\Phi_{NPDC}\rangle = C\Big(|0\rangle + \eta |\Phi_{2N}\rangle +{\eta^2\over 2}|\Phi_{4N}\rangle +...\Big),~~~~~\label{NPDC}\end{aligned}$$ with $$\begin{aligned} |\Phi_{2N}\rangle = \int d\omega_1d\omega_2\Phi(\omega_1,\omega_2)\hat a^{\dag}_H(\omega_1)\hat a^{\dag}_V(\omega_2)|0\rangle\end{aligned}$$ and $$\begin{aligned} &&|\Phi_{4N}\rangle = \int d\omega_1d\omega_2d\omega_1'd\omega_2' \Phi(\omega_1,\omega_2)\Phi(\omega_1',\omega_2')\times\cr&&\hskip 0.8in\times \hat a^{\dag}_H(\omega_1)\hat a^{\dag}_V(\omega_2)\hat a^{\dag}_H(\omega_1')\hat a^{\dag}_V(\omega_2')|0\rangle.~~~~~~\end{aligned}$$ From Eq.(\[PN5\]), it is straightforward to have $$\begin{aligned} &&P_{2N} = |\eta|^2\int d\omega_1d\omega_2|\Phi(\omega_1,\omega_2)|^2 ~~~~\label{P2N'}\end{aligned}$$ and $$\begin{aligned} &&P_{4N} = 2|\eta|^2({\cal A}+{\cal E}) = 2P_2^2 (1+{\cal E/A}).~~~~\label{P4N}\end{aligned}$$ Although there is no photon bunching at two-photon detection, we still have the pair bunching effect that depends on the ${\cal E/A}$ quantity. N-Photon Interference from an N-photon state ============================================ As seen in the previous section, a direct N-photon detection scheme cannot characterize the temporal indistinguishability in the general case. Therefore, we need to seek another method. Since the direct result of photon indistinguishability is the interference effect, our scheme will be an N-photon interference scheme. As a matter of fact, a Hong-Ou-Mandel interferometer [@hom] has already been used to measure two-photon indistinguishability from a type-II non-degenerate parametric down-conversion [@wam; @ser]. Our method proposed in the following will be a generalization of the Hong-Ou-Mandel interferometer from a two-photon case to an arbitrary N-photon case. NOON State Projection as a Measure for Distinguishability --------------------------------------------------------- The NOON state projection measurement was recently proposed to demonstrate an N-photon de Broglie wavelength without the need for a NOON state [@sun1]. It was demonstrated for $N= 4$ with states from parametric down-conversion [@sun2] and $N=6$ for a coherent state [@res] experimentally . The scheme is depicted in Fig.2 where the input is an arbitrary N-photon state of two polarization modes in the form of $$\begin{aligned} &&|\Psi_N\rangle = \sum_{k=0}^N c_k |N-k, k\rangle.~~~~\label{PsiN}\end{aligned}$$ The N-photon coincidence probability from the $N$ detectors is proportional to $$\begin{aligned} &&P_N \propto |\langle NOON|\Psi_N\rangle|^2.~~~~\label{P-N}\end{aligned}$$ If the input state is of the form of $|N-k,k\rangle (k\ne0, N)$, the output of the projection is zero. From the construction of the NOON state, we find this orthogonal projection is a result of N-photon interference and thus it can be used to characterize the temporal indistinguishability by the visibility in the interference. We will demonstrate this in the following sections. Three-photon case ----------------- Let us start with a three-photon state of the form $|2_H, 1_V\rangle$. So the three-photon NOON state projection measurement should yield null three-photon coincidence in the ideal case when all three photons are in one temporal mode. However, there may be some delay between the vertical photon and the two horizontal photons due to birefringence. Furthermore, the two horizontal photons may also be separated from each other. To account for the three scenarios described above, we cannot use the single-mode state of $|2_H, 1_V\rangle$ and have to resort to the multi-mode model discussed in Sect.III. A multi-mode three-photon polarization state for $|2_H, 1_V\rangle$ has the form of $$\begin{aligned} &&|\Phi_{3}\rangle = \int d\omega_1d\omega_2d\omega_3 \Phi(\omega_1,\omega_2,\omega_3)\times\cr&&\hskip 0.8in\times \hat a^{\dag}_H(\omega_1)\hat a^{\dag}_H(\omega_2)\hat a^{\dag}_V(\omega_3)|0\rangle.~~~~~~\label{Phi3}\end{aligned}$$ For simplicity of argument, we take $\Phi(\omega_1,\omega_2,\omega_3)$ in the form of Eq.(\[Phi-T\]): $$\begin{aligned} && \Phi(\omega_1,\omega_2,\omega_3)\cr&&\hskip 0.3in=\phi(\omega_1)e^{i\omega_1T_1} \phi(\omega_2)e^{i\omega_2T_2}\phi(\omega_3)e^{i\omega_3T_3}.~~~~~~\label{Phi3-T}\end{aligned}$$ We will use the un-normalized state because in practice, the state in Eq.(\[Phi3\]) can be generated by superposing a weak coherent state $|\alpha\rangle$ with a two-photon state $|\eta\rangle = |0\rangle+\eta|1_H,1_V\rangle$ from non-degenerate parametric down-conversion: $$\begin{aligned} &&|\Phi_{3}\rangle = |\alpha\rangle_H|\eta\rangle \approx |0\rangle + \alpha|1_H,0_V\rangle + \cr&&\hskip 0.5in +(\alpha^2/\sqrt{2})|2_H,0_V\rangle +\eta|1_H,1_V\rangle + \cr&&\hskip 0.6in + (\alpha^3/\sqrt{6})|3_H,0_V\rangle +\eta\alpha|2_H,1_V\rangle,~~~~~~\label{Phi3SM}\end{aligned}$$ where the states are in a single temporal mode and we only write out states up to three photons. A three-photon coincidence measure like that in the N-photon NOON state projection will only have contributions from the last two terms. By making the coherent state weak enough so that $|\eta|>>|\alpha|^2$, we are left with only $|2_H,1_V\rangle$ term. Since $|\alpha|, |\eta|<<1$, the three-photon state is not normalized. For the scenarios presented in the beginning, we can relate them to different values of $T_1,T_2,T_3$. So $T_1=T_2=T_3$ is for the case of three photons all in one single temporal mode. When $|T_3-T_1| >> 1/\Delta\omega, |T_3-T_2| >> 1/\Delta\omega$, the V-photon is far from the two H-photons. When $|T_2-T_1| >> 1/\Delta\omega$, the two H-photons are far apart. For the projection measurement in Fig.2 with $N=3$, we have the electric field operators at three detectors as $$\begin{aligned} \begin{cases} \hat E_0(t) = [\hat E_H(t) - \hat E_V(t)]/\sqrt{6}, \cr\hat E_1(t) = [\hat E_H(t) - e^{i2\pi/3}\hat E_V(t)]/\sqrt{6}, \cr\hat E_2(t) = [\hat E_H(t) - e^{i4\pi/3}\hat E_V(t)]/\sqrt{6}.\label{E3} \end{cases}\end{aligned}$$ To find the three-photon coincidence probability, we first calculate the time correlation function $$\begin{aligned} &&\Gamma^{(3)}(t_1,t_2,t_3) \cr&&\hskip 0.2in = \langle \hat E_0^{\dag}(t_3)\hat E_1^{\dag}(t_2) \hat E_2^{\dag}(t_1) \hat E_2(t_1)\hat E_1(t_2) \hat E_0(t_3)\rangle.~~~~~~\label{Ga}\end{aligned}$$ It is easy to calculate $\hat E_2(t_1)\hat E_1(t_2) \hat E_0(t_3)|\Phi_3\rangle$: $$\begin{aligned} \hat E_2(t_1)\hat E_1(t_2) \hat E_0(t_3)|\Phi_3\rangle= {-1\over 6\sqrt{6}} \bigg(\hat E_H\hat E_V\hat E_He^{i2\pi/3}+\hat E_V\hat E_H\hat E_H+\hat E_H\hat E_H\hat E_Ve^{i4\pi/3}\bigg)|\Phi_3\rangle.\label{E3}\end{aligned}$$ Here we dropped the terms that have no contribution. The order of the operators is kept for the time variables $t_3t_2t_1$. With the state in Eq.(\[Phi3\]) and $\Phi_3$ in Eq.(\[Phi3-T\]), it is straightforward to find $$\begin{aligned} &&\hat E_2(t_1)\hat E_1(t_2) \hat E_0(t_3)|\Phi_3\rangle = {-1\over 6\sqrt{6}} \bigg\{\Big[G(t_1,t_2,t_3)+G(t_2,t_1,t_3)\Big]e^{i4\pi/3} + \Big[G(t_1,t_3,t_2)+G(t_3,t_1,t_2)\Big] e^{i2\pi/3}+\cr &&\hskip 3in + \Big[G(t_2,t_3,t_1)+ G(t_3,t_2,t_1)\Big]\bigg\}|0\rangle,\label{E3G}\end{aligned}$$ where $$\begin{aligned} &&G(t_1,t_2,t_3) = {1\over \sqrt{(2\pi)^3}}\int d\omega_1 d\omega_2 d\omega_3 \Phi(\omega_1,\omega_2,\omega_3)\times\cr &&\hskip 1.7in \times e^{-i(\omega_1t_1+\omega_2t_2+\omega_3t_3)}\cr &&\hskip 0.73in= g(t_1-T_1)g(t_2-T_2)g(t_3-T_3)~~~~~~~~~~\label{G3}\end{aligned}$$ with $g(\tau)$ given in Eq.(\[g\]). The three-photon joint detection probability is an integral of the correlation function in Eq.(\[Ga\]) over all time variables: $$\begin{aligned} P_3 = \int dt_1dt_2dt_3 \Gamma^{(3)}(t_1,t_2,t_3).~~~~~~\label{P3}\end{aligned}$$ We are now ready to discuss the three scenarios presented in the beginning of this section. The interference effect is best measured by the visibility which is usually defined as the relative depth of modulation as compared to the situation when the interference effect is zero. In the three scenarios, we find the situation when the V-photon is far apart from the two H-photons corresponds to the case of no interference, which sets the reference line for evaluating the visibility defined by $$\begin{aligned} {\cal V}_3 ={| P_3-P_3(T_3=\infty)|\over P_3(T_3=\infty)}. \label{V3}\end{aligned}$$ Experimentally, we can scan $T_3$ from $T_3=\infty$ until we observe the dip in $P_3$ and use Eq.(\[V3\]) to calculate the visibility. Depending on the separation between the two H-photons, we actually only have two distinct cases: ($i$) the two H-photons are completely indistinguishable with $T_1=T_2\equiv T$; ($ii$) the two H-photons are well separated and distinguishable in time with $|T_1-T_2|>>1/\Delta\omega$. In case ($i$) with $T_1=T_2\equiv T$, we have the exchange symmetry $G(t_1,t_2,t_3) = G(t_2,t_1,t_3)$ and Eq.(\[P3\]) becomes after the time integral $$\begin{aligned} P_3 = 2{{\cal A}_3-{\cal E}_3(\Delta T)\over 36}~~~~~~\label{P33}\end{aligned}$$ with $\Delta T = T_3-T$ and $$\begin{aligned} {\cal A}_3 &\equiv & \int d\omega_1 d\omega_2 d\omega_3 |\Phi(\omega_1,\omega_2,\omega_3)|^2 \cr&=& \bigg(\int d\omega|\phi(\omega)|^2\bigg)^3,~~~~~\end{aligned}$$ $$\begin{aligned} {\cal E}_3 (\tau) &\equiv & \int d\omega|\phi(\omega)|^2 \bigg|\int d\omega|\phi(\omega)|^2e^{-i\omega\tau}\bigg|^2.~~~~~\end{aligned}$$ Note that ${\cal E}_3(0) ={\cal A}_3$ and ${\cal E}_3(\infty)=0$. So from Eq.(\[V3\]), we have the visibility for case ($i$) as $$\begin{aligned} {\cal V}_3(i) ={| P_3(\Delta T=0)-P_3(\Delta T=\infty)|\over P_3(\Delta T=\infty)} = 1. \label{V3i}\end{aligned}$$ The 100% visibility corresponds to the single-mode discussion before. In case ($ii$) with $|T_1-T_2|>>1/\Delta\omega$, there is no overlap between $G(t_1,t_2,t_3)$ and $G(t_2,t_1,t_3)$ so that $\int dt_1dt_2dt_3 G^*(t_1,t_2,t_3)G(t_2,t_1,t_3) = 0$. We obtain after the time integral $$\begin{aligned} P_3 = {{\cal A}_3-{\cal E}_3(\Delta T_1)/2-{\cal E}_3(\Delta T_2)/2\over 36}~~~~~~\label{P35}\end{aligned}$$ with $\Delta T_1 = T_3 -T_1$ and $\Delta T_2 = T_3 -T_2$. So we will have two dips with half depth when $T_3$ scans through $T_1$ and $T_2$. The visibility of each dip is then 50%. In summary, we find that the scenario when the two H-photons are separated have a visibility of 50% while when the two H-photons are in one temporal mode, the interference visibility becomes 100%. Therefore, we can distinguish the two different scenarios in the three-photon case by measuring the visibility in the NOON state projection measurement. Recent experiment by Liu [*et al*]{} [@liu] realized the two scenarios described above and confirmed the corresponding visibility. Next, we will generalize this result to an N-photon state. $N+1$-photon case ----------------- Let us now generalize the conclusion in the previous section to the case of $|1_H,N_V\rangle$ with an arbitrary integer $N$. The most general scenario in this case is when the single horizontal photon (H) is indistinguishable from $m$ vertical photons (V) while other $N-m$ V-photons are well separated in time from the $m+1$ photons (the case of $1HmV+(N-m)V$ or $1HmV$ for short). The multi-mode description of this state has the form of $$\begin{aligned} |\Phi(1HmV)\rangle = \int d\omega_1 d\omega_2...d\omega_{N+1} \Phi(\omega_1,..., \omega_N; \omega_{N+1})\hat a_V^{\dag}(\omega_1)...\hat a_V^{\dag}(\omega_N)\hat a_H^{\dag}(\omega_{N+1})|vac\rangle,~~~~~~\label{9-86}\end{aligned}$$ with $$\begin{aligned} \Phi(\omega_1,..., \omega_N; \omega_{N+1}) = \phi(\omega_1)e^{i\omega_1T_1} ... \phi(\omega_N)e^{i\omega_NT_N} \phi(\omega_{N+1})e^{i\omega_{N+1}T_{N+1}}.~~~~~~~\end{aligned}$$ Here we take $\Phi$ in the form of Eq.(\[Phi-T\]) for ease of calculation. When $m$ H-photons are in the same temporal mode with the V-photon, we have $T_1=...=T_m=T_{N+1}\equiv T$. But the other $N-m$ V-photons are well separated from these $m+1$ photons. This leads to $|T_j-T_k|>>1/\Delta\omega$ with $j=1,2,...,m, N+1$ and $k=m+1,...,N$ and the orthogonal relation: $$\begin{aligned} &&\int dt_1dt_2 g^*(t_1-T_j)g^*(t_2-T_k)\cr && \hskip 0.6in \times g(t_1-T_k)g(t_2-T_j) = 0.\label{9-90}\end{aligned}$$ Now we are ready to evaluate the joint $N+1$-photon probability $P_{N+1}$ in the NOON-state projection measurement scheme with an input state of $|\Phi(1HmV)\rangle$ in Eq.(\[9-86\]). $P_{N+1}$ is a time integral of the correlation function from $(N+1)$ detectors: $$\begin{aligned} &&\Gamma^{(N)}(t_1,t_2,...,t_N) \cr &&\hskip 0.3in = \langle \Phi(1HmV)|\hat E^{\dag}_{N+1}(t_{N+1})...\hat E^{\dag}_1(t_1) \cr && \hskip 0.8 in \times \hat E_1(t_1)...\hat E_{N+1}(t_{N+1}) |\Phi(1HmV)\rangle,~~~~~\label{9-91}\end{aligned}$$ with $$\begin{aligned} \hat E_j(t) \propto \hat E_V(t) - \hat E_H(t)e^{i\delta_j}+...,\label{9-92}\end{aligned}$$ where $$\begin{aligned} \hat E_{H,V}(t) ={1\over \sqrt{2\pi}}\int d \omega \hat a_{H,V}(\omega)e^{-i\omega t}.\end{aligned}$$ It is easy to first evaluate $\hat E_1(t_1)...\hat E_{N+1}(t_{N+1})$ $|\Phi(1HmV) \rangle$. After expanding the product, we find only $N+1$ nonzero terms of the form $$\begin{aligned} -\sum_{k=1}^{N+1} e^{i \delta_k} \hat E_V(t_1)...\hat E_H(t_k)... \hat E_V(t_{N+1})|\Phi(1HmV)\rangle.~~~~\end{aligned}$$ For the state $|\Phi(1HmV)\rangle$ in Eq.(\[9-86\]), we have $$\begin{aligned} &&\hat E_V(t_1)...\hat E_H(t_k)... \hat E_V(t_{N+1})|\Phi(1HmV)\rangle \cr &&\hskip 0.8in ={\mathcal G}(P_{k, N+1}\{t_1, ..., t_{N+1}\})|vac\rangle,~~~~~\end{aligned}$$ with $$\begin{aligned} {\cal G}(t_1, ..., t_{N}; t_{N+1})= \sum_P G(P\{t_1, ..., t_{N}\}; t_{N+1}),\label{9-96}\end{aligned}$$ and $$\begin{aligned} G(t_1, ..., t_{N}; t_{N+1})= \prod_{s=1}^{N+1}g(t_s-T_s), \label{GN+1}\end{aligned}$$ where $P_{k, N+1}$ exchanges $t_k$ with $t_{N+1}$ and $P$ is a permutation of $t_1, ..., t_{N}$. For the case of $1VmH$, we have $$\begin{aligned} G(t_1, ..., t_{N+1})= g(t_{N+1}-T) \prod_{s=1}^{m}g(t_s-T)\prod_{l=m+1}^{N}g(t_l-T_l), \label{G1VmH}\end{aligned}$$ so that $G(t_1, ..., t_{N+1})$ has exchange symmetry in $t_1,...,t_m, t_{N+1}$. The overall $(N+1)$-photon coincidence probability is then given by $$\begin{aligned} P_{N+1}(1HmV) &\propto &\int dt_1...dt_{N+1} \bigg| \sum_{k=1}^{N+1} e^{i\delta_k} {\cal G}(P_{k, N+1}\{t_1, ..., t_{N+1}\})\bigg|^2\cr &=& \sum_{k,j} e^{i(\delta_k - \delta_j)} \int dt_1...dt_{N+1} {\cal G}(P_{k, N+1}\{t_1, ..., t_{N+1}\}) {\cal G}^*(P_{j, N+1}\{t_1, ..., t_{N+1}\}).\label{9-97}\end{aligned}$$ Diagonal terms of $k=j$ in the double sum are all same because the integration is over all time variables: $$\begin{aligned} \int dt_1...dt_{N+1} \big| {\cal G}(P_{k, N+1}\{t_1, ..., t_{N+1}\})\big|^2 = \int dt_1...dt_{N+1} \big|{\cal G}(t_1, ..., t_{N+1})\big|^2.\end{aligned}$$ Furthermore, $$\begin{aligned} \int dt_1...dt_{N+1} \big|{\cal G}(t_1, ..., t_{N+1})\big|^2 &=& \int dt_1...dt_Ndt_{N+1} \bigg|\sum_P P[g(t_1-T_1)...g(t_N-T_N)]g(t_{N+1}-T_{N+1})\bigg|^2\cr &=& \int dt_1...dt_N \bigg|\sum_P P[g(t_1-T_1)...g(t_N-T_N)]\bigg|^2,\end{aligned}$$ where we used the normalization relation in Eq.(\[g-norm\]). From Eqs.(\[PN2\]–\[PN4\]), we find that it is simply $N!{\cal N}$ with ${\cal N}$ given in Eq.(\[NT\]). So the diagonal terms of $k=j$ in Eq.(\[9-97\]) are summed to be $(N+1)N!{\cal N}$. The cross terms in the double sum in Eq.(\[9-97\]) are given by $$\begin{aligned} \sum_{k\ne j}e^{i(\delta_k-\delta_j)} \int dt_1...dt_{N+1} {\cal G}(P_{k, N+1}\{t_1, ..., t_{N+1}\}) {\cal G}^*(P_{j, N+1}\{t_1, ..., t_{N+1}\}).\end{aligned}$$ Let us consider one arbitrary term in the sum. The time integral part can be rewritten as $$\begin{aligned} \sum_{P} \int dt_1...dt_{N+1} G(P\{t_1, ...,t_{k-1}, t_{N+1} ,t_{k+1},..., t_{N}\}; t_k) \sum_{P^{\prime}} G^*(P^{\prime}\{t_1, ...,t_{j-1}, t_{N+1} ,t_{j+1} , ..., t_{N}\}; t_j).\end{aligned}$$ Since $k\ne j$, the variable set $\{t_1, ...,t_{k-1}, t_{N+1}, t_{k+1}$, $..., t_{N}\}$ is different from $\{t_1, ...,t_{j-1}, t_{N+1} ,t_{j+1} , ..., t_{N}\}$ only at $t_j$ and $t_k$. For those $P$s such that $P\{t_1, ...,t_{k-1}, t_{N+1} ,t_{k+1},..., t_{N}\}$ moves $t_{j}$ to the first $m$ positions in the variable set $\{t_1, ..., t_{N}\}$, the symmetry between $t_1,...,t_m$ and $t_{N+1}$ in the function $G(t_1, ..., t_{N}; t_{N+1})$ in Eq.(\[G1VmH\]) will make $G(P\{t_1, ...,t_{k-1}, t_{N+1} ,t_{k+1},..., t_{N}\}; t_k) = G(P\{t_1, ...,t_{j-1}, t_{N+1} ,t_{j+1} , ..., t_{N}\}; t_j)$. There are totally $m(N-1)!$ such permutations and they all lead the time integral to $$\begin{aligned} \int dt_1...dt_{N+1} G(t_1, ...,t_{j-1}, t_{N+1} ,t_{j+1},..., t_{N}\}; t_j) \sum_{P^{\prime}}G^*(P^{\prime}\{t_1, ...,t_{j-1}, t_{N+1} ,t_{j+1} , ..., t_{N}\}; t_j).\end{aligned}$$ By Eq.(\[PN4a\]), it is simply ${\cal N}$. For the other permutations that move $t_j$ to the position of $t_{m+1}, ..., t_N$, it cannot be interchanged with $t_k$ because $T \ne T_s (s=m+1, ..., N)$. Furthermore, by the orthogonal relation in Eq.(\[9-90\]), the time integral is simply zero. Therefore, the cross terms are equal to $$\begin{aligned} \int dt_1...dt_{N+1} \sum_{k\ne j}e^{i(\delta_k-\delta_j)} {\cal G}(P_{1k}\{t_1, ..., t_{N+1}\}) {\cal G}^*(P_{1j}\{t_1, ..., t_{N+1}\}) = m (N-1)!{\cal N} \sum_{k\ne j} e^{i(\delta_k-\delta_j)}.\label{9-102}\end{aligned}$$ But because $\sum_k e^{i\delta_k} = 0$, we have $$\begin{aligned} \sum_{k\ne j} e^{i(\delta_k-\delta_j)} &&= \bigg(\sum_{k,j} - \sum_{k=j}\bigg) e^{i(\delta_k-\delta_j)} \cr &&= \sum_k e^{i\delta_k} \sum_j e^{-i\delta_j} -(N+1) \cr &&=-(N+1).\label{9-103}\end{aligned}$$ So the final result is $$\begin{aligned} P_{N+1}(1HmV) &\propto & {\cal N} (N+1)(N-1)!(N- m ) \cr &=& (N+1)!{\cal N}\Big(1-{m\over N}\Big).~~~~~\label{9-104}\end{aligned}$$ For the generalized Hong-Ou-Mandel interferometer, we scan the delay of the H-photon relative to the V-photons. When it does not overlap with any of the V-photons, no interference occurs and $P_{N+1}$ is a straight line which corresponds to $m=0$ in Eq.(\[9-104\]) with $P_{N+1}(\infty) = (N+1)!{\cal N}$. The value in Eq.(\[9-104\]) corresponds to the case when the delay is zero between the $m$ V-photons and the one H-photon and a local maximum interference is achieved. So the visibility is $$\begin{aligned} {\cal V}_{N+1}(1HmV) &\equiv &{P_{N+1}(\infty) - P_{N+1}(1HmV) \over P_{N+1}(\infty)} \cr &=& {m\over N}.\label{VN+1}\end{aligned}$$ Note that this visibility only depends on $N$ and $m$, i.e., the total number $N$ of V-photons and the number $m$ of V-photons that overlap with the single H-photon. It is independent of the normalization factor ${\cal N}$ or how the other $N-m$ photons distribute in time. So for a temporal distribution of well separated groups of V-photons shown in Fig.3a, as we scan the location of the single H-photon, we will have more dips of various visibility (Fig.3b) and the visibility is $m/N$ when the single H-photon overlaps with the group of $m$ V-photons that are in one temporal mode and are well separated from other V-photons. In general for a temporal distribution with $m$ partially overlapping V-photons, the visibility will be a value less than $m/N$. Therefore, the experimentally measurable visibility of the dips can be used to characterize the degree of temporal indistinguishability of an N-photon state. The general case of $|k_H, N_V\rangle$ with $k>1$ ================================================= For a more general case of input state of $|k_H, N_V\rangle$ with $k>1$, there are many scenarios for the temporal distribution of the photons. We will start with the four-photon case of $k=N=2$. Four-photon case of $|2_V, 2_H\rangle$ -------------------------------------- This situation was discussed in Ref.[@sun1] for $4\times 1$ case and $2HV\times 2HV$ case. It was shown that ${\cal V}_4 (4\times 1) =1$ and ${\cal V}_4 (2\times 2) =1/3$. But there are other scenarios like $2H1V+1V$ and $1H1V+1H+1V$. We will consider a simpler model to include these two scenarios so as to complete the distinguishability discussion in the four-photon case. For simplicity, we will again only discuss an un-normalized independent four-photon state of the form $$\begin{aligned} &&|\Phi_{4}\rangle = \int d\omega_1d\omega_2d\omega_3d\omega_4 \Phi(\omega_1,\omega_2,\omega_3,\omega_4)\times\cr&&\hskip 0.8in\times \hat a^{\dag}_H(\omega_1)\hat a^{\dag}_H(\omega_2)\hat a^{\dag}_V(\omega_3)a^{\dag}_V(\omega_4)|0\rangle.~~~~~~\label{Phi4}\end{aligned}$$ with $\Phi(\omega_1,\omega_2,\omega_3,\omega_4)$ in the form of Eq.(\[Phi-T\]): $$\begin{aligned} && \Phi(\omega_1,\omega_2,\omega_3,\omega_4)=\phi(\omega_1)e^{i\omega_1T_1} \phi(\omega_2)e^{i\omega_2T_2}\times\cr&&\hskip 1.2in\times\phi(\omega_3)e^{i\omega_3T_3} \phi(\omega_4)e^{i\omega_4T_4}.~~~~~~\label{Phi4-T}\end{aligned}$$ For the NOON state projection measurement with $N=4$, the field operators at the four detectors are related to the input field operators as $$\begin{aligned} \begin{cases} \hat E_0(t) = [\hat E_H(t) - \hat E_V(t)]/2+... ,\cr\hat E_1(t) = [\hat E_H(t) +\hat E_V(t)]/2+..., \cr\hat E_2(t) = [\hat E_H(t) - i\hat E_V(t)]/2+..., \cr\hat E_3(t) = [\hat E_H(t) + i\hat E_V(t)]/2+...,~~~~~~\label{E4} \end{cases}\end{aligned}$$ where we omit the vacuum modes. The four-photon detection probability at the four detectors is related to the following correlation function: $$\begin{aligned} \Gamma^{(4)}(t_1,t_2,t_3,t_4) = \langle \hat E_0^{\dag}(t_4)\hat E_1^{\dag}(t_3) \hat E_2^{\dag}(t_2) \hat E_3^{\dag}(t_1)\hat E_3(t_1) \hat E_2(t_2)\hat E_1(t_3) \hat E_0(t_4)\rangle.~~~~~~\label{Ga4}\end{aligned}$$ Again, it is easy to first calculate $\hat E_3(t_1) \hat E_2(t_2) \hat E_1(t_3)\hat E_0(t_4) |\Phi_4\rangle$. For this, we expand $\hat E_3(t_1) \hat E_2(t_2) \hat E_1(t_3)\hat E_0(t_4)$: $$\begin{aligned} \hat E_3(t_1)\hat E_2(t_2)\hat E_1(t_3)\hat E_0(t_4) = [(VVHH-HHVV) + i(VHVH+HVHV) -i(HVVH+VHHV)]/16, ~~~~~\label{EHV}\end{aligned}$$ where $H=\hat E_H, V=\hat E_V$ and we keep the time ordering. For the state $|\Phi_4\rangle$ in Eq.(\[Phi4\]), we have $$\begin{aligned} HHVV|\Phi_4\rangle = [G(t_1,t_2,t_3,t_4)+ G(t_2,t_1,t_3,t_4)+ G(t_1,t_2,t_4,t_3)+G(t_2,t_1,t_4,t_3)]|0\rangle, ~~~~~~\label{HHVV}\end{aligned}$$ $$\begin{aligned} VVHH|\Phi_4\rangle = [G(t_3,t_4,t_1,t_2)+ G(t_4,t_3,t_1,t_2)+ [G(t_3,t_4,t_2,t_1)+G(t_4,t_3,t_2,t_1)]|0\rangle, ~~~~~~\label{VVHH}\end{aligned}$$ $$\begin{aligned} HVHV|\Phi_4\rangle = [G(t_1,t_3,t_2,t_4)+ G(t_3,t_1,t_2,t_4)+ G(t_1,t_3,t_4,t_2)+G(t_3,t_1,t_4,t_2)]|0\rangle, ~~~~~~\label{HVHV}\end{aligned}$$ $$\begin{aligned} VHVH|\Phi_4\rangle = [G(t_2,t_4,t_1,t_3)+ G(t_2,t_4,t_3,t_1)+ G(t_4,t_2,t_1,t_3)+G(t_4,t_2,t_3,t_1)]|0\rangle, ~~~~~~\label{VHVH}\end{aligned}$$ $$\begin{aligned} HVVH|\Phi_4\rangle = [G(t_1,t_4,t_2,t_3)+ G(t_1,t_4,t_3,t_2)+ G(t_4,t_1,t_2,t_3)+G(t_4,t_1,t_3,t_2)]|0\rangle, ~~~~~~\label{HVHV}\end{aligned}$$ $$\begin{aligned} VHHV|\Phi_4\rangle = [G(t_2,t_3,t_1,t_4)+ G(t_2,t_3,t_4,t_1)+ G(t_3,t_2,t_1,t_4)+G(t_3,t_2,t_4,t_1)]|0\rangle ~~~~~~\label{VHHV}\end{aligned}$$ with $$\begin{aligned} G(t_1,t_2,t_3,t_4)= {1\over (2\pi)^2}\int d\omega_1 d\omega_2d\omega_3d\omega_4 \Phi(\omega_1,\omega_2,\omega_3,\omega_4)\times e^{-i(\omega_1t_1+\omega_2t_2+\omega_3t_3+\omega_4t_4)}.~~~~~~\label{G4}\end{aligned}$$ For the $\Phi$-function given in Eq.(\[Phi4-T\]), the above $G$-function is simply $$\begin{aligned} &&G(t_1,t_2,t_3,t_4)\cr &&\hskip 0.2 in= g(t_1-T_1)g(t_2-T_2)g(t_3-T_3)g(t_4-T_4).~~~~~~\label{G4-T}\end{aligned}$$ Four-photon coincidence probability is proportional to a time integral of the correlation function $\Gamma^{(4)}$: $$\begin{aligned} P_4 = \int dt_1dt_2dt_3dt_4 \Gamma^{(4)}(t_1,t_2,t_3,t_4).~~~~~~\label{P4}\end{aligned}$$ Next, we will evaluate $P_4$ for various scenarios of photon distinguishability. To describe the four scenarios discussed in the beginning of this section, we introduce three delay parameters: $\Delta T, \Delta T_V, \Delta T_H$ so that $T_2 = T_1+\Delta T_H, T_3 = T_1+\Delta T, T_4=T_3+\Delta T_V$. Therefore, $\Delta T$ is for the delay between the H-photons and the V-photons and $\Delta T_{H(V)}$ for the delay between the two H(V)-photons. When $\Delta T=\pm\infty$, there is no overlap between the H- and V-photons and no interference occurs. This sets up the baseline for evaluating the visibility of interference. We start with the $4\times 1$ case: ([*i*]{}) $\Delta T_H =0 = \Delta T_V$. There is an exchange symmetry between $t_1, t_2$ and between $t_3,t_4$ in $G(t_1,t_2,t_3,t_4)$ with $$\begin{aligned} G(t_1,t_2,t_3,t_4)= g(t_1-T_1)g(t_2-T_1)g(t_3-T_1-\Delta T)g(t_4-T_1-\Delta T).~~~~~~\label{G4-Ti}\end{aligned}$$ So we have $$\begin{aligned} &&\hat E_3(t_1)\hat E_2(t_2)\hat E_1(t_3)\hat E_0(t_4)|\Phi_4\rangle ={1\over4}\Big \{[G(t_1,t_2,t_3,t_4)-G(t_3,t_4, t_1,t_2)] + i[G(t_1,t_3,t_2,t_4)+G(t_2,t_4, t_1,t_3)] \cr &&\hskip 2.8in -i[G(t_1,t_4,t_3,t_2)+G(t_3,t_2, t_1,t_4)]\Big\}|0\rangle. ~~~~\label{E-4}\end{aligned}$$ After the time integral, we obtain $$\begin{aligned} P_4(\Delta T) = {1\over 8}\big[3{\cal A}_4 -4{\cal E}_4^{(1)}(\Delta T) + {\cal E}_4^{(2)}(\Delta T)\big]\end{aligned}$$ with $$\begin{aligned} {\cal A}_4 = \bigg(\int d\omega|\phi(\omega)|^2\bigg)^4,\end{aligned}$$ $$\begin{aligned} {\cal E}_4^{(1)}(\tau) = \bigg(\int d\omega|\phi(\omega)|^2e^{-i\omega\tau}\int d\omega|\phi(\omega)|^2\bigg)^2,\end{aligned}$$ $$\begin{aligned} {\cal E}_4^{(2)}(\tau) = \bigg(\int d\omega|\phi(\omega)|^2e^{-i\omega\tau}\bigg)^4.\end{aligned}$$ Note that ${\cal E}_4^{(1)}(0)={\cal E}_4^{(2)}(0)={\cal A}_4$ and ${\cal E}_4^{(1)}(\infty)={\cal E}_4^{(2)}(\infty)=0$. As we scan the relative delay $\Delta T$ between the H- and V-photons, the four-photon coincidence will show an interference dip all the way to zero when $\Delta T=0$, which corresponds to the case of $T_1=T_2=T_3=T_4$ or the $4\times 1$ case. So the visibility is 100% for the $4\times 1$ case. ([*ii*]{}) $\Delta T_H=0$ but $\Delta T_V >>1/\Delta\omega$. In this case, the two V-photons are well separated and we have $$\begin{aligned} &&G(t_1,t_2,t_3,t_4)= g(t_1-T_1)g(t_2-T_1) g(t_3-T_1-\Delta T) g(t_4-T_1-\Delta T_V-\Delta T).~~~~~~\label{G4-Tii}\end{aligned}$$ When $\Delta T=0$, there is an exchange symmetry between $\{t_1,t_2,t_3\}$ in $G(t_1,t_2,t_3,t_4)$. This is the $2H1V+1V$ case. But for arbitrary $\Delta T$, there is only a permutation symmetry between $t_1,t_2$ in $G(t_1,t_2,t_3,t_4)$. Then we have $$\begin{aligned} &&\hat E_3(t_1)\hat E_2(t_2)\hat E_1(t_3)\hat E_0(t_4)|\Phi_4\rangle \cr &&\hskip 0.5in={1\over8}\Big \{[G(t_1,t_2,t_3,t_4)+ G(t_1,t_2,t_4,t_3)-G(t_3,t_4, t_1,t_2)-G(t_3,t_4, t_2,t_1)] \cr &&\hskip 1in + i[G(t_1,t_3,t_2,t_4)+G(t_1,t_3,t_4,t_2)+G(t_2,t_4, t_1,t_3)+G(t_2,t_4, t_3,t_1)] \cr &&\hskip 1.2in-i[G(t_1,t_4,t_3,t_2)+G(t_3,t_2, t_1,t_4,)+G(t_1,t_4,t_2,t_3)+G(t_3,t_2, t_4,t_1,)]\Big\}|0\rangle. ~~~~\label{E-4ii}\end{aligned}$$ When $\Delta T=\pm\infty$, there is no overlap between all the terms in Eq.(\[E-4ii\]) so that all the cross terms are zero after the time integral in Eq.(\[P4\]). So we have $$\begin{aligned} P_4(\Delta T=\pm\infty) = 3{\cal A}_4/16.\end{aligned}$$ On the other hand, when $\Delta T=0$, there is an exchange symmetry between $\{t_1,t_2,t_3\}$ in $G(t_1,t_2,t_3,t_4)$. So Eq.(\[E-4ii\]) becomes $$\begin{aligned} &&\hat E_3(t_1)\hat E_2(t_2)\hat E_1(t_3)\hat E_0(t_4)|\Phi_4\rangle \cr &&\hskip 0.2in={1\over8}\Big [G(t_1,t_2,t_3,t_4)+ G(t_1,t_2,t_4,t_3)\cr &&\hskip 0.5in-G(t_3,t_4, t_1,t_2)-G(t_3,t_4, t_2,t_1) \Big]|0\rangle ~~~~\label{E-4ii0}\end{aligned}$$ and there is no overlap between all four terms above. After the time integral, we obtain $$\begin{aligned} P_4(\Delta T=0) = {\cal A}_4/16.\end{aligned}$$ So the visibility is $$\begin{aligned} {\cal V}_4(2H1V+1V) = 2/3.\end{aligned}$$ for the $2H1V+1V$ case. In fact, there is another $2H1V+1V$ case when the two H-photons overlaps with the other V-photon and $\Delta T=-\Delta T_V$. In this case, we have the exchange symmetry between $\{t_1,t_2,t_4\}$ in $G(t_1,t_2,t_3,t_4)$ so that $$\begin{aligned} P_4(\Delta T=-\Delta T_V) = {\cal A}_4/16,\end{aligned}$$ which also gives ${\cal V}_4(2H1V+1V)=2/3$. ([*iii*]{}) $\Delta T_H=\Delta T_V \equiv T >>1/\Delta\omega$. This is the $1H1V+1H1V$ or the $2\times 2$ case when $\Delta T=0$ and we have the exchange symmetry between $\{t_1,t_3\}$ and between $\{t_2,t_4\}$. But for $\Delta T=\pm\infty$, there is no overlap between any two of the 24 terms in Eqs.(\[HHVV\]–\[VHHV\]). So we have after the time integral in Eq.(\[P4\]) $$\begin{aligned} P_4(\Delta T=\pm\infty) = 24{\cal A}_4/16^2 = 3{\cal A}_4/32.\label{infty}\end{aligned}$$ When $\Delta T=0$, on the other hand, we have $$\begin{aligned} &&\hat E_3(t_1)\hat E_2(t_2)\hat E_1(t_3)\hat E_0(t_4)|\Phi_4\rangle ={i\over 8}\Big [G(t_1,t_3,t_4,t_2)+G(t_3,t_1, t_2,t_4)-G(t_1,t_4, t_3,t_2)-G(t_4,t_1,t_2,t_3)\Big]|0\rangle. ~~~~~~~\label{E-4iii}\end{aligned}$$ The above four terms have no overlap so that we obtain $$\begin{aligned} P_4(\Delta T=0) = 4{\cal A}_4/8^2 = {\cal A}_4/16.\end{aligned}$$ Therefore, the visibility for the $2\times2$ case is simply $$\begin{aligned} {\cal V}_4(2\times2) = 1/3.\end{aligned}$$ ([*iv*]{}) $|\Delta T_H-\Delta T_V|>>1/\Delta\omega$ and $|\Delta T_H|, |\Delta T_V|>>1/\Delta\omega$. As we scan $\Delta T$, there is an exchange symmetry only in one pair of the variables between $\{t_1,t_2\}$ and $\{t_3,t_4\}$, that is, between $\{t_1,t_3\}$ when $\Delta T=0$, or between $\{t_1,t_4\}$ when $\Delta T=-\Delta T_V$, or between $\{t_2,t_3\}$ when $\Delta T=\Delta T_H$, or between $\{t_2,t_4\}$ when $\Delta T=\Delta T_H-\Delta T_V$. This is the $(1H1V+1H+1V)$ case. In all these cases, 8 out of 24 terms in Eqs.(\[HHVV\]-\[VHHV\]) are cancelled in Eq.(\[EHV\]) and the remaining ones are orthogonal to each other so that we have $$\begin{aligned} P_4(\Delta T=0) = 16{\cal A}_4/16^2 = {\cal A}_4/16.\end{aligned}$$ The situation when $\Delta T=\pm \infty$ is same as Eq.(\[infty\]). Therefore the visibility is $$\begin{aligned} {\cal V}_4(1H1V+1H+1V) = 1/3.\end{aligned}$$ for the $1H1V+1H+1V$ case. These are all likely distinct scenarios. We summarize the visibility in Table I. Although visibility is derived with a specific $\Phi$-function in Eq.(\[Phi4-T\]), in general, visibility is the same regardless of the form of $\Phi$ as long as it is such that $G(t_1,t_2,t_3,t_4)$ has the required exchange symmetry in each scenario listed above. The intermediate situations will not have any symmetry in $G(t_1,t_2,t_3,t_4)$ and thus have very complicated dependence on the various permutations of $\Phi(\omega_1,\omega_2,\omega_3,\omega_4)$. Ref.[@sun1] discussed the intermediate scenario from the $2\times 2$ case to the $4\times 1$ case. Indeed, the visibility depends on the quantity ${\cal E/A}$, which defines the degree of pair distinguishability. Xiang [*et al*]{} [@xia] realized the $2\times 2$ and the $4\times 1$ cases experimentally and confirmed the visibility in Table I. ------ -- --------- -- --------- -- --------- 2H2V 2H1V+1V 1HV+1HV 1HV+H+V 1 2/3 1/3 1/3 ------ -- --------- -- --------- -- --------- : \[tab:table1\]Visibility for 2 H-photons and 2 V-photons input The Special Cases of $|2_H, 3_V\rangle$, $|2_H, 4_V\rangle$, and $|3_H, 3_V\rangle$ ------------------------------------------------------- Following the same line of derivation but in a much more complicated fashion, we may find the visibility for all the scenarios for the input states of $|2_H, 3_V\rangle$, $|2_H, 4_V\rangle$, and $|3_H, 3_V\rangle$. We list the likely scenarios below and tabulate the visibility for each scenarios in Tables II-IV. ### The Case of $|2_H, 3_V\rangle$ The case of $|2_H, 3_V\rangle$ has 8 different scenarios. They are \(i) $~|2H3V\rangle, |2H2V+1V\rangle, |2H1V+2V\rangle$, and \(ii) $|1H3V+1H\rangle, |1H2V+1H1V\rangle, |1H2V+1H+1V\rangle$, $|1H1V+1H1V+1V\rangle, |1H1V+1H+2V\rangle$. Their visibilities are listed in Table II. ------ ------ ------ ------ ------ ------ ------ ------ 2H3V 2H2V 2H1V 1H3V 1H2V 1H2V HV+V HV+V +V +2V +H +HV +H+V +HV +H+V 1 5/6 1/2 3/4 5/12 1/2 1/3 1/4 ------ ------ ------ ------ ------ ------ ------ ------ : Visibility for 2 H-photons and 3 V-photons input ### The Case of $|2_H, 4_V\rangle$ The case of $|2_H, 4_V\rangle$ has 12 different scenarios. They are \(i) $~|2H4V\rangle, |2H3V+1V\rangle, |2H2V+2V\rangle, |2H1V+3V\rangle$, and \(ii) $|1H4V+1H\rangle, |1H3V+1H1V\rangle, |1H3V+1H+1V\rangle,|1H2V+1H2V\rangle,$ $|1H2V+1H1V+1V\rangle, |1H2V+1H+2V\rangle, |1H1V+1H1V+2V\rangle$, $|1H1V+1H+3V\rangle$. The scenarios with different visibility are listed in Table III. $|1H2V+1H1V+1V\rangle$ and $|1H2V+1H+2V\rangle$ have the same visibility of 2/5 as $|1H2V+1H2V\rangle$. [cccccccccc]{} 2H4V&2H3V&2H2V&2H1V &1H4V&1H3V&1H3V&1H2V&2$\times$HV&1H1V\ &+V&+2V&+3V &+H&+HV& +H+V&+1H2V&+2V&+1H+3V\ 1&9/10&7/10&2/5&4/5&1/2&3/5&2/5&3/10&1/5\ In general, they follow the trend that smaller visibility corresponds to less photon overlapping. However, there are exceptions: $1H2V+HV$ has less visibility than $1H2V+1H+V$ in Table II and $1H3V+HV$ has less visibility than $1H3V+1H+1V$ in Table III. So the runaway $HV$ does not help when $H$ and $V$ overlap in these cases. ### The Case of $|3_H, 3_V\rangle$ There are totally 11 different scenarios in the special case of $|3_H, 3_V\rangle$: \(i) $~|3H3V\rangle, |3H2V+V\rangle, |3H1V+2V\rangle$; \(ii) $|2H2V+1H1V\rangle, |2H2V+1H+1V\rangle, |2H1V+1H2V\rangle$, $|2H1V+1H1V+1V\rangle, |2H1V+1H+2V\rangle$; \(iii) $|1H1V+1H1V+1H1V\rangle, |1H1V+1H1V+1H+1V\rangle$, $|1H1V+1H+1V+1H+1V\rangle$. In Table IV, we list the visibility for most of the scenarios. $|2H1V+1H1V+1V\rangle$ and $|2H1V+1H+2V\rangle$ have the same visibility of 2/5 as $|2H1V+1H2V\rangle$ and are not listed. As can be seen, anomaly occurs for $|2H2V+1H1V\rangle$ and $|2H2V+1H+1V\rangle$ where visibility is bigger for the case with less photon overlap. The scenarios of $|3H3V\rangle$, $|2H2V+1H1V\rangle$, and $|3\times HV\rangle$ were observed experimentally by Xiang [*et al.*]{} [@xia] with the corresponding visibility in Table IV. ------ ------ ------ ------ ------ ------- ------------- ------------- -------- 3H3V 3H2V 3H1V 2H2V 2H2V 2H1V HV$\times$3 HV$\times$2 HV+V +V +2V +HV +H+V +1H2V +H+V +H+H+V 1 9/10 3/5 3/5 7/10 2/5 2/5 3/10 1/5 ------ ------ ------ ------ ------ ------- ------------- ------------- -------- General Formula for the Visibility ---------------------------------- The most general case is when the input state is in the form of $|k_H, N_V\rangle$ with $k\le N$. The most general scenario is when the $k$ H-photons don’t overlap in time but rather are split into $r$ temporally well separated subgroups with $k_j$ indistinguishable photons in the $j$th group and $k_1+...+ k_r = k$. We also divide the $N$ V-photons into $r+1$ subgroups with $m_j$ V-photons overlap in time with the $j$th H-photon group. The rest $N-m_1-...-m_r$ V-photons are in a separate group by themselves. The wave function for these $N+k$ photons will satisfy the permutation symmetry relation similar to Eq.(\[sym\]) for the overlapping photons and the orthogonal relation similar to Eq.(\[orth\]) for the well separated photons. The derivation of the general formula for the visibility in the $(N+k)$-photon NOON-state projection measurement is very complicated and lengthy. It follows the general line of argument as that leading to Eq.(\[VN+1\]). We will present the detailed procedure elsewhere [@ou3] but only give the result as $$\begin{aligned} {\cal V}_{N+k} = \sum_{l=1}^k (-1)^{l-1}\sum_{i_1...i_r \atop i_1+...+i_r=l}^{l} \bigg({l!\over i_1!...i_r!}\bigg) {C_{k_1}^{i_1}...C_{k_r}^{i_r}~ m_1^{(i_1)}...m_r^{(i_r)}\over (N+k-1)...(N+k-l)},~~~~~~~~\label{VN+k}\end{aligned}$$ where $m^{(0)}=0=m^{(m)}$, $m^{(i)} \equiv m(m-1)...(m-i+1)$, and $C_N^M\equiv (N+M)!/N!M!$. For the special case of $k=1$, Eq.(\[VN+k\]) recovers the expression in Eq.(\[VN+1\]). Furthermore, we can easily check that the formula in Eq.(\[VN+k\]) indeed leads to the visibility values in Tables I-IV. Conclusion and Discussion ========================= The complementary principle of quantum interference is demonstrated in a quantitative way in multi-photon interference where photons can be categorized by their temporal distinguishability. The temporal indistinguishability of photons in turn can be characterized by the permutation symmetry in the multi-photon wave function while the temporal distinguishability by the orthogonality of the permuted wave functions. Generalization to other degrees of freedom such as spatial modes is straightforward. Although the above conclusions were made on photons, they should apply to any bosons as well as fermions so long as the occupation number of each mode is less than or equal to one. This work was supported by the US National Science Foundation under Grant No. 0245421 and No.0427647. The author would like to thank Mr. F. W. Sun for stimulating discussion. [99]{} L. Mandel and E. Wolf, Rev. Mod. Phys. [**37**]{}, 231 (1965). R. J. Glauber, Phys. Rev. [**130**]{}, 2529 (1963); [**131**]{}, 2766 (1963). L. Mandel, Phys. Rev. [**144**]{}, 1071 (1966). M. Born and E. Wolf, [*Principle of Optics*]{}, (Pergamon, Oxford, 1st ed., 1959; 7th ed., 1999). P. W. Shor, in Proceedings of the 35th Annual Symposium on the Foundations of Computer Science, Los Alamitos, CA (IEEE Computer Society Press, New York, 1994), p. 124., L. K. Grover, Phys. Rev. Lett. [**79**]{}, 325 (1997). H. F. Hofmann, Phys. Rev. A [**70**]{}, 023812 (2004). E. Knill, R. Laflamme, and G. J. Milburn, Nature (London) [**409**]{}, 46 (2001). P. Michler [*et al.*]{}, Science [**290**]{}, 2282 (2000). C. Santori, M. Pelton, G. Solomon, Y. Dale, and Y. Yamamoto, Phys. Rev. Lett. [**86**]{}, 1502 (2001). K. Tsujino, H. F. Hofmann, S. Takeuchi, and K. Sasaki, Phys. Rev. Lett. [**92**]{}, 153602 (2004). Z. Y. Ou, Phys. Rev. A [**72**]{}, 053814 (2005). Z. Y. Ou, J.-K. Rhee, and L. J. Wang Phys. Rev. Lett. [**83**]{}, 959 (1999). Z. Y. Ou, J.-K. Rhee, and L. J. Wang, Phys. Rev. A [**60**]{}, 593 (1999). F. W. Sun, Z. Y. Ou, and G. C. Guo, Phys. Rev. A [**73**]{}, 023808 (2006). F. W. Sun, B. H. Liu, Y. F. Huang, Z. Y. Ou, and G. C. Guo, Phys. Rev. A [**74**]{}, 033812 (2006). K. J. Resch, K. L. Pregnell, R. Prevedel, A. Gilchrist, G. J. Pryde, J. L. O’Brien, and A. G. White, quant-ph/0511214. W. P. Grice and I. A. Walmsley, Phys. Rev. A [**56**]{}, 1627 (1997). C. K. Hong, Z. Y. Ou, and L. Mandel, , 2044 (1987). M. Atatüre, A. V. Sergienko, B. M. Jost, B. E. A. Saleh, and M. C. Teich, Phys. Rev. Lett. [**83**]{}, 1323 (1999). B. H. Liu, F. W. Sun, Y. X. Gong, Y. F. Huang, Z. Y. Ou, and G. C. Guo, submitted (2006); quant-ph/0606118. G. Y. Xiang, Y. F. Huang, F. W. Sun, P. Zhang, Z. Y. Ou, and G. C. Guo, Phys. Rev. Lett. [**97**]{}, 023604 (2006). Z. Y. Ou, to be published (2007).
{ "pile_set_name": "ArXiv" }
--- author: - Casey Papovich - the GOODS and MIPS GTO teams date: 'Received  2005 month day; accepted  2006  month day' title: 'A Spitzer View of Massive Galaxies at $z\sim 1-3$' --- Introduction ============ Most ($\sim$50%) of the stellar mass in galaxies today formed during the short time between $z$$\sim$3 and 1 (e.g., Dickinson et al. 2003, Rudnick et al. 2003). Although much of this stellar mass density resides in massive galaxies, which appear at epochs prior to $z$$\sim$1–2 (e.g., Bell et al. 2004, McCarthy 2004), it is still unclear when and where the stars in these galaxies formed. It may be that galaxies “downsize”, forming most of their stars in their current configuration at early cosmological times, with lower mass galaxies continuing to form stars to the present epoch (e.g., Bauer et al. 2005, Juneau et al. 2005). Alternatively, stars may form predominantly in low–mass galaxies at high redshifts, which then merge over time to form large, massive galaxies at more recent times (Bauer et al. 1998, Kauffmann & Charlot 1998, Cimatti et al. 2002). Understanding the formation and evolution of massive galaxies has been challenged by difficulties in constructing complete samples of star–forming and massive galaxies at $z$${\gtrsim}$ 1. Massive galaxies at $z$${\lesssim}$1 exist on a fairly prominent red sequence (e.g., Blanton et al. 2003, Bell et al. 2004, Faber et al. 2005), and are largely devoid of ongoing star formation, evolve passively, and contain up to half of the stellar–mass density. However, current hierarchical models predict colors for massive galaxies at $z$$\sim$0 that are too blue compared to observations (e.g., Somerville, Primack, & Faber 2001, Davé et al. 2005). Some recent models suppress star formation at late times in massive galaxies by truncating star formation at some mass threshold, or by using feedback from strong AGN (e.g., Granato et al. 2001, Davé et al. 2005, Croton et al.2006). Most of the massive galaxies at $z$${\lesssim}$1 have red colors with formation epochs $z_f$${\gtrsim}$ 2. The morphologies of the most optically luminous (and most massive) galaxies transforms from “normal” early–type galaxies at $z$$\sim$ 1 to irregular systems at $z$$\sim$ 2–3 (e.g., Papovich et al. 2005). Therefore, we need to study the properties of massive galaxies at these earlier epochs. In these proceedings, I discuss recent [*Spitzer*]{} observations at 3–24  of massive galaxies at $z$$\sim$1.5–3 in the southern Great Observatories Origins Deep (GOODS–S) field. The [*Spitzer*]{}IRAC and MIPS observations provide constraints on the star–formation and AGN activity in massive galaxies at these epochs. I discuss implications the analyses of these data have for the stellar–mass assembly rates and formation epochs of massive galaxies. I also review some the uncertainties on these bolometric corrections between the observed [*Spitzer*]{} 24  data and total IR luminosities in $z$$\sim$ 1.5–3 galaxies and prospects for future improvements. Stellar Masses and Star Formation in High–$z$ Massive Galaxies ============================================================== GOODS is a multiwavelength survey of two 10$\times$15fields, one in the northern *Hubble* Deep Field, and one in the southern *Chandra* Deep Field. The GOODS datasets include (along with other things) [*HST*]{}/ACS and VLT/ISAAC imaging (Giavalisco et al. 2004), and recent [*Spitzer*]{} imaging. I make use of these data for the work described here, as well as data from [*Spitzer*]{}/MIPS 24  in this field from time allocated to the MIPS GTOs (e.g., Papovich et al. 2004). In these proceedings, I primarily focus on so–called distant red galaxies (DRGs) selected with $>$2.3 mag (see Franx et al. 2003). This color criterion identifies both galaxies at $z$$\sim$2–3.5 whose light is dominated by a passively evolving stellar population older than $\sim$250 Myr (i.e., with a strong Balmer/4000 Å break between the $J$ and –bands), and also star–forming galaxies at these redshifts whose light is heavily reddened by dust (Förster–Schreiber et al. 2004, Labbé et al. 2005, Papovich et al. 2006). For the GOODS–S data, the $>$2.3 mag selection is approximately complete to stellar mass ${\hbox{$\mathcal{M}$}}$$\geq$$10^{11}$  for passively evolving galaxies, and we find 153 DRGs to ${\hbox{$K_s$}}$$\leq$23.2 mag, spanning 0.8$\leq$$z$$\leq$3.7 with $\langle z\rangle $$\simeq$2.2 (see Papovich et al. 2006). ![Stellar masses of galaxies inferred by fitting two–component models to the galaxies’ rest–frame UV–to–near-IR data. The inset bars show the mean errors as a function of mass. The short–dashed line shows the characteristic present–day stellar mass (Cole et al. 2001); the long–dashed line shows the stellar mass limit for a passively evolving stellar population formed at $z$$\sim$${\hbox{$\infty$}}$ with ${\hbox{$K_s$}}$=23.2 mag. ](lirz.ps){width="0.95\linewidth"} ![Stellar masses of galaxies inferred by fitting two–component models to the galaxies’ rest–frame UV–to–near-IR data. The inset bars show the mean errors as a function of mass. The short–dashed line shows the characteristic present–day stellar mass (Cole et al. 2001); the long–dashed line shows the stellar mass limit for a passively evolving stellar population formed at $z$$\sim$${\hbox{$\infty$}}$ with ${\hbox{$K_s$}}$=23.2 mag. ](massz_2comp.ps){width="0.95\linewidth"} More than 50% of the DRGs have 24  detections with $f_\nu(24{\hbox{$\mu$m}})$$\geq$50 . Daddi et al. (2005) find a similar 24 –detection rate for massive galaxies at 1.5${\lesssim}$$z$${\lesssim}$2.5 selected via their $BzK$ colors. Interestingly, this implies that the majority of massive galaxies at $z$$\sim$2 emit strongly in the thermal IR — *they are either actively forming stars, supermassive blackholes, or both at this epoch*. The 24  emission at $z$$\sim$1.5–3 probes the mid–IR ($\sim$5–10 ), which broadly correlates with the total IR, ${\hbox{$L_{\mathrm{IR}}$}}$$\equiv$$L(8-1000{\hbox{$\mu$m}})$ (Chary & Elbaz 2001). Figure 1 shows the inferred  for the DRGs using the Dale et al. (2002) models to convert the observed mid–IR to total IR luminosity. There is inherent uncertainty in this conversion, which I discuss in § 3. The 24  flux densities for the $z$$\sim$1.5–3 DRGs yield ${\hbox{$L_{\mathrm{IR}}$}}$$\approx$$10^{11.5-13}$ , which if attributed to star–formation corresponds to SFRs of $\approx$100–1000 yr$^{-1}$ (e.g., Kennicutt 1998). Nearly all of the DRGs are detected in the deep [*Spitzer*]{}/IRAC data, implying they have substantial stellar masses. In Papovich et al.(2006), we modeled the DRG stellar populations by comparing their ACS, ISAAC, and IRAC    photometry to a suite of stellar–population synthesis models (Bruzual & Charlot 2003), varying the age, star–formation history, and dust content. We use the model stellar-mass–to–light ratios to estimate the galaxies’ stellar mass. We first allow for star–formation histories with the SFR parameterized as a decaying exponential with an $e$–folding time, $\tau$, ranging from short $\tau$’s (burst of star–formation) to long $\tau$’s (constant star–formation). We also use models with a two–component star–formation history characterized by a passively evolving stellar population formed in a “burst” at $z_\mathrm{form}$=$\infty$, summed with the exponentially–decaying–SFR model. Although the modeling loosely constrains the ages, dust content, and star–formation histories of the DRGs, it provides relatively robust estimates of the galaxies’ stellar masses (see also Förster–Schreiber et al. 2004). Typical uncertainties for the stellar masses for the full DRG sample are 0.1–0.3 dex. Figure 2 shows the stellar masses inferred for the DRGs by fitting the two–component models. Figure 3 shows the specific SFRs ($\Psi/{\hbox{$\mathcal{M}$}}$) derived from the masses and SFRs for the DRGs, where the SFRs are calculated from the summed UV and IR emission. The figure also shows the specific SFRs for galaxies at lower redshift from COMBO–17 (Wolf et al. 2003), which overlaps with the GTO [*Spitzer*]{} 24  imaging. The SFRs for the COMBO–17 galaxies are calculated in an analogous manner as for the DRGs. The massive DRGs at 1.5$\leq$$z$$\leq$3 have high specific SFRs: DRGs with ${\hbox{$\mathcal{M}$}}$$>$$10^{11}$  and 1.5$\leq$$z$$\leq$3 have $\Psi/{\hbox{$\mathcal{M}$}}$$\sim$0.2–10 Gyr$^{-1}$ (excluding X–ray sources). In contrast, at $z$${\lesssim}$ 0.75 there is an apparent lack of galaxies with high specific SFRs and high stellar masses: galaxies with ${\hbox{$\mathcal{M}$}}$$\geq$$10^{11}$  have $\Psi/{\hbox{$\mathcal{M}$}}$$\sim$0.1–1 Gyr$^{-1}$. We define the integrated specific SFR as the ratio of the sum of the SFRs, $\Psi_i$, to the sum of their stellar masses, ${\hbox{$\mathcal{M}$}}_i$, $\Upsilon$$\equiv$$\sum_i \Psi_i / {\sum_i{\hbox{$\mathcal{M}$}}_i}$, summed over all $i$ galaxies. This is essentially just the ratio of the SFR density to the stellar mass density for a volume–limited sample of galaxies. Figure 4 shows the integrated specific SFRs for DRGs at $z$$\sim$1.5–3.0 and COMBO–17 at $z\sim 0.4$ and 0.7 with ${\hbox{$\mathcal{M}$}}\geq 10^{11}$  (see Papovich et al. 2006). The data point for the DRGs includes all objects with ${\hbox{$\mathcal{M}$}}$$\geq$$10^{11}$ , and assumes that 24 –undetected DRGs have no star formation. The error–box lower bound shows what happens if we exclude objects with X–ray detections or IR colors indicative of AGN. The error–box upper bound shows what happens if we calculate SFRs for the 24 –*undetected* DRGs assuming they have $f_\nu(24{\hbox{$\mu$m}})$=60 , the 50% completeness limit. The integrated specific SFR in galaxies with $>$$10^{11}$ declines by more than an order of magnitude from $z$$\sim$1.5–3 to $z$${\lesssim}$0.7. Our results indicate that the relative star–formation in massive galaxies is reduced for $z$${\lesssim}$1 as galaxies with lower stellar masses have higher specific SFRs, supporting the so–called “downsizing” paradigm. ![Evolution of the integrated specific SFR, the ratio of the total SFR to the total stellar mass (from Papovich et al. 2006). The curves show the expected evolution from the global SFR density (solid lines, Cole et al. 2001, thick line includes correction for dust extinction; dashed line, Hernquist & Springel 2003). Data points show results for galaxies with $\geq$$10^{11}$ . Filled circle corresponds to the DRGs; filled diamonds correspond to the COMBO–17 galaxies. The inset error bar shows an estimate on the systematics.](specsfr_model.ps){width="0.7\linewidth"} Figure 4 also shows the specific SFR integrated over all galaxies (not just the most massive); this is the ratio of the cosmic SFR density to its integral, $\Upsilon$$=$$\dot{\rho}_\ast / \int \dot{\rho}_\ast\, dt$. There is a decrease in the global specific SFR with decreasing redshift. The evolution in the integrated specific SFR in massive galaxies is accelerated relative to the global value. Galaxies with ${\hbox{$\mathcal{M}$}}$$\ge$$10^{11}$  were forming stars at or slightly above the rate integrated over all galaxies at $z$$\sim$1.5–3. In contrast, by $z{\lesssim}1$ galaxies with ${\hbox{$\mathcal{M}$}}$ $\ge$$10^{11}$  have an integrated specific SFR much lower than the global value. *Thus, by $z$${\lesssim}$1.5 massive galaxies have formed most of their stellar mass, and lower–mass galaxies dominate the cosmic SFR density* (see also Papovich et al. 2006). Uncertainties on Mid-IR–Derived IR Luminosities =============================================== [**3.1 AGN Contribution to the mid–IR Emission**]{} Many of the massive galaxies at $z$$\sim$1.5–3 are detected in the deep X–ray data. In figures 1 and 3, the X-ray–detected DRGs tend to have the highest inferred IR luminosities and specific SFRs. Many of these objects have ${\hbox{$L_{\mathrm{IR}}$}}$$\geq$$10^{13}$ , comparable to PG quasars, which have warm thermal dust temperatures (Haas et al.2003). The X-ray to optical flux ratios in these objects imply the presence of an AGN with $L_X$${\gtrsim}$$10^{42}$ erg s$^{-1}$ (for $z$ ${\gtrsim}$1.5, see Alexander et al. 2003, Papovich et al. 2006). In addition, many authors are finding AGN candidates based on rest–frame near–IR colors from [*Spitzer*]{} observations. The X–ray, UV, and optical emission in these AGN is heavily obscured by gas and dust, and they are often missed in deep X–ray surveys (e.g., Donley et al.2005, Stern et al. 2005, Alonso–Herrero et al. 2006, Barmby et al.2006). Roughly 25% of the non–X-ray detected DRGs in the GOODS–S field satisfy the IRAC color–selection for AGN from Stern et al.(2005). Of these, roughly one–half have ACS–through–IRAC colors consistent with dust–enshrouded AGN. Figure 5 shows examples of these galaxies with and with–out X–ray detections. In all cases, the galaxies show a clear flux excess in their IRAC  and  photometry, presumably arising from the AGN. Combined with the 15% of DRGs detected in the X–rays, up to 25% of the DRG population host AGN. If AGN contribute to the observed 24  emission in galaxies at $z$$\sim$1.5–3, then they can affect the inferred IR luminosities. For example, although the Chary & Elbaz (2001) and Dale & Helou (2002) IR templates include galxaies with ${\hbox{$L_{\mathrm{IR}}$}}$${\gtrsim}$ $10^{13}$ , using an IR template for Mrk 231 — with a known AGN and warmer dust temperature — would reduce the inferred IR luminosity for galaxies at $z$$\sim$ 1.5–3 by factors of $\sim$2–5. To limit the effects of any bias in the inferred IR luminosities caused by AGN, we restricted the error box on the integrated specific SFR for the DRGs in figure 4 to only those galaxies with no X–ray or IR indications for AGN. Even when AGN are present, it is unclear whether star–formation or the AGN activity domninates the bolometric IR luminosity. Alexander et al. (2005) demonstrated that a high fraction ($\sim$80%) of sub–mm galaxies are detected in 2 Msec *Chandra* X–ray data. However, the X-ray–detected sub–mm galaxies have IR to X–ray luminosity ratios up to an order of magnitude higher than what is expected for AGN alone. This suggests that both star–formation and AGN contribute to the bolometric emission. Similarly, Frayer et al.(2003) report a near–IR spectrum of a sub–mm galaxy SMM J04431+0210 at $z$$\sim$ 2.5. This galaxy has $J-{\hbox{$K_s$}}$$\simeq$ 3.2 mag, qualifying as a DRG. The spectrum shows that the galaxy nucleus has a low to \[N[II]{}\] flux ratio consistent with ionization from an AGN. However,  is spatially resolved in their spectrum and the \[N[II]{}\] line strength drops off in the off–nucleus spectrum. This implies extended star–formation beyond the nucleus, which presumably contributes to the inferred IR luminosity. Therefore, it seems that both AGN and star–formation occur simultaneously in high–redshift IR–detected galaxies and both probably contribute to the IR emission. The high AGN occurrence in DRGs and sub–mm galaxies provides some evidence that massive galaxies at $z$$\sim$1.5–3 simultaneously form stars and grow supermassive black holes. Although speculative, the presence of AGN in these galaxies may provide the impetus for the present–day relation between black hole and bulge mass, and/or provide the feedback necessary to squelch star–formation (e.g., Kauffmann et al. 2004), moving the galaxies onto the red sequence. Emission line ratios from high–resolution spectroscopy at near–IR and/or mid–IR wavelengths may help constrain the ionization state of the galaxies’ gas, identifying the fraction of galaxies with strong AGN activity as a function of mass at these redshifts. [**3.2 Uncertainties in the Shape of the IR Spectral Energy Distribution**]{} Although the mid–IR (5–15 ) emission broadly correlates with the total IR luminosity (e.g., Chary & Elbaz 2001), there exists considerable scatter. Partly this arises because galaxies of a given IR luminosity show a range in the strength of their mid–IR emission (from polycyclic aromatic hydrocarbons, PAHs) and absorption features (e.g., Armus et al. 2004). This variation is reflected in a comparison of various model IR spectral energy distributions in the literature (e.g., see Le Floc’h et al. 2005). For example, if we used the IR templates of Chary & Elbaz (2001) instead of those of Dale & Helou (2002), then we would derive IR luminosities a factor of 2–3 higher for $z$$\sim$1.5–3 galaxies with ${\hbox{$L_{\mathrm{IR}}$}}$$\sim$$10^{12.5-13}$ . In a recent study of galaxies at $z$$\leq$1.2 detected at *ISO* 15  and [*Spitzer*]{} 24  in the northern GOODS field, Marcillac et al.(2006) found that, compared to Chary & Elbaz (2001), the IR models of Dale & Helou (2002) provide a tighter correlation between derived from the mid–IR and radio flux densities with a scatter of 40%, suggesting the latter may better reflect reality. Some scatter is inherent in converting the 24  flux densities to total IR luminosity. Several studies of the mid–IR colors of galaxies to $z$$\sim$1 show that their 15 and 16 to 24  colors have more scatter than predicted by simple models that map a single IR template to a given IR luminosity (e.g., Elbaz et al. 2005, Teplitz et al. 2005, Marcillac et al. 2006). Chapman et al. (2003) find that the temperature–luminosity distribution in local IR–luminous galaxies has a factor of 2–3 scatter in the IR–luminosity for galaxies of a given dust temperature. However, Daddi et al. (2005) found that the Chary & Elbaz (2001) IR model with ${\hbox{$L_{\mathrm{IR}}$}}$=$10^{12.2}$  fits the average spectral energy distribution of 24 –detected $BzK$ objects at $\langle z \rangle$=1.9 in the northern GOODS field. Therefore, *while on average high–redshift galaxies have IR spectral energy distributions consistent with the models, individually there remains significant scatter*. Summary ======= In this contribution, I discussed star–formation and AGN activity in massive galaxies (${\gtrsim}$$10^{11}$ ) at $z$$\sim$1–3 using observations from the [*Spitzer*]{} Space Telescope at 3–24 . Interestingly, the majority (${\gtrsim}$50%) of these objects have $f_\nu(24{\hbox{$\mu$m}})$$\geq$50 , which if attributed to star formation implies SFRs of $\simeq$100–1000  yr$^{-1}$. Galaxies at $z$$\sim$1.5–3 with ${\hbox{$\mathcal{M}$}}$$\geq$$10^{11}$  have specific SFRs equal to or exceeding the global average value. In contrast, galaxies at $z$$\sim$0.3–0.75 with these masses have specific SFRs less than the global average, and more than 10$\times$ lower than that at $z$$\sim$1.5–3. Therefore, by $z$${\lesssim}$1.5 massive galaxies have formed most of their stellar mass, and lower–mass galaxies dominate the SFR density. As many as 25% of the massive galaxies at $z$${\gtrsim}$1.5 host AGN. The high AGN occurrence in massive galaxies at $z$$\sim$ 1.5–3 provides evidence that they are simultaneously forming stars and growing supermassive black holes. This may provide the impetus for the present–day black-hole–bulge-mass relation and/or provide the feedback necessary to squelch star–formation in such galaxies, moving them onto the red sequence. The largest source of uncertainty results from systematic errors on the bolometric corrections between the observed [*Spitzer*]{} 24 data and total IR luminosity. While on average high–redshift galaxies have IR spectral energy distributions consistent with the models, individually there remains significant scatter. Future work is needed to improve our understanding of the distribution between the mid–IR (rest–frame 5–15 ) emission and total IR luminosity in galaxies at $z$$\sim$1.5–3. It may be possible to use “average” 24–to–70 and 160  colors of distant galaxies to further constrain the distribution and scatter of galaxies’ IR spectral energy distributions. The upcoming *Herschel* Space Observatory (and eventually *SAFIR*) will mitigate this problem by measuring the far–IR emission of distant galaxies directly. I wish to thank the conference organizers for their hard work in planning an outstanding meeting in an extremely beautiful locale. I look forward to the next “Extreme Starbursts” meeting. I am grateful for my colleagues on the MIPS GTO and GOODS teams for this research; I am thankful for their continued collaboration. In particular I am indebted to L. Moustakas, M. Dickinson, E. Le Floc’h, E. Daddi, and G. Rieke. I also am grateful for the AAS International Travel Grant, which made the trip to this meeting possible. Support for this work was provided by NASA through the Spitzer Space Telescope Fellowship Program, through a contract issued by the Jet Propulsion Laboratory (JPL), California Institute of Technology (Caltech) under a contract with NASA. [99]{} Alexander, D. M., Bauer, F. E., Brandt, W. N., et al. 2003, , 126, 539 Alexander, D. M., Bauer, F. E., Chapman, S. C., [et al.]{}. 2005, , 632, 736 Alonso–Herrero, A., et al. 2006, , in press (astro–ph/0511507) Armus, L., Charmandaris, V., Spoon, H. W. W., et al. 2004, , 154, 178 Barmby, P., Alonso–Herrero, A., Donley, J. L., et al. 2006, , in press (astro–ph/0512618) Bauer, A. E., Drory, N., Hill, G. J., & Feulner, G. 2005, , 621, L89 Bauer, C. M., Cole, S., Frenk, C. S., & Lacey, C. G. 1998, , 498, 504 Bell, E. F., Wolf, C., Meisenheimer, K., [et al.]{} 2004, , 608, 752 Blanton, M. R., Hogg, D. W., Bahcall, N. A., et al. 2003, , 594, 186 Bruzual, G. A., & Charlot, S. 2003, , 344, 1000 Calzetti, D., Armus, L., Bohlin, R. C., et al. 2000, , 533, 682 Chapman, S. C., Helou, G., Lewis, G. F., & Dale, D. A. 2003, , 588, 186 Chary, R. R., & Elbaz, D.2001, , 556, 562 Cimatti, A., Daddi, E., Mignoli, M., et al. 2002, , 381, L68 Cole, S., Norberg, P., Baugh, C. M., et al. 2001, , 326, 255 Croton, D., Springel, V., White, S. D. M., et al. 2006, , 365, 11 Daddi, E., Dickinson, M., Chary, R., [et al.]{} 2005b, , 631, L13 Dale, D. A., & Helou, G. 2002, , 576, 159 Davé, R., Finlator, K., Hernquist, L., et al. 2005, preprint (astro–ph/0510625) De Lucia, G., Springel, V., White, S. D. M., et al. 2005, preprint (astro–ph/0509725) Dickinson, M., Papovich, C., Ferguson, H. C., & Budavári, T. 2003, , 587, 25 Donley, J. L., Rieke, G. H., Rigby, J. R., Pérez–Gonzalez, P. G. 2005, , 634, 169 Elbaz, D., Le Floc’h, E., Dole, H., & Marcillac, D., 2005 , 434, 1 Faber, S. M., Willmer, C. N. A., Wolf, C., et al. 2005, preprint (astro–ph/0506044) Förster–Schreiber, N. M., van Dokkum, P. G., Franx, M., [et al.]{} 2004, , 616, 40 Franx, M., Labbé, I., Rudnick, G., et al. 2003, , 587, 79 Frayer, D., Armus, L, Scoville, N. Z., et al. 2003, , 126, 73 Giavalisco, M., Dickinson, M., Ferguson, H. C., et al. 2004, , 600, L93 Granato, G. L., Silva, L., Monaco, P., [et al.]{} 2001, , 324, 757 Haas, M., Klaas, U., Müller, S. A. H., et al. 2003, , 402, 87 Juneau, S., Glazebrook, K., Crampton, D., et al.  2005, , 619, L135 Kauffmann, G., & Charlot, S. 1998, , 294, 705 Kauffmann, G., White, S. D. M., Heckman, T. M., et al. 2004, , 353, 713 Kennicutt, R. C., Jr. 1998, , 36 189 Labbé, I., Huang, J., Franx, M., et al. 2005, , 624, L81 Le Floc’h, E., Papovich, C., Dole, H., et al. 2005, , 632, 169 Marcillac, D., Elbaz, D., Chary, R., et al. 2006, , in press McCarthy, P. 2004, , 42, 477 Papovich, C., Dole, H., Egami E., et al. 2004, , 154, 70 Papovich, C., Dickinson, M., Giavalisco, M., Conselice, C. J., & Ferguson, H. C. 2005, , 631, 101 Papovich, C., Moustakas, L. A., Dickinson, M., et al. 2006, , in press (astro–ph/0511289) Rudnick, G., Rix, H.–W., Franx, M., et al. 2003, , 599, 847 Stern, D., Eisenhardt, P., Gorjian, V., et al. 2005, , 631, 163 Somerville, R. S., Primack, J., & Faber, S. M. 2001, , Teplitz, H., Charmandaris, V., Chary, R., [et al.]{} 2005, , 634, 128 Wolf, C., Meisenheimer, K., Rix, H.–W., et al. 2003, , 401, 73 \[lastpage\]
{ "pile_set_name": "ArXiv" }
--- author: - 'Yi Cai,' - 'Tao Han,' - 'Tong Li,' - and Richard Ruiz bibliography: - 'colliderLNVrefs.bib' title: | **Lepton Number Violation:\ Seesaw Models and Their Collider Tests** --- Introduction ============ Neutrino flavor oscillation experiments from astrophysical and terrestrial sources provide overwhelming evidence that neutrinos have small but nonzero masses. Current observations paint a picture consistent with a mixing structure parameterized by the $3\times 3$ Pontecorvo-Maki-Nakagawa-Sakata (PMNS) matrix [@Pontecorvo:1957qd; @Pontecorvo:1957cp; @Maki:1962mu] with at least two massive neutrinos. This is contrary to the Standard Model of particle physics (SM) [@Weinberg:1967tq], which allows three massless neutrinos and hence no flavor oscillations. Consequently, to accommodate these observations, the SM must [@Ma:1998dn] be extended to a more complete theory by new degrees of freedom. One could of course introduce right-handed (RH) neutrino states (${\nu_R}$) and construct Dirac mass terms, $m_D^{} \overline {\nu_L} \nu_R$, in the same fashion as for all the other elementary fermions in the SM. However, in this minimal construction, the new states do not carry any SM gauge charges, and thus these “sterile neutrinos” have the capacity to be Majorana fermions [@Majorana:1937vz]. The most significant consequence of this would be the existence of the RH Majorana mass term $M_R\overline{(\nu_R)^c} \nu_R$ and the explicit violation of lepton number ($L$). In light of this prospect, a grand frontier opens for theoretical model-building with rich and new phenomenology at the energy scales accessible by collider experiments, and which we will review in this article. Generically, if we integrate out the new states, presumably much heavier than the electroweak (EW) scale, the new physics may be parameterized at leading order through the dimension-5 lepton number violating operator [@Weinberg:1979sa], the so-called “Weinberg operator,” $$\mathcal{L}_5 = \frac{\alpha}{\Lambda} ~ (LH)(LH) \xrightarrow{\rm EWSB} \mathcal{L}_5 \ni \frac{\alpha v^2_0}{2\Lambda}\ \overline{(\nu_L)^c}\ \nu_L , \label{Weinberg}$$ where $L$ and $H$ are, respectively, the SM left-handed (LH) lepton doublet and Higgs doublet, with vacuum expectation value (vev) $v_0 \approx 246$ GeV. After electroweak (EW) symmetry breaking (EWSB), $\mathcal{L}_5$ generates a Majorana mass term for neutrinos. One significance of Eq. (\[Weinberg\]) is the fact that its ultraviolet (UV) completions are severely restricted. For example: extending the SM field content minimally, [*i.e.*]{}, by only a single SM multiplet, permits only three [@Ma:1998dn] tree-level completions of Eq. (\[Weinberg\]), a set of constructions famously known as the Type I [@Minkowski:1977sc; @Yanagida:1979as; @GellMann:1980vs; @Glashow:1979nm; @Mohapatra:1979ia; @Shrock:1980ct; @Schechter:1980gr], Type II [@Konetschny:1977bn; @Cheng:1980qt; @Lazarides:1980nt; @Schechter:1980gr; @Mohapatra:1980yp], and Type III [@Foot:1988aq] Seesaw mechanisms. These minimal mechanisms can be summarized with the following: [**Minimal Type I Seesaw**]{} [@Minkowski:1977sc; @Yanagida:1979as; @GellMann:1980vs; @Glashow:1979nm; @Mohapatra:1979ia; @Shrock:1980ct; @Schechter:1980gr]: In the minimal Type I Seesaw, one hypothesizes the existence of a right-handed (RH) neutrino $\nu_R$, which transforms as a singlet, i.e., as $(1,1,0)$, under the SM gauge group SU(3)$_c \otimes$SU(2)$_L \otimes$U(1)$_Y$, that possesses a RH Majorana mass $M_{\nu_R}$ and interacts with a single generation of SM leptons through a Yukawa coupling $y_\nu$. After mass mixing and assuming $M_{\nu_R}\gg y_\nu v_0$, the light neutrino mass eigenvalue $m_\nu$ is given by $m_\nu \sim y^2_\nu v_0^2/ M_{\nu_R}$. If $y_\nu \simeq 1$, then to obtain a light neutrino mass of order an eV, $M_{\nu_R}$ is required to be of order $10^{14}-10^{15}$ GeV. $M_{\nu_R}$ can be made much lower though by balancing against a correspondingly lower $y_\nu$. [**Minimal Type II Seesaw**]{} [@Konetschny:1977bn; @Cheng:1980qt; @Lazarides:1980nt; @Schechter:1980gr; @Mohapatra:1980yp]: The minimal Type II Seesaw features the introduction of a Higgs field $\Delta$ with mass $M_{\Delta}$ in a triplet representation of SU$(2)_L$, and hence transforms as $(1,3,2)$ under the SM gauge group. In this mechanism, light neutrino masses are given by LH Majorana masses $m_\nu \approx Y_\nu v_{\Delta}$, where $v_{\Delta}$ is the vev of the neutral component of the new scalar triplet and $Y_\nu$ is the corresponding Yukawa coupling. Due to mixing between the SM Higgs doublet and the new scalar triplet via a dimensionful parameter $\mu$, EWSB leads to a relation $v_{\Delta} \sim \mu v_0^2/M_{\Delta}^2$. In this case the new scale $\Lambda$ is replaced by $M_{\Delta}^2/\mu$. With $Y_\nu \approx 1$ and $\mu \sim M_{\Delta}$, the scale is also $10^{14}- 10^{15}$ GeV. Again, $M_\Delta$ can be of TeV scale if $Y_\nu$ is small or $\mu \ll M_\Delta$. It is noteworthy that in the Type II Seesaw no RH neutrinos are needed to explain the observed neutrino masses and mixing. [**Minimal Type III Seesaw**]{} [@Foot:1988aq]: The minimal Type III Seesaw is similar to the other two cases in that one introduces the fermionic multiplet $\Sigma_L$ that is a triplet (adjoint representation) under SU$(2)_L$ and transforms as $(1,3,0)$ under the SM gauge group. The resulting mass matrix for neutrinos has the same form as in Type I Seesaw, but in addition features heavy leptons that are electrically charged. The new physics scale $\Lambda$ in Eq. (\[Weinberg\]) is replaced by the mass of the leptons $M_\Sigma$, which can also be as low as a TeV if balanced with a small Yukawa coupling. However, to fully reproduce oscillation data, at least two of the three known neutrinos need nonzero masses. This requires a nontrivial Yukawa coupling matrix for neutrinos if appealing to any of the aforementioned Seesaws mechanisms, and, if invoking the Type I or III Seesaws, extending the SM by at least two generations of multiplets [@Wyler:1982dd], which need not be in the same SM gauge representation. In light of this, one sees that Weinberg’s assumption of a high-scale Seesaw [@Weinberg:1979sa] is not necessary to generate tiny neutrino masses in connection with lepton $(L)$ number violation. For example: the so-called Inverse [@Mohapatra:1986aw; @Mohapatra:1986bd; @Bernabeu:1987gr; @Gavela:2009cd] or Linear [@Akhmedov:1995ip; @Akhmedov:1995vm] variants of the Type I and III Seesaw models, their generic extensions as well as hybridizations, [*i.e.*]{}, the combination of two or more Seesaw mechanisms, can naturally lead to mass scales associated with neutrino mass-generation accessible at present-day experiments, and in particular, collider experiments. A qualitative feature of these low-scale Seesaws is that light neutrino masses are proportional to the scale of $L$ violation, as opposed to inversely related as in high-scale Seesaws [@Moffat:2017feq]. The Weinberg operator in Eq. (\[Weinberg\]) is the lowest order and simplest parameterization of neutrino mass generation using only the SM particle spectrum and its gauge symmetries. Beyond its tree-level realizations, neutrino Majorana masses may alternatively be generated radiatively. Suppression by loop factors may provide a partial explanation for the smallness of neutrino masses and again allow much lower mass scales associated with neutrino mass-generation. The first of such models was proposed at one-loop in Refs. [@Zee:1980ai; @Hall:1983id], at two-loop order in Refs. [@Cheng:1980qt; @Zee:1985id; @Babu:1988ki], and at three-loop order in Ref. [@Krauss:2002px]. A key feature of radiative neutrino mass models is the absence of tree-level contributions to neutrino masses either because there the necessary particles, such as SM singlet fermion as in Type I Seesaw, are not present or because relevant couplings are forbidden by additional symmetries. Consequently, it is necessary that the new field multiplets run in the loop(s) that generate neutrino masses. As observing lepton number violation would imply the existence of Majorana masses for neutrinos [@Schechter:1981bd; @Hirsch:2006yk; @Duerr:2011zd], confirming the existence of this new mass scale would, in addition, verify the presence of a Seesaw mechanism. To this end, there have been on-going efforts in several directions, most notably the neutrinoless double beta $(0\nu\beta\beta)$-decay experiments, both current [@KamLAND-Zen:2016pfg; @Agostini:2017iyd; @Alfonso:2015wka; @Albert:2014awa] and upcoming [@Arnaboldi:2002du; @Arnold:2010tu; @Alvarez:2012flf], as well as proposed general purpose fixed-target facilities [@Alekhin:2015byh; @Anelli:2015pba]. Complementary to this are on-going searches for lepton number violating processes at collider experiments, which focus broadly on rare meson decays [@Liventsev:2013zz; @Aaij:2012zr; @Aaij:2014aba], heavy neutral leptons in Type I-like models [@Sirunyan:2017yrk; @Khachatryan:2016jqo; @Khachatryan:2016olu; @Khachatryan:2015gha; @Aad:2015xaa], heavy bosons in Type II-like models [@ATLAS:2016pbt; @ATLAS:2014kca; @Chatrchyan:2012ya], heavy charged leptons in Type III-like models [@Aad:2015cxa; @CMS:2012ra; @Sirunyan:2017qkz], and lepton number violating contact interactions [@CMS:2016blm; @Sirunyan:2017xnz]. Furthermore, accurate measurements of the PMNS matrix elements and stringent limits on the neutrino masses themselves provide crucial information and knowledge of lepton flavor mixing that could shed light on the construction of Seesaw models. In this context, we present a review of searches for lepton number violation at current and future collider experiments. Along with the current bounds from the experiments at LEP, Belle, LHCb and ATLAS/CMS at 8 and 13 TeV, we present studies for the 13 and 14 TeV LHC. Where available, we also include results for a future 100 TeV hadron collider, an $ep$ collider (LHeC), and a future high-energy $e^+e^-$ collider. We consider a number of tree- and loop-level Seesaw models, including, as phenomenological benchmarks, the canonical Type I, II, and III Seesaw mechanisms, their extensions and hybridizations, and radiative Seesaw formulations in $pp$, $ep$, and $ee$ collisions. We note that the classification of collider signatures based on the canonical Seesaws is actually highly suitable, as the same underlying extended and hybrid Seesaw mechanism can be molded to produce wildly varying collider predictions. We do not attempt to cover the full aspects of UV-complete models for each type. This review is only limited to a selective, but representative, presentation of tests of Seesaw models at collider experiments. For complementary reviews, we refer readers to Refs. [@Gluza:2002vs; @Barger:2003qi; @Mohapatra:2006gs; @Rodejohann:2011mu; @Chen:2011de; @Atre:2009rg; @Deppisch:2015qwa] and references therein. This review is organized according to the following: In Sec. \[sec:nuparameters\] we first show the PMNS matrix and summarize the mixing and mass-difference parameters from neutrino oscillation data. With those constraints, we also show the allowed mass spectra for the three massive neutrino scheme. Our presentation is agnostic, phenomenological, and categorized according to collider signature, i.e., according to the presence of Majorana neutrinos (Type I) as in Sec. \[sec:type1\], doubly charged scalars (Type II) as in Sec. \[sec:type2\], new heavy charged/neutral leptons (Type III) as in Sec. \[sec:type3\], and new Higgs, diquarks and leptoquarks in Sec. \[sec:loop\]. Particular focus is given to state-of-the-art computations, newly available Monte Carlo tools, and new collider signatures that offer expanded coverage of Seesaw parameter spaces at current and future colliders. Finally in Sec. \[sec:con\] we summarize our main results. Neutrino Mass and Oscillation Parameters {#sec:nuparameters} ======================================== In order to provide a general guidance for model construction and collider searches, we first summarize the neutrino mass and mixing parameters in light of oscillation data. Neutrino mixing can be parameterized by the PMNS matrix [@Pontecorvo:1957qd; @Pontecorvo:1957cp; @Maki:1962mu] as $$U_{PMNS} = \begin{pmatrix} 1 & 0 & 0\\ 0 & c_{23} & s_{23} \\ 0 & -s_{23} & c_{23} \end{pmatrix} \begin{pmatrix} c_{13} & 0 & e^{-i\delta}s_{13} \\ 0 & 1 & 0 \\ -e^{i\delta}s_{13} & 0 & c_{13} \end{pmatrix} \begin{pmatrix} c_{12} & s_{12} & 0 \\ -s_{12} & c_{12} & 0 \\ 0 & 0 & 1 \end{pmatrix} \text{diag} (e^{i \Phi_1/2}, 1, e^{i \Phi_2/2})$$ = ( [lll]{} c\_[12]{} c\_[13]{} & c\_[13]{} s\_[12]{} & e\^[-]{} s\_[13]{}\ -c\_[12]{} s\_[13]{} s\_[23]{} e\^-c\_[23]{} s\_[12]{} & c\_[12]{} c\_[23]{}-e\^ s\_[12]{} s\_[13]{} s\_[23]{} & c\_[13]{} s\_[23]{}\ s\_[12]{} s\_[23]{}-e\^ c\_[12]{} c\_[23]{} s\_[13]{} & -c\_[23]{} s\_[12]{} s\_[13]{} e\^-c\_[12]{} s\_[23]{} & c\_[13]{} c\_[23]{} ) (e\^[i \_1/2]{}, 1, e\^[i \_2/2]{}), where $s_{ij}\equiv\sin{\theta_{ij}}$, $c_{ij}\equiv\cos{\theta_{ij}}$, $0 \le \theta_{ij} \le \pi/2$, and $0 \le \delta, \Phi_i \le 2\pi$, with $\delta$ being the Dirac CP phase and $\Phi_i$ the Majorana phases. While the PMNS is a well-defined $3\times3$ unitary matrix, throughout this review we use the term generically to describe the $3\times3$ active-light mixing that may not, in general, be unitary. [l|cc|cc|c]{} & & & Any Ordering\ & bfp $\pm 1\sigma$ & $3\sigma$ range & bfp $\pm 1\sigma$ & $3\sigma$ range & $3\sigma$ range\ ------------------------------------------------------------------------ $\sin^2\theta_{12}$ & $0.306_{-0.012}^{+0.012}$ & $0.271 \to 0.345$ & $0.306_{-0.012}^{+0.012}$ & $0.271 \to 0.345$ & $0.271 \to 0.345$\ $\theta_{12}/^\circ$ & $33.56_{-0.75}^{+0.77}$ & $31.38 \to 35.99$ & $33.56_{-0.75}^{+0.77}$ & $31.38 \to 35.99$ & $31.38 \to 35.99$\ $\sin^2\theta_{23}$ & $0.441_{-0.021}^{+0.027}$ & $0.385 \to 0.635$ & $0.587_{-0.024}^{+0.020}$ & $0.393 \to 0.640$ & $0.385 \to 0.638$\ $\theta_{23}/^\circ$ & $41.6_{-1.2}^{+1.5}$ & $38.4 \to 52.8$ & $50.0_{-1.4}^{+1.1}$ & $38.8 \to 53.1$ & $38.4 \to 53.0$\ $\sin^2\theta_{13}$ & $0.02166_{-0.00075}^{+0.00075}$ & $0.01934 \to 0.02392$ & $0.02179_{-0.00076}^{+0.00076}$ & $0.01953 \to 0.02408$ & $0.01934 \to 0.02397$\ $\theta_{13}/^\circ$ & $8.46_{-0.15}^{+0.15}$ & $7.99 \to 8.90$ & $8.49_{-0.15}^{+0.15}$ & $8.03 \to 8.93$ & $7.99 \to 8.91$\ $\delta_\text{CP}/^\circ$ & $261_{-59}^{+51}$ & $\hphantom{00}0 \to 360$ & $277_{-46}^{+40}$ & $145 \to 391$ & $\hphantom{00}0 \to 360$\ $\dfrac{{\Delta m^2}_{21}}{10^{-5}~{\ensuremath{\text{eV}^2}}}$ & $7.50_{-0.17}^{+0.19}$ & $7.03 \to 8.09$ & $7.50_{-0.17}^{+0.19}$ & $7.03 \to 8.09$ & $7.03 \to 8.09$\ $\dfrac{{\Delta m^2}_{3\ell}}{10^{-3}~{\ensuremath{\text{eV}^2}}}$ & $+2.524_{-0.040}^{+0.039}$ & $+2.407 \to +2.643$ & $-2.514_{-0.041}^{+0.038}$ & $-2.635 \to -2.399$ & $\begin{bmatrix} +2.407 \to +2.643\\[-2pt] -2.629 \to -2.405 \end{bmatrix}$\ The neutrino mixing matrix is very different from the quark-sector Cabbibo-Kobayashi-Maskawa (CKM) matrix, in that most of the PMNS mixing angles are large whereas CKM angles are small-to-negligible. In recent years, several reactor experiments, such as Daya Bay [@An:2012eh], Double Chooz [@Abe:2011fz], and RENO [@Ahn:2012nd], have reported non-zero measurements of $\theta_{13}$ by searching for the disappearance of antielectron neutrinos. Among these reactor experiments, Daya Bay gives the most conclusive result with $\sin^22\theta_{13}\approx 0.084$ or $\theta_{13}\approx8.4^\circ$ [@An:2015rpe; @Esteban:2016qun], the smallest entry of the PMNS matrix. More recently, there have been reports on indications of a non-zero Dirac CP phase, with $\delta \approx 3\pi/2$ [@Abe:2013hdq; @Adamson:2016tbq; @Abe:2017uxa]. However, it cannot presently be excluded that evidence for such a large Dirac phase may instead be evidence for sterile neutrinos or new neutral currents [@Forero:2016cmb; @Ge:2016dlx; @deGouvea:2016pom; @Miranda:2016wdr]. Neutrino oscillation experiments can help to extract the size of the mass-squared splitting between three neutrino mass eigenstates. The sign of $\Delta m_{31}^2 = m_{3}^2 - m_{1}^2$, however, still remains unknown at this time. It can be either positive, commonly referred as the Normal Hierarchy (NH), or negative and referred to as the Inverted Hierarchy (IH). The terms Normal Ordering (NO) and Inverted Ordering (IO) are also often used in the literature in lieu of NH and IH, respectively. Taking into account the reactor data from the antineutrino disappearance experiments mentioned above together with other disappearance and appearance measurement, the latest global fit of the neutrino masses and mixing parameters from the `NuFit` collaboration [@Esteban:2016qun], are listed in Table \[tab:nufit\] for NH (left) and IH (center). The tightest constraint on the sum of neutrino masses comes from cosmological data. Combining Planck+WMAP+highL+BAO data, this yields at $95\%$ confidence level (CL) [@Ade:2015xua] \_[i=1]{}\^3 m\_i &lt;  0.230  . Given this and the measured neutrino mass splittings, we show in Fig. \[mabs\] the three active neutrino mass spectra as a function of the lowest neutrino mass in (a) NH and (b) IH. With the potential sensitivity of the sum of neutrino masses being close to $0.1\ev$ in the near future ($5-7$ years) [@Lesgourgues:2014zoa], upcoming cosmological probes will not be able to settle the issue of the neutrino mass hierarchy. However, the improved measurement $\sim 0.01\ev$ over a longer term ($7-15$ years) [@Hamann:2012fe; @Lesgourgues:2014zoa] would be sensitive enough to determine the absolute mass scale of a heavier neutrino spectrum. In addition, there are multiple proposed experiments aiming to determine the neutrino mass hierarchy. The Deep Underground Neutrino Experiment (DUNE) will detect neutrino beams from the Long Baseline Neutrino Facility (LBNF), and probe the Dirac CP-phase and mass hierarchy. With a baseline of 1300 km, DUNE is able to determine the mass hierarchy with at least $5\sigma$ significance [@Acciarri:2016crz]. The Jiangmen Underground Neutrino Observatory (JUNO) plans to precisely measure the reactor antielectron neutrinos and improve the accuracy of $\Delta m_{21}^2$, $\Delta m_{32}^2$ and $\sin^2\theta_{12}$ to 1% level [@An:2015jdp]. The Hyper-Kamiokande (Hyper-K) experiment, an upgrade of the T2K experiment, can measure the precision of $\delta$ to be $7^\circ-21^\circ$ and reach $3~(5)\sigma$ significance for mass hierarchy determination after 5 (10) years exposure [@Abe:2015zbg]. Finally, the Karlsruhe Tritium Neutrino experiment (KATRIN), a tritium $\beta$ decay experiment, aims to measure the effective “electron-neutrino mass” with sub-eV sensitivity [@Osipowicz:2001sq]. The Type I Seesaw and Lepton Number Violation at Colliders {#sec:type1} ========================================================== We begin our presentation of collider searches for lepton number violation in the context of Type I Seesaw models. After describing the canonical Type I mechanism [@Minkowski:1977sc; @Yanagida:1979as; @GellMann:1980vs; @Glashow:1979nm; @Mohapatra:1979ia] and its phenomenological decoupling at collider scales in Sec. \[sec:type1Canon\], we discuss various representative, low-scale models that incorporate the Type I mechanism and its extensions. We then present collider searches for lepton number violation mediated by Majorana neutrinos $(N)$, which is the characteristic feature of Type I-based scenarios, in Sec. \[sec:typeIhybrid\]. This is further categorized according to associated phenomena of increasing complexity: $N$ production via massive Abelian gauge bosons is reviewed in Sec. \[sec:type1Abelian\], via massive non-Abelian gauge bosons in Sec. \[sec:lrsmCollider\], and via dimension-six operators in Sec. \[sec:neftTests\]. Type I Seesaw Models -------------------- ### The Canonical Type I Seesaw Mechanism {#sec:type1Canon} In the canonical Type I Seesaw mechanism one hypothesizes a single RH neutral leptonic state, $N_R\sim (1,1,0)$, in addition to the SM matter content. However, reproducing neutrino oscillation data requires more degrees of freedom. Therefore, for our purposes, we assume $i =1,\dots,3$ LH states and $j=1,\dots,n$ RH states. Following the notation of Refs. [@Atre:2009rg; @Han:2012vk], the full theory is $$\mathcal{L}_{\rm Type~I} = \mathcal{L}_{\rm SM} + \mathcal{L}_{N~{\rm Kin}} + \mathcal{L}_{N}, \label{eq:type1Lag}$$ where $\mathcal{L}_{\rm SM}$ is the SM Lagrangian, $\mathcal{L}_{N~{\rm Kin}}$ is $N_R$’s kinetic term, and interactions and mass terms, $$\begin{aligned} {\cal L}_N &=& -\overline{L} \ Y_\nu^D \ \tilde{H} \ N_{R} \ - \ {1\over 2} \overline{(N^c)}_{L} \ M_R \ N_{R} + \ \text{H.c.}\end{aligned}$$ $L$ and $H$ are the SM LH lepton and Higgs doublets, respectively, and $\tilde{H}=i \sigma_2 H^*$. Once $H$ settles on the vev $\langle H\rangle =v_0/\sqrt 2$, neutrinos acquire Dirac masses $m_D = Y_\nu^D \ v_0 / \sqrt{2}$ and we have $$\begin{aligned} {\cal L}_N \ni -{1\over 2} \ \left( \overline{\nu}_{L} \ m_D \ N_{R} \ + \ \overline{(N^c)}_{L} \ m_D^T \ (\nu^c)_{R} \ + \ \overline{(N^c)}_{L} \ M_R \ N_{R} \right) \ + \ \text{H.c.}\end{aligned}$$ After introducing a unitary transformation into $m~(m')$ light (heavy) mass eigenstates, $$\begin{aligned} \left( \begin{array}{c} \nu \\ N^c \\ \end{array} \right)_L = \mathbb{N}\left( \begin{array}{c} \nu_m \\ N^c_{m'} \\ \end{array} \right)_L, \ \ \ \mathbb{N}= \left( \begin{array}{cc} U & V \\ X & Y \\ \end{array} \right), \label{eq:nuMixDefs}\end{aligned}$$ one obtains the diagonalized mass matrix for neutrinos $$\begin{aligned} \mathbb{N}^\dagger \left( \begin{array}{cc} 0 & m_D \\ m_{D}^{T} & M \\ \end{array} \right) \mathbb{N}^\ast &=& \left( \begin{array}{cc} m_\nu & 0 \\ 0 & M_N \\ \end{array} \right), \label{eq:type1NuMixMatrix}\end{aligned}$$ with mass eigenvalues $m_\nu=diag(m_1,m_2,m_3)$ and $M_N=diag(M_1,\cdots, M_{m'})$. In the limit $m_D\ll M_R$, the light $(m_\nu)$ and heavy $(M_N)$ neutrino masses are respectively $$\begin{aligned} m_\nu\approx - m_D M^{-1}_R m_D^T \quad\text{and}\quad M_N\approx M_R.\end{aligned}$$ The mixing elements typically scale like $$\begin{aligned} UU^\dagger\approx I- m_\nu M_N^{-1}, \ \ \ VV^\dagger\approx m_\nu M_N^{-1}, \label{eq:massscale}\end{aligned}$$ with the unitarity condition $UU^\dagger+VV^\dagger=I$. With another matrix $U_\ell$ diagonalizing the charged lepton mass matrix, we have the approximate neutrino mass mixing matrix $U_{PMNS}$ and the matrix $V_{\ell N}$, which transits heavy neutrinos to charged leptons. These are given by $$\begin{aligned} U_\ell^\dagger U\equiv U_{PMNS}, \ \ \ U_\ell^\dagger V\equiv V_{\ell N}, \quad\text{and}\quad U_{PMNS}U_{PMNS}^\dagger+V_{\ell N}V_{\ell N}^\dagger=I.\end{aligned}$$ The decomposition of active neutrino states into a general number of massive eigenstates is then given by [@Atre:2009rg; @Han:2012vk], $\nu_{\ell} = \sum_{m=1}^{3} U_{\ell m}\nu_{m} + \sum_{m'=1}^{n}V_{\ell m'} N^{c}_{m'}.$ From this, the SM EW boson couplings to heavy mass eigenstates (in the mixed mass-flavor basis) are $$\begin{aligned} \mathcal{L}_{\rm Int.} = &-& \frac{g}{\sqrt{2}}W^+_\mu \sum_{\ell=e}^\tau \left( \sum_{m=1}^3 ~\overline{\nu_m} ~U_{\ell m}^* + \sum_{m'=1}^n \overline{N^c_{m'}} ~V_{\ell N_{m'}}^* \right)\gamma^\mu P_L\ell^-\nonumber\\ &-& \frac{g}{2\cos\theta_W}Z_\mu \sum_{\ell=e}^\tau \left( \sum_{m=1}^3 ~\overline{\nu_m} ~U_{\ell m}^* + \sum_{m'=1}^n \overline{N^c_{m'}} ~V_{\ell N_{m'}}^* \right)\gamma^\mu P_L\nu_\ell\nonumber\\ &-& \frac{g}{2M_W} h \sum_{\ell=e}^\tau \sum_{m'=1}^n ~m_{N_{m'}}\overline{N^c_{m'}} ~V_{\ell N_{m'}}^* P_L\nu_\ell + \text{H.c.} \label{eq:typeIEWLag}\end{aligned}$$ There is a particular utility of using this mixed mass-flavor basis in collider searches for heavy neutrinos. Empirically, $\vert V_{\ell N_{m'}}\vert \lesssim 10^{-2}$ [@delAguila:2008pw; @Antusch:2014woa; @deGouvea:2015euy; @Fernandez-Martinez:2016lgt], which means pair production of $N_{m'}$ via EW processes is suppressed by $\vert V_{\ell N_{m'}}\vert^2 \lesssim 10^{-4}$ relative to single production of $N_{m'}$. Moreover, in collider processes involving $\nu_{m}-N_{m'}$ vertices, one sums over $\nu_{m}$ either because it is an internal particle or an undetected external state. This summation effectively undoes the decomposition of one neutrino interaction state for neutral current vertices, resulting in the basis above. In phenomenological analyses, it is common practice to consider only the lightest heavy neutrino mass eigenstate, i.e., $N_{m'=4}$, to reduce the effective number of independent model parameters. In such cases, the mass eigenstate is denoted simply as $N$ and one reports sensitivity on the associated mixing element, labeled as $\vert V_{\ell N} \vert $ or $\vert V_{\ell 4} \vert$, and which are equivalent to $\vert V_{\ell N_{m'=4}} \vert.$ Throughout this text, the $\vert V_{\ell N} \vert$ notation is adopted where possible. From Eq. (\[eq:type1NuMixMatrix\]), an important relation among neutrino masses can be derived. Namely, that $$\begin{aligned} U^\ast_{PMNS}m_\nu U^\dagger_{PMNS}+V^\ast_{\ell N}M_NV^\dagger_{\ell N}=0 \ . \label{typei}\end{aligned}$$ Here the masses and mixing of the light neutrinos in the first term are measurable from the oscillation experiments, and the second term contains the masses and mixing of the new heavy neutrinos. We now consider a simple case: degenerate heavy neutrinos with mass $M_N = diag(M_1,\cdots, M_{m'}) = M_N \mathds{I}_{m'}$. Using this assumption, we obtain from Eq. (\[typei\]), $$\begin{aligned} M_N\sum_N (V_{\ell N}^\ast)^2=(U^\ast_{PMNS}m_\nu U^\dagger_{PMNS})_{\ell\ell} \ . \label{dege}\end{aligned}$$ Using the oscillation data in Table \[tab:nufit\] as inputs[^1], we display in Fig. \[fig:vvmcase1\] the normalized mixing of each lepton flavor in this scenario[^2]. Interestingly, one can see the characteristic features: $$\begin{aligned} \sum_N |V_{eN}|^2 & \ll \sum_N |V_{\mu N}|^2, \sum_N |V_{\tau N}|^2 & \quad {\rm for \ NH}, \\ \sum_N |V_{eN}|^2 & > \sum_N |V_{\mu N}|^2, \sum_N |V_{\tau N}|^2 & \quad {\rm for \ IH}.\end{aligned}$$ As shown in Fig. \[fig:brnlwcase1\], a corresponding pattern also emerges in the branching fraction[^3] of the degenerate neutrinos decaying into charged leptons plus a $W$ boson, $$\begin{aligned} &{\rm BR}(\mu^\pm W^\mp), {\rm BR}(\tau^\pm W^\mp)\sim (20-30)\%\gg {\rm BR}(e^\pm W^\mp)\sim (3-4)\% \ & {\rm for} \ {\rm NH},\\ &{\rm BR}(e^\pm W^\mp)\sim 27\% > {\rm BR}(\mu^\pm W^\mp), {\rm BR}(\tau^\pm W^\mp)\sim (10-20)\% \ &{\rm for} \ {\rm IH},\end{aligned}$$ with ${\rm BR}(\ell^\pm W^\mp)={\rm BR}(N_i\to \ell^+W^-+\ell^-W^+)$. These patterns show a rather general feature that ratios of Seesaw partner observables, , cross sections and branching fractions, encode information on light neutrinos, such as their mass hierarchy [@Perez:2009mu; @Ibarra:2011xn]. Hence, one can distinguish between competing light neutrino mass and mixing patterns with high energy observables. ![Branching fractions of process $N_i\rightarrow \ell^+W^- + \ell^-W^+$ versus the lightest neutrino mass for NH and IH in the case $\Omega=I$ with $M_i=300 \ \rm{GeV}$ and $m_h=125 \ \rm{GeV}$, assuming vanishing Majorana phases. []{data-label="fig:nibr"}](nibr.jpg){width="11cm"} More generally, the $V_{\ell N}$ in Eq. (\[typei\]) can be formally solved in terms of an arbitrary orthogonal complex matrix $\Omega$, known as the Casas-Ibarra parametrization [@Casas:2001sr], using the ansatz $$\begin{aligned} V_{\ell N}=U_{PMNS} ~m_\nu^{1/2}\Omega M_N^{-1/2}, \label{omega}\end{aligned}$$ with the orthogonality condition $\Omega \Omega^T=I$. For the simplest incarnation of a unity matrix $\Omega=I$, the $|V_{\ell N_{m'}}|^2$ are proportional to one and only one light neutrino mass, and thus the branching ratio of $N_{m'}\to \ell^\pm W^\mp$ for each lepton flavor is independent of neutrino mass and universal for both NH and IH [@Perez:2009mu]. Nevertheless, one can still differentiate between the three heavy neutrinos according to the decay rates to their leading decay channels. As shown in Fig. \[fig:nibr\] for $\Omega=I$, one sees $$\begin{aligned} &{\rm BR}(e^\pm W^\mp)\sim 40\%>{\rm BR}(\mu^\pm W^\mp),{\rm BR}(\tau^\pm W^\mp)\sim (4-15)\% \ & {\rm for} \ N_1,\\ &{\rm BR}(e^\pm W^\mp)\sim 20\%\approx {\rm BR}(\mu^\pm W^\mp)\approx {\rm BR}(\tau^\pm W^\mp)\sim (10-30)\% & \ {\rm for} \ N_2,\\ &{\rm BR}(\mu^\pm W^\mp),{\rm BR}(\tau^\pm W^\mp)\sim (15-40)\%\gg {\rm BR}(e^\pm W^\mp)\sim 1\% \ & {\rm for} \ N_3.\end{aligned}$$ A realistic Dirac mass matrix can be quite arbitrary with three complex angles parameterizing the orthogonal matrix $\Omega$. However, the arbitrariness of the Dirac mass matrix is not a universal feature of Seesaw models; the neutrino Yukawa matrix in the Type II Seesaw, for example, is much more constrained. Beyond this, Fig. \[fig:vvmcase1\] also shows another general feature of minimal, high-scale Seesaw constructions, namely that the active-sterile mixing $\vert V_{\ell N}\vert$ is vanishingly small. For a heavy neutrino mass of $M_N \sim 100$ GeV, Eq. (\[dege\]) implies $\vert V_{\ell N}\vert^2 \sim 10^{-14} - 10^{-12}$. This leads to the well-known decoupling of observable lepton number violation in the minimal, high-scale Type I Seesaw scenario at colliders experiments [@Pilaftsis:1991ug; @Kersten:2007vk; @Moffat:2017feq]. For low-scale Type I Seesaws, such decoupling of observable lepton number violation also occurs: Due to the allowed arbitrariness of the matrix $\Omega$ in Eq. (\[omega\]), it is possible to construct $\Omega$ and $M_N$ with particular entry patterns or symmetry structures, also known as “textures” in the literature, such that $V_{\ell N}$ is nonzero but $m_\nu$ vanishes. Light neutrino masses can then be generated as perturbations from these textures. In Ref. [@Moffat:2017feq] it was proved that such delicate (and potentially fine-tuned [@AristizabalSierra:2011mn; @Lopez-Pavon:2015cga; @Fernandez-Martinez:2015hxa]) constructions result in small neutrino masses being proportional to small $L$-violating parameters, instead of being inversely proportional as in the high-scale case. Subsequently, in low-scale Seesaw scenarios that assume only fermionic gauge singlets, tiny neutrino masses is equivalent to an approximate conservation of lepton number, and leads to the suppression of observable $L$ violation in high energy processes. Hence, any observation of lepton number violation (and Seesaw partners in general) at collider experiments implies a much richer neutrino mass-generation scheme than just the canonical, high-scale Type I Seesaw. ### Type I+II Hybrid Seesaw Mechanism {#sec:typeIHybrid} While the discovery of lepton number violation in, say, $0\nu\beta\beta$ or hadron collisions would imply the Majorana nature of neutrinos [@Schechter:1981bd; @Hirsch:2006yk; @Duerr:2011zd], it would be less clear which mechanism or mechanisms are driving light neutrino masses to their sub-eV values. This is because in the most general case neutrinos possess both LH and RH Majorana masses in addition to Dirac masses. In such hybrid Seesaw models, two or more “canonical” tree- and loop-level mechanisms are combined and, so to speak, may give rise to phenomenology that is greater than the sum of its parts. A well-studied hybrid model is the Type I+II Seesaw mechanism, wherein the light neutrino mass matrix $M_{\nu}$, when $M_D M_R^{-1}\ll1$, is given by [@Chen:2005jm; @Akhmedov:2006de; @Akhmedov:2006yp; @Chao:2007mz; @Chao:2009ef; @Gu:2008yj; @Chao:2008mq] $$M_{\nu}^{light} = M_L - M_D M_R^{-1} M_D^T. \label{eq:hybridNuMass}$$ Here, the Dirac and Majorana mass terms, $M_D$, $M_R$, have their respective origins according to the Type I model, whereas $M_L$ originates from the Type II mechanism; see Sec. \[sec:type2\] for details. In this scenario, sub-eV neutrino masses can arise not only from parametrically small Type I and II masses but additionally from an incomplete cancellation of the two terms [@Akhmedov:2006de; @Akhmedov:2006yp; @Chao:2007mz]. While a significant or even moderate cancellation requires a high-degree of fine tuning and is radiatively instable [@Chao:2008mq], this situation cannot theoretically be ruled out *a priori*. For a one-generation mechanism, the relative minus sign in Eq. (\[eq:hybridNuMass\]) is paramount for such a cancellation; however, in a multi-generation scheme, it is not as crucial as $M_D$ is, in general, complex and can absorb the sign through a phase rotation. Moreover, this fine-tuning scenario is a caveat of the aforementioned decoupling of $L$-violation in a minimal Type I Seesaw from LHC phenomenology [@Pilaftsis:1991ug; @Kersten:2007vk; @Moffat:2017feq]. As we will discuss shortly, regardless of its providence, if such a situation were to be realized in nature, then vibrant and rich collider signatures emerges. ### Type I Seesaw in U$(1)_X$ Gauge Extensions of the Standard Model {#sec:type1U1X} Another manner in which the decoupling of heavy Majorana neutrinos $N$ from collider experiments can be avoided is through the introduction of new gauge symmetries, under which $N$ is charged. One such example is the well-studied U$(1)_{X}$ Abelian gauge extension of the SM [@Langacker:1980js; @Hewett:1988xc; @Faraggi:1990ita; @Accomando:2010fz; @Faraggi:2015iaa], where U$(1)_X$ is a linear combination of U$(1)_Y$ and U$(1)_{B-L}$ after the spontaneous breaking of electroweak symmetry and $B-L$ (baryon minus lepton number) symmetries. In this class of models, RH neutrinos are introduced to cancel gauge anomalies and realize a Type I Seesaw mechanism. Generally, such a theory can be described by modifying the SM covariant derivatives by [@Salvioni:2009mt] $$\begin{aligned} D_\mu \ni ig_1YB_\mu \quad\to\quad D_\mu \ni ig_1YB_\mu + i (\tilde{g}Y + g_1' Y_{BL}) B_\mu', \label{U1}\end{aligned}$$ where $B_\mu (Y)$ and $B_\mu' (Y_{BL})$ are the gauge fields (quantum numbers) of U$(1)_Y$ and U$(1)_{B-L}$, respectively. The most economical extension with vanishing mixing between U$(1)_Y$ and U$(1)_{B-L}$, , U$(1)_X={\rm U}(1)_{B-L}$ and $\tilde{g}=0$ in Eq. (\[U1\]), introduces three RH neutrinos and a new complex scalar $S$ that are all charged under the new gauge group but remain singlets under the SM symmetries [@Carlson:1986cu; @Buchmuller:1991ce; @Abbas:2007ag]. In this extension one can then construct the neutrino Yukawa interactions $$\begin{aligned} {\cal L}_I^Y & = & - \bar{L}_L \ Y_\nu^D \ \tilde{H} \ N_R \ -\frac{1}{2}Y_\nu^M \ \overline{(N^c)_L} \ N_R \ S+ \ \rm{H.c.} \label{Y}\end{aligned}$$ Once the Higgs $S$ acquires the vacuum expectation value $\langle S\rangle = v_S/\sqrt{2}$, $B-L$ is broken, spontaneously generating the RH Majorana mass matrix $M_N=Y_\nu^M v_S/\sqrt{2}$ from Eq. (\[Y\]). It is interesting to note that the scalar vev provides a dynamical mechanism for the heavy, RH Majorana mass generation, , a Type I Seesaw via a Type II mechanism; see Sec. \[sec:type2\] for more details. The Seesaw formula and the mixing between the SM charged leptons and heavy neutrinos here are exactly the same as those in the canonical Type I Seesaw. The mass of neutral gauge field $B_\mu'$, $M_{Z'}=M_{Z_{B-L}}=2g_{BL}v_S$, is generated from $S$’ kinetic term, $\left( D_\mu S \right)^\dagger \left( D^\mu S \right)$ with $D_\mu S = \partial_\mu S \ + \ i 2 g_{BL} B_\mu' S$. Note that in the minimal model, $g_{BL}=g_1'$. As in other extended scalar scenarios, the quadratic term $H^\dagger H S^\dagger S$ in the scalar potential results in the SM Higgs $H$ and $S$ interaction states mixing into two CP-even mass eigenstates, $H_1$ and $H_2$. ### Type I+II Hybrid Seesaw in Left-Right Symmetric Model {#sec:hybrid} As discussed in Sec. \[sec:typeIHybrid\], it may be the case that light neutrino masses result from an interplay of multiple Seesaw mechanisms. For example: the Type I+II hybrid mechanism with light neutrino masses given by Eq. (\[eq:hybridNuMass\]). It is also worth observing two facts: First, in the absence of Majorana masses, the minimum fermionic field content for a Type I+II Seesaw automatically obeys an accidental global U$(1)_{B-L}$ symmetry. Second, with three RH neutrinos, all fermions can be sorted into either SU$(2)_L$ doublets (as in the SM) or SU$(2)_R$ doublets, its RH analogue. As the hallmark of the Type II model (see Sec. \[sec:type2\]) is the spontaneous generation of LH Majorana masses from a scalar SU$(2)_L$ triplet $\Delta_L$, it is conceivable that RH neutrino Majorana masses could also be generated spontaneously, but from a scalar SU$(2)_R$ triplet $\Delta_R$. (This is similar to the spontaneous breaking of U$(1)_{B-L}$ in Sec. \[sec:type1U1X\].) This realization of the Type I+II Seesaw is known as the Left-Right Symmetric Model (LRSM) [@Pati:1974yy; @Mohapatra:1974hk; @Mohapatra:1974gc; @Senjanovic:1975rk; @Senjanovic:1978ev], and remains one of the best-motivated and well-studied extensions of the SM. For recent, dedicated reviews, see Ref. [@Duka:1999uc; @Mohapatra:2006gs; @Senjanovic:2016bya]. The high energy symmetries of the LRSM is based on the extended gauge group $$\mathcal{G}_{\rm LRSM} = {{\rm SU}(3)_{c}}\otimes{{\rm SU}(2)_{L}}\otimes{{\rm SU}(2)_{R}}\otimes{{\rm U}(1)_{B-L}}, \label{eq:lrsmGaugeGrp}$$ or its embeddings, and conjectures that elementary states, in the UV limit, participate in LH and RH chiral currents with equal strength. While the original formulation of model supposes a generalized parity $\mathcal{P}_X=\mathcal{P}$ that enforces an exchange symmetry between fields charged under ${{\rm SU}(2)_{L}}$ and ${{\rm SU}(2)_{R}}$, it is also possible to achieve this symmetry via a generalized charge conjugation $\mathcal{P}_X=\mathcal{C}$ [@Maiezza:2010ic]. For fermionic and scalar multiplets $Q_{L,R}$ and $\Phi$, the exchange relationships are [@Maiezza:2010ic], $$\mathcal{P}:\left\{\begin{matrix} Q_L \leftrightarrow Q_R \\ \Phi \leftrightarrow \Phi^\dagger \end{matrix}\right., \quad\text{and}\quad \mathcal{C}:\left\{\begin{matrix} Q_L \leftrightarrow (Q_R)^c \\ \Phi \leftrightarrow \Phi^T \end{matrix}\right., \quad\text{where}\quad (Q_R)^c = C\gamma^0Q_R^*.$$ A non-trivial, low-energy consequence of these complementary formulations of the LRSM is the relationship between the LH CKM matrix in the SM, $V_{ij}^{\rm L}$, and its RH analogue, $V_{ij}^{\rm R}$. For generalized conjugation, one has $\vert V_{ij}^{\rm R} \vert = \vert V_{ij}^{\rm L}\vert$, whereas $\vert V_{ij}^{\rm R} \vert \approx \vert V_{ij}^{\rm L}\vert+ \mathcal{O}(m_b/m_t)$ for generalized parity [@Zhang:2007fn; @Zhang:2007da; @Maiezza:2010ic; @Senjanovic:2014pva; @Senjanovic:2015yea]. Moreover, LR parity also establishes a connection between the Dirac and Majorana masses in the leptonic sector [@Nemevsek:2012iq; @Senjanovic:2016vxw]. Under generalized parity, for example, the Dirac $(Y^D_{1,2})$ and Majorana $(Y_{L,R})$ Yukawa matrices must satisfy [@Senjanovic:2016vxw], $$Y_{1,2}^D = Y_{1,2}^{D\dagger} \quad\text{and}\quad Y_L = Y_R.$$ Such relationships in the LRSM remove the arbitrariness of neutrino Dirac mass matrices, as discussed in Sec. \[sec:type1Canon\], and permits one to calculate $\Omega$, even for nonzero $\Delta_L$ vev [@Akhmedov:2008tb; @Nemevsek:2012iq]. However, the potential cancellation between Type I and II Seesaw masses in Eq. \[eq:hybridNuMass\] still remains. In addition to the canonical formulation of the LRSM are several alternatives. For example: It is possible to instead generate LH and RH Majorana neutrino masses radiatively in the absence of triplet scalars [@FileviezPerez:2017zwm; @FileviezPerez:2016erl]. One can gauge baryon number and lepton number independently, which, for an anomaly-free theory, gives rise to vector-like leptons and a Type III Seesaw mechanism [@FileviezPerez:2008sr; @Duerr:2013opa] (see Sec. \[sec:type3\]), as well as embed the model into an $R$-parity-violating Supersymmetric framework [@FileviezPerez:2008sx; @Everett:2009vy]. Despite the large scalar sector of the LRSM (two complex triplets and one complex bidoublet), and hence a litany of neutral and charged Higgses, the symmetry structure in Eq. (\[eq:lrsmGaugeGrp\]) confines the number in independent degrees of freedom to 18 [@Deshpande:1990ip; @Duka:1999uc]. These consist of three mass scales $\mu_{1,\dots,3}$, 14 dimensionless couplings $\lambda_{1,\dots,4}$, $\rho_{1,\dots,4}$, $\alpha_{1,\dots,3}$, $\beta_{1,\dots,3}$, and one CP-violating phase, $\delta_2$. For further discussions on the spontaneous breakdown of CP in LR scenarios, see also Refs. [@Senjanovic:1978ev; @Basecq:1985sx; @Kiers:2002cz]. With explicit CP conservation, the minimization conditions on the scalar potential give rise to the so-called LRSM vev Seesaw relationship [@Deshpande:1990ip], $$v_L = \frac{\beta_2 k_1^2 + \beta_1 k_1 k_2 + \beta_3 k_2^2}{(2\rho_1 - \rho_3)v_R}, \label{eq:vevSeesaw}$$ where, $v_{L,R}$ and $k_{1,2}$ are the vevs of $\Delta_{L,R}$ and the Higgs bidoublet $\Phi$, respectively, with $v_L^2 \ll k_1^2 + k_2^2 \approx (246\gev)^2 \ll v_R$. In the LRSM, the bidoublet $\Phi$ fulfills the role of the SM Higgs to generate the known Dirac masses of elementary fermions and permits a neutral scalar $h_i$ with mass $m_{h_i}\approx125$ GeV and SM-like couplings. In the absence of egregious fine-tuning, , $\rho_3 \not\approx 2\rho_1$, Eq. (\[eq:vevSeesaw\]) suggests that $v_L$ in the LRSM is inherently small because, in addition to $k_{1},k_{2}\ll v_R$, custodial symmetry is respected (up to hypercharge corrections) when all $\beta_i$ are identically zero [@Mitra:2016kov]. Consistent application of such naturalness arguments reveals a lower bound on the scalar potential parameters [@Mitra:2016kov], $$\begin{aligned} \rho_{1,2,4} > \frac{g^2_R}{4}\left(\frac{m_{\rm FCNH}}{M_{W_R}}\right)^2, &\quad& \rho_{3} > g^2_R\left(\frac{m_{\rm FCNH}}{M_{W_R}}\right)^2 + 2\rho_1 \sim 6\rho_1, \\ \alpha_{1,\dots,3} > g^2_R\left(\frac{m_{\rm FCNH}}{M_{W_R}}\right)^2, &\quad& \mu_{1,2}^2 > (m_{\rm FCNH})^2, \quad \mu_3^2 > \frac{1}{2}(m_{\rm FCNH})^2,\end{aligned}$$ where $M_{W_R}$ and $g_R$ are the mass and coupling of the $W_R^\pm$ gauge boson associated with ${{\rm SU}(2)_{R}}$, and $m_{\rm FCNH}$ is the mass scale of the LRSM scalar sector participating in flavor-changing neutral transitions. Present searches for neutron EDMs [@Zhang:2007da; @Zhang:2007fn; @Xu:2009nt; @Maiezza:2014ala] and FCNCs [@Chakrabortty:2012pp; @Bambhaniya:2013wza; @Maiezza:2016bzp; @Bertolini:2014sua; @Maiezza:2014ala] require $m_{\rm FCNH} > 10-20$ TeV at 90% CL. Subsequently, in the absence of FCNC-suppressing mechanisms, $\rho_{i} > 1$ for LHC-scale $W_R$. Thus, discovering LRSM at the LHC may suggest a strongly coupled scalar sector. Conversely, for $\rho_i<1$ and $m_{\rm FCNH}\sim15~(20)\tev$, one finds $M_{W_R}\gtrsim10~(12)\tev$, scales that are within the reach of future hadron colliders [@Arkani-Hamed:2015vfh; @Golling:2016gvc; @Mitra:2016kov]. For more detailed discussions on the perturbativity and stability of the LRSM scalar section, see Refs. [@Bertolini:2009qj; @Bertolini:2009es; @Bertolini:2012im; @Maiezza:2016bzp; @Mitra:2016kov; @Mohapatra:2014qva; @Maiezza:2016ybz] and references therein. After $\Delta_R$ acquires a vev and LR symmetry is broken spontaneously, the neutral component of ${{\rm SU}(2)_{R}}$, , $W_R^{3}$, and the ${{\rm U}(1)_{B-L}}$ boson, , $X_{B-L}$, mix into the massive eigenstate $Z'_{\rm LRSM}$ (sometimes labeled $Z_R$) and the orthogonal, massless vector boson $B$. $B$ is recognized as the gauge field associated with weak hypercharge in the SM, the generators of which are built from the remnants of ${{\rm SU}(2)_{R}}$ and ${{\rm U}(1)_{B-L}}$. The relation between electric charge $Q$, weak left/right isospin $T_{L/R}^3$, baryon minus lepton number $B$-$L$, and weak hypercharge $Y$ is given by $$Q = T_{L}^3 + T_{R}^3 + \frac{(B-L)}{2} \equiv T_{L}^3 + \frac{Y}{2}, \quad\text{with}\quad Y = 2T_{R}^3 + (B-L).$$ This in turn implies that the remaining components of ${{\rm SU}(2)_{R}}$, $W_R^{1}$ and $W_R^{2}$, combine into the state $W_R^\pm$ with electric charge $Q^{W_R}=\pm1$ and mass $M_{W_R}=g_R v_R/\sqrt{2}$. After EWSB, it is possible for the massive $W_R$ and $W_L$ gauge fields to mix, with the mixing angle $\xi_{\rm LR}$ given by $\tan2\xi_{\rm LR} = 2k_1 k_2/(v_R^2 - v_L^2)\lesssim 2 v_{\rm SM}^2/v_R^2$. Neutral meson mass splittings [@Beall:1981ze; @Langacker:1989xa; @Bertolini:2014sua; @Maiezza:2010ic; @Bernard:2015boz; @Buras:2013ooa] coupled with improved lattice calculations, e.g. [@Mescia:2012fg; @Garron:2016mva], Weak CPV [@Maiezza:2010ic; @Cirigliano:2016yhc; @Buras:2013ooa], EDMs [@Maiezza:2010ic; @Zhang:2007da; @Zhang:2007fn; @Buras:2013ooa], and CP violation in the electron EDM [@Nemevsek:2012iq], are particularly sensitive to this mixing, implying the competitive bound of $M_{W_R} \gtrsim 3$ TeV at 95% CL [@Bertolini:2014sua]. This forces $W_L-W_R$ mixing to be, $\tan2\xi_{\rm LR}/2 \approx \xi_{\rm LR} \lesssim M_W^2/M_{W_R}^2 < 7-7.5\times10^{-4}.$ A similar conclusion can be reached on $Z- Z'_{LRSM}$ mixing. Subsequently, the light and heavy mass eigenstates of LRSM gauge bosons, $W_1^\pm,~W_2^\pm,~Z_1,~Z_2$, where $M_{V_1}<M_{V_2}$, are closely aligned with their gauge states. In other words, to a very good approximation, $W_1\approx W_{\rm SM}$, $Z_1\approx Z_{\rm SM}$, $W_2\approx W_R$ and $Z'\approx Z'_{LRSM}$ (or sometimes $Z'\approx Z_R$). The mass relation between the LR gauge bosons is $M_{Z_R} = \sqrt{2\cos^2\theta_W / \cos2\theta_W}M_{W_R}\approx (1.7)\times M_{W_R}$, and implies that bounds on one mass results in indirect bounds on the second mass; see, for example, Ref. [@Lindner:2016lpp]. ### Heavy Neutrino Effective Field Theory {#sec:neft} It is possible that the coupling of TeV-scale Majorana neutrinos to the SM sector is dominated by new states with masses that are hierarchically larger than the heavy neutrino mass or the reach of present-day collider experiments. For example: Scalar SU$(2)_R$ triplets in the Left-Right Symmetric Model may acquire vevs $\mathcal{O}(10)$ TeV, resulting in new gauge bosons that are kinematically accessible at the LHC but, due to $\mathcal{O}(10^{-3}-10^{-2})$ triplet Yukawa couplings, give rise to EW-scale RH Majorana neutrino masses. In such a pathological but realistic scenario, the LHC phenomenology appears as a canonical Type I Seesaw mechanism despite originating from a different Seesaw mechanism [@Ruiz:2017nip]. While it is generally accepted that such mimicry can occur among Seesaws, few explicit examples exist in the literature and further investigation is encouraged. For such situations, it is possible to parameterize the effects of super-heavy degrees of freedom using the Heavy Neutrino Effective Field Theory (NEFT) framework [@delAguila:2008ir]. NEFT is an extension of the usual SM Effective Field Theory (SMEFT) [@Burges:1983zg; @Leung:1984ni; @Buchmuller:1985jz; @Grzadkowski:2010es], whereby instead of augmenting the SM Lagrangian with higher dimension operators one starts from the Type I Seesaw Lagrangian in Eq. (\[eq:type1Lag\]) and builds operators using that field content. Including all SU$(3)$ $\otimes$ SU$(2)_L$ $\otimes$ U$(1)_Y$-invariant, operators of mass dimension $d>4$, the NEFT Lagrangian before EWSB is given by $$\begin{aligned} \mathcal{L}_{\rm NEFT} &=& \mathcal{L}_{\rm Type~I} + \sum_{d=5}\sum_{i} \frac{\alpha_i^{(d)}}{\Lambda^{(d-4)}}\mathcal{O}_{i}^{(d)}.\end{aligned}$$ Here, $\mathcal{O}_{i}^{(d)}$ are dimension $d$, Lorentz and gauge invariant permutations of Type I fields, and $\alpha_i^{(d)}\ll4\pi$ are the corresponding Wilson coefficients. The list of $\mathcal{O}_{i}^{(d)}$ are known explicitly for $d=5$ [@Aparici:2009fh; @Elgaard-Clausen:2017xkq], 6 [@delAguila:2008ir; @Elgaard-Clausen:2017xkq], and 7 [@Bhattacharya:2015vja; @Liao:2016qyd; @Elgaard-Clausen:2017xkq], and can be built for larger $d$ following Refs. [@Henning:2015alf; @Kobach:2016ami; @Liao:2017amb]. After EWSB, fermions should then be decomposed into their mass eigenstates via quark and lepton mixing. For example: among the $d=6$, four-fermion contact operations $\mathcal{O}_i^{(6)}$ that contribute to heavy $N$ production in hadron colliders (see Eq. (\[eq:heavyNDYCC\])) in the interaction/gauge basis are [@delAguila:2008ir] $$\begin{aligned} \mathcal{O}_V^{(6)} = \left(\overline{d}\gamma^\mu P_R u\right)\left(\overline{e}\gamma_\mu P_R N_R\right) \quad\text{and}\quad \mathcal{O}_{S3}^{(6)} = \left(\overline{Q}\gamma^\mu P_R N_R\right)\varepsilon\left(\overline{L}\gamma_\mu P_R d\right).\end{aligned}$$ In terms of light $(\nu_m)$ and heavy $(N_{m'})$ mass eigenstates and using Eq. (\[eq:nuMixDefs\]), one can generically [@Atre:2009rg; @Han:2012vk] decompose the heavy neutrino interaction state $N_\ell$ as $N_{\ell} = \sum_{m=1}^{3} X_{\ell m}\nu_{m}^c + \sum_{m'=1}^{n}Y_{\ell N_{m'}} N_{m'},$ with $\vert Y_{\ell N_{m'}} \vert$ of order the elements of $U_{PMNS}$. Inserting this into the preceding operators gives quantities in terms of leptonic mass eigenstates: $$\begin{aligned} \mathcal{O}_V^{(6)} &=& \sum_{m=1}^3\left(\overline{d}\gamma^\mu P_R u\right)\left(\overline{\ell}\gamma_\mu P_R~X_{\ell m}~\nu_{m}^c\right) ~+~ \sum_{m'=1}\left(\overline{d}\gamma^\mu P_R u\right)\left(\overline{\ell}\gamma_\mu P_R~Y_{\ell N_{m'}}~N_{m'}\right), \quad\text{and}\quad\nonumber\\ \mathcal{O}_{S3}^{(6)} &=& \sum_{m=1}^3\left(\overline{Q}\gamma^\mu P_R ~X_{\ell m} \nu_{m}^c \right)\left(\overline{\ell}\gamma_\mu P_R d\right) ~+~ \sum_{m'=1}\left(\overline{Q}\gamma^\mu P_R ~Y_{\ell N_{m'}} N_{m'} \right)\left(\overline{\ell}\gamma_\mu P_R d\right).\end{aligned}$$ After EWSB, a similar decomposition for quarks gauge states in terms of CKM matrix elements and mass eigenstates should be applied. For more information on such decompositions, see, , [@Ruiz:2017nip] and references therein. It should be noted that after integrating out the heavy $N$ field, the marginal operators at $d>5$ generated from the Type I Lagrangian are not the same operators generated by integrating the analogous Seesaw partner in the Type II and III scenarios [@Abada:2007ux; @delAguila:2012nu]. Heavy Neutrinos at Colliders {#sec:typeIhybrid} ---------------------------- The connection between low-scale Seesaw models and colliders is made no clearer than in searches for heavy neutrinos, both Majorana and (pseudo-)Dirac, in the context of Type I-based scenarios. While extensive, the topic’s body of literature is still progressing in several directions. This is particularly true for the development of collider signatures, Monte Carlo tools, and high-order perturbative corrections. Together, these advancements greatly improve sensitivity to neutrinos and their mixing structures at collider experiments. We now review the various searches for $L$-violating collider processes facilitated by Majorana neutrinos $N$. We start with low-mass (Sec. \[sec:type1LowMass\]) and high-mass (Secs. \[sec:type1HighMassPP\] and \[sec:type1HighMassEX\]) neutrinos in the context of Type I-based hybrid scenarios, before moving onto Abelian (Sec. \[sec:type1Abelian\]) and non-Abelian (Sec. \[sec:lrsmCollider\]) gauge extensions, and finally the semi-model independent NEFT framework (Sec. \[sec:neftTests\]). Lepton number violating collider processes involving pseudo-Dirac neutrinos are, by construction, suppressed [@Wolfenstein:1981kw; @Petcov:1982ya; @Leung:1983ti; @Valle:1983dk; @Moffat:2017feq]. Thus, a discussion of their phenomenology is outside the scope of this review and we refer readers to thorough reviews such as Refs. [@Ibarra:2011xn; @Weiland:2013wha; @Antusch:2016ejd]. ### Low-Mass Heavy Neutrinos at $pp$ and $ee$ Colliders {#sec:type1LowMass} For Majorana neutrinos below the $M_W$ mass scale, lepton number violating processes may manifest in numerous ways, including rare decays of mesons, baryons, $\mu$ and $\tau$ leptons, and even SM electroweak bosons. Specifically, one may discover $L$ violation in three-body meson decays to lighter mesons $M_1^\pm \to M_2^\mp \ell_1^\pm \ell_2^\pm$ [@Littenberg:1991ek; @Littenberg:2000fg; @Dib:2000wm; @Atre:2005eb; @Atre:2009rg; @Cvetic:2010rw; @Dib:2014iga; @Cvetic:2014nla; @Cvetic:2015ura; @Cvetic:2015naa; @Cvetic:2016fbv; @Dib:2014pga; @Quintero:2016iwi; @Milanes:2016rzr; @Wang:2014lda; @Dong:2013raa; @Asaka:2016rwd], such as that shown in Fig. \[fig:lnvBDecayDiagram\]; four-body meson decays to lighter mesons $M_1^\pm \to M_2^\mp M_3^0 \ell_1^\pm \ell_2^\pm $  [@Quintero:2016iwi; @Milanes:2016rzr; @Castro:2013jsn; @Yuan:2013yba; @Cvetic:2017vwl]; four-body meson decays to leptons $M^\pm \to \ell_1^\pm \ell_1^\pm \ell_2^\mp \nu$  [@Cvetic:2015naa; @Cvetic:2016fbv; @Cvetic:2017vwl; @Cvetic:2012hd; @Cvetic:2013eza]; five-body meson decays [@Cvetic:2017vwl]; four-body baryon decays to mesons, $B \to M \ell_1^\pm \ell_2^\pm$ [@Mejia-Guisao:2017nzx]; three-body $\tau$ decay to mesons, $\tau^\pm \to \ell^\mp M_1^\pm M_2^\pm$ [@Quintero:2016iwi; @Kobach:2014hea; @Zamora-Saa:2016ito]; four-body $\tau$ decays to mesons, $\tau^\pm \to \ell_1^\pm \ell_1^\pm M^\mp \nu$ [@Quintero:2016iwi; @Kobach:2014hea; @Yuan:2017xdp; @Castro:2012gi; @Mandal:2016hpr]; four-body $W$ boson decays, $W^\pm \to \ell_1^\pm \ell_1^\pm \ell_2^\mp \nu$ [@BarShalom:2006bv; @Dib:2017iva; @Dib:2017vux; @Dib:2016wge; @Dib:2015oka]; Higgs boson decays, $h \to NN \to \ell_1^\pm \ell_2^\pm + X$ [@BhupalDev:2012zg; @Gago:2015vma; @Caputo:2017pit; @Das:2017zjc]. and even top quark decays, $t \to b W^{+*} \to b \ell_1^+ N \to b \ell_1^+ \ell_2^\pm q \overline{q'}$ [@Weinberg:1979sa; @BarShalom:2006bv; @Si:2008jd; @Quintero:2011yh]. The $W$ boson case is notable as azimuthal and polar distributions [@Han:2012vk] and endpoint kinematics [@Dib:2016wge] can differentiate between $L$ conservation and non-conservation. Of the various collider searches for GeV-scale $N$, great complementarity is afforded by $B$-factories. As shown in \[fig:lnvBDecayLimits\], an analysis of Belle I [@Liventsev:2013zz] and LHCb Run I [@Aaij:2012zr; @Aaij:2014aba] searches for $L$-violating final states from meson decays excluded [@Shuve:2016muy] $\vert V_{\mu N}\vert^2\gtrsim3\times10^{-5}$ for $M_N = 1-5$ GeV. Along these same lines, the observability of displaced decays of heavy neutrinos [@Helo:2013esa; @Blondel:2014bra; @Izaguirre:2015pga; @Gago:2015vma; @Antusch:2017hhu; @Antusch:2016vyf] and so-called “neutrino-antineutrino oscillations” [@Anamiati:2016uxp; @Antusch:2017ebe; @Das:2017hmg; @Antusch:2017pkq] (in analogy to $\mathcal{B}-\overline{\mathcal{B}}$ oscillations) and have also been discussed. Indirectly, the presence of heavy Majorana neutrinos can appear in precision EW measurements as deviations from lepton flavor unitarity and universality, and is ideally suited for $e^+e^-$ colliders [@delAguila:2008pw; @Gomez-Ceballos:2013zzn; @Antusch:2014woa; @deGouvea:2015euy; @Antusch:2015mia; @Fernandez-Martinez:2016lgt; @Antusch:2016ejd], such as the International Linear Collider (ILC)  [@Baer:2013cma; @Fujii:2015jha], Circular $e^-e^+$ Collider (CepC) [@CEPC-SPPCStudyGroup:2015csa], and Future Circular Collider-$ee$ (FCC-ee) [@Gomez-Ceballos:2013zzn]. An especially famous example of this is the number of active, light neutrino flavors $N_\nu$, which can be inferred from the $Z$ boson’s invisible width $\Gamma^Z_{\rm Inv}$. At lepton colliders, $\Gamma^Z_{\rm Inv}$ can be determined in two different ways: The first is from line-shape measurements of the $Z$ resonance as a function of $\sqrt{s}$, and is measured to be $N_\nu^{\rm Line}=2.9840\pm0.0082$ [@ALEPH:2005ab]. The second is from searches for invisible $Z$ decays, , $e^+e^- \to Z\gamma$, and is found to be $N_\nu^{\rm Inv}=2.92\pm0.05$ [@Beringer:1900zz]. Provocatively, both measurements deviate from the SM prediction of $N_\nu^{\rm SM}=3$ at the $2\sigma$ level. It is unclear if deviations from $N_\nu^{\rm SM}$ are the result of experimental uncertainty or indicate the presence of, for example, heavy sterile neutrinos [@Jarlskog:1990kt; @Blondel:2014bra]. Nonetheless, a future $Z$-pole machine can potentially clarify this discrepancy [@Blondel:2014bra]. For investigations into EW constraints on heavy neutrinos, see Refs. [@delAguila:2008pw; @Antusch:2014woa; @deGouvea:2015euy; @Fernandez-Martinez:2016lgt]. ### High-Mass Heavy Neutrinos at $pp$ Colliders {#sec:type1HighMassPP} Collider searches for heavy Majorana neutrinos with masses above $M_W$ have long been of interest to the community [@Keung:1983uu; @Gronau:1984ct; @Willenbrock:1985tj; @Petcov:1984nf], with exceptionally notable works appearing in the early 1990s [@Pilaftsis:1991ug; @Dicus:1991fk; @Dicus:1991wj; @Datta:1991mf; @Datta:1993nm] and late-2000s [@Kersten:2007vk; @Atre:2009rg; @Han:2006ip; @Bray:2007ru; @delAguila:2006bda; @delAguila:2007qnc; @delAguila:2008cj; @delAguila:2009bb]. In the past decade, among the biggest advancements in Seesaw phenomenology is the treatment of collider signatures for such hefty $N$ in Type I-based models. While coupled to concurrent developments in Monte Carlo simulation packages, the progression has been driven by attempts to reconcile conflicting reports of heavy neutrino production cross sections for the LHC. This was at last resolved in Refs. [@Alva:2014gxa; @Degrande:2016aje], wherein new, infrared- and collinear- (IRC-)safe definitions for inclusive and semi-inclusive[^4] production channels were introduced. The significance of such collider signatures is that they are well-defined at all orders in $\alpha_s$, and hence correspond to physical observables. We now summarize this extensive body of literature, emphasizing recent results. ![Born diagrams for heavy neutrino $(N)$ production via (a) Drell-Yan, (b) gluon fusion, and (c) electroweak vector boson fusion; from Ref. [@Ruiz:2017yyf] and drawn using `JaxoDraw` [@Binosi:2008ig]. []{data-label="fig:heavyNDiagrams"}](heavyNThresh_feynman_MultiProd_1706_02298.pdf){width="96.00000%"} For Majorana neutrinos with $M_N > M_W$, the most extensively studied  [@Keung:1983uu; @Gronau:1984ct; @Datta:1991mf; @Han:2006ip; @delAguila:2006bda; @delAguila:2007qnc; @Bray:2007ru; @Chao:2009ef; @delAguila:2008cj; @delAguila:2009bb; @Atre:2009rg; @Antusch:2016ejd; @Das:2017hmg; @Chen:2011hc] collider production mechanism is the $L$-violating, charged current (CC) Drell-Yan (DY) process [@Keung:1983uu], shown in Fig. \[fig:heavyNDiagrams\](a), and given by $$q_1 ~\overline{q}_2 \rightarrow W^{\pm *} \rightarrow N ~\ell^\pm_1, ~\quad\text{with}\quad N \to \ell_2^\pm W^{\mp } \to \ell_2^\pm q_1' ~\overline{q'}_2. \label{eq:heavyNDYCC}$$ A comparison of Fig. \[fig:heavyNDiagrams\](a) to the meson decay diagram of Fig. \[fig:lnvBDecayDiagram\] immediately reveals that Eq. (\[eq:heavyNDYCC\]) is the former’s high momentum transfer completion. Subsequently, much of the aforementioned kinematical properties related to $L$-violating meson decays also hold for the CC DY channel [@Han:2012vk; @Dev:2015kca]. Among the earliest studies are those likewise focusing on neutral current (NC) DY production [@Gronau:1984ct; @Willenbrock:1985tj; @Dicus:1991wj; @Datta:1991mf; @Datta:1993nm], again shown in Fig. \[fig:heavyNDiagrams\](a), and given by $$\begin{aligned} q ~\overline{q} \rightarrow Z^* \rightarrow N ~\overset{(-)}{\nu_\ell}, \label{eq:heavyNDYNC}\end{aligned}$$ as well as the gluon fusion mechanism [@Willenbrock:1985tj; @Dicus:1991wj], shown in Fig. \[fig:heavyNDiagrams\](b), and given by $$g ~g \rightarrow Z^*/h^* \rightarrow N ~ \overset{(-)}{\nu_\ell}. \label{eq:heavyNGF}$$ Interestingly, despite gluon fusion being formally an $\mathcal{O}(\alpha_s^2)$ correction to Eq. (\[eq:heavyNDYNC\]), it is non-interfering, separately gauge invariant, and the subject of renewed interest [@Hessler:2014ssa; @Degrande:2016aje; @Ruiz:2017yyf]. Moreover, in accordance to the Goldstone Equivalence Theorem [@Chanowitz:1985hj; @Lee:1977yc], the $ggZ^*$ contribution has been shown [@Hessler:2014ssa; @Ruiz:2017yyf] to be as large as the $ggh^*$ contribution, and therefore should not be neglected. Pair production of $N$ via $s$-channel scattering [@Willenbrock:1985tj; @Datta:1991mf], , $gg \to N N$, or weak boson scattering [@Dicus:1991fk; @Datta:1993nm; @Han:2006ip], , $W^\pm W^\mp \to NN$, have also been discussed, but are relatively suppressed compared to single production by an additional mixing factor of $\vert V_{\ell N_{m'}}\vert^2 \lesssim 10^{-4}$. A recent, noteworthy development is the interest in semi-inclusive and exclusive production of heavy neutrinos at hadron colliders, , $N$ production in association with jets. In particular, several studies have investigated the semi-inclusive, photon-initiated vector boson fusion (VBF) process [@Datta:1993nm; @Dev:2013wba; @Alva:2014gxa; @Degrande:2016aje], shown in Fig. \[fig:heavyNDiagrams\](c), and given by $$q ~\gamma \rightarrow N ~\ell^\pm ~q', \label{eq:heavyNvbfIncl}$$ and its deeply inelastic, $\mathcal{O}(\alpha)$ radiative correction [@Datta:1993nm; @Dev:2013wba; @Bambhaniya:2014hla; @Alva:2014gxa; @Ng:2015hba; @Arganda:2015ija; @Degrande:2016aje; @Andres:2017daw], $$q_1 ~q_2 \xrightarrow{W\gamma+WZ \to N\ell^\pm} N ~\ell^\pm ~q'_1 ~q'_2. \label{eq:heavyNvbfExcl}$$ At $\mathcal{O}(\alpha^4)$ (here we do not distinguish between $\alpha$ and $\alpha_W$), the full, gauge invariant set of diagrams, which includes the sub-leading $W^\pm Z\to N\ell^\pm$ scattering, is given in Fig. \[fig:heavyNVBF2\]. ![Born diagrams for the $\mathcal{O}(\alpha^4)$ heavy neutrino $(N)$ production process $q_1q_2\to N\ell^\pm q_1' q_2'$ [@Alva:2014gxa].[]{data-label="fig:heavyNVBF2"}](feynman_WA_HeavyN_DIS_1411_7305.pdf){width="96.00000%"} Treatment of the VBF channel is somewhat subtle in that it receives contributions from collinear QED radiation off the proton [@Dev:2013wba], collinear QED radiation off initial-states quarks [@Alva:2014gxa], and QED radiation in the deeply inelastic/high momentum transfer limit [@Datta:1993nm]. For example: In the top line of diagrams in Fig. \[fig:heavyNVBF2\], one sees that in the collinear limit of the $q_2 \to \gamma^* q_2'$ splitting, the virtual $\gamma^*$ goes on-shell and the splitting factorizes into a photon parton distribution function (PDF), recovering the process in Eq. (\[eq:heavyNvbfIncl\]) [@Alva:2014gxa; @Degrande:2016aje]. As these sub-channels are different kinematic limits of the same process, care is needed when combining channels so as to not double count regions of phase space. While ingredients to the VBF channel have been known for some time, consistent schemes to combine/match the processes are more recent [@Alva:2014gxa; @Degrande:2016aje]. Moreover, for inclusive studies, Ref. [@Degrande:2016aje] showed that the use of Eq. (\[eq:heavyNvbfIncl\]) in conjunction with a $\gamma$-PDF containing both elastic and inelastic contributions [@Martin:2014nqa] can reproduce the fully matched calculation of Ref. [@Alva:2014gxa] within the $\mathcal{O}(20\%)$ uncertainty resulting from missing NLO in QED terms. Neglecting the collinear $q_2 \to \gamma^* q_2'$ splitting accounts for the unphysical cross sections reported in Refs. [@Dev:2013wba; @Deppisch:2015qwa]. Presently, recommended PDF sets containing such $\gamma$-PDFs include: MMHT QED (no available `lhaid`)  [@Martin:2014nqa; @Harland-Lang:2016kog], NNPDF 3.1+LUXqed (`lhaid=324900`)  [@Bertone:2017bme], LUXqed17+PDF4LHC15 (`lhaid=82200`)  [@Manohar:2017eqh; @Manohar:2016nzj], and CT14 QED Inclusive (`lhaid = 13300`)  [@Schmidt:2015zda]. Qualitatively, the MMHT [@Martin:2014nqa] and LUXqed [@Manohar:2017eqh; @Manohar:2016nzj] treatments of photon PDFs are the most rigorous. In analogy to the gluon fusion and NC DY, Eq. (\[eq:heavyNvbfIncl\]) (and hence Eq. (\[eq:heavyNvbfExcl\])) is a non-interfering, $\mathcal{O}(\alpha)$ correction to the CC DY process. Thus, the CC DY and VBF channels can be summed coherently. In addition to these channels, the semi-inclusive, associated $n$-jet production mode, $$p ~p ~\to W^* ~+~ \geq nj ~+~ X \quad\to\quad N ~\ell^\pm ~+~ \geq n j ~+~ X, \quad\text{for}\quad n\in \mathbb{N}, \label{eq:heavyNnj}$$ has also appeared in the recent literature [@Dev:2013wba; @Das:2014jxa; @Degrande:2016aje]. As with VBF, much care is needed to correctly model Eq. (\[eq:heavyNnj\]). As reported in Refs. [@Ruiz:2015zca; @Degrande:2016aje], the production of heavy leptons in association with QCD jets is nuanced due to the presence of additional $t$-channel propagators that can lead to artificially large cross sections if matrix element poles are not sufficiently regulated. (It is not enough to simply remove the divergences with phase space cuts.) After phase space integration, these propagators give rise to logarithmic dependence on the various process scales. Generically [@Collins:1984kg; @Ruiz:2015zca], the cross section for heavy lepton and jets in Eq. (\[eq:heavyNnj\]) scales as: $$\sigma(pp\rightarrow N\ell^\pm+nj +X) \sim \sum^n_{k=1} \alpha_s^k(Q^2)\log^{(2k-1)}\left(\frac{Q^2}{q_T^2}\right). \label{eq:heavyNnjetXSec}$$ Here, $Q\sim M_N$ is the scale of the hard scattering process, $q_T = \sqrt{\vert \vec{q}_T \vert^2}$, and $\vec{q}_T \equiv \sum_k^n \vec{p}_{T,k}^{~j}$, is the $(N\ell)$-system’s transverse momentum, which recoils against the vector sum of all jet $\vec{p}_T$. It is clear for a fixed $M_N$ that too low jet $p_T$ cuts can lead to too small $q_T$ and cause numerically large (collinear) logarithms such that $\log(M_N^2/q_T^2) \gg 1/\alpha_s(M_N)$, spoiling the perturbative convergence of Eq. (\[eq:heavyNnjetXSec\]). Similarly, for a fixed $q_T$, arbitrarily large $M_N$ can again spoil perturbative convergence. As noted in Refs. [@Alva:2014gxa; @Degrande:2016aje], neglecting this fact has led to conflicting predictions in several studies on heavy neutrino production in $pp$ collisions. It is possible [@Degrande:2016aje], however, to tune $p_T$ cuts on jets with varying $M_N$ to enforce the validity of Eq. (\[eq:heavyNnjetXSec\]). Within the Collins-Soper-Sterman (CSS) resummation formalism [@Collins:1984kg], Eq. (\[eq:heavyNnjetXSec\]) is trustworthy when $\alpha_s(Q^2)$ is perturbative and $q_T\sim Q$, , $$\log (Q/\Lambda_{\rm QCD}) \gg1 \quad\text{and}\quad \alpha_s(Q)\log^2 (Q^2/q_T^2) \lesssim 1.$$ Noting that at 1-loop $\alpha_s(Q)$ can be written as $1/\alpha_s(Q) \approx (\beta_0/2\pi)\log (Q/\Lambda_{\rm QCD})$, and setting $Q=M_N$, one can invert the second CSS condition and obtain a consistency relationship [@Degrande:2016aje]: $$q_T = \vert \vec{q}_T \vert = \left\vert \sum_{k=1}^n \vec{p}_{T,k}^{~j} \right\vert \gtrsim M_N \times e^{-(1/2)\sqrt{(\beta_0/2\pi)\log(M_N/\Lambda_{\rm QCD})}}. \label{eq:cssConsistency}$$ This stipulates a minimum $q_T$ needed for semi-inclusive processes like Eq. (\[eq:heavyNnjetXSec\]) to be valid in perturbation theory. When $q_T$ of the $(N\ell)$-system is dominated by a single, hard radiation, Eq. (\[eq:cssConsistency\]) is consequential: In this approximation, $q_T \approx \vert \vec{p}_{T,1}^{~j} \vert$ and Eq. (\[eq:cssConsistency\]) suggests a *scale-dependent*, minimum jet $p_T$ cut to ensure that specifically the semi-inclusive $pp\to N\ell +\geq 1j+X$ cross section is well-defined in perturbation theory. Numerically, this is sizable: for $M_N = 30~(300)~[3000]~{\rm GeV}$, one requires that $\vert \vec{p}_{T,1}^{~j} \vert \gtrsim 9~(65)~[540]$ GeV, or alternatively $\vert \vec{p}_{T,1}^{~j} \vert \gtrsim 0.3~(0.22)~[0.18]\times M_N$, and indicates that naïve application of fiducial $p_T^j$ cuts for the LHC do not readily apply for $\sqrt{s} = $ 27-100 TeV scenarios, where one can probe much larger $M_N$. The perturbative stability of this approach is demonstrated by the (roughly) flat $K$-factor of $K^{\rm NLO}\approx1.2$ for the semi-inclusive $pp\to N\ell^\pm+1j$ process, shown in the lower panel of Fig. \[fig:heavyNXSecVsM\]. Hence, the artificially large $N$ production cross sections reported in Refs. [@Dev:2013wba; @Das:2014jxa; @Deppisch:2015qwa] can be attributed to a loss of perturbative control over their calculation, not the presence of an enhancement mechanism. Upon the appropriate replacement of $M_N$, Eq. (\[eq:cssConsistency\]) holds for other color-singlet processes [@Degrande:2016aje], including mono-jet searches, and is consistent with explicit $p_T$ resummations of high-mass lepton [@Ruiz:2015zca] and slepton [@Dreiner:2006sv; @Chen:2006ep] production. A characteristic of heavy neutrino production cross sections is that the active-sterile mixing, $\vert V_{\ell N}\vert$, factorizes out of the partonic and hadronic scattering expressions. Exploiting this one can define [@Han:2006ip] a “bare” cross section $\sigma_{0}$, given by $$\sigma_{0}(pp\rightarrow N+X) ~\equiv~ \sigma(pp\rightarrow N+X) / \vert V_{\ell N}\vert^2.$$ Assuming resonant production of $N$, a similar expression can be extracted at the $N$ decay level, $$\sigma_{0}(pp \to \ell^\pm_1 \ell^\pm_2 + X) ~\equiv~ \sigma(pp \to \ell^\pm_1 \ell^\pm_2 + X) / S_{\ell_1 \ell_2}, \quad S_{\ell_1 \ell_2} = \cfrac{\vert V_{\ell_1 N}\vert^2\vert V_{\ell_2 N}\vert^2 }{\sum_{\ell=e}^\tau \vert V_{\ell N}\vert^2}.$$ These definitions, which hold at higher orders in $\alpha_s$ [@Ruiz:2015zca; @Degrande:2016aje], allow one to make cross section predictions and comparisons independent of a particular flavor model, including those that largely conserve lepton number, such as the inverse and linear Seesaws. It also allows for a straightforward reinterpretation of limits on collider cross sections as limits on $S_{\ell_1 \ell_2}$, or $\vert V_{\ell N}\vert$ with additional but generic assumptions. An exception to this factorizablity is the case of nearly degenerate neutrinos with total widths that are comparable to their mass splitting [@Bray:2007ru; @Fuchs:2014ola; @Fuchs:2016swt; @Anamiati:2016uxp]. \ Figure \[fig:heavyNXSec\] shows a comparison of the leading, single $N$ hadronic production cross sections, divided by active-heavy mixing $\vert V_{\ell N}\vert^2$, as a function of (a) heavy neutrino mass $M_N$ at $\sqrt{s}=14$ [@Degrande:2016aje] and (b) collider energy $\sqrt{s}$ up to 100 TeV for $M_N = 500,~1000$ GeV [@Ruiz:2017yyf]. The various accuracies reported reflect the maturity of modern Seesaw calculations. Presently, state-of-the-art predictions for single $N$ production modes are automated up to NLO+PS in QCD for the Drell-Yan and VBF channels [@Ruiz:2015gsa; @Degrande:2016aje], amongst others, and known up to N$^3$LL(threshold) for the gluon fusion channel [@Ruiz:2017yyf]. With Monte Carlo packages, predictions are available at LO with multi-leg merging (MLM) [@delAguila:2007qnc; @Alpgen:HeavyN; @Degrande:2016aje; @FeynRules:HeavyNnlo] as well as up to NLO with parton shower matching and merging [@Degrande:2016aje; @FeynRules:HeavyNnlo]. The NLO in QCD-accurate [@Degrande:2014vpa], `HeavyNnlo` universal FeynRules object (UFO) [@Degrande:2011ua] model file is available from Refs. [@Degrande:2016aje; @FeynRules:HeavyNnlo]. Model files built using `FeynRules` [@Christensen:2008py; @Degrande:2011ua; @Alloul:2013bka] construct and evaluate $L$-violating currents following the Feynman rules convention of Ref. [@Denner:1992me]. A brief comment is needed regarding choosing MLM+PS or NLO+PS computations: To produce MLM Monte Carlo samples, one must sum semi-inclusive channels with successively higher leg multiplicities in accordance with Eqs. (\[eq:heavyNnjetXSec\])-(\[eq:cssConsistency\]) and correct for phase space double-counting. However, such MLM samples are formally LO in $\mathcal{O}(\alpha_s)$ because of missing virtual corrections. NLO+PS is formally more accurate, under better perturbative control (due to explicit cancellation of infrared singularities), and thus is recommended for modeling heavy $N$ at colliders. Such computations are possible with modern, general-purpose event generators, such as Herwig [@Bellm:2015jjp], MadGraph5\_aMC@NLO [@Alwall:2014hca], and Sherpa [@Gleisberg:2008ta]. At the 13 and 14 TeV LHC, heavy $N$ production is dominated by charged-current mechanisms for phenomenologically relevant mass scales, , $M_N \lesssim 700$ GeV [@Alva:2014gxa]. At more energetic colliders, however, the growth in the gluon-gluon luminosity increases the $gg\to N\nu$ cross section faster than the CC DY channel. In particular, at $\sqrt{s} = 20-30$ TeV, neutral-current mechanisms surpass charged-current modes for heavy $N$ production with $M_N = 500-1000$ GeV [@Ruiz:2017yyf]. As seen in the sub-panel of Fig. \[fig:heavyNXSecVsM\], NLO in QCD contributions only modify inclusive, DY-type cross section normalizations by $+20$-to-$+30\%$ and VBF negligibly, indicating that the prescriptions of Ref. [@Degrande:2016aje] are sufficient to ensure perturbative control over a wide-range of scales. One should emphasize that while VBF normalizations do not appreciably change under QCD corrections [@Han:1992hr], VBF kinematics do change considerably [@Degrande:2016aje; @Degrande:2015xnm; @Cacciari:2015jma; @Dreyer:2016oyx]. The numerical impact, however, is observable-dependent and can be large if new kinematic channels are opened at higher orders of $\alpha_s$. In comparison to this, the sub-panel of Fig. \[fig:heavyNXSecVsS\] shows that QCD corrections to gluon fusion are huge ($+150$-to-$+200\%$), but convergent and consistent with SM Higgs, heavy Higgs, and heavy pseudoscalar production [@Bonvini:2014qga; @Anastasiou:2016cez; @Anastasiou:2015ema]; for additional details, see Ref. [@Ruiz:2017yyf]. With these computational advancements, considerable collider sensitivity to $L$-violating processes in the Type I Seesaw has been reached. In Fig. \[100TeVdiscovery.fig\] is the expected sensitivity to active-sterile neutrino mixing via the combined CC DY+VBF channels and in same-sign $\mu^\pm\mu^\pm+X$ final-state. With $\mathcal{L}=1$ ab$^{-1}$ of data for $M_N > M_W$ at $\sqrt{s} = $ 14 (100) TeV, one can exclude at $2\sigma$ $S_{\mu\mu} \approx \vert V_{\mu N}\vert^2 \gtrsim 10^{-4}~(10^{-5})$ [@Alva:2014gxa]. This is assuming the 2013 Snowmass benchmark detector configuration for $\sqrt{s}=100$ TeV [@Avetisyan:2013onh]. Sensitivity to the $e^\pm e^\pm$ and $e^\pm\mu^\pm$ channels is comparable, up to detector (in)efficiencies for electrons and muons. As shown in Fig. \[fig:lhcTypeILimits\], with $\mathcal{L}\approx20$ fb$^{-1}$ at 8 TeV, the ATLAS and CMS experiments have excluded at 95% CLs $\vert V_{\ell N}\vert^2 \gtrsim 10^{-3} - 10^{-1}$ for $M_N = 100 - 450$ GeV [@Sirunyan:2017yrk; @Khachatryan:2016jqo; @Khachatryan:2016olu; @Khachatryan:2015gha; @Aad:2015xaa]. For heavier $M_N$, quarks from the on-shell $W$ boson decay can form a single jet instead of the usual two-jet configuration. In such cases, well-known “fat jet” techniques can be used [@ATLAS:2012ak; @Cox:2017eme]. Upon discovery of $L$-violating processes involving heavy neutrinos, among the most pressing quantities to measure are $N$’s chiral couplings to other fields [@Han:2012vk; @Dev:2015kca], its flavor structure [@Chen:2011hc; @Nemevsek:2012iq; @Anamiati:2016uxp; @Das:2017hmg], and a potential determination if the signal is actually made of multiple, nearly degenerate $N$ [@Chao:2009ef; @Antusch:2017ebe]. \ ### High-Mass Heavy Neutrinos at $ep$ Colliders {#sec:type1HighMassEX} Complementary to searches for $L$ violation in $pp$ collisions are the prospects for heavy $N$ production at $ep$ deeply inelastic scattering (DIS) colliders [@Buchmuller:1990vh; @Buchmuller:1991tu; @Blaksley:2011ey; @Duarte:2014zea; @Banerjee:2015gca; @Mondal:2016kof; @Antusch:2016ejd; @Ingelman:1993ve; @Liang:2010gm], such as proposed Large Hadron-electron Collider (LHeC)  [@AbelleiraFernandez:2012cc], or a $\mu p$ analogue [@Blaksley:2011ey]. As shown in Fig. \[fig:heavyNDISproduction\], DIS production of Majorana neutrinos can occur in multiple ways, including (a) $W$ exchange and (b) $W\gamma$ fusion. For treatment of initial-state photons from electron beams, see Ref. [@Frixione:1993yw]. Search strategies for Majorana neutrinos at DIS experiments typically rely on production via the former since $e\gamma\to NW$ associated production can suffer from large phase space suppression, especially at lower beam energies. On the other hand, at higher beam energies, the latter process can provide additional polarization information on $N$ and its decays [@Antusch:2016ejd]. At DIS facilities, one usually searches for $L$ violation by requiring that $N$ decays to a charged lepton of opposite sign from the original beam configuration, , $$\ell_1^\pm ~q_i ~\to ~N ~q_f, \quad\text{with}\quad N \to \ell_2^\mp W^\pm \to ~\ell_2^\mp ~q ~\overline{q'}, \label{eq:heavyNdis}$$ which is only possible of $N$ is Majorana and is relatively free of SM backgrounds. As in the $pp$ case, the existence of a high-$p_T$ charged lepton without accompanying MET (at the partonic level) greatly reduces SM backgrounds. At the hadronic level, this translates to requiring one charged lepton and three high-$p_T$ jets: two that arise from the decay of $N$, which scale as $p_T^j \sim M_N/4$, and the third from the $W$ exchange, which scales as $p_T^j \sim M_W/2$. However, it was recently noted [@Mattelaer:2016ynf] that tagging this third jet is not necessary to reconstruct and identify the heavy neutrino, and that a more inclusive search may prove more sensitive. Although Eq. (\[eq:heavyNdis\]) represents the so-called “golden channel”, searches for $N\to Z/h+\nu$ decays, but which do not manifestly violate lepton number, have also been proposed [@Ingelman:1993ve]. While the lower beam energies translate to a lower mass reach for $M_N$, large luminosity targets and relative cleaner hadronic environment result in a better sensitivity than the LHC to smaller active-sterile mixing for smaller neutrino Majorana masses. In Fig. \[fig:heavyNDISproduction\], one sees the expected 90% CL active-sterile mixing $\vert\theta\vert^2$ (or $\vert V_{\ell N}\vert^2$) sensitivity assuming (c) $ep$ configuration with $E_e = 150$ GeV and (d) $\mu p$ configuration with $E_\mu = 2$ TeV. For $\mathcal{L}\sim \mathcal{O}(100)$ fb$^{-1}$, one can probe $\vert V_{\ell N}\vert^2 \sim 10^{-5}-10^{-3}$ for $M_N = 250-750$ GeV [@Blaksley:2011ey]. ### Heavy Neutrinos and U$(1)_{X}$ Gauge Extensions at Colliders {#sec:type1Abelian} Due to the small mixing between the heavy neutrinos and the SM leptons in minimal Type I Seesaw scenarios, typically of the order $|V_{\ell N}|^2\sim \mathcal{O}(m_\nu/M_N)$, the predicted rates for collider-scale lepton number violation is prohibitively small. With a new gauge interaction, say, from U$(1)_{B-L}$, the gauge boson $Z'=Z_{BL}$ can be produced copiously in $pp$ and $p\bar{p}$ collisions via gauge interactions in quark annihilation [@Salvioni:2009mt; @Jenkins:1987ue; @Carena:2004xs; @Emam:2007dy; @Langacker:2008yv; @Iso:2009nw; @Basso:2010pe; @Deppisch:2013cya] and at Linear Colliders in $e^+e^-$ annihilation [@Freitas:2004hq; @Basso:2009hf; @Iso:2009nw; @Ramirez-Sanchez:2016ugz], $$q \bar{q} \to Z' \to N N \quad\text{and}\quad e^+e^- \to Z' \to N N.$$ $Z_{BL}$’s subsequent decay to a pair of heavy Majorana neutrinos may lead to a large sample of events without involving the suppression from a small active-sterile mixing angles  [@delAguila:2007ua; @Huitu:2008gf; @Basso:2008iv; @AguilarSaavedra:2009ik; @Perez:2009mu; @Li:2010rb; @AguilarSaavedra:2012gf; @Accomando:2016sge; @Accomando:2017qcs]. As a function of $M_{Z_{BL}}$, Fig. \[fig:type1BLxSecNLONNLL\] shows the NLO+NLL(Thresh.) $pp\to Z_{BL}\to \ell^+\ell^-$ production and decay rate for $\sqrt{s}=13$ TeV and representative values of coupling $g_{BL}$. As a function of Majorana neutrino mass $M_{N_1}$, Fig. \[fig:type1BLxSecLO\] shows the LO $pp\to Z_{BL}\to NN$ production and decay rate for $\sqrt{s}=14$ TeV and 100 TeV and representative $M_{Z_{BL}}$. As $N$ is Majorana, the mixing-induced decays modes $N\to \ell^\pm W^\mp, \nu Z, \nu h$ open for $M_{N_1}>M_W,M_Z,M_h$, respectively. Taking these into account, followed by the leptonic and/or hadronic decays of $W$, $Z$ and $h$, the detectable signatures include the lepton number violating, same-sign dileptons, $NN\to \ell^\pm\ell^\pm W^\mp W^\mp \to \ell^\pm\ell^\pm+nj$ [@Perez:2009mu; @Cox:2017eme]; final states with three charged leptons, $\ell^\pm\ell^\pm\ell^\mp+nj+$MET [@Basso:2008iv; @Kang:2015uoc; @Accomando:2017qcs]; and four-charged lepton, $\ell^\pm\ell^\pm\ell^\mp\ell^\mp+$MET [@Huitu:2008gf; @Abdelalim:2014cxa]. Assuming that only the third generation fermions are charged under $B-L$ symmetry, the HL-LHC can probe $Z'$ mass up to 2.2 TeV and heavy neutrino mass in the range of $0.2-1.1$ TeV as shown in Fig. \[typei-sensitivity\] [@Cox:2017eme]. For super-heavy $Z_{BL}$, [*e.g.*]{}, $M_{Z_{BL}}\gtrsim5$ TeV $\gg M_{N}$, one should note that at the 13 TeV LHC, a nontrivial contribution of the total $pp\to Z_{BL} \to NN$ cross section comes from the kinematical threshold region, where the $(NN)$ system’s invariant mass is near $m_{NN}\sim 2M_N$ and $Z_{BL}^*$ is far off-shell. This implies that the $L$-violating process $pp\to NN \to \ell^\pm\ell^\pm+nj$ can still proceed despite $Z_{BL}$ being kinematically inaccessible [@Ruiz:2017nip]. For more details, see Sec. \[sec:neftTests\]. Additionally, for such heavy $Z_{BL}$ that are resonantly produced, the emergent $N$ are highly boosted with Lorentz factors of $\gamma \sim M_{Z_{BL}}/2M_N$. For $M_N\ll M_{Z_{BL}}$, this leads to highly collimated decay products, with separations scaling as $\Delta R\sim 2/\gamma \sim 4M_N/M_{Z_{BL}}$, and eventually the formation of lepton jets [@Izaguirre:2015pga; @Dube:2017jgo], , collimated clusters of light, charged leptons and electromagnetic radiation, and neutrino jets [@Mitra:2016kov; @Mattelaer:2016ynf; @Cox:2017eme; @Nemevsek:2018bbt], , collimated clusters of electromagnetic and hadronic activity from decays of high-$p_T$ heavy neutrinos. Leading Order-accurate Monte Carlo simulations for tree-level processes involving $Z'$ bosons and heavy neutrinos in $U(1)_X$ theories are possible using the `SM+B-L` FeynRules UFO model [@FeynRules:BL; @Basso:2011na; @Basso:2008iv]. At NLO+PS accuracy, Monte Carlo simulations can be performed using the `Effective LRSM at NLO in QCD` UFO model[@Mattelaer:2016ynf; @FeynRules:LRSMnlo], and, for light, long-lived neutrinos and arbitrary $Z'$ boson couplings, the `SM + W’ and Z’ at NLO in QCD` UFO model [@Fuks:2017vtl; @FeynRules:WZPrimeAtNLO]. In $B-L$ models, heavy neutrino pairs can also be produced through the gluon fusion process mediated by the two $H_1$ and $H_2$ [@Basso:2010yz; @Pruna:2011me; @Accomando:2016rpc; @Accomando:2017qcs], and given by $$\begin{aligned} gg\to H_1, H_2\to N N.\end{aligned}$$ For long-lived heavy neutrinos with $M_N\lesssim 200$ GeV, this process becomes important compared to the channel mediated by $Z'$. Fig. \[typei-xsechiggs\] (a) shows that for $M_{H_2}<500$ GeV, $M_N<200$ GeV, and $M_{Z'}=5$ TeV, the cross section $\sigma(pp\to H_2\to NN)$ can be above 1 fb at the $\sqrt{s}=13$ TeV LHC. For $M_N<60$ GeV, decays of the SM-like Higgs $H_1$ also contributes to neutrino pair production. Summing over the contributions via $H_1$ and $H_2$ the total cross section can reach about 700 fb for $M_{H_2}<150$ GeV as shown in Fig. \[typei-xsechiggs\] (b). Owing to this extensive phenomenology, collider experiments are broadly sensitive to $Z'$ bosons from U$(1)_{BL}$ gauge theories. For example: Searches at LEP-II have set the lower bound of $M_{Z'}/g_{BL}\gtrsim 6$ TeV [@Carena:2004xs]. For more generic $Z'$ (including $Z_R$ in LRSM models), comparable limits from combined LEP+EW precision data have been derived in Ref. [@delAguila:2010mx; @delAguila:2011yd]. Direct searches for a $Z'$ with SM-like couplings to fermions exclude $M_{Z'}<2.9$ TeV at 95% CLs by ATLAS [@Aad:2014cka] and CMS [@Khachatryan:2014fba] at $\sqrt{s}=8$ TeV. $Z_{BL}$ gauge bosons with the benchmark coupling $g_1'=g_{BL}$ are stringently constrained by searches for dilepton resonances at the LHC, with $M_{Z'}\lesssim 2.1-3.75$ TeV excluded at 95% CLs for $g_{BL}=0.15-0.95$, as seen in Fig. \[fig:type1BLxSecNLONNLL\] [@Klasen:2016qux]. Searches for $Z'$ decays to dijets at the LHC have exclude $M_{Z'}<1.5-3.5$ TeV for $g_{BL}=0.07-0.27$ [@Sirunyan:2016iap; @Aaboud:2017yvp]. Fig. \[B-LcouplingmassATLAS\] (a) shows that ATLAS excludes $M_{Z'}<4.5$ TeV at $\sqrt{s}=13$ TeV. Further constraints are given in the plane of coupling strength $\gamma'=g_{BL}/g_Z$ vs. $M_{Z'}$ by ATLAS at $\sqrt{s}=13$ TeV with 36.1 fb$^{-1}$ [@Aaboud:2017buh] as shown in the lower curve of Fig. \[B-LcouplingmassATLAS\](b). For $\sqrt{s}=27$ TeV, early projections show that with $\mathcal{L}=1~(3)$ ab$^{-1}$, $M_{Z'}\lesssim19~(20)$ TeV can be probed in the dijet channel [@Chekanov:2017pnx]. ### Heavy Neutrinos and the Left-Right Symmetric Model at Colliders {#sec:lrsmCollider} In addition to the broad triplet scalar phenomenology discussed later in Sec. \[sec:type2Collider\], the LRSM predicts at low scales massive $W_R^{\pm}$ and $Z_R$ gauge bosons that couple appreciably to SM fields as well as to heavy Majorana neutrinos $N$. The existence of these exotic states leads to a rich collider phenomenology that we now address, focusing, of course, on lepton number violating final states. The collider phenomenology for $Z_R$ searches is very comparable to that for $Z'$ gauge bosons in U$(1)_X$ theories  [@delAguila:2007ua; @Huitu:2008gf; @Basso:2008iv; @AguilarSaavedra:2009ik; @Perez:2009mu; @Li:2010rb; @AguilarSaavedra:2012gf; @Accomando:2016sge; @Accomando:2017qcs], and thus we refer readers to Sec. \[sec:type1Abelian\] for more generic collider phenomenology. In the LRSM, for $M_N<M_{W_R}$ or $M_N<M_{Z_R}/2$, the most remarkable collider processes are the single and pair production of heavy Majorana neutrinos $N$ through resonant charged and neutral ${{\rm SU}(2)_{R}}$ currents, $$\begin{aligned} q \overline{q'} \to W_R^\pm \to ~N_i ~\ell^\pm \quad\text{and}\quad q \overline{q'} \to Z_R \to ~N_i ~N_j.\end{aligned}$$ As first observed in Ref. [@Keung:1983uu], $N_i$ can decay into $L$-violating final-states, giving rise to the collider signatures, $$pp \to W_R^\pm \to ~N_i ~\ell^\pm \to \ell_1^\pm ~\ell_2^\pm + nj \quad\text{and}\quad pp \to Z_R \to ~N_i ~N_j \to \ell_1^\pm ~\ell_2^\pm + nj. \label{eq:lrsmppssllnj}$$ In the minimal/manifest LRSM, the decay of $N_i$ proceeds primarily via off-shell three-body right-handed currents, as shown in Fig. \[fig:mLRSM\], due to mixing suppression to left-handed currents. In a generic LRSM scenario, the naïve mixing suppression of $\vert V_{\ell N}\vert^2 \sim\mathcal{O}(m_\nu / M_N)$ is not guaranteed due to the interplay between the Types I and II Seesaws, , as in Refs [@Anamiati:2016uxp; @Das:2017hmg]. (However, heavy-light neutrino mixing in the LRSM is much less free than in pure Type I scenarios due to constraints on Dirac and RH masses from LR parity; see Sec. \[sec:hybrid\] for more details.) Subsequently, if $\vert V_{\ell N}\vert$ is not too far from present bounds (see, ,  [@Fernandez-Martinez:2016lgt]), then decays of $N_i$ to on-shell EW bosons, as shown in Fig. \[fig:aLRSM\], can occur with rates comparable to decays via off-shell $W_R^*$ [@Han:2012vk]. The inverse process [@Chen:2013fna], , $N_i$ production via off-shell EW currents and decay via off-shell RH currents as well as vector boson scattering involving $t$-channel $W_R$ and $Z_R$ bosons [@Dev:2016dja] are in theory also possible but insatiably phase space-suppressed. For $M_N>M_{W_R},M_{Z_R}$, resonant $N$ production via off-shell ${{\rm SU}(2)_{R}}$ currents is also possible, and is analogous to the production through off-shell, ${{\rm SU}(2)_{L}}$ currents in Eqs. (\[eq:heavyNDYCC\])-(\[eq:heavyNDYNC\]). As $M_{W_R},M_{Z_R}$ are bound to be above a few-to-several TeV, the relevant collider phenomenology is largely the same as when $M_N<M_{W_R},M_{Z_R}$ [@Chakrabortty:2012pp], and hence will not be individually discussed. Aside from the mere possibility of $L$ violation, what makes these channels so exceptional, if they exist, are their production rates. Up to symmetry-breaking corrections, the RH gauge coupling is $g_R \approx g_L \approx 0.65$, which is not a small number. In Fig. \[fig:lrsmXSec\], we show for $\sqrt{s}=13$ and 100 TeV the production rate for resonant $W_R$ at various accuracies as a function of mass [@Mitra:2016kov]; rates for $Z_R$ are marginally smaller due to slight coupling suppression. As in other Seesaw scenarios, much recent progress has gone into advancing the precision of integrated and differential predictions for the LRSM: The inclusive production of $W_R$ and $Z_R$ are now known up to NLO+NNLL(Thresh) [@Mitra:2016kov], automated at NLO+NLL(Thresh+$k_T$) [@Fuks:2013vua; @Jezo:2014wra], automated at NNLO  [@Gavin:2010az; @Gavin:2012sy], and differentially has been automated at NLO with parton shower matching for Monte Carlo simulations [@Mattelaer:2016ynf]. For $\sqrt{\tau_0} = M_{W_R/Z_R}/\sqrt{s} \gtrsim 0.3$, threshold corrections become as large as (N)NLO corrections, which span roughly $+20\%$ to $+30\%$, and have an important impact cross section normalizations [@Mitra:2016kov; @Appell:1988ie]. For example: The inclusive $W_R$ cross section at LO (NLO+NNLL) for $M_{W_R}=5$ TeV is $\sigma\sim0.7~(1.7)$ fb. After $\mathcal{L}=1{{\rm ~ab^{-1}}}$ and assuming a combined branching-detection efficiency-selection acceptance of BR$\times\varepsilon\times\mathcal{A}=2\%$, the number of observed events is $N\sim14~(34)$. For simple Gaussian statistics with a zero background hypothesis, this is the difference between a $6\sigma$ “discovery” and $4\sigma$ “evidence”. Clearly, the HL-LHC program is much more sensitive to ultra-high-mass resonances than previously argued. For the collider processes in Eq. (\[eq:lrsmppssllnj\]), such estimations of branching, acceptance/selection, and background rates resemble actual rates; see, , [@Keung:1983uu; @Ferrari:2000sp; @Das:2012ii; @Han:2012vk; @Mitra:2016kov; @Chen:2013fna; @Dev:2016dja; @Goh:2009fm]. For $M_{W_R},~M_{Z_R}\gg M_{N}$, one finds generically that BR$(W_R \to \ell^\pm N_i)\sim 1/(1+3N_c)\sim\mathcal{O}(10\%)$, BR$(Z_R \to N_i N_j)\sim\mathcal{O}(10\%)$, and, for the lightest heavy $N_i$ in this limit, BR$(N_1 \to \ell^\pm X)\sim\mathcal{O}(100\%)$. Trigger rates for multi-TeV, stable charged leptons $(e,\mu)$ at ATLAS and CMS exceed $80\%-95\%$, but conversely, the momentum resolution for such energetic muons is severely degraded; for additional information, see [@Aad:2015xaa; @ATLAS:2016ecs; @Khachatryan:2014dka; @Khachatryan:2016jww] and references therein. As in searches for Majorana neutrinos in the previous Type I-based scenarios, the final-states in Eq. (\[eq:lrsmppssllnj\]) possess same-sign, high-$p_T$ charged leptons without accompanying MET at the partonic level [@Keung:1983uu; @Han:2006ip; @Ferrari:2000sp]. For the LRSM, this is particularly distinct since the kinematics of the signal process scale with the TeV-scale $W_R$ and $Z_R$ masses. Accordingly, top quark and EW background processes that can mimic the fiducial collider definition correspondingly must carry *multi*-TeV system invariant masses, and are inherently more phase space suppressed than the signal processes at the LHC [@Ferrari:2000sp]. Consequently, so long as $M_N \lesssim M_{W_R},~M_{Z_R} \ll \sqrt{s}$, $s$-channel production of $W_R$ and $Z_R$ remains the most promising mechanism for discovering $L$ violation in the LRSM at hadron colliders. In Fig. \[fig:lrsmDisc\] we show the discovery potential at 14 TeV LHC of $W_R$ and $N$ in (a) the minimal LRSM as in Fig. \[fig:mLRSM\] after $\mathcal{L}=30{{\rm ~fb^{-1}}}$ [@Das:2012ii] and (b) the agnostic mixing scenario as in Fig. \[fig:aLRSM\] [@Han:2012vk]. Final-states involving $\tau$ leptons are also possible, but inherently suffer from the difficult signal event reconstruction and larger backgrounds due to partonic-level MET induced by $\tau$ decays [@AguilarSaavedra:2012fu]. Unfortunately, direct searches at the $\sqrt{s}=7/8$ TeV LHC via the DY channels have yielded no evidence for lepton number violating processes mediated by $W_R$ and $Z_R$ gauge bosons from the LRSM [@ATLAS:2012ak; @Aad:2015xaa; @Khachatryan:2014dka; @CMS:2012zv]. As shown in Fig. \[fig:atlasLRSMExcl\], searches for $W_R/Z_R$ in the $e^\pm e^\pm+nj$ and $\mu^\pm\mu^\pm + nj$ final state have excluded, approximately, $M_{W_R/Z_R}\lesssim1.5-2.5$ TeV and $M_N\lesssim2$ TeV. However, sensitivity to the $e^\pm e^\pm+nj$ greatly diminishes for $M_N\ll M_{W_R/Z_R}$. Interestingly, for $M_N\ll M_{W_R}, M_{Z_R}$, decays of $N$ become highly boosted and its decay products, , $\ell^\pm_2 q \overline{q'}$, become highly collimated. In such cases, the isolation criterion for electrons (and some muons) in detector experiments fail, particularly when $\sqrt{r_N} = M_N / M_{W_R} < 0.1$ [@Ferrari:2000sp; @Han:2012vk; @Aad:2015xaa; @Mitra:2016kov]. Instead of requiring the identification of two well-isolated charged leptons for the processes given in Eq. (\[eq:lrsmppssllnj\]), one can instead consider the $N$-decay system as a single, high-$p_T$ *neutrino jet* [@Mitra:2016kov; @Mattelaer:2016ynf]. The hadronic-level collider signature is then $$\begin{aligned} p p \to W_R \to \ell^\pm ~N \to \ell^\pm ~j_N,\end{aligned}$$ where the neutrino jet $j_N$ is comprised of three “partons”, $(\ell_2,q,\overline{q'})$, with an invariant mass of $m_j \sim M_N$. (Neutrino jets are distinct from so-called “lepton jets” [@Izaguirre:2015pga], which are built from collimated charged leptons and largely absent of hadrons.) This alternative topology for $M_N\ll M_{W_R}$ recovers the lost sensitivity of the same-sign dilepton final state, as seen in Fig. \[fig:lrsmNjetDiscov\]. Inevitably, for $N$ masses below the EW scale, rare $L$-violating decay modes also of SM particles open. In particular, for $M_N$ below the top quark mass $m_t$, one has the rare decay mode, $t \to b W_R^{+*} \to b \ell_1^+ N \to b \ell_1^+ \ell_2^\pm q \overline{q'}$ [@Si:2008jd]. Such processes, however, can be especially difficult to distinguish from rare SM processes, , $t\to W b \ell^+\ell^-$ [@Quintero:2014lqa], particularly due to the large jet combinatorics. \ For too small $M_N/M_{W_R}$ ratio, the lifetime for $N$, which scales as $\tau_N \sim M_{W_R}^4/M_N^5$, can become quite long. In such instances, the decays of $N$ are no longer prompt and searches for $pp\to W_R\to N\ell$ map onto searches for Sequential Standard Model $W'$ bosons [@Maiezza:2015lza; @Fuks:2017vtl]. Likewise, searches for $L$-violating top quark decays become searches for RH currents in $t\to b\ell\not\!p_T$ decays. For intermediate lifetimes, displaced vertex searches become relevant [@Helo:2013esa; @Anamiati:2016uxp; @Das:2017hmg; @Cottin:2018kmq; @Nemevsek:2018bbt]. \ Another recent avenue of exploration is the reassessment for resonant production of $W_R$ and $Z_R$ in Eq. (\[eq:lrsmppssllnj\]). In the limit where $M_{W_R}\gtrsim \sqrt{s}$ but $M_N\ll \sqrt{s}$, resonant production of $N$, and hence a lepton number violating final state, is still possible despite $W_R$ being kinematically inaccessible [@Ruiz:2017nip]. In such cases, $N$ is produced near mass threshold with $p_T^N \sim M_N$ instead of the usual $p_T^N \sim M_{W_R}/2$. The same-sign leptons discovery channel is then kinematically and topologically identical to Type I Seesaw searches, and hence is actively searched for at the LHC, despite this kinematic regime not being well-studied in the literature. Reinterpretation of observed and expected sensitivities at the 14 and 100 TeV LHC are shown in Fig. \[fig:lrsmRecast\]. One sees that with the anticipated cache of LHC data, $M_{W_R}\lesssim 9$ TeV can be excluded for $M_N\lesssim 1$ TeV. In addition to the aforementioned DY and VBF channels, there has been recent attention [@Chen:2014qda; @Frank:2010cj; @Dev:2016dja; @Mattelaer:2016ynf] given to the production of LRSM scalar and vector bosons in association with heavy flavor quarks, , $$g \overset{(-)}{b} \to \overset{(-)}{t} W_R^\pm ~\text{or}~ \overset{(-)}{t}H^\pm_R \quad\text{and}\quad g g \to t\overline{t}Z_R ~\text{or}~~t\overline{t}H_R^0.$$ As in the SM, such processes are critical in measuring the couplings of gauge bosons to quarks as well as determining heavy flavor PDFs. However, also as in the SM, care is needed in calculating the rates of these processes when $M_R \gg m_b,~m_t$. Here, $M_R$ is generically the mass of the RH scalar or vector boson. As discussed just after Eq. (\[eq:heavyNnj\]), it has been noted recently in Ref. [@Mattelaer:2016ynf] that such associated processes possess logarithmic dependence on the outgoing top quarks’ kinematics, , that the inclusive cross section scales as $\sigma\sim \alpha_s^k\log^{2k-1}\left(M_R^2/(m_t^2+p_T^{t~2})\right)$. Subsequently, for $M_R \gtrsim 1-2$ TeV, these logarithms grow numerically large since $\log^2(M_R^2/m_t^2)\gtrsim 1/\alpha_s$ and can spoil the perturbativity convergence of fixed order predictions. For example, the (N)NLO $K$-factor of $K^{\rm (N)NLO}\gtrsim 1.6-2.0$ claimed in Ref. [@Dev:2016dja] indicate a loss of perturbative control, not an enhancement, and leads to a significant overestimation of their cross sections. As in the case of EW boson production in association with heavy flavors [@Dicus:1988cx; @Maltoni:2012pa], the correct treatment requires either a matching/subtraction scheme with top quark PDFs to remove double counting of phase space configurations [@Dawson:2014pea; @Han:2014nja] or kinematic requirements on the associated top quarks/heavy quark jets, , Eq. (\[eq:cssConsistency\]) [@Degrande:2016aje]. In all of these various estimates for discovery potential, it is important to also keep in mind what can be learned from observing $L$ violation and LR symmetry at the LHC or a future collider, including $ep$ machines [@Kaya:2015tia; @Mattelaer:2016ynf; @Lindner:2016lxq; @Buchmuller:1992wm; @Mondal:2016czu; @Mondal:2015zba; @Sarmiento-Alvarado:2014eha; @Dev:2015vra]. Primary goals post-discovery include: determination of $W_R$ and $Z_R$ chiral coupling to fermions [@Han:2012vk; @Gopalakrishna:2010xm; @Nemevsek:2012iq], which can be quantified for quarks and leptons independently [@Han:2012vk], determination of the leptonic and quark mixing [@Vasquez:2014mxa; @Senjanovic:2016vxw; @Nemevsek:2012iq; @Tello:2010am; @Gluza:2015goa; @Gluza:2016qqv; @Anamiati:2016uxp; @Das:2017hmg], as well as potential CP violation [@Bajc:2009ft; @Anamiati:2016uxp; @Gluza:2015goa; @Gluza:2016qqv; @Das:2017hmg]. We emphasize that the discovery of TeV-scale LRSM could have profound implications on high-scale baryo- and leptogenesis [@GellMann:1980vs; @Frere:2008ct; @Dhuria:2015cfa; @Deppisch:2013jxa; @Harz:2015fwa] as well as searches for $0\nu\beta\beta$  [@Tello:2010am; @Barry:2013xxa; @Nemevsek:2012iq; @Peng:2015haa; @Lindner:2016lpp]. The latter instance is particularly noteworthy as the relationship between $m^{ee}_\nu$ and $m_{\nu_1}$ in the LRSM is different because of the new mediating fields [@Tello:2010am]. We finish this section by noting our many omissions, in particular: supersymmetric extensions of the LRSM, [*e.g.*]{}, [@Frank:1995dh; @Demir:2009nq]; embeddings into larger internal symmetry structures, [*e.g.*]{}, [@Goh:2009fm; @Appelquist:2003uu]; as well as generic extensions with additional vector-like or mirror quarks, [*e.g.*]{}, [@Goh:2009fm; @deAlmeida:2010qb]. While each of these extensions have their phenomenological uniquenesses, their collider signatures are broadly indistinguishable from the minimal LRSM scenario. With regard to Type I-based Seesaws in extra dimensional frameworks, it is worthwhile to note that it has recently [@Agashe:2015izu; @Agashe:2016ttz; @Agashe:2017ann] been observed that in warped five-dimensional models, a more careful organization of Kaluza-Klein states and basis decomposition results in an inverse Seesaw mechanism as opposed to a canonical Type I-like Seesaw mechanism, as conventionally believed. Again, this leads to greatly suppressed $L$ violation at collider experiments. ### Heavy Neutrino Effective Field Theory at Colliders {#sec:neftTests} As discussed in Sec. \[sec:neft\], the production and decay of Majorana neutrinos in colliders may occur through contact interactions if mediating degrees of freedom are much heavier than the hard scattering process scale. Such scenarios have recently become a popular topic [@Bhattacharya:2015vja; @Liao:2016qyd; @Ruiz:2017nip; @Leonardi:2015qna; @Duarte:2016miz; @Duarte:2016caz; @Duarte:2015iba; @Duarte:2014zea; @Caputo:2017pit], in part because of the considerable sensitivity afforded by collider experiments. This is particularly true for $L$-violating final-states in $pp$ collisions, which naturally have small experimental backgrounds. As shown in Fig. \[fig:neftLimits\], for various operators, searches for $L$-violating process $pp\to N\ell^\pm_1 \to \ell^\pm_1 \ell^\pm_2+X$ by the ATLAS and CMS experiments have set wide limits on the effective mass scale of $\Lambda > 1-5$ TeV for $M_N = 100$ GeV$ -$4.5 TeV [@Leonardi:2015qna; @Ruiz:2017nip; @CMS:2016blm]. Projections for $\sqrt{s}=14~(100)$ TeV after $\mathcal{L}=1~(10)$ ab$^{-1}$ show that $\Lambda\lesssim9~(40)$ TeV can be achieved [@Ruiz:2017nip]. These search strategies are also applicable for the more general situation where $L$ violation is mediated entirely via SMEFT operators [@Abada:2007ux; @delAguila:2012nu] as introduced in Sec. \[sec:neft\]. The Type II Seesaw and Lepton Number Violation at Colliders {#sec:type2} =========================================================== In this section we review lepton number violating collider signatures associated with the Type II Seesaw mechanism [@Konetschny:1977bn; @Gelmini:1980re; @Cheng:1980qt; @Lazarides:1980nt; @Schechter:1980gr; @Mohapatra:1980yp] and its extensions. The Type II model is unique among the original tree-level realizations of the Weinberg operator in that lepton number is spontaneously broken; in the original formulations of the Type I and III Seesaws, lepton number violation is explicit by means of a Majorana mass allowed by gauge invariance. In Sec. \[sec:type2Models\], we summarize the main highlights of the canonical Type II Seesaw and other Type II-based scenarios. We then review in Sec. \[sec:type2Collider\] collider searches for lepton number violation mediated by exotically charged scalars $(H^\pm,H^{\pm\pm})$, which is the characteristic feature of Type II-based scenarios. \ Type II Seesaw Models {#sec:type2Models} --------------------- In the Type II mechanism [@Konetschny:1977bn; @Gelmini:1980re; @Cheng:1980qt; @Lazarides:1980nt; @Schechter:1980gr; @Mohapatra:1980yp], tiny neutrino masses arise through the Yukawa interaction, $$\begin{aligned} \Delta\mathcal{L}_{II}^m = - \overline{L^c} ~Y_\nu~ i\sigma_2 \ \Delta_L \ L+~\Hc, \label{eq:typeIIYukawa}\end{aligned}$$ between the SM LH lepton doublet $L$, its charge conjugate, and an SU$(2)_L$ scalar triplet (adjoint representation) $\Delta_L$ with mass $M_\Delta$ and Yukawa coupling $Y_\nu$. More precisely, the new scalar transforms as $(1,3,1)$ under the full SM gauge symmetry and possesses lepton number $L=-2$, thereby ensuring that Eq. (\[eq:typeIIYukawa\]) conserves lepton number before EWSB. Due to its hypercharge and $L$ assignments, $\Delta_L$ does not couple to quarks at tree-level. It does, however, couple to the SM Higgs doublet, particularly through the doublet-triplet mixing operator $$\Delta\mathcal{L}_{H\Delta_L} \ni \mu H^T \ i\sigma_2 \ \Delta_L^\dagger H+~\Hc$$ The importance of this term is that after minimizing the full Type II scalar potential $V_{\rm Type~II}$, $\Delta_L$ acquires a small vev $v_\Delta$ that in turn induces a LH Majorana mass for SM neutrinos, given by $$\begin{aligned} M_\nu=\sqrt{2}Y_\nu v_\Delta \quad\text{with}\quad v_\Delta= \langle\Delta_L\rangle = {\mu v_0^2\over \sqrt{2} M_\Delta^2}.\end{aligned}$$ In the above, $v_0 = \sqrt{2}\langle H \rangle$ is the vev of the SM Higgs and $v_0^2+v_\Delta^2 = (\sqrt{2}G_F)^{-1}\approx (246 \ {\rm GeV})^2$. As a result of being spontaneously broken by $\Delta_L$, tiny $0.1\ev$ neutrino masses follow from the combination of three scales: $\mu$, $v_0$, and $M_\Delta$. In addition, after EWSB, there are seven physical Higgses, including the singly and doubly electrically charged $H^\pm$ and $H^{\pm\pm}$ with masses $M_{H^\pm,H^{\pm\pm}} \sim M_\Delta$. As $v_\Delta$ contributes to EWSB at tree-level, and hence the EW $\rho/T$-parameter, $v_\Delta$ is constrained by precision EW observables, with present limits placing $v_\Delta\lesssim\mathcal{O}(1$ GeV) [@Perez:2008ha; @Dutta:2014dba; @Han:2005nk; @Melfo:2011nx; @Arhrib:2011uy; @Kanemura:2012rs; @Chen:2008jg; @Chen:2005jx; @Das:2016bir]. The impact of triplet scalars on the naturalness of the SM-like Higgs at 125 GeV has also been studied [@Dev:2013ff; @Arhrib:2011uy; @Dev:2017ouk]. The simultaneous sensitivity of $M_\nu$ to collider, neutrino mass measurement, and neutrino oscillation experiments is one of the clearest examples of their complementarity and necessity to understanding neutrinos physics. ![Constraints on the squared coupling $Y_+^i\equiv \sum_j |\Gamma_{+}^{ji}|^2 v^2_{\Delta}$, versus the lowest neutrino mass for NH (a) and IH (b). []{data-label="Yi"}](Ynh.jpg "fig:"){width="6cm"}\ ![Constraints on the squared coupling $Y_+^i\equiv \sum_j |\Gamma_{+}^{ji}|^2 v^2_{\Delta}$, versus the lowest neutrino mass for NH (a) and IH (b). []{data-label="Yi"}](Yih.jpg "fig:"){width="6cm"}\ For SM-like Yukawas $Y_\nu \sim 10^{-6} - 1$, one finds that $v_\Delta \sim 0.1\ev - 100\kev$ are needed in order to reproduce $0.1$ eV neutrino masses. Subsequently, for $\mu \sim M_\Delta$, then $M_\Delta \sim \mu \sim 10^8 -10^{14}$ GeV, and for $\mu \sim v_0$, then $M_\Delta\sim 10^5 - 10^8$ GeV. In either case, these scales are too high for present-day experiments. However, as nonzero $\mu$ is associated with both lepton number and custodial symmetry non-conservation, one may expect it to be small [@Senjanovic:1978ev] and natural, in the t’Hooft sense [@tHooft:1979rat]. Imposing technical naturalness can have dramatic impact on LHC phenomenology: for example, if $\mu \sim 1$ MeV (keV), then $M_\Delta\sim 10^2 - 10^5~(10^1-10^4)$ GeV, scales well within the LHC’s energy budget. Moreover, this also indicates that proposed future hadron collider experiments [@Arkani-Hamed:2015vfh; @Golling:2016gvc] will be sensitive to MeV-to-GeV values of the scalar-doublet mixing parameter $\mu$, independent of precision Higgs coupling measurements, which are presently at the 10% level [@Khachatryan:2016vau]. Assuming Higgs coupling deviations of $\mathcal{O}(\mu/M_h)$, this implies the weak 7/8 TeV LHC limit of $\mu\lesssim\mathcal{O}$(10 GeV). While not yet competitive with constraints from EW precision data, improvements on Higgs coupling measurements will be greatly improved over the LHC’s lifetime. After decomposition of leptons into their mass eigenstates, the Yukawa interactions of the singly and doubly charged Higgses are $$\begin{aligned} \nu_L^T \ C \ \Gamma_+ \ H^+ \ \ell_L, \quad&:&\quad \Gamma_+ = \cos \theta_+ \ \frac{m_\nu^{diag}}{v_{\Delta}} \ U_{PMNS}^\dagger, \ \ \ \theta_+ \approx {\sqrt 2v_\Delta\over v_0}, \\ \ell_L^T \ C \ \Gamma_{++} \ H^{++} \ \ell_L \quad&:&\quad \Gamma_{++} = {M_\nu\over \sqrt{2} v_\Delta} = U_{PMNS}^* \ \frac{m_{\nu}^{diag}}{\sqrt{2} \ v_{\Delta}} \ U_{PMNS}^{\dagger}.\end{aligned}$$ The constrained neutrino mass matrix $M_\nu=\sqrt{2}v_\Delta \Gamma_{++}$ and squared Yukawa coupling $Y_+^i\equiv \sum_j |\Gamma_{+}^{ji}|^2 v^2_{\Delta}$ with vanishing Majorana phases are shown in Figs. \[mii\] and \[Yi\] respectively. The results reveal the following mass and Yukawa patterns: $$\begin{aligned} M_\nu^{22},M_\nu^{33}\gg M_\nu^{11} \quad\text{and}\quad Y_+^2,Y_+^3\gg Y_+^1 \qquad & {\rm for \ NH}; \\ M_\nu^{11}\gg M_\nu^{22},M_\nu^{33} \quad\text{and}\quad Y_+^1\gg Y_+^2,Y_+^3 \qquad & {\rm for \ IH}.\end{aligned}$$ ![Scatter plots for the $H^{++}$ decay branching fractions to the flavor-diagonal like-sign dileptons versus the lowest neutrino mass for NH (a) and IH (b) with $\Phi_1 = \Phi_2 = 0$.[]{data-label="brii"}](Hppbrnhdiag.jpg "fig:"){width="7.cm"}\ ![Scatter plots for the $H^{++}$ decay branching fractions to the flavor-diagonal like-sign dileptons versus the lowest neutrino mass for NH (a) and IH (b) with $\Phi_1 = \Phi_2 = 0$.[]{data-label="brii"}](Hppbrihdiag.jpg "fig:"){width="7.cm"}\ ![$H^{++}$ decay to the flavor-off-diagonal like-sign dileptons versus the lowest neutrino mass for NH (a) and IH (b) with $\Phi_1 = \Phi_2 = 0$. []{data-label="brij"}](Hppbrnhoffdiag.jpg "fig:"){width="7.cm"}\ ![$H^{++}$ decay to the flavor-off-diagonal like-sign dileptons versus the lowest neutrino mass for NH (a) and IH (b) with $\Phi_1 = \Phi_2 = 0$. []{data-label="brij"}](Hppbrihoffdiag.jpg "fig:"){width="7.cm"}\ Below $v_{\Delta}\approx 10^{-4}$ GeV, the doubly charged Higgs $H^{\pm\pm}$ decays dominantly to same-sign lepton pairs. For vanishing Majorana phases $\Phi_1 = \Phi_2 = 0$, we show in Figs. \[brii\] and \[brij\] the branching fraction of the decays into same-flavor and different-flavor leptonic final states, respectively. Relations among the branching fractions of the lepton number violating Higgs decays of both the singly- and doubly-charged Higgs in the NH and IH, with vanishing Majorana phases, are summarized in Table. \[relationii\]. Relations ---- ----------------------------------------------------------------------------------------------------------------------- NH BR$( H^{++} \to \tau^+ \tau^+ / \mu^+ \mu^+ ) \sim (20-40)\% \gg$ BR$( H^{++} \to e^+ e^+ )\sim (0.1-0.6)\%$ BR$( H^{++} \to \mu^+ \tau^+ )\sim (30-40)\% \gg $ BR$(H^{++} \to e^+ \mu^+ / e^+ \tau^+ )\lesssim 5\%$ BR$( H^{+} \to \tau^+ \bar{\nu} / \mu^+ \bar{\nu} )\sim (30-60)\% \gg $ BR$(H^{+} \to e^+ \bar{\nu} )\sim (2.5-3)\%$ IH BR$( H^{++} \to e^+ e^+ )\sim 50\% \ > \ $ BR$( H^{++} \to \mu^+ \mu^+ / \tau^+ \tau^+ )\sim (6-20)\%$ BR$( H^{++} \to \mu^+ \tau^+ )\sim (20-30)\% \ \gg \ $ BR$( H^{++} \to e^+ \mu^+ / e^+ \tau^+ )\sim (0.1-4)\%$ BR$( H^{+} \to e^+ \bar{\nu} )\sim 50\% \ > \ $ BR$( H^{+} \to \mu^+ \bar{\nu} / \bar{\nu} )\sim (20-30)\%$ : Relations among the branching fractions of the lepton number violating Higgs decays for the neutrino mass patterns of NH and IH, with vanishing Majorana phases. []{data-label="relationii"} The impact of Majorana phases can be substantial in doubly charged Higgs decays [@Akeroyd:2007zv; @Garayoa:2007fw]. In the case of the IH, a large cancellation among the relevant channels occurs due to the phase at $\Phi_1=\pi$. As a result, in this scenario, the dominant channels swap from $H^{++} \to e^+e^+,\ \mu^{+} \tau^+$ when $\Phi_1 \approx 0$ to $H^{++} \to e^+ \mu^+,\ e^{+} \tau^+$ when $\Phi_1 \approx \pi$, as shown in Fig. \[Majorana2\]. Therefore this qualitative change can be made use of to extract the value of the Majorana phase $\Phi_1$. In the NH case, however, the dependence of the decay branching fractions on the phase is rather weak because of the lack of a subtle cancellation [@Perez:2008ha]. ![Scatter plots of the same (a) and different (b) flavor leptonic branching fractions for the $H^{++}$ decay versus the Majorana phase $\Phi_1$ for the IH with $m_3=0$ and $\Phi_2 \in (0, 2\pi)$.[]{data-label="Majorana2"}](Hppbrphiihdiag.jpg "fig:"){width="6cm"}\ ![Scatter plots of the same (a) and different (b) flavor leptonic branching fractions for the $H^{++}$ decay versus the Majorana phase $\Phi_1$ for the IH with $m_3=0$ and $\Phi_2 \in (0, 2\pi)$.[]{data-label="Majorana2"}](Hppbrphiihoffdiag.jpg "fig:"){width="6cm"}\ The Type II mechanism can be embedded in a number of extended gauge scenarios, for example the LRSM as discussed in Sec. \[sec:hybrid\], as well as GUTs, such as $(331)$ theories [@Cuypers:1996ia; @Tully:2000kk; @Fonseca:2016xsy; @Cogollo:2008zc] and the extensions of minimal SU$(5)$ [@Georgi:1974sy]. For $(331)$ models, one finds the presence of bileptons [@Frampton:1992wt; @Pisano:1991ee], , gauge bosons with $L = \pm2$ charges and hence $Q=\pm2$ electric charges. In a realistic extension of the Georgi-Glashow model, a scalar 15-dimensional representation is added [@Dorsner:2005fq] and the scalar triplet stays in the $\textbf{15}$ representation together with scalar leptoquark $\Phi\sim (3,2,1/6)$. The SU$(5)$ symmetry thus indicates that the couplings of the leptoquark to matter gain the same Yukawas $Y_\nu$ responsible for neutrino mass matrix [@FileviezPerez:2008dw]. Extensions with vector-like leptons in nontrivial SU$(2)_L$ representations are also possible [@Bahrami:2013bsa]. Unsurprisingly, the phenomenology [@Cuypers:1996ia; @Fonseca:2016xsy; @Nepomuceno:2016jyr; @Corcella:2017dns; @Meirose:2011cs] and direct search constraints [@Nepomuceno:2016jyr; @Corcella:2017dns] for $L$-violating, doubly charged vector bosons are similar to $L$-violating, doubly charged scalar bosons, which we now discuss. Triplet Higgs Scalars at Colliders {#sec:type2Collider} ---------------------------------- ### Triplet Higgs Scalars and the Type II Seesaw at Colliders ![ Born-level diagrams depicting Type II triplet scalar production in $pp$ collisions via (a) the DY mechanism, (b) same-sign $W^\pm W^\pm$ scattering, and (c) $\gamma\gamma$ fusion. []{data-label="fig:feynmanTypeIIScalars"}](feynman_TypeII_ScalarProduction.pdf "fig:"){width=".9\textwidth"} \[fig:feynmanTypeIIScalars\] If kinematically accessible, the canonical and well-studied [@Rizzo:1981xx; @Akeroyd:2005gt; @Perez:2008ha; @Bambhaniya:2013wza] triplet scalars production channels at hadron colliders are the neutral and charged current DY processes, given by $$\begin{aligned} pp\to \gamma^*/Z^* \to H^{++} H^{--}, \ \ \ pp\to W^{\pm*} \to H^{\pm\pm} H^{\mp},\end{aligned}$$ and shown in Fig. \[fig:feynmanTypeIIScalars\](a). Unlike Type I models, scalars in the Type II Seesaw couple to EW bosons directly via gauge couplings. Subsequently, their production rates are sizable and can be predicted as a function of mass without additional input. In Fig. \[fig:typeiiScalarXSec\] we show the LO pair production cross section of triplet scalars via the (a) neutral and (b) charged current DY process at $\sqrt{s} = 14$ and 100 TeV. NLO in QCD corrections to these processes are well-known [@Muhlleitner:2003me] and span $K^{\rm NLO } = \sigma^{\rm NLO}/\sigma^{\rm LO} = 1.1 - 1.3$ away from boundaries of collider phase space; moreover, due to the color-structure of DY-like processes, inclusive kinematics of very heavy scalar triplets are Born-like and thus naïve normalization of kinematics by $K^{\rm NLO }$ gives reliable estimates of both NLO- and NLO+PS-accurate results [@Ruiz:2015zca; @Fuks:2017vtl]. For $M_{H^{\pm\pm}} = 1$ TeV, one finds that the LO pair production rates can reach $\sigma\sim0.1~(10)$ fb at $\sqrt{s}=14$ (100) TeV, indicating $\mathcal{O}(10^2)~(\mathcal{O}(10^4))$ of events with the ab$^{-1}$-scale data sets expected at the respective collider program. \ In addition to the DY channels are: single production of charged Higgses via weak boson scatter, as shown in Fig. \[fig:feynmanTypeIIScalars\](b) and investigated in [@Han:2005nk; @Chen:2008qb]; charged Higgs pair production via $\gamma\gamma$ scattering, as shown in Fig. \[fig:feynmanTypeIIScalars\](c), studied in [@Drees:1994zx; @Han:2007bk; @Bambhaniya:2015wna; @Dutta:2014dba; @Babu:2016rcr], and computed at $\sqrt{s}=14$ TeV [@Han:2007bk] in Fig. \[fig:type2\_DblyChargedHiggsXSec\]; as well as pair production through weak boson scattering, as studied in  [@Bambhaniya:2015wna; @Dutta:2014dba] and computed for the 14 TeV LHC [@Dutta:2014dba] in Fig. \[fig:type2\_vbfHiggs\_ProdXSec\]. As in the case of $W\gamma$ scattering in heavy $N$ production in Sec. \[sec:type1\], there is renewed interest [@Bambhaniya:2015wna] in the $\gamma\gamma$-mechanisms due to the new availability of photon PDFs that include both elastic and (deeply) inelastic contributions, , NNPDF 2.3 and 3.0 QED PDF sets [@Ball:2013hta; @Ball:2014uwa]. However, care should be taken in drawing conclusions based on these specific PDF sets due to the (presently) large $\gamma$-PDF uncertainty, particularly at large Bjorken-$x$ where this can reach greater than $100\%$ [@Ball:2013hta]. For example: As shown in Fig. \[fig:type2\_DblyChargedHiggsXSec\], $\gamma\gamma$ production is unambiguously sub-leading to the DY mechanism and only contributes about $10\%$ despite recent claims to the contrary [@Babu:2016rcr; @Ghosh:2017jbw]. The collinear behavior and the factorization scale dependence of the incoming photons must be treated with great care. As more data is collected and $\gamma$-PDF methodology further matures, one anticipates these uncertainties to greatly shrink; for further discussions of $\gamma$-PDFs, see Refs. [@Martin:2014nqa; @Alva:2014gxa; @Harland-Lang:2016kog; @Degrande:2016aje; @Manohar:2016nzj; @Manohar:2017eqh]. For a list of recommended $\gamma$-PDFs, see the discussion just above Eq. (\[eq:heavyNnj\]). \ Similar to the $\gamma\gamma$ channel, production of triplet scalars from gluon fusion is sub-leading with respect to DY due to multiple vanishing contributions [@delAguila:1990yw; @Hessler:2014ssa] and despite an expectedly large QCD correction of $K^{\rm N^3LL} = \sigma^{\rm N^3LL}/\sigma^{\rm LO}\sim2.5-3$ [@Ruiz:2017yyf]. If triplet scalar couplings to the SM-like Higgs are not too small and if sufficiently light, then such scalars may appear in pairs as rare decays of the 125 GeV scalar boson [@Nemevsek:2016enw]. Likewise, if neutral triplet scalars mix appreciably with the SM-like Higgs, then single production via gluon fusion is also possible [@Nemevsek:2016enw]; one should note that in such cases, the QCD $K$-factors calculated in Ref. [@Ruiz:2017yyf] are applicable. A noteworthy direction of progress in searches for triplet scalars at colliders are the implementation of exotically charged scalars into FeynRules model files. In particular, lepton number violating scalars are available in the `LNV-Scalars` [@delAguila:2013mia; @FeynRules:LNVScalars] model file as well as in a full implementation of LRSM at LO accuracy [@Roitgrund:2014zka; @FeynRules:LRSMlo]; the Georgi-Machacek model [@Georgi:1985nv] is also available at NLO in QCD accuracy [@Degrande:2015xnm; @FeynRules:GMnlo]. These permit simulation of triplet scalar production in inclusive $\ell\ell/\ell p/pp$ collisions using modern, general-purpose event generators, such as Herwig [@Bellm:2015jjp], MadGraph5\_aMC@NLO [@Alwall:2014hca], and Sherpa [@Gleisberg:2008ta]. \ Due to the unknown Yukawa structure in Eq. (\[eq:typeIIYukawa\]), the decays of the triplet scalars to SM states are much more ambiguous than their production. Subsequently, branching rates of $H^{\pm}\to \ell^\pm \nu$ and $H^{\pm\pm}\to\ell^\pm_1\ell^\pm_2$ are often taken as phenomenological parameters in analyses and experimental searches. When taking such a model-agnostic approach, it may be necessary to also consider the lifetimes of scalar triplets: In a pure Type II scenario, for $M_{H^{\pm\pm}}<270$ GeV and sub-MeV values of the triplet vev $v_L$, the proper decay length of $H^{\pm\pm}$ can exceed 10 $\mu$m [@Han:2005nk]. As a result, exotically charged triplet scalars may manifest at collider experiments in searches for long-lived, multi-charged particles such as Refs. [@Aad:2013pqd; @Aad:2015oga; @CMS:2012aza; @Barrie:2017eyd]. For prompt decays of triplet scalars, the discovery potential at hadron colliders is quantified in Fig. \[fig:typeiiDisc\]. In particular, following the analysis of Ref. [@Perez:2008ha], Figs. \[fig:type2\_TripletContour14TeV\] and \[fig:type2\_TripletContour100TeV\] show event contours in the ${\rm BR}(H^{++}\to \mu^+\mu^+)$ vs. $M_{H^{\pm\pm}}$ plane after $\mathcal{L}=300~(3000){{\rm ~fb^{-1}}}$ of data at $\sqrt{s}=14$ TeV and 100 TeV, respectively. At the $2\sigma$ level, one finds the sensitivity to doubly charged Higgs is about $M_{H^{\pm\pm}}=0.75 ~(1.1)$ TeV at 14 TeV and $M_{H^{\pm\pm}}=2~(3.5)$ TeV at 100 TeV. In Figs. \[fig:type2\_vbfHiggs\_emuSig\] and \[fig:type2\_vbfHiggs\_tauSig\], one similarly has the signal significance $\sigma = S/\sqrt{S+B}$ after $\mathcal{L}=1$ and $3{{\rm ~ab^{-1}}}$ at the 14 TeV LHC for VBF production of doubly charged Higgs pairs and their decays to $e^\pm\mu^\pm$ and $\tau^\pm\tau^\pm$ final-states, respectively [@Dutta:2014dba]. Upon the fortuitous discovery of a doubly charged scalar, however, will require also observing other charged scalars to determine its precise weak isospin and hypercharge quantum numbers [@Bambhaniya:2013wza; @delAguila:2013yaa; @delAguila:2013mia]. In light of such sensitivity at hadron colliders, it is unsurprising then that null results from searches at the 7/8/13 TeV LHC [@Chatrchyan:2012ya; @ATLAS:2014kca; @ATLAS:2017iqw; @CMS:2017pet] have placed stringent constraints on EW-scale triplet scalar masses, assuming benchmark branching rates. As seen in Fig. \[fig:type2ExclATLAS\], results from the ATLAS experiment in searches for doubly charged Higgs pairs decaying to leptons, after collecting $\mathcal{L}=36{{\rm ~fb^{-1}}}$ of data at 13 TeV, have ruled out $M_{H^{\pm\pm}} > 600-900$ GeV at 95% CLs in both the (a) single-flavor and (b) mixed light-lepton final states [@ATLAS:2017iqw]. Comparable limits have been reached by the CMS experiment [@CMS:2017pet]. At future $e^-e^+$ colliders, triplet scalars can appear in $t$-channel exchanges, inducing charged lepton flavor violation (cLFV) and forward-backward asymmetries [@Nomura:2017abh]; in three-body decays of taus that are absent of light-neutrinos in the final state, , $\tau^\pm \to \ell^\mp H^{\pm\pm *} \to \ell^\mp \mu^\pm \mu^\pm$ [@Hays:2017ekz]; and, of course, in pairs via $s$-channel gauge currents [@Frank:1995ex]. In the event of such observations, the nontrivial conversion of an $e^-e^+$ beam into an $e^-e^-/e^-\mu^-/\mu^-\mu^-$ facility could provide complimentary information on scalar triplet Yukawa couplings by means of the “inverse” $0\nu\beta\beta$ processes, $\ell^-_i\ell^-_j \to W^-_{L/R} W^-_{L/R}$ [@Rizzo:1982kn; @Rodejohann:2010jh; @Barry:2012ga]. ### Triplet Higgs Scalars and the Left-Right Symmetric Model at Colliders {#sec:type2LRSM} Turning to scalars in the LRSM, as introduced in Sec. \[sec:hybrid\], it was recently observed [@Maiezza:2015lza; @Nemevsek:2016enw] that in a certain class of neutrino mass models, decays of the SM-like Higgs boson $h(125~{\rm GeV})$ to heavy neutrino pairs, $h\to NN$, may occur much more readily than previously thought. The significance of this reaction is one’s ability to confirm neutrino masses are generated, in part, through EWSB. It would also indicate sensitivity to the scalar sector responsible for generating RH Majorana masses. Interactions between SM particles and $N$ typically proceed through heavy-light neutrino mixing, $\vert V_{\ell N}\vert$, which, is a numerically small quantity. As $h\to NN$ involves two $N$, the issue is compounded and usually renders the decay rate prohibitively small in a pure Type I scenario. For $H\in\{H^0,~H^\pm,~H^{\pm\pm}\}$ predicted in Type I+II Seesaws, and in particular the LRSM, the situation is more interesting: it may be that $h(125\gev)$ and the RH neutral scalars mix sufficiently that decays to relatively light $(2M_N < 125\gev)$ heavy neutrino pairs are possible [@Maiezza:2015lza]. This is allowed as $H$ can couple appreciable to $N$ and the mixing between $H^0$ and $h$ is much less constrained. Subsequently, the naïve neutrino mixing suppression is avoided by exploiting that $h\to N N$ decays can proceed instead through $H^0-h$ mixing. In a similar vein, it may be possible for $h$ to decay to triplet pairs and subsequently to $N$ or same-sign charged leptons, or for single $H^0$ production to proceed directly [@Nemevsek:2016enw]. Such processes are shown diagrammatically in Fig. \[fig:majoranaHiggsDiagrams\]. As a result, the $L$-violating Higgs decays, $$\begin{aligned} h(125\gev) &\to& ~N ~N ~\to ~W_R^{\pm *} ~W_R^{\pm *} ~\ell_1^\mp ~\ell_2^\mp ~\to ~\ell_1^\mp ~\ell_2^\mp ~+~ nj, \\ h(125\gev) &\to& ~H^0 ~H^0 ~\to ~4N ~\to ~\ell_1^\pm ~\ell_2^\mp ~\ell_3^\pm ~\ell_4^\mp ~+~ nj, \\ h(125\gev) &\to& H^{++} ~H^{--} ~\to ~\ell_1^\pm ~\ell_2^\pm ~\ell_3^\mp ~\ell_4^\mp,\end{aligned}$$ are not only possible, but also provide complementary coverage of low-mass $N$ scenarios that are outside the reach of $0\nu\beta\beta$ experiments and direct searches for $W_R$ at colliders. The sensitivity of such modes are summarized in Fig. \[fig:majoranaHiggs\] [@Maiezza:2015lza; @Nemevsek:2016enw]. The associated production channels, $$p p \to H^{0,\pm\pm}~W_R^\mp \quad\text{and}\quad pp\to H^0 Z_R,$$ are also possible. However, as in the SM, these channels are $s$-channel and phase space suppressed, which lead to prohibitively small cross sections in light of present mass limits [@Bambhaniya:2013wza]. Lastly, one should note that the search for such Higgs decays is not limited to hadron colliders. As presently designed future lepton colliders are aimed at operating as Higgs factories, searches for such $L$-violating Higgs decays [@Atwood:2007zza; @Ren:2008sz; @Yue:2010zzb] at such facilities represent an attractive discovery prospect. In this context, a relatively understudied topic is the possible manifestation of Seesaw in precision measurements of the known SM-like Higgs boson [@BhupalDev:2012zg; @Maiezza:2015lza; @Banks:2008xt]. Some related studies also exist in the literature such as for generic pheno [@Han:2007bk; @delAguila:2013mia; @Han:2007bk]; for little Higgs [@Hektor:2007uu; @Han:2005nk]; and for decay ratios and mixing patterns of exotically charged Higgs[@Kadastik:2007yd; @Chun:2003ej]. The Type III Seesaw and Lepton Number Violation at Colliders {#sec:type3} ============================================================ We now turn to collider searches for lepton number violation in the context of the Type III Seesaw mechanism [@Foot:1988aq] as well as its embedding in GUTs and other SM extensions. In some sense, the Type III model is the fermionic version of the Type II scenario, namely that Seesaw partner fermions couple to the SM via both weak gauge and Yukawa couplings. Subsequently, much of the Type III collider phenomenology resembles that of Type I-based models. However, quantitatively, the presence of gauge couplings lead to a very different outlook and level of sensitivity. We now summarize the main highlights of the canonical Type III Seesaw (Sec. \[sec:type3Canon\]), Type III-based models (Sec. \[sec:type3Extend\]), and then review their $L$-violating collider phenomenology (Sec. \[sec:type3Collider\]). As with the previous Seesaw scenarios, a discussion of cLFV is outside the scope of this review. For recent summaries on cLFV in the Type III Seesaw, see Refs. [@Abada:2007ux; @Abada:2008ea; @Eboli:2011ia; @Agostinho:2017biv] and references therein. Type III Seesaw Models {#sec:type3Theory} ---------------------- ### The Canonical Type III Seesaw Mechanism {#sec:type3Canon} In addition to the SM field content, the Type III Seesaw [@Foot:1988aq] consists of SU$(2)_L$ triplet (adjoint) leptons, $$\begin{aligned} &&\Sigma_L= \Sigma^a_L \sigma^a = \left( \begin{array}{cc} \Sigma_L^0/\sqrt{2} & \Sigma^+_L \\ \Sigma_L^- & -\Sigma_L^0/\sqrt{2} \\ \end{array} \right) \ , \quad \Sigma^\pm_L \equiv \frac{\Sigma^1_L \mp i\Sigma^2_L}{\sqrt{2}}, \quad \Sigma_L^0=\Sigma_L^3,\end{aligned}$$ which transform as $(1,3,0)$ under the SM gauge group. Here $\Sigma^\pm_L$ have U$(1)_{\rm EM}$ charges $Q=\pm1$, and the $\sigma^{a}$ for $a=1,\dots,3$, are the usual Pauli SU$(2)$ matrices. The RH conjugate fields are related by $$\Sigma_R^c = \begin{pmatrix} \Sigma^{0 c}_R/\sqrt{2} & \Sigma^{-c}_R \\ \Sigma^{+c} & -\Sigma^{0 c}_R/\sqrt{2} \end{pmatrix}, \quad\text{for}\quad \psi_R^c \equiv (\psi^c)_R = (\psi_L)^c.$$ The Type III Lagrangian is given by the sum of the SM Lagrangian, the triplet’s kinetic and mass terms, $$\mathcal{L}_{T} = \frac{1}{2}{\mathop{\rm Tr}}\left[\overline{\Sigma_L}i\not\!\!D\Sigma_L\right] -\left(\frac{M_\Sigma}{2}\overline{\Sigma_L^0}\Sigma_R^{0c} + M_\Sigma\overline{\Sigma^-_L}\Sigma^{+c}_R + \text{H.c.}\right), \label{eq:typeIIIMassLag}$$ and the triplet’s Yukawa coupling to the SM LH lepton $(L)$ and Higgs $(H)$ doublet fields, $$\mathcal{L}_Y = -Y_\Sigma \overline{L} ~\Sigma_R^c ~i\sigma^2 H^* + \text{H.c.} \label{eq:typeIIIYukawa}$$ From Eq. (\[eq:typeIIIYukawa\]), one can deduce the emergence of a Yukawa coupling between the charged SM leptons and the charged triplet leptons. This, in turn, induces a mass mixing among charged leptons that is similar to doublet-singlet and doublet-triplet neutrino mass mixing, and represents one of the more remarkable features of the Type III mechanism. The impact of EW fermion triplets on the SM Higgs, naturalness in the context of the Type III Seesaw has been discussed in Refs. [@Gogoladze:2008ak; @He:2012ub; @Gogoladze:2010in]. After expanding Eqs. (\[eq:typeIIIMassLag\])-(\[eq:typeIIIYukawa\]), the relevant charged lepton and neutrino mass terms are [@Li:2009mw] $$\begin{aligned} \mathcal{L}^m_{\rm III}= - &\left( \begin{array}{cc} \overline{l_R} & \overline{\Psi_R} \\ \end{array} \right) \left( \begin{array}{cc} m_l & 0 \\ Y_\Sigma v_0 & M_\Sigma \\ \end{array} \right) \left( \begin{array}{c} l_L \\ \Psi_L \\ \end{array} \right) \nonumber\\ & -\left( \begin{array}{cc} \overline{\nu_L^c} & \overline{\Sigma_L^{0c}} \\ \end{array} \right) \left( \begin{array}{cc} 0 & Y_\Sigma^Tv_0/2\sqrt{2} \\ Y_\Sigma v_0/2\sqrt{2} & M_\Sigma/2 \\ \end{array} \right) \left( \begin{array}{c} \nu_L \\ \Sigma^0_L \\ \end{array} \right) + \text{H.c.}, \label{mass-matrix}\end{aligned}$$ with $\Psi_L\equiv\Sigma_L^-$, $\Psi_R\equiv \Sigma_L^{+c}$, and $\Psi = \Psi_L + \Psi_R$. After introducing unitarity matrices to transit light doublet and heavy triplet lepton fields as below $$\begin{aligned} &&\left( \begin{array}{c} l_{L,R} \\ \Psi_{L,R} \\ \end{array} \right) =U_{L,R} \left( \begin{array}{c} l_{mL,R} \\ \Psi_{mL,R} \\ \end{array} \right), \ \ \ \left( \begin{array}{c} \nu_{L} \\ \Sigma^0_{L} \\ \end{array} \right) =U_0 \left( \begin{array}{c} \nu_{mL} \\ \Sigma^0_{mL} \\ \end{array} \right),\\ &&U_L\equiv \left( \begin{array}{cc} U_{Lll} & U_{Ll\Psi} \\ U_{L\Psi l} & U_{L\Psi\Psi} \\ \end{array} \right), \ \ U_R\equiv \left( \begin{array}{cc} U_{Rll} & U_{Rl\Psi} \\ U_{R\Psi l} & U_{R\Psi\Psi} \\ \end{array} \right), \ \ U_0\equiv \left( \begin{array}{cc} U_{0\nu\nu} & U_{0\nu\Sigma} \\ U_{0\Sigma \nu} & U_{0\Sigma\Sigma} \\ \end{array} \right),\end{aligned}$$ one obtains the diagonal mass matrices and mass eigenvalues for neutrinos and charged leptons, $$\begin{aligned} {\rm diag}(\mathcal{N}) = U_0^\dagger \left( \begin{array}{cc} 0 & Y_\Sigma^\dagger v_0/\sqrt{2} \\ Y_\Sigma^\ast v_0/\sqrt{2} & M_\Sigma^\ast \\ \end{array} \right) U_0^\ast = \left( \begin{array}{cc} m_\nu^{diag} & 0 \\ 0 & M_{N}^{diag} \\ \end{array} \right), \\ {\rm diag}(\mathcal{E}) = \ U_L^\dagger \left( \begin{array}{cc} m_l^\dagger & Y_\Sigma^\dagger v_0 \\ 0 & M_\Sigma^\dagger\\ \end{array} \right) U_R = \left( \begin{array}{cc} m_l^{diag} & 0 \\ 0 & M_{E}^{diag} \\ \end{array} \right).\end{aligned}$$ The light neutrino mass eigenstates are denoted by $\nu_{j}$ for $j=1,\dots,3$; whereas the heavy neutral and charged leptons are respectively given by $N_{j'}$ and $E^\pm_{k'}$. In the literature, $N$ and $E^\pm$ are often denoted as $T^0,~T^\pm$ or $\Sigma^0,~\Sigma^\pm$. However, there is no standard convention as to what set of symbols are used to denote gauge and mass eigenstates. Where possible, we follow the convention of Ref.  [@Arhrib:2009mz] and generically denote triplet-doublet mixing by $Y_T$ and $\varepsilon_T$. This means that in the mass basis, triplet gauge states are given by $$\begin{aligned} \Psi^\pm = Y_T~E^\pm + \sqrt{2} \varepsilon_T~ \ell^\pm & \quad\text{and}\quad & \Psi^0 = Y_T~N + \varepsilon_T~ \nu_m, \nonumber\nonumber\\ \text{with}\quad \vert Y_T\vert\sim \mathcal{O}(1) & \quad\text{and}\quad & \vert\varepsilon_T\vert\sim {Y_\Sigma v_0\over \sqrt{2}M_\Sigma}\ll1.\end{aligned}$$ The resulting interaction Lagrangian, in the mass eigenbasis then contains [@Arhrib:2009mz] $$\begin{aligned} \mathcal{L}_{\rm Type~III}^{\rm Mass~Basis} &\ni& - \overline{E^-_{k'}}\left( e Y_T A_\mu\gamma^\mu + g\cos\theta_W Y_T Z_\mu\gamma^\mu \right)E^-_{k'} -g Y_T \overline{E^-_{k'}} W^-_\mu\gamma^\mu N_{j'} \nonumber \\ &-& \frac{e}{2s_wc_w}Z_\mu \left(\varepsilon_T \overline{N_{j'}} \gamma^\mu P_R \nu_j +\sqrt{2} \varepsilon_T \overline{E^-_{k'}} \gamma^\mu P_R \ell^-_k \right) \nonumber\\ &-& \frac{e}{s_w}W_\mu^+ \left(\varepsilon_T \overline{\nu_j} \gamma^\mu P_L E^-_{k'} + \frac{1}{\sqrt{2}} \varepsilon_T \overline{N_{j'}} \gamma^\mu P_R \ell^-_k \right) +{\rm H.c.} \label{int-typeiii}\end{aligned}$$ From this, one sees a second key feature of the Type III Seesaw, that gauge interactions between heavy lepton pairs proceeds largely through pure vector currents with axial-vector deviations (not shown) suppressed by $\mathcal{O}(\varepsilon_T^2)$ at the Lagrangian level. This follows from the triplet fermions vector-like nature. Similarly, the mixing-suppressed gauge couplings between heavy and light leptons proceeds through SM-like currents. Explicitly, the light and heavy neutrino mass eigenvalues are $$\begin{aligned} m_\nu\approx {Y_\Sigma^2v_0^2\over 2M_\Sigma}, \ \ \ M_{N}\approx M_\Sigma,\end{aligned}$$ and for the charged leptons are $$\begin{aligned} m_l-m_l{Y_\Sigma^2 v_0^2\over 2M_\Sigma^2}\approx m_l, \ \ \ M_E\approx M_\Sigma.\end{aligned}$$ This slight deviation in the light, charged leptons’ mass eigenvalues implies a similar variation in the anticipated Higgs coupling to the same charged leptons. At tree-level, the heavy leptons $N$ and $E^\pm$ are degenerate in mass, a relic of SU$(2)_L$ gauge invariance. However, after EWSB, and for $M_{\Sigma} \gtrsim 100$ GeV, radiative corrections split this degeneracy by [@Arhrib:2009mz], $$\begin{aligned} && \Delta M_T\equiv M_{E}-M_{N}=\frac{\alpha_W}{2\pi}\frac{M_W^2}{M_\Sigma} \left[f\left(\frac{M_\Sigma}{M_Z}\right)-f\left(\frac{M_\Sigma}{M_W}\right)\right] \approx 160{{\rm ~MeV}}, \\ && \text{where}\quad~f\left(y\right)=\frac{1}{4y^2}\log{y^2}- \left(1+\frac{1}{2y^2}\right) \sqrt{4y^2-1}\arctan{\sqrt{4y^2-1}},\end{aligned}$$ and opens the $E^\pm \to N \pi^\pm$ decay mode. Beyond this are the heavy lepton decays to EW bosons and light leptons that proceed through doublet-triplet lepton mixing. The mixings are governed by the elements in the unitary matrices $U_{L,R}$ and $U_0$. Expanding $U_{L,R}$ and $U_0$ up to order $Y_\Sigma^2v_0^2M_\Sigma^{-2}$, one gets the following results [@Abada:2008ea; @He:2009tf] $$\begin{aligned} &&U_{Lll} = 1- \epsilon\;,\;\;U_{Ll\Psi} = Y^\dagger_\Sigma M^{-1}_\Sigma v_0\;,\;\;\;\;\;\;\; U_{L\Psi l} = - M^{-1}_\Sigma Y_\Sigma v_0\;,\;\;\;\;\;\;U_{L\Psi\Psi} = 1-\epsilon'\;,\nonumber\\ &&U_{Rll} = 1\;,\;\;\;\;\;\;\;\;U_{Rl\Psi} = m_l Y^\dagger_\Sigma M^{-2}_\Sigma v_0\;,\;\;\; U_{R\Psi l} = - M^{-2}_\Sigma Y_\Sigma m_l v_0\;,\;\;U_{R\Psi\Psi} = 1\;,\nonumber\\ &&U_{0\nu\nu} = (1- \epsilon/2)U_{PMNS}\;,\;\; U_{0\nu \Sigma} = Y^\dagger_\Sigma M^{-1}_\Sigma v_0/\sqrt{2}\;,\;\; U_{0\Sigma \nu} = - M^{-1}_\Sigma Y_\Sigma U_{0\nu\nu} v_0/\sqrt{2}\;,\nonumber\\ &&U_{0\Sigma\Sigma } = 1-\epsilon'/2\;,\;\;\;\;\;\;\;\; \epsilon = Y^\dagger_\Sigma M^{-2}_\Sigma Y_\Sigma v_0^2/2\;,\;\;\;\; \epsilon' = M^{-1}_\Sigma Y_\Sigma Y^\dagger_\Sigma M^{-1}_\Sigma v_0^2/2\;.\nonumber \label{approx}\end{aligned}$$ To the order of $Y_\Sigma v_0M_\Sigma^{-1}$, the mixing between the SM charged leptons and triplet leptons, , $V_{\ell N}=-Y_\Sigma^\dagger v_0 M_\Sigma^{-1}/\sqrt{2}$, follows the same relation as Eq. (\[typei\]) in the Type I Seesaw [@Li:2009mw] and the couplings in the interactions in Eq. (\[int-typeiii\]) are all given by $V_{\ell N}$ [@AguilarSaavedra:2009ik; @Li:2009mw]. ![Triplet decay widths as function of the triplet mass and assuming $M_{h_{\rm SM}}=115{{\rm ~GeV}}$ [@Franceschini:2008pz]. []{data-label="fig:type3Width"}](type3_TripletLeptonWidth_0805_1613.pdf){width=".95\textwidth"} Hence, the partial widths for both the heavy charged lepton and heavy neutrino are proportional to $|V_{\ell N}|^2$. For $M_E\approx M_N\gg M_W, M_Z, M_{h}$, the partial widths behave like [@delAguila:2008cj; @AguilarSaavedra:2009ik] $$\begin{aligned} &&{1\over 2}\Gamma(N\to \sum_\ell \ell^+ W^-+\ell^-W^+)\approx \Gamma(N\to \sum_\nu\nu Z+\bar{\nu}Z)\approx \Gamma(N\to \sum_\nu \nu h+\bar{\nu}h)\nonumber \\ &\approx &{1\over 2} \Gamma(E^\pm\to \sum_\nu \overset{(-)}\nu W^\pm)\approx \Gamma(E^\pm\to \sum_\ell \ell^\pm Z)\approx \Gamma(E^\pm\to \sum_\ell \ell^\pm h) \nonumber\\ &\approx & {G_F\over 8\sqrt{2}\pi}\sum_\ell |V_{\ell N}|^2 M_\Sigma^3.\end{aligned}$$ Thus the heavy lepton branching ratios exhibit asymptotic behavior consistent with the Goldstone Equivalence Theorem [@Chanowitz:1985hj; @Lee:1977yc], and are given by the relations [@Franceschini:2008pz; @delAguila:2008cj; @AguilarSaavedra:2009ik; @Arhrib:2009mz], $$\begin{aligned} &&{1\over 2}{\rm BR}(N\to \sum_\ell \ell^+ W^-+\ell^-W^+)\approx {\rm BR}(N\to \sum_\nu\nu Z+\bar{\nu}Z)\approx {\rm BR}(N\to \sum_\nu \nu h+\bar{\nu}h)\nonumber \\ &\approx &{1\over 2} {\rm BR}(E^\pm\to \sum_\nu \overset{(-)}\nu W^\pm)\approx {\rm BR}(E^\pm\to \sum_\ell \ell^\pm Z)\approx {\rm BR}(E^\pm\to \sum_\ell \ell^\pm h) \approx {1\over 4}.\quad\end{aligned}$$ As displayed in Fig. \[fig:type3Width\] by Ref. [@Franceschini:2008pz], as the triplet mass grows, this asymptotic behavior can be seen explicitly in the triplet lepton partial widths. ### Type I+III Hybrid Seesaw in Grand Unified and Extended Gauge Theory {#sec:type3Extend} One plausible possibility to rescue the minimal grand unified theory, , SU$(5)$, is to introduce an adjoint $24_F$ fermion multiplet in addition to the original $10_F$ and $\bar{5}_F$ fermionic representations [@Ma:1998dn; @Bajc:2006ia]. As the $24_F$ contains both singlet and triplet fermions in this non-supersymmetric SU$(5)$, the SM gauge couplings unify and neutrino masses can generated through a hybridization of the Types I and III Seesaw mechanisms. The Yukawa interactions and Majorana masses in this Type I+III Seesaw read [@Arhrib:2009mz] $$\begin{aligned} \Delta\mathcal{L}_{\rm I+III}^Y=Y_S LHS + Y_T LHT-{M_S\over 2} SS - {M_T\over 2} TT +{\rm H.c.},\end{aligned}$$ where $S$ and $T=\left({T^-+T^+\over \sqrt{2}},{T^--T^+\over i\sqrt{2}},T^0\right)$ are the fermionic singlet and triplet fields, respectively, with masses $M_S$ and $M_T$. In the limit that $M_{S},M_{T}\gg Y_S v_0, Y_Tv_0$, the light neutrino masses are then given by the sum of the individual Type I and III contributions $$\begin{aligned} m_\nu=- (Y_S v_0/\sqrt{2})^2 M_S^{-1} - (Y_Tv_0/\sqrt{2})^2 M_T^{-1},\end{aligned}$$ The most remarkable prediction of this SU$(5)$ theory is that the unification constraint and the stability of proton require the triplet mass to be small: $M_T\lesssim 1$ TeV [@Bajc:2006ia; @Dorsner:2006fx]. Thus, in SU$(5)$ scenarios, the triplet leptons of this Type I+III Seesaw are within the LHC’s kinematic reach and can be tested via $L$-violating collider signatures [@Ma:1998dn; @Bajc:2007zf; @Perez:2007rm; @AristizabalSierra:2010mv; @Aguilar-Saavedra:2013twa; @Biggio:2010me]. Other GUT models that can accommodate the Type III Seesaw and potentially lead to collider-scale $L$-violation include variations of SO$(10)$ [@Chakrabortty:2010az] theories. It is also possible to embed the Type III scenario into extended gauge sectors, including Left-Right Symmetric theories [@FileviezPerez:2008sr; @Duerr:2013opa; @Gu:2011yx; @Mohapatra:2009fj], which also represents a Type I+II+III hybrid Seesaw hat trick. Additionally, Type III-based hybrid Seesaws can be triggered via fermions in other SU$(2)_L\times$U$(1)_Y$ representations [@Delgado:2011iz; @Ma:2013tda; @Yu:2015pwa; @Nomura:2017abu], The collider phenomenology in many of these cases is very comparable to that of the Type I and II Seesaws, as discussed in Secs. \[sec:type1\] and  \[sec:type2\], or the more traditional Type III scenario, which we now discuss. Heavy Charged Leptons and Neutrinos at Colliders {#sec:type3Collider} ------------------------------------------------ ### Heavy Charged Leptons and Neutrinos at $pp$ Colliders ![ Born level production of Type III lepton pairs via (a) Drell-Yan, (b) gluon fusion, and (c) photon fusion.[]{data-label="fig:feynman_Type3_MultiProd"}](feynman_TypeIII_MultiProd.pdf){width="95.00000%"} Due to the presence of both gauge and Yukawa couplings to SM fields, the collider phenomenology for triplet leptons is exceedingly rich. In hadron collisions, for example, pairs of heavy triplet leptons are produced dominantly via charged and neutral Drell-Yan (DY) currents, given by $$\begin{aligned} q\bar{q}'\to W^{\ast\pm}\to T^\pm T^0 \quad\text{and}\quad q\bar{q} \to \gamma^\ast/Z^\ast \to T^+ T^-,\end{aligned}$$ and shown in Fig. \[fig:feynman\_Type3\_MultiProd\](a). For the DY process, the total cross section is now known up to NLO and differentially at NLO+LL in $k_T$ resummation [@Ruiz:2015zca]. As function of mass, the $N\ell^\pm$ (singlet) as well as $T^+T^-$ and $T^\pm T^0$ (triplet) DY production cross sections at $\sqrt{s}=14$ and 100 TeV are displayed in Fig. \[fig:SeesawTripletXSecMass\]. While the three rates are naïvely comparable, one should assign a mixing factor of $\vert V_{\ell N}\vert^2 \lesssim 10^{-2}$ to the singlet production since it proceeds through active-sterile neutrino mixing, , Yukawa couplings, whereas triplet lepton pair production proceeds through gauge couplings. Heavy triplet leptons can also be produced singly in the association with light leptons and neutrinos, $$\begin{aligned} q\bar{q}'\to W^{\ast\pm}\to T^\pm \nu, ~T^0\ell^\pm \quad\text{and}\quad q\bar{q} \to \gamma^\ast/Z^\ast \to T^\pm \ell^\mp.\end{aligned}$$ As single production modes are proportional to the small [@delAguila:2008pw] doublet-triplet mixing, denoted by $|V_{\ell T}|$, these processes suffer from the same small signal rates at colliders as does singlet production in Type I-based Seesaws (see Sec. \[sec:type1Canon\]). However, as heavy-light lepton vertices also posses axial-vector contributions, new production channels are present, such as the gluon fusion mechanism [@Willenbrock:1985tj; @Dicus:1991wj; @Hessler:2014ssa; @Ruiz:2017yyf], shown in Fig. \[fig:feynman\_Type3\_MultiProd\](b) and given by $$\begin{aligned} g g \to Z^* / h^* \to T^\pm \ell^\mp. \label{eq:type3GF}\end{aligned}$$ It is noteworthy that the partonic expression for gluon fusion channels $gg \to Z^*/h^* \to T^\pm \ell^\mp$ is equal to the Type I analogue $gg\to N \nu_\ell$ [@Hessler:2014ssa], and hence so are its QCD corrections [@Ruiz:2017yyf]. Conversely, heavy triplet pair production through gluon fusion, , $gg \to T\overline{T}$, is zero since their couplings to weak bosons are vector-like, and hence vanish according to Furry’s Theorem [@Willenbrock:1985tj; @delAguila:1990yw; @Dicus:1991wj]. For $\sqrt{s} = 7-100$ TeV, the N$^3$LL(Threshold) corrections to the Born rates of Eq. (\[eq:type3GF\]) span $+160\%$ to $+260\%$ [@Ruiz:2017yyf]. Hence, for singly produced triplet leptons, the gluon fusion mechanism is dominant over the DY channel for $\sqrt{s}\gtrsim 20-25$ TeV, over a wide range of EW- and TeV-scale triplet masses [@Hessler:2014ssa; @Ruiz:2017yyf]. More exotic production channels also exist, such as the $\gamma\gamma \to T^+ T^-$ VBF channel, shown in Fig. \[fig:feynman\_Type3\_MultiProd\](c), as well as permutations involving $W$ and $Z$. However, the $\gamma\gamma$ contributions is sub-leading due to coupling and phase space suppression. For representative heavy lepton masses of $M_T = 500$ GeV and 1 TeV as well as doublet-triplet mixing of $|V_{\ell T}|^2 = 10^{-2}$, we display in Fig. \[fig:SeesawTripletXSecBeam\] the $pp\to T^+T^-$ and $T^\pm\ell^\mp$ production cross sections via various hadronic production mechanisms as a function of collider energy $\sqrt{s}$. In the figure, the dominance of pair production over single production is unambiguous. Interestingly, considering that the triplet mass splitting is $\Delta M_T \sim \mathcal{O}(200)$ MeV as stated above, one should not expect to discover the neutral current single production mode without also observing the charged channel almost simultaneously. Hence, despite sharing much common phenomenology, experimentally differentiating a Type I scenario from a Type III (or I+III) scenario is straightforward. Leading order-accurate Monte Carlo simulations for tree-level processes involving Type III leptons are possible with the `Type III Seesaw` FeynRules UFO model [@Biggio:2011ja; @Abada:2008ea; @FeynRules:TypeIII], as well as a Minimal Lepton Flavor Violation variant `MLFV Type III Seesaw` [@Eboli:2011ia; @Agostinho:2017biv; @FeynRules:Type3mlfv]. The models can be ported into modern, general-purpose event generators, such at Herwig [@Bellm:2015jjp], MadGraph5\_aMC@NLO [@Alwall:2014hca], and Sherpa [@Gleisberg:2008ta]. \ Hadron collider tests of the Type III Seesaw can be categorized according to the final-state lepton multiplicities, which include: the $L$-violating, same-sign dilepton and jets final state, $\ell_1^\pm \ell_2^\pm + nj$ [@Biggio:2011ja; @Bajc:2006ia; @Franceschini:2008pz; @Li:2009mw; @delAguila:2008cj; @Arhrib:2009mz; @AguilarSaavedra:2009ik; @Choubey:2009wp]; the four-lepton final state, $\ell_1^\pm \ell_2^\pm \ell_3^\mp \ell_4^\mp + nj$ [@Biggio:2011ja; @Li:2009mw; @delAguila:2008cj; @AguilarSaavedra:2009ik; @Franceschini:2008pz]; other charged lepton multiplicities [@Biggio:2011ja; @delAguila:2008cj; @AguilarSaavedra:2009ik; @Franceschini:2008pz; @Bandyopadhyay:2011aa]; and also displaced charged lepton vertices [@Franceschini:2008pz; @He:2009ua]. Other “displaced” signatures, include triplet lepton decays to displaced Higgs bosons [@Bandyopadhyay:2010wp]. Direct searches for Type III Seesaw partners at the $\sqrt{s}=7/8$ TeV [@CMS:2012ra; @Aad:2015cxa; @Aad:2015dha] and $\sqrt{s}=13$ TeV [@CMS:2016hmk; @CMS:2017wua; @Sirunyan:2017qkz] LHC have yet to show evidence of heavy leptons. As shown in Fig. \[fig:type3Searches\] (a), triplet masses below $M_T\lesssim 800$ GeV have been excluded at 95% CLs [@CMS:2017wua]. Figure \[fig:type3Searches\] (b) displays the discovery potential of triplet leptons at high-luminosity 100 TeV collider. One can discover triplet lepton as heavy as 4 (6.5) TeV with 300 (3000) fb$^{-1}$ integrated luminosity. The absence of triplet leptons in multi-lepton final states can also be interpreted as a constrain on doublet-triplet neutrino mixing. In Fig. \[fig:type3Searches\](c,d), one sees the exclusion contours of doublet-triplet neutrino mixing in $\vert V_{\mu N}\vert-\vert V_{eN}\vert$ and $\vert V_{\tau N}\vert-\vert V_{eN}\vert$ spaces after $\mathcal{L}=4.9$ fb$^{-1}$ of data at CMS (labels denote heavy neutral lepton mass in GeV) [@Aguilar-Saavedra:2013twa]. ### Heavy Charged Leptons and Neutrinos at $ee$ and $ep$ Colliders The triplet leptons can also be produced at the leptonic colliders like the ILC and the Compact Linear Collider (CLIC) [@Arhrib:2009mz; @Goswami:2017jqs], and the electron-hadron collider like LHeC [@Liang:2010gm]. Besides the similar s-channels as hadron colliders, at $e^+e^-$ colliders, the triplet lepton single and pair productions can also happen in $t$-channel via the exchange of $h,~W$, or $Z$ boson. Triplet leptons can also lead to anomalous pair production of SM weak bosons [@Yue:2010zzb]. Assuming $M_\Sigma=500$ GeV and $V_{eN}=0.05$, the cross sections of triplet lepton single and pair productions are shown in Fig. \[typeiii-eexsec\] (a). For the single production at 1 TeV $e^+e^-$ collider, the triplet lepton with mass up to about 950-980 GeV can be reached with 300 fb$^{-1}$. To discover the heavy charged lepton through $e^+e^-\to \Sigma^+\Sigma^-$ production at $\sqrt{s}=2$ TeV, the luminosity as low (high) as 60 (480) fb$^{-1}$ is needed as shown in Fig. \[typeiii-eexsec\] (b). Radiative Neutrino Mass Models and Lepton Number Violation at Colliders {#sec:loop} ======================================================================= A common feature of the Seesaw mechanisms discussed in the previous sessions is that they are all tree-level, UV completion of the dimension-5 Weinberg operator in of Eq. (\[Weinberg\]). Though economical and elegant, these models often imply subtle balancing between a Seesaw mass scale at a TeV or below and small Yukawa couplings, in the hope for them to be observable at current and near future experiments. In an altogether different paradigm, it may be the case that small neutrino masses are instead generated radiatively. In [*radiative neutrino mass models*]{}, loop and (heavy) mass factors can contribute to the suppression of light neutrino masses and partly explain their smallness. A key feature of radiative neutrino mass models is that the Weinberg operator is not generated at tree-level: For some models, this may be because the particles required to generate tree-level masses, , SM singlet fermions in Type I, triplet scalars in Type II, or triplet leptons in Type III, do not exist in the theory. For others, it may be the case that the required couplings are forbidden by new symmetries. Whatever the case, it is necessary that the new field multiplets run in the loops to generate neutrino masses. At one-loop, such models were first proposed in Refs. [@Zee:1980ai; @Hall:1983id], at two-loop in Refs. [@Cheng:1980qt; @Zee:1985id; @Babu:1988ki], and more recently at three-loop order in Ref. [@Krauss:2002px]. Besides these early works, a plethora of radiative mass models exist due to the relative ease with which unique loop topologies can be constructed at a given loop order, as well as the feasibility to accommodate loop contributions from various exotic particles, including leptoquarks, vector-like leptons and quarks, electrically charged scalars, and EW multiplets. For a recent, comprehensive review, see Ref. [@Cai:2017jrq]. However, the diversity of the exotic particles and interactions in radiative neutrino mass models make it neither feasible nor pragmatic to develop a simple and unique strategy to test these theories at colliders. Although some effort has been made to advance approaches to collider tests of radiative neutrino mass models more systematically [@Cai:2014kra; @AristizabalSierra:2007nf], it remains largely model-dependent. As a comprehensive summary of the literature for radiative neutrino mass models and their collider study is beyond the scope of this review, in this section, we focus on a small number of representative models with distinctive $L$-violating collider signatures. It is worth pointing out that some popular radiative neutrino mass models do not predict clear lepton number violation at collider scales. A prime example are the Scotogenic models [@Ma:2006km], a class of one-loop radiative neutrino mass scenario with a discrete $Z_2$ symmetry. Scotogenic models typically contain three SM singlet fermions $N_i$ with Majorana masses and are odd under the $Z_2$, whereas SM fields are even. The discrete symmetry forbids the mixing between the SM neutrinos and $N_i$ that one needs to trigger the Type I and III Seesaw mechanisms. As a result, collider strategies to search for lepton number violation mediated by heavy Majorana neutrinos as presented in Sec. \[sec:type1\] are not applicable to the Scotogenic model. Instead, collider tests of Scotogenic models include, for example, searches for the additional EW scalars [@Ho:2013hia; @Ho:2013spa; @Hessler:2016kwm; @Diaz:2016udz] that facilitate lepton number conserving processes. Subsequently, we avoid further discussing radiative models without collider-scale lepton number violation. Like in the previous sections, we first present in Sec. \[sec:radModels\] an overview of representative radiative models. Then, in Sec. \[sec:radAtColliders\], we review collider searches for lepton number violation associated with radiative neutrino mass models. Selected Radiative Neutrino Mass Models {#sec:radModels} --------------------------------------- ### The Zee-Babu Model The first radiative scenario we consider is the well-known Zee-Babu model, a two-loop radiative neutrino mass model proposed independently by Zee [@Zee:1985id] and Babu [@Babu:1988ki]. In the model, the SM field content is extended by including one singly-charged scalar ($h^\pm$) and one doubly-charged scalar ($k^{\pm\pm}$). Both scalars are singlets under $SU(3)_c\times SU(2)_L$, leading to the lepton number violating interaction Lagrangian $$\begin{aligned} \Delta \mathcal{L} = \bar{L} Y^\dagger e_R H + \bar{\tilde{L}} f L h^+ + \overline{e_R^c} g e_R k^{++} + \mu_{ZB} h^+h^+ k^{--} + {\rm H.c.},\end{aligned}$$ where $L~(H)$ is the SM LH lepton (Higgs) doublet. The $3\times 3$ Yukawa coupling matrices $f$ and $g$ are anti-symmetric and symmetric, respectively. The trilinear coupling $\mu_{ZB}$ contributes to the masses of the charged scalars at the loop level. For large values of $(\mu_{ZB}/m_{h^\pm})$ or $(\mu_{ZB}/m_{k^{\pm\pm}})$, where $m_{h^\pm,k^{\pm\pm}}$ are the masses of $h^\pm$ and $k^{\pm\pm}$, the scalar potential may have QED-breaking minima. This can be avoided by imposing the condition $\left| \mu_{ZB}\right| \ll 4\pi \; {\rm min}(m_h, m_k)$. The combined presence of $Y$, $f$, $g$ and $\mu_{ZB}$ collectively break lepton number and lead to the generation of a small Majorana neutrino mass. At lowest order, neutrino masses in the Zee-Babu model arise at two-loop order, as depicted in Fig. \[fig:zeebabu\]. The resulting neutrino mass matrix scales as $$\begin{aligned} \mathcal{M}_{\nu} \simeq \left(\frac{v^2 \mu_{ZB}}{96\pi^2 M^2}\right)~ f Y g^\dagger Y^T f^T \; ,\end{aligned}$$ where $M=\max(m_{h^\pm}, m_{k^{\pm\pm}})$ is the heaviest mass in the loop. Since $f$ is antisymmetric, the determinant of the neutrino mass matrix vanishes, $\det \mathcal{M}_\nu = 0$. Therefore the Zee-Babu models yields at least one exactly massless neutrino. An important consequence is that the heaviest neutrino mass is determined by the atmospheric mass difference, which can be estimated as $$\begin{aligned} m_\nu \approx 6.6\times 10^{-3} f^2 g \left(\frac{m_\tau^2}{M}\right) \approx 0.05~{\rm eV} \; , \label{eq:ZBRadMass}\end{aligned}$$ where $m_\tau\approx1.778$ GeV is the tau lepton mass. This implies the product $f^2 g$ can not be arbitrarily small, , for $M\sim 100~{\rm GeV}$, one finds $g^2 f\gtrsim 10^{-7}$. Subsequently, the parameter space of the Zee-Babu model is constrained by both neutrino oscillation data, low-energy experiments such as decays mediated $k^{\pm\pm}$ at tree level, and high-energy searches for direct pair production of $k^{\pm\pm}$. The study of $h^\pm$ is mostly similar to that of the singly-charged scalar in the Zee model [@Zee:1980ai], although the lepton number violating effects are not experimentally observable due to the missing information carried away by the light (Majorana) neutrino in the decay product. The doubly-charged scalar $k^{\pm\pm}$ can decay to a pair of same-sign leptons, which manifestly violates lepton number by $\Delta L = \pm2$, with a partial decay width given by $$\begin{aligned} \Gamma(k^{\pm\pm}\to \ell_a^\pm \ell_b^\pm) = \frac{\left|g_{ab}\right|^2}{4\pi (1+\delta_{ab})} m_k \;.\end{aligned}$$ If $m_{k^{\pm\pm}}> 2 m_{h^\pm},$ then the $k^{\pm\pm}\to h^\pm h^\pm$ decay mode opens with a partial decay width of $$\Gamma(k^{\pm\pm}\to h^\pm h^\pm) =\frac{m_{k^{\pm\pm}}}{8\pi} \left(\frac{\mu_{ZB}}{m_{k^{\pm\pm}}}\right)^2 \sqrt{1-\frac{4 m_{h^{\pm}}^2 }{m_{k^{\pm\pm}}^2}} \; .$$ Doubly-charged scalars, appear in many other radiative neutrino mass models, including the three-loop Cocktail Model [@Gustafsson:2012vj], whose eponymous mass-generating diagram is shown in the right panel of Fig. \[fig:radiative\]. The doubly-charged scalar couples to the SM lepton doublet and a singly-charged scalar in the same manner as in the Zee-Babu model, and thus again is similar to a Type II scenario. Radiative Type II Seesaw model [@Ma:2015xla] that generates neutrino mass at one-loop order contains an $SU(2)_L$ triplet scalar and thus also has similar LHC phenomenology as the tree-level Type II Seesaw mechanism [@Guo:2016dzl]. ### The Colored Zee-Babu Model with Leptoquark ![Feynman diagram for the generation of neutrino masses at two-loop order in the colored Zee-Babu model [@Kohda:2012sr]. []{data-label="fig:cZBM"}](cZBM.pdf){width="50.00000%"} In a particularly interesting variant of the Zee-Babu model, proposed in Ref. [@Kohda:2012sr], all particles in the neutrino mass-loop are charged under QCD. As shown in Fig. \[fig:cZBM\], the lepton doublet in the loop of the Zee-Babu model is replaced with down-type quark while the singly- and doubly-charged scalars are replaced with a leptoquark $S_{LQ}^{-\frac{1}{3}}$ and a diquark $S_{DQ}^{-\frac{2}{3}}$. Under the SM gauge group, the leptoquark and diquark quantum numbers are $$\begin{aligned} S_{LQ}^{-\frac{1}{3}} \quad:\quad (3,1,-\frac{1}{3}) \quad\text{and}\quad S_{DQ}^{-\frac{2}{3}} \quad:\quad(6, 1, -\frac{2}{3}) \; .\end{aligned}$$ The decay of the diquark $S_{DQ}^{-\frac{2}{3}}$ is analogous to that of the doubly-charged scalar $k^{\pm\pm}$ in that it can decay to a pair of same-sign down-type quarks or a pair of same-sign leptoquarks, if kinematically allowed. For the models mentioned above, we will only review the collider study with the characteristics different from the tree-level Seesaws in the following. Radiative Neutrino Mass Models at Colliders {#sec:radAtColliders} ------------------------------------------- ### Doubly-charged Scalar at the LHC As mentioned above, the Zee-Babu model contains two singlet charged scalars, $h^{\pm}$ and $k^{\pm\pm}$. Moreover, due to the presence of the doubly-charged scalar decay mode to two same-sign leptons $k^{\pm\pm}\to \ell^\pm \ell^\pm$ via the coupling $\mu_{ZB}$, collider searches for $L$-violating effects in the context of the Zee-Babu model are centered on $k^{\pm\pm}$ and its decays. Like the triplet Higgs in Type II Seesaw, the doubly-charged scalar $k^{\pm\pm}$ can be pair produced via the Drell-Yan process at the LHC if kinematically accessible and is given by $$\begin{aligned} p p \to \gamma^*/Z^* \to k^{++} k^{--}.\end{aligned}$$ This is the same process as shown in Fig. \[fig:feynmanTypeIIScalars\](a). However, an important distinction is that while $H^{\pm\pm}$ in the Type II Seesaw is an $SU(2)_L$ triplet, the $k^{\pm\pm}$ here is a singlet. As this quantum-number assignment leads to different $Z$ boson couplings, and hence different production cross section at colliders, it is a differentiating characteristic of the model. Note the $\gamma\gamma$ fusion processes, shown in Fig. \[fig:feynmanTypeIIScalars\], also applies to $k^{++}k^{--}$ pair production and leads to the same production cross section. Since the collider signal for pair produced $k^{\pm\pm}$ is the same as $H^{\pm\pm}$ in the Type II Seesaw, the search for doubly-charged scalar can be easily performed for both cases as shown in Fig. \[fig:type2ExclATLAS\]. Obviously, the constraint on the singlet is less stringent due to the absence of weak isospin interactions. With $36.1$ fb$^{-1}$ data at 13 TeV, ATLAS has excluded $k^{\pm\pm}$ mass lower than 656-761 GeV for ${\rm BR}(k^{\pm \pm}\to e^\pm e^\pm) + {\rm BR}(k^{\pm\pm} \to \mu^\pm \mu^\pm) = 1$ at 95% CLs [@ATLAS:2017iqw]. Low energy LFV experiments, especially $\mu\to e\gamma$, impose very stringent constraints on the parameter space of the Zee-Babu model. The MEG experiment [@Adam:2013mnn; @TheMEG:2016wtm] has placed an upper bound on the decay branching ratio ${\rm BR}(\mu\to e\gamma)<4.2\times 10^{-13}$, which can be roughly translated as [@Herrero-Garcia:2014hfa] $$\begin{aligned} \left|f_{13}^* f_{23}\right|^2 \frac{m_{k^{\pm\pm}}^2}{m_{h^\pm}^2} + 16 \left| \sum g_{1i}^* g_{i2} \right|^2 < 1.2\times 10^{-6} \left(\frac{m_k}{{\rm TeV}}\right)^4 \; .\end{aligned}$$ To satisfy LFV constraints, the doubly- and singly-charged scalar masses are pushed well above TeV, with $m_{k^{\pm\pm}}>1.3~(1.9)$ TeV and $m_{h^{\pm}}>1.3~(2.0)$ TeV for the NH (IH), assuming $\mu_{ZB}= min(m_{k^{\pm\pm}}, m_{h^{\pm}})$. This can be very easily relaxed, however, by choosing larger $\mu_{ZB}$ and balancing smaller Yukawa couplings to generate the right neutrino mass spectrum. A recent study has projected the sensitivities of the LHC with large luminosities by scaling the cross section bound by $1/\sqrt{\mathcal{L}}$ for two benchmark scenarios: one for NH and one for IH [@Alcaide:2017dcx]. The projected sensitivities are shown in Fig. \[fig:limit\_zeebabu\] for model parameters consistant with neutrino oscillation data. Note that these benchmarks are chosen to have $\mu_{ZB}=5 \min(m_{k^{\pm\pm}}, m_{h^{\pm}})$ such that the constraints from flavor experiments such as $\mu\to e\gamma$ are much less stringent at the price of a more fine-tuned the scalar potential. We can see that the NH benchmark is less constrained than the IH one when $m_{k^{\pm\pm}}<2 m_{h^{\pm}}$ because $k^{\pm\pm}$ has a smaller branching ratio to leptons. ### Leptoquark at the LHC ![$L$-violating processes at the LHC in the colored Zee-Babu model [@Kohda:2012sr].[]{data-label="fig:cZBM_LHC"}](cZBM_LHC.pdf){width="40.00000%"} In the colored Zee-Babu model, $L$-violating signals can be observed in events with pair produced leptoquarks $S_{LQ}^{-\frac{1}{3}}$ via $s$-channel diquark $S_{DQ}^{-\frac{2}{3}}$, shown in Fig. \[fig:cZBM\_LHC\], and given by, $$\begin{aligned} pp\to S_{DQ}^{-\frac{2}{3}*}\to S_{LQ}^{-\frac{1}{3}} S_{LQ}^{-\frac{1}{3}} \to u\ell^- u\ell^{\prime -} . \label{eq:radLQlnv}\end{aligned}$$ One benchmark has been briefly studied in Ref. [@Kohda:2012sr]. For leptoquark mass of 1 TeV and diquark mass of 4 TeV, a benchmark consistent with neutrino oscillation data and low energy experiments, the $L$-violating process in Eq. (\[eq:radLQlnv\]) can proceed with an LHC cross section of $0.18$ fb at $\sqrt{s}=14$ TeV. So far, there are no dedicated collider study for this model. In principle, however, one can recast ATLAS or CMS searches for heavy neutrinos, such as Refs. [@Aad:2015xaa; @Khachatryan:2014dka], to derive the limit on the model parameter space. Lepton number violating collider processes, $pp\to\ell^\pm\ell^\pm+nj$, involving charged scalars, leptoquarks and diquarks have also been studied for the LHC in Refs. [@Peng:2015haa; @Helo:2013ika; @Helo:2013dla]. Example diagrams are shown in Fig. \[fig:LNVLHC\]. Even though these studies are performed without a concrete neutrino mass model, they possess the most important ingredient of Majorana neutrino mass models: $L$ violation by two units, and therefore radiative neutrino mass models can be constructed from the relevant matter content. Some processes, however, are realized with a SM singlet fermion (for example the left panel of Fig. \[fig:LNVLHC\]), which implies the existence of a tree-level Seesaw. Other processes without SM singlet fermions, SU$(2)_L$ triplet scalars, or triplet fermions, such as the one on the right panel of Fig. \[fig:LNVLHC\], can be realized in a radiative neutrino mass model. Detailed kinematical analyses for resonant mass reconstruction would help to sort out the underlying dynamics. ### Correlation with Lepton Flavor Violation In radiative neutrino mass models the breaking of lepton number generally needs the simultaneous presence of multiple couplings. For example, in the Zee-Babu model, $Y$, $f$, $g$ and $\mu_{ZB}$ together break lepton number. The observation of pair produced $k^{\pm\pm}$ itself is insufficient to declare $L$ violation. In order to establish $L$ violation in the theory and thus probe the Majorana nature of the neutrinos, the couplings of $h^\pm$ to SM leptons and to $k^{\pm\pm}$ have to be studied at the same time. For the colored Zee-Babu model, the $L$ violation process shown in Fig. \[fig:cZBM\_LHC\] involves all couplings except the SM Yukawa necessary to break the lepton number. Note, however, the cross section for this process is proportional to the product of couplings and suppressed by the heavy exotic masses, which both contribute to the smallness of the neutrino masses. Thus the cross section for this processes must be kinematically suppressed. For radiative neutrino mass models with dark matter candidates, probing lepton number violation at colliders alone is generally much more difficult as the dark matter candidate appears as missing transverse energy just as neutrinos. Overall, the study of $L$-violation of radiative neutrino mass models can be performed either with the combination of different processes that test different subsets of the couplings or in a single process that involves all couplings at once whose production cross section is generally suppressed. On the contrary, radiative neutrino mass models contain LFV couplings and exotic particles that can be tested much easier than $L$ violation stated above. The search strategies for LFV couplings and new particles vary from model to model. It is definitely impossible to cover all and they are also not the focus of this review. Thus we will take a few simple examples to illustrate the searches. The leading LFV signals can be produced in a radiative neutrino mass model from the QCD pair production of the leptoquark $S_{LQ}^{-\frac{1}{3}}$ with its suitable subsequent decays such as $$\begin{aligned} pp ~\to~ S_{LQ}^{+\frac{1}{3}} S_{LQ}^{-\frac{1}{3}} ~\to~ \bar{t}\ell^+ t\ell^{\prime -}\end{aligned}$$ where $S_{LQ}^{+\frac{1}{3}}=\left(S_{LQ}^{-\frac{1}{3}}\right)^*$ and the top quarks decay hadronically. Note that the leptoquark pair can also decay to $\bar{b}\nu_{\ell} \ t\ell^{\prime -}$ or $\bar{b}\nu_{\ell} \ b \bar{\nu}_{\ell^\prime}$, where the LFV effects are not easy to disentangle at colliders due to the invisible neutrinos. However, these decay channels can result in final states $\ell^+\ell^{\prime -}X$, inclusive flavour off-diagonal charged lepton pair accompanied by missing transverse energy, jets , if the quarks decay to appropriate leptons. The same final states have been used to search for stop in SUSY theories and thus the results for stop searches at the LHC can be translated to that of the leptoquark $S_{LQ}^{-\frac{1}{3}}$, $m({S_{LQ}^{-\frac{1}{3}}})\gtrsim 600$ GeV [@Cai:2014kra] based on the ATLAS stop search at $\sqrt{s}=8$ TeV [@Aad:2014qaa] [^5]. No recast of stop searches has been performed for 13 TeV run at the time of this work. Besides leptoquarks, radiative neutrino mass models also comprise exotic particles such as vector-like quarks, vector-like leptons, charged scalar singlets (both singly- and doubly-charged) and higher-dimensional EW multiplets. For example, disappearing tracks can be used to search for higher-dimensional EW multiplet fermions whose mass splitting between the neutral and the singly-charged component is around 100 MeV. The current LHC searches have set a lower mass limit of 430 GeV at $95\%$ CL for a triplet fermion with a lifetime of about 0.2 ns [@ATLAS:2017bna; @Aad:2013yna; @CMS:2014gxa]. We refer the readers to the section about collider tests of radiative neutrino mass model in Ref. [@Cai:2017jrq] and the references therein for details. We want to stress, however, that even though $L$ violation in the radiative models is more complicated and challenging to search for in collider experiments, their observation is essential and conclusive to establish the Majorana nature of neutrinos. So once we find signals in either LFV processes or new particles searches, we should search for $L$ violation in specific radiative neutrino mass models that give these LFV processes or contain these new particles, in order to ultimately test the generation of neutrino masses. Summary and Conclusions {#sec:con} ======================= Exploring the origin of neutrinos’ tiny masses, their large mixing, and their Dirac or Majorana nature are among the most pressing issues in particle physics today. If one or more neutrino Seesaw mechanisms are realized in nature, it would be ultimately important to identify the new scales responsible for generating neutrino masses. Neutrino oscillation experiments, however, may not provide such information, and thus complementary pathways, such as collider experiments, are vital to understanding the nature of neutrinos. Observing lepton number violation at collider experiments would be a conclusive verdict for the existence of neutrino Majorana masses, but also direct evidence of a mass scale qualitatively distinct from those in the SM. In this context, we have reviewed tests of low-scale neutrino mass models at $pp$, $ep$, and $ee$ colliders, focusing particularly on searches for lepton number $(L)$ violation: We begin with summarizing present neutrino oscillation and cosmology data and their impact on the light neutrino mass spectra in Sec. \[sec:nuparameters\]. We then consider several representative scenarios as phenomenological benchmarks, including the characteristic Type I Seesaw in Sec. \[sec:type1\], the Type II Seesaw in Sec. \[sec:type2\], the Type III in Sec. \[sec:type3\], radiative constructions in Sec. \[sec:loop\], as well as extensions and hybridizations of these scenarios. We summarize the current status of experimental signatures featuring $L$ violation, and present anticipated coverage in the theory parameter space at current and future colliders. We emphasize new production and decay channels, their phenomenological relevance and treatment across different collider facilities. We also summarize available Monte Carlo tools available for studying Seesaw partners in collider environments. The Type I Seesaw is characterized by new right-handed, SM gauge singlet neutrinos, known also as “sterile neutrinos,” which mix with left-handed neutrinos via mass diagonalization. As this mixing scales with light neutrino masses and elements of the PMNS matrix, heavy neutrino decays to charged leptons may exhibit some predictable patterns if one adopts some simplifying assumptions for the mixing matrix, as shown for example in Figs. \[fig:brnlwcase1\] and \[fig:nibr\], that are correlated with neutrino oscillation data. The canonical high-scale Type I model, however, predicts tiny active-sterile mixing, with $\vert V_{\ell N}\vert^2 \sim m_\nu/M_N$, and thus that heavy $N$ decouple from collider experiments. Subsequently, observing lepton number violation in collider experiments, as discussed in Sec. \[sec:typeIhybrid\], implies a much richer neutrino mass-generation scheme than just the canonical, high-scale Type I Seesaw. In exploring the phenomenological parameter space, the 14 TeV LHC (and potential 100 TeV successor) and $\mathcal{L}=1$ ab$^{-1}$ integrated luminosity could reach at least $2\sigma$ sensitivity for heavy neutrino masses of $M_N\lesssim 500$ GeV ($1$ TeV) with a mixing $\vert V_{\ell N}\vert^2 \lesssim 10^{-3}$, as seen in Fig. \[100TeVdiscovery.fig\]. If $N$ is charged under another gauge group that also couples to the SM, as in $B$-$L$ or LR gauge extensions, then the discovery limit may be extended to $M_N \sim M_{Z'}, M_{W_R}$, when kinematically accessible; see Secs. \[sec:type1Abelian\] and Sec. \[sec:lrsmCollider\]. The Type II Seesaw is characterized by heavy SU$(2)_L$ triplet scalars, which result in new singly- and doubly-charged Higgs bosons. They can be copiously produced in pairs via SM electroweak gauge interactions if kinematically accessible at collider energies, and search for the doubly-charged Higgs bosons via the same-sign dilepton channel $H^{\pm\pm} \to \ell^\pm \ell^\pm$ is an on-going effort at the LHC. Current direct searches at 13 TeV bound triplet scalar masses to be above (roughly) 800 GeV. With anticipated LHC luminosity and energy upgrades, one can expect for the search to go beyond a TeV. Furthermore, if neutrino masses are dominantly from triplet Yukawa couplings, then the patterns of the neutrino mixing and mass relations from the oscillation experiments will correlate with the decays of the triplet Higgs bosons to charged leptons, as seen from the branching fraction predictions in Figs. \[brii\] and \[brij\] and in Table \[relationii\]. Since a Higgs triplet naturally exists in certain extensions beyond the SM, such as in Little Higgs theory, the LRSM, and GUT theories, the search for such signals may prove beneficial as discussed in \[sec:type2LRSM\]. The Type III Seesaw is characterized by heavy SU$(2)_L$ triplet leptons, which result in vector-like, charged and neutral leptons. Such multiplets can be realized in realistic GUT theories in hybridization with heavy singlet neutrinos from a Type I Seesaw. Drell-Yan pair production of heavy charged leptons at hadron colliders is sizable as it is governed by the SM gauge interactions. They can decay to the SM leptons plus EW bosons, leading to same-sign dilepton events. Direct searches for promptly decaying triplet leptons at the LHC set a lower bound on the triplet mass scale of around 800 GeV. A future 100 TeV $pp$ collider can extend the mass reach to at least several TeV, as seen in Fig. \[fig:type3Searches\]. Finally, neutrino masses can also be generated radiatively, which provides an attractive explanation for the smallness of neutrino masses with a plausibly low mass scale. Among the large collection of radiative neutrino mass models, the Zee-Babu model contains a doubly-charged SU$(2)_L$ singlet scalar with collider signal akin to the doubly-charged Higgs in the Type II Seesaw. ATLAS has excluded $k^{\pm\pm}$ mass below $660-760$ GeV assuming the benchmark decay rate $\sum_{\ell_i=e,\mu} {\rm BR}(k^{\pm\pm} \to \ell_1^\pm \ell_2^\pm )=1$. The high luminosity LHC is sensitive up to about a TeV for both $k^{\pm\pm}$ and its companion scalar $h^\pm$ in the Zee-Babu model with constraints from neutrino oscillation data and other low energy experiments. For the colored variant of the Zee-Babu model, a pair of same-sign leptoquark can be produced via an $s$-channel diquark at the LHC. Their subsequent decay lead to the lepton number violating same-sign dilepton plus jets final state, which still await dedicated studies. As a final remark, viable low-scale neutrino mass models often generate a rich flavor structure in the charged lepton sector that predict lepton flavor-violating transitions. Such processes are typically much more easily observable than lepton number violating processes, in part due to larger production and decay rates, and should be searched for in both high- and low-energy experiments. Acknowledgements {#sec:ack} ================ We thank Michael A. Schmidt for useful discussions. Past and present members of the IPPP are thanked for discussions. The work of TH was supported in part by the U.S. Department of Energy under grant No. DE-FG02- 95ER40896 and in part by the PITT PACC. The work of TL was supported in part by the Australian Research Council Centre of Excellence for Particle Physics at the Tera-scale. The work of RR was funded in part by the UK Science and Technology Facilities Council (STFC), the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreements No 690575 (InvisiblesPlus RISE) and No 674896 (InvisiblesPlus RISE). Support for the open accessibility of this work is provided by the Research Councils UK, external grant number ST/G000905/1.\ Republication of the various figures is granted under the terms of the Creative Commons Attribution Licenses, American Physical Society PRD License Nos.:  4234250091794,  4234250774584,  4234260138469,  4234270420615,  4234270625995,  4234270758210,  4234280275772,\  4234280455112,  4234291212539,  4234300360477,  4234301170893,  4234310175148, and PRL License No. 4234301302410. [^1]: This is done for simplicity since $U_{PMNS}$ in Table \[tab:nufit\] is unitary whereas here it is not; for more details, see [@Esteban:2016qun; @Parke:2015goa]. [^2]: $\sum_N (V_{\ell N}^\ast)^2=\sum_N |V_{\ell N}|^2$ only when all phases on the right-hand side of Eq. (\[dege\]) vanish [@Perez:2009mu]. [^3]: Where $\text{BR}(A\to X)\equiv \Gamma(A\to X)/\sum_Y\Gamma(A\to Y)$ for partial width $\Gamma(A\to Y)$. [^4]: A note on terminology: High-$p_T$ hadron collider observables, , fiducial distributions, are inherently inclusive with respect to jets with arbitrarily low $p_T$. In this sense, we refer to hadronic-level processes with a fixed multiplicity of jets satisfying kinematical requirements (and with an arbitrary number of additional jets that do not) as *exclusive,* , $pp\to W^\pm + 3j + X$; those with a minimum multiplicity meeting these requirements are labeled *semi-inclusive,* , $pp\to W^\pm + \geq 3j + X$; and those with an arbitrary number of jets are labeled *inclusive,* , $pp\to W^\pm + X$. Due to DGLAP-evolution, exclusive, partonic amplitudes convolved with PDFs are semi-inclusive at the hadronic level. [^5]: There are many dedicated leptoquark searches at the LHC [@Aaboud:2016qeg; @Aad:2015caa; @CMS:2016imw; @CMS:2016qhm; @CMS:2016hsa]. However, the leptoquarks searched only couple to one generation of fermions at a time and thus generate no LFV signals.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The Southern Galactic Plane Survey (SGPS) is a project to image the   line emission and 1.4 GHz continuum in the fourth quadrant of the Milky Way at high resolution using the Australia Telescope Compact Array (ATCA) and the Parkes Radio Telescope. In this paper we describe the survey details and goals, present $\lambda 21$-cm continuum data, and discuss  absorption and emission characteristics of the SGPS Test Region ($325\fdg5 \leq l \leq 333\fdg5$; $-0\fdg5 \leq b \leq +3\fdg5$). We explore the effects of massive stars on the interstellar medium (ISM) through a study of  shells and the  environments of  regions and supernova remnants. We find an  shell surrounding the   region RCW 94 which indicates that the region is embedded in a molecular cloud. We give lower limits for the kinematic distances to SNRs G327.4+0.4 and G330.2+1.0 of 4.3 kpc and 4.9 kpc, respectively. We find evidence of interaction with the surrounding  for both of these remnants. We also present images of a possible new SNR G328.6-0.0. Additionally, we have discovered two small  shells with no counterparts in continuum emission.' author: - 'N. M. McClure-Griffiths, A. J. Green, John M. Dickey, B. M. Gaensler, R. F. Haynes, & M. H. Wieringa' nocite: - '[@weaver74]' - '[@goss70]' - '[@egan98]' - '[@dickey92]' - '[@giacani00]' title: 'The Southern Galactic Plane Survey: The Test Region' --- Introduction {#sec:intro} ============ Surveys of the Milky Way allow us to look at the energetics and structure of our own Galaxy with spatial resolution that is unattainable in other galaxies. The inner Galaxy, in particular, provides a number of scientific opportunities. In the 1960’s and 70’s Galactic neutral hydrogen () and continuum were mapped using single dish radio telescopes with low spatial resolution (eg. Weaver & Williams 1974; Goss & Shaver 1970). As a result, much of what we know about the structure of the Galactic  is restricted to large scales. In other wave-bands Galactic surveys are much more up-to-date. In the infrared, X-ray, and H$\alpha$ the inner Galaxy has been mapped extensively, while  remains seriously under-sampled. Over the past five years the  atlas of the Galaxy has been greatly improved by the Dominion Radio Astrophysical Observatory’s (DRAO) Canadian Galactic Plane Survey (CGPS) which covers a longitude range of $l= 74$ to 147 using a combination of interferometer and single dish data to image the Galaxy at a resolution of one arcminute [@taylor99]. Despite the contributions of the CGPS to our knowledge of Galactic , the inner Galaxy remains neglected. We have recently begun the Southern Galactic Plane Survey (SGPS), a large-scale project to image the $\lambda 21$-cm continuum and  spectral line in the fourth quadrant of the Galactic Plane with high angular and velocity resolution [@dickey99; @mcgriff99]. The SGPS makes use of high spatial resolution data from the Australia Telescope Compact Array (ATCA) and short spacing information from the Parkes 64m single dish.[^1] The final project will provide a complete  dataset of $253\arcdeg \leq l \leq 358$ and $-1\fdg0 \leq b \leq +1\fdg0$ at an angular resolution of 2, and with velocity resolution of $\Delta v = 0.82~{\rm km~s^{-1}}$. In addition, we have extended the single dish coverage to $b=\pm 10\arcdeg$ in order to study large scale structures which protrude from the Galactic plane. This dataset is particularly useful for studying the structure and dynamics of the neutral medium, on which massive stars have a significant impact. O and B stars not only affect the medium through ionization, as in the case of  regions, they also perturb the medium through winds and at the end of their lives, as supernovae. The resultant supernova remnants (SNRs) plow shocks through the neutral medium, ionizing and compressing the medium and leaving a lasting impression. Not only do  regions and SNRs impact the ISM, the structure of the ISM - particularly density enhancements - affects the morphology of  regions and SNRs. After the SNR or  region has ceased to exist in continuum, the imprint may remain in the ISM in the form of an  shell. Because of the large range of spatial scales sampled with the combined Parkes and ATCA data, the SGPS is an ideal dataset in which to explore these effects. In this paper we introduce the details of the Southern Galactic Plane Survey with attention to  spectral line and $\lambda$21-cm continuum data from the SGPS Test Region ($325\fdg5 \leq l \leq 333\fdg5$; $-0\fdg5 \leq b \leq +3\fdg5$). Other scientific highlights from the SGPS are discussed elsewhere: two large  shells discovered in the Parkes data are presented in @mcgriff00b, preliminary images of  emission and absorption features are presented in @mcgriff00a, and the polarization properties of the Test Region are presented in @gaensler00a. Here we explore the connections between the  and $\lambda$21-cm continuum images of the Test Region. In §\[subsec:obj\] & \[sec:obs\] we describe the survey objectives, observing and data analysis strategies. In §\[sec:cont\] we discuss the $\lambda$21-cm continuum emission. absorption towards continuum sources is discussed in §\[sec:abs\]. We have chosen a representative sample of  regions and supernova remnants (SNRs) to study the relationship between the continuum emission from these objects and the surrounding  environments in §\[sec:emis\]. The Test Region is an excellent area to initiate such a study as it contains many  regions and SNRs, as well as extended emission structure. Using absorption and  morphological matches to the continuum emission, we seek to create a three-dimensional view of the Galaxy in this subregion. Survey Objectives {#subsec:obj} ----------------- The general goal of the Southern Galactic Plane Survey is to provide a dataset with which to study the structure and dynamics of the neutral hydrogen () in the inner Galaxy. Previous studies of the inner Galaxy have lacked the sensitivity and resolution necessary to study the physical processes of the interstellar medium (ISM) over a large range of spatial scales. Though the specific goals of the SGPS are numerous, we will highlight a few below: - The SGPS will allow us to address questions about the spatial distributions and scale heights of the warm and cool components of the neutral medium in order to understand the thermal phases of the interstellar medium. - Combining the Parkes and ATCA data, we will be able to probe the interstellar medium through a broad range of spatial scales in order to develop a statistical interpretation of the neutral ISM throughout the inner Galaxy. -  emission data over the large range of spatial scales available in the SGPS will allow us to detect a full sample of  shells with which to study not only the formation of shells - particularly of the largest, most enigmatic ones - but also their distribution in the Galaxy and global effect on Galactic structure. - A number of  self-absorption (HISA) features are apparent in the SGPS Test Region [@mcgriff00a]. HISA, where cold  clouds absorb the diffuse background  emission, is an excellent probe of the distribution of the coldest, most compact  clouds. Further exploration of these features will be presented in a subsequent paper [@dickey00]. - The dataset will provide  absorption spectra for distance estimates of many Galactic objects and may be useful in identifying those extragalactic sources located close to the Galactic plane. - The inclusion of full polarization information for the continuum data will allow us to explore the polarization structure of individual objects, such as supernova remnants, and to investigate the Galactic magnetic field structure using the polarization of the diffuse background emission. Results of the polarization properties of the Test Region appear in @gaensler00a. Observations and Analysis {#sec:obs} ========================= Observations of the SGPS Test Region were made with the Australia Telescope Compact Array (ATCA; Frater, Brooks, & Whiteoak 1992) and Parkes Radio Telescope. The ATCA is an east-west synthesis instrument near Narrabri NSW, with six 22 m antennas on a 6 km track. Five antennas are movable into configurations with baselines between 31 m and 6 km. The ATCA data consist of a 190 pointing mosaic covering $325\fdg5 \leq l \leq 333\fdg5$ and $-0\fdg5 \leq b \leq +3\fdg5$. These data were obtained during five separate observing sessions between April 1997 and April 1998. The observing dates and times are given in Table \[tab:obs\]. The observations were made with several compact array configurations - 750A, 750C, 750D, and 375 - in order to obtain maximum sensitivity to large scale structures. Each of the 190 pointings was observed in forty 30 s snapshots at a broad range of hour angles for good [*u-v*]{} coverage. The pointings for the Test Region were arranged on a square grid with 15separation (Nyquist) as determined by the ATCA primary beam FWHM, which is 33 at $\lambda$-21 cm. The pointing centers are plotted on the 21-cm continuum ATCA image of the SGPS Test Region shown in Fig. \[fig:centers\]. The ATCA feeds receive two orthogonal linear polarizations, $X$ and $Y$. All observations were recorded in a wideband continuum mode with 32 channels, each 4 MHz, across a 128 MHz total bandwidth with polarization products $XX$, $YY$, $XY$, and $YX$ to enable calculation of all four Stokes parameters; and simultaneously in a spectral line mode with polarization products $XX$ and $YY$ in 1024 channels across a 4 MHz total bandwidth. The continuum data are centered on $\nu = 1384$ MHz, whereas the spectral line data are centered on $\nu= 1420.0$ MHz and have channel separation of $\Delta v = 0.82~{\rm km~s^{-1}}$ ($3.9$ kHz). Data editing, calibration, and imaging of the ATCA data are carried out in the [miriad]{} data reduction package [@sault00]. The source PKS B1934-638 is used for flux density and bandpass calibration and was observed once per observing session. A flux density of 14.94 Jy at 1420 MHz is assumed [@reynolds94]. PKS B1934-638 has no detectable linear polarization and can therefore be used to solve for the polarization leakages. Polarization leakages are also calculated for the sources MRC B1613-586 and MRC B1431-48 which were observed approximately once every hour for calibration of the time-variation in complex antenna gains. The individual pointings were linearly combined and imaged using a standard grid-and-FFT scheme with super-uniform weighting. Like uniform weighting, super-uniform weighting minimizes sidelobe levels to improve the dynamic range and sensitivity to extended structure. Uniform weighting reduces to natural weighting, however, if the field of view is much larger than the primary beam, as is the case for large mosaics. Super-uniform weighting overcomes this limitation by decoupling the weighting from the field size [@sault00]. In this manner super-uniform weighting attempts to minimize sidelobe contributions from strong sources over a region smaller than the full field being imaged and therefore is typically more successful than uniform weighting on large mosaics [@sault96]. The specifics of the calibration and imaging of ATCA polarization data are described in [@gaensler00a], here we discuss only the Stokes I continuum and  data. Two  data cubes were produced. In one cube the continuum emission was subtracted from the  [*u-v*]{} data using the [miriad]{} task uvlin. The second cube, for use in absorption studies, contains continuum emission. The resultant synthesized beam for both cubes and the continuum image is $124\farcs9 \times 107\farcs5$ ($\alpha \times \delta$). For the data presented here the 6 km baselines of the ATCA were not used. However, the long baselines are retained for some absorption studies, such as those presented in @dickey00. Joint deconvolution was performed on the full linear mosaics using a maximum entropy algorithm implemented in [miriad]{} task mosmem [@sault96]. The method of joint deconvolution is very effective for maximizing the [*u-v*]{} coverage attainable in mosaiced observations [@sault96]. Despite deconvolution, some sidelobes are visible around strong point sources in the continuum and  images. Though mosaicing allows us to recover angular scales larger than normal interferometric observations by reducing the effective shortest projected baseline, we are nonetheless limited in practice to angular scales smaller than $\theta = \lambda / (d -D/2)$, where $d=30.6$ m is the shortest baseline of the ATCA, and $D= 22$ m is the diameter of a single antenna. This limits the ATCA data to angular scales smaller than $\theta \approx 36\arcmin$. In order to recover information on larger size scales, the ATCA mosaic must be combined with single dish data from the Parkes Radio Telescope. The Parkes Radio Telescope is a 64 m antenna situated near Parkes NSW, Australia. It has a thirteen beam, $\lambda$21-cm receiver package at prime focus [@staveley-smith96]. The SGPS Test Region was observed on 1998 December 15-16. The Parkes survey was subsequently expanded to $b=\pm 10\arcdeg$ for more complete coverage of large scale structures during additional observing sessions: 1998 June 18-21, 1999 September 18-27, and 2000 March 10-15. Observations were made by the process of mapping “on-the-fly” with the inner seven beams of the multibeam system. Data were recorded in 5 s samples, while scanning through three degrees in Galactic latitude. The data were taken in frequency switching mode using the narrowband back-end [@haynes98], with a total bandwidth of 8 MHz across 2048 channels. Though the Parkes data are frequency switched, total power information for each sample is retained. Each sample was divided by the previous frequency switched sample and the residual bandpass shape fitted with a series of Fourier components. The spectra were then multiplied by the mean of the reference signal over the spectrum to reconstruct the continuum emission with a flat baseline. Absolute brightness temperature calibration of the  line data was performed from observations of the IAU standard regions S6 and S9 [@williams73]. A detailed description of the observing strategy, calibration, and imaging procedures is found in @mcgriff00b. Off-line channels were used for continuum subtraction and to produce the continuum image using the AIPS task IMLIN. The Parkes data on the SGPS Test Region have an angular resolution of $\sim 15\arcmin$. The final, calibrated data have a bandwidth of 4 MHz with 1024 channels, such that the velocity resolution of $0.82$ matches the ATCA data. It should be noted that the data presented here have not been corrected for the effects of stray radiation. Stray radiation leakage from bright emission through the back sidelobes of a single dish beam introduces baseline errors which are typically between $0.5$ K and $2$ K [@kalberla80]. When compared to low latitude Galactic spectra, this is a small percentage, but it does nonetheless limit the sensitivity and accurate representation of extended features in the Parkes data. The data for the entire survey will have a first order stray radiation correction applied. A complete correction requires a low resolution survey of the entire sky with a known antenna pattern in order to reconstruct the stray radiation at every point on the sky, at any azimuth and elevation, and at any time of the year. Such a task is beyond the scope of this project. A first order correction, however, can be done by re-observing the survey region at different times of the year and comparing the spectra. The velocity shifts caused by the Earth’s motion around the Sun result in excess emission at different velocities. We have, therefore, re-observed the full survey region four times at three month intervals around the year and we will compare each spectrum. The minimum value at each velocity will be a reasonable upper limit to the stray radiation corrected profile, good to within $\sim 0.5$ K of $T_B$. The final step of imaging involves combining the Parkes and ATCA data. The data may be combined in the Fourier domain after deconvolution of the individual images or in the [*u-v*]{} plane prior to deconvolution. @stanimirovic99 showed that the results are comparable using either method, but that combining after deconvolution produced results that were typically more consistent than with other methods. Comparison of our data combined in both ways shows similar results. We have chosen, therefore, to combine the data in the Fourier domain after deconvolution. In this method, the interferometric data  and continuum data are imaged and deconvolved, the single-dish data are imaged and the clean interferometric and single-dish images are Fourier transformed and combined. This technique is implemented in the [miriad]{} task immerge. Slight differences in calibration can lead to the necessity of a relative calibration factor by which the single-dish dataset is multiplied before combination. This calibration factor is determined by comparing the datasets in the Fourier plane at every pixel and frequency in the range of overlapping spatial frequencies. In order to calculate the calibration factor both images must be deconvolved, a step which requires a good knowledge of the single-dish beam [@stanimirovic99]. Using a two dimensional Gaussian with FWHM $15\farcm5$ for the Parkes beam and by comparing the Parkes and ATCA continuum images of a strong, compact source in the Test Region, we calculated a relative calibration factor of 1.19. Two combined  data cubes were created, one containing the continuum emission for absorption studies and one which had the continuum subtracted for emission studies. The continuum images were combined in the same way as the cube and with the same calibration factor. The combined Parkes and ATCA 21-cm continuum image is shown in Fig. \[fig:21cont\]. The combined data are sensitive to all angular scales from the synthesized beam size, $124\farcs9 \times 107\farcs5$ ($\alpha \times \delta$), up to the image size, $8\arcdeg \times 4 \arcdeg$ ($l \times b$) for the Test Region. Because of the fine scale structure seen in the velocity domain, no Hanning smoothing was applied to the data. Each channel image has a velocity separation of $0.82$ . Channel images from the continuum subtracted combined data cube are shown in Fig. \[fig:chans1\]. Every fourth channel from $v=-127$  to $v=79$ is shown. The rms noise in the channel images is $\sim 2.4$ K of ${\rm T_B}$ for the ATCA data, $\sim 100$ mK of ${\rm T_B}$ for the Parkes data and $\sim 2.3$ K of ${\rm T_B}$ for the combined dataset. The rms noise in the continuum images is $\sim 5.5~{\rm mJy~beam^{-1}}$ for the ATCA data, $\sim 500~{\rm mJy~beam^{-1}}$ for the Parkes data (beam size $15\farcm5 \times 15\farcm5$), and $\sim 7~{\rm mJy~beam^{-1}}$ for the combined data. Continuum Emission {#sec:cont} ================== The combined $\lambda$21-cm continuum image of the SGPS Test Region is shown in Fig. \[fig:21cont\]. Most sources have been previously catalogued as  regions or SNRs [@avedisova97; @caswell87; @green00; @whiteoak96]. There are also many unresolved sources scattered throughout the Test Region.  absorption measurements towards many of these suggest that most are extragalactic. This region has been studied in H${\rm \alpha}$ by @georgelin94 as part of an extensive H${\rm \alpha}$ survey of the Southern Galactic Plane. As shown in the Fig. \[fig:diag\], a diagram of the expected velocities and spiral arms in the fourth quadrant, the Test Region line of sight crosses both the Sagittarius-Carina and Scutum-Crux spiral arms and runs tangent to the Norma arm at $l\approx 327\arcdeg$. As a result this region has a particularly high density of continuum sources. Discrete Sources {#subsec:sources} ---------------- We describe here the more prominent discrete sources in the SGPS Test Region. These sources are marked in on the MOST 843 MHz continuum image shown in Fig. \[fig:most\]. Several individual sources are discussed in detail below with comment given about their associated  emission. Starting at the lower longitude end, the first strong source is RCW 94 [@rodgers60; @shaver79] at $l=326\fdg3$, $b=+0\fdg8$, with an angular diameter of about 18. This ring-like structure is an  region, with strongest emission to the lower left. There is a smaller region adjoining the  region at $l=326\fdg4$, $b=+0\fdg9$. At $l=326\fdg7$, $b=+0\fdg8$ is another  region, RCW 95. Directly below RCW 95 is the brighter, extended  region G326.65+0.59 [@georgelin94]. Closer to the Plane at higher longitudes is a very extended thermal filamentary structure G326.96+0.03. This source has arcs of emission above and below a centralized bright knot. Because these sources are all at the same distance, we refer to the grouping of RCW 94, RCW 95, G326.65+0.59, and G326.96+0.03 as the RCW 94-95  region complex. Above the high longitude edge of the G326.96+0.03 arc is SNR G327.4+0.4, a large shell type SNR with enhanced limb brightening to the lower left. Further from the Plane than G327.4+0.4 there is a smaller, weaker supernova remnant, SNR G327.4+1.0. This source has a nearly closed arc extending to higher latitudes. At slightly higher longitudes there is a region of extended emission comprised of several thermal sources grouped at G327.83+0.11. At higher longitudes and lower latitudes than these sources there is another  region, G327.99-0.09. Near $l=328\arcdeg$ is the compact  region G328.31+0.45 and the extremely bright Crab-like SNR G328.4+0.2 [@gaensler00b]. At higher longitudes, the compact source G328.81-0.08 is classified as an  region in @caswell87 [ hereafter CH87] on the basis of a recombination line detection. However, examination of the Midcourse Space Experiment (MSX; Egan et al. 1998) band A ($6.8-10.8$ ) image shows only a small infrared source, IRAS 15550-5306, slightly offset from the center of G328.81-0.08. This infrared source has a FWHM of $\sim 30\arcsec$, whereas the offset $\lambda$21-cm source has a FWHM of $\sim 3\arcmin$. It is unclear whether the infrared source is the same as the radio source. About $30\arcmin$ from G328.81-0.08 there is an extended source centered at $l=328\fdg6$, $b=0\arcdeg$. Fig. \[fig:newsnr\] shows the combined Parkes and ATCA 1.4 GHz SGPS continuum image of this region and the MOST 843 MHz image of the same area. In the SGPS image the source has a mostly filled, double loop morphology of angular diameter $\sim 0\fdg5$, and internal filamentary structure. Only the edges of the source are observed as filaments at the $\sim 15~{\rm mJy~beam^{-1}}$ level in the MOST image. The smooth, extended emission observed in the SGPS image is not detected by MOST. There is no counterpart to this emission in the MSX band A image. The fact that no corresponding IR emission is detected from this source suggests that the emission is non-thermal and we propose that it is a new SNR candidate, G328.6-0.0. There are also several long overlapping thermal filaments visible in both of the radio continuum images and the MSX image. These thermal filaments do not appear to be directly associated with this double loop SNR candidate, G328.6-0.0, but extend to $b=0\fdg5$. Embedded within the left edge of the SNR candidate is an unresolved source at $l=328\fdg59$, $b=-0\fdg11$ which does not appear in the MSX image, suggesting that it too is either non-thermal or extragalactic. Near $l=329\arcdeg$ there are two  regions, G329.35+0.14 and G329.49+0.21. At larger longitudes there is the large supernova remnant, SNR G329.7+0.4, which accounts for much of the extended emission in this portion of the Test Region. The large ($\Delta \theta \sim 40\arcmin$) remnant consists of many loops. Above G329.7+0.4 is a large ($\sim 20\arcmin$) loop of thermal emission which is also visible in the MSX images. This may be an extended  envelope (EHE) as similarly suggested for the area around $l=312\arcdeg$ [@whiteoak94]. This region appears connected to large arcs of thermal emission extending above and below SNR G330.2+1.0. SNR G330.2+1.0 is a composite remnant with irregular emission that does not seem to form a clear shell. Between $l=330\fdg2$ and $l=330\fdg9$ there are few bright discrete sources. The area is largely filled with extended emission. Another large  complex is located around $l=331\arcdeg$, including the  regions: G331.03-0.15, G331.26-0.19, G331.52-0.07; and at the high longitude limit of the Test Region: SNR G332.0+0.2. SNR G332.0+0.2 is a mostly complete shell of angular diameter $\sim 10\arcmin$. Further above the Plane are two large  regions: G331.35+1.07 and G331.36+0.51. Extended Continuum Emission {#subsec:extended} --------------------------- Throughout the Test Region there is diffuse emission which decreases in intensity with increasing latitude. Fig. \[fig:contslice\] shows three slices across the Test Region continuum image, all at $l=328\fdg7$. The first slice is across the ATCA image alone, the second across the Parkes image alone, and the third is across the combined image. It is clear that the ATCA data is not sensitive to the large scale Galactic emission, but resolves the individual sources. The slice across the Parkes image shows the large-scale decrease in emission, but the low resolution does not clearly delineate the individual sources. The combined image, however, shows both the large-scale emission and the resolved sources. From this slice it is obvious that both single dish and interferometric data are necessary to understand the relationship between discrete and extended Galactic continuum emission. Though some of the extended emission in the Test Region is smooth and can be attributed to the diffuse Galactic background, there is also structure in the emission. It is not immediately obvious whether the structure can be attributed to the Galactic background, whether it is associated with discrete continuum sources, or whether it has an altogether different nature. This question was addressed by @whiteoak94 who carefully examined a $3\arcdeg \times 2\arcdeg$ region in the Plane around $l=312\arcdeg$ from the MOST Galactic Plane Survey [@green99]. They note low surface brightness, extended emission in the high resolution 843 MHz continuum images and suggest that this is associated with the ionized ISM. The emission they detect is thermal and they relate some of it to extended  envelopes (EHEs) around  regions. The MOST data are only sensitive to structures up to 30, so it is unclear whether there are larger-scale associations amongst some of these filamentary structures. The SGPS data are sensitive to all size scales up to about four degrees and are hence well-suited to study these features. The MOST and SGPS surveys complement each other very well. Comparison of the MOST 843 MHz image in Fig. \[fig:most\] and the SGPS ATCA continuum image in Fig. \[fig:centers\] shows good correlation between the two interferometric images, though the MOST images are at a slightly higher angular resolution ($43\farcs0 \times 51\farcs9, \alpha \times \delta$). In both images there are filaments, typically a few arcminutes in width and up to few degrees in length. With the inclusion of the Parkes data in Fig. \[fig:21cont\] we can observe how these filaments relate to the larger-scale diffuse emission. It appears that there are two categories of extended emission: one where filamentary structures are part of larger-scale, filled structures and one where the filaments are self-contained structures. An example of a structure which appears filled-in with the inclusion of the Parkes data is the possible SNR G328.6-0.0, as shown in Fig. \[fig:newsnr\] and described in §\[subsec:sources\]. In this case, the loops observed in the MOST and ATCA images are observed as part of a cohesive, filled structure in the SGPS image. Clearly, the filaments are the SNR edges which the interferometers can detect, while they cannot detect the large-scale smooth emission in the center of the remnant. By contrast, there are many loops and filaments near SNR G330.2+1.0 that appear filamentary in both the MOST and SGPS images, implying that they are not part of a larger, filled structure. These filaments are visible in the MSX images, implying that they are thermal. These structures may be sheets viewed edge-on, threads, or the edges of EHEs where the surfaces are too diffuse to be detected.  Absorption {#sec:abs} =========== In order to create a three-dimensional view of this portion of the Galaxy we have extracted  absorption spectra towards the brighter  regions and SNRs. The method we use to determine the HI absorption spectrum is based on averaging the spectra toward the brightest part of the continuum (“on-source spectra”), and subtracting an interpolated average of the spectra from the region surrounding the continuum source (“off-source spectra”; Dickey et al. 1992). The on-source spectra are selected based on the continuum image by setting a high threshold, typically 80% of the continuum peak, and including in the average only spectra toward pixels whose continuum brightness is above this high threshold. These on-source spectra are averaged with weighting factor equal to the continuum brightness in each pixel, which optimizes the signal-to-noise ratio in the resulting absorption spectrum. Similarly, the off-source spectra are selected toward pixels whose continuum brightness is below a low threshold, typically 20% of the continuum peak. The off-source spectra are not simply averaged, but interpolated to give a better prediction of the emission spectrum in the direction of the continuum peak. This interpolation is based on a simple bi-linear fit (least squares fitting a linear function of two dimensions) done independently for each spectral channel. The outer boundary for which spectra are included in the off-source interpolation is typically 7, but for the extended supernova remnant G327.4+0.2 we extend the outer boundary to 15. The high and low thresholds also have to be adjusted in some cases depending on the continuum flux, down to 70% for the high threshold of the faintest sources, and in the range 10% to 30% for the low threshold depending on the angular size of the continuum distribution. This interpolation process does not change the fundamental angular resolution of the survey, i.e. it is not an extrapolation on the [*u-v*]{} plane. So the spectra derived for both the absorption and the expected emission still correspond to one beam area of roughly 2 diameter. The absorption spectrum is then determined by subtracting the interpolated off-source spectrum from the averaged on-source spectrum. The resultant optical depth spectrum, $e^{-\tau}$, is the absorption spectrum divided by the continuum flux averaged over the pixels above the high threshold. The optical depth spectra towards RCW 94 and G326.65+0.59 are shown in Fig. \[fig:rcw94\_abs\] (bottom plot) with corresponding interpolated off-source emission spectra (upper plot). The errors in the absorption spectra are generally dominated by uncertainty in the interpolated emission, particularly at low latitudes where the emission is not smoothly distributed on scales of a few arcminutes. To estimate the error in the absorption we compute the predictions for the off-source spectra based on the bilinear fit to the emission and take the difference between these predictions and the actual spectra in each off-source pixel. The rms average of these differences gives an error envelope on the interpolated emission in the direction of the continuum peak. Dividing by the continuum flux averaged over the on-source pixels gives the optical depth error spectrum ($\pm 1 \sigma$), shown in dotted lines on Figs. \[fig:rcw94\_abs\] & \[fig:g327abs\]. For the absorption feature with most extreme velocity (most negative or most positive) we define $V_{\rm L}$, the lower limit on the distance. Following @frail90 we define the velocity corresponding to the upper distance limit, $V_{\rm U}$, as the first emission peak ($T_b > 35$ K) beyond $V_{\rm L}$ which does not correspond to any absorption features. Upper distance limits are estimates only, since the absence of an absorption line at the higher velocity is not conclusive evidence that the continuum source is nearer than the HI emission. A region of 21-cm emission may not show absorption because the cool gas may have a covering factor less than one (i.e. there may be gaps between absorbing clouds, filled only with warm gas which does not show detectable absorption). This is unlikely for an emission region with column density of $3 \times 10^{20}~{\rm cm^{-2}}$ or more, which is implied by a line of brightness temperature greater than 35 K. Emission lines stronger than this almost always show some absorption, so our upper limit distances should be mostly valid. We assume that the velocity errors are dominated by random cloud motions on the order of 6 [@dickey97]. Velocity limits and corresponding kinematic distances are given in Table \[tab:abs\] for sources brighter than $\sim 800~{\rm mJy~beam^{-1}}$. Included in the table for comparison are the radio recombination line velocities from CH87. Some individual sources are discussed in detail later. The absorption velocities in Table \[tab:abs\] clearly lie in two dominant distributions, one centered around $v=-50$  and another centered around $v=-90$ . Calculated isovelocity contours from the @fich89 rotation curve are plotted in Fig. \[fig:diag\] on the @taylor93 model of Galactic spiral arms to show the velocities covered by individual spiral arms and the range of expected velocities for any given line of sight. The Test Region line of sight is marked by the wedge. Fig. \[fig:diag\] shows that gas between $v=-20$  and $v=-30$ is located in the Sagittarius-Carina arm. The position of this arm in velocity space is well traced by  regions [@georgelin76; @caswell87]. Gas at velocities between $v=-50$  and $v=-75$  is located within the Scutum-Crux arm. As noted in CH87 there are many features around $v=-90$  for which the correspondence to a spiral feature is unclear. @georgelin76 assign these  regions to the Norma arm. At $l\approx 327\arcdeg$ the line of sight is nearly tangent to this arm, which accounts for the large number of sources seen there. The distribution centered at $v=-90$  has a large spread in velocity space, extending as far as the terminal velocity near $v=-110$ . This large velocity spread can also be explained by the line of sight remaining in the spiral arm for a significant distance near the arm tangent. Using  regions and diffuse H$\alpha$ emission, @georgelin94 similarly note velocity distributions at $v=-20$ , $v=-40$ , and $v=-65$ . They do not, however, detect H$\alpha$ emission near $v=-90$  because it is beyond the extinction limit. From the  absorption we determine new kinematic distances to two supernova remnants and confirm distances to a further one SNR and nine   regions. For SNR G328.4+0.2 our absorption spectrum looks very similar to @gaensler00b, who also found an extreme velocity of $v=28$ . The  region velocities all correspond with the velocities given in CH87. The new kinematic distances for SNR G327.4+0.4 and SNR G330.2+1.0 are given in Table \[tab:abs\] and are discussed in detail below. Distance Ambiguities for Individual Sources {#subsec:ambi} ------------------------------------------- Many of the sources presented here have distance ambiguities. Sources in the fourth quadrant with negative velocities are found inside the solar circle where each velocity corresponds to two distances. There are several methods for distinguishing between the two distances. H$\alpha$ emission is often used as an indicator, as it is severely absorbed at far distances. One can also make rather uncertain estimates based on associations with nearby objects of known distance, the emission spectrum towards the object, or the more likely linear size and luminosity. For  regions with recombination velocities, comparison of the  absorption velocity with the recombination line velocity can resolve the ambiguity [@kuchar94]. If the most extreme  absorption is at or near the recombination line velocity then the cloud is at the near distance. However, if the   absorption is seen beyond the recombination line velocity then the   region is at the far distance. For all of the  regions presented here we have recombination line velocities from CH87. In all cases there is no absorption significantly beyond the recombination line velocity, implying the near distance for these regions. The @georgelin94 H$\alpha$ survey of this region also resolved many of the distance ambiguities through associations with stellar distances. In particular, @georgelin94 favor the near distance for the star forming region associated with RCW 94 and RCW 95, as discussed below.  Emission Features {#sec:emis} ================== The RCW 94-95  Region Complex {#subsec:rcw} ----------------------------- As described in Section \[sec:cont\], the Test Region contains many catalogued  regions. RCW 94, RCW 95, and G326.65+0.59 are part of a large star-forming complex in the Scutum-Crux arm. ${\rm H_2CO}$ and hydrogen recombination line (H$109\alpha$ & H$110\alpha$) velocities of $v=-42$  and $v=-45$ have been measured for RCW 94 and 95, respectively (CH87). We have extracted  absorption profiles towards both RCW 94 and G326.65+0.59 (see Fig. \[fig:rcw94\_abs\]) which confirm these velocities, showing deep absorption out to $v=-47$  and $v=-43$ , respectively. The most extreme absorption line for G326.65+0.59, centered at $-46$  is slightly broader ($\Delta v \approx 15$ ) than that for RCW 94 ($\Delta v \approx 9$ ), suggesting a more turbulent region [@shaver79]. @georgelin94 also note the differing line widths of these two regions. We adopt the IAU standard values for the Sun’s orbital velocity, $\Theta_o = 220$ , Galactic center distance, $R_o = 8.5$ kpc, and use the rotation curve of @fich89 to calculate distances. Assuming a common velocity of $v=-45\pm6$ , we find distances of $3.1\pm 0.3$ kpc or $11.1\pm0.3$ kpc. @georgelin94 identify two populations of  regions within the Scutum-Crux arm, one at $v=-40$  and another at $v=-50$ . While they associate the RCW 94-95 complex with the $v=-40$  population, our absorption velocities indicate absorption out to nearly $v=-50$ . In order to resolve the distance ambiguity @georgelin94 have identified stars of spectral type O to B3 in the vicinity of  regions near $l=328\arcdeg$. They identify an O6f star, LSS 3386, in the vicinity of RCW 94 at a distance of $2.3$ kpc. They also identify an O7V star, BDMW123, at a distance of 3.3 kpc near RCW 95. While it is not clear whether these stars are [*the*]{} ionizing stars for these regions, their presence does seem to indicate a preference for the near distance of the distance ambiguity. We note, also, that at a distance of $3.1$ kpc RCW 94 has a physical diameter of $\sim 17$ pc, a typical diameter for an extended  region. Whereas at a distance of 11.1 kpc, the physical diameter would be $\sim 62$ pc, unusually large for an  region. The diameter at the larger distance, as well as the nearby massive stars noted in @georgelin94, are evidence favoring a distance of $3.1$ kpc. The  emission morphology in the RCW 94-95 region between velocities of $-35$  and $-50$  is complicated. We have detected an  shell surrounding RCW 94 centered at $v=-38$ . There is a ridge of centered at $l=326\fdg3$, $b=+0\fdg8$ that lies just outside the continuum emission contours. This shell is shown in Fig. \[fig:rcw94\], where the greyscale is the  channel image at $v=-38$  and the contours are 21-cm continuum emission. The shell is apparent from $v=-35$  to $v=-42$ . The shell itself is surrounded by a ring of decreased emission which, although not continuous, is also centered on the region. The emission shell has an average diameter of $\sim 24$ pc, a thickness of $\sim 5$ pc, and shows a brightness temperature increase from the interior to the shell edge of about a factor of two. The  shell morphology closely matches the morphology of the  regions, implying that the shell is indeed related to RCW 94. The lack of  inside to the shell is clearly due to ionization in the  region. The origins of the shell are somewhat less clear. Using the column density integrated through the range of velocities including the shell we estimate that the  mass of the shell is $\sim 170~{\rm M_{\odot}}$. That the shell extends over $\sim 7$  suggests that it may be expanding. The velocity gradient at this place in the Galaxy is $\sim 18~{\rm km~s^{-1}~kpc^{-1}}$. Therefore a static shell with velocity width $\Delta v \approx 7$  would have an extent of 380 pc along the line-of-sight. Since it is highly unlikely that the shell extends that far, we suggest that the velocity width is due to expansion such that $v_{exp} = \Delta v / 2$. Assuming $v_{exp}\sim4$ , we estimate the energy required to form this shell is on the order of $\sim 10^{51}$ ergs, which is consistent with the amount of energy expected from stellar winds over the lifetime of a single massive star. Because of the low expansion velocity, the formation energy for a shell whose expansion has stalled is comparable. We suggest that the  shell and depression around RCW 94 are the signatures of a molecular cloud encircling the  region. In this case, the  region appears to be embedded in a molecular cloud, displaying various stages of ionization and dissociation related to the interior stars. Interior to the $\sim 24$ pc inner shell radius the UV photons from the stars ionize the neutral gas, producing the  region. The stars photo-dissociate the surrounding molecular gas, producing an  shell which extends to a radius of $\sim 29$ pc. The  morphology correspondence with the continuum morphology especially supports this hypothesis. In particular, the region of dense  emission in the concave portion on the right-hand side of the  region indicates that the expansion of the photo-dissociation region (PDR) was impeded by a density enhancement in the external medium, presumably clumps of molecular material. The extension of the shell surrounding the compact source to the upper left, indicates that the shape of the shell is directly related to the shape of the  region, and that they are therefore correlated. Comparison with the CO images of @bronfman89 indicates molecular gas at the position and velocity of RCW 94. Immediately exterior to the  shell we can expect to see emission from polycyclic aromatic hydrocarbons (PAHs) at $6.2, 7.7, 8.6,$ or 11.3  [@simpson99]. Close examination of MSX band A data does reveal an increase in emission exterior to the  shell, which also supports the theory that this  region and its  shell are embedded in a molecular cloud. @gaensler00a explore the polarization properties of this region and find that the depolarization is consistent with being caused by an  region embedded in molecular gas with several layers of ionization and photo-dissociation. The  depression around RCW 94 appears to extend towards the Plane at lower longitudes where it traces the morphology of the large bow shaped structure, G326.96+0.03, seen at the bottom of the continuum image. This source is seen in the MOST images, as well as the MSX images, and therefore appears to be a thermal source. We measure  absorption towards the knot of emission at $l=326\fdg95$, $b=+0\fdg02$. Though the spectrum is rather noisy (Fig. \[fig:g327abs\] left), we see a strong absorption feature at $v=-60$ , which indicates that this region may be slightly more distant than RCW 94-95 ($d=3.9$ kpc), though still in the Scutum-Crux arm. There is an extended region of  emission to the left of these  regions which is much brighter than that surrounding RCW 94-95. The column density in this region, over the range of channels spanning the depression ($v=-38.4~{\rm km~s^{-1}}$ to $v=-33.45~{\rm km~s^{-1}}$), is a factor of two larger than it is surrounding the  shell. HI Shells {#subsec:shells} --------- Whereas  regions and SNRs draw a connection between the  line and continuum emission, the impact of massive stars on the ISM can also be seen with  shells, where no continuum object exists. These cavities survive much longer than the radiative lifetime of a SNR or an  region, allowing us to explore the lasting effects of massive stars on the ISM.  shells are often detected as voids in the  with brightened “walls” of swept-up material. These shells can range in size from tens of parsecs to kiloparsecs. The majority of the shells, especially the smaller ones, are caused by the combined effects of stellar winds and supernovae [@heiles84]. The ultimate destruction of an  shell occurs on the time-scale of tens of millions of years when they eventually dissipate as a result of turbulent motions in the ISM and shear due to differential rotation in the Galaxy. We have detected two small shells in the SGPS Test Region [@mcgriff00a]. The first of these appears as an  void at $l=329\fdg3$, $b=+0\fdg4$, $v=-108$ , the terminal velocity for this line of sight. The velocity implies a kinematic distance of $7.3$ kpc. Fig. \[fig:term\] is an  channel image at $v=-108$  showing a small shell of angular diameter is $\sim 0\fdg4$. At a distance of $7.3$ kpc the shell has a physical diameter of $\sim 50$ pc. Because of its position at the terminal velocity it is very difficult to distinguish the front and back caps, we detect only the front cap. It is not unusual to detect only one cap, though. There is only one detectable cap for a large majority of the shells catalogued by @heiles84. Detecting only one cap makes it difficult to estimate an expansion velocity. It may be that the shell is stalled or that the structure is mostly cylindrical and expanding in the plane of the sky. Though we cannot measure the expansion velocity, we interpret this structure as a stalled wind or supernova blown shell. The second shell is observed in the local gas at $v=-2.1$ , $l=330\fdg5$, and $b=+2\fdg12$. This shell is shown in Fig. \[fig:local\], a channel image of the  at $v=-2.1$ . The shell is remarkably circular with an angular diameter of $\sim 2\fdg5$. Because of its low velocity its distance is very uncertain, we estimate $D= 350 - 500$ pc, which implies a physical radius of only $\sim 15$ pc. Given its small size we speculate that this shell may have been formed by an old SNR. There are no associated features in the continuum image. Supernova Remnants {#subsec:snrs} ------------------ ### SNR G327.4+0.4 {#subsubsec:g327.4}  studies of supernova remnants offer a great deal of information.   absorption spectra allow us to place limits on the distances to SNRs, which lead to physical radii as well as age estimates. Examination of related  emission structures may help us to understand why shell-type remnant morphology is dominated by loops and knots, in addition to diffuse emission. The morphology of SNRs is undoubtedly related in some complicated way to the inhomogeneities in the ISM into which they expand, as well as to non-uniform magnetic fields. One might expect SNR continuum emission to trace the local ISM in such a way that bright emission may be correlated with density enhancements exterior to the SNR shell. In principle these ISM density enhancements would be apparent as brightness temperature enhancements in the neutral hydrogen at velocities similar to the systemic velocity of the remnant. However, in practice it has proven difficult to correlate the continuum emission with emission structures in the  (eg. Giacani et al.2000). We have extracted an absorption spectrum towards SNR G327.4+0.4 as shown in Fig. \[fig:g327abs\], right side. The positive and negative wiggles in the range $-85$ to $-95$  are a characteristic signature of variations in the terminal velocity, and cannot be trusted as real absorption. There is a strong absorption line centered at $v=-48$ , and a weaker, noisy line at about $-70$ . There is no absorption corresponding to the emission peak at $-80$ . We therefore adopt$V_{\rm L}=-70$  and $V_{\rm U}=-80$ . These velocities indicate a distance of $4.3\pm0.5$  kpc. This places the remnant on the far side of the Scutum-Crux Arm. SNR G327.4+04 is a multi-arc shell-type SNR. The continuum emission has a bright, sharply bounded rim to the lower left, while the emission on the upper right side is much more diffuse. If the limb brightening observed to the lower left is a consequence of the shock impacting a density enhancement, we might expect to see an  cloud exterior to the shell at the systemic velocity of the remnant. This is confirmed in the  channel images at $v\approx -70$ , where we see a ridge of emission just exterior to the continuum contours. Fig. \[fig:snr327.4\] is an average of two velocity channels centered at $v=-70$  with continuum contours overlaid. There is an increase in  density just exterior to the brightest portion of the SNR. In addition, the  to the upper right, exterior to the less bright edge of the remnant, is much more diffuse. In this case it does appear that the continuum morphology is related to the surrounding . SNR G327.4+0.4 has an angular diameter of $\sim 15\farcm3$, yielding a physical radius of $10.5 \pm 0.6$ pc at $4.3$ kpc. To reinforce the point that this remnant must be at the near distance, we note that at $10.0$ kpc the physical radius of the SNR would be an unusually large 22 pc. The radius allows us to estimate some fundamental parameters for the SNR. If the density of the medium into which the SNR expands is given by $n_o$ (in ${\rm cm ^{-3}}$), the mass swept up by the SNR is $\sim 120 n_o~{\rm M_{\odot}}$. As this mass is only a few times the presumed mass of the progenitor, we believe that the SNR is undergoing adiabatic expansion, but that it has only recently left the free expansion phase of evolution. Using the standard assumption that the SNR is in the Sedov-Taylor phase, we estimate the age $t_{\rm SNR} = (5.3\pm 0.8) (n_o/E_{51})^{1/2} \times 10^3$ yr, where $E_{51}$ is the input energy of the supernova explosion in units of $10^{51}$ ergs. If we assume typical values of $n_o=0.2$ and $E_{51}=1$ [@frail94], we find $t_{\rm SNR} = (2.4 \pm 0.3) \times 10^3$ yr. ### SNR G330.2+1.0 {#subsubsec:g330.2} We obtain an absorption spectrum towards SNR G330.2+1.0 which shows absorption out to $v=-80\pm 6$ , indicating a minimum distance of $d=4.9\pm 0.3$ kpc for the supernova remnant. As noted in §\[sec:cont\], this SNR has no clearly defined shell. There is continuum emission surrounding the brightened center. It is not clear how much of that emission is associated with the SNR. Fig. \[fig:snr330.2\] is an image at $v=-80$ overlaid with 21-cm continuum contours of SNR G330.2+1.0. At the low longitude end of G330.2+1.0 the lowest continuum contour extends in an arc away from the SNR center. Spatially offset from this edge the  emission follows the same arc. Similarly, the continuum contour closest to the Plane is bounded by  emission which traces the contour. The morphological similarities between the lowest continuum contour and the  emission to the right of the remnant suggests a possible correlation between the two. The overlap of  from two distances at this velocity, as well as the small angular size of the SNR make it difficult to confirm whether the apparent -continuum correlation is real. Conclusions {#sec:concl} =========== We have presented  and $\lambda 21$-cm data from the SGPS Test Region ($325\fdg5 \leq l \leq 333\fdg5$; $-0\fdg5 \leq b \leq 3\fdg5$), which are representative of the full Survey. These results highlight the interesting effects of massive stars on the ISM. The SGPS is ideal for studying the structure and dynamics of the  in the inner Galaxy as it is sensitive to a large range of angular scales ($2\arcmin \lesssim \theta \lesssim 2\arcdeg$). In the Test Region we have explored the  associated with three products of massive star life and death:  regions, SNRs, and shells. Using  absorption for systemic velocities and corresponding kinematic distances of  regions and SNRs we are able to create a three dimensional picture of the distribution of the continuum sources in this region of the Galaxy. We have highlighted several interesting  and $\lambda 21$-cm continuum emission features from the Test Region. The features in the continuum image include extended emission structures and a possible new SNR G328.6-0.0. Comparing the  and continuum, we found an  shell around the region RCW 94 which indicates that the  region is embedded in a molecular cloud. In this case we see the reciprocal effects of massive stars and the surrounding  during the stellar lifetime. The continuum emission morphology of the  region closely matches the morphology of the surrounding . We use  absorption towards the SNRs G327.4+0.4 and G330.2+1.0 to determine kinematic distances of 4.3 and 4.9 kpc, respectively.  at the systemic velocity of these remnants shows morphological similarities to the continuum emission. In particular, density enhancements were found exterior to regions of continuum limb-brightening for G327.4+0.4. We also found two small  shells with no counterparts in continuum emission. We use the sizes and lack of detectable expansion velocity to interpret these structures as stalled supernova or wind blown shells which are older than the radiative lifetimes of either  regions or SNRs. Deciphering  structure has always been challenging, but the recent availability of high resolution Galactic surveys such as the SGPS has improved the situation dramatically. Much of the inner Galaxy is completely filled with a variety of  structures including shells, worms, sheets, and filaments. Though it is extremely difficult to determine the origins of many of the structures using  emission data alone, combination with   absorption and radio continuum emission measurements enables us to determine a three-dimensional, dynamical picture of the ISM. We thank Veta Avedisova for supplying us with her extensive catalogue of star formation regions. This research has made use of the CDS SIMBAD database. JMD and NMM-G acknowledge support of NSF grant AST-9732695 to the University of Minnesota. NMM-G is supported by NASA Graduate Student Researchers Program (GSRP) Fellowship NGT 5-50250. BMG acknowledges the support of NASA through Hubble fellowship grant HST-HF-01107.01-A awarded by STScI, which is operated by AURA Inc. for NASA under contract NAS 5-26555. Avedisova, V. S., 1997, Baltic Astronomy, 6, 307 Bronfman, L, Alvarez, H., Cohen, R. S., & Thaddeus, P., 1989, , 71, 481 Caswell, J. L. & Haynes, R. F., , 171, 261 Dickey, J. M., McClure-Griffiths, N. M., Gaensler, B. M., Green, A. J., Haynes, R. F., Wieringa, M. H., 2000, in preparation Dickey, J. M., McClure-Griffiths, N., Gaensler, B., Green, A., Haynes, R., Wieringa, M., 1999, in New Perspectives on the Interstellar Medium, ASP Conf. Ser. 168, eds. A. R. Taylor, T. L. Landecker & G. Joncas, (San Francisco: ASP), 27 Dickey, J. M., 1997, , 488, 258 Dickey, J. M., Brinks. E., & Puche, D., 1992, , 385, 501 Egan, M. P., Shipman, R. F., Price, S. D., Carey, S. J., Clark, F. O., & Cohen, M. 1998, ApJ, 494, L199 Fich, M., Blitz, L., & Stark, A. A., 1989, , 342, 272 Frail, D. A., Goss, W. M., Whiteoak, J. B. Z., 1994, , 437, 781 Frail, D. A., & Weisberg, J. M., 1990, , 100, 743 Frater, R. H., Brooks, J. W., & Whiteoak, J. B., 1992, J. Electr. Electron. Eng. Aust., 12, 103 Gaensler, B. M., Dickey, J. M., McClure-Griffiths, N. M., Green, A. J., Wieringa, M. H., Haynes, R. F., 2000a, ApJ, 549, in press Gaensler, B. M., Dickel, J., Green, A. J., 2000b, , 542, 380 Georgelin, Y.M. & Georgelin, Y. P., 1976, , 49, 57 Georgelin, Y. M., Amram, P., Georgelin, Y. P., le Coarer, E., & Marcelin, M., 1994, , 108, 513 Giacani, E. B., Dubner, G. M., Green, A. J., Goss, W. M. & Gaensler, B. M. 2000, , 119, 281 Gooch, R. E., 1996, in Astronomical Data Analysis Software and Systems V, ASP Conf. Ser. 101, ed. G. H. Jacoby & J. Barnes, (San Francisco: ASP), 80 Goss, W. M. & Shaver, P. A., 1970, AuJPA, 14, 1 Green, A. J., Cram, L. E., Large, M. I., & Ye, T., 1999, , 122, 207 Green, D. A, 2000, A Catalogue of Galactic Supernova Remnants (2000 August version), (Cambridge: Mullard Radio Astronomy Observatory, Cavendish Laboratory) (http://www.mrao.cam.ac.uk/surveys/snrs/) Haynes, R. F., Staveley-Smith, L., Mebold, U., Kalberla, P., White, G., Jones, P., Dickey, J., Green, A., 1998, in New Views of the Magellanic Clouds, IAU Symp. 190, eds.Y.-H. Chu, N. Suntzeff, J. Hesser, & D. Bohlender, (San Francisco: ASP), 108 Heiles, C., 1984, , 55, 585 Kalberla, P. M. W., Mebold, U., and Reich, W., 1980, A&A, 82, 275 Kuchar, T. A. & Bania, T. M., 1994, , 436, 117 McClure-Griffiths, N.M., Dickey, J. M., Gaensler, B. M., Green, A. J., Haynes, R. F., Wieringa, M. H., 2000a, PASA, accepted McClure-Griffiths, N.M., Dickey, J. M., Gaensler, B. M., Green, A. J., Haynes, R. F., Wieringa, M. H., 2000b, , 119, 2828 McClure-Griffiths, N. M., Dickey, J. M., Gaensler, B. M., Green, A. J., Haynes, R. F., Wieringa, M. H., 1999, , 194, 0402 Rodgers, A. W., Campbell, C. T., Whiteoak, J. B., 1960, , 121, 103 Reynolds, J. E., 1994, ATNF Technical Document Series, (Sydney: Australia Telescope National Facility), AT/39.3/0400 (http://www.narrabri.atnf.csiro.au/observing/users\_guide/html/node215.html) Sault, R. J., Staveley-Smith, L. & Brouw, W. N., 1996, , 120, 375 Sault, R. J., Killeen, N.E. B., 1998, The Miriad User’s Guide, (Sydney: Australia Telescope National Facility) (http://www.atnf.csiro.au/computing/software/miriad/) Shaver, P. A., McGee, R. X., & Pottasch, S. R., 1979, , 280, 476 Simpson, J. P., Witteborn, F. C., Cohen, M., Price, S. D., 1999, in The Central Parsecs of the Galaxy, ASP Conf. Ser. 186, eds. H. Falcke, A. Cotera, W. J.Duschl, F. Melia, & M. J. Rieke, (San Francisco: ASP), 527 Stanimirović, S., 1999, PhD Thesis, University of Western Sydney, Nepean Stanimirović, S., Staveley-Smith, L., Dickey, J. M., Sault, R. J., & Snowden, S. L., 1998, , 301, 417 Staveley-Smith, L., Wilson, W. E., Bird, T. S., Disney, M. J., Ekers, R. D., Freeman, K. C., Haynes, R. F., Sinclair, M. W., Vaile, R. A., Webster, R. L., Wright, A. E., 1996, PASA, 13, 243 Taylor, A. R., 1999, in New Perspectives on the Interstellar Medium, ASP Conf. Ser. 168, eds. A. R. Taylor, T. L. Landecker & G.  Joncas, (San Francisco: ASP), 3 Taylor, J. H. & Cordes, J. M., 1993, , 411, 674 Weaver, H. F. & Williams, D. R. W., 1974, , 17, 251 Whiteoak, J. B. Z. & Green, A. J., 1996, , 118, 329 Whiteoak, J. B.Z., Cram, L. E., & Large, M. I., 1994, , 269, 294 Williams, D., 1973, , 8, 505 [lll]{} 1997 Apr 21 & 0.375 & 09:22 - 22:14\ 1997 Apr 22 & 0.375 & 08:00 - 22:05\ 1997 Apr 23 & 0.375 & 12:11 - 22:03\ 1997 Aug 11 & 0.75B & 05:33 - 16:06\ 1997 Aug 13 & 0.75B & 08:33 - 15:51\ 1997 Oct 25 & 0.75C & 20:53 - 11:12\ 1998 Mar 30 & 0.375 & 10:00 - 20:40\ 1998 Apr 20 & 0.75A & 08:06 - 20:59\ 1998 Apr 21 & 0.75A & 07:44 - 21:02\ 1998 Dec 14-15 & Parkes & 20:16 - 05:21\ 1998 Dec 15-16 & Parkes & 20:22 - 06:33\ 1998 Dec 16 & Parkes & 14:56 - 22:30\ [llrcccccc]{} RCW 94 & 326.45 & $+0.91$ & $-42$ & $-48$ & $-45$ & 2.9 & 3.3 & 6.2\ G326.65+0.59 & 326.66 & $+0.59$ & $-47$ & $-60$ & $-44$ & 3.2 & 3.9 & 6.0\ G326.96+0.03 & 326.95 & $+0.02$ & $-57$ & $-62$ & $-64$ & 3.7 & 4.0 & 5.6\ SNR G327.4+0.4 (Kes 27)& 327.34 & $+0.40$ & $-67$ & $-82$ & & 4.3 & 5.4 & 5.2\ G327.99-0.09 & 327.99 & $-0.09$ & $-52$ & $-68$ & $-45$ & 3.5 & 4.3 & 5.7\ G328.31+0.45 & 328.30 & $+0.44$ & $-96$ & $-101$ & $-97$ & 6.0 & 6.5 & 4.6\ SNR G328.4+0.2 (MSH 15-57)& 328.42 & $+0.22$ & $+28$ & & & 17.4 & & 11.1\ G329.35+0.14 & 329.34 & $+0.14$ & $-103$ & $-109$ & $-107$ & 6.4 & 7.3 & 4.4\ G329.49+0.21 & 329.47 & $+0.21$ & $-100$ & $-109$ & $-102$ & 6.1 & 7.3 & 4.5\ SNR G330.2+1.0 & 330.21 & $+0.97$ & $-80$ & & & 4.9(9.9) & & 4.9\ G331.03-0.15 & 331.05 & $-0.16$ & $-95$ & $-100$ & $-89$ & 5.5 & 5.9 & 4.5\ G331.26-0.19 & 331.27 & $-0.19$ & $-89$ & $-100$ & $-85$ & 5.3 & 5.9 & 4.5\ G331.52-0.07 & 331.52 & $-0.08$ & $-92$ & $-102$ & $-89$ & 5.5 & 6.0 & 4.4\ [^1]: The Parkes telescope and the ATCA are part of the Australia Telescope, which is funded by the Commonwealth of Australia for operation as a national facility managed by the Commonwealth Scientific and Industrial Research Organisation.
{ "pile_set_name": "ArXiv" }
--- abstract: 'The direct nuclear X–ray emission of the ultraluminous IRAS galaxy NGC 6240 has been observed for the first time by [*Beppo*]{}SAX. It is seen through an absorber with $N_{\rm H} \sim 2\times 10^{24}$ cm$^{-2}$. The 2-10 keV X–ray luminosity of $>10^{44}$ erg s$^{-1}$ definitely proves that a powerful AGN is hosted in this galaxy, despite the LINER optical appearance. Assuming the typical $L_{\rm X}/L_{\rm bol}$ ratio for quasars, our result implies that it is the AGN and not the starburst that dominates the energy output of NGC 6240.' author: - 'P. Vignati' - 'S. Molendi' - 'G. Matt' - 'M. Guainazzi' - 'L.A. Antonelli' - 'L. Bassani' - 'W.N. Brandt' - 'A.C. Fabian' - 'K. Iwasawa' - 'R. Maiolino' - 'G. Malaguti' - 'A. Marconi' - 'G.C. Perola' date: 'To appear in A&A Letters' title: BeppoSAX unveils the nuclear component in NGC 6240 --- Introduction ============ One of the major open problems with Ultraluminous Infrared Galaxies (ULIRG) is whether they host active nuclei at their centers. In at least a fraction of objects, there is circumstantial evidence that this is the case. One of the best examples in this respect is NGC 6240[^1], a well known source in which the presence of two nuclei (Fried & Schulz 1983) suggests an ongoing galaxy merger. Indications of the presence of an AGN come mainly from X–rays: the flat 2–10 keV X–ray spectrum and the prominent iron line complex observed by ASCA (Iwasawa & Comastri 1998) are very similar to those of NGC 1068 (Iwasawa, Fabian & Matt 1997; Matt et al. 1997 and references therein) and suggest reflection from cold and warm matter of an otherwise invisible nuclear component. On the other hand, in the optical spectrum all the diagnostic line ratios point towards a classification as a LINER (e.g. Veilleux et al. 1995), and the ISO SWS diagnostic diagram places NGC 6240 in the region of star formation dominance (Genzel et al. 1998). In soft X–rays the spectrum is fairly complex, and best modeled by multi-temperature plasma emission, very likely associated, at least partly, with the powerful starburst and related wind present in this galaxy, a hypothesis strengthened by the ROSAT HRI discovery of rather luminous ($\sim$10$^{42}$ erg cm$^{-2}$ s$^{-1}$) emission extended over tens of kpc (Komossa, Schulz & Greiner 1998; Schulz et al. 1998). On the other hand, the point–like soft X–ray component is possibly due to scattering of the nuclear radiation (Komossa et al. 1998). A direct measurement of the nuclear luminosity is needed to assess the relative importance of AGN and starburst emission. The only way to do that is to explore the hard X–ray band, which may reveal the direct nuclear component. If the absorbing matter is moderately thick to Compton scattering (i.e. $N_{\rm H}\sim$a few$\times10^{24}$ cm$^{-2}$), the photons emerge above 10 keV, where the most sensitive instrument presently available is the PDS onboard [*Beppo*]{}SAX. This has already added two sources (Circinus Galaxy, Matt et al. 1999; Mrk 3, Cappi et al. 1999) to the small list of “moderately–thick" Seyfert 2s, which was previously composed of only one object, NGC 4945 (Iwasawa et al. 1993; Done et al. 1996). Observation and data reduction ============================== [*Beppo*]{}SAX (Boella et al. 1997) observed NGC 6240 on 1998 August 14–17. In this paper, data from three instruments are discussed: the LECS (0.1–10 keV), the MECS (1.5–10 keV), and the PDS (13–300 keV). Data reduction and analysis follow the standard criteria, as described in Matt et al. (1997). The total effective exposure time was 52.5 ks in the LECS, 119.4 ks in the MECS and 114.4 ks in the PDS. We accumulated and examined images for the MECS instruments. Comparing the radial profiles in three energy ranges (1.5–4, 4–7.5, 7.5–10 keV) with the corresponding Point Spread Function, no evidence for extended emission is found. Light curves and spectra were extracted within regions of 4$^{\prime}$ and of 2$^{\prime}$ radius centered on the source for MECS and LECS, respectively (the small LECS extraction radius is chosen to avoid contamination from a nearby, soft source, not visible in the MECS). The light curves are consistent with being constant. The resulting count rate is (8.46$\pm$0.67)$\times$10$^{-3}$ counts s$^{-1}$ (0.1–4.5 keV) for the LECS and (2.46$\pm$0.07)$\times$10$^{-2}$ counts s$^{-1}$ (1.5–10 keV) in the MECS. No significant variability was found in the PDS light–curve either, with a count rate of 0.38$\pm$0.04 counts s$^{-1}$. We restricted the spectral analysis to the 0.1–4.5 keV and 1.5–10 keV energy bands for LECS and MECS respectively, where the latest released (September 1997) response matrices are best calibrated. Standard blank–sky files provided by the [*Beppo*]{}SAX Science Data Center (SDC) were used for the background subtraction. Spectral analysis and results ============================= In Fig. 1 we compare the $>$2 keV spectra of NGC 6240 and NGC 1068. Both spectra have been normalized to the Crab spectrum; since the latter is a power law with photon index $\simeq$ 2, we are basically showing the $\nu F_{\nu}$ spectrum. The MECS spectra of the two sources are similar in shape, not surprisingly as we know that in this energy range both sources have been successfully fitted by two reflection components, warm and cold (e.g. Iwasawa et al. 1997; Iwasawa & Comastri 1998). While below 10 keV the flux of NGC 6240 is the lowest, the opposite occurs in the PDS band. Clearly, in NGC 6240 a new component emerges at high energies. No known bright sources are present in the field of view of the PDS, and the probability of a serendipitous source in the $1.3^{\circ}$ PDS field with a flux equal or larger than that of NGC 6240 is $\simlt 10^{-5}$, if the 2–10 keV ASCA LogN–LogS (Cagnoni et al. 1998) is adopted, and a power law spectrum with photon index 1.8 is assumed. Thus the best explanation for this component is nuclear emission piercing through an absorber with column density $N_H$ a few$\times$10$^{24}$ cm $^{-2}$, i.e. Compton–thick but still permitting partial transmission above $\sim$10 keV. In Fig. 2, the broad band spectrum is fitted following Iwasawa & Comastri (1998): an excess is evident in the PDS band. We therefore modeled the continuum above 2 keV with: [*i*]{}) a power law, representing the intrinsic emission, with a photon index fixed at a typical value of $\Gamma$=1.8, absorbed by a column density allowed to be free (leaving the index as a free parameter, a value consistent with the adopted one, but very loosely constrained, is found); [*ii*]{}) a second, unabsorbed power law with the same index, standing for the warm reflection component; [*iii*]{}) a pure cold reflection component (PEXRAV model in XSPEC) generated by a power law with the same index. As we are interested here in the high energies, and because below 2 keV ASCA is superior to [*Beppo*]{}SAX as far as both sensitivity and spectral resolution are concerned, we content ourselves with adopting in this band the best fitting model of Iwasawa & Comastri (1998). It consists of a multi-temperature optically thin plasma, which probably originates in a powerful starburst. In particular we use a two–temperature MEKAL model, where the absorption of the coldest MEKAL was fixed at the Galactic value (N$_{\rm H}$ = 5.8$\times$ 10$^{20}$ cm$^{-2}$), while the absorption of the other MEKAL was allowed to be a free parameter. We applied the latter absorption to all the AGN components. The envisaged scenario is then one in which an external starburst region, associated with the cooler MEKAL component, absorbs emission from the inner regions of NGC 6240 (e.g. Fabian el al. 1998). A prominent and broad iron K line complex, very similar to that seen in NGC 1068, is also evident. Because of this similarity we assume that the line is actually a blend of: the 6.4 keV K$\alpha$ and 7.06 keV K$\beta$ fluorescent lines from cold iron (the intensity of the latter has been forced to be equal to 1/9 that of the former, as expected from atomic physics); the 6.7 and 6.97 K$\alpha$ recombination lines from He– and H–like iron. In summary the complete model, $F$, we used to fit our data can be written as: $$F = A_{\rm G}\{A_{\rm SB}[A_{\rm T}(PL) + R_{\rm W} + R_{\rm C} + M_{\rm H} + 4GL] + M_{\rm C}\},$$ where: $A_{\rm G}$ is the absorption associated with the Galactic column, $A_{\rm SB}$ is the absorption related to the starburst, $A_{\rm T}$ is the absorption acting on the nuclear emission (including the Compton cross section), possibly associated with the molecular torus shrouding the nucleus, $PL$ is the power–law modeling the nuclear component, $R_{\rm W}$ and $R_{\rm C}$ are respectively the warm (optically thin) and cold (optically thick) reflection components, $M_{\rm H}$ and $M_{\rm C}$ are respectively the cold and hot thermal components and $4GL$ are the four Gaussian lines modeling the iron blend. [|c|c|c|]{} Component & Parameter & Value   &   &  $A_{\rm SB}$ & N$_{\rm H}$ (10$^{22}$ cm$^{-2}$) & 1.99$^{+1.23}_{-0.67}$ $A_{\rm T}$ & N$_{\rm H}$ (10$^{22}$ cm$^{-2}$) & 218$^{+40}_{-27}$ $PL$ & $\Gamma$ & 1.8 (fixed)   & Flux at 1  keV$^a$ & 4.1($^{+1.3}_{-1.3}$)$\times10^{-2}$ $GL1$ & E$_{line}$ & 6.4 (fixed)   & I$_{line}$$^b$ & 2.23($^{+0.56}_{-0.51}$)$\times10^{-5}$ $GL2$ & E$_{line}$ & 6.7 (fixed)   & I$_{line}$$^b$ & 0.70($^{+0.74}_{-0.69}$)$\times10^{-5}$ $GL3$ & E$_{line}$ & 6.97 (fixed)   & I$_{line}$$^b$ & 0.53($^{+0.45}_{-0.43}$)$\times10^{-5}$ $GL4$ & E$_{line}$ & 7.06 (fixed)   & I$_{line}$$^b$ & 0.11$\times$I$_{line}$(6.4 keV) $R_{\rm W}$ & $\Gamma$ & 1.8 (fixed)   & Flux at 1  keV$^a$ & 0.24($^{+0.09}_{-0.09}$)$\times10^{-3}$ $R_{\rm C}$ & $\Gamma$ & 1.8 (fixed)   & Flux at 1  keV$^a$ & 2.4($^{+1.1}_{-1.2}$)$\times10^{-3}$ $M_{\rm H}$ & $\rm kT$ (keV) & 0.51$^{+0.14}_{-0.18}$   & Flux$^c$ & 9.4($^{+4.1}_{-2.2}$)$\times10^{-13}$ $M_{\rm C}$ & $\rm kT$ (keV) & 0.70$^{+0.19}_{-0.25}$   & Flux$^c$ & 1.3($^{+5.1}_{-0.7}$)$\times10^{-11}$   &   &     $^a$Photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$; $^b$Photons cm$^{-2}$ s$^{-1}$; $^c$Bolometric flux in erg cm$^{-2}$ s$^{-1}$ In Fig. 3 we show the deconvolved best fitting model with all the components. The best fit parameters are listed in Table 1. The fit is acceptable, ${\chi}^2$/d.o.f.=131/119. If we do not include the transmission component, and try therefore to account for the spectral component emerging in the PDS band with pure Compton reflection (as in NGC1068, Matt et al. 1997), the fit is significantly worse (${\rm \chi^2 =179/121}$ dof; see Fig. 2). It must be noted that the absorber of the nuclear emission, $A_{\rm T}$, is Compton–thick. A proper model, including Compton scattering within the matter (see Matt et al. 1999 and Matt, Pompilio & La Franca 1999 for details on this model) should have then been used, if this absorber has a significant covering factor (as for example for a torus geometry with no significant vertical density gradient). When this is done, however, the parameters of the various components do not differ significantly from those obtained with the simpler model, apart from the flux associated with the intrinsic emission, which turns out to be about one third that reported in Table 1. This is because a fraction of the photons emitted towards other directions are now scattered into the line of sight, and the nuclear luminosity required to explain the observed flux is therefore lower. The same fit but without $R_{\rm C}$ is also acceptable (${\rm \chi^2 = 133/120}$ dof). In this case, the observed 6.4 keV iron line would entirely originate in the thick absorbing medium, which is possible provided that the covering factor is not much smaller than unity, as we have verified by Monte Carlo simulations. Therefore, the presence of the cold reflection component is not certain in this source. If $R_{\rm W}$ is instead excluded, the fit is significantly worse (${\rm \chi^2 = 139/120}$ dof). Discussion ========== The main result of the [*Beppo*]{}SAX observation of NGC 6240 is the discovery and measurement of the nuclear emission. The 2–10 keV nuclear luminosity, after correction for absorption, is 3.6$\times$10$^{44}$ erg s$^{-1}$ (assuming $H_0$=50 km s$^{-1}$ Mpc$^{-1}$), or about one third of this if the absorber covers a large solid angle, like in a torus. As discussed above, the presence of cold reflection is not unambiguously established in this source. Let us, however, assume that it does exist, and take the best fitting parameters listed in Table 1. The ratio between the 2–10 keV cold reflection and intrinsic fluxes is about 0.3% for the higher value of the intrinsic luminosity and 3 times larger for the lower value. Assuming a torus geometry like that adopted by Ghisellini, Haardt & Matt (1994), and consistently the lower figure for the intrinsic luminosity, we derive an inclination angle of the torus axis with respect to the observer of $\sim$35$^{\circ}$ ($\sim$45$^{\circ}$ if the upper figure is instead adopted). It is worth noting that both NGC 6240 and Circinus Galaxy (Guainazzi et al. 1999) appear to be observed at moderate inclinations, while NGC 1068, which is totally obscured, is possibly seen almost edge-on (Greenhill et al. 1996, Matt et al. 1997; but see Kishimoto 1999 for a different point of view). This would suggest a decrease in the density of the X–ray absorbing medium with increasing distance from the equatorial plane. The ratio between 2–10 keV warm reflection and intrinsic fluxes is about 0.5% (1.5% for the lower luminosity). The He– and H–like lines have EW with respect to the warm reflection component of about 1 and 0.8 keV, respectively. From these values we estimate the column density of the warm mirror (e.g. Matt, Fabian & Brandt 1996; Guainazzi et al. 1999), $N_{\rm H,warm}$, to be between 10$^{21}$ and 10$^{22}$ atoms cm$^{-2}$, roughly consistent with the value estimated from the ratio between the normalizations of the warm reflection and of the intrinsic emission. The 2-10 keV X–ray luminosity is 10 to 30 times lower than the IR luminosity (calculated as prescribed by Mulchaey et al. 1994), consistent with the usual values for Seyferts (e.g. Mulchaey et al. 1994; Awaki 1997). In the case of NGC 6240, therefore, a ULIRG turns out by no means to be a peculiar AGN, as far as X–ray properties are concerned. If $L_{\rm 2-10~keV}/L_{\rm bol}\sim$0.03, as typical for QSO (Elvis et al. 1994), our result implies that it is the AGN that dominates the energy output, and not the starburst as deduced by Genzel et al. (1998) on the basis of the ISO spectrum. In the optical, NGC 6240 is classified as a LINER. Possibly, the NLR is obscured by the starburst region, whose column density, from our analysis, corresponds to A$_{\rm V}\sim$10. It is worth noting that NGC 6240 is not the first LINER that, when observed in X-rays, turns out to be an AGN (see e.g., NGC 4945, Done et al. 1996; NGC 1052, Guainazzi & Antonelli 1999). Differently from them (but similarly to two other ULIRG, IRAS 23060+0505: Brandt et al. 1997 and IRAS 20460+1925: Ogasaka et al. 1997), NGC 6240 has a high X–ray luminosity, and it may be called an obscured QSO. While it would be premature to draw conclusions on ULIRGs as a class, one cannot help noting that the so far elusive type 2 QSO should be better searched for in X–rays rather than in the optical band. We thank the referee, R. Antonucci, for valuable comments. We acknowledge the BeppoSAX SDC team for providing pre–processed event files and the support in data reduction. GM and GCP acknowledge financial support from ASI, WNB from NASA LTSA grant NAG5-8107. MG acknowledges an ESA Research Fellowship. BeppoSAX is a joint Italian-Dutch program. Awaki H., 1997, in “Emission Lines in Active Galaxies: New Methods and Techniques", ASP conf. series Vol 113, p. 44 Boella G., Butler R.C., Perola G.C., et al., 1997, A&AS 112, 299 Brandt W.N., Fabian A.C., Takahashi K., Fujimoto R., Yamashita A., Inoue H., Ogasaka Y., 1997, MNRAS 290, 617 Cagnoni I., Della Ceca R., Maccacaro T., 1998, ApJ 493, 54 Cappi M., Bassani L., Comastri A., et al., 1999, A&A 344, 857 Done C., Madejski G.M., Smith D.A., 1996, ApJ 463, 63 Fabian A.C., Barcons X., Almaini O., Iwasawa K., 1999, MNRAS 297, L11 Fried J.W., Schulz H., 1983, A&A 118, 166 Genzel R., Lutz D., Sturm E., et al., 1998, A&A 498, 579 Ghisellini G., Haardt F., Matt G., 1994, MNRAS 267, 743 Greenhill, L. J., Gwinn, C. R., Antonucci, R., Barvainis, R., 1996, ApJ 472, L21 Guainazzi M., Matt G., Antonelli L.A., et al., 1999, MNRAS in press (astro-ph/9905261) Guainazzi M., Antonelli, L.A., 1999, MNRAS 304, L15 Iwasawa K., Koyama K., Awaki H., et al., 1993, ApJ 409, 155 Iwasawa K., Fabian A.C., Matt G., 1997, MNRAS 289, 443 Iwasawa K., Comastri A., 1998, MNRAS 297, 1219 Kishimoto M., 1999, ApJ 518, 676 Komossa S., Schulz H., Greiner J., 1998, A&A 334, 110 Matt G., Brandt W.N,, Fabian A.C., 1996, MNRAS 280, 823 Matt G., Guainazzi M., Frontera F., et al., 1997, A&A 325, L13 Matt G., Pompilio F., La Franca F., 1999, New Astronomy 4/3, 191 Matt G., Guainazzi M., Maiolino R., et al., 1999, A&A 341, L39 Mulchaey J.S., Koratkar A., Ward M.J., 1994, ApJ 436, 586 309 Ogasaka Y., Inoue H., Brandt W.N., et al., 1997, PASJ 49, 179 Schulz H., Komossa S., Bergöfer T.W., Boer B., 1998, A&A 330, 823 Veilleux S., Kim D.-C., Sanders D.B., Mazzarella J.M., Soifer B.T., 1995, ApJSS 98, 171 [^1]: If $L_{\rm IR} \geq 10^{12} L_{\odot}$ is the limiting criterion for ULIRG, strictly speaking NGC 6240 belongs to this class only if $H_0 \leq$65 km s$^{-1}$ Mpc$^{-1}$; Genzel et al. (1998).
{ "pile_set_name": "ArXiv" }
--- abstract: 'A two-dimensional body, exhibiting a slight rotational movement, moves in a rarefied medium of particles which collide with it in a perfectly elastic way. In previously realized investigations by the first two authors, Plakhov & Gouveia (2007, Nonlinearity, 20), shapes of nonconvex bodies were sought which would maximize the braking force of the medium on their movement. Giving continuity to this study, new investigations have been undertaken which culminate in an outcome which represents a large qualitative advance relative to that which was achieved earlier. This result, now presented, consists of a two-dimensional shape which confers on the body a resistance which is very close to its theoretical supremum value. But its interest does not lie solely in the maximization of Newtonian resistance; on regarding its characteristics, other areas of application are seen to begin to appear which are thought to be capable of having great utility. The optimal shape which has been encountered resulted from numerical studies, thus it is the object of additional study of an analytical nature, where it proves some important properties which explain in great part its effectiveness.' author: - | Paulo D. F. Gouveia${\,}^a$\ `[email protected]` - | Alexander Plakhov${\,}^{b,c}$\ `[email protected]` - | Delfim F. M. Torres${\,}^{b}$\ `[email protected]` date: | ${}^a$Bragança Polytechnic Institute, 5301-854 Bragança, Portugal\ ${}^{b}$University of Aveiro, 3810-193 Aveiro, Portugal\ ${}^{c}$Aberystwyth University, SY23 3BZ Aberystwyth, UK title: 'Two-dimensional body of maximum mean resistance[^1]' --- **Keywords:** body of maximal resistance, billiards, Newton’s aerodynamic problem, retroreflector. **2000 MSC:** 74F10, 65D15, 70E15, 49K30, 49Q10. Introdution =========== One area of investigation in contemporary mathematics is concerned with the search for shapes of bodies, within predefined classes, which permit the minimization or maximization of the resistance to which they are subjected when they move in rarefied media. The first problem of this nature goes back to the decade of the 1680s, a time when Isaac Newton studied a problem of minimum resistance for a specific class of convex bodies, which moved in media of infinitesimal particles, rarefied to such a degree that it was possible to discount any interaction between the particles, and in which the interaction of these with the bodies could be described as perfectly elastic collisions [@newton1686]. More recently we have witnessed important developments in this area with the broadening of study to new classes of bodies and to media with characteristics which are less restrictive: problems of resistance in non-symmetrical bodies [@buttazzo95; @buttazzo97; @buttazzo93; @robert06; @robert01b], in nonconvex bodies of single collisions [@brock96; @buttazzo93; @comte01; @robert01] and multiple collisions [@Plakhov03b; @plakhov03; @Plakhov04], bodies of developable surfaces [@robert01], considering collisions with friction [@horstmann02] and in media with positive temperature [@Plakhov05]. However most studies which have been published have given special attention to classes of convex bodies. The convexity of a body is a sufficient condition for the resistance to be solely a function of singular collisions — all the particles collide at once with the body. This attribute allows us to considerably reduce the complexity of the problems dealt with. Even the various studies of classes of nonconvex bodies which have emerged, especially in the last decade, are based almost always on conditions that guarantee a single impact per particle — [@brock96; @buttazzo93; @comte01; @robert01]. Only very recently have there begun to emerge some studies supposing multiple reflections (see, e.g. [@Plakhov03b; @plakhov03; @Plakhov04]). In the class of convex bodies, the problem is normally reduced to the minimization of Newton’s functional — an analytical formula for the value of the resistance. But, in the context of nonconvex bodies, there is not any simple formula known for the calculation of the resistance. Even if it is extremely complex, in general, to deal analytically with problems of multiple collisions, for some specific problems of minimization the job has not been revealed to be particularly difficult, there even being some results already available [@Plakhov03b; @plakhov03]. If, on the other hand, we consider the problem of maximization, then in this case the solution becomes trivial — for any dimension, it is enough that the front part of the body is orthogonal to the direction of the movement. And what if the body exhibits, besides its translational movement, a slight rotational movement? When we think of this kind of problem, we have in mind, for example, artificial satellites, of relatively low orbits, which do not possess any control system which could stabilize their orientation, or other devices in similar conditions. In this situation we imagine that, over its path, the device rotates slowly around itself. The problem of resistance minimization for rotating nonconvex two-dimensional bodies has already been studied in [@Plakhov04; @ARMA]: it was shown that the maximal reduction of resistance, as compared with the convex case, is approximately $1.22\%$. In its turn, the problem of maximization of the average resistance of bodies in rotation is far from being trivial, in contrast with that which occurs when we deal with purely translational movement. This class of problems was, therefore, the object of study of the work carried out by the authors in [@Plakhov07:CM; @Plakhov07]: nonconvex shapes of bodies were investigated which would maximize the resistance that they would have to confront if they moved in rarefied media, and, simultaneously, exhibited a slight rotational movement. With the numerical study which was executed, various geometrical shapes were found which conferred on the bodies rather interesting values of resistance: but it was in later investigations, performed in the follow-up of this work, that the authors managed to arrive at the best of the results — a two-dimensional shape which confers on the body a resistance very near to its maximum theoretical limit. It is this latest result which now is presented here. The presentation of the work is organized in the following way. In section \[sec:defBidimens\], we begin by defining, for the two-dimensional case, the problem of maximization, which is the object of the present study. Then, in section \[sec:estNumer\], we describe the numerical study which was realized in the tracking of the body of maximum resistance and we present the main original result of this study: a two-dimensional shape which maximizes Newtonian resistance. The two-dimensional shape is then the object of study in section \[sec:caract\], where some important properties are shown which help to explain the value of resistance which it displays. In section \[sec:concl\], we present the main conclusions of our study and include some notes on possible working directions to undertake in the future. Finally, in appendices \[cha:condSuf3col\] and \[cha:min3col\], proofs of theorems \[teor:3ref\] and \[teor:min3ref\] are provided. Definition of the problem for the two-dimensional case {#sec:defBidimens} ====================================================== Consider a disc in slow and uniform rotation, moving in a direction parallel to its plane. We will designate the disc of radius $r$ by $C_r$ and its boundary by $\partial C_r$. We then remove small pieces of the disc along its perimeter, in an $\varepsilon$-neighborhood of $\partial C_r$, with $\varepsilon \in \mathbb{R}_+$ of value arbitrarily small when compared with the value of $r$. We are thus left with a new body $B$ defined by a subset of $C_r$ and characterized by a certain roughness along all its perimeter. The essential question which we put is the following: up to what point can the resistance of a body $B$ be augmented? More than getting to know the absolute value of this resistance, we are principally interested in learning what is the increase which can be obtained in relation to the smooth body (a perfectly circular contour, in this case), that is, learning the normalized value $$\label{eq:normalizacao} R(B)=\frac{\text{Resistance}(B)}{\text{Resistance}(C_r)} \text{.}$$ It is possible, from the beginning, to know some important reference values for the normalized resistance: $R(C_r)=1$ and the value of the resistance $R(B)$ will have to be found between $0.9878$ ([@Plakhov04; @ARMA]) and $1.5$. The value $1.5$ will be hypothetically achieved if all the particles are reflected by the body with the velocity $\mathbf{v}^+$ (velocity with which the particles separate definitively from the body) opposite to the velocity of incidence $\mathbf{v}$ (velocity with which the particles strike the body for the first time), $\mathbf{v}^+=-\mathbf{v}$, the situation in which the maximum momentum is transmitted to the body. It is also possible for us to know the resistance value of some elementary bodies of the type $B$. This is the case, for example, of discs with the contour entirely formed by rectangular indentations which are arbitrarily small or with the shape of rectangular isosceles triangles. As was proved in [@Plakhov07], these bodies are associated with resistances, respectively, of $R=1.25$ and $R=\sqrt{2}$. Apart from being defined in the disc $C_r$, it is assumed that the body to be maximized is a connected set $B\in\mathbb{R}^2$, with piecewise smooth boundary $\partial B$. Therefore, let us consider a billiard in $\mathbb{R}^2\setminus B$. An infinitesimal particle moves freely, until, upon colliding with the body $B$, it suffers various reflections (one at least) at regular points of its boundary $\partial B$, ending up by resuming free movement which separates it definitively from the body. Denote by $\text{conv}B$ the convex hull of $B$. The particle intercepts the $\partial(\text{conv}B)$ contour twice: when it enters into the set $\text{conv}B$ and in the moment that it leaves. $L=|\partial(\text{conv}B)|$ is considered the total length of the curve $\partial(\text{conv}B)$, and the velocity of the particle is in the first and second moments of interception represented by $\mathbf{v}$ and $\mathbf{v}^+$, and $x$ and $x^+$ the respective points where they occur. As well, the angles which the vectors $-\mathbf{v}$ and $\mathbf{v}^+$ make with the outer normal vector to the section of $\partial(\text{conv}B)$ between the points $x$ and $x^+$ are designated $\varphi$ and $\varphi^+$. They will be positive if they are defined in the anti-clockwise direction from the normal vector, and negative in the opposite case. With these directions, both $\varphi$ as well as $\varphi^+$ take values in the interval $[-\pi/2,\pi/2]$. Representing the cavities which characterize the contour of $B$ by subsets $\Omega_1,\Omega_2, \ldots$, which in their total make up the set $\text{conv}B \setminus B$, the normalized resistance of the body $B$ (equation ) takes the following form (cf. [@Plakhov07]): $$\label{eq:RB2} R(B) =\frac{|\partial(\text{conv}B)|}{|\partial C_r|} \left(\frac{L_0}{L}+\sum_{i \ne 0}{\frac{L_i}{L}R(\tilde{\Omega}_i)}\right) \text{,}$$ being $L_0=|\partial(\text{conv}B)\cap \partial B|$ the length of the convex part of the contour $\partial B$, $L_i=|\partial(\text{conv}B)\cap {\Omega}_i|$, with $i=1,2,\ldots$, the size of the opening of the cavity ${\Omega}_i$, and $R(\tilde{\Omega}_i)$ the resistance of the normalized cavity $\tilde{\Omega}_i$, in relation to a smooth segment of unitary size, with $$\label{eq:R} R(\tilde{\Omega}_i)=\frac{3}{8} \int_{-1/2}^{1/2}\int_{-\pi/2}^{\pi/2} \left( 1+\cos\left( \varphi^+(x,\varphi) -\varphi \right) \right) \cos \varphi\, \mathrm{d}\varphi\,\mathrm{d} x \text{.}$$ The function $\varphi^+$ should be seen as the angle of departure of a particle which interacts with a cavity $\tilde{\Omega}_i$ that has opening of unit size and is similar to $\Omega_i$, with the similarity factor $1/L_i$ — see illustration of figure \[fig:cavidadeOmega\]. ![Example of trajectory of a particle which interacts with a cavity $\tilde{\Omega}_i$.[]{data-label="fig:cavidadeOmega"}](fig1){width="0.35\columnwidth"} From equation , we understand that the resistance of $B$ can be seen as a weighted mean ($\sum_i L_i/L=1$) of the resistances of the individual cavities which characterize all its boundary (including resistance of the convex part of the boundary), multiplied by a factor which relates the perimeters of the bodies $\text{conv}B$ and $C_r$. Thus, maximizing the resistance of the $B$ body amounts to maximizing the perimeter of $\text{conv}B$ ($|\partial(\text{conv}B)|\le |\partial C_r|$) and the individual resistances of the cavities $\Omega_i$. Having found the optimal shape $\Omega^*$, which maximizes the functional , the body of maximum resistance $B$ will be that whose boundary is formed only by the concatenation of small cavities with this shape. We can therefore restrict our problem to the sub-class of bodies $B$ which have their boundary integrally covered with equal cavities, and in doing so admit, without any loss of generality, that each cavity $\Omega_i$ occupies the place of a circle arc of size $\varepsilon\ll r$. As with $L_i=2r\sin(\varepsilon/2r)$, the ratio between the perimeters takes the value $$\label{eq:RxPerimetros} \frac{|\partial(\text{conv}B)|}{|\partial C_r|} =\frac{ \sin(\varepsilon/2r)}{\varepsilon/2r} \approx 1- \frac{(\varepsilon/r)^2}{24}\text{,}$$ or that is, given a body $B$ of a boundary formed by cavities similar to $\Omega$, from  and , we conclude that the total resistance of the body will be equal to the resistance of the individual cavity $\Omega$, less a small fraction of this value, which can be neglected when $\varepsilon\ll r$, $$\label{eq:Raprox} R(B)\approx R(\Omega)-\frac{(\varepsilon/r)^2}{24}R(\Omega)\text{.}$$ Thus, our research has as its objective the finding of cavity shapes $\Omega$ which maximize the value of the functional , whose limit we know to be found in the interval $$\label{eq:RminMax} 1\le \text{sup}_\Omega R(\Omega) \le 1.5 \text{,}$$ as is easily proven using : if $\Omega$ is a smooth segment, $\varphi^+(x,\varphi)=-\varphi$ and $R(\Omega)=\frac{3}{8} \int_{-1/2}^{1/2}\int_{-\pi/2}^{\pi/2} \left( 1+\cos\left( 2\varphi \right) \right) \cos \varphi\, \mathrm{d}\varphi\,\mathrm{d} x=1 $; in the conditions of maximum resistance $\varphi^+(x,\varphi)$ $=\varphi$, thus $ R(\Omega)\le\frac{3}{8} \int_{-1/2}^{1/2}\int_{-\pi/2}^{\pi/2} 2 \cos \varphi\, \mathrm{d}\varphi\,\mathrm{d} x =1.5$. Numerical study of the problem {#sec:estNumer} ============================== In the class of problems which we are studying, only for some shapes of $\Omega$ which are very elementary is it possible to derive an analytical formula of their resistance , as we saw in the rectangular and triangular shapes previously referred to. For somewhat more elaborate shapes, the analytical calculation becomes rapidly too complex, if not impossible, given the great difficulty in knowing the function $\varphi^+: [-1/2,1/2]\times[-\pi/2,\pi/2]\rightarrow[-\pi/2,\pi/2]$, which as we know, is intimately related to the format of the cavity $\Omega$. Therefore, recourse to numerical computation emerges as the natural and inevitable approach in order to be able to investigate this class of problems. There have been developed various computational models which simulate the dynamics of billiard in the cavity. The algorithms of construction of these models, as well as the those responsible for the numerical calculation of the associated resistance, were implemented using the programming language C, given the computational effort involved (language C was created in 1972 by Dennis Ritchie; for its study we suggest, among the extensive documentation available, that which is the reference book of its language, written by Brian Kernighan and Dennis Ritchie himself, [@LangC]). The efficiency of the object code, generated by the compilers of C, allowed the numerical approximation of  to be made with a sufficiently elevated number of subdivisions of the intervals of integration — between some hundreds and various thousands (up to $5000$). The results were, because of this, obtained with a precision which reached in some cases $10^{-6}$. This precision was controlled by observation of the difference between successive approximations of the resistance $R$ which were obtained with the augmentation of the number of subdivisions. For the maximization for the resistance of the idealized models, there were used the global algorithms of optimization of the *toolbox* “*Genetic Algorithm and Direct Search*” (version 2.0.1 (R2006a), documented in [@toolbox]), a collection of functions which extends the optimization capacities of the MATLAB numerical computation system. The option for Genetic and Direct search methods is essentially owed to the fact that these do not require any information about the gradient of the objective function nor about derivatives of a higher order — as the analytical form of the resistance function is in general unknown (given that it depends on $\varphi^+(x,\varphi)$), this type of information, if it were necessary, would have to be obtained by numerical approximation, something which would greatly impede the optimization process. The MATLAB computation system (version 7.2 (R2006a)) was also chosen because it had functionalities which allowed it to be used for the objective function the subroutine compiled in C of resistance calculation, as well as the $\varphi^+(x,\varphi)$ function invoked in itself. “Double Parabola”: a two-dimensional shape which maximizes resistance {#sec:DuplaParab} --------------------------------------------------------------------- In the numerical study which the authors carried out in [@Plakhov07:CM; @Plakhov07], shapes of $\Omega_f$ defined by continuous and piecewise differentiable $f:[-1/2,1/2]\rightarrow \mathbb{R}^+$ functions were sought for: $$\label{eq:Omegaf} \Omega_f=\left\{(x,y):\,-1/2\le x \le 1/2,\; 0\le y \le f(x)\right\}\text{,}$$ with the interval $[-1/2, 1/2] \times \{ 0 \}$ being the opening. The search for the maximum resistance was begun in the class of continuous functions $f$ with derivative $f'$ piecewise constant, broadening later to the study of classes of functions with the second derivative $f''$ piecewise constant. In the first of the cases the contour of $\Omega_f$ is a polygonal line, and in the second, a curve composed of parabolic arcs. Not having been able with these shapes to exceed the value of resistance $R=1.44772$, we decided, in this new study, to extend the search to shapes different from those considered in . We studied shapes $\Omega^g$ defined by functions $x$ of $y$ of the following form: $$\label{eq:Omegag} \Omega^g=\left\{(x,y):\,0\le y \le h,\; -g(y)\le x \le g(y)\right\}\text{,}$$ where $h>0$ and $g:[0,h] \rightarrow \mathbb{R}_0^+$ is a continuous function with $g(0)=1/2$ and $g(h)=0$. The new problem of maximum resistance studied by us can therefore be formulated in the following way: > [*To find $\sup_g R(\Omega^g)$ in the continuous and piecewise differentiable functions $g:[0,h] \rightarrow \mathbb{R}_0^+$, such as $g(0)=1/2$ and $g(h)=0$, with $h>0$.* ]{} Similarly to the study which was carried out for the sets $\Omega_f$, in the search for shapes $\Omega^g$, the functions $g$ were considered piecewise linear and piecewise quadratic. If in the classes of linear functions it was not possible to achieve a gain in resistance relative to the results obtained for the sets $\Omega_f$, in the quadratic functions the results exceeded the highest expectations: there was found a shape of cavity $\Omega^g$ which presented the resistance $R=1.4965$, a value very close to its theoretical limit of $1.5$. There were also carried out some tests with polynomial functions of higher order or described by specific conical sections, but, not having verified any additional gain in the maximization of resistance, it was decided not to report the respective results. There therefore follows the description of the best result which was obtained, encountered in the class of quadratic functions $x=\pm g(y)$. The value of resistance of the sets $\Omega^g$ were studied, just as defined in , in the class of quadratic functions $$g_{h,\beta}(y) = \alpha y^2 + \beta y +1/2, \text{ for } 0 \le y \le h\,,$$ where $h>0$ and $\alpha= \frac{-\beta h -1/2}{h^2}$ (given that $g_{h,\beta}(h) =0$). In the optimization of the curve, the two parameters of the configuration were made to vary: $h$, the height of the $\partial\Omega^g$ curve, and $\beta$, in its slope at the origin ($g'(0)$). In this class of functions the algorithms of optimization converge rapidly towards a very interesting result: the maximum resistance was reached with $h=1.4142$ and $\beta=0.0000$, and assumed the value $R=1.4965$, that is, a value $49.65\%$ above the resistance of the rectilinear segment. This result seems to us really interesting: 1. it represents a considerable gain in the value of the resistance, relative to the best result obtained earlier (in [@Plakhov07:CM; @Plakhov07]), which was situated $44.77\%$ above the reference value; 2. The corresponding set $\Omega^g$ has a much more simple shape than that of set $\Omega_f$ associated with the best earlier result, since it is formed by two arcs of symmetrical parabolas, while the earlier one was made up of fourteen of these arcs; 3. this new resistance value is very near to its maximal theoretical limit, which, as is known, is found $50\%$ above the value of reference; 4. The optimal parameters appear to assume value which give to the set $\Omega^g$ a configuration with very special characteristics, as in what follows will be understood. Note that the optimal parameters appear to approximate the values $h=\sqrt{2}=1.41421\ldots$ and $\beta=0$. The following question can therefore be put: > [*Are these not the exact values of the optimal parameters?* ]{} The graphical representation of the function $R(h,\beta)$ through the level curves, figure \[fig:curvNivel\], are effectively in concordance with this possibility — note that the level curves appear perfectly centered on the $(\sqrt{2},0)$ coordinates; marked on the figure by “$+$” . --------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------- ![Level curves of the $R(h,\beta)$ function.[]{data-label="fig:curvNivel"}](fig2a "fig:"){height="0.3\columnwidth"} ![Level curves of the $R(h,\beta)$ function.[]{data-label="fig:curvNivel"}](fig2b "fig:"){height="0.3\columnwidth"} (a) (b) --------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------- Note also, in figure \[fig:Res\_h\], the resistance graph $R(h)$ for $\beta=0$, where it can equally be perceived that there is a surprising elevation of resistance when $h\rightarrow\sqrt{2}$. Thus the resistance of the $\Omega^{g_{h\beta}}$ cavity was numerically calculated with the exact values $h=\sqrt{2}$ and $\beta=0$, the result having confirmed the value $1.49650$. ![Resistance graphic $R(h)$ for $\beta=0$.[]{data-label="fig:Res_h"}](fig3){width="0.4\columnwidth"} There is yet one more reason which suggests also an affirmative response to the formulated question. The shape of the set $\Omega^{g_{h,\beta}}$ with $h=\sqrt{2}$ and $\beta=0$ is a particular case with which is associated special characteristics which could justify the elevated value of resistance presented. The two sections of the shape are similar arcs of two parabolas with the common horizontal axis and concavities turned one towards the other — see figure \[fig:parabOpt\]. But the particularity of the configuration resides in the fact that the axis of the parabolas coincides with the line of entry of the cavity (axis of $x$), and that the focus of each one coincides with the vertex of the other. ----------------------------------------------------------------------------------------------------------------------------- -- ----------------------------------------------------------------------------------------------------------------------------- ![(almost) Optimal 2D shape — the *Double Parabola*.[]{data-label="fig:parabOpt"}](fig4a "fig:"){height="0.25\columnwidth"} ![(almost) Optimal 2D shape — the *Double Parabola*.[]{data-label="fig:parabOpt"}](fig4b "fig:"){height="0.25\columnwidth"} (a) (b) ----------------------------------------------------------------------------------------------------------------------------- -- ----------------------------------------------------------------------------------------------------------------------------- This shape of cavity appears to effectively deal with a very particular case. In contrast with what happened with all the other shapes which were studied, the integrand function of functional  displays a rather smooth shape, presenting only a few small irregularities for $\varphi$ angles of little amplitude. Noting this characteristic, and taking into account that the integrand function almost does not depend on $x$, the resistance was calculated, for this shape in particular, using the rule of Simpson $1/3$ in the integration in order to $\varphi$. The double integration in the equation  was thus numerically approximated by the following expression: $$\label{eq:RnsSimpson} R=\frac{1}{2} \Delta x \Delta \varphi \sum_{i=N_x/2+1}^{N_x}\sum_{k=1}^{N_\varphi-1} w_k \left( 1+\cos\left( \varphi^+(x_i,\varphi_k) -\varphi_k \right) \right) \cos \varphi_k\text{,}$$ with $w_k=2$ for $k$ odd and $w_k=1$ for $k$ even, $x_i=- 1/2+(i-1/2) \Delta x$, $\Delta x=1/N_x$, $\varphi_k=-\pi/2+ k \Delta \varphi$ and $\Delta \varphi=\pi/N_\varphi$. $N_x$ and $N_\varphi$ are the number of sub-intervals to consider in the integration of the variables $x$ and $\varphi$ (both even numbers), respectively, and $\Delta x$ and $\Delta \varphi$ the increments for the correspondent discreet variables. Given that the shape $\Omega^{g_{\sqrt{2},0}}$ presents horizontal symmetry, the first summation of the expression considers only the second half of the interval of integration of the variable $x$. In order to be easily referred to, this shape of cavity (figure \[fig:parabOpt\]a) will be, from here on, named simply “*Double Parabola*”. Thus, in the context of this paper, the term “Double Parabola” should be always understood as the name of the cavity whose shape is described by two parabolas which, apart from being geometrically equal, find themselves “nested” in the particular position to which we have referred. Since the resistance of the Double Parabola assumes a value which is very close to its theoretical limit, in a final attempt to achieve this limit, it was resolved to extend the study even further to other classes of functions $g(y)$ which admit the Double Parabola as a particular case or which allow proximate configurations of this nearly optimal shape. In all these cases the best results were invariably obtained when the shape of the curves approximated the shape of the Double Parabola, without ever having overtaken the value $R=1.4965$. It was begun by considering functions $g(y)$ piecewise quadratic, including curves *splines*, without achieving interesting results; only for functions $g(y)$ of $2$ or $3$ segments was it possible to approximate the resistance and the shape of the Double Parabola. Cubic and bi-quadratic functions $g(y)$ were also considered[^2], but in both cases the process of optimization brought them proximate to the curves of quadratic order, with the coefficients of greater order taking values which were almost zero. The problem was studied in the class of conical sections, considering, for lateral facets of the cavity, two symmetrical arcs either of an ellipse or of a hyperbole. Also in these cases the arcs assumed a shape very close to the arcs of the parabolas. The Double Parabola being the best shape encountered, and dealing with a nearly optimal shape, in the section which follows it is the object of deeper study, of an essentially analytical nature, where the reasons for its good performance are sought. Characterization of the reflections in the shape “Double Parabola” {#sec:caract} ================================================================== Each one of the illustrations of figure \[fig:trajectorias\] shows, for the “Double Parabola”, a concrete trajectory, obtained with our computational model. -------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------- ![Example of trajectories obtained with the computational model.[]{data-label="fig:trajectorias"}](fig6a "fig:"){width="0.15\columnwidth"} ![Example of trajectories obtained with the computational model.[]{data-label="fig:trajectorias"}](fig6b "fig:"){width="0.15\columnwidth"} ![Example of trajectories obtained with the computational model.[]{data-label="fig:trajectorias"}](fig6c "fig:"){width="0.15\columnwidth"} \(a) $x=0.45$, $\varphi=75^\circ$. \(b) $x=0.45$, $\varphi=55^\circ$. \(c) $x=0.45$, $\varphi=35^\circ$. ![Example of trajectories obtained with the computational model.[]{data-label="fig:trajectorias"}](fig6d "fig:"){width="0.15\columnwidth"} ![Example of trajectories obtained with the computational model.[]{data-label="fig:trajectorias"}](fig6e "fig:"){width="0.15\columnwidth"} ![Example of trajectories obtained with the computational model.[]{data-label="fig:trajectorias"}](fig6f "fig:"){width="0.15\columnwidth"} \(d) $x=0.3$, $\varphi=75^\circ$. \(e) $x=0.0$, $\varphi=35^\circ$. \(f) $x=0.48$, $\varphi=5^\circ$. -------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------- It is comforting to verify that, with the exception of one trajectory, in all the others the particle emerges from the cavity with a velocity which is nearly opposite to that which was its entry velocity. This is the “symptom” which unequivocally characterizes a cavity of optimal performance. Even in the case of the trajectory of the illustration (f), the direction of the exit velocity appears not to vary greatly from that of entry. If we analyze the five first illustrations, we may verify that there exists something in common in the behavior of the particle: in describing the trajectory, the particle is always subject to three reflections. This appears to be a determinant characteristic for the approximation of the two angles of entry and exit. If, for example, we imagine three trajectories with proximate configurations, respectively, the trajectories (a), (b) and (c), but with the difference of not possessing the third reflection, the result would be completely different, as easily can be seen in the illustrations. Although this conviction is by nature essentially empirical, the results of the study which follow are heading in the direction of confirming that one very significant part of the “benign” trajectories — those in which the vectors velocity of entry and of exit are nearly parallel; we call them so because they represent positive contributions to the maximization of resistance — suppose exactly three reflections. We now will try to interpret another type of results obtained with our computational model, commencing with the graphical representation of the distribution of the pairs $(\varphi,\varphi^+)$ on the Cartesian plane — see figure \[fig:Dist\_Phi\_PhiPlus\]. This graph was produced with $10.000$ pairs of values $(x,\varphi)$, generated by a random process of uniform distribution. ![Distribution of the $(\varphi,\varphi^+)$ pairs on the Cartesian plane.[]{data-label="fig:Dist_Phi_PhiPlus"}](fig8){width="0.4\columnwidth"} The points concentrate themselves on the proximities of the diagonal $\varphi=\varphi^+$, which revealing of good behavior on the part of the cavity. In addition, with these results it is shown that the response of the cavity deteriorates as $\varphi$ approaches zero. Therefore, it begins to be understood that the “benign” trajectories have their origin essentially in entry angles of elevated amplitude. If we consider figure \[fig:Dist\_Phi\_PhiPlus\], there appears to exist an additional perturbation in the behavior of the cavity when the amplitude of the entry angle is inferior to about $20^\circ$, which means that some $(\varphi,\varphi^+)$ pairs become, in relation to the others, more dispersed and more distant from the diagonal $\varphi^+=\varphi$. We have already called attention to the possible importance of the three reflections in the degree of approximation verified in the angles $\varphi$ and $\varphi^+$. It occurs to us, therefore, to put the following question: is it not precisely the number of reflections that, on differentiating themselves from the $3$ occurrences, interfere so negatively with the behavior of the cavity? The investigations that follow will demonstrate, among other things, that our suspicion on this point has a basis. The following theorem says that for $\varphi$ outside some interval $(-\varphi_0,\varphi_0)$, the number of reflections is always three. The proof is presented in appendix \[cha:condSuf3col\]. \[teor:3ref\] For $\varphi$ entry angles superior (in absolute value) to $\varphi_0=\arctan\left(\frac{\sqrt{2}}{4}\right)\simeq 19.47^\circ$, the number of reflections to which the particle is subjected in the interior of the Double Parabola cavity is always equal to three, and they occur alternately on the left and right faces of the cavity, no matter what the entry position may be. As a way to verify that the deductions which we have made are effectively in concordance with the numerical results of the computational model which was developed, we present one more graph, figure \[fig:ncolisoes4\], produced with $10.000$ pairs of $(x,\varphi)$ values, generated randomly with uniform distribution. ![Distribution of the $(\varphi,nr)$ pairs on the Cartesian plane, being $nr$ the nº of reflections.[]{data-label="fig:ncolisoes4"}](fig9){width="0.4\columnwidth"} As can be observed in figure \[fig:ncolisoes4\], all the trajectories with $4$ or more reflections, among the $10.000$ considered, happened within the interval $(-\varphi_0,\varphi_0)$. Outside this interval (for $\left|\varphi\right|>\varphi_0$) the trajectories are always of three reflections. Additionally, we can verify that there isn’t any trajectory with less than three reflections. This numerical evidence is confirmed by the following theorem: \[teor:min3ref\] Any particle which enters in the cavity Double Parabola describes a trajectory with a minimum of $3$ reflections. The proof of theorem 2 is presented in appendix \[cha:min3col\]. Of the conclusions which we arrived at we can immediately come to the following corollary: in trajectories with $4$ or more reflections the angular difference $\left|\varphi-\varphi^+\right|$, no matter how much bigger it may be, will never be superior to $2\varphi_0\simeq 38.94^\circ$, a value which is much more inferior to the greatest angle which it is possible to form between two vectors ($180^\circ$). The proof of this corollary is simple: as a trajectory of $4$ or more reflections is always associated with a entry angle $-\varphi_0<\varphi<\varphi_0$, the exit angle will be situated necessarily in the same interval; taking into account the property of reversibility associated with the law of reflection which governs reflections, if just to be absurd we were to admit $\left|\varphi^+\right|>\varphi_0$, on inverting the direction of the particle, we would be in the position of having a trajectory of more than $3$ reflections with a $\varphi^+$ entry angle situated outside the interval $(-\varphi_0,\varphi_0)$, which would enter into contradiction with the initial postulate. Summarizing: - There is verified a great predominance of trajectories with $3$ reflections; - There are no trajectories of fewer than $3$ reflections; - The critical angle $\varphi_0$ has the value $\varphi_0=\arctan\left(\frac{\sqrt{2}}{4}\right)\simeq 19.47^\circ$; - Outside the interval $(-\varphi_0,\varphi_0)$, all the trajectories are of $3$ reflections; - In trajectories with $4$ or more reflections, the angular difference is delimited by $2\varphi_0$: $\left|\varphi-\varphi^+\right|<2\varphi_0$. Conclusion and future perspectives {#sec:concl} ================================== In the continuation of the study carried out previously by the authors in [@Plakhov07:CM; @Plakhov07], with the work now presented it has been possible to obtain an original result which appears to us to have great scope: the algorithms of optimization converged for a geometrical shape very close to the ideal shape — the *Double Parabola*. This concerns a form of roughness which confers a nearly maximal resistance (very close to the theoretical upper bound) to a disc which, not only travels in a translational movement but also rotates slowly around itself. In figure \[fig:discoOptimo\] one of these bodies is shown. ![(almost) Optimal 2D body.[]{data-label="fig:discoOptimo"}](fig16){width="0.25\columnwidth"} Noting that the contour of the presented body is integrally formed by $42$ cavities $\Omega$ with the shape of a Double Parabola, each one of which with a relative resistance of $1.49650$, from and we conclude that $R(B)=\frac{\sin(\pi/42)}{\pi/42}R(\Omega)\approx 1.4951$ is the total resistance of the body, a value $49.51\%$ above the value of resistance of the corresponding disc of smooth contour (the smallest disc which includes the body). We know that if the body were formed by a sufficiently elevated number of these cavities, its resistance would even reach the value $1.4965$, but the example presented is sufficient in order for us to understand how close we are to the known theoretical upper bound ($50\%$). Although the value of resistance of the Double Parabola had been determined numerically, an analytical study was done in section \[sec:caract\], with the objective of consolidating the presented results. We have managed to prove some important properties which help in the understanding of the elevated value of resistance which was obtained. We will try in the future to develop other theoretical studies which will allow us to consolidate this result even further. For example, an interesting open problem lies in delimiting the lag between the angles of entrance and exit for the trajectories of $3$ reflections — for the others (trajectories with $4$ or more reflections) we already know that $\left|\varphi-\varphi^+\right|<2\varphi_0 \simeq 2 \times 19.47^\circ$. The Double Parabola is effectively a result of great practical scope. Besides maximizing Newtonian resistance, it is exciting to verify that the potentialities of the Double Parabola shape found by us could also reveal themselves to be very interesting in other areas of practical interest. If we coat the interior part of the Double Parabola cavity with a polished “surface”, the trajectory of the light in its interior will be described by the principles of geometrical optics, in particular rectilinear propagation of light, laws of reflection and reversibility of light. Thus, as the computational models which were developed by us to simulate the dynamic of billiards in the interior of each one of the shapes studied (where collisions of particles are considered perfectly elastic) are equally valid when the problem becomes of an optical nature, we can also look at $2$D shape found by us in this new perspective. Given the characteristics of reflection which the Double Parabola shape presents we can rapidly conceive for it a natural propensity for being able to be used with success in the design of retroreflectors — see in [@gouvPhD] the exploratory study of its possible utilization in roadway signalization and the automobile industry. An incursion into the three-dimensional case, carried out in [@gouvPhD], also showed that the Double Parabola is a shape of cavity which is very special. Our conviction of its effectiveness was strongly reinforced when we obtained the best result for the 3D case. This result was achieved with a cavity whose surface is the area swept by the movement of the Double Parabola curve in the direction perpendicular to its plane. The value of its resistance ($R=1.80$) having been a little below the theoretical upper bound for the 3D case ($R=2$), to go beyond this value will be also an interesting challenge to consider in the future. For the 2D case we envision greater difficulty in going beyond the result which has already been reached — whether for the proximity which it has to the theoretical upper bound, or for the fact that we have already carried out, without success, a series of investigation with just this objective. Proof of theorem \[teor:3ref\] {#cha:condSuf3col} ============================== Consider a particle which enters into the cavity in $(x,0)$, with the vector velocity forming an angle $\varphi$ with the vertical axis, just as is found represented in the illustrations of figure \[fig:parabOpt3col\], where we assume that the axis of symmetry of the cavity is the axis of the $y$ and that its base $\overline{A_0A_1}$ is placed on the axis $x$. In this way, the position of the particle at entry of the cavity assumes only values in the interval $\left(-\frac{1}{2},\,\frac{1}{2}\right)\times\{0\}$. ![Set of illustrations to the study of the trajectory of particles with entry angle $\varphi>\varphi_0 \simeq 19.47^\circ$, in the cavity “Double Parabola”.[]{data-label="fig:parabOpt3col"}](fig17){width="0.6\columnwidth"} Given the symmetry of the cavity in relation to its vertical axis, it will be enough to analyze its behavior for $\varphi_0 <\varphi< 90^\circ$. The conclusions at which we arrive will be in this way equally valid for $-90^\circ<\varphi< -\varphi_0$. We will analyze therefore in detail and separately each one of the sub-trajectories which compose all the trajectory described by the movement of the particle in the interior of the cavity. *Sub-trajectory $\overrightarrow{B_0B_1}$* For $\varphi>\varphi_0$, we have the guarantee that the first reflection occurs in the parabolic curve of the left side of the cavity, just as can be easily deduced from the illustration (a). So that the particle collides with the left curve it is enough that the $\varphi$ angle is superior to $\arctan(x/\sqrt{2})$, a magnitude which has as upper bound $\varphi_0=\arctan(\sqrt{2}/4)$. We thus have the initial trajectory of the cavity represented in illustration (a) by vector $\overrightarrow{B_0B_1}$. *Sub-trajectory $\overrightarrow{B_1B_2}$* \[pg:subtrajB1B2\] After colliding in $B_1$, in agreement with the law of reflection, the particle follows trajectory $\overrightarrow{B_1B_2}$. We prove that $\overrightarrow{B_1B_2}$ has an ascendant path — illustration (a). We trace the straight line $\overline{A_1A_2}$, segment, parallel to the initial trajectory of the particle $\overline{B_0B_1}$, which passes through the focus of the left parabola ($A_1$). Because of the focal property of this parabola, a particle which takes the sub-trajectory $\overrightarrow{A_1A_2}$, after reflection at $A_2$, will follow a horizontal direction $\overline{A_2A_3}$ (proceeding after its trajectory, after a new reflection, in the direction of the focus $A_0$ of the second parabola). Upon the occurrence of the first reflection of the particle at $B_1$, a point of the curve necessarily positioned below $A_2$, the trajectory $\overrightarrow{B_1B_2}$, which it will follow straight away, will be on an ascendant path, since the derivative $\frac{\mathrm{d} y}{\mathrm{d} x}$ of the curve at this point ($B_1$) is superior to the derivative in $A_2$, where the trajectory followed was horizontal. Although we now know that $\overrightarrow{B_1B_2}$ takes an ascendant path, nothing yet guarantees to us that the second reflection happens necessarily in the parabola of the right side. If we are able to verify that for $\varphi=\varphi_0$ the second reflection is always on the right side, no what the entry position $x$ is, therefore, logically, the same will happen for any value $\varphi>\varphi_0$. This premise can be easily accepted with the help of illustration (a) of figure \[fig:parabOpt3colA\]: for any value of $\varphi>\varphi_0$, with the first reflection at a given point $B_1$, it is always possible to trace a trajectory for $\varphi=\varphi_0$ which presents the first reflection at the same point $B_1$; the second reflection at the curve of the right side being for the case $\varphi=\varphi_0$, necessarily the same will happen for the trajectory with $\varphi>\varphi_0$, since the angle of reflection will be less in this second case, just as is illustrated in the figure. Consequently, it will be enough for us to prove for $\varphi=\varphi_0$, that the second reflection always occurs in the parabola of the right side, so that the same is proven for any which is the $\varphi>\varphi_0$. ![Illustrations to the study of the second reflection.[]{data-label="fig:parabOpt3colA"}](fig18){width="0.57\columnwidth"} In illustration (b) of figure \[fig:parabOpt3colA\] the trajectory until the second reflection of a particle with the angle of entry $\varphi_0$ ($\overrightarrow{B_0B_1}$ and $\overrightarrow{B_1B_2}$) is shown. As one can conclude from the illustration, the $B_2$ reflection only will happen on the curve of the left side if the $\alpha$ angle is less than $\alpha'$. We have determined the value of the two angles. Being $(x_1,y_1)$ the coordinates of the point $B_1$, we will have $\tan(\alpha')=-x_1/(\sqrt{2}-y_1)$, thus $$\label{eq:alphalinha} \alpha' =\arctan\left(\frac{-(y_1^2/4-1/2)}{\sqrt{2}-y_1}\right) =\arctan\left(\frac{(2-y_1^2)/4}{\sqrt{2}-y_1}\right) =\arctan\left(\frac{\sqrt{2}+y_1}{4}\right)\text{.}$$ In order to arrive at the value of $\alpha$ we resolve the system of three equations, of unknown $\alpha$, $\theta$ and $\beta$, which are taken directly from the geometry of the actual figure $$\left\{ \begin{array}{l} \alpha + \beta + \theta = \pi\\ \beta = \varphi_0+\theta\\ \arctan \left(\frac{1}{2}y_1\right)+\varphi_0+\theta = \frac{\pi}{2} \end{array} \right.$$ The tangent line to the curve in $B_1$ makes with the vertical an angle whose tangent has as its value the derivative $\frac{\mathrm{d} x}{\mathrm{d} y}$ of the curve at that point (in $y=y_1$), where $\frac{\mathrm{d} x}{\mathrm{d} y} =\frac{\mathrm{d}}{\mathrm{d} y}(\frac{1}{4}y^2-\frac{1}{2})=\frac{1}{2}y$. Because of this, that angle emerges represented in the third of the equations by the magnitude $\arctan(y_1/2)$. By resolving the system, the following result is obtained for $\alpha$ $$\label{eq:alpha} \alpha = \varphi_0+2\arctan({y_1}/{2})=\arctan({\sqrt{2}}/{4})+2\arctan({y_1}/{2}) \text{.}$$ Finally we prove that $\alpha>\alpha'$, no matter what $y_1\in \, (0,\sqrt{2})$ is. Of the equations  and , it will be equivalent to proving $$\arctan(\sqrt{2}/4)+2\arctan(y_1/2)>\arctan((\sqrt{2}+y_1)/4)\text{.}$$ Given that $0<y_1<\sqrt{2}$, both of the members of the inequality represent angles situated in the first quadrant of the trigonometrical circle. Because of this we can maintain the inequality for the tangent of the respective angles. Applying the tangent to both of the members, after effecting some trigonometrical simplifications, we arrive at the following relation $$\frac{1}{4}\,\frac{4\sqrt{2}-\sqrt{2}y_1^2+16 y_1}{4-y_1^2-\sqrt{2} y_1}> \frac{\sqrt{2}+y_1}{4}$$ Which, with additional algebraic simplifications, takes the form $${y_1\left(\sqrt{2}y_1+14+y_1^2\right)}/({4-y_1^2-\sqrt{2} y_1})>0 \text{.}$$ As $0<y_1<\sqrt{2}$, it can easily be seen that both the numerator and the denominator of the fraction present in this latest inequality are positive magnitudes. Thus $\alpha>\alpha'$, which contradicts the condition which was necessary so that the reflection $B_2$ could occur on the curve of the left side, it therefore being proven, as we intended, that in no situation does the reflection $B_2$ of illustration (b) of figure \[fig:parabOpt3colA\] happen on the curve of the left side. Logically, we can therefore conclude that the same happens for whatever is $\varphi>\varphi_0$: the second reflection of the particle occurs always in the parabola of the right side. *Sub-trajectory $\overrightarrow{B_2B_3}$* \[pg:subtrajB2B3\] We prove that the sub-trajectory $\overrightarrow{B_2B_3}$ has a descendant path — illustration (b) of figure \[fig:parabOpt3col\]. Imagine, for this purpose, a sub-trajectory $\overrightarrow{A_0A_2}$, parallel to $\overline{B_1B_2}$ and which passes through the focus $A_0$. The sub-trajectory $\overrightarrow{A_2A_3}$ which will follow the reflection in $A_2$ — a point of the parabola on the right side situated below $B_2$ — will be horizontal. The derivative of the curve in $A_2$ being superior to the derivative value in $B_2$, the sub-trajectory $\overline{B_2B_3}$ will necessarily be of a descending nature. Even if we already know that the sub-trajectory is descendant, we have not yet shown that sub-trajectory in no situation conducts the particle directly to the exit of the cavity. Therefore follows the proof that the reflection $B_3$ always occurs in a position superior to $A_0$ — illustration (c) of figure \[fig:parabOpt3col\]. We trace $\overline{A_2B_2}$, a segment of the horizontal straight line which passes through the point of reflection $B_2$. If the particle followed this trajectory, it would collide at the same point $B_2$, but heads itself to $A_0$. Therefore, by the law of reflection, $B_3$ will have to be above $A_0$, since $\overline{B_1B_2}$ makes an angle with the normal vector at the curve in $B_2$ less than that formed by segment $\overline{A_2B_2}$. *Sub-trajectory $\overrightarrow{B_3B_4}$* We will now show that the sub-trajectory which follows the reflection at $B_3$ crosses the segment $\overline{A_0A_1}$, that is, directs itself to the outside of the cavity — illustration (d) of figure \[fig:parabOpt3col\]. We trace, therefore, $\overline{A_2B_3}$, a segment of horizontal straight line which passes through the point of reflection $B_3$. If the particle followed this trajectory, it would collide at $B_3$ and would head itself towards $A_1$. Therefore, by the law of reflection, the straight line where the sub-trajectory $\overrightarrow{B_3B_4}$ is placed will have necessarily to pass below $A_1$, since $\overline{B_2B_3}$ makes an angle with the normal vector at the curve in $B_3$ bigger than that formed by the segment $\overline{A_2B_3}$. We have shown that the sub-trajectory crosses the axis of the $x$ at a point situated to the left of $A_1$, but we have not yet shown that it occurs to the right of $A_0$. For that, we will have to prove that the third is the last of the reflections, that is, that in no situation does there occur a fourth reflection in the parabola of the left side. There follows this proof, of them all the most complex one. In order to prove that following the third reflection there occurs no other collision in the left parabola, we will show that a fourth collision — represented by $B_4$ in the illustration (a) of figure \[fig:parabOpt3colB\] — has its origin always in an entry angle $\varphi$ inferior to $\varphi_0$. ![Illustrations to the study of a hypothetical fourth reflection.[]{data-label="fig:parabOpt3colB"}](fig19){width="0.57\columnwidth"} We will thus study the trajectory of the particle in the inverse order of its progression: we commence by admitting the existence of the sub-trajectory $\overrightarrow{B_3B_4}$ of illustration (a) and we will analyze its implications in all the preceding trajectory. In illustration (a) of figure \[fig:parabOpt3colB\] are to be found represented the sub-trajectories $\overrightarrow{B_2B_3}$ and $\overrightarrow{B_3B_4}$. We begin by relating $\alpha_2$ with $\alpha_3$, the angles which the vectors $\overrightarrow{B_2B_3}$ and $\overrightarrow{B_3B_4}$, respectively, form with the vertical axis. For these purposes we resolved the system of three equations, of unknown $\alpha_2$, $\theta_3$ and $\beta_2$, which are taken from the geometry of the figure,[^3] $$\left\{ \begin{array}{l} \theta_3 =\beta_2 + \alpha_3\\ \arctan \left(\frac{1}{2}y_3\right)+\beta_2 = \frac{\pi}{2}\\ \alpha_2 +\theta_3+\beta_2 = \pi \end{array} \right.$$ obtaining $$\label{eq:alpha2} \alpha_2 =2\arctan({y_3}/{2})-\alpha_3 \text{,}$$ in which $\arctan(y_3/2)$ is the angle which the straight line tangent at the curve in $B_3$ makes with the vertical — the inclination of the straight line tangent is given by $\frac{\mathrm{d} x}{\mathrm{d} y}\left.\right|_{y=y_3} =\frac{\mathrm{d}}{\mathrm{d} y}(\frac{1}{4}y^2-\frac{1}{2})\left.\right|_{y=y_3} =\frac{1}{2}y_3$. In its turn, the angle $\alpha_3$ can be expressed in the following way: $\alpha_3 =\arctan(({x_3-x_4})/({y_3-y_4})) =\arctan(({\frac{1}{4}y_3^2-\frac{1}{4}y_4^2})/({y_3-y_4})) =\arctan({(y_3+y_4)}/{4})\text{,}$ which permits us to write $\alpha_2$ in function only of the ordinates $y_3$ and $y_4$ of the extremes of the vector $\overrightarrow{B_3B_4}$, $$\label{eq:alpha2b} \alpha_2 =2\arctan ({y_3}/{2})-\arctan(({y_3+y_4})/{4}) \text{.}$$ In order to be able to prove what we intend — impossibility of occurrence of the reflection $B_4$ — we need to find a lower bound for the ordinate of the position where each one of the four reflections occurs, or in other words, to determine $\left\{y_1^*,y_2^*,y_3^*,y_4^*\right\}$, just that $$\label{eq:minorantes} y_1\geq y_1^*,\;y_2\geq y_2^*,\;y_3\geq y_3^*,\;y_4\geq y_4^*,\; \;\forall (\varphi,x)\in(\varphi_0,{\pi}/{2})\times (-{1}/{2},{1}/{2})\text{.}$$ It can easily be understood that $y_4^*=0$. We will therefore determine the other three lower bounds, commencing with $y_2^*$. We know that $0<y_4<y_3$; therefore, from we take away that $$\label{eq:alpha2Limites} \arctan({y_3}/{2})<\alpha_2 < 2\arctan ({y_3}/{2})-\arctan ({y_3}/{4}) \text{.}$$ Being aware that $\alpha_2$ is situated in the first quadrant of the trigonometrical circle, we can maintain the inequalities for the tangent of the respective angles. After some algebraic simplifications, we obtain $$\label{eq:tanAlpha2Limites} {y_3}/{2}<\tan(\alpha_2) < {y_3\left(12+y_3^2\right)}/{16} \text{.}$$ The equation of the straight line which connects $B_2$ to $B_3$ takes the form $x=m(y-y_3)+x_3,$ with $m=\tan(\alpha_2)$ and $x_3=\frac{1}{4}y_3^2-\frac{1}{2}$. As we are interested in finding the ordinate of the point of interception of this straight line with the parabolic curve situated on the right side, with equation $x=-\frac{1}{4}y^2+\frac{1}{2} \text{,}$ we have to resolve the equation of second degree, in the variable $y$, which results in the elimination of the variable $x$ by combination of the two previous equations. The ordinate $y_2$, of the second reflection, being the positive root of the equation, takes the form $y_2 = -2 m+\sqrt{4 m^2+4 m y_3 - y_3^2+4} \text{ .}$ The magnitude $y_2$ is expressed in function of two variables, $m$ and $y_3$, which as we know assume only positive values. So as to accept more easily the deductions which we are going to make in the tracking of $y_2^*$, we will imagine, without any loss of generality, that $y_3$ is a fixed value. We begin by showing that the derivative of $y_2$ in order to the variable $m$, $$\label{eq:derivaday2} \frac{\mathrm{d} y_2}{\mathrm{d} m}= \frac{2\left(2m+y_3 -\sqrt{4 m^2+4m y_3 -y_3^2 + 4}\right)}{\sqrt{4 m^2+4m y_3 -y_3^2 + 4}} \text{ ,}$$ has a negative value for whatever value of $y_3$ is. As $y_3<\sqrt{2}$, inevitably $y_3^2<4$, thus the two radicands $(4m^2+4my_3-y_3^2+4)$ present in the equation  have always a positive value. The restriction $y_3<\sqrt{2}$ allows us still to successively deduce the following inequalities $$\begin{aligned} y_3^2<2 \Leftrightarrow 2 y_3^2<4 \Leftrightarrow y_3^2<4-y_3^2 \Leftrightarrow 4 m^2+4 m y_3 + y_3^2<4 m^2+4 m y_3 +4-y_3^2\\ \Leftrightarrow \left(2 m+ y_3\right)^2<4 m^2+4 m y_3 -y_3^2+4 \Leftrightarrow 2 m+ y_3<\sqrt{4 m^2+4 m y_3 -y_3^2+4} \text{ .}\end{aligned}$$ This last inequality confirms that $\frac{\mathrm{d} y_2}{\mathrm{d} m}<0$, for whatever $y_3$ may be. In this way, the value $y_2$ is so much less the greater is the value of $m$. As is examined in , $m<M={y_3\left(12+y_3^2\right)}/{16}$, thus $y_2 > -2 M+\sqrt{4 M^2+4 M y_3 - y_3^2+4} \text{ .}$ Substituting $M$, there is obtained, after some simplifications, $$\label{eq:y2minor} y_2 > f(y_3), \text{ with } f(y_3)=-\frac{3}{2} y_3-\frac{1}{8} y_3^3+\frac{1}{8} \sqrt{272 y_3^2+40 y_3^4+y_3^6+256} \text{ .}$$ In order to find the minimum value of $f(y_3)$ we begin by computing its derivative: $$\frac{\mathrm{d}}{\mathrm{d} y_3}f(y_3)= \frac{272 y_3+80 y_3^3+3 y_3^5-(12+3 y_3^2) \sqrt{272 y_3^2+40 y_3^4+y_3^6+256}} {8 \sqrt{272 y_3^2+40 y_3^4+y_3^6+256}} \text{.}$$ The radicands being clearly positive, we only have to concern ourselves with the numerator of the fraction. To find the roots of the derivative function $\frac{\mathrm{d}}{\mathrm{d} y_3}f(y_3)$ is equivalent because of this to resolving the equation $$\left(272 y_3+80 y_3^3+3 y_3^5\right)^2=\left(12+3 y_3^2\right)^2 \left(272 y_3^2+40 y_3^4+y_3^6+256\right) \text{,}$$ which can be simplified in the following: $$2304-1024 y_3^2-992 y_3^4-160 y_3^6-3 y_3^8=0 \text{.}$$ This polynomial equation has only one real positive root, of the value $\tilde{y}_3=\frac{2}{3} \sqrt{-51 + 6 \sqrt{79}}\simeq 1.017 \text{,}$ signifying that $f(y_3)$ has a global minimum in $\tilde{y}_3$, because, as we show in what follows, $\frac{\mathrm{d}^2}{\mathrm{d} y_3^2}f(y_3)>0$ and the function does not presents other points of stationarity. We thus show that $\frac{\mathrm{d}^2}{\mathrm{d} y_3^2}f(y_3)>0$, with $$\begin{aligned} \label{eq:y2segDeriv} \frac{\mathrm{d}^2}{\mathrm{d} y_3^2}f(y_3)= \frac{ 18240 y_3^4+2960 y_3^6+180 y_3^8+3 y_3^{10}+34816+30720 y_3^2 }{ 4 (272 y_3^2+40 y_3^4+y_3^6+256)^\frac{3}{2} } \nonumber\\ -\frac{ (816 y_3^3+120 y_3^5+3 y_3^7+768 y_3) \sqrt{272 y_3^2+40 y_3^4+y_3^6+256} }{ 4 (272 y_3^2+40 y_3^4+y_3^6+256)^\frac{3}{2} } \text{.}\end{aligned}$$ We show that $\frac{\mathrm{d}^2}{\mathrm{d} y_3^2}f(y_3)>0$ is equivalent to showing that the numerator of the first fraction is superior to the numerator of the second, in . After elevating the two terms to the square, we arrive at the inequality $$\begin{aligned} -3 y_3^{16}+36 y_3^{14}+9464 y_3^{12}+191616 y_3^{10}+1514112 y_3^8+5817344 y_3^6 \nonumber\\ +13535232 y_3^4+15532032 y_3^2+9469952>0 \text{.}\end{aligned}$$ We easily prove the veracity of this relation, given that we have a unique negative term ($-3 y_3^{16})$ which, for example, is inferior in absolute value to the constant term ($9469952$),  $y_3<\sqrt{2} \Rightarrow 3 y_3^{16}<768<9469952 \text{.}$ The proof is thus complete that $f(y_3)$ has a global minimum in $\tilde{y}_3$, of the value $f(\tilde{y}_3)=\frac{8}{9} \sqrt{-51 + 6 \sqrt{79}} \text{.}$ Accordingly, from , we finally conclude that $$\label{eq:y2min} y_2>y_2^*=\frac{8}{9} \sqrt{-51 + 6 \sqrt{79}} \simeq 1.356 \text{.}$$ A lower bound is then found for the height of the second reflection $B_2$ (illustration (a) of figure \[fig:parabOpt3colB\]). We now determine $y_3^*$, a lower bound for the height of the third reflection $B_3$. So that the reflection $B_2$ occurs in the parabola of the right side it is necessary that the angle $\alpha_2$ is greater than the angle formed between the vertical axis and the segment of straight line which unites $B_3$ with the superior vertex of the cavity, $$\label{eq:alpha2Maj} \alpha_2> \arctan \Big(\frac{-x_3}{\sqrt{2}-y_3}\Big) =\arctan \Big(\frac{-\frac{1}{4}y_3^2+\frac{1}{2}}{\sqrt{2}-y_3}\Big) =\arctan \Big(\frac{\sqrt{2}+y_3}{4}\Big) \text{.}$$ This inequality, in conjunction with the second relation of inequality of , allows us to write $$({\sqrt{2}+y_3})/{4}<\tan(\alpha_2) < {y_3\left(12+y_3^2\right)}/{16} \Rightarrow ({\sqrt{2}+y_3})/{4} < {y_3\left(12+y_3^2\right)}/{16} \text{,}$$ from which results the inequality $$\label{eq:ineq3g} y_3^3+ 8y_3+ 4\sqrt{2}>0 \text{.}$$ As the polynomial of the left hand-side of  has a positive derivative and admits a unique real root, we immediately conclude that it constitutes an inferior limit for $y_3$, this limit being $$\label{eq:y3min} y_3^*=\frac{1}{3}\left(54\sqrt{2}+6\sqrt{546}\right)^\frac{1}{3} -{8}{\left(54\sqrt{2}+6\sqrt{546}\right)^\frac{-1}{3}} \simeq 0.670 \text{.}$$ It is left to us to determine $y_1^*$, a lower bound for the value of $y_1$ — the ordinate where the first reflection occurs. To this end we resort to illustration (b) of figure \[fig:parabOpt3colB\], which gives us a more detailed representation of the part of the cavity where the first two reflections occur, $B_1$ and $B_2$. The scheme presented was constructed counting that the first reflection ($B_1$) occurs at a point which is more elevated than that of the third reflection ($B_3$). This is in fact the situation. This itself can be proven by showing that $\alpha_2$ is always smaller than the angle formed between the normal vector at the curve in $B_2$ and the vertical axis, or that is $\alpha_2<\frac{\pi}{2}-\arctan\left(\frac{1}{2}y_2\right) \text{.}$ Taking, in , at the upper limit of $\tan(\alpha_2)$ and having in mind that $y<\sqrt{2}$, we build the following sequence of inequalities which proves what is intended: $$\alpha_2 < \arctan \left(\frac{y_3\left(12+y_3^2\right)}{16}\right) < \overbrace{ \arctan \left(\frac{7\sqrt{2}}{8}\right) }^{\simeq 51.06^\circ}< \overbrace{\frac{\pi}{2}-\arctan\left(\frac{\sqrt{2}}{2}\right)}^{\simeq 54.74^\circ} <\frac{\pi}{2}-\arctan\left(\frac{1}{2}y_2\right) \text{.}$$ We now try to find $y_1^*$. We can define $y_1^*$ as being the ordinate of the point of interception of the left parabola with the semi-straight line with its origin at point $B_2$, positioning as low as possible ($y_2=y_2^*$), and with equal slope to the largest value permitted for the slope of the trajectory which preceded $B_2$ ($\overrightarrow{B_1B_2}$). The equation of the straight line which connects $B_1$ to $B_2$ takes the form $x=m(y-y_2)+x_2\text{,}$ with $m=\tan(\alpha_1)$ and $x_2=-\frac{1}{4}y_2^2+\frac{1}{2}$. As we are interested in finding the point of intersection of this straight line with the parabolic curve situated on the left side, with equation $x=\frac{1}{4}y^2-\frac{1}{2} \text{ ,}$ we will have to resolve the equation of the second degree, in the variable $y$, which results in the elimination of the variable $x$ by combination of the two previous equations. Although we have two positive real roots, we are only interested in the smaller of the two, which takes the form $$\label{eq:y1} y_1 = 2 m-\sqrt{4 m^2-4 m y_2 - y_2^2+4} \text{ .}$$ As we said, if we do $y_2=y_2^*$ and place the maximum slope to the straight line, which in the previous equations is equivalent to considering $m$ minimum, we obtain $y_1 = y_1^*$. Being $m=\tan(\alpha_1)$, we should determine the value of $\alpha_1$ through the system of equations $$\left\{ \begin{array}{l} \alpha_1 =2\theta_2 + \alpha_2\\ \arctan \left(\frac{1}{2}y_2\right)+\alpha_2 +\theta_2 = \frac{\pi}{2} \end{array} \right.$$ which is taken from the geometry of illustration (b) of figure \[fig:parabOpt3colB\]. Is obtained $$\label{eq:alpha1} \alpha_1 =\pi-2\arctan ({y_2}/{2})-\alpha_2 \text{.}$$ From this latest equality, from , and given that $y_2<\sqrt{2}$ and $y_3<\sqrt{2}$, we deduce that $\alpha_1 > \pi-2\arctan ({y_2}/{2})-\arctan ({y_3\left(12+y_3^2\right)}/{16})> \pi-2\arctan ({1/\sqrt{2}})-\arctan (7\sqrt{2}/8) \text{,}$ thus $m=\tan(\alpha_1) > \tan( \pi-2\arctan (1/{\sqrt{2}})-\arctan (7\sqrt{2}/8))= \frac{23}{20}\sqrt{2} \text{.}$ If in we do $m=\frac{23}{20}\sqrt{2}$ and $y_2 = y_2^*$ we then obtain the inferior limit for $y_1$ $$\label{eq:y1min} y_1^*= \frac{23}{10}\sqrt{2} -\frac{1}{90}\sqrt{ 444498-33120\sqrt{2}\sqrt{-51+6\sqrt{79}} -38400\sqrt{79} } \simeq 1.274 \text{.}$$ Summarizing, $(y_1^*,y_2^*,y_3^*,y_4^*)\simeq (1.274, 1.356, 0.670, 0) \text{.}$ With the help of illustration (b) of figure \[fig:parabOpt3colB\] we will, finally, analyze the entry angle $\varphi$ of the particle. With the system of equations ($\frac{1}{2}y_1=\frac{\mathrm{d} x}{\mathrm{d} y}\left.\right|_{y=y_1}$) $$\left\{ \begin{array}{l} 2\theta_1+\varphi +\alpha_1 = \pi\\ \arctan \left(\frac{1}{2}y_1\right)+\varphi +\theta_1 = \frac{\pi}{2} \end{array} \right.$$ and with equalities and we obtain, successively, $$\begin{aligned} \varphi &=&\alpha_1-2\arctan ({y_1}/{2}) \text{,}\nonumber\\ \label{eq:phiP2} \varphi &=& \pi-2\arctan ({y_1}/{2}) -2\arctan ({y_2}/{2})-\alpha_2 \text{,}\\ \varphi &=& \pi-2\arctan ({y_1}/{2}) -2\arctan ({y_2}/{2}) -2\arctan ({y_3}/{2})+\arctan ({(y_3+y_4)}/{4}) \text{.} \nonumber\end{aligned}$$ Taking , from we deduce that $\varphi < \pi-2\arctan \left({y_1}/{2}\right) -2\arctan \left({y_2}/{2}\right) -\arctan ({(\sqrt{2}+y_3)}/{4}) \text{.}$ In agreement with the definitions and with the values found in , and , we can conclude that $$\varphi < \pi-2\arctan ({y_1^*}/{2}) -2\arctan ({y_2^*}/{2}) -\arctan ({(\sqrt{2}+y_3^*)}/{4}) \simeq 19.18^\circ \text{,}$$ or that is $$\varphi < \varphi_0 \simeq 19.47^\circ \text{.}$$ With this we can finally conclude that it is impossible to have a fourth reflection, since for this to happen the particle would have to have entered in the cavity with an angle $\varphi < \varphi_0$, as we have just finished showing — something which would contradict our initial imposition, $\varphi > \varphi_0$. As the cavity presents symmetry in relation to its central vertical axis, the conclusion to which we have arrived is equally valid for $\varphi < -\varphi_0$, thus being proven that which we intended (theorem \[teor:3ref\]): > [*For $\left|\varphi\right|>\varphi_0$, there always occur three reflections, alternatively on the left and right facets of the Double Parabola cavity.* ]{} Proof of theorem \[teor:min3ref\] {#cha:min3col} ================================= For $\left|\varphi\right|>\varphi_0$ the statement of theorem \[teor:min3ref\] is already proved. It remains to consider the case $-\varphi_0<\varphi<\varphi_0$, but given the symmetry of the cavity we will only need to study the interval $0<\varphi<\varphi_0$. We will base our proof on some of the deductions which we made in appendix \[cha:condSuf3col\], the illustrations of figure \[fig:parabOpt3col\] being especially useful to us. We will also assume to be truthful the following premise: “If the second reflection happens on the same facet of the cavity where the first reflection occurred, there will necessarily be a third reflection”. We exempt ourselves from proving this principle because it appears evident to us. For $0<\varphi<\varphi_0$ the first reflection can just as well occur on the left-hand facet as on the right-hand side. We will analyze each one of the cases separately. *1st reflection on the left-hand side* Being $0<\varphi<\varphi_0$ we can have the first two reflections on the left-hand facet, it being in this case guaranteed, as we assume above, that $3$ or more reflections will exist. If on the other hand, the second reflection is on the right side, an initial part of the trajectory can always be represented by the first three illustrations of figure \[fig:parabOpt3col\] (assuming $0<\varphi<\varphi_0$), which guarantee, also in this case, the existence of a third reflection $B_3$. In order to prove what we have just finished saying, it will be enough to prove the ascendant nature of the sub-trajectory $\overrightarrow{B_1B_2}$. We establish on the parabola of the left side (illustration (a) of figure \[fig:parabOpt3col\]) the first point of reflection $B_1$. For whatever $B_1$ is it is always possible for us to trace an initial sub-trajectory $\overrightarrow{B_0B_1}$ with its origin in an entry angle $\varphi>\varphi_0$. As in appendix \[cha:condSuf3col\] (page ) we showed the sub-trajectory $\overrightarrow{B_1B_2}$ which would follow it to be ascendant, the same will necessarily come about for whatever $0<\varphi<\varphi_0$ may be, given that in this case $\overrightarrow{B_0B_1}$ will represent a more accentuated negative slope. Since in appendix \[cha:condSuf3col\] (page ) we characterized $\overrightarrow{B_2B_3}$ only with basis in the ascendant nature of the sub-trajectory preceding $\overrightarrow{B_1B_2}$, the conclusions to which we arrive for $\overrightarrow{B_2B_3}$ are equally valid for $0<\varphi<\varphi_0$. *1st reflection on the right side* Also in this case we can have the first two reflections on the right-hand facet, it being guaranteed that $3$ or more reflections will exist. If this does not occur, we will necessarily have a trajectory with the aspect of the trajectory $B_0B_1B_2B_3$ illustrated in the scheme of figure \[fig:parabOptMin3col\], where as well there are represented two auxiliary trajectories (the dotted lines), $A_0B_1A_2$ and $A_1B_2A_3$, which, on passing through the foci of the parabolas, present the sub-trajectory posterior to the reflection horizontal. ![Illustrative scheme to the study of the trajectory of particles with entry angle $0<\varphi<\varphi_0$.[]{data-label="fig:parabOptMin3col"}](fig20){width="0.255\columnwidth"} Having as a basis the laws of reflection, we can succinctly deduce the following: as the angle $\widehat{A_0B_1A_2}$ must be interior to the angle $\widehat{B_0B_1B_2}$, we conclude that $\overrightarrow{B_1B_2}$ is of a ascendant nature; as $\widehat{B_1B_2B_3}$ is necessarily an interior angle to $\widehat{A_1B_2A_3}$, we conclude that $B_3$ must be situated between $A_1$ and $A_3$, which guarantees the existence of a third reflection. Thus is the proof of the theorem \[teor:min3ref\] concluded. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by the *Centre for Research on Optimization and Control* (CEOC) from the *Portuguese Foundation for Science and Technology* (FCT), cofinanced by the *European Community Fund* (ECF) ; and by the FCT research project . [10]{} Isaac Newton. Philosophiae naturalis principia mathematica. , 1687. G. Buttazzo, V. Ferone, and B. Kawohl. Minimum problems over sets of concave functions and related questions. , 173:71–89, 1995. G. Buttazzo and P. Guasoni. Shape optimization problems over classes of convex domains. , 4(2):343–351, 1997. G. Buttazzo and B. Kawohl. On [N]{}ewton’s problem of minimal resistance. , 15:7–12, 1993. T. Lachand-Robert and E. Oudet. Minimizing within convex bodies using a convex hull method. , 16:368–379, 2006. T. Lachand-Robert and M. A. Peletier. An example of non-convex minimization and an application to [N]{}ewton’s problem of the body of least resistance. , 18:179–198, 2001. F. Brock, V. Ferone, and B. Kawohl. A symmetry problem in the calculus of variations. , 4:593–599, 1996. M. Comte and T. Lachand-Robert. ewton’s problem of the body of minimal resistance under a single-impact assumption. , 12:173–211, 2001. T. Lachand-Robert and M. A. Peletier. ewton’s problem of the body of minimal resistance in the class of convex developable functions. , 226:153–176, 2001. Alexander Yu. Plakhov. ewton’s problem of a body of minimal aerodynamic resistance. , 390(3):314–317, 2003. Alexander Yu. Plakhov. ewton’s problem of the body of minimal resistance with a bounded number of collisions. , 58(1):191–192, 2003. Alexander Yu. Plakhov. ewton’s problem of the body of minimum mean resistance. , 195(7–8):1017–1037, 2004. D. Horstmann, B. Kawohl, and P. Villaggio. ewton’s aerodynamic problem in the presence of friction. , 9:295–307, 2002. Alexander Yu. Plakhov and Delfim F. M. Torres. ewton’s aerodynamic problem in a medium of chaotically moving particles. , 196(6):111–160, 2005. Alexander Yu. Plakhov Billiards and two-dimensional problems of optimal resistance. , DOI 10.1007/s00205-008-0137-1 (33 pp.), 2008. Alexander Yu. Plakhov and Paulo D. F. Gouveia. . Cadernos de Matemática, CM07/I-12, Universidade de Aveiro, Abril 2007. Alexander Yu. Plakhov and Paulo D. F. Gouveia. Problems of maximal mean resistance on the plane. , 20:2271–2287, 2007. Brian W. Kernighan and Dennis M. Ritchie. . Prentice-Hall, second edition, 1988. Genetic algorithm and direct search toolbox user’s guide – for use with [MATLAB]{}. The MathWorks, Inc., 2004. D. G. Hook and P. R. McAree. Using sturm sequences to bracket real roots of polynomial equations. In [*Graphics gems*]{}, pages 416–422. Academic Press Professional, Inc., San Diego, CA, USA, 1990. Paulo D. F. Gouveia. . Ph.D. thesis, Univ. of Aveiro, Portugal, February 2008. [^1]: Accepted (April 16, 2009) for publication in the journal “Applied Mathematics and Computation”. [^2]: In the bi-quadratic curves, the point of interception of the trajectory of the particle with the boundary of the cavity is calculated by resolving an equation of the 4th degree. The roots of this equation were obtained numerically using the method described in [@Hook90]. The equations of inferior order were always resolved by utilising the known analytical formulas. [^3]: The variables denoted by $x_i$ and $y_i$, with $i =1,\ldots,4$, represent the coordinates of $i$-th point of reflection, identified by $B_i$.
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this short note we show that (I) in a QCD-like theory with four (rather than two) degenerate flavors $ud\,u''d''$, the $\pi\pi''$ scattering length is positive (attractive); and (II) in QCD with only two (u,d) degenerate flavors the I=2 (say, $\pi^+\pi^+$ hadronic) scattering length is, in the large $N_C$ limit, repulsive. $\pi(\pi'')$ are the lowest physical states coupling to $J^p = \bar{u}(x)\gamma_5d(x)$ and $J^{p''} = \bar{u}''(x)\gamma_5d''(x),$ respectively.' address: | School of Physics and Astronomy\ Tel Aviv University, Tel Aviv, Israel\ and\ Department of Physics and Astronomy\ University of South Carolina\ Columbia, South Carolina 29208 USA\ [ ]{} author: - 'S. Nussinov' date: October 2000 title: | QCD Inequalities, Large\ Scattering Lengths --- To probe the threshold $\pi\pi$ (or ($\pi'\pi'$)) scattering consider the Euclidean two-point correlators: $$Q(x) = <0|J^p(x)J^p(x)(J^p(0)J^p(0))^+|0>$$ $$Q'(x) = <0|J^p(x)J^{p'}(x)(J^p(0)J^{p'}(0))^+|0>$$ where $J^p = \bar{u}(x)\gamma_5d(x)$ and $J^{p'} = \bar{u}'(x)\gamma_5d'(x),$ respectively. The correlators have spectral decompositions $$Q(x) = \int_{\mu_0}^{+\infty} \sigma_Q(\mu^2) {\rm exp}(-\mu x)d\mu^2$$ $$Q'(x) = \int_{\mu_0}^{+\infty} \sigma_Q'(\mu^2) {\rm exp}(-\mu x)d\mu^2$$ with a common threshold $\mu_0 = 2m_\pi$ (or $\mu_0 = m_\pi + m_{\pi'}$). The asymptotic behavior of $Q(x)\, (Q'(x))$ in the limit $|x| \rightarrow \infty$ (or Euclidear time separating the points $(0 = (\vec{0},0)$ and $x = (\vec{0},t), \; t \rightarrow \infty)$ is controlled by the spectral density $\sigma_Q$ or $\sigma_{Q'}$ at $\mu = 2m_\pi$. The $\pi\pi$ (or $\pi\pi'$) scattering at threshold can be described using nonrelativistic methods. The Levinson’s theorem implies that the density of states in any convenient large radius R of the $\pi\pi \, (\pi\pi')$ system deviates from that of a free-noninteracting $\pi\pi \, (\pi\pi')$ pair by $$\frac{dn_{\pi\pi}}{d\mu^2} - \frac{dn_{\pi\pi}^{(0)}}{d\mu^2} = \frac{d\delta (k)^{(L=0)}}{dk} \approx a^{(L=0)}$$ with $k = 1/2 \sqrt{\mu^2-4m_\pi^2}$ the CMS momentum, and $\delta^{(0)}$ and $a^{(0)}$ are the S-wave phase, shift and scattering length of the $\pi\pi$ system. Analogously for $\pi\pi'$, $$\frac{dn_{\pi\pi'}}{d\mu^2} - \frac{dn_{\pi\pi'}^{(0)}}{d\mu^2} = \frac{d\delta (k)^{(L=0)}}{dk} \approx a'^{(L=0)} .$$ Because hadronic forces have finite range only the $L=0$ wave needs to be considered. By definition, $$\frac{dn_{\pi\pi}}{d\mu^2} \propto \sigma_Q(\mu^2); \; \frac{dn_{\pi\pi'}}{d\mu^2} \propto \sigma_Q'(\mu^2) .$$ To prove (I), we will show that $$Q'(x) \geq \; <0|J^p(x)J^{p^+}(0)|0> <0|J^{p'}(x)J^{p'^+}(0)|0>.$$ That is, the joint propagation of the four quarks—which asymptotically becomes the joint $\pi$ and $\pi'$ propagation from 0 to $x$—is enhanced relative to the product of the separate $J^pJ^p$ and $J^{p'}J^{p'}$ correlators. The latter is in fact the independent $\pi$ and $\pi'$ propagations between 0 and $x$. To prove Eq. (8), we use the path integral representation of the various correlators. Since all the four-quark flavors created at 0 and annihilated at $x$ are distinct, there is a unique “contraction” for $Q'(x)$ or Eq. (2) yielding: $$\begin{aligned} Q'(x) & = & \int d\mu(A)\;{\rm tr}\; \gamma_5S_{u_A}(x,0)\gamma_5S_{d_A}(0,x) \cdot {\rm tr}\; \gamma_5S_{u'_A}(x,0)\gamma_5S_{d'_A}(0,x)\\ \nonumber & = & \int d\mu(A) {\rm tr}(S_A^+(0,x)S_A(0,x)) \cdot {\rm tr}\; (S_A^+(0,x)S_A(0,x))\end{aligned}$$ where $$d\mu(A) = D(A_\mu(x)) \cdot e^{-SYM\{A_\mu(x)\}} \Pi_{u,d,u'd'} Det(D\!\!\!\!/_A^{q_i}+m_i)/Z .$$ is the positive[@Wei; @VandW] measure in the path integral. $D\!\!\!\!/_A = \gamma_\nu(\delta_\nu + gA_\mu^a\lambda_a)$ is the covariant Dirac operator in the $A_\mu^a(x)$ background, and $S_{i_A}(x) = <0|\frac{1}{D\!\!\!\!/_A+m_i^0}|x>$ is the propagator of the Fermion (i.e., quark of flavor i and bare mass $m_i^0$) in the $A_\mu^a(x)$ background. Finally the partition function Z in the denominator of Eq. (10) normalizes the path integral measure to $\int d\mu(A) = 1$. In deriving the second line of Eq. (9), we have used the “$\gamma_5$ conjugation” property of the Euclidean propagator, $$\gamma_5 \; S_A^i(0,x)\gamma_5 = [S_A^i(x,0)]^\dagger$$ with the $\dagger$ referring here to Hermitian conjugation in color-spinor space of the ${S_A^i(x,0)}_{\alpha\alpha',aa'}\;\\ 12~\times~12$ matrix with spinor $1 \leq \alpha, \alpha' \leq 4$ and color $1 \leq a,a' \leq N_C$ indices, which for simplicity we generally omit. The latter $\gamma_5$ conjugation also implies the positivity of the determinant of the Dirac operator $(D\!\!\!\!/_A+m^0)$, which is the key to the claimed positivity of the measure $d\mu(A)$. The proofs of all these appear in the original paper of Weingarten[@Wei] and of Vafa and Witten [@VandW] and in a recent comprehensive review[@NandL]. Finally, we utilized the fact that since $m_u^{0} = m_d^{0} = m_{u'}^{0} = m_{d'}^{0}$, all four propagators are actually the same. $$S_A^u(x,0) = S_A^d(x,0) = S_A^{u'}(x,0) = S_A^{d'}(x,0) \equiv S_A(x,0).$$ The inequality $$Q'(x) \geq P(x) P'(x)$$ with $$P(x) = <0|J^p(x)(J^p(0))^+|0> \; ; \; P'(x) = <0|J^{p'}(x)(J^{p'}(0))^+|0>$$ can be readily derived[@NandL]. To this end we compare $Q'(x)$—as given by the path integral (second line of Eq. (9))—and the product of path integrals for $P(x),P'(x)$: $$P(x) = \int d\mu(A) {\rm tr}(S_A^+(0,x)S_A(0,x))$$ and $$P'(x) = \int d\mu(A) {\rm tr}(S'^+_A(0,x) S'_A(0,x))$$ (which in the $m_u^{(0)} = m_{u'}^{(0)};m_d^{(0)} = m_{d'}^{(0)}$ case of interest are equal as $S_A = S'_A$). Let us denote $${\rm tr}\; S_A^+(0,x) S_A(0,x) \equiv \pi_A(x) \geq 0 .$$ The desired inequality (13) then is (as most QCD correlator inequalities are) just the Schwartz inequality; $$\int d\mu(A) \pi_A(x)\pi_A(x) \geq |\int d \mu(A) \pi_A(x)|^2.$$ One dramatic way in which Eq. (18) could be implemented is if a four-quark [*bound*]{} (scalar?) doubly-charged state $\bar{u}\gamma_5d \; \bar{u'}\gamma_5d$ existed below the $\pi^+\pi'^+ (= 2m_\pi)$ threshold[@Espr]. We are trying to mimic real QCD where we know that exotic $q\bar{q}q\bar{q}$ mesons—should they exist at all—are much heavier than $2m_\pi \approx$ 270 MeV. Hence we will [*assume*]{} that there are no such bound states below threshold. Using then Eqs. (5)-(7) above, our inequality (18) implies an attractive (positive) S-wave $\pi\pi'$ scattering length; namely, assertion I above. One can also directly argue by going to the nonrelativistic limit and viewing the path integral expression as an euclidean diffusive evolution, that the probability of returning to the origin $\vec{r}=0$ after some long time T is enhanced [*iff*]{} the interaction between the $\pi\pi'$ is attractive.[@NusSat] A related, more ambitious approach uses four-point inequalities. One might argue that the introduction of $\pi'$ (and the corresponding extra two flavors, $u'd'$, degenerate with $u$ and $d$) is unphysical and renders this result meaningless. Such an argument could be even more forcefully made against Weingarten’s proof[@Wei] of $m_N \geq m_\pi$ where degenerate yet distinct flavors were introduced. We believe that this is not the case. In particular, the augmentation of the flavor sector can be useful for picking up specific flavor contraction patterns in the real QCD with two (or three) light (almost) degenerate flavors. Thus, let us consider $\pi\pi$ scattering in QCD (see Fig. 1). \#1\#2[1.75\#2]{} We can have separate, independent propagation of two pions, represented in Fig. 1(a), where we have gluons exchanged only between $u\bar{d}$ (or $\bar{u}d$) in the same initial and final pions. For $\pi^+\pi^-$ scattering we have the flavor connected contraction depicted in Fig. 1(b), corresponding to $q\bar{q}$ annihilation. Likewise for $\pi^+\pi^+$ scattering, we have instead the quark exchange diagram, Fig. 1(c). Finally, for all pion pairs—and for the $\pi\pi'$ case—we can have the $\bar{q}q$ bubbles interaction also via gluon exchanges (at least two gluons are required because of the color neutrality), as indicated in Fig. 1(d). A key point is that for $\pi\pi'$ scattering we have the free separate propagation (Fig. 1(a)) and gluon exchanges (Fig. 1(d)): the $\pi$ and $\pi'$ have no common quarks and/or anti-quarks to allow for $\bar{q}q'$ annihilation and/or $\dot{q}q'$ exchange. The same arguments could be made in any vectorial theory and, in particular, in QED. This then conforms to the attractive perturbative two-photon exchange, $1/r^6$ van der Waals (VDW) potential and its “retarded” Casimir Polder version at large distances. For two polarizable atoms $A,B$ the latter is $$V_{AB}^{CP} = - \frac{\hbar c}{(4\pi)^3r^7} [23(\alpha^E_A \alpha^E_B + \alpha^M_A \alpha^M_B) -7(\alpha^E_A \alpha^M_B + \alpha^E_B \alpha^M_A)]$$ The VDW interaction is, by second-order perturbation, always attractive between two stationary systems in their ground state. The CP interaction is manifestly positive if A and B are dynamically the same, i.e., $\alpha^E_A = \alpha^E_B$ and $\alpha^M_A = \alpha^M_B$. Amusingly, the case for which we have been able to generalize this result satisfies both requirements. First, we know (and it can also be proven via QCD inequalities) that the pseudoscalar $\pi$ (and $\pi'$) are indeed the lightest mesons. Second, since $m_u^0 = m_d^0 (= m_{u'}^0 = m_{d'}^0)$ and the gauge interaction are flavor independent, the scattering “A and B”, i.e., $\pi$ and $\pi$ (or $\pi$ and $\pi'$) are dynamically the same. We note that $V_{CP}$ of Eq. (19) remains attractive so long as the objects A and B have similar ratios of electrical and magnetical polarizabilities. This is expected to be the case for the “chromo” polarizabilities of different hadrons, suggesting that $(J/\psi-A,Z)(D_s-A,Z)(D_s,B_u)$, etc., bound states of heavy/extended hadrons sharing no common quark flavors will form. We next turn to our second main goal; namely, point II above. With only $u$ and $d$ degenerate quark flavors we have for $\pi^+\pi^-$ ($\pi^+\pi^+$) scattering, or, for the correlators: $$<0|J^{{p_s}^+}(x)J^{{p_s}^-}(x)(J^{{p_s}^+}(0)J^{{p_s}^-}(0)^+|0> = Q_{\pi^+\pi^-}$$ $$<0|J^{{p_s}^+}(x)J^{{p_s}^+}(x)(J^{{p_s}^+}(0)J^{{p_s}^+}(0)^+|0> = Q_{\pi^+\pi}\; ,$$ the contribution of the additional contractions. These are $$Q_{\pi^+\pi^-}(x)|_{Annihilation} = \int \, d\nu(A) {\rm tr} \, S_A^{+d}(0,x)S_A^u(x,x)S_A^d(x,0)S_A^{u^+}(0,0)$$ and $$Q_{\pi^+\pi^+}(x)|_{Exchange} = - \int \, d\nu(A) {\rm tr}\{S_A^+(0,x)S(0,x)S_A^+(0,x)S_A^d(x,0)S_A(0,x)\},$$ respectively. The crucial minus sign in Eq. (23) stems from the need to permute two $\psi_u(x)$ (or two $\psi_u(0)$) noncommuting operators so as to arrive from the $\int \, d\nu(A) {\rm tr}\{S_A^+(0,x)S_A^+(0,x)\}^2$ contraction pattern of Fig. 1(a) + 1(d) to Eq. (23) and the contraction 1(c). In the large $N_C$ limit the contributions (22) and (23) to the correlators dominate. It is tempting to associate these—in the chiral, threshold limit—with the classical[@AandD] results, $a^{I=0} \approx 0,2\,m_\pi^{-1}$ and $a^{I=2} = -\frac{2}{7} \; a^{I=0}$. Higher-order corrections in the chiral expansions preserve the negataive (repulsive) $a^{I=2}$—in agreement with experimental measurements which unfortunately are rather poor for $a^{I=2}$. This is particularly gratifying in view that we have an opposite sign contribution (due to the multi-gluon, Fig. 1(d)) which does not, however, reverse the sign. Let us conclude with two comments. [*(i)*]{} The expressions (22) and (23) for $Q_{\pi^+\pi^-}$ and $Q_{\pi^+\pi^+}$ seem drastically different. However, in the threshold chiral limit they may both converge to similar objects. The point is that as $(x)$ (or $t$) tends to $\infty$, the relevant propagating hadronic system is totally dominated by the threshold pion states. In the spirit of “dual-string-QCD” approach, which may not be completely inappropriate[@Ros; @Love] we may then represent the soft pions as two nearby $u\bar{d}$ lines so that their mass proportional to the string bit length between them approaches zero and they are indeed point-like as appropriate for the chiral Lagrangian approach. Naively, we would then expect the diagrams corresponding to Figs. 1(b) and 1(c) to be essentially equal up to a factor (-1). This would be the case if in the VDM spirit we can describe the process merely via a $\rho$ exchange. In this case, $a^{I=2} = -\frac{1}{2}\,a^{I=1}$. To get a better ratio we need some $I=0$ scalar $t$ channel exchange which enhances $a_{\pi^+\pi^-}$ and suppresses $a_{\pi^+\pi^+}$. The magnitude of this term is fixed by requiring that it will have the same magnitude also in the crosses, s, channel, thereby obtaining the desired -2/7 ratio of the $I=2$ and $I=0$ scattering length. [*(ii)*]{} We can estimate the contribution of the multi-gluon exchanges to threshold $\pi\pi$ physics by the positive shift of $a^{I=0}$ and $a^{I=2}$ as compared with the original chiral Lagrangian estimate. It is tempting to identify the term, $$4\pi a^2_{gluon\,exchange}$$ with the asymptotic $\pi^+\pi^+$ cross section (controlled also by multi-gluon exchange) extrapolates all the way to threshold. We will pursue this in a future publication. The author would like to acknowledge the hospitality of the Department of Physics of the University of Maryland. [99]{} D. Weingarten, Phys. Rev. Lett. [**51**]{}, 1830 (1983). C. Vafa and E. Witten, Nucl. Phys. [**B234**]{}, 173 (1984). S. Nussinov and M. A. Lampert, “QCD Inequalities,” hep-ph/9911532. D. Espriu, M. Gross and J. F. Wehater, Phys. Lett. [**B146**]{}, 67 (1984). S. Nussinov and B. Sathiapalan, Nucl. Phys. [**B26**]{}, 285 (1985. S. Nussinov and M. Spiegelglas, Phys. Lett. [**B202**]{}, 265 (1988). R. Gupta, A. Patel and S. Sharpe, Phys. Rev. D [**48**]{}, 388 (1993). S. Adler and R. Dashen, [*Current algebra*]{} (Benjamin, New York, 1968). J. L. Rosner, Phys. Rev. Lett. [**21**]{}, 950 (1968). C. Lovelace, Phys. Lett. [**B28**]{}, 264 (1968).
{ "pile_set_name": "ArXiv" }
--- abstract: 'A quasiconformal tree is a doubling metric tree in which the diameter of each arc is bounded above by a fixed multiple of the distance between its endpoints. We study the geometry of these trees in two directions. First, we construct a catalog of metric trees in a purely combinatorial way, and show that every quasiconformal tree is bi-Lipschitz equivalent to one of the trees in our catalog. This is inspired by results of Herron-Meyer and Rohde for quasi-arcs. Second, we show that a quasiconformal tree bi-Lipschitz embeds in a Euclidean space if and only if its set of leaves admits such an embedding. In particular, all quasi-arcs bi-Lipschitz embed into some Euclidean space.' address: - | Department of Mathematical Sciences\ Ball State University\ Muncie, IN 47306 - | Department of Mathematics\ The University of Tennessee\ Knoxville, TN 37966 author: - 'Guy C. David' - Vyron Vellis bibliography: - 'quasitrees-ref.bib' title: 'Bi-Lipschitz geometry of quasiconformal trees' --- [^1] Introduction {#sec:intro} ============ In this paper, a *(metric) tree* is a compact, connected, locally connected metric space with the property that each pair of distinct points form the endpoints of a unique arc. In some sense, trees make up the simplest class of one-dimensional continua, and are ubiquitous in analysis and geometry. Within the class of all trees, an important role has been played by the class of *quasiconformal trees* studied in [@BM; @BM2; @Kinneberg]. By definition, these are trees $T$ that satisfy two simple geometric properties: - $T$ is *doubling*: there is a constant $N$ such that each ball in $T$ can be covered by $N$ balls of half the radius. - $T$ is *bounded turning*: there is a constant $C$ such that each pair of points $x,y \in T$ can be joined by a continuum whose diameter is at most $Cd(x,y)$. These conditions are both invariant under quasisymmetric mappings, making the class of quasiconformal trees a natural quasisymmetrically invariant class. We do not recall the definition of quasisymmetric mappings here (see [@Heinonen] or [@BM]), but merely note that they are an important generalization of conformal mappings to arbitrary metric spaces. Quasiconformal trees generalize two more well-known types of spaces. For one, quasiconformal trees that are simply topological arcs (i.e., have no branching) are called *quasi-arcs*, and have been studied in complex analysis and analysis on metric spaces for decades [@GH]. For example, the famous von Koch snowflake is a quasi-arc. A well-known result of Tukia and Väisälä [@TuVa] shows that quasi-arcs are exactly those spaces that are quasisymmetrically equivalent to the unit interval $[0,1]$. Quasiconformal trees also generalize (doubling) *geodesic trees*. Geodesic trees are trees in which, for each pair of points $x,y$, the unique arc joining them has (finite) length equal to $d(x,y)$. Thus, in a geodesic trees, all paths are “straight” (isometric to intervals in the real line), whereas paths in quasiconformal trees may be fractal, like the von Koch snowflake. Geodesic trees are of course standard objects of study in many parts of mathematics and computer science. Recently, Bonk and Meyer [@BM] generalized the result of Tukia and Väisälä mentioned above by showing that each quasiconformal tree is quasisymmetric to a geodesic tree. Rather than studying the quasisymmetric geometry of quasiconformal trees, this paper is concerned with the finer notion of bi-Lipschitz geometry. Recall that a mapping $f$ between two metric spaces is called *bi-Lipschitz* (or $L$-bi-Lipschitz to emphasize the constant) if there is a constant $L\geq 1$ such that $$L^{-1} d(x,y) \leq d(f(x),f(y)) \leq L d(x,y), \quad \text{ for all } x,y\in X.$$ Thus, bi-Lipschitz mappings preserve distances up to constant factors. All bi-Lipschitz mappings are quasisymmetric, but the converse is false. For example, one may parametrize the von Koch snowflake $K$ by a quasisymmetric map $[0,1]\rightarrow K$, but not by a bi-Lipschitz map. Given a metric space $X$, natural questions in the bi-Lipschitz world are: - Which metric spaces are *bi-Lipschitz equivalent* to $X$, i.e., admit a surjective bi-Lipschitz mapping onto $X$? - Does $X$ admit a *bi-Lipschitz embedding* into some Euclidean space $\R^n$, i.e., a bi-Lipschitz mapping from $X$ into $\R^n$? The first of these questions is about recognizing or providing models for spaces up to bi-Lipschitz equivalence, i.e., up to bounded distortion of their metrics. The second is about understanding which spaces can be viewed as subsets of Euclidean space up to bounded distortion, and in complete generality is a major problem in analysis, geometry, and computer science [@Heinonen_embed; @Ostrovskii]. We study both these questions for quasiconformal trees. Concerning the first, we give a “combinatorial model” for generating quasiconformal trees based on a purely discrete construction, and then show that every quasiconformal tree is bi-Lipschitz equivalent to one of our combinatorial constructions. This is in the vein of the combinatorial models for quasi-arcs up to bi-Lipschitz equivalence given by Rohde [@Rohde] and Herron-Meyer [@HM], although the construction for trees is more elaborate. Concerning the second question, we build on ideas from [@RV] to show that every **quasi-arc** admits a bi-Lipschitz embedding into some Euclidean space, and use this to show that the bi-Lipschitz embedding properties of quasiconformal trees are completely controlled by their sets of *leaves*. We leave open the main question of whether all quasi-trees admit bi-Lipschitz embeddings into Euclidean space; see below for additional background and discussion. We now discuss these ideas in more detail. Combinatorial models for quasiconformal trees up to bi-Lipschitz equivalence {#sec:introcomb} ---------------------------------------------------------------------------- We first give a way to define metric spaces using certain sequences of combinatorial graphs, that is, $G=(V,E)$ where $V$ is the vertex set and $E$ is the edge set. This is inspired by the ideas of [@Rohde; @HM] concerning quasi-arcs, with a number of new wrinkles in the case of trees. To simplify the presentation as much as possible, a number of definitions are postponed until Section \[sec:prelim\]. Let $A$ be an “alphabet”: a set of the form $\{1,\dots,n\}$ or $A=\N$. Denote by $\varepsilon$ the empty word and by $|w|$ the length of a word, i.e., the number of letters. Let $A^0 = \{\varepsilon\}$, and for each $k\in\N$ denote by $A^k$ the set of all words made from the alphabet $A$ of length exactly $k$. Define the set of finite words $$A^* = \bigcup_{k=0}^{\infty}A^k.$$ Denote also by $A^{\N}$ the set of infinite words formed by the alphabet $A$, and $A^\N_u\subseteq A^\N$ the set of all infinite words that begin with a given finite word $u\in A^*$. \[def:combdata\] We consider the following **combinatorial data** $\mathscr{C} = (A,(G_k)_{k\in\N})$ where: 1. \[eq:combdata1\] $A$ is a finite or infinite alphabet: $A=\{1,\dots, M\}$ for some integer $M\geq 2$, or $A=\mathbb{N}$. 2. \[eq:combdata2\] For each $k\in\mathbb{N}$, $G_k=(A^k, E_k)$ is a connected combinatorial graph on the vertex set $A^k$ with the following properties: 1. \[eq:combdata2a\] For each $w\in A^k$ the subgraph of $G_{k+1}$ induced by the vertex set $\{w1,\dots,wM\}$ is connected. 2. \[eq:combdata2b\] If $\{w,u\}\in E_k$, then there is a pair $(i,j)\in A\times A$ such that $\{wi ,uj\}\in E_{k+1}$. We next define a way to “move between” different infinite word sets $A^\N_u$ using the structure of the combinatorial data. Moves between $A^\N_u$ and $A^\N_v$ are always permitted if $u$ and $v$ are adjacent words of equal length, but in general we take into account the full scope of the combinatorial data. Thus, given combinatorial data $\mathscr{C} = (A,(G_k)_{k\in\N})$, we say that two infinite word sets $A^\N_{u_1}$ and $A^\N_{u_2}$ *combinatorially intersect*, and write $A^\N_{u_1} \wedge_{\mathscr{C}} A^\N_{u_2} \neq \emptyset$, if the following holds: $$\begin{aligned} &\text{For each } n > \max\{|u_1|,|u_2|\}, \text{ there exist words } w_1, w_2\in A^n\text{, beginning with } u_1 \text{ and } u_2\text{,}\label{eq:combintersection}\\&\text{respectively, that are adjacent in } G_n.\nonumber\end{aligned}$$ In other words, two word sets $A^{\N}_{u_1}$ and $A^{\N}_{u_2}$ combinatorially intersect if their restrictions to every sufficiently large finite level are adjacent. Below, in Definition \[def:combintersection\], we will give a precise definition of the set $A^\N_{u_1} \wedge_{\mathscr{C}} A^\N_{u_2}$ and show that its non-emptiness is equivalent to . Given this notion of combinatorial intersection, we can describe how to move between two infinite words: \[def:chain\] Given two words $w,w' \in A^{\N}$ we say that $\{A^{\N}_{w_1},\dots, A^{\N}_{w_n}\}$ is a *chain joining $w$ with $w'$* if $w\in A^{\N}_{w_1}$, $w'\in A^{\N}_{w_n}$ and for every $i=1,\dots,n-1$, we have $A^{\N}_{w_i}\wedge_{\mathscr{C}}A^{\N}_{w_{i+1}} \neq \emptyset$. Now that we have a way to move between two infinite words, we can define a distance on $A^\N$ by assigning costs to each chain with a “diameter function”: \[def:diam\] Given an alphabet $A$, a *diameter function* is a function $\D:A^* \to [0,1]$ that satisfies: 1. \[eq:diam1\] $\D(\varepsilon)=1$; 2. \[eq:diam3\] for each $w\in A^k$ and $i\in A$, $\Delta(wi)=0$ for all but finitely many $i\in A$; 3. \[eq:diam4\] $\lim_{n\to \infty}\max\{\D(w): w\in A^n\} = 0.$ The class of all diameter functions on $A$ is defined by $\mathscr{D}(A)$. Given $0< \delta_1 \leq \delta_2 \leq 1$ and finite $A$, we denote by $\mathscr{D}(A,\d_1, \d_2)$ the collection of all diameter functions on the alphabet $A$ such that, $$\text{for each $w\in A^*$ and $i,j \in A$}, \qquad \D(wi)=\D(wj)\quad \text{and}\quad\frac{\Delta(wi)}{\Delta(w)}\in\{\delta_1, \delta_2\}.$$ Note that, in Definition \[def:diam\], is automatic if $A$ is finite, and is automatic if $\D \in \mathscr{D}(A,\d_1,\d_2)$ and $\delta_2<1$. In , condition implies that the maximum is actually achieved, even if $A$ is infinite. Given combinatorial data $\mathscr{C} = (A,(G_k)_{k\in\N})$ and $\D\in \mathscr{D}(A)$, we define a pseudometric $D_{\mathscr{C},\D}$ on $A^\mathbb{N}$ by: $$\label{eq:pseudometric} D_{\mathscr{C},\D}(w, u) = \inf \sum_{i=0}^N \Delta(v_i)$$ where the infimum is taken over all chains $\{A^\N_{v_i}\}$ joining $w$ with $u$. We prove in Lemma \[lem:pseudometric\] that $D_{\mathscr{C},\D}$ is indeed always a pseudometric on $A^\N$. Taking the quotient space $\mathcal{A} :=A^\N/\sim$ under the equivalence relation that identifies points with zero $D_{\mathscr{C},\D}$-distance, we obtain a metric space $$(\mathcal{A}, d_{\mathscr{C},\D}),$$ where $d_{\mathscr{C},\D}([w],[v]) = D_{\mathscr{C},\D}(w,v)$. Our main theorem on these combinatorial models is as follows: \[thm:maincombthm\] 1. \[eq:maincomb1\] If $\mathscr{C}$ defines combinatorial data and $\Delta\in \mathscr{D}(A)$, then the space $(\mathcal{A}, d_{\mathscr{C},\D})$ is compact, connected, and bounded turning with constant $C$=1. 2. \[eq:maincomb2\] If in addition each graph $G_k$ in the combinatorial data is a combinatorial tree, then the space $(\mathcal{A}, d_{\mathscr{C},\D})$ is a metric tree. 3. \[eq:maincomb3\] Conversely, if $X$ is an arbitrary quasiconformal tree, then there exist combinatorial data $\mathscr{C}$ and a diameter function $\D\in\mathscr{D}(A,K_1, K_2)$ such that $X$ is bi-Lipschitz equivalent to the space $(\mathcal{A}, d_{\mathscr{C},\D})$. The choice of alphabet, the constants $K_1$ and $K_2$, and the bi-Lipschitz constant depend only on the doubling and bounded turning constants of $X$, and ${\mathop\mathrm{diam}\nolimits}(X)$. Parts and of Theorem \[thm:maincombthm\] are proven in Proposition \[prop:comb-tree\], and part is proven (with a more detailed statement) in Theorem \[thm:main\]. The metric space $(\mathcal{A}, d_{\mathscr{C},\D})$ constructed from given combinatorial data and diameter function need not be doubling in general, even if the alphabet $A$ is finite, the graphs $G_k$ are all combinatorial trees, and the diameter function $\D$ lies in $\mathscr{D}(A,\delta_1,\delta_2)$ for $0<\delta_1<\delta_2<1$. However, in Proposition \[prop:doubling\] we give some sufficient conditions that imply that the space $(\mathcal{A}, d_{\mathscr{C},\D})$ is doubling. In Theorem \[thm:maincombthm\], the space $(\mathcal{A}, d_{\mathscr{C},\D})$ that we construct always satisfies these conditions. This is stated explicitly in Theorem \[thm:main\]. Bi-Lipschitz embeddings of quasi-arcs and quasiconformal trees {#sec:introembed} -------------------------------------------------------------- We now turn our attention to the problem of finding bi-Lipschitz embeddings of quasiconformal trees into Euclidean space. The most natural question is: \[q:embed\] Does every quasiconformal tree have a bi-Lipschitz embedding into some Euclidean space $\R^n$? We do not answer this question here. However, recalling our discussion that quasiconformal trees generalize both quasi-arcs and geodesic trees, some supporting evidence for a positive answer is provided by a theorem of Gupta-Krauthgammer-Lee [@GKL], which states that every doubling, *geodesic* tree admits such an embedding. (See also the alternative proof in [@LNP Theorem 2.12].) By adapting techniques of Romney and the second named author, we make progress in the case where the tree has no branching: \[prop:introqcircles\] Every quasi-arc admits a bi-Lipschitz embedding into some Euclidean space $\R^n$. Proposition \[prop:introqcircles\] is a simplified version of Proposition \[prop:qcircles\] below, where we identify the sharp dimension $n$ for the embedding. We note that Herron and Meyer proved Proposition \[prop:introqcircles\] in the special case of quasi-arcs with Assouad dimension less than 2; see [@HM Theorem C]. Using Proposition \[prop:qcircles\], and results of Lang-Plaut [@LP] and Seo [@Seo], we give a criteria that can answer Question \[q:embed\] in certain examples. If $X$ is a metric tree, we denote by $\LL(X)$ be the set of *leaves* of $X$, i.e., $$\LL(X) := \{x\in X: X\setminus \{x\} \text{ is connected}\}.$$ \[thm:mainembed\] A quasiconformal tree $X$ admits a bi-Lipschitz embedding into some Euclidean space if and only if $\LL(X)$ admits a bi-Lipschitz embedding into some Euclidean space. Theorem \[thm:mainembed\] is a simplified version of the quantitative statement of Theorem \[thm:mainembed2\]. An equivalent reformulation of Theorem \[thm:mainembed\] is that a subset $E$ of a quasiconformal tree $X$ admits a bi-Lipschitz embedding into some Euclidean space if and only if the minimal sub-tree of $X$ containing $E$ does. Outline of the paper -------------------- In Section \[sec:prelim\], we review some elementary notions from graph theory concerning combinatorial graphs and trees. In Section \[sec:combtrees\], we provide more details on our combinatorial models and prove parts and of Theorem \[thm:maincombthm\]. In Section \[sec:doubling\], we work in the case of combinatorial trees and identify conditions on $A$, $\mathscr{C}$, and $\D$ that guarantee that the metric tree $(\mathcal{A}, d_{\mathscr{C},\D})$ is doubling. In Section \[sec:characterization\] we prove a more detailed version of part of Theorem \[thm:maincombthm\]. The basic idea is to construct an $n$-adic decomposition $(X_w)_{w\in \{1,\dots,n\}^*}$ of a given quasiconformal tree $X$, for some $n\geq 2$ that satisfies the following properties: 1. Each $X_w$ is the union of its children $X_{w1},\dots,X_{wn}$, which are themselves trees. Each two of the children intersect in at most one point, which has valency 2 in $X$. 2. Each child $X_{wi}$ of $X_w$ has diameter comparable to that of $X_w$. 3. Any two points $x,y$ on $X_w\cap \overline{X\setminus X_w}$ have distance comparable to the diameter of $X_w$. This is accomplished by performing certain subdivisions and gluings on top of a construction of Bonk and Meyer [@BM]. Once we have such a decomposition, we can build combinatorial data $\mathscr{C}$ and a diameter function $\D$ such that $(\mathcal{A}, d_{\mathscr{C},\D})$ is bi-Lipschitz equivalent to $X$. Finally, in Section \[sec:embed\], we prove a quantitative version of Proposition \[prop:introqcircles\] and then apply a bi-Lipschitz welding result of Lang and Plaut [@LP] and a bi-Lipschitz embedding characterization of Seo [@Seo] to complete the proof of Theorem \[thm:mainembed\]. Preliminaries {#sec:prelim} ============= In this section, we introduce some further preliminary definitions and results related to the combinatorial models defined in Section \[sec:introcomb\]. Words ----- Recall from Section \[sec:introcomb\] that we start with an alphabet $A = \{1, \dots, M\}$, for some integer $M\geq 2$, or $A=\N$. In addition to the sets $A^*$, $A^\N$, $A^\N_u$ defined above, we also set a few other pieces of notation. For $w\in A^*$ and $k\geq |w|$ define $$A^k_w = \{wu : u\in A^{k-|w|}\}, \quad A^*_w = \{wu : u\in A^{*}\}, \quad A^{\N}_w = \{wu : u\in A^{\N}\}.$$ Given $n\in\N$ and $w\in A^{\N}$ denote by $w(n)$ the unique word $u\in A^n$ such that $w=uw'$ for some $w'\in A^{\N}$. Similarly, if $n\in\mathbb{N}$ and $w\in A^{*}$, $w(n)$ denotes the initial subword of $w$ with length $n$, and we set $w(n)=w$ if $n\geq |w|$. Finally, given $k\in\N$ and $u \in A^k$ denote by $u^{\uparrow}$ the unique element of $A^{k-1}$ such that $u \in A^k_{u^{\uparrow}}$. Combinatorial graphs and trees ------------------------------ Definition \[def:combdata\] above uses some graph theory terminology. A *combinatorial graph* is a pair $G=(V,E)$ of a finite or countable vertex set $V$ and an edge set $$E \subset \{\{v,v'\} : v,v' \in V\text{ and }v\neq v'\}.$$ If $\{v,v'\}\in E$, we say that the vertices $v$ and $v'$ are *adjacent* in $G$. A combinatorial graph $G' = (V',E')$ is a *subgraph* of $G=(V,E)$ (and we write $G\subset G'$) if $V'\subset V$ and $E'\subset E$. We commonly generate subgraphs of $G=(V,E)$ by starting with a vertex set $V'\subset V$ and considering the *subgraph of $G$ induced by $V'$*: the graph $G'=(V',E')$ where $E'$ is the set of all edges between two vertices of $V'$. A *path* in $G$ is a set $ \g = \{\{v_1,v_2\}, \dots, \{v_{n-1},v_n\}\} \subset E$; in this case we say that $\g$ joins $v_1$, $v_n$. The path $ \g = \{\{v_1,v_2\}, \dots, \{v_{n-1},v_n\}\}$ is a *combinatorial arc* or *simple path* if for all $i, j \in \{1,\dots,n\}$, $v_i = v_j$ if and only if $i=j$; in this case we say that the endpoints of the arc $\g$ are the points $v_1,v_n$. A combinatorial graph $G = (V,E)$ is connected, if for any distinct $v,v' \in V$ there exists a path $\g$ in $G$ that joins $v$ with $v'$. A *component* of a combinatorial graph $G$ is a maximal connected subgraph of $G$. A graph $T = (V,E)$ is a *combinatorial tree* if for any distinct $v,v'$ there exists unique combinatorial arc $\g$ whose endpoints are $v$ and $v'$. Given a combinatorial tree $T = (V,E)$ and a point $v\in V$, define the valencies $$\text{Val}(T,v) := {\operatorname{card}}\{e \in E : v\in e\} \qquad\text{and}\qquad \text{Val}(T) := \max_{v\in V} \text{Val}(T,v)$$ and the set of leaves $\text{Leaves}(T) := \{v\in V : \text{Val}(T,v) = 1\}$. Here ${\operatorname{card}}$ denotes the cardinality of a finite or countable set, taking values in $\mathbb{N} \cup \{\infty\}$. Given a combinatorial graph $G=(V,E)$ and a vertex $v\in V$, we write $G\setminus\{v\}$ to be the subgraph of $G$ induced by $V\setminus\{v\}$. Note that, if $T$ is a tree, then every component of $T\setminus\{v\}$ is a tree. A model for bounded turning metric spaces and trees {#sec:combtrees} =================================================== Combinatorial data ------------------ Recall the notion of combinatorial data $\mathscr{C} = (A, (G_k)_{k\in\N})$ from Definition \[def:combdata\], where $A$ is an alphabet and $G_k=(A^k,E_k)$ are combinatorial graphs on the vertex sets $A^k$, satisfying certain axioms. **For the remainder of Section \[sec:combtrees\], we fix combinatorial data $\mathscr{C}= (A, (G_k)_{k\in\N})$.** Our first lemma gives some basic structural properties of these graphs. In particular, if each $G_k$ is a combinatorial tree, then the pair $(i,j) \in A\times A$ of Definition \[def:combdata\] is unique. \[lem:adjacent\] Let $k\geq j$ and $v\neq w\in A^j$. 1. If $v$ and $w$ are adjacent in $G_j$, then there are words $v'$ and $w'$ in $A^{k-j}$ such that $vv'$ and $ww'$ are adjacent in $G_k$. 2. If $G_k$ is a combinatorial tree and there are words $v'$ and $w'$ in $A^{k-j}$ such that $vv'$ and $ww'$ are adjacent in $G_k$, then $v$ and $w$ are adjacent in $G_j$. 3. If $G_k$ is a combinatorial tree and $v$ and $w$ are adjacent in $G_j$, then there is a unique pair of words $(v', w')$ in $A^{k-j} \times A^{k-j}$ such that $vv'$ and $ww'$ are adjacent in $G_k$. The first statement is an immediate consequence of in Definition \[def:combdata\], and induction on $k-j$. For the second, suppose that $v$ and $w$ were not adjacent in $G_j$, under these assumptions. Let $v=u_0, u_1, \dots, u_{n-1}, u_n=w$ be a path from $v$ to $w$ in $G_j$. Note that $n\geq 2$. Then, by the first statement in the lemma and part of Definition \[def:combdata\], there is a simple path from $vv'\in A^k_v$ to $ww'\in A^k_w$ in $G_k$ of the form $$\text{ elements of } A^k_{u_0}, \text{ elements of } A^k_{u_1}, \dots, \text{ elements of } A^k_{u_n}.$$ On the other hand, there is also an adjacency between $vv'$ and $ww'$ in $G_k$. This contradicts the assumption that $G_k$ is a tree. For the third claim, the existence of $v'$ and $w'$ follows from (1). Suppose that the uniqueness failed. We consider the following two possible cases. Suppose first that there are two distinct $v',v'' \in A$ and $w'\in A$ such that both $vv'$ and $vv''$ are adjacent to $ww'$. Then there exists two combinatorial arcs in $G_{k}$ that join $vv'$ with $vv''$; one through the vertices of $G_{k}$ restricted on $A^{k}_v$ (by (2a) in Definition \[def:combdata\]) and another is $\{\{vv',ww'\},\{ww', vv''\}\}$. This contradicts the fact that $G_{k}$ is a tree. The other possibility is that there are two distinct $v',v'' \in A$ and two distinct $w',w''\in A$ such that $vv'$ is adjacent to $ww'$, and $vv''$ are adjacent to $ww''$. Then there exist two combinatorial arcs in $G_{k}$ that join $vv'$ with $vv''$; one through the vertices of $G_{k}$ restricted on $A^{k}_v$ (by (2a) in Definition \[def:combdata\]) and another through the vertices of $G_{k}$ restricted on $A^{k}_w$ along with edges $\{vv',ww''\}$ and $\{vv'',ww''\}$. This again contradicts the fact that $G_{k}$ is a tree. Combinatorial intersection and chains ------------------------------------- Recall the notion of combinatorial intersection $A^\N_u \wedge_\mathscr{C} A^\N_v$ defined in in Section \[sec:introcomb\]. There, we only defined what it means for this set to be non-empty, but here we actually give a meaning to the set itself. \[def:combintersection\] Given $u_1,u_2 \in A^*$, define $$\begin{aligned} A^{\N}_{u_1} \wedge_{\mathscr{C}} A^{\N}_{u_2} := &\{w \in A^{\N}_{u_1} : \forall n > \max\{|u_1|,|u_2|\} \text{ there exists $u\in A^n_{u_2}$ with $\{w(n),u\}\in E_n\}$}\\ \cup &\{w \in A^{\N}_{u_2} : \forall n >\max\{|u_1|,|u_2|\} \text{ there exists $u\in A^n_{u_1}$ with $\{w(n),u\}\in E_n\}$}.\nonumber\end{aligned}$$ The set $A^{\N}_{u_1} \wedge_{\mathscr{C}} A^{\N}_{u_2}$ is called the *combinatorial intersection* of $A^{\N}_{u_1}$ and $A^{\N}_{u_2}$. We now show that this definition agrees with that in , and give an equivalent reformulation in the case of trees. \[lem:intersection1\] Let $u_1,u_2 \in A^*$. The following are equivalent. 1. The set $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$ is non-empty. 2. For every $k>\max\{|u_1|,|u_2|\}$ there exists $v_1\in A^k_{u_1}$ and $v_2 \in A^k_{u_2}$ such that $\{v_1,v_2\} \in E_k$. If each graph $G_k$ is a combinatorial tree, then (1) and (2) are also equivalent to the following. 1. There exists $k> \max\{|u_1|,|u_2|\}$ and $v_1\in A^k_{u_1}$, $v \in A^k_{u_2}$ such that $\{v_1,v_2\} \in E_k$. We start by showing the equivalence of (1) and (2). That (1) implies (2) follows immediately from the definition of $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$. To show that (2) implies (1), let $k_0=\max\{|u_1|,|u_2|\}$ and choose $u_1i_1 \in A^{k_0+1}_{u_1},u_2 j_1 \in A^{k_0+1}_{u_2}$ such that $\{u_1i_1,u_2j_1\} \in E_{k_0+1}$. By (2b) in Definition \[def:combdata\], given that $\{u_1i_1\cdots i_{n-k}, u_2j_1\cdots j_{n-k}\} \in E_n$ for some $n\geq k+1$, there exist $i_{n-k+1},j_{n-k+1}\in A$ such that $\{u_1i_1\cdots i_{n-k+1}, u_2j_1\cdots j_{n-k+1}\} \in E_{n+1}$. Set now $$w_1 = u_1i_1i_2\cdots \quad\text{and}\quad w_2=u_2j_1j_2\cdots$$ and note that both $w_1$ and $w_2$ are in $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$. Assume now that each graph $G_k$ is a combinatorial tree. Clearly (2) implies (3) so it suffices to show that (3) implies (2). Assume there is an integer $k_0\geq \max\{|u_1|,|u_2|\}$ and words $w_1 \in A_{u_1}^{k_0}$ and $w_2\in A^{k_0}_{u_2}$ such that $\{w_1,w_2\} \in E_{k_0}$. If $k\geq k_0$, then by Lemma \[lem:adjacent\](1), there exist $v_1 \in A^{k}_{w_1}$ and $v_2\in A^{k}_{w_2}$ (hence $v_1 \in A^{k}_{u_1}$ and $v_2\in A^k_{u_2}$) such that $\{v_1,v_2\}\in E_k$. If $k$ is an integer with $ \max\{|u_1|,|u_2|\} \leq k \leq k_0$, then by Lemma \[lem:adjacent\](2), there exist $v_1 \in A^{k}_{u_1}$ and $v_2\in A^{k}_{u_2}$ such that $w_1\in A^k_{v_1}$, $w_2\in A^k_{v_2}$ and $\{v_1,v_2\}\in E_k$. Therefore, (2) holds. The next lemma gives a description of the set $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$ in the case that each $G_k$ is a combinatorial tree. \[lem:intersection2\] Let $u_1,u_2 \in A^*$ with $|u_1|\leq |u_2|$, let $k_1=|u_1|$ and let $u'_2 = u_2(k_1)$. 1. If $u_2'=u_1$ (that is, $u_2 \in A^{*}_{u_1}$), then $A^{\N}_{u_2} \subset A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$. Suppose additionally that each $G_k$ is a combinatorial tree. Then: 1. If $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2} \neq \emptyset$, then either $\{u_1,u_2'\} \in E_{k_1}$ or $u_1=u_2'$. 2. If $\{u_1,u_2'\} \in E_{k_1}$, then $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$ contains exactly two elements; one in $A^{\N}_{u_1}$ and one in $A^{\N}_{u_2}$. The converse is also true. Let $u_1,u_2,v \in A^*$ and $k_1\in\N$ be as in the statement and let $k_2=|u_2|$. To prove (1), assume that $u_2'=u_1$, that is, $u_2 \in A^{k_2}_{u_1}$. Let $w\in A^{\N}_{u_2}$. By Definition \[def:combdata\](2a), the subgraph of $G_{k_2+1}$ induced by $A^{k_2+1}_{u_1}$ is connected. Fix $v\in A^{k_2+1}_{u_1}$ adjacent to $w(k_2+1)$. Applying Definition \[def:combdata\](2b) we find a sequence $\{i_1,i_2\dots\} \subset A$ such that for each $n\in\N$, $vi_1\cdots i_n$ is adjacent to $w(k_2+n+1)$. Since $vi_1\cdots i_n \in A^{k_2+n+1}_{u_1}$, by definition, $w \in A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$. Assume now for the rest of the proof that each $G_k$ is a combinatorial tree. To prove (2), assume that $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2} \neq \emptyset$. By Lemma \[lem:intersection1\](2), we have that there exists $v_1\in A^{k_2+1}_{u_1}$ and $v_2 \in A^{k_2+1}_{u_2}$ such that $\{v_1,v_2\} \in E_{k_2+1}$. Applying Lemma \[lem:adjacent\](2), we have that either $u_1=u_2'$ or $\{u_1,u_2'\} \in E_{k_1}$. To prove (3), assume that $\{u_1,u_2'\} \in E_{k_1}$ and let $v_1$ and $v_2$ be as in the proof of (2). That is, $v_1 \in A^{k_2+1}_{u_1}$, $v_2\in A^{k_2+1}_{u_2}$, and $\{v_1,v_2\}\in E_{k_2+1}$. By Definition (2b) of \[def:combdata\], there exist $i_1,i_2,\dots \in A$ and $j_1,j_2,\dots,\in A$ such that for all $m\in\N$, $\{v_1i_1\cdots i_m, v_2j_1\cdots j_m\} \in E_{k_2+1+m}$. It follows that the words $w_1= v_1i_1i_2\cdots \in A^{\N}_{u_1}$ and $w_2= v_2j_1j_2\cdots \in A^{\N}_{u_2}$ are in $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$. Suppose now that there exist two distinct $w_1',w_1 \in A^{\N}_{u_1}$ such that $w_1',w_1 \in A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$. Let $l>k_2$ be an integer such that $w_1(l)\neq w_1'(l)$. By Lemma \[lem:intersection1\](2), there exist $v, v'\in A^\N_{u_2} \subseteq A^\N_{u'_2}$ such that $\{w_1(l),v\}$ and $\{w'_1(l),v'\}$ are in $E_l$. This contradicts the uniqueness statement of Lemma \[lem:adjacent\](3). Finally, for the converse of (3) simply note that if $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$ contains exactly two elements, then by (2), either $u_1=u_2'$, or $u_1$ is adjacent to $u_2'$. However, the former is false since in that case, by (1), $A^{\N}_{u_1}\wedge_{\mathscr{C}}A^{\N}_{u_2}$ would be an infinite set. We now study chains, as defined in Definition \[def:chain\] of Section \[sec:introcomb\]. The following lemma shows that, if each $G_k$ in the combinatorial data is a combinatorial tree, that chains must respect the “between-ness” relation in each $G_k$. \[lem:chain\] Suppose that each graph $G_k$ is a combinatorial tree. Let $w_1,w_2,w_3 \in A^k$, and let $w_2$ be on the unique combinatorial arc in $G_k$ that joins $w_1$ and $w_3$. If $u_1\in A^{\N}_{w_1}$ and $u_3 \in A^{\N}_{w_3}$, then for every chain $\{A^{\N}_{v_1},\dots, A^{\N}_{v_n}\}$ joining $u_1$ with $u_3$, there exists $v\in A^*$ and $v_i$ such that $A^{\N}_{w_2v}\subset A^{\N}_{v_i}$. We may assume that the three words $w_1, w_2, w_3$ are distinct, otherwise the lemma is trivial. As a start, we note that $u_1$ has an initial $w_1$ substring and an initial $v_1$ substring, so either $v_1$ is an initial substring of $w_1$ or vice versa. A similar consideration applies to $u_3$, $v_n$, and $w_3$. For each $i\in 1,\dots, n$, we define a subset $P_i \subseteq A_k = V(G_k)$ as follows: If $|v_i|<k$, then let $P_i = A^k_{v_i}$. If $|v_i|\geq k$, then let $P_i = \{v_i(k)\}$. In either case, $P_i$ induces a connected subgraph of $G_k$. $P_1$ contains $w_1$ and $P_n$ contains $w_3$. If $|v_1|<k$, then $v_1$ is an initial substring of $w_1$, and so $P_1 = A^k_{v_1} \ni w_1$. If $|v_1|\geq k$, then $w_1=v_1(k)\in P_1$. By the same argument, $P_n$ contains $w_3$. For each $i\in \{1,\dots, n-1\}$, either $P_i \cap P_{i+1}\neq \emptyset$ or there is an edge $\{a,b\}\in E_k$ with $a\in P_i$ and $b\in P_{i+1}$. Assume without loss of generality that $|v_i|\geq |v_{i+1}|$. Since $\{v_i\}$ is a chain, $A^\N_{v_i} \wedge_{\mathscr{C}} A^\N_{v_{i+1}}\neq \emptyset$. Case 1: If $|v_{i+1}| \leq |v_i| < k$, then $P_i=A^k_{v_i}$ and $P_{i+1}=A^k_{v_{i+1}}\supseteq A^k_{v_{i+1}v'}$. These contain adjacent elements by Lemma \[lem:intersection1\](2). Case 2: If $k \leq |v_{i+1}| \leq |v_i|$, then $P_i=\{v_i(k)\}$ and $P_{i+1}=\{v_{i+1}(k)\}$. By Lemma \[lem:intersection1\](3) and Lemma \[lem:adjacent\](2), the elements $v_i(k)$ and $v_{i+1}(k)$ are either equal or adjacent in $G_k$. Case 3: If $|v_{i+1}| \leq k \leq |v_i|$, then $P_i=\{v_i(k)\}$ and $P_{i+1}=A^k_{v_{i+1}}$. We have $|v_i|=|v_{i+1}v'|$ for some $v'\in A^*$. If $v_i(k)= (v_{i+1}v')(k)$, then $P_i \subseteq P_{i+1}$. Otherwise, Lemma \[lem:intersection1\](2) and Lemma \[lem:adjacent\](2) say that $v_i(k)$ is adjacent to $(v_{i+1}v')(k)$. Since $|v_{i+1}|<k$, the vertex $(v_{i+1}v')(k)\in A^k_{v_{i+1}} = P_{i+1}$. This completes the proof of the claim. Thus, the union of the sets $P_1, P_2, \dots, P_n$ induces a connected subgraph of $G_k$ that contains $w_1$ and $w_3$. It therefore must contain $w_2$, so $w_2\in P_i$ for some $i$. If $|v_i|< k$, then this means that $w_2\in P_i = A^k_{v_i}$. Thus $A^\N_{w_2}\subseteq A^\N_{v_i}$, which proves the lemma in this case. If $|v_i|\geq k$, then $w_2\in P_i = \{v_i(k)\}$. Thus, $w_2v=v_i$ for some word $v$, which proves the lemma in this case. Diameter functions and metrics {#sec:diameter} ------------------------------ Recall the notion of a diameter function $\D$ on an alphabet $A$ (and the class $\mathscr{D}(A)$ of all diameter functions on $A$) from Definition \[def:diam\]. **For the remainder of Section \[sec:combtrees\], we fix a diameter function $\D\in \mathscr{D}(A)$.** Given $\mathscr{C}$ and $\D$, we defined the distance $D_{\mathscr{C},\D}$ on $A^\N$ in by taking an infimum over chains. We first prove that $D_{\mathscr{C},\D}$ is indeed a pseudometric as claimed. \[lem:pseudometric\] The function $D_{\mathscr{C},\D}$ is a pseudometric on $A^\mathbb{N}$. First, notice that for any $w\in A^{\N}$ and any $n\in\N$, $\{A_{w(n)}^{\N}\}$ is a chain that joins $w$ with $w$. Thus, $$D_{\mathscr{C},\D}(w,w) \leq \D(w(n)) \leq \max_{v\in A^n}\D(v),$$ which vanishes as $n\to \infty$ by property in Definition \[def:diam\]. Hence, $D_{\mathscr{C},\D}(w,w)=0$. The symmetry of $D_{\mathscr{C},\D}$ is trivial, as any chain joining $w$ with $u$ is also a chain joining $u$ with $w$. For the triangle inequality, fix $\e>0$. Let $\{A_{w_1}^{\N},\dots,A_{w_n}^{\N}\}$ be a chain joining $w$ with $u$ and let $\{A_{u_1}^{\N},\dots,A_{u_m}^{\N}\}$ be a chain joining $u$ with $v$ such that $$\sum_{i=1}^n \D(w_i) < \frac{\e}{2} + D_{\mathscr{C},\D}(w,u) \qquad\text{and}\qquad \sum_{j=1}^n \D(u_j) < \frac{\e}{2} + D_{\mathscr{C},\D}(u,v).$$ By Lemma \[lem:intersection2\](1), we have that $A^{\N}_{w_n}\wedge_{\mathscr{C}}A^{\N}_{u_1} \neq \emptyset$, and so $\{A_{w_1}^{\N},\dots,A_{w_n}^{\N},A_{u_1}^{\N},\dots,A_{u_m}^{\N}\}$ is a chain joining $w$ with $v$. Thus, $D_{\mathscr{C},\D}(w,v) \leq D_{\mathscr{C},\D}(w,u) + D_{\mathscr{C},\D}(u,v) + \e$. As $\e$ was chosen arbitrarily, the lemma follows. We now describe more precisely the associated metric space $(\mathcal{A}, d_{\mathscr{C},\D})$ associated to a given combinatorial data $\mathscr{C}$ and diameter function $\D$ on $A$, introduced briefly in Section \[sec:introcomb\]. To turn $D_{\mathscr{C},\D}$ into a metric, we define a relation on $A^{\N}$. In particular, we write $w\sim u$ (for convenience we drop the dependence on $\mathscr{C},\D$) if and only if $D_{\mathscr{C},\D}(w,u)=0$. Since $D_{\mathscr{C},\D}$ is a pseudometric, it follows that $\sim$ is an equivalence relation. Using this identification, we define $$\mathcal{A}= A^{\N}/\sim \qquad\text{and}\qquad \mathcal{A}_w= A^{\N}_w/\sim \quad\text{for each $w\in A^*$}.$$ Based on $D_{\mathscr{C},\D}$, we define a function $d_{\mathscr{C},\D}$ on $\mathcal{A}\times\mathcal{A}$ in the usual way: if $[w],[u] \in \mathcal{A}$, then set $$d_{\mathscr{C},\D}([w],[u]) := D_{\mathscr{C},\D}(w,u).$$ The function $d_{\mathscr{C},\D}$ is well-defined. To see why this is true, let $w,w',u, \in A^{\N}$ such that $[w]=[w']$. By Lemma \[lem:pseudometric\] we have that $D_{\mathscr{C},\D}(w,u) \leq D_{\mathscr{C},\D}(w,w') + D_{\mathscr{C},\D}(w',u) = D_{\mathscr{C},\D}(w',u)$. Similarly, $D_{\mathscr{C},\D}(w',u) \leq D_{\mathscr{C},\D}(w,u)$ and, thus, $D_{\mathscr{C},\D}(w',u) = D_{\mathscr{C},\D}(w,u)$. \[lem:dist\] The function $d_{\mathscr{C},\D}$ is a metric on $\mathcal{A}$ and for each $w \in A^*$, ${\mathop\mathrm{diam}\nolimits}{\mathcal{A}_{w}} \leq \D(w)$. We first show that $d_{\mathscr{C},\D}$ is a metric. It is clear that $d_{\mathscr{C},\D}$ is non-negative, symmetric and $d_{\mathscr{C},\D}([w],[u]) = 0$ if and only if $[w]= [u]$ in $\mathcal{A}$. The triangle inequality follows from Lemma \[lem:pseudometric\]. Let $w\in A^*$ and $[u_1],[u_2]\in \mathcal{A}_w$. We may choose $u_1$ and $u_2$ in $A_w$. The set $\{A^{\N}_w\}$ is then a chain joining $u_1$ with $u_2$ and $d_{\mathscr{C},\D}([u_1], [u_2]) \leq \D(w)$. Therefore, ${\mathop\mathrm{diam}\nolimits}{\mathcal{A}_{w}^{\N}} \leq \D(w)$. We use standard metric space terminology when discussing $(\mathcal{A}, d_{\mathscr{C},\D})$. In particular, if $[w]\in\mathcal{A}$ and $r>0$, we write $B([w],r)$ for the open ball centered at $[w]$ of radius $r$ in this space. Bounded turning spaces {#sec:BTtrees} ---------------------- We now work towards the following proposition, which proves parts and of Theorem \[thm:maincombthm\]. \[prop:comb-tree\] The metric space $(\mathcal{A},d_{\mathscr{C},\D})$ is compact, path-connected, and $1$-bounded turning. Moreover, if each combinatorial graph $G_k$ is a combinatorial tree, then the metric space is a tree. (Here we are using the shorthand “$C$-bounded turning” for “bounded turning with constant $C$”.) The separate statements of Proposition \[prop:comb-tree\] are proven in Lemmas \[lem:compact\], \[lem:connected\], \[lem:bt\], and \[lem:tree\]. \[lem:cover\] Fix $w\in A^*$. Let $$I = \{i\in A: \Delta(wi)>0\}.$$ If ${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_w)>0$, then $$\mathcal{A}_w \subseteq \bigcup_{i\in I} \mathcal{A}_{wi}.$$ The assumption that ${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_w)>0$ implies that $I$ is non-empty. Let $k=|w|$. Consider any $[v]\in{\mathcal{A}}_w$; without loss of generality, $v(k)=w$. We will show that $[v]=[u]$ for some $u\in \cup_{i\in I} A^\N_{wi}$. If $v(k+1)\in \{wi:i\in I\}$, then we are done, so suppose it is not. Then there is a simple path $$u_1, u_2, \dots, u_n$$ in the combinatorial tree $G_{k+1}$ such that $u_1=v(k+1)$, $u_n=wi$ for some $i\in I$, and $u_j\notin \{wi:i\in I\}$ for $1\leq j \leq n-1$. By Lemma \[lem:adjacent\], there is a sequence $i_1, i_2, \dots$ in $A$ such that $u_n i_1 \dots i_m$ is adjacent to some element of $A_{u_{n-1}}^{k+m}$ for each $m$. Let $$u = u_n i_1 i_2 \dots \in A^\mathbb{N}.$$ Then, for each $m\geq 1$, the chain $$\{A^\mathbb{N}_{u_1}, \dots, A^\mathbb{N}_{u_{n-1}}, A^\mathbb{N}_{u(k+m)}\}$$ joins $v$ to $u \in {\mathcal{A}}_{wi}$. Note that $\D(u_j)=0$ for $1\leq j \leq n-1$. Therefore, $$D_{\mathscr{C},\D}(v, u) \leq \Delta(u(k+m)) \leq \max\{\Delta(r): r\in A^{k+m}\} \rightarrow 0 \text{ as } m\rightarrow \infty.$$ It follows that $[v]=[u]\in {\mathcal{A}}_{wi}$. This completes the proof. We can now prove a slightly stronger version of the first statement in Proposition \[prop:comb-tree\]. \[lem:compact\] For each $w\in A^*$, the metric space $(\mathcal{A}_w,d_{\mathscr{C},\D})$ is compact. In particular, taking $w=\varepsilon$ we see that $(\mathcal{A},d_{\mathscr{C},\D})$ is compact, as required in Proposition \[prop:comb-tree\]. We show that $(\mathcal{A}_w,d_{\mathscr{C},\D})$ is sequentially compact. Let $([w_n])$ be a sequence in $\mathcal{A}_w$. Suppose that this sequence has no convergent subsequence. This implies that ${\mathop\mathrm{diam}\nolimits}({\mathcal{A}}_w)>0$, otherwise $([w_n])$ would be constant. Let $$I_1 = \{i\in A: \Delta(wi)>0\}.$$ Note that $I_1$ is finite by Definition \[def:diam\]. Thus, by Lemma \[lem:cover\] there exists $i_1\in I_1$ and a subsequence $([w^1_n])$ of $([w_n])$ in $\mathcal{A}_{wi_1}$. We proceed by induction to construct sets $I_m\subseteq A$, indices $i_m\in I_m$, and subsequences $([w^m_n])$ of $([w_n])$ contained in ${\mathcal{A}}_{wi_1i_2\dots i_m}$. Assuming that there is a subsequence $([w_n^m]) \subseteq \mathcal{A}_{wi_1\cdots i_m}$, let $$I_{m+1} = \{ i\in A: \Delta(wi_1\cdots i_m i)>0\},$$ which is finite as above. As above, ${\mathop\mathrm{diam}\nolimits}({\mathcal{A}}_{wi_1\cdots i_m})>0$, otherwise $([w_n^m])$ would be constant, hence convergent. Thus, by Lemma \[lem:cover\], there is $i_{m+1}\in I_{m+1}\subseteq A$ and a subsequence $([w^{m+1}_n])$ of $([w_n^m])$ in $\mathcal{A}_{wi_1\cdots i_{m+1}}$. Set $u = wi_1i_2\cdots \in A^{\N}$ and consider the subsequence $([w^n_n])$ of $([w_n])$. Then, $d_{\mathscr{C},\D}([w_{n}^n],[u]) \leq \D(u(n)) \to 0$ as $n\to\infty$, contradicting our assumption. Thus, $(\mathcal{A}_w,d_{\mathscr{C},\D})$ is compact. We now work towards the connectedness properties. The following definition is convenient: An *$\e$-path* in a metric space $(X,d)$ is a finite sequence $(x_1, \dots, x_n)$ such that $d(x_i, x_{i+1})\leq\e$ for each $i\in\{1,\dots, n-1\}$. We say that the $\e$-path *joins* $a$ and $b$ if $a=x_1$ and $b=x_n$. \[lem:discreteBT\] Let $[w_1],[w_2]\in{\mathcal{A}}$ with $d_{\mathscr{C},\D}([w_1],[w_2])<r$, and let $\epsilon>0$. Then there is a $\epsilon$-path joining $[w_1]$ and $[w_2]$ of diameter less than $r$. Fix $[w_1]$, $[w_2]$, $r>0$, and $\e>0$ as in the statement of the lemma. Let $\{A^\mathbb{N}_{u_1},\dots,A^\mathbb{N}_{u_k}\}$ be a chain joining $w_1$ with $w_2$ such that $$\sum_{i=1}^{k}\D(u_i) \leq d_{\mathscr{C},\D}([w_1],[w_2]) + \frac{r - d_{\mathscr{C},\D}([w_1],[w_2])}{2} < r.$$ Note that for any $i,j\in \{1,\dots,k\}$ and any $w_i \in A^{\N}_{u_i}$ and $w_j\in A^{\N}_{u_j}$, we may use a subset of this same chain to join them, and so obtain $$\label{eq:diambound} d_{\mathscr{C},\D}([w_i],[w_j]) <r.$$ By property in Definition \[def:diam\], there exists $m\in\N$ such that $\D(u) \leq \e/2$ for all $u\in A^m$. By the properties of $G_m$, there exists a path $$\g = \{\{u_1',u_2'\}, \dots \{u_{n-1}',u_n'\}\} \subset \bigcup_{i=1}^k A^m_{u_i}$$ such that $w_1\in A^{\N}_{u_1'}$ and $w_2 \in A^{\N}_{u_n'}$. For each $i\in\{1,\dots,n\}$ let $v_i = u_i'1^{\infty}$ and let $v_0=w_1$ and $v_{n+1}=w_2$. Then for each $i=1,\dots,n-1$, $$d_{\mathscr{C},\D}([v_i],[v_{i+1}]) \leq \D(u_i') + \D(u_{i+1}') \leq \e/2 + \e/2 = \e$$ and similarly $d_{\mathscr{C},\D}([w],[v_{1}]) \leq \D(u_1') \leq \e$ and $d_{\mathscr{C},\D}([w_0],[v_{n}]) \leq \D(u_n') \leq \e$. Thus, $(v_0, v_1, \dots, v_{n+1})$ is an $\e$-path joining $w_1$ to $w_2$. Its diameter is less than $r$ by . The following lemma completes the proof of the topological properties in Proposition \[prop:comb-tree\]. \[lem:connected\] The metric space $(\mathcal{A},d_{\mathscr{C},\D})$ has the property that $\overline{B([w_0],r)}$ is connected for each $[w_0]\in{\mathcal{A}}$ and $r>0$. In particular, the space is connected, locally connected, and path-connected. The second sentence follows from the first: connectedness by taking $r=1\geq{\mathop\mathrm{diam}\nolimits}({\mathcal{A}})$, local connectedness by, e.g., [@Whyburn (15.1)], and path-connectedness by the Hahn-Mazurkiewicz Theorem and Lemma \[lem:compact\]. For the first sentence, fix $w_0 \in A^{\N}$ and $r>0$. To show that $\overline{B([w_0],r)}$ is connected, it suffices to show that for any $\epsilon>0$, each $[w]\in \overline{B([w_0],r)}$ can be joined to $[w_0]$ by an $\epsilon$-path contained in $ \overline{B([w_0],r)}$. The point $[w]$ is less than $\epsilon$-distance away from an element $[w']$ of $B([w_0],r)$. There is an $\epsilon$-path joining $[w_0]$ to $[w']$ inside $B([w_0],r)$, by Lemma \[lem:discreteBT\]. Since $d_{\mathscr{C},\D}([w'],[w])<\epsilon$, appending $[w]$ to this path yields an $\epsilon$-path joining $[w_0]$ to $[w]$ inside $\overline{B([w_0],r)}$. \[lem:bt\] The metric space $(\mathcal{A},d_{\mathscr{C},\D})$ is $1$-bounded turning. Let $[w_1], [w_2]\in {\mathcal{A}}$, with $r=d_{\mathscr{C},\D}([w_1],[w_2])>0$. Let $\epsilon>0$. By Lemma \[lem:discreteBT\], there is an $\epsilon$-path $(v_0, v_1, \dots, v_n)$ joining $[w_1]$ to $[w_2]$ with diameter at most $r+\epsilon$. Define a compact set $K_\epsilon \subseteq {\mathcal{A}}$ by $$K_\epsilon = \cup_{j=0}^n \overline{B([v_j],2\epsilon)}.$$ Note that each ball in this union is connected, by Lemma \[lem:connected\]. Since $\overline{B([v_j],2\epsilon)} \cap \overline{B([v_{j+1}],2\epsilon)}\neq \emptyset$ for each $j=0\dots n-1$, it follows that $K_\epsilon$ is also connected. Moreover, $$\label{eq:diamK} {\mathop\mathrm{diam}\nolimits}(K_\epsilon) \leq r+5\epsilon.$$ The sets $K_{1}, K_{1/2}, K_{1/3}$ are each compact, connected, and contain both $[w_1]$ and $[w_2]$. They therefore admit a subsequence that converges in the Hausdorff metric to a compact, connected set that contains $[w_1]$ and $[w_2]$. By , this set has diameter $r$. This completes the proof. Metric Trees ------------ We now prove the second half of Proposition \[prop:comb-tree\], namely, that if each combinatorial graph in our data is in fact a combinatorial tree, then the resulting metric space is a metric tree. **Thus, for the remainder of Section \[sec:combtrees\], we assume that each combinatorial graph $G_k$ is a metric tree, and we rename the graphs $T_k$ to reflect this.** \[lem:tree1\] Suppose that $w,w',w_0 \in A^k$ and $w_0$ is on the unique combinatorial arc in $T_k$ that joins $w$ with $w'$. If there exist $u\in A^{\N}_{w}$ and $u'\in A^{\N}_{w'}$ such that $[u]=[u']$, then $[u]\in \mathcal{A}_{w_0}$ Let $w,w',w_0$ be as in the statement of the lemma. We claim that for any $\e>0$ sufficiently small, there exists $v \in A^{\N}_{w_0}$ such that $D_{\mathscr{C},\D}(u,v) < \e$. Assuming this claim, by Lemma \[lem:compact\], it follows that there exists $u_0 \in \mathcal{A}_{w_0}$ such that $D_{\mathscr{C},\D}(u,u_0)=0$ and we obtain that $[u]\in \mathcal{A}_{w_0}$. To prove the claim, fix $\e>0$. Since $D_{\mathscr{C},\D}(u,u') = 0$, there exists a chain $\{A^{\N}_{w_1},\dots,A^{\N}_{w_m}\}$ that joins $u$ with $u'$ such that $ \sum_{l=1}^m\D(w_l) < \e.$ By Lemma \[lem:chain\], there exist $l_0 \in \{1,\dots,m\}$ and $v \in A^{\N}_{w_0}\cap A^{\N}_{w_{l_0}}$. It follows that $ \sum_{l=1}^{l_0}\D(w_l) < \e$ and so $D_{\mathscr{C},\D}(u,v) < \e$. As $\epsilon>0$ was arbitrary, this proves the initial claim and hence the lemma. \[lem:tree\] The metric space $(\mathcal{A},d_{\mathscr{C},\D})$ is a metric tree. First of all, since $(\mathcal{A},d_{\mathscr{C},\D})$ is Hausdorff and path-connected, it is also arcwise connected. Let $[w_1], [w_2]$ be two distinct arbitrary points in ${\mathcal{A}}$. We will show that there is a point of ${\mathcal{A}}\setminus\{[w_1],[w_2]\}$ (in fact, a whole continuum) that every path $\g$ from $[w_1]$ to $[w_2]$ must contain. This clearly implies that there can be no simple closed path containing $[w_1]$ and $[w_2]$, and therefore that ${\mathcal{A}}$ is a metric tree. (See [@Charatonik Theorem 1.1] for various characterizations of metric trees, called dendrites there, from which we are using characterization (20).) For each $n\in\N$ let $$\{ v_{n,1}, \dots, v_{n,m(n)} \} \subseteq A^n,$$ be all the vertices of $T_n$ lying on the unique combinatorial arc that joins $w_1(n)$ with $w_2(n)$, ordered so that $v_{n,1}=w_{1}(n)$, $v_{n,m(n)}=w_2(n)$, and $\{v_{n,i},v_{n,i+1}\} \in E_n$ for all $i=1,\dots,m(n)-1$. By by Definition \[def:combdata\], properties and , we have that for all $n\in\N$, $$\bigcup_{i=1}^{m(n+1)}\mathcal{A}_{v_{n+1,i}} \subseteq \bigcup_{i=1}^{m(n)}\mathcal{A}_{v_{n,i}}.$$ Let $$K_n:= \bigcup_{i=1}^{m(n)}\mathcal{A}_{v_{n,i}}\subseteq \mathcal{A},\text{ and }$$ $$K := \bigcap_{n=1}^{\infty} K_n \subseteq {\mathcal{A}}$$ Note that the above sets are all compact by Lemma \[lem:compact\]. We have that $[w_1],[w_2] \in K$. Given $\e>0$, there exists $n\in\N$ such that $\sup_{w\in A^n}\D(w)<\e/2$. Then, $$d_{\mathscr{C},\D}([w_1], K_n) \leq D_{\mathscr{C},\D}(w_1,v_{n,1}1^{\infty}) \leq \D(w_1(n)) + \D(v_{n,1}) <\e.$$ Therefore, $[w_1]\in K$. Similarly, $[w_2]\in K$. \[claim:Kcontinuum\] The set $K$ contains a continuum that joins $[w_1]$ with $[w_2]$. For any $\d>0$, there exists $n\in \N$ such that $\sup_{w\in A^n}\D(w)<\d/2$. By compactness of $\mathcal{A}$, for any $i =1,\dots, m(n)$ there exists $[v_i] \in \mathcal{A}_{v_{n,i}} \cap K$. It is then immediate that $([w_1], [v_1], [v_2],\dots,[v_{m(n)}], [w_2])$ is a $\d$-path joining $[w_1]$ with $[w_2]$. As $\delta$ was arbitrary, this shows that the connected component of $K$ containing $[w_1]$ also contains $[w_2]$. The set $K$ is contained in every path $\g$ from $[w_1]$ to $[w_2]$ in ${\mathcal{A}}$. Fix such a path $\g$ and let $\e>0$ and $[v_0] \in K$. Choose $n\in \N$ such that $\sup_{w\in A^n}\D(w)<\e$. Let $i \in \{1,\dots,m(n)\}$ such that $[v_0] \in \mathcal{A}_{v_{n,i}}$. Let $\{T_{n,j} =(V_j, E_j)\}_j$ enumerate the components of $T_n \setminus\{v_{n,i}\}$. For each $j$, let $X_j = \bigcup_{w\in V_j}\mathcal{A}_w$. These are compact sets whose union contains $\mathcal{A} \setminus \mathcal{A}_{v_{n,i}}$. Moreover, by Lemma \[lem:tree1\], the sets $\{X_j\}$ have the property that $X_j \cap X_{j'} \subseteq {\mathcal{A}}_{v_{n,i}}$ whenever $j\neq j'$. If neither of $[w_1]$ or $[w_2]$ is contained in $\mathcal{A}_{v_{n,i}}$, then $w_1(n)$ and $w_2(n)$ are contained in different subgraphs $T_{n,j}$ and hence $[w_1],[w_2]$ are contained in different sets $X_j$. In either case, the path $\g$ must intersect $\mathcal{A}_{v_{n,i}}$. Thus, $$d_{\mathscr{C},\D}(\g,[v_0]) \leq \D(w_{n,i}) < \e.$$ Since $\epsilon>0$ was arbitrary, we have $[v_0]\in \g$. Thus, every path in ${\mathcal{A}}$ from $[w_1]$ to $[w_2]$ contains $K$, which contains a fixed continuum joining $[w_1]$ and $[w_2]$. In particular, any two such paths must intersect somewhere other than their endpoints. This shows that ${\mathcal{A}}$ is a metric tree. \[rem:arc\] Given $w_1,w_2 \in A^{\N}$, let $K\subset \mathcal{A}$ be as in the proof of Lemma \[lem:tree\]. We showed above that $K$ contains a continuum that joins $[w_1]$ with $[w_2]$ and, conversely, that every path in $\mathcal{A}$ that joins $[w_1]$ with $[w_2]$ contains $K$. Therefore, $K$ is the unique arc that joins $[w_1]$ with $[w_2]$ in $\mathcal{A}$. Together, Lemmas \[lem:connected\], \[lem:bt\], and \[lem:tree\] prove Proposition \[prop:comb-tree\]. Doubling metric trees {#sec:doubling} ===================== Recall that a metric space is $C$-doubling if there exists a constant $C\geq 1$ such that for any $x\in X$ and $r>0$, the ball $B(x,r)$ can be covered by at most $C$ balls of radius $r/2$. Our goal here is to give some sufficient conditions for our combinatorial construction to yield a doubling metric tree. **For the remainder of Section \[sec:doubling\], we assume that $A$ is an alphabet and $\mathscr{C} = (A,(T_k)_{k\in\N})$ is combinatorial data as in Definition \[def:combdata\], with the additional assumption that each graph $T_k$ is a combinatorial tree.** Given $u\in A^*$ we define the *combinatorial boundary* of $A^{\N}_u$ by $$\partial_{\mathscr{C}}A^{\N}_u := A^\N_u \cap \bigcup_{v\in A^{|u|}\setminus\{u\}} \left( A^{\N}_v \wedge_{\mathscr{C}}A^{\N}_u \right).$$ In other words, $w\in \partial_{\mathscr{C}}A^{\N}_u$ if and only if $w\in A^{\N}_u$ and for every $n>|u|$, there exists $u' \in A^n \setminus A^n_u$ with $\{w(n),u'\}\in E_{n}$. \[prop:doubling\] Fix $N,n_0\in \N$, $c>1$, and $\d_1, \d_2\in (0,1)$. There exists $C>1$, depending only on these constants, with the following property. Assume that: 1. \[P1\] ${\operatorname{card}}{A} \leq N$. 2. \[P2\] $\text{Val}(T_k)\leq n_0$ for all $k\in\N$. 3. \[P3\] For all $w \in A^*$ and $i\in A$, $\d_1\D(w) \leq \D(wi) \leq \d_2\D(w)$. 4. \[P4\] For all $u\in A^*$ and all distinct $w_1,w_2 \in \partial_{\mathscr{C}}A^{\N}_u$ we have $d_{\mathscr{C},\D}([w_1],[w_2]) \geq c^{-1}\D(u).$ Then $(\mathcal{A},d_{\mathscr{C},\D})$ is $C$-doubling. Recall the definition of a parent word $u^{\uparrow}$. For the proof of Proposition \[prop:doubling\], we make the following definition. Given $r>0$ define $$A^*(r) := \left\{w\in A^* : \D(w)< r\text{ and }\D(w^{\uparrow}) \geq r\right\}.$$ \[rem:partition\] The set $A^*(r)$ induces a partition on $A^{\N}$. Namely, $A^{\N} = \bigcup_{u\in A^{*}(r)}A_u^{\N}$ and for distinct $w,u \in A^{*}(r)$ we have $A^{\N}_w \cap A^{\N}_u = \emptyset$. \[lem:partition\] Let $A$ and $\mathscr{C}$ satisfy \[P2\]. Then, for each $r>0$ and for each $w\in A^*(r)$, there exist at most $n_0$ words $u \in A^*(r) \setminus \{w\}$ such that $A^{\N}_w\wedge_{\mathscr{C}}A^{\N}_u \neq \emptyset$. Let $r>0$ and $w\in A^*(r)$. To prove the claim, let $u_1,\dots,u_n$ be words in $A^*(r) \setminus \{w\}$ such that $A^{\N}_w\wedge_{\mathscr{C}}A^{\N}_{u_i} \neq \emptyset$ for each $i$. Let $k_0 = |w|$. If $|u_i|< k_0$, then by Lemmas \[lem:intersection1\] and \[lem:intersection2\], there exists a unique $u_i' \in A^{k_0}_{u_i}$ such that $\{u_i'\}\in E_{k_0}$. If $|u_i| \geq k_0$, then let $u_i' = u_i(k_0)$. and by Lemma \[lem:adjacent\], we have that $\{w,u_i'\} \in E_{k_0}$. We claim that if $i\neq j$, then $u_i'\neq u_j'$. Assuming the claim, by \[P2\] we have that $n \leq n_0$. To prove the claim we fix distinct $i,j \in \{1,\dots,n\}$ and consider three possible cases. *Case 1.* Suppose that $|u_i|\geq k_0$ and $|u_j|\geq k_0$. For a contradiction, assume that $u_i' = u_j' = u'$. By Remark \[rem:partition\] we have that $u' \neq w$. Therefore, by Lemma \[lem:intersection2\], $\{u',w\} \in E_{k_0}$. Let $k = \max\{|u_i|,|u_j|\}$. By Lemma \[lem:adjacent\], there exist unique $w'' \in A^k$ and unique $u''\in A^k_{u'}$ such that $\{w'',u''\}\in E_k$. By Remark \[rem:partition\], either $u'' \not\in A_{u_i}^k$ or $u'' \not\in A_{u_j}^k$. Assuming the former (without loss of generality), by Lemma \[lem:intersection2\], we have $A^{\N}_{w}\wedge_{\mathscr{C}}A^{\N}_{u_i}= \emptyset$ which is a contradiction. *Case 2.* Suppose that $|u_i|\leq k_0$ and $|u_j|\leq k_0$. For a contradiction, assume that $u_i' = u_j' = u'$. Then $A^\N_{u_i}\cap A^\N_{u_j} \neq \emptyset$, which contradicts Remark \[rem:partition\]. *Case 3.* Suppose that $|u_i|\leq k_0$ and $|u_j|\geq k_0$. By Remark \[rem:partition\], $u_i' \neq w$. Now apply the arguments of Case 1 to the triple $u_i'$, $w$, and $u_j$. Let $[w]\in \mathcal{A}$ and $r>0$. To prove the proposition, it suffices to prove that the doubling property holds for the ball $B([w],r)$ if $r<c^{-1}{\mathop\mathrm{diam}\nolimits}{\mathcal{A}}$. Let $u_0$ be the unique element of $A^*(c\delta_1^{-1}r)$ such that $w\in A^\N_{u_0}$. \[claim:doubling1\] There exist at most $n_0$ words $u \in A^*(c\d_1^{-1}r) \setminus \{u_0\}$ such that $A^{\N}_{u_0}\wedge_{\mathscr{C}}A^{\N}_u \neq \emptyset$, and each such word $u$ satisfies $$c\delta_1^{-1}r > \Delta(u) \geq cr.$$ By Lemma \[lem:partition\], there exist at most $n_0$ such words $u \in A^*(c\d_1^{-1}r) \setminus \{u_0\}$. Moreover, by \[P3\], for each $u \in A^*(c\d_1^{-1}r)$, $$c\d_1^{-1}r > \D(u) \geq \d_1\D(u^{\uparrow}) \geq cr.\qedhere$$ \[claim:doubling2\] If $u\in A^*(c\d_1^{-1}r)$ and $A^{\N}_{u_0}\wedge_{\mathscr{C}}A^{\N}_u = \emptyset$, then for any $w'\in A^{\N}_u$ we have $d_{\mathscr{C},\D}([w],[w']) \geq r$. Let $\gamma \subset \mathcal{A}$ be the unique arc with endpoints $[w]$ and $[w']$. For each $k$, let $P_k$ be the simple path in $T_k$ from $w(k)$ to $w'(k)$. Let $n=\max\{|u|,|u_0|\}$. Then $P_n$ must contain a vertex $v\in A^n\setminus (A^n_{u_0} \cup A^n_{u})$, otherwise $A^{\N}_{u_0}\wedge_{\mathscr{C}}A^{\N}_u \neq \emptyset$. Consider the following two possible cases. *Case 1.* Suppose that $v\in A^*(c\d_1^{-1}r)$ or $v$ has a descendent in $A^*(c\d_1^{-1}r)$. Then $v$ is adjacent to two distinct vertices $v_1$ and $v_2$ of $P_n$. For $i=1,2$, let $w_i\in A^\mathbb{N}_v$ be such that $w_i(k)\in P_k$ and is adjacent to an element of $A^k_{v_i}$ for each $k\geq n$. Then $w_1$ and $w_2$ are distinct elements of $\partial_{\mathscr{C}}A^{\N}_{v}$. Moreover, by Remark \[rem:arc\], both $[w_1]$ and $[w_2]$ are in $\gamma$. Therefore, by the 1-bounded turning property of $\mathcal{A}$ and by \[P4\], $$d_{\mathscr{C},\D}([w],[w']) \geq {\mathop\mathrm{diam}\nolimits}{\gamma} \geq d_{\mathscr{C},\D}([w_1],[w_2]) \geq c^{-1}\D(v) \geq r.$$ *Case 2.* Suppose that $v$ is contained in $A^*_{v'}$ for some $v' \in A^*(c\d_1^{-1}r)$. Let $m =|v'|$. Note that $P_m$ must contain $v'$ and that $A^{\N}_{v'}$ is disjoint from both $A^{\N}_{w(m)}$ and $A^{\N}_{w'(m)}$. Therefore, $v'$ is adjacent to two distinct vertices of $P_m$. Now working as in Case 1, we obtain that $d_{\mathscr{C},\D}([w],[w']) \geq c^{-1}\D(v') \geq r$. \[claim:doubling3\] Let $u \in A^*(c\d_1^{-1}r)$ and let $k$ be the smallest positive integer such that $$k\geq \frac{\log((2c)^{-1}\d_1)}{\log(\d_2)}.$$ Then $${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_v) < r/2$$ for each $v\in A^{|u|+k}_u$. By the upper bound in \[P3\] we have that for every $v \in A^{|u|+k}_u$, $${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_v) \leq \D(v) \leq \d_2^k\D(u) < \d_2^k\d_1^{-1}cr \leq r/2.\qedhere$$ Let $\{u_1,\dots,u_p\}$ be all the words $u\in A^*(c\d_1^{-1}r)\setminus\{u_0\}$ such that $A^{\N}_{u_0}\wedge_{\mathscr{C}}A^{\N}_u \neq \emptyset.$ By Claim \[claim:doubling2\], $$B([w],r) \subseteq \bigcup_{i=0}^p\mathcal{A}_{u_i}.$$ Claim \[claim:doubling1\] implies that $p\leq n_0$. Claim \[claim:doubling3\] and imply that each of the sets $\mathcal{A}_{u_i}$ in this union can be covered by at most $N^k$ sets of diameter $< r/2$, hence $N^k$ balls of radius $r/2$. This completes the proof. Of the assumptions in Proposition \[prop:doubling\], the hardest to check is \[P4\]. We now give some sufficient conditions for \[P4\] which are easier to verify. For the next lemma we use the following notation. Consider combinatorial data $\mathscr{C} = (A,(T_k)_{k\in\N})$ as fixed at the beginning of this section. For each $k\in\N$ and $w\in A^k$, let $\partial_{\mathscr{C}} A^{k+1}_w$ be all words $u \in A^{k+1}_w$ for which there exists $u' \in A^{k+1}\setminus A^{k+1}_w$ with $\{u,u'\}\in E_{k+1}$. \[lem:suffdoubling\] Let $\mathscr{C} = (A,(T_k)_{k\in\N})$ be combinatorial data as fixed at the beginning of this section, and let $\D \in \mathscr{D}(A)$. Assume that the following conditions hold for each $k\geq 0$. 1. Suppose that $w,u,u' \in A^k$ are distinct with $\{w,u\}, \{w,u'\}\in E_k$. If $wi,wj,ul,u'l' \in A^{k+1}$ with $\{wi,ul\}, \{wj,u'l'\}\in E_{k+1}$, then $i\neq j$. 2. For any $w\in A^k$, and any distinct $u,u' \in \partial_{\mathscr{C}}A^{k+1}_w$, the arc $\{\{u,u_1\},\dots,\{u_n,u'\}\}$ joining $u$ with $u'$ in $T_{k+1}$ satisfies $$\D(u) + \D(u_1) + \cdots + \D(u_n) + \D(u') \geq \D(w).$$ Then \[P4\] of Proposition \[prop:doubling\] holds with $c=1$. In particular, ${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_u)=\Delta(u)$ for each $u\in A^*$ with at least two neighbors in $T_{|u|}$. For the proof of the lemma, given a chain $\mathcal{C} = \{A^{\N}_{u_1},\dots,A^{\N}_{u_n}\}$ joining two words in $A^\N$, we define the *depth* of $\mathcal{C}$ to be the number $\text{Depth}(\mathcal{C}) := \max\{|u_1|,\dots,|u_n|\}$ and the *$\D$-length of $\mathcal{C}$* to be $$\ell(\mathcal{C}) := \sum_{i=1}^n\D(u_i).$$ Fix $k\in\N$ and $w \in A^k$. Let $w_1, w_2$ be distinct elements of the combinatorial boundary $\partial_{\mathscr{C}}A^\mathbb{N}_w$. We will show that $d_{\mathscr{C},\D}([w_1],[w_2]) = \D(w)$. By definition of the combinatorial boundary and Lemma \[lem:intersection2\](3), there are distinct $u_1,u_2 \in A^k$, both adjacent to $w$, such that for each $i=1,2$, the combinatorial intersection $ A^{\N}_w\wedge_{\mathscr{C}}A^{\N}_{u_i}$ contains exactly two elements: $w_i$ which is in $A^{\N}_w$, and $w_i'$ which is in $A^{\N}_{u_i}$. On the one hand, $\{A^{\N}_w\}$ is a chain joining $w_1$ with $w_2$, so $d_{\mathscr{C},\D}([w_1],[w_2]) \leq \D(w)$. For the opposite inequality, fix $\mathcal{C} = \{A^{\N}_{v_1},\dots,A^{\N}_{v_n}\}$ to be a chain in $A^{\N}$ joining $w_1$ with $w_2$. We start by doing four reductions. First, if $A^{\N}_w \subset A^{\N}_{v_i}$ for some $i$, then we can replace $\mathcal{C}$ with $\mathcal{C}' = \{A^{\N}_w\}$ which has smaller $\D$-length. Therefore we may assume that for all $i$, either $A^{\N}_w\cap A^{\N}_{v_i} = \emptyset$, or $A^{\N}_{v_i} \subset A^{\N}_w$. Second, dropping some of the sets in the chain, if necessary, we may assume that $A^{\N}_{v_i}\subset A^{\N}_w$ for all $i$. Third, if $A^{\N}_{v_i} \subset A^{\N}_{v_j}$, then we can drop $A^{\N}_{v_i}$. Fourth, it follows from Lemma \[lem:chain\] that for any $l\geq \text{Depth}(\mathcal{C})$, $$\bigcup_{v\in P_l}A^{\N}_v \subset A^{\N}_{v_1}\cup \cdots\cup A^{\N}_{v_n},$$ where $P_{l}$ denotes the combinatorial arc in $T_{l}$ that joins $w'_1(l)$ with $w'_2(l)$. Therefore, the subset $$\{A^{\N}_{v_i} : A^{\N}_{v_i} \cap \bigcup_{v\in P_l}A^{\N}_v \neq \emptyset\}$$ forms a chain joining $w_1$ and $w_2$. In other words, we may drop all sets $A^\mathbb{N}_{v_i}$ from the chain such that $A^{\N}_{v_i} \cap \bigcup_{v\in P_l}A^{\N}_v = \emptyset$. The four reductions imply that we may assume that for all $i$, 1. $A^{\N}_{v_i}\subset A^{\N}_w$; 2. if $j\neq i$, then $A^{\N}_{v_i}\cap A^{\N}_{v_j} = \emptyset$; 3. for all $l\geq \text{Depth}(\mathcal{C})$, there exists $v\in P_l$ such that $A^{\N}_v \subset A^{\N}_{v_i}$; and 4. for all $l\geq \text{Depth}(\mathcal{C})$, $\bigcup_{v\in P_l}A^{\N}_v \subset A^{\N}_{v_1}\cup \cdots\cup A^{\N}_{v_n}$. Let $k_0 = \text{Depth}(\mathcal{C})$ and $i_0\in \{1,\dots,n\}$ such that $|v_{i_0}|=k_0$. If $k_0 = |w|$, then $\mathcal{C}=\{A^{\N}_w\}$ and the $\D$-length of $\mathcal{C}$ is equal to $\D(w)$. Assume now that $k_0 > |w|$. Then $v_{i_0}^{\uparrow}$ is contained in $P_{k_0-1}$. Moreover, $v_{i_0}^\uparrow$ has valency 2 in $P_{k_0-1}$, because the endpoints of $P_{k_0-1}$ are in $A^{k_0-1}_{u_i}$ and not in $A^{k_0-1}_w$. By (iii) and assumption (1) of the lemma, $A^{k_0}_{v_{i_0}^{\uparrow}} \cap P_{k_0}$ has at least two elements. By (ii), (iv) and the assumption that $|v_{i_0}| = \text{Depth}(\mathcal{C})$, each element of $A^{k_0}_{v_{i_0}^{\uparrow}} \cap P_{k_0}$ must be in $\{v_1, \dots, v_n\}$. Enumerate them as $\{v_{j_1}, v_{j_2}, \dots, v_{j_p}\}$. Since $v_{i_0}^\uparrow$ has valency 2 in $P_{k_0-1}$, the elements of $\{v_{j_1}, v_{j_2}, \dots, v_{j_p}\}$ contain the vertices of a simple path joining two points of $\partial_{\mathscr{C}}A^{k_0}_{v_{i_0}^{\uparrow}}$. But then, by assumption (2) of the lemma, $$\D(v_{j_1}) + \cdots + \D(v_{j_p}) \geq \D(v_{i_0}^{\uparrow})$$ and we can replace $\mathcal{C}$ with the chain $$\mathcal{C} \cup \{ A^{\N}_{v_{i_0}^{\uparrow}}\} \setminus \{ A^{\N}_{v_{i}} : v_i \in A_{v_{i_0}^{\uparrow}}^{k_0} \}$$ which has at most the $\D$-length of $\mathcal{C}$. Working in similar fashion, we can show that if $\text{Depth}(\mathcal{C}) > |w|$, then there exists a chain $\mathcal{C}'$ joining $w_1$ with $w_2$ such that $\text{Depth}(\mathcal{C}') = \text{Depth}(\mathcal{C}) - 1$ and has at most the $\D$-length of $\mathcal{C}$. Applying a backwards induction on the depth of $\mathcal{C}$, we obtain that $$\ell(\mathcal{C}) \geq \ell(\{A^\mathbb{N}_w\}) = \D(w).$$ Therefore, $d_{\mathscr{C},\D}([w_1],[w_2]) \geq \D(w)$. For the final statement in the lemma, any $u\in A^k$ with two distinct neighbors must have at least two distinct words in its combinatorial boundary, and so $${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_u) \geq \D(w)$$ by the first part of the lemma. The reverse inequality follows from Lemma \[lem:dist\]. A simple case in which the conditions of Proposition \[prop:doubling\] can be verified, using Lemma \[lem:suffdoubling\], is in the case of quasi-arcs. Let $\mathscr{C} = (A,(T_k)_{k\in\N})$ be combinatorial data such that ${\operatorname{card}}{A} = N \geq 2$ and each $T_k = (A^k,E_k)$ is a combinatorial arc. Let $\D \in \mathscr{D}(A)$ satisfy \[P3\] and assume that, for all $k\geq 0$ and $w\in A^k$, $$\label{eq:longchain} \sum_{wi \in A_w^{k+1}}\D(wi) \geq \D(w).$$ Then, $(\mathcal{A},d_{\mathscr{C},\D})$ is a doubling bounded turning arc. First, since ${\operatorname{card}}{A}= N$ and each $T_k$ is a combinatorial arc, conditions \[P1\] and \[P2\] of Proposition \[prop:doubling\] are immediately satisfied. Since ${\operatorname{card}}{A}\geq 2$, assumption (1) of Lemma \[lem:suffdoubling\] is satisfied and by (\[eq:longchain\]), assumption (2) of Lemma \[lem:suffdoubling\] is satisfied. Hence, by Lemma \[lem:suffdoubling\] and Proposition \[prop:comb-tree\], $(\mathcal{A},d_{\mathscr{C},\D})$ is doubling and bounded turning. It remains to show that $(\mathcal{A},d_{\mathscr{C},\D})$ is an arc. By design, there exists exactly two words $w_1,w_2 \in A^{\N}$ such that for all $n\in\N$, the valency of $w_i(n)$ in $T_n$ is 1. Recalling the definition of $K$ from the proof of Lemma \[lem:tree\], we note that $K = \mathcal{A}$. Therefore, $(\mathcal{A},d_{\mathscr{C},\D})$ is an arc. Characterization of quasiconformal trees {#sec:characterization} ======================================== We now claim that our combinatorial constructions above describe all quasiconformal trees up to bi-Lipschitz equivalence. The following result proves part of Theorem \[thm:maincombthm\], while providing additional details, and is the goal of this section. \[thm:main\] Let $(X,d)$ be an $N$-doubling, $C$-bounded turning tree. Then for any $M\in\mathbb{N}$ sufficiently large, $K_1>0$ sufficiently small, and $K_2\in [\frac{1}{2},1)$, there exist: 1. an alphabet $A=\{1,\dots,M\}$, 2. combinatorial data $\mathscr{C}=(A,(T_k)_{k\in\N})$ with each $T_k$ a combinatorial tree, 3. a diameter function $\D \in \mathscr{D}(A,K_1, K_2)$ such that $(\mathcal{A},d_{\mathscr{C},\D})$ is bi-Lipschitz equivalent to $X$. The sufficient condition on $M$ depends only on $N$ and $C$. The sufficient condition on $K_1$ depends only on $M$, $N$, and $C$. The bi-Lipschitz constant depends only on $N$, $C$, $K_2/K_1$, and ${\mathop\mathrm{diam}\nolimits}(X)$. Moreover, $(\mathscr{C},\Delta)$ satisfies the conditions of Proposition \[prop:doubling\]. We first make some small reductions. If $X$ is a single point, then Theorem \[thm:main\] is easy. For example, one may take $M=2$, $\D\in\mathscr{D}\left(A,\frac{1}{3}, \frac{1}{3}\right)$, and each $T_k$ a combinatorial arc. Thus, we may assume that ${\mathop\mathrm{diam}\nolimits}(X)>0$ and so, by rescaling, that ${\mathop\mathrm{diam}\nolimits}(X)=1$. We may also assume that the bounded turning constant $C$ is equal to $1$, by replacing the metric $d$ on $X$ with a bi-Lipschitz equivalent $1$-bounded turning metric. (See [@BM Lemma 2.5].) **All these assumptions are in force for the remainder of Section \[sec:characterization\]. Thus, we fix an $N$-doubling, $1$-bounded turning metric tree $X$ of diameter $1$.** Subdividing into a uniform number of pieces ------------------------------------------- To prove Theorem \[thm:main\], we use a construction of Bonk-Meyer [@BM] to decompose the tree $X$ into suitable pieces. We then modify this construction to decompose $X$ into an equal number of pieces at each scale. We first summarize the results we need from [@BM Section 5]. \[prop:BM\] Let $\delta>0$ sufficiently small, depending on $N$. Then, for each $n\in\mathbb{N}$, there is a constant $M(N,\delta)\in\mathbb{N}$ and $\delta^n$-separated sets $V_n\subseteq X$ satisfying $$V_1 \subseteq V_2 \subseteq \dots$$ with the following properties. Write $\mathcal{T}_n$ for the collection of closures of components of $X\setminus V_n$. Then: 1. Each $T\in\mathcal{T}_n$ is a connected subset (hence subtree) of $X$ with $\emptyset\neq T \cap\overline{X\setminus T} \subseteq V_n$. 2. Distinct elements $T,T'\in\mathcal{T}_n$ have at most one point in common, and such a common point is an element of $V_n$. 3. Each element of $V_n$ is in exactly two elements of $\mathcal{T}_n$. 4. Each element of $\mathcal{T}_{n+1}$ ($n\geq 1$) is in exactly one element of $\mathcal{T}_{n}$, and each element of $\mathcal{T}_{n}$ is the union of all elements of $\mathcal{T}_{n+1}$ inside it. 5. We have $\delta^n \leq {\mathop\mathrm{diam}\nolimits}(T) \leq K\delta^n$ for each $T\in\mathcal{T}_n$, where $K$ is a constant depending only on $N$. 6. Each element of $\mathcal{T}_{n}$ contains at least two and at most $M(N,\delta)$ elements of $\mathcal{T}_{n+1}$. 7. Each element of $\mathcal{T}_{n}$ intersects at most $M(N,\delta)$ other elements of $\mathcal{T}_n$. The first four items appear explicitly in [@BM Lemma 5.1]. The fifth appears in [@BM Equation (5.3)]. The existence of the upper bound $M(N,\delta)$ in (6) and (7) is an immediate consequence of (1)-(5) and the doubling property, as in [@BM Lemma 5.7]. The lower bound of two in (6) follows from (4) and (5) if $\delta<1/K$. Bonk and Meyer refer to the elements of $\mathcal{T}_n$ as “$n$-tiles”, but we will reserve the word “tiles” for the modifications we construct below. Before that, we observe that adjacency graphs induced by these sets form combinatorial trees. \[lem:tiletree\] Let $X$ be a metric tree. Let $\mathcal{S}$ be a finite collection of compact, connected subsets of $X$ such that $ \cup_{S\in\mathcal{S}} S = X$ and no point of $X$ is in more than two different sets of $\mathcal{S}$. Then the graph $G$ such that $$V(G) = \{ S\in\mathcal{S}\},$$ $$E(G) = \{ \{S,S'\}\subseteq V(G) : S\neq S' \text{ and } S \cap S' \neq \emptyset\}$$ is a combinatorial tree. The connectedness of $G$ follows easily from the facts that $X$ is connected, all $S\in\mathcal{S}$ are compact, and $\cup_{S\in\mathcal{S}} S = X$. To see that $G$ is a combinatorial tree, we will use the following simple equivalent characterization of combinatorial trees: A connected finite graph is a combinatorial tree if and only if the removal of any edge disconnects it. Thus, suppose that the removal of an edge $\{S,S'\}$ from $G$ left it connected. Let $S=S_0, S_1, \dots, S_n=S'$ be the ordered vertices along a simple path from $S$ to $S'$ in $G$ avoiding this edge; note that $n\geq 2$. Let $x\in S\cap S'$, $p\in S\cap S_1$, and $q\in S'\cap S_{n-1}$. The points $x$, $p$, and $q$ are distinct, by the assumption that no point is in more than two elements of $\mathcal{S}$. Similarly, $x$ is disjoint from $S_{i}$ for each $1\leq i \leq n-1$. There is an arc from $p$ to $q$ in $S \cup S'$, which must pass through $x$. Since $X$ is a metric tree, $p$ and $q$ must be in distinct connected components of $X\setminus \{x\}$. On the other hand, $\cup_{i=1}^{n-1} S_i$ is a connected subset of $X\setminus \{x\}$ containing both and we reach a contradiction. We now modify the construction of Proposition \[prop:BM\] so that each tile has an equal number of children. This requires us to give up some control on the diameters of the tiles. However, it is crucial to retain the property that the boundary points of a given tile are “well-separated”, in the sense that the distance between two distinct boundary points of a tile is always comparable to the diameter of the tile. This is property (6) of Lemma \[lem:tiles\] below. Fix $\delta$ sufficiently small, depending on $N$, so that Proposition \[prop:BM\] holds, and so that in addition $K\delta< 1/2$, where $K$ is the constant from Proposition \[prop:BM\](5). Thus, we have constants $K=K(N)$ and $M(N,\delta)$ from Proposition \[prop:BM\], items (5) and (6). \[lem:tiles\] Let $M\geq M(N,\delta)$, $K_1\in (0,K^{-1}\delta^{\log_2(M)+1}]$, and $K_2\in [\frac{1}{2},1)$. Let $A=\{1,\dots,M\}$. Then there is a collection of closed subsets $X_w\subset X$, for all $w\in A^*$, satisfying the following properties. 1. For each $w\in A^*$, $X_w$ is a connected subset (hence subtree) of $X$, and $X_{\varepsilon} =X$. 2. For each $w\in A^*$ and $i\in A$, $X_{wi}\subseteq X_w$. Moreover, $X_w = \bigcup_{i\in A}X_{wi}$. 3. For each $w\in A^*$ and $i\in A$, $$K_1{\mathop\mathrm{diam}\nolimits}{X_{w}} \leq {\mathop\mathrm{diam}\nolimits}{X_{wi}} \leq K_2{\mathop\mathrm{diam}\nolimits}{X_w}.$$ 4. For each $w\in A^{*}\setminus\{\varepsilon\}$ and every $x\in X_{w}\cap \overline{X\setminus X_w}$, we have that $x$ is a leaf of $X_w$ and contained in $X_{w'}$ for exactly one $w'\in A^{|w|}\setminus \{w\}$. 5. For every $w\in A^*$ and distinct $i,j \in A$ we have that $X_{wi}\cap X_{wj}$ is either a point or empty. 6. There exists $K_3 \in (0,1)$ such that for all $w\in A^*$ and for all distinct $x,y \in X_{w}\cap \overline{X\setminus X_w}$, we have $$d(x,y) \geq K_3 {\mathop\mathrm{diam}\nolimits}{X_w}.$$ Fix $\delta, M, K, K_1, K_2$ as above, and let $A=\{1,\dots,M\}$. We first relabel the collections $\mathcal{T}_n$ constructed in Proposition \[prop:BM\]. Set $T_{\varepsilon}=X$. We write $\mathcal{T}_1 = \{T_1,\dots,T_{m_{\varepsilon}}\}$. Assume now that for some $n\in\N$ and some $w\in \N^n$ we have defined $T_w$ to be an element of $\mathcal{T}_n$. Then we write $\{T_{w1},\dots, T_{wm_w}\}$ to be the elements of $\mathcal{T}_{n+1}$ contained in $T_w$. By Proposition \[prop:BM\](6), we have $2 \leq m_w \leq M$. Therefore, for every $T_w$ defined, we have $w\in A^*$. We set $\mathcal{W}$ to be the set of all words $w$ in $A^*$ for which $T_w$ has been defined. Given integer $n\geq 0$ and $w\in A^*$ we denote $\mathcal{W}^n = \mathcal{W}\cap A^n$, $\mathcal{W}_w = \mathcal{W} \cap A^*_w$, and $\mathcal{W}^n_w = \mathcal{W}\cap A^n_w$. We now define the family $\{X_w\}_{w\in A^*}$ in an inductive manner. <span style="font-variant:small-caps;">Step 0.</span> Set $X_{\varepsilon} = T_{\varepsilon} = X$. <span style="font-variant:small-caps;">Inductive hypothesis.</span> Suppose that for some integer $k\geq 0$ we have defined closed sets $\{X_w\}_{w\in A^k}$ such that the properties of the lemma up to level $k$ hold, with $K_3=\delta/K$. That is, we assume that the following conditions hold. 1. For each $l\leq k$ and $w\in A^l$, $X_w$ is a connected subset of $X$. 2. For each $l\leq k-1$, $w\in A^{l}$, and $i\in A$, we have $X_{wi}\subseteq X_w$. Moreover, $X_w = \bigcup_{i\in A}X_{wi}$. 3. For each $l\leq k-1$, $w\in A^{l}$, and $i\in A$, we have $$K_1{\mathop\mathrm{diam}\nolimits}{X_{w}} \leq {\mathop\mathrm{diam}\nolimits}{X_{wi}} \leq K_2{\mathop\mathrm{diam}\nolimits}{X_w}.$$ 4. For each $l\leq k$, $w\in A^{l}\setminus\{\varepsilon\}$ and every $x\in X_{w}\cap \overline{X\setminus X_w}$, we have that $x$ is a leaf of $X_w$ and contained in $X_{w'}$ for exactly one $w'\in A^{l}\setminus \{w\}$. 5. For each $l\leq k-1$, $w\in A^l$ and distinct $i,j \in A$ we have that $X_{wi}\cap X_{wj}$ is either a point or empty. 6. For each $l\leq k$, $w\in A^l$, and distinct $x,y \in X_{w}\cap \overline{X\setminus X_w}$, we have $$d(x,y) \geq K_3 {\mathop\mathrm{diam}\nolimits}{X_w}.$$ In addition, we make the following inductive assumption: 1. For each $w\in A^k$ there exists $u\in\mathcal{W}$ and distinct $ui_1,\dots, ui_q \in \mathcal{W}_u^{|u|+1}$ such that $X_w = \bigcup_{j=1}^q T_{ui_j}$. Note that (7) holds when $k=0$. <span style="font-variant:small-caps;">Inductive step.</span> We now describe the construction of the sets $\{X_w\}_{w\in A^{k+1}}$. Fix a word $w\in A^k$. By assumption (7), $X_w = T_{ui_1}\cup\cdots \cup T_{u i_q}$. For simplicity, we assume that $i_j = j$ for all $j$. By Proposition \[prop:BM\](6), $q \leq M$. *Case 1: $q=M$.* In this case we set $X_{wj}=T_{uj}$ for $j=1,\dots,M$. *Case 2: $q < M$.* Let $n$ be the smallest integer such that $$\label{eq:tilesum} \sum_{j=1}^q{\operatorname{card}}(\mathcal{W}^{n+|u|}_{uj}) \geq M.$$ By Proposition \[prop:BM\](6), $2 \leq n \leq \log_2{M}+1$. *Case 2.1 : the sum in (\[eq:tilesum\]) is equal to $M$.* In this case we set $$\{X_{wi} : i\in A\} := \left\{T_{v} : v \in \bigcup_{j=1}^q\mathcal{W}^{n+|u|}_{uj}\right\}.$$ *Case 2.2: the sum in (\[eq:tilesum\]) is strictly greater than $M$.* Enumerate the elements of $\bigcup_{j=1}^q\mathcal{W}^{n-1+|u|}_{uj} =\{u_1,\dots,u_r\}$ so that for each $i\in\{1,\dots,r\}$ the set $$T_{u_i}\cap \overline{X_w \setminus (T_{u_1}\cup\cdots \cup T_{u_{i}})}$$ contains only one point. In other words, the sets $X_w\setminus T_{u_1}$, $(X_w\setminus T_{u_1})\setminus T_{u_2}$, etc. are connected. That this is possible follows from Lemma \[lem:tiletree\] and the fact that every finite combinatorial tree has a leaf. By minimality of $n$, we have that $r<M$. Now let $m$ be the smallest integer in $\{1,\dots,r\}$ such that $$\label{eq:tilesum2} \sum_{i=1}^m {\operatorname{card}}(\mathcal{W}_{u_i}^{n+|u|}) + (r-m) \geq M.$$ Note that if $m=r$, then holds by , so such a minimal $m$ exists. *Case 2.2.1: the sum in (\[eq:tilesum2\]) is equal to $M$*. Then by minimality of $n$ we have $m<r$ and we set $$\{X_{wi} : i\in A\} := \left\{T_{v} : v \in \bigcup_{j=1}^m\mathcal{W}^{n+|u|}_{u_i} \cup\{u_{m+1},\dots, u_r\}\right\}.$$ *Case 2.2.2: the sum in (\[eq:tilesum2\]) is strictly greater than $M$.* As before, enumerate the elements of $\mathcal{W}_{u_m}^{n+|u|} = \{u_mi_1, \dots, u_mi_l\}$ so that for each $j\in\{1,\dots,l\}$ the set $$T_{u_mi_j}\cap \overline{T_{u_m} \setminus (T_{u_mi_1}\cup\cdots \cup T_{u_{m}i_j})}$$ contains only one point. By the minimality of $m$ (and the fact that $r<M$) we have $$\sum_{i=1}^{m-1} {\operatorname{card}}(\mathcal{W}_{u_i}^{n+|u|}) + (r-(m-1)) \leq M-1$$ and so $$\label{eq:p} \sum_{i=1}^{m-1} {\operatorname{card}}(\mathcal{W}_{u_i}^{n+|u|}) + (r-m) \leq M-2.$$ Let $$p= M-1 - (r-m) - \sum_{i=1}^{m-1} {\operatorname{card}}(\mathcal{W}_{u_i}^{n+|u|}).$$ Note that $p\geq 1$ by . Moreover, $p\leq l-1={\operatorname{card}}{\mathcal{W}_{u_m}^{n+|u|}}-1$, otherwise $$\sum_{i=1}^{m} {\operatorname{card}}(\mathcal{W}_{u_i}^{n+|u|}) + (r-m) \leq M-1,$$ contradicting . Define now $$\mathcal{U} := \bigcup_{i=1}^{m-1}\mathcal{W}_{u_i}^{n+|u|} \cup \{u_mi_1,\dots, w_mi_p\} \cup \{w_{m+1},\dots,w_r\}.$$ Note that ${\operatorname{card}}(\mathcal{U}) = M-1$ by choice of $p$. Set $$\{X_{wi} : i\in A\} := \{T_v : v\in\mathcal{U}\} \cup \{\overline{T_{u_m} \setminus (T_{u_mi_1}\cup\cdots \cup T_{u_{m}i_p})}\}.$$ To complete the inductive step and the proof of Lemma \[lem:tiles\], it remains to check that properties (1)–(7) above are satisfied up to level $k+1$. Firstly, by design, properties (2) and (7) are immediately satisfied. Property (1) follows from Proposition \[prop:BM\](1) and the design of the enumerations $\{u_1,\dots,u_r\}$ and $\{u_mi_1, \dots, u_mi_l\}$. Properties (4), (5) for level $k+1$ follows from properties (4), (5) for level $k$, Proposition \[prop:BM\](3), and the design of the two enumerations $\{u_1,\dots,u_r\}$ and $\{u_mi_1, \dots, u_mi_l\}$. For property (3), fix $w\in A^k$ and $i\in A$. By (7), there exists $u\in \mathcal{W}^l$ and $uj \in \mathcal{W}^{l+1}$ such that $T_{uj} \subset X_w \subset T_u$. By the design above, there exists $v \in \mathcal{W}^{l+n}_u$ and $vj' \in \mathcal{W}^{l+n+1}_u$ such that $T_{vj'} \subset X_{wi} \subset T_{v}$ and $2 \leq n \leq \log_2{M}+1$. Therefore, applying Proposition \[prop:BM\](5), $$K_1\leq K^{-1}\d^{\log_2{M}+1} \leq \frac{{\mathop\mathrm{diam}\nolimits}{X_{wi}}}{{\mathop\mathrm{diam}\nolimits}{X_w}} \leq K\delta \leq K_2.$$ Finally, for property (6), fix $w\in A^{k+1}$ and distinct $x,y \in X_w \cap \overline{X \setminus X_w}$. By (7), we know that $X_w = T_{u i_1}\cup \cdots \cup T_{u i_n}$ for some $u\in \mathcal{W}^l$ and $u_{i_1}, \dots, u_{i_n} \in \mathcal{W}^{l+1}$. By Proposition \[prop:BM\](1), $x,y$ have distance at least $\delta^{l+1}$ so $${\mathop\mathrm{dist}\nolimits}(x,y) \geq \d^{l+1} \geq (\d/K) {\mathop\mathrm{diam}\nolimits}{T_u} \geq (\d/K){\mathop\mathrm{diam}\nolimits}{X_w}. \qedhere$$ We call the sets $X_w$ constructed in Lemma \[lem:tiles\] “tiles”. We observe that these new tiles also maintain the property that they can only touch a controlled number of tiles of the same scale: \[lem:tiledoubling\] There is a constant $n_0$, depending only on the doubling constant of $X$ and the constants from Lemma \[lem:tiles\], such that if $w\in A^*$, then $${\operatorname{card}}\{ v\in A^{|w|}: v\neq w, X_v \cap X_w \neq \emptyset\} \leq n_0.$$ Let $$W = \{ v\in A^{|w|}: v\neq w, X_v \cap X_w \neq \emptyset\}.$$ For each $v\in W$, Lemma \[lem:tiles\](5) implies that $X_w\cap X_v$ is a single point, which we call $x_{v}\in X_w \cap X_v$. Moreover, properties (4) and (6) of Lemma \[lem:tiles\] imply that if $v,v'\in W$ and $v\neq v'$, then $$d(x_v, x_{v'}) \geq K_3 {\mathop\mathrm{diam}\nolimits}(X_w).$$ Since all the points $\{x_v: v\in W\}$ are contained in $X_w$, the doubling property of $X$ completes the proof. Definition of combinatorial data {#subsec:charcombdata} -------------------------------- Fix $\delta$ as above Lemma \[lem:tiles\] and apply Lemma \[lem:tiles\] with fixed parameters $M\in\N$ and $K_1, K_2\in (0,1)$ as in the statement of that lemma. Let $A=\{1, \dots, M\}$. We define combinatorial data $\mathscr{C}=(A,(T_k)_{k\in\N})$ by setting $T_k=(A^k, E_k)$, where two words $v,w$ of $A^k$ are adjacent if and only if $X_v \cap X_w \neq \emptyset$. $\mathscr{C}$ satisfies the conditions of Definition \[def:combdata\], and each graph $T_k$ is a combinatorial tree. Property of Definition \[def:combdata\] is immediate. That $T_k$ is a (connected) combinatorial tree follows from Lemma \[lem:tiletree\]. Property of Definition \[def:combdata\] holds similarly, taking $X=X_w$, which is connected, and again using Lemma \[lem:tiletree\]. For Property , consider $\{w,u\}\in E_k$. Then there is a point $x\in X_w\cap X_u$. By Lemma \[lem:tiles\](2), there are words $wi$ and $uj$ such that $x\in X_{wi} \cap X_{uj}$, and therefore $\{wi, uj\}\in E_{k+1}$. One basic consequence of this construction of combinatorial data is the following. \[lem:tiletouch\] If $w,u\in A^*$ and $A^\N_w \wedge_\mathscr{C} A^N_u \neq \emptyset$, then $X_w \cap X_u \neq \emptyset$. Let $w,u\in A^*$ with $A^\N_w \wedge_\mathscr{C} A^N_u \neq \emptyset$. By Lemma \[lem:intersection1\], there are then $k\in\mathbb{N}$, $w'\in A^{k}_w$, and $u'\in A^{k}_u$ with $\{w',u'\}\in E_k$. It follows from the definition of $\mathscr{C}$ that $X_{w'} \cap X_{u'}\neq \emptyset$, $X_{w'}\subseteq X_w$, and $X_{u'}\subseteq X_u$. This proves the lemma. Definition of diameter function ------------------------------- We continue to use the quasiconformal tree $X$ fixed at the start of Section \[sec:characterization\], and the constants $M,K_1, K_2$ and combinatorial data $\mathscr{C}=(A, (T_k)_{k\in\N})$ fixed at the start of Section \[subsec:charcombdata\]. We now define a diameter function $\D\in \mathscr{D}(A,K_1,K_2)$ with the following two rules. - $\D(\varepsilon) =1$. - Suppose that for some $w\in A^*$ we have defined $\D(w)$. 1. If $\D(w) \leq {\mathop\mathrm{diam}\nolimits}{X_w}$, then we define $\D(wi)= K_2\D(w)$ for all $i\in A$. 2. If $\D(w) > {\mathop\mathrm{diam}\nolimits}{X_w}$, then we define $\D(wi)=K_1 \D(w)$ for all $i\in A$. This satisfies Definition \[def:diam\], with property (3) following from the fact that $K_1<K_2<1$. We now show that $\Delta(w)$ is always comparable to ${\mathop\mathrm{diam}\nolimits}(X_w)$. This argument is very similar to the proof of Theorem A in [@HM 4.1]. For all $w\in A^*$, $$\label{eq:comparable} (K_2/K_1)^{-1} \D(w) \leq {\mathop\mathrm{diam}\nolimits}(X_w) \leq (K_2/K_1)\D(w).$$ By Lemma \[lem:tiles\](3) we have for all $w\in A^*$ $$K_1{\mathop\mathrm{diam}\nolimits}(X_w) \leq {\mathop\mathrm{diam}\nolimits}(X_{wi}) \leq K_2{\mathop\mathrm{diam}\nolimits}(X_w).$$ Note that holds for $w=\varepsilon$, since $\Delta(\varepsilon)={\mathop\mathrm{diam}\nolimits}(X_\varepsilon)=1$. Assume by induction that we have a word $w$ such that holds. Consider any $i\in A$. There are two possibilities. *Case 1*: $\D(w) \leq {\mathop\mathrm{diam}\nolimits}(X_w)$. In this case, we have $$\D(wi) = K_2\D(w) \leq K_2{\mathop\mathrm{diam}\nolimits}(X_w) \leq (K_2/K_1){\mathop\mathrm{diam}\nolimits}(X_{wi})$$ and $${\mathop\mathrm{diam}\nolimits}(X_{wi}) \leq K_2{\mathop\mathrm{diam}\nolimits}(X_w) \leq K_2(K_2/K_1)\D(w) =(K_2/K_1)\D(wi),$$ which together prove for the word $wi$ in case 1. *Case 2*: $\D(w) > {\mathop\mathrm{diam}\nolimits}(X_w)$. In this case, we have $$\D(wi) = K_1\D(w) \leq K_1(K_2/K_1){\mathop\mathrm{diam}\nolimits}(X_w) \leq (K_2/K_1){\mathop\mathrm{diam}\nolimits}(X_{wi})$$ and $${\mathop\mathrm{diam}\nolimits}(X_{wi}) \leq K_2{\mathop\mathrm{diam}\nolimits}(X_w) < K_2\D(w) = (K_2/K_1)\D(wi),$$ which together prove for the word $wi$ in case 2. As in Section \[sec:diameter\], let $\sim$ be the equivalence relation on $A^{\N}$ induced by the diameter function $\D$ and let $\mathcal{A}=A^{\N}/\sim$ and $\mathcal{A}_w = A^{\N}_w/\sim$. Proof of Theorem \[thm:main\] ----------------------------- A consequence of Lemma \[lem:tiles\](2) is that, for each $x\in X$, there exists an infinite word $w_x\in A^{\mathbb{N}}$ such that $x \in X_{w(n)}$ for all $n\in\N$. We therefore define a map $f\colon X \rightarrow \mathcal{A}$ by $f(x) = [w_x]$. \[lem:welldef\] The map $f: X \to \mathcal{A}$ defined above is well-defined and surjective. Suppose that there exist two words $w,u \in A^\N$ such that for all $n\in\N$, $x\in X_{w(n)}\cap X_{u(n)}$. Then, by the construction of the combinatorial data $\mathscr{C}$, for each $n\in\N$ we have $\{w(n),u(n)\} \in E_n$. (Recall that $E_n$ is the set of edges of $T_n$.) Thus, for each $n\in\N$, the set $\{A^{\N}_{w(n)}, A^{\N}_{u(n)}\}$ is a chain that joins $w$ with $u$, and so $d_{\mathscr{C},\D}([w],[u]) \leq \D(w(n)) + \D(u(n)) \rightarrow 0$ as $n\rightarrow\infty$. We therefore have that $d_{\mathscr{C},\D}([w],[u]) = 0$, which implies that $[w]=[u]$. This shows that $f$ is well-defined. To show that $f$ is surjective, consider an arbitary $[u]\in\mathcal{A}$. We have nested compact tiles $$X_{u(1)} \supseteq X_{u(2)} \supseteq X_{u(3)} \dots$$ in $X$. Let $x\in \cap_{n\in\mathbb{N}} X_{u(n)}$. If $f(x)=w\in\mathcal{A}$, then by definition of $f$ we have $$x\in X_{w(n)} \cap X_{u(n)} \text{ for all } n\in\mathbb{N}.$$ As before, $u(n)$ and $w(n)$ are adjacent in $T_n$ for each $n$, and hence again $$d_{\mathscr{C},\Delta}([u],[w]) \leq \Delta(u(n))+\Delta(w(n))\rightarrow 0.$$ Thus, $[u]=[w]=f(x)$ and $f$ is surjective. The proof of Theorem \[thm:main\] concludes with the next two results. \[prop:bilipschitz\] The map $f: (X,d) \to (\mathcal{A},d_{\mathscr{C},\D})$ is bi-Lipschitz, with constant depending only on $K_1$, $K_2$, and $K_3$. Fix $x,y \in X$. We first claim that $d_{\mathscr{C},\D}(f(x),f(y)) \geq \frac{K_1}{K_2} d(x,y)$. Suppose that $f(x)=[w]$ and $f(y) = [u]$. Let $\{ A^{\N}_{w_1},\dots, A^{\N}_{w_m}\}$ be a chain joining $w$ with $u$. Since $w\in A^\N_{w_1}$, we have $w_1=w(|w_1|)$ and therefore $x\in X_{w_1}$; similarly, $y\in X_{w_m}$. We also have $X_{w_i}\cap X_{w_{i+1}} \neq \emptyset$ for each $i\in\{1,\dots,m-1\}$, by Lemma \[lem:tiletouch\]. Therefore, using the triangle inequality and , we have $$\label{eq:diameters} \sum_{i=1}^m \D(w_i) \geq \frac{K_1}{K_2}\sum_{i=1}^m {\mathop\mathrm{diam}\nolimits}{X_{w_i}} \geq \frac{K_1}{K_2}d(x,y).$$ Taking the infimum over all possible chains, we obtain $d_{\mathscr{C},\D}(f(x),f(y)) \geq \frac{K_1}{K_2}d(x,y)$, as desired. We now claim that $$\label{eq:lipschitz} d_{\mathscr{C},\D}(f(x),f(y)) \lesssim d(x,y),$$ with implied constant depending only on $K_1, K_2, K_3$. Let $w_0$ be a word in $\mathcal{W}$ of maximal length such that $x,y \in X_{w_0}$. Then, there exists distinct $i,j \in A$ such that $w_0i,w_0j \in \mathcal{W}$, $x\in X_{w_0i}$ and $y\in X_{w_0j}$. Set $k=|w_0|$. We consider the following two possible cases. Suppose first that $X_{w_0i}\cap X_{w_0j} = \emptyset$. Let $\gamma$ be the unique arc in $X$ with endpoints $x,y$. Note that $\gamma\subseteq X_{w_0}$ as $X_{w_0}$ is connected. Assuming $X_{w_0i}\cap X_{w_0j} = \emptyset$, it follows that $\gamma \setminus (X_{w_0i}\cup X_{w_0j})$ is a non-empty relatively open subset of $\gamma$. There must therefore exist some $l \in A\setminus\{i,j\}$ such that $\gamma \cap \partial X_{w_0l}$ contains two distinct points $v,v'$ of $\partial X_{w_0l}$. By the $1$-bounded turning property of $X$ and Lemma \[lem:tiles\](6), $$d(x,y) \geq {\mathop\mathrm{diam}\nolimits}{\g} \geq d(v,v') \geq K_3{\mathop\mathrm{diam}\nolimits}(X_{w_0l}).$$ On the other hand, $f(x),f(y) \in \mathcal{A}_{w_0}$ and so, by Lemma \[lem:dist\] and , we have: $$d_{\mathscr{C},\D}(f(x),f(y)) \leq {\mathop\mathrm{diam}\nolimits}{\mathcal{A}_{w_0}} \leq \D(w_0) \leq \frac{K_2}{K_1}{\mathop\mathrm{diam}\nolimits}(X_{w_0}).$$ Therefore, using Lemma \[lem:tiles\](3), $$d(x,y) \geq K_3{\mathop\mathrm{diam}\nolimits}(X_{w_0l}) \geq K_3 K_1 {\mathop\mathrm{diam}\nolimits}(X_{w_0}) \geq \frac{K_1^2 K_3}{K_2} d_{\mathscr{C},\D}(f(x),f(y)).$$ This completes the proof of in the case where $X_{w_0i}\cap X_{w_0j} = \emptyset$. Suppose now that $X_{w_0i}\cap X_{w_0j} \neq \emptyset$. Find words $w,u \in A^*$ of maximal lengths such that $w_0w, w_0u \in \mathcal{W}^*$, $x\in X_{w_0w}$, $y\in X_{w_0u}$ and $X_{w_0w}\cap X_{w_0u} \neq \emptyset$. Then there exist $w_0wi, w_0uj \in \mathcal{W}^*$ such that $X_{w_0wi}\cap X_{w_0u} = \emptyset$, $X_{w_0uj}\cap X_{w_0w} = \emptyset$, $x\in X_{w_0wi}$ and $y\in X_{w_0uj}$. Let $z$ be the unique point of $X_{w_0w}\cap X_{w_0u}$ and again set $\gamma$ to be the unique arc from $x$ to $y$ in $X$, which must pass through $z$. Choose $k\in A$ such that $z\in X_{w_0wk}$. Note that $k\neq i$ by the maximality of $w$, and that $z\in \partial X_{w_0 w k}$. The sub-arc of $\gamma$ from $x$ to $z$ must also contain a point $v\in \partial X_{w_0 w k}$ distinct from $z$, by Lemma \[lem:tiles\](4). Hence, again by $1$-bounded turning and Lemma \[lem:tiles\](6), we have $$d(x,z) \geq d(v,z) \geq K_3 {\mathop\mathrm{diam}\nolimits}(X_{w_0 w k}).$$ Similarly, $$d(y,z) \geq K_3 {\mathop\mathrm{diam}\nolimits}(X_{w_0 u l}),$$ for some $l\in A$. By the $1$-bounded turning property and Lemma \[lem:tiles\](3), $$d(x,y) \geq \frac{1}{2}(d(x,z)+d(y,z)) \geq \frac{1}{2}K_3({\mathop\mathrm{diam}\nolimits}(X_{w_0 w k}) + {\mathop\mathrm{diam}\nolimits}(X_{w_0 u l})) \geq \frac{1}{2}K_3K_1({\mathop\mathrm{diam}\nolimits}(X_{w_0 w}) + {\mathop\mathrm{diam}\nolimits}(X_{w_0 u})).$$ On the other hand, $f(x) \in \mathcal{A}_{w_0w}$, $f(y) \in \mathcal{A}_{w_0u}$ and $\{A^{\N}_{w_0w}, A^{\N}_{w_0u}\}$ is a chain joining $f(x)$ and $f(y)$. Therefore, $$d_{\mathscr{C},\D}(f(x),f(y)) \leq {\mathop\mathrm{diam}\nolimits}{\mathcal{A}_{w_0w}} + {\mathop\mathrm{diam}\nolimits}{\mathcal{A}_{w_0w}} \leq \D(w_0w) + \D(w_0u) \leq \frac{K_1}{K_2}({\mathop\mathrm{diam}\nolimits}(X_{w_0 w}) + {\mathop\mathrm{diam}\nolimits}(X_{w_0 u})).$$ Therefore, $$d_{\mathscr{C},\D}(f(x),f(y)) \leq \frac{2}{K_2 K_3} d(x,y).$$ This completes the proof of and hence of the proposition. Finally, to prove the “moreover” piece of Theorem \[thm:main\], we now show: The combinatorial data $\mathscr{C}$ and diameter function $\Delta$ defined above satisfy the conditions of Proposition \[prop:doubling\] for some choice of $N,n_0,c,\delta_1,\delta_2$ Property \[P1\] of Proposition \[prop:doubling\] follows from our choice of a finite alphabet $A=\{1,\dots,M\}$. Property \[P2\] follows from Lemma \[lem:tiledoubling\] and the definition of the combinatorial trees $T_k$ in our combinatorial data. Property \[P3\] is immediate from our construction of $\Delta$, with $\delta_1=K_1$ and $\delta_2=K_2$. It remains to verify Property \[P4\] of Proposition \[prop:doubling\]. Consider $u\in A^*$ and distinct $w_1, w_2\in\partial_{\mathscr{C}}A^\mathbb{N}_u$. Thus, for each $i\in\{1,2\}$ there is $v_i\in A^{|u|}\setminus\{u\}$ such that $w_i \in A^\N_u \wedge_{\mathscr{C}} A^\N_{v_i}$. For each $i\in\{1,2\}$, let $x_i\in X$ denote the point $$x_i = \cap_{n=0}^\infty X_{w_i(n)}.$$ By definition, we have $f(x_i) = w_i$. Notice that $x_1$ and $x_2$ are both in $X_u$ as $w_i\in A^\N_u$. We first claim that, for $i\in\{1,2\}$, $$\label{eq:boundarypoint} x_i \in X_{v_i} \cap X_u \subseteq \partial X_u.$$ As $w_i\in A^\N_u \wedge_{\mathscr{C}} A^\N_{v_i}$, we can find for each $n>|u|$ a word $v^i_n\in A^n_{v_i}$ such that $\{w_i(n), v^i_n\}\in E_n$. It follows from the definition of $\mathscr{C}$ that $$\emptyset \neq X_{w_i(n)} \cap X_{v^i_n} \subseteq X_{w_i(n)} \cap X_{v_i}$$ for all $n>|u|$. Hence, $${\mathop\mathrm{dist}\nolimits}(x_i, X_{v_i}) \leq {\mathop\mathrm{diam}\nolimits}(X_{w_i(n)}) \rightarrow 0 \text{ as } n\rightarrow\infty,$$ and so $x_i \in X_u \cap X_{v_i} \subseteq \partial X_u$. We next claim that $x_1\neq x_2$. Suppose to the contrary that $x_1=x_2=x$, and choose $n>|u|$ such that $w_1(n)\neq w_2(n)$. Then $X_{w_1(n)}$ and $X_{w_2(n)}$ are distinct subsets of $X_u$ with $x\in X_{w_1(n)} \cap X_{w_2(n)}$. In addition, we showed in that $x\in X_{v_1}$. It follows that there is an element $v\in A^n_{v_1}$ with $x\in X_{v}$. The word $v$, beginning as it does with $v_1\neq u$, is distinct from both $w_1(n)$ and $w_2(n)$, and so the three words $v$, $w_1(n)$, and $w_2$ are distinct and of the same length $n$. Moreover, $x\in X_{w_1(n)} \cap X_{w_2(n)} \cap X_v$. However, this contradicts Lemma \[lem:tiles\](4). Thus, $x_1$ and $x_2$ are distinct elements of $\partial X_u$. By Lemma \[lem:tiles\](6) and , $$d(x_1,x_2) \geq K_3 {\mathop\mathrm{diam}\nolimits}(X_u) \geq (K_3K_1/K_2) \Delta(u).$$ By Proposition \[prop:bilipschitz\], $f$ is bi-Lipschitz with constant depending only on $K_1, K_2, K_3$. Therefore $$d_{\mathscr{C},\Delta}([w_1],[w_2]) = d_{\mathscr{C},\Delta}(f(x_1),f(x_2)) \geq c \Delta(u),$$ for some $c$ depending only on $K_1, K_2, K_3$. This completes the proof. Bi-Lipschitz embedabbility of quasiconformal trees {#sec:embed} ================================================== This section is devoted to the proof of the following quantitative version of Theorem \[thm:mainembed\]. \[thm:mainembed2\] Let $X$ be a $C$-doubling, $c$-bounded turning tree. Assume that $\LL(X)$ admits an $L$-bi-Lipschitz embedding into some $\R^M$. Then $X$ admits an $L'$-bi-lipschitz embedding into some $\R^N$. Here $N$ and $L'$ depend only on $C$, $c$, $M$ and $L'$. The proof of Theorem \[thm:mainembed2\] consists of two steps. In Section \[sec:quasiarcs\] we prove the special case of embedabbility of quasi-arcs, i.e., quasiconformal trees in which the set of leaves consists of exactly two points. This is done in Proposition \[prop:qcircles\] below, which is a stronger version of Proposition \[prop:introqcircles\] from the introduction. Then, in Section \[sec:Seo-method\], we employ a bi-Lipschitz welding theorem of Lang and Plaut [@LP] and a characterization of metric spaces admitting bi-Lipschitz embedding into Euclidean spaces by Seo [@Seo] to complete the proof of Theorem \[thm:mainembed2\]. Bi-Lipschitz embeddability of quasi-arcs {#sec:quasiarcs} ---------------------------------------- The main result of this subsection is the following special case of Theorem \[thm:mainembed\] where the leaf set $\mathcal{L}(X)$ consists of only two points. In particular, this gives a detailed, sharp version of Proposition \[prop:introqcircles\]. We first introduce a piece of terminology: A metric space $X$ is *$(C,s)$-homogeneous*, for some $C,s\geq 0$, if every subset of diameter $d$ can be covered by at most $C\epsilon^{-s}$ sets of diameter at most $\epsilon d$. In particular, every doubling metric space is $(C,s)$-homogeneous for some $C$ and $s$ depending on the doubling constant [@Heinonen Section 10.13]. \[prop:qcircles\] Given $s\geq 1$, $C>0$ and $c\geq 1$, there exists $L=L(c,C,s)>1$ with the following property. If $\G = ([0,1],d)$ is $c$-bounded turning and $(C,s)$-homogeneous, then it is $L$-bi-Lipschitz embeddable in $\R^{\lfloor s \rfloor + 1}$. Proposition \[prop:qcircles\] generalizes Theorem C in [@HM], where it was assumed that $s<2$. We remark that the dimension $\lfloor s \rfloor + 1$ in Proposition \[prop:qcircles\] is sharp when $s>1$, in the sense that there exists a $1$-bounded turning, $(C,s)$-homogeneous metric $d$ on $[0,1]$ (namely the snowflaked Euclidean metric $|\cdot|^{1/s}$) such that $([0,1],d)$ can not be bi-Lipschitz embedded in $\R^{\lfloor s \rfloor}$. For the proof of Proposition \[prop:qcircles\], we may assume that ${\mathop\mathrm{diam}\nolimits}{\G} = 1$. The proof uses a construction of Herron and Meyer [@HM] and a bi-Lipschitz embedding method of Romney-Vellis [@RV] (see also [@BH] and [@Wu]). Let $M \in \{2,3,\dots\}$ and $A=\{1,\dots,M\}$. Let $\mathscr{C}_M=(A,(G_k)_{k\in\N})$ where for each $k\in\N$ the graph $G_k$ is a simple path with the following two rules: 1. For each $w\in A^*$ and $i\in\{1,\dots,M-1\}$ we have that $wi$ is adjacent to $wi'$, where $i'=i+1$. 2. If $wiv,wjv'\in A^*$ with $i<j$, $|v|=|v'|$ and $wiv$ is adjacent to $wjv'$, then $wivM$ is adjacent to $wjv'1$. In other words, each word in $A^k$ is simply adjacent to the following word in lexicographic order in $G_k$. As before, given a diameter function $\mathscr{D}\in\mathscr{D}(A)$, we write $\mathcal{A} = A^{\N}/\sim$ and for each $w\in A^*$, $\mathcal{A}_w = A^{\N}_w/\sim$. The following lemma summarizes some properties of this construction when the diameter function is chosen with certain parameters. \[lem:arcproperties\] Consider combinatorial data $\mathscr{C}_M$ as above, and let $\delta\in (M^{-1},1]$ and $\D \in \mathscr{D}(A,M^{-1},\d)$. Let $\mathcal{A}$ be the associated metric space. 1. Suppose $v,v'\in A^k$ with $v$ preceding $v'$ in lexicographic order. Then $\mathcal{A}_{v} \cap \mathcal{A}_{v'}\neq \emptyset$ if and only if $v$ and $v'$ are adjacent in $G_k$. 2. In case (1), $[vM^\infty]=[v'1^\infty]$ is the unique element of $\mathcal{A}_{v} \cap \mathcal{A}_{v'}\neq \emptyset$. 3. For each $v\in A^*$, the set $\mathcal{A}_v$ is a topological arc with $M^{-1}\Delta(v)\leq{\mathop\mathrm{diam}\nolimits}{\mathcal{A}_v}\leq \Delta(v)$. We begin with (1). Suppose $v,v'\in A^k$, with $v$ preceding $v'$ in lexicographic order, and $\mathcal{A}_{v} \cap \mathcal{A}_{v'}\neq \emptyset$. This means that there are infinite words $w,w'$ with $[vw]=[v'w']$. Suppose $v$ and $v'$ were not adjacent; let $u$ be a word on the simple path $T_k$ between them. Let $n\in\mathbb{N}$ be such that $\Delta(t)<\frac{1}{2}\Delta(u)$ for all $t\in A^n$. Each $t\in A^n_u$ is on the unique simple path between $(vw)(n)$ and $(v'w')(n)$ in $T_{n}$. By Lemma \[lem:tree1\], $[vw]$ and $[v'w']$ are both in $\mathcal{A}_{t}$ for each $t\in A^n_u$. In particular, all $\mathcal{A}_t$ for $t\in A^n_u$ share a common point. Therefore $${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_u) \leq 2\max\{{\mathop\mathrm{diam}\nolimits}(\mathcal{A}_{t}) : t\in A^n_u\} < \Delta(u).$$ On the other hand, our combinatorial data $\mathscr{C}_M$ satisfies the assumptions of Lemma \[lem:suffdoubling\]. Therefore ${\mathop\mathrm{diam}\nolimits}(\mathcal{A}_u)=\Delta(u)$, which is a contradiction. This proves the “forward direction” of (1). For the other direction, it is immediate from the construction of $\mathscr{C}_M$ that if $v$ and $v'$ are adjacent in $T_k$, with $v$ lexicographically preceding, then for each $n\in\mathbb{N}$ $$d_{\mathscr{C},\Delta}(vM^\infty,v'1^\infty) \leq \Delta(vM^n) + \Delta(v'1^n) \rightarrow 0 \text{ as } n\rightarrow \infty,$$ and so $[vM^\infty]=[v'1^\infty] \in \mathcal{A}_v \cap \mathcal{A}_{v'}$. For (2), suppose there was a point $p$ other than $[vM^\infty]=[v'1^\infty]$ in $\mathcal{A}_v \cap \mathcal{A}_{v'}$. Then there would be an infinite word $w\in A^\mathbb{N}$, $w\neq M^\infty$, such that $p=[vw]$. Choose $n$ such that the $n$th letter of $w$ is not $M$. Then $vw(n)$ and $v'1^n$ are not adjacent in $T_{k+n}$, but $[vw]\in\mathcal{A}_{vw(n)} \cap \mathcal{A}_{v'1^n}$. This contradicts (1). For fact (3), it is an immediate consequence of Remark \[rem:arc\] that each $\mathcal{A}_v$ is a topological arc. The diameter of $\mathcal{A}_v$ is at most $\D(v)$ by Lemma \[lem:dist\]. If $v$ has at least two neighbors in $T_{|v|}$, then ${\mathop\mathrm{diam}\nolimits}(T_{|v|})=\Delta(v)$ be Lemma \[lem:suffdoubling\]. Otherwise, $vi$ has at least two neighbors in $T_{|v|+1}$ for some $i\in A$, and so $${\mathop\mathrm{diam}\nolimits}{\mathcal{A}_v} \geq {\mathop\mathrm{diam}\nolimits}{\mathcal{A}_{vi}} \geq \Delta(vi) \geq M^{-1} \Delta(v).\qedhere$$ If $d\in (M^{-1},1)$ and $\D \in \mathscr{D}(A,M^{-1},\d)$, then the space $(\mathcal{A},d_{\mathscr{C}_M,\D})$ is $s$-homogeneous with $s=\log(M)/\log(1/\d)$. The following result can be obtained following the arguments of Theorem B of [@HM] essentially verbatim; we provide a brief reference to the necessary arguments. \[lem:HM\] Let $s\geq 1$, $c\geq 1$, and $\G$ a $c$-bounded turning and $s$-homogeneous metric arc with ${\mathop\mathrm{diam}\nolimits}{G}=1$. Then for any $M\in\{2,3,\dots\}$ and any $\d \in (M^{-1/s},1)$, there exists $\D \in \mathscr{D}(A,M^{-1},\d)$ and an $L$-bi-Lipschitz homeomorphism $f : \G \to (\mathcal{A},d_{\mathscr{C}_M,\D})$. The constant $L$ depends only on $c$, $s$, and $M$. Following exactly the procedure on [@HM p. 622], we divide $\G$ into $M$ sub-arcs of equal diameter, then iterate this procedure on each sub-arc. Letting $\mathscr{C}_M = (A=\{1, \dots, M\}, G_k)$ as above, this yields an assignment to each element $w\in A^*$ of an arc $\gamma_w\subseteq\G$, with nesting and adjacency properties reflecting that of $\mathscr{C}_M$ and $\sup_{w\in A^k} {\mathop\mathrm{diam}\nolimits}(\gamma_w) \rightarrow 0$ as $k\rightarrow\infty$. The argument on [@HM p. 622-623] provides a diameter function $\D\in\mathscr{D}(A,M^{-1},\d)$ such that $$\D(w) \approx {\mathop\mathrm{diam}\nolimits}(\gamma_w),$$ with implied constant depending only on $c$, $s$, and $M$. Defining $F\colon {\mathcal{A}}\rightarrow \G$ by $F([w]) = \cap_{k=1}^\infty \gamma_{w(k)}$, we see exactly as in Lemma \[lem:welldef\] and Proposition \[prop:bilipschitz\] of the present paper that $F$ is well-defined, surjective, and bi-Lipschitz. Taking $f=F^{-1}$ completes the proof. We now fix parameters $M$ and $\delta$ that will enable us to use a construction from [@RV]. Given $s\geq 1$, let - $n$ be the minimal integer satisfying $n>(\lfloor s \rfloor + 1 -s)^{-1}$, - $p=\lfloor s \rfloor - 1 + \frac{n-1}{n} = \lfloor s \rfloor - \frac{1}{n}>0$, - $M_0 = 9^{n(\lfloor s \rfloor + 1)}$, - $M = M_0^{1+p}$, and - $\delta = M_0^{-1}$. The above parameters all depend on $s$, but we suppress this in the notation. Observe that $\delta > M^{-1/s} \geq M^{-1}$ in all cases, and in fact $\delta$ is an integer multiple of $M^{-1}$. Only $\delta$ and $M$ will play a direct role below. Given Proposition \[lem:HM\], the proof of Proposition \[prop:qcircles\] now reduces to the following lemma. \[lem:snowembed\] Let $s\geq 1$ and choose $M$ and $\delta$ as above. Let $\D\in \mathscr{D}(A,M^{-1},\delta)$. Then there is a bi-Lipschitz embedding of $(\mathcal{A},d_{\mathscr{C}_M,\D})$ into $\R^{\lfloor s \rfloor + 1}$ with bi-Lipschitz constant depending only on $M$, $\d$ and $s$, and thus only on $s$. The construction of the embedding follows ideas and notation from [@RV]. We fix parameters $M$ and $\delta$ as in the statement of Lemma \[lem:snowembed\] and write $\mathscr{C}=\mathscr{C}_M$. We also fix $\D\in \mathscr{D}(A,M^{-1},\delta)$ for the remainder of this subsection. Let $$\begin{aligned} I &= [0,1]\times\{0\}^{\lfloor s \rfloor}\\ L&=\left( (\{0\}\times[0,1/2])\cup([0,1/2]\times\{1/2\})\right)\times\{0\}^{\lfloor s \rfloor-1}\end{aligned}$$ with the convention that $E\times\{0\}^0 = E$. An $I$-segment (resp. $L$-segment) is the image of $I$ (resp. $L$) under a similarity mapping of $\R^{\lfloor s \rfloor + 1}$, and is parallel to the coordinate axes. Given an $I$- or $L$- segment $\tau$ with length $\ell$ and endpoints $x^*,y^*$, we define the *cubic thickening* $Q(\tau)$ of $\tau$ to be the union of all closed cubes parallel to coordinate axes, of side length $(1-2\delta)\ell$ and centered on points $z\in\tau$ such that $$\max\{|z-x^*|,|z-y^*|\} \geq \ell(1-2\delta)/2.$$ Define also $\mathcal{C}(\tau)$ to be the closed cube which is parallel to coordinate axes, has side length $\ell$, and is centered on the midpoint of $\tau$. The intersection $Q(\tau) \cap \partial\mathcal{C}(\tau)$ has exactly two components which we call the *entrances* of $Q(\tau)$. For each $\tau\in\{I,L\}$ we define two polygonal arcs $\mathcal{J}(\tau)$ and $\mathcal{J}_0(\tau)$ in the following lemma. \[lem:RV\] Given $\tau \in \{I,L\}$ there exist two polygonal arcs $\mathcal{J}(\tau)$ and $\mathcal{J}_0(\tau)$, each contained in $Q(\tau)$, whose endpoints are the same as those of $\tau$ and that satisfy the following properties. 1. The arcs $\mathcal{J}(\tau), \mathcal{J}_0(\tau)$ consist of $M$-many $I$-segments and $L$-segments $\sigma_{i}$, $i\in \{1,\dots,M\}$, labeled according to their order in $\mathcal{J}(\tau)$ with $\sigma_1$ containing the origin. Each $\sigma_i$ in $\mathcal{J}(\tau)$ has length $\delta$ and each $\sigma_i$ in $\mathcal{J}_0(\tau)$ has length $M^{-1}$. 2. The segments $\sigma_{1}$ and $\sigma_{M}$ are $I$-segments. 3. For all $i\in \{1,\dots,M-1\}$, $Q(\sigma_i)\cap Q(\sigma_{i+1})$ is an entrance of $Q(\sigma_i)$ and an entrance of $Q(\sigma_{i+1})$. If $i,j \in \{1,\dots,M\}$, with $|i-j|>1$, then $Q(\sigma_i)\cap Q(\sigma_j) = \emptyset$. 4. If $E_1,E_2$ are the entrances of $Q(\tau)$, then an entrance of $Q(\sigma_1)$ is contained in $E_1$ and an entrance of $Q(\sigma_M)$ is contained in $E_2$. Moreover, for any $i\in\{2,\dots,M-1\}$, $Q(\sigma_i)\cap \partial Q(\tau) = \emptyset$. The constructions of $\mathcal{J}_0(I)$ and $\mathcal{J}_0(L)$ are quite simple. Write $I = \bigcup_{m=1}^{M}\sigma_m$ with $$\sigma_m =\left[\frac{m-1}{M},\frac{m}{M}\right]\times\{0\}^{\lfloor s \rfloor} \subset \R^{\lfloor s \rfloor + 1}$$ and set $\mathcal{J}_{0}(I) = \bigcup_{m=1}^M \sigma_m = I$. Similarly write $L = \bigcup_{m=1}^{M}\sigma_m$ where $\sigma_{m}$ is an $L$-segment if $m=\frac{M+1}{2}$ and an $I$-segment otherwise and each $\sigma_m$ has length $1/M$. Set $\mathcal{J}_{0}(L) = \bigcup_{m=1}^M \sigma_m$. The constructions of $\mathcal{J}(I)$ and $\mathcal{J}(L)$ are more complicated and can be found in [@RV 6.1, 6.2] (where they are denoted as $J_I(N,n)$ and $J_L(N,n)$, respectively). Without describing the construction, we briefly explain how our parameters match with those of [@RV]. The parameter $N$ appearing on [@RV p. 4] matches our $\lfloor s \rfloor -1$. Our parameters $p$ and $n$ match the ones given there. Our parameter $M_0$ corresponds to $M$ on [@RV p. 4], and our parameter $M$ corresponds to $M^{1+p}$ on [@RV p. 5]. We record a few more simple consequences of properties (J1)–(J4). \[lem:pipelines\] Consider $\tau \in \{I,L\}$, $J \in \{\mathcal{J}(\tau) ,\mathcal{J}_0(\tau)\}$. Recall that $J$ is a union of sets $\{\sigma_i\}_{i=1}^M$, each of which is an $I$-segment or $L$-segment. Then: 1. For each $i\in \{1, \dots, M\}$, $Q(\sigma_i) \subset Q(\tau)$. 2. For each $i\in\{2,\dots,M-1\}$, $${\mathop\mathrm{dist}\nolimits}(Q(\sigma_i),\partial Q(\tau)) \geq M^{-2} .$$ 3. If $i,j \in \{1,\dots,M\}$ with $|i-j|>1$, then, $${\mathop\mathrm{dist}\nolimits}(Q(\sigma_i), Q(\sigma_j)) \geq M^{-2} .$$ 4. Let $E$ be the entrance of $Q(\tau)$ that contains an endpoint of $\sigma_1$ (resp. endpoint of $\sigma_M$) and let $P$ be the $\lfloor s \rfloor$-dimensional plane that contains $E$. Then for all $i\in\{2,\dots,M\}$ (resp. $i\in\{1,\dots,M-1\}$) $${\mathop\mathrm{dist}\nolimits}(Q(\sigma_i),P) \geq M^{-2}.$$ All four statements are obvious in the case $J=\mathcal{J}_0(\tau)$, so we now assume that $J=\mathcal{J}(\tau)$. Statement (1) is an immediate consequence of the fact that $J\subseteq Q(\tau)$ and property (J4) of Lemma \[lem:RV\]. For the remaining three properties, it is useful to first observe that, since $\delta$ is an integer multiple of $M^{-1}$, the sets $Q(\tau)$ and $Q(\sigma_i)$ are each unions of axis-parallel cubes whose vertices lie on the $M^{-2}$-scale grid $M^{-2}\mathbb{Z}^{\lfloor s \rfloor + 1}.$ Statements (2) and (4) follow immediately from this observation and (J4). Statement (3) follows immediately from this observation and (J3). We now use Lemma \[lem:RV\] to construct arcs in $\R^{\lfloor s \rfloor +1}$ that mimic the metric properties of the combinatorial construction $\mathscr{C}, \D$ fixed below the statement of Lemma \[lem:snowembed\]. \[lem:defnoftau\] For each $w\in A^*$ there exists an $I$- or $L$-segment $\tau_w$ with the following properties. 1. If $w,u \in A^k$ are adjacent, then $\tau_w$ and $\tau_u$ intersect at an endpoint while $Q(\tau_w)\cap Q(\tau_u)$ is contained in an entrance of $Q(\tau_w)$ and an entrance of $Q(\tau_u)$. If $w,u \in A^k$ are distinct but not adjacent, then $Q(\tau_w)\cap Q(\tau_u)$ and $\tau_w\cap\tau_u$ are empty. 2. For any $w\in A^*$, there exists $\tau\in \{I,L\}$ such that $\tau_w$ and $Q(\tau_w)$ are scaled copies of $\tau$ and $Q(\tau)$, respectively, by a factor of $\D(w)$. The construction is done in an inductive manner. Let $\tau_{\varepsilon}:=I \subset \R^{\lfloor s \rfloor + 1}$. Property (1) of the lemma is vacuous in this base case, while property (2) is immediate. Assume now that for some integer $k\geq 0$ we have defined $I$- and $L$-segments $\tau_w$ (for all $j\leq k$ and $w\in A^j$) satisfying the properties of the lemma. Fix $w\in A^{k}$, and let $u$ be the preceding vertex of $A^k$ in lexicographic order, assuming for the moment that such a vertex exists. Let $E$ be the entrance of $Q(\tau_w)$ that intersects an entrance of $Q(\tau_u)$. Suppose that $\tau_w$ is a rescaled copy of $\tau \in \{I,L\}$. Let $\phi_w : \R^{\lfloor s\rfloor + 1} \to \R^{\lfloor s\rfloor + 1}$ be a similarity map such that $Q(\tau)$ is mapped onto $Q(\tau_w)$, the entrance of $Q(\tau)$ that contains the origin is mapped onto the entrance of $Q(\tau_w)$ that contains $Q(\tau_w)\cap Q(\tau_u)$, and the other entrance of $Q(\tau)$ is mapped to the other entrance of $Q(\tau_w)$. If there is no $u\in A^k$ preceding $w$ in lexicographic order, then $w=1^k$ for some $k\geq 0$. In that case, if $k=0$ we set $\phi_w$ to be the identity, and if $k\geq 1$ we set $u=1^{k-1}2$ and do the analogous construction of $\phi_w$ to arrange that the entrance of $Q(\tau)$ that does not contain the origin is mapped onto the entrance of $Q(\tau_w)$ that contains $Q(\tau_w)\cap Q(\tau_u)$. We now define $\tau_{wi}$ for each $i\in A$: - If $\D(w1)=M^{-1}\D(w)$, then for each $i\in A$ set $\tau_{wi} = \phi_w(\sigma_i)$ where $\sigma_i \subset\mathcal{J}_0(\tau)$. - If $\D(w1) = \delta\D(w)$, then for each $i\in A$ set $\tau_{wi} = \phi_w(\sigma_i)$ where $\sigma_i \subset\mathcal{J}(\tau)$. This completes the definition of the arcs $\tau_{w}$ for all $w\in A^{k+1}$. We now prove that the family $\{\tau_w : w\in A^{k+1}\}$ satisfies properties (1) and (2) of the lemma. For property (2) of the lemma, by design, and the inductive hypothesis (2), for all $i\in A$ $${\mathop\mathrm{diam}\nolimits}{\tau_{wi}} = \D(w){\mathop\mathrm{diam}\nolimits}{\sigma_i} = \D(wi){\mathop\mathrm{diam}\nolimits}{Q(\tau')}$$ where $\tau' \in \{I,L\}$. Therefore, property (2) holds for $k+1$. We now turn to the proof of (1). Let $w \in A^{k}$ and $i\in A$. Let also $u\in A^k$ and $j\in A$. We consider two cases. *Case 1.* Assume that $w=u$ and $i\neq j$. If $wi$ is adjacent to $wj$, then by design of paths $\mathcal{J}(\tau)$ and $\mathcal{J}_0(\tau)$, we have that $\tau_{wi}$ and $\tau_{wj}$ share an endpoint and by (J3) $Q(\tau_{wi})\cap Q(\tau_{wj})$ is a common entrance of $Q(\tau_{wi})$ and $Q(\tau_{wj})$. If $wi$ is not adjacent to $wj$, then again by (J3) $Q(\tau_{wi})\cap Q(\tau_{wj}) = \emptyset$ which also implies that $\tau_{wi}\cap \tau_{wj} = \emptyset$. *Case 2.* Assume that $u\neq w$. The proof splits in two subcases. *Case 2.1*. Assume that $i\not\in\{1,M\}$. Then $wi$ is not adjacent to $uj$ and by (J4) $Q(\tau_{wi})$ is contained in the interior of $Q(\tau_{w})$ which is disjoint from $Q_{u}$ by the inductive hypothesis. Therefore, $Q(\tau_{wi})\cap Q(\tau_{uj})$ and $\tau_{wi}\cap\tau_{uj}$ are both empty. *Case 2.2* Assume that $i\in\{1,M\}$. Without loss of generality, we assume that $i=1$; the case $i=M$ is similar. By design $Q(\tau_{wi})$ intersects one entrance of $Q(\tau_w)$ but not the other. Therefore, if $u$ is not adjacent to $w$ or if it is adjacent to $w$ but is preceded by $w$, then the inductive hypothesis implies that $Q(\tau_{w1})\cap Q(\tau_{uj})$ and $\tau_{w1}\cap\tau_{uj}$ are both empty. Assume now that $u$ is adjacent to $w$ and precedes $w$. Then, the only $j\in A$ for which $Q(\tau_{uj})$ intersects the entrance of $Q(\tau_u)$ which contains $Q(\tau_{w})\cap Q(\tau_{u})$ is $j=M$. In this case, $\tau_{uM}\cap \tau_{w1}$ is the common endpoint of $\tau_w$ and $\tau_u$. Therefore, $Q(\tau_{wi})\cap Q(\tau_{uj})$ is nonempty and is contained in an entrance of $Q(\tau_{wi})$ and an entrance of $Q(\tau_{uj})$. Lemma \[lem:defnoftau\](2) implies that for all $w\in A^*$, $$\label{eq:est1} 2^{-1/2}\D(w) \leq {\mathop\mathrm{diam}\nolimits}{\tau_w} \leq \D(w).$$ \[lem:pipes2\] Let $w,u \in A^k$ be adjacent words, with $w$ preceding $u$ in lexicographic order. If $i\in A\setminus \{M\}$ or if $j\in A\setminus \{1\}$, then $${\mathop\mathrm{dist}\nolimits}(Q(\tau_{wi}),Q(\tau_{uj})) \gtrsim_s \max\{\D(w),\D(u)\}.$$ Set $E = Q(\tau_{w})\cap Q(\tau_u)$. By Lemma \[lem:defnoftau\], $E$ is contained in an entrance of $Q(\tau_w)$ and in an entrance of $Q(\tau_u)$. Let $P$ be the $\lfloor s\rfloor$-dimensional plane in $\R^{\lfloor s\rfloor+1}$ that contains $E$. Then, $P$ separates the interior of $Q(\tau_{wi})$ from the interior $Q(\tau_{uj})$. By Lemma \[lem:pipelines\], $${\mathop\mathrm{dist}\nolimits}(Q(\tau_{wi}),Q(\tau_{uj})) \geq \max\{{\mathop\mathrm{dist}\nolimits}(Q(\tau_{wi}),P), {\mathop\mathrm{dist}\nolimits}(Q(\tau_{uj}),P)\} \gtrsim_s \max\{\D(w),\D(u)\}. \qedhere$$ For each $w\in A^*$ and $k\geq |w|$, set $$\mathcal{Q}^{(k)}_w := \bigcup_{u \in A^{k}_w}Q(\tau_u),\quad \mathcal{Q}^{(k)} := \bigcup_{u\in A^k}Q(\tau_u),\quad \mathcal{Q}_w := \bigcap_{n\geq |w|}\mathcal{Q}^{(n)}_w.$$ By , if $w \in A^{\N}$, then $\lim_{n\to\infty}{\mathop\mathrm{diam}\nolimits}{Q(\tau_{w(n)})} \leq \lim_{n\to\infty}(\lfloor s \rfloor + 1)^{1/2}\delta^{n} = 0$. For each $w\in A^{\N}$ denote by $x_w$ the unique point $$\{x_w\} := \bigcap_{n\in\N} Q(\tau_{w(n)}) = \bigcap_{n\in\N}\mathcal{Q}_{w(n)}.$$ Define a map $F : (\mathcal{A},d_{\mathscr{C},\D}) \to \mathcal{Q}_{\varepsilon} \subset\R^{\lfloor s \rfloor + 1}$ by $F([w]) =x_w$. $F$ is well-defined, and $F(\mathcal{A}_w) = \mathcal{Q}_w$ for all $w\in A^*$. Let $[w]=[v]\in\mathcal{A}$, with $w\neq v$. By Lemma \[lem:arcproperties\], there is a $n\in\mathbb{N}$ and $u,u'$ adjacent in $A^n$ such that $w=uM^\infty$ and $v=u'1^\infty$ (or vice versa). For each $n\in\N$, $Q(\tau_{uM^n})$ intersects with $Q(\tau_{u'1^n})$ on a common entrance. Denote by $p$ the unique point in $\bigcap_{n\in\N}(Q(\tau_{uM^n})\cap Q(\tau_{u'1^n}))$. Then $Q(\tau_{w(k)})$ and $Q(\tau_{v(k)})$ both contain $p$ for all $k$, and hence $F([v])=F([w])=p$. So $F$ is well-defined. For the second part, fix $n\in\mathbb{N}$ and $w\in A^n$. For $k\geq n$, note that $\{\mathcal{Q}_w^{(k)}\}$ converges in Hausdorff distance to $\mathcal{Q}_w$. By construction, each point of $F([\mathcal{A}_w])$ is contained in the Hausdorff limit of the sets $\mathcal{Q}_w^{(k)}$, and hence in $\mathcal{Q}_w$. Thus, $F(\mathcal{A}_w) \subseteq \mathcal{Q}_w$. For the other inclusion, fix $p\in \mathcal{Q}_w$. Let $v_0=w$. For each $k\geq 1$, we inductively set $v_k\in A^{|w|+k}_{v_{k-1}} \subseteq A^{|w|+k}_w$ to be a word with $p\in \mathcal{Q}_{v_k}$. Let $v$ be the infinite word such that $v(|w|+k)=v_k$ for all $k\geq 0$. Then immediately $p=F([v])$. Therefore, $\mathcal{Q}_w\subseteq F(\mathcal{A}_w)$. It remains to show now that $F$ is $L$-bi-Lipschitz with $L$ depending only on $s$. Fix distinct $[w],[w']\in \mathcal{A}$. Without loss of generality, assume that $w$ precedes $w'$ in lexicographic order. Let $\sigma$ be the unique arc in $\mathcal{A}$ whose endpoints are $[w]$ and $[w']$. Let also $w_0\in A^*$ be the longest word such that $[w],[w']\in \mathcal{A}_{w_0}$. Let also $i,j \in A$ such that $[w]\in \mathcal{A}_{w_0i}$ and $[w']\in \mathcal{A}_{w_0j}$. By maximality of $w_0$ we have that $i\neq j$. We consider the following possible two cases. *Case 1.* Suppose that $|i-j|>1$. On one hand, there exists $i' \in A$ such that $\mathcal{A}_{w_0i'} \subset \sigma$ which implies that $$M^{-1}\D(w_0) \leq \D(w_0i) \leq {\mathop\mathrm{diam}\nolimits}{\sigma} = d_{\mathscr{C},\D}([w],[w']) \leq \D(w_0).$$ On the other hand, $F([w]) \in Q(\tau_{w_0i})$, $F([w']) \in Q(\tau_{w_0j})$ and by Lemma \[lem:pipelines\], $$M^{-2}\D(w_0) \leq {\mathop\mathrm{dist}\nolimits}(Q(\tau_{w_0i}),Q(\tau_{w_0j})) \leq |F([w])-F([w'])| \leq {\mathop\mathrm{diam}\nolimits}{Q(\tau_{w_0})} \leq (\lfloor s\rfloor +1)^{1/2}\D(w_0).$$ Therefore, $d_{\mathscr{C},\D}([w],[w']) \approx_{s} \D(w_0) \approx_s |F([w])-F([w'])|$. This completes the proof in Case 1. *Case 2.* Suppose that $|i-j|=1$. Without loss of generality, assume that $j=i+1$. Let $k$ and $l$ be the unique integers such that $$\mathcal{A}_{w_0iM^k}\cup \mathcal{A}_{w_0j1^l} \subset \sigma \subset \mathcal{A}_{w_0iM^{k-1}}\cup \mathcal{A}_{w_0j1^{l-1}}.$$ Let also $i',j'\in A$ such that $[w] \in \mathcal{A}_{w_0iM^{k-1}i'}$ and $[w'] \in \mathcal{A}_{w_0j1^{l-1}j'}$. Note that $i' \neq M$ while $j' \neq 1$. On one hand, using the $1$-bounded turning property of $({\mathcal{A}}, d_{\mathscr{C},\D})$ and Lemma \[lem:arcproperties\], we have $$\begin{aligned} \max\{\D(w_0iM^k), \D(w_0j1^l)\} &\leq M d_{\mathscr{C},\D}([w],[w'])\\ &\leq M {\mathop\mathrm{diam}\nolimits}(\mathcal{A}_{w_0iM^{k-1}}\cup \mathcal{A}_{w_0j1^{l-1}})\\ &\leq 2M\max\{ \D(w_0iM^{k-1}) , \D(w_0j1^{l-1})\} \\ &\leq 2M^2\max\{\D(w_0iM^k), \D(w_0j1^l)\}.\end{aligned}$$ On the other hand, by Lemma \[lem:pipes2\], $$\begin{aligned} |F([w])-F([w'])| &\lesssim \max\{{\mathop\mathrm{diam}\nolimits}{Q(\tau_{w_0iM^{k-1}})}, {\mathop\mathrm{diam}\nolimits}{ Q(\tau_{w_0j1^{l-1}})}\}\\ &\lesssim \max\{\D(w_0iM^{k-1}), \D(w_0j1^{l-1})\}\\ &\lesssim_s {\mathop\mathrm{dist}\nolimits}( Q(\tau_{w_0iM^{k-1}i'}), Q(\tau_{w_0iM^{k-1}i'}))\\ &\leq |F([w])-F([w'])|.\end{aligned}$$ Therefore, $$|F([w])-F([w'])| \approx_s \max\{\D(w_0iM^{k-1}), \D(w_0j1^{l-1})\} \approx_{s} d_{\mathscr{C},\D}([w],[w']).$$ This completes the proof in Case 2 and the proof of the lemma. Proof of Theorem \[thm:mainembed2\] {#sec:Seo-method} ----------------------------------- Here we prove Theorem \[thm:mainembed\] using two bi-Lipschitz embedding results of Lang and Plaut [@LP] and of Seo [@Seo]. The first result says that one can “glue" two bi-Lipschitz embeddings into a single embedding. \[thm:welding\] Let $X$ be a metric space and let $X_1, X_2 \subset X$ be closed subsets such that $X = X_1 \cup X_2$. If $X_1$ $L_1$-bi-Lipschitz embeds in $\R^{n_1}$ and $X_2$ $L_2$-bi-Lipschitz embeds in $\R^{n_2}$, then $X$ $L$-bi-Lipschitz embeds in $\R^{n_1+n_2+1}$ with $L$ depending on $L_1$, $L_2$, $n_1$ and $n_2$. Using Theorem \[thm:welding\] we show that balls of $X$ that are appropriately far from $\LL(X)$ admit a bi-Lipschitz embedding into some Euclidean space quantitatively. \[lem:Whitneycube\] Let $X$ be a doubling, bounded turning tree. For every $0<\beta<1$, there exist $L$ and $N$ depending only on the doubling constant of $X$, the bounded turning constant of $X$, and $\beta$ such that if $B(x,r)$ is a ball with $x\in X\setminus \LL(X)$ and $ r < \beta{\mathop\mathrm{dist}\nolimits}(x,\LL(X))$, then $B(x,r)$ admits an $L$-bi-Lipschitz embedding into $\R^{N}$. Fix $0<\beta<1$. Let $B=\overline{B}(x,r)$ be a ball with $x\in X\setminus\LL(X)$ and $ r < \beta{\mathop\mathrm{dist}\nolimits}(x,\LL(X))$. Let $D$ denote the doubling constant of $X$ and $H$ the bounded turning constant. We will argue that $B$ is contained in a union of at most $K=K(\beta, D,H)$ quasi-arcs. By Proposition \[prop:qcircles\] and Theorem \[thm:welding\], the latter implies that $B$ admits an $L$-bi-Lipschitz embedding into $\R^{N}$ with $N$ and $L$ depending only on $K$ and $D$, hence only on $\beta$, $D$ and $H$. Let $\Gamma$ be the collection of all arcs in $X$ that join $x$ to a leaf of $X$. For each $\gamma\in \Gamma$, parametrize it by a continuous $\gamma\colon [0,1]\rightarrow X$ such that $\gamma(0)=x$ and $\gamma(1)\in\LL(X)$. Let $x_\gamma=\gamma(t_\gamma)$, where $$t_\gamma = \sup \{t\in [0,1] : \gamma(t)\in B\}.$$ In other words, $x_\gamma$ is the “last” point on $\gamma$ contained in $B$. Similarly, let $y_\gamma$ denote the last point on $\gamma$ contained in $\overline{B}(x,r/\beta)$. Note that $B$ and $\overline{B}(x,r/\beta)$ are disjoint from $\LL(X)$ by assumption, so the points $x_\gamma$ and $y_\gamma$ must exist for each $\gamma\in\Gamma$. Two properties of these points are clear: 1. If $x_\gamma \neq x_{\gamma'}$, then $y_\gamma \neq y_{\gamma'}$. In particular, $$\label{eq:card} \text{card}{\{x_{\g} : \g \in \G\}} \leq \text{card}{\{y_{\g} : \g \in \G\}}.$$ 2. We have $d(x_\gamma, x)=r$ and $d(y_\gamma,x)=r/\beta$ for each $\gamma\in \Gamma$. Finally, let $\Gamma_0$ be the collection of arcs joining $x$ to $x_\gamma$, as $\gamma$ ranges in $\Gamma$. We will show that $\Gamma_0$ contains a controlled finite number of distinct elements, by showing that the collection $\{x_\gamma : \gamma\in \Gamma\}$ contains a controlled number of distinct elements. Since $B$ is contained in the union of all arcs of $\Gamma_0$, this will complete the proof. Suppose $\gamma, \gamma'\in \Gamma$ have $x_\gamma \neq x_{\gamma'}$. We then claim that $$d(y_\gamma, y_{\gamma'}) \geq \eta r,$$ for some constant $\eta$ depending only on $D$ and $H$. Indeed, the arc $[y_\gamma, y_{\gamma'}]$ must contain $x_\gamma$, and hence its diameter is at least $$d(y_\gamma, x_\gamma) \geq \left(\frac{1}{\beta} - 1\right)r,$$ and so $$d(y_\gamma, y_{\gamma'}) \geq \frac{1}{H}{\mathop\mathrm{diam}\nolimits}([y_\gamma, y_{\gamma'}]) \geq \frac{1}{H} \left(\frac{1}{\beta} - 1\right)r = \eta r.$$ The total number of different arcs in $\Gamma_0$ is controlled by the total number of distinct $x_\gamma$, which is controlled by $\text{card}\{y_\gamma: \gamma\in\Gamma\}$ by . The points $y_\gamma$ form an $\eta r$-separated set in $B(x,r/\beta)$, and so the cardinality of this set is bounded by a constant $K$ depending only on $\eta$, $\beta$, and the doubling constant $D$. The second bi-Lipschitz embedding result that we need is Seo’s general bi-Lipschitz embeddability criterion [@Seo]. In fact, we use a simplified version of Seo’s result presented by Romney in [@Romney Theorem 2.2]. Before stating the result we recall a generalized notion of Whitney decomposition for metric measure spaces due to Christ [@Christ] and Seo [@Seo]. \[def:ChristWhitney\] Let $(X,d,\mu)$ be a metric measure space and let $\Omega$ be an open proper subset of $X$. A collection $\mathscr{Q}$ of open subsets of $\Omega$ is a *Christ-Whitney decomposition* of $\Omega$ if there exist constants $\d\in(0,1)$, $C_1>c_0>0$, and $a\geq 4$ such that the following properties are satisfied: 1. $\bigcup_{Q\in\mathscr{Q}} Q$ is dense in $\Omega$. 2. For every $Q,Q' \in \mathscr{Q}$ with $Q\neq Q'$ we have $Q\cap Q' = \emptyset$. 3. For every $Q\in \mathscr{Q}$, there exists $x\in \Omega$ and $k\in\mathbb{Z}$ such that $$B(x,c_0 \d^k) \subset Q \subset B(x,C_1\d^k).$$ 4. For every $Q\in \mathscr{Q}$, $$(a-2)C_1\delta^k \leq {\mathop\mathrm{dist}\nolimits}(Q,X\setminus\Omega) \leq \left(\frac{a C_1}{\delta}\right)\delta^k.$$ \[lem:CWdecomp\] Let $X$ be a doubling metric space and $Y$ be a nonempty closed proper subset of $X$. Then $X\setminus Y$ has a Christ-Whitney decomposition, with constants $\delta, c_0, C_1, a$ absolute. \[thm:seo\] Let $X$ be a complete metric space. Then $X$ admits an $L$-bi-Lipschitz embedding into some Euclidean space $\R^M$ if and only if the following conditions hold for some constants $L_1, L_2, M_1, M_2$: 1. \[eq:seo1\] $X$ is doubling. 2. \[eq:seo2\] There is a non-empty closed subset of $Y\subseteq X$ which admits an $L_1$-bi-Lipschitz embedding into some $\R^{M_1}$. 3. \[eq:seo3\] There is a Christ-Whitney decomposition of $X\setminus Y$ such that each cube admits an $L_2$-bi-Lipschitz embedding into some $\R^{M_2}$. The distortion $L$ and target dimension $M$ of the embedding of $X$ depend only on the doubling constant of $\mu$, $M_1$, $M_2$, and $L_1$, $L_2$. It suffices to show that $X$ satisfies the conditions of Theorem \[thm:seo\] with $Y = \overline{\LL(X)}$. The doubling property in Theorem \[thm:seo\] is satisfied by assumption. We assume that $\LL(X)$, hence $Y$, admits a bi-Lipschitz embedding into some $\R^{M_1}$, so is assumed to hold in Theorem \[thm:main\]. It remains to prove . By Lemma \[lem:CWdecomp\] there exists a Christ-Whitney decomposition $\mathscr{Q}$ for some constants $\d\in(0,1)$, $C_1>c_0>0$, and $a\geq 4$. Let $Q\in\mathscr{Q}$ be an arbitrary cube of this decomposition. The doubling property of $X$ implies that there exists $N\in\N$, depending only on the doubling constant of $X$ and the constants of the Christ-Whitney decomposition, and at most $N$ balls $B_1,\dots,B_n$ with centers on $Q$ and of radius $\frac{1}{3}{\mathop\mathrm{dist}\nolimits}(Q,Y)$, such that $Q\subset B_1\cup\cdots\cup B_n$. In particular, the balls $B_i$ each satisfy the assumptions of Lemma \[lem:Whitneycube\] with $\beta=\frac{1}{2}$. Thus, by Lemma \[lem:Whitneycube\], each $B_i$ admits an $L'$-bi-Lipschitz embedding into $\R^{M'}$, where $L'$ and $M'$ depend only on the doubling and bounded turning constants of $X$. By Theorem \[thm:welding\], $Q\subseteq B_1\cup\cdots\cup B_n$ admits an $L_2$-bi-Lipschitz embedding into $\R^{M_2}$, where $L_2$ and $M_2$ depend only on the doubling and bounded turning constants of $X$. This verifies condition of Theorem \[thm:seo\] and completes the proof of Theorem \[thm:mainembed2\]. [^1]: G. C. David was partially supported by NSF DMS grant 1758709. V. Vellis was partially supported by NSF DMS grants 1800731 and 1952510.
{ "pile_set_name": "ArXiv" }
--- abstract: 'Dilation and thermopower measurements on YbAgGe, a heavy-fermion antiferromagnet, clarify and refine the magnetic field-temperature ($H$-$T$) phase diagram and reveal a field-induced phase with $T$-linear resistivity. On the low-$H$ side of this phase we find evidence for a first-order transition and suggest that YbAgGe at 4.5 T may be close to a quantum critical end point. On the high-$H$ side our results are consistent with a second-order transition suppressed to a quantum critical point near 7.2 T. We discuss these results in light of global phase diagrams proposed for Kondo lattice systems.' author: - 'G. M. Schmiedeshoff' - 'E. D. Mun' - 'A. W. Lounsbury' - 'S. J. Tracy' - 'E. C. Palm' - 'S. T. Hannahs' - 'J. -H. Park' - 'T. P. Murphy' - 'S. L. Bud’ko' - 'P. C. Canfield' title: Multiple regions of quantum criticality in YbAgGe --- 327[Sr$_3$Ru$_2$O$_7$]{} When a classical second-order phase transition is suppressed to zero temperature $T$ by a tuning parameter (such as pressure, doping, or magnetic field $H$), a quantum critical point (QCP) can occur [@Sachdev1999]; suppression of a first-order phase transition can lead to a quantum critical end point (QCEP) [@Millis2002]. In the vicinity of a QCP or QCEP quantum fluctuations associated with the zero-point energies of the adjacent $T=0$ phases can persist to remarkably high temperatures. These fluctuations can dramatically affect the interactions between particles, leading to unusual thermodynamic and transport properties [@Sachdev1999; @Stewart2001; @Millis2002; @Lohneysen2007] and to novel states of matter [@Coleman2005; @Gegenwart2008a; @Schofield2010]. YbAgGe, a stoichiometric heavy-fermion (HF) antiferromagnet, crystallizes in a hexagonal ZrNiAl-type structure [@Poettgen1997]. The zero-field electronic specific heat coefficient falls in the range 0.15-1.0 J/mol K$^2$ and the Kondo temperature is 20-25 K [@Budko2004a]. The Yb ions form a quasi-Kagomé lattice in which magnetic coupling, geometric frustration and Kondo interactions compete in a manner which allows the suppression of low temperature AF order by modest applied magnetic fields, fields that tune the quantum critical behavior [@Si2010a; @Kim2010] and lead to a complex magnetic phase diagram [@Gignoux2001]. The phase diagram of YbAgGe, summarizing earlier measurements [@Budko2008a] with $H$ perpendicular to the $c$-axis, is shown in Fig. \[fig01\]a. Solid (dashed) lines are guides to the eye for thermodynamic phase boundaries (‘Hall lines’ denoting features in the field-dependent Hall resistivity [@Budko2005a; @Budko2005b]) labeled by numerals 1-3 (4-6), lower case letters $a$-$f$ label phases or regions of the phase diagram. Commensurate AF order is observed in the $a$-phase [@Fak2005a] where a sharp first-order phase transition manifests along phase line 1 [@Budko2004a], and incommensurate AF order has been reported in the $b$-phase [@Fak2006a]. Relatively broad features manifest along the higher temperature part of phase line 3 [@Budko2004a], but below about 0.5 K the features sharpen and neutron scattering measurements reveal a return to commensurate AF order [@McMorrow2008]. Based on the suppression of phase line 3 near 4.5 T and its near coincidence with Hall line 4 below 0.3 K, a field-induced QCP was proposed [@Budko2004a; @Budko2005a; @Budko2005b]. Intriguingly, several other features of this phase diagram remained mysterious. ![(color online) Representative longitudinal dilation data with $H$ applied parallel to the $ab$-axis. Panel (a) shows the low temperature thermal expansion in the fixed fields listed next to arrows denoting the appropriate axes for each data set (the data at 0 T/1.2 T have been shifted by $+10/-9.0\times{10}^{-6}$/T for clarity). The $T_i$ label features associated with the phase boundaries or crossovers of Fig. 1 (see text). A semi-log plot of $\alpha_{ab}/T$ vs. $T$ in the vicinity of 4.5 T is shown in the inset where the direction of increasing $H$ is shown by a curved arrow and the solid line is a guide to the eye for the low temperature 4.7 T data. Panel (b) shows the magnetostriction at 387 mK. The $H_j$ label features in the data where $j$ denotes the guides to the eye of Fig. 1a (see text). The inset shows an expanded view of the transitions below 6 T at 17 mK. \[fig02\]](f2.eps){height="4.0"} The low-$T$ electrical resistivity is large [@Budko2004a] and varies like $T^n$ with $n\simeq{1}$ in region $d$, smoothly increasing from $\simeq{1}$ to $\simeq{2}$ in region $e$, and $\simeq{2}$ in region $f$ [@Niklowitz2006]. Such non-Fermi liquid (nFL) behavior is expected near a QCP [@Stewart2001; @Lohneysen2007], but the broad field range with $n\simeq{1}$ and the recovery of Fermi liquid (FL) behavior in fields so far above that of the QCP are surprising. Further, a logarithmic divergence of the specific heat appears most clearly for $H\sim$ 7 T [@Budko2004a], near Hall line 5, raising the possibility of (at least) one other phase being suppressed in fields well above that of the proposed QCP near 4.5 T. In this paper we describe longitudinal dilation and transverse thermoelectric power (TEP) measurements on YbAgGe that shed light on some of these mysteries and paint a fuller picture of quantum criticality in Yb-based HFs. ![(color online) Representative transverse TEP data with $H$ applied parallel to the $ab$-axis and the temperature gradient in the basal plane, perpendicular to $H$. Panel (a) shows the temperature dependence of $S$ across phase boundaries or crossover regions at the fields shown. The $T_i$ label features associated with the phases (or regions) of Fig. 1. Panel (b) shows $S/T$ vs. $T$ in applied fields on a semi-log plot. The solid line through the 8 T data is a guide to the eye. The field dependence of $S$ at 0.4 K is shown in the inset where the $H_i$ label features in the data and $i$ denotes the guides to the eye of Fig. 1a. \[fig03\]](f3.eps){height="4.0"} Single crystals of YbAgGe were grown from an AgGe-rich ternary solution [@Budko2004a]. Longitudinal dilation and transverse TEP measurements were made with techniques described elsewhere [@Schmiedeshoff2006b; @Mun2010a]. $H$ was applied normal to the $c$-axis. The coefficients of linear thermal expansion and linear magnetostriction along the $ab$-axis (approximately parallel to \[210\]) are $\alpha_{ab} = \partial(\ln{L_{ab}})/\partial{T}$ and $\lambda_{ab} = \partial(\ln{L_{ab}})/\partial{H}$ respectively where $L_{ab}$ is the thickness of the sample, 1-2 mm. The Seebeck coefficient is $S = -\Delta{V}/\Delta{T}$, where $\Delta{V}$ is the potential difference across the sample when a temperature gradient is applied. Typical data are shown in Figs. \[fig02\] and \[fig03\] where nFL behavior is also illustrated. In our dilation data, first-order phase transitions are identified by their peak-like shape and the presence of thermal or magnetic hysteresis ([*e.g.*]{} features labeled $T_a$ and $H_{1-4}$ in Fig. 2); second-order transitions do not show hysteresis within experimental uncertainty and usually show a step-like shape as illustrated by the dashed lines through the 3.4 and 6.0 T data of Fig. 2a. Pronounced features in the 3.4 and 6.0 T data of Fig. 2a, labeled $T_{c,d}$ correlate well with extrema in $\lambda_{ab}$ and are used to construct the phase diagram. The first-order transitions observed in the magnetostriction exhibit varying levels of magnetic hysteresis. This hysteresis becomes quite pronounced as the temperature falls below about 0.2 K, as illustrated in the inset of Fig. 2b where the features associated with $H_1$ and $H_3$ change dramatically in the $dH/dt < 0$ data. For all of our $T \leq$ 0.2 K magnetostriction data the sample was zero-field-cooled from temperatures above 0.3 K. Fig. 1b was constructed from $dH/dt > 0$ data. The phase diagram assembled from our dilation and TEP measurements is shown in Fig. 1b where the solid and dashed lines are those of Fig. 1a, except for the dotted portion of phase line 3 where we observe no thermodynamic features. In comparison with the phase diagram of Fig. 1a we find: that phase line 2 joins phase line 3 to surround the $c$-phase; that Hall lines 4-6 are all associated with thermodynamic phase transitions as $T \rightarrow 0$; and that phase transitions anchoring Hall lines 4 and 5 define the low temperature boundaries of a field-induced $d$-phase (the high temperature limits of the $n\simeq{1}$ region of the resistivity agree reasonably well with the top of the $d$-phase dome [@Niklowitz2006]). The phase transitions are second-order across the tops of the $c$- and $d$-phase domes and first-order along phase line 3 (Hall line 4) below 0.3 K (0.2 K). The overlap of the phase line data determined from $\alpha_{ab}$, $\lambda_{ab}$, and $S(T,H)$ is extensive along the $c$-phase boundary, there is some overlap on the high-field side of the $d$-phase but no overlap is observed on the low-field side where a small gap appears between the $\alpha_{ab}$ and $\lambda_{ab}$ data, a gap Hall line 4 passes through. Broad extrema in $\alpha_{ab}(T)$ and $S(T)$, labeled with $T_b$ in Figs. \[fig02\]a and \[fig03\]a respectively, may represent a cross-over region between higher-$T$ fluctuations and the lower-$T$ incommensurate AF order of the $b$-phase, leading to the uncertain nature of the phase diagram in this region. Several signatures of quantum criticality appear in our data. A change in the sign of the thermal expansion at a QCP has been predicted [@Garst2005], the phase diagram coordinates where $\alpha_{ab}$ passes through zero are plotted as solid triangles in Fig. \[fig01\]b. These ‘lines of zeros’ pass along the tops and high-$H$ sides of the $c$- and $d$-phases and presumably extend to $T = 0$ near 4.5 and 7.2 T. Hall line 4 (5) correlates with the line of zeros associated with the $c$-phase ($d$-phase) as $T\rightarrow{0}$. A sign change in $S(T)$ correlates with the field-induced QCP in YbRh$_2$Si$_2$ where it is attributed to an abrupt change in the Fermi surface [@Hartmann2010a], though such a sign change does not appear to be a universal QCP signature [@Kim2010]. $S/T$ as $T \rightarrow 0$ is predicted to reach its maximum value as the QCP is approached and its symmetry with respect to the QCP can help distinguish between theoretical models [@Kim2010]. Sign changes in $S(T,H)$ for YbAgGe are plotted as solid diamonds in Fig. 1b and correlate with Hall line 4. Maxima in $S(H)$ are plotted as solid stars in Fig. 1b and correlate with Hall line 5 where the largest values of $S/T$ as $T \rightarrow 0$ also appear. ![(color online) The temperature dependence of the Grüneisen parameter $\Gamma_{ab}$ of YbAgGe in applied magnetic fields. The curved arrow represents the direction of increasing field. (The apparent, slight downturn of the 8.0 T data at the low-$T$ end is comparable to the noise in these data.) The inset shows the low temperature end of the 4.5 T data on a log-log plot. The solid line is a guide to the eye representing a $T^{-5/2}$ temperature dependence. \[fig04\]](f4.eps){width="2.9"} The Grüneisen parameter, characterizing the volume dependence of the energy scales in the system (which should be dominated by the quantum critical contribution as $T \rightarrow 0$), is defined as $\Gamma_{ab} = V_m\alpha_{ab}/\kappa{C_p}$ where $V_m$ is the molar volume [@Poettgen1997], $\kappa$ is the compressibility [@Sengupta2010], and $C_p$ is the specific heat [@Tokiwa2006]. $\Gamma_{ab}$ is predicted to diverge as a QCP is approached [@Zhu2003], our results are shown in Fig. 4. The largest values of $\Gamma_{ab}$, comparable to that of other HF compounds [@deVisser1990], occur at 4.5 T. At this field the low-$T$ upturn in $\Gamma_{ab}(T)$ is consistent with the onset of Grüneisen divergence, perhaps with a power-law temperature dependence as suggested by the data in the inset of Fig. 4 (though the temperature range over which divergent behavior is observed is too limited, at this time, for a definitive quantitative analysis). $\Gamma_{ab}(T)$ at 7, 8, and 9 T also show low-$T$ upturns suggesting that Grüneisen divergence may develop at lower temperatures; we cannot, as yet, explain the large magnitudes nor the shapes of $\Gamma_{ab}(T)$ above 0.5 K in this field range. It will be necessary to extend these measurements to lower temperatures, higher fields, and at a higher density of field values. We take the presence of nFL behavior (both earlier [@Budko2008a] and current: the logarithmic temperature dependences of $\alpha_{ab}/T$ and $S/T$ suggested by the data in the inset of Fig. 2a and the main panel of Fig. 3b respectively), the features in the Hall resistivity [@Budko2005a; @Budko2005b] (Hall lines 4 and 5), the zeros in $\alpha_{ab}(T)$, and the features and zeros in $S(T,H)$ as strong evidence for (at least) two regions of quantum criticality in YbAgGe near $H_{c1} = 4.5$ T (also supported by the onset of Grüneisen divergence) and $H_{c2} = 7.2$ T. If a QCP at $H_{c1}$ is due to the suppression of the AF transition characterized by the full phase line 3, as previously suggested [@Budko2004a], then one would expect a continuous transition as $T \rightarrow 0$ [@Sachdev1999], perhaps similar to that observed in YbRh$_2$Si$_2$ [@Gegenwart2008a]. However, the transitions along phase line 3 are first-order (see Fig. 2b) as $T \rightarrow 0$. A step-like feature developing in the low-$T$ magnetization near $H_{c1}$ [@Tokiwa2006] and the thermodynamic structure of the $c$- and $d$-phases (two second-order phase lines approaching/joining a first-order phase line extending to $T=0$) are similar to that expected for the spin-flop class of metamagnetism [@Stryjewski1977; @Aharony1983]. Metamagnets can exhibit a wide variety of multicritical behavior, including intermediate anisotropy scenarios leading to critical end points [@Vilfan1986]. We suggest that YbAgGe may be close to a QCEP near $H_{c1}$, possibly similar to that observed in 327 [@Grigera2001a] but with spin-flop metamagnetism driving the quantum criticality. The phase transitions at the base of Hall line 5 are clearly continuous as shown in the main panel of Fig. 2b while $n\simeq{1}$ and $\alpha_{ab}(T)$ passes through zero here. In this region of the phase diagram both the specific heat and $S/T$ are logarithmically divergent. We propose that a QCP occurs near $H_{c2} = 7.2$ T, driving the pronounced nFL behavior in this region of the phase diagram. Theoretical efforts, characterizing several quantum critical materials on a global phase diagram incorporating Kondo coupling and degree of magnetic frustration, suggest that YbAgGe may evolve from AF order ($d$-phase) through a spin-liquid phase ($e$-phase) before a FL ($f$-phase) is recovered [@Si2010a; @Custers2010a]. Recent work on other Yb-based HF compounds suggests [@Custers2010a; @Coleman2010] that a quantum critical [*phase*]{}, bounded by QCPs at $H_{c2}$ and a hypothetical spin-liquid/heavy FL QCP [@Si2010a] near 12 T, may underlie the spin-liquid. The linear $T$-dependence and large magnitude of the resistivity suggest strange-metal behavior for the $d$-phase [@McGreevy2010], and as a spin-flop phase it will carry a net magnetic moment [@Stryjewski1977], though the small step in $M(H)$, about $0.1 \mu_B/$Yb [@Tokiwa2006], suggests that an underlying AF symmetry may still be present. The theoretical identification of the $e$-phase as a spin-liquid is supported by the large magnetostriction [@Ruff2010] we observe. The proposed global phase diagram thus seems appropriate for YbAgGe. Elastic and inelastic neutron scattering would be profitable though challenging microscopic probes of these high-$H$ phases. In conclusion, dilation and TEP measurements reveal high-$H$ phase boundaries in YbAgGe that delineate the region of $T$-linear resistivity. On the low-$H$ side this phase appears to be close to a QCEP near 4.5 T, associated with a first-order, most likely metamagnetic, phase transition. On the high-$H$ side this phase appears to end in the continuous suppression of a second-order transition ending in a QCP near 7.2 T, explaining the pronounced nFL behavior nearby. Even with the identification of this field-stabilized phase, there remains a clear nFL region over which the resistivity varies as $T^n$ with $n$ continuously changing from $\simeq{1}$ to $\simeq{2}$ as $H$ increases. Theory suggests a quantum critical phase and/or spin-liquid in this region but more evidence, theoretical as well as experimental, is needed. This work was supported by the National Science Foundation under DMR-1006118. Work at Ames Laboratory was supported by the U.S. Department of Energy, Basic Energy Sciences, under Contract No. DE-AC02-07CH11358. Work at the National High Magnetic Field Laboratory was supported under the auspices of the National Science Foundation, the State of Florida, and the U.S. Department of Energy. [35]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ** (, , ). , , ****, (). , ****, (). , , ****, (). , ****, (). , , ****, (). , ****, (). , , ****, (). , , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , , ****, (). , , ****, (). , , ****, (). , , ****, (). . , , ****, (). , , ****, (). , , ****, (). , ****, (). , , ****, (). , , ****, (). , , ****, (). , , ****, (). , , ****, (). , ****, (). , ****, (). , ****, (). , , ****, (). , , ****, (). , ****, (). , ****, (). , , ****, ().
{ "pile_set_name": "ArXiv" }
--- abstract: 'With the growing complexity of deep learning methods adopted in practical applications, there is an increasing and stringent need to explain and interpret the decisions of such methods. In this work, we focus on explainable AI and propose a novel generic and model-agnostic framework for synthesizing input exemplars that maximize a desired response from a machine learning model. To this end, we use a generative model, which acts as a prior for generating data, and traverse its latent space using a novel evolutionary strategy with momentum updates. Our framework is generic because $(i)$ it can employ any underlying generator, e.g. Variational Auto-Encoders (VAEs) or Generative Adversarial Networks (GANs), and $(ii)$ it can be applied to any input data, e.g. images, text samples or tabular data. Since we use a zero-order optimization method, our framework is model-agnostic, in the sense that the machine learning model that we aim to explain is a black-box. We stress out that our novel framework *does not* require access or knowledge of the internal structure or the training data of the black-box model. We conduct experiments with two generative models, VAEs and GANs, and synthesize exemplars for various data formats, image, text and tabular, demonstrating that our framework is generic. We also employ our prototype synthetization framework on various black-box models, for which we only know the input and the output formats, showing that it is model-agnostic. Moreover, we compare our framework (available at <https://github.com/antoniobarbalau/exemplar>) with a model-dependent approach based on gradient descent, proving that our framework obtains equally-good exemplars in a shorter time.' author: - 'Antonio Barbalau$^{1,3}$' - 'Adrian Cosma$^{2, 3}$' - Radu Tudor Ionescu$^1$ - 'Marius Popescu$^{1,3}$' bibliography: - 'refs.bib' title: 'A Generic and Model-Agnostic Exemplar Synthetization Framework for Explainable AI' --- Introduction ============ Related Work {#sec_related} ============ Method {#sec_method} ====== Experiments {#sec_exp} =========== Conclusion {#sec_conclusion} ========== In this paper, we proposed a novel evolutionary strategy that incorporates momentum for generating exemplars for black-box models. Our framework requires an underlying generator, but it does not back-propagate gradients through the black-box model or the generator. We conducted experiments, showing that our approach can produce exemplars for three data types: image, text and tabular. Furthermore, our experiments indicate that our idea of incorporating momentum into a standard evolutionary strategy is useful, reducing the number of model calls by $19\%$. The empirical results demonstrate that our optimization algorithm converges faster than gradient descent with momentum, while providing similar or even more realistic exemplars. Given that our method does not require access to the weights or the training data of the black-box model, we believe it has a boarder applicability than gradient descent methods such as [@Nguyen-NIPS-2016].
{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove a new global existence result for the asymptotically flat, spherically symmetric Einstein-Vlasov system which describes in the framework of general relativity an ensemble of particles which interact by gravity. The data are such that initially all the particles are moving radially outward and that this property can be bootstrapped. The resulting non-vacuum spacetime is future geodesically complete.' author: - | H[å]{}kan Andréasson\ Department of Mathematics, Chalmers,\ S-41296 Göteborg, Sweden\ email: [email protected]\  \ Markus Kunze\ Fachbereich Mathematik\ Universität Duisburg-Essen\ D-45117 Essen, Germany\ email: [email protected]\  \ Gerhard Rein\ Fakultät für Mathematik und Physik\ Universität Bayreuth,\ D-95440 Bayreuth, Germany\ email: [email protected] title: 'Global existence for the spherically symmetric Einstein-Vlasov system with outgoing matter' --- Introduction ============ Global properties of non-vacuum spacetimes are one of the central themes of the mathematical analysis of general relativity. A matter model which has proven particularly suitable for such an analysis is the collisionless gas as described by the Vlasov equation. Here the matter content of spacetime is represented by a large ensemble of particles with a density function $f$ on phase space, i.e., on the mass shell of the tangent bundle of the spacetime manifold. The particles move along time-like geodesics which is equivalent to $f$ satisfying the corresponding continuity equation, i.e., the Vlasov equation, and the energy momentum tensor is then expressed in terms of $f.$ For a general introduction to the resulting Einstein-Vlasov system we refer to [@An1] and [@Rl7]. A major result for the asymptotically flat case of this system is a global existence result for small, spherically symmetric initial data [@RR1]. In addition, there is numerical evidence [@AR1; @OC; @RRS2] that at least in the spherically symmetric case the weak cosmic censorship conjecture holds: The gravitational collapse of regular initial data leads to a black hole rather than a naked singularity. In view of the small data result mentioned above and the lack of proof of global existence for general data (in Schwarzschild coordinates or in maximal-areal coordinates) it is natural to ask if one can find an open set of initial data which are not small in the sense of [@RR1] but for which global existence holds. In the present paper we do find initial data with this property by establishing global existence for data which are such that initially all the particles in the ensemble move radially outward sufficiently fast so that they can be shown to keep on moving outward for all future time. This configuration results in estimates on the geometry of spacetime which imply that it is future geodesically complete. The basic mechanism of the proof is completely different from the one for small initial data which relied on the dispersive properties of the Vlasov matter model when the fields are small. Since its actual implementation for the Einstein-Vlasov system quite hides this mechanism, it should be useful to explain it first in the case of the much simpler Vlasov-Poisson system which is the Newtonian limit of the Einstein-Vlasov system. Due to the spherical symmetry the maximal force experienced by a particle at distance $r$ from the origin is $-M/r^2$ in the Vlasov-Poisson case, where $M>0$ is the total mass of the ensemble. Hence along any particle trajectory $$\frac{d}{dt} \left(\frac{1}{2} w^2 - \frac{M}{r}\right) = w \dot w + \frac{M}{r^2}\dot r = w\,\left(\dot w + \frac{M}{r^2}\right) \geq 0,$$ as long as its radial velocity $\dot r = w= x\cdot v/r \geq 0$. Hence $$\frac{1}{2} w^2(t) - \frac{M}{r(t)} \geq \frac{1}{2} w^2(0) - \frac{M}{r(0)}$$ and $$\frac{1}{2} w^2(t) \geq \frac{1}{2} w^2(0) - \frac{M}{r(0)}$$ on any time interval on which $w(t)$ remains non-negative. Now let $w_->0$ be an initial lower bound for the radial velocities of the particles in the ensemble, $r_->0$ an initial lower bound for their distance from the origin, and assume that $$W_- := \frac{1}{2} w_-^2 - \frac{M}{r_-} > 0.$$ Then as long as a particle is moving outward, $$w(t) > W_-,\ r(t) > r_- + W_- t.$$ But this implies that all the particles keep moving outward for all future time. For the Vlasov-Poisson system global existence and uniqueness of smooth solutions has been established for general initial data [@Pf; @LP], but the above argument can easily be extended to yield global existence for spherically symmetric initial data as specified above. For the spherically symmetric Einstein-Vlasov system we find a substitute for the pseudo-energy $\frac{1}{2} w^2 - \frac{M}{r}$ used above, but when taking its derivative along a particle trajectory it turns out to be important to use suitable coordinates on spacetime, namely maximal-areal coordinates, in order to see that under suitable circumstances this quantity does not decrease along radially outgoing trajectories. Once this is shown a bootstrap argument implies that all the particles keep moving radially outward on the future directed maximal interval of existence of the solution. We can then use a result from [@Rl7] to conclude that the solution is global in the time coordinate used, and the obtained control of the metric quantities turns out to be sufficient to establish future geodesic completeness. The Einstein-Vlasov system ========================== We consider the system in the spherically symmetric, asymptotically flat case and write the metric in the following form: $$ds^{2}=-(\alpha^2-a^2\beta^2)dt^2+2a^2\beta dtdr+a^2 dr^2 + r^2\left(d\theta^2 + \sin^2\theta\, d\phi^2\right).$$ Here the metric coefficients $\alpha, \beta$, and $a$ depend on $t\in {\mathbb R}$ and $r\geq 0$, $\alpha$ and $a$ are positive, and the polar angles $\theta\in[0,\pi]$ and $\phi\in[0,2\pi]$ parameterize the unit sphere. The radial coordinate $r$ is thus the area radius. Let $K^{a}_{b}$ be the second fundamental form and define \[kdef\] =K\^\_=. By imposing the maximal gauge condition, which means that each hypersurface of constant $t$ has vanishing mean curvature, we obtain the following field equations where subscripts denote partial derivatives with respect to $r$ or $t$: $$\begin{aligned} a_{r} & = & \frac{3}{2}a^{3}r \kappa^2+4\pi r a^{3}\rho+\frac{a}{2r}(1-a^{2}), \label{ee1}\\ \kappa_{r} & = & -\frac{3}{r}\kappa-4\pi a\jmath, \label{ee2}\\ a_{t} & = & 2\alpha a \kappa+(a\beta)_{r}, \label{ee3}\\ \displaystyle\alpha_{rr} & = & \alpha_{r}\left(\frac{a_{r}}{a}-\frac{2}{r}\right) +\frac{2\alpha}{r^2}\left(2r\frac{a_{r}}{a}+a^2-1\right)+4\pi a^2\alpha(S-3\rho). \label{ee4}\end{aligned}$$ The Vlasov equation takes the form $$\label{vlasov} \partial_{t}f+\left(\frac{\alpha w}{a E}-\beta\right)\partial_{r}f +\left(-\frac{\alpha_{r} E}{a}-2\alpha \kappa w+ \frac{\alpha L} {a r^3 E}\right)\partial_{w}f=0,$$ where $$\label{E} E=E(r,w,L)=\sqrt{1+w^{2}+L/r^{2}}.$$ The variables $w$ and $L$ can be thought of as the momentum in the radial direction and the square of the angular momentum respectively, see [@R] for more details. The matter quantities which appear in the field equations are defined by $$\begin{aligned} \rho(t,r) & = & \frac{\pi}{r^{2}} \int_{-\infty}^{\infty}\int_{0}^{\infty} E f(t,r,w,L) \,dL\,dw, \label{rho-def} \\ \jmath(t,r) & = & \frac{\pi}{r^{2}}\int_{-\infty}^{\infty}\int_{0}^{\infty} w f(t,r,w,L)\,dL\,dw, \label{j-def} \\ S(t,r) & = & \frac{\pi}{r^{2}}\int_{-\infty}^{\infty}\int_{0}^{\infty} \frac{w^2 + L /r^{2}}{E}\, f(t,r,w,L)\,dL\,dw. \label{S-def}\end{aligned}$$ We impose the following boundary conditions which ensure asymptotic flatness and a regular center: $$\label{bdryc} a(t,0)=a(t,\infty)=\alpha(t,\infty)=1.$$ The equations (\[ee1\])–(\[S-def\]) together with the relation (\[kdef\]) and the boundary conditions (\[bdryc\]) constitute the Einstein-Vlasov system for a spherically symmetric, asymptotically flat spacetime in maximal-areal coordinates. First we note that the phase space density $f$ is constant along solutions of the characteristic system $$\begin{aligned} \dot r & = & \frac{\alpha(\tau,r) w}{a(\tau,r) E} -\beta(\tau,r), \label{charsys1} \\ \dot w & = & -\frac{\alpha_{r}(\tau,r) E}{a(\tau,r)} -2\alpha(\tau,r) \kappa(\tau,r) w +\frac{\alpha(\tau,r) L}{a(\tau,r) r^3 E}, \label{charsys2} \\ \dot L & = & 0, \label{charsys3}\end{aligned}$$ of the Vlasov equation. If $\tau\mapsto (R,W,L)(\tau,t,r,w,L)$ denotes the solution of the characteristic system with $(R,W,L)(t,t,r,w,L)=(r,w,L)$ then $$f(t,r,w,L)=\fn((R,W,L)(0,t,r,w,L)),$$ with $\fn = f_{|t=0}$ the initial data for $f$. In particular, $f\geq 0$ provided this is true for the initial data as we will assume throughout. Next we notice that (\[ee2\]) can be rewritten as $$\left(r^3 \kappa\right)_r = - 4 \pi r^3 a \jmath,$$ and upon integration $$\label{kap-form} \kappa(t, r)=-\frac{4\pi}{r^3} \int_0^r a(t, s)\jmath(t, s) s^3\,ds.$$ The Hawking mass $m$ is given by $$\label{m} m=\frac{r}{2}\left(1+\frac{\beta^2}{\alpha^2}-\frac{1}{a^2}\right).$$ We also introduce the quantity $$\label{mudef} \mu=\frac{r}{2}\left(1-\frac{1}{a^2}\right)$$ and note that by (\[ee1\]), $\mu$ can be expressed as $$\label{murewr} \mu(t,r)=\int_{0}^{r}\left(4\pi \rho(t,s) +\frac{3}{2}\kappa^2(t,s)\right) s^2\,ds.$$ Assuming that the matter is compactly supported initially and hence also for later times, (\[kap-form\]) implies that $\kappa(t,r)\sim r^{-3}$ for $r$ large. Hence the limits as $r$ tends to $\infty$ of $m$ and $\mu$ are equal, so that the ADM mass $M$ is $$\label{M-def} M=\int_0^{\infty}\left(4\pi \rho(t,r)+\frac{3}{2} \kappa^2(t,r)\right) r^2 dr.$$ The ADM mass is conserved, and $\mu(t, r)\le M$. Finally we notice that using (\[ee1\]) the second order equation (\[ee4\]) can be rewritten in the form $$\left(\frac{r^2}{a}\alpha_r \right)_r =4\pi r^2 a \alpha (\rho + S) + 6 r^2 a \alpha \kappa^2,$$ and upon integration, $$\label{alphar1} \alpha_r(t, r)=\frac{a(t, r)}{r^2}\int_{0}^{r}\left(4\pi\alpha a (\rho+S)+6 a \alpha \kappa^2\right) s^2\,ds.$$ Thus $\alpha$ is monotonically increasing outwards, and from (\[bdryc\]) it follows that $$\label{alphaleq1} 0<\alpha\leq 1.$$ A direct computation using (\[ee4\]), (\[ee1\]), (\[mudef\]), and (\[murewr\]) implies that $$\left(\frac{r^2}{a^2}\alpha_r - \alpha \mu \right)_r = 4\pi r^2 S \alpha + \frac{9}{2} r^2 \kappa^2 \alpha -4\pi r^3\rho\alpha_r - \frac{3}{2} r^3 \kappa^2 \alpha_r.$$ With the boundary conditions $(\alpha \mu)_{|r=0} = 0 = \left(r^2 a^{-2}\alpha_r\right)_{|r=0}$ this yields the identity $$\label{alphar2} \frac{r^2}{a^2(t,r)}\,\alpha_r(t,r) =(\alpha \mu)(t,r) +\int_0^r\left[4 \pi S \alpha + \frac{9}{2} \kappa^2 \alpha -4\pi s \rho \alpha_r - \frac{3}{2} s \kappa^2 \alpha_r\right] s^2 ds,$$ which turns out to be very convenient in what follows. The main result =============== Let $\fn\,$ be some initial data of ADM mass $M$ such that $$\supp \fn \subset ]r_-, r_+[\times [w_-, \infty[\times [0, L_+]\,.$$ We wish to specify these parameters in such a way that for the corresponding solution of the Einstein-Vlasov system all particles have radial velocity $w > W_-$ with an additional parameter $W_->0$. To this end we abbreviate $$\label{xieta-def} \xi = \frac{2M}{r_-},\ \eta = \frac{W_-}{\sqrt{1+W_-^2+L_+/r_-^2}},$$ and we require that the above parameters satisfy the following conditions: $$\begin{aligned} \label{master1} && \left(2\min\left\{\frac{r_-}{r_+},\frac{1-3\xi}{\sqrt{1-\xi}+ \xi/2} \eta\right\} -1 -\frac{9}{4}\,\frac{\xi}{1-\xi} -\frac{3}{2}\,\frac{ \xi}{\sqrt{1-\xi}} -\frac{L_+}{W_- r_-^2}\right) W_-^2 \nonumber \\ && \qquad\qquad -\left(1+\frac{L_+ }{W_- r_-^2} + \frac{9}{4}\,\frac{\xi}{1-\xi}\right) \left(1+\frac{L_+ }{r_-^2}\right) -\frac{L_+ }{r_-^2}\frac{1}{1-\xi} > 0\end{aligned}$$ and $$\label{master2} w_-^2> \frac{1}{1-\xi}\left(W_-^2+\xi\left(1+\frac{L_+}{r_-^2}\right)\right).$$ After stating our main result we will show that there exist parameters satisfying these conditions, and that there exist initial data with such parameters. [**Theorem**]{} [*Suppose that $0<r_-<r_+$, $w_->0$, $L_+>0$, $W_->0$, and $M>0$ satisfy the conditions (\[master1\]) and (\[master2\]), and consider a solution $f$ on a maximal existence interval $[0, T[$ such that the initial data $\fn=f_{|t=0}$ is compactly supported, smooth, non-negative, has ADM mass $M$, and satisfies $$\label{supp-as} \supp\fn\subset ]r_-, r_+[\times [w_-, \infty[\times [0, L_+]\,.$$ Then $T=\infty$, and the resulting spacetime is future geodesically complete.* ]{} The proof of this theorem is given in Section \[proof-sec\] below. To show that parameters $\xi, r_\pm, w_-, L_+, W_-$ exist which satisfy (\[master1\]) and (\[master2\]) we proceed as follows. Let $$D(\xi,\eta)=2 \frac{1-3\xi}{\sqrt{1-\xi}+ \xi/2} \eta -1 -\frac{9}{4}\,\frac{\xi}{1-\xi} -\frac{3}{2}\,\frac{ \xi}{\sqrt{1-\xi}},\quad\xi\in [0, 1[,\ \eta >0.$$ Since $D(0,1)=1$ and $D(1/3,1) < 0$ there exists some $\xi^\ast \in ]0,1/3[$ such that $D(\xi,1)>0$ for $\xi\in [0, \xi^\ast[$ and $D(\xi^\ast,1)=0$; numerically, $\xi^\ast\approx 0.0994$. So firstly we assume that $$\xi<\xi^\ast.$$ Secondly, we choose $r_+>r_->0$ such that $$2 \frac{r_-}{r_+} > 1 + \frac{9}{4}\,\frac{\xi}{1-\xi} + \frac{3}{2}\,\frac{ \xi}{\sqrt{1-\xi}},$$ where we notice that the right hand side is increasing in $\xi$ and its value for $\xi=\xi^\ast$ is $\approx 1.405$ so that the inequality can be satisfied with $r_+>r_->0$. Now we fix $L_+>0$ arbitrarily. Since $\eta \to 1$ as $W_- \to \infty$ we can choose $W_->0$ sufficiently large so that $$D(\xi,\eta) - \frac{L_+ }{W_- r_-^2} > 0$$ and $$2 \frac{r_-}{r_+} > 1 + \frac{9}{4}\,\frac{\xi}{1-\xi} + \frac{3}{2}\,\frac{ \xi}{\sqrt{1-\xi}} + \frac{L_+ }{W_- r_-^2}.$$ Then the factor multiplying $W_-^2$ in the condition (\[master1\]) is positive, and by further increasing $W_-$ if necessary condition (\[master1\]) is satisfied. Finally, we take $w_->0$ sufficiently large so that (\[master2\]) holds as well. Let us now fix a set of parameters $\xi, r_\pm, L_+, w_-, W_-$ such that (\[master1\]) and (\[master2\]) are satisfied. We want to argue that an initial condition with these parameter values does exist. In order to do so we fix some compactly supported, smooth function $\open{g}\,$ such that $$\supp \open{g}\subset ]r_-, r_+[\times [w_-, \infty[\times [0, L_+]\,,$$ and we consider $\fn=A\open{g}\,$ for some amplitude $A>0$. Quantities induced by $\fn$ will be denoted by the corresponding superscript, and they depend on $A$. By continuity, we can make sure that $\open{a}(r)\le 2$ for all $A>0$ sufficiently small; the dependence of the initial data for the metric quantities on the initial matter terms and hence on $\fn$ will be studied in [@AKR]. Now let $$I=\int_0^\infty4\pi\open{\rho}(s) s^2\,ds.$$ Then by (\[kap-form\]), $$|\open{\kappa}(r)| \le\frac{4\pi}{r^3}\int_0^r\open{a}(s)\,|\open{\jmath}(s)|\,s^3\,ds \le \frac{8\pi}{r^2}\int_0^r\open{\rho}(s) s^2\,ds =\frac{2I}{r^2}\, .$$ Since $\open{\kappa}(r)=0$ for $r\le r_-$ this implies that $$I \leq M = \int_{r_-}^\infty \left(4\pi\open{\rho}(r) +\frac{3}{2}\open{\kappa}^{\,\,2}(r)\right) r^2 dr \le I + \frac{6 I^2}{ r_-}.$$ Since the quantity $I$ depends linearly on the amplitude $A$, $2M(\fn\,)/r_- = \xi < \xi^\ast$ for $A>0$ sufficiently small, and arguing again by continuity any value of $\xi \in ]0,\xi^\ast[$ should be realized by adjusting $A$ properly. Since $\fn$ has the same support as $\open{g\,}$, all conditions required in the theorem then hold for $\fn$. Proof of the theorem {#proof-sec} ==================== [*Step 1: The bootstrap argument.*]{}\ Our first and crucial step in the proof of the theorem is a bootstrap argument which shows that the condition $w>W_-$ which holds on the support of $f$ initially will persist on the maximal forward existence interval of the solution. To begin with we show that for all $t\in [0,T[$ and $r\geq r_-$, $$\label{gote} a(t,r) < a_+=\frac{1}{\sqrt{1-\xi}}\ \mbox{and}\ \alpha(t,r) > \alpha_-= \frac{1-3\xi}{1-\xi};$$ eventually we shall see that these bounds hold for all $r\geq 0$. We recall that $\xi=2M/r_-$. By (\[mudef\]), $$\frac{1}{a^2(t, r)}=1-\frac{2\mu(t, r)}{r} > 1-\frac{2M}{r_-}=1-\xi$$ for $r\geq r_-$; the inequality is strict because the solution is non-trivial and hence $\kappa$ does not vanish identically. By (\[alphar1\]), (\[alphaleq1\]), and (\[M-def\]), $$\label{alpha-r} \alpha_{r}(t, r) \le\frac{a_+^2}{r^2}\int_{0}^{r}\left(4\pi (\rho+S)+6 \kappa^2\right) s^2 ds < \frac{4 a_+^2 M}{r^2}\,.$$ Since $\alpha(t, \infty)=1$, this implies that $$\alpha(t, r) > 1-\frac{4 a_+^2 M}{r_-} =\frac{1-3\xi}{1-\xi}.$$ Next we define for $t\in [0, \infty[$, $$\begin{aligned} R_-(t) & = & r_- + \frac{\alpha_-}{a_+} \,\frac{W_-}{\sqrt{1+W_-^2 +L_+/r_-^2}}\,t, \label{R-def} \\ R_+(t) & = & r_+ + \left(1 + \frac{a_+ M}{r_-}\right) t, \label{R+def}\end{aligned}$$ as well as $$\begin{aligned} \label{gamma-def} \gamma & = & \min_{t\in [0, \infty[}\frac{R_-(t)}{R_+(t)} =\min\left\{\frac{r_-}{ r_+},\frac{\alpha_-}{a_+(1+ a_+ M/r_-)} \frac{W_-}{\sqrt{1+W_-^2 +L_+/r_-^2}}\right\} \nonumber \\ & = & \min\left\{\frac{r_-}{ r_+}, \frac{1-3\xi}{ \sqrt{1-\xi}+ \xi/2} \,\eta\right\},\end{aligned}$$ where we recall the abbreviations in (\[xieta-def\]) and (\[gote\]). Let $[0, t_0[$ denote the maximal time interval such that for $t\in [0, t_0[$ and $(r, w, L)\in \supp f(t)$, $$\label{rwbd} R_-(t)<r<R_+(t),\quad w>W_-\,.$$ By (\[master2\]) and (\[supp-as\]) this holds for $t=0$, and by continuity there is a maximal time interval $[0, t_0[$ with $t_0>0$ on which (\[rwbd\]) persists. Suppose that $t_0\in ]0, T[$. Then we must have equality at $t=t_0$ in at least one of the inequalities in (\[rwbd\]). But in the following part of the proof we show that (\[rwbd\]) persists on $[0, t_0]$. Thus necessarily $t_0=T$. Consider a characteristic such that $f(t, r(t), w(t), L) > 0$. Denoting the parameter along the curve by $t$ it is calculated that \[dds1\] & = & - 2 (1-) E w - 4 w\^2 + 6 w\^2\ & & + 2 - 6 L + 2 E w - 2 , and we want to show that this quantity remains positive on the time interval $[0,t_0[$. By (\[j-def\]) and (\[rwbd\]), $\jmath\ge 0$ for $t\in [0,t_0[$. Hence (\[kdef\]) and (\[kap-form\]) lead to the estimate \[betaeq\] (t,r)=- \_0\^r 4a(t,s) (t,s) s\^3 ds0, and we can drop the fifth and seventh term in (\[dds1\]). Using the estimate $$\frac{1}{a^2(t,r)} = 1-\frac{2\mu(t,r)}{r} \geq 1-\frac{2 M}{r}$$ and the fact that $\alpha_r\ge 0$, cf. (\[alphar1\]), we find that \[dds2\] - 2 E w - 4 w\^2 +(3Mw\^2+) + 2 E w. By (\[alphar2\]) and in view of $\alpha(t, s)\le\alpha(t, r)$, $$\alpha_r(t,r)\leq\frac{M}{r^2} (\alpha a^2)(t,r) +\frac{(\alpha a^2)(t,r)}{r^2} \int_0^r\left(4 \pi S + \frac{9}{2} \kappa^2 \right) s^2 ds.$$ Inserting this into (\[dds2\]) and using $a\geq 1$ and (\[betaeq\]), \[dds3\] & & ; notice that the first term coming from the above estimate for $\alpha_r$ is exactly cancelled by the last term in (\[dds2\]), which is crucial. Now the estimates $w^2/E\le w$ and $E\ge 1$ imply that \[Sest\] S(t,r) & & (t,r)+\_[-]{}\^ \_0\^f(t, w, r, L)dLdw\ & & (t, r)+ \_[-]{}\^\_0\^ wf(t, w, r, L)dLdw\ & & (1+)(t,r). By (\[kap-form\]), $\kappa(t,r)=0$ for $r\leq R_-(t)$, and $$\label{kappa-est} |\kappa(t,r)| \leq a_+ \frac{4\pi}{r^2} \int_0^r \jmath(t, s) s^2 ds \leq a_+ \frac{4\pi}{r^2} \int_0^r\rho(t, s) s^2 ds \leq \frac{a_+ M}{r^2}.$$ Let $$J(t,r)=\int_0^r 4\pi \jmath(t,s) s^2 ds \geq 0.$$ Then using the fact that $J$ is increasing in $r$ and $R_-(t)\ge r_-$ for $t\in [0,t_0[$, $$\label{kintest} \int_0^r\kappa^2 s^2 ds \leq a_+^2 M\int_{R_-(t)}^r\frac{1}{s^2} \bigg(\int_0^s 4\pi \jmath \tau^2 d\tau \bigg)\,ds \leq\frac{a_+^2 M}{R_-(t)}\,J \leq\frac{a_+^2 M}{r_-}\,J.$$ By the definition of $\gamma$, see (\[gamma-def\]), \[wrongjest\] \_0\^r 4(t,s)s\^3 ds J(t,r). We insert the estimates (\[Sest\]), (\[kintest\]), and (\[wrongjest\]) into (\[dds3\]) and use the fact that $r\geq R_-(t) \geq r_-$ along the characteristic under consideration to obtain the estimate $$\begin{aligned} \frac{d}{dt}\left[\left(1-\frac{2 M}{r}\right) E^2\right] & \ge & \frac{2\alpha}{r^2} J \biggl[2 \gamma w^2- \left(1+\frac{L_+}{W_-}\frac{1}{r_-^2}\right) E w -\frac{9}{2}\frac{a_+^2 M}{r_-}\,E w \\ & & \hspace{12em} -\left(3 M w^2 +\frac{L_+ }{r_-}\right) \frac{a_+ }{r_-}\biggr].\end{aligned}$$ Due to the estimate $E w \leq E^2 = 1+w^2 +L/r^2$ this implies that $$\begin{aligned} \frac{d}{dt}\left[\left(1-\frac{2 M}{r}\right) E^2\right] &\geq& \frac{2\alpha}{r^2} J\biggl[\left(2\gamma-1 -\frac{9}{2}\frac{a_+^2 M}{r_-}-\frac{L_+ }{W_- r_-^2} -\frac{3a_+ M}{r_-}\right) w^2 \\ && {} -\bigg(1+\frac{L_+ }{W_- r_-^2}+\frac{9}{2}\,\frac{a_+^2 M}{r_-}\bigg) \bigg(1+\frac{L_+ }{r_-^2}\bigg)-\frac{L_+ a_+ }{r_-^2}\biggr].\end{aligned}$$ If we recall the definition of $\gamma$ in (\[gamma-def\]) we see that the condition (\[master1\]) implies that the factor multiplying $w^2$ in the right hand side of this estimate is positive. Hence the estimate is still true if we replace $w$ by $W_-$, since by assumption $w^2>W_-^2$ on the time interval $[0,t_0[$. Again recalling the condition (\[master1\]) we conclude that $$\frac{d}{dt}\left[\left(1-\frac{2 M}{r}\right) E^2\right] > 0$$ as long as $t\in[0,t_0[$. Hence for any characteristic in $\supp f$ on this time interval, $$\left(1-\frac{2 M}{r(t)}\right)\left(1+w^2(t)+\frac{L}{r^2(t)}\right) \ge \left(1-\frac{2 M}{r(0)}\right)\left(1+w^2(0) + \frac{L}{r^2(0)}\right).$$ Thus $$w^2(t)+\frac{L}{r^2(t)} \geq\left(1-\frac{2 M}{r(0)}\right)\left(w^2(0)+\frac{L}{r^2(0)}\right) -\frac{2 M}{r(0)}\,.$$ Since $\dot{r}\ge 0$, $r(t)\ge r(0) > r_-$. Together with $w(0)\ge w_-$ this implies that $$\begin{aligned} w^2(t) & \ge & \left(1-\frac{2M}{r_-}\right) w_-^2 -\frac{2 M}{r_-}\left(1+\frac{L_+}{r_-^2}\right) \\ & = & (1-\xi)\, w_-^2 -\xi\,\left(1+\frac{L_+}{r_-^2}\right)>W_-^2\,,\end{aligned}$$ the latter because of (\[master2\]). Therefore $w(t)>W_-$ for $t\in [0, t_0]$. From (\[betaeq\]) it follows that $$\beta(t,r) \geq - \frac{a_+}{r} \int_0^r 4\pi \jmath s^2\,ds \ge -\frac{a_+ M}{r_-}.$$ Thus by (\[charsys1\]), $$\frac{\alpha_-}{a_+} \frac{W_-}{\sqrt{1+W_-^2 +L_+/r_-^2}} \leq \dot r \leq 1+\frac{a_+ M}{r_-}.$$ Since $r(0)\in ]r_-, r_+[$, this implies that $R_-(t)<r(t)<R_+(t)$ for $t\in [0, t_0]$; recall (\[R-def\]) and (\[R+def\]). Therefore (\[rwbd\]) holds on $[0, t_0]$, and $t_0=T$. [*Step 2: Global existence in maximal-areal coordinates*]{}.\ The next step in the proof of the theorem is to show that the estimates obtained in Step 1 on the maximal existence interval $[0,T[$ imply that $T=\infty$. In [@Rl7 Thm. 2.1] sufficient conditions for global existence in maximal-isotropic coordinates are obtained. It is straightforward to see that these conditions are also sufficient in our case, i.e., if we can show that the support of the momenta and the metric function $a$ do not blow up, then $T=\infty$ follows; the local well-posedness in maximal-areal coordinates and continuation of solutions will be studied more systematically in [@AKR]. At points $(t, r, \theta, \phi)$ with $r<R_{-}(t)$ there is no matter. Hence by (\[kap-form\]), (\[alphar1\]), (\[ee1\]), and (\[bdryc\]) this implies that $$\label{vacuum} a_r(t,r) = \alpha_r (t,r)=\kappa(t,r) = 0\ \mbox{for}\ r<R_-(t).$$ In particular the bounds in (\[gote\]) hold for all $r\geq 0$ and $t \in [0, T[$. In view of the latter it is sufficient to verify that $$\label{Ebound} \sup\{ E \mid (r, w, L)\in \supp f(t),\ t\in [0,T[ \} < \infty.$$ In what follows $C>0$ denotes a constant which only depends on the parameters specified in (\[master1\]) and (\[master2\]) and which may change from line to line. By (\[E\]) and (\[charsys1\])–(\[charsys3\]) it follows that along characteristics, $$\label{dotE} \dot E=-\frac{\alpha_r w}{a} -\frac{\alpha\kappa}{E}\Big(2w^2-\frac{L}{r^2}\Big).$$ Since by (\[rwbd\]) and (\[R-def\]) $r(t)\geq R_-(t)\geq C(1+t)$ for any characteristic in $\supp f$, (\[kappa-est\]) and (\[alpha-r\]) imply that $$|\kappa(t, r)|+ \alpha_{r}(t, r) \leq \frac{C}{(1+t)^2}.$$ We point out that in this particular case when $w>0$ the first term of the right hand side of (\[dotE\]) can be dropped since it is non-positive but this more general estimate will be needed later. Since $\alpha\le 1$ and $a\geq 1$ we thus obtain the estimate $$\label{Edotest} |\dot E|\leq C \frac{E}{(1+t)^2},$$ and Gronwall’s lemma implies the desired bound (\[Ebound\]). Thus $T=\infty$. [*Step 3: Future geodesic completeness.*]{}\ In order to analyze this question polar coordinates such as we use on the spatial $t=\mathrm{const}$ surfaces are not convenient, and we introduce new coordinates $$x^0 = t,\ x^1 = r\sin\theta\cos\phi,\ x^2 = r\sin\theta\sin\phi,\ x^3 = r \cos\theta,$$ which can be thought of as the corresponding Cartesian coordinates. In these coordinates the metric becomes $$g_{00} = - \alpha^2 + a^2 \beta^2,\ g_{0i} = a^2 \beta \frac{x_i}{r},\ g_{ij} = \delta_{ij} + (a^2-1)\frac{x_i x_j}{r^2},$$ where Latin indices $i,j$ run from $1$ to $3$ and $x_i = \delta_{ij}x^j$. Let us now consider an arbitrary future directed, time-like or null geodesic, i.e., a solution $(x^\gamma (s),p^\gamma(s))$ of the geodesic equations $$\begin{aligned} \frac{dx^{\gamma}}{ds} & = & p^{\gamma}, \\ \frac{dp^{\gamma}}{ds} & = & - \Gamma^{\gamma}_{\delta\epsilon}\,p^{\delta}p^\epsilon,\end{aligned}$$ where Greek indices $\gamma,\delta,\epsilon$ run from $0$ to $3$, $\Gamma^{\gamma}_{\delta\epsilon}$ are the Christoffel symbols, and $$p^0 >0,\ g_{\gamma\delta} p^\gamma p^\delta = - m^2 \leq 0;$$ $m^2$ is conserved along any geodesic, and $p^0$ cannot change sign. Such a geodesic exists on a maximally extended interval $[0,s_+[$, and future geodesic completeness means that $s_+=\infty$ for all such geodesics. The following relations between the variables $r$, $w$, $L$, and $p^\gamma$ hold: $$\begin{aligned} E &=& \alpha \, p^0,\\ w &=& \left(\frac{x_i p^i}{r} + \beta p^0 \right)\, a,\\ \frac{L}{r^2} &=& \delta_{ij} p^i p^j - \left(\frac{x_i p^i}{r}\right)^2,\end{aligned}$$ where we now re-define $$E=E(r,w,L)=\sqrt{m^2+w^{2}+L/r^{2}}.$$ That we had $m=1$ before means that our system describes an ensemble of particles of rest mass $1$, but for geodesic completeness we have to consider any time-like or null geodesic, i.e., we have to allow any $m\geq 0$. Since $dt/ds = p^0 > 0$ we can re-parameterize the geodesic by coordinate time $t\in [0,t_+[$. Along the geodesic again (\[dotE\]) holds. By (\[vacuum\]), $\dot{E}=0$ as long as $r \leq R_-(t)$ so that also (\[Edotest\]) holds along the geodesic, and $E\le C$ for some constant $C$. Since by (\[gote\]) $\alpha$ is bounded from below, $p^0$ remains bounded on $[0,t_+[$. Since both $a$ and $\beta$ are bounded as well also the $p^i$ remain bounded, and hence $t_+=\infty$. But since $dt/ds = p^0$ is bounded, this implies that $s_+=\infty$, and the proof of the theorem is complete. [$\Box$]{} [AAAA]{} , The Einstein-Vlasov System/Kinetic Theory, [*Living Rev. Relativ.*]{} [**8**]{} (2005). , A numerical investigation of the stability of steady states and critical phenomena for the spherically symmetric Einstein-Vlasov system. [*Class. Quantum Grav.*]{} [**23**]{}, 3659–3677 (2006). , The spherically symmetric Einstein-Vlasov system in maximal areal coordinates. In preparation. , Propagation of moments and regularity for the 3-dimensional Vlasov-Poisson system, [*Invent. Math.*]{}  [**105**]{}, 415–430 (1991). , Critical phenomena at the threshold of black hole formation for collisionless matter in spherical symmetry, [*Phys. Rev. D.*]{} [**65**]{}, 024007 (2002). , Global classical solutions of the Vlasov-Poisson system in three dimensions for general initial data, [*J. Differential Equations*]{}  [**95**]{}, 281–303 (1992). , The Vlasov-Einstein system with surface symmetry, [*Habilitationsschrift*]{}, Munich (1995). , Global existence of solutions of the spherically symmetric Vlasov-Einstein system with small initial data, [*Commun. Math. Phys.*]{} [**150**]{}, 561–583 (1992). Erratum: [*Commun. Math. Phys.*]{} [**176**]{}, 475–478 (1996). , Critical collapse of collisionless matter—a numerical investigation, [*Phys. Rev. D*]{} [**58**]{}, 044007 (1998). , An introduction to the Einstein-Vlasov system, [*Banach Center Publ. *]{}[**41**]{}, 35–68 (1997).
{ "pile_set_name": "ArXiv" }
--- abstract: 'We address simultaneous inference for mixed parameters which are the key ingredients in small area estimation. We assume linear mixed model framework. Firstly, we analyse statistical properties of a max-type statistic and use it to construct simultaneous prediction intervals as well as to implement multiple testing procedure. Secondly, we derive bands based on the volume-of-tube formula. In addition, we adapt some of the simultaneous inference methods from regression and nonparametric curve estimation and compare them with our approaches. Simultaneous intervals are necessary to compare clusters since the presently available intervals are not statistically valid for such analysis. The proposed testing procedures can be used to validate certain statements about the set of mixed parameters or to test pairwise differences. Our proposal is accompanied by simulation experiments and a data example on small area household incomes. Both of them demonstrate an excellent performance and utility of our techniques.' author: - 'Katarzyna Reluga[^1], María José Lombardía[^2] andStefan Andreas Sperlich[^3]' bibliography: - 'cite.bib' title: '**Simultaneous inference for mixed and small area parameters**' --- \#1 0 [0]{} 1 [0]{} **Simultaneous inference for mixed and small area parameters** [*Keywords:*]{} linear mixed models, max-type statistic, multiple testing, simultaneous confidence intervals, small area estimation, unit level model Introduction {#sec:intro} ============ The family of linear mixed effects models (LMM) developed by [@henderson1950] have been extensively applied in the statistical analysis of correlated or clustered data structures. This modelling framework arises naturally in many fields such as medicine, biology, sociology, etc. Under LMM we suppose that the extra between-cluster variation is captured by cluster-specific random effects. Empirical best predictors (EBLUP) as well as empirical Bayes (EB) estimators are recognized methods to obtain mixed model predictions. They are essential elements of the small area estimation (SAE) which concentrates on the statistically valid analysis of mixed parameters. To assess the accuracy of a predicted mixed parameter, it is crucial to measure its variability. Traditionally one would provide the mean squared error (MSE) or prediction intervals which are more informative from the perspective of practitioners. Both tools were broadly discussed in the literature (cfr., monograph on SAE of [@rao2015small]). However, the focus was almost always on the fixed effects, or cluster-wise inference for the mixed effects, whereas simultaneous inference was largely neglected. Consider the prediction intervals of mixed parameters under a cluster-wise setting; the coverage probability of $100(1-\alpha )$ intervals refers to the mean across all clusters. This implies that, by construction, about $100\alpha$ percent of the provided intervals do not contain the true parameter. As a consequence, these cluster-wise prediction (or often called confidence) intervals are not appropriate for addressing either a joint analysis or a comparison of clusters. Yet, such comparisons are of great interest in many applied domains of SAE, for example, when statistical offices report to policy makers, or within public health centres carrying out studies on demographic groups. The use of uniform methods is highly relevant for practitioners who try to find significant discrepancies between clusters or make decisions on resource re-allocations for different areas. Alternatively, statisticians may prefer to carry out formal significance tests to disprove or support simultaneous hypotheses regarding certain characteristic. We propose statistical tools which serve these two purposes. The aim of this paper is to close the gap between the needs of practitioners and what the present literature provides. We construct simultaneous prediction interval (SPI) for a mixed parameter using the max-type statistic which is readily applicable in the multiple testing (MT) procedure. Despite the unquestionable utility of maximal distribution in the context of LMM, no one to the best of our knowledge, has investigated its theoretical and empirical properties. Our results are different from those derived by [@sun1999confidence] or [@maringwa2008application] within the framework of longitudinal studies. They proposed to apply, respectively, the volume-of-tube formula and Monte Carlo (MC) sampling to construct simultaneous bands for linear combinations of fixed effects only. In contrast, we investigate a more complex problem of looking at mixed effects, i.e., the combination of fixed and random effects. Our proposal also differs from the derivation of [@krivobokova2010simultaneous] who employed a mixed model representation for penalized splines to construct uniform bands for one-dimensional regression curves. Finally, our results are distinct from those of [@ganesh2009simultaneous] who addressed the problem of SPI in SAE from a Bayesian perspective, creating credible bands, but only for the special case of the area level model. On the contrary, we consider a general LMM under frequentist framework. In addition, our contribution in the area of MT is an appealing, practical methodology used under LMM for the first time. Employment of the max-type statistic might be considered as a complement to the study of [@kramlinger2018marginal] who thoroughly investigated the issue of MT and confidence sets based on (asymptotic) chi-square statistics. In the classical literature (without mixed effects), max-type and chi-square statistics have always been considered as complements and are both well established in the practitioners’ toolbox. We believe that it is equally valid for mixed models. The former are more popular for confidence sets as they can provide SPI in Tukey’s sense, whereas chi-square statistics are widely recognized for MT. The remainder of the paper is organized as follows. In Section \[sec:LMM\] we introduce the modelling framework and the parameter of interest. The SPI constructions as well as their consistency are outlined in Section \[sec:SPI\]. A simulation study and a data example are provided in Sections \[sec:simulations\] and \[sec:data\_example\] respectively. The conclusions can be found in Section \[sec:conclusions\]. Some technical details are deferred to Appendix \[sec:appendixA\]. Linear mixed model inference {#sec:LMM} ============================ Consider a classical linear mixed model formulation $\bm{y}=\bm{X\beta}+\bm{Z u}+\bm{e}$, where $\bm{X}$, $\bm{Z}$ are known, full column rank matrices for a fixed and a random part, $\bm{\beta}$ is a vector of fixed effects, $\bm{u}$ is a vector of random effects, and stochastic errors are denoted by $\bm{e}$. We assume $\bm{u}$ and $\bm{e}$ to be independent with $\bm{u}{\overset{ind}{\sim}} N_{q}{(\bm{0},\bm{G})}$ and $\bm{e}{\overset{ind}{\sim}} N_{n}{(\bm{0},\bm{R})}$. More specifically, we investigate LMM with block diagonal covariance matrix (LMMb): $$\label{eq:LMM_b} \bm{y}_d=\bm{X}_d\bm{\beta}+\bm{Z}_d\bm{u}_d+\bm{e}_d,\quad d=1,\dots,D,$$ where $n_d$ is the number of units in the $d^{th}$ cluster (or area), $\bm{y}_d\in \mathbb{R}^{n_d}$, $\bm{X}_d \in \mathbb{R}^{n_d\times(p+1)}$ and $\bm{Z}_d\in\mathbb{R}^{n_d\times q_d}$. Furthermore, $D$ is the total number of clusters, $\bm{\beta}\in\mathbb{R}^{p+1}$ an unknown common vector of regression coefficients, $\bm{u}_d{\overset{ind}{\sim}} N_{q_d}{(\bm{0},\bm{G}_d)}$ and $\bm{e}_d{\overset{ind}{\sim}} \prescript{}{n_d}{(\bm{0},\bm{R}_d)}$, $n=\sum_{d=1}^D n_d$. We assume that $\bm{G}_d=\bm{G}_d(\bm{\theta})\in\mathbb{R}^{q_d\times q_d}$ and $\bm{R}_d=\bm{R}_d(\bm{\theta})\in\mathbb{R}^{n_d\times n_d}$ which depend on variance parameters $\bm{\theta}=(\theta_1,...,\theta_h)^t$. LMM can be easily retrieved applying the notation introduced by [@prased_rao], p. 168. Under this setup we suppose that the variance-covariance matrix $\bm{V}$ is nonsingular $\forall \theta_i$, $i=1,\dots,h$ and $$\label{eq:moments_LMM} \mathbb{E}(\bm{y})=\bm{X\beta} \quad \text{ and } \quad \mathbb{V}\mathrm{ar}(\bm{y})=\bm{R} + \bm{Z G}\bm{Z}^t\eqqcolon\bm{V}(\bm{\theta})=\bm{V}.$$ Two important examples of , extensively used in SAE, are the *nested error regression model* (NERM) of [@battese1988error] and the *Fay-Herriot model* (FHM) of [@fay_herriot]. The former is defined as $$\label{eq:NERM} y_{dj}=\bm{x}^t_{dj}\bm{\beta}+u_d+e_{dj},\quad d=1,\dots,D,\quad j=1,\dots,n_d,$$ where $y_{dj}$ is the quantity of interest for the $j^{th}$ unit in the $d^{th}$ cluster, $\bm{x}_{dj}=(1,x_{dj1},\dots,x_{djp})^t$, $u_d{\overset{iid}{\sim}} N(0,\sigma^2_u)$ and $e_{dj}{\overset{iid}{\sim}} N(0,\sigma^2_e)$ for $d=1,\dots,D$, $j=1,\dots,n_d$. Here $\bm{y}_d=(y_{d1},\dots,y_{dn_d})$, $\bm{X}_d=\mathrm{col}_{1\leqslant j\leqslant n_d}\bm{x}^t_{dj}$, $q_d=1$, $Z_d=\bm{1}_{n_d}$ with $\bm{1}_{n_d}$ a $n_d$ vector of ones, $\bm{e}_d=(e_{d1},\dots,e_{dn_d})^t$, $\bm{\theta}=(\sigma_{e}^2,\sigma_u^2)^t$, $\bm{R}_d(\bm{\theta})=\sigma_{e}^2\bm{I}_{n_d}$ with $\bm{I}_{n_d}$ a $n_d\times n_d$ identity matrix, and $\bm{G}_d(\bm{\theta})=\sigma_{u}^2$. FHM is identified as $$\label{eq:FH} y_{d}=\bm{x}^t_{d}\bm{\beta}+u_d +e_{d},\quad d=1,\dots,D ,$$ where $\bm{x}_{d}=(1,x_{d1},\dots,x_{dp})^t$, $u_d{\overset{iid}{\sim}} N(0,\sigma^2_u)$ and $e_{d}{\overset{iid}{\sim}} N(0,\sigma^2_{e_{d}})$ with $\sigma^2_{e_{d}}$ ($d=1,\dots,D$) being *known*. In this case, $n_d=q_d=1$, $Z_d=1$, $\bm{\theta}=\sigma^2_u$, $\bm{R}_d(\sigma^2_u)=\sigma^2_{e_{d}}$. We are interested in simultaneous inference for a general mixed parameter which is essential in SAE $$\label{eq:mu} \mu_d=\bm{k}^t_{d}\bm{\beta}+\bm{m}^t_{d}\bm{u}_d,\quad d=1,\dots,D,$$ with $\bm{k}_{d}\in \mathbb{R}^{p+1}$ and $\bm{m}_{d}\in \mathbb{R}^{q_d}$ known. In the data example and simulations $\mu_d$ is a cluster conditional mean, but other parameters of interest can be explored as well. [@henderson] developed the best linear unbiased predictor (BLUP) of a linear combination of random effects $\bm{u}$ and fixed effects $\bm{\beta}$ for a completely known covariance matrix $\bm{V}$. Applying his idea we obtain the BLUP estimator for : $$\label{eq:mu_hat_tilde} \tilde{\mu}_d\coloneqq\tilde{\mu}_d(\bm{\theta})=\bm{k}^t_{d}\tilde{\bm{\beta}}+\bm{m}^t_{d}\tilde{\bm{u}}_d, \quad d=1,\dots,D,$$ where $\bm{\theta}=(\theta_1,\dots,\theta_h)^t$, $\tilde{\bm{\beta}}=\tilde{\bm{\beta}}(\bm{\theta})=(\bm{X}^t\bm{V}^{-1}\bm{X})^{-1}\bm{X}^t\bm{V}^{-1}\bm{y}$ and $\tilde{\bm{u}}_d=\tilde{\bm{u}}_d(\bm{\theta})=\bm{G}_d \bm{Z}^t_d \bm{V}_d^{-1}(\bm{y}_d-\bm{X}_d \tilde{\bm{\beta}})$. In practice $\bm{\theta}$ is unknown, hence we use $\hat{\bm{\theta}}\coloneqq\hat{\bm{\theta}}(\bm{y})$ which yields the EBLUP $$\label{eq:mu_hat_e} \hat{\mu}_d := \hat{\mu}_d(\hat{\bm{\theta}})= \bm{k}^t_{d}\hat{\bm{\beta}}+\bm{m}^t_{d}\hat{\bm{u}}_d,\quad d=1,\dots,D$$ with $\hat{\bm{\theta}}=(\hat{\theta}_1,\dots,\hat{\theta}_h)^t$, $\hat{\bm{\beta}}=\hat{\bm{\beta}}(\hat{\bm{\theta}})$ and $\hat{\bm{u}}=\hat{\bm{u}}(\hat{\bm{\theta}})$. Having assumed certain conditions on the distributions of random effects and errors, as well as the variance components $\bm{\theta}$ (see assumptions in Appendix \[sec:RC\]), [@kackar_unbias] proved that the two-stage procedure provides an unbiased estimator for $\mu_d$. To construct a studentized max-type statistic, it is important to assess the variability of prediction. The most common measurement of uncertainty is the mean squared error $\mathrm{MSE}(\hat{\mu}_d)=\mathbb{E}(\hat{\mu}_d-\mu_d)$. Here $\mathbb{E}$ denotes the expectation with respect to model . We can decompose the MSE into three terms $$\label{eq:MSE} \mathrm{MSE}(\hat{\mu}_d )=\mathrm{MSE}(\tilde{\mu}_d )+\mathbb{E}\left(\hat{\mu}_d -\tilde{\mu}_d \right)^2+2\mathbb{E}\left\{(\tilde{\mu}_d -\mu_d)(\hat{\mu}_d -\tilde{\mu}_d )\right\}.$$ $\mathrm{MSE}(\tilde{\mu}_d )$ accounts for the variability when the variance components $\bm{\theta}$ are known. Assuming LMMb and $\bm{b}_d^t=\bm{k}^t_d-\bm{a}_d^t\bm{X}_d$ with $\bm{a}^t_d=\bm{m}^t_d\bm{G}\bm{Z}^t_d\bm{V}^{-1}_d$, $\mathrm{MSE}(\tilde{\mu}_d)$ reduces to $$\label{eq:MSE_b_first_term} \bm{m}^t_d(\bm{G}_d-\bm{G}_d\bm{Z}^t_d \bm{V}_d^{-1}\bm{Z}_d\bm{G}_d)\bm{m}_d+\bm{b}_d^t\left(\sum_{d=1}^{D}\bm{X}^t_{d}\bm{V}^{-1}_d\bm{X}_d\right)^{-1}\bm{b}_d =: g_{1d}(\bm{\theta}) + g_{2d}(\bm{\theta}),$$ where $g_{1d}$ accounts for the variability of $\tilde{\mu}_d$ once $\bm{\beta}$ is known and $g_{2d}$ for the estimator $\tilde{\bm{\beta}}$. The second term in is intractable and there exists a vast literature which investigates a precise estimation of this quantity, cfr., [@rao2015small] for details. As far as the third term is concerned, it actually disappears under normality of errors and random effects; therefore it is rarely considered. Following [@chatterjee2008parametric], we suggest a construction of accurate SPI using only $\bm{g}_1(\hat{\bm{\theta}})=( g_{11}(\hat{\bm{\theta}}),\dots,g_{1D}(\hat{\bm{\theta}}))^t$, where $g_{1d}(\hat{\bm{\theta}})$ is defined in with $\bm{\theta}$ replaced by a consistent estimator. SPI and MT for mixed parameter {#sec:SPI} ============================== Simultaneous confidence intervals (SCI) have been discussed extensively in nonparametric statistics when one is interested in the estimation of model $y_{dj}=m(x_{dj}) + \varepsilon_{dj}$ with $m$ an $r$ times differentiable function, $d=1,\dots, D$ and $j=1,\dots,n_d$. The asymptotic distribution of $ \sup_{a\leqslant x \leqslant b}\left\lvert\hat{m}(x)-m(x)\right\lvert$ has been intensively investigated. [@bickel1973some] considered the distribution of $\sup_{a\leqslant x \leqslant b}|W(x)|$ with $W(x)$ a standard Gaussian process; [@hall1991convergence] proved its poor convergence rate ($(\log n)^{-1}$ with $n$ the number of observations). Therefore [@claeskens2003bootstrap] proposed bootstrap approximations. Another approach is to use the connection between the tail probabilities of the Gaussian random field and the volume-of-tube formula of [@weyl1939volume] as examined by [@krivobokova2010simultaneous]. Various approaches have been put forward to tackle the problem of confidence bands for regression surfaces, i.e., to find critical values $f_1$ and $f_2$ such that $P(f_1\leqslant f(\bm{x})\leqslant f_2, \forall \bm{x}\in \mathcal{X} )=1-\alpha$ where $f(\bm{x})=\beta_0+\beta_1x_1+\dots+\beta_px_p$ and $\mathcal{X}\subseteq \mathbb{R}^p$. The most straightforward tool is to use Bonferroni t-statistics based on the Bonferroni inequality, i.e., to choose the critical value as $\alpha/2(p+1)$ quantile from a t-distribution with $n-(p+1)$ degrees of freedom. In Section \[sec:bonfer\] we show how to adapt Bonferroni intervals to be suitable under LMM setting. On the other hand, [@working1929applications] solved this problem completely for $x\in \mathbb{R}$, i.e., an unconstrained one-dimensional domain. Construction of [@scheffe1953method] type SCIs was also considered; these were originally developed for models with homoscedastic independent errors and they generalize the analysis of [@working1929applications]. The author suggested creating simultaneous intervals for a regression space $f(\bm{x})$, assuming $\bm{x} \in \mathcal{X}\equiv \mathbb{R}^{p}$, i.e., an unconstrained multidimensional domain of interest. Though constructions for infinite domains are mathematically less intricate, in majority of cases we deal with a $p$-dimensional rectangle $\mathcal{X} \subset \mathbb{R}^p$. Thus, any methodology using a constrained region should provide narrower bands, among others the volume-of-tube based approach which has been used under the parametric setting too by [@sun1993tail], [@sun1994simultaneous] or [@sun1999confidence]. Last but not least, one can obtain Bayesian simultaneous credible bands applying Markov Chain MC. These are conceptually different from frequentist bands. We concentrate on the construction of SPI and MT procedures for the mixed parameter considering a confidence region $\mathcal{I}_{1-\alpha}=\bigtimes_{d=1}^{D}\mathcal{I}_{d,1-\alpha}$ such that $P(\mu_d \in \mathcal{I}_{1-\alpha}\; \forall d\in[D])=1-\alpha$, $[D]=\{1,\dots,D\}$. This is equivalent to finding a critical value $c_{S_0}(1-\alpha)$ which satisfies $$\label{eq:S_proba} \alpha={P}\left(\left\lvert \frac{\hat{\mu}_d - \mu_d}{\hat{\sigma}(\hat{\mu}_d )} \right\rvert \geqslant c_{S_0}(1-\alpha) \; \forall d\in[D]\right) ={P}\left(\max_{d=1,\dots, D}\left\lvert \frac{\hat{\mu}_d - \mu_d}{\hat{\sigma}(\hat{\mu}_d )} \right\rvert \geqslant c_{S_0}(1-\alpha) \right).$$ The critical value $c_{S_0}(1-\alpha)$ is in fact the $(1-\alpha)^{th}$-quantile of the studentized statistic $$\label{eq:s} S_0\coloneqq \max_{d=1,\dots,D}\left\lvert S_{0d} \right\rvert,\quad S_{0d}=\frac{\hat{\mu}_d - \mu_d}{\hat{\sigma}(\hat{\mu}_d )} ,$$ that is, $$\label{eq:c_S_0} c_{S_0}(1-\alpha)\coloneqq \inf\{t\in \mathbb{R}: \mathrm{P}(S_0\leqslant t)\geqslant 1 -\alpha\}.$$ It follows that with probability $1-\alpha$, a region defined as $$\label{eq:unif_band_S} \mathcal{I}^{S}_{1-\alpha}= \bigtimes_{d=1}^D \mathcal{I}^{S}_{d,1-\alpha}, \quad\text{where}\quad\mathcal{I}^{S}_{d,1-\alpha}= \left[\hat{\mu}_d \pm c_{S_0}(1-\alpha) \hat{\sigma}(\hat{\mu}_d ) \right],$$ covers all mixed parameters. Since the probability density function (pdf) of $S_0$ is right skewed, we need to consider its upper quantile and construct symmetric $\mathcal{I}^{S}_{d,1-\alpha}$, $d\in [D]$. This approach might be regarded as a variation of the studentized maximum modulus method of [@tukey1952various; @tukey1953u] and [@roy1953simultaneous]. Our methodology is readily applicable for hypothesis testing. Consider the problem $$\label{eq:test_proc} H_0:\bm{A\mu}=\bm{h}\quad vs.\quad H_1:\bm{A\mu}\neq\bm{h},$$ $\bm{A}\in \mathbb{R}^{D'\times D}$ with $D'\leqslant D$ and $\bm{h}\in \mathbb{R}^{D'}$ a vector of constants. A test based on the max-type statistic $t_{H}$ rejects $H_0$ at the $\alpha$-level if $t_{H}\geqslant c_{H_0}(1-\alpha)$ with $c_{H_0}(1-\alpha)\coloneqq \inf\{t\in \mathbb{R}:\mathrm{P}(S_{H_0}\leqslant t)\geqslant 1-\alpha \}$, where $$\label{eq:mult_test_quant} t_{H}\coloneqq \max_{d=1,\dots, D'}\left \lvert t_{H_d} \right\rvert, \ S_{H_0}\coloneqq \max_{d=1,\dots,D}\left\lvert S_{H_0d} \right\rvert, \ t_{H_d} = \frac{\hat{\mu}^H_d-h_d}{\hat{\sigma}(\hat{\mu}^H_d)} \;\text{and}\; S_{H_0d}=\frac{\mu^H_d - h_d}{\hat{\sigma}(\mu^H_d )},$$ $\bm{\mu}^H=(\mu^H_1,\dots,\mu^H_{D'})^t\coloneqq\bm{A}\bm{\mu}\in \mathbb{R}^{D'}$ and $\hat{\bm{\mu}}^H$ its estimated counterpart. In other words, $\bm{h}\notin \mathcal{I}_{1-\alpha}^{H_0}$ with $$\mathcal{I}_{1-\alpha}^{H_0}= \bigtimes_{d=1}^D \mathcal{I}_{d,1-\alpha}^{H_0}, \quad\text{where}\quad \mathcal{I}_{d,1-\alpha}^{H_0}= \left[\hat{\mu}^H_d \pm c_{H_0}(1-\alpha) \hat{\sigma}(\hat{\mu}^H_d )\right].$$ In practice, a standard problem is to test for statistical differences with respect to some characteristic, see an example from the empirical study in Section \[sec:data\_example\]. Our test is based on the single step procedure and exhibits a weak control of a family-wise error (FWER). If one aims at testing multiple hypotheses with a strong control of FWER, then the step-down technique of [@romano2005exact] can be implemented. Appendix \[sec:Appendix\_step\_down\] includes its detailed description and some numerical results within the context of LMM. Since $\mu_d$, $d=1,\dots, D$ is unknown, we need to approximate the critical values in the above equations to obtain operational prediction intervals or testing procedures. In two following subsections we develop bootstrap and the volume-of-tube based approaches to tackle this problem. Afterwards we investigate three alternatives which are the extensions of the existing methods used in the context of regression and nonparametric curve fitting. Bootstrap procedure {#sec:bootstrap_SPI} ------------------- A straightforward way to approximate $c_{S_0}(1-\alpha)$ is to use bootstrap which circumvents a direct application of the normal asymptotic distribution. It can also provide much faster convergence, cfr., [@chatterjee2008parametric]. Let $B$ be the number of bootstrap samples $(\bm{y}^{*(b)},\bm{X}, \bm{Z})$, $b=1,\dots, B$, and $$\label{eq:s_crit_bbot} S^{*(b)}_B\coloneqq\max_{d=1,\dots,D}\left\lvert S^{*(b)}_{Bd} \right\rvert, \quad S^{*(b)}_{Bd} =\frac{\hat{\mu}^{*(b)}_d-\mu^{*(b)}_d} {\hat{\sigma}^*(\hat{\mu}^{*(b)}_d)},$$ the bootstrap analogues of . We will show that the critical value might be successfully estimated by the $(1-\alpha)^{th}$-quantile of , i.e., $$\label{eq:c_S_B} c_{BS}(1-\alpha)\coloneqq \inf\{t^*\in \mathbb{R}: \mathrm{P}(S^*_B\leqslant t^*)\geqslant 1 -\alpha\}.$$ Then, the bootstrap SPI is defined as $$\label{eq:unif_band_boot} \mathcal{I}^{BS}_{1-\alpha}= \bigtimes_{d=D}^D \mathcal{I}^{BS}_{d,1-\alpha}, \quad \text{where}\quad \mathcal{I}^{BS}_{d,1-\alpha}=\left[\hat{\mu}_d \pm c_{BS}(1-\alpha) \hat{\sigma}(\hat{\mu}_d )\right].$$ The choice of $\hat{\sigma}(\hat{\mu}_d)=\sqrt{g_{1d}(\hat{\bm{\theta}})}$ is motivated by the asymptotic analysis of [@chatterjee2008parametric]. Validity of the above bootstrap method is shown following Theorem 3.1 of these authors (henceforth Theorem CLL, given in Appendix \[sec:prop\_1\]) and a Gaussian approximation result (GAR) of [@chernozhukov2013gaussian]. For further technical details consult the proof of Proposition \[prop:con\_max\] in Appendix \[sec:prop\_1\]. \[prop:con\_max\] Suppose assumptions of Theorem CLL hold and $B(\log(Dn))/n\leq C_1 n^{-c_1}$ for some constants $B$, $c_1$, $C_1>0$. Let $\mathrm{P}^*$ denote the probability measure induced by parametric bootstrap. Then, $$\sup\limits_{q\in\mathbb{R}}\left \lvert\mathrm{P}^*\left\{ \max_{d=1,\dots, D} w(S^*_{Bd})\leqslant q \right\} - \mathrm{P}\left\{ \max_{d=1,\dots, D} w(S_{0d})\leqslant q \right\} \right\rvert=o_P(1)$$ for $w(t)=t$, $w(t)=-t$ and $w(t)=|t|$. An important implication of Proposition \[prop:con\_max\] is the coverage probability of $\mathcal{I}_{1-\alpha}^{BS}$. \[cor:convergence\_interval\] Under Proposition \[prop:con\_max\] it holds that $$P\left( \mu_d \in \mathcal{I}^{BS}_{1-\alpha}\;\forall d\in[D]\right)\xrightarrow[]{n\rightarrow \infty} 1-\alpha \ .$$ According to the theoretical developments for max-type statistics, the Kolmogorov distance defined in Proposition \[prop:con\_max\] converges to $0$ at best at polynomial rate $(\log(\cdot))^{c_2}/n^{c_3}$, where $(\cdot)$ is the number of parameters for which we wish to obtain the maximum (in our case $D$), and $c_2, c_3$ some constants, see e.g. [@chernozhukov2013gaussian] and [@dezeure2017high]. We are not aware of results for max-type statistics that one could employ to obtain second order correctness as specified in [@chatterjee2008parametric] without a $\log(\cdot)$ term. Observe that a slight modification of the bootstrap distribution $S^*_B$, i.e., $$S^{*(b)}_{BH_0}\coloneqq\max_{d=1,\dots,D}\left\lvert S^{*(b)}_{BH_0d} \right\rvert, \quad S^{*(b)}_{BH_0d} =\frac{\hat{\mu}^{*H(b)}_d-\mu^{*H(b)}_d} {\hat{\sigma}^*(\hat{\mu}^{*H(b)}_d)},$$ with $$\bm{\mu}^{*H(b)}=(\mu^{*H(b)}_1,\dots,\mu^{*H(b)}_{D'})^t\coloneqq\bm{A}\bm{\mu}^{*(b)} \in \mathbb{R}^{D'} .$$ and its estimated version $$\hat{\bm{\mu}}^{*H(b)}=(\bm{a}^t_1(\bm{k}^t_{1}\hat{\bm{\beta}}^{*(b)}+\bm{m}^t_{1}\hat{\bm{u}}_1^{*(b)}),\dots, \bm{a}^t_1(\bm{k}^t_{D}\hat{\bm{\beta}}^{*(b)}+\bm{m}^t_{D}\hat{\bm{u}}_D^{*(b)}) )^t\coloneqq\bm{A}\hat{\bm{\mu}}^{*(b)}$$ with $\bm{a}_d\in\mathbb{R}^D$ the rows of $\bm{A}$, can be applied to find a bootstrap approximation $c_{BH_0}(1-\alpha)\coloneqq \inf\{t\in \mathbb{R}:\mathrm{P}(S^*_{BH_0}\leqslant t)\geqslant 1-\alpha \}$ of the critical value $c_{H_0}(1-\alpha)$, and be used in a test. In Appendix \[sec:Appendix\_step\_down\] we show how one can use Proposition \[prop:con\_max\] in the step-down procedure. An indisputable advantage of the bootstrap SPIs is their generality: as soon as we can mimic a data generating process for our model, it can be applied and easily implemented to construct SPI and carry out MTs for any kind of estimator. On the other hand, it is more computer intensive than any analytical derivation. The volume-of-tube procedure {#sec:vot_sec} ---------------------------- Consider LMMb defined in ; one way to obtain BLUP estimates for $\bm{\beta}$ and $\bm{u}$ is to solve the so called “mixed model equations” of [@henderson1950] which can be written in matrix form $$\label{eq:mom_eq_mat} \begin{bmatrix} \bm{X}^t\bm{R}^{-1}\bm{X} & \bm{X}^t\bm{R}^{-1}\bm{Z} \\ \bm{Z}^t\bm{R}^{-1}\bm{X} & \bm{Z}^t\bm{R}^{-1}\bm{Z} + \bm{G}^{-1} \\ \end{bmatrix} \begin{bmatrix} \tilde{\bm{\beta}}\\ \tilde{\bm{u}} \end{bmatrix} = \begin{bmatrix} \bm{X}^t\bm{R}^{-1}\bm{y}\\ \bm{Z}^t\bm{R}^{-1}\bm{y} \end{bmatrix}.$$ This can be expressed following the notation of [@gilmour1995average], that is $$\label{eq:mom_eq_compact} \bm{K}\tilde{\bm{\phi}}=\bm{C}^t\bm{R}^{-1}\bm{y},$$ with $\tilde{\bm{\phi}}=\left(\tilde{\bm{\beta}}^t,\tilde{\bm{u}}^t\right)^t$, $\bm{C}=\left[\bm{X}\:\bm{Z}\right]$, $\bm{K}=\bm{C}^t\bm{R}^{-1}\bm{C}+\bm{G}^{+}$ with $\bm{G}^{+}=\begin{bmatrix}\bm{0}_{(p+1)\times (p+1)} & \bm{0}_{(p+1)\times D}\\\bm{0}_{D\times (p+1)} & \bm{G}^{-1}_{D\times D}\end{bmatrix}$. From we obtain a straightforward formula for the estimates $\tilde{\bm{\phi}}=\bm{K}^{-1}\bm{C}^t\bm{R}^{-1}\bm{y}$ and $\hat{\bm{\phi}}=\tilde{\bm{\phi}}(\hat{\bm{\theta}})$. For some $\bm{x}=\left(1,x_1,\dots,x_{p}\right)^t$ with $x_1,\dots,x_p\in \mathcal{X}\subset \mathbb{R}^{p}$, $\bm{z}=\left(z_{1},\dots,z_q\right)^t\in\mathcal{Z}\subset \mathbb{R}^q$ and $\bm{c}=(\bm{x}^t,\bm{z}^t)^t\in\mathcal{X}\times\mathcal{Z}\eqqcolon\mathcal{C}$ we have $$\begin{aligned} & \bm{x}^t\bm{\beta}+\bm{z}^t\bm{u}= \bm{c}^t\tilde{\bm{\phi}}\equiv\bm{l}(\bm{x},\bm{\theta})^t\bm{y}=\sum_{i=1}^{n}\bm{l}_i(\bm{x},\bm{\theta})y_i , \mbox{ where } & \\ & \bm{l}^t(\bm{x},\bm{\theta})=(\bm{l}_1(\bm{x},\bm{\theta}),\bm{l}_2(\bm{x},\bm{\theta}),\dots,\bm{l}_n(\bm{x},\bm{\theta}))^t =\bm{c}^t(\bm{C}^t\bm{R}^{-1}\bm{C}+\bm{G}^{+})^{-1}\bm{C}^t\bm{R}^{-1} , &\end{aligned}$$ i.e., $\bm{l}(\bm{x},\bm{\theta})$ is an $n$-vector, and the BLUP fitted values are $\tilde{\bm{y}}=\bm{L}\bm{y}$ where $\bm{L}=\bm{C}(\bm{C}^t\bm{R}^{-1}\bm{C}+\bm{G}^{+})^{-1}\bm{C}^t\bm{R}^{-1}$ which is also called a ridge regression formulation of the BLUP. Having reformulated the LMMb, we extend the approach of [@faraway1995simultaneous], [@sun1999confidence] and [@krivobokova2010simultaneous]. They used, under simpler settings (cfr., comments in Introduction \[sec:intro\]), the volume-of-tube formula to approximate the tail probabilities of the Gaussian random field, and obtain an approximation of the critical value $c_{S_0}(1-\alpha)$. Observe that $\bm{l}^t(\bm{x},\bm{\theta})=\bm{c}^t\left(\bm{C}^t\bm{C}+\sigma_{e}^{2}\bm{G}^{+}\right)^{-1}\bm{C}^t$. Assuming normality for errors and random effects we have $$\label{eq:Z} Z=\frac{\bm{c}^t\left(\tilde{\bm{\phi}}-\bm{\phi}\right)}{\sqrt{\mathbb{}\mathbb{V}\mathrm{ar} \left\{\bm{c}^t\left(\tilde{\bm{\phi}}-\bm{\phi}\right)\right\} }} =\frac{\bm{c}^t\left(\tilde{\bm{\phi}}-\bm{\phi}\right)}{\sqrt{\sigma_e^2 \bm{c}^t \left(\bm{C}^t\bm{C}+\sigma_{e}^{2}\bm{G}^{+}\right)^{-1} \bm{c} }}\sim\mathrm{N}(0,1).$$ We conclude that $Z$ is a nonsingular Gaussian random field with mean 0 and variance 1. Consequently, the following expressions can be retrieved from $$\begin{aligned} \label{eq:l_M_e_M} \bm{l}_M(\bm{x},\bm{\theta})\coloneqq\left(\bm{C}^t\bm{C}+\sigma^{2}_e\bm{G}^{+}\right)^{-1/2} \bm{c} & & \text{and} & & \bm{e}_M(\bm{x},\bm{\theta})\coloneqq\left(\bm{C}^t\bm{C}+\sigma^{2}_e\bm{G}^{+}\right)^{1/2}\left(\tilde{\bm{\phi}}-\bm{\phi}\right)\end{aligned}$$ such that ${\left\lVert\bm{l}_M(\bm{x},\bm{\theta})\right\rVert}=\bm{c}^t \left(\bm{C}^t\bm{C}+\sigma^{2}_e\bm{G}^{+}\right)^{-1} \bm{c}$. The problem of finding a $(1-\alpha )$ SPI for $\hat{\mu}_d $  $\forall d\in [D]$ boils down to the choice of a critical value $c_{VT}(1-\alpha)$ $$\label{eq:alpha_first} \alpha=\mathrm{P}\left(\left\lvert\hat{\mu}_d - \mu_d\right\lvert \geqslant c_{VT}(1-\alpha)\hat{\sigma}_{e}{\left\lVert\bm{l}_M(\bm{x},\hat{\bm{\theta}})\right\rVert},\text{ for some }\bm{c}\in \mathcal{C}\right).$$ Let $\bm{l}_M=\bm{l}(\bm{x},\bm{\theta})^t_M$, $\hat{\bm{l}}_M=\bm{l}(\bm{x},\bm{\hat{\theta}})_M^t$, $\bm{e}_M=\bm{e}_M(\bm{x},\bm{\theta})$, $\hat{\bm{e}}_M=\bm{e}_M(\bm{x},\hat{\bm{\theta}})$ and $\lambda_M=\hat{\bm{e}}_M-\bm{e}_M$. \[prop:P\_tube\] Suppose that $\sigma^2_e$ is estimated by some consistent estimator. Define $$\begin{aligned} \label{eq:constants} \mathcal{Q}= \frac{\bm{l}_M}{{\left\lVert\bm{l}_M\right\rVert}},& & & & \xi=\inf\limits_{\bm{c}\in\mathcal{C}}\frac{||\hat{\bm{l}}_M||}{||\bm{l}_M||} & & \text{and} & & \eta=\sup\limits_{\bm{c}\in\mathcal{C}}\frac{\left|\left(\hat{\bm{l}}_M-\bm{l}_M\right)^t\bm{e}_M + \hat{\bm{l}}_M(\hat{\bm{e}}_M-\bm{e}_M) \right|}{ \sigma_e||\bm{l}_M||}. \end{aligned}$$ We assume that $\exists$ $\xi_{0}>0$, $\eta_{0}>0$ such that $\mathrm{P}(\xi\leqslant\xi_{0})=o(\alpha)$ and $\mathrm{P}(\eta\leqslant\eta_{0})=o(\alpha)$ as $n\rightarrow\infty$ and $\alpha\rightarrow 0$. Thus one can approximate as follows:\ for $p=1$ $$\begin{aligned} \alpha & \leqslant & \frac{\kappa_0}{\pi}\left[\left\{ 1+\frac{c_{VT}(1-\alpha)^{2}\xi_{0}^2}{\nu} \right\}^{-\nu/2}+ \eta_0\frac{2^{1/2}c_{VT}(1-\alpha)\xi_{0}\Gamma\left(\frac{\nu+1}{2}\right)} {\nu^{1/2}\Gamma\left(\frac{\nu}{2}\right)} \right. \\ & & \left. \left\{1+\frac{c_{VT}(1-\alpha)^2\xi_{0}^{2}}{\nu} \right\}^{-(\nu+1)/2}\right] + \mathcal{E} \mathrm{P}\{|t_v|> c_{VT}(1-\alpha)\xi_{0}\} ; \end{aligned}$$ for $p=2$ $$\begin{aligned} \alpha & \leqslant & \frac{\kappa_0}{ 2^{1/2} \pi^{3/2} }\left[ \frac{2^{1/2} c_{VT}(1-\alpha)\xi_{0}\Gamma\left(\frac{\nu+1}{2}\right)} {\nu^{1/2}\Gamma\left(\frac{\nu}{2}\right)} \left\{1+\frac{c_{VT}(1-\alpha)^2\xi_{0}^{2}}{\nu} \right\}^{-(\nu+1)/2} - \eta_0\frac{ c_{VT}(1-\alpha)\xi_{0}}{\nu^{1/2}} \right. \\ & & \left\{ 1+\frac{c_{VT}(1-\alpha)^{2}\xi_{0}^2}{\nu} \right\}^{-\nu/2} + \left. \eta_0 \frac{ 2 c_{VT}(1-\alpha) \xi_{0}^{2} \Gamma\left(\frac{\nu+2}{2}\right) }{\nu \Gamma\left(\frac{\nu}{2}\right)} \left\{1+\frac{c_{VT}(1-\alpha)^2\xi_{0}^{2}}{\nu} \right\}^{-(\nu+2)/2} \right] \\ & &+\frac{\zeta_0}{2\pi} \left[ \left\{ 1+\frac{c_{VT}(1-\alpha)^{2}\xi_{0}^2}{\nu} \right\}^{-\nu/2} + \eta_0 \frac{2^{1/2} c_{VT}(1-\alpha)\xi_{0}\Gamma\left(\frac{\nu+1}{2}\right)} {\nu^{1/2}\Gamma\left(\frac{\nu}{2}\right)} \right] + 2\mathcal{E}\cdot \mathrm{P}(|t_v|> c_{VT}(1-\alpha) \xi_0); \end{aligned}$$ for $p\geqslant3$ $$\begin{aligned} \alpha & \leqslant & \frac{\kappa_0 \Gamma ( \frac{p+1}{2} )}{\pi^{(p+1)/2}} \mathrm{P} \left\{ \mathrm{F}_{p+1,\nu} \geqslant \frac{\left( c_{VT}(1-\alpha)\xi_0 -\eta_{0} \frac{\sigma_e}{\hat{\sigma}_e} \right)^2}{p+1} \right\} + \frac{\zeta_0}{2} \frac{\Gamma ( \frac{p}{2} )}{\pi^{p/2}} \mathrm{P} \left\{ \mathrm{F}_{p,\nu} \geqslant \frac{\left( c_{VT}(1-\alpha)\xi_0 -\eta_{0} \frac{\sigma_e}{\hat{\sigma}_e} \right)^2 }{p} \right\} \\ & & + \frac{\kappa_2+\zeta_1+m_0}{2\pi} \frac{\Gamma ( \frac{p-1}{2} )}{\pi^{(p-1)/2}} \mathrm{P} \left\{\mathrm{F}_{p-1,\nu} \geqslant \frac{\left( c_{VT}(1-\alpha)\xi_0 -\eta_{0} \frac{\sigma_e}{\hat{\sigma}_e} \right)^2 }{p-1} \right\} ; \end{aligned}$$ where $t_{\nu}$ is a t-distributed random variable with $\nu$ degrees of freedom, $F_{d_1,d_2}$ an F-distributed random variable with parameters $d_1$ and $d_2$, $\kappa_0=\int_{\bm{c}\in\mathcal{C}}^{}{\left\lVert\mathcal{Q}'(\bm{x})\right\rVert}d{\bm{x}}$ the volume of the manifold $\mathcal{M}=\left\{\mathcal{Q}(\bm{c}),\bm{c}\in \mathcal{C}\right\}$, $\zeta_0$ the boundary area of $\mathcal{M}$. Further, $\kappa_2$ and $\zeta_1$ measure the curvatures of $\mathcal{M}$ and $\partial\mathcal{M}$ respectively, whereas $m_0$ measures the rotation angles of $\partial^2 \mathcal{M}$, and finally, $\mathcal{E}$ is the Euler-Poincaré characteristic of $\mathcal{M}$. The proof of Proposition \[prop:P\_tube\] is included into Appendix \[sec:Appendix\_Prop\_2\]. Having retrieved the critical value, we can construct the volume-of-tube SPI $$\label{eq:unif_band_VoT} \mathcal{I}^{VT}_{1-\alpha}= \bigtimes_{d=1}^D \mathcal{I}^{VT}_{d,1-\alpha}, \quad \text{where}\quad \mathcal{I}^{VT}_{d,1-\alpha}= \left[\hat{\mu}_d \pm c_{VT}(1-\alpha)\hat{\sigma}_{e}||\hat{\bm{l}}_M||\right].$$ The approximations from Proposition \[prop:P\_tube\] are a little conservative, i.e., the coverage probability is higher than the nominal $1-\alpha$ (although still lower than for Sheffé’s bands); it approaches $1-\alpha$ as $n\rightarrow\infty$ and $\alpha\rightarrow 0$. We immediately see that the approximation formulas contain several constants. Numerical approximation of the constants describing the geometry of the manifold $\mathcal{M}$ may not pose a major problem, but it is not clear how we should estimate $\xi_0$ and $\eta_0$ under LMMb. Some ideas were derived for simpler one-dimensional models. [@sun1999confidence] proposed a derivative and a perturbation method to estimate correction constant $\xi_{0}$, whereas [@sun1994simultaneous] suggested estimating $\eta_0$ nonparametrically. It is unclear, though, how to extend their implementations for LMM setting. Bootstrap approximation can be regarded as an alternative. However, in this case it would be easier to use bootstrap directly as described in Section \[sec:bootstrap\_SPI\]. Finally, the application of the volume-of-tube formula results in two sources of errors, from the approximation itself and from the estimation of the constants, making the approximation less reliable. These conclusions led us to a technique based on MC sampling in Section \[sec:MC\_SPI\]. Alternative methods {#sec:other_methods} ------------------- ### Monte Carlo procedure {#sec:MC_SPI} Consider mixed model equations from Section \[sec:vot\_sec\]. [@ruppert2003semiparametric] proposed a simple numerical approach to construct confidence bands of one-dimensional nonparametric curves using an empirical approximation of , i.e., $$\label{eq:rand_norm_app} \begin{bmatrix} \hat{\bm{\beta}}-\bm{\beta} \\ \hat{\bm{u}}-\bm{u} \end{bmatrix} {\overset{approx}{\sim}} N\left\{ \bm{0},\left(\bm{C}^t\hat{\bm{R}}^{-1}\bm{C}+\hat{\bm{G}}^{+}\right)^{-1}\right\}.$$ We apply to simulate the distribution of and set $$\label{eq:hat_s} S_0=\max_{d=1,\dots,D}\left\lvert S_{0d} \right\rvert \approx \max_{d=1,\dots,D} \frac {\left|\bar{\bm{c}}_d^t \begin{bmatrix} \hat{\bm{\beta}}-\bm{\beta} \\ \hat{\bm{u}}_d-\bm{u}_d \end{bmatrix} \right|}{\hat{\sigma}(\hat{\mu}_d )} \eqqcolon \max_{d=1,\dots,D}\left\lvert S_{MCd} \right\rvert=S_{MC},$$ where $\bar{\bm{c}}_d=(\bm{k}_d^t,\bm{m}_d^t)^t$. Afterwards, we draw $K$ realisations from and estimate the critical value $c_{S_0}(1-\alpha)$ by the $([(1-\alpha)K]+1)^{th}$ order statistic of $S_{MC}$. Finally, we construct MC SPI $$\label{eq:unif_band_MC} \mathcal{I}^{MC}_{1-\alpha}= \bigtimes \mathcal{I}^{MC}_{d,1-\alpha} \quad \text{where}\quad \mathcal{I}^{MC}_{d,1-\alpha}= \left[\hat{\mu}_d \pm c_{MC}(1-\alpha) \hat{\sigma}(\hat{\mu}_d )\right].$$ Similarly to the bootstrap method we can obtain a critical value for MT. The consistency of $\mathcal{I}^{MC}_{1-\alpha}$ follows immediately from which is a standard result for mixed models and $Z=N(0,1)+O(n^{-1/2})$. GAR of [@chernozhukov2013gaussian] might be invoked to prove the consistency for the maxima. MC SPI are easy to implement and less computer intensive than bootstrap. Yet, it is limited to normally distributed random effects and errors. Since the normal approximation does not hold for non-linear models, MC SPI might be applied under LMM only. ### Bonferroni procedure {#sec:bonfer} Classical simultaneous inference has been considered via Bonferroni adjustment. Since $(\tilde{\mu}_d-\mu_d)/\sigma(\tilde{\mu}_d)$ is a Gaussian pivot, the critical value to construct SPI or MT is selected as $c_{BO}(1-\alpha)=\Phi^{-1}(1-\alpha/2D)$. Having retrieved the value of interest, Bonferroni SPIs are $$\label{eq:unif_band_BO} \mathcal{I}^{BO}_{1-\alpha}= \bigtimes_{d=1}^D \mathcal{I}^{BO}_{d,1-\alpha}, \quad\text{where}\quad \mathcal{I}^{BO}_{d,1-\alpha}= \left[\hat{\mu}_d \pm c_{BO}(1-\alpha) \hat{\sigma}(\hat{\mu}_d )\right].$$ The same critical value might be used in MT procedure , yet it provides a weak control of FWER. Using Bonferroni methodology we do not approximate the true distribution of statistics $S_0$ in . Hence, we assume that $\sqrt{\mathrm{mse}(\hat{\mu}_d)}=\hat{\sigma}(\hat{\mu}_d)$, which is an estimated version of MSE defined in . An application of this procedure is extremely simple and does not require any computational effort. Therefore it will be our benchmark under independence. However, the results of [@romano2005exact] confirm that the Bonferroni-Holm method performs poorly for correlated random variables, e.g., when allowing for spatio-temporal dependence of random effects and/or errors as in [@hobza2018small]. We use a quantile from the normal instead of the t-distribution, because the number of mixed parameters is allowed to grow to infinity, and the latter distribution converges to the former, cfr., the high-dimensional regression setting in [@chernozhukov2013gaussian]. ### Beran procedure [@beran1988balanced] developed a procedure to obtain balanced simultaneous intervals with an overall coverage probability $1-\alpha$ within the context of the models without random effects. His technique is based on so called *roots* and bootstrapping to approximate their cumulative distribution functions (cdfs). We can follow Beran’s methodology and evaluate its performance under LMM. Suppose that $S_d$ is a root and consider $\max_{d=1,\dots, D}S_d$ as defined in . Let $F_{S_d}$ and $F_{S}$ be their respective cdfs. Furthermore, we denote with $F^{-1}_{S_d}(a) $ and $F^{-1}_{S}(a)$ the largest $a^{th}$ quantiles of $F_{S_d}$ and $F_{S}$. [@beran1988balanced] suggested bootstrap approximations $F^*_{S_d}$ and $F^*_{S}$ to obtain $D$ critical values defined as $c_{BEd}(1-\alpha)=F^{*-1}_{S_d}\{F^{*-1}_{S}(1-\alpha)\}$. The details can be found in his paper. The SPI is then $$\label{eq:unif_band_Be} \mathcal{I}^{BE}_{1-\alpha}= \bigtimes_{d=1}^D \mathcal{I}^{BE}_{d,1-\alpha}, \quad \text{where} \quad \mathcal{I}^{BE}_{d,1-\alpha} =\left[\hat{\mu}_d \pm c_{BEd}(1-\alpha) \hat{\sigma}(\hat{\mu}_d )\right].$$ Critical values $c_{BEd}(1-\alpha)$ are not directly applicable for MT. Moreover, Beran’s method is equally computer intensive as bootstrap SPI, but in comparison to the former it might provide a poor coverage rate as its convergence in sup-norm is not guaranteed, cfr., results in Section \[sec:simulations\]. Simulation study {#sec:simulations} ================ A simulation study is carried out to investigate the finite sample properties of bootstrap (BS), Monte Carlo (MC), Beran (BE) and Bonferroni (BO) intervals as well as to evaluate the empirical power of MT for various scenarios. In particular, we analyse the results under NERM and FHM which are prominent examples of LMM used in SAE. As far as the former is taken into account, we set $x_{dj1}=1$, $x_{dj2}\sim U(0,1)$ $\forall$ $d\in[D]$ and $j\in[n_d]$, whereas under FHM $x_{d1}=1$, $x_{d2}\sim U(0,1)$ $\forall d\in [D]$ with $\bm{\beta}=(1,1)^t$ for both models. When it comes to the estimators of $\hat{\sigma}(\hat{\mu}_d)$, [@prased_rao] derived simplified versions of $g_{1d}$ defined in which are given by $g^N_{1d}(\hat{\bm{\theta}})=\hat{\sigma}^2_{u}/(\hat{\sigma}^2_{u}+\hat{\sigma}^2_{e}/n_d)(\hat{\sigma}^2_{e}/n_d)$ and $g^F_{1d}(\hat{\bm{\theta}})=\hat{\sigma}^2_{u}\sigma^2_{e}/(\hat{\sigma}^2_{u}+\sigma^2_{e})$ for NERM and FHM respectively. The total number of simulation runs is $I=2500$ and within each we draw $B=1000$ bootstrap samples. In addition, $D\in \{15,30,60,90\}$, i.e., we consider small to medium sample sizes. Once NERM is considered, we set $n_d=5$ $\forall d\in [D]$, $e_{dj}\sim N(0,\sigma^2_e)$, $u_{d}\sim N(0,\sigma^2_u)$ such that the intraclass correlation coefficient (ICC) equals 1/3, 1/2 or 2/3 (the exact numbers are given in the first column of Table \[tab:simulations\]). Since the results hardly differ when estimating $\bm{\theta}$ using REML or MM, we skip the latter. Regarding FHM, we apply a similar setting as in [@datta2005measuring]. We suppose that the random effects and error terms are independent, centred and normally distributed with unknown variance $\sigma_u^2=1$ and known $\sigma^2_{e_d}$ as follows. Each fifth part of the total number of clusters is assigned to a different value for $\sigma^2_{e_d}$; in Scenario 1: $0.7$, $0.6$, $0.5$, $0.4$, $0.3$ and in Scenario 2: $2.0$, $0.6$, $0.5$, $0.4$, $0.2$. Variance $\sigma_u^2 $ is estimated using REML, Henderson’s method 3 [@prased_rao], and the method of [@fay_herriot]. We present results only for the former as all the others performed in a similar way. Simulated scenarios are optimal for Bonferroni procedure, therefore we can treat it as a benchmark (cfr., comments in Section \[sec:bonfer\]). Let $g^{N*(b)}_{1d}(\hat{\bm{\theta}}^{*(b)}) = \hat{\sigma}^{2*(b)}_{u}/(\hat{\sigma}^{2*(b)}_{u} + \hat{\sigma}^{2*(b)}_{e}/n_d)(\hat{\sigma}^{2*(b)}_{e}/n_d)$ be the bootstrap equivalent of $g^N_{1d}(\hat{\bm{\theta}})$. We describe a parametric bootstrap procedure that yields promising results when constructing SPI. Under NERM the algorithm is 1. From the original sample, obtain consistent estimators $\hat{\bm{\beta}}$ and $\hat{\bm{\theta}}=(\hat{\sigma}^2_e, \hat{\sigma}^2_u)$. 2. Generate $D$ independent copies of a variable $W_1\sim N(0,1)$. Construct the vector $\bm{u}^*=(u^*_1,u^*_2,\dots,u^*_D)$ with elements $u^*_d=\hat{\sigma}_uW_1$, $d=[D]$. 3. Generate $N$ independent copies of a variable $W_2\sim N(0,1)$. Construct the vector $\bm{e}^*=(e^*_1,e^*_2,\dots,e^*_D)$ with elements $e^*_d=\hat{\sigma}_{e}W_2$, $d=[D]$. 4. Create a bootstrap sample $\bm{y}^*=\bm{X}\hat{\bm{\beta}}+\bm{u}^*+\bm{e}^*$. 5. Fit the nested error regression model  to the bootstrap sample from Step 4 and obtain bootstrap estimates $\hat{\bm{\beta}}^*$, $\hat{\bm{\theta}}^*=(\hat\sigma^{2*}_e,\hat\sigma^{2*}_u)$, $\mu_{dj}^*$ and $\hat{\mu}_{dj}^*$. 6. Repeat Steps 2-5 $B$ times. Calculate $S^{*(b)}_B$, $b=1,\dots,B$, using $g^{N*(b)}_{1d}(\hat{\bm{\theta}}^{*(b)})$. Obtain $c_{BS}(1-\alpha)$ and $\mathcal{I}^{BS}_{1-\alpha}$ as defined in Section \[sec:bootstrap\_SPI\]. To implement it under FHM, we need to slightly modify step 1 defining $\hat{\bm{\theta}}=\hat{\sigma}^2_u$ and $g^{F{*(b)}}_{1d}(\hat{\bm{\theta}}^{*(b)})=\hat{\sigma}^{2*(b)}_{u}\sigma^2_{e}/(\hat{\sigma}^{2*(b)}_{u}+\sigma^2_{e})$. Additionally, we need to replace step 3 by 1. Generate $D$ independent copies of a variable $W_2\sim N(0,1)$. Construct the vector $\bm{e}^*=(e^*_1,e^*_2,\dots,e^*_D)$ with elements $e^*_d={\sigma}_{e_d}W_2$, $d=[D]$. We use several criteria to evaluate performance of different methods to construct SPIs. First of all, we compute an empirical coverage probability (ECP) $$ECP =\frac{1}{I}\sum_{k=1}^{I}\mathbbm{1}\{\mu^{(k)}_{d}\in \mathcal{I}^{P}_{1-\alpha} \,\, \forall d \in [D] \} \quad \text{ where } P = BS, MC, BE \text{ or } BO,$$ that is, a percentage of times the parameters for all clusters are inside their SPI for $\alpha=0.05$. (our simulation experiment was carried out also for $\alpha=0.1$ and $\alpha=0.01$, nevertheless we did not include them, because they led to the same conclusions). The second comparison criterion is the average width (WS) of different SPIs. This quantity is calculated for each cluster over the widths of the intervals from $I$ simulations, and averaged over all clusters to obtain an aggregated indicator $$WS=\frac{1}{DI} \sum_{d=1}^{D}\sum_{k=1}^{I}\rho^{(k)}_d,\; \rho^{(k)}_d=2c^{(k)}_{P}(1-\alpha)\hat{\sigma}^{(k)}(\hat{\mu}_d ), \quad \text{ where P = BS, MC, BE or BO.}$$ Lower values of $WS$ are preferable. Finally, we compute the variance of widths over the simulations, and average them over all clusters, i.e., $$\label{eq:VU_MC} VS=\frac{1}{D(I-1)} \sum_{d=1}^{D} \sum_{k=1}^{I} \left(\rho^{(k)}_d-\bar{\rho}_d \right)^2,\; \bar{\rho}_d=\sum_{k=1}^{I} \rho^{(k)}_d/I.$$ We prefer lower values of $VS$, as they indicate that the length of intervals is stable. \(i) In practice, $c_{BS}(1-\alpha)$ and $c_{BE}(1-\alpha)$ are approximated by $[\{(1-\alpha)B\}+1]^{th}$ order statistics of the empirical bootstrap distributions. (ii) To construct $\mathcal{I}^{MC}_{1-\alpha}$ we can use $\bm{g}_{1}$ or the variance expression from the denominator in . Since the numerical differences between them were negligible, results are presented only for the latter. [max width=]{} ------------------ ---- ------ ------ ------ ------ --------------- --------------- --------------- --------------- D BS MC BE BO BS MC BE BO 15 95.4 92.9 93.6 93.8 1.876 (0.031) 1.754 (0.022) 1.803 (0.026) 1.794 (0.024) $\sigma^2_e=0.5$ 30 95.2 93.9 92.5 94.4 1.947 (0.015) 1.890 (0.013) 1.871 (0.013) 1.910 (0.013) $\sigma^2_u=1$ 60 94.9 93.7 88.7 94.2 2.041 (0.008) 2.011 (0.007) 1.936 (0.007) 2.023 (0.007) 90 95.2 94.4 84.4 94.9 2.101 (0.006) 2.079 (0.005) 1.926 (0.005) 2.088 (0.005) 15 96.7 91.2 93.8 94.4 2.695 (0.113) 2.358 (0.046) 2.488 (0.052) 2.488 (0.049) $\sigma^2_e=1$ 30 95.5 92.8 92.5 94.4 2.671 (0.027) 2.552 (0.024) 2.567 (0.024) 2.608 (0.024) $\sigma^2_u=1$ 60 95.0 93.7 89.1 94.5 2.774 (0.014) 2.719 (0.012) 2.631 (0.012) 2.750 (0.012) 90 95.2 94.2 83.2 94.8 2.850 (0.010) 2.811 (0.009) 2.614 (0.008) 2.833 (0.009) 15 98.3 87.3 92.3 96.5 2.816 (0.205) 2.156 (0.065) 2.362 (0.046) 2.488 (0.087) $\sigma^2_e=1$ 30 97.3 90.6 94.2 94.8 2.641 (0.050) 2.346 (0.032) 2.469 (0.023) 2.485 (0.022) $\sigma^2_u=0.5$ 60 95.3 92.7 89.7 94.5 2.616 (0.012) 2.513 (0.015) 2.478 (0.012) 2.577 (0.012) 90 95.0 93.0 83.9 94.6 2.663 (0.010) 2.597 (0.010) 2.441 (0.008) 2.643 (0.009) ------------------ ---- ------ ------ ------ ------ --------------- --------------- --------------- --------------- : ECP (in %), WS and VS under NERM. The nominal coverage probability is $95\%$.[]{data-label="tab:simulations"} Table \[tab:simulations\] shows the numerical results of the criteria evaluating performance of methods to construct SPI for a mixed parameter. Under all scenarios, BS attains the nominal level of 95% even for a small number of clusters ($D=15$). Yet, due to the overestimation of variability of the clusters, this method suffers from an overcoverage when $ICC=1/3$ for $D=15$ and $D=30$. BO yields satisfactory results too, which does not surprise as we consider independent mixed parameter with normal random effects and errors. Nonetheless it attains the nominal level only for $D=15$ and $ICC=2/3$ falling short in all other cases. MC has worse performance, the convergence to the nominal level is slower with ECP oscillating around 94% only for $D=60$ and $D=90$ when $ICC=2/3$ or $ICC=1/2$. In addition, it does not attain the nominal level under the third scenario. Last but not least, BE diverges. This undesirable effect might be mitigated using double bootstrap which is more computer intensive and its implementation does not assure significantly better numerical performance. The second part of Table \[tab:simulations\] summarizes results for WS and VS. As expected, the width increases with growing $D$. Note that BS and BO converge to approximately the same value for $D=90$. When we consider VS, we conclude that BS is certainly more variable than any other method for $D=15$, but this difference decreases with increasing $D$. In SAE, which focuses on the statistical analysis of mixed parameters, undercoverage is considered a more severe type of error than overcoverage, partly due to the difficulties to detect and alleviate it [@yoshimori2015numerical]. Often, overcoverage is a result of an excessive variability in the small samples which is illustrated in Table \[tab:simulations\] by VS. Having this in mind, we conclude that BS is the most satisfactory method. ![image](me_c){width="75.00000%"} \[fig:me\] Figure \[fig:me\] displays 95% bootstrap SPI in light blue and [@chatterjee2008parametric] cluster-wise prediction intervals (CPI) in blue. The critical values for the latter have been selected using parametric bootstrap such that each CPI covers its $\mu_d$, $d=1,\dots, D$ with a probability $0.95$, cfr., [@chatterjee2008parametric] for details. In comparison to CPI, SPI covers all clusters with a certain probability. Black dots represent the true mixed parameters $\mu_{d}$. Out of the 30, 3 are clearly outside of their CPI, and another 4 on the boundary. Recall that this does not happen by chance but by construction: for $100(1-\alpha)$% CPI about $100\alpha$% (often many more) of the cluster parameters are not in their intervals. In contrast, our SPIs contain all of the true mixed parameters. Moreover, the SPIs do not seem to be excessively wide but just as wide as necessary: one cluster mean is right at the upper bound of its SPI. CPI and SPI are methodologically different and constructed to cover distinct sets with a certain probability. Hence, one can argue that their direct comparison is flawed and should not be investigated. We do not claim otherwise; rather, this illustration serves to highlight why SPIs are highly relevant in practice. Figures \[fig:me\] and \[fig:interval\_est\] demonstrate that the cluster-wise inference leads to completely erroneous conclusions once applied to perform joint statements and comparisons. Secondly, we hope to convince practitioners to apply SPI if their goal is to compare mixed parameters across clusters. \ \ \ \[fig:test\] Regarding MT, Figure \[fig:test\] displays the empirical power of bootstrap and MC based max-type tests for $H_0:\bm{\mu}=\bm{h}$ vs $H_1: \bm{\mu}=\bm{h}+\bm{1}_{D}\delta$. In the simulations, we simply set $\bm{h}:=\bm{\mu}$ and under $H_1$ we added a constant $\delta\in[-2,2]$ to each element of $\bm{h}$. As expected, ICC influences the Type II error – the curves are the steepest and almost not distinguishable for $ICC=2/3$. The bootstrap test performs significantly better when $ICC=1/2$ and $ICC=1/3$ for small and medium $D$. In contrast, MC based tests do not attain the nominal level under $H_0$ for small sample sizes. For larger $D$ the curves almost coincide under three scenarios. Let us finally turn to the analysis under FHM. Since the performance of MT under FHM leads to similar conclusions as for NERM, we restrict ourselves to presenting ECP, WS and VS for different SPIs. Table \[tab:simulationsFH\] displays numerical results. Bootstrap SPIs suffer from overcoverage for small $D$, similarly to Bonferroni SPIs, which is probably caused by the same reasons as in case of NERM. Nevertheless, Bonferroni fails to achieve the nominal level for a large number of clusters. Beran’s SPIs diverge and MC SPIs undercover for $D=60$ and $D=90$. [max width=]{} ----- ---- ------ ------ ------ ------ --------------- --------------- --------------- --------------- D BS MC BE BO BS MC BE BO S.1 15 97.3 95.6 96.7 96.5 3.728 (0.016) 3.516 (0.024) 3.672 (0.040) 3.691 (0.019) 30 96.6 95.2 94.8 96.6 3.792 (0.017) 3.664 (0.023) 3.688 (0.045) 3.818 (0.013) 60 95.7 92.6 89.6 93.9 3.973 (0.014) 3.804 (0.031) 3.760 (0.103) 3.873 (0.027) 90 95.2 93.3 84 94.4 4.024 (0.016) 3.920 (0.025) 3.694 (0.090) 3.970 (0.022) S.2 15 98.0 95.7 97.6 95.9 4.073 (0.034) 3.749 (0.061) 4.005 (0.076) 3.962 (0.096) 30 97.1 95.6 94.8 96.1 3.795 (0.017) 3.667 (0.023) 3.690 (0.046) 4.028 (0.040) 60 97.4 93.4 91.6 94.9 4.198 (0.035) 3.956 (0.067) 3.981 (0.147) 4.029 (0.064) 90 96.6 93.9 86.1 94.6 4.218 (0.037) 4.006 (0.054) 3.880 (0.123) 4.119 (0.053) ----- ---- ------ ------ ------ ------ --------------- --------------- --------------- --------------- : ECP (in %), WS and VS under FHM. The nominal coverage probability is $95\%$.[]{data-label="tab:simulationsFH"} As a matter of fact, it should be mentioned that bootstrap SPI are not only versatile, but contrary to the other methods, also robust to severe departures from normality or random effects and errors (cfr., [@reluga2019simultaneous]). Empirical analysis of income data in Galicia {#sec:data_example} ============================================ We apply our proposal within the context of SAE and construct SPI for average household income in 52 counties of Galicia, Northern Spain. It is of great interest for Statistical Offices and politicians alike to gather information about the average income of individuals and households. Moreover, it is equally important to obtain this information at the county level of so-called *comarcas* to be able to adjust regional policies and resource allocations. We focus on the household income making use of the Structural Survey for Homes of Galicia in 2015 with 9203 households in total, but with $n_d<20$ in several clusters. The survey contains information about the total income as well as different characteristics on individual and household level. The dependent variable refers to the total yearly household income which consists of paid work, own professional activity and miscellaneous benefits. It is well known that income data are right skewed. Unsurprisingly, our dependent variable exhibits this feature too. We transform it by $y^{\log}_{dj}=\log(y_{dj}+c)$, where constant $c>0$ minimizes the Fisher skewness of the model residuals with $y^{\log}_{dj}$ as a response. It is selected from a grid within the range of household incomes $[\min(y_{dj}), \max(y_{dj})]$. Appendix \[sec:Appendix\_app\_Galicia\] contains a table with retained covariates and a description of the model selection process. We aim at the construction of SPI for the average monthly household income over 52 comarcas of Galicia, i.e., we consider with $\bm{k}_d=\bar{\bm{x}}_d$, where $\bar{\bm{x}}_d=n_d^{-1}\sum_{j=1}^{n_d}\bm{x}_{dj}$, $\bm{m}_d=1$, $u_d\in\mathbb{R}$ $\forall d\in [D]$. The estimation of this indicator at the small area level is of particular interest for the Statistical Office of Galicia (IGE). \[fig:kd\_qq\_R\] After a variable selection process, we try to validate the normality assumption for errors and random effects. The transformation of the responses seems to lead to a reasonable model adequacy presented at the kernel density plot of Cholesky residuals [@jacqmin2007robustness] in the left panel of Figure \[fig:kd\_qq\_R\]. The uncorrelated Cholesky residuals are constructed by multiplying $\bm{y}-\bm{X}\hat{\bm{\beta}}$ by the Cholesky square root of the variance matrix. However, a departure from normality is visible in the QQ plot displayed in the second panel. The pattern of heavy-tails is stark. This is also detected by the Shapiro-Francia test [@shapiro1972approximate] with a p-value of $p<0.001$. An empirical analysis of [@reluga2019simultaneous] demonstrated that the deviations from normality of errors has a strong impact on the coverage probability of MC SPI. On the other hand, bootstrap based SPI is robust to those departures, and provides a good coverage probability. In the light of these comments and the simulation results, Figure \[fig:kd\_qq\_R\] suggests focusing on bootstrap SPI. Regarding the normality of random effects, the right panel of Figure \[fig:kd\_qq\_R\] displays a diagnostic plot of [@lange1989assessing] using standardized empirical Bayes estimates of the random effects in a weighted normal QQ plot; it supports the adequacy of the normality assumption. ![image](widths_intC){width="90.00000%"} \[fig:interval\_est\] Figure \[fig:interval\_est\] displays bootstrap CPI proposed by [@chatterjee2008parametric] and SPI of the log of mean incomes in the counties of Galicia. We divide the plot in Figure \[fig:interval\_est\] into 5 panels based on the number of units in each comarca. We can see a lot of variability over the estimates for the comarcas. Evaluating the results of CPI (blue) versus SPI (light blue), it is apparent that the cluster-wise prediction intervals are not adequate to address either a joint consideration or a comparison of the comarcas. If we consider, for example, the comarcas of *A Mariña Oriental* and [*Chantada*]{} (5th and 6th regions of the second panel in the black rectangle), the CPIs indicate significantly different incomes, whereas the SPIs do not support this claim. Moreover, there are other comarcas (practically in each panel) for which CPIs would insinuate significant differences whereas statistically valid SPIs do not confirm that conclusion. Nevertheless, SPIs are not unnecessarily wide for practical use. We detect significant and valid differences between several interval estimates. Figure \[fig:maps\] present maps with lower and upper limits of bootstrap SPIs. The boundaries are classified into one of five categories which were built using $0.2$, $0.4$, $0.6$ and $0.8$ quantiles of point estimators. One detects a substantial variation of the household income over the counties. Lower and upper boundary of the interval estimates for the comarcas of [*A Coruña*]{}, [*Santiago de Compostela*]{} and *Lugo* (with a large number of units) are classified into the richest category; they are indicated in the last panel of Figure \[fig:interval\_est\]. [*O Ribeiro*]{}, being located in the inner zone where inhabitants live off the agriculture, seems to be the poorest region; lower and upper boundary of its interval estimate lie within the first category (marked with ellipsoid in the second panel of Figure \[fig:interval\_est\]). Furthermore, we can see that in the south-east Galicia, there is a group of relatively poor comarcas, with the exception of Ourense. Similar conclusions can be drawn from the publications of Galician Institute of Statistics. CPIs should not be used to make such maps as this would suggest that we were allowed to compare them. Moreover, CPIs would not contain at least 3 (for $95\%$ intervals) of the true mixed parameters. ![image](maps_col_2){width="80.00000%"} \[fig:maps\] Our max-type statistic might also be used for hypothesis testing. Suppose that our goal is to statistically validate or disprove the hypothesis that there is no statistical difference in the mean household income between households having male and female household heads (hh). To test this hypothesis we take clusters created from the cross section of comarcas and hh-gender. For the time being, we consider $\bm{\mu}\in \mathbb{R}^{104}$ and test $H_0:\bm{A\mu}=\bm{0}_{104}$ versus $H_1:\bm{A\mu}\neq\bm{0}_{104}$, where $\bm{A}\in \mathbb{R}^{52\times104}$ with rows that are composed of $104$-dimensional vectors $\bm{a}$ with $1$ on the $2d-1$ place, $-1$ on the $2d$ place but $0$ otherwise, where $d$ stands for a particular comarca. The test statistic is $t_{H} =\max_{d=1,\dots, D}|t_{H_d}|=10.495$ whereas the bootstrap critical value is $c_{BH_0}(1-\alpha)=8.673$. That is, we clearly reject $H_0$ of no difference. Conclusions {#sec:conclusions} =========== We employed the max-type statistic to construct SPI and MT procedure for a mixed parameter under LMMb and applied it within the framework of SAE. We theoretically derived two novel techniques based on bootstrap and the volume-of-tube formula; the latter did not turn out to be operational. Furthermore, we carried out an empirical comparison with the methods adapted from the setting without random effects and nonparametric curve estimation. Though slightly conservative for very small samples, bootstrap SPIs yield satisfactory results in all simulation studies. For the sake of brevity, we only presented the scenarios under normality and independent parameter estimates. These assumptions are optimal for Bonferroni procedure which was less reliable than our approach. In addition, our bootstrap based max-type statistic is readily applicable for testing multiple statistical hypotheses. Our empirical studies reveal its superiority over another resampling based test. It is clear that accounting for the variability of all clusters makes SPI wider than CPI. Nonetheless, only SPIs are statistically valid for comparisons between clusters. Moreover, if we conducted several surveys, SPIs would contain all true parameters in $100(1-\alpha)\%$ of studies, whereas CPIs would not cover about $D\alpha$ of them in each survey. Finally, as soon as we wish to make joint statements, the application of our methodology is indispensable in empirical studies, which was shown in the analysis of household income in the comarcas of Galicia. Our max-type statistic simultaneous inference can be easily extended to account for more complex data structures with several random effects, heteroscedasticity, spatial or temporal correlation which arises naturally in e.g., longitudinal studies. Moreover, our methodology can be developed to be applied under generalized linear, semi- or nonparametric mixed models. In these cases we would need to apply estimators and wild bootstrap as in, e.g., [@lombardiasperlich2008]. Appendix {#sec:appendixA} ======== Regularity conditions {#sec:RC} --------------------- 1. $\bm{X}_d$ and $\bm{Z}_d$ are uniformly bounded such that $\sum_{d=1}^{D}\bm{X}^t_d \bm{V}_d^{-1}\bm{X}_d=\left\{O(D)\right\}_{(p+1)\times (p+1)}$. 2. Covariance matrices $\bm{G}_d$ and $\bm{R}_d$ have a linear structure, that is $\bm{G}_d=\sum_{j=0}^{h}\theta_i\bm{J}_{dj}\bm{J}_{dj}^t$ and $\bm{R}_d=\sum_{j=0}^{h}\theta_i\bm{H}_{dj}\bm{H}_{dj}^t$, where $d=1,\dots, D$, $j=0,\dots,s$ $\theta_0=1$, $\bm{J}_{dj}$ and $\bm{H}_{dj}$ are known of order $n_d\times q_d$ and $q_d\times q_d$ respectively; in addition, the elements of $\bm{J}_{dj}$ and $\bm{H}_{dj}$ are uniformly bounded known constants such that $\bm{G}_d$ and $\bm{R}_d$ are positive definite matrices. In certain cases, $\bm{J}_{dj}$ and $\bm{H}_{dj}$ can be null matrices; 3. Rate of convergence: $\sup_{d\geqslant1}n_d<<\infty$ and $\sup_{d\geqslant1}q_d<<\infty$. 4. To ensure the nonsingularity of $\bm{\Sigma}_{\bm{\theta}}$, $ 0<\inf_{d\leqslant1}\sigma^2_{e_d} \leqslant \sup_{d\leqslant1}\sigma^2_{e_d}<\infty$ and $\sigma^2_u\in(0,\infty)$; 5. $\bm{b}_d^t=\bm{k}^t_d-\bm{a}_d^t\bm{X}_d$ with $b_{di}=O(1)$ for $i=1,\dots, p+1$; 6. $\{\frac{\partial}{\partial \theta_j}\bm{a}_d^t\bm{X}_d\}_i=O(1)$ for $j=1,\dots,h$ and $i=1,\dots, p+1$; 7. $\hat{\bm{\theta}}$ satisfies: $(i)$ $\hat{\bm{\theta}}-\bm{\theta}=O_p(m^{-1/2})$, $(ii)$ $\hat{\bm{\theta}}(\bm{y})=\hat{\bm{\theta}}(-\bm{y})$ and $(iii)$ $\hat{\bm{\theta}}(\bm{y}+\bm{Xr})=\hat{\bm{\theta}}(\bm{y})$ for any $\bm{r}\in \mathbb{R}^{p+1}$. Furthermore, we will evoke Assumptions 1-4 from [@chatterjee2008parametric], which are quite technical but largely irrelevant in practice. Proof of Proposition 1 {#sec:prop_1} ---------------------- We concentrate on the consistency of the bootstrap SPI. [@richardson1994asymptotic] proved the consistency and asymptotic normality of fixed parameters $\hat{\bm{\beta}}$ and variance components $\hat{\bm{\theta}}$, whereas the consistency of their bootstrap counterparts was proven by [@gonzalez2008bootstrap]. To demonstrate Proposition \[prop:con\_max\], we make use of Theorem CLL of [@chatterjee2008parametric] and two further propositions. In this section, we use some notations from [@chatterjee2008parametric] whenever they are not in conflict with the notations in the rest of the paper, otherwise we shall keep ours. \[theorem:CLL\] (Theorem 3.1 of [@chatterjee2008parametric]). Let $T=\bm{f}^t(\bm{X\beta}+\bm{Zu})$, $\bm{f}\in\mathbb{R}^{n}$ a vector and a conditional distribution of $T$ given $\bm{y}$ be $N(\mu_T,\sigma^2_T)$, where $$\begin{split} \mu_T&=\bm{f}^t\bm{Z}\bm{\beta}+\bm{f}^t\bm{ZG}\bm{Z}^t\bm{V}^{-1}(\bm{y}-\bm{X\bm{\beta}}) \\ \sigma^2_T&=\bm{f}^t\bm{Z}(\bm{G}-\bm{GZ}^t\bm{V}^{-1}\bm{ZG})\bm{Z}^t\bm{f} . \end{split}$$ Let $T^*=\bm{f}^t(\bm{X}\hat{\bm{\beta}}+\bm{Z}\bm{u}^*)$ be a bootstrap equivalent of $T$, $\mathcal{L}_n$ the cdf of $\hat{\sigma}_T^{-1}(T-\hat{\mu}_T)$ with $\bm{\theta}$ replaced by $\hat{\bm{\theta}}$, and $\mathcal{L}^*_n$ the cdf of $\hat{\sigma}^{-1*}(T^*-\hat{\mu}^*_T)$. Suppose Assumptions 1-4 of [@chatterjee2008parametric] hold and $p+1+h=o(n)$. Then, $$\sup\limits_{q\in\mathbb{R}} \left\lvert\mathcal{L}_n(q)-\mathcal{L}^*_n(q)\right\rvert=O_P\{(p+h+1)^3n^{-3/2}\}.$$ Moreover, $\mathcal{L}_n(q)$ admits a short asymptotic expansion, i.e., $\mathcal{L}_n(q)=\Phi(q)+h^2n^{-1}\gamma(q,\bm{\beta},\bm{\theta})+O(h^3n^{-3/2})$ where $\Phi(q)$ stands for a cdf of a standard Gaussian random variable. \[remark:GLS\_vs\_OLS\] [@chatterjee2008parametric] estimated fixed parameters using an ordinary least squares method, whereas in our paper we used a generalized least squares. As pointed out by the authors, an asymptotic expansion still holds as soon as the weighting matrices are smooth functions of $\bm{\theta}$, which we assume in R.2. \[remark:multivariate\_individual\] In their analysis, Theorem \[theorem:CLL\] was demonstrated for one dimensional parameter, yet a similar asymptotic expansion would apply to a multivariate parameter with some notational and algebraic adjustments, see Remark 4 in [@chatterjee2008parametric]. Without loss of generality, we suppose $\mu_d=T=\bm{f}_1^t(\bm{X\beta}+\bm{Zu})$ for some vector $\bm{f}_1\in\mathbb{R}^n$. Notice that in case of our application in Section \[sec:data\_example\] this claim is satisfied with $\bm{f}_1=(0,\dots,1/n_d,\dots,1/n_d,\dots ,0)^t$. In other cases some algebraical changes in the proof of Theorem \[theorem:CLL\] might be needed. Further, define a multivariate parameter $\bm{T}=\bm{F}(\bm{X\beta}+\bm{Zu})$ and suppose that there exists a matrix $\bm{F}_1\in \mathbb{R}^{D\times n}$ such that $\bm{\mu}=\bm{T}=\bm{F}_1(\bm{X\beta}+\bm{Zu})$, $\bm{\mu}=(\mu_1, \mu_2,\dots, \mu_D)^t$ and $\bm{S}_0=\hat{\bm{\sigma}}^{-1} (\bm{\mu}-\hat{\bm{\mu}})$ with $\hat{\bm{\sigma}}^{2}$, which is a matrix equivalent of $\hat{\sigma}^2_T$ and $\bm{S}_0$ composed of $S_{0d}$ defined in . Let $\bm{S}^*=\bm{\sigma}^{-1^*} (\bm{\mu}^*-\hat{\bm{\mu}}^*)$ be the bootstrap version of $\bm{S}_0$. By Theorem \[theorem:CLL\] and Remark \[remark:multivariate\_individual\] it follows that $$\label{eq:L_M_expansion} \mathcal{L}_M(\bm{q})=\bm{\Phi}(\bm{q})+n^{-1}h^2\bm{\gamma}(\bm{q},\bm{\beta}, \bm{\theta})+O(h^3n^{-3/2}),$$ where $\mathcal{L}_M(\bm{q})=\mathrm{P}(\bm{S}_0\leqslant\bm{q})$. To show that Proposition \[prop:con\_max\] is satisfied, we establish the validity of two propositions and employ GAR of [@chernozhukov2013gaussian]. The derivation proceeds within the same lines as the one in [@zhang2017simultaneous]. \[prop:max\_norm\] Assume that Theorem \[theorem:CLL\] holds and let $v_d\sim N(0,1)$, $d=1,\dots,D$, $\bm{v}=(v_1,\dots,v_D)^t$. Then, $$\sup\limits_{q\in\mathbb{R}}\left \lvert\mathrm{P}\left\{\max_{d=1,\dots, D} w(S_{0d})\leqslant q \right\} - \mathrm{P}\left\{ \max_{d=1,\dots, D} w(v_d)\leqslant q \right\} \right\rvert=o(1)$$ where $w$ as defined in Proposition \[prop:con\_max\]. *Proof*: To apply GAR, suppose that $S_{0d}=n^{-1/2}\sum_{i=1}^{n}S_{0di}$, i.e., $S_{0d}$ has a linear structure. Proposition \[prop:con\_max\] is satisfied if we demonstrate Corollary 2.1 of [@chernozhukov2013gaussian] by validating their Condition (E.1), i.e., showing that $$(i)\;c_0\leqslant \mathbb{E}(S_{0di})\leqslant C_0\quad \text{and} \quad (ii)\;\max_{k=1,2}\mathbb{E}(\vert S_{0dj}\rvert^{2+k}/B^k)+\mathbb{E}\{\exp(\lvert S_{0di} \rvert/B)\}\leqslant 4$$ uniformly over $d$, with $c_0$, $C_0$ and B some constants. First of all, notice that $\sigma^2_e$ and $\sigma^2_u$ are finite. Therefore $(i)$ would hold for $c_0(\sigma^2_e+\sigma^2_u)\leqslant \mathbb{E}(S_{di})\leqslant C_0(\sigma^2_e+\sigma^2_u)$. Moreover, due to the normality of $e_{dj}$ and $u_d$, $d=1,\dots, D$, $j=1,\dots, n_d$, all the moments of $S_{0di}$ exist. Hence we can find $c_1$ such that $\max_{k=1,2} \mathbb{E}(|S_{0dj}|^{2+k})\leqslant c_1$ as well as $C_1$ such that $\mathbb{E}(|S_{0dj}|^{i})\leqslant C_1$, $i\in \mathbb{N}$. Thus, the second term in $(ii)$ is $$\mathbb{E}\{\exp(\lvert S_{0di} /C_2\rvert)\}=1+\sum_{k=1}^{\infty}\frac{\mathbb{E}(|S_{0dj}|^k)}{k!C_2^k} \leqslant 1+\sum_{k=1}^{\infty} \frac{(C_1/C_2)^k}{k!}=\exp(C_1/C_2)<\infty$$ for $C_2$ some constant. Since both terms of $(ii)$ are bounded, we can find a sufficiently large $B$ such that the second condition of (E.1) holds. Furthermore, $|S_{0d}|=\max(S_{0d}, -S_{0d})$ implies that $w_d(t)=|t|$ is valid and Proposition \[prop:max\_norm\] is proven. The parametric bootstrap counterpart is given in Proposition \[prop:max\_norm\_boot\]. \[prop:max\_norm\_boot\] Assume that Theorem \[theorem:CLL\] and Proposition \[prop:max\_norm\] hold. Then, $$\sup\limits_{q\in\mathbb{R}}\left \lvert\mathrm{P}^*\left\{ \max_{d=1,\dots, D} w(S^*_{Bd})\leqslant q \right\} - \mathrm{P}\left\{ \max_{d=1,\dots, D} w(v_d)\leqslant q \right\} \right\rvert=o_P(1)$$ where $w$ is as defined in Proposition \[prop:con\_max\]. *Proof*: A similar reasoning as in Proposition \[prop:max\_norm\] holds, replacing $\mathrm{P}$, $S_{0d}$, $\sigma^2_e$ and $\sigma^2_u$ by $\mathrm{P}^*$, $S^*_{Bd}$, $\hat{\sigma}^2_e$ and $\hat{\sigma}^2_u$. The consistency of $\mathcal{I}_{1-\alpha}^{BS}$ and Proposition \[prop:con\_max\] can also be based on a heuristic argument of [@hall1990simultaneous] combined with the asymptotic expansion in Theorem \[theorem:CLL\] of [@chatterjee2008parametric]. Ideally, $c_{S_0}(1-\alpha)$ would be determined from $$\pi\left\{c_{S_0}(1-\alpha)\right\}=\mathrm{P}\left\{ -c_{S_0}(1-\alpha)\leqslant S_{0d} \leqslant c_{S_0}(1-\alpha) \; \forall d \in [D]\right\}=1-\alpha.$$ Since $\pi\left\{c_{S_0}(1-\alpha)\right\}$ is unknown, we approximate it by bootstrap $$\pi^*\{c_{BS}(1-\alpha)\}=\mathrm{P}\left\{ -c_{BS}(1-\alpha)\leqslant S^*_{Bd} \leqslant c_{BS}(1-\alpha) \; \forall d \in [D]|\mathcal{W}\right\}=1-\alpha$$ with $\mathcal{W}=\{(y_{dj},\bm{x}_{dj},\bm{z}_{dj}),d\in[D],j\in[n_d] \} $. If we prove that $\pi$ and $\pi^*$ are asymptotically close up to the order $O_P((h^2n^{-3/2})$, it implies the same order of accuracy for $\mathcal{I}^{BS}_{1-\alpha}$. Define $$R=\{\bm{x}\times \bm{z}\in\mathcal{X}\times\mathcal{Z}:-c_{S_0}(1-\alpha)\leqslant S_{0d} \leqslant c_{S_0}(1-\alpha) \, \forall d \in [D]\}$$ which can be represented as a finite number of unions and intersection of convex sets. This number is bounded uniformly for $D\geqslant2$ and $c_{S_0}(1-\alpha)>0$. Observe that $\pi=\int_{R} d\mathcal{L}_n$. Theorem \[theorem:CLL\] shows that for all continuity points $q$ the [cdf]{}’s $\mathcal{L}_n$ and $\mathcal{L}^*_n$ converge to the same limit at the desired speed, and the same speed is maintained in the convergence of $\pi$ and $\pi^*$. Since $\pi$ is defined as an integral of $d\mathcal{L}_n$ over $R$, a direct consequence of Lemma \[theorem:CLL\] is $\sup_{k\in\mathbb{R}}| \pi(k)-\pi^*(k)|=O_P(h^2n^{-3/2})$. The step-down testing procedure: description and numerical results {#sec:Appendix_step_down} ================================================================== The testing procedure described in Section \[sec:SPI\] controls weakly for FWER. If one aims at testing multiple hypotheses like $$\label{eq:test_ind} H_{0d}:\mu_d=h_d \text{ for all } d\in [D] \text{ vs. } H_{1d}:\mu_d\neq h_d, \text{ for some } d \in [D] ,$$ then the step-down technique of [@romano2005exact]can be used assuring a strong control of the FWER. More specifically, let $\Omega$ be the space for all data-generating processes and $w$ be the true one. Each $H_{0d}$ is equivalent to $\omega\in\Omega_d$ for some $\Omega_d \subseteq \Omega$. For any $\vartheta\subseteq[D]$, denote $\Omega^{\vartheta}=(\cap_{d\in \vartheta}\Omega_d)\cap (\cap_{d \notin \vartheta}\Omega^c_d)$ with $\Omega^c_d= \Omega \textbackslash \Omega_d$. The strong control of the FWER implies that $$\label{eq:cond_Propo_RW} \sup\limits_{\vartheta \subseteq [D] } \sup\limits_{\omega \in \Omega^{\vartheta}} \mathrm{P}_{\omega}(\text{reject at least one hypothesis } H_{0d}, d \in \vartheta) \leqslant \alpha + o(1).$$ The step-down procedure of [@romano2005exact] can be summarised in the following way; consider $t_{H_d}$, $d\in[D]$ defined in Section \[sec:SPI\] and denote by $c_{\vartheta}(1-\alpha)$ an estimate for the $(1-\alpha)$-quantile of $\max_{d\in \vartheta} |t_{H_d}|$. Let $\vartheta(1)=[D]$ and reject all hypotheses $H_{0d}$ for which $t_{H_d}>c_{\vartheta{1}}(1-\alpha)$. If no hypothesis is rejected, then stop. Otherwise, let $\vartheta(2)$ be the set of null hypotheses not rejected at the first step. At step $k$, let $\vartheta(k)\subseteq[D]$ be the set of hypotheses not rejected at step $k-1$. Reject all hypotheses $H_{0d}$, $d\in \vartheta(k)$ for which $t_{H_d}>c_{\vartheta(k)}(1-\alpha)$. If no hypothesis is rejected, stop. Proceed in this way until the algorithm stops. The authors proved the strong control of the FWER, that is $$\label{eq:cond_1_RW} c_{\vartheta}(1-\alpha)\leqslant c_{\vartheta'}(1-\alpha), \text{ for } \vartheta \subseteq \vartheta',$$ $$\label{eq:cond_2_RW} \sup\limits_{\vartheta \subseteq [D] } \sup\limits_{\omega \in \Omega^{\vartheta}} \mathrm{P}_{\omega}(\max\limits_{d \in \vartheta }|t_{H_d}| > c_{\vartheta}(1-\alpha)) \leqslant \alpha + o(1).$$ A direct consequence of Lemma \[theorem:CLL\] is a following corollary regarding a practical method to obtain appropriate critical values $c_{\vartheta}(1-\alpha)$ in the step down-procedure. Under assumptions of Lemma \[theorem:CLL\] the step-down procedure of [@romano2005exact] with bootstrap approximation of critical value $c_\vartheta(1-\alpha)$ provides a strong control of the FWER, satisfying . [max width=]{} ---------------------------------- -------- -------- -------- -------- -------- -------- -------- -------- $D=15$ $D=30$ $D=60$ $D=90$ $D=15$ $D=30$ $D=60$ $D=90$ $\sigma^2_e=0.5$, $\sigma^2_u=1$ 0.043 0.042 0.049 0.048 0.048 0.044 0.508 0.046 $\sigma^2_e=1$, $\sigma^2_u=1$ 0.039 0.037 0.045 0.042 0.042 0.042 0.046 0.044 $\sigma^2_e=1$, $\sigma^2_u=0.5$ 0.047 0.040 0.038 0.037 0.025 0.039 0.044 0.040 ---------------------------------- -------- -------- -------- -------- -------- -------- -------- -------- : Empirical FWER under NERM. The nominal FWER is $0.05$.[]{data-label="tab:RW_simulations"} We compare the finite sample performance of the step-down method of [@romano2005exact] with the Bonferroni procedure which seems to be the most serious competitor of the bootstrap method. To this end, we consider multiple two-sided testing hypotheses $H_{0,d}: \mu_d=h_d$ for all $d\in[D]$ for three scenarios under NERM. We assume that $\mu_d=h_d+1$ for $d=1,\dots,D/5$. The performance criterion is an empirical FWER whereas the true nominal level of FWER is set to $\alpha=0.05$. The analysis of Table \[tab:RW\_simulations\] confirms that bootstrap step-down procedure controls strongly for the FWER, whereas Bonferroni method fails to do it under $\sigma^2_e=0.5$, $\sigma^2_u=1$ for $D=60$. Except for the third scenario and D=15, bootstrap procedure provides lower or equally good FWER. Proof of Proposition 2 {#sec:Appendix_Prop_2} ====================== To derive approximation formulas in Proposition 2, we suppose for the moment that the manifold $\mathcal{M}=\left\{\mathcal{Q}(\bm{c}),\bm{c}\in \mathcal{C}\right\}$ has no boundary, that is the Euler-Poincaré characteristic $\mathcal{E}=0$. In addition, $\bm{l}_M$, $\hat{\bm{l}}_M$, $\bm{e}_M$, $\hat{\bm{e}}_{M}$, $\lambda=\hat{\bm{e}}_M-\bm{e}_M$ and the other assumptions remain as defined in Section \[sec:vot\_sec\]. For some $\bm{x}\in \mathcal{X}$, $\bm{z}\in\mathcal{Z}$ and $\bm{c}\in\mathcal{C}$ the difference between $\bm{x}^t\bm{\beta}+\bm{z}^t\bm{u}$ and its estimate can be bounded by $$\label{eq:bound_dif} \left|\bm{x}^t\bm{\beta}+\bm{z}^t\bm{u}-\hat{\bm{l}}^t\bm{y}\right|= \left|\hat{\bm{l}}_M^t \hat{\bm{e}}_M\right|=\left|\bm{l}_M^t \bm{e}_M + \left(\hat{\bm{l}}_M-\bm{l}_M\right)^t\bm{e}_M + \hat{\bm{l}}^t_M\lambda_M \right|\leqslant \left| \bm{l}_M^t \bm{e}_M \right| + \eta(\bm{x}),$$ where $\eta(\bm{x})=\left|\left(\hat{\bm{l}}_M-\bm{l}_M\right)^t\bm{e}_M + \hat{\bm{l}}^t_M\lambda_M \right|$. If we normalize the first term on the right hand side, it is straightforward to see that $$Z=\frac{\left\langle\bm{l}_M,\bm{e}_M\right\rangle}{\sigma_e||\bm{l}_M||}= \left\langle\frac{\bm{l}}{||\bm{l}_M||},\frac{\bm{e}_M}{\sigma_{e}}\right\rangle= \left\langle\mathcal{Q},\frac{\bm{e}_M}{\sigma_{e}}\right\rangle,$$ which coincides with the formula of a Gaussian random field in Section 3.2. Here, $\bm{e}_M/\sigma_{e}$ is a $n$-vector of normally distributed random variables. Following the derivation of [@sun1994simultaneous] and [@sun1999confidence], one needs to choose $c_{VT}(1-\alpha)$ such that $$\begin{split} \alpha&= \mathrm{P}\left(\sup_{\bm{c}\in\mathcal{C}} \frac{\left|\bm{x}^t\bm{\beta}+\bm{z}^t\bm{u}-\hat{\bm{l}}^t\bm{y}\right|}{\hat{\sigma}_e||\hat{\bm{l}}_M||} > c_{VT}(1-\alpha) \right)\\ &\leqslant \mathrm{P}\left[ \sup_{\bm{c}\in\mathcal{C}} \left[ \left\{ \left\lvert\frac{\left\langle\bm{l}_M,\bm{e}_M\right\rangle}{\sigma_e||\bm{l}_M||}\right\rvert + \frac{\eta(\bm{x})}{ \sigma_e||\bm{l}_M||} \right\} \frac{||\bm{l}_M||}{||\hat{\bm{l}}_M||} \right] > c_{VT}(1-\alpha)\frac{\hat{\sigma}_e}{\sigma_e} \right]\\ &\leqslant \mathrm{P}\left\{\sup_{\bm{c}\in\mathcal{C}} \left|Z\right| > c_{VT}(1-\alpha)\frac{\hat{\sigma}_e}{\sigma_e} \xi -\eta \right\} =2\mathrm{P}\left\{\sup_{\bm{c}\in\mathcal{C}} Z > c_{VT}(1-\alpha)\frac{\hat{\sigma}_e}{\sigma_e} \xi -\eta \right\} , \end{split}$$ where $\xi=\inf\limits_{\bm{c}\in\mathcal{C}}\frac{||\hat{\bm{l}}_M||}{||\bm{l}_M||}$ is the minimum of the ratio between estimated $\hat{\bm{l}}_M$ and the true $\bm{l}_M$, and $\eta=\sup\limits_{\bm{c}\in\mathcal{C}}\frac{\eta(\bm{x})}{ \sigma_e||\bm{l}_M||}$ accounts for the difference in the estimation of variance parameters. When $\bm{\theta}$ is obtained using some consistent estimator (for example REML or MM), then [@jiang1998asymptotic] proved that $\hat{\sigma}_e$ is asymptotically independent of $Z$ and $$\xi=1+o_p(1), \quad \eta=o_p(1) \quad \text{as} \quad n\rightarrow\infty.$$ Therefore, $\xi$ and $\eta$ can be bounded by positive constants $\xi\leqslant\xi_0$ and $\eta\leqslant\eta_{0}$ in probability as $n\rightarrow\infty$, and we obtain the approximation $$\label{eq:Proba} \alpha\leqslant 2\mathrm{P}\left(\sup_{\bm{c}\in\mathcal{C}} Z > \frac{c_{VT}(1-\alpha)}{\nu^{1/2}} \frac{\hat{\sigma}_e\nu^{1/2}}{\sigma_e} \xi_0 -\eta_0 \right)\ + o(\alpha).$$ Under our setting, the variable $\nu^{1/2}\hat{\sigma}/\sigma$ is approximately $\chi$ distributed with $\nu$ degrees of freedom and a pdf $$f(w,\nu)=\frac{w^{\nu-1}e^{-w^2/2}}{2^{\nu/2-1}\Gamma(\nu/2)}.$$ Since we deal with a Guassian random field, we can adjust [@sun1993tail]’s approximations formulas to retrieve a value of $c_{VT}(1-\alpha)$ for $p=1$, $p=2$ and $p\geqslant3$. First of all, let us focus on the cases $p=1$ and $p=2$, where we need to find a solution of $$\alpha\leqslant 2\int_{0}^{\infty} \mathrm{P}\left\{\sup_{\bm{c}\in\mathcal{C}} Z > \frac{c_{VT}(1-\alpha)\xi_0 }{\nu^{1/2}} w -\eta_0 \right\}\ f(w,\nu)dw+o(\alpha).$$ We develop two expressions using Taylor expansions. Let $c'_{VT}(1-\alpha)=c_{VT}(1-\alpha)\xi_0$ and $\frac{c'_{VT}(1-\alpha)w}{\nu^{1/2}}=j(w)$. Then for any $\eta \rightarrow 0$ we have $$\begin{aligned} & & \exp\left\{-\frac{1}{2}\left(j(w)-\eta\right)^2\right\} =\exp\left\{-\frac{1}{2}j^2(w)\right\}+ \exp\left\{-\frac{1}{2}j^2(w)\right\}j(w)\eta+o(\eta) \\ & & =A_1+A_2\eta+o(\eta) , \mbox{ and } \\ & & \left\{j(w)-\eta\right\}\exp\left\{-\frac{1}{2}\left(j(w)-\eta\right)^2\right\} =j(w)\exp\left\{-\frac{1}{2}j^2(w)\right\} + \left[j'(w)\exp\left\{-\frac{1}{2}j^2(w)\right\} \right. \\ & & \left. -j^2(w) \exp\left\{-\frac{1}{2}j^2(w)\right\} \right](-\eta)+o(\eta) =j(w)\exp\left\{-\frac{1}{2}j^2(w)\right\}-\frac{c'_{VT}(1-\alpha)\eta}{\nu^{1/2}}\exp\left\{-\frac{1}{2}j^2(w)\right\}\\ & & +j^2(w)\exp\left\{-\frac{1}{2}j^2(w)\right\}\eta+o(\eta) =A_2-A_1 \frac{c'_{VT}(1-\alpha) \eta}{\nu^{1/2}} + A_3 \eta+o(\eta) .\end{aligned}$$ Using a $\chi$ distribution, we have for ${\cal A}_j := \int_{0}^{\infty} A_j f(w,\nu) dw$, $j=1,2,3$: $$\begin{aligned} {\cal A}_1 &=& \int_{0}^{\infty}\exp\left\{ -\frac{1}{2} j^2(w) \right\} f(w,\nu) dw = \int_{0}^{\infty}\exp\left\{ - \frac{ c'^2_{VT}(1-\alpha)}{2\nu} w^2 \right\} f(w,\nu)dw \\ & = & \left\{ 1+\frac{c_{VT}(1-\alpha)^{2}\xi_0^2}{\nu} \right\}^{-\nu/2}, \\ {\cal A}_2 &=& \int_{0}^{\infty}\exp\left\{ -\frac{1}{2} j^2(w) \right\} j(w) f(w,\nu) dw = \frac{ c'_{VT}(1-\alpha)}{\nu^{1/2}}\int_{0}^{\infty}\exp\left\{- \frac{ c'^2_{VT}(1-\alpha)}{2\nu} w^2 \right\} w f(w,\nu)dw \\ & = & \frac{ c_{VT}(1-\alpha)\xi_0}{\nu^{1/2}} \frac{2^{1/2} \Gamma\left(\frac{\nu+1}{2}\right) }{\Gamma\left(\frac{\nu}{2}\right) \left\{1+\frac{c_{VT}(1-\alpha)^2\xi_{0}^{2}}{\nu} \right\}^{(\nu+1)/2} }, \\ {\cal A}_3 &=& \int_{0}^{\infty}\exp\left\{ -\frac{1}{2} j^2(w) \right\} j^2(w) f(w,\nu) dw = \frac{ c'^2_{VT}(1-\alpha)}{\nu}\int_{0}^{\infty}\exp\left\{ - \frac{ c'^2_{VT}(1-\alpha) }{2\nu} w^2 \right\} w^2 f(w,\nu)dw \\ & = & \frac{ c^2_{VT}(1-\alpha) \xi_0^{2} }{\nu} \frac{2 \Gamma\left(\frac{\nu+2}{2}\right) }{\Gamma\left(\frac{\nu}{2}\right) \left\{1+\frac{c_{VT}(1-\alpha)^2\xi_{0}^{2}}{\nu} \right\}^{(\nu+2)/2} } .\end{aligned}$$ To find an approximation for the model with $p\geqslant 3$, we modify a following expression $$\begin{aligned} \alpha&\leqslant & 2 \mathrm{P}\left\{\sup_{\bm{c}\in\mathcal{C}} Z > \frac{c_{VT}(1-\alpha)}{\nu^{1/2} } \frac{\hat{\sigma}_e\nu^{1/2}}{\sigma_e} \xi_0 -\eta_0 \right\} =2 \mathrm{P}\left\{\sup_{\bm{c}\in\mathcal{C}} Z > \frac{c_{VT}(1-\alpha)}{\nu^{1/2} } \frac{\hat{\sigma}_e\nu^{1/2}}{\sigma_e} \xi_0 - \eta_0 \frac{\nu^{1/2} \hat{\sigma}_e \sigma_e }{\nu^{1/2} \sigma_e \hat{\sigma}_e } \right\} \\ \\ &=&2 \mathrm{P}\left[\sup_{\bm{c}\in\mathcal{C}} Z > \frac{\hat{\sigma}_e\nu^{1/2}}{\sigma_e \nu^{1/2} } \left\{ c_{VT}(1-\alpha) \xi_0 - \eta_0 \frac{\sigma_e }{ \hat{\sigma}_e } \right\} \right] .\end{aligned}$$ Having calculated all of the necessary terms, we obtain final approximations:\ 1. for $p=1$: $$\begin{aligned} \alpha&\leqslant & 2\int_{0}^{\infty} P\left\{\sup_{\bm{c}\in\mathcal{C}} Z > \frac{c_{VT}(1-\alpha)\xi_0 }{\nu^{1/2}} w -\eta_0 \right\} \ f(w,\nu)dw+o(\alpha) \\ & \approx & \int_{0}^{\infty}\frac{\kappa_0}{\pi}\exp\left[ -\frac{1}{2} \left\{ \frac{ c'_{VT}(1-\alpha)w}{\nu^{1/2}} -\eta_0\right\}\right]f(w,\nu) \end{aligned}$$ 2. for $p=2$ $$\begin{aligned} \alpha&\leqslant & 2\int_{0}^{\infty} \mathrm{P}\left\{\sup_{\bm{c}\in\mathcal{C}} Z > \frac{c_{VT}(1-\alpha)\xi_0 }{\nu^{1/2}} w -\eta_0 \right\}\ f(w,\nu)dw+o(\alpha) \\ &\approx & \int_{0}^{\infty} \left[ \frac{\kappa_0 \left\{ c'_{VT}(1-\alpha) w / \nu^{1/2} -\eta_0 \right\} }{2^{1/2} \pi^{3/2} } \exp\left[ -\frac{1}{2} \left\{\frac{ c'_{VT}(1-\alpha)w}{\nu^{1/2}} -\eta_0\right\}^2\right] \right. \\ & & \left. + \frac{\zeta_0}{2\pi}\exp\left[ -\frac{1}{2} \left\{ \frac{ c'_{VT}(1-\alpha)w}{\nu^{1/2}} - \eta_0\right\}^2\right]\right] f(w,\nu) dw \end{aligned}$$ 3. for $p\geqslant3$. $$\begin{aligned} \alpha &\leqslant & 2 P\left[\sup_{\bm{c}\in\mathcal{C}} Z > \frac{\hat{\sigma}_e\nu^{1/2}}{\sigma_e\nu^{1/2}} \left\{ c_{VT}(1-\alpha) \xi_0 - \eta_0 \frac{\sigma_e }{ \hat{\sigma}_e } \right\} \right] \end{aligned}$$ The results in Proposition 2 follow immediately from these expressions. For the manifold $\mathcal{M}$ with the boundary we added the boundary’s correction terms $\mathcal{E}P\{|t_v|> c_{VT}(1-\alpha)\}$ and $2\mathcal{E}P\{|t_v|>c_{VT}(1-\alpha)\}$ for $p=1$ and $p=2$ respectively. Details of the empirical analysis of income data in Galicia {#sec:Appendix_app_Galicia} =========================================================== Table \[tab:desc\_stat\] displays the covariates with their standard deviations as well as the estimated REML coefficients with standard errors and p-values. The variable selection was carried out applying a method introduced by [@lombardia2017mixed]: a generalized AIC (xGAIC), which uses a quasi-likelihood and generalized degrees of freedom. Two types of covariates were selected by xGAIC – on the household level and on the individual level of the head of household. [max width=]{} Mean Stdev $\hat{\beta}$ S.E. p-value ----------- ------------------------------------------------------------------- -------- --------------- -------- --------- ------- Linc $\log(\text{household income} + 9834.658)$ 10.308 0.407 - - - SizeM = 1 if $<$ 10 000 0.362 0.481 -0.035 0.008 0.000 Type1 = 1 if households consists of 1 person 0.208 0.406 -0.615 0.011 0.000 Type2 = 1 if households consists of more than 1 person (without centre) 0.023 0.149 -0.301 0.021 0.000 Type3 = 1 if households consists of a couple with children 0.304 0.460 -0.198 0.010 0.000 Type4 = 1 if households consists of a couple without children 0.246 0.431 -0.338 0.010 0.000 Type5 = 1 if households consists of a single parent 0.093 0.290 -0.360 0.013 0.000 Ten1 = 1 if property without mortgage 0.663 0.473 0.096 0.009 0.000 Ten2 = 1 if property with mortgage 0.168 0.374 0.150 0.010 0.000 Dif1 = 1 if some difficulties coming to the end of a month 0.445 0.497 -0.171 0.006 0.000 Dif2 = 1 if a lot of difficulties coming to the end of a month 0.123 0.328 -0.389 0.010 0.000 Age1 = 1 if $\leqslant$ 44 0.238 0.426 -0.026 0.010 0.009 Age2 = 1 if 45 $\leqslant$ age $\leqslant$ 64 0.377 0.485 0.031 0.008 0.000 Educ1 = 1 if a primary education 0.232 0.422 -0.303 0.011 0.000 Educ2 = 1 if a secondary education 0.515 0.500 -0.231 0.008 0.000 Sex = 1 if male 0.625 0.484 0.021 0.007 0.001 Birth1 = 1 if born in Galicia 0.903 0.296 0.086 0.014 0.000 Birth2 = 1 if born in the rest of Spain 0.049 0.215 0.120 0.019 0.000 Intercept - - 10.742 0.022 0.000 : Descriptive statistics and coefficient estimates with standard errors and p-values.[]{data-label="tab:desc_stat"} [^1]: Katarzyna Reluga is a PhD candidate and a teaching assistant at the University of Geneva, GSEM, RCS, Switzerland. E-mail: [email protected]. [^2]: María José Lombardía is a Professor at the University of A Coruña, CITIC, Spain. E-mail: [email protected]. [^3]: Stefan Andreas Sperlich is a Professor at the University of Geneva, GSEM, RCS, Switzerland. E-mail: [email protected].\ [[The authors gratefully acknowledge support from]{} the MINECO grants MTM2017-82724-R and MTM2014-52876-R, the Xunta de Galicia (Grupos de Referencia Competitiva ED431C-2016-015 and Centro Singular de Investigación de Galicia ED431G/01), all of them through the ERDF. [We would like to thank]{} the Galician Institute of Statistics (IGE) for the transfer of the data. Finally, we thank D. Flores, W. Gonzalez Manteiga, E. López-Vizcaíno, D. Morales, T. Schmid, N. Salvati and S. Ranjbar for helpful discussions.]{}
{ "pile_set_name": "ArXiv" }
--- abstract: 'In a combined theoretical and experimental work, we investigate X-ray Absorption Near-Edge Structure (XANES) spectroscopy of the I $L_3$ and the Pb $M_5$ edges of the methylammonium lead iodide () hybrid inorganic-organic perovskite and its binary phase . The absorption onsets are dominated by bound excitons with sizable binding energies of a few hundred meV and pronounced anisotropy. The spectra of both materials exhibit remarkable similarities, suggesting that the fingerprints of core excitations in are essentially given by its inorganic component, with negligible influence from the organic groups. The theoretical analysis complementing experimental observations provides the conceptual insights required for a full characterization of this complex material.' author: - Christian Vorwerk - Claudia Hartmann - Caterina Cocchi - Golnaz Sadoughi - 'Severin N. Habisreutinger' - Roberto Félix - 'Regan G. Wilks' - 'Henry J. Snaith' - Marcus Bär - Claudia Draxl title: 'Exciton-Dominated Core-Level Absorption Spectra of Hybrid Organic-Inorganic Lead Halide Perovskites' --- ![image](./toc3.pdf){width="1.0\linewidth"} Hybrid organic-inorganic halide perovskites have emerged in the last few years as photovoltaic materials with remarkable efficiencies. Starting from the seminal work by Kojima *et al.* in 2009 [@MAPbI3seminal], solar cells based on hybrid perovskites have reached power conversion efficiencies (PCE) close to that of silicon wafer-based devices [@solar; @solar2; @solar3; @yin+15jmca], making them the *“next big thing in photovoltaics”* [@bigthing]. This astonishingly fast development has sparked great interest in the structural, electronic, and optical properties of the hybrid organic-inorganic halide perovskites [@Stoumpous; @exp_excitonbind; @bren+16natrm; @Schulz2014], also stimulating the theoretical community [@MAPbI3optic; @Zu2014; @MAPbI3opticalGW; @MAPbI3opticalGW2; @MAPbI3opticalDFT; @Menendez; @li-rink16prb; @gao+16prb; @Kresse2016; @Giustino2015; @Giustino2017; @MAPbI3GW2]. While the PCE achieved in laboratory environments is encouraging in view of commercial applications, the long-term stability of the samples remains a critical issue [@BertrandeChemMater2015; @YangInorgChem2017; @LeijtensAdvEnergyMat2015; @TiepAdvEnergyMat2016]. For methylammonium lead iodide (), one of the most studied metal-halide perovskites, previous X-ray photoelectron spectroscopy studies have identified sample degradation due to light irridation, both in the visible [@degradation:light] and in the X-ray [@degradation:x-ray; @bestpractices] region, as well as due to exposure to air and moisture [@degradation:water]. Under these conditions, may decompose into the binary phase . Additional decomposition to elemental metallic Pb and under X-ray illumination was recently reported [@degradation:light; @YangInorgChem2017]. Access to the local electronic and chemical structure can be achieved with X-ray Absorption Near-Edge Structure (XANES) spectroscopy. In this technique, the absorption of X-ray radiation in resonance with an atomic absorption edge yields information about the local environment surrounding the absorbing atom. Although XANES provides relevant species-specific information about phase separation and the electronic structure of the sample, the interpretation and rationalization of the spectra requires additional insight. In particular, an open question is whether the presence of degradation products, such as the binary phase , can be determined by X-ray absorption spectroscopy. The required analysis is provided by *ab initio* many-body theory, which combines an accurate description of the electronic structure of the system with an explicit treatment of excitonic effects, that can be crucial in the absorption of X-ray radiation from core electrons. Theoretical studies based on this methodology have been performed to determine the response of to visible light [@MAPbI3optic; @MAPbI3GW2; @Kresse2016; @Zu2014]. In optical spectroscopy, transitions from valence to conduction bands are probed, and valence excitons may form due to the Coloumb interaction between these states. Core level spectroscopy provides complementary information. With the selective excitation of the initial deep-lying states, core level spectroscopy allows a direct probing of the local environment of the excited species [@Weine2009; @Cocchi2016; @Fossard2016; @Schwartz2016; @Cocchi2015]. First-principle theory, in turn, enables a thorough characterization of the excitations in terms of band contributions and spatial extension. In this Letter, we present the results of a joint theoretical and experimental work, where we investigate XANES and core excitations from the I $L_{3}$ and the Pb $M_{5}$ edges in and . By exploring differences and similarities between the XANES of the perovskite and of its binary phase , we discuss the role of the inorganic part of the hybrid compound in determining the absorption behavior of . We perform a careful analysis of the measured and computed spectra, focusing on the main features at the absorption onset, where excitonic effects are especially pronounced. A detailed inspection of the character of the lowest-energy electron-hole pairs reveals their strong anisotropy and their sizable binding energy of a few hundreds meV. ![\[fig1:L3\] a) Calculated X-ray absorption near-edge spectra of the I $L_3$ edge of (red) and (black). A Lorentzian broadening of 0.5 eV is used in the calculations. b) Experimental XANES from the $L_3$ egde. For a direct comparison, calculated spectra with an increased broadening of 1.5 eV are included. The spectra of are offset by 0.5 for better readability. In both panels the difference between the spectra of and is shown in purple. The normalization of the calculated and experimental spectra is described in the Supporting Information. ](./L3_comparison2.pdf){width="0.425\linewidth"} We start our analysis by examining the XANES of and from the I $L_3$ edge, shown in Fig. \[fig1:L3\]. Due to the large spin-orbit splitting of 298.51 eV between I $2p_{3/2}$ and $2p_{1/2}$ electrons, the $L_2$ and $L_3$ edges can be treated independently within our theoretical framework (see Fig. S2 in the Supporting Information \[SI\]). The spectra of both materials (Fig. \[fig1:L3\]a) are characterized by the three main features, labeled A, B, and C, with only some differences in the intensity distribution. Our many-body perturbation theory (MBPT) results reveal that the feature A is formed by two bound excitons with binding energies of 480 meV in and 450 meV in . At higher transition energies, where peaks B and C appear, more pronounced differences in terms of oscillator strength are visible in the calculated XANES of the two compounds. The features A and B stem from interband transitions to the unoccupied I $d$ bands, and thus the intensity of these excitations reflects the differences in the electronic structure of the two materials. Note that in the calculated XANES a relatively small Lorentzian broadening of 0.5 eV is chosen to better resolve the fine structure of the absorption rise and the presence of the two above-mentioned resonances. A detailed analysis of the element-projected density of states (DOS) of both and is reported in the SI (Fig. S1). In the experimental spectrum (Fig. \[fig1:L3\]b), the fine structure is masked by the considerable intrinsic lifetime broadening, which amounts to at least 3.08 eV at the I $L_3$ edge [@natwidth; @atomdata]. For a better comparison with the experimental data in Fig. \[fig1:L3\]b, the calculated spectra are plotted with an increased broadening of 1.5 eV. Additional information illustrating the good agreement between theory and experiment is obtained by considering the difference between the spectra of and , as shown in Fig. \[fig1:L3\] in purple color. In this way, peaks A and B become visible also in the experimental results (Fig. \[fig1:L3\]b). Indeed, the computed spectra (Fig. \[fig1:L3\]a) reproduce not only the relative position of the maxima, but also the minimum at about 4561 eV. The overall magnitude of the computed signal is larger than in the experimental one due to the small lifetime broadening. ![\[fig3:M5\] X-ray absorption near-edge spectra of the Pb $M_5$ edge of (red) and (black) as obtained from (a) theory and (b) experiment. In panel (a), a Lorentzian broadening of 0.5 eV is included. For a direct comparison, panel (b) contains calculated spectra with an increased Lorentzian broadening of 1.5 eV. The spectra of are offset in intensity by 0.5 for better readability.](./M5_comparison.pdf){width="0.425\linewidth"} In contrast to the results from the I $L_3$ edge spectra discussed above, the Pb $M_5$ edge spectra of and shown in Fig. \[fig3:M5\] exhibit a distinct pre-edge feature both in theory (Fig. \[fig3:M5\]a) and experiment (Fig. \[fig3:M5\]b). This pronounced peak at 2485 eV has a clear excitonic nature. From MBPT calculations we identify this peak as formed by several bound electron-hole pairs with binding energies up to 860 meV in and 950 meV in (see also the discussion below related to Fig. \[fig5:BSEvsIP\]). At higher energies, around 2488 eV, we find an additional peak in the calculated spectrum of , which does not appear in that of . From the analysis of the projected DOS (see SI, Fig. S1), we can clarify that this weak peak in the perovskite originates from transitions to the Pb $p$ states above the band gap, which are present only in the electronic structure of but absent in the one of . In the experimental spectrum of (Fig. \[fig3:M5\]b) this peak is not visible, most likely due to the large lifetime broadening of about 3 eV [@atomdata], that characterizes also the excitations from the Pb $3d$ electrons. For direct comparison, Fig. \[fig3:M5\]b also displays calculated spectra with an increased Lorentzian broadening of 1.5 eV. As expected, the feature B is masked by the broadening. The measured XANES of displays an additional pre-edge peak at about 2481 eV, which does not appear in the spectrum of and is not reproduced by theory. Additional calculations (details reported in Fig. S3 in the SI) rule out effects due to a phase transition in the perovskite sample, as well as to the presence of metallic elemental Pb, which can be formed upon X-ray illumination, as reported in the recent literature [@degradation:light; @degradation:x-ray; @bestpractices]. X-ray transitions ascribed to - and - bonds are reported to be in this spectral range [@XANESLead], however, hard X-ray photoemission data (not shown) does not show any evidence for the presence of either sulfur and/or - at the surface of the studied samples. Therefore, some uncertainty remains on the origin of the pre-peak in the Pb $M_5$ edge spectrum of . The analysis of the XANES presented above shows remarkable similarities between the spectra of and for both considered absorption edges. While, on the one hand, this result may discourage the use of X-ray absorption techniques to detect the presence of residual in the hybrid perovskite samples, on the other hand, it confirms that the spectral fingerprints of are to a large extent determined by its inorganic cage. Our *ab initio* many-body approach can be exploited to further characterize the electron-hole pairs that dominate the absorption onset of the XANES of and from both I $L_3$ and Pb $M_5$ edges. In Fig. \[fig5:BSEvsIP\], we report the spectra of computed with and without the inclusion of electron-hole correlation, namely by solving the Bethe-Salpeter equation (BSE) and within the independent-particle approximation (IPA). In both panels, the dashed line indicates the onset of the IPA spectrum. For the bound excitons in the BSE spectra, the binding energy $E_b$ is obtained as the difference with respect to the IPA onset. As already mentioned above, we find $E_b=480$ meV and $E_b=860$ meV for the lowest-energy exciton in the I $L_3$ and Pb $M_5$ edge XANES of , respectively. The binding energies of the lowest-energy bound exciton in the Pb $M_5$ edge of is very similar to the one in , as reported recently [@Vorwerk2017]. Excitons in core spectra are considerably more strongly bound than those in the optical spectra, resulting in significantly larger binding energies compared to those for of valence excitons, for which binding energies of 40 meV have been reported for based on BSE calculations [@Zu2014]. This binding energy difference is mainly due to the localization of the involved core states. ![\[fig5:BSEvsIP\]X-ray absorption near-edge spectrum from the a) I $L_3$ and b) Pb $M_5$ edge of as obtained from the solution of the Bethe-Salpeter equation (BSE, red line) and within the independent-particle approximation (IPA, shaded area). The onset of the IPA spectrum is marked by the vertical dashed line, while the position of the lowest-energy bound exciton from BSE by the solid red line.](./IP_comparison.pdf){width="0.425\linewidth"} ![\[fig4:M\_contr\] Real-space distribution of the first exciton of in the XANES from a) the I $L_3$ and b) the Pb $M_5$ edge. The position of the core hole is fixed at the I and Pb atoms, respectively, and marked in both cases by a large dot in the color of the corresponding atom (purple for I and grey for Pb). c) Reciprocal-space analysis the first exciton of from c) the I $L_3$ and d) the Pb $M_5$ edge. The size of the red circles quantifies the contribution of each electronic state to the exciton. The Fermi energy is set to zero in the mid-gap and marked by a dashed line. e) Projected density of states of , with relevant contributions from the I $s$ and $d$ states, as well as Pb $p$ and $f$ states.](./exciton_contribution.pdf){width="0.98\linewidth"} In Fig. \[fig4:M\_contr\], we show the real- and reciprocal-space representation of the lowest-energy excitons in the XANES of from the I $L_3$ and the Pb $M_5$ edge. The excitonic wave-functions displayed in Fig. \[fig4:M\_contr\]a,b are remarkably similar. With the hole fixed at the position of the I and Pb atom, respectively (more details in the SI), the electron distribution is extended along the Pb-I bond ($y$ axis, see Fig. \[fig4:M\_contr\]a) over several unit cells, while in the directions perpendicular to this bond it is localized within one unit cell. This anisotropy is not visible in the absorption spectra, because the excitons extending along the bond are almost degenerate. The probability density associated with the excitonic wave-function in Fig. \[fig4:M\_contr\]a exhibits depletion along the Pb-I bond, due to the character of the hybridized Pb-I anti-bonding state formed at the bottom of the conduction band. While the wave-function distribution around the absorbing I atom is nearly spherically symmetric, the one around the neighboring Pb and I atoms reveals the contribution of the anisiotropic $p$ states. A similar character is found also for the first bound exciton in the spectrum of from the Pb $M_5$-edge (Fig. \[fig4:M\_contr\]b). Since the Pb atoms in the center of the octahedron are bound to six I atoms along the three Cartesian directions, a number of quasi-degenerate electron-hole pairs appear. Excitons with the largest binding energies, *i.e.*, with the lowest excitation energy, are two-fold degenerate. These two degenerate excitons are therefore symmetric along the Pb-I bonds perpendicular to the molecular axis. In Fig. \[fig4:M\_contr\]b we show the real-space distribution of the exciton extending within the $(\bar{1}00)$ plane. Bound excitons generated by radiation with polarization parallel to it have slightly smaller binding energies. Again, the anisotropic anti-bonding character directly stemming from the targeted unoccupied electronic states is clearly visible. The distribution of this electron-hole pair is also very similar to the one obtained for the Pb $M_{4,5}$-edge spectrum of  [@Vorwerk2017]. In that case, an analogous distribution along the Pb-I bonding direction is shown. However, a higher degree of in-plane localization appears in , where the exciton is confined within a single layer. The real-space distribution of the excitons is obviously reflected also in reciprocal space, as shown by the band-structure plots in Fig. \[fig4:M\_contr\]c,d. The contributions to both excitons arise mostly from the bottom of the conduction band at the R point, where the direct band gap of appears. Much smaller contributions originate from the R-X path. This distribution of core excitons in reciprocal space agrees very well with the one shown for bound excitons in the optical spectrum of obtained on the same level of theory [@Kresse2016]. The two lowest unoccupied bands, where most of the excitonic weight is localized, are characterized by a strong hybridization between Pb and I $p$ states, as shown in the projected DOS in Fig. \[fig4:M\_contr\]e. This hybridization is reflected in the difference in peak intensity between the two excitons: Transitions from the Pb $3d_{5/2}$ ($M_5$ edge) spectrum target directly the Pb $p$ contributions to the conduction band, giving rise to the intense peak in the spectrum. On the other hand, transitions from the I $2p_{3/2}$ levels ($L_3$ spectrum) are dipole-allowed only for the unoccupied I states with $s$ and $d$ character, which contribute to a much smaller extent to the lowest conduction band (Fig. \[fig4:M\_contr\]e). This explains the relatively low oscillator strength of the first peak in the I $L_3$ spectrum of (Fig. \[fig1:L3\]a) compared to the one at the onset of the $M_5$ edge (Fig. \[fig3:M5\]a). To summarize, in a joint theoretical and experimental work we have studied core level excitations from the I $L_3$ and the Pb $M_5$ edge of and its binary phase . We have shown that these two materials exhibit very similar XANES at both edges, suggesting the dominant contribution from the inorganic cage to the core excitations of . Differences between the spectra concern mainly minor features, which can hardly be exploited to detect the presence of the binary phase in hybrid perovskite samples. The absorption onset in the XANES of from both considered edges is characterized by pronounced excitonic effects. A detailed theoretical analysis shows that the bound excitons in the core spectra extend along the Pb-I bond for several unit cells, while being confined in the other directions. They originate from excitations targeting the lowest-energy conduction bands around high-symmetry points in the Brillouin zone. Moreover, the intensity of the peaks reveals the hybridization of iodine and lead derived states. Overall, this combined experimental and theoretical analysis of the core excitations offers a new perspective on the electronic structure in this complex material and its excitations. Experimental Methods ==================== *Theoretical Methods* – Core-level absorption spectra are obtained in a two-step process: First, the electronic structure is calculated from density-functional theory (DFT) [@hohe-kohn64pr; @kohn-sham65pr] employing the generalized gradient approximation for the exchange-correlation functional, as developed by Perdew, Burke and Ernzerhof [@perd+96prl]. Second, *ab initio* X-ray spectra are obtained in the framework of many-body perturbation theory (MBPT)[@Vorwerk2017] from the solution of the Bethe-Salpeter equation (BSE) [@hank-sham80prb; @Strinati]. In this approach, the many body problem is mapped into an effective two-particle Hamiltonian entering the eigenvalue equation $\hat{H}^{BSE}_{cu \mathbf{k}, c'u' \mathbf{k}'}A^{\lambda}_{c' u' \mathbf{k}'}=E^{\lambda}A^{\lambda}_{c u \mathbf{k}}$, where $(cu \mathbf{k})$ denotes any transition from a core ($c$) to an unoccupied state ($u$) at $\mathbf{k}$. The electron-hole Hamiltonian $\hat{H}^{BSE}=\hat{H}^{diag}+\hat{H}^{x}-\hat{H}^{d}$ accounts for the diagonal term $\hat{H}^{diag}$, which describes single-particle transitions in the independent-particle approximation, the exchange term $\hat{H}^{x}$, which describes the repulsive exchange interaction through the short-range Coulomb potential, and the direct term $H^{d}$, which includes the attractive screened Coulomb interaction between the electron and the hole. Scissors operators are applied to include the quasi-particle correction to the Kohn-Sham states from DFT. Their values of 86 eV for the I $L_3$ and 52 eV for the Pb $M_5$ edge are chosen to align the computed spectra to the experimental ones. Additional details about the BSE and its implementation within an all-electron full-potential framework are reported in Refs. . The eigenstates $A^{\lambda}$ provide information about the excitonic wavefunction, both in real space, $\Phi^{\lambda}(\mathbf{r}_e,\mathbf{r}_h)=\sum_{c u \mathbf{k}}A^{\lambda}_{cu \mathbf{k}}\psi_{u \mathbf{k}}(\mathbf{r}_e)\psi_{c \mathbf{k}}(\mathbf{r}_h)$, and in reciprocal space through the *exciton weights*, defined as $w^{\lambda}_{u \mathbf{k}}=\sum_{c}|A^{\lambda}_{cu \mathbf{k}}|^2$. Furthermore, the transition coefficients are obtained as $t^{\lambda}_{i}=\sum_{cu\mathbf{k}}A^{\lambda}_{cu \mathbf{k}}\frac{\langle c\mathbf{k}|p_i| c\mathbf{k} \rangle}{\epsilon_{u \mathbf{k}}-\epsilon_c}$, where $\epsilon_{u \mathbf{k}}$ and $\epsilon_{c}$ are the Kohn-Sham eigenvalues of the conduction and core state respectively. The binding energy of bound excitons is obtained as the difference between the transition energy $E^{\lambda}$ and the onset of independent-particle transitions. The macroscopic dielectric function $\epsilon_M$ is calculated as $\epsilon_M^{ij}(\omega)=\delta_{ij}+\sum_{\lambda}\frac{t^{\lambda}_{i}\left[t^{\lambda}_j\right]^{*}}{\omega-E^{\lambda}+\textrm{i}\Gamma}$. All calculations are performed with the all-electron, full-potential code `exciting` [@exciting] implementing DFT and MBPT and including an explicit treatment of core electrons. This is done by adopting the linearized augmented plane-wave (LAPW) basis set. Computational details are reported in the SI. *Experimental Methods* – perovskite thin films of 300 nm nominal thickness were prepared on compact /FTO/glass substrates at University of Oxford following the standard “one-pot” preparation approach [@solar; @lowtemp]. The compact layers were prepared by spin-coating an acidic solution of titanium isopropoxide dissolved in ethanol at 2000 rpm for 60s on fluorine-doped tin oxide (FTO) substrates (Pilkington, TEC7) followed by drying at $150\; ^{\circ}\textrm{C}$ and annealing at $500 \; ^{\circ}\textrm{C}$ for 45 min. The precursor solution for the perovskite was produced by dissolving methylammonium iodide (, “MAI”) and lead (II) chloride () in anhydrous N,N-dimethylformamide (DMF) in a 3:1 molar ratio with a final concentration of 2.64 mol/l MAI and 0.88 mol/l . This solution was spin-coated onto compact at 2000 rpm in a nitrogen-filled glovebox for 45 s. After spin-coating, the films were left to dry at room temperature inside the glovebox to allow the solvent to slowly evaporate, followed by an annealing step for 2.5 h at $90\;^{\circ}\textrm{C}$ (this step is needed for the crystallization and formation of the perovskite structure). After preparation, samples were sealed in a container under inert gas and transferred from the University of Oxford to the Helmholtz Zentrum Berlin für Materialien und Energie GmbH (HZB), where they were again unpacked and mounted on sample holders in a -purged glovebox. The samples were introduced into the load lock of the High Kinetic Energy Photoelectron Spectrometer (HiKE) endstation [@Gorgoi200948] (see below) with a N$_2$-filled glovebag to minimize exposure to ambient air. As reference sample commercially available powder (Sigma-Aldrich, 99.999% trace metals basis) was used. The powder was mounted in air onto the sample holder by pressing it onto double-sided carbon tape. XANES measurements of the I L$_{2,3}$ and Pb $M_{4,5}$ edges were carried out in the HiKE endstation [@Gorgoi200948] located at the BESSY II KMC-1 beamline [@Schaefers2007] at Helmholtz-Zentrum Berlin (HZB). For the I $L_{2,3}$ edge XANES spectrum the excitation energy was scanned through the range of 4500 – 4950 eV and for Pb $M_{4,5}$ – edge through 2400 – 2650 eV recording the I $\textrm{L}_{\alpha+\beta}$ and the Pb $\textrm{M}_{\alpha,\beta}$ emissions in partial fluorescence yield (PFY) mode, respectively. The (selected) fluorescence photons were detected with a Bruker XFlash4010 silicon drift detector with a beryllium window. The energy steps for the scanning of the excitation energy were varied for different energy ranges. For the I $L_3$ edge in Fig. \[fig1:L3\], the data points were measured with an energy step of 0.1 eV for and 0.5 eV for in the shown energy range. For the Pb $M_5$ edge in Fig. \[fig3:M5\], the energy step in the shown excitation energy range was 0.2 eV for and 0.1 eV for . For the energy calibration of the photon energy Au $4f$ peaks were always measured on a clean, electrically grounded Au foil, using the starting and ending excitation energies of the I $L_{2,3}$ and Pb $M_{4,5}$ - edge measurements. The Au $4f$ peaks were fitted with the fitting program “Fityk” [@fityk] version 0.9.8 using as an approximation a linear background and Voigt functions to fit the spin-orbit split doublet, by fixing the area ratio according to the multiplicy (2j+1) and coupling the peak shape. The spin-orbit splitting for Au $4f$ was set fixed to 3.67 eV [@handbook]. The photon energies were then calibrated by setting the Au $4f_{7/2}$ binding energy to 84.00 eV. Work partly funded by the German Research Foundation (DFG) through the Collaborative Research Center 951 HIOS and the GraFOx Leibniz ScienceCampus. Funding from the Helmholtz Energieallianz is also appreciated. C.V. acknowledges financial support from the Humboldt Research Track Scholarship of the Humboldt Universität zu Berlin. C.H. acknowledges support from the Potsdam University Helmholtz Zentrum Berlin graduate school HyPerCells: Perovskites Basic Research for High Efficiency Solar Cells. C.C. acknowledges support from the Berliner Chancengleichheitsprogramm and IRIS Adlershof. C.H., R.F., R.G.W., and M.B. additionally acknowledge funding from the Helmholtz Association (VH-NG- 423). We thank Helmholtz Zentrum Berlin for the allocation of synchrotron radiation beamtime for XANES measurements. We thank Karsten Hannewald for fruitful discussions in the early stage of the project. The Supporting Information contains additional details regarding the structural and electronic properties of and , as well as further information about the theoretical spectra. Details of the first-principles calculations are also reported and Refs.  are cited. [10]{} Kojima, A.; Teshima, K.; Shirai, Y.; Miyasaka, T. Organometal Halide Perovskites as Visible-Light Sensitizers for Photovoltaic Cells. *J. Am. Chem. Soc. * **2009**, *131*, 6050–6051 Lee, M. M.; Teuscher, J.; Miyasaka, T.; Murakami, T. N.; Snaith, H. J. Efficient Hybrid Solar Cells Based on Meso-Superstructured Organometal Halide Perovskites. *Science* **2012**, *338*, 643–647 McMeekin, D. P.; Sadoughi, G.; Rehman, W.; Eperon, G. E.; Saliba, M.; H[ö]{}rantner, M. T.; Haghighirad, A.; Sakai, N.; Korte, L.; Rech, B. A Mixed-Cation Lead Mixed-Halide Perovskite Absorber for Tandem Solar Cells. *Science* **2016**, *351*, 151–155 Yang, W. S.; Noh, J. H.; Jeon, N. J.; Kim, Y. C.; Ryu, S.; Seo, J.; Seok, S. I. High-performance Photovoltaic Perovskite Layers Fabricated Through Intramolecular Exchange. *Science* **2015**, *348*, 1234–1237 Yin, W.-J.; Yang, J.-H.; Kang, J.; Yan, Y.; Wei, S.-H. Halide Perovskite Materials for Solar Cells: A Theoretical Review. *J. Mater. Chem. A* **2015**, *3*, 8926–8942 Bisquert, J. The Swift Surge of Perovskite Photovoltaics. *J. Phys. Chem. Lett.* **2013**, *4*, 2597–2598 Stoumpos, C. C.; Malliakas, C. D.; Kanatzidis, M. G. Semiconducting Tin and Lead Iodide Perovskites with Organic Cations: Phase Transitions, High Mobilities, and Near-Infrared Photoluminescent Properties. *Inorg. Chem. * **2013**, *52*, 9019–9038 Miyata, A.; Mitioglu, A.; Plochocka, P.; Portugall, O.; Wang, J. T.-W.; Stranks, S. D.; Snaith, H. J.; Nicholas, R. J. [Direct Measurement of the Exciton Binding Energy and Effective Masses for Charge Carriers in Organic-Inorganic Tri-Halide Perovskites]{}. *Nat. Phys.* **2015**, *11*, 582–587 Brenner, T. M.; Egger, D. A.; Kronik, L.; Hodes, G.; Cahen, D. Hybrid Organic-Inorganic Perovskites: Low-Cost Semiconductors with Intriguing Charge-Transport Properties. *Nat. Rev. Mater.* **2016**, *1*, 15007 Schulz, P.; Edri, E.; Kirmayer, S.; Hodes, G.; Cahen, D.; Kahn, A. Interface Energetics in Organo-Metal Halide Perovskite-Based Photovoltaic Cells. *Energy Environ. Sci.* **2014**, *7*, 1377–1381 Ahmed, T.; [La-o-vorakiat]{}, C.; Salim, T.; Lam, Y. M.; Chia, E. E. M.; Zhu, J.-X. Optical Properties of Organometallic Perovskite: An Ab Initio Study Using Relativistic GW Correction and Bethe-Salpeter Equation. *EPL (Europhysics Letters)* **2014**, *108*, 67015 Zhu, X.; Su, H.; Marcus, R. A.; Michel-Beyerle, M. E. Computed and Experimental Absorption Spectra of the Perovskite $\mathrm{CH}_3\mathrm{NH}_3\mathrm{PbI}_3$. *J. Phys. Chem. Lett.* **2014**, *5*, 3061–3065, PMID: 26278260 Leguy, A. M. A.; Azarhoosh, P.; Alonso, M. I.; Campoy-Quiles, M.; Weber, O. J.; Yao, J.; Bryant, D.; Weller, M. T.; Nelson, J.; Walsh, A. Experimental and Theoretical Optical Properties of Methylammonium Lead Halide Perovskites. *Nanoscale* **2016**, *8*, 6317–6327 Mosconi, E.; Umari, P.; De Angelis, F. Electronic and Optical Properties of MAPbX3 Perovskites (X = I[,]{} Br[,]{} Cl): A Unified DFT and GW Theoretical Analysis. *Phys. Chem. Chem. Phys.* **2016**, *18*, 27158–27164 Lang, L.; Yang, J.-H.; Liu, H.-R.; Xiang, H.; Gong, X. First-Principles Study on the Electronic and Optical Properties of Cubic ABX$_3$ Halide Perovskites. *Phys. Lett. A* **2014**, *378*, 290 – 293 Menéndez-Proupin, E.; Palacios, P.; Wahnón, P.; Conesa, J. C. Self-Consistent Relativistic Band Structure of the $\mathrm{CH}{}_{3}\mathrm{NH}{}_{3}\mathrm{PbI}{}_{3}$ Perovskite. *Phys. Rev. B* **2014**, *90*, 045207 Li, J.; Rinke, P. Atomic Structure of Metal-Halide Perovskites from First Principles: The Chicken-And-Egg Paradox of the Organic-Inorganic Interaction. *Phys. Rev. B* **2016**, *94*, 045201 Gao, W.; Gao, X.; Abtew, T. A.; Sun, Y.-Y.; Zhang, S.; Zhang, P. Quasiparticle Band Gap of Organic-Inorganic Hybrid Perovskites: Crystal Structure, Spin-Orbit Coupling, and Self-Energy Effects. *Phys. Rev. B* **2016**, *93*, 085202 Bokdam, M.; Sander, T.; Stroppa, A.; Picozzi, S.; Sarma, D. D.; Franchini, C.; Kresse, G. Role of Polar Phonons in the Photo Excited State of Metal Halide Perovskites. *Sci. Rep. * **2016**, *6*, 28618 Filip, M. R.; Verdi, C.; Giustino, F. GW Band Structures and Carrier Effective Masses of $\mathrm{CH}_3\mathrm{NH}_3\mathrm{PbI}_3$ and Hypothetical Perovskites of the Type $\mathrm{APbI}_3$: A = $\mathrm{NH}_4$, $\mathrm{PH}_4$, $\mathrm{AsH}_4$, and $\mathrm{SbH4}_4$. *J. Phys. Chem. C* **2015**, *119*, 25209–25219 Volonakis, G.; Haghighirad, A. A.; Snaith, H. J.; Giustino, F. Route to Stable Lead-Free Double Perovskites with the Electronic Structure of $\mathrm{CH}_3\mathrm{NH}_3\mathrm{PbI}_3$: A Case for Mixed-Cation \[Cs/$\mathrm{CH}_3\mathrm{NH}_3$/$\mathrm{CH(NH}_2\mathrm{)}_2]_2\mathrm{InBiBr}_6$. *J. Phys. Chem. Lett.* **2017**, *8*, 3917–3924, PMID: 28745888 Umari, P.; Mosconi, E.; De Angelis, F. [Relativistic GW Calculations on CH$_3$NH$_3$PbI$_3$ and CH$_3$NH$_3$SnI$_3$ Perovskites for Solar Cell Applications]{}. *Sci. Rep. * **2014**, *4*, 4467 Philippe, B.; Park, B.-W.; Lindblad, R.; Oscarsson, J.; Ahmadi, S.; Johansson, E. M. J.; Rensmo, H. Chemical and Electronic Structure Characterization of Lead Halide Perovskites and Stability Behavior under Different Exposures – A Photoelectron Spectroscopy Investigation. *Chem. Mater. * **2015**, *27*, 1720–1731 Yang, J.; Kelly, T. L. Decomposition and Cell Failure Mechanisms in Lead Halide Perovskite Solar Cells. *Inorg. Chem. * **2017**, *56*, 92–101 Leijtens, T.; Eperon, G. E.; Noel, N. K.; Habisreutinger, S. N.; Petrozza, A.; Snaith, H. J. Stability of Metal Halide Perovskite Solar Cells. *Adv. Energy Mater. * **2015**, *5*, 1500963 Tiep, N. H.; Ku, Z.; Fan, H. J. Recent Advances in Improving the Stability of Perovskite Solar Cells. *Adv. Energy Mater. * **2016**, *6*, 1501420 Li, Y.; Xu, X.; Wang, C.; Ecker, B.; Yang, J.; Huang, J.; Gao, Y. Light-Induced Degradation of CH$_3$NH$_3$PbI$_3$ Hybrid Perovskite Thin Film. *J. Phys. Chem. C* **2017**, *121*, 3904–3910 Motoki, K.; Miyazawa, Y.; Kobayashi, D.; Ikegami, M.; Miyasaka, T.; Yamamoto, T.; Hirose, K. Degradation of CH$_3$NH$_3$PbI$_3$ Perovskite Due to Soft X-ray Irradiation as Analyzed by an X-ray Photoelectron Spectroscopy Time-Dependent Measurement Method. *J. Appl. Phys. * **2017**, *121*, 085501 Hoye, R. L. Z.; Schulz, P.; Schelhas, L. T.; Holder, A. M.; Stone, K. H.; Perkins, J. D.; Vigil-Fowler, D.; Siol, S.; Scanlon, D. O.; Zakutayev, A. Perovskite-Inspired Photovoltaic Materials: Toward Best Practices in Materials Characterization and Calculations. *Chem. Mater. * **2017**, *29*, 1964–1988 Li, Y.; Xu, X.; Wang, C.; Wang, C.; Xie, F.; Yang, J.; Gao, Y. Degradation by Exposure of Coevaporated CH$_3$NH$_3$PbI$_3$ Thin Films. *J. Phys. Chem. C* **2015**, *119*, 23996–24002 Olovsson, W.; Tanaka, I.; Mizoguchi, T.; Puschnig, P.; Ambrosch-Draxl, C. All-Electron Bethe-Salpeter Calculations for Shallow-Core X-Ray Absorption Near-Edge Structures. *Phys. Rev. B* **2009**, *79*, 041102 Cocchi, C.; Zschiesche, H.; Nabok, D.; Mogilatenko, A.; Albrecht, M.; Galazka, Z.; Kirmse, H.; Draxl, C.; Koch, C. T. Atomic Signatures of Local Environment from Core-Level Spectroscopy in $\ensuremath{\beta}\text{\ensuremath{-}}{\mathrm{Ga}}_{2}{\mathrm{O}}_{3}$. *Phys. Rev. B* **2016**, *94*, 075147 Fossard, F.; Hug, G.; Gilmore, K.; Kas, J. J.; Rehr, J. J.; Vila, F. D.; Shirley, E. L. Quantitative First-Principles Calculations of Valence and Core Excitation Spectra of Solid ${\mathrm{C}}_{60}$. *Phys. Rev. B* **2017**, *95*, 115112 Schwartz, C.; Nordlund, D.; Weng, T.-C.; Sokaras, D.; Mansfield, L.; Krishnapriyan, A. S.; Ramanathan, K.; Hurst, K. E.; Prendergast, D.; Christensen, S. T. Electronic Structure Study of the CdS Buffer Layer in CIGS Solar Cells by X-Ray Asorption Spectroscopy: Experiment and Theory. *Sol. Energy Mater. Sol. Cells* **2016**, *149*, 275 – 283 Cocchi, C.; Draxl, C. Bound Excitons and Many-Body Effects in X-Ray Absorption Spectra of Azobenzene-Functionalized Self-Assembled Monolayers. *Phys. Rev. B* **2015**, *92*, 205105 Krause, M. O.; Oliver, J. H. Natural Widths of Atomic K and L Levels, K$\alpha$ X-ray Lines and Several KLL Auger Lines. *J. Phys. Chem. Ref. Data* **1979**, *8*, 329–338 Keski-Rahkonen, O.; Krause, M. O. Total and Partial Atomic-Level Widths. *At. Data Nucl. Data Tables * **1974**, *14*, 139 – 146 Bovenkamp, G. L.; Prange, A.; Roy, A.; Schumacher, W.; Hormes, J. X-Ray Absorption Near Edge Structure Spectra as a Basis for the Speciation of Lead. *J. Phys. Conf. Ser.* **2009**, *190*, 012190 Vorwerk, C.; Cocchi, C.; Draxl, C. Addressing Electron-Hole Correlation in Core Excitations of Solids: An All-Electron Many-Body Approach from First Principles. *Phys. Rev. B* **2017**, *95*, 155121 Hohenberg, P.; Kohn, W. Inhomogeneus Electron Gas. *Phys. Rev. * **1964**, *136*, B864–B871 Kohn, W.; Sham, L. J. Self-Consistent Equations Including Exchange and Correlation Effects. *Phys. Rev. * **1965**, *140*, A1133–A1138 Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized Gradient Approximation Made Simple. *Phys. Rev. Lett. * **1996**, *77*, 3865–3868 Hanke, W.; Sham, L. J. Many-Particle Effects in the Optical Spectrum of a Semiconductor. *Phys. Rev. B* **1980**, *21*, 4656–4673 Strinati, G. Application of the Green’s Function Method to the Study of the Optical Properties of Semiconductors. *Riv. Nuovo Cimento * **1988**, *11* Puschnig, P.; Ambrosch-Draxl, C. Optical Absorption Spectra of Semiconductors and Insulators Including Electron-Hole Correlations: An Ab Initio Study within the LAPW Method. *Phys. Rev. B* **2002**, *66*, 165105 Sagmeister, S.; Ambrosch-Draxl, C. Time-Dependent Density Functional Theory versus [[BetheSalpeter]{}]{} Equation: An All-Electron Study. *Phys. Chem. Chem. Phys. * **2009**, *11*, 4451–4457 Gulans, A.; Kontur, S.; Meisenbichler, C.; Nabok, D.; Pavone, P.; Rigamonti, S.; Sagmeister, S.; Werner, U.; Draxl, C. Exciting: a Full-Potential All-Electron Package Implementing Density-Functional Theory and Many-Body Perturbation Theory. *J. Phys.: Condens. Matter. * **2014**, *26*, 363202 Ball, J. M.; Lee, M. M.; Hey, A.; Snaith, H. J. Low-Temperature Processed Meso-Superstructured to Thin-Film Perovskite Solar Cells. *Energy Environ. Sci.* **2013**, *6*, 1739–1743 Gorgoi, M.; Svensson, S.; Schäfers, F.; Öhrwall, G.; Mertin, M.; Bressler, P.; Karis, O.; Siegbahn, H.; Sandell, A.; Rensmo, H. The High Kinetic Energy Photoelectron Spectroscopy Facility at BESSY Progress and First Results. *Nucl. Instr. Meth. Phys. Res. A* **2009**, *601*, 48 – 53 Schaefers, F.; Mertin, M.; Gorgoi, M. KMC-1: A High Resolution and High Flux Soft X-ray Beamline at BESSY. *Rev. Sci. Instrum. * **2007**, *78*, 123102 Wojdyr, M. [[*Fityk*]{}: a General-Purpose Peak Fitting Program]{}. *J. Appl. Cryst. * **2010**, *43*, 1126–1128 Moulder, J. F.; Chastain, J. *Handbook of X-ray Photoelectron Spectroscopy: a Reference Book of Standard Spectra for Identification and Interpretation of XPS Data*; Physical Electronics Division, Perkin-Elmer Corp.: Eden Prairie, Minn, 1995 Brivio, F.; Walker, A. B.; Walsh, A. Structural and Electronic Properties of Hybrid Perovskites for High-Efficiency Thin-Film Photovoltaics from First-Principles. *APL Mater.* **2013**, *1*, 042111 Westenbrink, H. G.; Terpstra, P. On the Crystal Structure of Lead-Iodide. *Proc. K. Ned. Akad. Wet. * **1926**, *29*, 431–442 Davey, W. P. Precision Measurements of the Lattice Constants of Twelve Common Metals. *Phys. Rev. * **1925**, *25*, 753–761 Kawamura, Y.; Mashiyama, H.; Hasebe, K. Structural Study on Cubic–Tetragonal Transition of CH$_3$NH$_3$PbI$_3$. *J. Phys. Soc. Jpn. * **2002**, *71*, 1694–1697
{ "pile_set_name": "ArXiv" }
--- abstract: 'In this work we provide a framework for modeling compact stars in which the interior matter distribution obeys a generalised Chaplygin equation of state. The interior geometry of the stellar object is described by a spherically symmetric line element which is simultaneously comoving and isotropic with the exterior spacetime being vacuum. We are able to integrate the Einstein field equations and present closed form solutions which adequately describe compact strange star candidates like Her X-1, RX J 1856-37, PSRJ 1614-2230 and SAX J1808.4-3658.' author: - Piyali Bhar - Megan Govender - Ranjan Sharma date: 'Received: date / Accepted: date' title: Modeling Anisotropic Stars Obeying Chaplygin Equation of State --- [example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore Introduction ============ The search for exact solutions of the Einstein field equations has generated a rich field of models describing relativistic compact objects. Since the pioneering work of Schwarzschild who obtained the first interior solution describing a uniform density sphere, the modeling of relativistic stars has moved from the regime of toy models to sophisticated, realistic stellar structures. With the discovery of pulsars, neutron stars and strange stars there was a need to obtain relativistic analogues of Newtonian stars, particularly when the densities of the stellar material was of the order of $10^{14}$gm cm$^{-3}$. The simplistic model of a static uniform density star has been generalised to include the effects of pressure anisotropy, electric charge, scalar field, dark energy and the cosmological constant, on the gross physical properties of compact objects[@rayso; @ma1; @ma2]. Models of relativistic fluid spheres have also been obtained within the framework of higher order theories of gravity including the Randall-Sundrum brane scenario, Einstein-Gauss-Bonnet gravity and Lovelock formalisms[@sud1; @sud2; @abbas; @nare]. In order to close the system of equations governing the gravitational and thermodynamical behaviour of bounded objects, various techniques were employed by researchers working in this field of study: (i) imposition of symmetry, (ii) adhoc assumptions of the gravitational potentials, (iii) specific choices of the fall-off behaviour of the pressure, density or the anisotropy, to name a few[@dev1; @dev2; @iv1]. To construct a stellar model, a physically motivated route, in general, is to impose an equation of state which relates the pressure as a thermodynamical function of the density, ie., $p = p(\rho)$. Most of the earlier works were centered on imposing a linear equation of state of the form $p = \alpha \rho$ where $\alpha$ is a constant. This was later generalised to $p = \alpha \rho - \beta$, where $\beta = \alpha \rho_s$ and $\rho_s$ is the surface density[@sm1]. The conditions were relaxed by allowing for anisotropic pressure. Note that works in fundamental particle physics led to the MIT-bag model which hinged on an equation of state of the form $p = \alpha \rho - 4B$ where $B$ is the Bag constant. The linear equation of state was further generalised to the quadratic equation of state of the form $p = \alpha \rho - \beta + \sigma \rho^2$[@sifiso]. One of the first successful attempts to obtain a generalisation of the Newtonian polytrope was achieved by Buchdahl[@buch] in which he obtained a pseudo-relativistic version of the Lane-Emden polytrope of index $5$. Herrera and Barreto presented a general formalism to generate relativistic polytropes with anisotropic pressure in Schwarzschild coordinates[@poly]. Their findings also prompted further investigations into the origins of anisotropy, cracking in relativistic stellar models and stability[@azam]. In order to fine-tune these models with observations, some researchers employed a mixed polytrope equation of state in which two or more species of particles made up the stellar fluid. The inclusion of charge within the stellar core led to a plethora of static stellar models in which the role of the electromagnetic field on the stability, mass-radius ratio and redshift was demonstrated[@r1; @ra1; @ra2]. Issues surrounding the black hole horizon paradox necessitated the search for alternative models of black holes free of horizons. The gravastar model was first proposed by Mazur and Mottola [@mazur] which sought to address many of the problems encountered during the final stages of gravitational collapse. Dynamically the model hinged on the phenomenon that during the latter epoch of gravitational collapse spacetime itself would undergo phase transitions which would halt collapse. The emerging picture of a gravastar was that of a layered composite: a de Sitter interior filled with constant positive (dark) energy density $\rho$ featuring an isotropic negative pressure $\rho = -p >0$. This layer is then connected via three intermediate layers to an exterior vacuum Schwarzschild solution. The intermediate relatively thin shell is composed of stiff matter ($p = \rho$). Stability of the composite profile is achieved by utilising two infinitesimally-thin shells endowed with surface densities $\sigma_\pm$ and surface tensions $\vartheta_\pm$. An interesting model was proposed by Usmani [*et al*]{}[@ra2] in which they generalised the Mazur-Mottola gravastar picture to include charge. In addition, the interior of the gravastar admitted conformal motion. They were able to show that charged interior de Sitter void must generate the gravitational mass. This mass is accountable for the attractive force that counter-balances the electromagnetic repulsion due to the presence of charge during the collapse process[@rahaman1]. In a more recent model, Banerjee [*et al*]{}[@brig] presented a Braneworld generalisation of a gravastar admitting conformal motion. Motivated by the existence of dark energy, Lobo and coworkers[@lobo1] have proposed stellar models, the so-called ‘dark stars’ in which the equation of state of is of the form $p = \alpha \rho$ in which $-1 < \alpha < -1/3$. It has been proposed that in the phantom regime ($\alpha = -1$), the extremely high pressures may invoke a topological change rendering the dark energy star to a wormhole. An interesting proposal regarding dark energy and dark matter is treating them as different manifestations of a single entity. This proposal leads to the Chaplygin gas model in which the equation of state derives from string theory. Various applications of the Chaplygin gas model have been pursued in order to account for cosmological observations such as acceleration of the cosmic fluid and structure formation. The Chaplygin equation of state has been subsequently modified to a more generalized Chaplygin gas equation of state. The generalised Chaplygin equation of state has been employed to model dark stars which are remnants of continued gravitational collapse. The idea here is that the dark energy provides sufficient repulsion to halt collapse leading to stable bounded configurations free of horizons and singularities[@p1; @p2]. This paper is structured as follows: In section $2$ we introduce the field equations necessary for the modeling of a spherically symmetric star within the framework of general relativity. In section $3$. we present a particular solution describing the interior of the star in which the matter content obeys a generalised Chaplygin equation of state. The junction conditions required for the smooth matching of the interior spacetime to the exterior Schwarzschild solution are worked out in section $4$. A detailed physical analysis of the geometrical and thermodynamical behaviour of our model is presented in section $5$. We discuss the stability, energy conditions and mass-radius relation in sections $6$, $7$ and $8$, respectively. We conclude with a discussion of our results in section $9$. Spherically symmetric spacetime =============================== We consider a model which represents a static spherically symmetric anisotropic fluid configuration obeying a generalised Chaplygin equation of state. The interior spacetime is described by a spherically symmetric line element which is simultaneously comoving and isotropic $$ds^2 = -A^2(r)dt^2 + B^2(r)\left[dr^2 + r^2d\Omega^2\right], \label{metric}$$ where $d\Omega^2 = d\theta^2 + \sin^2{\theta}d\phi^2$ and the metric functions, $A(r)$ and $B(r)$ are yet to be determined. For our model the energy-momentum tensor for the stellar fluid is $$T_{ab} = {\mbox diag}\left(-\rho, p_r, p_t, p_t\right),\label{2}$$ where $\rho$, $p_r$ and $p_t$ are the proper energy density, radial pressure and tangential pressure, respectively. The fluid four–velocity ${\bf u}$ is comoving and is given by $$u^a = \displaystyle\frac{1}{A} \delta^{a}_0 \,. \label{2'}$$ The Einstein field equations for the line element (\[metric\]) are $$\label{g3a} 8\pi\rho = - \frac{1}{B^2} \left( 2\frac{B''}{B} - \frac{{B'}^2}{B^2} + \frac{4}{r}\frac{B'}{B} \right),$$ $$\label{g3b} 8\pi p_r = \frac{1}{B^2} \left[\frac{{B'}^2}{B^2} + 2\frac{A'}{A}\frac{B'}{B} + \frac{2}{r} \left(\frac{A'}{A} + \frac{B'}{B}\right) \right],$$ $$\label{g3c} 8\pi p_t=\frac{1}{B^2} \left[\frac{A''}{A} + \frac{B''}{B}- \frac{{B'}^2}{B^2}+\frac{1}{r} \left(\frac{A'}{A}+\frac{B'}{B}\right)\right],$$ where primes denote differentiation with respect to the radial coordinate $r$. We have utilized geometrized units in deriving the above system of equations in which the coupling constant and the speed of light are taken to be unity. The mass of the spherical object is given by $$\label{mass} m(r)= 4\pi\int_0^r\omega^2 \rho(\omega)d\omega,$$ where $\omega$ is an integration variable. In order to close the system of equations, we assume that the interior matter distribution obeys a generalised Chaplygin equation of state of the form $$\label{eos} p_r = H\rho - \frac{K}{\rho}$$ where $H$ and $K$ are positive constants. Substituting (\[g3a\]) and (\[g3b\]) in (\[eos\]) we obtain $$\label{master} \frac{A'}{A} = \frac{1}{2}\left(\frac{B'}{B}+\frac{1}{r}\right)^{-1}[G(r)-F(r)],$$ where $$G(r) =\frac{(8\pi)^{2}KB^{4}}{ 2\frac{B''}{B} - \frac{{B'}^2}{B^2} + \frac{4}{r}\frac{B'}{B}},$$ $$F(r)=2H\frac{B''}{B}+(1-H)\left(\frac{B'}{B}\right)^{2}+(1+2H)\frac{2}{r}\frac{B'}{B}.$$ On integrating (\[master\]) we obtain $$\label{g5} A= d \exp \left[\frac{1}{2}\int H(r) dr\right],$$ where $$\label{h} H(r) = \left(\frac{B'}{B}+\frac{1}{r}\right)^{-1}[G(r)-F(r)],$$ and $d$ is a constant of integration. Therefore, the line element (\[metric\]) can now be written as $$\label{g6} ds^2 = -d^2 \exp \left[\int H(r) dr\right] dt^2 + B^2 \left[dr^2 + r^2d\Omega^2\right],$$ where $H(r)$ is given in (\[h\]). Hence, any solution describing a static spherically symmetric anisotropic matter distribution obeying a generalised Chaplygin equation of state in isotropic coordinates can be easily determined by a single generating function $B(r)$. Generating solutions ==================== In order to close the system of equations several choices for $B(r)$ can be made. It is interesting to note that the choice of the metric potential $B(r)$ determines the gravitational and thermodynamical behaviour of the model. Hence the choice of $B(r)$ must satisfy all the requirements for a realistic stellar model. Recent work by Naidu and Govender[@ng2016] have shown that the end-state of gravitational collapse resulting from a dynamically unstable static core is ‘sensitive’ to the choice of the initial metric functions. They show that for the same $B(r)$ but with two distinct initially static cores; (i) vanishing radial pressure within the static configuration and (ii) uniform density interior, the final outcome of dissipative collapse leads to very different temperature profiles. Following Govender and Thirukkanesh[@gov1] we utilise the physically motivated choice for $B(r)$ as $$\label{b} B(r) = \frac{a}{\sqrt{1 + br^2}}$$ where $a$ and $b$ are constants. One can easily verify that the gravitational potential $B$ in (\[b\]) satisfies the regularity conditions, $B(0)$ = constant and $B'(r) = 1$ at the origin. The same expression of $B(r)$ was previously utilized to model compact objects in curvature coordinates by Schwarzschild [@sch], Einstein [@ein] and de Sitter [@des] and more recently in comoving coordinates by Govender and Thirukkanesh [@gov1] and Thirukkanesh [*et al.*]{} [@thir].\ With this choice of $B(r)$, we obtain from (\[g5\]) $$A(r) = d\exp\left[\left\{\frac{1+H}{4}-\frac{16a^{4}K\pi^{2}}{b^{2}}\right\}(1+br^{2})\right](1 + br^2)^{\frac{1+5H}{4}}(6 + br^2)^{\frac{80a^4K\pi^{2}}{b^2}},$$ where $d$ is a constant of integration.\ Subsequently, the field equations yield $$\begin{aligned} \rho &=& \frac{b(6 + br^2)}{8\pi a^2(1 + br^2)}, \\ p_r &=& \frac{bH}{8\pi a^2}\left(\frac{6 + br^2}{1 + br^2}\right) - \frac{8\pi a^2K}{b}\left(\frac{1 + br^2}{6 + br^2}\right),\\ p_t &=& \frac{C_1+C_2r^2+C_3r^4+C_4r^6+C_5r^8 C_6r^{10}}{32\pi a^2 b^2 (6 + b r^2)^2 (1 + b r^2)},\end{aligned}$$ where, $C_i's$ (i=1,2,...6) are given by $$C_1=96 b (9 b^2 H - 16 a^4 K \pi^2)$$ $$C_2=16 [9 b^4 \{3 + H (10 + 9 H)\} - 96 a^4 b^2 (5 + 3 H) K \pi^2 + 256 a^8 K^2 \pi^4]$$ $$C_3=8 b [3 b^4 \{18 + H (47 + 36 H)\} - 32 a^4 b^2 (47 + 42 H) K \pi^2 + 2048 a^8 K^2 \pi^4]$$ $$C_4=8 b^2 [b^4 \{18 + H (43 + 27 H)\} - 16 a^4 b^2 (57 + 61 H) K \pi^2 + 3072 a^8 K^2 \pi^4]$$ $$C_5=4 b^3 \{b^2 (1 + H) - 64 a^4 K \pi^2\} \{b^2 (5 + 6 H) - 64 a^4 K \pi^2\}$$ $$C_6=b^4 \{b^2 (1 + H) - 64 a^4 K \pi^2\}^2.$$ We define the anisotropic factor as $$\Delta=p_t-p_r$$ which is repulsive in nature if $\Delta>0$ and attractive if $\Delta<0$. Matching Conditions =================== In this section we match the interior spacetime $({\cal M}_{-})$ to the exterior spacetime $({\cal M}_{+})$ described by the exterior Schwarzschild solution in comoving isotropic coordinates[@bonnor1] $$\begin{aligned} ds^2 &=& - \frac{\left( 1 - \frac{M}{2r} \right)^2}{\left( 1 + \frac{M}{2r} \right)^2} dt^2 + \nonumber\\ && \left( 1 + \frac{M}{2r} \right)^{4} [dr^2 + r^2 (d\theta^2 +\sin^2\theta d\phi^2)], \label{exterior-metric}\end{aligned}$$ where $M$ is the mass within a sphere of radius $R$. Matching of interior metric (\[metric\]) and exterior metric (\[exterior-metric\]) at the boundary $r=R$ leads to the constraints $$\begin{aligned} \label{g12} A(R) &=&\frac{\left(1 -\frac{M}{2R}\right)}{\left(1 +\frac{M}{2R}\right)},\\ \label{g13}B(R) &=&\left(1 +\frac{M}{2R}\right)^2,\end{aligned}$$ where $$\label{masss} M=m(R)=\frac{bR^3+15\left(R- \frac{\arctan [\sqrt{b}R]}{\sqrt{b}}\right)}{6a^2}.$$ The condition (\[g12\]) imposes the following restriction on the constant of integration $$\begin{aligned} d&=&\frac{12a^2R -\left[bR^3 +15\left(R- \frac{\arctan [\sqrt{b}R]}{\sqrt{b}}\right)\right]}{12a^2R +\left[bR^3 +15\left(R- \frac{\arctan [\sqrt{b}R]}{\sqrt{b}}\right)\right]} \nonumber\\ &&\times\exp\left[\left\{\frac{16a^{4}K\pi^{2}}{b^{2}}-\frac{1+H}{4}\right\}(1+bR^{2})\right](1 + bR^2)^{-\frac{1+5H}{4}}(6 + bR^2)^{-\frac{80a^4K\pi^{2}}{b^2}}\nonumber\\\end{aligned}$$ The condition (\[g13\]) implies $$\frac{a}{\sqrt{1+b R^2}}= \left[1+\frac{bR^3+15\left(R- \frac{\arctan [\sqrt{b}R]}{\sqrt{b}}\right)}{12a^2R}\right]^2,$$ which imposes a restriction on the parameters $a$ and $b$ which can be determined if we specify the radius of the sphere. To examine the behabiour of the model parameters like matter density, radial and transverse pressure etc. we assume $a = 1.3997$, $b=0.009$, and $H=0.294$. By using the matching conditions together with $p_r (r=R) = 0$, we obtain the constant $K = 2.04558 \times 10^{-7}$ for a star of radius $6.7~$km. The mass of the stellar configuration turns out to be $0.789~ M_{\odot}$ which is very close to the observed mass of the strange star candidate Her X-1[@rawls]. Physical Analysis ================= We are now in a position to discuss the the physical features of the model generated in the preceding section. In order to describe a realistic stellar structure our model must satisfy the following physical requirements : 1. Regularity of the gravitational potentials at the origin: In our model, $A^2(0)=d^2 6^{\frac{160a^4K \pi^{2}}{b^2}}e^{[b^2(1+{H})-64a^4K\pi^{2}]/2b^2},~ B^2(0)=a^2$ which are constants and $(A^2(r))'=(B^2(r))'=0 $ at the origin $r=0$, which indicates that the gravitational potentials are regular at the origin. 2. Positive definiteness of the energy density and pressure at the centre: Since $\rho(0)= \frac{3b}{4\pi a^2}$, the energy density is positive and regular at the origin. We also have $p_r(0)=\frac{9b^2H - 16\pi^{2}a^4K}{12\pi a^2b}$. To ensure that the radial pressure is positive at the center we must have $\frac{K}{H} < \left(\frac{3b}{4\pi a^2}\right)^{2}$. Moreover, $\displaystyle \frac{d\rho}{dr} = -\frac{5b^2r}{4\pi a^2(1+br^2)^2} <0$ i.e., the energy density is a decreasing function of $r$. We also note that $\displaystyle\frac{dp_r}{dr} = -\frac{80a^2K \pi r}{(6 + br^2)^2} - \frac{5b^2Hr}{4\pi a^{2}(1 + br^2)^2}$, which implies that $p_r$ is a decreasing function of $r$. 3. Continuity of the extrinsic curvature across the matching hyper-surface, $K^{-}_{ij} = K^{+}_{ij}$: Continuity of the extrinsic curvature across the matching hyper-surface, $r=R$ yields $$(p_r)_{(r=R)} = 0;$$ which gives $$R=\sqrt{\frac{2(32 a^4 K \pi^2+20 a^2 b \pi \sqrt{HK}-3b^{2}H)}{b(b^2 H - 64 a^4 K \pi^2)}},$$ which is finite for appropriate choice of parameters $a, b, H$ and $K$. 4. Ratio of trace of stress tensor to energy density $(p_r+2p_t)/\rho$: Fulfillment of the requirement that the ratio of trace of stress tensor to energy density $(p_r+2p_t)/\rho$ should decrease radially outward is shown graphically in Fig. (\[diag\]). 5. Velocity of sound: For causality to be obeyed the radial and transverse velocities of sound should be in between \[$0,~1$\]. The radial velocity $(v_{sr}^{2})$ and transverse velocity $(v_{st}^{2})$ of sound can be obtained as $$v_{sr}^{2}=\frac{dp_r}{d\rho},$$ $$v_{st}^{2}=\frac{dp_t}{d\rho}.$$ Due to the complexity, we illustrate the causality conditions with the help of graphical representations. Fig. (\[sv1\]) and Fig. (\[sv2\])clearly show that $0 <v_{sr}^{2}\leq 1$ and $0<v_{st}^{2} \leq 1$ everywhere within the stellar configuration. 6. Stability: Following Heintzmann and Hillebrandt[@hein], a model of anisotropic compact star will be stable if $\Gamma>\frac{4}{3}$ everywhere within the stellar interior where the adiabatic index $\Gamma$ is defined as $$\Gamma=\frac{\rho+p_r}{p_r}\frac{dp_r}{d\rho}.$$ Fig. (\[gamma\]) shows that values of the adiabatic index which clearly indicates that the particular configuration developed in this paper is stable. 7. Energy conditions: A realistic star should satisfy the energy conditions namely, the Weak Energy Condition (WEC), Null Energy Condition (NEC) and Strong Energy Condition (SEC) as given below: $$\label{ec1} (i)\,\,\mbox{Null energy condition (NEC)}: \rho \geq 0,$$ $$\label{ec2} (ii) \,\,\mbox{Weak energy condition (WEC)}: \rho-p_r \geq 0, \rho-p_t \geq 0,$$ $$\label{ec3} (iii)\,\,\mbox{Strong energy condition (SEC)}: \rho-p_r-2p_t \geq 0.$$ For the specific stellar configuration developed here, validity of the inequalities (\[ec1\])-(\[ec3\]) have been shown with the help of graphical representations in Fig. (\[ec\]). ![Matter density $\rho$ is plotted against the radial distance $r$ inside the fluid for a particular configuration with $a=1.3997$, $b=0.009$, $H=0.294$ and $K=2.04558\times10^{-7}$](rho.eps) .\[rho\] ![Radial pressure $p_r$ is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="pr"}](pr.eps) ![Transverse pressure $p_t$ is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="pt"}](pt.eps) ![$\frac{p_r+2p_t}{rho}$ is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="diag"}](diag.eps) ![Anisotropic factor is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="delta"}](delta.eps) ![Square of radial velocity of sound is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="sv1"}](sv1.eps) ![Square of transverse velocity of sound is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="sv2"}](sv2.eps) ![The adiabatic index $\Gamma$ is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="gamma"}](gamma.eps) ![Mass function is is plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="mass"}](mass.eps) ![Weak, Null and Strong energy conditions are plotted against the radial distance $r$ inside the stellar interior by employing the same values of the constants as mentioned in Fig. (\[rho\]).[]{data-label="ec"}](ec.eps) Mass-radius relation ==================== The mass function in our model is obtained from (\[mass\]) as $$m(r)=\frac{br^3+15\left(r-\frac{\arctan [\sqrt{b}r]}{\sqrt{b}}\right)}{6a^2} \label{mass1}$$ The profile of the mass function is shown in Fig. (\[mass\]). Since at $r\rightarrow 0$ we have $m(r)\rightarrow 0$, it implies that the mass function is free from any central singularity. Buchdahl[@buchdahl59] obtained an upper bound on the mass to radius ratio i.e., compactness $u$ of a relativistic star of a compact star such that $M/r < 4/9$. In our model, we have $$u(r)=\frac{br^3+15\left(r-\frac{\arctan [\sqrt{b}r]}{\sqrt{b}}\right)}{6a^2 r}. \label{ur}$$ The compactness $u$ for different compact star models are given in Table 2. The table shows that compactness of configurations are wihin the Buchdahl limit[@buchdahl59]. We have also determined the surface redshift using the formula $$1+z_s=\left(1-\frac{2m}{R}\right)^{-\frac{1}{2}},$$ which for our model turns out to be $$z_s=\left(1-\frac{bR^3+15\left(r-\frac{\arctan [\sqrt{b}R]}{\sqrt{b}}\right)}{3a^2 R}\right)^{-\frac{1}{2}}-1.$$ The values of the surface redshift parameter for different stellar configurations are given in Table  1. For an isotropic star, in the absence of a cosmological constant, Buchdahl[@buchdahl59] and Straumann [@stra] have shown that $z_s \leq 2$. Böhmer and Harko[@bh] showed that for an anisotropic star, in the presence of a cosmological constant, the surface redshift can take a much higher value $z_s \leq 5$. The restriction was subsequently modified by Ivanov [@ivanov] who showed that the maximum permissible value could be as high as $z_s = 5.211$. In our case, we have $z_s \leq 1$ for different compact star models developed in this paper. Compact Star $R~$(km) a b H K $M~$($M_{\odot}$) ------------------ ---------- -------- -------- ------- ------------- ------------------- Her X - 1 6.7 1.3997 0.009 0.294 0.00012910 0.789 RX J 1856-37 6.006 1.6859 0.024 0.21 0.00020280 0.904 PSRJ 1614-2230 11.3664 2.0952 0.0138 0.29 0.000022416 1.86 SAX J1808.4-3658 7.07 2.1015 0.036 0.29 0.000149572 1.157 : Values of the constants $a$, $b$, $H$ and $K$ for different compact star models. ------------------ ------------------------ ------------------------ ----------------------- -------- -------- -- Compact Star central density surface density central pressure $u$ $z_s$ gm cm$^{-3}$ gm cm$^{-3}$ dyne cm$^{-2}$ Her X - 1 $1.4805\times10^{15}$ $1.1255\times10^{15}$ $1.6535\times10^{35}$ 0.1737 0.2379 RX J 1856-37 $2.7215\times10^{15}$ $1.6691\times10^{15}$ $3.2088\times10^{35}$ 0.2219 0.3409 PSRJ 1614-2230 $1.0131\times 10^{15}$ $0.4722\times 10^{15}$ $2.0698\times10^{35}$ 0.2413 0.3905 SAX J1808.4-3658 $2.6271\times10^{15}$ $1.2199\times 10^{15}$ $5.3784\times10^{35}$ 0.2414 0.3906 ------------------ ------------------------ ------------------------ ----------------------- -------- -------- -- : Values of physical parameters for different compact star models. Discussions =========== In this paper we have presented a new model of a compact star in isotropic coordinates which is free from central singularity with the exterior being the vacuum Schwarzschild spacetime. To solve the Einstein field equations we have employed the modified Chaplygin equation of state which is inspired by the current observation of the expanding universe and its connection to the existence of dark energy. By considering the observed radius of the compact star Her X-1 as an input parameter (we have assumed $R=6.7~$km), we have analyzed physical viability of our model. We note that the metric coefficients are free from any singularity. The variations of $\rho$, $\frac{p_r+2p_t}{\rho}$ ,$p_r,~p_t$ are plotted in Fig. (\[rho\]), (\[diag\]), (\[pr\]) and (\[pt\]), respectively which clearly show that matter density, radial and transverse pressure are positive inside the stellar interior and they are monotonically decreasing functions of radial coordinate. At the boundary of the star the matter density and transverse pressure are non-negative and the radial pressure vanishes as expected. From the plot of $\frac{p_r+2p_t}{\rho}$ (Fig. (\[diag\])) we note that it is non-zero and monotonically decreasing from the center towards the stellar surface. The anisotropic factor $\Delta=p_t-p_r$ is plotted against $r$ in Fig. (\[delta\]). Since $\Delta > 0$, the anisotropic factor is repulsive in nature in our model which is a desirable feature of a compact star[@gm]. Moreover, at the center of the star $\Delta$ vanishes which is also an essential feature of a realistic star. In order to investigate the relevance of our model in the study of compact stars, we have considered a number of compact stars, namely Her X-1, RX J 1856-37, PSRJ 1614-2230 and SAX J1808.4-3658 and showed that for the estimated radii the star’s masses determined from our model are very close to the observed masses (see Ref. [@maurya; @tj]). This leads naturally to the proposition that the solution obtained in this paper can be used as a viable model for describing ultra-compact stars. [99]{} S. K. Maurya, Y. K. Gupta, S. Ray and S. R. Chowdhury, [*Eur. Phys. J.*]{} C [**75**]{}, 389 (2015). P. Mafa Takisa and S. D. Maharaj, [*Astrophys. Space Sci.*]{} [**354**]{}, 463 (2014). S. D. Maharaj, J. M. Sunzu and S. Ray, [*Eur. Phys. J. Plus*]{} [**129**]{}, 3 (2014). S. Hansraj, B. Chilambwe and S. D Maharaj, [*Eur. Phys. J.*]{} C [**75**]{}, 277 (2015). S. D. Maharaj, B. Chilambwe and S. Hansraj, [*Phys. Rev.*]{} D [**91**]{}, 084049 (2015). G. Abbas, D. Momeni, M. A. Ali, R. Myrzakulov and S. Qaisar, [*Astrophys. Space Sci.*]{} [**357**]{}, 158 (2015). N. K. Dadhich, A. Molina and A. Khugaev, [*Phys. Rev.*]{} D [**81**]{}, 104026 (2010). K. Dev, and M. Gleiser, [*Gen. Relativ. Gravit.*]{} [**34**]{}, 1793 (2002). K. Dev, and M. Gleiser, M., [*Gen. Relativ. Gravit.*]{} [**35**]{}, 1435 (2003). B. V. Ivanov, [*J. Math. Phys.*]{} [**43**]{}, 1029 (2002). R. Sharma and S. D. Maharaj, [*Mon. Not. R. Astron. Soc.*]{} [**375**]{}, 1265 (2007). S. A. Ngubelanga, S. D. Maharaj and S. Ray, [*Astrophys. Space Sci.*]{} [**357**]{}, 74 (2015). H. A. Buchdahl, [*Phys. Rev.*]{} D [**116**]{}, 1027 (1959). L. Herrera and W. Barreto, [*Phys. Rev.*]{} D [**88**]{}, 084022 (2013)., [*Eur. Phys. J.*]{} C [**75**]{}, 123 (2015). M. Azam, S. A. Mardan and M. A. Rehman, [*Astrophys. Space Sci.*]{} [**359**]{}, 14 (2015). S. Ray, D. Ray and R. N. Tiwari, [*Astrophys. Space Sci.*]{} [**199**]{}, 333 (1993). S. Ray, B. Das, F. Rahaman and S. Ray, [*Int. J. Mod. Phys.*]{} D [**16**]{}, 1745 (2007). A. A. Usmani, F. Rahaman, S. Ray, K. K. Nandi, P . K. F. Kuhfittig, Sk A. Rakib and Z. Hasan, [*Phys. Lett.*]{} B[**701**]{}, 388 (2011). P. O. Mazur and E. Mottola, [*Proc. Nat. Acad. Sci.*]{} [**101**]{}, 9545 (2004). F. Rahaman et. al, [*Phys. Lett.*]{} B [**707**]{}, 319 (2012). A. Banerjee, F. Rahaman, S. Islam and M. Govender, [*Eur. Phys. J.*]{} C, (2015)(submitted). F. S. N. Lobo, [*Class. Quantum Grav.*]{} [**23**]{}, 1525 (2006). P. Bhar, [*Eur. Phys. J.*]{} C [**75**]{}, 123 (2015). P. Bhar, [*Astrophys. Space Sci.*]{} [**359**]{}, 41 (2015). K. Schwarzschild, [*Sitz. Deut. Akad. Wiss. Math.-Phys. Berlin*]{}, [**24**]{}, 424 (1916). A. Einstein, [*Sitz. Deut. Akad. Wiss. Math.-Phys.Berlin*]{} [**8**]{}, 142 (1917). W. de Sitter, [*Proc. Roy. Acad. Amst.*]{} [**19**]{}, 1217 (1917). N. F. Naidu and M. Govender, gr-qc/1602.02874 (2016). M. Govender and S. Thirukkanesh, [*Astrophys. Space Sci.*]{} [**358**]{}, 39 (2015). S. Thirukkanesh, M. Govender and D. B. Lortan, [*Int. J. Mod. Phys.*]{} D[**24**]{}, 1550002 (2015). H. Andréasson, [*Commun. Math. Phys.*]{} [**288**]{}, 715 (2009). M. L. Rawls et al., [*ApJ*]{} [**730**]{}, 25 (2011). W. B. Bonnor, A. K. G. de Oliveira and N. O. Santos, [*Phys. Rep.*]{}, [**181**]{}, 269 (1989). H.A. Buchdahl, [*Phys. Rev.*]{}, [**116**]{}, 1027 (1959. N. Straumann, [*General Relativity and Relativistic Astrophysics*]{} (Springer, Berlin, 1984) 43. C.G. Böhmer, T. Harko, [*Class. Quantum Gravit.*]{} [**23**]{}, 6479 (2006). B.V. Ivanov, [*Phys. Rev.*]{} D [**65**]{}, 104011 (2002) M. K. Gokhroo and A. L. Mehra, [*Gen. Relativ. Gravit.*]{} [**26**]{}, 75 (1994). R. Tikekar, and K. Jotania, [*Pramana J. Phys.*]{} [**68**]{}, 397 (2007). S. K. Maurya, Y. K. Gupta, Saibal Ray and Sourav Roy Chowdhury, [*Eur. Phys. J.*]{} C [**75**]{}, 389 (2015). Heintzmann, H., Hillebrandt, W.: Astron. Astrophys. 38, 51 (1975)
{ "pile_set_name": "ArXiv" }
--- abstract: 'Suppose $q$ is a prime power and $f\!\in\!{\mathbb{F}}_q[x]$ is a univariate polynomial with exactly $t$ monomial terms and degree $<\!q-1$. To establish a finite field analogue of Descartes’ Rule, Bi, Cheng, and Rojas (2013) proved an upper bound of $2(q-1)^{\frac{t-2}{t-1}}$ on the number of cosets in ${\mathbb{F}}^*_q$ needed to cover the roots of $f$ in ${\mathbb{F}}^*_q$. Here, we give explicit $f$ with root structure approaching this bound: For $q$ a $(t-1)$-st power of a prime we give an explicit $t$-nomial vanishing on $q^{\frac{t-2}{t-1}}$ distinct cosets of ${\mathbb{F}}^*_q$. Over prime fields ${\mathbb{F}}_p$, computational data we provide suggests that it is harder to construct explicit sparse polynomials with many roots. Nevertheless, assuming the Generalized Riemann Hypothesis, we find explicit trinomials having $\Omega\!\left(\frac{\log p}{\log \log p}\right)$ distinct roots in ${\mathbb{F}}_p$.' address: - 'School of Computer Science, University of Oklahoma, Norman, OK  73019' - 'Department of Mathematical Sciences, Clemson University, Clemson, SC  29634-0975' - 'TAMU 3368, College Station, TX  77843-3368' - 'Department of Mathematics, University of California, Irvine, CA  92697-3875 ' author: - Qi Cheng - Shuhong Gao - 'J. Maurice Rojas' - Daqing Wan title: | \ Sparse Univariate Polynomials\ with Many Roots Over Finite Fields --- Introduction ============ How can one best bound the complexity of an algebraic set in terms of the complexity of its defining polynomials? Over the complex numbers (or any algebraically closed field), Bézout’s Theorem [@bezout] bounds the number of roots, for a system of multivariate polynomials, in terms of the degrees of the polynomials. Over finite fields, Weil’s famous mid-20${^{\text{\underline{th}}}}$ century result [@weil] bounds the number of points on a curve in terms of the genus of the curve (which can also be bounded in terms of degree). These bounds are optimal for dense polynomials. For sparse polynomials, over fields that are not algebraically closed, these bounds can be much larger than necessary. For example, Descartes’ Rule [@descartes] tells us that a univariate real polynomial with exactly $t$ monomial terms always has less than $2t$ real roots, even though the terms may have arbitrarily large degree. Is there an analogue of Descartes’ Rule over finite fields? Despite the wealth of beautiful and deep 20${^{\text{\underline{th}}}}$-century results on point-counting for curves and higher-dimensional varieties over finite fields, varieties defined by sparse [*univariate*]{} polynomials were all but ignored until [@cfklls] (see Lemma 7 there, in particular). Aside from their own intrinsic interest, refined univariate root counts over finite fields are useful in applications such as cryptography (see, e.g., [@cfklls]), the efficient generation of pseudo-random sequences (see, e.g., [@bomb]), and refined estimates for certain exponential sums over finite fields [@bourgain Proof of Theorem 4]. For instance, estimates on the number of roots of univariate tetranomials over a finite field were a key step in establishing the uniformity of the [*Diffie-Helman distribution*]{} [@cfklls Proof of Thm. 8, Sec. 4] — a quantitative statement important in justifying the security of cryptosystems based on the Discrete Logarithm Problem. We are thus interested in the number of roots of sparse univariate polynomials over finite fields. The polynomial $x^q-x$ having two terms and exactly $q$ roots in ${\mathbb{F}}_q$ might suggest that there is no finite field analogue of Descartes’ rule. However, the roots of $x^q-x$ consist of $0$ and the roots of $x^{q-1}-1$, and the latter roots form the unit group ${\mathbb{F}}^*_q\!:=\!{\mathbb{F}}\setminus \{0\}$. For an arbitrary binomial $ax^n + bx^m \in {\mathbb{F}}_q[x]$ with $n\!>\!m$ and $a$ and $b$ nonzero, the roots consist of $0$ (if $m>0$) and the roots of $x^{n-m} + b/a$. Note that the number of roots of $x^{n-m} + b/a$ in ${\mathbb{F}}_q$ is either $0$ or $\gcd(n-m,q-1)$. In the latter case, the roots form a coset of a subgroup of ${\mathbb{F}}^*_q$. For polynomials with three or more terms, the number of roots quickly becomes mysterious and difficult, and, as we shall demonstrate in this paper, may exhibit very different behaviors in the two extreme cases where (a) $q$ is a large power of a prime, and (b) $q$ is a large prime. To fix notation, we call a polynomial $f(x)\!=\!c_1+c_2x^{e_2}+\cdots+c_tx^{e_t}\!\in\!{\mathbb{F}}_q[x]$ with $e_1\!<\!e_2\!<\cdots<\!e_t\!<\!q-1$ and $c_i\neq 0$ for all $i$ a [*(univariate) $t$-nomial*]{}. The best current upper bounds on the number of roots of $f$ in ${\mathbb{F}}_q$, as a function of $q$, $t$, and the coset structure of the roots of $f$, can be summarized as follows, using $|\cdot|$ for set cardinality: \[thm:z\] Let $f\!\in\!{\mathbb{F}}_q[x]$ be any univariate $t$-nomial with degree $<\!q-1$ and exponent set $\{e_1,\ldots,e_t\}$ containing $0$. Set $\delta(f)\!:=\!\gcd(e_1,\ldots,e_t,q-1)$, $Z(f)\!:=\!\left\{x\!\in\!{\mathbb{F}}_q\; | \; f(x)\!=\!0\right\}$, $R(f)\!:=\!\left|Z(f)\right|$, and let $C(f)$ denote the maximum cardinality of any coset (of any subgroup of ${\mathbb{F}}^*_q)$ contained in $Z(f)$. Then:\ 0. (Special case of [@karpshpar Thm. 1]) $R(f)\!\leq\!\frac{t-1}{t}(q-1)$.\ 1. [@bcr2 Thm. 1.1] $Z(f)$ is a union of no more than $2\left(\frac{q-1}{\delta(f)} \right)^{\frac{t-2}{t-1}}$ cosets, each associated to one of two subgroups $H_1\!\subseteq\!H_2$ of ${\mathbb{F}}^*_q$, where $|H_1|\!=\!\delta(f)$, $|H_2|\!\geq\!\delta(f)\left(\frac{q-1}{\delta(f)}\right)^{1/(t-1)}$, and $|H_2|$ can be determined within $2^{O(t)} (\log q)^{O(1)}$ bit operations.\ 2. [@ko Thm. 1.2] For $t\!=\!3$ we have $R(f)\!\leq\!\delta(f)\left\lfloor \frac{1}{2}+\sqrt{\frac{q-1}{\delta(f)}}\right\rfloor$ and, if we have in addition that $q$ is a square and $\delta(f)\!=\!1$, then $R(f)\!\leq\!\sqrt{q}$.\ 3. (See [@kelley Thms. 2.2 & 2.3]) For any $t\!\geq\!2$ we have $R(f)\!\leq\!2(q-1)^{\frac{t-2}{t-1}}C(f)^{1/(t-1)}$.\    \[1\][Furthermore, $\displaystyle{C(f)\leq\max\{k\!\in\!{\mathbb{N}}\; : \; k|(q-1) \text{ and, for all } i, \text{ there is a } j\!\neq\!i \text{ with } k|(e_i-e_j)\}}$. [$\blacksquare$]{}]{} For any fixed $t\!\geq\!2$, [*Dirichlet’s Theorem*]{} (see, e.g., [@bs Thm. 8.4.1, Pg. 215]) implies that there are infinitely many prime $q$ with $t|(q-1)$. For such pairs $(q,t)$ the bound from Assertion (0) is tight: The roots of $$f(x)\!=\!\frac{x^{q-1}-1}{x^{(q-1)/t}-1}\!=\!1 +x^{\frac{1}{t}(q-1)}+\cdots+x^{\frac{t-1}{t}(q-1)},$$ are the disjoint union of $t-1$ cosets of size $\delta(f)\!=\!\frac{q-1}{t}$. (There are no $H_2$-cosets for this $t$-nomial.) However, Assertions (1) and (3) tell us that we can get even sharper bounds by making use of the structure of the cosets inside $Z(f)$. For instance, when $t\!=\!3$ and $\delta(f)\!=\!1$, Assertion (2) yields the upper bound $\sqrt{q}$, which is smaller than $\frac{2}{3}(q-1)$ for $q\!\geq\!5$. While Assertion (3) might sometimes not improve on the upper bound $\frac{t-1}{t}(q-1)$, it is often the case that $C(f)$ is provably small enough for a significant improvement. For instance, when $e_1\!=\!0$ and $\gcd(e_i,q-1)\!=\!1$ for all $i\!\geq\!2$, we have $C(f)\!=\!1$ and then $R(f)\!\leq\!2(q-1)^{\frac{t-2}{t-1}}$. Our first main result is two explicit families of $t$-nomials revealing that Assertions (1)–(3) are close to optimal for [*non*]{}-prime $q$. \[thm:opt\] Let $t,u,p\!\in\!{\mathbb{N}}$ with $t\!\geq\!2$ and $p$ prime. If $q\!=\!p^{(t-1)u}$ then the polynomial $$r_{t,u,p}(x):=1+x+x^{p^u}+\cdots+x^{p^{(t-2)u}}$$ has $\delta(r_{t,u,p})\!=\!C(r_{t,u,p})\!=\!1$ and exactly $q^{(t-2)/(t-1)}$ roots in ${\mathbb{F}}_q$. Furthermore, if $q\!=\!p^{tu}$, then the polynomial $$g_{t,u,p}(x):=1+x+x^{1+p^u}+\cdots +x^{1+p^u+\cdots +p^{(t-2)u}}$$ has $\delta(g_{t,u,p})\!=\!1$, $C(g_{t,u,p})\!\leq\!\lfloor t/2\rfloor$, and exactly $q^{(t-2)/t}+\cdots+q^{1/t}+1$ roots in ${\mathbb{F}}_q$. Theorem \[thm:opt\] is proved in Section \[sec:trick\] below. The polynomials $r_{t,u,p}$ show that the bounds from Assertions (1) and (3) of Theorem \[thm:z\] [*are within a factor of $2$ of being optimal*]{}, at least for $\delta(f)\!=\!C(f)\!=\!1$ and a positive density of prime powers $q$. Note in particular that $r_{3,u,p}$ shows that Assertion (2) of Theorem \[thm:z\] is [*optimal*]{} for square $q$ and $\delta(f)\!=\!1$. (See [@ko Thm. 1.3] for a different set of extremal trinomials when $q$ is an odd square.) The second family $g_{t,u,p}$ reveals similar behavior for a different family of prime powers. Optimally bounding the maximal number of roots (or cosets of roots) for the case of [*prime*]{} $q$ is already more subtle in the trinomial case: We are unaware of any earlier lower bound that is a strictly increasing function of $q$. (Note, for instance, that $1+x^{(q-1)/3}+x^{2(q-1)/3}$ vanishes on just cosets of roots, assuming $q\!=\!1$ mod $3$.) In Section \[sec:harder\], we shall prove the following theorem. \[thm:grh\] For any $n\!\geq\!2$ and prime $p\!\geq\!n+2$, consider $h_{n,p}(x)\!:=\!x^n-x-1$ as an element of ${\mathbb{F}}_p[x]$. Then $\delta(h_{n,p})\!=\!C(h_{n,p})\!=\!1$, and there exists an infinite family of primes $p$ such that the trinomial $h_{n,p}$ has exactly $n$ distinct roots in ${\mathbb{F}}_p$ where 1. $n\!=\!\Omega\!\left(\frac{\log\log p}{\log\log \log p} \right)$ unconditionally, and 2. $n\!=\!\Omega\!\left(\frac{\log p}{\log \log p}\right)$ if the Generalized Riemann Hypothesis (GRH) is true. This theorem is proved by using results from algebraic number theory on the splitting of primes in number fields. In particular, we use a classic estimate of Lagarias, Montgomery, and Odlyzko [@lmo] (reviewed in Section \[sec:harder\] below) on the least prime ideal possessing a Frobenius element exhibiting a particular Galois conjugacy class. The latter result is in turn heavily based on the effective Chebotarev Density Theorem of Lagarias and Odlyzko [@lo]. Cohen [@Coh70; @Coh72] applied an effective function field analogue of the Chebotarev Density Theorem to trinomials of the form $x^n + x + b$, where $b$ varies in ${\mathbb{F}}_p$ and $p$ goes to infinity. His results tell us that when $p$ is large, there always exist $b\!\in\!{\mathbb{F}}_p$ so that $x^n + x + b$ splits completely over ${\mathbb{F}}_p$, and the least such $p$ satisfies $n\!=\!\Omega\!\left(\frac{\log p}{\log \log p}\right)$ unconditionally. This gives the existence of $b\in {\mathbb{F}}_p$ (and infinitely many primes $p$) with $x^n+x+b$ having many roots in ${\mathbb{F}}_p$, but gives no information on how to easily find such $b$. The novelty of Theorem \[thm:grh\] is thus an even simpler family of trinomials with number of roots in ${\mathbb{F}}_p$ increasing as $p$ runs over an infinite sequence of primes. Assuming GRH, the rate of increase matches Cohen’s lower bound. Could the existence of trinomials $f$ with $\delta(f)\!=\!1$ and, say, $\Omega\!\left(\sqrt{p}\right)$ roots in ${\mathbb{F}}_p$ (as one might conjecture from Theorem \[thm:opt\]) be obstructed by $p$ not being a square? We feel that $p$ being prime is such an obstruction and, based on experimental results below, we conjecture the following upper bound: \[conj:logp\] There is an absolute constant $c\!\geq\!1$ such that, for any prime $p\!\geq\!3$ and $\gamma,e_2,e_3\!\in\!\{1,\ldots,p-2\}$, with $e_3\!>\!e_2\!>\!0$ and $\gcd(e_2,e_3,p-1)\!=\!1$, the trinomial $\gamma+x^{e_2}+x^{e_3}$ has no more than $(\log p)^c$ roots in ${\mathbb{F}}_p$. (See also [@ko Conj. 1.5 & Sec. 4] for other refined conjectures and heuristics in this direction.) It is a simple exercise to show that, to compute the maximal number of roots in ${\mathbb{F}}_p$ of trinomials with $\delta(f)\!=\!1$, one can restrict to the family of trinomials in the conjecture. For any $n\!\in\!{\mathbb{N}}$, let $p_n$ be the least prime for which there exists a univariate trinomial $f_n$ with $\delta(f_n)\!=\!1$ and exactly $n$ distinct roots in ${\mathbb{F}}_{p_n}$. We did a computer search to find the values of $p_n$ for $1\!\leq\!n\!\leq 16$. They are... \[1\] $n$ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 ------- --- --- ---- ---- ---- ----- ----- ----- ----- ------ ------ ------ ------- ------- ------- ------ $p_n$ 3 5 11 23 47 151 173 349 619 1201 2753 4801 10867 16633 71237 8581 For example, $p_{16}\!=\!8581$ because $-364+363x+x^{2729}$ has exactly $16$ roots in ${\mathbb{F}}_{8581}$, and [*no*]{} trinomial $f\!\in\!{\mathbb{F}}_p[x]$ with $p\!<\!8581$ and $\delta(f)\!=\!1$ has more than $15$ roots in ${\mathbb{F}}_p$. In the appendix, we give representative trinomials for each $p_n$ above. To get a feel for how the maximal number of roots of a trinomial grows with the field size, let us compare the graphs (drawn darker) of the functions $0.91\log x$ and $1.77\log x$ with the \[1\][piecewise linear curve (drawn lighter) going through the sequence of points $((p_1,1),\ldots,(p_{12},12)$,]{} $(p_{16},16),(p_{13},13), (p_{14},14), (p_{15},15))$ as shown in Figure \[Fig1\] below. We used some simple [Maple]{} and [Sage]{} programs that took a few hours to find most of the data above. The value of $p_{15}$ took [C]{} code (written by Zander Kelley) running on a supercomputer for 2 days. Quantitative results on sparse polynomials over finite fields sometimes admit stronger analogues involving complex roots of unity. For instance, [@bcr1; @bcr2] and [@cheng] deal with the complexity of deciding whether a sparse polynomial vanishes on a (finite) subgroup of, respectively, ${\mathbb{F}}^*_q$ or ${\mathbb{C}}^*$. It is thus interesting to observe a recent complex analogue to our trinomial root counts over ${\mathbb{F}}_q$: Theobald and de Wolff [@tt] proved that, if $\gcd(e_2,e_3)\!=\!1$, a trinomial $c_1+c_2x^{e_2}+c_3x^{e_3}\!\in\!{\mathbb{C}}[x]$ can have at most $2$ complex roots of the same absolute value. So any such trinomial has at most $2$ roots in the [*torsion subgroup*]{} $\{\zeta\!\in\!{\mathbb{C}}\; | \; \zeta^n\!=\!1 \text{ for some } n\!\in\!{\mathbb{N}}\}$ of ${\mathbb{C}}^*$. This upper bound is sharp: Consider $(x-1)(x-\zeta)$ for any $\zeta\!\not\in\!\{\pm 1\}$ satisfying $\zeta^n\!=\!1$ for some $n\!\geq\!3$. Main Lower Bounds in the Prime Power Case {#sec:trick} ========================================= [**Proof of Theorem \[thm:opt\]**]{}: To establish the root count for $r_{t,u,p}$ it clearly suffices to prove that $r_{t,u,p}$ divides $x^q-x$ or, equivalently, $x^{p^{(t-1)u}}\!=\!x$ in the ring $R\!:=\!{\mathbb{F}}_p[x]/\langle r_{t,u,p}(x)\rangle$. Toward this end, observe that $x^{p^{(t-1)u}} = \left(x^{p^{(t-2)u}} \right)^{p^u} = \left(-1-x^{p^0}-\cdots-x^{p^{(t-3)u}}\right)^{p^u}$ in $R$. Since $(a+b)^{p^u}\!=\!a^{p^u}+b^{p^u}$ in any ring of characteristic $p$, we thus obtain by induction on $t$ that $x^{p^{(t-1)u}}=(-1)^{p^u}\left(1+x^{p^u}+\cdots+p^{(t-2)u}\right)$ in $R$. The last factor is merely $r_{t,u,p}(x)-x$, so we obtain $x^{p^{(t-1)u}}=(-1)^{p^u}(-x) = (-1)^{1+p^u}x$ in $R$. Since $(-1)^{1+p^k}\!=\!1$ in ${\mathbb{F}}_p$ for all primes $p$, we have thus established the root count for $r_{t,u,p}$. That $\delta(r_{t,u,p})\!=\!1$ is clear since $r_{t,u,p}$ has a nonzero constant term and $1$ as one of its exponents. Likewise, $\delta(g_{t,u,p})\!=\!1$. That $C(r_{t,u,p})\!=\!1$ is clear because the lowest exponents of $r_{t,u,p}$ are $0$ and $1$, the rest are powers of $p$, and $\gcd(p^k,q-1)\!=\!1$ for all $k\!\in\!{\mathbb{N}}$. We postpone proving our upper bound on $C(g_{t,u,p})$ until after we prove our stated root count for $g_{t,u,p}$. Consider now the set $S$ of elements in ${\mathbb{F}}_{q}$ whose trace to ${\mathbb{F}}_{p^u}$ is zero, that is, $$S:=\left\{ a \in {\mathbb{F}}_q \; | \; a + a^{p^u} + a^{p^{2u}} + \cdots + a^{p^{(t-1)u}} = 0\right\}.$$ Then $S$ has $q/p^u\!=\!p^{(t-1)u}$ elements and is a vector space of dimension $t-1$ over ${\mathbb{F}}_{p^u}$. Let $a \in S$ be nonzero. We show that $a^{p^u -1}$ is a root of $g_{t,u,p}$: $$g_{t,u,p}(a^{p^u -1})= 1 + \sum_{i=1}^{t-2} a^{(p^u -1) (p^{iu} + \cdots + p^u + 1)} = 1 + \sum_{i=1}^{t-2} a^{p^{(i+1)u} - 1} = \frac{1}{a} \sum_{i=0}^{t-1} a^{p^{iu}} =0.$$ Now note that, for any $a\!\in\!S$ and any nonzero $w\!\in\!{\mathbb{F}}_{p^u}$, the element $aw$ is also in $S$. Also, for $a,b\!\in\!{\mathbb{F}}_q$, $a^{p^u-1}\!=\!b^{p^u-1}$ if and only if $b\!=\!aw$ for some nonzero $w\!\in\!{\mathbb{F}}_{p^u}$. Therefore, when $a$ runs through $S\!\setminus\!\{0\}$, the element $a^{p^u-1}$ yields $(p^{(t-1)u} - 1)/(p^u -1)\!=\! 1 + p^u + \cdots + p^{(t-2)u}$ roots for $g_{t,u,p}$. To finally prove our upper bound on $C(g_{t,u,p})$, note that, for $j\!>\!i$, $$1+p^u+\cdots+p^{(j-2)u}\!=\!\frac{p^{(j-1)u}-1}{p^u-1},$$ and $$\frac{p^{(j-1)u}-1}{p^u-1}-\frac{p^{(i-1)u}-1}{p^u-1} \!=\!p^{(i-1)u}\left(\frac{p^{(j-i)u}-1}{p^u-1}\right).$$ So for $i\!\geq\!2$, $$\begin{aligned} \max\limits_{j\in\{1,\ldots,t\}\setminus\{i\}} \gcd\!\left(p^{(i-1)u}\left(\frac{p^{(j-i)u}-1}{p^u-1}\right), p^{tu}-1\right) & = & \max\limits_{j\in\{1,\ldots,t\}\setminus\{i\}} \gcd\!\left(\frac{p^{|j-i|u}-1}{p^u-1},p^{tu}-1\right)\\ & = & \max\limits_{\ell\in\{1,\ldots,\max\{i-1,t-i\}\}} \gcd\!\left(\frac{p^{\ell u}-1}{p^u-1},p^{tu}-1\right).\end{aligned}$$ Hence,   $D(g_{t,u,p}) := \min\limits_{i\in\{1,\ldots,t\}} \max\limits_{\ell\in\{1,\ldots,\max\{i-1,t-i\}\}} \gcd\!\left(\frac{p^{\ell u}-1}{p^u-1},p^{tu}-1\right)$ $$\begin{aligned} & = & \max\limits_{\ell\in\{1,\ldots,\lfloor t/2\rfloor \}\}} \gcd\!\left(\frac{p^{\ell u}-1}{p^u-1},p^{tu}-1\right)\\ & = & \max\limits_{\ell\in\{1,\ldots,\lfloor t/2\rfloor \}\}} \gcd\!\left(\frac{p^{\ell u}-1}{p^u-1},\left(\frac{p^{tu}-1}{p^u-1}\right) (p^u-1)\right)\\ & \leq & \max\limits_{\ell\in\{1,\ldots,\lfloor t/2\rfloor \}\}} \gcd\!\left(\frac{p^{\ell u}-1}{p^u-1},\frac{p^{tu}-1}{p^u-1}\right) \gcd\!\left(\frac{p^{\ell u}-1}{p^u-1},p^u-1\right),\\ & = & \max\limits_{\ell\in\{1,\ldots,\lfloor t/2\rfloor \}\}} \left(\frac{p^{\gcd(\ell,t) u}-1}{p^u-1}\right)\gcd\!\left(\ell,p^u-1\right). \end{aligned}$$ The last equality follows easily from two elementary facts: (1) $\gcd(x^\ell-1,x^t-1)\!=\!x^{\gcd(\ell,t)}-1$, and (2) $(x-1)(x^{\ell-2}+2x^{\ell-3}+\cdots+(\ell-2)x +(\ell-1))\!=\! x^{\ell-1}+\cdots+x^2+x-(\ell-1)$. So $D(g_{t,u,p})\!\leq\!\max\limits_{\ell\in\{1,\ldots,\lfloor t/2\rfloor \}} 1\cdot \ell \leq \lfloor t/2\rfloor$. By [@kelley Prop. 2.4] and [@kelley Thm. 2.2],$D(g_{t,u,p})\!\geq\!C(g_{t,u,p})$, so we are done. [$\blacksquare$]{} Main Lower Bounds in the Prime Case {#sec:harder} =================================== We’ll need several results from algebraic number theory. First, let $K$ be any number field, i.e., a finite algebraic extension of ${\mathbb{Q}}$. Let $d_{K}$ denotes the discriminant of $K$ over ${\mathbb{Q}}$, and ${\mathcal{O}}_K$ the ring of algebraic integers of $K$, i.e., those elements of $K$ with monic minimal polynomial in ${\mathbb{Z}}[x]$. We need to know the size of the smallest prime $p \in {\mathbb{Z}}$ that splits completely in ${\mathcal{O}}_K$. There are various bounds in the literature that are proved via some effective version of the Chebotarev Density Theorem. For instance: \[thm:least\] (See [@lo Cor. 1.2 & pp. 461–462] and [@lmo Thm. 1.1].) If $f\!\in\!{\mathbb{Z}}[x]$ is any irreducible polynomial of degree $n$ then the least prime $p$ for which the reduction of $f$ mod $p$ has $n$ distinct roots in ${\mathbb{F}}_p$ is (unconditionally) no greater than $d^{O(1)}_K$, where $K\!\subset\!{\mathbb{C}}$ is the splitting field of $f$. Furthermore, if GRH is true, then $p\!=\!O((\log d_K)^2)$. The papers [@lo; @lmo] in fact work in much greater generality: Our limited paraphrase above is simply the special case where one is looking for a prime yielding a Frobenius element corresponding to the identity element of the Galois group of $f$ over ${\mathbb{Q}}$. The best recent estimates indicate that, in the unconditional case of Theorem \[thm:least\], we can take the $O$-constant to be $40$, for sufficiently large $d_K$ [@kadiring]. Also, for an abelian extension $K$ over ${\mathbb{Q}}$, Pollack [@Poll] gives a much better bound (in the unconditional case): $p\!=\!O_{{\varepsilon},K}\!\left(d^{\frac{1}{4}+ {\varepsilon}}_K\right)$ where ${\varepsilon}>0$ is arbitrary, and the implied $O$-constant depends only on ${\varepsilon}$ and the degree of $K$ over ${\mathbb{Q}}$. We will also need good bounds on discriminants of number fields. In the following theorem, the lower bound is due to Minkowski [@minkowski] and the upper bound follows easily from work of Tôyama [@toyama] (by induction on the number of composita generated by the distinct roots of $f$). (See, e.g., [@bs pp. 259–260].) \[thm:disc\] For any number field $K$ of degree $n$ over ${\mathbb{Q}}$, we have $ d_K\!\geq\!\frac{n^{2n}}{(n!)^2}\!\geq\! \frac{(\pi e^2/4)^n}{2\pi n}\!>\!\frac{5.8^n}{6.3n}$. Also, if $K$ has minimal polynomial $f\!\in\!{\mathbb{Q}}[x]$ and $L$ is the splitting field of $f$, then $d_L$ divides $d_K^{(n-1)!+(n-2)! n +\cdots+1! n^{n-2}+0!n^{n-1}}$. [$\blacksquare$]{} [**Proof of Theorem \[thm:grh\]**]{}: Clearly, $\delta(h_{n,p})\!=\!1$. Also, since $\gcd(n,n-1)\!=\!1$, it is clear that $C(h_{n,p})\!=\!1$. Now, for any $n \geq 2$, the trinomial $h_n:=x^n -x -1$ is irreducible over ${\mathbb{Q}}$ [@selmer]. Let $\alpha \in {\mathbb{C}}$ be any root of $h_n$ and let $K = {\mathbb{Q}}(\alpha)$, so that $[K:{\mathbb{Q}}]=n$. Then $d_K$ divides the resultant of $h_n$ and $h'_n$ [@bs Thm. 8.7.1, pg. 228]. The resultant of $h_n$ and $h'_n$ can then be computed explicitly to be $(-1)^{\frac{(n+2)(n-1)}{2}} \left(n^n+(-1)^n (n-1)^{n-1}\right)$ [@swan]. Hence $d_K$ divides $n^n+(-1)^n (n-1)^{n-1}$. (We note that an elegant modern development of trinomial discriminants can be found in Chapter 12 of [@gkz94]; see Formula 1.38 on Page 406 in particular). Let $L$ be the the splitting field of $h_n$. Then $L$ has degree at most $n!$ and, by Theorem \[thm:disc\], $d_K\!>\!\frac{5.8^n}{6.3n}$ and $d_L$ divides $\left(n^n+ (-1)^n (n-1)^{n-1}\right)^{(n-1)!+(n-2)! n + \cdots+1! n^{n-2}+0! n^{n-1}}$. Note that, for $n \geq 3$, we have $n^n+(-1)^n (n-1)^{n-1}\! \leq n^n+ (n-1)^{n-1}\! \leq\!e^{n\log n + \frac{4}{27}}$. Also, by Stirling’s Estimate [@rudin Pg. 200], $n!\!<\!e\sqrt{n}\left(\frac{n}{e}\right)^n$ (for all $n\!\geq\!1$), so we have $$\begin{aligned} & & (n-1)! +n(n-2)!+\cdots+2!n^{n-3}+1!n^{n-2}+0!n^{n-1}\\ & &\ \ < \ e\sqrt{n-1}\left(\frac{n-1}{e}\right)^{n-1}+ e\sqrt{n-2}\left(\frac{n-2}{e}\right)^{n-2}n+ \cdots+e\sqrt{1}\left(\frac{1}{e}\right)^1 n^{n-2}+ n^{n-1}\\ & & \ \ < \ e\sqrt{n}\left(1+\frac{1}{e}+\cdots+\frac{1}{e^{n-1}}\right)n^{n-1} \!=\!\left(\frac{e}{1-1/e}\right)n^{n-1/2}\!<\!4.31n^{n-1/2},\end{aligned}$$ and thus $d_L\!<\!e^{4.5n^{n+1/2}\log n}$. Theorem \[thm:least\] then tells us that there is a prime $p \in {\mathbb{Z}}$ so that $h_n$ splits completely modulo $p$ with no repeated roots where 1. $p=e^{O(n^{n+1/2}\log n)}= e^{e^{(n+1/2+o(1))\log n}}$ unconditionally, and 2. $p=O((n^{n+1/2}\log n)^2)=e^{(2n+1+o(1))\log n}$ if GRH is true. So then we must have 1. $(n+1/2+o(1))\log n\!\geq\!\log\log p$ unconditionally, and 2. $(2n+1+o(1))\log n\!\geq\!\log p$ if GRH is true. Considering the growth order of the inverse function of $x\log x$, we obtain our stated asymptotic lower bounds in Theorem \[thm:grh\]. [$\blacksquare$]{} We used the family of trinomials $x^n-x-1$ mainly for the sake of simplifying our proof. Many other families would likely exhibit the same behavior, albeit with some additional intricacies in applying prime ideal bounds. However, the deeper question is to understand the structure of [*truly*]{} extremal trinomials over prime fields, such as those appearing in the Appendix below. Acknowledgements {#acknowledgements .unnumbered} ================ We thank the American Institute of Mathematics for their splendid hospitality and support of our AIM SQuaRE project, which formed the genesis of this paper. Special thanks go to Zander Kelley (and the Texas A&M Supercomputer facility) for computing $p_{15}$, and pointing out the paper [@karpshpar]. Thanks also to Kiran Kedlaya for sharing his cython code (which helped push our computational experiments further) and for pointing out an error in an earlier computation of $p_{12}$. We also thank E. Mehmet Kiral, Timo de Wolff, and Matthew P. Young for useful discussions, and the referees for their insightful comments. Appendix: Some Extremal Trinomials {#appendix-some-extremal-trinomials .unnumbered} ================================== We list in Figure \[Fig2\], for $n\!\in\!\{1,\ldots,16\}$, trinomials $f_n$ with $\delta(f_n)\!=\!1$ and $f_n$ having exactly $n$ distinct roots in ${\mathbb{F}}_{p_n}$, with $p_n$ the smallest prime admitting such a trinomial. $n$ $f_n$ $p_n$ ------ -------------------------- --------- $1$ $1+x-2x^2$ $3$ $2$ $1+x-2x^2$ $5$ $3$ $1-3x+2x^3$ $11$ $4$ $-2+x+x^4$ $23$ $5$ $1+4x-5x^8$ $47$ $6$ $1+24x-25x^{33}$ $151$ $7$ $-2+x+x^{34}$ $173$ $8$ $1+23x-24x^{21}$ $349$ $9$ $-71+70x+x^{184}$ $619$ $10$ $1+5x-6x^{152}$ $1201$ $11$ $-797+796x+x^{67}$ $2753$ $12$ $-82+81x+x^{1318}$ $4801$ $13$ $-1226+1225x+x^{225}$ $10867$ $14$ $-39+38x+x^{2264}$ $16633$ $15$ $29574-29573x-x^{27103}$ $71237$ $16$ $-364+363x+x^{2729}$ $8581$ In particular, for each $n\!\in\!\{1,\ldots,16\}$, a full search was done so that the trinomial $f_n$ below has the least degree among all trinomials over ${\mathbb{F}}_{p_n}$ having exactly $n$ roots in ${\mathbb{F}}_{p_n}$. (It happens to be the case that, for $n\!\in\!\{1,\ldots,16\}$, we can also pick the middle degree monomial of $f_n$ to be $x$.) By rescaling the variable as necessary, we have forced $1$ to be among the roots of each of the trinomials below. It is easily checked via the last part of Assertion (3) of Theorem \[thm:z\] that $C(f_n)\!=\!1$ for each $n\!\in\!\{1,\ldots,16\}$. The least prime $p_{17}$ for which there is a trinomial $f_{17}$ with $\delta(f_{17})\!=\!1$ and exactly $17$ roots in ${\mathbb{F}}_{p_{17}}$ is currently unknown (as of July 2016). Better and faster code should hopefully change this situation soon. [A]{} Eric Bach and Jeff Shallit, [*Algorithmic Number Theory, Vol. I: Efficient Algorithms,*]{} MIT Press, Cambridge, MA, 1996. Etiene Bézout, [*General Theory of Algebraic Equations,*]{} translated from the original French by Eric Feron, Princeton University Press, 2006. Jingguo Bi; Qi Cheng; and J. Maurice Rojas, [*“Sub-Linear Root Detection, and New Hardness Results, for Sparse Polynomials Over Finite Fields,”*]{} proceedings of ISSAC (International Symposium on Symbolic and Algebraic Computation, June 26–29, Boston, MA), pp. 61–68, ACM Press, 2013. Jingguo Bi; Qi Cheng; and J. Maurice Rojas, [*“Sub-Linear Root Detection, and New Hardness Results, for Sparse Polynomials Over Finite Fields,”*]{} SIAM J. Comput., accepted. Enrico Bombieri; Jean Bourgain; and Sergei Konyagin, [*“Roots of polynomials in subgroups of ${\mathbb{F}}^*_p$ and applications to congruences,”*]{} Int. Math. Res. Not. IMRN 2009, no. 5, pp. 802–834. Jean Bourgain, [*“Estimates on Exponential Sums Related to the Diffie-Hellman Distributions,”*]{} Geom. funct. anal., Vol. 15 (2005), pp. 1–34. Ran Canetti; John B. Friedlander; Sergei Konyagin; Michael Larsen; Daniel Lieman; and Igor E. Shparlinski, [*“On the statistical properties of Diffie-Hellman distributions,”*]{} Israel J. Math. 120 (2000), pp. 23–46. Qi Cheng, [*“Derandomization of Sparse Cyclotomic Integer Zero Testing,”*]{} proceedings of FOCS 2007, pp. 74–80, IEEE press, 2007. Steve D. Cohen, [*“The distribution of polynomials over finite fields,”*]{} Acta. Arith., [**17**]{} (1970), pp. 255–271. Steve D. Cohen, [*“Uniform distribution of polynomials over finite fields,”*]{} J. London Math. Soc. [**6**]{} (1972), pp. 93–102. Israel M. Gel’fand; Mikhail M. Kapranov; and Andrei V. Zelevinsky, [*Discriminants, Resultants and Multidimensional Determinants,*]{} Birkhäuser, Boston, 1994. Habiba Kadiri and Nathan Ng, [*“Bound for the least prime ideal in the Chebotarev density theorem,”*]{} in preparation, University of Lethbridge, Alberta, Canada, 2014. Marek Karpinski and Igor Shparlinski, [*“On some approximation problems concerning sparse polynomials over finite fields,”*]{} Theoretical Computer Science 157 (1996), pp. 259-266. Alexander Kelley, [*“Roots of Sparse Polynomials over a Finite Field,”*]{} in Proceedings of Twelfth Algorithmic Number Theory Symposium (ANTS-XII, University of Kaiserslautern, August 29 – September 2, 2016), to appear. Also available as Math ArXiV preprint [1602.00208]{} . Alexander Kelley and Sean Owen, [*“Estimating the Number Of Roots of Trinomials over Finite Fields,”*]{} submitted to special issue of Journal of Symbolic Computation dedicated to MEGA 2014. Jeff Lagarias and Andrew Odlyzko, [*“Effective Versions of the Chebotarev Density Theorem,”*]{} Algebraic Number Fields: $L$-functions and Galois Properties (Proc. Sympos. Univ. Durham, Durham, 1975), 409–464, Academic Press, London, 1977. Jeff Lagarias; Hugh Montgomery; and Andrew Odlyzko, [*“A bound for the least prime ideal in the Chebotarev density theorem,”*]{} Inventiones Math., 54 (1979), pp. 271–296. Hermann Minkowski, [*“Théorèmes arithmétiques,”*]{} C. R. Acad. Sci. Paris [**112**]{} (1891), pp. 209–212. Paul Pollack, [*The smallest prime that splits completely in an abelian number field*]{}, Proc.  American Mathematical Society, [**142**]{} (2014), no. 6, pp. 1925–1934. Walter Rudin, [*Principles of Mathematical Analysis,*]{} 3${^{\text{\underline{rd}}}}$ edition, McGraw-Hill, 1976. Ernst S. Selmer, [*“On the Irreducibility of Certain Trinomials,”*]{} Math. Scan. [**4**]{} (1956), pp. 287–302. David Eugene Smith and Marcia L. Latham, [*The Geometry of René Descartes,*]{} translated from the French and Latin (with a facsimile of Descartes’ 1637 French edition), Dover Publications Inc., New York (1954). Richard G. Swan, [*“Factorization of Polynomials over Finite Fields,”*]{} Pacific Journal of Mathematics, Vol. 12, No. 3, March 1962. Thorsten Theobald and Timo de Wolff, [*“Norms of Roots of Trinomials,”*]{} Math ArXiV preprint [1411.6552]{} . Hiraku Tôyama, [*“A note on the different of the composed field,”*]{} Kodai Math. Sem. Report [**7**]{} (1955), pp. 43–44. André Weil, [*“Numbers of solutions of equations in finite fields,”*]{} Bull. Amer. Math. Soc. 55, (1949), pp. 497-–508.
{ "pile_set_name": "ArXiv" }
--- author: - 'S. M. Albrecht' - 'A. P. Higginbotham' - 'M. Madsen' - 'F. Kuemmeth' - 'T. S. Jespersen' - 'J. Nyg[å]{}rd' - 'P. Krogstrup' - 'C. M. Marcus' title: Exponential Protection of Zero Modes in Majorana Islands --- [^1] [^2] **Majorana zero modes are quasiparticle excitations in condensed matter systems that have been proposed as building blocks of fault-tolerant quantum computers [@Kitaev:2003jw]. They are expected to exhibit non-Abelian particle statistics, in contrast to the usual statistics of fermions and bosons, enabling quantum operations to be performed by braiding isolated modes around one another [@Kitaev:2003jw; @Nayak:2008dp]. Quantum braiding operations are topologically protected insofar as these modes are pinned near zero energy, and the pinning is predicted to be exponential as the modes become spatially separated [@Read:392241; @Kitaev:2001kla]. Following theoretical proposals [@Lutchyn:2010hpa; @Oreg:2010gk], several experiments have identified signatures of Majorana modes in proximitized nanowires [@Mourik:2012je; @Rokhinson:2012ep; @Das:2012hi; @Deng:2012gn; @Churchill:2013cq] and atomic chains [@NadjPerge:2014ey], with small mode-splitting potentially explained by hybridization of Majoranas [@DasSarma:2012kt; @Stanescu:2013je; @Rainis:2013zxa]. Here, we use Coulomb-blockade spectroscopy in an InAs nanowire segment with epitaxial aluminum, which forms a proximity-induced superconducting Coulomb island (a ÔMajorana islandÕ) that is isolated from normal-metal leads by tunnel barriers, to measure the splitting of near-zero-energy Majorana modes. We observe exponential suppression of energy splitting with increasing wire length. For short devices of a few hundred nanometers, sub-gap state energies oscillate as the magnetic field is varied, as is expected for hybridized Majorana modes. Splitting decreases by a factor of about ten for each half a micrometer of increased wire length. For devices longer than about one micrometer, transport in strong magnetic fields occurs through a zero-energy state that is energetically isolated from a continuum, yielding uniformly spaced Coulomb-blockade conductance peaks, consistent with teleportation via Majorana modes[@Fu:2010ho; @Hutzen:2012gg]. Our results help to explain the trivial-to-topological transition in finite systems and to quantify the scaling of topological protection with end-mode separation.** \[fig1\] ![image](fig1.pdf){width="189mm"} The set of structures we investigate consist of InAs nanowires grown by molecular beam epitaxy in the $[0001]$ wurtzite direction with an epitaxial Al shell on two facets of the hexagonal cross section [@2015NatMa..14..400K]. The Al shell was removed except in a small segment of length $L$ and isolated from normal metal (Ti/Au) leads by electrostatic gate-controlled barriers (Fig. 1a). Charging energy, $E_ \mathrm{C} $, of the device ranges from greater than to less than the superconducting gap of Al ($\sim$ 0.2 meV). The thin Al shell ($8-10~\mathrm{nm}$ thickness on the two facets) gives a large critical field, $B_\mathrm{c}$, before superconductivity is destroyed: for fields along the wire axis, $B_\mathrm{c,||} \sim 1~ \mathrm{T} $; out of the plane of the substrate but roughly in the plane of the two Al-covered facets, $B_\mathrm{c,\perp} \sim 700~ \mathrm{mT} $ (Fig. 1b). The very high achieved critical fields make these wires a suitable platform for investigating topological superconductivity [@2015NatMa..14..400K]. Five devices over a range of Al shell lengths $L \sim 0.3-1.5\,\mu$m were measured (see Methods for device layouts). Charge occupation and tunnel coupling to the leads were tuned via electrostatic gates. Differential conductance, $g$, in the Coulomb blockade regime (high-resistance barriers) was measured using standard ac lock-in techniques in a dilution refrigerator (electron temperature $\sim 50~ \mathrm{mK} $). Figure 1c shows $g$ as a function of gate voltage, $V_\mathrm{G}$, and source-drain bias, $V_\mathrm{SD}$. For the $L = 790~\mathrm{nm}$ device, the zero-field data (top panel) show a series of evenly spaced Coulomb diamonds with a characteristic negative-differential conductance (NDC) region at higher bias. NDC is known from metallic superconductor islands [@Hekking:1993gc; @Hergenrother:1994ei] and has recently been reported in a proximitized semiconductor device similar to those investigated here [@Higginbotham:2015wc]. The zero-magnetic-field diamonds reflect transport via Cooper pairs, with gate voltage period proportional to $2e$, the charge of a Cooper pair. At moderate magnetic fields (Fig. 1c, middle panel), the large diamonds shrink and a second set of diamonds appears, yielding even-odd spacing of Coulomb blockade zero-bias conductance peaks [@Eiles:1993uh], as seen in the cuts in Fig. 1d. At larger magnetic fields (Fig. 1c, lower panel) Coulomb diamonds are again periodic, now with precisely half the spacing of the zero-field diamonds, corresponding to $1e$ periodicity. NDC is absent, and resonant structure is visible within each diamond, indicating transport through discrete resonances at low bias and a continuum at high bias (see magnification in Fig. 1c). Coulomb blockade conductance peaks at high magnetic field (see Fig. 1d for zero bias cuts) with regular $1e$ periodicity (half the zero-field spacing) accompanied by a discrete subgap spectrum is a proposed signature of electron teleportation by Majorana end states [@Fu:2010ho; @Hutzen:2012gg]. We designate as a ‘Majorana island’ (MI) the ungrounded tunneling device in this high-field regime, where a subgap state near zero energy, energetically isolated from a continuum, leads to 1$e$-periodic Coulomb blockade conductance peaks. Zero-bias conductance can be qualitatively understood within a simple zero-temperature model where the energy of the superconducting island—with or without subgap states (Fig. 1d)—is given by a series of shifted parabolas, $E_N( N_\mathrm{G} ) = E_ \mathrm{C}(N_ \mathrm{G} - N )^2 + p_N E_0$, where $N_ \mathrm{G} = CV_ \mathrm{G}/e$ is the gate-induced charge (electron charge $e$ and gate capacitance $C$) [@Tuominen:1992vy; @Eiles:1993uh; @Hekking:1993gc; @Lafarge:1993zz; @Hergenrother:1994ei; @1994MPLB....8.1007M]. $E_0$ is the energy of the lowest quasiparticle state, which is filled for odd parity ($p_{N}=1$, odd $N$), and empty for even parity ($p_N = 0$, even $N$) [@Higginbotham:2015wc]. Transport occurs when the ground state has a charge degeneracy, i.e., when the $E_N$ parabolas intersect. For $E_0 > E_\mathrm{C}$, the ground state always has even parity; transport in this regime occurs via tunneling of Cooper pairs at degeneracies of the even-$N$ parabolas. This is the regime of the $2e$-periodic Coulomb blockade peaks seen at low magnetic fields (Fig. 1d, blue). The odd charge state is spinful and can be lowered by Zeeman energy when a magnetic field is applied. For sufficiently large field, such that $E_0 < E_\mathrm{C}$, an odd-$N$ ground state emerges. This transition from 2$e$ charging to 1$e$ charging is seen experimentally as the splitting of the $2e$-periodic Coulomb diamonds into the even-odd double-diamond pattern in Fig. 1d (green trace). In this regime the Coulomb peak spacing is proportional to $E_\mathrm{C} + 2 E_0$ for even diamonds and $E_\mathrm{C} - 2 E_0$ for odd diamonds [@Tuominen:1992vy; @Lafarge:1993zz]. For the particular case of a zero-energy Majorana state, $E_0 = 0$, peak spacing is regular and $1e$-periodic. This regime is observed at higher fields (Fig. 1d, red), though not sufficiently high to destroy superconductivity. Coulomb peak spacings are measured as a function of magnetic field, allowing the state energy, $E_0( B )$, to be extracted. An example, showing 10 consecutive peaks for the $L = 0.9~\mathrm{\mu m}$ device, is shown in Fig. 2a. The peaks are $2e$-periodic at $B=0$, start splitting around $\sim 95~ \mathrm{mT}$, and become $1e$-periodic at $\sim110~\mathrm{mT}$, well below the spectroscopically observed closing of the superconducting gap at $B_ \mathrm{c} \sim 600~\mathrm{mT}$ (see Methods). This points towards the presence of a state close to zero energy within the superconducting regime over a range of $\sim 500~\mathrm{mT}$. Separately averaging even and odd Coulomb peak spacings, $\langle S_ \mathrm{e,o} \rangle$, over an ensemble of adjacent peaks reveals oscillations around the $1e$-periodic value as a function of applied magnetic field. This is consistent with an oscillating state energy $E_0$ due to hybridized Majorana modes [@Rainis:2013zxa; @DasSarma:2012kt; @Stanescu:2013je]. For the $L = 0.9~\mathrm{\mu m}$ device (Fig. 2b), peak spacing oscillations yield an energy oscillation amplitude $A = 7.0 \pm 1.5 ~\mu \mathrm{eV}$, converted from gate voltage to energy using the gate lever arm, $\eta$, extracted independently from the slope of the Coulomb diamonds. For the $L = 1.5~\mu \mathrm{m} $ device (Fig. 2c) average Coulomb peak spacing oscillations based on 22 consecutive peaks yield a barely resolvable amplitude, $A = 1.2 \pm 0.5 ~\mu \mathrm{eV} $. Oscillation amplitudes for the five measured devices (see Methods for device details), are shown in Fig. 2d along with a two-parameter fit to an exponential function, $ A = A_0 e^{ -L / \xi}$, giving $A_0 = 300~ \mathrm{\mu eV} $ and $\xi = 260~ \mathrm{nm} $ as fit parameters. The data fits well to the predicted exponential form that characterizes the topological protection of Majorana modes [@Read:392241; @Kitaev:2001kla; @DasSarma:2012kt]. \[fig2\] ![](fig2.pdf "fig:"){width="89mm"} \[fig3\] ![](fig3.pdf "fig:"){width="89mm"} Excited states of the MI are probed using finite-bias transport spectroscopy. This technique requires a fixed gate voltage, chosen such that at zero bias the electrochemical potential of the leads aligns with the middle of the spectroscopic gap of the MI. With this choice, the conductance at source-drain bias $V_\mathrm{SD}$ is due to states at energy $e V_\mathrm{SD} / 2$. A conductance peak at zero bias corresponds to a zero-energy state. In the case shown in Figs. 3(a,b), the gate voltage is tuned using the characteristic finite-bias conductance spectra for a short InAs/Al island, investigated previously in Ref. [@Higginbotham:2015wc]. Ground-state energies determined by finite-bias spectroscopy match those extracted from zero-bias peak spacings (see Methods Fig. S7). Bias spectroscopy shows discrete zero-energy states emerging at sufficient applied field over a range of device lengths. In a short device (Fig. 3c), the discrete state moves linearly in magnetic field, passing through zero and merging with a continuum at $V_\mathrm{SD} \sim 100~\mathrm{\mu eV}$. This merging is expected for Majorana systems in the short-length limit, where quenching of spin-orbit coupling results in unprotected parity crossings and state intersections at high energy [@Stanescu:2013je]. Rather than passing directly through zero, the first zero crossing extends for $40~\mathrm{mT}$, which is not understood. Medium-length devices show the subgap state bending back toward zero after zero crossings (Fig. 3d), in agreement with theoretical predictions for the emergence of Majorana behavior with increasing system length [@Stanescu:2013je; @Rainis:2013zxa]. For a long device ($L = 1.5\, \mu$m), bias spectroscopy shows a zero-energy state separated from a continuum at higher bias (Fig. 3e). The zero-energy state is present over a field range of $120~\mathrm{mT}$, with an associated energy gap $(30~\mathrm{\mu eV} ) / k_\mathrm{B} = 0.35 ~\mathrm{K}$. The evolution from unprotected parity crossings, to energetically isolated oscillating states, and then to a fixed zero-energy state, with increasing device length is consistent with the expected crossover from a strongly overlapping precursor of split Majoranas to a topologically protected Majorana state locked at zero energy [@Stanescu:2013je; @Rainis:2013zxa]. Note that in the data in Fig. 3e, the signal from the discrete state disappears above $B_{||} = 320~\mathrm{mT}$. This is not expected within a simple Majorana picture. Even though the zero-bias peak disappears, the peak spacing remains $1e$-periodic (see Methods). The observed effective g-factors, $g \sim 20-50 $, extracted from the addition spectrum and bias spectroscopy (see Methods), are large compared to previous studies on InAs nanowires [@Csonka:2008gz; @Schroer:2011ev; @Das:2012hi], perhaps resulting from field focusing from the Al shell. The measured gap to the continuum at zero magnetic field is consistent with the gap of aluminum $\Delta_ \mathrm{Al} \sim 180 ~\mu \mathrm{eV} $, and is roughly the same in all devices. The discrete subgap states (Fig. 3c-e) have zero-field energy less than but comparable to the gap, ranging from $E_0 \left( B = 0 \right) \sim 50-160 ~\mu \mathrm{eV} $, consistent with expectations for half-shell geometries [@Cole:2015vf]. The measured gap between the near-zero-energy state and the continuum in the high-field (topological) regime, $\Delta_\mathrm{T}\sim 30~\mathrm{\mu e V}$, along with the coherence length extracted from the exponential fit to the length-dependent splitting (Fig. 2d), $\xi \sim 260~\mathrm{nm}$, are consistent with topological superconductivity. Within this picture, at low magnetic fields, the gap and coherence length are related to the strength of spin-orbit coupling, yielding a value $\alpha_\mathrm{SO} \sim \xi \cdot \Delta_\mathrm{T} = 8\times 10^{-2}~\mathrm{eV \cdot \AA}$ that is consistent with previously reported values in InAs nanowires [@Fasth:2007ie; @Das:2012hi]. For a single subband picture, this implies a Fermi velocity that is lower than expected, suggesting that more than one subband is occupied under the Al shell, though we are not able to extract the number of modes directly. \[fig4\] ![](fig4.pdf "fig:"){width="80mm"} Finally, we consider the magnetic field dependence of Coulomb blockade peak heights (as opposed to spacings), as seen in Fig. 4. We found in most devices that below the field $B^{*}$ where $2e$-periodic peaks split, all peaks were uniformly high amplitude. Above $B^{*}$, peak heights were rapidly suppressed and remained low up to a second characteristic field, $B^{**}$, coincident with $1e$ periodicity (i.e., the field where even-odd spacing differences vanished). Above $B^{**}$, peak heights recovered. In the longer wires, peaks were nearly absent between $B^{*}$ and $B^{**}$, as seen in Fig. 4c. We interpret these observations as follows: In the present lead-wire-lead geometry, transport above $B^{*}$ involves single electrons entering one end of the wire and leaving from the other. The onset of uniform spacing with the reappearance of high peaks above $B^{**}$ indicates the emergence of a state (or states) at zero energy with strong wave function support at both ends of the wire. This is consistent with teleportation of electrons from one end of the wire to the other via a Majorana mode [@Fu:2010ho; @Hutzen:2012gg], though not necessarily a unique signature [@2015PhRvB..92b0511S]. Thus while the simultaneous brightening of peaks with their becoming uniformly spaced at $B^{**}$ suggests a subgap/Majorana mode moving to the ends of the wire as it moves to zero energy, we cannot rule out other forms of end-localized zero-energy states that could appear above a critical field. In summary, we have studied Majorana islands composed of InAs nanowires covered on two facets with epitaxial Al, for a range of device lengths. Zero-energy states are observed for wires of all lengths away from zero field. Oscillating energy splittings, measured using Coulomb blockade spectroscopy, are exponentially suppressed with wire length, with a characteristic length $\xi = 260~\mathrm{nm}$. This constitutes an explicit demonstration of exponential protection of zero-energy modes. Finite-bias measurements show transport through a discrete zero-energy state, with a measured topological gap $\Delta_\mathrm{T} = 30~\mathrm{\mu eV}$ for long devices. The extracted $\Delta_\mathrm{T}$ and $\xi$ are consistent with known parameters for InAs nanowires and the emergence of topological superconductivity. Brightening of Coulomb peaks at the field where spacing becomes uniform for longer devices suggests the presence of a robust delocalized state connecting the leads, and provides experimental support for electron teleportation via Majorana modes. **Acknowledgements** We thank K. Flensberg, M. Leijnse, M. Deng, W. Chang, and R. Lutchyn for valuable discussions, and G. Ungaretti, S. Upadhyay, C. S[ø]{}rensen, M. von Soosten, and D. Sherman for contributions to growth and fabrication. Research supported by Microsoft Project Q, the Danish National Research Foundation, the Lundbeck Foundation, the Carlsberg Foundation, and the European Commission. CMM acknowledges support from the Villum Foundation. **Author contributions** P.K., T.S.J. and J.N. developed the nanowire materials. S.M.A. fabricated the devices. S.M.A., A.P.H. and M.M. carried out the measurements with input from F.K., T.S.J. and C.M.M. Data analysis was done by S.M.A., A.P.H. and M.M. All authors contributed to interpreting the data. The manuscript was written by S.M.A., A.P.H. and C.M.M. with suggestions from all other authors. [10]{} url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} . ** ****, (). , , , & . ** ****, (). & . ** ****, (). . ** ****, (). , & . ** ****, (). , & . ** ****, (). *et al.* . ** ****, (). , & . ** ****, (). *et al.* . ** ****, (). *et al.* . ** ****, (). *et al.* . ** ****, (). *et al.* . ** ****, (). , & . ** ****, (). , & . ** ****, (). , , & . ** ****, (). . ** ****, (). , , , & . ** ****, (). *et al.* . ** ****, (). , , & . ** ****, (). , & . ** ****, (). *et al.* . ** ****, (). , & . ** ****, (). , , & . ** ****, (). , , , & . ** ****, (). , & . ** ****, (). , , & . ** ****, (). , , & . ** ****, (). , & . ** ****, (). , , , & . ** ****, (). , & . ** ****, (). Methods ======= Sample preparation ------------------ The InAs nanowires with epitaxial Al shell were grown via a two-step process by molecular beam epitaxy. First, the InAs nanowires were grown using the vapor-liquid-solid method with Au as a catalyst at $420^\circ$C. Second, after cooling the system to $-30^\circ$C the Al was grown on two facets of the hexagonal cross section [@2015NatMa..14..400K]. Afterwards the nanowires were deposited on degenerately doped Si substrates with 100-500 nm thick thermal oxides using either wet or dry deposition techniques. Wet deposition involves sonicating a growth substrate of nanowires in methanol for a few seconds, then putting several drops of the nanowire-methanol solution onto the chip surface using a pipette. Dry deposition was done by bringing a small piece of cleanroom wipe in touch with the growth substrate, then afterwards swiping it onto the chip surface. We find that while wet deposition results in a more uniform dispersion of nanowires on the chip surface, dry deposition is faster and less wasteful with nanowires. Selective removal of the Al shell was done by patterning etch windows using electron beam lithography on both sides of the nanowire, plasma cleaning the surface of the nanowire using oxygen, then etching the Al using a Transene Al Etchant D with an etching time of 10 seconds at $50^\circ$C. Depending on the device, ohmic contacts to the InAs core were fabricated using either ion milling or sulfur passivation to remove surface oxides. Ion milling was done for times ranging from 85 s to 110 s using a Kaufman & Robinson KDC 40 4-CM DC Ion Source with an acceleration voltage of 120V and an ion beam current density of 0.5 mA/$\mathrm{cm^2}$ at the chip surface. Sulfur passivation was done using a 2.1% solution of $\mathrm{(NH_4)_2S}$ in DI water with 0.15 M dissolved elemental sulfur at $40^\circ$C for 20 minutes. This was followed by the deposition of $5~ \mathrm{nm}$ of Ti as a sticking layer and $ 70-100~ \mathrm{nm} $ of Au for the ohmic contact. We found that ion milling resulted in more stable devices. Side and plunger gates were lithographically defined in the same fabrication step as the ohmic contacts in order to increase device yield. PMMA was used as resist in all lithography steps. Device Geometries ----------------- Gate patterns of the five measured devices are shown in ED Fig. \[figgeo\]. With the exception of the $L = 0.9 ~ \mathrm{\mu m} $ device, all measurements involving gate dependence are tuned through resonances using the plunger gate on either the Al side or the uncoated InAs side. For the $ L = 0.9 ~ \mathrm{\mu m} $ device, the lower left side gate is used to tune through resonances of the quantum dot, because the central plunger gate was not bonded during the cool down. ![image](figm1.pdf){width="183mm"} Measurements ------------ Transport measurements were carried out in an Oxford Triton dilution refrigerator with a base electron temperature of $T \sim 50~\mathrm{mK} $ and a 6-1-1 T vector magnet. Differential conductance, $g = dI/dV_\mathrm{SD} $, was measured using the AC-lockin technique with an excitation voltage in the range 2-6 $\mathrm{\mu V} $. Peak spacing data summary ------------------------- ![image](figm2.pdf){width="160mm"} The exponential curve in Fig. 2d (main text) is derived from even-odd peak spacing measurements in the high critical field directions, $B_{||}$ and $B_\perp$, summarized in ED Fig. \[fig:eoSummary\]. Suppression of spacing fluctuations with increased device length is clearly visible. The measured $A$ is indicated by black arrows in the inset, and the values are recorded in ED Table \[tab:Ecsummary\] for each device length, along with charging energies and lever arms. For $L=330~\mathrm{nm}$, Coulomb peak fluctuations became uncorrelated after several peaks. To obtain a large statistical ensemble, fluctuations were averaged over five sets of Coulomb peaks taken in different device tunings. ED Fig. \[fig:eoSummary\]a shows data from a single set of peaks, and ED Table \[tab:Ecsummary\] reports the full ensemble average. In a transverse magnetic field applied in the low critical field direction, $B_\mathrm{tr}$, shown in ED Fig. \[fig:eoSummary\]f-i, the oscillations are absent, with the exception of an initial overshoot for $L = 0.9~\mathrm{\mu m}$ at $B_\mathrm{tr} = 55~\mathrm{mT}$ (ED Fig. \[fig:eoSummary\]i), before the system is driven into the normal state at $B_\mathrm{tr} \sim 65~\mathrm{mT}$. L \[$\mathrm{nm}$\]   $E_\mathrm{C}$ \[$\mathrm{meV}$\]   $\eta$ \[$\mathrm{eV/V}$\]   $A$ \[$\mu \mathrm{eV}$\]   ----------------------- ------------------------------------- ------------------------------ ----------------------------- 330 1.6 0.048 106 400 0.40 0.012 60 790 0.14 0.008 14 950 0.054 0.0016 7 1540 0.022 0.002 1.2 : \[tab:Ecsummary\] Device length, $L$, charging energy, $E_\mathrm{C}$, lever arm, $\eta$, and characteristic amplitude, $A$, for the five measured devices. ![\[figBC\]](figm3.pdf){width="89mm"} ![image](figm4.pdf){width="150mm"} ![image](figm5.pdf){width="183mm"} Magnetic Field Orientation -------------------------- The direction of the nanowire on the chip was found by orienting the magnetic field from a vector magnet in the chip plane and spectroscopically measuring the anisotropy of the critical magnetic field. By comparing to the wire-direction based on optical and electron micrographs, we estimate an angular precision of $\pm ~3$ degrees. Critical Field Measurements --------------------------- The observed $2e$-$1e$ splitting at $B_{||} \sim 95~ \mathrm{mT} $ is compared to the closing of the superconducting gap at a considerably higher critical field, $B_{c,||}$, in ED Fig. \[figBC\]. Bias spectroscopy in ED Fig. \[figBC\]b shows a closing of the superconducting gap at $B_{c,||} \sim 600~\mathrm{mT}$, more than 500 mT after the onset of evenly spaced $1e$-periodic Coulomb peaks. The change from $2e$ to $1e$-periodicity at $B_{||} \sim 100~ \mathrm{mT} $ in ED Fig. \[figBC\]a coincides with a reduction in the measured Coulomb gap in ED Fig. \[figBC\]b, reflecting the transition from Cooper pair charging (energy penalty $2E_ \mathrm{C} $) to single-electron charging (energy penalty $E_ \mathrm{C} $). The measurement in ED Fig. \[figBC\]b was taken in a Coulomb valley at the gate voltage $V_ \mathrm{G}= -14.92~ \mathrm{V}$. Averaging of peak spacings -------------------------- In Fig. 2b of the main text, we show the extracted average peak spacing for several even and odd Coulomb valleys. A high resolution measurement of the $2e$-$1e$ splitting is shown in ED Fig. \[figpeaks\]a. The individual even and odd valleys, $S_ \mathrm{e,o} $ in ED Fig. \[figpeaks\]b, exhibit the same oscillating behavior but show a small deviation from the average between $100-125~ \mathrm{mT} $, which might be attributable to $g$-factor fluctuations for successive charge occupations of the quantum dot. Below 100 mT the fluctuations are very small, giving an indication of instrumental noise in the measurement. Angle Dependence ---------------- Angle dependence of the anti-crossing of the state with the continuum for $L = 400~ \mathrm{nm} $ is shown in ED Fig. \[figangle\]. We focus on magnetic fields, $B_\alpha$, with angles, $\alpha$, in the plane perpendicular to the nanowire direction. The measurements show a pronounced anti-crossing between the sub-gap state and an excitation continuum ($\alpha = 112.5^\circ$ and $\alpha = 135^\circ$) that is significantly reduced for $\alpha = 67.5^\circ$. Interpreting angle dependence is complicated by the anisotropy of g-factor and critical field. The critical field is maximized for $\alpha=120^\circ$, and is reduced drastically for near-perpendicular field alignment ($\alpha = 22.5^\circ$). The observed g-factors are highly dependent on field orientation and device tuning. For the $L = 400~ \mathrm{nm} $ device shown in ED Fig. \[figangle\], we found an approximately sinusoidal variation in $g$-factor by a factor of 2, with maximum g-factor occurring near $\alpha = 90^\circ$. Choice of gate voltage for bias spectroscopy -------------------------------------------- ![image](figm6.pdf){width="183mm"} For bias spectroscopy, the gate voltage is fixed either by interpreting Coulomb diamonds, as discussed in the main text, or from even-odd peak spacings. While details of the bias spectroscopy, such as locations of zero-crossing, depend on the choice of gate voltage, general features such as slopes, typical fluctuation amplitude, and the presence of a robust excitation gap are not strongly affected by the choice of gate voltage (ED Fig. \[figSilvanos\]). Comparison of addition energies and finite bias spectroscopy ------------------------------------------------------------ ![[]{data-label="figMethodComp"}](figm7.pdf) ![[]{data-label="figCMode"}](figm8.pdf) Peak spacings are used to measure the energy of the lowest-lying state. The same information is present in the bias spectroscopy, and gives consistent results, as shown in ED Fig. \[figMethodComp\]. Bias spectroscopy of long device {#sec:longBias} -------------------------------- Common-mode fluctuations in Coulomb peak position were observed in the longest ($L=1.5~\mathrm{\mu m}$) device, as shown in ED Fig. \[figCMode\]a. The fluctuations evidently correspond to a shift in the electrochemical potential of the dot, likely due to a nearby, field-dependent charge trap. The fluctuations are small compared to charging energy, but complicate the application of bias spectroscopy which needs to be performed at fixed electrochemical potential. To correct for the fluctuations, we introduce an effective gate voltage, $$V_\mathrm{G,eff}( B ) = V_\mathrm{G} + \delta V( B ),$$ that removes the common-mode peak motion. The offset voltage is zero at low field, when Coulomb peaks are $2e$ periodic ($\delta V( B ) = 0$ for $B \leq 175~\mathrm{mT}$). At high field, $\delta V( B )$ is chosen so that the reference Coulomb peak (labeled in ED Fig. \[figCMode\]b) occurs at constant $V_\mathrm{G,eff}$. All nonzero $\delta V( B )$ are listed in ED Table \[tab:deltaV\]. B (mT)   $\delta V~(\mathrm{mV})$ ---------- -------------------------- 180 0.25 230 0.25 235 0.25 240 0.25 245 0.25 250 0.25 255 0.25 260 0.5 265 0.5 270 0.5 275 0.5 280 0.75 300 0.25 305 0.75 310 1.25 315 1.5 320 1.75 325 1.75 330 1.75 335 1.75 340 1.75 345 1.75 350 1.75 355 1.75 360 1.75 365 1.75 370 1.75 375 1.75 380 1.75 385 1.75 390 1.75 395 1.75 400 1.75 : \[tab:deltaV\] All nonzero offset voltage values, $\delta V( B )$, for $L=1.5~\mathrm{\mu m}$ device. Offset is defined for $B=0, 5, 10, ..., 400 ~ \mathrm{mT}$. As shown in ED Fig. \[figCMode\]b, this procedure removes the common-mode peak motion. In the case of the $1.5~\mathrm{\mu m}$ device, bias spectroscopy is performed at fixed $V_\mathrm{G,eff}$, which allows us to infer the energy of the sub-gap state at fixed electrochemical potential. Zero-energy state at successive Coulomb peaks --------------------------------------------- The zero-energy state is robust over many successive Coulomb peaks, as shown in ED Fig. \[figDiamonds\]. The full bias spectroscopy as a function of field is also reproducible over several peaks, as shown in ED Fig. \[fig4zbps\]. ![[]{data-label="figDiamonds"}](figm9.pdf){width="50.00000%"} ![[]{data-label="fig4zbps"}](figm10.pdf) ![[]{data-label="figgfac"}](figm11.pdf) Measured g-factors ------------------ As can be seen in ED Fig. \[figgfac\] the state energy does not move linearly in magnetic field. A non-linear behavior with magnetic field is expected in the presence of strong spin-orbit coupling and a finite critical field. If the behavior was strictly linear one would expect $$B^{**} = \frac{E_0}{E_0 - E_\mathrm{C}} B^*,$$ because the peak splitting at $B^*$ occurs when $E_0(B=0) - E_\mathrm{Z} = E_ \mathrm{C} $ and the state is at zero energy at $B^{**}$ when $E_\mathrm{Z} = E_0(B=0)$ (see Fig. 4 in the main text for reference). The non-linear behavior of $E_0(B)$ at higher magnetic fields approaching $B^{**}$ renders this unsuitable for an accurate measurement of the state energy at zero field. In the low field regime where the state energy is approximately linear with magnetic field we calculate an effective g-factor. Using this slope it is possible to give a rough estimate of the state energy $E_0(B=0)$ assuming linear behavior and extrapolating the state energy to zero magnetic field. For bias spectroscopy it should be noted that for gate voltages in the middle of the spectroscopic gap (see main text) transport through a state at $V_ \mathrm{SD} = V_0$ indicates a state energy $E_0 = e V_ \mathrm{0} /2$. An example for $L = 330~ \mathrm{nm} $ is shown in ED Fig. \[figgfac\]a. Using the addition spectrum, the state energy can be calculated from the peak spacing $S$ according to $E_0 = \left(\eta S - E_ \mathrm{C}\right)/ 2 $. Examples of extracted effective g-factors in the linear range are shown in ED Fig. \[figgfac\]b,c. [^1]: These authors contributed equally to this work. [^2]: These authors contributed equally to this work.
{ "pile_set_name": "ArXiv" }
--- abstract: | This paper proposes a model-free approach to analyze panel data with heterogeneous dynamic structures across observational units. We first compute the sample mean, autocovariances, and autocorrelations for each unit, and then estimate the parameters of interest based on their empirical distributions. We then investigate the asymptotic properties of our estimators using double asymptotics and propose split-panel jackknife bias correction and inference based on the cross-sectional bootstrap. We illustrate the usefulness of our procedures by studying the deviation dynamics of the law of one price. Monte Carlo simulations confirm that the proposed bias correction is effective and yields valid inference in small samples. *Keywords*: Panel data, heterogeneity, functional central limit theorem, jackknife, bootstrap. *JEL Classification*: C13, C14, C23. author: - 'Ryo Okui[^1]' - 'Takahide Yanagi[^2]' bibliography: - 'panhet.bib' date: 'January, 2019' title: 'Panel Data Analysis with Heterogeneous Dynamics[^3]' --- Introduction ============ Understanding the dynamics of a potentially heterogeneous variable is an important research consideration in economics. For instance, some researchers have investigated the price deviation of the law of one price (LOP) using panel data analyses, and a recent finding by @CruciniShintaniTsuruga15 indicates that time-series persistence and volatility measures for the LOP deviation are heterogeneous across US cities and goods. Other examples include income (e.g., [@BrowningEjrnesAlvarez10]) and productivity (e.g., [@HsiehKlenow09] and [@GandhiNavarroRivers16]) dynamics, for which there is a considerable body of research.[^4] We propose easy-to-implement procedures for analyzing the heterogeneous dynamic structure of panel data, $\{\{y_{it}\}_{t=1}^T\}_{i=1}^N$, without assuming any specific model.[^5] To this end, we investigate the cross-sectional distributional properties of the mean, autocovariances, and autocorrelations of $y_{it}$ with heterogeneous units. Our model-free approach is especially useful when empirical researchers are reluctant to assume specific models for heterogeneity given the threat of problems with misspecification. Despite the voluminous literature on dynamic panel data analyses, many studies assume specific models for the dynamics (such as autoregressive (AR) models) and homogeneity in the dynamics, allowing heterogeneity only in the mean of the process.[^6] While several other studies also consider either heterogeneous dynamics or model-free analyses, we are unaware of any specific study that proposes panel data analysis for heterogeneous dynamics without specifying a particular model. The distributional properties of the heterogeneous mean, autocovariances, and autocorrelations provide various pieces of information and are perhaps the most basic descriptive statistics for dynamics. Indeed, a typical first step in analyzing time-series data is to examine these properties. As we demonstrate in this study, the distributions of the heterogeneous mean, autocovariances, and autocorrelations can also be useful descriptive statistics for understanding heterogeneous dynamics in panel data. For example, it would be interesting to examine the degree of heterogeneity of the LOP deviations across items such as goods and services. In this case, the mean and variance of the heterogeneous means measure the amount and dispersion of the long-run LOP deviations across items. We can also examine the correlation of the heterogeneous means and autocorrelations that shows whether the magnitudes of the long-run LOP deviations relate to the speed of price adjustment toward the long-run LOP deviations. Moreover, our analysis provides the entire distribution of heterogeneity for the long-run LOP deviations or the adjustment speed. This entire distribution could be useful to investigate, for example, whether goods and services possess different dynamics. We derive the asymptotic properties of the empirical distributions of the estimated means, autocovariances, and autocorrelations based on double asymptotics, under which both the number of cross-sectional observations, $N$, and the length of the time series, $T$, tend to infinity. By using empirical process theory (e.g., [@vanderVaartWellner96]) and the inversion theorem for characteristic functions (e.g., [@gil1951note] and [@wendel1961non]), we show that the empirical distributions converge weakly to Gaussian processes under a condition for the relative magnitudes of $N$ and $T$ that is slightly “stronger” than $N^3 / T^4 \to 0$. The proof is challenging because our empirical distributions are biased estimators and depend on both $N$ and $T$, so that we cannot directly apply standard empirical process techniques. We show that the estimation error in the estimated mean, autocovariances, and autocorrelations for each unit biases the empirical distributions whose convergence rates depend on $T$. We also derive the asymptotic distributions of the estimators for other distributional properties (e.g., the quantile function) using the functional delta method. When we can write the parameter of interest as the expected value of a smooth function of the heterogeneous mean and/or autocovariances, we derive the exact order of the bias. This class of parameters includes the mean, variance, and other moments (such as correlations) of the heterogeneous mean, autocovariances, and/or autocorrelations. Importantly, we can analytically evaluate the bias, and find it has two sources. The first is the incidental parameter problem originally discussed in @NeymanScott48. The second arises when the smooth function is nonlinear. We show the asymptotic distribution of the estimator under the condition $N / T ^2 \to 0$ under which both biases are negligible. As $T$ is often small compared with $N$ in microeconometric applications, we propose to reduce the biases using @DhaeneJochmans15’s ([-@DhaeneJochmans15]) split-panel jackknife. The jackknife bias-corrected estimator is asymptotically unbiased under a weaker condition on the relative magnitudes of $N$ and $T$ and does not inflate the asymptotic variance. We propose to use the cross-sectional bootstrapping (e.g., [@GoncalvesKaffo14]) to test hypotheses and construct confidence intervals. The bootstrap distribution is asymptotically equivalent to the distribution of the estimator, but fails to capture the bias. We thus recommend the bootstrap based on the jackknife bias-corrected estimator because this would not suffer from large bias. As an empirical illustration, we examine the speed of price adjustment toward the long-run LOP deviation using a panel data set of various items for different US cities. We find statistically significant evidence that long-run LOP deviations in the item–city pairs with more persistent dynamics tend to be small and suffer from relatively large shocks. We also find formal statistical evidence that the distribution of the LOP adjustment speed for goods differs from that for services. We also conduct Monte Carlo simulations. They demonstrate noticeable performances of the bootstrap inference based on the jackknife bias-corrected estimation in small samples. #### Paper organization Section \[sec-literature\] reviews the studies related to this paper, Section \[sec-setting\] explains the setting, and Section \[sec-procedures\] introduces the procedures. In Section \[sec-asymptotics\], we derive the asymptotics of the distribution estimators, while Section \[sec-smooth\] considers the asymptotics for estimating the expected value of a smooth function. Section \[sec-extensions\] presents Kolmogorov–Smirnov (KS)-type tests based on the distribution estimators and Sections \[sec-application\] and \[sec-LOP-montecarlo\] develop the application and simulations. Section \[sec-conclusion\] concludes and Appendix \[sec-appendix\] contains the technical proofs. The supplementary appendix provides remarks on higher-order bias correction, a test for parametric specifications, other extensions, and several mathematical proofs, applications for income and productivity, and additional Monte Carlo simulations. Related studies {#sec-literature} =============== This paper most closely relates to the literature on heterogeneous panel AR models, which capture the heterogeneity in the dynamics by allowing for unit-specific AR coefficients. @PesaranSmith95, @HsiaoPesaranTahmiscioglu99, @PesaranShinSmith99, @PhillipsMoon99, and @Pesaran06 provide such analyses. The mean group estimator in @PesaranSmith95 is identical to our estimator (without bias correction) for the mean of the heterogeneous first-order autocorrelation if their AR(1) model does not contain exogenous covariates. @HsiaoPesaranTahmiscioglu99 show that the mean group estimator is asymptotically unbiased under $N/T^2\to 0$, which is the condition we obtain without the bias correction. Building on this literature, we aim to estimate the entire distributions of the unit-specific heterogeneous mean, autocovariances, and autocorrelations without model specifications. Researchers have developed econometric methods for investigating features of heterogeneity other than the mean. For example, @Hospido2015 investigated the variances of individual and job effects that additively affect the income process. As another example, @BotosaruSasaki2018 estimated the volatility function in a permanent–transitory model of income dynamics. However, these studies assume some models and have different motivations from ours. Elsewhere, @MavroeidisSasakiWelch15 identify and estimate the distribution of the AR coefficients in heterogeneous panel AR models. The advantage of their approach is that $T$ can be fixed. While we impose $T \to \infty$, our method is much simpler to implement. By contrast, the estimation method in @MavroeidisSasakiWelch15 requires the maximization of a kernel-weighting function written as an integration over multiple variables. The theoretical results for our distribution estimation relate to [@JochmansWeidner2018] who consider estimating the distribution of a true quantity based on a noisy measurement (e.g., an estimated quantity). They derive the formula for the bias of their empirical distribution estimator under the assumption that their observations exhibit Gaussian errors.[^7] In contrast, the present paper does not specify parametric distributions for our observations, at the cost of not showing the exact formula for the bias of the distribution estimator. Our results and theirs are thus complementary and thereby represent individual contributions. In a similar motivation to us, [@OkuiYanagi2018] develop nonparametric kernel-smoothing estimation based on the estimated means, autocovariances, and autocorrelations for cross-sectional units in panel data. There are several theoretical differences between the two papers, and they develop different proof techniques and obtain different results (see Remark \[remark:kernel\] for details). Moreover, an important practical issue also arises in the kernel estimation as the cross-sectional bootstrap is not suitable for the kernel estimation. This issue comes from the well-known result that the bootstrap cannot capture kernel-smoothing bias, so that they propose an alternative valid inference. Several studies propose model-free methods to investigate panel data dynamics. For example, using long panel data, @Okui10a [@Okui11; @Okui14] estimates autocovariances, and @LeeOkuiShintani13 consider infinite-order panel AR models. Because we can represent a stationary time series with an infinite-order AR process under mild conditions, their approach is essentially model-free. However, these studies assume homogeneous dynamics. While not directly connected, this study also relates to the recent literature on random coefficients or nonparametric panel data models with nonadditive unobserved heterogeneity.[^8] For example, @ArellanoBonhomme12 consider linear random coefficients models for panel data and discuss the identification and estimation of the distribution of random coefficients using deconvolution techniques. @Chamberlain92 and @GrahamPowell12 consider a model similar to that of @ArellanoBonhomme12, but focus on the means of random coefficients. @FernandezValLee13 examine moment restriction models with random coefficients using the generalized method of moments estimation. Their analysis of the smooth function of unit-specific effects closely relates to our analysis of the smooth function of means and autocovariances, at least in terms of technique. Finally, @Evdokimov09 and @Freyberger17 consider nonparametric panel regression models with unit-specific and interactive fixed effects, respectively, entering the unspecified structural function, but they do not infer heterogeneous dynamic structures. Settings {#sec-setting} ======== We observe panel data $\{\{y_{it}\}_{t=1}^{T}\}_{i=1}^{N}$ where $y_{it}$ is a scalar random variable, $i$ a cross-sectional unit, and $t$ a time period. We assume that $\{y_{it}\}_{t=1}^T$ is independent across units. We assume that the law of $\{y_{it}\}_{t=1}^T$ is stationary over time, but its dynamic structure may be heterogeneous. Specifically, we consider the following data-generating process (DGP) to model the heterogeneous dynamic structure, in a spirit similar to @GalvaoKato14. The unobserved unit-specific effect $\alpha_i$ is independently drawn from a distribution common to all units. We then draw the time series $\{y_{it}\}_{t=1}^T $ for unit $i$ from some distribution $\mathcal{L}(\{y_{it}\}_{t=1}^T; \alpha_i)$ that may depend on $\alpha_i$. The dynamic structure of $y_{it}$ can be heterogeneous because the realized value of $\alpha_i$ can vary across units. For example, in an application of the LOP deviations, $\alpha_i$ might represent unobservable permanent trade costs specific to item $i$. Note that $\alpha_i$ is an abstract parameter used to model heterogeneity in the dynamics across units and does not directly appear in the actual implementation of the procedure. For notational simplicity, we denote “$\cdot | \alpha_{i}$” by “$\cdot | i$”; that is, “conditional on $\alpha_{i}$” becomes “conditional on $i$” below. To infer the properties of heterogeneity in a model-free manner, we aim to develop statistical tools to analyze the cross-sectional distributions of the heterogeneous means, autocovariances, and autocorrelations of $y_{it}$. The mean for unit $i$ is $\mu_{i} \coloneqq E(y_{it}|i)$. Note that $\mu_i$ is a random variable whose realization differs across units. Because we assume stationarity, $\mu_{i}$ is constant over time. Let $\gamma_{k,i} \coloneqq E((y_{it} - \mu_i) (y_{i,t-k} - \mu_i)|i)$ and $\rho_{k,i} \coloneqq \gamma_{k, i} / \gamma_{0, i}$ be the $k$-th conditional autocovariance and autocorrelation of $y_{it}$ given $\alpha_{i}$, respectively. Note that $\gamma_{0,i}$ is the variance for unit $i$. To understand the possibly heterogeneous dynamics, we estimate the quantities that characterize the distributions of $\mu_i$, $\gamma_{k,i}$, and $\rho_{k,i}$. Below, we often use the notation $\xi_i$ to represent one of $\mu_i$, $\gamma_{k,i}$, and $\rho_{k,i}$. We consider cases in which both $N$ and $T$ are large. For example, in our empirical illustration for the LOP deviations, we use panel data where $(N,T) = (2248, 72)$. A large $N$ allows us to estimate consistently the cross-sectional distributions of $\mu_i$, $\gamma_{k,i}$, and $\rho_{k,i}$. We require a large $T$ to identify and estimate $\mu_i$, $\gamma_{k,i}$, and $\rho_{k,i}$ based on the time series for each unit. Our setting is very general and includes many situations. \[example:AR\] The panel AR(1) model with heterogeneous coefficients, as in @PesaranSmith95 and others, is a special case of our setting. This model is $y_{it} = c_i + \phi_i y_{i,t-1} + u_{it}$, where $c_i$ and $\phi_i$ are the unit-specific parameters, and $u_{it}$ follows a strong white noise process with variance $\sigma^2$. In this case, $\alpha_i = (c_i, \phi_i)$, $\mu_i = c_i / (1- \phi_i)$, $\gamma_{k,i} = \sigma^2 \phi_i^k/ (1- \phi_i^2)$, and $\rho_{k,i}=\phi_i^k$. Our setting also includes cases in which the true DGP follows some nonlinear process. Suppose that $y_{it}$ is generated by $y_{it} = m(\alpha_{i}, u_{it})$, where $m$ is some function and $u_{it}$ is stationary over time and independent across units. In this case, $\mu_i = E( m(\alpha_{i}, \epsilon_{it}) | \alpha_i )$ and $\gamma_{k, i}$ and $\rho_{k,i}$ are the $k$-th-order autocovariance and autocorrelation of $w_{it} = y_{it} - \mu_i$ given $\alpha_i$, respectively. We focus on estimating the heterogeneous mean, autocovariance, and autocorrelation structure and do not aim to recover the underlying structural form of the DGP. We understand that addressing several important economic questions requires knowledge of the structural function of the dynamics. Nonetheless, the distributions of the heterogeneous means, autocovariances, and autocorrelations can be estimated relatively easily without imposing strong assumptions and can provide valuable information, even if our ultimate goal is to identify the structural form. For example, if our procedure reveals a positive correlation between the time-series variance and the time-series persistence for the LOP deviations, the structural form on the LOP deviations should be specified so as to allow for such correlation. Procedures {#sec-procedures} ========== In this section, we present the statistical procedures to estimate the distributional characteristics of the heterogeneous mean, autocovariances, and autocorrelations. We first estimate the mean $\mu_i$, autocovariances $\gamma_{k, i}$, and autocorrelations $\rho_{k,i}$ using the sample analogs $\hat \mu_i \coloneqq \bar y_i \coloneqq T^{-1} \sum_{t=1}^{T}y_{it}$, $\hat \gamma_{k,i} \coloneqq (T-k)^{-1} \sum_{t=k+1}^T (y_{it} - \bar y_i) (y_{i,t-k} - \bar y_{i})$, and $\hat \rho_{k,i} \coloneqq \hat \gamma_{k,i} / \hat \gamma_{0,i}$. We write $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$ as the corresponding estimator of $\xi_i = \mu_i$, $\gamma_{k,i}$, or $\rho_{k,i}$, respectively. We then compute the empirical distribution of $\{ \hat \xi_i\}_{i=1}^N$: $$\begin{aligned} \label{eq-em-dis} \mathbb{F}_N^{\hat \xi} (a) \coloneqq \frac{1}{N} \sum_{i=1}^{N} \mathbf{1}(\hat \xi_i \leq a),\end{aligned}$$ where $\mathbf{1} (\cdot)$ is the indicator function and $a \in \mathbb{R}$. This empirical cumulative distribution function (CDF) is interesting in its own right because it is an estimator of the cross-sectional CDF of $\xi_i$, say $F_0^{\xi} (a) \coloneqq \Pr (\xi_i \le a)$. We can estimate other distributional quantities or test some hypotheses based on the empirical distribution of $\hat \xi_i$. For example, we can consider estimating the $\tau$-th quantile $q_{\tau}^{\xi} \coloneqq \inf \{ a \in \mathbb{R} : F_0^{\xi} (a) \ge \tau \}$ by the empirical quantile $\hat q_{\tau}^{\hat \xi} \coloneqq \inf \{ a \in \mathbb{R} : \mathbb{F}_N^{\hat \xi} (a) \ge \tau \}$. We can also test the difference in the heterogeneous dynamic structures across distinct groups and parametric specifications for the heterogeneous means, autocovariances, or autocorrelations based on the empirical distribution. We develop such tests based on KS-type statistics in Section \[sec-extensions\] and the supplementary appendix. We can also estimate a function of moments of the heterogeneous mean and autocovariances straightforwardly. Let $S \coloneqq h(E( g(\theta_i) ) )$ be the parameter of interest, where $\theta_i$ is an $l \times 1$ vector whose elements belong to a subset of $(\mu_i, \gamma_{0,i}, \gamma_{1,i},\dots)$, and $g: \mathbb{R}^l \to \mathbb{R}^m$ and $h: \mathbb{R}^m \to \mathbb{R}^n$ are vector-valued functions. We can estimate $S$ by the sample analog: $$\begin{aligned} \label{eq-H} \hat S \coloneqq h\left( \frac{1}{N} \sum_{i=1}^N g (\hat \theta_i )\right),\end{aligned}$$ where $\hat \theta_i$ is the estimator corresponding to $\theta_i$. For example, we can estimate the correlation between the mean and variance using the sample correlation of $\hat \mu_i$ and $\hat \gamma_{0, i}$. Note that we do not need to consider a function of autocorrelations separately given the autocorrelations are functions of the autocovariances. Section \[sec-smooth\] investigates the asymptotics for $\hat S$ when $g$ and $h$ are *smooth*, and we analytically evaluate the bias of order $O(1/T)$ in $\hat S$. @DhaeneJochmans15’s ([-@DhaeneJochmans15]) split-panel jackknife can reduce the bias in $\hat S$ if $g$ and $h$ are smooth. For example, the half-panel jackknife (HPJ) bias correction can delete the bias of order $O(1/T)$. Suppose that $T$ is even.[^9] We divide the panel data into two subpanels: $\{ \{ y_{it}\}_{t=1}^{T/2}\}_{i=1}^N$ and $\{ \{ y_{it}\}_{t=T/2 +1}^{T}\}_{i=1}^N$. Let $\hat S^{(1)}$ and $\hat S^{(2)}$ be the estimators of $S$ computed using $\{\{ y_{it} \}_{t=1}^{T/2}\}_{i=1}^{N}$ and $\{\{ y_{it} \}_{t=T/2+1}^{T}\}_{i=1}^{N}$, respectively. Let $\bar S \coloneqq (\hat S^{(1)} + \hat S^{(2)} )/2$. The HPJ bias-corrected estimator of $S$ is $$\begin{aligned} \label{eq-G-HPJ} \hat S^H \coloneqq \hat S - (\bar S - \hat S) = 2 \hat S - \bar S. \end{aligned}$$ The HPJ estimates the bias in $\hat S$ by $\bar S - \hat S$, and $\hat S^H$ does not show bias of order $O(1/T)$. We may also consider a higher-order jackknife bias correction to eliminate the bias of an order higher than $O(1/T)$, as discussed in @DhaeneJochmans15. In particular, we consider the third-order jackknife (TOJ) in the empirical application and simulations. However, we must modify the formula in @DhaeneJochmans15 for TOJ to correct higher-order biases in our setting. See the supplementary appendix for the details of this modification. Our Monte Carlo results below indicate that TOJ can be more successful than HPJ when higher-order biases are severe. However, in some cases, TOJ eliminates biases at the cost of deteriorating precision of estimation. Hence, we recommend adopting both HPJ and TOJ in practical situations. For statistical inference of parameter $S$, we suggest using the cross-sectional bootstrap to approximate the distribution of the bias-corrected estimator. Here, we present the algorithm for the HPJ estimator. The cross-sectional bootstrap regards each time series as the unit of observation and approximates the distribution of statistics under the empirical distribution of ($\hat \theta_i, \hat \theta_i^{(1)}, \hat \theta_i^{(2)}$), where $\hat \theta_i^{(1)}$ and $\hat \theta_i^{(2)}$ denote the estimates of $\theta_i$ from the first and second subpanels, respectively. The algorithm is: 1. Randomly draw $(\hat \theta_1^*, \hat \theta_1^{*(1)}, \hat \theta_1^{*(2)}), \dots, (\hat \theta_N^*, \hat \theta_N^{*(1)}, \hat \theta_N^{*(2)})$ from $\{ (\hat \theta_i, \hat \theta_i^{(1)}, \hat \theta_i^{(2)})\}_{i=1}^N$ with replacement.[^10] 2. Compute the statistics of interest, say $\vartheta$, using $(\hat \theta_1^*, \hat \theta_1^{*(1)}, \hat \theta_1^{*(2)}), \dots, ( \hat \theta_N^*, \hat \theta_N^{*(1)}, \hat \theta_N^{*(2)})$ 3. Repeat 1 and 2 $B$ times. Let $\vartheta^* (b)$ be the statistics of interest computed in the $b$-th bootstrap. 4. Compute the quantities of interest using the empirical distribution of $\{ \vartheta^* (b) \}_{b=1}^B$. For example, suppose that we are interested in constructing a 95% confidence interval for a scalar parameter $S$. We obtain the bootstrap approximation of the distribution of $\vartheta = \hat S^H - S$. Let $\hat S^{H*} (b)$ be the HPJ estimate of $S$ obtained with the $b$-th bootstrap sample. We then compute the 2.5% and 97.5% quantiles, denoted as $q^*_{0.025}$ and $q^*_{0.975}$, respectively, of the empirical distribution of $\{ \vartheta^* (b) \}_{b=1}^B$ with $\vartheta^* (b) = \hat S^{H*}(b) - \hat S^{H}$. The bootstrap 95% confidence interval for $S$ is $[ \hat S^H - q^*_{0.975}, \hat S^H - q^*_{0.025}]$. Asymptotic analysis for the distribution estimators {#sec-asymptotics} =================================================== This section presents the asymptotic properties of the distribution estimator in . We first show the uniform consistency of the empirical distribution and then derive the functional central limit theorem (functional CLT). We also show that the functional delta method can apply in this case. All analyses presented below are under double asymptotics ($N, T \to \infty$).[^11] While our setting is fully nonparametric as introduced in Section \[sec-setting\], the following representation is useful for our theoretical analysis. Let $w_{it} \coloneqq y_{it} - E(y_{it}|i) = y_{it} - \mu_i$. By construction, $y_{it} = \mu_{i} + w_{it}$ and $E(w_{it}|i)=0$ for any $i$ and $t$. Note also that $\gamma_{k,i} = E(w_{it} w_{i,t-k}|i)$. Assumptions ----------- Because we use empirical process techniques, it is convenient to rewrite the empirical distributions as empirical processes indexed by a class of indicator functions. Let $\mathbb{P}_{N}^{\hat \xi} \coloneqq N^{-1} \sum_{i=1}^{N} \delta_{\hat \xi_i}$ be the empirical measure of $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$, where $\delta_{\hat \xi_i}$ is the probability distribution degenerated at $\hat \xi_i$. Let $\mathcal{F} \coloneqq \{ \mathbf{1}_{(-\infty, a]} : a \in \mathbb{R} \}$ be the class of indicator functions where $\mathbf{1}_{(-\infty,a]} (x) \coloneqq \mathbf{1} ( x \le a )$. We denote the probability measure of $\xi_i$ as $P_0^{\xi}$. In this notation, the empirical distribution function $\mathbb{F}^{\hat \xi}_{N}$ in is an empirical process indexed by $\mathcal{F}$, and $\mathbb{P}_N^{\hat \xi} f = \mathbb{F}_N^{\hat \xi} (a)$ for $f = \mathbf{1}_{(-\infty, a]}$. Similarly, $P_0^{\xi} f = F_0^{\xi}(a) = \Pr(\xi_i \leq a)$ for $f = \mathbf{1}_{(-\infty, a]}$. We often use shorthand notations such as $\mathbb{P}_N = \mathbb{P}_N^{\hat \xi}$, $\mathbb{F}_N = \mathbb{F}^{\hat \xi}_N$, $P_0 = P_0^{\xi}$, and $F_0 = F_0^{\xi}$ by omitting the superscripts $\hat \xi$ and $\xi$. Throughout the study, we assume the following summarizes the conditions in Section \[sec-setting\]. \[as-basic\] The sample space of $\alpha_i$ is some Polish space and $y_{it} \in \mathbb{R}$ is a scalar real random variable. $\{(\{y_{it}\}_{t=1}^T, \alpha_i)\}_{i=1}^N$ is independently and identically distributed (i.i.d.) across $i$. We stress that the i.i.d. assumption does not restrict the heterogeneous dynamics. The following assumptions depend on natural numbers $r_m$ and $r_d$, which will be specified in the theorems that use this assumption. For a strictly stationary stochastic process $\{X_t\}_{t=1}^\infty$, define $\alpha$-mixing coefficients as $\alpha (m) = \sup_{A \in \mathcal{M}_1^k, B\in \mathcal{M}_{k+m}^{\infty}} |\Pr (A\cap B) - \Pr(A) \Pr(B) |$, where $\mathcal{M}_a^b$ denotes the $\sigma$-algebra generated by $X_j$ for $a \leq j \leq b$, and call the process $\alpha$-mixing if $\alpha (m) \to 0$ as $m\to \infty$. \[as-mixing-c\] For each $i$, $\{y_{it}\}_{t=1}^{\infty}$ is strictly stationary and $\alpha$-mixing given $\alpha_i$, with mixing coefficients $\{\alpha (m|i)\}_{m=0}^\infty$. There exists a natural number $r_m$ and a sequence $\{ \alpha (m) \}_{m=0}^\infty$ such that for any $i$ and $m$, $\alpha (m|i) \le \alpha (m)$ and $\sum_{m=0}^{\infty} (m+1)^{r_m/2-1} \alpha(m) ^{\delta / (r_m+\delta)} < \infty$ for some $\delta>0$. \[as-w-moment-c\] There exists a natural number $r_d$ such that $E|w_{it}|^{r_d+\delta} < \infty$ for some $\delta > 0$. Assumptions \[as-mixing-c\] and \[as-w-moment-c\] are mild regularity conditions on the process of $y_{it}$. Assumption \[as-mixing-c\] is a mixing condition depending on $r_m$ and restricts the degree of persistence of $y_{it}$ across time. It also imposes stationarity on $\{y_{it}\}_{t=1}^{\infty}$, which in particular implies that the initial values are generated from the stationary distribution. Note that a large $T$ could also guarantee that such an initial value condition (see, e.g., Section 4.3.2 in [@Hsiao2014]) is negligible in our analysis. Nonetheless, we impose this condition to simplify the analysis. Assumption \[as-w-moment-c\] requires that $w_{it}$ has some moment higher than the $r_d$-th order. These assumptions are satisfied, for example, when $y_{it}$ follows a heterogeneous stationary panel ARMA model with Gaussian innovations. We also introduce Assumptions \[as-mu-con\], \[as-gamma-con\], and \[as-rho-con\] for the uniform consistency and functional CLTs of the empirical distributions $\mathbb{F}_N^{\hat \mu}$, $\mathbb{F}_N^{\hat \gamma_k}$, and $\mathbb{F}_N^{\hat \rho_k}$, respectively. Condition a) in each assumption is introduced for the uniform consistency, and the remaining conditions are required for the functional CLTs. We define $\bar w_i \coloneqq T^{-1} \sum_{t=1}^T w_{it}$. \[as-mu-con\] a) The random variable $\mu_i$ is continuously distributed. b) The CDF of $\mu_i$ is thrice boundedly differentiable. c) The CDF of $\hat \mu_i$ is thrice boundedly differentiable uniformly over $T$. d) There exists some fixed $M < \infty$ such that $E[(\bar w_i)^2 | \mu_i = \cdot ] \le M / T$. \[as-gamma-con\] a) The random variable $\gamma_{k,i}$ is continuously distributed. b) The CDF of $\gamma_{k,i}$ is thrice boundedly differentiable. c) The CDF of $\hat \gamma_{k,i}$ is thrice boundedly differentiable uniformly over $T$. d) There exists some fixed $M < \infty$ such that $E[(\bar w_i)^2 | \gamma_{k,i} = \cdot] \le M/T$ and $E[(\hat \gamma_{k,i} - \gamma_{k,i})^2 | \gamma_{k,i} = \cdot] \le M / T$. \[as-rho-con\] a) The random variable $\rho_{k,i}$ is continuously distributed. b) The CDF of $\rho_{k,i}$ is thrice boundedly differentiable. c) The CDF of $\hat \rho_{k,i}$ is thrice boundedly differentiable uniformly over $T$. d) There exists some fixed $M < \infty$ such that $E[(\bar w_i)^2 | \rho_{k,i} = \cdot] \le M/T$, $E[(\hat \gamma_{k,i} - \gamma_{k,i})^2 | \rho_{k,i} = \cdot] \le M/T$, and $E[(\hat \gamma_{0,i} - \gamma_{0,i})^2 | \rho_{k,i} = \cdot] \le M/T$. e) There exist some fixed $\varepsilon > 0$ and $M < \infty$ such that $\hat \gamma_{0,i} > \varepsilon$, $\gamma_{0,i} > \varepsilon$, $|\hat \gamma_{k,i}| < M$, and $|\gamma_{k,i}| < M$ almost surely. Assumption \[as-mu-con\] states that $\mu_i$ and $\hat \mu_i $ are continuous random variables. This assumption is restrictive in the sense that it does not allow a discrete distribution of $\mu_i$ or no heterogeneity in the mean (i.e., $\mu_i$ is homogeneous such that $\mu_i = \mu$ for some constant $\mu$ for any $i$).[^12] [^13] The uniform consistency and functional CLT could not hold without the continuity of $\mu_i$. The assumption also imposes restrictions on the distribution of the noise $\bar w_{i}$. These assumptions are satisfied when data are generated by Gaussian ARMA processes with continuously distributed parameters. That said, we may be able to relax the continuity of $\hat \mu_i$ (the estimated version of $\mu_i$) in condition c), but this requires different proofs to evaluate the order of the bias for the functional CLT. Assumption \[as-mu-con\] also restricts the order of the conditional moment of $\bar w_i$. Note that Lemma \[lem-moment-w\] shows that the order of $E(\bar w_i^2)$ is $1/T$ and this assumption states that the same order holds for the conditional counterpart. Assumptions \[as-gamma-con\] and \[as-rho-con\] are similar to Assumption \[as-mu-con\], except Assumption \[as-rho-con\].e restricts variances $\gamma_{0, i}$ and $\hat \gamma_{0,i}$ that are bounded away from zero and autocovariances $\gamma_{k,i}$ and $\hat \gamma_{k,i}$ that are bounded. We need these additional conditions to examine the empirical distribution for $\hat \rho_{k,i}$. Uniform consistency ------------------- The following theorem establishes the uniform consistency of the distribution estimator. \[thm-gc\] Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-mu-con\].a hold for $r_m = 2$ and $r_d = 2$ if $\hat \xi_i = \hat \mu_i$ and $\xi_i = \mu_i$; that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-gamma-con\].a hold for $r_m = 4$ and $r_d = 4$ if $\hat \xi_i = \hat \gamma_{k,i}$ and $\xi_i = \gamma_{k,i}$; and that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-rho-con\].a hold for $r_m = 4$ and $r_d = 4$ if $\hat \xi_i = \hat \rho_{k,i}$ and $\xi_i = \rho_{k,i}$. When $N,T \to \infty$, the class $\mathcal{F}$ is $P_0$-Glivenko–Cantelli in the sense that $\sup_{f \in \mathcal{F}} | \mathbb{P}_N f - P_{0} f | \stackrel{as}{\longrightarrow} 0$ where $\stackrel{as}{\longrightarrow}$ signifies the almost sure convergence. Note that Theorem \[thm-gc\] cannot be directly shown by the usual Glivenko–Cantelli theorem (e.g., Theorem 19.1 in [@vanderVaart98]) because the true distribution of $\hat \xi_i$ changes as $T$ increases. Nonetheless, our proof follows similar steps to those of the usual Glivenko–Cantelli theorem. Functional central limit theorem -------------------------------- We present the functional CLTs for the empirical distributions of $\hat \mu_i$, $\hat \gamma_{k,i}$, and $\hat \rho_{k,i}$. We aim to derive the asymptotic law of $\sqrt{N}(\mathbb{P}_{N} f - P_{0} f)$ where $f \in \mathcal{F}$. We can also obtain the asymptotic distribution of other quantities via the functional delta method based on this result. The functional CLT for $\mathbb{P}_N$ holds under a similar set of assumptions for the uniform consistency, but we need all of the conditions in Assumption \[as-mu-con\], \[as-gamma-con\], or \[as-rho-con\] to evaluate the order of the bias. We also require a condition on the relative magnitudes of $N$ and $T$ asymptotically to eliminate the bias. Let $\ell^{\infty}(\mathcal{F})$ be the collection of all bounded real functions on $\mathcal{F}$. \[thm-fclt\] Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-mu-con\] hold for $r_m = 4$ and $r_d = 4$ if $\hat \xi_i = \hat \mu_i$ and $\xi_i = \mu_i$; that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-gamma-con\] hold for $r_m = 8$ and $r_d = 8$ if $\hat \xi_i = \hat \gamma_{k,i}$ and $\xi_i = \gamma_{k,i}$; and that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-rho-con\] hold for $r_m = 8$ and $r_d = 8$ if $\hat \xi_i = \hat \rho_{k,i}$ and $\xi_i = \rho_{k,i}$. When $N,T \to \infty$ with $N^{3+\epsilon}/T^4 \to 0$ for some $\epsilon \in (0, 1/3)$, we have $$\begin{aligned} \sqrt{N}(\mathbb{P}_N - P_0) \leadsto \mathbb{G}_{P_0} \qquad \mbox{in} \quad \ell^{\infty}(\mathcal{F}), \end{aligned}$$ where $\leadsto$ signifies weak convergence, $\mathbb{G}_{P_0}$ is a Gaussian process with zero mean and covariance function $E(\mathbb{G}_{P_0}(f_{i})\mathbb{G}_{P_0}(f_j))=F_0(a_i \wedge a_j) - F_0(a_i) F_0(a_j)$ with $f_i = \mathbf{1}_{(-\infty, a_i]}$ and $f_{j} = \mathbf{1}_{(-\infty, a_j]}$ for $a_i, a_j \in \mathbb{R}$ and $a_i \wedge a_j$ is the minimum of $a_i$ and $a_j$. The asymptotic law of the empirical process is Gaussian, which is identical to the limiting distribution for the empirical process constructed using the true $\xi_i = \mu_i$, $\gamma_{k,i}$, or $\rho_{k,i}$. However, this result requires that $N^{3+\epsilon}/T^4 \to 0$ for some $\epsilon$ such that $0 < \epsilon < 1/3$, which allows us to ignore the estimation error in $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$ asymptotically. Note that the condition, $N^{3+\epsilon}/T^4 \to 0$, is almost equivalent to $N^3 / T^4 \to 0$ because we can select an arbitrarily small $\epsilon > 0$. We provide a brief summary of the proof and explain why we require the condition $N^{3+\epsilon}/T^4 \to 0$. The key to understanding the mechanism behind the requirement that $N^{3+\epsilon}/T^4 \to 0$ is to recognize that $E (\mathbb{P}_N f) \neq P_0 f$. That is, $\mathbb{P}_N f$ is not an unbiased estimator for $P_0 f$. As a result, we cannot directly apply the existing results for the empirical process to derive the asymptotic distribution. Let $P_T = P_T^{\hat \xi}$ be the (true) probability measure of $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$. Note that $P_T$ depends on $T$ and $P_T \neq P_0$, and observe that $E(\mathbb{P}_N f) = P_T f$. Let $\mathbb{G}_{N, P_T} \coloneqq \sqrt{N}(\mathbb{P}_N - P_T)$. We observe that $$\begin{aligned} \sqrt{N}(\mathbb{P}_N f - P_0 f) = & \ \mathbb{G}_{N,P_T} f \label{fclt_1} \\ & + \sqrt{N} (P_T f - P_0 f). \label{fclt_2}\end{aligned}$$ For $\mathbb{G}_{N,P_T}$ in , we can directly apply the uniform CLT for the empirical process based on triangular arrays ([@vanderVaartWellner96 Lemma 2.8.7]) and obtain $\mathbb{G}_{N, P_T} \leadsto \mathbb{G}_{P_0}$ in $\ell^{\infty}(\mathcal{F})$ as $N \to \infty$. This part of the proof is standard. We require the condition $N^{3+\epsilon}/T^4 \to 0$ to eliminate the effect of the bias term $\sqrt{N}(P_T f - P_0 f )$ in . In the proof of the theorem, we show that $$\begin{aligned} \sqrt{N} \sup_{f\in\mathcal{F}} \left| P_{T} f - P_{0}f \right| = O\left(\frac{\sqrt{N}}{T^{2/(3+\epsilon)}}\right),\end{aligned}$$ for any $0 < \epsilon < 1/3$. The result is based on the evaluation that the difference between $P_T$ and $P_0$ is of order $O(1 / T^{2/(3+\epsilon)})$. This order is obtained by evaluating the characteristic functions of $\hat \xi_i$ and $\xi_i$, and applying the inversion theorem ([@gil1951note] and [@wendel1961non]). We note that the condition, $0 < \epsilon < 1/3$, is used to ensure the integrability of integrals for the inversion theorem (see the proof for details). As a result, we can establish the weak convergence under the condition $N^{3+\epsilon}/T^4 \to 0$ for a sufficiently small $0 < \epsilon < 1/3$. When we additionally assume that $\hat \xi_i$ exhibits a Gaussian error, we can derive the exact bias that is of order $O(1/T)$ and show the same limiting law as in Theorem \[thm-fclt\] under the weaker condition on the relative magnitudes that $N / T^2 \to 0$. In this case, we can also validate the HPJ bias correction for the distribution estimator. The proof utilizing the Gaussian assumption (and a location–scale assumption) can be found in @JochmansWeidner2018 in a general setting for noisy measurement, and we do not explore such a proof here. However, we stress that our proof for Theorem \[thm-fclt\] is distinct from theirs because we do not assume Gaussianity nor any parametric specification for $\hat \xi_i$ and $\xi_i$ and it requires a quite different proof technique. Functional delta method ----------------------- We can derive the asymptotic distribution of an estimator that is a function of the empirical distribution using the functional delta method. Suppose that we are interested in the asymptotics of $\phi(\mathbb{P}_{N})$ for a functional $\phi:D(\mathcal{F}) \to \mathbb{R}$ where $D(\mathcal{F})$ is the collection of all càdlàg real functions of $\mathcal{F}$. For example, the $\tau$-th quantile $\phi (P_0) = q_{\tau} = F_0^{-1}(\tau) = \inf\{a \in \mathbb{R} : F_0 (a) \geq \tau \}$ for $\tau \in (0,1)$ may be estimated by the empirical quantile of $\hat \xi_i$: $\phi(\mathbb{P}_{N}) = \hat q_{\tau}= \mathbb{F}_N^{-1} (\tau) = \inf\{a \in \mathbb{R} : \mathbb{F}_N (a) \geq \tau \}$. More generally, we can estimate the quantile process $F_0^{-1}$ using the empirical quantile process $\mathbb{F}_N^{-1}$. The derivation of the asymptotic distribution of $\phi(\mathbb{P}_{N})$ is a direct application of the functional delta method (e.g., [@vanderVaartWellner96 Theorems 3.9.4]) and Theorem \[thm-fclt\]. We summarize this result in the following corollary.[^14] \[cor-delta\] Suppose that the assumptions in Theorem \[thm-fclt\] hold. Suppose that $\phi:D(\mathcal{F}) \subset \ell^{\infty}(\mathcal{F}) \to \mathbb{E}$ is Hadamard differentiable at $P_0$ with the derivative $\phi'_{P_0}$ where $\mathbb{E}$ is a normed linear space. When $N,T \to \infty$ with $N^{3+\epsilon}/T^4 \to 0$ for some $\epsilon \in (0, 1/3)$, we have $\sqrt{N} ( \phi(\mathbb{P}_N) - \phi(P_{0}) ) \leadsto \phi'_{P_{0}}(\mathbb{G}_{P_{0}})$. As an example, we can use this result to derive the asymptotic distribution of $\hat q_{\tau}$. The form $\phi'_{P_0}$ for $\hat q_{\tau}$ is available in Example 20.5 in @vanderVaart98 and indicates that $\sqrt{N} (\hat q_{\tau} - q_{\tau}) \leadsto \mathcal{N} ( 0, \tau (1-\tau) / (f (q_\tau))^2)$ where $\mathcal{N}(\mu, \sigma^2)$ is the normal distribution with mean $\mu$ and variance $\sigma^2$ and $f = f^{\xi}$ is the density function of $\xi_i$. We can also derive the asymptotic law of the empirical quantile process $\mathbb{F}_N^{-1}$. If $f$ is continuous and positive in the interval $[F_0^{-1}(p) - \varepsilon, F_0^{-1}(q) + \varepsilon]$ for some $0 < p < q < 1$ and $\varepsilon > 0$, then Corollary \[cor-delta\] means that $$\begin{aligned} \sqrt{N} \left( \mathbb{F}_N^{-1} - F_0^{-1} \right) \leadsto - \frac{\mathbb{G}_{P_0} \circ F_0 \left( F_0^{-1} \right)}{f \left( F_0^{-1} \right)} \qquad \mbox{in} \quad \ell^{\infty}[p, q].\end{aligned}$$ This process is known to be Gaussian with zero mean and a known covariance function (e.g., Example 3.9.24 in [@vanderVaartWellner96]). Function of the expected value of a smooth function of the heterogeneous mean and/or autocovariances {#sec-smooth} ==================================================================================================== In this section, we consider the estimation of a function of the expected value of a smooth function of the heterogeneous mean and/or autocovariances. We also develop the asymptotic justifications of the HPJ bias correction and the cross-sectional bootstrap inference. Asymptotic results {#sec-ghat} ------------------ We derive the asymptotic properties of $\hat S = h(N^{-1} \sum_{i=1}^{N} g( \hat \theta_i ))$ in as the estimator of $S = h(E(g(\theta_i)))$. Define $G \coloneqq E(g(\theta_i))$ and $\hat G \coloneqq N^{-1} \sum_{i=1}^N g(\hat \theta_i)$ such that $S = h(G)$ and $\hat S = h(\hat G)$. We make the following assumptions to develop the asymptotic properties of $\hat S$. \[as-h-c\] The function $h: \mathbb{R}^m \to \mathbb{R}^n$ is continuous in a neighborhood of $G$. \[as-h-cd\] The function $h: \mathbb{R}^m \to \mathbb{R}^n$ is continuously differentiable in a neighborhood of $G$. The matrix of the first derivatives $\nabla h(G)\coloneqq (\nabla h_1(G)^\top, \nabla h_2(G)^\top, \dots, \nabla h_n(G)^\top)^\top$ is of full row rank. \[as-multi-smooth\] The function $g=(g_1,g_2,\dots,g_m):\mathcal{O} \to \mathbb{R}^m$ is twice-continuously differentiable where $\mathcal{O} \subset \mathbb{R}^{l}$ is a convex open subset. The covariance matrix $\Gamma \coloneqq E[(g(\theta_i) - E(g(\theta_i)))(g(\theta_i) - E(g(\theta_i)))^\top]$ exists and is nonsingular. For any $p=1,2,\dots,m$, the elements of the Hessian matrix of $g_p$ are bounded functions. For any $p=1,2,\dots,m$, the function $g_p$ satisfies $E[((\partial/ \partial z_{j}) g_p(z)|_{z=\theta_i})^4] < \infty$ for any $j=1,2,\dots,m$. These assumptions impose conditions on the smoothness of $h$ and $g$ and the existence of moments. Assumption \[as-h-c\] applies to the continuous mapping theorem for the proof of consistency. Assumption \[as-h-cd\] is stronger than Assumption \[as-h-c\] and is used for the application of the delta method to derive the asymptotic distribution. Assumption \[as-multi-smooth\] states that the function $g$ is sufficiently smooth. This assumption is satisfied when the parameter of interest is the mean (i.e., $g(a) = a$) or the $p$-th order moment (i.e., $g(a) =a ^p$), for example. However, this assumption is not satisfied when estimating the CDF (i.e., $g(a) = \mathbf{1} (a \leq c)$ for some $c\in\mathbb{R}$) or quantiles. The existence of the first derivative is crucial for analyzing the asymptotic property of $\hat S$. The second derivative is useful for evaluating the order of the asymptotic bias. Assumption \[as-multi-smooth\] also guarantees that the asymptotic variance exists, which rules out homogeneous dynamics, i.e., it excludes the case where $\theta_i = \theta$ for constant $\theta$ for any $i$ (see the supplementary appendix for the asymptotic results for homogeneous dynamics). The following theorem demonstrates the asymptotic properties of $\hat S$. \[thm-h\] Let $r^* = 4$ if $\theta_i = \mu_i$ such that $S =h( E(g(\mu_i)))$ for some $h$ and $g$, and $r^*=8$ if $\theta_i$ contains $\gamma_{k,i}$ for some $k$. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], \[as-h-c\], and \[as-multi-smooth\] hold for $r_m=4$ and $r_d=r^*$. When $N,T \to \infty$, it holds that $\hat S \stackrel{p}{\longrightarrow} S$. Moreover, suppose that Assumption \[as-h-cd\] also holds. When $N, T \to \infty$ with $N/T^2 \to 0$, it holds that $$\begin{aligned} \sqrt{N} (\hat S - S) \leadsto \mathcal{N} \left(0, \nabla h(G) \Gamma (\nabla h(G))^\top \right).\end{aligned}$$ The estimator $\hat S$ is consistent when both $N$ and $T$ tend to infinity and is asymptotically normal with mean zero when $N/T^2 \to 0$. Importantly, in contrast to the discussion in Section \[sec-asymptotics\], the distribution of $\theta_i$ need not be continuous and can be discrete as long as it is not degenerate (homogeneous). The remarkable result is that the asymptotically unbiased estimation holds under $N/T^2 \to 0$. This condition is weaker than that for $\mathbb{P}_N$, which is $N^{3 + \epsilon}/ T^{4} \to 0$ for some $0 < \epsilon < 1/3$. This result comes from the smoothness of $g$ and the fact that $\hat \theta_i$ is first-order unbiased for $\theta_i$. @FernandezValLee13 also observe similar asymptotic results for estimators of smooth functions of heterogeneous quantities in a different context. To obtain a better understanding of the results in the theorem, we first consider the case in which $\theta_i = \mu_i$ such that $l=1$, $h$ is an identity function such that $n=1$, and $g$ is a scalar function such that $m=1$. Denote our parameter of interest as $G^{\mu} \coloneqq E(g(\mu_i))$ and let $\hat G^{\hat \mu} \coloneqq N^{-1}\sum_{i=1}^N g(\hat \mu_i)$. By Taylor’s theorem and $\hat \mu_i = \mu_i + \bar w_i$, we observe the following expansion: $$\begin{aligned} \sqrt{N}\left( \hat G^{\hat \mu} - G^{\mu} \right) = \frac{1}{\sqrt{N}} \sum_{i=1}^N \Big( g(\mu_i) - E \big( g(\mu_i) \big) \Big) + \frac{1}{\sqrt{N}} \sum_{i=1}^N \bar w_{i} g'(\mu_i) + \frac{1}{2\sqrt{N}} \sum_{i=1}^N (\bar w_i)^2 g''(\tilde{\mu}_i),\label{eq-mu-expansion}\end{aligned}$$ where $\tilde{\mu}_i$ is between $\mu_i$ and $\hat \mu_i$. The second term in has a mean of zero and is of order $O_{p}(1/\sqrt{T})$. The fact that it has a mean of zero is the key reason that a milder condition, $N/T^2 \to 0$, is sufficient for the asymptotically unbiased estimation of $G^{\mu}$. The third term corresponds to the bias caused by the nonlinearity of $g$. When $g$ is linear, this term does not appear and the parameter can be estimated without any restriction on the relative magnitudes of $N$ and $T$. The nonlinearity bias is of order $O_{p}(\sqrt{N}/T)$. We use the condition $N/T^2 \to 0$ to eliminate the effect of this bias. When our parameter of interest involves $\gamma_{k,i}$ for some $k$, we encounter an additional source of bias. Let us consider the case in which $\theta_i = \gamma_{k,i}$ for some $k$ such that $l=1$, $h$ is an identity function such that $n=1$, and $g$ is a scalar function such that $m=1$. We denote our parameter of interest as $G^{\gamma_{k}} \coloneqq E(g(\gamma_{k,i}))$ and let $\hat G^{\hat \gamma_k} \coloneqq N^{-1}\sum_{i=1}^N g(\hat \gamma_{k,i})$. We can expand $\hat \gamma_{k,i}$ as follows: $$\begin{aligned} \hat \gamma_{k,i} = \gamma_{k, i} + \frac{1}{T-k} \sum_{t=k+1}^T (w_{it} w_{i,t-k} - \gamma_{k,i}) - ( \bar w_i)^2 + o_{p}\left(\frac{1}{T}\right).\end{aligned}$$ Note that the second term has a mean of zero, although it is of order $O_p(1/\sqrt{T})$. The third term $(\bar w_i)^2$ is the estimation error in $\bar y_i$ ($=\hat \mu_i$) and is of order $O_p(1/T)$, and causes the incidental parameter bias [@NeymanScott48; @Nickell1981]. By Taylor’s theorem and the expansion of $\hat \gamma_{k, i}$, we have $$\begin{aligned} & \sqrt{N}(\hat G^{\hat \gamma_k} - G^{\gamma_k} ) \nonumber \\ =& \ \frac{1}{\sqrt{N}} \sum_{i=1}^N \Big( g(\gamma_{k,i}) - E \big( g(\gamma_{k,i}) \big) \Big) \label{gex-1}\\ & \ + \frac{1}{\sqrt{N}} \sum_{i=1}^N \left( \frac{1}{T-k} \sum_{t=k+1}^T w_{it} w_{i,t-k} - \gamma_{k,i}\right) g'(\gamma_{k,i}) \label{gex-2}\\ & \ - \frac{1}{\sqrt{N}} \sum_{i=1}^N (\bar w_{i})^2 g'(\gamma_{k,i}) + \frac{1}{2\sqrt{N}} \sum_{i=1}^N (\hat \gamma_{k,i} - \gamma_{k,i})^2 g^{\prime\prime}(\tilde \gamma_{k,i}) + o_{p}\left( \frac{\sqrt{N}}{T}\right) \label{gex-3}, \end{aligned}$$ where $\tilde \gamma_{k,i}$ is between $\hat \gamma_{k,i}$ and $\gamma_{k,i}$. In contrast to $\hat G^{\hat \mu}$, this $\hat G^{\hat \gamma_k}$ has an incidental parameter bias corresponding to the first term in . This bias is of order $O_{p}(\sqrt{N}/T)$ and does not appear in the expansion of $\hat G^{\hat \mu}$. This term makes the condition $N/T^2\to 0$ necessary, even when $g$ is linear. The other terms are similar to those in the expansion of $\hat G^{\hat \mu}$. The term on the right-hand side of yields the asymptotic normality of $\hat G^{\hat \gamma_k}$. The term in has a mean of zero and is of order $O_{p}(1/\sqrt{T})$. The second term in is the nonlinearity bias term that also appears in $\hat G^{\hat \mu}$, which is also of order $O_{p}(\sqrt{N}/T)$. \[remark:kernel\] The analysis here is not applicable to kernel-smoothing estimation, and asymptotic analyses for kernel-smoothing estimators require different proof techniques. To see this, we consider the kernel estimator for the density of $\mu_i$, say $(Nh)^{-1} \sum_{i=1}^N K((x - \hat \mu_i) / h)$, where $K$ is a kernel function and $h \to 0$ is bandwidth. The summand $K((x - \cdot) / h)$ depends on the bandwidth $h$, which shrinks to zero as the sample size increases, so that the shape of the summand changes depending on the sample size, unlike the summand $g(\cdot)$ here. As a result, the kernel estimation requires much more careful investigations for nonlinearity bias terms. @OkuiYanagi2018 formally demonstrate this issue for the kernel density and CDF estimation and find that their relative magnitude conditions of $N$ and $T$ differ from the condition $N / T^2 \to 0$ here and vary in the number of nonlinearity bias terms that can be evaluated. Split-panel jackknife bias correction ------------------------------------- We provide a theoretical justification for the HPJ bias-corrected estimator in , which we base on the bias-correction method proposed by @DhaeneJochmans15. We make the following additional assumptions to study the HPJ estimator of $S$. \[as-g-moment-2\] The function $g=(g_1,g_2,\dots,g_m):\mathcal{O} \to \mathbb{R}^m$ is thrice differentiable. The covariance matrix $\Gamma = E[(g(\theta_i) - E(g(\theta_i)))(g(\theta_i) - E(g(\theta_i)))^\top]$ exists and is nonsingular. For any $p=1,2,\dots,m$, the function $g_p$ satisfies $E[((\partial/ \partial z_{j}) g_p(z)|_{z=\theta_i})^4] < \infty$ for any $j=1,2,\dots,m$, and $E[((\partial^2 / \partial z_{j_1} \partial z_{j_2})g_p(z)|_{z=\theta_i})^4]<\infty$ for any $j_1, j_2 =1,2,\dots,m$. All third-order derivatives of $g$ are bounded. Assumption \[as-g-moment-2\] requires that $g$ is thrice differentiable, contrary to Assumption \[as-multi-smooth\] and imposes stronger moment conditions. We require this condition to conduct a higher-order expansion of $\hat S$. The following theorem shows the asymptotic normality of the HPJ estimator. \[thm-hpj\] Let $r^* = 8$ if $\theta_i = \mu_i$ such that $S = h( E(g(\mu_i)))$ for some $h$ and $g$, and $r^*=16$ if $\theta_i$ contains $\gamma_{k,i}$ for some $k$. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], \[as-h-cd\], and \[as-g-moment-2\] are satisfied for $r_m = 8$ and $r_d =r^*$. When $N,T\to \infty$ with $N/T^2 \to \nu$ for some $\nu \in [0,\infty)$, it holds that $$\begin{aligned} \sqrt{N} (\hat S^H - S) \leadsto \mathcal{N} \left(0, \nabla h(G) \Gamma (\nabla h(G))^\top \right). \end{aligned}$$ The HPJ estimator is asymptotically unbiased, even when $N/T^2 \to 0$ is not satisfied. Moreover, this bias correction does not inflate the asymptotic variance. The reason why the HPJ works is the same as @DhaeneJochmans15, and the detail can be found in the proof of Theorem \[thm-hpj\]. Cross-sectional bootstrap ------------------------- In this section, we present the justification for the use of the cross-sectional bootstrap introduced in Section \[sec-procedures\]. The first theorem concerns $\hat S$ and the second theorem discusses the case where $\hat S^H$. We also provide a theorem for distribution function estimators. We require several additional assumptions. The following assumption is required for Lyapunov’s conditions for $\hat G^*$, which is the estimator of $G$ obtained with the bootstrap sample. Note that $\hat S^*= h(\hat G^*) = h(N^{-1} \sum_{i=1}^N g(\hat \theta_i^*))$ where $\hat \theta_i^*$ is the estimator of $\theta_i$ based on the bootstrap sample. \[as-boot\] The function $g=(g_1,g_2,\dots,g_m):\mathcal{O} \to \mathbb{R}^m$ is twice-continuously differentiable. The covariance matrix of $g(\theta_i)$, $\Gamma$, exists and is nonsingular. The elements of the Hessian matrices of $g_p$ for $p=1,2,\dots, m$, $g_{p_1} (\cdot) g_{p_2} (\cdot)$ for $p_1, p_2 = 1,2,\dots, m$, and $(g(\cdot) ^\top g(\cdot))$ are bounded. For any $p=1,2,\dots,m$, the function $g_p$ satisfies $E[( (\partial/\partial z_{j})g_p(z)|_{z=\theta_i})^4]<\infty$ for any $j=1,2,\dots,l$. For any $p_1, p_2 = 1,2,\dots, m$, $E[ ( (\partial / \partial z_j) g_{p_1}(z) |_{z=\theta_i} g_{p_2} (\theta_i))^2] < \infty$. For any $j=1,2,\dots, l$, $E[ ( g(\theta_i)^\top g(\theta_i) (\partial / \partial z_j) g_{p_1}(z) |_{z=\theta_i} g_{p_2} (\theta_i))^2] < \infty$ is satisfied. The following theorem states that the bootstrap distribution converges to the asymptotic distribution of $\hat S$, but fails to capture the bias term. Let $P^*$ be the bootstrap distribution (that is identical here to the empirical distribution of $\hat \theta_i$, or as below, $\hat \xi_i$). \[thm-bootstrap\] Let $r^* = 4$ if $\theta_i = \mu_i$ such that $S =h( E(g(\mu_i)))$ for some $h$ and $g$, and $r^*=8$ if $\theta_i$ contains $\gamma_{k,i}$ for some $k$. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], \[as-h-cd\], and \[as-boot\] hold for $r_m = 4$ and $r_d =r^*$. When $N, T \to \infty$, we have $$\begin{aligned} \sup_{x \in \mathbb{R}} \left| P^* \left( \sqrt{N} (\hat S^* - \hat S) \le x \right) - \Pr \left( \mathcal{N} \left(0, \nabla h(G) \Gamma (\nabla h(G))^\top \right) \le x \right) \right| \stackrel{p}{\longrightarrow} 0. \end{aligned}$$ The bootstrap does not capture the bias properties of $\hat G$ shown in Section \[sec-ghat\]. This implies that when $T$ is small, we must be cautious about using the bootstrap to make statistical inference. @GalvaoKato14, @GoncalvesKaffo14, and @Kaffo14 also observe similar issues. We can also show that the bootstrap can approximate the asymptotic distribution of the HPJ estimator. The proof is analogous to the proof of Theorem \[thm-bootstrap\], and is thus omitted. Let $r^* = 4$ if $\theta_i = \mu_i$ such that $S =h( E(g(\mu_i)))$ for some $h$ and $g$, and $r^* = 8$ if $\theta_i$ contains $\gamma_{k,i}$ for some $k$. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], \[as-h-cd\], and \[as-boot\] are satisfied for $r_m = 4$ and $r_d =r^*$. When $N,T \to \infty$, we have $$\begin{aligned} \sup_{x \in \mathbb{R}} \left| P^* \left( \sqrt{N} (\hat S^{H*} - \hat S^H) \le x \right) - \Pr \left( \mathcal{N} \left(0, \nabla h(G) \Gamma (\nabla h(G))^\top \right) \le x \right) \right| \stackrel{p}{\longrightarrow} 0. \end{aligned}$$ The cross-sectional bootstrap can approximate the asymptotic distribution of the HPJ estimator correctly under the condition that $N/T^2 $ does not diverge. Because the HPJ estimator has a smaller bias, the bootstrap approximation is more appropriate for the HPJ estimator. Lastly, we show the pointwise validity of the bootstrap for the estimator of the distribution function evaluated at some point $a \in \mathbb{R}$.[^15] Let $\xi_i$ be one of $\mu_i$, $\gamma_{k,i}$, and $\rho_{k,i}$ with the distribution function $F_0 = F_0^\xi$, and $\hat \xi_i$ be the corresponding estimator. The pointwise estimator of $F_0(a)$ at $a \in \mathbb{R}$ is $\mathbb{F}_N (a) = N^{-1} \sum_{i=1}^N \mathbf{1}(\hat \xi_i \le a)$. The bootstrap estimator is $\mathbb{F}_N^* (a) \coloneqq N^{-1} \sum_{i=1}^N \mathbf{1}(\hat \xi_i^* \le a)$. \[thm-dist-bootstrap\] Suppose that the assumptions in Theorem \[thm-gc\] hold. When $N,T \to \infty$, it holds that $$\begin{aligned} \sup_{x \in \mathbb{R}} \left| P^* \left(\sqrt{N}\left(\mathbb{F}_N^{*} (a) - \mathbb{F}_N(a)\right) \le x \right) - \Pr\Big( \mathcal{N} \Big(0, F(a)(1 - F(a)) \Big) \le x \Big) \right| \stackrel{p}{\longrightarrow} 0. \end{aligned}$$ If the rate condition $N^{3 + \epsilon} / T^4 \to 0$ is satisfied for some $\epsilon \in (0, 1/3)$, the bootstrap distribution consistently estimates the asymptotic distribution of $\mathbb{F}_N(a)$. Difference in degrees of heterogeneity {#sec-extensions} ====================================== We develop a two-sample KS test as an application of the convergence for the distribution estimator in Section \[sec-asymptotics\]. Specifically, we develop a test to examine whether the distributions of $\mu_i$, $\gamma_{k,i}$, and $\rho_{k,i}$ differ across distinct groups. In many applications, it would be interesting to see whether distinct groups possess different heterogeneous structures. For example, when studying the LOP deviation, we may want to know whether the distribution of LOP adjustment speed differs between goods and services. We develop a testing procedure for such hypotheses without any parametric specification. We consider two panel data sets for two different groups: $\{\{y_{it,(1)} \}_{t=1}^{T_{1}} \}_{i=1}^{N_1}$ and $\{\{y_{it,(2)}\}_{t=1}^{T_{2}}\}_{i=1}^{N_2}$. We allow $T_1 \neq T_2$ and/or $N_1 \neq N_2$. Define $y_{i,(1)} \coloneqq \{y_{it,(1)}\}_{t=1}^{T_{1}}$ and $y_{i,(2)} \coloneqq \{y_{it,(2)}\}_{t=1}^{T_{2}}$. We estimate the distributions of the mean, autocovariances, or autocorrelations for each group. Let $\xi_{i,(a)} = \mu_{i,(a)}$, $\gamma_{k,i,(a)}$, or $\rho_{k,i,(a)}$ be the true quantity for group $a=1,2$. Let $\hat \xi_{i,(a)} = \hat \mu_{i,(a)}$, $\hat \gamma_{k,i,(a)}$, or $\hat \rho_{k,i,(a)}$ be the corresponding estimator of $\xi_{i,(a)}$. We denote the probability distribution of $\xi_{i,(a)}$ by $P_{0,(a)} = P_{0,(a)}^{\xi}$ and the empirical distribution of $\hat \xi_{i,(a)}$ by $\mathbb{P}_{N_a,(a)} = \mathbb{P}_{N_a,(a)}^{\hat \xi}$ for $a=1,2$. We focus on the following hypothesis to examine the difference in the degrees of heterogeneity between the two groups. $$\begin{aligned} H_{0} : P_{0,(1)} = P_{0,(2)} \; \mbox{ v.s. } \; H_{1} : P_{0,(1)} \neq P_{0,(2)}.\end{aligned}$$ Under the null hypothesis $H_0$, the distributions are identical for the two groups. We investigate the hypothesis using the following two-sample KS statistic based on our empirical distribution estimators. $$\begin{aligned} KS_2 & \coloneqq \sqrt{\frac{N_{1}N_{2}}{N_1 + N_2}} \left\| \mathbb{P}_{N_1,(1)} - \mathbb{P}_{N_2,(2)} \right\|_{\infty} = \sqrt{\frac{N_{1}N_{2}}{N_1 + N_2}} \sup_{f \in \mathcal{F}} \left| \mathbb{P}_{N_1,(1)} f - \mathbb{P}_{N_2,(2)} f \right|,\end{aligned}$$ where $\| \cdot \|_{\infty}$ is the uniform norm. This measures the distance between the empirical distributions of the two groups and differs from the usual two-sample KS statistic in that it is based on the empirical distributions of the estimates. We introduce the following assumption about the data sets. \[as-ks\] Each of $\{\{y_{it,(1)} \}_{t=1}^{T_{1}} \}_{i=1}^{N_1}$ and $\{\{y_{it,(2)}\}_{t=1}^{T_{2}}\}_{i=1}^{N_2}$ satisfies Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-mu-con\] with $r_m = 4$, $r_d = 4$ when $\hat \xi_{i,(a)} = \hat \mu_{i,(a)}$ and $\xi_{i,(a)} = \mu_{i,(a)}$; Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-gamma-con\] with $r_m = 8$, $r_d = 8$ when $\hat \xi_{i,(a)} = \hat \gamma_{k,i,(a)}$ and $\xi_{i,(a)} = \gamma_{k,i,(a)}$; and Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-rho-con\] with $r_m = 8$, $r_d = 8$ when $\hat \xi_{i,(a)} = \hat \rho_{k,i,(a)}$ and $\xi_{i,(a)} = \rho_{k,i,(a)}$. $(y_{1,(1)}, \dots, y_{N_1,(1)})$ and $(y_{1,(2)}, \dots, y_{N_2, (2)})$ are independent. We need the assumptions introduced in the previous sections along with the independence assumption, which implies that our test cannot be used to determine the equivalence of the distributions of two variables from the same units. Our test is intended to compare the distributions of the same variable from different groups. It is also important to note that this independence assumption may collapse when there are some time effects. For example, when the time periods of the two panel data sets overlap, the panel data sets can be dependent given the presence of common time trends. The asymptotic null distribution of $KS_2$ is derived using Theorem \[thm-fclt\]. \[thm-KS2\] Suppose that Assumption \[as-ks\] is satisfied. When $N_{1},T_{1} \to \infty$ with $N_{1}^{3+\epsilon}/T_{1}^4 \to 0$ and $N_{2},T_{2} \to \infty$ with $N_{2}^{3+\epsilon}/T_{2}^4 \to 0$ for some $\epsilon \in (0, 1/3)$ and $N_1 / (N_1 + N_2) \to \lambda$ for some $\lambda \in (0,1)$, it holds that $KS_2$ converges in a distribution to $\| \mathbb{G}_{P_{0,(1)}} \|_{\infty}$ under $H_{0}$. The asymptotic null distribution of $KS_2$ is the uniform norm of a Gaussian process. We require the conditions $N_1^{3 + \epsilon} / T_1^{4} \to 0$ and $N_2^{3 + \epsilon} / T_2^{4} \to 0$ to use the result of Theorem \[thm-fclt\]. The condition $N_1/(N_1+N_2) \to \lambda$ implies that $N_1$ is not much greater or less than $N_2$ and guarantees the existence of the asymptotic null distribution. Note that the asymptotic distribution does not depend on $P_{0,(1)}$, and critical values can be computed readily. @Kolmogorov1933 and @Smirnov1944 (for easy reference see, e.g., Theorem 6.10 in [@Shao03] or Section 2.1.5 in [@Serfling02]) showed that $$\begin{aligned} \label{KS-01} \Pr(\| \mathbb{G}_{P_{0,(1)}} \|_{\infty} \leq a) = 1- 2\sum_{j=1}^{\infty} (-1)^{j-1} \exp\left(-2 j^2 a^2 \right),\end{aligned}$$ for any continuous distribution $P_{0,(1)}$, with $a > 0$. The right-hand side of does not depend on $P_{0,(1)}$. Moreover, the critical values are readily available in many statistical software packages and it is easy to implement our tests. When the true distributions of the estimated quantities, $\hat \xi_{i,(1)}$ and $\hat \xi_{i,(2)}$, are the same, i.e., when $P_{T_1,(1)}^{\hat \xi} = P_{T_2,(2)}^{\hat \xi}$, neither the condition $N_1^{3 + \epsilon} / T_1^4 \to 0$ nor $N_2^{3 + \epsilon} / T_2^4 \to 0$ is needed to establish Theorem \[thm-KS2\]. In particular, when $T_1 = T_2$ and the mean and dynamic structures of the two groups are completely identical under the null hypothesis, we can test the null hypothesis $H_0$ without restricting the relative order of $N_a$ and $T_a$ for $a=1,2$. In this case, both of the distribution function estimates suffer from the same bias, which is canceled out in $KS_2$ under the null hypothesis. Note that we still need the condition $N_1/(N_1 + N_2) \to \lambda \in (0,1)$. Empirical application {#sec-application} ===================== We apply our procedures to panel data on prices in US cities. The speed of price adjustment toward the long-run law of one price (LOP) has important implications in economics. [@AndersonVanWincoop04] survey the literature on price adjustment and trade costs. Several studies focus on the properties of heterogeneity in price deviations from the LOP based on model specifications. For example, [@EngelRogers01] and [@ParsleyWei01] examine the heterogeneity in the time-series volatility of the LOP deviation, and [@CruciniShintaniTsuruga15] consider the heterogeneous properties for the time-series persistence of the LOP deviation. We investigate the heterogeneous properties of the LOP deviations across cities and items using our procedures. We examine whether the LOP deviations dynamics are heterogeneous depending on the item-specific unobserved component such as city- or item-specific permanent trade costs or the item category (e.g., goods or services). Our model-free empirical results complement the findings in existing studies by investigating the heterogeneous properties of the permanent amount, time-series volatility, and persistence of the LOP deviations across cities and items. Data ---- We use data from the American Chamber of Commerce Researchers Association Cost of Living Index produced by the Council of Community and Economic Research.[^16] [@ParsleyWei96], [@YazganYilmazkuday11], and [@CruciniShintaniTsuruga15] use the same data set as [@LeeOkuiShintani13] do for their empirical illustration. The data set contains quarterly price series of 48 consumer price index categorized goods and services for 52 US cities from 1990Q1 to 2007Q4.[^17] The LOP deviation for item $k$ in city $i$ at time $t$ is $y_{i k t}=\ln p_{i k t}-\ln p_{0kt}$ where $p_{ikt}$ is the price of item $k$ in city $i$ at time $t$ and $p_{0kt}$ is that for the benchmark city of Albuquerque, NM. In this empirical application, we regard each item–city pair as a cross-sectional unit, implying a focus on the heterogeneity of the dynamic structures of the LOP deviations across item–city pairs. The number of units is $N=2448$ ($= 48 \times 51$) and the length of the time series is $T=72$ ($=18 \times 4$). Results ------- Table \[table-accra\] summarizes the estimates based on the empirical distribution without bias correction (ED) and the HPJ and TOJ estimates for the distributional features of the heterogeneous means, variances, and first-order autocorrelations of the LOP deviations. The estimates of mean, standard deviation (std), 25% quantile (Q25), median (Q50), and 75% quantile (Q75) for each quantity are presented. We also estimate the correlations between these three quantities. The 95% confidence intervals are computed using the cross-sectional bootstrap. ------------------------------------ -------------------- ------------------ -------------------- -------------------- ------------------ [**Distribution of $\mu$**]{}   ED -0.037 0.131 -0.121 -0.040 0.033   95% CI \[-0.042, -0.032\] \[0.127, 0.136\] \[-0.129, -0.115\] \[-0.046, -0.035\] \[0.027, 0.041\]   HPJ -0.037 0.120 -0.115 -0.040 0.024   95% CI \[-0.042, -0.032\] \[0.115, 0.125\] \[-0.125, -0.108\] \[-0.050, -0.033\] \[0.018, 0.034\]   TOJ -0.037 0.114 -0.110 -0.040 0.014   95% CI \[-0.042, -0.032\] \[0.108, 0.119\] \[-0.127, -0.097\] \[-0.055, -0.030\] \[0.003, 0.034\] [**Distribution of $\gamma_0$**]{}   ED 0.021 0.021 0.010 0.016 0.025   95% CI \[0.020, 0.022\] \[0.018, 0.024\] \[0.009, 0.010\] \[0.015, 0.017\] \[0.024, 0.026\]   HPJ 0.024 0.019 0.013 0.019 0.029   95% CI \[0.023, 0.025\] \[0.017, 0.021\] \[0.012, 0.013\] \[0.018, 0.020\] \[0.028, 0.031\]   TOJ 0.026 0.017 0.015 0.021 0.032   95% CI \[0.024, 0.027\] \[0.014, 0.020\] \[0.014, 0.016\] \[0.020, 0.023\] \[0.029, 0.034\] [**Distribution of $\rho_1$**]{}   ED 0.531 0.206 0.386 0.543 0.691   95% CI \[0.523, 0.539\] \[0.201, 0.211\] \[0.372, 0.393\] \[0.533, 0.555\] \[0.679, 0.702\]   HPJ 0.627 0.159 0.519 0.644 0.748   95% CI \[0.616, 0.636\] \[0.150, 0.168\] \[0.497, 0.532\] \[0.628, 0.664\] \[0.729, 0.767\]   TOJ 0.663 0.134 0.58 0.686 0.760   95% CI \[0.645, 0.679\] \[0.119, 0.149\] \[0.533, 0.612\] \[0.651, 0.723\] \[0.723, 0.796\] ------------------------------------ -------------------- ------------------ -------------------- -------------------- ------------------ : Distribution of price dynamics[]{data-label="table-accra"} -------- ------------------ -------------------- ------------------- ED 0.046 -0.128 0.113 95% CI \[0.002, 0.092\] \[-0.166, -0.087\] \[0.048, 0.177\] HPJ 0.082 -0.160 0.106 95% CI \[0.009, 0.143\] \[-0.218, -0.104\] \[-0.003, 0.205\] TOJ 0.123 -0.170 0.020 95% CI \[0.025, 0.233\] \[-0.260, -0.086\] \[-0.151, 0.167\] -------- ------------------ -------------------- ------------------- : Distribution of price dynamics[]{data-label="table-accra"} The results show that the bias-corrected estimates can substantially differ from the ED estimates, even though this data set has a relatively long time series. In particular, both HPJ and TOJ estimates imply more volatile and persistent dynamics than those implied by the ED estimates. This result demonstrates that the bias correction is important even when $T$ is relatively large. LOP deviations exhibit significant heterogeneity across item–city pairs, as shown in the estimates of the standard deviations and quantiles of the heterogeneous means, variances, and first-order autocorrelations. The standard deviation estimate of the heterogeneous mean indicates a substantial degree of permanent price differences across cities and items. Likewise, the magnitude of the variance in price differences shows large heterogeneity. Interestingly, the positive correlation between the means and the variances implies that the larger the permanent LOP deviation is, the larger the variance of the deviation tends to be. The results for the first-order autocorrelations indicate that the LOP deviations are serially positively correlated. The amount of heterogeneity implied by the bias-corrected estimates is less than that implied by the ED estimate, but all estimates imply that the first-order autocorrelations have a substantial degree of heterogeneity. The first-order autocorrelations are negatively correlated with the mean LOP deviation, and the correlation between first-order autocorrelations and variances is slightly positive. This result indicates that item–city pairs with persistent price difference tend to have small permanent price differences but tend to suffer from relatively large shocks. We also examine whether the distribution of the LOP deviations dynamics differs between goods and services, similarly to prior works that point out different price dynamics between goods and services (e.g., [@ParsleyWei96] and [@NakamuraSteinsson08]). Table \[table-items\] describes the classification of goods and services in our analysis. Table \[table-accra-gs\] summarizes the estimation results, which indicate that price dynamics for goods are markedly different from those of services. In particular, prices for services tend to have more persistent dynamics. In fact, the value of the two-sample KS test for the first-order autocorrelations is 0.353, with a $p$-value of 0. This provides statistical evidence that the speed of price adjustment for services is slower than that for goods. ---------- ------------------------------------------------------------------------------- Goods T-bone steak, Ground beef, Frying chicken, Chunk light tuna, Whole milk, Eggs, Margarine, Parmesan cheese, Potatoes, Bananas, Lettuce, Bread, Coffee, Sugar, Corn flakes, Sweet peas, Peaches, Shortening, Frozen corn, Soft drink, Beer, Wine, Facial tissues, Dishwashing powder, Men’s dress shirt, Shampoo, Toothpaste, Tennis balls. Services Hamburger sandwich, Pizza, Fried chicken, Total home energy cost, Telephone, Apartment, Home purchase price, Mortgage rate, Monthly payment, Dry cleaning, Major appliance repair, Auto maintenance, Gasoline, Doctor office visit, Dentist office visit, Haircut, Beauty salon, Newspaper subscription, Movie, Bowling. ---------- ------------------------------------------------------------------------------- : Items[]{data-label="table-items"} Note: The service category includes those that may be considered as goods, but whose prices are likely to include the cost of a service. Our results are robust to minor modifications to the classification. ------------------------------------ -------------------- ------------------ -------------------- -------------------- ------------------- [**Distribution of $\mu$**]{}   ED -0.042 0.120 -0.127 -0.051 0.026   95% CI \[-0.048, -0.036\] \[0.114, 0.125\] \[-0.134, -0.119\] \[-0.059, -0.045\] \[0.016, 0.032\]   HPJ -0.042 0.107 -0.121 -0.053 0.019   95% CI \[-0.049, -0.036\] \[0.100, 0.113\] \[-0.131, -0.109\] \[-0.064, -0.046\] \[0.006, 0.027\]   TOJ -0.042 0.101 -0.115 -0.054 0.016   95% CI \[-0.048, -0.036\] \[0.094, 0.107\] \[-0.132, -0.094\] \[-0.073, -0.040\] \[-0.005, 0.035\] [**Distribution of $\gamma_0$**]{}   ED 0.025 0.021 0.013 0.019 0.029   95% CI \[0.024, 0.026\] \[0.019, 0.024\] \[0.012, 0.014\] \[0.019, 0.020\] \[0.028, 0.031\]   HPJ 0.028 0.022 0.016 0.023 0.033   95% CI \[0.027, 0.030\] \[0.019, 0.025\] \[0.015, 0.017\] \[0.022, 0.024\] \[0.031, 0.035\]   TOJ 0.030 0.021 0.018 0.025 0.033   95% CI \[0.029, 0.032\] \[0.018, 0.025\] \[0.017, 0.020\] \[0.023, 0.027\] \[0.030, 0.038\] [**Distribution of $\rho_1$**]{}   ED 0.474 0.188 0.348 0.480 0.604   95% CI \[0.465, 0.484\] \[0.181, 0.194\] \[0.333, 0.360\] \[0.466, 0.491\] \[0.591, 0.617\]   HPJ 0.577 0.140 0.491 0.591 0.673   95% CI \[0.564, 0.591\] \[0.130, 0.150\] \[0.463, 0.509\] \[0.564, 0.608\] \[0.654, 0.697\]   TOJ 0.604 0.106 0.553 0.630 0.686   95% CI \[0.583, 0.625\] \[0.086, 0.125\] \[0.502, 0.594\] \[0.581, 0.668\] \[0.649, 0.740\] ------------------------------------ -------------------- ------------------ -------------------- -------------------- ------------------- : Distribution of price dynamics for goods and services[]{data-label="table-accra-gs"} ------------------------------------ -------------------- ------------------ -------------------- -------------------- ------------------ [**Distribution of $\mu$**]{}   ED -0.030 0.146 -0.113 -0.023 0.046   95% CI \[-0.039, -0.022\] \[0.138, 0.154\] \[-0.126, -0.101\] \[-0.031, -0.015\] \[0.034, 0.057\]   HPJ -0.030 0.138 -0.104 -0.024 0.038   95% CI \[-0.039, -0.021\] \[0.129, 0.146\] \[-0.118, -0.091\] \[-0.035, -0.013\] \[0.024, 0.055\]   TOJ -0.030 0.131 -0.093 -0.029 0.028   95% CI \[-0.039, -0.021\] \[0.123, 0.140\] \[-0.117, -0.068\] \[-0.046, -0.011\] \[0.005, 0.055\] [**Distribution of $\gamma_0$**]{}   ED 0.015 0.019 0.006 0.011 0.018   95% CI \[0.014, 0.016\] \[0.011, 0.025\] \[0.006, 0.007\] \[0.010, 0.012\] \[0.018, 0.020\]   HPJ 0.018 0.016 0.008 0.014 0.021   95% CI \[0.016, 0.019\] \[0.011, 0.022\] \[0.007, 0.009\] \[0.013, 0.016\] \[0.020, 0.025\]   TOJ 0.019 0.018 0.009 0.016 0.023   95% CI \[0.018, 0.021\] \[0.009, 0.027\] \[0.008, 0.010\] \[0.014, 0.018\] \[0.020, 0.028\] [**Distribution of $\rho_1$**]{}   ED 0.610 0.204 0.479 0.655 0.764   95% CI \[0.598, 0.624\] \[0.194, 0.213\] \[0.453, 0.510\] \[0.643, 0.669\] \[0.754, 0.778\]   HPJ 0.696 0.161 0.604 0.749 0.809   95% CI \[0.679, 0.713\] \[0.145, 0.174\] \[0.565, 0.655\] \[0.731, 0.769\] \[0.790, 0.828\]   TOJ 0.745 0.130 0.684 0.806 0.822   95% CI \[0.721, 0.769\] \[0.103, 0.155\] \[0.613, 0.778\] \[0.766, 0.851\] \[0.781, 0.867\] ------------------------------------ -------------------- ------------------ -------------------- -------------------- ------------------ : Distribution of price dynamics for goods and services[]{data-label="table-accra-gs"} Our findings are informative in their own right and are in line with existing results. For example, @CruciniShintaniTsuruga15 find significant heterogeneity in LOP deviation dynamics by considering city–city pairs in addition to the city–item pairs. @choi2007heterogeneity find that the speed of price adjustment is heterogeneous, even among tradable goods using Japanese data. As existing studies comparing goods and services, @ParsleyWei96 and @NakamuraSteinsson08 find that services exhibit slower price adjustments and less frequent price changes. Those findings are based on model specifications for heterogeneity, so that our results complement them in a model-free manner with formal statistical procedures. There could be several potential sources of significant heterogeneity in LOP deviation dynamics. For example, there may be some item- and/or city-specific unobservables, such as permanent trade costs and productivity shocks, which can be sources of heterogeneity across items and cities. As another example, an information difference across items and/or cities may lead to heterogeneous dynamic structures. From this viewpoint, @CruciniShintaniTsuruga15 relate heterogeneity to an information difference across managers in different cities based on the noisy information model. The empirical findings here are useful even when the ultimate goal of an application on the LOP deviations is a structural estimation based on some model specifications. For example, our findings here illustrate the importance of taking into account heterogeneity and the type of heterogeneity that needs to be considered in structural estimation. In particular, we demonstrate that goods and services exhibit different heterogeneous dynamic structures, so that empirical researchers should consider different heterogeneity for goods and services. Our recommendation is thus to implement the model-free procedure for understanding the properties of heterogeneous dynamics even when investigating the underlying mechanism and their implications based on structural estimation. Monte Carlo simulation {#sec-LOP-montecarlo} ====================== This section presents the Monte Carlo simulation results. We conduct the simulation using [R]{} with 5,000 replications. Design ------ For $N = 250, 1000, 4000$ and $T = 12, 24, 48$, we generate simulated data using an AR(1) process $$\begin{aligned} y_{it}= (1-\phi_i)\varsigma_i + \phi_i y_{i,t-1} + \sqrt{(1-\phi_i^2)\sigma_i^2} u_{it},\end{aligned}$$ where $u_{it} \sim i.i.d. \ \mathcal{N}(0,1)$. The initial observations are generated by $y_{i0} \sim i.i.d. \ \mathcal{N}(\varsigma_i, \sigma_i^2)$ and $u_{i0} \sim i.i.d. \ \mathcal{N}(0, 1)$. Note that this DGP satisfies $\mu_i = \varsigma_i$, $\gamma_{0,i}=\sigma^2_i$, and $\rho_{1,i}=\phi_i$. The unit-specific random variables $\varsigma_i, \phi_i$, and $\sigma_i^2$ are generated by the truncated normal distribution: $$\begin{pmatrix} \varsigma_i \\ \sigma_i^2 \\ \phi_i \end{pmatrix} \sim i.i.d. \; \mathcal{N} \begin{pmatrix} \begin{pmatrix} -1 \\ 1.5 \\ 0.4 \end{pmatrix} ,& \begin{pmatrix} 1 & 0.2 \cdot 1 \cdot 0.7 & -0.3 \cdot 1 \cdot 0.2 \\ 0.2 \cdot 1 \cdot 0.7 & 0.7^2 & 0.4 \cdot 0.7 \cdot 0.2 \\ -0.3 \cdot 1 \cdot 0.2 & 0.4 \cdot 0.7 \cdot 0.2 & 0.2^2 \end{pmatrix} \end{pmatrix} ,$$ conditional on $\sigma_i^2 > 0$ and $|\phi_i| < 1$. #### Parameters. We estimate the means, standard deviations, 25%, 50%, and 75% quantiles, and correlation coefficients of $\mu_i, \gamma_{0,i}$, and $\rho_{1,i}$. #### Estimators. We consider three estimators: the empirical distribution (ED) without bias correction, the HPJ bias-corrected estimator, and the TOJ bias-corrected estimator. Results ------- Tables \[table-monte-mean\], \[table-monte-acov\], \[table-monte-acor\], and \[table-monte-cor\] summarize the results of the Monte Carlo simulation and provide the bias and the root mean squared error (rmse) of each estimator and the coverage probability (cp) of the 95% confidence interval based on the cross-sectional bootstrap. The column labeled “true” displays the true value of the corresponding quantity. ------------ ---------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- $\mu$ mean $-0.993$ $ 250$ $12$ $ 0.000$ $0.071$ $0.948$ $ 0.000$ $0.071$ $0.947$ $ 0.000$ $0.071$ $0.950$ $-0.993$ $ 250$ $24$ $ 0.001$ $0.067$ $0.945$ $ 0.001$ $0.067$ $0.947$ $ 0.001$ $0.067$ $0.947$ $-0.993$ $ 250$ $48$ $ 0.000$ $0.064$ $0.955$ $ 0.000$ $0.064$ $0.955$ $ 0.000$ $0.064$ $0.956$ $-0.993$ $1000$ $12$ $ 0.000$ $0.035$ $0.954$ $ 0.000$ $0.035$ $0.952$ $ 0.000$ $0.035$ $0.954$ $-0.993$ $1000$ $24$ $-0.001$ $0.034$ $0.945$ $-0.001$ $0.034$ $0.947$ $-0.001$ $0.034$ $0.946$ $-0.993$ $1000$ $48$ $-0.001$ $0.033$ $0.946$ $-0.001$ $0.033$ $0.947$ $-0.001$ $0.033$ $0.946$ $-0.993$ $4000$ $12$ $ 0.000$ $0.018$ $0.947$ $ 0.000$ $0.018$ $0.949$ $ 0.000$ $0.018$ $0.945$ $-0.993$ $4000$ $24$ $ 0.000$ $0.017$ $0.948$ $ 0.000$ $0.017$ $0.951$ $ 0.000$ $0.017$ $0.948$ $-0.993$ $4000$ $48$ $ 0.000$ $0.017$ $0.944$ $ 0.000$ $0.017$ $0.943$ $ 0.000$ $0.017$ $0.940$ $\mu$ std $ 0.997$ $ 250$ $12$ $ 0.135$ $0.145$ $0.248$ $ 0.042$ $0.071$ $0.901$ $ 0.005$ $0.064$ $0.946$ $ 0.997$ $ 250$ $24$ $ 0.076$ $0.091$ $0.676$ $ 0.016$ $0.055$ $0.939$ $ 0.000$ $0.057$ $0.934$ $ 0.997$ $ 250$ $48$ $ 0.040$ $0.061$ $0.891$ $ 0.004$ $0.049$ $0.943$ $-0.001$ $0.050$ $0.939$ $ 0.997$ $1000$ $12$ $ 0.136$ $0.139$ $0.000$ $ 0.044$ $0.052$ $0.673$ $ 0.007$ $0.033$ $0.942$ $ 0.997$ $1000$ $24$ $ 0.076$ $0.080$ $0.121$ $ 0.015$ $0.030$ $0.922$ $-0.001$ $0.028$ $0.946$ $ 0.997$ $1000$ $48$ $ 0.040$ $0.046$ $0.613$ $ 0.004$ $0.025$ $0.949$ $-0.001$ $0.025$ $0.944$ $ 0.997$ $4000$ $12$ $ 0.137$ $0.137$ $0.000$ $ 0.044$ $0.047$ $0.110$ $ 0.007$ $0.017$ $0.931$ $ 0.997$ $4000$ $24$ $ 0.076$ $0.077$ $0.000$ $ 0.016$ $0.020$ $0.780$ $-0.001$ $0.014$ $0.945$ $ 0.997$ $4000$ $48$ $ 0.041$ $0.042$ $0.067$ $ 0.005$ $0.013$ $0.934$ $-0.001$ $0.013$ $0.946$ $\mu$ 25%Q $-1.666$ $ 250$ $12$ $-0.095$ $0.133$ $0.831$ $-0.030$ $0.125$ $0.960$ $-0.001$ $0.200$ $0.989$ $-1.666$ $ 250$ $24$ $-0.054$ $0.103$ $0.910$ $-0.010$ $0.108$ $0.964$ $ 0.003$ $0.169$ $0.992$ $-1.666$ $ 250$ $48$ $-0.026$ $0.091$ $0.941$ $ 0.001$ $0.103$ $0.966$ $ 0.006$ $0.152$ $0.993$ $-1.666$ $1000$ $12$ $-0.099$ $0.109$ $0.431$ $-0.034$ $0.069$ $0.928$ $-0.005$ $0.099$ $0.976$ $-1.666$ $1000$ $24$ $-0.056$ $0.072$ $0.754$ $-0.012$ $0.056$ $0.960$ $ 0.001$ $0.088$ $0.981$ $-1.666$ $1000$ $48$ $-0.029$ $0.053$ $0.896$ $-0.003$ $0.052$ $0.958$ $ 0.002$ $0.077$ $0.982$ $-1.666$ $4000$ $12$ $-0.101$ $0.103$ $0.010$ $-0.036$ $0.048$ $0.778$ $-0.007$ $0.051$ $0.958$ $-1.666$ $4000$ $24$ $-0.056$ $0.060$ $0.291$ $-0.012$ $0.030$ $0.937$ $ 0.002$ $0.044$ $0.970$ $-1.666$ $4000$ $48$ $-0.030$ $0.037$ $0.717$ $-0.003$ $0.027$ $0.948$ $ 0.002$ $0.039$ $0.966$ $\mu$ 50%Q $-0.993$ $ 250$ $12$ $-0.027$ $0.092$ $0.934$ $-0.018$ $0.113$ $0.958$ $-0.006$ $0.180$ $0.988$ $-0.993$ $ 250$ $24$ $-0.019$ $0.085$ $0.943$ $-0.010$ $0.102$ $0.965$ $-0.006$ $0.157$ $0.991$ $-0.993$ $ 250$ $48$ $-0.011$ $0.082$ $0.949$ $-0.004$ $0.097$ $0.961$ $ 0.000$ $0.143$ $0.990$ $-0.993$ $1000$ $12$ $-0.028$ $0.052$ $0.904$ $-0.019$ $0.059$ $0.942$ $-0.010$ $0.092$ $0.972$ $-0.993$ $1000$ $24$ $-0.019$ $0.046$ $0.926$ $-0.009$ $0.053$ $0.952$ $-0.001$ $0.082$ $0.978$ $-0.993$ $1000$ $48$ $-0.012$ $0.042$ $0.940$ $-0.005$ $0.048$ $0.961$ $-0.002$ $0.072$ $0.979$ $-0.993$ $4000$ $12$ $-0.029$ $0.036$ $0.746$ $-0.020$ $0.035$ $0.887$ $-0.011$ $0.048$ $0.951$ $-0.993$ $4000$ $24$ $-0.018$ $0.028$ $0.860$ $-0.009$ $0.027$ $0.944$ $-0.003$ $0.041$ $0.964$ $-0.993$ $4000$ $48$ $-0.011$ $0.024$ $0.904$ $-0.004$ $0.025$ $0.950$ $-0.001$ $0.036$ $0.967$ $\mu$ 75%Q $-0.320$ $ 250$ $12$ $ 0.064$ $0.118$ $0.909$ $ 0.008$ $0.129$ $0.965$ $-0.010$ $0.214$ $0.988$ $-0.320$ $ 250$ $24$ $ 0.035$ $0.099$ $0.940$ $ 0.003$ $0.116$ $0.971$ $-0.002$ $0.183$ $0.993$ $-0.320$ $ 250$ $48$ $ 0.014$ $0.089$ $0.955$ $-0.003$ $0.106$ $0.969$ $-0.005$ $0.161$ $0.994$ $-0.320$ $1000$ $12$ $ 0.067$ $0.083$ $0.733$ $ 0.012$ $0.064$ $0.956$ $-0.007$ $0.105$ $0.978$ $-0.320$ $1000$ $24$ $ 0.034$ $0.058$ $0.893$ $ 0.001$ $0.059$ $0.957$ $-0.005$ $0.094$ $0.981$ $-0.320$ $1000$ $48$ $ 0.015$ $0.047$ $0.936$ $-0.003$ $0.054$ $0.960$ $-0.004$ $0.082$ $0.985$ $-0.320$ $4000$ $12$ $ 0.068$ $0.073$ $0.223$ $ 0.013$ $0.035$ $0.935$ $-0.005$ $0.054$ $0.963$ $-0.320$ $4000$ $24$ $ 0.035$ $0.043$ $0.667$ $ 0.002$ $0.029$ $0.953$ $-0.004$ $0.047$ $0.967$ $-0.320$ $4000$ $48$ $ 0.017$ $0.029$ $0.877$ $ 0.000$ $0.027$ $0.951$ $-0.001$ $0.041$ $0.969$ ------------ ---------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- : Monte Carlo simulation results for $\mu$[]{data-label="table-monte-mean"} ----------------- --------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- $\gamma_0$ mean $1.529$ $ 250$ $12$ $-0.291$ $0.296$ $0.001$ $-0.072$ $0.101$ $0.804$ $ 0.000$ $0.099$ $0.938$ $1.529$ $ 250$ $24$ $-0.159$ $0.167$ $0.151$ $-0.025$ $0.063$ $0.920$ $ 0.002$ $0.073$ $0.945$ $1.529$ $ 250$ $48$ $-0.082$ $0.094$ $0.596$ $-0.007$ $0.050$ $0.946$ $ 0.002$ $0.056$ $0.948$ $1.529$ $1000$ $12$ $-0.291$ $0.292$ $0.000$ $-0.073$ $0.081$ $0.453$ $-0.002$ $0.048$ $0.948$ $1.529$ $1000$ $24$ $-0.157$ $0.159$ $0.000$ $-0.023$ $0.037$ $0.871$ $ 0.005$ $0.036$ $0.949$ $1.529$ $1000$ $48$ $-0.082$ $0.086$ $0.077$ $-0.007$ $0.026$ $0.940$ $ 0.001$ $0.029$ $0.947$ $1.529$ $4000$ $12$ $-0.291$ $0.292$ $0.000$ $-0.073$ $0.075$ $0.022$ $-0.002$ $0.025$ $0.943$ $1.529$ $4000$ $24$ $-0.158$ $0.158$ $0.000$ $-0.024$ $0.028$ $0.628$ $ 0.004$ $0.019$ $0.944$ $1.529$ $4000$ $48$ $-0.082$ $0.083$ $0.000$ $-0.007$ $0.015$ $0.906$ $ 0.002$ $0.015$ $0.945$ $\gamma_0$ std $0.668$ $ 250$ $12$ $ 0.215$ $0.225$ $0.022$ $ 0.165$ $0.188$ $0.462$ $ 0.090$ $0.191$ $0.913$ $0.668$ $ 250$ $24$ $ 0.144$ $0.152$ $0.103$ $ 0.074$ $0.097$ $0.787$ $ 0.020$ $0.107$ $0.938$ $0.668$ $ 250$ $48$ $ 0.087$ $0.095$ $0.384$ $ 0.029$ $0.054$ $0.912$ $ 0.002$ $0.067$ $0.936$ $0.668$ $1000$ $12$ $ 0.216$ $0.219$ $0.000$ $ 0.165$ $0.171$ $0.010$ $ 0.090$ $0.124$ $0.811$ $0.668$ $1000$ $24$ $ 0.147$ $0.149$ $0.000$ $ 0.077$ $0.084$ $0.244$ $ 0.024$ $0.060$ $0.928$ $0.668$ $1000$ $48$ $ 0.087$ $0.090$ $0.003$ $ 0.029$ $0.037$ $0.765$ $ 0.002$ $0.034$ $0.943$ $0.668$ $4000$ $12$ $ 0.217$ $0.217$ $0.000$ $ 0.165$ $0.167$ $0.000$ $ 0.091$ $0.100$ $0.401$ $0.668$ $4000$ $24$ $ 0.146$ $0.147$ $0.000$ $ 0.076$ $0.077$ $0.000$ $ 0.022$ $0.035$ $0.875$ $0.668$ $4000$ $48$ $ 0.088$ $0.088$ $0.000$ $ 0.029$ $0.031$ $0.264$ $ 0.002$ $0.017$ $0.949$ $\gamma_0$ 25%Q $1.055$ $ 250$ $12$ $-0.451$ $0.453$ $0.000$ $-0.221$ $0.233$ $0.209$ $-0.085$ $0.150$ $0.923$ $1.055$ $ 250$ $24$ $-0.272$ $0.277$ $0.001$ $-0.093$ $0.121$ $0.788$ $-0.017$ $0.129$ $0.979$ $1.055$ $ 250$ $48$ $-0.152$ $0.161$ $0.220$ $-0.033$ $0.081$ $0.942$ $ 0.000$ $0.118$ $0.986$ $1.055$ $1000$ $12$ $-0.453$ $0.454$ $0.000$ $-0.224$ $0.227$ $0.000$ $-0.089$ $0.109$ $0.742$ $1.055$ $1000$ $24$ $-0.274$ $0.275$ $0.000$ $-0.094$ $0.102$ $0.327$ $-0.018$ $0.066$ $0.958$ $1.055$ $1000$ $48$ $-0.154$ $0.156$ $0.000$ $-0.034$ $0.050$ $0.864$ $ 0.000$ $0.059$ $0.974$ $1.055$ $4000$ $12$ $-0.454$ $0.454$ $0.000$ $-0.225$ $0.226$ $0.000$ $-0.090$ $0.095$ $0.198$ $1.055$ $4000$ $24$ $-0.274$ $0.275$ $0.000$ $-0.095$ $0.097$ $0.003$ $-0.018$ $0.036$ $0.926$ $1.055$ $4000$ $48$ $-0.154$ $0.155$ $0.000$ $-0.034$ $0.039$ $0.551$ $ 0.000$ $0.030$ $0.964$ $\gamma_0$ 50%Q $1.515$ $ 250$ $12$ $-0.472$ $0.476$ $0.000$ $-0.182$ $0.206$ $0.549$ $-0.054$ $0.175$ $0.960$ $1.515$ $ 250$ $24$ $-0.274$ $0.281$ $0.006$ $-0.076$ $0.115$ $0.873$ $-0.015$ $0.152$ $0.976$ $1.515$ $ 250$ $48$ $-0.150$ $0.161$ $0.286$ $-0.027$ $0.082$ $0.951$ $-0.001$ $0.128$ $0.987$ $1.515$ $1000$ $12$ $-0.473$ $0.474$ $0.000$ $-0.183$ $0.189$ $0.048$ $-0.054$ $0.100$ $0.912$ $1.515$ $1000$ $24$ $-0.274$ $0.276$ $0.000$ $-0.075$ $0.087$ $0.605$ $-0.014$ $0.075$ $0.966$ $1.515$ $1000$ $48$ $-0.151$ $0.154$ $0.000$ $-0.028$ $0.049$ $0.894$ $-0.003$ $0.066$ $0.972$ $1.515$ $4000$ $12$ $-0.474$ $0.474$ $0.000$ $-0.184$ $0.185$ $0.000$ $-0.056$ $0.070$ $0.730$ $1.515$ $4000$ $24$ $-0.275$ $0.275$ $0.000$ $-0.076$ $0.079$ $0.071$ $-0.014$ $0.040$ $0.943$ $1.515$ $4000$ $48$ $-0.152$ $0.152$ $0.000$ $-0.029$ $0.035$ $0.696$ $-0.003$ $0.033$ $0.959$ $\gamma_0$ 75%Q $1.982$ $ 250$ $12$ $-0.325$ $0.337$ $0.093$ $-0.037$ $0.146$ $0.948$ $ 0.029$ $0.260$ $0.980$ $1.982$ $ 250$ $24$ $-0.169$ $0.187$ $0.468$ $-0.011$ $0.116$ $0.965$ $ 0.002$ $0.209$ $0.987$ $1.982$ $ 250$ $48$ $-0.085$ $0.112$ $0.773$ $-0.003$ $0.100$ $0.966$ $-0.001$ $0.178$ $0.988$ $1.982$ $1000$ $12$ $-0.324$ $0.327$ $0.000$ $-0.036$ $0.079$ $0.925$ $ 0.029$ $0.132$ $0.964$ $1.982$ $1000$ $24$ $-0.165$ $0.169$ $0.027$ $-0.005$ $0.059$ $0.956$ $ 0.009$ $0.109$ $0.967$ $1.982$ $1000$ $48$ $-0.084$ $0.092$ $0.377$ $-0.003$ $0.050$ $0.958$ $-0.001$ $0.090$ $0.970$ $1.982$ $4000$ $12$ $-0.324$ $0.324$ $0.000$ $-0.036$ $0.050$ $0.826$ $ 0.029$ $0.071$ $0.938$ $1.982$ $4000$ $24$ $-0.165$ $0.166$ $0.000$ $-0.006$ $0.031$ $0.944$ $ 0.009$ $0.055$ $0.957$ $1.982$ $4000$ $48$ $-0.083$ $0.085$ $0.006$ $-0.002$ $0.025$ $0.950$ $ 0.000$ $0.045$ $0.962$ ----------------- --------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- : Monte Carlo simulation results for $\gamma_0$[]{data-label="table-monte-acov"} --------------- --------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- $\rho_1$ mean $0.397$ $ 250$ $12$ $-0.199$ $0.200$ $0.000$ $ 0.008$ $0.030$ $0.939$ $ 0.004$ $0.054$ $0.949$ $0.397$ $ 250$ $24$ $-0.097$ $0.098$ $0.000$ $ 0.005$ $0.021$ $0.944$ $ 0.004$ $0.033$ $0.951$ $0.397$ $ 250$ $48$ $-0.047$ $0.049$ $0.087$ $ 0.002$ $0.016$ $0.945$ $ 0.001$ $0.021$ $0.948$ $0.397$ $1000$ $12$ $-0.200$ $0.200$ $0.000$ $ 0.007$ $0.016$ $0.916$ $ 0.004$ $0.028$ $0.944$ $0.397$ $1000$ $24$ $-0.097$ $0.097$ $0.000$ $ 0.005$ $0.011$ $0.919$ $ 0.003$ $0.017$ $0.946$ $0.397$ $1000$ $48$ $-0.047$ $0.048$ $0.000$ $ 0.002$ $0.008$ $0.939$ $ 0.000$ $0.011$ $0.946$ $0.397$ $4000$ $12$ $-0.199$ $0.200$ $0.000$ $ 0.008$ $0.010$ $0.825$ $ 0.003$ $0.014$ $0.941$ $0.397$ $4000$ $24$ $-0.097$ $0.097$ $0.000$ $ 0.005$ $0.007$ $0.818$ $ 0.004$ $0.009$ $0.928$ $0.397$ $4000$ $48$ $-0.047$ $0.048$ $0.000$ $ 0.002$ $0.005$ $0.907$ $ 0.000$ $0.005$ $0.941$ $\rho_1$ std $0.198$ $ 250$ $12$ $ 0.109$ $0.109$ $0.000$ $ 0.045$ $0.050$ $0.496$ $-0.030$ $0.055$ $0.889$ $0.198$ $ 250$ $24$ $ 0.058$ $0.059$ $0.000$ $ 0.007$ $0.018$ $0.931$ $-0.012$ $0.034$ $0.924$ $0.198$ $ 250$ $48$ $ 0.029$ $0.031$ $0.128$ $ 0.001$ $0.013$ $0.945$ $-0.003$ $0.021$ $0.940$ $0.198$ $1000$ $12$ $ 0.109$ $0.109$ $0.000$ $ 0.045$ $0.047$ $0.018$ $-0.030$ $0.038$ $0.739$ $0.198$ $1000$ $24$ $ 0.058$ $0.059$ $0.000$ $ 0.008$ $0.012$ $0.845$ $-0.011$ $0.019$ $0.885$ $0.198$ $1000$ $48$ $ 0.030$ $0.030$ $0.000$ $ 0.001$ $0.006$ $0.950$ $-0.002$ $0.010$ $0.946$ $0.198$ $4000$ $12$ $ 0.109$ $0.109$ $0.000$ $ 0.045$ $0.046$ $0.000$ $-0.030$ $0.032$ $0.265$ $0.198$ $4000$ $24$ $ 0.058$ $0.058$ $0.000$ $ 0.008$ $0.009$ $0.545$ $-0.011$ $0.014$ $0.697$ $0.198$ $4000$ $48$ $ 0.030$ $0.030$ $0.000$ $ 0.001$ $0.003$ $0.940$ $-0.002$ $0.006$ $0.916$ $\rho_1$ 25%Q $0.263$ $ 250$ $12$ $-0.275$ $0.277$ $0.000$ $-0.011$ $0.053$ $0.957$ $ 0.079$ $0.133$ $0.926$ $0.263$ $ 250$ $24$ $-0.135$ $0.138$ $0.000$ $ 0.005$ $0.038$ $0.957$ $ 0.015$ $0.079$ $0.975$ $0.263$ $ 250$ $48$ $-0.066$ $0.069$ $0.105$ $ 0.003$ $0.030$ $0.964$ $ 0.002$ $0.056$ $0.985$ $0.263$ $1000$ $12$ $-0.277$ $0.277$ $0.000$ $-0.013$ $0.029$ $0.926$ $ 0.079$ $0.096$ $0.721$ $0.263$ $1000$ $24$ $-0.137$ $0.137$ $0.000$ $ 0.003$ $0.019$ $0.951$ $ 0.010$ $0.041$ $0.956$ $0.263$ $1000$ $48$ $-0.067$ $0.067$ $0.000$ $ 0.003$ $0.015$ $0.954$ $ 0.002$ $0.028$ $0.971$ $0.263$ $4000$ $12$ $-0.277$ $0.277$ $0.000$ $-0.012$ $0.018$ $0.840$ $ 0.079$ $0.084$ $0.170$ $0.263$ $4000$ $24$ $-0.137$ $0.137$ $0.000$ $ 0.003$ $0.010$ $0.938$ $ 0.011$ $0.022$ $0.925$ $0.263$ $4000$ $48$ $-0.067$ $0.067$ $0.000$ $ 0.003$ $0.008$ $0.939$ $ 0.002$ $0.014$ $0.959$ $\rho_1$ 50%Q $0.397$ $ 250$ $12$ $-0.182$ $0.184$ $0.000$ $ 0.020$ $0.049$ $0.944$ $-0.020$ $0.098$ $0.967$ $0.397$ $ 250$ $24$ $-0.085$ $0.088$ $0.019$ $ 0.013$ $0.036$ $0.940$ $ 0.007$ $0.068$ $0.972$ $0.397$ $ 250$ $48$ $-0.040$ $0.045$ $0.407$ $ 0.005$ $0.027$ $0.958$ $ 0.001$ $0.049$ $0.982$ $0.397$ $1000$ $12$ $-0.183$ $0.183$ $0.000$ $ 0.019$ $0.030$ $0.868$ $-0.024$ $0.055$ $0.938$ $0.397$ $1000$ $24$ $-0.086$ $0.086$ $0.000$ $ 0.011$ $0.020$ $0.903$ $ 0.004$ $0.034$ $0.962$ $0.397$ $1000$ $48$ $-0.041$ $0.042$ $0.008$ $ 0.004$ $0.014$ $0.940$ $-0.001$ $0.025$ $0.968$ $0.397$ $4000$ $12$ $-0.183$ $0.183$ $0.000$ $ 0.019$ $0.022$ $0.608$ $-0.025$ $0.035$ $0.837$ $0.397$ $4000$ $24$ $-0.085$ $0.086$ $0.000$ $ 0.012$ $0.014$ $0.710$ $ 0.004$ $0.018$ $0.946$ $0.397$ $4000$ $48$ $-0.041$ $0.041$ $0.000$ $ 0.004$ $0.008$ $0.900$ $ 0.000$ $0.013$ $0.951$ $\rho_1$ 75%Q $0.531$ $ 250$ $12$ $-0.106$ $0.109$ $0.011$ $ 0.045$ $0.063$ $0.866$ $ 0.012$ $0.096$ $0.969$ $0.531$ $ 250$ $24$ $-0.047$ $0.051$ $0.393$ $ 0.012$ $0.036$ $0.951$ $-0.003$ $0.068$ $0.980$ $0.531$ $ 250$ $48$ $-0.021$ $0.028$ $0.782$ $ 0.004$ $0.028$ $0.961$ $-0.002$ $0.051$ $0.982$ $0.531$ $1000$ $12$ $-0.106$ $0.106$ $0.000$ $ 0.045$ $0.050$ $0.502$ $ 0.013$ $0.049$ $0.954$ $0.531$ $1000$ $24$ $-0.046$ $0.047$ $0.009$ $ 0.014$ $0.022$ $0.879$ $ 0.000$ $0.034$ $0.968$ $0.531$ $1000$ $48$ $-0.021$ $0.023$ $0.417$ $ 0.004$ $0.014$ $0.946$ $-0.002$ $0.025$ $0.969$ $0.531$ $4000$ $12$ $-0.105$ $0.105$ $0.000$ $ 0.045$ $0.046$ $0.021$ $ 0.013$ $0.027$ $0.920$ $0.531$ $4000$ $24$ $-0.045$ $0.046$ $0.000$ $ 0.014$ $0.017$ $0.608$ $ 0.000$ $0.017$ $0.956$ $0.531$ $4000$ $48$ $-0.020$ $0.021$ $0.009$ $ 0.004$ $0.008$ $0.907$ $-0.001$ $0.013$ $0.964$ --------------- --------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- : Monte Carlo simulation results for $\rho_1$[]{data-label="table-monte-acor"} ------------------------ ---------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- $\mu$ vs $\gamma_0$ $ 0.193$ $ 250$ $12$ $-0.142$ $0.156$ $0.377$ $-0.103$ $0.136$ $0.762$ $-0.066$ $0.152$ $0.892$ $ 0.193$ $ 250$ $24$ $-0.098$ $0.117$ $0.638$ $-0.055$ $0.097$ $0.884$ $-0.026$ $0.112$ $0.930$ $ 0.193$ $ 250$ $48$ $-0.061$ $0.088$ $0.826$ $-0.024$ $0.078$ $0.934$ $-0.006$ $0.090$ $0.941$ $ 0.193$ $1000$ $12$ $-0.142$ $0.146$ $0.008$ $-0.103$ $0.112$ $0.352$ $-0.068$ $0.095$ $0.799$ $ 0.193$ $1000$ $24$ $-0.100$ $0.105$ $0.111$ $-0.057$ $0.070$ $0.692$ $-0.028$ $0.062$ $0.912$ $ 0.193$ $1000$ $48$ $-0.062$ $0.070$ $0.485$ $-0.025$ $0.045$ $0.889$ $-0.006$ $0.045$ $0.939$ $ 0.193$ $4000$ $12$ $-0.142$ $0.143$ $0.000$ $-0.103$ $0.105$ $0.002$ $-0.068$ $0.075$ $0.467$ $ 0.193$ $4000$ $24$ $-0.100$ $0.101$ $0.000$ $-0.056$ $0.060$ $0.200$ $-0.027$ $0.038$ $0.825$ $ 0.193$ $4000$ $48$ $-0.062$ $0.064$ $0.021$ $-0.025$ $0.031$ $0.725$ $-0.007$ $0.023$ $0.938$ $\mu$ vs $\rho_1$ $ 0.405$ $ 250$ $12$ $-0.245$ $0.253$ $0.014$ $-0.161$ $0.191$ $0.619$ $-0.087$ $0.200$ $0.911$ $ 0.405$ $ 250$ $24$ $-0.158$ $0.169$ $0.207$ $-0.072$ $0.110$ $0.863$ $-0.017$ $0.134$ $0.940$ $ 0.405$ $ 250$ $48$ $-0.091$ $0.107$ $0.634$ $-0.024$ $0.075$ $0.934$ $ 0.003$ $0.096$ $0.941$ $ 0.405$ $1000$ $12$ $-0.245$ $0.247$ $0.000$ $-0.163$ $0.170$ $0.086$ $-0.092$ $0.128$ $0.823$ $ 0.405$ $1000$ $24$ $-0.158$ $0.160$ $0.000$ $-0.070$ $0.082$ $0.604$ $-0.014$ $0.067$ $0.941$ $ 0.405$ $1000$ $48$ $-0.091$ $0.095$ $0.095$ $-0.023$ $0.042$ $0.897$ $ 0.004$ $0.048$ $0.943$ $ 0.405$ $4000$ $12$ $-0.245$ $0.246$ $0.000$ $-0.163$ $0.165$ $0.000$ $-0.091$ $0.101$ $0.465$ $ 0.405$ $4000$ $24$ $-0.158$ $0.158$ $0.000$ $-0.070$ $0.073$ $0.074$ $-0.014$ $0.036$ $0.927$ $ 0.405$ $4000$ $48$ $-0.090$ $0.091$ $0.000$ $-0.023$ $0.029$ $0.745$ $ 0.004$ $0.024$ $0.946$ $\gamma_0$ vs $\rho_1$ $-0.286$ $ 250$ $12$ $ 0.367$ $0.373$ $0.000$ $ 0.326$ $0.343$ $0.141$ $ 0.240$ $0.302$ $0.740$ $-0.286$ $ 250$ $24$ $ 0.271$ $0.280$ $0.019$ $ 0.175$ $0.200$ $0.565$ $ 0.076$ $0.172$ $0.917$ $-0.286$ $ 250$ $48$ $ 0.172$ $0.184$ $0.242$ $ 0.072$ $0.112$ $0.870$ $ 0.011$ $0.122$ $0.950$ $-0.286$ $1000$ $12$ $ 0.367$ $0.369$ $0.000$ $ 0.327$ $0.332$ $0.000$ $ 0.241$ $0.258$ $0.247$ $-0.286$ $1000$ $24$ $ 0.271$ $0.273$ $0.000$ $ 0.174$ $0.181$ $0.054$ $ 0.078$ $0.110$ $0.825$ $-0.286$ $1000$ $48$ $ 0.170$ $0.174$ $0.000$ $ 0.070$ $0.083$ $0.637$ $ 0.008$ $0.063$ $0.940$ $-0.286$ $4000$ $12$ $ 0.368$ $0.368$ $0.000$ $ 0.328$ $0.329$ $0.000$ $ 0.242$ $0.246$ $0.000$ $-0.286$ $4000$ $24$ $ 0.271$ $0.271$ $0.000$ $ 0.174$ $0.176$ $0.000$ $ 0.077$ $0.086$ $0.479$ $-0.286$ $4000$ $48$ $ 0.171$ $0.172$ $0.000$ $ 0.070$ $0.074$ $0.094$ $ 0.009$ $0.032$ $0.938$ ------------------------ ---------- -------- ------ -- ---------- --------- --------- -- ---------- --------- --------- -- ---------- --------- --------- : Monte Carlo simulation results for correlations[]{data-label="table-monte-cor"} The simulation result demonstrates that our asymptotic analyses provide information about the finite-sample behavior and the importance of bias correction. First, ED has large biases in some parameters of interest, such as the quantities $\gamma_{0,i}$ and $Cor(\gamma_{0,i}, \rho_{1,i})$. Second, for many parameters, the coverage probabilities of ED differ significantly from 0.95 because of these large biases. Third, the biases and coverage probabilities of ED can improve when $T$ is large, although significant biases can remain, even with a large $T$. These results recommend the importance of developing a bias-correction method. The split-panel jackknife bias correction reduces biases and improves coverage probabilities for many parameters. HPJ can work well, especially when biases in ED are large. The coverage probabilities of HPJ are satisfactory for about half of the cases, in particular those with large $T$ and those in which the parameter of interest is $\mu_i$ and $\rho_{1,i}$. Conversely, when $T$ is small and when the parameter of interest is $\gamma_{0,i}$ or std, they are not satisfactory. We suspect that large higher-order biases caused by a small $T$ or highly nonlinear parameters may be present in those cases in which HPJ does not work well. For such cases, TOJ can further improve both the biases and coverage probabilities, which can be expected by our discussion for higher-order jackknife in the supplementary appendix. In contrast, in some cases TOJ eliminates biases at the inevitable cost of inflation of standard deviations, which may lead to wider confidence intervals, and the coverage probabilities for TOJ may be over 0.95 when estimating some quantiles. In summary, our recommendation based on these simulation results is to employ split-panel jackknife bias-corrected estimation. When HPJ and TOJ estimates are close to each other, both estimates could be reliable. In contrast, when both estimates differ due to a severe higher-order bias, we could rely on TOJ, especially when estimating highly nonlinear parameters, while being cautious about the precision of point estimates. ED is not recommended. Conclusion {#sec-conclusion} ========== This paper proposes methods to analyze heterogeneous dynamic structures using panel data. Our methods are easily implemented without requiring a model specification. We first compute the sample mean, autocovariances, and autocorrelations for each unit. We then use these to estimate the parameters of interest, such as the distribution function, the quantile function, and the other moments of the heterogeneous mean, autocovariances, and/or autocorrelations. We establish conditions on the relative magnitudes of $N$ and $T$ under which the estimator for the distribution function does not suffer from asymptotic bias. When the parameter of interest can be written as the expected value of a smooth function of the heterogeneous mean and/or autocovariances, the bias of the estimator is of order $O(1/T)$ and can be reduced using the split-panel jackknife bias correction. In addition, we develop inference based on the cross-sectional bootstrap and provide an extension based on the proposed procedures involving the testing of differences in heterogeneous dynamic structures across distinct groups. We apply our procedures to the dynamics of LOP deviations in different US cities for various items and obtain new empirical evidence for significant heterogeneity. The results of the Monte Carlo simulations demonstrate the desirable properties of the proposed procedures. #### Future work. Several future research topics are possible. First, it would be interesting to develop a formal testing procedure to examine whether the dynamics are heterogeneous. We are currently working on this extension. Second, it would be interesting to examine quantities in addition to means and autocovariances. For example, [@ArellanoBlundellBonhomme15] highlight the importance of nonlinearity and conditional skewness in earnings and consumption dynamics. Third, while we developed our analysis for stationary panel data, it would be interesting to consider nonstationary panel data. Finally, while we focus only on balanced panel data, an analysis based on unbalanced panel data would be useful. Appendix: Proofs and technical lemmas {#sec-appendix} ===================================== This appendix presents the proofs of the theorems and the technical lemmas used to prove the theorems. Section \[subsec-proof-theorem\] contains the proofs for the theorems in the main text. The technical lemmas are given in Section \[subsec-technical-lemma\]. Proofs of theorems {#subsec-proof-theorem} ------------------ This section contains the proofs of the theorems in the main text. We repeatedly cite the results in @vanderVaartWellner96, subsequently abbreviated as VW. We also denote a generic constant by $M < \infty$ throughout. ### Proof of Theorem \[thm-gc\] Let $\mathbb{P}_N = \mathbb{P}_N^{\hat \xi}$, $P_T = P_T^{\hat \xi}$, and $P_0 = P_0^{\xi}$ be the probability measures defined in the main body for $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$ or $\hat \rho_{k,i}$, and $\xi_i = \mu_i$, $\gamma_{k,i}$ or $\rho_{k,i}$, respectively. Let $\mathbb{F}_N$, $F_T$ and $F_0$ be the corresponding CDFs. By the triangle inequality, we have $\sup_{f \in \mathcal{F}} | \mathbb{P}_{N}f - P_{0}f | \leq \sup_{f \in \mathcal{F}} | \mathbb{P}_{N}f - P_Tf | + \sup_{f \in \mathcal{F}} | P_T f - P_{0}f |$. For the second term on the right-hand side, Corollary \[lem-second-mu\] or Lemma \[lem-gamma-moment\], or \[lem-rho-moment\] for $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$, respectively, implies that $\hat \xi_i$ converges to $\xi_i$ in mean square convergence and thus also implies that $\hat \xi_i$ converges to $\xi_i$ in distribution. Hence, Lemma 2.11 in @vanderVaart98 implies that $\sup_{f \in \mathcal{F}} | P_T f - P_{0}f | \to 0$ as $\xi_i$ is continuously distributed by Assumption \[as-mu-con\].a, \[as-gamma-con\].a, or \[as-rho-con\].a. We then show that the first term almost surely converges to $0$. Note that, for $f=\mathbf{1}_{(-\infty,a]}$, $\mathbb{P}_{N}f = \mathbb{F}_{N}(a)$, and $E(\mathbb{F}_{N}(a)) = \Pr(\hat \xi_i \leq a)=P_T f$. We first fix a monotonic sequence $T= T(N)$ such that $T\to \infty$ as $N\to \infty$, which transforms our sample into triangular arrays. We use the strong law of large numbers for triangular arrays (see, e.g., [@Hu1989 Theorem 2]). This is possible because under Assumption \[as-basic\], $\mathbf{1}(\hat \xi_i \leq a)$ for any $a \in \mathbb{R}$ is i.i.d. across units, the condition (1.5) in @Hu1989 is clearly satisfied, and the condition (1.6) in @Hu1989 is also satisfied when we set $X=2$ in condition (1.6). Thus, we have $\mathbb{F}_N(a) - \Pr(\hat \xi_i \leq a) \stackrel{as}{\longrightarrow} 0$ and $\mathbb{F}_N(a-) - \Pr(\hat \xi_i < a) \stackrel{as}{\longrightarrow} 0$ for every $a \in \mathbb{R}$, when $T(N) \to \infty$ as $N \to \infty$. Given a fixed $\varepsilon > 0$, there exists a partition $-\infty = a_0 < a_1 <\cdots< a_L =\infty$ such that $\Pr(\xi_i < a_l) - \Pr(\xi_i \leq a_{l-1}) < \varepsilon /3$ for every $l$. We showed $ \sup_{f \in \mathcal{F}} \left| P_T f - P_{0}f \right| \to 0$, which implies that for sufficiently large $N,T$, $\sup_{f \in \mathcal{F}} \left| P_T f - P_{0}f \right| < \varepsilon /3$. Therefore, we have $\Pr(\hat \xi_i < a_l) - \Pr(\hat \xi_i \leq a_{l-1}) < \varepsilon$ for every $l$. The rest of the proof is the same as that for Theorem 19.1 in @vanderVaart98. For $a_{l-1} \leq a < a_{l}$, $$\begin{aligned} \mathbb{F}_N (a) - \Pr(\hat \xi_i \leq a) &\leq \mathbb{F}_N (a_l-) - \Pr(\hat \xi_i < a_l) + \varepsilon,\\ \mathbb{F}_N (a) - \Pr(\hat \xi_i \leq a) &\geq \mathbb{F}_N (a_{l-1}-) - \Pr(\hat \xi_i < a_{l-1}) - \varepsilon.\end{aligned}$$ Accordingly, we have $\limsup_{N,T \to \infty} (\sup_{f \in \mathcal{F}} |\mathbb{P}_{N}f - P_T f|) \leq \varepsilon$ almost surely. This is true for every $\varepsilon > 0$, and we thus get $\sup_{f \in \mathcal{F}} | \mathbb{P}_{N}f - P_T f | \stackrel{as}{\longrightarrow} 0$. We note that this result holds for all monotonic diagonal paths $N \to \infty, T(N) \to \infty$. As stated in REMARKS (a) in @PhillipsMoon99, it thus holds under double asymptotics $N,T \to \infty$. Consequently, we obtain the desired result by the continuous mapping theorem. ### Proof of Theorem \[thm-fclt\] The proof is based on the decomposition in and . To study the asymptotic behavior of , we use Lemma 2.8.7 in VW. We first fix a monotonic sequence $T= T(N)$ such that $T(N) \to \infty$ as $N\to \infty$, making our sample into triangular arrays. By Theorem 2.8.3, Example 2.5.4, and Example 2.3.4 in VW, the class $\mathcal{F}$ is Donsker and pre-Gaussian uniformly in $\{P_T\}$. Thus, we need to check the conditions $(2.8.5)$ and $(2.8.6)$ in VW. The condition $(2.8.6)$ in VW is immediate for the envelope function $F=1$ (constant). We check the condition $(2.8.5)$ in VW. Let $\rho_{P_T}$ and $\rho_{P_{0}}$ be the variance semi-metrics with respect to $P_T$ and $P_{0}$, respectively. Then, $$\begin{aligned} &\sup_{f,g \in \mathcal{F}} | \rho_{P_T}(f,g) - \rho_{P_{0}}(f,g) | \\ =& \sup_{f,g \in \mathcal{F}} | \sqrt{P_T((f-g)-P_T(f-g))^{2}} - \sqrt{P_{0}((f-g)-P_{0}(f-g))^{2}} | \\ =& \sup_{a,a' \in \mathbb{R}} | \sqrt{P_T(\mathbf{1}_{(-\infty, a]} - \mathbf{1}_{(-\infty, a']} - P_T(\mathbf{1}_{(-\infty, a]} -\mathbf{1}_{(-\infty, a']}))^{2}} \\ & - \sqrt{P_{0}(\mathbf{1}_{(-\infty, a]} - \mathbf{1}_{(-\infty, a']} - P_{0}(\mathbf{1}_{(-\infty, a]} -\mathbf{1}_{(-\infty, a']}))^{2}} | \\ \leq& \sup_{a,a' \in \mathbb{R}} | P_T(\mathbf{1}_{(-\infty, a]} - \mathbf{1}_{(-\infty, a']} - P_T(\mathbf{1}_{(-\infty, a]} -\mathbf{1}_{(-\infty, a']}))^{2}\\ & - P_{0}(\mathbf{1}_{(-\infty, a]} - \mathbf{1}_{(-\infty, a']} - P_{0}(\mathbf{1}_{(-\infty, a]} -\mathbf{1}_{(-\infty, a']}))^{2}|^{1/2},\end{aligned}$$ where the first inequality follows from the triangle inequality. Without loss of generality, we assume $a >a'$. Then, by simple algebra, $$\begin{aligned} & \sup_{f,g \in \mathcal{F}} | \rho_{P_T}(f,g) - \rho_{P_{0}}(f,g) | \\ \leq & \sup_{a, a' \in \mathbb{R}} \left| (P_T\mathbf{1}_{(-\infty,a]} - P_{0}\mathbf{1}_{(-\infty,a]}) - (P_T\mathbf{1}_{(-\infty,a]}\mathbf{1}_{(-\infty,a']} - P_{0}\mathbf{1}_{(-\infty,a]}\mathbf{1}_{(-\infty,a']}) \right. \\ & + (P_T\mathbf{1}_{(-\infty,a']} - P_{0}\mathbf{1}_{(-\infty,a']}) -((P_T\mathbf{1}_{(-\infty,a]})^2 - (P_{0}\mathbf{1}_{(-\infty,a]})^2) \\ & - ((P_T\mathbf{1}_{(-\infty,a']})^2 - (P_{0}\mathbf{1}_{(-\infty,a]'})^2) + 2(P_T\mathbf{1}_{(-\infty,a]}P_T\mathbf{1}_{(-\infty,a']} - P_T\mathbf{1}_{(-\infty,a]}P_{0}\mathbf{1}_{(-\infty,a']})\\ & \left. + \; 2(P_T\mathbf{1}_{(-\infty,a]}P_{0}\mathbf{1}_{(-\infty,a']} - P_{0}\mathbf{1}_{(-\infty,a]}P_{0}\mathbf{1}_{(-\infty,a']}) \right| ^{1/2}\\ \leq & \; 11 \sup_{a \in \mathbb{R}} \left|P_T\mathbf{1}_{(-\infty,a]} - P_{0}\mathbf{1}_{(-\infty,a]}\right|^{1/2}\\ \rightarrow & \; 0,\end{aligned}$$ where the last conclusion follows from Lemma 2.11 in @vanderVaart98, and $\hat \xi_i \stackrel{p}{\longrightarrow} \xi_i$, which follows from Corollary \[lem-second-mu\] or Lemma \[lem-gamma-moment\], or \[lem-rho-moment\] for $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$, respectively. Therefore, condition (2.8.5) in VW is satisfied. Therefore, by Lemma 2.8.7 in VW, we show that $$\begin{aligned} \mathbb{G}_{N,P_T} \leadsto \mathbb{G}_{P_{0}} \qquad \mbox{in} \quad\ell^{\infty}(\mathcal{F}). \label{eq-w-conv}\end{aligned}$$ Note that holds for all monotonic diagonal paths $T(N) \to \infty$ as $N \to \infty$. As in REMARKS (a) of @PhillipsMoon99, thus holds under double asymptotics $N,T \to \infty$. Next, we study the asymptotic behavior of : $\sqrt{N}(P_Tf-P_{0}f)$. Because the nonstochastic function sequence $P_Tf-P_{0}f$ is uniformly bounded in $f \in \mathcal{F}$, we should consider the convergence rate of $\sup_{f \in \mathcal{F}} |P_Tf - P_{0}f |$. Lemma \[lem-dis-diff-mu\], \[lem-dis-diff-gamma\], or \[lem-dis-diff-rho\] for $\hat \xi_i = \hat \mu_i$, $\hat \gamma_{k,i}$, or $\hat \rho_{k,i}$, respectively, shows that $\sup_{f \in \mathcal{F}} | P_Tf - P_{0}f | = O( T^{-2/(3+\epsilon)})$ for any $\epsilon \in (0, 1/3)$. Therefore, given $N^{3 + \epsilon}/T^{4} \to 0$ for some $\epsilon \in (0, 1/3)$, the desired result holds by Slutsky’s theorem. ### Proofs of Theorems \[thm-h\]-\[thm-dist-bootstrap\] These proofs are included in the supplement. ### Proof of Theorem \[thm-KS2\] We first observe that $$\begin{aligned} KS_2 = \left\| \sqrt{\frac{N_1N_2}{N_1+N_2}} (\mathbb{P}_{N_1,(1)} - P_{0,(1)}) - \sqrt{\frac{N_1N_2}{N_1+N_2}} (\mathbb{P}_{N_2,(2)} - P_{0,(2)} )+ \sqrt{\frac{N_1N_2}{N_1+N_2}}(P_{0,(1)} - P_{0,(2)}) \right\|_{\infty}.\end{aligned}$$ Note that, under Assumption \[as-ks\], $\sqrt{N_1}(\mathbb{P}_{N_1,(1)} - P_{0,(1)})$ and $\sqrt{N_2}(\mathbb{P}_{N_2,(2)} - P_{0,(2)})$ jointly converge in distribution to independent Brownian processes $\mathbb{G}_{P_{0,(1)}}$ and $\mathbb{G}_{P_{0,(2)}}$ given $N_1,T_1\to\infty$ with $N_1^{3+ \epsilon} /T_1^{4} \to 0$ and $N_2,T_2\to\infty$ with $N_2^{3 + \epsilon} /T_2^{4} \to 0$ for some $\epsilon \in (0, 1/3)$ by Theorem \[thm-fclt\]. Therefore, under $H_0:P_{0,(1)} = P_{0,(2)}$, $KS_2$ converges in distribution to $\| \sqrt{1-\lambda}\mathbb{G}_{P_{0,(1)}} - \sqrt{\lambda}\mathbb{G}_{P_{0,(2)}} \|_{\infty}$ by the continuous mapping theorem given $N_1 / (N_1 + N_2) \to \lambda \in (0,1)$. It is easy to see that the distribution of the limit random variable $\sqrt{1-\lambda}\mathbb{G}_{P_{0,(1)}} - \sqrt{\lambda}\mathbb{G}_{P_{0,(2)}}$ is identical to that of $\mathbb{G}_{P_{0,(1)}}$ under $H_0$. Thus, we have the desired result. Technical lemmas {#subsec-technical-lemma} ---------------- \[lem-mixing-var\] Let $\{\upsilon_t\}_{t=1}^{\infty}$ denote a stationary process taking values in $\mathbb{R}$ and let $\alpha(m)$ denote its $\alpha$-mixing coefficients. Suppose that $ E(|\upsilon_1| ^q) <\infty$ and $\sum_{m=1}^\infty \alpha(m)^{1-2/q} <\infty$ for some $q>2$. Then, we have $var \left( \sum_{t=1}^T \upsilon_t \right)\le C T$ with $C=12 (E(|\upsilon_1|^q))^{2/q} \sum_{m=0}^{\infty} \alpha (m)^{1-2/q}$. The proof is available in @GalvaoKato14 (the discussion after Theorem C.1). \[lem-yokoyama\] Let $\{\upsilon_t\}_{t=1}^{\infty}$ denote a strictly stationary $\alpha$-mixing process taking values in $\mathbb{R}$, and let $\alpha(m)$ denote its $\alpha$-mixing coefficients. Suppose that $E (\upsilon_t) =0$ and for some constants $\delta > 0$ and $r>2$, $E(|\upsilon_1 | ^{r+\delta}) <\infty$. If $\sum_{m=0}^{\infty} (m+1)^{r/2 -1} \alpha (m)^{\delta/ (r+\delta)} < \infty$, then there exists a constant $C$ independent of $T$ such that $E ( | \sum_{t=1}^T \upsilon_t |^r ) \le C T^{r/2}$. The proof is available in @Yokoyama1980. \[lem-moment-w\] Let $r$ be an even natural number. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], and \[as-w-moment-c\] hold for $r_m = r$ and $r_d =r$. Then, it holds that $E((\bar w_i)^r )\le C T^{ - r / 2 }$. The proof is included in the supplement. Because $\hat \mu_i - \mu_i = \bar y_i - \mu_i = \bar w_i $, we obtain the following result as a corollary. \[lem-second-mu\] Let $r$ be an even natural number. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], and \[as-w-moment-c\] hold for $r_m = r$ and $r_d =r$. Then we have $E( (\hat \mu_i -\mu_i)^{r}) = O(T^{-r/2})$. \[lem-mixing\] Let $r$ be an even natural number. Suppose that Assumptions \[as-basic\] and \[as-mixing-c\] hold for $r_m = r$. Then, $\{w_{it} w_{i,t-k}\}_{t=k+1}^{\infty}$ for a fixed $k$ given $\alpha_i$ is strictly stationary and $\alpha$-mixing and its mixing coefficients $\{\alpha_{k} (m|i)\}_{m=0}^{\infty}$ possess the following properties: there exists a sequence $\{\alpha_k (m)\}_{m=0}^{\infty}$ such that for any $i$ and $m$, $\alpha_k (m|i) \le \alpha_k (m)$ and $\sum_{m=0}^{\infty} (m+1)^{r/2-1} \alpha_k (m) ^{\delta / (r+\delta)} < \infty$ for some $\delta>0$. The proof is similar to the proof of Theorem 14.1 in @Davidson1994. Clearly, for any $i$ and any $0 \leq m < k$, $\alpha_k (m|i) \leq 1$, and that for any $i$ and any $m \geq k$, $\alpha_k (m|i) \le \alpha (m-k |i) \leq \alpha(m-k)$ by the definition of $\alpha$-mixing coefficients and Assumption \[as-mixing-c\]. Thus, we have $\sum_{m=0}^{\infty} (m+1)^{r/2-1} \alpha_k (m)^{\delta / (r+\delta)} \le \sum_{m=0}^{k-1} (m+1)^{r/2-1} + \sum_{m=k}^{\infty} (m+1)^{r/2-1} \alpha (m-k)^{\delta / (r+\delta)} < \infty$ under Assumption \[as-mixing-c\]. \[lem-moment-wk\] Let $r$ be an even natural number. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], and \[as-w-moment-c\] hold for $r_m = r$ and $r_d = 2r$. Then, it holds that $E((\sum_{t=k+1}^T (w_{it}w_{i,t-k} - \gamma_{k,i}))^r) \leq C T^{ r / 2 }$ for some constant $C$. In view of Lemma \[lem-mixing\], the lemma follows the same line as that for Lemma \[lem-moment-w\]. \[lem-gamma-moment\] Let $r$ be an even natural number. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], and \[as-w-moment-c\] hold for $r_m = 2r$ and $r_d=2r$. Then, we have $E((\hat\gamma_{k,i}-\gamma_{k,i})^{r}) = O(T^{-r/2})$. The proof is included in the supplement. \[lem-rho-moment\] Let $r$ be an even natural number. Suppose that Assumptions \[as-basic\], \[as-mixing-c\], and \[as-w-moment-c\] hold for $r_m = 2r$ and $r_d=2r$ and that $\gamma_{0,i} > \epsilon$ almost surely for some constant $\epsilon>0$. We have $E((\hat\rho_{k,i}-\rho_{k,i})^{r}) = O(T^{-r/2})$. We observe that $E(( \hat \rho_{k,i} - \rho_{k,i})^r ) = E(( \gamma_{0,i}^{-1} ( \hat \gamma_{k,i} - \gamma_{k,i} ) - \gamma_{0,i}^{-1} \hat \rho_{k,i} ( \hat \gamma_{0,i} - \gamma_{0,i} ) )^r)$. By Loéve’s $c_r$ inequality, we only need to examine the $r$-order moment of each term in parentheses on the right-hand side. We have $E( ( \gamma_{0,i}^{-1} ( \hat \gamma_{k,i} - \gamma_{k,i}) )^r ) \le M E((\hat\gamma_{k,i}-\gamma_{k,i})^{r})$ for some $M < \infty$ by the assumption that $\gamma_{0,i} >\epsilon$. Lemma \[lem-gamma-moment\] implies that $E((\hat\gamma_{k,i}-\gamma_{k,i})^{r}) = T^{-r/2}$. For the second term, it holds that $E ( ( \gamma_{0,i}^{-1} \hat \rho_{k,i} ( \hat \gamma_{0,i} - \gamma_{0,i}))^r)\le M E((\hat\gamma_{0,i}-\gamma_{0,i})^{r})$ for some $M < \infty$ by the assumption that $\gamma_{0,i} >\epsilon$ and the fact that $\vert \hat \rho_{k,i} \vert \le 1$. Lemma \[lem-gamma-moment\] implies that $E((\hat\gamma_{0,i}-\gamma_{0,i})^{r}) = T^{-r/2}$. We thus have the desired result. \[lem-dis-diff-mu\] Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-mu-con\] hold for $r_m = 4$ and $r_d = 4$. Let $P_T = P_T^{\hat \mu}$ and $P_0 = P_0^{\mu}$ be the probability measures of $\hat \mu_i$ and $\mu_i$, respectively. It holds that $\sup_{f \in \mathcal{F}} | P_T f - P_0 f | = O ( T^{-2 / (3 + 2 \epsilon)} )$ for any $\epsilon \in (0, 1/3)$. The proof is based on the comparison between the characteristic functions of $\hat \mu_j$ and $\mu_j$. In the proof, we use the index $j$ instead of the index $i$ because $i$ is reserved for the imaginary number. We introduce the sum of $\hat \mu_j$ and a Gaussian noise to guarantee that terms in the expansion of the characteristic function below are integrable. Consider $\hat \mu_j = \mu_j + \bar w_j$, $\tilde \mu_j \coloneqq \hat \mu_j + z = \mu_j + \bar w_j + z$, and $\check \mu_j \coloneqq \mu_j + z$ where $z \sim \mathcal{N}(0,\sigma^2)$ for some $\sigma^2 > 0$ and $z$ is independent of $(\alpha_j, \{ w_{jt} \}_{t=1}^T)$. Below we consider a situation where $\sigma^2 \to 0$ depending on $T \to \infty$. Let $P_T$, $\tilde P_T$, $\check P$, and $P_0$ be the probability measures of $\hat \mu_{j}$, $\tilde \mu_j$, $\check \mu_j$, and $\mu_{j}$, respectively. We observe that $$\begin{aligned} \label{eq:difference} \sup_{f \in \mathcal{F}} | P_T f - P_0 f | \le \sup_{f \in \mathcal{F}} | P_T f - \tilde P_T f | + \sup_{f \in \mathcal{F}} | \tilde P_T f - \check P f | + \sup_{f \in \mathcal{F}} | \check P f - P_0 f |. \end{aligned}$$ We examine each term on the right-hand side. For $f = \mathbf{1}_{(-\infty, a]}$, we write the CDFs $F_T(a) = P_T f = \Pr(\hat \mu_j \le a)$, $\tilde F_T(a) = \tilde P_T f = \Pr(\tilde \mu_j \le a)$, $\check F(a) = \check P f = \Pr(\check \mu_j \le a)$, and $F_0(a) = P_0 f = \Pr(\mu_j \le a)$. We first examine the first term in . For any $a \in \mathbb{R}$, we observe that $\tilde F_T (a ) - F_T (a) = E [ E( \mathbf{1} \{ \hat \mu_j + z \le a \}| z ) ] - F_T (a) = E [ F_T (a - z)] - F_T (a)$ because of the law of iterated expectations and the independence between $z$ and $\hat \mu_j$. We consider the third-order Taylor expansion of $F_T (a-z)$ around $z= 0$: $F_T ( a-z) = F_T ( a) - z F_T' (a) + z^2 F_T^{''} (a) / 2 - z^3 F_T^{'''} (\tilde a) / 3!$ where $\tilde a$ is between $a-z$ and $a$. Noting that $E(z) = 0$, $E(z^2) = \sigma^2$, and $E|z|^3 = O(\sigma^3)$, we obtain that $\left| \tilde F_T (a ) - F_T (a) \right| = \left| E [ F_T (a - z) ] - F_T (a) \right| = O(\sigma^2)$ uniformly over $a \in \mathbb{R}$ by Assumption \[as-mu-con\].c. Hence, we have $\sup_{f \in \mathcal{F}} | P_T f - \tilde P_T f | = O(\sigma^2)$. We then examine the third term in . The proof is the same as for the first term. Given $z$ is independent of $\mu_j$, we observe that $\check F (a ) - F_0 (a) = E [ F_0 (a - z)] - F_0 (a)$. The third-order Taylor expansion of $F_0 (a-z)$ around $z= 0$ is $F_0 ( a-z) = F_0 ( a) - z F_0' (a) + z^2 F_0^{''} (a) / 2 - z^3 F_0^{'''} (\tilde a) / 3!$. By Assumption \[as-mu-con\].b, we obtain that $\left| \check F (a ) - F_0 (a) \right| = \left| E [ F_0 (a - z) ] - F_0 (a) \right| = O(\sigma^2)$ uniformly over $a \in \mathbb{R}$. Hence, we have $\sup_{f \in \mathcal{F}} | \check P f - P_0 f | = O(\sigma^2)$. Next, we evaluate the second term in . To this end, we first expand the characteristic functions of $\check \mu_j$ and $\tilde \mu_j$. By the independence between $z$ and $(\alpha_j, \{w_{jt}\}_{t=1}^T)$, we have $$\begin{aligned} \psi_{\check \mu} (\zeta) &\coloneqq E[ \exp (i \zeta \check \mu_j )] = E[\exp (i \zeta z)] E[ \exp (i \zeta \mu_j )] = \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \psi_{\mu}(\zeta), \\ \psi_{\tilde \mu}(\zeta) &\coloneqq E[ \exp (i \zeta \tilde \mu_j )] = E[\exp (i \zeta z)] E[ \exp (i \zeta \hat \mu_j )] = \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \psi_{\hat \mu}(\zeta), \end{aligned}$$ where $\psi_{\mu}(\zeta) \coloneqq E[ \exp (i \zeta \mu_j) ]$ and $\psi_{\hat \mu}(\zeta) \coloneqq E[ \exp (i \zeta \hat \mu_j) ]$ are the characteristic functions of $\mu_j$ and $\hat \mu_j$. For the characteristic function of $\hat \mu_j$, we observe that $\psi_{\hat \mu}(\zeta) = E[ \exp (i \zeta \hat \mu_j ) ] = E [ \exp (i \zeta \mu_j ) \exp ( i \zeta \bar w_j ) ]$. By Taylor’s theorem, it holds that $\exp ( i \zeta \bar w_j ) = 1 + i \zeta \bar w_j - \zeta^2 (\bar w_j)^2 / 2 - i \zeta^3 ( \bar w_j )^3 \exp (i \zeta \tilde w_j) / 3!$ where $\tilde w_j$ is between 0 and $\bar w_j$.[^18] Therefore, it holds that $$\begin{aligned} E[ \exp (i \zeta \hat \mu_j ) ] =& E[ \exp (i \zeta \mu_j) ] + i \zeta E[ \bar w_j \exp (i \zeta \mu_j) ] \\ &- \frac{1}{2} \zeta^2 E[ (\bar w_j)^2 \exp (i \zeta \mu_j) ] - \frac{1}{3!} i \zeta^3 E[ ( \bar w_j )^3 \exp (i \zeta \tilde w_j) \exp (i \zeta \mu_j) ]\\ =& \psi_{\mu}(\zeta) - \frac{1}{2} \zeta^2 E[ (\bar w_j)^2 \exp (i \zeta \mu_j) ] - \frac{1}{3!} i \zeta^3 E[ ( \bar w_j )^3 \exp (i \zeta \tilde w_j) \exp (i \zeta \mu_j) ], \end{aligned}$$ where $E[ \bar w_j \exp (i \zeta \mu_j) ] =0$ follows from $E(\bar w_j | j) = 0$. Hence, it holds that $$\begin{aligned} \psi_{\tilde \mu}(\zeta) = \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \left( \psi_{\mu}(\zeta) - \frac{1}{2} \zeta^2 E[ (\bar w_j)^2 \exp (i \zeta \mu_j) ] - \frac{1}{3!} i \zeta^3 E[ ( \bar w_j )^3 \exp (i \zeta \tilde w_j) \exp (i \zeta \mu_j) ] \right). \end{aligned}$$ We use the inversion theorem ([@gil1951note] and [@wendel1961non]) to bound $\check F (a) - \tilde F_T(a)$ for any $a \in \mathbb{R}$. $$\label{eq:inverse} \begin{split} \check F_T(a) - \tilde F (a) =& \left( \frac{1}{2} + \frac{1}{2 \pi} \int_{-\infty}^{\infty} \frac{e^{i\zeta a} \psi_{\check \mu}(-\zeta) - e^{-i \zeta a} \psi_{\check \mu}(\zeta)}{i \zeta} d\zeta \right) \\ & - \left(\frac{1}{2} + \frac{1}{2 \pi} \int_{-\infty}^{\infty} \frac{e^{i\zeta a} \psi_{\tilde \mu}(-\zeta) - e^{-i \zeta a} \psi_{\tilde \mu}(\zeta)}{i \zeta} d\zeta \right) \\ =& \frac{1}{ \pi } \int_{-\infty}^{\infty} \frac{e^{-i \zeta a}}{i \zeta} [\psi_{\tilde \mu}(\zeta) - \psi_{\check \mu}(\zeta)] d\zeta \\ =& \frac{1}{\pi} \int_{-\infty}^{\infty} \frac{e^{-i \zeta a}}{i \zeta} \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \left( - \frac{1}{2} \zeta^2 E \left[ (\bar w_j)^2 \exp (i \zeta \mu_j) \right] \right) d \zeta \\ & + \frac{1}{\pi} \int_{-\infty}^{\infty} \frac{e^{-i \zeta a}}{i\zeta} \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \left( - \frac{1}{3!} i \zeta^3 E \left[ ( \bar w_j )^3 \exp (i \zeta \tilde w_j) \exp (i \zeta \mu_j) \right] \right) d \zeta. \end{split}$$ We examine each of the two terms. First, we consider the first term in the last line of . $$\label{eq:integral} \begin{split} & \frac{1}{\pi} \int_{-\infty}^{\infty} \frac{e^{-i \zeta a} }{i \zeta} \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \left( - \frac{1}{2} \zeta^2 E \left[ (\bar w_j)^2 \exp (i \zeta \mu_j) \right) \right] d \zeta \\ = & E \left[ (\bar w_j)^2 \frac{i}{2 \pi} \int_{-\infty}^{\infty} \exp(i \zeta (\mu_j-a )) \zeta \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) d\zeta \right], \end{split}$$ by Fubini’s theorem. Here, it holds that $$\begin{aligned} \frac{i}{2 \pi} \int_{-\infty}^{\infty} \exp(i \zeta (\mu_j-a )) \zeta \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) d\zeta = \frac{a - \mu_j}{\sqrt{2 \pi} \sigma^3} \exp \left( - \frac{(a - \mu_j)^2}{2 \sigma^2} \right). \end{aligned}$$ Thus, equation can be written as $$\begin{aligned} \text{Equation \eqref{eq:integral}} = E\left[ (\bar w_j)^2 \frac{a - \mu_j}{\sqrt{2 \pi} \sigma^3} \exp \left( - \frac{(a - \mu_j)^2}{2 \sigma^2} \right) \right]. \end{aligned}$$ To proceed, we define the shorthand notation $Z_j \coloneqq a - \mu_j$. We consider any nonrandom $A_\sigma > 0$ that satisfies $A _\sigma \to 0$ as $\sigma \to 0$. We then have $$\label{eq:splitA} \begin{split} \left| E\left[ (\bar w_j)^2 \frac{Z_j}{\sigma^3} \exp \left( - \frac{Z_j^2}{2 \sigma^2} \right) \right] \right| \le & \frac{1}{\sigma^3}E\left[ (\bar w_j)^2 |Z_j| \exp \left( - \frac{Z_j^2}{2 \sigma^2} \right) \right] \\ =& \frac{1}{\sigma^3}E\left[ (\bar w_j)^2 |Z_j| \exp \left( - \frac{Z_j^2}{2 \sigma^2} \right) \mathbf{1}(|Z_j| \le A_{\sigma}) \right] \\ & + \frac{1}{\sigma^3}E\left[ (\bar w_j)^2 |Z_j| \exp \left( - \frac{Z_j^2}{2 \sigma^2} \right) \mathbf{1}(|Z_j| > A_{\sigma}) \right]. \\ \end{split}$$ For the first term in the last line of , it holds that $$\begin{aligned} \frac{1}{\sigma^3} E\left[ (\bar w_j)^2 |Z_j| \exp \left( - \frac{Z_j^2}{2 \sigma^2} \right) \mathbf{1}(|Z_j| \le A_{\sigma}) \right] \le& \frac{A_{\sigma}}{\sigma^3} \exp(0) E \left[ (\bar w_j)^2 \mathbf{1}(|Z_j| \le A_{\sigma}) \right] \\ =& \frac{A_{\sigma}}{\sigma^3} E \left[ E\left[ (\bar w_j)^2 | \mu_j \right] \mathbf{1}(|Z_j| \le A_{\sigma}) \right] \\ \le& \frac{A_{\sigma}}{\sigma^3} \frac{M}{T} E \left[ \mathbf{1}(|Z_j| \le A_{\sigma}) \right] \\ =& O\left( \frac{A_{\sigma}^{2}}{\sigma^3 T} \right), \end{aligned}$$ where the second inequality follows by Assumption \[as-mu-con\].d and the last equality holds by $E[\mathbf{1}(|Z_j| \le A_{\sigma})] = \Pr(|Z_j| \le A_{\sigma}) = \int_0^{A_{\sigma}} f_Z(z) dz = O(A_{\sigma})$ based on the bounded density of $\mu_j$. For the second term in the last line of , we have $$\begin{aligned} & \frac{1}{\sigma^3}E\left[ (\bar w_j)^2 |Z_j| \exp \left( - \frac{Z_j^2}{2 \sigma^2} \right) \mathbf{1}(|Z_j| > A_{\sigma}) \right] \\ & \le \frac{1}{\sigma^3} \sqrt{E\left[ (\bar w_j)^4 \right]} \sqrt{E\left[ Z_j^2 \exp^2 \left( - \frac{Z_j^2}{2 \sigma^2} \right) \mathbf{1}(|Z_j| > A_{\sigma}) \right]}\\ & \le \frac{1}{\sigma^3} \sqrt{E\left[ (\bar w_j)^4 \right]} \sqrt{E(Z_j^2) \exp^2 \left( - \frac{A_{\sigma}^2}{2 \sigma^2} \right) }\\ &= O\left( \frac{1}{\sigma^3} \right) \cdot O\left( \frac{1}{T} \right) \cdot O\left( \exp \left( - \frac{A_{\sigma}^2}{\sigma^2} \right) \right) = O\left( \frac{1}{\sigma^3 T} \exp \left( - \frac{A_{\sigma}^2}{\sigma^2} \right) \right), \end{aligned}$$ where we used the Cauchy–Schwarz inequality, Lemma \[lem-moment-w\], and $E(Z_j^2) = O(1)$. Note that, if $A_{\sigma} = \sigma^{1 - \epsilon'}$ for any $0< \epsilon' < 1$, it holds that $\exp(-A_{\sigma}^2 / \sigma^2) / A_{\sigma}^{2} = o(1)$ as $\sigma \to 0$. Therefore, by setting $A_{\sigma} = \sigma^{1 - \epsilon'}$ for any $0< \epsilon' < 1$, we obtain that $$\begin{aligned} \text{Equation \eqref{eq:integral}} = O\left( \frac{A_{\sigma}^{2}}{\sigma^3 T} \right) + O\left( \frac{1}{\sigma^3 T} \exp \left( - \frac{A_{\sigma}^2}{\sigma^2} \right) \right) = O\left( \frac{A_{\sigma}^{2}}{\sigma^3 T} \right) = O\left( \frac{1}{\sigma^{1 + 2 \epsilon'} T} \right). \end{aligned}$$ Next, we consider the second term in the last line of . $$\begin{aligned} & \frac{1}{\pi} \int_{-\infty}^{\infty} \frac{e^{-i \zeta a} }{i\zeta} \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) \left( - \frac{1}{3!} i \zeta^3 E\left[ ( \bar w_j )^3 \exp (i \zeta \tilde w_j) \exp (i \zeta \mu_j) \right] \right) d \zeta \\ & = -\frac{1}{ 3! \pi} E \left[ (\bar w_j)^3 \int_{-\infty}^{\infty} \exp(i \zeta (\mu_j + \tilde w_j -a )) \zeta^2 \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) d\zeta \right]. \end{aligned}$$ Note that $|\exp(i \zeta (\mu_j + \tilde w_j - a))| \le 1$ and $E |\bar w_j|^3 = O(T^{-3/2})$ by Lemma \[lem-moment-w\] and Hölder’s inequality. Given $(2 \pi)^{-1/2} \sigma \exp (- \sigma^2 \zeta^2 /2) $ is the density function of $\mathcal{N}(0, 1/ \sigma^2)$, we thus obtain that $$\begin{aligned} \left| \frac{1}{ 3! \pi} E \left[ (\bar w_j)^3 \int_{-\infty}^{\infty} \exp(i \zeta (\mu_j + \tilde w_j -a )) \zeta^2 \exp \left( -\frac{1}{2} \sigma^2 \zeta^2 \right) d\zeta \right] \right| = O \left( \frac{1}{\sigma^3 T^{3/2} }\right). \end{aligned}$$ In sum, we have shown the order of the second term in : $$\begin{aligned} \left| \tilde F_T(a) - \check F (a) \right| = O\left( \frac{1}{\sigma^{1 + 2 \epsilon'} T} + \frac{1}{\sigma^3 T^{3/2} }\right), \end{aligned}$$ uniformly over $a \in \mathbb{R}$ for any $0< \epsilon' < 1$. Based on the above results, we have shown that $$\begin{aligned} \sup_{f \in \mathcal{F}} \left| P_T f - P_0 f \right| = O\left( \sigma^2 + \frac{1}{\sigma^{1 + 2 \epsilon'} T} + \frac{1}{\sigma^3 T^{3/2} }\right). \end{aligned}$$ When we set $\sigma = 1 / T^{1 / (3 + 2\epsilon')}$ with any $0 < \epsilon' < 1 / 6$, we obtain the following convergence result: $$\begin{aligned} \sup_{f \in \mathcal{F}} \left| P_T f - P_0 f \right| = O\left( \sigma^2 + \frac{1}{\sigma^{1 + 2 \epsilon'} T} \right) = O \left( \frac{1}{T^{2 / (3 + \epsilon)}} \right), \end{aligned}$$ for any $0 < \epsilon = 2\epsilon' < 1/3$. \[lem-dis-diff-gamma\] Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-gamma-con\] hold for $r_m = 8$ and $r_d = 8$. Let $P_T = P_T^{\hat \gamma_k}$ and $P_0 = P_0^{\gamma_k}$ be the probability measures of $\hat \gamma_{k,i}$ and $\gamma_{k,i}$, respectively. It holds that $\sup_{f \in \mathcal{F}} | P_T f - P_0 f | = O ( T^{-2 / (3 + \epsilon)} )$ for any $\epsilon \in (0, 1/3)$. The proof is included in the supplement. \[lem-dis-diff-rho\] Suppose that Assumptions \[as-basic\], \[as-mixing-c\], \[as-w-moment-c\], and \[as-rho-con\] hold for $r_m = 8$ and $r_d = 8$. Let $P_T = P_T^{\hat \rho_k}$ and $P_0 = P_0^{\rho_k}$ be the probability measures of $\hat \rho_{k,i}$ and $\rho_{k,i}$, respectively. It holds that $\sup_{f \in \mathcal{F}} | P_T f - P_0 f | = O ( T^{-2 / (3 + \epsilon)})$ for any $\epsilon \in (0, 1/3)$. The proof is included in the supplement. Acknowledgments {#acknowledgments .unnumbered} =============== The authors greatly appreciate the assistance of Yuya Sasaki and Mototsugu Shintani in providing the panel data for Chilean firms and US prices, respectively, and for their helpful discussions. The authors would also like to thank the editor (Oliver Linton) and the associate editor, three anonymous referees, Debopam Bhattacharya, Richard Blundell, Maurice Bun, Yoosoon Chang, Andrew Chesher, Yasunori Fujikoshi, Kazuhiko Hayakawa, Toshio Honda, Hidehiko Ichimura, Koen Jochmans, Artūras Juodis, Hiroaki Kaido, Hiroyuki Kasahara, Kengo Kato, Toru Kitagawa, Yuichi Kitamura, Frank Kleibergen, Roger Koenker, Eiji Kurozumi, Simon Lee, Alexei Onatski, Taisuke Otsu, Hashem Pesaran, Katsumi Shimotsu, Richard Smith, Mans Söderbom, Liangjun Su, Martin Weidner, Yoon-Jae Whang, Takashi Yamagata, Yohei Yamamoto, and seminar participants at various institutes for their helpful comments and useful discussions. Part of this research was conducted while Okui was at Vrije Universiteit Amsterdam and Kyoto University and while Yanagi was at Hitotsubashi University. All remaining errors are ours. Okui acknowledges financial support from the Japan Society for the Promotion of Science (JSPS) under KAKENHI Grant Nos. 22330067, 25780151, 25285067, 15H03329, and 16K03598. Yanagi recognizes the financial support of Grant-in-Aid for JSPS Fellows No. 252035 and KAKENHI Grant Nos. 15H06214 and 17K13715. [^1]: NYU Shanghai, 1555 Century Avenue, Pudong, Shanghai, China, 200122; and Department of Economics, University of Gothenburg, P.O. Box 640, SE-405 30 Gothenburg, Sweden. Tel: +86-21-2059-6157. Email: <[email protected]> [^2]: Graduate School of Economics, Kyoto University, Yoshida Honmachi, Sakyo, Kyoto, 606-8501, Japan. Email: <[email protected]> [^3]: First Version: February 2014. This paper was previously circulated under the title “Dynamic Panel Data Analysis when the Dynamics are Heterogeneous.” [^4]: For example, several researchers have pointed out the heterogeneous dynamics of the income process. These include @MeghirPistaferri04 and @Hospido2012, which both suggest the importance of heterogeneity in income process volatility, and @BotosaruSasaki2018, which develops a nonparametric procedure to estimate the volatility function in the permanent–transitory model of income dynamics. [^5]: The proposed procedures are readily available via an [R]{} package from the authors’ websites. [^6]: See @Arellano03b and @Baltagi08 for excellent reviews of the existing studies on dynamic panel data analyses. [^7]: An inspection of their proof leads us to surmise that their results may hold more generally as long as we assume that the distribution of the standardized quantity is homogeneous (that is, the estimated quantities satisfy a location–scale assumption). In our setting, heterogeneity can appear in more general ways and this generality is important because we are interested in heterogeneous dynamics. Indeed, we suspect that it is very difficult, if not impossible, to arrive at a setting in which all means, autocovariances, and autocorrelations exhibit heterogeneity and satisfy a location–scale assumption simultaneously. [^8]: There are also studies, such as @ChernozhukovFernandezValLuo2018, that aim to investigate the heterogeneity caused by observable covariates. However, these methods do not use the panel feature of the data and are distinct from the literature to which the present paper belongs. [^9]: If $T$ is odd, we define $\bar S = (\hat S^{(1,1)} + \hat S^{(2,1)} + \hat S^{(1,2)} + \hat S^{(2,2)} )/4$ as in @DhaeneJochmans15 [page 9], where $\hat S^{(1,1)}$, $\hat S^{(2,1)}$, $\hat S^{(1,2)}$, and $\hat S^{(2,2)}$ are the estimators of $S$ computed using $\{\{ y_{it} \}_{t=1}^{\lceil T/2 \rceil}\}_{i=1}^{N}$, $\{\{ y_{it} \}_{t=\lceil T/2 \rceil+1}^{T}\}_{i=1}^{N}$, $\{\{ y_{it} \}_{t=1}^{\lfloor T/2 \rfloor}\}_{i=1}^{N}$, and $\{\{ y_{it} \}_{t=\lfloor T/2 \rfloor +1}^{T}\}_{i=1}^{N}$, respectively. Here, $\lceil \cdot \rceil$ and $\lfloor \cdot \rfloor$ are the ceiling and floor functions, respectively. We note that the asymptotic properties of the HPJ estimator for odd $T$ are the same as those for even $T$. [^10]: If bootstrap is used to approximate the distribution of $\hat S-S$ not $\hat S^H-S$, then we just need to resample from $\{ \hat \theta_1,\dots, \hat \theta_N \}$. If inference is based on the TOJ estimator, estimates of $\theta_i $ from other subpanels are also required. [^11]: More precisely, we consider the case in which $T=T(N)$, where $T(N)$ is increasing in $N$ and $T(N) \to \infty$ as $N \to \infty$, but the exact form of the function $T(N)$ is left unspecified, except the condition imposed in each theorem. Note that analyses under sequential asymptotics where $N \to \infty$ after $T \to \infty$ ignore estimation errors in $\hat \mu_i$, $\hat \gamma_{k,i}$, and $\hat \rho_{k,i}$, and they fail to capture the bias. We do not consider the sequential asymptotics where $T \to \infty$ after $N \to \infty$. [^12]: Discrete heterogeneity is considered in, for example, [@BonhommeManresa15] and [@SuShiPhillips14] for linear panel data analyses. [^13]: We might consider testing homogeneity in a formal manner by extending the testing procedures in @pesaran2008testing to our model-free context. The construction of test statistics and the derivation of their asymptotic distributions for such extensions are nontrivial tasks, and this topic is left for future work. [^14]: In Corollary \[cor-delta\], we can change the Hadamard differentiability of $\phi:D(\mathcal{F}) \subset \ell^{\infty}(\mathcal{F}) \to \mathbb{E}$ to the Hadamard differentiability of $\phi: \ell^{\infty}(\mathcal{F}) \subset D(\mathbb{\bar R}) \to \mathbb{R}$ tangentially to a set of continuous functions in $D(\mathbb{\bar R})$, where $D(\mathbb{\bar R})$ is the Banach space of all càdlàg functions $z:\mathbb{\bar {R}} \to \mathbb{R}$ on $\mathbb{\bar R}$ equipped with the uniform norm. See Lemma 3.9.20 and Example 3.9.21 in @vanderVaartWellner96 for details. [^15]: While a uniform validity of the cross-sectional bootstrap for the distribution estimator would be desirable, this investigation is challenging because it requires new empirical process techniques. For now, we leave it as an interesting future research topic. [^16]: Mototsugu Shintani kindly provided us with the dataset ready for analysis. [^17]: While the original data source contains price information for more items in more cities, we restrict the observations to obtain a balanced panel data set, as in [@CruciniShintaniTsuruga15]. [^18]: Strictly speaking, this expansion may not hold as the mean value expression of the remainder term for Taylor’s theorem for complex functions may not exist. However, as $\exp(i \zeta \bar w_j) = \sin(\zeta \bar w_j) + i \cos(\zeta \bar w_j)$, applying Taylor’s theorem for real functions to $\cos:\mathbb{R} \to [-1, 1]$ and $\sin: \mathbb{R} \to [-1, 1]$ leads to $$\begin{aligned} \exp ( i \zeta \bar w_j ) = 1 + i \zeta \bar w_j - \frac{1}{2} \zeta^2 (\bar w_j)^2 - \frac{1}{6} \zeta^3 ( \bar w_j )^3 \big( \cos(c_1) - i \sin(c_2) \big), \end{aligned}$$ where $c_1$ and $c_2$ are located between 0 and $\zeta \bar w_j$ but $c_1 \neq c_2$ in general. Given $\cos(\cdot)$ and $\sin(\cdot)$ are bounded functions, we can obtain the same result in the main body based on this observation.
{ "pile_set_name": "ArXiv" }